diff options
Diffstat (limited to 'src/jumper/SkJumper_generated_win.S')
| -rw-r--r-- | src/jumper/SkJumper_generated_win.S | 35222 |
1 files changed, 17571 insertions, 17651 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index b230d46b3a..f65f7c253e 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -99,14 +99,14 @@ _sk_seed_shader_hsw LABEL PROC DB 197,249,110,194 ; vmovd %edx,%xmm0 DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 197,252,88,5,139,200,2,0 ; vaddps 0x2c88b(%rip),%ymm0,%ymm0 # 2c9e0 <_sk_overlay_sse2_8bit+0x488> + DB 197,252,88,5,139,199,2,0 ; vaddps 0x2c78b(%rip),%ymm0,%ymm0 # 2c8e0 <_sk_overlay_sse2_8bit+0x431> DB 197,249,110,201 ; vmovd %ecx,%xmm1 DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,61,209,2,0 ; vbroadcastss 0x2d13d(%rip),%ymm2 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,21,29,208,2,0 ; vbroadcastss 0x2d01d(%rip),%ymm2 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,50,209,2,0 ; vbroadcastss 0x2d132(%rip),%ymm2 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,21,18,208,2,0 ; vbroadcastss 0x2d012(%rip),%ymm2 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 @@ -119,33 +119,33 @@ _sk_dither_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,121,110,194 ; vmovd %edx,%xmm8 DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 - DB 197,61,254,5,93,200,2,0 ; vpaddd 0x2c85d(%rip),%ymm8,%ymm8 # 2ca00 <_sk_overlay_sse2_8bit+0x4a8> + DB 197,61,254,5,93,199,2,0 ; vpaddd 0x2c75d(%rip),%ymm8,%ymm8 # 2c900 <_sk_overlay_sse2_8bit+0x451> DB 197,121,110,201 ; vmovd %ecx,%xmm9 DB 196,66,125,88,201 ; vpbroadcastd %xmm9,%ymm9 DB 196,65,53,239,200 ; vpxor %ymm8,%ymm9,%ymm9 - DB 196,98,125,88,21,246,208,2,0 ; vpbroadcastd 0x2d0f6(%rip),%ymm10 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> + DB 196,98,125,88,21,214,207,2,0 ; vpbroadcastd 0x2cfd6(%rip),%ymm10 # 2d190 <_sk_overlay_sse2_8bit+0xce1> DB 196,65,53,219,218 ; vpand %ymm10,%ymm9,%ymm11 DB 196,193,37,114,243,5 ; vpslld $0x5,%ymm11,%ymm11 DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,193,45,114,242,4 ; vpslld $0x4,%ymm10,%ymm10 - DB 196,98,125,88,37,219,208,2,0 ; vpbroadcastd 0x2d0db(%rip),%ymm12 # 2d2b4 <_sk_overlay_sse2_8bit+0xd5c> - DB 196,98,125,88,45,214,208,2,0 ; vpbroadcastd 0x2d0d6(%rip),%ymm13 # 2d2b8 <_sk_overlay_sse2_8bit+0xd60> + DB 196,98,125,88,37,187,207,2,0 ; vpbroadcastd 0x2cfbb(%rip),%ymm12 # 2d194 <_sk_overlay_sse2_8bit+0xce5> + DB 196,98,125,88,45,182,207,2,0 ; vpbroadcastd 0x2cfb6(%rip),%ymm13 # 2d198 <_sk_overlay_sse2_8bit+0xce9> DB 196,65,53,219,245 ; vpand %ymm13,%ymm9,%ymm14 DB 196,193,13,114,246,2 ; vpslld $0x2,%ymm14,%ymm14 + DB 196,65,37,235,222 ; vpor %ymm14,%ymm11,%ymm11 DB 196,65,61,219,237 ; vpand %ymm13,%ymm8,%ymm13 DB 196,65,21,254,237 ; vpaddd %ymm13,%ymm13,%ymm13 + DB 196,65,21,235,210 ; vpor %ymm10,%ymm13,%ymm10 DB 196,65,53,219,204 ; vpand %ymm12,%ymm9,%ymm9 DB 196,193,53,114,209,1 ; vpsrld $0x1,%ymm9,%ymm9 DB 196,65,61,219,196 ; vpand %ymm12,%ymm8,%ymm8 DB 196,193,61,114,208,2 ; vpsrld $0x2,%ymm8,%ymm8 - DB 196,65,21,235,210 ; vpor %ymm10,%ymm13,%ymm10 DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 - DB 196,65,37,235,214 ; vpor %ymm14,%ymm11,%ymm10 - DB 196,65,61,235,194 ; vpor %ymm10,%ymm8,%ymm8 + DB 196,65,61,235,195 ; vpor %ymm11,%ymm8,%ymm8 DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,136,208,2,0 ; vbroadcastss 0x2d088(%rip),%ymm9 # 2d2bc <_sk_overlay_sse2_8bit+0xd64> - DB 196,98,125,24,21,131,208,2,0 ; vbroadcastss 0x2d083(%rip),%ymm10 # 2d2c0 <_sk_overlay_sse2_8bit+0xd68> + DB 196,98,125,24,13,104,207,2,0 ; vbroadcastss 0x2cf68(%rip),%ymm9 # 2d19c <_sk_overlay_sse2_8bit+0xced> + DB 196,98,125,24,21,99,207,2,0 ; vbroadcastss 0x2cf63(%rip),%ymm10 # 2d1a0 <_sk_overlay_sse2_8bit+0xcf1> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 DB 196,98,125,24,0 ; vbroadcastss (%rax),%ymm8 DB 196,65,44,89,192 ; vmulps %ymm8,%ymm10,%ymm8 @@ -175,7 +175,7 @@ _sk_uniform_color_hsw LABEL PROC PUBLIC _sk_black_color_hsw _sk_black_color_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,11,208,2,0 ; vbroadcastss 0x2d00b(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,235,206,2,0 ; vbroadcastss 0x2ceeb(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 @@ -184,7 +184,7 @@ _sk_black_color_hsw LABEL PROC PUBLIC _sk_white_color_hsw _sk_white_color_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,242,207,2,0 ; vbroadcastss 0x2cff2(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,210,206,2,0 ; vbroadcastss 0x2ced2(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 @@ -222,7 +222,7 @@ _sk_clear_hsw LABEL PROC PUBLIC _sk_srcatop_hsw _sk_srcatop_hsw LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,145,207,2,0 ; vbroadcastss 0x2cf91(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,113,206,2,0 ; vbroadcastss 0x2ce71(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,226,61,184,196 ; vfmadd231ps %ymm4,%ymm8,%ymm0 DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 @@ -236,7 +236,7 @@ _sk_srcatop_hsw LABEL PROC PUBLIC _sk_dstatop_hsw _sk_dstatop_hsw LABEL PROC - DB 196,98,125,24,5,96,207,2,0 ; vbroadcastss 0x2cf60(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,64,206,2,0 ; vbroadcastss 0x2ce40(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,226,101,184,196 ; vfmadd231ps %ymm4,%ymm3,%ymm0 @@ -269,7 +269,7 @@ _sk_dstin_hsw LABEL PROC PUBLIC _sk_srcout_hsw _sk_srcout_hsw LABEL PROC - DB 196,98,125,24,5,3,207,2,0 ; vbroadcastss 0x2cf03(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,227,205,2,0 ; vbroadcastss 0x2cde3(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -280,7 +280,7 @@ _sk_srcout_hsw LABEL PROC PUBLIC _sk_dstout_hsw _sk_dstout_hsw LABEL PROC - DB 196,226,125,24,5,226,206,2,0 ; vbroadcastss 0x2cee2(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,194,205,2,0 ; vbroadcastss 0x2cdc2(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3 DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0 DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1 @@ -291,7 +291,7 @@ _sk_dstout_hsw LABEL PROC PUBLIC _sk_srcover_hsw _sk_srcover_hsw LABEL PROC - DB 196,98,125,24,5,193,206,2,0 ; vbroadcastss 0x2cec1(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,161,205,2,0 ; vbroadcastss 0x2cda1(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0 DB 196,194,85,184,200 ; vfmadd231ps %ymm8,%ymm5,%ymm1 @@ -302,7 +302,7 @@ _sk_srcover_hsw LABEL PROC PUBLIC _sk_dstover_hsw _sk_dstover_hsw LABEL PROC - DB 196,98,125,24,5,156,206,2,0 ; vbroadcastss 0x2ce9c(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,124,205,2,0 ; vbroadcastss 0x2cd7c(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 DB 196,226,61,168,205 ; vfmadd213ps %ymm5,%ymm8,%ymm1 @@ -322,7 +322,7 @@ _sk_modulate_hsw LABEL PROC PUBLIC _sk_multiply_hsw _sk_multiply_hsw LABEL PROC - DB 196,98,125,24,5,99,206,2,0 ; vbroadcastss 0x2ce63(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,67,205,2,0 ; vbroadcastss 0x2cd43(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -343,7 +343,7 @@ _sk_multiply_hsw LABEL PROC PUBLIC _sk_plus__hsw _sk_plus__hsw LABEL PROC DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0 - DB 196,98,125,24,5,18,206,2,0 ; vbroadcastss 0x2ce12(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,242,204,2,0 ; vbroadcastss 0x2ccf2(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 197,244,88,205 ; vaddps %ymm5,%ymm1,%ymm1 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 @@ -369,7 +369,7 @@ _sk_screen_hsw LABEL PROC PUBLIC _sk_xor__hsw _sk_xor__hsw LABEL PROC - DB 196,98,125,24,5,189,205,2,0 ; vbroadcastss 0x2cdbd(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,157,204,2,0 ; vbroadcastss 0x2cc9d(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -401,7 +401,7 @@ _sk_darken_hsw LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,65,205,2,0 ; vbroadcastss 0x2cd41(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,33,204,2,0 ; vbroadcastss 0x2cc21(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -424,7 +424,7 @@ _sk_lighten_hsw LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,236,204,2,0 ; vbroadcastss 0x2ccec(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,204,203,2,0 ; vbroadcastss 0x2cbcc(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -450,7 +450,7 @@ _sk_difference_hsw LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,139,204,2,0 ; vbroadcastss 0x2cc8b(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,107,203,2,0 ; vbroadcastss 0x2cb6b(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -470,7 +470,7 @@ _sk_exclusion_hsw LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,69,204,2,0 ; vbroadcastss 0x2cc45(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,37,203,2,0 ; vbroadcastss 0x2cb25(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 196,194,69,184,216 ; vfmadd231ps %ymm8,%ymm7,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -478,7 +478,7 @@ _sk_exclusion_hsw LABEL PROC PUBLIC _sk_colorburn_hsw _sk_colorburn_hsw LABEL PROC - DB 196,98,125,24,5,47,204,2,0 ; vbroadcastss 0x2cc2f(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,15,203,2,0 ; vbroadcastss 0x2cb0f(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 @@ -534,7 +534,7 @@ _sk_colorburn_hsw LABEL PROC PUBLIC _sk_colordodge_hsw _sk_colordodge_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,63,203,2,0 ; vbroadcastss 0x2cb3f(%rip),%ymm9 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,13,31,202,2,0 ; vbroadcastss 0x2ca1f(%rip),%ymm9 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9 @@ -585,7 +585,7 @@ _sk_colordodge_hsw LABEL PROC PUBLIC _sk_hardlight_hsw _sk_hardlight_hsw LABEL PROC - DB 196,98,125,24,5,89,202,2,0 ; vbroadcastss 0x2ca59(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,57,201,2,0 ; vbroadcastss 0x2c939(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -634,7 +634,7 @@ _sk_hardlight_hsw LABEL PROC PUBLIC _sk_overlay_hsw _sk_overlay_hsw LABEL PROC - DB 196,98,125,24,5,141,201,2,0 ; vbroadcastss 0x2c98d(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,109,200,2,0 ; vbroadcastss 0x2c86d(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -694,10 +694,10 @@ _sk_softlight_hsw LABEL PROC DB 196,65,20,88,197 ; vaddps %ymm13,%ymm13,%ymm8 DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,66,61,168,192 ; vfmadd213ps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,29,164,200,2,0 ; vbroadcastss 0x2c8a4(%rip),%ymm11 # 2d2c4 <_sk_overlay_sse2_8bit+0xd6c> + DB 196,98,125,24,29,132,199,2,0 ; vbroadcastss 0x2c784(%rip),%ymm11 # 2d1a4 <_sk_overlay_sse2_8bit+0xcf5> DB 196,65,20,88,227 ; vaddps %ymm11,%ymm13,%ymm12 DB 196,65,28,89,192 ; vmulps %ymm8,%ymm12,%ymm8 - DB 196,98,125,24,37,149,200,2,0 ; vbroadcastss 0x2c895(%rip),%ymm12 # 2d2c8 <_sk_overlay_sse2_8bit+0xd70> + DB 196,98,125,24,37,117,199,2,0 ; vbroadcastss 0x2c775(%rip),%ymm12 # 2d1a8 <_sk_overlay_sse2_8bit+0xcf9> DB 196,66,21,184,196 ; vfmadd231ps %ymm12,%ymm13,%ymm8 DB 196,65,124,82,245 ; vrsqrtps %ymm13,%ymm14 DB 196,65,124,83,246 ; vrcpps %ymm14,%ymm14 @@ -707,7 +707,7 @@ _sk_softlight_hsw LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,68,200,2,0 ; vbroadcastss 0x2c844(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,36,199,2,0 ; vbroadcastss 0x2c724(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,65,60,92,237 ; vsubps %ymm13,%ymm8,%ymm13 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,98,125,168,235 ; vfmadd213ps %ymm3,%ymm0,%ymm13 @@ -820,11 +820,11 @@ _sk_hue_hsw LABEL PROC DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10 DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10 DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - DB 196,98,125,24,53,87,198,2,0 ; vbroadcastss 0x2c657(%rip),%ymm14 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> - DB 196,98,125,24,61,82,198,2,0 ; vbroadcastss 0x2c652(%rip),%ymm15 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,53,55,197,2,0 ; vbroadcastss 0x2c537(%rip),%ymm14 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> + DB 196,98,125,24,61,50,197,2,0 ; vbroadcastss 0x2c532(%rip),%ymm15 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13 - DB 196,226,125,24,5,67,198,2,0 ; vbroadcastss 0x2c643(%rip),%ymm0 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,226,125,24,5,35,197,2,0 ; vbroadcastss 0x2c523(%rip),%ymm0 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13 DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10 DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -879,7 +879,7 @@ _sk_hue_hsw LABEL PROC DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9 DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,13,4,197,2,0 ; vbroadcastss 0x2c504(%rip),%ymm1 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,13,228,195,2,0 ; vbroadcastss 0x2c3e4(%rip),%ymm1 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11 @@ -933,11 +933,11 @@ _sk_saturation_hsw LABEL PROC DB 196,65,28,89,210 ; vmulps %ymm10,%ymm12,%ymm10 DB 196,65,44,94,214 ; vdivps %ymm14,%ymm10,%ymm10 DB 196,67,45,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm10,%ymm12 - DB 196,98,125,24,53,49,196,2,0 ; vbroadcastss 0x2c431(%rip),%ymm14 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> - DB 196,98,125,24,61,44,196,2,0 ; vbroadcastss 0x2c42c(%rip),%ymm15 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,53,17,195,2,0 ; vbroadcastss 0x2c311(%rip),%ymm14 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> + DB 196,98,125,24,61,12,195,2,0 ; vbroadcastss 0x2c30c(%rip),%ymm15 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,66,93,184,238 ; vfmadd231ps %ymm14,%ymm4,%ymm13 - DB 196,226,125,24,5,29,196,2,0 ; vbroadcastss 0x2c41d(%rip),%ymm0 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,226,125,24,5,253,194,2,0 ; vbroadcastss 0x2c2fd(%rip),%ymm0 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 196,98,77,184,232 ; vfmadd231ps %ymm0,%ymm6,%ymm13 DB 196,65,116,89,215 ; vmulps %ymm15,%ymm1,%ymm10 DB 196,66,53,184,214 ; vfmadd231ps %ymm14,%ymm9,%ymm10 @@ -992,7 +992,7 @@ _sk_saturation_hsw LABEL PROC DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 DB 196,65,36,95,200 ; vmaxps %ymm8,%ymm11,%ymm9 DB 196,65,116,95,192 ; vmaxps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,13,222,194,2,0 ; vbroadcastss 0x2c2de(%rip),%ymm1 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,13,190,193,2,0 ; vbroadcastss 0x2c1be(%rip),%ymm1 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,116,92,215 ; vsubps %ymm7,%ymm1,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,116,92,219 ; vsubps %ymm3,%ymm1,%ymm11 @@ -1020,11 +1020,11 @@ _sk_color_hsw LABEL PROC DB 197,108,89,199 ; vmulps %ymm7,%ymm2,%ymm8 DB 197,116,89,215 ; vmulps %ymm7,%ymm1,%ymm10 DB 197,52,89,223 ; vmulps %ymm7,%ymm9,%ymm11 - DB 196,98,125,24,45,141,194,2,0 ; vbroadcastss 0x2c28d(%rip),%ymm13 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> - DB 196,98,125,24,53,136,194,2,0 ; vbroadcastss 0x2c288(%rip),%ymm14 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,45,109,193,2,0 ; vbroadcastss 0x2c16d(%rip),%ymm13 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> + DB 196,98,125,24,53,104,193,2,0 ; vbroadcastss 0x2c168(%rip),%ymm14 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,84,89,230 ; vmulps %ymm14,%ymm5,%ymm12 DB 196,66,93,184,229 ; vfmadd231ps %ymm13,%ymm4,%ymm12 - DB 196,98,125,24,61,121,194,2,0 ; vbroadcastss 0x2c279(%rip),%ymm15 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,98,125,24,61,89,193,2,0 ; vbroadcastss 0x2c159(%rip),%ymm15 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 196,66,77,184,231 ; vfmadd231ps %ymm15,%ymm6,%ymm12 DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9 DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -1080,7 +1080,7 @@ _sk_color_hsw LABEL PROC DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1 DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8 DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9 - DB 196,226,125,24,5,47,193,2,0 ; vbroadcastss 0x2c12f(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,15,192,2,0 ; vbroadcastss 0x2c00f(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -1108,11 +1108,11 @@ _sk_luminosity_hsw LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,213 ; vmulps %ymm5,%ymm3,%ymm10 DB 197,100,89,222 ; vmulps %ymm6,%ymm3,%ymm11 - DB 196,98,125,24,45,222,192,2,0 ; vbroadcastss 0x2c0de(%rip),%ymm13 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> - DB 196,98,125,24,53,217,192,2,0 ; vbroadcastss 0x2c0d9(%rip),%ymm14 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,45,190,191,2,0 ; vbroadcastss 0x2bfbe(%rip),%ymm13 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> + DB 196,98,125,24,53,185,191,2,0 ; vbroadcastss 0x2bfb9(%rip),%ymm14 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,116,89,230 ; vmulps %ymm14,%ymm1,%ymm12 DB 196,66,109,184,229 ; vfmadd231ps %ymm13,%ymm2,%ymm12 - DB 196,98,125,24,61,202,192,2,0 ; vbroadcastss 0x2c0ca(%rip),%ymm15 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,98,125,24,61,170,191,2,0 ; vbroadcastss 0x2bfaa(%rip),%ymm15 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 196,66,53,184,231 ; vfmadd231ps %ymm15,%ymm9,%ymm12 DB 196,65,44,89,206 ; vmulps %ymm14,%ymm10,%ymm9 DB 196,66,61,184,205 ; vfmadd231ps %ymm13,%ymm8,%ymm9 @@ -1168,7 +1168,7 @@ _sk_luminosity_hsw LABEL PROC DB 196,193,116,95,206 ; vmaxps %ymm14,%ymm1,%ymm1 DB 196,65,44,95,198 ; vmaxps %ymm14,%ymm10,%ymm8 DB 196,65,124,95,206 ; vmaxps %ymm14,%ymm0,%ymm9 - DB 196,226,125,24,5,128,191,2,0 ; vbroadcastss 0x2bf80(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,96,190,2,0 ; vbroadcastss 0x2be60(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -1198,17 +1198,17 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 72,133,255 ; test %rdi,%rdi DB 15,133,179,0,0,0 ; jne 1444 <_sk_srcover_rgba_8888_hsw+0xd3> DB 196,129,126,111,60,152 ; vmovdqu (%r8,%r11,4),%ymm7 - DB 197,197,219,37,129,182,2,0 ; vpand 0x2b681(%rip),%ymm7,%ymm4 # 2ca20 <_sk_overlay_sse2_8bit+0x4c8> + DB 197,197,219,37,129,181,2,0 ; vpand 0x2b581(%rip),%ymm7,%ymm4 # 2c920 <_sk_overlay_sse2_8bit+0x471> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,69,0,45,148,182,2,0 ; vpshufb 0x2b694(%rip),%ymm7,%ymm5 # 2ca40 <_sk_overlay_sse2_8bit+0x4e8> + DB 196,226,69,0,45,148,181,2,0 ; vpshufb 0x2b594(%rip),%ymm7,%ymm5 # 2c940 <_sk_overlay_sse2_8bit+0x491> DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,69,0,53,167,182,2,0 ; vpshufb 0x2b6a7(%rip),%ymm7,%ymm6 # 2ca60 <_sk_overlay_sse2_8bit+0x508> + DB 196,226,69,0,53,167,181,2,0 ; vpshufb 0x2b5a7(%rip),%ymm7,%ymm6 # 2c960 <_sk_overlay_sse2_8bit+0x4b1> DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 DB 197,197,114,215,24 ; vpsrld $0x18,%ymm7,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,221,190,2,0 ; vbroadcastss 0x2bedd(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,189,189,2,0 ; vbroadcastss 0x2bdbd(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,252,190,2,0 ; vbroadcastss 0x2befc(%rip),%ymm9 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,13,220,189,2,0 ; vbroadcastss 0x2bddc(%rip),%ymm9 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,194,93,184,192 ; vfmadd231ps %ymm8,%ymm4,%ymm0 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 @@ -1330,7 +1330,7 @@ _sk_clamp_0_hsw LABEL PROC PUBLIC _sk_clamp_1_hsw _sk_clamp_1_hsw LABEL PROC - DB 196,98,125,24,5,234,188,2,0 ; vbroadcastss 0x2bcea(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,202,187,2,0 ; vbroadcastss 0x2bbca(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 @@ -1340,7 +1340,7 @@ _sk_clamp_1_hsw LABEL PROC PUBLIC _sk_clamp_a_hsw _sk_clamp_a_hsw LABEL PROC - DB 196,98,125,24,5,201,188,2,0 ; vbroadcastss 0x2bcc9(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,169,187,2,0 ; vbroadcastss 0x2bba9(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0 DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1 @@ -1350,7 +1350,7 @@ _sk_clamp_a_hsw LABEL PROC PUBLIC _sk_clamp_a_dst_hsw _sk_clamp_a_dst_hsw LABEL PROC - DB 196,98,125,24,5,171,188,2,0 ; vbroadcastss 0x2bcab(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,139,187,2,0 ; vbroadcastss 0x2bb8b(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,68,93,248 ; vminps %ymm8,%ymm7,%ymm7 DB 197,220,93,231 ; vminps %ymm7,%ymm4,%ymm4 DB 197,212,93,239 ; vminps %ymm7,%ymm5,%ymm5 @@ -1377,7 +1377,7 @@ _sk_swap_rb_hsw LABEL PROC PUBLIC _sk_invert_hsw _sk_invert_hsw LABEL PROC - DB 196,98,125,24,5,102,188,2,0 ; vbroadcastss 0x2bc66(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,70,187,2,0 ; vbroadcastss 0x2bb46(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 @@ -1423,7 +1423,7 @@ PUBLIC _sk_unpremul_hsw _sk_unpremul_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,246,187,2,0 ; vbroadcastss 0x2bbf6(%rip),%ymm10 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,21,214,186,2,0 ; vbroadcastss 0x2bad6(%rip),%ymm10 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10 DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 @@ -1434,16 +1434,16 @@ _sk_unpremul_hsw LABEL PROC PUBLIC _sk_from_srgb_hsw _sk_from_srgb_hsw LABEL PROC - DB 196,98,125,24,5,3,188,2,0 ; vbroadcastss 0x2bc03(%rip),%ymm8 # 2d2dc <_sk_overlay_sse2_8bit+0xd84> + DB 196,98,125,24,5,227,186,2,0 ; vbroadcastss 0x2bae3(%rip),%ymm8 # 2d1bc <_sk_overlay_sse2_8bit+0xd0d> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,245,187,2,0 ; vbroadcastss 0x2bbf5(%rip),%ymm11 # 2d2e0 <_sk_overlay_sse2_8bit+0xd88> - DB 196,98,125,24,37,216,187,2,0 ; vbroadcastss 0x2bbd8(%rip),%ymm12 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,29,213,186,2,0 ; vbroadcastss 0x2bad5(%rip),%ymm11 # 2d1c0 <_sk_overlay_sse2_8bit+0xd11> + DB 196,98,125,24,37,184,186,2,0 ; vbroadcastss 0x2bab8(%rip),%ymm12 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,125,168,235 ; vfmadd213ps %ymm11,%ymm0,%ymm13 - DB 196,98,125,24,53,221,187,2,0 ; vbroadcastss 0x2bbdd(%rip),%ymm14 # 2d2e4 <_sk_overlay_sse2_8bit+0xd8c> + DB 196,98,125,24,53,189,186,2,0 ; vbroadcastss 0x2babd(%rip),%ymm14 # 2d1c4 <_sk_overlay_sse2_8bit+0xd15> DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13 - DB 196,98,125,24,21,211,187,2,0 ; vbroadcastss 0x2bbd3(%rip),%ymm10 # 2d2e8 <_sk_overlay_sse2_8bit+0xd90> + DB 196,98,125,24,21,179,186,2,0 ; vbroadcastss 0x2bab3(%rip),%ymm10 # 2d1c8 <_sk_overlay_sse2_8bit+0xd19> DB 196,193,124,194,194,1 ; vcmpltps %ymm10,%ymm0,%ymm0 DB 196,195,21,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm13,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -1464,16 +1464,16 @@ _sk_from_srgb_hsw LABEL PROC PUBLIC _sk_from_srgb_dst_hsw _sk_from_srgb_dst_hsw LABEL PROC - DB 196,98,125,24,5,107,187,2,0 ; vbroadcastss 0x2bb6b(%rip),%ymm8 # 2d2dc <_sk_overlay_sse2_8bit+0xd84> + DB 196,98,125,24,5,75,186,2,0 ; vbroadcastss 0x2ba4b(%rip),%ymm8 # 2d1bc <_sk_overlay_sse2_8bit+0xd0d> DB 196,65,92,89,200 ; vmulps %ymm8,%ymm4,%ymm9 DB 197,92,89,212 ; vmulps %ymm4,%ymm4,%ymm10 - DB 196,98,125,24,29,93,187,2,0 ; vbroadcastss 0x2bb5d(%rip),%ymm11 # 2d2e0 <_sk_overlay_sse2_8bit+0xd88> - DB 196,98,125,24,37,64,187,2,0 ; vbroadcastss 0x2bb40(%rip),%ymm12 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,29,61,186,2,0 ; vbroadcastss 0x2ba3d(%rip),%ymm11 # 2d1c0 <_sk_overlay_sse2_8bit+0xd11> + DB 196,98,125,24,37,32,186,2,0 ; vbroadcastss 0x2ba20(%rip),%ymm12 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,93,168,235 ; vfmadd213ps %ymm11,%ymm4,%ymm13 - DB 196,98,125,24,53,69,187,2,0 ; vbroadcastss 0x2bb45(%rip),%ymm14 # 2d2e4 <_sk_overlay_sse2_8bit+0xd8c> + DB 196,98,125,24,53,37,186,2,0 ; vbroadcastss 0x2ba25(%rip),%ymm14 # 2d1c4 <_sk_overlay_sse2_8bit+0xd15> DB 196,66,45,168,238 ; vfmadd213ps %ymm14,%ymm10,%ymm13 - DB 196,98,125,24,21,59,187,2,0 ; vbroadcastss 0x2bb3b(%rip),%ymm10 # 2d2e8 <_sk_overlay_sse2_8bit+0xd90> + DB 196,98,125,24,21,27,186,2,0 ; vbroadcastss 0x2ba1b(%rip),%ymm10 # 2d1c8 <_sk_overlay_sse2_8bit+0xd19> DB 196,193,92,194,226,1 ; vcmpltps %ymm10,%ymm4,%ymm4 DB 196,195,21,74,225,64 ; vblendvps %ymm4,%ymm9,%ymm13,%ymm4 DB 196,65,84,89,200 ; vmulps %ymm8,%ymm5,%ymm9 @@ -1495,19 +1495,19 @@ _sk_from_srgb_dst_hsw LABEL PROC PUBLIC _sk_to_srgb_hsw _sk_to_srgb_hsw LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,223,186,2,0 ; vbroadcastss 0x2badf(%rip),%ymm8 # 2d2ec <_sk_overlay_sse2_8bit+0xd94> + DB 196,98,125,24,5,191,185,2,0 ; vbroadcastss 0x2b9bf(%rip),%ymm8 # 2d1cc <_sk_overlay_sse2_8bit+0xd1d> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,213,186,2,0 ; vbroadcastss 0x2bad5(%rip),%ymm11 # 2d2f0 <_sk_overlay_sse2_8bit+0xd98> - DB 196,98,125,24,37,208,186,2,0 ; vbroadcastss 0x2bad0(%rip),%ymm12 # 2d2f4 <_sk_overlay_sse2_8bit+0xd9c> + DB 196,98,125,24,29,181,185,2,0 ; vbroadcastss 0x2b9b5(%rip),%ymm11 # 2d1d0 <_sk_overlay_sse2_8bit+0xd21> + DB 196,98,125,24,37,176,185,2,0 ; vbroadcastss 0x2b9b0(%rip),%ymm12 # 2d1d4 <_sk_overlay_sse2_8bit+0xd25> DB 196,65,124,40,236 ; vmovaps %ymm12,%ymm13 DB 196,66,53,168,235 ; vfmadd213ps %ymm11,%ymm9,%ymm13 - DB 196,98,125,24,53,193,186,2,0 ; vbroadcastss 0x2bac1(%rip),%ymm14 # 2d2f8 <_sk_overlay_sse2_8bit+0xda0> + DB 196,98,125,24,53,161,185,2,0 ; vbroadcastss 0x2b9a1(%rip),%ymm14 # 2d1d8 <_sk_overlay_sse2_8bit+0xd29> DB 196,66,53,168,238 ; vfmadd213ps %ymm14,%ymm9,%ymm13 - DB 196,98,125,24,61,183,186,2,0 ; vbroadcastss 0x2bab7(%rip),%ymm15 # 2d2fc <_sk_overlay_sse2_8bit+0xda4> + DB 196,98,125,24,61,151,185,2,0 ; vbroadcastss 0x2b997(%rip),%ymm15 # 2d1dc <_sk_overlay_sse2_8bit+0xd2d> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,20,89,201 ; vmulps %ymm9,%ymm13,%ymm9 - DB 196,98,125,24,45,163,186,2,0 ; vbroadcastss 0x2baa3(%rip),%ymm13 # 2d300 <_sk_overlay_sse2_8bit+0xda8> + DB 196,98,125,24,45,131,185,2,0 ; vbroadcastss 0x2b983(%rip),%ymm13 # 2d1e0 <_sk_overlay_sse2_8bit+0xd31> DB 196,193,124,194,197,1 ; vcmpltps %ymm13,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -1539,26 +1539,26 @@ _sk_rgb_to_hsl_hsw LABEL PROC DB 197,116,93,202 ; vminps %ymm2,%ymm1,%ymm9 DB 196,65,124,93,201 ; vminps %ymm9,%ymm0,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,190,185,2,0 ; vbroadcastss 0x2b9be(%rip),%ymm11 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,29,158,184,2,0 ; vbroadcastss 0x2b89e(%rip),%ymm11 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 197,116,194,234,1 ; vcmpltps %ymm2,%ymm1,%ymm13 - DB 196,98,125,24,53,255,185,2,0 ; vbroadcastss 0x2b9ff(%rip),%ymm14 # 2d304 <_sk_overlay_sse2_8bit+0xdac> + DB 196,98,125,24,53,223,184,2,0 ; vbroadcastss 0x2b8df(%rip),%ymm14 # 2d1e4 <_sk_overlay_sse2_8bit+0xd35> DB 196,65,4,87,255 ; vxorps %ymm15,%ymm15,%ymm15 DB 196,67,5,74,238,208 ; vblendvps %ymm13,%ymm14,%ymm15,%ymm13 DB 196,66,37,168,229 ; vfmadd213ps %ymm13,%ymm11,%ymm12 DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 DB 197,124,92,233 ; vsubps %ymm1,%ymm0,%ymm13 - DB 196,98,125,24,53,230,185,2,0 ; vbroadcastss 0x2b9e6(%rip),%ymm14 # 2d30c <_sk_overlay_sse2_8bit+0xdb4> + DB 196,98,125,24,53,198,184,2,0 ; vbroadcastss 0x2b8c6(%rip),%ymm14 # 2d1ec <_sk_overlay_sse2_8bit+0xd3d> DB 196,66,37,168,238 ; vfmadd213ps %ymm14,%ymm11,%ymm13 - DB 196,98,125,24,53,212,185,2,0 ; vbroadcastss 0x2b9d4(%rip),%ymm14 # 2d308 <_sk_overlay_sse2_8bit+0xdb0> + DB 196,98,125,24,53,180,184,2,0 ; vbroadcastss 0x2b8b4(%rip),%ymm14 # 2d1e8 <_sk_overlay_sse2_8bit+0xd39> DB 196,194,37,168,214 ; vfmadd213ps %ymm14,%ymm11,%ymm2 DB 197,188,194,201,0 ; vcmpeqps %ymm1,%ymm8,%ymm1 DB 196,227,21,74,202,16 ; vblendvps %ymm1,%ymm2,%ymm13,%ymm1 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,195,117,74,196,0 ; vblendvps %ymm0,%ymm12,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,29,75,185,2,0 ; vbroadcastss 0x2b94b(%rip),%ymm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,29,43,184,2,0 ; vbroadcastss 0x2b82b(%rip),%ymm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,193,116,89,211 ; vmulps %ymm11,%ymm1,%ymm2 DB 197,36,194,218,1 ; vcmpltps %ymm2,%ymm11,%ymm11 DB 196,65,12,92,224 ; vsubps %ymm8,%ymm14,%ymm12 @@ -1568,7 +1568,7 @@ _sk_rgb_to_hsl_hsw LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,199,128 ; vblendvps %ymm8,%ymm15,%ymm0,%ymm0 DB 196,195,117,74,207,128 ; vblendvps %ymm8,%ymm15,%ymm1,%ymm1 - DB 196,98,125,24,5,122,185,2,0 ; vbroadcastss 0x2b97a(%rip),%ymm8 # 2d310 <_sk_overlay_sse2_8bit+0xdb8> + DB 196,98,125,24,5,90,184,2,0 ; vbroadcastss 0x2b85a(%rip),%ymm8 # 2d1f0 <_sk_overlay_sse2_8bit+0xd41> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1583,30 +1583,30 @@ _sk_hsl_to_rgb_hsw LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,233 ; vmovaps %ymm1,%ymm5 DB 197,252,40,224 ; vmovaps %ymm0,%ymm4 - DB 196,98,125,24,5,209,184,2,0 ; vbroadcastss 0x2b8d1(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,177,183,2,0 ; vbroadcastss 0x2b7b1(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,84,89,210 ; vmulps %ymm2,%ymm5,%ymm10 DB 196,65,84,92,218 ; vsubps %ymm10,%ymm5,%ymm11 DB 196,67,45,74,203,144 ; vblendvps %ymm9,%ymm11,%ymm10,%ymm9 DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 - DB 196,98,125,24,13,16,185,2,0 ; vbroadcastss 0x2b910(%rip),%ymm9 # 2d308 <_sk_overlay_sse2_8bit+0xdb0> + DB 196,98,125,24,13,240,183,2,0 ; vbroadcastss 0x2b7f0(%rip),%ymm9 # 2d1e8 <_sk_overlay_sse2_8bit+0xd39> DB 196,66,109,170,202 ; vfmsub213ps %ymm10,%ymm2,%ymm9 - DB 196,98,125,24,29,14,185,2,0 ; vbroadcastss 0x2b90e(%rip),%ymm11 # 2d314 <_sk_overlay_sse2_8bit+0xdbc> + DB 196,98,125,24,29,238,183,2,0 ; vbroadcastss 0x2b7ee(%rip),%ymm11 # 2d1f4 <_sk_overlay_sse2_8bit+0xd45> DB 196,65,92,88,219 ; vaddps %ymm11,%ymm4,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,45,224,184,2,0 ; vbroadcastss 0x2b8e0(%rip),%ymm13 # 2d304 <_sk_overlay_sse2_8bit+0xdac> + DB 196,98,125,24,45,192,183,2,0 ; vbroadcastss 0x2b7c0(%rip),%ymm13 # 2d1e4 <_sk_overlay_sse2_8bit+0xd35> DB 196,193,4,89,197 ; vmulps %ymm13,%ymm15,%ymm0 - DB 196,98,125,24,53,218,184,2,0 ; vbroadcastss 0x2b8da(%rip),%ymm14 # 2d30c <_sk_overlay_sse2_8bit+0xdb4> + DB 196,98,125,24,53,186,183,2,0 ; vbroadcastss 0x2b7ba(%rip),%ymm14 # 2d1ec <_sk_overlay_sse2_8bit+0xd3d> DB 197,12,92,224 ; vsubps %ymm0,%ymm14,%ymm12 DB 196,66,37,168,225 ; vfmadd213ps %ymm9,%ymm11,%ymm12 - DB 196,226,125,24,29,212,184,2,0 ; vbroadcastss 0x2b8d4(%rip),%ymm3 # 2d318 <_sk_overlay_sse2_8bit+0xdc0> + DB 196,226,125,24,29,180,183,2,0 ; vbroadcastss 0x2b7b4(%rip),%ymm3 # 2d1f8 <_sk_overlay_sse2_8bit+0xd49> DB 196,193,100,194,255,2 ; vcmpleps %ymm15,%ymm3,%ymm7 DB 196,195,29,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm12,%ymm7 DB 196,65,60,194,231,2 ; vcmpleps %ymm15,%ymm8,%ymm12 DB 196,227,45,74,255,192 ; vblendvps %ymm12,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,37,171,184,2,0 ; vbroadcastss 0x2b8ab(%rip),%ymm12 # 2d310 <_sk_overlay_sse2_8bit+0xdb8> + DB 196,98,125,24,37,139,183,2,0 ; vbroadcastss 0x2b78b(%rip),%ymm12 # 2d1f0 <_sk_overlay_sse2_8bit+0xd41> DB 196,65,28,194,255,2 ; vcmpleps %ymm15,%ymm12,%ymm15 DB 196,194,37,168,193 ; vfmadd213ps %ymm9,%ymm11,%ymm0 DB 196,99,125,74,255,240 ; vblendvps %ymm15,%ymm7,%ymm0,%ymm15 @@ -1622,7 +1622,7 @@ _sk_hsl_to_rgb_hsw LABEL PROC DB 197,156,194,192,2 ; vcmpleps %ymm0,%ymm12,%ymm0 DB 196,194,37,168,249 ; vfmadd213ps %ymm9,%ymm11,%ymm7 DB 196,227,69,74,201,0 ; vblendvps %ymm0,%ymm1,%ymm7,%ymm1 - DB 196,226,125,24,5,95,184,2,0 ; vbroadcastss 0x2b85f(%rip),%ymm0 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> + DB 196,226,125,24,5,63,183,2,0 ; vbroadcastss 0x2b73f(%rip),%ymm0 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 DB 196,227,125,8,224,1 ; vroundps $0x1,%ymm0,%ymm4 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 @@ -1672,10 +1672,10 @@ _sk_scale_u8_hsw LABEL PROC DB 72,133,255 ; test %rdi,%rdi DB 117,58 ; jne 1bb2 <_sk_scale_u8_hsw+0x52> DB 196,2,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm8 - DB 197,57,219,5,202,184,2,0 ; vpand 0x2b8ca(%rip),%xmm8,%xmm8 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 197,57,219,5,186,183,2,0 ; vpand 0x2b7ba(%rip),%xmm8,%xmm8 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,135,183,2,0 ; vbroadcastss 0x2b787(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,103,182,2,0 ; vbroadcastss 0x2b667(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -1724,7 +1724,7 @@ _sk_scale_u8_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf001c5c <_sk_overlay_sse2_8bit+0xffffffffdefd5704> + DB 233,255,255,255,222 ; jmpq ffffffffdf001c5c <_sk_overlay_sse2_8bit+0xffffffffdefd57ad> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -1748,20 +1748,20 @@ _sk_scale_565_hsw LABEL PROC DB 15,133,161,0,0,0 ; jne 1d28 <_sk_scale_565_hsw+0xc0> DB 196,1,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm8 DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 - DB 196,98,125,88,13,137,182,2,0 ; vpbroadcastd 0x2b689(%rip),%ymm9 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,98,125,88,13,105,181,2,0 ; vpbroadcastd 0x2b569(%rip),%ymm9 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 196,65,61,219,201 ; vpand %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,122,182,2,0 ; vbroadcastss 0x2b67a(%rip),%ymm10 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,98,125,24,21,90,181,2,0 ; vbroadcastss 0x2b55a(%rip),%ymm10 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,88,21,112,182,2,0 ; vpbroadcastd 0x2b670(%rip),%ymm10 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,98,125,88,21,80,181,2,0 ; vpbroadcastd 0x2b550(%rip),%ymm10 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,97,182,2,0 ; vbroadcastss 0x2b661(%rip),%ymm11 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,98,125,24,29,65,181,2,0 ; vbroadcastss 0x2b541(%rip),%ymm11 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,88,29,87,182,2,0 ; vpbroadcastd 0x2b657(%rip),%ymm11 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,98,125,88,29,55,181,2,0 ; vpbroadcastd 0x2b537(%rip),%ymm11 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 196,65,61,219,195 ; vpand %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,72,182,2,0 ; vbroadcastss 0x2b648(%rip),%ymm11 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,98,125,24,29,40,181,2,0 ; vbroadcastss 0x2b528(%rip),%ymm11 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,100,194,223,1 ; vcmpltps %ymm7,%ymm3,%ymm11 DB 196,65,44,93,224 ; vminps %ymm8,%ymm10,%ymm12 @@ -1848,10 +1848,10 @@ _sk_lerp_u8_hsw LABEL PROC DB 72,133,255 ; test %rdi,%rdi DB 117,78 ; jne 1e61 <_sk_lerp_u8_hsw+0x66> DB 196,2,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm8 - DB 197,57,219,5,47,182,2,0 ; vpand 0x2b62f(%rip),%xmm8,%xmm8 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 197,57,219,5,31,181,2,0 ; vpand 0x2b51f(%rip),%xmm8,%xmm8 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,236,180,2,0 ; vbroadcastss 0x2b4ec(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,204,179,2,0 ; vbroadcastss 0x2b3cc(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 @@ -1903,7 +1903,7 @@ _sk_lerp_u8_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 232,255,255,255,221 ; callq ffffffffde001f0c <_sk_overlay_sse2_8bit+0xffffffffddfd59b4> + DB 232,255,255,255,221 ; callq ffffffffde001f0c <_sk_overlay_sse2_8bit+0xffffffffddfd5a5d> DB 255 ; (bad) DB 255 ; (bad) DB 255,210 ; callq *%rdx @@ -1927,20 +1927,20 @@ _sk_lerp_565_hsw LABEL PROC DB 15,133,181,0,0,0 ; jne 1fec <_sk_lerp_565_hsw+0xd4> DB 196,1,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm8 DB 196,66,125,51,192 ; vpmovzxwd %xmm8,%ymm8 - DB 196,98,125,88,13,217,179,2,0 ; vpbroadcastd 0x2b3d9(%rip),%ymm9 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,98,125,88,13,185,178,2,0 ; vpbroadcastd 0x2b2b9(%rip),%ymm9 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 196,65,61,219,201 ; vpand %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,202,179,2,0 ; vbroadcastss 0x2b3ca(%rip),%ymm10 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,98,125,24,21,170,178,2,0 ; vbroadcastss 0x2b2aa(%rip),%ymm10 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,88,21,192,179,2,0 ; vpbroadcastd 0x2b3c0(%rip),%ymm10 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,98,125,88,21,160,178,2,0 ; vpbroadcastd 0x2b2a0(%rip),%ymm10 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 196,65,61,219,210 ; vpand %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,177,179,2,0 ; vbroadcastss 0x2b3b1(%rip),%ymm11 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,98,125,24,29,145,178,2,0 ; vbroadcastss 0x2b291(%rip),%ymm11 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,88,29,167,179,2,0 ; vpbroadcastd 0x2b3a7(%rip),%ymm11 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,98,125,88,29,135,178,2,0 ; vpbroadcastd 0x2b287(%rip),%ymm11 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 196,65,61,219,195 ; vpand %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,152,179,2,0 ; vbroadcastss 0x2b398(%rip),%ymm11 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,98,125,24,29,120,178,2,0 ; vbroadcastss 0x2b278(%rip),%ymm11 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,100,194,223,1 ; vcmpltps %ymm7,%ymm3,%ymm11 DB 196,65,44,93,224 ; vminps %ymm8,%ymm10,%ymm12 @@ -2010,23 +2010,26 @@ _sk_load_tables_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,133,255 ; test %rdi,%rdi - DB 117,103 ; jne 2101 <_sk_load_tables_hsw+0x71> + DB 117,116 ; jne 210e <_sk_load_tables_hsw+0x7e> DB 196,193,126,111,28,144 ; vmovdqu (%r8,%rdx,4),%ymm3 - DB 197,229,219,13,216,169,2,0 ; vpand 0x2a9d8(%rip),%ymm3,%ymm1 # 2ca80 <_sk_overlay_sse2_8bit+0x528> + DB 197,229,219,13,216,168,2,0 ; vpand 0x2a8d8(%rip),%ymm3,%ymm1 # 2c980 <_sk_overlay_sse2_8bit+0x4d1> DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 DB 76,139,64,8 ; mov 0x8(%rax),%r8 DB 76,139,72,16 ; mov 0x10(%rax),%r9 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,194,109,146,4,136 ; vgatherdps %ymm2,(%r8,%ymm1,4),%ymm0 - DB 196,226,101,0,21,216,169,2,0 ; vpshufb 0x2a9d8(%rip),%ymm3,%ymm2 # 2caa0 <_sk_overlay_sse2_8bit+0x548> + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 + DB 197,253,239,192 ; vpxor %ymm0,%ymm0,%ymm0 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,194,53,146,12,145 ; vgatherdps %ymm9,(%r9,%ymm2,4),%ymm1 + DB 196,194,53,146,4,136 ; vgatherdps %ymm9,(%r8,%ymm1,4),%ymm0 + DB 196,98,101,0,13,207,168,2,0 ; vpshufb 0x2a8cf(%rip),%ymm3,%ymm9 # 2c9a0 <_sk_overlay_sse2_8bit+0x4f1> + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 + DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 + DB 196,130,45,146,12,137 ; vgatherdps %ymm10,(%r9,%ymm9,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,98,101,0,13,224,169,2,0 ; vpshufb 0x2a9e0(%rip),%ymm3,%ymm9 # 2cac0 <_sk_overlay_sse2_8bit+0x568> + DB 196,98,101,0,13,211,168,2,0 ; vpshufb 0x2a8d3(%rip),%ymm3,%ymm9 # 2c9c0 <_sk_overlay_sse2_8bit+0x511> DB 196,162,61,146,20,136 ; vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,40,178,2,0 ; vbroadcastss 0x2b228(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,251,176,2,0 ; vbroadcastss 0x2b0fb(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2035,21 +2038,21 @@ _sk_load_tables_hsw LABEL PROC DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,139 ; ja 20a0 <_sk_load_tables_hsw+0x10> + DB 15,135,122,255,255,255 ; ja 20a0 <_sk_load_tables_hsw+0x10> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,140,0,0,0 ; lea 0x8c(%rip),%r10 # 21ac <_sk_load_tables_hsw+0x11c> + DB 76,141,21,139,0,0,0 ; lea 0x8b(%rip),%r10 # 21bc <_sk_load_tables_hsw+0x12c> DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9 DB 77,1,209 ; add %r10,%r9 DB 65,255,225 ; jmpq *%r9 DB 196,193,121,110,28,144 ; vmovd (%r8,%rdx,4),%xmm3 - DB 233,107,255,255,255 ; jmpq 20a0 <_sk_load_tables_hsw+0x10> + DB 233,90,255,255,255 ; jmpq 20a0 <_sk_load_tables_hsw+0x10> DB 196,193,121,110,68,144,8 ; vmovd 0x8(%r8,%rdx,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 DB 196,227,117,2,216,4 ; vpblendd $0x4,%ymm0,%ymm1,%ymm3 DB 196,193,122,126,4,144 ; vmovq (%r8,%rdx,4),%xmm0 DB 196,227,101,2,216,3 ; vpblendd $0x3,%ymm0,%ymm3,%ymm3 - DB 233,68,255,255,255 ; jmpq 20a0 <_sk_load_tables_hsw+0x10> + DB 233,51,255,255,255 ; jmpq 20a0 <_sk_load_tables_hsw+0x10> DB 196,193,121,110,68,144,24 ; vmovd 0x18(%r8,%rdx,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -2062,25 +2065,24 @@ _sk_load_tables_hsw LABEL PROC DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3 DB 196,193,122,111,4,144 ; vmovdqu (%r8,%rdx,4),%xmm0 DB 196,227,125,2,219,240 ; vpblendd $0xf0,%ymm3,%ymm0,%ymm3 - DB 233,245,254,255,255 ; jmpq 20a0 <_sk_load_tables_hsw+0x10> - DB 144 ; nop - DB 126,255 ; jle 21ad <_sk_load_tables_hsw+0x11d> + DB 233,228,254,255,255 ; jmpq 20a0 <_sk_load_tables_hsw+0x10> + DB 127,255 ; jg 21bd <_sk_load_tables_hsw+0x12d> DB 255 ; (bad) - DB 255,159,255,255,255,137 ; lcall *-0x76000001(%rdi) + DB 255,160,255,255,255,138 ; jmpq *-0x75000001(%rax) DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 238 ; out %al,(%dx) + DB 239 ; out %eax,(%dx) DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 218,255 ; (bad) + DB 219,255 ; (bad) DB 255 ; (bad) - DB 255,198 ; inc %esi + DB 255,199 ; inc %edi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff - DB 176,255 ; mov $0xff,%al + DB 177,255 ; mov $0xff,%cl DB 255 ; (bad) DB 255 ; .byte 0xff @@ -2090,7 +2092,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,201,0,0,0 ; jne 22a7 <_sk_load_tables_u16_be_hsw+0xdf> + DB 15,133,215,0,0,0 ; jne 22c5 <_sk_load_tables_u16_be_hsw+0xed> DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8 DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2 DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3 @@ -2102,62 +2104,65 @@ _sk_load_tables_u16_be_hsw LABEL PROC DB 197,121,97,194 ; vpunpcklwd %xmm2,%xmm0,%xmm8 DB 197,121,105,202 ; vpunpckhwd %xmm2,%xmm0,%xmm9 DB 197,241,97,195 ; vpunpcklwd %xmm3,%xmm1,%xmm0 - DB 197,241,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm3 + DB 197,113,105,243 ; vpunpckhwd %xmm3,%xmm1,%xmm14 DB 197,185,108,200 ; vpunpcklqdq %xmm0,%xmm8,%xmm1 - DB 197,185,109,208 ; vpunpckhqdq %xmm0,%xmm8,%xmm2 - DB 197,49,108,195 ; vpunpcklqdq %xmm3,%xmm9,%xmm8 - DB 197,121,111,21,33,178,2,0 ; vmovdqa 0x2b221(%rip),%xmm10 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 196,193,113,219,194 ; vpand %xmm10,%xmm1,%xmm0 + DB 197,57,109,192 ; vpunpckhqdq %xmm0,%xmm8,%xmm8 + DB 196,65,49,108,214 ; vpunpcklqdq %xmm14,%xmm9,%xmm10 + DB 197,121,111,29,0,177,2,0 ; vmovdqa 0x2b100(%rip),%xmm11 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,193,113,219,195 ; vpand %xmm11,%xmm1,%xmm0 DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1 - DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 + DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 + DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 DB 76,139,64,8 ; mov 0x8(%rax),%r8 DB 76,139,72,16 ; mov 0x10(%rax),%r9 - DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 - DB 196,194,29,146,4,136 ; vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0 - DB 196,193,105,219,202 ; vpand %xmm10,%xmm2,%xmm1 - DB 196,226,125,51,209 ; vpmovzxwd %xmm1,%ymm2 - DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 - DB 196,194,29,146,12,145 ; vgatherdps %ymm12,(%r9,%ymm2,4),%ymm1 + DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 + DB 197,253,239,192 ; vpxor %ymm0,%ymm0,%ymm0 + DB 196,194,21,146,4,136 ; vgatherdps %ymm13,(%r8,%ymm1,4),%ymm0 + DB 196,193,57,219,203 ; vpand %xmm11,%xmm8,%xmm1 + DB 196,98,125,51,193 ; vpmovzxwd %xmm1,%ymm8 + DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 + DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 + DB 196,130,21,146,12,129 ; vgatherdps %ymm13,(%r9,%ymm8,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,193,57,219,210 ; vpand %xmm10,%xmm8,%xmm2 - DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8 - DB 196,162,37,146,20,128 ; vgatherdps %ymm11,(%rax,%ymm8,4),%ymm2 - DB 197,177,109,219 ; vpunpckhqdq %xmm3,%xmm9,%xmm3 + DB 196,193,41,219,219 ; vpand %xmm11,%xmm10,%xmm3 + DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 + DB 196,226,29,146,20,152 ; vgatherdps %ymm12,(%rax,%ymm3,4),%ymm2 + DB 196,193,49,109,222 ; vpunpckhqdq %xmm14,%xmm9,%xmm3 DB 197,185,113,243,8 ; vpsllw $0x8,%xmm3,%xmm8 DB 197,225,113,211,8 ; vpsrlw $0x8,%xmm3,%xmm3 DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3 DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,158,176,2,0 ; vbroadcastss 0x2b09e(%rip),%ymm8 # 2d33c <_sk_overlay_sse2_8bit+0xde4> + DB 196,98,125,24,5,96,175,2,0 ; vbroadcastss 0x2af60(%rip),%ymm8 # 2d21c <_sk_overlay_sse2_8bit+0xd6d> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je 230d <_sk_load_tables_u16_be_hsw+0x145> + DB 116,85 ; je 232b <_sk_load_tables_u16_be_hsw+0x153> DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb 230d <_sk_load_tables_u16_be_hsw+0x145> + DB 114,72 ; jb 232b <_sk_load_tables_u16_be_hsw+0x153> DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je 231a <_sk_load_tables_u16_be_hsw+0x152> + DB 116,72 ; je 2338 <_sk_load_tables_u16_be_hsw+0x160> DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb 231a <_sk_load_tables_u16_be_hsw+0x152> + DB 114,59 ; jb 2338 <_sk_load_tables_u16_be_hsw+0x160> DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,9,255,255,255 ; je 21f9 <_sk_load_tables_u16_be_hsw+0x31> + DB 15,132,251,254,255,255 ; je 2209 <_sk_load_tables_u16_be_hsw+0x31> DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,248,254,255,255 ; jb 21f9 <_sk_load_tables_u16_be_hsw+0x31> + DB 15,130,234,254,255,255 ; jb 2209 <_sk_load_tables_u16_be_hsw+0x31> DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9 - DB 233,236,254,255,255 ; jmpq 21f9 <_sk_load_tables_u16_be_hsw+0x31> + DB 233,222,254,255,255 ; jmpq 2209 <_sk_load_tables_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,223,254,255,255 ; jmpq 21f9 <_sk_load_tables_u16_be_hsw+0x31> + DB 233,209,254,255,255 ; jmpq 2209 <_sk_load_tables_u16_be_hsw+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,214,254,255,255 ; jmpq 21f9 <_sk_load_tables_u16_be_hsw+0x31> + DB 233,200,254,255,255 ; jmpq 2209 <_sk_load_tables_u16_be_hsw+0x31> PUBLIC _sk_load_tables_rgb_u16_be_hsw _sk_load_tables_rgb_u16_be_hsw LABEL PROC @@ -2165,7 +2170,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,82 ; lea (%rdx,%rdx,2),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,193,0,0,0 ; jne 23f6 <_sk_load_tables_rgb_u16_be_hsw+0xd3> + DB 15,133,206,0,0,0 ; jne 2421 <_sk_load_tables_rgb_u16_be_hsw+0xe0> DB 196,1,122,111,28,72 ; vmovdqu (%r8,%r9,2),%xmm11 DB 196,129,122,111,92,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm3 DB 196,129,122,111,84,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm2 @@ -2173,113 +2178,118 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC DB 197,249,115,216,4 ; vpsrldq $0x4,%xmm0,%xmm0 DB 196,193,57,115,219,6 ; vpsrldq $0x6,%xmm11,%xmm8 DB 197,169,115,219,6 ; vpsrldq $0x6,%xmm3,%xmm10 - DB 197,153,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm12 + DB 197,241,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm1 DB 197,177,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm9 - DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 - DB 196,193,57,97,202 ; vpunpcklwd %xmm10,%xmm8,%xmm1 + DB 196,65,113,97,201 ; vpunpcklwd %xmm9,%xmm1,%xmm9 DB 197,233,97,192 ; vpunpcklwd %xmm0,%xmm2,%xmm0 - DB 196,193,25,97,209 ; vpunpcklwd %xmm9,%xmm12,%xmm2 - DB 197,97,97,193 ; vpunpcklwd %xmm1,%xmm3,%xmm8 - DB 197,225,105,201 ; vpunpckhwd %xmm1,%xmm3,%xmm1 - DB 197,249,97,218 ; vpunpcklwd %xmm2,%xmm0,%xmm3 - DB 197,249,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm0 - DB 197,185,108,211 ; vpunpcklqdq %xmm3,%xmm8,%xmm2 - DB 197,185,109,219 ; vpunpckhqdq %xmm3,%xmm8,%xmm3 - DB 197,113,108,192 ; vpunpcklqdq %xmm0,%xmm1,%xmm8 - DB 197,121,111,13,176,176,2,0 ; vmovdqa 0x2b0b0(%rip),%xmm9 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 196,193,105,219,193 ; vpand %xmm9,%xmm2,%xmm0 - DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1 + DB 196,193,57,97,210 ; vpunpcklwd %xmm10,%xmm8,%xmm2 + DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 + DB 197,225,97,202 ; vpunpcklwd %xmm2,%xmm3,%xmm1 + DB 197,225,105,210 ; vpunpckhwd %xmm2,%xmm3,%xmm2 + DB 196,193,121,97,217 ; vpunpcklwd %xmm9,%xmm0,%xmm3 + DB 196,193,121,105,193 ; vpunpckhwd %xmm9,%xmm0,%xmm0 + DB 197,105,108,192 ; vpunpcklqdq %xmm0,%xmm2,%xmm8 + DB 197,241,108,195 ; vpunpcklqdq %xmm3,%xmm1,%xmm0 + DB 197,241,109,203 ; vpunpckhqdq %xmm3,%xmm1,%xmm1 + DB 197,249,111,29,128,175,2,0 ; vmovdqa 0x2af80(%rip),%xmm3 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 197,249,219,195 ; vpand %xmm3,%xmm0,%xmm0 + DB 196,98,125,51,200 ; vpmovzxwd %xmm0,%ymm9 DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 + DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 DB 76,139,64,8 ; mov 0x8(%rax),%r8 DB 76,139,72,16 ; mov 0x10(%rax),%r9 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,194,109,146,4,136 ; vgatherdps %ymm2,(%r8,%ymm1,4),%ymm0 - DB 196,193,97,219,201 ; vpand %xmm9,%xmm3,%xmm1 - DB 196,226,125,51,209 ; vpmovzxwd %xmm1,%ymm2 - DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,194,101,146,12,145 ; vgatherdps %ymm3,(%r9,%ymm2,4),%ymm1 + DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 + DB 197,253,239,192 ; vpxor %ymm0,%ymm0,%ymm0 + DB 196,130,37,146,4,136 ; vgatherdps %ymm11,(%r8,%ymm9,4),%ymm0 + DB 197,241,219,203 ; vpand %xmm3,%xmm1,%xmm1 + DB 196,98,125,51,201 ; vpmovzxwd %xmm1,%ymm9 + DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 + DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 + DB 196,130,37,146,12,137 ; vgatherdps %ymm11,(%r9,%ymm9,4),%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,193,57,219,209 ; vpand %xmm9,%xmm8,%xmm2 - DB 196,226,125,51,218 ; vpmovzxwd %xmm2,%ymm3 + DB 197,185,219,219 ; vpand %xmm3,%xmm8,%xmm3 + DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 DB 196,226,45,146,20,152 ; vgatherdps %ymm10,(%rax,%ymm3,4),%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,184,174,2,0 ; vbroadcastss 0x2aeb8(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,109,173,2,0 ; vbroadcastss 0x2ad6d(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0 - DB 196,1,121,196,92,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm11 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,1,121,196,92,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm11 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,32 ; jne 242f <_sk_load_tables_rgb_u16_be_hsw+0x10c> + DB 117,31 ; jne 2459 <_sk_load_tables_rgb_u16_be_hsw+0x118> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 233,59,255,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 233,47,255,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> DB 196,129,121,110,68,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm0 - DB 196,1,121,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm0,%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,1,121,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm0,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,49 ; jb 247a <_sk_load_tables_rgb_u16_be_hsw+0x157> + DB 114,48 ; jb 24a3 <_sk_load_tables_rgb_u16_be_hsw+0x162> DB 196,129,121,110,68,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm0 - DB 196,129,121,196,92,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm0,%xmm3 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,129,121,196,92,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm0,%xmm3 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 117,50 ; jne 2495 <_sk_load_tables_rgb_u16_be_hsw+0x172> + DB 117,48 ; jne 24bd <_sk_load_tables_rgb_u16_be_hsw+0x17c> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 233,240,254,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 233,229,254,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 - DB 233,213,254,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 233,203,254,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> DB 196,129,121,110,68,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm0 - DB 196,1,121,196,84,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm0,%xmm10 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,1,121,196,84,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm0,%xmm10 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,40 ; jb 24d7 <_sk_load_tables_rgb_u16_be_hsw+0x1b4> + DB 114,39 ; jb 24fe <_sk_load_tables_rgb_u16_be_hsw+0x1bd> DB 196,129,121,110,68,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm0 - DB 196,129,121,196,84,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm0,%xmm2 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,129,121,196,84,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm0,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 117,32 ; jne 24e9 <_sk_load_tables_rgb_u16_be_hsw+0x1c6> + DB 117,30 ; jne 250f <_sk_load_tables_rgb_u16_be_hsw+0x1ce> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 - DB 233,147,254,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 + DB 233,138,254,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 - DB 233,129,254,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 233,121,254,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> DB 196,129,121,110,68,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm0 - DB 196,1,121,196,100,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm0,%xmm12 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,129,121,196,76,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm0,%xmm1 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,25 ; jb 251c <_sk_load_tables_rgb_u16_be_hsw+0x1f9> + DB 114,25 ; jb 2542 <_sk_load_tables_rgb_u16_be_hsw+0x201> DB 196,129,121,110,68,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm0 - DB 196,129,121,196,68,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm0,%xmm0 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 233,78,254,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 196,129,121,196,68,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm0,%xmm0 + DB 233,70,254,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 233,69,254,255,255 ; jmpq 236a <_sk_load_tables_rgb_u16_be_hsw+0x47> + DB 233,61,254,255,255 ; jmpq 2388 <_sk_load_tables_rgb_u16_be_hsw+0x47> PUBLIC _sk_byte_tables_hsw _sk_byte_tables_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,168,173,2,0 ; vbroadcastss 0x2ada8(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,98,172,2,0 ; vbroadcastss 0x2ac62(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 DB 196,65,249,126,200 ; vmovq %xmm9,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 76,139,16 ; mov (%rax),%r10 - DB 196,131,121,32,4,10,0 ; vpinsrb $0x0,(%r10,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,193,121,110,193 ; vmovd %r9d,%xmm0 DB 196,67,249,22,201,1 ; vpextrq $0x1,%xmm9,%r9 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,121,32,20,2,1 ; vpinsrb $0x1,(%r10,%r8,1),%xmm0,%xmm10 + DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d + DB 196,67,121,32,208,1 ; vpinsrb $0x1,%r8d,%xmm0,%xmm10 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,32 ; shr $0x20,%r9 DB 196,99,125,57,200,1 ; vextracti128 $0x1,%ymm9,%xmm0 @@ -2306,10 +2316,12 @@ _sk_byte_tables_hsw LABEL PROC DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,193,249,126,201 ; vmovq %xmm1,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,4,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,194 ; vmovd %r10d,%xmm0 DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,121,32,4,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,121,32,193,1 ; vpinsrb $0x1,%r9d,%xmm0,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 @@ -2333,28 +2345,30 @@ _sk_byte_tables_hsw LABEL PROC DB 196,195,121,32,200,7 ; vpinsrb $0x7,%r8d,%xmm0,%xmm1 DB 76,139,64,16 ; mov 0x10(%rax),%r8 DB 196,193,108,89,192 ; vmulps %ymm8,%ymm2,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 197,253,91,208 ; vcvtps2dq %ymm0,%ymm2 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,194 ; vmovd %r10d,%xmm0 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,105,32,20,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm2,%xmm2 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,121,32,193,1 ; vpinsrb $0x1,%r9d,%xmm0,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,210,1 ; vextracti128 $0x1,%ymm2,%xmm2 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 196,195,105,32,209,2 ; vpinsrb $0x2,%r9d,%xmm2,%xmm2 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 196,195,105,32,210,3 ; vpinsrb $0x3,%r10d,%xmm2,%xmm2 + DB 196,195,121,32,194,3 ; vpinsrb $0x3,%r10d,%xmm0,%xmm0 DB 69,137,202 ; mov %r9d,%r10d DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 196,195,105,32,210,4 ; vpinsrb $0x4,%r10d,%xmm2,%xmm2 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,195,121,32,194,4 ; vpinsrb $0x4,%r10d,%xmm0,%xmm0 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,233,32 ; shr $0x20,%r9 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 196,195,105,32,193,5 ; vpinsrb $0x5,%r9d,%xmm2,%xmm0 + DB 196,195,121,32,193,5 ; vpinsrb $0x5,%r9d,%xmm0,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 196,195,121,32,193,6 ; vpinsrb $0x6,%r9d,%xmm0,%xmm0 @@ -2363,28 +2377,30 @@ _sk_byte_tables_hsw LABEL PROC DB 196,195,121,32,208,7 ; vpinsrb $0x7,%r8d,%xmm0,%xmm2 DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 196,193,100,89,192 ; vmulps %ymm8,%ymm3,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 197,253,91,216 ; vcvtps2dq %ymm0,%ymm3 + DB 196,193,249,126,216 ; vmovq %xmm3,%r8 DB 68,137,192 ; mov %r8d,%eax - DB 196,195,121,32,28,1,0 ; vpinsrb $0x0,(%r9,%rax,1),%xmm0,%xmm3 - DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 197,249,110,192 ; vmovd %eax,%xmm0 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,97,32,28,1,1 ; vpinsrb $0x1,(%r9,%r8,1),%xmm3,%xmm3 - DB 65,137,192 ; mov %eax,%r8d - DB 72,193,232,32 ; shr $0x20,%rax - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 - DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d - DB 196,195,97,32,216,2 ; vpinsrb $0x2,%r8d,%xmm3,%xmm3 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 196,227,121,32,192,1 ; vpinsrb $0x1,%eax,%xmm0,%xmm0 + DB 68,137,208 ; mov %r10d,%eax + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,227,125,57,219,1 ; vextracti128 $0x1,%ymm3,%xmm3 DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,227,97,32,216,3 ; vpinsrb $0x3,%eax,%xmm3,%xmm3 + DB 196,227,121,32,192,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm0 + DB 196,193,249,126,216 ; vmovq %xmm3,%r8 + DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax + DB 196,227,121,32,192,3 ; vpinsrb $0x3,%eax,%xmm0,%xmm0 DB 68,137,192 ; mov %r8d,%eax DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,227,97,32,216,4 ; vpinsrb $0x4,%eax,%xmm3,%xmm3 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,227,121,32,192,4 ; vpinsrb $0x4,%eax,%xmm0,%xmm0 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 DB 73,193,232,32 ; shr $0x20,%r8 DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax - DB 196,227,97,32,192,5 ; vpinsrb $0x5,%eax,%xmm3,%xmm0 + DB 196,227,121,32,192,5 ; vpinsrb $0x5,%eax,%xmm0,%xmm0 DB 68,137,208 ; mov %r10d,%eax DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,227,121,32,216,6 ; vpinsrb $0x6,%eax,%xmm0,%xmm3 @@ -2392,7 +2408,7 @@ _sk_byte_tables_hsw LABEL PROC DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,145,171,2,0 ; vbroadcastss 0x2ab91(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,48,170,2,0 ; vbroadcastss 0x2aa30(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 @@ -2420,10 +2436,12 @@ _sk_byte_tables_rgb_hsw LABEL PROC DB 196,65,249,126,200 ; vmovq %xmm9,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 76,139,16 ; mov (%rax),%r10 - DB 196,131,121,32,4,10,0 ; vpinsrb $0x0,(%r10,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,193,121,110,193 ; vmovd %r9d,%xmm0 DB 196,67,249,22,201,1 ; vpextrq $0x1,%xmm9,%r9 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,121,32,20,2,1 ; vpinsrb $0x1,(%r10,%r8,1),%xmm0,%xmm10 + DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d + DB 196,67,121,32,208,1 ; vpinsrb $0x1,%r8d,%xmm0,%xmm10 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,32 ; shr $0x20,%r9 DB 196,99,125,57,200,1 ; vextracti128 $0x1,%ymm9,%xmm0 @@ -2450,10 +2468,12 @@ _sk_byte_tables_rgb_hsw LABEL PROC DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,193,249,126,201 ; vmovq %xmm1,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,4,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,194 ; vmovd %r10d,%xmm0 DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,121,32,4,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,121,32,193,1 ; vpinsrb $0x1,%r9d,%xmm0,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 @@ -2477,28 +2497,30 @@ _sk_byte_tables_rgb_hsw LABEL PROC DB 196,195,121,32,200,7 ; vpinsrb $0x7,%r8d,%xmm0,%xmm1 DB 76,139,72,16 ; mov 0x10(%rax),%r9 DB 197,188,89,194 ; vmulps %ymm2,%ymm8,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 197,253,91,208 ; vcvtps2dq %ymm0,%ymm2 + DB 196,193,249,126,208 ; vmovq %xmm2,%r8 DB 68,137,192 ; mov %r8d,%eax - DB 196,195,121,32,20,1,0 ; vpinsrb $0x0,(%r9,%rax,1),%xmm0,%xmm2 - DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 197,249,110,192 ; vmovd %eax,%xmm0 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,105,32,20,1,1 ; vpinsrb $0x1,(%r9,%r8,1),%xmm2,%xmm2 - DB 65,137,192 ; mov %eax,%r8d - DB 72,193,232,32 ; shr $0x20,%rax - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 - DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d - DB 196,195,105,32,208,2 ; vpinsrb $0x2,%r8d,%xmm2,%xmm2 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 196,227,121,32,192,1 ; vpinsrb $0x1,%eax,%xmm0,%xmm0 + DB 68,137,208 ; mov %r10d,%eax + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,227,125,57,210,1 ; vextracti128 $0x1,%ymm2,%xmm2 DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,227,105,32,208,3 ; vpinsrb $0x3,%eax,%xmm2,%xmm2 + DB 196,227,121,32,192,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm0 + DB 196,193,249,126,208 ; vmovq %xmm2,%r8 + DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax + DB 196,227,121,32,192,3 ; vpinsrb $0x3,%eax,%xmm0,%xmm0 DB 68,137,192 ; mov %r8d,%eax DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,227,105,32,208,4 ; vpinsrb $0x4,%eax,%xmm2,%xmm2 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,227,121,32,192,4 ; vpinsrb $0x4,%eax,%xmm0,%xmm0 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,232,32 ; shr $0x20,%r8 DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax - DB 196,227,105,32,192,5 ; vpinsrb $0x5,%eax,%xmm2,%xmm0 + DB 196,227,121,32,192,5 ; vpinsrb $0x5,%eax,%xmm0,%xmm0 DB 68,137,208 ; mov %r10d,%eax DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,227,121,32,208,6 ; vpinsrb $0x6,%eax,%xmm0,%xmm2 @@ -2506,7 +2528,7 @@ _sk_byte_tables_rgb_hsw LABEL PROC DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax DB 196,194,125,49,193 ; vpmovzxbd %xmm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,121,169,2,0 ; vbroadcastss 0x2a979(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,4,168,2,0 ; vbroadcastss 0x2a804(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 @@ -2530,6 +2552,7 @@ _sk_table_r_hsw LABEL PROC DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,125,91,192 ; vcvtps2dq %ymm0,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 196,130,53,146,4,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2546,6 +2569,7 @@ _sk_table_g_hsw LABEL PROC DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,125,91,193 ; vcvtps2dq %ymm1,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,130,53,146,12,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2562,6 +2586,7 @@ _sk_table_b_hsw LABEL PROC DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2 DB 197,125,91,194 ; vcvtps2dq %ymm2,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 196,130,53,146,20,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2578,6 +2603,7 @@ _sk_table_a_hsw LABEL PROC DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3 DB 197,125,91,195 ; vcvtps2dq %ymm3,%ymm8 DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 196,130,53,146,28,128 ; vgatherdps %ymm9,(%r8,%ymm8,4),%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2595,41 +2621,41 @@ _sk_parametric_r_hsw LABEL PROC DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10 DB 196,226,125,24,0 ; vbroadcastss (%rax),%ymm0 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,111,168,2,0 ; vbroadcastss 0x2a86f(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> - DB 196,98,125,24,45,106,168,2,0 ; vbroadcastss 0x2a86a(%rip),%ymm13 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,234,166,2,0 ; vbroadcastss 0x2a6ea(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> + DB 196,98,125,24,45,229,166,2,0 ; vbroadcastss 0x2a6e5(%rip),%ymm13 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,192,167,2,0 ; vbroadcastss 0x2a7c0(%rip),%ymm13 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,45,59,166,2,0 ; vbroadcastss 0x2a63b(%rip),%ymm13 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,82,168,2,0 ; vbroadcastss 0x2a852(%rip),%ymm13 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,45,205,166,2,0 ; vbroadcastss 0x2a6cd(%rip),%ymm13 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,72,168,2,0 ; vbroadcastss 0x2a848(%rip),%ymm11 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,29,195,166,2,0 ; vbroadcastss 0x2a6c3(%rip),%ymm11 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,62,168,2,0 ; vbroadcastss 0x2a83e(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,185,166,2,0 ; vbroadcastss 0x2a6b9(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,52,168,2,0 ; vbroadcastss 0x2a834(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,175,166,2,0 ; vbroadcastss 0x2a6af(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,21,168,2,0 ; vbroadcastss 0x2a815(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,144,166,2,0 ; vbroadcastss 0x2a690(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,11,168,2,0 ; vbroadcastss 0x2a80b(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,134,166,2,0 ; vbroadcastss 0x2a686(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11 - DB 196,226,125,24,5,1,168,2,0 ; vbroadcastss 0x2a801(%rip),%ymm0 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,226,125,24,5,124,166,2,0 ; vbroadcastss 0x2a67c(%rip),%ymm0 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,247,167,2,0 ; vbroadcastss 0x2a7f7(%rip),%ymm10 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,21,114,166,2,0 ; vbroadcastss 0x2a672(%rip),%ymm10 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0 DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,21,234,167,2,0 ; vbroadcastss 0x2a7ea(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,101,166,2,0 ; vbroadcastss 0x2a665(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 + DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,1,167,2,0 ; vbroadcastss 0x2a701(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,125,165,2,0 ; vbroadcastss 0x2a57d(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2647,41 +2673,41 @@ _sk_parametric_g_hsw LABEL PROC DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10 DB 196,226,125,24,8 ; vbroadcastss (%rax),%ymm1 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,73,167,2,0 ; vbroadcastss 0x2a749(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> - DB 196,98,125,24,45,68,167,2,0 ; vbroadcastss 0x2a744(%rip),%ymm13 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,197,165,2,0 ; vbroadcastss 0x2a5c5(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> + DB 196,98,125,24,45,192,165,2,0 ; vbroadcastss 0x2a5c0(%rip),%ymm13 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,154,166,2,0 ; vbroadcastss 0x2a69a(%rip),%ymm13 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,45,22,165,2,0 ; vbroadcastss 0x2a516(%rip),%ymm13 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,44,167,2,0 ; vbroadcastss 0x2a72c(%rip),%ymm13 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,45,168,165,2,0 ; vbroadcastss 0x2a5a8(%rip),%ymm13 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,34,167,2,0 ; vbroadcastss 0x2a722(%rip),%ymm11 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,29,158,165,2,0 ; vbroadcastss 0x2a59e(%rip),%ymm11 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,24,167,2,0 ; vbroadcastss 0x2a718(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,148,165,2,0 ; vbroadcastss 0x2a594(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,14,167,2,0 ; vbroadcastss 0x2a70e(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,138,165,2,0 ; vbroadcastss 0x2a58a(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,239,166,2,0 ; vbroadcastss 0x2a6ef(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,107,165,2,0 ; vbroadcastss 0x2a56b(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,229,166,2,0 ; vbroadcastss 0x2a6e5(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,97,165,2,0 ; vbroadcastss 0x2a561(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11 - DB 196,226,125,24,13,219,166,2,0 ; vbroadcastss 0x2a6db(%rip),%ymm1 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,226,125,24,13,87,165,2,0 ; vbroadcastss 0x2a557(%rip),%ymm1 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,209,166,2,0 ; vbroadcastss 0x2a6d1(%rip),%ymm10 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,21,77,165,2,0 ; vbroadcastss 0x2a54d(%rip),%ymm10 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 196,98,125,24,21,196,166,2,0 ; vbroadcastss 0x2a6c4(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,64,165,2,0 ; vbroadcastss 0x2a540(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 + DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1 DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,219,165,2,0 ; vbroadcastss 0x2a5db(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,88,164,2,0 ; vbroadcastss 0x2a458(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2699,41 +2725,41 @@ _sk_parametric_b_hsw LABEL PROC DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10 DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,35,166,2,0 ; vbroadcastss 0x2a623(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> - DB 196,98,125,24,45,30,166,2,0 ; vbroadcastss 0x2a61e(%rip),%ymm13 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,160,164,2,0 ; vbroadcastss 0x2a4a0(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> + DB 196,98,125,24,45,155,164,2,0 ; vbroadcastss 0x2a49b(%rip),%ymm13 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,116,165,2,0 ; vbroadcastss 0x2a574(%rip),%ymm13 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,45,241,163,2,0 ; vbroadcastss 0x2a3f1(%rip),%ymm13 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,6,166,2,0 ; vbroadcastss 0x2a606(%rip),%ymm13 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,45,131,164,2,0 ; vbroadcastss 0x2a483(%rip),%ymm13 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,252,165,2,0 ; vbroadcastss 0x2a5fc(%rip),%ymm11 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,29,121,164,2,0 ; vbroadcastss 0x2a479(%rip),%ymm11 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,242,165,2,0 ; vbroadcastss 0x2a5f2(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,111,164,2,0 ; vbroadcastss 0x2a46f(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,232,165,2,0 ; vbroadcastss 0x2a5e8(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,101,164,2,0 ; vbroadcastss 0x2a465(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,201,165,2,0 ; vbroadcastss 0x2a5c9(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,70,164,2,0 ; vbroadcastss 0x2a446(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,191,165,2,0 ; vbroadcastss 0x2a5bf(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,60,164,2,0 ; vbroadcastss 0x2a43c(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11 - DB 196,226,125,24,21,181,165,2,0 ; vbroadcastss 0x2a5b5(%rip),%ymm2 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,226,125,24,21,50,164,2,0 ; vbroadcastss 0x2a432(%rip),%ymm2 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,171,165,2,0 ; vbroadcastss 0x2a5ab(%rip),%ymm10 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,21,40,164,2,0 ; vbroadcastss 0x2a428(%rip),%ymm10 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2 DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 - DB 196,98,125,24,21,158,165,2,0 ; vbroadcastss 0x2a59e(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,27,164,2,0 ; vbroadcastss 0x2a41b(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 + DB 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,181,164,2,0 ; vbroadcastss 0x2a4b5(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,51,163,2,0 ; vbroadcastss 0x2a333(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2751,41 +2777,41 @@ _sk_parametric_a_hsw LABEL PROC DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10 DB 196,226,125,24,24 ; vbroadcastss (%rax),%ymm3 DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11 - DB 196,98,125,24,37,253,164,2,0 ; vbroadcastss 0x2a4fd(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> - DB 196,98,125,24,45,248,164,2,0 ; vbroadcastss 0x2a4f8(%rip),%ymm13 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,123,163,2,0 ; vbroadcastss 0x2a37b(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> + DB 196,98,125,24,45,118,163,2,0 ; vbroadcastss 0x2a376(%rip),%ymm13 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,78,164,2,0 ; vbroadcastss 0x2a44e(%rip),%ymm13 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,45,204,162,2,0 ; vbroadcastss 0x2a2cc(%rip),%ymm13 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10 - DB 196,98,125,24,45,224,164,2,0 ; vbroadcastss 0x2a4e0(%rip),%ymm13 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,45,94,163,2,0 ; vbroadcastss 0x2a35e(%rip),%ymm13 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13 - DB 196,98,125,24,29,214,164,2,0 ; vbroadcastss 0x2a4d6(%rip),%ymm11 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,29,84,163,2,0 ; vbroadcastss 0x2a354(%rip),%ymm11 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11 - DB 196,98,125,24,37,204,164,2,0 ; vbroadcastss 0x2a4cc(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,74,163,2,0 ; vbroadcastss 0x2a34a(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,194,164,2,0 ; vbroadcastss 0x2a4c2(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,64,163,2,0 ; vbroadcastss 0x2a340(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10 DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,163,164,2,0 ; vbroadcastss 0x2a4a3(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,33,163,2,0 ; vbroadcastss 0x2a321(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,153,164,2,0 ; vbroadcastss 0x2a499(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,23,163,2,0 ; vbroadcastss 0x2a317(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11 - DB 196,226,125,24,29,143,164,2,0 ; vbroadcastss 0x2a48f(%rip),%ymm3 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,226,125,24,29,13,163,2,0 ; vbroadcastss 0x2a30d(%rip),%ymm3 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,133,164,2,0 ; vbroadcastss 0x2a485(%rip),%ymm10 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,21,3,163,2,0 ; vbroadcastss 0x2a303(%rip),%ymm10 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3 DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3 - DB 196,98,125,24,21,120,164,2,0 ; vbroadcastss 0x2a478(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,246,162,2,0 ; vbroadcastss 0x2a2f6(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 + DB 197,172,88,219 ; vaddps %ymm3,%ymm10,%ymm3 DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,143,163,2,0 ; vbroadcastss 0x2a38f(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,14,162,2,0 ; vbroadcastss 0x2a20e(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2802,35 +2828,35 @@ _sk_gamma_hsw LABEL PROC DB 197,252,40,241 ; vmovaps %ymm1,%ymm6 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,124,91,208 ; vcvtdq2ps %ymm0,%ymm10 - DB 196,98,125,24,29,220,163,2,0 ; vbroadcastss 0x2a3dc(%rip),%ymm11 # 2d340 <_sk_overlay_sse2_8bit+0xde8> - DB 196,226,125,24,45,215,163,2,0 ; vbroadcastss 0x2a3d7(%rip),%ymm5 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,29,91,162,2,0 ; vbroadcastss 0x2a25b(%rip),%ymm11 # 2d220 <_sk_overlay_sse2_8bit+0xd71> + DB 196,226,125,24,45,86,162,2,0 ; vbroadcastss 0x2a256(%rip),%ymm5 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 197,124,84,205 ; vandps %ymm5,%ymm0,%ymm9 - DB 196,226,125,24,37,46,163,2,0 ; vbroadcastss 0x2a32e(%rip),%ymm4 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,37,173,161,2,0 ; vbroadcastss 0x2a1ad(%rip),%ymm4 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,52,86,228 ; vorps %ymm4,%ymm9,%ymm12 - DB 196,98,125,24,45,193,163,2,0 ; vbroadcastss 0x2a3c1(%rip),%ymm13 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,45,64,162,2,0 ; vbroadcastss 0x2a240(%rip),%ymm13 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,66,37,168,213 ; vfmadd213ps %ymm13,%ymm11,%ymm10 - DB 196,98,125,24,13,183,163,2,0 ; vbroadcastss 0x2a3b7(%rip),%ymm9 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,13,54,162,2,0 ; vbroadcastss 0x2a236(%rip),%ymm9 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,66,29,188,209 ; vfnmadd231ps %ymm9,%ymm12,%ymm10 DB 197,124,91,246 ; vcvtdq2ps %ymm6,%ymm14 DB 196,66,37,168,245 ; vfmadd213ps %ymm13,%ymm11,%ymm14 DB 197,124,91,255 ; vcvtdq2ps %ymm7,%ymm15 DB 196,66,37,168,253 ; vfmadd213ps %ymm13,%ymm11,%ymm15 - DB 196,98,125,24,29,155,163,2,0 ; vbroadcastss 0x2a39b(%rip),%ymm11 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,29,26,162,2,0 ; vbroadcastss 0x2a21a(%rip),%ymm11 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,65,28,88,227 ; vaddps %ymm11,%ymm12,%ymm12 - DB 196,98,125,24,45,145,163,2,0 ; vbroadcastss 0x2a391(%rip),%ymm13 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,45,16,162,2,0 ; vbroadcastss 0x2a210(%rip),%ymm13 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 196,65,20,94,228 ; vdivps %ymm12,%ymm13,%ymm12 DB 196,65,44,92,212 ; vsubps %ymm12,%ymm10,%ymm10 DB 196,98,125,24,32 ; vbroadcastss (%rax),%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 DB 196,67,125,8,194,1 ; vroundps $0x1,%ymm10,%ymm8 DB 196,65,44,92,192 ; vsubps %ymm8,%ymm10,%ymm8 - DB 196,226,125,24,21,109,163,2,0 ; vbroadcastss 0x2a36d(%rip),%ymm2 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,226,125,24,21,236,161,2,0 ; vbroadcastss 0x2a1ec(%rip),%ymm2 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 197,44,88,210 ; vaddps %ymm2,%ymm10,%ymm10 - DB 196,226,125,24,29,100,163,2,0 ; vbroadcastss 0x2a364(%rip),%ymm3 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,226,125,24,29,227,161,2,0 ; vbroadcastss 0x2a1e3(%rip),%ymm3 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,98,61,188,211 ; vfnmadd231ps %ymm3,%ymm8,%ymm10 - DB 196,226,125,24,13,90,163,2,0 ; vbroadcastss 0x2a35a(%rip),%ymm1 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,226,125,24,13,217,161,2,0 ; vbroadcastss 0x2a1d9(%rip),%ymm1 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,65,116,92,192 ; vsubps %ymm8,%ymm1,%ymm8 - DB 196,226,125,24,5,80,163,2,0 ; vbroadcastss 0x2a350(%rip),%ymm0 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,226,125,24,5,207,161,2,0 ; vbroadcastss 0x2a1cf(%rip),%ymm0 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 196,65,124,94,192 ; vdivps %ymm8,%ymm0,%ymm8 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 197,204,84,245 ; vandps %ymm5,%ymm6,%ymm6 @@ -2861,7 +2887,7 @@ _sk_gamma_hsw LABEL PROC DB 197,244,92,205 ; vsubps %ymm5,%ymm1,%ymm1 DB 197,252,94,193 ; vdivps %ymm1,%ymm0,%ymm0 DB 197,236,88,192 ; vaddps %ymm0,%ymm2,%ymm0 - DB 196,226,125,24,13,193,162,2,0 ; vbroadcastss 0x2a2c1(%rip),%ymm1 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,226,125,24,13,64,161,2,0 ; vbroadcastss 0x2a140(%rip),%ymm1 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 197,188,89,209 ; vmulps %ymm1,%ymm8,%ymm2 DB 197,204,89,217 ; vmulps %ymm1,%ymm6,%ymm3 DB 197,252,89,225 ; vmulps %ymm1,%ymm0,%ymm4 @@ -2879,26 +2905,26 @@ _sk_gamma_hsw LABEL PROC PUBLIC _sk_lab_to_xyz_hsw _sk_lab_to_xyz_hsw LABEL PROC - DB 196,98,125,24,5,121,162,2,0 ; vbroadcastss 0x2a279(%rip),%ymm8 # 2d36c <_sk_overlay_sse2_8bit+0xe14> - DB 196,98,125,24,13,220,161,2,0 ; vbroadcastss 0x2a1dc(%rip),%ymm9 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> - DB 196,98,125,24,21,107,162,2,0 ; vbroadcastss 0x2a26b(%rip),%ymm10 # 2d370 <_sk_overlay_sse2_8bit+0xe18> + DB 196,98,125,24,5,248,160,2,0 ; vbroadcastss 0x2a0f8(%rip),%ymm8 # 2d24c <_sk_overlay_sse2_8bit+0xd9d> + DB 196,98,125,24,13,91,160,2,0 ; vbroadcastss 0x2a05b(%rip),%ymm9 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> + DB 196,98,125,24,21,234,160,2,0 ; vbroadcastss 0x2a0ea(%rip),%ymm10 # 2d250 <_sk_overlay_sse2_8bit+0xda1> DB 196,194,53,168,202 ; vfmadd213ps %ymm10,%ymm9,%ymm1 DB 196,194,53,168,210 ; vfmadd213ps %ymm10,%ymm9,%ymm2 - DB 196,98,125,24,13,92,162,2,0 ; vbroadcastss 0x2a25c(%rip),%ymm9 # 2d374 <_sk_overlay_sse2_8bit+0xe1c> + DB 196,98,125,24,13,219,160,2,0 ; vbroadcastss 0x2a0db(%rip),%ymm9 # 2d254 <_sk_overlay_sse2_8bit+0xda5> DB 196,66,125,184,200 ; vfmadd231ps %ymm8,%ymm0,%ymm9 - DB 196,226,125,24,5,82,162,2,0 ; vbroadcastss 0x2a252(%rip),%ymm0 # 2d378 <_sk_overlay_sse2_8bit+0xe20> + DB 196,226,125,24,5,209,160,2,0 ; vbroadcastss 0x2a0d1(%rip),%ymm0 # 2d258 <_sk_overlay_sse2_8bit+0xda9> DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 - DB 196,98,125,24,5,73,162,2,0 ; vbroadcastss 0x2a249(%rip),%ymm8 # 2d37c <_sk_overlay_sse2_8bit+0xe24> + DB 196,98,125,24,5,200,160,2,0 ; vbroadcastss 0x2a0c8(%rip),%ymm8 # 2d25c <_sk_overlay_sse2_8bit+0xdad> DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8 - DB 196,98,125,24,13,63,162,2,0 ; vbroadcastss 0x2a23f(%rip),%ymm9 # 2d380 <_sk_overlay_sse2_8bit+0xe28> + DB 196,98,125,24,13,190,160,2,0 ; vbroadcastss 0x2a0be(%rip),%ymm9 # 2d260 <_sk_overlay_sse2_8bit+0xdb1> DB 196,98,109,172,200 ; vfnmadd213ps %ymm0,%ymm2,%ymm9 DB 196,193,60,89,200 ; vmulps %ymm8,%ymm8,%ymm1 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 - DB 196,226,125,24,21,44,162,2,0 ; vbroadcastss 0x2a22c(%rip),%ymm2 # 2d384 <_sk_overlay_sse2_8bit+0xe2c> + DB 196,226,125,24,21,171,160,2,0 ; vbroadcastss 0x2a0ab(%rip),%ymm2 # 2d264 <_sk_overlay_sse2_8bit+0xdb5> DB 197,108,194,209,1 ; vcmpltps %ymm1,%ymm2,%ymm10 - DB 196,98,125,24,29,34,162,2,0 ; vbroadcastss 0x2a222(%rip),%ymm11 # 2d388 <_sk_overlay_sse2_8bit+0xe30> + DB 196,98,125,24,29,161,160,2,0 ; vbroadcastss 0x2a0a1(%rip),%ymm11 # 2d268 <_sk_overlay_sse2_8bit+0xdb9> DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,37,24,162,2,0 ; vbroadcastss 0x2a218(%rip),%ymm12 # 2d38c <_sk_overlay_sse2_8bit+0xe34> + DB 196,98,125,24,37,151,160,2,0 ; vbroadcastss 0x2a097(%rip),%ymm12 # 2d26c <_sk_overlay_sse2_8bit+0xdbd> DB 196,65,60,89,196 ; vmulps %ymm12,%ymm8,%ymm8 DB 196,99,61,74,193,160 ; vblendvps %ymm10,%ymm1,%ymm8,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -2913,9 +2939,9 @@ _sk_lab_to_xyz_hsw LABEL PROC DB 196,65,52,88,203 ; vaddps %ymm11,%ymm9,%ymm9 DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 DB 196,227,53,74,208,32 ; vblendvps %ymm2,%ymm0,%ymm9,%ymm2 - DB 196,226,125,24,5,205,161,2,0 ; vbroadcastss 0x2a1cd(%rip),%ymm0 # 2d390 <_sk_overlay_sse2_8bit+0xe38> + DB 196,226,125,24,5,76,160,2,0 ; vbroadcastss 0x2a04c(%rip),%ymm0 # 2d270 <_sk_overlay_sse2_8bit+0xdc1> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,196,161,2,0 ; vbroadcastss 0x2a1c4(%rip),%ymm8 # 2d394 <_sk_overlay_sse2_8bit+0xe3c> + DB 196,98,125,24,5,67,160,2,0 ; vbroadcastss 0x2a043(%rip),%ymm8 # 2d274 <_sk_overlay_sse2_8bit+0xdc5> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2929,12 +2955,12 @@ _sk_load_a8_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,52 ; jne 3225 <_sk_load_a8_hsw+0x4c> + DB 117,52 ; jne 3286 <_sk_load_a8_hsw+0x4c> DB 196,130,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm0 - DB 197,249,219,5,81,162,2,0 ; vpand 0x2a251(%rip),%xmm0,%xmm0 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 197,249,219,5,224,160,2,0 ; vpand 0x2a0e0(%rip),%xmm0,%xmm0 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,15,161,2,0 ; vbroadcastss 0x2a10f(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,142,159,2,0 ; vbroadcastss 0x29f8e(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -2946,15 +2972,15 @@ _sk_load_a8_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,190 ; ja 31f7 <_sk_load_a8_hsw+0x1e> + DB 119,190 ; ja 3258 <_sk_load_a8_hsw+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # 32bc <_sk_load_a8_hsw+0xe3> + DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 331c <_sk_load_a8_hsw+0xe2> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 235,159 ; jmp 31f7 <_sk_load_a8_hsw+0x1e> + DB 235,159 ; jmp 3258 <_sk_load_a8_hsw+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0 @@ -2962,7 +2988,7 @@ _sk_load_a8_hsw LABEL PROC DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,121,2,193,1 ; vpblendd $0x1,%xmm1,%xmm0,%xmm0 - DB 233,119,255,255,255 ; jmpq 31f7 <_sk_load_a8_hsw+0x1e> + DB 233,119,255,255,255 ; jmpq 3258 <_sk_load_a8_hsw+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0 @@ -2973,22 +2999,23 @@ _sk_load_a8_hsw LABEL PROC DB 196,129,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,113,2,192,12 ; vpblendd $0xc,%xmm0,%xmm1,%xmm0 - DB 233,60,255,255,255 ; jmpq 31f7 <_sk_load_a8_hsw+0x1e> - DB 144 ; nop - DB 145 ; xchg %eax,%ecx + DB 233,60,255,255,255 ; jmpq 3258 <_sk_load_a8_hsw+0x1e> + DB 146 ; xchg %eax,%edx DB 255 ; (bad) DB 255 ; (bad) - DB 255,171,255,255,255,156 ; ljmp *-0x63000001(%rbx) + DB 255,172,255,255,255,157,255 ; ljmp *-0x620001(%rdi,%rdi,8) DB 255 ; (bad) DB 255 ; (bad) + DB 234 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf0032cc <_sk_overlay_sse2_8bit+0xffffffffdefd6d74> DB 255 ; (bad) DB 255 ; (bad) - DB 255,211 ; callq *%rbx + DB 223,255 ; (bad) + DB 255 ; (bad) + DB 255,212 ; callq *%rsp DB 255 ; (bad) DB 255 ; (bad) - DB 255,196 ; inc %esp + DB 255,197 ; inc %ebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -3002,12 +3029,12 @@ _sk_load_a8_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,52 ; jne 3324 <_sk_load_a8_dst_hsw+0x4c> + DB 117,52 ; jne 3384 <_sk_load_a8_dst_hsw+0x4c> DB 196,130,121,48,36,24 ; vpmovzxbw (%r8,%r11,1),%xmm4 - DB 197,217,219,37,82,161,2,0 ; vpand 0x2a152(%rip),%xmm4,%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 197,217,219,37,226,159,2,0 ; vpand 0x29fe2(%rip),%xmm4,%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 196,226,125,51,228 ; vpmovzxwd %xmm4,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,16,160,2,0 ; vbroadcastss 0x2a010(%rip),%ymm5 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,45,144,158,2,0 ; vbroadcastss 0x29e90(%rip),%ymm5 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,220,89,253 ; vmulps %ymm5,%ymm4,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 @@ -3019,15 +3046,15 @@ _sk_load_a8_dst_hsw LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,190 ; ja 32f6 <_sk_load_a8_dst_hsw+0x1e> + DB 119,190 ; ja 3356 <_sk_load_a8_dst_hsw+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,121,0,0,0 ; lea 0x79(%rip),%r10 # 33bc <_sk_load_a8_dst_hsw+0xe4> + DB 76,141,21,121,0,0,0 ; lea 0x79(%rip),%r10 # 341c <_sk_load_a8_dst_hsw+0xe4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 235,159 ; jmp 32f6 <_sk_load_a8_dst_hsw+0x1e> + DB 235,159 ; jmp 3356 <_sk_load_a8_dst_hsw+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,2 ; vpinsrw $0x2,%eax,%xmm4,%xmm4 @@ -3035,7 +3062,7 @@ _sk_load_a8_dst_hsw LABEL PROC DB 197,249,110,232 ; vmovd %eax,%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,89,2,229,1 ; vpblendd $0x1,%xmm5,%xmm4,%xmm4 - DB 233,119,255,255,255 ; jmpq 32f6 <_sk_load_a8_dst_hsw+0x1e> + DB 233,119,255,255,255 ; jmpq 3356 <_sk_load_a8_dst_hsw+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,6 ; vpinsrw $0x6,%eax,%xmm4,%xmm4 @@ -3046,7 +3073,7 @@ _sk_load_a8_dst_hsw LABEL PROC DB 196,129,121,110,44,24 ; vmovd (%r8,%r11,1),%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,81,2,228,12 ; vpblendd $0xc,%xmm4,%xmm5,%xmm4 - DB 233,60,255,255,255 ; jmpq 32f6 <_sk_load_a8_dst_hsw+0x1e> + DB 233,60,255,255,255 ; jmpq 3356 <_sk_load_a8_dst_hsw+0x1e> DB 102,144 ; xchg %ax,%ax DB 144 ; nop DB 255 ; (bad) @@ -3055,7 +3082,7 @@ _sk_load_a8_dst_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 232,255,255,255,221 ; callq ffffffffde0033cc <_sk_overlay_sse2_8bit+0xffffffffddfd6e74> + DB 232,255,255,255,221 ; callq ffffffffde00342c <_sk_overlay_sse2_8bit+0xffffffffddfd6f7d> DB 255 ; (bad) DB 255 ; (bad) DB 255,210 ; callq *%rdx @@ -3092,8 +3119,10 @@ _sk_gather_a8_hsw LABEL PROC DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,4,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 71,15,182,28,24 ; movzbl (%r8,%r11,1),%r11d + DB 196,193,121,110,195 ; vmovd %r11d,%xmm0 + DB 196,195,121,32,194,1 ; vpinsrb $0x1,%r10d,%xmm0,%xmm0 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax @@ -3108,7 +3137,7 @@ _sk_gather_a8_hsw LABEL PROC DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,139,158,2,0 ; vbroadcastss 0x29e8b(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,4,157,2,0 ; vbroadcastss 0x29d04(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -3128,57 +3157,56 @@ _sk_store_a8_hsw LABEL PROC DB 77,15,175,193 ; imul %r9,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,12,158,2,0 ; vbroadcastss 0x29e0c(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,133,156,2,0 ; vbroadcastss 0x29c85(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8 DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 - DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 + DB 197,57,103,192 ; vpackuswb %xmm0,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 34f5 <_sk_store_a8_hsw+0x45> - DB 196,1,123,17,4,24 ; vmovsd %xmm8,(%r8,%r11,1) + DB 117,10 ; jne 355b <_sk_store_a8_hsw+0x44> + DB 196,1,121,214,4,24 ; vmovq %xmm8,(%r8,%r11,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 34f1 <_sk_store_a8_hsw+0x41> + DB 119,236 ; ja 3557 <_sk_store_a8_hsw+0x40> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,87,0,0,0 ; lea 0x57(%rip),%r10 # 356c <_sk_store_a8_hsw+0xbc> + DB 76,141,21,85,0,0,0 ; lea 0x55(%rip),%r10 # 35d0 <_sk_store_a8_hsw+0xb9> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,3,121,20,4,24,0 ; vpextrb $0x0,%xmm8,(%r8,%r11,1) - DB 235,202 ; jmp 34f1 <_sk_store_a8_hsw+0x41> + DB 235,202 ; jmp 3557 <_sk_store_a8_hsw+0x40> DB 196,3,121,20,68,24,2,4 ; vpextrb $0x4,%xmm8,0x2(%r8,%r11,1) - DB 196,98,57,0,5,40,159,2,0 ; vpshufb 0x29f28(%rip),%xmm8,%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xf08> + DB 196,98,57,0,5,178,157,2,0 ; vpshufb 0x29db2(%rip),%xmm8,%xmm8 # 2d350 <_sk_overlay_sse2_8bit+0xea1> DB 196,3,121,21,4,24,0 ; vpextrw $0x0,%xmm8,(%r8,%r11,1) - DB 235,176 ; jmp 34f1 <_sk_store_a8_hsw+0x41> + DB 235,176 ; jmp 3557 <_sk_store_a8_hsw+0x40> DB 196,3,121,20,68,24,6,12 ; vpextrb $0xc,%xmm8,0x6(%r8,%r11,1) DB 196,3,121,20,68,24,5,10 ; vpextrb $0xa,%xmm8,0x5(%r8,%r11,1) DB 196,3,121,20,68,24,4,8 ; vpextrb $0x8,%xmm8,0x4(%r8,%r11,1) - DB 196,98,57,0,5,14,159,2,0 ; vpshufb 0x29f0e(%rip),%xmm8,%xmm8 # 2d470 <_sk_overlay_sse2_8bit+0xf18> + DB 196,98,57,0,5,152,157,2,0 ; vpshufb 0x29d98(%rip),%xmm8,%xmm8 # 2d360 <_sk_overlay_sse2_8bit+0xeb1> DB 196,1,121,126,4,24 ; vmovd %xmm8,(%r8,%r11,1) - DB 235,135 ; jmp 34f1 <_sk_store_a8_hsw+0x41> - DB 102,144 ; xchg %ax,%ax - DB 178,255 ; mov $0xff,%dl + DB 235,135 ; jmp 3557 <_sk_store_a8_hsw+0x40> + DB 180,255 ; mov $0xff,%ah DB 255 ; (bad) - DB 255,195 ; inc %ebx + DB 255,197 ; inc %ebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 187,255,255,255,237 ; mov $0xedffffff,%ebx + DB 189,255,255,255,239 ; mov $0xefffffff,%ebp DB 255 ; (bad) DB 255 ; (bad) - DB 255,229 ; jmpq *%rbp + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 221,255 ; (bad) + DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,213 ; callq *%rbp + DB 255,215 ; callq *%rdi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -3192,15 +3220,15 @@ _sk_load_g8_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,57 ; jne 35d9 <_sk_load_g8_hsw+0x51> + DB 117,57 ; jne 363d <_sk_load_g8_hsw+0x51> DB 196,130,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm0 - DB 197,249,219,5,162,158,2,0 ; vpand 0x29ea2(%rip),%xmm0,%xmm0 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 197,249,219,5,46,157,2,0 ; vpand 0x29d2e(%rip),%xmm0,%xmm0 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,96,157,2,0 ; vbroadcastss 0x29d60(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,220,155,2,0 ; vbroadcastss 0x29bdc(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,221,156,2,0 ; vbroadcastss 0x29cdd(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,89,155,2,0 ; vbroadcastss 0x29b59(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax @@ -3209,15 +3237,15 @@ _sk_load_g8_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,185 ; ja 35a6 <_sk_load_g8_hsw+0x1e> + DB 119,185 ; ja 360a <_sk_load_g8_hsw+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # 3670 <_sk_load_g8_hsw+0xe8> + DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # 36d4 <_sk_load_g8_hsw+0xe8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 235,154 ; jmp 35a6 <_sk_load_g8_hsw+0x1e> + DB 235,154 ; jmp 360a <_sk_load_g8_hsw+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0 @@ -3225,7 +3253,7 @@ _sk_load_g8_hsw LABEL PROC DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,121,2,193,1 ; vpblendd $0x1,%xmm1,%xmm0,%xmm0 - DB 233,114,255,255,255 ; jmpq 35a6 <_sk_load_g8_hsw+0x1e> + DB 233,114,255,255,255 ; jmpq 360a <_sk_load_g8_hsw+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0 @@ -3236,7 +3264,7 @@ _sk_load_g8_hsw LABEL PROC DB 196,129,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,113,2,192,12 ; vpblendd $0xc,%xmm0,%xmm1,%xmm0 - DB 233,55,255,255,255 ; jmpq 35a6 <_sk_load_g8_hsw+0x1e> + DB 233,55,255,255,255 ; jmpq 360a <_sk_load_g8_hsw+0x1e> DB 144 ; nop DB 145 ; xchg %eax,%ecx DB 255 ; (bad) @@ -3245,7 +3273,7 @@ _sk_load_g8_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf003680 <_sk_overlay_sse2_8bit+0xffffffffdefd7128> + DB 233,255,255,255,222 ; jmpq ffffffffdf0036e4 <_sk_overlay_sse2_8bit+0xffffffffdefd7235> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -3265,15 +3293,15 @@ _sk_load_g8_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,57 ; jne 36dd <_sk_load_g8_dst_hsw+0x51> + DB 117,57 ; jne 3741 <_sk_load_g8_dst_hsw+0x51> DB 196,130,121,48,36,24 ; vpmovzxbw (%r8,%r11,1),%xmm4 - DB 197,217,219,37,158,157,2,0 ; vpand 0x29d9e(%rip),%xmm4,%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 197,217,219,37,42,156,2,0 ; vpand 0x29c2a(%rip),%xmm4,%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 196,226,125,51,228 ; vpmovzxwd %xmm4,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,92,156,2,0 ; vbroadcastss 0x29c5c(%rip),%ymm5 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,45,216,154,2,0 ; vbroadcastss 0x29ad8(%rip),%ymm5 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,220,89,229 ; vmulps %ymm5,%ymm4,%ymm4 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,61,217,155,2,0 ; vbroadcastss 0x29bd9(%rip),%ymm7 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,61,85,154,2,0 ; vbroadcastss 0x29a55(%rip),%ymm7 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,236 ; vmovaps %ymm4,%ymm5 DB 197,252,40,244 ; vmovaps %ymm4,%ymm6 DB 255,224 ; jmpq *%rax @@ -3282,15 +3310,15 @@ _sk_load_g8_dst_hsw LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,185 ; ja 36aa <_sk_load_g8_dst_hsw+0x1e> + DB 119,185 ; ja 370e <_sk_load_g8_dst_hsw+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # 3774 <_sk_load_g8_dst_hsw+0xe8> + DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # 37d8 <_sk_load_g8_dst_hsw+0xe8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 235,154 ; jmp 36aa <_sk_load_g8_dst_hsw+0x1e> + DB 235,154 ; jmp 370e <_sk_load_g8_dst_hsw+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,2 ; vpinsrw $0x2,%eax,%xmm4,%xmm4 @@ -3298,7 +3326,7 @@ _sk_load_g8_dst_hsw LABEL PROC DB 197,249,110,232 ; vmovd %eax,%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,89,2,229,1 ; vpblendd $0x1,%xmm5,%xmm4,%xmm4 - DB 233,114,255,255,255 ; jmpq 36aa <_sk_load_g8_dst_hsw+0x1e> + DB 233,114,255,255,255 ; jmpq 370e <_sk_load_g8_dst_hsw+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,6 ; vpinsrw $0x6,%eax,%xmm4,%xmm4 @@ -3309,7 +3337,7 @@ _sk_load_g8_dst_hsw LABEL PROC DB 196,129,121,110,44,24 ; vmovd (%r8,%r11,1),%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,81,2,228,12 ; vpblendd $0xc,%xmm4,%xmm5,%xmm4 - DB 233,55,255,255,255 ; jmpq 36aa <_sk_load_g8_dst_hsw+0x1e> + DB 233,55,255,255,255 ; jmpq 370e <_sk_load_g8_dst_hsw+0x1e> DB 144 ; nop DB 145 ; xchg %eax,%ecx DB 255 ; (bad) @@ -3318,7 +3346,7 @@ _sk_load_g8_dst_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf003784 <_sk_overlay_sse2_8bit+0xffffffffdefd722c> + DB 233,255,255,255,222 ; jmpq ffffffffdf0037e8 <_sk_overlay_sse2_8bit+0xffffffffdefd7339> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -3355,8 +3383,10 @@ _sk_gather_g8_hsw LABEL PROC DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,131,121,32,4,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0 - DB 196,131,121,32,4,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 71,15,182,28,24 ; movzbl (%r8,%r11,1),%r11d + DB 196,193,121,110,195 ; vmovd %r11d,%xmm0 + DB 196,195,121,32,194,1 ; vpinsrb $0x1,%r10d,%xmm0,%xmm0 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax @@ -3371,10 +3401,10 @@ _sk_gather_g8_hsw LABEL PROC DB 196,227,121,32,192,7 ; vpinsrb $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,211,154,2,0 ; vbroadcastss 0x29ad3(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,72,153,2,0 ; vbroadcastss 0x29948(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,80,154,2,0 ; vbroadcastss 0x29a50(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,197,152,2,0 ; vbroadcastss 0x298c5(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx @@ -3393,69 +3423,72 @@ _sk_load_565_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,114 ; jne 38fa <_sk_load_565_hsw+0x8d> + DB 117,114 ; jne 3965 <_sk_load_565_hsw+0x8d> DB 196,129,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,136,154,2,0 ; vpbroadcastd 0x29a88(%rip),%ymm0 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,226,125,88,5,253,152,2,0 ; vpbroadcastd 0x298fd(%rip),%ymm0 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,123,154,2,0 ; vbroadcastss 0x29a7b(%rip),%ymm1 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,226,125,24,13,240,152,2,0 ; vbroadcastss 0x298f0(%rip),%ymm1 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,114,154,2,0 ; vpbroadcastd 0x29a72(%rip),%ymm1 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,226,125,88,13,231,152,2,0 ; vpbroadcastd 0x298e7(%rip),%ymm1 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,101,154,2,0 ; vbroadcastss 0x29a65(%rip),%ymm3 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,226,125,24,29,218,152,2,0 ; vbroadcastss 0x298da(%rip),%ymm3 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,92,154,2,0 ; vpbroadcastd 0x29a5c(%rip),%ymm3 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,88,29,209,152,2,0 ; vpbroadcastd 0x298d1(%rip),%ymm3 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,79,154,2,0 ; vbroadcastss 0x29a4f(%rip),%ymm3 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,226,125,24,29,196,152,2,0 ; vbroadcastss 0x298c4(%rip),%ymm3 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,180,153,2,0 ; vbroadcastss 0x299b4(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,41,152,2,0 ; vbroadcastss 0x29829(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,128 ; ja 388e <_sk_load_565_hsw+0x21> + DB 119,128 ; ja 38f9 <_sk_load_565_hsw+0x21> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,99,0,0,0 ; lea 0x63(%rip),%r10 # 397c <_sk_load_565_hsw+0x10f> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 39e8 <_sk_load_565_hsw+0x110> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 233,94,255,255,255 ; jmpq 388e <_sk_load_565_hsw+0x21> + DB 233,94,255,255,255 ; jmpq 38f9 <_sk_load_565_hsw+0x21> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,110,12,88 ; vmovd (%r8,%r11,2),%xmm1 DB 196,227,121,2,193,1 ; vpblendd $0x1,%xmm1,%xmm0,%xmm0 - DB 233,65,255,255,255 ; jmpq 388e <_sk_load_565_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 38f9 <_sk_load_565_hsw+0x21> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,122,126,12,88 ; vmovq (%r8,%r11,2),%xmm1 DB 196,227,113,2,192,12 ; vpblendd $0xc,%xmm0,%xmm1,%xmm0 - DB 233,20,255,255,255 ; jmpq 388e <_sk_load_565_hsw+0x21> - DB 102,144 ; xchg %ax,%ax - DB 166 ; cmpsb %es:(%rdi),%ds:(%rsi) + DB 233,20,255,255,255 ; jmpq 38f9 <_sk_load_565_hsw+0x21> + DB 15,31,0 ; nopl (%rax) + DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) - DB 255,180,255,255,255,237,255 ; pushq -0x120001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,229 ; jmpq *%rbp DB 255 ; (bad) + DB 236 ; in (%dx),%al DB 255 ; (bad) DB 255 ; (bad) - DB 221,255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255 ; (bad) + DB 255 ; (bad) + DB 220,255 ; fdivr %st,%st(7) + DB 255 ; (bad) + DB 255,208 ; callq *%rax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -3470,53 +3503,53 @@ _sk_load_565_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,114 ; jne 3a25 <_sk_load_565_dst_hsw+0x8d> + DB 117,114 ; jne 3a91 <_sk_load_565_dst_hsw+0x8d> DB 196,129,122,111,36,88 ; vmovdqu (%r8,%r11,2),%xmm4 DB 196,226,125,51,244 ; vpmovzxwd %xmm4,%ymm6 - DB 196,226,125,88,37,93,153,2,0 ; vpbroadcastd 0x2995d(%rip),%ymm4 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,226,125,88,37,209,151,2,0 ; vpbroadcastd 0x297d1(%rip),%ymm4 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,205,219,228 ; vpand %ymm4,%ymm6,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,80,153,2,0 ; vbroadcastss 0x29950(%rip),%ymm5 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,226,125,24,45,196,151,2,0 ; vbroadcastss 0x297c4(%rip),%ymm5 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 197,220,89,229 ; vmulps %ymm5,%ymm4,%ymm4 - DB 196,226,125,88,45,71,153,2,0 ; vpbroadcastd 0x29947(%rip),%ymm5 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,226,125,88,45,187,151,2,0 ; vpbroadcastd 0x297bb(%rip),%ymm5 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 197,205,219,237 ; vpand %ymm5,%ymm6,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,125,24,61,58,153,2,0 ; vbroadcastss 0x2993a(%rip),%ymm7 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,226,125,24,61,174,151,2,0 ; vbroadcastss 0x297ae(%rip),%ymm7 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 197,212,89,239 ; vmulps %ymm7,%ymm5,%ymm5 - DB 196,226,125,88,61,49,153,2,0 ; vpbroadcastd 0x29931(%rip),%ymm7 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,88,61,165,151,2,0 ; vpbroadcastd 0x297a5(%rip),%ymm7 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,205,219,247 ; vpand %ymm7,%ymm6,%ymm6 DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 - DB 196,226,125,24,61,36,153,2,0 ; vbroadcastss 0x29924(%rip),%ymm7 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,226,125,24,61,152,151,2,0 ; vbroadcastss 0x29798(%rip),%ymm7 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 197,204,89,247 ; vmulps %ymm7,%ymm6,%ymm6 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,61,137,152,2,0 ; vbroadcastss 0x29889(%rip),%ymm7 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,61,253,150,2,0 ; vbroadcastss 0x296fd(%rip),%ymm7 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,128 ; ja 39b9 <_sk_load_565_dst_hsw+0x21> + DB 119,128 ; ja 3a25 <_sk_load_565_dst_hsw+0x21> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 3aa8 <_sk_load_565_dst_hsw+0x110> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 3b14 <_sk_load_565_dst_hsw+0x110> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 233,94,255,255,255 ; jmpq 39b9 <_sk_load_565_dst_hsw+0x21> + DB 233,94,255,255,255 ; jmpq 3a25 <_sk_load_565_dst_hsw+0x21> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,121,110,44,88 ; vmovd (%r8,%r11,2),%xmm5 DB 196,227,89,2,229,1 ; vpblendd $0x1,%xmm5,%xmm4,%xmm4 - DB 233,65,255,255,255 ; jmpq 39b9 <_sk_load_565_dst_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 3a25 <_sk_load_565_dst_hsw+0x21> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,122,126,44,88 ; vmovq (%r8,%r11,2),%xmm5 DB 196,227,81,2,228,12 ; vpblendd $0xc,%xmm4,%xmm5,%xmm4 - DB 233,20,255,255,255 ; jmpq 39b9 <_sk_load_565_dst_hsw+0x21> + DB 233,20,255,255,255 ; jmpq 3a25 <_sk_load_565_dst_hsw+0x21> DB 15,31,0 ; nopl (%rax) DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) @@ -3583,23 +3616,23 @@ _sk_gather_565_hsw LABEL PROC DB 65,15,183,4,88 ; movzwl (%r8,%rbx,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 - DB 196,226,125,88,5,165,151,2,0 ; vpbroadcastd 0x297a5(%rip),%ymm0 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,226,125,88,5,25,150,2,0 ; vpbroadcastd 0x29619(%rip),%ymm0 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,237,219,192 ; vpand %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,152,151,2,0 ; vbroadcastss 0x29798(%rip),%ymm1 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,226,125,24,13,12,150,2,0 ; vbroadcastss 0x2960c(%rip),%ymm1 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,143,151,2,0 ; vpbroadcastd 0x2978f(%rip),%ymm1 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,226,125,88,13,3,150,2,0 ; vpbroadcastd 0x29603(%rip),%ymm1 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,130,151,2,0 ; vbroadcastss 0x29782(%rip),%ymm3 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,226,125,24,29,246,149,2,0 ; vbroadcastss 0x295f6(%rip),%ymm3 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,88,29,121,151,2,0 ; vpbroadcastd 0x29779(%rip),%ymm3 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,88,29,237,149,2,0 ; vpbroadcastd 0x295ed(%rip),%ymm3 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,237,219,211 ; vpand %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,108,151,2,0 ; vbroadcastss 0x2976c(%rip),%ymm3 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,226,125,24,29,224,149,2,0 ; vbroadcastss 0x295e0(%rip),%ymm3 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,209,150,2,0 ; vbroadcastss 0x296d1(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,69,149,2,0 ; vbroadcastss 0x29545(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 @@ -3615,11 +3648,11 @@ _sk_store_565_hsw LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,149,151,2,0 ; vbroadcastss 0x29795(%rip),%ymm8 # 2d398 <_sk_overlay_sse2_8bit+0xe40> + DB 196,98,125,24,5,9,150,2,0 ; vbroadcastss 0x29609(%rip),%ymm8 # 2d278 <_sk_overlay_sse2_8bit+0xdc9> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,53,114,241,11 ; vpslld $0xb,%ymm9,%ymm9 - DB 196,98,125,24,21,128,151,2,0 ; vbroadcastss 0x29780(%rip),%ymm10 # 2d39c <_sk_overlay_sse2_8bit+0xe44> + DB 196,98,125,24,21,244,149,2,0 ; vbroadcastss 0x295f4(%rip),%ymm10 # 2d27c <_sk_overlay_sse2_8bit+0xdcd> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,45,114,242,5 ; vpslld $0x5,%ymm10,%ymm10 @@ -3630,7 +3663,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 3c5a <_sk_store_565_hsw+0x76> + DB 117,10 ; jne 3cc6 <_sk_store_565_hsw+0x76> DB 196,1,122,127,4,88 ; vmovdqu %xmm8,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3638,22 +3671,22 @@ _sk_store_565_hsw LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 3c56 <_sk_store_565_hsw+0x72> + DB 119,236 ; ja 3cc2 <_sk_store_565_hsw+0x72> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,67,0,0,0 ; lea 0x43(%rip),%r10 # 3cb8 <_sk_store_565_hsw+0xd4> + DB 76,141,21,67,0,0,0 ; lea 0x43(%rip),%r10 # 3d24 <_sk_store_565_hsw+0xd4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,3,121,21,4,88,0 ; vpextrw $0x0,%xmm8,(%r8,%r11,2) - DB 235,207 ; jmp 3c56 <_sk_store_565_hsw+0x72> + DB 235,207 ; jmp 3cc2 <_sk_store_565_hsw+0x72> DB 196,3,121,21,68,88,4,2 ; vpextrw $0x2,%xmm8,0x4(%r8,%r11,2) DB 196,1,121,126,4,88 ; vmovd %xmm8,(%r8,%r11,2) - DB 235,191 ; jmp 3c56 <_sk_store_565_hsw+0x72> + DB 235,191 ; jmp 3cc2 <_sk_store_565_hsw+0x72> DB 196,3,121,21,68,88,12,6 ; vpextrw $0x6,%xmm8,0xc(%r8,%r11,2) DB 196,3,121,21,68,88,10,5 ; vpextrw $0x5,%xmm8,0xa(%r8,%r11,2) DB 196,3,121,21,68,88,8,4 ; vpextrw $0x4,%xmm8,0x8(%r8,%r11,2) DB 196,1,121,214,4,88 ; vmovq %xmm8,(%r8,%r11,2) - DB 235,159 ; jmp 3c56 <_sk_store_565_hsw+0x72> + DB 235,159 ; jmp 3cc2 <_sk_store_565_hsw+0x72> DB 144 ; nop DB 198 ; (bad) DB 255 ; (bad) @@ -3689,28 +3722,28 @@ _sk_load_4444_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,138,0,0,0 ; jne 3d7d <_sk_load_4444_hsw+0xa9> + DB 15,133,138,0,0,0 ; jne 3de9 <_sk_load_4444_hsw+0xa9> DB 196,129,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,153,150,2,0 ; vpbroadcastd 0x29699(%rip),%ymm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xe48> + DB 196,226,125,88,5,13,149,2,0 ; vpbroadcastd 0x2950d(%rip),%ymm0 # 2d280 <_sk_overlay_sse2_8bit+0xdd1> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,140,150,2,0 ; vbroadcastss 0x2968c(%rip),%ymm1 # 2d3a4 <_sk_overlay_sse2_8bit+0xe4c> + DB 196,226,125,24,13,0,149,2,0 ; vbroadcastss 0x29500(%rip),%ymm1 # 2d284 <_sk_overlay_sse2_8bit+0xdd5> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,131,150,2,0 ; vpbroadcastd 0x29683(%rip),%ymm1 # 2d3a8 <_sk_overlay_sse2_8bit+0xe50> + DB 196,226,125,88,13,247,148,2,0 ; vpbroadcastd 0x294f7(%rip),%ymm1 # 2d288 <_sk_overlay_sse2_8bit+0xdd9> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,118,150,2,0 ; vbroadcastss 0x29676(%rip),%ymm2 # 2d3ac <_sk_overlay_sse2_8bit+0xe54> + DB 196,226,125,24,21,234,148,2,0 ; vbroadcastss 0x294ea(%rip),%ymm2 # 2d28c <_sk_overlay_sse2_8bit+0xddd> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,109,150,2,0 ; vpbroadcastd 0x2966d(%rip),%ymm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xe58> + DB 196,226,125,88,21,225,148,2,0 ; vpbroadcastd 0x294e1(%rip),%ymm2 # 2d290 <_sk_overlay_sse2_8bit+0xde1> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,96,150,2,0 ; vbroadcastss 0x29660(%rip),%ymm8 # 2d3b4 <_sk_overlay_sse2_8bit+0xe5c> + DB 196,98,125,24,5,212,148,2,0 ; vbroadcastss 0x294d4(%rip),%ymm8 # 2d294 <_sk_overlay_sse2_8bit+0xde5> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,86,150,2,0 ; vpbroadcastd 0x29656(%rip),%ymm8 # 2d3b8 <_sk_overlay_sse2_8bit+0xe60> + DB 196,98,125,88,5,202,148,2,0 ; vpbroadcastd 0x294ca(%rip),%ymm8 # 2d298 <_sk_overlay_sse2_8bit+0xde9> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,72,150,2,0 ; vbroadcastss 0x29648(%rip),%ymm8 # 2d3bc <_sk_overlay_sse2_8bit+0xe64> + DB 196,98,125,24,5,188,148,2,0 ; vbroadcastss 0x294bc(%rip),%ymm8 # 2d29c <_sk_overlay_sse2_8bit+0xded> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3719,27 +3752,27 @@ _sk_load_4444_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,100,255,255,255 ; ja 3cf9 <_sk_load_4444_hsw+0x25> + DB 15,135,100,255,255,255 ; ja 3d65 <_sk_load_4444_hsw+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 3e04 <_sk_load_4444_hsw+0x130> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 3e70 <_sk_load_4444_hsw+0x130> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 233,66,255,255,255 ; jmpq 3cf9 <_sk_load_4444_hsw+0x25> + DB 233,66,255,255,255 ; jmpq 3d65 <_sk_load_4444_hsw+0x25> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,110,12,88 ; vmovd (%r8,%r11,2),%xmm1 DB 196,227,121,2,193,1 ; vpblendd $0x1,%xmm1,%xmm0,%xmm0 - DB 233,37,255,255,255 ; jmpq 3cf9 <_sk_load_4444_hsw+0x25> + DB 233,37,255,255,255 ; jmpq 3d65 <_sk_load_4444_hsw+0x25> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,122,126,12,88 ; vmovq (%r8,%r11,2),%xmm1 DB 196,227,113,2,192,12 ; vpblendd $0xc,%xmm0,%xmm1,%xmm0 - DB 233,248,254,255,255 ; jmpq 3cf9 <_sk_load_4444_hsw+0x25> + DB 233,248,254,255,255 ; jmpq 3d65 <_sk_load_4444_hsw+0x25> DB 15,31,0 ; nopl (%rax) DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) @@ -3773,28 +3806,28 @@ _sk_load_4444_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,138,0,0,0 ; jne 3ec9 <_sk_load_4444_dst_hsw+0xa9> + DB 15,133,138,0,0,0 ; jne 3f35 <_sk_load_4444_dst_hsw+0xa9> DB 196,129,122,111,36,88 ; vmovdqu (%r8,%r11,2),%xmm4 DB 196,226,125,51,252 ; vpmovzxwd %xmm4,%ymm7 - DB 196,226,125,88,37,77,149,2,0 ; vpbroadcastd 0x2954d(%rip),%ymm4 # 2d3a0 <_sk_overlay_sse2_8bit+0xe48> + DB 196,226,125,88,37,193,147,2,0 ; vpbroadcastd 0x293c1(%rip),%ymm4 # 2d280 <_sk_overlay_sse2_8bit+0xdd1> DB 197,197,219,228 ; vpand %ymm4,%ymm7,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,64,149,2,0 ; vbroadcastss 0x29540(%rip),%ymm5 # 2d3a4 <_sk_overlay_sse2_8bit+0xe4c> + DB 196,226,125,24,45,180,147,2,0 ; vbroadcastss 0x293b4(%rip),%ymm5 # 2d284 <_sk_overlay_sse2_8bit+0xdd5> DB 197,220,89,229 ; vmulps %ymm5,%ymm4,%ymm4 - DB 196,226,125,88,45,55,149,2,0 ; vpbroadcastd 0x29537(%rip),%ymm5 # 2d3a8 <_sk_overlay_sse2_8bit+0xe50> + DB 196,226,125,88,45,171,147,2,0 ; vpbroadcastd 0x293ab(%rip),%ymm5 # 2d288 <_sk_overlay_sse2_8bit+0xdd9> DB 197,197,219,237 ; vpand %ymm5,%ymm7,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,125,24,53,42,149,2,0 ; vbroadcastss 0x2952a(%rip),%ymm6 # 2d3ac <_sk_overlay_sse2_8bit+0xe54> + DB 196,226,125,24,53,158,147,2,0 ; vbroadcastss 0x2939e(%rip),%ymm6 # 2d28c <_sk_overlay_sse2_8bit+0xddd> DB 197,212,89,238 ; vmulps %ymm6,%ymm5,%ymm5 - DB 196,226,125,88,53,33,149,2,0 ; vpbroadcastd 0x29521(%rip),%ymm6 # 2d3b0 <_sk_overlay_sse2_8bit+0xe58> + DB 196,226,125,88,53,149,147,2,0 ; vpbroadcastd 0x29395(%rip),%ymm6 # 2d290 <_sk_overlay_sse2_8bit+0xde1> DB 197,197,219,246 ; vpand %ymm6,%ymm7,%ymm6 DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 - DB 196,98,125,24,5,20,149,2,0 ; vbroadcastss 0x29514(%rip),%ymm8 # 2d3b4 <_sk_overlay_sse2_8bit+0xe5c> + DB 196,98,125,24,5,136,147,2,0 ; vbroadcastss 0x29388(%rip),%ymm8 # 2d294 <_sk_overlay_sse2_8bit+0xde5> DB 196,193,76,89,240 ; vmulps %ymm8,%ymm6,%ymm6 - DB 196,98,125,88,5,10,149,2,0 ; vpbroadcastd 0x2950a(%rip),%ymm8 # 2d3b8 <_sk_overlay_sse2_8bit+0xe60> + DB 196,98,125,88,5,126,147,2,0 ; vpbroadcastd 0x2937e(%rip),%ymm8 # 2d298 <_sk_overlay_sse2_8bit+0xde9> DB 196,193,69,219,248 ; vpand %ymm8,%ymm7,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,252,148,2,0 ; vbroadcastss 0x294fc(%rip),%ymm8 # 2d3bc <_sk_overlay_sse2_8bit+0xe64> + DB 196,98,125,24,5,112,147,2,0 ; vbroadcastss 0x29370(%rip),%ymm8 # 2d29c <_sk_overlay_sse2_8bit+0xded> DB 196,193,68,89,248 ; vmulps %ymm8,%ymm7,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3803,27 +3836,27 @@ _sk_load_4444_dst_hsw LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,100,255,255,255 ; ja 3e45 <_sk_load_4444_dst_hsw+0x25> + DB 15,135,100,255,255,255 ; ja 3eb1 <_sk_load_4444_dst_hsw+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 3f50 <_sk_load_4444_dst_hsw+0x130> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 3fbc <_sk_load_4444_dst_hsw+0x130> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 233,66,255,255,255 ; jmpq 3e45 <_sk_load_4444_dst_hsw+0x25> + DB 233,66,255,255,255 ; jmpq 3eb1 <_sk_load_4444_dst_hsw+0x25> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,121,110,44,88 ; vmovd (%r8,%r11,2),%xmm5 DB 196,227,89,2,229,1 ; vpblendd $0x1,%xmm5,%xmm4,%xmm4 - DB 233,37,255,255,255 ; jmpq 3e45 <_sk_load_4444_dst_hsw+0x25> + DB 233,37,255,255,255 ; jmpq 3eb1 <_sk_load_4444_dst_hsw+0x25> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,122,126,44,88 ; vmovq (%r8,%r11,2),%xmm5 DB 196,227,81,2,228,12 ; vpblendd $0xc,%xmm4,%xmm5,%xmm4 - DB 233,248,254,255,255 ; jmpq 3e45 <_sk_load_4444_dst_hsw+0x25> + DB 233,248,254,255,255 ; jmpq 3eb1 <_sk_load_4444_dst_hsw+0x25> DB 15,31,0 ; nopl (%rax) DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) @@ -3890,25 +3923,25 @@ _sk_gather_4444_hsw LABEL PROC DB 65,15,183,4,88 ; movzwl (%r8,%rbx,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 DB 196,226,125,51,216 ; vpmovzxwd %xmm0,%ymm3 - DB 196,226,125,88,5,121,147,2,0 ; vpbroadcastd 0x29379(%rip),%ymm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xe48> + DB 196,226,125,88,5,237,145,2,0 ; vpbroadcastd 0x291ed(%rip),%ymm0 # 2d280 <_sk_overlay_sse2_8bit+0xdd1> DB 197,229,219,192 ; vpand %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,108,147,2,0 ; vbroadcastss 0x2936c(%rip),%ymm1 # 2d3a4 <_sk_overlay_sse2_8bit+0xe4c> + DB 196,226,125,24,13,224,145,2,0 ; vbroadcastss 0x291e0(%rip),%ymm1 # 2d284 <_sk_overlay_sse2_8bit+0xdd5> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,88,13,99,147,2,0 ; vpbroadcastd 0x29363(%rip),%ymm1 # 2d3a8 <_sk_overlay_sse2_8bit+0xe50> + DB 196,226,125,88,13,215,145,2,0 ; vpbroadcastd 0x291d7(%rip),%ymm1 # 2d288 <_sk_overlay_sse2_8bit+0xdd9> DB 197,229,219,201 ; vpand %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,86,147,2,0 ; vbroadcastss 0x29356(%rip),%ymm2 # 2d3ac <_sk_overlay_sse2_8bit+0xe54> + DB 196,226,125,24,21,202,145,2,0 ; vbroadcastss 0x291ca(%rip),%ymm2 # 2d28c <_sk_overlay_sse2_8bit+0xddd> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,21,77,147,2,0 ; vpbroadcastd 0x2934d(%rip),%ymm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xe58> + DB 196,226,125,88,21,193,145,2,0 ; vpbroadcastd 0x291c1(%rip),%ymm2 # 2d290 <_sk_overlay_sse2_8bit+0xde1> DB 197,229,219,210 ; vpand %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,64,147,2,0 ; vbroadcastss 0x29340(%rip),%ymm8 # 2d3b4 <_sk_overlay_sse2_8bit+0xe5c> + DB 196,98,125,24,5,180,145,2,0 ; vbroadcastss 0x291b4(%rip),%ymm8 # 2d294 <_sk_overlay_sse2_8bit+0xde5> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,54,147,2,0 ; vpbroadcastd 0x29336(%rip),%ymm8 # 2d3b8 <_sk_overlay_sse2_8bit+0xe60> + DB 196,98,125,88,5,170,145,2,0 ; vpbroadcastd 0x291aa(%rip),%ymm8 # 2d298 <_sk_overlay_sse2_8bit+0xde9> DB 196,193,101,219,216 ; vpand %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,40,147,2,0 ; vbroadcastss 0x29328(%rip),%ymm8 # 2d3bc <_sk_overlay_sse2_8bit+0xe64> + DB 196,98,125,24,5,156,145,2,0 ; vbroadcastss 0x2919c(%rip),%ymm8 # 2d29c <_sk_overlay_sse2_8bit+0xded> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -3926,7 +3959,7 @@ _sk_store_4444_hsw LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,253,146,2,0 ; vbroadcastss 0x292fd(%rip),%ymm8 # 2d3c0 <_sk_overlay_sse2_8bit+0xe68> + DB 196,98,125,24,5,113,145,2,0 ; vbroadcastss 0x29171(%rip),%ymm8 # 2d2a0 <_sk_overlay_sse2_8bit+0xdf1> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,53,114,241,12 ; vpslld $0xc,%ymm9,%ymm9 @@ -3944,7 +3977,7 @@ _sk_store_4444_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 4126 <_sk_store_4444_hsw+0x82> + DB 117,10 ; jne 4192 <_sk_store_4444_hsw+0x82> DB 196,1,122,127,4,88 ; vmovdqu %xmm8,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -3952,22 +3985,22 @@ _sk_store_4444_hsw LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 4122 <_sk_store_4444_hsw+0x7e> + DB 119,236 ; ja 418e <_sk_store_4444_hsw+0x7e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,67,0,0,0 ; lea 0x43(%rip),%r10 # 4184 <_sk_store_4444_hsw+0xe0> + DB 76,141,21,67,0,0,0 ; lea 0x43(%rip),%r10 # 41f0 <_sk_store_4444_hsw+0xe0> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,3,121,21,4,88,0 ; vpextrw $0x0,%xmm8,(%r8,%r11,2) - DB 235,207 ; jmp 4122 <_sk_store_4444_hsw+0x7e> + DB 235,207 ; jmp 418e <_sk_store_4444_hsw+0x7e> DB 196,3,121,21,68,88,4,2 ; vpextrw $0x2,%xmm8,0x4(%r8,%r11,2) DB 196,1,121,126,4,88 ; vmovd %xmm8,(%r8,%r11,2) - DB 235,191 ; jmp 4122 <_sk_store_4444_hsw+0x7e> + DB 235,191 ; jmp 418e <_sk_store_4444_hsw+0x7e> DB 196,3,121,21,68,88,12,6 ; vpextrw $0x6,%xmm8,0xc(%r8,%r11,2) DB 196,3,121,21,68,88,10,5 ; vpextrw $0x5,%xmm8,0xa(%r8,%r11,2) DB 196,3,121,21,68,88,8,4 ; vpextrw $0x4,%xmm8,0x8(%r8,%r11,2) DB 196,1,121,214,4,88 ; vmovq %xmm8,(%r8,%r11,2) - DB 235,159 ; jmp 4122 <_sk_store_4444_hsw+0x7e> + DB 235,159 ; jmp 418e <_sk_store_4444_hsw+0x7e> DB 144 ; nop DB 198 ; (bad) DB 255 ; (bad) @@ -4003,16 +4036,16 @@ _sk_load_8888_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,86 ; jne 4212 <_sk_load_8888_hsw+0x72> + DB 117,86 ; jne 427e <_sk_load_8888_hsw+0x72> DB 196,129,126,111,28,152 ; vmovdqu (%r8,%r11,4),%ymm3 - DB 197,229,219,5,22,137,2,0 ; vpand 0x28916(%rip),%ymm3,%ymm0 # 2cae0 <_sk_overlay_sse2_8bit+0x588> + DB 197,229,219,5,170,135,2,0 ; vpand 0x287aa(%rip),%ymm3,%ymm0 # 2c9e0 <_sk_overlay_sse2_8bit+0x531> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,73,145,2,0 ; vbroadcastss 0x29149(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,189,143,2,0 ; vbroadcastss 0x28fbd(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,27,137,2,0 ; vpshufb 0x2891b(%rip),%ymm3,%ymm1 # 2cb00 <_sk_overlay_sse2_8bit+0x5a8> + DB 196,226,101,0,13,175,135,2,0 ; vpshufb 0x287af(%rip),%ymm3,%ymm1 # 2ca00 <_sk_overlay_sse2_8bit+0x551> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,41,137,2,0 ; vpshufb 0x28929(%rip),%ymm3,%ymm2 # 2cb20 <_sk_overlay_sse2_8bit+0x5c8> + DB 196,226,101,0,21,189,135,2,0 ; vpshufb 0x287bd(%rip),%ymm3,%ymm2 # 2ca20 <_sk_overlay_sse2_8bit+0x571> DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 @@ -4025,21 +4058,21 @@ _sk_load_8888_hsw LABEL PROC DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,156 ; ja 41c2 <_sk_load_8888_hsw+0x22> + DB 119,156 ; ja 422e <_sk_load_8888_hsw+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 42b8 <_sk_load_8888_hsw+0x118> + DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 4324 <_sk_load_8888_hsw+0x118> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,28,152 ; vmovd (%r8,%r11,4),%xmm3 - DB 235,128 ; jmp 41c2 <_sk_load_8888_hsw+0x22> + DB 235,128 ; jmp 422e <_sk_load_8888_hsw+0x22> DB 196,129,121,110,68,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 DB 196,227,117,2,216,4 ; vpblendd $0x4,%ymm0,%ymm1,%ymm3 DB 196,129,122,126,4,152 ; vmovq (%r8,%r11,4),%xmm0 DB 196,227,101,2,216,3 ; vpblendd $0x3,%ymm0,%ymm3,%ymm3 - DB 233,89,255,255,255 ; jmpq 41c2 <_sk_load_8888_hsw+0x22> + DB 233,89,255,255,255 ; jmpq 422e <_sk_load_8888_hsw+0x22> DB 196,129,121,110,68,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -4052,7 +4085,7 @@ _sk_load_8888_hsw LABEL PROC DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3 DB 196,129,122,111,4,152 ; vmovdqu (%r8,%r11,4),%xmm0 DB 196,227,125,2,219,240 ; vpblendd $0xf0,%ymm3,%ymm0,%ymm3 - DB 233,10,255,255,255 ; jmpq 41c2 <_sk_load_8888_hsw+0x22> + DB 233,10,255,255,255 ; jmpq 422e <_sk_load_8888_hsw+0x22> DB 130 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -4084,16 +4117,16 @@ _sk_load_8888_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,86 ; jne 4346 <_sk_load_8888_dst_hsw+0x72> + DB 117,86 ; jne 43b2 <_sk_load_8888_dst_hsw+0x72> DB 196,129,126,111,60,152 ; vmovdqu (%r8,%r11,4),%ymm7 - DB 197,197,219,37,66,136,2,0 ; vpand 0x28842(%rip),%ymm7,%ymm4 # 2cb40 <_sk_overlay_sse2_8bit+0x5e8> + DB 197,197,219,37,214,134,2,0 ; vpand 0x286d6(%rip),%ymm7,%ymm4 # 2ca40 <_sk_overlay_sse2_8bit+0x591> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,98,125,24,5,21,144,2,0 ; vbroadcastss 0x29015(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,137,142,2,0 ; vbroadcastss 0x28e89(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,92,89,224 ; vmulps %ymm8,%ymm4,%ymm4 - DB 196,226,69,0,45,71,136,2,0 ; vpshufb 0x28847(%rip),%ymm7,%ymm5 # 2cb60 <_sk_overlay_sse2_8bit+0x608> + DB 196,226,69,0,45,219,134,2,0 ; vpshufb 0x286db(%rip),%ymm7,%ymm5 # 2ca60 <_sk_overlay_sse2_8bit+0x5b1> DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 DB 196,193,84,89,232 ; vmulps %ymm8,%ymm5,%ymm5 - DB 196,226,69,0,53,85,136,2,0 ; vpshufb 0x28855(%rip),%ymm7,%ymm6 # 2cb80 <_sk_overlay_sse2_8bit+0x628> + DB 196,226,69,0,53,233,134,2,0 ; vpshufb 0x286e9(%rip),%ymm7,%ymm6 # 2ca80 <_sk_overlay_sse2_8bit+0x5d1> DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 DB 196,193,76,89,240 ; vmulps %ymm8,%ymm6,%ymm6 DB 197,197,114,215,24 ; vpsrld $0x18,%ymm7,%ymm7 @@ -4106,21 +4139,21 @@ _sk_load_8888_dst_hsw LABEL PROC DB 197,197,239,255 ; vpxor %ymm7,%ymm7,%ymm7 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,156 ; ja 42f6 <_sk_load_8888_dst_hsw+0x22> + DB 119,156 ; ja 4362 <_sk_load_8888_dst_hsw+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 43ec <_sk_load_8888_dst_hsw+0x118> + DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 4458 <_sk_load_8888_dst_hsw+0x118> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,60,152 ; vmovd (%r8,%r11,4),%xmm7 - DB 235,128 ; jmp 42f6 <_sk_load_8888_dst_hsw+0x22> + DB 235,128 ; jmp 4362 <_sk_load_8888_dst_hsw+0x22> DB 196,129,121,110,100,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm4 DB 196,226,121,89,228 ; vpbroadcastq %xmm4,%xmm4 DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 DB 196,227,85,2,252,4 ; vpblendd $0x4,%ymm4,%ymm5,%ymm7 DB 196,129,122,126,36,152 ; vmovq (%r8,%r11,4),%xmm4 DB 196,227,69,2,252,3 ; vpblendd $0x3,%ymm4,%ymm7,%ymm7 - DB 233,89,255,255,255 ; jmpq 42f6 <_sk_load_8888_dst_hsw+0x22> + DB 233,89,255,255,255 ; jmpq 4362 <_sk_load_8888_dst_hsw+0x22> DB 196,129,121,110,100,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm4 DB 196,226,125,89,228 ; vpbroadcastq %xmm4,%ymm4 DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 @@ -4133,7 +4166,7 @@ _sk_load_8888_dst_hsw LABEL PROC DB 196,227,69,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm7 DB 196,129,122,111,36,152 ; vmovdqu (%r8,%r11,4),%xmm4 DB 196,227,93,2,255,240 ; vpblendd $0xf0,%ymm7,%ymm4,%ymm7 - DB 233,10,255,255,255 ; jmpq 42f6 <_sk_load_8888_dst_hsw+0x22> + DB 233,10,255,255,255 ; jmpq 4362 <_sk_load_8888_dst_hsw+0x22> DB 130 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -4165,15 +4198,16 @@ _sk_gather_8888_hsw LABEL PROC DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 + DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 DB 196,194,117,144,28,128 ; vpgatherdd %ymm1,(%r8,%ymm0,4),%ymm3 - DB 197,229,219,5,106,135,2,0 ; vpand 0x2876a(%rip),%ymm3,%ymm0 # 2cba0 <_sk_overlay_sse2_8bit+0x648> + DB 197,229,219,5,250,133,2,0 ; vpand 0x285fa(%rip),%ymm3,%ymm0 # 2caa0 <_sk_overlay_sse2_8bit+0x5f1> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,221,142,2,0 ; vbroadcastss 0x28edd(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,77,141,2,0 ; vbroadcastss 0x28d4d(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,226,101,0,13,111,135,2,0 ; vpshufb 0x2876f(%rip),%ymm3,%ymm1 # 2cbc0 <_sk_overlay_sse2_8bit+0x668> + DB 196,226,101,0,13,255,133,2,0 ; vpshufb 0x285ff(%rip),%ymm3,%ymm1 # 2cac0 <_sk_overlay_sse2_8bit+0x611> DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,226,101,0,21,125,135,2,0 ; vpshufb 0x2877d(%rip),%ymm3,%ymm2 # 2cbe0 <_sk_overlay_sse2_8bit+0x688> + DB 196,226,101,0,21,13,134,2,0 ; vpshufb 0x2860d(%rip),%ymm3,%ymm2 # 2cae0 <_sk_overlay_sse2_8bit+0x631> DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 @@ -4191,7 +4225,7 @@ _sk_store_8888_hsw LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,58,142,2,0 ; vbroadcastss 0x28e3a(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,170,140,2,0 ; vbroadcastss 0x28caa(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -4207,7 +4241,7 @@ _sk_store_8888_hsw LABEL PROC DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 44f6 <_sk_store_8888_hsw+0x78> + DB 117,10 ; jne 4566 <_sk_store_8888_hsw+0x78> DB 196,1,126,127,4,152 ; vmovdqu %ymm8,(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -4215,17 +4249,17 @@ _sk_store_8888_hsw LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 44f2 <_sk_store_8888_hsw+0x74> + DB 119,236 ; ja 4562 <_sk_store_8888_hsw+0x74> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # 4564 <_sk_store_8888_hsw+0xe6> + DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # 45d4 <_sk_store_8888_hsw+0xe6> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,121,126,4,152 ; vmovd %xmm8,(%r8,%r11,4) - DB 235,208 ; jmp 44f2 <_sk_store_8888_hsw+0x74> + DB 235,208 ; jmp 4562 <_sk_store_8888_hsw+0x74> DB 196,3,121,22,68,152,8,2 ; vpextrd $0x2,%xmm8,0x8(%r8,%r11,4) DB 196,1,121,214,4,152 ; vmovq %xmm8,(%r8,%r11,4) - DB 235,192 ; jmp 44f2 <_sk_store_8888_hsw+0x74> + DB 235,192 ; jmp 4562 <_sk_store_8888_hsw+0x74> DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,3,121,22,76,152,24,2 ; vpextrd $0x2,%xmm9,0x18(%r8,%r11,4) DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 @@ -4233,7 +4267,7 @@ _sk_store_8888_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,1,121,126,76,152,16 ; vmovd %xmm9,0x10(%r8,%r11,4) DB 196,1,122,127,4,152 ; vmovdqu %xmm8,(%r8,%r11,4) - DB 235,143 ; jmp 44f2 <_sk_store_8888_hsw+0x74> + DB 235,143 ; jmp 4562 <_sk_store_8888_hsw+0x74> DB 144 ; nop DB 182,255 ; mov $0xff,%dh DB 255 ; (bad) @@ -4266,16 +4300,16 @@ _sk_load_bgra_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,86 ; jne 45f2 <_sk_load_bgra_hsw+0x72> + DB 117,86 ; jne 4662 <_sk_load_bgra_hsw+0x72> DB 196,129,126,111,28,152 ; vmovdqu (%r8,%r11,4),%ymm3 - DB 197,229,219,5,86,134,2,0 ; vpand 0x28656(%rip),%ymm3,%ymm0 # 2cc00 <_sk_overlay_sse2_8bit+0x6a8> + DB 197,229,219,5,230,132,2,0 ; vpand 0x284e6(%rip),%ymm3,%ymm0 # 2cb00 <_sk_overlay_sse2_8bit+0x651> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,105,141,2,0 ; vbroadcastss 0x28d69(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,217,139,2,0 ; vbroadcastss 0x28bd9(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,208 ; vmulps %ymm8,%ymm0,%ymm2 - DB 196,226,101,0,5,91,134,2,0 ; vpshufb 0x2865b(%rip),%ymm3,%ymm0 # 2cc20 <_sk_overlay_sse2_8bit+0x6c8> + DB 196,226,101,0,5,235,132,2,0 ; vpshufb 0x284eb(%rip),%ymm3,%ymm0 # 2cb20 <_sk_overlay_sse2_8bit+0x671> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 DB 196,193,124,89,200 ; vmulps %ymm8,%ymm0,%ymm1 - DB 196,226,101,0,5,105,134,2,0 ; vpshufb 0x28669(%rip),%ymm3,%ymm0 # 2cc40 <_sk_overlay_sse2_8bit+0x6e8> + DB 196,226,101,0,5,249,132,2,0 ; vpshufb 0x284f9(%rip),%ymm3,%ymm0 # 2cb40 <_sk_overlay_sse2_8bit+0x691> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 @@ -4288,21 +4322,21 @@ _sk_load_bgra_hsw LABEL PROC DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,156 ; ja 45a2 <_sk_load_bgra_hsw+0x22> + DB 119,156 ; ja 4612 <_sk_load_bgra_hsw+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 4698 <_sk_load_bgra_hsw+0x118> + DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 4708 <_sk_load_bgra_hsw+0x118> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,28,152 ; vmovd (%r8,%r11,4),%xmm3 - DB 235,128 ; jmp 45a2 <_sk_load_bgra_hsw+0x22> + DB 235,128 ; jmp 4612 <_sk_load_bgra_hsw+0x22> DB 196,129,121,110,68,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 DB 196,227,117,2,216,4 ; vpblendd $0x4,%ymm0,%ymm1,%ymm3 DB 196,129,122,126,4,152 ; vmovq (%r8,%r11,4),%xmm0 DB 196,227,101,2,216,3 ; vpblendd $0x3,%ymm0,%ymm3,%ymm3 - DB 233,89,255,255,255 ; jmpq 45a2 <_sk_load_bgra_hsw+0x22> + DB 233,89,255,255,255 ; jmpq 4612 <_sk_load_bgra_hsw+0x22> DB 196,129,121,110,68,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -4315,7 +4349,7 @@ _sk_load_bgra_hsw LABEL PROC DB 196,227,101,56,216,1 ; vinserti128 $0x1,%xmm0,%ymm3,%ymm3 DB 196,129,122,111,4,152 ; vmovdqu (%r8,%r11,4),%xmm0 DB 196,227,125,2,219,240 ; vpblendd $0xf0,%ymm3,%ymm0,%ymm3 - DB 233,10,255,255,255 ; jmpq 45a2 <_sk_load_bgra_hsw+0x22> + DB 233,10,255,255,255 ; jmpq 4612 <_sk_load_bgra_hsw+0x22> DB 130 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -4347,16 +4381,16 @@ _sk_load_bgra_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,86 ; jne 4726 <_sk_load_bgra_dst_hsw+0x72> + DB 117,86 ; jne 4796 <_sk_load_bgra_dst_hsw+0x72> DB 196,129,126,111,60,152 ; vmovdqu (%r8,%r11,4),%ymm7 - DB 197,197,219,37,130,133,2,0 ; vpand 0x28582(%rip),%ymm7,%ymm4 # 2cc60 <_sk_overlay_sse2_8bit+0x708> + DB 197,197,219,37,18,132,2,0 ; vpand 0x28412(%rip),%ymm7,%ymm4 # 2cb60 <_sk_overlay_sse2_8bit+0x6b1> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,98,125,24,5,53,140,2,0 ; vbroadcastss 0x28c35(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,165,138,2,0 ; vbroadcastss 0x28aa5(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,92,89,240 ; vmulps %ymm8,%ymm4,%ymm6 - DB 196,226,69,0,37,135,133,2,0 ; vpshufb 0x28587(%rip),%ymm7,%ymm4 # 2cc80 <_sk_overlay_sse2_8bit+0x728> + DB 196,226,69,0,37,23,132,2,0 ; vpshufb 0x28417(%rip),%ymm7,%ymm4 # 2cb80 <_sk_overlay_sse2_8bit+0x6d1> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 196,193,92,89,232 ; vmulps %ymm8,%ymm4,%ymm5 - DB 196,226,69,0,37,149,133,2,0 ; vpshufb 0x28595(%rip),%ymm7,%ymm4 # 2cca0 <_sk_overlay_sse2_8bit+0x748> + DB 196,226,69,0,37,37,132,2,0 ; vpshufb 0x28425(%rip),%ymm7,%ymm4 # 2cba0 <_sk_overlay_sse2_8bit+0x6f1> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 196,193,92,89,224 ; vmulps %ymm8,%ymm4,%ymm4 DB 197,197,114,215,24 ; vpsrld $0x18,%ymm7,%ymm7 @@ -4369,21 +4403,21 @@ _sk_load_bgra_dst_hsw LABEL PROC DB 197,197,239,255 ; vpxor %ymm7,%ymm7,%ymm7 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,156 ; ja 46d6 <_sk_load_bgra_dst_hsw+0x22> + DB 119,156 ; ja 4746 <_sk_load_bgra_dst_hsw+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 47cc <_sk_load_bgra_dst_hsw+0x118> + DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 483c <_sk_load_bgra_dst_hsw+0x118> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,60,152 ; vmovd (%r8,%r11,4),%xmm7 - DB 235,128 ; jmp 46d6 <_sk_load_bgra_dst_hsw+0x22> + DB 235,128 ; jmp 4746 <_sk_load_bgra_dst_hsw+0x22> DB 196,129,121,110,100,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm4 DB 196,226,121,89,228 ; vpbroadcastq %xmm4,%xmm4 DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 DB 196,227,85,2,252,4 ; vpblendd $0x4,%ymm4,%ymm5,%ymm7 DB 196,129,122,126,36,152 ; vmovq (%r8,%r11,4),%xmm4 DB 196,227,69,2,252,3 ; vpblendd $0x3,%ymm4,%ymm7,%ymm7 - DB 233,89,255,255,255 ; jmpq 46d6 <_sk_load_bgra_dst_hsw+0x22> + DB 233,89,255,255,255 ; jmpq 4746 <_sk_load_bgra_dst_hsw+0x22> DB 196,129,121,110,100,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm4 DB 196,226,125,89,228 ; vpbroadcastq %xmm4,%ymm4 DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 @@ -4396,7 +4430,7 @@ _sk_load_bgra_dst_hsw LABEL PROC DB 196,227,69,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm7 DB 196,129,122,111,36,152 ; vmovdqu (%r8,%r11,4),%xmm4 DB 196,227,93,2,255,240 ; vpblendd $0xf0,%ymm7,%ymm4,%ymm7 - DB 233,10,255,255,255 ; jmpq 46d6 <_sk_load_bgra_dst_hsw+0x22> + DB 233,10,255,255,255 ; jmpq 4746 <_sk_load_bgra_dst_hsw+0x22> DB 130 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -4428,15 +4462,16 @@ _sk_gather_bgra_hsw LABEL PROC DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 + DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 DB 196,194,117,144,28,128 ; vpgatherdd %ymm1,(%r8,%ymm0,4),%ymm3 - DB 197,229,219,5,170,132,2,0 ; vpand 0x284aa(%rip),%ymm3,%ymm0 # 2ccc0 <_sk_overlay_sse2_8bit+0x768> + DB 197,229,219,5,54,131,2,0 ; vpand 0x28336(%rip),%ymm3,%ymm0 # 2cbc0 <_sk_overlay_sse2_8bit+0x711> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,253,138,2,0 ; vbroadcastss 0x28afd(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,105,137,2,0 ; vbroadcastss 0x28969(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,208 ; vmulps %ymm8,%ymm0,%ymm2 - DB 196,226,101,0,5,175,132,2,0 ; vpshufb 0x284af(%rip),%ymm3,%ymm0 # 2cce0 <_sk_overlay_sse2_8bit+0x788> + DB 196,226,101,0,5,59,131,2,0 ; vpshufb 0x2833b(%rip),%ymm3,%ymm0 # 2cbe0 <_sk_overlay_sse2_8bit+0x731> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 DB 196,193,124,89,200 ; vmulps %ymm8,%ymm0,%ymm1 - DB 196,226,101,0,5,189,132,2,0 ; vpshufb 0x284bd(%rip),%ymm3,%ymm0 # 2cd00 <_sk_overlay_sse2_8bit+0x7a8> + DB 196,226,101,0,5,73,131,2,0 ; vpshufb 0x28349(%rip),%ymm3,%ymm0 # 2cc00 <_sk_overlay_sse2_8bit+0x751> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 @@ -4454,7 +4489,7 @@ _sk_store_bgra_hsw LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,90,138,2,0 ; vbroadcastss 0x28a5a(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,198,136,2,0 ; vbroadcastss 0x288c6(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,65,108,89,200 ; vmulps %ymm8,%ymm2,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -4470,7 +4505,7 @@ _sk_store_bgra_hsw LABEL PROC DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 48d6 <_sk_store_bgra_hsw+0x78> + DB 117,10 ; jne 494a <_sk_store_bgra_hsw+0x78> DB 196,1,126,127,4,152 ; vmovdqu %ymm8,(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -4478,17 +4513,17 @@ _sk_store_bgra_hsw LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 48d2 <_sk_store_bgra_hsw+0x74> + DB 119,236 ; ja 4946 <_sk_store_bgra_hsw+0x74> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # 4944 <_sk_store_bgra_hsw+0xe6> + DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # 49b8 <_sk_store_bgra_hsw+0xe6> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,121,126,4,152 ; vmovd %xmm8,(%r8,%r11,4) - DB 235,208 ; jmp 48d2 <_sk_store_bgra_hsw+0x74> + DB 235,208 ; jmp 4946 <_sk_store_bgra_hsw+0x74> DB 196,3,121,22,68,152,8,2 ; vpextrd $0x2,%xmm8,0x8(%r8,%r11,4) DB 196,1,121,214,4,152 ; vmovq %xmm8,(%r8,%r11,4) - DB 235,192 ; jmp 48d2 <_sk_store_bgra_hsw+0x74> + DB 235,192 ; jmp 4946 <_sk_store_bgra_hsw+0x74> DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,3,121,22,76,152,24,2 ; vpextrd $0x2,%xmm9,0x18(%r8,%r11,4) DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 @@ -4496,7 +4531,7 @@ _sk_store_bgra_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,1,121,126,76,152,16 ; vmovd %xmm9,0x10(%r8,%r11,4) DB 196,1,122,127,4,152 ; vmovdqu %xmm8,(%r8,%r11,4) - DB 235,143 ; jmp 48d2 <_sk_store_bgra_hsw+0x74> + DB 235,143 ; jmp 4946 <_sk_store_bgra_hsw+0x74> DB 144 ; nop DB 182,255 ; mov $0xff,%dh DB 255 ; (bad) @@ -4529,7 +4564,7 @@ _sk_load_f16_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,101 ; jne 49e1 <_sk_load_f16_hsw+0x81> + DB 117,101 ; jne 4a55 <_sk_load_f16_hsw+0x81> DB 196,65,121,16,4,192 ; vmovupd (%r8,%rax,8),%xmm8 DB 196,193,121,16,84,192,16 ; vmovupd 0x10(%r8,%rax,8),%xmm2 DB 196,193,121,16,92,192,32 ; vmovupd 0x20(%r8,%rax,8),%xmm3 @@ -4555,29 +4590,29 @@ _sk_load_f16_hsw LABEL PROC DB 196,65,123,16,4,192 ; vmovsd (%r8,%rax,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je 4a47 <_sk_load_f16_hsw+0xe7> + DB 116,85 ; je 4abb <_sk_load_f16_hsw+0xe7> DB 196,65,57,22,68,192,8 ; vmovhpd 0x8(%r8,%rax,8),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb 4a47 <_sk_load_f16_hsw+0xe7> + DB 114,72 ; jb 4abb <_sk_load_f16_hsw+0xe7> DB 196,193,123,16,84,192,16 ; vmovsd 0x10(%r8,%rax,8),%xmm2 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je 4a54 <_sk_load_f16_hsw+0xf4> + DB 116,72 ; je 4ac8 <_sk_load_f16_hsw+0xf4> DB 196,193,105,22,84,192,24 ; vmovhpd 0x18(%r8,%rax,8),%xmm2,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb 4a54 <_sk_load_f16_hsw+0xf4> + DB 114,59 ; jb 4ac8 <_sk_load_f16_hsw+0xf4> DB 196,193,123,16,92,192,32 ; vmovsd 0x20(%r8,%rax,8),%xmm3 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,109,255,255,255 ; je 4997 <_sk_load_f16_hsw+0x37> + DB 15,132,109,255,255,255 ; je 4a0b <_sk_load_f16_hsw+0x37> DB 196,193,97,22,92,192,40 ; vmovhpd 0x28(%r8,%rax,8),%xmm3,%xmm3 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,92,255,255,255 ; jb 4997 <_sk_load_f16_hsw+0x37> + DB 15,130,92,255,255,255 ; jb 4a0b <_sk_load_f16_hsw+0x37> DB 196,65,122,126,76,192,48 ; vmovq 0x30(%r8,%rax,8),%xmm9 - DB 233,80,255,255,255 ; jmpq 4997 <_sk_load_f16_hsw+0x37> + DB 233,80,255,255,255 ; jmpq 4a0b <_sk_load_f16_hsw+0x37> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,67,255,255,255 ; jmpq 4997 <_sk_load_f16_hsw+0x37> + DB 233,67,255,255,255 ; jmpq 4a0b <_sk_load_f16_hsw+0x37> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,58,255,255,255 ; jmpq 4997 <_sk_load_f16_hsw+0x37> + DB 233,58,255,255,255 ; jmpq 4a0b <_sk_load_f16_hsw+0x37> PUBLIC _sk_load_f16_dst_hsw _sk_load_f16_dst_hsw LABEL PROC @@ -4589,7 +4624,7 @@ _sk_load_f16_dst_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,101 ; jne 4ade <_sk_load_f16_dst_hsw+0x81> + DB 117,101 ; jne 4b52 <_sk_load_f16_dst_hsw+0x81> DB 196,65,121,16,4,192 ; vmovupd (%r8,%rax,8),%xmm8 DB 196,193,121,16,116,192,16 ; vmovupd 0x10(%r8,%rax,8),%xmm6 DB 196,193,121,16,124,192,32 ; vmovupd 0x20(%r8,%rax,8),%xmm7 @@ -4615,29 +4650,29 @@ _sk_load_f16_dst_hsw LABEL PROC DB 196,65,123,16,4,192 ; vmovsd (%r8,%rax,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je 4b44 <_sk_load_f16_dst_hsw+0xe7> + DB 116,85 ; je 4bb8 <_sk_load_f16_dst_hsw+0xe7> DB 196,65,57,22,68,192,8 ; vmovhpd 0x8(%r8,%rax,8),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb 4b44 <_sk_load_f16_dst_hsw+0xe7> + DB 114,72 ; jb 4bb8 <_sk_load_f16_dst_hsw+0xe7> DB 196,193,123,16,116,192,16 ; vmovsd 0x10(%r8,%rax,8),%xmm6 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je 4b51 <_sk_load_f16_dst_hsw+0xf4> + DB 116,72 ; je 4bc5 <_sk_load_f16_dst_hsw+0xf4> DB 196,193,73,22,116,192,24 ; vmovhpd 0x18(%r8,%rax,8),%xmm6,%xmm6 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb 4b51 <_sk_load_f16_dst_hsw+0xf4> + DB 114,59 ; jb 4bc5 <_sk_load_f16_dst_hsw+0xf4> DB 196,193,123,16,124,192,32 ; vmovsd 0x20(%r8,%rax,8),%xmm7 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,109,255,255,255 ; je 4a94 <_sk_load_f16_dst_hsw+0x37> + DB 15,132,109,255,255,255 ; je 4b08 <_sk_load_f16_dst_hsw+0x37> DB 196,193,65,22,124,192,40 ; vmovhpd 0x28(%r8,%rax,8),%xmm7,%xmm7 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,92,255,255,255 ; jb 4a94 <_sk_load_f16_dst_hsw+0x37> + DB 15,130,92,255,255,255 ; jb 4b08 <_sk_load_f16_dst_hsw+0x37> DB 196,65,122,126,76,192,48 ; vmovq 0x30(%r8,%rax,8),%xmm9 - DB 233,80,255,255,255 ; jmpq 4a94 <_sk_load_f16_dst_hsw+0x37> + DB 233,80,255,255,255 ; jmpq 4b08 <_sk_load_f16_dst_hsw+0x37> DB 197,193,87,255 ; vxorpd %xmm7,%xmm7,%xmm7 DB 197,201,87,246 ; vxorpd %xmm6,%xmm6,%xmm6 - DB 233,67,255,255,255 ; jmpq 4a94 <_sk_load_f16_dst_hsw+0x37> + DB 233,67,255,255,255 ; jmpq 4b08 <_sk_load_f16_dst_hsw+0x37> DB 197,193,87,255 ; vxorpd %xmm7,%xmm7,%xmm7 - DB 233,58,255,255,255 ; jmpq 4a94 <_sk_load_f16_dst_hsw+0x37> + DB 233,58,255,255,255 ; jmpq 4b08 <_sk_load_f16_dst_hsw+0x37> PUBLIC _sk_gather_f16_hsw _sk_gather_f16_hsw LABEL PROC @@ -4649,8 +4684,10 @@ _sk_gather_f16_hsw LABEL PROC DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,194,237,144,28,192 ; vpgatherdq %ymm2,(%r8,%xmm0,8),%ymm3 + DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 + DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 + DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 + DB 196,194,189,144,28,192 ; vpgatherdq %ymm8,(%r8,%xmm0,8),%ymm3 DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 DB 196,194,245,144,20,192 ; vpgatherdq %ymm1,(%r8,%xmm0,8),%ymm2 DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0 @@ -4696,31 +4733,31 @@ _sk_store_f16_hsw LABEL PROC DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9 DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,31 ; jne 4c5f <_sk_store_f16_hsw+0x7b> - DB 196,65,120,17,28,192 ; vmovups %xmm11,(%r8,%rax,8) - DB 196,65,120,17,84,192,16 ; vmovups %xmm10,0x10(%r8,%rax,8) - DB 196,65,120,17,76,192,32 ; vmovups %xmm9,0x20(%r8,%rax,8) + DB 117,31 ; jne 4cdc <_sk_store_f16_hsw+0x7b> + DB 196,65,122,127,28,192 ; vmovdqu %xmm11,(%r8,%rax,8) + DB 196,65,122,127,84,192,16 ; vmovdqu %xmm10,0x10(%r8,%rax,8) + DB 196,65,122,127,76,192,32 ; vmovdqu %xmm9,0x20(%r8,%rax,8) DB 196,65,122,127,68,192,48 ; vmovdqu %xmm8,0x30(%r8,%rax,8) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,192 ; vmovq %xmm11,(%r8,%rax,8) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 4c5b <_sk_store_f16_hsw+0x77> + DB 116,240 ; je 4cd8 <_sk_store_f16_hsw+0x77> DB 196,65,121,23,92,192,8 ; vmovhpd %xmm11,0x8(%r8,%rax,8) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 4c5b <_sk_store_f16_hsw+0x77> + DB 114,227 ; jb 4cd8 <_sk_store_f16_hsw+0x77> DB 196,65,121,214,84,192,16 ; vmovq %xmm10,0x10(%r8,%rax,8) - DB 116,218 ; je 4c5b <_sk_store_f16_hsw+0x77> + DB 116,218 ; je 4cd8 <_sk_store_f16_hsw+0x77> DB 196,65,121,23,84,192,24 ; vmovhpd %xmm10,0x18(%r8,%rax,8) DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,205 ; jb 4c5b <_sk_store_f16_hsw+0x77> + DB 114,205 ; jb 4cd8 <_sk_store_f16_hsw+0x77> DB 196,65,121,214,76,192,32 ; vmovq %xmm9,0x20(%r8,%rax,8) - DB 116,196 ; je 4c5b <_sk_store_f16_hsw+0x77> + DB 116,196 ; je 4cd8 <_sk_store_f16_hsw+0x77> DB 196,65,121,23,76,192,40 ; vmovhpd %xmm9,0x28(%r8,%rax,8) DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,183 ; jb 4c5b <_sk_store_f16_hsw+0x77> + DB 114,183 ; jb 4cd8 <_sk_store_f16_hsw+0x77> DB 196,65,121,214,68,192,48 ; vmovq %xmm8,0x30(%r8,%rax,8) - DB 235,174 ; jmp 4c5b <_sk_store_f16_hsw+0x77> + DB 235,174 ; jmp 4cd8 <_sk_store_f16_hsw+0x77> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_hsw LABEL PROC @@ -4733,7 +4770,7 @@ _sk_load_u16_be_hsw LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,204,0,0,0 ; jne 4da0 <_sk_load_u16_be_hsw+0xf3> + DB 15,133,204,0,0,0 ; jne 4e1d <_sk_load_u16_be_hsw+0xf3> DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -4752,7 +4789,7 @@ _sk_load_u16_be_hsw LABEL PROC DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,21,7,134,2,0 ; vbroadcastss 0x28607(%rip),%ymm10 # 2d33c <_sk_overlay_sse2_8bit+0xde4> + DB 196,98,125,24,21,106,132,2,0 ; vbroadcastss 0x2846a(%rip),%ymm10 # 2d21c <_sk_overlay_sse2_8bit+0xd6d> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -4780,29 +4817,29 @@ _sk_load_u16_be_hsw LABEL PROC DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je 4e06 <_sk_load_u16_be_hsw+0x159> + DB 116,85 ; je 4e83 <_sk_load_u16_be_hsw+0x159> DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb 4e06 <_sk_load_u16_be_hsw+0x159> + DB 114,72 ; jb 4e83 <_sk_load_u16_be_hsw+0x159> DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je 4e13 <_sk_load_u16_be_hsw+0x166> + DB 116,72 ; je 4e90 <_sk_load_u16_be_hsw+0x166> DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb 4e13 <_sk_load_u16_be_hsw+0x166> + DB 114,59 ; jb 4e90 <_sk_load_u16_be_hsw+0x166> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,6,255,255,255 ; je 4cef <_sk_load_u16_be_hsw+0x42> + DB 15,132,6,255,255,255 ; je 4d6c <_sk_load_u16_be_hsw+0x42> DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,245,254,255,255 ; jb 4cef <_sk_load_u16_be_hsw+0x42> + DB 15,130,245,254,255,255 ; jb 4d6c <_sk_load_u16_be_hsw+0x42> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,233,254,255,255 ; jmpq 4cef <_sk_load_u16_be_hsw+0x42> + DB 233,233,254,255,255 ; jmpq 4d6c <_sk_load_u16_be_hsw+0x42> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,220,254,255,255 ; jmpq 4cef <_sk_load_u16_be_hsw+0x42> + DB 233,220,254,255,255 ; jmpq 4d6c <_sk_load_u16_be_hsw+0x42> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,211,254,255,255 ; jmpq 4cef <_sk_load_u16_be_hsw+0x42> + DB 233,211,254,255,255 ; jmpq 4d6c <_sk_load_u16_be_hsw+0x42> PUBLIC _sk_load_rgb_u16_be_hsw _sk_load_rgb_u16_be_hsw LABEL PROC @@ -4817,7 +4854,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 72,141,4,64 ; lea (%rax,%rax,2),%rax DB 72,193,248,32 ; sar $0x20,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,203,0,0,0 ; jne 4f12 <_sk_load_rgb_u16_be_hsw+0xf6> + DB 15,133,205,0,0,0 ; jne 4f91 <_sk_load_rgb_u16_be_hsw+0xf8> DB 196,65,122,111,28,64 ; vmovdqu (%r8,%rax,2),%xmm11 DB 196,193,122,111,92,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm3 DB 196,193,122,111,84,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm2 @@ -4825,102 +4862,102 @@ _sk_load_rgb_u16_be_hsw LABEL PROC DB 197,249,115,216,4 ; vpsrldq $0x4,%xmm0,%xmm0 DB 196,193,57,115,219,6 ; vpsrldq $0x6,%xmm11,%xmm8 DB 197,169,115,219,6 ; vpsrldq $0x6,%xmm3,%xmm10 - DB 197,153,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm12 + DB 197,241,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm1 DB 197,177,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm9 - DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 - DB 196,193,57,97,202 ; vpunpcklwd %xmm10,%xmm8,%xmm1 + DB 196,193,113,97,201 ; vpunpcklwd %xmm9,%xmm1,%xmm1 DB 197,233,97,192 ; vpunpcklwd %xmm0,%xmm2,%xmm0 - DB 196,193,25,97,209 ; vpunpcklwd %xmm9,%xmm12,%xmm2 - DB 197,97,97,193 ; vpunpcklwd %xmm1,%xmm3,%xmm8 - DB 197,97,105,201 ; vpunpckhwd %xmm1,%xmm3,%xmm9 - DB 197,249,97,202 ; vpunpcklwd %xmm2,%xmm0,%xmm1 - DB 197,121,105,210 ; vpunpckhwd %xmm2,%xmm0,%xmm10 - DB 197,185,108,193 ; vpunpcklqdq %xmm1,%xmm8,%xmm0 - DB 197,225,113,240,8 ; vpsllw $0x8,%xmm0,%xmm3 + DB 196,193,57,97,210 ; vpunpcklwd %xmm10,%xmm8,%xmm2 + DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 + DB 197,97,97,194 ; vpunpcklwd %xmm2,%xmm3,%xmm8 + DB 197,225,105,210 ; vpunpckhwd %xmm2,%xmm3,%xmm2 + DB 197,249,97,217 ; vpunpcklwd %xmm1,%xmm0,%xmm3 + DB 197,249,105,193 ; vpunpckhwd %xmm1,%xmm0,%xmm0 + DB 197,233,108,208 ; vpunpcklqdq %xmm0,%xmm2,%xmm2 + DB 197,185,108,195 ; vpunpcklqdq %xmm3,%xmm8,%xmm0 + DB 197,241,113,240,8 ; vpsllw $0x8,%xmm0,%xmm1 DB 197,249,113,208,8 ; vpsrlw $0x8,%xmm0,%xmm0 - DB 197,225,235,192 ; vpor %xmm0,%xmm3,%xmm0 + DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,29,122,132,2,0 ; vbroadcastss 0x2847a(%rip),%ymm3 # 2d33c <_sk_overlay_sse2_8bit+0xde4> - DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 197,185,109,201 ; vpunpckhqdq %xmm1,%xmm8,%xmm1 - DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 + DB 196,98,125,24,13,217,130,2,0 ; vbroadcastss 0x282d9(%rip),%ymm9 # 2d21c <_sk_overlay_sse2_8bit+0xd6d> + DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 197,185,109,203 ; vpunpckhqdq %xmm3,%xmm8,%xmm1 + DB 197,225,113,241,8 ; vpsllw $0x8,%xmm1,%xmm3 DB 197,241,113,209,8 ; vpsrlw $0x8,%xmm1,%xmm1 - DB 197,233,235,201 ; vpor %xmm1,%xmm2,%xmm1 + DB 197,225,235,201 ; vpor %xmm1,%xmm3,%xmm1 DB 196,226,125,51,201 ; vpmovzxwd %xmm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,193,49,108,210 ; vpunpcklqdq %xmm10,%xmm9,%xmm2 - DB 197,185,113,242,8 ; vpsllw $0x8,%xmm2,%xmm8 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 197,225,113,242,8 ; vpsllw $0x8,%xmm2,%xmm3 DB 197,233,113,210,8 ; vpsrlw $0x8,%xmm2,%xmm2 - DB 197,185,235,210 ; vpor %xmm2,%xmm8,%xmm2 + DB 197,225,235,210 ; vpor %xmm2,%xmm3,%xmm2 DB 196,226,125,51,210 ; vpmovzxwd %xmm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,156,131,2,0 ; vbroadcastss 0x2839c(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,253,129,2,0 ; vbroadcastss 0x281fd(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 - DB 196,65,121,196,92,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm11 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,65,121,196,92,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm11 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,32 ; jne 4f4b <_sk_load_rgb_u16_be_hsw+0x12f> + DB 117,31 ; jne 4fc9 <_sk_load_rgb_u16_be_hsw+0x130> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 233,49,255,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 233,48,255,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> DB 196,193,121,110,68,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm0 - DB 196,65,121,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm0,%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,65,121,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm0,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,49 ; jb 4f96 <_sk_load_rgb_u16_be_hsw+0x17a> + DB 114,48 ; jb 5013 <_sk_load_rgb_u16_be_hsw+0x17a> DB 196,193,121,110,68,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm0 - DB 196,193,121,196,92,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm0,%xmm3 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,193,121,196,92,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm0,%xmm3 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 117,50 ; jne 4fb1 <_sk_load_rgb_u16_be_hsw+0x195> + DB 117,48 ; jne 502d <_sk_load_rgb_u16_be_hsw+0x194> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 233,230,254,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 233,230,254,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 - DB 233,203,254,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 233,204,254,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> DB 196,193,121,110,68,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm0 - DB 196,65,121,196,84,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm0,%xmm10 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,65,121,196,84,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm0,%xmm10 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,40 ; jb 4ff3 <_sk_load_rgb_u16_be_hsw+0x1d7> + DB 114,39 ; jb 506e <_sk_load_rgb_u16_be_hsw+0x1d5> DB 196,193,121,110,68,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm0 - DB 196,193,121,196,84,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm0,%xmm2 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,193,121,196,84,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm0,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 117,32 ; jne 5005 <_sk_load_rgb_u16_be_hsw+0x1e9> + DB 117,30 ; jne 507f <_sk_load_rgb_u16_be_hsw+0x1e6> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 - DB 233,137,254,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 + DB 233,139,254,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 - DB 233,119,254,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 233,122,254,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> DB 196,193,121,110,68,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm0 - DB 196,65,121,196,100,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm0,%xmm12 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,193,121,196,76,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm0,%xmm1 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,25 ; jb 5038 <_sk_load_rgb_u16_be_hsw+0x21c> + DB 114,25 ; jb 50b2 <_sk_load_rgb_u16_be_hsw+0x219> DB 196,193,121,110,68,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm0 - DB 196,193,121,196,68,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm0,%xmm0 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 233,68,254,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 196,193,121,196,68,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm0,%xmm0 + DB 233,71,254,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 233,59,254,255,255 ; jmpq 4e7c <_sk_load_rgb_u16_be_hsw+0x60> + DB 233,62,254,255,255 ; jmpq 4ef9 <_sk_load_rgb_u16_be_hsw+0x60> PUBLIC _sk_store_u16_be_hsw _sk_store_u16_be_hsw LABEL PROC @@ -4932,7 +4969,7 @@ _sk_store_u16_be_hsw LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax - DB 196,98,125,24,5,92,131,2,0 ; vbroadcastss 0x2835c(%rip),%ymm8 # 2d3c4 <_sk_overlay_sse2_8bit+0xe6c> + DB 196,98,125,24,5,194,129,2,0 ; vbroadcastss 0x281c2(%rip),%ymm8 # 2d2a4 <_sk_overlay_sse2_8bit+0xdf5> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10 @@ -4970,39 +5007,39 @@ _sk_store_u16_be_hsw LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,31 ; jne 514c <_sk_store_u16_be_hsw+0x10b> - DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2) - DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2) - DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2) + DB 117,31 ; jne 51c6 <_sk_store_u16_be_hsw+0x10b> + DB 196,65,122,127,28,64 ; vmovdqu %xmm11,(%r8,%rax,2) + DB 196,65,122,127,84,64,16 ; vmovdqu %xmm10,0x10(%r8,%rax,2) + DB 196,65,122,127,76,64,32 ; vmovdqu %xmm9,0x20(%r8,%rax,2) DB 196,65,122,127,68,64,48 ; vmovdqu %xmm8,0x30(%r8,%rax,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 5148 <_sk_store_u16_be_hsw+0x107> + DB 116,240 ; je 51c2 <_sk_store_u16_be_hsw+0x107> DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 5148 <_sk_store_u16_be_hsw+0x107> + DB 114,227 ; jb 51c2 <_sk_store_u16_be_hsw+0x107> DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je 5148 <_sk_store_u16_be_hsw+0x107> + DB 116,218 ; je 51c2 <_sk_store_u16_be_hsw+0x107> DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2) DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,205 ; jb 5148 <_sk_store_u16_be_hsw+0x107> + DB 114,205 ; jb 51c2 <_sk_store_u16_be_hsw+0x107> DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je 5148 <_sk_store_u16_be_hsw+0x107> + DB 116,196 ; je 51c2 <_sk_store_u16_be_hsw+0x107> DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2) DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,183 ; jb 5148 <_sk_store_u16_be_hsw+0x107> + DB 114,183 ; jb 51c2 <_sk_store_u16_be_hsw+0x107> DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp 5148 <_sk_store_u16_be_hsw+0x107> + DB 235,174 ; jmp 51c2 <_sk_store_u16_be_hsw+0x107> PUBLIC _sk_load_f32_hsw _sk_load_f32_hsw LABEL PROC + DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 197,253,87,192 ; vxorpd %ymm0,%ymm0,%ymm0 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,135,101,1,0,0 ; ja 530f <_sk_load_f32_hsw+0x175> - DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp + DB 15,135,94,1,0,0 ; ja 5389 <_sk_load_f32_hsw+0x175> DB 197,253,17,36,36 ; vmovupd %ymm4,(%rsp) DB 197,253,17,108,36,32 ; vmovupd %ymm5,0x20(%rsp) DB 197,253,17,116,36,64 ; vmovupd %ymm6,0x40(%rsp) @@ -5014,7 +5051,7 @@ _sk_load_f32_hsw LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 77,99,201 ; movslq %r9d,%r9 - DB 76,141,21,52,1,0,0 ; lea 0x134(%rip),%r10 # 5320 <_sk_load_f32_hsw+0x186> + DB 76,141,21,50,1,0,0 ; lea 0x132(%rip),%r10 # 5398 <_sk_load_f32_hsw+0x184> DB 73,99,4,186 ; movslq (%r10,%rdi,4),%rax DB 76,1,208 ; add %r10,%rax DB 197,221,87,228 ; vxorpd %ymm4,%ymm4,%ymm4 @@ -5063,7 +5100,6 @@ _sk_load_f32_hsw LABEL PROC DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 - DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 196,193,116,20,210 ; vunpcklps %ymm10,%ymm1,%ymm2 DB 196,65,116,21,194 ; vunpckhps %ymm10,%ymm1,%ymm8 DB 197,228,20,200 ; vunpcklps %ymm0,%ymm3,%ymm1 @@ -5073,37 +5109,34 @@ _sk_load_f32_hsw LABEL PROC DB 197,189,20,211 ; vunpcklpd %ymm3,%ymm8,%ymm2 DB 197,189,21,219 ; vunpckhpd %ymm3,%ymm8,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax DB 197,229,87,219 ; vxorpd %ymm3,%ymm3,%ymm3 DB 196,65,45,87,210 ; vxorpd %ymm10,%ymm10,%ymm10 DB 197,245,87,201 ; vxorpd %ymm1,%ymm1,%ymm1 - DB 235,203 ; jmp 52e9 <_sk_load_f32_hsw+0x14f> - DB 102,144 ; xchg %ax,%ax - DB 25,255 ; sbb %edi,%edi - DB 255 ; (bad) - DB 255,159,255,255,255,138 ; lcall *-0x75000001(%rdi) - DB 255 ; (bad) + DB 235,196 ; jmp 535c <_sk_load_f32_hsw+0x148> + DB 27,255 ; sbb %edi,%edi DB 255 ; (bad) - DB 255,117,255 ; pushq -0x1(%rbp) + DB 255,161,255,255,255,140 ; jmpq *-0x73000001(%rcx) DB 255 ; (bad) - DB 255,90,255 ; lcall *-0x1(%rdx) DB 255 ; (bad) - DB 255,68,255,255 ; incl -0x1(%rdi,%rdi,8) - DB 255,51 ; pushq (%rbx) + DB 255,119,255 ; pushq -0x1(%rdi) DB 255 ; (bad) + DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) + DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) - DB 255,38 ; jmpq *(%rsi) + DB 255,53,255,255,255,40 ; pushq 0x28ffffff(%rip) # 290053b4 <_sk_overlay_sse2_8bit+0x28fd8f05> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_f32_dst_hsw _sk_load_f32_dst_hsw LABEL PROC + DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 197,221,87,228 ; vxorpd %ymm4,%ymm4,%ymm4 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,135,101,1,0,0 ; ja 54b5 <_sk_load_f32_dst_hsw+0x175> - DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp + DB 15,135,94,1,0,0 ; ja 552d <_sk_load_f32_dst_hsw+0x175> DB 197,253,17,4,36 ; vmovupd %ymm0,(%rsp) DB 197,253,17,76,36,32 ; vmovupd %ymm1,0x20(%rsp) DB 197,253,17,84,36,64 ; vmovupd %ymm2,0x40(%rsp) @@ -5115,7 +5148,7 @@ _sk_load_f32_dst_hsw LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 77,99,201 ; movslq %r9d,%r9 - DB 76,141,21,50,1,0,0 ; lea 0x132(%rip),%r10 # 54c4 <_sk_load_f32_dst_hsw+0x184> + DB 76,141,21,50,1,0,0 ; lea 0x132(%rip),%r10 # 553c <_sk_load_f32_dst_hsw+0x184> DB 73,99,4,186 ; movslq (%r10,%rdi,4),%rax DB 76,1,208 ; add %r10,%rax DB 197,253,87,192 ; vxorpd %ymm0,%ymm0,%ymm0 @@ -5164,7 +5197,6 @@ _sk_load_f32_dst_hsw LABEL PROC DB 197,252,16,84,36,64 ; vmovups 0x40(%rsp),%ymm2 DB 197,252,16,76,36,32 ; vmovups 0x20(%rsp),%ymm1 DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 - DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 196,193,84,20,242 ; vunpcklps %ymm10,%ymm5,%ymm6 DB 196,65,84,21,194 ; vunpckhps %ymm10,%ymm5,%ymm8 DB 197,196,20,236 ; vunpcklps %ymm4,%ymm7,%ymm5 @@ -5174,11 +5206,12 @@ _sk_load_f32_dst_hsw LABEL PROC DB 197,189,20,247 ; vunpcklpd %ymm7,%ymm8,%ymm6 DB 197,189,21,255 ; vunpckhpd %ymm7,%ymm8,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax DB 197,197,87,255 ; vxorpd %ymm7,%ymm7,%ymm7 DB 196,65,45,87,210 ; vxorpd %ymm10,%ymm10,%ymm10 DB 197,213,87,237 ; vxorpd %ymm5,%ymm5,%ymm5 - DB 235,203 ; jmp 548f <_sk_load_f32_dst_hsw+0x14f> + DB 235,196 ; jmp 5500 <_sk_load_f32_dst_hsw+0x148> DB 27,255 ; sbb %edi,%edi DB 255 ; (bad) DB 255,161,255,255,255,140 ; jmpq *-0x73000001(%rcx) @@ -5189,7 +5222,7 @@ _sk_load_f32_dst_hsw LABEL PROC DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) - DB 255,53,255,255,255,40 ; pushq 0x28ffffff(%rip) # 290054e0 <_sk_overlay_sse2_8bit+0x28fd8f88> + DB 255,53,255,255,255,40 ; pushq 0x28ffffff(%rip) # 29005558 <_sk_overlay_sse2_8bit+0x28fd90a9> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -5213,7 +5246,7 @@ _sk_store_f32_hsw LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,255 ; test %rdi,%rdi - DB 117,55 ; jne 5563 <_sk_store_f32_hsw+0x7f> + DB 117,55 ; jne 55db <_sk_store_f32_hsw+0x7f> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -5226,22 +5259,22 @@ _sk_store_f32_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 555f <_sk_store_f32_hsw+0x7b> + DB 116,240 ; je 55d7 <_sk_store_f32_hsw+0x7b> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 555f <_sk_store_f32_hsw+0x7b> + DB 114,227 ; jb 55d7 <_sk_store_f32_hsw+0x7b> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 555f <_sk_store_f32_hsw+0x7b> + DB 116,218 ; je 55d7 <_sk_store_f32_hsw+0x7b> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,205 ; jb 555f <_sk_store_f32_hsw+0x7b> + DB 114,205 ; jb 55d7 <_sk_store_f32_hsw+0x7b> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 555f <_sk_store_f32_hsw+0x7b> + DB 116,195 ; je 55d7 <_sk_store_f32_hsw+0x7b> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,181 ; jb 555f <_sk_store_f32_hsw+0x7b> + DB 114,181 ; jb 55d7 <_sk_store_f32_hsw+0x7b> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 555f <_sk_store_f32_hsw+0x7b> + DB 235,171 ; jmp 55d7 <_sk_store_f32_hsw+0x7b> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -5308,7 +5341,7 @@ _sk_mirror_x_hsw LABEL PROC DB 196,65,124,92,218 ; vsubps %ymm10,%ymm0,%ymm11 DB 196,193,58,88,192 ; vaddss %xmm8,%xmm8,%xmm0 DB 196,98,125,24,192 ; vbroadcastss %xmm0,%ymm8 - DB 197,178,89,5,23,124,2,0 ; vmulss 0x27c17(%rip),%xmm9,%xmm0 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 197,178,89,5,127,122,2,0 ; vmulss 0x27a7f(%rip),%xmm9,%xmm0 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,226,125,24,192 ; vbroadcastss %xmm0,%ymm0 DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 DB 196,227,125,8,192,1 ; vroundps $0x1,%ymm0,%ymm0 @@ -5333,7 +5366,7 @@ _sk_mirror_y_hsw LABEL PROC DB 196,65,116,92,218 ; vsubps %ymm10,%ymm1,%ymm11 DB 196,193,58,88,200 ; vaddss %xmm8,%xmm8,%xmm1 DB 196,98,125,24,193 ; vbroadcastss %xmm1,%ymm8 - DB 197,178,89,13,179,123,2,0 ; vmulss 0x27bb3(%rip),%xmm9,%xmm1 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 197,178,89,13,27,122,2,0 ; vmulss 0x27a1b(%rip),%xmm9,%xmm1 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,226,125,24,201 ; vbroadcastss %xmm1,%ymm1 DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 DB 196,227,125,8,201,1 ; vroundps $0x1,%ymm1,%ymm1 @@ -5353,7 +5386,7 @@ PUBLIC _sk_clamp_x_1_hsw _sk_clamp_x_1_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,104,123,2,0 ; vbroadcastss 0x27b68(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,208,121,2,0 ; vbroadcastss 0x279d0(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -5367,9 +5400,9 @@ _sk_repeat_x_1_hsw LABEL PROC PUBLIC _sk_mirror_x_1_hsw _sk_mirror_x_1_hsw LABEL PROC - DB 196,98,125,24,5,95,123,2,0 ; vbroadcastss 0x27b5f(%rip),%ymm8 # 2d2c4 <_sk_overlay_sse2_8bit+0xd6c> + DB 196,98,125,24,5,199,121,2,0 ; vbroadcastss 0x279c7(%rip),%ymm8 # 2d1a4 <_sk_overlay_sse2_8bit+0xcf5> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,53,123,2,0 ; vbroadcastss 0x27b35(%rip),%ymm9 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,13,157,121,2,0 ; vbroadcastss 0x2799d(%rip),%ymm9 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -5383,11 +5416,11 @@ _sk_mirror_x_1_hsw LABEL PROC PUBLIC _sk_luminance_to_alpha_hsw _sk_luminance_to_alpha_hsw LABEL PROC - DB 196,226,125,24,29,33,124,2,0 ; vbroadcastss 0x27c21(%rip),%ymm3 # 2d3c8 <_sk_overlay_sse2_8bit+0xe70> - DB 196,98,125,24,5,28,124,2,0 ; vbroadcastss 0x27c1c(%rip),%ymm8 # 2d3cc <_sk_overlay_sse2_8bit+0xe74> + DB 196,226,125,24,29,137,122,2,0 ; vbroadcastss 0x27a89(%rip),%ymm3 # 2d2a8 <_sk_overlay_sse2_8bit+0xdf9> + DB 196,98,125,24,5,132,122,2,0 ; vbroadcastss 0x27a84(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xdfd> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 196,226,125,184,203 ; vfmadd231ps %ymm3,%ymm0,%ymm1 - DB 196,226,125,24,29,13,124,2,0 ; vbroadcastss 0x27c0d(%rip),%ymm3 # 2d3d0 <_sk_overlay_sse2_8bit+0xe78> + DB 196,226,125,24,29,117,122,2,0 ; vbroadcastss 0x27a75(%rip),%ymm3 # 2d2b0 <_sk_overlay_sse2_8bit+0xe01> DB 196,226,109,168,217 ; vfmadd213ps %ymm1,%ymm2,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -5570,9 +5603,9 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC DB 76,139,64,8 ; mov 0x8(%rax),%r8 DB 77,137,202 ; mov %r9,%r10 DB 73,255,202 ; dec %r10 - DB 120,7 ; js 5abb <_sk_evenly_spaced_gradient_hsw+0x18> + DB 120,7 ; js 5b33 <_sk_evenly_spaced_gradient_hsw+0x18> DB 196,193,242,42,202 ; vcvtsi2ss %r10,%xmm1,%xmm1 - DB 235,22 ; jmp 5ad1 <_sk_evenly_spaced_gradient_hsw+0x2e> + DB 235,22 ; jmp 5b49 <_sk_evenly_spaced_gradient_hsw+0x2e> DB 77,137,211 ; mov %r10,%r11 DB 73,209,235 ; shr %r11 DB 65,131,226,1 ; and $0x1,%r10d @@ -5581,66 +5614,74 @@ _sk_evenly_spaced_gradient_hsw LABEL PROC DB 197,242,88,201 ; vaddss %xmm1,%xmm1,%xmm1 DB 196,226,125,24,201 ; vbroadcastss %xmm1,%ymm1 DB 197,244,89,200 ; vmulps %ymm0,%ymm1,%ymm1 - DB 197,126,91,217 ; vcvttps2dq %ymm1,%ymm11 + DB 197,126,91,225 ; vcvttps2dq %ymm1,%ymm12 DB 73,131,249,8 ; cmp $0x8,%r9 - DB 119,70 ; ja 5b2a <_sk_evenly_spaced_gradient_hsw+0x87> - DB 196,66,37,22,0 ; vpermps (%r8),%ymm11,%ymm8 + DB 119,73 ; ja 5ba5 <_sk_evenly_spaced_gradient_hsw+0x8a> + DB 196,66,29,22,0 ; vpermps (%r8),%ymm12,%ymm8 DB 76,139,64,40 ; mov 0x28(%rax),%r8 - DB 196,66,37,22,8 ; vpermps (%r8),%ymm11,%ymm9 + DB 196,66,29,22,8 ; vpermps (%r8),%ymm12,%ymm9 DB 76,139,64,16 ; mov 0x10(%rax),%r8 DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 196,194,37,22,8 ; vpermps (%r8),%ymm11,%ymm1 + DB 196,194,29,22,8 ; vpermps (%r8),%ymm12,%ymm1 DB 76,139,64,48 ; mov 0x30(%rax),%r8 - DB 196,66,37,22,16 ; vpermps (%r8),%ymm11,%ymm10 - DB 196,194,37,22,17 ; vpermps (%r9),%ymm11,%ymm2 + DB 196,66,29,22,24 ; vpermps (%r8),%ymm12,%ymm11 + DB 196,194,29,22,17 ; vpermps (%r9),%ymm12,%ymm2 DB 76,139,64,56 ; mov 0x38(%rax),%r8 - DB 196,66,37,22,32 ; vpermps (%r8),%ymm11,%ymm12 + DB 196,66,29,22,40 ; vpermps (%r8),%ymm12,%ymm13 DB 76,139,64,32 ; mov 0x20(%rax),%r8 - DB 196,194,37,22,24 ; vpermps (%r8),%ymm11,%ymm3 + DB 196,194,29,22,24 ; vpermps (%r8),%ymm12,%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 196,98,37,22,40 ; vpermps (%rax),%ymm11,%ymm13 - DB 235,110 ; jmp 5b98 <_sk_evenly_spaced_gradient_hsw+0xf5> + DB 196,98,29,22,16 ; vpermps (%rax),%ymm12,%ymm10 + DB 233,147,0,0,0 ; jmpq 5c38 <_sk_evenly_spaced_gradient_hsw+0x11d> DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 + DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,2,117,146,4,152 ; vgatherdps %ymm1,(%r8,%ymm11,4),%ymm8 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,2,117,146,4,160 ; vgatherdps %ymm1,(%r8,%ymm12,4),%ymm8 DB 76,139,64,40 ; mov 0x28(%rax),%r8 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,2,117,146,12,152 ; vgatherdps %ymm1,(%r8,%ymm11,4),%ymm9 + DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 + DB 196,2,117,146,12,160 ; vgatherdps %ymm1,(%r8,%ymm12,4),%ymm9 DB 76,139,64,16 ; mov 0x10(%rax),%r8 DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,130,109,146,12,152 ; vgatherdps %ymm2,(%r8,%ymm11,4),%ymm1 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 + DB 196,130,109,146,12,160 ; vgatherdps %ymm2,(%r8,%ymm12,4),%ymm1 DB 76,139,64,48 ; mov 0x30(%rax),%r8 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,2,109,146,20,152 ; vgatherdps %ymm2,(%r8,%ymm11,4),%ymm10 + DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11 + DB 196,2,109,146,28,160 ; vgatherdps %ymm2,(%r8,%ymm12,4),%ymm11 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,130,101,146,20,153 ; vgatherdps %ymm3,(%r9,%ymm11,4),%ymm2 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 + DB 196,130,101,146,20,161 ; vgatherdps %ymm3,(%r9,%ymm12,4),%ymm2 DB 76,139,64,56 ; mov 0x38(%rax),%r8 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,2,101,146,36,152 ; vgatherdps %ymm3,(%r8,%ymm11,4),%ymm12 + DB 196,65,20,87,237 ; vxorps %ymm13,%ymm13,%ymm13 + DB 196,2,101,146,44,160 ; vgatherdps %ymm3,(%r8,%ymm12,4),%ymm13 DB 76,139,64,32 ; mov 0x20(%rax),%r8 - DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 - DB 196,130,21,146,28,152 ; vgatherdps %ymm13,(%r8,%ymm11,4),%ymm3 + DB 196,65,5,118,255 ; vpcmpeqd %ymm15,%ymm15,%ymm15 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,130,5,146,28,160 ; vgatherdps %ymm15,(%r8,%ymm12,4),%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 196,34,13,146,44,152 ; vgatherdps %ymm14,(%rax,%ymm11,4),%ymm13 + DB 196,34,13,146,20,160 ; vgatherdps %ymm14,(%rax,%ymm12,4),%ymm10 DB 196,66,125,168,193 ; vfmadd213ps %ymm9,%ymm0,%ymm8 - DB 196,194,125,168,202 ; vfmadd213ps %ymm10,%ymm0,%ymm1 - DB 196,194,125,168,212 ; vfmadd213ps %ymm12,%ymm0,%ymm2 - DB 196,194,125,168,221 ; vfmadd213ps %ymm13,%ymm0,%ymm3 + DB 196,194,125,168,203 ; vfmadd213ps %ymm11,%ymm0,%ymm1 + DB 196,194,125,168,213 ; vfmadd213ps %ymm13,%ymm0,%ymm2 + DB 196,194,125,168,218 ; vfmadd213ps %ymm10,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,124,41,192 ; vmovaps %ymm8,%ymm0 DB 255,224 ; jmpq *%rax PUBLIC _sk_gauss_a_to_rgba_hsw _sk_gauss_a_to_rgba_hsw LABEL PROC - DB 196,226,125,24,5,23,120,2,0 ; vbroadcastss 0x27817(%rip),%ymm0 # 2d3d4 <_sk_overlay_sse2_8bit+0xe7c> - DB 196,226,125,24,13,18,120,2,0 ; vbroadcastss 0x27812(%rip),%ymm1 # 2d3d8 <_sk_overlay_sse2_8bit+0xe80> + DB 196,226,125,24,5,87,118,2,0 ; vbroadcastss 0x27657(%rip),%ymm0 # 2d2b4 <_sk_overlay_sse2_8bit+0xe05> + DB 196,226,125,24,13,82,118,2,0 ; vbroadcastss 0x27652(%rip),%ymm1 # 2d2b8 <_sk_overlay_sse2_8bit+0xe09> DB 196,226,101,168,200 ; vfmadd213ps %ymm0,%ymm3,%ymm1 - DB 196,226,125,24,5,8,120,2,0 ; vbroadcastss 0x27808(%rip),%ymm0 # 2d3dc <_sk_overlay_sse2_8bit+0xe84> + DB 196,226,125,24,5,72,118,2,0 ; vbroadcastss 0x27648(%rip),%ymm0 # 2d2bc <_sk_overlay_sse2_8bit+0xe0d> DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0 - DB 196,226,125,24,13,254,119,2,0 ; vbroadcastss 0x277fe(%rip),%ymm1 # 2d3e0 <_sk_overlay_sse2_8bit+0xe88> + DB 196,226,125,24,13,62,118,2,0 ; vbroadcastss 0x2763e(%rip),%ymm1 # 2d2c0 <_sk_overlay_sse2_8bit+0xe11> DB 196,226,101,184,200 ; vfmadd231ps %ymm0,%ymm3,%ymm1 - DB 196,226,125,24,5,244,119,2,0 ; vbroadcastss 0x277f4(%rip),%ymm0 # 2d3e4 <_sk_overlay_sse2_8bit+0xe8c> + DB 196,226,125,24,5,52,118,2,0 ; vbroadcastss 0x27634(%rip),%ymm0 # 2d2c4 <_sk_overlay_sse2_8bit+0xe15> DB 196,226,101,184,193 ; vfmadd231ps %ymm1,%ymm3,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -5653,46 +5694,54 @@ _sk_gradient_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 73,131,248,1 ; cmp $0x1,%r8 - DB 15,134,180,0,0,0 ; jbe 5cc8 <_sk_gradient_hsw+0xc3> + DB 15,134,220,0,0,0 ; jbe 5d90 <_sk_gradient_hsw+0xeb> DB 76,139,72,72 ; mov 0x48(%rax),%r9 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 65,186,1,0,0,0 ; mov $0x1,%r10d - DB 196,226,125,24,21,133,118,2,0 ; vbroadcastss 0x27685(%rip),%ymm2 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> + DB 196,226,125,24,21,197,116,2,0 ; vbroadcastss 0x274c5(%rip),%ymm2 # 2d190 <_sk_overlay_sse2_8bit+0xce1> DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9 DB 196,130,125,24,28,145 ; vbroadcastss (%r9,%r10,4),%ymm3 DB 197,228,194,216,2 ; vcmpleps %ymm0,%ymm3,%ymm3 DB 196,227,117,74,218,48 ; vblendvps %ymm3,%ymm2,%ymm1,%ymm3 - DB 196,65,101,254,201 ; vpaddd %ymm9,%ymm3,%ymm9 + DB 197,53,254,203 ; vpaddd %ymm3,%ymm9,%ymm9 DB 73,255,194 ; inc %r10 DB 77,57,208 ; cmp %r10,%r8 - DB 117,226 ; jne 5c30 <_sk_gradient_hsw+0x2b> + DB 117,227 ; jne 5cd0 <_sk_gradient_hsw+0x2b> DB 76,139,72,8 ; mov 0x8(%rax),%r9 DB 73,131,248,8 ; cmp $0x8,%r8 - DB 118,121 ; jbe 5cd1 <_sk_gradient_hsw+0xcc> + DB 15,134,158,0,0,0 ; jbe 5d99 <_sk_gradient_hsw+0xf4> DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 + DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,2,117,146,4,137 ; vgatherdps %ymm1,(%r9,%ymm9,4),%ymm8 DB 76,139,64,40 ; mov 0x28(%rax),%r8 + DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 DB 196,2,117,146,20,136 ; vgatherdps %ymm1,(%r8,%ymm9,4),%ymm10 DB 76,139,64,16 ; mov 0x10(%rax),%r8 DB 76,139,72,24 ; mov 0x18(%rax),%r9 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 DB 196,130,109,146,12,136 ; vgatherdps %ymm2,(%r8,%ymm9,4),%ymm1 DB 76,139,64,48 ; mov 0x30(%rax),%r8 + DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,2,109,146,28,136 ; vgatherdps %ymm2,(%r8,%ymm9,4),%ymm11 + DB 196,2,109,146,36,136 ; vgatherdps %ymm2,(%r8,%ymm9,4),%ymm12 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 DB 196,130,101,146,20,137 ; vgatherdps %ymm3,(%r9,%ymm9,4),%ymm2 DB 76,139,64,56 ; mov 0x38(%rax),%r8 + DB 196,65,20,87,237 ; vxorps %ymm13,%ymm13,%ymm13 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,2,101,146,36,136 ; vgatherdps %ymm3,(%r8,%ymm9,4),%ymm12 + DB 196,2,101,146,44,136 ; vgatherdps %ymm3,(%r8,%ymm9,4),%ymm13 DB 76,139,64,32 ; mov 0x20(%rax),%r8 - DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 - DB 196,130,21,146,28,136 ; vgatherdps %ymm13,(%r8,%ymm9,4),%ymm3 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,65,5,118,255 ; vpcmpeqd %ymm15,%ymm15,%ymm15 + DB 196,130,5,146,28,136 ; vgatherdps %ymm15,(%r8,%ymm9,4),%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 196,34,13,146,44,136 ; vgatherdps %ymm14,(%rax,%ymm9,4),%ymm13 - DB 235,77 ; jmp 5d15 <_sk_gradient_hsw+0x110> + DB 196,34,13,146,28,136 ; vgatherdps %ymm14,(%rax,%ymm9,4),%ymm11 + DB 235,77 ; jmp 5ddd <_sk_gradient_hsw+0x138> DB 76,139,72,8 ; mov 0x8(%rax),%r9 DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 196,66,53,22,1 ; vpermps (%r9),%ymm9,%ymm8 @@ -5702,18 +5751,18 @@ _sk_gradient_hsw LABEL PROC DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 196,194,53,22,8 ; vpermps (%r8),%ymm9,%ymm1 DB 76,139,64,48 ; mov 0x30(%rax),%r8 - DB 196,66,53,22,24 ; vpermps (%r8),%ymm9,%ymm11 + DB 196,66,53,22,32 ; vpermps (%r8),%ymm9,%ymm12 DB 196,194,53,22,17 ; vpermps (%r9),%ymm9,%ymm2 DB 76,139,64,56 ; mov 0x38(%rax),%r8 - DB 196,66,53,22,32 ; vpermps (%r8),%ymm9,%ymm12 + DB 196,66,53,22,40 ; vpermps (%r8),%ymm9,%ymm13 DB 76,139,64,32 ; mov 0x20(%rax),%r8 DB 196,194,53,22,24 ; vpermps (%r8),%ymm9,%ymm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax - DB 196,98,53,22,40 ; vpermps (%rax),%ymm9,%ymm13 + DB 196,98,53,22,24 ; vpermps (%rax),%ymm9,%ymm11 DB 196,66,125,168,194 ; vfmadd213ps %ymm10,%ymm0,%ymm8 - DB 196,194,125,168,203 ; vfmadd213ps %ymm11,%ymm0,%ymm1 - DB 196,194,125,168,212 ; vfmadd213ps %ymm12,%ymm0,%ymm2 - DB 196,194,125,168,221 ; vfmadd213ps %ymm13,%ymm0,%ymm3 + DB 196,194,125,168,204 ; vfmadd213ps %ymm12,%ymm0,%ymm1 + DB 196,194,125,168,213 ; vfmadd213ps %ymm13,%ymm0,%ymm2 + DB 196,194,125,168,219 ; vfmadd213ps %ymm11,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,124,41,192 ; vmovaps %ymm8,%ymm0 DB 255,224 ; jmpq *%rax @@ -5748,24 +5797,24 @@ _sk_xy_to_unit_angle_hsw LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,56,118,2,0 ; vbroadcastss 0x27638(%rip),%ymm13 # 2d3e8 <_sk_overlay_sse2_8bit+0xe90> - DB 196,98,125,24,53,51,118,2,0 ; vbroadcastss 0x27633(%rip),%ymm14 # 2d3ec <_sk_overlay_sse2_8bit+0xe94> + DB 196,98,125,24,45,80,116,2,0 ; vbroadcastss 0x27450(%rip),%ymm13 # 2d2c8 <_sk_overlay_sse2_8bit+0xe19> + DB 196,98,125,24,53,75,116,2,0 ; vbroadcastss 0x2744b(%rip),%ymm14 # 2d2cc <_sk_overlay_sse2_8bit+0xe1d> DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14 - DB 196,98,125,24,45,41,118,2,0 ; vbroadcastss 0x27629(%rip),%ymm13 # 2d3f0 <_sk_overlay_sse2_8bit+0xe98> + DB 196,98,125,24,45,65,116,2,0 ; vbroadcastss 0x27441(%rip),%ymm13 # 2d2d0 <_sk_overlay_sse2_8bit+0xe21> DB 196,66,29,184,238 ; vfmadd231ps %ymm14,%ymm12,%ymm13 - DB 196,98,125,24,53,31,118,2,0 ; vbroadcastss 0x2761f(%rip),%ymm14 # 2d3f4 <_sk_overlay_sse2_8bit+0xe9c> + DB 196,98,125,24,53,55,116,2,0 ; vbroadcastss 0x27437(%rip),%ymm14 # 2d2d4 <_sk_overlay_sse2_8bit+0xe25> DB 196,66,29,184,245 ; vfmadd231ps %ymm13,%ymm12,%ymm14 DB 196,65,36,89,222 ; vmulps %ymm14,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,10,118,2,0 ; vbroadcastss 0x2760a(%rip),%ymm10 # 2d3f8 <_sk_overlay_sse2_8bit+0xea0> + DB 196,98,125,24,21,34,116,2,0 ; vbroadcastss 0x27422(%rip),%ymm10 # 2d2d8 <_sk_overlay_sse2_8bit+0xe29> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,160,116,2,0 ; vbroadcastss 0x274a0(%rip),%ymm10 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,21,184,114,2,0 ; vbroadcastss 0x272b8(%rip),%ymm10 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,138,116,2,0 ; vbroadcastss 0x2748a(%rip),%ymm10 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,21,162,114,2,0 ; vbroadcastss 0x272a2(%rip),%ymm10 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -5789,23 +5838,23 @@ _sk_xy_to_2pt_conical_quadratic_max_hsw LABEL PROC DB 197,50,89,80,44 ; vmulss 0x2c(%rax),%xmm9,%xmm10 DB 196,66,125,24,210 ; vbroadcastss %xmm10,%ymm10 DB 197,44,88,208 ; vaddps %ymm0,%ymm10,%ymm10 - DB 196,98,125,24,29,138,117,2,0 ; vbroadcastss 0x2758a(%rip),%ymm11 # 2d3fc <_sk_overlay_sse2_8bit+0xea4> + DB 196,98,125,24,29,162,115,2,0 ; vbroadcastss 0x273a2(%rip),%ymm11 # 2d2dc <_sk_overlay_sse2_8bit+0xe2d> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 DB 197,116,89,217 ; vmulps %ymm1,%ymm1,%ymm11 DB 196,98,125,184,216 ; vfmadd231ps %ymm0,%ymm0,%ymm11 DB 196,193,50,89,193 ; vmulss %xmm9,%xmm9,%xmm0 DB 196,226,125,24,192 ; vbroadcastss %xmm0,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,13,105,117,2,0 ; vbroadcastss 0x27569(%rip),%ymm9 # 2d400 <_sk_overlay_sse2_8bit+0xea8> + DB 196,98,125,24,13,129,115,2,0 ; vbroadcastss 0x27381(%rip),%ymm9 # 2d2e0 <_sk_overlay_sse2_8bit+0xe31> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,194,45,184,194 ; vfmadd231ps %ymm10,%ymm10,%ymm0 DB 197,252,81,192 ; vsqrtps %ymm0,%ymm0 DB 196,98,125,24,64,36 ; vbroadcastss 0x24(%rax),%ymm8 - DB 196,98,125,24,13,76,117,2,0 ; vbroadcastss 0x2754c(%rip),%ymm9 # 2d404 <_sk_overlay_sse2_8bit+0xeac> + DB 196,98,125,24,13,100,115,2,0 ; vbroadcastss 0x27364(%rip),%ymm9 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> DB 196,65,44,87,201 ; vxorps %ymm9,%ymm10,%ymm9 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,221,115,2,0 ; vbroadcastss 0x273dd(%rip),%ymm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,29,245,113,2,0 ; vbroadcastss 0x271f5(%rip),%ymm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 196,65,44,89,208 ; vmulps %ymm8,%ymm10,%ymm10 DB 197,180,92,192 ; vsubps %ymm0,%ymm9,%ymm0 @@ -5822,23 +5871,23 @@ _sk_xy_to_2pt_conical_quadratic_min_hsw LABEL PROC DB 197,50,89,80,44 ; vmulss 0x2c(%rax),%xmm9,%xmm10 DB 196,66,125,24,210 ; vbroadcastss %xmm10,%ymm10 DB 197,44,88,208 ; vaddps %ymm0,%ymm10,%ymm10 - DB 196,98,125,24,29,242,116,2,0 ; vbroadcastss 0x274f2(%rip),%ymm11 # 2d3fc <_sk_overlay_sse2_8bit+0xea4> + DB 196,98,125,24,29,10,115,2,0 ; vbroadcastss 0x2730a(%rip),%ymm11 # 2d2dc <_sk_overlay_sse2_8bit+0xe2d> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 DB 197,116,89,217 ; vmulps %ymm1,%ymm1,%ymm11 DB 196,98,125,184,216 ; vfmadd231ps %ymm0,%ymm0,%ymm11 DB 196,193,50,89,193 ; vmulss %xmm9,%xmm9,%xmm0 DB 196,226,125,24,192 ; vbroadcastss %xmm0,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,13,209,116,2,0 ; vbroadcastss 0x274d1(%rip),%ymm9 # 2d400 <_sk_overlay_sse2_8bit+0xea8> + DB 196,98,125,24,13,233,114,2,0 ; vbroadcastss 0x272e9(%rip),%ymm9 # 2d2e0 <_sk_overlay_sse2_8bit+0xe31> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,194,45,184,194 ; vfmadd231ps %ymm10,%ymm10,%ymm0 DB 197,252,81,192 ; vsqrtps %ymm0,%ymm0 DB 196,98,125,24,64,36 ; vbroadcastss 0x24(%rax),%ymm8 - DB 196,98,125,24,13,180,116,2,0 ; vbroadcastss 0x274b4(%rip),%ymm9 # 2d404 <_sk_overlay_sse2_8bit+0xeac> + DB 196,98,125,24,13,204,114,2,0 ; vbroadcastss 0x272cc(%rip),%ymm9 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> DB 196,65,44,87,201 ; vxorps %ymm9,%ymm10,%ymm9 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,69,115,2,0 ; vbroadcastss 0x27345(%rip),%ymm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,29,93,113,2,0 ; vbroadcastss 0x2715d(%rip),%ymm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 196,65,44,89,208 ; vmulps %ymm8,%ymm10,%ymm10 DB 197,180,92,192 ; vsubps %ymm0,%ymm9,%ymm0 @@ -5854,14 +5903,14 @@ _sk_xy_to_2pt_conical_linear_hsw LABEL PROC DB 197,58,89,72,44 ; vmulss 0x2c(%rax),%xmm8,%xmm9 DB 196,66,125,24,201 ; vbroadcastss %xmm9,%ymm9 DB 197,52,88,200 ; vaddps %ymm0,%ymm9,%ymm9 - DB 196,98,125,24,21,96,116,2,0 ; vbroadcastss 0x27460(%rip),%ymm10 # 2d3fc <_sk_overlay_sse2_8bit+0xea4> + DB 196,98,125,24,21,120,114,2,0 ; vbroadcastss 0x27278(%rip),%ymm10 # 2d2dc <_sk_overlay_sse2_8bit+0xe2d> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 DB 197,116,89,209 ; vmulps %ymm1,%ymm1,%ymm10 DB 196,98,125,184,208 ; vfmadd231ps %ymm0,%ymm0,%ymm10 DB 196,193,58,89,192 ; vmulss %xmm8,%xmm8,%xmm0 DB 196,226,125,24,192 ; vbroadcastss %xmm0,%ymm0 DB 197,172,92,192 ; vsubps %ymm0,%ymm10,%ymm0 - DB 196,98,125,24,5,67,116,2,0 ; vbroadcastss 0x27443(%rip),%ymm8 # 2d404 <_sk_overlay_sse2_8bit+0xeac> + DB 196,98,125,24,5,91,114,2,0 ; vbroadcastss 0x2725b(%rip),%ymm8 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> DB 196,193,124,87,192 ; vxorps %ymm8,%ymm0,%ymm0 DB 196,193,124,94,193 ; vdivps %ymm9,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax @@ -5895,7 +5944,7 @@ _sk_apply_vector_mask_hsw LABEL PROC PUBLIC _sk_save_xy_hsw _sk_save_xy_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,131,114,2,0 ; vbroadcastss 0x27283(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,155,112,2,0 ; vbroadcastss 0x2709b(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9 DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 @@ -5925,9 +5974,9 @@ _sk_accumulate_hsw LABEL PROC PUBLIC _sk_bilinear_nx_hsw _sk_bilinear_nx_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,115,115,2,0 ; vbroadcastss 0x27373(%rip),%ymm0 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,5,139,113,2,0 ; vbroadcastss 0x2718b(%rip),%ymm0 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,10,114,2,0 ; vbroadcastss 0x2720a(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,34,112,2,0 ; vbroadcastss 0x27022(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -5936,7 +5985,7 @@ _sk_bilinear_nx_hsw LABEL PROC PUBLIC _sk_bilinear_px_hsw _sk_bilinear_px_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,234,113,2,0 ; vbroadcastss 0x271ea(%rip),%ymm0 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,5,2,112,2,0 ; vbroadcastss 0x27002(%rip),%ymm0 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -5946,9 +5995,9 @@ _sk_bilinear_px_hsw LABEL PROC PUBLIC _sk_bilinear_ny_hsw _sk_bilinear_ny_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,42,115,2,0 ; vbroadcastss 0x2732a(%rip),%ymm1 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,13,66,113,2,0 ; vbroadcastss 0x27142(%rip),%ymm1 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,192,113,2,0 ; vbroadcastss 0x271c0(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,216,111,2,0 ; vbroadcastss 0x26fd8(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -5957,7 +6006,7 @@ _sk_bilinear_ny_hsw LABEL PROC PUBLIC _sk_bilinear_py_hsw _sk_bilinear_py_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,160,113,2,0 ; vbroadcastss 0x271a0(%rip),%ymm1 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,13,184,111,2,0 ; vbroadcastss 0x26fb8(%rip),%ymm1 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -5967,13 +6016,13 @@ _sk_bilinear_py_hsw LABEL PROC PUBLIC _sk_bicubic_n3x_hsw _sk_bicubic_n3x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,227,114,2,0 ; vbroadcastss 0x272e3(%rip),%ymm0 # 2d40c <_sk_overlay_sse2_8bit+0xeb4> + DB 196,226,125,24,5,251,112,2,0 ; vbroadcastss 0x270fb(%rip),%ymm0 # 2d2ec <_sk_overlay_sse2_8bit+0xe3d> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,118,113,2,0 ; vbroadcastss 0x27176(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,142,111,2,0 ; vbroadcastss 0x26f8e(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,211,113,2,0 ; vbroadcastss 0x271d3(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> - DB 196,98,125,24,29,190,114,2,0 ; vbroadcastss 0x272be(%rip),%ymm11 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,235,111,2,0 ; vbroadcastss 0x26feb(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> + DB 196,98,125,24,29,214,112,2,0 ; vbroadcastss 0x270d6(%rip),%ymm11 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,36,89,193 ; vmulps %ymm9,%ymm11,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -5983,16 +6032,16 @@ _sk_bicubic_n3x_hsw LABEL PROC PUBLIC _sk_bicubic_n1x_hsw _sk_bicubic_n1x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,149,114,2,0 ; vbroadcastss 0x27295(%rip),%ymm0 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,5,173,112,2,0 ; vbroadcastss 0x270ad(%rip),%ymm0 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,44,113,2,0 ; vbroadcastss 0x2712c(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,68,111,2,0 ; vbroadcastss 0x26f44(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,134,114,2,0 ; vbroadcastss 0x27286(%rip),%ymm9 # 2d414 <_sk_overlay_sse2_8bit+0xebc> - DB 196,98,125,24,21,129,114,2,0 ; vbroadcastss 0x27281(%rip),%ymm10 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,13,158,112,2,0 ; vbroadcastss 0x2709e(%rip),%ymm9 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> + DB 196,98,125,24,21,153,112,2,0 ; vbroadcastss 0x27099(%rip),%ymm10 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,3,113,2,0 ; vbroadcastss 0x27103(%rip),%ymm9 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,13,27,111,2,0 ; vbroadcastss 0x26f1b(%rip),%ymm9 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,105,114,2,0 ; vbroadcastss 0x27269(%rip),%ymm10 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,21,129,112,2,0 ; vbroadcastss 0x27081(%rip),%ymm10 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 DB 197,124,17,144,128,0,0,0 ; vmovups %ymm10,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -6001,14 +6050,14 @@ _sk_bicubic_n1x_hsw LABEL PROC PUBLIC _sk_bicubic_p1x_hsw _sk_bicubic_p1x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,217,112,2,0 ; vbroadcastss 0x270d9(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,241,110,2,0 ; vbroadcastss 0x26ef1(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,51,114,2,0 ; vbroadcastss 0x27233(%rip),%ymm10 # 2d414 <_sk_overlay_sse2_8bit+0xebc> - DB 196,98,125,24,29,46,114,2,0 ; vbroadcastss 0x2722e(%rip),%ymm11 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,21,75,112,2,0 ; vbroadcastss 0x2704b(%rip),%ymm10 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> + DB 196,98,125,24,29,70,112,2,0 ; vbroadcastss 0x27046(%rip),%ymm11 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,31,114,2,0 ; vbroadcastss 0x2721f(%rip),%ymm8 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,5,55,112,2,0 ; vbroadcastss 0x27037(%rip),%ymm8 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,66,53,184,195 ; vfmadd231ps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -6017,12 +6066,12 @@ _sk_bicubic_p1x_hsw LABEL PROC PUBLIC _sk_bicubic_p3x_hsw _sk_bicubic_p3x_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,251,113,2,0 ; vbroadcastss 0x271fb(%rip),%ymm0 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,226,125,24,5,19,112,2,0 ; vbroadcastss 0x27013(%rip),%ymm0 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,236,112,2,0 ; vbroadcastss 0x270ec(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> - DB 196,98,125,24,29,215,113,2,0 ; vbroadcastss 0x271d7(%rip),%ymm11 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,4,111,2,0 ; vbroadcastss 0x26f04(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> + DB 196,98,125,24,29,239,111,2,0 ; vbroadcastss 0x26fef(%rip),%ymm11 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,52,89,195 ; vmulps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -6032,13 +6081,13 @@ _sk_bicubic_p3x_hsw LABEL PROC PUBLIC _sk_bicubic_n3y_hsw _sk_bicubic_n3y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,178,113,2,0 ; vbroadcastss 0x271b2(%rip),%ymm1 # 2d40c <_sk_overlay_sse2_8bit+0xeb4> + DB 196,226,125,24,13,202,111,2,0 ; vbroadcastss 0x26fca(%rip),%ymm1 # 2d2ec <_sk_overlay_sse2_8bit+0xe3d> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,68,112,2,0 ; vbroadcastss 0x27044(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,92,110,2,0 ; vbroadcastss 0x26e5c(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,161,112,2,0 ; vbroadcastss 0x270a1(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> - DB 196,98,125,24,29,140,113,2,0 ; vbroadcastss 0x2718c(%rip),%ymm11 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,185,110,2,0 ; vbroadcastss 0x26eb9(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> + DB 196,98,125,24,29,164,111,2,0 ; vbroadcastss 0x26fa4(%rip),%ymm11 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,36,89,193 ; vmulps %ymm9,%ymm11,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -6048,16 +6097,16 @@ _sk_bicubic_n3y_hsw LABEL PROC PUBLIC _sk_bicubic_n1y_hsw _sk_bicubic_n1y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,99,113,2,0 ; vbroadcastss 0x27163(%rip),%ymm1 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,13,123,111,2,0 ; vbroadcastss 0x26f7b(%rip),%ymm1 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,249,111,2,0 ; vbroadcastss 0x26ff9(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,17,110,2,0 ; vbroadcastss 0x26e11(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,83,113,2,0 ; vbroadcastss 0x27153(%rip),%ymm9 # 2d414 <_sk_overlay_sse2_8bit+0xebc> - DB 196,98,125,24,21,78,113,2,0 ; vbroadcastss 0x2714e(%rip),%ymm10 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,13,107,111,2,0 ; vbroadcastss 0x26f6b(%rip),%ymm9 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> + DB 196,98,125,24,21,102,111,2,0 ; vbroadcastss 0x26f66(%rip),%ymm10 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,66,61,168,209 ; vfmadd213ps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,13,208,111,2,0 ; vbroadcastss 0x26fd0(%rip),%ymm9 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,13,232,109,2,0 ; vbroadcastss 0x26de8(%rip),%ymm9 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,66,61,184,202 ; vfmadd231ps %ymm10,%ymm8,%ymm9 - DB 196,98,125,24,21,54,113,2,0 ; vbroadcastss 0x27136(%rip),%ymm10 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,21,78,111,2,0 ; vbroadcastss 0x26f4e(%rip),%ymm10 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,66,61,184,209 ; vfmadd231ps %ymm9,%ymm8,%ymm10 DB 197,124,17,144,160,0,0,0 ; vmovups %ymm10,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -6066,14 +6115,14 @@ _sk_bicubic_n1y_hsw LABEL PROC PUBLIC _sk_bicubic_p1y_hsw _sk_bicubic_p1y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,166,111,2,0 ; vbroadcastss 0x26fa6(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,190,109,2,0 ; vbroadcastss 0x26dbe(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,255,112,2,0 ; vbroadcastss 0x270ff(%rip),%ymm10 # 2d414 <_sk_overlay_sse2_8bit+0xebc> - DB 196,98,125,24,29,250,112,2,0 ; vbroadcastss 0x270fa(%rip),%ymm11 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,21,23,111,2,0 ; vbroadcastss 0x26f17(%rip),%ymm10 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> + DB 196,98,125,24,29,18,111,2,0 ; vbroadcastss 0x26f12(%rip),%ymm11 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,66,53,168,218 ; vfmadd213ps %ymm10,%ymm9,%ymm11 DB 196,66,53,168,216 ; vfmadd213ps %ymm8,%ymm9,%ymm11 - DB 196,98,125,24,5,235,112,2,0 ; vbroadcastss 0x270eb(%rip),%ymm8 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,5,3,111,2,0 ; vbroadcastss 0x26f03(%rip),%ymm8 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,66,53,184,195 ; vfmadd231ps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -6082,12 +6131,12 @@ _sk_bicubic_p1y_hsw LABEL PROC PUBLIC _sk_bicubic_p3y_hsw _sk_bicubic_p3y_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,199,112,2,0 ; vbroadcastss 0x270c7(%rip),%ymm1 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,226,125,24,13,223,110,2,0 ; vbroadcastss 0x26edf(%rip),%ymm1 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,183,111,2,0 ; vbroadcastss 0x26fb7(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> - DB 196,98,125,24,29,162,112,2,0 ; vbroadcastss 0x270a2(%rip),%ymm11 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,207,109,2,0 ; vbroadcastss 0x26dcf(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> + DB 196,98,125,24,29,186,110,2,0 ; vbroadcastss 0x26eba(%rip),%ymm11 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,66,61,168,218 ; vfmadd213ps %ymm10,%ymm8,%ymm11 DB 196,65,52,89,195 ; vmulps %ymm11,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -6166,12 +6215,12 @@ _sk_callback_hsw LABEL PROC PUBLIC _sk_clut_3D_hsw _sk_clut_3D_hsw LABEL PROC - DB 72,129,236,88,2,0,0 ; sub $0x258,%rsp - DB 197,252,17,188,36,32,2,0,0 ; vmovups %ymm7,0x220(%rsp) - DB 197,254,127,180,36,0,2,0,0 ; vmovdqu %ymm6,0x200(%rsp) - DB 197,252,17,172,36,224,1,0,0 ; vmovups %ymm5,0x1e0(%rsp) - DB 197,252,17,164,36,192,1,0,0 ; vmovups %ymm4,0x1c0(%rsp) - DB 197,252,17,156,36,160,1,0,0 ; vmovups %ymm3,0x1a0(%rsp) + DB 72,129,236,184,3,0,0 ; sub $0x3b8,%rsp + DB 197,252,17,188,36,128,3,0,0 ; vmovups %ymm7,0x380(%rsp) + DB 197,252,17,180,36,96,3,0,0 ; vmovups %ymm6,0x360(%rsp) + DB 197,252,17,172,36,64,3,0,0 ; vmovups %ymm5,0x340(%rsp) + DB 197,252,17,164,36,32,3,0,0 ; vmovups %ymm4,0x320(%rsp) + DB 197,252,17,156,36,0,3,0,0 ; vmovups %ymm3,0x300(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -6179,214 +6228,262 @@ _sk_clut_3D_hsw LABEL PROC DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 - DB 197,252,17,84,36,96 ; vmovups %ymm2,0x60(%rsp) DB 197,254,91,234 ; vcvttps2dq %ymm2,%ymm5 + DB 197,252,40,242 ; vmovaps %ymm2,%ymm6 + DB 197,252,17,180,36,64,1,0,0 ; vmovups %ymm6,0x140(%rsp) DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 197,254,127,28,36 ; vmovdqu %ymm3,(%rsp) DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d DB 196,193,121,110,225 ; vmovd %r9d,%xmm4 DB 196,226,125,88,228 ; vpbroadcastd %xmm4,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 197,220,89,201 ; vmulps %ymm1,%ymm4,%ymm1 - DB 197,252,17,76,36,64 ; vmovups %ymm1,0x40(%rsp) - DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 - DB 197,254,127,140,36,128,1,0,0 ; vmovdqu %ymm1,0x180(%rsp) - DB 196,226,117,64,203 ; vpmulld %ymm3,%ymm1,%ymm1 - DB 197,254,127,140,36,224,0,0,0 ; vmovdqu %ymm1,0xe0(%rsp) - DB 197,117,254,229 ; vpaddd %ymm5,%ymm1,%ymm12 - DB 197,125,111,205 ; vmovdqa %ymm5,%ymm9 - DB 197,126,127,140,36,96,1,0,0 ; vmovdqu %ymm9,0x160(%rsp) - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,117,64,202 ; vpmulld %ymm2,%ymm1,%ymm1 - DB 196,226,125,88,233 ; vpbroadcastd %xmm1,%ymm5 + DB 197,254,91,225 ; vcvttps2dq %ymm1,%ymm4 + DB 197,254,127,164,36,224,2,0,0 ; vmovdqu %ymm4,0x2e0(%rsp) + DB 197,124,40,201 ; vmovaps %ymm1,%ymm9 + DB 197,124,17,140,36,96,2,0,0 ; vmovups %ymm9,0x260(%rsp) + DB 196,98,101,64,228 ; vpmulld %ymm4,%ymm3,%ymm12 + DB 197,157,254,205 ; vpaddd %ymm5,%ymm12,%ymm1 + DB 197,125,111,197 ; vmovdqa %ymm5,%ymm8 + DB 197,126,127,132,36,0,2,0,0 ; vmovdqu %ymm8,0x200(%rsp) + DB 196,193,121,110,224 ; vmovd %r8d,%xmm4 + DB 196,226,93,64,210 ; vpmulld %ymm2,%ymm4,%ymm2 + DB 196,226,125,88,226 ; vpbroadcastd %xmm2,%ymm4 DB 68,139,64,8 ; mov 0x8(%rax),%r8d DB 65,255,200 ; dec %r8d - DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 - DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 197,244,89,192 ; vmulps %ymm0,%ymm1,%ymm0 + DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 + DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2 + DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 + DB 197,236,89,192 ; vmulps %ymm0,%ymm2,%ymm0 DB 197,254,91,208 ; vcvttps2dq %ymm0,%ymm2 - DB 196,226,109,64,205 ; vpmulld %ymm5,%ymm2,%ymm1 - DB 197,254,127,76,36,32 ; vmovdqu %ymm1,0x20(%rsp) + DB 197,254,127,148,36,192,2,0,0 ; vmovdqu %ymm2,0x2c0(%rsp) + DB 197,124,40,208 ; vmovaps %ymm0,%ymm10 + DB 197,124,17,148,36,64,2,0,0 ; vmovups %ymm10,0x240(%rsp) + DB 196,98,93,64,218 ; vpmulld %ymm2,%ymm4,%ymm11 DB 72,139,0 ; mov (%rax),%rax - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 - DB 196,226,125,88,61,128,110,2,0 ; vpbroadcastd 0x26e80(%rip),%ymm7 # 2d424 <_sk_overlay_sse2_8bit+0xecc> - DB 196,226,117,64,207 ; vpmulld %ymm7,%ymm1,%ymm1 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,98,77,146,4,136 ; vgatherdps %ymm6,(%rax,%ymm1,4),%ymm8 - DB 196,98,125,88,21,244,108,2,0 ; vpbroadcastd 0x26cf4(%rip),%ymm10 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> - DB 196,193,117,254,242 ; vpaddd %ymm10,%ymm1,%ymm6 - DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 - DB 196,226,37,146,28,176 ; vgatherdps %ymm11,(%rax,%ymm6,4),%ymm3 - DB 196,98,125,88,29,227,108,2,0 ; vpbroadcastd 0x26ce3(%rip),%ymm11 # 2d2b8 <_sk_overlay_sse2_8bit+0xd60> - DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,36,136 ; vgatherdps %ymm6,(%rax,%ymm1,4),%ymm4 - DB 196,98,125,24,45,51,110,2,0 ; vbroadcastss 0x26e33(%rip),%ymm13 # 2d420 <_sk_overlay_sse2_8bit+0xec8> - DB 196,193,124,88,245 ; vaddps %ymm13,%ymm0,%ymm6 - DB 197,254,91,246 ; vcvttps2dq %ymm6,%ymm6 - DB 196,226,77,64,205 ; vpmulld %ymm5,%ymm6,%ymm1 - DB 197,254,127,140,36,128,0,0,0 ; vmovdqu %ymm1,0x80(%rsp) - DB 196,193,117,254,236 ; vpaddd %ymm12,%ymm1,%ymm5 - DB 196,226,85,64,239 ; vpmulld %ymm7,%ymm5,%ymm5 - DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 - DB 196,98,29,146,60,168 ; vgatherdps %ymm12,(%rax,%ymm5,4),%ymm15 - DB 196,65,85,254,226 ; vpaddd %ymm10,%ymm5,%ymm12 - DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 - DB 196,162,13,146,12,160 ; vgatherdps %ymm14,(%rax,%ymm12,4),%ymm1 - DB 196,193,85,254,235 ; vpaddd %ymm11,%ymm5,%ymm5 - DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 - DB 196,226,29,146,52,168 ; vgatherdps %ymm12,(%rax,%ymm5,4),%ymm6 - DB 197,252,91,234 ; vcvtdq2ps %ymm2,%ymm5 - DB 197,124,92,245 ; vsubps %ymm5,%ymm0,%ymm14 - DB 196,193,4,92,192 ; vsubps %ymm8,%ymm15,%ymm0 - DB 196,194,13,168,192 ; vfmadd213ps %ymm8,%ymm14,%ymm0 - DB 197,252,17,132,36,64,1,0,0 ; vmovups %ymm0,0x140(%rsp) - DB 197,244,92,195 ; vsubps %ymm3,%ymm1,%ymm0 - DB 196,226,13,168,195 ; vfmadd213ps %ymm3,%ymm14,%ymm0 - DB 197,252,17,132,36,32,1,0,0 ; vmovups %ymm0,0x120(%rsp) - DB 197,204,92,196 ; vsubps %ymm4,%ymm6,%ymm0 - DB 196,226,13,168,196 ; vfmadd213ps %ymm4,%ymm14,%ymm0 - DB 197,252,17,132,36,0,1,0,0 ; vmovups %ymm0,0x100(%rsp) - DB 197,148,88,68,36,64 ; vaddps 0x40(%rsp),%ymm13,%ymm0 - DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 196,226,125,64,4,36 ; vpmulld (%rsp),%ymm0,%ymm0 - DB 197,254,127,132,36,160,0,0,0 ; vmovdqu %ymm0,0xa0(%rsp) - DB 196,193,125,254,193 ; vpaddd %ymm9,%ymm0,%ymm0 - DB 197,126,111,76,36,32 ; vmovdqu 0x20(%rsp),%ymm9 - DB 196,193,125,254,201 ; vpaddd %ymm9,%ymm0,%ymm1 - DB 196,226,117,64,207 ; vpmulld %ymm7,%ymm1,%ymm1 - DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,36,136 ; vgatherdps %ymm3,(%rax,%ymm1,4),%ymm4 - DB 196,193,117,254,218 ; vpaddd %ymm10,%ymm1,%ymm3 + DB 197,165,254,193 ; vpaddd %ymm1,%ymm11,%ymm0 + DB 196,98,125,88,61,137,108,2,0 ; vpbroadcastd 0x26c89(%rip),%ymm15 # 2d304 <_sk_overlay_sse2_8bit+0xe55> + DB 196,194,125,64,199 ; vpmulld %ymm15,%ymm0,%ymm0 + DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 + DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 196,226,109,146,44,128 ; vgatherdps %ymm2,(%rax,%ymm0,4),%ymm5 + DB 197,252,17,172,36,160,2,0,0 ; vmovups %ymm5,0x2a0(%rsp) + DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 + DB 196,193,125,250,213 ; vpsubd %ymm13,%ymm0,%ymm2 + DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,226,85,146,52,152 ; vgatherdps %ymm5,(%rax,%ymm3,4),%ymm6 - DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 - DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,20,136 ; vgatherdps %ymm3,(%rax,%ymm1,4),%ymm2 - DB 197,126,111,164,36,128,0,0,0 ; vmovdqu 0x80(%rsp),%ymm12 - DB 197,157,254,192 ; vpaddd %ymm0,%ymm12,%ymm0 - DB 196,226,125,64,199 ; vpmulld %ymm7,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,28,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm3 - DB 196,193,125,254,202 ; vpaddd %ymm10,%ymm0,%ymm1 + DB 196,226,85,146,60,144 ; vgatherdps %ymm5,(%rax,%ymm2,4),%ymm7 + DB 197,252,17,188,36,128,2,0,0 ; vmovups %ymm7,0x280(%rsp) + DB 196,226,125,88,61,215,106,2,0 ; vpbroadcastd 0x26ad7(%rip),%ymm7 # 2d198 <_sk_overlay_sse2_8bit+0xce9> + DB 197,253,254,199 ; vpaddd %ymm7,%ymm0,%ymm0 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,98,85,146,4,136 ; vgatherdps %ymm5,(%rax,%ymm1,4),%ymm8 - DB 196,193,125,254,195 ; vpaddd %ymm11,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,44,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm5 - DB 197,228,92,196 ; vsubps %ymm4,%ymm3,%ymm0 - DB 196,226,13,168,196 ; vfmadd213ps %ymm4,%ymm14,%ymm0 - DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) - DB 197,188,92,198 ; vsubps %ymm6,%ymm8,%ymm0 - DB 196,226,13,168,198 ; vfmadd213ps %ymm6,%ymm14,%ymm0 - DB 197,252,17,132,36,192,0,0,0 ; vmovups %ymm0,0xc0(%rsp) - DB 197,84,92,250 ; vsubps %ymm2,%ymm5,%ymm15 - DB 196,98,13,168,250 ; vfmadd213ps %ymm2,%ymm14,%ymm15 - DB 197,148,88,68,36,96 ; vaddps 0x60(%rsp),%ymm13,%ymm0 - DB 197,126,91,232 ; vcvttps2dq %ymm0,%ymm13 - DB 197,149,254,132,36,224,0,0,0 ; vpaddd 0xe0(%rsp),%ymm13,%ymm0 - DB 197,181,254,200 ; vpaddd %ymm0,%ymm9,%ymm1 - DB 196,226,117,64,207 ; vpmulld %ymm7,%ymm1,%ymm1 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,226,109,146,44,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm5 - DB 196,193,117,254,210 ; vpaddd %ymm10,%ymm1,%ymm2 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,28,144 ; vgatherdps %ymm6,(%rax,%ymm2,4),%ymm3 - DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,226,109,146,52,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm6 - DB 197,157,254,192 ; vpaddd %ymm0,%ymm12,%ymm0 - DB 196,226,125,64,199 ; vpmulld %ymm7,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,20,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm2 - DB 196,193,125,254,202 ; vpaddd %ymm10,%ymm0,%ymm1 - DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,226,53,146,36,136 ; vgatherdps %ymm9,(%rax,%ymm1,4),%ymm4 - DB 196,193,125,254,195 ; vpaddd %ymm11,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,98,117,146,4,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm8 - DB 197,236,92,205 ; vsubps %ymm5,%ymm2,%ymm1 - DB 196,226,13,168,205 ; vfmadd213ps %ymm5,%ymm14,%ymm1 - DB 197,92,92,203 ; vsubps %ymm3,%ymm4,%ymm9 - DB 196,98,13,168,203 ; vfmadd213ps %ymm3,%ymm14,%ymm9 - DB 197,188,92,198 ; vsubps %ymm6,%ymm8,%ymm0 - DB 196,226,13,168,198 ; vfmadd213ps %ymm6,%ymm14,%ymm0 - DB 197,149,254,148,36,160,0,0,0 ; vpaddd 0xa0(%rsp),%ymm13,%ymm2 - DB 197,237,254,92,36,32 ; vpaddd 0x20(%rsp),%ymm2,%ymm3 - DB 196,226,101,64,223 ; vpmulld %ymm7,%ymm3,%ymm3 + DB 196,226,85,146,20,128 ; vgatherdps %ymm5,(%rax,%ymm0,4),%ymm2 + DB 197,252,17,148,36,160,1,0,0 ; vmovups %ymm2,0x1a0(%rsp) + DB 196,226,125,24,5,27,108,2,0 ; vbroadcastss 0x26c1b(%rip),%ymm0 # 2d300 <_sk_overlay_sse2_8bit+0xe51> + DB 197,172,88,232 ; vaddps %ymm0,%ymm10,%ymm5 + DB 197,254,91,237 ; vcvttps2dq %ymm5,%ymm5 + DB 196,226,93,64,213 ; vpmulld %ymm5,%ymm4,%ymm2 + DB 197,237,254,201 ; vpaddd %ymm1,%ymm2,%ymm1 + DB 196,194,117,64,207 ; vpmulld %ymm15,%ymm1,%ymm1 + DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 196,226,93,146,44,136 ; vgatherdps %ymm4,(%rax,%ymm1,4),%ymm5 + DB 197,252,17,172,36,32,2,0,0 ; vmovups %ymm5,0x220(%rsp) + DB 196,193,117,250,229 ; vpsubd %ymm13,%ymm1,%ymm4 + DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 + DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 + DB 196,98,85,146,20,160 ; vgatherdps %ymm5,(%rax,%ymm4,4),%ymm10 + DB 197,124,17,148,36,224,1,0,0 ; vmovups %ymm10,0x1e0(%rsp) + DB 197,245,254,207 ; vpaddd %ymm7,%ymm1,%ymm1 + DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 196,226,93,146,44,136 ; vgatherdps %ymm4,(%rax,%ymm1,4),%ymm5 + DB 197,252,17,172,36,192,1,0,0 ; vmovups %ymm5,0x1c0(%rsp) + DB 197,180,88,200 ; vaddps %ymm0,%ymm9,%ymm1 + DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 + DB 196,98,101,64,201 ; vpmulld %ymm1,%ymm3,%ymm9 + DB 196,193,53,254,200 ; vpaddd %ymm8,%ymm9,%ymm1 + DB 196,193,117,254,219 ; vpaddd %ymm11,%ymm1,%ymm3 + DB 196,194,101,64,223 ; vpmulld %ymm15,%ymm3,%ymm3 + DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 DB 196,226,93,146,44,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm5 - DB 196,193,101,254,226 ; vpaddd %ymm10,%ymm3,%ymm4 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,98,77,146,4,160 ; vgatherdps %ymm6,(%rax,%ymm4,4),%ymm8 - DB 196,193,101,254,219 ; vpaddd %ymm11,%ymm3,%ymm3 + DB 197,252,17,172,36,128,1,0,0 ; vmovups %ymm5,0x180(%rsp) + DB 196,193,101,250,229 ; vpsubd %ymm13,%ymm3,%ymm4 + DB 196,65,61,239,192 ; vpxor %ymm8,%ymm8,%ymm8 + DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 + DB 196,98,85,146,4,160 ; vgatherdps %ymm5,(%rax,%ymm4,4),%ymm8 + DB 197,124,17,132,36,96,1,0,0 ; vmovups %ymm8,0x160(%rsp) + DB 197,229,254,223 ; vpaddd %ymm7,%ymm3,%ymm3 + DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 - DB 196,226,93,146,52,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm6 - DB 197,157,254,210 ; vpaddd %ymm2,%ymm12,%ymm2 - DB 196,226,109,64,215 ; vpmulld %ymm7,%ymm2,%ymm2 + DB 196,226,93,146,44,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm5 + DB 197,252,17,172,36,32,1,0,0 ; vmovups %ymm5,0x120(%rsp) + DB 197,237,254,201 ; vpaddd %ymm1,%ymm2,%ymm1 + DB 196,194,117,64,207 ; vpmulld %ymm15,%ymm1,%ymm1 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 - DB 196,193,109,254,218 ; vpaddd %ymm10,%ymm2,%ymm3 - DB 197,197,118,255 ; vpcmpeqd %ymm7,%ymm7,%ymm7 - DB 196,98,69,146,20,152 ; vgatherdps %ymm7,(%rax,%ymm3,4),%ymm10 - DB 196,193,109,254,211 ; vpaddd %ymm11,%ymm2,%ymm2 + DB 196,226,101,146,36,136 ; vgatherdps %ymm3,(%rax,%ymm1,4),%ymm4 + DB 197,252,17,164,36,0,1,0,0 ; vmovups %ymm4,0x100(%rsp) + DB 196,193,117,250,221 ; vpsubd %ymm13,%ymm1,%ymm3 + DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 196,226,93,146,44,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm5 + DB 197,252,17,172,36,224,0,0,0 ; vmovups %ymm5,0xe0(%rsp) + DB 197,245,254,207 ; vpaddd %ymm7,%ymm1,%ymm1 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,60,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm7 - DB 197,220,92,213 ; vsubps %ymm5,%ymm4,%ymm2 - DB 196,226,13,168,213 ; vfmadd213ps %ymm5,%ymm14,%ymm2 - DB 196,193,44,92,216 ; vsubps %ymm8,%ymm10,%ymm3 - DB 196,194,13,168,216 ; vfmadd213ps %ymm8,%ymm14,%ymm3 - DB 197,196,92,230 ; vsubps %ymm6,%ymm7,%ymm4 - DB 196,226,13,168,230 ; vfmadd213ps %ymm6,%ymm14,%ymm4 - DB 197,252,91,172,36,128,1,0,0 ; vcvtdq2ps 0x180(%rsp),%ymm5 - DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 - DB 197,204,92,237 ; vsubps %ymm5,%ymm6,%ymm5 - DB 197,252,16,188,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm7 - DB 197,252,16,52,36 ; vmovups (%rsp),%ymm6 - DB 197,204,92,247 ; vsubps %ymm7,%ymm6,%ymm6 - DB 196,226,85,168,247 ; vfmadd213ps %ymm7,%ymm5,%ymm6 - DB 197,124,16,132,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm8 - DB 197,252,16,188,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm7 - DB 196,193,68,92,248 ; vsubps %ymm8,%ymm7,%ymm7 - DB 196,194,85,168,248 ; vfmadd213ps %ymm8,%ymm5,%ymm7 - DB 197,124,16,148,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm10 - DB 196,65,4,92,194 ; vsubps %ymm10,%ymm15,%ymm8 - DB 196,66,85,168,194 ; vfmadd213ps %ymm10,%ymm5,%ymm8 - DB 197,236,92,209 ; vsubps %ymm1,%ymm2,%ymm2 - DB 196,226,85,168,209 ; vfmadd213ps %ymm1,%ymm5,%ymm2 - DB 196,193,100,92,201 ; vsubps %ymm9,%ymm3,%ymm1 - DB 196,194,85,168,201 ; vfmadd213ps %ymm9,%ymm5,%ymm1 - DB 197,220,92,216 ; vsubps %ymm0,%ymm4,%ymm3 - DB 196,226,85,168,216 ; vfmadd213ps %ymm0,%ymm5,%ymm3 - DB 197,252,91,132,36,96,1,0,0 ; vcvtdq2ps 0x160(%rsp),%ymm0 + DB 196,226,101,146,36,136 ; vgatherdps %ymm3,(%rax,%ymm1,4),%ymm4 + DB 197,252,17,164,36,192,0,0,0 ; vmovups %ymm4,0xc0(%rsp) + DB 197,204,88,192 ; vaddps %ymm0,%ymm6,%ymm0 + DB 197,126,91,208 ; vcvttps2dq %ymm0,%ymm10 + DB 196,193,29,254,218 ; vpaddd %ymm10,%ymm12,%ymm3 + DB 197,165,254,195 ; vpaddd %ymm3,%ymm11,%ymm0 + DB 196,194,125,64,199 ; vpmulld %ymm15,%ymm0,%ymm0 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 + DB 196,226,117,146,36,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm4 + DB 197,252,17,164,36,160,0,0,0 ; vmovups %ymm4,0xa0(%rsp) + DB 196,193,125,250,237 ; vpsubd %ymm13,%ymm0,%ymm5 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 + DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 + DB 196,226,77,146,12,168 ; vgatherdps %ymm6,(%rax,%ymm5,4),%ymm1 + DB 197,252,17,140,36,128,0,0,0 ; vmovups %ymm1,0x80(%rsp) + DB 197,253,254,239 ; vpaddd %ymm7,%ymm0,%ymm5 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 + DB 196,226,77,146,4,168 ; vgatherdps %ymm6,(%rax,%ymm5,4),%ymm0 + DB 197,252,17,68,36,96 ; vmovups %ymm0,0x60(%rsp) + DB 197,237,254,219 ; vpaddd %ymm3,%ymm2,%ymm3 + DB 196,194,101,64,223 ; vpmulld %ymm15,%ymm3,%ymm3 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 + DB 196,226,85,146,4,152 ; vgatherdps %ymm5,(%rax,%ymm3,4),%ymm0 + DB 197,252,17,68,36,64 ; vmovups %ymm0,0x40(%rsp) + DB 196,193,101,250,197 ; vpsubd %ymm13,%ymm3,%ymm0 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 + DB 196,226,13,146,36,128 ; vgatherdps %ymm14,(%rax,%ymm0,4),%ymm4 + DB 197,229,254,199 ; vpaddd %ymm7,%ymm3,%ymm0 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 + DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 + DB 196,226,13,146,12,128 ; vgatherdps %ymm14,(%rax,%ymm0,4),%ymm1 + DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp) + DB 196,193,53,254,202 ; vpaddd %ymm10,%ymm9,%ymm1 + DB 196,65,117,254,211 ; vpaddd %ymm11,%ymm1,%ymm10 + DB 196,66,45,64,247 ; vpmulld %ymm15,%ymm10,%ymm14 + DB 196,65,45,239,210 ; vpxor %ymm10,%ymm10,%ymm10 + DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12 + DB 196,34,29,146,20,176 ; vgatherdps %ymm12,(%rax,%ymm14,4),%ymm10 + DB 196,65,13,250,229 ; vpsubd %ymm13,%ymm14,%ymm12 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 + DB 196,162,37,146,4,160 ; vgatherdps %ymm11,(%rax,%ymm12,4),%ymm0 + DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) + DB 197,13,254,223 ; vpaddd %ymm7,%ymm14,%ymm11 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 + DB 196,34,13,146,4,152 ; vgatherdps %ymm14,(%rax,%ymm11,4),%ymm8 + DB 197,237,254,201 ; vpaddd %ymm1,%ymm2,%ymm1 + DB 196,194,117,64,207 ; vpmulld %ymm15,%ymm1,%ymm1 + DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11 + DB 196,65,13,118,246 ; vpcmpeqd %ymm14,%ymm14,%ymm14 + DB 196,98,13,146,28,136 ; vgatherdps %ymm14,(%rax,%ymm1,4),%ymm11 + DB 196,65,117,250,245 ; vpsubd %ymm13,%ymm1,%ymm14 + DB 196,65,5,239,255 ; vpxor %ymm15,%ymm15,%ymm15 + DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 + DB 196,34,53,146,60,176 ; vgatherdps %ymm9,(%rax,%ymm14,4),%ymm15 + DB 197,245,254,207 ; vpaddd %ymm7,%ymm1,%ymm1 + DB 197,197,239,255 ; vpxor %ymm7,%ymm7,%ymm7 + DB 196,226,21,146,60,136 ; vgatherdps %ymm13,(%rax,%ymm1,4),%ymm7 + DB 197,252,91,140,36,192,2,0,0 ; vcvtdq2ps 0x2c0(%rsp),%ymm1 + DB 197,252,16,132,36,64,2,0,0 ; vmovups 0x240(%rsp),%ymm0 + DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 + DB 197,252,16,132,36,160,2,0,0 ; vmovups 0x2a0(%rsp),%ymm0 + DB 197,252,16,156,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm3 + DB 197,100,92,200 ; vsubps %ymm0,%ymm3,%ymm9 + DB 196,98,117,168,200 ; vfmadd213ps %ymm0,%ymm1,%ymm9 + DB 197,252,16,132,36,128,2,0,0 ; vmovups 0x280(%rsp),%ymm0 + DB 197,252,16,156,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm3 + DB 197,100,92,232 ; vsubps %ymm0,%ymm3,%ymm13 + DB 196,98,117,168,232 ; vfmadd213ps %ymm0,%ymm1,%ymm13 + DB 197,252,16,148,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm2 + DB 197,252,16,132,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm0 + DB 197,124,92,242 ; vsubps %ymm2,%ymm0,%ymm14 + DB 196,98,117,168,242 ; vfmadd213ps %ymm2,%ymm1,%ymm14 + DB 197,252,16,132,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm0 + DB 197,252,16,148,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm2 + DB 197,108,92,224 ; vsubps %ymm0,%ymm2,%ymm12 + DB 196,98,117,168,224 ; vfmadd213ps %ymm0,%ymm1,%ymm12 + DB 197,252,16,132,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm0 + DB 197,252,16,148,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm2 + DB 197,236,92,232 ; vsubps %ymm0,%ymm2,%ymm5 + DB 196,226,117,168,232 ; vfmadd213ps %ymm0,%ymm1,%ymm5 + DB 197,252,16,132,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm0 + DB 197,252,16,148,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm2 + DB 197,236,92,240 ; vsubps %ymm0,%ymm2,%ymm6 + DB 196,226,117,168,240 ; vfmadd213ps %ymm0,%ymm1,%ymm6 + DB 197,252,16,148,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm2 + DB 197,252,16,68,36,64 ; vmovups 0x40(%rsp),%ymm0 + DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0 + DB 196,226,117,168,194 ; vfmadd213ps %ymm2,%ymm1,%ymm0 + DB 197,252,16,156,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm3 + DB 197,220,92,211 ; vsubps %ymm3,%ymm4,%ymm2 + DB 196,226,117,168,211 ; vfmadd213ps %ymm3,%ymm1,%ymm2 DB 197,252,16,100,36,96 ; vmovups 0x60(%rsp),%ymm4 - DB 197,220,92,224 ; vsubps %ymm0,%ymm4,%ymm4 - DB 197,236,92,198 ; vsubps %ymm6,%ymm2,%ymm0 - DB 196,226,93,168,198 ; vfmadd213ps %ymm6,%ymm4,%ymm0 - DB 197,244,92,207 ; vsubps %ymm7,%ymm1,%ymm1 - DB 196,226,93,168,207 ; vfmadd213ps %ymm7,%ymm4,%ymm1 - DB 196,193,100,92,208 ; vsubps %ymm8,%ymm3,%ymm2 - DB 196,194,93,168,208 ; vfmadd213ps %ymm8,%ymm4,%ymm2 + DB 197,252,16,92,36,32 ; vmovups 0x20(%rsp),%ymm3 + DB 197,228,92,220 ; vsubps %ymm4,%ymm3,%ymm3 + DB 196,226,117,168,220 ; vfmadd213ps %ymm4,%ymm1,%ymm3 + DB 196,65,36,92,218 ; vsubps %ymm10,%ymm11,%ymm11 + DB 196,66,117,168,218 ; vfmadd213ps %ymm10,%ymm1,%ymm11 + DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 + DB 197,4,92,212 ; vsubps %ymm4,%ymm15,%ymm10 + DB 196,98,117,168,212 ; vfmadd213ps %ymm4,%ymm1,%ymm10 + DB 196,193,68,92,248 ; vsubps %ymm8,%ymm7,%ymm7 + DB 196,194,117,168,248 ; vfmadd213ps %ymm8,%ymm1,%ymm7 + DB 197,252,91,140,36,224,2,0,0 ; vcvtdq2ps 0x2e0(%rsp),%ymm1 + DB 197,124,16,132,36,96,2,0,0 ; vmovups 0x260(%rsp),%ymm8 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 196,193,28,92,225 ; vsubps %ymm9,%ymm12,%ymm4 + DB 196,194,117,168,225 ; vfmadd213ps %ymm9,%ymm1,%ymm4 + DB 196,193,84,92,237 ; vsubps %ymm13,%ymm5,%ymm5 + DB 196,194,117,168,237 ; vfmadd213ps %ymm13,%ymm1,%ymm5 + DB 196,193,76,92,246 ; vsubps %ymm14,%ymm6,%ymm6 + DB 196,194,117,168,246 ; vfmadd213ps %ymm14,%ymm1,%ymm6 + DB 197,36,92,192 ; vsubps %ymm0,%ymm11,%ymm8 + DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8 + DB 197,44,92,202 ; vsubps %ymm2,%ymm10,%ymm9 + DB 196,98,117,168,202 ; vfmadd213ps %ymm2,%ymm1,%ymm9 + DB 197,196,92,211 ; vsubps %ymm3,%ymm7,%ymm2 + DB 196,226,117,168,211 ; vfmadd213ps %ymm3,%ymm1,%ymm2 + DB 197,252,91,132,36,0,2,0,0 ; vcvtdq2ps 0x200(%rsp),%ymm0 + DB 197,252,16,140,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm1 + DB 197,244,92,216 ; vsubps %ymm0,%ymm1,%ymm3 + DB 197,188,92,196 ; vsubps %ymm4,%ymm8,%ymm0 + DB 196,226,101,168,196 ; vfmadd213ps %ymm4,%ymm3,%ymm0 + DB 197,180,92,205 ; vsubps %ymm5,%ymm9,%ymm1 + DB 196,226,101,168,205 ; vfmadd213ps %ymm5,%ymm3,%ymm1 + DB 197,236,92,214 ; vsubps %ymm6,%ymm2,%ymm2 + DB 196,226,101,168,214 ; vfmadd213ps %ymm6,%ymm3,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,156,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm3 - DB 197,252,16,164,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm4 - DB 197,252,16,172,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm5 - DB 197,252,16,180,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm6 - DB 197,252,16,188,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm7 - DB 72,129,196,88,2,0,0 ; add $0x258,%rsp + DB 197,252,16,156,36,0,3,0,0 ; vmovups 0x300(%rsp),%ymm3 + DB 197,252,16,164,36,32,3,0,0 ; vmovups 0x320(%rsp),%ymm4 + DB 197,252,16,172,36,64,3,0,0 ; vmovups 0x340(%rsp),%ymm5 + DB 197,252,16,180,36,96,3,0,0 ; vmovups 0x360(%rsp),%ymm6 + DB 197,252,16,188,36,128,3,0,0 ; vmovups 0x380(%rsp),%ymm7 + DB 72,129,196,184,3,0,0 ; add $0x3b8,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_clut_4D_hsw _sk_clut_4D_hsw LABEL PROC - DB 72,129,236,56,4,0,0 ; sub $0x438,%rsp - DB 197,252,17,188,36,0,4,0,0 ; vmovups %ymm7,0x400(%rsp) - DB 197,252,17,180,36,224,3,0,0 ; vmovups %ymm6,0x3e0(%rsp) - DB 197,252,17,172,36,192,3,0,0 ; vmovups %ymm5,0x3c0(%rsp) - DB 197,252,17,164,36,160,3,0,0 ; vmovups %ymm4,0x3a0(%rsp) + DB 72,129,236,120,7,0,0 ; sub $0x778,%rsp + DB 197,252,17,188,36,64,7,0,0 ; vmovups %ymm7,0x740(%rsp) + DB 197,252,17,180,36,32,7,0,0 ; vmovups %ymm6,0x720(%rsp) + DB 197,252,17,172,36,0,7,0,0 ; vmovups %ymm5,0x700(%rsp) + DB 197,252,17,164,36,224,6,0,0 ; vmovups %ymm4,0x6e0(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,20 ; mov 0x14(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -6394,36 +6491,35 @@ _sk_clut_4D_hsw LABEL PROC DB 196,226,125,88,228 ; vpbroadcastd %xmm4,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 197,220,89,219 ; vmulps %ymm3,%ymm4,%ymm3 - DB 197,252,17,156,36,32,1,0,0 ; vmovups %ymm3,0x120(%rsp) + DB 197,252,17,156,36,160,1,0,0 ; vmovups %ymm3,0x1a0(%rsp) DB 197,254,91,243 ; vcvttps2dq %ymm3,%ymm6 - DB 197,254,127,180,36,0,1,0,0 ; vmovdqu %ymm6,0x100(%rsp) - DB 196,193,121,110,224 ; vmovd %r8d,%xmm4 DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 - DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3 - DB 197,254,127,92,36,64 ; vmovdqu %ymm3,0x40(%rsp) + DB 196,193,121,110,224 ; vmovd %r8d,%xmm4 + DB 196,98,125,88,252 ; vpbroadcastd %xmm4,%ymm15 DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d DB 196,193,121,110,233 ; vmovd %r9d,%xmm5 DB 196,226,125,88,237 ; vpbroadcastd %xmm5,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 DB 197,212,89,210 ; vmulps %ymm2,%ymm5,%ymm2 - DB 197,252,17,148,36,224,0,0,0 ; vmovups %ymm2,0xe0(%rsp) - DB 197,254,91,210 ; vcvttps2dq %ymm2,%ymm2 - DB 197,254,127,148,36,128,3,0,0 ; vmovdqu %ymm2,0x380(%rsp) - DB 196,226,109,64,211 ; vpmulld %ymm3,%ymm2,%ymm2 - DB 197,254,127,84,36,96 ; vmovdqu %ymm2,0x60(%rsp) - DB 197,237,254,222 ; vpaddd %ymm6,%ymm2,%ymm3 - DB 197,254,127,156,36,128,0,0,0 ; vmovdqu %ymm3,0x80(%rsp) + DB 197,254,91,234 ; vcvttps2dq %ymm2,%ymm5 + DB 197,254,127,172,36,192,6,0,0 ; vmovdqu %ymm5,0x6c0(%rsp) + DB 197,124,40,242 ; vmovaps %ymm2,%ymm14 + DB 197,124,17,180,36,96,6,0,0 ; vmovups %ymm14,0x660(%rsp) + DB 196,226,5,64,213 ; vpmulld %ymm5,%ymm15,%ymm2 + DB 197,254,127,84,36,32 ; vmovdqu %ymm2,0x20(%rsp) + DB 197,237,254,230 ; vpaddd %ymm6,%ymm2,%ymm4 + DB 197,125,111,230 ; vmovdqa %ymm6,%ymm12 + DB 197,126,127,164,36,64,6,0,0 ; vmovdqu %ymm12,0x640(%rsp) DB 196,193,121,110,232 ; vmovd %r8d,%xmm5 - DB 196,226,85,64,236 ; vpmulld %ymm4,%ymm5,%ymm5 - DB 196,226,125,88,253 ; vpbroadcastd %xmm5,%ymm7 - DB 197,254,127,188,36,160,0,0,0 ; vmovdqu %ymm7,0xa0(%rsp) + DB 196,226,85,64,235 ; vpmulld %ymm3,%ymm5,%ymm5 + DB 196,226,125,88,221 ; vpbroadcastd %xmm5,%ymm3 DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d DB 196,193,121,110,241 ; vmovd %r9d,%xmm6 DB 196,226,125,88,246 ; vpbroadcastd %xmm6,%ymm6 DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 - DB 197,204,89,225 ; vmulps %ymm1,%ymm6,%ymm4 + DB 197,204,89,241 ; vmulps %ymm1,%ymm6,%ymm6 DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,226,117,64,205 ; vpmulld %ymm5,%ymm1,%ymm1 DB 68,139,64,8 ; mov 0x8(%rax),%r8d @@ -6432,364 +6528,479 @@ _sk_clut_4D_hsw LABEL PROC DB 196,226,125,88,237 ; vpbroadcastd %xmm5,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 DB 197,212,89,232 ; vmulps %ymm0,%ymm5,%ymm5 - DB 197,254,91,196 ; vcvttps2dq %ymm4,%ymm0 - DB 197,254,127,132,36,96,3,0,0 ; vmovdqu %ymm0,0x360(%rsp) - DB 197,252,17,164,36,224,2,0,0 ; vmovups %ymm4,0x2e0(%rsp) - DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1 + DB 197,254,91,198 ; vcvttps2dq %ymm6,%ymm0 + DB 197,254,127,132,36,160,6,0,0 ; vmovdqu %ymm0,0x6a0(%rsp) + DB 197,252,17,180,36,128,6,0,0 ; vmovups %ymm6,0x680(%rsp) + DB 196,226,125,88,209 ; vpbroadcastd %xmm1,%ymm2 DB 72,139,0 ; mov (%rax),%rax - DB 196,226,125,64,199 ; vpmulld %ymm7,%ymm0,%ymm0 + DB 196,226,101,64,192 ; vpmulld %ymm0,%ymm3,%ymm0 DB 197,254,127,4,36 ; vmovdqu %ymm0,(%rsp) - DB 197,253,254,243 ; vpaddd %ymm3,%ymm0,%ymm6 - DB 197,126,91,245 ; vcvttps2dq %ymm5,%ymm14 - DB 196,98,13,64,233 ; vpmulld %ymm1,%ymm14,%ymm13 - DB 197,149,254,198 ; vpaddd %ymm6,%ymm13,%ymm0 - DB 196,98,125,88,21,172,105,2,0 ; vpbroadcastd 0x269ac(%rip),%ymm10 # 2d424 <_sk_overlay_sse2_8bit+0xecc> + DB 197,253,254,252 ; vpaddd %ymm4,%ymm0,%ymm7 + DB 197,254,91,197 ; vcvttps2dq %ymm5,%ymm0 + DB 197,252,40,205 ; vmovaps %ymm5,%ymm1 + DB 197,252,17,140,36,32,1,0,0 ; vmovups %ymm1,0x120(%rsp) + DB 197,254,127,132,36,128,1,0,0 ; vmovdqu %ymm0,0x180(%rsp) + DB 196,98,109,64,200 ; vpmulld %ymm0,%ymm2,%ymm9 + DB 197,181,254,199 ; vpaddd %ymm7,%ymm9,%ymm0 + DB 196,98,125,88,21,157,102,2,0 ; vpbroadcastd 0x2669d(%rip),%ymm10 # 2d304 <_sk_overlay_sse2_8bit+0xe55> DB 196,194,125,64,194 ; vpmulld %ymm10,%ymm0,%ymm0 + DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,98,85,146,4,128 ; vgatherdps %ymm5,(%rax,%ymm0,4),%ymm8 + DB 197,124,17,132,36,96,1,0,0 ; vmovups %ymm8,0x160(%rsp) + DB 196,65,21,118,237 ; vpcmpeqd %ymm13,%ymm13,%ymm13 + DB 196,193,125,250,237 ; vpsubd %ymm13,%ymm0,%ymm5 DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,98,61,146,60,128 ; vgatherdps %ymm8,(%rax,%ymm0,4),%ymm15 - DB 196,98,125,88,29,31,104,2,0 ; vpbroadcastd 0x2681f(%rip),%ymm11 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> - DB 196,65,125,254,195 ; vpaddd %ymm11,%ymm0,%ymm8 - DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9 - DB 196,162,53,146,20,128 ; vgatherdps %ymm9,(%rax,%ymm8,4),%ymm2 - DB 196,98,125,88,37,14,104,2,0 ; vpbroadcastd 0x2680e(%rip),%ymm12 # 2d2b8 <_sk_overlay_sse2_8bit+0xd60> - DB 196,193,125,254,196 ; vpaddd %ymm12,%ymm0,%ymm0 + DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11 + DB 196,98,61,146,28,168 ; vgatherdps %ymm8,(%rax,%ymm5,4),%ymm11 + DB 197,124,17,156,36,64,1,0,0 ; vmovups %ymm11,0x140(%rsp) + DB 196,98,125,88,29,232,100,2,0 ; vpbroadcastd 0x264e8(%rip),%ymm11 # 2d198 <_sk_overlay_sse2_8bit+0xce9> + DB 196,193,125,254,195 ; vpaddd %ymm11,%ymm0,%ymm0 DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,226,61,146,60,128 ; vgatherdps %ymm8,(%rax,%ymm0,4),%ymm7 - DB 196,98,125,24,13,93,105,2,0 ; vbroadcastss 0x2695d(%rip),%ymm9 # 2d420 <_sk_overlay_sse2_8bit+0xec8> - DB 196,65,84,88,193 ; vaddps %ymm9,%ymm5,%ymm8 + DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 + DB 196,226,61,146,44,128 ; vgatherdps %ymm8,(%rax,%ymm0,4),%ymm5 + DB 197,252,17,172,36,160,0,0,0 ; vmovups %ymm5,0xa0(%rsp) + DB 196,226,125,24,5,42,102,2,0 ; vbroadcastss 0x2662a(%rip),%ymm0 # 2d300 <_sk_overlay_sse2_8bit+0xe51> + DB 197,116,88,192 ; vaddps %ymm0,%ymm1,%ymm8 DB 196,65,126,91,192 ; vcvttps2dq %ymm8,%ymm8 - DB 196,226,61,64,193 ; vpmulld %ymm1,%ymm8,%ymm0 - DB 197,254,127,132,36,192,0,0,0 ; vmovdqu %ymm0,0xc0(%rsp) - DB 197,253,254,206 ; vpaddd %ymm6,%ymm0,%ymm1 + DB 196,194,109,64,232 ; vpmulld %ymm8,%ymm2,%ymm5 + DB 197,213,254,207 ; vpaddd %ymm7,%ymm5,%ymm1 DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,98,77,146,4,136 ; vgatherdps %ymm6,(%rax,%ymm1,4),%ymm8 - DB 196,193,117,254,243 ; vpaddd %ymm11,%ymm1,%ymm6 + DB 197,197,118,255 ; vpcmpeqd %ymm7,%ymm7,%ymm7 + DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 + DB 196,226,69,146,20,136 ; vgatherdps %ymm7,(%rax,%ymm1,4),%ymm2 + DB 197,252,17,148,36,0,1,0,0 ; vmovups %ymm2,0x100(%rsp) + DB 196,193,117,250,253 ; vpsubd %ymm13,%ymm1,%ymm7 + DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 + DB 196,226,61,146,20,184 ; vgatherdps %ymm8,(%rax,%ymm7,4),%ymm2 + DB 197,252,17,148,36,224,0,0,0 ; vmovups %ymm2,0xe0(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 + DB 197,197,118,255 ; vpcmpeqd %ymm7,%ymm7,%ymm7 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 + DB 196,226,69,146,20,136 ; vgatherdps %ymm7,(%rax,%ymm1,4),%ymm2 + DB 197,252,17,148,36,192,0,0,0 ; vmovups %ymm2,0xc0(%rsp) + DB 197,204,88,200 ; vaddps %ymm0,%ymm6,%ymm1 + DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 + DB 196,226,101,64,241 ; vpmulld %ymm1,%ymm3,%ymm6 + DB 197,205,254,204 ; vpaddd %ymm4,%ymm6,%ymm1 + DB 196,193,117,254,209 ; vpaddd %ymm9,%ymm1,%ymm2 + DB 196,194,109,64,210 ; vpmulld %ymm10,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,4,176 ; vgatherdps %ymm3,(%rax,%ymm6,4),%ymm0 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 + DB 197,221,239,228 ; vpxor %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,224,5,0,0 ; vmovups %ymm4,0x5e0(%rsp) + DB 196,193,109,250,221 ; vpsubd %ymm13,%ymm2,%ymm3 + DB 197,197,118,255 ; vpcmpeqd %ymm7,%ymm7,%ymm7 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,69,146,36,152 ; vgatherdps %ymm7,(%rax,%ymm3,4),%ymm4 + DB 197,252,17,164,36,128,0,0,0 ; vmovups %ymm4,0x80(%rsp) + DB 196,193,109,254,211 ; vpaddd %ymm11,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,52,136 ; vgatherdps %ymm3,(%rax,%ymm1,4),%ymm6 - DB 196,193,124,91,206 ; vcvtdq2ps %ymm14,%ymm1 - DB 197,84,92,241 ; vsubps %ymm1,%ymm5,%ymm14 - DB 196,193,60,92,207 ; vsubps %ymm15,%ymm8,%ymm1 - DB 196,194,13,168,207 ; vfmadd213ps %ymm15,%ymm14,%ymm1 - DB 197,252,17,140,36,64,3,0,0 ; vmovups %ymm1,0x340(%rsp) - DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0 - DB 196,226,13,168,194 ; vfmadd213ps %ymm2,%ymm14,%ymm0 - DB 197,252,17,132,36,32,3,0,0 ; vmovups %ymm0,0x320(%rsp) - DB 197,204,92,199 ; vsubps %ymm7,%ymm6,%ymm0 - DB 196,226,13,168,199 ; vfmadd213ps %ymm7,%ymm14,%ymm0 - DB 197,252,17,132,36,0,3,0,0 ; vmovups %ymm0,0x300(%rsp) - DB 197,124,17,76,36,32 ; vmovups %ymm9,0x20(%rsp) - DB 196,193,92,88,193 ; vaddps %ymm9,%ymm4,%ymm0 - DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 196,226,125,64,140,36,160,0,0,0 ; vpmulld 0xa0(%rsp),%ymm0,%ymm1 - DB 197,245,254,132,36,128,0,0,0 ; vpaddd 0x80(%rsp),%ymm1,%ymm0 - DB 197,125,111,249 ; vmovdqa %ymm1,%ymm15 - DB 196,193,125,254,205 ; vpaddd %ymm13,%ymm0,%ymm1 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,100,36,96 ; vmovups %ymm4,0x60(%rsp) + DB 197,213,254,201 ; vpaddd %ymm1,%ymm5,%ymm1 DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 - DB 196,193,117,254,211 ; vpaddd %ymm11,%ymm1,%ymm2 - DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,226,85,146,52,144 ; vgatherdps %ymm5,(%rax,%ymm2,4),%ymm6 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 + DB 197,252,17,92,36,64 ; vmovups %ymm3,0x40(%rsp) + DB 196,193,117,250,213 ; vpsubd %ymm13,%ymm1,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,32,6,0,0 ; vmovups %ymm4,0x620(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,226,109,146,44,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm5 - DB 197,254,111,164,36,192,0,0,0 ; vmovdqu 0xc0(%rsp),%ymm4 - DB 197,221,254,192 ; vpaddd %ymm0,%ymm4,%ymm0 - DB 196,194,125,64,194 ; vpmulld %ymm10,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,20,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm2 - DB 196,193,125,254,203 ; vpaddd %ymm11,%ymm0,%ymm1 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 + DB 197,252,17,156,36,0,6,0,0 ; vmovups %ymm3,0x600(%rsp) + DB 197,140,88,200 ; vaddps %ymm0,%ymm14,%ymm1 + DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 + DB 196,98,5,64,249 ; vpmulld %ymm1,%ymm15,%ymm15 + DB 196,193,5,254,204 ; vpaddd %ymm12,%ymm15,%ymm1 + DB 197,126,111,4,36 ; vmovdqu (%rsp),%ymm8 + DB 196,193,117,254,208 ; vpaddd %ymm8,%ymm1,%ymm2 + DB 196,193,109,254,217 ; vpaddd %ymm9,%ymm2,%ymm3 + DB 196,194,101,64,218 ; vpmulld %ymm10,%ymm3,%ymm3 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 + DB 196,226,93,146,60,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm7 + DB 197,252,17,188,36,224,3,0,0 ; vmovups %ymm7,0x3e0(%rsp) + DB 196,193,101,250,229 ; vpsubd %ymm13,%ymm3,%ymm4 DB 197,197,118,255 ; vpcmpeqd %ymm7,%ymm7,%ymm7 - DB 196,98,69,146,4,136 ; vgatherdps %ymm7,(%rax,%ymm1,4),%ymm8 - DB 196,193,125,254,196 ; vpaddd %ymm12,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,60,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm7 - DB 197,236,92,195 ; vsubps %ymm3,%ymm2,%ymm0 - DB 196,226,13,168,195 ; vfmadd213ps %ymm3,%ymm14,%ymm0 - DB 197,252,17,132,36,160,0,0,0 ; vmovups %ymm0,0xa0(%rsp) - DB 197,188,92,198 ; vsubps %ymm6,%ymm8,%ymm0 - DB 196,226,13,168,198 ; vfmadd213ps %ymm6,%ymm14,%ymm0 - DB 197,252,17,132,36,128,0,0,0 ; vmovups %ymm0,0x80(%rsp) - DB 197,196,92,197 ; vsubps %ymm5,%ymm7,%ymm0 - DB 196,226,13,168,197 ; vfmadd213ps %ymm5,%ymm14,%ymm0 - DB 197,252,17,132,36,192,2,0,0 ; vmovups %ymm0,0x2c0(%rsp) - DB 197,180,88,132,36,224,0,0,0 ; vaddps 0xe0(%rsp),%ymm9,%ymm0 - DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 196,226,125,64,68,36,64 ; vpmulld 0x40(%rsp),%ymm0,%ymm0 - DB 197,254,127,132,36,0,2,0,0 ; vmovdqu %ymm0,0x200(%rsp) - DB 197,253,254,132,36,0,1,0,0 ; vpaddd 0x100(%rsp),%ymm0,%ymm0 - DB 197,253,254,12,36 ; vpaddd (%rsp),%ymm0,%ymm1 - DB 196,193,117,254,213 ; vpaddd %ymm13,%ymm1,%ymm2 - DB 196,65,125,111,205 ; vmovdqa %ymm13,%ymm9 + DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 + DB 196,98,69,146,52,160 ; vgatherdps %ymm7,(%rax,%ymm4,4),%ymm14 + DB 197,124,17,180,36,192,5,0,0 ; vmovups %ymm14,0x5c0(%rsp) + DB 196,193,101,254,219 ; vpaddd %ymm11,%ymm3,%ymm3 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 + DB 196,226,93,146,60,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm7 + DB 197,252,17,188,36,64,5,0,0 ; vmovups %ymm7,0x540(%rsp) + DB 197,213,254,210 ; vpaddd %ymm2,%ymm5,%ymm2 DB 196,194,109,64,210 ; vpmulld %ymm10,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,44,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm5 - DB 196,193,109,254,219 ; vpaddd %ymm11,%ymm2,%ymm3 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,60,152 ; vgatherdps %ymm6,(%rax,%ymm3,4),%ymm7 - DB 196,193,109,254,212 ; vpaddd %ymm12,%ymm2,%ymm2 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,160,5,0,0 ; vmovups %ymm4,0x5a0(%rsp) + DB 196,193,109,250,221 ; vpsubd %ymm13,%ymm2,%ymm3 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 + DB 196,98,93,146,52,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm14 + DB 197,124,17,180,36,128,5,0,0 ; vmovups %ymm14,0x580(%rsp) + DB 196,193,109,254,211 ; vpaddd %ymm11,%ymm2,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,96,5,0,0 ; vmovups %ymm4,0x560(%rsp) + DB 197,205,254,201 ; vpaddd %ymm1,%ymm6,%ymm1 + DB 196,193,117,254,209 ; vpaddd %ymm9,%ymm1,%ymm2 + DB 196,194,109,64,210 ; vpmulld %ymm10,%ymm2,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,32,5,0,0 ; vmovups %ymm4,0x520(%rsp) + DB 196,193,109,250,221 ; vpsubd %ymm13,%ymm2,%ymm3 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 197,196,87,255 ; vxorps %ymm7,%ymm7,%ymm7 + DB 196,226,93,146,60,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm7 + DB 197,252,17,188,36,0,5,0,0 ; vmovups %ymm7,0x500(%rsp) + DB 196,193,109,254,211 ; vpaddd %ymm11,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,52,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm6 - DB 197,221,254,201 ; vpaddd %ymm1,%ymm4,%ymm1 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,128,3,0,0 ; vmovups %ymm4,0x380(%rsp) + DB 197,213,254,201 ; vpaddd %ymm1,%ymm5,%ymm1 DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 - DB 196,193,117,254,211 ; vpaddd %ymm11,%ymm1,%ymm2 - DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,98,61,146,44,144 ; vgatherdps %ymm8,(%rax,%ymm2,4),%ymm13 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,98,109,146,4,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm8 - DB 197,228,92,205 ; vsubps %ymm5,%ymm3,%ymm1 - DB 196,226,13,168,205 ; vfmadd213ps %ymm5,%ymm14,%ymm1 - DB 197,252,17,76,36,64 ; vmovups %ymm1,0x40(%rsp) - DB 197,148,92,207 ; vsubps %ymm7,%ymm13,%ymm1 - DB 196,226,13,168,207 ; vfmadd213ps %ymm7,%ymm14,%ymm1 - DB 197,252,17,140,36,160,2,0,0 ; vmovups %ymm1,0x2a0(%rsp) - DB 197,188,92,206 ; vsubps %ymm6,%ymm8,%ymm1 - DB 196,226,13,168,206 ; vfmadd213ps %ymm6,%ymm14,%ymm1 - DB 197,252,17,140,36,128,2,0,0 ; vmovups %ymm1,0x280(%rsp) - DB 197,126,127,188,36,128,1,0,0 ; vmovdqu %ymm15,0x180(%rsp) - DB 197,133,254,192 ; vpaddd %ymm0,%ymm15,%ymm0 - DB 196,193,125,254,201 ; vpaddd %ymm9,%ymm0,%ymm1 - DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 + DB 197,252,17,156,36,224,4,0,0 ; vmovups %ymm3,0x4e0(%rsp) + DB 196,193,117,250,213 ; vpsubd %ymm13,%ymm1,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,192,4,0,0 ; vmovups %ymm4,0x4c0(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 - DB 196,193,117,254,211 ; vpaddd %ymm11,%ymm1,%ymm2 - DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,226,85,146,52,144 ; vgatherdps %ymm5,(%rax,%ymm2,4),%ymm6 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 - DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,226,109,146,44,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm5 - DB 197,221,254,192 ; vpaddd %ymm0,%ymm4,%ymm0 - DB 196,194,125,64,194 ; vpmulld %ymm10,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,20,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm2 - DB 196,193,125,254,203 ; vpaddd %ymm11,%ymm0,%ymm1 - DB 197,197,118,255 ; vpcmpeqd %ymm7,%ymm7,%ymm7 - DB 196,98,69,146,4,136 ; vgatherdps %ymm7,(%rax,%ymm1,4),%ymm8 - DB 196,193,125,254,196 ; vpaddd %ymm12,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,60,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm7 - DB 197,236,92,195 ; vsubps %ymm3,%ymm2,%ymm0 - DB 196,226,13,168,195 ; vfmadd213ps %ymm3,%ymm14,%ymm0 - DB 197,252,17,132,36,96,2,0,0 ; vmovups %ymm0,0x260(%rsp) - DB 197,188,92,198 ; vsubps %ymm6,%ymm8,%ymm0 - DB 196,226,13,168,198 ; vfmadd213ps %ymm6,%ymm14,%ymm0 - DB 197,252,17,132,36,64,2,0,0 ; vmovups %ymm0,0x240(%rsp) - DB 197,196,92,197 ; vsubps %ymm5,%ymm7,%ymm0 - DB 196,226,13,168,197 ; vfmadd213ps %ymm5,%ymm14,%ymm0 - DB 197,252,17,132,36,32,2,0,0 ; vmovups %ymm0,0x220(%rsp) - DB 197,252,16,68,36,32 ; vmovups 0x20(%rsp),%ymm0 - DB 197,252,88,132,36,32,1,0,0 ; vaddps 0x120(%rsp),%ymm0,%ymm0 - DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 197,254,127,132,36,160,1,0,0 ; vmovdqu %ymm0,0x1a0(%rsp) - DB 197,125,254,68,36,96 ; vpaddd 0x60(%rsp),%ymm0,%ymm8 - DB 197,189,254,12,36 ; vpaddd (%rsp),%ymm8,%ymm1 + DB 197,252,17,156,36,160,4,0,0 ; vmovups %ymm3,0x4a0(%rsp) + DB 197,252,88,132,36,160,1,0,0 ; vaddps 0x1a0(%rsp),%ymm0,%ymm0 + DB 197,254,91,248 ; vcvttps2dq %ymm0,%ymm7 + DB 197,197,254,68,36,32 ; vpaddd 0x20(%rsp),%ymm7,%ymm0 + DB 197,189,254,200 ; vpaddd %ymm0,%ymm8,%ymm1 DB 197,181,254,209 ; vpaddd %ymm1,%ymm9,%ymm2 - DB 197,125,127,207 ; vmovdqa %ymm9,%ymm7 DB 196,194,109,64,210 ; vpmulld %ymm10,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 - DB 196,193,109,254,219 ; vpaddd %ymm11,%ymm2,%ymm3 - DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,98,85,146,44,152 ; vgatherdps %ymm5,(%rax,%ymm3,4),%ymm13 - DB 196,193,109,254,212 ; vpaddd %ymm12,%ymm2,%ymm2 + DB 197,252,17,100,36,32 ; vmovups %ymm4,0x20(%rsp) + DB 196,193,109,250,221 ; vpsubd %ymm13,%ymm2,%ymm3 + DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 + DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 + DB 196,98,93,146,52,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm14 + DB 197,124,17,180,36,128,4,0,0 ; vmovups %ymm14,0x480(%rsp) + DB 196,193,109,254,211 ; vpaddd %ymm11,%ymm2,%ymm2 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,44,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm5 - DB 197,126,111,140,36,192,0,0,0 ; vmovdqu 0xc0(%rsp),%ymm9 - DB 197,181,254,201 ; vpaddd %ymm1,%ymm9,%ymm1 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,96,4,0,0 ; vmovups %ymm4,0x460(%rsp) + DB 197,213,254,201 ; vpaddd %ymm1,%ymm5,%ymm1 DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 - DB 196,193,117,254,211 ; vpaddd %ymm11,%ymm1,%ymm2 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,4,144 ; vgatherdps %ymm6,(%rax,%ymm2,4),%ymm0 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 + DB 197,252,17,156,36,64,4,0,0 ; vmovups %ymm3,0x440(%rsp) + DB 196,193,117,250,213 ; vpsubd %ymm13,%ymm1,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,32,4,0,0 ; vmovups %ymm4,0x420(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,226,109,146,52,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm6 - DB 197,228,92,204 ; vsubps %ymm4,%ymm3,%ymm1 - DB 196,226,13,168,204 ; vfmadd213ps %ymm4,%ymm14,%ymm1 - DB 197,252,17,76,36,96 ; vmovups %ymm1,0x60(%rsp) - DB 196,193,124,92,197 ; vsubps %ymm13,%ymm0,%ymm0 - DB 196,194,13,168,197 ; vfmadd213ps %ymm13,%ymm14,%ymm0 - DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp) - DB 197,204,92,197 ; vsubps %ymm5,%ymm6,%ymm0 - DB 196,226,13,168,197 ; vfmadd213ps %ymm5,%ymm14,%ymm0 - DB 197,252,17,132,36,224,1,0,0 ; vmovups %ymm0,0x1e0(%rsp) - DB 196,193,5,254,192 ; vpaddd %ymm8,%ymm15,%ymm0 - DB 197,253,254,207 ; vpaddd %ymm7,%ymm0,%ymm1 - DB 197,125,111,255 ; vmovdqa %ymm7,%ymm15 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 + DB 197,252,17,156,36,0,4,0,0 ; vmovups %ymm3,0x400(%rsp) + DB 197,205,254,192 ; vpaddd %ymm0,%ymm6,%ymm0 + DB 196,193,125,254,201 ; vpaddd %ymm9,%ymm0,%ymm1 DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 - DB 196,193,117,254,211 ; vpaddd %ymm11,%ymm1,%ymm2 - DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,226,85,146,52,144 ; vgatherdps %ymm5,(%rax,%ymm2,4),%ymm6 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 + DB 197,252,17,156,36,160,3,0,0 ; vmovups %ymm3,0x3a0(%rsp) + DB 196,193,117,250,213 ; vpsubd %ymm13,%ymm1,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,164,36,192,3,0,0 ; vmovups %ymm4,0x3c0(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 - DB 196,226,109,146,60,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm7 - DB 197,181,254,192 ; vpaddd %ymm0,%ymm9,%ymm0 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 + DB 197,252,17,156,36,0,3,0,0 ; vmovups %ymm3,0x300(%rsp) + DB 197,213,254,192 ; vpaddd %ymm0,%ymm5,%ymm0 DB 196,194,125,64,194 ; vpmulld %ymm10,%ymm0,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 DB 196,226,117,146,20,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm2 - DB 196,193,125,254,203 ; vpaddd %ymm11,%ymm0,%ymm1 - DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 - DB 196,98,85,146,44,136 ; vgatherdps %ymm5,(%rax,%ymm1,4),%ymm13 - DB 196,193,125,254,196 ; vpaddd %ymm12,%ymm0,%ymm0 + DB 197,252,17,148,36,96,3,0,0 ; vmovups %ymm2,0x360(%rsp) + DB 196,193,125,250,205 ; vpsubd %ymm13,%ymm0,%ymm1 + DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 + DB 197,252,17,156,36,64,3,0,0 ; vmovups %ymm3,0x340(%rsp) + DB 196,193,125,254,195 ; vpaddd %ymm11,%ymm0,%ymm0 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,36,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm4 - DB 197,236,92,195 ; vsubps %ymm3,%ymm2,%ymm0 - DB 196,226,13,168,195 ; vfmadd213ps %ymm3,%ymm14,%ymm0 - DB 197,252,17,132,36,192,1,0,0 ; vmovups %ymm0,0x1c0(%rsp) - DB 197,148,92,198 ; vsubps %ymm6,%ymm13,%ymm0 - DB 196,226,13,168,198 ; vfmadd213ps %ymm6,%ymm14,%ymm0 - DB 197,252,17,132,36,64,1,0,0 ; vmovups %ymm0,0x140(%rsp) - DB 197,220,92,199 ; vsubps %ymm7,%ymm4,%ymm0 - DB 196,226,13,168,199 ; vfmadd213ps %ymm7,%ymm14,%ymm0 - DB 197,252,17,132,36,96,1,0,0 ; vmovups %ymm0,0x160(%rsp) - DB 197,254,111,132,36,160,1,0,0 ; vmovdqu 0x1a0(%rsp),%ymm0 - DB 197,253,254,172,36,0,2,0,0 ; vpaddd 0x200(%rsp),%ymm0,%ymm5 - DB 197,213,254,4,36 ; vpaddd (%rsp),%ymm5,%ymm0 - DB 196,193,125,254,207 ; vpaddd %ymm15,%ymm0,%ymm1 + DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 + DB 196,226,117,146,20,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm2 + DB 197,252,17,148,36,32,3,0,0 ; vmovups %ymm2,0x320(%rsp) + DB 197,133,254,199 ; vpaddd %ymm7,%ymm15,%ymm0 + DB 196,193,125,254,200 ; vpaddd %ymm8,%ymm0,%ymm1 + DB 196,193,117,254,209 ; vpaddd %ymm9,%ymm1,%ymm2 + DB 196,194,109,64,210 ; vpmulld %ymm10,%ymm2,%ymm2 + DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 + DB 197,252,17,36,36 ; vmovups %ymm4,(%rsp) + DB 196,193,109,250,221 ; vpsubd %ymm13,%ymm2,%ymm3 + DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,226,61,146,36,152 ; vgatherdps %ymm8,(%rax,%ymm3,4),%ymm4 + DB 197,252,17,164,36,224,2,0,0 ; vmovups %ymm4,0x2e0(%rsp) + DB 196,193,109,254,211 ; vpaddd %ymm11,%ymm2,%ymm2 + DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,61,146,28,144 ; vgatherdps %ymm8,(%rax,%ymm2,4),%ymm3 + DB 197,252,17,156,36,96,2,0,0 ; vmovups %ymm3,0x260(%rsp) + DB 197,213,254,201 ; vpaddd %ymm1,%ymm5,%ymm1 DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,60,136 ; vgatherdps %ymm6,(%rax,%ymm1,4),%ymm7 - DB 196,193,117,254,243 ; vpaddd %ymm11,%ymm1,%ymm6 + DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 + DB 197,252,17,156,36,192,2,0,0 ; vmovups %ymm3,0x2c0(%rsp) + DB 196,193,117,250,213 ; vpsubd %ymm13,%ymm1,%ymm2 DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,98,61,146,44,176 ; vgatherdps %ymm8,(%rax,%ymm6,4),%ymm13 - DB 196,193,117,254,204 ; vpaddd %ymm12,%ymm1,%ymm1 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,20,136 ; vgatherdps %ymm6,(%rax,%ymm1,4),%ymm2 - DB 197,181,254,192 ; vpaddd %ymm0,%ymm9,%ymm0 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,61,146,28,144 ; vgatherdps %ymm8,(%rax,%ymm2,4),%ymm3 + DB 197,252,17,156,36,160,2,0,0 ; vmovups %ymm3,0x2a0(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 + DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 + DB 196,226,109,146,28,136 ; vgatherdps %ymm2,(%rax,%ymm1,4),%ymm3 + DB 197,252,17,156,36,128,2,0,0 ; vmovups %ymm3,0x280(%rsp) + DB 197,205,254,216 ; vpaddd %ymm0,%ymm6,%ymm3 + DB 196,193,101,254,193 ; vpaddd %ymm9,%ymm3,%ymm0 DB 196,194,125,64,194 ; vpmulld %ymm10,%ymm0,%ymm0 + DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 + DB 196,226,109,146,12,128 ; vgatherdps %ymm2,(%rax,%ymm0,4),%ymm1 + DB 197,252,17,140,36,64,2,0,0 ; vmovups %ymm1,0x240(%rsp) + DB 196,193,125,250,229 ; vpsubd %ymm13,%ymm0,%ymm4 DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,52,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm6 + DB 196,65,29,239,228 ; vpxor %ymm12,%ymm12,%ymm12 + DB 196,98,117,146,36,160 ; vgatherdps %ymm1,(%rax,%ymm4,4),%ymm12 DB 196,193,125,254,203 ; vpaddd %ymm11,%ymm0,%ymm1 - DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 - DB 196,226,61,146,28,136 ; vgatherdps %ymm8,(%rax,%ymm1,4),%ymm3 - DB 196,193,125,254,196 ; vpaddd %ymm12,%ymm0,%ymm0 - DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 - DB 196,226,117,146,36,128 ; vgatherdps %ymm1,(%rax,%ymm0,4),%ymm4 - DB 197,204,92,199 ; vsubps %ymm7,%ymm6,%ymm0 - DB 196,226,13,168,199 ; vfmadd213ps %ymm7,%ymm14,%ymm0 - DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) - DB 196,193,100,92,205 ; vsubps %ymm13,%ymm3,%ymm1 - DB 196,194,13,168,205 ; vfmadd213ps %ymm13,%ymm14,%ymm1 - DB 197,92,92,194 ; vsubps %ymm2,%ymm4,%ymm8 - DB 196,98,13,168,194 ; vfmadd213ps %ymm2,%ymm14,%ymm8 - DB 197,213,254,148,36,128,1,0,0 ; vpaddd 0x180(%rsp),%ymm5,%ymm2 - DB 196,193,109,254,223 ; vpaddd %ymm15,%ymm2,%ymm3 - DB 196,194,101,64,218 ; vpmulld %ymm10,%ymm3,%ymm3 - DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 - DB 196,226,93,146,44,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm5 - DB 196,193,101,254,227 ; vpaddd %ymm11,%ymm3,%ymm4 - DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 - DB 196,226,77,146,60,160 ; vgatherdps %ymm6,(%rax,%ymm4,4),%ymm7 - DB 196,193,101,254,220 ; vpaddd %ymm12,%ymm3,%ymm3 DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 - DB 196,226,93,146,52,152 ; vgatherdps %ymm4,(%rax,%ymm3,4),%ymm6 - DB 197,181,254,210 ; vpaddd %ymm2,%ymm9,%ymm2 - DB 196,194,109,64,210 ; vpmulld %ymm10,%ymm2,%ymm2 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,226,93,146,4,136 ; vgatherdps %ymm4,(%rax,%ymm1,4),%ymm0 + DB 197,252,17,132,36,32,2,0,0 ; vmovups %ymm0,0x220(%rsp) + DB 197,213,254,203 ; vpaddd %ymm3,%ymm5,%ymm1 + DB 196,194,117,64,202 ; vpmulld %ymm10,%ymm1,%ymm1 DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,226,101,146,36,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm4 - DB 196,193,109,254,219 ; vpaddd %ymm11,%ymm2,%ymm3 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,226,101,146,4,136 ; vgatherdps %ymm3,(%rax,%ymm1,4),%ymm0 + DB 197,252,17,132,36,0,2,0,0 ; vmovups %ymm0,0x200(%rsp) + DB 196,193,117,250,221 ; vpsubd %ymm13,%ymm1,%ymm3 DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 - DB 196,98,45,146,28,152 ; vgatherdps %ymm10,(%rax,%ymm3,4),%ymm11 - DB 196,193,109,254,212 ; vpaddd %ymm12,%ymm2,%ymm2 - DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3 - DB 196,98,101,146,20,144 ; vgatherdps %ymm3,(%rax,%ymm2,4),%ymm10 - DB 197,220,92,213 ; vsubps %ymm5,%ymm4,%ymm2 - DB 196,226,13,168,213 ; vfmadd213ps %ymm5,%ymm14,%ymm2 - DB 197,164,92,223 ; vsubps %ymm7,%ymm11,%ymm3 - DB 196,226,13,168,223 ; vfmadd213ps %ymm7,%ymm14,%ymm3 - DB 197,172,92,230 ; vsubps %ymm6,%ymm10,%ymm4 - DB 196,226,13,168,230 ; vfmadd213ps %ymm6,%ymm14,%ymm4 - DB 197,252,91,172,36,96,3,0,0 ; vcvtdq2ps 0x360(%rsp),%ymm5 - DB 197,252,16,180,36,224,2,0,0 ; vmovups 0x2e0(%rsp),%ymm6 - DB 197,204,92,237 ; vsubps %ymm5,%ymm6,%ymm5 - DB 197,252,16,188,36,64,3,0,0 ; vmovups 0x340(%rsp),%ymm7 - DB 197,252,16,180,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm6 - DB 197,204,92,247 ; vsubps %ymm7,%ymm6,%ymm6 - DB 196,226,85,168,247 ; vfmadd213ps %ymm7,%ymm5,%ymm6 - DB 197,124,16,148,36,32,3,0,0 ; vmovups 0x320(%rsp),%ymm10 - DB 197,252,16,188,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm7 - DB 196,193,68,92,250 ; vsubps %ymm10,%ymm7,%ymm7 - DB 196,194,85,168,250 ; vfmadd213ps %ymm10,%ymm5,%ymm7 - DB 197,124,16,156,36,0,3,0,0 ; vmovups 0x300(%rsp),%ymm11 - DB 197,124,16,148,36,192,2,0,0 ; vmovups 0x2c0(%rsp),%ymm10 - DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 - DB 196,66,85,168,211 ; vfmadd213ps %ymm11,%ymm5,%ymm10 - DB 197,124,16,100,36,64 ; vmovups 0x40(%rsp),%ymm12 - DB 197,124,16,156,36,96,2,0,0 ; vmovups 0x260(%rsp),%ymm11 - DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,66,85,168,220 ; vfmadd213ps %ymm12,%ymm5,%ymm11 - DB 197,124,16,172,36,160,2,0,0 ; vmovups 0x2a0(%rsp),%ymm13 - DB 197,124,16,164,36,64,2,0,0 ; vmovups 0x240(%rsp),%ymm12 - DB 196,65,28,92,229 ; vsubps %ymm13,%ymm12,%ymm12 - DB 196,66,85,168,229 ; vfmadd213ps %ymm13,%ymm5,%ymm12 - DB 197,124,16,180,36,128,2,0,0 ; vmovups 0x280(%rsp),%ymm14 - DB 197,124,16,172,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm13 - DB 196,65,20,92,238 ; vsubps %ymm14,%ymm13,%ymm13 - DB 196,66,85,168,238 ; vfmadd213ps %ymm14,%ymm5,%ymm13 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,226,45,146,4,152 ; vgatherdps %ymm10,(%rax,%ymm3,4),%ymm0 + DB 197,252,17,132,36,224,1,0,0 ; vmovups %ymm0,0x1e0(%rsp) + DB 196,193,117,254,203 ; vpaddd %ymm11,%ymm1,%ymm1 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,226,21,146,4,136 ; vgatherdps %ymm13,(%rax,%ymm1,4),%ymm0 + DB 197,252,17,132,36,192,1,0,0 ; vmovups %ymm0,0x1c0(%rsp) + DB 197,252,91,140,36,128,1,0,0 ; vcvtdq2ps 0x180(%rsp),%ymm1 + DB 197,252,16,132,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm0 + DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 + DB 197,252,16,132,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm0 + DB 197,252,16,148,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm2 + DB 197,108,92,248 ; vsubps %ymm0,%ymm2,%ymm15 + DB 196,98,117,168,248 ; vfmadd213ps %ymm0,%ymm1,%ymm15 + DB 197,252,16,132,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm0 + DB 197,252,16,148,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm2 + DB 197,108,92,240 ; vsubps %ymm0,%ymm2,%ymm14 + DB 196,98,117,168,240 ; vfmadd213ps %ymm0,%ymm1,%ymm14 + DB 197,252,16,172,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm5 + DB 197,252,16,132,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm0 + DB 197,252,92,197 ; vsubps %ymm5,%ymm0,%ymm0 + DB 196,226,117,168,197 ; vfmadd213ps %ymm5,%ymm1,%ymm0 + DB 197,252,17,132,36,128,1,0,0 ; vmovups %ymm0,0x180(%rsp) + DB 197,252,16,68,36,64 ; vmovups 0x40(%rsp),%ymm0 + DB 197,252,16,148,36,224,5,0,0 ; vmovups 0x5e0(%rsp),%ymm2 + DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0 + DB 196,226,117,168,194 ; vfmadd213ps %ymm2,%ymm1,%ymm0 + DB 197,252,17,132,36,96,1,0,0 ; vmovups %ymm0,0x160(%rsp) + DB 197,252,16,132,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm0 + DB 197,252,16,148,36,32,6,0,0 ; vmovups 0x620(%rsp),%ymm2 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,17,148,36,64,1,0,0 ; vmovups %ymm2,0x140(%rsp) DB 197,252,16,68,36,96 ; vmovups 0x60(%rsp),%ymm0 - DB 197,124,16,180,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm14 - DB 197,12,92,240 ; vsubps %ymm0,%ymm14,%ymm14 - DB 196,98,85,168,240 ; vfmadd213ps %ymm0,%ymm5,%ymm14 + DB 197,252,16,148,36,0,6,0,0 ; vmovups 0x600(%rsp),%ymm2 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,17,148,36,32,1,0,0 ; vmovups %ymm2,0x120(%rsp) + DB 197,252,16,132,36,160,5,0,0 ; vmovups 0x5a0(%rsp),%ymm0 + DB 197,252,16,148,36,224,3,0,0 ; vmovups 0x3e0(%rsp),%ymm2 + DB 197,124,92,210 ; vsubps %ymm2,%ymm0,%ymm10 + DB 196,98,117,168,210 ; vfmadd213ps %ymm2,%ymm1,%ymm10 + DB 197,252,16,132,36,192,5,0,0 ; vmovups 0x5c0(%rsp),%ymm0 + DB 197,252,16,148,36,128,5,0,0 ; vmovups 0x580(%rsp),%ymm2 + DB 197,108,92,216 ; vsubps %ymm0,%ymm2,%ymm11 + DB 196,98,117,168,216 ; vfmadd213ps %ymm0,%ymm1,%ymm11 + DB 197,252,16,132,36,96,5,0,0 ; vmovups 0x560(%rsp),%ymm0 + DB 197,252,16,148,36,64,5,0,0 ; vmovups 0x540(%rsp),%ymm2 + DB 197,124,92,202 ; vsubps %ymm2,%ymm0,%ymm9 + DB 196,98,117,168,202 ; vfmadd213ps %ymm2,%ymm1,%ymm9 + DB 197,252,16,132,36,32,5,0,0 ; vmovups 0x520(%rsp),%ymm0 + DB 197,252,16,148,36,224,4,0,0 ; vmovups 0x4e0(%rsp),%ymm2 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,17,148,36,0,1,0,0 ; vmovups %ymm2,0x100(%rsp) + DB 197,252,16,132,36,0,5,0,0 ; vmovups 0x500(%rsp),%ymm0 + DB 197,252,16,148,36,192,4,0,0 ; vmovups 0x4c0(%rsp),%ymm2 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,17,148,36,224,0,0,0 ; vmovups %ymm2,0xe0(%rsp) + DB 197,252,16,132,36,160,4,0,0 ; vmovups 0x4a0(%rsp),%ymm0 + DB 197,252,16,148,36,128,3,0,0 ; vmovups 0x380(%rsp),%ymm2 + DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0 + DB 196,226,117,168,194 ; vfmadd213ps %ymm2,%ymm1,%ymm0 + DB 197,252,17,132,36,192,0,0,0 ; vmovups %ymm0,0xc0(%rsp) DB 197,252,16,68,36,32 ; vmovups 0x20(%rsp),%ymm0 - DB 197,124,16,140,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm9 - DB 197,52,92,248 ; vsubps %ymm0,%ymm9,%ymm15 - DB 196,98,85,168,248 ; vfmadd213ps %ymm0,%ymm5,%ymm15 - DB 197,252,16,132,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm0 - DB 197,124,16,140,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm9 - DB 197,52,92,200 ; vsubps %ymm0,%ymm9,%ymm9 - DB 196,98,85,168,200 ; vfmadd213ps %ymm0,%ymm5,%ymm9 + DB 197,252,16,148,36,64,4,0,0 ; vmovups 0x440(%rsp),%ymm2 + DB 197,236,92,248 ; vsubps %ymm0,%ymm2,%ymm7 + DB 196,226,117,168,248 ; vfmadd213ps %ymm0,%ymm1,%ymm7 + DB 197,252,16,132,36,128,4,0,0 ; vmovups 0x480(%rsp),%ymm0 + DB 197,252,16,148,36,32,4,0,0 ; vmovups 0x420(%rsp),%ymm2 + DB 197,108,92,192 ; vsubps %ymm0,%ymm2,%ymm8 + DB 196,98,117,168,192 ; vfmadd213ps %ymm0,%ymm1,%ymm8 + DB 197,252,16,132,36,96,4,0,0 ; vmovups 0x460(%rsp),%ymm0 + DB 197,252,16,148,36,0,4,0,0 ; vmovups 0x400(%rsp),%ymm2 + DB 197,236,92,240 ; vsubps %ymm0,%ymm2,%ymm6 + DB 196,226,117,168,240 ; vfmadd213ps %ymm0,%ymm1,%ymm6 + DB 197,252,16,132,36,160,3,0,0 ; vmovups 0x3a0(%rsp),%ymm0 + DB 197,252,16,148,36,96,3,0,0 ; vmovups 0x360(%rsp),%ymm2 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,17,148,36,160,0,0,0 ; vmovups %ymm2,0xa0(%rsp) + DB 197,252,16,132,36,192,3,0,0 ; vmovups 0x3c0(%rsp),%ymm0 + DB 197,252,16,148,36,64,3,0,0 ; vmovups 0x340(%rsp),%ymm2 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,17,148,36,128,0,0,0 ; vmovups %ymm2,0x80(%rsp) + DB 197,252,16,132,36,32,3,0,0 ; vmovups 0x320(%rsp),%ymm0 + DB 197,252,16,148,36,0,3,0,0 ; vmovups 0x300(%rsp),%ymm2 + DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0 + DB 196,226,117,168,194 ; vfmadd213ps %ymm2,%ymm1,%ymm0 + DB 197,252,17,68,36,96 ; vmovups %ymm0,0x60(%rsp) DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 + DB 197,252,16,148,36,192,2,0,0 ; vmovups 0x2c0(%rsp),%ymm2 DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 - DB 196,226,85,168,208 ; vfmadd213ps %ymm0,%ymm5,%ymm2 - DB 197,228,92,217 ; vsubps %ymm1,%ymm3,%ymm3 - DB 196,226,85,168,217 ; vfmadd213ps %ymm1,%ymm5,%ymm3 - DB 196,193,92,92,200 ; vsubps %ymm8,%ymm4,%ymm1 - DB 196,194,85,168,200 ; vfmadd213ps %ymm8,%ymm5,%ymm1 - DB 197,252,91,132,36,128,3,0,0 ; vcvtdq2ps 0x380(%rsp),%ymm0 - DB 197,252,16,164,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm4 - DB 197,220,92,192 ; vsubps %ymm0,%ymm4,%ymm0 - DB 197,164,92,230 ; vsubps %ymm6,%ymm11,%ymm4 - DB 196,226,125,168,230 ; vfmadd213ps %ymm6,%ymm0,%ymm4 - DB 197,156,92,239 ; vsubps %ymm7,%ymm12,%ymm5 - DB 196,226,125,168,239 ; vfmadd213ps %ymm7,%ymm0,%ymm5 - DB 196,193,20,92,242 ; vsubps %ymm10,%ymm13,%ymm6 - DB 196,194,125,168,242 ; vfmadd213ps %ymm10,%ymm0,%ymm6 - DB 196,193,108,92,214 ; vsubps %ymm14,%ymm2,%ymm2 - DB 196,194,125,168,214 ; vfmadd213ps %ymm14,%ymm0,%ymm2 - DB 196,193,100,92,223 ; vsubps %ymm15,%ymm3,%ymm3 - DB 196,194,125,168,223 ; vfmadd213ps %ymm15,%ymm0,%ymm3 - DB 196,193,116,92,249 ; vsubps %ymm9,%ymm1,%ymm7 - DB 196,194,125,168,249 ; vfmadd213ps %ymm9,%ymm0,%ymm7 - DB 197,252,91,132,36,0,1,0,0 ; vcvtdq2ps 0x100(%rsp),%ymm0 - DB 197,252,16,140,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm1 - DB 197,116,92,192 ; vsubps %ymm0,%ymm1,%ymm8 - DB 197,236,92,196 ; vsubps %ymm4,%ymm2,%ymm0 - DB 196,226,61,168,196 ; vfmadd213ps %ymm4,%ymm8,%ymm0 - DB 197,228,92,205 ; vsubps %ymm5,%ymm3,%ymm1 - DB 196,226,61,168,205 ; vfmadd213ps %ymm5,%ymm8,%ymm1 - DB 197,196,92,214 ; vsubps %ymm6,%ymm7,%ymm2 - DB 196,226,61,168,214 ; vfmadd213ps %ymm6,%ymm8,%ymm2 + DB 196,226,117,168,208 ; vfmadd213ps %ymm0,%ymm1,%ymm2 + DB 197,252,16,132,36,224,2,0,0 ; vmovups 0x2e0(%rsp),%ymm0 + DB 197,252,16,156,36,160,2,0,0 ; vmovups 0x2a0(%rsp),%ymm3 + DB 197,228,92,216 ; vsubps %ymm0,%ymm3,%ymm3 + DB 196,226,117,168,216 ; vfmadd213ps %ymm0,%ymm1,%ymm3 + DB 197,252,16,132,36,128,2,0,0 ; vmovups 0x280(%rsp),%ymm0 + DB 197,252,16,172,36,96,2,0,0 ; vmovups 0x260(%rsp),%ymm5 + DB 197,252,92,229 ; vsubps %ymm5,%ymm0,%ymm4 + DB 196,226,117,168,229 ; vfmadd213ps %ymm5,%ymm1,%ymm4 + DB 197,252,16,132,36,64,2,0,0 ; vmovups 0x240(%rsp),%ymm0 + DB 197,252,16,172,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm5 + DB 197,84,92,232 ; vsubps %ymm0,%ymm5,%ymm13 + DB 196,98,117,168,232 ; vfmadd213ps %ymm0,%ymm1,%ymm13 + DB 197,252,16,132,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm0 + DB 196,193,124,92,196 ; vsubps %ymm12,%ymm0,%ymm0 + DB 196,194,117,168,196 ; vfmadd213ps %ymm12,%ymm1,%ymm0 + DB 197,252,17,68,36,64 ; vmovups %ymm0,0x40(%rsp) + DB 197,252,16,132,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm0 + DB 197,252,16,172,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm5 + DB 197,84,92,224 ; vsubps %ymm0,%ymm5,%ymm12 + DB 196,98,117,168,224 ; vfmadd213ps %ymm0,%ymm1,%ymm12 + DB 197,252,91,140,36,160,6,0,0 ; vcvtdq2ps 0x6a0(%rsp),%ymm1 + DB 197,252,16,132,36,128,6,0,0 ; vmovups 0x680(%rsp),%ymm0 + DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 + DB 197,252,16,132,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm0 + DB 196,193,124,92,199 ; vsubps %ymm15,%ymm0,%ymm0 + DB 196,194,117,168,199 ; vfmadd213ps %ymm15,%ymm1,%ymm0 + DB 197,252,16,172,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm5 + DB 196,65,84,92,254 ; vsubps %ymm14,%ymm5,%ymm15 + DB 196,66,117,168,254 ; vfmadd213ps %ymm14,%ymm1,%ymm15 + DB 197,252,16,172,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm5 + DB 197,124,16,180,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm14 + DB 197,12,92,245 ; vsubps %ymm5,%ymm14,%ymm14 + DB 196,98,117,168,245 ; vfmadd213ps %ymm5,%ymm1,%ymm14 + DB 197,252,16,172,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm5 + DB 196,193,84,92,234 ; vsubps %ymm10,%ymm5,%ymm5 + DB 196,194,117,168,234 ; vfmadd213ps %ymm10,%ymm1,%ymm5 + DB 197,124,16,148,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm10 + DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 + DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10 + DB 197,124,16,156,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm11 + DB 196,65,36,92,217 ; vsubps %ymm9,%ymm11,%ymm11 + DB 196,66,117,168,217 ; vfmadd213ps %ymm9,%ymm1,%ymm11 + DB 197,124,16,140,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm9 + DB 197,52,92,207 ; vsubps %ymm7,%ymm9,%ymm9 + DB 196,98,117,168,207 ; vfmadd213ps %ymm7,%ymm1,%ymm9 + DB 197,252,16,188,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm7 + DB 196,193,68,92,248 ; vsubps %ymm8,%ymm7,%ymm7 + DB 196,194,117,168,248 ; vfmadd213ps %ymm8,%ymm1,%ymm7 + DB 197,124,16,68,36,96 ; vmovups 0x60(%rsp),%ymm8 + DB 197,60,92,198 ; vsubps %ymm6,%ymm8,%ymm8 + DB 196,98,117,168,198 ; vfmadd213ps %ymm6,%ymm1,%ymm8 + DB 197,148,92,242 ; vsubps %ymm2,%ymm13,%ymm6 + DB 196,226,117,168,242 ; vfmadd213ps %ymm2,%ymm1,%ymm6 + DB 197,252,16,84,36,64 ; vmovups 0x40(%rsp),%ymm2 + DB 197,236,92,211 ; vsubps %ymm3,%ymm2,%ymm2 + DB 196,226,117,168,211 ; vfmadd213ps %ymm3,%ymm1,%ymm2 + DB 197,156,92,220 ; vsubps %ymm4,%ymm12,%ymm3 + DB 196,226,117,168,220 ; vfmadd213ps %ymm4,%ymm1,%ymm3 + DB 197,252,91,140,36,192,6,0,0 ; vcvtdq2ps 0x6c0(%rsp),%ymm1 + DB 197,252,16,164,36,96,6,0,0 ; vmovups 0x660(%rsp),%ymm4 + DB 197,220,92,201 ; vsubps %ymm1,%ymm4,%ymm1 + DB 197,212,92,224 ; vsubps %ymm0,%ymm5,%ymm4 + DB 196,226,117,168,224 ; vfmadd213ps %ymm0,%ymm1,%ymm4 + DB 196,193,44,92,239 ; vsubps %ymm15,%ymm10,%ymm5 + DB 196,194,117,168,239 ; vfmadd213ps %ymm15,%ymm1,%ymm5 + DB 196,65,36,92,214 ; vsubps %ymm14,%ymm11,%ymm10 + DB 196,66,117,168,214 ; vfmadd213ps %ymm14,%ymm1,%ymm10 + DB 196,193,76,92,193 ; vsubps %ymm9,%ymm6,%ymm0 + DB 196,194,117,168,193 ; vfmadd213ps %ymm9,%ymm1,%ymm0 + DB 197,236,92,215 ; vsubps %ymm7,%ymm2,%ymm2 + DB 196,226,117,168,215 ; vfmadd213ps %ymm7,%ymm1,%ymm2 + DB 196,193,100,92,216 ; vsubps %ymm8,%ymm3,%ymm3 + DB 196,194,117,168,216 ; vfmadd213ps %ymm8,%ymm1,%ymm3 + DB 197,252,91,140,36,64,6,0,0 ; vcvtdq2ps 0x640(%rsp),%ymm1 + DB 197,252,16,180,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm6 + DB 197,204,92,241 ; vsubps %ymm1,%ymm6,%ymm6 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 196,226,77,168,196 ; vfmadd213ps %ymm4,%ymm6,%ymm0 + DB 197,236,92,205 ; vsubps %ymm5,%ymm2,%ymm1 + DB 196,226,77,168,205 ; vfmadd213ps %ymm5,%ymm6,%ymm1 + DB 196,193,100,92,210 ; vsubps %ymm10,%ymm3,%ymm2 + DB 196,194,77,168,210 ; vfmadd213ps %ymm10,%ymm6,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,236,96,2,0 ; vbroadcastss 0x260ec(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> - DB 197,252,16,164,36,160,3,0,0 ; vmovups 0x3a0(%rsp),%ymm4 - DB 197,252,16,172,36,192,3,0,0 ; vmovups 0x3c0(%rsp),%ymm5 - DB 197,252,16,180,36,224,3,0,0 ; vmovups 0x3e0(%rsp),%ymm6 - DB 197,252,16,188,36,0,4,0,0 ; vmovups 0x400(%rsp),%ymm7 - DB 72,129,196,56,4,0,0 ; add $0x438,%rsp + DB 196,226,125,24,29,218,90,2,0 ; vbroadcastss 0x25ada(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> + DB 197,252,16,164,36,224,6,0,0 ; vmovups 0x6e0(%rsp),%ymm4 + DB 197,252,16,172,36,0,7,0,0 ; vmovups 0x700(%rsp),%ymm5 + DB 197,252,16,180,36,32,7,0,0 ; vmovups 0x720(%rsp),%ymm6 + DB 197,252,16,188,36,64,7,0,0 ; vmovups 0x740(%rsp),%ymm7 + DB 72,129,196,120,7,0,0 ; add $0x778,%rsp DB 255,224 ; jmpq *%rax DB 144 ; nop - DB 144 ; nop - DB 144 ; nop PUBLIC _sk_start_pipeline_avx _sk_start_pipeline_avx LABEL PROC @@ -6822,13 +7033,13 @@ _sk_start_pipeline_avx LABEL PROC DB 73,137,246 ; mov %rsi,%r14 DB 76,137,141,8,255,255,255 ; mov %r9,-0xf8(%rbp) DB 76,57,203 ; cmp %r9,%rbx - DB 115,101 ; jae 72d4 <_sk_start_pipeline_avx+0xe4> + DB 115,101 ; jae 77c4 <_sk_start_pipeline_avx+0xe4> DB 72,139,133,24,255,255,255 ; mov -0xe8(%rbp),%rax DB 72,141,64,8 ; lea 0x8(%rax),%rax DB 72,137,133,16,255,255,255 ; mov %rax,-0xf0(%rbp) DB 72,139,149,24,255,255,255 ; mov -0xe8(%rbp),%rdx DB 76,57,189,16,255,255,255 ; cmp %r15,-0xf0(%rbp) - DB 119,38 ; ja 72b7 <_sk_start_pipeline_avx+0xc7> + DB 119,38 ; ja 77a7 <_sk_start_pipeline_avx+0xc7> DB 76,139,165,24,255,255,255 ; mov -0xe8(%rbp),%r12 DB 49,255 ; xor %edi,%edi DB 76,137,246 ; mov %r14,%rsi @@ -6839,16 +7050,16 @@ _sk_start_pipeline_avx LABEL PROC DB 73,131,196,16 ; add $0x10,%r12 DB 77,57,252 ; cmp %r15,%r12 DB 73,137,212 ; mov %rdx,%r12 - DB 118,225 ; jbe 7298 <_sk_start_pipeline_avx+0xa8> + DB 118,225 ; jbe 7788 <_sk_start_pipeline_avx+0xa8> DB 76,137,255 ; mov %r15,%rdi DB 72,41,215 ; sub %rdx,%rdi - DB 116,9 ; je 72c8 <_sk_start_pipeline_avx+0xd8> + DB 116,9 ; je 77b8 <_sk_start_pipeline_avx+0xd8> DB 76,137,246 ; mov %r14,%rsi DB 72,137,217 ; mov %rbx,%rcx DB 65,255,213 ; callq *%r13 DB 72,255,195 ; inc %rbx DB 72,59,157,8,255,255,255 ; cmp -0xf8(%rbp),%rbx - DB 117,173 ; jne 7281 <_sk_start_pipeline_avx+0x91> + DB 117,173 ; jne 7771 <_sk_start_pipeline_avx+0x91> DB 197,248,40,181,32,255,255,255 ; vmovaps -0xe0(%rbp),%xmm6 DB 197,248,40,189,48,255,255,255 ; vmovaps -0xd0(%rbp),%xmm7 DB 197,120,40,133,64,255,255,255 ; vmovaps -0xc0(%rbp),%xmm8 @@ -6881,15 +7092,15 @@ _sk_seed_shader_avx LABEL PROC DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 197,252,88,5,213,89,2,0 ; vaddps 0x259d5(%rip),%ymm0,%ymm0 # 2cd20 <_sk_overlay_sse2_8bit+0x7c8> + DB 197,252,88,5,229,83,2,0 ; vaddps 0x253e5(%rip),%ymm0,%ymm0 # 2cc20 <_sk_overlay_sse2_8bit+0x771> DB 197,249,110,201 ; vmovd %ecx,%xmm1 DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,65,95,2,0 ; vbroadcastss 0x25f41(%rip),%ymm2 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,21,49,89,2,0 ; vbroadcastss 0x25931(%rip),%ymm2 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,54,95,2,0 ; vbroadcastss 0x25f36(%rip),%ymm2 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,21,38,89,2,0 ; vbroadcastss 0x25926(%rip),%ymm2 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 @@ -6899,58 +7110,61 @@ _sk_seed_shader_avx LABEL PROC PUBLIC _sk_dither_avx _sk_dither_avx LABEL PROC - DB 72,173 ; lods %ds:(%rsi),%rax DB 197,121,110,194 ; vmovd %edx,%xmm8 DB 196,65,121,112,192,0 ; vpshufd $0x0,%xmm8,%xmm8 - DB 197,57,254,13,224,96,2,0 ; vpaddd 0x260e0(%rip),%xmm8,%xmm9 # 2d480 <_sk_overlay_sse2_8bit+0xf28> - DB 197,57,254,5,232,96,2,0 ; vpaddd 0x260e8(%rip),%xmm8,%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xf38> - DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 - DB 197,121,110,201 ; vmovd %ecx,%xmm9 - DB 196,65,121,112,201,0 ; vpshufd $0x0,%xmm9,%xmm9 - DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 - DB 196,65,52,87,208 ; vxorps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,228,94,2,0 ; vbroadcastss 0x25ee4(%rip),%ymm11 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> - DB 196,65,44,84,203 ; vandps %ymm11,%ymm10,%ymm9 - DB 196,193,25,114,241,5 ; vpslld $0x5,%xmm9,%xmm12 - DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 - DB 196,193,49,114,241,5 ; vpslld $0x5,%xmm9,%xmm9 - DB 196,67,29,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm12,%ymm9 - DB 196,65,60,84,219 ; vandps %ymm11,%ymm8,%ymm11 - DB 196,193,25,114,243,4 ; vpslld $0x4,%xmm11,%xmm12 - DB 196,67,125,25,219,1 ; vextractf128 $0x1,%ymm11,%xmm11 - DB 196,193,33,114,243,4 ; vpslld $0x4,%xmm11,%xmm11 - DB 196,67,29,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm12,%ymm11 - DB 196,98,125,24,37,165,94,2,0 ; vbroadcastss 0x25ea5(%rip),%ymm12 # 2d2b4 <_sk_overlay_sse2_8bit+0xd5c> - DB 196,98,125,24,45,160,94,2,0 ; vbroadcastss 0x25ea0(%rip),%ymm13 # 2d2b8 <_sk_overlay_sse2_8bit+0xd60> - DB 196,65,44,84,245 ; vandps %ymm13,%ymm10,%ymm14 - DB 196,193,1,114,246,2 ; vpslld $0x2,%xmm14,%xmm15 - DB 196,67,125,25,246,1 ; vextractf128 $0x1,%ymm14,%xmm14 - DB 196,193,9,114,246,2 ; vpslld $0x2,%xmm14,%xmm14 - DB 196,67,5,24,246,1 ; vinsertf128 $0x1,%xmm14,%ymm15,%ymm14 - DB 196,65,60,84,237 ; vandps %ymm13,%ymm8,%ymm13 - DB 196,65,17,254,253 ; vpaddd %xmm13,%xmm13,%xmm15 - DB 196,67,125,25,237,1 ; vextractf128 $0x1,%ymm13,%xmm13 - DB 196,65,17,254,237 ; vpaddd %xmm13,%xmm13,%xmm13 - DB 196,67,5,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm15,%ymm13 - DB 196,65,44,84,212 ; vandps %ymm12,%ymm10,%ymm10 - DB 196,193,1,114,210,1 ; vpsrld $0x1,%xmm10,%xmm15 + DB 197,57,254,13,226,90,2,0 ; vpaddd 0x25ae2(%rip),%xmm8,%xmm9 # 2d370 <_sk_overlay_sse2_8bit+0xec1> + DB 197,57,254,5,234,90,2,0 ; vpaddd 0x25aea(%rip),%xmm8,%xmm8 # 2d380 <_sk_overlay_sse2_8bit+0xed1> + DB 196,67,53,24,208,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm10 + DB 197,121,110,217 ; vmovd %ecx,%xmm11 + DB 196,65,121,112,219,0 ; vpshufd $0x0,%xmm11,%xmm11 + DB 196,67,37,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 + DB 196,65,36,87,218 ; vxorps %ymm10,%ymm11,%ymm11 + DB 196,98,125,24,21,214,88,2,0 ; vbroadcastss 0x258d6(%rip),%ymm10 # 2d190 <_sk_overlay_sse2_8bit+0xce1> + DB 196,65,36,84,210 ; vandps %ymm10,%ymm11,%ymm10 + DB 196,193,25,114,242,5 ; vpslld $0x5,%xmm10,%xmm12 DB 196,67,125,25,210,1 ; vextractf128 $0x1,%ymm10,%xmm10 - DB 196,193,41,114,210,1 ; vpsrld $0x1,%xmm10,%xmm10 - DB 196,67,5,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm15,%ymm10 - DB 196,65,60,84,196 ; vandps %ymm12,%ymm8,%ymm8 - DB 196,193,25,114,208,2 ; vpsrld $0x2,%xmm8,%xmm12 - DB 196,67,125,25,192,1 ; vextractf128 $0x1,%ymm8,%xmm8 - DB 196,193,57,114,208,2 ; vpsrld $0x2,%xmm8,%xmm8 - DB 196,67,29,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm12,%ymm8 + DB 196,193,41,114,242,5 ; vpslld $0x5,%xmm10,%xmm10 + DB 196,67,29,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm12,%ymm10 + DB 197,123,18,37,209,98,2,0 ; vmovddup 0x262d1(%rip),%xmm12 # 2dbb0 <_sk_overlay_sse2_8bit+0x1701> + DB 196,65,49,219,236 ; vpand %xmm12,%xmm9,%xmm13 + DB 196,193,17,114,245,4 ; vpslld $0x4,%xmm13,%xmm13 + DB 196,65,57,219,228 ; vpand %xmm12,%xmm8,%xmm12 + DB 196,193,25,114,244,4 ; vpslld $0x4,%xmm12,%xmm12 + DB 196,67,21,24,228,1 ; vinsertf128 $0x1,%xmm12,%ymm13,%ymm12 + DB 196,98,125,24,45,148,88,2,0 ; vbroadcastss 0x25894(%rip),%ymm13 # 2d198 <_sk_overlay_sse2_8bit+0xce9> + DB 196,65,36,84,237 ; vandps %ymm13,%ymm11,%ymm13 + DB 196,193,9,114,245,2 ; vpslld $0x2,%xmm13,%xmm14 + DB 196,67,125,25,237,1 ; vextractf128 $0x1,%ymm13,%xmm13 + DB 196,193,17,114,245,2 ; vpslld $0x2,%xmm13,%xmm13 + DB 196,67,13,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 + DB 197,123,18,53,143,98,2,0 ; vmovddup 0x2628f(%rip),%xmm14 # 2dbb8 <_sk_overlay_sse2_8bit+0x1709> + DB 196,65,49,219,254 ; vpand %xmm14,%xmm9,%xmm15 + DB 196,65,1,254,255 ; vpaddd %xmm15,%xmm15,%xmm15 + DB 196,65,57,219,246 ; vpand %xmm14,%xmm8,%xmm14 + DB 196,65,9,254,246 ; vpaddd %xmm14,%xmm14,%xmm14 + DB 196,67,5,24,246,1 ; vinsertf128 $0x1,%xmm14,%ymm15,%ymm14 + DB 196,65,12,86,228 ; vorps %ymm12,%ymm14,%ymm12 + DB 196,98,125,24,53,67,88,2,0 ; vbroadcastss 0x25843(%rip),%ymm14 # 2d194 <_sk_overlay_sse2_8bit+0xce5> + DB 196,65,36,84,222 ; vandps %ymm14,%ymm11,%ymm11 + DB 196,193,9,114,211,1 ; vpsrld $0x1,%xmm11,%xmm14 + DB 196,67,125,25,219,1 ; vextractf128 $0x1,%ymm11,%xmm11 + DB 196,193,33,114,211,1 ; vpsrld $0x1,%xmm11,%xmm11 + DB 196,67,13,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm14,%ymm11 DB 196,65,20,86,219 ; vorps %ymm11,%ymm13,%ymm11 - DB 196,65,36,86,192 ; vorps %ymm8,%ymm11,%ymm8 - DB 196,65,52,86,206 ; vorps %ymm14,%ymm9,%ymm9 - DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8 + DB 197,123,18,45,69,98,2,0 ; vmovddup 0x26245(%rip),%xmm13 # 2dbc0 <_sk_overlay_sse2_8bit+0x1711> + DB 196,65,49,219,205 ; vpand %xmm13,%xmm9,%xmm9 + DB 196,65,57,219,197 ; vpand %xmm13,%xmm8,%xmm8 + DB 196,193,49,114,209,2 ; vpsrld $0x2,%xmm9,%xmm9 + DB 196,193,57,114,208,2 ; vpsrld $0x2,%xmm8,%xmm8 + DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 + DB 196,65,28,86,192 ; vorps %ymm8,%ymm12,%ymm8 DB 196,65,60,86,194 ; vorps %ymm10,%ymm8,%ymm8 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 196,65,60,86,195 ; vorps %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,11,94,2,0 ; vbroadcastss 0x25e0b(%rip),%ymm9 # 2d2bc <_sk_overlay_sse2_8bit+0xd64> + DB 196,98,125,24,13,230,87,2,0 ; vbroadcastss 0x257e6(%rip),%ymm9 # 2d19c <_sk_overlay_sse2_8bit+0xced> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,1,94,2,0 ; vbroadcastss 0x25e01(%rip),%ymm9 # 2d2c0 <_sk_overlay_sse2_8bit+0xd68> + DB 196,98,125,24,13,220,87,2,0 ; vbroadcastss 0x257dc(%rip),%ymm9 # 2d1a0 <_sk_overlay_sse2_8bit+0xcf1> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 196,98,125,24,8 ; vbroadcastss (%rax),%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 @@ -6980,7 +7194,7 @@ _sk_uniform_color_avx LABEL PROC PUBLIC _sk_black_color_avx _sk_black_color_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,137,93,2,0 ; vbroadcastss 0x25d89(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,100,87,2,0 ; vbroadcastss 0x25764(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2 @@ -6989,7 +7203,7 @@ _sk_black_color_avx LABEL PROC PUBLIC _sk_white_color_avx _sk_white_color_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,112,93,2,0 ; vbroadcastss 0x25d70(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,75,87,2,0 ; vbroadcastss 0x2574b(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 @@ -7027,7 +7241,7 @@ _sk_clear_avx LABEL PROC PUBLIC _sk_srcatop_avx _sk_srcatop_avx LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,15,93,2,0 ; vbroadcastss 0x25d0f(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,234,86,2,0 ; vbroadcastss 0x256ea(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -7046,7 +7260,7 @@ _sk_srcatop_avx LABEL PROC PUBLIC _sk_dstatop_avx _sk_dstatop_avx LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 - DB 196,98,125,24,13,205,92,2,0 ; vbroadcastss 0x25ccd(%rip),%ymm9 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,13,168,86,2,0 ; vbroadcastss 0x256a8(%rip),%ymm9 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,52,92,207 ; vsubps %ymm7,%ymm9,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 @@ -7082,7 +7296,7 @@ _sk_dstin_avx LABEL PROC PUBLIC _sk_srcout_avx _sk_srcout_avx LABEL PROC - DB 196,98,125,24,5,104,92,2,0 ; vbroadcastss 0x25c68(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,67,86,2,0 ; vbroadcastss 0x25643(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -7093,7 +7307,7 @@ _sk_srcout_avx LABEL PROC PUBLIC _sk_dstout_avx _sk_dstout_avx LABEL PROC - DB 196,226,125,24,5,71,92,2,0 ; vbroadcastss 0x25c47(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,34,86,2,0 ; vbroadcastss 0x25622(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3 DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0 DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1 @@ -7104,7 +7318,7 @@ _sk_dstout_avx LABEL PROC PUBLIC _sk_srcover_avx _sk_srcover_avx LABEL PROC - DB 196,98,125,24,5,38,92,2,0 ; vbroadcastss 0x25c26(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,1,86,2,0 ; vbroadcastss 0x25601(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -7119,7 +7333,7 @@ _sk_srcover_avx LABEL PROC PUBLIC _sk_dstover_avx _sk_dstover_avx LABEL PROC - DB 196,98,125,24,5,245,91,2,0 ; vbroadcastss 0x25bf5(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,208,85,2,0 ; vbroadcastss 0x255d0(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0 @@ -7143,7 +7357,7 @@ _sk_modulate_avx LABEL PROC PUBLIC _sk_multiply_avx _sk_multiply_avx LABEL PROC - DB 196,98,125,24,5,176,91,2,0 ; vbroadcastss 0x25bb0(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,139,85,2,0 ; vbroadcastss 0x2558b(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -7172,7 +7386,7 @@ _sk_multiply_avx LABEL PROC PUBLIC _sk_plus__avx _sk_plus__avx LABEL PROC DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0 - DB 196,98,125,24,5,63,91,2,0 ; vbroadcastss 0x25b3f(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,26,85,2,0 ; vbroadcastss 0x2551a(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 197,244,88,205 ; vaddps %ymm5,%ymm1,%ymm1 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 @@ -7202,7 +7416,7 @@ _sk_screen_avx LABEL PROC PUBLIC _sk_xor__avx _sk_xor__avx LABEL PROC - DB 196,98,125,24,5,222,90,2,0 ; vbroadcastss 0x25ade(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,185,84,2,0 ; vbroadcastss 0x254b9(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -7237,7 +7451,7 @@ _sk_darken_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,90,90,2,0 ; vbroadcastss 0x25a5a(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,53,84,2,0 ; vbroadcastss 0x25435(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -7261,7 +7475,7 @@ _sk_lighten_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,2,90,2,0 ; vbroadcastss 0x25a02(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,221,83,2,0 ; vbroadcastss 0x253dd(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -7288,7 +7502,7 @@ _sk_difference_avx LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,158,89,2,0 ; vbroadcastss 0x2599e(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,121,83,2,0 ; vbroadcastss 0x25379(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -7309,7 +7523,7 @@ _sk_exclusion_avx LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,85,89,2,0 ; vbroadcastss 0x25955(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,48,83,2,0 ; vbroadcastss 0x25330(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,199 ; vmulps %ymm7,%ymm8,%ymm8 DB 197,188,88,219 ; vaddps %ymm3,%ymm8,%ymm3 @@ -7318,7 +7532,7 @@ _sk_exclusion_avx LABEL PROC PUBLIC _sk_colorburn_avx _sk_colorburn_avx LABEL PROC - DB 196,98,125,24,5,60,89,2,0 ; vbroadcastss 0x2593c(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,23,83,2,0 ; vbroadcastss 0x25317(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 @@ -7378,7 +7592,7 @@ _sk_colorburn_avx LABEL PROC PUBLIC _sk_colordodge_avx _sk_colordodge_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,64,88,2,0 ; vbroadcastss 0x25840(%rip),%ymm9 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,13,27,82,2,0 ; vbroadcastss 0x2521b(%rip),%ymm9 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9 @@ -7433,7 +7647,7 @@ _sk_colordodge_avx LABEL PROC PUBLIC _sk_hardlight_avx _sk_hardlight_avx LABEL PROC - DB 196,98,125,24,5,75,87,2,0 ; vbroadcastss 0x2574b(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,38,81,2,0 ; vbroadcastss 0x25126(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -7486,7 +7700,7 @@ _sk_hardlight_avx LABEL PROC PUBLIC _sk_overlay_avx _sk_overlay_avx LABEL PROC - DB 196,98,125,24,5,112,86,2,0 ; vbroadcastss 0x25670(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,75,80,2,0 ; vbroadcastss 0x2504b(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -7551,10 +7765,10 @@ _sk_softlight_avx LABEL PROC DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,65,60,89,216 ; vmulps %ymm8,%ymm8,%ymm11 DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,29,115,85,2,0 ; vbroadcastss 0x25573(%rip),%ymm11 # 2d2c4 <_sk_overlay_sse2_8bit+0xd6c> + DB 196,98,125,24,29,78,79,2,0 ; vbroadcastss 0x24f4e(%rip),%ymm11 # 2d1a4 <_sk_overlay_sse2_8bit+0xcf5> DB 196,65,28,88,235 ; vaddps %ymm11,%ymm12,%ymm13 DB 196,65,20,89,192 ; vmulps %ymm8,%ymm13,%ymm8 - DB 196,98,125,24,45,100,85,2,0 ; vbroadcastss 0x25564(%rip),%ymm13 # 2d2c8 <_sk_overlay_sse2_8bit+0xd70> + DB 196,98,125,24,45,63,79,2,0 ; vbroadcastss 0x24f3f(%rip),%ymm13 # 2d1a8 <_sk_overlay_sse2_8bit+0xcf9> DB 196,65,28,89,245 ; vmulps %ymm13,%ymm12,%ymm14 DB 196,65,12,88,192 ; vaddps %ymm8,%ymm14,%ymm8 DB 196,65,124,82,244 ; vrsqrtps %ymm12,%ymm14 @@ -7565,7 +7779,7 @@ _sk_softlight_avx LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,14,85,2,0 ; vbroadcastss 0x2550e(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,233,78,2,0 ; vbroadcastss 0x24ee9(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,65,60,92,228 ; vsubps %ymm12,%ymm8,%ymm12 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 @@ -7692,12 +7906,12 @@ _sk_hue_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,236,82,2,0 ; vbroadcastss 0x252ec(%rip),%ymm14 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,53,199,76,2,0 ; vbroadcastss 0x24cc7(%rip),%ymm14 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,226,82,2,0 ; vbroadcastss 0x252e2(%rip),%ymm15 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,61,189,76,2,0 ; vbroadcastss 0x24cbd(%rip),%ymm15 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,211,82,2,0 ; vbroadcastss 0x252d3(%rip),%ymm0 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,226,125,24,5,174,76,2,0 ; vbroadcastss 0x24cae(%rip),%ymm0 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -7758,7 +7972,7 @@ _sk_hue_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,128,81,2,0 ; vbroadcastss 0x25180(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,91,75,2,0 ; vbroadcastss 0x24b5b(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -7815,12 +8029,12 @@ _sk_saturation_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,164,80,2,0 ; vbroadcastss 0x250a4(%rip),%ymm14 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,53,127,74,2,0 ; vbroadcastss 0x24a7f(%rip),%ymm14 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,154,80,2,0 ; vbroadcastss 0x2509a(%rip),%ymm15 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,61,117,74,2,0 ; vbroadcastss 0x24a75(%rip),%ymm15 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,139,80,2,0 ; vbroadcastss 0x2508b(%rip),%ymm0 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,226,125,24,5,102,74,2,0 ; vbroadcastss 0x24a66(%rip),%ymm0 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -7881,7 +8095,7 @@ _sk_saturation_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,56,79,2,0 ; vbroadcastss 0x24f38(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,19,73,2,0 ; vbroadcastss 0x24913(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -7910,12 +8124,12 @@ _sk_color_avx LABEL PROC DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp) DB 197,124,89,199 ; vmulps %ymm7,%ymm0,%ymm8 DB 197,116,89,207 ; vmulps %ymm7,%ymm1,%ymm9 - DB 196,98,125,24,45,228,78,2,0 ; vbroadcastss 0x24ee4(%rip),%ymm13 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,45,191,72,2,0 ; vbroadcastss 0x248bf(%rip),%ymm13 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,92,89,213 ; vmulps %ymm13,%ymm4,%ymm10 - DB 196,98,125,24,53,218,78,2,0 ; vbroadcastss 0x24eda(%rip),%ymm14 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,53,181,72,2,0 ; vbroadcastss 0x248b5(%rip),%ymm14 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,84,89,222 ; vmulps %ymm14,%ymm5,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,203,78,2,0 ; vbroadcastss 0x24ecb(%rip),%ymm15 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,98,125,24,61,166,72,2,0 ; vbroadcastss 0x248a6(%rip),%ymm15 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 196,65,76,89,223 ; vmulps %ymm15,%ymm6,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -7978,7 +8192,7 @@ _sk_color_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,102,77,2,0 ; vbroadcastss 0x24d66(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,65,71,2,0 ; vbroadcastss 0x24741(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,84,36,32 ; vmulps 0x20(%rsp),%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -8008,12 +8222,12 @@ _sk_luminosity_avx LABEL PROC DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 - DB 196,98,125,24,45,14,77,2,0 ; vbroadcastss 0x24d0e(%rip),%ymm13 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,45,233,70,2,0 ; vbroadcastss 0x246e9(%rip),%ymm13 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,108,89,213 ; vmulps %ymm13,%ymm2,%ymm10 - DB 196,98,125,24,53,4,77,2,0 ; vbroadcastss 0x24d04(%rip),%ymm14 # 2d2d0 <_sk_overlay_sse2_8bit+0xd78> + DB 196,98,125,24,53,223,70,2,0 ; vbroadcastss 0x246df(%rip),%ymm14 # 2d1b0 <_sk_overlay_sse2_8bit+0xd01> DB 196,65,116,89,222 ; vmulps %ymm14,%ymm1,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,245,76,2,0 ; vbroadcastss 0x24cf5(%rip),%ymm15 # 2d2d4 <_sk_overlay_sse2_8bit+0xd7c> + DB 196,98,125,24,61,208,70,2,0 ; vbroadcastss 0x246d0(%rip),%ymm15 # 2d1b4 <_sk_overlay_sse2_8bit+0xd05> DB 196,65,28,89,223 ; vmulps %ymm15,%ymm12,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -8076,7 +8290,7 @@ _sk_luminosity_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,144,75,2,0 ; vbroadcastss 0x24b90(%rip),%ymm0 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,5,107,69,2,0 ; vbroadcastss 0x2456b(%rip),%ymm0 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -8107,9 +8321,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,42,1,0,0 ; jne 88b4 <_sk_srcover_rgba_8888_avx+0x14a> + DB 15,133,42,1,0,0 ; jne 8db9 <_sk_srcover_rgba_8888_avx+0x14a> DB 196,1,125,16,4,152 ; vmovupd (%r8,%r11,4),%ymm8 - DB 197,125,40,13,168,69,2,0 ; vmovapd 0x245a8(%rip),%ymm9 # 2cd40 <_sk_overlay_sse2_8bit+0x7e8> + DB 197,125,40,13,163,63,2,0 ; vmovapd 0x23fa3(%rip),%ymm9 # 2cc40 <_sk_overlay_sse2_8bit+0x791> DB 196,193,61,84,225 ; vandpd %ymm9,%ymm8,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 196,193,81,114,208,8 ; vpsrld $0x8,%xmm8,%xmm5 @@ -8127,9 +8341,9 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 197,193,114,215,24 ; vpsrld $0x18,%xmm7,%xmm7 DB 196,227,61,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm8,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,179,74,2,0 ; vbroadcastss 0x24ab3(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,142,68,2,0 ; vbroadcastss 0x2448e(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,210,74,2,0 ; vbroadcastss 0x24ad2(%rip),%ymm9 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,13,173,68,2,0 ; vbroadcastss 0x244ad(%rip),%ymm9 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,60,89,212 ; vmulps %ymm4,%ymm8,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 @@ -8162,7 +8376,7 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,65,53,86,202 ; vorpd %ymm10,%ymm9,%ymm9 DB 196,65,61,86,193 ; vorpd %ymm9,%ymm8,%ymm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 88ec <_sk_srcover_rgba_8888_avx+0x182> + DB 117,66 ; jne 8df1 <_sk_srcover_rgba_8888_avx+0x182> DB 196,1,124,17,4,152 ; vmovups %ymm8,(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8171,33 +8385,33 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,195,254,255,255 ; ja 8790 <_sk_srcover_rgba_8888_avx+0x26> + DB 15,135,195,254,255,255 ; ja 8c95 <_sk_srcover_rgba_8888_avx+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,4,1,0,0 ; lea 0x104(%rip),%r10 # 89dc <_sk_srcover_rgba_8888_avx+0x272> + DB 76,141,21,3,1,0,0 ; lea 0x103(%rip),%r10 # 8ee0 <_sk_srcover_rgba_8888_avx+0x271> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,122,16,4,152 ; vmovss (%r8,%r11,4),%xmm8 - DB 233,164,254,255,255 ; jmpq 8790 <_sk_srcover_rgba_8888_avx+0x26> + DB 233,164,254,255,255 ; jmpq 8c95 <_sk_srcover_rgba_8888_avx+0x26> DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,180 ; ja 88b0 <_sk_srcover_rgba_8888_avx+0x146> + DB 119,180 ; ja 8db5 <_sk_srcover_rgba_8888_avx+0x146> DB 65,15,182,193 ; movzbl %r9b,%eax - DB 76,141,13,241,0,0,0 ; lea 0xf1(%rip),%r9 # 89f8 <_sk_srcover_rgba_8888_avx+0x28e> + DB 76,141,13,240,0,0,0 ; lea 0xf0(%rip),%r9 # 8efc <_sk_srcover_rgba_8888_avx+0x28d> DB 73,99,4,129 ; movslq (%r9,%rax,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax DB 196,1,121,126,4,152 ; vmovd %xmm8,(%r8,%r11,4) - DB 235,152 ; jmp 88b0 <_sk_srcover_rgba_8888_avx+0x146> + DB 235,152 ; jmp 8db5 <_sk_srcover_rgba_8888_avx+0x146> DB 196,129,121,110,100,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm4 DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 196,99,85,12,196,4 ; vblendps $0x4,%ymm4,%ymm5,%ymm8 DB 196,129,123,16,36,152 ; vmovsd (%r8,%r11,4),%xmm4 DB 196,99,61,13,196,1 ; vblendpd $0x1,%ymm4,%ymm8,%ymm8 - DB 233,81,254,255,255 ; jmpq 8790 <_sk_srcover_rgba_8888_avx+0x26> + DB 233,81,254,255,255 ; jmpq 8c95 <_sk_srcover_rgba_8888_avx+0x26> DB 196,129,121,110,100,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm4 DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 DB 196,227,125,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm0,%ymm4 @@ -8211,10 +8425,10 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,99,61,24,196,1 ; vinsertf128 $0x1,%xmm4,%ymm8,%ymm8 DB 196,129,121,16,36,152 ; vmovupd (%r8,%r11,4),%xmm4 DB 196,67,93,13,192,12 ; vblendpd $0xc,%ymm8,%ymm4,%ymm8 - DB 233,252,253,255,255 ; jmpq 8790 <_sk_srcover_rgba_8888_avx+0x26> + DB 233,252,253,255,255 ; jmpq 8c95 <_sk_srcover_rgba_8888_avx+0x26> DB 196,3,121,22,68,152,8,2 ; vpextrd $0x2,%xmm8,0x8(%r8,%r11,4) DB 196,1,121,214,4,152 ; vmovq %xmm8,(%r8,%r11,4) - DB 233,9,255,255,255 ; jmpq 88b0 <_sk_srcover_rgba_8888_avx+0x146> + DB 233,9,255,255,255 ; jmpq 8db5 <_sk_srcover_rgba_8888_avx+0x146> DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,3,121,22,76,152,24,2 ; vpextrd $0x2,%xmm9,0x18(%r8,%r11,4) DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 @@ -8222,35 +8436,33 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,1,122,17,76,152,16 ; vmovss %xmm9,0x10(%r8,%r11,4) DB 196,1,121,17,4,152 ; vmovupd %xmm8,(%r8,%r11,4) - DB 233,213,254,255,255 ; jmpq 88b0 <_sk_srcover_rgba_8888_avx+0x146> - DB 144 ; nop - DB 5,255,255,255,82 ; add $0x52ffffff,%eax - DB 255 ; (bad) + DB 233,213,254,255,255 ; jmpq 8db5 <_sk_srcover_rgba_8888_avx+0x146> + DB 6 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 60,255 ; cmp $0xff,%al + DB 255,83,255 ; callq *-0x1(%rbx) DB 255 ; (bad) - DB 255,167,255,255,255,147 ; jmpq *-0x6c000001(%rdi) DB 255 ; (bad) + DB 61,255,255,255,168 ; cmp $0xa8ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 127,255 ; jg 89f1 <_sk_srcover_rgba_8888_avx+0x287> + DB 255,148,255,255,255,128,255 ; callq *-0x7f0001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,99,255 ; jmpq *-0x1(%rbx) + DB 255,100,255,255 ; jmpq *-0x1(%rdi,%rdi,8) + DB 255,25 ; lcall *(%rcx) DB 255 ; (bad) - DB 255,24 ; lcall *(%rax) DB 255 ; (bad) + DB 255,165,255,255,255,157 ; jmpq *-0x62000001(%rbp) DB 255 ; (bad) - DB 255,164,255,255,255,156,255 ; jmpq *-0x630001(%rdi,%rdi,8) DB 255 ; (bad) DB 255 ; (bad) - DB 216,255 ; fdivr %st(7),%st + DB 217,255 ; fcos DB 255 ; (bad) - DB 255,203 ; dec %ebx + DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 189,255,255,255,175 ; mov $0xafffffff,%ebp + DB 190,255,255,255,176 ; mov $0xb0ffffff,%esi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8267,7 +8479,7 @@ _sk_clamp_0_avx LABEL PROC PUBLIC _sk_clamp_1_avx _sk_clamp_1_avx LABEL PROC - DB 196,98,125,24,5,114,72,2,0 ; vbroadcastss 0x24872(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,78,66,2,0 ; vbroadcastss 0x2424e(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 @@ -8277,7 +8489,7 @@ _sk_clamp_1_avx LABEL PROC PUBLIC _sk_clamp_a_avx _sk_clamp_a_avx LABEL PROC - DB 196,98,125,24,5,81,72,2,0 ; vbroadcastss 0x24851(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,45,66,2,0 ; vbroadcastss 0x2422d(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0 DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1 @@ -8287,7 +8499,7 @@ _sk_clamp_a_avx LABEL PROC PUBLIC _sk_clamp_a_dst_avx _sk_clamp_a_dst_avx LABEL PROC - DB 196,98,125,24,5,51,72,2,0 ; vbroadcastss 0x24833(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,15,66,2,0 ; vbroadcastss 0x2420f(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,68,93,248 ; vminps %ymm8,%ymm7,%ymm7 DB 197,220,93,231 ; vminps %ymm7,%ymm4,%ymm4 DB 197,212,93,239 ; vminps %ymm7,%ymm5,%ymm5 @@ -8314,7 +8526,7 @@ _sk_swap_rb_avx LABEL PROC PUBLIC _sk_invert_avx _sk_invert_avx LABEL PROC - DB 196,98,125,24,5,238,71,2,0 ; vbroadcastss 0x247ee(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,202,65,2,0 ; vbroadcastss 0x241ca(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 @@ -8360,7 +8572,7 @@ PUBLIC _sk_unpremul_avx _sk_unpremul_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,126,71,2,0 ; vbroadcastss 0x2477e(%rip),%ymm10 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,21,90,65,2,0 ; vbroadcastss 0x2415a(%rip),%ymm10 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10 DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 @@ -8371,17 +8583,17 @@ _sk_unpremul_avx LABEL PROC PUBLIC _sk_from_srgb_avx _sk_from_srgb_avx LABEL PROC - DB 196,98,125,24,5,139,71,2,0 ; vbroadcastss 0x2478b(%rip),%ymm8 # 2d2dc <_sk_overlay_sse2_8bit+0xd84> + DB 196,98,125,24,5,103,65,2,0 ; vbroadcastss 0x24167(%rip),%ymm8 # 2d1bc <_sk_overlay_sse2_8bit+0xd0d> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,105,71,2,0 ; vbroadcastss 0x24769(%rip),%ymm11 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,29,69,65,2,0 ; vbroadcastss 0x24145(%rip),%ymm11 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12 - DB 196,98,125,24,45,111,71,2,0 ; vbroadcastss 0x2476f(%rip),%ymm13 # 2d2e0 <_sk_overlay_sse2_8bit+0xd88> + DB 196,98,125,24,45,75,65,2,0 ; vbroadcastss 0x2414b(%rip),%ymm13 # 2d1c0 <_sk_overlay_sse2_8bit+0xd11> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,96,71,2,0 ; vbroadcastss 0x24760(%rip),%ymm12 # 2d2e4 <_sk_overlay_sse2_8bit+0xd8c> + DB 196,98,125,24,37,60,65,2,0 ; vbroadcastss 0x2413c(%rip),%ymm12 # 2d1c4 <_sk_overlay_sse2_8bit+0xd15> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,53,86,71,2,0 ; vbroadcastss 0x24756(%rip),%ymm14 # 2d2e8 <_sk_overlay_sse2_8bit+0xd90> + DB 196,98,125,24,53,50,65,2,0 ; vbroadcastss 0x24132(%rip),%ymm14 # 2d1c8 <_sk_overlay_sse2_8bit+0xd19> DB 196,193,124,194,198,1 ; vcmpltps %ymm14,%ymm0,%ymm0 DB 196,195,45,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm10,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -8405,17 +8617,17 @@ _sk_from_srgb_avx LABEL PROC PUBLIC _sk_from_srgb_dst_avx _sk_from_srgb_dst_avx LABEL PROC - DB 196,98,125,24,5,223,70,2,0 ; vbroadcastss 0x246df(%rip),%ymm8 # 2d2dc <_sk_overlay_sse2_8bit+0xd84> + DB 196,98,125,24,5,187,64,2,0 ; vbroadcastss 0x240bb(%rip),%ymm8 # 2d1bc <_sk_overlay_sse2_8bit+0xd0d> DB 196,65,92,89,200 ; vmulps %ymm8,%ymm4,%ymm9 DB 197,92,89,212 ; vmulps %ymm4,%ymm4,%ymm10 - DB 196,98,125,24,29,189,70,2,0 ; vbroadcastss 0x246bd(%rip),%ymm11 # 2d2cc <_sk_overlay_sse2_8bit+0xd74> + DB 196,98,125,24,29,153,64,2,0 ; vbroadcastss 0x24099(%rip),%ymm11 # 2d1ac <_sk_overlay_sse2_8bit+0xcfd> DB 196,65,92,89,227 ; vmulps %ymm11,%ymm4,%ymm12 - DB 196,98,125,24,45,195,70,2,0 ; vbroadcastss 0x246c3(%rip),%ymm13 # 2d2e0 <_sk_overlay_sse2_8bit+0xd88> + DB 196,98,125,24,45,159,64,2,0 ; vbroadcastss 0x2409f(%rip),%ymm13 # 2d1c0 <_sk_overlay_sse2_8bit+0xd11> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,180,70,2,0 ; vbroadcastss 0x246b4(%rip),%ymm12 # 2d2e4 <_sk_overlay_sse2_8bit+0xd8c> + DB 196,98,125,24,37,144,64,2,0 ; vbroadcastss 0x24090(%rip),%ymm12 # 2d1c4 <_sk_overlay_sse2_8bit+0xd15> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,53,170,70,2,0 ; vbroadcastss 0x246aa(%rip),%ymm14 # 2d2e8 <_sk_overlay_sse2_8bit+0xd90> + DB 196,98,125,24,53,134,64,2,0 ; vbroadcastss 0x24086(%rip),%ymm14 # 2d1c8 <_sk_overlay_sse2_8bit+0xd19> DB 196,193,92,194,230,1 ; vcmpltps %ymm14,%ymm4,%ymm4 DB 196,195,45,74,225,64 ; vblendvps %ymm4,%ymm9,%ymm10,%ymm4 DB 196,65,84,89,200 ; vmulps %ymm8,%ymm5,%ymm9 @@ -8440,20 +8652,20 @@ _sk_from_srgb_dst_avx LABEL PROC PUBLIC _sk_to_srgb_avx _sk_to_srgb_avx LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,63,70,2,0 ; vbroadcastss 0x2463f(%rip),%ymm8 # 2d2ec <_sk_overlay_sse2_8bit+0xd94> + DB 196,98,125,24,5,27,64,2,0 ; vbroadcastss 0x2401b(%rip),%ymm8 # 2d1cc <_sk_overlay_sse2_8bit+0xd1d> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,57,70,2,0 ; vbroadcastss 0x24639(%rip),%ymm11 # 2d2f4 <_sk_overlay_sse2_8bit+0xd9c> + DB 196,98,125,24,29,21,64,2,0 ; vbroadcastss 0x24015(%rip),%ymm11 # 2d1d4 <_sk_overlay_sse2_8bit+0xd25> DB 196,65,52,89,227 ; vmulps %ymm11,%ymm9,%ymm12 - DB 196,98,125,24,45,39,70,2,0 ; vbroadcastss 0x24627(%rip),%ymm13 # 2d2f0 <_sk_overlay_sse2_8bit+0xd98> + DB 196,98,125,24,45,3,64,2,0 ; vbroadcastss 0x24003(%rip),%ymm13 # 2d1d0 <_sk_overlay_sse2_8bit+0xd21> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,52,89,228 ; vmulps %ymm12,%ymm9,%ymm12 - DB 196,98,125,24,53,28,70,2,0 ; vbroadcastss 0x2461c(%rip),%ymm14 # 2d2f8 <_sk_overlay_sse2_8bit+0xda0> + DB 196,98,125,24,53,248,63,2,0 ; vbroadcastss 0x23ff8(%rip),%ymm14 # 2d1d8 <_sk_overlay_sse2_8bit+0xd29> DB 196,65,28,88,230 ; vaddps %ymm14,%ymm12,%ymm12 - DB 196,98,125,24,61,18,70,2,0 ; vbroadcastss 0x24612(%rip),%ymm15 # 2d2fc <_sk_overlay_sse2_8bit+0xda4> + DB 196,98,125,24,61,238,63,2,0 ; vbroadcastss 0x23fee(%rip),%ymm15 # 2d1dc <_sk_overlay_sse2_8bit+0xd2d> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 - DB 196,98,125,24,37,254,69,2,0 ; vbroadcastss 0x245fe(%rip),%ymm12 # 2d300 <_sk_overlay_sse2_8bit+0xda8> + DB 196,98,125,24,37,218,63,2,0 ; vbroadcastss 0x23fda(%rip),%ymm12 # 2d1e0 <_sk_overlay_sse2_8bit+0xd31> DB 196,193,124,194,196,1 ; vcmpltps %ymm12,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -8488,7 +8700,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,116,93,202 ; vminps %ymm2,%ymm1,%ymm9 DB 196,65,124,93,201 ; vminps %ymm9,%ymm0,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,10,69,2,0 ; vbroadcastss 0x2450a(%rip),%ymm11 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,29,230,62,2,0 ; vbroadcastss 0x23ee6(%rip),%ymm11 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 196,65,28,89,227 ; vmulps %ymm11,%ymm12,%ymm12 @@ -8498,19 +8710,19 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 DB 196,193,116,89,203 ; vmulps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,55,69,2,0 ; vbroadcastss 0x24537(%rip),%ymm11 # 2d30c <_sk_overlay_sse2_8bit+0xdb4> + DB 196,98,125,24,29,19,63,2,0 ; vbroadcastss 0x23f13(%rip),%ymm11 # 2d1ec <_sk_overlay_sse2_8bit+0xd3d> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,37,69,2,0 ; vbroadcastss 0x24525(%rip),%ymm11 # 2d308 <_sk_overlay_sse2_8bit+0xdb0> + DB 196,98,125,24,29,1,63,2,0 ; vbroadcastss 0x23f01(%rip),%ymm11 # 2d1e8 <_sk_overlay_sse2_8bit+0xd39> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,227,117,74,202,224 ; vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,13,69,2,0 ; vbroadcastss 0x2450d(%rip),%ymm2 # 2d304 <_sk_overlay_sse2_8bit+0xdac> + DB 196,226,125,24,21,233,62,2,0 ; vbroadcastss 0x23ee9(%rip),%ymm2 # 2d1e4 <_sk_overlay_sse2_8bit+0xd35> DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 DB 196,227,117,74,194,0 ; vblendvps %ymm0,%ymm2,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,37,136,68,2,0 ; vbroadcastss 0x24488(%rip),%ymm12 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,37,100,62,2,0 ; vbroadcastss 0x23e64(%rip),%ymm12 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,193,116,89,212 ; vmulps %ymm12,%ymm1,%ymm2 DB 197,28,194,226,1 ; vcmpltps %ymm2,%ymm12,%ymm12 DB 196,65,36,92,216 ; vsubps %ymm8,%ymm11,%ymm11 @@ -8520,7 +8732,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,198,128 ; vblendvps %ymm8,%ymm14,%ymm0,%ymm0 DB 196,195,117,74,206,128 ; vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - DB 196,98,125,24,5,183,68,2,0 ; vbroadcastss 0x244b7(%rip),%ymm8 # 2d310 <_sk_overlay_sse2_8bit+0xdb8> + DB 196,98,125,24,5,147,62,2,0 ; vbroadcastss 0x23e93(%rip),%ymm8 # 2d1f0 <_sk_overlay_sse2_8bit+0xd41> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8535,7 +8747,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,225 ; vmovaps %ymm1,%ymm4 DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 - DB 196,98,125,24,5,14,68,2,0 ; vbroadcastss 0x2440e(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,234,61,2,0 ; vbroadcastss 0x23dea(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,92,89,210 ; vmulps %ymm2,%ymm4,%ymm10 DB 196,65,92,92,218 ; vsubps %ymm10,%ymm4,%ymm11 @@ -8543,23 +8755,23 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 DB 197,108,88,202 ; vaddps %ymm2,%ymm2,%ymm9 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,29,80,68,2,0 ; vbroadcastss 0x24450(%rip),%ymm11 # 2d314 <_sk_overlay_sse2_8bit+0xdbc> + DB 196,98,125,24,29,44,62,2,0 ; vbroadcastss 0x23e2c(%rip),%ymm11 # 2d1f4 <_sk_overlay_sse2_8bit+0xd45> DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,37,34,68,2,0 ; vbroadcastss 0x24422(%rip),%ymm12 # 2d304 <_sk_overlay_sse2_8bit+0xdac> + DB 196,98,125,24,37,254,61,2,0 ; vbroadcastss 0x23dfe(%rip),%ymm12 # 2d1e4 <_sk_overlay_sse2_8bit+0xd35> DB 196,193,4,89,196 ; vmulps %ymm12,%ymm15,%ymm0 - DB 196,98,125,24,45,28,68,2,0 ; vbroadcastss 0x2441c(%rip),%ymm13 # 2d30c <_sk_overlay_sse2_8bit+0xdb4> + DB 196,98,125,24,45,248,61,2,0 ; vbroadcastss 0x23df8(%rip),%ymm13 # 2d1ec <_sk_overlay_sse2_8bit+0xd3d> DB 197,20,92,240 ; vsubps %ymm0,%ymm13,%ymm14 DB 196,65,36,89,246 ; vmulps %ymm14,%ymm11,%ymm14 DB 196,65,52,88,246 ; vaddps %ymm14,%ymm9,%ymm14 - DB 196,226,125,24,13,17,68,2,0 ; vbroadcastss 0x24411(%rip),%ymm1 # 2d318 <_sk_overlay_sse2_8bit+0xdc0> + DB 196,226,125,24,13,237,61,2,0 ; vbroadcastss 0x23ded(%rip),%ymm1 # 2d1f8 <_sk_overlay_sse2_8bit+0xd49> DB 196,193,116,194,255,2 ; vcmpleps %ymm15,%ymm1,%ymm7 DB 196,195,13,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm14,%ymm7 DB 196,65,60,194,247,2 ; vcmpleps %ymm15,%ymm8,%ymm14 DB 196,227,45,74,255,224 ; vblendvps %ymm14,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,53,232,67,2,0 ; vbroadcastss 0x243e8(%rip),%ymm14 # 2d310 <_sk_overlay_sse2_8bit+0xdb8> + DB 196,98,125,24,53,196,61,2,0 ; vbroadcastss 0x23dc4(%rip),%ymm14 # 2d1f0 <_sk_overlay_sse2_8bit+0xd41> DB 196,65,12,194,255,2 ; vcmpleps %ymm15,%ymm14,%ymm15 DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -8578,7 +8790,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,164,89,247 ; vmulps %ymm7,%ymm11,%ymm6 DB 197,180,88,246 ; vaddps %ymm6,%ymm9,%ymm6 DB 196,227,77,74,237,0 ; vblendvps %ymm0,%ymm5,%ymm6,%ymm5 - DB 196,226,125,24,5,146,67,2,0 ; vbroadcastss 0x24392(%rip),%ymm0 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> + DB 196,226,125,24,5,110,61,2,0 ; vbroadcastss 0x23d6e(%rip),%ymm0 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> DB 197,228,88,192 ; vaddps %ymm0,%ymm3,%ymm0 DB 196,227,125,8,216,1 ; vroundps $0x1,%ymm0,%ymm3 DB 197,252,92,195 ; vsubps %ymm3,%ymm0,%ymm0 @@ -8628,15 +8840,15 @@ _sk_scale_u8_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,74 ; jne 9095 <_sk_scale_u8_avx+0x62> + DB 117,75 ; jne 959a <_sk_scale_u8_avx+0x63> DB 196,2,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm8 - DB 197,57,219,5,247,67,2,0 ; vpand 0x243f7(%rip),%xmm8,%xmm8 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 + DB 197,57,219,5,227,61,2,0 ; vpand 0x23de3(%rip),%xmm8,%xmm8 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,66,121,51,200 ; vpmovzxwd %xmm8,%xmm9 + DB 196,65,121,112,192,78 ; vpshufd $0x4e,%xmm8,%xmm8 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 - DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 + DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,164,66,2,0 ; vbroadcastss 0x242a4(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,127,60,2,0 ; vbroadcastss 0x23c7f(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -8649,15 +8861,15 @@ _sk_scale_u8_avx LABEL PROC DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,167 ; ja 9051 <_sk_scale_u8_avx+0x1e> + DB 119,166 ; ja 9555 <_sk_scale_u8_avx+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,123,0,0,0 ; lea 0x7b(%rip),%r10 # 9130 <_sk_scale_u8_avx+0xfd> + DB 76,141,21,122,0,0,0 ; lea 0x7a(%rip),%r10 # 9634 <_sk_scale_u8_avx+0xfd> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,121,110,192 ; vmovd %eax,%xmm8 - DB 235,136 ; jmp 9051 <_sk_scale_u8_avx+0x1e> + DB 235,135 ; jmp 9555 <_sk_scale_u8_avx+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 197,57,196,192,2 ; vpinsrw $0x2,%eax,%xmm8,%xmm8 @@ -8665,7 +8877,7 @@ _sk_scale_u8_avx LABEL PROC DB 197,121,110,200 ; vmovd %eax,%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 DB 196,67,57,14,193,3 ; vpblendw $0x3,%xmm9,%xmm8,%xmm8 - DB 233,95,255,255,255 ; jmpq 9051 <_sk_scale_u8_avx+0x1e> + DB 233,94,255,255,255 ; jmpq 9555 <_sk_scale_u8_avx+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 197,57,196,192,6 ; vpinsrw $0x6,%eax,%xmm8,%xmm8 @@ -8676,21 +8888,22 @@ _sk_scale_u8_avx LABEL PROC DB 196,1,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 DB 196,67,49,14,192,240 ; vpblendw $0xf0,%xmm8,%xmm9,%xmm8 - DB 233,35,255,255,255 ; jmpq 9051 <_sk_scale_u8_avx+0x1e> - DB 102,144 ; xchg %ax,%ax - DB 142,255 ; mov %edi,%? + DB 233,34,255,255,255 ; jmpq 9555 <_sk_scale_u8_avx+0x1e> + DB 144 ; nop + DB 143 ; (bad) DB 255 ; (bad) - DB 255,169,255,255,255,153 ; ljmp *-0x66000001(%rcx) + DB 255 ; (bad) + DB 255,170,255,255,255,154 ; ljmp *-0x65000001(%rdx) DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 232,255,255,255,221 ; callq ffffffffde009140 <_sk_overlay_sse2_8bit+0xffffffffddfdcbe8> + DB 233,255,255,255,222 ; jmpq ffffffffdf009644 <_sk_overlay_sse2_8bit+0xffffffffdefdd195> DB 255 ; (bad) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255,211 ; callq *%rbx DB 255 ; (bad) DB 255 ; (bad) - DB 255,194 ; inc %edx + DB 255,195 ; inc %ebx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8705,26 +8918,26 @@ _sk_scale_565_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,177,0,0,0 ; jne 921c <_sk_scale_565_avx+0xd0> + DB 15,133,178,0,0,0 ; jne 9721 <_sk_scale_565_avx+0xd1> DB 196,1,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm8 - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 + DB 196,66,121,51,200 ; vpmovzxwd %xmm8,%xmm9 + DB 196,65,121,112,192,78 ; vpshufd $0x4e,%xmm8,%xmm8 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 - DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - DB 196,98,125,24,13,149,65,2,0 ; vbroadcastss 0x24195(%rip),%ymm9 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 + DB 196,98,125,24,13,112,59,2,0 ; vbroadcastss 0x23b70(%rip),%ymm9 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 196,65,60,84,201 ; vandps %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,134,65,2,0 ; vbroadcastss 0x24186(%rip),%ymm10 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,98,125,24,21,97,59,2,0 ; vbroadcastss 0x23b61(%rip),%ymm10 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,124,65,2,0 ; vbroadcastss 0x2417c(%rip),%ymm10 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,98,125,24,21,87,59,2,0 ; vbroadcastss 0x23b57(%rip),%ymm10 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 196,65,60,84,210 ; vandps %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,109,65,2,0 ; vbroadcastss 0x2416d(%rip),%ymm11 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,98,125,24,29,72,59,2,0 ; vbroadcastss 0x23b48(%rip),%ymm11 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,29,99,65,2,0 ; vbroadcastss 0x24163(%rip),%ymm11 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,98,125,24,29,62,59,2,0 ; vbroadcastss 0x23b3e(%rip),%ymm11 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 196,65,60,84,195 ; vandps %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,84,65,2,0 ; vbroadcastss 0x24154(%rip),%ymm11 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,98,125,24,29,47,59,2,0 ; vbroadcastss 0x23b2f(%rip),%ymm11 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,100,194,223,1 ; vcmpltps %ymm7,%ymm3,%ymm11 DB 196,65,44,93,224 ; vminps %ymm8,%ymm10,%ymm12 @@ -8743,44 +8956,42 @@ _sk_scale_565_avx LABEL PROC DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,60,255,255,255 ; ja 9171 <_sk_scale_565_avx+0x25> + DB 15,135,59,255,255,255 ; ja 9675 <_sk_scale_565_avx+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 92a4 <_sk_scale_565_avx+0x158> + DB 76,141,21,99,0,0,0 ; lea 0x63(%rip),%r10 # 97a8 <_sk_scale_565_avx+0x158> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,121,110,192 ; vmovd %eax,%xmm8 - DB 233,26,255,255,255 ; jmpq 9171 <_sk_scale_565_avx+0x25> + DB 233,25,255,255,255 ; jmpq 9675 <_sk_scale_565_avx+0x25> DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 196,1,57,196,68,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,121,110,12,88 ; vmovd (%r8,%r11,2),%xmm9 DB 196,67,57,14,193,3 ; vpblendw $0x3,%xmm9,%xmm8,%xmm8 - DB 233,252,254,255,255 ; jmpq 9171 <_sk_scale_565_avx+0x25> + DB 233,251,254,255,255 ; jmpq 9675 <_sk_scale_565_avx+0x25> DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 196,1,57,196,68,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,57,196,68,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,57,196,68,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,122,126,12,88 ; vmovq (%r8,%r11,2),%xmm9 DB 196,67,49,14,192,240 ; vpblendw $0xf0,%xmm8,%xmm9,%xmm8 - DB 233,206,254,255,255 ; jmpq 9171 <_sk_scale_565_avx+0x25> - DB 144 ; nop - DB 165 ; movsl %ds:(%rsi),%es:(%rdi) - DB 255 ; (bad) + DB 233,205,254,255,255 ; jmpq 9675 <_sk_scale_565_avx+0x25> + DB 166 ; cmpsb %es:(%rdi),%ds:(%rsi) DB 255 ; (bad) - DB 255,192 ; inc %eax DB 255 ; (bad) + DB 255,193 ; inc %ecx DB 255 ; (bad) - DB 255,179,255,255,255,238 ; pushq -0x11000001(%rbx) DB 255 ; (bad) + DB 255,180,255,255,255,239,255 ; pushq -0x100001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) + DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8813,15 +9024,15 @@ _sk_lerp_u8_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,110 ; jne 9385 <_sk_lerp_u8_avx+0x86> + DB 117,111 ; jne 988a <_sk_lerp_u8_avx+0x87> DB 196,2,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm8 - DB 197,57,219,5,43,65,2,0 ; vpand 0x2412b(%rip),%xmm8,%xmm8 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 + DB 197,57,219,5,23,59,2,0 ; vpand 0x23b17(%rip),%xmm8,%xmm8 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,66,121,51,200 ; vpmovzxwd %xmm8,%xmm9 + DB 196,65,121,112,192,78 ; vpshufd $0x4e,%xmm8,%xmm8 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 - DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 + DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,216,63,2,0 ; vbroadcastss 0x23fd8(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,179,57,2,0 ; vbroadcastss 0x239b3(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 @@ -8842,15 +9053,15 @@ _sk_lerp_u8_avx LABEL PROC DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,131 ; ja 931d <_sk_lerp_u8_avx+0x1e> + DB 119,130 ; ja 9821 <_sk_lerp_u8_avx+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,127,0,0,0 ; lea 0x7f(%rip),%r10 # 9424 <_sk_lerp_u8_avx+0x125> + DB 76,141,21,126,0,0,0 ; lea 0x7e(%rip),%r10 # 9928 <_sk_lerp_u8_avx+0x125> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,121,110,192 ; vmovd %eax,%xmm8 - DB 233,97,255,255,255 ; jmpq 931d <_sk_lerp_u8_avx+0x1e> + DB 233,96,255,255,255 ; jmpq 9821 <_sk_lerp_u8_avx+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 197,57,196,192,2 ; vpinsrw $0x2,%eax,%xmm8,%xmm8 @@ -8858,7 +9069,7 @@ _sk_lerp_u8_avx LABEL PROC DB 197,121,110,200 ; vmovd %eax,%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 DB 196,67,57,14,193,3 ; vpblendw $0x3,%xmm9,%xmm8,%xmm8 - DB 233,56,255,255,255 ; jmpq 931d <_sk_lerp_u8_avx+0x1e> + DB 233,55,255,255,255 ; jmpq 9821 <_sk_lerp_u8_avx+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 197,57,196,192,6 ; vpinsrw $0x6,%eax,%xmm8,%xmm8 @@ -8869,23 +9080,21 @@ _sk_lerp_u8_avx LABEL PROC DB 196,1,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm9 DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 DB 196,67,49,14,192,240 ; vpblendw $0xf0,%xmm8,%xmm9,%xmm8 - DB 233,252,254,255,255 ; jmpq 931d <_sk_lerp_u8_avx+0x1e> - DB 15,31,0 ; nopl (%rax) - DB 138,255 ; mov %bh,%bh - DB 255 ; (bad) - DB 255,168,255,255,255,152 ; ljmp *-0x67000001(%rax) + DB 233,251,254,255,255 ; jmpq 9821 <_sk_lerp_u8_avx+0x1e> + DB 102,144 ; xchg %ax,%ax + DB 139,255 ; mov %edi,%edi DB 255 ; (bad) + DB 255,169,255,255,255,153 ; ljmp *-0x66000001(%rcx) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) + DB 232,255,255,255,221 ; callq ffffffffde009938 <_sk_overlay_sse2_8bit+0xffffffffddfdd489> DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) - DB 255,193 ; inc %ecx + DB 255,194 ; inc %edx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8900,26 +9109,26 @@ _sk_lerp_565_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,213,0,0,0 ; jne 9534 <_sk_lerp_565_avx+0xf4> + DB 15,133,214,0,0,0 ; jne 9a39 <_sk_lerp_565_avx+0xf5> DB 196,1,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm8 - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 + DB 196,66,121,51,200 ; vpmovzxwd %xmm8,%xmm9 + DB 196,65,121,112,192,78 ; vpshufd $0x4e,%xmm8,%xmm8 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 - DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - DB 196,98,125,24,13,161,62,2,0 ; vbroadcastss 0x23ea1(%rip),%ymm9 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 + DB 196,98,125,24,13,124,56,2,0 ; vbroadcastss 0x2387c(%rip),%ymm9 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 196,65,60,84,201 ; vandps %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,146,62,2,0 ; vbroadcastss 0x23e92(%rip),%ymm10 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,98,125,24,21,109,56,2,0 ; vbroadcastss 0x2386d(%rip),%ymm10 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,136,62,2,0 ; vbroadcastss 0x23e88(%rip),%ymm10 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,98,125,24,21,99,56,2,0 ; vbroadcastss 0x23863(%rip),%ymm10 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 196,65,60,84,210 ; vandps %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,121,62,2,0 ; vbroadcastss 0x23e79(%rip),%ymm11 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,98,125,24,29,84,56,2,0 ; vbroadcastss 0x23854(%rip),%ymm11 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,29,111,62,2,0 ; vbroadcastss 0x23e6f(%rip),%ymm11 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,98,125,24,29,74,56,2,0 ; vbroadcastss 0x2384a(%rip),%ymm11 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 196,65,60,84,195 ; vandps %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,96,62,2,0 ; vbroadcastss 0x23e60(%rip),%ymm11 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,98,125,24,29,59,56,2,0 ; vbroadcastss 0x2383b(%rip),%ymm11 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,100,194,223,1 ; vcmpltps %ymm7,%ymm3,%ymm11 DB 196,65,44,93,224 ; vminps %ymm8,%ymm10,%ymm12 @@ -8946,44 +9155,42 @@ _sk_lerp_565_avx LABEL PROC DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,24,255,255,255 ; ja 9465 <_sk_lerp_565_avx+0x25> + DB 15,135,23,255,255,255 ; ja 9969 <_sk_lerp_565_avx+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 95bc <_sk_lerp_565_avx+0x17c> + DB 76,141,21,99,0,0,0 ; lea 0x63(%rip),%r10 # 9ac0 <_sk_lerp_565_avx+0x17c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,121,110,192 ; vmovd %eax,%xmm8 - DB 233,246,254,255,255 ; jmpq 9465 <_sk_lerp_565_avx+0x25> + DB 233,245,254,255,255 ; jmpq 9969 <_sk_lerp_565_avx+0x25> DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 196,1,57,196,68,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,121,110,12,88 ; vmovd (%r8,%r11,2),%xmm9 DB 196,67,57,14,193,3 ; vpblendw $0x3,%xmm9,%xmm8,%xmm8 - DB 233,216,254,255,255 ; jmpq 9465 <_sk_lerp_565_avx+0x25> + DB 233,215,254,255,255 ; jmpq 9969 <_sk_lerp_565_avx+0x25> DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 196,1,57,196,68,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,57,196,68,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,57,196,68,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm8,%xmm8 DB 196,1,122,126,12,88 ; vmovq (%r8,%r11,2),%xmm9 DB 196,67,49,14,192,240 ; vpblendw $0xf0,%xmm8,%xmm9,%xmm8 - DB 233,170,254,255,255 ; jmpq 9465 <_sk_lerp_565_avx+0x25> - DB 144 ; nop - DB 165 ; movsl %ds:(%rsi),%es:(%rdi) - DB 255 ; (bad) + DB 233,169,254,255,255 ; jmpq 9969 <_sk_lerp_565_avx+0x25> + DB 166 ; cmpsb %es:(%rdi),%ds:(%rsi) DB 255 ; (bad) - DB 255,192 ; inc %eax DB 255 ; (bad) + DB 255,193 ; inc %ecx DB 255 ; (bad) - DB 255,179,255,255,255,238 ; pushq -0x11000001(%rbx) DB 255 ; (bad) + DB 255,180,255,255,255,239,255 ; pushq -0x100001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) + DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8995,9 +9202,9 @@ _sk_load_tables_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,7,2,0,0 ; jne 97f6 <_sk_load_tables_avx+0x21e> + DB 15,133,7,2,0,0 ; jne 9cfa <_sk_load_tables_avx+0x21e> DB 196,65,125,16,20,144 ; vmovupd (%r8,%rdx,4),%ymm10 - DB 197,125,40,13,99,55,2,0 ; vmovapd 0x23763(%rip),%ymm9 # 2cd60 <_sk_overlay_sse2_8bit+0x808> + DB 197,125,40,13,95,49,2,0 ; vmovapd 0x2315f(%rip),%ymm9 # 2cc60 <_sk_overlay_sse2_8bit+0x7b1> DB 196,193,45,84,201 ; vandpd %ymm9,%ymm10,%ymm1 DB 196,227,125,25,200,1 ; vextractf128 $0x1,%ymm1,%xmm0 DB 196,193,249,126,192 ; vmovq %xmm0,%r8 @@ -9089,7 +9296,7 @@ _sk_load_tables_avx LABEL PROC DB 196,193,65,114,213,24 ; vpsrld $0x18,%xmm13,%xmm7 DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,226,125,24,61,59,59,2,0 ; vbroadcastss 0x23b3b(%rip),%ymm7 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,61,23,53,2,0 ; vbroadcastss 0x23517(%rip),%ymm7 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7 @@ -9100,21 +9307,21 @@ _sk_load_tables_avx LABEL PROC DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,230,253,255,255 ; ja 95f5 <_sk_load_tables_avx+0x1d> + DB 15,135,230,253,255,255 ; ja 9af9 <_sk_load_tables_avx+0x1d> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,146,0,0,0 ; lea 0x92(%rip),%r10 # 98ac <_sk_load_tables_avx+0x2d4> + DB 76,141,21,146,0,0,0 ; lea 0x92(%rip),%r10 # 9db0 <_sk_load_tables_avx+0x2d4> DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9 DB 77,1,209 ; add %r10,%r9 DB 65,255,225 ; jmpq *%r9 DB 196,65,122,16,20,144 ; vmovss (%r8,%rdx,4),%xmm10 - DB 233,198,253,255,255 ; jmpq 95f5 <_sk_load_tables_avx+0x1d> + DB 233,198,253,255,255 ; jmpq 9af9 <_sk_load_tables_avx+0x1d> DB 196,193,121,110,68,144,8 ; vmovd 0x8(%r8,%rdx,4),%xmm0 DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,99,117,12,208,4 ; vblendps $0x4,%ymm0,%ymm1,%ymm10 DB 196,193,123,16,4,144 ; vmovsd (%r8,%rdx,4),%xmm0 DB 196,99,45,13,208,1 ; vblendpd $0x1,%ymm0,%ymm10,%ymm10 - DB 233,159,253,255,255 ; jmpq 95f5 <_sk_load_tables_avx+0x1d> + DB 233,159,253,255,255 ; jmpq 9af9 <_sk_load_tables_avx+0x1d> DB 196,193,121,110,68,144,24 ; vmovd 0x18(%r8,%rdx,4),%xmm0 DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 @@ -9128,9 +9335,9 @@ _sk_load_tables_avx LABEL PROC DB 196,99,45,24,208,1 ; vinsertf128 $0x1,%xmm0,%ymm10,%ymm10 DB 196,193,121,16,4,144 ; vmovupd (%r8,%rdx,4),%xmm0 DB 196,67,125,13,210,12 ; vblendpd $0xc,%ymm10,%ymm0,%ymm10 - DB 233,74,253,255,255 ; jmpq 95f5 <_sk_load_tables_avx+0x1d> + DB 233,74,253,255,255 ; jmpq 9af9 <_sk_load_tables_avx+0x1d> DB 144 ; nop - DB 120,255 ; js 98ad <_sk_load_tables_avx+0x2d5> + DB 120,255 ; js 9db1 <_sk_load_tables_avx+0x2d5> DB 255 ; (bad) DB 255,153,255,255,255,131 ; lcall *-0x7c000001(%rcx) DB 255 ; (bad) @@ -9153,13 +9360,14 @@ _sk_load_tables_avx LABEL PROC PUBLIC _sk_load_tables_u16_be_avx _sk_load_tables_u16_be_avx LABEL PROC - DB 72,131,236,56 ; sub $0x38,%rsp + DB 72,131,236,88 ; sub $0x58,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 197,252,17,60,36 ; vmovups %ymm7,(%rsp) - DB 15,133,87,2,0,0 ; jne 9b3e <_sk_load_tables_u16_be_avx+0x276> + DB 197,252,17,124,36,32 ; vmovups %ymm7,0x20(%rsp) + DB 197,252,17,52,36 ; vmovups %ymm6,(%rsp) + DB 15,133,78,2,0,0 ; jne a03f <_sk_load_tables_u16_be_avx+0x273> DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8 DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2 DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3 @@ -9171,139 +9379,137 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 197,121,97,202 ; vpunpcklwd %xmm2,%xmm0,%xmm9 DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8 DB 197,241,97,195 ; vpunpcklwd %xmm3,%xmm1,%xmm0 - DB 197,113,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm11 - DB 197,177,108,200 ; vpunpcklqdq %xmm0,%xmm9,%xmm1 - DB 197,49,109,224 ; vpunpckhqdq %xmm0,%xmm9,%xmm12 - DB 197,121,111,21,28,59,2,0 ; vmovdqa 0x23b1c(%rip),%xmm10 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 196,193,113,219,202 ; vpand %xmm10,%xmm1,%xmm1 - DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2 + DB 197,113,105,227 ; vpunpckhwd %xmm3,%xmm1,%xmm12 + DB 197,177,108,208 ; vpunpcklqdq %xmm0,%xmm9,%xmm2 + DB 197,49,109,232 ; vpunpckhqdq %xmm0,%xmm9,%xmm13 + DB 196,65,57,108,212 ; vpunpcklqdq %xmm12,%xmm8,%xmm10 + DB 197,121,111,29,253,52,2,0 ; vmovdqa 0x234fd(%rip),%xmm11 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,193,105,219,195 ; vpand %xmm11,%xmm2,%xmm0 + DB 197,249,112,208,78 ; vpshufd $0x4e,%xmm0,%xmm2 + DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 DB 196,193,249,126,208 ; vmovq %xmm2,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 76,139,80,8 ; mov 0x8(%rax),%r10 DB 196,129,122,16,28,138 ; vmovss (%r10,%r9,4),%xmm3 DB 196,195,249,22,209,1 ; vpextrq $0x1,%xmm2,%r9 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,97,33,44,130,16 ; vinsertps $0x10,(%r10,%r8,4),%xmm3,%xmm13 + DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 + DB 73,193,232,30 ; shr $0x1e,%r8 + DB 196,3,97,33,12,2,16 ; vinsertps $0x10,(%r10,%r8,1),%xmm3,%xmm9 DB 69,137,200 ; mov %r9d,%r8d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 DB 196,129,122,16,28,130 ; vmovss (%r10,%r8,4),%xmm3 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 - DB 196,129,122,16,4,138 ; vmovss (%r10,%r9,4),%xmm0 + DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,129,122,16,12,10 ; vmovss (%r10,%r9,1),%xmm1 DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,30 ; shr $0x1e,%r8 DB 196,129,122,16,20,138 ; vmovss (%r10,%r9,4),%xmm2 - DB 196,195,249,22,201,1 ; vpextrq $0x1,%xmm1,%r9 - DB 196,131,105,33,12,2,16 ; vinsertps $0x10,(%r10,%r8,1),%xmm2,%xmm1 + DB 196,195,249,22,193,1 ; vpextrq $0x1,%xmm0,%r9 + DB 73,193,232,30 ; shr $0x1e,%r8 + DB 196,131,105,33,4,2,16 ; vinsertps $0x10,(%r10,%r8,1),%xmm2,%xmm0 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 196,129,122,16,20,130 ; vmovss (%r10,%r8,4),%xmm2 + DB 196,131,121,33,20,130,32 ; vinsertps $0x20,(%r10,%r8,4),%xmm0,%xmm2 DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 196,227,17,33,219,32 ; vinsertps $0x20,%xmm3,%xmm13,%xmm3 - DB 196,99,97,33,232,48 ; vinsertps $0x30,%xmm0,%xmm3,%xmm13 - DB 196,99,113,33,242,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm14 - DB 196,1,122,16,60,10 ; vmovss (%r10,%r9,1),%xmm15 - DB 196,193,25,219,210 ; vpand %xmm10,%xmm12,%xmm2 - DB 196,193,105,105,193 ; vpunpckhwd %xmm9,%xmm2,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 196,227,49,33,195,32 ; vinsertps $0x20,%xmm3,%xmm9,%xmm0 + DB 196,99,121,33,241,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm14 + DB 196,3,105,33,12,10,48 ; vinsertps $0x30,(%r10,%r9,1),%xmm2,%xmm9 + DB 196,193,17,219,203 ; vpand %xmm11,%xmm13,%xmm1 + DB 197,249,112,209,78 ; vpshufd $0x4e,%xmm1,%xmm2 + DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,3,113,33,36,136,16 ; vinsertps $0x10,(%r8,%r9,4),%xmm1,%xmm12 + DB 196,129,122,16,28,144 ; vmovss (%r8,%r10,4),%xmm3 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,226,121,51,193 ; vpmovzxwd %xmm1,%xmm0 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,3,97,33,60,8,16 ; vinsertps $0x10,(%r8,%r9,1),%xmm3,%xmm15 DB 69,137,209 ; mov %r10d,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,226,121,51,194 ; vpmovzxwd %xmm2,%xmm0 DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,129,122,16,28,144 ; vmovss (%r8,%r10,4),%xmm3 + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,129,122,16,28,16 ; vmovss (%r8,%r10,1),%xmm3 DB 69,137,202 ; mov %r9d,%r10d - DB 73,193,233,30 ; shr $0x1e,%r9 DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1 DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 - DB 196,131,113,33,12,8,16 ; vinsertps $0x10,(%r8,%r9,1),%xmm1,%xmm1 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,131,113,33,4,8,16 ; vinsertps $0x10,(%r8,%r9,1),%xmm1,%xmm0 DB 69,137,209 ; mov %r10d,%r9d - DB 196,129,122,16,60,136 ; vmovss (%r8,%r9,4),%xmm7 - DB 196,195,9,33,199,48 ; vinsertps $0x30,%xmm15,%xmm14,%xmm0 - DB 196,65,57,108,243 ; vpunpcklqdq %xmm11,%xmm8,%xmm14 - DB 196,195,125,24,197,1 ; vinsertf128 $0x1,%xmm13,%ymm0,%ymm0 DB 73,193,234,30 ; shr $0x1e,%r10 - DB 196,227,25,33,210,32 ; vinsertps $0x20,%xmm2,%xmm12,%xmm2 - DB 196,227,105,33,219,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm3 - DB 196,99,113,33,239,32 ; vinsertps $0x20,%xmm7,%xmm1,%xmm13 - DB 196,1,122,16,60,16 ; vmovss (%r8,%r10,1),%xmm15 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 196,193,9,219,250 ; vpand %xmm10,%xmm14,%xmm7 - DB 196,193,65,105,209 ; vpunpckhwd %xmm9,%xmm7,%xmm2 - DB 196,193,249,126,208 ; vmovq %xmm2,%r8 - DB 68,137,192 ; mov %r8d,%eax - DB 196,193,122,16,12,129 ; vmovss (%r9,%rax,4),%xmm1 - DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,113,33,20,129,16 ; vinsertps $0x10,(%r9,%r8,4),%xmm1,%xmm10 - DB 65,137,192 ; mov %eax,%r8d - DB 72,193,232,32 ; shr $0x20,%rax - DB 196,226,121,51,207 ; vpmovzxwd %xmm7,%xmm1 - DB 196,1,122,16,52,129 ; vmovss (%r9,%r8,4),%xmm14 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 - DB 196,65,122,16,36,129 ; vmovss (%r9,%rax,4),%xmm12 + DB 196,131,121,33,4,136,32 ; vinsertps $0x20,(%r8,%r9,4),%xmm0,%xmm0 + DB 196,3,121,33,44,16,48 ; vinsertps $0x30,(%r8,%r10,1),%xmm0,%xmm13 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 + DB 196,193,41,219,195 ; vpand %xmm11,%xmm10,%xmm0 + DB 197,249,112,200,78 ; vpshufd $0x4e,%xmm0,%xmm1 + DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 + DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 + DB 68,137,200 ; mov %r9d,%eax + DB 196,193,122,16,12,130 ; vmovss (%r10,%rax,4),%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,3,113,33,20,10,16 ; vinsertps $0x10,(%r10,%r9,1),%xmm1,%xmm10 DB 68,137,192 ; mov %r8d,%eax + DB 196,65,122,16,28,130 ; vmovss (%r10,%rax,4),%xmm11 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 DB 73,193,232,30 ; shr $0x1e,%r8 - DB 196,193,122,16,20,129 ; vmovss (%r9,%rax,4),%xmm2 - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 - DB 196,131,105,33,20,1,16 ; vinsertps $0x10,(%r9,%r8,1),%xmm2,%xmm2 - DB 68,137,208 ; mov %r10d,%eax - DB 196,193,122,16,60,129 ; vmovss (%r9,%rax,4),%xmm7 - DB 196,195,17,33,207,48 ; vinsertps $0x30,%xmm15,%xmm13,%xmm1 - DB 73,193,234,30 ; shr $0x1e,%r10 - DB 196,1,122,16,44,17 ; vmovss (%r9,%r10,1),%xmm13 - DB 196,227,117,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm1,%ymm1 - DB 196,195,41,33,222,32 ; vinsertps $0x20,%xmm14,%xmm10,%xmm3 - DB 196,195,97,33,220,48 ; vinsertps $0x30,%xmm12,%xmm3,%xmm3 - DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 - DB 196,195,105,33,213,48 ; vinsertps $0x30,%xmm13,%xmm2,%xmm2 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 - DB 196,193,57,109,219 ; vpunpckhqdq %xmm11,%xmm8,%xmm3 - DB 197,193,113,243,8 ; vpsllw $0x8,%xmm3,%xmm7 + DB 196,129,122,16,60,2 ; vmovss (%r10,%r8,1),%xmm7 + DB 68,137,200 ; mov %r9d,%eax + DB 196,193,122,16,4,130 ; vmovss (%r10,%rax,4),%xmm0 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,131,121,33,4,10,16 ; vinsertps $0x10,(%r10,%r9,1),%xmm0,%xmm0 + DB 68,137,216 ; mov %r11d,%eax + DB 196,195,121,33,4,130,32 ; vinsertps $0x20,(%r10,%rax,4),%xmm0,%xmm0 + DB 73,193,235,30 ; shr $0x1e,%r11 + DB 196,131,121,33,52,26,48 ; vinsertps $0x30,(%r10,%r11,1),%xmm0,%xmm6 + DB 196,195,53,24,198,1 ; vinsertf128 $0x1,%xmm14,%ymm9,%ymm0 + DB 196,227,1,33,202,32 ; vinsertps $0x20,%xmm2,%xmm15,%xmm1 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 196,227,21,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm13,%ymm1 + DB 196,195,41,33,211,32 ; vinsertps $0x20,%xmm11,%xmm10,%xmm2 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 + DB 196,227,77,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm6,%ymm2 + DB 196,193,57,109,220 ; vpunpckhqdq %xmm12,%xmm8,%xmm3 + DB 197,201,113,243,8 ; vpsllw $0x8,%xmm3,%xmm6 DB 197,225,113,211,8 ; vpsrlw $0x8,%xmm3,%xmm3 - DB 197,193,235,219 ; vpor %xmm3,%xmm7,%xmm3 - DB 196,193,97,105,249 ; vpunpckhwd %xmm9,%xmm3,%xmm7 + DB 197,201,235,219 ; vpor %xmm3,%xmm6,%xmm3 + DB 196,226,121,51,243 ; vpmovzxwd %xmm3,%xmm6 + DB 197,249,112,219,78 ; vpshufd $0x4e,%xmm3,%xmm3 DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 - DB 196,227,101,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm3,%ymm3 + DB 196,227,77,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm6,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,226,125,24,61,15,56,2,0 ; vbroadcastss 0x2380f(%rip),%ymm7 # 2d33c <_sk_overlay_sse2_8bit+0xde4> - DB 197,228,89,223 ; vmulps %ymm7,%ymm3,%ymm3 + DB 196,226,125,24,53,244,49,2,0 ; vbroadcastss 0x231f4(%rip),%ymm6 # 2d21c <_sk_overlay_sse2_8bit+0xd6d> + DB 197,228,89,222 ; vmulps %ymm6,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,60,36 ; vmovups (%rsp),%ymm7 - DB 72,131,196,56 ; add $0x38,%rsp + DB 197,252,16,52,36 ; vmovups (%rsp),%ymm6 + DB 197,252,16,124,36,32 ; vmovups 0x20(%rsp),%ymm7 + DB 72,131,196,88 ; add $0x58,%rsp DB 255,224 ; jmpq *%rax DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je 9ba4 <_sk_load_tables_u16_be_avx+0x2dc> + DB 116,85 ; je a0a5 <_sk_load_tables_u16_be_avx+0x2d9> DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb 9ba4 <_sk_load_tables_u16_be_avx+0x2dc> + DB 114,72 ; jb a0a5 <_sk_load_tables_u16_be_avx+0x2d9> DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je 9bb1 <_sk_load_tables_u16_be_avx+0x2e9> + DB 116,72 ; je a0b2 <_sk_load_tables_u16_be_avx+0x2e6> DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb 9bb1 <_sk_load_tables_u16_be_avx+0x2e9> + DB 114,59 ; jb a0b2 <_sk_load_tables_u16_be_avx+0x2e6> DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,123,253,255,255 ; je 9902 <_sk_load_tables_u16_be_avx+0x3a> + DB 15,132,132,253,255,255 ; je 9e0c <_sk_load_tables_u16_be_avx+0x40> DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,106,253,255,255 ; jb 9902 <_sk_load_tables_u16_be_avx+0x3a> + DB 15,130,115,253,255,255 ; jb 9e0c <_sk_load_tables_u16_be_avx+0x40> DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9 - DB 233,94,253,255,255 ; jmpq 9902 <_sk_load_tables_u16_be_avx+0x3a> + DB 233,103,253,255,255 ; jmpq 9e0c <_sk_load_tables_u16_be_avx+0x40> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,81,253,255,255 ; jmpq 9902 <_sk_load_tables_u16_be_avx+0x3a> + DB 233,90,253,255,255 ; jmpq 9e0c <_sk_load_tables_u16_be_avx+0x40> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,72,253,255,255 ; jmpq 9902 <_sk_load_tables_u16_be_avx+0x3a> + DB 233,81,253,255,255 ; jmpq 9e0c <_sk_load_tables_u16_be_avx+0x40> PUBLIC _sk_load_tables_rgb_u16_be_avx _sk_load_tables_rgb_u16_be_avx LABEL PROC @@ -9314,7 +9520,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 72,133,255 ; test %rdi,%rdi DB 197,252,17,124,36,32 ; vmovups %ymm7,0x20(%rsp) DB 197,252,17,52,36 ; vmovups %ymm6,(%rsp) - DB 15,133,75,2,0,0 ; jne 9e26 <_sk_load_tables_rgb_u16_be_avx+0x26c> + DB 15,133,57,2,0,0 ; jne a315 <_sk_load_tables_rgb_u16_be_avx+0x25a> DB 196,1,122,111,28,72 ; vmovdqu (%r8,%r9,2),%xmm11 DB 196,129,122,111,92,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm3 DB 196,129,122,111,84,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm2 @@ -9322,189 +9528,187 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 197,249,115,216,4 ; vpsrldq $0x4,%xmm0,%xmm0 DB 196,193,57,115,219,6 ; vpsrldq $0x6,%xmm11,%xmm8 DB 197,169,115,219,6 ; vpsrldq $0x6,%xmm3,%xmm10 - DB 197,153,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm12 + DB 197,241,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm1 DB 197,177,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm9 - DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 - DB 196,193,57,97,202 ; vpunpcklwd %xmm10,%xmm8,%xmm1 + DB 196,65,113,97,201 ; vpunpcklwd %xmm9,%xmm1,%xmm9 DB 197,233,97,192 ; vpunpcklwd %xmm0,%xmm2,%xmm0 - DB 196,193,25,97,209 ; vpunpcklwd %xmm9,%xmm12,%xmm2 - DB 197,97,97,193 ; vpunpcklwd %xmm1,%xmm3,%xmm8 - DB 197,225,105,201 ; vpunpckhwd %xmm1,%xmm3,%xmm1 - DB 197,249,97,218 ; vpunpcklwd %xmm2,%xmm0,%xmm3 - DB 197,249,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm0 - DB 197,185,108,211 ; vpunpcklqdq %xmm3,%xmm8,%xmm2 - DB 197,57,109,219 ; vpunpckhqdq %xmm3,%xmm8,%xmm11 - DB 197,113,108,192 ; vpunpcklqdq %xmm0,%xmm1,%xmm8 - DB 197,121,111,13,10,56,2,0 ; vmovdqa 0x2380a(%rip),%xmm9 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 196,193,105,219,193 ; vpand %xmm9,%xmm2,%xmm0 - DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 + DB 196,193,57,97,210 ; vpunpcklwd %xmm10,%xmm8,%xmm2 + DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 + DB 197,225,97,202 ; vpunpcklwd %xmm2,%xmm3,%xmm1 + DB 197,225,105,210 ; vpunpckhwd %xmm2,%xmm3,%xmm2 + DB 196,193,121,97,217 ; vpunpcklwd %xmm9,%xmm0,%xmm3 + DB 196,193,121,105,193 ; vpunpckhwd %xmm9,%xmm0,%xmm0 + DB 197,105,108,216 ; vpunpcklqdq %xmm0,%xmm2,%xmm11 + DB 197,241,108,211 ; vpunpcklqdq %xmm3,%xmm1,%xmm2 + DB 197,241,109,203 ; vpunpckhqdq %xmm3,%xmm1,%xmm1 + DB 197,121,111,53,247,49,2,0 ; vmovdqa 0x231f7(%rip),%xmm14 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,193,105,219,214 ; vpand %xmm14,%xmm2,%xmm2 + DB 197,249,112,218,78 ; vpshufd $0x4e,%xmm2,%xmm3 + DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 + DB 196,193,249,126,216 ; vmovq %xmm3,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 76,139,80,8 ; mov 0x8(%rax),%r10 - DB 196,129,122,16,20,138 ; vmovss (%r10,%r9,4),%xmm2 - DB 196,195,249,22,201,1 ; vpextrq $0x1,%xmm1,%r9 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,105,33,36,130,16 ; vinsertps $0x10,(%r10,%r8,4),%xmm2,%xmm12 + DB 196,129,122,16,4,138 ; vmovss (%r10,%r9,4),%xmm0 + DB 196,195,249,22,217,1 ; vpextrq $0x1,%xmm3,%r9 + DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 + DB 73,193,232,30 ; shr $0x1e,%r8 + DB 196,3,121,33,4,2,16 ; vinsertps $0x10,(%r10,%r8,1),%xmm0,%xmm8 DB 69,137,200 ; mov %r9d,%r8d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,129,122,16,20,130 ; vmovss (%r10,%r8,4),%xmm2 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 - DB 196,129,122,16,12,138 ; vmovss (%r10,%r9,4),%xmm1 + DB 196,1,122,16,20,130 ; vmovss (%r10,%r8,4),%xmm10 + DB 196,193,249,126,208 ; vmovq %xmm2,%r8 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,1,122,16,36,10 ; vmovss (%r10,%r9,1),%xmm12 DB 69,137,193 ; mov %r8d,%r9d + DB 196,129,122,16,4,138 ; vmovss (%r10,%r9,4),%xmm0 + DB 196,195,249,22,209,1 ; vpextrq $0x1,%xmm2,%r9 DB 73,193,232,30 ; shr $0x1e,%r8 - DB 196,129,122,16,28,138 ; vmovss (%r10,%r9,4),%xmm3 - DB 196,195,249,22,193,1 ; vpextrq $0x1,%xmm0,%r9 - DB 196,131,97,33,28,2,16 ; vinsertps $0x10,(%r10,%r8,1),%xmm3,%xmm3 + DB 196,131,121,33,4,2,16 ; vinsertps $0x10,(%r10,%r8,1),%xmm0,%xmm0 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 196,129,122,16,4,130 ; vmovss (%r10,%r8,4),%xmm0 + DB 196,131,121,33,4,130,32 ; vinsertps $0x20,(%r10,%r8,4),%xmm0,%xmm0 DB 76,139,64,16 ; mov 0x10(%rax),%r8 - DB 196,227,25,33,210,32 ; vinsertps $0x20,%xmm2,%xmm12,%xmm2 - DB 196,227,105,33,201,48 ; vinsertps $0x30,%xmm1,%xmm2,%xmm1 - DB 196,129,122,16,20,10 ; vmovss (%r10,%r9,1),%xmm2 - DB 196,65,33,219,225 ; vpand %xmm9,%xmm11,%xmm12 - DB 196,65,25,105,218 ; vpunpckhwd %xmm10,%xmm12,%xmm11 - DB 196,65,249,126,217 ; vmovq %xmm11,%r9 + DB 196,3,121,33,12,10,48 ; vinsertps $0x30,(%r10,%r9,1),%xmm0,%xmm9 + DB 196,193,113,219,198 ; vpand %xmm14,%xmm1,%xmm0 + DB 197,249,112,200,78 ; vpshufd $0x4e,%xmm0,%xmm1 + DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,1,122,16,44,144 ; vmovss (%r8,%r10,4),%xmm13 - DB 196,67,249,22,218,1 ; vpextrq $0x1,%xmm11,%r10 - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,3,17,33,28,136,16 ; vinsertps $0x10,(%r8,%r9,4),%xmm13,%xmm11 + DB 196,129,122,16,20,144 ; vmovss (%r8,%r10,4),%xmm2 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,3,105,33,60,8,16 ; vinsertps $0x10,(%r8,%r9,1),%xmm2,%xmm15 DB 69,137,209 ; mov %r10d,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,66,121,51,244 ; vpmovzxwd %xmm12,%xmm14 - DB 196,1,122,16,44,136 ; vmovss (%r8,%r9,4),%xmm13 - DB 196,65,249,126,241 ; vmovq %xmm14,%r9 - DB 196,1,122,16,36,144 ; vmovss (%r8,%r10,4),%xmm12 + DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,129,122,16,28,16 ; vmovss (%r8,%r10,1),%xmm3 DB 69,137,202 ; mov %r9d,%r10d + DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 DB 73,193,233,30 ; shr $0x1e,%r9 - DB 196,1,122,16,60,144 ; vmovss (%r8,%r10,4),%xmm15 - DB 196,67,249,22,242,1 ; vpextrq $0x1,%xmm14,%r10 - DB 196,3,1,33,52,8,16 ; vinsertps $0x10,(%r8,%r9,1),%xmm15,%xmm14 + DB 196,131,113,33,4,8,16 ; vinsertps $0x10,(%r8,%r9,1),%xmm1,%xmm0 DB 69,137,209 ; mov %r10d,%r9d - DB 196,1,122,16,60,136 ; vmovss (%r8,%r9,4),%xmm15 - DB 196,227,97,33,192,32 ; vinsertps $0x20,%xmm0,%xmm3,%xmm0 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 73,193,234,30 ; shr $0x1e,%r10 - DB 196,129,122,16,52,16 ; vmovss (%r8,%r10,1),%xmm6 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 196,65,57,219,193 ; vpand %xmm9,%xmm8,%xmm8 - DB 196,193,57,105,210 ; vpunpckhwd %xmm10,%xmm8,%xmm2 - DB 196,193,249,126,208 ; vmovq %xmm2,%r8 - DB 68,137,192 ; mov %r8d,%eax - DB 196,193,122,16,12,129 ; vmovss (%r9,%rax,4),%xmm1 - DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,113,33,12,129,16 ; vinsertps $0x10,(%r9,%r8,4),%xmm1,%xmm9 - DB 65,137,192 ; mov %eax,%r8d - DB 72,193,232,32 ; shr $0x20,%rax - DB 196,194,121,51,200 ; vpmovzxwd %xmm8,%xmm1 - DB 196,1,122,16,4,129 ; vmovss (%r9,%r8,4),%xmm8 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 - DB 196,65,122,16,20,129 ; vmovss (%r9,%rax,4),%xmm10 + DB 196,131,121,33,4,136,32 ; vinsertps $0x20,(%r8,%r9,4),%xmm0,%xmm0 + DB 196,3,121,33,44,16,48 ; vinsertps $0x30,(%r8,%r10,1),%xmm0,%xmm13 + DB 76,139,80,24 ; mov 0x18(%rax),%r10 + DB 196,193,33,219,198 ; vpand %xmm14,%xmm11,%xmm0 + DB 197,249,112,200,78 ; vpshufd $0x4e,%xmm0,%xmm1 + DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 + DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 + DB 68,137,200 ; mov %r9d,%eax + DB 196,193,122,16,12,130 ; vmovss (%r10,%rax,4),%xmm1 + DB 196,195,249,22,195,1 ; vpextrq $0x1,%xmm0,%r11 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,3,113,33,28,10,16 ; vinsertps $0x10,(%r10,%r9,1),%xmm1,%xmm11 DB 68,137,192 ; mov %r8d,%eax + DB 196,65,122,16,52,130 ; vmovss (%r10,%rax,4),%xmm14 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 DB 73,193,232,30 ; shr $0x1e,%r8 - DB 196,193,122,16,20,129 ; vmovss (%r9,%rax,4),%xmm2 - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 - DB 196,131,105,33,20,1,16 ; vinsertps $0x10,(%r9,%r8,1),%xmm2,%xmm2 - DB 68,137,208 ; mov %r10d,%eax - DB 196,193,122,16,60,129 ; vmovss (%r9,%rax,4),%xmm7 - DB 196,195,33,33,205,32 ; vinsertps $0x20,%xmm13,%xmm11,%xmm1 - DB 73,193,234,30 ; shr $0x1e,%r10 - DB 196,1,122,16,28,17 ; vmovss (%r9,%r10,1),%xmm11 - DB 196,195,113,33,204,48 ; vinsertps $0x30,%xmm12,%xmm1,%xmm1 - DB 196,195,9,33,223,32 ; vinsertps $0x20,%xmm15,%xmm14,%xmm3 - DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 - DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 - DB 196,195,49,33,216,32 ; vinsertps $0x20,%xmm8,%xmm9,%xmm3 - DB 196,195,97,33,218,48 ; vinsertps $0x30,%xmm10,%xmm3,%xmm3 - DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 - DB 196,195,105,33,211,48 ; vinsertps $0x30,%xmm11,%xmm2,%xmm2 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 + DB 196,129,122,16,60,2 ; vmovss (%r10,%r8,1),%xmm7 + DB 68,137,200 ; mov %r9d,%eax + DB 196,193,122,16,4,130 ; vmovss (%r10,%rax,4),%xmm0 + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,131,121,33,4,10,16 ; vinsertps $0x10,(%r10,%r9,1),%xmm0,%xmm0 + DB 68,137,216 ; mov %r11d,%eax + DB 196,195,121,33,4,130,32 ; vinsertps $0x20,(%r10,%rax,4),%xmm0,%xmm0 + DB 73,193,235,30 ; shr $0x1e,%r11 + DB 196,131,121,33,52,26,48 ; vinsertps $0x30,(%r10,%r11,1),%xmm0,%xmm6 + DB 196,195,57,33,194,32 ; vinsertps $0x20,%xmm10,%xmm8,%xmm0 + DB 196,195,121,33,196,48 ; vinsertps $0x30,%xmm12,%xmm0,%xmm0 + DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 + DB 196,227,1,33,202,32 ; vinsertps $0x20,%xmm2,%xmm15,%xmm1 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 196,227,21,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm13,%ymm1 + DB 196,195,33,33,214,32 ; vinsertps $0x20,%xmm14,%xmm11,%xmm2 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 + DB 196,227,77,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm6,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,151,52,2,0 ; vbroadcastss 0x23497(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,136,46,2,0 ; vbroadcastss 0x22e88(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,16,52,36 ; vmovups (%rsp),%ymm6 DB 197,252,16,124,36,32 ; vmovups 0x20(%rsp),%ymm7 DB 72,131,196,88 ; add $0x58,%rsp DB 255,224 ; jmpq *%rax DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0 - DB 196,1,121,196,92,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm11 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,1,121,196,92,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm11 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,32 ; jne 9e5f <_sk_load_tables_rgb_u16_be_avx+0x2a5> + DB 117,31 ; jne a34d <_sk_load_tables_rgb_u16_be_avx+0x292> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 233,177,253,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 233,196,253,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> DB 196,129,121,110,68,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm0 - DB 196,1,121,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm0,%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,1,121,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm0,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,49 ; jb 9eaa <_sk_load_tables_rgb_u16_be_avx+0x2f0> + DB 114,48 ; jb a397 <_sk_load_tables_rgb_u16_be_avx+0x2dc> DB 196,129,121,110,68,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm0 - DB 196,129,121,196,92,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm0,%xmm3 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,129,121,196,92,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm0,%xmm3 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 117,50 ; jne 9ec5 <_sk_load_tables_rgb_u16_be_avx+0x30b> + DB 117,48 ; jne a3b1 <_sk_load_tables_rgb_u16_be_avx+0x2f6> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 233,102,253,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 233,122,253,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 - DB 233,75,253,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 233,96,253,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> DB 196,129,121,110,68,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm0 - DB 196,1,121,196,84,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm0,%xmm10 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,1,121,196,84,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm0,%xmm10 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,40 ; jb 9f07 <_sk_load_tables_rgb_u16_be_avx+0x34d> + DB 114,39 ; jb a3f2 <_sk_load_tables_rgb_u16_be_avx+0x337> DB 196,129,121,110,68,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm0 - DB 196,129,121,196,84,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm0,%xmm2 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,129,121,196,84,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm0,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 117,32 ; jne 9f19 <_sk_load_tables_rgb_u16_be_avx+0x35f> + DB 117,30 ; jne a403 <_sk_load_tables_rgb_u16_be_avx+0x348> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 - DB 233,9,253,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 + DB 233,31,253,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 - DB 233,247,252,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 233,14,253,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> DB 196,129,121,110,68,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm0 - DB 196,1,121,196,100,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm0,%xmm12 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,129,121,196,76,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm0,%xmm1 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,25 ; jb 9f4c <_sk_load_tables_rgb_u16_be_avx+0x392> + DB 114,25 ; jb a436 <_sk_load_tables_rgb_u16_be_avx+0x37b> DB 196,129,121,110,68,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm0 - DB 196,129,121,196,68,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm0,%xmm0 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 233,196,252,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 196,129,121,196,68,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm0,%xmm0 + DB 233,219,252,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 233,187,252,255,255 ; jmpq 9c10 <_sk_load_tables_rgb_u16_be_avx+0x56> + DB 233,210,252,255,255 ; jmpq a111 <_sk_load_tables_rgb_u16_be_avx+0x56> PUBLIC _sk_byte_tables_avx _sk_byte_tables_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,120,51,2,0 ; vbroadcastss 0x23378(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,110,45,2,0 ; vbroadcastss 0x22d6e(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,125,91,200 ; vcvtps2dq %ymm0,%ymm9 DB 196,65,249,126,200 ; vmovq %xmm9,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 76,139,16 ; mov (%rax),%r10 - DB 196,131,121,32,4,10,0 ; vpinsrb $0x0,(%r10,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,193,121,110,193 ; vmovd %r9d,%xmm0 DB 196,67,249,22,201,1 ; vpextrq $0x1,%xmm9,%r9 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,121,32,20,2,1 ; vpinsrb $0x1,(%r10,%r8,1),%xmm0,%xmm10 + DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d + DB 196,67,121,32,208,1 ; vpinsrb $0x1,%r8d,%xmm0,%xmm10 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,32 ; shr $0x20,%r9 DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 @@ -9514,10 +9718,12 @@ _sk_byte_tables_avx LABEL PROC DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d DB 196,67,49,32,201,3 ; vpinsrb $0x3,%r9d,%xmm9,%xmm9 DB 69,137,193 ; mov %r8d,%r9d - DB 196,3,121,32,20,10,0 ; vpinsrb $0x0,(%r10,%r9,1),%xmm0,%xmm10 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,65,121,110,209 ; vmovd %r9d,%xmm10 DB 196,195,249,22,193,1 ; vpextrq $0x1,%xmm0,%r9 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,41,32,4,2,1 ; vpinsrb $0x1,(%r10,%r8,1),%xmm10,%xmm0 + DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d + DB 196,195,41,32,192,1 ; vpinsrb $0x1,%r8d,%xmm10,%xmm0 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,32 ; shr $0x20,%r9 DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d @@ -9529,10 +9735,12 @@ _sk_byte_tables_avx LABEL PROC DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,202 ; vmovd %r10d,%xmm1 DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,113,32,12,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm1,%xmm1 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,113,32,201,1 ; vpinsrb $0x1,%r9d,%xmm1,%xmm1 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 @@ -9542,10 +9750,12 @@ _sk_byte_tables_avx LABEL PROC DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d DB 196,67,113,32,226,3 ; vpinsrb $0x3,%r10d,%xmm1,%xmm12 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,202 ; vmovd %r10d,%xmm1 DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,113,32,4,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm1,%xmm0 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,113,32,193,1 ; vpinsrb $0x1,%r9d,%xmm1,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d @@ -9554,26 +9764,30 @@ _sk_byte_tables_avx LABEL PROC DB 196,67,121,32,232,3 ; vpinsrb $0x3,%r8d,%xmm0,%xmm13 DB 76,139,64,16 ; mov 0x10(%rax),%r8 DB 196,193,108,89,200 ; vmulps %ymm8,%ymm2,%ymm1 - DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 - DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 197,253,91,209 ; vcvtps2dq %ymm1,%ymm2 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,202 ; vmovd %r10d,%xmm1 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,105,32,20,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm2,%xmm2 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,113,32,201,1 ; vpinsrb $0x1,%r9d,%xmm1,%xmm1 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 + DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 196,195,105,32,209,2 ; vpinsrb $0x2,%r9d,%xmm2,%xmm2 - DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,195,113,32,201,2 ; vpinsrb $0x2,%r9d,%xmm1,%xmm1 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d - DB 196,67,105,32,218,3 ; vpinsrb $0x3,%r10d,%xmm2,%xmm11 + DB 196,67,113,32,218,3 ; vpinsrb $0x3,%r10d,%xmm1,%xmm11 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,202 ; vmovd %r10d,%xmm1 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,105,32,12,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm2,%xmm1 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,113,32,201,1 ; vpinsrb $0x1,%r9d,%xmm1,%xmm1 DB 69,137,209 ; mov %r10d,%r9d DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 196,195,113,32,201,2 ; vpinsrb $0x2,%r9d,%xmm1,%xmm1 @@ -9582,39 +9796,43 @@ _sk_byte_tables_avx LABEL PROC DB 196,67,113,32,240,3 ; vpinsrb $0x3,%r8d,%xmm1,%xmm14 DB 76,139,72,24 ; mov 0x18(%rax),%r9 DB 196,193,100,89,200 ; vmulps %ymm8,%ymm3,%ymm1 - DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 + DB 197,253,91,217 ; vcvtps2dq %ymm1,%ymm3 + DB 196,193,249,126,216 ; vmovq %xmm3,%r8 DB 68,137,192 ; mov %r8d,%eax - DB 196,195,121,32,28,1,0 ; vpinsrb $0x0,(%r9,%rax,1),%xmm0,%xmm3 - DB 196,227,249,22,200,1 ; vpextrq $0x1,%xmm1,%rax + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 197,249,110,200 ; vmovd %eax,%xmm1 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,97,32,28,1,1 ; vpinsrb $0x1,(%r9,%r8,1),%xmm3,%xmm3 - DB 65,137,192 ; mov %eax,%r8d - DB 72,193,232,32 ; shr $0x20,%rax - DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 - DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d - DB 196,195,97,32,216,2 ; vpinsrb $0x2,%r8d,%xmm3,%xmm3 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 + DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 196,227,113,32,200,1 ; vpinsrb $0x1,%eax,%xmm1,%xmm1 + DB 68,137,208 ; mov %r10d,%eax + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,227,125,25,219,1 ; vextractf128 $0x1,%ymm3,%xmm3 DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,99,97,32,192,3 ; vpinsrb $0x3,%eax,%xmm3,%xmm8 + DB 196,227,113,32,200,2 ; vpinsrb $0x2,%eax,%xmm1,%xmm1 + DB 196,193,249,126,216 ; vmovq %xmm3,%r8 + DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax + DB 196,99,113,32,192,3 ; vpinsrb $0x3,%eax,%xmm1,%xmm8 DB 68,137,192 ; mov %r8d,%eax - DB 196,195,121,32,4,1,0 ; vpinsrb $0x0,(%r9,%rax,1),%xmm0,%xmm0 - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 197,249,110,200 ; vmovd %eax,%xmm1 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,121,32,4,1,1 ; vpinsrb $0x1,(%r9,%r8,1),%xmm0,%xmm0 + DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 196,227,113,32,200,1 ; vpinsrb $0x1,%eax,%xmm1,%xmm1 DB 68,137,208 ; mov %r10d,%eax DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,99,121,32,248,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm15 + DB 196,99,113,32,248,2 ; vpinsrb $0x2,%eax,%xmm1,%xmm15 DB 73,193,234,32 ; shr $0x20,%r10 DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax - DB 196,194,121,49,193 ; vpmovzxbd %xmm9,%xmm0 - DB 196,194,121,49,202 ; vpmovzxbd %xmm10,%xmm1 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 + DB 196,194,121,49,201 ; vpmovzxbd %xmm9,%xmm1 + DB 196,194,121,49,194 ; vpmovzxbd %xmm10,%xmm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 DB 196,194,121,49,204 ; vpmovzxbd %xmm12,%xmm1 DB 196,194,121,49,213 ; vpmovzxbd %xmm13,%xmm2 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,98,125,24,13,102,49,2,0 ; vbroadcastss 0x23166(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,38,43,2,0 ; vbroadcastss 0x22b26(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 @@ -9646,10 +9864,12 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 196,65,249,126,200 ; vmovq %xmm9,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 76,139,16 ; mov (%rax),%r10 - DB 196,131,121,32,4,10,0 ; vpinsrb $0x0,(%r10,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,193,121,110,193 ; vmovd %r9d,%xmm0 DB 196,67,249,22,201,1 ; vpextrq $0x1,%xmm9,%r9 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,3,121,32,20,2,1 ; vpinsrb $0x1,(%r10,%r8,1),%xmm0,%xmm10 + DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d + DB 196,67,121,32,208,1 ; vpinsrb $0x1,%r8d,%xmm0,%xmm10 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,32 ; shr $0x20,%r9 DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 @@ -9659,10 +9879,12 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d DB 196,67,49,32,201,3 ; vpinsrb $0x3,%r9d,%xmm9,%xmm9 DB 69,137,193 ; mov %r8d,%r9d - DB 196,3,121,32,20,10,0 ; vpinsrb $0x0,(%r10,%r9,1),%xmm0,%xmm10 + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 196,65,121,110,209 ; vmovd %r9d,%xmm10 DB 196,195,249,22,193,1 ; vpextrq $0x1,%xmm0,%r9 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,41,32,4,2,1 ; vpinsrb $0x1,(%r10,%r8,1),%xmm10,%xmm0 + DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d + DB 196,195,41,32,192,1 ; vpinsrb $0x1,%r8d,%xmm10,%xmm0 DB 69,137,200 ; mov %r9d,%r8d DB 73,193,233,32 ; shr $0x20,%r9 DB 71,15,182,4,2 ; movzbl (%r10,%r8,1),%r8d @@ -9674,10 +9896,12 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,193,249,126,201 ; vmovq %xmm1,%r9 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,4,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,194 ; vmovd %r10d,%xmm0 DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,121,32,4,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,121,32,193,1 ; vpinsrb $0x1,%r9d,%xmm0,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 73,193,234,32 ; shr $0x20,%r10 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 @@ -9687,10 +9911,12 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d DB 196,67,121,32,210,3 ; vpinsrb $0x3,%r10d,%xmm0,%xmm10 DB 69,137,202 ; mov %r9d,%r10d - DB 196,131,121,32,4,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 196,193,121,110,194 ; vmovd %r10d,%xmm0 DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,131,121,32,4,8,1 ; vpinsrb $0x1,(%r8,%r9,1),%xmm0,%xmm0 + DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d + DB 196,195,121,32,193,1 ; vpinsrb $0x1,%r9d,%xmm0,%xmm0 DB 69,137,209 ; mov %r10d,%r9d DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 196,195,121,32,193,2 ; vpinsrb $0x2,%r9d,%xmm0,%xmm0 @@ -9699,26 +9925,30 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 196,67,121,32,224,3 ; vpinsrb $0x3,%r8d,%xmm0,%xmm12 DB 76,139,72,16 ; mov 0x10(%rax),%r9 DB 197,188,89,194 ; vmulps %ymm2,%ymm8,%ymm0 - DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 197,253,91,208 ; vcvtps2dq %ymm0,%ymm2 + DB 196,193,249,126,208 ; vmovq %xmm2,%r8 DB 68,137,192 ; mov %r8d,%eax - DB 196,195,121,32,20,1,0 ; vpinsrb $0x0,(%r9,%rax,1),%xmm0,%xmm2 - DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 197,249,110,192 ; vmovd %eax,%xmm0 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,105,32,20,1,1 ; vpinsrb $0x1,(%r9,%r8,1),%xmm2,%xmm2 - DB 65,137,192 ; mov %eax,%r8d - DB 72,193,232,32 ; shr $0x20,%rax - DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d - DB 196,195,105,32,208,2 ; vpinsrb $0x2,%r8d,%xmm2,%xmm2 - DB 196,193,249,126,192 ; vmovq %xmm0,%r8 + DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 196,227,121,32,192,1 ; vpinsrb $0x1,%eax,%xmm0,%xmm0 + DB 68,137,208 ; mov %r10d,%eax + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax - DB 196,99,105,32,192,3 ; vpinsrb $0x3,%eax,%xmm2,%xmm8 + DB 196,227,121,32,192,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm0 + DB 196,193,249,126,208 ; vmovq %xmm2,%r8 + DB 67,15,182,4,17 ; movzbl (%r9,%r10,1),%eax + DB 196,99,121,32,192,3 ; vpinsrb $0x3,%eax,%xmm0,%xmm8 DB 68,137,192 ; mov %r8d,%eax - DB 196,195,121,32,12,1,0 ; vpinsrb $0x0,(%r9,%rax,1),%xmm0,%xmm1 - DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax + DB 197,249,110,192 ; vmovd %eax,%xmm0 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,131,113,32,4,1,1 ; vpinsrb $0x1,(%r9,%r8,1),%xmm1,%xmm0 + DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax + DB 196,227,121,32,192,1 ; vpinsrb $0x1,%eax,%xmm0,%xmm0 DB 68,137,208 ; mov %r10d,%eax DB 65,15,182,4,1 ; movzbl (%r9,%rax,1),%eax DB 196,99,121,32,232,2 ; vpinsrb $0x2,%eax,%xmm0,%xmm13 @@ -9728,7 +9958,7 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 196,194,121,49,203 ; vpmovzxbd %xmm11,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,67,47,2,0 ; vbroadcastss 0x22f43(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,219,40,2,0 ; vbroadcastss 0x228db(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,194,121,49,202 ; vpmovzxbd %xmm10,%xmm1 DB 196,194,121,49,212 ; vpmovzxbd %xmm12,%xmm2 @@ -9955,44 +10185,44 @@ _sk_parametric_r_avx LABEL PROC DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,216 ; vcvtdq2ps %ymm0,%ymm11 - DB 196,98,125,24,37,212,43,2,0 ; vbroadcastss 0x22bd4(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> + DB 196,98,125,24,37,108,37,2,0 ; vbroadcastss 0x2256c(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,202,43,2,0 ; vbroadcastss 0x22bca(%rip),%ymm12 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,98,37,2,0 ; vbroadcastss 0x22562(%rip),%ymm12 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,32,43,2,0 ; vbroadcastss 0x22b20(%rip),%ymm12 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,37,184,36,2,0 ; vbroadcastss 0x224b8(%rip),%ymm12 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,178,43,2,0 ; vbroadcastss 0x22bb2(%rip),%ymm12 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,37,74,37,2,0 ; vbroadcastss 0x2254a(%rip),%ymm12 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,168,43,2,0 ; vbroadcastss 0x22ba8(%rip),%ymm12 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,37,64,37,2,0 ; vbroadcastss 0x22540(%rip),%ymm12 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,153,43,2,0 ; vbroadcastss 0x22b99(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,49,37,2,0 ; vbroadcastss 0x22531(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,143,43,2,0 ; vbroadcastss 0x22b8f(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,39,37,2,0 ; vbroadcastss 0x22527(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,115,43,2,0 ; vbroadcastss 0x22b73(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,11,37,2,0 ; vbroadcastss 0x2250b(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,105,43,2,0 ; vbroadcastss 0x22b69(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,1,37,2,0 ; vbroadcastss 0x22501(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,90,43,2,0 ; vbroadcastss 0x22b5a(%rip),%ymm11 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,98,125,24,29,242,36,2,0 ; vbroadcastss 0x224f2(%rip),%ymm11 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,80,43,2,0 ; vbroadcastss 0x22b50(%rip),%ymm11 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,29,232,36,2,0 ; vbroadcastss 0x224e8(%rip),%ymm11 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,65,43,2,0 ; vbroadcastss 0x22b41(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,217,36,2,0 ; vbroadcastss 0x224d9(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 + DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,88,42,2,0 ; vbroadcastss 0x22a58(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,241,35,2,0 ; vbroadcastss 0x223f1(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10012,44 +10242,44 @@ _sk_parametric_g_avx LABEL PROC DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,217 ; vcvtdq2ps %ymm1,%ymm11 - DB 196,98,125,24,37,153,42,2,0 ; vbroadcastss 0x22a99(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> + DB 196,98,125,24,37,50,36,2,0 ; vbroadcastss 0x22432(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,143,42,2,0 ; vbroadcastss 0x22a8f(%rip),%ymm12 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,40,36,2,0 ; vbroadcastss 0x22428(%rip),%ymm12 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,229,41,2,0 ; vbroadcastss 0x229e5(%rip),%ymm12 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,37,126,35,2,0 ; vbroadcastss 0x2237e(%rip),%ymm12 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,119,42,2,0 ; vbroadcastss 0x22a77(%rip),%ymm12 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,37,16,36,2,0 ; vbroadcastss 0x22410(%rip),%ymm12 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,109,42,2,0 ; vbroadcastss 0x22a6d(%rip),%ymm12 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,37,6,36,2,0 ; vbroadcastss 0x22406(%rip),%ymm12 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,94,42,2,0 ; vbroadcastss 0x22a5e(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,247,35,2,0 ; vbroadcastss 0x223f7(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,84,42,2,0 ; vbroadcastss 0x22a54(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,237,35,2,0 ; vbroadcastss 0x223ed(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1 DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,56,42,2,0 ; vbroadcastss 0x22a38(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,209,35,2,0 ; vbroadcastss 0x223d1(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,46,42,2,0 ; vbroadcastss 0x22a2e(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,199,35,2,0 ; vbroadcastss 0x223c7(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,31,42,2,0 ; vbroadcastss 0x22a1f(%rip),%ymm11 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,98,125,24,29,184,35,2,0 ; vbroadcastss 0x223b8(%rip),%ymm11 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,21,42,2,0 ; vbroadcastss 0x22a15(%rip),%ymm11 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,29,174,35,2,0 ; vbroadcastss 0x223ae(%rip),%ymm11 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,6,42,2,0 ; vbroadcastss 0x22a06(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,159,35,2,0 ; vbroadcastss 0x2239f(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 + DB 197,172,88,201 ; vaddps %ymm1,%ymm10,%ymm1 DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,29,41,2,0 ; vbroadcastss 0x2291d(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,183,34,2,0 ; vbroadcastss 0x222b7(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10069,44 +10299,44 @@ _sk_parametric_b_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,218 ; vcvtdq2ps %ymm2,%ymm11 - DB 196,98,125,24,37,94,41,2,0 ; vbroadcastss 0x2295e(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> + DB 196,98,125,24,37,248,34,2,0 ; vbroadcastss 0x222f8(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,84,41,2,0 ; vbroadcastss 0x22954(%rip),%ymm12 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,238,34,2,0 ; vbroadcastss 0x222ee(%rip),%ymm12 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,170,40,2,0 ; vbroadcastss 0x228aa(%rip),%ymm12 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,37,68,34,2,0 ; vbroadcastss 0x22244(%rip),%ymm12 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,60,41,2,0 ; vbroadcastss 0x2293c(%rip),%ymm12 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,37,214,34,2,0 ; vbroadcastss 0x222d6(%rip),%ymm12 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,50,41,2,0 ; vbroadcastss 0x22932(%rip),%ymm12 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,37,204,34,2,0 ; vbroadcastss 0x222cc(%rip),%ymm12 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,35,41,2,0 ; vbroadcastss 0x22923(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,189,34,2,0 ; vbroadcastss 0x222bd(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,25,41,2,0 ; vbroadcastss 0x22919(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,179,34,2,0 ; vbroadcastss 0x222b3(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2 DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,253,40,2,0 ; vbroadcastss 0x228fd(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,151,34,2,0 ; vbroadcastss 0x22297(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,243,40,2,0 ; vbroadcastss 0x228f3(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,141,34,2,0 ; vbroadcastss 0x2228d(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,228,40,2,0 ; vbroadcastss 0x228e4(%rip),%ymm11 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,98,125,24,29,126,34,2,0 ; vbroadcastss 0x2227e(%rip),%ymm11 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,218,40,2,0 ; vbroadcastss 0x228da(%rip),%ymm11 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,29,116,34,2,0 ; vbroadcastss 0x22274(%rip),%ymm11 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,203,40,2,0 ; vbroadcastss 0x228cb(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,101,34,2,0 ; vbroadcastss 0x22265(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 + DB 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,226,39,2,0 ; vbroadcastss 0x227e2(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,125,33,2,0 ; vbroadcastss 0x2217d(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10126,44 +10356,44 @@ _sk_parametric_a_avx LABEL PROC DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,219 ; vcvtdq2ps %ymm3,%ymm11 - DB 196,98,125,24,37,35,40,2,0 ; vbroadcastss 0x22823(%rip),%ymm12 # 2d340 <_sk_overlay_sse2_8bit+0xde8> + DB 196,98,125,24,37,190,33,2,0 ; vbroadcastss 0x221be(%rip),%ymm12 # 2d220 <_sk_overlay_sse2_8bit+0xd71> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,25,40,2,0 ; vbroadcastss 0x22819(%rip),%ymm12 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,37,180,33,2,0 ; vbroadcastss 0x221b4(%rip),%ymm12 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,111,39,2,0 ; vbroadcastss 0x2276f(%rip),%ymm12 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,37,10,33,2,0 ; vbroadcastss 0x2210a(%rip),%ymm12 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,1,40,2,0 ; vbroadcastss 0x22801(%rip),%ymm12 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,37,156,33,2,0 ; vbroadcastss 0x2219c(%rip),%ymm12 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,247,39,2,0 ; vbroadcastss 0x227f7(%rip),%ymm12 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,37,146,33,2,0 ; vbroadcastss 0x22192(%rip),%ymm12 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,232,39,2,0 ; vbroadcastss 0x227e8(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,131,33,2,0 ; vbroadcastss 0x22183(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,222,39,2,0 ; vbroadcastss 0x227de(%rip),%ymm12 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,37,121,33,2,0 ; vbroadcastss 0x22179(%rip),%ymm12 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3 DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,194,39,2,0 ; vbroadcastss 0x227c2(%rip),%ymm11 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,29,93,33,2,0 ; vbroadcastss 0x2215d(%rip),%ymm11 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,184,39,2,0 ; vbroadcastss 0x227b8(%rip),%ymm11 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,98,125,24,29,83,33,2,0 ; vbroadcastss 0x22153(%rip),%ymm11 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,169,39,2,0 ; vbroadcastss 0x227a9(%rip),%ymm11 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,98,125,24,29,68,33,2,0 ; vbroadcastss 0x22144(%rip),%ymm11 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,159,39,2,0 ; vbroadcastss 0x2279f(%rip),%ymm11 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,98,125,24,29,58,33,2,0 ; vbroadcastss 0x2213a(%rip),%ymm11 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,144,39,2,0 ; vbroadcastss 0x22790(%rip),%ymm10 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,98,125,24,21,43,33,2,0 ; vbroadcastss 0x2212b(%rip),%ymm10 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 - DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 + DB 197,172,88,219 ; vaddps %ymm3,%ymm10,%ymm3 DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,167,38,2,0 ; vbroadcastss 0x226a7(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,67,32,2,0 ; vbroadcastss 0x22043(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10180,34 +10410,34 @@ _sk_gamma_avx LABEL PROC DB 197,252,40,233 ; vmovaps %ymm1,%ymm5 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,91,200 ; vcvtdq2ps %ymm0,%ymm1 - DB 196,98,125,24,5,240,38,2,0 ; vbroadcastss 0x226f0(%rip),%ymm8 # 2d340 <_sk_overlay_sse2_8bit+0xde8> + DB 196,98,125,24,5,140,32,2,0 ; vbroadcastss 0x2208c(%rip),%ymm8 # 2d220 <_sk_overlay_sse2_8bit+0xd71> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,13,230,38,2,0 ; vbroadcastss 0x226e6(%rip),%ymm9 # 2d344 <_sk_overlay_sse2_8bit+0xdec> + DB 196,98,125,24,13,130,32,2,0 ; vbroadcastss 0x22082(%rip),%ymm9 # 2d224 <_sk_overlay_sse2_8bit+0xd75> DB 196,193,124,84,193 ; vandps %ymm9,%ymm0,%ymm0 - DB 196,226,125,24,37,60,38,2,0 ; vbroadcastss 0x2263c(%rip),%ymm4 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,37,216,31,2,0 ; vbroadcastss 0x21fd8(%rip),%ymm4 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,252,86,196 ; vorps %ymm4,%ymm0,%ymm0 - DB 196,98,125,24,21,207,38,2,0 ; vbroadcastss 0x226cf(%rip),%ymm10 # 2d348 <_sk_overlay_sse2_8bit+0xdf0> + DB 196,98,125,24,21,107,32,2,0 ; vbroadcastss 0x2206b(%rip),%ymm10 # 2d228 <_sk_overlay_sse2_8bit+0xd79> DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,29,197,38,2,0 ; vbroadcastss 0x226c5(%rip),%ymm11 # 2d34c <_sk_overlay_sse2_8bit+0xdf4> + DB 196,98,125,24,29,97,32,2,0 ; vbroadcastss 0x22061(%rip),%ymm11 # 2d22c <_sk_overlay_sse2_8bit+0xd7d> DB 196,193,124,89,211 ; vmulps %ymm11,%ymm0,%ymm2 DB 197,244,92,202 ; vsubps %ymm2,%ymm1,%ymm1 - DB 196,98,125,24,37,183,38,2,0 ; vbroadcastss 0x226b7(%rip),%ymm12 # 2d350 <_sk_overlay_sse2_8bit+0xdf8> + DB 196,98,125,24,37,83,32,2,0 ; vbroadcastss 0x22053(%rip),%ymm12 # 2d230 <_sk_overlay_sse2_8bit+0xd81> DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,45,173,38,2,0 ; vbroadcastss 0x226ad(%rip),%ymm13 # 2d354 <_sk_overlay_sse2_8bit+0xdfc> + DB 196,98,125,24,45,73,32,2,0 ; vbroadcastss 0x22049(%rip),%ymm13 # 2d234 <_sk_overlay_sse2_8bit+0xd85> DB 197,148,94,192 ; vdivps %ymm0,%ymm13,%ymm0 DB 197,244,92,192 ; vsubps %ymm0,%ymm1,%ymm0 DB 196,98,125,24,48 ; vbroadcastss (%rax),%ymm14 DB 196,193,124,89,198 ; vmulps %ymm14,%ymm0,%ymm0 DB 196,227,125,8,200,1 ; vroundps $0x1,%ymm0,%ymm1 DB 197,252,92,241 ; vsubps %ymm1,%ymm0,%ymm6 - DB 196,98,125,24,61,140,38,2,0 ; vbroadcastss 0x2268c(%rip),%ymm15 # 2d358 <_sk_overlay_sse2_8bit+0xe00> + DB 196,98,125,24,61,40,32,2,0 ; vbroadcastss 0x22028(%rip),%ymm15 # 2d238 <_sk_overlay_sse2_8bit+0xd89> DB 196,193,124,88,199 ; vaddps %ymm15,%ymm0,%ymm0 - DB 196,226,125,24,29,130,38,2,0 ; vbroadcastss 0x22682(%rip),%ymm3 # 2d35c <_sk_overlay_sse2_8bit+0xe04> + DB 196,226,125,24,29,30,32,2,0 ; vbroadcastss 0x2201e(%rip),%ymm3 # 2d23c <_sk_overlay_sse2_8bit+0xd8d> DB 197,204,89,203 ; vmulps %ymm3,%ymm6,%ymm1 DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 - DB 196,226,125,24,21,117,38,2,0 ; vbroadcastss 0x22675(%rip),%ymm2 # 2d360 <_sk_overlay_sse2_8bit+0xe08> + DB 196,226,125,24,21,17,32,2,0 ; vbroadcastss 0x22011(%rip),%ymm2 # 2d240 <_sk_overlay_sse2_8bit+0xd91> DB 197,236,92,198 ; vsubps %ymm6,%ymm2,%ymm0 - DB 196,226,125,24,53,108,38,2,0 ; vbroadcastss 0x2266c(%rip),%ymm6 # 2d364 <_sk_overlay_sse2_8bit+0xe0c> + DB 196,226,125,24,53,8,32,2,0 ; vbroadcastss 0x22008(%rip),%ymm6 # 2d244 <_sk_overlay_sse2_8bit+0xd95> DB 197,204,94,192 ; vdivps %ymm0,%ymm6,%ymm0 DB 197,244,88,192 ; vaddps %ymm0,%ymm1,%ymm0 DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) @@ -10249,7 +10479,7 @@ _sk_gamma_avx LABEL PROC DB 197,236,92,212 ; vsubps %ymm4,%ymm2,%ymm2 DB 197,204,94,210 ; vdivps %ymm2,%ymm6,%ymm2 DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,176,37,2,0 ; vbroadcastss 0x225b0(%rip),%ymm2 # 2d368 <_sk_overlay_sse2_8bit+0xe10> + DB 196,226,125,24,21,76,31,2,0 ; vbroadcastss 0x21f4c(%rip),%ymm2 # 2d248 <_sk_overlay_sse2_8bit+0xd99> DB 197,236,89,28,36 ; vmulps (%rsp),%ymm2,%ymm3 DB 197,252,89,226 ; vmulps %ymm2,%ymm0,%ymm4 DB 197,244,89,210 ; vmulps %ymm2,%ymm1,%ymm2 @@ -10267,31 +10497,31 @@ _sk_gamma_avx LABEL PROC PUBLIC _sk_lab_to_xyz_avx _sk_lab_to_xyz_avx LABEL PROC - DB 196,98,125,24,5,99,37,2,0 ; vbroadcastss 0x22563(%rip),%ymm8 # 2d36c <_sk_overlay_sse2_8bit+0xe14> + DB 196,98,125,24,5,255,30,2,0 ; vbroadcastss 0x21eff(%rip),%ymm8 # 2d24c <_sk_overlay_sse2_8bit+0xd9d> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,193,36,2,0 ; vbroadcastss 0x224c1(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,93,30,2,0 ; vbroadcastss 0x21e5d(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,13,75,37,2,0 ; vbroadcastss 0x2254b(%rip),%ymm9 # 2d370 <_sk_overlay_sse2_8bit+0xe18> + DB 196,98,125,24,13,231,30,2,0 ; vbroadcastss 0x21ee7(%rip),%ymm9 # 2d250 <_sk_overlay_sse2_8bit+0xda1> DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 196,193,108,88,209 ; vaddps %ymm9,%ymm2,%ymm2 - DB 196,98,125,24,5,55,37,2,0 ; vbroadcastss 0x22537(%rip),%ymm8 # 2d374 <_sk_overlay_sse2_8bit+0xe1c> + DB 196,98,125,24,5,211,30,2,0 ; vbroadcastss 0x21ed3(%rip),%ymm8 # 2d254 <_sk_overlay_sse2_8bit+0xda5> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,45,37,2,0 ; vbroadcastss 0x2252d(%rip),%ymm8 # 2d378 <_sk_overlay_sse2_8bit+0xe20> + DB 196,98,125,24,5,201,30,2,0 ; vbroadcastss 0x21ec9(%rip),%ymm8 # 2d258 <_sk_overlay_sse2_8bit+0xda9> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,35,37,2,0 ; vbroadcastss 0x22523(%rip),%ymm8 # 2d37c <_sk_overlay_sse2_8bit+0xe24> + DB 196,98,125,24,5,191,30,2,0 ; vbroadcastss 0x21ebf(%rip),%ymm8 # 2d25c <_sk_overlay_sse2_8bit+0xdad> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1 - DB 196,98,125,24,5,21,37,2,0 ; vbroadcastss 0x22515(%rip),%ymm8 # 2d380 <_sk_overlay_sse2_8bit+0xe28> + DB 196,98,125,24,5,177,30,2,0 ; vbroadcastss 0x21eb1(%rip),%ymm8 # 2d260 <_sk_overlay_sse2_8bit+0xdb1> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,252,92,210 ; vsubps %ymm2,%ymm0,%ymm2 DB 197,116,89,193 ; vmulps %ymm1,%ymm1,%ymm8 DB 196,65,116,89,192 ; vmulps %ymm8,%ymm1,%ymm8 - DB 196,98,125,24,13,254,36,2,0 ; vbroadcastss 0x224fe(%rip),%ymm9 # 2d384 <_sk_overlay_sse2_8bit+0xe2c> + DB 196,98,125,24,13,154,30,2,0 ; vbroadcastss 0x21e9a(%rip),%ymm9 # 2d264 <_sk_overlay_sse2_8bit+0xdb5> DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,243,36,2,0 ; vbroadcastss 0x224f3(%rip),%ymm11 # 2d388 <_sk_overlay_sse2_8bit+0xe30> + DB 196,98,125,24,29,143,30,2,0 ; vbroadcastss 0x21e8f(%rip),%ymm11 # 2d268 <_sk_overlay_sse2_8bit+0xdb9> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,37,233,36,2,0 ; vbroadcastss 0x224e9(%rip),%ymm12 # 2d38c <_sk_overlay_sse2_8bit+0xe34> + DB 196,98,125,24,37,133,30,2,0 ; vbroadcastss 0x21e85(%rip),%ymm12 # 2d26c <_sk_overlay_sse2_8bit+0xdbd> DB 196,193,116,89,204 ; vmulps %ymm12,%ymm1,%ymm1 DB 196,67,117,74,192,160 ; vblendvps %ymm10,%ymm8,%ymm1,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -10306,9 +10536,9 @@ _sk_lab_to_xyz_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,193,108,89,212 ; vmulps %ymm12,%ymm2,%ymm2 DB 196,227,109,74,208,144 ; vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - DB 196,226,125,24,5,159,36,2,0 ; vbroadcastss 0x2249f(%rip),%ymm0 # 2d390 <_sk_overlay_sse2_8bit+0xe38> + DB 196,226,125,24,5,59,30,2,0 ; vbroadcastss 0x21e3b(%rip),%ymm0 # 2d270 <_sk_overlay_sse2_8bit+0xdc1> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,150,36,2,0 ; vbroadcastss 0x22496(%rip),%ymm8 # 2d394 <_sk_overlay_sse2_8bit+0xe3c> + DB 196,98,125,24,5,50,30,2,0 ; vbroadcastss 0x21e32(%rip),%ymm8 # 2d274 <_sk_overlay_sse2_8bit+0xdc5> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10322,15 +10552,15 @@ _sk_load_a8_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne af61 <_sk_load_a8_avx+0x5a> + DB 117,68 ; jne b4a7 <_sk_load_a8_avx+0x5c> DB 196,130,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm0 - DB 197,249,219,5,35,37,2,0 ; vpand 0x22523(%rip),%xmm0,%xmm0 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 197,249,219,5,207,30,2,0 ; vpand 0x21ecf(%rip),%xmm0,%xmm0 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,211,35,2,0 ; vbroadcastss 0x223d3(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,109,29,2,0 ; vbroadcastss 0x21d6d(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -10342,15 +10572,15 @@ _sk_load_a8_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,176 ; ja af25 <_sk_load_a8_avx+0x1e> + DB 119,174 ; ja b469 <_sk_load_a8_avx+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # aff8 <_sk_load_a8_avx+0xf1> + DB 76,141,21,122,0,0,0 ; lea 0x7a(%rip),%r10 # b540 <_sk_load_a8_avx+0xf5> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 235,145 ; jmp af25 <_sk_load_a8_avx+0x1e> + DB 235,143 ; jmp b469 <_sk_load_a8_avx+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0 @@ -10358,7 +10588,7 @@ _sk_load_a8_avx LABEL PROC DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,121,14,193,3 ; vpblendw $0x3,%xmm1,%xmm0,%xmm0 - DB 233,105,255,255,255 ; jmpq af25 <_sk_load_a8_avx+0x1e> + DB 233,103,255,255,255 ; jmpq b469 <_sk_load_a8_avx+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0 @@ -10369,22 +10599,24 @@ _sk_load_a8_avx LABEL PROC DB 196,129,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,113,14,192,240 ; vpblendw $0xf0,%xmm0,%xmm1,%xmm0 - DB 233,46,255,255,255 ; jmpq af25 <_sk_load_a8_avx+0x1e> - DB 144 ; nop - DB 145 ; xchg %eax,%ecx + DB 233,44,255,255,255 ; jmpq b469 <_sk_load_a8_avx+0x1e> + DB 15,31,0 ; nopl (%rax) + DB 143 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 255,171,255,255,255,156 ; ljmp *-0x63000001(%rbx) + DB 255,169,255,255,255,154 ; ljmp *-0x65000001(%rcx) DB 255 ; (bad) DB 255 ; (bad) + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf00b008 <_sk_overlay_sse2_8bit+0xffffffffdefdeab0> DB 255 ; (bad) DB 255 ; (bad) - DB 255,211 ; callq *%rbx + DB 220,255 ; fdivr %st,%st(7) + DB 255 ; (bad) + DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) - DB 255,196 ; inc %esp + DB 255,194 ; inc %edx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -10398,15 +10630,15 @@ _sk_load_a8_dst_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne b06e <_sk_load_a8_dst_avx+0x5a> + DB 117,68 ; jne b5b8 <_sk_load_a8_dst_avx+0x5c> DB 196,130,121,48,36,24 ; vpmovzxbw (%r8,%r11,1),%xmm4 - DB 197,217,219,37,22,36,2,0 ; vpand 0x22416(%rip),%xmm4,%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 197,209,239,237 ; vpxor %xmm5,%xmm5,%xmm5 - DB 197,217,105,237 ; vpunpckhwd %xmm5,%xmm4,%xmm5 + DB 197,217,219,37,190,29,2,0 ; vpand 0x21dbe(%rip),%xmm4,%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,226,121,51,236 ; vpmovzxwd %xmm4,%xmm5 + DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,198,34,2,0 ; vbroadcastss 0x222c6(%rip),%ymm5 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,45,92,28,2,0 ; vbroadcastss 0x21c5c(%rip),%ymm5 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,220,89,253 ; vmulps %ymm5,%ymm4,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 @@ -10418,15 +10650,15 @@ _sk_load_a8_dst_avx LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,176 ; ja b032 <_sk_load_a8_dst_avx+0x1e> + DB 119,174 ; ja b57a <_sk_load_a8_dst_avx+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # b104 <_sk_load_a8_dst_avx+0xf0> + DB 76,141,21,121,0,0,0 ; lea 0x79(%rip),%r10 # b650 <_sk_load_a8_dst_avx+0xf4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 235,145 ; jmp b032 <_sk_load_a8_dst_avx+0x1e> + DB 235,143 ; jmp b57a <_sk_load_a8_dst_avx+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,2 ; vpinsrw $0x2,%eax,%xmm4,%xmm4 @@ -10434,7 +10666,7 @@ _sk_load_a8_dst_avx LABEL PROC DB 197,249,110,232 ; vmovd %eax,%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,89,14,229,3 ; vpblendw $0x3,%xmm5,%xmm4,%xmm4 - DB 233,105,255,255,255 ; jmpq b032 <_sk_load_a8_dst_avx+0x1e> + DB 233,103,255,255,255 ; jmpq b57a <_sk_load_a8_dst_avx+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,6 ; vpinsrw $0x6,%eax,%xmm4,%xmm4 @@ -10445,29 +10677,29 @@ _sk_load_a8_dst_avx LABEL PROC DB 196,129,121,110,44,24 ; vmovd (%r8,%r11,1),%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,81,14,228,240 ; vpblendw $0xf0,%xmm4,%xmm5,%xmm4 - DB 233,46,255,255,255 ; jmpq b032 <_sk_load_a8_dst_avx+0x1e> - DB 146 ; xchg %eax,%edx - DB 255 ; (bad) + DB 233,44,255,255,255 ; jmpq b57a <_sk_load_a8_dst_avx+0x1e> + DB 102,144 ; xchg %ax,%ax + DB 144 ; nop DB 255 ; (bad) - DB 255,172,255,255,255,157,255 ; ljmp *-0x620001(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,170,255,255,255,155 ; ljmp *-0x64000001(%rdx) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) DB 255 ; (bad) + DB 232,255,255,255,221 ; callq ffffffffde00b660 <_sk_overlay_sse2_8bit+0xffffffffddfdf1b1> DB 255 ; (bad) - DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,212 ; callq *%rsp + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) - DB 255,197 ; inc %ebp + DB 255,195 ; inc %ebx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_gather_a8_avx _sk_gather_a8_avx LABEL PROC + DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 65,84 ; push %r12 @@ -10490,29 +10722,33 @@ _sk_gather_a8_avx LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,131,121,32,4,48,0 ; vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0 - DB 196,195,121,32,4,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0 - DB 67,15,182,28,32 ; movzbl (%r8,%r12,1),%ebx - DB 196,227,121,32,195,2 ; vpinsrb $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,182,28,56 ; movzbl (%r8,%r15,1),%ebx + DB 71,15,182,60,56 ; movzbl (%r8,%r15,1),%r15d + DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp + DB 197,249,110,197 ; vmovd %ebp,%xmm0 + DB 196,195,121,32,199,1 ; vpinsrb $0x1,%r15d,%xmm0,%xmm0 + DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp + DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 + DB 65,15,182,28,24 ; movzbl (%r8,%rbx,1),%ebx DB 196,227,121,32,195,3 ; vpinsrb $0x3,%ebx,%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,131,121,32,12,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1 - DB 196,131,113,32,12,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 + DB 67,15,182,44,16 ; movzbl (%r8,%r10,1),%ebp + DB 67,15,182,28,24 ; movzbl (%r8,%r11,1),%ebx + DB 197,249,110,203 ; vmovd %ebx,%xmm1 + DB 196,227,113,32,205,1 ; vpinsrb $0x1,%ebp,%xmm1,%xmm1 + DB 67,15,182,44,8 ; movzbl (%r8,%r9,1),%ebp + DB 196,227,113,32,205,2 ; vpinsrb $0x2,%ebp,%xmm1,%xmm1 DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax DB 196,227,113,32,200,3 ; vpinsrb $0x3,%eax,%xmm1,%xmm1 DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,45,33,2,0 ; vbroadcastss 0x2212d(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,180,26,2,0 ; vbroadcastss 0x21ab4(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -10522,6 +10758,7 @@ _sk_gather_a8_avx LABEL PROC DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_store_a8_avx @@ -10532,56 +10769,57 @@ _sk_store_a8_avx LABEL PROC DB 77,15,175,193 ; imul %r9,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,174,32,2,0 ; vbroadcastss 0x220ae(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,52,26,2,0 ; vbroadcastss 0x21a34(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8 DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 - DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 + DB 197,57,103,192 ; vpackuswb %xmm0,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne b253 <_sk_store_a8_avx+0x45> - DB 196,1,123,17,4,24 ; vmovsd %xmm8,(%r8,%r11,1) + DB 117,10 ; jne b7ac <_sk_store_a8_avx+0x44> + DB 196,1,121,214,4,24 ; vmovq %xmm8,(%r8,%r11,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja b24f <_sk_store_a8_avx+0x41> + DB 119,236 ; ja b7a8 <_sk_store_a8_avx+0x40> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,85,0,0,0 ; lea 0x55(%rip),%r10 # b2c8 <_sk_store_a8_avx+0xba> + DB 76,141,21,88,0,0,0 ; lea 0x58(%rip),%r10 # b824 <_sk_store_a8_avx+0xbc> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,3,121,20,4,24,0 ; vpextrb $0x0,%xmm8,(%r8,%r11,1) - DB 235,202 ; jmp b24f <_sk_store_a8_avx+0x41> + DB 235,202 ; jmp b7a8 <_sk_store_a8_avx+0x40> DB 196,3,121,20,68,24,2,4 ; vpextrb $0x4,%xmm8,0x2(%r8,%r11,1) - DB 196,98,57,0,5,202,33,2,0 ; vpshufb 0x221ca(%rip),%xmm8,%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xf08> + DB 196,98,57,0,5,97,27,2,0 ; vpshufb 0x21b61(%rip),%xmm8,%xmm8 # 2d350 <_sk_overlay_sse2_8bit+0xea1> DB 196,3,121,21,4,24,0 ; vpextrw $0x0,%xmm8,(%r8,%r11,1) - DB 235,176 ; jmp b24f <_sk_store_a8_avx+0x41> + DB 235,176 ; jmp b7a8 <_sk_store_a8_avx+0x40> DB 196,3,121,20,68,24,6,12 ; vpextrb $0xc,%xmm8,0x6(%r8,%r11,1) DB 196,3,121,20,68,24,5,10 ; vpextrb $0xa,%xmm8,0x5(%r8,%r11,1) DB 196,3,121,20,68,24,4,8 ; vpextrb $0x8,%xmm8,0x4(%r8,%r11,1) - DB 196,98,57,0,5,176,33,2,0 ; vpshufb 0x221b0(%rip),%xmm8,%xmm8 # 2d470 <_sk_overlay_sse2_8bit+0xf18> + DB 196,98,57,0,5,71,27,2,0 ; vpshufb 0x21b47(%rip),%xmm8,%xmm8 # 2d360 <_sk_overlay_sse2_8bit+0xeb1> DB 196,1,121,126,4,24 ; vmovd %xmm8,(%r8,%r11,1) - DB 235,135 ; jmp b24f <_sk_store_a8_avx+0x41> - DB 180,255 ; mov $0xff,%ah + DB 235,135 ; jmp b7a8 <_sk_store_a8_avx+0x40> + DB 15,31,0 ; nopl (%rax) + DB 177,255 ; mov $0xff,%cl DB 255 ; (bad) - DB 255,197 ; inc %ebp + DB 255,194 ; inc %edx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 189,255,255,255,239 ; mov $0xefffffff,%ebp + DB 186,255,255,255,236 ; mov $0xecffffff,%edx DB 255 ; (bad) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 223,255 ; (bad) + DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) - DB 255,215 ; callq *%rdi + DB 255,212 ; callq *%rsp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -10595,18 +10833,18 @@ _sk_load_g8_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,71 ; jne b343 <_sk_load_g8_avx+0x5f> + DB 117,73 ; jne b8a1 <_sk_load_g8_avx+0x61> DB 196,130,121,48,4,24 ; vpmovzxbw (%r8,%r11,1),%xmm0 - DB 197,249,219,5,70,33,2,0 ; vpand 0x22146(%rip),%xmm0,%xmm0 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 197,249,219,5,218,26,2,0 ; vpand 0x21ada(%rip),%xmm0,%xmm0 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,246,31,2,0 ; vbroadcastss 0x21ff6(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,120,25,2,0 ; vbroadcastss 0x21978(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,115,31,2,0 ; vbroadcastss 0x21f73(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,245,24,2,0 ; vbroadcastss 0x218f5(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 255,224 ; jmpq *%rax @@ -10615,15 +10853,15 @@ _sk_load_g8_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,171 ; ja b302 <_sk_load_g8_avx+0x1e> + DB 119,169 ; ja b85e <_sk_load_g8_avx+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,122,0,0,0 ; lea 0x7a(%rip),%r10 # b3dc <_sk_load_g8_avx+0xf8> + DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # b938 <_sk_load_g8_avx+0xf8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 235,140 ; jmp b302 <_sk_load_g8_avx+0x1e> + DB 235,138 ; jmp b85e <_sk_load_g8_avx+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,2 ; vpinsrw $0x2,%eax,%xmm0,%xmm0 @@ -10631,7 +10869,7 @@ _sk_load_g8_avx LABEL PROC DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,121,14,193,3 ; vpblendw $0x3,%xmm1,%xmm0,%xmm0 - DB 233,100,255,255,255 ; jmpq b302 <_sk_load_g8_avx+0x1e> + DB 233,98,255,255,255 ; jmpq b85e <_sk_load_g8_avx+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 197,249,196,192,6 ; vpinsrw $0x6,%eax,%xmm0,%xmm0 @@ -10642,24 +10880,22 @@ _sk_load_g8_avx LABEL PROC DB 196,129,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm1 DB 196,226,121,48,201 ; vpmovzxbw %xmm1,%xmm1 DB 196,227,113,14,192,240 ; vpblendw $0xf0,%xmm0,%xmm1,%xmm0 - DB 233,41,255,255,255 ; jmpq b302 <_sk_load_g8_avx+0x1e> - DB 15,31,0 ; nopl (%rax) - DB 143 ; (bad) - DB 255 ; (bad) + DB 233,39,255,255,255 ; jmpq b85e <_sk_load_g8_avx+0x1e> + DB 144 ; nop + DB 145 ; xchg %eax,%ecx DB 255 ; (bad) - DB 255,169,255,255,255,154 ; ljmp *-0x65000001(%rcx) DB 255 ; (bad) + DB 255,171,255,255,255,156 ; ljmp *-0x63000001(%rbx) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) + DB 233,255,255,255,222 ; jmpq ffffffffdf00b948 <_sk_overlay_sse2_8bit+0xffffffffdefdf499> DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,211 ; callq *%rbx DB 255 ; (bad) DB 255 ; (bad) - DB 255,194 ; inc %edx + DB 255,196 ; inc %esp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -10673,18 +10909,18 @@ _sk_load_g8_dst_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,71 ; jne b457 <_sk_load_g8_dst_avx+0x5f> + DB 117,73 ; jne b9b5 <_sk_load_g8_dst_avx+0x61> DB 196,130,121,48,36,24 ; vpmovzxbw (%r8,%r11,1),%xmm4 - DB 197,217,219,37,50,32,2,0 ; vpand 0x22032(%rip),%xmm4,%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 197,209,239,237 ; vpxor %xmm5,%xmm5,%xmm5 - DB 197,217,105,237 ; vpunpckhwd %xmm5,%xmm4,%xmm5 + DB 197,217,219,37,198,25,2,0 ; vpand 0x219c6(%rip),%xmm4,%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 196,226,121,51,236 ; vpmovzxwd %xmm4,%xmm5 + DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 + DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,226,30,2,0 ; vbroadcastss 0x21ee2(%rip),%ymm5 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,45,100,24,2,0 ; vbroadcastss 0x21864(%rip),%ymm5 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,220,89,229 ; vmulps %ymm5,%ymm4,%ymm4 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,61,95,30,2,0 ; vbroadcastss 0x21e5f(%rip),%ymm7 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,61,225,23,2,0 ; vbroadcastss 0x217e1(%rip),%ymm7 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,236 ; vmovaps %ymm4,%ymm5 DB 197,252,40,244 ; vmovaps %ymm4,%ymm6 DB 255,224 ; jmpq *%rax @@ -10693,15 +10929,15 @@ _sk_load_g8_dst_avx LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,171 ; ja b416 <_sk_load_g8_dst_avx+0x1e> + DB 119,169 ; ja b972 <_sk_load_g8_dst_avx+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,122,0,0,0 ; lea 0x7a(%rip),%r10 # b4f0 <_sk_load_g8_dst_avx+0xf8> + DB 76,141,21,120,0,0,0 ; lea 0x78(%rip),%r10 # ba4c <_sk_load_g8_dst_avx+0xf8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 235,140 ; jmp b416 <_sk_load_g8_dst_avx+0x1e> + DB 235,138 ; jmp b972 <_sk_load_g8_dst_avx+0x1e> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,2 ; vpinsrw $0x2,%eax,%xmm4,%xmm4 @@ -10709,7 +10945,7 @@ _sk_load_g8_dst_avx LABEL PROC DB 197,249,110,232 ; vmovd %eax,%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,89,14,229,3 ; vpblendw $0x3,%xmm5,%xmm4,%xmm4 - DB 233,100,255,255,255 ; jmpq b416 <_sk_load_g8_dst_avx+0x1e> + DB 233,98,255,255,255 ; jmpq b972 <_sk_load_g8_dst_avx+0x1e> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 197,217,196,224,6 ; vpinsrw $0x6,%eax,%xmm4,%xmm4 @@ -10720,30 +10956,29 @@ _sk_load_g8_dst_avx LABEL PROC DB 196,129,121,110,44,24 ; vmovd (%r8,%r11,1),%xmm5 DB 196,226,121,48,237 ; vpmovzxbw %xmm5,%xmm5 DB 196,227,81,14,228,240 ; vpblendw $0xf0,%xmm4,%xmm5,%xmm4 - DB 233,41,255,255,255 ; jmpq b416 <_sk_load_g8_dst_avx+0x1e> - DB 15,31,0 ; nopl (%rax) - DB 143 ; (bad) - DB 255 ; (bad) + DB 233,39,255,255,255 ; jmpq b972 <_sk_load_g8_dst_avx+0x1e> + DB 144 ; nop + DB 145 ; xchg %eax,%ecx DB 255 ; (bad) - DB 255,169,255,255,255,154 ; ljmp *-0x65000001(%rcx) DB 255 ; (bad) + DB 255,171,255,255,255,156 ; ljmp *-0x63000001(%rbx) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) + DB 233,255,255,255,222 ; jmpq ffffffffdf00ba5c <_sk_overlay_sse2_8bit+0xffffffffdefdf5ad> DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,211 ; callq *%rbx DB 255 ; (bad) DB 255 ; (bad) - DB 255,194 ; inc %edx + DB 255,196 ; inc %esp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_gather_g8_avx _sk_gather_g8_avx LABEL PROC + DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 65,84 ; push %r12 @@ -10766,38 +11001,43 @@ _sk_gather_g8_avx LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,131,121,32,4,48,0 ; vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0 - DB 196,195,121,32,4,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0 - DB 67,15,182,28,32 ; movzbl (%r8,%r12,1),%ebx - DB 196,227,121,32,195,2 ; vpinsrb $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,182,28,56 ; movzbl (%r8,%r15,1),%ebx + DB 71,15,182,60,56 ; movzbl (%r8,%r15,1),%r15d + DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp + DB 197,249,110,197 ; vmovd %ebp,%xmm0 + DB 196,195,121,32,199,1 ; vpinsrb $0x1,%r15d,%xmm0,%xmm0 + DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp + DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0 + DB 65,15,182,28,24 ; movzbl (%r8,%rbx,1),%ebx DB 196,227,121,32,195,3 ; vpinsrb $0x3,%ebx,%xmm0,%xmm0 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 - DB 196,131,121,32,12,24,0 ; vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1 - DB 196,131,113,32,12,16,1 ; vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1 - DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx - DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1 + DB 67,15,182,44,16 ; movzbl (%r8,%r10,1),%ebp + DB 67,15,182,28,24 ; movzbl (%r8,%r11,1),%ebx + DB 197,249,110,203 ; vmovd %ebx,%xmm1 + DB 196,227,113,32,205,1 ; vpinsrb $0x1,%ebp,%xmm1,%xmm1 + DB 67,15,182,44,8 ; movzbl (%r8,%r9,1),%ebp + DB 196,227,113,32,205,2 ; vpinsrb $0x2,%ebp,%xmm1,%xmm1 DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax DB 196,227,113,32,200,3 ; vpinsrb $0x3,%eax,%xmm1,%xmm1 DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,65,29,2,0 ; vbroadcastss 0x21d41(%rip),%ymm1 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,226,125,24,13,184,22,2,0 ; vbroadcastss 0x216b8(%rip),%ymm1 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,190,28,2,0 ; vbroadcastss 0x21cbe(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,53,22,2,0 ; vbroadcastss 0x21635(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_load_565_avx @@ -10810,56 +11050,56 @@ _sk_load_565_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,128,0,0,0 ; jne b69e <_sk_load_565_avx+0x9f> + DB 15,133,130,0,0,0 ; jne bc0a <_sk_load_565_avx+0xa1> DB 196,129,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm0 - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,228,28,2,0 ; vbroadcastss 0x21ce4(%rip),%ymm0 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,227,117,24,208,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm2 + DB 196,226,125,24,5,88,22,2,0 ; vbroadcastss 0x21658(%rip),%ymm0 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,215,28,2,0 ; vbroadcastss 0x21cd7(%rip),%ymm1 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,226,125,24,13,75,22,2,0 ; vbroadcastss 0x2164b(%rip),%ymm1 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,206,28,2,0 ; vbroadcastss 0x21cce(%rip),%ymm1 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,226,125,24,13,66,22,2,0 ; vbroadcastss 0x21642(%rip),%ymm1 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,193,28,2,0 ; vbroadcastss 0x21cc1(%rip),%ymm3 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,226,125,24,29,53,22,2,0 ; vbroadcastss 0x21635(%rip),%ymm3 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,184,28,2,0 ; vbroadcastss 0x21cb8(%rip),%ymm3 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,24,29,44,22,2,0 ; vbroadcastss 0x2162c(%rip),%ymm3 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,171,28,2,0 ; vbroadcastss 0x21cab(%rip),%ymm3 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,226,125,24,29,31,22,2,0 ; vbroadcastss 0x2161f(%rip),%ymm3 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,16,28,2,0 ; vbroadcastss 0x21c10(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,132,21,2,0 ; vbroadcastss 0x21584(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,110,255,255,255 ; ja b624 <_sk_load_565_avx+0x25> + DB 15,135,108,255,255,255 ; ja bb8e <_sk_load_565_avx+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,99,0,0,0 ; lea 0x63(%rip),%r10 # b724 <_sk_load_565_avx+0x125> + DB 76,141,21,99,0,0,0 ; lea 0x63(%rip),%r10 # bc90 <_sk_load_565_avx+0x127> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 233,76,255,255,255 ; jmpq b624 <_sk_load_565_avx+0x25> + DB 233,74,255,255,255 ; jmpq bb8e <_sk_load_565_avx+0x25> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,110,12,88 ; vmovd (%r8,%r11,2),%xmm1 DB 196,227,121,14,193,3 ; vpblendw $0x3,%xmm1,%xmm0,%xmm0 - DB 233,47,255,255,255 ; jmpq b624 <_sk_load_565_avx+0x25> + DB 233,45,255,255,255 ; jmpq bb8e <_sk_load_565_avx+0x25> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,122,126,12,88 ; vmovq (%r8,%r11,2),%xmm1 DB 196,227,113,14,192,240 ; vpblendw $0xf0,%xmm0,%xmm1,%xmm0 - DB 233,2,255,255,255 ; jmpq b624 <_sk_load_565_avx+0x25> + DB 233,0,255,255,255 ; jmpq bb8e <_sk_load_565_avx+0x25> DB 102,144 ; xchg %ax,%ax DB 166 ; cmpsb %es:(%rdi),%ds:(%rsi) DB 255 ; (bad) @@ -10890,73 +11130,75 @@ _sk_load_565_dst_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,128,0,0,0 ; jne b7df <_sk_load_565_dst_avx+0x9f> + DB 15,133,130,0,0,0 ; jne bd4d <_sk_load_565_dst_avx+0xa1> DB 196,129,122,111,36,88 ; vmovdqu (%r8,%r11,2),%xmm4 - DB 197,209,239,237 ; vpxor %xmm5,%xmm5,%xmm5 - DB 197,217,105,237 ; vpunpckhwd %xmm5,%xmm4,%xmm5 + DB 196,226,121,51,236 ; vpmovzxwd %xmm4,%xmm5 + DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,245,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm6 - DB 196,226,125,24,37,163,27,2,0 ; vbroadcastss 0x21ba3(%rip),%ymm4 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,227,85,24,244,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm6 + DB 196,226,125,24,37,21,21,2,0 ; vbroadcastss 0x21515(%rip),%ymm4 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,204,84,228 ; vandps %ymm4,%ymm6,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,150,27,2,0 ; vbroadcastss 0x21b96(%rip),%ymm5 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,226,125,24,45,8,21,2,0 ; vbroadcastss 0x21508(%rip),%ymm5 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 197,220,89,229 ; vmulps %ymm5,%ymm4,%ymm4 - DB 196,226,125,24,45,141,27,2,0 ; vbroadcastss 0x21b8d(%rip),%ymm5 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,226,125,24,45,255,20,2,0 ; vbroadcastss 0x214ff(%rip),%ymm5 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 197,204,84,237 ; vandps %ymm5,%ymm6,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,125,24,61,128,27,2,0 ; vbroadcastss 0x21b80(%rip),%ymm7 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,226,125,24,61,242,20,2,0 ; vbroadcastss 0x214f2(%rip),%ymm7 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 197,212,89,239 ; vmulps %ymm7,%ymm5,%ymm5 - DB 196,226,125,24,61,119,27,2,0 ; vbroadcastss 0x21b77(%rip),%ymm7 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,24,61,233,20,2,0 ; vbroadcastss 0x214e9(%rip),%ymm7 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,204,84,247 ; vandps %ymm7,%ymm6,%ymm6 DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 - DB 196,226,125,24,61,106,27,2,0 ; vbroadcastss 0x21b6a(%rip),%ymm7 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,226,125,24,61,220,20,2,0 ; vbroadcastss 0x214dc(%rip),%ymm7 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 197,204,89,247 ; vmulps %ymm7,%ymm6,%ymm6 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,61,207,26,2,0 ; vbroadcastss 0x21acf(%rip),%ymm7 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,61,65,20,2,0 ; vbroadcastss 0x21441(%rip),%ymm7 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,110,255,255,255 ; ja b765 <_sk_load_565_dst_avx+0x25> + DB 15,135,108,255,255,255 ; ja bcd1 <_sk_load_565_dst_avx+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,98,0,0,0 ; lea 0x62(%rip),%r10 # b864 <_sk_load_565_dst_avx+0x124> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # bdd4 <_sk_load_565_dst_avx+0x128> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 233,76,255,255,255 ; jmpq b765 <_sk_load_565_dst_avx+0x25> + DB 233,74,255,255,255 ; jmpq bcd1 <_sk_load_565_dst_avx+0x25> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,121,110,44,88 ; vmovd (%r8,%r11,2),%xmm5 DB 196,227,89,14,229,3 ; vpblendw $0x3,%xmm5,%xmm4,%xmm4 - DB 233,47,255,255,255 ; jmpq b765 <_sk_load_565_dst_avx+0x25> + DB 233,45,255,255,255 ; jmpq bcd1 <_sk_load_565_dst_avx+0x25> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,122,126,44,88 ; vmovq (%r8,%r11,2),%xmm5 DB 196,227,81,14,228,240 ; vpblendw $0xf0,%xmm4,%xmm5,%xmm4 - DB 233,2,255,255,255 ; jmpq b765 <_sk_load_565_dst_avx+0x25> - DB 144 ; nop - DB 167 ; cmpsl %es:(%rdi),%ds:(%rsi) + DB 233,0,255,255,255 ; jmpq bcd1 <_sk_load_565_dst_avx+0x25> + DB 15,31,0 ; nopl (%rax) + DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,193 ; inc %ecx DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) - DB 255,181,255,255,255,238 ; pushq -0x11000001(%rbp) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 236 ; in (%dx),%al DB 255 ; (bad) DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255 ; (bad) + DB 220,255 ; fdivr %st,%st(7) + DB 255 ; (bad) + DB 255,208 ; callq *%rax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -10986,19 +11228,19 @@ _sk_gather_565_avx LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,15,183,28,88 ; movzwl (%r8,%rbx,2),%ebx - DB 67,15,183,44,112 ; movzwl (%r8,%r14,2),%ebp + DB 71,15,183,60,120 ; movzwl (%r8,%r15,2),%r15d + DB 67,15,183,44,96 ; movzwl (%r8,%r12,2),%ebp DB 197,249,110,197 ; vmovd %ebp,%xmm0 - DB 197,249,196,195,1 ; vpinsrw $0x1,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,96 ; movzwl (%r8,%r12,2),%ebx - DB 197,249,196,195,2 ; vpinsrw $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,120 ; movzwl (%r8,%r15,2),%ebx + DB 196,193,121,196,199,1 ; vpinsrw $0x1,%r15d,%xmm0,%xmm0 + DB 67,15,183,44,112 ; movzwl (%r8,%r14,2),%ebp + DB 197,249,196,197,2 ; vpinsrw $0x2,%ebp,%xmm0,%xmm0 + DB 65,15,183,28,88 ; movzwl (%r8,%rbx,2),%ebx DB 197,249,196,195,3 ; vpinsrw $0x3,%ebx,%xmm0,%xmm0 DB 67,15,183,44,88 ; movzwl (%r8,%r11,2),%ebp DB 197,249,196,197,4 ; vpinsrw $0x4,%ebp,%xmm0,%xmm0 @@ -11008,27 +11250,27 @@ _sk_gather_565_avx LABEL PROC DB 197,249,196,197,6 ; vpinsrw $0x6,%ebp,%xmm0,%xmm0 DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,202,25,2,0 ; vbroadcastss 0x219ca(%rip),%ymm0 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,227,117,24,208,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm2 + DB 196,226,125,24,5,55,19,2,0 ; vbroadcastss 0x21337(%rip),%ymm0 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,189,25,2,0 ; vbroadcastss 0x219bd(%rip),%ymm1 # 2d328 <_sk_overlay_sse2_8bit+0xdd0> + DB 196,226,125,24,13,42,19,2,0 ; vbroadcastss 0x2132a(%rip),%ymm1 # 2d208 <_sk_overlay_sse2_8bit+0xd59> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,180,25,2,0 ; vbroadcastss 0x219b4(%rip),%ymm1 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,226,125,24,13,33,19,2,0 ; vbroadcastss 0x21321(%rip),%ymm1 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,167,25,2,0 ; vbroadcastss 0x219a7(%rip),%ymm3 # 2d330 <_sk_overlay_sse2_8bit+0xdd8> + DB 196,226,125,24,29,20,19,2,0 ; vbroadcastss 0x21314(%rip),%ymm3 # 2d210 <_sk_overlay_sse2_8bit+0xd61> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,158,25,2,0 ; vbroadcastss 0x2199e(%rip),%ymm3 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,24,29,11,19,2,0 ; vbroadcastss 0x2130b(%rip),%ymm3 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,145,25,2,0 ; vbroadcastss 0x21991(%rip),%ymm3 # 2d338 <_sk_overlay_sse2_8bit+0xde0> + DB 196,226,125,24,29,254,18,2,0 ; vbroadcastss 0x212fe(%rip),%ymm3 # 2d218 <_sk_overlay_sse2_8bit+0xd69> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,246,24,2,0 ; vbroadcastss 0x218f6(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,99,18,2,0 ; vbroadcastss 0x21263(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 @@ -11045,14 +11287,14 @@ _sk_store_565_avx LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,185,25,2,0 ; vbroadcastss 0x219b9(%rip),%ymm8 # 2d398 <_sk_overlay_sse2_8bit+0xe40> + DB 196,98,125,24,5,38,19,2,0 ; vbroadcastss 0x21326(%rip),%ymm8 # 2d278 <_sk_overlay_sse2_8bit+0xdc9> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,11 ; vpslld $0xb,%xmm9,%xmm10 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 DB 196,193,49,114,241,11 ; vpslld $0xb,%xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,98,125,24,21,146,25,2,0 ; vbroadcastss 0x21992(%rip),%ymm10 # 2d39c <_sk_overlay_sse2_8bit+0xe44> + DB 196,98,125,24,21,255,18,2,0 ; vbroadcastss 0x212ff(%rip),%ymm10 # 2d27c <_sk_overlay_sse2_8bit+0xdcd> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,33,114,242,5 ; vpslld $0x5,%xmm10,%xmm11 @@ -11066,7 +11308,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne ba5a <_sk_store_565_avx+0x9a> + DB 117,10 ; jne bfcd <_sk_store_565_avx+0x9a> DB 196,1,122,127,4,88 ; vmovdqu %xmm8,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11074,44 +11316,42 @@ _sk_store_565_avx LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja ba56 <_sk_store_565_avx+0x96> + DB 119,236 ; ja bfc9 <_sk_store_565_avx+0x96> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,67,0,0,0 ; lea 0x43(%rip),%r10 # bab8 <_sk_store_565_avx+0xf8> + DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # c02c <_sk_store_565_avx+0xf9> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,3,121,21,4,88,0 ; vpextrw $0x0,%xmm8,(%r8,%r11,2) - DB 235,207 ; jmp ba56 <_sk_store_565_avx+0x96> + DB 235,207 ; jmp bfc9 <_sk_store_565_avx+0x96> DB 196,3,121,21,68,88,4,2 ; vpextrw $0x2,%xmm8,0x4(%r8,%r11,2) DB 196,1,121,126,4,88 ; vmovd %xmm8,(%r8,%r11,2) - DB 235,191 ; jmp ba56 <_sk_store_565_avx+0x96> + DB 235,191 ; jmp bfc9 <_sk_store_565_avx+0x96> DB 196,3,121,21,68,88,12,6 ; vpextrw $0x6,%xmm8,0xc(%r8,%r11,2) DB 196,3,121,21,68,88,10,5 ; vpextrw $0x5,%xmm8,0xa(%r8,%r11,2) DB 196,3,121,21,68,88,8,4 ; vpextrw $0x4,%xmm8,0x8(%r8,%r11,2) DB 196,1,121,214,4,88 ; vmovq %xmm8,(%r8,%r11,2) - DB 235,159 ; jmp ba56 <_sk_store_565_avx+0x96> - DB 144 ; nop - DB 198 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,215 ; callq *%rdi + DB 235,159 ; jmp bfc9 <_sk_store_565_avx+0x96> + DB 102,144 ; xchg %ax,%ax + DB 197,255,255 ; (bad) + DB 255,214 ; callq *%rsi DB 255 ; (bad) DB 255 ; (bad) - DB 255,207 ; dec %edi + DB 255,206 ; dec %esi DB 255 ; (bad) DB 255 ; (bad) - DB 255,247 ; push %rdi + DB 255,246 ; push %rsi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 239 ; out %eax,(%dx) + DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi + DB 255,230 ; jmpq *%rsi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 223,255 ; (bad) + DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -11125,31 +11365,31 @@ _sk_load_4444_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,152,0,0,0 ; jne bb8b <_sk_load_4444_avx+0xb7> + DB 15,133,154,0,0,0 ; jne c101 <_sk_load_4444_avx+0xb9> DB 196,129,122,111,4,88 ; vmovdqu (%r8,%r11,2),%xmm0 - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,139,24,2,0 ; vbroadcastss 0x2188b(%rip),%ymm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xe48> + DB 196,227,117,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm3 + DB 196,226,125,24,5,245,17,2,0 ; vbroadcastss 0x211f5(%rip),%ymm0 # 2d280 <_sk_overlay_sse2_8bit+0xdd1> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,126,24,2,0 ; vbroadcastss 0x2187e(%rip),%ymm1 # 2d3a4 <_sk_overlay_sse2_8bit+0xe4c> + DB 196,226,125,24,13,232,17,2,0 ; vbroadcastss 0x211e8(%rip),%ymm1 # 2d284 <_sk_overlay_sse2_8bit+0xdd5> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,117,24,2,0 ; vbroadcastss 0x21875(%rip),%ymm1 # 2d3a8 <_sk_overlay_sse2_8bit+0xe50> + DB 196,226,125,24,13,223,17,2,0 ; vbroadcastss 0x211df(%rip),%ymm1 # 2d288 <_sk_overlay_sse2_8bit+0xdd9> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,104,24,2,0 ; vbroadcastss 0x21868(%rip),%ymm2 # 2d3ac <_sk_overlay_sse2_8bit+0xe54> + DB 196,226,125,24,21,210,17,2,0 ; vbroadcastss 0x211d2(%rip),%ymm2 # 2d28c <_sk_overlay_sse2_8bit+0xddd> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,95,24,2,0 ; vbroadcastss 0x2185f(%rip),%ymm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xe58> + DB 196,226,125,24,21,201,17,2,0 ; vbroadcastss 0x211c9(%rip),%ymm2 # 2d290 <_sk_overlay_sse2_8bit+0xde1> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,82,24,2,0 ; vbroadcastss 0x21852(%rip),%ymm8 # 2d3b4 <_sk_overlay_sse2_8bit+0xe5c> + DB 196,98,125,24,5,188,17,2,0 ; vbroadcastss 0x211bc(%rip),%ymm8 # 2d294 <_sk_overlay_sse2_8bit+0xde5> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,72,24,2,0 ; vbroadcastss 0x21848(%rip),%ymm8 # 2d3b8 <_sk_overlay_sse2_8bit+0xe60> + DB 196,98,125,24,5,178,17,2,0 ; vbroadcastss 0x211b2(%rip),%ymm8 # 2d298 <_sk_overlay_sse2_8bit+0xde9> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,58,24,2,0 ; vbroadcastss 0x2183a(%rip),%ymm8 # 2d3bc <_sk_overlay_sse2_8bit+0xe64> + DB 196,98,125,24,5,164,17,2,0 ; vbroadcastss 0x211a4(%rip),%ymm8 # 2d29c <_sk_overlay_sse2_8bit+0xded> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11158,44 +11398,46 @@ _sk_load_4444_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,86,255,255,255 ; ja baf9 <_sk_load_4444_avx+0x25> + DB 15,135,84,255,255,255 ; ja c06d <_sk_load_4444_avx+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,98,0,0,0 ; lea 0x62(%rip),%r10 # bc10 <_sk_load_4444_avx+0x13c> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # c188 <_sk_load_4444_avx+0x140> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 233,52,255,255,255 ; jmpq baf9 <_sk_load_4444_avx+0x25> + DB 233,50,255,255,255 ; jmpq c06d <_sk_load_4444_avx+0x25> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,110,12,88 ; vmovd (%r8,%r11,2),%xmm1 DB 196,227,121,14,193,3 ; vpblendw $0x3,%xmm1,%xmm0,%xmm0 - DB 233,23,255,255,255 ; jmpq baf9 <_sk_load_4444_avx+0x25> + DB 233,21,255,255,255 ; jmpq c06d <_sk_load_4444_avx+0x25> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,129,121,196,68,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,121,196,68,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm0,%xmm0 DB 196,129,122,126,12,88 ; vmovq (%r8,%r11,2),%xmm1 DB 196,227,113,14,192,240 ; vpblendw $0xf0,%xmm0,%xmm1,%xmm0 - DB 233,234,254,255,255 ; jmpq baf9 <_sk_load_4444_avx+0x25> - DB 144 ; nop - DB 167 ; cmpsl %es:(%rdi),%ds:(%rsi) + DB 233,232,254,255,255 ; jmpq c06d <_sk_load_4444_avx+0x25> + DB 15,31,0 ; nopl (%rax) + DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,193 ; inc %ecx DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) - DB 255,181,255,255,255,238 ; pushq -0x11000001(%rbp) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 236 ; in (%dx),%al DB 255 ; (bad) DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255 ; (bad) + DB 220,255 ; fdivr %st,%st(7) + DB 255 ; (bad) + DB 255,208 ; callq *%rax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -11210,31 +11452,31 @@ _sk_load_4444_dst_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,152,0,0,0 ; jne bce3 <_sk_load_4444_dst_avx+0xb7> + DB 15,133,154,0,0,0 ; jne c25d <_sk_load_4444_dst_avx+0xb9> DB 196,129,122,111,36,88 ; vmovdqu (%r8,%r11,2),%xmm4 - DB 197,209,239,237 ; vpxor %xmm5,%xmm5,%xmm5 - DB 197,217,105,237 ; vpunpckhwd %xmm5,%xmm4,%xmm5 + DB 196,226,121,51,236 ; vpmovzxwd %xmm4,%xmm5 + DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,253,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm7 - DB 196,226,125,24,37,51,23,2,0 ; vbroadcastss 0x21733(%rip),%ymm4 # 2d3a0 <_sk_overlay_sse2_8bit+0xe48> + DB 196,227,85,24,252,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm7 + DB 196,226,125,24,37,153,16,2,0 ; vbroadcastss 0x21099(%rip),%ymm4 # 2d280 <_sk_overlay_sse2_8bit+0xdd1> DB 197,196,84,228 ; vandps %ymm4,%ymm7,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,226,125,24,45,38,23,2,0 ; vbroadcastss 0x21726(%rip),%ymm5 # 2d3a4 <_sk_overlay_sse2_8bit+0xe4c> + DB 196,226,125,24,45,140,16,2,0 ; vbroadcastss 0x2108c(%rip),%ymm5 # 2d284 <_sk_overlay_sse2_8bit+0xdd5> DB 197,220,89,229 ; vmulps %ymm5,%ymm4,%ymm4 - DB 196,226,125,24,45,29,23,2,0 ; vbroadcastss 0x2171d(%rip),%ymm5 # 2d3a8 <_sk_overlay_sse2_8bit+0xe50> + DB 196,226,125,24,45,131,16,2,0 ; vbroadcastss 0x21083(%rip),%ymm5 # 2d288 <_sk_overlay_sse2_8bit+0xdd9> DB 197,196,84,237 ; vandps %ymm5,%ymm7,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,226,125,24,53,16,23,2,0 ; vbroadcastss 0x21710(%rip),%ymm6 # 2d3ac <_sk_overlay_sse2_8bit+0xe54> + DB 196,226,125,24,53,118,16,2,0 ; vbroadcastss 0x21076(%rip),%ymm6 # 2d28c <_sk_overlay_sse2_8bit+0xddd> DB 197,212,89,238 ; vmulps %ymm6,%ymm5,%ymm5 - DB 196,226,125,24,53,7,23,2,0 ; vbroadcastss 0x21707(%rip),%ymm6 # 2d3b0 <_sk_overlay_sse2_8bit+0xe58> + DB 196,226,125,24,53,109,16,2,0 ; vbroadcastss 0x2106d(%rip),%ymm6 # 2d290 <_sk_overlay_sse2_8bit+0xde1> DB 197,196,84,246 ; vandps %ymm6,%ymm7,%ymm6 DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 - DB 196,98,125,24,5,250,22,2,0 ; vbroadcastss 0x216fa(%rip),%ymm8 # 2d3b4 <_sk_overlay_sse2_8bit+0xe5c> + DB 196,98,125,24,5,96,16,2,0 ; vbroadcastss 0x21060(%rip),%ymm8 # 2d294 <_sk_overlay_sse2_8bit+0xde5> DB 196,193,76,89,240 ; vmulps %ymm8,%ymm6,%ymm6 - DB 196,98,125,24,5,240,22,2,0 ; vbroadcastss 0x216f0(%rip),%ymm8 # 2d3b8 <_sk_overlay_sse2_8bit+0xe60> + DB 196,98,125,24,5,86,16,2,0 ; vbroadcastss 0x21056(%rip),%ymm8 # 2d298 <_sk_overlay_sse2_8bit+0xde9> DB 196,193,68,84,248 ; vandps %ymm8,%ymm7,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,226,22,2,0 ; vbroadcastss 0x216e2(%rip),%ymm8 # 2d3bc <_sk_overlay_sse2_8bit+0xe64> + DB 196,98,125,24,5,72,16,2,0 ; vbroadcastss 0x21048(%rip),%ymm8 # 2d29c <_sk_overlay_sse2_8bit+0xded> DB 196,193,68,89,248 ; vmulps %ymm8,%ymm7,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11243,44 +11485,46 @@ _sk_load_4444_dst_avx LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,86,255,255,255 ; ja bc51 <_sk_load_4444_dst_avx+0x25> + DB 15,135,84,255,255,255 ; ja c1c9 <_sk_load_4444_dst_avx+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,98,0,0,0 ; lea 0x62(%rip),%r10 # bd68 <_sk_load_4444_dst_avx+0x13c> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # c2e4 <_sk_load_4444_dst_avx+0x140> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 233,52,255,255,255 ; jmpq bc51 <_sk_load_4444_dst_avx+0x25> + DB 233,50,255,255,255 ; jmpq c1c9 <_sk_load_4444_dst_avx+0x25> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,121,110,44,88 ; vmovd (%r8,%r11,2),%xmm5 DB 196,227,89,14,229,3 ; vpblendw $0x3,%xmm5,%xmm4,%xmm4 - DB 233,23,255,255,255 ; jmpq bc51 <_sk_load_4444_dst_avx+0x25> + DB 233,21,255,255,255 ; jmpq c1c9 <_sk_load_4444_dst_avx+0x25> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,129,89,196,100,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,10,5 ; vpinsrw $0x5,0xa(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,89,196,100,88,8,4 ; vpinsrw $0x4,0x8(%r8,%r11,2),%xmm4,%xmm4 DB 196,129,122,126,44,88 ; vmovq (%r8,%r11,2),%xmm5 DB 196,227,81,14,228,240 ; vpblendw $0xf0,%xmm4,%xmm5,%xmm4 - DB 233,234,254,255,255 ; jmpq bc51 <_sk_load_4444_dst_avx+0x25> - DB 144 ; nop - DB 167 ; cmpsl %es:(%rdi),%ds:(%rsi) + DB 233,232,254,255,255 ; jmpq c1c9 <_sk_load_4444_dst_avx+0x25> + DB 15,31,0 ; nopl (%rax) + DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,193 ; inc %ecx DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) - DB 255,181,255,255,255,238 ; pushq -0x11000001(%rbp) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 236 ; in (%dx),%al DB 255 ; (bad) DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255 ; (bad) + DB 220,255 ; fdivr %st,%st(7) + DB 255 ; (bad) + DB 255,208 ; callq *%rax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -11310,19 +11554,19 @@ _sk_gather_4444_avx LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,15,183,28,88 ; movzwl (%r8,%rbx,2),%ebx - DB 67,15,183,44,112 ; movzwl (%r8,%r14,2),%ebp + DB 71,15,183,60,120 ; movzwl (%r8,%r15,2),%r15d + DB 67,15,183,44,96 ; movzwl (%r8,%r12,2),%ebp DB 197,249,110,197 ; vmovd %ebp,%xmm0 - DB 197,249,196,195,1 ; vpinsrw $0x1,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,96 ; movzwl (%r8,%r12,2),%ebx - DB 197,249,196,195,2 ; vpinsrw $0x2,%ebx,%xmm0,%xmm0 - DB 67,15,183,28,120 ; movzwl (%r8,%r15,2),%ebx + DB 196,193,121,196,199,1 ; vpinsrw $0x1,%r15d,%xmm0,%xmm0 + DB 67,15,183,44,112 ; movzwl (%r8,%r14,2),%ebp + DB 197,249,196,197,2 ; vpinsrw $0x2,%ebp,%xmm0,%xmm0 + DB 65,15,183,28,88 ; movzwl (%r8,%rbx,2),%ebx DB 197,249,196,195,3 ; vpinsrw $0x3,%ebx,%xmm0,%xmm0 DB 67,15,183,44,88 ; movzwl (%r8,%r11,2),%ebp DB 197,249,196,197,4 ; vpinsrw $0x4,%ebp,%xmm0,%xmm0 @@ -11332,29 +11576,29 @@ _sk_gather_4444_avx LABEL PROC DB 197,249,196,197,6 ; vpinsrw $0x6,%ebp,%xmm0,%xmm0 DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax DB 197,249,196,192,7 ; vpinsrw $0x7,%eax,%xmm0,%xmm0 - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,66,21,2,0 ; vbroadcastss 0x21542(%rip),%ymm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xe48> + DB 196,227,117,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm3 + DB 196,226,125,24,5,163,14,2,0 ; vbroadcastss 0x20ea3(%rip),%ymm0 # 2d280 <_sk_overlay_sse2_8bit+0xdd1> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,53,21,2,0 ; vbroadcastss 0x21535(%rip),%ymm1 # 2d3a4 <_sk_overlay_sse2_8bit+0xe4c> + DB 196,226,125,24,13,150,14,2,0 ; vbroadcastss 0x20e96(%rip),%ymm1 # 2d284 <_sk_overlay_sse2_8bit+0xdd5> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,44,21,2,0 ; vbroadcastss 0x2152c(%rip),%ymm1 # 2d3a8 <_sk_overlay_sse2_8bit+0xe50> + DB 196,226,125,24,13,141,14,2,0 ; vbroadcastss 0x20e8d(%rip),%ymm1 # 2d288 <_sk_overlay_sse2_8bit+0xdd9> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,31,21,2,0 ; vbroadcastss 0x2151f(%rip),%ymm2 # 2d3ac <_sk_overlay_sse2_8bit+0xe54> + DB 196,226,125,24,21,128,14,2,0 ; vbroadcastss 0x20e80(%rip),%ymm2 # 2d28c <_sk_overlay_sse2_8bit+0xddd> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,22,21,2,0 ; vbroadcastss 0x21516(%rip),%ymm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xe58> + DB 196,226,125,24,21,119,14,2,0 ; vbroadcastss 0x20e77(%rip),%ymm2 # 2d290 <_sk_overlay_sse2_8bit+0xde1> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,9,21,2,0 ; vbroadcastss 0x21509(%rip),%ymm8 # 2d3b4 <_sk_overlay_sse2_8bit+0xe5c> + DB 196,98,125,24,5,106,14,2,0 ; vbroadcastss 0x20e6a(%rip),%ymm8 # 2d294 <_sk_overlay_sse2_8bit+0xde5> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,255,20,2,0 ; vbroadcastss 0x214ff(%rip),%ymm8 # 2d3b8 <_sk_overlay_sse2_8bit+0xe60> + DB 196,98,125,24,5,96,14,2,0 ; vbroadcastss 0x20e60(%rip),%ymm8 # 2d298 <_sk_overlay_sse2_8bit+0xde9> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,241,20,2,0 ; vbroadcastss 0x214f1(%rip),%ymm8 # 2d3bc <_sk_overlay_sse2_8bit+0xe64> + DB 196,98,125,24,5,82,14,2,0 ; vbroadcastss 0x20e52(%rip),%ymm8 # 2d29c <_sk_overlay_sse2_8bit+0xded> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -11373,7 +11617,7 @@ _sk_store_4444_avx LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,197,20,2,0 ; vbroadcastss 0x214c5(%rip),%ymm8 # 2d3c0 <_sk_overlay_sse2_8bit+0xe68> + DB 196,98,125,24,5,38,14,2,0 ; vbroadcastss 0x20e26(%rip),%ymm8 # 2d2a0 <_sk_overlay_sse2_8bit+0xdf1> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,12 ; vpslld $0xc,%xmm9,%xmm10 @@ -11400,7 +11644,7 @@ _sk_store_4444_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne bf94 <_sk_store_4444_avx+0xb8> + DB 117,10 ; jne c513 <_sk_store_4444_avx+0xb8> DB 196,1,122,127,4,88 ; vmovdqu %xmm8,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11408,44 +11652,41 @@ _sk_store_4444_avx LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja bf90 <_sk_store_4444_avx+0xb4> + DB 119,236 ; ja c50f <_sk_store_4444_avx+0xb4> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,69,0,0,0 ; lea 0x45(%rip),%r10 # bff4 <_sk_store_4444_avx+0x118> + DB 76,141,21,66,0,0,0 ; lea 0x42(%rip),%r10 # c570 <_sk_store_4444_avx+0x115> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,3,121,21,4,88,0 ; vpextrw $0x0,%xmm8,(%r8,%r11,2) - DB 235,207 ; jmp bf90 <_sk_store_4444_avx+0xb4> + DB 235,207 ; jmp c50f <_sk_store_4444_avx+0xb4> DB 196,3,121,21,68,88,4,2 ; vpextrw $0x2,%xmm8,0x4(%r8,%r11,2) DB 196,1,121,126,4,88 ; vmovd %xmm8,(%r8,%r11,2) - DB 235,191 ; jmp bf90 <_sk_store_4444_avx+0xb4> + DB 235,191 ; jmp c50f <_sk_store_4444_avx+0xb4> DB 196,3,121,21,68,88,12,6 ; vpextrw $0x6,%xmm8,0xc(%r8,%r11,2) DB 196,3,121,21,68,88,10,5 ; vpextrw $0x5,%xmm8,0xa(%r8,%r11,2) DB 196,3,121,21,68,88,8,4 ; vpextrw $0x4,%xmm8,0x8(%r8,%r11,2) DB 196,1,121,214,4,88 ; vmovq %xmm8,(%r8,%r11,2) - DB 235,159 ; jmp bf90 <_sk_store_4444_avx+0xb4> - DB 15,31,0 ; nopl (%rax) - DB 196 ; (bad) - DB 255 ; (bad) + DB 235,159 ; jmp c50f <_sk_store_4444_avx+0xb4> + DB 199 ; (bad) DB 255 ; (bad) - DB 255,213 ; callq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 255,205 ; dec %ebp + DB 216,255 ; fdivr %st(7),%st DB 255 ; (bad) + DB 255,208 ; callq *%rax DB 255 ; (bad) - DB 255,245 ; push %rbp DB 255 ; (bad) DB 255 ; (bad) + DB 248 ; clc DB 255 ; (bad) - DB 237 ; in (%dx),%eax DB 255 ; (bad) + DB 255,240 ; push %rax DB 255 ; (bad) - DB 255,229 ; jmpq *%rbp DB 255 ; (bad) DB 255 ; (bad) + DB 232,255,255,255,224 ; callq ffffffffe100c588 <_sk_overlay_sse2_8bit+0xffffffffe0fe00d9> DB 255 ; (bad) - DB 221,255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -11459,12 +11700,12 @@ _sk_load_8888_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,135,0,0,0 ; jne c0b7 <_sk_load_8888_avx+0xa7> + DB 15,133,135,0,0,0 ; jne c633 <_sk_load_8888_avx+0xa7> DB 196,1,125,16,12,152 ; vmovupd (%r8,%r11,4),%ymm9 - DB 197,125,40,21,66,13,2,0 ; vmovapd 0x20d42(%rip),%ymm10 # 2cd80 <_sk_overlay_sse2_8bit+0x828> + DB 197,125,40,21,198,6,2,0 ; vmovapd 0x206c6(%rip),%ymm10 # 2cc80 <_sk_overlay_sse2_8bit+0x7d1> DB 196,193,53,84,194 ; vandpd %ymm10,%ymm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,208,18,2,0 ; vbroadcastss 0x212d0(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,52,12,2,0 ; vbroadcastss 0x20c34(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 196,193,113,114,209,8 ; vpsrld $0x8,%xmm9,%xmm1 DB 196,99,125,25,203,1 ; vextractf128 $0x1,%ymm9,%xmm3 @@ -11491,21 +11732,21 @@ _sk_load_8888_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,102,255,255,255 ; ja c036 <_sk_load_8888_avx+0x26> + DB 15,135,102,255,255,255 ; ja c5b2 <_sk_load_8888_avx+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # c16c <_sk_load_8888_avx+0x15c> + DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # c6e8 <_sk_load_8888_avx+0x15c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,122,16,12,152 ; vmovss (%r8,%r11,4),%xmm9 - DB 233,71,255,255,255 ; jmpq c036 <_sk_load_8888_avx+0x26> + DB 233,71,255,255,255 ; jmpq c5b2 <_sk_load_8888_avx+0x26> DB 196,129,121,110,68,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm0 DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,99,117,12,200,4 ; vblendps $0x4,%ymm0,%ymm1,%ymm9 DB 196,129,123,16,4,152 ; vmovsd (%r8,%r11,4),%xmm0 DB 196,99,53,13,200,1 ; vblendpd $0x1,%ymm0,%ymm9,%ymm9 - DB 233,32,255,255,255 ; jmpq c036 <_sk_load_8888_avx+0x26> + DB 233,32,255,255,255 ; jmpq c5b2 <_sk_load_8888_avx+0x26> DB 196,129,121,110,68,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm0 DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 @@ -11519,9 +11760,9 @@ _sk_load_8888_avx LABEL PROC DB 196,99,53,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 DB 196,129,121,16,4,152 ; vmovupd (%r8,%r11,4),%xmm0 DB 196,67,125,13,201,12 ; vblendpd $0xc,%ymm9,%ymm0,%ymm9 - DB 233,203,254,255,255 ; jmpq c036 <_sk_load_8888_avx+0x26> + DB 233,203,254,255,255 ; jmpq c5b2 <_sk_load_8888_avx+0x26> DB 144 ; nop - DB 120,255 ; js c16d <_sk_load_8888_avx+0x15d> + DB 120,255 ; js c6e9 <_sk_load_8888_avx+0x15d> DB 255 ; (bad) DB 255,153,255,255,255,131 ; lcall *-0x7c000001(%rcx) DB 255 ; (bad) @@ -11552,12 +11793,12 @@ _sk_load_8888_dst_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,135,0,0,0 ; jne c22f <_sk_load_8888_dst_avx+0xa7> + DB 15,133,135,0,0,0 ; jne c7ab <_sk_load_8888_dst_avx+0xa7> DB 196,1,125,16,12,152 ; vmovupd (%r8,%r11,4),%ymm9 - DB 197,125,40,21,234,11,2,0 ; vmovapd 0x20bea(%rip),%ymm10 # 2cda0 <_sk_overlay_sse2_8bit+0x848> + DB 197,125,40,21,110,5,2,0 ; vmovapd 0x2056e(%rip),%ymm10 # 2cca0 <_sk_overlay_sse2_8bit+0x7f1> DB 196,193,53,84,226 ; vandpd %ymm10,%ymm9,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,98,125,24,5,88,17,2,0 ; vbroadcastss 0x21158(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,188,10,2,0 ; vbroadcastss 0x20abc(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,92,89,224 ; vmulps %ymm8,%ymm4,%ymm4 DB 196,193,81,114,209,8 ; vpsrld $0x8,%xmm9,%xmm5 DB 196,99,125,25,207,1 ; vextractf128 $0x1,%ymm9,%xmm7 @@ -11584,21 +11825,21 @@ _sk_load_8888_dst_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,102,255,255,255 ; ja c1ae <_sk_load_8888_dst_avx+0x26> + DB 15,135,102,255,255,255 ; ja c72a <_sk_load_8888_dst_avx+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # c2e4 <_sk_load_8888_dst_avx+0x15c> + DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # c860 <_sk_load_8888_dst_avx+0x15c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,122,16,12,152 ; vmovss (%r8,%r11,4),%xmm9 - DB 233,71,255,255,255 ; jmpq c1ae <_sk_load_8888_dst_avx+0x26> + DB 233,71,255,255,255 ; jmpq c72a <_sk_load_8888_dst_avx+0x26> DB 196,129,121,110,100,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm4 DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 196,99,85,12,204,4 ; vblendps $0x4,%ymm4,%ymm5,%ymm9 DB 196,129,123,16,36,152 ; vmovsd (%r8,%r11,4),%xmm4 DB 196,99,53,13,204,1 ; vblendpd $0x1,%ymm4,%ymm9,%ymm9 - DB 233,32,255,255,255 ; jmpq c1ae <_sk_load_8888_dst_avx+0x26> + DB 233,32,255,255,255 ; jmpq c72a <_sk_load_8888_dst_avx+0x26> DB 196,129,121,110,100,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm4 DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 DB 196,227,125,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm0,%ymm4 @@ -11612,9 +11853,9 @@ _sk_load_8888_dst_avx LABEL PROC DB 196,99,53,24,204,1 ; vinsertf128 $0x1,%xmm4,%ymm9,%ymm9 DB 196,129,121,16,36,152 ; vmovupd (%r8,%r11,4),%xmm4 DB 196,67,93,13,201,12 ; vblendpd $0xc,%ymm9,%ymm4,%ymm9 - DB 233,203,254,255,255 ; jmpq c1ae <_sk_load_8888_dst_avx+0x26> + DB 233,203,254,255,255 ; jmpq c72a <_sk_load_8888_dst_avx+0x26> DB 144 ; nop - DB 120,255 ; js c2e5 <_sk_load_8888_dst_avx+0x15d> + DB 120,255 ; js c861 <_sk_load_8888_dst_avx+0x15d> DB 255 ; (bad) DB 255,153,255,255,255,131 ; lcall *-0x7c000001(%rcx) DB 255 ; (bad) @@ -11655,29 +11896,29 @@ _sk_gather_8888_avx LABEL PROC DB 196,225,249,126,200 ; vmovq %xmm1,%rax DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx DB 65,137,222 ; mov %ebx,%r14d + DB 72,193,235,32 ; shr $0x20,%rbx DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d - DB 72,193,235,32 ; shr $0x20,%rbx DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,121,110,4,176 ; vmovd (%r8,%r14,4),%xmm0 - DB 196,195,121,34,4,152,1 ; vpinsrd $0x1,(%r8,%rbx,4),%xmm0,%xmm0 + DB 196,129,121,110,4,152 ; vmovd (%r8,%r11,4),%xmm0 + DB 196,131,121,34,4,144,1 ; vpinsrd $0x1,(%r8,%r10,4),%xmm0,%xmm0 DB 196,131,121,34,4,160,2 ; vpinsrd $0x2,(%r8,%r12,4),%xmm0,%xmm0 DB 196,3,121,34,4,184,3 ; vpinsrd $0x3,(%r8,%r15,4),%xmm0,%xmm8 DB 196,129,121,110,4,136 ; vmovd (%r8,%r9,4),%xmm0 DB 196,195,121,34,4,128,1 ; vpinsrd $0x1,(%r8,%rax,4),%xmm0,%xmm0 - DB 196,131,121,34,4,152,2 ; vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 - DB 196,131,121,34,28,144,3 ; vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 + DB 196,131,121,34,4,176,2 ; vpinsrd $0x2,(%r8,%r14,4),%xmm0,%xmm0 + DB 196,195,121,34,28,152,3 ; vpinsrd $0x3,(%r8,%rbx,4),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,14,10,2,0 ; vmovaps 0x20a0e(%rip),%ymm10 # 2cdc0 <_sk_overlay_sse2_8bit+0x868> + DB 197,124,40,21,146,3,2,0 ; vmovaps 0x20392(%rip),%ymm10 # 2ccc0 <_sk_overlay_sse2_8bit+0x811> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,92,15,2,0 ; vbroadcastss 0x20f5c(%rip),%ymm9 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,13,192,8,2,0 ; vbroadcastss 0x208c0(%rip),%ymm9 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -11712,7 +11953,7 @@ _sk_store_8888_avx LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,140,14,2,0 ; vbroadcastss 0x20e8c(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,240,7,2,0 ; vbroadcastss 0x207f0(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -11737,7 +11978,7 @@ _sk_store_8888_avx LABEL PROC DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8 DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne c4da <_sk_store_8888_avx+0xae> + DB 117,10 ; jne ca56 <_sk_store_8888_avx+0xae> DB 196,1,124,17,4,152 ; vmovups %ymm8,(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -11745,17 +11986,17 @@ _sk_store_8888_avx LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja c4d6 <_sk_store_8888_avx+0xaa> + DB 119,236 ; ja ca52 <_sk_store_8888_avx+0xaa> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # c548 <_sk_store_8888_avx+0x11c> + DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # cac4 <_sk_store_8888_avx+0x11c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,121,126,4,152 ; vmovd %xmm8,(%r8,%r11,4) - DB 235,208 ; jmp c4d6 <_sk_store_8888_avx+0xaa> + DB 235,208 ; jmp ca52 <_sk_store_8888_avx+0xaa> DB 196,3,121,22,68,152,8,2 ; vpextrd $0x2,%xmm8,0x8(%r8,%r11,4) DB 196,1,121,214,4,152 ; vmovq %xmm8,(%r8,%r11,4) - DB 235,192 ; jmp c4d6 <_sk_store_8888_avx+0xaa> + DB 235,192 ; jmp ca52 <_sk_store_8888_avx+0xaa> DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,3,121,22,76,152,24,2 ; vpextrd $0x2,%xmm9,0x18(%r8,%r11,4) DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 @@ -11763,7 +12004,7 @@ _sk_store_8888_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,1,122,17,76,152,16 ; vmovss %xmm9,0x10(%r8,%r11,4) DB 196,1,121,17,4,152 ; vmovupd %xmm8,(%r8,%r11,4) - DB 235,143 ; jmp c4d6 <_sk_store_8888_avx+0xaa> + DB 235,143 ; jmp ca52 <_sk_store_8888_avx+0xaa> DB 144 ; nop DB 182,255 ; mov $0xff,%dh DB 255 ; (bad) @@ -11796,12 +12037,12 @@ _sk_load_bgra_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,135,0,0,0 ; jne c60b <_sk_load_bgra_avx+0xa7> + DB 15,133,135,0,0,0 ; jne cb87 <_sk_load_bgra_avx+0xa7> DB 196,1,125,16,12,152 ; vmovupd (%r8,%r11,4),%ymm9 - DB 197,125,40,21,78,8,2,0 ; vmovapd 0x2084e(%rip),%ymm10 # 2cde0 <_sk_overlay_sse2_8bit+0x888> + DB 197,125,40,21,210,1,2,0 ; vmovapd 0x201d2(%rip),%ymm10 # 2cce0 <_sk_overlay_sse2_8bit+0x831> DB 196,193,53,84,202 ; vandpd %ymm10,%ymm9,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,98,125,24,5,124,13,2,0 ; vbroadcastss 0x20d7c(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,224,6,2,0 ; vbroadcastss 0x206e0(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,116,89,208 ; vmulps %ymm8,%ymm1,%ymm2 DB 196,193,113,114,209,8 ; vpsrld $0x8,%xmm9,%xmm1 DB 196,99,125,25,203,1 ; vextractf128 $0x1,%ymm9,%xmm3 @@ -11828,21 +12069,21 @@ _sk_load_bgra_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,102,255,255,255 ; ja c58a <_sk_load_bgra_avx+0x26> + DB 15,135,102,255,255,255 ; ja cb06 <_sk_load_bgra_avx+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # c6c0 <_sk_load_bgra_avx+0x15c> + DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # cc3c <_sk_load_bgra_avx+0x15c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,122,16,12,152 ; vmovss (%r8,%r11,4),%xmm9 - DB 233,71,255,255,255 ; jmpq c58a <_sk_load_bgra_avx+0x26> + DB 233,71,255,255,255 ; jmpq cb06 <_sk_load_bgra_avx+0x26> DB 196,129,121,110,68,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm0 DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,99,117,12,200,4 ; vblendps $0x4,%ymm0,%ymm1,%ymm9 DB 196,129,123,16,4,152 ; vmovsd (%r8,%r11,4),%xmm0 DB 196,99,53,13,200,1 ; vblendpd $0x1,%ymm0,%ymm9,%ymm9 - DB 233,32,255,255,255 ; jmpq c58a <_sk_load_bgra_avx+0x26> + DB 233,32,255,255,255 ; jmpq cb06 <_sk_load_bgra_avx+0x26> DB 196,129,121,110,68,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm0 DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 @@ -11856,9 +12097,9 @@ _sk_load_bgra_avx LABEL PROC DB 196,99,53,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 DB 196,129,121,16,4,152 ; vmovupd (%r8,%r11,4),%xmm0 DB 196,67,125,13,201,12 ; vblendpd $0xc,%ymm9,%ymm0,%ymm9 - DB 233,203,254,255,255 ; jmpq c58a <_sk_load_bgra_avx+0x26> + DB 233,203,254,255,255 ; jmpq cb06 <_sk_load_bgra_avx+0x26> DB 144 ; nop - DB 120,255 ; js c6c1 <_sk_load_bgra_avx+0x15d> + DB 120,255 ; js cc3d <_sk_load_bgra_avx+0x15d> DB 255 ; (bad) DB 255,153,255,255,255,131 ; lcall *-0x7c000001(%rcx) DB 255 ; (bad) @@ -11889,12 +12130,12 @@ _sk_load_bgra_dst_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,135,0,0,0 ; jne c783 <_sk_load_bgra_dst_avx+0xa7> + DB 15,133,135,0,0,0 ; jne ccff <_sk_load_bgra_dst_avx+0xa7> DB 196,1,125,16,12,152 ; vmovupd (%r8,%r11,4),%ymm9 - DB 197,125,40,21,246,6,2,0 ; vmovapd 0x206f6(%rip),%ymm10 # 2ce00 <_sk_overlay_sse2_8bit+0x8a8> + DB 197,125,40,21,122,0,2,0 ; vmovapd 0x2007a(%rip),%ymm10 # 2cd00 <_sk_overlay_sse2_8bit+0x851> DB 196,193,53,84,234 ; vandpd %ymm10,%ymm9,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,98,125,24,5,4,12,2,0 ; vbroadcastss 0x20c04(%rip),%ymm8 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,5,104,5,2,0 ; vbroadcastss 0x20568(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,84,89,240 ; vmulps %ymm8,%ymm5,%ymm6 DB 196,193,81,114,209,8 ; vpsrld $0x8,%xmm9,%xmm5 DB 196,99,125,25,207,1 ; vextractf128 $0x1,%ymm9,%xmm7 @@ -11921,21 +12162,21 @@ _sk_load_bgra_dst_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,102,255,255,255 ; ja c702 <_sk_load_bgra_dst_avx+0x26> + DB 15,135,102,255,255,255 ; ja cc7e <_sk_load_bgra_dst_avx+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # c838 <_sk_load_bgra_dst_avx+0x15c> + DB 76,141,21,145,0,0,0 ; lea 0x91(%rip),%r10 # cdb4 <_sk_load_bgra_dst_avx+0x15c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,122,16,12,152 ; vmovss (%r8,%r11,4),%xmm9 - DB 233,71,255,255,255 ; jmpq c702 <_sk_load_bgra_dst_avx+0x26> + DB 233,71,255,255,255 ; jmpq cc7e <_sk_load_bgra_dst_avx+0x26> DB 196,129,121,110,100,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm4 DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 DB 196,99,85,12,204,4 ; vblendps $0x4,%ymm4,%ymm5,%ymm9 DB 196,129,123,16,36,152 ; vmovsd (%r8,%r11,4),%xmm4 DB 196,99,53,13,204,1 ; vblendpd $0x1,%ymm4,%ymm9,%ymm9 - DB 233,32,255,255,255 ; jmpq c702 <_sk_load_bgra_dst_avx+0x26> + DB 233,32,255,255,255 ; jmpq cc7e <_sk_load_bgra_dst_avx+0x26> DB 196,129,121,110,100,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm4 DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 DB 196,227,125,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm0,%ymm4 @@ -11949,9 +12190,9 @@ _sk_load_bgra_dst_avx LABEL PROC DB 196,99,53,24,204,1 ; vinsertf128 $0x1,%xmm4,%ymm9,%ymm9 DB 196,129,121,16,36,152 ; vmovupd (%r8,%r11,4),%xmm4 DB 196,67,93,13,201,12 ; vblendpd $0xc,%ymm9,%ymm4,%ymm9 - DB 233,203,254,255,255 ; jmpq c702 <_sk_load_bgra_dst_avx+0x26> + DB 233,203,254,255,255 ; jmpq cc7e <_sk_load_bgra_dst_avx+0x26> DB 144 ; nop - DB 120,255 ; js c839 <_sk_load_bgra_dst_avx+0x15d> + DB 120,255 ; js cdb5 <_sk_load_bgra_dst_avx+0x15d> DB 255 ; (bad) DB 255,153,255,255,255,131 ; lcall *-0x7c000001(%rcx) DB 255 ; (bad) @@ -11992,29 +12233,29 @@ _sk_gather_bgra_avx LABEL PROC DB 196,225,249,126,200 ; vmovq %xmm1,%rax DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 + DB 196,193,249,126,194 ; vmovq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx DB 65,137,222 ; mov %ebx,%r14d + DB 72,193,235,32 ; shr $0x20,%rbx DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d - DB 72,193,235,32 ; shr $0x20,%rbx DB 73,193,239,32 ; shr $0x20,%r15 - DB 196,129,121,110,4,176 ; vmovd (%r8,%r14,4),%xmm0 - DB 196,195,121,34,4,152,1 ; vpinsrd $0x1,(%r8,%rbx,4),%xmm0,%xmm0 + DB 196,129,121,110,4,152 ; vmovd (%r8,%r11,4),%xmm0 + DB 196,131,121,34,4,144,1 ; vpinsrd $0x1,(%r8,%r10,4),%xmm0,%xmm0 DB 196,131,121,34,4,160,2 ; vpinsrd $0x2,(%r8,%r12,4),%xmm0,%xmm0 DB 196,3,121,34,4,184,3 ; vpinsrd $0x3,(%r8,%r15,4),%xmm0,%xmm8 DB 196,129,121,110,4,136 ; vmovd (%r8,%r9,4),%xmm0 DB 196,195,121,34,4,128,1 ; vpinsrd $0x1,(%r8,%rax,4),%xmm0,%xmm0 - DB 196,131,121,34,4,152,2 ; vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 - DB 196,131,121,34,28,144,3 ; vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 + DB 196,131,121,34,4,176,2 ; vpinsrd $0x2,(%r8,%r14,4),%xmm0,%xmm0 + DB 196,195,121,34,28,152,3 ; vpinsrd $0x3,(%r8,%rbx,4),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,13,26,5,2,0 ; vmovaps 0x2051a(%rip),%ymm9 # 2ce20 <_sk_overlay_sse2_8bit+0x8c8> + DB 197,124,40,13,158,254,1,0 ; vmovaps 0x1fe9e(%rip),%ymm9 # 2cd20 <_sk_overlay_sse2_8bit+0x871> DB 196,193,124,84,193 ; vandps %ymm9,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,21,8,10,2,0 ; vbroadcastss 0x20a08(%rip),%ymm10 # 2d320 <_sk_overlay_sse2_8bit+0xdc8> + DB 196,98,125,24,21,108,3,2,0 ; vbroadcastss 0x2036c(%rip),%ymm10 # 2d200 <_sk_overlay_sse2_8bit+0xd51> DB 196,193,124,89,210 ; vmulps %ymm10,%ymm0,%ymm2 DB 196,193,121,114,208,8 ; vpsrld $0x8,%xmm8,%xmm0 DB 197,241,114,211,8 ; vpsrld $0x8,%xmm3,%xmm1 @@ -12049,7 +12290,7 @@ _sk_store_bgra_avx LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 196,98,125,24,5,56,9,2,0 ; vbroadcastss 0x20938(%rip),%ymm8 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 196,98,125,24,5,156,2,2,0 ; vbroadcastss 0x2029c(%rip),%ymm8 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 196,65,108,89,200 ; vmulps %ymm8,%ymm2,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -12074,7 +12315,7 @@ _sk_store_bgra_avx LABEL PROC DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8 DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne ca2e <_sk_store_bgra_avx+0xae> + DB 117,10 ; jne cfaa <_sk_store_bgra_avx+0xae> DB 196,1,124,17,4,152 ; vmovups %ymm8,(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -12082,17 +12323,17 @@ _sk_store_bgra_avx LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja ca2a <_sk_store_bgra_avx+0xaa> + DB 119,236 ; ja cfa6 <_sk_store_bgra_avx+0xaa> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # ca9c <_sk_store_bgra_avx+0x11c> + DB 76,141,21,83,0,0,0 ; lea 0x53(%rip),%r10 # d018 <_sk_store_bgra_avx+0x11c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,1,121,126,4,152 ; vmovd %xmm8,(%r8,%r11,4) - DB 235,208 ; jmp ca2a <_sk_store_bgra_avx+0xaa> + DB 235,208 ; jmp cfa6 <_sk_store_bgra_avx+0xaa> DB 196,3,121,22,68,152,8,2 ; vpextrd $0x2,%xmm8,0x8(%r8,%r11,4) DB 196,1,121,214,4,152 ; vmovq %xmm8,(%r8,%r11,4) - DB 235,192 ; jmp ca2a <_sk_store_bgra_avx+0xaa> + DB 235,192 ; jmp cfa6 <_sk_store_bgra_avx+0xaa> DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,3,121,22,76,152,24,2 ; vpextrd $0x2,%xmm9,0x18(%r8,%r11,4) DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 @@ -12100,7 +12341,7 @@ _sk_store_bgra_avx LABEL PROC DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,1,122,17,76,152,16 ; vmovss %xmm9,0x10(%r8,%r11,4) DB 196,1,121,17,4,152 ; vmovupd %xmm8,(%r8,%r11,4) - DB 235,143 ; jmp ca2a <_sk_store_bgra_avx+0xaa> + DB 235,143 ; jmp cfa6 <_sk_store_bgra_avx+0xaa> DB 144 ; nop DB 182,255 ; mov $0xff,%dh DB 255 ; (bad) @@ -12137,128 +12378,109 @@ _sk_load_f16_avx LABEL PROC DB 197,252,17,124,36,96 ; vmovups %ymm7,0x60(%rsp) DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) - DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) - DB 15,133,147,2,0,0 ; jne cd89 <_sk_load_f16_avx+0x2d1> + DB 197,252,17,36,36 ; vmovups %ymm4,(%rsp) + DB 15,133,33,2,0,0 ; jne d293 <_sk_load_f16_avx+0x25f> DB 196,65,121,16,4,192 ; vmovupd (%r8,%rax,8),%xmm8 DB 196,193,121,16,84,192,16 ; vmovupd 0x10(%r8,%rax,8),%xmm2 - DB 196,193,121,16,76,192,32 ; vmovupd 0x20(%r8,%rax,8),%xmm1 + DB 196,193,121,16,92,192,32 ; vmovupd 0x20(%r8,%rax,8),%xmm3 DB 196,65,122,111,76,192,48 ; vmovdqu 0x30(%r8,%rax,8),%xmm9 DB 197,185,97,194 ; vpunpcklwd %xmm2,%xmm8,%xmm0 DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 - DB 196,193,113,97,217 ; vpunpcklwd %xmm9,%xmm1,%xmm3 - DB 196,193,113,105,201 ; vpunpckhwd %xmm9,%xmm1,%xmm1 - DB 197,121,97,250 ; vpunpcklwd %xmm2,%xmm0,%xmm15 + DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 + DB 196,193,97,105,217 ; vpunpckhwd %xmm9,%xmm3,%xmm3 + DB 197,121,97,242 ; vpunpcklwd %xmm2,%xmm0,%xmm14 DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8 - DB 197,225,97,209 ; vpunpcklwd %xmm1,%xmm3,%xmm2 - DB 197,97,105,201 ; vpunpckhwd %xmm1,%xmm3,%xmm9 - DB 197,129,108,194 ; vpunpcklqdq %xmm2,%xmm15,%xmm0 - DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 - DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 + DB 197,113,97,251 ; vpunpcklwd %xmm3,%xmm1,%xmm15 + DB 197,113,105,203 ; vpunpckhwd %xmm3,%xmm1,%xmm9 + DB 196,193,9,108,199 ; vpunpcklqdq %xmm15,%xmm14,%xmm0 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,98,125,24,37,213,8,2,0 ; vbroadcastss 0x208d5(%rip),%ymm12 # 2d428 <_sk_overlay_sse2_8bit+0xed0> - DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 - DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 - DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,153,8,2,0 ; vbroadcastss 0x20899(%rip),%xmm11 # 2d404 <_sk_overlay_sse2_8bit+0xeac> - DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,179,8,2,0 ; vbroadcastss 0x208b3(%rip),%xmm13 # 2d42c <_sk_overlay_sse2_8bit+0xed4> - DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 - DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 - DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 - DB 196,99,45,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm10,%ymm10 - DB 197,225,114,241,16 ; vpslld $0x10,%xmm1,%xmm3 - DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 + DB 197,123,18,21,253,10,2,0 ; vmovddup 0x20afd(%rip),%xmm10 # 2dbc8 <_sk_overlay_sse2_8bit+0x1719> + DB 196,193,121,219,218 ; vpand %xmm10,%xmm0,%xmm3 + DB 196,98,121,24,29,47,2,2,0 ; vbroadcastss 0x2022f(%rip),%xmm11 # 2d308 <_sk_overlay_sse2_8bit+0xe59> + DB 197,33,102,227 ; vpcmpgtd %xmm3,%xmm11,%xmm12 + DB 196,193,113,219,210 ; vpand %xmm10,%xmm1,%xmm2 + DB 197,33,102,234 ; vpcmpgtd %xmm2,%xmm11,%xmm13 + DB 196,195,21,24,252,1 ; vinsertf128 $0x1,%xmm12,%ymm13,%ymm7 + DB 197,123,18,37,220,10,2,0 ; vmovddup 0x20adc(%rip),%xmm12 # 2dbd0 <_sk_overlay_sse2_8bit+0x1721> + DB 196,193,113,219,204 ; vpand %xmm12,%xmm1,%xmm1 DB 197,241,114,241,16 ; vpslld $0x10,%xmm1,%xmm1 - DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 - DB 197,249,114,240,13 ; vpslld $0xd,%xmm0,%xmm0 - DB 196,193,97,114,246,13 ; vpslld $0xd,%xmm14,%xmm3 - DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 - DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0 - DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 - DB 196,226,121,24,29,105,8,2,0 ; vbroadcastss 0x20869(%rip),%xmm3 # 2d430 <_sk_overlay_sse2_8bit+0xed8> - DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 + DB 196,193,121,219,196 ; vpand %xmm12,%xmm0,%xmm0 + DB 197,249,114,240,16 ; vpslld $0x10,%xmm0,%xmm0 + DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 + DB 197,241,235,202 ; vpor %xmm2,%xmm1,%xmm1 + DB 197,233,114,243,13 ; vpslld $0xd,%xmm3,%xmm2 + DB 197,249,235,194 ; vpor %xmm2,%xmm0,%xmm0 + DB 196,226,121,24,29,233,1,2,0 ; vbroadcastss 0x201e9(%rip),%xmm3 # 2d30c <_sk_overlay_sse2_8bit+0xe5d> DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 - DB 196,195,125,74,198,160 ; vblendvps %ymm10,%ymm14,%ymm0,%ymm0 - DB 197,129,109,202 ; vpunpckhqdq %xmm2,%xmm15,%xmm1 - DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 - DB 197,241,105,212 ; vpunpckhwd %xmm4,%xmm1,%xmm2 - DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 - DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,193,116,84,212 ; vandps %ymm12,%ymm1,%ymm2 - DB 197,244,87,202 ; vxorps %ymm2,%ymm1,%ymm1 - DB 196,195,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm10 - DB 196,193,40,87,251 ; vxorps %xmm11,%xmm10,%xmm7 - DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 - DB 196,193,112,87,243 ; vxorps %xmm11,%xmm1,%xmm6 - DB 197,145,102,246 ; vpcmpgtd %xmm6,%xmm13,%xmm6 - DB 196,227,77,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 - DB 197,193,114,242,16 ; vpslld $0x10,%xmm2,%xmm7 - DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 + DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 + DB 196,227,125,74,196,112 ; vblendvps %ymm7,%ymm4,%ymm0,%ymm0 + DB 196,193,9,109,207 ; vpunpckhqdq %xmm15,%xmm14,%xmm1 + DB 196,226,121,51,209 ; vpmovzxwd %xmm1,%xmm2 + DB 197,249,112,201,78 ; vpshufd $0x4e,%xmm1,%xmm1 + DB 196,98,121,51,249 ; vpmovzxwd %xmm1,%xmm15 + DB 196,193,1,219,250 ; vpand %xmm10,%xmm15,%xmm7 + DB 197,33,102,247 ; vpcmpgtd %xmm7,%xmm11,%xmm14 + DB 196,193,105,219,202 ; vpand %xmm10,%xmm2,%xmm1 + DB 197,33,102,233 ; vpcmpgtd %xmm1,%xmm11,%xmm13 + DB 196,67,21,24,238,1 ; vinsertf128 $0x1,%xmm14,%ymm13,%ymm13 + DB 196,193,105,219,212 ; vpand %xmm12,%xmm2,%xmm2 DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 - DB 196,227,69,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm7,%ymm2 DB 197,241,114,241,13 ; vpslld $0xd,%xmm1,%xmm1 - DB 196,193,65,114,242,13 ; vpslld $0xd,%xmm10,%xmm7 - DB 196,227,117,24,207,1 ; vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 - DB 197,244,86,202 ; vorps %ymm2,%ymm1,%ymm1 - DB 196,227,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm2 + DB 197,233,235,201 ; vpor %xmm1,%xmm2,%xmm1 + DB 196,193,1,219,212 ; vpand %xmm12,%xmm15,%xmm2 + DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 + DB 197,193,114,247,13 ; vpslld $0xd,%xmm7,%xmm7 + DB 197,233,235,215 ; vpor %xmm7,%xmm2,%xmm2 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,195,117,74,206,96 ; vblendvps %ymm6,%ymm14,%ymm1,%ymm1 + DB 196,227,117,74,204,208 ; vblendvps %ymm13,%ymm4,%ymm1,%ymm1 DB 196,193,57,108,209 ; vpunpcklqdq %xmm9,%xmm8,%xmm2 - DB 197,233,105,244 ; vpunpckhwd %xmm4,%xmm2,%xmm6 - DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - DB 196,193,108,84,244 ; vandps %ymm12,%ymm2,%ymm6 - DB 197,236,87,214 ; vxorps %ymm6,%ymm2,%ymm2 - DB 196,227,125,25,215,1 ; vextractf128 $0x1,%ymm2,%xmm7 - DB 196,193,64,87,235 ; vxorps %xmm11,%xmm7,%xmm5 - DB 197,145,102,237 ; vpcmpgtd %xmm5,%xmm13,%xmm5 - DB 196,193,104,87,227 ; vxorps %xmm11,%xmm2,%xmm4 - DB 197,145,102,228 ; vpcmpgtd %xmm4,%xmm13,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 197,209,114,246,16 ; vpslld $0x10,%xmm6,%xmm5 - DB 196,227,125,25,246,1 ; vextractf128 $0x1,%ymm6,%xmm6 - DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 - DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 + DB 196,226,121,51,250 ; vpmovzxwd %xmm2,%xmm7 + DB 197,249,112,210,78 ; vpshufd $0x4e,%xmm2,%xmm2 + DB 196,98,121,51,234 ; vpmovzxwd %xmm2,%xmm13 + DB 196,65,17,219,250 ; vpand %xmm10,%xmm13,%xmm15 + DB 196,65,33,102,247 ; vpcmpgtd %xmm15,%xmm11,%xmm14 + DB 196,193,65,219,210 ; vpand %xmm10,%xmm7,%xmm2 + DB 197,161,102,242 ; vpcmpgtd %xmm2,%xmm11,%xmm6 + DB 196,195,77,24,246,1 ; vinsertf128 $0x1,%xmm14,%ymm6,%ymm6 + DB 196,193,65,219,252 ; vpand %xmm12,%xmm7,%xmm7 + DB 197,193,114,247,16 ; vpslld $0x10,%xmm7,%xmm7 DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 - DB 197,201,114,247,13 ; vpslld $0xd,%xmm7,%xmm6 - DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - DB 197,236,86,213 ; vorps %ymm5,%ymm2,%ymm2 - DB 196,227,125,25,213,1 ; vextractf128 $0x1,%ymm2,%xmm5 + DB 197,193,235,210 ; vpor %xmm2,%xmm7,%xmm2 + DB 196,193,17,219,252 ; vpand %xmm12,%xmm13,%xmm7 + DB 197,193,114,247,16 ; vpslld $0x10,%xmm7,%xmm7 + DB 196,193,81,114,247,13 ; vpslld $0xd,%xmm15,%xmm5 + DB 197,193,235,237 ; vpor %xmm5,%xmm7,%xmm5 DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 196,227,109,24,213,1 ; vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 - DB 196,195,109,74,214,64 ; vblendvps %ymm4,%ymm14,%ymm2,%ymm2 - DB 196,193,57,109,225 ; vpunpckhqdq %xmm9,%xmm8,%xmm4 - DB 196,193,89,105,234 ; vpunpckhwd %xmm10,%xmm4,%xmm5 - DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 196,193,92,84,236 ; vandps %ymm12,%ymm4,%ymm5 - DB 197,220,87,229 ; vxorps %ymm5,%ymm4,%ymm4 - DB 196,227,125,25,230,1 ; vextractf128 $0x1,%ymm4,%xmm6 - DB 196,193,72,87,251 ; vxorps %xmm11,%xmm6,%xmm7 - DB 197,17,102,199 ; vpcmpgtd %xmm7,%xmm13,%xmm8 - DB 196,193,88,87,251 ; vxorps %xmm11,%xmm4,%xmm7 - DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 - DB 196,195,69,24,248,1 ; vinsertf128 $0x1,%xmm8,%ymm7,%ymm7 - DB 197,185,114,245,16 ; vpslld $0x10,%xmm5,%xmm8 - DB 196,227,125,25,237,1 ; vextractf128 $0x1,%ymm5,%xmm5 + DB 196,227,109,74,212,96 ; vblendvps %ymm6,%ymm4,%ymm2,%ymm2 + DB 196,193,57,109,233 ; vpunpckhqdq %xmm9,%xmm8,%xmm5 + DB 196,226,121,51,245 ; vpmovzxwd %xmm5,%xmm6 + DB 197,249,112,237,78 ; vpshufd $0x4e,%xmm5,%xmm5 + DB 196,226,121,51,237 ; vpmovzxwd %xmm5,%xmm5 + DB 196,65,81,219,202 ; vpand %xmm10,%xmm5,%xmm9 + DB 196,65,33,102,193 ; vpcmpgtd %xmm9,%xmm11,%xmm8 + DB 196,193,73,219,250 ; vpand %xmm10,%xmm6,%xmm7 + DB 197,33,102,215 ; vpcmpgtd %xmm7,%xmm11,%xmm10 + DB 196,67,45,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm10,%ymm8 + DB 196,193,73,219,244 ; vpand %xmm12,%xmm6,%xmm6 + DB 196,193,81,219,236 ; vpand %xmm12,%xmm5,%xmm5 + DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 + DB 197,193,114,247,13 ; vpslld $0xd,%xmm7,%xmm7 + DB 197,201,235,247 ; vpor %xmm7,%xmm6,%xmm6 DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 - DB 196,227,61,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm8,%ymm5 - DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 - DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 - DB 196,227,93,24,230,1 ; vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 - DB 197,220,86,229 ; vorps %ymm5,%ymm4,%ymm4 - DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 + DB 196,193,65,114,241,13 ; vpslld $0xd,%xmm9,%xmm7 + DB 197,209,235,239 ; vpor %xmm7,%xmm5,%xmm5 DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 - DB 197,217,254,219 ; vpaddd %xmm3,%xmm4,%xmm3 + DB 197,201,254,219 ; vpaddd %xmm3,%xmm6,%xmm3 DB 196,227,101,24,221,1 ; vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 - DB 196,195,101,74,222,112 ; vblendvps %ymm7,%ymm14,%ymm3,%ymm3 + DB 196,227,101,74,220,128 ; vblendvps %ymm8,%ymm4,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 @@ -12269,29 +12491,29 @@ _sk_load_f16_avx LABEL PROC DB 196,65,123,16,4,192 ; vmovsd (%r8,%rax,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je cdef <_sk_load_f16_avx+0x337> + DB 116,85 ; je d2f9 <_sk_load_f16_avx+0x2c5> DB 196,65,57,22,68,192,8 ; vmovhpd 0x8(%r8,%rax,8),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb cdef <_sk_load_f16_avx+0x337> + DB 114,72 ; jb d2f9 <_sk_load_f16_avx+0x2c5> DB 196,193,123,16,84,192,16 ; vmovsd 0x10(%r8,%rax,8),%xmm2 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je cdfc <_sk_load_f16_avx+0x344> + DB 116,72 ; je d306 <_sk_load_f16_avx+0x2d2> DB 196,193,105,22,84,192,24 ; vmovhpd 0x18(%r8,%rax,8),%xmm2,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb cdfc <_sk_load_f16_avx+0x344> - DB 196,193,123,16,76,192,32 ; vmovsd 0x20(%r8,%rax,8),%xmm1 + DB 114,59 ; jb d306 <_sk_load_f16_avx+0x2d2> + DB 196,193,123,16,92,192,32 ; vmovsd 0x20(%r8,%rax,8),%xmm3 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,63,253,255,255 ; je cb11 <_sk_load_f16_avx+0x59> - DB 196,193,113,22,76,192,40 ; vmovhpd 0x28(%r8,%rax,8),%xmm1,%xmm1 + DB 15,132,177,253,255,255 ; je d08d <_sk_load_f16_avx+0x59> + DB 196,193,97,22,92,192,40 ; vmovhpd 0x28(%r8,%rax,8),%xmm3,%xmm3 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,46,253,255,255 ; jb cb11 <_sk_load_f16_avx+0x59> + DB 15,130,160,253,255,255 ; jb d08d <_sk_load_f16_avx+0x59> DB 196,65,122,126,76,192,48 ; vmovq 0x30(%r8,%rax,8),%xmm9 - DB 233,34,253,255,255 ; jmpq cb11 <_sk_load_f16_avx+0x59> - DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 + DB 233,148,253,255,255 ; jmpq d08d <_sk_load_f16_avx+0x59> + DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,21,253,255,255 ; jmpq cb11 <_sk_load_f16_avx+0x59> - DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 - DB 233,12,253,255,255 ; jmpq cb11 <_sk_load_f16_avx+0x59> + DB 233,135,253,255,255 ; jmpq d08d <_sk_load_f16_avx+0x59> + DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 + DB 233,126,253,255,255 ; jmpq d08d <_sk_load_f16_avx+0x59> PUBLIC _sk_load_f16_dst_avx _sk_load_f16_dst_avx LABEL PROC @@ -12307,128 +12529,109 @@ _sk_load_f16_dst_avx LABEL PROC DB 197,252,17,92,36,96 ; vmovups %ymm3,0x60(%rsp) DB 197,252,17,84,36,64 ; vmovups %ymm2,0x40(%rsp) DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp) - DB 197,254,127,4,36 ; vmovdqu %ymm0,(%rsp) - DB 15,133,147,2,0,0 ; jne d0d6 <_sk_load_f16_dst_avx+0x2d1> + DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) + DB 15,133,33,2,0,0 ; jne d56e <_sk_load_f16_dst_avx+0x25f> DB 196,65,121,16,4,192 ; vmovupd (%r8,%rax,8),%xmm8 DB 196,193,121,16,116,192,16 ; vmovupd 0x10(%r8,%rax,8),%xmm6 - DB 196,193,121,16,108,192,32 ; vmovupd 0x20(%r8,%rax,8),%xmm5 + DB 196,193,121,16,124,192,32 ; vmovupd 0x20(%r8,%rax,8),%xmm7 DB 196,65,122,111,76,192,48 ; vmovdqu 0x30(%r8,%rax,8),%xmm9 DB 197,185,97,230 ; vpunpcklwd %xmm6,%xmm8,%xmm4 DB 197,185,105,246 ; vpunpckhwd %xmm6,%xmm8,%xmm6 - DB 196,193,81,97,249 ; vpunpcklwd %xmm9,%xmm5,%xmm7 - DB 196,193,81,105,233 ; vpunpckhwd %xmm9,%xmm5,%xmm5 - DB 197,89,97,254 ; vpunpcklwd %xmm6,%xmm4,%xmm15 + DB 196,193,65,97,233 ; vpunpcklwd %xmm9,%xmm7,%xmm5 + DB 196,193,65,105,249 ; vpunpckhwd %xmm9,%xmm7,%xmm7 + DB 197,89,97,246 ; vpunpcklwd %xmm6,%xmm4,%xmm14 DB 197,89,105,198 ; vpunpckhwd %xmm6,%xmm4,%xmm8 - DB 197,193,97,245 ; vpunpcklwd %xmm5,%xmm7,%xmm6 - DB 197,65,105,205 ; vpunpckhwd %xmm5,%xmm7,%xmm9 - DB 197,129,108,230 ; vpunpcklqdq %xmm6,%xmm15,%xmm4 - DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 197,217,105,232 ; vpunpckhwd %xmm0,%xmm4,%xmm5 + DB 197,81,97,255 ; vpunpcklwd %xmm7,%xmm5,%xmm15 + DB 197,81,105,207 ; vpunpckhwd %xmm7,%xmm5,%xmm9 + DB 196,193,9,108,231 ; vpunpcklqdq %xmm15,%xmm14,%xmm4 + DB 196,226,121,51,236 ; vpmovzxwd %xmm4,%xmm5 + DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 196,98,125,24,37,136,5,2,0 ; vbroadcastss 0x20588(%rip),%ymm12 # 2d428 <_sk_overlay_sse2_8bit+0xed0> - DB 196,193,92,84,236 ; vandps %ymm12,%ymm4,%ymm5 - DB 197,220,87,229 ; vxorps %ymm5,%ymm4,%ymm4 - DB 196,195,125,25,230,1 ; vextractf128 $0x1,%ymm4,%xmm14 - DB 196,98,121,24,29,76,5,2,0 ; vbroadcastss 0x2054c(%rip),%xmm11 # 2d404 <_sk_overlay_sse2_8bit+0xeac> - DB 196,193,8,87,251 ; vxorps %xmm11,%xmm14,%xmm7 - DB 196,98,121,24,45,102,5,2,0 ; vbroadcastss 0x20566(%rip),%xmm13 # 2d42c <_sk_overlay_sse2_8bit+0xed4> - DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 - DB 196,65,88,87,211 ; vxorps %xmm11,%xmm4,%xmm10 - DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 - DB 196,99,45,24,215,1 ; vinsertf128 $0x1,%xmm7,%ymm10,%ymm10 - DB 197,193,114,245,16 ; vpslld $0x10,%xmm5,%xmm7 - DB 196,227,125,25,237,1 ; vextractf128 $0x1,%ymm5,%xmm5 + DB 197,123,18,21,34,8,2,0 ; vmovddup 0x20822(%rip),%xmm10 # 2dbc8 <_sk_overlay_sse2_8bit+0x1719> + DB 196,193,89,219,250 ; vpand %xmm10,%xmm4,%xmm7 + DB 196,98,121,24,29,84,255,1,0 ; vbroadcastss 0x1ff54(%rip),%xmm11 # 2d308 <_sk_overlay_sse2_8bit+0xe59> + DB 197,33,102,231 ; vpcmpgtd %xmm7,%xmm11,%xmm12 + DB 196,193,81,219,242 ; vpand %xmm10,%xmm5,%xmm6 + DB 197,33,102,238 ; vpcmpgtd %xmm6,%xmm11,%xmm13 + DB 196,195,21,24,220,1 ; vinsertf128 $0x1,%xmm12,%ymm13,%ymm3 + DB 197,123,18,37,1,8,2,0 ; vmovddup 0x20801(%rip),%xmm12 # 2dbd0 <_sk_overlay_sse2_8bit+0x1721> + DB 196,193,81,219,236 ; vpand %xmm12,%xmm5,%xmm5 DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 - DB 196,227,69,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm7,%ymm5 - DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 - DB 196,193,65,114,246,13 ; vpslld $0xd,%xmm14,%xmm7 - DB 196,227,93,24,231,1 ; vinsertf128 $0x1,%xmm7,%ymm4,%ymm4 - DB 197,220,86,229 ; vorps %ymm5,%ymm4,%ymm4 - DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 - DB 196,226,121,24,61,28,5,2,0 ; vbroadcastss 0x2051c(%rip),%xmm7 # 2d430 <_sk_overlay_sse2_8bit+0xed8> - DB 197,209,254,239 ; vpaddd %xmm7,%xmm5,%xmm5 + DB 196,193,89,219,228 ; vpand %xmm12,%xmm4,%xmm4 + DB 197,217,114,244,16 ; vpslld $0x10,%xmm4,%xmm4 + DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 + DB 197,209,235,238 ; vpor %xmm6,%xmm5,%xmm5 + DB 197,201,114,247,13 ; vpslld $0xd,%xmm7,%xmm6 + DB 197,217,235,230 ; vpor %xmm6,%xmm4,%xmm4 + DB 196,226,121,24,61,14,255,1,0 ; vbroadcastss 0x1ff0e(%rip),%xmm7 # 2d30c <_sk_overlay_sse2_8bit+0xe5d> DB 197,217,254,231 ; vpaddd %xmm7,%xmm4,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 - DB 196,195,93,74,230,160 ; vblendvps %ymm10,%ymm14,%ymm4,%ymm4 - DB 197,129,109,238 ; vpunpckhqdq %xmm6,%xmm15,%xmm5 - DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 197,209,105,240 ; vpunpckhwd %xmm0,%xmm5,%xmm6 - DB 196,226,121,51,237 ; vpmovzxwd %xmm5,%xmm5 - DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 - DB 196,193,84,84,244 ; vandps %ymm12,%ymm5,%ymm6 - DB 197,212,87,238 ; vxorps %ymm6,%ymm5,%ymm5 - DB 196,195,125,25,234,1 ; vextractf128 $0x1,%ymm5,%xmm10 - DB 196,193,40,87,219 ; vxorps %xmm11,%xmm10,%xmm3 - DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 - DB 196,193,80,87,211 ; vxorps %xmm11,%xmm5,%xmm2 - DB 197,145,102,210 ; vpcmpgtd %xmm2,%xmm13,%xmm2 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 - DB 197,225,114,246,16 ; vpslld $0x10,%xmm6,%xmm3 - DB 196,227,125,25,246,1 ; vextractf128 $0x1,%ymm6,%xmm6 - DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 - DB 196,227,101,24,222,1 ; vinsertf128 $0x1,%xmm6,%ymm3,%ymm3 - DB 197,209,114,245,13 ; vpslld $0xd,%xmm5,%xmm5 - DB 196,193,73,114,242,13 ; vpslld $0xd,%xmm10,%xmm6 - DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 - DB 197,212,86,219 ; vorps %ymm3,%ymm5,%ymm3 - DB 196,227,125,25,221,1 ; vextractf128 $0x1,%ymm3,%xmm5 + DB 197,209,254,239 ; vpaddd %xmm7,%xmm5,%xmm5 + DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 + DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 + DB 196,227,93,74,224,48 ; vblendvps %ymm3,%ymm0,%ymm4,%ymm4 + DB 196,193,9,109,223 ; vpunpckhqdq %xmm15,%xmm14,%xmm3 + DB 196,226,121,51,235 ; vpmovzxwd %xmm3,%xmm5 + DB 197,249,112,219,78 ; vpshufd $0x4e,%xmm3,%xmm3 + DB 196,98,121,51,251 ; vpmovzxwd %xmm3,%xmm15 + DB 196,193,1,219,242 ; vpand %xmm10,%xmm15,%xmm6 + DB 197,33,102,246 ; vpcmpgtd %xmm6,%xmm11,%xmm14 + DB 196,193,81,219,218 ; vpand %xmm10,%xmm5,%xmm3 + DB 197,33,102,235 ; vpcmpgtd %xmm3,%xmm11,%xmm13 + DB 196,67,21,24,238,1 ; vinsertf128 $0x1,%xmm14,%ymm13,%ymm13 + DB 196,193,81,219,236 ; vpand %xmm12,%xmm5,%xmm5 + DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 + DB 197,225,114,243,13 ; vpslld $0xd,%xmm3,%xmm3 + DB 197,209,235,219 ; vpor %xmm3,%xmm5,%xmm3 + DB 196,193,1,219,236 ; vpand %xmm12,%xmm15,%xmm5 + DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 + DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 + DB 197,209,235,238 ; vpor %xmm6,%xmm5,%xmm5 DB 197,209,254,239 ; vpaddd %xmm7,%xmm5,%xmm5 DB 197,225,254,223 ; vpaddd %xmm7,%xmm3,%xmm3 DB 196,227,101,24,221,1 ; vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 - DB 196,195,101,74,238,32 ; vblendvps %ymm2,%ymm14,%ymm3,%ymm5 - DB 196,193,57,108,209 ; vpunpcklqdq %xmm9,%xmm8,%xmm2 - DB 197,233,105,216 ; vpunpckhwd %xmm0,%xmm2,%xmm3 - DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 - DB 196,193,108,84,220 ; vandps %ymm12,%ymm2,%ymm3 - DB 197,236,87,211 ; vxorps %ymm3,%ymm2,%ymm2 - DB 196,227,125,25,214,1 ; vextractf128 $0x1,%ymm2,%xmm6 - DB 196,193,72,87,203 ; vxorps %xmm11,%xmm6,%xmm1 - DB 197,145,102,201 ; vpcmpgtd %xmm1,%xmm13,%xmm1 - DB 196,193,104,87,195 ; vxorps %xmm11,%xmm2,%xmm0 - DB 197,145,102,192 ; vpcmpgtd %xmm0,%xmm13,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 197,241,114,243,16 ; vpslld $0x10,%xmm3,%xmm1 - DB 196,227,125,25,219,1 ; vextractf128 $0x1,%ymm3,%xmm3 - DB 197,225,114,243,16 ; vpslld $0x10,%xmm3,%xmm3 - DB 196,227,117,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm1,%ymm1 - DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 - DB 197,225,114,246,13 ; vpslld $0xd,%xmm6,%xmm3 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 - DB 197,236,86,201 ; vorps %ymm1,%ymm2,%ymm1 - DB 196,227,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm2 - DB 197,233,254,215 ; vpaddd %xmm7,%xmm2,%xmm2 + DB 196,227,101,74,232,208 ; vblendvps %ymm13,%ymm0,%ymm3,%ymm5 + DB 196,193,57,108,217 ; vpunpcklqdq %xmm9,%xmm8,%xmm3 + DB 196,226,121,51,243 ; vpmovzxwd %xmm3,%xmm6 + DB 197,249,112,219,78 ; vpshufd $0x4e,%xmm3,%xmm3 + DB 196,98,121,51,235 ; vpmovzxwd %xmm3,%xmm13 + DB 196,65,17,219,250 ; vpand %xmm10,%xmm13,%xmm15 + DB 196,65,33,102,247 ; vpcmpgtd %xmm15,%xmm11,%xmm14 + DB 196,193,73,219,218 ; vpand %xmm10,%xmm6,%xmm3 + DB 197,161,102,211 ; vpcmpgtd %xmm3,%xmm11,%xmm2 + DB 196,195,109,24,214,1 ; vinsertf128 $0x1,%xmm14,%ymm2,%ymm2 + DB 196,193,73,219,244 ; vpand %xmm12,%xmm6,%xmm6 + DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 + DB 197,225,114,243,13 ; vpslld $0xd,%xmm3,%xmm3 + DB 197,201,235,219 ; vpor %xmm3,%xmm6,%xmm3 + DB 196,193,17,219,244 ; vpand %xmm12,%xmm13,%xmm6 + DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 + DB 196,193,113,114,247,13 ; vpslld $0xd,%xmm15,%xmm1 + DB 197,201,235,201 ; vpor %xmm1,%xmm6,%xmm1 DB 197,241,254,207 ; vpaddd %xmm7,%xmm1,%xmm1 - DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,195,117,74,246,0 ; vblendvps %ymm0,%ymm14,%ymm1,%ymm6 - DB 196,193,57,109,193 ; vpunpckhqdq %xmm9,%xmm8,%xmm0 - DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 - DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 - DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 - DB 196,227,125,25,194,1 ; vextractf128 $0x1,%ymm0,%xmm2 - DB 196,193,104,87,219 ; vxorps %xmm11,%xmm2,%xmm3 - DB 197,17,102,195 ; vpcmpgtd %xmm3,%xmm13,%xmm8 - DB 196,193,120,87,219 ; vxorps %xmm11,%xmm0,%xmm3 - DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 - DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 - DB 197,185,114,241,16 ; vpslld $0x10,%xmm1,%xmm8 - DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 + DB 197,225,254,223 ; vpaddd %xmm7,%xmm3,%xmm3 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 + DB 196,227,117,74,240,32 ; vblendvps %ymm2,%ymm0,%ymm1,%ymm6 + DB 196,193,57,109,201 ; vpunpckhqdq %xmm9,%xmm8,%xmm1 + DB 196,226,121,51,209 ; vpmovzxwd %xmm1,%xmm2 + DB 197,249,112,201,78 ; vpshufd $0x4e,%xmm1,%xmm1 + DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 + DB 196,65,113,219,202 ; vpand %xmm10,%xmm1,%xmm9 + DB 196,65,33,102,193 ; vpcmpgtd %xmm9,%xmm11,%xmm8 + DB 196,193,105,219,218 ; vpand %xmm10,%xmm2,%xmm3 + DB 197,33,102,211 ; vpcmpgtd %xmm3,%xmm11,%xmm10 + DB 196,67,45,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm10,%ymm8 + DB 196,193,105,219,212 ; vpand %xmm12,%xmm2,%xmm2 + DB 196,193,113,219,204 ; vpand %xmm12,%xmm1,%xmm1 + DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 + DB 197,225,114,243,13 ; vpslld $0xd,%xmm3,%xmm3 + DB 197,233,235,211 ; vpor %xmm3,%xmm2,%xmm2 DB 197,241,114,241,16 ; vpslld $0x10,%xmm1,%xmm1 - DB 196,227,61,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 - DB 197,249,114,240,13 ; vpslld $0xd,%xmm0,%xmm0 - DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 - DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0 - DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 + DB 196,193,97,114,241,13 ; vpslld $0xd,%xmm9,%xmm3 + DB 197,241,235,203 ; vpor %xmm3,%xmm1,%xmm1 DB 197,241,254,207 ; vpaddd %xmm7,%xmm1,%xmm1 - DB 197,249,254,199 ; vpaddd %xmm7,%xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,195,125,74,254,48 ; vblendvps %ymm3,%ymm14,%ymm0,%ymm7 + DB 197,233,254,215 ; vpaddd %xmm7,%xmm2,%xmm2 + DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 + DB 196,227,117,74,248,128 ; vblendvps %ymm8,%ymm0,%ymm1,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 DB 197,252,16,76,36,32 ; vmovups 0x20(%rsp),%ymm1 @@ -12439,29 +12642,29 @@ _sk_load_f16_dst_avx LABEL PROC DB 196,65,123,16,4,192 ; vmovsd (%r8,%rax,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je d13c <_sk_load_f16_dst_avx+0x337> + DB 116,85 ; je d5d4 <_sk_load_f16_dst_avx+0x2c5> DB 196,65,57,22,68,192,8 ; vmovhpd 0x8(%r8,%rax,8),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb d13c <_sk_load_f16_dst_avx+0x337> + DB 114,72 ; jb d5d4 <_sk_load_f16_dst_avx+0x2c5> DB 196,193,123,16,116,192,16 ; vmovsd 0x10(%r8,%rax,8),%xmm6 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je d149 <_sk_load_f16_dst_avx+0x344> + DB 116,72 ; je d5e1 <_sk_load_f16_dst_avx+0x2d2> DB 196,193,73,22,116,192,24 ; vmovhpd 0x18(%r8,%rax,8),%xmm6,%xmm6 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb d149 <_sk_load_f16_dst_avx+0x344> - DB 196,193,123,16,108,192,32 ; vmovsd 0x20(%r8,%rax,8),%xmm5 + DB 114,59 ; jb d5e1 <_sk_load_f16_dst_avx+0x2d2> + DB 196,193,123,16,124,192,32 ; vmovsd 0x20(%r8,%rax,8),%xmm7 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,63,253,255,255 ; je ce5e <_sk_load_f16_dst_avx+0x59> - DB 196,193,81,22,108,192,40 ; vmovhpd 0x28(%r8,%rax,8),%xmm5,%xmm5 + DB 15,132,177,253,255,255 ; je d368 <_sk_load_f16_dst_avx+0x59> + DB 196,193,65,22,124,192,40 ; vmovhpd 0x28(%r8,%rax,8),%xmm7,%xmm7 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,46,253,255,255 ; jb ce5e <_sk_load_f16_dst_avx+0x59> + DB 15,130,160,253,255,255 ; jb d368 <_sk_load_f16_dst_avx+0x59> DB 196,65,122,126,76,192,48 ; vmovq 0x30(%r8,%rax,8),%xmm9 - DB 233,34,253,255,255 ; jmpq ce5e <_sk_load_f16_dst_avx+0x59> - DB 197,209,87,237 ; vxorpd %xmm5,%xmm5,%xmm5 + DB 233,148,253,255,255 ; jmpq d368 <_sk_load_f16_dst_avx+0x59> + DB 197,193,87,255 ; vxorpd %xmm7,%xmm7,%xmm7 DB 197,201,87,246 ; vxorpd %xmm6,%xmm6,%xmm6 - DB 233,21,253,255,255 ; jmpq ce5e <_sk_load_f16_dst_avx+0x59> - DB 197,209,87,237 ; vxorpd %xmm5,%xmm5,%xmm5 - DB 233,12,253,255,255 ; jmpq ce5e <_sk_load_f16_dst_avx+0x59> + DB 233,135,253,255,255 ; jmpq d368 <_sk_load_f16_dst_avx+0x59> + DB 197,193,87,255 ; vxorpd %xmm7,%xmm7,%xmm7 + DB 233,126,253,255,255 ; jmpq d368 <_sk_load_f16_dst_avx+0x59> PUBLIC _sk_gather_f16_avx _sk_gather_f16_avx LABEL PROC @@ -12469,11 +12672,11 @@ _sk_gather_f16_avx LABEL PROC DB 65,86 ; push %r14 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp - DB 197,252,17,124,36,96 ; vmovups %ymm7,0x60(%rsp) - DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) - DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) - DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) + DB 72,131,236,120 ; sub $0x78,%rsp + DB 197,252,17,124,36,64 ; vmovups %ymm7,0x40(%rsp) + DB 197,252,17,116,36,32 ; vmovups %ymm6,0x20(%rsp) + DB 197,252,17,44,36 ; vmovups %ymm5,(%rsp) + DB 197,252,40,236 ; vmovaps %ymm4,%ymm5 DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 @@ -12492,146 +12695,127 @@ _sk_gather_f16_avx LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 - DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx DB 65,137,222 ; mov %ebx,%r14d - DB 196,195,249,22,199,1 ; vpextrq $0x1,%xmm0,%r15 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,199 ; vmovq %xmm0,%r15 DB 69,137,252 ; mov %r15d,%r12d DB 73,193,239,32 ; shr $0x20,%r15 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,193,122,126,4,216 ; vmovq (%r8,%rbx,8),%xmm0 - DB 196,129,122,126,12,240 ; vmovq (%r8,%r14,8),%xmm1 + DB 196,129,122,126,4,248 ; vmovq (%r8,%r15,8),%xmm0 + DB 196,129,122,126,12,224 ; vmovq (%r8,%r12,8),%xmm1 DB 197,113,108,200 ; vpunpcklqdq %xmm0,%xmm1,%xmm9 - DB 196,129,122,126,12,248 ; vmovq (%r8,%r15,8),%xmm1 - DB 196,129,122,126,20,224 ; vmovq (%r8,%r12,8),%xmm2 + DB 196,193,122,126,12,216 ; vmovq (%r8,%rbx,8),%xmm1 + DB 196,129,122,126,20,240 ; vmovq (%r8,%r14,8),%xmm2 DB 197,233,108,201 ; vpunpcklqdq %xmm1,%xmm2,%xmm1 DB 196,129,122,126,20,208 ; vmovq (%r8,%r10,8),%xmm2 DB 196,129,122,126,28,216 ; vmovq (%r8,%r11,8),%xmm3 - DB 197,97,108,210 ; vpunpcklqdq %xmm2,%xmm3,%xmm10 + DB 197,225,108,210 ; vpunpcklqdq %xmm2,%xmm3,%xmm2 DB 196,65,122,126,4,192 ; vmovq (%r8,%rax,8),%xmm8 DB 196,129,122,126,28,200 ; vmovq (%r8,%r9,8),%xmm3 DB 196,193,97,108,216 ; vpunpcklqdq %xmm8,%xmm3,%xmm3 DB 197,177,97,193 ; vpunpcklwd %xmm1,%xmm9,%xmm0 DB 197,177,105,201 ; vpunpckhwd %xmm1,%xmm9,%xmm1 - DB 197,169,97,211 ; vpunpcklwd %xmm3,%xmm10,%xmm2 - DB 197,169,105,219 ; vpunpckhwd %xmm3,%xmm10,%xmm3 - DB 197,121,97,249 ; vpunpcklwd %xmm1,%xmm0,%xmm15 + DB 197,105,97,203 ; vpunpcklwd %xmm3,%xmm2,%xmm9 + DB 197,233,105,211 ; vpunpckhwd %xmm3,%xmm2,%xmm2 + DB 197,121,97,241 ; vpunpcklwd %xmm1,%xmm0,%xmm14 DB 197,121,105,193 ; vpunpckhwd %xmm1,%xmm0,%xmm8 - DB 197,233,97,203 ; vpunpcklwd %xmm3,%xmm2,%xmm1 - DB 197,105,105,203 ; vpunpckhwd %xmm3,%xmm2,%xmm9 - DB 197,129,108,193 ; vpunpcklqdq %xmm1,%xmm15,%xmm0 - DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 - DB 197,249,105,210 ; vpunpckhwd %xmm2,%xmm0,%xmm2 + DB 197,49,97,250 ; vpunpcklwd %xmm2,%xmm9,%xmm15 + DB 197,49,105,202 ; vpunpckhwd %xmm2,%xmm9,%xmm9 + DB 196,193,9,108,199 ; vpunpcklqdq %xmm15,%xmm14,%xmm0 + DB 196,226,121,51,208 ; vpmovzxwd %xmm0,%xmm2 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - DB 196,98,125,24,37,201,1,2,0 ; vbroadcastss 0x201c9(%rip),%ymm12 # 2d428 <_sk_overlay_sse2_8bit+0xed0> - DB 196,193,124,84,212 ; vandps %ymm12,%ymm0,%ymm2 - DB 197,252,87,194 ; vxorps %ymm2,%ymm0,%ymm0 - DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,141,1,2,0 ; vbroadcastss 0x2018d(%rip),%xmm11 # 2d404 <_sk_overlay_sse2_8bit+0xeac> - DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,167,1,2,0 ; vbroadcastss 0x201a7(%rip),%xmm13 # 2d42c <_sk_overlay_sse2_8bit+0xed4> - DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 - DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 - DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 - DB 196,99,45,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm10,%ymm10 - DB 197,225,114,242,16 ; vpslld $0x10,%xmm2,%xmm3 - DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 + DB 197,123,18,21,218,4,2,0 ; vmovddup 0x204da(%rip),%xmm10 # 2dbc8 <_sk_overlay_sse2_8bit+0x1719> + DB 196,193,121,219,218 ; vpand %xmm10,%xmm0,%xmm3 + DB 196,98,121,24,29,12,252,1,0 ; vbroadcastss 0x1fc0c(%rip),%xmm11 # 2d308 <_sk_overlay_sse2_8bit+0xe59> + DB 197,33,102,227 ; vpcmpgtd %xmm3,%xmm11,%xmm12 + DB 196,193,105,219,202 ; vpand %xmm10,%xmm2,%xmm1 + DB 197,33,102,233 ; vpcmpgtd %xmm1,%xmm11,%xmm13 + DB 196,195,21,24,228,1 ; vinsertf128 $0x1,%xmm12,%ymm13,%ymm4 + DB 197,123,18,37,185,4,2,0 ; vmovddup 0x204b9(%rip),%xmm12 # 2dbd0 <_sk_overlay_sse2_8bit+0x1721> + DB 196,193,105,219,212 ; vpand %xmm12,%xmm2,%xmm2 DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 - DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 - DB 197,249,114,240,13 ; vpslld $0xd,%xmm0,%xmm0 - DB 196,193,97,114,246,13 ; vpslld $0xd,%xmm14,%xmm3 - DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 - DB 197,252,86,194 ; vorps %ymm2,%ymm0,%ymm0 - DB 196,227,125,25,194,1 ; vextractf128 $0x1,%ymm0,%xmm2 - DB 196,226,121,24,29,93,1,2,0 ; vbroadcastss 0x2015d(%rip),%xmm3 # 2d430 <_sk_overlay_sse2_8bit+0xed8> - DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 + DB 196,193,121,219,196 ; vpand %xmm12,%xmm0,%xmm0 + DB 197,249,114,240,16 ; vpslld $0x10,%xmm0,%xmm0 + DB 197,241,114,241,13 ; vpslld $0xd,%xmm1,%xmm1 + DB 197,233,235,201 ; vpor %xmm1,%xmm2,%xmm1 + DB 197,233,114,243,13 ; vpslld $0xd,%xmm3,%xmm2 + DB 197,249,235,194 ; vpor %xmm2,%xmm0,%xmm0 + DB 196,226,121,24,29,198,251,1,0 ; vbroadcastss 0x1fbc6(%rip),%xmm3 # 2d30c <_sk_overlay_sse2_8bit+0xe5d> DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 - DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 - DB 196,195,125,74,198,160 ; vblendvps %ymm10,%ymm14,%ymm0,%ymm0 - DB 197,129,109,201 ; vpunpckhqdq %xmm1,%xmm15,%xmm1 - DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 - DB 197,241,105,212 ; vpunpckhwd %xmm4,%xmm1,%xmm2 - DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 - DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,193,116,84,212 ; vandps %ymm12,%ymm1,%ymm2 - DB 197,244,87,202 ; vxorps %ymm2,%ymm1,%ymm1 - DB 196,195,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm10 - DB 196,193,40,87,251 ; vxorps %xmm11,%xmm10,%xmm7 - DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 - DB 196,193,112,87,243 ; vxorps %xmm11,%xmm1,%xmm6 - DB 197,145,102,246 ; vpcmpgtd %xmm6,%xmm13,%xmm6 - DB 196,227,77,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm6,%ymm6 - DB 197,193,114,242,16 ; vpslld $0x10,%xmm2,%xmm7 - DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 + DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 196,65,20,87,237 ; vxorps %ymm13,%ymm13,%ymm13 + DB 196,195,125,74,197,64 ; vblendvps %ymm4,%ymm13,%ymm0,%ymm0 + DB 196,193,9,109,207 ; vpunpckhqdq %xmm15,%xmm14,%xmm1 + DB 196,226,121,51,209 ; vpmovzxwd %xmm1,%xmm2 + DB 197,249,112,201,78 ; vpshufd $0x4e,%xmm1,%xmm1 + DB 196,98,121,51,249 ; vpmovzxwd %xmm1,%xmm15 + DB 196,193,1,219,226 ; vpand %xmm10,%xmm15,%xmm4 + DB 197,33,102,244 ; vpcmpgtd %xmm4,%xmm11,%xmm14 + DB 196,193,105,219,202 ; vpand %xmm10,%xmm2,%xmm1 + DB 197,161,102,249 ; vpcmpgtd %xmm1,%xmm11,%xmm7 + DB 196,195,69,24,254,1 ; vinsertf128 $0x1,%xmm14,%ymm7,%ymm7 + DB 196,193,105,219,212 ; vpand %xmm12,%xmm2,%xmm2 DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 - DB 196,227,69,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm7,%ymm2 DB 197,241,114,241,13 ; vpslld $0xd,%xmm1,%xmm1 - DB 196,193,65,114,242,13 ; vpslld $0xd,%xmm10,%xmm7 - DB 196,227,117,24,207,1 ; vinsertf128 $0x1,%xmm7,%ymm1,%ymm1 - DB 197,244,86,202 ; vorps %ymm2,%ymm1,%ymm1 - DB 196,227,125,25,202,1 ; vextractf128 $0x1,%ymm1,%xmm2 + DB 197,233,235,201 ; vpor %xmm1,%xmm2,%xmm1 + DB 196,193,1,219,212 ; vpand %xmm12,%xmm15,%xmm2 + DB 197,233,114,242,16 ; vpslld $0x10,%xmm2,%xmm2 + DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 + DB 197,233,235,212 ; vpor %xmm4,%xmm2,%xmm2 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 196,195,117,74,206,96 ; vblendvps %ymm6,%ymm14,%ymm1,%ymm1 + DB 196,195,117,74,205,112 ; vblendvps %ymm7,%ymm13,%ymm1,%ymm1 DB 196,193,57,108,209 ; vpunpcklqdq %xmm9,%xmm8,%xmm2 - DB 197,233,105,244 ; vpunpckhwd %xmm4,%xmm2,%xmm6 - DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - DB 196,193,108,84,244 ; vandps %ymm12,%ymm2,%ymm6 - DB 197,236,87,214 ; vxorps %ymm6,%ymm2,%ymm2 - DB 196,227,125,25,215,1 ; vextractf128 $0x1,%ymm2,%xmm7 - DB 196,193,64,87,235 ; vxorps %xmm11,%xmm7,%xmm5 - DB 197,145,102,237 ; vpcmpgtd %xmm5,%xmm13,%xmm5 - DB 196,193,104,87,227 ; vxorps %xmm11,%xmm2,%xmm4 - DB 197,145,102,228 ; vpcmpgtd %xmm4,%xmm13,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 197,209,114,246,16 ; vpslld $0x10,%xmm6,%xmm5 - DB 196,227,125,25,246,1 ; vextractf128 $0x1,%ymm6,%xmm6 - DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 - DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 + DB 196,226,121,51,226 ; vpmovzxwd %xmm2,%xmm4 + DB 197,249,112,210,78 ; vpshufd $0x4e,%xmm2,%xmm2 + DB 196,98,121,51,250 ; vpmovzxwd %xmm2,%xmm15 + DB 196,193,1,219,250 ; vpand %xmm10,%xmm15,%xmm7 + DB 197,33,102,247 ; vpcmpgtd %xmm7,%xmm11,%xmm14 + DB 196,193,89,219,210 ; vpand %xmm10,%xmm4,%xmm2 + DB 197,161,102,242 ; vpcmpgtd %xmm2,%xmm11,%xmm6 + DB 196,195,77,24,246,1 ; vinsertf128 $0x1,%xmm14,%ymm6,%ymm6 + DB 196,193,89,219,228 ; vpand %xmm12,%xmm4,%xmm4 + DB 197,217,114,244,16 ; vpslld $0x10,%xmm4,%xmm4 DB 197,233,114,242,13 ; vpslld $0xd,%xmm2,%xmm2 - DB 197,201,114,247,13 ; vpslld $0xd,%xmm7,%xmm6 - DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - DB 197,236,86,213 ; vorps %ymm5,%ymm2,%ymm2 - DB 196,227,125,25,213,1 ; vextractf128 $0x1,%ymm2,%xmm5 - DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 + DB 197,217,235,210 ; vpor %xmm2,%xmm4,%xmm2 + DB 196,193,1,219,228 ; vpand %xmm12,%xmm15,%xmm4 + DB 197,217,114,244,16 ; vpslld $0x10,%xmm4,%xmm4 + DB 197,193,114,247,13 ; vpslld $0xd,%xmm7,%xmm7 + DB 197,217,235,231 ; vpor %xmm7,%xmm4,%xmm4 + DB 197,217,254,227 ; vpaddd %xmm3,%xmm4,%xmm4 DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 - DB 196,227,109,24,213,1 ; vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 - DB 196,195,109,74,214,64 ; vblendvps %ymm4,%ymm14,%ymm2,%ymm2 + DB 196,227,109,24,212,1 ; vinsertf128 $0x1,%xmm4,%ymm2,%ymm2 + DB 196,195,109,74,213,96 ; vblendvps %ymm6,%ymm13,%ymm2,%ymm2 DB 196,193,57,109,225 ; vpunpckhqdq %xmm9,%xmm8,%xmm4 - DB 196,193,89,105,234 ; vpunpckhwd %xmm10,%xmm4,%xmm5 + DB 196,226,121,51,244 ; vpmovzxwd %xmm4,%xmm6 + DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,121,51,228 ; vpmovzxwd %xmm4,%xmm4 - DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 196,193,92,84,236 ; vandps %ymm12,%ymm4,%ymm5 - DB 197,220,87,229 ; vxorps %ymm5,%ymm4,%ymm4 - DB 196,227,125,25,230,1 ; vextractf128 $0x1,%ymm4,%xmm6 - DB 196,193,72,87,251 ; vxorps %xmm11,%xmm6,%xmm7 - DB 197,17,102,199 ; vpcmpgtd %xmm7,%xmm13,%xmm8 - DB 196,193,88,87,251 ; vxorps %xmm11,%xmm4,%xmm7 - DB 197,145,102,255 ; vpcmpgtd %xmm7,%xmm13,%xmm7 - DB 196,195,69,24,248,1 ; vinsertf128 $0x1,%xmm8,%ymm7,%ymm7 - DB 197,185,114,245,16 ; vpslld $0x10,%xmm5,%xmm8 - DB 196,227,125,25,237,1 ; vextractf128 $0x1,%ymm5,%xmm5 - DB 197,209,114,245,16 ; vpslld $0x10,%xmm5,%xmm5 - DB 196,227,61,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm8,%ymm5 - DB 197,217,114,244,13 ; vpslld $0xd,%xmm4,%xmm4 - DB 197,201,114,246,13 ; vpslld $0xd,%xmm6,%xmm6 - DB 196,227,93,24,230,1 ; vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 - DB 197,220,86,229 ; vorps %ymm5,%ymm4,%ymm4 - DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 - DB 197,209,254,235 ; vpaddd %xmm3,%xmm5,%xmm5 - DB 197,217,254,219 ; vpaddd %xmm3,%xmm4,%xmm3 - DB 196,227,101,24,221,1 ; vinsertf128 $0x1,%xmm5,%ymm3,%ymm3 - DB 196,195,101,74,222,112 ; vblendvps %ymm7,%ymm14,%ymm3,%ymm3 - DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 - DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 - DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 - DB 197,252,16,124,36,96 ; vmovups 0x60(%rsp),%ymm7 - DB 72,129,196,152,0,0,0 ; add $0x98,%rsp + DB 196,65,89,219,202 ; vpand %xmm10,%xmm4,%xmm9 + DB 196,65,33,102,193 ; vpcmpgtd %xmm9,%xmm11,%xmm8 + DB 196,193,73,219,250 ; vpand %xmm10,%xmm6,%xmm7 + DB 197,33,102,215 ; vpcmpgtd %xmm7,%xmm11,%xmm10 + DB 196,67,45,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm10,%ymm8 + DB 196,193,73,219,244 ; vpand %xmm12,%xmm6,%xmm6 + DB 196,193,89,219,228 ; vpand %xmm12,%xmm4,%xmm4 + DB 197,201,114,246,16 ; vpslld $0x10,%xmm6,%xmm6 + DB 197,193,114,247,13 ; vpslld $0xd,%xmm7,%xmm7 + DB 197,201,235,247 ; vpor %xmm7,%xmm6,%xmm6 + DB 197,217,114,244,16 ; vpslld $0x10,%xmm4,%xmm4 + DB 196,193,65,114,241,13 ; vpslld $0xd,%xmm9,%xmm7 + DB 197,217,235,231 ; vpor %xmm7,%xmm4,%xmm4 + DB 197,217,254,227 ; vpaddd %xmm3,%xmm4,%xmm4 + DB 197,201,254,219 ; vpaddd %xmm3,%xmm6,%xmm3 + DB 196,227,101,24,220,1 ; vinsertf128 $0x1,%xmm4,%ymm3,%ymm3 + DB 196,195,101,74,221,128 ; vblendvps %ymm8,%ymm13,%ymm3,%ymm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 197,252,40,229 ; vmovaps %ymm5,%ymm4 + DB 197,252,16,44,36 ; vmovups (%rsp),%ymm5 + DB 197,252,16,116,36,32 ; vmovups 0x20(%rsp),%ymm6 + DB 197,252,16,124,36,64 ; vmovups 0x40(%rsp),%ymm7 + DB 72,131,196,120 ; add $0x78,%rsp DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 @@ -12640,105 +12824,103 @@ _sk_gather_f16_avx LABEL PROC PUBLIC _sk_store_f16_avx _sk_store_f16_avx LABEL PROC - DB 72,129,236,216,0,0,0 ; sub $0xd8,%rsp - DB 197,252,17,188,36,160,0,0,0 ; vmovups %ymm7,0xa0(%rsp) - DB 197,252,17,180,36,128,0,0,0 ; vmovups %ymm6,0x80(%rsp) - DB 197,252,17,108,36,96 ; vmovups %ymm5,0x60(%rsp) - DB 197,252,17,100,36,64 ; vmovups %ymm4,0x40(%rsp) - DB 196,98,125,24,13,58,255,1,0 ; vbroadcastss 0x1ff3a(%rip),%ymm9 # 2d404 <_sk_overlay_sse2_8bit+0xeac> - DB 196,65,124,84,209 ; vandps %ymm9,%ymm0,%ymm10 - DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) - DB 196,65,124,87,218 ; vxorps %ymm10,%ymm0,%ymm11 - DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12 - DB 196,98,121,24,5,76,255,1,0 ; vbroadcastss 0x1ff4c(%rip),%xmm8 # 2d434 <_sk_overlay_sse2_8bit+0xedc> - DB 196,65,57,102,236 ; vpcmpgtd %xmm12,%xmm8,%xmm13 - DB 196,65,57,102,243 ; vpcmpgtd %xmm11,%xmm8,%xmm14 - DB 196,67,13,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 - DB 196,193,9,114,210,16 ; vpsrld $0x10,%xmm10,%xmm14 - DB 196,67,125,25,210,1 ; vextractf128 $0x1,%ymm10,%xmm10 - DB 196,193,41,114,210,16 ; vpsrld $0x10,%xmm10,%xmm10 - DB 196,67,13,24,242,1 ; vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 - DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11 + DB 72,129,236,184,0,0,0 ; sub $0xb8,%rsp + DB 197,252,17,188,36,128,0,0,0 ; vmovups %ymm7,0x80(%rsp) + DB 197,252,17,116,36,96 ; vmovups %ymm6,0x60(%rsp) + DB 197,252,17,108,36,64 ; vmovups %ymm5,0x40(%rsp) + DB 197,252,17,100,36,32 ; vmovups %ymm4,0x20(%rsp) + DB 197,252,40,225 ; vmovaps %ymm1,%ymm4 + DB 196,98,125,24,5,254,249,1,0 ; vbroadcastss 0x1f9fe(%rip),%ymm8 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> + DB 196,65,124,84,216 ; vandps %ymm8,%ymm0,%ymm11 + DB 196,98,125,24,21,28,250,1,0 ; vbroadcastss 0x1fa1c(%rip),%ymm10 # 2d310 <_sk_overlay_sse2_8bit+0xe61> + DB 196,65,124,84,226 ; vandps %ymm10,%ymm0,%ymm12 + DB 196,67,125,25,229,1 ; vextractf128 $0x1,%ymm12,%xmm13 + DB 196,98,121,24,13,12,250,1,0 ; vbroadcastss 0x1fa0c(%rip),%xmm9 # 2d314 <_sk_overlay_sse2_8bit+0xe65> + DB 196,65,49,102,245 ; vpcmpgtd %xmm13,%xmm9,%xmm14 + DB 196,65,49,102,252 ; vpcmpgtd %xmm12,%xmm9,%xmm15 + DB 196,67,5,24,246,1 ; vinsertf128 $0x1,%xmm14,%ymm15,%ymm14 + DB 196,67,125,25,223,1 ; vextractf128 $0x1,%ymm11,%xmm15 + DB 196,193,1,114,215,16 ; vpsrld $0x10,%xmm15,%xmm15 + DB 196,193,33,114,211,16 ; vpsrld $0x10,%xmm11,%xmm11 + DB 196,193,17,114,213,13 ; vpsrld $0xd,%xmm13,%xmm13 + DB 196,65,17,254,239 ; vpaddd %xmm15,%xmm13,%xmm13 DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12 - DB 196,98,125,24,21,19,255,1,0 ; vbroadcastss 0x1ff13(%rip),%ymm10 # 2d438 <_sk_overlay_sse2_8bit+0xee0> - DB 196,65,12,86,242 ; vorps %ymm10,%ymm14,%ymm14 - DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15 - DB 196,65,1,254,228 ; vpaddd %xmm12,%xmm15,%xmm12 - DB 196,65,9,254,219 ; vpaddd %xmm11,%xmm14,%xmm11 - DB 196,67,37,24,228,1 ; vinsertf128 $0x1,%xmm12,%ymm11,%ymm12 - DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 - DB 196,99,29,74,224,208 ; vblendvps %ymm13,%ymm0,%ymm12,%ymm12 - DB 196,65,116,84,233 ; vandps %ymm9,%ymm1,%ymm13 - DB 197,252,17,76,36,32 ; vmovups %ymm1,0x20(%rsp) - DB 196,65,116,87,245 ; vxorps %ymm13,%ymm1,%ymm14 + DB 196,65,25,254,227 ; vpaddd %xmm11,%xmm12,%xmm12 + DB 196,98,121,24,29,207,249,1,0 ; vbroadcastss 0x1f9cf(%rip),%xmm11 # 2d318 <_sk_overlay_sse2_8bit+0xe69> + DB 196,65,17,254,235 ; vpaddd %xmm11,%xmm13,%xmm13 + DB 196,65,25,254,227 ; vpaddd %xmm11,%xmm12,%xmm12 + DB 196,67,29,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm12,%ymm13 + DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 + DB 196,99,21,74,233,224 ; vblendvps %ymm14,%ymm1,%ymm13,%ymm13 + DB 196,65,92,84,242 ; vandps %ymm10,%ymm4,%ymm14 DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15 - DB 196,193,57,102,255 ; vpcmpgtd %xmm15,%xmm8,%xmm7 - DB 196,65,57,102,222 ; vpcmpgtd %xmm14,%xmm8,%xmm11 - DB 196,227,37,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm11,%ymm7 - DB 196,193,33,114,213,16 ; vpsrld $0x10,%xmm13,%xmm11 - DB 196,99,125,25,238,1 ; vextractf128 $0x1,%ymm13,%xmm6 + DB 196,193,49,102,255 ; vpcmpgtd %xmm15,%xmm9,%xmm7 + DB 196,65,49,102,230 ; vpcmpgtd %xmm14,%xmm9,%xmm12 + DB 196,227,29,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm12,%ymm7 + DB 197,252,17,36,36 ; vmovups %ymm4,(%rsp) + DB 196,65,92,84,224 ; vandps %ymm8,%ymm4,%ymm12 + DB 196,99,125,25,230,1 ; vextractf128 $0x1,%ymm12,%xmm6 DB 197,201,114,214,16 ; vpsrld $0x10,%xmm6,%xmm6 - DB 196,227,37,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm11,%ymm6 - DB 196,193,33,114,215,13 ; vpsrld $0xd,%xmm15,%xmm11 - DB 196,193,76,86,242 ; vorps %ymm10,%ymm6,%ymm6 - DB 196,227,125,25,245,1 ; vextractf128 $0x1,%ymm6,%xmm5 - DB 196,193,81,254,235 ; vpaddd %xmm11,%xmm5,%xmm5 + DB 196,193,81,114,215,13 ; vpsrld $0xd,%xmm15,%xmm5 + DB 197,124,40,250 ; vmovaps %ymm2,%ymm15 + DB 197,209,254,238 ; vpaddd %xmm6,%xmm5,%xmm5 + DB 196,193,73,114,212,16 ; vpsrld $0x10,%xmm12,%xmm6 DB 196,193,89,114,214,13 ; vpsrld $0xd,%xmm14,%xmm4 - DB 197,201,254,228 ; vpaddd %xmm4,%xmm6,%xmm4 + DB 197,217,254,230 ; vpaddd %xmm6,%xmm4,%xmm4 + DB 196,193,81,254,235 ; vpaddd %xmm11,%xmm5,%xmm5 + DB 196,193,89,254,227 ; vpaddd %xmm11,%xmm4,%xmm4 DB 196,227,93,24,229,1 ; vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 - DB 196,99,93,74,232,112 ; vblendvps %ymm7,%ymm0,%ymm4,%ymm13 - DB 196,193,108,84,225 ; vandps %ymm9,%ymm2,%ymm4 - DB 197,236,87,236 ; vxorps %ymm4,%ymm2,%ymm5 - DB 196,227,125,25,238,1 ; vextractf128 $0x1,%ymm5,%xmm6 - DB 197,185,102,254 ; vpcmpgtd %xmm6,%xmm8,%xmm7 - DB 197,57,102,221 ; vpcmpgtd %xmm5,%xmm8,%xmm11 - DB 196,227,37,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm11,%ymm7 - DB 197,161,114,212,16 ; vpsrld $0x10,%xmm4,%xmm11 - DB 196,227,125,25,228,1 ; vextractf128 $0x1,%ymm4,%xmm4 - DB 197,217,114,212,16 ; vpsrld $0x10,%xmm4,%xmm4 - DB 196,227,37,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm11,%ymm4 - DB 197,201,114,214,13 ; vpsrld $0xd,%xmm6,%xmm6 - DB 196,193,92,86,226 ; vorps %ymm10,%ymm4,%ymm4 - DB 196,227,125,25,225,1 ; vextractf128 $0x1,%ymm4,%xmm1 - DB 197,241,254,206 ; vpaddd %xmm6,%xmm1,%xmm1 + DB 196,99,93,74,241,112 ; vblendvps %ymm7,%ymm1,%ymm4,%ymm14 + DB 196,193,4,84,226 ; vandps %ymm10,%ymm15,%ymm4 + DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 + DB 197,177,102,245 ; vpcmpgtd %xmm5,%xmm9,%xmm6 + DB 197,177,102,252 ; vpcmpgtd %xmm4,%xmm9,%xmm7 + DB 196,227,69,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm7,%ymm6 + DB 196,193,4,84,248 ; vandps %ymm8,%ymm15,%ymm7 + DB 196,227,125,25,250,1 ; vextractf128 $0x1,%ymm7,%xmm2 + DB 197,233,114,210,16 ; vpsrld $0x10,%xmm2,%xmm2 DB 197,209,114,213,13 ; vpsrld $0xd,%xmm5,%xmm5 + DB 197,209,254,210 ; vpaddd %xmm2,%xmm5,%xmm2 + DB 197,209,114,215,16 ; vpsrld $0x10,%xmm7,%xmm5 + DB 197,217,114,212,13 ; vpsrld $0xd,%xmm4,%xmm4 DB 197,217,254,229 ; vpaddd %xmm5,%xmm4,%xmm4 - DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 - DB 196,99,117,74,216,112 ; vblendvps %ymm7,%ymm0,%ymm1,%ymm11 - DB 196,193,100,84,225 ; vandps %ymm9,%ymm3,%ymm4 - DB 197,228,87,236 ; vxorps %ymm4,%ymm3,%ymm5 - DB 196,227,125,25,238,1 ; vextractf128 $0x1,%ymm5,%xmm6 - DB 197,185,102,254 ; vpcmpgtd %xmm6,%xmm8,%xmm7 - DB 197,57,102,197 ; vpcmpgtd %xmm5,%xmm8,%xmm8 - DB 196,227,61,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm8,%ymm7 - DB 197,185,114,212,16 ; vpsrld $0x10,%xmm4,%xmm8 - DB 196,227,125,25,228,1 ; vextractf128 $0x1,%ymm4,%xmm4 - DB 197,217,114,212,16 ; vpsrld $0x10,%xmm4,%xmm4 - DB 196,227,61,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm8,%ymm4 - DB 196,193,92,86,226 ; vorps %ymm10,%ymm4,%ymm4 - DB 197,201,114,214,13 ; vpsrld $0xd,%xmm6,%xmm6 - DB 196,227,125,25,225,1 ; vextractf128 $0x1,%ymm4,%xmm1 - DB 197,241,254,206 ; vpaddd %xmm6,%xmm1,%xmm1 + DB 196,193,105,254,211 ; vpaddd %xmm11,%xmm2,%xmm2 + DB 196,193,89,254,227 ; vpaddd %xmm11,%xmm4,%xmm4 + DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 + DB 196,99,109,74,225,96 ; vblendvps %ymm6,%ymm1,%ymm2,%ymm12 + DB 196,193,100,84,226 ; vandps %ymm10,%ymm3,%ymm4 + DB 196,227,125,25,229,1 ; vextractf128 $0x1,%ymm4,%xmm5 + DB 197,177,102,245 ; vpcmpgtd %xmm5,%xmm9,%xmm6 + DB 197,177,102,252 ; vpcmpgtd %xmm4,%xmm9,%xmm7 + DB 196,227,69,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm7,%ymm6 + DB 196,193,100,84,248 ; vandps %ymm8,%ymm3,%ymm7 + DB 196,227,125,25,250,1 ; vextractf128 $0x1,%ymm7,%xmm2 + DB 197,233,114,210,16 ; vpsrld $0x10,%xmm2,%xmm2 DB 197,209,114,213,13 ; vpsrld $0xd,%xmm5,%xmm5 + DB 197,209,254,210 ; vpaddd %xmm2,%xmm5,%xmm2 + DB 197,209,114,215,16 ; vpsrld $0x10,%xmm7,%xmm5 + DB 197,217,114,212,13 ; vpsrld $0xd,%xmm4,%xmm4 DB 197,217,254,229 ; vpaddd %xmm5,%xmm4,%xmm4 - DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 - DB 196,227,117,74,200,112 ; vblendvps %ymm7,%ymm0,%ymm1,%ymm1 - DB 196,99,125,25,224,1 ; vextractf128 $0x1,%ymm12,%xmm0 - DB 196,226,25,43,192 ; vpackusdw %xmm0,%xmm12,%xmm0 - DB 196,99,125,25,236,1 ; vextractf128 $0x1,%ymm13,%xmm4 - DB 196,226,17,43,228 ; vpackusdw %xmm4,%xmm13,%xmm4 - DB 196,99,125,25,221,1 ; vextractf128 $0x1,%ymm11,%xmm5 - DB 196,226,33,43,245 ; vpackusdw %xmm5,%xmm11,%xmm6 + DB 196,193,105,254,211 ; vpaddd %xmm11,%xmm2,%xmm2 + DB 196,193,89,254,227 ; vpaddd %xmm11,%xmm4,%xmm4 + DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 + DB 196,227,109,74,201,96 ; vblendvps %ymm6,%ymm1,%ymm2,%ymm1 + DB 196,99,125,25,234,1 ; vextractf128 $0x1,%ymm13,%xmm2 + DB 196,226,17,43,210 ; vpackusdw %xmm2,%xmm13,%xmm2 + DB 196,99,125,25,244,1 ; vextractf128 $0x1,%ymm14,%xmm4 + DB 196,226,9,43,228 ; vpackusdw %xmm4,%xmm14,%xmm4 + DB 196,99,125,25,229,1 ; vextractf128 $0x1,%ymm12,%xmm5 + DB 196,226,25,43,245 ; vpackusdw %xmm5,%xmm12,%xmm6 DB 196,227,125,25,205,1 ; vextractf128 $0x1,%ymm1,%xmm5 DB 196,226,113,43,205 ; vpackusdw %xmm5,%xmm1,%xmm1 - DB 197,249,97,236 ; vpunpcklwd %xmm4,%xmm0,%xmm5 - DB 197,249,105,196 ; vpunpckhwd %xmm4,%xmm0,%xmm0 + DB 197,233,97,236 ; vpunpcklwd %xmm4,%xmm2,%xmm5 + DB 197,233,105,212 ; vpunpckhwd %xmm4,%xmm2,%xmm2 DB 197,201,97,225 ; vpunpcklwd %xmm1,%xmm6,%xmm4 DB 197,201,105,201 ; vpunpckhwd %xmm1,%xmm6,%xmm1 DB 197,81,98,220 ; vpunpckldq %xmm4,%xmm5,%xmm11 DB 197,81,106,212 ; vpunpckhdq %xmm4,%xmm5,%xmm10 - DB 197,121,98,201 ; vpunpckldq %xmm1,%xmm0,%xmm9 - DB 197,121,106,193 ; vpunpckhdq %xmm1,%xmm0,%xmm8 + DB 197,105,98,201 ; vpunpckldq %xmm1,%xmm2,%xmm9 + DB 197,105,106,193 ; vpunpckhdq %xmm1,%xmm2,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax DB 76,99,72,8 ; movslq 0x8(%rax),%r9 DB 76,99,193 ; movslq %ecx,%r8 @@ -12747,38 +12929,38 @@ _sk_store_f16_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,79 ; jne d722 <_sk_store_f16_avx+0x286> - DB 196,65,120,17,28,192 ; vmovups %xmm11,(%r8,%rax,8) - DB 196,65,120,17,84,192,16 ; vmovups %xmm10,0x10(%r8,%rax,8) - DB 196,65,120,17,76,192,32 ; vmovups %xmm9,0x20(%r8,%rax,8) + DB 117,74 ; jne db25 <_sk_store_f16_avx+0x26e> + DB 196,65,122,127,28,192 ; vmovdqu %xmm11,(%r8,%rax,8) + DB 196,65,122,127,84,192,16 ; vmovdqu %xmm10,0x10(%r8,%rax,8) + DB 196,65,122,127,76,192,32 ; vmovdqu %xmm9,0x20(%r8,%rax,8) DB 196,65,122,127,68,192,48 ; vmovdqu %xmm8,0x30(%r8,%rax,8) DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 - DB 197,252,16,76,36,32 ; vmovups 0x20(%rsp),%ymm1 - DB 197,252,16,100,36,64 ; vmovups 0x40(%rsp),%ymm4 - DB 197,252,16,108,36,96 ; vmovups 0x60(%rsp),%ymm5 - DB 197,252,16,180,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm6 - DB 197,252,16,188,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm7 - DB 72,129,196,216,0,0,0 ; add $0xd8,%rsp + DB 197,252,16,12,36 ; vmovups (%rsp),%ymm1 + DB 197,124,41,250 ; vmovaps %ymm15,%ymm2 + DB 197,252,16,100,36,32 ; vmovups 0x20(%rsp),%ymm4 + DB 197,252,16,108,36,64 ; vmovups 0x40(%rsp),%ymm5 + DB 197,252,16,116,36,96 ; vmovups 0x60(%rsp),%ymm6 + DB 197,252,16,188,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm7 + DB 72,129,196,184,0,0,0 ; add $0xb8,%rsp DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,192 ; vmovq %xmm11,(%r8,%rax,8) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,192 ; je d6ee <_sk_store_f16_avx+0x252> + DB 116,197 ; je daf6 <_sk_store_f16_avx+0x23f> DB 196,65,121,23,92,192,8 ; vmovhpd %xmm11,0x8(%r8,%rax,8) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,179 ; jb d6ee <_sk_store_f16_avx+0x252> + DB 114,184 ; jb daf6 <_sk_store_f16_avx+0x23f> DB 196,65,121,214,84,192,16 ; vmovq %xmm10,0x10(%r8,%rax,8) - DB 116,170 ; je d6ee <_sk_store_f16_avx+0x252> + DB 116,175 ; je daf6 <_sk_store_f16_avx+0x23f> DB 196,65,121,23,84,192,24 ; vmovhpd %xmm10,0x18(%r8,%rax,8) DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,157 ; jb d6ee <_sk_store_f16_avx+0x252> + DB 114,162 ; jb daf6 <_sk_store_f16_avx+0x23f> DB 196,65,121,214,76,192,32 ; vmovq %xmm9,0x20(%r8,%rax,8) - DB 116,148 ; je d6ee <_sk_store_f16_avx+0x252> + DB 116,153 ; je daf6 <_sk_store_f16_avx+0x23f> DB 196,65,121,23,76,192,40 ; vmovhpd %xmm9,0x28(%r8,%rax,8) DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,135 ; jb d6ee <_sk_store_f16_avx+0x252> + DB 114,140 ; jb daf6 <_sk_store_f16_avx+0x23f> DB 196,65,121,214,68,192,48 ; vmovq %xmm8,0x30(%r8,%rax,8) - DB 233,123,255,255,255 ; jmpq d6ee <_sk_store_f16_avx+0x252> + DB 235,131 ; jmp daf6 <_sk_store_f16_avx+0x23f> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC @@ -12791,7 +12973,7 @@ _sk_load_u16_be_avx LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,253,0,0,0 ; jne d897 <_sk_load_u16_be_avx+0x124> + DB 15,133,12,1,0,0 ; jne dca6 <_sk_load_u16_be_avx+0x133> DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -12800,76 +12982,79 @@ _sk_load_u16_be_avx LABEL PROC DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2 DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1 DB 196,193,97,105,217 ; vpunpckhwd %xmm9,%xmm3,%xmm3 - DB 197,121,97,202 ; vpunpcklwd %xmm2,%xmm0,%xmm9 + DB 197,121,97,210 ; vpunpcklwd %xmm2,%xmm0,%xmm10 DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8 DB 197,241,97,211 ; vpunpcklwd %xmm3,%xmm1,%xmm2 - DB 197,113,105,227 ; vpunpckhwd %xmm3,%xmm1,%xmm12 - DB 197,177,108,194 ; vpunpcklqdq %xmm2,%xmm9,%xmm0 + DB 197,113,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm11 + DB 197,169,108,194 ; vpunpcklqdq %xmm2,%xmm10,%xmm0 DB 197,241,113,240,8 ; vpsllw $0x8,%xmm0,%xmm1 DB 197,249,113,208,8 ; vpsrlw $0x8,%xmm0,%xmm0 DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 - DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,49,251,1,0 ; vbroadcastss 0x1fb31(%rip),%ymm11 # 2d33c <_sk_overlay_sse2_8bit+0xde4> - DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 - DB 197,177,109,202 ; vpunpckhqdq %xmm2,%xmm9,%xmm1 + DB 196,98,125,24,13,17,246,1,0 ; vbroadcastss 0x1f611(%rip),%ymm9 # 2d21c <_sk_overlay_sse2_8bit+0xd6d> + DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 197,169,109,202 ; vpunpckhqdq %xmm2,%xmm10,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 DB 197,241,113,209,8 ; vpsrlw $0x8,%xmm1,%xmm1 DB 197,233,235,201 ; vpor %xmm1,%xmm2,%xmm1 - DB 196,193,113,105,210 ; vpunpckhwd %xmm10,%xmm1,%xmm2 + DB 196,226,121,51,209 ; vpmovzxwd %xmm1,%xmm2 + DB 197,249,112,201,78 ; vpshufd $0x4e,%xmm1,%xmm1 DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 - DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 + DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,193,116,89,203 ; vmulps %ymm11,%ymm1,%ymm1 - DB 196,193,57,108,212 ; vpunpcklqdq %xmm12,%xmm8,%xmm2 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 196,193,57,108,211 ; vpunpcklqdq %xmm11,%xmm8,%xmm2 DB 197,225,113,242,8 ; vpsllw $0x8,%xmm2,%xmm3 DB 197,233,113,210,8 ; vpsrlw $0x8,%xmm2,%xmm2 DB 197,225,235,210 ; vpor %xmm2,%xmm3,%xmm2 - DB 196,193,105,105,218 ; vpunpckhwd %xmm10,%xmm2,%xmm3 + DB 196,226,121,51,218 ; vpmovzxwd %xmm2,%xmm3 + DB 197,249,112,210,78 ; vpshufd $0x4e,%xmm2,%xmm2 DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 - DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 - DB 196,193,57,109,220 ; vpunpckhqdq %xmm12,%xmm8,%xmm3 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 196,193,57,109,219 ; vpunpckhqdq %xmm11,%xmm8,%xmm3 DB 197,185,113,243,8 ; vpsllw $0x8,%xmm3,%xmm8 DB 197,225,113,211,8 ; vpsrlw $0x8,%xmm3,%xmm3 DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3 - DB 196,65,97,105,194 ; vpunpckhwd %xmm10,%xmm3,%xmm8 + DB 196,98,121,51,195 ; vpmovzxwd %xmm3,%xmm8 + DB 197,249,112,219,78 ; vpshufd $0x4e,%xmm3,%xmm3 DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 - DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 + DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,193,100,89,219 ; vmulps %ymm11,%ymm3,%ymm3 + DB 196,193,100,89,217 ; vmulps %ymm9,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,85 ; je d8fd <_sk_load_u16_be_avx+0x18a> + DB 116,85 ; je dd0c <_sk_load_u16_be_avx+0x199> DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,72 ; jb d8fd <_sk_load_u16_be_avx+0x18a> + DB 114,72 ; jb dd0c <_sk_load_u16_be_avx+0x199> DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 116,72 ; je d90a <_sk_load_u16_be_avx+0x197> + DB 116,72 ; je dd19 <_sk_load_u16_be_avx+0x1a6> DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,59 ; jb d90a <_sk_load_u16_be_avx+0x197> + DB 114,59 ; jb dd19 <_sk_load_u16_be_avx+0x1a6> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 15,132,213,254,255,255 ; je d7b5 <_sk_load_u16_be_avx+0x42> + DB 15,132,198,254,255,255 ; je dbb5 <_sk_load_u16_be_avx+0x42> DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,130,196,254,255,255 ; jb d7b5 <_sk_load_u16_be_avx+0x42> + DB 15,130,181,254,255,255 ; jb dbb5 <_sk_load_u16_be_avx+0x42> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,184,254,255,255 ; jmpq d7b5 <_sk_load_u16_be_avx+0x42> + DB 233,169,254,255,255 ; jmpq dbb5 <_sk_load_u16_be_avx+0x42> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,171,254,255,255 ; jmpq d7b5 <_sk_load_u16_be_avx+0x42> + DB 233,156,254,255,255 ; jmpq dbb5 <_sk_load_u16_be_avx+0x42> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,162,254,255,255 ; jmpq d7b5 <_sk_load_u16_be_avx+0x42> + DB 233,147,254,255,255 ; jmpq dbb5 <_sk_load_u16_be_avx+0x42> PUBLIC _sk_load_rgb_u16_be_avx _sk_load_rgb_u16_be_avx LABEL PROC @@ -12884,7 +13069,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 72,141,4,64 ; lea (%rax,%rax,2),%rax DB 72,193,248,32 ; sar $0x20,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,241,0,0,0 ; jne da2f <_sk_load_rgb_u16_be_avx+0x11c> + DB 15,133,253,0,0,0 ; jne de4a <_sk_load_rgb_u16_be_avx+0x128> DB 196,65,122,111,28,64 ; vmovdqu (%r8,%rax,2),%xmm11 DB 196,193,122,111,92,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm3 DB 196,193,122,111,84,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm2 @@ -12892,109 +13077,111 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 197,249,115,216,4 ; vpsrldq $0x4,%xmm0,%xmm0 DB 196,193,57,115,219,6 ; vpsrldq $0x6,%xmm11,%xmm8 DB 197,169,115,219,6 ; vpsrldq $0x6,%xmm3,%xmm10 - DB 197,153,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm12 + DB 197,241,115,218,6 ; vpsrldq $0x6,%xmm2,%xmm1 DB 197,177,115,216,6 ; vpsrldq $0x6,%xmm0,%xmm9 - DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 - DB 196,193,57,97,202 ; vpunpcklwd %xmm10,%xmm8,%xmm1 + DB 196,193,113,97,201 ; vpunpcklwd %xmm9,%xmm1,%xmm1 DB 197,233,97,192 ; vpunpcklwd %xmm0,%xmm2,%xmm0 - DB 196,193,25,97,209 ; vpunpcklwd %xmm9,%xmm12,%xmm2 - DB 197,97,97,193 ; vpunpcklwd %xmm1,%xmm3,%xmm8 - DB 197,97,105,201 ; vpunpckhwd %xmm1,%xmm3,%xmm9 - DB 197,249,97,202 ; vpunpcklwd %xmm2,%xmm0,%xmm1 - DB 197,121,105,210 ; vpunpckhwd %xmm2,%xmm0,%xmm10 - DB 197,185,108,193 ; vpunpcklqdq %xmm1,%xmm8,%xmm0 - DB 197,225,113,240,8 ; vpsllw $0x8,%xmm0,%xmm3 + DB 196,193,57,97,210 ; vpunpcklwd %xmm10,%xmm8,%xmm2 + DB 197,161,97,219 ; vpunpcklwd %xmm3,%xmm11,%xmm3 + DB 197,97,97,194 ; vpunpcklwd %xmm2,%xmm3,%xmm8 + DB 197,225,105,210 ; vpunpckhwd %xmm2,%xmm3,%xmm2 + DB 197,249,97,217 ; vpunpcklwd %xmm1,%xmm0,%xmm3 + DB 197,249,105,193 ; vpunpckhwd %xmm1,%xmm0,%xmm0 + DB 197,233,108,208 ; vpunpcklqdq %xmm0,%xmm2,%xmm2 + DB 197,185,108,195 ; vpunpcklqdq %xmm3,%xmm8,%xmm0 + DB 197,241,113,240,8 ; vpsllw $0x8,%xmm0,%xmm1 DB 197,249,113,208,8 ; vpsrlw $0x8,%xmm0,%xmm0 - DB 197,225,235,192 ; vpor %xmm0,%xmm3,%xmm0 - DB 196,65,33,239,219 ; vpxor %xmm11,%xmm11,%xmm11 - DB 196,193,121,105,211 ; vpunpckhwd %xmm11,%xmm0,%xmm2 + DB 197,241,235,192 ; vpor %xmm0,%xmm1,%xmm0 + DB 196,226,121,51,200 ; vpmovzxwd %xmm0,%xmm1 + DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 - DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,21,115,249,1,0 ; vbroadcastss 0x1f973(%rip),%ymm2 # 2d33c <_sk_overlay_sse2_8bit+0xde4> - DB 197,252,89,194 ; vmulps %ymm2,%ymm0,%ymm0 - DB 197,185,109,201 ; vpunpckhqdq %xmm1,%xmm8,%xmm1 + DB 196,98,125,24,13,64,244,1,0 ; vbroadcastss 0x1f440(%rip),%ymm9 # 2d21c <_sk_overlay_sse2_8bit+0xd6d> + DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 197,185,109,203 ; vpunpckhqdq %xmm3,%xmm8,%xmm1 DB 197,225,113,241,8 ; vpsllw $0x8,%xmm1,%xmm3 DB 197,241,113,209,8 ; vpsrlw $0x8,%xmm1,%xmm1 DB 197,225,235,201 ; vpor %xmm1,%xmm3,%xmm1 - DB 196,193,113,105,219 ; vpunpckhwd %xmm11,%xmm1,%xmm3 + DB 196,226,121,51,217 ; vpmovzxwd %xmm1,%xmm3 + DB 197,249,112,201,78 ; vpshufd $0x4e,%xmm1,%xmm1 DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1 - DB 196,227,117,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm1,%ymm1 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,193,49,108,218 ; vpunpcklqdq %xmm10,%xmm9,%xmm3 - DB 197,185,113,243,8 ; vpsllw $0x8,%xmm3,%xmm8 - DB 197,225,113,211,8 ; vpsrlw $0x8,%xmm3,%xmm3 - DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3 - DB 196,65,97,105,195 ; vpunpckhwd %xmm11,%xmm3,%xmm8 - DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 - DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 - DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 197,225,113,242,8 ; vpsllw $0x8,%xmm2,%xmm3 + DB 197,233,113,210,8 ; vpsrlw $0x8,%xmm2,%xmm2 + DB 197,225,235,210 ; vpor %xmm2,%xmm3,%xmm2 + DB 196,226,121,51,218 ; vpmovzxwd %xmm2,%xmm3 + DB 197,249,112,210,78 ; vpshufd $0x4e,%xmm2,%xmm2 + DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 + DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,127,248,1,0 ; vbroadcastss 0x1f87f(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,226,125,24,29,68,243,1,0 ; vbroadcastss 0x1f344(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 255,224 ; jmpq *%rax DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 - DB 196,65,121,196,92,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm11 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,65,121,196,92,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm11 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,32 ; jne da68 <_sk_load_rgb_u16_be_avx+0x155> + DB 117,31 ; jne de82 <_sk_load_rgb_u16_be_avx+0x160> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 - DB 233,11,255,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 233,0,255,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> DB 196,193,121,110,68,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm0 - DB 196,65,121,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm0,%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,65,121,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm0,%xmm8 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,49 ; jb dab3 <_sk_load_rgb_u16_be_avx+0x1a0> + DB 114,48 ; jb decc <_sk_load_rgb_u16_be_avx+0x1aa> DB 196,193,121,110,68,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm0 - DB 196,193,121,196,92,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm0,%xmm3 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,193,121,196,92,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm0,%xmm3 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 117,50 ; jne dace <_sk_load_rgb_u16_be_avx+0x1bb> + DB 117,48 ; jne dee6 <_sk_load_rgb_u16_be_avx+0x1c4> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 - DB 233,192,254,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 233,182,254,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 - DB 233,165,254,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 233,156,254,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> DB 196,193,121,110,68,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm0 - DB 196,65,121,196,84,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm0,%xmm10 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,65,121,196,84,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm0,%xmm10 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,40 ; jb db10 <_sk_load_rgb_u16_be_avx+0x1fd> + DB 114,39 ; jb df27 <_sk_load_rgb_u16_be_avx+0x205> DB 196,193,121,110,68,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm0 - DB 196,193,121,196,84,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm0,%xmm2 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,193,121,196,84,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm0,%xmm2 DB 72,131,255,5 ; cmp $0x5,%rdi - DB 117,32 ; jne db22 <_sk_load_rgb_u16_be_avx+0x20f> + DB 117,30 ; jne df38 <_sk_load_rgb_u16_be_avx+0x216> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 - DB 233,99,254,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 + DB 233,91,254,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12 + DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 - DB 233,81,254,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 233,74,254,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> DB 196,193,121,110,68,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm0 - DB 196,65,121,196,100,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm0,%xmm12 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 + DB 196,193,121,196,76,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm0,%xmm1 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,25 ; jb db55 <_sk_load_rgb_u16_be_avx+0x242> + DB 114,25 ; jb df6b <_sk_load_rgb_u16_be_avx+0x249> DB 196,193,121,110,68,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm0 - DB 196,193,121,196,68,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm0,%xmm0 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 - DB 233,30,254,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 196,193,121,196,68,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm0,%xmm0 + DB 233,23,254,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 - DB 233,21,254,255,255 ; jmpq d973 <_sk_load_rgb_u16_be_avx+0x60> + DB 233,14,254,255,255 ; jmpq dd82 <_sk_load_rgb_u16_be_avx+0x60> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC @@ -13006,7 +13193,7 @@ _sk_store_u16_be_avx LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax - DB 196,98,125,24,5,63,248,1,0 ; vbroadcastss 0x1f83f(%rip),%ymm8 # 2d3c4 <_sk_overlay_sse2_8bit+0xe6c> + DB 196,98,125,24,5,9,243,1,0 ; vbroadcastss 0x1f309(%rip),%ymm8 # 2d2a4 <_sk_overlay_sse2_8bit+0xdf5> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10 @@ -13044,39 +13231,39 @@ _sk_store_u16_be_avx LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,31 ; jne dc69 <_sk_store_u16_be_avx+0x10b> - DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2) - DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2) - DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2) + DB 117,31 ; jne e07f <_sk_store_u16_be_avx+0x10b> + DB 196,65,122,127,28,64 ; vmovdqu %xmm11,(%r8,%rax,2) + DB 196,65,122,127,84,64,16 ; vmovdqu %xmm10,0x10(%r8,%rax,2) + DB 196,65,122,127,76,64,32 ; vmovdqu %xmm9,0x20(%r8,%rax,2) DB 196,65,122,127,68,64,48 ; vmovdqu %xmm8,0x30(%r8,%rax,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je dc65 <_sk_store_u16_be_avx+0x107> + DB 116,240 ; je e07b <_sk_store_u16_be_avx+0x107> DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb dc65 <_sk_store_u16_be_avx+0x107> + DB 114,227 ; jb e07b <_sk_store_u16_be_avx+0x107> DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je dc65 <_sk_store_u16_be_avx+0x107> + DB 116,218 ; je e07b <_sk_store_u16_be_avx+0x107> DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2) DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,205 ; jb dc65 <_sk_store_u16_be_avx+0x107> + DB 114,205 ; jb e07b <_sk_store_u16_be_avx+0x107> DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je dc65 <_sk_store_u16_be_avx+0x107> + DB 116,196 ; je e07b <_sk_store_u16_be_avx+0x107> DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2) DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,183 ; jb dc65 <_sk_store_u16_be_avx+0x107> + DB 114,183 ; jb e07b <_sk_store_u16_be_avx+0x107> DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp dc65 <_sk_store_u16_be_avx+0x107> + DB 235,174 ; jmp e07b <_sk_store_u16_be_avx+0x107> PUBLIC _sk_load_f32_avx _sk_load_f32_avx LABEL PROC + DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 197,253,87,192 ; vxorpd %ymm0,%ymm0,%ymm0 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,135,101,1,0,0 ; ja de2c <_sk_load_f32_avx+0x175> - DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp + DB 15,135,94,1,0,0 ; ja e242 <_sk_load_f32_avx+0x175> DB 197,253,17,36,36 ; vmovupd %ymm4,(%rsp) DB 197,253,17,108,36,32 ; vmovupd %ymm5,0x20(%rsp) DB 197,253,17,116,36,64 ; vmovupd %ymm6,0x40(%rsp) @@ -13088,7 +13275,7 @@ _sk_load_f32_avx LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 77,99,201 ; movslq %r9d,%r9 - DB 76,141,21,51,1,0,0 ; lea 0x133(%rip),%r10 # de3c <_sk_load_f32_avx+0x185> + DB 76,141,21,53,1,0,0 ; lea 0x135(%rip),%r10 # e254 <_sk_load_f32_avx+0x187> DB 73,99,4,186 ; movslq (%r10,%rdi,4),%rax DB 76,1,208 ; add %r10,%rax DB 197,221,87,228 ; vxorpd %ymm4,%ymm4,%ymm4 @@ -13137,7 +13324,6 @@ _sk_load_f32_avx LABEL PROC DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 DB 197,252,16,36,36 ; vmovups (%rsp),%ymm4 - DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 196,193,116,20,210 ; vunpcklps %ymm10,%ymm1,%ymm2 DB 196,65,116,21,194 ; vunpckhps %ymm10,%ymm1,%ymm8 DB 197,228,20,200 ; vunpcklps %ymm0,%ymm3,%ymm1 @@ -13147,37 +13333,39 @@ _sk_load_f32_avx LABEL PROC DB 197,189,20,211 ; vunpcklpd %ymm3,%ymm8,%ymm2 DB 197,189,21,219 ; vunpckhpd %ymm3,%ymm8,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax DB 197,229,87,219 ; vxorpd %ymm3,%ymm3,%ymm3 DB 196,65,45,87,210 ; vxorpd %ymm10,%ymm10,%ymm10 DB 197,245,87,201 ; vxorpd %ymm1,%ymm1,%ymm1 - DB 235,203 ; jmp de06 <_sk_load_f32_avx+0x14f> - DB 144 ; nop - DB 26,255 ; sbb %bh,%bh + DB 235,196 ; jmp e215 <_sk_load_f32_avx+0x148> + DB 15,31,0 ; nopl (%rax) + DB 24,255 ; sbb %bh,%bh DB 255 ; (bad) - DB 255,160,255,255,255,139 ; jmpq *-0x74000001(%rax) + DB 255,158,255,255,255,137 ; lcall *-0x76000001(%rsi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,118,255 ; pushq -0x1(%rsi) + DB 255,116,255,255 ; pushq -0x1(%rdi,%rdi,8) + DB 255,89,255 ; lcall *-0x1(%rcx) DB 255 ; (bad) - DB 255,91,255 ; lcall *-0x1(%rbx) + DB 255,67,255 ; incl -0x1(%rbx) DB 255 ; (bad) - DB 255,69,255 ; incl -0x1(%rbp) + DB 255,50 ; pushq (%rdx) DB 255 ; (bad) - DB 255,52,255 ; pushq (%rdi,%rdi,8) DB 255 ; (bad) - DB 255,39 ; jmpq *(%rdi) + DB 255 ; .byte 0xff + DB 37 ; .byte 0x25 DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_f32_dst_avx _sk_load_f32_dst_avx LABEL PROC + DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 197,221,87,228 ; vxorpd %ymm4,%ymm4,%ymm4 DB 72,131,255,7 ; cmp $0x7,%rdi - DB 15,135,101,1,0,0 ; ja dfd1 <_sk_load_f32_dst_avx+0x175> - DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp + DB 15,135,94,1,0,0 ; ja e3e9 <_sk_load_f32_dst_avx+0x175> DB 197,253,17,4,36 ; vmovupd %ymm0,(%rsp) DB 197,253,17,76,36,32 ; vmovupd %ymm1,0x20(%rsp) DB 197,253,17,84,36,64 ; vmovupd %ymm2,0x40(%rsp) @@ -13189,7 +13377,7 @@ _sk_load_f32_dst_avx LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 77,99,201 ; movslq %r9d,%r9 - DB 76,141,21,50,1,0,0 ; lea 0x132(%rip),%r10 # dfe0 <_sk_load_f32_dst_avx+0x184> + DB 76,141,21,50,1,0,0 ; lea 0x132(%rip),%r10 # e3f8 <_sk_load_f32_dst_avx+0x184> DB 73,99,4,186 ; movslq (%r10,%rdi,4),%rax DB 76,1,208 ; add %r10,%rax DB 197,253,87,192 ; vxorpd %ymm0,%ymm0,%ymm0 @@ -13238,7 +13426,6 @@ _sk_load_f32_dst_avx LABEL PROC DB 197,252,16,84,36,64 ; vmovups 0x40(%rsp),%ymm2 DB 197,252,16,76,36,32 ; vmovups 0x20(%rsp),%ymm1 DB 197,252,16,4,36 ; vmovups (%rsp),%ymm0 - DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 196,193,84,20,242 ; vunpcklps %ymm10,%ymm5,%ymm6 DB 196,65,84,21,194 ; vunpckhps %ymm10,%ymm5,%ymm8 DB 197,196,20,236 ; vunpcklps %ymm4,%ymm7,%ymm5 @@ -13248,11 +13435,12 @@ _sk_load_f32_dst_avx LABEL PROC DB 197,189,20,247 ; vunpcklpd %ymm7,%ymm8,%ymm6 DB 197,189,21,255 ; vunpckhpd %ymm7,%ymm8,%ymm7 DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax DB 197,197,87,255 ; vxorpd %ymm7,%ymm7,%ymm7 DB 196,65,45,87,210 ; vxorpd %ymm10,%ymm10,%ymm10 DB 197,213,87,237 ; vxorpd %ymm5,%ymm5,%ymm5 - DB 235,203 ; jmp dfab <_sk_load_f32_dst_avx+0x14f> + DB 235,196 ; jmp e3bc <_sk_load_f32_dst_avx+0x148> DB 27,255 ; sbb %edi,%edi DB 255 ; (bad) DB 255,161,255,255,255,140 ; jmpq *-0x73000001(%rcx) @@ -13263,7 +13451,7 @@ _sk_load_f32_dst_avx LABEL PROC DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) - DB 255,53,255,255,255,40 ; pushq 0x28ffffff(%rip) # 2900dffc <_sk_overlay_sse2_8bit+0x28fe1aa4> + DB 255,53,255,255,255,40 ; pushq 0x28ffffff(%rip) # 2900e414 <_sk_overlay_sse2_8bit+0x28fe1f65> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -13287,35 +13475,35 @@ _sk_store_f32_avx LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,255 ; test %rdi,%rdi - DB 117,55 ; jne e07f <_sk_store_f32_avx+0x7f> + DB 117,55 ; jne e497 <_sk_store_f32_avx+0x7f> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 DB 196,67,61,6,195,49 ; vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 - DB 196,65,125,17,36,128 ; vmovupd %ymm12,(%r8,%rax,4) - DB 196,65,125,17,108,128,32 ; vmovupd %ymm13,0x20(%r8,%rax,4) + DB 196,65,124,17,36,128 ; vmovups %ymm12,(%r8,%rax,4) + DB 196,65,124,17,108,128,32 ; vmovups %ymm13,0x20(%r8,%rax,4) DB 196,65,125,17,76,128,64 ; vmovupd %ymm9,0x40(%r8,%rax,4) DB 196,65,125,17,68,128,96 ; vmovupd %ymm8,0x60(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je e07b <_sk_store_f32_avx+0x7b> + DB 116,240 ; je e493 <_sk_store_f32_avx+0x7b> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb e07b <_sk_store_f32_avx+0x7b> + DB 114,227 ; jb e493 <_sk_store_f32_avx+0x7b> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je e07b <_sk_store_f32_avx+0x7b> + DB 116,218 ; je e493 <_sk_store_f32_avx+0x7b> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,255,5 ; cmp $0x5,%rdi - DB 114,205 ; jb e07b <_sk_store_f32_avx+0x7b> + DB 114,205 ; jb e493 <_sk_store_f32_avx+0x7b> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je e07b <_sk_store_f32_avx+0x7b> + DB 116,195 ; je e493 <_sk_store_f32_avx+0x7b> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,255,7 ; cmp $0x7,%rdi - DB 114,181 ; jb e07b <_sk_store_f32_avx+0x7b> + DB 114,181 ; jb e493 <_sk_store_f32_avx+0x7b> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp e07b <_sk_store_f32_avx+0x7b> + DB 235,171 ; jmp e493 <_sk_store_f32_avx+0x7b> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -13390,15 +13578,15 @@ _sk_repeat_y_avx LABEL PROC PUBLIC _sk_mirror_x_avx _sk_mirror_x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,121,110,0 ; vmovd (%rax),%xmm8 + DB 197,122,16,0 ; vmovss (%rax),%xmm8 DB 197,122,16,72,4 ; vmovss 0x4(%rax),%xmm9 - DB 196,65,121,112,208,0 ; vpshufd $0x0,%xmm8,%xmm10 + DB 196,67,121,4,208,0 ; vpermilps $0x0,%xmm8,%xmm10 DB 196,67,45,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 DB 196,65,124,92,218 ; vsubps %ymm10,%ymm0,%ymm11 DB 196,193,58,88,192 ; vaddss %xmm8,%xmm8,%xmm0 DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 DB 196,99,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm8 - DB 197,178,89,5,163,240,1,0 ; vmulss 0x1f0a3(%rip),%xmm9,%xmm0 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 197,178,89,5,107,235,1,0 ; vmulss 0x1eb6b(%rip),%xmm9,%xmm0 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 @@ -13422,15 +13610,15 @@ _sk_mirror_x_avx LABEL PROC PUBLIC _sk_mirror_y_avx _sk_mirror_y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,121,110,0 ; vmovd (%rax),%xmm8 + DB 197,122,16,0 ; vmovss (%rax),%xmm8 DB 197,122,16,72,4 ; vmovss 0x4(%rax),%xmm9 - DB 196,65,121,112,208,0 ; vpshufd $0x0,%xmm8,%xmm10 + DB 196,67,121,4,208,0 ; vpermilps $0x0,%xmm8,%xmm10 DB 196,67,45,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 DB 196,65,116,92,218 ; vsubps %ymm10,%ymm1,%ymm11 DB 196,193,58,88,200 ; vaddss %xmm8,%xmm8,%xmm1 DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,99,117,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm8 - DB 197,178,89,13,22,240,1,0 ; vmulss 0x1f016(%rip),%xmm9,%xmm1 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 197,178,89,13,222,234,1,0 ; vmulss 0x1eade(%rip),%xmm9,%xmm1 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 @@ -13455,7 +13643,7 @@ PUBLIC _sk_clamp_x_1_avx _sk_clamp_x_1_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,176,239,1,0 ; vbroadcastss 0x1efb0(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,120,234,1,0 ; vbroadcastss 0x1ea78(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -13469,9 +13657,9 @@ _sk_repeat_x_1_avx LABEL PROC PUBLIC _sk_mirror_x_1_avx _sk_mirror_x_1_avx LABEL PROC - DB 196,98,125,24,5,167,239,1,0 ; vbroadcastss 0x1efa7(%rip),%ymm8 # 2d2c4 <_sk_overlay_sse2_8bit+0xd6c> + DB 196,98,125,24,5,111,234,1,0 ; vbroadcastss 0x1ea6f(%rip),%ymm8 # 2d1a4 <_sk_overlay_sse2_8bit+0xcf5> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,125,239,1,0 ; vbroadcastss 0x1ef7d(%rip),%ymm9 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,13,69,234,1,0 ; vbroadcastss 0x1ea45(%rip),%ymm9 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -13485,12 +13673,12 @@ _sk_mirror_x_1_avx LABEL PROC PUBLIC _sk_luminance_to_alpha_avx _sk_luminance_to_alpha_avx LABEL PROC - DB 196,226,125,24,29,105,240,1,0 ; vbroadcastss 0x1f069(%rip),%ymm3 # 2d3c8 <_sk_overlay_sse2_8bit+0xe70> + DB 196,226,125,24,29,49,235,1,0 ; vbroadcastss 0x1eb31(%rip),%ymm3 # 2d2a8 <_sk_overlay_sse2_8bit+0xdf9> DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,29,96,240,1,0 ; vbroadcastss 0x1f060(%rip),%ymm3 # 2d3cc <_sk_overlay_sse2_8bit+0xe74> + DB 196,226,125,24,29,40,235,1,0 ; vbroadcastss 0x1eb28(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xdfd> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,83,240,1,0 ; vbroadcastss 0x1f053(%rip),%ymm1 # 2d3d0 <_sk_overlay_sse2_8bit+0xe78> + DB 196,226,125,24,13,27,235,1,0 ; vbroadcastss 0x1eb1b(%rip),%ymm1 # 2d2b0 <_sk_overlay_sse2_8bit+0xe01> DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 DB 197,252,88,217 ; vaddps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -13723,9 +13911,9 @@ _sk_evenly_spaced_gradient_avx LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 72,139,104,8 ; mov 0x8(%rax),%rbp DB 72,255,203 ; dec %rbx - DB 120,7 ; js e724 <_sk_evenly_spaced_gradient_avx+0x28> + DB 120,7 ; js eb3c <_sk_evenly_spaced_gradient_avx+0x28> DB 196,225,242,42,203 ; vcvtsi2ss %rbx,%xmm1,%xmm1 - DB 235,21 ; jmp e739 <_sk_evenly_spaced_gradient_avx+0x3d> + DB 235,21 ; jmp eb51 <_sk_evenly_spaced_gradient_avx+0x3d> DB 73,137,216 ; mov %rbx,%r8 DB 73,209,232 ; shr %r8 DB 131,227,1 ; and $0x1,%ebx @@ -13882,18 +14070,18 @@ _sk_evenly_spaced_gradient_avx LABEL PROC PUBLIC _sk_gauss_a_to_rgba_avx _sk_gauss_a_to_rgba_avx LABEL PROC - DB 196,226,125,24,5,104,233,1,0 ; vbroadcastss 0x1e968(%rip),%ymm0 # 2d3d8 <_sk_overlay_sse2_8bit+0xe80> + DB 196,226,125,24,5,48,228,1,0 ; vbroadcastss 0x1e430(%rip),%ymm0 # 2d2b8 <_sk_overlay_sse2_8bit+0xe09> DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 196,226,125,24,13,87,233,1,0 ; vbroadcastss 0x1e957(%rip),%ymm1 # 2d3d4 <_sk_overlay_sse2_8bit+0xe7c> + DB 196,226,125,24,13,31,228,1,0 ; vbroadcastss 0x1e41f(%rip),%ymm1 # 2d2b4 <_sk_overlay_sse2_8bit+0xe05> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,78,233,1,0 ; vbroadcastss 0x1e94e(%rip),%ymm1 # 2d3dc <_sk_overlay_sse2_8bit+0xe84> + DB 196,226,125,24,13,22,228,1,0 ; vbroadcastss 0x1e416(%rip),%ymm1 # 2d2bc <_sk_overlay_sse2_8bit+0xe0d> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,65,233,1,0 ; vbroadcastss 0x1e941(%rip),%ymm1 # 2d3e0 <_sk_overlay_sse2_8bit+0xe88> + DB 196,226,125,24,13,9,228,1,0 ; vbroadcastss 0x1e409(%rip),%ymm1 # 2d2c0 <_sk_overlay_sse2_8bit+0xe11> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,52,233,1,0 ; vbroadcastss 0x1e934(%rip),%ymm1 # 2d3e4 <_sk_overlay_sse2_8bit+0xe8c> + DB 196,226,125,24,13,252,227,1,0 ; vbroadcastss 0x1e3fc(%rip),%ymm1 # 2d2c4 <_sk_overlay_sse2_8bit+0xe15> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -13914,24 +14102,24 @@ _sk_gradient_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,80 ; jb eb35 <_sk_gradient_avx+0x71> + DB 114,81 ; jb ef4e <_sk_gradient_avx+0x72> DB 72,139,88,72 ; mov 0x48(%rax),%rbx DB 73,255,200 ; dec %r8 DB 72,131,195,4 ; add $0x4,%rbx DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 - DB 196,98,125,24,21,178,231,1,0 ; vbroadcastss 0x1e7b2(%rip),%ymm10 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> + DB 196,98,125,24,21,122,226,1,0 ; vbroadcastss 0x1e27a(%rip),%ymm10 # 2d190 <_sk_overlay_sse2_8bit+0xce1> DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,98,125,24,3 ; vbroadcastss (%rbx),%ymm8 DB 197,60,194,192,2 ; vcmpleps %ymm0,%ymm8,%ymm8 DB 196,67,53,74,194,128 ; vblendvps %ymm8,%ymm10,%ymm9,%ymm8 DB 196,99,125,25,194,1 ; vextractf128 $0x1,%ymm8,%xmm2 DB 196,227,125,25,203,1 ; vextractf128 $0x1,%ymm1,%xmm3 - DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 - DB 197,185,254,201 ; vpaddd %xmm1,%xmm8,%xmm1 + DB 197,225,254,210 ; vpaddd %xmm2,%xmm3,%xmm2 + DB 196,193,113,254,200 ; vpaddd %xmm8,%xmm1,%xmm1 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 72,131,195,4 ; add $0x4,%rbx DB 73,255,200 ; dec %r8 - DB 117,205 ; jne eb02 <_sk_gradient_avx+0x3e> + DB 117,204 ; jne ef1a <_sk_gradient_avx+0x3e> DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 73,193,232,32 ; shr $0x20,%r8 @@ -14110,27 +14298,27 @@ _sk_xy_to_unit_angle_avx LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,218,228,1,0 ; vbroadcastss 0x1e4da(%rip),%ymm13 # 2d3e8 <_sk_overlay_sse2_8bit+0xe90> + DB 196,98,125,24,45,161,223,1,0 ; vbroadcastss 0x1dfa1(%rip),%ymm13 # 2d2c8 <_sk_overlay_sse2_8bit+0xe19> DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,208,228,1,0 ; vbroadcastss 0x1e4d0(%rip),%ymm14 # 2d3ec <_sk_overlay_sse2_8bit+0xe94> + DB 196,98,125,24,53,151,223,1,0 ; vbroadcastss 0x1df97(%rip),%ymm14 # 2d2cc <_sk_overlay_sse2_8bit+0xe1d> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,193,228,1,0 ; vbroadcastss 0x1e4c1(%rip),%ymm14 # 2d3f0 <_sk_overlay_sse2_8bit+0xe98> + DB 196,98,125,24,53,136,223,1,0 ; vbroadcastss 0x1df88(%rip),%ymm14 # 2d2d0 <_sk_overlay_sse2_8bit+0xe21> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,229 ; vmulps %ymm13,%ymm12,%ymm12 - DB 196,98,125,24,45,178,228,1,0 ; vbroadcastss 0x1e4b2(%rip),%ymm13 # 2d3f4 <_sk_overlay_sse2_8bit+0xe9c> + DB 196,98,125,24,45,121,223,1,0 ; vbroadcastss 0x1df79(%rip),%ymm13 # 2d2d4 <_sk_overlay_sse2_8bit+0xe25> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,157,228,1,0 ; vbroadcastss 0x1e49d(%rip),%ymm10 # 2d3f8 <_sk_overlay_sse2_8bit+0xea0> + DB 196,98,125,24,21,100,223,1,0 ; vbroadcastss 0x1df64(%rip),%ymm10 # 2d2d8 <_sk_overlay_sse2_8bit+0xe29> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,51,227,1,0 ; vbroadcastss 0x1e333(%rip),%ymm10 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,21,250,221,1,0 ; vbroadcastss 0x1ddfa(%rip),%ymm10 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,29,227,1,0 ; vbroadcastss 0x1e31d(%rip),%ymm10 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,21,228,221,1,0 ; vbroadcastss 0x1dde4(%rip),%ymm10 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -14156,7 +14344,7 @@ _sk_xy_to_2pt_conical_quadratic_max_avx LABEL PROC DB 196,67,121,4,210,0 ; vpermilps $0x0,%xmm10,%xmm10 DB 196,67,45,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 DB 197,44,88,208 ; vaddps %ymm0,%ymm10,%ymm10 - DB 196,98,125,24,29,19,228,1,0 ; vbroadcastss 0x1e413(%rip),%ymm11 # 2d3fc <_sk_overlay_sse2_8bit+0xea4> + DB 196,98,125,24,29,218,222,1,0 ; vbroadcastss 0x1deda(%rip),%ymm11 # 2d2dc <_sk_overlay_sse2_8bit+0xe2d> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 DB 197,252,89,192 ; vmulps %ymm0,%ymm0,%ymm0 DB 197,116,89,217 ; vmulps %ymm1,%ymm1,%ymm11 @@ -14165,17 +14353,17 @@ _sk_xy_to_2pt_conical_quadratic_max_avx LABEL PROC DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,13,231,227,1,0 ; vbroadcastss 0x1e3e7(%rip),%ymm9 # 2d400 <_sk_overlay_sse2_8bit+0xea8> + DB 196,98,125,24,13,174,222,1,0 ; vbroadcastss 0x1deae(%rip),%ymm9 # 2d2e0 <_sk_overlay_sse2_8bit+0xe31> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,65,44,89,194 ; vmulps %ymm10,%ymm10,%ymm8 DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 DB 197,252,81,192 ; vsqrtps %ymm0,%ymm0 DB 196,98,125,24,64,36 ; vbroadcastss 0x24(%rax),%ymm8 - DB 196,98,125,24,13,197,227,1,0 ; vbroadcastss 0x1e3c5(%rip),%ymm9 # 2d404 <_sk_overlay_sse2_8bit+0xeac> + DB 196,98,125,24,13,140,222,1,0 ; vbroadcastss 0x1de8c(%rip),%ymm9 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> DB 196,65,44,87,201 ; vxorps %ymm9,%ymm10,%ymm9 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,86,226,1,0 ; vbroadcastss 0x1e256(%rip),%ymm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,29,29,221,1,0 ; vbroadcastss 0x1dd1d(%rip),%ymm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 196,65,60,89,210 ; vmulps %ymm10,%ymm8,%ymm10 DB 197,180,92,192 ; vsubps %ymm0,%ymm9,%ymm0 @@ -14193,7 +14381,7 @@ _sk_xy_to_2pt_conical_quadratic_min_avx LABEL PROC DB 196,67,121,4,210,0 ; vpermilps $0x0,%xmm10,%xmm10 DB 196,67,45,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 DB 197,44,88,208 ; vaddps %ymm0,%ymm10,%ymm10 - DB 196,98,125,24,29,101,227,1,0 ; vbroadcastss 0x1e365(%rip),%ymm11 # 2d3fc <_sk_overlay_sse2_8bit+0xea4> + DB 196,98,125,24,29,44,222,1,0 ; vbroadcastss 0x1de2c(%rip),%ymm11 # 2d2dc <_sk_overlay_sse2_8bit+0xe2d> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 DB 197,252,89,192 ; vmulps %ymm0,%ymm0,%ymm0 DB 197,116,89,217 ; vmulps %ymm1,%ymm1,%ymm11 @@ -14202,17 +14390,17 @@ _sk_xy_to_2pt_conical_quadratic_min_avx LABEL PROC DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 - DB 196,98,125,24,13,57,227,1,0 ; vbroadcastss 0x1e339(%rip),%ymm9 # 2d400 <_sk_overlay_sse2_8bit+0xea8> + DB 196,98,125,24,13,0,222,1,0 ; vbroadcastss 0x1de00(%rip),%ymm9 # 2d2e0 <_sk_overlay_sse2_8bit+0xe31> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 196,65,44,89,194 ; vmulps %ymm10,%ymm10,%ymm8 DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 DB 197,252,81,192 ; vsqrtps %ymm0,%ymm0 DB 196,98,125,24,64,36 ; vbroadcastss 0x24(%rax),%ymm8 - DB 196,98,125,24,13,23,227,1,0 ; vbroadcastss 0x1e317(%rip),%ymm9 # 2d404 <_sk_overlay_sse2_8bit+0xeac> + DB 196,98,125,24,13,222,221,1,0 ; vbroadcastss 0x1ddde(%rip),%ymm9 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> DB 196,65,44,87,201 ; vxorps %ymm9,%ymm10,%ymm9 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,168,225,1,0 ; vbroadcastss 0x1e1a8(%rip),%ymm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,29,111,220,1,0 ; vbroadcastss 0x1dc6f(%rip),%ymm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 196,65,60,89,210 ; vmulps %ymm10,%ymm8,%ymm10 DB 197,180,92,192 ; vsubps %ymm0,%ymm9,%ymm0 @@ -14229,7 +14417,7 @@ _sk_xy_to_2pt_conical_linear_avx LABEL PROC DB 196,67,121,4,201,0 ; vpermilps $0x0,%xmm9,%xmm9 DB 196,67,53,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 DB 197,52,88,200 ; vaddps %ymm0,%ymm9,%ymm9 - DB 196,98,125,24,21,189,226,1,0 ; vbroadcastss 0x1e2bd(%rip),%ymm10 # 2d3fc <_sk_overlay_sse2_8bit+0xea4> + DB 196,98,125,24,21,132,221,1,0 ; vbroadcastss 0x1dd84(%rip),%ymm10 # 2d2dc <_sk_overlay_sse2_8bit+0xe2d> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 DB 197,252,89,192 ; vmulps %ymm0,%ymm0,%ymm0 DB 197,116,89,209 ; vmulps %ymm1,%ymm1,%ymm10 @@ -14238,7 +14426,7 @@ _sk_xy_to_2pt_conical_linear_avx LABEL PROC DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,172,92,192 ; vsubps %ymm0,%ymm10,%ymm0 - DB 196,98,125,24,5,149,226,1,0 ; vbroadcastss 0x1e295(%rip),%ymm8 # 2d404 <_sk_overlay_sse2_8bit+0xeac> + DB 196,98,125,24,5,92,221,1,0 ; vbroadcastss 0x1dd5c(%rip),%ymm8 # 2d2e4 <_sk_overlay_sse2_8bit+0xe35> DB 196,193,124,87,192 ; vxorps %ymm8,%ymm0,%ymm0 DB 196,193,124,94,193 ; vdivps %ymm9,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax @@ -14273,7 +14461,7 @@ _sk_apply_vector_mask_avx LABEL PROC PUBLIC _sk_save_xy_avx _sk_save_xy_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,209,224,1,0 ; vbroadcastss 0x1e0d1(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,152,219,1,0 ; vbroadcastss 0x1db98(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9 DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 @@ -14306,9 +14494,9 @@ _sk_accumulate_avx LABEL PROC PUBLIC _sk_bilinear_nx_avx _sk_bilinear_nx_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,185,225,1,0 ; vbroadcastss 0x1e1b9(%rip),%ymm0 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,5,128,220,1,0 ; vbroadcastss 0x1dc80(%rip),%ymm0 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,80,224,1,0 ; vbroadcastss 0x1e050(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,23,219,1,0 ; vbroadcastss 0x1db17(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -14317,7 +14505,7 @@ _sk_bilinear_nx_avx LABEL PROC PUBLIC _sk_bilinear_px_avx _sk_bilinear_px_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,48,224,1,0 ; vbroadcastss 0x1e030(%rip),%ymm0 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,5,247,218,1,0 ; vbroadcastss 0x1daf7(%rip),%ymm0 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -14327,9 +14515,9 @@ _sk_bilinear_px_avx LABEL PROC PUBLIC _sk_bilinear_ny_avx _sk_bilinear_ny_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,112,225,1,0 ; vbroadcastss 0x1e170(%rip),%ymm1 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,13,55,220,1,0 ; vbroadcastss 0x1dc37(%rip),%ymm1 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,6,224,1,0 ; vbroadcastss 0x1e006(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,205,218,1,0 ; vbroadcastss 0x1dacd(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -14338,7 +14526,7 @@ _sk_bilinear_ny_avx LABEL PROC PUBLIC _sk_bilinear_py_avx _sk_bilinear_py_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,230,223,1,0 ; vbroadcastss 0x1dfe6(%rip),%ymm1 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,226,125,24,13,173,218,1,0 ; vbroadcastss 0x1daad(%rip),%ymm1 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -14348,14 +14536,14 @@ _sk_bilinear_py_avx LABEL PROC PUBLIC _sk_bicubic_n3x_avx _sk_bicubic_n3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,41,225,1,0 ; vbroadcastss 0x1e129(%rip),%ymm0 # 2d40c <_sk_overlay_sse2_8bit+0xeb4> + DB 196,226,125,24,5,240,219,1,0 ; vbroadcastss 0x1dbf0(%rip),%ymm0 # 2d2ec <_sk_overlay_sse2_8bit+0xe3d> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,188,223,1,0 ; vbroadcastss 0x1dfbc(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,131,218,1,0 ; vbroadcastss 0x1da83(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,13,225,1,0 ; vbroadcastss 0x1e10d(%rip),%ymm10 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,212,219,1,0 ; vbroadcastss 0x1dbd4(%rip),%ymm10 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,11,224,1,0 ; vbroadcastss 0x1e00b(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> + DB 196,98,125,24,21,210,218,1,0 ; vbroadcastss 0x1dad2(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -14365,19 +14553,19 @@ _sk_bicubic_n3x_avx LABEL PROC PUBLIC _sk_bicubic_n1x_avx _sk_bicubic_n1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,214,224,1,0 ; vbroadcastss 0x1e0d6(%rip),%ymm0 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,5,157,219,1,0 ; vbroadcastss 0x1db9d(%rip),%ymm0 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,109,223,1,0 ; vbroadcastss 0x1df6d(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,52,218,1,0 ; vbroadcastss 0x1da34(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,203,224,1,0 ; vbroadcastss 0x1e0cb(%rip),%ymm9 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,13,146,219,1,0 ; vbroadcastss 0x1db92(%rip),%ymm9 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,185,224,1,0 ; vbroadcastss 0x1e0b9(%rip),%ymm10 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,98,125,24,21,128,219,1,0 ; vbroadcastss 0x1db80(%rip),%ymm10 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,58,223,1,0 ; vbroadcastss 0x1df3a(%rip),%ymm10 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,21,1,218,1,0 ; vbroadcastss 0x1da01(%rip),%ymm10 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,155,224,1,0 ; vbroadcastss 0x1e09b(%rip),%ymm9 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,13,98,219,1,0 ; vbroadcastss 0x1db62(%rip),%ymm9 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -14386,17 +14574,17 @@ _sk_bicubic_n1x_avx LABEL PROC PUBLIC _sk_bicubic_p1x_avx _sk_bicubic_p1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,11,223,1,0 ; vbroadcastss 0x1df0b(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,210,217,1,0 ; vbroadcastss 0x1d9d2(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,105,224,1,0 ; vbroadcastss 0x1e069(%rip),%ymm10 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,21,48,219,1,0 ; vbroadcastss 0x1db30(%rip),%ymm10 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,87,224,1,0 ; vbroadcastss 0x1e057(%rip),%ymm11 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,98,125,24,29,30,219,1,0 ; vbroadcastss 0x1db1e(%rip),%ymm11 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,66,224,1,0 ; vbroadcastss 0x1e042(%rip),%ymm9 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,13,9,219,1,0 ; vbroadcastss 0x1db09(%rip),%ymm9 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -14405,13 +14593,13 @@ _sk_bicubic_p1x_avx LABEL PROC PUBLIC _sk_bicubic_p3x_avx _sk_bicubic_p3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,30,224,1,0 ; vbroadcastss 0x1e01e(%rip),%ymm0 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,226,125,24,5,229,218,1,0 ; vbroadcastss 0x1dae5(%rip),%ymm0 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,3,224,1,0 ; vbroadcastss 0x1e003(%rip),%ymm10 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,202,218,1,0 ; vbroadcastss 0x1daca(%rip),%ymm10 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,1,223,1,0 ; vbroadcastss 0x1df01(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> + DB 196,98,125,24,21,200,217,1,0 ; vbroadcastss 0x1d9c8(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -14421,14 +14609,14 @@ _sk_bicubic_p3x_avx LABEL PROC PUBLIC _sk_bicubic_n3y_avx _sk_bicubic_n3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,208,223,1,0 ; vbroadcastss 0x1dfd0(%rip),%ymm1 # 2d40c <_sk_overlay_sse2_8bit+0xeb4> + DB 196,226,125,24,13,151,218,1,0 ; vbroadcastss 0x1da97(%rip),%ymm1 # 2d2ec <_sk_overlay_sse2_8bit+0xe3d> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,98,222,1,0 ; vbroadcastss 0x1de62(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,41,217,1,0 ; vbroadcastss 0x1d929(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,179,223,1,0 ; vbroadcastss 0x1dfb3(%rip),%ymm10 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,122,218,1,0 ; vbroadcastss 0x1da7a(%rip),%ymm10 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,177,222,1,0 ; vbroadcastss 0x1deb1(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> + DB 196,98,125,24,21,120,217,1,0 ; vbroadcastss 0x1d978(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -14438,19 +14626,19 @@ _sk_bicubic_n3y_avx LABEL PROC PUBLIC _sk_bicubic_n1y_avx _sk_bicubic_n1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,124,223,1,0 ; vbroadcastss 0x1df7c(%rip),%ymm1 # 2d408 <_sk_overlay_sse2_8bit+0xeb0> + DB 196,226,125,24,13,67,218,1,0 ; vbroadcastss 0x1da43(%rip),%ymm1 # 2d2e8 <_sk_overlay_sse2_8bit+0xe39> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,18,222,1,0 ; vbroadcastss 0x1de12(%rip),%ymm8 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> + DB 196,98,125,24,5,217,216,1,0 ; vbroadcastss 0x1d8d9(%rip),%ymm8 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,112,223,1,0 ; vbroadcastss 0x1df70(%rip),%ymm9 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,13,55,218,1,0 ; vbroadcastss 0x1da37(%rip),%ymm9 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,94,223,1,0 ; vbroadcastss 0x1df5e(%rip),%ymm10 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,98,125,24,21,37,218,1,0 ; vbroadcastss 0x1da25(%rip),%ymm10 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,223,221,1,0 ; vbroadcastss 0x1dddf(%rip),%ymm10 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,21,166,216,1,0 ; vbroadcastss 0x1d8a6(%rip),%ymm10 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,64,223,1,0 ; vbroadcastss 0x1df40(%rip),%ymm9 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,13,7,218,1,0 ; vbroadcastss 0x1da07(%rip),%ymm9 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -14459,17 +14647,17 @@ _sk_bicubic_n1y_avx LABEL PROC PUBLIC _sk_bicubic_p1y_avx _sk_bicubic_p1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,176,221,1,0 ; vbroadcastss 0x1ddb0(%rip),%ymm8 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 196,98,125,24,5,119,216,1,0 ; vbroadcastss 0x1d877(%rip),%ymm8 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,13,223,1,0 ; vbroadcastss 0x1df0d(%rip),%ymm10 # 2d418 <_sk_overlay_sse2_8bit+0xec0> + DB 196,98,125,24,21,212,217,1,0 ; vbroadcastss 0x1d9d4(%rip),%ymm10 # 2d2f8 <_sk_overlay_sse2_8bit+0xe49> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,251,222,1,0 ; vbroadcastss 0x1defb(%rip),%ymm11 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,98,125,24,29,194,217,1,0 ; vbroadcastss 0x1d9c2(%rip),%ymm11 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,230,222,1,0 ; vbroadcastss 0x1dee6(%rip),%ymm9 # 2d41c <_sk_overlay_sse2_8bit+0xec4> + DB 196,98,125,24,13,173,217,1,0 ; vbroadcastss 0x1d9ad(%rip),%ymm9 # 2d2fc <_sk_overlay_sse2_8bit+0xe4d> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -14478,13 +14666,13 @@ _sk_bicubic_p1y_avx LABEL PROC PUBLIC _sk_bicubic_p3y_avx _sk_bicubic_p3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,194,222,1,0 ; vbroadcastss 0x1dec2(%rip),%ymm1 # 2d414 <_sk_overlay_sse2_8bit+0xebc> + DB 196,226,125,24,13,137,217,1,0 ; vbroadcastss 0x1d989(%rip),%ymm1 # 2d2f4 <_sk_overlay_sse2_8bit+0xe45> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,166,222,1,0 ; vbroadcastss 0x1dea6(%rip),%ymm10 # 2d410 <_sk_overlay_sse2_8bit+0xeb8> + DB 196,98,125,24,21,109,217,1,0 ; vbroadcastss 0x1d96d(%rip),%ymm10 # 2d2f0 <_sk_overlay_sse2_8bit+0xe41> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,164,221,1,0 ; vbroadcastss 0x1dda4(%rip),%ymm10 # 2d31c <_sk_overlay_sse2_8bit+0xdc4> + DB 196,98,125,24,21,107,216,1,0 ; vbroadcastss 0x1d86b(%rip),%ymm10 # 2d1fc <_sk_overlay_sse2_8bit+0xd4d> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -14520,8 +14708,8 @@ _sk_callback_avx LABEL PROC DB 196,227,93,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm4,%ymm5 DB 196,227,101,6,201,49 ; vperm2f128 $0x31,%ymm1,%ymm3,%ymm1 DB 196,227,93,6,192,49 ; vperm2f128 $0x31,%ymm0,%ymm4,%ymm0 - DB 197,253,17,83,8 ; vmovupd %ymm2,0x8(%rbx) - DB 197,253,17,107,40 ; vmovupd %ymm5,0x28(%rbx) + DB 197,252,17,83,8 ; vmovups %ymm2,0x8(%rbx) + DB 197,252,17,107,40 ; vmovups %ymm5,0x28(%rbx) DB 197,253,17,75,72 ; vmovupd %ymm1,0x48(%rbx) DB 197,253,17,67,104 ; vmovupd %ymm0,0x68(%rbx) DB 72,133,255 ; test %rdi,%rdi @@ -14563,14 +14751,16 @@ _sk_callback_avx LABEL PROC PUBLIC _sk_clut_3D_avx _sk_clut_3D_avx LABEL PROC + DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,136,2,0,0 ; sub $0x288,%rsp - DB 197,252,17,188,36,96,2,0,0 ; vmovups %ymm7,0x260(%rsp) - DB 197,252,17,180,36,64,2,0,0 ; vmovups %ymm6,0x240(%rsp) - DB 197,252,17,172,36,32,2,0,0 ; vmovups %ymm5,0x220(%rsp) - DB 197,252,17,164,36,0,2,0,0 ; vmovups %ymm4,0x200(%rsp) - DB 197,252,17,156,36,224,1,0,0 ; vmovups %ymm3,0x1e0(%rsp) + DB 72,129,236,88,2,0,0 ; sub $0x258,%rsp + DB 197,252,17,188,36,32,2,0,0 ; vmovups %ymm7,0x220(%rsp) + DB 197,252,17,180,36,0,2,0,0 ; vmovups %ymm6,0x200(%rsp) + DB 197,252,17,172,36,224,1,0,0 ; vmovups %ymm5,0x1e0(%rsp) + DB 197,252,17,164,36,192,1,0,0 ; vmovups %ymm4,0x1c0(%rsp) + DB 197,252,17,156,36,160,1,0,0 ; vmovups %ymm3,0x1a0(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -14579,7 +14769,7 @@ _sk_clut_3D_avx LABEL PROC DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 - DB 197,252,17,148,36,128,1,0,0 ; vmovups %ymm2,0x180(%rsp) + DB 197,252,17,148,36,64,1,0,0 ; vmovups %ymm2,0x140(%rsp) DB 197,254,91,218 ; vcvttps2dq %ymm2,%ymm3 DB 68,139,72,12 ; mov 0xc(%rax),%r9d DB 69,141,81,255 ; lea -0x1(%r9),%r10d @@ -14588,813 +14778,792 @@ _sk_clut_3D_avx LABEL PROC DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 - DB 197,252,17,140,36,128,0,0,0 ; vmovups %ymm1,0x80(%rsp) + DB 197,252,17,76,36,64 ; vmovups %ymm1,0x40(%rsp) DB 197,254,91,225 ; vcvttps2dq %ymm1,%ymm4 DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 - DB 197,121,112,233,0 ; vpshufd $0x0,%xmm1,%xmm13 - DB 196,194,89,64,213 ; vpmulld %xmm13,%xmm4,%xmm2 - DB 197,253,111,236 ; vmovdqa %ymm4,%ymm5 - DB 197,254,127,172,36,224,0,0,0 ; vmovdqu %ymm5,0xe0(%rsp) - DB 197,249,127,148,36,32,1,0,0 ; vmovdqa %xmm2,0x120(%rsp) - DB 197,105,254,211 ; vpaddd %xmm3,%xmm2,%xmm10 - DB 197,254,127,156,36,96,1,0,0 ; vmovdqu %ymm3,0x160(%rsp) + DB 197,121,112,201,0 ; vpshufd $0x0,%xmm1,%xmm9 + DB 196,227,125,25,226,1 ; vextractf128 $0x1,%ymm4,%xmm2 + DB 197,253,111,244 ; vmovdqa %ymm4,%ymm6 + DB 197,254,127,180,36,192,0,0,0 ; vmovdqu %ymm6,0xc0(%rsp) + DB 196,226,49,64,226 ; vpmulld %xmm2,%xmm9,%xmm4 + DB 197,249,127,164,36,16,1,0,0 ; vmovdqa %xmm4,0x110(%rsp) + DB 196,227,125,25,223,1 ; vextractf128 $0x1,%ymm3,%xmm7 + DB 197,249,127,124,36,96 ; vmovdqa %xmm7,0x60(%rsp) + DB 197,253,111,235 ; vmovdqa %ymm3,%ymm5 + DB 197,254,127,172,36,32,1,0,0 ; vmovdqu %ymm5,0x120(%rsp) DB 196,193,121,110,209 ; vmovd %r9d,%xmm2 DB 196,226,105,64,201 ; vpmulld %xmm1,%xmm2,%xmm1 - DB 197,121,112,241,0 ; vpshufd $0x0,%xmm1,%xmm14 + DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 DB 68,139,64,8 ; mov 0x8(%rax),%r8d DB 65,255,200 ; dec %r8d DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 197,116,89,192 ; vmulps %ymm0,%ymm1,%ymm8 - DB 196,193,126,91,200 ; vcvttps2dq %ymm8,%ymm1 - DB 196,194,113,64,198 ; vpmulld %xmm14,%xmm1,%xmm0 - DB 197,253,111,241 ; vmovdqa %ymm1,%ymm6 - DB 197,254,127,116,36,32 ; vmovdqu %ymm6,0x20(%rsp) - DB 197,249,127,68,36,96 ; vmovdqa %xmm0,0x60(%rsp) + DB 197,116,89,232 ; vmulps %ymm0,%ymm1,%ymm13 + DB 196,65,126,91,221 ; vcvttps2dq %ymm13,%ymm11 + DB 196,99,125,25,216,1 ; vextractf128 $0x1,%ymm11,%xmm0 + DB 196,226,97,64,192 ; vpmulld %xmm0,%xmm3,%xmm0 + DB 197,249,127,68,36,48 ; vmovdqa %xmm0,0x30(%rsp) DB 72,139,0 ; mov (%rax),%rax - DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 - DB 196,98,121,24,29,90,220,1,0 ; vbroadcastss 0x1dc5a(%rip),%xmm11 # 2d424 <_sk_overlay_sse2_8bit+0xecc> - DB 196,194,121,64,203 ; vpmulld %xmm11,%xmm0,%xmm1 - DB 196,193,249,126,200 ; vmovq %xmm1,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,4,136 ; vmovss (%rax,%r9,4),%xmm0 - DB 196,163,121,33,4,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm0,%xmm0 - DB 196,163,121,33,4,152,32 ; vinsertps $0x20,(%rax,%r11,4),%xmm0,%xmm0 - DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 - DB 196,227,121,33,226,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm4 - DB 196,227,125,25,232,1 ; vextractf128 $0x1,%ymm5,%xmm0 - DB 196,194,121,64,213 ; vpmulld %xmm13,%xmm0,%xmm2 - DB 197,249,127,148,36,0,1,0,0 ; vmovdqa %xmm2,0x100(%rsp) - DB 196,227,125,25,240,1 ; vextractf128 $0x1,%ymm6,%xmm0 - DB 196,194,121,64,238 ; vpmulld %xmm14,%xmm0,%xmm5 - DB 197,249,127,108,36,80 ; vmovdqa %xmm5,0x50(%rsp) - DB 196,227,125,25,216,1 ; vextractf128 $0x1,%ymm3,%xmm0 - DB 197,248,41,4,36 ; vmovaps %xmm0,(%rsp) - DB 197,233,254,192 ; vpaddd %xmm0,%xmm2,%xmm0 - DB 197,209,254,208 ; vpaddd %xmm0,%xmm5,%xmm2 - DB 196,194,105,64,211 ; vpmulld %xmm11,%xmm2,%xmm2 + DB 197,217,254,207 ; vpaddd %xmm7,%xmm4,%xmm1 + DB 197,249,254,193 ; vpaddd %xmm1,%xmm0,%xmm0 + DB 196,226,121,24,37,15,215,1,0 ; vbroadcastss 0x1d70f(%rip),%xmm4 # 2d304 <_sk_overlay_sse2_8bit+0xe55> + DB 196,98,121,64,228 ; vpmulld %xmm4,%xmm0,%xmm12 + DB 196,226,49,64,198 ; vpmulld %xmm6,%xmm9,%xmm0 + DB 197,249,127,132,36,0,1,0,0 ; vmovdqa %xmm0,0x100(%rsp) + DB 197,249,254,197 ; vpaddd %xmm5,%xmm0,%xmm0 + DB 196,194,97,64,211 ; vpmulld %xmm11,%xmm3,%xmm2 + DB 197,249,127,84,36,32 ; vmovdqa %xmm2,0x20(%rsp) + DB 197,233,254,208 ; vpaddd %xmm0,%xmm2,%xmm2 + DB 196,226,105,64,212 ; vpmulld %xmm4,%xmm2,%xmm2 DB 196,193,121,126,208 ; vmovd %xmm2,%r8d DB 196,195,121,22,209,1 ; vpextrd $0x1,%xmm2,%r9d DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 - DB 196,163,65,33,60,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,163,65,33,60,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm7,%xmm7 - DB 196,227,93,24,223,1 ; vinsertf128 $0x1,%xmm7,%ymm4,%ymm3 - DB 197,252,17,156,36,192,0,0,0 ; vmovups %ymm3,0xc0(%rsp) - DB 196,98,121,24,13,33,218,1,0 ; vbroadcastss 0x1da21(%rip),%xmm9 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> - DB 196,193,105,254,249 ; vpaddd %xmm9,%xmm2,%xmm7 - DB 196,195,249,22,248,1 ; vpextrq $0x1,%xmm7,%r8 - DB 196,193,249,126,249 ; vmovq %xmm7,%r9 - DB 196,193,113,254,249 ; vpaddd %xmm9,%xmm1,%xmm7 - DB 196,193,249,126,250 ; vmovq %xmm7,%r10 - DB 196,195,249,22,251,1 ; vpextrq $0x1,%xmm7,%r11 + DB 196,65,121,126,231 ; vmovd %xmm12,%r15d + DB 196,67,121,22,230,1 ; vpextrd $0x1,%xmm12,%r14d + DB 196,67,121,22,228,2 ; vpextrd $0x2,%xmm12,%r12d + DB 196,99,121,22,227,3 ; vpextrd $0x3,%xmm12,%ebx + DB 196,161,122,16,44,184 ; vmovss (%rax,%r15,4),%xmm5 + DB 196,163,81,33,44,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm5,%xmm5 + DB 196,163,81,33,44,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm5,%xmm5 + DB 196,227,81,33,44,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm5,%xmm5 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,163,73,33,52,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm6,%xmm6 + DB 196,163,73,33,52,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm6,%xmm6 + DB 196,99,77,24,245,1 ; vinsertf128 $0x1,%xmm5,%ymm6,%ymm14 + DB 197,201,118,246 ; vpcmpeqd %xmm6,%xmm6,%xmm6 + DB 197,233,250,238 ; vpsubd %xmm6,%xmm2,%xmm5 + DB 196,195,249,22,232,1 ; vpextrq $0x1,%xmm5,%r8 + DB 196,193,249,126,233 ; vmovq %xmm5,%r9 + DB 197,153,250,238 ; vpsubd %xmm6,%xmm12,%xmm5 + DB 196,195,249,22,234,1 ; vpextrq $0x1,%xmm5,%r10 + DB 196,225,249,126,235 ; vmovq %xmm5,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,44,152 ; vmovss (%rax,%r11,4),%xmm5 + DB 196,227,81,33,44,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm5,%xmm5 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,81,33,238,32 ; vinsertps $0x20,%xmm6,%xmm5,%xmm5 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,81,33,238,48 ; vinsertps $0x30,%xmm6,%xmm5,%xmm5 DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 - DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 - DB 196,227,65,33,246,32 ; vinsertps $0x20,%xmm6,%xmm7,%xmm6 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 - DB 68,137,211 ; mov %r10d,%ebx - DB 69,137,216 ; mov %r11d,%r8d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 - DB 196,227,65,33,237,32 ; vinsertps $0x20,%xmm5,%xmm7,%xmm5 - DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,161,122,16,60,152 ; vmovss (%rax,%r11,4),%xmm7 - DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 - DB 196,99,85,24,254,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm15 - DB 196,98,121,24,37,150,217,1,0 ; vbroadcastss 0x1d996(%rip),%xmm12 # 2d2b8 <_sk_overlay_sse2_8bit+0xd60> - DB 196,193,105,254,212 ; vpaddd %xmm12,%xmm2,%xmm2 + DB 196,227,77,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm6,%ymm5 + DB 196,98,121,24,21,120,212,1,0 ; vbroadcastss 0x1d478(%rip),%xmm10 # 2d198 <_sk_overlay_sse2_8bit+0xce9> + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,113,254,204 ; vpaddd %xmm12,%xmm1,%xmm1 - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 - DB 196,227,105,33,213,32 ; vinsertps $0x20,%xmm5,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 - DB 196,227,105,33,213,48 ; vinsertps $0x30,%xmm5,%xmm2,%xmm2 - DB 196,227,109,24,225,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm4 - DB 196,226,125,24,13,106,218,1,0 ; vbroadcastss 0x1da6a(%rip),%ymm1 # 2d420 <_sk_overlay_sse2_8bit+0xec8> - DB 197,188,88,209 ; vaddps %ymm1,%ymm8,%ymm2 - DB 197,252,17,140,36,160,0,0,0 ; vmovups %ymm1,0xa0(%rsp) - DB 197,254,91,210 ; vcvttps2dq %ymm2,%ymm2 - DB 196,227,125,25,213,1 ; vextractf128 $0x1,%ymm2,%xmm5 - DB 196,194,81,64,222 ; vpmulld %xmm14,%xmm5,%xmm3 - DB 197,249,127,92,36,112 ; vmovdqa %xmm3,0x70(%rsp) - DB 196,194,105,64,214 ; vpmulld %xmm14,%xmm2,%xmm2 - DB 197,249,127,84,36,64 ; vmovdqa %xmm2,0x40(%rsp) - DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 - DB 196,194,105,64,235 ; vpmulld %xmm11,%xmm2,%xmm5 - DB 196,193,249,126,233 ; vmovq %xmm5,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,235,1 ; vpextrq $0x1,%xmm5,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,193,25,254,210 ; vpaddd %xmm10,%xmm12,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,105,33,214,32 ; vinsertps $0x20,%xmm6,%xmm2,%xmm2 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 DB 196,227,105,33,214,48 ; vinsertps $0x30,%xmm6,%xmm2,%xmm2 - DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 - DB 196,194,121,64,195 ; vpmulld %xmm11,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx - DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,163,73,33,52,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm6,%xmm6 - DB 196,163,73,33,52,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm6,%xmm6 - DB 196,163,73,33,52,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm6,%xmm6 - DB 196,227,109,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 - DB 196,193,121,254,241 ; vpaddd %xmm9,%xmm0,%xmm6 - DB 196,195,249,22,240,1 ; vpextrq $0x1,%xmm6,%r8 - DB 196,193,249,126,241 ; vmovq %xmm6,%r9 - DB 196,193,81,254,241 ; vpaddd %xmm9,%xmm5,%xmm6 - DB 196,193,249,126,242 ; vmovq %xmm6,%r10 - DB 196,227,249,22,243,1 ; vpextrq $0x1,%xmm6,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,52,152 ; vmovss (%rax,%r11,4),%xmm6 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 - DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,65,33,219,32 ; vinsertps $0x20,%xmm3,%xmm7,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 - DB 196,99,101,24,214,1 ; vinsertf128 $0x1,%xmm6,%ymm3,%ymm10 - DB 196,193,121,254,196 ; vpaddd %xmm12,%xmm0,%xmm0 + DB 196,227,77,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm6,%ymm2 + DB 196,226,125,24,61,76,213,1,0 ; vbroadcastss 0x1d54c(%rip),%ymm7 # 2d300 <_sk_overlay_sse2_8bit+0xe51> + DB 197,148,88,247 ; vaddps %ymm7,%ymm13,%ymm6 + DB 197,124,40,231 ; vmovaps %ymm7,%ymm12 + DB 197,124,17,164,36,128,0,0,0 ; vmovups %ymm12,0x80(%rsp) + DB 197,254,91,246 ; vcvttps2dq %ymm6,%ymm6 + DB 196,227,125,25,247,1 ; vextractf128 $0x1,%ymm6,%xmm7 + DB 196,226,97,64,255 ; vpmulld %xmm7,%xmm3,%xmm7 + DB 197,249,127,124,36,16 ; vmovdqa %xmm7,0x10(%rsp) + DB 196,226,97,64,222 ; vpmulld %xmm6,%xmm3,%xmm3 + DB 197,249,127,28,36 ; vmovdqa %xmm3,(%rsp) + DB 197,225,254,192 ; vpaddd %xmm0,%xmm3,%xmm0 + DB 197,193,254,201 ; vpaddd %xmm1,%xmm7,%xmm1 + DB 196,226,113,64,204 ; vpmulld %xmm4,%xmm1,%xmm1 + DB 196,226,121,64,220 ; vpmulld %xmm4,%xmm0,%xmm3 + DB 196,193,121,126,217 ; vmovd %xmm3,%r9d + DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d + DB 196,195,121,22,218,2 ; vpextrd $0x2,%xmm3,%r10d + DB 196,195,121,22,219,3 ; vpextrd $0x3,%xmm3,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 + DB 196,163,73,33,52,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm6,%xmm6 + DB 196,163,73,33,52,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm6,%xmm6 + DB 196,163,73,33,52,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm6,%xmm6 + DB 196,99,77,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm6,%ymm8 + DB 197,201,118,246 ; vpcmpeqd %xmm6,%xmm6,%xmm6 + DB 197,225,250,198 ; vpsubd %xmm6,%xmm3,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,81,254,196 ; vpaddd %xmm12,%xmm5,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 197,241,250,198 ; vpsubd %xmm6,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 + DB 196,227,121,33,199,32 ; vinsertps $0x20,%xmm7,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 - DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,121,33,199,48 ; vinsertps $0x30,%xmm7,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 + DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 + DB 196,33,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm15 + DB 196,195,65,33,255,32 ; vinsertps $0x20,%xmm15,%xmm7,%xmm7 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,65,33,246,48 ; vinsertps $0x30,%xmm6,%xmm7,%xmm6 + DB 196,227,77,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm6,%ymm0 + DB 196,193,97,254,218 ; vpaddd %xmm10,%xmm3,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 - DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 - DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 - DB 197,252,91,92,36,32 ; vcvtdq2ps 0x20(%rsp),%ymm3 - DB 197,60,92,243 ; vsubps %ymm3,%ymm8,%ymm14 - DB 197,252,16,156,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm3 - DB 197,236,92,211 ; vsubps %ymm3,%ymm2,%ymm2 - DB 197,140,89,210 ; vmulps %ymm2,%ymm14,%ymm2 - DB 197,228,88,210 ; vaddps %ymm2,%ymm3,%ymm2 - DB 197,252,17,84,36,32 ; vmovups %ymm2,0x20(%rsp) - DB 196,193,44,92,215 ; vsubps %ymm15,%ymm10,%ymm2 - DB 197,140,89,210 ; vmulps %ymm2,%ymm14,%ymm2 - DB 197,132,88,210 ; vaddps %ymm2,%ymm15,%ymm2 - DB 197,252,17,148,36,192,1,0,0 ; vmovups %ymm2,0x1c0(%rsp) - DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 - DB 197,140,89,192 ; vmulps %ymm0,%ymm14,%ymm0 - DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 - DB 197,252,17,132,36,160,1,0,0 ; vmovups %ymm0,0x1a0(%rsp) - DB 197,244,88,132,36,128,0,0,0 ; vaddps 0x80(%rsp),%ymm1,%ymm0 + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 + DB 196,193,124,91,219 ; vcvtdq2ps %ymm11,%ymm3 + DB 197,20,92,219 ; vsubps %ymm3,%ymm13,%ymm11 + DB 196,193,60,92,222 ; vsubps %ymm14,%ymm8,%ymm3 + DB 197,164,89,219 ; vmulps %ymm3,%ymm11,%ymm3 + DB 197,140,88,219 ; vaddps %ymm3,%ymm14,%ymm3 + DB 197,252,17,156,36,96,1,0,0 ; vmovups %ymm3,0x160(%rsp) + DB 197,252,92,197 ; vsubps %ymm5,%ymm0,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 197,212,88,192 ; vaddps %ymm0,%ymm5,%ymm0 + DB 197,252,17,132,36,160,0,0,0 ; vmovups %ymm0,0xa0(%rsp) + DB 197,244,92,194 ; vsubps %ymm2,%ymm1,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 197,236,88,192 ; vaddps %ymm0,%ymm2,%ymm0 + DB 197,252,17,132,36,128,1,0,0 ; vmovups %ymm0,0x180(%rsp) + DB 197,156,88,68,36,64 ; vaddps 0x40(%rsp),%ymm12,%ymm0 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 196,194,121,64,205 ; vpmulld %xmm13,%xmm0,%xmm1 - DB 197,249,127,140,36,192,0,0,0 ; vmovdqa %xmm1,0xc0(%rsp) + DB 196,226,49,64,200 ; vpmulld %xmm0,%xmm9,%xmm1 + DB 197,249,127,140,36,224,0,0,0 ; vmovdqa %xmm1,0xe0(%rsp) DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 196,194,121,64,213 ; vpmulld %xmm13,%xmm0,%xmm2 - DB 197,249,127,148,36,80,1,0,0 ; vmovdqa %xmm2,0x150(%rsp) - DB 197,241,254,180,36,96,1,0,0 ; vpaddd 0x160(%rsp),%xmm1,%xmm6 - DB 197,249,111,100,36,96 ; vmovdqa 0x60(%rsp),%xmm4 - DB 197,201,254,196 ; vpaddd %xmm4,%xmm6,%xmm0 - DB 196,194,121,64,235 ; vpmulld %xmm11,%xmm0,%xmm5 - DB 196,193,249,126,233 ; vmovq %xmm5,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,235,1 ; vpextrq $0x1,%xmm5,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,227,121,33,201,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm1 - DB 197,233,254,60,36 ; vpaddd (%rsp),%xmm2,%xmm7 - DB 197,121,111,108,36,80 ; vmovdqa 0x50(%rsp),%xmm13 - DB 196,193,65,254,197 ; vpaddd %xmm13,%xmm7,%xmm0 - DB 196,194,121,64,195 ; vpmulld %xmm11,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx + DB 196,226,49,64,192 ; vpmulld %xmm0,%xmm9,%xmm0 + DB 197,249,127,132,36,240,0,0,0 ; vmovdqa %xmm0,0xf0(%rsp) + DB 197,249,254,116,36,96 ; vpaddd 0x60(%rsp),%xmm0,%xmm6 + DB 197,121,111,100,36,48 ; vmovdqa 0x30(%rsp),%xmm12 + DB 196,193,73,254,196 ; vpaddd %xmm12,%xmm6,%xmm0 + DB 196,98,121,64,196 ; vpmulld %xmm4,%xmm0,%xmm8 + DB 197,241,254,140,36,32,1,0,0 ; vpaddd 0x120(%rsp),%xmm1,%xmm1 + DB 197,249,111,108,36,32 ; vmovdqa 0x20(%rsp),%xmm5 + DB 197,241,254,197 ; vpaddd %xmm5,%xmm1,%xmm0 + DB 196,226,121,64,196 ; vpmulld %xmm4,%xmm0,%xmm0 + DB 196,193,121,126,193 ; vmovd %xmm0,%r9d DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 + DB 196,195,121,22,194,2 ; vpextrd $0x2,%xmm0,%r10d + DB 196,195,121,22,195,3 ; vpextrd $0x3,%xmm0,%r11d + DB 196,65,121,126,199 ; vmovd %xmm8,%r15d + DB 196,67,121,22,198,1 ; vpextrd $0x1,%xmm8,%r14d + DB 196,67,121,22,196,2 ; vpextrd $0x2,%xmm8,%r12d + DB 196,99,121,22,195,3 ; vpextrd $0x3,%xmm8,%ebx + DB 196,161,122,16,60,184 ; vmovss (%rax,%r15,4),%xmm7 + DB 196,163,65,33,60,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm7,%xmm7 + DB 196,163,65,33,60,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm7,%xmm7 + DB 196,227,65,33,60,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm7,%xmm7 + DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 - DB 196,163,105,33,20,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 - DB 197,252,17,12,36 ; vmovups %ymm1,(%rsp) - DB 196,193,121,254,201 ; vpaddd %xmm9,%xmm0,%xmm1 - DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 - DB 196,193,249,126,201 ; vmovq %xmm1,%r9 - DB 196,193,81,254,201 ; vpaddd %xmm9,%xmm5,%xmm1 - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 + DB 196,163,105,33,20,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm2,%xmm2 + DB 196,227,109,24,215,1 ; vinsertf128 $0x1,%xmm7,%ymm2,%ymm2 + DB 197,252,17,84,36,96 ; vmovups %ymm2,0x60(%rsp) + DB 197,225,118,219 ; vpcmpeqd %xmm3,%xmm3,%xmm3 + DB 197,249,250,211 ; vpsubd %xmm3,%xmm0,%xmm2 + DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 197,185,250,211 ; vpsubd %xmm3,%xmm8,%xmm2 + DB 196,65,17,118,237 ; vpcmpeqd %xmm13,%xmm13,%xmm13 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 + DB 196,227,105,33,215,32 ; vinsertps $0x20,%xmm7,%xmm2,%xmm2 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 + DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,65,33,219,32 ; vinsertps $0x20,%xmm3,%xmm7,%xmm3 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,99,109,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm10 - DB 196,193,121,254,196 ; vpaddd %xmm12,%xmm0,%xmm0 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 + DB 196,99,101,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm9 + DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,81,254,196 ; vpaddd %xmm12,%xmm5,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,57,254,194 ; vpaddd %xmm10,%xmm8,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm5 - DB 197,121,111,124,36,64 ; vmovdqa 0x40(%rsp),%xmm15 - DB 197,129,254,198 ; vpaddd %xmm6,%xmm15,%xmm0 - DB 196,194,121,64,243 ; vpmulld %xmm11,%xmm0,%xmm6 - DB 196,193,249,126,241 ; vmovq %xmm6,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,243,1 ; vpextrq $0x1,%xmm6,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,227,121,33,210,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm2 - DB 197,193,254,68,36,112 ; vpaddd 0x70(%rsp),%xmm7,%xmm0 - DB 196,194,121,64,195 ; vpmulld %xmm11,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx - DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 - DB 196,163,97,33,28,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,163,97,33,28,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,227,109,24,251,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm7 - DB 196,193,121,254,209 ; vpaddd %xmm9,%xmm0,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,73,254,209 ; vpaddd %xmm9,%xmm6,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,12,136 ; vmovss (%rax,%r9,4),%xmm1 - DB 196,227,97,33,201,32 ; vinsertps $0x20,%xmm1,%xmm3,%xmm1 + DB 196,99,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm8 + DB 197,121,111,60,36 ; vmovdqa (%rsp),%xmm15 + DB 197,129,254,193 ; vpaddd %xmm1,%xmm15,%xmm0 + DB 197,121,111,116,36,16 ; vmovdqa 0x10(%rsp),%xmm14 + DB 197,137,254,206 ; vpaddd %xmm6,%xmm14,%xmm1 + DB 196,226,113,64,244 ; vpmulld %xmm4,%xmm1,%xmm6 + DB 196,226,121,64,204 ; vpmulld %xmm4,%xmm0,%xmm1 + DB 196,193,121,126,201 ; vmovd %xmm1,%r9d + DB 196,195,121,22,200,1 ; vpextrd $0x1,%xmm1,%r8d + DB 196,195,121,22,202,2 ; vpextrd $0x2,%xmm1,%r10d + DB 196,195,121,22,203,3 ; vpextrd $0x3,%xmm1,%r11d + DB 196,193,121,126,247 ; vmovd %xmm6,%r15d + DB 196,195,121,22,246,1 ; vpextrd $0x1,%xmm6,%r14d + DB 196,195,121,22,244,2 ; vpextrd $0x2,%xmm6,%r12d + DB 196,227,121,22,243,3 ; vpextrd $0x3,%xmm6,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 + DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 + DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 + DB 196,163,105,33,20,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm2,%xmm2 + DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 + DB 196,193,113,250,213 ; vpsubd %xmm13,%xmm1,%xmm2 + DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 196,193,73,250,213 ; vpsubd %xmm13,%xmm6,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 - DB 196,99,117,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm8 - DB 196,193,121,254,196 ; vpaddd %xmm12,%xmm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,73,254,196 ; vpaddd %xmm12,%xmm6,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,12,176 ; vmovss (%rax,%r14,4),%xmm1 - DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,97,33,223,32 ; vinsertps $0x20,%xmm7,%xmm3,%xmm3 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,163,113,33,12,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 + DB 196,99,101,24,234,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm13 + DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 + DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 196,193,73,254,202 ; vpaddd %xmm10,%xmm6,%xmm1 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 197,252,16,20,36 ; vmovups (%rsp),%ymm2 - DB 197,196,92,202 ; vsubps %ymm2,%ymm7,%ymm1 - DB 197,140,89,201 ; vmulps %ymm1,%ymm14,%ymm1 - DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 - DB 196,193,60,92,210 ; vsubps %ymm10,%ymm8,%ymm2 - DB 197,140,89,210 ; vmulps %ymm2,%ymm14,%ymm2 - DB 197,172,88,210 ; vaddps %ymm2,%ymm10,%ymm2 - DB 197,252,92,197 ; vsubps %ymm5,%ymm0,%ymm0 - DB 197,140,89,192 ; vmulps %ymm0,%ymm14,%ymm0 - DB 197,212,88,192 ; vaddps %ymm0,%ymm5,%ymm0 - DB 197,252,91,156,36,224,0,0,0 ; vcvtdq2ps 0xe0(%rsp),%ymm3 - DB 197,252,16,172,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm5 - DB 197,212,92,219 ; vsubps %ymm3,%ymm5,%ymm3 - DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) - DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 - DB 197,244,92,205 ; vsubps %ymm5,%ymm1,%ymm1 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,212,88,201 ; vaddps %ymm1,%ymm5,%ymm1 - DB 197,252,17,140,36,128,0,0,0 ; vmovups %ymm1,0x80(%rsp) - DB 197,252,16,172,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm5 - DB 197,236,92,205 ; vsubps %ymm5,%ymm2,%ymm1 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,212,88,201 ; vaddps %ymm1,%ymm5,%ymm1 - DB 197,252,17,140,36,224,0,0,0 ; vmovups %ymm1,0xe0(%rsp) - DB 197,252,16,140,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm1 - DB 197,252,92,193 ; vsubps %ymm1,%ymm0,%ymm0 - DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,244,88,192 ; vaddps %ymm0,%ymm1,%ymm0 - DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp) - DB 197,252,16,132,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm0 - DB 197,252,88,132,36,128,1,0,0 ; vaddps 0x180(%rsp),%ymm0,%ymm0 - DB 197,126,91,192 ; vcvttps2dq %ymm0,%ymm8 - DB 197,185,254,188,36,32,1,0,0 ; vpaddd 0x120(%rsp),%xmm8,%xmm7 - DB 197,217,254,199 ; vpaddd %xmm7,%xmm4,%xmm0 - DB 196,194,121,64,243 ; vpmulld %xmm11,%xmm0,%xmm6 - DB 196,193,249,126,241 ; vmovq %xmm6,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,243,1 ; vpextrq $0x1,%xmm6,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,227,121,33,217,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm3 - DB 196,99,125,25,192,1 ; vextractf128 $0x1,%ymm8,%xmm0 - DB 197,248,41,132,36,160,0,0,0 ; vmovaps %xmm0,0xa0(%rsp) - DB 197,249,254,140,36,0,1,0,0 ; vpaddd 0x100(%rsp),%xmm0,%xmm1 - DB 197,145,254,193 ; vpaddd %xmm1,%xmm13,%xmm0 - DB 196,194,121,64,195 ; vpmulld %xmm11,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx - DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 - DB 196,163,105,33,20,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,227,101,24,226,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm4 - DB 196,193,121,254,209 ; vpaddd %xmm9,%xmm0,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,73,254,209 ; vpaddd %xmm9,%xmm6,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 - DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 - DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 - DB 196,99,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm10 - DB 196,193,121,254,196 ; vpaddd %xmm12,%xmm0,%xmm0 + DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 + DB 197,252,16,84,36,96 ; vmovups 0x60(%rsp),%ymm2 + DB 197,252,92,194 ; vsubps %ymm2,%ymm0,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 197,236,88,208 ; vaddps %ymm0,%ymm2,%ymm2 + DB 196,193,20,92,193 ; vsubps %ymm9,%ymm13,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 197,180,88,216 ; vaddps %ymm0,%ymm9,%ymm3 + DB 196,193,116,92,192 ; vsubps %ymm8,%ymm1,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 197,188,88,200 ; vaddps %ymm0,%ymm8,%ymm1 + DB 197,252,91,132,36,192,0,0,0 ; vcvtdq2ps 0xc0(%rsp),%ymm0 + DB 197,252,16,116,36,64 ; vmovups 0x40(%rsp),%ymm6 + DB 197,76,92,232 ; vsubps %ymm0,%ymm6,%ymm13 + DB 197,252,16,132,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm0 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 197,148,89,210 ; vmulps %ymm2,%ymm13,%ymm2 + DB 197,252,88,194 ; vaddps %ymm2,%ymm0,%ymm0 + DB 197,252,17,68,36,64 ; vmovups %ymm0,0x40(%rsp) + DB 197,252,16,132,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm0 + DB 197,228,92,208 ; vsubps %ymm0,%ymm3,%ymm2 + DB 197,148,89,210 ; vmulps %ymm2,%ymm13,%ymm2 + DB 197,252,88,194 ; vaddps %ymm2,%ymm0,%ymm0 + DB 197,252,17,132,36,192,0,0,0 ; vmovups %ymm0,0xc0(%rsp) + DB 197,252,16,132,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm0 + DB 197,244,92,200 ; vsubps %ymm0,%ymm1,%ymm1 + DB 197,148,89,201 ; vmulps %ymm1,%ymm13,%ymm1 + DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 + DB 197,252,17,132,36,160,0,0,0 ; vmovups %ymm0,0xa0(%rsp) + DB 197,252,16,132,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm0 + DB 197,252,88,140,36,64,1,0,0 ; vaddps 0x140(%rsp),%ymm0,%ymm1 + DB 197,126,91,201 ; vcvttps2dq %ymm1,%ymm9 + DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 + DB 197,249,127,132,36,128,0,0,0 ; vmovdqa %xmm0,0x80(%rsp) + DB 197,249,254,180,36,16,1,0,0 ; vpaddd 0x110(%rsp),%xmm0,%xmm6 + DB 197,177,254,156,36,0,1,0,0 ; vpaddd 0x100(%rsp),%xmm9,%xmm3 + DB 197,153,254,206 ; vpaddd %xmm6,%xmm12,%xmm1 + DB 196,226,113,64,204 ; vpmulld %xmm4,%xmm1,%xmm1 + DB 197,209,254,211 ; vpaddd %xmm3,%xmm5,%xmm2 + DB 196,226,105,64,212 ; vpmulld %xmm4,%xmm2,%xmm2 + DB 196,193,121,126,209 ; vmovd %xmm2,%r9d + DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d + DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d + DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,60,184 ; vmovss (%rax,%r15,4),%xmm7 + DB 196,163,65,33,60,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm7,%xmm7 + DB 196,163,65,33,60,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm7,%xmm7 + DB 196,227,65,33,60,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm7,%xmm7 + DB 196,161,122,16,4,136 ; vmovss (%rax,%r9,4),%xmm0 + DB 196,163,121,33,4,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm0,%xmm0 + DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 + DB 196,163,121,33,4,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm0,%xmm0 + DB 196,99,125,24,231,1 ; vinsertf128 $0x1,%xmm7,%ymm0,%ymm12 + DB 197,209,118,237 ; vpcmpeqd %xmm5,%xmm5,%xmm5 + DB 197,233,250,197 ; vpsubd %xmm5,%xmm2,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,73,254,196 ; vpaddd %xmm12,%xmm6,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 197,241,250,197 ; vpsubd %xmm5,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 + DB 196,227,121,33,199,32 ; vinsertps $0x20,%xmm7,%xmm0,%xmm0 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,121,33,199,48 ; vinsertps $0x30,%xmm7,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 + DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,65,33,237,32 ; vinsertps $0x20,%xmm5,%xmm7,%xmm5 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 + DB 196,227,85,24,248,1 ; vinsertf128 $0x1,%xmm0,%ymm5,%ymm7 + DB 196,193,105,254,194 ; vpaddd %xmm10,%xmm2,%xmm0 + DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 196,193,113,254,194 ; vpaddd %xmm10,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,12,176 ; vmovss (%rax,%r14,4),%xmm1 + DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,12,144 ; vmovss (%rax,%r10,4),%xmm1 + DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 + DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 + DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 + DB 196,99,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm8 + DB 197,129,254,195 ; vpaddd %xmm3,%xmm15,%xmm0 + DB 197,137,254,206 ; vpaddd %xmm6,%xmm14,%xmm1 + DB 196,226,113,64,204 ; vpmulld %xmm4,%xmm1,%xmm1 + DB 196,226,121,64,212 ; vpmulld %xmm4,%xmm0,%xmm2 + DB 196,193,121,126,209 ; vmovd %xmm2,%r9d + DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d + DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d + DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,99,109,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm13 - DB 197,129,254,199 ; vpaddd %xmm7,%xmm15,%xmm0 - DB 196,194,121,64,243 ; vpmulld %xmm11,%xmm0,%xmm6 - DB 196,193,249,126,241 ; vmovq %xmm6,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,243,1 ; vpextrq $0x1,%xmm6,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm3 + DB 197,209,118,237 ; vpcmpeqd %xmm5,%xmm5,%xmm5 + DB 197,233,250,197 ; vpsubd %xmm5,%xmm2,%xmm0 + DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 197,241,250,197 ; vpsubd %xmm5,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,44,176 ; vmovss (%rax,%r14,4),%xmm5 + DB 196,227,121,33,197,32 ; vinsertps $0x20,%xmm5,%xmm0,%xmm0 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,121,33,197,48 ; vinsertps $0x30,%xmm5,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 197,121,111,124,36,112 ; vmovdqa 0x70(%rsp),%xmm15 - DB 197,129,254,201 ; vpaddd %xmm1,%xmm15,%xmm1 - DB 196,194,113,64,251 ; vpmulld %xmm11,%xmm1,%xmm7 - DB 197,249,126,251 ; vmovd %xmm7,%ebx - DB 196,195,121,22,248,1 ; vpextrd $0x1,%xmm7,%r8d - DB 196,195,121,22,249,2 ; vpextrd $0x2,%xmm7,%r9d - DB 196,195,121,22,250,3 ; vpextrd $0x3,%xmm7,%r10d - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,125,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm1 - DB 196,193,65,254,193 ; vpaddd %xmm9,%xmm7,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 + DB 196,163,81,33,44,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm5,%xmm5 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,81,33,238,32 ; vinsertps $0x20,%xmm6,%xmm5,%xmm5 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,81,33,238,48 ; vinsertps $0x30,%xmm6,%xmm5,%xmm5 + DB 196,227,85,24,240,1 ; vinsertf128 $0x1,%xmm0,%ymm5,%ymm6 + DB 196,193,105,254,194 ; vpaddd %xmm10,%xmm2,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,73,254,193 ; vpaddd %xmm9,%xmm6,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,113,254,194 ; vpaddd %xmm10,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,12,176 ; vmovss (%rax,%r14,4),%xmm1 + DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,12,144 ; vmovss (%rax,%r10,4),%xmm1 + DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 + DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 196,193,100,92,204 ; vsubps %ymm12,%ymm3,%ymm1 + DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 + DB 197,28,88,225 ; vaddps %ymm1,%ymm12,%ymm12 + DB 197,204,92,207 ; vsubps %ymm7,%ymm6,%ymm1 + DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 + DB 197,68,88,241 ; vaddps %ymm1,%ymm7,%ymm14 + DB 196,193,124,92,192 ; vsubps %ymm8,%ymm0,%ymm0 + DB 197,164,89,192 ; vmulps %ymm0,%ymm11,%ymm0 + DB 197,60,88,192 ; vaddps %ymm0,%ymm8,%ymm8 + DB 197,249,111,132,36,128,0,0,0 ; vmovdqa 0x80(%rsp),%xmm0 + DB 197,249,254,180,36,240,0,0,0 ; vpaddd 0xf0(%rsp),%xmm0,%xmm6 + DB 197,177,254,140,36,224,0,0,0 ; vpaddd 0xe0(%rsp),%xmm9,%xmm1 + DB 197,241,254,68,36,32 ; vpaddd 0x20(%rsp),%xmm1,%xmm0 + DB 197,201,254,84,36,48 ; vpaddd 0x30(%rsp),%xmm6,%xmm2 + DB 196,98,105,64,204 ; vpmulld %xmm4,%xmm2,%xmm9 + DB 196,226,121,64,220 ; vpmulld %xmm4,%xmm0,%xmm3 + DB 196,193,121,126,217 ; vmovd %xmm3,%r9d + DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d + DB 196,195,121,22,218,2 ; vpextrd $0x2,%xmm3,%r10d + DB 196,195,121,22,219,3 ; vpextrd $0x3,%xmm3,%r11d + DB 196,65,121,126,207 ; vmovd %xmm9,%r15d + DB 196,67,121,22,206,1 ; vpextrd $0x1,%xmm9,%r14d + DB 196,67,121,22,204,2 ; vpextrd $0x2,%xmm9,%r12d + DB 196,99,121,22,203,3 ; vpextrd $0x3,%xmm9,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 + DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 + DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 + DB 196,163,105,33,20,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm2,%xmm2 DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 - DB 196,193,65,254,212 ; vpaddd %xmm12,%xmm7,%xmm2 + DB 197,209,118,237 ; vpcmpeqd %xmm5,%xmm5,%xmm5 + DB 197,225,250,213 ; vpsubd %xmm5,%xmm3,%xmm2 DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,73,254,212 ; vpaddd %xmm12,%xmm6,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 197,177,250,213 ; vpsubd %xmm5,%xmm9,%xmm2 + DB 196,65,1,118,255 ; vpcmpeqd %xmm15,%xmm15,%xmm15 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,44,176 ; vmovss (%rax,%r14,4),%xmm5 + DB 196,227,105,33,213,32 ; vinsertps $0x20,%xmm5,%xmm2,%xmm2 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 - DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 - DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 - DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 - DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 - DB 197,140,89,201 ; vmulps %ymm1,%ymm14,%ymm1 - DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,17,140,36,32,1,0,0 ; vmovups %ymm1,0x120(%rsp) - DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 197,140,89,192 ; vmulps %ymm0,%ymm14,%ymm0 - DB 197,172,88,192 ; vaddps %ymm0,%ymm10,%ymm0 - DB 197,252,17,132,36,0,1,0,0 ; vmovups %ymm0,0x100(%rsp) - DB 196,193,108,92,197 ; vsubps %ymm13,%ymm2,%ymm0 - DB 197,140,89,192 ; vmulps %ymm0,%ymm14,%ymm0 - DB 197,20,88,208 ; vaddps %ymm0,%ymm13,%ymm10 - DB 197,249,111,132,36,160,0,0,0 ; vmovdqa 0xa0(%rsp),%xmm0 - DB 197,249,254,180,36,80,1,0,0 ; vpaddd 0x150(%rsp),%xmm0,%xmm6 - DB 197,185,254,132,36,192,0,0,0 ; vpaddd 0xc0(%rsp),%xmm8,%xmm0 - DB 197,249,254,76,36,96 ; vpaddd 0x60(%rsp),%xmm0,%xmm1 - DB 196,66,113,64,235 ; vpmulld %xmm11,%xmm1,%xmm13 - DB 196,65,249,126,233 ; vmovq %xmm13,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,99,249,22,235,1 ; vpextrq $0x1,%xmm13,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,105,33,213,48 ; vinsertps $0x30,%xmm5,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm1,%xmm1 - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 197,201,254,84,36,80 ; vpaddd 0x50(%rsp),%xmm6,%xmm2 - DB 196,194,105,64,211 ; vpmulld %xmm11,%xmm2,%xmm2 - DB 197,249,126,211 ; vmovd %xmm2,%ebx - DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d - DB 196,195,121,22,209,2 ; vpextrd $0x2,%xmm2,%r9d - DB 196,195,121,22,210,3 ; vpextrd $0x3,%xmm2,%r10d - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 - DB 196,163,97,33,28,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,163,97,33,28,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,227,117,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm1,%ymm1 - DB 196,193,105,254,217 ; vpaddd %xmm9,%xmm2,%xmm3 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 + DB 196,163,81,33,44,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm5,%xmm5 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,81,33,239,32 ; vinsertps $0x20,%xmm7,%xmm5,%xmm5 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 + DB 196,227,85,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm5,%ymm2 + DB 196,193,97,254,218 ; vpaddd %xmm10,%xmm3,%xmm3 DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,17,254,217 ; vpaddd %xmm9,%xmm13,%xmm3 - DB 196,193,249,126,218 ; vmovq %xmm3,%r10 - DB 196,227,249,22,219,1 ; vpextrq $0x1,%xmm3,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,49,254,218 ; vpaddd %xmm10,%xmm9,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 - DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 DB 196,161,122,16,44,176 ; vmovss (%rax,%r14,4),%xmm5 DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 - DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 - DB 196,163,81,33,44,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm5,%xmm5 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 - DB 196,227,81,33,239,32 ; vinsertps $0x20,%xmm7,%xmm5,%xmm5 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 - DB 196,227,85,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm5,%ymm5 - DB 196,193,105,254,212 ; vpaddd %xmm12,%xmm2,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,17,254,212 ; vpaddd %xmm12,%xmm13,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 + DB 196,163,81,33,44,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm5,%xmm5 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,81,33,239,32 ; vinsertps $0x20,%xmm7,%xmm5,%xmm5 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 - DB 196,227,97,33,223,32 ; vinsertps $0x20,%xmm7,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 - DB 196,99,101,24,234,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm13 - DB 197,249,254,68,36,64 ; vpaddd 0x40(%rsp),%xmm0,%xmm0 - DB 197,129,254,214 ; vpaddd %xmm6,%xmm15,%xmm2 - DB 196,194,105,64,243 ; vpmulld %xmm11,%xmm2,%xmm6 - DB 196,194,121,64,195 ; vpmulld %xmm11,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 + DB 196,99,85,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm5,%ymm9 + DB 197,241,254,12,36 ; vpaddd (%rsp),%xmm1,%xmm1 + DB 197,201,254,92,36,16 ; vpaddd 0x10(%rsp),%xmm6,%xmm3 + DB 196,226,97,64,220 ; vpmulld %xmm4,%xmm3,%xmm3 + DB 196,226,113,64,228 ; vpmulld %xmm4,%xmm1,%xmm4 + DB 196,193,121,126,225 ; vmovd %xmm4,%r9d + DB 196,195,121,22,224,1 ; vpextrd $0x1,%xmm4,%r8d + DB 196,195,121,22,226,2 ; vpextrd $0x2,%xmm4,%r10d + DB 196,195,121,22,227,3 ; vpextrd $0x3,%xmm4,%r11d + DB 196,193,121,126,223 ; vmovd %xmm3,%r15d + DB 196,195,121,22,222,1 ; vpextrd $0x1,%xmm3,%r14d + DB 196,195,121,22,220,2 ; vpextrd $0x2,%xmm3,%r12d + DB 196,227,121,22,219,3 ; vpextrd $0x3,%xmm3,%ebx + DB 196,161,122,16,12,184 ; vmovss (%rax,%r15,4),%xmm1 + DB 196,163,113,33,12,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm1,%xmm1 + DB 196,163,113,33,12,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm1,%xmm1 + DB 196,227,113,33,12,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 + DB 196,163,81,33,44,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm5,%xmm5 + DB 196,163,81,33,44,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm5,%xmm5 + DB 196,163,81,33,44,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm5,%xmm5 + DB 196,227,85,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm5,%ymm1 + DB 196,193,89,250,239 ; vpsubd %xmm15,%xmm4,%xmm5 + DB 196,195,249,22,232,1 ; vpextrq $0x1,%xmm5,%r8 + DB 196,193,249,126,233 ; vmovq %xmm5,%r9 + DB 196,193,97,250,239 ; vpsubd %xmm15,%xmm3,%xmm5 + DB 196,195,249,22,234,1 ; vpextrq $0x1,%xmm5,%r10 + DB 196,225,249,126,235 ; vmovq %xmm5,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,44,152 ; vmovss (%rax,%r11,4),%xmm5 + DB 196,227,81,33,44,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm5,%xmm5 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,81,33,238,32 ; vinsertps $0x20,%xmm6,%xmm5,%xmm5 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,81,33,238,48 ; vinsertps $0x30,%xmm6,%xmm5,%xmm5 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 197,249,126,243 ; vmovd %xmm6,%ebx - DB 196,195,121,22,240,1 ; vpextrd $0x1,%xmm6,%r8d - DB 196,195,121,22,241,2 ; vpextrd $0x2,%xmm6,%r9d - DB 196,195,121,22,242,3 ; vpextrd $0x3,%xmm6,%r10d - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 - DB 196,163,97,33,28,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,163,97,33,28,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,227,109,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm3 - DB 196,193,73,254,209 ; vpaddd %xmm9,%xmm6,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,121,254,209 ; vpaddd %xmm9,%xmm0,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,105,33,212,32 ; vinsertps $0x20,%xmm4,%xmm2,%xmm2 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,105,33,212,48 ; vinsertps $0x30,%xmm4,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 - DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 - DB 196,193,73,254,228 ; vpaddd %xmm12,%xmm6,%xmm4 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 + DB 196,227,77,24,245,1 ; vinsertf128 $0x1,%xmm5,%ymm6,%ymm6 + DB 196,193,89,254,226 ; vpaddd %xmm10,%xmm4,%xmm4 DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,121,254,196 ; vpaddd %xmm12,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 + DB 196,193,97,254,218 ; vpaddd %xmm10,%xmm3,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,121,33,196,32 ; vinsertps $0x20,%xmm4,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,121,33,196,48 ; vinsertps $0x30,%xmm4,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 196,227,93,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm4,%ymm0 - DB 197,228,92,217 ; vsubps %ymm1,%ymm3,%ymm3 - DB 197,140,89,219 ; vmulps %ymm3,%ymm14,%ymm3 - DB 197,244,88,203 ; vaddps %ymm3,%ymm1,%ymm1 - DB 197,236,92,213 ; vsubps %ymm5,%ymm2,%ymm2 - DB 197,140,89,210 ; vmulps %ymm2,%ymm14,%ymm2 - DB 197,212,88,210 ; vaddps %ymm2,%ymm5,%ymm2 - DB 196,193,124,92,197 ; vsubps %ymm13,%ymm0,%ymm0 - DB 197,140,89,192 ; vmulps %ymm0,%ymm14,%ymm0 - DB 197,148,88,192 ; vaddps %ymm0,%ymm13,%ymm0 - DB 197,252,16,164,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm4 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 + DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,89,33,229,32 ; vinsertps $0x20,%xmm5,%xmm4,%xmm4 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 + DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 + DB 196,227,93,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm3 + DB 197,244,92,200 ; vsubps %ymm0,%ymm1,%ymm1 + DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 + DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 + DB 197,204,92,202 ; vsubps %ymm2,%ymm6,%ymm1 + DB 197,164,89,201 ; vmulps %ymm1,%ymm11,%ymm1 + DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 + DB 196,193,100,92,209 ; vsubps %ymm9,%ymm3,%ymm2 + DB 197,164,89,210 ; vmulps %ymm2,%ymm11,%ymm2 + DB 197,180,88,210 ; vaddps %ymm2,%ymm9,%ymm2 + DB 196,193,124,92,196 ; vsubps %ymm12,%ymm0,%ymm0 + DB 197,148,89,192 ; vmulps %ymm0,%ymm13,%ymm0 + DB 197,156,88,192 ; vaddps %ymm0,%ymm12,%ymm0 + DB 196,193,116,92,206 ; vsubps %ymm14,%ymm1,%ymm1 + DB 197,148,89,201 ; vmulps %ymm1,%ymm13,%ymm1 + DB 197,140,88,201 ; vaddps %ymm1,%ymm14,%ymm1 + DB 196,193,108,92,208 ; vsubps %ymm8,%ymm2,%ymm2 + DB 197,148,89,210 ; vmulps %ymm2,%ymm13,%ymm2 + DB 197,188,88,210 ; vaddps %ymm2,%ymm8,%ymm2 + DB 197,252,91,156,36,32,1,0,0 ; vcvtdq2ps 0x120(%rsp),%ymm3 + DB 197,252,16,164,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm4 + DB 197,220,92,219 ; vsubps %ymm3,%ymm4,%ymm3 + DB 197,252,16,100,36,64 ; vmovups 0x40(%rsp),%ymm4 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,16,164,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm4 DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 - DB 197,252,16,28,36 ; vmovups (%rsp),%ymm3 DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,16,164,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm4 + DB 197,252,16,164,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm4 DB 197,236,92,212 ; vsubps %ymm4,%ymm2,%ymm2 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 DB 197,220,88,210 ; vaddps %ymm2,%ymm4,%ymm2 - DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0 - DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,172,88,216 ; vaddps %ymm0,%ymm10,%ymm3 - DB 197,252,91,132,36,96,1,0,0 ; vcvtdq2ps 0x160(%rsp),%ymm0 - DB 197,252,16,164,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm4 - DB 197,220,92,224 ; vsubps %ymm0,%ymm4,%ymm4 - DB 197,252,16,172,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm5 - DB 197,244,92,197 ; vsubps %ymm5,%ymm1,%ymm0 - DB 197,220,89,192 ; vmulps %ymm0,%ymm4,%ymm0 - DB 197,212,88,192 ; vaddps %ymm0,%ymm5,%ymm0 - DB 197,252,16,172,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm5 - DB 197,236,92,205 ; vsubps %ymm5,%ymm2,%ymm1 - DB 197,220,89,201 ; vmulps %ymm1,%ymm4,%ymm1 - DB 197,212,88,201 ; vaddps %ymm1,%ymm5,%ymm1 - DB 197,252,16,108,36,32 ; vmovups 0x20(%rsp),%ymm5 - DB 197,228,92,213 ; vsubps %ymm5,%ymm3,%ymm2 - DB 197,220,89,210 ; vmulps %ymm2,%ymm4,%ymm2 - DB 197,212,88,210 ; vaddps %ymm2,%ymm5,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,156,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm3 - DB 197,252,16,164,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm4 - DB 197,252,16,172,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm5 - DB 197,252,16,180,36,64,2,0,0 ; vmovups 0x240(%rsp),%ymm6 - DB 197,252,16,188,36,96,2,0,0 ; vmovups 0x260(%rsp),%ymm7 - DB 72,129,196,136,2,0,0 ; add $0x288,%rsp + DB 197,252,16,156,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm3 + DB 197,252,16,164,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm4 + DB 197,252,16,172,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm5 + DB 197,252,16,180,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm6 + DB 197,252,16,188,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm7 + DB 72,129,196,88,2,0,0 ; add $0x258,%rsp DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax PUBLIC _sk_clut_4D_avx _sk_clut_4D_avx LABEL PROC + DB 65,87 ; push %r15 DB 65,86 ; push %r14 + DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,40,3,0,0 ; sub $0x328,%rsp - DB 197,252,17,188,36,0,3,0,0 ; vmovups %ymm7,0x300(%rsp) - DB 197,252,17,180,36,224,2,0,0 ; vmovups %ymm6,0x2e0(%rsp) - DB 197,252,17,172,36,192,2,0,0 ; vmovups %ymm5,0x2c0(%rsp) - DB 197,252,17,164,36,160,2,0,0 ; vmovups %ymm4,0x2a0(%rsp) + DB 72,129,236,88,3,0,0 ; sub $0x358,%rsp + DB 197,252,17,188,36,32,3,0,0 ; vmovups %ymm7,0x320(%rsp) + DB 197,252,17,180,36,0,3,0,0 ; vmovups %ymm6,0x300(%rsp) + DB 197,252,17,172,36,224,2,0,0 ; vmovups %ymm5,0x2e0(%rsp) + DB 197,252,17,164,36,192,2,0,0 ; vmovups %ymm4,0x2c0(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,20 ; mov 0x14(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -15402,9 +15571,8 @@ _sk_clut_4D_avx LABEL PROC DB 197,249,112,228,0 ; vpshufd $0x0,%xmm4,%xmm4 DB 196,227,93,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm4,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 197,220,89,219 ; vmulps %ymm3,%ymm4,%ymm3 - DB 197,252,17,156,36,128,2,0,0 ; vmovups %ymm3,0x280(%rsp) - DB 197,254,91,227 ; vcvttps2dq %ymm3,%ymm4 + DB 197,92,89,203 ; vmulps %ymm3,%ymm4,%ymm9 + DB 197,124,17,140,36,96,2,0,0 ; vmovups %ymm9,0x260(%rsp) DB 68,139,72,16 ; mov 0x10(%rax),%r9d DB 69,141,81,255 ; lea -0x1(%r9),%r10d DB 196,193,121,110,218 ; vmovd %r10d,%xmm3 @@ -15412,20 +15580,19 @@ _sk_clut_4D_avx LABEL PROC DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 - DB 197,252,17,148,36,160,1,0,0 ; vmovups %ymm2,0x1a0(%rsp) - DB 197,254,91,234 ; vcvttps2dq %ymm2,%ymm5 - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 197,249,112,242,0 ; vpshufd $0x0,%xmm2,%xmm6 - DB 196,226,81,64,222 ; vpmulld %xmm6,%xmm5,%xmm3 - DB 197,249,111,254 ; vmovdqa %xmm6,%xmm7 - DB 197,249,127,124,36,96 ; vmovdqa %xmm7,0x60(%rsp) - DB 197,254,127,172,36,0,2,0,0 ; vmovdqu %ymm5,0x200(%rsp) - DB 197,249,127,156,36,64,1,0,0 ; vmovdqa %xmm3,0x140(%rsp) - DB 197,225,254,244 ; vpaddd %xmm4,%xmm3,%xmm6 - DB 197,249,127,180,36,224,0,0,0 ; vmovdqa %xmm6,0xe0(%rsp) - DB 197,254,127,164,36,96,2,0,0 ; vmovdqu %ymm4,0x260(%rsp) - DB 196,193,121,110,217 ; vmovd %r9d,%xmm3 - DB 196,226,97,64,210 ; vpmulld %xmm2,%xmm3,%xmm2 + DB 197,252,17,148,36,128,1,0,0 ; vmovups %ymm2,0x180(%rsp) + DB 197,254,91,218 ; vcvttps2dq %ymm2,%ymm3 + DB 196,227,125,25,218,1 ; vextractf128 $0x1,%ymm3,%xmm2 + DB 197,253,111,251 ; vmovdqa %ymm3,%ymm7 + DB 197,254,127,188,36,192,1,0,0 ; vmovdqu %ymm7,0x1c0(%rsp) + DB 196,193,121,110,216 ; vmovd %r8d,%xmm3 + DB 197,249,112,227,0 ; vpshufd $0x0,%xmm3,%xmm4 + DB 196,226,89,64,242 ; vpmulld %xmm2,%xmm4,%xmm6 + DB 197,249,111,236 ; vmovdqa %xmm4,%xmm5 + DB 197,249,127,108,36,64 ; vmovdqa %xmm5,0x40(%rsp) + DB 197,249,127,180,36,32,1,0,0 ; vmovdqa %xmm6,0x120(%rsp) + DB 196,193,121,110,209 ; vmovd %r9d,%xmm2 + DB 196,226,105,64,211 ; vpmulld %xmm3,%xmm2,%xmm2 DB 197,121,112,194,0 ; vpshufd $0x0,%xmm2,%xmm8 DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -15433,1561 +15600,1530 @@ _sk_clut_4D_avx LABEL PROC DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3 DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 197,228,89,217 ; vmulps %ymm1,%ymm3,%ymm3 - DB 197,252,17,156,36,0,1,0,0 ; vmovups %ymm3,0x100(%rsp) + DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 + DB 197,252,17,140,36,0,1,0,0 ; vmovups %ymm1,0x100(%rsp) + DB 197,254,91,217 ; vcvttps2dq %ymm1,%ymm3 + DB 196,227,125,25,217,1 ; vextractf128 $0x1,%ymm3,%xmm1 + DB 197,125,111,211 ; vmovdqa %ymm3,%ymm10 + DB 197,126,127,148,36,64,1,0,0 ; vmovdqu %ymm10,0x140(%rsp) + DB 196,226,57,64,217 ; vpmulld %xmm1,%xmm8,%xmm3 + DB 197,249,127,156,36,96,1,0,0 ; vmovdqa %xmm3,0x160(%rsp) DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 DB 196,226,113,64,202 ; vpmulld %xmm2,%xmm1,%xmm1 + DB 197,249,112,225,0 ; vpshufd $0x0,%xmm1,%xmm4 DB 68,139,64,8 ; mov 0x8(%rax),%r8d DB 65,255,200 ; dec %r8d - DB 196,193,121,110,208 ; vmovd %r8d,%xmm2 - DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2 - DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 - DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 197,108,89,232 ; vmulps %ymm0,%ymm2,%ymm13 - DB 197,254,91,211 ; vcvttps2dq %ymm3,%ymm2 - DB 197,249,112,193,0 ; vpshufd $0x0,%xmm1,%xmm0 + DB 196,193,121,110,200 ; vmovd %r8d,%xmm1 + DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 + DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 + DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 + DB 197,116,89,248 ; vmulps %ymm0,%ymm1,%ymm15 + DB 196,65,126,91,223 ; vcvttps2dq %ymm15,%ymm11 + DB 196,99,125,25,216,1 ; vextractf128 $0x1,%ymm11,%xmm0 + DB 196,226,89,64,208 ; vpmulld %xmm0,%xmm4,%xmm2 + DB 197,249,127,20,36 ; vmovdqa %xmm2,(%rsp) DB 72,139,0 ; mov (%rax),%rax - DB 196,194,105,64,200 ; vpmulld %xmm8,%xmm2,%xmm1 - DB 197,125,111,210 ; vmovdqa %ymm2,%ymm10 - DB 197,126,127,148,36,96,1,0,0 ; vmovdqu %ymm10,0x160(%rsp) - DB 197,249,127,140,36,128,1,0,0 ; vmovdqa %xmm1,0x180(%rsp) - DB 197,241,254,214 ; vpaddd %xmm6,%xmm1,%xmm2 - DB 196,65,126,91,229 ; vcvttps2dq %ymm13,%ymm12 - DB 196,226,25,64,200 ; vpmulld %xmm0,%xmm12,%xmm1 - DB 197,249,127,12,36 ; vmovdqa %xmm1,(%rsp) - DB 197,241,254,202 ; vpaddd %xmm2,%xmm1,%xmm1 - DB 196,98,121,24,61,239,202,1,0 ; vbroadcastss 0x1caef(%rip),%xmm15 # 2d424 <_sk_overlay_sse2_8bit+0xecc> - DB 196,194,113,64,247 ; vpmulld %xmm15,%xmm1,%xmm6 - DB 196,193,249,126,240 ; vmovq %xmm6,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 196,195,249,22,242,1 ; vpextrq $0x1,%xmm6,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,12,136 ; vmovss (%rax,%r9,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,152,32 ; vinsertps $0x20,(%rax,%r11,4),%xmm1,%xmm1 - DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 - DB 196,99,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm9 - DB 196,227,125,25,233,1 ; vextractf128 $0x1,%ymm5,%xmm1 - DB 196,226,113,64,239 ; vpmulld %xmm7,%xmm1,%xmm5 - DB 197,249,127,172,36,32,1,0,0 ; vmovdqa %xmm5,0x120(%rsp) - DB 196,99,125,25,209,1 ; vextractf128 $0x1,%ymm10,%xmm1 - DB 196,194,113,64,216 ; vpmulld %xmm8,%xmm1,%xmm3 - DB 197,249,127,156,36,144,0,0,0 ; vmovdqa %xmm3,0x90(%rsp) - DB 196,99,125,25,225,1 ; vextractf128 $0x1,%ymm12,%xmm1 - DB 196,226,113,64,248 ; vpmulld %xmm0,%xmm1,%xmm7 - DB 197,249,127,124,36,48 ; vmovdqa %xmm7,0x30(%rsp) - DB 196,227,125,25,225,1 ; vextractf128 $0x1,%ymm4,%xmm1 - DB 197,248,41,76,36,64 ; vmovaps %xmm1,0x40(%rsp) - DB 197,209,254,201 ; vpaddd %xmm1,%xmm5,%xmm1 - DB 197,249,127,140,36,192,1,0,0 ; vmovdqa %xmm1,0x1c0(%rsp) - DB 197,225,254,201 ; vpaddd %xmm1,%xmm3,%xmm1 - DB 197,193,254,217 ; vpaddd %xmm1,%xmm7,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 196,193,121,126,216 ; vmovd %xmm3,%r8d - DB 196,195,121,22,217,1 ; vpextrd $0x1,%xmm3,%r9d - DB 196,195,121,22,218,2 ; vpextrd $0x2,%xmm3,%r10d - DB 196,195,121,22,219,3 ; vpextrd $0x3,%xmm3,%r11d + DB 196,193,126,91,201 ; vcvttps2dq %ymm9,%ymm1 + DB 197,254,127,140,36,64,2,0,0 ; vmovdqu %ymm1,0x240(%rsp) + DB 196,227,125,25,200,1 ; vextractf128 $0x1,%ymm1,%xmm0 + DB 197,249,127,132,36,224,0,0,0 ; vmovdqa %xmm0,0xe0(%rsp) + DB 197,201,254,192 ; vpaddd %xmm0,%xmm6,%xmm0 + DB 197,249,127,132,36,160,0,0,0 ; vmovdqa %xmm0,0xa0(%rsp) + DB 197,225,254,216 ; vpaddd %xmm0,%xmm3,%xmm3 + DB 197,233,254,195 ; vpaddd %xmm3,%xmm2,%xmm0 + DB 196,98,121,24,13,221,197,1,0 ; vbroadcastss 0x1c5dd(%rip),%xmm9 # 2d304 <_sk_overlay_sse2_8bit+0xe55> + DB 196,194,121,64,209 ; vpmulld %xmm9,%xmm0,%xmm2 + DB 196,226,81,64,199 ; vpmulld %xmm7,%xmm5,%xmm0 + DB 197,249,127,132,36,224,1,0,0 ; vmovdqa %xmm0,0x1e0(%rsp) + DB 197,121,254,241 ; vpaddd %xmm1,%xmm0,%xmm14 + DB 196,194,57,64,194 ; vpmulld %xmm10,%xmm8,%xmm0 + DB 197,249,127,68,36,112 ; vmovdqa %xmm0,0x70(%rsp) + DB 196,193,121,254,206 ; vpaddd %xmm14,%xmm0,%xmm1 + DB 196,194,89,64,195 ; vpmulld %xmm11,%xmm4,%xmm0 + DB 197,249,127,68,36,16 ; vmovdqa %xmm0,0x10(%rsp) + DB 197,249,254,193 ; vpaddd %xmm1,%xmm0,%xmm0 + DB 196,194,121,64,193 ; vpmulld %xmm9,%xmm0,%xmm0 + DB 196,193,121,126,192 ; vmovd %xmm0,%r8d + DB 196,195,121,22,193,1 ; vpextrd $0x1,%xmm0,%r9d + DB 196,195,121,22,194,2 ; vpextrd $0x2,%xmm0,%r10d + DB 196,195,121,22,195,3 ; vpextrd $0x3,%xmm0,%r11d + DB 196,193,121,126,215 ; vmovd %xmm2,%r15d + DB 196,195,121,22,214,1 ; vpextrd $0x1,%xmm2,%r14d + DB 196,195,121,22,212,2 ; vpextrd $0x2,%xmm2,%r12d + DB 196,227,121,22,211,3 ; vpextrd $0x3,%xmm2,%ebx + DB 196,161,122,16,44,184 ; vmovss (%rax,%r15,4),%xmm5 + DB 196,163,81,33,44,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm5,%xmm5 + DB 196,163,81,33,44,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm5,%xmm5 + DB 196,227,81,33,44,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm5,%xmm5 DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 DB 196,163,65,33,60,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm7,%xmm7 DB 196,163,65,33,60,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm7,%xmm7 - DB 196,99,53,24,247,1 ; vinsertf128 $0x1,%xmm7,%ymm9,%ymm14 - DB 196,98,121,24,21,157,200,1,0 ; vbroadcastss 0x1c89d(%rip),%xmm10 # 2d2b0 <_sk_overlay_sse2_8bit+0xd58> - DB 196,65,97,254,202 ; vpaddd %xmm10,%xmm3,%xmm9 - DB 196,67,249,22,200,1 ; vpextrq $0x1,%xmm9,%r8 - DB 196,65,249,126,201 ; vmovq %xmm9,%r9 - DB 196,65,73,254,202 ; vpaddd %xmm10,%xmm6,%xmm9 - DB 196,65,249,126,202 ; vmovq %xmm9,%r10 - DB 196,67,249,22,203,1 ; vpextrq $0x1,%xmm9,%r11 + DB 196,99,69,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm7,%ymm13 + DB 197,201,118,246 ; vpcmpeqd %xmm6,%xmm6,%xmm6 + DB 197,249,250,238 ; vpsubd %xmm6,%xmm0,%xmm5 + DB 196,195,249,22,232,1 ; vpextrq $0x1,%xmm5,%r8 + DB 196,193,249,126,233 ; vmovq %xmm5,%r9 + DB 197,233,250,238 ; vpsubd %xmm6,%xmm2,%xmm5 + DB 196,195,249,22,234,1 ; vpextrq $0x1,%xmm5,%r10 + DB 196,225,249,126,235 ; vmovq %xmm5,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,44,152 ; vmovss (%rax,%r11,4),%xmm5 + DB 196,227,81,33,44,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm5,%xmm5 + DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 + DB 196,227,81,33,239,32 ; vinsertps $0x20,%xmm7,%xmm5,%xmm5 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,65,33,228,32 ; vinsertps $0x20,%xmm4,%xmm7,%xmm4 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,65,33,246,32 ; vinsertps $0x20,%xmm6,%xmm7,%xmm6 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 68,137,211 ; mov %r10d,%ebx - DB 69,137,216 ; mov %r11d,%r8d + DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 + DB 196,227,77,24,237,1 ; vinsertf128 $0x1,%xmm5,%ymm6,%ymm5 + DB 196,98,121,24,21,54,195,1,0 ; vbroadcastss 0x1c336(%rip),%xmm10 # 2d198 <_sk_overlay_sse2_8bit+0xce9> + DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 + DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 196,193,105,254,194 ; vpaddd %xmm10,%xmm2,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 + DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 - DB 196,227,65,33,237,32 ; vinsertps $0x20,%xmm5,%xmm7,%xmm5 - DB 73,193,235,32 ; shr $0x20,%r11 - DB 196,161,122,16,60,152 ; vmovss (%rax,%r11,4),%xmm7 - DB 196,227,81,33,239,48 ; vinsertps $0x30,%xmm7,%xmm5,%xmm5 - DB 196,99,85,24,220,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm11 - DB 196,98,121,24,13,18,200,1,0 ; vbroadcastss 0x1c812(%rip),%xmm9 # 2d2b8 <_sk_overlay_sse2_8bit+0xd60> - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 + DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,105,33,214,32 ; vinsertps $0x20,%xmm6,%xmm2,%xmm2 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,105,33,214,48 ; vinsertps $0x30,%xmm6,%xmm2,%xmm2 + DB 196,227,109,24,208,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm2 + DB 196,98,125,24,37,10,196,1,0 ; vbroadcastss 0x1c40a(%rip),%ymm12 # 2d300 <_sk_overlay_sse2_8bit+0xe51> + DB 196,193,4,88,196 ; vaddps %ymm12,%ymm15,%ymm0 + DB 197,124,17,164,36,128,0,0,0 ; vmovups %ymm12,0x80(%rsp) + DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 + DB 196,227,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm6 + DB 196,226,89,64,246 ; vpmulld %xmm6,%xmm4,%xmm6 + DB 197,249,127,116,36,48 ; vmovdqa %xmm6,0x30(%rsp) + DB 196,226,89,64,192 ; vpmulld %xmm0,%xmm4,%xmm0 + DB 197,249,127,68,36,32 ; vmovdqa %xmm0,0x20(%rsp) + DB 197,249,254,193 ; vpaddd %xmm1,%xmm0,%xmm0 + DB 197,201,254,203 ; vpaddd %xmm3,%xmm6,%xmm1 + DB 196,194,113,64,201 ; vpmulld %xmm9,%xmm1,%xmm1 + DB 196,194,121,64,225 ; vpmulld %xmm9,%xmm0,%xmm4 + DB 196,193,121,126,225 ; vmovd %xmm4,%r9d + DB 196,195,121,22,224,1 ; vpextrd $0x1,%xmm4,%r8d + DB 196,195,121,22,226,2 ; vpextrd $0x2,%xmm4,%r10d + DB 196,195,121,22,227,3 ; vpextrd $0x3,%xmm4,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 + DB 197,201,118,246 ; vpcmpeqd %xmm6,%xmm6,%xmm6 + DB 197,217,250,222 ; vpsubd %xmm6,%xmm4,%xmm3 DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,73,254,217 ; vpaddd %xmm9,%xmm6,%xmm3 - DB 196,193,249,126,218 ; vmovq %xmm3,%r10 - DB 196,227,249,22,219,1 ; vpextrq $0x1,%xmm3,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 197,241,250,222 ; vpsubd %xmm6,%xmm1,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 - DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 - DB 196,227,89,33,229,32 ; vinsertps $0x20,%xmm5,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 - DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 - DB 196,227,93,24,243,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm6 - DB 196,226,125,24,37,230,200,1,0 ; vbroadcastss 0x1c8e6(%rip),%ymm4 # 2d420 <_sk_overlay_sse2_8bit+0xec8> - DB 197,148,88,220 ; vaddps %ymm4,%ymm13,%ymm3 - DB 197,252,40,252 ; vmovaps %ymm4,%ymm7 - DB 197,252,17,188,36,160,0,0,0 ; vmovups %ymm7,0xa0(%rsp) - DB 197,254,91,219 ; vcvttps2dq %ymm3,%ymm3 - DB 196,227,125,25,220,1 ; vextractf128 $0x1,%ymm3,%xmm4 - DB 196,226,89,64,224 ; vpmulld %xmm0,%xmm4,%xmm4 - DB 197,249,127,100,36,32 ; vmovdqa %xmm4,0x20(%rsp) - DB 196,226,97,64,192 ; vpmulld %xmm0,%xmm3,%xmm0 - DB 197,249,127,68,36,16 ; vmovdqa %xmm0,0x10(%rsp) - DB 197,249,254,194 ; vpaddd %xmm2,%xmm0,%xmm0 - DB 196,194,121,64,215 ; vpmulld %xmm15,%xmm0,%xmm2 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,121,33,219,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm3 - DB 197,217,254,193 ; vpaddd %xmm1,%xmm4,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx - DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,101,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm3 - DB 196,193,121,254,202 ; vpaddd %xmm10,%xmm0,%xmm1 - DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 - DB 196,193,249,126,201 ; vmovq %xmm1,%r9 - DB 196,193,105,254,202 ; vpaddd %xmm10,%xmm2,%xmm1 - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 + DB 196,227,77,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm6,%ymm3 + DB 196,193,89,254,226 ; vpaddd %xmm10,%xmm4,%xmm4 + DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 + DB 196,193,249,126,225 ; vmovq %xmm4,%r9 + DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 DB 196,227,113,33,204,32 ; vinsertps $0x20,%xmm4,%xmm1,%xmm1 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,113,33,204,48 ; vinsertps $0x30,%xmm4,%xmm1,%xmm1 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,44,136 ; vmovss (%rax,%r9,4),%xmm5 - DB 196,227,89,33,229,32 ; vinsertps $0x20,%xmm5,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 - DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 - DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,193 ; vpaddd %xmm9,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,113,33,204,48 ; vinsertps $0x30,%xmm4,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,105,33,212,32 ; vinsertps $0x20,%xmm4,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,105,33,212,48 ; vinsertps $0x30,%xmm4,%xmm2,%xmm2 - DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 - DB 196,193,124,91,212 ; vcvtdq2ps %ymm12,%ymm2 - DB 197,148,92,234 ; vsubps %ymm2,%ymm13,%ymm5 - DB 196,193,100,92,214 ; vsubps %ymm14,%ymm3,%ymm2 - DB 197,212,89,210 ; vmulps %ymm2,%ymm5,%ymm2 - DB 197,12,88,234 ; vaddps %ymm2,%ymm14,%ymm13 - DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 - DB 197,212,89,201 ; vmulps %ymm1,%ymm5,%ymm1 - DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1 - DB 197,252,17,140,36,224,1,0,0 ; vmovups %ymm1,0x1e0(%rsp) - DB 197,252,92,198 ; vsubps %ymm6,%ymm0,%ymm0 - DB 197,212,89,192 ; vmulps %ymm0,%ymm5,%ymm0 - DB 197,204,88,192 ; vaddps %ymm0,%ymm6,%ymm0 + DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 + DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 + DB 196,193,124,91,227 ; vcvtdq2ps %ymm11,%ymm4 + DB 197,4,92,252 ; vsubps %ymm4,%ymm15,%ymm15 + DB 196,193,124,92,197 ; vsubps %ymm13,%ymm0,%ymm0 + DB 197,132,89,192 ; vmulps %ymm0,%ymm15,%ymm0 + DB 197,148,88,192 ; vaddps %ymm0,%ymm13,%ymm0 + DB 197,252,17,132,36,160,1,0,0 ; vmovups %ymm0,0x1a0(%rsp) + DB 197,228,92,197 ; vsubps %ymm5,%ymm3,%ymm0 + DB 197,132,89,192 ; vmulps %ymm0,%ymm15,%ymm0 + DB 197,212,88,192 ; vaddps %ymm0,%ymm5,%ymm0 DB 197,252,17,132,36,192,0,0,0 ; vmovups %ymm0,0xc0(%rsp) + DB 197,244,92,194 ; vsubps %ymm2,%ymm1,%ymm0 + DB 197,132,89,192 ; vmulps %ymm0,%ymm15,%ymm0 + DB 197,108,88,232 ; vaddps %ymm0,%ymm2,%ymm13 DB 197,124,16,156,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm11 - DB 197,164,88,199 ; vaddps %ymm7,%ymm11,%ymm0 + DB 196,193,36,88,196 ; vaddps %ymm12,%ymm11,%ymm0 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 196,194,121,64,200 ; vpmulld %xmm8,%xmm0,%xmm1 - DB 197,249,127,140,36,128,0,0,0 ; vmovdqa %xmm1,0x80(%rsp) + DB 196,226,57,64,200 ; vpmulld %xmm0,%xmm8,%xmm1 + DB 197,249,127,76,36,96 ; vmovdqa %xmm1,0x60(%rsp) DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 196,194,121,64,216 ; vpmulld %xmm8,%xmm0,%xmm3 - DB 197,249,127,92,36,112 ; vmovdqa %xmm3,0x70(%rsp) - DB 197,241,254,188,36,224,0,0,0 ; vpaddd 0xe0(%rsp),%xmm1,%xmm7 - DB 197,193,254,4,36 ; vpaddd (%rsp),%xmm7,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm1,%xmm1 - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 197,225,254,148,36,192,1,0,0 ; vpaddd 0x1c0(%rsp),%xmm3,%xmm2 - DB 197,233,254,92,36,48 ; vpaddd 0x30(%rsp),%xmm2,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,163,89,33,36,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm4,%xmm4 - DB 196,163,89,33,36,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,163,89,33,36,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,99,117,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm1,%ymm12 - DB 196,193,97,254,202 ; vpaddd %xmm10,%xmm3,%xmm1 + DB 196,226,57,64,192 ; vpmulld %xmm0,%xmm8,%xmm0 + DB 197,249,127,68,36,80 ; vmovdqa %xmm0,0x50(%rsp) + DB 197,249,254,172,36,160,0,0,0 ; vpaddd 0xa0(%rsp),%xmm0,%xmm5 + DB 196,193,113,254,198 ; vpaddd %xmm14,%xmm1,%xmm0 + DB 197,121,111,36,36 ; vmovdqa (%rsp),%xmm12 + DB 196,193,81,254,204 ; vpaddd %xmm12,%xmm5,%xmm1 + DB 196,194,113,64,209 ; vpmulld %xmm9,%xmm1,%xmm2 + DB 197,121,111,68,36,16 ; vmovdqa 0x10(%rsp),%xmm8 + DB 196,193,121,254,200 ; vpaddd %xmm8,%xmm0,%xmm1 + DB 196,194,113,64,241 ; vpmulld %xmm9,%xmm1,%xmm6 + DB 196,193,121,126,241 ; vmovd %xmm6,%r9d + DB 196,195,121,22,240,1 ; vpextrd $0x1,%xmm6,%r8d + DB 196,195,121,22,242,2 ; vpextrd $0x2,%xmm6,%r10d + DB 196,195,121,22,243,3 ; vpextrd $0x3,%xmm6,%r11d + DB 196,193,121,126,215 ; vmovd %xmm2,%r15d + DB 196,195,121,22,214,1 ; vpextrd $0x1,%xmm2,%r14d + DB 196,195,121,22,212,2 ; vpextrd $0x2,%xmm2,%r12d + DB 196,227,121,22,211,3 ; vpextrd $0x3,%xmm2,%ebx + DB 196,161,122,16,12,184 ; vmovss (%rax,%r15,4),%xmm1 + DB 196,163,113,33,12,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm1,%xmm1 + DB 196,163,113,33,12,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm1,%xmm1 + DB 196,227,113,33,12,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,99,101,24,241,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm14 + DB 197,225,118,219 ; vpcmpeqd %xmm3,%xmm3,%xmm3 + DB 197,201,250,203 ; vpsubd %xmm3,%xmm6,%xmm1 DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 DB 196,193,249,126,201 ; vmovq %xmm1,%r9 - DB 196,193,121,254,202 ; vpaddd %xmm10,%xmm0,%xmm1 - DB 196,193,249,126,202 ; vmovq %xmm1,%r10 - DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 197,233,250,203 ; vpsubd %xmm3,%xmm2,%xmm1 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,113,33,204,32 ; vinsertps $0x20,%xmm4,%xmm1,%xmm1 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,113,33,204,48 ; vinsertps $0x30,%xmm4,%xmm1,%xmm1 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 + DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 196,227,93,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm4,%ymm1 - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 + DB 196,193,73,254,218 ; vpaddd %xmm10,%xmm6,%xmm3 DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,99,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm8 - DB 197,121,111,116,36,16 ; vmovdqa 0x10(%rsp),%xmm14 - DB 197,137,254,199 ; vpaddd %xmm7,%xmm14,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,227,101,24,242,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm6 + DB 197,249,254,68,36,32 ; vpaddd 0x20(%rsp),%xmm0,%xmm0 + DB 197,209,254,84,36,48 ; vpaddd 0x30(%rsp),%xmm5,%xmm2 + DB 196,194,105,64,209 ; vpmulld %xmm9,%xmm2,%xmm2 + DB 196,194,121,64,233 ; vpmulld %xmm9,%xmm0,%xmm5 + DB 196,193,121,126,233 ; vmovd %xmm5,%r9d + DB 196,195,121,22,232,1 ; vpextrd $0x1,%xmm5,%r8d + DB 196,195,121,22,234,2 ; vpextrd $0x2,%xmm5,%r10d + DB 196,195,121,22,235,3 ; vpextrd $0x3,%xmm5,%r11d + DB 196,193,121,126,215 ; vmovd %xmm2,%r15d + DB 196,195,121,22,214,1 ; vpextrd $0x1,%xmm2,%r14d + DB 196,195,121,22,212,2 ; vpextrd $0x2,%xmm2,%r12d + DB 196,227,121,22,211,3 ; vpextrd $0x3,%xmm2,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 + DB 197,217,118,228 ; vpcmpeqd %xmm4,%xmm4,%xmm4 + DB 197,209,250,220 ; vpsubd %xmm4,%xmm5,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 197,233,250,220 ; vpsubd %xmm4,%xmm2,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 + DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,228,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm4 - DB 197,233,254,84,36,32 ; vpaddd 0x20(%rsp),%xmm2,%xmm2 - DB 196,194,105,64,223 ; vpmulld %xmm15,%xmm2,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 - DB 196,163,105,33,20,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 - DB 196,193,97,254,226 ; vpaddd %xmm10,%xmm3,%xmm4 - DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 - DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,121,254,226 ; vpaddd %xmm10,%xmm0,%xmm4 - DB 196,193,249,126,226 ; vmovq %xmm4,%r10 - DB 196,227,249,22,227,1 ; vpextrq $0x1,%xmm4,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,163,73,33,52,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm6,%xmm6 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 - DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 + DB 196,227,93,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm3 + DB 196,193,81,254,226 ; vpaddd %xmm10,%xmm5,%xmm4 + DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 + DB 196,193,249,126,225 ; vmovq %xmm4,%r9 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 - DB 196,227,77,24,252,1 ; vinsertf128 $0x1,%xmm4,%ymm6,%ymm7 - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 - DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 - DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 + DB 196,227,105,33,212,32 ; vinsertps $0x20,%xmm4,%xmm2,%xmm2 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,105,33,212,48 ; vinsertps $0x30,%xmm4,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 - DB 196,193,108,92,212 ; vsubps %ymm12,%ymm2,%ymm2 - DB 197,212,89,210 ; vmulps %ymm2,%ymm5,%ymm2 - DB 197,156,88,210 ; vaddps %ymm2,%ymm12,%ymm2 - DB 197,196,92,217 ; vsubps %ymm1,%ymm7,%ymm3 - DB 197,212,89,219 ; vmulps %ymm3,%ymm5,%ymm3 + DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,89,33,229,32 ; vinsertps $0x20,%xmm5,%xmm4,%xmm4 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 + DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 + DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 + DB 196,193,124,92,198 ; vsubps %ymm14,%ymm0,%ymm0 + DB 197,132,89,192 ; vmulps %ymm0,%ymm15,%ymm0 + DB 197,140,88,192 ; vaddps %ymm0,%ymm14,%ymm0 + DB 197,228,92,217 ; vsubps %ymm1,%ymm3,%ymm3 + DB 197,132,89,219 ; vmulps %ymm3,%ymm15,%ymm3 DB 197,244,88,203 ; vaddps %ymm3,%ymm1,%ymm1 - DB 196,193,124,92,192 ; vsubps %ymm8,%ymm0,%ymm0 - DB 197,212,89,192 ; vmulps %ymm0,%ymm5,%ymm0 - DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 - DB 197,252,91,156,36,96,1,0,0 ; vcvtdq2ps 0x160(%rsp),%ymm3 + DB 197,236,92,214 ; vsubps %ymm6,%ymm2,%ymm2 + DB 197,132,89,210 ; vmulps %ymm2,%ymm15,%ymm2 + DB 197,204,88,210 ; vaddps %ymm2,%ymm6,%ymm2 + DB 197,252,91,156,36,64,1,0,0 ; vcvtdq2ps 0x140(%rsp),%ymm3 DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 - DB 197,252,17,156,36,96,1,0,0 ; vmovups %ymm3,0x160(%rsp) - DB 196,193,108,92,213 ; vsubps %ymm13,%ymm2,%ymm2 - DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 - DB 197,148,88,210 ; vaddps %ymm2,%ymm13,%ymm2 - DB 197,252,17,148,36,0,1,0,0 ; vmovups %ymm2,0x100(%rsp) - DB 197,252,16,148,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm2 - DB 197,244,92,202 ; vsubps %ymm2,%ymm1,%ymm1 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 - DB 197,252,17,140,36,224,0,0,0 ; vmovups %ymm1,0xe0(%rsp) - DB 197,252,16,140,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm1 - DB 197,252,92,193 ; vsubps %ymm1,%ymm0,%ymm0 + DB 197,252,17,156,36,64,1,0,0 ; vmovups %ymm3,0x140(%rsp) + DB 197,252,16,164,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm4 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,244,88,192 ; vaddps %ymm0,%ymm1,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,17,132,36,0,1,0,0 ; vmovups %ymm0,0x100(%rsp) + DB 197,252,16,164,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm4 + DB 197,244,92,196 ; vsubps %ymm4,%ymm1,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 DB 197,252,17,132,36,192,0,0,0 ; vmovups %ymm0,0xc0(%rsp) - DB 197,252,16,132,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm0 - DB 197,252,88,132,36,160,1,0,0 ; vaddps 0x1a0(%rsp),%ymm0,%ymm0 + DB 196,193,108,92,197 ; vsubps %ymm13,%ymm2,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,148,88,192 ; vaddps %ymm0,%ymm13,%ymm0 + DB 197,252,17,132,36,160,0,0,0 ; vmovups %ymm0,0xa0(%rsp) + DB 197,252,16,132,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm0 + DB 197,252,88,132,36,128,1,0,0 ; vaddps 0x180(%rsp),%ymm0,%ymm0 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 - DB 197,249,111,84,36,96 ; vmovdqa 0x60(%rsp),%xmm2 - DB 196,226,113,64,218 ; vpmulld %xmm2,%xmm1,%xmm3 - DB 197,249,127,156,36,224,1,0,0 ; vmovdqa %xmm3,0x1e0(%rsp) - DB 196,226,121,64,194 ; vpmulld %xmm2,%xmm0,%xmm0 - DB 197,249,127,68,36,96 ; vmovdqa %xmm0,0x60(%rsp) - DB 197,249,254,132,36,96,2,0,0 ; vpaddd 0x260(%rsp),%xmm0,%xmm0 - DB 197,249,127,132,36,80,2,0,0 ; vmovdqa %xmm0,0x250(%rsp) - DB 197,249,254,148,36,128,1,0,0 ; vpaddd 0x180(%rsp),%xmm0,%xmm2 - DB 197,233,254,4,36 ; vpaddd (%rsp),%xmm2,%xmm0 - DB 196,194,121,64,247 ; vpmulld %xmm15,%xmm0,%xmm6 - DB 196,193,249,126,241 ; vmovq %xmm6,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,243,1 ; vpextrq $0x1,%xmm6,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 197,249,111,84,36,64 ; vmovdqa 0x40(%rsp),%xmm2 + DB 196,226,105,64,201 ; vpmulld %xmm1,%xmm2,%xmm1 + DB 197,249,127,140,36,160,1,0,0 ; vmovdqa %xmm1,0x1a0(%rsp) + DB 196,226,105,64,192 ; vpmulld %xmm0,%xmm2,%xmm0 + DB 197,249,127,68,36,64 ; vmovdqa %xmm0,0x40(%rsp) + DB 197,241,254,140,36,224,0,0,0 ; vpaddd 0xe0(%rsp),%xmm1,%xmm1 + DB 197,249,127,140,36,48,2,0,0 ; vmovdqa %xmm1,0x230(%rsp) + DB 197,241,254,172,36,96,1,0,0 ; vpaddd 0x160(%rsp),%xmm1,%xmm5 + DB 196,193,81,254,204 ; vpaddd %xmm12,%xmm5,%xmm1 + DB 196,194,113,64,225 ; vpmulld %xmm9,%xmm1,%xmm4 + DB 197,121,254,180,36,64,2,0,0 ; vpaddd 0x240(%rsp),%xmm0,%xmm14 + DB 197,137,254,84,36,112 ; vpaddd 0x70(%rsp),%xmm14,%xmm2 + DB 196,193,105,254,248 ; vpaddd %xmm8,%xmm2,%xmm7 + DB 196,194,65,64,249 ; vpmulld %xmm9,%xmm7,%xmm7 + DB 196,193,121,126,249 ; vmovd %xmm7,%r9d + DB 196,195,121,22,248,1 ; vpextrd $0x1,%xmm7,%r8d + DB 196,195,121,22,250,2 ; vpextrd $0x2,%xmm7,%r10d + DB 196,195,121,22,251,3 ; vpextrd $0x3,%xmm7,%r11d + DB 196,193,121,126,231 ; vmovd %xmm4,%r15d + DB 196,195,121,22,230,1 ; vpextrd $0x1,%xmm4,%r14d + DB 196,195,121,22,228,2 ; vpextrd $0x2,%xmm4,%r12d + DB 196,227,121,22,227,3 ; vpextrd $0x3,%xmm4,%ebx + DB 196,161,122,16,28,184 ; vmovss (%rax,%r15,4),%xmm3 + DB 196,163,97,33,28,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm3,%xmm3 + DB 196,163,97,33,28,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm3,%xmm3 + DB 196,227,97,33,28,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 + DB 196,163,73,33,52,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm6,%xmm6 + DB 196,163,73,33,52,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm6,%xmm6 + DB 196,163,73,33,52,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm6,%xmm6 + DB 196,227,77,24,203,1 ; vinsertf128 $0x1,%xmm3,%ymm6,%ymm1 + DB 197,249,118,192 ; vpcmpeqd %xmm0,%xmm0,%xmm0 + DB 197,193,250,216 ; vpsubd %xmm0,%xmm7,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 197,217,250,216 ; vpsubd %xmm0,%xmm4,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,227,121,33,225,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm4 - DB 197,225,254,68,36,64 ; vpaddd 0x40(%rsp),%xmm3,%xmm0 - DB 197,249,127,132,36,64,2,0,0 ; vmovdqa %xmm0,0x240(%rsp) - DB 197,249,254,140,36,144,0,0,0 ; vpaddd 0x90(%rsp),%xmm0,%xmm1 - DB 197,121,111,108,36,48 ; vmovdqa 0x30(%rsp),%xmm13 - DB 196,193,113,254,221 ; vpaddd %xmm13,%xmm1,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,163,65,33,60,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm7,%xmm7 - DB 196,163,65,33,60,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm7,%xmm7 - DB 196,163,65,33,60,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,227,93,24,199,1 ; vinsertf128 $0x1,%xmm7,%ymm4,%ymm0 - DB 197,252,17,68,36,64 ; vmovups %ymm0,0x40(%rsp) - DB 196,193,97,254,250 ; vpaddd %xmm10,%xmm3,%xmm7 - DB 196,195,249,22,248,1 ; vpextrq $0x1,%xmm7,%r8 - DB 196,193,249,126,249 ; vmovq %xmm7,%r9 - DB 196,193,73,254,250 ; vpaddd %xmm10,%xmm6,%xmm7 - DB 196,193,249,126,250 ; vmovq %xmm7,%r10 - DB 196,227,249,22,251,1 ; vpextrq $0x1,%xmm7,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,60,152 ; vmovss (%rax,%r11,4),%xmm7 - DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 - DB 196,161,122,16,4,176 ; vmovss (%rax,%r14,4),%xmm0 - DB 196,227,65,33,192,32 ; vinsertps $0x20,%xmm0,%xmm7,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,4,144 ; vmovss (%rax,%r10,4),%xmm0 + DB 196,227,73,33,192,32 ; vinsertps $0x20,%xmm0,%xmm6,%xmm0 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,227,121,33,199,48 ; vinsertps $0x30,%xmm7,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,65,33,228,32 ; vinsertps $0x20,%xmm4,%xmm7,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 196,99,93,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm4,%ymm8 - DB 196,193,97,254,193 ; vpaddd %xmm9,%xmm3,%xmm0 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,121,33,198,48 ; vinsertps $0x30,%xmm6,%xmm0,%xmm0 + DB 196,99,125,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm13 + DB 196,193,65,254,194 ; vpaddd %xmm10,%xmm7,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,73,254,193 ; vpaddd %xmm9,%xmm6,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,89,254,194 ; vpaddd %xmm10,%xmm4,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,99,101,24,224,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm12 - DB 197,137,254,194 ; vpaddd %xmm2,%xmm14,%xmm0 - DB 196,194,121,64,215 ; vpmulld %xmm15,%xmm0,%xmm2 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 197,121,111,92,36,32 ; vmovdqa 0x20(%rsp),%xmm11 - DB 197,161,254,201 ; vpaddd %xmm1,%xmm11,%xmm1 - DB 196,194,113,64,223 ; vpmulld %xmm15,%xmm1,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,125,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm1 - DB 196,193,97,254,194 ; vpaddd %xmm10,%xmm3,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,194 ; vpaddd %xmm10,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,121,33,196,32 ; vinsertps $0x20,%xmm4,%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,121,33,196,48 ; vinsertps $0x30,%xmm4,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 196,227,93,24,240,1 ; vinsertf128 $0x1,%xmm0,%ymm4,%ymm6 - DB 196,193,97,254,193 ; vpaddd %xmm9,%xmm3,%xmm0 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 196,99,101,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm11 + DB 197,121,111,100,36,32 ; vmovdqa 0x20(%rsp),%xmm12 + DB 197,153,254,194 ; vpaddd %xmm2,%xmm12,%xmm0 + DB 197,121,111,68,36,48 ; vmovdqa 0x30(%rsp),%xmm8 + DB 197,185,254,213 ; vpaddd %xmm5,%xmm8,%xmm2 + DB 196,194,105,64,225 ; vpmulld %xmm9,%xmm2,%xmm4 + DB 196,194,121,64,233 ; vpmulld %xmm9,%xmm0,%xmm5 + DB 196,193,121,126,233 ; vmovd %xmm5,%r9d + DB 196,195,121,22,232,1 ; vpextrd $0x1,%xmm5,%r8d + DB 196,195,121,22,234,2 ; vpextrd $0x2,%xmm5,%r10d + DB 196,195,121,22,235,3 ; vpextrd $0x3,%xmm5,%r11d + DB 196,193,121,126,231 ; vmovd %xmm4,%r15d + DB 196,195,121,22,230,1 ; vpextrd $0x1,%xmm4,%r14d + DB 196,195,121,22,228,2 ; vpextrd $0x2,%xmm4,%r12d + DB 196,227,121,22,227,3 ; vpextrd $0x3,%xmm4,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 + DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 + DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 + DB 196,163,105,33,20,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm2,%xmm2 + DB 196,227,109,24,208,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm2 + DB 197,225,118,219 ; vpcmpeqd %xmm3,%xmm3,%xmm3 + DB 197,209,250,195 ; vpsubd %xmm3,%xmm5,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,193 ; vpaddd %xmm9,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 197,217,250,195 ; vpsubd %xmm3,%xmm4,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 - DB 197,252,16,84,36,64 ; vmovups 0x40(%rsp),%ymm2 - DB 197,244,92,202 ; vsubps %ymm2,%ymm1,%ymm1 - DB 197,212,89,201 ; vmulps %ymm1,%ymm5,%ymm1 - DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 - DB 197,252,17,76,36,64 ; vmovups %ymm1,0x40(%rsp) - DB 196,193,76,92,200 ; vsubps %ymm8,%ymm6,%ymm1 - DB 197,212,89,201 ; vmulps %ymm1,%ymm5,%ymm1 - DB 197,188,88,201 ; vaddps %ymm1,%ymm8,%ymm1 - DB 197,252,17,140,36,192,1,0,0 ; vmovups %ymm1,0x1c0(%rsp) - DB 196,193,124,92,196 ; vsubps %ymm12,%ymm0,%ymm0 - DB 197,212,89,192 ; vmulps %ymm0,%ymm5,%ymm0 - DB 197,28,88,192 ; vaddps %ymm0,%ymm12,%ymm8 - DB 197,249,111,132,36,80,2,0,0 ; vmovdqa 0x250(%rsp),%xmm0 - DB 197,249,254,180,36,128,0,0,0 ; vpaddd 0x80(%rsp),%xmm0,%xmm6 - DB 197,201,254,4,36 ; vpaddd (%rsp),%xmm6,%xmm0 - DB 196,194,121,64,215 ; vpmulld %xmm15,%xmm0,%xmm2 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,227,121,33,217,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm3 - DB 197,249,111,132,36,64,2,0,0 ; vmovdqa 0x240(%rsp),%xmm0 - DB 197,249,254,76,36,112 ; vpaddd 0x70(%rsp),%xmm0,%xmm1 - DB 196,193,113,254,197 ; vpaddd %xmm13,%xmm1,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx - DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,163,89,33,36,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm4,%xmm4 - DB 196,163,89,33,36,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,163,89,33,36,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,99,101,24,236,1 ; vinsertf128 $0x1,%xmm4,%ymm3,%ymm13 - DB 196,193,121,254,226 ; vpaddd %xmm10,%xmm0,%xmm4 - DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 - DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,105,254,226 ; vpaddd %xmm10,%xmm2,%xmm4 - DB 196,193,249,126,226 ; vmovq %xmm4,%r10 - DB 196,227,249,22,227,1 ; vpextrq $0x1,%xmm4,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 - DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 - DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,65,33,219,32 ; vinsertps $0x20,%xmm3,%xmm7,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 - DB 196,99,101,24,244,1 ; vinsertf128 $0x1,%xmm4,%ymm3,%ymm14 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 196,227,101,24,248,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm7 + DB 196,193,81,254,194 ; vpaddd %xmm10,%xmm5,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,193 ; vpaddd %xmm9,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,89,254,194 ; vpaddd %xmm10,%xmm4,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 + DB 197,236,92,209 ; vsubps %ymm1,%ymm2,%ymm2 + DB 197,132,89,210 ; vmulps %ymm2,%ymm15,%ymm2 + DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1 + DB 197,252,17,140,36,224,0,0,0 ; vmovups %ymm1,0xe0(%rsp) + DB 196,193,68,92,213 ; vsubps %ymm13,%ymm7,%ymm2 + DB 197,132,89,210 ; vmulps %ymm2,%ymm15,%ymm2 + DB 197,148,88,202 ; vaddps %ymm2,%ymm13,%ymm1 + DB 197,252,17,140,36,160,2,0,0 ; vmovups %ymm1,0x2a0(%rsp) + DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 + DB 197,132,89,192 ; vmulps %ymm0,%ymm15,%ymm0 + DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 + DB 197,252,17,132,36,128,2,0,0 ; vmovups %ymm0,0x280(%rsp) + DB 197,249,111,132,36,48,2,0,0 ; vmovdqa 0x230(%rsp),%xmm0 + DB 197,249,254,108,36,80 ; vpaddd 0x50(%rsp),%xmm0,%xmm5 + DB 197,137,254,100,36,96 ; vpaddd 0x60(%rsp),%xmm14,%xmm4 + DB 197,209,254,4,36 ; vpaddd (%rsp),%xmm5,%xmm0 + DB 196,66,121,64,217 ; vpmulld %xmm9,%xmm0,%xmm11 + DB 197,121,111,108,36,16 ; vmovdqa 0x10(%rsp),%xmm13 + DB 196,193,89,254,197 ; vpaddd %xmm13,%xmm4,%xmm0 + DB 196,194,121,64,209 ; vpmulld %xmm9,%xmm0,%xmm2 + DB 196,193,121,126,209 ; vmovd %xmm2,%r9d + DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d + DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d + DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d + DB 196,65,121,126,223 ; vmovd %xmm11,%r15d + DB 196,67,121,22,222,1 ; vpextrd $0x1,%xmm11,%r14d + DB 196,67,121,22,220,2 ; vpextrd $0x2,%xmm11,%r12d + DB 196,99,121,22,219,3 ; vpextrd $0x3,%xmm11,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,12,136 ; vmovss (%rax,%r9,4),%xmm1 + DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 + DB 196,163,113,33,12,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm1,%xmm1 + DB 196,163,113,33,12,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm1,%xmm1 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 197,225,118,219 ; vpcmpeqd %xmm3,%xmm3,%xmm3 + DB 197,233,250,203 ; vpsubd %xmm3,%xmm2,%xmm1 + DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 197,161,250,203 ; vpsubd %xmm3,%xmm11,%xmm1 + DB 197,193,118,255 ; vpcmpeqd %xmm7,%xmm7,%xmm7 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,99,109,24,224,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm12 - DB 197,201,254,68,36,16 ; vpaddd 0x10(%rsp),%xmm6,%xmm0 - DB 196,194,121,64,215 ; vpmulld %xmm15,%xmm0,%xmm2 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 + DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,193,33,254,210 ; vpaddd %xmm10,%xmm11,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 + DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,121,33,219,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm3 - DB 197,161,254,193 ; vpaddd %xmm1,%xmm11,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 197,249,126,195 ; vmovd %xmm0,%ebx - DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d - DB 196,195,121,22,193,2 ; vpextrd $0x2,%xmm0,%r9d - DB 196,195,121,22,194,3 ; vpextrd $0x3,%xmm0,%r10d - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 - DB 196,193,121,254,218 ; vpaddd %xmm10,%xmm0,%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 196,99,101,24,218,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm11 + DB 197,153,254,212 ; vpaddd %xmm4,%xmm12,%xmm2 + DB 197,185,254,221 ; vpaddd %xmm5,%xmm8,%xmm3 + DB 196,194,97,64,225 ; vpmulld %xmm9,%xmm3,%xmm4 + DB 196,194,105,64,233 ; vpmulld %xmm9,%xmm2,%xmm5 + DB 196,193,121,126,233 ; vmovd %xmm5,%r9d + DB 196,195,121,22,232,1 ; vpextrd $0x1,%xmm5,%r8d + DB 196,195,121,22,234,2 ; vpextrd $0x2,%xmm5,%r10d + DB 196,195,121,22,235,3 ; vpextrd $0x3,%xmm5,%r11d + DB 196,193,121,126,231 ; vmovd %xmm4,%r15d + DB 196,195,121,22,230,1 ; vpextrd $0x1,%xmm4,%r14d + DB 196,195,121,22,228,2 ; vpextrd $0x2,%xmm4,%r12d + DB 196,227,121,22,227,3 ; vpextrd $0x3,%xmm4,%ebx + DB 196,161,122,16,20,184 ; vmovss (%rax,%r15,4),%xmm2 + DB 196,163,105,33,20,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm2,%xmm2 + DB 196,163,105,33,20,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm2,%xmm2 + DB 196,227,105,33,20,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 + DB 197,209,250,223 ; vpsubd %xmm7,%xmm5,%xmm3 DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,105,254,218 ; vpaddd %xmm10,%xmm2,%xmm3 - DB 196,193,249,126,218 ; vmovq %xmm3,%r10 - DB 196,227,249,22,219,1 ; vpextrq $0x1,%xmm3,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 197,217,250,223 ; vpsubd %xmm7,%xmm4,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 + DB 196,99,77,24,243,1 ; vinsertf128 $0x1,%xmm3,%ymm6,%ymm14 + DB 196,193,81,254,218 ; vpaddd %xmm10,%xmm5,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 196,193,89,254,218 ; vpaddd %xmm10,%xmm4,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 - DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 196,227,93,24,243,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm6 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,193 ; vpaddd %xmm9,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 + DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,89,33,229,32 ; vinsertps $0x20,%xmm5,%xmm4,%xmm4 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 - DB 196,193,116,92,205 ; vsubps %ymm13,%ymm1,%ymm1 - DB 197,252,17,172,36,32,2,0,0 ; vmovups %ymm5,0x220(%rsp) - DB 197,212,89,201 ; vmulps %ymm1,%ymm5,%ymm1 - DB 197,148,88,201 ; vaddps %ymm1,%ymm13,%ymm1 - DB 196,193,76,92,214 ; vsubps %ymm14,%ymm6,%ymm2 - DB 197,212,89,210 ; vmulps %ymm2,%ymm5,%ymm2 - DB 197,140,88,210 ; vaddps %ymm2,%ymm14,%ymm2 - DB 196,193,124,92,196 ; vsubps %ymm12,%ymm0,%ymm0 - DB 197,212,89,192 ; vmulps %ymm0,%ymm5,%ymm0 - DB 197,156,88,192 ; vaddps %ymm0,%ymm12,%ymm0 - DB 197,252,16,100,36,64 ; vmovups 0x40(%rsp),%ymm4 + DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 + DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 + DB 196,227,93,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm3 + DB 197,236,92,208 ; vsubps %ymm0,%ymm2,%ymm2 + DB 197,124,17,188,36,0,2,0,0 ; vmovups %ymm15,0x200(%rsp) + DB 197,132,89,210 ; vmulps %ymm2,%ymm15,%ymm2 + DB 197,252,88,194 ; vaddps %ymm2,%ymm0,%ymm0 + DB 197,140,92,209 ; vsubps %ymm1,%ymm14,%ymm2 + DB 197,132,89,210 ; vmulps %ymm2,%ymm15,%ymm2 + DB 197,244,88,202 ; vaddps %ymm2,%ymm1,%ymm1 + DB 196,193,100,92,211 ; vsubps %ymm11,%ymm3,%ymm2 + DB 197,132,89,210 ; vmulps %ymm2,%ymm15,%ymm2 + DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 + DB 197,252,16,164,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm4 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 197,252,16,156,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm3 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,16,164,36,160,2,0,0 ; vmovups 0x2a0(%rsp),%ymm4 DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 - DB 197,252,16,156,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm3 DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,16,164,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm4 + DB 197,252,16,164,36,128,2,0,0 ; vmovups 0x280(%rsp),%ymm4 DB 197,236,92,212 ; vsubps %ymm4,%ymm2,%ymm2 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 DB 197,220,88,210 ; vaddps %ymm2,%ymm4,%ymm2 - DB 196,193,124,92,192 ; vsubps %ymm8,%ymm0,%ymm0 - DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 - DB 197,252,91,156,36,0,2,0,0 ; vcvtdq2ps 0x200(%rsp),%ymm3 - DB 197,252,16,164,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm4 + DB 197,252,91,156,36,192,1,0,0 ; vcvtdq2ps 0x1c0(%rsp),%ymm3 + DB 197,252,16,164,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm4 DB 197,220,92,219 ; vsubps %ymm3,%ymm4,%ymm3 - DB 197,252,17,92,36,64 ; vmovups %ymm3,0x40(%rsp) + DB 197,252,17,156,36,224,0,0,0 ; vmovups %ymm3,0xe0(%rsp) DB 197,252,16,164,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm4 - DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,17,140,36,160,1,0,0 ; vmovups %ymm1,0x1a0(%rsp) - DB 197,252,16,164,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm4 - DB 197,236,92,204 ; vsubps %ymm4,%ymm2,%ymm1 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,17,140,36,0,2,0,0 ; vmovups %ymm1,0x200(%rsp) - DB 197,252,16,140,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm1 - DB 197,252,92,193 ; vsubps %ymm1,%ymm0,%ymm0 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,17,132,36,128,1,0,0 ; vmovups %ymm0,0x180(%rsp) + DB 197,252,16,164,36,192,0,0,0 ; vmovups 0xc0(%rsp),%ymm4 + DB 197,244,92,196 ; vsubps %ymm4,%ymm1,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,17,132,36,192,1,0,0 ; vmovups %ymm0,0x1c0(%rsp) + DB 197,252,16,140,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm1 + DB 197,236,92,193 ; vsubps %ymm1,%ymm2,%ymm0 DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 DB 197,244,88,192 ; vaddps %ymm0,%ymm1,%ymm0 DB 197,252,17,132,36,0,1,0,0 ; vmovups %ymm0,0x100(%rsp) - DB 197,252,16,132,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm0 - DB 197,252,88,132,36,128,2,0,0 ; vaddps 0x280(%rsp),%ymm0,%ymm0 + DB 197,252,16,132,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm0 + DB 197,252,88,132,36,96,2,0,0 ; vaddps 0x260(%rsp),%ymm0,%ymm0 DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 - DB 197,249,254,140,36,64,1,0,0 ; vpaddd 0x140(%rsp),%xmm0,%xmm1 - DB 197,249,127,140,36,160,0,0,0 ; vmovdqa %xmm1,0xa0(%rsp) - DB 197,253,111,224 ; vmovdqa %ymm0,%ymm4 - DB 197,254,127,164,36,192,0,0,0 ; vmovdqu %ymm4,0xc0(%rsp) - DB 197,241,254,148,36,128,1,0,0 ; vpaddd 0x180(%rsp),%xmm1,%xmm2 - DB 197,249,111,44,36 ; vmovdqa (%rsp),%xmm5 - DB 197,209,254,194 ; vpaddd %xmm2,%xmm5,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm1,%xmm1 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,113,33,251,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm7 - DB 196,227,125,25,225,1 ; vextractf128 $0x1,%ymm4,%xmm1 - DB 197,248,41,140,36,224,0,0,0 ; vmovaps %xmm1,0xe0(%rsp) - DB 197,113,254,180,36,32,1,0,0 ; vpaddd 0x120(%rsp),%xmm1,%xmm14 - DB 197,137,254,140,36,144,0,0,0 ; vpaddd 0x90(%rsp),%xmm14,%xmm1 - DB 197,241,254,92,36,48 ; vpaddd 0x30(%rsp),%xmm1,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,163,89,33,36,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm4,%xmm4 - DB 196,163,89,33,36,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,163,89,33,36,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,99,69,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm7,%ymm12 - DB 196,193,97,254,226 ; vpaddd %xmm10,%xmm3,%xmm4 - DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 - DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,121,254,226 ; vpaddd %xmm10,%xmm0,%xmm4 - DB 196,225,249,126,227 ; vmovq %xmm4,%rbx - DB 196,195,249,22,226,1 ; vpextrq $0x1,%xmm4,%r10 - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 - DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 65,137,216 ; mov %ebx,%r8d - DB 69,137,209 ; mov %r10d,%r9d + DB 197,254,127,132,36,192,0,0,0 ; vmovdqu %ymm0,0xc0(%rsp) + DB 197,249,254,148,36,224,1,0,0 ; vpaddd 0x1e0(%rsp),%xmm0,%xmm2 + DB 197,249,127,148,36,128,0,0,0 ; vmovdqa %xmm2,0x80(%rsp) + DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 + DB 197,249,127,132,36,160,0,0,0 ; vmovdqa %xmm0,0xa0(%rsp) + DB 197,121,254,188,36,32,1,0,0 ; vpaddd 0x120(%rsp),%xmm0,%xmm15 + DB 197,129,254,172,36,96,1,0,0 ; vpaddd 0x160(%rsp),%xmm15,%xmm5 + DB 197,249,111,36,36 ; vmovdqa (%rsp),%xmm4 + DB 197,217,254,197 ; vpaddd %xmm5,%xmm4,%xmm0 + DB 196,194,121,64,201 ; vpmulld %xmm9,%xmm0,%xmm1 + DB 197,233,254,84,36,112 ; vpaddd 0x70(%rsp),%xmm2,%xmm2 + DB 197,145,254,194 ; vpaddd %xmm2,%xmm13,%xmm0 + DB 196,194,121,64,193 ; vpmulld %xmm9,%xmm0,%xmm0 + DB 196,193,121,126,193 ; vmovd %xmm0,%r9d + DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d + DB 196,195,121,22,194,2 ; vpextrd $0x2,%xmm0,%r10d + DB 196,195,121,22,195,3 ; vpextrd $0x3,%xmm0,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,52,184 ; vmovss (%rax,%r15,4),%xmm6 + DB 196,163,73,33,52,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm6,%xmm6 + DB 196,163,73,33,52,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm6,%xmm6 + DB 196,227,73,33,52,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm6,%xmm6 + DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 + DB 196,163,65,33,60,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm7,%xmm7 + DB 196,163,65,33,60,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm7,%xmm7 + DB 196,163,65,33,60,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm7,%xmm7 + DB 196,99,69,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm7,%ymm14 + DB 197,225,118,219 ; vpcmpeqd %xmm3,%xmm3,%xmm3 + DB 197,249,250,243 ; vpsubd %xmm3,%xmm0,%xmm6 + DB 196,195,249,22,240,1 ; vpextrq $0x1,%xmm6,%r8 + DB 196,193,249,126,241 ; vmovq %xmm6,%r9 + DB 197,241,250,243 ; vpsubd %xmm3,%xmm1,%xmm6 + DB 196,65,17,118,237 ; vpcmpeqd %xmm13,%xmm13,%xmm13 + DB 196,195,249,22,242,1 ; vpextrq $0x1,%xmm6,%r10 + DB 196,225,249,126,243 ; vmovq %xmm6,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,52,152 ; vmovss (%rax,%r11,4),%xmm6 DB 196,227,73,33,52,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm6,%xmm6 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 + DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 DB 73,193,234,32 ; shr $0x20,%r10 DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 - DB 196,99,77,24,196,1 ; vinsertf128 $0x1,%xmm4,%ymm6,%ymm8 - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 - DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 - DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 + DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,65,33,219,32 ; vinsertps $0x20,%xmm3,%xmm7,%xmm3 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,99,101,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm13 - DB 197,121,111,92,36,16 ; vmovdqa 0x10(%rsp),%xmm11 - DB 197,161,254,194 ; vpaddd %xmm2,%xmm11,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,97,33,223,48 ; vinsertps $0x30,%xmm7,%xmm3,%xmm3 + DB 196,227,101,24,254,1 ; vinsertf128 $0x1,%xmm6,%ymm3,%ymm7 + DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 + DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,193,113,254,194 ; vpaddd %xmm10,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,12,176 ; vmovss (%rax,%r14,4),%xmm1 + DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,12,144 ; vmovss (%rax,%r10,4),%xmm1 + DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 197,241,254,76,36,32 ; vpaddd 0x20(%rsp),%xmm1,%xmm1 - DB 196,194,113,64,223 ; vpmulld %xmm15,%xmm1,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 - DB 196,193,97,254,210 ; vpaddd %xmm10,%xmm3,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,121,254,210 ; vpaddd %xmm10,%xmm0,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,105,33,212,32 ; vinsertps $0x20,%xmm4,%xmm2,%xmm2 + DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,105,33,212,48 ; vinsertps $0x30,%xmm4,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 + DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 196,99,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm8 + DB 197,121,111,100,36,32 ; vmovdqa 0x20(%rsp),%xmm12 + DB 197,153,254,194 ; vpaddd %xmm2,%xmm12,%xmm0 + DB 197,121,111,92,36,48 ; vmovdqa 0x30(%rsp),%xmm11 + DB 197,161,254,205 ; vpaddd %xmm5,%xmm11,%xmm1 + DB 196,194,113,64,201 ; vpmulld %xmm9,%xmm1,%xmm1 + DB 196,194,121,64,193 ; vpmulld %xmm9,%xmm0,%xmm0 + DB 196,193,121,126,193 ; vmovd %xmm0,%r9d + DB 196,195,121,22,192,1 ; vpextrd $0x1,%xmm0,%r8d + DB 196,195,121,22,194,2 ; vpextrd $0x2,%xmm0,%r10d + DB 196,195,121,22,195,3 ; vpextrd $0x3,%xmm0,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,20,184 ; vmovss (%rax,%r15,4),%xmm2 + DB 196,163,105,33,20,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm2,%xmm2 + DB 196,163,105,33,20,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm2,%xmm2 + DB 196,227,105,33,20,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm2,%xmm2 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 + DB 196,193,121,250,221 ; vpsubd %xmm13,%xmm0,%xmm3 DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 - DB 196,193,116,92,204 ; vsubps %ymm12,%ymm1,%ymm1 - DB 197,252,16,156,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm3 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,28,88,225 ; vaddps %ymm1,%ymm12,%ymm12 - DB 196,193,108,92,200 ; vsubps %ymm8,%ymm2,%ymm1 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,188,88,201 ; vaddps %ymm1,%ymm8,%ymm1 - DB 197,252,17,140,36,64,1,0,0 ; vmovups %ymm1,0x140(%rsp) - DB 196,193,124,92,197 ; vsubps %ymm13,%ymm0,%ymm0 - DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,148,88,192 ; vaddps %ymm0,%ymm13,%ymm0 - DB 197,252,17,132,36,32,1,0,0 ; vmovups %ymm0,0x120(%rsp) - DB 197,249,111,132,36,160,0,0,0 ; vmovdqa 0xa0(%rsp),%xmm0 - DB 197,249,254,148,36,128,0,0,0 ; vpaddd 0x80(%rsp),%xmm0,%xmm2 - DB 197,233,254,197 ; vpaddd %xmm5,%xmm2,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,193,113,250,221 ; vpsubd %xmm13,%xmm1,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,44,176 ; vmovss (%rax,%r14,4),%xmm5 + DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm1,%xmm1 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,113,33,227,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm4 - DB 197,137,254,76,36,112 ; vpaddd 0x70(%rsp),%xmm14,%xmm1 - DB 197,121,111,68,36,48 ; vmovdqa 0x30(%rsp),%xmm8 - DB 196,193,113,254,216 ; vpaddd %xmm8,%xmm1,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,163,73,33,52,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm6,%xmm6 - DB 196,163,73,33,52,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm6,%xmm6 - DB 196,163,73,33,52,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm6,%xmm6 - DB 196,99,93,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm4,%ymm14 - DB 196,193,97,254,226 ; vpaddd %xmm10,%xmm3,%xmm4 - DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 - DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,121,254,226 ; vpaddd %xmm10,%xmm0,%xmm4 - DB 196,193,249,126,226 ; vmovq %xmm4,%r10 - DB 196,227,249,22,227,1 ; vpextrq $0x1,%xmm4,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 - DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 - DB 196,227,89,33,230,32 ; vinsertps $0x20,%xmm6,%xmm4,%xmm4 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 + DB 196,163,81,33,44,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm5,%xmm5 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,81,33,238,32 ; vinsertps $0x20,%xmm6,%xmm5,%xmm5 DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,227,89,33,230,48 ; vinsertps $0x30,%xmm6,%xmm4,%xmm4 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,163,73,33,52,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm6,%xmm6 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 - DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 + DB 196,227,81,33,238,48 ; vinsertps $0x30,%xmm6,%xmm5,%xmm5 + DB 196,227,85,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm5,%ymm5 + DB 196,193,121,254,194 ; vpaddd %xmm10,%xmm0,%xmm0 + DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 196,193,113,254,194 ; vpaddd %xmm10,%xmm1,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 - DB 196,227,77,24,244,1 ; vinsertf128 $0x1,%xmm4,%ymm6,%ymm6 - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 - DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 - DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,12,176 ; vmovss (%rax,%r14,4),%xmm1 + DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,12,144 ; vmovss (%rax,%r10,4),%xmm1 + DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 + DB 196,163,113,33,12,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm1,%xmm1 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1 + DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 196,193,108,92,206 ; vsubps %ymm14,%ymm2,%ymm1 + DB 197,252,16,148,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm2 + DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 + DB 197,12,88,241 ; vaddps %ymm1,%ymm14,%ymm14 + DB 197,212,92,207 ; vsubps %ymm7,%ymm5,%ymm1 + DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 + DB 197,196,88,201 ; vaddps %ymm1,%ymm7,%ymm1 + DB 197,252,17,140,36,32,1,0,0 ; vmovups %ymm1,0x120(%rsp) + DB 196,193,124,92,192 ; vsubps %ymm8,%ymm0,%ymm0 + DB 197,236,89,192 ; vmulps %ymm0,%ymm2,%ymm0 + DB 197,60,88,192 ; vaddps %ymm0,%ymm8,%ymm8 + DB 197,129,254,68,36,80 ; vpaddd 0x50(%rsp),%xmm15,%xmm0 + DB 197,249,111,140,36,128,0,0,0 ; vmovdqa 0x80(%rsp),%xmm1 + DB 197,241,254,124,36,96 ; vpaddd 0x60(%rsp),%xmm1,%xmm7 + DB 197,249,254,204 ; vpaddd %xmm4,%xmm0,%xmm1 + DB 196,194,113,64,201 ; vpmulld %xmm9,%xmm1,%xmm1 + DB 197,121,111,124,36,16 ; vmovdqa 0x10(%rsp),%xmm15 + DB 196,193,65,254,215 ; vpaddd %xmm15,%xmm7,%xmm2 + DB 196,194,105,64,209 ; vpmulld %xmm9,%xmm2,%xmm2 + DB 196,193,121,126,209 ; vmovd %xmm2,%r9d + DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d + DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d + DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,28,184 ; vmovss (%rax,%r15,4),%xmm3 + DB 196,163,97,33,28,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm3,%xmm3 + DB 196,163,97,33,28,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm3,%xmm3 + DB 196,227,97,33,28,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm3,%xmm3 DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,99,101,24,232,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm13 - DB 197,161,254,194 ; vpaddd %xmm2,%xmm11,%xmm0 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 196,163,89,33,36,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm4,%xmm4 + DB 196,163,89,33,36,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm4,%xmm4 + DB 196,163,89,33,36,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm4,%xmm4 + DB 196,227,93,24,227,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm4 + DB 197,209,118,237 ; vpcmpeqd %xmm5,%xmm5,%xmm5 + DB 197,233,250,221 ; vpsubd %xmm5,%xmm2,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 197,241,250,221 ; vpsubd %xmm5,%xmm1,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,44,176 ; vmovss (%rax,%r14,4),%xmm5 + DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 197,249,111,108,36,32 ; vmovdqa 0x20(%rsp),%xmm5 - DB 197,209,254,201 ; vpaddd %xmm1,%xmm5,%xmm1 - DB 196,194,113,64,223 ; vpmulld %xmm15,%xmm1,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 - DB 196,193,97,254,210 ; vpaddd %xmm10,%xmm3,%xmm2 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 + DB 196,163,81,33,44,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm5,%xmm5 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,81,33,238,32 ; vinsertps $0x20,%xmm6,%xmm5,%xmm5 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,81,33,238,48 ; vinsertps $0x30,%xmm6,%xmm5,%xmm5 + DB 196,227,85,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm5,%ymm5 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,121,254,210 ; vpaddd %xmm10,%xmm0,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 - DB 196,227,105,33,212,32 ; vinsertps $0x20,%xmm4,%xmm2,%xmm2 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,227,105,33,212,48 ; vinsertps $0x30,%xmm4,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 - DB 196,163,89,33,36,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 - DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 196,227,93,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm4,%ymm2 - DB 196,193,97,254,217 ; vpaddd %xmm9,%xmm3,%xmm3 + DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,99,109,24,233,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm13 + DB 197,153,254,215 ; vpaddd %xmm7,%xmm12,%xmm2 + DB 197,161,254,192 ; vpaddd %xmm0,%xmm11,%xmm0 + DB 196,65,121,111,227 ; vmovdqa %xmm11,%xmm12 + DB 196,194,121,64,201 ; vpmulld %xmm9,%xmm0,%xmm1 + DB 196,194,105,64,209 ; vpmulld %xmm9,%xmm2,%xmm2 + DB 196,193,121,126,209 ; vmovd %xmm2,%r9d + DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d + DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d + DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d + DB 196,193,121,126,207 ; vmovd %xmm1,%r15d + DB 196,195,121,22,206,1 ; vpextrd $0x1,%xmm1,%r14d + DB 196,195,121,22,204,2 ; vpextrd $0x2,%xmm1,%r12d + DB 196,227,121,22,203,3 ; vpextrd $0x3,%xmm1,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 + DB 197,201,118,246 ; vpcmpeqd %xmm6,%xmm6,%xmm6 + DB 197,233,250,222 ; vpsubd %xmm6,%xmm2,%xmm3 DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 DB 196,193,249,126,217 ; vmovq %xmm3,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 197,241,250,222 ; vpsubd %xmm6,%xmm1,%xmm3 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 + DB 196,227,77,24,243,1 ; vinsertf128 $0x1,%xmm3,%ymm6,%ymm6 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 + DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 196,193,113,254,202 ; vpaddd %xmm10,%xmm1,%xmm1 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 + DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 + DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 + DB 73,193,232,32 ; shr $0x20,%r8 DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 - DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 - DB 196,227,101,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm0 - DB 196,193,116,92,206 ; vsubps %ymm14,%ymm1,%ymm1 - DB 197,252,16,156,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm3 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,140,88,201 ; vaddps %ymm1,%ymm14,%ymm1 - DB 197,236,92,214 ; vsubps %ymm6,%ymm2,%ymm2 - DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 - DB 197,204,88,210 ; vaddps %ymm2,%ymm6,%ymm2 - DB 196,193,124,92,197 ; vsubps %ymm13,%ymm0,%ymm0 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 197,252,16,156,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm3 DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,148,88,192 ; vaddps %ymm0,%ymm13,%ymm0 - DB 196,193,116,92,204 ; vsubps %ymm12,%ymm1,%ymm1 - DB 197,252,16,156,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm3 - DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,156,88,201 ; vaddps %ymm1,%ymm12,%ymm1 - DB 197,252,17,140,36,160,0,0,0 ; vmovups %ymm1,0xa0(%rsp) - DB 197,252,16,164,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm4 - DB 197,236,92,204 ; vsubps %ymm4,%ymm2,%ymm1 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,204,92,213 ; vsubps %ymm5,%ymm6,%ymm2 + DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 + DB 197,212,88,210 ; vaddps %ymm2,%ymm5,%ymm2 + DB 196,193,116,92,205 ; vsubps %ymm13,%ymm1,%ymm1 DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,17,140,36,64,1,0,0 ; vmovups %ymm1,0x140(%rsp) - DB 197,252,16,140,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm1 - DB 197,252,92,193 ; vsubps %ymm1,%ymm0,%ymm0 + DB 197,148,88,201 ; vaddps %ymm1,%ymm13,%ymm1 + DB 196,193,124,92,198 ; vsubps %ymm14,%ymm0,%ymm0 + DB 197,252,16,156,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm3 DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,244,88,192 ; vaddps %ymm0,%ymm1,%ymm0 + DB 197,140,88,192 ; vaddps %ymm0,%ymm14,%ymm0 + DB 197,252,17,132,36,128,0,0,0 ; vmovups %ymm0,0x80(%rsp) + DB 197,252,16,164,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm4 + DB 197,236,92,196 ; vsubps %ymm4,%ymm2,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 DB 197,252,17,132,36,32,1,0,0 ; vmovups %ymm0,0x120(%rsp) + DB 196,193,116,92,192 ; vsubps %ymm8,%ymm1,%ymm0 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 + DB 197,252,17,132,36,224,1,0,0 ; vmovups %ymm0,0x1e0(%rsp) DB 197,254,111,132,36,192,0,0,0 ; vmovdqu 0xc0(%rsp),%ymm0 - DB 197,121,254,116,36,96 ; vpaddd 0x60(%rsp),%xmm0,%xmm14 - DB 197,249,111,132,36,224,0,0,0 ; vmovdqa 0xe0(%rsp),%xmm0 - DB 197,121,254,172,36,224,1,0,0 ; vpaddd 0x1e0(%rsp),%xmm0,%xmm13 - DB 197,137,254,148,36,128,1,0,0 ; vpaddd 0x180(%rsp),%xmm14,%xmm2 - DB 197,233,254,4,36 ; vpaddd (%rsp),%xmm2,%xmm0 - DB 196,194,121,64,247 ; vpmulld %xmm15,%xmm0,%xmm6 - DB 196,193,249,126,241 ; vmovq %xmm6,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,243,1 ; vpextrq $0x1,%xmm6,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 197,121,254,116,36,64 ; vpaddd 0x40(%rsp),%xmm0,%xmm14 + DB 197,249,111,132,36,160,0,0,0 ; vmovdqa 0xa0(%rsp),%xmm0 + DB 197,121,254,172,36,160,1,0,0 ; vpaddd 0x1a0(%rsp),%xmm0,%xmm13 + DB 197,145,254,188,36,96,1,0,0 ; vpaddd 0x160(%rsp),%xmm13,%xmm7 + DB 197,137,254,108,36,112 ; vpaddd 0x70(%rsp),%xmm14,%xmm5 + DB 197,193,254,12,36 ; vpaddd (%rsp),%xmm7,%xmm1 + DB 196,194,113,64,209 ; vpmulld %xmm9,%xmm1,%xmm2 + DB 196,193,81,254,207 ; vpaddd %xmm15,%xmm5,%xmm1 + DB 196,194,113,64,225 ; vpmulld %xmm9,%xmm1,%xmm4 + DB 196,193,121,126,225 ; vmovd %xmm4,%r9d + DB 196,195,121,22,224,1 ; vpextrd $0x1,%xmm4,%r8d + DB 196,195,121,22,226,2 ; vpextrd $0x2,%xmm4,%r10d + DB 196,195,121,22,227,3 ; vpextrd $0x3,%xmm4,%r11d + DB 196,193,121,126,215 ; vmovd %xmm2,%r15d + DB 196,195,121,22,214,1 ; vpextrd $0x1,%xmm2,%r14d + DB 196,195,121,22,212,2 ; vpextrd $0x2,%xmm2,%r12d + DB 196,227,121,22,211,3 ; vpextrd $0x3,%xmm2,%ebx + DB 196,161,122,16,12,184 ; vmovss (%rax,%r15,4),%xmm1 + DB 196,163,113,33,12,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm1,%xmm1 + DB 196,163,113,33,12,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm1,%xmm1 + DB 196,227,113,33,12,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm3 + DB 197,249,118,192 ; vpcmpeqd %xmm0,%xmm0,%xmm0 + DB 197,217,250,200 ; vpsubd %xmm0,%xmm4,%xmm1 + DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 + DB 196,193,249,126,201 ; vmovq %xmm1,%r9 + DB 197,233,250,200 ; vpsubd %xmm0,%xmm2,%xmm1 + DB 196,65,1,118,255 ; vpcmpeqd %xmm15,%xmm15,%xmm15 + DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10 + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,12,152 ; vmovss (%rax,%r11,4),%xmm1 + DB 196,227,113,33,12,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 + DB 196,227,113,33,206,32 ; vinsertps $0x20,%xmm6,%xmm1,%xmm1 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,113,33,206,48 ; vinsertps $0x30,%xmm6,%xmm1,%xmm1 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 - DB 197,145,254,140,36,144,0,0,0 ; vpaddd 0x90(%rsp),%xmm13,%xmm1 - DB 196,193,113,254,216 ; vpaddd %xmm8,%xmm1,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,163,89,33,36,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm4,%xmm4 - DB 196,163,89,33,36,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,163,89,33,36,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm4,%xmm4 - DB 196,99,125,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm0,%ymm12 - DB 196,193,97,254,226 ; vpaddd %xmm10,%xmm3,%xmm4 - DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 - DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,73,254,226 ; vpaddd %xmm10,%xmm6,%xmm4 - DB 196,193,249,126,226 ; vmovq %xmm4,%r10 - DB 196,227,249,22,227,1 ; vpextrq $0x1,%xmm4,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 - DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 - DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,4,144 ; vmovss (%rax,%r10,4),%xmm0 + DB 196,227,73,33,192,32 ; vinsertps $0x20,%xmm0,%xmm6,%xmm0 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,4,136 ; vmovss (%rax,%r9,4),%xmm0 - DB 196,227,65,33,192,32 ; vinsertps $0x20,%xmm0,%xmm7,%xmm0 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,121,33,199,48 ; vinsertps $0x30,%xmm7,%xmm0,%xmm0 - DB 196,227,125,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm0,%ymm4 - DB 196,193,97,254,193 ; vpaddd %xmm9,%xmm3,%xmm0 + DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 + DB 196,227,121,33,198,48 ; vinsertps $0x30,%xmm6,%xmm0,%xmm0 + DB 196,227,125,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm1 + DB 196,193,89,254,194 ; vpaddd %xmm10,%xmm4,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,73,254,193 ; vpaddd %xmm9,%xmm6,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,105,254,194 ; vpaddd %xmm10,%xmm2,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 - DB 196,227,121,33,195,32 ; vinsertps $0x20,%xmm3,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 + DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 - DB 196,99,101,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm3,%ymm11 - DB 197,233,254,68,36,16 ; vpaddd 0x10(%rsp),%xmm2,%xmm0 - DB 196,194,121,64,215 ; vpmulld %xmm15,%xmm0,%xmm2 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,161,122,16,20,144 ; vmovss (%rax,%r10,4),%xmm2 + DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,4,128 ; vmovss (%rax,%r8,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,163,121,33,4,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm0,%xmm0 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,121,33,195,48 ; vinsertps $0x30,%xmm3,%xmm0,%xmm0 - DB 197,209,254,201 ; vpaddd %xmm1,%xmm5,%xmm1 - DB 196,194,113,64,223 ; vpmulld %xmm15,%xmm1,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,12,152 ; vmovss (%rax,%rbx,4),%xmm1 - DB 196,163,113,33,12,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm1,%xmm1 - DB 196,163,113,33,12,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm1,%xmm1 - DB 196,163,113,33,12,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,227,125,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm1 - DB 196,193,97,254,194 ; vpaddd %xmm10,%xmm3,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 + DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,105,33,212,32 ; vinsertps $0x20,%xmm4,%xmm2,%xmm2 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 + DB 196,227,105,33,212,48 ; vinsertps $0x30,%xmm4,%xmm2,%xmm2 + DB 196,99,109,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm11 + DB 197,121,111,68,36,32 ; vmovdqa 0x20(%rsp),%xmm8 + DB 197,185,254,197 ; vpaddd %xmm5,%xmm8,%xmm0 + DB 197,153,254,215 ; vpaddd %xmm7,%xmm12,%xmm2 + DB 196,194,105,64,233 ; vpmulld %xmm9,%xmm2,%xmm5 + DB 196,194,121,64,225 ; vpmulld %xmm9,%xmm0,%xmm4 + DB 196,193,121,126,225 ; vmovd %xmm4,%r9d + DB 196,195,121,22,224,1 ; vpextrd $0x1,%xmm4,%r8d + DB 196,195,121,22,226,2 ; vpextrd $0x2,%xmm4,%r10d + DB 196,195,121,22,227,3 ; vpextrd $0x3,%xmm4,%r11d + DB 196,193,121,126,239 ; vmovd %xmm5,%r15d + DB 196,195,121,22,238,1 ; vpextrd $0x1,%xmm5,%r14d + DB 196,195,121,22,236,2 ; vpextrd $0x2,%xmm5,%r12d + DB 196,227,121,22,235,3 ; vpextrd $0x3,%xmm5,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 + DB 196,163,105,33,20,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm2,%xmm2 + DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 + DB 196,163,105,33,20,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm2,%xmm2 + DB 196,227,109,24,208,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm2 + DB 196,193,89,250,199 ; vpsubd %xmm15,%xmm4,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,194 ; vpaddd %xmm10,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,81,250,199 ; vpsubd %xmm15,%xmm5,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 DB 196,161,122,16,52,176 ; vmovss (%rax,%r14,4),%xmm6 DB 196,227,121,33,198,32 ; vinsertps $0x20,%xmm6,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,227,121,33,198,48 ; vinsertps $0x30,%xmm6,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,52,128 ; vmovss (%rax,%r8,4),%xmm6 - DB 196,163,73,33,52,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm6,%xmm6 - DB 196,161,122,16,60,136 ; vmovss (%rax,%r9,4),%xmm7 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,121,33,198,48 ; vinsertps $0x30,%xmm6,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 + DB 196,163,73,33,52,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm6,%xmm6 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 DB 196,227,73,33,247,32 ; vinsertps $0x20,%xmm7,%xmm6,%xmm6 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 DB 196,227,77,24,240,1 ; vinsertf128 $0x1,%xmm0,%ymm6,%ymm6 - DB 196,193,97,254,193 ; vpaddd %xmm9,%xmm3,%xmm0 + DB 196,193,89,254,194 ; vpaddd %xmm10,%xmm4,%xmm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 196,193,105,254,193 ; vpaddd %xmm9,%xmm2,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,81,254,194 ; vpaddd %xmm10,%xmm5,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2 - DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d + DB 196,227,121,33,4,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,36,176 ; vmovss (%rax,%r14,4),%xmm4 + DB 196,227,121,33,196,32 ; vinsertps $0x20,%xmm4,%xmm0,%xmm0 DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm2,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 - DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0 - DB 196,193,116,92,204 ; vsubps %ymm12,%ymm1,%ymm1 - DB 197,252,16,148,36,32,2,0,0 ; vmovups 0x220(%rsp),%ymm2 - DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 - DB 197,156,88,201 ; vaddps %ymm1,%ymm12,%ymm1 - DB 197,252,17,140,36,128,1,0,0 ; vmovups %ymm1,0x180(%rsp) - DB 197,204,92,204 ; vsubps %ymm4,%ymm6,%ymm1 - DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 - DB 197,92,88,225 ; vaddps %ymm1,%ymm4,%ymm12 - DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 - DB 197,236,89,192 ; vmulps %ymm0,%ymm2,%ymm0 - DB 197,252,40,234 ; vmovaps %ymm2,%ymm5 - DB 197,36,88,216 ; vaddps %ymm0,%ymm11,%ymm11 - DB 197,145,254,76,36,112 ; vpaddd 0x70(%rsp),%xmm13,%xmm1 - DB 197,137,254,132,36,128,0,0,0 ; vpaddd 0x80(%rsp),%xmm14,%xmm0 - DB 197,249,254,28,36 ; vpaddd (%rsp),%xmm0,%xmm3 - DB 196,66,97,64,247 ; vpmulld %xmm15,%xmm3,%xmm14 - DB 196,65,249,126,241 ; vmovq %xmm14,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,99,249,22,243,1 ; vpextrq $0x1,%xmm14,%rbx - DB 65,137,218 ; mov %ebx,%r10d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,121,33,196,48 ; vinsertps $0x30,%xmm4,%xmm0,%xmm0 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 69,137,194 ; mov %r8d,%r10d DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 - DB 196,227,97,33,228,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm4 - DB 196,193,113,254,216 ; vpaddd %xmm8,%xmm1,%xmm3 - DB 196,194,97,64,223 ; vpmulld %xmm15,%xmm3,%xmm3 - DB 197,249,126,219 ; vmovd %xmm3,%ebx - DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d - DB 196,195,121,22,217,2 ; vpextrd $0x2,%xmm3,%r9d - DB 196,195,121,22,218,3 ; vpextrd $0x3,%xmm3,%r10d - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,163,73,33,52,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm6,%xmm6 - DB 196,163,73,33,52,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm6,%xmm6 - DB 196,163,73,33,52,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm6,%xmm6 - DB 196,99,93,24,198,1 ; vinsertf128 $0x1,%xmm6,%ymm4,%ymm8 - DB 196,193,97,254,226 ; vpaddd %xmm10,%xmm3,%xmm4 - DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 - DB 196,193,249,126,225 ; vmovq %xmm4,%r9 - DB 196,193,9,254,226 ; vpaddd %xmm10,%xmm14,%xmm4 - DB 196,193,249,126,226 ; vmovq %xmm4,%r10 - DB 196,227,249,22,227,1 ; vpextrq $0x1,%xmm4,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 DB 196,163,89,33,36,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm4,%xmm4 - DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 - DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,89,33,229,32 ; vinsertps $0x20,%xmm5,%xmm4,%xmm4 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 - DB 196,163,65,33,60,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm7,%xmm7 - DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 - DB 196,227,65,33,210,32 ; vinsertps $0x20,%xmm2,%xmm7,%xmm2 + DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 + DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 + DB 196,227,93,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm4,%ymm0 + DB 197,236,92,211 ; vsubps %ymm3,%ymm2,%ymm2 + DB 197,252,16,164,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm4 + DB 197,220,89,210 ; vmulps %ymm2,%ymm4,%ymm2 + DB 197,228,88,210 ; vaddps %ymm2,%ymm3,%ymm2 + DB 197,252,17,148,36,96,1,0,0 ; vmovups %ymm2,0x160(%rsp) + DB 197,204,92,209 ; vsubps %ymm1,%ymm6,%ymm2 + DB 197,220,89,210 ; vmulps %ymm2,%ymm4,%ymm2 + DB 197,116,88,250 ; vaddps %ymm2,%ymm1,%ymm15 + DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 + DB 197,220,89,192 ; vmulps %ymm0,%ymm4,%ymm0 + DB 197,124,40,228 ; vmovaps %ymm4,%ymm12 + DB 197,36,88,216 ; vaddps %ymm0,%ymm11,%ymm11 + DB 197,145,254,124,36,80 ; vpaddd 0x50(%rsp),%xmm13,%xmm7 + DB 197,137,254,68,36,96 ; vpaddd 0x60(%rsp),%xmm14,%xmm0 + DB 197,249,254,76,36,16 ; vpaddd 0x10(%rsp),%xmm0,%xmm1 + DB 197,193,254,20,36 ; vpaddd (%rsp),%xmm7,%xmm2 + DB 196,66,105,64,241 ; vpmulld %xmm9,%xmm2,%xmm14 + DB 196,194,113,64,209 ; vpmulld %xmm9,%xmm1,%xmm2 + DB 196,193,121,126,209 ; vmovd %xmm2,%r9d + DB 196,195,121,22,208,1 ; vpextrd $0x1,%xmm2,%r8d + DB 196,195,121,22,210,2 ; vpextrd $0x2,%xmm2,%r10d + DB 196,195,121,22,211,3 ; vpextrd $0x3,%xmm2,%r11d + DB 196,65,121,126,247 ; vmovd %xmm14,%r15d + DB 196,67,121,22,246,1 ; vpextrd $0x1,%xmm14,%r14d + DB 196,67,121,22,244,2 ; vpextrd $0x2,%xmm14,%r12d + DB 196,99,121,22,243,3 ; vpextrd $0x3,%xmm14,%ebx + DB 196,161,122,16,12,184 ; vmovss (%rax,%r15,4),%xmm1 + DB 196,163,113,33,12,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm1,%xmm1 + DB 196,163,113,33,12,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm1,%xmm1 + DB 196,227,113,33,12,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm1,%xmm1 + DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm3,%xmm3 + DB 196,163,97,33,28,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm3,%xmm3 + DB 196,227,101,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm3,%ymm1 + DB 197,217,118,228 ; vpcmpeqd %xmm4,%xmm4,%xmm4 + DB 197,233,250,220 ; vpsubd %xmm4,%xmm2,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 197,137,250,220 ; vpsubd %xmm4,%xmm14,%xmm3 + DB 197,201,118,246 ; vpcmpeqd %xmm6,%xmm6,%xmm6 + DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10 + DB 196,225,249,126,219 ; vmovq %xmm3,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 - DB 196,227,105,33,215,48 ; vinsertps $0x30,%xmm7,%xmm2,%xmm2 - DB 196,227,109,24,252,1 ; vinsertf128 $0x1,%xmm4,%ymm2,%ymm7 - DB 196,193,97,254,209 ; vpaddd %xmm9,%xmm3,%xmm2 + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,28,152 ; vmovss (%rax,%r11,4),%xmm3 + DB 196,227,97,33,28,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm3,%xmm3 + DB 196,161,122,16,44,176 ; vmovss (%rax,%r14,4),%xmm5 + DB 196,227,97,33,221,32 ; vinsertps $0x20,%xmm5,%xmm3,%xmm3 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,44,144 ; vmovss (%rax,%r10,4),%xmm5 + DB 196,227,97,33,221,48 ; vinsertps $0x30,%xmm5,%xmm3,%xmm3 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,44,152 ; vmovss (%rax,%rbx,4),%xmm5 + DB 196,163,81,33,44,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm5,%xmm5 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,81,33,228,32 ; vinsertps $0x20,%xmm4,%xmm5,%xmm4 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,44,128 ; vmovss (%rax,%r8,4),%xmm5 + DB 196,227,89,33,229,48 ; vinsertps $0x30,%xmm5,%xmm4,%xmm4 + DB 196,227,93,24,235,1 ; vinsertf128 $0x1,%xmm3,%ymm4,%ymm5 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,9,254,209 ; vpaddd %xmm9,%xmm14,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 196,193,9,254,210 ; vpaddd %xmm10,%xmm14,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx + DB 73,193,233,32 ; shr $0x20,%r9 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,36,152 ; vmovss (%rax,%rbx,4),%xmm4 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 DB 196,99,101,24,242,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm14 - DB 197,249,254,68,36,16 ; vpaddd 0x10(%rsp),%xmm0,%xmm0 - DB 197,241,254,76,36,32 ; vpaddd 0x20(%rsp),%xmm1,%xmm1 - DB 196,194,113,64,207 ; vpmulld %xmm15,%xmm1,%xmm1 - DB 196,194,121,64,199 ; vpmulld %xmm15,%xmm0,%xmm0 - DB 196,193,249,126,193 ; vmovq %xmm0,%r9 - DB 69,137,200 ; mov %r9d,%r8d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,218 ; mov %ebx,%r10d + DB 197,185,254,192 ; vpaddd %xmm0,%xmm8,%xmm0 + DB 197,193,254,84,36,48 ; vpaddd 0x30(%rsp),%xmm7,%xmm2 + DB 196,194,105,64,209 ; vpmulld %xmm9,%xmm2,%xmm2 + DB 196,194,121,64,217 ; vpmulld %xmm9,%xmm0,%xmm3 + DB 196,193,121,126,217 ; vmovd %xmm3,%r9d + DB 196,195,121,22,216,1 ; vpextrd $0x1,%xmm3,%r8d + DB 196,195,121,22,218,2 ; vpextrd $0x2,%xmm3,%r10d + DB 196,195,121,22,219,3 ; vpextrd $0x3,%xmm3,%r11d + DB 196,193,121,126,215 ; vmovd %xmm2,%r15d + DB 196,195,121,22,214,1 ; vpextrd $0x1,%xmm2,%r14d + DB 196,195,121,22,212,2 ; vpextrd $0x2,%xmm2,%r12d + DB 196,227,121,22,211,3 ; vpextrd $0x3,%xmm2,%ebx + DB 196,161,122,16,4,184 ; vmovss (%rax,%r15,4),%xmm0 + DB 196,163,121,33,4,176,16 ; vinsertps $0x10,(%rax,%r14,4),%xmm0,%xmm0 + DB 196,163,121,33,4,160,32 ; vinsertps $0x20,(%rax,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,4,152,48 ; vinsertps $0x30,(%rax,%rbx,4),%xmm0,%xmm0 + DB 196,161,122,16,36,136 ; vmovss (%rax,%r9,4),%xmm4 + DB 196,163,89,33,36,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm4,%xmm4 + DB 196,163,89,33,36,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm4,%xmm4 + DB 196,163,89,33,36,152,48 ; vinsertps $0x30,(%rax,%r11,4),%xmm4,%xmm4 + DB 196,227,93,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm4,%ymm0 + DB 197,225,250,230 ; vpsubd %xmm6,%xmm3,%xmm4 + DB 196,195,249,22,224,1 ; vpextrq $0x1,%xmm4,%r8 + DB 196,193,249,126,225 ; vmovq %xmm4,%r9 + DB 197,233,250,230 ; vpsubd %xmm6,%xmm2,%xmm4 + DB 196,195,249,22,226,1 ; vpextrq $0x1,%xmm4,%r10 + DB 196,225,249,126,227 ; vmovq %xmm4,%rbx + DB 65,137,219 ; mov %ebx,%r11d DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d + DB 196,161,122,16,36,152 ; vmovss (%rax,%r11,4),%xmm4 + DB 196,227,89,33,36,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm4,%xmm4 + DB 196,161,122,16,60,176 ; vmovss (%rax,%r14,4),%xmm7 + DB 196,227,89,33,231,32 ; vinsertps $0x20,%xmm7,%xmm4,%xmm4 + DB 73,193,234,32 ; shr $0x20,%r10 + DB 196,161,122,16,60,144 ; vmovss (%rax,%r10,4),%xmm7 + DB 196,227,89,33,231,48 ; vinsertps $0x30,%xmm7,%xmm4,%xmm4 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 196,161,122,16,20,128 ; vmovss (%rax,%r8,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 - DB 196,163,105,33,20,144,32 ; vinsertps $0x20,(%rax,%r10,4),%xmm2,%xmm2 - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 197,249,126,203 ; vmovd %xmm1,%ebx - DB 196,195,121,22,200,1 ; vpextrd $0x1,%xmm1,%r8d - DB 196,195,121,22,201,2 ; vpextrd $0x2,%xmm1,%r9d - DB 196,195,121,22,202,3 ; vpextrd $0x3,%xmm1,%r10d - DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 - DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 - DB 196,163,97,33,28,136,32 ; vinsertps $0x20,(%rax,%r9,4),%xmm3,%xmm3 - DB 196,163,97,33,28,144,48 ; vinsertps $0x30,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,227,109,24,227,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm4 - DB 196,193,113,254,210 ; vpaddd %xmm10,%xmm1,%xmm2 - DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8 - DB 196,193,249,126,209 ; vmovq %xmm2,%r9 - DB 196,193,121,254,210 ; vpaddd %xmm10,%xmm0,%xmm2 - DB 196,193,249,126,210 ; vmovq %xmm2,%r10 - DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx - DB 69,137,203 ; mov %r9d,%r11d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,60,152 ; vmovss (%rax,%rbx,4),%xmm7 + DB 196,163,65,33,60,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm7,%xmm7 + DB 196,161,122,16,52,144 ; vmovss (%rax,%r10,4),%xmm6 + DB 196,227,65,33,246,32 ; vinsertps $0x20,%xmm6,%xmm7,%xmm6 + DB 73,193,232,32 ; shr $0x20,%r8 + DB 196,161,122,16,60,128 ; vmovss (%rax,%r8,4),%xmm7 + DB 196,227,73,33,247,48 ; vinsertps $0x30,%xmm7,%xmm6,%xmm6 + DB 196,227,77,24,252,1 ; vinsertf128 $0x1,%xmm4,%ymm6,%ymm7 + DB 196,193,97,254,218 ; vpaddd %xmm10,%xmm3,%xmm3 + DB 196,195,249,22,216,1 ; vpextrq $0x1,%xmm3,%r8 + DB 196,193,249,126,217 ; vmovq %xmm3,%r9 + DB 196,193,105,254,210 ; vpaddd %xmm10,%xmm2,%xmm2 + DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10 + DB 196,225,249,126,211 ; vmovq %xmm2,%rbx + DB 65,137,219 ; mov %ebx,%r11d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 69,137,214 ; mov %r10d,%r14d DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 - DB 196,163,105,33,20,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm2,%xmm2 + DB 196,227,105,33,20,152,16 ; vinsertps $0x10,(%rax,%rbx,4),%xmm2,%xmm2 DB 196,161,122,16,28,176 ; vmovss (%rax,%r14,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,28,128 ; vmovss (%rax,%r8,4),%xmm3 - DB 196,163,97,33,28,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm3,%xmm3 - DB 196,161,122,16,52,136 ; vmovss (%rax,%r9,4),%xmm6 - DB 196,227,97,33,222,32 ; vinsertps $0x20,%xmm6,%xmm3,%xmm3 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,52,152 ; vmovss (%rax,%rbx,4),%xmm6 - DB 196,227,97,33,222,48 ; vinsertps $0x30,%xmm6,%xmm3,%xmm3 - DB 196,227,101,24,218,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm3 - DB 196,193,113,254,201 ; vpaddd %xmm9,%xmm1,%xmm1 - DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 - DB 196,193,249,126,201 ; vmovq %xmm1,%r9 - DB 196,193,121,254,193 ; vpaddd %xmm9,%xmm0,%xmm0 - DB 196,193,249,126,194 ; vmovq %xmm0,%r10 - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 69,137,203 ; mov %r9d,%r11d + DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 + DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 + DB 68,137,203 ; mov %r9d,%ebx DB 73,193,233,32 ; shr $0x20,%r9 - DB 69,137,198 ; mov %r8d,%r14d - DB 196,161,122,16,4,152 ; vmovss (%rax,%r11,4),%xmm0 - DB 196,163,121,33,4,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm0,%xmm0 - DB 196,161,122,16,12,176 ; vmovss (%rax,%r14,4),%xmm1 - DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0 + DB 69,137,194 ; mov %r8d,%r10d + DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 + DB 196,161,122,16,36,144 ; vmovss (%rax,%r10,4),%xmm4 + DB 196,227,97,33,220,32 ; vinsertps $0x20,%xmm4,%xmm3,%xmm3 DB 73,193,232,32 ; shr $0x20,%r8 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0 - DB 69,137,208 ; mov %r10d,%r8d - DB 65,137,217 ; mov %ebx,%r9d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 196,161,122,16,12,128 ; vmovss (%rax,%r8,4),%xmm1 - DB 196,163,113,33,12,144,16 ; vinsertps $0x10,(%rax,%r10,4),%xmm1,%xmm1 - DB 196,161,122,16,20,136 ; vmovss (%rax,%r9,4),%xmm2 - DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 197,250,16,20,152 ; vmovss (%rax,%rbx,4),%xmm2 - DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 196,193,92,92,200 ; vsubps %ymm8,%ymm4,%ymm1 - DB 197,212,89,201 ; vmulps %ymm1,%ymm5,%ymm1 - DB 197,188,88,201 ; vaddps %ymm1,%ymm8,%ymm1 - DB 197,228,92,215 ; vsubps %ymm7,%ymm3,%ymm2 - DB 197,212,89,210 ; vmulps %ymm2,%ymm5,%ymm2 - DB 197,196,88,210 ; vaddps %ymm2,%ymm7,%ymm2 - DB 196,193,124,92,198 ; vsubps %ymm14,%ymm0,%ymm0 - DB 197,212,89,192 ; vmulps %ymm0,%ymm5,%ymm0 - DB 197,140,88,192 ; vaddps %ymm0,%ymm14,%ymm0 - DB 197,252,16,164,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm4 - DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 - DB 197,252,16,156,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm3 + DB 196,161,122,16,36,128 ; vmovss (%rax,%r8,4),%xmm4 + DB 196,227,97,33,220,48 ; vinsertps $0x30,%xmm4,%xmm3,%xmm3 + DB 196,227,101,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 + DB 197,252,92,193 ; vsubps %ymm1,%ymm0,%ymm0 + DB 197,156,89,192 ; vmulps %ymm0,%ymm12,%ymm0 + DB 197,244,88,192 ; vaddps %ymm0,%ymm1,%ymm0 + DB 197,196,92,205 ; vsubps %ymm5,%ymm7,%ymm1 + DB 197,156,89,201 ; vmulps %ymm1,%ymm12,%ymm1 + DB 197,212,88,201 ; vaddps %ymm1,%ymm5,%ymm1 + DB 196,193,108,92,214 ; vsubps %ymm14,%ymm2,%ymm2 + DB 197,156,89,210 ; vmulps %ymm2,%ymm12,%ymm2 + DB 197,140,88,210 ; vaddps %ymm2,%ymm14,%ymm2 + DB 197,252,16,164,36,96,1,0,0 ; vmovups 0x160(%rsp),%ymm4 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 197,252,16,156,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm3 + DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 196,193,116,92,207 ; vsubps %ymm15,%ymm1,%ymm1 DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 - DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 196,193,108,92,212 ; vsubps %ymm12,%ymm2,%ymm2 + DB 197,132,88,201 ; vaddps %ymm1,%ymm15,%ymm1 + DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 - DB 197,156,88,210 ; vaddps %ymm2,%ymm12,%ymm2 - DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 + DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2 + DB 197,252,16,164,36,128,0,0,0 ; vmovups 0x80(%rsp),%ymm4 + DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 + DB 197,252,16,156,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm3 DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0 - DB 197,252,16,164,36,160,0,0,0 ; vmovups 0xa0(%rsp),%ymm4 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,16,164,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm4 DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 - DB 197,252,16,92,36,64 ; vmovups 0x40(%rsp),%ymm3 DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 - DB 197,252,16,164,36,64,1,0,0 ; vmovups 0x140(%rsp),%ymm4 + DB 197,252,16,164,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm4 DB 197,236,92,212 ; vsubps %ymm4,%ymm2,%ymm2 DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 DB 197,220,88,210 ; vaddps %ymm2,%ymm4,%ymm2 - DB 197,252,16,164,36,32,1,0,0 ; vmovups 0x120(%rsp),%ymm4 + DB 197,252,91,156,36,64,2,0,0 ; vcvtdq2ps 0x240(%rsp),%ymm3 + DB 197,252,16,164,36,96,2,0,0 ; vmovups 0x260(%rsp),%ymm4 + DB 197,220,92,219 ; vsubps %ymm3,%ymm4,%ymm3 + DB 197,252,16,164,36,128,1,0,0 ; vmovups 0x180(%rsp),%ymm4 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 197,220,88,216 ; vaddps %ymm0,%ymm4,%ymm3 - DB 197,252,91,132,36,96,2,0,0 ; vcvtdq2ps 0x260(%rsp),%ymm0 - DB 197,252,16,164,36,128,2,0,0 ; vmovups 0x280(%rsp),%ymm4 - DB 197,220,92,224 ; vsubps %ymm0,%ymm4,%ymm4 - DB 197,252,16,172,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm5 - DB 197,244,92,197 ; vsubps %ymm5,%ymm1,%ymm0 - DB 197,220,89,192 ; vmulps %ymm0,%ymm4,%ymm0 - DB 197,212,88,192 ; vaddps %ymm0,%ymm5,%ymm0 - DB 197,252,16,172,36,0,2,0,0 ; vmovups 0x200(%rsp),%ymm5 - DB 197,236,92,205 ; vsubps %ymm5,%ymm2,%ymm1 - DB 197,220,89,201 ; vmulps %ymm1,%ymm4,%ymm1 - DB 197,212,88,201 ; vaddps %ymm1,%ymm5,%ymm1 - DB 197,252,16,172,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm5 - DB 197,228,92,213 ; vsubps %ymm5,%ymm3,%ymm2 - DB 197,220,89,210 ; vmulps %ymm2,%ymm4,%ymm2 - DB 197,212,88,210 ; vaddps %ymm2,%ymm5,%ymm2 + DB 197,220,88,192 ; vaddps %ymm0,%ymm4,%ymm0 + DB 197,252,16,164,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm4 + DB 197,244,92,204 ; vsubps %ymm4,%ymm1,%ymm1 + DB 197,228,89,201 ; vmulps %ymm1,%ymm3,%ymm1 + DB 197,220,88,201 ; vaddps %ymm1,%ymm4,%ymm1 + DB 197,252,16,164,36,0,1,0,0 ; vmovups 0x100(%rsp),%ymm4 + DB 197,236,92,212 ; vsubps %ymm4,%ymm2,%ymm2 + DB 197,228,89,210 ; vmulps %ymm2,%ymm3,%ymm2 + DB 197,220,88,210 ; vaddps %ymm2,%ymm4,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,171,169,1,0 ; vbroadcastss 0x1a9ab(%rip),%ymm3 # 2d2ac <_sk_overlay_sse2_8bit+0xd54> - DB 197,252,16,164,36,160,2,0,0 ; vmovups 0x2a0(%rsp),%ymm4 - DB 197,252,16,172,36,192,2,0,0 ; vmovups 0x2c0(%rsp),%ymm5 - DB 197,252,16,180,36,224,2,0,0 ; vmovups 0x2e0(%rsp),%ymm6 - DB 197,252,16,188,36,0,3,0,0 ; vmovups 0x300(%rsp),%ymm7 - DB 72,129,196,40,3,0,0 ; add $0x328,%rsp + DB 196,226,125,24,29,16,165,1,0 ; vbroadcastss 0x1a510(%rip),%ymm3 # 2d18c <_sk_overlay_sse2_8bit+0xcdd> + DB 197,252,16,164,36,192,2,0,0 ; vmovups 0x2c0(%rsp),%ymm4 + DB 197,252,16,172,36,224,2,0,0 ; vmovups 0x2e0(%rsp),%ymm5 + DB 197,252,16,180,36,0,3,0,0 ; vmovups 0x300(%rsp),%ymm6 + DB 197,252,16,188,36,32,3,0,0 ; vmovups 0x320(%rsp),%ymm7 + DB 72,129,196,88,3,0,0 ; add $0x358,%rsp DB 91 ; pop %rbx + DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax PUBLIC _sk_start_pipeline_sse41 @@ -17021,13 +17157,13 @@ _sk_start_pipeline_sse41 LABEL PROC DB 73,137,246 ; mov %rsi,%r14 DB 76,137,141,8,255,255,255 ; mov %r9,-0xf8(%rbp) DB 76,57,203 ; cmp %r9,%rbx - DB 115,101 ; jae 12a13 <_sk_start_pipeline_sse41+0xe2> + DB 115,101 ; jae 12d92 <_sk_start_pipeline_sse41+0xe2> DB 72,139,133,24,255,255,255 ; mov -0xe8(%rbp),%rax DB 72,141,64,4 ; lea 0x4(%rax),%rax DB 72,137,133,16,255,255,255 ; mov %rax,-0xf0(%rbp) DB 76,57,189,16,255,255,255 ; cmp %r15,-0xf0(%rbp) DB 72,139,149,24,255,255,255 ; mov -0xe8(%rbp),%rdx - DB 119,38 ; ja 129f6 <_sk_start_pipeline_sse41+0xc5> + DB 119,38 ; ja 12d75 <_sk_start_pipeline_sse41+0xc5> DB 76,139,165,24,255,255,255 ; mov -0xe8(%rbp),%r12 DB 49,255 ; xor %edi,%edi DB 76,137,246 ; mov %r14,%rsi @@ -17038,16 +17174,16 @@ _sk_start_pipeline_sse41 LABEL PROC DB 73,131,196,8 ; add $0x8,%r12 DB 77,57,252 ; cmp %r15,%r12 DB 73,137,212 ; mov %rdx,%r12 - DB 118,225 ; jbe 129d7 <_sk_start_pipeline_sse41+0xa6> + DB 118,225 ; jbe 12d56 <_sk_start_pipeline_sse41+0xa6> DB 76,137,255 ; mov %r15,%rdi DB 72,41,215 ; sub %rdx,%rdi - DB 116,9 ; je 12a07 <_sk_start_pipeline_sse41+0xd6> + DB 116,9 ; je 12d86 <_sk_start_pipeline_sse41+0xd6> DB 76,137,246 ; mov %r14,%rsi DB 72,137,217 ; mov %rbx,%rcx DB 65,255,213 ; callq *%r13 DB 72,255,195 ; inc %rbx DB 72,59,157,8,255,255,255 ; cmp -0xf8(%rbp),%rbx - DB 117,173 ; jne 129c0 <_sk_start_pipeline_sse41+0x8f> + DB 117,173 ; jne 12d3f <_sk_start_pipeline_sse41+0x8f> DB 15,40,181,32,255,255,255 ; movaps -0xe0(%rbp),%xmm6 DB 15,40,189,48,255,255,255 ; movaps -0xd0(%rbp),%xmm7 DB 68,15,40,133,64,255,255,255 ; movaps -0xc0(%rbp),%xmm8 @@ -17078,13 +17214,13 @@ _sk_seed_shader_sse41 LABEL PROC DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,88,5,35,170,1,0 ; addps 0x1aa23(%rip),%xmm0 # 2d4a0 <_sk_overlay_sse2_8bit+0xf48> + DB 15,88,5,148,165,1,0 ; addps 0x1a594(%rip),%xmm0 # 2d390 <_sk_overlay_sse2_8bit+0xee1> DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,88,13,32,170,1,0 ; addps 0x1aa20(%rip),%xmm1 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,88,13,145,165,1,0 ; addps 0x1a591(%rip),%xmm1 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,39,170,1,0 ; movaps 0x1aa27(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,152,165,1,0 ; movaps 0x1a598(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -17096,36 +17232,36 @@ PUBLIC _sk_dither_sse41 _sk_dither_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 102,68,15,110,194 ; movd %edx,%xmm8 - DB 102,69,15,112,200,0 ; pshufd $0x0,%xmm8,%xmm9 - DB 102,68,15,254,13,192,169,1,0 ; paddd 0x1a9c0(%rip),%xmm9 # 2d480 <_sk_overlay_sse2_8bit+0xf28> - DB 102,68,15,110,193 ; movd %ecx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 - DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 - DB 102,68,15,111,21,247,169,1,0 ; movdqa 0x1a9f7(%rip),%xmm10 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 + DB 102,68,15,254,5,49,165,1,0 ; paddd 0x1a531(%rip),%xmm8 # 2d370 <_sk_overlay_sse2_8bit+0xec1> + DB 102,68,15,110,201 ; movd %ecx,%xmm9 + DB 102,69,15,112,201,0 ; pshufd $0x0,%xmm9,%xmm9 + DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 + DB 102,68,15,111,21,104,165,1,0 ; movdqa 0x1a568(%rip),%xmm10 # 2d3c0 <_sk_overlay_sse2_8bit+0xf11> + DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,65,15,114,243,5 ; pslld $0x5,%xmm11 - DB 102,69,15,219,209 ; pand %xmm9,%xmm10 + DB 102,69,15,219,208 ; pand %xmm8,%xmm10 DB 102,65,15,114,242,4 ; pslld $0x4,%xmm10 - DB 102,68,15,111,37,227,169,1,0 ; movdqa 0x1a9e3(%rip),%xmm12 # 2d4e0 <_sk_overlay_sse2_8bit+0xf88> - DB 102,68,15,111,45,234,169,1,0 ; movdqa 0x1a9ea(%rip),%xmm13 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 + DB 102,68,15,111,37,84,165,1,0 ; movdqa 0x1a554(%rip),%xmm12 # 2d3d0 <_sk_overlay_sse2_8bit+0xf21> + DB 102,68,15,111,45,91,165,1,0 ; movdqa 0x1a55b(%rip),%xmm13 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,69,15,111,241 ; movdqa %xmm9,%xmm14 DB 102,69,15,219,245 ; pand %xmm13,%xmm14 DB 102,65,15,114,246,2 ; pslld $0x2,%xmm14 - DB 102,69,15,219,233 ; pand %xmm9,%xmm13 + DB 102,69,15,235,243 ; por %xmm11,%xmm14 + DB 102,69,15,219,232 ; pand %xmm8,%xmm13 DB 102,69,15,254,237 ; paddd %xmm13,%xmm13 - DB 102,69,15,219,196 ; pand %xmm12,%xmm8 - DB 102,65,15,114,208,1 ; psrld $0x1,%xmm8 - DB 102,69,15,219,204 ; pand %xmm12,%xmm9 - DB 102,65,15,114,209,2 ; psrld $0x2,%xmm9 DB 102,69,15,235,234 ; por %xmm10,%xmm13 - DB 102,69,15,235,233 ; por %xmm9,%xmm13 - DB 102,69,15,235,243 ; por %xmm11,%xmm14 - DB 102,69,15,235,245 ; por %xmm13,%xmm14 - DB 102,69,15,235,240 ; por %xmm8,%xmm14 - DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8 - DB 68,15,89,5,165,169,1,0 ; mulps 0x1a9a5(%rip),%xmm8 # 2d500 <_sk_overlay_sse2_8bit+0xfa8> - DB 68,15,88,5,173,169,1,0 ; addps 0x1a9ad(%rip),%xmm8 # 2d510 <_sk_overlay_sse2_8bit+0xfb8> + DB 102,69,15,219,204 ; pand %xmm12,%xmm9 + DB 102,65,15,114,209,1 ; psrld $0x1,%xmm9 + DB 102,69,15,219,196 ; pand %xmm12,%xmm8 + DB 102,65,15,114,208,2 ; psrld $0x2,%xmm8 + DB 102,69,15,235,197 ; por %xmm13,%xmm8 + DB 102,69,15,235,198 ; por %xmm14,%xmm8 + DB 102,69,15,235,193 ; por %xmm9,%xmm8 + DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 + DB 68,15,89,5,22,165,1,0 ; mulps 0x1a516(%rip),%xmm8 # 2d3f0 <_sk_overlay_sse2_8bit+0xf41> + DB 68,15,88,5,30,165,1,0 ; addps 0x1a51e(%rip),%xmm8 # 2d400 <_sk_overlay_sse2_8bit+0xf51> DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 @@ -17163,7 +17299,7 @@ _sk_uniform_color_sse41 LABEL PROC PUBLIC _sk_black_color_sse41 _sk_black_color_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,227,168,1,0 ; movaps 0x1a8e3(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,84,164,1,0 ; movaps 0x1a454(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -17172,7 +17308,7 @@ _sk_black_color_sse41 LABEL PROC PUBLIC _sk_white_color_sse41 _sk_white_color_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,5,207,168,1,0 ; movaps 0x1a8cf(%rip),%xmm0 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,5,64,164,1,0 ; movaps 0x1a440(%rip),%xmm0 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,40,216 ; movaps %xmm0,%xmm3 @@ -17210,7 +17346,7 @@ _sk_clear_sse41 LABEL PROC PUBLIC _sk_srcatop_sse41 _sk_srcatop_sse41 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,127,168,1,0 ; movaps 0x1a87f(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,240,163,1,0 ; movaps 0x1a3f0(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -17233,7 +17369,7 @@ PUBLIC _sk_dstatop_sse41 _sk_dstatop_sse41 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 DB 68,15,89,196 ; mulps %xmm4,%xmm8 - DB 68,15,40,13,50,168,1,0 ; movaps 0x1a832(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,163,163,1,0 ; movaps 0x1a3a3(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -17274,7 +17410,7 @@ _sk_dstin_sse41 LABEL PROC PUBLIC _sk_srcout_sse41 _sk_srcout_sse41 LABEL PROC - DB 68,15,40,5,198,167,1,0 ; movaps 0x1a7c6(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,55,163,1,0 ; movaps 0x1a337(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -17285,7 +17421,7 @@ _sk_srcout_sse41 LABEL PROC PUBLIC _sk_dstout_sse41 _sk_dstout_sse41 LABEL PROC - DB 68,15,40,5,166,167,1,0 ; movaps 0x1a7a6(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,23,163,1,0 ; movaps 0x1a317(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -17300,7 +17436,7 @@ _sk_dstout_sse41 LABEL PROC PUBLIC _sk_srcover_sse41 _sk_srcover_sse41 LABEL PROC - DB 68,15,40,5,121,167,1,0 ; movaps 0x1a779(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,234,162,1,0 ; movaps 0x1a2ea(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -17318,7 +17454,7 @@ _sk_srcover_sse41 LABEL PROC PUBLIC _sk_dstover_sse41 _sk_dstover_sse41 LABEL PROC - DB 68,15,40,5,61,167,1,0 ; movaps 0x1a73d(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,174,162,1,0 ; movaps 0x1a2ae(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -17342,7 +17478,7 @@ _sk_modulate_sse41 LABEL PROC PUBLIC _sk_multiply_sse41 _sk_multiply_sse41 LABEL PROC - DB 68,15,40,5,1,167,1,0 ; movaps 0x1a701(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,114,162,1,0 ; movaps 0x1a272(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -17378,7 +17514,7 @@ _sk_multiply_sse41 LABEL PROC PUBLIC _sk_plus__sse41 _sk_plus__sse41 LABEL PROC DB 15,88,196 ; addps %xmm4,%xmm0 - DB 68,15,40,5,130,166,1,0 ; movaps 0x1a682(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,243,161,1,0 ; movaps 0x1a1f3(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 15,88,205 ; addps %xmm5,%xmm1 DB 65,15,93,200 ; minps %xmm8,%xmm1 @@ -17416,7 +17552,7 @@ _sk_screen_sse41 LABEL PROC PUBLIC _sk_xor__sse41 _sk_xor__sse41 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,14,166,1,0 ; movaps 0x1a60e(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,127,161,1,0 ; movaps 0x1a17f(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -17462,7 +17598,7 @@ _sk_darken_sse41 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,95,209 ; maxps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,105,165,1,0 ; movaps 0x1a569(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,218,160,1,0 ; movaps 0x1a0da(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -17494,7 +17630,7 @@ _sk_lighten_sse41 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,254,164,1,0 ; movaps 0x1a4fe(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,111,160,1,0 ; movaps 0x1a06f(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -17529,7 +17665,7 @@ _sk_difference_sse41 LABEL PROC DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,136,164,1,0 ; movaps 0x1a488(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,249,159,1,0 ; movaps 0x19ff9(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -17555,7 +17691,7 @@ _sk_exclusion_sse41 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,56,164,1,0 ; movaps 0x1a438(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,169,159,1,0 ; movaps 0x19fa9(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -17565,7 +17701,7 @@ _sk_exclusion_sse41 LABEL PROC PUBLIC _sk_colorburn_sse41 _sk_colorburn_sse41 LABEL PROC - DB 68,15,40,29,31,164,1,0 ; movaps 0x1a41f(%rip),%xmm11 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,29,144,159,1,0 ; movaps 0x19f90(%rip),%xmm11 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,227 ; movaps %xmm11,%xmm12 DB 68,15,92,231 ; subps %xmm7,%xmm12 DB 69,15,40,204 ; movaps %xmm12,%xmm9 @@ -17647,7 +17783,7 @@ _sk_colorburn_sse41 LABEL PROC PUBLIC _sk_colordodge_sse41 _sk_colordodge_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,21,233,162,1,0 ; movaps 0x1a2e9(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,90,158,1,0 ; movaps 0x19e5a(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -17728,7 +17864,7 @@ _sk_hardlight_sse41 LABEL PROC DB 15,40,244 ; movaps %xmm4,%xmm6 DB 15,40,227 ; movaps %xmm3,%xmm4 DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 68,15,40,21,177,161,1,0 ; movaps 0x1a1b1(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,34,157,1,0 ; movaps 0x19d22(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,40,234 ; movaps %xmm10,%xmm5 DB 15,92,239 ; subps %xmm7,%xmm5 DB 15,40,197 ; movaps %xmm5,%xmm0 @@ -17810,7 +17946,7 @@ PUBLIC _sk_overlay_sse41 _sk_overlay_sse41 LABEL PROC DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 68,15,40,240 ; movaps %xmm0,%xmm14 - DB 68,15,40,21,131,160,1,0 ; movaps 0x1a083(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,244,155,1,0 ; movaps 0x19bf4(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 65,15,40,195 ; movaps %xmm11,%xmm0 @@ -17894,7 +18030,7 @@ _sk_softlight_sse41 LABEL PROC DB 15,40,198 ; movaps %xmm6,%xmm0 DB 15,94,199 ; divps %xmm7,%xmm0 DB 65,15,84,193 ; andps %xmm9,%xmm0 - DB 15,40,13,70,159,1,0 ; movaps 0x19f46(%rip),%xmm1 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,13,183,154,1,0 ; movaps 0x19ab7(%rip),%xmm1 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 68,15,40,240 ; movaps %xmm0,%xmm14 @@ -17907,10 +18043,10 @@ _sk_softlight_sse41 LABEL PROC DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,89,210 ; mulps %xmm2,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 68,15,40,45,116,159,1,0 ; movaps 0x19f74(%rip),%xmm13 # 2d520 <_sk_overlay_sse2_8bit+0xfc8> + DB 68,15,40,45,229,154,1,0 ; movaps 0x19ae5(%rip),%xmm13 # 2d410 <_sk_overlay_sse2_8bit+0xf61> DB 69,15,88,245 ; addps %xmm13,%xmm14 DB 68,15,89,242 ; mulps %xmm2,%xmm14 - DB 68,15,40,37,116,159,1,0 ; movaps 0x19f74(%rip),%xmm12 # 2d530 <_sk_overlay_sse2_8bit+0xfd8> + DB 68,15,40,37,229,154,1,0 ; movaps 0x19ae5(%rip),%xmm12 # 2d420 <_sk_overlay_sse2_8bit+0xf71> DB 69,15,89,252 ; mulps %xmm12,%xmm15 DB 69,15,88,254 ; addps %xmm14,%xmm15 DB 15,40,198 ; movaps %xmm6,%xmm0 @@ -18096,12 +18232,12 @@ _sk_hue_sse41 LABEL PROC DB 68,15,84,208 ; andps %xmm0,%xmm10 DB 15,84,200 ; andps %xmm0,%xmm1 DB 68,15,84,232 ; andps %xmm0,%xmm13 - DB 15,40,5,218,156,1,0 ; movaps 0x19cda(%rip),%xmm0 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 15,40,5,75,152,1,0 ; movaps 0x1984b(%rip),%xmm0 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 68,15,89,224 ; mulps %xmm0,%xmm12 - DB 15,40,21,223,156,1,0 ; movaps 0x19cdf(%rip),%xmm2 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 15,40,21,80,152,1,0 ; movaps 0x19850(%rip),%xmm2 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 15,89,250 ; mulps %xmm2,%xmm7 DB 65,15,88,252 ; addps %xmm12,%xmm7 - DB 68,15,40,53,224,156,1,0 ; movaps 0x19ce0(%rip),%xmm14 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,53,81,152,1,0 ; movaps 0x19851(%rip),%xmm14 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 68,15,40,252 ; movaps %xmm4,%xmm15 DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 68,15,88,255 ; addps %xmm7,%xmm15 @@ -18184,7 +18320,7 @@ _sk_hue_sse41 LABEL PROC DB 65,15,88,214 ; addps %xmm14,%xmm2 DB 15,40,196 ; movaps %xmm4,%xmm0 DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 - DB 68,15,40,13,245,154,1,0 ; movaps 0x19af5(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,102,150,1,0 ; movaps 0x19666(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,40,225 ; movaps %xmm9,%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 15,40,68,36,48 ; movaps 0x30(%rsp),%xmm0 @@ -18278,14 +18414,14 @@ _sk_saturation_sse41 LABEL PROC DB 68,15,84,215 ; andps %xmm7,%xmm10 DB 68,15,84,223 ; andps %xmm7,%xmm11 DB 68,15,84,199 ; andps %xmm7,%xmm8 - DB 15,40,21,24,154,1,0 ; movaps 0x19a18(%rip),%xmm2 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 15,40,21,137,149,1,0 ; movaps 0x19589(%rip),%xmm2 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 15,40,221 ; movaps %xmm5,%xmm3 DB 15,89,218 ; mulps %xmm2,%xmm3 - DB 15,40,13,27,154,1,0 ; movaps 0x19a1b(%rip),%xmm1 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 15,40,13,140,149,1,0 ; movaps 0x1958c(%rip),%xmm1 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 15,40,254 ; movaps %xmm6,%xmm7 DB 15,89,249 ; mulps %xmm1,%xmm7 DB 15,88,251 ; addps %xmm3,%xmm7 - DB 68,15,40,45,26,154,1,0 ; movaps 0x19a1a(%rip),%xmm13 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,45,139,149,1,0 ; movaps 0x1958b(%rip),%xmm13 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 69,15,89,245 ; mulps %xmm13,%xmm14 DB 68,15,88,247 ; addps %xmm7,%xmm14 DB 65,15,40,218 ; movaps %xmm10,%xmm3 @@ -18366,7 +18502,7 @@ _sk_saturation_sse41 LABEL PROC DB 65,15,88,253 ; addps %xmm13,%xmm7 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 102,68,15,56,20,223 ; blendvps %xmm0,%xmm7,%xmm11 - DB 68,15,40,13,48,152,1,0 ; movaps 0x19830(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,161,147,1,0 ; movaps 0x193a1(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 68,15,92,204 ; subps %xmm4,%xmm9 DB 15,40,60,36 ; movaps (%rsp),%xmm7 @@ -18421,14 +18557,14 @@ _sk_color_sse41 LABEL PROC DB 15,40,231 ; movaps %xmm7,%xmm4 DB 68,15,89,244 ; mulps %xmm4,%xmm14 DB 15,89,204 ; mulps %xmm4,%xmm1 - DB 68,15,40,13,229,151,1,0 ; movaps 0x197e5(%rip),%xmm9 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,13,86,147,1,0 ; movaps 0x19356(%rip),%xmm9 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 65,15,40,250 ; movaps %xmm10,%xmm7 DB 65,15,89,249 ; mulps %xmm9,%xmm7 - DB 68,15,40,21,229,151,1,0 ; movaps 0x197e5(%rip),%xmm10 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 68,15,40,21,86,147,1,0 ; movaps 0x19356(%rip),%xmm10 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 65,15,40,219 ; movaps %xmm11,%xmm3 DB 65,15,89,218 ; mulps %xmm10,%xmm3 DB 15,88,223 ; addps %xmm7,%xmm3 - DB 68,15,40,29,226,151,1,0 ; movaps 0x197e2(%rip),%xmm11 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,29,83,147,1,0 ; movaps 0x19353(%rip),%xmm11 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 69,15,40,236 ; movaps %xmm12,%xmm13 DB 69,15,89,235 ; mulps %xmm11,%xmm13 DB 68,15,88,235 ; addps %xmm3,%xmm13 @@ -18513,7 +18649,7 @@ _sk_color_sse41 LABEL PROC DB 65,15,88,251 ; addps %xmm11,%xmm7 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 102,15,56,20,207 ; blendvps %xmm0,%xmm7,%xmm1 - DB 68,15,40,13,238,149,1,0 ; movaps 0x195ee(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,95,145,1,0 ; movaps 0x1915f(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,92,196 ; subps %xmm4,%xmm0 DB 68,15,89,192 ; mulps %xmm0,%xmm8 @@ -18565,13 +18701,13 @@ _sk_luminosity_sse41 LABEL PROC DB 69,15,89,216 ; mulps %xmm8,%xmm11 DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,89,205 ; mulps %xmm5,%xmm9 - DB 68,15,40,5,176,149,1,0 ; movaps 0x195b0(%rip),%xmm8 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,5,33,145,1,0 ; movaps 0x19121(%rip),%xmm8 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 65,15,89,192 ; mulps %xmm8,%xmm0 - DB 68,15,40,21,180,149,1,0 ; movaps 0x195b4(%rip),%xmm10 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 68,15,40,21,37,145,1,0 ; movaps 0x19125(%rip),%xmm10 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 15,40,233 ; movaps %xmm1,%xmm5 DB 65,15,89,234 ; mulps %xmm10,%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 - DB 68,15,40,37,178,149,1,0 ; movaps 0x195b2(%rip),%xmm12 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,37,35,145,1,0 ; movaps 0x19123(%rip),%xmm12 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 68,15,40,242 ; movaps %xmm2,%xmm14 DB 69,15,89,244 ; mulps %xmm12,%xmm14 DB 68,15,88,245 ; addps %xmm5,%xmm14 @@ -18656,7 +18792,7 @@ _sk_luminosity_sse41 LABEL PROC DB 65,15,88,244 ; addps %xmm12,%xmm6 DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 102,68,15,56,20,206 ; blendvps %xmm0,%xmm6,%xmm9 - DB 15,40,5,184,147,1,0 ; movaps 0x193b8(%rip),%xmm0 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,5,41,143,1,0 ; movaps 0x18f29(%rip),%xmm0 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,92,215 ; subps %xmm7,%xmm2 DB 15,89,226 ; mulps %xmm2,%xmm4 @@ -18700,23 +18836,23 @@ _sk_srcover_rgba_8888_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,219,0,0,0 ; jne 14273 <_sk_srcover_rgba_8888_sse41+0xfb> + DB 15,133,219,0,0,0 ; jne 145f2 <_sk_srcover_rgba_8888_sse41+0xfb> DB 243,65,15,111,60,128 ; movdqu (%r8,%rax,4),%xmm7 DB 72,133,255 ; test %rdi,%rdi - DB 102,15,111,37,199,147,1,0 ; movdqa 0x193c7(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,37,56,143,1,0 ; movdqa 0x18f38(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,56,0,45,195,147,1,0 ; pshufb 0x193c3(%rip),%xmm5 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,45,52,143,1,0 ; pshufb 0x18f34(%rip),%xmm5 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,56,0,53,195,147,1,0 ; pshufb 0x193c3(%rip),%xmm6 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,53,52,143,1,0 ; pshufb 0x18f34(%rip),%xmm6 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 DB 102,15,114,215,24 ; psrld $0x18,%xmm7 DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 - DB 68,15,40,5,224,146,1,0 ; movaps 0x192e0(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,81,142,1,0 ; movaps 0x18e51(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 - DB 68,15,40,37,180,147,1,0 ; movaps 0x193b4(%rip),%xmm12 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,37,37,143,1,0 ; movaps 0x18f25(%rip),%xmm12 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 65,15,89,196 ; mulps %xmm12,%xmm0 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -18742,7 +18878,7 @@ _sk_srcover_rgba_8888_sse41 LABEL PROC DB 102,15,114,240,24 ; pslld $0x18,%xmm0 DB 102,15,235,194 ; por %xmm2,%xmm0 DB 102,15,235,193 ; por %xmm1,%xmm0 - DB 117,88 ; jne 142b1 <_sk_srcover_rgba_8888_sse41+0x139> + DB 117,88 ; jne 14630 <_sk_srcover_rgba_8888_sse41+0x139> DB 243,65,15,127,4,128 ; movdqu %xmm0,(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,193 ; movaps %xmm9,%xmm0 @@ -18753,32 +18889,32 @@ _sk_srcover_rgba_8888_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,91 ; je 142db <_sk_srcover_rgba_8888_sse41+0x163> + DB 116,91 ; je 1465a <_sk_srcover_rgba_8888_sse41+0x163> DB 102,15,239,255 ; pxor %xmm7,%xmm7 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,22 ; je 142a0 <_sk_srcover_rgba_8888_sse41+0x128> + DB 116,22 ; je 1461f <_sk_srcover_rgba_8888_sse41+0x128> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,10,255,255,255 ; jne 1419e <_sk_srcover_rgba_8888_sse41+0x26> + DB 15,133,10,255,255,255 ; jne 1451d <_sk_srcover_rgba_8888_sse41+0x26> DB 102,65,15,110,100,128,8 ; movd 0x8(%r8,%rax,4),%xmm4 DB 102,15,112,252,69 ; pshufd $0x45,%xmm4,%xmm7 DB 243,65,15,126,36,128 ; movq (%r8,%rax,4),%xmm4 DB 102,15,58,14,252,15 ; pblendw $0xf,%xmm4,%xmm7 - DB 233,237,254,255,255 ; jmpq 1419e <_sk_srcover_rgba_8888_sse41+0x26> + DB 233,237,254,255,255 ; jmpq 1451d <_sk_srcover_rgba_8888_sse41+0x26> DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,40 ; je 142e6 <_sk_srcover_rgba_8888_sse41+0x16e> + DB 116,40 ; je 14665 <_sk_srcover_rgba_8888_sse41+0x16e> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,15 ; je 142d3 <_sk_srcover_rgba_8888_sse41+0x15b> + DB 116,15 ; je 14652 <_sk_srcover_rgba_8888_sse41+0x15b> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,149 ; jne 1425f <_sk_srcover_rgba_8888_sse41+0xe7> + DB 117,149 ; jne 145de <_sk_srcover_rgba_8888_sse41+0xe7> DB 102,65,15,58,22,68,128,8,2 ; pextrd $0x2,%xmm0,0x8(%r8,%rax,4) DB 102,65,15,214,4,128 ; movq %xmm0,(%r8,%rax,4) - DB 235,132 ; jmp 1425f <_sk_srcover_rgba_8888_sse41+0xe7> + DB 235,132 ; jmp 145de <_sk_srcover_rgba_8888_sse41+0xe7> DB 102,65,15,110,60,128 ; movd (%r8,%rax,4),%xmm7 - DB 233,184,254,255,255 ; jmpq 1419e <_sk_srcover_rgba_8888_sse41+0x26> + DB 233,184,254,255,255 ; jmpq 1451d <_sk_srcover_rgba_8888_sse41+0x26> DB 102,65,15,126,4,128 ; movd %xmm0,(%r8,%rax,4) - DB 233,110,255,255,255 ; jmpq 1425f <_sk_srcover_rgba_8888_sse41+0xe7> + DB 233,110,255,255,255 ; jmpq 145de <_sk_srcover_rgba_8888_sse41+0xe7> PUBLIC _sk_clamp_0_sse41 _sk_clamp_0_sse41 LABEL PROC @@ -18792,7 +18928,7 @@ _sk_clamp_0_sse41 LABEL PROC PUBLIC _sk_clamp_1_sse41 _sk_clamp_1_sse41 LABEL PROC - DB 68,15,40,5,175,145,1,0 ; movaps 0x191af(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,32,141,1,0 ; movaps 0x18d20(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -18802,7 +18938,7 @@ _sk_clamp_1_sse41 LABEL PROC PUBLIC _sk_clamp_a_sse41 _sk_clamp_a_sse41 LABEL PROC - DB 15,93,29,148,145,1,0 ; minps 0x19194(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,29,5,141,1,0 ; minps 0x18d05(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -18811,7 +18947,7 @@ _sk_clamp_a_sse41 LABEL PROC PUBLIC _sk_clamp_a_dst_sse41 _sk_clamp_a_dst_sse41 LABEL PROC - DB 15,93,61,128,145,1,0 ; minps 0x19180(%rip),%xmm7 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,61,241,140,1,0 ; minps 0x18cf1(%rip),%xmm7 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,93,231 ; minps %xmm7,%xmm4 DB 15,93,239 ; minps %xmm7,%xmm5 DB 15,93,247 ; minps %xmm7,%xmm6 @@ -18840,7 +18976,7 @@ _sk_swap_rb_sse41 LABEL PROC PUBLIC _sk_invert_sse41 _sk_invert_sse41 LABEL PROC - DB 68,15,40,5,60,145,1,0 ; movaps 0x1913c(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,173,140,1,0 ; movaps 0x18cad(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,200 ; subps %xmm0,%xmm9 DB 69,15,40,208 ; movaps %xmm8,%xmm10 @@ -18892,7 +19028,7 @@ _sk_premul_dst_sse41 LABEL PROC PUBLIC _sk_unpremul_sse41 _sk_unpremul_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,198,144,1,0 ; movaps 0x190c6(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,55,140,1,0 ; movaps 0x18c37(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,94,203 ; divps %xmm3,%xmm9 DB 68,15,194,195,4 ; cmpneqps %xmm3,%xmm8 DB 69,15,84,193 ; andps %xmm9,%xmm8 @@ -18904,20 +19040,20 @@ _sk_unpremul_sse41 LABEL PROC PUBLIC _sk_from_srgb_sse41 _sk_from_srgb_sse41 LABEL PROC - DB 68,15,40,29,145,145,1,0 ; movaps 0x19191(%rip),%xmm11 # 2d5b0 <_sk_overlay_sse2_8bit+0x1058> + DB 68,15,40,29,2,141,1,0 ; movaps 0x18d02(%rip),%xmm11 # 2d4a0 <_sk_overlay_sse2_8bit+0xff1> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 DB 68,15,40,208 ; movaps %xmm0,%xmm10 DB 69,15,89,210 ; mulps %xmm10,%xmm10 - DB 68,15,40,37,9,145,1,0 ; movaps 0x19109(%rip),%xmm12 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,37,122,140,1,0 ; movaps 0x18c7a(%rip),%xmm12 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,121,145,1,0 ; movaps 0x19179(%rip),%xmm13 # 2d5c0 <_sk_overlay_sse2_8bit+0x1068> + DB 68,15,40,45,234,140,1,0 ; movaps 0x18cea(%rip),%xmm13 # 2d4b0 <_sk_overlay_sse2_8bit+0x1001> DB 69,15,88,197 ; addps %xmm13,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 - DB 68,15,40,53,121,145,1,0 ; movaps 0x19179(%rip),%xmm14 # 2d5d0 <_sk_overlay_sse2_8bit+0x1078> + DB 68,15,40,53,234,140,1,0 ; movaps 0x18cea(%rip),%xmm14 # 2d4c0 <_sk_overlay_sse2_8bit+0x1011> DB 69,15,88,198 ; addps %xmm14,%xmm8 - DB 68,15,40,61,125,145,1,0 ; movaps 0x1917d(%rip),%xmm15 # 2d5e0 <_sk_overlay_sse2_8bit+0x1088> + DB 68,15,40,61,238,140,1,0 ; movaps 0x18cee(%rip),%xmm15 # 2d4d0 <_sk_overlay_sse2_8bit+0x1021> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 68,15,40,209 ; movaps %xmm1,%xmm10 @@ -18952,19 +19088,19 @@ PUBLIC _sk_from_srgb_dst_sse41 _sk_from_srgb_dst_sse41 LABEL PROC DB 68,15,40,204 ; movaps %xmm4,%xmm9 DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,29,200,144,1,0 ; movaps 0x190c8(%rip),%xmm11 # 2d5b0 <_sk_overlay_sse2_8bit+0x1058> + DB 68,15,40,29,57,140,1,0 ; movaps 0x18c39(%rip),%xmm11 # 2d4a0 <_sk_overlay_sse2_8bit+0xff1> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 69,15,89,211 ; mulps %xmm11,%xmm10 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 - DB 68,15,40,37,65,144,1,0 ; movaps 0x19041(%rip),%xmm12 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,37,178,139,1,0 ; movaps 0x18bb2(%rip),%xmm12 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 65,15,89,228 ; mulps %xmm12,%xmm4 - DB 68,15,40,45,181,144,1,0 ; movaps 0x190b5(%rip),%xmm13 # 2d5c0 <_sk_overlay_sse2_8bit+0x1068> + DB 68,15,40,45,38,140,1,0 ; movaps 0x18c26(%rip),%xmm13 # 2d4b0 <_sk_overlay_sse2_8bit+0x1001> DB 65,15,88,229 ; addps %xmm13,%xmm4 DB 15,89,224 ; mulps %xmm0,%xmm4 - DB 68,15,40,53,182,144,1,0 ; movaps 0x190b6(%rip),%xmm14 # 2d5d0 <_sk_overlay_sse2_8bit+0x1078> + DB 68,15,40,53,39,140,1,0 ; movaps 0x18c27(%rip),%xmm14 # 2d4c0 <_sk_overlay_sse2_8bit+0x1011> DB 65,15,88,230 ; addps %xmm14,%xmm4 - DB 68,15,40,61,186,144,1,0 ; movaps 0x190ba(%rip),%xmm15 # 2d5e0 <_sk_overlay_sse2_8bit+0x1088> + DB 68,15,40,61,43,140,1,0 ; movaps 0x18c2b(%rip),%xmm15 # 2d4d0 <_sk_overlay_sse2_8bit+0x1021> DB 69,15,194,207,1 ; cmpltps %xmm15,%xmm9 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 102,65,15,56,20,226 ; blendvps %xmm0,%xmm10,%xmm4 @@ -19007,22 +19143,22 @@ _sk_to_srgb_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,209 ; movaps %xmm1,%xmm2 DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 - DB 68,15,40,29,43,144,1,0 ; movaps 0x1902b(%rip),%xmm11 # 2d5f0 <_sk_overlay_sse2_8bit+0x1098> + DB 68,15,40,29,156,139,1,0 ; movaps 0x18b9c(%rip),%xmm11 # 2d4e0 <_sk_overlay_sse2_8bit+0x1031> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 - DB 68,15,40,37,43,144,1,0 ; movaps 0x1902b(%rip),%xmm12 # 2d600 <_sk_overlay_sse2_8bit+0x10a8> + DB 68,15,40,37,156,139,1,0 ; movaps 0x18b9c(%rip),%xmm12 # 2d4f0 <_sk_overlay_sse2_8bit+0x1041> DB 69,15,40,248 ; movaps %xmm8,%xmm15 DB 69,15,89,252 ; mulps %xmm12,%xmm15 - DB 68,15,40,21,43,144,1,0 ; movaps 0x1902b(%rip),%xmm10 # 2d610 <_sk_overlay_sse2_8bit+0x10b8> + DB 68,15,40,21,156,139,1,0 ; movaps 0x18b9c(%rip),%xmm10 # 2d500 <_sk_overlay_sse2_8bit+0x1051> DB 69,15,88,250 ; addps %xmm10,%xmm15 DB 69,15,89,248 ; mulps %xmm8,%xmm15 - DB 68,15,40,45,43,144,1,0 ; movaps 0x1902b(%rip),%xmm13 # 2d620 <_sk_overlay_sse2_8bit+0x10c8> + DB 68,15,40,45,156,139,1,0 ; movaps 0x18b9c(%rip),%xmm13 # 2d510 <_sk_overlay_sse2_8bit+0x1061> DB 69,15,88,253 ; addps %xmm13,%xmm15 - DB 68,15,40,53,47,144,1,0 ; movaps 0x1902f(%rip),%xmm14 # 2d630 <_sk_overlay_sse2_8bit+0x10d8> + DB 68,15,40,53,160,139,1,0 ; movaps 0x18ba0(%rip),%xmm14 # 2d520 <_sk_overlay_sse2_8bit+0x1071> DB 69,15,88,198 ; addps %xmm14,%xmm8 DB 69,15,83,192 ; rcpps %xmm8,%xmm8 DB 69,15,89,199 ; mulps %xmm15,%xmm8 - DB 68,15,40,61,43,144,1,0 ; movaps 0x1902b(%rip),%xmm15 # 2d640 <_sk_overlay_sse2_8bit+0x10e8> + DB 68,15,40,61,156,139,1,0 ; movaps 0x18b9c(%rip),%xmm15 # 2d530 <_sk_overlay_sse2_8bit+0x1081> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 68,15,82,202 ; rsqrtps %xmm2,%xmm9 @@ -19076,7 +19212,7 @@ _sk_rgb_to_hsl_sse41 LABEL PROC DB 68,15,93,224 ; minps %xmm0,%xmm12 DB 65,15,40,203 ; movaps %xmm11,%xmm1 DB 65,15,92,204 ; subps %xmm12,%xmm1 - DB 68,15,40,53,223,141,1,0 ; movaps 0x18ddf(%rip),%xmm14 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,53,80,137,1,0 ; movaps 0x18950(%rip),%xmm14 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,94,241 ; divps %xmm1,%xmm14 DB 69,15,40,211 ; movaps %xmm11,%xmm10 DB 69,15,194,208,0 ; cmpeqps %xmm8,%xmm10 @@ -19085,27 +19221,27 @@ _sk_rgb_to_hsl_sse41 LABEL PROC DB 65,15,89,198 ; mulps %xmm14,%xmm0 DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 68,15,194,250,1 ; cmpltps %xmm2,%xmm15 - DB 68,15,84,61,70,143,1,0 ; andps 0x18f46(%rip),%xmm15 # 2d650 <_sk_overlay_sse2_8bit+0x10f8> + DB 68,15,84,61,183,138,1,0 ; andps 0x18ab7(%rip),%xmm15 # 2d540 <_sk_overlay_sse2_8bit+0x1091> DB 68,15,88,248 ; addps %xmm0,%xmm15 DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,194,193,0 ; cmpeqps %xmm9,%xmm0 DB 65,15,92,208 ; subps %xmm8,%xmm2 DB 65,15,89,214 ; mulps %xmm14,%xmm2 - DB 68,15,40,45,57,143,1,0 ; movaps 0x18f39(%rip),%xmm13 # 2d660 <_sk_overlay_sse2_8bit+0x1108> + DB 68,15,40,45,170,138,1,0 ; movaps 0x18aaa(%rip),%xmm13 # 2d550 <_sk_overlay_sse2_8bit+0x10a1> DB 65,15,88,213 ; addps %xmm13,%xmm2 DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 69,15,89,198 ; mulps %xmm14,%xmm8 - DB 68,15,88,5,53,143,1,0 ; addps 0x18f35(%rip),%xmm8 # 2d670 <_sk_overlay_sse2_8bit+0x1118> + DB 68,15,88,5,166,138,1,0 ; addps 0x18aa6(%rip),%xmm8 # 2d560 <_sk_overlay_sse2_8bit+0x10b1> DB 102,68,15,56,20,194 ; blendvps %xmm0,%xmm2,%xmm8 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 102,69,15,56,20,199 ; blendvps %xmm0,%xmm15,%xmm8 - DB 68,15,89,5,45,143,1,0 ; mulps 0x18f2d(%rip),%xmm8 # 2d680 <_sk_overlay_sse2_8bit+0x1128> + DB 68,15,89,5,158,138,1,0 ; mulps 0x18a9e(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x10c1> DB 69,15,40,203 ; movaps %xmm11,%xmm9 DB 69,15,194,204,4 ; cmpneqps %xmm12,%xmm9 DB 69,15,84,193 ; andps %xmm9,%xmm8 DB 69,15,92,235 ; subps %xmm11,%xmm13 DB 69,15,88,220 ; addps %xmm12,%xmm11 - DB 15,40,5,65,141,1,0 ; movaps 0x18d41(%rip),%xmm0 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,40,5,178,136,1,0 ; movaps 0x188b2(%rip),%xmm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,40,211 ; movaps %xmm11,%xmm2 DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,194,194,1 ; cmpltps %xmm2,%xmm0 @@ -19126,7 +19262,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,100,36,32 ; movaps %xmm4,0x20(%rsp) DB 15,41,92,36,16 ; movaps %xmm3,0x10(%rsp) DB 68,15,40,208 ; movaps %xmm0,%xmm10 - DB 68,15,40,13,243,140,1,0 ; movaps 0x18cf3(%rip),%xmm9 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,13,100,136,1,0 ; movaps 0x18864(%rip),%xmm9 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,194,194,2 ; cmpleps %xmm2,%xmm0 DB 15,40,217 ; movaps %xmm1,%xmm3 @@ -19139,19 +19275,19 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 15,41,20,36 ; movaps %xmm2,(%rsp) DB 69,15,88,192 ; addps %xmm8,%xmm8 DB 68,15,92,197 ; subps %xmm5,%xmm8 - DB 68,15,40,53,159,142,1,0 ; movaps 0x18e9f(%rip),%xmm14 # 2d690 <_sk_overlay_sse2_8bit+0x1138> + DB 68,15,40,53,16,138,1,0 ; movaps 0x18a10(%rip),%xmm14 # 2d580 <_sk_overlay_sse2_8bit+0x10d1> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 102,65,15,58,8,198,1 ; roundps $0x1,%xmm14,%xmm0 DB 68,15,92,240 ; subps %xmm0,%xmm14 - DB 68,15,40,29,152,142,1,0 ; movaps 0x18e98(%rip),%xmm11 # 2d6a0 <_sk_overlay_sse2_8bit+0x1148> + DB 68,15,40,29,9,138,1,0 ; movaps 0x18a09(%rip),%xmm11 # 2d590 <_sk_overlay_sse2_8bit+0x10e1> DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 15,40,245 ; movaps %xmm5,%xmm6 DB 65,15,92,240 ; subps %xmm8,%xmm6 - DB 15,40,61,49,142,1,0 ; movaps 0x18e31(%rip),%xmm7 # 2d650 <_sk_overlay_sse2_8bit+0x10f8> + DB 15,40,61,162,137,1,0 ; movaps 0x189a2(%rip),%xmm7 # 2d540 <_sk_overlay_sse2_8bit+0x1091> DB 69,15,40,238 ; movaps %xmm14,%xmm13 DB 68,15,89,239 ; mulps %xmm7,%xmm13 - DB 15,40,29,66,142,1,0 ; movaps 0x18e42(%rip),%xmm3 # 2d670 <_sk_overlay_sse2_8bit+0x1118> + DB 15,40,29,179,137,1,0 ; movaps 0x189b3(%rip),%xmm3 # 2d560 <_sk_overlay_sse2_8bit+0x10b1> DB 68,15,40,227 ; movaps %xmm3,%xmm12 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 68,15,89,230 ; mulps %xmm6,%xmm12 @@ -19161,7 +19297,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 68,15,40,253 ; movaps %xmm5,%xmm15 DB 102,69,15,56,20,252 ; blendvps %xmm0,%xmm12,%xmm15 - DB 68,15,40,37,33,142,1,0 ; movaps 0x18e21(%rip),%xmm12 # 2d680 <_sk_overlay_sse2_8bit+0x1128> + DB 68,15,40,37,146,137,1,0 ; movaps 0x18992(%rip),%xmm12 # 2d570 <_sk_overlay_sse2_8bit+0x10c1> DB 65,15,40,196 ; movaps %xmm12,%xmm0 DB 65,15,194,198,2 ; cmpleps %xmm14,%xmm0 DB 68,15,89,238 ; mulps %xmm6,%xmm13 @@ -19195,7 +19331,7 @@ _sk_hsl_to_rgb_sse41 LABEL PROC DB 65,15,40,198 ; movaps %xmm14,%xmm0 DB 15,40,20,36 ; movaps (%rsp),%xmm2 DB 102,15,56,20,202 ; blendvps %xmm0,%xmm2,%xmm1 - DB 68,15,88,21,186,141,1,0 ; addps 0x18dba(%rip),%xmm10 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,88,21,43,137,1,0 ; addps 0x1892b(%rip),%xmm10 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 102,65,15,58,8,194,1 ; roundps $0x1,%xmm10,%xmm0 DB 68,15,92,208 ; subps %xmm0,%xmm10 DB 69,15,194,218,2 ; cmpleps %xmm10,%xmm11 @@ -19247,11 +19383,11 @@ _sk_scale_u8_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,52 ; jne 149e1 <_sk_scale_u8_sse41+0x4c> + DB 117,52 ; jne 14d60 <_sk_scale_u8_sse41+0x4c> DB 102,71,15,56,49,4,16 ; pmovzxbd (%r8,%r10,1),%xmm8 - DB 102,68,15,219,5,179,139,1,0 ; pand 0x18bb3(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,68,15,219,5,36,135,1,0 ; pand 0x18724(%rip),%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,247,140,1,0 ; mulps 0x18cf7(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,89,5,104,136,1,0 ; mulps 0x18868(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -19262,12 +19398,12 @@ _sk_scale_u8_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,59 ; je 14a29 <_sk_scale_u8_sse41+0x94> + DB 116,59 ; je 14da8 <_sk_scale_u8_sse41+0x94> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 14a10 <_sk_scale_u8_sse41+0x7b> + DB 116,23 ; je 14d8f <_sk_scale_u8_sse41+0x7b> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,181 ; jne 149b4 <_sk_scale_u8_sse41+0x1f> + DB 117,181 ; jne 14d33 <_sk_scale_u8_sse41+0x1f> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 @@ -19275,10 +19411,10 @@ _sk_scale_u8_sse41 LABEL PROC DB 102,68,15,110,200 ; movd %eax,%xmm9 DB 102,69,15,56,49,201 ; pmovzxbd %xmm9,%xmm9 DB 102,69,15,58,14,193,15 ; pblendw $0xf,%xmm9,%xmm8 - DB 235,139 ; jmp 149b4 <_sk_scale_u8_sse41+0x1f> + DB 235,139 ; jmp 14d33 <_sk_scale_u8_sse41+0x1f> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 - DB 233,124,255,255,255 ; jmpq 149b4 <_sk_scale_u8_sse41+0x1f> + DB 233,124,255,255,255 ; jmpq 14d33 <_sk_scale_u8_sse41+0x1f> PUBLIC _sk_scale_565_sse41 _sk_scale_565_sse41 LABEL PROC @@ -19291,19 +19427,19 @@ _sk_scale_565_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,159,0,0,0 ; jne 14afb <_sk_scale_565_sse41+0xc3> + DB 15,133,159,0,0,0 ; jne 14e7a <_sk_scale_565_sse41+0xc3> DB 102,71,15,56,51,28,80 ; pmovzxwd (%r8,%r10,2),%xmm11 - DB 102,15,111,5,101,140,1,0 ; movdqa 0x18c65(%rip),%xmm0 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,5,214,135,1,0 ; movdqa 0x187d6(%rip),%xmm0 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,65,15,219,195 ; pand %xmm11,%xmm0 DB 68,15,91,200 ; cvtdq2ps %xmm0,%xmm9 - DB 68,15,89,13,100,140,1,0 ; mulps 0x18c64(%rip),%xmm9 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,5,108,140,1,0 ; movdqa 0x18c6c(%rip),%xmm0 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 68,15,89,13,213,135,1,0 ; mulps 0x187d5(%rip),%xmm9 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,5,221,135,1,0 ; movdqa 0x187dd(%rip),%xmm0 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,65,15,219,195 ; pand %xmm11,%xmm0 DB 68,15,91,208 ; cvtdq2ps %xmm0,%xmm10 - DB 68,15,89,21,107,140,1,0 ; mulps 0x18c6b(%rip),%xmm10 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,68,15,219,29,114,140,1,0 ; pand 0x18c72(%rip),%xmm11 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 68,15,89,21,220,135,1,0 ; mulps 0x187dc(%rip),%xmm10 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,68,15,219,29,227,135,1,0 ; pand 0x187e3(%rip),%xmm11 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11 - DB 68,15,89,29,118,140,1,0 ; mulps 0x18c76(%rip),%xmm11 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 68,15,89,29,231,135,1,0 ; mulps 0x187e7(%rip),%xmm11 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 15,40,195 ; movaps %xmm3,%xmm0 DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 DB 69,15,40,226 ; movaps %xmm10,%xmm12 @@ -19328,22 +19464,22 @@ _sk_scale_565_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,60 ; je 14b44 <_sk_scale_565_sse41+0x10c> + DB 116,60 ; je 14ec3 <_sk_scale_565_sse41+0x10c> DB 102,69,15,239,219 ; pxor %xmm11,%xmm11 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,26 ; je 14b2d <_sk_scale_565_sse41+0xf5> + DB 116,26 ; je 14eac <_sk_scale_565_sse41+0xf5> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,70,255,255,255 ; jne 14a63 <_sk_scale_565_sse41+0x2b> + DB 15,133,70,255,255,255 ; jne 14de2 <_sk_scale_565_sse41+0x2b> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,68,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm11 DB 102,67,15,110,4,80 ; movd (%r8,%r10,2),%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 DB 102,68,15,58,14,216,15 ; pblendw $0xf,%xmm0,%xmm11 - DB 233,31,255,255,255 ; jmpq 14a63 <_sk_scale_565_sse41+0x2b> + DB 233,31,255,255,255 ; jmpq 14de2 <_sk_scale_565_sse41+0x2b> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,68,15,110,216 ; movd %eax,%xmm11 - DB 233,16,255,255,255 ; jmpq 14a63 <_sk_scale_565_sse41+0x2b> + DB 233,16,255,255,255 ; jmpq 14de2 <_sk_scale_565_sse41+0x2b> PUBLIC _sk_lerp_1_float_sse41 _sk_lerp_1_float_sse41 LABEL PROC @@ -19374,11 +19510,11 @@ _sk_lerp_u8_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,72 ; jne 14beb <_sk_lerp_u8_sse41+0x60> + DB 117,72 ; jne 14f6a <_sk_lerp_u8_sse41+0x60> DB 102,71,15,56,49,4,16 ; pmovzxbd (%r8,%r10,1),%xmm8 - DB 102,68,15,219,5,189,137,1,0 ; pand 0x189bd(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,68,15,219,5,46,133,1,0 ; pand 0x1852e(%rip),%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,1,139,1,0 ; mulps 0x18b01(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,89,5,114,134,1,0 ; mulps 0x18672(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -19396,12 +19532,12 @@ _sk_lerp_u8_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,62 ; je 14c36 <_sk_lerp_u8_sse41+0xab> + DB 116,62 ; je 14fb5 <_sk_lerp_u8_sse41+0xab> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 14c1a <_sk_lerp_u8_sse41+0x8f> + DB 116,23 ; je 14f99 <_sk_lerp_u8_sse41+0x8f> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,161 ; jne 14baa <_sk_lerp_u8_sse41+0x1f> + DB 117,161 ; jne 14f29 <_sk_lerp_u8_sse41+0x1f> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 @@ -19409,10 +19545,10 @@ _sk_lerp_u8_sse41 LABEL PROC DB 102,68,15,110,200 ; movd %eax,%xmm9 DB 102,69,15,56,49,201 ; pmovzxbd %xmm9,%xmm9 DB 102,69,15,58,14,193,15 ; pblendw $0xf,%xmm9,%xmm8 - DB 233,116,255,255,255 ; jmpq 14baa <_sk_lerp_u8_sse41+0x1f> + DB 233,116,255,255,255 ; jmpq 14f29 <_sk_lerp_u8_sse41+0x1f> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 - DB 233,101,255,255,255 ; jmpq 14baa <_sk_lerp_u8_sse41+0x1f> + DB 233,101,255,255,255 ; jmpq 14f29 <_sk_lerp_u8_sse41+0x1f> PUBLIC _sk_lerp_565_sse41 _sk_lerp_565_sse41 LABEL PROC @@ -19425,19 +19561,19 @@ _sk_lerp_565_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,165,0,0,0 ; jne 14d0e <_sk_lerp_565_sse41+0xc9> + DB 15,133,165,0,0,0 ; jne 1508d <_sk_lerp_565_sse41+0xc9> DB 102,71,15,56,51,20,80 ; pmovzxwd (%r8,%r10,2),%xmm10 - DB 102,15,111,5,88,138,1,0 ; movdqa 0x18a58(%rip),%xmm0 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,5,201,133,1,0 ; movdqa 0x185c9(%rip),%xmm0 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,65,15,219,194 ; pand %xmm10,%xmm0 DB 68,15,91,200 ; cvtdq2ps %xmm0,%xmm9 - DB 68,15,89,13,87,138,1,0 ; mulps 0x18a57(%rip),%xmm9 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,5,95,138,1,0 ; movdqa 0x18a5f(%rip),%xmm0 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 68,15,89,13,200,133,1,0 ; mulps 0x185c8(%rip),%xmm9 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,5,208,133,1,0 ; movdqa 0x185d0(%rip),%xmm0 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,65,15,219,194 ; pand %xmm10,%xmm0 DB 68,15,91,216 ; cvtdq2ps %xmm0,%xmm11 - DB 68,15,89,29,94,138,1,0 ; mulps 0x18a5e(%rip),%xmm11 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,68,15,219,21,101,138,1,0 ; pand 0x18a65(%rip),%xmm10 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 68,15,89,29,207,133,1,0 ; mulps 0x185cf(%rip),%xmm11 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,68,15,219,21,214,133,1,0 ; pand 0x185d6(%rip),%xmm10 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 - DB 68,15,89,21,105,138,1,0 ; mulps 0x18a69(%rip),%xmm10 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 68,15,89,21,218,133,1,0 ; mulps 0x185da(%rip),%xmm10 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 15,40,195 ; movaps %xmm3,%xmm0 DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -19465,33 +19601,33 @@ _sk_lerp_565_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,60 ; je 14d57 <_sk_lerp_565_sse41+0x112> + DB 116,60 ; je 150d6 <_sk_lerp_565_sse41+0x112> DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,26 ; je 14d40 <_sk_lerp_565_sse41+0xfb> + DB 116,26 ; je 150bf <_sk_lerp_565_sse41+0xfb> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,64,255,255,255 ; jne 14c70 <_sk_lerp_565_sse41+0x2b> + DB 15,133,64,255,255,255 ; jne 14fef <_sk_lerp_565_sse41+0x2b> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,68,15,112,208,69 ; pshufd $0x45,%xmm0,%xmm10 DB 102,67,15,110,4,80 ; movd (%r8,%r10,2),%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 DB 102,68,15,58,14,208,15 ; pblendw $0xf,%xmm0,%xmm10 - DB 233,25,255,255,255 ; jmpq 14c70 <_sk_lerp_565_sse41+0x2b> + DB 233,25,255,255,255 ; jmpq 14fef <_sk_lerp_565_sse41+0x2b> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,68,15,110,208 ; movd %eax,%xmm10 - DB 233,10,255,255,255 ; jmpq 14c70 <_sk_lerp_565_sse41+0x2b> + DB 233,10,255,255,255 ; jmpq 14fef <_sk_lerp_565_sse41+0x2b> PUBLIC _sk_load_tables_sse41 _sk_load_tables_sse41 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,20,1,0,0 ; jne 14e88 <_sk_load_tables_sse41+0x122> + DB 15,133,17,1,0,0 ; jne 15207 <_sk_load_tables_sse41+0x122> DB 243,69,15,111,4,144 ; movdqu (%r8,%rdx,4),%xmm8 - DB 65,86 ; push %r14 - DB 83 ; push %rbx - DB 102,15,111,5,235,135,1,0 ; movdqa 0x187eb(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,5,92,131,1,0 ; movdqa 0x1835c(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 DB 102,73,15,126,193 ; movq %xmm0,%r9 @@ -19506,7 +19642,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,66,15,58,33,4,155,32 ; insertps $0x20,(%rbx,%r11,4),%xmm0 DB 102,66,15,58,33,4,3,48 ; insertps $0x30,(%rbx,%r8,1),%xmm0 DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 - DB 102,15,56,0,13,166,135,1,0 ; pshufb 0x187a6(%rip),%xmm1 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,13,23,131,1,0 ; pshufb 0x18317(%rip),%xmm1 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 DB 102,72,15,126,203 ; movq %xmm1,%rbx DB 68,15,182,203 ; movzbl %bl,%r9d @@ -19521,7 +19657,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 DB 76,139,64,24 ; mov 0x18(%rax),%r8 DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 - DB 102,15,56,0,21,98,135,1,0 ; pshufb 0x18762(%rip),%xmm2 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,21,211,130,1,0 ; pshufb 0x182d3(%rip),%xmm2 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx DB 102,72,15,126,208 ; movq %xmm2,%rax DB 68,15,182,200 ; movzbl %al,%r9d @@ -19536,7 +19672,7 @@ _sk_load_tables_sse41 LABEL PROC DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8 DB 65,15,91,216 ; cvtdq2ps %xmm8,%xmm3 - DB 15,89,29,63,136,1,0 ; mulps 0x1883f(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,29,176,131,1,0 ; mulps 0x183b0(%rip),%xmm3 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 @@ -19544,38 +19680,38 @@ _sk_load_tables_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,52 ; je 14ec9 <_sk_load_tables_sse41+0x163> + DB 116,52 ; je 15248 <_sk_load_tables_sse41+0x163> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 14eb7 <_sk_load_tables_sse41+0x151> + DB 116,23 ; je 15236 <_sk_load_tables_sse41+0x151> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,208,254,255,255 ; jne 14d7a <_sk_load_tables_sse41+0x14> + DB 15,133,211,254,255,255 ; jne 150fc <_sk_load_tables_sse41+0x17> DB 102,65,15,110,68,144,8 ; movd 0x8(%r8,%rdx,4),%xmm0 DB 102,68,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm8 DB 243,65,15,126,4,144 ; movq (%r8,%rdx,4),%xmm0 DB 102,68,15,58,14,192,15 ; pblendw $0xf,%xmm0,%xmm8 - DB 233,177,254,255,255 ; jmpq 14d7a <_sk_load_tables_sse41+0x14> + DB 233,180,254,255,255 ; jmpq 150fc <_sk_load_tables_sse41+0x17> DB 102,69,15,110,4,144 ; movd (%r8,%rdx,4),%xmm8 - DB 233,166,254,255,255 ; jmpq 14d7a <_sk_load_tables_sse41+0x14> + DB 233,169,254,255,255 ; jmpq 150fc <_sk_load_tables_sse41+0x17> PUBLIC _sk_load_tables_u16_be_sse41 _sk_load_tables_u16_be_sse41 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,95,1,0,0 ; jne 15049 <_sk_load_tables_u16_be_sse41+0x175> + DB 15,133,92,1,0,0 ; jne 153c8 <_sk_load_tables_u16_be_sse41+0x175> DB 102,67,15,16,4,72 ; movupd (%r8,%r9,2),%xmm0 DB 243,67,15,111,76,72,16 ; movdqu 0x10(%r8,%r9,2),%xmm1 - DB 65,86 ; push %r14 - DB 83 ; push %rbx DB 102,68,15,40,200 ; movapd %xmm0,%xmm9 DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,68,15,111,5,17,136,1,0 ; movdqa 0x18811(%rip),%xmm8 # 2d730 <_sk_overlay_sse2_8bit+0x11d8> + DB 102,68,15,111,5,130,131,1,0 ; movdqa 0x18382(%rip),%xmm8 # 2d620 <_sk_overlay_sse2_8bit+0x1171> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 @@ -19593,7 +19729,7 @@ _sk_load_tables_u16_be_sse41 LABEL PROC DB 102,15,58,33,194,32 ; insertps $0x20,%xmm2,%xmm0 DB 243,66,15,16,20,3 ; movss (%rbx,%r8,1),%xmm2 DB 102,15,58,33,194,48 ; insertps $0x30,%xmm2,%xmm0 - DB 102,15,56,0,13,192,135,1,0 ; pshufb 0x187c0(%rip),%xmm1 # 2d740 <_sk_overlay_sse2_8bit+0x11e8> + DB 102,15,56,0,13,49,131,1,0 ; pshufb 0x18331(%rip),%xmm1 # 2d630 <_sk_overlay_sse2_8bit+0x1181> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 DB 102,72,15,126,203 ; movq %xmm1,%rbx @@ -19629,31 +19765,33 @@ _sk_load_tables_u16_be_sse41 LABEL PROC DB 102,65,15,235,216 ; por %xmm8,%xmm3 DB 102,15,56,51,219 ; pmovzxwd %xmm3,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,14,135,1,0 ; mulps 0x1870e(%rip),%xmm3 # 2d750 <_sk_overlay_sse2_8bit+0x11f8> + DB 15,89,29,127,130,1,0 ; mulps 0x1827f(%rip),%xmm3 # 2d640 <_sk_overlay_sse2_8bit+0x1191> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 DB 255,224 ; jmpq *%rax DB 242,67,15,16,4,72 ; movsd (%r8,%r9,2),%xmm0 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,13 ; jne 15062 <_sk_load_tables_u16_be_sse41+0x18e> + DB 117,13 ; jne 153e1 <_sk_load_tables_u16_be_sse41+0x18e> DB 243,15,126,192 ; movq %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,149,254,255,255 ; jmpq 14ef7 <_sk_load_tables_u16_be_sse41+0x23> + DB 233,152,254,255,255 ; jmpq 15279 <_sk_load_tables_u16_be_sse41+0x26> DB 102,67,15,22,68,72,8 ; movhpd 0x8(%r8,%r9,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,128,254,255,255 ; jb 14ef7 <_sk_load_tables_u16_be_sse41+0x23> + DB 15,130,131,254,255,255 ; jb 15279 <_sk_load_tables_u16_be_sse41+0x26> DB 243,67,15,126,76,72,16 ; movq 0x10(%r8,%r9,2),%xmm1 - DB 233,116,254,255,255 ; jmpq 14ef7 <_sk_load_tables_u16_be_sse41+0x23> + DB 233,119,254,255,255 ; jmpq 15279 <_sk_load_tables_u16_be_sse41+0x26> PUBLIC _sk_load_tables_rgb_u16_be_sse41 _sk_load_tables_rgb_u16_be_sse41 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,82 ; lea (%rdx,%rdx,2),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,79,1,0,0 ; jne 151e4 <_sk_load_tables_rgb_u16_be_sse41+0x161> + DB 15,133,76,1,0,0 ; jne 15563 <_sk_load_tables_rgb_u16_be_sse41+0x161> DB 243,67,15,111,20,72 ; movdqu (%r8,%r9,2),%xmm2 DB 243,67,15,111,76,72,8 ; movdqu 0x8(%r8,%r9,2),%xmm1 DB 102,15,115,217,4 ; psrldq $0x4,%xmm1 @@ -19661,13 +19799,11 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 102,65,15,115,217,6 ; psrldq $0x6,%xmm9 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 - DB 65,86 ; push %r14 - DB 83 ; push %rbx - DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 + DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 - DB 102,68,15,111,5,87,134,1,0 ; movdqa 0x18657(%rip),%xmm8 # 2d730 <_sk_overlay_sse2_8bit+0x11d8> + DB 102,68,15,111,5,200,129,1,0 ; movdqa 0x181c8(%rip),%xmm8 # 2d620 <_sk_overlay_sse2_8bit+0x1171> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 @@ -19685,7 +19821,7 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 102,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm0 DB 243,66,15,16,28,3 ; movss (%rbx,%r8,1),%xmm3 DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0 - DB 102,15,56,0,13,38,134,1,0 ; pshufb 0x18626(%rip),%xmm1 # 2d760 <_sk_overlay_sse2_8bit+0x1208> + DB 102,15,56,0,13,119,129,1,0 ; pshufb 0x18177(%rip),%xmm1 # 2d630 <_sk_overlay_sse2_8bit+0x1181> DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 DB 102,72,15,126,203 ; movq %xmm1,%rbx @@ -19716,106 +19852,116 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC DB 243,65,15,16,28,24 ; movss (%r8,%rbx,1),%xmm3 DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,225,130,1,0 ; movaps 0x182e1(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,82,126,1,0 ; movaps 0x17e52(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 91 ; pop %rbx DB 65,94 ; pop %r14 DB 255,224 ; jmpq *%rax DB 102,67,15,110,20,72 ; movd (%r8,%r9,2),%xmm2 - DB 102,67,15,196,84,72,4,2 ; pinsrw $0x2,0x4(%r8,%r9,2),%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 102,67,15,196,84,72,4,2 ; pinsrw $0x2,0x4(%r8,%r9,2),%xmm2 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,14 ; jne 1520a <_sk_load_tables_rgb_u16_be_sse41+0x187> + DB 117,14 ; jne 15589 <_sk_load_tables_rgb_u16_be_sse41+0x187> DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 - DB 233,177,254,255,255 ; jmpq 150bb <_sk_load_tables_rgb_u16_be_sse41+0x38> + DB 233,180,254,255,255 ; jmpq 1543d <_sk_load_tables_rgb_u16_be_sse41+0x3b> DB 102,71,15,110,76,72,6 ; movd 0x6(%r8,%r9,2),%xmm9 - DB 102,71,15,196,76,72,10,2 ; pinsrw $0x2,0xa(%r8,%r9,2),%xmm9 DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 102,71,15,196,76,72,10,2 ; pinsrw $0x2,0xa(%r8,%r9,2),%xmm9 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,24 ; jb 1523b <_sk_load_tables_rgb_u16_be_sse41+0x1b8> + DB 114,24 ; jb 155ba <_sk_load_tables_rgb_u16_be_sse41+0x1b8> DB 102,67,15,110,76,72,12 ; movd 0xc(%r8,%r9,2),%xmm1 - DB 102,67,15,196,76,72,16,2 ; pinsrw $0x2,0x10(%r8,%r9,2),%xmm1 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,128,254,255,255 ; jmpq 150bb <_sk_load_tables_rgb_u16_be_sse41+0x38> + DB 102,67,15,196,76,72,16,2 ; pinsrw $0x2,0x10(%r8,%r9,2),%xmm1 + DB 233,131,254,255,255 ; jmpq 1543d <_sk_load_tables_rgb_u16_be_sse41+0x3b> DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,119,254,255,255 ; jmpq 150bb <_sk_load_tables_rgb_u16_be_sse41+0x38> + DB 233,122,254,255,255 ; jmpq 1543d <_sk_load_tables_rgb_u16_be_sse41+0x3b> PUBLIC _sk_byte_tables_sse41 _sk_byte_tables_sse41 LABEL PROC - DB 65,87 ; push %r15 + DB 85 ; push %rbp DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,77,131,1,0 ; movaps 0x1834d(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,191,126,1,0 ; movaps 0x17ebf(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8 DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,137,211 ; mov %r10d,%r11d + DB 77,137,194 ; mov %r8,%r10 DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,139,56 ; mov (%rax),%r15 - DB 76,139,112,8 ; mov 0x8(%rax),%r14 - DB 102,67,15,58,32,4,31,0 ; pinsrb $0x0,(%r15,%r11,1),%xmm0 - DB 102,67,15,58,32,4,23,1 ; pinsrb $0x1,(%r15,%r10,1),%xmm0 - DB 67,15,182,28,15 ; movzbl (%r15,%r9,1),%ebx - DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0 - DB 67,15,182,28,7 ; movzbl (%r15,%r8,1),%ebx - DB 102,15,58,32,195,3 ; pinsrb $0x3,%ebx,%xmm0 + DB 102,73,15,126,192 ; movq %xmm0,%r8 + DB 69,137,195 ; mov %r8d,%r11d + DB 77,137,198 ; mov %r8,%r14 + DB 73,193,238,32 ; shr $0x20,%r14 + DB 72,139,24 ; mov (%rax),%rbx + DB 76,139,64,8 ; mov 0x8(%rax),%r8 + DB 70,15,182,52,51 ; movzbl (%rbx,%r14,1),%r14d + DB 66,15,182,44,27 ; movzbl (%rbx,%r11,1),%ebp + DB 102,15,110,197 ; movd %ebp,%xmm0 + DB 102,65,15,58,32,198,1 ; pinsrb $0x1,%r14d,%xmm0 + DB 66,15,182,44,11 ; movzbl (%rbx,%r9,1),%ebp + DB 102,15,58,32,197,2 ; pinsrb $0x2,%ebp,%xmm0 + DB 66,15,182,44,19 ; movzbl (%rbx,%r10,1),%ebp + DB 102,15,58,32,197,3 ; pinsrb $0x3,%ebp,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,14,132,1,0 ; movaps 0x1840e(%rip),%xmm9 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,13,117,127,1,0 ; movaps 0x17f75(%rip),%xmm9 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 - DB 102,72,15,58,22,203,1 ; pextrq $0x1,%xmm1,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 102,72,15,58,22,205,1 ; pextrq $0x1,%xmm1,%rbp + DB 65,137,233 ; mov %ebp,%r9d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 65,137,218 ; mov %ebx,%r10d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 69,137,202 ; mov %r9d,%r10d - DB 73,193,233,32 ; shr $0x20,%r9 - DB 102,67,15,58,32,12,22,0 ; pinsrb $0x0,(%r14,%r10,1),%xmm1 - DB 102,67,15,58,32,12,14,1 ; pinsrb $0x1,(%r14,%r9,1),%xmm1 - DB 71,15,182,4,6 ; movzbl (%r14,%r8,1),%r8d - DB 102,65,15,58,32,200,2 ; pinsrb $0x2,%r8d,%xmm1 - DB 65,15,182,28,30 ; movzbl (%r14,%rbx,1),%ebx - DB 102,15,58,32,203,3 ; pinsrb $0x3,%ebx,%xmm1 + DB 69,15,182,28,24 ; movzbl (%r8,%rbx,1),%r11d + DB 67,15,182,28,16 ; movzbl (%r8,%r10,1),%ebx + DB 102,15,110,203 ; movd %ebx,%xmm1 + DB 102,65,15,58,32,203,1 ; pinsrb $0x1,%r11d,%xmm1 + DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx + DB 102,15,58,32,203,2 ; pinsrb $0x2,%ebx,%xmm1 + DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp + DB 102,15,58,32,205,3 ; pinsrb $0x3,%ebp,%xmm1 DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 76,139,80,16 ; mov 0x10(%rax),%r10 + DB 76,139,72,16 ; mov 0x10(%rax),%r9 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2 - DB 102,73,15,58,22,208,1 ; pextrq $0x1,%xmm2,%r8 - DB 69,137,193 ; mov %r8d,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 65,137,219 ; mov %ebx,%r11d + DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx + DB 65,137,216 ; mov %ebx,%r8d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,67,15,58,32,20,26,0 ; pinsrb $0x0,(%r10,%r11,1),%xmm2 - DB 102,65,15,58,32,20,26,1 ; pinsrb $0x1,(%r10,%rbx,1),%xmm2 - DB 67,15,182,28,10 ; movzbl (%r10,%r9,1),%ebx - DB 102,15,58,32,211,2 ; pinsrb $0x2,%ebx,%xmm2 - DB 67,15,182,28,2 ; movzbl (%r10,%r8,1),%ebx - DB 102,15,58,32,211,3 ; pinsrb $0x3,%ebx,%xmm2 + DB 102,72,15,126,213 ; movq %xmm2,%rbp + DB 65,137,234 ; mov %ebp,%r10d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 69,15,182,28,41 ; movzbl (%r9,%rbp,1),%r11d + DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp + DB 102,15,110,213 ; movd %ebp,%xmm2 + DB 102,65,15,58,32,211,1 ; pinsrb $0x1,%r11d,%xmm2 + DB 67,15,182,44,1 ; movzbl (%r9,%r8,1),%ebp + DB 102,15,58,32,213,2 ; pinsrb $0x2,%ebp,%xmm2 + DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp + DB 102,15,58,32,213,3 ; pinsrb $0x3,%ebp,%xmm2 DB 102,15,56,49,210 ; pmovzxbd %xmm2,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 + DB 72,139,64,24 ; mov 0x18(%rax),%rax DB 65,15,89,216 ; mulps %xmm8,%xmm3 DB 102,15,91,219 ; cvtps2dq %xmm3,%xmm3 - DB 102,72,15,58,22,219,1 ; pextrq $0x1,%xmm3,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 102,72,15,58,22,221,1 ; pextrq $0x1,%xmm3,%rbp + DB 65,137,232 ; mov %ebp,%r8d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 102,72,15,126,219 ; movq %xmm3,%rbx + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,72,15,126,216 ; movq %xmm3,%rax - DB 65,137,194 ; mov %eax,%r10d - DB 72,193,232,32 ; shr $0x20,%rax - DB 102,67,15,58,32,28,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm3 - DB 102,65,15,58,32,28,1,1 ; pinsrb $0x1,(%r9,%rax,1),%xmm3 - DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax - DB 102,15,58,32,216,2 ; pinsrb $0x2,%eax,%xmm3 - DB 65,15,182,4,25 ; movzbl (%r9,%rbx,1),%eax + DB 68,15,182,20,24 ; movzbl (%rax,%rbx,1),%r10d + DB 66,15,182,28,8 ; movzbl (%rax,%r9,1),%ebx + DB 102,15,110,219 ; movd %ebx,%xmm3 + DB 102,65,15,58,32,218,1 ; pinsrb $0x1,%r10d,%xmm3 + DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx + DB 102,15,58,32,219,2 ; pinsrb $0x2,%ebx,%xmm3 + DB 15,182,4,40 ; movzbl (%rax,%rbp,1),%eax DB 102,15,58,32,216,3 ; pinsrb $0x3,%eax,%xmm3 DB 102,15,56,49,219 ; pmovzxbd %xmm3,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 @@ -19823,12 +19969,12 @@ _sk_byte_tables_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_byte_tables_rgb_sse41 _sk_byte_tables_rgb_sse41 LABEL PROC - DB 65,87 ; push %r15 + DB 85 ; push %rbp DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax @@ -19847,49 +19993,55 @@ _sk_byte_tables_rgb_sse41 LABEL PROC DB 69,137,195 ; mov %r8d,%r11d DB 77,137,198 ; mov %r8,%r14 DB 73,193,238,32 ; shr $0x20,%r14 - DB 76,139,56 ; mov (%rax),%r15 + DB 72,139,24 ; mov (%rax),%rbx DB 76,139,64,8 ; mov 0x8(%rax),%r8 - DB 102,67,15,58,32,4,31,0 ; pinsrb $0x0,(%r15,%r11,1),%xmm0 - DB 102,67,15,58,32,4,55,1 ; pinsrb $0x1,(%r15,%r14,1),%xmm0 - DB 67,15,182,28,15 ; movzbl (%r15,%r9,1),%ebx - DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0 - DB 67,15,182,28,23 ; movzbl (%r15,%r10,1),%ebx - DB 102,15,58,32,195,3 ; pinsrb $0x3,%ebx,%xmm0 + DB 70,15,182,52,51 ; movzbl (%rbx,%r14,1),%r14d + DB 66,15,182,44,27 ; movzbl (%rbx,%r11,1),%ebp + DB 102,15,110,197 ; movd %ebp,%xmm0 + DB 102,65,15,58,32,198,1 ; pinsrb $0x1,%r14d,%xmm0 + DB 66,15,182,44,11 ; movzbl (%rbx,%r9,1),%ebp + DB 102,15,58,32,197,2 ; pinsrb $0x2,%ebp,%xmm0 + DB 66,15,182,44,19 ; movzbl (%rbx,%r10,1),%ebp + DB 102,15,58,32,197,3 ; pinsrb $0x3,%ebp,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,13,122,130,1,0 ; movaps 0x1827a(%rip),%xmm9 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,13,209,125,1,0 ; movaps 0x17dd1(%rip),%xmm9 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 - DB 102,72,15,58,22,203,1 ; pextrq $0x1,%xmm1,%rbx - DB 65,137,217 ; mov %ebx,%r9d + DB 102,72,15,58,22,205,1 ; pextrq $0x1,%xmm1,%rbp + DB 65,137,233 ; mov %ebp,%r9d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 65,137,218 ; mov %ebx,%r10d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,202 ; movq %xmm1,%r10 - DB 69,137,211 ; mov %r10d,%r11d - DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,58,32,12,24,0 ; pinsrb $0x0,(%r8,%r11,1),%xmm1 - DB 102,67,15,58,32,12,16,1 ; pinsrb $0x1,(%r8,%r10,1),%xmm1 - DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d - DB 102,65,15,58,32,201,2 ; pinsrb $0x2,%r9d,%xmm1 - DB 65,15,182,28,24 ; movzbl (%r8,%rbx,1),%ebx - DB 102,15,58,32,203,3 ; pinsrb $0x3,%ebx,%xmm1 + DB 69,15,182,28,24 ; movzbl (%r8,%rbx,1),%r11d + DB 67,15,182,28,16 ; movzbl (%r8,%r10,1),%ebx + DB 102,15,110,203 ; movd %ebx,%xmm1 + DB 102,65,15,58,32,203,1 ; pinsrb $0x1,%r11d,%xmm1 + DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx + DB 102,15,58,32,203,2 ; pinsrb $0x2,%ebx,%xmm1 + DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp + DB 102,15,58,32,205,3 ; pinsrb $0x3,%ebp,%xmm1 DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 76,139,72,16 ; mov 0x10(%rax),%r9 + DB 72,139,64,16 ; mov 0x10(%rax),%rax DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2 - DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx - DB 65,137,216 ; mov %ebx,%r8d + DB 102,72,15,58,22,213,1 ; pextrq $0x1,%xmm2,%rbp + DB 65,137,232 ; mov %ebp,%r8d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 102,72,15,126,211 ; movq %xmm2,%rbx + DB 65,137,217 ; mov %ebx,%r9d DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,72,15,126,208 ; movq %xmm2,%rax - DB 65,137,194 ; mov %eax,%r10d - DB 72,193,232,32 ; shr $0x20,%rax - DB 102,67,15,58,32,20,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm2 - DB 102,65,15,58,32,20,1,1 ; pinsrb $0x1,(%r9,%rax,1),%xmm2 - DB 67,15,182,4,1 ; movzbl (%r9,%r8,1),%eax - DB 102,15,58,32,208,2 ; pinsrb $0x2,%eax,%xmm2 - DB 65,15,182,4,25 ; movzbl (%r9,%rbx,1),%eax + DB 68,15,182,20,24 ; movzbl (%rax,%rbx,1),%r10d + DB 66,15,182,28,8 ; movzbl (%rax,%r9,1),%ebx + DB 102,15,110,211 ; movd %ebx,%xmm2 + DB 102,65,15,58,32,210,1 ; pinsrb $0x1,%r10d,%xmm2 + DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx + DB 102,15,58,32,211,2 ; pinsrb $0x2,%ebx,%xmm2 + DB 15,182,4,40 ; movzbl (%rax,%rbp,1),%eax DB 102,15,58,32,208,3 ; pinsrb $0x3,%eax,%xmm2 DB 102,15,56,49,210 ; pmovzxbd %xmm2,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 @@ -19897,7 +20049,7 @@ _sk_byte_tables_rgb_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 - DB 65,95 ; pop %r15 + DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax PUBLIC _sk_table_r_sse41 @@ -20025,31 +20177,31 @@ _sk_parametric_r_sse41 LABEL PROC DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,194 ; cvtdq2ps %xmm10,%xmm8 - DB 68,15,89,5,108,128,1,0 ; mulps 0x1806c(%rip),%xmm8 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,21,116,128,1,0 ; andps 0x18074(%rip),%xmm10 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,21,156,125,1,0 ; orps 0x17d9c(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,5,116,128,1,0 ; addps 0x18074(%rip),%xmm8 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 68,15,40,37,124,128,1,0 ; movaps 0x1807c(%rip),%xmm12 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,5,172,123,1,0 ; mulps 0x17bac(%rip),%xmm8 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,21,180,123,1,0 ; andps 0x17bb4(%rip),%xmm10 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,21,236,120,1,0 ; orps 0x178ec(%rip),%xmm10 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,5,180,123,1,0 ; addps 0x17bb4(%rip),%xmm8 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 68,15,40,37,188,123,1,0 ; movaps 0x17bbc(%rip),%xmm12 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 - DB 68,15,88,21,124,128,1,0 ; addps 0x1807c(%rip),%xmm10 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 68,15,40,37,132,128,1,0 ; movaps 0x18084(%rip),%xmm12 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,21,188,123,1,0 ; addps 0x17bbc(%rip),%xmm10 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 68,15,40,37,196,123,1,0 ; movaps 0x17bc4(%rip),%xmm12 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,196 ; subps %xmm12,%xmm8 DB 69,15,89,195 ; mulps %xmm11,%xmm8 DB 102,69,15,58,8,208,1 ; roundps $0x1,%xmm8,%xmm10 DB 69,15,40,216 ; movaps %xmm8,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,5,113,128,1,0 ; addps 0x18071(%rip),%xmm8 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 68,15,40,21,121,128,1,0 ; movaps 0x18079(%rip),%xmm10 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,5,177,123,1,0 ; addps 0x17bb1(%rip),%xmm8 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 68,15,40,21,185,123,1,0 ; movaps 0x17bb9(%rip),%xmm10 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 69,15,89,211 ; mulps %xmm11,%xmm10 DB 69,15,92,194 ; subps %xmm10,%xmm8 - DB 68,15,40,21,121,128,1,0 ; movaps 0x18079(%rip),%xmm10 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,21,185,123,1,0 ; movaps 0x17bb9(%rip),%xmm10 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 68,15,40,29,125,128,1,0 ; movaps 0x1807d(%rip),%xmm11 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,29,189,123,1,0 ; movaps 0x17bbd(%rip),%xmm11 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 69,15,94,218 ; divps %xmm10,%xmm11 DB 69,15,88,216 ; addps %xmm8,%xmm11 - DB 68,15,89,29,125,128,1,0 ; mulps 0x1807d(%rip),%xmm11 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 68,15,89,29,189,123,1,0 ; mulps 0x17bbd(%rip),%xmm11 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10 DB 243,68,15,16,64,20 ; movss 0x14(%rax),%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 @@ -20057,7 +20209,7 @@ _sk_parametric_r_sse41 LABEL PROC DB 102,69,15,56,20,193 ; blendvps %xmm0,%xmm9,%xmm8 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,4,125,1,0 ; minps 0x17d04(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,93,5,84,120,1,0 ; minps 0x17854(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -20085,31 +20237,31 @@ _sk_parametric_g_sse41 LABEL PROC DB 68,15,88,217 ; addps %xmm1,%xmm11 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 68,15,89,37,69,127,1,0 ; mulps 0x17f45(%rip),%xmm12 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,29,77,127,1,0 ; andps 0x17f4d(%rip),%xmm11 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,29,117,124,1,0 ; orps 0x17c75(%rip),%xmm11 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,37,77,127,1,0 ; addps 0x17f4d(%rip),%xmm12 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 15,40,13,86,127,1,0 ; movaps 0x17f56(%rip),%xmm1 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,37,133,122,1,0 ; mulps 0x17a85(%rip),%xmm12 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,29,141,122,1,0 ; andps 0x17a8d(%rip),%xmm11 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,29,197,119,1,0 ; orps 0x177c5(%rip),%xmm11 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,37,141,122,1,0 ; addps 0x17a8d(%rip),%xmm12 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 15,40,13,150,122,1,0 ; movaps 0x17a96(%rip),%xmm1 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,88,29,86,127,1,0 ; addps 0x17f56(%rip),%xmm11 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 15,40,13,95,127,1,0 ; movaps 0x17f5f(%rip),%xmm1 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,29,150,122,1,0 ; addps 0x17a96(%rip),%xmm11 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 15,40,13,159,122,1,0 ; movaps 0x17a9f(%rip),%xmm1 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 65,15,94,203 ; divps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10 DB 69,15,40,220 ; movaps %xmm12,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,37,76,127,1,0 ; addps 0x17f4c(%rip),%xmm12 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 15,40,13,85,127,1,0 ; movaps 0x17f55(%rip),%xmm1 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,37,140,122,1,0 ; addps 0x17a8c(%rip),%xmm12 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 15,40,13,149,122,1,0 ; movaps 0x17a95(%rip),%xmm1 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 68,15,40,21,85,127,1,0 ; movaps 0x17f55(%rip),%xmm10 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,21,149,122,1,0 ; movaps 0x17a95(%rip),%xmm10 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,13,90,127,1,0 ; movaps 0x17f5a(%rip),%xmm1 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 15,40,13,154,122,1,0 ; movaps 0x17a9a(%rip),%xmm1 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 65,15,94,202 ; divps %xmm10,%xmm1 DB 65,15,88,204 ; addps %xmm12,%xmm1 - DB 15,89,13,91,127,1,0 ; mulps 0x17f5b(%rip),%xmm1 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 15,89,13,155,122,1,0 ; mulps 0x17a9b(%rip),%xmm1 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,68,15,91,209 ; cvtps2dq %xmm1,%xmm10 DB 243,15,16,72,20 ; movss 0x14(%rax),%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 @@ -20117,7 +20269,7 @@ _sk_parametric_g_sse41 LABEL PROC DB 102,65,15,56,20,201 ; blendvps %xmm0,%xmm9,%xmm1 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,95,200 ; maxps %xmm0,%xmm1 - DB 15,93,13,230,123,1,0 ; minps 0x17be6(%rip),%xmm1 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,13,54,119,1,0 ; minps 0x17736(%rip),%xmm1 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -20145,31 +20297,31 @@ _sk_parametric_b_sse41 LABEL PROC DB 68,15,88,218 ; addps %xmm2,%xmm11 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 68,15,89,37,39,126,1,0 ; mulps 0x17e27(%rip),%xmm12 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,29,47,126,1,0 ; andps 0x17e2f(%rip),%xmm11 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,29,87,123,1,0 ; orps 0x17b57(%rip),%xmm11 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,37,47,126,1,0 ; addps 0x17e2f(%rip),%xmm12 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 15,40,21,56,126,1,0 ; movaps 0x17e38(%rip),%xmm2 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,37,103,121,1,0 ; mulps 0x17967(%rip),%xmm12 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,29,111,121,1,0 ; andps 0x1796f(%rip),%xmm11 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,29,167,118,1,0 ; orps 0x176a7(%rip),%xmm11 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,37,111,121,1,0 ; addps 0x1796f(%rip),%xmm12 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 15,40,21,120,121,1,0 ; movaps 0x17978(%rip),%xmm2 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,88,29,56,126,1,0 ; addps 0x17e38(%rip),%xmm11 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 15,40,21,65,126,1,0 ; movaps 0x17e41(%rip),%xmm2 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,29,120,121,1,0 ; addps 0x17978(%rip),%xmm11 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 15,40,21,129,121,1,0 ; movaps 0x17981(%rip),%xmm2 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 65,15,94,211 ; divps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10 DB 69,15,40,220 ; movaps %xmm12,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,37,46,126,1,0 ; addps 0x17e2e(%rip),%xmm12 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 15,40,21,55,126,1,0 ; movaps 0x17e37(%rip),%xmm2 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,37,110,121,1,0 ; addps 0x1796e(%rip),%xmm12 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 15,40,21,119,121,1,0 ; movaps 0x17977(%rip),%xmm2 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 68,15,92,226 ; subps %xmm2,%xmm12 - DB 68,15,40,21,55,126,1,0 ; movaps 0x17e37(%rip),%xmm10 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,21,119,121,1,0 ; movaps 0x17977(%rip),%xmm10 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,21,60,126,1,0 ; movaps 0x17e3c(%rip),%xmm2 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 15,40,21,124,121,1,0 ; movaps 0x1797c(%rip),%xmm2 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 65,15,94,210 ; divps %xmm10,%xmm2 DB 65,15,88,212 ; addps %xmm12,%xmm2 - DB 15,89,21,61,126,1,0 ; mulps 0x17e3d(%rip),%xmm2 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 15,89,21,125,121,1,0 ; mulps 0x1797d(%rip),%xmm2 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,68,15,91,210 ; cvtps2dq %xmm2,%xmm10 DB 243,15,16,80,20 ; movss 0x14(%rax),%xmm2 DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 @@ -20177,7 +20329,7 @@ _sk_parametric_b_sse41 LABEL PROC DB 102,65,15,56,20,209 ; blendvps %xmm0,%xmm9,%xmm2 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,95,208 ; maxps %xmm0,%xmm2 - DB 15,93,21,200,122,1,0 ; minps 0x17ac8(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,21,24,118,1,0 ; minps 0x17618(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -20205,31 +20357,31 @@ _sk_parametric_a_sse41 LABEL PROC DB 68,15,88,219 ; addps %xmm3,%xmm11 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12 - DB 68,15,89,37,9,125,1,0 ; mulps 0x17d09(%rip),%xmm12 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,29,17,125,1,0 ; andps 0x17d11(%rip),%xmm11 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,29,57,122,1,0 ; orps 0x17a39(%rip),%xmm11 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,37,17,125,1,0 ; addps 0x17d11(%rip),%xmm12 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 15,40,29,26,125,1,0 ; movaps 0x17d1a(%rip),%xmm3 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,37,73,120,1,0 ; mulps 0x17849(%rip),%xmm12 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,29,81,120,1,0 ; andps 0x17851(%rip),%xmm11 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,29,137,117,1,0 ; orps 0x17589(%rip),%xmm11 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,37,81,120,1,0 ; addps 0x17851(%rip),%xmm12 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 15,40,29,90,120,1,0 ; movaps 0x1785a(%rip),%xmm3 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,88,29,26,125,1,0 ; addps 0x17d1a(%rip),%xmm11 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 15,40,29,35,125,1,0 ; movaps 0x17d23(%rip),%xmm3 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,29,90,120,1,0 ; addps 0x1785a(%rip),%xmm11 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 15,40,29,99,120,1,0 ; movaps 0x17863(%rip),%xmm3 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 65,15,94,219 ; divps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10 DB 69,15,40,220 ; movaps %xmm12,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 68,15,88,37,16,125,1,0 ; addps 0x17d10(%rip),%xmm12 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 15,40,29,25,125,1,0 ; movaps 0x17d19(%rip),%xmm3 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,37,80,120,1,0 ; addps 0x17850(%rip),%xmm12 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 15,40,29,89,120,1,0 ; movaps 0x17859(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 65,15,89,219 ; mulps %xmm11,%xmm3 DB 68,15,92,227 ; subps %xmm3,%xmm12 - DB 68,15,40,21,25,125,1,0 ; movaps 0x17d19(%rip),%xmm10 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,21,89,120,1,0 ; movaps 0x17859(%rip),%xmm10 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,211 ; subps %xmm11,%xmm10 - DB 15,40,29,30,125,1,0 ; movaps 0x17d1e(%rip),%xmm3 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 15,40,29,94,120,1,0 ; movaps 0x1785e(%rip),%xmm3 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 65,15,94,218 ; divps %xmm10,%xmm3 DB 65,15,88,220 ; addps %xmm12,%xmm3 - DB 15,89,29,31,125,1,0 ; mulps 0x17d1f(%rip),%xmm3 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 15,89,29,95,120,1,0 ; mulps 0x1785f(%rip),%xmm3 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,68,15,91,211 ; cvtps2dq %xmm3,%xmm10 DB 243,15,16,88,20 ; movss 0x14(%rax),%xmm3 DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 @@ -20237,7 +20389,7 @@ _sk_parametric_a_sse41 LABEL PROC DB 102,65,15,56,20,217 ; blendvps %xmm0,%xmm9,%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,95,216 ; maxps %xmm0,%xmm3 - DB 15,93,29,170,121,1,0 ; minps 0x179aa(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,29,250,116,1,0 ; minps 0x174fa(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -20253,20 +20405,20 @@ _sk_gamma_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,91,194 ; cvtdq2ps %xmm2,%xmm0 - DB 15,89,5,38,124,1,0 ; mulps 0x17c26(%rip),%xmm0 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,40,53,46,124,1,0 ; movaps 0x17c2e(%rip),%xmm14 # 2d780 <_sk_overlay_sse2_8bit+0x1228> + DB 15,89,5,102,119,1,0 ; mulps 0x17766(%rip),%xmm0 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,40,53,110,119,1,0 ; movaps 0x1776e(%rip),%xmm14 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> DB 65,15,84,214 ; andps %xmm14,%xmm2 - DB 68,15,40,37,82,121,1,0 ; movaps 0x17952(%rip),%xmm12 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,37,162,116,1,0 ; movaps 0x174a2(%rip),%xmm12 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,86,212 ; orps %xmm12,%xmm2 - DB 68,15,40,21,38,124,1,0 ; movaps 0x17c26(%rip),%xmm10 # 2d790 <_sk_overlay_sse2_8bit+0x1238> + DB 68,15,40,21,102,119,1,0 ; movaps 0x17766(%rip),%xmm10 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> DB 65,15,88,194 ; addps %xmm10,%xmm0 - DB 68,15,40,29,42,124,1,0 ; movaps 0x17c2a(%rip),%xmm11 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,40,29,106,119,1,0 ; movaps 0x1776a(%rip),%xmm11 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 15,40,226 ; movaps %xmm2,%xmm4 DB 65,15,89,227 ; mulps %xmm11,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 68,15,40,13,40,124,1,0 ; movaps 0x17c28(%rip),%xmm9 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> + DB 68,15,40,13,104,119,1,0 ; movaps 0x17768(%rip),%xmm9 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> DB 65,15,88,209 ; addps %xmm9,%xmm2 - DB 68,15,40,45,44,124,1,0 ; movaps 0x17c2c(%rip),%xmm13 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,40,45,108,119,1,0 ; movaps 0x1776c(%rip),%xmm13 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 65,15,40,229 ; movaps %xmm13,%xmm4 DB 15,94,226 ; divps %xmm2,%xmm4 DB 72,173 ; lods %ds:(%rsi),%rax @@ -20277,21 +20429,21 @@ _sk_gamma_sse41 LABEL PROC DB 102,15,58,8,208,1 ; roundps $0x1,%xmm0,%xmm2 DB 15,40,224 ; movaps %xmm0,%xmm4 DB 15,92,226 ; subps %xmm2,%xmm4 - DB 15,40,53,15,124,1,0 ; movaps 0x17c0f(%rip),%xmm6 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> + DB 15,40,53,79,119,1,0 ; movaps 0x1774f(%rip),%xmm6 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> DB 15,88,198 ; addps %xmm6,%xmm0 - DB 15,40,61,37,124,1,0 ; movaps 0x17c25(%rip),%xmm7 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 15,40,61,101,119,1,0 ; movaps 0x17765(%rip),%xmm7 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 15,40,239 ; movaps %xmm7,%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 DB 15,40,212 ; movaps %xmm4,%xmm2 - DB 15,40,37,5,124,1,0 ; movaps 0x17c05(%rip),%xmm4 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 15,40,37,69,119,1,0 ; movaps 0x17745(%rip),%xmm4 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 15,89,212 ; mulps %xmm4,%xmm2 DB 15,92,194 ; subps %xmm2,%xmm0 - DB 68,15,40,61,23,124,1,0 ; movaps 0x17c17(%rip),%xmm15 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,61,87,119,1,0 ; movaps 0x17757(%rip),%xmm15 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 65,15,40,215 ; movaps %xmm15,%xmm2 DB 15,94,213 ; divps %xmm5,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 15,89,5,115,123,1,0 ; mulps 0x17b73(%rip),%xmm0 # 2d770 <_sk_overlay_sse2_8bit+0x1218> + DB 15,89,5,179,118,1,0 ; mulps 0x176b3(%rip),%xmm0 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> DB 65,15,84,206 ; andps %xmm14,%xmm1 DB 65,15,86,204 ; orps %xmm12,%xmm1 DB 65,15,88,194 ; addps %xmm10,%xmm0 @@ -20317,8 +20469,8 @@ _sk_gamma_sse41 LABEL PROC DB 15,94,206 ; divps %xmm6,%xmm1 DB 15,88,200 ; addps %xmm0,%xmm1 DB 15,91,195 ; cvtdq2ps %xmm3,%xmm0 - DB 15,89,5,20,123,1,0 ; mulps 0x17b14(%rip),%xmm0 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 15,84,29,29,123,1,0 ; andps 0x17b1d(%rip),%xmm3 # 2d780 <_sk_overlay_sse2_8bit+0x1228> + DB 15,89,5,84,118,1,0 ; mulps 0x17654(%rip),%xmm0 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 15,84,29,93,118,1,0 ; andps 0x1765d(%rip),%xmm3 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> DB 65,15,86,220 ; orps %xmm12,%xmm3 DB 65,15,88,194 ; addps %xmm10,%xmm0 DB 68,15,89,219 ; mulps %xmm3,%xmm11 @@ -20336,7 +20488,7 @@ _sk_gamma_sse41 LABEL PROC DB 15,92,253 ; subps %xmm5,%xmm7 DB 68,15,94,255 ; divps %xmm7,%xmm15 DB 68,15,88,248 ; addps %xmm0,%xmm15 - DB 15,40,5,101,123,1,0 ; movaps 0x17b65(%rip),%xmm0 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 15,40,5,165,118,1,0 ; movaps 0x176a5(%rip),%xmm0 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 68,15,89,248 ; mulps %xmm0,%xmm15 @@ -20355,29 +20507,29 @@ _sk_gamma_sse41 LABEL PROC PUBLIC _sk_lab_to_xyz_sse41 _sk_lab_to_xyz_sse41 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,89,5,50,123,1,0 ; mulps 0x17b32(%rip),%xmm8 # 2d820 <_sk_overlay_sse2_8bit+0x12c8> - DB 68,15,40,13,170,120,1,0 ; movaps 0x178aa(%rip),%xmm9 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,89,5,114,118,1,0 ; mulps 0x17672(%rip),%xmm8 # 2d700 <_sk_overlay_sse2_8bit+0x1251> + DB 68,15,40,13,250,115,1,0 ; movaps 0x173fa(%rip),%xmm9 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 65,15,89,201 ; mulps %xmm9,%xmm1 - DB 15,40,5,47,123,1,0 ; movaps 0x17b2f(%rip),%xmm0 # 2d830 <_sk_overlay_sse2_8bit+0x12d8> + DB 15,40,5,111,118,1,0 ; movaps 0x1766f(%rip),%xmm0 # 2d710 <_sk_overlay_sse2_8bit+0x1261> DB 15,88,200 ; addps %xmm0,%xmm1 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 68,15,88,5,45,123,1,0 ; addps 0x17b2d(%rip),%xmm8 # 2d840 <_sk_overlay_sse2_8bit+0x12e8> - DB 68,15,89,5,53,123,1,0 ; mulps 0x17b35(%rip),%xmm8 # 2d850 <_sk_overlay_sse2_8bit+0x12f8> - DB 15,89,13,62,123,1,0 ; mulps 0x17b3e(%rip),%xmm1 # 2d860 <_sk_overlay_sse2_8bit+0x1308> + DB 68,15,88,5,109,118,1,0 ; addps 0x1766d(%rip),%xmm8 # 2d720 <_sk_overlay_sse2_8bit+0x1271> + DB 68,15,89,5,117,118,1,0 ; mulps 0x17675(%rip),%xmm8 # 2d730 <_sk_overlay_sse2_8bit+0x1281> + DB 15,89,13,126,118,1,0 ; mulps 0x1767e(%rip),%xmm1 # 2d740 <_sk_overlay_sse2_8bit+0x1291> DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 15,89,21,67,123,1,0 ; mulps 0x17b43(%rip),%xmm2 # 2d870 <_sk_overlay_sse2_8bit+0x1318> + DB 15,89,21,131,118,1,0 ; mulps 0x17683(%rip),%xmm2 # 2d750 <_sk_overlay_sse2_8bit+0x12a1> DB 69,15,40,208 ; movaps %xmm8,%xmm10 DB 68,15,92,210 ; subps %xmm2,%xmm10 DB 68,15,40,217 ; movaps %xmm1,%xmm11 DB 69,15,89,219 ; mulps %xmm11,%xmm11 DB 68,15,89,217 ; mulps %xmm1,%xmm11 - DB 68,15,40,13,55,123,1,0 ; movaps 0x17b37(%rip),%xmm9 # 2d880 <_sk_overlay_sse2_8bit+0x1328> + DB 68,15,40,13,119,118,1,0 ; movaps 0x17677(%rip),%xmm9 # 2d760 <_sk_overlay_sse2_8bit+0x12b1> DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 15,40,21,55,123,1,0 ; movaps 0x17b37(%rip),%xmm2 # 2d890 <_sk_overlay_sse2_8bit+0x1338> + DB 15,40,21,119,118,1,0 ; movaps 0x17677(%rip),%xmm2 # 2d770 <_sk_overlay_sse2_8bit+0x12c1> DB 15,88,202 ; addps %xmm2,%xmm1 - DB 68,15,40,37,60,123,1,0 ; movaps 0x17b3c(%rip),%xmm12 # 2d8a0 <_sk_overlay_sse2_8bit+0x1348> + DB 68,15,40,37,124,118,1,0 ; movaps 0x1767c(%rip),%xmm12 # 2d780 <_sk_overlay_sse2_8bit+0x12d1> DB 65,15,89,204 ; mulps %xmm12,%xmm1 DB 102,65,15,56,20,203 ; blendvps %xmm0,%xmm11,%xmm1 DB 69,15,40,216 ; movaps %xmm8,%xmm11 @@ -20396,8 +20548,8 @@ _sk_lab_to_xyz_sse41 LABEL PROC DB 65,15,89,212 ; mulps %xmm12,%xmm2 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 102,65,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm2 - DB 15,89,13,245,122,1,0 ; mulps 0x17af5(%rip),%xmm1 # 2d8b0 <_sk_overlay_sse2_8bit+0x1358> - DB 15,89,21,254,122,1,0 ; mulps 0x17afe(%rip),%xmm2 # 2d8c0 <_sk_overlay_sse2_8bit+0x1368> + DB 15,89,13,53,118,1,0 ; mulps 0x17635(%rip),%xmm1 # 2d790 <_sk_overlay_sse2_8bit+0x12e1> + DB 15,89,21,62,118,1,0 ; mulps 0x1763e(%rip),%xmm2 # 2d7a0 <_sk_overlay_sse2_8bit+0x12f1> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,40,200 ; movaps %xmm8,%xmm1 @@ -20412,11 +20564,11 @@ _sk_load_a8_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,39 ; jne 15e0c <_sk_load_a8_sse41+0x3f> + DB 117,39 ; jne 161ac <_sk_load_a8_sse41+0x3f> DB 102,67,15,56,49,4,16 ; pmovzxbd (%r8,%r10,1),%xmm0 - DB 102,15,219,5,124,119,1,0 ; pand 0x1777c(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,5,204,114,1,0 ; pand 0x172cc(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,194,120,1,0 ; mulps 0x178c2(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,29,18,116,1,0 ; mulps 0x17412(%rip),%xmm3 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -20425,12 +20577,12 @@ _sk_load_a8_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 15e4e <_sk_load_a8_sse41+0x81> + DB 116,53 ; je 161ee <_sk_load_a8_sse41+0x81> DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 15e38 <_sk_load_a8_sse41+0x6b> + DB 116,21 ; je 161d8 <_sk_load_a8_sse41+0x6b> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,195 ; jne 15dec <_sk_load_a8_sse41+0x1f> + DB 117,195 ; jne 1618c <_sk_load_a8_sse41+0x1f> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 @@ -20438,10 +20590,10 @@ _sk_load_a8_sse41 LABEL PROC DB 102,15,110,200 ; movd %eax,%xmm1 DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1 DB 102,15,58,14,193,15 ; pblendw $0xf,%xmm1,%xmm0 - DB 235,158 ; jmp 15dec <_sk_load_a8_sse41+0x1f> + DB 235,158 ; jmp 1618c <_sk_load_a8_sse41+0x1f> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 - DB 235,147 ; jmp 15dec <_sk_load_a8_sse41+0x1f> + DB 235,147 ; jmp 1618c <_sk_load_a8_sse41+0x1f> PUBLIC _sk_load_a8_dst_sse41 _sk_load_a8_dst_sse41 LABEL PROC @@ -20452,11 +20604,11 @@ _sk_load_a8_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,39 ; jne 15e98 <_sk_load_a8_dst_sse41+0x3f> + DB 117,39 ; jne 16238 <_sk_load_a8_dst_sse41+0x3f> DB 102,67,15,56,49,36,16 ; pmovzxbd (%r8,%r10,1),%xmm4 - DB 102,15,219,37,240,118,1,0 ; pand 0x176f0(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,37,64,114,1,0 ; pand 0x17240(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,252 ; cvtdq2ps %xmm4,%xmm7 - DB 15,89,61,54,120,1,0 ; mulps 0x17836(%rip),%xmm7 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,61,134,115,1,0 ; mulps 0x17386(%rip),%xmm7 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,228 ; xorps %xmm4,%xmm4 DB 102,15,239,237 ; pxor %xmm5,%xmm5 @@ -20465,12 +20617,12 @@ _sk_load_a8_dst_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 15eda <_sk_load_a8_dst_sse41+0x81> + DB 116,53 ; je 1627a <_sk_load_a8_dst_sse41+0x81> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 15ec4 <_sk_load_a8_dst_sse41+0x6b> + DB 116,21 ; je 16264 <_sk_load_a8_dst_sse41+0x6b> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,195 ; jne 15e78 <_sk_load_a8_dst_sse41+0x1f> + DB 117,195 ; jne 16218 <_sk_load_a8_dst_sse41+0x1f> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -20478,10 +20630,10 @@ _sk_load_a8_dst_sse41 LABEL PROC DB 102,15,110,232 ; movd %eax,%xmm5 DB 102,15,56,49,237 ; pmovzxbd %xmm5,%xmm5 DB 102,15,58,14,229,15 ; pblendw $0xf,%xmm5,%xmm4 - DB 235,158 ; jmp 15e78 <_sk_load_a8_dst_sse41+0x1f> + DB 235,158 ; jmp 16218 <_sk_load_a8_dst_sse41+0x1f> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 - DB 235,147 ; jmp 15e78 <_sk_load_a8_dst_sse41+0x1f> + DB 235,147 ; jmp 16218 <_sk_load_a8_dst_sse41+0x1f> PUBLIC _sk_gather_a8_sse41 _sk_gather_a8_sse41 LABEL PROC @@ -20499,15 +20651,17 @@ _sk_gather_a8_sse41 LABEL PROC DB 102,73,15,126,194 ; movq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,58,32,4,24,0 ; pinsrb $0x0,(%r8,%r11,1),%xmm0 - DB 102,67,15,58,32,4,16,1 ; pinsrb $0x1,(%r8,%r10,1),%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 71,15,182,28,24 ; movzbl (%r8,%r11,1),%r11d + DB 102,65,15,110,195 ; movd %r11d,%xmm0 + DB 102,65,15,58,32,194,1 ; pinsrb $0x1,%r10d,%xmm0 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 102,65,15,58,32,193,2 ; pinsrb $0x2,%r9d,%xmm0 DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax DB 102,15,58,32,192,3 ; pinsrb $0x3,%eax,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,107,119,1,0 ; mulps 0x1776b(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,29,181,114,1,0 ; mulps 0x172b5(%rip),%xmm3 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -20523,13 +20677,13 @@ _sk_store_a8_sse41 LABEL PROC DB 77,15,175,193 ; imul %r9,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 - DB 68,15,40,5,29,118,1,0 ; movaps 0x1761d(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,103,113,1,0 ; movaps 0x17167(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 68,15,89,195 ; mulps %xmm3,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 - DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 - DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 + DB 102,68,15,56,43,192 ; packusdw %xmm0,%xmm8 + DB 102,68,15,103,192 ; packuswb %xmm0,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 15fad <_sk_store_a8_sse41+0x49> + DB 117,17 ; jne 16353 <_sk_store_a8_sse41+0x49> DB 102,68,15,126,192 ; movd %xmm8,%eax DB 67,137,4,16 ; mov %eax,(%r8,%r10,1) DB 72,173 ; lods %ds:(%rsi),%rax @@ -20539,17 +20693,17 @@ _sk_store_a8_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,41 ; je 15fe9 <_sk_store_a8_sse41+0x85> + DB 116,41 ; je 1638f <_sk_store_a8_sse41+0x85> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,15 ; je 15fd5 <_sk_store_a8_sse41+0x71> + DB 116,15 ; je 1637b <_sk_store_a8_sse41+0x71> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,217 ; jne 15fa5 <_sk_store_a8_sse41+0x41> + DB 117,217 ; jne 1634b <_sk_store_a8_sse41+0x41> DB 102,71,15,58,20,68,16,2,8 ; pextrb $0x8,%xmm8,0x2(%r8,%r10,1) - DB 102,68,15,56,0,5,241,120,1,0 ; pshufb 0x178f1(%rip),%xmm8 # 2d8d0 <_sk_overlay_sse2_8bit+0x1378> + DB 102,68,15,56,0,5,43,116,1,0 ; pshufb 0x1742b(%rip),%xmm8 # 2d7b0 <_sk_overlay_sse2_8bit+0x1301> DB 102,71,15,58,21,4,16,0 ; pextrw $0x0,%xmm8,(%r8,%r10,1) - DB 235,188 ; jmp 15fa5 <_sk_store_a8_sse41+0x41> + DB 235,188 ; jmp 1634b <_sk_store_a8_sse41+0x41> DB 102,71,15,58,20,4,16,0 ; pextrb $0x0,%xmm8,(%r8,%r10,1) - DB 235,178 ; jmp 15fa5 <_sk_store_a8_sse41+0x41> + DB 235,178 ; jmp 1634b <_sk_store_a8_sse41+0x41> PUBLIC _sk_load_g8_sse41 _sk_load_g8_sse41 LABEL PROC @@ -20560,25 +20714,25 @@ _sk_load_g8_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,42 ; jne 16035 <_sk_load_g8_sse41+0x42> + DB 117,42 ; jne 163db <_sk_load_g8_sse41+0x42> DB 102,67,15,56,49,4,16 ; pmovzxbd (%r8,%r10,1),%xmm0 - DB 102,15,219,5,86,117,1,0 ; pand 0x17556(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,5,160,112,1,0 ; pand 0x170a0(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,156,118,1,0 ; mulps 0x1769c(%rip),%xmm0 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,5,230,113,1,0 ; mulps 0x171e6(%rip),%xmm0 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,147,116,1,0 ; movaps 0x17493(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,221,111,1,0 ; movaps 0x16fdd(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 16077 <_sk_load_g8_sse41+0x84> + DB 116,53 ; je 1641d <_sk_load_g8_sse41+0x84> DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 16061 <_sk_load_g8_sse41+0x6e> + DB 116,21 ; je 16407 <_sk_load_g8_sse41+0x6e> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,192 ; jne 16012 <_sk_load_g8_sse41+0x1f> + DB 117,192 ; jne 163b8 <_sk_load_g8_sse41+0x1f> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 @@ -20586,10 +20740,10 @@ _sk_load_g8_sse41 LABEL PROC DB 102,15,110,200 ; movd %eax,%xmm1 DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1 DB 102,15,58,14,193,15 ; pblendw $0xf,%xmm1,%xmm0 - DB 235,155 ; jmp 16012 <_sk_load_g8_sse41+0x1f> + DB 235,155 ; jmp 163b8 <_sk_load_g8_sse41+0x1f> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 - DB 235,144 ; jmp 16012 <_sk_load_g8_sse41+0x1f> + DB 235,144 ; jmp 163b8 <_sk_load_g8_sse41+0x1f> PUBLIC _sk_load_g8_dst_sse41 _sk_load_g8_dst_sse41 LABEL PROC @@ -20600,25 +20754,25 @@ _sk_load_g8_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,42 ; jne 160c4 <_sk_load_g8_dst_sse41+0x42> + DB 117,42 ; jne 1646a <_sk_load_g8_dst_sse41+0x42> DB 102,67,15,56,49,36,16 ; pmovzxbd (%r8,%r10,1),%xmm4 - DB 102,15,219,37,199,116,1,0 ; pand 0x174c7(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,37,17,112,1,0 ; pand 0x17011(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,37,13,118,1,0 ; mulps 0x1760d(%rip),%xmm4 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,37,87,113,1,0 ; mulps 0x17157(%rip),%xmm4 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,61,4,116,1,0 ; movaps 0x17404(%rip),%xmm7 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,61,78,111,1,0 ; movaps 0x16f4e(%rip),%xmm7 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,40,244 ; movaps %xmm4,%xmm6 DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 16106 <_sk_load_g8_dst_sse41+0x84> + DB 116,53 ; je 164ac <_sk_load_g8_dst_sse41+0x84> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 160f0 <_sk_load_g8_dst_sse41+0x6e> + DB 116,21 ; je 16496 <_sk_load_g8_dst_sse41+0x6e> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,192 ; jne 160a1 <_sk_load_g8_dst_sse41+0x1f> + DB 117,192 ; jne 16447 <_sk_load_g8_dst_sse41+0x1f> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -20626,10 +20780,10 @@ _sk_load_g8_dst_sse41 LABEL PROC DB 102,15,110,232 ; movd %eax,%xmm5 DB 102,15,56,49,237 ; pmovzxbd %xmm5,%xmm5 DB 102,15,58,14,229,15 ; pblendw $0xf,%xmm5,%xmm4 - DB 235,155 ; jmp 160a1 <_sk_load_g8_dst_sse41+0x1f> + DB 235,155 ; jmp 16447 <_sk_load_g8_dst_sse41+0x1f> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 - DB 235,144 ; jmp 160a1 <_sk_load_g8_dst_sse41+0x1f> + DB 235,144 ; jmp 16447 <_sk_load_g8_dst_sse41+0x1f> PUBLIC _sk_gather_g8_sse41 _sk_gather_g8_sse41 LABEL PROC @@ -20647,17 +20801,19 @@ _sk_gather_g8_sse41 LABEL PROC DB 102,73,15,126,194 ; movq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,58,32,4,24,0 ; pinsrb $0x0,(%r8,%r11,1),%xmm0 - DB 102,67,15,58,32,4,16,1 ; pinsrb $0x1,(%r8,%r10,1),%xmm0 + DB 71,15,182,20,16 ; movzbl (%r8,%r10,1),%r10d + DB 71,15,182,28,24 ; movzbl (%r8,%r11,1),%r11d + DB 102,65,15,110,195 ; movd %r11d,%xmm0 + DB 102,65,15,58,32,194,1 ; pinsrb $0x1,%r10d,%xmm0 DB 71,15,182,12,8 ; movzbl (%r8,%r9,1),%r9d DB 102,65,15,58,32,193,2 ; pinsrb $0x2,%r9d,%xmm0 DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax DB 102,15,58,32,192,3 ; pinsrb $0x3,%eax,%xmm0 DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,63,117,1,0 ; mulps 0x1753f(%rip),%xmm0 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,5,131,112,1,0 ; mulps 0x17083(%rip),%xmm0 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,54,115,1,0 ; movaps 0x17336(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,122,110,1,0 ; movaps 0x16e7a(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -20672,41 +20828,41 @@ _sk_load_565_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,80 ; jne 161fd <_sk_load_565_sse41+0x6b> + DB 117,80 ; jne 165a9 <_sk_load_565_sse41+0x6b> DB 102,67,15,56,51,20,80 ; pmovzxwd (%r8,%r10,2),%xmm2 - DB 102,15,111,5,20,117,1,0 ; movdqa 0x17514(%rip),%xmm0 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,5,88,112,1,0 ; movdqa 0x17058(%rip),%xmm0 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,22,117,1,0 ; mulps 0x17516(%rip),%xmm0 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,13,30,117,1,0 ; movdqa 0x1751e(%rip),%xmm1 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 15,89,5,90,112,1,0 ; mulps 0x1705a(%rip),%xmm0 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,13,98,112,1,0 ; movdqa 0x17062(%rip),%xmm1 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,32,117,1,0 ; mulps 0x17520(%rip),%xmm1 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,15,219,21,40,117,1,0 ; pand 0x17528(%rip),%xmm2 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 15,89,13,100,112,1,0 ; mulps 0x17064(%rip),%xmm1 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,15,219,21,108,112,1,0 ; pand 0x1706c(%rip),%xmm2 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,46,117,1,0 ; mulps 0x1752e(%rip),%xmm2 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 15,89,21,114,112,1,0 ; mulps 0x17072(%rip),%xmm2 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,197,114,1,0 ; movaps 0x172c5(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,9,110,1,0 ; movaps 0x16e09(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 1623f <_sk_load_565_sse41+0xad> + DB 116,53 ; je 165eb <_sk_load_565_sse41+0xad> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 16229 <_sk_load_565_sse41+0x97> + DB 116,21 ; je 165d5 <_sk_load_565_sse41+0x97> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,154 ; jne 161b4 <_sk_load_565_sse41+0x22> + DB 117,154 ; jne 16560 <_sk_load_565_sse41+0x22> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,208,69 ; pshufd $0x45,%xmm0,%xmm2 DB 102,67,15,110,4,80 ; movd (%r8,%r10,2),%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 DB 102,15,58,14,208,15 ; pblendw $0xf,%xmm0,%xmm2 - DB 233,117,255,255,255 ; jmpq 161b4 <_sk_load_565_sse41+0x22> + DB 233,117,255,255,255 ; jmpq 16560 <_sk_load_565_sse41+0x22> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,208 ; movd %eax,%xmm2 - DB 233,103,255,255,255 ; jmpq 161b4 <_sk_load_565_sse41+0x22> + DB 233,103,255,255,255 ; jmpq 16560 <_sk_load_565_sse41+0x22> PUBLIC _sk_load_565_dst_sse41 _sk_load_565_dst_sse41 LABEL PROC @@ -20718,41 +20874,41 @@ _sk_load_565_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,80 ; jne 162b8 <_sk_load_565_dst_sse41+0x6b> + DB 117,80 ; jne 16664 <_sk_load_565_dst_sse41+0x6b> DB 102,67,15,56,51,52,80 ; pmovzxwd (%r8,%r10,2),%xmm6 - DB 102,15,111,37,89,116,1,0 ; movdqa 0x17459(%rip),%xmm4 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,37,157,111,1,0 ; movdqa 0x16f9d(%rip),%xmm4 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,37,91,116,1,0 ; mulps 0x1745b(%rip),%xmm4 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,45,99,116,1,0 ; movdqa 0x17463(%rip),%xmm5 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 15,89,37,159,111,1,0 ; mulps 0x16f9f(%rip),%xmm4 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,45,167,111,1,0 ; movdqa 0x16fa7(%rip),%xmm5 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,219,238 ; pand %xmm6,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,45,101,116,1,0 ; mulps 0x17465(%rip),%xmm5 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,15,219,53,109,116,1,0 ; pand 0x1746d(%rip),%xmm6 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 15,89,45,169,111,1,0 ; mulps 0x16fa9(%rip),%xmm5 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,15,219,53,177,111,1,0 ; pand 0x16fb1(%rip),%xmm6 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,89,53,115,116,1,0 ; mulps 0x17473(%rip),%xmm6 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 15,89,53,183,111,1,0 ; mulps 0x16fb7(%rip),%xmm6 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,61,10,114,1,0 ; movaps 0x1720a(%rip),%xmm7 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,61,78,109,1,0 ; movaps 0x16d4e(%rip),%xmm7 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 162fa <_sk_load_565_dst_sse41+0xad> + DB 116,53 ; je 166a6 <_sk_load_565_dst_sse41+0xad> DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 162e4 <_sk_load_565_dst_sse41+0x97> + DB 116,21 ; je 16690 <_sk_load_565_dst_sse41+0x97> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,154 ; jne 1626f <_sk_load_565_dst_sse41+0x22> + DB 117,154 ; jne 1661b <_sk_load_565_dst_sse41+0x22> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,244,69 ; pshufd $0x45,%xmm4,%xmm6 DB 102,67,15,110,36,80 ; movd (%r8,%r10,2),%xmm4 DB 102,15,56,51,228 ; pmovzxwd %xmm4,%xmm4 DB 102,15,58,14,244,15 ; pblendw $0xf,%xmm4,%xmm6 - DB 233,117,255,255,255 ; jmpq 1626f <_sk_load_565_dst_sse41+0x22> + DB 233,117,255,255,255 ; jmpq 1661b <_sk_load_565_dst_sse41+0x22> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,240 ; movd %eax,%xmm6 - DB 233,103,255,255,255 ; jmpq 1626f <_sk_load_565_dst_sse41+0x22> + DB 233,103,255,255,255 ; jmpq 1661b <_sk_load_565_dst_sse41+0x22> PUBLIC _sk_gather_565_sse41 _sk_gather_565_sse41 LABEL PROC @@ -20770,26 +20926,28 @@ _sk_gather_565_sse41 LABEL PROC DB 102,73,15,126,194 ; movq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,4,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm0 - DB 102,67,15,196,4,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm0 + DB 71,15,183,20,80 ; movzwl (%r8,%r10,2),%r10d + DB 71,15,183,28,88 ; movzwl (%r8,%r11,2),%r11d + DB 102,65,15,110,195 ; movd %r11d,%xmm0 + DB 102,65,15,196,194,1 ; pinsrw $0x1,%r10d,%xmm0 DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d DB 102,65,15,196,193,2 ; pinsrw $0x2,%r9d,%xmm0 DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax DB 102,15,196,192,3 ; pinsrw $0x3,%eax,%xmm0 DB 102,15,56,51,208 ; pmovzxwd %xmm0,%xmm2 - DB 102,15,111,5,94,115,1,0 ; movdqa 0x1735e(%rip),%xmm0 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,5,155,110,1,0 ; movdqa 0x16e9b(%rip),%xmm0 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,96,115,1,0 ; mulps 0x17360(%rip),%xmm0 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,13,104,115,1,0 ; movdqa 0x17368(%rip),%xmm1 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 15,89,5,157,110,1,0 ; mulps 0x16e9d(%rip),%xmm0 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,13,165,110,1,0 ; movdqa 0x16ea5(%rip),%xmm1 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,106,115,1,0 ; mulps 0x1736a(%rip),%xmm1 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,15,219,21,114,115,1,0 ; pand 0x17372(%rip),%xmm2 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 15,89,13,167,110,1,0 ; mulps 0x16ea7(%rip),%xmm1 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,15,219,21,175,110,1,0 ; pand 0x16eaf(%rip),%xmm2 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,120,115,1,0 ; mulps 0x17378(%rip),%xmm2 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 15,89,21,181,110,1,0 ; mulps 0x16eb5(%rip),%xmm2 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,15,113,1,0 ; movaps 0x1710f(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,76,108,1,0 ; movaps 0x16c4c(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_sse41 @@ -20801,12 +20959,12 @@ _sk_store_565_sse41 LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 68,15,40,5,15,117,1,0 ; movaps 0x1750f(%rip),%xmm8 # 2d8e0 <_sk_overlay_sse2_8bit+0x1388> + DB 68,15,40,5,60,112,1,0 ; movaps 0x1703c(%rip),%xmm8 # 2d7c0 <_sk_overlay_sse2_8bit+0x1311> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 DB 102,65,15,114,241,11 ; pslld $0xb,%xmm9 - DB 68,15,40,21,4,117,1,0 ; movaps 0x17504(%rip),%xmm10 # 2d8f0 <_sk_overlay_sse2_8bit+0x1398> + DB 68,15,40,21,49,112,1,0 ; movaps 0x17031(%rip),%xmm10 # 2d7d0 <_sk_overlay_sse2_8bit+0x1321> DB 68,15,89,209 ; mulps %xmm1,%xmm10 DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10 DB 102,65,15,114,242,5 ; pslld $0x5,%xmm10 @@ -20814,27 +20972,27 @@ _sk_store_565_sse41 LABEL PROC DB 68,15,89,194 ; mulps %xmm2,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,69,15,86,194 ; orpd %xmm10,%xmm8 - DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 + DB 102,68,15,56,43,192 ; packusdw %xmm0,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 16423 <_sk_store_565_sse41+0x70> - DB 242,69,15,17,4,64 ; movsd %xmm8,(%r8,%rax,2) + DB 117,10 ; jne 167d6 <_sk_store_565_sse41+0x70> + DB 102,69,15,214,4,64 ; movq %xmm8,(%r8,%rax,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,69,15,56,51,192 ; pmovzxwd %xmm8,%xmm8 DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,35 ; je 16459 <_sk_store_565_sse41+0xa6> + DB 116,35 ; je 1680c <_sk_store_565_sse41+0xa6> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,15 ; je 1644b <_sk_store_565_sse41+0x98> + DB 116,15 ; je 167fe <_sk_store_565_sse41+0x98> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,221 ; jne 1641f <_sk_store_565_sse41+0x6c> + DB 117,221 ; jne 167d2 <_sk_store_565_sse41+0x6c> DB 102,69,15,58,21,68,64,4,4 ; pextrw $0x4,%xmm8,0x4(%r8,%rax,2) DB 242,69,15,112,192,232 ; pshuflw $0xe8,%xmm8,%xmm8 DB 102,69,15,126,4,64 ; movd %xmm8,(%r8,%rax,2) - DB 235,198 ; jmp 1641f <_sk_store_565_sse41+0x6c> + DB 235,198 ; jmp 167d2 <_sk_store_565_sse41+0x6c> DB 102,69,15,58,21,4,64,0 ; pextrw $0x0,%xmm8,(%r8,%rax,2) - DB 235,188 ; jmp 1641f <_sk_store_565_sse41+0x6c> + DB 235,188 ; jmp 167d2 <_sk_store_565_sse41+0x6c> PUBLIC _sk_load_4444_sse41 _sk_load_4444_sse41 LABEL PROC @@ -20846,44 +21004,44 @@ _sk_load_4444_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,95 ; jne 164dd <_sk_load_4444_sse41+0x7a> + DB 117,95 ; jne 16890 <_sk_load_4444_sse41+0x7a> DB 102,67,15,56,51,28,80 ; pmovzxwd (%r8,%r10,2),%xmm3 - DB 102,15,111,5,115,116,1,0 ; movdqa 0x17473(%rip),%xmm0 # 2d900 <_sk_overlay_sse2_8bit+0x13a8> + DB 102,15,111,5,160,111,1,0 ; movdqa 0x16fa0(%rip),%xmm0 # 2d7e0 <_sk_overlay_sse2_8bit+0x1331> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,117,116,1,0 ; mulps 0x17475(%rip),%xmm0 # 2d910 <_sk_overlay_sse2_8bit+0x13b8> - DB 102,15,111,13,125,116,1,0 ; movdqa 0x1747d(%rip),%xmm1 # 2d920 <_sk_overlay_sse2_8bit+0x13c8> + DB 15,89,5,162,111,1,0 ; mulps 0x16fa2(%rip),%xmm0 # 2d7f0 <_sk_overlay_sse2_8bit+0x1341> + DB 102,15,111,13,170,111,1,0 ; movdqa 0x16faa(%rip),%xmm1 # 2d800 <_sk_overlay_sse2_8bit+0x1351> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,127,116,1,0 ; mulps 0x1747f(%rip),%xmm1 # 2d930 <_sk_overlay_sse2_8bit+0x13d8> - DB 102,15,111,21,135,116,1,0 ; movdqa 0x17487(%rip),%xmm2 # 2d940 <_sk_overlay_sse2_8bit+0x13e8> + DB 15,89,13,172,111,1,0 ; mulps 0x16fac(%rip),%xmm1 # 2d810 <_sk_overlay_sse2_8bit+0x1361> + DB 102,15,111,21,180,111,1,0 ; movdqa 0x16fb4(%rip),%xmm2 # 2d820 <_sk_overlay_sse2_8bit+0x1371> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,137,116,1,0 ; mulps 0x17489(%rip),%xmm2 # 2d950 <_sk_overlay_sse2_8bit+0x13f8> - DB 102,15,219,29,145,116,1,0 ; pand 0x17491(%rip),%xmm3 # 2d960 <_sk_overlay_sse2_8bit+0x1408> + DB 15,89,21,182,111,1,0 ; mulps 0x16fb6(%rip),%xmm2 # 2d830 <_sk_overlay_sse2_8bit+0x1381> + DB 102,15,219,29,190,111,1,0 ; pand 0x16fbe(%rip),%xmm3 # 2d840 <_sk_overlay_sse2_8bit+0x1391> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,151,116,1,0 ; mulps 0x17497(%rip),%xmm3 # 2d970 <_sk_overlay_sse2_8bit+0x1418> + DB 15,89,29,196,111,1,0 ; mulps 0x16fc4(%rip),%xmm3 # 2d850 <_sk_overlay_sse2_8bit+0x13a1> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 1651f <_sk_load_4444_sse41+0xbc> + DB 116,53 ; je 168d2 <_sk_load_4444_sse41+0xbc> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 16509 <_sk_load_4444_sse41+0xa6> + DB 116,21 ; je 168bc <_sk_load_4444_sse41+0xa6> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,139 ; jne 16485 <_sk_load_4444_sse41+0x22> + DB 117,139 ; jne 16838 <_sk_load_4444_sse41+0x22> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 102,67,15,110,4,80 ; movd (%r8,%r10,2),%xmm0 DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0 DB 102,15,58,14,216,15 ; pblendw $0xf,%xmm0,%xmm3 - DB 233,102,255,255,255 ; jmpq 16485 <_sk_load_4444_sse41+0x22> + DB 233,102,255,255,255 ; jmpq 16838 <_sk_load_4444_sse41+0x22> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,216 ; movd %eax,%xmm3 - DB 233,88,255,255,255 ; jmpq 16485 <_sk_load_4444_sse41+0x22> + DB 233,88,255,255,255 ; jmpq 16838 <_sk_load_4444_sse41+0x22> PUBLIC _sk_load_4444_dst_sse41 _sk_load_4444_dst_sse41 LABEL PROC @@ -20895,44 +21053,44 @@ _sk_load_4444_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,95 ; jne 165a7 <_sk_load_4444_dst_sse41+0x7a> + DB 117,95 ; jne 1695a <_sk_load_4444_dst_sse41+0x7a> DB 102,67,15,56,51,60,80 ; pmovzxwd (%r8,%r10,2),%xmm7 - DB 102,15,111,37,169,115,1,0 ; movdqa 0x173a9(%rip),%xmm4 # 2d900 <_sk_overlay_sse2_8bit+0x13a8> + DB 102,15,111,37,214,110,1,0 ; movdqa 0x16ed6(%rip),%xmm4 # 2d7e0 <_sk_overlay_sse2_8bit+0x1331> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,37,171,115,1,0 ; mulps 0x173ab(%rip),%xmm4 # 2d910 <_sk_overlay_sse2_8bit+0x13b8> - DB 102,15,111,45,179,115,1,0 ; movdqa 0x173b3(%rip),%xmm5 # 2d920 <_sk_overlay_sse2_8bit+0x13c8> + DB 15,89,37,216,110,1,0 ; mulps 0x16ed8(%rip),%xmm4 # 2d7f0 <_sk_overlay_sse2_8bit+0x1341> + DB 102,15,111,45,224,110,1,0 ; movdqa 0x16ee0(%rip),%xmm5 # 2d800 <_sk_overlay_sse2_8bit+0x1351> DB 102,15,219,239 ; pand %xmm7,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,45,181,115,1,0 ; mulps 0x173b5(%rip),%xmm5 # 2d930 <_sk_overlay_sse2_8bit+0x13d8> - DB 102,15,111,53,189,115,1,0 ; movdqa 0x173bd(%rip),%xmm6 # 2d940 <_sk_overlay_sse2_8bit+0x13e8> + DB 15,89,45,226,110,1,0 ; mulps 0x16ee2(%rip),%xmm5 # 2d810 <_sk_overlay_sse2_8bit+0x1361> + DB 102,15,111,53,234,110,1,0 ; movdqa 0x16eea(%rip),%xmm6 # 2d820 <_sk_overlay_sse2_8bit+0x1371> DB 102,15,219,247 ; pand %xmm7,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,89,53,191,115,1,0 ; mulps 0x173bf(%rip),%xmm6 # 2d950 <_sk_overlay_sse2_8bit+0x13f8> - DB 102,15,219,61,199,115,1,0 ; pand 0x173c7(%rip),%xmm7 # 2d960 <_sk_overlay_sse2_8bit+0x1408> + DB 15,89,53,236,110,1,0 ; mulps 0x16eec(%rip),%xmm6 # 2d830 <_sk_overlay_sse2_8bit+0x1381> + DB 102,15,219,61,244,110,1,0 ; pand 0x16ef4(%rip),%xmm7 # 2d840 <_sk_overlay_sse2_8bit+0x1391> DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 - DB 15,89,61,205,115,1,0 ; mulps 0x173cd(%rip),%xmm7 # 2d970 <_sk_overlay_sse2_8bit+0x1418> + DB 15,89,61,250,110,1,0 ; mulps 0x16efa(%rip),%xmm7 # 2d850 <_sk_overlay_sse2_8bit+0x13a1> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,53 ; je 165e9 <_sk_load_4444_dst_sse41+0xbc> + DB 116,53 ; je 1699c <_sk_load_4444_dst_sse41+0xbc> DB 102,15,239,255 ; pxor %xmm7,%xmm7 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 165d3 <_sk_load_4444_dst_sse41+0xa6> + DB 116,21 ; je 16986 <_sk_load_4444_dst_sse41+0xa6> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,139 ; jne 1654f <_sk_load_4444_dst_sse41+0x22> + DB 117,139 ; jne 16902 <_sk_load_4444_dst_sse41+0x22> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,252,69 ; pshufd $0x45,%xmm4,%xmm7 DB 102,67,15,110,36,80 ; movd (%r8,%r10,2),%xmm4 DB 102,15,56,51,228 ; pmovzxwd %xmm4,%xmm4 DB 102,15,58,14,252,15 ; pblendw $0xf,%xmm4,%xmm7 - DB 233,102,255,255,255 ; jmpq 1654f <_sk_load_4444_dst_sse41+0x22> + DB 233,102,255,255,255 ; jmpq 16902 <_sk_load_4444_dst_sse41+0x22> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,248 ; movd %eax,%xmm7 - DB 233,88,255,255,255 ; jmpq 1654f <_sk_load_4444_dst_sse41+0x22> + DB 233,88,255,255,255 ; jmpq 16902 <_sk_load_4444_dst_sse41+0x22> PUBLIC _sk_gather_4444_sse41 _sk_gather_4444_sse41 LABEL PROC @@ -20950,28 +21108,30 @@ _sk_gather_4444_sse41 LABEL PROC DB 102,73,15,126,194 ; movq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,4,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm0 - DB 102,67,15,196,4,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm0 + DB 71,15,183,20,80 ; movzwl (%r8,%r10,2),%r10d + DB 71,15,183,28,88 ; movzwl (%r8,%r11,2),%r11d + DB 102,65,15,110,195 ; movd %r11d,%xmm0 + DB 102,65,15,196,194,1 ; pinsrw $0x1,%r10d,%xmm0 DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d DB 102,65,15,196,193,2 ; pinsrw $0x2,%r9d,%xmm0 DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax DB 102,15,196,192,3 ; pinsrw $0x3,%eax,%xmm0 DB 102,15,56,51,216 ; pmovzxwd %xmm0,%xmm3 - DB 102,15,111,5,159,114,1,0 ; movdqa 0x1729f(%rip),%xmm0 # 2d900 <_sk_overlay_sse2_8bit+0x13a8> + DB 102,15,111,5,197,109,1,0 ; movdqa 0x16dc5(%rip),%xmm0 # 2d7e0 <_sk_overlay_sse2_8bit+0x1331> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,161,114,1,0 ; mulps 0x172a1(%rip),%xmm0 # 2d910 <_sk_overlay_sse2_8bit+0x13b8> - DB 102,15,111,13,169,114,1,0 ; movdqa 0x172a9(%rip),%xmm1 # 2d920 <_sk_overlay_sse2_8bit+0x13c8> + DB 15,89,5,199,109,1,0 ; mulps 0x16dc7(%rip),%xmm0 # 2d7f0 <_sk_overlay_sse2_8bit+0x1341> + DB 102,15,111,13,207,109,1,0 ; movdqa 0x16dcf(%rip),%xmm1 # 2d800 <_sk_overlay_sse2_8bit+0x1351> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,171,114,1,0 ; mulps 0x172ab(%rip),%xmm1 # 2d930 <_sk_overlay_sse2_8bit+0x13d8> - DB 102,15,111,21,179,114,1,0 ; movdqa 0x172b3(%rip),%xmm2 # 2d940 <_sk_overlay_sse2_8bit+0x13e8> + DB 15,89,13,209,109,1,0 ; mulps 0x16dd1(%rip),%xmm1 # 2d810 <_sk_overlay_sse2_8bit+0x1361> + DB 102,15,111,21,217,109,1,0 ; movdqa 0x16dd9(%rip),%xmm2 # 2d820 <_sk_overlay_sse2_8bit+0x1371> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,181,114,1,0 ; mulps 0x172b5(%rip),%xmm2 # 2d950 <_sk_overlay_sse2_8bit+0x13f8> - DB 102,15,219,29,189,114,1,0 ; pand 0x172bd(%rip),%xmm3 # 2d960 <_sk_overlay_sse2_8bit+0x1408> + DB 15,89,21,219,109,1,0 ; mulps 0x16ddb(%rip),%xmm2 # 2d830 <_sk_overlay_sse2_8bit+0x1381> + DB 102,15,219,29,227,109,1,0 ; pand 0x16de3(%rip),%xmm3 # 2d840 <_sk_overlay_sse2_8bit+0x1391> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,195,114,1,0 ; mulps 0x172c3(%rip),%xmm3 # 2d970 <_sk_overlay_sse2_8bit+0x1418> + DB 15,89,29,233,109,1,0 ; mulps 0x16de9(%rip),%xmm3 # 2d850 <_sk_overlay_sse2_8bit+0x13a1> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -20984,7 +21144,7 @@ _sk_store_4444_sse41 LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 68,15,40,5,177,114,1,0 ; movaps 0x172b1(%rip),%xmm8 # 2d980 <_sk_overlay_sse2_8bit+0x1428> + DB 68,15,40,5,215,109,1,0 ; movaps 0x16dd7(%rip),%xmm8 # 2d860 <_sk_overlay_sse2_8bit+0x13b1> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -21002,27 +21162,27 @@ _sk_store_4444_sse41 LABEL PROC DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,69,15,86,193 ; orpd %xmm9,%xmm8 DB 102,69,15,86,194 ; orpd %xmm10,%xmm8 - DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 + DB 102,68,15,56,43,192 ; packusdw %xmm0,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 16735 <_sk_store_4444_sse41+0x84> - DB 242,69,15,17,4,64 ; movsd %xmm8,(%r8,%rax,2) + DB 117,10 ; jne 16aef <_sk_store_4444_sse41+0x84> + DB 102,69,15,214,4,64 ; movq %xmm8,(%r8,%rax,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,69,15,56,51,192 ; pmovzxwd %xmm8,%xmm8 DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,35 ; je 1676b <_sk_store_4444_sse41+0xba> + DB 116,35 ; je 16b25 <_sk_store_4444_sse41+0xba> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,15 ; je 1675d <_sk_store_4444_sse41+0xac> + DB 116,15 ; je 16b17 <_sk_store_4444_sse41+0xac> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,221 ; jne 16731 <_sk_store_4444_sse41+0x80> + DB 117,221 ; jne 16aeb <_sk_store_4444_sse41+0x80> DB 102,69,15,58,21,68,64,4,4 ; pextrw $0x4,%xmm8,0x4(%r8,%rax,2) DB 242,69,15,112,192,232 ; pshuflw $0xe8,%xmm8,%xmm8 DB 102,69,15,126,4,64 ; movd %xmm8,(%r8,%rax,2) - DB 235,198 ; jmp 16731 <_sk_store_4444_sse41+0x80> + DB 235,198 ; jmp 16aeb <_sk_store_4444_sse41+0x80> DB 102,69,15,58,21,4,64,0 ; pextrw $0x0,%xmm8,(%r8,%rax,2) - DB 235,188 ; jmp 16731 <_sk_store_4444_sse41+0x80> + DB 235,188 ; jmp 16aeb <_sk_store_4444_sse41+0x80> PUBLIC _sk_load_8888_sse41 _sk_load_8888_sse41 LABEL PROC @@ -21034,19 +21194,19 @@ _sk_load_8888_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,89 ; jne 167ea <_sk_load_8888_sse41+0x75> + DB 117,89 ; jne 16ba4 <_sk_load_8888_sse41+0x75> DB 243,65,15,111,28,128 ; movdqu (%r8,%rax,4),%xmm3 - DB 102,15,111,5,209,109,1,0 ; movdqa 0x16dd1(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,5,7,105,1,0 ; movdqa 0x16907(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,18,111,1,0 ; movaps 0x16f12(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,72,106,1,0 ; movaps 0x16a48(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,193,109,1,0 ; pshufb 0x16dc1(%rip),%xmm1 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,13,247,104,1,0 ; pshufb 0x168f7(%rip),%xmm1 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,189,109,1,0 ; pshufb 0x16dbd(%rip),%xmm2 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,21,243,104,1,0 ; pshufb 0x168f3(%rip),%xmm2 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -21057,19 +21217,19 @@ _sk_load_8888_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,45 ; je 16824 <_sk_load_8888_sse41+0xaf> + DB 116,45 ; je 16bde <_sk_load_8888_sse41+0xaf> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 16813 <_sk_load_8888_sse41+0x9e> + DB 116,18 ; je 16bcd <_sk_load_8888_sse41+0x9e> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,144 ; jne 16797 <_sk_load_8888_sse41+0x22> + DB 117,144 ; jne 16b51 <_sk_load_8888_sse41+0x22> DB 102,65,15,110,68,128,8 ; movd 0x8(%r8,%rax,4),%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 243,65,15,126,4,128 ; movq (%r8,%rax,4),%xmm0 DB 102,15,58,14,216,15 ; pblendw $0xf,%xmm0,%xmm3 - DB 233,115,255,255,255 ; jmpq 16797 <_sk_load_8888_sse41+0x22> + DB 233,115,255,255,255 ; jmpq 16b51 <_sk_load_8888_sse41+0x22> DB 102,65,15,110,28,128 ; movd (%r8,%rax,4),%xmm3 - DB 233,104,255,255,255 ; jmpq 16797 <_sk_load_8888_sse41+0x22> + DB 233,104,255,255,255 ; jmpq 16b51 <_sk_load_8888_sse41+0x22> PUBLIC _sk_load_8888_dst_sse41 _sk_load_8888_dst_sse41 LABEL PROC @@ -21081,19 +21241,19 @@ _sk_load_8888_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,89 ; jne 168a4 <_sk_load_8888_dst_sse41+0x75> + DB 117,89 ; jne 16c5e <_sk_load_8888_dst_sse41+0x75> DB 243,65,15,111,60,128 ; movdqu (%r8,%rax,4),%xmm7 - DB 102,15,111,37,23,109,1,0 ; movdqa 0x16d17(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,37,77,104,1,0 ; movdqa 0x1684d(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 68,15,40,5,88,110,1,0 ; movaps 0x16e58(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,142,105,1,0 ; movaps 0x1698e(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,224 ; mulps %xmm8,%xmm4 DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,56,0,45,7,109,1,0 ; pshufb 0x16d07(%rip),%xmm5 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,45,61,104,1,0 ; pshufb 0x1683d(%rip),%xmm5 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 DB 65,15,89,232 ; mulps %xmm8,%xmm5 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,56,0,53,3,109,1,0 ; pshufb 0x16d03(%rip),%xmm6 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,53,57,104,1,0 ; pshufb 0x16839(%rip),%xmm6 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 DB 65,15,89,240 ; mulps %xmm8,%xmm6 DB 102,15,114,215,24 ; psrld $0x18,%xmm7 @@ -21104,19 +21264,19 @@ _sk_load_8888_dst_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,45 ; je 168de <_sk_load_8888_dst_sse41+0xaf> + DB 116,45 ; je 16c98 <_sk_load_8888_dst_sse41+0xaf> DB 102,15,239,255 ; pxor %xmm7,%xmm7 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 168cd <_sk_load_8888_dst_sse41+0x9e> + DB 116,18 ; je 16c87 <_sk_load_8888_dst_sse41+0x9e> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,144 ; jne 16851 <_sk_load_8888_dst_sse41+0x22> + DB 117,144 ; jne 16c0b <_sk_load_8888_dst_sse41+0x22> DB 102,65,15,110,100,128,8 ; movd 0x8(%r8,%rax,4),%xmm4 DB 102,15,112,252,69 ; pshufd $0x45,%xmm4,%xmm7 DB 243,65,15,126,36,128 ; movq (%r8,%rax,4),%xmm4 DB 102,15,58,14,252,15 ; pblendw $0xf,%xmm4,%xmm7 - DB 233,115,255,255,255 ; jmpq 16851 <_sk_load_8888_dst_sse41+0x22> + DB 233,115,255,255,255 ; jmpq 16c0b <_sk_load_8888_dst_sse41+0x22> DB 102,65,15,110,60,128 ; movd (%r8,%rax,4),%xmm7 - DB 233,104,255,255,255 ; jmpq 16851 <_sk_load_8888_dst_sse41+0x22> + DB 233,104,255,255,255 ; jmpq 16c0b <_sk_load_8888_dst_sse41+0x22> PUBLIC _sk_gather_8888_sse41 _sk_gather_8888_sse41 LABEL PROC @@ -21138,17 +21298,17 @@ _sk_gather_8888_sse41 LABEL PROC DB 102,65,15,58,34,28,128,1 ; pinsrd $0x1,(%r8,%rax,4),%xmm3 DB 102,67,15,58,34,28,152,2 ; pinsrd $0x2,(%r8,%r11,4),%xmm3 DB 102,67,15,58,34,28,144,3 ; pinsrd $0x3,(%r8,%r10,4),%xmm3 - DB 102,15,111,5,39,108,1,0 ; movdqa 0x16c27(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,5,93,103,1,0 ; movdqa 0x1675d(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,104,109,1,0 ; movaps 0x16d68(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,158,104,1,0 ; movaps 0x1689e(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,56,0,13,23,108,1,0 ; pshufb 0x16c17(%rip),%xmm1 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,13,77,103,1,0 ; pshufb 0x1674d(%rip),%xmm1 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,56,0,21,19,108,1,0 ; pshufb 0x16c13(%rip),%xmm2 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,21,73,103,1,0 ; pshufb 0x16749(%rip),%xmm2 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -21166,7 +21326,7 @@ _sk_store_8888_sse41 LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 68,15,40,5,237,107,1,0 ; movaps 0x16bed(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,35,103,1,0 ; movaps 0x16723(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -21185,23 +21345,23 @@ _sk_store_8888_sse41 LABEL PROC DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 102,69,15,235,194 ; por %xmm10,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 16a13 <_sk_store_8888_sse41+0x7f> + DB 117,10 ; jne 16dcd <_sk_store_8888_sse41+0x7f> DB 243,69,15,127,4,128 ; movdqu %xmm8,(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,29 ; je 16a3d <_sk_store_8888_sse41+0xa9> + DB 116,29 ; je 16df7 <_sk_store_8888_sse41+0xa9> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,15 ; je 16a35 <_sk_store_8888_sse41+0xa1> + DB 116,15 ; je 16def <_sk_store_8888_sse41+0xa1> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,227 ; jne 16a0f <_sk_store_8888_sse41+0x7b> + DB 117,227 ; jne 16dc9 <_sk_store_8888_sse41+0x7b> DB 102,69,15,58,22,68,128,8,2 ; pextrd $0x2,%xmm8,0x8(%r8,%rax,4) DB 102,69,15,214,4,128 ; movq %xmm8,(%r8,%rax,4) - DB 235,210 ; jmp 16a0f <_sk_store_8888_sse41+0x7b> + DB 235,210 ; jmp 16dc9 <_sk_store_8888_sse41+0x7b> DB 102,69,15,126,4,128 ; movd %xmm8,(%r8,%rax,4) - DB 235,202 ; jmp 16a0f <_sk_store_8888_sse41+0x7b> + DB 235,202 ; jmp 16dc9 <_sk_store_8888_sse41+0x7b> PUBLIC _sk_load_bgra_sse41 _sk_load_bgra_sse41 LABEL PROC @@ -21213,19 +21373,19 @@ _sk_load_bgra_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,89 ; jne 16aba <_sk_load_bgra_sse41+0x75> + DB 117,89 ; jne 16e74 <_sk_load_bgra_sse41+0x75> DB 243,65,15,111,28,128 ; movdqu (%r8,%rax,4),%xmm3 - DB 102,15,111,5,1,107,1,0 ; movdqa 0x16b01(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,5,55,102,1,0 ; movdqa 0x16637(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,208 ; cvtdq2ps %xmm0,%xmm2 - DB 68,15,40,5,66,108,1,0 ; movaps 0x16c42(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,120,103,1,0 ; movaps 0x16778(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,56,0,5,241,106,1,0 ; pshufb 0x16af1(%rip),%xmm0 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,5,39,102,1,0 ; pshufb 0x16627(%rip),%xmm0 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,56,0,5,237,106,1,0 ; pshufb 0x16aed(%rip),%xmm0 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,5,35,102,1,0 ; pshufb 0x16623(%rip),%xmm0 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -21236,19 +21396,19 @@ _sk_load_bgra_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,45 ; je 16af4 <_sk_load_bgra_sse41+0xaf> + DB 116,45 ; je 16eae <_sk_load_bgra_sse41+0xaf> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 16ae3 <_sk_load_bgra_sse41+0x9e> + DB 116,18 ; je 16e9d <_sk_load_bgra_sse41+0x9e> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,144 ; jne 16a67 <_sk_load_bgra_sse41+0x22> + DB 117,144 ; jne 16e21 <_sk_load_bgra_sse41+0x22> DB 102,65,15,110,68,128,8 ; movd 0x8(%r8,%rax,4),%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 243,65,15,126,4,128 ; movq (%r8,%rax,4),%xmm0 DB 102,15,58,14,216,15 ; pblendw $0xf,%xmm0,%xmm3 - DB 233,115,255,255,255 ; jmpq 16a67 <_sk_load_bgra_sse41+0x22> + DB 233,115,255,255,255 ; jmpq 16e21 <_sk_load_bgra_sse41+0x22> DB 102,65,15,110,28,128 ; movd (%r8,%rax,4),%xmm3 - DB 233,104,255,255,255 ; jmpq 16a67 <_sk_load_bgra_sse41+0x22> + DB 233,104,255,255,255 ; jmpq 16e21 <_sk_load_bgra_sse41+0x22> PUBLIC _sk_load_bgra_dst_sse41 _sk_load_bgra_dst_sse41 LABEL PROC @@ -21260,19 +21420,19 @@ _sk_load_bgra_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,89 ; jne 16b74 <_sk_load_bgra_dst_sse41+0x75> + DB 117,89 ; jne 16f2e <_sk_load_bgra_dst_sse41+0x75> DB 243,65,15,111,60,128 ; movdqu (%r8,%rax,4),%xmm7 - DB 102,15,111,37,71,106,1,0 ; movdqa 0x16a47(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,37,125,101,1,0 ; movdqa 0x1657d(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,244 ; cvtdq2ps %xmm4,%xmm6 - DB 68,15,40,5,136,107,1,0 ; movaps 0x16b88(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,190,102,1,0 ; movaps 0x166be(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,240 ; mulps %xmm8,%xmm6 DB 102,15,111,231 ; movdqa %xmm7,%xmm4 - DB 102,15,56,0,37,55,106,1,0 ; pshufb 0x16a37(%rip),%xmm4 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,37,109,101,1,0 ; pshufb 0x1656d(%rip),%xmm4 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,236 ; cvtdq2ps %xmm4,%xmm5 DB 65,15,89,232 ; mulps %xmm8,%xmm5 DB 102,15,111,231 ; movdqa %xmm7,%xmm4 - DB 102,15,56,0,37,51,106,1,0 ; pshufb 0x16a33(%rip),%xmm4 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,37,105,101,1,0 ; pshufb 0x16569(%rip),%xmm4 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 65,15,89,224 ; mulps %xmm8,%xmm4 DB 102,15,114,215,24 ; psrld $0x18,%xmm7 @@ -21283,19 +21443,19 @@ _sk_load_bgra_dst_sse41 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,45 ; je 16bae <_sk_load_bgra_dst_sse41+0xaf> + DB 116,45 ; je 16f68 <_sk_load_bgra_dst_sse41+0xaf> DB 102,15,239,255 ; pxor %xmm7,%xmm7 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 16b9d <_sk_load_bgra_dst_sse41+0x9e> + DB 116,18 ; je 16f57 <_sk_load_bgra_dst_sse41+0x9e> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,144 ; jne 16b21 <_sk_load_bgra_dst_sse41+0x22> + DB 117,144 ; jne 16edb <_sk_load_bgra_dst_sse41+0x22> DB 102,65,15,110,100,128,8 ; movd 0x8(%r8,%rax,4),%xmm4 DB 102,15,112,252,69 ; pshufd $0x45,%xmm4,%xmm7 DB 243,65,15,126,36,128 ; movq (%r8,%rax,4),%xmm4 DB 102,15,58,14,252,15 ; pblendw $0xf,%xmm4,%xmm7 - DB 233,115,255,255,255 ; jmpq 16b21 <_sk_load_bgra_dst_sse41+0x22> + DB 233,115,255,255,255 ; jmpq 16edb <_sk_load_bgra_dst_sse41+0x22> DB 102,65,15,110,60,128 ; movd (%r8,%rax,4),%xmm7 - DB 233,104,255,255,255 ; jmpq 16b21 <_sk_load_bgra_dst_sse41+0x22> + DB 233,104,255,255,255 ; jmpq 16edb <_sk_load_bgra_dst_sse41+0x22> PUBLIC _sk_gather_bgra_sse41 _sk_gather_bgra_sse41 LABEL PROC @@ -21317,17 +21477,17 @@ _sk_gather_bgra_sse41 LABEL PROC DB 102,65,15,58,34,28,128,1 ; pinsrd $0x1,(%r8,%rax,4),%xmm3 DB 102,67,15,58,34,28,152,2 ; pinsrd $0x2,(%r8,%r11,4),%xmm3 DB 102,67,15,58,34,28,144,3 ; pinsrd $0x3,(%r8,%r10,4),%xmm3 - DB 102,15,111,5,87,105,1,0 ; movdqa 0x16957(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,5,141,100,1,0 ; movdqa 0x1648d(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,208 ; cvtdq2ps %xmm0,%xmm2 - DB 68,15,40,5,152,106,1,0 ; movaps 0x16a98(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,206,101,1,0 ; movaps 0x165ce(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,56,0,5,71,105,1,0 ; pshufb 0x16947(%rip),%xmm0 # 2d580 <_sk_overlay_sse2_8bit+0x1028> + DB 102,15,56,0,5,125,100,1,0 ; pshufb 0x1647d(%rip),%xmm0 # 2d470 <_sk_overlay_sse2_8bit+0xfc1> DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,56,0,5,67,105,1,0 ; pshufb 0x16943(%rip),%xmm0 # 2d590 <_sk_overlay_sse2_8bit+0x1038> + DB 102,15,56,0,5,121,100,1,0 ; pshufb 0x16479(%rip),%xmm0 # 2d480 <_sk_overlay_sse2_8bit+0xfd1> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 @@ -21345,7 +21505,7 @@ _sk_store_bgra_sse41 LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 68,15,40,5,29,105,1,0 ; movaps 0x1691d(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,83,100,1,0 ; movaps 0x16453(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 68,15,40,202 ; movaps %xmm2,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -21364,23 +21524,23 @@ _sk_store_bgra_sse41 LABEL PROC DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 102,69,15,235,194 ; por %xmm10,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 16ce3 <_sk_store_bgra_sse41+0x7f> + DB 117,10 ; jne 1709d <_sk_store_bgra_sse41+0x7f> DB 243,69,15,127,4,128 ; movdqu %xmm8,(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,29 ; je 16d0d <_sk_store_bgra_sse41+0xa9> + DB 116,29 ; je 170c7 <_sk_store_bgra_sse41+0xa9> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,15 ; je 16d05 <_sk_store_bgra_sse41+0xa1> + DB 116,15 ; je 170bf <_sk_store_bgra_sse41+0xa1> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,227 ; jne 16cdf <_sk_store_bgra_sse41+0x7b> + DB 117,227 ; jne 17099 <_sk_store_bgra_sse41+0x7b> DB 102,69,15,58,22,68,128,8,2 ; pextrd $0x2,%xmm8,0x8(%r8,%rax,4) DB 102,69,15,214,4,128 ; movq %xmm8,(%r8,%rax,4) - DB 235,210 ; jmp 16cdf <_sk_store_bgra_sse41+0x7b> + DB 235,210 ; jmp 17099 <_sk_store_bgra_sse41+0x7b> DB 102,69,15,126,4,128 ; movd %xmm8,(%r8,%rax,4) - DB 235,202 ; jmp 16cdf <_sk_store_bgra_sse41+0x7b> + DB 235,202 ; jmp 17099 <_sk_store_bgra_sse41+0x7b> PUBLIC _sk_load_f16_sse41 _sk_load_f16_sse41 LABEL PROC @@ -21392,80 +21552,81 @@ _sk_load_f16_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,62,1,0,0 ; jne 16e73 <_sk_load_f16_sse41+0x15e> + DB 15,133,68,1,0,0 ; jne 17233 <_sk_load_f16_sse41+0x164> DB 102,65,15,16,4,192 ; movupd (%r8,%rax,8),%xmm0 DB 243,65,15,111,76,192,16 ; movdqu 0x10(%r8,%rax,8),%xmm1 DB 102,68,15,40,200 ; movapd %xmm0,%xmm9 DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 - DB 102,68,15,97,216 ; punpcklwd %xmm0,%xmm11 + DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 + DB 102,68,15,97,224 ; punpcklwd %xmm0,%xmm12 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1 - DB 102,68,15,111,5,34,108,1,0 ; movdqa 0x16c22(%rip),%xmm8 # 2d990 <_sk_overlay_sse2_8bit+0x1438> - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,65,15,219,208 ; pand %xmm8,%xmm2 - DB 102,15,239,202 ; pxor %xmm2,%xmm1 - DB 102,15,111,29,29,108,1,0 ; movdqa 0x16c1d(%rip),%xmm3 # 2d9a0 <_sk_overlay_sse2_8bit+0x1448> - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,65,15,56,51,212 ; pmovzxwd %xmm12,%xmm2 + DB 102,68,15,111,5,72,103,1,0 ; movdqa 0x16748(%rip),%xmm8 # 2d870 <_sk_overlay_sse2_8bit+0x13c1> + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,65,15,219,200 ; pand %xmm8,%xmm1 + DB 102,68,15,111,21,70,103,1,0 ; movdqa 0x16746(%rip),%xmm10 # 2d880 <_sk_overlay_sse2_8bit+0x13d1> + DB 102,65,15,219,210 ; pand %xmm10,%xmm2 + DB 102,15,111,29,73,103,1,0 ; movdqa 0x16749(%rip),%xmm3 # 2d890 <_sk_overlay_sse2_8bit+0x13e1> + DB 102,15,114,241,16 ; pslld $0x10,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 102,15,56,63,195 ; pmaxud %xmm3,%xmm0 - DB 102,15,118,193 ; pcmpeqd %xmm1,%xmm0 - DB 102,15,114,241,13 ; pslld $0xd,%xmm1 - DB 102,15,235,202 ; por %xmm2,%xmm1 - DB 102,68,15,111,21,9,108,1,0 ; movdqa 0x16c09(%rip),%xmm10 # 2d9b0 <_sk_overlay_sse2_8bit+0x1458> - DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 - DB 102,15,219,193 ; pand %xmm1,%xmm0 - DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11 - DB 102,69,15,56,51,219 ; pmovzxwd %xmm11,%xmm11 - DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 - DB 102,65,15,219,208 ; pand %xmm8,%xmm2 - DB 102,68,15,239,218 ; pxor %xmm2,%xmm11 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,65,15,111,203 ; movdqa %xmm11,%xmm1 - DB 102,15,56,63,203 ; pmaxud %xmm3,%xmm1 - DB 102,65,15,118,203 ; pcmpeqd %xmm11,%xmm1 - DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,68,15,235,218 ; por %xmm2,%xmm11 - DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 - DB 102,65,15,219,203 ; pand %xmm11,%xmm1 - DB 102,69,15,56,51,217 ; pmovzxwd %xmm9,%xmm11 - DB 102,69,15,111,227 ; movdqa %xmm11,%xmm12 + DB 102,15,118,194 ; pcmpeqd %xmm2,%xmm0 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,15,235,209 ; por %xmm1,%xmm2 + DB 102,68,15,111,29,53,103,1,0 ; movdqa 0x16735(%rip),%xmm11 # 2d8a0 <_sk_overlay_sse2_8bit+0x13f1> + DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 + DB 102,15,219,194 ; pand %xmm2,%xmm0 + DB 102,65,15,112,204,238 ; pshufd $0xee,%xmm12,%xmm1 + DB 102,15,56,51,209 ; pmovzxwd %xmm1,%xmm2 + DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 DB 102,69,15,219,224 ; pand %xmm8,%xmm12 - DB 102,69,15,239,220 ; pxor %xmm12,%xmm11 + DB 102,65,15,219,210 ; pand %xmm10,%xmm2 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,56,63,203 ; pmaxud %xmm3,%xmm1 + DB 102,15,118,202 ; pcmpeqd %xmm2,%xmm1 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,65,15,235,212 ; por %xmm12,%xmm2 + DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 + DB 102,15,219,202 ; pand %xmm2,%xmm1 + DB 102,69,15,56,51,225 ; pmovzxwd %xmm9,%xmm12 + DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 + DB 102,69,15,219,232 ; pand %xmm8,%xmm13 + DB 102,69,15,219,226 ; pand %xmm10,%xmm12 + DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 + DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 DB 102,15,56,63,211 ; pmaxud %xmm3,%xmm2 - DB 102,65,15,118,211 ; pcmpeqd %xmm11,%xmm2 - DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 - DB 102,65,15,219,211 ; pand %xmm11,%xmm2 - DB 102,65,15,115,217,8 ; psrldq $0x8,%xmm9 + DB 102,65,15,118,212 ; pcmpeqd %xmm12,%xmm2 + DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 102,69,15,254,227 ; paddd %xmm11,%xmm12 + DB 102,65,15,219,212 ; pand %xmm12,%xmm2 + DB 102,69,15,112,201,238 ; pshufd $0xee,%xmm9,%xmm9 DB 102,69,15,56,51,201 ; pmovzxwd %xmm9,%xmm9 DB 102,69,15,219,193 ; pand %xmm9,%xmm8 - DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 + DB 102,69,15,219,202 ; pand %xmm10,%xmm9 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,56,63,217 ; pmaxud %xmm9,%xmm3 DB 102,65,15,118,217 ; pcmpeqd %xmm9,%xmm3 DB 102,65,15,114,241,13 ; pslld $0xd,%xmm9 DB 102,69,15,235,200 ; por %xmm8,%xmm9 - DB 102,69,15,254,202 ; paddd %xmm10,%xmm9 + DB 102,69,15,254,203 ; paddd %xmm11,%xmm9 DB 102,65,15,219,217 ; pand %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 242,65,15,16,4,192 ; movsd (%r8,%rax,8),%xmm0 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,13 ; jne 16e8c <_sk_load_f16_sse41+0x177> + DB 117,13 ; jne 1724c <_sk_load_f16_sse41+0x17d> DB 243,15,126,192 ; movq %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,182,254,255,255 ; jmpq 16d42 <_sk_load_f16_sse41+0x2d> + DB 233,176,254,255,255 ; jmpq 170fc <_sk_load_f16_sse41+0x2d> DB 102,65,15,22,68,192,8 ; movhpd 0x8(%r8,%rax,8),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,161,254,255,255 ; jb 16d42 <_sk_load_f16_sse41+0x2d> + DB 15,130,155,254,255,255 ; jb 170fc <_sk_load_f16_sse41+0x2d> DB 243,65,15,126,76,192,16 ; movq 0x10(%r8,%rax,8),%xmm1 - DB 233,149,254,255,255 ; jmpq 16d42 <_sk_load_f16_sse41+0x2d> + DB 233,143,254,255,255 ; jmpq 170fc <_sk_load_f16_sse41+0x2d> PUBLIC _sk_load_f16_dst_sse41 _sk_load_f16_dst_sse41 LABEL PROC @@ -21477,80 +21638,81 @@ _sk_load_f16_dst_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,62,1,0,0 ; jne 1700b <_sk_load_f16_dst_sse41+0x15e> + DB 15,133,68,1,0,0 ; jne 173d1 <_sk_load_f16_dst_sse41+0x164> DB 102,65,15,16,36,192 ; movupd (%r8,%rax,8),%xmm4 DB 243,65,15,111,108,192,16 ; movdqu 0x10(%r8,%rax,8),%xmm5 DB 102,68,15,40,204 ; movapd %xmm4,%xmm9 DB 102,68,15,97,205 ; punpcklwd %xmm5,%xmm9 DB 102,15,105,229 ; punpckhwd %xmm5,%xmm4 - DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 - DB 102,68,15,97,220 ; punpcklwd %xmm4,%xmm11 + DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 + DB 102,68,15,97,228 ; punpcklwd %xmm4,%xmm12 DB 102,68,15,105,204 ; punpckhwd %xmm4,%xmm9 - DB 102,65,15,56,51,235 ; pmovzxwd %xmm11,%xmm5 - DB 102,68,15,111,5,138,106,1,0 ; movdqa 0x16a8a(%rip),%xmm8 # 2d990 <_sk_overlay_sse2_8bit+0x1438> - DB 102,15,111,245 ; movdqa %xmm5,%xmm6 - DB 102,65,15,219,240 ; pand %xmm8,%xmm6 - DB 102,15,239,238 ; pxor %xmm6,%xmm5 - DB 102,15,111,61,133,106,1,0 ; movdqa 0x16a85(%rip),%xmm7 # 2d9a0 <_sk_overlay_sse2_8bit+0x1448> - DB 102,15,114,246,16 ; pslld $0x10,%xmm6 - DB 102,15,111,229 ; movdqa %xmm5,%xmm4 + DB 102,65,15,56,51,244 ; pmovzxwd %xmm12,%xmm6 + DB 102,68,15,111,5,170,101,1,0 ; movdqa 0x165aa(%rip),%xmm8 # 2d870 <_sk_overlay_sse2_8bit+0x13c1> + DB 102,15,111,238 ; movdqa %xmm6,%xmm5 + DB 102,65,15,219,232 ; pand %xmm8,%xmm5 + DB 102,68,15,111,21,168,101,1,0 ; movdqa 0x165a8(%rip),%xmm10 # 2d880 <_sk_overlay_sse2_8bit+0x13d1> + DB 102,65,15,219,242 ; pand %xmm10,%xmm6 + DB 102,15,111,61,171,101,1,0 ; movdqa 0x165ab(%rip),%xmm7 # 2d890 <_sk_overlay_sse2_8bit+0x13e1> + DB 102,15,114,245,16 ; pslld $0x10,%xmm5 + DB 102,15,111,230 ; movdqa %xmm6,%xmm4 DB 102,15,56,63,231 ; pmaxud %xmm7,%xmm4 - DB 102,15,118,229 ; pcmpeqd %xmm5,%xmm4 - DB 102,15,114,245,13 ; pslld $0xd,%xmm5 - DB 102,15,235,238 ; por %xmm6,%xmm5 - DB 102,68,15,111,21,113,106,1,0 ; movdqa 0x16a71(%rip),%xmm10 # 2d9b0 <_sk_overlay_sse2_8bit+0x1458> - DB 102,65,15,254,234 ; paddd %xmm10,%xmm5 - DB 102,15,219,229 ; pand %xmm5,%xmm4 - DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11 - DB 102,69,15,56,51,219 ; pmovzxwd %xmm11,%xmm11 - DB 102,65,15,111,243 ; movdqa %xmm11,%xmm6 - DB 102,65,15,219,240 ; pand %xmm8,%xmm6 - DB 102,68,15,239,222 ; pxor %xmm6,%xmm11 - DB 102,15,114,246,16 ; pslld $0x10,%xmm6 - DB 102,65,15,111,235 ; movdqa %xmm11,%xmm5 - DB 102,15,56,63,239 ; pmaxud %xmm7,%xmm5 - DB 102,65,15,118,235 ; pcmpeqd %xmm11,%xmm5 - DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,68,15,235,222 ; por %xmm6,%xmm11 - DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 - DB 102,65,15,219,235 ; pand %xmm11,%xmm5 - DB 102,69,15,56,51,217 ; pmovzxwd %xmm9,%xmm11 - DB 102,69,15,111,227 ; movdqa %xmm11,%xmm12 + DB 102,15,118,230 ; pcmpeqd %xmm6,%xmm4 + DB 102,15,114,246,13 ; pslld $0xd,%xmm6 + DB 102,15,235,245 ; por %xmm5,%xmm6 + DB 102,68,15,111,29,151,101,1,0 ; movdqa 0x16597(%rip),%xmm11 # 2d8a0 <_sk_overlay_sse2_8bit+0x13f1> + DB 102,65,15,254,243 ; paddd %xmm11,%xmm6 + DB 102,15,219,230 ; pand %xmm6,%xmm4 + DB 102,65,15,112,236,238 ; pshufd $0xee,%xmm12,%xmm5 + DB 102,15,56,51,245 ; pmovzxwd %xmm5,%xmm6 + DB 102,68,15,111,230 ; movdqa %xmm6,%xmm12 DB 102,69,15,219,224 ; pand %xmm8,%xmm12 - DB 102,69,15,239,220 ; pxor %xmm12,%xmm11 + DB 102,65,15,219,242 ; pand %xmm10,%xmm6 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,65,15,111,243 ; movdqa %xmm11,%xmm6 + DB 102,15,111,238 ; movdqa %xmm6,%xmm5 + DB 102,15,56,63,239 ; pmaxud %xmm7,%xmm5 + DB 102,15,118,238 ; pcmpeqd %xmm6,%xmm5 + DB 102,15,114,246,13 ; pslld $0xd,%xmm6 + DB 102,65,15,235,244 ; por %xmm12,%xmm6 + DB 102,65,15,254,243 ; paddd %xmm11,%xmm6 + DB 102,15,219,238 ; pand %xmm6,%xmm5 + DB 102,69,15,56,51,225 ; pmovzxwd %xmm9,%xmm12 + DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 + DB 102,69,15,219,232 ; pand %xmm8,%xmm13 + DB 102,69,15,219,226 ; pand %xmm10,%xmm12 + DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 + DB 102,65,15,111,244 ; movdqa %xmm12,%xmm6 DB 102,15,56,63,247 ; pmaxud %xmm7,%xmm6 - DB 102,65,15,118,243 ; pcmpeqd %xmm11,%xmm6 - DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 - DB 102,65,15,219,243 ; pand %xmm11,%xmm6 - DB 102,65,15,115,217,8 ; psrldq $0x8,%xmm9 + DB 102,65,15,118,244 ; pcmpeqd %xmm12,%xmm6 + DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 102,69,15,254,227 ; paddd %xmm11,%xmm12 + DB 102,65,15,219,244 ; pand %xmm12,%xmm6 + DB 102,69,15,112,201,238 ; pshufd $0xee,%xmm9,%xmm9 DB 102,69,15,56,51,201 ; pmovzxwd %xmm9,%xmm9 DB 102,69,15,219,193 ; pand %xmm9,%xmm8 - DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 + DB 102,69,15,219,202 ; pand %xmm10,%xmm9 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,56,63,249 ; pmaxud %xmm9,%xmm7 DB 102,65,15,118,249 ; pcmpeqd %xmm9,%xmm7 DB 102,65,15,114,241,13 ; pslld $0xd,%xmm9 DB 102,69,15,235,200 ; por %xmm8,%xmm9 - DB 102,69,15,254,202 ; paddd %xmm10,%xmm9 + DB 102,69,15,254,203 ; paddd %xmm11,%xmm9 DB 102,65,15,219,249 ; pand %xmm9,%xmm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 242,65,15,16,36,192 ; movsd (%r8,%rax,8),%xmm4 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,13 ; jne 17024 <_sk_load_f16_dst_sse41+0x177> + DB 117,13 ; jne 173ea <_sk_load_f16_dst_sse41+0x17d> DB 243,15,126,228 ; movq %xmm4,%xmm4 DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 233,182,254,255,255 ; jmpq 16eda <_sk_load_f16_dst_sse41+0x2d> + DB 233,176,254,255,255 ; jmpq 1729a <_sk_load_f16_dst_sse41+0x2d> DB 102,65,15,22,100,192,8 ; movhpd 0x8(%r8,%rax,8),%xmm4 DB 102,15,239,237 ; pxor %xmm5,%xmm5 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,161,254,255,255 ; jb 16eda <_sk_load_f16_dst_sse41+0x2d> + DB 15,130,155,254,255,255 ; jb 1729a <_sk_load_f16_dst_sse41+0x2d> DB 243,65,15,126,108,192,16 ; movq 0x10(%r8,%rax,8),%xmm5 - DB 233,149,254,255,255 ; jmpq 16eda <_sk_load_f16_dst_sse41+0x2d> + DB 233,143,254,255,255 ; jmpq 1729a <_sk_load_f16_dst_sse41+0x2d> PUBLIC _sk_gather_f16_sse41 _sk_gather_f16_sse41 LABEL PROC @@ -21577,59 +21739,60 @@ _sk_gather_f16_sse41 LABEL PROC DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9 DB 102,15,105,209 ; punpckhwd %xmm1,%xmm2 - DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 - DB 102,68,15,97,218 ; punpcklwd %xmm2,%xmm11 + DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 + DB 102,68,15,97,226 ; punpcklwd %xmm2,%xmm12 DB 102,68,15,105,202 ; punpckhwd %xmm2,%xmm9 - DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1 - DB 102,68,15,111,5,197,104,1,0 ; movdqa 0x168c5(%rip),%xmm8 # 2d990 <_sk_overlay_sse2_8bit+0x1438> - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,65,15,219,208 ; pand %xmm8,%xmm2 - DB 102,15,239,202 ; pxor %xmm2,%xmm1 - DB 102,15,111,29,192,104,1,0 ; movdqa 0x168c0(%rip),%xmm3 # 2d9a0 <_sk_overlay_sse2_8bit+0x1448> - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,65,15,56,51,212 ; pmovzxwd %xmm12,%xmm2 + DB 102,68,15,111,5,223,99,1,0 ; movdqa 0x163df(%rip),%xmm8 # 2d870 <_sk_overlay_sse2_8bit+0x13c1> + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,65,15,219,200 ; pand %xmm8,%xmm1 + DB 102,68,15,111,21,221,99,1,0 ; movdqa 0x163dd(%rip),%xmm10 # 2d880 <_sk_overlay_sse2_8bit+0x13d1> + DB 102,65,15,219,210 ; pand %xmm10,%xmm2 + DB 102,15,111,29,224,99,1,0 ; movdqa 0x163e0(%rip),%xmm3 # 2d890 <_sk_overlay_sse2_8bit+0x13e1> + DB 102,15,114,241,16 ; pslld $0x10,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 102,15,56,63,195 ; pmaxud %xmm3,%xmm0 - DB 102,15,118,193 ; pcmpeqd %xmm1,%xmm0 - DB 102,15,114,241,13 ; pslld $0xd,%xmm1 - DB 102,15,235,202 ; por %xmm2,%xmm1 - DB 102,68,15,111,21,172,104,1,0 ; movdqa 0x168ac(%rip),%xmm10 # 2d9b0 <_sk_overlay_sse2_8bit+0x1458> - DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 - DB 102,15,219,193 ; pand %xmm1,%xmm0 - DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11 - DB 102,69,15,56,51,219 ; pmovzxwd %xmm11,%xmm11 - DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 - DB 102,65,15,219,208 ; pand %xmm8,%xmm2 - DB 102,68,15,239,218 ; pxor %xmm2,%xmm11 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,65,15,111,203 ; movdqa %xmm11,%xmm1 - DB 102,15,56,63,203 ; pmaxud %xmm3,%xmm1 - DB 102,65,15,118,203 ; pcmpeqd %xmm11,%xmm1 - DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,68,15,235,218 ; por %xmm2,%xmm11 - DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 - DB 102,65,15,219,203 ; pand %xmm11,%xmm1 - DB 102,69,15,56,51,217 ; pmovzxwd %xmm9,%xmm11 - DB 102,69,15,111,227 ; movdqa %xmm11,%xmm12 + DB 102,15,118,194 ; pcmpeqd %xmm2,%xmm0 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,15,235,209 ; por %xmm1,%xmm2 + DB 102,68,15,111,29,204,99,1,0 ; movdqa 0x163cc(%rip),%xmm11 # 2d8a0 <_sk_overlay_sse2_8bit+0x13f1> + DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 + DB 102,15,219,194 ; pand %xmm2,%xmm0 + DB 102,65,15,112,204,238 ; pshufd $0xee,%xmm12,%xmm1 + DB 102,15,56,51,209 ; pmovzxwd %xmm1,%xmm2 + DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 DB 102,69,15,219,224 ; pand %xmm8,%xmm12 - DB 102,69,15,239,220 ; pxor %xmm12,%xmm11 + DB 102,65,15,219,210 ; pand %xmm10,%xmm2 DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,56,63,203 ; pmaxud %xmm3,%xmm1 + DB 102,15,118,202 ; pcmpeqd %xmm2,%xmm1 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,65,15,235,212 ; por %xmm12,%xmm2 + DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 + DB 102,15,219,202 ; pand %xmm2,%xmm1 + DB 102,69,15,56,51,225 ; pmovzxwd %xmm9,%xmm12 + DB 102,69,15,111,236 ; movdqa %xmm12,%xmm13 + DB 102,69,15,219,232 ; pand %xmm8,%xmm13 + DB 102,69,15,219,226 ; pand %xmm10,%xmm12 + DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 + DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 DB 102,15,56,63,211 ; pmaxud %xmm3,%xmm2 - DB 102,65,15,118,211 ; pcmpeqd %xmm11,%xmm2 - DB 102,65,15,114,243,13 ; pslld $0xd,%xmm11 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 102,69,15,254,218 ; paddd %xmm10,%xmm11 - DB 102,65,15,219,211 ; pand %xmm11,%xmm2 - DB 102,65,15,115,217,8 ; psrldq $0x8,%xmm9 + DB 102,65,15,118,212 ; pcmpeqd %xmm12,%xmm2 + DB 102,65,15,114,244,13 ; pslld $0xd,%xmm12 + DB 102,69,15,235,229 ; por %xmm13,%xmm12 + DB 102,69,15,254,227 ; paddd %xmm11,%xmm12 + DB 102,65,15,219,212 ; pand %xmm12,%xmm2 + DB 102,69,15,112,201,238 ; pshufd $0xee,%xmm9,%xmm9 DB 102,69,15,56,51,201 ; pmovzxwd %xmm9,%xmm9 DB 102,69,15,219,193 ; pand %xmm9,%xmm8 - DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 + DB 102,69,15,219,202 ; pand %xmm10,%xmm9 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,56,63,217 ; pmaxud %xmm9,%xmm3 DB 102,65,15,118,217 ; pcmpeqd %xmm9,%xmm3 DB 102,65,15,114,241,13 ; pslld $0xd,%xmm9 DB 102,69,15,235,200 ; por %xmm8,%xmm9 - DB 102,69,15,254,202 ; paddd %xmm10,%xmm9 + DB 102,69,15,254,203 ; paddd %xmm11,%xmm9 DB 102,65,15,219,217 ; pand %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -21643,75 +21806,79 @@ _sk_store_f16_sse41 LABEL PROC DB 73,193,224,3 ; shl $0x3,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 102,68,15,111,21,208,103,1,0 ; movdqa 0x167d0(%rip),%xmm10 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 102,68,15,111,216 ; movdqa %xmm0,%xmm11 - DB 102,69,15,219,218 ; pand %xmm10,%xmm11 + DB 102,68,15,111,13,244,98,1,0 ; movdqa 0x162f4(%rip),%xmm9 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> + DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 + DB 102,69,15,219,225 ; pand %xmm9,%xmm12 + DB 102,68,15,111,29,241,98,1,0 ; movdqa 0x162f1(%rip),%xmm11 # 2d8c0 <_sk_overlay_sse2_8bit+0x1411> DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 - DB 102,69,15,239,235 ; pxor %xmm11,%xmm13 - DB 102,68,15,111,13,195,103,1,0 ; movdqa 0x167c3(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1478> - DB 102,65,15,114,211,16 ; psrld $0x10,%xmm11 - DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 - DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 102,68,15,111,21,238,98,1,0 ; movdqa 0x162ee(%rip),%xmm10 # 2d8d0 <_sk_overlay_sse2_8bit+0x1421> + DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 + DB 102,69,15,111,197 ; movdqa %xmm13,%xmm8 + DB 102,69,15,56,63,194 ; pmaxud %xmm10,%xmm8 + DB 102,69,15,118,197 ; pcmpeqd %xmm13,%xmm8 DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 102,68,15,111,37,180,103,1,0 ; movdqa 0x167b4(%rip),%xmm12 # 2d9e0 <_sk_overlay_sse2_8bit+0x1488> - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 102,69,15,254,221 ; paddd %xmm13,%xmm11 - DB 102,69,15,223,195 ; pandn %xmm11,%xmm8 - DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 - DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 - DB 102,69,15,219,218 ; pand %xmm10,%xmm11 + DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 + DB 102,68,15,111,37,212,98,1,0 ; movdqa 0x162d4(%rip),%xmm12 # 2d8e0 <_sk_overlay_sse2_8bit+0x1431> + DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 + DB 102,69,15,219,197 ; pand %xmm13,%xmm8 + DB 102,68,15,56,43,192 ; packusdw %xmm0,%xmm8 + DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 + DB 102,69,15,219,233 ; pand %xmm9,%xmm13 DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,69,15,239,243 ; pxor %xmm11,%xmm14 - DB 102,65,15,114,211,16 ; psrld $0x10,%xmm11 - DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 - DB 102,69,15,102,238 ; pcmpgtd %xmm14,%xmm13 + DB 102,69,15,219,243 ; pand %xmm11,%xmm14 + DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13 + DB 102,69,15,111,254 ; movdqa %xmm14,%xmm15 + DB 102,69,15,56,63,250 ; pmaxud %xmm10,%xmm15 + DB 102,69,15,118,254 ; pcmpeqd %xmm14,%xmm15 DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,235,220 ; por %xmm12,%xmm11 - DB 102,69,15,254,222 ; paddd %xmm14,%xmm11 - DB 102,69,15,223,235 ; pandn %xmm11,%xmm13 - DB 102,69,15,56,43,237 ; packusdw %xmm13,%xmm13 + DB 102,69,15,254,245 ; paddd %xmm13,%xmm14 + DB 102,69,15,254,244 ; paddd %xmm12,%xmm14 + DB 102,69,15,219,254 ; pand %xmm14,%xmm15 + DB 102,68,15,56,43,248 ; packusdw %xmm0,%xmm15 + DB 102,69,15,97,199 ; punpcklwd %xmm15,%xmm8 DB 102,68,15,111,242 ; movdqa %xmm2,%xmm14 - DB 102,69,15,219,242 ; pand %xmm10,%xmm14 + DB 102,69,15,219,241 ; pand %xmm9,%xmm14 DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 - DB 102,69,15,239,254 ; pxor %xmm14,%xmm15 + DB 102,69,15,219,251 ; pand %xmm11,%xmm15 DB 102,65,15,114,214,16 ; psrld $0x10,%xmm14 - DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 - DB 102,69,15,102,223 ; pcmpgtd %xmm15,%xmm11 + DB 102,69,15,111,239 ; movdqa %xmm15,%xmm13 + DB 102,69,15,56,63,234 ; pmaxud %xmm10,%xmm13 + DB 102,69,15,118,239 ; pcmpeqd %xmm15,%xmm13 DB 102,65,15,114,215,13 ; psrld $0xd,%xmm15 - DB 102,69,15,235,244 ; por %xmm12,%xmm14 - DB 102,69,15,254,247 ; paddd %xmm15,%xmm14 - DB 102,69,15,223,222 ; pandn %xmm14,%xmm11 - DB 102,69,15,56,43,219 ; packusdw %xmm11,%xmm11 - DB 102,68,15,219,211 ; pand %xmm3,%xmm10 - DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,69,15,239,242 ; pxor %xmm10,%xmm14 - DB 102,65,15,114,210,16 ; psrld $0x10,%xmm10 - DB 102,69,15,102,206 ; pcmpgtd %xmm14,%xmm9 - DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,235,212 ; por %xmm12,%xmm10 - DB 102,69,15,254,214 ; paddd %xmm14,%xmm10 - DB 102,69,15,223,202 ; pandn %xmm10,%xmm9 - DB 102,69,15,56,43,201 ; packusdw %xmm9,%xmm9 - DB 102,69,15,97,197 ; punpcklwd %xmm13,%xmm8 - DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11 + DB 102,69,15,254,254 ; paddd %xmm14,%xmm15 + DB 102,69,15,254,252 ; paddd %xmm12,%xmm15 + DB 102,69,15,219,239 ; pand %xmm15,%xmm13 + DB 102,68,15,56,43,232 ; packusdw %xmm0,%xmm13 + DB 102,68,15,219,203 ; pand %xmm3,%xmm9 + DB 102,68,15,219,219 ; pand %xmm3,%xmm11 + DB 102,65,15,114,209,16 ; psrld $0x10,%xmm9 + DB 102,69,15,56,63,211 ; pmaxud %xmm11,%xmm10 + DB 102,69,15,118,211 ; pcmpeqd %xmm11,%xmm10 + DB 102,65,15,114,211,13 ; psrld $0xd,%xmm11 + DB 102,69,15,254,217 ; paddd %xmm9,%xmm11 + DB 102,69,15,254,220 ; paddd %xmm12,%xmm11 + DB 102,69,15,219,211 ; pand %xmm11,%xmm10 + DB 102,68,15,56,43,208 ; packusdw %xmm0,%xmm10 + DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 - DB 102,69,15,98,203 ; punpckldq %xmm11,%xmm9 + DB 102,69,15,98,205 ; punpckldq %xmm13,%xmm9 DB 72,133,255 ; test %rdi,%rdi - DB 117,21 ; jne 17322 <_sk_store_f16_sse41+0x152> + DB 117,21 ; jne 1770a <_sk_store_f16_sse41+0x16e> DB 69,15,17,12,192 ; movups %xmm9,(%r8,%rax,8) - DB 102,69,15,106,195 ; punpckhdq %xmm11,%xmm8 + DB 102,69,15,106,197 ; punpckhdq %xmm13,%xmm8 DB 243,69,15,127,68,192,16 ; movdqu %xmm8,0x10(%r8,%rax,8) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,69,15,214,12,192 ; movq %xmm9,(%r8,%rax,8) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 1731e <_sk_store_f16_sse41+0x14e> + DB 116,240 ; je 17706 <_sk_store_f16_sse41+0x16a> DB 102,69,15,23,76,192,8 ; movhpd %xmm9,0x8(%r8,%rax,8) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 1731e <_sk_store_f16_sse41+0x14e> - DB 102,69,15,106,195 ; punpckhdq %xmm11,%xmm8 + DB 114,227 ; jb 17706 <_sk_store_f16_sse41+0x16a> + DB 102,69,15,106,197 ; punpckhdq %xmm13,%xmm8 DB 102,69,15,214,68,192,16 ; movq %xmm8,0x10(%r8,%rax,8) - DB 235,213 ; jmp 1731e <_sk_store_f16_sse41+0x14e> + DB 235,213 ; jmp 17706 <_sk_store_f16_sse41+0x16a> PUBLIC _sk_load_u16_be_sse41 _sk_load_u16_be_sse41 LABEL PROC @@ -21724,7 +21891,7 @@ _sk_load_u16_be_sse41 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,185,0,0,0 ; jne 17429 <_sk_load_u16_be_sse41+0xe0> + DB 15,133,185,0,0,0 ; jne 17811 <_sk_load_u16_be_sse41+0xe0> DB 102,65,15,16,4,64 ; movupd (%r8,%rax,2),%xmm0 DB 243,65,15,111,76,64,16 ; movdqu 0x10(%r8,%rax,2),%xmm1 DB 102,15,40,208 ; movapd %xmm0,%xmm2 @@ -21740,7 +21907,7 @@ _sk_load_u16_be_sse41 LABEL PROC DB 102,15,235,200 ; por %xmm0,%xmm1 DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,148,99,1,0 ; movaps 0x16394(%rip),%xmm8 # 2d750 <_sk_overlay_sse2_8bit+0x11f8> + DB 68,15,40,5,156,94,1,0 ; movaps 0x15e9c(%rip),%xmm8 # 2d640 <_sk_overlay_sse2_8bit+0x1191> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -21768,16 +21935,16 @@ _sk_load_u16_be_sse41 LABEL PROC DB 255,224 ; jmpq *%rax DB 242,65,15,16,4,64 ; movsd (%r8,%rax,2),%xmm0 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,13 ; jne 17442 <_sk_load_u16_be_sse41+0xf9> + DB 117,13 ; jne 1782a <_sk_load_u16_be_sse41+0xf9> DB 243,15,126,192 ; movq %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,59,255,255,255 ; jmpq 1737d <_sk_load_u16_be_sse41+0x34> + DB 233,59,255,255,255 ; jmpq 17765 <_sk_load_u16_be_sse41+0x34> DB 102,65,15,22,68,64,8 ; movhpd 0x8(%r8,%rax,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,38,255,255,255 ; jb 1737d <_sk_load_u16_be_sse41+0x34> + DB 15,130,38,255,255,255 ; jb 17765 <_sk_load_u16_be_sse41+0x34> DB 243,65,15,126,76,64,16 ; movq 0x10(%r8,%rax,2),%xmm1 - DB 233,26,255,255,255 ; jmpq 1737d <_sk_load_u16_be_sse41+0x34> + DB 233,26,255,255,255 ; jmpq 17765 <_sk_load_u16_be_sse41+0x34> PUBLIC _sk_load_rgb_u16_be_sse41 _sk_load_rgb_u16_be_sse41 LABEL PROC @@ -21792,7 +21959,7 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 72,141,4,64 ; lea (%rax,%rax,2),%rax DB 72,193,248,32 ; sar $0x20,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,170,0,0,0 ; jne 17538 <_sk_load_rgb_u16_be_sse41+0xd5> + DB 15,133,170,0,0,0 ; jne 17920 <_sk_load_rgb_u16_be_sse41+0xd5> DB 243,65,15,111,20,64 ; movdqu (%r8,%rax,2),%xmm2 DB 243,65,15,111,92,64,8 ; movdqu 0x8(%r8,%rax,2),%xmm3 DB 102,15,115,219,4 ; psrldq $0x4,%xmm3 @@ -21800,8 +21967,8 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,115,217,6 ; psrldq $0x6,%xmm1 - DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 + DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 @@ -21812,7 +21979,7 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 102,15,235,200 ; por %xmm0,%xmm1 DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,99,98,1,0 ; movaps 0x16263(%rip),%xmm8 # 2d750 <_sk_overlay_sse2_8bit+0x11f8> + DB 68,15,40,5,107,93,1,0 ; movaps 0x15d6b(%rip),%xmm8 # 2d640 <_sk_overlay_sse2_8bit+0x1191> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -21829,27 +21996,27 @@ _sk_load_rgb_u16_be_sse41 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,138,95,1,0 ; movaps 0x15f8a(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,146,90,1,0 ; movaps 0x15a92(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax DB 102,65,15,110,20,64 ; movd (%r8,%rax,2),%xmm2 - DB 102,65,15,196,84,64,4,2 ; pinsrw $0x2,0x4(%r8,%rax,2),%xmm2 DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,65,15,196,84,64,4,2 ; pinsrw $0x2,0x4(%r8,%rax,2),%xmm2 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,13 ; jne 1755d <_sk_load_rgb_u16_be_sse41+0xfa> + DB 117,13 ; jne 17945 <_sk_load_rgb_u16_be_sse41+0xfa> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,85,255,255,255 ; jmpq 174b2 <_sk_load_rgb_u16_be_sse41+0x4f> + DB 233,85,255,255,255 ; jmpq 1789a <_sk_load_rgb_u16_be_sse41+0x4f> DB 102,65,15,110,68,64,6 ; movd 0x6(%r8,%rax,2),%xmm0 - DB 102,65,15,196,68,64,10,2 ; pinsrw $0x2,0xa(%r8,%rax,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,65,15,196,68,64,10,2 ; pinsrw $0x2,0xa(%r8,%rax,2),%xmm0 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,24 ; jb 1758e <_sk_load_rgb_u16_be_sse41+0x12b> + DB 114,24 ; jb 17976 <_sk_load_rgb_u16_be_sse41+0x12b> DB 102,65,15,110,92,64,12 ; movd 0xc(%r8,%rax,2),%xmm3 - DB 102,65,15,196,92,64,16,2 ; pinsrw $0x2,0x10(%r8,%rax,2),%xmm3 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,36,255,255,255 ; jmpq 174b2 <_sk_load_rgb_u16_be_sse41+0x4f> + DB 102,65,15,196,92,64,16,2 ; pinsrw $0x2,0x10(%r8,%rax,2),%xmm3 + DB 233,36,255,255,255 ; jmpq 1789a <_sk_load_rgb_u16_be_sse41+0x4f> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 233,27,255,255,255 ; jmpq 174b2 <_sk_load_rgb_u16_be_sse41+0x4f> + DB 233,27,255,255,255 ; jmpq 1789a <_sk_load_rgb_u16_be_sse41+0x4f> PUBLIC _sk_store_u16_be_sse41 _sk_store_u16_be_sse41 LABEL PROC @@ -21861,44 +22028,44 @@ _sk_store_u16_be_sse41 LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax - DB 68,15,40,21,51,100,1,0 ; movaps 0x16433(%rip),%xmm10 # 2d9f0 <_sk_overlay_sse2_8bit+0x1498> + DB 68,15,40,21,75,95,1,0 ; movaps 0x15f4b(%rip),%xmm10 # 2d8f0 <_sk_overlay_sse2_8bit+0x1441> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 - DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8 + DB 102,68,15,56,43,192 ; packusdw %xmm0,%xmm8 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9 DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 69,15,89,202 ; mulps %xmm10,%xmm9 - DB 102,69,15,91,217 ; cvtps2dq %xmm9,%xmm11 - DB 102,69,15,56,43,219 ; packusdw %xmm11,%xmm11 - DB 102,69,15,111,203 ; movdqa %xmm11,%xmm9 - DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9 - DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 - DB 102,69,15,235,217 ; por %xmm9,%xmm11 + DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 + DB 102,68,15,56,43,200 ; packusdw %xmm0,%xmm9 + DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 + DB 102,65,15,113,243,8 ; psllw $0x8,%xmm11 + DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 + DB 102,69,15,235,203 ; por %xmm11,%xmm9 + DB 102,69,15,97,193 ; punpcklwd %xmm9,%xmm8 DB 68,15,40,202 ; movaps %xmm2,%xmm9 DB 69,15,89,202 ; mulps %xmm10,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 - DB 102,69,15,56,43,201 ; packusdw %xmm9,%xmm9 - DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 - DB 102,65,15,113,244,8 ; psllw $0x8,%xmm12 + DB 102,68,15,56,43,200 ; packusdw %xmm0,%xmm9 + DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 + DB 102,65,15,113,243,8 ; psllw $0x8,%xmm11 DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 - DB 102,69,15,235,204 ; por %xmm12,%xmm9 + DB 102,69,15,235,203 ; por %xmm11,%xmm9 DB 68,15,89,211 ; mulps %xmm3,%xmm10 DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10 - DB 102,69,15,56,43,210 ; packusdw %xmm10,%xmm10 - DB 102,69,15,111,226 ; movdqa %xmm10,%xmm12 - DB 102,65,15,113,244,8 ; psllw $0x8,%xmm12 + DB 102,68,15,56,43,208 ; packusdw %xmm0,%xmm10 + DB 102,69,15,111,218 ; movdqa %xmm10,%xmm11 + DB 102,65,15,113,243,8 ; psllw $0x8,%xmm11 DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 - DB 102,69,15,235,212 ; por %xmm12,%xmm10 - DB 102,69,15,97,195 ; punpcklwd %xmm11,%xmm8 + DB 102,69,15,235,211 ; por %xmm11,%xmm10 DB 102,69,15,97,202 ; punpcklwd %xmm10,%xmm9 DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 DB 102,69,15,98,209 ; punpckldq %xmm9,%xmm10 DB 72,133,255 ; test %rdi,%rdi - DB 117,21 ; jne 1768b <_sk_store_u16_be_sse41+0xf4> + DB 117,21 ; jne 17a73 <_sk_store_u16_be_sse41+0xf4> DB 69,15,17,20,64 ; movups %xmm10,(%r8,%rax,2) DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 DB 243,69,15,127,68,64,16 ; movdqu %xmm8,0x10(%r8,%rax,2) @@ -21906,13 +22073,13 @@ _sk_store_u16_be_sse41 LABEL PROC DB 255,224 ; jmpq *%rax DB 102,69,15,214,20,64 ; movq %xmm10,(%r8,%rax,2) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 17687 <_sk_store_u16_be_sse41+0xf0> + DB 116,240 ; je 17a6f <_sk_store_u16_be_sse41+0xf0> DB 102,69,15,23,84,64,8 ; movhpd %xmm10,0x8(%r8,%rax,2) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 17687 <_sk_store_u16_be_sse41+0xf0> + DB 114,227 ; jb 17a6f <_sk_store_u16_be_sse41+0xf0> DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 DB 102,69,15,214,68,64,16 ; movq %xmm8,0x10(%r8,%rax,2) - DB 235,213 ; jmp 17687 <_sk_store_u16_be_sse41+0xf0> + DB 235,213 ; jmp 17a6f <_sk_store_u16_be_sse41+0xf0> PUBLIC _sk_load_f32_sse41 _sk_load_f32_sse41 LABEL PROC @@ -21926,7 +22093,7 @@ _sk_load_f32_sse41 LABEL PROC DB 73,99,193 ; movslq %r9d,%rax DB 69,15,16,4,128 ; movups (%r8,%rax,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 1771d <_sk_load_f32_sse41+0x6b> + DB 117,66 ; jne 17b05 <_sk_load_f32_sse41+0x6b> DB 65,15,16,68,128,16 ; movups 0x10(%r8,%rax,4),%xmm0 DB 65,15,16,92,128,32 ; movups 0x20(%r8,%rax,4),%xmm3 DB 69,15,16,76,128,48 ; movups 0x30(%r8,%rax,4),%xmm9 @@ -21946,17 +22113,17 @@ _sk_load_f32_sse41 LABEL PROC DB 255,224 ; jmpq *%rax DB 69,15,87,201 ; xorps %xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,8 ; jne 1772f <_sk_load_f32_sse41+0x7d> + DB 117,8 ; jne 17b17 <_sk_load_f32_sse41+0x7d> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 - DB 235,190 ; jmp 176ed <_sk_load_f32_sse41+0x3b> + DB 235,190 ; jmp 17ad5 <_sk_load_f32_sse41+0x3b> DB 65,15,16,68,128,16 ; movups 0x10(%r8,%rax,4),%xmm0 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,8 ; jb 17743 <_sk_load_f32_sse41+0x91> + DB 114,8 ; jb 17b2b <_sk_load_f32_sse41+0x91> DB 65,15,16,92,128,32 ; movups 0x20(%r8,%rax,4),%xmm3 - DB 235,170 ; jmp 176ed <_sk_load_f32_sse41+0x3b> + DB 235,170 ; jmp 17ad5 <_sk_load_f32_sse41+0x3b> DB 15,87,219 ; xorps %xmm3,%xmm3 - DB 235,165 ; jmp 176ed <_sk_load_f32_sse41+0x3b> + DB 235,165 ; jmp 17ad5 <_sk_load_f32_sse41+0x3b> PUBLIC _sk_load_f32_dst_sse41 _sk_load_f32_dst_sse41 LABEL PROC @@ -21970,7 +22137,7 @@ _sk_load_f32_dst_sse41 LABEL PROC DB 73,99,193 ; movslq %r9d,%rax DB 69,15,16,4,128 ; movups (%r8,%rax,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 177b3 <_sk_load_f32_dst_sse41+0x6b> + DB 117,66 ; jne 17b9b <_sk_load_f32_dst_sse41+0x6b> DB 65,15,16,100,128,16 ; movups 0x10(%r8,%rax,4),%xmm4 DB 65,15,16,124,128,32 ; movups 0x20(%r8,%rax,4),%xmm7 DB 69,15,16,76,128,48 ; movups 0x30(%r8,%rax,4),%xmm9 @@ -21990,17 +22157,17 @@ _sk_load_f32_dst_sse41 LABEL PROC DB 255,224 ; jmpq *%rax DB 69,15,87,201 ; xorps %xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,8 ; jne 177c5 <_sk_load_f32_dst_sse41+0x7d> + DB 117,8 ; jne 17bad <_sk_load_f32_dst_sse41+0x7d> DB 15,87,255 ; xorps %xmm7,%xmm7 DB 15,87,228 ; xorps %xmm4,%xmm4 - DB 235,190 ; jmp 17783 <_sk_load_f32_dst_sse41+0x3b> + DB 235,190 ; jmp 17b6b <_sk_load_f32_dst_sse41+0x3b> DB 65,15,16,100,128,16 ; movups 0x10(%r8,%rax,4),%xmm4 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,8 ; jb 177d9 <_sk_load_f32_dst_sse41+0x91> + DB 114,8 ; jb 17bc1 <_sk_load_f32_dst_sse41+0x91> DB 65,15,16,124,128,32 ; movups 0x20(%r8,%rax,4),%xmm7 - DB 235,170 ; jmp 17783 <_sk_load_f32_dst_sse41+0x3b> + DB 235,170 ; jmp 17b6b <_sk_load_f32_dst_sse41+0x3b> DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 235,165 ; jmp 17783 <_sk_load_f32_dst_sse41+0x3b> + DB 235,165 ; jmp 17b6b <_sk_load_f32_dst_sse41+0x3b> PUBLIC _sk_store_f32_sse41 _sk_store_f32_sse41 LABEL PROC @@ -22027,7 +22194,7 @@ _sk_store_f32_sse41 LABEL PROC DB 102,69,15,20,203 ; unpcklpd %xmm11,%xmm9 DB 102,69,15,17,36,128 ; movupd %xmm12,(%r8,%rax,4) DB 72,133,255 ; test %rdi,%rdi - DB 117,29 ; jne 1785b <_sk_store_f32_sse41+0x7d> + DB 117,29 ; jne 17c43 <_sk_store_f32_sse41+0x7d> DB 102,69,15,21,211 ; unpckhpd %xmm11,%xmm10 DB 69,15,17,68,128,16 ; movups %xmm8,0x10(%r8,%rax,4) DB 102,69,15,17,76,128,32 ; movupd %xmm9,0x20(%r8,%rax,4) @@ -22035,12 +22202,12 @@ _sk_store_f32_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,246 ; je 17857 <_sk_store_f32_sse41+0x79> + DB 116,246 ; je 17c3f <_sk_store_f32_sse41+0x79> DB 69,15,17,68,128,16 ; movups %xmm8,0x10(%r8,%rax,4) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,234 ; jb 17857 <_sk_store_f32_sse41+0x79> + DB 114,234 ; jb 17c3f <_sk_store_f32_sse41+0x79> DB 102,69,15,17,76,128,32 ; movupd %xmm9,0x20(%r8,%rax,4) - DB 235,225 ; jmp 17857 <_sk_store_f32_sse41+0x79> + DB 235,225 ; jmp 17c3f <_sk_store_f32_sse41+0x79> PUBLIC _sk_clamp_x_sse41 _sk_clamp_x_sse41 LABEL PROC @@ -22120,7 +22287,7 @@ _sk_mirror_x_sse41 LABEL PROC DB 65,15,92,194 ; subps %xmm10,%xmm0 DB 243,69,15,88,192 ; addss %xmm8,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 243,68,15,89,13,37,89,1,0 ; mulss 0x15925(%rip),%xmm9 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 243,68,15,89,13,29,84,1,0 ; mulss 0x1541d(%rip),%xmm9 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 @@ -22149,7 +22316,7 @@ _sk_mirror_y_sse41 LABEL PROC DB 65,15,92,202 ; subps %xmm10,%xmm1 DB 243,69,15,88,192 ; addss %xmm8,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 - DB 243,68,15,89,13,179,88,1,0 ; mulss 0x158b3(%rip),%xmm9 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 243,68,15,89,13,171,83,1,0 ; mulss 0x153ab(%rip),%xmm9 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,89,201 ; mulps %xmm1,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 @@ -22172,7 +22339,7 @@ PUBLIC _sk_clamp_x_1_sse41 _sk_clamp_x_1_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,118,90,1,0 ; minps 0x15a76(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,93,5,126,85,1,0 ; minps 0x1557e(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -22186,9 +22353,9 @@ _sk_repeat_x_1_sse41 LABEL PROC PUBLIC _sk_mirror_x_1_sse41 _sk_mirror_x_1_sse41 LABEL PROC - DB 68,15,40,5,183,90,1,0 ; movaps 0x15ab7(%rip),%xmm8 # 2d520 <_sk_overlay_sse2_8bit+0xfc8> + DB 68,15,40,5,191,85,1,0 ; movaps 0x155bf(%rip),%xmm8 # 2d410 <_sk_overlay_sse2_8bit+0xf61> DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,13,59,90,1,0 ; movaps 0x15a3b(%rip),%xmm9 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,13,67,85,1,0 ; movaps 0x15543(%rip),%xmm9 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 102,69,15,58,8,201,1 ; roundps $0x1,%xmm9,%xmm9 DB 69,15,88,201 ; addps %xmm9,%xmm9 @@ -22203,10 +22370,10 @@ _sk_mirror_x_1_sse41 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse41 _sk_luminance_to_alpha_sse41 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,90,95,1,0 ; mulps 0x15f5a(%rip),%xmm0 # 2da00 <_sk_overlay_sse2_8bit+0x14a8> - DB 15,89,13,99,95,1,0 ; mulps 0x15f63(%rip),%xmm1 # 2da10 <_sk_overlay_sse2_8bit+0x14b8> + DB 15,89,5,114,90,1,0 ; mulps 0x15a72(%rip),%xmm0 # 2d900 <_sk_overlay_sse2_8bit+0x1451> + DB 15,89,13,123,90,1,0 ; mulps 0x15a7b(%rip),%xmm1 # 2d910 <_sk_overlay_sse2_8bit+0x1461> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,105,95,1,0 ; mulps 0x15f69(%rip),%xmm3 # 2da20 <_sk_overlay_sse2_8bit+0x14c8> + DB 15,89,29,129,90,1,0 ; mulps 0x15a81(%rip),%xmm3 # 2d920 <_sk_overlay_sse2_8bit+0x1471> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -22502,9 +22669,9 @@ _sk_evenly_spaced_gradient_sse41 LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 76,139,112,8 ; mov 0x8(%rax),%r14 DB 72,255,203 ; dec %rbx - DB 120,7 ; js 17f57 <_sk_evenly_spaced_gradient_sse41+0x18> + DB 120,7 ; js 1833f <_sk_evenly_spaced_gradient_sse41+0x18> DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 - DB 235,21 ; jmp 17f6c <_sk_evenly_spaced_gradient_sse41+0x2d> + DB 235,21 ; jmp 18354 <_sk_evenly_spaced_gradient_sse41+0x2d> DB 73,137,216 ; mov %rbx,%r8 DB 73,209,232 ; shr %r8 DB 131,227,1 ; and $0x1,%ebx @@ -22591,15 +22758,15 @@ _sk_evenly_spaced_gradient_sse41 LABEL PROC PUBLIC _sk_gauss_a_to_rgba_sse41 _sk_gauss_a_to_rgba_sse41 LABEL PROC - DB 15,40,5,24,89,1,0 ; movaps 0x15918(%rip),%xmm0 # 2da30 <_sk_overlay_sse2_8bit+0x14d8> + DB 15,40,5,48,84,1,0 ; movaps 0x15430(%rip),%xmm0 # 2d930 <_sk_overlay_sse2_8bit+0x1481> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,30,89,1,0 ; addps 0x1591e(%rip),%xmm0 # 2da40 <_sk_overlay_sse2_8bit+0x14e8> + DB 15,88,5,54,84,1,0 ; addps 0x15436(%rip),%xmm0 # 2d940 <_sk_overlay_sse2_8bit+0x1491> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,36,89,1,0 ; addps 0x15924(%rip),%xmm0 # 2da50 <_sk_overlay_sse2_8bit+0x14f8> + DB 15,88,5,60,84,1,0 ; addps 0x1543c(%rip),%xmm0 # 2d950 <_sk_overlay_sse2_8bit+0x14a1> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,42,89,1,0 ; addps 0x1592a(%rip),%xmm0 # 2da60 <_sk_overlay_sse2_8bit+0x1508> + DB 15,88,5,66,84,1,0 ; addps 0x15442(%rip),%xmm0 # 2d960 <_sk_overlay_sse2_8bit+0x14b1> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,48,89,1,0 ; addps 0x15930(%rip),%xmm0 # 2da70 <_sk_overlay_sse2_8bit+0x1518> + DB 15,88,5,72,84,1,0 ; addps 0x15448(%rip),%xmm0 # 2d970 <_sk_overlay_sse2_8bit+0x14c1> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -22608,26 +22775,24 @@ _sk_gauss_a_to_rgba_sse41 LABEL PROC PUBLIC _sk_gradient_sse41 _sk_gradient_sse41 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,51 ; jb 1818f <_sk_gradient_sse41+0x42> + DB 114,41 ; jb 18570 <_sk_gradient_sse41+0x3b> DB 76,139,72,72 ; mov 0x48(%rax),%r9 DB 73,255,200 ; dec %r8 DB 73,131,193,4 ; add $0x4,%r9 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,21,94,83,1,0 ; movaps 0x1535e(%rip),%xmm2 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 243,65,15,16,25 ; movss (%r9),%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 - DB 15,84,218 ; andps %xmm2,%xmm3 - DB 102,15,254,203 ; paddd %xmm3,%xmm1 + DB 243,65,15,16,17 ; movss (%r9),%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,194,208,2 ; cmpleps %xmm0,%xmm2 + DB 102,15,250,202 ; psubd %xmm2,%xmm1 DB 73,131,193,4 ; add $0x4,%r9 DB 73,255,200 ; dec %r8 - DB 117,227 ; jne 18172 <_sk_gradient_sse41+0x25> - DB 65,86 ; push %r14 - DB 83 ; push %rbx + DB 117,230 ; jne 18556 <_sk_gradient_sse41+0x21> DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 73,193,232,32 ; shr $0x20,%r8 @@ -22752,26 +22917,26 @@ _sk_xy_to_unit_angle_sse41 LABEL PROC DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,40,236 ; movaps %xmm12,%xmm13 DB 69,15,89,237 ; mulps %xmm13,%xmm13 - DB 68,15,40,21,157,86,1,0 ; movaps 0x1569d(%rip),%xmm10 # 2da80 <_sk_overlay_sse2_8bit+0x1528> + DB 68,15,40,21,191,81,1,0 ; movaps 0x151bf(%rip),%xmm10 # 2d980 <_sk_overlay_sse2_8bit+0x14d1> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,161,86,1,0 ; addps 0x156a1(%rip),%xmm10 # 2da90 <_sk_overlay_sse2_8bit+0x1538> + DB 68,15,88,21,195,81,1,0 ; addps 0x151c3(%rip),%xmm10 # 2d990 <_sk_overlay_sse2_8bit+0x14e1> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,165,86,1,0 ; addps 0x156a5(%rip),%xmm10 # 2daa0 <_sk_overlay_sse2_8bit+0x1548> + DB 68,15,88,21,199,81,1,0 ; addps 0x151c7(%rip),%xmm10 # 2d9a0 <_sk_overlay_sse2_8bit+0x14f1> DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,21,169,86,1,0 ; addps 0x156a9(%rip),%xmm10 # 2dab0 <_sk_overlay_sse2_8bit+0x1558> + DB 68,15,88,21,203,81,1,0 ; addps 0x151cb(%rip),%xmm10 # 2d9b0 <_sk_overlay_sse2_8bit+0x1501> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,194,195,1 ; cmpltps %xmm11,%xmm0 - DB 68,15,40,29,168,86,1,0 ; movaps 0x156a8(%rip),%xmm11 # 2dac0 <_sk_overlay_sse2_8bit+0x1568> + DB 68,15,40,29,202,81,1,0 ; movaps 0x151ca(%rip),%xmm11 # 2d9c0 <_sk_overlay_sse2_8bit+0x1511> DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 102,69,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm10 DB 69,15,194,200,1 ; cmpltps %xmm8,%xmm9 - DB 68,15,40,29,129,80,1,0 ; movaps 0x15081(%rip),%xmm11 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,29,147,75,1,0 ; movaps 0x14b93(%rip),%xmm11 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 102,69,15,56,20,211 ; blendvps %xmm0,%xmm11,%xmm10 DB 15,40,193 ; movaps %xmm1,%xmm0 DB 65,15,194,192,1 ; cmpltps %xmm8,%xmm0 - DB 68,15,40,13,115,80,1,0 ; movaps 0x15073(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,133,75,1,0 ; movaps 0x14b85(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 102,69,15,56,20,209 ; blendvps %xmm0,%xmm9,%xmm10 DB 69,15,194,194,7 ; cmpordps %xmm10,%xmm8 @@ -22801,7 +22966,7 @@ _sk_xy_to_2pt_conical_quadratic_max_sse41 LABEL PROC DB 243,69,15,89,203 ; mulss %xmm11,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,89,13,28,86,1,0 ; mulps 0x1561c(%rip),%xmm9 # 2dad0 <_sk_overlay_sse2_8bit+0x1578> + DB 68,15,89,13,62,81,1,0 ; mulps 0x1513e(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1521> DB 15,89,192 ; mulps %xmm0,%xmm0 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 @@ -22809,7 +22974,7 @@ _sk_xy_to_2pt_conical_quadratic_max_sse41 LABEL PROC DB 243,69,15,89,219 ; mulss %xmm11,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 68,15,89,21,7,86,1,0 ; mulps 0x15607(%rip),%xmm10 # 2dae0 <_sk_overlay_sse2_8bit+0x1588> + DB 68,15,89,21,41,81,1,0 ; mulps 0x15129(%rip),%xmm10 # 2d9e0 <_sk_overlay_sse2_8bit+0x1531> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 @@ -22818,8 +22983,8 @@ _sk_xy_to_2pt_conical_quadratic_max_sse41 LABEL PROC DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 68,15,87,13,191,84,1,0 ; xorps 0x154bf(%rip),%xmm9 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 68,15,89,5,167,79,1,0 ; mulps 0x14fa7(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,87,13,209,79,1,0 ; xorps 0x14fd1(%rip),%xmm9 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> + DB 68,15,89,5,185,74,1,0 ; mulps 0x14ab9(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 @@ -22838,7 +23003,7 @@ _sk_xy_to_2pt_conical_quadratic_min_sse41 LABEL PROC DB 243,69,15,89,203 ; mulss %xmm11,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,89,13,126,85,1,0 ; mulps 0x1557e(%rip),%xmm9 # 2dad0 <_sk_overlay_sse2_8bit+0x1578> + DB 68,15,89,13,160,80,1,0 ; mulps 0x150a0(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1521> DB 15,89,192 ; mulps %xmm0,%xmm0 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 @@ -22846,7 +23011,7 @@ _sk_xy_to_2pt_conical_quadratic_min_sse41 LABEL PROC DB 243,69,15,89,219 ; mulss %xmm11,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 68,15,89,21,105,85,1,0 ; mulps 0x15569(%rip),%xmm10 # 2dae0 <_sk_overlay_sse2_8bit+0x1588> + DB 68,15,89,21,139,80,1,0 ; mulps 0x1508b(%rip),%xmm10 # 2d9e0 <_sk_overlay_sse2_8bit+0x1531> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 @@ -22855,8 +23020,8 @@ _sk_xy_to_2pt_conical_quadratic_min_sse41 LABEL PROC DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 68,15,87,13,33,84,1,0 ; xorps 0x15421(%rip),%xmm9 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 68,15,89,5,9,79,1,0 ; mulps 0x14f09(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,87,13,51,79,1,0 ; xorps 0x14f33(%rip),%xmm9 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> + DB 68,15,89,5,27,74,1,0 ; mulps 0x14a1b(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 @@ -22872,7 +23037,7 @@ _sk_xy_to_2pt_conical_linear_sse41 LABEL PROC DB 243,69,15,89,200 ; mulss %xmm8,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,89,13,241,84,1,0 ; mulps 0x154f1(%rip),%xmm9 # 2dad0 <_sk_overlay_sse2_8bit+0x1578> + DB 68,15,89,13,19,80,1,0 ; mulps 0x15013(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1521> DB 15,89,192 ; mulps %xmm0,%xmm0 DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 69,15,89,210 ; mulps %xmm10,%xmm10 @@ -22880,7 +23045,7 @@ _sk_xy_to_2pt_conical_linear_sse41 LABEL PROC DB 243,69,15,89,192 ; mulss %xmm8,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 65,15,92,192 ; subps %xmm8,%xmm0 - DB 15,87,5,189,83,1,0 ; xorps 0x153bd(%rip),%xmm0 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> + DB 15,87,5,207,78,1,0 ; xorps 0x14ecf(%rip),%xmm0 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> DB 65,15,94,193 ; divps %xmm9,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -22917,7 +23082,7 @@ _sk_apply_vector_mask_sse41 LABEL PROC PUBLIC _sk_save_xy_sse41 _sk_save_xy_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,67,78,1,0 ; movaps 0x14e43(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,5,85,73,1,0 ; movaps 0x14955(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -22957,8 +23122,8 @@ _sk_bilinear_nx_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,245,83,1,0 ; addps 0x153f5(%rip),%xmm0 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,189,77,1,0 ; movaps 0x14dbd(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,5,23,79,1,0 ; addps 0x14f17(%rip),%xmm0 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,207,72,1,0 ; movaps 0x148cf(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -22969,7 +23134,7 @@ _sk_bilinear_px_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,140,77,1,0 ; addps 0x14d8c(%rip),%xmm0 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,88,5,158,72,1,0 ; addps 0x1489e(%rip),%xmm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -22979,8 +23144,8 @@ _sk_bilinear_ny_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,174,83,1,0 ; addps 0x153ae(%rip),%xmm1 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,118,77,1,0 ; movaps 0x14d76(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,13,208,78,1,0 ; addps 0x14ed0(%rip),%xmm1 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,136,72,1,0 ; movaps 0x14888(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -22991,7 +23156,7 @@ _sk_bilinear_py_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,68,77,1,0 ; addps 0x14d44(%rip),%xmm1 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,88,13,86,72,1,0 ; addps 0x14856(%rip),%xmm1 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -23001,13 +23166,13 @@ _sk_bicubic_n3x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,119,83,1,0 ; addps 0x15377(%rip),%xmm0 # 2db00 <_sk_overlay_sse2_8bit+0x15a8> - DB 68,15,40,13,47,77,1,0 ; movaps 0x14d2f(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,5,153,78,1,0 ; addps 0x14e99(%rip),%xmm0 # 2da00 <_sk_overlay_sse2_8bit+0x1551> + DB 68,15,40,13,65,72,1,0 ; movaps 0x14841(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,107,83,1,0 ; mulps 0x1536b(%rip),%xmm9 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,13,3,79,1,0 ; addps 0x14f03(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,13,141,78,1,0 ; mulps 0x14e8d(%rip),%xmm9 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,13,21,74,1,0 ; addps 0x14a15(%rip),%xmm9 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -23018,16 +23183,16 @@ _sk_bicubic_n1x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,34,83,1,0 ; addps 0x15322(%rip),%xmm0 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,234,76,1,0 ; movaps 0x14cea(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,5,68,78,1,0 ; addps 0x14e44(%rip),%xmm0 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,252,71,1,0 ; movaps 0x147fc(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,62,83,1,0 ; movaps 0x1533e(%rip),%xmm8 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,5,96,78,1,0 ; movaps 0x14e60(%rip),%xmm8 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,66,83,1,0 ; addps 0x15342(%rip),%xmm8 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,5,100,78,1,0 ; addps 0x14e64(%rip),%xmm8 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,182,76,1,0 ; addps 0x14cb6(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,88,5,200,71,1,0 ; addps 0x147c8(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,58,83,1,0 ; addps 0x1533a(%rip),%xmm8 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,5,92,78,1,0 ; addps 0x14e5c(%rip),%xmm8 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -23035,17 +23200,17 @@ _sk_bicubic_n1x_sse41 LABEL PROC PUBLIC _sk_bicubic_p1x_sse41 _sk_bicubic_p1x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,148,76,1,0 ; movaps 0x14c94(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,5,166,71,1,0 ; movaps 0x147a6(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,72,64 ; movups 0x40(%rax),%xmm9 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,21,240,82,1,0 ; movaps 0x152f0(%rip),%xmm10 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,21,18,78,1,0 ; movaps 0x14e12(%rip),%xmm10 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,244,82,1,0 ; addps 0x152f4(%rip),%xmm10 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,21,22,78,1,0 ; addps 0x14e16(%rip),%xmm10 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,240,82,1,0 ; addps 0x152f0(%rip),%xmm10 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,21,18,78,1,0 ; addps 0x14e12(%rip),%xmm10 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,144,128,0,0,0 ; movups %xmm10,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -23055,11 +23220,11 @@ _sk_bicubic_p3x_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,195,82,1,0 ; addps 0x152c3(%rip),%xmm0 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 15,88,5,229,77,1,0 ; addps 0x14de5(%rip),%xmm0 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,147,82,1,0 ; mulps 0x15293(%rip),%xmm8 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,5,43,78,1,0 ; addps 0x14e2b(%rip),%xmm8 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,5,181,77,1,0 ; mulps 0x14db5(%rip),%xmm8 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,5,61,73,1,0 ; addps 0x1493d(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -23070,13 +23235,13 @@ _sk_bicubic_n3y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,89,82,1,0 ; addps 0x15259(%rip),%xmm1 # 2db00 <_sk_overlay_sse2_8bit+0x15a8> - DB 68,15,40,13,17,76,1,0 ; movaps 0x14c11(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,13,123,77,1,0 ; addps 0x14d7b(%rip),%xmm1 # 2da00 <_sk_overlay_sse2_8bit+0x1551> + DB 68,15,40,13,35,71,1,0 ; movaps 0x14723(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,77,82,1,0 ; mulps 0x1524d(%rip),%xmm9 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,13,229,77,1,0 ; addps 0x14de5(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,13,111,77,1,0 ; mulps 0x14d6f(%rip),%xmm9 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,13,247,72,1,0 ; addps 0x148f7(%rip),%xmm9 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -23087,16 +23252,16 @@ _sk_bicubic_n1y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,3,82,1,0 ; addps 0x15203(%rip),%xmm1 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,203,75,1,0 ; movaps 0x14bcb(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,13,37,77,1,0 ; addps 0x14d25(%rip),%xmm1 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,221,70,1,0 ; movaps 0x146dd(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,31,82,1,0 ; movaps 0x1521f(%rip),%xmm8 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,5,65,77,1,0 ; movaps 0x14d41(%rip),%xmm8 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,35,82,1,0 ; addps 0x15223(%rip),%xmm8 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,5,69,77,1,0 ; addps 0x14d45(%rip),%xmm8 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,151,75,1,0 ; addps 0x14b97(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,88,5,169,70,1,0 ; addps 0x146a9(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,27,82,1,0 ; addps 0x1521b(%rip),%xmm8 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,5,61,77,1,0 ; addps 0x14d3d(%rip),%xmm8 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -23104,17 +23269,17 @@ _sk_bicubic_n1y_sse41 LABEL PROC PUBLIC _sk_bicubic_p1y_sse41 _sk_bicubic_p1y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,117,75,1,0 ; movaps 0x14b75(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,5,135,70,1,0 ; movaps 0x14687(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,72,96 ; movups 0x60(%rax),%xmm9 DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 68,15,40,21,208,81,1,0 ; movaps 0x151d0(%rip),%xmm10 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,21,242,76,1,0 ; movaps 0x14cf2(%rip),%xmm10 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,212,81,1,0 ; addps 0x151d4(%rip),%xmm10 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,21,246,76,1,0 ; addps 0x14cf6(%rip),%xmm10 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,208,81,1,0 ; addps 0x151d0(%rip),%xmm10 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,21,242,76,1,0 ; addps 0x14cf2(%rip),%xmm10 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,144,160,0,0,0 ; movups %xmm10,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -23124,11 +23289,11 @@ _sk_bicubic_p3y_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,162,81,1,0 ; addps 0x151a2(%rip),%xmm1 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 15,88,13,196,76,1,0 ; addps 0x14cc4(%rip),%xmm1 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,114,81,1,0 ; mulps 0x15172(%rip),%xmm8 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,5,10,77,1,0 ; addps 0x14d0a(%rip),%xmm8 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,5,148,76,1,0 ; mulps 0x14c94(%rip),%xmm8 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,5,28,72,1,0 ; addps 0x1481c(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -23206,265 +23371,263 @@ _sk_clut_3D_sse41 LABEL PROC DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,200,0,0,0 ; sub $0xc8,%rsp - DB 15,41,188,36,176,0,0,0 ; movaps %xmm7,0xb0(%rsp) - DB 15,41,180,36,160,0,0,0 ; movaps %xmm6,0xa0(%rsp) - DB 15,41,172,36,144,0,0,0 ; movaps %xmm5,0x90(%rsp) - DB 15,41,164,36,128,0,0,0 ; movaps %xmm4,0x80(%rsp) - DB 15,41,92,36,112 ; movaps %xmm3,0x70(%rsp) - DB 72,137,76,36,24 ; mov %rcx,0x18(%rsp) + DB 72,129,236,184,0,0,0 ; sub $0xb8,%rsp + DB 15,41,188,36,160,0,0,0 ; movaps %xmm7,0xa0(%rsp) + DB 15,41,180,36,144,0,0,0 ; movaps %xmm6,0x90(%rsp) + DB 15,41,172,36,128,0,0,0 ; movaps %xmm5,0x80(%rsp) + DB 15,41,100,36,112 ; movaps %xmm4,0x70(%rsp) + DB 15,41,92,36,96 ; movaps %xmm3,0x60(%rsp) + DB 15,40,216 ; movaps %xmm0,%xmm3 + DB 72,137,76,36,8 ; mov %rcx,0x8(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d - DB 102,65,15,110,217 ; movd %r9d,%xmm3 - DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 - DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,218 ; mulps %xmm2,%xmm3 - DB 243,15,91,227 ; cvttps2dq %xmm3,%xmm4 - DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 68,15,41,68,36,96 ; movaps %xmm8,0x60(%rsp) + DB 102,65,15,110,193 ; movd %r9d,%xmm0 + DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 + DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 + DB 15,89,194 ; mulps %xmm2,%xmm0 + DB 15,41,68,36,80 ; movaps %xmm0,0x50(%rsp) + DB 243,15,91,240 ; cvttps2dq %xmm0,%xmm6 + DB 15,40,37,151,75,1,0 ; movaps 0x14b97(%rip),%xmm4 # 2da50 <_sk_overlay_sse2_8bit+0x15a1> + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 15,41,68,36,48 ; movaps %xmm0,0x30(%rsp) DB 102,65,15,110,208 ; movd %r8d,%xmm2 - DB 102,15,112,234,0 ; pshufd $0x0,%xmm2,%xmm5 - DB 102,15,127,108,36,32 ; movdqa %xmm5,0x20(%rsp) + DB 102,68,15,112,202,0 ; pshufd $0x0,%xmm2,%xmm9 DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d - DB 102,65,15,110,217 ; movd %r9d,%xmm3 - DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 - DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,217 ; mulps %xmm1,%xmm3 - DB 243,68,15,91,235 ; cvttps2dq %xmm3,%xmm13 - DB 102,68,15,127,108,36,64 ; movdqa %xmm13,0x40(%rsp) - DB 15,40,251 ; movaps %xmm3,%xmm7 - DB 15,41,60,36 ; movaps %xmm7,(%rsp) - DB 102,68,15,56,64,237 ; pmulld %xmm5,%xmm13 - DB 102,65,15,111,205 ; movdqa %xmm13,%xmm1 - DB 102,15,254,204 ; paddd %xmm4,%xmm1 - DB 102,68,15,111,212 ; movdqa %xmm4,%xmm10 - DB 102,68,15,127,84,36,80 ; movdqa %xmm10,0x50(%rsp) - DB 102,65,15,110,216 ; movd %r8d,%xmm3 - DB 102,15,56,64,218 ; pmulld %xmm2,%xmm3 - DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 + DB 102,65,15,110,233 ; movd %r9d,%xmm5 + DB 102,15,112,237,0 ; pshufd $0x0,%xmm5,%xmm5 + DB 15,91,253 ; cvtdq2ps %xmm5,%xmm7 + DB 15,89,249 ; mulps %xmm1,%xmm7 + DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp) + DB 243,15,91,239 ; cvttps2dq %xmm7,%xmm5 + DB 102,15,127,108,36,32 ; movdqa %xmm5,0x20(%rsp) + DB 15,40,207 ; movaps %xmm7,%xmm1 + DB 15,88,204 ; addps %xmm4,%xmm1 + DB 243,68,15,91,209 ; cvttps2dq %xmm1,%xmm10 + DB 102,69,15,56,64,209 ; pmulld %xmm9,%xmm10 + DB 102,68,15,56,64,205 ; pmulld %xmm5,%xmm9 + DB 102,65,15,110,192 ; movd %r8d,%xmm0 + DB 102,15,56,64,194 ; pmulld %xmm2,%xmm0 DB 68,139,64,8 ; mov 0x8(%rax),%r8d DB 65,255,200 ; dec %r8d - DB 102,65,15,110,208 ; movd %r8d,%xmm2 - DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 - DB 68,15,91,242 ; cvtdq2ps %xmm2,%xmm14 - DB 68,15,89,240 ; mulps %xmm0,%xmm14 - DB 243,65,15,91,198 ; cvttps2dq %xmm14,%xmm0 - DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 - DB 102,68,15,56,64,227 ; pmulld %xmm3,%xmm12 - DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 - DB 102,15,254,209 ; paddd %xmm1,%xmm2 - DB 102,15,111,37,232,79,1,0 ; movdqa 0x14fe8(%rip),%xmm4 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,56,64,212 ; pmulld %xmm4,%xmm2 - DB 102,68,15,111,220 ; movdqa %xmm4,%xmm11 - DB 102,68,15,111,61,69,73,1,0 ; movdqa 0x14945(%rip),%xmm15 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,65,15,254,231 ; paddd %xmm15,%xmm4 - DB 102,73,15,58,22,224,1 ; pextrq $0x1,%xmm4,%r8 - DB 102,73,15,126,225 ; movq %xmm4,%r9 + DB 102,65,15,110,200 ; movd %r8d,%xmm1 + DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 + DB 68,15,91,217 ; cvtdq2ps %xmm1,%xmm11 + DB 68,15,89,219 ; mulps %xmm3,%xmm11 + DB 102,65,15,111,209 ; movdqa %xmm9,%xmm2 + DB 102,15,111,238 ; movdqa %xmm6,%xmm5 + DB 102,15,127,108,36,64 ; movdqa %xmm5,0x40(%rsp) + DB 102,15,254,213 ; paddd %xmm5,%xmm2 + DB 102,68,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm8 + DB 65,15,88,227 ; addps %xmm11,%xmm4 + DB 243,68,15,91,244 ; cvttps2dq %xmm4,%xmm14 + DB 102,69,15,56,64,240 ; pmulld %xmm8,%xmm14 + DB 243,65,15,91,195 ; cvttps2dq %xmm11,%xmm0 + DB 102,68,15,56,64,192 ; pmulld %xmm0,%xmm8 + DB 102,65,15,111,216 ; movdqa %xmm8,%xmm3 + DB 102,15,254,218 ; paddd %xmm2,%xmm3 + DB 102,68,15,111,37,238,74,1,0 ; movdqa 0x14aee(%rip),%xmm12 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,65,15,56,64,220 ; pmulld %xmm12,%xmm3 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,250,204 ; psubd %xmm4,%xmm1 + DB 102,69,15,118,237 ; pcmpeqd %xmm13,%xmm13 + DB 102,73,15,58,22,200,1 ; pextrq $0x1,%xmm1,%r8 + DB 102,73,15,126,201 ; movq %xmm1,%r9 DB 72,139,0 ; mov (%rax),%rax DB 69,137,202 ; mov %r9d,%r10d DB 73,193,233,32 ; shr $0x20,%r9 DB 69,137,195 ; mov %r8d,%r11d - DB 243,66,15,16,36,144 ; movss (%rax,%r10,4),%xmm4 - DB 102,66,15,58,33,36,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm4 - DB 243,66,15,16,44,152 ; movss (%rax,%r11,4),%xmm5 - DB 102,15,58,33,229,32 ; insertps $0x20,%xmm5,%xmm4 + DB 243,66,15,16,12,144 ; movss (%rax,%r10,4),%xmm1 + DB 102,66,15,58,33,12,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm1 + DB 243,66,15,16,36,152 ; movss (%rax,%r11,4),%xmm4 + DB 102,15,58,33,204,32 ; insertps $0x20,%xmm4,%xmm1 DB 73,193,232,32 ; shr $0x20,%r8 - DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 - DB 102,15,58,33,229,48 ; insertps $0x30,%xmm5,%xmm4 - DB 102,65,15,126,208 ; movd %xmm2,%r8d - DB 102,65,15,58,22,209,1 ; pextrd $0x1,%xmm2,%r9d - DB 102,65,15,58,22,210,2 ; pextrd $0x2,%xmm2,%r10d - DB 102,65,15,58,22,211,3 ; pextrd $0x3,%xmm2,%r11d - DB 102,68,15,111,13,246,72,1,0 ; movdqa 0x148f6(%rip),%xmm9 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,65,15,254,209 ; paddd %xmm9,%xmm2 - DB 102,73,15,58,22,214,1 ; pextrq $0x1,%xmm2,%r14 - DB 102,72,15,126,211 ; movq %xmm2,%rbx + DB 243,66,15,16,36,128 ; movss (%rax,%r8,4),%xmm4 + DB 102,15,58,33,204,48 ; insertps $0x30,%xmm4,%xmm1 + DB 102,65,15,126,216 ; movd %xmm3,%r8d + DB 102,65,15,58,22,217,1 ; pextrd $0x1,%xmm3,%r9d + DB 102,65,15,58,22,218,2 ; pextrd $0x2,%xmm3,%r10d + DB 102,65,15,58,22,219,3 ; pextrd $0x3,%xmm3,%r11d + DB 102,15,111,37,242,67,1,0 ; movdqa 0x143f2(%rip),%xmm4 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,254,220 ; paddd %xmm4,%xmm3 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 + DB 102,73,15,58,22,222,1 ; pextrq $0x1,%xmm3,%r14 + DB 102,72,15,126,219 ; movq %xmm3,%rbx DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx DB 69,137,244 ; mov %r14d,%r12d - DB 243,66,15,16,52,184 ; movss (%rax,%r15,4),%xmm6 - DB 102,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm6 - DB 243,66,15,16,20,160 ; movss (%rax,%r12,4),%xmm2 - DB 102,15,58,33,242,32 ; insertps $0x20,%xmm2,%xmm6 + DB 243,66,15,16,36,184 ; movss (%rax,%r15,4),%xmm4 + DB 102,15,58,33,36,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm4 + DB 243,66,15,16,28,160 ; movss (%rax,%r12,4),%xmm3 + DB 102,15,58,33,227,32 ; insertps $0x20,%xmm3,%xmm4 DB 73,193,238,32 ; shr $0x20,%r14 - DB 243,66,15,16,20,176 ; movss (%rax,%r14,4),%xmm2 - DB 102,15,58,33,242,48 ; insertps $0x30,%xmm2,%xmm6 - DB 15,40,45,11,79,1,0 ; movaps 0x14f0b(%rip),%xmm5 # 2db50 <_sk_overlay_sse2_8bit+0x15f8> - DB 68,15,88,197 ; addps %xmm5,%xmm8 - DB 68,15,41,68,36,48 ; movaps %xmm8,0x30(%rsp) - DB 15,40,215 ; movaps %xmm7,%xmm2 - DB 15,88,213 ; addps %xmm5,%xmm2 - DB 65,15,88,238 ; addps %xmm14,%xmm5 - DB 243,68,15,91,197 ; cvttps2dq %xmm5,%xmm8 - DB 102,68,15,56,64,195 ; pmulld %xmm3,%xmm8 - DB 102,65,15,254,200 ; paddd %xmm8,%xmm1 - DB 102,65,15,111,251 ; movdqa %xmm11,%xmm7 - DB 102,15,56,64,207 ; pmulld %xmm7,%xmm1 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,65,15,254,223 ; paddd %xmm15,%xmm3 + DB 243,66,15,16,28,176 ; movss (%rax,%r14,4),%xmm3 + DB 102,15,58,33,227,48 ; insertps $0x30,%xmm3,%xmm4 + DB 102,65,15,254,214 ; paddd %xmm14,%xmm2 + DB 102,65,15,56,64,212 ; pmulld %xmm12,%xmm2 + DB 102,15,111,218 ; movdqa %xmm2,%xmm3 + DB 102,65,15,250,221 ; psubd %xmm13,%xmm3 + DB 102,69,15,118,255 ; pcmpeqd %xmm15,%xmm15 DB 102,73,15,58,22,222,1 ; pextrq $0x1,%xmm3,%r14 DB 102,72,15,126,219 ; movq %xmm3,%rbx DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx DB 69,137,244 ; mov %r14d,%r12d - DB 243,66,15,16,44,184 ; movss (%rax,%r15,4),%xmm5 - DB 102,15,58,33,44,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm5 - DB 243,66,15,16,28,160 ; movss (%rax,%r12,4),%xmm3 - DB 102,15,58,33,235,32 ; insertps $0x20,%xmm3,%xmm5 + DB 243,66,15,16,28,184 ; movss (%rax,%r15,4),%xmm3 + DB 102,15,58,33,28,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm3 + DB 243,66,15,16,52,160 ; movss (%rax,%r12,4),%xmm6 + DB 102,15,58,33,222,32 ; insertps $0x20,%xmm6,%xmm3 DB 73,193,238,32 ; shr $0x20,%r14 - DB 243,66,15,16,28,176 ; movss (%rax,%r14,4),%xmm3 - DB 102,15,58,33,235,48 ; insertps $0x30,%xmm3,%xmm5 - DB 102,65,15,126,205 ; movd %xmm1,%r13d - DB 102,65,15,58,22,206,1 ; pextrd $0x1,%xmm1,%r14d - DB 102,65,15,58,22,207,2 ; pextrd $0x2,%xmm1,%r15d - DB 102,65,15,58,22,204,3 ; pextrd $0x3,%xmm1,%r12d - DB 102,65,15,254,201 ; paddd %xmm9,%xmm1 - DB 102,72,15,58,22,203,1 ; pextrq $0x1,%xmm1,%rbx - DB 102,72,15,126,205 ; movq %xmm1,%rbp + DB 243,66,15,16,52,176 ; movss (%rax,%r14,4),%xmm6 + DB 102,15,58,33,222,48 ; insertps $0x30,%xmm6,%xmm3 + DB 102,65,15,126,213 ; movd %xmm2,%r13d + DB 102,65,15,58,22,214,1 ; pextrd $0x1,%xmm2,%r14d + DB 102,65,15,58,22,215,2 ; pextrd $0x2,%xmm2,%r15d + DB 102,65,15,58,22,212,3 ; pextrd $0x3,%xmm2,%r12d + DB 102,15,254,215 ; paddd %xmm7,%xmm2 + DB 102,68,15,111,239 ; movdqa %xmm7,%xmm13 + DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx + DB 102,72,15,126,213 ; movq %xmm2,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 - DB 102,15,58,33,28,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm3 + DB 243,15,16,60,136 ; movss (%rax,%rcx,4),%xmm7 + DB 102,15,58,33,60,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm7 DB 137,217 ; mov %ebx,%ecx - DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 - DB 102,15,58,33,217,32 ; insertps $0x20,%xmm1,%xmm3 + DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 + DB 102,15,58,33,250,32 ; insertps $0x20,%xmm2,%xmm7 DB 72,193,235,32 ; shr $0x20,%rbx - DB 243,15,16,12,152 ; movss (%rax,%rbx,4),%xmm1 - DB 102,15,58,33,217,48 ; insertps $0x30,%xmm1,%xmm3 - DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,92,240 ; subps %xmm0,%xmm14 - DB 243,66,15,16,4,128 ; movss (%rax,%r8,4),%xmm0 - DB 102,66,15,58,33,4,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm0 - DB 102,66,15,58,33,4,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm0 - DB 102,66,15,58,33,4,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm0 - DB 243,66,15,16,12,168 ; movss (%rax,%r13,4),%xmm1 - DB 102,66,15,58,33,12,176,16 ; insertps $0x10,(%rax,%r14,4),%xmm1 - DB 102,66,15,58,33,12,184,32 ; insertps $0x20,(%rax,%r15,4),%xmm1 - DB 102,66,15,58,33,12,160,48 ; insertps $0x30,(%rax,%r12,4),%xmm1 - DB 15,92,200 ; subps %xmm0,%xmm1 - DB 65,15,89,206 ; mulps %xmm14,%xmm1 - DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,92,236 ; subps %xmm4,%xmm5 - DB 65,15,89,238 ; mulps %xmm14,%xmm5 - DB 15,88,236 ; addps %xmm4,%xmm5 - DB 15,92,222 ; subps %xmm6,%xmm3 - DB 65,15,89,222 ; mulps %xmm14,%xmm3 - DB 15,88,222 ; addps %xmm6,%xmm3 - DB 243,68,15,91,218 ; cvttps2dq %xmm2,%xmm11 - DB 102,68,15,56,64,92,36,32 ; pmulld 0x20(%rsp),%xmm11 - DB 102,65,15,111,227 ; movdqa %xmm11,%xmm4 - DB 102,65,15,254,226 ; paddd %xmm10,%xmm4 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,65,15,254,196 ; paddd %xmm12,%xmm0 - DB 102,15,56,64,199 ; pmulld %xmm7,%xmm0 - DB 102,68,15,111,215 ; movdqa %xmm7,%xmm10 - DB 102,15,126,193 ; movd %xmm0,%ecx - DB 102,15,58,22,197,1 ; pextrd $0x1,%xmm0,%ebp - DB 102,65,15,58,22,192,2 ; pextrd $0x2,%xmm0,%r8d - DB 102,15,58,22,195,3 ; pextrd $0x3,%xmm0,%ebx - DB 243,15,16,52,136 ; movss (%rax,%rcx,4),%xmm6 - DB 102,15,58,33,52,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm6 - DB 102,66,15,58,33,52,128,32 ; insertps $0x20,(%rax,%r8,4),%xmm6 - DB 243,15,16,20,152 ; movss (%rax,%rbx,4),%xmm2 - DB 102,15,58,33,242,48 ; insertps $0x30,%xmm2,%xmm6 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,65,15,254,215 ; paddd %xmm15,%xmm2 - DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx - DB 102,72,15,126,213 ; movq %xmm2,%rbp - DB 137,235 ; mov %ebp,%ebx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d DB 243,15,16,20,152 ; movss (%rax,%rbx,4),%xmm2 + DB 102,15,58,33,250,48 ; insertps $0x30,%xmm2,%xmm7 + DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 + DB 68,15,92,216 ; subps %xmm0,%xmm11 + DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 + DB 102,66,15,58,33,20,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm2 + DB 102,66,15,58,33,20,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm2 + DB 102,66,15,58,33,20,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm2 + DB 243,66,15,16,4,168 ; movss (%rax,%r13,4),%xmm0 + DB 102,66,15,58,33,4,176,16 ; insertps $0x10,(%rax,%r14,4),%xmm0 + DB 102,66,15,58,33,4,184,32 ; insertps $0x20,(%rax,%r15,4),%xmm0 + DB 102,66,15,58,33,4,160,48 ; insertps $0x30,(%rax,%r12,4),%xmm0 + DB 15,92,194 ; subps %xmm2,%xmm0 + DB 65,15,89,195 ; mulps %xmm11,%xmm0 + DB 15,88,194 ; addps %xmm2,%xmm0 + DB 15,92,217 ; subps %xmm1,%xmm3 + DB 65,15,89,219 ; mulps %xmm11,%xmm3 + DB 15,88,217 ; addps %xmm1,%xmm3 + DB 15,92,252 ; subps %xmm4,%xmm7 + DB 65,15,89,251 ; mulps %xmm11,%xmm7 + DB 15,88,252 ; addps %xmm4,%xmm7 + DB 102,65,15,111,226 ; movdqa %xmm10,%xmm4 + DB 102,15,254,229 ; paddd %xmm5,%xmm4 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,65,15,254,200 ; paddd %xmm8,%xmm1 + DB 102,65,15,56,64,204 ; pmulld %xmm12,%xmm1 + DB 102,15,126,201 ; movd %xmm1,%ecx + DB 102,15,58,22,205,1 ; pextrd $0x1,%xmm1,%ebp + DB 102,65,15,58,22,200,2 ; pextrd $0x2,%xmm1,%r8d + DB 102,15,58,22,203,3 ; pextrd $0x3,%xmm1,%ebx + DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 DB 102,15,58,33,20,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm2 - DB 243,66,15,16,60,128 ; movss (%rax,%r8,4),%xmm7 - DB 102,15,58,33,215,32 ; insertps $0x20,%xmm7,%xmm2 + DB 102,66,15,58,33,20,128,32 ; insertps $0x20,(%rax,%r8,4),%xmm2 + DB 243,15,16,52,152 ; movss (%rax,%rbx,4),%xmm6 + DB 102,15,58,33,214,48 ; insertps $0x30,%xmm6,%xmm2 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,65,15,250,247 ; psubd %xmm15,%xmm6 + DB 102,69,15,118,255 ; pcmpeqd %xmm15,%xmm15 + DB 102,72,15,58,22,241,1 ; pextrq $0x1,%xmm6,%rcx + DB 102,72,15,126,245 ; movq %xmm6,%rbp + DB 65,137,232 ; mov %ebp,%r8d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 137,203 ; mov %ecx,%ebx + DB 243,66,15,16,52,128 ; movss (%rax,%r8,4),%xmm6 + DB 102,15,58,33,52,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm6 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,15,58,33,245,32 ; insertps $0x20,%xmm5,%xmm6 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,60,136 ; movss (%rax,%rcx,4),%xmm7 - DB 102,15,58,33,215,48 ; insertps $0x30,%xmm7,%xmm2 - DB 102,65,15,254,193 ; paddd %xmm9,%xmm0 - DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx - DB 102,72,15,126,197 ; movq %xmm0,%rbp - DB 137,235 ; mov %ebp,%ebx + DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 + DB 102,15,58,33,245,48 ; insertps $0x30,%xmm5,%xmm6 + DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 + DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx + DB 102,72,15,126,205 ; movq %xmm1,%rbp + DB 65,137,232 ; mov %ebp,%r8d DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,4,152 ; movss (%rax,%rbx,4),%xmm0 - DB 102,15,58,33,4,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm0 - DB 243,66,15,16,60,128 ; movss (%rax,%r8,4),%xmm7 - DB 102,15,58,33,199,32 ; insertps $0x20,%xmm7,%xmm0 + DB 137,203 ; mov %ecx,%ebx + DB 243,66,15,16,12,128 ; movss (%rax,%r8,4),%xmm1 + DB 102,15,58,33,12,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm1 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,15,58,33,205,32 ; insertps $0x20,%xmm5,%xmm1 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,60,136 ; movss (%rax,%rcx,4),%xmm7 - DB 102,15,58,33,199,48 ; insertps $0x30,%xmm7,%xmm0 - DB 102,65,15,254,224 ; paddd %xmm8,%xmm4 - DB 102,65,15,56,64,226 ; pmulld %xmm10,%xmm4 - DB 102,15,111,252 ; movdqa %xmm4,%xmm7 - DB 102,65,15,254,255 ; paddd %xmm15,%xmm7 - DB 102,72,15,58,22,249,1 ; pextrq $0x1,%xmm7,%rcx - DB 102,72,15,126,253 ; movq %xmm7,%rbp - DB 137,235 ; mov %ebp,%ebx + DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 + DB 102,15,58,33,205,48 ; insertps $0x30,%xmm5,%xmm1 + DB 102,65,15,254,230 ; paddd %xmm14,%xmm4 + DB 102,65,15,56,64,228 ; pmulld %xmm12,%xmm4 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,65,15,250,239 ; psubd %xmm15,%xmm5 + DB 102,72,15,58,22,233,1 ; pextrq $0x1,%xmm5,%rcx + DB 102,72,15,126,237 ; movq %xmm5,%rbp + DB 65,137,232 ; mov %ebp,%r8d DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,68,15,16,20,152 ; movss (%rax,%rbx,4),%xmm10 - DB 102,68,15,58,33,20,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm10 - DB 243,66,15,16,60,128 ; movss (%rax,%r8,4),%xmm7 - DB 102,68,15,58,33,215,32 ; insertps $0x20,%xmm7,%xmm10 + DB 137,203 ; mov %ecx,%ebx + DB 243,70,15,16,36,128 ; movss (%rax,%r8,4),%xmm12 + DB 102,68,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm12 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,68,15,58,33,229,32 ; insertps $0x20,%xmm5,%xmm12 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,60,136 ; movss (%rax,%rcx,4),%xmm7 - DB 102,68,15,58,33,215,48 ; insertps $0x30,%xmm7,%xmm10 - DB 102,65,15,126,225 ; movd %xmm4,%r9d + DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 + DB 102,68,15,58,33,229,48 ; insertps $0x30,%xmm5,%xmm12 + DB 102,65,15,126,224 ; movd %xmm4,%r8d DB 102,65,15,58,22,227,1 ; pextrd $0x1,%xmm4,%r11d DB 102,65,15,58,22,226,2 ; pextrd $0x2,%xmm4,%r10d - DB 102,65,15,58,22,224,3 ; pextrd $0x3,%xmm4,%r8d - DB 102,65,15,111,249 ; movdqa %xmm9,%xmm7 - DB 102,15,254,231 ; paddd %xmm7,%xmm4 + DB 102,65,15,58,22,225,3 ; pextrd $0x3,%xmm4,%r9d + DB 102,65,15,254,229 ; paddd %xmm13,%xmm4 DB 102,72,15,58,22,225,1 ; pextrq $0x1,%xmm4,%rcx DB 102,72,15,126,227 ; movq %xmm4,%rbx - DB 137,221 ; mov %ebx,%ebp + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d - DB 243,68,15,16,12,168 ; movss (%rax,%rbp,4),%xmm9 - DB 102,68,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm9 - DB 243,66,15,16,36,176 ; movss (%rax,%r14,4),%xmm4 - DB 102,68,15,58,33,204,32 ; insertps $0x20,%xmm4,%xmm9 + DB 137,205 ; mov %ecx,%ebp + DB 243,70,15,16,60,176 ; movss (%rax,%r14,4),%xmm15 + DB 102,68,15,58,33,60,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm15 + DB 243,15,16,36,168 ; movss (%rax,%rbp,4),%xmm4 + DB 102,68,15,58,33,252,32 ; insertps $0x20,%xmm4,%xmm15 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 - DB 102,68,15,58,33,204,48 ; insertps $0x30,%xmm4,%xmm9 - DB 243,66,15,16,36,136 ; movss (%rax,%r9,4),%xmm4 + DB 102,68,15,58,33,252,48 ; insertps $0x30,%xmm4,%xmm15 + DB 243,66,15,16,36,128 ; movss (%rax,%r8,4),%xmm4 DB 102,66,15,58,33,36,152,16 ; insertps $0x10,(%rax,%r11,4),%xmm4 DB 102,66,15,58,33,36,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm4 - DB 102,66,15,58,33,36,128,48 ; insertps $0x30,(%rax,%r8,4),%xmm4 - DB 15,92,230 ; subps %xmm6,%xmm4 - DB 65,15,89,230 ; mulps %xmm14,%xmm4 - DB 15,88,230 ; addps %xmm6,%xmm4 - DB 68,15,92,210 ; subps %xmm2,%xmm10 - DB 69,15,89,214 ; mulps %xmm14,%xmm10 - DB 68,15,88,210 ; addps %xmm2,%xmm10 - DB 68,15,92,200 ; subps %xmm0,%xmm9 - DB 69,15,89,206 ; mulps %xmm14,%xmm9 - DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 15,91,68,36,64 ; cvtdq2ps 0x40(%rsp),%xmm0 - DB 15,40,20,36 ; movaps (%rsp),%xmm2 - DB 15,92,208 ; subps %xmm0,%xmm2 - DB 15,41,20,36 ; movaps %xmm2,(%rsp) - DB 15,92,225 ; subps %xmm1,%xmm4 - DB 15,89,226 ; mulps %xmm2,%xmm4 - DB 15,88,225 ; addps %xmm1,%xmm4 - DB 68,15,92,213 ; subps %xmm5,%xmm10 - DB 68,15,89,210 ; mulps %xmm2,%xmm10 - DB 68,15,88,213 ; addps %xmm5,%xmm10 - DB 68,15,92,203 ; subps %xmm3,%xmm9 - DB 68,15,89,202 ; mulps %xmm2,%xmm9 - DB 68,15,88,203 ; addps %xmm3,%xmm9 - DB 243,15,91,68,36,48 ; cvttps2dq 0x30(%rsp),%xmm0 - DB 102,68,15,254,232 ; paddd %xmm0,%xmm13 - DB 102,65,15,111,212 ; movdqa %xmm12,%xmm2 - DB 102,65,15,254,213 ; paddd %xmm13,%xmm2 - DB 102,15,111,45,192,75,1,0 ; movdqa 0x14bc0(%rip),%xmm5 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,56,64,213 ; pmulld %xmm5,%xmm2 - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,65,15,254,207 ; paddd %xmm15,%xmm1 + DB 102,66,15,58,33,36,136,48 ; insertps $0x30,(%rax,%r9,4),%xmm4 + DB 15,92,226 ; subps %xmm2,%xmm4 + DB 65,15,89,227 ; mulps %xmm11,%xmm4 + DB 15,88,226 ; addps %xmm2,%xmm4 + DB 68,15,92,230 ; subps %xmm6,%xmm12 + DB 69,15,89,227 ; mulps %xmm11,%xmm12 + DB 68,15,88,230 ; addps %xmm6,%xmm12 + DB 68,15,92,249 ; subps %xmm1,%xmm15 + DB 69,15,89,251 ; mulps %xmm11,%xmm15 + DB 68,15,88,249 ; addps %xmm1,%xmm15 + DB 15,91,76,36,32 ; cvtdq2ps 0x20(%rsp),%xmm1 + DB 68,15,40,108,36,16 ; movaps 0x10(%rsp),%xmm13 + DB 68,15,92,233 ; subps %xmm1,%xmm13 + DB 15,92,224 ; subps %xmm0,%xmm4 + DB 65,15,89,229 ; mulps %xmm13,%xmm4 + DB 15,88,224 ; addps %xmm0,%xmm4 + DB 68,15,92,227 ; subps %xmm3,%xmm12 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 68,15,88,227 ; addps %xmm3,%xmm12 + DB 68,15,92,255 ; subps %xmm7,%xmm15 + DB 69,15,89,253 ; mulps %xmm13,%xmm15 + DB 68,15,88,255 ; addps %xmm7,%xmm15 + DB 243,15,91,84,36,48 ; cvttps2dq 0x30(%rsp),%xmm2 + DB 102,68,15,254,202 ; paddd %xmm2,%xmm9 + DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 + DB 102,65,15,254,193 ; paddd %xmm9,%xmm0 + DB 102,15,111,45,253,70,1,0 ; movdqa 0x146fd(%rip),%xmm5 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,56,64,197 ; pmulld %xmm5,%xmm0 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,15,118,246 ; pcmpeqd %xmm6,%xmm6 + DB 102,15,250,206 ; psubd %xmm6,%xmm1 DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx DB 102,72,15,126,205 ; movq %xmm1,%rbp DB 137,235 ; mov %ebp,%ebx @@ -23477,115 +23640,117 @@ _sk_clut_3D_sse41 LABEL PROC DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 DB 102,15,58,33,203,48 ; insertps $0x30,%xmm3,%xmm1 - DB 102,65,15,126,208 ; movd %xmm2,%r8d - DB 102,65,15,58,22,209,1 ; pextrd $0x1,%xmm2,%r9d - DB 102,65,15,58,22,210,2 ; pextrd $0x2,%xmm2,%r10d - DB 102,65,15,58,22,211,3 ; pextrd $0x3,%xmm2,%r11d - DB 102,15,254,215 ; paddd %xmm7,%xmm2 - DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx - DB 102,72,15,126,211 ; movq %xmm2,%rbx + DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,65,15,58,22,193,1 ; pextrd $0x1,%xmm0,%r9d + DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d + DB 102,65,15,58,22,195,3 ; pextrd $0x3,%xmm0,%r11d + DB 102,15,111,61,14,64,1,0 ; movdqa 0x1400e(%rip),%xmm7 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,254,199 ; paddd %xmm7,%xmm0 + DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx + DB 102,72,15,126,195 ; movq %xmm0,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,20,168 ; movss (%rax,%rbp,4),%xmm2 - DB 102,15,58,33,20,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm2 + DB 243,15,16,4,168 ; movss (%rax,%rbp,4),%xmm0 + DB 102,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm0 DB 243,66,15,16,28,176 ; movss (%rax,%r14,4),%xmm3 - DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 + DB 102,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm0 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 - DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 - DB 102,65,15,111,240 ; movdqa %xmm8,%xmm6 - DB 102,68,15,254,238 ; paddd %xmm6,%xmm13 - DB 102,68,15,56,64,237 ; pmulld %xmm5,%xmm13 - DB 102,65,15,111,221 ; movdqa %xmm13,%xmm3 - DB 102,65,15,254,223 ; paddd %xmm15,%xmm3 + DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0 + DB 102,69,15,254,206 ; paddd %xmm14,%xmm9 + DB 102,68,15,56,64,205 ; pmulld %xmm5,%xmm9 + DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 + DB 102,15,250,222 ; psubd %xmm6,%xmm3 DB 102,72,15,58,22,217,1 ; pextrq $0x1,%xmm3,%rcx DB 102,72,15,126,219 ; movq %xmm3,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,68,15,16,4,168 ; movss (%rax,%rbp,4),%xmm8 - DB 102,68,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm8 + DB 243,15,16,52,168 ; movss (%rax,%rbp,4),%xmm6 + DB 102,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm6 DB 243,66,15,16,28,176 ; movss (%rax,%r14,4),%xmm3 - DB 102,68,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm8 + DB 102,15,58,33,243,32 ; insertps $0x20,%xmm3,%xmm6 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 - DB 102,68,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm8 - DB 102,69,15,126,237 ; movd %xmm13,%r13d - DB 102,69,15,58,22,239,1 ; pextrd $0x1,%xmm13,%r15d - DB 102,69,15,58,22,236,2 ; pextrd $0x2,%xmm13,%r12d - DB 102,69,15,58,22,238,3 ; pextrd $0x3,%xmm13,%r14d - DB 102,68,15,254,239 ; paddd %xmm7,%xmm13 - DB 102,76,15,58,22,235,1 ; pextrq $0x1,%xmm13,%rbx - DB 102,76,15,126,237 ; movq %xmm13,%rbp + DB 102,15,58,33,243,48 ; insertps $0x30,%xmm3,%xmm6 + DB 102,69,15,126,205 ; movd %xmm9,%r13d + DB 102,69,15,58,22,207,1 ; pextrd $0x1,%xmm9,%r15d + DB 102,69,15,58,22,204,2 ; pextrd $0x2,%xmm9,%r12d + DB 102,69,15,58,22,206,3 ; pextrd $0x3,%xmm9,%r14d + DB 102,68,15,254,207 ; paddd %xmm7,%xmm9 + DB 102,76,15,58,22,203,1 ; pextrq $0x1,%xmm9,%rbx + DB 102,76,15,126,205 ; movq %xmm9,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,68,15,16,44,136 ; movss (%rax,%rcx,4),%xmm13 - DB 102,68,15,58,33,44,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm13 + DB 243,15,16,60,136 ; movss (%rax,%rcx,4),%xmm7 + DB 102,15,58,33,60,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm7 DB 137,217 ; mov %ebx,%ecx DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 - DB 102,68,15,58,33,235,32 ; insertps $0x20,%xmm3,%xmm13 + DB 102,15,58,33,251,32 ; insertps $0x20,%xmm3,%xmm7 DB 72,193,235,32 ; shr $0x20,%rbx DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 - DB 102,68,15,58,33,235,48 ; insertps $0x30,%xmm3,%xmm13 - DB 243,66,15,16,28,128 ; movss (%rax,%r8,4),%xmm3 - DB 102,66,15,58,33,28,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm3 - DB 102,66,15,58,33,28,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm3 - DB 102,66,15,58,33,28,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm3 - DB 243,66,15,16,60,168 ; movss (%rax,%r13,4),%xmm7 - DB 102,66,15,58,33,60,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm7 - DB 102,66,15,58,33,60,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm7 - DB 102,66,15,58,33,60,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm7 - DB 15,92,251 ; subps %xmm3,%xmm7 - DB 65,15,89,254 ; mulps %xmm14,%xmm7 - DB 15,88,251 ; addps %xmm3,%xmm7 - DB 68,15,92,193 ; subps %xmm1,%xmm8 - DB 69,15,89,198 ; mulps %xmm14,%xmm8 - DB 68,15,88,193 ; addps %xmm1,%xmm8 - DB 68,15,92,234 ; subps %xmm2,%xmm13 - DB 69,15,89,238 ; mulps %xmm14,%xmm13 - DB 68,15,88,234 ; addps %xmm2,%xmm13 - DB 102,68,15,254,216 ; paddd %xmm0,%xmm11 - DB 102,69,15,254,227 ; paddd %xmm11,%xmm12 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 102,68,15,56,64,225 ; pmulld %xmm1,%xmm12 - DB 102,65,15,111,196 ; movdqa %xmm12,%xmm0 - DB 102,65,15,254,199 ; paddd %xmm15,%xmm0 + DB 102,15,58,33,251,48 ; insertps $0x30,%xmm3,%xmm7 + DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 + DB 102,66,15,58,33,44,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm5 + DB 102,66,15,58,33,44,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm5 + DB 102,66,15,58,33,44,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm5 + DB 243,66,15,16,28,168 ; movss (%rax,%r13,4),%xmm3 + DB 102,66,15,58,33,28,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm3 + DB 102,66,15,58,33,28,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm3 + DB 102,66,15,58,33,28,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm3 + DB 15,92,221 ; subps %xmm5,%xmm3 + DB 65,15,89,219 ; mulps %xmm11,%xmm3 + DB 15,88,221 ; addps %xmm5,%xmm3 + DB 15,92,241 ; subps %xmm1,%xmm6 + DB 65,15,89,243 ; mulps %xmm11,%xmm6 + DB 15,88,241 ; addps %xmm1,%xmm6 + DB 15,92,248 ; subps %xmm0,%xmm7 + DB 65,15,89,251 ; mulps %xmm11,%xmm7 + DB 15,88,248 ; addps %xmm0,%xmm7 + DB 102,68,15,254,210 ; paddd %xmm2,%xmm10 + DB 102,69,15,254,194 ; paddd %xmm10,%xmm8 + DB 102,15,111,13,57,69,1,0 ; movdqa 0x14539(%rip),%xmm1 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,68,15,56,64,193 ; pmulld %xmm1,%xmm8 + DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 + DB 102,15,118,237 ; pcmpeqd %xmm5,%xmm5 + DB 102,15,250,197 ; psubd %xmm5,%xmm0 DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,197 ; movq %xmm0,%rbp DB 137,235 ; mov %ebp,%ebx DB 72,193,237,32 ; shr $0x20,%rbp DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 - DB 102,15,58,33,28,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm3 + DB 243,68,15,16,12,152 ; movss (%rax,%rbx,4),%xmm9 + DB 102,68,15,58,33,12,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm9 DB 243,66,15,16,4,128 ; movss (%rax,%r8,4),%xmm0 - DB 102,15,58,33,216,32 ; insertps $0x20,%xmm0,%xmm3 + DB 102,68,15,58,33,200,32 ; insertps $0x20,%xmm0,%xmm9 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 - DB 102,15,58,33,216,48 ; insertps $0x30,%xmm0,%xmm3 - DB 102,69,15,126,224 ; movd %xmm12,%r8d - DB 102,69,15,58,22,225,1 ; pextrd $0x1,%xmm12,%r9d - DB 102,69,15,58,22,226,2 ; pextrd $0x2,%xmm12,%r10d - DB 102,69,15,58,22,227,3 ; pextrd $0x3,%xmm12,%r11d - DB 102,15,111,21,27,67,1,0 ; movdqa 0x1431b(%rip),%xmm2 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,68,15,254,226 ; paddd %xmm2,%xmm12 - DB 102,76,15,58,22,225,1 ; pextrq $0x1,%xmm12,%rcx - DB 102,76,15,126,227 ; movq %xmm12,%rbx + DB 102,68,15,58,33,200,48 ; insertps $0x30,%xmm0,%xmm9 + DB 102,69,15,126,192 ; movd %xmm8,%r8d + DB 102,69,15,58,22,193,1 ; pextrd $0x1,%xmm8,%r9d + DB 102,69,15,58,22,194,2 ; pextrd $0x2,%xmm8,%r10d + DB 102,69,15,58,22,195,3 ; pextrd $0x3,%xmm8,%r11d + DB 102,15,111,21,68,62,1,0 ; movdqa 0x13e44(%rip),%xmm2 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,68,15,254,194 ; paddd %xmm2,%xmm8 + DB 102,76,15,58,22,193,1 ; pextrq $0x1,%xmm8,%rcx + DB 102,76,15,126,195 ; movq %xmm8,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,44,168 ; movss (%rax,%rbp,4),%xmm5 - DB 102,15,58,33,44,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm5 + DB 243,68,15,16,4,168 ; movss (%rax,%rbp,4),%xmm8 + DB 102,68,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm8 DB 243,66,15,16,4,176 ; movss (%rax,%r14,4),%xmm0 - DB 102,15,58,33,232,32 ; insertps $0x20,%xmm0,%xmm5 + DB 102,68,15,58,33,192,32 ; insertps $0x20,%xmm0,%xmm8 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 - DB 102,15,58,33,232,48 ; insertps $0x30,%xmm0,%xmm5 - DB 102,68,15,254,222 ; paddd %xmm6,%xmm11 - DB 102,68,15,56,64,217 ; pmulld %xmm1,%xmm11 - DB 102,69,15,254,251 ; paddd %xmm11,%xmm15 - DB 102,76,15,58,22,249,1 ; pextrq $0x1,%xmm15,%rcx - DB 102,76,15,126,251 ; movq %xmm15,%rbx + DB 102,68,15,58,33,192,48 ; insertps $0x30,%xmm0,%xmm8 + DB 102,69,15,254,214 ; paddd %xmm14,%xmm10 + DB 102,68,15,56,64,209 ; pmulld %xmm1,%xmm10 + DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 + DB 102,15,250,197 ; psubd %xmm5,%xmm0 + DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx + DB 102,72,15,126,195 ; movq %xmm0,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d @@ -23596,13 +23761,13 @@ _sk_clut_3D_sse41 LABEL PROC DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 DB 102,15,58,33,200,48 ; insertps $0x30,%xmm0,%xmm1 - DB 102,69,15,126,221 ; movd %xmm11,%r13d - DB 102,69,15,58,22,223,1 ; pextrd $0x1,%xmm11,%r15d - DB 102,69,15,58,22,220,2 ; pextrd $0x2,%xmm11,%r12d - DB 102,69,15,58,22,222,3 ; pextrd $0x3,%xmm11,%r14d - DB 102,68,15,254,218 ; paddd %xmm2,%xmm11 - DB 102,76,15,58,22,219,1 ; pextrq $0x1,%xmm11,%rbx - DB 102,76,15,126,221 ; movq %xmm11,%rbp + DB 102,69,15,126,213 ; movd %xmm10,%r13d + DB 102,69,15,58,22,215,1 ; pextrd $0x1,%xmm10,%r15d + DB 102,69,15,58,22,212,2 ; pextrd $0x2,%xmm10,%r12d + DB 102,69,15,58,22,214,3 ; pextrd $0x3,%xmm10,%r14d + DB 102,68,15,254,210 ; paddd %xmm2,%xmm10 + DB 102,76,15,58,22,211,1 ; pextrq $0x1,%xmm10,%rbx + DB 102,76,15,126,213 ; movq %xmm10,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 @@ -23613,53 +23778,52 @@ _sk_clut_3D_sse41 LABEL PROC DB 72,193,235,32 ; shr $0x20,%rbx DB 243,15,16,4,152 ; movss (%rax,%rbx,4),%xmm0 DB 102,15,58,33,208,48 ; insertps $0x30,%xmm0,%xmm2 - DB 243,66,15,16,52,128 ; movss (%rax,%r8,4),%xmm6 - DB 102,66,15,58,33,52,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm6 - DB 102,66,15,58,33,52,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm6 - DB 102,66,15,58,33,52,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm6 + DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 + DB 102,66,15,58,33,44,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm5 + DB 102,66,15,58,33,44,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm5 + DB 102,66,15,58,33,44,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm5 DB 243,66,15,16,4,168 ; movss (%rax,%r13,4),%xmm0 DB 102,66,15,58,33,4,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm0 DB 102,66,15,58,33,4,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm0 DB 102,66,15,58,33,4,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm0 - DB 15,92,198 ; subps %xmm6,%xmm0 - DB 65,15,89,198 ; mulps %xmm14,%xmm0 - DB 15,88,198 ; addps %xmm6,%xmm0 - DB 15,92,203 ; subps %xmm3,%xmm1 - DB 65,15,89,206 ; mulps %xmm14,%xmm1 - DB 15,88,203 ; addps %xmm3,%xmm1 - DB 15,92,213 ; subps %xmm5,%xmm2 - DB 65,15,89,214 ; mulps %xmm14,%xmm2 - DB 15,88,213 ; addps %xmm5,%xmm2 - DB 15,92,199 ; subps %xmm7,%xmm0 - DB 15,40,28,36 ; movaps (%rsp),%xmm3 - DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,199 ; addps %xmm7,%xmm0 - DB 65,15,92,200 ; subps %xmm8,%xmm1 - DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 65,15,92,213 ; subps %xmm13,%xmm2 - DB 15,89,211 ; mulps %xmm3,%xmm2 - DB 65,15,88,213 ; addps %xmm13,%xmm2 - DB 15,91,92,36,80 ; cvtdq2ps 0x50(%rsp),%xmm3 - DB 15,40,108,36,96 ; movaps 0x60(%rsp),%xmm5 + DB 15,92,197 ; subps %xmm5,%xmm0 + DB 65,15,89,195 ; mulps %xmm11,%xmm0 + DB 15,88,197 ; addps %xmm5,%xmm0 + DB 65,15,92,201 ; subps %xmm9,%xmm1 + DB 65,15,89,203 ; mulps %xmm11,%xmm1 + DB 65,15,88,201 ; addps %xmm9,%xmm1 + DB 65,15,92,208 ; subps %xmm8,%xmm2 + DB 65,15,89,211 ; mulps %xmm11,%xmm2 + DB 65,15,88,208 ; addps %xmm8,%xmm2 + DB 15,92,195 ; subps %xmm3,%xmm0 + DB 65,15,89,197 ; mulps %xmm13,%xmm0 + DB 15,88,195 ; addps %xmm3,%xmm0 + DB 15,92,206 ; subps %xmm6,%xmm1 + DB 65,15,89,205 ; mulps %xmm13,%xmm1 + DB 15,88,206 ; addps %xmm6,%xmm1 + DB 15,92,215 ; subps %xmm7,%xmm2 + DB 65,15,89,213 ; mulps %xmm13,%xmm2 + DB 15,88,215 ; addps %xmm7,%xmm2 + DB 15,91,92,36,64 ; cvtdq2ps 0x40(%rsp),%xmm3 + DB 15,40,108,36,80 ; movaps 0x50(%rsp),%xmm5 DB 15,92,235 ; subps %xmm3,%xmm5 DB 15,92,196 ; subps %xmm4,%xmm0 DB 15,89,197 ; mulps %xmm5,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 - DB 65,15,92,202 ; subps %xmm10,%xmm1 + DB 65,15,92,204 ; subps %xmm12,%xmm1 DB 15,89,205 ; mulps %xmm5,%xmm1 - DB 65,15,88,202 ; addps %xmm10,%xmm1 - DB 65,15,92,209 ; subps %xmm9,%xmm2 + DB 65,15,88,204 ; addps %xmm12,%xmm1 + DB 65,15,92,215 ; subps %xmm15,%xmm2 DB 15,89,213 ; mulps %xmm5,%xmm2 - DB 65,15,88,209 ; addps %xmm9,%xmm2 + DB 65,15,88,215 ; addps %xmm15,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 72,139,76,36,24 ; mov 0x18(%rsp),%rcx - DB 15,40,92,36,112 ; movaps 0x70(%rsp),%xmm3 - DB 15,40,164,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm4 - DB 15,40,172,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm5 - DB 15,40,180,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm6 - DB 15,40,188,36,176,0,0,0 ; movaps 0xb0(%rsp),%xmm7 - DB 72,129,196,200,0,0,0 ; add $0xc8,%rsp + DB 72,139,76,36,8 ; mov 0x8(%rsp),%rcx + DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 + DB 15,40,100,36,112 ; movaps 0x70(%rsp),%xmm4 + DB 15,40,172,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm5 + DB 15,40,180,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm6 + DB 15,40,188,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm7 + DB 72,129,196,184,0,0,0 ; add $0xb8,%rsp DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,93 ; pop %r13 @@ -23676,511 +23840,521 @@ _sk_clut_4D_sse41 LABEL PROC DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 72,129,236,24,1,0,0 ; sub $0x118,%rsp - DB 15,41,188,36,0,1,0,0 ; movaps %xmm7,0x100(%rsp) - DB 15,41,180,36,240,0,0,0 ; movaps %xmm6,0xf0(%rsp) - DB 15,41,172,36,224,0,0,0 ; movaps %xmm5,0xe0(%rsp) - DB 15,41,164,36,208,0,0,0 ; movaps %xmm4,0xd0(%rsp) + DB 72,129,236,8,1,0,0 ; sub $0x108,%rsp + DB 15,41,188,36,240,0,0,0 ; movaps %xmm7,0xf0(%rsp) + DB 15,41,180,36,224,0,0,0 ; movaps %xmm6,0xe0(%rsp) + DB 15,41,172,36,208,0,0,0 ; movaps %xmm5,0xd0(%rsp) + DB 15,41,164,36,192,0,0,0 ; movaps %xmm4,0xc0(%rsp) + DB 68,15,40,193 ; movaps %xmm1,%xmm8 DB 72,137,76,36,120 ; mov %rcx,0x78(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,20 ; mov 0x14(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d - DB 102,65,15,110,225 ; movd %r9d,%xmm4 - DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,227 ; mulps %xmm3,%xmm4 - DB 243,15,91,236 ; cvttps2dq %xmm4,%xmm5 - DB 102,15,127,172,36,128,0,0,0 ; movdqa %xmm5,0x80(%rsp) - DB 68,15,40,212 ; movaps %xmm4,%xmm10 - DB 68,15,41,148,36,192,0,0,0 ; movaps %xmm10,0xc0(%rsp) - DB 102,65,15,110,224 ; movd %r8d,%xmm4 - DB 102,15,112,244,0 ; pshufd $0x0,%xmm4,%xmm6 - DB 102,15,127,180,36,160,0,0,0 ; movdqa %xmm6,0xa0(%rsp) + DB 102,65,15,110,201 ; movd %r9d,%xmm1 + DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 + DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 + DB 15,89,203 ; mulps %xmm3,%xmm1 + DB 15,41,140,36,176,0,0,0 ; movaps %xmm1,0xb0(%rsp) + DB 243,15,91,249 ; cvttps2dq %xmm1,%xmm7 + DB 15,40,29,123,66,1,0 ; movaps 0x1427b(%rip),%xmm3 # 2da50 <_sk_overlay_sse2_8bit+0x15a1> + DB 15,88,203 ; addps %xmm3,%xmm1 + DB 15,41,76,36,32 ; movaps %xmm1,0x20(%rsp) + DB 102,65,15,110,232 ; movd %r8d,%xmm5 + DB 102,68,15,112,237,0 ; pshufd $0x0,%xmm5,%xmm13 DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d - DB 102,65,15,110,217 ; movd %r9d,%xmm3 - DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 - DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,218 ; mulps %xmm2,%xmm3 - DB 243,68,15,91,227 ; cvttps2dq %xmm3,%xmm12 - DB 102,68,15,127,100,36,64 ; movdqa %xmm12,0x40(%rsp) - DB 15,40,251 ; movaps %xmm3,%xmm7 - DB 15,41,124,36,96 ; movaps %xmm7,0x60(%rsp) - DB 102,68,15,56,64,230 ; pmulld %xmm6,%xmm12 - DB 102,65,15,111,220 ; movdqa %xmm12,%xmm3 - DB 102,15,254,221 ; paddd %xmm5,%xmm3 - DB 102,65,15,110,208 ; movd %r8d,%xmm2 - DB 102,15,56,64,212 ; pmulld %xmm4,%xmm2 - DB 102,15,112,234,0 ; pshufd $0x0,%xmm2,%xmm5 - DB 102,15,127,108,36,16 ; movdqa %xmm5,0x10(%rsp) + DB 102,65,15,110,201 ; movd %r9d,%xmm1 + DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 + DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 + DB 15,89,202 ; mulps %xmm2,%xmm1 + DB 15,41,76,36,96 ; movaps %xmm1,0x60(%rsp) + DB 243,15,91,209 ; cvttps2dq %xmm1,%xmm2 + DB 102,15,127,148,36,160,0,0,0 ; movdqa %xmm2,0xa0(%rsp) + DB 15,88,203 ; addps %xmm3,%xmm1 + DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 + DB 102,65,15,56,64,205 ; pmulld %xmm13,%xmm1 + DB 102,15,127,76,36,16 ; movdqa %xmm1,0x10(%rsp) + DB 102,68,15,56,64,234 ; pmulld %xmm2,%xmm13 + DB 102,65,15,110,240 ; movd %r8d,%xmm6 + DB 102,15,56,64,245 ; pmulld %xmm5,%xmm6 DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d - DB 102,65,15,110,225 ; movd %r9d,%xmm4 - DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 68,15,91,244 ; cvtdq2ps %xmm4,%xmm14 - DB 68,15,89,241 ; mulps %xmm1,%xmm14 - DB 102,65,15,110,224 ; movd %r8d,%xmm4 - DB 102,15,56,64,226 ; pmulld %xmm2,%xmm4 - DB 68,139,64,8 ; mov 0x8(%rax),%r8d - DB 65,255,200 ; dec %r8d - DB 102,65,15,110,200 ; movd %r8d,%xmm1 + DB 102,65,15,110,201 ; movd %r9d,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 - DB 68,15,91,249 ; cvtdq2ps %xmm1,%xmm15 - DB 68,15,89,248 ; mulps %xmm0,%xmm15 - DB 243,65,15,91,198 ; cvttps2dq %xmm14,%xmm0 + DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2 + DB 65,15,89,208 ; mulps %xmm8,%xmm2 + DB 68,139,72,8 ; mov 0x8(%rax),%r9d + DB 65,255,201 ; dec %r9d + DB 102,65,15,110,201 ; movd %r9d,%xmm1 + DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 + DB 68,15,91,225 ; cvtdq2ps %xmm1,%xmm12 + DB 68,15,89,224 ; mulps %xmm0,%xmm12 + DB 102,65,15,111,229 ; movdqa %xmm13,%xmm4 + DB 102,15,127,188,36,144,0,0,0 ; movdqa %xmm7,0x90(%rsp) + DB 102,15,254,231 ; paddd %xmm7,%xmm4 + DB 102,65,15,110,232 ; movd %r8d,%xmm5 + DB 102,15,56,64,238 ; pmulld %xmm6,%xmm5 + DB 102,68,15,112,246,0 ; pshufd $0x0,%xmm6,%xmm14 + DB 15,40,194 ; movaps %xmm2,%xmm0 + DB 15,88,195 ; addps %xmm3,%xmm0 + DB 243,68,15,91,192 ; cvttps2dq %xmm0,%xmm8 + DB 102,69,15,56,64,198 ; pmulld %xmm14,%xmm8 + DB 102,68,15,127,68,36,80 ; movdqa %xmm8,0x50(%rsp) + DB 243,15,91,194 ; cvttps2dq %xmm2,%xmm0 DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) - DB 102,15,56,64,197 ; pmulld %xmm5,%xmm0 - DB 102,15,127,68,36,80 ; movdqa %xmm0,0x50(%rsp) - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,254,203 ; paddd %xmm3,%xmm1 - DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 243,69,15,91,207 ; cvttps2dq %xmm15,%xmm9 - DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 - DB 102,68,15,56,64,236 ; pmulld %xmm4,%xmm13 - DB 102,65,15,111,197 ; movdqa %xmm13,%xmm0 - DB 102,68,15,127,108,36,48 ; movdqa %xmm13,0x30(%rsp) - DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,56,64,5,104,70,1,0 ; pmulld 0x14668(%rip),%xmm0 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,68,15,111,29,207,63,1,0 ; movdqa 0x13fcf(%rip),%xmm11 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,73,15,58,22,208,1 ; pextrq $0x1,%xmm2,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 + DB 102,68,15,56,64,240 ; pmulld %xmm0,%xmm14 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,15,254,204 ; paddd %xmm4,%xmm1 + DB 102,68,15,112,253,0 ; pshufd $0x0,%xmm5,%xmm15 + DB 65,15,88,220 ; addps %xmm12,%xmm3 + DB 243,68,15,91,203 ; cvttps2dq %xmm3,%xmm9 + DB 102,69,15,56,64,207 ; pmulld %xmm15,%xmm9 + DB 243,65,15,91,236 ; cvttps2dq %xmm12,%xmm5 + DB 102,68,15,56,64,253 ; pmulld %xmm5,%xmm15 + DB 102,65,15,111,223 ; movdqa %xmm15,%xmm3 + DB 102,15,254,217 ; paddd %xmm1,%xmm3 + DB 102,68,15,111,29,117,65,1,0 ; movdqa 0x14175(%rip),%xmm11 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,65,15,56,64,219 ; pmulld %xmm11,%xmm3 + DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,15,250,240 ; psubd %xmm0,%xmm6 + DB 102,73,15,58,22,240,1 ; pextrq $0x1,%xmm6,%r8 + DB 102,73,15,126,241 ; movq %xmm6,%r9 DB 72,139,0 ; mov (%rax),%rax DB 69,137,202 ; mov %r9d,%r10d DB 73,193,233,32 ; shr $0x20,%r9 DB 69,137,195 ; mov %r8d,%r11d - DB 243,66,15,16,20,144 ; movss (%rax,%r10,4),%xmm2 - DB 102,66,15,58,33,20,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm2 - DB 243,66,15,16,44,152 ; movss (%rax,%r11,4),%xmm5 - DB 102,15,58,33,213,32 ; insertps $0x20,%xmm5,%xmm2 + DB 243,66,15,16,60,144 ; movss (%rax,%r10,4),%xmm7 + DB 102,66,15,58,33,60,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm7 + DB 243,66,15,16,52,152 ; movss (%rax,%r11,4),%xmm6 + DB 102,15,58,33,254,32 ; insertps $0x20,%xmm6,%xmm7 DB 73,193,232,32 ; shr $0x20,%r8 - DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 - DB 102,15,58,33,213,48 ; insertps $0x30,%xmm5,%xmm2 - DB 102,65,15,126,192 ; movd %xmm0,%r8d - DB 102,65,15,58,22,193,1 ; pextrd $0x1,%xmm0,%r9d - DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d - DB 102,65,15,58,22,195,3 ; pextrd $0x3,%xmm0,%r11d - DB 102,68,15,111,5,128,63,1,0 ; movdqa 0x13f80(%rip),%xmm8 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,65,15,254,192 ; paddd %xmm8,%xmm0 - DB 102,73,15,58,22,198,1 ; pextrq $0x1,%xmm0,%r14 - DB 102,72,15,126,195 ; movq %xmm0,%rbx + DB 243,66,15,16,52,128 ; movss (%rax,%r8,4),%xmm6 + DB 102,15,58,33,254,48 ; insertps $0x30,%xmm6,%xmm7 + DB 102,65,15,126,216 ; movd %xmm3,%r8d + DB 102,65,15,58,22,217,1 ; pextrd $0x1,%xmm3,%r9d + DB 102,65,15,58,22,218,2 ; pextrd $0x2,%xmm3,%r10d + DB 102,65,15,58,22,219,3 ; pextrd $0x3,%xmm3,%r11d + DB 102,68,15,111,21,125,58,1,0 ; movdqa 0x13a7d(%rip),%xmm10 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,218 ; paddd %xmm10,%xmm3 + DB 102,73,15,58,22,222,1 ; pextrq $0x1,%xmm3,%r14 + DB 102,72,15,126,219 ; movq %xmm3,%rbx DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx DB 69,137,244 ; mov %r14d,%r12d - DB 243,66,15,16,44,184 ; movss (%rax,%r15,4),%xmm5 - DB 102,15,58,33,44,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm5 - DB 243,66,15,16,4,160 ; movss (%rax,%r12,4),%xmm0 - DB 102,15,58,33,232,32 ; insertps $0x20,%xmm0,%xmm5 + DB 243,66,15,16,4,184 ; movss (%rax,%r15,4),%xmm0 + DB 102,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm0 + DB 243,66,15,16,28,160 ; movss (%rax,%r12,4),%xmm3 + DB 102,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm0 DB 73,193,238,32 ; shr $0x20,%r14 - DB 243,66,15,16,4,176 ; movss (%rax,%r14,4),%xmm0 - DB 102,15,58,33,232,48 ; insertps $0x30,%xmm0,%xmm5 - DB 15,40,53,149,69,1,0 ; movaps 0x14595(%rip),%xmm6 # 2db50 <_sk_overlay_sse2_8bit+0x15f8> - DB 68,15,88,214 ; addps %xmm6,%xmm10 - DB 68,15,41,148,36,176,0,0,0 ; movaps %xmm10,0xb0(%rsp) - DB 15,88,254 ; addps %xmm6,%xmm7 - DB 15,41,188,36,144,0,0,0 ; movaps %xmm7,0x90(%rsp) - DB 65,15,40,198 ; movaps %xmm14,%xmm0 - DB 15,88,198 ; addps %xmm6,%xmm0 - DB 65,15,88,247 ; addps %xmm15,%xmm6 - DB 243,15,91,246 ; cvttps2dq %xmm6,%xmm6 - DB 102,15,56,64,244 ; pmulld %xmm4,%xmm6 - DB 102,15,254,206 ; paddd %xmm6,%xmm1 - DB 102,68,15,111,214 ; movdqa %xmm6,%xmm10 - DB 102,68,15,127,84,36,32 ; movdqa %xmm10,0x20(%rsp) - DB 102,15,56,64,13,96,69,1,0 ; pmulld 0x14560(%rip),%xmm1 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,65,15,254,227 ; paddd %xmm11,%xmm4 - DB 102,73,15,58,22,230,1 ; pextrq $0x1,%xmm4,%r14 - DB 102,72,15,126,227 ; movq %xmm4,%rbx + DB 243,66,15,16,28,176 ; movss (%rax,%r14,4),%xmm3 + DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0 + DB 102,68,15,127,76,36,48 ; movdqa %xmm9,0x30(%rsp) + DB 102,65,15,254,201 ; paddd %xmm9,%xmm1 + DB 102,65,15,56,64,203 ; pmulld %xmm11,%xmm1 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,250,29,171,64,1,0 ; psubd 0x140ab(%rip),%xmm3 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,73,15,58,22,222,1 ; pextrq $0x1,%xmm3,%r14 + DB 102,72,15,126,219 ; movq %xmm3,%rbx DB 65,137,223 ; mov %ebx,%r15d DB 72,193,235,32 ; shr $0x20,%rbx DB 69,137,244 ; mov %r14d,%r12d - DB 243,66,15,16,52,184 ; movss (%rax,%r15,4),%xmm6 - DB 102,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm6 - DB 243,66,15,16,36,160 ; movss (%rax,%r12,4),%xmm4 - DB 102,15,58,33,244,32 ; insertps $0x20,%xmm4,%xmm6 + DB 243,66,15,16,28,184 ; movss (%rax,%r15,4),%xmm3 + DB 102,15,58,33,28,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm3 + DB 243,66,15,16,52,160 ; movss (%rax,%r12,4),%xmm6 + DB 102,15,58,33,222,32 ; insertps $0x20,%xmm6,%xmm3 DB 73,193,238,32 ; shr $0x20,%r14 - DB 243,66,15,16,36,176 ; movss (%rax,%r14,4),%xmm4 - DB 102,15,58,33,244,48 ; insertps $0x30,%xmm4,%xmm6 + DB 243,66,15,16,52,176 ; movss (%rax,%r14,4),%xmm6 + DB 102,15,58,33,222,48 ; insertps $0x30,%xmm6,%xmm3 DB 102,65,15,126,205 ; movd %xmm1,%r13d DB 102,65,15,58,22,206,1 ; pextrd $0x1,%xmm1,%r14d DB 102,65,15,58,22,207,2 ; pextrd $0x2,%xmm1,%r15d DB 102,65,15,58,22,204,3 ; pextrd $0x3,%xmm1,%r12d - DB 102,65,15,254,200 ; paddd %xmm8,%xmm1 - DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 + DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 DB 102,72,15,58,22,203,1 ; pextrq $0x1,%xmm1,%rbx DB 102,72,15,126,205 ; movq %xmm1,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 - DB 102,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm4 - DB 137,217 ; mov %ebx,%ecx DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 - DB 102,15,58,33,225,32 ; insertps $0x20,%xmm1,%xmm4 + DB 102,15,58,33,12,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm1 + DB 137,217 ; mov %ebx,%ecx + DB 243,15,16,52,136 ; movss (%rax,%rcx,4),%xmm6 + DB 102,15,58,33,206,32 ; insertps $0x20,%xmm6,%xmm1 DB 72,193,235,32 ; shr $0x20,%rbx - DB 243,15,16,12,152 ; movss (%rax,%rbx,4),%xmm1 - DB 102,15,58,33,225,48 ; insertps $0x30,%xmm1,%xmm4 - DB 65,15,91,201 ; cvtdq2ps %xmm9,%xmm1 - DB 68,15,92,249 ; subps %xmm1,%xmm15 - DB 243,66,15,16,12,128 ; movss (%rax,%r8,4),%xmm1 - DB 102,66,15,58,33,12,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm1 - DB 102,66,15,58,33,12,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm1 - DB 102,66,15,58,33,12,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm1 - DB 243,66,15,16,60,168 ; movss (%rax,%r13,4),%xmm7 - DB 102,66,15,58,33,60,176,16 ; insertps $0x10,(%rax,%r14,4),%xmm7 - DB 102,66,15,58,33,60,184,32 ; insertps $0x20,(%rax,%r15,4),%xmm7 - DB 102,66,15,58,33,60,160,48 ; insertps $0x30,(%rax,%r12,4),%xmm7 - DB 15,92,249 ; subps %xmm1,%xmm7 - DB 65,15,89,255 ; mulps %xmm15,%xmm7 - DB 15,88,249 ; addps %xmm1,%xmm7 - DB 15,92,242 ; subps %xmm2,%xmm6 - DB 65,15,89,247 ; mulps %xmm15,%xmm6 - DB 15,88,242 ; addps %xmm2,%xmm6 - DB 15,92,229 ; subps %xmm5,%xmm4 - DB 65,15,89,231 ; mulps %xmm15,%xmm4 - DB 15,88,229 ; addps %xmm5,%xmm4 - DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 - DB 102,15,56,64,68,36,16 ; pmulld 0x10(%rsp),%xmm0 - DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) - DB 102,15,254,216 ; paddd %xmm0,%xmm3 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 - DB 102,15,111,5,50,68,1,0 ; movdqa 0x14432(%rip),%xmm0 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,56,64,200 ; pmulld %xmm0,%xmm1 - DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,68,15,111,5,139,61,1,0 ; movdqa 0x13d8b(%rip),%xmm8 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,192 ; paddd %xmm8,%xmm0 - DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx - DB 102,72,15,126,197 ; movq %xmm0,%rbp - DB 137,235 ; mov %ebp,%ebx + DB 243,15,16,52,152 ; movss (%rax,%rbx,4),%xmm6 + DB 102,15,58,33,206,48 ; insertps $0x30,%xmm6,%xmm1 + DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 + DB 68,15,92,229 ; subps %xmm5,%xmm12 + DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 + DB 102,66,15,58,33,44,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm5 + DB 102,66,15,58,33,44,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm5 + DB 102,66,15,58,33,44,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm5 + DB 243,66,15,16,52,168 ; movss (%rax,%r13,4),%xmm6 + DB 102,66,15,58,33,52,176,16 ; insertps $0x10,(%rax,%r14,4),%xmm6 + DB 102,66,15,58,33,52,184,32 ; insertps $0x20,(%rax,%r15,4),%xmm6 + DB 102,66,15,58,33,52,160,48 ; insertps $0x30,(%rax,%r12,4),%xmm6 + DB 15,92,245 ; subps %xmm5,%xmm6 + DB 65,15,89,244 ; mulps %xmm12,%xmm6 + DB 15,88,245 ; addps %xmm5,%xmm6 + DB 15,92,223 ; subps %xmm7,%xmm3 + DB 65,15,89,220 ; mulps %xmm12,%xmm3 + DB 15,88,223 ; addps %xmm7,%xmm3 + DB 15,92,200 ; subps %xmm0,%xmm1 + DB 65,15,89,204 ; mulps %xmm12,%xmm1 + DB 15,88,200 ; addps %xmm0,%xmm1 + DB 102,65,15,254,224 ; paddd %xmm8,%xmm4 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,68,15,127,124,36,64 ; movdqa %xmm15,0x40(%rsp) + DB 102,65,15,254,199 ; paddd %xmm15,%xmm0 + DB 102,65,15,56,64,195 ; pmulld %xmm11,%xmm0 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,250,45,139,63,1,0 ; psubd 0x13f8b(%rip),%xmm5 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 + DB 102,72,15,58,22,233,1 ; pextrq $0x1,%xmm5,%rcx + DB 102,72,15,126,237 ; movq %xmm5,%rbp + DB 65,137,232 ; mov %ebp,%r8d DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,4,152 ; movss (%rax,%rbx,4),%xmm0 - DB 102,15,58,33,4,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm0 - DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 - DB 102,15,58,33,194,32 ; insertps $0x20,%xmm2,%xmm0 + DB 137,203 ; mov %ecx,%ebx + DB 243,66,15,16,60,128 ; movss (%rax,%r8,4),%xmm7 + DB 102,15,58,33,60,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm7 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,15,58,33,253,32 ; insertps $0x20,%xmm5,%xmm7 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,15,58,33,194,48 ; insertps $0x30,%xmm2,%xmm0 - DB 102,65,15,126,200 ; movd %xmm1,%r8d - DB 102,65,15,58,22,201,1 ; pextrd $0x1,%xmm1,%r9d - DB 102,65,15,58,22,202,2 ; pextrd $0x2,%xmm1,%r10d - DB 102,65,15,58,22,203,3 ; pextrd $0x3,%xmm1,%r11d - DB 102,65,15,254,203 ; paddd %xmm11,%xmm1 - DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx - DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 137,221 ; mov %ebx,%ebp + DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 + DB 102,15,58,33,253,48 ; insertps $0x30,%xmm5,%xmm7 + DB 102,65,15,126,193 ; movd %xmm0,%r9d + DB 102,65,15,58,22,195,1 ; pextrd $0x1,%xmm0,%r11d + DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d + DB 102,65,15,58,22,192,3 ; pextrd $0x3,%xmm0,%r8d + DB 102,65,15,254,194 ; paddd %xmm10,%xmm0 + DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx + DB 102,72,15,126,195 ; movq %xmm0,%rbx + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,12,168 ; movss (%rax,%rbp,4),%xmm1 - DB 102,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm1 - DB 243,66,15,16,20,176 ; movss (%rax,%r14,4),%xmm2 - DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1 + DB 137,205 ; mov %ecx,%ebp + DB 243,66,15,16,4,176 ; movss (%rax,%r14,4),%xmm0 + DB 102,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm0 + DB 243,15,16,44,168 ; movss (%rax,%rbp,4),%xmm5 + DB 102,15,58,33,197,32 ; insertps $0x20,%xmm5,%xmm0 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 - DB 102,65,15,254,218 ; paddd %xmm10,%xmm3 - DB 102,65,15,56,64,217 ; pmulld %xmm9,%xmm3 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,65,15,254,208 ; paddd %xmm8,%xmm2 - DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx - DB 102,72,15,126,211 ; movq %xmm2,%rbx + DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 + DB 102,15,58,33,197,48 ; insertps $0x30,%xmm5,%xmm0 + DB 102,65,15,254,225 ; paddd %xmm9,%xmm4 + DB 102,65,15,56,64,227 ; pmulld %xmm11,%xmm4 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,65,15,250,232 ; psubd %xmm8,%xmm5 + DB 102,72,15,58,22,233,1 ; pextrq $0x1,%xmm5,%rcx + DB 102,72,15,126,235 ; movq %xmm5,%rbx + DB 65,137,222 ; mov %ebx,%r14d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 137,205 ; mov %ecx,%ebp + DB 243,70,15,16,4,176 ; movss (%rax,%r14,4),%xmm8 + DB 102,68,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm8 + DB 243,15,16,44,168 ; movss (%rax,%rbp,4),%xmm5 + DB 102,68,15,58,33,197,32 ; insertps $0x20,%xmm5,%xmm8 + DB 72,193,233,32 ; shr $0x20,%rcx + DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 + DB 102,68,15,58,33,197,48 ; insertps $0x30,%xmm5,%xmm8 + DB 102,65,15,126,231 ; movd %xmm4,%r15d + DB 102,65,15,58,22,228,1 ; pextrd $0x1,%xmm4,%r12d + DB 102,65,15,58,22,229,2 ; pextrd $0x2,%xmm4,%r13d + DB 102,65,15,58,22,230,3 ; pextrd $0x3,%xmm4,%r14d + DB 102,65,15,254,226 ; paddd %xmm10,%xmm4 + DB 102,72,15,58,22,225,1 ; pextrq $0x1,%xmm4,%rcx + DB 102,72,15,126,227 ; movq %xmm4,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d - DB 243,68,15,16,20,168 ; movss (%rax,%rbp,4),%xmm10 - DB 102,68,15,58,33,20,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm10 - DB 243,66,15,16,20,176 ; movss (%rax,%r14,4),%xmm2 - DB 102,68,15,58,33,210,32 ; insertps $0x20,%xmm2,%xmm10 + DB 243,68,15,16,12,168 ; movss (%rax,%rbp,4),%xmm9 + DB 102,68,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm9 + DB 137,203 ; mov %ecx,%ebx + DB 243,15,16,36,152 ; movss (%rax,%rbx,4),%xmm4 + DB 102,68,15,58,33,204,32 ; insertps $0x20,%xmm4,%xmm9 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,68,15,58,33,210,48 ; insertps $0x30,%xmm2,%xmm10 - DB 102,65,15,126,221 ; movd %xmm3,%r13d - DB 102,65,15,58,22,223,1 ; pextrd $0x1,%xmm3,%r15d - DB 102,65,15,58,22,220,2 ; pextrd $0x2,%xmm3,%r12d - DB 102,65,15,58,22,222,3 ; pextrd $0x3,%xmm3,%r14d - DB 102,65,15,254,219 ; paddd %xmm11,%xmm3 - DB 102,72,15,58,22,219,1 ; pextrq $0x1,%xmm3,%rbx + DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 + DB 102,68,15,58,33,204,48 ; insertps $0x30,%xmm4,%xmm9 + DB 243,66,15,16,36,136 ; movss (%rax,%r9,4),%xmm4 + DB 102,66,15,58,33,36,152,16 ; insertps $0x10,(%rax,%r11,4),%xmm4 + DB 102,66,15,58,33,36,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm4 + DB 102,66,15,58,33,36,128,48 ; insertps $0x30,(%rax,%r8,4),%xmm4 + DB 243,66,15,16,44,184 ; movss (%rax,%r15,4),%xmm5 + DB 102,66,15,58,33,44,160,16 ; insertps $0x10,(%rax,%r12,4),%xmm5 + DB 102,66,15,58,33,44,168,32 ; insertps $0x20,(%rax,%r13,4),%xmm5 + DB 102,66,15,58,33,44,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm5 + DB 15,92,236 ; subps %xmm4,%xmm5 + DB 65,15,89,236 ; mulps %xmm12,%xmm5 + DB 15,88,236 ; addps %xmm4,%xmm5 + DB 68,15,92,199 ; subps %xmm7,%xmm8 + DB 69,15,89,196 ; mulps %xmm12,%xmm8 + DB 68,15,88,199 ; addps %xmm7,%xmm8 + DB 68,15,92,200 ; subps %xmm0,%xmm9 + DB 69,15,89,204 ; mulps %xmm12,%xmm9 + DB 68,15,88,200 ; addps %xmm0,%xmm9 + DB 15,91,4,36 ; cvtdq2ps (%rsp),%xmm0 + DB 15,92,208 ; subps %xmm0,%xmm2 + DB 15,41,20,36 ; movaps %xmm2,(%rsp) + DB 15,92,238 ; subps %xmm6,%xmm5 + DB 15,89,234 ; mulps %xmm2,%xmm5 + DB 15,88,238 ; addps %xmm6,%xmm5 + DB 68,15,92,195 ; subps %xmm3,%xmm8 + DB 68,15,89,194 ; mulps %xmm2,%xmm8 + DB 68,15,88,195 ; addps %xmm3,%xmm8 + DB 68,15,92,201 ; subps %xmm1,%xmm9 + DB 68,15,89,202 ; mulps %xmm2,%xmm9 + DB 68,15,88,201 ; addps %xmm1,%xmm9 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,254,132,36,144,0,0,0 ; paddd 0x90(%rsp),%xmm0 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,65,15,254,206 ; paddd %xmm14,%xmm1 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,65,15,254,223 ; paddd %xmm15,%xmm3 + DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 + DB 102,15,56,64,218 ; pmulld %xmm2,%xmm3 + DB 102,15,126,217 ; movd %xmm3,%ecx + DB 102,15,58,22,221,1 ; pextrd $0x1,%xmm3,%ebp + DB 102,65,15,58,22,216,2 ; pextrd $0x2,%xmm3,%r8d + DB 102,15,58,22,219,3 ; pextrd $0x3,%xmm3,%ebx + DB 243,68,15,16,20,136 ; movss (%rax,%rcx,4),%xmm10 + DB 102,68,15,58,33,20,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm10 + DB 102,70,15,58,33,20,128,32 ; insertps $0x20,(%rax,%r8,4),%xmm10 + DB 243,15,16,36,152 ; movss (%rax,%rbx,4),%xmm4 + DB 102,68,15,58,33,212,48 ; insertps $0x30,%xmm4,%xmm10 + DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,15,118,246 ; pcmpeqd %xmm6,%xmm6 + DB 102,15,250,230 ; psubd %xmm6,%xmm4 + DB 102,72,15,58,22,225,1 ; pextrq $0x1,%xmm4,%rcx + DB 102,72,15,126,229 ; movq %xmm4,%rbp + DB 65,137,232 ; mov %ebp,%r8d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 137,203 ; mov %ecx,%ebx + DB 243,70,15,16,28,128 ; movss (%rax,%r8,4),%xmm11 + DB 102,68,15,58,33,28,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm11 + DB 243,15,16,36,152 ; movss (%rax,%rbx,4),%xmm4 + DB 102,68,15,58,33,220,32 ; insertps $0x20,%xmm4,%xmm11 + DB 72,193,233,32 ; shr $0x20,%rcx + DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 + DB 102,68,15,58,33,220,48 ; insertps $0x30,%xmm4,%xmm11 + DB 102,15,111,61,110,54,1,0 ; movdqa 0x1366e(%rip),%xmm7 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,254,223 ; paddd %xmm7,%xmm3 + DB 102,72,15,58,22,217,1 ; pextrq $0x1,%xmm3,%rcx DB 102,72,15,126,221 ; movq %xmm3,%rbp - DB 137,233 ; mov %ebp,%ecx + DB 65,137,232 ; mov %ebp,%r8d DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,68,15,16,12,136 ; movss (%rax,%rcx,4),%xmm9 - DB 102,68,15,58,33,12,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm9 - DB 137,217 ; mov %ebx,%ecx - DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,68,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm9 + DB 137,203 ; mov %ecx,%ebx + DB 243,66,15,16,36,128 ; movss (%rax,%r8,4),%xmm4 + DB 102,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm4 + DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 + DB 102,15,58,33,227,32 ; insertps $0x20,%xmm3,%xmm4 + DB 72,193,233,32 ; shr $0x20,%rcx + DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 + DB 102,15,58,33,227,48 ; insertps $0x30,%xmm3,%xmm4 + DB 102,68,15,111,124,36,48 ; movdqa 0x30(%rsp),%xmm15 + DB 102,65,15,254,207 ; paddd %xmm15,%xmm1 + DB 102,15,56,64,202 ; pmulld %xmm2,%xmm1 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,250,222 ; psubd %xmm6,%xmm3 + DB 102,72,15,58,22,217,1 ; pextrq $0x1,%xmm3,%rcx + DB 102,72,15,126,221 ; movq %xmm3,%rbp + DB 65,137,232 ; mov %ebp,%r8d + DB 72,193,237,32 ; shr $0x20,%rbp + DB 137,203 ; mov %ecx,%ebx + DB 243,66,15,16,52,128 ; movss (%rax,%r8,4),%xmm6 + DB 102,15,58,33,52,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm6 + DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 + DB 102,15,58,33,243,32 ; insertps $0x20,%xmm3,%xmm6 + DB 72,193,233,32 ; shr $0x20,%rcx + DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 + DB 102,15,58,33,243,48 ; insertps $0x30,%xmm3,%xmm6 + DB 102,65,15,126,201 ; movd %xmm1,%r9d + DB 102,65,15,58,22,203,1 ; pextrd $0x1,%xmm1,%r11d + DB 102,65,15,58,22,202,2 ; pextrd $0x2,%xmm1,%r10d + DB 102,65,15,58,22,200,3 ; pextrd $0x3,%xmm1,%r8d + DB 102,15,254,207 ; paddd %xmm7,%xmm1 + DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 243,15,16,20,152 ; movss (%rax,%rbx,4),%xmm2 - DB 102,68,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm9 - DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 - DB 102,66,15,58,33,20,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm2 - DB 102,66,15,58,33,20,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm2 - DB 102,66,15,58,33,20,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm2 - DB 243,70,15,16,4,168 ; movss (%rax,%r13,4),%xmm8 - DB 102,70,15,58,33,4,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm8 - DB 102,70,15,58,33,4,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm8 - DB 102,70,15,58,33,4,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm8 - DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 69,15,89,199 ; mulps %xmm15,%xmm8 - DB 68,15,88,194 ; addps %xmm2,%xmm8 - DB 68,15,92,208 ; subps %xmm0,%xmm10 - DB 69,15,89,215 ; mulps %xmm15,%xmm10 - DB 68,15,88,208 ; addps %xmm0,%xmm10 - DB 68,15,92,201 ; subps %xmm1,%xmm9 - DB 69,15,89,207 ; mulps %xmm15,%xmm9 - DB 68,15,88,201 ; addps %xmm1,%xmm9 - DB 15,91,4,36 ; cvtdq2ps (%rsp),%xmm0 - DB 68,15,92,240 ; subps %xmm0,%xmm14 - DB 68,15,41,52,36 ; movaps %xmm14,(%rsp) - DB 68,15,92,199 ; subps %xmm7,%xmm8 - DB 69,15,89,198 ; mulps %xmm14,%xmm8 - DB 68,15,88,199 ; addps %xmm7,%xmm8 - DB 68,15,92,214 ; subps %xmm6,%xmm10 - DB 69,15,89,214 ; mulps %xmm14,%xmm10 - DB 68,15,88,214 ; addps %xmm6,%xmm10 - DB 68,15,92,204 ; subps %xmm4,%xmm9 - DB 69,15,89,206 ; mulps %xmm14,%xmm9 - DB 68,15,88,204 ; addps %xmm4,%xmm9 - DB 243,68,15,91,180,36,144,0,0,0 ; cvttps2dq 0x90(%rsp),%xmm14 - DB 102,68,15,56,64,180,36,160,0,0,0 ; pmulld 0xa0(%rsp),%xmm14 - DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 - DB 102,15,254,164,36,128,0,0,0 ; paddd 0x80(%rsp),%xmm4 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 137,205 ; mov %ecx,%ebp + DB 243,66,15,16,60,176 ; movss (%rax,%r14,4),%xmm7 + DB 102,15,58,33,60,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm7 + DB 243,15,16,12,168 ; movss (%rax,%rbp,4),%xmm1 + DB 102,15,58,33,249,32 ; insertps $0x20,%xmm1,%xmm7 + DB 72,193,233,32 ; shr $0x20,%rcx + DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 + DB 102,15,58,33,249,48 ; insertps $0x30,%xmm1,%xmm7 + DB 243,66,15,16,28,136 ; movss (%rax,%r9,4),%xmm3 + DB 102,66,15,58,33,28,152,16 ; insertps $0x10,(%rax,%r11,4),%xmm3 + DB 102,66,15,58,33,28,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm3 + DB 102,66,15,58,33,28,128,48 ; insertps $0x30,(%rax,%r8,4),%xmm3 + DB 65,15,92,218 ; subps %xmm10,%xmm3 + DB 68,15,41,164,36,128,0,0,0 ; movaps %xmm12,0x80(%rsp) + DB 65,15,89,220 ; mulps %xmm12,%xmm3 + DB 65,15,88,218 ; addps %xmm10,%xmm3 + DB 65,15,92,243 ; subps %xmm11,%xmm6 + DB 65,15,89,244 ; mulps %xmm12,%xmm6 + DB 65,15,88,243 ; addps %xmm11,%xmm6 + DB 15,92,252 ; subps %xmm4,%xmm7 + DB 65,15,89,252 ; mulps %xmm12,%xmm7 + DB 15,88,252 ; addps %xmm4,%xmm7 DB 102,15,254,68,36,80 ; paddd 0x50(%rsp),%xmm0 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,68,15,111,108,36,48 ; movdqa 0x30(%rsp),%xmm13 - DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 - DB 102,15,111,45,248,65,1,0 ; movdqa 0x141f8(%rip),%xmm5 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,56,64,205 ; pmulld %xmm5,%xmm1 - DB 102,15,126,201 ; movd %xmm1,%ecx - DB 102,15,58,22,205,1 ; pextrd $0x1,%xmm1,%ebp - DB 102,65,15,58,22,200,2 ; pextrd $0x2,%xmm1,%r8d - DB 102,15,58,22,203,3 ; pextrd $0x3,%xmm1,%ebx - DB 243,15,16,52,136 ; movss (%rax,%rcx,4),%xmm6 - DB 102,15,58,33,52,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm6 - DB 102,66,15,58,33,52,128,32 ; insertps $0x20,(%rax,%r8,4),%xmm6 - DB 243,15,16,20,152 ; movss (%rax,%rbx,4),%xmm2 - DB 102,15,58,33,242,48 ; insertps $0x30,%xmm2,%xmm6 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,111,29,33,59,1,0 ; movdqa 0x13b21(%rip),%xmm3 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx - DB 102,72,15,126,213 ; movq %xmm2,%rbp - DB 137,235 ; mov %ebp,%ebx + DB 102,15,254,76,36,64 ; paddd 0x40(%rsp),%xmm1 + DB 102,68,15,111,218 ; movdqa %xmm2,%xmm11 + DB 102,65,15,56,64,203 ; pmulld %xmm11,%xmm1 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,69,15,118,228 ; pcmpeqd %xmm12,%xmm12 + DB 102,65,15,250,228 ; psubd %xmm12,%xmm4 + DB 102,72,15,58,22,225,1 ; pextrq $0x1,%xmm4,%rcx + DB 102,72,15,126,229 ; movq %xmm4,%rbp + DB 65,137,232 ; mov %ebp,%r8d DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,60,152 ; movss (%rax,%rbx,4),%xmm7 - DB 102,15,58,33,60,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm7 - DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 - DB 102,15,58,33,250,32 ; insertps $0x20,%xmm2,%xmm7 + DB 137,203 ; mov %ecx,%ebx + DB 243,66,15,16,36,128 ; movss (%rax,%r8,4),%xmm4 + DB 102,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm4 + DB 243,15,16,20,152 ; movss (%rax,%rbx,4),%xmm2 + DB 102,15,58,33,226,32 ; insertps $0x20,%xmm2,%xmm4 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,15,58,33,250,48 ; insertps $0x30,%xmm2,%xmm7 - DB 102,65,15,254,203 ; paddd %xmm11,%xmm1 + DB 102,15,58,33,226,48 ; insertps $0x30,%xmm2,%xmm4 + DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,65,15,58,22,201,1 ; pextrd $0x1,%xmm1,%r9d + DB 102,65,15,58,22,202,2 ; pextrd $0x2,%xmm1,%r10d + DB 102,65,15,58,22,203,3 ; pextrd $0x3,%xmm1,%r11d + DB 102,68,15,111,21,174,52,1,0 ; movdqa 0x134ae(%rip),%xmm10 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx - DB 102,72,15,126,205 ; movq %xmm1,%rbp - DB 137,235 ; mov %ebp,%ebx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,12,152 ; movss (%rax,%rbx,4),%xmm1 - DB 102,15,58,33,12,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm1 - DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 65,137,222 ; mov %ebx,%r14d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 137,205 ; mov %ecx,%ebp + DB 243,66,15,16,12,176 ; movss (%rax,%r14,4),%xmm1 + DB 102,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm1 + DB 243,15,16,20,168 ; movss (%rax,%rbp,4),%xmm2 DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 - DB 102,68,15,111,92,36,32 ; movdqa 0x20(%rsp),%xmm11 - DB 102,65,15,254,195 ; paddd %xmm11,%xmm0 - DB 102,15,56,64,197 ; pmulld %xmm5,%xmm0 + DB 102,65,15,254,199 ; paddd %xmm15,%xmm0 + DB 102,65,15,56,64,195 ; pmulld %xmm11,%xmm0 DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,254,211 ; paddd %xmm3,%xmm2 + DB 102,65,15,250,212 ; psubd %xmm12,%xmm2 DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx - DB 102,72,15,126,213 ; movq %xmm2,%rbp - DB 137,235 ; mov %ebp,%ebx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 - DB 102,15,58,33,28,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm3 - DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 - DB 102,15,58,33,218,32 ; insertps $0x20,%xmm2,%xmm3 - DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,15,58,33,218,48 ; insertps $0x30,%xmm2,%xmm3 - DB 102,65,15,126,193 ; movd %xmm0,%r9d - DB 102,65,15,58,22,195,1 ; pextrd $0x1,%xmm0,%r11d - DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d - DB 102,65,15,58,22,192,3 ; pextrd $0x3,%xmm0,%r8d - DB 102,15,254,5,73,58,1,0 ; paddd 0x13a49(%rip),%xmm0 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx - DB 102,72,15,126,195 ; movq %xmm0,%rbx - DB 137,221 ; mov %ebx,%ebp + DB 102,72,15,126,211 ; movq %xmm2,%rbx + DB 65,137,222 ; mov %ebx,%r14d DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,44,168 ; movss (%rax,%rbp,4),%xmm5 - DB 102,15,58,33,44,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm5 - DB 243,66,15,16,4,176 ; movss (%rax,%r14,4),%xmm0 - DB 102,15,58,33,232,32 ; insertps $0x20,%xmm0,%xmm5 - DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 - DB 102,15,58,33,232,48 ; insertps $0x30,%xmm0,%xmm5 - DB 243,66,15,16,20,136 ; movss (%rax,%r9,4),%xmm2 - DB 102,66,15,58,33,20,152,16 ; insertps $0x10,(%rax,%r11,4),%xmm2 - DB 102,66,15,58,33,20,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm2 - DB 102,66,15,58,33,20,128,48 ; insertps $0x30,(%rax,%r8,4),%xmm2 - DB 15,92,214 ; subps %xmm6,%xmm2 - DB 65,15,89,215 ; mulps %xmm15,%xmm2 - DB 15,88,214 ; addps %xmm6,%xmm2 - DB 15,92,223 ; subps %xmm7,%xmm3 - DB 65,15,89,223 ; mulps %xmm15,%xmm3 - DB 15,88,223 ; addps %xmm7,%xmm3 - DB 15,92,233 ; subps %xmm1,%xmm5 - DB 65,15,89,239 ; mulps %xmm15,%xmm5 - DB 15,88,233 ; addps %xmm1,%xmm5 - DB 102,15,254,100,36,16 ; paddd 0x10(%rsp),%xmm4 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,65,15,254,197 ; paddd %xmm13,%xmm0 - DB 102,68,15,111,45,41,64,1,0 ; movdqa 0x14029(%rip),%xmm13 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,65,15,56,64,197 ; pmulld %xmm13,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,53,135,57,1,0 ; movdqa 0x13987(%rip),%xmm6 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,206 ; paddd %xmm6,%xmm1 - DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx - DB 102,72,15,126,205 ; movq %xmm1,%rbp - DB 137,235 ; mov %ebp,%ebx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,12,152 ; movss (%rax,%rbx,4),%xmm1 - DB 102,15,58,33,12,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm1 - DB 243,66,15,16,52,128 ; movss (%rax,%r8,4),%xmm6 - DB 102,15,58,33,206,32 ; insertps $0x20,%xmm6,%xmm1 + DB 137,205 ; mov %ecx,%ebp + DB 243,70,15,16,60,176 ; movss (%rax,%r14,4),%xmm15 + DB 102,68,15,58,33,60,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm15 + DB 243,15,16,20,168 ; movss (%rax,%rbp,4),%xmm2 + DB 102,68,15,58,33,250,32 ; insertps $0x20,%xmm2,%xmm15 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,52,136 ; movss (%rax,%rcx,4),%xmm6 - DB 102,15,58,33,206,48 ; insertps $0x30,%xmm6,%xmm1 - DB 102,65,15,126,192 ; movd %xmm0,%r8d - DB 102,65,15,58,22,193,1 ; pextrd $0x1,%xmm0,%r9d - DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d - DB 102,65,15,58,22,195,3 ; pextrd $0x3,%xmm0,%r11d - DB 102,15,111,61,69,57,1,0 ; movdqa 0x13945(%rip),%xmm7 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,254,199 ; paddd %xmm7,%xmm0 + DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 + DB 102,68,15,58,33,250,48 ; insertps $0x30,%xmm2,%xmm15 + DB 102,65,15,126,199 ; movd %xmm0,%r15d + DB 102,65,15,58,22,196,1 ; pextrd $0x1,%xmm0,%r12d + DB 102,65,15,58,22,197,2 ; pextrd $0x2,%xmm0,%r13d + DB 102,65,15,58,22,198,3 ; pextrd $0x3,%xmm0,%r14d + DB 102,65,15,254,194 ; paddd %xmm10,%xmm0 DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,195 ; movq %xmm0,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,4,168 ; movss (%rax,%rbp,4),%xmm0 - DB 102,15,58,33,4,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm0 - DB 243,66,15,16,52,176 ; movss (%rax,%r14,4),%xmm6 - DB 102,15,58,33,198,32 ; insertps $0x20,%xmm6,%xmm0 - DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,52,136 ; movss (%rax,%rcx,4),%xmm6 - DB 102,15,58,33,198,48 ; insertps $0x30,%xmm6,%xmm0 - DB 102,65,15,254,227 ; paddd %xmm11,%xmm4 - DB 102,65,15,56,64,229 ; pmulld %xmm13,%xmm4 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,68,15,111,45,205,56,1,0 ; movdqa 0x138cd(%rip),%xmm13 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,245 ; paddd %xmm13,%xmm6 - DB 102,72,15,58,22,241,1 ; pextrq $0x1,%xmm6,%rcx - DB 102,72,15,126,243 ; movq %xmm6,%rbx - DB 137,221 ; mov %ebx,%ebp - DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d DB 243,68,15,16,28,168 ; movss (%rax,%rbp,4),%xmm11 DB 102,68,15,58,33,28,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm11 - DB 243,66,15,16,52,176 ; movss (%rax,%r14,4),%xmm6 - DB 102,68,15,58,33,222,32 ; insertps $0x20,%xmm6,%xmm11 + DB 137,203 ; mov %ecx,%ebx + DB 243,15,16,4,152 ; movss (%rax,%rbx,4),%xmm0 + DB 102,68,15,58,33,216,32 ; insertps $0x20,%xmm0,%xmm11 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,52,136 ; movss (%rax,%rcx,4),%xmm6 - DB 102,68,15,58,33,222,48 ; insertps $0x30,%xmm6,%xmm11 - DB 102,65,15,126,229 ; movd %xmm4,%r13d - DB 102,65,15,58,22,231,1 ; pextrd $0x1,%xmm4,%r15d - DB 102,65,15,58,22,228,2 ; pextrd $0x2,%xmm4,%r12d - DB 102,65,15,58,22,230,3 ; pextrd $0x3,%xmm4,%r14d - DB 102,15,254,231 ; paddd %xmm7,%xmm4 - DB 102,72,15,58,22,227,1 ; pextrq $0x1,%xmm4,%rbx - DB 102,72,15,126,229 ; movq %xmm4,%rbp - DB 137,233 ; mov %ebp,%ecx - DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,15,16,60,136 ; movss (%rax,%rcx,4),%xmm7 - DB 102,15,58,33,60,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm7 - DB 137,217 ; mov %ebx,%ecx - DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 - DB 102,15,58,33,252,32 ; insertps $0x20,%xmm4,%xmm7 - DB 72,193,235,32 ; shr $0x20,%rbx - DB 243,15,16,36,152 ; movss (%rax,%rbx,4),%xmm4 - DB 102,15,58,33,252,48 ; insertps $0x30,%xmm4,%xmm7 - DB 243,66,15,16,36,128 ; movss (%rax,%r8,4),%xmm4 - DB 102,66,15,58,33,36,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm4 - DB 102,66,15,58,33,36,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm4 - DB 102,66,15,58,33,36,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm4 - DB 243,66,15,16,52,168 ; movss (%rax,%r13,4),%xmm6 - DB 102,66,15,58,33,52,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm6 - DB 102,66,15,58,33,52,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm6 - DB 102,66,15,58,33,52,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm6 - DB 15,92,244 ; subps %xmm4,%xmm6 - DB 65,15,89,247 ; mulps %xmm15,%xmm6 - DB 15,88,244 ; addps %xmm4,%xmm6 + DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 + DB 102,68,15,58,33,216,48 ; insertps $0x30,%xmm0,%xmm11 + DB 243,66,15,16,4,128 ; movss (%rax,%r8,4),%xmm0 + DB 102,66,15,58,33,4,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm0 + DB 102,66,15,58,33,4,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm0 + DB 102,66,15,58,33,4,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm0 + DB 243,70,15,16,20,184 ; movss (%rax,%r15,4),%xmm10 + DB 102,70,15,58,33,20,160,16 ; insertps $0x10,(%rax,%r12,4),%xmm10 + DB 102,70,15,58,33,20,168,32 ; insertps $0x20,(%rax,%r13,4),%xmm10 + DB 102,70,15,58,33,20,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm10 + DB 68,15,92,208 ; subps %xmm0,%xmm10 + DB 68,15,40,164,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm12 + DB 69,15,89,212 ; mulps %xmm12,%xmm10 + DB 68,15,88,208 ; addps %xmm0,%xmm10 + DB 68,15,92,252 ; subps %xmm4,%xmm15 + DB 69,15,89,252 ; mulps %xmm12,%xmm15 + DB 68,15,88,252 ; addps %xmm4,%xmm15 DB 68,15,92,217 ; subps %xmm1,%xmm11 - DB 69,15,89,223 ; mulps %xmm15,%xmm11 + DB 69,15,89,220 ; mulps %xmm12,%xmm11 DB 68,15,88,217 ; addps %xmm1,%xmm11 - DB 15,92,248 ; subps %xmm0,%xmm7 - DB 65,15,89,255 ; mulps %xmm15,%xmm7 - DB 15,88,248 ; addps %xmm0,%xmm7 - DB 15,92,242 ; subps %xmm2,%xmm6 + DB 68,15,92,211 ; subps %xmm3,%xmm10 DB 15,40,4,36 ; movaps (%rsp),%xmm0 - DB 15,89,240 ; mulps %xmm0,%xmm6 - DB 15,88,242 ; addps %xmm2,%xmm6 - DB 68,15,92,219 ; subps %xmm3,%xmm11 + DB 68,15,89,208 ; mulps %xmm0,%xmm10 + DB 68,15,88,211 ; addps %xmm3,%xmm10 + DB 68,15,92,254 ; subps %xmm6,%xmm15 + DB 68,15,89,248 ; mulps %xmm0,%xmm15 + DB 68,15,88,254 ; addps %xmm6,%xmm15 + DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 68,15,89,216 ; mulps %xmm0,%xmm11 - DB 68,15,88,219 ; addps %xmm3,%xmm11 - DB 15,92,253 ; subps %xmm5,%xmm7 - DB 15,89,248 ; mulps %xmm0,%xmm7 - DB 15,88,253 ; addps %xmm5,%xmm7 - DB 15,91,68,36,64 ; cvtdq2ps 0x40(%rsp),%xmm0 + DB 68,15,88,223 ; addps %xmm7,%xmm11 + DB 15,91,132,36,160,0,0,0 ; cvtdq2ps 0xa0(%rsp),%xmm0 DB 15,40,76,36,96 ; movaps 0x60(%rsp),%xmm1 DB 15,92,200 ; subps %xmm0,%xmm1 DB 15,41,76,36,96 ; movaps %xmm1,0x60(%rsp) - DB 65,15,92,240 ; subps %xmm8,%xmm6 - DB 15,89,241 ; mulps %xmm1,%xmm6 - DB 65,15,88,240 ; addps %xmm8,%xmm6 - DB 69,15,92,218 ; subps %xmm10,%xmm11 + DB 68,15,92,213 ; subps %xmm5,%xmm10 + DB 68,15,89,209 ; mulps %xmm1,%xmm10 + DB 68,15,88,213 ; addps %xmm5,%xmm10 + DB 69,15,92,248 ; subps %xmm8,%xmm15 + DB 68,15,89,249 ; mulps %xmm1,%xmm15 + DB 69,15,88,248 ; addps %xmm8,%xmm15 + DB 69,15,92,217 ; subps %xmm9,%xmm11 DB 68,15,89,217 ; mulps %xmm1,%xmm11 - DB 69,15,88,218 ; addps %xmm10,%xmm11 - DB 65,15,92,249 ; subps %xmm9,%xmm7 - DB 15,89,249 ; mulps %xmm1,%xmm7 - DB 65,15,88,249 ; addps %xmm9,%xmm7 - DB 243,15,91,132,36,176,0,0,0 ; cvttps2dq 0xb0(%rsp),%xmm0 - DB 102,15,127,68,36,64 ; movdqa %xmm0,0x40(%rsp) - DB 102,68,15,254,224 ; paddd %xmm0,%xmm12 - DB 102,15,111,84,36,80 ; movdqa 0x50(%rsp),%xmm2 - DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 - DB 102,68,15,111,76,36,48 ; movdqa 0x30(%rsp),%xmm9 - DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 - DB 102,15,254,194 ; paddd %xmm2,%xmm0 - DB 102,15,111,29,215,61,1,0 ; movdqa 0x13dd7(%rip),%xmm3 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,56,64,195 ; pmulld %xmm3,%xmm0 + DB 69,15,88,217 ; addps %xmm9,%xmm11 + DB 243,15,91,68,36,32 ; cvttps2dq 0x20(%rsp),%xmm0 + DB 102,15,127,68,36,32 ; movdqa %xmm0,0x20(%rsp) + DB 102,68,15,254,232 ; paddd %xmm0,%xmm13 + DB 102,65,15,111,222 ; movdqa %xmm14,%xmm3 + DB 102,65,15,254,221 ; paddd %xmm13,%xmm3 + DB 102,15,111,124,36,64 ; movdqa 0x40(%rsp),%xmm7 + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,15,254,195 ; paddd %xmm3,%xmm0 + DB 102,15,111,21,66,57,1,0 ; movdqa 0x13942(%rip),%xmm2 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,56,64,194 ; pmulld %xmm2,%xmm0 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,69,15,111,213 ; movdqa %xmm13,%xmm10 - DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 + DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 + DB 102,65,15,250,200 ; psubd %xmm8,%xmm1 DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx DB 102,72,15,126,205 ; movq %xmm1,%rbp DB 137,235 ; mov %ebp,%ebx DB 72,193,237,32 ; shr $0x20,%rbp DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,36,152 ; movss (%rax,%rbx,4),%xmm4 - DB 102,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm4 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,15,58,33,44,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm5 DB 243,66,15,16,12,128 ; movss (%rax,%r8,4),%xmm1 - DB 102,15,58,33,225,32 ; insertps $0x20,%xmm1,%xmm4 + DB 102,15,58,33,233,32 ; insertps $0x20,%xmm1,%xmm5 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 - DB 102,15,58,33,225,48 ; insertps $0x30,%xmm1,%xmm4 + DB 102,15,58,33,233,48 ; insertps $0x30,%xmm1,%xmm5 DB 102,65,15,126,192 ; movd %xmm0,%r8d DB 102,65,15,58,22,193,1 ; pextrd $0x1,%xmm0,%r9d DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d DB 102,65,15,58,22,195,3 ; pextrd $0x3,%xmm0,%r11d - DB 102,15,111,45,246,54,1,0 ; movdqa 0x136f6(%rip),%xmm5 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,254,197 ; paddd %xmm5,%xmm0 + DB 102,15,111,37,81,50,1,0 ; movdqa 0x13251(%rip),%xmm4 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,254,196 ; paddd %xmm4,%xmm0 + DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx + DB 102,72,15,126,195 ; movq %xmm0,%rbx + DB 137,221 ; mov %ebx,%ebp + DB 72,193,235,32 ; shr $0x20,%rbx + DB 65,137,206 ; mov %ecx,%r14d + DB 243,15,16,52,168 ; movss (%rax,%rbp,4),%xmm6 + DB 102,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm6 + DB 243,66,15,16,4,176 ; movss (%rax,%r14,4),%xmm0 + DB 102,15,58,33,240,32 ; insertps $0x20,%xmm0,%xmm6 + DB 72,193,233,32 ; shr $0x20,%rcx + DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 + DB 102,15,58,33,240,48 ; insertps $0x30,%xmm0,%xmm6 + DB 102,68,15,111,76,36,48 ; movdqa 0x30(%rsp),%xmm9 + DB 102,65,15,254,217 ; paddd %xmm9,%xmm3 + DB 102,15,56,64,218 ; pmulld %xmm2,%xmm3 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,65,15,250,192 ; psubd %xmm8,%xmm0 DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,195 ; movq %xmm0,%rbx DB 137,221 ; mov %ebx,%ebp @@ -24193,164 +24367,151 @@ _sk_clut_4D_sse41 LABEL PROC DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 DB 102,15,58,33,193,48 ; insertps $0x30,%xmm1,%xmm0 - DB 102,68,15,111,68,36,32 ; movdqa 0x20(%rsp),%xmm8 - DB 102,65,15,254,208 ; paddd %xmm8,%xmm2 - DB 102,15,56,64,211 ; pmulld %xmm3,%xmm2 - DB 102,68,15,111,235 ; movdqa %xmm3,%xmm13 - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 - DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx - DB 102,72,15,126,203 ; movq %xmm1,%rbx - DB 137,221 ; mov %ebx,%ebp - DB 72,193,235,32 ; shr $0x20,%rbx - DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,12,168 ; movss (%rax,%rbp,4),%xmm1 - DB 102,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm1 - DB 243,66,15,16,28,176 ; movss (%rax,%r14,4),%xmm3 - DB 102,15,58,33,203,32 ; insertps $0x20,%xmm3,%xmm1 - DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 - DB 102,15,58,33,203,48 ; insertps $0x30,%xmm3,%xmm1 - DB 102,65,15,126,213 ; movd %xmm2,%r13d - DB 102,65,15,58,22,215,1 ; pextrd $0x1,%xmm2,%r15d - DB 102,65,15,58,22,212,2 ; pextrd $0x2,%xmm2,%r12d - DB 102,65,15,58,22,214,3 ; pextrd $0x3,%xmm2,%r14d - DB 102,15,254,213 ; paddd %xmm5,%xmm2 - DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx - DB 102,72,15,126,213 ; movq %xmm2,%rbp + DB 102,65,15,126,221 ; movd %xmm3,%r13d + DB 102,65,15,58,22,223,1 ; pextrd $0x1,%xmm3,%r15d + DB 102,65,15,58,22,220,2 ; pextrd $0x2,%xmm3,%r12d + DB 102,65,15,58,22,222,3 ; pextrd $0x3,%xmm3,%r14d + DB 102,15,254,220 ; paddd %xmm4,%xmm3 + DB 102,68,15,111,196 ; movdqa %xmm4,%xmm8 + DB 102,72,15,58,22,219,1 ; pextrq $0x1,%xmm3,%rbx + DB 102,72,15,126,221 ; movq %xmm3,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,15,58,33,20,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm2 - DB 137,217 ; mov %ebx,%ecx DB 243,15,16,28,136 ; movss (%rax,%rcx,4),%xmm3 - DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2 + DB 102,15,58,33,28,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm3 + DB 137,217 ; mov %ebx,%ecx + DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 + DB 102,15,58,33,217,32 ; insertps $0x20,%xmm1,%xmm3 DB 72,193,235,32 ; shr $0x20,%rbx - DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 - DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2 - DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 - DB 102,66,15,58,33,44,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm5 - DB 102,66,15,58,33,44,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm5 - DB 102,66,15,58,33,44,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm5 - DB 243,66,15,16,28,168 ; movss (%rax,%r13,4),%xmm3 - DB 102,66,15,58,33,28,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm3 - DB 102,66,15,58,33,28,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm3 - DB 102,66,15,58,33,28,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm3 - DB 15,92,221 ; subps %xmm5,%xmm3 - DB 65,15,89,223 ; mulps %xmm15,%xmm3 - DB 15,88,221 ; addps %xmm5,%xmm3 - DB 15,92,204 ; subps %xmm4,%xmm1 - DB 65,15,89,207 ; mulps %xmm15,%xmm1 - DB 15,88,204 ; addps %xmm4,%xmm1 - DB 15,92,208 ; subps %xmm0,%xmm2 - DB 65,15,89,215 ; mulps %xmm15,%xmm2 - DB 15,88,208 ; addps %xmm0,%xmm2 - DB 102,68,15,254,100,36,16 ; paddd 0x10(%rsp),%xmm12 - DB 102,65,15,111,228 ; movdqa %xmm12,%xmm4 - DB 102,65,15,254,225 ; paddd %xmm9,%xmm4 - DB 102,65,15,56,64,229 ; pmulld %xmm13,%xmm4 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,65,15,254,194 ; paddd %xmm10,%xmm0 - DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx - DB 102,72,15,126,197 ; movq %xmm0,%rbp + DB 243,15,16,12,152 ; movss (%rax,%rbx,4),%xmm1 + DB 102,15,58,33,217,48 ; insertps $0x30,%xmm1,%xmm3 + DB 243,66,15,16,12,128 ; movss (%rax,%r8,4),%xmm1 + DB 102,66,15,58,33,12,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm1 + DB 102,66,15,58,33,12,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm1 + DB 102,66,15,58,33,12,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm1 + DB 243,66,15,16,36,168 ; movss (%rax,%r13,4),%xmm4 + DB 102,66,15,58,33,36,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm4 + DB 102,66,15,58,33,36,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm4 + DB 102,66,15,58,33,36,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm4 + DB 15,92,225 ; subps %xmm1,%xmm4 + DB 65,15,89,228 ; mulps %xmm12,%xmm4 + DB 15,88,225 ; addps %xmm1,%xmm4 + DB 15,92,197 ; subps %xmm5,%xmm0 + DB 65,15,89,196 ; mulps %xmm12,%xmm0 + DB 15,88,197 ; addps %xmm5,%xmm0 + DB 15,92,222 ; subps %xmm6,%xmm3 + DB 65,15,89,220 ; mulps %xmm12,%xmm3 + DB 15,88,222 ; addps %xmm6,%xmm3 + DB 102,68,15,254,108,36,80 ; paddd 0x50(%rsp),%xmm13 + DB 102,65,15,111,205 ; movdqa %xmm13,%xmm1 + DB 102,15,254,207 ; paddd %xmm7,%xmm1 + DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 + DB 102,65,15,56,64,204 ; pmulld %xmm12,%xmm1 + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,15,250,21,109,55,1,0 ; psubd 0x1376d(%rip),%xmm2 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx + DB 102,72,15,126,213 ; movq %xmm2,%rbp DB 137,235 ; mov %ebp,%ebx DB 72,193,237,32 ; shr $0x20,%rbp DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,4,152 ; movss (%rax,%rbx,4),%xmm0 - DB 102,15,58,33,4,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm0 - DB 243,66,15,16,44,128 ; movss (%rax,%r8,4),%xmm5 - DB 102,15,58,33,197,32 ; insertps $0x20,%xmm5,%xmm0 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,15,58,33,44,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm5 + DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 + DB 102,15,58,33,234,32 ; insertps $0x20,%xmm2,%xmm5 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,44,136 ; movss (%rax,%rcx,4),%xmm5 - DB 102,15,58,33,197,48 ; insertps $0x30,%xmm5,%xmm0 - DB 102,65,15,126,224 ; movd %xmm4,%r8d - DB 102,65,15,58,22,225,1 ; pextrd $0x1,%xmm4,%r9d - DB 102,65,15,58,22,226,2 ; pextrd $0x2,%xmm4,%r10d - DB 102,65,15,58,22,227,3 ; pextrd $0x3,%xmm4,%r11d - DB 102,68,15,111,13,42,53,1,0 ; movdqa 0x1352a(%rip),%xmm9 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,65,15,254,225 ; paddd %xmm9,%xmm4 - DB 102,72,15,58,22,225,1 ; pextrq $0x1,%xmm4,%rcx - DB 102,72,15,126,227 ; movq %xmm4,%rbx + DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 + DB 102,15,58,33,234,48 ; insertps $0x30,%xmm2,%xmm5 + DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,65,15,58,22,201,1 ; pextrd $0x1,%xmm1,%r9d + DB 102,65,15,58,22,202,2 ; pextrd $0x2,%xmm1,%r10d + DB 102,65,15,58,22,203,3 ; pextrd $0x3,%xmm1,%r11d + DB 102,65,15,254,200 ; paddd %xmm8,%xmm1 + DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx + DB 102,72,15,126,203 ; movq %xmm1,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,44,168 ; movss (%rax,%rbp,4),%xmm5 - DB 102,15,58,33,44,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm5 - DB 243,66,15,16,36,176 ; movss (%rax,%r14,4),%xmm4 - DB 102,15,58,33,236,32 ; insertps $0x20,%xmm4,%xmm5 + DB 243,15,16,52,168 ; movss (%rax,%rbp,4),%xmm6 + DB 102,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm6 + DB 243,66,15,16,12,176 ; movss (%rax,%r14,4),%xmm1 + DB 102,15,58,33,241,32 ; insertps $0x20,%xmm1,%xmm6 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 - DB 102,15,58,33,236,48 ; insertps $0x30,%xmm4,%xmm5 - DB 102,69,15,254,224 ; paddd %xmm8,%xmm12 - DB 102,69,15,56,64,229 ; pmulld %xmm13,%xmm12 - DB 102,65,15,111,228 ; movdqa %xmm12,%xmm4 - DB 102,65,15,254,226 ; paddd %xmm10,%xmm4 - DB 102,72,15,58,22,225,1 ; pextrq $0x1,%xmm4,%rcx - DB 102,72,15,126,227 ; movq %xmm4,%rbx + DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 + DB 102,15,58,33,241,48 ; insertps $0x30,%xmm1,%xmm6 + DB 102,69,15,254,233 ; paddd %xmm9,%xmm13 + DB 102,65,15,111,249 ; movdqa %xmm9,%xmm7 + DB 102,69,15,56,64,236 ; pmulld %xmm12,%xmm13 + DB 102,65,15,111,205 ; movdqa %xmm13,%xmm1 + DB 102,15,250,13,185,54,1,0 ; psubd 0x136b9(%rip),%xmm1 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx + DB 102,72,15,126,203 ; movq %xmm1,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,68,15,16,20,168 ; movss (%rax,%rbp,4),%xmm10 - DB 102,68,15,58,33,20,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm10 - DB 243,66,15,16,36,176 ; movss (%rax,%r14,4),%xmm4 - DB 102,68,15,58,33,212,32 ; insertps $0x20,%xmm4,%xmm10 + DB 243,68,15,16,12,168 ; movss (%rax,%rbp,4),%xmm9 + DB 102,68,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm9 + DB 243,66,15,16,12,176 ; movss (%rax,%r14,4),%xmm1 + DB 102,68,15,58,33,201,32 ; insertps $0x20,%xmm1,%xmm9 DB 72,193,233,32 ; shr $0x20,%rcx - DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 - DB 102,68,15,58,33,212,48 ; insertps $0x30,%xmm4,%xmm10 - DB 102,69,15,126,229 ; movd %xmm12,%r13d - DB 102,69,15,58,22,231,1 ; pextrd $0x1,%xmm12,%r15d - DB 102,69,15,58,22,228,2 ; pextrd $0x2,%xmm12,%r12d - DB 102,69,15,58,22,230,3 ; pextrd $0x3,%xmm12,%r14d - DB 102,69,15,254,225 ; paddd %xmm9,%xmm12 - DB 102,76,15,58,22,227,1 ; pextrq $0x1,%xmm12,%rbx - DB 102,76,15,126,229 ; movq %xmm12,%rbp + DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 + DB 102,68,15,58,33,201,48 ; insertps $0x30,%xmm1,%xmm9 + DB 102,69,15,126,237 ; movd %xmm13,%r13d + DB 102,69,15,58,22,239,1 ; pextrd $0x1,%xmm13,%r15d + DB 102,69,15,58,22,236,2 ; pextrd $0x2,%xmm13,%r12d + DB 102,69,15,58,22,238,3 ; pextrd $0x3,%xmm13,%r14d + DB 102,69,15,254,232 ; paddd %xmm8,%xmm13 + DB 102,76,15,58,22,235,1 ; pextrq $0x1,%xmm13,%rbx + DB 102,76,15,126,237 ; movq %xmm13,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp DB 243,68,15,16,4,136 ; movss (%rax,%rcx,4),%xmm8 DB 102,68,15,58,33,4,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm8 DB 137,217 ; mov %ebx,%ecx - DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 - DB 102,68,15,58,33,196,32 ; insertps $0x20,%xmm4,%xmm8 + DB 243,15,16,12,136 ; movss (%rax,%rcx,4),%xmm1 + DB 102,68,15,58,33,193,32 ; insertps $0x20,%xmm1,%xmm8 DB 72,193,235,32 ; shr $0x20,%rbx - DB 243,15,16,36,152 ; movss (%rax,%rbx,4),%xmm4 - DB 102,68,15,58,33,196,48 ; insertps $0x30,%xmm4,%xmm8 - DB 243,70,15,16,36,128 ; movss (%rax,%r8,4),%xmm12 - DB 102,70,15,58,33,36,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm12 - DB 102,70,15,58,33,36,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm12 - DB 102,70,15,58,33,36,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm12 - DB 243,66,15,16,36,168 ; movss (%rax,%r13,4),%xmm4 - DB 102,66,15,58,33,36,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm4 - DB 102,66,15,58,33,36,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm4 - DB 102,66,15,58,33,36,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm4 - DB 65,15,92,228 ; subps %xmm12,%xmm4 - DB 65,15,89,231 ; mulps %xmm15,%xmm4 - DB 65,15,88,228 ; addps %xmm12,%xmm4 - DB 68,15,92,208 ; subps %xmm0,%xmm10 - DB 69,15,89,215 ; mulps %xmm15,%xmm10 - DB 68,15,88,208 ; addps %xmm0,%xmm10 - DB 68,15,92,197 ; subps %xmm5,%xmm8 - DB 69,15,89,199 ; mulps %xmm15,%xmm8 - DB 68,15,88,197 ; addps %xmm5,%xmm8 - DB 15,92,227 ; subps %xmm3,%xmm4 - DB 15,40,4,36 ; movaps (%rsp),%xmm0 - DB 15,89,224 ; mulps %xmm0,%xmm4 - DB 15,88,227 ; addps %xmm3,%xmm4 - DB 68,15,92,209 ; subps %xmm1,%xmm10 - DB 68,15,89,208 ; mulps %xmm0,%xmm10 - DB 68,15,88,209 ; addps %xmm1,%xmm10 - DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 68,15,89,192 ; mulps %xmm0,%xmm8 - DB 68,15,88,194 ; addps %xmm2,%xmm8 - DB 102,68,15,254,116,36,64 ; paddd 0x40(%rsp),%xmm14 - DB 102,15,111,68,36,80 ; movdqa 0x50(%rsp),%xmm0 - DB 102,65,15,254,198 ; paddd %xmm14,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,254,76,36,48 ; paddd 0x30(%rsp),%xmm1 - DB 102,65,15,56,64,205 ; pmulld %xmm13,%xmm1 + DB 243,15,16,12,152 ; movss (%rax,%rbx,4),%xmm1 + DB 102,68,15,58,33,193,48 ; insertps $0x30,%xmm1,%xmm8 + DB 243,66,15,16,12,128 ; movss (%rax,%r8,4),%xmm1 + DB 102,66,15,58,33,12,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm1 + DB 102,66,15,58,33,12,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm1 + DB 102,66,15,58,33,12,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm1 + DB 243,70,15,16,36,168 ; movss (%rax,%r13,4),%xmm12 + DB 102,70,15,58,33,36,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm12 + DB 102,70,15,58,33,36,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm12 + DB 102,70,15,58,33,36,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm12 + DB 68,15,92,225 ; subps %xmm1,%xmm12 + DB 68,15,40,172,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm13 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 68,15,88,225 ; addps %xmm1,%xmm12 + DB 68,15,92,205 ; subps %xmm5,%xmm9 + DB 69,15,89,205 ; mulps %xmm13,%xmm9 + DB 68,15,88,205 ; addps %xmm5,%xmm9 + DB 68,15,92,198 ; subps %xmm6,%xmm8 + DB 69,15,89,197 ; mulps %xmm13,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 68,15,92,228 ; subps %xmm4,%xmm12 + DB 15,40,12,36 ; movaps (%rsp),%xmm1 + DB 68,15,89,225 ; mulps %xmm1,%xmm12 + DB 68,15,88,228 ; addps %xmm4,%xmm12 + DB 68,15,92,200 ; subps %xmm0,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 68,15,88,200 ; addps %xmm0,%xmm9 + DB 68,15,92,195 ; subps %xmm3,%xmm8 + DB 68,15,89,193 ; mulps %xmm1,%xmm8 + DB 68,15,88,195 ; addps %xmm3,%xmm8 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,254,68,36,32 ; paddd 0x20(%rsp),%xmm0 + DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) + DB 102,68,15,254,240 ; paddd %xmm0,%xmm14 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,15,111,108,36,64 ; movdqa 0x40(%rsp),%xmm5 + DB 102,15,254,205 ; paddd %xmm5,%xmm1 + DB 102,15,111,37,77,53,1,0 ; movdqa 0x1354d(%rip),%xmm4 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,56,64,204 ; pmulld %xmm4,%xmm1 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,68,15,111,37,95,51,1,0 ; movdqa 0x1335f(%rip),%xmm12 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,196 ; paddd %xmm12,%xmm0 + DB 102,15,118,246 ; pcmpeqd %xmm6,%xmm6 + DB 102,15,250,198 ; psubd %xmm6,%xmm0 DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,197 ; movq %xmm0,%rbp DB 137,235 ; mov %ebp,%ebx @@ -24367,7 +24528,7 @@ _sk_clut_4D_sse41 LABEL PROC DB 102,65,15,58,22,201,1 ; pextrd $0x1,%xmm1,%r9d DB 102,65,15,58,22,202,2 ; pextrd $0x2,%xmm1,%r10d DB 102,65,15,58,22,203,3 ; pextrd $0x3,%xmm1,%r11d - DB 102,15,111,29,28,51,1,0 ; movdqa 0x1331c(%rip),%xmm3 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> + DB 102,15,111,29,94,46,1,0 ; movdqa 0x12e5e(%rip),%xmm3 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> DB 102,15,254,203 ; paddd %xmm3,%xmm1 DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx DB 102,72,15,126,203 ; movq %xmm1,%rbx @@ -24381,99 +24542,99 @@ _sk_clut_4D_sse41 LABEL PROC DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1 - DB 102,15,254,108,36,32 ; paddd 0x20(%rsp),%xmm5 - DB 102,65,15,56,64,237 ; pmulld %xmm13,%xmm5 - DB 102,15,111,213 ; movdqa %xmm5,%xmm2 - DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 + DB 102,68,15,254,247 ; paddd %xmm7,%xmm14 + DB 102,68,15,56,64,244 ; pmulld %xmm4,%xmm14 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 + DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 + DB 102,15,250,214 ; psubd %xmm6,%xmm2 DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx DB 102,72,15,126,211 ; movq %xmm2,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,68,15,16,12,168 ; movss (%rax,%rbp,4),%xmm9 - DB 102,68,15,58,33,12,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm9 + DB 243,15,16,52,168 ; movss (%rax,%rbp,4),%xmm6 + DB 102,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm6 DB 243,66,15,16,20,176 ; movss (%rax,%r14,4),%xmm2 - DB 102,68,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm9 + DB 102,15,58,33,242,32 ; insertps $0x20,%xmm2,%xmm6 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,68,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm9 - DB 102,15,111,213 ; movdqa %xmm5,%xmm2 - DB 102,65,15,126,213 ; movd %xmm2,%r13d - DB 102,65,15,58,22,215,1 ; pextrd $0x1,%xmm2,%r15d - DB 102,65,15,58,22,212,2 ; pextrd $0x2,%xmm2,%r12d - DB 102,65,15,58,22,214,3 ; pextrd $0x3,%xmm2,%r14d - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx - DB 102,72,15,126,213 ; movq %xmm2,%rbp + DB 102,15,58,33,242,48 ; insertps $0x30,%xmm2,%xmm6 + DB 102,69,15,126,245 ; movd %xmm14,%r13d + DB 102,69,15,58,22,247,1 ; pextrd $0x1,%xmm14,%r15d + DB 102,69,15,58,22,244,2 ; pextrd $0x2,%xmm14,%r12d + DB 102,69,15,58,22,246,3 ; pextrd $0x3,%xmm14,%r14d + DB 102,68,15,254,243 ; paddd %xmm3,%xmm14 + DB 102,76,15,58,22,243,1 ; pextrq $0x1,%xmm14,%rbx + DB 102,76,15,126,245 ; movq %xmm14,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp - DB 243,68,15,16,36,136 ; movss (%rax,%rcx,4),%xmm12 - DB 102,68,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm12 + DB 243,15,16,36,136 ; movss (%rax,%rcx,4),%xmm4 + DB 102,15,58,33,36,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm4 DB 137,217 ; mov %ebx,%ecx DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 - DB 102,68,15,58,33,226,32 ; insertps $0x20,%xmm2,%xmm12 + DB 102,15,58,33,226,32 ; insertps $0x20,%xmm2,%xmm4 DB 72,193,235,32 ; shr $0x20,%rbx DB 243,15,16,20,152 ; movss (%rax,%rbx,4),%xmm2 - DB 102,68,15,58,33,226,48 ; insertps $0x30,%xmm2,%xmm12 + DB 102,15,58,33,226,48 ; insertps $0x30,%xmm2,%xmm4 DB 243,66,15,16,20,128 ; movss (%rax,%r8,4),%xmm2 DB 102,66,15,58,33,20,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm2 DB 102,66,15,58,33,20,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm2 DB 102,66,15,58,33,20,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm2 - DB 243,70,15,16,44,168 ; movss (%rax,%r13,4),%xmm13 - DB 102,70,15,58,33,44,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm13 - DB 102,70,15,58,33,44,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm13 - DB 102,70,15,58,33,44,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm13 - DB 68,15,92,234 ; subps %xmm2,%xmm13 - DB 69,15,89,239 ; mulps %xmm15,%xmm13 - DB 68,15,88,234 ; addps %xmm2,%xmm13 - DB 68,15,92,200 ; subps %xmm0,%xmm9 - DB 69,15,89,207 ; mulps %xmm15,%xmm9 - DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 69,15,89,231 ; mulps %xmm15,%xmm12 - DB 68,15,88,225 ; addps %xmm1,%xmm12 - DB 102,68,15,254,116,36,16 ; paddd 0x10(%rsp),%xmm14 - DB 102,15,111,68,36,48 ; movdqa 0x30(%rsp),%xmm0 - DB 102,65,15,254,198 ; paddd %xmm14,%xmm0 - DB 102,15,56,64,5,24,56,1,0 ; pmulld 0x13818(%rip),%xmm0 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,254,5,124,49,1,0 ; paddd 0x1317c(%rip),%xmm0 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> + DB 243,66,15,16,28,168 ; movss (%rax,%r13,4),%xmm3 + DB 102,66,15,58,33,28,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm3 + DB 102,66,15,58,33,28,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm3 + DB 102,66,15,58,33,28,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm3 + DB 15,92,218 ; subps %xmm2,%xmm3 + DB 65,15,89,221 ; mulps %xmm13,%xmm3 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 15,92,240 ; subps %xmm0,%xmm6 + DB 65,15,89,245 ; mulps %xmm13,%xmm6 + DB 15,88,240 ; addps %xmm0,%xmm6 + DB 15,92,225 ; subps %xmm1,%xmm4 + DB 65,15,89,229 ; mulps %xmm13,%xmm4 + DB 15,88,225 ; addps %xmm1,%xmm4 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,254,68,36,80 ; paddd 0x50(%rsp),%xmm0 + DB 102,15,254,232 ; paddd %xmm0,%xmm5 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,56,64,239 ; pmulld %xmm7,%xmm5 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,250,5,126,51,1,0 ; psubd 0x1337e(%rip),%xmm0 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,197 ; movq %xmm0,%rbp DB 137,235 ; mov %ebp,%ebx DB 72,193,237,32 ; shr $0x20,%rbp DB 65,137,200 ; mov %ecx,%r8d - DB 243,15,16,28,152 ; movss (%rax,%rbx,4),%xmm3 - DB 102,15,58,33,28,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm3 + DB 243,15,16,44,152 ; movss (%rax,%rbx,4),%xmm5 + DB 102,15,58,33,44,168,16 ; insertps $0x10,(%rax,%rbp,4),%xmm5 DB 243,66,15,16,4,128 ; movss (%rax,%r8,4),%xmm0 - DB 102,15,58,33,216,32 ; insertps $0x20,%xmm0,%xmm3 + DB 102,15,58,33,232,32 ; insertps $0x20,%xmm0,%xmm5 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 - DB 102,15,58,33,216,48 ; insertps $0x30,%xmm0,%xmm3 + DB 102,15,58,33,232,48 ; insertps $0x30,%xmm0,%xmm5 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,65,15,126,192 ; movd %xmm0,%r8d DB 102,65,15,58,22,193,1 ; pextrd $0x1,%xmm0,%r9d DB 102,65,15,58,22,194,2 ; pextrd $0x2,%xmm0,%r10d DB 102,65,15,58,22,195,3 ; pextrd $0x3,%xmm0,%r11d - DB 102,15,111,213 ; movdqa %xmm5,%xmm2 - DB 102,15,254,194 ; paddd %xmm2,%xmm0 + DB 102,15,254,5,140,44,1,0 ; paddd 0x12c8c(%rip),%xmm0 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,195 ; movq %xmm0,%rbx DB 137,221 ; mov %ebx,%ebp DB 72,193,235,32 ; shr $0x20,%rbx DB 65,137,206 ; mov %ecx,%r14d - DB 243,15,16,44,168 ; movss (%rax,%rbp,4),%xmm5 - DB 102,15,58,33,44,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm5 + DB 243,68,15,16,52,168 ; movss (%rax,%rbp,4),%xmm14 + DB 102,68,15,58,33,52,152,16 ; insertps $0x10,(%rax,%rbx,4),%xmm14 DB 243,66,15,16,4,176 ; movss (%rax,%r14,4),%xmm0 - DB 102,15,58,33,232,32 ; insertps $0x20,%xmm0,%xmm5 + DB 102,68,15,58,33,240,32 ; insertps $0x20,%xmm0,%xmm14 DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 - DB 102,15,58,33,232,48 ; insertps $0x30,%xmm0,%xmm5 - DB 102,68,15,254,116,36,32 ; paddd 0x20(%rsp),%xmm14 - DB 102,68,15,56,64,53,93,55,1,0 ; pmulld 0x1375d(%rip),%xmm14 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,111,5,197,48,1,0 ; movdqa 0x130c5(%rip),%xmm0 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,198 ; paddd %xmm14,%xmm0 + DB 102,68,15,58,33,240,48 ; insertps $0x30,%xmm0,%xmm14 + DB 102,15,254,84,36,48 ; paddd 0x30(%rsp),%xmm2 + DB 102,15,56,64,215 ; pmulld %xmm7,%xmm2 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,250,5,197,50,1,0 ; psubd 0x132c5(%rip),%xmm0 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx DB 102,72,15,126,195 ; movq %xmm0,%rbx DB 137,221 ; mov %ebx,%ebp @@ -24486,13 +24647,14 @@ _sk_clut_4D_sse41 LABEL PROC DB 72,193,233,32 ; shr $0x20,%rcx DB 243,15,16,4,136 ; movss (%rax,%rcx,4),%xmm0 DB 102,15,58,33,200,48 ; insertps $0x30,%xmm0,%xmm1 - DB 102,69,15,126,245 ; movd %xmm14,%r13d - DB 102,69,15,58,22,247,1 ; pextrd $0x1,%xmm14,%r15d - DB 102,69,15,58,22,244,2 ; pextrd $0x2,%xmm14,%r12d - DB 102,69,15,58,22,246,3 ; pextrd $0x3,%xmm14,%r14d - DB 102,68,15,254,242 ; paddd %xmm2,%xmm14 - DB 102,76,15,58,22,243,1 ; pextrq $0x1,%xmm14,%rbx - DB 102,76,15,126,245 ; movq %xmm14,%rbp + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,65,15,126,197 ; movd %xmm0,%r13d + DB 102,65,15,58,22,199,1 ; pextrd $0x1,%xmm0,%r15d + DB 102,65,15,58,22,196,2 ; pextrd $0x2,%xmm0,%r12d + DB 102,65,15,58,22,198,3 ; pextrd $0x3,%xmm0,%r14d + DB 102,15,254,5,211,43,1,0 ; paddd 0x12bd3(%rip),%xmm0 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,72,15,58,22,195,1 ; pextrq $0x1,%xmm0,%rbx + DB 102,72,15,126,197 ; movq %xmm0,%rbp DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp DB 243,15,16,20,136 ; movss (%rax,%rcx,4),%xmm2 @@ -24503,63 +24665,63 @@ _sk_clut_4D_sse41 LABEL PROC DB 72,193,235,32 ; shr $0x20,%rbx DB 243,15,16,4,152 ; movss (%rax,%rbx,4),%xmm0 DB 102,15,58,33,208,48 ; insertps $0x30,%xmm0,%xmm2 - DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 - DB 102,70,15,58,33,52,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm14 - DB 102,70,15,58,33,52,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm14 - DB 102,70,15,58,33,52,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm14 + DB 243,66,15,16,60,128 ; movss (%rax,%r8,4),%xmm7 + DB 102,66,15,58,33,60,136,16 ; insertps $0x10,(%rax,%r9,4),%xmm7 + DB 102,66,15,58,33,60,144,32 ; insertps $0x20,(%rax,%r10,4),%xmm7 + DB 102,66,15,58,33,60,152,48 ; insertps $0x30,(%rax,%r11,4),%xmm7 DB 243,66,15,16,4,168 ; movss (%rax,%r13,4),%xmm0 DB 102,66,15,58,33,4,184,16 ; insertps $0x10,(%rax,%r15,4),%xmm0 DB 102,66,15,58,33,4,160,32 ; insertps $0x20,(%rax,%r12,4),%xmm0 DB 102,66,15,58,33,4,176,48 ; insertps $0x30,(%rax,%r14,4),%xmm0 - DB 65,15,92,198 ; subps %xmm14,%xmm0 - DB 65,15,89,199 ; mulps %xmm15,%xmm0 - DB 65,15,88,198 ; addps %xmm14,%xmm0 - DB 15,92,203 ; subps %xmm3,%xmm1 - DB 65,15,89,207 ; mulps %xmm15,%xmm1 - DB 15,88,203 ; addps %xmm3,%xmm1 - DB 15,92,213 ; subps %xmm5,%xmm2 - DB 65,15,89,215 ; mulps %xmm15,%xmm2 - DB 15,88,213 ; addps %xmm5,%xmm2 - DB 65,15,92,197 ; subps %xmm13,%xmm0 - DB 15,40,28,36 ; movaps (%rsp),%xmm3 + DB 15,92,199 ; subps %xmm7,%xmm0 + DB 65,15,89,197 ; mulps %xmm13,%xmm0 + DB 15,88,199 ; addps %xmm7,%xmm0 + DB 15,92,205 ; subps %xmm5,%xmm1 + DB 65,15,89,205 ; mulps %xmm13,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 65,15,92,214 ; subps %xmm14,%xmm2 + DB 65,15,89,213 ; mulps %xmm13,%xmm2 + DB 65,15,88,214 ; addps %xmm14,%xmm2 + DB 15,92,195 ; subps %xmm3,%xmm0 + DB 15,40,44,36 ; movaps (%rsp),%xmm5 + DB 15,89,197 ; mulps %xmm5,%xmm0 + DB 15,88,195 ; addps %xmm3,%xmm0 + DB 15,92,206 ; subps %xmm6,%xmm1 + DB 15,89,205 ; mulps %xmm5,%xmm1 + DB 15,88,206 ; addps %xmm6,%xmm1 + DB 15,92,212 ; subps %xmm4,%xmm2 + DB 15,89,213 ; mulps %xmm5,%xmm2 + DB 15,88,212 ; addps %xmm4,%xmm2 + DB 65,15,92,196 ; subps %xmm12,%xmm0 + DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 65,15,88,197 ; addps %xmm13,%xmm0 + DB 65,15,88,196 ; addps %xmm12,%xmm0 DB 65,15,92,201 ; subps %xmm9,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 DB 65,15,88,201 ; addps %xmm9,%xmm1 - DB 65,15,92,212 ; subps %xmm12,%xmm2 - DB 15,89,211 ; mulps %xmm3,%xmm2 - DB 65,15,88,212 ; addps %xmm12,%xmm2 - DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 - DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,196 ; addps %xmm4,%xmm0 - DB 65,15,92,202 ; subps %xmm10,%xmm1 - DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 65,15,88,202 ; addps %xmm10,%xmm1 DB 65,15,92,208 ; subps %xmm8,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 DB 65,15,88,208 ; addps %xmm8,%xmm2 - DB 15,91,156,36,128,0,0,0 ; cvtdq2ps 0x80(%rsp),%xmm3 - DB 15,40,164,36,192,0,0,0 ; movaps 0xc0(%rsp),%xmm4 + DB 15,91,156,36,144,0,0,0 ; cvtdq2ps 0x90(%rsp),%xmm3 + DB 15,40,164,36,176,0,0,0 ; movaps 0xb0(%rsp),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 - DB 15,92,198 ; subps %xmm6,%xmm0 + DB 65,15,92,194 ; subps %xmm10,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 - DB 15,88,198 ; addps %xmm6,%xmm0 - DB 65,15,92,203 ; subps %xmm11,%xmm1 + DB 65,15,88,194 ; addps %xmm10,%xmm0 + DB 65,15,92,207 ; subps %xmm15,%xmm1 DB 15,89,204 ; mulps %xmm4,%xmm1 - DB 65,15,88,203 ; addps %xmm11,%xmm1 - DB 15,92,215 ; subps %xmm7,%xmm2 + DB 65,15,88,207 ; addps %xmm15,%xmm1 + DB 65,15,92,211 ; subps %xmm11,%xmm2 DB 15,89,212 ; mulps %xmm4,%xmm2 - DB 15,88,215 ; addps %xmm7,%xmm2 + DB 65,15,88,211 ; addps %xmm11,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,61,47,1,0 ; movaps 0x12f3d(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,139,42,1,0 ; movaps 0x12a8b(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,139,76,36,120 ; mov 0x78(%rsp),%rcx - DB 15,40,164,36,208,0,0,0 ; movaps 0xd0(%rsp),%xmm4 - DB 15,40,172,36,224,0,0,0 ; movaps 0xe0(%rsp),%xmm5 - DB 15,40,180,36,240,0,0,0 ; movaps 0xf0(%rsp),%xmm6 - DB 15,40,188,36,0,1,0,0 ; movaps 0x100(%rsp),%xmm7 - DB 72,129,196,24,1,0,0 ; add $0x118,%rsp + DB 15,40,164,36,192,0,0,0 ; movaps 0xc0(%rsp),%xmm4 + DB 15,40,172,36,208,0,0,0 ; movaps 0xd0(%rsp),%xmm5 + DB 15,40,180,36,224,0,0,0 ; movaps 0xe0(%rsp),%xmm6 + DB 15,40,188,36,240,0,0,0 ; movaps 0xf0(%rsp),%xmm7 + DB 72,129,196,8,1,0,0 ; add $0x108,%rsp DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,93 ; pop %r13 @@ -24599,13 +24761,13 @@ _sk_start_pipeline_sse2 LABEL PROC DB 73,137,246 ; mov %rsi,%r14 DB 76,137,141,8,255,255,255 ; mov %r9,-0xf8(%rbp) DB 76,57,203 ; cmp %r9,%rbx - DB 115,101 ; jae 1a69d <_sk_start_pipeline_sse2+0xe2> + DB 115,101 ; jae 1aa3f <_sk_start_pipeline_sse2+0xe2> DB 72,139,133,24,255,255,255 ; mov -0xe8(%rbp),%rax DB 72,141,64,4 ; lea 0x4(%rax),%rax DB 72,137,133,16,255,255,255 ; mov %rax,-0xf0(%rbp) DB 76,57,189,16,255,255,255 ; cmp %r15,-0xf0(%rbp) DB 72,139,149,24,255,255,255 ; mov -0xe8(%rbp),%rdx - DB 119,38 ; ja 1a680 <_sk_start_pipeline_sse2+0xc5> + DB 119,38 ; ja 1aa22 <_sk_start_pipeline_sse2+0xc5> DB 76,139,165,24,255,255,255 ; mov -0xe8(%rbp),%r12 DB 49,255 ; xor %edi,%edi DB 76,137,246 ; mov %r14,%rsi @@ -24616,16 +24778,16 @@ _sk_start_pipeline_sse2 LABEL PROC DB 73,131,196,8 ; add $0x8,%r12 DB 77,57,252 ; cmp %r15,%r12 DB 73,137,212 ; mov %rdx,%r12 - DB 118,225 ; jbe 1a661 <_sk_start_pipeline_sse2+0xa6> + DB 118,225 ; jbe 1aa03 <_sk_start_pipeline_sse2+0xa6> DB 76,137,255 ; mov %r15,%rdi DB 72,41,215 ; sub %rdx,%rdi - DB 116,9 ; je 1a691 <_sk_start_pipeline_sse2+0xd6> + DB 116,9 ; je 1aa33 <_sk_start_pipeline_sse2+0xd6> DB 76,137,246 ; mov %r14,%rsi DB 72,137,217 ; mov %rbx,%rcx DB 65,255,213 ; callq *%r13 DB 72,255,195 ; inc %rbx DB 72,59,157,8,255,255,255 ; cmp -0xf8(%rbp),%rbx - DB 117,173 ; jne 1a64a <_sk_start_pipeline_sse2+0x8f> + DB 117,173 ; jne 1a9ec <_sk_start_pipeline_sse2+0x8f> DB 15,40,181,32,255,255,255 ; movaps -0xe0(%rbp),%xmm6 DB 15,40,189,48,255,255,255 ; movaps -0xd0(%rbp),%xmm7 DB 68,15,40,133,64,255,255,255 ; movaps -0xc0(%rbp),%xmm8 @@ -24656,13 +24818,13 @@ _sk_seed_shader_sse2 LABEL PROC DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,88,5,153,45,1,0 ; addps 0x12d99(%rip),%xmm0 # 2d4a0 <_sk_overlay_sse2_8bit+0xf48> + DB 15,88,5,231,40,1,0 ; addps 0x128e7(%rip),%xmm0 # 2d390 <_sk_overlay_sse2_8bit+0xee1> DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,88,13,150,45,1,0 ; addps 0x12d96(%rip),%xmm1 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,88,13,228,40,1,0 ; addps 0x128e4(%rip),%xmm1 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,21,157,45,1,0 ; movaps 0x12d9d(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,235,40,1,0 ; movaps 0x128eb(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,87,237 ; xorps %xmm5,%xmm5 @@ -24674,36 +24836,36 @@ PUBLIC _sk_dither_sse2 _sk_dither_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 102,68,15,110,194 ; movd %edx,%xmm8 - DB 102,69,15,112,200,0 ; pshufd $0x0,%xmm8,%xmm9 - DB 102,68,15,254,13,54,45,1,0 ; paddd 0x12d36(%rip),%xmm9 # 2d480 <_sk_overlay_sse2_8bit+0xf28> - DB 102,68,15,110,193 ; movd %ecx,%xmm8 DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8 - DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 - DB 102,68,15,111,21,109,45,1,0 ; movdqa 0x12d6d(%rip),%xmm10 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 + DB 102,68,15,254,5,132,40,1,0 ; paddd 0x12884(%rip),%xmm8 # 2d370 <_sk_overlay_sse2_8bit+0xec1> + DB 102,68,15,110,201 ; movd %ecx,%xmm9 + DB 102,69,15,112,201,0 ; pshufd $0x0,%xmm9,%xmm9 + DB 102,69,15,239,200 ; pxor %xmm8,%xmm9 + DB 102,68,15,111,21,187,40,1,0 ; movdqa 0x128bb(%rip),%xmm10 # 2d3c0 <_sk_overlay_sse2_8bit+0xf11> + DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 DB 102,69,15,219,218 ; pand %xmm10,%xmm11 DB 102,65,15,114,243,5 ; pslld $0x5,%xmm11 - DB 102,69,15,219,209 ; pand %xmm9,%xmm10 + DB 102,69,15,219,208 ; pand %xmm8,%xmm10 DB 102,65,15,114,242,4 ; pslld $0x4,%xmm10 - DB 102,68,15,111,37,89,45,1,0 ; movdqa 0x12d59(%rip),%xmm12 # 2d4e0 <_sk_overlay_sse2_8bit+0xf88> - DB 102,68,15,111,45,96,45,1,0 ; movdqa 0x12d60(%rip),%xmm13 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 + DB 102,68,15,111,37,167,40,1,0 ; movdqa 0x128a7(%rip),%xmm12 # 2d3d0 <_sk_overlay_sse2_8bit+0xf21> + DB 102,68,15,111,45,174,40,1,0 ; movdqa 0x128ae(%rip),%xmm13 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,69,15,111,241 ; movdqa %xmm9,%xmm14 DB 102,69,15,219,245 ; pand %xmm13,%xmm14 DB 102,65,15,114,246,2 ; pslld $0x2,%xmm14 - DB 102,69,15,219,233 ; pand %xmm9,%xmm13 + DB 102,69,15,235,243 ; por %xmm11,%xmm14 + DB 102,69,15,219,232 ; pand %xmm8,%xmm13 DB 102,69,15,254,237 ; paddd %xmm13,%xmm13 - DB 102,69,15,219,196 ; pand %xmm12,%xmm8 - DB 102,65,15,114,208,1 ; psrld $0x1,%xmm8 - DB 102,69,15,219,204 ; pand %xmm12,%xmm9 - DB 102,65,15,114,209,2 ; psrld $0x2,%xmm9 DB 102,69,15,235,234 ; por %xmm10,%xmm13 - DB 102,69,15,235,233 ; por %xmm9,%xmm13 - DB 102,69,15,235,243 ; por %xmm11,%xmm14 - DB 102,69,15,235,245 ; por %xmm13,%xmm14 - DB 102,69,15,235,240 ; por %xmm8,%xmm14 - DB 69,15,91,198 ; cvtdq2ps %xmm14,%xmm8 - DB 68,15,89,5,27,45,1,0 ; mulps 0x12d1b(%rip),%xmm8 # 2d500 <_sk_overlay_sse2_8bit+0xfa8> - DB 68,15,88,5,35,45,1,0 ; addps 0x12d23(%rip),%xmm8 # 2d510 <_sk_overlay_sse2_8bit+0xfb8> + DB 102,69,15,219,204 ; pand %xmm12,%xmm9 + DB 102,65,15,114,209,1 ; psrld $0x1,%xmm9 + DB 102,69,15,219,196 ; pand %xmm12,%xmm8 + DB 102,65,15,114,208,2 ; psrld $0x2,%xmm8 + DB 102,69,15,235,197 ; por %xmm13,%xmm8 + DB 102,69,15,235,198 ; por %xmm14,%xmm8 + DB 102,69,15,235,193 ; por %xmm9,%xmm8 + DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 + DB 68,15,89,5,105,40,1,0 ; mulps 0x12869(%rip),%xmm8 # 2d3f0 <_sk_overlay_sse2_8bit+0xf41> + DB 68,15,88,5,113,40,1,0 ; addps 0x12871(%rip),%xmm8 # 2d400 <_sk_overlay_sse2_8bit+0xf51> DB 243,68,15,16,16 ; movss (%rax),%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 @@ -24741,7 +24903,7 @@ _sk_uniform_color_sse2 LABEL PROC PUBLIC _sk_black_color_sse2 _sk_black_color_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,89,44,1,0 ; movaps 0x12c59(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,167,39,1,0 ; movaps 0x127a7(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -24750,7 +24912,7 @@ _sk_black_color_sse2 LABEL PROC PUBLIC _sk_white_color_sse2 _sk_white_color_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,5,69,44,1,0 ; movaps 0x12c45(%rip),%xmm0 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,5,147,39,1,0 ; movaps 0x12793(%rip),%xmm0 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,40,216 ; movaps %xmm0,%xmm3 @@ -24788,7 +24950,7 @@ _sk_clear_sse2 LABEL PROC PUBLIC _sk_srcatop_sse2 _sk_srcatop_sse2 LABEL PROC DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 68,15,40,5,245,43,1,0 ; movaps 0x12bf5(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,67,39,1,0 ; movaps 0x12743(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -24811,7 +24973,7 @@ PUBLIC _sk_dstatop_sse2 _sk_dstatop_sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 DB 68,15,89,196 ; mulps %xmm4,%xmm8 - DB 68,15,40,13,168,43,1,0 ; movaps 0x12ba8(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,246,38,1,0 ; movaps 0x126f6(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 @@ -24852,7 +25014,7 @@ _sk_dstin_sse2 LABEL PROC PUBLIC _sk_srcout_sse2 _sk_srcout_sse2 LABEL PROC - DB 68,15,40,5,60,43,1,0 ; movaps 0x12b3c(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,138,38,1,0 ; movaps 0x1268a(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 @@ -24863,7 +25025,7 @@ _sk_srcout_sse2 LABEL PROC PUBLIC _sk_dstout_sse2 _sk_dstout_sse2 LABEL PROC - DB 68,15,40,5,28,43,1,0 ; movaps 0x12b1c(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,106,38,1,0 ; movaps 0x1266a(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -24878,7 +25040,7 @@ _sk_dstout_sse2 LABEL PROC PUBLIC _sk_srcover_sse2 _sk_srcover_sse2 LABEL PROC - DB 68,15,40,5,239,42,1,0 ; movaps 0x12aef(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,61,38,1,0 ; movaps 0x1263d(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -24896,7 +25058,7 @@ _sk_srcover_sse2 LABEL PROC PUBLIC _sk_dstover_sse2 _sk_dstover_sse2 LABEL PROC - DB 68,15,40,5,179,42,1,0 ; movaps 0x12ab3(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,1,38,1,0 ; movaps 0x12601(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -24920,7 +25082,7 @@ _sk_modulate_sse2 LABEL PROC PUBLIC _sk_multiply_sse2 _sk_multiply_sse2 LABEL PROC - DB 68,15,40,5,119,42,1,0 ; movaps 0x12a77(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,197,37,1,0 ; movaps 0x125c5(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 69,15,40,209 ; movaps %xmm9,%xmm10 @@ -24956,7 +25118,7 @@ _sk_multiply_sse2 LABEL PROC PUBLIC _sk_plus__sse2 _sk_plus__sse2 LABEL PROC DB 15,88,196 ; addps %xmm4,%xmm0 - DB 68,15,40,5,248,41,1,0 ; movaps 0x129f8(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,70,37,1,0 ; movaps 0x12546(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 15,88,205 ; addps %xmm5,%xmm1 DB 65,15,93,200 ; minps %xmm8,%xmm1 @@ -24994,7 +25156,7 @@ _sk_screen_sse2 LABEL PROC PUBLIC _sk_xor__sse2 _sk_xor__sse2 LABEL PROC DB 68,15,40,195 ; movaps %xmm3,%xmm8 - DB 15,40,29,132,41,1,0 ; movaps 0x12984(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,210,36,1,0 ; movaps 0x124d2(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,40,203 ; movaps %xmm3,%xmm9 DB 68,15,92,207 ; subps %xmm7,%xmm9 DB 65,15,89,193 ; mulps %xmm9,%xmm0 @@ -25040,7 +25202,7 @@ _sk_darken_sse2 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,95,209 ; maxps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,223,40,1,0 ; movaps 0x128df(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,45,36,1,0 ; movaps 0x1242d(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -25072,7 +25234,7 @@ _sk_lighten_sse2 LABEL PROC DB 68,15,89,206 ; mulps %xmm6,%xmm9 DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,116,40,1,0 ; movaps 0x12874(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,194,35,1,0 ; movaps 0x123c2(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -25107,7 +25269,7 @@ _sk_difference_sse2 LABEL PROC DB 65,15,93,209 ; minps %xmm9,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,254,39,1,0 ; movaps 0x127fe(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,76,35,1,0 ; movaps 0x1234c(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -25133,7 +25295,7 @@ _sk_exclusion_sse2 LABEL PROC DB 15,89,214 ; mulps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 68,15,92,194 ; subps %xmm2,%xmm8 - DB 15,40,21,174,39,1,0 ; movaps 0x127ae(%rip),%xmm2 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,21,252,34,1,0 ; movaps 0x122fc(%rip),%xmm2 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,215 ; mulps %xmm7,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -25144,7 +25306,7 @@ _sk_exclusion_sse2 LABEL PROC PUBLIC _sk_colorburn_sse2 _sk_colorburn_sse2 LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 - DB 68,15,40,13,145,39,1,0 ; movaps 0x12791(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,223,34,1,0 ; movaps 0x122df(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -25237,7 +25399,7 @@ _sk_colorburn_sse2 LABEL PROC PUBLIC _sk_colordodge_sse2 _sk_colordodge_sse2 LABEL PROC DB 68,15,40,200 ; movaps %xmm0,%xmm9 - DB 68,15,40,21,57,38,1,0 ; movaps 0x12639(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,135,33,1,0 ; movaps 0x12187(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,218 ; movaps %xmm10,%xmm11 DB 68,15,92,223 ; subps %xmm7,%xmm11 DB 65,15,40,195 ; movaps %xmm11,%xmm0 @@ -25328,7 +25490,7 @@ _sk_hardlight_sse2 LABEL PROC DB 15,41,52,36 ; movaps %xmm6,(%rsp) DB 15,40,245 ; movaps %xmm5,%xmm6 DB 15,40,236 ; movaps %xmm4,%xmm5 - DB 68,15,40,29,223,36,1,0 ; movaps 0x124df(%rip),%xmm11 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,29,45,32,1,0 ; movaps 0x1202d(%rip),%xmm11 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,211 ; movaps %xmm11,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,194 ; movaps %xmm10,%xmm8 @@ -25415,7 +25577,7 @@ PUBLIC _sk_overlay_sse2 _sk_overlay_sse2 LABEL PROC DB 68,15,40,193 ; movaps %xmm1,%xmm8 DB 68,15,40,232 ; movaps %xmm0,%xmm13 - DB 68,15,40,13,154,35,1,0 ; movaps 0x1239a(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,232,30,1,0 ; movaps 0x11ee8(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 68,15,92,215 ; subps %xmm7,%xmm10 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -25505,7 +25667,7 @@ _sk_softlight_sse2 LABEL PROC DB 68,15,40,213 ; movaps %xmm5,%xmm10 DB 68,15,94,215 ; divps %xmm7,%xmm10 DB 69,15,84,212 ; andps %xmm12,%xmm10 - DB 68,15,40,13,68,34,1,0 ; movaps 0x12244(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,146,29,1,0 ; movaps 0x11d92(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,249 ; movaps %xmm9,%xmm15 DB 69,15,92,250 ; subps %xmm10,%xmm15 DB 69,15,40,218 ; movaps %xmm10,%xmm11 @@ -25518,10 +25680,10 @@ _sk_softlight_sse2 LABEL PROC DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 DB 65,15,88,194 ; addps %xmm10,%xmm0 - DB 68,15,40,53,110,34,1,0 ; movaps 0x1226e(%rip),%xmm14 # 2d520 <_sk_overlay_sse2_8bit+0xfc8> + DB 68,15,40,53,188,29,1,0 ; movaps 0x11dbc(%rip),%xmm14 # 2d410 <_sk_overlay_sse2_8bit+0xf61> DB 69,15,88,222 ; addps %xmm14,%xmm11 DB 68,15,89,216 ; mulps %xmm0,%xmm11 - DB 68,15,40,21,110,34,1,0 ; movaps 0x1226e(%rip),%xmm10 # 2d530 <_sk_overlay_sse2_8bit+0xfd8> + DB 68,15,40,21,188,29,1,0 ; movaps 0x11dbc(%rip),%xmm10 # 2d420 <_sk_overlay_sse2_8bit+0xf71> DB 69,15,89,234 ; mulps %xmm10,%xmm13 DB 69,15,88,235 ; addps %xmm11,%xmm13 DB 15,88,228 ; addps %xmm4,%xmm4 @@ -25666,7 +25828,7 @@ _sk_hue_sse2 LABEL PROC DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 68,15,89,211 ; mulps %xmm3,%xmm10 - DB 68,15,40,5,250,31,1,0 ; movaps 0x11ffa(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,72,27,1,0 ; movaps 0x11b48(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,216 ; movaps %xmm8,%xmm11 DB 15,40,207 ; movaps %xmm7,%xmm1 DB 68,15,92,217 ; subps %xmm1,%xmm11 @@ -25714,12 +25876,12 @@ _sk_hue_sse2 LABEL PROC DB 69,15,84,206 ; andps %xmm14,%xmm9 DB 69,15,84,214 ; andps %xmm14,%xmm10 DB 65,15,84,214 ; andps %xmm14,%xmm2 - DB 68,15,40,61,183,31,1,0 ; movaps 0x11fb7(%rip),%xmm15 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,61,5,27,1,0 ; movaps 0x11b05(%rip),%xmm15 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 65,15,89,231 ; mulps %xmm15,%xmm4 - DB 15,40,5,188,31,1,0 ; movaps 0x11fbc(%rip),%xmm0 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 15,40,5,10,27,1,0 ; movaps 0x11b0a(%rip),%xmm0 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 15,89,240 ; mulps %xmm0,%xmm6 DB 15,88,244 ; addps %xmm4,%xmm6 - DB 68,15,40,53,190,31,1,0 ; movaps 0x11fbe(%rip),%xmm14 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,53,12,27,1,0 ; movaps 0x11b0c(%rip),%xmm14 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 68,15,40,239 ; movaps %xmm7,%xmm13 DB 69,15,89,238 ; mulps %xmm14,%xmm13 DB 68,15,88,238 ; addps %xmm6,%xmm13 @@ -25897,14 +26059,14 @@ _sk_saturation_sse2 LABEL PROC DB 68,15,84,211 ; andps %xmm3,%xmm10 DB 68,15,84,203 ; andps %xmm3,%xmm9 DB 15,84,195 ; andps %xmm3,%xmm0 - DB 68,15,40,5,11,29,1,0 ; movaps 0x11d0b(%rip),%xmm8 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,5,89,24,1,0 ; movaps 0x11859(%rip),%xmm8 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 15,40,214 ; movaps %xmm6,%xmm2 DB 65,15,89,208 ; mulps %xmm8,%xmm2 - DB 15,40,13,13,29,1,0 ; movaps 0x11d0d(%rip),%xmm1 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 15,40,13,91,24,1,0 ; movaps 0x1185b(%rip),%xmm1 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 15,40,221 ; movaps %xmm5,%xmm3 DB 15,89,217 ; mulps %xmm1,%xmm3 DB 15,88,218 ; addps %xmm2,%xmm3 - DB 68,15,40,37,12,29,1,0 ; movaps 0x11d0c(%rip),%xmm12 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,37,90,24,1,0 ; movaps 0x1185a(%rip),%xmm12 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 69,15,89,236 ; mulps %xmm12,%xmm13 DB 68,15,88,235 ; addps %xmm3,%xmm13 DB 65,15,40,210 ; movaps %xmm10,%xmm2 @@ -25949,7 +26111,7 @@ _sk_saturation_sse2 LABEL PROC DB 15,40,223 ; movaps %xmm7,%xmm3 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,221 ; mulps %xmm5,%xmm3 - DB 68,15,40,5,193,27,1,0 ; movaps 0x11bc1(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,15,23,1,0 ; movaps 0x1170f(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,40,224 ; movaps %xmm8,%xmm4 DB 68,15,92,199 ; subps %xmm7,%xmm8 DB 15,88,253 ; addps %xmm5,%xmm7 @@ -26050,14 +26212,14 @@ _sk_color_sse2 LABEL PROC DB 68,15,40,213 ; movaps %xmm5,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 DB 65,15,40,208 ; movaps %xmm8,%xmm2 - DB 68,15,40,45,201,26,1,0 ; movaps 0x11ac9(%rip),%xmm13 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,45,23,22,1,0 ; movaps 0x11617(%rip),%xmm13 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 68,15,40,198 ; movaps %xmm6,%xmm8 DB 69,15,89,197 ; mulps %xmm13,%xmm8 - DB 68,15,40,53,201,26,1,0 ; movaps 0x11ac9(%rip),%xmm14 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 68,15,40,53,23,22,1,0 ; movaps 0x11617(%rip),%xmm14 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 65,15,40,195 ; movaps %xmm11,%xmm0 DB 65,15,89,198 ; mulps %xmm14,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,29,197,26,1,0 ; movaps 0x11ac5(%rip),%xmm11 # 2d560 <_sk_overlay_sse2_8bit+0x1008> + DB 68,15,40,29,19,22,1,0 ; movaps 0x11613(%rip),%xmm11 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> DB 69,15,89,227 ; mulps %xmm11,%xmm12 DB 68,15,88,224 ; addps %xmm0,%xmm12 DB 65,15,40,193 ; movaps %xmm9,%xmm0 @@ -26065,7 +26227,7 @@ _sk_color_sse2 LABEL PROC DB 69,15,40,250 ; movaps %xmm10,%xmm15 DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 68,15,88,248 ; addps %xmm0,%xmm15 - DB 68,15,40,5,1,26,1,0 ; movaps 0x11a01(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,79,21,1,0 ; movaps 0x1154f(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,40,224 ; movaps %xmm8,%xmm4 DB 15,92,226 ; subps %xmm2,%xmm4 DB 15,89,252 ; mulps %xmm4,%xmm7 @@ -26201,15 +26363,15 @@ _sk_luminosity_sse2 LABEL PROC DB 68,15,40,205 ; movaps %xmm5,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 DB 15,89,222 ; mulps %xmm6,%xmm3 - DB 68,15,40,37,131,24,1,0 ; movaps 0x11883(%rip),%xmm12 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,37,209,19,1,0 ; movaps 0x113d1(%rip),%xmm12 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 68,15,40,199 ; movaps %xmm7,%xmm8 DB 69,15,89,196 ; mulps %xmm12,%xmm8 - DB 68,15,40,45,131,24,1,0 ; movaps 0x11883(%rip),%xmm13 # 2d550 <_sk_overlay_sse2_8bit+0xff8> + DB 68,15,40,45,209,19,1,0 ; movaps 0x113d1(%rip),%xmm13 # 2d440 <_sk_overlay_sse2_8bit+0xf91> DB 68,15,40,241 ; movaps %xmm1,%xmm14 DB 69,15,89,245 ; mulps %xmm13,%xmm14 DB 69,15,88,240 ; addps %xmm8,%xmm14 - DB 68,15,40,29,127,24,1,0 ; movaps 0x1187f(%rip),%xmm11 # 2d560 <_sk_overlay_sse2_8bit+0x1008> - DB 68,15,40,5,215,23,1,0 ; movaps 0x117d7(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,29,205,19,1,0 ; movaps 0x113cd(%rip),%xmm11 # 2d450 <_sk_overlay_sse2_8bit+0xfa1> + DB 68,15,40,5,37,19,1,0 ; movaps 0x11325(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,248 ; movaps %xmm8,%xmm15 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 68,15,92,248 ; subps %xmm0,%xmm15 @@ -26349,10 +26511,10 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,228,0,0,0 ; jne 1bfd6 <_sk_srcover_rgba_8888_sse2+0x104> + DB 15,133,228,0,0,0 ; jne 1c378 <_sk_srcover_rgba_8888_sse2+0x104> DB 243,69,15,111,4,128 ; movdqu (%r8,%rax,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 102,15,111,53,109,22,1,0 ; movdqa 0x1166d(%rip),%xmm6 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,53,187,17,1,0 ; movdqa 0x111bb(%rip),%xmm6 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,224 ; movdqa %xmm8,%xmm4 DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 @@ -26366,9 +26528,9 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 15,91,247 ; cvtdq2ps %xmm7,%xmm6 DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8 DB 65,15,91,248 ; cvtdq2ps %xmm8,%xmm7 - DB 68,15,40,5,125,21,1,0 ; movaps 0x1157d(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,203,16,1,0 ; movaps 0x110cb(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,92,195 ; subps %xmm3,%xmm8 - DB 68,15,40,37,81,22,1,0 ; movaps 0x11651(%rip),%xmm12 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,37,159,17,1,0 ; movaps 0x1119f(%rip),%xmm12 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 65,15,89,196 ; mulps %xmm12,%xmm0 DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,89,204 ; mulps %xmm4,%xmm9 @@ -26394,7 +26556,7 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 102,15,114,240,24 ; pslld $0x18,%xmm0 DB 102,15,235,194 ; por %xmm2,%xmm0 DB 102,15,235,193 ; por %xmm1,%xmm0 - DB 117,84 ; jne 1c010 <_sk_srcover_rgba_8888_sse2+0x13e> + DB 117,84 ; jne 1c3b2 <_sk_srcover_rgba_8888_sse2+0x13e> DB 243,65,15,127,4,128 ; movdqu %xmm0,(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,193 ; movaps %xmm9,%xmm0 @@ -26405,32 +26567,32 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,90 ; je 1c03d <_sk_srcover_rgba_8888_sse2+0x16b> + DB 116,90 ; je 1c3df <_sk_srcover_rgba_8888_sse2+0x16b> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 1c005 <_sk_srcover_rgba_8888_sse2+0x133> + DB 116,23 ; je 1c3a7 <_sk_srcover_rgba_8888_sse2+0x133> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,0,255,255,255 ; jne 1bef8 <_sk_srcover_rgba_8888_sse2+0x26> + DB 15,133,0,255,255,255 ; jne 1c29a <_sk_srcover_rgba_8888_sse2+0x26> DB 102,65,15,110,100,128,8 ; movd 0x8(%r8,%rax,4),%xmm4 DB 102,68,15,112,196,69 ; pshufd $0x45,%xmm4,%xmm8 DB 102,69,15,18,4,128 ; movlpd (%r8,%rax,4),%xmm8 - DB 233,232,254,255,255 ; jmpq 1bef8 <_sk_srcover_rgba_8888_sse2+0x26> + DB 233,232,254,255,255 ; jmpq 1c29a <_sk_srcover_rgba_8888_sse2+0x26> DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,43 ; je 1c048 <_sk_srcover_rgba_8888_sse2+0x176> + DB 116,43 ; je 1c3ea <_sk_srcover_rgba_8888_sse2+0x176> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 1c035 <_sk_srcover_rgba_8888_sse2+0x163> + DB 116,18 ; je 1c3d7 <_sk_srcover_rgba_8888_sse2+0x163> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,153 ; jne 1bfc2 <_sk_srcover_rgba_8888_sse2+0xf0> + DB 117,153 ; jne 1c364 <_sk_srcover_rgba_8888_sse2+0xf0> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,65,15,126,76,128,8 ; movd %xmm1,0x8(%r8,%rax,4) DB 102,65,15,214,4,128 ; movq %xmm0,(%r8,%rax,4) - DB 235,133 ; jmp 1bfc2 <_sk_srcover_rgba_8888_sse2+0xf0> + DB 235,133 ; jmp 1c364 <_sk_srcover_rgba_8888_sse2+0xf0> DB 102,69,15,110,4,128 ; movd (%r8,%rax,4),%xmm8 - DB 233,176,254,255,255 ; jmpq 1bef8 <_sk_srcover_rgba_8888_sse2+0x26> + DB 233,176,254,255,255 ; jmpq 1c29a <_sk_srcover_rgba_8888_sse2+0x26> DB 102,65,15,126,4,128 ; movd %xmm0,(%r8,%rax,4) - DB 233,111,255,255,255 ; jmpq 1bfc2 <_sk_srcover_rgba_8888_sse2+0xf0> + DB 233,111,255,255,255 ; jmpq 1c364 <_sk_srcover_rgba_8888_sse2+0xf0> PUBLIC _sk_clamp_0_sse2 _sk_clamp_0_sse2 LABEL PROC @@ -26444,7 +26606,7 @@ _sk_clamp_0_sse2 LABEL PROC PUBLIC _sk_clamp_1_sse2 _sk_clamp_1_sse2 LABEL PROC - DB 68,15,40,5,77,20,1,0 ; movaps 0x1144d(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,155,15,1,0 ; movaps 0x10f9b(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,93,192 ; minps %xmm8,%xmm0 DB 65,15,93,200 ; minps %xmm8,%xmm1 DB 65,15,93,208 ; minps %xmm8,%xmm2 @@ -26454,7 +26616,7 @@ _sk_clamp_1_sse2 LABEL PROC PUBLIC _sk_clamp_a_sse2 _sk_clamp_a_sse2 LABEL PROC - DB 15,93,29,50,20,1,0 ; minps 0x11432(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,29,128,15,1,0 ; minps 0x10f80(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -26463,7 +26625,7 @@ _sk_clamp_a_sse2 LABEL PROC PUBLIC _sk_clamp_a_dst_sse2 _sk_clamp_a_dst_sse2 LABEL PROC - DB 15,93,61,30,20,1,0 ; minps 0x1141e(%rip),%xmm7 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,93,61,108,15,1,0 ; minps 0x10f6c(%rip),%xmm7 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,93,231 ; minps %xmm7,%xmm4 DB 15,93,239 ; minps %xmm7,%xmm5 DB 15,93,247 ; minps %xmm7,%xmm6 @@ -26492,7 +26654,7 @@ _sk_swap_rb_sse2 LABEL PROC PUBLIC _sk_invert_sse2 _sk_invert_sse2 LABEL PROC - DB 68,15,40,5,218,19,1,0 ; movaps 0x113da(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,5,40,15,1,0 ; movaps 0x10f28(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 68,15,92,200 ; subps %xmm0,%xmm9 DB 69,15,40,208 ; movaps %xmm8,%xmm10 @@ -26544,7 +26706,7 @@ _sk_premul_dst_sse2 LABEL PROC PUBLIC _sk_unpremul_sse2 _sk_unpremul_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 - DB 68,15,40,13,100,19,1,0 ; movaps 0x11364(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,178,14,1,0 ; movaps 0x10eb2(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,94,203 ; divps %xmm3,%xmm9 DB 68,15,194,195,4 ; cmpneqps %xmm3,%xmm8 DB 69,15,84,193 ; andps %xmm9,%xmm8 @@ -26556,20 +26718,20 @@ _sk_unpremul_sse2 LABEL PROC PUBLIC _sk_from_srgb_sse2 _sk_from_srgb_sse2 LABEL PROC - DB 68,15,40,5,47,20,1,0 ; movaps 0x1142f(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1058> + DB 68,15,40,5,125,15,1,0 ; movaps 0x10f7d(%rip),%xmm8 # 2d4a0 <_sk_overlay_sse2_8bit+0xff1> DB 68,15,40,232 ; movaps %xmm0,%xmm13 DB 69,15,89,232 ; mulps %xmm8,%xmm13 DB 68,15,40,216 ; movaps %xmm0,%xmm11 DB 69,15,89,219 ; mulps %xmm11,%xmm11 - DB 68,15,40,13,167,19,1,0 ; movaps 0x113a7(%rip),%xmm9 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,13,245,14,1,0 ; movaps 0x10ef5(%rip),%xmm9 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,241 ; mulps %xmm9,%xmm14 - DB 68,15,40,21,23,20,1,0 ; movaps 0x11417(%rip),%xmm10 # 2d5c0 <_sk_overlay_sse2_8bit+0x1068> + DB 68,15,40,21,101,15,1,0 ; movaps 0x10f65(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0x1001> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 69,15,89,243 ; mulps %xmm11,%xmm14 - DB 68,15,40,29,23,20,1,0 ; movaps 0x11417(%rip),%xmm11 # 2d5d0 <_sk_overlay_sse2_8bit+0x1078> + DB 68,15,40,29,101,15,1,0 ; movaps 0x10f65(%rip),%xmm11 # 2d4c0 <_sk_overlay_sse2_8bit+0x1011> DB 69,15,88,243 ; addps %xmm11,%xmm14 - DB 68,15,40,37,27,20,1,0 ; movaps 0x1141b(%rip),%xmm12 # 2d5e0 <_sk_overlay_sse2_8bit+0x1088> + DB 68,15,40,37,105,15,1,0 ; movaps 0x10f69(%rip),%xmm12 # 2d4d0 <_sk_overlay_sse2_8bit+0x1021> DB 65,15,194,196,1 ; cmpltps %xmm12,%xmm0 DB 68,15,84,232 ; andps %xmm0,%xmm13 DB 65,15,85,198 ; andnps %xmm14,%xmm0 @@ -26603,20 +26765,20 @@ _sk_from_srgb_sse2 LABEL PROC PUBLIC _sk_from_srgb_dst_sse2 _sk_from_srgb_dst_sse2 LABEL PROC - DB 68,15,40,5,108,19,1,0 ; movaps 0x1136c(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1058> + DB 68,15,40,5,186,14,1,0 ; movaps 0x10eba(%rip),%xmm8 # 2d4a0 <_sk_overlay_sse2_8bit+0xff1> DB 68,15,40,236 ; movaps %xmm4,%xmm13 DB 69,15,89,232 ; mulps %xmm8,%xmm13 DB 68,15,40,220 ; movaps %xmm4,%xmm11 DB 69,15,89,219 ; mulps %xmm11,%xmm11 - DB 68,15,40,13,228,18,1,0 ; movaps 0x112e4(%rip),%xmm9 # 2d540 <_sk_overlay_sse2_8bit+0xfe8> + DB 68,15,40,13,50,14,1,0 ; movaps 0x10e32(%rip),%xmm9 # 2d430 <_sk_overlay_sse2_8bit+0xf81> DB 68,15,40,244 ; movaps %xmm4,%xmm14 DB 69,15,89,241 ; mulps %xmm9,%xmm14 - DB 68,15,40,21,84,19,1,0 ; movaps 0x11354(%rip),%xmm10 # 2d5c0 <_sk_overlay_sse2_8bit+0x1068> + DB 68,15,40,21,162,14,1,0 ; movaps 0x10ea2(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0x1001> DB 69,15,88,242 ; addps %xmm10,%xmm14 DB 69,15,89,243 ; mulps %xmm11,%xmm14 - DB 68,15,40,29,84,19,1,0 ; movaps 0x11354(%rip),%xmm11 # 2d5d0 <_sk_overlay_sse2_8bit+0x1078> + DB 68,15,40,29,162,14,1,0 ; movaps 0x10ea2(%rip),%xmm11 # 2d4c0 <_sk_overlay_sse2_8bit+0x1011> DB 69,15,88,243 ; addps %xmm11,%xmm14 - DB 68,15,40,37,88,19,1,0 ; movaps 0x11358(%rip),%xmm12 # 2d5e0 <_sk_overlay_sse2_8bit+0x1088> + DB 68,15,40,37,166,14,1,0 ; movaps 0x10ea6(%rip),%xmm12 # 2d4d0 <_sk_overlay_sse2_8bit+0x1021> DB 65,15,194,228,1 ; cmpltps %xmm12,%xmm4 DB 68,15,84,236 ; andps %xmm4,%xmm13 DB 65,15,85,230 ; andnps %xmm14,%xmm4 @@ -26651,22 +26813,22 @@ _sk_from_srgb_dst_sse2 LABEL PROC PUBLIC _sk_to_srgb_sse2 _sk_to_srgb_sse2 LABEL PROC DB 68,15,82,232 ; rsqrtps %xmm0,%xmm13 - DB 68,15,40,5,229,18,1,0 ; movaps 0x112e5(%rip),%xmm8 # 2d5f0 <_sk_overlay_sse2_8bit+0x1098> + DB 68,15,40,5,51,14,1,0 ; movaps 0x10e33(%rip),%xmm8 # 2d4e0 <_sk_overlay_sse2_8bit+0x1031> DB 68,15,40,240 ; movaps %xmm0,%xmm14 DB 69,15,89,240 ; mulps %xmm8,%xmm14 - DB 68,15,40,13,229,18,1,0 ; movaps 0x112e5(%rip),%xmm9 # 2d600 <_sk_overlay_sse2_8bit+0x10a8> + DB 68,15,40,13,51,14,1,0 ; movaps 0x10e33(%rip),%xmm9 # 2d4f0 <_sk_overlay_sse2_8bit+0x1041> DB 69,15,40,253 ; movaps %xmm13,%xmm15 DB 69,15,89,249 ; mulps %xmm9,%xmm15 - DB 68,15,40,21,229,18,1,0 ; movaps 0x112e5(%rip),%xmm10 # 2d610 <_sk_overlay_sse2_8bit+0x10b8> + DB 68,15,40,21,51,14,1,0 ; movaps 0x10e33(%rip),%xmm10 # 2d500 <_sk_overlay_sse2_8bit+0x1051> DB 69,15,88,250 ; addps %xmm10,%xmm15 DB 69,15,89,253 ; mulps %xmm13,%xmm15 - DB 68,15,40,29,229,18,1,0 ; movaps 0x112e5(%rip),%xmm11 # 2d620 <_sk_overlay_sse2_8bit+0x10c8> + DB 68,15,40,29,51,14,1,0 ; movaps 0x10e33(%rip),%xmm11 # 2d510 <_sk_overlay_sse2_8bit+0x1061> DB 69,15,88,251 ; addps %xmm11,%xmm15 - DB 68,15,40,37,233,18,1,0 ; movaps 0x112e9(%rip),%xmm12 # 2d630 <_sk_overlay_sse2_8bit+0x10d8> + DB 68,15,40,37,55,14,1,0 ; movaps 0x10e37(%rip),%xmm12 # 2d520 <_sk_overlay_sse2_8bit+0x1071> DB 69,15,88,236 ; addps %xmm12,%xmm13 DB 69,15,83,237 ; rcpps %xmm13,%xmm13 DB 69,15,89,239 ; mulps %xmm15,%xmm13 - DB 68,15,40,61,229,18,1,0 ; movaps 0x112e5(%rip),%xmm15 # 2d640 <_sk_overlay_sse2_8bit+0x10e8> + DB 68,15,40,61,51,14,1,0 ; movaps 0x10e33(%rip),%xmm15 # 2d530 <_sk_overlay_sse2_8bit+0x1081> DB 65,15,194,199,1 ; cmpltps %xmm15,%xmm0 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,197 ; andnps %xmm13,%xmm0 @@ -26716,7 +26878,7 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 68,15,93,216 ; minps %xmm0,%xmm11 DB 65,15,40,202 ; movaps %xmm10,%xmm1 DB 65,15,92,203 ; subps %xmm11,%xmm1 - DB 68,15,40,45,168,16,1,0 ; movaps 0x110a8(%rip),%xmm13 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,45,246,11,1,0 ; movaps 0x10bf6(%rip),%xmm13 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 68,15,94,233 ; divps %xmm1,%xmm13 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,194,192,0 ; cmpeqps %xmm8,%xmm0 @@ -26725,30 +26887,30 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,40,241 ; movaps %xmm9,%xmm14 DB 68,15,194,242,1 ; cmpltps %xmm2,%xmm14 - DB 68,15,84,53,14,18,1,0 ; andps 0x1120e(%rip),%xmm14 # 2d650 <_sk_overlay_sse2_8bit+0x10f8> + DB 68,15,84,53,92,13,1,0 ; andps 0x10d5c(%rip),%xmm14 # 2d540 <_sk_overlay_sse2_8bit+0x1091> DB 69,15,88,244 ; addps %xmm12,%xmm14 DB 69,15,40,250 ; movaps %xmm10,%xmm15 DB 69,15,194,249,0 ; cmpeqps %xmm9,%xmm15 DB 65,15,92,208 ; subps %xmm8,%xmm2 DB 65,15,89,213 ; mulps %xmm13,%xmm2 - DB 68,15,40,37,1,18,1,0 ; movaps 0x11201(%rip),%xmm12 # 2d660 <_sk_overlay_sse2_8bit+0x1108> + DB 68,15,40,37,79,13,1,0 ; movaps 0x10d4f(%rip),%xmm12 # 2d550 <_sk_overlay_sse2_8bit+0x10a1> DB 65,15,88,212 ; addps %xmm12,%xmm2 DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 69,15,89,197 ; mulps %xmm13,%xmm8 - DB 68,15,88,5,253,17,1,0 ; addps 0x111fd(%rip),%xmm8 # 2d670 <_sk_overlay_sse2_8bit+0x1118> + DB 68,15,88,5,75,13,1,0 ; addps 0x10d4b(%rip),%xmm8 # 2d560 <_sk_overlay_sse2_8bit+0x10b1> DB 65,15,84,215 ; andps %xmm15,%xmm2 DB 69,15,85,248 ; andnps %xmm8,%xmm15 DB 68,15,86,250 ; orps %xmm2,%xmm15 DB 68,15,84,240 ; andps %xmm0,%xmm14 DB 65,15,85,199 ; andnps %xmm15,%xmm0 DB 65,15,86,198 ; orps %xmm14,%xmm0 - DB 15,89,5,238,17,1,0 ; mulps 0x111ee(%rip),%xmm0 # 2d680 <_sk_overlay_sse2_8bit+0x1128> + DB 15,89,5,60,13,1,0 ; mulps 0x10d3c(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x10c1> DB 69,15,40,194 ; movaps %xmm10,%xmm8 DB 69,15,194,195,4 ; cmpneqps %xmm11,%xmm8 DB 65,15,84,192 ; andps %xmm8,%xmm0 DB 69,15,92,226 ; subps %xmm10,%xmm12 DB 69,15,88,211 ; addps %xmm11,%xmm10 - DB 68,15,40,13,1,16,1,0 ; movaps 0x11001(%rip),%xmm9 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,13,79,11,1,0 ; movaps 0x10b4f(%rip),%xmm9 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,40,210 ; movaps %xmm10,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 68,15,194,202,1 ; cmpltps %xmm2,%xmm9 @@ -26771,7 +26933,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,41,92,36,32 ; movaps %xmm3,0x20(%rsp) DB 68,15,40,218 ; movaps %xmm2,%xmm11 DB 15,40,240 ; movaps %xmm0,%xmm6 - DB 68,15,40,13,172,15,1,0 ; movaps 0x10fac(%rip),%xmm9 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,13,250,10,1,0 ; movaps 0x10afa(%rip),%xmm9 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,40,209 ; movaps %xmm9,%xmm10 DB 69,15,194,211,2 ; cmpleps %xmm11,%xmm10 DB 15,40,193 ; movaps %xmm1,%xmm0 @@ -26788,28 +26950,28 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 69,15,88,211 ; addps %xmm11,%xmm10 DB 69,15,88,219 ; addps %xmm11,%xmm11 DB 69,15,92,218 ; subps %xmm10,%xmm11 - DB 15,40,5,70,17,1,0 ; movaps 0x11146(%rip),%xmm0 # 2d690 <_sk_overlay_sse2_8bit+0x1138> + DB 15,40,5,148,12,1,0 ; movaps 0x10c94(%rip),%xmm0 # 2d580 <_sk_overlay_sse2_8bit+0x10d1> DB 15,88,198 ; addps %xmm6,%xmm0 DB 243,15,91,200 ; cvttps2dq %xmm0,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,40,216 ; movaps %xmm0,%xmm3 DB 15,194,217,1 ; cmpltps %xmm1,%xmm3 - DB 15,84,29,94,15,1,0 ; andps 0x10f5e(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,84,29,172,10,1,0 ; andps 0x10aac(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,92,193 ; subps %xmm1,%xmm0 - DB 68,15,40,45,48,17,1,0 ; movaps 0x11130(%rip),%xmm13 # 2d6a0 <_sk_overlay_sse2_8bit+0x1148> + DB 68,15,40,45,126,12,1,0 ; movaps 0x10c7e(%rip),%xmm13 # 2d590 <_sk_overlay_sse2_8bit+0x10e1> DB 69,15,40,197 ; movaps %xmm13,%xmm8 DB 68,15,194,192,2 ; cmpleps %xmm0,%xmm8 DB 69,15,40,242 ; movaps %xmm10,%xmm14 DB 69,15,92,243 ; subps %xmm11,%xmm14 DB 65,15,40,217 ; movaps %xmm9,%xmm3 DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 - DB 15,40,21,240,16,1,0 ; movaps 0x110f0(%rip),%xmm2 # 2d680 <_sk_overlay_sse2_8bit+0x1128> + DB 15,40,21,62,12,1,0 ; movaps 0x10c3e(%rip),%xmm2 # 2d570 <_sk_overlay_sse2_8bit+0x10c1> DB 68,15,40,250 ; movaps %xmm2,%xmm15 DB 68,15,194,248,2 ; cmpleps %xmm0,%xmm15 - DB 15,40,13,176,16,1,0 ; movaps 0x110b0(%rip),%xmm1 # 2d650 <_sk_overlay_sse2_8bit+0x10f8> + DB 15,40,13,254,11,1,0 ; movaps 0x10bfe(%rip),%xmm1 # 2d540 <_sk_overlay_sse2_8bit+0x1091> DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,40,45,198,16,1,0 ; movaps 0x110c6(%rip),%xmm5 # 2d670 <_sk_overlay_sse2_8bit+0x1118> + DB 15,40,45,20,12,1,0 ; movaps 0x10c14(%rip),%xmm5 # 2d560 <_sk_overlay_sse2_8bit+0x10b1> DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,92,224 ; subps %xmm0,%xmm4 DB 65,15,89,230 ; mulps %xmm14,%xmm4 @@ -26832,7 +26994,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,222 ; movaps %xmm6,%xmm3 DB 15,194,216,1 ; cmpltps %xmm0,%xmm3 - DB 15,84,29,187,14,1,0 ; andps 0x10ebb(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,84,29,9,10,1,0 ; andps 0x10a09(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,195 ; subps %xmm3,%xmm0 DB 68,15,40,230 ; movaps %xmm6,%xmm12 DB 68,15,92,224 ; subps %xmm0,%xmm12 @@ -26862,12 +27024,12 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,40,60,36 ; movaps (%rsp),%xmm7 DB 15,40,231 ; movaps %xmm7,%xmm4 DB 15,85,227 ; andnps %xmm3,%xmm4 - DB 15,88,53,52,16,1,0 ; addps 0x11034(%rip),%xmm6 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 15,88,53,130,11,1,0 ; addps 0x10b82(%rip),%xmm6 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 243,15,91,198 ; cvttps2dq %xmm6,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,222 ; movaps %xmm6,%xmm3 DB 15,194,216,1 ; cmpltps %xmm0,%xmm3 - DB 15,84,29,47,14,1,0 ; andps 0x10e2f(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,84,29,125,9,1,0 ; andps 0x1097d(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,92,195 ; subps %xmm3,%xmm0 DB 15,92,240 ; subps %xmm0,%xmm6 DB 15,89,206 ; mulps %xmm6,%xmm1 @@ -26928,13 +27090,13 @@ _sk_scale_u8_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,61 ; jne 1c796 <_sk_scale_u8_sse2+0x55> + DB 117,61 ; jne 1cb38 <_sk_scale_u8_sse2+0x55> DB 102,71,15,110,4,16 ; movd (%r8,%r10,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,219,5,254,13,1,0 ; pand 0x10dfe(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,68,15,219,5,76,9,1,0 ; pand 0x1094c(%rip),%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,66,15,1,0 ; mulps 0x10f42(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,89,5,144,10,1,0 ; mulps 0x10a90(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 @@ -26945,12 +27107,12 @@ _sk_scale_u8_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,61 ; je 1c7e0 <_sk_scale_u8_sse2+0x9f> + DB 116,61 ; je 1cb82 <_sk_scale_u8_sse2+0x9f> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 1c7c5 <_sk_scale_u8_sse2+0x84> + DB 116,23 ; je 1cb67 <_sk_scale_u8_sse2+0x84> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,181 ; jne 1c769 <_sk_scale_u8_sse2+0x28> + DB 117,181 ; jne 1cb0b <_sk_scale_u8_sse2+0x28> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 @@ -26959,10 +27121,10 @@ _sk_scale_u8_sse2 LABEL PROC DB 102,68,15,96,200 ; punpcklbw %xmm0,%xmm9 DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 DB 242,69,15,16,193 ; movsd %xmm9,%xmm8 - DB 235,137 ; jmp 1c769 <_sk_scale_u8_sse2+0x28> + DB 235,137 ; jmp 1cb0b <_sk_scale_u8_sse2+0x28> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 - DB 233,122,255,255,255 ; jmpq 1c769 <_sk_scale_u8_sse2+0x28> + DB 233,122,255,255,255 ; jmpq 1cb0b <_sk_scale_u8_sse2+0x28> PUBLIC _sk_scale_565_sse2 _sk_scale_565_sse2 LABEL PROC @@ -26974,20 +27136,20 @@ _sk_scale_565_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,173,0,0,0 ; jne 1c8bb <_sk_scale_565_sse2+0xcc> + DB 15,133,173,0,0,0 ; jne 1cc5d <_sk_scale_565_sse2+0xcc> DB 243,71,15,126,20,80 ; movq (%r8,%r10,2),%xmm10 DB 102,68,15,97,208 ; punpcklwd %xmm0,%xmm10 - DB 102,68,15,111,5,174,14,1,0 ; movdqa 0x10eae(%rip),%xmm8 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,68,15,111,5,252,9,1,0 ; movdqa 0x109fc(%rip),%xmm8 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,69,15,219,194 ; pand %xmm10,%xmm8 DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,173,14,1,0 ; mulps 0x10ead(%rip),%xmm8 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,68,15,111,13,180,14,1,0 ; movdqa 0x10eb4(%rip),%xmm9 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 68,15,89,5,251,9,1,0 ; mulps 0x109fb(%rip),%xmm8 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,68,15,111,13,2,10,1,0 ; movdqa 0x10a02(%rip),%xmm9 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,69,15,219,202 ; pand %xmm10,%xmm9 DB 69,15,91,201 ; cvtdq2ps %xmm9,%xmm9 - DB 68,15,89,13,179,14,1,0 ; mulps 0x10eb3(%rip),%xmm9 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,68,15,219,21,186,14,1,0 ; pand 0x10eba(%rip),%xmm10 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 68,15,89,13,1,10,1,0 ; mulps 0x10a01(%rip),%xmm9 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,68,15,219,21,8,10,1,0 ; pand 0x10a08(%rip),%xmm10 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 - DB 68,15,89,21,190,14,1,0 ; mulps 0x10ebe(%rip),%xmm10 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 68,15,89,21,12,10,1,0 ; mulps 0x10a0c(%rip),%xmm10 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 68,15,40,219 ; movaps %xmm3,%xmm11 DB 68,15,194,223,1 ; cmpltps %xmm7,%xmm11 DB 69,15,40,225 ; movaps %xmm9,%xmm12 @@ -27014,22 +27176,22 @@ _sk_scale_565_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,59 ; je 1c903 <_sk_scale_565_sse2+0x114> + DB 116,60 ; je 1cca6 <_sk_scale_565_sse2+0x115> DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,27 ; je 1c8ee <_sk_scale_565_sse2+0xff> + DB 116,27 ; je 1cc90 <_sk_scale_565_sse2+0xff> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,60,255,255,255 ; jne 1c819 <_sk_scale_565_sse2+0x2a> + DB 15,133,60,255,255,255 ; jne 1cbbb <_sk_scale_565_sse2+0x2a> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,208,69 ; pshufd $0x45,%xmm8,%xmm10 DB 102,71,15,110,4,80 ; movd (%r8,%r10,2),%xmm8 - DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 + DB 242,69,15,112,192,212 ; pshuflw $0xd4,%xmm8,%xmm8 DB 242,69,15,16,208 ; movsd %xmm8,%xmm10 - DB 233,22,255,255,255 ; jmpq 1c819 <_sk_scale_565_sse2+0x2a> + DB 233,21,255,255,255 ; jmpq 1cbbb <_sk_scale_565_sse2+0x2a> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,68,15,110,208 ; movd %eax,%xmm10 - DB 233,7,255,255,255 ; jmpq 1c819 <_sk_scale_565_sse2+0x2a> + DB 233,6,255,255,255 ; jmpq 1cbbb <_sk_scale_565_sse2+0x2a> PUBLIC _sk_lerp_1_float_sse2 _sk_lerp_1_float_sse2 LABEL PROC @@ -27060,13 +27222,13 @@ _sk_lerp_u8_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,81 ; jne 1c9b3 <_sk_lerp_u8_sse2+0x69> + DB 117,81 ; jne 1cd56 <_sk_lerp_u8_sse2+0x69> DB 102,71,15,110,4,16 ; movd (%r8,%r10,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,219,5,245,11,1,0 ; pand 0x10bf5(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,68,15,219,5,66,7,1,0 ; pand 0x10742(%rip),%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,57,13,1,0 ; mulps 0x10d39(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,89,5,134,8,1,0 ; mulps 0x10886(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 15,92,196 ; subps %xmm4,%xmm0 DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 @@ -27084,12 +27246,12 @@ _sk_lerp_u8_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,64 ; je 1ca00 <_sk_lerp_u8_sse2+0xb6> + DB 116,64 ; je 1cda3 <_sk_lerp_u8_sse2+0xb6> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 1c9e2 <_sk_lerp_u8_sse2+0x98> + DB 116,23 ; je 1cd85 <_sk_lerp_u8_sse2+0x98> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,161 ; jne 1c972 <_sk_lerp_u8_sse2+0x28> + DB 117,161 ; jne 1cd15 <_sk_lerp_u8_sse2+0x28> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 @@ -27098,10 +27260,10 @@ _sk_lerp_u8_sse2 LABEL PROC DB 102,68,15,96,200 ; punpcklbw %xmm0,%xmm9 DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 DB 242,69,15,16,193 ; movsd %xmm9,%xmm8 - DB 233,114,255,255,255 ; jmpq 1c972 <_sk_lerp_u8_sse2+0x28> + DB 233,114,255,255,255 ; jmpq 1cd15 <_sk_lerp_u8_sse2+0x28> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 - DB 233,99,255,255,255 ; jmpq 1c972 <_sk_lerp_u8_sse2+0x28> + DB 233,99,255,255,255 ; jmpq 1cd15 <_sk_lerp_u8_sse2+0x28> PUBLIC _sk_lerp_565_sse2 _sk_lerp_565_sse2 LABEL PROC @@ -27113,20 +27275,20 @@ _sk_lerp_565_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,173,0,0,0 ; jne 1cadb <_sk_lerp_565_sse2+0xcc> + DB 15,133,173,0,0,0 ; jne 1ce7e <_sk_lerp_565_sse2+0xcc> DB 243,71,15,126,4,80 ; movq (%r8,%r10,2),%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 - DB 102,68,15,111,13,142,12,1,0 ; movdqa 0x10c8e(%rip),%xmm9 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,68,15,111,13,219,7,1,0 ; movdqa 0x107db(%rip),%xmm9 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,69,15,219,200 ; pand %xmm8,%xmm9 DB 69,15,91,201 ; cvtdq2ps %xmm9,%xmm9 - DB 68,15,89,13,141,12,1,0 ; mulps 0x10c8d(%rip),%xmm9 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,68,15,111,21,148,12,1,0 ; movdqa 0x10c94(%rip),%xmm10 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 68,15,89,13,218,7,1,0 ; mulps 0x107da(%rip),%xmm9 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,68,15,111,21,225,7,1,0 ; movdqa 0x107e1(%rip),%xmm10 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,69,15,219,208 ; pand %xmm8,%xmm10 DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11 - DB 68,15,89,29,147,12,1,0 ; mulps 0x10c93(%rip),%xmm11 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,68,15,219,5,154,12,1,0 ; pand 0x10c9a(%rip),%xmm8 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 68,15,89,29,224,7,1,0 ; mulps 0x107e0(%rip),%xmm11 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,68,15,219,5,231,7,1,0 ; pand 0x107e7(%rip),%xmm8 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 - DB 68,15,89,5,158,12,1,0 ; mulps 0x10c9e(%rip),%xmm8 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 68,15,89,5,235,7,1,0 ; mulps 0x107eb(%rip),%xmm8 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 68,15,40,211 ; movaps %xmm3,%xmm10 DB 68,15,194,215,1 ; cmpltps %xmm7,%xmm10 DB 69,15,40,227 ; movaps %xmm11,%xmm12 @@ -27155,33 +27317,33 @@ _sk_lerp_565_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,59 ; je 1cb23 <_sk_lerp_565_sse2+0x114> + DB 116,60 ; je 1cec7 <_sk_lerp_565_sse2+0x115> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,27 ; je 1cb0e <_sk_lerp_565_sse2+0xff> + DB 116,27 ; je 1ceb1 <_sk_lerp_565_sse2+0xff> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,60,255,255,255 ; jne 1ca39 <_sk_lerp_565_sse2+0x2a> + DB 15,133,60,255,255,255 ; jne 1cddc <_sk_lerp_565_sse2+0x2a> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 102,69,15,112,192,69 ; pshufd $0x45,%xmm8,%xmm8 DB 102,71,15,110,12,80 ; movd (%r8,%r10,2),%xmm9 - DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9 + DB 242,69,15,112,201,212 ; pshuflw $0xd4,%xmm9,%xmm9 DB 242,69,15,16,193 ; movsd %xmm9,%xmm8 - DB 233,22,255,255,255 ; jmpq 1ca39 <_sk_lerp_565_sse2+0x2a> + DB 233,21,255,255,255 ; jmpq 1cddc <_sk_lerp_565_sse2+0x2a> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 - DB 233,7,255,255,255 ; jmpq 1ca39 <_sk_lerp_565_sse2+0x2a> + DB 233,6,255,255,255 ; jmpq 1cddc <_sk_lerp_565_sse2+0x2a> PUBLIC _sk_load_tables_sse2 _sk_load_tables_sse2 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,32,1,0,0 ; jne 1cc60 <_sk_load_tables_sse2+0x12e> + DB 15,133,31,1,0,0 ; jne 1d006 <_sk_load_tables_sse2+0x130> DB 243,69,15,111,12,144 ; movdqu (%r8,%rdx,4),%xmm9 - DB 65,86 ; push %r14 - DB 83 ; push %rbx - DB 102,68,15,111,5,30,10,1,0 ; movdqa 0x10a1e(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,68,15,111,5,106,5,1,0 ; movdqa 0x1056a(%rip),%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 @@ -27194,12 +27356,12 @@ _sk_load_tables_sse2 LABEL PROC DB 72,139,88,8 ; mov 0x8(%rax),%rbx DB 76,139,112,16 ; mov 0x10(%rax),%r14 DB 243,66,15,16,12,11 ; movss (%rbx,%r9,1),%xmm1 - DB 243,66,15,16,4,3 ; movss (%rbx,%r8,1),%xmm0 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 - DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 DB 15,20,193 ; unpcklps %xmm1,%xmm0 + DB 243,66,15,16,12,3 ; movss (%rbx,%r8,1),%xmm1 + DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,15,20,194 ; unpcklpd %xmm2,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 DB 102,65,15,219,200 ; pand %xmm8,%xmm1 @@ -27211,12 +27373,12 @@ _sk_load_tables_sse2 LABEL PROC DB 69,15,182,208 ; movzbl %r8b,%r10d DB 73,193,232,30 ; shr $0x1e,%r8 DB 243,65,15,16,20,30 ; movss (%r14,%rbx,1),%xmm2 - DB 243,67,15,16,12,6 ; movss (%r14,%r8,1),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 - DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 + DB 243,67,15,16,20,6 ; movss (%r14,%r8,1),%xmm2 + DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 DB 76,139,64,24 ; mov 0x18(%rax),%r8 DB 102,65,15,111,209 ; movdqa %xmm9,%xmm2 DB 102,15,114,210,16 ; psrld $0x10,%xmm2 @@ -27228,16 +27390,16 @@ _sk_load_tables_sse2 LABEL PROC DB 72,193,232,30 ; shr $0x1e,%rax DB 68,15,182,211 ; movzbl %bl,%r10d DB 72,193,235,30 ; shr $0x1e,%rbx - DB 243,69,15,16,4,0 ; movss (%r8,%rax,1),%xmm8 - DB 243,65,15,16,20,24 ; movss (%r8,%rbx,1),%xmm2 - DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 + DB 243,65,15,16,28,0 ; movss (%r8,%rax,1),%xmm3 DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 - DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 + DB 243,69,15,16,4,24 ; movss (%r8,%rbx,1),%xmm8 + DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 65,15,20,216 ; unpcklps %xmm8,%xmm3 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 102,65,15,114,209,24 ; psrld $0x18,%xmm9 DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3 - DB 15,89,29,103,10,1,0 ; mulps 0x10a67(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,29,177,5,1,0 ; mulps 0x105b1(%rip),%xmm3 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 @@ -27245,91 +27407,90 @@ _sk_load_tables_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,45 ; je 1cc9a <_sk_load_tables_sse2+0x168> + DB 116,45 ; je 1d040 <_sk_load_tables_sse2+0x16a> DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,23 ; je 1cc8f <_sk_load_tables_sse2+0x15d> + DB 116,23 ; je 1d035 <_sk_load_tables_sse2+0x15f> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 15,133,196,254,255,255 ; jne 1cb46 <_sk_load_tables_sse2+0x14> + DB 15,133,197,254,255,255 ; jne 1ceed <_sk_load_tables_sse2+0x17> DB 102,65,15,110,68,144,8 ; movd 0x8(%r8,%rdx,4),%xmm0 DB 102,68,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm9 DB 102,69,15,18,12,144 ; movlpd (%r8,%rdx,4),%xmm9 - DB 233,172,254,255,255 ; jmpq 1cb46 <_sk_load_tables_sse2+0x14> + DB 233,173,254,255,255 ; jmpq 1ceed <_sk_load_tables_sse2+0x17> DB 102,69,15,110,12,144 ; movd (%r8,%rdx,4),%xmm9 - DB 233,161,254,255,255 ; jmpq 1cb46 <_sk_load_tables_sse2+0x14> + DB 233,162,254,255,255 ; jmpq 1ceed <_sk_load_tables_sse2+0x17> PUBLIC _sk_load_tables_u16_be_sse2 _sk_load_tables_u16_be_sse2 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,149,0,0,0,0 ; lea 0x0(,%rdx,4),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,97,1,0,0 ; jne 1ce1c <_sk_load_tables_u16_be_sse2+0x177> + DB 15,133,90,1,0,0 ; jne 1d1be <_sk_load_tables_u16_be_sse2+0x173> DB 102,67,15,16,4,72 ; movupd (%r8,%r9,2),%xmm0 - DB 102,67,15,16,76,72,16 ; movupd 0x10(%r8,%r9,2),%xmm1 - DB 65,86 ; push %r14 - DB 83 ; push %rbx + DB 243,67,15,111,76,72,16 ; movdqu 0x10(%r8,%r9,2),%xmm1 DB 102,68,15,40,200 ; movapd %xmm0,%xmm9 DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9 - DB 102,68,15,111,21,64,10,1,0 ; movdqa 0x10a40(%rip),%xmm10 # 2d730 <_sk_overlay_sse2_8bit+0x11d8> - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,65,15,219,194 ; pand %xmm10,%xmm0 + DB 102,15,111,21,139,5,1,0 ; movdqa 0x1058b(%rip),%xmm2 # 2d620 <_sk_overlay_sse2_8bit+0x1171> + DB 102,15,112,217,238 ; pshufd $0xee,%xmm1,%xmm3 + DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,65,15,97,192 ; punpcklwd %xmm8,%xmm0 - DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 - DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 69,15,182,200 ; movzbl %r8b,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,15,182,218 ; movzbl %r10b,%r11d - DB 73,193,234,30 ; shr $0x1e,%r10 - DB 72,139,88,8 ; mov 0x8(%rax),%rbx - DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,28,19 ; movss (%rbx,%r10,1),%xmm3 - DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 - DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 102,15,115,217,8 ; psrldq $0x8,%xmm1 - DB 102,65,15,219,202 ; pand %xmm10,%xmm1 DB 102,65,15,97,200 ; punpcklwd %xmm8,%xmm1 - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,73,15,126,192 ; movq %xmm0,%r8 DB 102,73,15,126,201 ; movq %xmm1,%r9 DB 69,15,182,209 ; movzbl %r9b,%r10d DB 73,193,233,30 ; shr $0x1e,%r9 - DB 243,67,15,16,20,14 ; movss (%r14,%r9,1),%xmm2 - DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 - DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 + DB 69,15,182,216 ; movzbl %r8b,%r11d + DB 73,193,232,30 ; shr $0x1e,%r8 + DB 72,139,88,8 ; mov 0x8(%rax),%rbx + DB 76,139,112,16 ; mov 0x10(%rax),%r14 + DB 243,66,15,16,12,11 ; movss (%rbx,%r9,1),%xmm1 + DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 + DB 15,20,193 ; unpcklps %xmm1,%xmm0 + DB 243,70,15,16,20,3 ; movss (%rbx,%r8,1),%xmm10 + DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 + DB 65,15,20,202 ; unpcklps %xmm10,%xmm1 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 102,15,219,218 ; pand %xmm2,%xmm3 + DB 102,65,15,97,216 ; punpcklwd %xmm8,%xmm3 + DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 + DB 102,73,15,126,200 ; movq %xmm1,%r8 + DB 102,72,15,126,219 ; movq %xmm3,%rbx + DB 68,15,182,203 ; movzbl %bl,%r9d + DB 72,193,235,30 ; shr $0x1e,%rbx + DB 69,15,182,208 ; movzbl %r8b,%r10d + DB 73,193,232,30 ; shr $0x1e,%r8 + DB 243,65,15,16,28,30 ; movss (%r14,%rbx,1),%xmm3 + DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 DB 15,20,203 ; unpcklps %xmm3,%xmm1 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 102,69,15,219,209 ; pand %xmm9,%xmm10 - DB 102,69,15,97,208 ; punpcklwd %xmm8,%xmm10 - DB 102,65,15,112,210,78 ; pshufd $0x4e,%xmm10,%xmm2 - DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,76,15,126,208 ; movq %xmm10,%rax - DB 68,15,182,208 ; movzbl %al,%r10d + DB 243,71,15,16,20,6 ; movss (%r14,%r8,1),%xmm10 + DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 65,15,20,218 ; unpcklps %xmm10,%xmm3 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 + DB 76,139,64,24 ; mov 0x18(%rax),%r8 + DB 102,65,15,219,209 ; pand %xmm9,%xmm2 + DB 102,65,15,97,208 ; punpcklwd %xmm8,%xmm2 + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,72,15,126,219 ; movq %xmm3,%rbx + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 68,15,182,200 ; movzbl %al,%r9d DB 72,193,232,30 ; shr $0x1e,%rax - DB 243,69,15,16,20,1 ; movss (%r9,%rax,1),%xmm10 - DB 243,65,15,16,20,153 ; movss (%r9,%rbx,4),%xmm2 - DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 - DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3 + DB 68,15,182,211 ; movzbl %bl,%r10d + DB 72,193,235,30 ; shr $0x1e,%rbx + DB 243,65,15,16,28,0 ; movss (%r8,%rax,1),%xmm3 + DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 65,15,20,210 ; unpcklps %xmm10,%xmm2 + DB 243,69,15,16,20,24 ; movss (%r8,%rbx,1),%xmm10 + DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 65,15,20,218 ; unpcklps %xmm10,%xmm3 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3 DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9 DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9 @@ -27337,124 +27498,123 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 102,65,15,235,217 ; por %xmm9,%xmm3 DB 102,65,15,97,216 ; punpcklwd %xmm8,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,59,9,1,0 ; mulps 0x1093b(%rip),%xmm3 # 2d750 <_sk_overlay_sse2_8bit+0x11f8> + DB 15,89,29,137,4,1,0 ; mulps 0x10489(%rip),%xmm3 # 2d640 <_sk_overlay_sse2_8bit+0x1191> DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx DB 65,94 ; pop %r14 DB 255,224 ; jmpq *%rax DB 242,67,15,16,4,72 ; movsd (%r8,%r9,2),%xmm0 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,17 ; jne 1ce39 <_sk_load_tables_u16_be_sse2+0x194> - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 233,143,254,255,255 ; jmpq 1ccc8 <_sk_load_tables_u16_be_sse2+0x23> + DB 117,13 ; jne 1d1d7 <_sk_load_tables_u16_be_sse2+0x18c> + DB 243,15,126,192 ; movq %xmm0,%xmm0 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 233,154,254,255,255 ; jmpq 1d071 <_sk_load_tables_u16_be_sse2+0x26> DB 102,67,15,22,68,72,8 ; movhpd 0x8(%r8,%r9,2),%xmm0 - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,122,254,255,255 ; jb 1ccc8 <_sk_load_tables_u16_be_sse2+0x23> - DB 242,67,15,16,76,72,16 ; movsd 0x10(%r8,%r9,2),%xmm1 - DB 233,110,254,255,255 ; jmpq 1ccc8 <_sk_load_tables_u16_be_sse2+0x23> + DB 15,130,133,254,255,255 ; jb 1d071 <_sk_load_tables_u16_be_sse2+0x26> + DB 243,67,15,126,76,72,16 ; movq 0x10(%r8,%r9,2),%xmm1 + DB 233,121,254,255,255 ; jmpq 1d071 <_sk_load_tables_u16_be_sse2+0x26> PUBLIC _sk_load_tables_rgb_u16_be_sse2 _sk_load_tables_rgb_u16_be_sse2 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,82 ; lea (%rdx,%rdx,2),%r9 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,80,1,0,0 ; jne 1cfbc <_sk_load_tables_rgb_u16_be_sse2+0x162> - DB 243,71,15,111,28,72 ; movdqu (%r8,%r9,2),%xmm11 + DB 15,133,78,1,0,0 ; jne 1d35b <_sk_load_tables_rgb_u16_be_sse2+0x163> + DB 243,67,15,111,20,72 ; movdqu (%r8,%r9,2),%xmm2 DB 243,67,15,111,76,72,8 ; movdqu 0x8(%r8,%r9,2),%xmm1 DB 102,15,115,217,4 ; psrldq $0x4,%xmm1 - DB 102,69,15,111,211 ; movdqa %xmm11,%xmm10 + DB 102,68,15,111,210 ; movdqa %xmm2,%xmm10 DB 102,65,15,115,218,6 ; psrldq $0x6,%xmm10 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 - DB 65,86 ; push %r14 - DB 83 ; push %rbx - DB 102,68,15,97,217 ; punpcklwd %xmm1,%xmm11 DB 102,68,15,97,208 ; punpcklwd %xmm0,%xmm10 - DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 + DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 102,65,15,97,194 ; punpcklwd %xmm10,%xmm0 - DB 102,68,15,111,5,126,8,1,0 ; movdqa 0x1087e(%rip),%xmm8 # 2d730 <_sk_overlay_sse2_8bit+0x11d8> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,68,15,111,5,205,3,1,0 ; movdqa 0x103cd(%rip),%xmm8 # 2d620 <_sk_overlay_sse2_8bit+0x1171> DB 102,65,15,219,192 ; pand %xmm8,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 69,15,182,200 ; movzbl %r8b,%r9d - DB 73,193,232,32 ; shr $0x20,%r8 - DB 102,73,15,126,194 ; movq %xmm0,%r10 - DB 69,15,182,218 ; movzbl %r10b,%r11d - DB 73,193,234,30 ; shr $0x1e,%r10 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,15,182,209 ; movzbl %r9b,%r10d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 69,15,182,216 ; movzbl %r8b,%r11d + DB 73,193,232,30 ; shr $0x1e,%r8 DB 72,139,88,8 ; mov 0x8(%rax),%rbx DB 76,139,112,16 ; mov 0x10(%rax),%r14 - DB 243,66,15,16,28,19 ; movss (%rbx,%r10,1),%xmm3 - DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 - DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 243,66,15,16,4,155 ; movss (%rbx,%r11,4),%xmm0 - DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 + DB 243,66,15,16,28,11 ; movss (%rbx,%r9,1),%xmm3 + DB 243,66,15,16,4,147 ; movss (%rbx,%r10,4),%xmm0 DB 15,20,195 ; unpcklps %xmm3,%xmm0 + DB 243,70,15,16,28,3 ; movss (%rbx,%r8,1),%xmm11 + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 65,15,20,219 ; unpcklps %xmm11,%xmm3 + DB 102,15,20,195 ; unpcklpd %xmm3,%xmm0 DB 102,65,15,219,200 ; pand %xmm8,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 69,15,182,209 ; movzbl %r9b,%r10d - DB 73,193,233,30 ; shr $0x1e,%r9 - DB 243,67,15,16,20,14 ; movss (%r14,%r9,1),%xmm2 - DB 243,65,15,16,12,158 ; movss (%r14,%rbx,4),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 - DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,73,15,126,216 ; movq %xmm3,%r8 + DB 102,72,15,126,203 ; movq %xmm1,%rbx + DB 68,15,182,203 ; movzbl %bl,%r9d + DB 72,193,235,30 ; shr $0x1e,%rbx + DB 69,15,182,208 ; movzbl %r8b,%r10d + DB 73,193,232,30 ; shr $0x1e,%r8 + DB 243,65,15,16,28,30 ; movss (%r14,%rbx,1),%xmm3 + DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 DB 15,20,203 ; unpcklps %xmm3,%xmm1 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 76,139,72,24 ; mov 0x18(%rax),%r9 - DB 102,69,15,105,218 ; punpckhwd %xmm10,%xmm11 - DB 102,69,15,219,216 ; pand %xmm8,%xmm11 - DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11 - DB 102,65,15,112,211,78 ; pshufd $0x4e,%xmm11,%xmm2 - DB 102,72,15,126,211 ; movq %xmm2,%rbx - DB 68,15,182,195 ; movzbl %bl,%r8d - DB 72,193,235,32 ; shr $0x20,%rbx - DB 102,76,15,126,216 ; movq %xmm11,%rax - DB 68,15,182,208 ; movzbl %al,%r10d + DB 243,71,15,16,28,6 ; movss (%r14,%r8,1),%xmm11 + DB 243,67,15,16,28,150 ; movss (%r14,%r10,4),%xmm3 + DB 65,15,20,219 ; unpcklps %xmm11,%xmm3 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 + DB 76,139,64,24 ; mov 0x18(%rax),%r8 + DB 102,65,15,105,210 ; punpckhwd %xmm10,%xmm2 + DB 102,65,15,219,208 ; pand %xmm8,%xmm2 + DB 102,65,15,97,209 ; punpcklwd %xmm9,%xmm2 + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,72,15,126,219 ; movq %xmm3,%rbx + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 68,15,182,200 ; movzbl %al,%r9d DB 72,193,232,30 ; shr $0x1e,%rax - DB 243,69,15,16,4,1 ; movss (%r9,%rax,1),%xmm8 - DB 243,65,15,16,20,153 ; movss (%r9,%rbx,4),%xmm2 - DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 - DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2 - DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3 + DB 68,15,182,211 ; movzbl %bl,%r10d + DB 72,193,235,30 ; shr $0x1e,%rbx + DB 243,65,15,16,28,0 ; movss (%r8,%rax,1),%xmm3 + DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 + DB 243,69,15,16,4,24 ; movss (%r8,%rbx,1),%xmm8 + DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3 + DB 65,15,20,216 ; unpcklps %xmm8,%xmm3 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,9,5,1,0 ; movaps 0x10509(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,90,0,1,0 ; movaps 0x1005a(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 91 ; pop %rbx DB 65,94 ; pop %r14 DB 255,224 ; jmpq *%rax - DB 102,71,15,110,28,72 ; movd (%r8,%r9,2),%xmm11 - DB 102,71,15,196,92,72,4,2 ; pinsrw $0x2,0x4(%r8,%r9,2),%xmm11 + DB 102,67,15,110,20,72 ; movd (%r8,%r9,2),%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 102,67,15,196,84,72,4,2 ; pinsrw $0x2,0x4(%r8,%r9,2),%xmm2 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,14 ; jne 1cfe2 <_sk_load_tables_rgb_u16_be_sse2+0x188> + DB 117,14 ; jne 1d381 <_sk_load_tables_rgb_u16_be_sse2+0x189> DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 233,176,254,255,255 ; jmpq 1ce92 <_sk_load_tables_rgb_u16_be_sse2+0x38> + DB 233,178,254,255,255 ; jmpq 1d233 <_sk_load_tables_rgb_u16_be_sse2+0x3b> DB 102,71,15,110,84,72,6 ; movd 0x6(%r8,%r9,2),%xmm10 - DB 102,71,15,196,84,72,10,2 ; pinsrw $0x2,0xa(%r8,%r9,2),%xmm10 DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 102,71,15,196,84,72,10,2 ; pinsrw $0x2,0xa(%r8,%r9,2),%xmm10 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,24 ; jb 1d013 <_sk_load_tables_rgb_u16_be_sse2+0x1b9> + DB 114,24 ; jb 1d3b2 <_sk_load_tables_rgb_u16_be_sse2+0x1ba> DB 102,67,15,110,76,72,12 ; movd 0xc(%r8,%r9,2),%xmm1 - DB 102,67,15,196,76,72,16,2 ; pinsrw $0x2,0x10(%r8,%r9,2),%xmm1 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,127,254,255,255 ; jmpq 1ce92 <_sk_load_tables_rgb_u16_be_sse2+0x38> + DB 102,67,15,196,76,72,16,2 ; pinsrw $0x2,0x10(%r8,%r9,2),%xmm1 + DB 233,129,254,255,255 ; jmpq 1d233 <_sk_load_tables_rgb_u16_be_sse2+0x3b> DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,118,254,255,255 ; jmpq 1ce92 <_sk_load_tables_rgb_u16_be_sse2+0x38> + DB 233,120,254,255,255 ; jmpq 1d233 <_sk_load_tables_rgb_u16_be_sse2+0x3b> PUBLIC _sk_byte_tables_sse2 _sk_byte_tables_sse2 LABEL PROC @@ -27462,7 +27622,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 65,86 ; push %r14 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,118,5,1,0 ; movaps 0x10576(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,199,0,1,0 ; movaps 0x100c7(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,73,15,126,192 ; movq %xmm0,%r8 @@ -27484,13 +27644,13 @@ _sk_byte_tables_sse2 LABEL PROC DB 66,15,182,28,19 ; movzbl (%rbx,%r10,1),%ebx DB 193,227,8 ; shl $0x8,%ebx DB 68,9,203 ; or %r9d,%ebx - DB 102,15,196,195,0 ; pinsrw $0x0,%ebx,%xmm0 + DB 102,15,110,195 ; movd %ebx,%xmm0 DB 102,15,196,197,1 ; pinsrw $0x1,%ebp,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,21,32,6,1,0 ; movaps 0x10620(%rip),%xmm10 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,21,114,1,1,0 ; movaps 0x10172(%rip),%xmm10 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -27509,7 +27669,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp DB 193,229,8 ; shl $0x8,%ebp DB 68,9,205 ; or %r9d,%ebp - DB 102,15,196,205,0 ; pinsrw $0x0,%ebp,%xmm1 + DB 102,15,110,205 ; movd %ebp,%xmm1 DB 102,15,196,203,1 ; pinsrw $0x1,%ebx,%xmm1 DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 @@ -27533,7 +27693,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx DB 193,227,8 ; shl $0x8,%ebx DB 68,9,195 ; or %r8d,%ebx - DB 102,15,196,211,0 ; pinsrw $0x0,%ebx,%xmm2 + DB 102,15,110,211 ; movd %ebx,%xmm2 DB 102,15,196,213,1 ; pinsrw $0x1,%ebp,%xmm2 DB 102,65,15,96,209 ; punpcklbw %xmm9,%xmm2 DB 102,65,15,97,209 ; punpcklwd %xmm9,%xmm2 @@ -27557,7 +27717,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 15,182,4,40 ; movzbl (%rax,%rbp,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 68,9,192 ; or %r8d,%eax - DB 102,15,196,216,0 ; pinsrw $0x0,%eax,%xmm3 + DB 102,15,110,216 ; movd %eax,%xmm3 DB 102,15,196,219,1 ; pinsrw $0x1,%ebx,%xmm3 DB 102,65,15,96,217 ; punpcklbw %xmm9,%xmm3 DB 102,65,15,97,217 ; punpcklwd %xmm9,%xmm3 @@ -27601,13 +27761,13 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 66,15,182,28,19 ; movzbl (%rbx,%r10,1),%ebx DB 193,227,8 ; shl $0x8,%ebx DB 68,9,203 ; or %r9d,%ebx - DB 102,15,196,195,0 ; pinsrw $0x0,%ebx,%xmm0 + DB 102,15,110,195 ; movd %ebx,%xmm0 DB 102,15,196,197,1 ; pinsrw $0x1,%ebp,%xmm0 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,21,92,4,1,0 ; movaps 0x1045c(%rip),%xmm10 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,21,178,255,0,0 ; movaps 0xffb2(%rip),%xmm10 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,194 ; mulps %xmm10,%xmm0 DB 65,15,89,200 ; mulps %xmm8,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 @@ -27626,7 +27786,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp DB 193,229,8 ; shl $0x8,%ebp DB 68,9,205 ; or %r9d,%ebp - DB 102,15,196,205,0 ; pinsrw $0x0,%ebp,%xmm1 + DB 102,15,110,205 ; movd %ebp,%xmm1 DB 102,15,196,203,1 ; pinsrw $0x1,%ebx,%xmm1 DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 @@ -27650,7 +27810,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 15,182,4,40 ; movzbl (%rax,%rbp,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 68,9,192 ; or %r8d,%eax - DB 102,15,196,208,0 ; pinsrw $0x0,%eax,%xmm2 + DB 102,15,110,208 ; movd %eax,%xmm2 DB 102,15,196,211,1 ; pinsrw $0x1,%ebx,%xmm2 DB 102,65,15,96,209 ; punpcklbw %xmm9,%xmm2 DB 102,65,15,97,209 ; punpcklwd %xmm9,%xmm2 @@ -27681,12 +27841,12 @@ _sk_table_r_sse2 LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,4,128 ; movss (%r8,%rax,4),%xmm0 - DB 68,15,20,192 ; unpcklps %xmm0,%xmm8 DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 - DB 65,15,20,193 ; unpcklps %xmm9,%xmm0 DB 65,15,20,192 ; unpcklps %xmm8,%xmm0 + DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 69,15,20,200 ; unpcklps %xmm8,%xmm9 + DB 102,65,15,20,193 ; unpcklpd %xmm9,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -27709,12 +27869,12 @@ _sk_table_g_sse2 LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,12,128 ; movss (%r8,%rax,4),%xmm1 - DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 DB 243,67,15,16,12,152 ; movss (%r8,%r11,4),%xmm1 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 - DB 65,15,20,201 ; unpcklps %xmm9,%xmm1 DB 65,15,20,200 ; unpcklps %xmm8,%xmm1 + DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 69,15,20,200 ; unpcklps %xmm8,%xmm9 + DB 102,65,15,20,201 ; unpcklpd %xmm9,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -27737,12 +27897,12 @@ _sk_table_b_sse2 LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,20,128 ; movss (%r8,%rax,4),%xmm2 - DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 - DB 65,15,20,209 ; unpcklps %xmm9,%xmm2 DB 65,15,20,208 ; unpcklps %xmm8,%xmm2 + DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 69,15,20,200 ; unpcklps %xmm8,%xmm9 + DB 102,65,15,20,209 ; unpcklpd %xmm9,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -27765,12 +27925,12 @@ _sk_table_a_sse2 LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 243,71,15,16,4,144 ; movss (%r8,%r10,4),%xmm8 - DB 243,65,15,16,28,128 ; movss (%r8,%rax,4),%xmm3 - DB 68,15,20,195 ; unpcklps %xmm3,%xmm8 DB 243,67,15,16,28,152 ; movss (%r8,%r11,4),%xmm3 - DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 - DB 65,15,20,217 ; unpcklps %xmm9,%xmm3 DB 65,15,20,216 ; unpcklps %xmm8,%xmm3 + DB 243,69,15,16,4,128 ; movss (%r8,%rax,4),%xmm8 + DB 243,71,15,16,12,136 ; movss (%r8,%r9,4),%xmm9 + DB 69,15,20,200 ; unpcklps %xmm8,%xmm9 + DB 102,65,15,20,217 ; unpcklpd %xmm9,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -27795,15 +27955,15 @@ _sk_parametric_r_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,58,2,1,0 ; mulps 0x1023a(%rip),%xmm9 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,21,66,2,1,0 ; andps 0x10242(%rip),%xmm10 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,21,106,255,0,0 ; orps 0xff6a(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,13,66,2,1,0 ; addps 0x10242(%rip),%xmm9 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 68,15,40,37,74,2,1,0 ; movaps 0x1024a(%rip),%xmm12 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,13,126,253,0,0 ; mulps 0xfd7e(%rip),%xmm9 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,21,134,253,0,0 ; andps 0xfd86(%rip),%xmm10 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,21,190,250,0,0 ; orps 0xfabe(%rip),%xmm10 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,13,134,253,0,0 ; addps 0xfd86(%rip),%xmm9 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 68,15,40,37,142,253,0,0 ; movaps 0xfd8e(%rip),%xmm12 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,74,2,1,0 ; addps 0x1024a(%rip),%xmm10 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 68,15,40,37,82,2,1,0 ; movaps 0x10252(%rip),%xmm12 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,21,142,253,0,0 ; addps 0xfd8e(%rip),%xmm10 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 68,15,40,37,150,253,0,0 ; movaps 0xfd96(%rip),%xmm12 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -27811,22 +27971,22 @@ _sk_parametric_r_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,44,255,0,0 ; movaps 0xff2c(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,128,250,0,0 ; movaps 0xfa80(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,32,2,1,0 ; addps 0x10220(%rip),%xmm9 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 68,15,40,37,40,2,1,0 ; movaps 0x10228(%rip),%xmm12 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,13,100,253,0,0 ; addps 0xfd64(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 68,15,40,37,108,253,0,0 ; movaps 0xfd6c(%rip),%xmm12 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,40,2,1,0 ; movaps 0x10228(%rip),%xmm12 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,37,108,253,0,0 ; movaps 0xfd6c(%rip),%xmm12 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,44,2,1,0 ; movaps 0x1022c(%rip),%xmm13 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,45,112,253,0,0 ; movaps 0xfd70(%rip),%xmm13 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,44,2,1,0 ; mulps 0x1022c(%rip),%xmm13 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 68,15,89,45,112,253,0,0 ; mulps 0xfd70(%rip),%xmm13 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -27860,15 +28020,15 @@ _sk_parametric_g_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,252,0,1,0 ; mulps 0x100fc(%rip),%xmm9 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,21,4,1,1,0 ; andps 0x10104(%rip),%xmm10 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,21,44,254,0,0 ; orps 0xfe2c(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,13,4,1,1,0 ; addps 0x10104(%rip),%xmm9 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 68,15,40,37,12,1,1,0 ; movaps 0x1010c(%rip),%xmm12 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,13,64,252,0,0 ; mulps 0xfc40(%rip),%xmm9 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,21,72,252,0,0 ; andps 0xfc48(%rip),%xmm10 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,21,128,249,0,0 ; orps 0xf980(%rip),%xmm10 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,13,72,252,0,0 ; addps 0xfc48(%rip),%xmm9 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 68,15,40,37,80,252,0,0 ; movaps 0xfc50(%rip),%xmm12 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,12,1,1,0 ; addps 0x1010c(%rip),%xmm10 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 68,15,40,37,20,1,1,0 ; movaps 0x10114(%rip),%xmm12 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,21,80,252,0,0 ; addps 0xfc50(%rip),%xmm10 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 68,15,40,37,88,252,0,0 ; movaps 0xfc58(%rip),%xmm12 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -27876,22 +28036,22 @@ _sk_parametric_g_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,238,253,0,0 ; movaps 0xfdee(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,66,249,0,0 ; movaps 0xf942(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,226,0,1,0 ; addps 0x100e2(%rip),%xmm9 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 68,15,40,37,234,0,1,0 ; movaps 0x100ea(%rip),%xmm12 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,13,38,252,0,0 ; addps 0xfc26(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 68,15,40,37,46,252,0,0 ; movaps 0xfc2e(%rip),%xmm12 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,234,0,1,0 ; movaps 0x100ea(%rip),%xmm12 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,37,46,252,0,0 ; movaps 0xfc2e(%rip),%xmm12 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,238,0,1,0 ; movaps 0x100ee(%rip),%xmm13 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,45,50,252,0,0 ; movaps 0xfc32(%rip),%xmm13 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,238,0,1,0 ; mulps 0x100ee(%rip),%xmm13 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 68,15,89,45,50,252,0,0 ; mulps 0xfc32(%rip),%xmm13 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -27925,15 +28085,15 @@ _sk_parametric_b_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,190,255,0,0 ; mulps 0xffbe(%rip),%xmm9 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,21,198,255,0,0 ; andps 0xffc6(%rip),%xmm10 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,21,238,252,0,0 ; orps 0xfcee(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,13,198,255,0,0 ; addps 0xffc6(%rip),%xmm9 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 68,15,40,37,206,255,0,0 ; movaps 0xffce(%rip),%xmm12 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,13,2,251,0,0 ; mulps 0xfb02(%rip),%xmm9 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,21,10,251,0,0 ; andps 0xfb0a(%rip),%xmm10 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,21,66,248,0,0 ; orps 0xf842(%rip),%xmm10 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,13,10,251,0,0 ; addps 0xfb0a(%rip),%xmm9 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 68,15,40,37,18,251,0,0 ; movaps 0xfb12(%rip),%xmm12 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,206,255,0,0 ; addps 0xffce(%rip),%xmm10 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 68,15,40,37,214,255,0,0 ; movaps 0xffd6(%rip),%xmm12 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,21,18,251,0,0 ; addps 0xfb12(%rip),%xmm10 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 68,15,40,37,26,251,0,0 ; movaps 0xfb1a(%rip),%xmm12 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -27941,22 +28101,22 @@ _sk_parametric_b_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,176,252,0,0 ; movaps 0xfcb0(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,4,248,0,0 ; movaps 0xf804(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,164,255,0,0 ; addps 0xffa4(%rip),%xmm9 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 68,15,40,37,172,255,0,0 ; movaps 0xffac(%rip),%xmm12 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,13,232,250,0,0 ; addps 0xfae8(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 68,15,40,37,240,250,0,0 ; movaps 0xfaf0(%rip),%xmm12 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,172,255,0,0 ; movaps 0xffac(%rip),%xmm12 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,37,240,250,0,0 ; movaps 0xfaf0(%rip),%xmm12 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,176,255,0,0 ; movaps 0xffb0(%rip),%xmm13 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,45,244,250,0,0 ; movaps 0xfaf4(%rip),%xmm13 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,176,255,0,0 ; mulps 0xffb0(%rip),%xmm13 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 68,15,89,45,244,250,0,0 ; mulps 0xfaf4(%rip),%xmm13 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -27990,15 +28150,15 @@ _sk_parametric_a_sse2 LABEL PROC DB 69,15,88,209 ; addps %xmm9,%xmm10 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9 - DB 68,15,89,13,128,254,0,0 ; mulps 0xfe80(%rip),%xmm9 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 68,15,84,21,136,254,0,0 ; andps 0xfe88(%rip),%xmm10 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,86,21,176,251,0,0 ; orps 0xfbb0(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> - DB 68,15,88,13,136,254,0,0 ; addps 0xfe88(%rip),%xmm9 # 2d790 <_sk_overlay_sse2_8bit+0x1238> - DB 68,15,40,37,144,254,0,0 ; movaps 0xfe90(%rip),%xmm12 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,89,13,196,249,0,0 ; mulps 0xf9c4(%rip),%xmm9 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 68,15,84,21,204,249,0,0 ; andps 0xf9cc(%rip),%xmm10 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,86,21,4,247,0,0 ; orps 0xf704(%rip),%xmm10 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> + DB 68,15,88,13,204,249,0,0 ; addps 0xf9cc(%rip),%xmm9 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> + DB 68,15,40,37,212,249,0,0 ; movaps 0xf9d4(%rip),%xmm12 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 69,15,89,226 ; mulps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,88,21,144,254,0,0 ; addps 0xfe90(%rip),%xmm10 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> - DB 68,15,40,37,152,254,0,0 ; movaps 0xfe98(%rip),%xmm12 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,88,21,212,249,0,0 ; addps 0xf9d4(%rip),%xmm10 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> + DB 68,15,40,37,220,249,0,0 ; movaps 0xf9dc(%rip),%xmm12 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 69,15,94,226 ; divps %xmm10,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 DB 69,15,89,203 ; mulps %xmm11,%xmm9 @@ -28006,22 +28166,22 @@ _sk_parametric_a_sse2 LABEL PROC DB 69,15,91,226 ; cvtdq2ps %xmm10,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,194,236,1 ; cmpltps %xmm12,%xmm13 - DB 68,15,40,21,114,251,0,0 ; movaps 0xfb72(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,198,246,0,0 ; movaps 0xf6c6(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,84,234 ; andps %xmm10,%xmm13 DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,229 ; subps %xmm13,%xmm12 DB 69,15,40,233 ; movaps %xmm9,%xmm13 DB 69,15,92,236 ; subps %xmm12,%xmm13 - DB 68,15,88,13,102,254,0,0 ; addps 0xfe66(%rip),%xmm9 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> - DB 68,15,40,37,110,254,0,0 ; movaps 0xfe6e(%rip),%xmm12 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 68,15,88,13,170,249,0,0 ; addps 0xf9aa(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> + DB 68,15,40,37,178,249,0,0 ; movaps 0xf9b2(%rip),%xmm12 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 69,15,89,229 ; mulps %xmm13,%xmm12 DB 69,15,92,204 ; subps %xmm12,%xmm9 - DB 68,15,40,37,110,254,0,0 ; movaps 0xfe6e(%rip),%xmm12 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 68,15,40,37,178,249,0,0 ; movaps 0xf9b2(%rip),%xmm12 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 69,15,92,229 ; subps %xmm13,%xmm12 - DB 68,15,40,45,114,254,0,0 ; movaps 0xfe72(%rip),%xmm13 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,45,182,249,0,0 ; movaps 0xf9b6(%rip),%xmm13 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 69,15,94,236 ; divps %xmm12,%xmm13 DB 69,15,88,233 ; addps %xmm9,%xmm13 - DB 68,15,89,45,114,254,0,0 ; mulps 0xfe72(%rip),%xmm13 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 68,15,89,45,182,249,0,0 ; mulps 0xf9b6(%rip),%xmm13 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 102,69,15,91,205 ; cvtps2dq %xmm13,%xmm9 DB 243,68,15,16,96,20 ; movss 0x14(%rax),%xmm12 DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12 @@ -28045,19 +28205,19 @@ _sk_gamma_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,91,194 ; cvtdq2ps %xmm2,%xmm0 - DB 15,89,5,122,253,0,0 ; mulps 0xfd7a(%rip),%xmm0 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 15,84,21,131,253,0,0 ; andps 0xfd83(%rip),%xmm2 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 68,15,40,53,171,250,0,0 ; movaps 0xfaab(%rip),%xmm14 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,89,5,190,248,0,0 ; mulps 0xf8be(%rip),%xmm0 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 15,84,21,199,248,0,0 ; andps 0xf8c7(%rip),%xmm2 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 68,15,40,53,255,245,0,0 ; movaps 0xf5ff(%rip),%xmm14 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,86,214 ; orps %xmm14,%xmm2 - DB 68,15,40,37,127,253,0,0 ; movaps 0xfd7f(%rip),%xmm12 # 2d790 <_sk_overlay_sse2_8bit+0x1238> + DB 68,15,40,37,195,248,0,0 ; movaps 0xf8c3(%rip),%xmm12 # 2d670 <_sk_overlay_sse2_8bit+0x11c1> DB 65,15,88,196 ; addps %xmm12,%xmm0 - DB 68,15,40,29,131,253,0,0 ; movaps 0xfd83(%rip),%xmm11 # 2d7a0 <_sk_overlay_sse2_8bit+0x1248> + DB 68,15,40,29,199,248,0,0 ; movaps 0xf8c7(%rip),%xmm11 # 2d680 <_sk_overlay_sse2_8bit+0x11d1> DB 15,40,226 ; movaps %xmm2,%xmm4 DB 65,15,89,227 ; mulps %xmm11,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 68,15,40,21,129,253,0,0 ; movaps 0xfd81(%rip),%xmm10 # 2d7b0 <_sk_overlay_sse2_8bit+0x1258> + DB 68,15,40,21,197,248,0,0 ; movaps 0xf8c5(%rip),%xmm10 # 2d690 <_sk_overlay_sse2_8bit+0x11e1> DB 65,15,88,210 ; addps %xmm10,%xmm2 - DB 68,15,40,45,133,253,0,0 ; movaps 0xfd85(%rip),%xmm13 # 2d7c0 <_sk_overlay_sse2_8bit+0x1268> + DB 68,15,40,45,201,248,0,0 ; movaps 0xf8c9(%rip),%xmm13 # 2d6a0 <_sk_overlay_sse2_8bit+0x11f1> DB 65,15,40,229 ; movaps %xmm13,%xmm4 DB 15,94,226 ; divps %xmm2,%xmm4 DB 72,173 ; lods %ds:(%rsi),%rax @@ -28069,27 +28229,27 @@ _sk_gamma_sse2 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 15,40,224 ; movaps %xmm0,%xmm4 DB 15,194,226,1 ; cmpltps %xmm2,%xmm4 - DB 68,15,40,13,85,250,0,0 ; movaps 0xfa55(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,13,169,245,0,0 ; movaps 0xf5a9(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 65,15,84,225 ; andps %xmm9,%xmm4 DB 15,92,212 ; subps %xmm4,%xmm2 DB 15,40,224 ; movaps %xmm0,%xmm4 DB 15,92,226 ; subps %xmm2,%xmm4 - DB 15,40,53,81,253,0,0 ; movaps 0xfd51(%rip),%xmm6 # 2d7d0 <_sk_overlay_sse2_8bit+0x1278> + DB 15,40,53,149,248,0,0 ; movaps 0xf895(%rip),%xmm6 # 2d6b0 <_sk_overlay_sse2_8bit+0x1201> DB 15,88,198 ; addps %xmm6,%xmm0 - DB 15,40,61,103,253,0,0 ; movaps 0xfd67(%rip),%xmm7 # 2d7f0 <_sk_overlay_sse2_8bit+0x1298> + DB 15,40,61,171,248,0,0 ; movaps 0xf8ab(%rip),%xmm7 # 2d6d0 <_sk_overlay_sse2_8bit+0x1221> DB 15,40,239 ; movaps %xmm7,%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 DB 15,40,212 ; movaps %xmm4,%xmm2 - DB 15,40,37,71,253,0,0 ; movaps 0xfd47(%rip),%xmm4 # 2d7e0 <_sk_overlay_sse2_8bit+0x1288> + DB 15,40,37,139,248,0,0 ; movaps 0xf88b(%rip),%xmm4 # 2d6c0 <_sk_overlay_sse2_8bit+0x1211> DB 15,89,212 ; mulps %xmm4,%xmm2 DB 15,92,194 ; subps %xmm2,%xmm0 - DB 68,15,40,61,89,253,0,0 ; movaps 0xfd59(%rip),%xmm15 # 2d800 <_sk_overlay_sse2_8bit+0x12a8> + DB 68,15,40,61,157,248,0,0 ; movaps 0xf89d(%rip),%xmm15 # 2d6e0 <_sk_overlay_sse2_8bit+0x1231> DB 65,15,40,215 ; movaps %xmm15,%xmm2 DB 15,94,213 ; divps %xmm5,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 15,89,5,181,252,0,0 ; mulps 0xfcb5(%rip),%xmm0 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 15,84,13,190,252,0,0 ; andps 0xfcbe(%rip),%xmm1 # 2d780 <_sk_overlay_sse2_8bit+0x1228> + DB 15,89,5,249,247,0,0 ; mulps 0xf7f9(%rip),%xmm0 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 15,84,13,2,248,0,0 ; andps 0xf802(%rip),%xmm1 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> DB 65,15,86,206 ; orps %xmm14,%xmm1 DB 65,15,88,196 ; addps %xmm12,%xmm0 DB 15,40,233 ; movaps %xmm1,%xmm5 @@ -28119,9 +28279,9 @@ _sk_gamma_sse2 LABEL PROC DB 15,94,206 ; divps %xmm6,%xmm1 DB 15,88,200 ; addps %xmm0,%xmm1 DB 15,91,195 ; cvtdq2ps %xmm3,%xmm0 - DB 15,89,5,68,252,0,0 ; mulps 0xfc44(%rip),%xmm0 # 2d770 <_sk_overlay_sse2_8bit+0x1218> - DB 15,84,29,77,252,0,0 ; andps 0xfc4d(%rip),%xmm3 # 2d780 <_sk_overlay_sse2_8bit+0x1228> - DB 15,86,29,118,249,0,0 ; orps 0xf976(%rip),%xmm3 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,89,5,136,247,0,0 ; mulps 0xf788(%rip),%xmm0 # 2d650 <_sk_overlay_sse2_8bit+0x11a1> + DB 15,84,29,145,247,0,0 ; andps 0xf791(%rip),%xmm3 # 2d660 <_sk_overlay_sse2_8bit+0x11b1> + DB 15,86,29,202,244,0,0 ; orps 0xf4ca(%rip),%xmm3 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,88,196 ; addps %xmm12,%xmm0 DB 68,15,89,219 ; mulps %xmm3,%xmm11 DB 65,15,92,195 ; subps %xmm11,%xmm0 @@ -28143,7 +28303,7 @@ _sk_gamma_sse2 LABEL PROC DB 15,92,253 ; subps %xmm5,%xmm7 DB 68,15,94,255 ; divps %xmm7,%xmm15 DB 68,15,88,248 ; addps %xmm0,%xmm15 - DB 15,40,5,131,252,0,0 ; movaps 0xfc83(%rip),%xmm0 # 2d810 <_sk_overlay_sse2_8bit+0x12b8> + DB 15,40,5,199,247,0,0 ; movaps 0xf7c7(%rip),%xmm0 # 2d6f0 <_sk_overlay_sse2_8bit+0x1241> DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 68,15,89,248 ; mulps %xmm0,%xmm15 @@ -28161,29 +28321,29 @@ _sk_gamma_sse2 LABEL PROC PUBLIC _sk_lab_to_xyz_sse2 _sk_lab_to_xyz_sse2 LABEL PROC - DB 15,89,5,85,252,0,0 ; mulps 0xfc55(%rip),%xmm0 # 2d820 <_sk_overlay_sse2_8bit+0x12c8> - DB 68,15,40,5,205,249,0,0 ; movaps 0xf9cd(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 15,89,5,153,247,0,0 ; mulps 0xf799(%rip),%xmm0 # 2d700 <_sk_overlay_sse2_8bit+0x1251> + DB 68,15,40,5,33,245,0,0 ; movaps 0xf521(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 65,15,89,200 ; mulps %xmm8,%xmm1 - DB 68,15,40,13,81,252,0,0 ; movaps 0xfc51(%rip),%xmm9 # 2d830 <_sk_overlay_sse2_8bit+0x12d8> + DB 68,15,40,13,149,247,0,0 ; movaps 0xf795(%rip),%xmm9 # 2d710 <_sk_overlay_sse2_8bit+0x1261> DB 65,15,88,201 ; addps %xmm9,%xmm1 DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 65,15,88,209 ; addps %xmm9,%xmm2 - DB 15,88,5,78,252,0,0 ; addps 0xfc4e(%rip),%xmm0 # 2d840 <_sk_overlay_sse2_8bit+0x12e8> - DB 15,89,5,87,252,0,0 ; mulps 0xfc57(%rip),%xmm0 # 2d850 <_sk_overlay_sse2_8bit+0x12f8> - DB 15,89,13,96,252,0,0 ; mulps 0xfc60(%rip),%xmm1 # 2d860 <_sk_overlay_sse2_8bit+0x1308> + DB 15,88,5,146,247,0,0 ; addps 0xf792(%rip),%xmm0 # 2d720 <_sk_overlay_sse2_8bit+0x1271> + DB 15,89,5,155,247,0,0 ; mulps 0xf79b(%rip),%xmm0 # 2d730 <_sk_overlay_sse2_8bit+0x1281> + DB 15,89,13,164,247,0,0 ; mulps 0xf7a4(%rip),%xmm1 # 2d740 <_sk_overlay_sse2_8bit+0x1291> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,21,102,252,0,0 ; mulps 0xfc66(%rip),%xmm2 # 2d870 <_sk_overlay_sse2_8bit+0x1318> + DB 15,89,21,170,247,0,0 ; mulps 0xf7aa(%rip),%xmm2 # 2d750 <_sk_overlay_sse2_8bit+0x12a1> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 68,15,92,202 ; subps %xmm2,%xmm9 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 DB 68,15,89,225 ; mulps %xmm1,%xmm12 - DB 15,40,21,91,252,0,0 ; movaps 0xfc5b(%rip),%xmm2 # 2d880 <_sk_overlay_sse2_8bit+0x1328> + DB 15,40,21,159,247,0,0 ; movaps 0xf79f(%rip),%xmm2 # 2d760 <_sk_overlay_sse2_8bit+0x12b1> DB 68,15,40,194 ; movaps %xmm2,%xmm8 DB 69,15,194,196,1 ; cmpltps %xmm12,%xmm8 - DB 68,15,40,21,90,252,0,0 ; movaps 0xfc5a(%rip),%xmm10 # 2d890 <_sk_overlay_sse2_8bit+0x1338> + DB 68,15,40,21,158,247,0,0 ; movaps 0xf79e(%rip),%xmm10 # 2d770 <_sk_overlay_sse2_8bit+0x12c1> DB 65,15,88,202 ; addps %xmm10,%xmm1 - DB 68,15,40,29,94,252,0,0 ; movaps 0xfc5e(%rip),%xmm11 # 2d8a0 <_sk_overlay_sse2_8bit+0x1348> + DB 68,15,40,29,162,247,0,0 ; movaps 0xf7a2(%rip),%xmm11 # 2d780 <_sk_overlay_sse2_8bit+0x12d1> DB 65,15,89,203 ; mulps %xmm11,%xmm1 DB 69,15,84,224 ; andps %xmm8,%xmm12 DB 68,15,85,193 ; andnps %xmm1,%xmm8 @@ -28207,8 +28367,8 @@ _sk_lab_to_xyz_sse2 LABEL PROC DB 15,84,194 ; andps %xmm2,%xmm0 DB 65,15,85,209 ; andnps %xmm9,%xmm2 DB 15,86,208 ; orps %xmm0,%xmm2 - DB 68,15,89,5,14,252,0,0 ; mulps 0xfc0e(%rip),%xmm8 # 2d8b0 <_sk_overlay_sse2_8bit+0x1358> - DB 15,89,21,23,252,0,0 ; mulps 0xfc17(%rip),%xmm2 # 2d8c0 <_sk_overlay_sse2_8bit+0x1368> + DB 68,15,89,5,82,247,0,0 ; mulps 0xf752(%rip),%xmm8 # 2d790 <_sk_overlay_sse2_8bit+0x12e1> + DB 15,89,21,91,247,0,0 ; mulps 0xf75b(%rip),%xmm2 # 2d7a0 <_sk_overlay_sse2_8bit+0x12f1> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -28222,13 +28382,13 @@ _sk_load_a8_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,46 ; jne 1dcf7 <_sk_load_a8_sse2+0x46> + DB 117,46 ; jne 1e093 <_sk_load_a8_sse2+0x46> DB 102,67,15,110,4,16 ; movd (%r8,%r10,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,5,145,248,0,0 ; pand 0xf891(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,5,229,243,0,0 ; pand 0xf3e5(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,215,249,0,0 ; mulps 0xf9d7(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,29,43,245,0,0 ; mulps 0xf52b(%rip),%xmm3 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,87,201 ; xorpd %xmm1,%xmm1 @@ -28237,12 +28397,12 @@ _sk_load_a8_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,54 ; je 1dd3a <_sk_load_a8_sse2+0x89> + DB 116,54 ; je 1e0d6 <_sk_load_a8_sse2+0x89> DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1dd23 <_sk_load_a8_sse2+0x72> + DB 116,21 ; je 1e0bf <_sk_load_a8_sse2+0x72> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,195 ; jne 1dcd7 <_sk_load_a8_sse2+0x26> + DB 117,195 ; jne 1e073 <_sk_load_a8_sse2+0x26> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 @@ -28251,10 +28411,10 @@ _sk_load_a8_sse2 LABEL PROC DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 235,157 ; jmp 1dcd7 <_sk_load_a8_sse2+0x26> + DB 235,157 ; jmp 1e073 <_sk_load_a8_sse2+0x26> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 - DB 235,146 ; jmp 1dcd7 <_sk_load_a8_sse2+0x26> + DB 235,146 ; jmp 1e073 <_sk_load_a8_sse2+0x26> PUBLIC _sk_load_a8_dst_sse2 _sk_load_a8_dst_sse2 LABEL PROC @@ -28265,13 +28425,13 @@ _sk_load_a8_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,46 ; jne 1dd8b <_sk_load_a8_dst_sse2+0x46> + DB 117,46 ; jne 1e127 <_sk_load_a8_dst_sse2+0x46> DB 102,67,15,110,36,16 ; movd (%r8,%r10,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,219,37,253,247,0,0 ; pand 0xf7fd(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,37,81,243,0,0 ; pand 0xf351(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,252 ; cvtdq2ps %xmm4,%xmm7 - DB 15,89,61,67,249,0,0 ; mulps 0xf943(%rip),%xmm7 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,61,151,244,0,0 ; mulps 0xf497(%rip),%xmm7 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,228 ; xorps %xmm4,%xmm4 DB 102,15,87,237 ; xorpd %xmm5,%xmm5 @@ -28280,12 +28440,12 @@ _sk_load_a8_dst_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,54 ; je 1ddce <_sk_load_a8_dst_sse2+0x89> + DB 116,54 ; je 1e16a <_sk_load_a8_dst_sse2+0x89> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1ddb7 <_sk_load_a8_dst_sse2+0x72> + DB 116,21 ; je 1e153 <_sk_load_a8_dst_sse2+0x72> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,195 ; jne 1dd6b <_sk_load_a8_dst_sse2+0x26> + DB 117,195 ; jne 1e107 <_sk_load_a8_dst_sse2+0x26> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -28294,10 +28454,10 @@ _sk_load_a8_dst_sse2 LABEL PROC DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 235,157 ; jmp 1dd6b <_sk_load_a8_dst_sse2+0x26> + DB 235,157 ; jmp 1e107 <_sk_load_a8_dst_sse2+0x26> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 - DB 235,146 ; jmp 1dd6b <_sk_load_a8_dst_sse2+0x26> + DB 235,146 ; jmp 1e107 <_sk_load_a8_dst_sse2+0x26> PUBLIC _sk_gather_a8_sse2 _sk_gather_a8_sse2 LABEL PROC @@ -28329,13 +28489,13 @@ _sk_gather_a8_sse2 LABEL PROC DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 68,9,200 ; or %r9d,%eax - DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 + DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,65,15,196,194,1 ; pinsrw $0x1,%r10d,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,29,82,248,0,0 ; mulps 0xf852(%rip),%xmm3 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,29,167,243,0,0 ; mulps 0xf3a7(%rip),%xmm3 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 @@ -28344,50 +28504,48 @@ _sk_gather_a8_sse2 LABEL PROC PUBLIC _sk_store_a8_sse2 _sk_store_a8_sse2 LABEL PROC - DB 72,131,236,40 ; sub $0x28,%rsp + DB 72,131,236,4 ; sub $0x4,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 76,99,72,8 ; movslq 0x8(%rax),%r9 DB 76,99,193 ; movslq %ecx,%r8 DB 77,15,175,193 ; imul %r9,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 - DB 68,15,40,5,4,247,0,0 ; movaps 0xf704(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,89,242,0,0 ; movaps 0xf259(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 68,15,89,195 ; mulps %xmm3,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 - DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 - DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 + DB 102,68,15,107,192 ; packssdw %xmm0,%xmm8 + DB 102,68,15,103,192 ; packuswb %xmm0,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 1ded1 <_sk_store_a8_sse2+0x54> + DB 117,17 ; jne 1e26c <_sk_store_a8_sse2+0x54> DB 102,68,15,126,192 ; movd %xmm8,%eax DB 67,137,4,16 ; mov %eax,(%r8,%r10,1) DB 72,173 ; lods %ds:(%rsi),%rax - DB 72,131,196,40 ; add $0x28,%rsp + DB 72,131,196,4 ; add $0x4,%rsp DB 255,224 ; jmpq *%rax DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,59 ; je 1df23 <_sk_store_a8_sse2+0xa6> + DB 116,54 ; je 1e2b9 <_sk_store_a8_sse2+0xa1> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,22 ; je 1df04 <_sk_store_a8_sse2+0x87> + DB 116,17 ; je 1e29a <_sk_store_a8_sse2+0x82> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,213 ; jne 1dec9 <_sk_store_a8_sse2+0x4c> - DB 102,68,15,127,68,36,16 ; movdqa %xmm8,0x10(%rsp) - DB 138,68,36,24 ; mov 0x18(%rsp),%al + DB 117,213 ; jne 1e264 <_sk_store_a8_sse2+0x4c> + DB 102,65,15,197,192,4 ; pextrw $0x4,%xmm8,%eax DB 67,136,68,16,2 ; mov %al,0x2(%r8,%r10,1) - DB 102,68,15,219,5,99,246,0,0 ; pand 0xf663(%rip),%xmm8 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,68,15,219,5,189,241,0,0 ; pand 0xf1bd(%rip),%xmm8 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 DB 102,68,15,126,192 ; movd %xmm8,%eax DB 102,67,137,4,16 ; mov %ax,(%r8,%r10,1) - DB 235,166 ; jmp 1dec9 <_sk_store_a8_sse2+0x4c> - DB 102,68,15,127,4,36 ; movdqa %xmm8,(%rsp) - DB 138,4,36 ; mov (%rsp),%al + DB 235,171 ; jmp 1e264 <_sk_store_a8_sse2+0x4c> + DB 102,68,15,126,192 ; movd %xmm8,%eax DB 67,136,4,16 ; mov %al,(%r8,%r10,1) - DB 235,151 ; jmp 1dec9 <_sk_store_a8_sse2+0x4c> + DB 235,160 ; jmp 1e264 <_sk_store_a8_sse2+0x4c> PUBLIC _sk_load_g8_sse2 _sk_load_g8_sse2 LABEL PROC @@ -28398,27 +28556,27 @@ _sk_load_g8_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,49 ; jne 1df7b <_sk_load_g8_sse2+0x49> + DB 117,49 ; jne 1e30d <_sk_load_g8_sse2+0x49> DB 102,67,15,110,4,16 ; movd (%r8,%r10,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,5,16,246,0,0 ; pand 0xf610(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,5,110,241,0,0 ; pand 0xf16e(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,86,247,0,0 ; mulps 0xf756(%rip),%xmm0 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,5,180,242,0,0 ; mulps 0xf2b4(%rip),%xmm0 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,77,245,0,0 ; movaps 0xf54d(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,171,240,0,0 ; movaps 0xf0ab(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,54 ; je 1dfbe <_sk_load_g8_sse2+0x8c> + DB 116,54 ; je 1e350 <_sk_load_g8_sse2+0x8c> DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1dfa7 <_sk_load_g8_sse2+0x75> + DB 116,21 ; je 1e339 <_sk_load_g8_sse2+0x75> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,192 ; jne 1df58 <_sk_load_g8_sse2+0x26> + DB 117,192 ; jne 1e2ea <_sk_load_g8_sse2+0x26> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 @@ -28427,10 +28585,10 @@ _sk_load_g8_sse2 LABEL PROC DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 235,154 ; jmp 1df58 <_sk_load_g8_sse2+0x26> + DB 235,154 ; jmp 1e2ea <_sk_load_g8_sse2+0x26> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 - DB 235,143 ; jmp 1df58 <_sk_load_g8_sse2+0x26> + DB 235,143 ; jmp 1e2ea <_sk_load_g8_sse2+0x26> PUBLIC _sk_load_g8_dst_sse2 _sk_load_g8_dst_sse2 LABEL PROC @@ -28441,27 +28599,27 @@ _sk_load_g8_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,49 ; jne 1e012 <_sk_load_g8_dst_sse2+0x49> + DB 117,49 ; jne 1e3a4 <_sk_load_g8_dst_sse2+0x49> DB 102,67,15,110,36,16 ; movd (%r8,%r10,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,219,37,121,245,0,0 ; pand 0xf579(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,219,37,215,240,0,0 ; pand 0xf0d7(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,37,191,246,0,0 ; mulps 0xf6bf(%rip),%xmm4 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,37,29,242,0,0 ; mulps 0xf21d(%rip),%xmm4 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,61,182,244,0,0 ; movaps 0xf4b6(%rip),%xmm7 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,61,20,240,0,0 ; movaps 0xf014(%rip),%xmm7 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,40,244 ; movaps %xmm4,%xmm6 DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,54 ; je 1e055 <_sk_load_g8_dst_sse2+0x8c> + DB 116,54 ; je 1e3e7 <_sk_load_g8_dst_sse2+0x8c> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1e03e <_sk_load_g8_dst_sse2+0x75> + DB 116,21 ; je 1e3d0 <_sk_load_g8_dst_sse2+0x75> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,192 ; jne 1dfef <_sk_load_g8_dst_sse2+0x26> + DB 117,192 ; jne 1e381 <_sk_load_g8_dst_sse2+0x26> DB 67,15,182,68,16,2 ; movzbl 0x2(%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -28470,10 +28628,10 @@ _sk_load_g8_dst_sse2 LABEL PROC DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 235,154 ; jmp 1dfef <_sk_load_g8_dst_sse2+0x26> + DB 235,154 ; jmp 1e381 <_sk_load_g8_dst_sse2+0x26> DB 67,15,182,4,16 ; movzbl (%r8,%r10,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 - DB 235,143 ; jmp 1dfef <_sk_load_g8_dst_sse2+0x26> + DB 235,143 ; jmp 1e381 <_sk_load_g8_dst_sse2+0x26> PUBLIC _sk_gather_g8_sse2 _sk_gather_g8_sse2 LABEL PROC @@ -28505,15 +28663,15 @@ _sk_gather_g8_sse2 LABEL PROC DB 65,15,182,4,0 ; movzbl (%r8,%rax,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 68,9,200 ; or %r9d,%eax - DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 + DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,65,15,196,194,1 ; pinsrw $0x1,%r10d,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,203,245,0,0 ; mulps 0xf5cb(%rip),%xmm0 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 15,89,5,42,241,0,0 ; mulps 0xf12a(%rip),%xmm0 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,194,243,0,0 ; movaps 0xf3c2(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,33,239,0,0 ; movaps 0xef21(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax @@ -28528,42 +28686,42 @@ _sk_load_565_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,83 ; jne 1e174 <_sk_load_565_sse2+0x6e> + DB 117,83 ; jne 1e505 <_sk_load_565_sse2+0x6e> DB 243,67,15,126,20,80 ; movq (%r8,%r10,2),%xmm2 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,157,245,0,0 ; movdqa 0xf59d(%rip),%xmm0 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,5,252,240,0,0 ; movdqa 0xf0fc(%rip),%xmm0 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,159,245,0,0 ; mulps 0xf59f(%rip),%xmm0 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,13,167,245,0,0 ; movdqa 0xf5a7(%rip),%xmm1 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 15,89,5,254,240,0,0 ; mulps 0xf0fe(%rip),%xmm0 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,13,6,241,0,0 ; movdqa 0xf106(%rip),%xmm1 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,169,245,0,0 ; mulps 0xf5a9(%rip),%xmm1 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,15,219,21,177,245,0,0 ; pand 0xf5b1(%rip),%xmm2 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 15,89,13,8,241,0,0 ; mulps 0xf108(%rip),%xmm1 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,15,219,21,16,241,0,0 ; pand 0xf110(%rip),%xmm2 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,183,245,0,0 ; mulps 0xf5b7(%rip),%xmm2 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 15,89,21,22,241,0,0 ; mulps 0xf116(%rip),%xmm2 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,78,243,0,0 ; movaps 0xf34e(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,173,238,0,0 ; movaps 0xeead(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,50 ; je 1e1b3 <_sk_load_565_sse2+0xad> + DB 116,51 ; je 1e545 <_sk_load_565_sse2+0xae> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1e1a0 <_sk_load_565_sse2+0x9a> + DB 116,21 ; je 1e531 <_sk_load_565_sse2+0x9a> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,154 ; jne 1e12b <_sk_load_565_sse2+0x25> + DB 117,154 ; jne 1e4bc <_sk_load_565_sse2+0x25> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,208,69 ; pshufd $0x45,%xmm0,%xmm2 DB 102,67,15,110,4,80 ; movd (%r8,%r10,2),%xmm0 - DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 + DB 242,15,112,192,212 ; pshuflw $0xd4,%xmm0,%xmm0 DB 242,15,16,208 ; movsd %xmm0,%xmm2 - DB 233,120,255,255,255 ; jmpq 1e12b <_sk_load_565_sse2+0x25> + DB 233,119,255,255,255 ; jmpq 1e4bc <_sk_load_565_sse2+0x25> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,208 ; movd %eax,%xmm2 - DB 233,106,255,255,255 ; jmpq 1e12b <_sk_load_565_sse2+0x25> + DB 233,105,255,255,255 ; jmpq 1e4bc <_sk_load_565_sse2+0x25> PUBLIC _sk_load_565_dst_sse2 _sk_load_565_dst_sse2 LABEL PROC @@ -28575,42 +28733,42 @@ _sk_load_565_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,83 ; jne 1e22f <_sk_load_565_dst_sse2+0x6e> + DB 117,83 ; jne 1e5c1 <_sk_load_565_dst_sse2+0x6e> DB 243,67,15,126,52,80 ; movq (%r8,%r10,2),%xmm6 DB 102,15,97,240 ; punpcklwd %xmm0,%xmm6 - DB 102,15,111,37,226,244,0,0 ; movdqa 0xf4e2(%rip),%xmm4 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,37,64,240,0,0 ; movdqa 0xf040(%rip),%xmm4 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,37,228,244,0,0 ; mulps 0xf4e4(%rip),%xmm4 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,45,236,244,0,0 ; movdqa 0xf4ec(%rip),%xmm5 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 15,89,37,66,240,0,0 ; mulps 0xf042(%rip),%xmm4 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,45,74,240,0,0 ; movdqa 0xf04a(%rip),%xmm5 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,219,238 ; pand %xmm6,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,45,238,244,0,0 ; mulps 0xf4ee(%rip),%xmm5 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,15,219,53,246,244,0,0 ; pand 0xf4f6(%rip),%xmm6 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 15,89,45,76,240,0,0 ; mulps 0xf04c(%rip),%xmm5 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,15,219,53,84,240,0,0 ; pand 0xf054(%rip),%xmm6 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,89,53,252,244,0,0 ; mulps 0xf4fc(%rip),%xmm6 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 15,89,53,90,240,0,0 ; mulps 0xf05a(%rip),%xmm6 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,61,147,242,0,0 ; movaps 0xf293(%rip),%xmm7 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,61,241,237,0,0 ; movaps 0xedf1(%rip),%xmm7 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,50 ; je 1e26e <_sk_load_565_dst_sse2+0xad> + DB 116,51 ; je 1e601 <_sk_load_565_dst_sse2+0xae> DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1e25b <_sk_load_565_dst_sse2+0x9a> + DB 116,21 ; je 1e5ed <_sk_load_565_dst_sse2+0x9a> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,154 ; jne 1e1e6 <_sk_load_565_dst_sse2+0x25> + DB 117,154 ; jne 1e578 <_sk_load_565_dst_sse2+0x25> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,244,69 ; pshufd $0x45,%xmm4,%xmm6 DB 102,67,15,110,36,80 ; movd (%r8,%r10,2),%xmm4 - DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 + DB 242,15,112,228,212 ; pshuflw $0xd4,%xmm4,%xmm4 DB 242,15,16,244 ; movsd %xmm4,%xmm6 - DB 233,120,255,255,255 ; jmpq 1e1e6 <_sk_load_565_dst_sse2+0x25> + DB 233,119,255,255,255 ; jmpq 1e578 <_sk_load_565_dst_sse2+0x25> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,240 ; movd %eax,%xmm6 - DB 233,106,255,255,255 ; jmpq 1e1e6 <_sk_load_565_dst_sse2+0x25> + DB 233,105,255,255,255 ; jmpq 1e578 <_sk_load_565_dst_sse2+0x25> PUBLIC _sk_gather_565_sse2 _sk_gather_565_sse2 LABEL PROC @@ -28634,27 +28792,29 @@ _sk_gather_565_sse2 LABEL PROC DB 102,73,15,126,194 ; movq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,20,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm2 - DB 102,67,15,196,20,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm2 + DB 71,15,183,20,80 ; movzwl (%r8,%r10,2),%r10d + DB 71,15,183,28,88 ; movzwl (%r8,%r11,2),%r11d + DB 102,65,15,110,211 ; movd %r11d,%xmm2 + DB 102,65,15,196,210,1 ; pinsrw $0x1,%r10d,%xmm2 DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d DB 102,65,15,196,209,2 ; pinsrw $0x2,%r9d,%xmm2 DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax DB 102,15,196,208,3 ; pinsrw $0x3,%eax,%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,5,206,243,0,0 ; movdqa 0xf3ce(%rip),%xmm0 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,5,36,239,0,0 ; movdqa 0xef24(%rip),%xmm0 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,208,243,0,0 ; mulps 0xf3d0(%rip),%xmm0 # 2d6e0 <_sk_overlay_sse2_8bit+0x1188> - DB 102,15,111,13,216,243,0,0 ; movdqa 0xf3d8(%rip),%xmm1 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 15,89,5,38,239,0,0 ; mulps 0xef26(%rip),%xmm0 # 2d5d0 <_sk_overlay_sse2_8bit+0x1121> + DB 102,15,111,13,46,239,0,0 ; movdqa 0xef2e(%rip),%xmm1 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,218,243,0,0 ; mulps 0xf3da(%rip),%xmm1 # 2d700 <_sk_overlay_sse2_8bit+0x11a8> - DB 102,15,219,21,226,243,0,0 ; pand 0xf3e2(%rip),%xmm2 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 15,89,13,48,239,0,0 ; mulps 0xef30(%rip),%xmm1 # 2d5f0 <_sk_overlay_sse2_8bit+0x1141> + DB 102,15,219,21,56,239,0,0 ; pand 0xef38(%rip),%xmm2 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,232,243,0,0 ; mulps 0xf3e8(%rip),%xmm2 # 2d720 <_sk_overlay_sse2_8bit+0x11c8> + DB 15,89,21,62,239,0,0 ; mulps 0xef3e(%rip),%xmm2 # 2d610 <_sk_overlay_sse2_8bit+0x1161> DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,127,241,0,0 ; movaps 0xf17f(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,213,236,0,0 ; movaps 0xecd5(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax PUBLIC _sk_store_565_sse2 @@ -28666,12 +28826,12 @@ _sk_store_565_sse2 LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 - DB 68,15,40,5,127,245,0,0 ; movaps 0xf57f(%rip),%xmm8 # 2d8e0 <_sk_overlay_sse2_8bit+0x1388> + DB 68,15,40,5,197,240,0,0 ; movaps 0xf0c5(%rip),%xmm8 # 2d7c0 <_sk_overlay_sse2_8bit+0x1311> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 DB 102,65,15,114,241,11 ; pslld $0xb,%xmm9 - DB 68,15,40,21,116,245,0,0 ; movaps 0xf574(%rip),%xmm10 # 2d8f0 <_sk_overlay_sse2_8bit+0x1398> + DB 68,15,40,21,186,240,0,0 ; movaps 0xf0ba(%rip),%xmm10 # 2d7d0 <_sk_overlay_sse2_8bit+0x1321> DB 68,15,89,209 ; mulps %xmm1,%xmm10 DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10 DB 102,65,15,114,242,5 ; pslld $0x5,%xmm10 @@ -28681,29 +28841,29 @@ _sk_store_565_sse2 LABEL PROC DB 102,69,15,86,194 ; orpd %xmm10,%xmm8 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 - DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 + DB 102,68,15,107,192 ; packssdw %xmm0,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 1e3be <_sk_store_565_sse2+0x7b> - DB 242,71,15,17,4,80 ; movsd %xmm8,(%r8,%r10,2) + DB 117,10 ; jne 1e758 <_sk_store_565_sse2+0x7b> + DB 102,71,15,214,4,80 ; movq %xmm8,(%r8,%r10,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,38 ; je 1e3f6 <_sk_store_565_sse2+0xb3> + DB 116,38 ; je 1e790 <_sk_store_565_sse2+0xb3> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 1e3e8 <_sk_store_565_sse2+0xa5> + DB 116,18 ; je 1e782 <_sk_store_565_sse2+0xa5> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,222 ; jne 1e3ba <_sk_store_565_sse2+0x77> + DB 117,222 ; jne 1e754 <_sk_store_565_sse2+0x77> DB 102,65,15,197,192,4 ; pextrw $0x4,%xmm8,%eax DB 102,67,137,68,80,4 ; mov %ax,0x4(%r8,%r10,2) DB 242,69,15,112,192,232 ; pshuflw $0xe8,%xmm8,%xmm8 DB 102,71,15,126,4,80 ; movd %xmm8,(%r8,%r10,2) - DB 235,196 ; jmp 1e3ba <_sk_store_565_sse2+0x77> + DB 235,196 ; jmp 1e754 <_sk_store_565_sse2+0x77> DB 102,68,15,126,192 ; movd %xmm8,%eax DB 102,67,137,4,80 ; mov %ax,(%r8,%r10,2) - DB 235,184 ; jmp 1e3ba <_sk_store_565_sse2+0x77> + DB 235,184 ; jmp 1e754 <_sk_store_565_sse2+0x77> PUBLIC _sk_load_4444_sse2 _sk_load_4444_sse2 LABEL PROC @@ -28715,45 +28875,45 @@ _sk_load_4444_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,98 ; jne 1e47f <_sk_load_4444_sse2+0x7d> + DB 117,98 ; jne 1e819 <_sk_load_4444_sse2+0x7d> DB 243,67,15,126,28,80 ; movq (%r8,%r10,2),%xmm3 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,209,244,0,0 ; movdqa 0xf4d1(%rip),%xmm0 # 2d900 <_sk_overlay_sse2_8bit+0x13a8> + DB 102,15,111,5,23,240,0,0 ; movdqa 0xf017(%rip),%xmm0 # 2d7e0 <_sk_overlay_sse2_8bit+0x1331> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,211,244,0,0 ; mulps 0xf4d3(%rip),%xmm0 # 2d910 <_sk_overlay_sse2_8bit+0x13b8> - DB 102,15,111,13,219,244,0,0 ; movdqa 0xf4db(%rip),%xmm1 # 2d920 <_sk_overlay_sse2_8bit+0x13c8> + DB 15,89,5,25,240,0,0 ; mulps 0xf019(%rip),%xmm0 # 2d7f0 <_sk_overlay_sse2_8bit+0x1341> + DB 102,15,111,13,33,240,0,0 ; movdqa 0xf021(%rip),%xmm1 # 2d800 <_sk_overlay_sse2_8bit+0x1351> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,221,244,0,0 ; mulps 0xf4dd(%rip),%xmm1 # 2d930 <_sk_overlay_sse2_8bit+0x13d8> - DB 102,15,111,21,229,244,0,0 ; movdqa 0xf4e5(%rip),%xmm2 # 2d940 <_sk_overlay_sse2_8bit+0x13e8> + DB 15,89,13,35,240,0,0 ; mulps 0xf023(%rip),%xmm1 # 2d810 <_sk_overlay_sse2_8bit+0x1361> + DB 102,15,111,21,43,240,0,0 ; movdqa 0xf02b(%rip),%xmm2 # 2d820 <_sk_overlay_sse2_8bit+0x1371> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,231,244,0,0 ; mulps 0xf4e7(%rip),%xmm2 # 2d950 <_sk_overlay_sse2_8bit+0x13f8> - DB 102,15,219,29,239,244,0,0 ; pand 0xf4ef(%rip),%xmm3 # 2d960 <_sk_overlay_sse2_8bit+0x1408> + DB 15,89,21,45,240,0,0 ; mulps 0xf02d(%rip),%xmm2 # 2d830 <_sk_overlay_sse2_8bit+0x1381> + DB 102,15,219,29,53,240,0,0 ; pand 0xf035(%rip),%xmm3 # 2d840 <_sk_overlay_sse2_8bit+0x1391> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,245,244,0,0 ; mulps 0xf4f5(%rip),%xmm3 # 2d970 <_sk_overlay_sse2_8bit+0x1418> + DB 15,89,29,59,240,0,0 ; mulps 0xf03b(%rip),%xmm3 # 2d850 <_sk_overlay_sse2_8bit+0x13a1> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,50 ; je 1e4be <_sk_load_4444_sse2+0xbc> + DB 116,51 ; je 1e859 <_sk_load_4444_sse2+0xbd> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1e4ab <_sk_load_4444_sse2+0xa9> + DB 116,21 ; je 1e845 <_sk_load_4444_sse2+0xa9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,139 ; jne 1e427 <_sk_load_4444_sse2+0x25> + DB 117,139 ; jne 1e7c1 <_sk_load_4444_sse2+0x25> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 102,67,15,110,4,80 ; movd (%r8,%r10,2),%xmm0 - DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 + DB 242,15,112,192,212 ; pshuflw $0xd4,%xmm0,%xmm0 DB 242,15,16,216 ; movsd %xmm0,%xmm3 - DB 233,105,255,255,255 ; jmpq 1e427 <_sk_load_4444_sse2+0x25> + DB 233,104,255,255,255 ; jmpq 1e7c1 <_sk_load_4444_sse2+0x25> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,216 ; movd %eax,%xmm3 - DB 233,91,255,255,255 ; jmpq 1e427 <_sk_load_4444_sse2+0x25> + DB 233,90,255,255,255 ; jmpq 1e7c1 <_sk_load_4444_sse2+0x25> PUBLIC _sk_load_4444_dst_sse2 _sk_load_4444_dst_sse2 LABEL PROC @@ -28765,45 +28925,45 @@ _sk_load_4444_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 DB 72,133,255 ; test %rdi,%rdi - DB 117,98 ; jne 1e549 <_sk_load_4444_dst_sse2+0x7d> + DB 117,98 ; jne 1e8e4 <_sk_load_4444_dst_sse2+0x7d> DB 243,67,15,126,60,80 ; movq (%r8,%r10,2),%xmm7 DB 102,15,97,248 ; punpcklwd %xmm0,%xmm7 - DB 102,15,111,37,7,244,0,0 ; movdqa 0xf407(%rip),%xmm4 # 2d900 <_sk_overlay_sse2_8bit+0x13a8> + DB 102,15,111,37,76,239,0,0 ; movdqa 0xef4c(%rip),%xmm4 # 2d7e0 <_sk_overlay_sse2_8bit+0x1331> DB 102,15,219,231 ; pand %xmm7,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,37,9,244,0,0 ; mulps 0xf409(%rip),%xmm4 # 2d910 <_sk_overlay_sse2_8bit+0x13b8> - DB 102,15,111,45,17,244,0,0 ; movdqa 0xf411(%rip),%xmm5 # 2d920 <_sk_overlay_sse2_8bit+0x13c8> + DB 15,89,37,78,239,0,0 ; mulps 0xef4e(%rip),%xmm4 # 2d7f0 <_sk_overlay_sse2_8bit+0x1341> + DB 102,15,111,45,86,239,0,0 ; movdqa 0xef56(%rip),%xmm5 # 2d800 <_sk_overlay_sse2_8bit+0x1351> DB 102,15,219,239 ; pand %xmm7,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,45,19,244,0,0 ; mulps 0xf413(%rip),%xmm5 # 2d930 <_sk_overlay_sse2_8bit+0x13d8> - DB 102,15,111,53,27,244,0,0 ; movdqa 0xf41b(%rip),%xmm6 # 2d940 <_sk_overlay_sse2_8bit+0x13e8> + DB 15,89,45,88,239,0,0 ; mulps 0xef58(%rip),%xmm5 # 2d810 <_sk_overlay_sse2_8bit+0x1361> + DB 102,15,111,53,96,239,0,0 ; movdqa 0xef60(%rip),%xmm6 # 2d820 <_sk_overlay_sse2_8bit+0x1371> DB 102,15,219,247 ; pand %xmm7,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,89,53,29,244,0,0 ; mulps 0xf41d(%rip),%xmm6 # 2d950 <_sk_overlay_sse2_8bit+0x13f8> - DB 102,15,219,61,37,244,0,0 ; pand 0xf425(%rip),%xmm7 # 2d960 <_sk_overlay_sse2_8bit+0x1408> + DB 15,89,53,98,239,0,0 ; mulps 0xef62(%rip),%xmm6 # 2d830 <_sk_overlay_sse2_8bit+0x1381> + DB 102,15,219,61,106,239,0,0 ; pand 0xef6a(%rip),%xmm7 # 2d840 <_sk_overlay_sse2_8bit+0x1391> DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 - DB 15,89,61,43,244,0,0 ; mulps 0xf42b(%rip),%xmm7 # 2d970 <_sk_overlay_sse2_8bit+0x1418> + DB 15,89,61,112,239,0,0 ; mulps 0xef70(%rip),%xmm7 # 2d850 <_sk_overlay_sse2_8bit+0x13a1> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,50 ; je 1e588 <_sk_load_4444_dst_sse2+0xbc> + DB 116,51 ; je 1e924 <_sk_load_4444_dst_sse2+0xbd> DB 102,15,239,255 ; pxor %xmm7,%xmm7 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,21 ; je 1e575 <_sk_load_4444_dst_sse2+0xa9> + DB 116,21 ; je 1e910 <_sk_load_4444_dst_sse2+0xa9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,139 ; jne 1e4f1 <_sk_load_4444_dst_sse2+0x25> + DB 117,139 ; jne 1e88c <_sk_load_4444_dst_sse2+0x25> DB 67,15,183,68,80,4 ; movzwl 0x4(%r8,%r10,2),%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,252,69 ; pshufd $0x45,%xmm4,%xmm7 DB 102,67,15,110,36,80 ; movd (%r8,%r10,2),%xmm4 - DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 + DB 242,15,112,228,212 ; pshuflw $0xd4,%xmm4,%xmm4 DB 242,15,16,252 ; movsd %xmm4,%xmm7 - DB 233,105,255,255,255 ; jmpq 1e4f1 <_sk_load_4444_dst_sse2+0x25> + DB 233,104,255,255,255 ; jmpq 1e88c <_sk_load_4444_dst_sse2+0x25> DB 67,15,183,4,80 ; movzwl (%r8,%r10,2),%eax DB 102,15,110,248 ; movd %eax,%xmm7 - DB 233,91,255,255,255 ; jmpq 1e4f1 <_sk_load_4444_dst_sse2+0x25> + DB 233,90,255,255,255 ; jmpq 1e88c <_sk_load_4444_dst_sse2+0x25> PUBLIC _sk_gather_4444_sse2 _sk_gather_4444_sse2 LABEL PROC @@ -28827,29 +28987,31 @@ _sk_gather_4444_sse2 LABEL PROC DB 102,73,15,126,194 ; movq %xmm0,%r10 DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 - DB 102,67,15,196,28,88,0 ; pinsrw $0x0,(%r8,%r11,2),%xmm3 - DB 102,67,15,196,28,80,1 ; pinsrw $0x1,(%r8,%r10,2),%xmm3 + DB 71,15,183,20,80 ; movzwl (%r8,%r10,2),%r10d + DB 71,15,183,28,88 ; movzwl (%r8,%r11,2),%r11d + DB 102,65,15,110,219 ; movd %r11d,%xmm3 + DB 102,65,15,196,218,1 ; pinsrw $0x1,%r10d,%xmm3 DB 71,15,183,12,72 ; movzwl (%r8,%r9,2),%r9d DB 102,65,15,196,217,2 ; pinsrw $0x2,%r9d,%xmm3 DB 65,15,183,4,64 ; movzwl (%r8,%rax,2),%eax DB 102,15,196,216,3 ; pinsrw $0x3,%eax,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,5,228,242,0,0 ; movdqa 0xf2e4(%rip),%xmm0 # 2d900 <_sk_overlay_sse2_8bit+0x13a8> + DB 102,15,111,5,33,238,0,0 ; movdqa 0xee21(%rip),%xmm0 # 2d7e0 <_sk_overlay_sse2_8bit+0x1331> DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,5,230,242,0,0 ; mulps 0xf2e6(%rip),%xmm0 # 2d910 <_sk_overlay_sse2_8bit+0x13b8> - DB 102,15,111,13,238,242,0,0 ; movdqa 0xf2ee(%rip),%xmm1 # 2d920 <_sk_overlay_sse2_8bit+0x13c8> + DB 15,89,5,35,238,0,0 ; mulps 0xee23(%rip),%xmm0 # 2d7f0 <_sk_overlay_sse2_8bit+0x1341> + DB 102,15,111,13,43,238,0,0 ; movdqa 0xee2b(%rip),%xmm1 # 2d800 <_sk_overlay_sse2_8bit+0x1351> DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,13,240,242,0,0 ; mulps 0xf2f0(%rip),%xmm1 # 2d930 <_sk_overlay_sse2_8bit+0x13d8> - DB 102,15,111,21,248,242,0,0 ; movdqa 0xf2f8(%rip),%xmm2 # 2d940 <_sk_overlay_sse2_8bit+0x13e8> + DB 15,89,13,45,238,0,0 ; mulps 0xee2d(%rip),%xmm1 # 2d810 <_sk_overlay_sse2_8bit+0x1361> + DB 102,15,111,21,53,238,0,0 ; movdqa 0xee35(%rip),%xmm2 # 2d820 <_sk_overlay_sse2_8bit+0x1371> DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,21,250,242,0,0 ; mulps 0xf2fa(%rip),%xmm2 # 2d950 <_sk_overlay_sse2_8bit+0x13f8> - DB 102,15,219,29,2,243,0,0 ; pand 0xf302(%rip),%xmm3 # 2d960 <_sk_overlay_sse2_8bit+0x1408> + DB 15,89,21,55,238,0,0 ; mulps 0xee37(%rip),%xmm2 # 2d830 <_sk_overlay_sse2_8bit+0x1381> + DB 102,15,219,29,63,238,0,0 ; pand 0xee3f(%rip),%xmm3 # 2d840 <_sk_overlay_sse2_8bit+0x1391> DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,29,8,243,0,0 ; mulps 0xf308(%rip),%xmm3 # 2d970 <_sk_overlay_sse2_8bit+0x1418> + DB 15,89,29,69,238,0,0 ; mulps 0xee45(%rip),%xmm3 # 2d850 <_sk_overlay_sse2_8bit+0x13a1> DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -28862,7 +29024,7 @@ _sk_store_4444_sse2 LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,210 ; movslq %edx,%r10 - DB 68,15,40,5,246,242,0,0 ; movaps 0xf2f6(%rip),%xmm8 # 2d980 <_sk_overlay_sse2_8bit+0x1428> + DB 68,15,40,5,51,238,0,0 ; movaps 0xee33(%rip),%xmm8 # 2d860 <_sk_overlay_sse2_8bit+0x13b1> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -28882,29 +29044,29 @@ _sk_store_4444_sse2 LABEL PROC DB 102,69,15,86,194 ; orpd %xmm10,%xmm8 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 - DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 + DB 102,68,15,107,192 ; packssdw %xmm0,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 1e6fb <_sk_store_4444_sse2+0x8f> - DB 242,71,15,17,4,80 ; movsd %xmm8,(%r8,%r10,2) + DB 117,10 ; jne 1ea9e <_sk_store_4444_sse2+0x8f> + DB 102,71,15,214,4,80 ; movq %xmm8,(%r8,%r10,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 102,68,15,97,192 ; punpcklwd %xmm0,%xmm8 DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,38 ; je 1e733 <_sk_store_4444_sse2+0xc7> + DB 116,38 ; je 1ead6 <_sk_store_4444_sse2+0xc7> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,18 ; je 1e725 <_sk_store_4444_sse2+0xb9> + DB 116,18 ; je 1eac8 <_sk_store_4444_sse2+0xb9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,222 ; jne 1e6f7 <_sk_store_4444_sse2+0x8b> + DB 117,222 ; jne 1ea9a <_sk_store_4444_sse2+0x8b> DB 102,65,15,197,192,4 ; pextrw $0x4,%xmm8,%eax DB 102,67,137,68,80,4 ; mov %ax,0x4(%r8,%r10,2) DB 242,69,15,112,192,232 ; pshuflw $0xe8,%xmm8,%xmm8 DB 102,71,15,126,4,80 ; movd %xmm8,(%r8,%r10,2) - DB 235,196 ; jmp 1e6f7 <_sk_store_4444_sse2+0x8b> + DB 235,196 ; jmp 1ea9a <_sk_store_4444_sse2+0x8b> DB 102,68,15,126,192 ; movd %xmm8,%eax DB 102,67,137,4,80 ; mov %ax,(%r8,%r10,2) - DB 235,184 ; jmp 1e6f7 <_sk_store_4444_sse2+0x8b> + DB 235,184 ; jmp 1ea9a <_sk_store_4444_sse2+0x8b> PUBLIC _sk_load_8888_sse2 _sk_load_8888_sse2 LABEL PROC @@ -28916,13 +29078,13 @@ _sk_load_8888_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,98 ; jne 1e7bd <_sk_load_8888_sse2+0x7e> + DB 117,98 ; jne 1eb60 <_sk_load_8888_sse2+0x7e> DB 243,69,15,111,12,128 ; movdqu (%r8,%rax,4),%xmm9 - DB 102,15,111,21,7,238,0,0 ; movdqa 0xee07(%rip),%xmm2 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,21,84,233,0,0 ; movdqa 0xe954(%rip),%xmm2 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,67,239,0,0 ; movaps 0xef43(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,144,234,0,0 ; movaps 0xea90(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -28942,18 +29104,18 @@ _sk_load_8888_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,41 ; je 1e7f3 <_sk_load_8888_sse2+0xb4> + DB 116,41 ; je 1eb96 <_sk_load_8888_sse2+0xb4> DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,19 ; je 1e7e8 <_sk_load_8888_sse2+0xa9> + DB 116,19 ; je 1eb8b <_sk_load_8888_sse2+0xa9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,134 ; jne 1e761 <_sk_load_8888_sse2+0x22> + DB 117,134 ; jne 1eb04 <_sk_load_8888_sse2+0x22> DB 102,65,15,110,68,128,8 ; movd 0x8(%r8,%rax,4),%xmm0 DB 102,68,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm9 DB 102,69,15,18,12,128 ; movlpd (%r8,%rax,4),%xmm9 - DB 233,110,255,255,255 ; jmpq 1e761 <_sk_load_8888_sse2+0x22> + DB 233,110,255,255,255 ; jmpq 1eb04 <_sk_load_8888_sse2+0x22> DB 102,69,15,110,12,128 ; movd (%r8,%rax,4),%xmm9 - DB 233,99,255,255,255 ; jmpq 1e761 <_sk_load_8888_sse2+0x22> + DB 233,99,255,255,255 ; jmpq 1eb04 <_sk_load_8888_sse2+0x22> PUBLIC _sk_load_8888_dst_sse2 _sk_load_8888_dst_sse2 LABEL PROC @@ -28965,13 +29127,13 @@ _sk_load_8888_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,98 ; jne 1e87c <_sk_load_8888_dst_sse2+0x7e> + DB 117,98 ; jne 1ec1f <_sk_load_8888_dst_sse2+0x7e> DB 243,69,15,111,12,128 ; movdqu (%r8,%rax,4),%xmm9 - DB 102,15,111,53,72,237,0,0 ; movdqa 0xed48(%rip),%xmm6 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,53,149,232,0,0 ; movdqa 0xe895(%rip),%xmm6 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,225 ; movdqa %xmm9,%xmm4 DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 68,15,40,5,132,238,0,0 ; movaps 0xee84(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,209,233,0,0 ; movaps 0xe9d1(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,224 ; mulps %xmm8,%xmm4 DB 102,65,15,111,233 ; movdqa %xmm9,%xmm5 DB 102,15,114,213,8 ; psrld $0x8,%xmm5 @@ -28991,18 +29153,18 @@ _sk_load_8888_dst_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,41 ; je 1e8b2 <_sk_load_8888_dst_sse2+0xb4> + DB 116,41 ; je 1ec55 <_sk_load_8888_dst_sse2+0xb4> DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,19 ; je 1e8a7 <_sk_load_8888_dst_sse2+0xa9> + DB 116,19 ; je 1ec4a <_sk_load_8888_dst_sse2+0xa9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,134 ; jne 1e820 <_sk_load_8888_dst_sse2+0x22> + DB 117,134 ; jne 1ebc3 <_sk_load_8888_dst_sse2+0x22> DB 102,65,15,110,100,128,8 ; movd 0x8(%r8,%rax,4),%xmm4 DB 102,68,15,112,204,69 ; pshufd $0x45,%xmm4,%xmm9 DB 102,69,15,18,12,128 ; movlpd (%r8,%rax,4),%xmm9 - DB 233,110,255,255,255 ; jmpq 1e820 <_sk_load_8888_dst_sse2+0x22> + DB 233,110,255,255,255 ; jmpq 1ebc3 <_sk_load_8888_dst_sse2+0x22> DB 102,69,15,110,12,128 ; movd (%r8,%rax,4),%xmm9 - DB 233,99,255,255,255 ; jmpq 1e820 <_sk_load_8888_dst_sse2+0x22> + DB 233,99,255,255,255 ; jmpq 1ebc3 <_sk_load_8888_dst_sse2+0x22> PUBLIC _sk_gather_8888_sse2 _sk_gather_8888_sse2 LABEL PROC @@ -29027,17 +29189,17 @@ _sk_gather_8888_sse2 LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 102,67,15,110,4,144 ; movd (%r8,%r10,4),%xmm0 - DB 102,65,15,110,12,128 ; movd (%r8,%rax,4),%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 DB 102,71,15,110,12,152 ; movd (%r8,%r11,4),%xmm9 - DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 - DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,21,50,236,0,0 ; movdqa 0xec32(%rip),%xmm2 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,65,15,110,4,128 ; movd (%r8,%rax,4),%xmm0 + DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 + DB 102,15,98,200 ; punpckldq %xmm0,%xmm1 + DB 102,68,15,108,201 ; punpcklqdq %xmm1,%xmm9 + DB 102,15,111,21,127,231,0,0 ; movdqa 0xe77f(%rip),%xmm2 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 68,15,40,5,110,237,0,0 ; movaps 0xed6e(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,187,232,0,0 ; movaps 0xe8bb(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -29064,7 +29226,7 @@ _sk_store_8888_sse2 LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 68,15,40,5,239,235,0,0 ; movaps 0xebef(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,60,231,0,0 ; movaps 0xe73c(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -29083,24 +29245,24 @@ _sk_store_8888_sse2 LABEL PROC DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 102,69,15,235,194 ; por %xmm10,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 1ea11 <_sk_store_8888_sse2+0x7f> + DB 117,10 ; jne 1edb4 <_sk_store_8888_sse2+0x7f> DB 243,69,15,127,4,128 ; movdqu %xmm8,(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,33 ; je 1ea3f <_sk_store_8888_sse2+0xad> + DB 116,33 ; je 1ede2 <_sk_store_8888_sse2+0xad> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,19 ; je 1ea37 <_sk_store_8888_sse2+0xa5> + DB 116,19 ; je 1edda <_sk_store_8888_sse2+0xa5> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,227 ; jne 1ea0d <_sk_store_8888_sse2+0x7b> + DB 117,227 ; jne 1edb0 <_sk_store_8888_sse2+0x7b> DB 102,69,15,112,200,78 ; pshufd $0x4e,%xmm8,%xmm9 DB 102,69,15,126,76,128,8 ; movd %xmm9,0x8(%r8,%rax,4) DB 102,69,15,214,4,128 ; movq %xmm8,(%r8,%rax,4) - DB 235,206 ; jmp 1ea0d <_sk_store_8888_sse2+0x7b> + DB 235,206 ; jmp 1edb0 <_sk_store_8888_sse2+0x7b> DB 102,69,15,126,4,128 ; movd %xmm8,(%r8,%rax,4) - DB 235,198 ; jmp 1ea0d <_sk_store_8888_sse2+0x7b> + DB 235,198 ; jmp 1edb0 <_sk_store_8888_sse2+0x7b> PUBLIC _sk_load_bgra_sse2 _sk_load_bgra_sse2 LABEL PROC @@ -29112,13 +29274,13 @@ _sk_load_bgra_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,98 ; jne 1eac5 <_sk_load_bgra_sse2+0x7e> + DB 117,98 ; jne 1ee68 <_sk_load_bgra_sse2+0x7e> DB 243,69,15,111,12,128 ; movdqu (%r8,%rax,4),%xmm9 - DB 102,15,111,5,255,234,0,0 ; movdqa 0xeaff(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,5,76,230,0,0 ; movdqa 0xe64c(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,219,200 ; pand %xmm0,%xmm1 DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2 - DB 68,15,40,5,59,236,0,0 ; movaps 0xec3b(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,136,231,0,0 ; movaps 0xe788(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -29138,18 +29300,18 @@ _sk_load_bgra_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,41 ; je 1eafb <_sk_load_bgra_sse2+0xb4> + DB 116,41 ; je 1ee9e <_sk_load_bgra_sse2+0xb4> DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,19 ; je 1eaf0 <_sk_load_bgra_sse2+0xa9> + DB 116,19 ; je 1ee93 <_sk_load_bgra_sse2+0xa9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,134 ; jne 1ea69 <_sk_load_bgra_sse2+0x22> + DB 117,134 ; jne 1ee0c <_sk_load_bgra_sse2+0x22> DB 102,65,15,110,68,128,8 ; movd 0x8(%r8,%rax,4),%xmm0 DB 102,68,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm9 DB 102,69,15,18,12,128 ; movlpd (%r8,%rax,4),%xmm9 - DB 233,110,255,255,255 ; jmpq 1ea69 <_sk_load_bgra_sse2+0x22> + DB 233,110,255,255,255 ; jmpq 1ee0c <_sk_load_bgra_sse2+0x22> DB 102,69,15,110,12,128 ; movd (%r8,%rax,4),%xmm9 - DB 233,99,255,255,255 ; jmpq 1ea69 <_sk_load_bgra_sse2+0x22> + DB 233,99,255,255,255 ; jmpq 1ee0c <_sk_load_bgra_sse2+0x22> PUBLIC _sk_load_bgra_dst_sse2 _sk_load_bgra_dst_sse2 LABEL PROC @@ -29161,13 +29323,13 @@ _sk_load_bgra_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 117,98 ; jne 1eb84 <_sk_load_bgra_dst_sse2+0x7e> + DB 117,98 ; jne 1ef27 <_sk_load_bgra_dst_sse2+0x7e> DB 243,69,15,111,12,128 ; movdqu (%r8,%rax,4),%xmm9 - DB 102,15,111,37,64,234,0,0 ; movdqa 0xea40(%rip),%xmm4 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,15,111,37,141,229,0,0 ; movdqa 0xe58d(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,233 ; movdqa %xmm9,%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 15,91,245 ; cvtdq2ps %xmm5,%xmm6 - DB 68,15,40,5,124,235,0,0 ; movaps 0xeb7c(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,201,230,0,0 ; movaps 0xe6c9(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,240 ; mulps %xmm8,%xmm6 DB 102,65,15,111,233 ; movdqa %xmm9,%xmm5 DB 102,15,114,213,8 ; psrld $0x8,%xmm5 @@ -29187,18 +29349,18 @@ _sk_load_bgra_dst_sse2 LABEL PROC DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,41 ; je 1ebba <_sk_load_bgra_dst_sse2+0xb4> + DB 116,41 ; je 1ef5d <_sk_load_bgra_dst_sse2+0xb4> DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,19 ; je 1ebaf <_sk_load_bgra_dst_sse2+0xa9> + DB 116,19 ; je 1ef52 <_sk_load_bgra_dst_sse2+0xa9> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,134 ; jne 1eb28 <_sk_load_bgra_dst_sse2+0x22> + DB 117,134 ; jne 1eecb <_sk_load_bgra_dst_sse2+0x22> DB 102,65,15,110,100,128,8 ; movd 0x8(%r8,%rax,4),%xmm4 DB 102,68,15,112,204,69 ; pshufd $0x45,%xmm4,%xmm9 DB 102,69,15,18,12,128 ; movlpd (%r8,%rax,4),%xmm9 - DB 233,110,255,255,255 ; jmpq 1eb28 <_sk_load_bgra_dst_sse2+0x22> + DB 233,110,255,255,255 ; jmpq 1eecb <_sk_load_bgra_dst_sse2+0x22> DB 102,69,15,110,12,128 ; movd (%r8,%rax,4),%xmm9 - DB 233,99,255,255,255 ; jmpq 1eb28 <_sk_load_bgra_dst_sse2+0x22> + DB 233,99,255,255,255 ; jmpq 1eecb <_sk_load_bgra_dst_sse2+0x22> PUBLIC _sk_gather_bgra_sse2 _sk_gather_bgra_sse2 LABEL PROC @@ -29223,17 +29385,17 @@ _sk_gather_bgra_sse2 LABEL PROC DB 69,137,211 ; mov %r10d,%r11d DB 73,193,234,32 ; shr $0x20,%r10 DB 102,67,15,110,4,144 ; movd (%r8,%r10,4),%xmm0 - DB 102,65,15,110,12,128 ; movd (%r8,%rax,4),%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 DB 102,71,15,110,12,152 ; movd (%r8,%r11,4),%xmm9 - DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 - DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 - DB 102,15,111,5,42,233,0,0 ; movdqa 0xe92a(%rip),%xmm0 # 2d570 <_sk_overlay_sse2_8bit+0x1018> + DB 102,65,15,110,4,128 ; movd (%r8,%rax,4),%xmm0 + DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 + DB 102,15,98,200 ; punpckldq %xmm0,%xmm1 + DB 102,68,15,108,201 ; punpcklqdq %xmm1,%xmm9 + DB 102,15,111,5,119,228,0,0 ; movdqa 0xe477(%rip),%xmm0 # 2d460 <_sk_overlay_sse2_8bit+0xfb1> DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,219,200 ; pand %xmm0,%xmm1 DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2 - DB 68,15,40,5,102,234,0,0 ; movaps 0xea66(%rip),%xmm8 # 2d6c0 <_sk_overlay_sse2_8bit+0x1168> + DB 68,15,40,5,179,229,0,0 ; movaps 0xe5b3(%rip),%xmm8 # 2d5b0 <_sk_overlay_sse2_8bit+0x1101> DB 65,15,89,208 ; mulps %xmm8,%xmm2 DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -29260,7 +29422,7 @@ _sk_store_bgra_sse2 LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 68,15,40,5,231,232,0,0 ; movaps 0xe8e7(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x1048> + DB 68,15,40,5,52,228,0,0 ; movaps 0xe434(%rip),%xmm8 # 2d490 <_sk_overlay_sse2_8bit+0xfe1> DB 68,15,40,202 ; movaps %xmm2,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 @@ -29279,24 +29441,24 @@ _sk_store_bgra_sse2 LABEL PROC DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 102,69,15,235,194 ; por %xmm10,%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 1ed19 <_sk_store_bgra_sse2+0x7f> + DB 117,10 ; jne 1f0bc <_sk_store_bgra_sse2+0x7f> DB 243,69,15,127,4,128 ; movdqu %xmm8,(%r8,%rax,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,3 ; and $0x3,%r9b DB 65,128,249,1 ; cmp $0x1,%r9b - DB 116,33 ; je 1ed47 <_sk_store_bgra_sse2+0xad> + DB 116,33 ; je 1f0ea <_sk_store_bgra_sse2+0xad> DB 65,128,249,2 ; cmp $0x2,%r9b - DB 116,19 ; je 1ed3f <_sk_store_bgra_sse2+0xa5> + DB 116,19 ; je 1f0e2 <_sk_store_bgra_sse2+0xa5> DB 65,128,249,3 ; cmp $0x3,%r9b - DB 117,227 ; jne 1ed15 <_sk_store_bgra_sse2+0x7b> + DB 117,227 ; jne 1f0b8 <_sk_store_bgra_sse2+0x7b> DB 102,69,15,112,200,78 ; pshufd $0x4e,%xmm8,%xmm9 DB 102,69,15,126,76,128,8 ; movd %xmm9,0x8(%r8,%rax,4) DB 102,69,15,214,4,128 ; movq %xmm8,(%r8,%rax,4) - DB 235,206 ; jmp 1ed15 <_sk_store_bgra_sse2+0x7b> + DB 235,206 ; jmp 1f0b8 <_sk_store_bgra_sse2+0x7b> DB 102,69,15,126,4,128 ; movd %xmm8,(%r8,%rax,4) - DB 235,198 ; jmp 1ed15 <_sk_store_bgra_sse2+0x7b> + DB 235,198 ; jmp 1f0b8 <_sk_store_bgra_sse2+0x7b> PUBLIC _sk_load_f16_sse2 _sk_load_f16_sse2 LABEL PROC @@ -29308,89 +29470,78 @@ _sk_load_f16_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,98,1,0,0 ; jne 1eed1 <_sk_load_f16_sse2+0x182> + DB 15,133,47,1,0,0 ; jne 1f241 <_sk_load_f16_sse2+0x14f> DB 102,65,15,16,4,192 ; movupd (%r8,%rax,8),%xmm0 - DB 102,65,15,16,76,192,16 ; movupd 0x10(%r8,%rax,8),%xmm1 + DB 243,65,15,111,76,192,16 ; movdqu 0x10(%r8,%rax,8),%xmm1 DB 102,68,15,40,192 ; movapd %xmm0,%xmm8 DB 102,68,15,97,193 ; punpcklwd %xmm1,%xmm8 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 - DB 102,68,15,97,240 ; punpcklwd %xmm0,%xmm14 + DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 + DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 DB 102,68,15,105,192 ; punpckhwd %xmm0,%xmm8 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 - DB 102,68,15,111,13,223,235,0,0 ; movdqa 0xebdf(%rip),%xmm9 # 2d990 <_sk_overlay_sse2_8bit+0x1438> - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,65,15,219,193 ; pand %xmm9,%xmm0 - DB 102,15,239,200 ; pxor %xmm0,%xmm1 - DB 102,15,114,240,16 ; pslld $0x10,%xmm0 - DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 - DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 - DB 102,68,15,235,232 ; por %xmm0,%xmm13 - DB 102,68,15,111,29,212,235,0,0 ; movdqa 0xebd4(%rip),%xmm11 # 2d9b0 <_sk_overlay_sse2_8bit+0x1458> - DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,214,235,0,0 ; movdqa 0xebd6(%rip),%xmm12 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,121,237,0,0 ; movdqa 0xed79(%rip),%xmm3 # 2db70 <_sk_overlay_sse2_8bit+0x1618> + DB 102,68,15,111,13,30,231,0,0 ; movdqa 0xe71e(%rip),%xmm9 # 2d870 <_sk_overlay_sse2_8bit+0x13c1> + DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 + DB 102,69,15,219,225 ; pand %xmm9,%xmm12 + DB 102,68,15,111,29,27,231,0,0 ; movdqa 0xe71b(%rip),%xmm11 # 2d880 <_sk_overlay_sse2_8bit+0x13d1> + DB 102,65,15,219,203 ; pand %xmm11,%xmm1 + DB 102,15,111,29,30,231,0,0 ; movdqa 0xe71e(%rip),%xmm3 # 2d890 <_sk_overlay_sse2_8bit+0x13e1> + DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0 - DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 - DB 102,65,15,115,222,8 ; psrldq $0x8,%xmm14 - DB 102,69,15,97,242 ; punpcklwd %xmm10,%xmm14 - DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 - DB 102,65,15,219,201 ; pand %xmm9,%xmm1 - DB 102,68,15,239,241 ; pxor %xmm1,%xmm14 - DB 102,15,114,241,16 ; pslld $0x10,%xmm1 - DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 - DB 102,15,114,242,13 ; pslld $0xd,%xmm2 - DB 102,15,235,209 ; por %xmm1,%xmm2 - DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,69,15,239,244 ; pxor %xmm12,%xmm14 + DB 102,15,114,241,13 ; pslld $0xd,%xmm1 + DB 102,65,15,235,204 ; por %xmm12,%xmm1 + DB 102,68,15,111,37,13,231,0,0 ; movdqa 0xe70d(%rip),%xmm12 # 2d8a0 <_sk_overlay_sse2_8bit+0x13f1> + DB 102,65,15,254,204 ; paddd %xmm12,%xmm1 + DB 102,15,223,193 ; pandn %xmm1,%xmm0 + DB 102,65,15,105,210 ; punpckhwd %xmm10,%xmm2 + DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 + DB 102,69,15,219,233 ; pand %xmm9,%xmm13 + DB 102,65,15,219,211 ; pand %xmm11,%xmm2 + DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,102,206 ; pcmpgtd %xmm14,%xmm1 + DB 102,15,102,202 ; pcmpgtd %xmm2,%xmm1 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,65,15,235,213 ; por %xmm13,%xmm2 + DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 DB 102,15,223,202 ; pandn %xmm2,%xmm1 DB 102,69,15,111,232 ; movdqa %xmm8,%xmm13 DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 - DB 102,65,15,111,213 ; movdqa %xmm13,%xmm2 - DB 102,65,15,219,209 ; pand %xmm9,%xmm2 - DB 102,68,15,239,234 ; pxor %xmm2,%xmm13 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 DB 102,69,15,111,245 ; movdqa %xmm13,%xmm14 - DB 102,65,15,114,246,13 ; pslld $0xd,%xmm14 - DB 102,68,15,235,242 ; por %xmm2,%xmm14 - DB 102,69,15,254,243 ; paddd %xmm11,%xmm14 - DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 + DB 102,69,15,219,241 ; pand %xmm9,%xmm14 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 102,65,15,114,246,16 ; pslld $0x10,%xmm14 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 DB 102,65,15,102,213 ; pcmpgtd %xmm13,%xmm2 - DB 102,65,15,223,214 ; pandn %xmm14,%xmm2 - DB 102,65,15,115,216,8 ; psrldq $0x8,%xmm8 - DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8 + DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 + DB 102,69,15,235,238 ; por %xmm14,%xmm13 + DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 + DB 102,65,15,223,213 ; pandn %xmm13,%xmm2 + DB 102,69,15,105,194 ; punpckhwd %xmm10,%xmm8 DB 102,69,15,219,200 ; pand %xmm8,%xmm9 - DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 + DB 102,69,15,219,195 ; pand %xmm11,%xmm8 DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 - DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 - DB 102,65,15,114,242,13 ; pslld $0xd,%xmm10 - DB 102,69,15,235,209 ; por %xmm9,%xmm10 - DB 102,69,15,254,211 ; paddd %xmm11,%xmm10 - DB 102,69,15,239,196 ; pxor %xmm12,%xmm8 DB 102,65,15,102,216 ; pcmpgtd %xmm8,%xmm3 - DB 102,65,15,223,218 ; pandn %xmm10,%xmm3 + DB 102,65,15,114,240,13 ; pslld $0xd,%xmm8 + DB 102,69,15,235,193 ; por %xmm9,%xmm8 + DB 102,69,15,254,196 ; paddd %xmm12,%xmm8 + DB 102,65,15,223,216 ; pandn %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 242,65,15,16,4,192 ; movsd (%r8,%rax,8),%xmm0 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,17 ; jne 1eeee <_sk_load_f16_sse2+0x19f> - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 233,142,254,255,255 ; jmpq 1ed7c <_sk_load_f16_sse2+0x2d> + DB 117,13 ; jne 1f25a <_sk_load_f16_sse2+0x168> + DB 243,15,126,192 ; movq %xmm0,%xmm0 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 233,197,254,255,255 ; jmpq 1f11f <_sk_load_f16_sse2+0x2d> DB 102,65,15,22,68,192,8 ; movhpd 0x8(%r8,%rax,8),%xmm0 - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,121,254,255,255 ; jb 1ed7c <_sk_load_f16_sse2+0x2d> - DB 242,65,15,16,76,192,16 ; movsd 0x10(%r8,%rax,8),%xmm1 - DB 233,109,254,255,255 ; jmpq 1ed7c <_sk_load_f16_sse2+0x2d> + DB 15,130,176,254,255,255 ; jb 1f11f <_sk_load_f16_sse2+0x2d> + DB 243,65,15,126,76,192,16 ; movq 0x10(%r8,%rax,8),%xmm1 + DB 233,164,254,255,255 ; jmpq 1f11f <_sk_load_f16_sse2+0x2d> PUBLIC _sk_load_f16_dst_sse2 _sk_load_f16_dst_sse2 LABEL PROC @@ -29402,89 +29553,78 @@ _sk_load_f16_dst_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,98,1,0,0 ; jne 1f091 <_sk_load_f16_dst_sse2+0x182> + DB 15,133,47,1,0,0 ; jne 1f3ca <_sk_load_f16_dst_sse2+0x14f> DB 102,65,15,16,36,192 ; movupd (%r8,%rax,8),%xmm4 - DB 102,65,15,16,108,192,16 ; movupd 0x10(%r8,%rax,8),%xmm5 + DB 243,65,15,111,108,192,16 ; movdqu 0x10(%r8,%rax,8),%xmm5 DB 102,68,15,40,196 ; movapd %xmm4,%xmm8 DB 102,68,15,97,197 ; punpcklwd %xmm5,%xmm8 DB 102,15,105,229 ; punpckhwd %xmm5,%xmm4 - DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 - DB 102,68,15,97,244 ; punpcklwd %xmm4,%xmm14 + DB 102,65,15,111,240 ; movdqa %xmm8,%xmm6 + DB 102,15,97,244 ; punpcklwd %xmm4,%xmm6 DB 102,68,15,105,196 ; punpckhwd %xmm4,%xmm8 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,65,15,111,238 ; movdqa %xmm14,%xmm5 + DB 102,15,111,238 ; movdqa %xmm6,%xmm5 DB 102,65,15,97,234 ; punpcklwd %xmm10,%xmm5 - DB 102,68,15,111,13,31,234,0,0 ; movdqa 0xea1f(%rip),%xmm9 # 2d990 <_sk_overlay_sse2_8bit+0x1438> - DB 102,15,111,229 ; movdqa %xmm5,%xmm4 - DB 102,65,15,219,225 ; pand %xmm9,%xmm4 - DB 102,15,239,236 ; pxor %xmm4,%xmm5 - DB 102,15,114,244,16 ; pslld $0x10,%xmm4 - DB 102,68,15,111,237 ; movdqa %xmm5,%xmm13 - DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 - DB 102,68,15,235,236 ; por %xmm4,%xmm13 - DB 102,68,15,111,29,20,234,0,0 ; movdqa 0xea14(%rip),%xmm11 # 2d9b0 <_sk_overlay_sse2_8bit+0x1458> - DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,22,234,0,0 ; movdqa 0xea16(%rip),%xmm12 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 102,65,15,239,236 ; pxor %xmm12,%xmm5 - DB 102,15,111,61,185,235,0,0 ; movdqa 0xebb9(%rip),%xmm7 # 2db70 <_sk_overlay_sse2_8bit+0x1618> + DB 102,68,15,111,13,149,229,0,0 ; movdqa 0xe595(%rip),%xmm9 # 2d870 <_sk_overlay_sse2_8bit+0x13c1> + DB 102,68,15,111,229 ; movdqa %xmm5,%xmm12 + DB 102,69,15,219,225 ; pand %xmm9,%xmm12 + DB 102,68,15,111,29,146,229,0,0 ; movdqa 0xe592(%rip),%xmm11 # 2d880 <_sk_overlay_sse2_8bit+0x13d1> + DB 102,65,15,219,235 ; pand %xmm11,%xmm5 + DB 102,15,111,61,149,229,0,0 ; movdqa 0xe595(%rip),%xmm7 # 2d890 <_sk_overlay_sse2_8bit+0x13e1> + DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 DB 102,15,111,231 ; movdqa %xmm7,%xmm4 DB 102,15,102,229 ; pcmpgtd %xmm5,%xmm4 - DB 102,65,15,223,229 ; pandn %xmm13,%xmm4 - DB 102,65,15,115,222,8 ; psrldq $0x8,%xmm14 - DB 102,69,15,97,242 ; punpcklwd %xmm10,%xmm14 - DB 102,65,15,111,238 ; movdqa %xmm14,%xmm5 - DB 102,65,15,219,233 ; pand %xmm9,%xmm5 - DB 102,68,15,239,245 ; pxor %xmm5,%xmm14 - DB 102,15,114,245,16 ; pslld $0x10,%xmm5 - DB 102,65,15,111,246 ; movdqa %xmm14,%xmm6 - DB 102,15,114,246,13 ; pslld $0xd,%xmm6 - DB 102,15,235,245 ; por %xmm5,%xmm6 - DB 102,65,15,254,243 ; paddd %xmm11,%xmm6 - DB 102,69,15,239,244 ; pxor %xmm12,%xmm14 + DB 102,15,114,245,13 ; pslld $0xd,%xmm5 + DB 102,65,15,235,236 ; por %xmm12,%xmm5 + DB 102,68,15,111,37,132,229,0,0 ; movdqa 0xe584(%rip),%xmm12 # 2d8a0 <_sk_overlay_sse2_8bit+0x13f1> + DB 102,65,15,254,236 ; paddd %xmm12,%xmm5 + DB 102,15,223,229 ; pandn %xmm5,%xmm4 + DB 102,65,15,105,242 ; punpckhwd %xmm10,%xmm6 + DB 102,68,15,111,238 ; movdqa %xmm6,%xmm13 + DB 102,69,15,219,233 ; pand %xmm9,%xmm13 + DB 102,65,15,219,243 ; pand %xmm11,%xmm6 + DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,65,15,102,238 ; pcmpgtd %xmm14,%xmm5 + DB 102,15,102,238 ; pcmpgtd %xmm6,%xmm5 + DB 102,15,114,246,13 ; pslld $0xd,%xmm6 + DB 102,65,15,235,245 ; por %xmm13,%xmm6 + DB 102,65,15,254,244 ; paddd %xmm12,%xmm6 DB 102,15,223,238 ; pandn %xmm6,%xmm5 DB 102,69,15,111,232 ; movdqa %xmm8,%xmm13 DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 - DB 102,65,15,111,245 ; movdqa %xmm13,%xmm6 - DB 102,65,15,219,241 ; pand %xmm9,%xmm6 - DB 102,68,15,239,238 ; pxor %xmm6,%xmm13 - DB 102,15,114,246,16 ; pslld $0x10,%xmm6 DB 102,69,15,111,245 ; movdqa %xmm13,%xmm14 - DB 102,65,15,114,246,13 ; pslld $0xd,%xmm14 - DB 102,68,15,235,246 ; por %xmm6,%xmm14 - DB 102,69,15,254,243 ; paddd %xmm11,%xmm14 - DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 + DB 102,69,15,219,241 ; pand %xmm9,%xmm14 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 102,65,15,114,246,16 ; pslld $0x10,%xmm14 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 DB 102,65,15,102,245 ; pcmpgtd %xmm13,%xmm6 - DB 102,65,15,223,246 ; pandn %xmm14,%xmm6 - DB 102,65,15,115,216,8 ; psrldq $0x8,%xmm8 - DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8 + DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 + DB 102,69,15,235,238 ; por %xmm14,%xmm13 + DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 + DB 102,65,15,223,245 ; pandn %xmm13,%xmm6 + DB 102,69,15,105,194 ; punpckhwd %xmm10,%xmm8 DB 102,69,15,219,200 ; pand %xmm8,%xmm9 - DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 + DB 102,69,15,219,195 ; pand %xmm11,%xmm8 DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 - DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 - DB 102,65,15,114,242,13 ; pslld $0xd,%xmm10 - DB 102,69,15,235,209 ; por %xmm9,%xmm10 - DB 102,69,15,254,211 ; paddd %xmm11,%xmm10 - DB 102,69,15,239,196 ; pxor %xmm12,%xmm8 DB 102,65,15,102,248 ; pcmpgtd %xmm8,%xmm7 - DB 102,65,15,223,250 ; pandn %xmm10,%xmm7 + DB 102,65,15,114,240,13 ; pslld $0xd,%xmm8 + DB 102,69,15,235,193 ; por %xmm9,%xmm8 + DB 102,69,15,254,196 ; paddd %xmm12,%xmm8 + DB 102,65,15,223,248 ; pandn %xmm8,%xmm7 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 242,65,15,16,36,192 ; movsd (%r8,%rax,8),%xmm4 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,17 ; jne 1f0ae <_sk_load_f16_dst_sse2+0x19f> - DB 102,15,87,237 ; xorpd %xmm5,%xmm5 - DB 102,15,20,229 ; unpcklpd %xmm5,%xmm4 - DB 102,15,87,237 ; xorpd %xmm5,%xmm5 - DB 233,142,254,255,255 ; jmpq 1ef3c <_sk_load_f16_dst_sse2+0x2d> + DB 117,13 ; jne 1f3e3 <_sk_load_f16_dst_sse2+0x168> + DB 243,15,126,228 ; movq %xmm4,%xmm4 + DB 102,15,239,237 ; pxor %xmm5,%xmm5 + DB 233,197,254,255,255 ; jmpq 1f2a8 <_sk_load_f16_dst_sse2+0x2d> DB 102,65,15,22,100,192,8 ; movhpd 0x8(%r8,%rax,8),%xmm4 - DB 102,15,87,237 ; xorpd %xmm5,%xmm5 + DB 102,15,239,237 ; pxor %xmm5,%xmm5 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,121,254,255,255 ; jb 1ef3c <_sk_load_f16_dst_sse2+0x2d> - DB 242,65,15,16,108,192,16 ; movsd 0x10(%r8,%rax,8),%xmm5 - DB 233,109,254,255,255 ; jmpq 1ef3c <_sk_load_f16_dst_sse2+0x2d> + DB 15,130,176,254,255,255 ; jb 1f2a8 <_sk_load_f16_dst_sse2+0x2d> + DB 243,65,15,126,108,192,16 ; movq 0x10(%r8,%rax,8),%xmm5 + DB 233,164,254,255,255 ; jmpq 1f2a8 <_sk_load_f16_dst_sse2+0x2d> PUBLIC _sk_gather_f16_sse2 _sk_gather_f16_sse2 LABEL PROC @@ -29517,73 +29657,125 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 DB 102,68,15,97,194 ; punpcklwd %xmm2,%xmm8 DB 102,15,105,202 ; punpckhwd %xmm2,%xmm1 - DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 - DB 102,68,15,97,241 ; punpcklwd %xmm1,%xmm14 + DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 + DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,68,15,105,193 ; punpckhwd %xmm1,%xmm8 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1 - DB 102,68,15,111,13,25,232,0,0 ; movdqa 0xe819(%rip),%xmm9 # 2d990 <_sk_overlay_sse2_8bit+0x1438> - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,65,15,219,193 ; pand %xmm9,%xmm0 - DB 102,15,239,200 ; pxor %xmm0,%xmm1 - DB 102,15,114,240,16 ; pslld $0x10,%xmm0 - DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 - DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 - DB 102,68,15,235,232 ; por %xmm0,%xmm13 - DB 102,68,15,111,29,14,232,0,0 ; movdqa 0xe80e(%rip),%xmm11 # 2d9b0 <_sk_overlay_sse2_8bit+0x1458> - DB 102,69,15,254,235 ; paddd %xmm11,%xmm13 - DB 102,68,15,111,37,16,232,0,0 ; movdqa 0xe810(%rip),%xmm12 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 102,65,15,239,204 ; pxor %xmm12,%xmm1 - DB 102,15,111,29,179,233,0,0 ; movdqa 0xe9b3(%rip),%xmm3 # 2db70 <_sk_overlay_sse2_8bit+0x1618> + DB 102,68,15,111,13,198,227,0,0 ; movdqa 0xe3c6(%rip),%xmm9 # 2d870 <_sk_overlay_sse2_8bit+0x13c1> + DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 + DB 102,69,15,219,225 ; pand %xmm9,%xmm12 + DB 102,68,15,111,29,195,227,0,0 ; movdqa 0xe3c3(%rip),%xmm11 # 2d880 <_sk_overlay_sse2_8bit+0x13d1> + DB 102,65,15,219,203 ; pand %xmm11,%xmm1 + DB 102,15,111,29,198,227,0,0 ; movdqa 0xe3c6(%rip),%xmm3 # 2d890 <_sk_overlay_sse2_8bit+0x13e1> + DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0 - DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 - DB 102,65,15,115,222,8 ; psrldq $0x8,%xmm14 - DB 102,69,15,97,242 ; punpcklwd %xmm10,%xmm14 - DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 - DB 102,65,15,219,201 ; pand %xmm9,%xmm1 - DB 102,68,15,239,241 ; pxor %xmm1,%xmm14 - DB 102,15,114,241,16 ; pslld $0x10,%xmm1 - DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 - DB 102,15,114,242,13 ; pslld $0xd,%xmm2 - DB 102,15,235,209 ; por %xmm1,%xmm2 - DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,69,15,239,244 ; pxor %xmm12,%xmm14 + DB 102,15,114,241,13 ; pslld $0xd,%xmm1 + DB 102,65,15,235,204 ; por %xmm12,%xmm1 + DB 102,68,15,111,37,181,227,0,0 ; movdqa 0xe3b5(%rip),%xmm12 # 2d8a0 <_sk_overlay_sse2_8bit+0x13f1> + DB 102,65,15,254,204 ; paddd %xmm12,%xmm1 + DB 102,15,223,193 ; pandn %xmm1,%xmm0 + DB 102,65,15,105,210 ; punpckhwd %xmm10,%xmm2 + DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 + DB 102,69,15,219,233 ; pand %xmm9,%xmm13 + DB 102,65,15,219,211 ; pand %xmm11,%xmm2 + DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,65,15,102,206 ; pcmpgtd %xmm14,%xmm1 + DB 102,15,102,202 ; pcmpgtd %xmm2,%xmm1 + DB 102,15,114,242,13 ; pslld $0xd,%xmm2 + DB 102,65,15,235,213 ; por %xmm13,%xmm2 + DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 DB 102,15,223,202 ; pandn %xmm2,%xmm1 DB 102,69,15,111,232 ; movdqa %xmm8,%xmm13 DB 102,69,15,97,234 ; punpcklwd %xmm10,%xmm13 - DB 102,65,15,111,213 ; movdqa %xmm13,%xmm2 - DB 102,65,15,219,209 ; pand %xmm9,%xmm2 - DB 102,68,15,239,234 ; pxor %xmm2,%xmm13 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 DB 102,69,15,111,245 ; movdqa %xmm13,%xmm14 - DB 102,65,15,114,246,13 ; pslld $0xd,%xmm14 - DB 102,68,15,235,242 ; por %xmm2,%xmm14 - DB 102,69,15,254,243 ; paddd %xmm11,%xmm14 - DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 + DB 102,69,15,219,241 ; pand %xmm9,%xmm14 + DB 102,69,15,219,235 ; pand %xmm11,%xmm13 + DB 102,65,15,114,246,16 ; pslld $0x10,%xmm14 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 DB 102,65,15,102,213 ; pcmpgtd %xmm13,%xmm2 - DB 102,65,15,223,214 ; pandn %xmm14,%xmm2 - DB 102,65,15,115,216,8 ; psrldq $0x8,%xmm8 - DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8 + DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13 + DB 102,69,15,235,238 ; por %xmm14,%xmm13 + DB 102,69,15,254,236 ; paddd %xmm12,%xmm13 + DB 102,65,15,223,213 ; pandn %xmm13,%xmm2 + DB 102,69,15,105,194 ; punpckhwd %xmm10,%xmm8 DB 102,69,15,219,200 ; pand %xmm8,%xmm9 - DB 102,69,15,239,193 ; pxor %xmm9,%xmm8 + DB 102,69,15,219,195 ; pand %xmm11,%xmm8 DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 - DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 - DB 102,65,15,114,242,13 ; pslld $0xd,%xmm10 - DB 102,69,15,235,209 ; por %xmm9,%xmm10 - DB 102,69,15,254,211 ; paddd %xmm11,%xmm10 - DB 102,69,15,239,196 ; pxor %xmm12,%xmm8 DB 102,65,15,102,216 ; pcmpgtd %xmm8,%xmm3 - DB 102,65,15,223,218 ; pandn %xmm10,%xmm3 + DB 102,65,15,114,240,13 ; pslld $0xd,%xmm8 + DB 102,69,15,235,193 ; por %xmm9,%xmm8 + DB 102,69,15,254,196 ; paddd %xmm12,%xmm8 + DB 102,65,15,223,216 ; pandn %xmm8,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_store_f16_sse2 _sk_store_f16_sse2 LABEL PROC + DB 72,131,236,24 ; sub $0x18,%rsp + DB 15,41,60,36 ; movaps %xmm7,(%rsp) + DB 102,68,15,111,29,6,227,0,0 ; movdqa 0xe306(%rip),%xmm11 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> + DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 + DB 102,69,15,219,195 ; pand %xmm11,%xmm8 + DB 102,68,15,111,21,3,227,0,0 ; movdqa 0xe303(%rip),%xmm10 # 2d8c0 <_sk_overlay_sse2_8bit+0x1411> + DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 + DB 102,69,15,219,242 ; pand %xmm10,%xmm14 + DB 102,15,111,61,1,227,0,0 ; movdqa 0xe301(%rip),%xmm7 # 2d8d0 <_sk_overlay_sse2_8bit+0x1421> + DB 102,68,15,111,37,168,228,0,0 ; movdqa 0xe4a8(%rip),%xmm12 # 2da80 <_sk_overlay_sse2_8bit+0x15d1> + DB 102,68,15,111,248 ; movdqa %xmm0,%xmm15 + DB 102,65,15,114,247,3 ; pslld $0x3,%xmm15 + DB 102,69,15,219,252 ; pand %xmm12,%xmm15 + DB 102,69,15,254,248 ; paddd %xmm8,%xmm15 + DB 102,68,15,111,45,90,223,0,0 ; movdqa 0xdf5a(%rip),%xmm13 # 2d550 <_sk_overlay_sse2_8bit+0x10a1> + DB 102,69,15,254,253 ; paddd %xmm13,%xmm15 + DB 102,65,15,114,231,16 ; psrad $0x10,%xmm15 + DB 102,68,15,111,199 ; movdqa %xmm7,%xmm8 + DB 102,69,15,102,198 ; pcmpgtd %xmm14,%xmm8 + DB 102,69,15,223,199 ; pandn %xmm15,%xmm8 + DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 + DB 102,69,15,219,243 ; pand %xmm11,%xmm14 + DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 + DB 102,65,15,114,247,3 ; pslld $0x3,%xmm15 + DB 102,69,15,219,252 ; pand %xmm12,%xmm15 + DB 102,69,15,254,254 ; paddd %xmm14,%xmm15 + DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 + DB 102,69,15,219,242 ; pand %xmm10,%xmm14 + DB 102,68,15,111,207 ; movdqa %xmm7,%xmm9 + DB 102,69,15,102,206 ; pcmpgtd %xmm14,%xmm9 + DB 102,68,15,107,192 ; packssdw %xmm0,%xmm8 + DB 102,69,15,254,253 ; paddd %xmm13,%xmm15 + DB 102,65,15,114,231,16 ; psrad $0x10,%xmm15 + DB 102,69,15,223,207 ; pandn %xmm15,%xmm9 + DB 102,68,15,107,200 ; packssdw %xmm0,%xmm9 + DB 102,69,15,97,193 ; punpcklwd %xmm9,%xmm8 + DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 + DB 102,69,15,219,203 ; pand %xmm11,%xmm9 + DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 + DB 102,65,15,114,247,3 ; pslld $0x3,%xmm15 + DB 102,69,15,219,252 ; pand %xmm12,%xmm15 + DB 102,69,15,254,249 ; paddd %xmm9,%xmm15 + DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 + DB 102,69,15,219,202 ; pand %xmm10,%xmm9 + DB 102,68,15,111,247 ; movdqa %xmm7,%xmm14 + DB 102,69,15,102,241 ; pcmpgtd %xmm9,%xmm14 + DB 102,69,15,254,253 ; paddd %xmm13,%xmm15 + DB 102,65,15,114,231,16 ; psrad $0x10,%xmm15 + DB 102,69,15,223,247 ; pandn %xmm15,%xmm14 + DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9 + DB 102,65,15,114,241,3 ; pslld $0x3,%xmm9 + DB 102,69,15,219,204 ; pand %xmm12,%xmm9 + DB 102,68,15,219,219 ; pand %xmm3,%xmm11 + DB 102,69,15,254,203 ; paddd %xmm11,%xmm9 + DB 102,69,15,254,205 ; paddd %xmm13,%xmm9 + DB 102,68,15,219,211 ; pand %xmm3,%xmm10 + DB 102,65,15,102,250 ; pcmpgtd %xmm10,%xmm7 + DB 102,65,15,114,225,16 ; psrad $0x10,%xmm9 + DB 102,65,15,223,249 ; pandn %xmm9,%xmm7 + DB 102,68,15,107,240 ; packssdw %xmm0,%xmm14 + DB 102,15,107,248 ; packssdw %xmm0,%xmm7 + DB 102,68,15,97,247 ; punpcklwd %xmm7,%xmm14 DB 72,173 ; lods %ds:(%rsi),%rax DB 76,99,72,8 ; movslq 0x8(%rax),%r9 DB 76,99,193 ; movslq %ecx,%r8 @@ -29591,83 +29783,28 @@ _sk_store_f16_sse2 LABEL PROC DB 73,193,224,3 ; shl $0x3,%r8 DB 76,3,0 ; add (%rax),%r8 DB 72,99,194 ; movslq %edx,%rax - DB 102,68,15,111,21,9,231,0,0 ; movdqa 0xe709(%rip),%xmm10 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 - DB 102,69,15,219,226 ; pand %xmm10,%xmm12 - DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 - DB 102,69,15,239,236 ; pxor %xmm12,%xmm13 - DB 102,68,15,111,13,252,230,0,0 ; movdqa 0xe6fc(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1478> - DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 - DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 - DB 102,69,15,102,197 ; pcmpgtd %xmm13,%xmm8 - DB 102,65,15,114,213,13 ; psrld $0xd,%xmm13 - DB 102,68,15,111,29,237,230,0,0 ; movdqa 0xe6ed(%rip),%xmm11 # 2d9e0 <_sk_overlay_sse2_8bit+0x1488> - DB 102,69,15,235,227 ; por %xmm11,%xmm12 - DB 102,69,15,254,229 ; paddd %xmm13,%xmm12 - DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,65,15,114,228,16 ; psrad $0x10,%xmm12 - DB 102,69,15,223,196 ; pandn %xmm12,%xmm8 - DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 - DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 - DB 102,69,15,219,226 ; pand %xmm10,%xmm12 - DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,69,15,239,244 ; pxor %xmm12,%xmm14 - DB 102,65,15,114,212,16 ; psrld $0x10,%xmm12 - DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 - DB 102,69,15,102,238 ; pcmpgtd %xmm14,%xmm13 - DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,235,227 ; por %xmm11,%xmm12 - DB 102,69,15,254,230 ; paddd %xmm14,%xmm12 - DB 102,65,15,114,244,16 ; pslld $0x10,%xmm12 - DB 102,65,15,114,228,16 ; psrad $0x10,%xmm12 - DB 102,69,15,223,236 ; pandn %xmm12,%xmm13 - DB 102,69,15,107,237 ; packssdw %xmm13,%xmm13 - DB 102,68,15,111,242 ; movdqa %xmm2,%xmm14 - DB 102,69,15,219,242 ; pand %xmm10,%xmm14 - DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 - DB 102,69,15,239,254 ; pxor %xmm14,%xmm15 - DB 102,65,15,114,214,16 ; psrld $0x10,%xmm14 - DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 - DB 102,69,15,102,231 ; pcmpgtd %xmm15,%xmm12 - DB 102,65,15,114,215,13 ; psrld $0xd,%xmm15 - DB 102,69,15,235,243 ; por %xmm11,%xmm14 - DB 102,69,15,254,247 ; paddd %xmm15,%xmm14 - DB 102,65,15,114,246,16 ; pslld $0x10,%xmm14 - DB 102,65,15,114,230,16 ; psrad $0x10,%xmm14 - DB 102,69,15,223,230 ; pandn %xmm14,%xmm12 - DB 102,69,15,107,228 ; packssdw %xmm12,%xmm12 - DB 102,68,15,219,211 ; pand %xmm3,%xmm10 - DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,69,15,239,242 ; pxor %xmm10,%xmm14 - DB 102,65,15,114,210,16 ; psrld $0x10,%xmm10 - DB 102,69,15,102,206 ; pcmpgtd %xmm14,%xmm9 - DB 102,65,15,114,214,13 ; psrld $0xd,%xmm14 - DB 102,69,15,235,211 ; por %xmm11,%xmm10 - DB 102,69,15,254,214 ; paddd %xmm14,%xmm10 - DB 102,65,15,114,242,16 ; pslld $0x10,%xmm10 - DB 102,65,15,114,226,16 ; psrad $0x10,%xmm10 - DB 102,69,15,223,202 ; pandn %xmm10,%xmm9 - DB 102,69,15,107,201 ; packssdw %xmm9,%xmm9 - DB 102,69,15,97,197 ; punpcklwd %xmm13,%xmm8 - DB 102,69,15,97,225 ; punpcklwd %xmm9,%xmm12 - DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 - DB 102,69,15,98,204 ; punpckldq %xmm12,%xmm9 DB 72,133,255 ; test %rdi,%rdi - DB 117,21 ; jne 1f415 <_sk_store_f16_sse2+0x17e> - DB 69,15,17,12,192 ; movups %xmm9,(%r8,%rax,8) - DB 102,69,15,106,196 ; punpckhdq %xmm12,%xmm8 + DB 117,40 ; jne 1f72b <_sk_store_f16_sse2+0x192> + DB 102,65,15,111,248 ; movdqa %xmm8,%xmm7 + DB 102,65,15,98,254 ; punpckldq %xmm14,%xmm7 + DB 243,65,15,127,60,192 ; movdqu %xmm7,(%r8,%rax,8) + DB 102,69,15,106,198 ; punpckhdq %xmm14,%xmm8 DB 243,69,15,127,68,192,16 ; movdqu %xmm8,0x10(%r8,%rax,8) DB 72,173 ; lods %ds:(%rsi),%rax + DB 15,40,60,36 ; movaps (%rsp),%xmm7 + DB 72,131,196,24 ; add $0x18,%rsp DB 255,224 ; jmpq *%rax + DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 + DB 102,69,15,98,206 ; punpckldq %xmm14,%xmm9 DB 102,69,15,214,12,192 ; movq %xmm9,(%r8,%rax,8) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 1f411 <_sk_store_f16_sse2+0x17a> + DB 116,222 ; je 1f71f <_sk_store_f16_sse2+0x186> DB 102,69,15,23,76,192,8 ; movhpd %xmm9,0x8(%r8,%rax,8) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 1f411 <_sk_store_f16_sse2+0x17a> - DB 102,69,15,106,196 ; punpckhdq %xmm12,%xmm8 + DB 114,209 ; jb 1f71f <_sk_store_f16_sse2+0x186> + DB 102,69,15,106,198 ; punpckhdq %xmm14,%xmm8 DB 102,69,15,214,68,192,16 ; movq %xmm8,0x10(%r8,%rax,8) - DB 235,213 ; jmp 1f411 <_sk_store_f16_sse2+0x17a> + DB 235,195 ; jmp 1f71f <_sk_store_f16_sse2+0x186> PUBLIC _sk_load_u16_be_sse2 _sk_load_u16_be_sse2 LABEL PROC @@ -29680,9 +29817,9 @@ _sk_load_u16_be_sse2 LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,190,0,0,0 ; jne 1f521 <_sk_load_u16_be_sse2+0xe5> + DB 15,133,190,0,0,0 ; jne 1f841 <_sk_load_u16_be_sse2+0xe5> DB 102,65,15,16,4,64 ; movupd (%r8,%rax,2),%xmm0 - DB 102,65,15,16,76,64,16 ; movupd 0x10(%r8,%rax,2),%xmm1 + DB 243,65,15,111,76,64,16 ; movdqu 0x10(%r8,%rax,2),%xmm1 DB 102,15,40,208 ; movapd %xmm0,%xmm2 DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 @@ -29697,7 +29834,7 @@ _sk_load_u16_be_sse2 LABEL PROC DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 68,15,40,5,156,226,0,0 ; movaps 0xe29c(%rip),%xmm8 # 2d750 <_sk_overlay_sse2_8bit+0x11f8> + DB 68,15,40,5,108,222,0,0 ; movaps 0xde6c(%rip),%xmm8 # 2d640 <_sk_overlay_sse2_8bit+0x1191> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -29725,17 +29862,16 @@ _sk_load_u16_be_sse2 LABEL PROC DB 255,224 ; jmpq *%rax DB 242,65,15,16,4,64 ; movsd (%r8,%rax,2),%xmm0 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,17 ; jne 1f53e <_sk_load_u16_be_sse2+0x102> - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 - DB 233,50,255,255,255 ; jmpq 1f470 <_sk_load_u16_be_sse2+0x34> + DB 117,13 ; jne 1f85a <_sk_load_u16_be_sse2+0xfe> + DB 243,15,126,192 ; movq %xmm0,%xmm0 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 233,54,255,255,255 ; jmpq 1f790 <_sk_load_u16_be_sse2+0x34> DB 102,65,15,22,68,64,8 ; movhpd 0x8(%r8,%rax,2),%xmm0 - DB 102,15,87,201 ; xorpd %xmm1,%xmm1 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 15,130,29,255,255,255 ; jb 1f470 <_sk_load_u16_be_sse2+0x34> - DB 242,65,15,16,76,64,16 ; movsd 0x10(%r8,%rax,2),%xmm1 - DB 233,17,255,255,255 ; jmpq 1f470 <_sk_load_u16_be_sse2+0x34> + DB 15,130,33,255,255,255 ; jb 1f790 <_sk_load_u16_be_sse2+0x34> + DB 243,65,15,126,76,64,16 ; movq 0x10(%r8,%rax,2),%xmm1 + DB 233,21,255,255,255 ; jmpq 1f790 <_sk_load_u16_be_sse2+0x34> PUBLIC _sk_load_rgb_u16_be_sse2 _sk_load_rgb_u16_be_sse2 LABEL PROC @@ -29750,7 +29886,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 72,141,4,64 ; lea (%rax,%rax,2),%rax DB 72,193,248,32 ; sar $0x20,%rax DB 72,133,255 ; test %rdi,%rdi - DB 15,133,175,0,0,0 ; jne 1f639 <_sk_load_rgb_u16_be_sse2+0xda> + DB 15,133,175,0,0,0 ; jne 1f955 <_sk_load_rgb_u16_be_sse2+0xda> DB 243,65,15,111,20,64 ; movdqu (%r8,%rax,2),%xmm2 DB 243,65,15,111,92,64,8 ; movdqu 0x8(%r8,%rax,2),%xmm3 DB 102,15,115,219,4 ; psrldq $0x4,%xmm3 @@ -29758,8 +29894,8 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,115,217,6 ; psrldq $0x6,%xmm1 - DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 + DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 @@ -29771,7 +29907,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,65,15,97,200 ; punpcklwd %xmm8,%xmm1 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 68,15,40,13,98,225,0,0 ; movaps 0xe162(%rip),%xmm9 # 2d750 <_sk_overlay_sse2_8bit+0x11f8> + DB 68,15,40,13,54,221,0,0 ; movaps 0xdd36(%rip),%xmm9 # 2d640 <_sk_overlay_sse2_8bit+0x1191> DB 65,15,89,193 ; mulps %xmm9,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -29788,27 +29924,27 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 65,15,89,209 ; mulps %xmm9,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,137,222,0,0 ; movaps 0xde89(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,40,29,93,218,0,0 ; movaps 0xda5d(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 255,224 ; jmpq *%rax DB 102,65,15,110,20,64 ; movd (%r8,%rax,2),%xmm2 - DB 102,65,15,196,84,64,4,2 ; pinsrw $0x2,0x4(%r8,%rax,2),%xmm2 DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,65,15,196,84,64,4,2 ; pinsrw $0x2,0x4(%r8,%rax,2),%xmm2 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,13 ; jne 1f65e <_sk_load_rgb_u16_be_sse2+0xff> + DB 117,13 ; jne 1f97a <_sk_load_rgb_u16_be_sse2+0xff> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,80,255,255,255 ; jmpq 1f5ae <_sk_load_rgb_u16_be_sse2+0x4f> + DB 233,80,255,255,255 ; jmpq 1f8ca <_sk_load_rgb_u16_be_sse2+0x4f> DB 102,65,15,110,68,64,6 ; movd 0x6(%r8,%rax,2),%xmm0 - DB 102,65,15,196,68,64,10,2 ; pinsrw $0x2,0xa(%r8,%rax,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,65,15,196,68,64,10,2 ; pinsrw $0x2,0xa(%r8,%rax,2),%xmm0 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,24 ; jb 1f68f <_sk_load_rgb_u16_be_sse2+0x130> + DB 114,24 ; jb 1f9ab <_sk_load_rgb_u16_be_sse2+0x130> DB 102,65,15,110,92,64,12 ; movd 0xc(%r8,%rax,2),%xmm3 - DB 102,65,15,196,92,64,16,2 ; pinsrw $0x2,0x10(%r8,%rax,2),%xmm3 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 233,31,255,255,255 ; jmpq 1f5ae <_sk_load_rgb_u16_be_sse2+0x4f> + DB 102,65,15,196,92,64,16,2 ; pinsrw $0x2,0x10(%r8,%rax,2),%xmm3 + DB 233,31,255,255,255 ; jmpq 1f8ca <_sk_load_rgb_u16_be_sse2+0x4f> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 233,22,255,255,255 ; jmpq 1f5ae <_sk_load_rgb_u16_be_sse2+0x4f> + DB 233,22,255,255,255 ; jmpq 1f8ca <_sk_load_rgb_u16_be_sse2+0x4f> PUBLIC _sk_store_u16_be_sse2 _sk_store_u16_be_sse2 LABEL PROC @@ -29820,52 +29956,52 @@ _sk_store_u16_be_sse2 LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 73,99,193 ; movslq %r9d,%rax - DB 68,15,40,21,50,227,0,0 ; movaps 0xe332(%rip),%xmm10 # 2d9f0 <_sk_overlay_sse2_8bit+0x1498> + DB 68,15,40,21,22,223,0,0 ; movaps 0xdf16(%rip),%xmm10 # 2d8f0 <_sk_overlay_sse2_8bit+0x1441> DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 69,15,89,194 ; mulps %xmm10,%xmm8 DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8 DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8 DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8 - DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8 + DB 102,68,15,107,192 ; packssdw %xmm0,%xmm8 DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9 DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 DB 102,69,15,235,193 ; por %xmm9,%xmm8 DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 69,15,89,202 ; mulps %xmm10,%xmm9 - DB 102,69,15,91,217 ; cvtps2dq %xmm9,%xmm11 - DB 102,65,15,114,243,16 ; pslld $0x10,%xmm11 - DB 102,65,15,114,227,16 ; psrad $0x10,%xmm11 - DB 102,69,15,107,219 ; packssdw %xmm11,%xmm11 - DB 102,69,15,111,203 ; movdqa %xmm11,%xmm9 - DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9 - DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 - DB 102,69,15,235,217 ; por %xmm9,%xmm11 + DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 + DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 + DB 102,65,15,114,225,16 ; psrad $0x10,%xmm9 + DB 102,68,15,107,200 ; packssdw %xmm0,%xmm9 + DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 + DB 102,65,15,113,243,8 ; psllw $0x8,%xmm11 + DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 + DB 102,69,15,235,203 ; por %xmm11,%xmm9 + DB 102,69,15,97,193 ; punpcklwd %xmm9,%xmm8 DB 68,15,40,202 ; movaps %xmm2,%xmm9 DB 69,15,89,202 ; mulps %xmm10,%xmm9 DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9 DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9 DB 102,65,15,114,225,16 ; psrad $0x10,%xmm9 - DB 102,69,15,107,201 ; packssdw %xmm9,%xmm9 - DB 102,69,15,111,225 ; movdqa %xmm9,%xmm12 - DB 102,65,15,113,244,8 ; psllw $0x8,%xmm12 + DB 102,68,15,107,200 ; packssdw %xmm0,%xmm9 + DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 + DB 102,65,15,113,243,8 ; psllw $0x8,%xmm11 DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 - DB 102,69,15,235,204 ; por %xmm12,%xmm9 + DB 102,69,15,235,203 ; por %xmm11,%xmm9 DB 68,15,89,211 ; mulps %xmm3,%xmm10 DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10 DB 102,65,15,114,242,16 ; pslld $0x10,%xmm10 DB 102,65,15,114,226,16 ; psrad $0x10,%xmm10 - DB 102,69,15,107,210 ; packssdw %xmm10,%xmm10 - DB 102,69,15,111,226 ; movdqa %xmm10,%xmm12 - DB 102,65,15,113,244,8 ; psllw $0x8,%xmm12 + DB 102,68,15,107,208 ; packssdw %xmm0,%xmm10 + DB 102,69,15,111,218 ; movdqa %xmm10,%xmm11 + DB 102,65,15,113,243,8 ; psllw $0x8,%xmm11 DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 - DB 102,69,15,235,212 ; por %xmm12,%xmm10 - DB 102,69,15,97,195 ; punpcklwd %xmm11,%xmm8 + DB 102,69,15,235,211 ; por %xmm11,%xmm10 DB 102,69,15,97,202 ; punpcklwd %xmm10,%xmm9 DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10 DB 102,69,15,98,209 ; punpckldq %xmm9,%xmm10 DB 72,133,255 ; test %rdi,%rdi - DB 117,21 ; jne 1f7b8 <_sk_store_u16_be_sse2+0x120> + DB 117,21 ; jne 1fad4 <_sk_store_u16_be_sse2+0x120> DB 69,15,17,20,64 ; movups %xmm10,(%r8,%rax,2) DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 DB 243,69,15,127,68,64,16 ; movdqu %xmm8,0x10(%r8,%rax,2) @@ -29873,13 +30009,13 @@ _sk_store_u16_be_sse2 LABEL PROC DB 255,224 ; jmpq *%rax DB 102,69,15,214,20,64 ; movq %xmm10,(%r8,%rax,2) DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,240 ; je 1f7b4 <_sk_store_u16_be_sse2+0x11c> + DB 116,240 ; je 1fad0 <_sk_store_u16_be_sse2+0x11c> DB 102,69,15,23,84,64,8 ; movhpd %xmm10,0x8(%r8,%rax,2) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,227 ; jb 1f7b4 <_sk_store_u16_be_sse2+0x11c> + DB 114,227 ; jb 1fad0 <_sk_store_u16_be_sse2+0x11c> DB 102,69,15,106,193 ; punpckhdq %xmm9,%xmm8 DB 102,69,15,214,68,64,16 ; movq %xmm8,0x10(%r8,%rax,2) - DB 235,213 ; jmp 1f7b4 <_sk_store_u16_be_sse2+0x11c> + DB 235,213 ; jmp 1fad0 <_sk_store_u16_be_sse2+0x11c> PUBLIC _sk_load_f32_sse2 _sk_load_f32_sse2 LABEL PROC @@ -29893,7 +30029,7 @@ _sk_load_f32_sse2 LABEL PROC DB 73,99,193 ; movslq %r9d,%rax DB 69,15,16,4,128 ; movups (%r8,%rax,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 1f84a <_sk_load_f32_sse2+0x6b> + DB 117,66 ; jne 1fb66 <_sk_load_f32_sse2+0x6b> DB 65,15,16,68,128,16 ; movups 0x10(%r8,%rax,4),%xmm0 DB 65,15,16,92,128,32 ; movups 0x20(%r8,%rax,4),%xmm3 DB 69,15,16,76,128,48 ; movups 0x30(%r8,%rax,4),%xmm9 @@ -29913,17 +30049,17 @@ _sk_load_f32_sse2 LABEL PROC DB 255,224 ; jmpq *%rax DB 69,15,87,201 ; xorps %xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,8 ; jne 1f85c <_sk_load_f32_sse2+0x7d> + DB 117,8 ; jne 1fb78 <_sk_load_f32_sse2+0x7d> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 - DB 235,190 ; jmp 1f81a <_sk_load_f32_sse2+0x3b> + DB 235,190 ; jmp 1fb36 <_sk_load_f32_sse2+0x3b> DB 65,15,16,68,128,16 ; movups 0x10(%r8,%rax,4),%xmm0 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,8 ; jb 1f870 <_sk_load_f32_sse2+0x91> + DB 114,8 ; jb 1fb8c <_sk_load_f32_sse2+0x91> DB 65,15,16,92,128,32 ; movups 0x20(%r8,%rax,4),%xmm3 - DB 235,170 ; jmp 1f81a <_sk_load_f32_sse2+0x3b> + DB 235,170 ; jmp 1fb36 <_sk_load_f32_sse2+0x3b> DB 15,87,219 ; xorps %xmm3,%xmm3 - DB 235,165 ; jmp 1f81a <_sk_load_f32_sse2+0x3b> + DB 235,165 ; jmp 1fb36 <_sk_load_f32_sse2+0x3b> PUBLIC _sk_load_f32_dst_sse2 _sk_load_f32_dst_sse2 LABEL PROC @@ -29937,7 +30073,7 @@ _sk_load_f32_dst_sse2 LABEL PROC DB 73,99,193 ; movslq %r9d,%rax DB 69,15,16,4,128 ; movups (%r8,%rax,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 1f8e0 <_sk_load_f32_dst_sse2+0x6b> + DB 117,66 ; jne 1fbfc <_sk_load_f32_dst_sse2+0x6b> DB 65,15,16,100,128,16 ; movups 0x10(%r8,%rax,4),%xmm4 DB 65,15,16,124,128,32 ; movups 0x20(%r8,%rax,4),%xmm7 DB 69,15,16,76,128,48 ; movups 0x30(%r8,%rax,4),%xmm9 @@ -29957,17 +30093,17 @@ _sk_load_f32_dst_sse2 LABEL PROC DB 255,224 ; jmpq *%rax DB 69,15,87,201 ; xorps %xmm9,%xmm9 DB 72,131,255,1 ; cmp $0x1,%rdi - DB 117,8 ; jne 1f8f2 <_sk_load_f32_dst_sse2+0x7d> + DB 117,8 ; jne 1fc0e <_sk_load_f32_dst_sse2+0x7d> DB 15,87,255 ; xorps %xmm7,%xmm7 DB 15,87,228 ; xorps %xmm4,%xmm4 - DB 235,190 ; jmp 1f8b0 <_sk_load_f32_dst_sse2+0x3b> + DB 235,190 ; jmp 1fbcc <_sk_load_f32_dst_sse2+0x3b> DB 65,15,16,100,128,16 ; movups 0x10(%r8,%rax,4),%xmm4 DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,8 ; jb 1f906 <_sk_load_f32_dst_sse2+0x91> + DB 114,8 ; jb 1fc22 <_sk_load_f32_dst_sse2+0x91> DB 65,15,16,124,128,32 ; movups 0x20(%r8,%rax,4),%xmm7 - DB 235,170 ; jmp 1f8b0 <_sk_load_f32_dst_sse2+0x3b> + DB 235,170 ; jmp 1fbcc <_sk_load_f32_dst_sse2+0x3b> DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 235,165 ; jmp 1f8b0 <_sk_load_f32_dst_sse2+0x3b> + DB 235,165 ; jmp 1fbcc <_sk_load_f32_dst_sse2+0x3b> PUBLIC _sk_store_f32_sse2 _sk_store_f32_sse2 LABEL PROC @@ -29994,7 +30130,7 @@ _sk_store_f32_sse2 LABEL PROC DB 102,69,15,20,203 ; unpcklpd %xmm11,%xmm9 DB 102,69,15,17,36,128 ; movupd %xmm12,(%r8,%rax,4) DB 72,133,255 ; test %rdi,%rdi - DB 117,29 ; jne 1f988 <_sk_store_f32_sse2+0x7d> + DB 117,29 ; jne 1fca4 <_sk_store_f32_sse2+0x7d> DB 102,69,15,21,211 ; unpckhpd %xmm11,%xmm10 DB 69,15,17,68,128,16 ; movups %xmm8,0x10(%r8,%rax,4) DB 102,69,15,17,76,128,32 ; movupd %xmm9,0x20(%r8,%rax,4) @@ -30002,12 +30138,12 @@ _sk_store_f32_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 72,131,255,1 ; cmp $0x1,%rdi - DB 116,246 ; je 1f984 <_sk_store_f32_sse2+0x79> + DB 116,246 ; je 1fca0 <_sk_store_f32_sse2+0x79> DB 69,15,17,68,128,16 ; movups %xmm8,0x10(%r8,%rax,4) DB 72,131,255,3 ; cmp $0x3,%rdi - DB 114,234 ; jb 1f984 <_sk_store_f32_sse2+0x79> + DB 114,234 ; jb 1fca0 <_sk_store_f32_sse2+0x79> DB 102,69,15,17,76,128,32 ; movupd %xmm9,0x20(%r8,%rax,4) - DB 235,225 ; jmp 1f984 <_sk_store_f32_sse2+0x79> + DB 235,225 ; jmp 1fca0 <_sk_store_f32_sse2+0x79> PUBLIC _sk_clamp_x_sse2 _sk_clamp_x_sse2 LABEL PROC @@ -30047,7 +30183,7 @@ _sk_repeat_x_sse2 LABEL PROC DB 243,69,15,91,194 ; cvttps2dq %xmm10,%xmm8 DB 69,15,91,216 ; cvtdq2ps %xmm8,%xmm11 DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10 - DB 68,15,84,21,159,218,0,0 ; andps 0xda9f(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,84,21,115,214,0,0 ; andps 0xd673(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 @@ -30071,7 +30207,7 @@ _sk_repeat_y_sse2 LABEL PROC DB 243,69,15,91,194 ; cvttps2dq %xmm10,%xmm8 DB 69,15,91,216 ; cvtdq2ps %xmm8,%xmm11 DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10 - DB 68,15,84,21,69,218,0,0 ; andps 0xda45(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,84,21,25,214,0,0 ; andps 0xd619(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 69,15,92,218 ; subps %xmm10,%xmm11 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 @@ -30095,13 +30231,13 @@ _sk_mirror_x_sse2 LABEL PROC DB 65,15,92,193 ; subps %xmm9,%xmm0 DB 243,69,15,88,210 ; addss %xmm10,%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 243,68,15,89,29,210,215,0,0 ; mulss 0xd7d2(%rip),%xmm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 243,68,15,89,29,150,211,0,0 ; mulss 0xd396(%rip),%xmm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 68,15,89,216 ; mulps %xmm0,%xmm11 DB 243,69,15,91,195 ; cvttps2dq %xmm11,%xmm8 DB 69,15,91,224 ; cvtdq2ps %xmm8,%xmm12 DB 69,15,194,220,1 ; cmpltps %xmm12,%xmm11 - DB 68,15,84,29,203,217,0,0 ; andps 0xd9cb(%rip),%xmm11 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,84,29,159,213,0,0 ; andps 0xd59f(%rip),%xmm11 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 69,15,92,227 ; subps %xmm11,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 @@ -30128,13 +30264,13 @@ _sk_mirror_y_sse2 LABEL PROC DB 65,15,92,201 ; subps %xmm9,%xmm1 DB 243,69,15,88,210 ; addss %xmm10,%xmm10 DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10 - DB 243,68,15,89,29,77,215,0,0 ; mulss 0xd74d(%rip),%xmm11 # 2d2a8 <_sk_overlay_sse2_8bit+0xd50> + DB 243,68,15,89,29,17,211,0,0 ; mulss 0xd311(%rip),%xmm11 # 2d188 <_sk_overlay_sse2_8bit+0xcd9> DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 68,15,89,217 ; mulps %xmm1,%xmm11 DB 243,69,15,91,195 ; cvttps2dq %xmm11,%xmm8 DB 69,15,91,224 ; cvtdq2ps %xmm8,%xmm12 DB 69,15,194,220,1 ; cmpltps %xmm12,%xmm11 - DB 68,15,84,29,70,217,0,0 ; andps 0xd946(%rip),%xmm11 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,84,29,26,213,0,0 ; andps 0xd51a(%rip),%xmm11 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 69,15,92,227 ; subps %xmm11,%xmm12 DB 69,15,89,226 ; mulps %xmm10,%xmm12 @@ -30155,7 +30291,7 @@ PUBLIC _sk_clamp_x_1_sse2 _sk_clamp_x_1_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 68,15,95,192 ; maxps %xmm0,%xmm8 - DB 68,15,93,5,253,216,0,0 ; minps 0xd8fd(%rip),%xmm8 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,93,5,209,212,0,0 ; minps 0xd4d1(%rip),%xmm8 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,15,40,192 ; movaps %xmm8,%xmm0 DB 255,224 ; jmpq *%rax @@ -30166,7 +30302,7 @@ _sk_repeat_x_1_sse2 LABEL PROC DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8 DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,194,200,1 ; cmpltps %xmm8,%xmm9 - DB 68,15,84,13,219,216,0,0 ; andps 0xd8db(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,84,13,175,212,0,0 ; andps 0xd4af(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,193 ; subps %xmm9,%xmm8 DB 65,15,92,192 ; subps %xmm8,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax @@ -30174,14 +30310,14 @@ _sk_repeat_x_1_sse2 LABEL PROC PUBLIC _sk_mirror_x_1_sse2 _sk_mirror_x_1_sse2 LABEL PROC - DB 68,15,40,5,39,217,0,0 ; movaps 0xd927(%rip),%xmm8 # 2d520 <_sk_overlay_sse2_8bit+0xfc8> + DB 68,15,40,5,251,212,0,0 ; movaps 0xd4fb(%rip),%xmm8 # 2d410 <_sk_overlay_sse2_8bit+0xf61> DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,13,171,216,0,0 ; movaps 0xd8ab(%rip),%xmm9 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,13,127,212,0,0 ; movaps 0xd47f(%rip),%xmm9 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 68,15,89,200 ; mulps %xmm0,%xmm9 DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10 DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,84,13,161,216,0,0 ; andps 0xd8a1(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,84,13,117,212,0,0 ; andps 0xd475(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,87,219 ; xorps %xmm11,%xmm11 DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,88,210 ; addps %xmm10,%xmm10 @@ -30195,10 +30331,10 @@ _sk_mirror_x_1_sse2 LABEL PROC PUBLIC _sk_luminance_to_alpha_sse2 _sk_luminance_to_alpha_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 15,89,5,183,221,0,0 ; mulps 0xddb7(%rip),%xmm0 # 2da00 <_sk_overlay_sse2_8bit+0x14a8> - DB 15,89,13,192,221,0,0 ; mulps 0xddc0(%rip),%xmm1 # 2da10 <_sk_overlay_sse2_8bit+0x14b8> + DB 15,89,5,155,217,0,0 ; mulps 0xd99b(%rip),%xmm0 # 2d900 <_sk_overlay_sse2_8bit+0x1451> + DB 15,89,13,164,217,0,0 ; mulps 0xd9a4(%rip),%xmm1 # 2d910 <_sk_overlay_sse2_8bit+0x1461> DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,29,198,221,0,0 ; mulps 0xddc6(%rip),%xmm3 # 2da20 <_sk_overlay_sse2_8bit+0x14c8> + DB 15,89,29,170,217,0,0 ; mulps 0xd9aa(%rip),%xmm3 # 2d920 <_sk_overlay_sse2_8bit+0x1471> DB 15,88,217 ; addps %xmm1,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -30494,9 +30630,9 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 76,139,112,8 ; mov 0x8(%rax),%r14 DB 72,255,203 ; dec %rbx - DB 120,7 ; js 200fa <_sk_evenly_spaced_gradient_sse2+0x18> + DB 120,7 ; js 20416 <_sk_evenly_spaced_gradient_sse2+0x18> DB 243,72,15,42,203 ; cvtsi2ss %rbx,%xmm1 - DB 235,21 ; jmp 2010f <_sk_evenly_spaced_gradient_sse2+0x2d> + DB 235,21 ; jmp 2042b <_sk_evenly_spaced_gradient_sse2+0x2d> DB 73,137,216 ; mov %rbx,%r8 DB 73,209,232 ; shr %r8 DB 131,227,1 ; and $0x1,%ebx @@ -30507,75 +30643,75 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 15,89,200 ; mulps %xmm0,%xmm1 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,73,15,126,210 ; movq %xmm2,%r10 - DB 69,137,208 ; mov %r10d,%r8d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 102,73,15,126,209 ; movq %xmm2,%r9 + DB 69,137,200 ; mov %r9d,%r8d + DB 73,193,233,32 ; shr $0x20,%r9 DB 102,73,15,126,203 ; movq %xmm1,%r11 - DB 69,137,217 ; mov %r11d,%r9d + DB 69,137,218 ; mov %r11d,%r10d DB 73,193,235,32 ; shr $0x20,%r11 DB 243,67,15,16,12,158 ; movss (%r14,%r11,4),%xmm1 - DB 243,67,15,16,20,150 ; movss (%r14,%r10,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,71,15,16,4,142 ; movss (%r14,%r9,4),%xmm8 - DB 243,67,15,16,20,134 ; movss (%r14,%r8,4),%xmm2 - DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 + DB 243,71,15,16,4,150 ; movss (%r14,%r10,4),%xmm8 DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 + DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 + DB 243,67,15,16,20,134 ; movss (%r14,%r8,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,68,15,20,194 ; unpcklpd %xmm2,%xmm8 DB 72,139,88,40 ; mov 0x28(%rax),%rbx DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 - DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 + DB 243,70,15,16,12,147 ; movss (%rbx,%r10,4),%xmm9 DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,68,15,20,202 ; unpcklpd %xmm2,%xmm9 DB 72,139,88,16 ; mov 0x10(%rax),%rbx DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 - DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 DB 72,139,88,48 ; mov 0x30(%rax),%rbx DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 243,70,15,16,20,139 ; movss (%rbx,%r9,4),%xmm10 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 - DB 68,15,20,211 ; unpcklps %xmm3,%xmm10 + DB 243,70,15,16,20,147 ; movss (%rbx,%r10,4),%xmm10 DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 243,70,15,16,28,155 ; movss (%rbx,%r11,4),%xmm11 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 - DB 68,15,20,218 ; unpcklps %xmm2,%xmm11 DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 102,68,15,20,211 ; unpcklpd %xmm3,%xmm10 + DB 72,139,88,24 ; mov 0x18(%rax),%rbx + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 65,15,20,211 ; unpcklps %xmm11,%xmm2 - DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 - DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 DB 243,70,15,16,28,139 ; movss (%rbx,%r9,4),%xmm11 DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 65,15,20,219 ; unpcklps %xmm11,%xmm3 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 72,139,88,56 ; mov 0x38(%rax),%rbx + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 243,70,15,16,28,147 ; movss (%rbx,%r10,4),%xmm11 DB 68,15,20,219 ; unpcklps %xmm3,%xmm11 - DB 69,15,20,220 ; unpcklps %xmm12,%xmm11 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 65,15,20,220 ; unpcklps %xmm12,%xmm3 + DB 102,68,15,20,219 ; unpcklpd %xmm3,%xmm11 DB 72,139,88,32 ; mov 0x20(%rax),%rbx DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 - DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 - DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 - DB 65,15,20,221 ; unpcklps %xmm13,%xmm3 DB 65,15,20,220 ; unpcklps %xmm12,%xmm3 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 + DB 69,15,20,236 ; unpcklps %xmm12,%xmm13 + DB 102,65,15,20,221 ; unpcklpd %xmm13,%xmm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 243,70,15,16,36,152 ; movss (%rax,%r11,4),%xmm12 DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 - DB 69,15,20,229 ; unpcklps %xmm13,%xmm12 - DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 - DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 - DB 69,15,20,238 ; unpcklps %xmm14,%xmm13 DB 69,15,20,236 ; unpcklps %xmm12,%xmm13 + DB 243,70,15,16,36,136 ; movss (%rax,%r9,4),%xmm12 + DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 + DB 69,15,20,244 ; unpcklps %xmm12,%xmm14 + DB 102,69,15,20,238 ; unpcklpd %xmm14,%xmm13 DB 68,15,89,192 ; mulps %xmm0,%xmm8 DB 69,15,88,193 ; addps %xmm9,%xmm8 DB 15,89,200 ; mulps %xmm0,%xmm1 @@ -30592,15 +30728,15 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC PUBLIC _sk_gauss_a_to_rgba_sse2 _sk_gauss_a_to_rgba_sse2 LABEL PROC - DB 15,40,5,149,215,0,0 ; movaps 0xd795(%rip),%xmm0 # 2da30 <_sk_overlay_sse2_8bit+0x14d8> + DB 15,40,5,114,211,0,0 ; movaps 0xd372(%rip),%xmm0 # 2d930 <_sk_overlay_sse2_8bit+0x1481> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,155,215,0,0 ; addps 0xd79b(%rip),%xmm0 # 2da40 <_sk_overlay_sse2_8bit+0x14e8> + DB 15,88,5,120,211,0,0 ; addps 0xd378(%rip),%xmm0 # 2d940 <_sk_overlay_sse2_8bit+0x1491> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,161,215,0,0 ; addps 0xd7a1(%rip),%xmm0 # 2da50 <_sk_overlay_sse2_8bit+0x14f8> + DB 15,88,5,126,211,0,0 ; addps 0xd37e(%rip),%xmm0 # 2d950 <_sk_overlay_sse2_8bit+0x14a1> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,167,215,0,0 ; addps 0xd7a7(%rip),%xmm0 # 2da60 <_sk_overlay_sse2_8bit+0x1508> + DB 15,88,5,132,211,0,0 ; addps 0xd384(%rip),%xmm0 # 2d960 <_sk_overlay_sse2_8bit+0x14b1> DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,5,173,215,0,0 ; addps 0xd7ad(%rip),%xmm0 # 2da70 <_sk_overlay_sse2_8bit+0x1518> + DB 15,88,5,138,211,0,0 ; addps 0xd38a(%rip),%xmm0 # 2d970 <_sk_overlay_sse2_8bit+0x14c1> DB 72,173 ; lods %ds:(%rsi),%rax DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -30609,97 +30745,95 @@ _sk_gauss_a_to_rgba_sse2 LABEL PROC PUBLIC _sk_gradient_sse2 _sk_gradient_sse2 LABEL PROC + DB 65,86 ; push %r14 + DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,51 ; jb 20312 <_sk_gradient_sse2+0x42> + DB 114,41 ; jb 2062e <_sk_gradient_sse2+0x3b> DB 76,139,72,72 ; mov 0x48(%rax),%r9 DB 73,255,200 ; dec %r8 DB 73,131,193,4 ; add $0x4,%r9 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,21,219,209,0,0 ; movaps 0xd1db(%rip),%xmm2 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 243,65,15,16,25 ; movss (%r9),%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 - DB 15,84,218 ; andps %xmm2,%xmm3 - DB 102,15,254,203 ; paddd %xmm3,%xmm1 + DB 243,65,15,16,17 ; movss (%r9),%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,194,208,2 ; cmpleps %xmm0,%xmm2 + DB 102,15,250,202 ; psubd %xmm2,%xmm1 DB 73,131,193,4 ; add $0x4,%r9 DB 73,255,200 ; dec %r8 - DB 117,227 ; jne 202f5 <_sk_gradient_sse2+0x25> - DB 65,86 ; push %r14 - DB 83 ; push %rbx + DB 117,230 ; jne 20614 <_sk_gradient_sse2+0x21> DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,73,15,126,210 ; movq %xmm2,%r10 - DB 69,137,208 ; mov %r10d,%r8d - DB 73,193,234,32 ; shr $0x20,%r10 + DB 102,73,15,126,209 ; movq %xmm2,%r9 + DB 69,137,200 ; mov %r9d,%r8d + DB 73,193,233,32 ; shr $0x20,%r9 DB 102,73,15,126,203 ; movq %xmm1,%r11 - DB 69,137,217 ; mov %r11d,%r9d + DB 69,137,218 ; mov %r11d,%r10d DB 73,193,235,32 ; shr $0x20,%r11 DB 72,139,88,8 ; mov 0x8(%rax),%rbx DB 76,139,112,16 ; mov 0x10(%rax),%r14 DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,70,15,16,4,139 ; movss (%rbx,%r9,4),%xmm8 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 - DB 68,15,20,194 ; unpcklps %xmm2,%xmm8 + DB 243,70,15,16,4,147 ; movss (%rbx,%r10,4),%xmm8 DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,68,15,20,194 ; unpcklpd %xmm2,%xmm8 DB 72,139,88,40 ; mov 0x28(%rax),%rbx DB 243,66,15,16,12,155 ; movss (%rbx,%r11,4),%xmm1 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 - DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 - DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 + DB 243,70,15,16,12,147 ; movss (%rbx,%r10,4),%xmm9 DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,68,15,20,202 ; unpcklpd %xmm2,%xmm9 DB 243,67,15,16,20,158 ; movss (%r14,%r11,4),%xmm2 DB 243,67,15,16,12,150 ; movss (%r14,%r10,4),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,67,15,16,12,142 ; movss (%r14,%r9,4),%xmm1 - DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 - DB 15,20,203 ; unpcklps %xmm3,%xmm1 DB 15,20,202 ; unpcklps %xmm2,%xmm1 + DB 243,67,15,16,20,142 ; movss (%r14,%r9,4),%xmm2 + DB 243,67,15,16,28,134 ; movss (%r14,%r8,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 DB 72,139,88,48 ; mov 0x30(%rax),%rbx DB 243,66,15,16,20,155 ; movss (%rbx,%r11,4),%xmm2 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 243,70,15,16,20,139 ; movss (%rbx,%r9,4),%xmm10 - DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 - DB 68,15,20,211 ; unpcklps %xmm3,%xmm10 + DB 243,70,15,16,20,147 ; movss (%rbx,%r10,4),%xmm10 DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 72,139,88,24 ; mov 0x18(%rax),%rbx - DB 243,70,15,16,28,155 ; movss (%rbx,%r11,4),%xmm11 - DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 - DB 68,15,20,218 ; unpcklps %xmm2,%xmm11 DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 102,68,15,20,211 ; unpcklpd %xmm3,%xmm10 + DB 72,139,88,24 ; mov 0x18(%rax),%rbx + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 65,15,20,211 ; unpcklps %xmm11,%xmm2 - DB 72,139,88,56 ; mov 0x38(%rax),%rbx - DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 - DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 - DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 DB 243,70,15,16,28,139 ; movss (%rbx,%r9,4),%xmm11 DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 65,15,20,219 ; unpcklps %xmm11,%xmm3 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 72,139,88,56 ; mov 0x38(%rax),%rbx + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 243,70,15,16,28,147 ; movss (%rbx,%r10,4),%xmm11 DB 68,15,20,219 ; unpcklps %xmm3,%xmm11 - DB 69,15,20,220 ; unpcklps %xmm12,%xmm11 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 65,15,20,220 ; unpcklps %xmm12,%xmm3 + DB 102,68,15,20,219 ; unpcklpd %xmm3,%xmm11 DB 72,139,88,32 ; mov 0x20(%rax),%rbx DB 243,70,15,16,36,155 ; movss (%rbx,%r11,4),%xmm12 DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 - DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 - DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 - DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 - DB 65,15,20,221 ; unpcklps %xmm13,%xmm3 DB 65,15,20,220 ; unpcklps %xmm12,%xmm3 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 + DB 69,15,20,236 ; unpcklps %xmm12,%xmm13 + DB 102,65,15,20,221 ; unpcklpd %xmm13,%xmm3 DB 72,139,64,64 ; mov 0x40(%rax),%rax DB 243,70,15,16,36,152 ; movss (%rax,%r11,4),%xmm12 DB 243,70,15,16,44,144 ; movss (%rax,%r10,4),%xmm13 - DB 69,15,20,229 ; unpcklps %xmm13,%xmm12 - DB 243,70,15,16,44,136 ; movss (%rax,%r9,4),%xmm13 - DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 - DB 69,15,20,238 ; unpcklps %xmm14,%xmm13 DB 69,15,20,236 ; unpcklps %xmm12,%xmm13 + DB 243,70,15,16,36,136 ; movss (%rax,%r9,4),%xmm12 + DB 243,70,15,16,52,128 ; movss (%rax,%r8,4),%xmm14 + DB 69,15,20,244 ; unpcklps %xmm12,%xmm14 + DB 102,69,15,20,238 ; unpcklpd %xmm14,%xmm13 DB 68,15,89,192 ; mulps %xmm0,%xmm8 DB 69,15,88,193 ; addps %xmm9,%xmm8 DB 15,89,200 ; mulps %xmm0,%xmm1 @@ -30762,29 +30896,29 @@ _sk_xy_to_unit_angle_sse2 LABEL PROC DB 69,15,94,220 ; divps %xmm12,%xmm11 DB 69,15,40,227 ; movaps %xmm11,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 - DB 68,15,40,45,58,213,0,0 ; movaps 0xd53a(%rip),%xmm13 # 2da80 <_sk_overlay_sse2_8bit+0x1528> + DB 68,15,40,45,26,209,0,0 ; movaps 0xd11a(%rip),%xmm13 # 2d980 <_sk_overlay_sse2_8bit+0x14d1> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,62,213,0,0 ; addps 0xd53e(%rip),%xmm13 # 2da90 <_sk_overlay_sse2_8bit+0x1538> + DB 68,15,88,45,30,209,0,0 ; addps 0xd11e(%rip),%xmm13 # 2d990 <_sk_overlay_sse2_8bit+0x14e1> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,66,213,0,0 ; addps 0xd542(%rip),%xmm13 # 2daa0 <_sk_overlay_sse2_8bit+0x1548> + DB 68,15,88,45,34,209,0,0 ; addps 0xd122(%rip),%xmm13 # 2d9a0 <_sk_overlay_sse2_8bit+0x14f1> DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,45,70,213,0,0 ; addps 0xd546(%rip),%xmm13 # 2dab0 <_sk_overlay_sse2_8bit+0x1558> + DB 68,15,88,45,38,209,0,0 ; addps 0xd126(%rip),%xmm13 # 2d9b0 <_sk_overlay_sse2_8bit+0x1501> DB 69,15,89,235 ; mulps %xmm11,%xmm13 DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9 - DB 68,15,40,21,69,213,0,0 ; movaps 0xd545(%rip),%xmm10 # 2dac0 <_sk_overlay_sse2_8bit+0x1568> + DB 68,15,40,21,37,209,0,0 ; movaps 0xd125(%rip),%xmm10 # 2d9c0 <_sk_overlay_sse2_8bit+0x1511> DB 69,15,92,213 ; subps %xmm13,%xmm10 DB 69,15,84,209 ; andps %xmm9,%xmm10 DB 69,15,85,205 ; andnps %xmm13,%xmm9 DB 69,15,86,202 ; orps %xmm10,%xmm9 DB 68,15,194,192,1 ; cmpltps %xmm0,%xmm8 - DB 68,15,40,21,24,207,0,0 ; movaps 0xcf18(%rip),%xmm10 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,21,232,202,0,0 ; movaps 0xcae8(%rip),%xmm10 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,92,209 ; subps %xmm9,%xmm10 DB 69,15,84,208 ; andps %xmm8,%xmm10 DB 69,15,85,193 ; andnps %xmm9,%xmm8 DB 69,15,86,194 ; orps %xmm10,%xmm8 DB 68,15,40,201 ; movaps %xmm1,%xmm9 DB 68,15,194,200,1 ; cmpltps %xmm0,%xmm9 - DB 68,15,40,21,7,207,0,0 ; movaps 0xcf07(%rip),%xmm10 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,21,215,202,0,0 ; movaps 0xcad7(%rip),%xmm10 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,208 ; subps %xmm8,%xmm10 DB 69,15,84,209 ; andps %xmm9,%xmm10 DB 69,15,85,200 ; andnps %xmm8,%xmm9 @@ -30815,7 +30949,7 @@ _sk_xy_to_2pt_conical_quadratic_max_sse2 LABEL PROC DB 243,69,15,89,203 ; mulss %xmm11,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,89,13,174,212,0,0 ; mulps 0xd4ae(%rip),%xmm9 # 2dad0 <_sk_overlay_sse2_8bit+0x1578> + DB 68,15,89,13,142,208,0,0 ; mulps 0xd08e(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1521> DB 15,89,192 ; mulps %xmm0,%xmm0 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 @@ -30823,7 +30957,7 @@ _sk_xy_to_2pt_conical_quadratic_max_sse2 LABEL PROC DB 243,69,15,89,219 ; mulss %xmm11,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 68,15,89,21,153,212,0,0 ; mulps 0xd499(%rip),%xmm10 # 2dae0 <_sk_overlay_sse2_8bit+0x1588> + DB 68,15,89,21,121,208,0,0 ; mulps 0xd079(%rip),%xmm10 # 2d9e0 <_sk_overlay_sse2_8bit+0x1531> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 @@ -30832,8 +30966,8 @@ _sk_xy_to_2pt_conical_quadratic_max_sse2 LABEL PROC DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 68,15,87,13,81,211,0,0 ; xorps 0xd351(%rip),%xmm9 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 68,15,89,5,57,206,0,0 ; mulps 0xce39(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,87,13,33,207,0,0 ; xorps 0xcf21(%rip),%xmm9 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> + DB 68,15,89,5,9,202,0,0 ; mulps 0xca09(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 @@ -30852,7 +30986,7 @@ _sk_xy_to_2pt_conical_quadratic_min_sse2 LABEL PROC DB 243,69,15,89,203 ; mulss %xmm11,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,89,13,16,212,0,0 ; mulps 0xd410(%rip),%xmm9 # 2dad0 <_sk_overlay_sse2_8bit+0x1578> + DB 68,15,89,13,240,207,0,0 ; mulps 0xcff0(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1521> DB 15,89,192 ; mulps %xmm0,%xmm0 DB 68,15,40,225 ; movaps %xmm1,%xmm12 DB 69,15,89,228 ; mulps %xmm12,%xmm12 @@ -30860,7 +30994,7 @@ _sk_xy_to_2pt_conical_quadratic_min_sse2 LABEL PROC DB 243,69,15,89,219 ; mulss %xmm11,%xmm11 DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11 DB 69,15,92,227 ; subps %xmm11,%xmm12 - DB 68,15,89,21,251,211,0,0 ; mulps 0xd3fb(%rip),%xmm10 # 2dae0 <_sk_overlay_sse2_8bit+0x1588> + DB 68,15,89,21,219,207,0,0 ; mulps 0xcfdb(%rip),%xmm10 # 2d9e0 <_sk_overlay_sse2_8bit+0x1531> DB 69,15,89,212 ; mulps %xmm12,%xmm10 DB 65,15,40,193 ; movaps %xmm9,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 @@ -30869,8 +31003,8 @@ _sk_xy_to_2pt_conical_quadratic_min_sse2 LABEL PROC DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 65,15,40,194 ; movaps %xmm10,%xmm0 DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 68,15,87,13,179,210,0,0 ; xorps 0xd2b3(%rip),%xmm9 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> - DB 68,15,89,5,155,205,0,0 ; mulps 0xcd9b(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,87,13,131,206,0,0 ; xorps 0xce83(%rip),%xmm9 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> + DB 68,15,89,5,107,201,0,0 ; mulps 0xc96b(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 65,15,89,192 ; mulps %xmm8,%xmm0 DB 69,15,92,202 ; subps %xmm10,%xmm9 DB 69,15,89,200 ; mulps %xmm8,%xmm9 @@ -30886,7 +31020,7 @@ _sk_xy_to_2pt_conical_linear_sse2 LABEL PROC DB 243,69,15,89,200 ; mulss %xmm8,%xmm9 DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9 DB 68,15,88,200 ; addps %xmm0,%xmm9 - DB 68,15,89,13,131,211,0,0 ; mulps 0xd383(%rip),%xmm9 # 2dad0 <_sk_overlay_sse2_8bit+0x1578> + DB 68,15,89,13,99,207,0,0 ; mulps 0xcf63(%rip),%xmm9 # 2d9d0 <_sk_overlay_sse2_8bit+0x1521> DB 15,89,192 ; mulps %xmm0,%xmm0 DB 68,15,40,209 ; movaps %xmm1,%xmm10 DB 69,15,89,210 ; mulps %xmm10,%xmm10 @@ -30894,7 +31028,7 @@ _sk_xy_to_2pt_conical_linear_sse2 LABEL PROC DB 243,69,15,89,192 ; mulss %xmm8,%xmm8 DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 DB 65,15,92,192 ; subps %xmm8,%xmm0 - DB 15,87,5,79,210,0,0 ; xorps 0xd24f(%rip),%xmm0 # 2d9c0 <_sk_overlay_sse2_8bit+0x1468> + DB 15,87,5,31,206,0,0 ; xorps 0xce1f(%rip),%xmm0 # 2d8b0 <_sk_overlay_sse2_8bit+0x1401> DB 65,15,94,193 ; divps %xmm9,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -30931,7 +31065,7 @@ _sk_apply_vector_mask_sse2 LABEL PROC PUBLIC _sk_save_xy_sse2 _sk_save_xy_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,213,204,0,0 ; movaps 0xccd5(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,5,165,200,0,0 ; movaps 0xc8a5(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 15,17,0 ; movups %xmm0,(%rax) DB 68,15,40,200 ; movaps %xmm0,%xmm9 DB 69,15,88,200 ; addps %xmm8,%xmm9 @@ -30939,7 +31073,7 @@ _sk_save_xy_sse2 LABEL PROC DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10 DB 69,15,40,217 ; movaps %xmm9,%xmm11 DB 69,15,194,218,1 ; cmpltps %xmm10,%xmm11 - DB 68,15,40,37,192,204,0,0 ; movaps 0xccc0(%rip),%xmm12 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 68,15,40,37,144,200,0,0 ; movaps 0xc890(%rip),%xmm12 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,84,220 ; andps %xmm12,%xmm11 DB 69,15,92,211 ; subps %xmm11,%xmm10 DB 69,15,92,202 ; subps %xmm10,%xmm9 @@ -30982,8 +31116,8 @@ _sk_bilinear_nx_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,89,210,0,0 ; addps 0xd259(%rip),%xmm0 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,33,204,0,0 ; movaps 0xcc21(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,5,57,206,0,0 ; addps 0xce39(%rip),%xmm0 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,241,199,0,0 ; movaps 0xc7f1(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -30994,7 +31128,7 @@ _sk_bilinear_px_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,240,203,0,0 ; addps 0xcbf0(%rip),%xmm0 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,88,5,192,199,0,0 ; addps 0xc7c0(%rip),%xmm0 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -31004,8 +31138,8 @@ _sk_bilinear_ny_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,18,210,0,0 ; addps 0xd212(%rip),%xmm1 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,218,203,0,0 ; movaps 0xcbda(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,13,242,205,0,0 ; addps 0xcdf2(%rip),%xmm1 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,170,199,0,0 ; movaps 0xc7aa(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -31016,7 +31150,7 @@ _sk_bilinear_py_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,168,203,0,0 ; addps 0xcba8(%rip),%xmm1 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 15,88,13,120,199,0,0 ; addps 0xc778(%rip),%xmm1 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -31026,13 +31160,13 @@ _sk_bicubic_n3x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,219,209,0,0 ; addps 0xd1db(%rip),%xmm0 # 2db00 <_sk_overlay_sse2_8bit+0x15a8> - DB 68,15,40,13,147,203,0,0 ; movaps 0xcb93(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,5,187,205,0,0 ; addps 0xcdbb(%rip),%xmm0 # 2da00 <_sk_overlay_sse2_8bit+0x1551> + DB 68,15,40,13,99,199,0,0 ; movaps 0xc763(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,207,209,0,0 ; mulps 0xd1cf(%rip),%xmm9 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,13,103,205,0,0 ; addps 0xcd67(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,13,175,205,0,0 ; mulps 0xcdaf(%rip),%xmm9 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,13,55,201,0,0 ; addps 0xc937(%rip),%xmm9 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,128,0,0,0 ; movups %xmm9,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -31043,16 +31177,16 @@ _sk_bicubic_n1x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,134,209,0,0 ; addps 0xd186(%rip),%xmm0 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,78,203,0,0 ; movaps 0xcb4e(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,5,102,205,0,0 ; addps 0xcd66(%rip),%xmm0 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,30,199,0,0 ; movaps 0xc71e(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,162,209,0,0 ; movaps 0xd1a2(%rip),%xmm8 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,5,130,205,0,0 ; movaps 0xcd82(%rip),%xmm8 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,166,209,0,0 ; addps 0xd1a6(%rip),%xmm8 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,5,134,205,0,0 ; addps 0xcd86(%rip),%xmm8 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,26,203,0,0 ; addps 0xcb1a(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,88,5,234,198,0,0 ; addps 0xc6ea(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,158,209,0,0 ; addps 0xd19e(%rip),%xmm8 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,5,126,205,0,0 ; addps 0xcd7e(%rip),%xmm8 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -31060,17 +31194,17 @@ _sk_bicubic_n1x_sse2 LABEL PROC PUBLIC _sk_bicubic_p1x_sse2 _sk_bicubic_p1x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,248,202,0,0 ; movaps 0xcaf8(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,5,200,198,0,0 ; movaps 0xc6c8(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,72,64 ; movups 0x40(%rax),%xmm9 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 68,15,40,21,84,209,0,0 ; movaps 0xd154(%rip),%xmm10 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,21,52,205,0,0 ; movaps 0xcd34(%rip),%xmm10 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,88,209,0,0 ; addps 0xd158(%rip),%xmm10 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,21,56,205,0,0 ; addps 0xcd38(%rip),%xmm10 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,84,209,0,0 ; addps 0xd154(%rip),%xmm10 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,21,52,205,0,0 ; addps 0xcd34(%rip),%xmm10 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,144,128,0,0,0 ; movups %xmm10,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -31080,11 +31214,11 @@ _sk_bicubic_p3x_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,0 ; movups (%rax),%xmm0 DB 68,15,16,64,64 ; movups 0x40(%rax),%xmm8 - DB 15,88,5,39,209,0,0 ; addps 0xd127(%rip),%xmm0 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 15,88,5,7,205,0,0 ; addps 0xcd07(%rip),%xmm0 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,247,208,0,0 ; mulps 0xd0f7(%rip),%xmm8 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,5,143,204,0,0 ; addps 0xcc8f(%rip),%xmm8 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,5,215,204,0,0 ; mulps 0xccd7(%rip),%xmm8 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,5,95,200,0,0 ; addps 0xc85f(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,128,0,0,0 ; movups %xmm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -31095,13 +31229,13 @@ _sk_bicubic_n3y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,189,208,0,0 ; addps 0xd0bd(%rip),%xmm1 # 2db00 <_sk_overlay_sse2_8bit+0x15a8> - DB 68,15,40,13,117,202,0,0 ; movaps 0xca75(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,13,157,204,0,0 ; addps 0xcc9d(%rip),%xmm1 # 2da00 <_sk_overlay_sse2_8bit+0x1551> + DB 68,15,40,13,69,198,0,0 ; movaps 0xc645(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 DB 69,15,40,193 ; movaps %xmm9,%xmm8 DB 69,15,89,192 ; mulps %xmm8,%xmm8 - DB 68,15,89,13,177,208,0,0 ; mulps 0xd0b1(%rip),%xmm9 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,13,73,204,0,0 ; addps 0xcc49(%rip),%xmm9 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,13,145,204,0,0 ; mulps 0xcc91(%rip),%xmm9 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,13,25,200,0,0 ; addps 0xc819(%rip),%xmm9 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,200 ; mulps %xmm8,%xmm9 DB 68,15,17,136,160,0,0,0 ; movups %xmm9,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -31112,16 +31246,16 @@ _sk_bicubic_n1y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,103,208,0,0 ; addps 0xd067(%rip),%xmm1 # 2daf0 <_sk_overlay_sse2_8bit+0x1598> - DB 68,15,40,13,47,202,0,0 ; movaps 0xca2f(%rip),%xmm9 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> + DB 15,88,13,71,204,0,0 ; addps 0xcc47(%rip),%xmm1 # 2d9f0 <_sk_overlay_sse2_8bit+0x1541> + DB 68,15,40,13,255,197,0,0 ; movaps 0xc5ff(%rip),%xmm9 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> DB 69,15,92,200 ; subps %xmm8,%xmm9 - DB 68,15,40,5,131,208,0,0 ; movaps 0xd083(%rip),%xmm8 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,5,99,204,0,0 ; movaps 0xcc63(%rip),%xmm8 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,135,208,0,0 ; addps 0xd087(%rip),%xmm8 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,5,103,204,0,0 ; addps 0xcc67(%rip),%xmm8 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,251,201,0,0 ; addps 0xc9fb(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,88,5,203,197,0,0 ; addps 0xc5cb(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 - DB 68,15,88,5,127,208,0,0 ; addps 0xd07f(%rip),%xmm8 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,5,95,204,0,0 ; addps 0xcc5f(%rip),%xmm8 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -31129,17 +31263,17 @@ _sk_bicubic_n1y_sse2 LABEL PROC PUBLIC _sk_bicubic_p1y_sse2 _sk_bicubic_p1y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 68,15,40,5,217,201,0,0 ; movaps 0xc9d9(%rip),%xmm8 # 2d4b0 <_sk_overlay_sse2_8bit+0xf58> + DB 68,15,40,5,169,197,0,0 ; movaps 0xc5a9(%rip),%xmm8 # 2d3a0 <_sk_overlay_sse2_8bit+0xef1> DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,72,96 ; movups 0x60(%rax),%xmm9 DB 65,15,88,200 ; addps %xmm8,%xmm1 - DB 68,15,40,21,52,208,0,0 ; movaps 0xd034(%rip),%xmm10 # 2db20 <_sk_overlay_sse2_8bit+0x15c8> + DB 68,15,40,21,20,204,0,0 ; movaps 0xcc14(%rip),%xmm10 # 2da20 <_sk_overlay_sse2_8bit+0x1571> DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,56,208,0,0 ; addps 0xd038(%rip),%xmm10 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 68,15,88,21,24,204,0,0 ; addps 0xcc18(%rip),%xmm10 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,89,209 ; mulps %xmm9,%xmm10 DB 69,15,88,208 ; addps %xmm8,%xmm10 DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,21,52,208,0,0 ; addps 0xd034(%rip),%xmm10 # 2db40 <_sk_overlay_sse2_8bit+0x15e8> + DB 68,15,88,21,20,204,0,0 ; addps 0xcc14(%rip),%xmm10 # 2da40 <_sk_overlay_sse2_8bit+0x1591> DB 68,15,17,144,160,0,0,0 ; movups %xmm10,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -31149,11 +31283,11 @@ _sk_bicubic_p3y_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 15,16,72,32 ; movups 0x20(%rax),%xmm1 DB 68,15,16,64,96 ; movups 0x60(%rax),%xmm8 - DB 15,88,13,6,208,0,0 ; addps 0xd006(%rip),%xmm1 # 2db30 <_sk_overlay_sse2_8bit+0x15d8> + DB 15,88,13,230,203,0,0 ; addps 0xcbe6(%rip),%xmm1 # 2da30 <_sk_overlay_sse2_8bit+0x1581> DB 69,15,40,200 ; movaps %xmm8,%xmm9 DB 69,15,89,201 ; mulps %xmm9,%xmm9 - DB 68,15,89,5,214,207,0,0 ; mulps 0xcfd6(%rip),%xmm8 # 2db10 <_sk_overlay_sse2_8bit+0x15b8> - DB 68,15,88,5,110,203,0,0 ; addps 0xcb6e(%rip),%xmm8 # 2d6b0 <_sk_overlay_sse2_8bit+0x1158> + DB 68,15,89,5,182,203,0,0 ; mulps 0xcbb6(%rip),%xmm8 # 2da10 <_sk_overlay_sse2_8bit+0x1561> + DB 68,15,88,5,62,199,0,0 ; addps 0xc73e(%rip),%xmm8 # 2d5a0 <_sk_overlay_sse2_8bit+0x10f1> DB 69,15,89,193 ; mulps %xmm9,%xmm8 DB 68,15,17,128,160,0,0,0 ; movups %xmm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -31225,12 +31359,13 @@ _sk_callback_sse2 LABEL PROC PUBLIC _sk_clut_3D_sse2 _sk_clut_3D_sse2 LABEL PROC - DB 72,129,236,184,0,0,0 ; sub $0xb8,%rsp - DB 15,41,188,36,160,0,0,0 ; movaps %xmm7,0xa0(%rsp) - DB 15,41,180,36,144,0,0,0 ; movaps %xmm6,0x90(%rsp) - DB 15,41,172,36,128,0,0,0 ; movaps %xmm5,0x80(%rsp) - DB 15,41,100,36,112 ; movaps %xmm4,0x70(%rsp) - DB 15,41,92,36,96 ; movaps %xmm3,0x60(%rsp) + DB 83 ; push %rbx + DB 72,129,236,192,0,0,0 ; sub $0xc0,%rsp + DB 15,41,188,36,176,0,0,0 ; movaps %xmm7,0xb0(%rsp) + DB 15,41,180,36,160,0,0,0 ; movaps %xmm6,0xa0(%rsp) + DB 15,41,172,36,144,0,0,0 ; movaps %xmm5,0x90(%rsp) + DB 15,41,164,36,128,0,0,0 ; movaps %xmm4,0x80(%rsp) + DB 15,41,92,36,112 ; movaps %xmm3,0x70(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -31238,25 +31373,27 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 DB 15,89,218 ; mulps %xmm2,%xmm3 - DB 15,40,235 ; movaps %xmm3,%xmm5 + DB 243,68,15,91,235 ; cvttps2dq %xmm3,%xmm13 + DB 68,15,40,219 ; movaps %xmm3,%xmm11 + DB 68,15,41,92,36,96 ; movaps %xmm11,0x60(%rsp) DB 102,65,15,110,208 ; movd %r8d,%xmm2 DB 102,68,15,112,250,0 ; pshufd $0x0,%xmm2,%xmm15 DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d DB 102,65,15,110,217 ; movd %r9d,%xmm3 DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 - DB 68,15,91,211 ; cvtdq2ps %xmm3,%xmm10 - DB 68,15,89,209 ; mulps %xmm1,%xmm10 - DB 243,65,15,91,226 ; cvttps2dq %xmm10,%xmm4 - DB 102,15,127,36,36 ; movdqa %xmm4,(%rsp) - DB 68,15,41,84,36,16 ; movaps %xmm10,0x10(%rsp) - DB 102,15,112,204,245 ; pshufd $0xf5,%xmm4,%xmm1 - DB 102,15,111,220 ; movdqa %xmm4,%xmm3 - DB 102,65,15,244,223 ; pmuludq %xmm15,%xmm3 - DB 102,68,15,112,203,232 ; pshufd $0xe8,%xmm3,%xmm9 + DB 15,91,251 ; cvtdq2ps %xmm3,%xmm7 + DB 15,89,249 ; mulps %xmm1,%xmm7 + DB 243,15,91,239 ; cvttps2dq %xmm7,%xmm5 + DB 102,15,127,44,36 ; movdqa %xmm5,(%rsp) + DB 15,41,124,36,16 ; movaps %xmm7,0x10(%rsp) + DB 102,15,112,205,245 ; pshufd $0xf5,%xmm5,%xmm1 DB 102,65,15,244,207 ; pmuludq %xmm15,%xmm1 + DB 102,65,15,111,223 ; movdqa %xmm15,%xmm3 + DB 102,15,244,221 ; pmuludq %xmm5,%xmm3 + DB 102,68,15,112,211,232 ; pshufd $0xe8,%xmm3,%xmm10 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,68,15,98,201 ; punpckldq %xmm1,%xmm9 + DB 102,68,15,98,209 ; punpckldq %xmm1,%xmm10 DB 102,65,15,110,200 ; movd %r8d,%xmm1 DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 DB 102,15,112,209,0 ; pshufd $0x0,%xmm1,%xmm2 @@ -31264,581 +31401,543 @@ _sk_clut_3D_sse2 LABEL PROC DB 65,255,200 ; dec %r8d DB 102,65,15,110,200 ; movd %r8d,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 - DB 68,15,91,233 ; cvtdq2ps %xmm1,%xmm13 - DB 68,15,89,232 ; mulps %xmm0,%xmm13 - DB 243,69,15,91,229 ; cvttps2dq %xmm13,%xmm12 - DB 102,65,15,112,196,245 ; pshufd $0xf5,%xmm12,%xmm0 - DB 102,65,15,111,204 ; movdqa %xmm12,%xmm1 - DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,68,15,112,241,232 ; pshufd $0xe8,%xmm1,%xmm14 - DB 102,68,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm8 - DB 102,65,15,244,192 ; pmuludq %xmm8,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,68,15,98,240 ; punpckldq %xmm0,%xmm14 - DB 243,15,91,197 ; cvttps2dq %xmm5,%xmm0 - DB 102,15,127,68,36,48 ; movdqa %xmm0,0x30(%rsp) - DB 15,40,229 ; movaps %xmm5,%xmm4 - DB 15,41,100,36,80 ; movaps %xmm4,0x50(%rsp) - DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 - DB 102,15,254,216 ; paddd %xmm0,%xmm3 - DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 - DB 102,15,254,195 ; paddd %xmm3,%xmm0 - DB 102,15,111,45,46,206,0,0 ; movdqa 0xce2e(%rip),%xmm5 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,244,197 ; pmuludq %xmm5,%xmm0 + DB 68,15,91,241 ; cvtdq2ps %xmm1,%xmm14 + DB 68,15,89,240 ; mulps %xmm0,%xmm14 + DB 243,69,15,91,206 ; cvttps2dq %xmm14,%xmm9 + DB 102,65,15,112,193,245 ; pshufd $0xf5,%xmm9,%xmm0 + DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,65,15,244,201 ; pmuludq %xmm9,%xmm1 + DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,244,205 ; pmuludq %xmm5,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,98,216 ; punpckldq %xmm0,%xmm3 + DB 102,15,127,92,36,48 ; movdqa %xmm3,0x30(%rsp) + DB 102,65,15,111,202 ; movdqa %xmm10,%xmm1 + DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 + DB 102,68,15,127,108,36,80 ; movdqa %xmm13,0x50(%rsp) + DB 102,15,254,217 ; paddd %xmm1,%xmm3 + DB 102,68,15,111,5,19,202,0,0 ; movdqa 0xca13(%rip),%xmm8 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,112,195,245 ; pshufd $0xf5,%xmm3,%xmm0 + DB 102,65,15,244,216 ; pmuludq %xmm8,%xmm3 + DB 102,65,15,244,192 ; pmuludq %xmm8,%xmm0 + DB 102,15,112,224,232 ; pshufd $0xe8,%xmm0,%xmm4 + DB 102,15,112,195,232 ; pshufd $0xe8,%xmm3,%xmm0 + DB 102,15,98,196 ; punpckldq %xmm4,%xmm0 + DB 102,65,15,126,216 ; movd %xmm3,%r8d + DB 102,15,112,216,229 ; pshufd $0xe5,%xmm0,%xmm3 + DB 102,65,15,126,217 ; movd %xmm3,%r9d + DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 + DB 102,65,15,126,218 ; movd %xmm3,%r10d + DB 102,15,112,216,231 ; pshufd $0xe7,%xmm0,%xmm3 + DB 102,65,15,126,219 ; movd %xmm3,%r11d + DB 72,139,24 ; mov (%rax),%rbx + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 243,66,15,16,36,147 ; movss (%rbx,%r10,4),%xmm4 + DB 15,20,227 ; unpcklps %xmm3,%xmm4 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,66,15,16,52,131 ; movss (%rbx,%r8,4),%xmm6 + DB 15,20,243 ; unpcklps %xmm3,%xmm6 + DB 102,15,20,244 ; unpcklpd %xmm4,%xmm6 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,250,220 ; psubd %xmm4,%xmm3 + DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,73,15,126,217 ; movq %xmm3,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 15,20,220 ; unpcklps %xmm4,%xmm3 + DB 69,137,193 ; mov %r8d,%r9d + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 + DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 102,15,20,221 ; unpcklpd %xmm5,%xmm3 + DB 102,68,15,111,37,218,194,0,0 ; movdqa 0xc2da(%rip),%xmm12 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,196 ; paddd %xmm12,%xmm0 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 243,66,15,16,44,147 ; movss (%rbx,%r10,4),%xmm5 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 69,137,193 ; mov %r8d,%r9d + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,20,236 ; unpcklpd %xmm4,%xmm5 + DB 15,40,37,255,200,0,0 ; movaps 0xc8ff(%rip),%xmm4 # 2da50 <_sk_overlay_sse2_8bit+0x15a1> + DB 68,15,88,220 ; addps %xmm4,%xmm11 + DB 68,15,41,92,36,64 ; movaps %xmm11,0x40(%rsp) + DB 68,15,40,223 ; movaps %xmm7,%xmm11 + DB 68,15,88,220 ; addps %xmm4,%xmm11 + DB 65,15,88,230 ; addps %xmm14,%xmm4 + DB 243,15,91,228 ; cvttps2dq %xmm4,%xmm4 + DB 102,15,112,252,245 ; pshufd $0xf5,%xmm4,%xmm7 + DB 102,15,244,250 ; pmuludq %xmm2,%xmm7 + DB 102,15,244,212 ; pmuludq %xmm4,%xmm2 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 + DB 102,15,112,215,232 ; pshufd $0xe8,%xmm7,%xmm2 + DB 102,15,98,194 ; punpckldq %xmm2,%xmm0 + DB 102,15,127,68,36,32 ; movdqa %xmm0,0x20(%rsp) + DB 102,15,254,200 ; paddd %xmm0,%xmm1 + DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 + DB 102,65,15,244,200 ; pmuludq %xmm8,%xmm1 + DB 102,65,15,244,208 ; pmuludq %xmm8,%xmm2 + DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 + DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 + DB 102,15,98,212 ; punpckldq %xmm4,%xmm2 DB 102,65,15,126,200 ; movd %xmm1,%r8d - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,112,202,229 ; pshufd $0xe5,%xmm2,%xmm1 DB 102,65,15,126,201 ; movd %xmm1,%r9d - DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,112,202,78 ; pshufd $0x4e,%xmm2,%xmm1 DB 102,65,15,126,202 ; movd %xmm1,%r10d - DB 76,139,24 ; mov (%rax),%r11 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,233 ; unpcklps %xmm1,%xmm5 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,12,139 ; movss (%r11,%r9,4),%xmm1 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,254,13,49,199,0,0 ; paddd 0xc731(%rip),%xmm1 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 102,15,112,202,231 ; pshufd $0xe7,%xmm2,%xmm1 + DB 102,15,126,200 ; movd %xmm1,%eax + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,36,147 ; movss (%rbx,%r10,4),%xmm4 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,66,15,16,60,131 ; movss (%rbx,%r8,4),%xmm7 DB 15,20,249 ; unpcklps %xmm1,%xmm7 - DB 102,15,254,5,5,199,0,0 ; paddd 0xc705(%rip),%xmm0 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,73,15,126,200 ; movq %xmm1,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,15,20,252 ; unpcklpd %xmm4,%xmm7 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,250,13,116,200,0,0 ; psubd 0xc874(%rip),%xmm1 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 15,40,45,26,205,0,0 ; movaps 0xcd1a(%rip),%xmm5 # 2db50 <_sk_overlay_sse2_8bit+0x15f8> - DB 15,88,229 ; addps %xmm5,%xmm4 - DB 15,41,100,36,64 ; movaps %xmm4,0x40(%rsp) - DB 68,15,88,213 ; addps %xmm5,%xmm10 - DB 65,15,88,237 ; addps %xmm13,%xmm5 - DB 243,15,91,237 ; cvttps2dq %xmm5,%xmm5 - DB 102,15,244,213 ; pmuludq %xmm5,%xmm2 - DB 102,15,112,237,245 ; pshufd $0xf5,%xmm5,%xmm5 - DB 102,65,15,244,232 ; pmuludq %xmm8,%xmm5 - DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 - DB 102,15,112,213,232 ; pshufd $0xe8,%xmm5,%xmm2 - DB 102,15,98,194 ; punpckldq %xmm2,%xmm0 - DB 102,15,127,68,36,32 ; movdqa %xmm0,0x20(%rsp) - DB 102,15,254,216 ; paddd %xmm0,%xmm3 - DB 102,15,112,227,245 ; pshufd $0xf5,%xmm3,%xmm4 - DB 102,15,111,5,227,204,0,0 ; movdqa 0xcce3(%rip),%xmm0 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 - DB 102,15,112,211,232 ; pshufd $0xe8,%xmm3,%xmm2 - DB 102,15,244,224 ; pmuludq %xmm0,%xmm4 - DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 - DB 102,15,112,220,232 ; pshufd $0xe8,%xmm4,%xmm3 - DB 102,15,98,211 ; punpckldq %xmm3,%xmm2 - DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 - DB 102,65,15,126,217 ; movd %xmm3,%r9d - DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 - DB 102,65,15,126,216 ; movd %xmm3,%r8d - DB 102,15,112,218,231 ; pshufd $0xe7,%xmm2,%xmm3 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,44,139 ; movss (%r11,%r9,4),%xmm5 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 102,15,111,218 ; movdqa %xmm2,%xmm3 - DB 102,15,254,29,234,197,0,0 ; paddd 0xc5ea(%rip),%xmm3 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,112,235,78 ; pshufd $0x4e,%xmm3,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,4,131 ; movss (%r11,%r8,4),%xmm0 - DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 - DB 102,15,254,21,190,197,0,0 ; paddd 0xc5be(%rip),%xmm2 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> + DB 102,15,20,224 ; unpcklpd %xmm0,%xmm4 + DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,213 ; unpcklps %xmm5,%xmm2 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 15,20,208 ; unpcklps %xmm0,%xmm2 - DB 65,15,91,196 ; cvtdq2ps %xmm12,%xmm0 - DB 68,15,92,232 ; subps %xmm0,%xmm13 - DB 15,92,230 ; subps %xmm6,%xmm4 - DB 65,15,89,229 ; mulps %xmm13,%xmm4 - DB 15,88,230 ; addps %xmm6,%xmm4 - DB 15,92,223 ; subps %xmm7,%xmm3 - DB 65,15,89,221 ; mulps %xmm13,%xmm3 - DB 15,88,223 ; addps %xmm7,%xmm3 - DB 15,92,209 ; subps %xmm1,%xmm2 - DB 65,15,89,213 ; mulps %xmm13,%xmm2 - DB 15,88,209 ; addps %xmm1,%xmm2 - DB 243,65,15,91,194 ; cvttps2dq %xmm10,%xmm0 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 102,15,20,209 ; unpcklpd %xmm1,%xmm2 + DB 65,15,91,193 ; cvtdq2ps %xmm9,%xmm0 + DB 68,15,92,240 ; subps %xmm0,%xmm14 + DB 15,92,254 ; subps %xmm6,%xmm7 + DB 65,15,89,254 ; mulps %xmm14,%xmm7 + DB 15,88,254 ; addps %xmm6,%xmm7 + DB 15,92,227 ; subps %xmm3,%xmm4 + DB 65,15,89,230 ; mulps %xmm14,%xmm4 + DB 15,88,227 ; addps %xmm3,%xmm4 + DB 15,92,213 ; subps %xmm5,%xmm2 + DB 65,15,89,214 ; mulps %xmm14,%xmm2 + DB 15,88,213 ; addps %xmm5,%xmm2 + DB 243,65,15,91,195 ; cvttps2dq %xmm11,%xmm0 DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,65,15,244,199 ; pmuludq %xmm15,%xmm0 DB 102,65,15,244,207 ; pmuludq %xmm15,%xmm1 - DB 102,68,15,112,224,232 ; pshufd $0xe8,%xmm0,%xmm12 + DB 102,68,15,244,248 ; pmuludq %xmm0,%xmm15 + DB 102,69,15,112,231,232 ; pshufd $0xe8,%xmm15,%xmm12 DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 DB 102,68,15,98,224 ; punpckldq %xmm0,%xmm12 DB 102,65,15,111,244 ; movdqa %xmm12,%xmm6 - DB 102,15,254,116,36,48 ; paddd 0x30(%rsp),%xmm6 - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,65,15,254,198 ; paddd %xmm14,%xmm0 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,65,15,111,232 ; movdqa %xmm8,%xmm5 - DB 102,15,244,197 ; pmuludq %xmm5,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,244,205 ; pmuludq %xmm5,%xmm1 - DB 102,68,15,111,221 ; movdqa %xmm5,%xmm11 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,65,15,254,245 ; paddd %xmm13,%xmm6 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 + DB 102,15,254,76,36,48 ; paddd 0x30(%rsp),%xmm1 + DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 + DB 102,65,15,244,200 ; pmuludq %xmm8,%xmm1 + DB 102,65,15,244,192 ; pmuludq %xmm8,%xmm0 + DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11 + DB 102,15,112,216,232 ; pshufd $0xe8,%xmm0,%xmm3 + DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 + DB 102,15,98,195 ; punpckldq %xmm3,%xmm0 + DB 102,65,15,126,200 ; movd %xmm1,%r8d DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,65,15,126,201 ; movd %xmm1,%r9d DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,65,15,126,202 ; movd %xmm1,%r10d DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,60,139 ; movss (%r11,%r9,4),%xmm7 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,69,15,16,4,131 ; movss (%r11,%rax,4),%xmm8 - DB 243,67,15,16,12,131 ; movss (%r11,%r8,4),%xmm1 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,70,15,16,4,131 ; movss (%rbx,%r8,4),%xmm8 DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 - DB 68,15,20,199 ; unpcklps %xmm7,%xmm8 + DB 102,68,15,20,195 ; unpcklpd %xmm3,%xmm8 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,68,15,111,21,134,196,0,0 ; movdqa 0xc486(%rip),%xmm10 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 - DB 102,15,112,249,78 ; pshufd $0x4e,%xmm1,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,69,15,118,201 ; pcmpeqd %xmm9,%xmm9 + DB 102,65,15,250,201 ; psubd %xmm9,%xmm1 + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,73,15,126,216 ; movq %xmm3,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 102,15,111,45,85,196,0,0 ; movdqa 0xc455(%rip),%xmm5 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,254,197 ; paddd %xmm5,%xmm0 - DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,233 ; unpcklps %xmm1,%xmm5 + DB 102,15,20,221 ; unpcklpd %xmm5,%xmm3 + DB 102,68,15,111,45,76,192,0,0 ; movdqa 0xc04c(%rip),%xmm13 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,197 ; paddd %xmm13,%xmm0 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,73,15,126,200 ; movq %xmm1,%r8 + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,4,147 ; movss (%r11,%r10,4),%xmm0 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 DB 15,20,232 ; unpcklps %xmm0,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,197 ; unpcklps %xmm5,%xmm0 + DB 102,15,20,205 ; unpcklpd %xmm5,%xmm1 DB 102,15,254,116,36,32 ; paddd 0x20(%rsp),%xmm6 - DB 102,15,112,238,245 ; pshufd $0xf5,%xmm6,%xmm5 - DB 102,65,15,111,251 ; movdqa %xmm11,%xmm7 - DB 102,15,244,247 ; pmuludq %xmm7,%xmm6 - DB 102,15,112,246,232 ; pshufd $0xe8,%xmm6,%xmm6 - DB 102,15,244,239 ; pmuludq %xmm7,%xmm5 - DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 - DB 102,15,98,245 ; punpckldq %xmm5,%xmm6 - DB 102,15,112,238,229 ; pshufd $0xe5,%xmm6,%xmm5 + DB 102,15,112,198,245 ; pshufd $0xf5,%xmm6,%xmm0 + DB 102,65,15,244,243 ; pmuludq %xmm11,%xmm6 + DB 102,65,15,244,195 ; pmuludq %xmm11,%xmm0 + DB 102,15,112,232,232 ; pshufd $0xe8,%xmm0,%xmm5 + DB 102,15,112,198,232 ; pshufd $0xe8,%xmm6,%xmm0 + DB 102,15,98,197 ; punpckldq %xmm5,%xmm0 + DB 102,65,15,126,240 ; movd %xmm6,%r8d + DB 102,15,112,232,229 ; pshufd $0xe5,%xmm0,%xmm5 DB 102,65,15,126,233 ; movd %xmm5,%r9d - DB 102,15,112,238,78 ; pshufd $0x4e,%xmm6,%xmm5 - DB 102,65,15,126,232 ; movd %xmm5,%r8d - DB 102,15,112,238,231 ; pshufd $0xe7,%xmm6,%xmm5 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 + DB 102,65,15,126,234 ; movd %xmm5,%r10d + DB 102,15,112,232,231 ; pshufd $0xe7,%xmm0,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,60,139 ; movss (%r11,%r9,4),%xmm7 - DB 15,20,253 ; unpcklps %xmm5,%xmm7 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 243,69,15,16,60,131 ; movss (%r11,%rax,4),%xmm15 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,52,147 ; movss (%rbx,%r10,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 + DB 243,70,15,16,60,131 ; movss (%rbx,%r8,4),%xmm15 DB 68,15,20,253 ; unpcklps %xmm5,%xmm15 - DB 68,15,20,255 ; unpcklps %xmm7,%xmm15 - DB 102,15,111,238 ; movdqa %xmm6,%xmm5 - DB 102,65,15,254,234 ; paddd %xmm10,%xmm5 - DB 102,15,112,253,78 ; pshufd $0x4e,%xmm5,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,233 ; movq %xmm5,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,68,15,20,254 ; unpcklpd %xmm6,%xmm15 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,65,15,250,233 ; psubd %xmm9,%xmm5 + DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,60,147 ; movss (%r11,%r10,4),%xmm7 - DB 15,20,239 ; unpcklps %xmm7,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,28,131 ; movss (%r11,%rax,4),%xmm11 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 68,15,20,223 ; unpcklps %xmm7,%xmm11 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,70,15,16,28,139 ; movss (%rbx,%r9,4),%xmm11 DB 68,15,20,221 ; unpcklps %xmm5,%xmm11 - DB 102,15,111,45,76,195,0,0 ; movdqa 0xc34c(%rip),%xmm5 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,254,245 ; paddd %xmm5,%xmm6 - DB 102,15,111,253 ; movdqa %xmm5,%xmm7 - DB 102,15,112,238,78 ; pshufd $0x4e,%xmm6,%xmm5 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 102,68,15,20,222 ; unpcklpd %xmm6,%xmm11 + DB 102,65,15,254,197 ; paddd %xmm13,%xmm0 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,241 ; movq %xmm6,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,52,147 ; movss (%r11,%r10,4),%xmm6 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,20,131 ; movss (%r11,%rax,4),%xmm10 - DB 243,67,15,16,52,131 ; movss (%r11,%r8,4),%xmm6 - DB 68,15,20,214 ; unpcklps %xmm6,%xmm10 - DB 68,15,20,213 ; unpcklps %xmm5,%xmm10 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 + DB 68,15,20,200 ; unpcklps %xmm0,%xmm9 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 102,68,15,20,205 ; unpcklpd %xmm5,%xmm9 DB 69,15,92,248 ; subps %xmm8,%xmm15 - DB 69,15,89,253 ; mulps %xmm13,%xmm15 + DB 69,15,89,254 ; mulps %xmm14,%xmm15 DB 69,15,88,248 ; addps %xmm8,%xmm15 - DB 68,15,92,217 ; subps %xmm1,%xmm11 - DB 69,15,89,221 ; mulps %xmm13,%xmm11 - DB 68,15,88,217 ; addps %xmm1,%xmm11 - DB 68,15,92,208 ; subps %xmm0,%xmm10 - DB 69,15,89,213 ; mulps %xmm13,%xmm10 - DB 68,15,88,208 ; addps %xmm0,%xmm10 + DB 68,15,92,219 ; subps %xmm3,%xmm11 + DB 69,15,89,222 ; mulps %xmm14,%xmm11 + DB 68,15,88,219 ; addps %xmm3,%xmm11 + DB 68,15,92,201 ; subps %xmm1,%xmm9 + DB 69,15,89,206 ; mulps %xmm14,%xmm9 + DB 68,15,88,201 ; addps %xmm1,%xmm9 DB 15,91,4,36 ; cvtdq2ps (%rsp),%xmm0 DB 15,40,76,36,16 ; movaps 0x10(%rsp),%xmm1 DB 15,92,200 ; subps %xmm0,%xmm1 DB 15,41,76,36,16 ; movaps %xmm1,0x10(%rsp) - DB 68,15,92,252 ; subps %xmm4,%xmm15 + DB 68,15,92,255 ; subps %xmm7,%xmm15 DB 68,15,89,249 ; mulps %xmm1,%xmm15 - DB 68,15,88,252 ; addps %xmm4,%xmm15 - DB 68,15,92,219 ; subps %xmm3,%xmm11 + DB 68,15,88,255 ; addps %xmm7,%xmm15 + DB 68,15,92,220 ; subps %xmm4,%xmm11 DB 68,15,89,217 ; mulps %xmm1,%xmm11 - DB 68,15,88,219 ; addps %xmm3,%xmm11 - DB 68,15,92,210 ; subps %xmm2,%xmm10 - DB 68,15,89,209 ; mulps %xmm1,%xmm10 - DB 68,15,88,210 ; addps %xmm2,%xmm10 + DB 68,15,88,220 ; addps %xmm4,%xmm11 + DB 68,15,92,202 ; subps %xmm2,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 68,15,88,202 ; addps %xmm2,%xmm9 DB 243,15,91,68,36,64 ; cvttps2dq 0x40(%rsp),%xmm0 DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) - DB 102,68,15,254,200 ; paddd %xmm0,%xmm9 - DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 - DB 102,65,15,254,201 ; paddd %xmm9,%xmm1 + DB 102,68,15,254,208 ; paddd %xmm0,%xmm10 + DB 102,15,111,116,36,48 ; movdqa 0x30(%rsp),%xmm6 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 + DB 102,65,15,254,202 ; paddd %xmm10,%xmm1 DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 - DB 102,15,111,5,238,200,0,0 ; movdqa 0xc8ee(%rip),%xmm0 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 + DB 102,15,111,29,17,197,0,0 ; movdqa 0xc511(%rip),%xmm3 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 + DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 - DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,15,112,202,232 ; pshufd $0xe8,%xmm2,%xmm1 - DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 + DB 102,15,98,218 ; punpckldq %xmm2,%xmm3 + DB 102,65,15,126,200 ; movd %xmm1,%r8d DB 102,15,112,203,229 ; pshufd $0xe5,%xmm3,%xmm1 DB 102,65,15,126,201 ; movd %xmm1,%r9d DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 - DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,65,15,126,202 ; movd %xmm1,%r10d DB 102,15,112,203,231 ; pshufd $0xe7,%xmm3,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,20,139 ; movss (%r11,%r9,4),%xmm2 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 + DB 102,15,20,202 ; unpcklpd %xmm2,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,111,5,246,193,0,0 ; movdqa 0xc1f6(%rip),%xmm0 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,208 ; paddd %xmm0,%xmm2 + DB 102,69,15,118,237 ; pcmpeqd %xmm13,%xmm13 + DB 102,65,15,250,213 ; psubd %xmm13,%xmm2 DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,226 ; unpcklps %xmm2,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,213 ; unpcklps %xmm5,%xmm2 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 15,20,212 ; unpcklps %xmm4,%xmm2 - DB 102,15,254,223 ; paddd %xmm7,%xmm3 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 102,15,20,213 ; unpcklpd %xmm5,%xmm2 + DB 102,15,111,5,225,189,0,0 ; movdqa 0xbde1(%rip),%xmm0 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,254,216 ; paddd %xmm0,%xmm3 DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,216 ; movq %xmm3,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 15,20,220 ; unpcklps %xmm4,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 102,15,111,124,36,32 ; movdqa 0x20(%rsp),%xmm7 - DB 102,68,15,254,207 ; paddd %xmm7,%xmm9 - DB 102,65,15,112,233,245 ; pshufd $0xf5,%xmm9,%xmm5 - DB 102,68,15,244,206 ; pmuludq %xmm6,%xmm9 - DB 102,65,15,112,217,232 ; pshufd $0xe8,%xmm9,%xmm3 - DB 102,15,244,238 ; pmuludq %xmm6,%xmm5 - DB 102,68,15,111,206 ; movdqa %xmm6,%xmm9 - DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 - DB 102,15,98,221 ; punpckldq %xmm5,%xmm3 - DB 102,15,112,235,229 ; pshufd $0xe5,%xmm3,%xmm5 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 102,15,20,221 ; unpcklpd %xmm5,%xmm3 + DB 102,68,15,254,84,36,32 ; paddd 0x20(%rsp),%xmm10 + DB 102,65,15,112,226,245 ; pshufd $0xf5,%xmm10,%xmm4 + DB 102,68,15,244,215 ; pmuludq %xmm7,%xmm10 + DB 102,15,244,231 ; pmuludq %xmm7,%xmm4 + DB 102,15,112,236,232 ; pshufd $0xe8,%xmm4,%xmm5 + DB 102,65,15,112,226,232 ; pshufd $0xe8,%xmm10,%xmm4 + DB 102,15,98,229 ; punpckldq %xmm5,%xmm4 + DB 102,69,15,126,208 ; movd %xmm10,%r8d + DB 102,15,112,236,229 ; pshufd $0xe5,%xmm4,%xmm5 DB 102,65,15,126,233 ; movd %xmm5,%r9d - DB 102,15,112,235,78 ; pshufd $0x4e,%xmm3,%xmm5 - DB 102,65,15,126,232 ; movd %xmm5,%r8d - DB 102,15,112,235,231 ; pshufd $0xe7,%xmm3,%xmm5 + DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 + DB 102,65,15,126,234 ; movd %xmm5,%r10d + DB 102,15,112,236,231 ; pshufd $0xe7,%xmm4,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,52,139 ; movss (%r11,%r9,4),%xmm6 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,69,15,16,4,131 ; movss (%r11,%rax,4),%xmm8 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,60,147 ; movss (%rbx,%r10,4),%xmm7 + DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 + DB 243,70,15,16,4,131 ; movss (%rbx,%r8,4),%xmm8 DB 68,15,20,197 ; unpcklps %xmm5,%xmm8 - DB 68,15,20,198 ; unpcklps %xmm6,%xmm8 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,15,254,232 ; paddd %xmm0,%xmm5 - DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 - DB 102,73,15,126,240 ; movq %xmm6,%r8 - DB 102,73,15,126,233 ; movq %xmm5,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,68,15,20,199 ; unpcklpd %xmm7,%xmm8 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,65,15,250,237 ; psubd %xmm13,%xmm5 + DB 102,15,112,253,78 ; pshufd $0x4e,%xmm5,%xmm7 + DB 102,73,15,126,248 ; movq %xmm7,%r8 + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,52,147 ; movss (%r11,%r10,4),%xmm6 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,4,131 ; movss (%r11,%r8,4),%xmm0 - DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 102,15,254,29,192,192,0,0 ; paddd 0xc0c0(%rip),%xmm3 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 - DB 102,73,15,126,192 ; movq %xmm0,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,70,15,16,44,139 ; movss (%rbx,%r9,4),%xmm13 + DB 68,15,20,237 ; unpcklps %xmm5,%xmm13 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 102,68,15,20,239 ; unpcklpd %xmm7,%xmm13 + DB 102,15,254,224 ; paddd %xmm0,%xmm4 + DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 + DB 102,73,15,126,232 ; movq %xmm5,%r8 + DB 102,72,15,126,224 ; movq %xmm4,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 + DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,252 ; unpcklps %xmm4,%xmm7 + DB 102,15,20,239 ; unpcklpd %xmm7,%xmm5 DB 68,15,92,193 ; subps %xmm1,%xmm8 - DB 69,15,89,197 ; mulps %xmm13,%xmm8 + DB 69,15,89,198 ; mulps %xmm14,%xmm8 DB 68,15,88,193 ; addps %xmm1,%xmm8 - DB 15,92,242 ; subps %xmm2,%xmm6 - DB 65,15,89,245 ; mulps %xmm13,%xmm6 - DB 15,88,242 ; addps %xmm2,%xmm6 - DB 15,92,236 ; subps %xmm4,%xmm5 - DB 65,15,89,237 ; mulps %xmm13,%xmm5 - DB 15,88,236 ; addps %xmm4,%xmm5 + DB 68,15,92,234 ; subps %xmm2,%xmm13 + DB 69,15,89,238 ; mulps %xmm14,%xmm13 + DB 68,15,88,234 ; addps %xmm2,%xmm13 + DB 15,92,235 ; subps %xmm3,%xmm5 + DB 65,15,89,238 ; mulps %xmm14,%xmm5 + DB 15,88,235 ; addps %xmm3,%xmm5 DB 102,68,15,254,36,36 ; paddd (%rsp),%xmm12 - DB 102,69,15,254,244 ; paddd %xmm12,%xmm14 - DB 102,65,15,112,206,245 ; pshufd $0xf5,%xmm14,%xmm1 - DB 102,65,15,111,209 ; movdqa %xmm9,%xmm2 - DB 102,68,15,244,242 ; pmuludq %xmm2,%xmm14 - DB 102,65,15,112,198,232 ; pshufd $0xe8,%xmm14,%xmm0 - DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,111,214 ; movdqa %xmm6,%xmm2 + DB 102,65,15,254,212 ; paddd %xmm12,%xmm2 + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,13,239,194,0,0 ; movdqa 0xc2ef(%rip),%xmm1 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,65,15,126,208 ; movd %xmm2,%r8d DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,65,15,126,201 ; movd %xmm1,%r9d DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,65,15,126,202 ; movd %xmm1,%r10d DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,20,139 ; movss (%r11,%r9,4),%xmm2 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,69,15,16,12,131 ; movss (%r11,%rax,4),%xmm9 - DB 243,67,15,16,12,131 ; movss (%r11,%r8,4),%xmm1 - DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 - DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,70,15,16,20,131 ; movss (%rbx,%r8,4),%xmm10 + DB 68,15,20,209 ; unpcklps %xmm1,%xmm10 + DB 102,68,15,20,210 ; unpcklpd %xmm2,%xmm10 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,254,13,186,191,0,0 ; paddd 0xbfba(%rip),%xmm1 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> + DB 102,15,250,13,148,194,0,0 ; psubd 0xc294(%rip),%xmm1 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 DB 102,73,15,126,208 ; movq %xmm2,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,52,131 ; movss (%r11,%rax,4),%xmm14 - DB 243,67,15,16,20,131 ; movss (%r11,%r8,4),%xmm2 - DB 68,15,20,242 ; unpcklps %xmm2,%xmm14 - DB 68,15,20,241 ; unpcklps %xmm1,%xmm14 - DB 102,15,254,5,140,191,0,0 ; paddd 0xbf8c(%rip),%xmm0 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,15,16,20,131 ; movss (%rbx,%rax,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,15,20,226 ; unpcklpd %xmm2,%xmm4 + DB 102,15,254,5,191,187,0,0 ; paddd 0xbbbf(%rip),%xmm0 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,73,15,126,200 ; movq %xmm1,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,12,131 ; movss (%r11,%r8,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 102,68,15,254,231 ; paddd %xmm7,%xmm12 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 102,15,20,217 ; unpcklpd %xmm1,%xmm3 + DB 102,68,15,254,100,36,32 ; paddd 0x20(%rsp),%xmm12 DB 102,65,15,112,196,245 ; pshufd $0xf5,%xmm12,%xmm0 - DB 102,68,15,244,228 ; pmuludq %xmm4,%xmm12 - DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 - DB 102,65,15,112,212,232 ; pshufd $0xe8,%xmm12,%xmm2 + DB 102,68,15,244,230 ; pmuludq %xmm6,%xmm12 + DB 102,15,244,198 ; pmuludq %xmm6,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,65,15,112,212,232 ; pshufd $0xe8,%xmm12,%xmm2 DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 + DB 102,69,15,126,224 ; movd %xmm12,%r8d DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 DB 102,65,15,126,193 ; movd %xmm0,%r9d DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 - DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,65,15,126,194 ; movd %xmm0,%r10d DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,12,139 ; movss (%r11,%r9,4),%xmm1 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 + DB 243,66,15,16,60,139 ; movss (%rbx,%r9,4),%xmm7 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 102,15,111,37,187,190,0,0 ; movdqa 0xbebb(%rip),%xmm4 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,226 ; paddd %xmm2,%xmm4 - DB 102,15,112,204,78 ; pshufd $0x4e,%xmm4,%xmm1 - DB 102,73,15,126,200 ; movq %xmm1,%r8 - DB 102,73,15,126,225 ; movq %xmm4,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,250,13,158,193,0,0 ; psubd 0xc19e(%rip),%xmm1 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,249,78 ; pshufd $0x4e,%xmm1,%xmm7 + DB 102,73,15,126,248 ; movq %xmm7,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 102,15,254,21,139,190,0,0 ; paddd 0xbe8b(%rip),%xmm2 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 - DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,60,131 ; movss (%rbx,%r8,4),%xmm7 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,247 ; unpcklps %xmm7,%xmm6 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 + DB 102,15,254,21,201,186,0,0 ; paddd 0xbac9(%rip),%xmm2 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,226 ; unpcklps %xmm2,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 15,20,212 ; unpcklps %xmm4,%xmm2 - DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 65,15,89,197 ; mulps %xmm13,%xmm0 - DB 65,15,88,193 ; addps %xmm9,%xmm0 - DB 65,15,92,206 ; subps %xmm14,%xmm1 - DB 65,15,89,205 ; mulps %xmm13,%xmm1 - DB 65,15,88,206 ; addps %xmm14,%xmm1 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 15,20,214 ; unpcklps %xmm6,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,52,131 ; movss (%rbx,%r8,4),%xmm6 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,254 ; unpcklps %xmm6,%xmm7 + DB 102,15,20,215 ; unpcklpd %xmm7,%xmm2 + DB 65,15,92,194 ; subps %xmm10,%xmm0 + DB 65,15,89,198 ; mulps %xmm14,%xmm0 + DB 65,15,88,194 ; addps %xmm10,%xmm0 + DB 15,92,204 ; subps %xmm4,%xmm1 + DB 65,15,89,206 ; mulps %xmm14,%xmm1 + DB 15,88,204 ; addps %xmm4,%xmm1 DB 15,92,211 ; subps %xmm3,%xmm2 - DB 65,15,89,213 ; mulps %xmm13,%xmm2 + DB 65,15,89,214 ; mulps %xmm14,%xmm2 DB 15,88,211 ; addps %xmm3,%xmm2 DB 65,15,92,192 ; subps %xmm8,%xmm0 DB 15,40,92,36,16 ; movaps 0x10(%rsp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 DB 65,15,88,192 ; addps %xmm8,%xmm0 - DB 15,92,206 ; subps %xmm6,%xmm1 + DB 65,15,92,205 ; subps %xmm13,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 15,88,206 ; addps %xmm6,%xmm1 + DB 65,15,88,205 ; addps %xmm13,%xmm1 DB 15,92,213 ; subps %xmm5,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 DB 15,88,213 ; addps %xmm5,%xmm2 - DB 15,91,92,36,48 ; cvtdq2ps 0x30(%rsp),%xmm3 - DB 15,40,100,36,80 ; movaps 0x50(%rsp),%xmm4 + DB 15,91,92,36,80 ; cvtdq2ps 0x50(%rsp),%xmm3 + DB 15,40,100,36,96 ; movaps 0x60(%rsp),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 DB 65,15,92,199 ; subps %xmm15,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -31846,25 +31945,27 @@ _sk_clut_3D_sse2 LABEL PROC DB 65,15,92,203 ; subps %xmm11,%xmm1 DB 15,89,204 ; mulps %xmm4,%xmm1 DB 65,15,88,203 ; addps %xmm11,%xmm1 - DB 65,15,92,210 ; subps %xmm10,%xmm2 + DB 65,15,92,209 ; subps %xmm9,%xmm2 DB 15,89,212 ; mulps %xmm4,%xmm2 - DB 65,15,88,210 ; addps %xmm10,%xmm2 + DB 65,15,88,209 ; addps %xmm9,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 - DB 15,40,100,36,112 ; movaps 0x70(%rsp),%xmm4 - DB 15,40,172,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm5 - DB 15,40,180,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm6 - DB 15,40,188,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm7 - DB 72,129,196,184,0,0,0 ; add $0xb8,%rsp + DB 15,40,92,36,112 ; movaps 0x70(%rsp),%xmm3 + DB 15,40,164,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm4 + DB 15,40,172,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm5 + DB 15,40,180,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm6 + DB 15,40,188,36,176,0,0,0 ; movaps 0xb0(%rsp),%xmm7 + DB 72,129,196,192,0,0,0 ; add $0xc0,%rsp + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax PUBLIC _sk_clut_4D_sse2 _sk_clut_4D_sse2 LABEL PROC - DB 72,129,236,56,1,0,0 ; sub $0x138,%rsp - DB 15,41,188,36,32,1,0,0 ; movaps %xmm7,0x120(%rsp) - DB 15,41,180,36,16,1,0,0 ; movaps %xmm6,0x110(%rsp) - DB 15,41,172,36,0,1,0,0 ; movaps %xmm5,0x100(%rsp) - DB 15,41,164,36,240,0,0,0 ; movaps %xmm4,0xf0(%rsp) + DB 83 ; push %rbx + DB 72,129,236,32,1,0,0 ; sub $0x120,%rsp + DB 15,41,188,36,16,1,0,0 ; movaps %xmm7,0x110(%rsp) + DB 15,41,180,36,0,1,0,0 ; movaps %xmm6,0x100(%rsp) + DB 15,41,172,36,240,0,0,0 ; movaps %xmm5,0xf0(%rsp) + DB 15,41,164,36,224,0,0,0 ; movaps %xmm4,0xe0(%rsp) DB 72,173 ; lods %ds:(%rsi),%rax DB 68,139,64,20 ; mov 0x14(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d @@ -31875,6 +31976,7 @@ _sk_clut_4D_sse2 LABEL PROC DB 15,40,236 ; movaps %xmm4,%xmm5 DB 102,65,15,110,216 ; movd %r8d,%xmm3 DB 102,15,112,251,0 ; pshufd $0x0,%xmm3,%xmm7 + DB 102,15,127,60,36 ; movdqa %xmm7,(%rsp) DB 68,139,64,16 ; mov 0x10(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d DB 102,65,15,110,225 ; movd %r9d,%xmm4 @@ -31882,1239 +31984,1163 @@ _sk_clut_4D_sse2 LABEL PROC DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,89,226 ; mulps %xmm2,%xmm4 DB 243,15,91,244 ; cvttps2dq %xmm4,%xmm6 - DB 102,15,127,116,36,96 ; movdqa %xmm6,0x60(%rsp) - DB 68,15,40,204 ; movaps %xmm4,%xmm9 - DB 68,15,41,140,36,160,0,0,0 ; movaps %xmm9,0xa0(%rsp) + DB 102,15,127,116,36,80 ; movdqa %xmm6,0x50(%rsp) + DB 68,15,40,252 ; movaps %xmm4,%xmm15 + DB 68,15,41,124,36,112 ; movaps %xmm15,0x70(%rsp) DB 102,15,112,214,245 ; pshufd $0xf5,%xmm6,%xmm2 - DB 102,15,111,230 ; movdqa %xmm6,%xmm4 - DB 102,15,127,124,36,80 ; movdqa %xmm7,0x50(%rsp) - DB 102,15,244,231 ; pmuludq %xmm7,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,244,215 ; pmuludq %xmm7,%xmm2 + DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,15,244,230 ; pmuludq %xmm6,%xmm4 + DB 102,68,15,112,244,232 ; pshufd $0xe8,%xmm4,%xmm14 DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,98,226 ; punpckldq %xmm2,%xmm4 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,127,180,36,208,0,0,0 ; movdqa %xmm6,0xd0(%rsp) + DB 102,68,15,98,242 ; punpckldq %xmm2,%xmm14 DB 102,65,15,110,208 ; movd %r8d,%xmm2 DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 DB 102,15,98,211 ; punpckldq %xmm3,%xmm2 - DB 102,68,15,112,250,0 ; pshufd $0x0,%xmm2,%xmm15 + DB 102,68,15,112,226,0 ; pshufd $0x0,%xmm2,%xmm12 DB 68,139,64,12 ; mov 0xc(%rax),%r8d DB 69,141,72,255 ; lea -0x1(%r8),%r9d DB 102,65,15,110,217 ; movd %r9d,%xmm3 DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 - DB 15,91,227 ; cvtdq2ps %xmm3,%xmm4 - DB 15,89,225 ; mulps %xmm1,%xmm4 - DB 243,15,91,220 ; cvttps2dq %xmm4,%xmm3 - DB 68,15,40,220 ; movaps %xmm4,%xmm11 + DB 68,15,91,219 ; cvtdq2ps %xmm3,%xmm11 + DB 68,15,89,217 ; mulps %xmm1,%xmm11 + DB 243,65,15,91,227 ; cvttps2dq %xmm11,%xmm4 + DB 102,15,127,100,36,64 ; movdqa %xmm4,0x40(%rsp) DB 68,15,41,92,36,16 ; movaps %xmm11,0x10(%rsp) - DB 102,15,127,92,36,64 ; movdqa %xmm3,0x40(%rsp) - DB 102,15,112,203,245 ; pshufd $0xf5,%xmm3,%xmm1 - DB 102,65,15,244,223 ; pmuludq %xmm15,%xmm3 + DB 102,15,112,204,245 ; pshufd $0xf5,%xmm4,%xmm1 + DB 102,65,15,244,204 ; pmuludq %xmm12,%xmm1 + DB 102,65,15,111,220 ; movdqa %xmm12,%xmm3 + DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,65,15,112,231,245 ; pshufd $0xf5,%xmm15,%xmm4 - DB 102,15,127,36,36 ; movdqa %xmm4,(%rsp) - DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 - DB 102,15,127,156,36,144,0,0,0 ; movdqa %xmm3,0x90(%rsp) + DB 102,15,127,156,36,128,0,0,0 ; movdqa %xmm3,0x80(%rsp) DB 102,65,15,110,200 ; movd %r8d,%xmm1 DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 + DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 DB 68,139,64,8 ; mov 0x8(%rax),%r8d DB 65,255,200 ; dec %r8d DB 102,65,15,110,208 ; movd %r8d,%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 - DB 68,15,91,242 ; cvtdq2ps %xmm2,%xmm14 - DB 68,15,89,240 ; mulps %xmm0,%xmm14 - DB 102,15,112,193,0 ; pshufd $0x0,%xmm1,%xmm0 - DB 243,69,15,91,230 ; cvttps2dq %xmm14,%xmm12 - DB 102,65,15,111,204 ; movdqa %xmm12,%xmm1 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,225,232 ; pshufd $0xe8,%xmm1,%xmm4 - DB 102,65,15,112,204,245 ; pshufd $0xf5,%xmm12,%xmm1 - DB 102,68,15,112,232,245 ; pshufd $0xf5,%xmm0,%xmm13 - DB 102,65,15,244,205 ; pmuludq %xmm13,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,225 ; punpckldq %xmm1,%xmm4 + DB 68,15,91,234 ; cvtdq2ps %xmm2,%xmm13 + DB 68,15,89,232 ; mulps %xmm0,%xmm13 + DB 243,69,15,91,205 ; cvttps2dq %xmm13,%xmm9 + DB 102,65,15,112,193,245 ; pshufd $0xf5,%xmm9,%xmm0 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,65,15,244,209 ; pmuludq %xmm9,%xmm2 + DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,98,224 ; punpckldq %xmm0,%xmm4 DB 102,15,127,100,36,32 ; movdqa %xmm4,0x20(%rsp) - DB 243,15,91,213 ; cvttps2dq %xmm5,%xmm2 - DB 102,15,127,148,36,176,0,0,0 ; movdqa %xmm2,0xb0(%rsp) + DB 243,15,91,197 ; cvttps2dq %xmm5,%xmm0 + DB 102,15,127,132,36,176,0,0,0 ; movdqa %xmm0,0xb0(%rsp) DB 68,15,40,197 ; movaps %xmm5,%xmm8 - DB 68,15,41,132,36,224,0,0,0 ; movaps %xmm8,0xe0(%rsp) - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,15,254,202 ; paddd %xmm2,%xmm1 + DB 68,15,41,132,36,208,0,0,0 ; movaps %xmm8,0xd0(%rsp) + DB 102,69,15,111,214 ; movdqa %xmm14,%xmm10 + DB 102,68,15,254,208 ; paddd %xmm0,%xmm10 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,254,209 ; paddd %xmm1,%xmm2 + DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 102,15,254,218 ; paddd %xmm2,%xmm3 - DB 102,68,15,111,21,87,194,0,0 ; movdqa 0xc257(%rip),%xmm10 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,112,235,245 ; pshufd $0xf5,%xmm3,%xmm5 - DB 102,65,15,244,218 ; pmuludq %xmm10,%xmm3 - DB 102,15,112,227,232 ; pshufd $0xe8,%xmm3,%xmm4 - DB 102,65,15,244,234 ; pmuludq %xmm10,%xmm5 - DB 102,15,112,221,232 ; pshufd $0xe8,%xmm5,%xmm3 - DB 102,15,98,227 ; punpckldq %xmm3,%xmm4 - DB 102,15,112,220,229 ; pshufd $0xe5,%xmm4,%xmm3 + DB 102,15,111,37,194,190,0,0 ; movdqa 0xbec2(%rip),%xmm4 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,112,195,245 ; pshufd $0xf5,%xmm3,%xmm0 + DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 + DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 + DB 102,15,112,224,232 ; pshufd $0xe8,%xmm0,%xmm4 + DB 102,15,112,195,232 ; pshufd $0xe8,%xmm3,%xmm0 + DB 102,15,98,196 ; punpckldq %xmm4,%xmm0 DB 102,65,15,126,216 ; movd %xmm3,%r8d - DB 102,15,112,220,78 ; pshufd $0x4e,%xmm4,%xmm3 + DB 102,15,112,216,229 ; pshufd $0xe5,%xmm0,%xmm3 DB 102,65,15,126,217 ; movd %xmm3,%r9d - DB 102,15,112,220,231 ; pshufd $0xe7,%xmm4,%xmm3 + DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 DB 102,65,15,126,218 ; movd %xmm3,%r10d - DB 76,139,24 ; mov (%rax),%r11 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,52,139 ; movss (%r11,%r9,4),%xmm6 - DB 15,20,222 ; unpcklps %xmm6,%xmm3 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 - DB 102,15,111,236 ; movdqa %xmm4,%xmm5 - DB 102,15,254,45,88,187,0,0 ; paddd 0xbb58(%rip),%xmm5 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 - DB 102,73,15,126,240 ; movq %xmm6,%r8 - DB 102,73,15,126,233 ; movq %xmm5,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,52,147 ; movss (%r11,%r10,4),%xmm6 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,247 ; unpcklps %xmm7,%xmm6 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 102,15,254,37,44,187,0,0 ; paddd 0xbb2c(%rip),%xmm4 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,225 ; movq %xmm4,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 - DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,239 ; unpcklps %xmm7,%xmm5 + DB 102,15,112,216,231 ; pshufd $0xe7,%xmm0,%xmm3 + DB 102,65,15,126,219 ; movd %xmm3,%r11d + DB 72,139,24 ; mov (%rax),%rbx + DB 243,66,15,16,28,155 ; movss (%rbx,%r11,4),%xmm3 + DB 243,66,15,16,36,147 ; movss (%rbx,%r10,4),%xmm4 + DB 15,20,227 ; unpcklps %xmm3,%xmm4 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,66,15,16,60,131 ; movss (%rbx,%r8,4),%xmm7 + DB 15,20,251 ; unpcklps %xmm3,%xmm7 + DB 102,15,20,252 ; unpcklpd %xmm4,%xmm7 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,250,220 ; psubd %xmm4,%xmm3 + DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,73,15,126,217 ; movq %xmm3,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 243,66,15,16,44,147 ; movss (%rbx,%r10,4),%xmm5 DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 15,40,61,65,193,0,0 ; movaps 0xc141(%rip),%xmm7 # 2db50 <_sk_overlay_sse2_8bit+0x15f8> - DB 68,15,88,199 ; addps %xmm7,%xmm8 + DB 69,137,193 ; mov %r8d,%r9d + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 + DB 102,15,254,5,140,183,0,0 ; paddd 0xb78c(%rip),%xmm0 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,73,15,126,193 ; movq %xmm0,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,32 ; shr $0x20,%r9 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 15,20,220 ; unpcklps %xmm4,%xmm3 + DB 69,137,193 ; mov %r8d,%r9d + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 102,15,20,222 ; unpcklpd %xmm6,%xmm3 + DB 15,40,53,182,189,0,0 ; movaps 0xbdb6(%rip),%xmm6 # 2da50 <_sk_overlay_sse2_8bit+0x15a1> + DB 68,15,88,198 ; addps %xmm6,%xmm8 DB 68,15,41,68,36,48 ; movaps %xmm8,0x30(%rsp) - DB 68,15,88,207 ; addps %xmm7,%xmm9 - DB 68,15,41,140,36,192,0,0,0 ; movaps %xmm9,0xc0(%rsp) + DB 68,15,88,254 ; addps %xmm6,%xmm15 + DB 68,15,41,188,36,192,0,0,0 ; movaps %xmm15,0xc0(%rsp) DB 69,15,40,195 ; movaps %xmm11,%xmm8 - DB 68,15,88,199 ; addps %xmm7,%xmm8 - DB 65,15,88,254 ; addps %xmm14,%xmm7 - DB 243,15,91,255 ; cvttps2dq %xmm7,%xmm7 - DB 102,15,244,199 ; pmuludq %xmm7,%xmm0 - DB 102,15,112,255,245 ; pshufd $0xf5,%xmm7,%xmm7 - DB 102,65,15,244,253 ; pmuludq %xmm13,%xmm7 - DB 102,68,15,112,216,232 ; pshufd $0xe8,%xmm0,%xmm11 - DB 102,15,112,199,232 ; pshufd $0xe8,%xmm7,%xmm0 - DB 102,68,15,98,216 ; punpckldq %xmm0,%xmm11 - DB 102,65,15,254,211 ; paddd %xmm11,%xmm2 - DB 102,68,15,127,156,36,128,0,0,0 ; movdqa %xmm11,0x80(%rsp) - DB 102,15,112,250,245 ; pshufd $0xf5,%xmm2,%xmm7 - DB 102,65,15,244,210 ; pmuludq %xmm10,%xmm2 - DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 - DB 102,65,15,244,250 ; pmuludq %xmm10,%xmm7 - DB 102,15,112,215,232 ; pshufd $0xe8,%xmm7,%xmm2 - DB 102,15,98,194 ; punpckldq %xmm2,%xmm0 - DB 102,15,112,208,229 ; pshufd $0xe5,%xmm0,%xmm2 - DB 102,65,15,126,209 ; movd %xmm2,%r9d - DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 65,15,88,245 ; addps %xmm13,%xmm6 + DB 243,15,91,246 ; cvttps2dq %xmm6,%xmm6 + DB 102,15,112,198,245 ; pshufd $0xf5,%xmm6,%xmm0 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,244,206 ; pmuludq %xmm6,%xmm1 + DB 102,68,15,112,249,232 ; pshufd $0xe8,%xmm1,%xmm15 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,68,15,98,248 ; punpckldq %xmm0,%xmm15 + DB 102,65,15,254,215 ; paddd %xmm15,%xmm2 + DB 102,68,15,127,124,36,96 ; movdqa %xmm15,0x60(%rsp) + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,13,105,189,0,0 ; movdqa 0xbd69(%rip),%xmm1 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,112,202,232 ; pshufd $0xe8,%xmm2,%xmm1 + DB 102,15,98,200 ; punpckldq %xmm0,%xmm1 DB 102,65,15,126,208 ; movd %xmm2,%r8d - DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,60,139 ; movss (%r11,%r9,4),%xmm7 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 + DB 102,15,112,193,229 ; pshufd $0xe5,%xmm1,%xmm0 + DB 102,65,15,126,193 ; movd %xmm0,%r9d + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,65,15,126,194 ; movd %xmm0,%r10d + DB 102,15,112,193,231 ; pshufd $0xe7,%xmm1,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,69,15,16,44,131 ; movss (%r11,%rax,4),%xmm13 - DB 243,67,15,16,20,131 ; movss (%r11,%r8,4),%xmm2 - DB 68,15,20,234 ; unpcklps %xmm2,%xmm13 - DB 68,15,20,239 ; unpcklps %xmm7,%xmm13 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,254,21,0,186,0,0 ; paddd 0xba00(%rip),%xmm2 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,112,250,78 ; pshufd $0x4e,%xmm2,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 243,66,15,16,52,131 ; movss (%rbx,%r8,4),%xmm6 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 102,15,20,242 ; unpcklpd %xmm2,%xmm6 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,250,5,20,189,0,0 ; psubd 0xbd14(%rip),%xmm0 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 + DB 102,73,15,126,208 ; movq %xmm2,%r8 + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,71,15,16,12,131 ; movss (%r11,%r8,4),%xmm9 - DB 65,15,20,209 ; unpcklps %xmm9,%xmm2 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 102,15,254,5,211,185,0,0 ; paddd 0xb9d3(%rip),%xmm0 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,112,248,78 ; pshufd $0x4e,%xmm0,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,68,15,16,28,131 ; movss (%rbx,%rax,4),%xmm11 + DB 68,15,20,216 ; unpcklps %xmm0,%xmm11 + DB 102,65,15,20,211 ; unpcklpd %xmm11,%xmm2 + DB 102,15,254,13,60,182,0,0 ; paddd 0xb63c(%rip),%xmm1 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,73,15,126,192 ; movq %xmm0,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,4,147 ; movss (%r11,%r10,4),%xmm0 - DB 15,20,248 ; unpcklps %xmm0,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 65,15,91,228 ; cvtdq2ps %xmm12,%xmm4 - DB 68,15,92,244 ; subps %xmm4,%xmm14 - DB 68,15,92,235 ; subps %xmm3,%xmm13 - DB 69,15,89,238 ; mulps %xmm14,%xmm13 - DB 68,15,88,235 ; addps %xmm3,%xmm13 - DB 15,92,214 ; subps %xmm6,%xmm2 - DB 65,15,89,214 ; mulps %xmm14,%xmm2 - DB 15,88,214 ; addps %xmm6,%xmm2 - DB 15,92,197 ; subps %xmm5,%xmm0 - DB 65,15,89,198 ; mulps %xmm14,%xmm0 - DB 69,15,40,206 ; movaps %xmm14,%xmm9 - DB 15,88,197 ; addps %xmm5,%xmm0 - DB 243,65,15,91,216 ; cvttps2dq %xmm8,%xmm3 - DB 102,68,15,244,251 ; pmuludq %xmm3,%xmm15 - DB 102,15,112,219,245 ; pshufd $0xf5,%xmm3,%xmm3 - DB 102,15,244,28,36 ; pmuludq (%rsp),%xmm3 - DB 102,65,15,112,231,232 ; pshufd $0xe8,%xmm15,%xmm4 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,20,204 ; unpcklpd %xmm4,%xmm1 + DB 65,15,91,193 ; cvtdq2ps %xmm9,%xmm0 + DB 68,15,92,232 ; subps %xmm0,%xmm13 + DB 15,92,247 ; subps %xmm7,%xmm6 + DB 65,15,89,245 ; mulps %xmm13,%xmm6 + DB 15,88,247 ; addps %xmm7,%xmm6 + DB 15,92,213 ; subps %xmm5,%xmm2 + DB 65,15,89,213 ; mulps %xmm13,%xmm2 + DB 15,88,213 ; addps %xmm5,%xmm2 + DB 15,92,203 ; subps %xmm3,%xmm1 + DB 65,15,89,205 ; mulps %xmm13,%xmm1 + DB 15,88,203 ; addps %xmm3,%xmm1 + DB 243,65,15,91,192 ; cvttps2dq %xmm8,%xmm0 + DB 102,15,112,216,245 ; pshufd $0xf5,%xmm0,%xmm3 + DB 102,65,15,244,220 ; pmuludq %xmm12,%xmm3 + DB 102,68,15,244,224 ; pmuludq %xmm0,%xmm12 + DB 102,65,15,112,228,232 ; pshufd $0xe8,%xmm12,%xmm4 + DB 102,15,112,195,232 ; pshufd $0xe8,%xmm3,%xmm0 + DB 102,15,98,224 ; punpckldq %xmm0,%xmm4 + DB 102,15,127,164,36,144,0,0,0 ; movdqa %xmm4,0x90(%rsp) + DB 102,68,15,254,212 ; paddd %xmm4,%xmm10 + DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 + DB 102,15,254,68,36,32 ; paddd 0x20(%rsp),%xmm0 + DB 102,15,112,216,245 ; pshufd $0xf5,%xmm0,%xmm3 + DB 102,15,111,37,16,188,0,0 ; movdqa 0xbc10(%rip),%xmm4 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 + DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 + DB 102,68,15,111,220 ; movdqa %xmm4,%xmm11 DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 + DB 102,15,112,224,232 ; pshufd $0xe8,%xmm0,%xmm4 DB 102,15,98,227 ; punpckldq %xmm3,%xmm4 - DB 102,15,127,36,36 ; movdqa %xmm4,(%rsp) - DB 102,15,254,204 ; paddd %xmm4,%xmm1 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,15,254,92,36,32 ; paddd 0x20(%rsp),%xmm3 - DB 102,15,112,227,245 ; pshufd $0xf5,%xmm3,%xmm4 - DB 102,68,15,111,53,143,191,0,0 ; movdqa 0xbf8f(%rip),%xmm14 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,65,15,244,222 ; pmuludq %xmm14,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,65,15,244,230 ; pmuludq %xmm14,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 - DB 102,15,98,220 ; punpckldq %xmm4,%xmm3 - DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 - DB 102,65,15,126,225 ; movd %xmm4,%r9d - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 - DB 102,65,15,126,224 ; movd %xmm4,%r8d - DB 102,15,112,227,231 ; pshufd $0xe7,%xmm3,%xmm4 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,139 ; movss (%r11,%r9,4),%xmm5 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,69,15,16,36,131 ; movss (%r11,%rax,4),%xmm12 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 68,15,20,228 ; unpcklps %xmm4,%xmm12 - DB 68,15,20,229 ; unpcklps %xmm5,%xmm12 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,111,61,151,184,0,0 ; movdqa 0xb897(%rip),%xmm7 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,231 ; paddd %xmm7,%xmm4 + DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,15,112,196,229 ; pshufd $0xe5,%xmm4,%xmm0 + DB 102,65,15,126,193 ; movd %xmm0,%r9d + DB 102,15,112,196,78 ; pshufd $0x4e,%xmm4,%xmm0 + DB 102,65,15,126,194 ; movd %xmm0,%r10d + DB 102,15,112,196,231 ; pshufd $0xe7,%xmm4,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,44,147 ; movss (%rbx,%r10,4),%xmm5 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 243,70,15,16,12,131 ; movss (%rbx,%r8,4),%xmm9 + DB 68,15,20,200 ; unpcklps %xmm0,%xmm9 + DB 102,68,15,20,205 ; unpcklpd %xmm5,%xmm9 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,69,15,118,228 ; pcmpeqd %xmm12,%xmm12 + DB 102,65,15,250,196 ; psubd %xmm12,%xmm0 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 + DB 102,73,15,126,232 ; movq %xmm5,%r8 + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 15,20,197 ; unpcklps %xmm5,%xmm0 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 102,15,20,199 ; unpcklpd %xmm7,%xmm0 + DB 102,68,15,111,5,220,180,0,0 ; movdqa 0xb4dc(%rip),%xmm8 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,224 ; paddd %xmm8,%xmm4 DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,225 ; movq %xmm4,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,224 ; movq %xmm4,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 102,68,15,111,5,102,184,0,0 ; movdqa 0xb866(%rip),%xmm8 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,65,15,254,216 ; paddd %xmm8,%xmm3 - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 - DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 - DB 15,20,220 ; unpcklps %xmm4,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 102,65,15,254,203 ; paddd %xmm11,%xmm1 - DB 102,15,112,217,245 ; pshufd $0xf5,%xmm1,%xmm3 - DB 102,65,15,244,206 ; pmuludq %xmm14,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,65,15,244,222 ; pmuludq %xmm14,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,15,98,203 ; punpckldq %xmm3,%xmm1 - DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 - DB 102,65,15,126,217 ; movd %xmm3,%r9d - DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 - DB 102,65,15,126,216 ; movd %xmm3,%r8d - DB 102,15,112,217,231 ; pshufd $0xe7,%xmm1,%xmm3 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,36,139 ; movss (%r11,%r9,4),%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,69,15,16,60,131 ; movss (%r11,%rax,4),%xmm15 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 68,15,20,251 ; unpcklps %xmm3,%xmm15 - DB 68,15,20,252 ; unpcklps %xmm4,%xmm15 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,15,254,223 ; paddd %xmm7,%xmm3 - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 - DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 102,15,20,231 ; unpcklpd %xmm7,%xmm4 + DB 102,69,15,254,215 ; paddd %xmm15,%xmm10 + DB 102,65,15,112,234,245 ; pshufd $0xf5,%xmm10,%xmm5 + DB 102,69,15,244,211 ; pmuludq %xmm11,%xmm10 + DB 102,65,15,244,235 ; pmuludq %xmm11,%xmm5 + DB 102,69,15,111,251 ; movdqa %xmm11,%xmm15 + DB 102,15,112,253,232 ; pshufd $0xe8,%xmm5,%xmm7 + DB 102,65,15,112,234,232 ; pshufd $0xe8,%xmm10,%xmm5 + DB 102,15,98,239 ; punpckldq %xmm7,%xmm5 + DB 102,69,15,126,208 ; movd %xmm10,%r8d + DB 102,15,112,253,229 ; pshufd $0xe5,%xmm5,%xmm7 + DB 102,65,15,126,249 ; movd %xmm7,%r9d + DB 102,15,112,253,78 ; pshufd $0x4e,%xmm5,%xmm7 + DB 102,65,15,126,250 ; movd %xmm7,%r10d + DB 102,15,112,253,231 ; pshufd $0xe7,%xmm5,%xmm7 + DB 102,15,126,248 ; movd %xmm7,%eax + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 243,66,15,16,28,147 ; movss (%rbx,%r10,4),%xmm3 + DB 15,20,223 ; unpcklps %xmm7,%xmm3 + DB 243,66,15,16,60,139 ; movss (%rbx,%r9,4),%xmm7 + DB 243,70,15,16,28,131 ; movss (%rbx,%r8,4),%xmm11 + DB 68,15,20,223 ; unpcklps %xmm7,%xmm11 + DB 102,68,15,20,219 ; unpcklpd %xmm3,%xmm11 + DB 102,15,111,221 ; movdqa %xmm5,%xmm3 + DB 102,65,15,250,220 ; psubd %xmm12,%xmm3 + DB 102,15,112,251,78 ; pshufd $0x4e,%xmm3,%xmm7 + DB 102,73,15,126,248 ; movq %xmm7,%r8 + DB 102,72,15,126,216 ; movq %xmm3,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,52,131 ; movss (%r11,%rax,4),%xmm14 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 68,15,20,247 ; unpcklps %xmm7,%xmm14 - DB 68,15,20,244 ; unpcklps %xmm4,%xmm14 - DB 102,65,15,254,200 ; paddd %xmm8,%xmm1 - DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 - DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,70,15,16,20,139 ; movss (%rbx,%r9,4),%xmm10 + DB 68,15,20,211 ; unpcklps %xmm3,%xmm10 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,251 ; unpcklps %xmm3,%xmm7 + DB 102,68,15,20,215 ; unpcklpd %xmm7,%xmm10 + DB 102,65,15,254,232 ; paddd %xmm8,%xmm5 + DB 102,15,112,221,78 ; pshufd $0x4e,%xmm5,%xmm3 + DB 102,73,15,126,216 ; movq %xmm3,%r8 + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,20,131 ; movss (%r11,%rax,4),%xmm10 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 68,15,20,215 ; unpcklps %xmm7,%xmm10 - DB 68,15,20,209 ; unpcklps %xmm1,%xmm10 - DB 69,15,92,252 ; subps %xmm12,%xmm15 - DB 68,15,41,76,36,112 ; movaps %xmm9,0x70(%rsp) - DB 69,15,89,249 ; mulps %xmm9,%xmm15 - DB 69,15,88,252 ; addps %xmm12,%xmm15 - DB 68,15,92,246 ; subps %xmm6,%xmm14 - DB 69,15,89,241 ; mulps %xmm9,%xmm14 - DB 68,15,88,246 ; addps %xmm6,%xmm14 - DB 68,15,92,213 ; subps %xmm5,%xmm10 - DB 69,15,89,209 ; mulps %xmm9,%xmm10 - DB 68,15,88,213 ; addps %xmm5,%xmm10 - DB 15,91,76,36,64 ; cvtdq2ps 0x40(%rsp),%xmm1 + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,70,15,16,36,139 ; movss (%rbx,%r9,4),%xmm12 + DB 68,15,20,227 ; unpcklps %xmm3,%xmm12 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,251 ; unpcklps %xmm3,%xmm7 + DB 102,68,15,20,231 ; unpcklpd %xmm7,%xmm12 + DB 69,15,92,217 ; subps %xmm9,%xmm11 + DB 68,15,41,172,36,160,0,0,0 ; movaps %xmm13,0xa0(%rsp) + DB 69,15,89,221 ; mulps %xmm13,%xmm11 + DB 69,15,88,217 ; addps %xmm9,%xmm11 + DB 68,15,92,208 ; subps %xmm0,%xmm10 + DB 69,15,89,213 ; mulps %xmm13,%xmm10 + DB 68,15,88,208 ; addps %xmm0,%xmm10 + DB 68,15,92,228 ; subps %xmm4,%xmm12 + DB 69,15,89,229 ; mulps %xmm13,%xmm12 + DB 68,15,88,228 ; addps %xmm4,%xmm12 + DB 15,91,68,36,64 ; cvtdq2ps 0x40(%rsp),%xmm0 DB 15,40,92,36,16 ; movaps 0x10(%rsp),%xmm3 - DB 15,92,217 ; subps %xmm1,%xmm3 + DB 15,92,216 ; subps %xmm0,%xmm3 DB 15,41,92,36,16 ; movaps %xmm3,0x10(%rsp) - DB 69,15,92,253 ; subps %xmm13,%xmm15 - DB 68,15,89,251 ; mulps %xmm3,%xmm15 - DB 69,15,88,253 ; addps %xmm13,%xmm15 - DB 68,15,92,242 ; subps %xmm2,%xmm14 - DB 68,15,89,243 ; mulps %xmm3,%xmm14 - DB 68,15,88,242 ; addps %xmm2,%xmm14 - DB 68,15,92,208 ; subps %xmm0,%xmm10 + DB 68,15,92,222 ; subps %xmm6,%xmm11 + DB 68,15,89,219 ; mulps %xmm3,%xmm11 + DB 68,15,88,222 ; addps %xmm6,%xmm11 + DB 68,15,92,210 ; subps %xmm2,%xmm10 DB 68,15,89,211 ; mulps %xmm3,%xmm10 - DB 68,15,88,208 ; addps %xmm0,%xmm10 + DB 68,15,88,210 ; addps %xmm2,%xmm10 + DB 68,15,92,225 ; subps %xmm1,%xmm12 + DB 68,15,89,227 ; mulps %xmm3,%xmm12 + DB 68,15,88,225 ; addps %xmm1,%xmm12 DB 243,15,91,132,36,192,0,0,0 ; cvttps2dq 0xc0(%rsp),%xmm0 DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,84,36,80 ; movdqa 0x50(%rsp),%xmm2 - DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,111,20,36 ; movdqa (%rsp),%xmm2 DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 + DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 + DB 102,68,15,112,202,232 ; pshufd $0xe8,%xmm2,%xmm9 DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 - DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 - DB 102,15,127,84,36,64 ; movdqa %xmm2,0x40(%rsp) - DB 102,68,15,111,218 ; movdqa %xmm2,%xmm11 - DB 102,68,15,254,156,36,176,0,0,0 ; paddd 0xb0(%rsp),%xmm11 - DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 - DB 102,15,254,132,36,144,0,0,0 ; paddd 0x90(%rsp),%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,68,15,111,108,36,32 ; movdqa 0x20(%rsp),%xmm13 - DB 102,65,15,254,205 ; paddd %xmm13,%xmm1 - DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 - DB 102,15,111,29,196,188,0,0 ; movdqa 0xbcc4(%rip),%xmm3 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9 + DB 102,68,15,127,12,36 ; movdqa %xmm9,(%rsp) + DB 102,68,15,254,140,36,176,0,0,0 ; paddd 0xb0(%rsp),%xmm9 + DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 + DB 102,15,254,132,36,128,0,0,0 ; paddd 0x80(%rsp),%xmm0 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,68,15,111,68,36,32 ; movdqa 0x20(%rsp),%xmm8 + DB 102,65,15,254,208 ; paddd %xmm8,%xmm2 + DB 102,15,112,202,245 ; pshufd $0xf5,%xmm2,%xmm1 + DB 102,65,15,111,223 ; movdqa %xmm15,%xmm3 DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 + DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 + DB 102,15,112,202,232 ; pshufd $0xe8,%xmm2,%xmm1 + DB 102,15,98,203 ; punpckldq %xmm3,%xmm1 + DB 102,65,15,126,208 ; movd %xmm2,%r8d DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 DB 102,65,15,126,209 ; movd %xmm2,%r9d DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,65,15,126,208 ; movd %xmm2,%r8d + DB 102,65,15,126,210 ; movd %xmm2,%r10d DB 102,15,112,209,231 ; pshufd $0xe7,%xmm1,%xmm2 DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,44,139 ; movss (%r11,%r9,4),%xmm5 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,69,15,16,12,131 ; movss (%r11,%rax,4),%xmm9 - DB 243,67,15,16,20,131 ; movss (%r11,%r8,4),%xmm2 - DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 - DB 68,15,20,205 ; unpcklps %xmm5,%xmm9 + DB 243,15,16,20,131 ; movss (%rbx,%rax,4),%xmm2 + DB 243,66,15,16,36,147 ; movss (%rbx,%r10,4),%xmm4 + DB 15,20,226 ; unpcklps %xmm2,%xmm4 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 102,15,20,220 ; unpcklpd %xmm4,%xmm3 DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,111,29,202,181,0,0 ; movdqa 0xb5ca(%rip),%xmm3 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,112,234,78 ; pshufd $0x4e,%xmm2,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,15,118,237 ; pcmpeqd %xmm5,%xmm5 + DB 102,15,250,213 ; psubd %xmm5,%xmm2 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,52,131 ; movss (%r11,%r8,4),%xmm6 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 15,20,213 ; unpcklps %xmm5,%xmm2 - DB 102,69,15,111,224 ; movdqa %xmm8,%xmm12 - DB 102,65,15,254,204 ; paddd %xmm12,%xmm1 - DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,20,131 ; movss (%rbx,%rax,4),%xmm2 + DB 243,66,15,16,60,139 ; movss (%rbx,%r9,4),%xmm7 + DB 15,20,250 ; unpcklps %xmm2,%xmm7 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 15,20,226 ; unpcklps %xmm2,%xmm4 + DB 102,15,20,252 ; unpcklpd %xmm4,%xmm7 + DB 102,68,15,111,61,43,178,0,0 ; movdqa 0xb22b(%rip),%xmm15 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,207 ; paddd %xmm15,%xmm1 + DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 + DB 102,73,15,126,208 ; movq %xmm2,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 102,15,111,164,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm4 - DB 102,15,254,196 ; paddd %xmm4,%xmm0 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 102,15,20,212 ; unpcklpd %xmm4,%xmm2 + DB 102,68,15,111,108,36,96 ; movdqa 0x60(%rsp),%xmm13 + DB 102,65,15,254,197 ; paddd %xmm13,%xmm0 DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,244,199 ; pmuludq %xmm7,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 + DB 102,15,244,198 ; pmuludq %xmm6,%xmm0 + DB 102,15,244,206 ; pmuludq %xmm6,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,65,15,126,201 ; movd %xmm1,%r9d - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,65,15,126,200 ; movd %xmm1,%r8d - DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,44,139 ; movss (%r11,%r9,4),%xmm5 - DB 15,20,233 ; unpcklps %xmm1,%xmm5 + DB 102,15,112,224,232 ; pshufd $0xe8,%xmm0,%xmm4 + DB 102,15,98,225 ; punpckldq %xmm1,%xmm4 + DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,15,112,196,229 ; pshufd $0xe5,%xmm4,%xmm0 + DB 102,65,15,126,193 ; movd %xmm0,%r9d + DB 102,15,112,196,78 ; pshufd $0x4e,%xmm4,%xmm0 + DB 102,65,15,126,194 ; movd %xmm0,%r10d + DB 102,15,112,196,231 ; pshufd $0xe7,%xmm4,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,69,15,16,4,131 ; movss (%r11,%rax,4),%xmm8 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 68,15,20,199 ; unpcklps %xmm7,%xmm8 - DB 68,15,20,197 ; unpcklps %xmm5,%xmm8 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,254,235 ; paddd %xmm3,%xmm5 - DB 102,15,112,253,78 ; pshufd $0x4e,%xmm5,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,233 ; movq %xmm5,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,52,147 ; movss (%rbx,%r10,4),%xmm6 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,250,197 ; psubd %xmm5,%xmm0 + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 - DB 15,20,253 ; unpcklps %xmm5,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 15,20,239 ; unpcklps %xmm7,%xmm5 - DB 102,65,15,254,196 ; paddd %xmm12,%xmm0 - DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 - DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 15,20,198 ; unpcklps %xmm6,%xmm0 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,52,131 ; movss (%rbx,%r8,4),%xmm6 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,238 ; unpcklps %xmm6,%xmm5 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 + DB 102,65,15,254,231 ; paddd %xmm15,%xmm4 + DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 + DB 102,73,15,126,232 ; movq %xmm5,%r8 + DB 102,72,15,126,224 ; movq %xmm4,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,4,147 ; movss (%r11,%r10,4),%xmm0 - DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 15,20,203 ; unpcklps %xmm3,%xmm1 - DB 69,15,92,193 ; subps %xmm9,%xmm8 - DB 15,40,68,36,112 ; movaps 0x70(%rsp),%xmm0 - DB 68,15,89,192 ; mulps %xmm0,%xmm8 - DB 69,15,88,193 ; addps %xmm9,%xmm8 - DB 15,92,234 ; subps %xmm2,%xmm5 - DB 15,89,232 ; mulps %xmm0,%xmm5 - DB 15,88,234 ; addps %xmm2,%xmm5 - DB 15,92,206 ; subps %xmm6,%xmm1 - DB 15,89,200 ; mulps %xmm0,%xmm1 - DB 68,15,40,224 ; movaps %xmm0,%xmm12 - DB 15,88,206 ; addps %xmm6,%xmm1 - DB 102,68,15,254,28,36 ; paddd (%rsp),%xmm11 - DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 - DB 102,65,15,254,213 ; paddd %xmm13,%xmm2 - DB 102,15,112,218,245 ; pshufd $0xf5,%xmm2,%xmm3 - DB 102,15,111,5,127,186,0,0 ; movdqa 0xba7f(%rip),%xmm0 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 - DB 102,15,112,250,232 ; pshufd $0xe8,%xmm2,%xmm7 - DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,36,139 ; movss (%rbx,%r9,4),%xmm4 + DB 15,20,229 ; unpcklps %xmm5,%xmm4 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 102,15,20,230 ; unpcklpd %xmm6,%xmm4 + DB 15,92,203 ; subps %xmm3,%xmm1 + DB 15,40,172,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm5 + DB 15,89,205 ; mulps %xmm5,%xmm1 + DB 15,88,203 ; addps %xmm3,%xmm1 + DB 15,92,199 ; subps %xmm7,%xmm0 + DB 15,89,197 ; mulps %xmm5,%xmm0 + DB 15,88,199 ; addps %xmm7,%xmm0 + DB 15,92,226 ; subps %xmm2,%xmm4 + DB 15,89,229 ; mulps %xmm5,%xmm4 + DB 15,88,226 ; addps %xmm2,%xmm4 + DB 102,68,15,254,140,36,144,0,0,0 ; paddd 0x90(%rsp),%xmm9 + DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 + DB 102,65,15,254,216 ; paddd %xmm8,%xmm3 + DB 102,15,112,211,245 ; pshufd $0xf5,%xmm3,%xmm2 + DB 102,15,111,45,52,183,0,0 ; movdqa 0xb734(%rip),%xmm5 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,221 ; pmuludq %xmm5,%xmm3 + DB 102,15,244,213 ; pmuludq %xmm5,%xmm2 + DB 102,15,112,234,232 ; pshufd $0xe8,%xmm2,%xmm5 DB 102,15,112,211,232 ; pshufd $0xe8,%xmm3,%xmm2 - DB 102,15,98,250 ; punpckldq %xmm2,%xmm7 - DB 102,15,112,215,229 ; pshufd $0xe5,%xmm7,%xmm2 - DB 102,65,15,126,209 ; movd %xmm2,%r9d - DB 102,15,112,215,78 ; pshufd $0x4e,%xmm7,%xmm2 - DB 102,65,15,126,208 ; movd %xmm2,%r8d - DB 102,15,112,215,231 ; pshufd $0xe7,%xmm7,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,28,139 ; movss (%r11,%r9,4),%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 102,15,126,248 ; movd %xmm7,%eax - DB 243,69,15,16,12,131 ; movss (%r11,%rax,4),%xmm9 - DB 243,67,15,16,20,131 ; movss (%r11,%r8,4),%xmm2 - DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 - DB 68,15,20,203 ; unpcklps %xmm3,%xmm9 - DB 102,15,111,215 ; movdqa %xmm7,%xmm2 - DB 102,15,111,5,137,179,0,0 ; movdqa 0xb389(%rip),%xmm0 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,208 ; paddd %xmm0,%xmm2 - DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 - DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 15,20,243 ; unpcklps %xmm3,%xmm6 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 - DB 102,68,15,111,45,88,179,0,0 ; movdqa 0xb358(%rip),%xmm13 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,65,15,254,253 ; paddd %xmm13,%xmm7 - DB 102,15,112,215,78 ; pshufd $0x4e,%xmm7,%xmm2 - DB 102,73,15,126,208 ; movq %xmm2,%r8 - DB 102,73,15,126,249 ; movq %xmm7,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax - DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 15,20,251 ; unpcklps %xmm3,%xmm7 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 - DB 102,68,15,254,220 ; paddd %xmm4,%xmm11 - DB 102,65,15,112,219,245 ; pshufd $0xf5,%xmm11,%xmm3 - DB 102,15,111,37,108,185,0,0 ; movdqa 0xb96c(%rip),%xmm4 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,68,15,244,220 ; pmuludq %xmm4,%xmm11 - DB 102,65,15,112,211,232 ; pshufd $0xe8,%xmm11,%xmm2 - DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,15,98,211 ; punpckldq %xmm3,%xmm2 + DB 102,15,98,213 ; punpckldq %xmm5,%xmm2 + DB 102,65,15,126,216 ; movd %xmm3,%r8d DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 DB 102,65,15,126,217 ; movd %xmm3,%r9d DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 - DB 102,65,15,126,216 ; movd %xmm3,%r8d + DB 102,65,15,126,218 ; movd %xmm3,%r10d DB 102,15,112,218,231 ; pshufd $0xe7,%xmm2,%xmm3 DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,69,15,16,28,131 ; movss (%r11,%rax,4),%xmm11 - DB 243,67,15,16,28,139 ; movss (%r11,%r9,4),%xmm3 - DB 65,15,20,219 ; unpcklps %xmm11,%xmm3 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,69,15,16,28,131 ; movss (%r11,%rax,4),%xmm11 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 68,15,20,220 ; unpcklps %xmm4,%xmm11 - DB 68,15,20,219 ; unpcklps %xmm3,%xmm11 - DB 102,15,111,218 ; movdqa %xmm2,%xmm3 - DB 102,15,254,216 ; paddd %xmm0,%xmm3 - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 - DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,66,15,16,44,147 ; movss (%rbx,%r10,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 243,70,15,16,4,131 ; movss (%rbx,%r8,4),%xmm8 + DB 68,15,20,198 ; unpcklps %xmm6,%xmm8 + DB 102,68,15,20,197 ; unpcklpd %xmm5,%xmm8 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,15,250,45,221,182,0,0 ; psubd 0xb6dd(%rip),%xmm5 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 - DB 15,20,220 ; unpcklps %xmm4,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 102,65,15,254,213 ; paddd %xmm13,%xmm2 - DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,60,139 ; movss (%rbx,%r9,4),%xmm7 + DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 102,15,20,254 ; unpcklpd %xmm6,%xmm7 + DB 102,65,15,254,215 ; paddd %xmm15,%xmm2 + DB 102,15,112,234,78 ; pshufd $0x4e,%xmm2,%xmm5 + DB 102,73,15,126,232 ; movq %xmm5,%r8 + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 15,20,213 ; unpcklps %xmm5,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 102,15,20,214 ; unpcklpd %xmm6,%xmm2 + DB 102,69,15,254,205 ; paddd %xmm13,%xmm9 + DB 102,65,15,112,233,245 ; pshufd $0xf5,%xmm9,%xmm5 + DB 102,15,111,29,59,182,0,0 ; movdqa 0xb63b(%rip),%xmm3 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,68,15,244,203 ; pmuludq %xmm3,%xmm9 + DB 102,15,244,235 ; pmuludq %xmm3,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,69,15,112,249,232 ; pshufd $0xe8,%xmm9,%xmm15 + DB 102,68,15,98,253 ; punpckldq %xmm5,%xmm15 + DB 102,69,15,126,200 ; movd %xmm9,%r8d + DB 102,65,15,112,239,229 ; pshufd $0xe5,%xmm15,%xmm5 + DB 102,65,15,126,233 ; movd %xmm5,%r9d + DB 102,65,15,112,239,78 ; pshufd $0x4e,%xmm15,%xmm5 + DB 102,65,15,126,234 ; movd %xmm5,%r10d + DB 102,65,15,112,239,231 ; pshufd $0xe7,%xmm15,%xmm5 + DB 102,15,126,232 ; movd %xmm5,%eax + DB 243,68,15,16,12,131 ; movss (%rbx,%rax,4),%xmm9 + DB 243,66,15,16,44,147 ; movss (%rbx,%r10,4),%xmm5 + DB 65,15,20,233 ; unpcklps %xmm9,%xmm5 + DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 + DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 + DB 69,15,20,233 ; unpcklps %xmm9,%xmm13 + DB 102,68,15,20,237 ; unpcklpd %xmm5,%xmm13 + DB 102,65,15,111,239 ; movdqa %xmm15,%xmm5 + DB 102,15,250,45,219,181,0,0 ; psubd 0xb5db(%rip),%xmm5 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,221,78 ; pshufd $0x4e,%xmm5,%xmm3 DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,44,131 ; movss (%r11,%rax,4),%xmm13 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 - DB 68,15,20,235 ; unpcklps %xmm3,%xmm13 - DB 68,15,20,234 ; unpcklps %xmm2,%xmm13 - DB 69,15,92,217 ; subps %xmm9,%xmm11 - DB 69,15,89,220 ; mulps %xmm12,%xmm11 - DB 69,15,88,217 ; addps %xmm9,%xmm11 - DB 15,92,198 ; subps %xmm6,%xmm0 - DB 65,15,89,196 ; mulps %xmm12,%xmm0 - DB 15,88,198 ; addps %xmm6,%xmm0 - DB 68,15,92,239 ; subps %xmm7,%xmm13 - DB 69,15,89,236 ; mulps %xmm12,%xmm13 - DB 68,15,88,239 ; addps %xmm7,%xmm13 - DB 69,15,92,216 ; subps %xmm8,%xmm11 - DB 15,40,92,36,16 ; movaps 0x10(%rsp),%xmm3 - DB 68,15,89,219 ; mulps %xmm3,%xmm11 - DB 69,15,88,216 ; addps %xmm8,%xmm11 - DB 15,92,197 ; subps %xmm5,%xmm0 - DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,197 ; addps %xmm5,%xmm0 - DB 15,40,208 ; movaps %xmm0,%xmm2 - DB 68,15,92,233 ; subps %xmm1,%xmm13 + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 15,20,243 ; unpcklps %xmm3,%xmm6 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 102,15,20,245 ; unpcklpd %xmm5,%xmm6 + DB 102,68,15,111,13,5,175,0,0 ; movdqa 0xaf05(%rip),%xmm9 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,69,15,254,249 ; paddd %xmm9,%xmm15 + DB 102,65,15,112,223,78 ; pshufd $0x4e,%xmm15,%xmm3 + DB 102,73,15,126,216 ; movq %xmm3,%r8 + DB 102,76,15,126,248 ; movq %xmm15,%rax + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,70,15,16,60,139 ; movss (%rbx,%r9,4),%xmm15 + DB 68,15,20,251 ; unpcklps %xmm3,%xmm15 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 102,68,15,20,253 ; unpcklpd %xmm5,%xmm15 + DB 69,15,92,232 ; subps %xmm8,%xmm13 + DB 15,40,156,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm3 DB 68,15,89,235 ; mulps %xmm3,%xmm13 + DB 69,15,88,232 ; addps %xmm8,%xmm13 + DB 15,92,247 ; subps %xmm7,%xmm6 + DB 15,89,243 ; mulps %xmm3,%xmm6 + DB 15,88,247 ; addps %xmm7,%xmm6 + DB 68,15,92,250 ; subps %xmm2,%xmm15 + DB 68,15,89,251 ; mulps %xmm3,%xmm15 + DB 68,15,40,195 ; movaps %xmm3,%xmm8 + DB 68,15,88,250 ; addps %xmm2,%xmm15 + DB 68,15,92,233 ; subps %xmm1,%xmm13 + DB 15,40,84,36,16 ; movaps 0x10(%rsp),%xmm2 + DB 68,15,89,234 ; mulps %xmm2,%xmm13 DB 68,15,88,233 ; addps %xmm1,%xmm13 - DB 15,91,68,36,96 ; cvtdq2ps 0x60(%rsp),%xmm0 - DB 15,40,140,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm1 - DB 15,92,200 ; subps %xmm0,%xmm1 - DB 15,41,140,36,160,0,0,0 ; movaps %xmm1,0xa0(%rsp) - DB 69,15,92,223 ; subps %xmm15,%xmm11 - DB 68,15,89,217 ; mulps %xmm1,%xmm11 - DB 69,15,88,223 ; addps %xmm15,%xmm11 - DB 68,15,41,92,36,96 ; movaps %xmm11,0x60(%rsp) - DB 65,15,92,214 ; subps %xmm14,%xmm2 - DB 15,89,209 ; mulps %xmm1,%xmm2 - DB 65,15,88,214 ; addps %xmm14,%xmm2 - DB 15,41,84,36,80 ; movaps %xmm2,0x50(%rsp) - DB 69,15,92,234 ; subps %xmm10,%xmm13 - DB 68,15,89,233 ; mulps %xmm1,%xmm13 - DB 69,15,88,234 ; addps %xmm10,%xmm13 + DB 65,15,40,205 ; movaps %xmm13,%xmm1 + DB 15,92,240 ; subps %xmm0,%xmm6 + DB 15,89,242 ; mulps %xmm2,%xmm6 + DB 15,88,240 ; addps %xmm0,%xmm6 + DB 68,15,92,252 ; subps %xmm4,%xmm15 + DB 68,15,89,250 ; mulps %xmm2,%xmm15 + DB 68,15,88,252 ; addps %xmm4,%xmm15 + DB 15,91,68,36,80 ; cvtdq2ps 0x50(%rsp),%xmm0 + DB 15,40,84,36,112 ; movaps 0x70(%rsp),%xmm2 + DB 15,92,208 ; subps %xmm0,%xmm2 + DB 15,41,84,36,112 ; movaps %xmm2,0x70(%rsp) + DB 65,15,92,203 ; subps %xmm11,%xmm1 + DB 15,89,202 ; mulps %xmm2,%xmm1 + DB 65,15,88,203 ; addps %xmm11,%xmm1 + DB 15,41,76,36,80 ; movaps %xmm1,0x50(%rsp) + DB 65,15,92,242 ; subps %xmm10,%xmm6 + DB 15,89,242 ; mulps %xmm2,%xmm6 + DB 65,15,88,242 ; addps %xmm10,%xmm6 + DB 15,41,116,36,64 ; movaps %xmm6,0x40(%rsp) + DB 69,15,92,252 ; subps %xmm12,%xmm15 + DB 68,15,89,250 ; mulps %xmm2,%xmm15 + DB 69,15,88,252 ; addps %xmm12,%xmm15 DB 243,15,91,68,36,48 ; cvttps2dq 0x30(%rsp),%xmm0 DB 102,15,127,68,36,48 ; movdqa %xmm0,0x30(%rsp) - DB 102,15,111,140,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm1 - DB 102,15,254,200 ; paddd %xmm0,%xmm1 - DB 102,15,111,132,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm0 - DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 - DB 102,68,15,111,84,36,32 ; movdqa 0x20(%rsp),%xmm10 - DB 102,65,15,111,202 ; movdqa %xmm10,%xmm1 - DB 102,15,254,200 ; paddd %xmm0,%xmm1 - DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 - DB 102,15,111,29,159,183,0,0 ; movdqa 0xb79f(%rip),%xmm3 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 + DB 102,68,15,254,240 ; paddd %xmm0,%xmm14 + DB 102,15,111,156,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm3 + DB 102,65,15,254,222 ; paddd %xmm14,%xmm3 + DB 102,68,15,111,92,36,32 ; movdqa 0x20(%rsp),%xmm11 + DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 + DB 102,15,254,195 ; paddd %xmm3,%xmm0 + DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 + DB 102,15,111,21,111,180,0,0 ; movdqa 0xb46f(%rip),%xmm2 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 + DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 - DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 - DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 - DB 102,65,15,126,209 ; movd %xmm2,%r9d - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,65,15,126,208 ; movd %xmm2,%r8d - DB 102,15,112,209,231 ; pshufd $0xe7,%xmm1,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,28,139 ; movss (%r11,%r9,4),%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,212 ; unpcklps %xmm4,%xmm2 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,15,111,61,166,176,0,0 ; movdqa 0xb0a6(%rip),%xmm7 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,223 ; paddd %xmm7,%xmm3 - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 + DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 + DB 102,15,98,209 ; punpckldq %xmm1,%xmm2 + DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 + DB 102,65,15,126,193 ; movd %xmm0,%r9d + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,65,15,126,194 ; movd %xmm0,%r10d + DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,36,147 ; movss (%rbx,%r10,4),%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 102,15,20,204 ; unpcklpd %xmm4,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,118,255 ; pcmpeqd %xmm7,%xmm7 + DB 102,15,250,199 ; psubd %xmm7,%xmm0 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 - DB 15,20,220 ; unpcklps %xmm4,%xmm3 - DB 102,68,15,111,13,117,176,0,0 ; movdqa 0xb075(%rip),%xmm9 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,65,15,254,201 ; paddd %xmm9,%xmm1 - DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 15,20,196 ; unpcklps %xmm4,%xmm0 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 + DB 102,65,15,254,209 ; paddd %xmm9,%xmm2 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 15,20,225 ; unpcklps %xmm1,%xmm4 - DB 102,68,15,111,188,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm15 - DB 102,65,15,254,199 ; paddd %xmm15,%xmm0 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,65,15,244,198 ; pmuludq %xmm14,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,65,15,244,206 ; pmuludq %xmm14,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,65,15,126,201 ; movd %xmm1,%r9d - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,65,15,126,200 ; movd %xmm1,%r8d - DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,44,139 ; movss (%r11,%r9,4),%xmm5 - DB 15,20,233 ; unpcklps %xmm1,%xmm5 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,52,131 ; movss (%r11,%r8,4),%xmm6 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,254,239 ; paddd %xmm7,%xmm5 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 102,15,20,213 ; unpcklpd %xmm5,%xmm2 + DB 102,68,15,111,84,36,96 ; movdqa 0x60(%rsp),%xmm10 + DB 102,65,15,254,218 ; paddd %xmm10,%xmm3 + DB 102,15,112,227,245 ; pshufd $0xf5,%xmm3,%xmm4 + DB 102,65,15,244,221 ; pmuludq %xmm13,%xmm3 + DB 102,65,15,244,229 ; pmuludq %xmm13,%xmm4 + DB 102,15,112,236,232 ; pshufd $0xe8,%xmm4,%xmm5 + DB 102,15,112,227,232 ; pshufd $0xe8,%xmm3,%xmm4 + DB 102,15,98,229 ; punpckldq %xmm5,%xmm4 + DB 102,65,15,126,216 ; movd %xmm3,%r8d + DB 102,15,112,220,229 ; pshufd $0xe5,%xmm4,%xmm3 + DB 102,65,15,126,217 ; movd %xmm3,%r9d + DB 102,15,112,220,78 ; pshufd $0x4e,%xmm4,%xmm3 + DB 102,65,15,126,218 ; movd %xmm3,%r10d + DB 102,15,112,220,231 ; pshufd $0xe7,%xmm4,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,66,15,16,44,147 ; movss (%rbx,%r10,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 15,20,222 ; unpcklps %xmm6,%xmm3 + DB 102,15,20,221 ; unpcklpd %xmm5,%xmm3 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,15,250,239 ; psubd %xmm7,%xmm5 + DB 102,69,15,118,228 ; pcmpeqd %xmm12,%xmm12 DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 DB 102,73,15,126,240 ; movq %xmm6,%r8 - DB 102,73,15,126,233 ; movq %xmm5,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,239 ; unpcklps %xmm7,%xmm5 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 102,65,15,254,193 ; paddd %xmm9,%xmm0 - DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,52,131 ; movss (%rbx,%r8,4),%xmm6 + DB 243,15,16,60,131 ; movss (%rbx,%rax,4),%xmm7 + DB 15,20,254 ; unpcklps %xmm6,%xmm7 + DB 102,15,20,239 ; unpcklpd %xmm7,%xmm5 + DB 102,65,15,254,225 ; paddd %xmm9,%xmm4 + DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 DB 102,73,15,126,240 ; movq %xmm6,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,224 ; movq %xmm4,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,4,147 ; movss (%r11,%r10,4),%xmm0 - DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,198 ; unpcklps %xmm6,%xmm0 - DB 15,92,202 ; subps %xmm2,%xmm1 - DB 69,15,40,196 ; movaps %xmm12,%xmm8 - DB 65,15,89,200 ; mulps %xmm8,%xmm1 - DB 15,88,202 ; addps %xmm2,%xmm1 - DB 15,92,235 ; subps %xmm3,%xmm5 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,60,139 ; movss (%rbx,%r9,4),%xmm7 + DB 15,20,252 ; unpcklps %xmm4,%xmm7 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 102,15,20,254 ; unpcklpd %xmm6,%xmm7 + DB 15,92,217 ; subps %xmm1,%xmm3 + DB 65,15,89,216 ; mulps %xmm8,%xmm3 + DB 15,88,217 ; addps %xmm1,%xmm3 + DB 15,92,232 ; subps %xmm0,%xmm5 DB 65,15,89,232 ; mulps %xmm8,%xmm5 - DB 15,88,235 ; addps %xmm3,%xmm5 - DB 15,92,196 ; subps %xmm4,%xmm0 - DB 65,15,89,192 ; mulps %xmm8,%xmm0 - DB 15,88,196 ; addps %xmm4,%xmm0 - DB 102,65,15,111,219 ; movdqa %xmm11,%xmm3 - DB 102,15,254,28,36 ; paddd (%rsp),%xmm3 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 - DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 - DB 102,15,112,218,245 ; pshufd $0xf5,%xmm2,%xmm3 - DB 102,65,15,244,214 ; pmuludq %xmm14,%xmm2 - DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 - DB 102,65,15,244,222 ; pmuludq %xmm14,%xmm3 - DB 102,69,15,111,230 ; movdqa %xmm14,%xmm12 - DB 102,15,112,211,232 ; pshufd $0xe8,%xmm3,%xmm2 - DB 102,15,98,226 ; punpckldq %xmm2,%xmm4 - DB 102,15,112,212,229 ; pshufd $0xe5,%xmm4,%xmm2 - DB 102,65,15,126,209 ; movd %xmm2,%r9d - DB 102,15,112,212,78 ; pshufd $0x4e,%xmm4,%xmm2 - DB 102,65,15,126,208 ; movd %xmm2,%r8d - DB 102,15,112,212,231 ; pshufd $0xe7,%xmm4,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,28,139 ; movss (%r11,%r9,4),%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,69,15,16,52,131 ; movss (%r11,%rax,4),%xmm14 - DB 243,67,15,16,52,131 ; movss (%r11,%r8,4),%xmm6 - DB 68,15,20,246 ; unpcklps %xmm6,%xmm14 - DB 68,15,20,243 ; unpcklps %xmm3,%xmm14 - DB 102,15,111,220 ; movdqa %xmm4,%xmm3 - DB 102,68,15,111,21,92,174,0,0 ; movdqa 0xae5c(%rip),%xmm10 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,218 ; paddd %xmm10,%xmm3 - DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 - DB 102,73,15,126,240 ; movq %xmm6,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 15,88,232 ; addps %xmm0,%xmm5 + DB 15,92,250 ; subps %xmm2,%xmm7 + DB 65,15,89,248 ; mulps %xmm8,%xmm7 + DB 15,88,250 ; addps %xmm2,%xmm7 + DB 102,68,15,254,180,36,144,0,0,0 ; paddd 0x90(%rsp),%xmm14 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,65,15,254,203 ; paddd %xmm11,%xmm1 + DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 + DB 102,65,15,244,205 ; pmuludq %xmm13,%xmm1 + DB 102,65,15,244,197 ; pmuludq %xmm13,%xmm0 + DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 + DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 + DB 102,15,98,194 ; punpckldq %xmm2,%xmm0 + DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,65,15,126,201 ; movd %xmm1,%r9d + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,65,15,126,202 ; movd %xmm1,%r10d + DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,126,200 ; movd %xmm1,%eax + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,36,147 ; movss (%rbx,%r10,4),%xmm4 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,70,15,16,28,131 ; movss (%rbx,%r8,4),%xmm11 + DB 68,15,20,217 ; unpcklps %xmm1,%xmm11 + DB 102,68,15,20,220 ; unpcklpd %xmm4,%xmm11 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,65,15,250,204 ; psubd %xmm12,%xmm1 + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,28,147 ; movss (%r11,%r10,4),%xmm3 - DB 15,20,243 ; unpcklps %xmm3,%xmm6 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,223 ; unpcklps %xmm7,%xmm3 - DB 15,20,222 ; unpcklps %xmm6,%xmm3 - DB 102,65,15,254,225 ; paddd %xmm9,%xmm4 - DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 - DB 102,73,15,126,240 ; movq %xmm6,%r8 - DB 102,73,15,126,225 ; movq %xmm4,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 15,20,204 ; unpcklps %xmm4,%xmm1 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 + DB 102,65,15,254,193 ; paddd %xmm9,%xmm0 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,52,131 ; movss (%r11,%rax,4),%xmm6 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,4,139 ; movss (%rbx,%r9,4),%xmm0 + DB 15,20,196 ; unpcklps %xmm4,%xmm0 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,231 ; unpcklps %xmm7,%xmm4 - DB 15,20,230 ; unpcklps %xmm6,%xmm4 - DB 102,65,15,111,211 ; movdqa %xmm11,%xmm2 - DB 102,65,15,254,215 ; paddd %xmm15,%xmm2 - DB 102,15,112,250,245 ; pshufd $0xf5,%xmm2,%xmm7 - DB 102,65,15,244,212 ; pmuludq %xmm12,%xmm2 - DB 102,15,112,242,232 ; pshufd $0xe8,%xmm2,%xmm6 - DB 102,65,15,244,252 ; pmuludq %xmm12,%xmm7 - DB 102,15,112,255,232 ; pshufd $0xe8,%xmm7,%xmm7 - DB 102,15,98,247 ; punpckldq %xmm7,%xmm6 - DB 102,15,112,254,229 ; pshufd $0xe5,%xmm6,%xmm7 - DB 102,65,15,126,249 ; movd %xmm7,%r9d - DB 102,15,112,254,78 ; pshufd $0x4e,%xmm6,%xmm7 - DB 102,65,15,126,248 ; movd %xmm7,%r8d - DB 102,15,112,254,231 ; pshufd $0xe7,%xmm6,%xmm7 - DB 102,15,126,248 ; movd %xmm7,%eax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,20,139 ; movss (%r11,%r9,4),%xmm2 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 + DB 102,15,20,198 ; unpcklpd %xmm6,%xmm0 + DB 102,69,15,254,242 ; paddd %xmm10,%xmm14 + DB 102,69,15,111,226 ; movdqa %xmm10,%xmm12 + DB 102,65,15,112,230,245 ; pshufd $0xf5,%xmm14,%xmm4 + DB 102,69,15,244,245 ; pmuludq %xmm13,%xmm14 + DB 102,65,15,244,229 ; pmuludq %xmm13,%xmm4 + DB 102,15,112,244,232 ; pshufd $0xe8,%xmm4,%xmm6 + DB 102,65,15,112,230,232 ; pshufd $0xe8,%xmm14,%xmm4 + DB 102,15,98,230 ; punpckldq %xmm6,%xmm4 + DB 102,69,15,126,240 ; movd %xmm14,%r8d + DB 102,15,112,244,229 ; pshufd $0xe5,%xmm4,%xmm6 + DB 102,65,15,126,241 ; movd %xmm6,%r9d + DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 + DB 102,65,15,126,242 ; movd %xmm6,%r10d + DB 102,15,112,244,231 ; pshufd $0xe7,%xmm4,%xmm6 DB 102,15,126,240 ; movd %xmm6,%eax - DB 243,69,15,16,36,131 ; movss (%r11,%rax,4),%xmm12 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 68,15,20,231 ; unpcklps %xmm7,%xmm12 - DB 68,15,20,226 ; unpcklps %xmm2,%xmm12 - DB 102,15,111,214 ; movdqa %xmm6,%xmm2 - DB 102,65,15,254,210 ; paddd %xmm10,%xmm2 - DB 102,15,112,250,78 ; pshufd $0x4e,%xmm2,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 + DB 15,20,214 ; unpcklps %xmm6,%xmm2 + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 243,70,15,16,44,131 ; movss (%rbx,%r8,4),%xmm13 + DB 68,15,20,238 ; unpcklps %xmm6,%xmm13 + DB 102,68,15,20,234 ; unpcklpd %xmm2,%xmm13 + DB 102,15,111,212 ; movdqa %xmm4,%xmm2 + DB 102,15,250,21,8,177,0,0 ; psubd 0xb108(%rip),%xmm2 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,60,147 ; movss (%r11,%r10,4),%xmm7 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,28,131 ; movss (%r11,%rax,4),%xmm11 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 68,15,20,223 ; unpcklps %xmm7,%xmm11 - DB 68,15,20,218 ; unpcklps %xmm2,%xmm11 - DB 102,65,15,254,241 ; paddd %xmm9,%xmm6 - DB 102,15,112,214,78 ; pshufd $0x4e,%xmm6,%xmm2 + DB 243,15,16,20,131 ; movss (%rbx,%rax,4),%xmm2 + DB 243,70,15,16,20,139 ; movss (%rbx,%r9,4),%xmm10 + DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 102,68,15,20,214 ; unpcklpd %xmm6,%xmm10 + DB 102,65,15,254,225 ; paddd %xmm9,%xmm4 + DB 102,15,112,212,78 ; pshufd $0x4e,%xmm4,%xmm2 DB 102,73,15,126,208 ; movq %xmm2,%r8 - DB 102,73,15,126,241 ; movq %xmm6,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,224 ; movq %xmm4,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,52,147 ; movss (%r11,%r10,4),%xmm6 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,20,131 ; movss (%r11,%rax,4),%xmm10 - DB 243,67,15,16,52,131 ; movss (%r11,%r8,4),%xmm6 - DB 68,15,20,214 ; unpcklps %xmm6,%xmm10 - DB 68,15,20,210 ; unpcklps %xmm2,%xmm10 - DB 69,15,92,230 ; subps %xmm14,%xmm12 - DB 69,15,89,224 ; mulps %xmm8,%xmm12 - DB 69,15,88,230 ; addps %xmm14,%xmm12 - DB 68,15,92,219 ; subps %xmm3,%xmm11 - DB 69,15,89,216 ; mulps %xmm8,%xmm11 - DB 68,15,88,219 ; addps %xmm3,%xmm11 - DB 68,15,92,212 ; subps %xmm4,%xmm10 + DB 243,15,16,20,131 ; movss (%rbx,%rax,4),%xmm2 + DB 243,70,15,16,12,139 ; movss (%rbx,%r9,4),%xmm9 + DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,20,131 ; movss (%rbx,%r8,4),%xmm2 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 102,68,15,20,206 ; unpcklpd %xmm6,%xmm9 + DB 69,15,92,235 ; subps %xmm11,%xmm13 + DB 69,15,89,232 ; mulps %xmm8,%xmm13 + DB 69,15,88,235 ; addps %xmm11,%xmm13 + DB 68,15,92,209 ; subps %xmm1,%xmm10 DB 69,15,89,208 ; mulps %xmm8,%xmm10 - DB 68,15,88,212 ; addps %xmm4,%xmm10 - DB 68,15,92,225 ; subps %xmm1,%xmm12 - DB 15,40,84,36,16 ; movaps 0x10(%rsp),%xmm2 - DB 68,15,89,226 ; mulps %xmm2,%xmm12 - DB 68,15,88,225 ; addps %xmm1,%xmm12 - DB 68,15,92,221 ; subps %xmm5,%xmm11 - DB 68,15,89,218 ; mulps %xmm2,%xmm11 - DB 68,15,88,221 ; addps %xmm5,%xmm11 - DB 68,15,92,208 ; subps %xmm0,%xmm10 - DB 68,15,89,210 ; mulps %xmm2,%xmm10 - DB 68,15,88,208 ; addps %xmm0,%xmm10 - DB 102,15,111,68,36,64 ; movdqa 0x40(%rsp),%xmm0 + DB 68,15,88,209 ; addps %xmm1,%xmm10 + DB 68,15,92,200 ; subps %xmm0,%xmm9 + DB 69,15,89,200 ; mulps %xmm8,%xmm9 + DB 69,15,40,216 ; movaps %xmm8,%xmm11 + DB 68,15,88,200 ; addps %xmm0,%xmm9 + DB 68,15,92,235 ; subps %xmm3,%xmm13 + DB 15,40,68,36,16 ; movaps 0x10(%rsp),%xmm0 + DB 68,15,89,232 ; mulps %xmm0,%xmm13 + DB 68,15,88,235 ; addps %xmm3,%xmm13 + DB 68,15,92,213 ; subps %xmm5,%xmm10 + DB 68,15,89,208 ; mulps %xmm0,%xmm10 + DB 68,15,88,213 ; addps %xmm5,%xmm10 + DB 68,15,92,207 ; subps %xmm7,%xmm9 + DB 68,15,89,200 ; mulps %xmm0,%xmm9 + DB 68,15,88,207 ; addps %xmm7,%xmm9 + DB 102,15,111,4,36 ; movdqa (%rsp),%xmm0 DB 102,15,254,68,36,48 ; paddd 0x30(%rsp),%xmm0 - DB 102,15,111,140,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm1 + DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) + DB 102,15,111,140,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm1 DB 102,15,254,200 ; paddd %xmm0,%xmm1 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,111,116,36,32 ; movdqa 0x20(%rsp),%xmm6 - DB 102,15,254,198 ; paddd %xmm6,%xmm0 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,68,15,111,116,36,32 ; movdqa 0x20(%rsp),%xmm14 + DB 102,65,15,254,198 ; paddd %xmm14,%xmm0 DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,29,204,178,0,0 ; movdqa 0xb2cc(%rip),%xmm3 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 + DB 102,68,15,111,5,229,175,0,0 ; movdqa 0xafe5(%rip),%xmm8 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,65,15,244,192 ; pmuludq %xmm8,%xmm0 + DB 102,65,15,244,200 ; pmuludq %xmm8,%xmm1 + DB 102,65,15,111,240 ; movdqa %xmm8,%xmm6 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 - DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 - DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 - DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 + DB 102,15,98,209 ; punpckldq %xmm1,%xmm2 + DB 102,65,15,126,192 ; movd %xmm0,%r8d DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 DB 102,65,15,126,193 ; movd %xmm0,%r9d DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 - DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,65,15,126,194 ; movd %xmm0,%r10d DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,12,139 ; movss (%r11,%r9,4),%xmm1 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,28,131 ; movss (%r11,%r8,4),%xmm3 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,68,15,111,5,210,171,0,0 ; movdqa 0xabd2(%rip),%xmm8 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,65,15,254,200 ; paddd %xmm8,%xmm1 + DB 102,15,118,255 ; pcmpeqd %xmm7,%xmm7 + DB 102,15,250,207 ; psubd %xmm7,%xmm1 DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 DB 15,20,203 ; unpcklps %xmm3,%xmm1 - DB 102,65,15,254,209 ; paddd %xmm9,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 102,15,20,205 ; unpcklpd %xmm5,%xmm1 + DB 102,68,15,111,5,179,168,0,0 ; movdqa 0xa8b3(%rip),%xmm8 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,65,15,254,208 ; paddd %xmm8,%xmm2 DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 DB 102,73,15,126,216 ; movq %xmm3,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 102,15,111,221 ; movdqa %xmm5,%xmm3 - DB 102,65,15,254,223 ; paddd %xmm15,%xmm3 - DB 102,15,112,227,245 ; pshufd $0xf5,%xmm3,%xmm4 - DB 102,65,15,244,222 ; pmuludq %xmm14,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,65,15,244,230 ; pmuludq %xmm14,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 - DB 102,15,98,220 ; punpckldq %xmm4,%xmm3 - DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 - DB 102,65,15,126,225 ; movd %xmm4,%r9d - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 102,15,20,213 ; unpcklpd %xmm5,%xmm2 + DB 102,65,15,254,228 ; paddd %xmm12,%xmm4 + DB 102,15,112,220,245 ; pshufd $0xf5,%xmm4,%xmm3 + DB 102,15,244,230 ; pmuludq %xmm6,%xmm4 + DB 102,15,244,222 ; pmuludq %xmm6,%xmm3 + DB 102,15,112,235,232 ; pshufd $0xe8,%xmm3,%xmm5 + DB 102,15,112,220,232 ; pshufd $0xe8,%xmm4,%xmm3 + DB 102,15,98,221 ; punpckldq %xmm5,%xmm3 DB 102,65,15,126,224 ; movd %xmm4,%r8d - DB 102,15,112,227,231 ; pshufd $0xe7,%xmm3,%xmm4 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,139 ; movss (%r11,%r9,4),%xmm5 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,69,15,16,52,131 ; movss (%r11,%rax,4),%xmm14 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 68,15,20,244 ; unpcklps %xmm4,%xmm14 - DB 68,15,20,245 ; unpcklps %xmm5,%xmm14 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,65,15,254,224 ; paddd %xmm8,%xmm4 - DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 - DB 102,73,15,126,232 ; movq %xmm5,%r8 - DB 102,73,15,126,225 ; movq %xmm4,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,15,112,235,229 ; pshufd $0xe5,%xmm3,%xmm5 + DB 102,65,15,126,233 ; movd %xmm5,%r9d + DB 102,15,112,235,78 ; pshufd $0x4e,%xmm3,%xmm5 + DB 102,65,15,126,234 ; movd %xmm5,%r10d + DB 102,15,112,235,231 ; pshufd $0xe7,%xmm3,%xmm5 + DB 102,15,126,232 ; movd %xmm5,%eax + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,52,147 ; movss (%rbx,%r10,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 + DB 243,70,15,16,36,131 ; movss (%rbx,%r8,4),%xmm12 + DB 68,15,20,229 ; unpcklps %xmm5,%xmm12 + DB 102,68,15,20,230 ; unpcklpd %xmm6,%xmm12 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,15,250,239 ; psubd %xmm7,%xmm5 + DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,232 ; movq %xmm5,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,44,147 ; movss (%r11,%r10,4),%xmm5 - DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,69,15,16,4,131 ; movss (%r11,%rax,4),%xmm8 - DB 243,67,15,16,44,131 ; movss (%r11,%r8,4),%xmm5 - DB 68,15,20,197 ; unpcklps %xmm5,%xmm8 - DB 68,15,20,196 ; unpcklps %xmm4,%xmm8 - DB 102,65,15,254,217 ; paddd %xmm9,%xmm3 - DB 102,69,15,111,249 ; movdqa %xmm9,%xmm15 - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 - DB 102,73,15,126,224 ; movq %xmm4,%r8 - DB 102,73,15,126,217 ; movq %xmm3,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,44,131 ; movss (%rbx,%rax,4),%xmm5 + DB 243,66,15,16,60,139 ; movss (%rbx,%r9,4),%xmm7 + DB 15,20,253 ; unpcklps %xmm5,%xmm7 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,44,131 ; movss (%rbx,%r8,4),%xmm5 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 102,15,20,254 ; unpcklpd %xmm6,%xmm7 + DB 102,65,15,254,216 ; paddd %xmm8,%xmm3 + DB 102,15,112,235,78 ; pshufd $0x4e,%xmm3,%xmm5 + DB 102,73,15,126,232 ; movq %xmm5,%r8 + DB 102,72,15,126,216 ; movq %xmm3,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,36,147 ; movss (%r11,%r10,4),%xmm4 - DB 15,20,220 ; unpcklps %xmm4,%xmm3 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,44,131 ; movss (%r11,%rax,4),%xmm5 - DB 243,67,15,16,36,131 ; movss (%r11,%r8,4),%xmm4 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 + DB 243,15,16,28,131 ; movss (%rbx,%rax,4),%xmm3 + DB 243,66,15,16,44,139 ; movss (%rbx,%r9,4),%xmm5 DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 68,15,92,240 ; subps %xmm0,%xmm14 - DB 15,40,92,36,112 ; movaps 0x70(%rsp),%xmm3 - DB 68,15,89,243 ; mulps %xmm3,%xmm14 - DB 68,15,88,240 ; addps %xmm0,%xmm14 - DB 68,15,92,193 ; subps %xmm1,%xmm8 - DB 68,15,89,195 ; mulps %xmm3,%xmm8 - DB 68,15,88,193 ; addps %xmm1,%xmm8 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,28,131 ; movss (%rbx,%r8,4),%xmm3 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,243 ; unpcklps %xmm3,%xmm6 + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 + DB 68,15,92,224 ; subps %xmm0,%xmm12 + DB 69,15,89,227 ; mulps %xmm11,%xmm12 + DB 68,15,88,224 ; addps %xmm0,%xmm12 + DB 15,92,249 ; subps %xmm1,%xmm7 + DB 65,15,89,251 ; mulps %xmm11,%xmm7 + DB 15,88,249 ; addps %xmm1,%xmm7 DB 15,92,234 ; subps %xmm2,%xmm5 - DB 15,89,235 ; mulps %xmm3,%xmm5 + DB 65,15,89,235 ; mulps %xmm11,%xmm5 DB 15,88,234 ; addps %xmm2,%xmm5 - DB 102,15,254,60,36 ; paddd (%rsp),%xmm7 - DB 102,15,254,247 ; paddd %xmm7,%xmm6 - DB 102,15,112,206,245 ; pshufd $0xf5,%xmm6,%xmm1 - DB 102,15,111,21,141,176,0,0 ; movdqa 0xb08d(%rip),%xmm2 # 2db60 <_sk_overlay_sse2_8bit+0x1608> - DB 102,15,244,242 ; pmuludq %xmm2,%xmm6 - DB 102,15,112,198,232 ; pshufd $0xe8,%xmm6,%xmm0 - DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,111,4,36 ; movdqa (%rsp),%xmm0 + DB 102,15,254,132,36,144,0,0,0 ; paddd 0x90(%rsp),%xmm0 + DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 + DB 102,15,254,208 ; paddd %xmm0,%xmm2 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,37,189,173,0,0 ; movdqa 0xadbd(%rip),%xmm4 # 2da60 <_sk_overlay_sse2_8bit+0x15b1> + DB 102,15,244,212 ; pmuludq %xmm4,%xmm2 + DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 + DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,65,15,126,208 ; movd %xmm2,%r8d DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,65,15,126,201 ; movd %xmm1,%r9d DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,65,15,126,200 ; movd %xmm1,%r8d + DB 102,65,15,126,202 ; movd %xmm1,%r10d DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,20,139 ; movss (%r11,%r9,4),%xmm2 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,66,15,16,20,147 ; movss (%rbx,%r10,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,69,15,16,12,131 ; movss (%r11,%rax,4),%xmm9 - DB 243,67,15,16,12,131 ; movss (%r11,%r8,4),%xmm1 - DB 68,15,20,201 ; unpcklps %xmm1,%xmm9 - DB 68,15,20,202 ; unpcklps %xmm2,%xmm9 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 243,70,15,16,4,131 ; movss (%rbx,%r8,4),%xmm8 + DB 68,15,20,193 ; unpcklps %xmm1,%xmm8 + DB 102,68,15,20,194 ; unpcklpd %xmm2,%xmm8 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,254,13,147,169,0,0 ; paddd 0xa993(%rip),%xmm1 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> + DB 102,15,250,13,102,173,0,0 ; psubd 0xad66(%rip),%xmm1 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 DB 102,73,15,126,208 ; movq %xmm2,%r8 - DB 102,73,15,126,201 ; movq %xmm1,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,36,131 ; movss (%r11,%rax,4),%xmm4 - DB 243,67,15,16,20,131 ; movss (%r11,%r8,4),%xmm2 - DB 15,20,226 ; unpcklps %xmm2,%xmm4 - DB 15,20,225 ; unpcklps %xmm1,%xmm4 - DB 102,65,15,254,199 ; paddd %xmm15,%xmm0 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 243,70,15,16,52,139 ; movss (%rbx,%r9,4),%xmm14 + DB 68,15,20,241 ; unpcklps %xmm1,%xmm14 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,12,131 ; movss (%rbx,%r8,4),%xmm1 + DB 243,15,16,20,131 ; movss (%rbx,%rax,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,68,15,20,242 ; unpcklpd %xmm2,%xmm14 + DB 102,15,254,5,143,166,0,0 ; paddd 0xa68f(%rip),%xmm0 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,73,15,126,200 ; movq %xmm1,%r8 - DB 102,73,15,126,193 ; movq %xmm0,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,28,131 ; movss (%r11,%rax,4),%xmm3 - DB 243,67,15,16,12,131 ; movss (%r11,%r8,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,28,139 ; movss (%rbx,%r9,4),%xmm3 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 102,68,15,111,255 ; movdqa %xmm7,%xmm15 - DB 102,68,15,254,188,36,128,0,0,0 ; paddd 0x80(%rsp),%xmm15 - DB 102,65,15,112,199,245 ; pshufd $0xf5,%xmm15,%xmm0 - DB 102,68,15,244,254 ; pmuludq %xmm6,%xmm15 - DB 102,15,244,198 ; pmuludq %xmm6,%xmm0 - DB 102,65,15,112,215,232 ; pshufd $0xe8,%xmm15,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 243,15,16,12,131 ; movss (%rbx,%rax,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 102,15,20,217 ; unpcklpd %xmm1,%xmm3 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 + DB 102,15,254,76,36,96 ; paddd 0x60(%rsp),%xmm1 + DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 + DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 + DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 + DB 102,65,15,126,200 ; movd %xmm1,%r8d DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 DB 102,65,15,126,193 ; movd %xmm0,%r9d DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 - DB 102,65,15,126,192 ; movd %xmm0,%r8d + DB 102,65,15,126,194 ; movd %xmm0,%r10d DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,12,139 ; movss (%r11,%r9,4),%xmm1 + DB 243,15,16,4,131 ; movss (%rbx,%rax,4),%xmm0 + DB 243,66,15,16,12,147 ; movss (%rbx,%r10,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,65,15,16,4,131 ; movss (%r11,%rax,4),%xmm0 - DB 243,67,15,16,60,131 ; movss (%r11,%r8,4),%xmm7 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 102,15,111,53,143,168,0,0 ; movdqa 0xa88f(%rip),%xmm6 # 2d4d0 <_sk_overlay_sse2_8bit+0xf78> - DB 102,15,254,242 ; paddd %xmm2,%xmm6 - DB 102,15,112,206,78 ; pshufd $0x4e,%xmm6,%xmm1 - DB 102,73,15,126,200 ; movq %xmm1,%r8 - DB 102,73,15,126,241 ; movq %xmm6,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,66,15,16,52,139 ; movss (%rbx,%r9,4),%xmm6 + DB 243,66,15,16,4,131 ; movss (%rbx,%r8,4),%xmm0 + DB 15,20,198 ; unpcklps %xmm6,%xmm0 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,250,13,110,172,0,0 ; psubd 0xac6e(%rip),%xmm1 # 2da70 <_sk_overlay_sse2_8bit+0x15c1> + DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 + DB 102,73,15,126,240 ; movq %xmm6,%r8 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,12,147 ; movss (%r11,%r10,4),%xmm1 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,12,131 ; movss (%r11,%rax,4),%xmm1 - DB 243,71,15,16,60,131 ; movss (%r11,%r8,4),%xmm15 - DB 65,15,20,207 ; unpcklps %xmm15,%xmm1 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 102,15,254,21,94,168,0,0 ; paddd 0xa85e(%rip),%xmm2 # 2d4f0 <_sk_overlay_sse2_8bit+0xf98> - DB 102,15,112,250,78 ; pshufd $0x4e,%xmm2,%xmm7 - DB 102,73,15,126,248 ; movq %xmm7,%r8 - DB 102,73,15,126,209 ; movq %xmm2,%r9 - DB 77,137,194 ; mov %r8,%r10 - DB 73,193,234,32 ; shr $0x20,%r10 - DB 76,137,200 ; mov %r9,%rax + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 243,66,15,16,12,139 ; movss (%rbx,%r9,4),%xmm1 + DB 15,20,206 ; unpcklps %xmm6,%xmm1 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,52,131 ; movss (%rbx,%r8,4),%xmm6 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 15,20,230 ; unpcklps %xmm6,%xmm4 + DB 102,15,20,204 ; unpcklpd %xmm4,%xmm1 + DB 102,15,254,21,153,165,0,0 ; paddd 0xa599(%rip),%xmm2 # 2d3e0 <_sk_overlay_sse2_8bit+0xf31> + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,73,15,126,224 ; movq %xmm4,%r8 + DB 102,72,15,126,208 ; movq %xmm2,%rax + DB 65,137,193 ; mov %eax,%r9d DB 72,193,232,32 ; shr $0x20,%rax - DB 243,65,15,16,60,131 ; movss (%r11,%rax,4),%xmm7 - DB 243,67,15,16,20,147 ; movss (%r11,%r10,4),%xmm2 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 - DB 68,137,200 ; mov %r9d,%eax - DB 69,137,192 ; mov %r8d,%r8d - DB 243,65,15,16,20,131 ; movss (%r11,%rax,4),%xmm2 - DB 243,67,15,16,52,131 ; movss (%r11,%r8,4),%xmm6 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 65,15,92,193 ; subps %xmm9,%xmm0 - DB 15,40,116,36,112 ; movaps 0x70(%rsp),%xmm6 - DB 15,89,198 ; mulps %xmm6,%xmm0 - DB 65,15,88,193 ; addps %xmm9,%xmm0 - DB 15,92,204 ; subps %xmm4,%xmm1 - DB 15,89,206 ; mulps %xmm6,%xmm1 - DB 15,88,204 ; addps %xmm4,%xmm1 + DB 243,15,16,36,131 ; movss (%rbx,%rax,4),%xmm4 + DB 243,66,15,16,20,139 ; movss (%rbx,%r9,4),%xmm2 + DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 68,137,192 ; mov %r8d,%eax + DB 73,193,232,32 ; shr $0x20,%r8 + DB 243,66,15,16,36,131 ; movss (%rbx,%r8,4),%xmm4 + DB 243,15,16,52,131 ; movss (%rbx,%rax,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 102,15,20,214 ; unpcklpd %xmm6,%xmm2 + DB 65,15,92,192 ; subps %xmm8,%xmm0 + DB 65,15,89,195 ; mulps %xmm11,%xmm0 + DB 65,15,88,192 ; addps %xmm8,%xmm0 + DB 65,15,92,206 ; subps %xmm14,%xmm1 + DB 65,15,89,203 ; mulps %xmm11,%xmm1 + DB 65,15,88,206 ; addps %xmm14,%xmm1 DB 15,92,211 ; subps %xmm3,%xmm2 - DB 15,89,214 ; mulps %xmm6,%xmm2 + DB 65,15,89,211 ; mulps %xmm11,%xmm2 DB 15,88,211 ; addps %xmm3,%xmm2 - DB 65,15,92,198 ; subps %xmm14,%xmm0 + DB 65,15,92,196 ; subps %xmm12,%xmm0 DB 15,40,92,36,16 ; movaps 0x10(%rsp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 65,15,88,198 ; addps %xmm14,%xmm0 - DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 65,15,88,196 ; addps %xmm12,%xmm0 + DB 15,92,207 ; subps %xmm7,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 65,15,88,200 ; addps %xmm8,%xmm1 + DB 15,88,207 ; addps %xmm7,%xmm1 DB 15,92,213 ; subps %xmm5,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 DB 15,88,213 ; addps %xmm5,%xmm2 - DB 65,15,92,196 ; subps %xmm12,%xmm0 - DB 15,40,156,36,160,0,0,0 ; movaps 0xa0(%rsp),%xmm3 + DB 65,15,92,197 ; subps %xmm13,%xmm0 + DB 15,40,92,36,112 ; movaps 0x70(%rsp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 65,15,88,196 ; addps %xmm12,%xmm0 - DB 65,15,92,203 ; subps %xmm11,%xmm1 + DB 65,15,88,197 ; addps %xmm13,%xmm0 + DB 65,15,92,202 ; subps %xmm10,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 65,15,88,203 ; addps %xmm11,%xmm1 - DB 65,15,92,210 ; subps %xmm10,%xmm2 + DB 65,15,88,202 ; addps %xmm10,%xmm1 + DB 65,15,92,209 ; subps %xmm9,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 - DB 65,15,88,210 ; addps %xmm10,%xmm2 + DB 65,15,88,209 ; addps %xmm9,%xmm2 DB 15,91,156,36,176,0,0,0 ; cvtdq2ps 0xb0(%rsp),%xmm3 - DB 15,40,164,36,224,0,0,0 ; movaps 0xe0(%rsp),%xmm4 + DB 15,40,164,36,208,0,0,0 ; movaps 0xd0(%rsp),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 - DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 + DB 15,40,92,36,80 ; movaps 0x50(%rsp),%xmm3 DB 15,92,195 ; subps %xmm3,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,88,195 ; addps %xmm3,%xmm0 - DB 15,40,92,36,80 ; movaps 0x50(%rsp),%xmm3 + DB 15,40,92,36,64 ; movaps 0x40(%rsp),%xmm3 DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,204 ; mulps %xmm4,%xmm1 DB 15,88,203 ; addps %xmm3,%xmm1 - DB 65,15,92,213 ; subps %xmm13,%xmm2 + DB 65,15,92,215 ; subps %xmm15,%xmm2 DB 15,89,212 ; mulps %xmm4,%xmm2 - DB 65,15,88,213 ; addps %xmm13,%xmm2 + DB 65,15,88,215 ; addps %xmm15,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,29,56,167,0,0 ; movaps 0xa738(%rip),%xmm3 # 2d4c0 <_sk_overlay_sse2_8bit+0xf68> - DB 15,40,164,36,240,0,0,0 ; movaps 0xf0(%rsp),%xmm4 - DB 15,40,172,36,0,1,0,0 ; movaps 0x100(%rsp),%xmm5 - DB 15,40,180,36,16,1,0,0 ; movaps 0x110(%rsp),%xmm6 - DB 15,40,188,36,32,1,0,0 ; movaps 0x120(%rsp),%xmm7 - DB 72,129,196,56,1,0,0 ; add $0x138,%rsp + DB 15,40,29,127,164,0,0 ; movaps 0xa47f(%rip),%xmm3 # 2d3b0 <_sk_overlay_sse2_8bit+0xf01> + DB 15,40,164,36,224,0,0,0 ; movaps 0xe0(%rsp),%xmm4 + DB 15,40,172,36,240,0,0,0 ; movaps 0xf0(%rsp),%xmm5 + DB 15,40,180,36,0,1,0,0 ; movaps 0x100(%rsp),%xmm6 + DB 15,40,188,36,16,1,0,0 ; movaps 0x110(%rsp),%xmm7 + DB 72,129,196,32,1,0,0 ; add $0x120,%rsp + DB 91 ; pop %rbx DB 255,224 ; jmpq *%rax DB 144 ; nop - DB 144 ; nop - DB 144 ; nop PUBLIC _sk_start_pipeline_hsw_8bit _sk_start_pipeline_hsw_8bit LABEL PROC @@ -33147,13 +33173,13 @@ _sk_start_pipeline_hsw_8bit LABEL PROC DB 73,137,246 ; mov %rsi,%r14 DB 76,137,141,8,255,255,255 ; mov %r9,-0xf8(%rbp) DB 76,57,203 ; cmp %r9,%rbx - DB 115,101 ; jae 22e98 <_sk_start_pipeline_hsw_8bit+0xe4> + DB 115,101 ; jae 23040 <_sk_start_pipeline_hsw_8bit+0xe4> DB 72,139,133,24,255,255,255 ; mov -0xe8(%rbp),%rax DB 72,141,64,16 ; lea 0x10(%rax),%rax DB 72,137,133,16,255,255,255 ; mov %rax,-0xf0(%rbp) DB 72,139,149,24,255,255,255 ; mov -0xe8(%rbp),%rdx DB 76,57,189,16,255,255,255 ; cmp %r15,-0xf0(%rbp) - DB 119,38 ; ja 22e7b <_sk_start_pipeline_hsw_8bit+0xc7> + DB 119,38 ; ja 23023 <_sk_start_pipeline_hsw_8bit+0xc7> DB 76,139,165,24,255,255,255 ; mov -0xe8(%rbp),%r12 DB 49,255 ; xor %edi,%edi DB 76,137,246 ; mov %r14,%rsi @@ -33164,16 +33190,16 @@ _sk_start_pipeline_hsw_8bit LABEL PROC DB 73,131,196,32 ; add $0x20,%r12 DB 77,57,252 ; cmp %r15,%r12 DB 73,137,212 ; mov %rdx,%r12 - DB 118,225 ; jbe 22e5c <_sk_start_pipeline_hsw_8bit+0xa8> + DB 118,225 ; jbe 23004 <_sk_start_pipeline_hsw_8bit+0xa8> DB 76,137,255 ; mov %r15,%rdi DB 72,41,215 ; sub %rdx,%rdi - DB 116,9 ; je 22e8c <_sk_start_pipeline_hsw_8bit+0xd8> + DB 116,9 ; je 23034 <_sk_start_pipeline_hsw_8bit+0xd8> DB 76,137,246 ; mov %r14,%rsi DB 72,137,217 ; mov %rbx,%rcx DB 65,255,213 ; callq *%r13 DB 72,255,195 ; inc %rbx DB 72,59,157,8,255,255,255 ; cmp -0xf8(%rbp),%rbx - DB 117,173 ; jne 22e45 <_sk_start_pipeline_hsw_8bit+0x91> + DB 117,173 ; jne 22fed <_sk_start_pipeline_hsw_8bit+0x91> DB 197,248,40,181,32,255,255,255 ; vmovaps -0xe0(%rbp),%xmm6 DB 197,248,40,189,48,255,255,255 ; vmovaps -0xd0(%rbp),%xmm7 DB 197,120,40,133,64,255,255,255 ; vmovaps -0xc0(%rbp),%xmm8 @@ -33211,7 +33237,7 @@ _sk_uniform_color_hsw_8bit LABEL PROC PUBLIC _sk_set_rgb_hsw_8bit _sk_set_rgb_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,250,16,37,202,163,0,0 ; vmovss 0xa3ca(%rip),%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 197,250,16,37,2,161,0,0 ; vmovss 0xa102(%rip),%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 197,218,89,40 ; vmulss (%rax),%xmm4,%xmm5 DB 196,97,250,44,197 ; vcvttss2si %xmm5,%r8 DB 197,218,89,104,4 ; vmulss 0x4(%rax),%xmm4,%xmm5 @@ -33224,7 +33250,7 @@ _sk_set_rgb_hsw_8bit LABEL PROC DB 68,9,200 ; or %r9d,%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 DB 196,226,125,88,228 ; vpbroadcastd %xmm4,%ymm4 - DB 197,253,111,45,247,158,0,0 ; vmovdqa 0x9ef7(%rip),%ymm5 # 2ce40 <_sk_overlay_sse2_8bit+0x8e8> + DB 197,253,111,45,79,156,0,0 ; vmovdqa 0x9c4f(%rip),%ymm5 # 2cd40 <_sk_overlay_sse2_8bit+0x891> DB 197,245,219,205 ; vpand %ymm5,%ymm1,%ymm1 DB 197,253,219,197 ; vpand %ymm5,%ymm0,%ymm0 DB 197,221,235,192 ; vpor %ymm0,%ymm4,%ymm0 @@ -33234,48 +33260,48 @@ _sk_set_rgb_hsw_8bit LABEL PROC PUBLIC _sk_premul_hsw_8bit _sk_premul_hsw_8bit LABEL PROC - DB 197,253,111,37,251,158,0,0 ; vmovdqa 0x9efb(%rip),%ymm4 # 2ce60 <_sk_overlay_sse2_8bit+0x908> + DB 197,253,111,37,83,156,0,0 ; vmovdqa 0x9c53(%rip),%ymm4 # 2cd60 <_sk_overlay_sse2_8bit+0x8b1> DB 196,226,125,0,236 ; vpshufb %ymm4,%ymm0,%ymm5 DB 196,226,117,0,228 ; vpshufb %ymm4,%ymm1,%ymm4 - DB 197,253,111,53,9,159,0,0 ; vmovdqa 0x9f09(%rip),%ymm6 # 2ce80 <_sk_overlay_sse2_8bit+0x928> - DB 197,221,235,230 ; vpor %ymm6,%ymm4,%ymm4 - DB 197,213,235,238 ; vpor %ymm6,%ymm5,%ymm5 - DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 196,98,125,48,206 ; vpmovzxbw %xmm6,%ymm9 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,98,125,48,197 ; vpmovzxbw %xmm5,%ymm8 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 197,125,111,5,65,156,0,0 ; vmovdqa 0x9c41(%rip),%ymm8 # 2cd80 <_sk_overlay_sse2_8bit+0x8d1> + DB 196,193,93,235,224 ; vpor %ymm8,%ymm4,%ymm4 + DB 196,193,85,235,232 ; vpor %ymm8,%ymm5,%ymm5 + DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,230,1 ; vextracti128 $0x1,%ymm4,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 197,221,213,225 ; vpmullw %ymm1,%ymm4,%ymm4 - DB 197,53,213,207 ; vpmullw %ymm7,%ymm9,%ymm9 + DB 197,205,213,247 ; vpmullw %ymm7,%ymm6,%ymm6 DB 197,213,213,232 ; vpmullw %ymm0,%ymm5,%ymm5 - DB 197,61,213,198 ; vpmullw %ymm6,%ymm8,%ymm8 - DB 197,189,253,246 ; vpaddw %ymm6,%ymm8,%ymm6 + DB 196,65,61,213,193 ; vpmullw %ymm9,%ymm8,%ymm8 + DB 196,65,61,253,193 ; vpaddw %ymm9,%ymm8,%ymm8 DB 197,213,253,192 ; vpaddw %ymm0,%ymm5,%ymm0 - DB 197,181,253,239 ; vpaddw %ymm7,%ymm9,%ymm5 + DB 197,205,253,239 ; vpaddw %ymm7,%ymm6,%ymm5 DB 197,221,253,201 ; vpaddw %ymm1,%ymm4,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,213,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm5 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,93,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm5 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_swap_rb_hsw_8bit _sk_swap_rb_hsw_8bit LABEL PROC - DB 197,253,111,37,129,158,0,0 ; vmovdqa 0x9e81(%rip),%ymm4 # 2cea0 <_sk_overlay_sse2_8bit+0x948> + DB 197,253,111,37,212,155,0,0 ; vmovdqa 0x9bd4(%rip),%ymm4 # 2cda0 <_sk_overlay_sse2_8bit+0x8f1> DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -33299,9 +33325,9 @@ _sk_load_8888_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 2306a <_sk_load_8888_hsw_8bit+0x2d> - DB 196,129,126,111,76,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm1 + DB 117,17 ; jne 23217 <_sk_load_8888_hsw_8bit+0x2d> DB 196,129,126,111,4,152 ; vmovdqu (%r8,%r11,4),%ymm0 + DB 196,129,126,111,76,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -33310,14 +33336,14 @@ _sk_load_8888_hsw_8bit LABEL PROC DB 197,253,239,192 ; vpxor %ymm0,%ymm0,%ymm0 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,228 ; ja 23066 <_sk_load_8888_hsw_8bit+0x29> + DB 119,228 ; ja 23213 <_sk_load_8888_hsw_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,23,1,0,0 ; lea 0x117(%rip),%r10 # 231a4 <_sk_load_8888_hsw_8bit+0x167> + DB 76,141,21,30,1,0,0 ; lea 0x11e(%rip),%r10 # 23358 <_sk_load_8888_hsw_8bit+0x16e> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,4,152 ; vmovd (%r8,%r11,4),%xmm0 - DB 235,200 ; jmp 23066 <_sk_load_8888_hsw_8bit+0x29> + DB 235,200 ; jmp 23213 <_sk_load_8888_hsw_8bit+0x29> DB 196,129,121,110,68,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33325,7 +33351,7 @@ _sk_load_8888_hsw_8bit LABEL PROC DB 196,130,121,53,36,152 ; vpmovzxdq (%r8,%r11,4),%xmm4 DB 197,249,112,228,232 ; vpshufd $0xe8,%xmm4,%xmm4 DB 196,227,125,2,196,3 ; vpblendd $0x3,%ymm4,%ymm0,%ymm0 - DB 235,159 ; jmp 23066 <_sk_load_8888_hsw_8bit+0x29> + DB 235,159 ; jmp 23213 <_sk_load_8888_hsw_8bit+0x29> DB 196,129,121,110,68,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33338,7 +33364,7 @@ _sk_load_8888_hsw_8bit LABEL PROC DB 196,227,125,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm0,%ymm0 DB 196,129,122,111,36,152 ; vmovdqu (%r8,%r11,4),%xmm4 DB 196,227,93,2,192,240 ; vpblendd $0xf0,%ymm0,%ymm4,%ymm0 - DB 233,80,255,255,255 ; jmpq 23066 <_sk_load_8888_hsw_8bit+0x29> + DB 233,80,255,255,255 ; jmpq 23213 <_sk_load_8888_hsw_8bit+0x29> DB 196,129,121,110,68,152,40 ; vmovd 0x28(%r8,%r11,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33347,7 +33373,8 @@ _sk_load_8888_hsw_8bit LABEL PROC DB 196,227,117,2,200,15 ; vpblendd $0xf,%ymm0,%ymm1,%ymm1 DB 196,129,121,110,68,152,32 ; vmovd 0x20(%r8,%r11,4),%xmm0 DB 196,227,117,2,200,1 ; vpblendd $0x1,%ymm0,%ymm1,%ymm1 - DB 233,20,255,255,255 ; jmpq 23060 <_sk_load_8888_hsw_8bit+0x23> + DB 196,129,126,111,4,152 ; vmovdqu (%r8,%r11,4),%ymm0 + DB 233,20,255,255,255 ; jmpq 23213 <_sk_load_8888_hsw_8bit+0x29> DB 196,129,121,110,68,152,56 ; vmovd 0x38(%r8,%r11,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33361,42 +33388,38 @@ _sk_load_8888_hsw_8bit LABEL PROC DB 196,129,126,111,4,152 ; vmovdqu (%r8,%r11,4),%ymm0 DB 196,129,122,111,100,152,32 ; vmovdqu 0x20(%r8,%r11,4),%xmm4 DB 196,227,93,2,201,240 ; vpblendd $0xf0,%ymm1,%ymm4,%ymm1 - DB 233,196,254,255,255 ; jmpq 23066 <_sk_load_8888_hsw_8bit+0x29> - DB 102,144 ; xchg %ax,%ax - DB 242,254 ; repnz (bad) - DB 255 ; (bad) - DB 255,16 ; callq *(%rax) + DB 233,190,254,255,255 ; jmpq 23213 <_sk_load_8888_hsw_8bit+0x29> + DB 15,31,0 ; nopl (%rax) + DB 235,254 ; jmp 23358 <_sk_load_8888_hsw_8bit+0x16e> DB 255 ; (bad) + DB 255,9 ; decl (%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 250 ; cli + DB 255,243 ; push %rbx DB 254 ; (bad) DB 255 ; (bad) - DB 255,97,255 ; jmpq *-0x1(%rcx) - DB 255 ; (bad) - DB 255,77,255 ; decl -0x1(%rbp) - DB 255 ; (bad) - DB 255 ; (bad) - DB 57,255 ; cmp %edi,%edi + DB 255,90,255 ; lcall *-0x1(%rdx) DB 255 ; (bad) - DB 255,35 ; jmpq *(%rbx) + DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) + DB 255,50 ; pushq (%rdx) DB 255 ; (bad) DB 255 ; (bad) - DB 188,254,255,255,150 ; mov $0x96fffffe,%esp + DB 255,28,255 ; lcall *(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,156,255,255,255,143,255 ; lcall *-0x700001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,136,255,255,255,114 ; decl 0x72ffffff(%rax) + DB 255,129,255,255,255,107 ; incl 0x6bffffff(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 190,255,255,255,168 ; mov $0xa8ffffff,%esi + DB 189,255,255,255,167 ; mov $0xa7ffffff,%ebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -33411,9 +33434,9 @@ _sk_load_8888_dst_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 2320d <_sk_load_8888_dst_hsw_8bit+0x2d> - DB 196,129,126,111,92,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm3 + DB 117,17 ; jne 233c1 <_sk_load_8888_dst_hsw_8bit+0x2d> DB 196,129,126,111,20,152 ; vmovdqu (%r8,%r11,4),%ymm2 + DB 196,129,126,111,92,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -33422,14 +33445,14 @@ _sk_load_8888_dst_hsw_8bit LABEL PROC DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,228 ; ja 23209 <_sk_load_8888_dst_hsw_8bit+0x29> + DB 119,228 ; ja 233bd <_sk_load_8888_dst_hsw_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,24,1,0,0 ; lea 0x118(%rip),%r10 # 23348 <_sk_load_8888_dst_hsw_8bit+0x168> + DB 76,141,21,28,1,0,0 ; lea 0x11c(%rip),%r10 # 23500 <_sk_load_8888_dst_hsw_8bit+0x16c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,20,152 ; vmovd (%r8,%r11,4),%xmm2 - DB 235,200 ; jmp 23209 <_sk_load_8888_dst_hsw_8bit+0x29> + DB 235,200 ; jmp 233bd <_sk_load_8888_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm2 DB 196,226,121,89,210 ; vpbroadcastq %xmm2,%xmm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33437,7 +33460,7 @@ _sk_load_8888_dst_hsw_8bit LABEL PROC DB 196,130,121,53,36,152 ; vpmovzxdq (%r8,%r11,4),%xmm4 DB 197,249,112,228,232 ; vpshufd $0xe8,%xmm4,%xmm4 DB 196,227,109,2,212,3 ; vpblendd $0x3,%ymm4,%ymm2,%ymm2 - DB 235,159 ; jmp 23209 <_sk_load_8888_dst_hsw_8bit+0x29> + DB 235,159 ; jmp 233bd <_sk_load_8888_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm2 DB 196,226,125,89,210 ; vpbroadcastq %xmm2,%ymm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33450,7 +33473,7 @@ _sk_load_8888_dst_hsw_8bit LABEL PROC DB 196,227,109,56,212,1 ; vinserti128 $0x1,%xmm4,%ymm2,%ymm2 DB 196,129,122,111,36,152 ; vmovdqu (%r8,%r11,4),%xmm4 DB 196,227,93,2,210,240 ; vpblendd $0xf0,%ymm2,%ymm4,%ymm2 - DB 233,80,255,255,255 ; jmpq 23209 <_sk_load_8888_dst_hsw_8bit+0x29> + DB 233,80,255,255,255 ; jmpq 233bd <_sk_load_8888_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,40 ; vmovd 0x28(%r8,%r11,4),%xmm2 DB 196,226,121,89,210 ; vpbroadcastq %xmm2,%xmm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33459,7 +33482,8 @@ _sk_load_8888_dst_hsw_8bit LABEL PROC DB 196,227,101,2,218,15 ; vpblendd $0xf,%ymm2,%ymm3,%ymm3 DB 196,129,121,110,84,152,32 ; vmovd 0x20(%r8,%r11,4),%xmm2 DB 196,227,101,2,218,1 ; vpblendd $0x1,%ymm2,%ymm3,%ymm3 - DB 233,20,255,255,255 ; jmpq 23203 <_sk_load_8888_dst_hsw_8bit+0x23> + DB 196,129,126,111,20,152 ; vmovdqu (%r8,%r11,4),%ymm2 + DB 233,20,255,255,255 ; jmpq 233bd <_sk_load_8888_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,56 ; vmovd 0x38(%r8,%r11,4),%xmm2 DB 196,226,125,89,210 ; vpbroadcastq %xmm2,%ymm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33473,42 +33497,39 @@ _sk_load_8888_dst_hsw_8bit LABEL PROC DB 196,129,126,111,20,152 ; vmovdqu (%r8,%r11,4),%ymm2 DB 196,129,122,111,100,152,32 ; vmovdqu 0x20(%r8,%r11,4),%xmm4 DB 196,227,93,2,219,240 ; vpblendd $0xf0,%ymm3,%ymm4,%ymm3 - DB 233,196,254,255,255 ; jmpq 23209 <_sk_load_8888_dst_hsw_8bit+0x29> - DB 15,31,0 ; nopl (%rax) - DB 241 ; icebp + DB 233,190,254,255,255 ; jmpq 233bd <_sk_load_8888_dst_hsw_8bit+0x29> + DB 144 ; nop + DB 237 ; in (%dx),%eax DB 254 ; (bad) DB 255 ; (bad) - DB 255,15 ; decl (%rdi) - DB 255 ; (bad) + DB 255,11 ; decl (%rbx) DB 255 ; (bad) DB 255 ; (bad) - DB 249 ; stc + DB 255,245 ; push %rbp DB 254 ; (bad) DB 255 ; (bad) - DB 255,96,255 ; jmpq *-0x1(%rax) - DB 255 ; (bad) - DB 255,76,255,255 ; decl -0x1(%rdi,%rdi,8) - DB 255 ; (bad) - DB 56,255 ; cmp %bh,%bh + DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) + DB 255,72,255 ; decl -0x1(%rax) DB 255 ; (bad) - DB 255,34 ; jmpq *(%rdx) + DB 255,52,255 ; pushq (%rdi,%rdi,8) DB 255 ; (bad) + DB 255,30 ; lcall *(%rsi) DB 255 ; (bad) DB 255 ; (bad) - DB 187,254,255,255,149 ; mov $0x95fffffe,%ebx + DB 255,158,255,255,255,145 ; lcall *-0x6e000001(%rsi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,135,255,255,255,113 ; incl 0x71ffffff(%rdi) + DB 255,131,255,255,255,109 ; incl 0x6dffffff(%rbx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,229 ; jmpq *%rbp + DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) - DB 255,209 ; callq *%rcx + DB 255,211 ; callq *%rbx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 189,255,255,255,167 ; mov $0xa7ffffff,%ebp + DB 191,255,255,255,169 ; mov $0xa9ffffff,%edi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -33523,7 +33544,7 @@ _sk_store_8888_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 233b1 <_sk_store_8888_hsw_8bit+0x2d> + DB 117,17 ; jne 23569 <_sk_store_8888_hsw_8bit+0x2d> DB 196,129,126,127,4,152 ; vmovdqu %ymm0,(%r8,%r11,4) DB 196,129,126,127,76,152,32 ; vmovdqu %ymm1,0x20(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -33532,17 +33553,17 @@ _sk_store_8888_hsw_8bit LABEL PROC DB 65,128,225,15 ; and $0xf,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,236 ; ja 233ad <_sk_store_8888_hsw_8bit+0x29> + DB 119,236 ; ja 23565 <_sk_store_8888_hsw_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,176,0,0,0 ; lea 0xb0(%rip),%r10 # 2347c <_sk_store_8888_hsw_8bit+0xf8> + DB 76,141,21,176,0,0,0 ; lea 0xb0(%rip),%r10 # 23634 <_sk_store_8888_hsw_8bit+0xf8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,126,4,152 ; vmovd %xmm0,(%r8,%r11,4) - DB 235,208 ; jmp 233ad <_sk_store_8888_hsw_8bit+0x29> + DB 235,208 ; jmp 23565 <_sk_store_8888_hsw_8bit+0x29> DB 196,131,121,22,68,152,8,2 ; vpextrd $0x2,%xmm0,0x8(%r8,%r11,4) DB 196,129,121,214,4,152 ; vmovq %xmm0,(%r8,%r11,4) - DB 235,192 ; jmp 233ad <_sk_store_8888_hsw_8bit+0x29> + DB 235,192 ; jmp 23565 <_sk_store_8888_hsw_8bit+0x29> DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 DB 196,131,121,22,100,152,24,2 ; vpextrd $0x2,%xmm4,0x18(%r8,%r11,4) DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 @@ -33550,12 +33571,12 @@ _sk_store_8888_hsw_8bit LABEL PROC DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 DB 196,129,121,126,100,152,16 ; vmovd %xmm4,0x10(%r8,%r11,4) DB 196,129,122,127,4,152 ; vmovdqu %xmm0,(%r8,%r11,4) - DB 235,143 ; jmp 233ad <_sk_store_8888_hsw_8bit+0x29> + DB 235,143 ; jmp 23565 <_sk_store_8888_hsw_8bit+0x29> DB 196,131,121,22,76,152,40,2 ; vpextrd $0x2,%xmm1,0x28(%r8,%r11,4) DB 196,131,121,22,76,152,36,1 ; vpextrd $0x1,%xmm1,0x24(%r8,%r11,4) DB 196,129,121,126,76,152,32 ; vmovd %xmm1,0x20(%r8,%r11,4) DB 196,129,126,127,4,152 ; vmovdqu %ymm0,(%r8,%r11,4) - DB 233,109,255,255,255 ; jmpq 233ad <_sk_store_8888_hsw_8bit+0x29> + DB 233,109,255,255,255 ; jmpq 23565 <_sk_store_8888_hsw_8bit+0x29> DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 DB 196,131,121,22,100,152,56,2 ; vpextrd $0x2,%xmm4,0x38(%r8,%r11,4) DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 @@ -33564,7 +33585,7 @@ _sk_store_8888_hsw_8bit LABEL PROC DB 196,129,121,126,100,152,48 ; vmovd %xmm4,0x30(%r8,%r11,4) DB 196,129,126,127,4,152 ; vmovdqu %ymm0,(%r8,%r11,4) DB 196,129,122,127,76,152,32 ; vmovdqu %xmm1,0x20(%r8,%r11,4) - DB 233,50,255,255,255 ; jmpq 233ad <_sk_store_8888_hsw_8bit+0x29> + DB 233,50,255,255,255 ; jmpq 23565 <_sk_store_8888_hsw_8bit+0x29> DB 144 ; nop DB 89 ; pop %rcx DB 255 ; (bad) @@ -33577,7 +33598,7 @@ _sk_store_8888_hsw_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 127,255 ; jg 23491 <_sk_store_8888_hsw_8bit+0x10d> + DB 127,255 ; jg 23649 <_sk_store_8888_hsw_8bit+0x10d> DB 255 ; (bad) DB 255,113,255 ; pushq -0x1(%rcx) DB 255 ; (bad) @@ -33613,10 +33634,10 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,35 ; jne 234f7 <_sk_load_bgra_hsw_8bit+0x3f> - DB 196,129,126,111,76,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm1 + DB 117,35 ; jne 236af <_sk_load_bgra_hsw_8bit+0x3f> DB 196,129,126,111,4,152 ; vmovdqu (%r8,%r11,4),%ymm0 - DB 197,253,111,37,215,153,0,0 ; vmovdqa 0x99d7(%rip),%ymm4 # 2cec0 <_sk_overlay_sse2_8bit+0x968> + DB 196,129,126,111,76,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm1 + DB 197,253,111,37,31,151,0,0 ; vmovdqa 0x971f(%rip),%ymm4 # 2cdc0 <_sk_overlay_sse2_8bit+0x911> DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -33627,14 +33648,14 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 197,253,239,192 ; vpxor %ymm0,%ymm0,%ymm0 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,210 ; ja 234e1 <_sk_load_bgra_hsw_8bit+0x29> + DB 119,210 ; ja 23699 <_sk_load_bgra_hsw_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,22,1,0,0 ; lea 0x116(%rip),%r10 # 23630 <_sk_load_bgra_hsw_8bit+0x178> + DB 76,141,21,30,1,0,0 ; lea 0x11e(%rip),%r10 # 237f0 <_sk_load_bgra_hsw_8bit+0x180> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,4,152 ; vmovd (%r8,%r11,4),%xmm0 - DB 235,182 ; jmp 234e1 <_sk_load_bgra_hsw_8bit+0x29> + DB 235,182 ; jmp 23699 <_sk_load_bgra_hsw_8bit+0x29> DB 196,129,121,110,68,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33642,7 +33663,7 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 196,130,121,53,36,152 ; vpmovzxdq (%r8,%r11,4),%xmm4 DB 197,249,112,228,232 ; vpshufd $0xe8,%xmm4,%xmm4 DB 196,227,125,2,196,3 ; vpblendd $0x3,%ymm4,%ymm0,%ymm0 - DB 235,141 ; jmp 234e1 <_sk_load_bgra_hsw_8bit+0x29> + DB 235,141 ; jmp 23699 <_sk_load_bgra_hsw_8bit+0x29> DB 196,129,121,110,68,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33655,7 +33676,7 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 196,227,125,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm0,%ymm0 DB 196,129,122,111,36,152 ; vmovdqu (%r8,%r11,4),%xmm4 DB 196,227,93,2,192,240 ; vpblendd $0xf0,%ymm0,%ymm4,%ymm0 - DB 233,62,255,255,255 ; jmpq 234e1 <_sk_load_bgra_hsw_8bit+0x29> + DB 233,62,255,255,255 ; jmpq 23699 <_sk_load_bgra_hsw_8bit+0x29> DB 196,129,121,110,68,152,40 ; vmovd 0x28(%r8,%r11,4),%xmm0 DB 196,226,121,89,192 ; vpbroadcastq %xmm0,%xmm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33664,7 +33685,8 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 196,227,117,2,200,15 ; vpblendd $0xf,%ymm0,%ymm1,%ymm1 DB 196,129,121,110,68,152,32 ; vmovd 0x20(%r8,%r11,4),%xmm0 DB 196,227,117,2,200,1 ; vpblendd $0x1,%ymm0,%ymm1,%ymm1 - DB 233,2,255,255,255 ; jmpq 234db <_sk_load_bgra_hsw_8bit+0x23> + DB 196,129,126,111,4,152 ; vmovdqu (%r8,%r11,4),%ymm0 + DB 233,2,255,255,255 ; jmpq 23699 <_sk_load_bgra_hsw_8bit+0x29> DB 196,129,121,110,68,152,56 ; vmovd 0x38(%r8,%r11,4),%xmm0 DB 196,226,125,89,192 ; vpbroadcastq %xmm0,%ymm0 DB 197,245,239,201 ; vpxor %ymm1,%ymm1,%ymm1 @@ -33678,40 +33700,38 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 196,129,126,111,4,152 ; vmovdqu (%r8,%r11,4),%ymm0 DB 196,129,122,111,100,152,32 ; vmovdqu 0x20(%r8,%r11,4),%xmm4 DB 196,227,93,2,201,240 ; vpblendd $0xf0,%ymm1,%ymm4,%ymm1 - DB 233,178,254,255,255 ; jmpq 234e1 <_sk_load_bgra_hsw_8bit+0x29> - DB 144 ; nop - DB 243,254 ; repz (bad) - DB 255 ; (bad) - DB 255,17 ; callq *(%rcx) + DB 233,172,254,255,255 ; jmpq 23699 <_sk_load_bgra_hsw_8bit+0x29> + DB 15,31,0 ; nopl (%rax) + DB 235,254 ; jmp 237f0 <_sk_load_bgra_hsw_8bit+0x180> DB 255 ; (bad) + DB 255,9 ; decl (%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 251 ; sti + DB 255,243 ; push %rbx DB 254 ; (bad) DB 255 ; (bad) - DB 255,98,255 ; jmpq *-0x1(%rdx) - DB 255 ; (bad) - DB 255,78,255 ; decl -0x1(%rsi) + DB 255,90,255 ; lcall *-0x1(%rdx) DB 255 ; (bad) + DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) - DB 58,255 ; cmp %bh,%bh + DB 255,50 ; pushq (%rdx) DB 255 ; (bad) - DB 255,36,255 ; jmpq *(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,171,254,255,255,151 ; ljmp *-0x68000002(%rbx) + DB 255,28,255 ; lcall *(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,156,255,255,255,143,255 ; lcall *-0x700001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,137,255,255,255,115 ; decl 0x73ffffff(%rcx) + DB 255,129,255,255,255,107 ; incl 0x6bffffff(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 255,211 ; callq *%rbx + DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 191,255,255,255,169 ; mov $0xa9ffffff,%edi + DB 189,255,255,255,167 ; mov $0xa7ffffff,%ebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -33726,10 +33746,10 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,35 ; jne 236ab <_sk_load_bgra_dst_hsw_8bit+0x3f> - DB 196,129,126,111,92,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm3 + DB 117,35 ; jne 2386b <_sk_load_bgra_dst_hsw_8bit+0x3f> DB 196,129,126,111,20,152 ; vmovdqu (%r8,%r11,4),%ymm2 - DB 197,253,111,37,67,152,0,0 ; vmovdqa 0x9843(%rip),%ymm4 # 2cee0 <_sk_overlay_sse2_8bit+0x988> + DB 196,129,126,111,92,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm3 + DB 197,253,111,37,131,149,0,0 ; vmovdqa 0x9583(%rip),%ymm4 # 2cde0 <_sk_overlay_sse2_8bit+0x931> DB 196,226,109,0,212 ; vpshufb %ymm4,%ymm2,%ymm2 DB 196,226,101,0,220 ; vpshufb %ymm4,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -33740,14 +33760,14 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,210 ; ja 23695 <_sk_load_bgra_dst_hsw_8bit+0x29> + DB 119,210 ; ja 23855 <_sk_load_bgra_dst_hsw_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,22,1,0,0 ; lea 0x116(%rip),%r10 # 237e4 <_sk_load_bgra_dst_hsw_8bit+0x178> + DB 76,141,21,30,1,0,0 ; lea 0x11e(%rip),%r10 # 239ac <_sk_load_bgra_dst_hsw_8bit+0x180> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,110,20,152 ; vmovd (%r8,%r11,4),%xmm2 - DB 235,182 ; jmp 23695 <_sk_load_bgra_dst_hsw_8bit+0x29> + DB 235,182 ; jmp 23855 <_sk_load_bgra_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm2 DB 196,226,121,89,210 ; vpbroadcastq %xmm2,%xmm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33755,7 +33775,7 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 196,130,121,53,36,152 ; vpmovzxdq (%r8,%r11,4),%xmm4 DB 197,249,112,228,232 ; vpshufd $0xe8,%xmm4,%xmm4 DB 196,227,109,2,212,3 ; vpblendd $0x3,%ymm4,%ymm2,%ymm2 - DB 235,141 ; jmp 23695 <_sk_load_bgra_dst_hsw_8bit+0x29> + DB 235,141 ; jmp 23855 <_sk_load_bgra_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm2 DB 196,226,125,89,210 ; vpbroadcastq %xmm2,%ymm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33768,7 +33788,7 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 196,227,109,56,212,1 ; vinserti128 $0x1,%xmm4,%ymm2,%ymm2 DB 196,129,122,111,36,152 ; vmovdqu (%r8,%r11,4),%xmm4 DB 196,227,93,2,210,240 ; vpblendd $0xf0,%ymm2,%ymm4,%ymm2 - DB 233,62,255,255,255 ; jmpq 23695 <_sk_load_bgra_dst_hsw_8bit+0x29> + DB 233,62,255,255,255 ; jmpq 23855 <_sk_load_bgra_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,40 ; vmovd 0x28(%r8,%r11,4),%xmm2 DB 196,226,121,89,210 ; vpbroadcastq %xmm2,%xmm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33777,7 +33797,8 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 196,227,101,2,218,15 ; vpblendd $0xf,%ymm2,%ymm3,%ymm3 DB 196,129,121,110,84,152,32 ; vmovd 0x20(%r8,%r11,4),%xmm2 DB 196,227,101,2,218,1 ; vpblendd $0x1,%ymm2,%ymm3,%ymm3 - DB 233,2,255,255,255 ; jmpq 2368f <_sk_load_bgra_dst_hsw_8bit+0x23> + DB 196,129,126,111,20,152 ; vmovdqu (%r8,%r11,4),%ymm2 + DB 233,2,255,255,255 ; jmpq 23855 <_sk_load_bgra_dst_hsw_8bit+0x29> DB 196,129,121,110,84,152,56 ; vmovd 0x38(%r8,%r11,4),%xmm2 DB 196,226,125,89,210 ; vpbroadcastq %xmm2,%ymm2 DB 197,229,239,219 ; vpxor %ymm3,%ymm3,%ymm3 @@ -33791,40 +33812,38 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 196,129,126,111,20,152 ; vmovdqu (%r8,%r11,4),%ymm2 DB 196,129,122,111,100,152,32 ; vmovdqu 0x20(%r8,%r11,4),%xmm4 DB 196,227,93,2,219,240 ; vpblendd $0xf0,%ymm3,%ymm4,%ymm3 - DB 233,178,254,255,255 ; jmpq 23695 <_sk_load_bgra_dst_hsw_8bit+0x29> - DB 144 ; nop - DB 243,254 ; repz (bad) - DB 255 ; (bad) - DB 255,17 ; callq *(%rcx) + DB 233,172,254,255,255 ; jmpq 23855 <_sk_load_bgra_dst_hsw_8bit+0x29> + DB 15,31,0 ; nopl (%rax) + DB 235,254 ; jmp 239ac <_sk_load_bgra_dst_hsw_8bit+0x180> DB 255 ; (bad) + DB 255,9 ; decl (%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 251 ; sti + DB 255,243 ; push %rbx DB 254 ; (bad) DB 255 ; (bad) - DB 255,98,255 ; jmpq *-0x1(%rdx) - DB 255 ; (bad) - DB 255,78,255 ; decl -0x1(%rsi) + DB 255,90,255 ; lcall *-0x1(%rdx) DB 255 ; (bad) + DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) - DB 58,255 ; cmp %bh,%bh + DB 255,50 ; pushq (%rdx) DB 255 ; (bad) - DB 255,36,255 ; jmpq *(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,171,254,255,255,151 ; ljmp *-0x68000002(%rbx) + DB 255,28,255 ; lcall *(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,156,255,255,255,143,255 ; lcall *-0x700001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,137,255,255,255,115 ; decl 0x73ffffff(%rcx) + DB 255,129,255,255,255,107 ; incl 0x6bffffff(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 255,211 ; callq *%rbx + DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 191,255,255,255,169 ; mov $0xa9ffffff,%edi + DB 189,255,255,255,167 ; mov $0xa7ffffff,%ebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -33838,11 +33857,11 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 197,253,111,37,193,150,0,0 ; vmovdqa 0x96c1(%rip),%ymm4 # 2cf00 <_sk_overlay_sse2_8bit+0x9a8> + DB 197,253,111,37,249,147,0,0 ; vmovdqa 0x93f9(%rip),%ymm4 # 2ce00 <_sk_overlay_sse2_8bit+0x951> DB 196,226,117,0,236 ; vpshufb %ymm4,%ymm1,%ymm5 DB 196,226,125,0,228 ; vpshufb %ymm4,%ymm0,%ymm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 2385f <_sk_store_bgra_hsw_8bit+0x3f> + DB 117,17 ; jne 23a27 <_sk_store_bgra_hsw_8bit+0x3f> DB 196,129,126,127,36,152 ; vmovdqu %ymm4,(%r8,%r11,4) DB 196,129,126,127,108,152,32 ; vmovdqu %ymm5,0x20(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -33851,17 +33870,17 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 65,128,225,15 ; and $0xf,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,236 ; ja 2385b <_sk_store_bgra_hsw_8bit+0x3b> + DB 119,236 ; ja 23a23 <_sk_store_bgra_hsw_8bit+0x3b> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,178,0,0,0 ; lea 0xb2(%rip),%r10 # 2392c <_sk_store_bgra_hsw_8bit+0x10c> + DB 76,141,21,178,0,0,0 ; lea 0xb2(%rip),%r10 # 23af4 <_sk_store_bgra_hsw_8bit+0x10c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,129,121,126,36,152 ; vmovd %xmm4,(%r8,%r11,4) - DB 235,208 ; jmp 2385b <_sk_store_bgra_hsw_8bit+0x3b> + DB 235,208 ; jmp 23a23 <_sk_store_bgra_hsw_8bit+0x3b> DB 196,131,121,22,100,152,8,2 ; vpextrd $0x2,%xmm4,0x8(%r8,%r11,4) DB 196,129,121,214,36,152 ; vmovq %xmm4,(%r8,%r11,4) - DB 235,192 ; jmp 2385b <_sk_store_bgra_hsw_8bit+0x3b> + DB 235,192 ; jmp 23a23 <_sk_store_bgra_hsw_8bit+0x3b> DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,131,121,22,108,152,24,2 ; vpextrd $0x2,%xmm5,0x18(%r8,%r11,4) DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 @@ -33869,12 +33888,12 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,129,121,126,108,152,16 ; vmovd %xmm5,0x10(%r8,%r11,4) DB 196,129,122,127,36,152 ; vmovdqu %xmm4,(%r8,%r11,4) - DB 235,143 ; jmp 2385b <_sk_store_bgra_hsw_8bit+0x3b> + DB 235,143 ; jmp 23a23 <_sk_store_bgra_hsw_8bit+0x3b> DB 196,131,121,22,108,152,40,2 ; vpextrd $0x2,%xmm5,0x28(%r8,%r11,4) DB 196,131,121,22,108,152,36,1 ; vpextrd $0x1,%xmm5,0x24(%r8,%r11,4) DB 196,129,121,126,108,152,32 ; vmovd %xmm5,0x20(%r8,%r11,4) DB 196,129,126,127,36,152 ; vmovdqu %ymm4,(%r8,%r11,4) - DB 233,109,255,255,255 ; jmpq 2385b <_sk_store_bgra_hsw_8bit+0x3b> + DB 233,109,255,255,255 ; jmpq 23a23 <_sk_store_bgra_hsw_8bit+0x3b> DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 DB 196,131,121,22,116,152,56,2 ; vpextrd $0x2,%xmm6,0x38(%r8,%r11,4) DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 @@ -33883,7 +33902,7 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 196,129,121,126,116,152,48 ; vmovd %xmm6,0x30(%r8,%r11,4) DB 196,129,126,127,36,152 ; vmovdqu %ymm4,(%r8,%r11,4) DB 196,129,122,127,108,152,32 ; vmovdqu %xmm5,0x20(%r8,%r11,4) - DB 233,50,255,255,255 ; jmpq 2385b <_sk_store_bgra_hsw_8bit+0x3b> + DB 233,50,255,255,255 ; jmpq 23a23 <_sk_store_bgra_hsw_8bit+0x3b> DB 15,31,0 ; nopl (%rax) DB 87 ; push %rdi DB 255 ; (bad) @@ -33896,7 +33915,7 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 125,255 ; jge 23941 <_sk_store_bgra_hsw_8bit+0x121> + DB 125,255 ; jge 23b09 <_sk_store_bgra_hsw_8bit+0x121> DB 255 ; (bad) DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) @@ -33907,7 +33926,7 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 23959 <_sk_store_bgra_hsw_8bit+0x139> + DB 235,255 ; jmp 23b21 <_sk_store_bgra_hsw_8bit+0x139> DB 255 ; (bad) DB 255 ; (bad) DB 222,255 ; fdivrp %st,%st(7) @@ -33929,33 +33948,33 @@ _sk_store_565_hsw_8bit LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 197,253,111,37,154,149,0,0 ; vmovdqa 0x959a(%rip),%ymm4 # 2cf20 <_sk_overlay_sse2_8bit+0x9c8> + DB 196,226,125,88,37,205,151,0,0 ; vpbroadcastd 0x97cd(%rip),%ymm4 # 2d31c <_sk_overlay_sse2_8bit+0xe6d> DB 197,253,219,236 ; vpand %ymm4,%ymm0,%ymm5 DB 197,245,219,228 ; vpand %ymm4,%ymm1,%ymm4 DB 197,221,114,244,8 ; vpslld $0x8,%ymm4,%ymm4 DB 197,213,114,245,8 ; vpslld $0x8,%ymm5,%ymm5 - DB 197,253,111,53,160,149,0,0 ; vmovdqa 0x95a0(%rip),%ymm6 # 2cf40 <_sk_overlay_sse2_8bit+0x9e8> + DB 196,226,125,88,53,182,151,0,0 ; vpbroadcastd 0x97b6(%rip),%ymm6 # 2d320 <_sk_overlay_sse2_8bit+0xe71> DB 197,253,219,254 ; vpand %ymm6,%ymm0,%ymm7 DB 197,245,219,246 ; vpand %ymm6,%ymm1,%ymm6 DB 197,205,114,214,5 ; vpsrld $0x5,%ymm6,%ymm6 - DB 197,197,114,215,5 ; vpsrld $0x5,%ymm7,%ymm7 - DB 197,125,111,5,166,149,0,0 ; vmovdqa 0x95a6(%rip),%ymm8 # 2cf60 <_sk_overlay_sse2_8bit+0xa08> - DB 196,65,117,219,200 ; vpand %ymm8,%ymm1,%ymm9 - DB 196,65,125,219,192 ; vpand %ymm8,%ymm0,%ymm8 - DB 196,193,61,114,208,19 ; vpsrld $0x13,%ymm8,%ymm8 - DB 196,193,53,114,209,19 ; vpsrld $0x13,%ymm9,%ymm9 - DB 197,213,235,239 ; vpor %ymm7,%ymm5,%ymm5 DB 197,221,235,230 ; vpor %ymm6,%ymm4,%ymm4 - DB 196,193,93,235,225 ; vpor %ymm9,%ymm4,%ymm4 - DB 196,193,85,235,232 ; vpor %ymm8,%ymm5,%ymm5 - DB 197,253,111,53,150,149,0,0 ; vmovdqa 0x9596(%rip),%ymm6 # 2cf80 <_sk_overlay_sse2_8bit+0xa28> + DB 197,205,114,215,5 ; vpsrld $0x5,%ymm7,%ymm6 + DB 197,213,235,238 ; vpor %ymm6,%ymm5,%ymm5 + DB 196,226,125,88,53,151,151,0,0 ; vpbroadcastd 0x9797(%rip),%ymm6 # 2d324 <_sk_overlay_sse2_8bit+0xe75> + DB 197,245,219,254 ; vpand %ymm6,%ymm1,%ymm7 + DB 197,253,219,246 ; vpand %ymm6,%ymm0,%ymm6 + DB 197,205,114,214,19 ; vpsrld $0x13,%ymm6,%ymm6 + DB 197,213,235,238 ; vpor %ymm6,%ymm5,%ymm5 + DB 197,205,114,215,19 ; vpsrld $0x13,%ymm7,%ymm6 + DB 197,221,235,230 ; vpor %ymm6,%ymm4,%ymm4 + DB 197,253,111,53,113,146,0,0 ; vmovdqa 0x9271(%rip),%ymm6 # 2ce20 <_sk_overlay_sse2_8bit+0x971> DB 196,226,85,0,238 ; vpshufb %ymm6,%ymm5,%ymm5 DB 196,227,253,0,237,232 ; vpermq $0xe8,%ymm5,%ymm5 DB 196,226,93,0,230 ; vpshufb %ymm6,%ymm4,%ymm4 DB 196,227,253,0,228,232 ; vpermq $0xe8,%ymm4,%ymm4 DB 196,227,85,56,228,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 23a15 <_sk_store_565_hsw_8bit+0xad> + DB 117,10 ; jne 23bda <_sk_store_565_hsw_8bit+0xaa> DB 196,129,126,127,36,88 ; vmovdqu %ymm4,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -33963,29 +33982,29 @@ _sk_store_565_hsw_8bit LABEL PROC DB 65,128,225,15 ; and $0xf,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,236 ; ja 23a11 <_sk_store_565_hsw_8bit+0xa9> + DB 119,236 ; ja 23bd6 <_sk_store_565_hsw_8bit+0xa6> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,176,0,0,0 ; lea 0xb0(%rip),%r10 # 23ae0 <_sk_store_565_hsw_8bit+0x178> + DB 76,141,21,179,0,0,0 ; lea 0xb3(%rip),%r10 # 23ca8 <_sk_store_565_hsw_8bit+0x178> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,131,121,21,36,88,0 ; vpextrw $0x0,%xmm4,(%r8,%r11,2) - DB 235,207 ; jmp 23a11 <_sk_store_565_hsw_8bit+0xa9> + DB 235,207 ; jmp 23bd6 <_sk_store_565_hsw_8bit+0xa6> DB 196,131,121,21,100,88,4,2 ; vpextrw $0x2,%xmm4,0x4(%r8,%r11,2) DB 196,129,121,126,36,88 ; vmovd %xmm4,(%r8,%r11,2) - DB 235,191 ; jmp 23a11 <_sk_store_565_hsw_8bit+0xa9> + DB 235,191 ; jmp 23bd6 <_sk_store_565_hsw_8bit+0xa6> DB 196,131,121,21,100,88,12,6 ; vpextrw $0x6,%xmm4,0xc(%r8,%r11,2) DB 196,131,121,21,100,88,10,5 ; vpextrw $0x5,%xmm4,0xa(%r8,%r11,2) DB 196,131,121,21,100,88,8,4 ; vpextrw $0x4,%xmm4,0x8(%r8,%r11,2) DB 196,129,121,214,36,88 ; vmovq %xmm4,(%r8,%r11,2) - DB 235,159 ; jmp 23a11 <_sk_store_565_hsw_8bit+0xa9> + DB 235,159 ; jmp 23bd6 <_sk_store_565_hsw_8bit+0xa6> DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,131,121,21,108,88,20,2 ; vpextrw $0x2,%xmm5,0x14(%r8,%r11,2) DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,131,121,21,108,88,18,1 ; vpextrw $0x1,%xmm5,0x12(%r8,%r11,2) DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,131,121,21,108,88,16,0 ; vpextrw $0x0,%xmm5,0x10(%r8,%r11,2) - DB 235,55 ; jmp 23ad5 <_sk_store_565_hsw_8bit+0x16d> + DB 235,55 ; jmp 23c9a <_sk_store_565_hsw_8bit+0x16a> DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,131,121,21,108,88,28,6 ; vpextrw $0x6,%xmm5,0x1c(%r8,%r11,2) DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 @@ -33995,39 +34014,39 @@ _sk_store_565_hsw_8bit LABEL PROC DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,129,121,214,108,88,16 ; vmovq %xmm5,0x10(%r8,%r11,2) DB 196,129,122,127,36,88 ; vmovdqu %xmm4,(%r8,%r11,2) - DB 233,49,255,255,255 ; jmpq 23a11 <_sk_store_565_hsw_8bit+0xa9> - DB 89 ; pop %rcx - DB 255 ; (bad) + DB 233,49,255,255,255 ; jmpq 23bd6 <_sk_store_565_hsw_8bit+0xa6> + DB 15,31,0 ; nopl (%rax) + DB 86 ; push %rsi DB 255 ; (bad) - DB 255,106,255 ; ljmp *-0x1(%rdx) DB 255 ; (bad) - DB 255,98,255 ; jmpq *-0x1(%rdx) + DB 255,103,255 ; jmpq *-0x1(%rdi) DB 255 ; (bad) - DB 255,138,255,255,255,130 ; decl -0x7d000001(%rdx) + DB 255,95,255 ; lcall *-0x1(%rdi) DB 255 ; (bad) + DB 255,135,255,255,255,127 ; incl 0x7fffffff(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 122,255 ; jp 23af5 <_sk_store_565_hsw_8bit+0x18d> + DB 255,119,255 ; pushq -0x1(%rdi) DB 255 ; (bad) - DB 255,114,255 ; pushq -0x1(%rdx) + DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) - DB 255,245 ; push %rbp + DB 255,242 ; push %rdx DB 255 ; (bad) DB 255 ; (bad) - DB 255,174,255,255,255,160 ; ljmp *-0x5f000001(%rsi) + DB 255,171,255,255,255,157 ; ljmp *-0x62000001(%rbx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,146,255,255,255,232 ; callq *-0x17000001(%rdx) + DB 255,143,255,255,255,229 ; decl -0x1a000001(%rdi) DB 255 ; (bad) DB 255 ; (bad) + DB 255,215 ; callq *%rdi DB 255 ; (bad) - DB 218,255 ; (bad) DB 255 ; (bad) - DB 255,204 ; dec %esp + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 190 ; .byte 0xbe + DB 187 ; .byte 0xbb DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -34042,41 +34061,41 @@ _sk_load_565_dst_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,222,0,0,0 ; jne 23c19 <_sk_load_565_dst_hsw_8bit+0xfd> + DB 15,133,222,0,0,0 ; jne 23de1 <_sk_load_565_dst_hsw_8bit+0xfd> DB 196,129,126,111,20,88 ; vmovdqu (%r8,%r11,2),%ymm2 DB 196,227,125,57,211,1 ; vextracti128 $0x1,%ymm2,%xmm3 DB 196,226,125,51,219 ; vpmovzxwd %xmm3,%ymm3 DB 196,226,125,51,210 ; vpmovzxwd %xmm2,%ymm2 - DB 196,226,125,88,37,202,151,0,0 ; vpbroadcastd 0x97ca(%rip),%ymm4 # 2d324 <_sk_overlay_sse2_8bit+0xdcc> + DB 196,226,125,88,37,226,148,0,0 ; vpbroadcastd 0x94e2(%rip),%ymm4 # 2d204 <_sk_overlay_sse2_8bit+0xd55> DB 197,237,219,236 ; vpand %ymm4,%ymm2,%ymm5 DB 197,229,219,228 ; vpand %ymm4,%ymm3,%ymm4 DB 197,221,114,212,8 ; vpsrld $0x8,%ymm4,%ymm4 DB 197,213,114,213,8 ; vpsrld $0x8,%ymm5,%ymm5 - DB 196,226,125,88,53,191,151,0,0 ; vpbroadcastd 0x97bf(%rip),%ymm6 # 2d334 <_sk_overlay_sse2_8bit+0xddc> + DB 196,226,125,88,53,215,148,0,0 ; vpbroadcastd 0x94d7(%rip),%ymm6 # 2d214 <_sk_overlay_sse2_8bit+0xd65> DB 197,237,219,254 ; vpand %ymm6,%ymm2,%ymm7 DB 197,229,219,246 ; vpand %ymm6,%ymm3,%ymm6 DB 197,205,114,246,19 ; vpslld $0x13,%ymm6,%ymm6 - DB 197,197,114,247,19 ; vpslld $0x13,%ymm7,%ymm7 - DB 197,213,235,239 ; vpor %ymm7,%ymm5,%ymm5 DB 197,221,235,230 ; vpor %ymm6,%ymm4,%ymm4 + DB 197,205,114,247,19 ; vpslld $0x13,%ymm7,%ymm6 + DB 197,213,235,238 ; vpor %ymm6,%ymm5,%ymm5 DB 197,205,114,212,5 ; vpsrld $0x5,%ymm4,%ymm6 DB 197,197,114,213,5 ; vpsrld $0x5,%ymm5,%ymm7 - DB 196,98,125,88,5,154,152,0,0 ; vpbroadcastd 0x989a(%rip),%ymm8 # 2d43c <_sk_overlay_sse2_8bit+0xee4> + DB 196,98,125,88,5,190,149,0,0 ; vpbroadcastd 0x95be(%rip),%ymm8 # 2d328 <_sk_overlay_sse2_8bit+0xe79> DB 196,193,69,219,248 ; vpand %ymm8,%ymm7,%ymm7 DB 196,193,77,219,240 ; vpand %ymm8,%ymm6,%ymm6 - DB 196,98,125,88,5,119,151,0,0 ; vpbroadcastd 0x9777(%rip),%ymm8 # 2d32c <_sk_overlay_sse2_8bit+0xdd4> + DB 196,98,125,88,5,143,148,0,0 ; vpbroadcastd 0x948f(%rip),%ymm8 # 2d20c <_sk_overlay_sse2_8bit+0xd5d> DB 196,65,101,219,200 ; vpand %ymm8,%ymm3,%ymm9 DB 196,65,109,219,192 ; vpand %ymm8,%ymm2,%ymm8 DB 196,193,61,114,240,5 ; vpslld $0x5,%ymm8,%ymm8 DB 196,193,53,114,241,5 ; vpslld $0x5,%ymm9,%ymm9 - DB 196,98,125,88,21,108,152,0,0 ; vpbroadcastd 0x986c(%rip),%ymm10 # 2d440 <_sk_overlay_sse2_8bit+0xee8> + DB 196,98,125,88,21,144,149,0,0 ; vpbroadcastd 0x9590(%rip),%ymm10 # 2d32c <_sk_overlay_sse2_8bit+0xe7d> DB 196,193,101,219,218 ; vpand %ymm10,%ymm3,%ymm3 DB 196,193,109,219,210 ; vpand %ymm10,%ymm2,%ymm2 DB 197,237,114,210,1 ; vpsrld $0x1,%ymm2,%ymm2 + DB 196,193,109,235,208 ; vpor %ymm8,%ymm2,%ymm2 DB 197,229,114,211,1 ; vpsrld $0x1,%ymm3,%ymm3 DB 196,193,101,235,217 ; vpor %ymm9,%ymm3,%ymm3 - DB 196,193,109,235,208 ; vpor %ymm8,%ymm2,%ymm2 - DB 196,98,125,88,5,73,152,0,0 ; vpbroadcastd 0x9849(%rip),%ymm8 # 2d444 <_sk_overlay_sse2_8bit+0xeec> + DB 196,98,125,88,5,109,149,0,0 ; vpbroadcastd 0x956d(%rip),%ymm8 # 2d330 <_sk_overlay_sse2_8bit+0xe81> DB 197,237,235,213 ; vpor %ymm5,%ymm2,%ymm2 DB 197,229,235,220 ; vpor %ymm4,%ymm3,%ymm3 DB 196,193,109,235,208 ; vpor %ymm8,%ymm2,%ymm2 @@ -34090,21 +34109,21 @@ _sk_load_565_dst_hsw_8bit LABEL PROC DB 197,237,239,210 ; vpxor %ymm2,%ymm2,%ymm2 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 15,135,16,255,255,255 ; ja 23b41 <_sk_load_565_dst_hsw_8bit+0x25> + DB 15,135,16,255,255,255 ; ja 23d09 <_sk_load_565_dst_hsw_8bit+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,28,1,0,0 ; lea 0x11c(%rip),%r10 # 23d58 <_sk_load_565_dst_hsw_8bit+0x23c> + DB 76,141,21,28,1,0,0 ; lea 0x11c(%rip),%r10 # 23f20 <_sk_load_565_dst_hsw_8bit+0x23c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 197,249,110,208 ; vmovd %eax,%xmm2 - DB 233,238,254,255,255 ; jmpq 23b41 <_sk_load_565_dst_hsw_8bit+0x25> + DB 233,238,254,255,255 ; jmpq 23d09 <_sk_load_565_dst_hsw_8bit+0x25> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,129,105,196,92,88,4,2 ; vpinsrw $0x2,0x4(%r8,%r11,2),%xmm2,%xmm3 DB 196,227,101,56,210,1 ; vinserti128 $0x1,%xmm2,%ymm3,%ymm2 DB 196,129,121,110,28,88 ; vmovd (%r8,%r11,2),%xmm3 DB 196,227,109,2,211,1 ; vpblendd $0x1,%ymm3,%ymm2,%ymm2 - DB 233,203,254,255,255 ; jmpq 23b41 <_sk_load_565_dst_hsw_8bit+0x25> + DB 233,203,254,255,255 ; jmpq 23d09 <_sk_load_565_dst_hsw_8bit+0x25> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,129,105,196,92,88,12,6 ; vpinsrw $0x6,0xc(%r8,%r11,2),%xmm2,%xmm3 DB 196,227,101,56,210,1 ; vinserti128 $0x1,%xmm2,%ymm3,%ymm2 @@ -34114,7 +34133,7 @@ _sk_load_565_dst_hsw_8bit LABEL PROC DB 196,227,109,2,211,15 ; vpblendd $0xf,%ymm3,%ymm2,%ymm2 DB 196,129,122,126,28,88 ; vmovq (%r8,%r11,2),%xmm3 DB 196,227,109,2,211,3 ; vpblendd $0x3,%ymm3,%ymm2,%ymm2 - DB 233,140,254,255,255 ; jmpq 23b41 <_sk_load_565_dst_hsw_8bit+0x25> + DB 233,140,254,255,255 ; jmpq 23d09 <_sk_load_565_dst_hsw_8bit+0x25> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,129,105,196,92,88,20,2 ; vpinsrw $0x2,0x14(%r8,%r11,2),%xmm2,%xmm3 DB 196,227,109,56,211,1 ; vinserti128 $0x1,%xmm3,%ymm2,%ymm2 @@ -34126,7 +34145,7 @@ _sk_load_565_dst_hsw_8bit LABEL PROC DB 196,227,109,56,211,1 ; vinserti128 $0x1,%xmm3,%ymm2,%ymm2 DB 196,129,122,111,28,88 ; vmovdqu (%r8,%r11,2),%xmm3 DB 196,227,101,2,210,240 ; vpblendd $0xf0,%ymm2,%ymm3,%ymm2 - DB 233,65,254,255,255 ; jmpq 23b41 <_sk_load_565_dst_hsw_8bit+0x25> + DB 233,65,254,255,255 ; jmpq 23d09 <_sk_load_565_dst_hsw_8bit+0x25> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,129,105,196,92,88,28,6 ; vpinsrw $0x6,0x1c(%r8,%r11,2),%xmm2,%xmm3 DB 196,227,109,56,211,1 ; vinserti128 $0x1,%xmm3,%ymm2,%ymm2 @@ -34140,11 +34159,11 @@ _sk_load_565_dst_hsw_8bit LABEL PROC DB 196,129,122,111,36,88 ; vmovdqu (%r8,%r11,2),%xmm4 DB 196,227,93,56,219,1 ; vinserti128 $0x1,%xmm3,%ymm4,%ymm3 DB 196,227,101,2,210,192 ; vpblendd $0xc0,%ymm2,%ymm3,%ymm2 - DB 233,233,253,255,255 ; jmpq 23b41 <_sk_load_565_dst_hsw_8bit+0x25> + DB 233,233,253,255,255 ; jmpq 23d09 <_sk_load_565_dst_hsw_8bit+0x25> DB 237 ; in (%dx),%eax DB 254 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,251 ; decl -0x4000001(%rip) # fffffffffc023d60 <_sk_overlay_sse2_8bit+0xfffffffffbff7808> + DB 255,13,255,255,255,251 ; decl -0x4000001(%rip) # fffffffffc023f28 <_sk_overlay_sse2_8bit+0xfffffffffbff7a79> DB 254 ; (bad) DB 255 ; (bad) DB 255,76,255,255 ; decl -0x1(%rdi,%rdi,8) @@ -34185,7 +34204,7 @@ _sk_load_a8_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,35 ; jne 23dcf <_sk_load_a8_hsw_8bit+0x3b> + DB 117,35 ; jne 23f97 <_sk_load_a8_hsw_8bit+0x3b> DB 196,129,122,111,4,24 ; vmovdqu (%r8,%r11,1),%xmm0 DB 197,249,112,200,78 ; vpshufd $0x4e,%xmm0,%xmm1 DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1 @@ -34199,35 +34218,35 @@ _sk_load_a8_hsw_8bit LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,207 ; ja 23db2 <_sk_load_a8_hsw_8bit+0x1e> + DB 119,207 ; ja 23f7a <_sk_load_a8_hsw_8bit+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,194,0,0,0 ; lea 0xc2(%rip),%r10 # 23eb0 <_sk_load_a8_hsw_8bit+0x11c> + DB 76,141,21,194,0,0,0 ; lea 0xc2(%rip),%r10 # 24078 <_sk_load_a8_hsw_8bit+0x11c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 235,176 ; jmp 23db2 <_sk_load_a8_hsw_8bit+0x1e> + DB 235,176 ; jmp 23f7a <_sk_load_a8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,2,2 ; vpinsrb $0x2,0x2(%r8,%r11,1),%xmm0,%xmm0 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,227,121,14,193,1 ; vpblendw $0x1,%xmm1,%xmm0,%xmm0 - DB 235,147 ; jmp 23db2 <_sk_load_a8_hsw_8bit+0x1e> + DB 235,147 ; jmp 23f7a <_sk_load_a8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,6,6 ; vpinsrb $0x6,0x6(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,5,5 ; vpinsrb $0x5,0x5(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,4,4 ; vpinsrb $0x4,0x4(%r8,%r11,1),%xmm0,%xmm0 DB 196,129,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm1 DB 196,227,121,2,193,1 ; vpblendd $0x1,%xmm1,%xmm0,%xmm0 - DB 233,102,255,255,255 ; jmpq 23db2 <_sk_load_a8_hsw_8bit+0x1e> + DB 233,102,255,255,255 ; jmpq 23f7a <_sk_load_a8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,10,10 ; vpinsrb $0xa,0xa(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,9,9 ; vpinsrb $0x9,0x9(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,8,8 ; vpinsrb $0x8,0x8(%r8,%r11,1),%xmm0,%xmm0 DB 196,129,122,126,12,24 ; vmovq (%r8,%r11,1),%xmm1 DB 196,227,113,2,192,12 ; vpblendd $0xc,%xmm0,%xmm1,%xmm0 - DB 233,57,255,255,255 ; jmpq 23db2 <_sk_load_a8_hsw_8bit+0x1e> + DB 233,57,255,255,255 ; jmpq 23f7a <_sk_load_a8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,14,14 ; vpinsrb $0xe,0xe(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,13,13 ; vpinsrb $0xd,0xd(%r8,%r11,1),%xmm0,%xmm0 @@ -34235,7 +34254,7 @@ _sk_load_a8_hsw_8bit LABEL PROC DB 196,129,122,126,12,24 ; vmovq (%r8,%r11,1),%xmm1 DB 196,131,113,34,76,24,8,2 ; vpinsrd $0x2,0x8(%r8,%r11,1),%xmm1,%xmm1 DB 196,227,113,2,192,8 ; vpblendd $0x8,%xmm0,%xmm1,%xmm0 - DB 233,4,255,255,255 ; jmpq 23db2 <_sk_load_a8_hsw_8bit+0x1e> + DB 233,4,255,255,255 ; jmpq 23f7a <_sk_load_a8_hsw_8bit+0x1e> DB 102,144 ; xchg %ax,%ax DB 71,255 ; rex.RXB (bad) DB 255 ; (bad) @@ -34247,7 +34266,7 @@ _sk_load_a8_hsw_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 123,255 ; jnp 23ec5 <_sk_load_a8_hsw_8bit+0x131> + DB 123,255 ; jnp 2408d <_sk_load_a8_hsw_8bit+0x131> DB 255 ; (bad) DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) @@ -34281,7 +34300,7 @@ _sk_load_a8_dst_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,35 ; jne 23f27 <_sk_load_a8_dst_hsw_8bit+0x3b> + DB 117,35 ; jne 240ef <_sk_load_a8_dst_hsw_8bit+0x3b> DB 196,129,122,111,20,24 ; vmovdqu (%r8,%r11,1),%xmm2 DB 197,249,112,218,78 ; vpshufd $0x4e,%xmm2,%xmm3 DB 196,226,125,49,219 ; vpmovzxbd %xmm3,%ymm3 @@ -34295,35 +34314,35 @@ _sk_load_a8_dst_hsw_8bit LABEL PROC DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,207 ; ja 23f0a <_sk_load_a8_dst_hsw_8bit+0x1e> + DB 119,207 ; ja 240d2 <_sk_load_a8_dst_hsw_8bit+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,194,0,0,0 ; lea 0xc2(%rip),%r10 # 24008 <_sk_load_a8_dst_hsw_8bit+0x11c> + DB 76,141,21,194,0,0,0 ; lea 0xc2(%rip),%r10 # 241d0 <_sk_load_a8_dst_hsw_8bit+0x11c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,208 ; vmovd %eax,%xmm2 - DB 235,176 ; jmp 23f0a <_sk_load_a8_dst_hsw_8bit+0x1e> + DB 235,176 ; jmp 240d2 <_sk_load_a8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,2,2 ; vpinsrb $0x2,0x2(%r8,%r11,1),%xmm2,%xmm2 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax DB 197,249,110,216 ; vmovd %eax,%xmm3 DB 196,227,105,14,211,1 ; vpblendw $0x1,%xmm3,%xmm2,%xmm2 - DB 235,147 ; jmp 23f0a <_sk_load_a8_dst_hsw_8bit+0x1e> + DB 235,147 ; jmp 240d2 <_sk_load_a8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,6,6 ; vpinsrb $0x6,0x6(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,5,5 ; vpinsrb $0x5,0x5(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,4,4 ; vpinsrb $0x4,0x4(%r8,%r11,1),%xmm2,%xmm2 DB 196,129,121,110,28,24 ; vmovd (%r8,%r11,1),%xmm3 DB 196,227,105,2,211,1 ; vpblendd $0x1,%xmm3,%xmm2,%xmm2 - DB 233,102,255,255,255 ; jmpq 23f0a <_sk_load_a8_dst_hsw_8bit+0x1e> + DB 233,102,255,255,255 ; jmpq 240d2 <_sk_load_a8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,10,10 ; vpinsrb $0xa,0xa(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,9,9 ; vpinsrb $0x9,0x9(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,8,8 ; vpinsrb $0x8,0x8(%r8,%r11,1),%xmm2,%xmm2 DB 196,129,122,126,28,24 ; vmovq (%r8,%r11,1),%xmm3 DB 196,227,97,2,210,12 ; vpblendd $0xc,%xmm2,%xmm3,%xmm2 - DB 233,57,255,255,255 ; jmpq 23f0a <_sk_load_a8_dst_hsw_8bit+0x1e> + DB 233,57,255,255,255 ; jmpq 240d2 <_sk_load_a8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,14,14 ; vpinsrb $0xe,0xe(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,13,13 ; vpinsrb $0xd,0xd(%r8,%r11,1),%xmm2,%xmm2 @@ -34331,7 +34350,7 @@ _sk_load_a8_dst_hsw_8bit LABEL PROC DB 196,129,122,126,28,24 ; vmovq (%r8,%r11,1),%xmm3 DB 196,131,97,34,92,24,8,2 ; vpinsrd $0x2,0x8(%r8,%r11,1),%xmm3,%xmm3 DB 196,227,97,2,210,8 ; vpblendd $0x8,%xmm2,%xmm3,%xmm2 - DB 233,4,255,255,255 ; jmpq 23f0a <_sk_load_a8_dst_hsw_8bit+0x1e> + DB 233,4,255,255,255 ; jmpq 240d2 <_sk_load_a8_dst_hsw_8bit+0x1e> DB 102,144 ; xchg %ax,%ax DB 71,255 ; rex.RXB (bad) DB 255 ; (bad) @@ -34343,7 +34362,7 @@ _sk_load_a8_dst_hsw_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 123,255 ; jnp 2401d <_sk_load_a8_dst_hsw_8bit+0x131> + DB 123,255 ; jnp 241e5 <_sk_load_a8_dst_hsw_8bit+0x131> DB 255 ; (bad) DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) @@ -34376,17 +34395,17 @@ _sk_store_a8_hsw_8bit LABEL PROC DB 77,15,175,193 ; imul %r9,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 197,253,111,37,65,143,0,0 ; vmovdqa 0x8f41(%rip),%ymm4 # 2cfa0 <_sk_overlay_sse2_8bit+0xa48> + DB 197,253,111,37,25,140,0,0 ; vmovdqa 0x8c19(%rip),%ymm4 # 2ce40 <_sk_overlay_sse2_8bit+0x991> DB 196,226,117,0,236 ; vpshufb %ymm4,%ymm1,%ymm5 DB 196,227,253,0,237,232 ; vpermq $0xe8,%ymm5,%ymm5 - DB 197,249,111,53,14,155,0,0 ; vmovdqa 0x9b0e(%rip),%xmm6 # 2db80 <_sk_overlay_sse2_8bit+0x1628> + DB 197,249,111,53,86,152,0,0 ; vmovdqa 0x9856(%rip),%xmm6 # 2da90 <_sk_overlay_sse2_8bit+0x15e1> DB 196,226,81,0,238 ; vpshufb %xmm6,%xmm5,%xmm5 DB 196,226,125,0,228 ; vpshufb %ymm4,%ymm0,%ymm4 DB 196,227,253,0,228,232 ; vpermq $0xe8,%ymm4,%ymm4 DB 196,226,89,0,230 ; vpshufb %xmm6,%xmm4,%xmm4 DB 197,217,108,229 ; vpunpcklqdq %xmm5,%xmm4,%xmm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 2409a <_sk_store_a8_hsw_8bit+0x56> + DB 117,10 ; jne 24262 <_sk_store_a8_hsw_8bit+0x56> DB 196,129,122,127,36,24 ; vmovdqu %xmm4,(%r8,%r11,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -34394,34 +34413,34 @@ _sk_store_a8_hsw_8bit LABEL PROC DB 65,128,225,15 ; and $0xf,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,236 ; ja 24096 <_sk_store_a8_hsw_8bit+0x52> + DB 119,236 ; ja 2425e <_sk_store_a8_hsw_8bit+0x52> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,139,0,0,0 ; lea 0x8b(%rip),%r10 # 24140 <_sk_store_a8_hsw_8bit+0xfc> + DB 76,141,21,139,0,0,0 ; lea 0x8b(%rip),%r10 # 24308 <_sk_store_a8_hsw_8bit+0xfc> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 196,131,121,20,36,24,0 ; vpextrb $0x0,%xmm4,(%r8,%r11,1) - DB 235,207 ; jmp 24096 <_sk_store_a8_hsw_8bit+0x52> + DB 235,207 ; jmp 2425e <_sk_store_a8_hsw_8bit+0x52> DB 196,131,121,20,100,24,2,2 ; vpextrb $0x2,%xmm4,0x2(%r8,%r11,1) DB 196,131,121,21,36,24,0 ; vpextrw $0x0,%xmm4,(%r8,%r11,1) - DB 235,190 ; jmp 24096 <_sk_store_a8_hsw_8bit+0x52> + DB 235,190 ; jmp 2425e <_sk_store_a8_hsw_8bit+0x52> DB 196,131,121,20,100,24,6,6 ; vpextrb $0x6,%xmm4,0x6(%r8,%r11,1) DB 196,131,121,20,100,24,5,5 ; vpextrb $0x5,%xmm4,0x5(%r8,%r11,1) DB 196,131,121,20,100,24,4,4 ; vpextrb $0x4,%xmm4,0x4(%r8,%r11,1) DB 196,129,121,126,36,24 ; vmovd %xmm4,(%r8,%r11,1) - DB 235,158 ; jmp 24096 <_sk_store_a8_hsw_8bit+0x52> + DB 235,158 ; jmp 2425e <_sk_store_a8_hsw_8bit+0x52> DB 196,131,121,20,100,24,10,10 ; vpextrb $0xa,%xmm4,0xa(%r8,%r11,1) DB 196,131,121,20,100,24,9,9 ; vpextrb $0x9,%xmm4,0x9(%r8,%r11,1) DB 196,131,121,20,100,24,8,8 ; vpextrb $0x8,%xmm4,0x8(%r8,%r11,1) - DB 235,32 ; jmp 24132 <_sk_store_a8_hsw_8bit+0xee> + DB 235,32 ; jmp 242fa <_sk_store_a8_hsw_8bit+0xee> DB 196,131,121,20,100,24,14,14 ; vpextrb $0xe,%xmm4,0xe(%r8,%r11,1) DB 196,131,121,20,100,24,13,13 ; vpextrb $0xd,%xmm4,0xd(%r8,%r11,1) DB 196,131,121,20,100,24,12,12 ; vpextrb $0xc,%xmm4,0xc(%r8,%r11,1) DB 196,131,121,22,100,24,8,2 ; vpextrd $0x2,%xmm4,0x8(%r8,%r11,1) DB 196,129,121,214,36,24 ; vmovq %xmm4,(%r8,%r11,1) - DB 233,89,255,255,255 ; jmpq 24096 <_sk_store_a8_hsw_8bit+0x52> + DB 233,89,255,255,255 ; jmpq 2425e <_sk_store_a8_hsw_8bit+0x52> DB 15,31,0 ; nopl (%rax) - DB 126,255 ; jle 24141 <_sk_store_a8_hsw_8bit+0xfd> + DB 126,255 ; jle 24309 <_sk_store_a8_hsw_8bit+0xfd> DB 255 ; (bad) DB 255,143,255,255,255,135 ; decl -0x78000001(%rdi) DB 255 ; (bad) @@ -34465,15 +34484,15 @@ _sk_load_g8_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,61 ; jne 241d1 <_sk_load_g8_hsw_8bit+0x55> + DB 117,60 ; jne 24398 <_sk_load_g8_hsw_8bit+0x54> DB 196,129,122,111,4,24 ; vmovdqu (%r8,%r11,1),%xmm0 DB 196,226,125,49,200 ; vpmovzxbd %xmm0,%ymm1 DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 - DB 196,226,125,88,37,150,146,0,0 ; vpbroadcastd 0x9296(%rip),%ymm4 # 2d448 <_sk_overlay_sse2_8bit+0xef0> + DB 196,226,125,88,37,186,143,0,0 ; vpbroadcastd 0x8fba(%rip),%ymm4 # 2d334 <_sk_overlay_sse2_8bit+0xe85> DB 196,226,125,64,236 ; vpmulld %ymm4,%ymm0,%ymm5 DB 196,226,117,64,196 ; vpmulld %ymm4,%ymm1,%ymm0 - DB 196,226,125,88,13,127,146,0,0 ; vpbroadcastd 0x927f(%rip),%ymm1 # 2d444 <_sk_overlay_sse2_8bit+0xeec> + DB 197,253,111,13,212,138,0,0 ; vmovdqa 0x8ad4(%rip),%ymm1 # 2ce60 <_sk_overlay_sse2_8bit+0x9b1> DB 197,253,235,193 ; vpor %ymm1,%ymm0,%ymm0 DB 197,213,235,201 ; vpor %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -34483,35 +34502,35 @@ _sk_load_g8_hsw_8bit LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,181 ; ja 2419a <_sk_load_g8_hsw_8bit+0x1e> + DB 119,182 ; ja 24362 <_sk_load_g8_hsw_8bit+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,196,0,0,0 ; lea 0xc4(%rip),%r10 # 242b4 <_sk_load_g8_hsw_8bit+0x138> + DB 76,141,21,197,0,0,0 ; lea 0xc5(%rip),%r10 # 2447c <_sk_load_g8_hsw_8bit+0x138> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 - DB 235,150 ; jmp 2419a <_sk_load_g8_hsw_8bit+0x1e> + DB 235,151 ; jmp 24362 <_sk_load_g8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,2,2 ; vpinsrb $0x2,0x2(%r8,%r11,1),%xmm0,%xmm0 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax DB 197,249,110,200 ; vmovd %eax,%xmm1 DB 196,227,121,14,193,1 ; vpblendw $0x1,%xmm1,%xmm0,%xmm0 - DB 233,118,255,255,255 ; jmpq 2419a <_sk_load_g8_hsw_8bit+0x1e> + DB 233,119,255,255,255 ; jmpq 24362 <_sk_load_g8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,6,6 ; vpinsrb $0x6,0x6(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,5,5 ; vpinsrb $0x5,0x5(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,4,4 ; vpinsrb $0x4,0x4(%r8,%r11,1),%xmm0,%xmm0 DB 196,129,121,110,12,24 ; vmovd (%r8,%r11,1),%xmm1 DB 196,227,121,2,193,1 ; vpblendd $0x1,%xmm1,%xmm0,%xmm0 - DB 233,73,255,255,255 ; jmpq 2419a <_sk_load_g8_hsw_8bit+0x1e> + DB 233,74,255,255,255 ; jmpq 24362 <_sk_load_g8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,10,10 ; vpinsrb $0xa,0xa(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,9,9 ; vpinsrb $0x9,0x9(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,8,8 ; vpinsrb $0x8,0x8(%r8,%r11,1),%xmm0,%xmm0 DB 196,129,122,126,12,24 ; vmovq (%r8,%r11,1),%xmm1 DB 196,227,113,2,192,12 ; vpblendd $0xc,%xmm0,%xmm1,%xmm0 - DB 233,28,255,255,255 ; jmpq 2419a <_sk_load_g8_hsw_8bit+0x1e> + DB 233,29,255,255,255 ; jmpq 24362 <_sk_load_g8_hsw_8bit+0x1e> DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 196,131,121,32,68,24,14,14 ; vpinsrb $0xe,0xe(%r8,%r11,1),%xmm0,%xmm0 DB 196,131,121,32,68,24,13,13 ; vpinsrb $0xd,0xd(%r8,%r11,1),%xmm0,%xmm0 @@ -34519,37 +34538,39 @@ _sk_load_g8_hsw_8bit LABEL PROC DB 196,129,122,126,12,24 ; vmovq (%r8,%r11,1),%xmm1 DB 196,131,113,34,76,24,8,2 ; vpinsrd $0x2,0x8(%r8,%r11,1),%xmm1,%xmm1 DB 196,227,113,2,192,8 ; vpblendd $0x8,%xmm0,%xmm1,%xmm0 - DB 233,231,254,255,255 ; jmpq 2419a <_sk_load_g8_hsw_8bit+0x1e> - DB 144 ; nop - DB 69,255 ; rex.RB (bad) + DB 233,232,254,255,255 ; jmpq 24362 <_sk_load_g8_hsw_8bit+0x1e> + DB 102,144 ; xchg %ax,%ax + DB 68,255 ; rex.R (bad) DB 255 ; (bad) - DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) - DB 255,80,255 ; callq *-0x1(%rax) + DB 255,91,255 ; lcall *-0x1(%rbx) DB 255 ; (bad) - DB 255,140,255,255,255,132,255 ; decl -0x7b0001(%rdi,%rdi,8) + DB 255,79,255 ; decl -0x1(%rdi) DB 255 ; (bad) + DB 255,139,255,255,255,131 ; decl -0x7c000001(%rbx) DB 255 ; (bad) - DB 124,255 ; jl 242c9 <_sk_load_g8_hsw_8bit+0x14d> DB 255 ; (bad) - DB 255,112,255 ; pushq -0x1(%rax) DB 255 ; (bad) + DB 123,255 ; jnp 24491 <_sk_load_g8_hsw_8bit+0x14d> DB 255 ; (bad) - DB 185,255,255,255,177 ; mov $0xb1ffffff,%ecx + DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,169,255,255,255,157 ; ljmp *-0x62000001(%rcx) + DB 184,255,255,255,176 ; mov $0xb0ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,168,255,255,255,156 ; ljmp *-0x63000001(%rax) DB 255 ; (bad) DB 255 ; (bad) + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,214 ; callq *%rsi DB 255 ; (bad) + DB 221,255 ; (bad) DB 255 ; (bad) - DB 255,202 ; dec %edx + DB 255,213 ; callq *%rbp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -34563,15 +34584,15 @@ _sk_load_g8_dst_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,61 ; jne 24345 <_sk_load_g8_dst_hsw_8bit+0x55> + DB 117,60 ; jne 2450c <_sk_load_g8_dst_hsw_8bit+0x54> DB 196,129,122,111,20,24 ; vmovdqu (%r8,%r11,1),%xmm2 DB 196,226,125,49,218 ; vpmovzxbd %xmm2,%ymm3 DB 197,249,112,210,78 ; vpshufd $0x4e,%xmm2,%xmm2 DB 196,226,125,49,210 ; vpmovzxbd %xmm2,%ymm2 - DB 196,226,125,88,37,34,145,0,0 ; vpbroadcastd 0x9122(%rip),%ymm4 # 2d448 <_sk_overlay_sse2_8bit+0xef0> + DB 196,226,125,88,37,70,142,0,0 ; vpbroadcastd 0x8e46(%rip),%ymm4 # 2d334 <_sk_overlay_sse2_8bit+0xe85> DB 196,226,109,64,236 ; vpmulld %ymm4,%ymm2,%ymm5 DB 196,226,101,64,212 ; vpmulld %ymm4,%ymm3,%ymm2 - DB 196,226,125,88,29,11,145,0,0 ; vpbroadcastd 0x910b(%rip),%ymm3 # 2d444 <_sk_overlay_sse2_8bit+0xeec> + DB 197,253,111,29,128,137,0,0 ; vmovdqa 0x8980(%rip),%ymm3 # 2ce80 <_sk_overlay_sse2_8bit+0x9d1> DB 197,237,235,211 ; vpor %ymm3,%ymm2,%ymm2 DB 197,213,235,219 ; vpor %ymm3,%ymm5,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -34581,35 +34602,35 @@ _sk_load_g8_dst_hsw_8bit LABEL PROC DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,181 ; ja 2430e <_sk_load_g8_dst_hsw_8bit+0x1e> + DB 119,182 ; ja 244d6 <_sk_load_g8_dst_hsw_8bit+0x1e> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,196,0,0,0 ; lea 0xc4(%rip),%r10 # 24428 <_sk_load_g8_dst_hsw_8bit+0x138> + DB 76,141,21,197,0,0,0 ; lea 0xc5(%rip),%r10 # 245f0 <_sk_load_g8_dst_hsw_8bit+0x138> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,208 ; vmovd %eax,%xmm2 - DB 235,150 ; jmp 2430e <_sk_load_g8_dst_hsw_8bit+0x1e> + DB 235,151 ; jmp 244d6 <_sk_load_g8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,2,2 ; vpinsrb $0x2,0x2(%r8,%r11,1),%xmm2,%xmm2 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax DB 197,249,110,216 ; vmovd %eax,%xmm3 DB 196,227,105,14,211,1 ; vpblendw $0x1,%xmm3,%xmm2,%xmm2 - DB 233,118,255,255,255 ; jmpq 2430e <_sk_load_g8_dst_hsw_8bit+0x1e> + DB 233,119,255,255,255 ; jmpq 244d6 <_sk_load_g8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,6,6 ; vpinsrb $0x6,0x6(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,5,5 ; vpinsrb $0x5,0x5(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,4,4 ; vpinsrb $0x4,0x4(%r8,%r11,1),%xmm2,%xmm2 DB 196,129,121,110,28,24 ; vmovd (%r8,%r11,1),%xmm3 DB 196,227,105,2,211,1 ; vpblendd $0x1,%xmm3,%xmm2,%xmm2 - DB 233,73,255,255,255 ; jmpq 2430e <_sk_load_g8_dst_hsw_8bit+0x1e> + DB 233,74,255,255,255 ; jmpq 244d6 <_sk_load_g8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,10,10 ; vpinsrb $0xa,0xa(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,9,9 ; vpinsrb $0x9,0x9(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,8,8 ; vpinsrb $0x8,0x8(%r8,%r11,1),%xmm2,%xmm2 DB 196,129,122,126,28,24 ; vmovq (%r8,%r11,1),%xmm3 DB 196,227,97,2,210,12 ; vpblendd $0xc,%xmm2,%xmm3,%xmm2 - DB 233,28,255,255,255 ; jmpq 2430e <_sk_load_g8_dst_hsw_8bit+0x1e> + DB 233,29,255,255,255 ; jmpq 244d6 <_sk_load_g8_dst_hsw_8bit+0x1e> DB 197,233,239,210 ; vpxor %xmm2,%xmm2,%xmm2 DB 196,131,105,32,84,24,14,14 ; vpinsrb $0xe,0xe(%r8,%r11,1),%xmm2,%xmm2 DB 196,131,105,32,84,24,13,13 ; vpinsrb $0xd,0xd(%r8,%r11,1),%xmm2,%xmm2 @@ -34617,37 +34638,39 @@ _sk_load_g8_dst_hsw_8bit LABEL PROC DB 196,129,122,126,28,24 ; vmovq (%r8,%r11,1),%xmm3 DB 196,131,97,34,92,24,8,2 ; vpinsrd $0x2,0x8(%r8,%r11,1),%xmm3,%xmm3 DB 196,227,97,2,210,8 ; vpblendd $0x8,%xmm2,%xmm3,%xmm2 - DB 233,231,254,255,255 ; jmpq 2430e <_sk_load_g8_dst_hsw_8bit+0x1e> - DB 144 ; nop - DB 69,255 ; rex.RB (bad) + DB 233,232,254,255,255 ; jmpq 244d6 <_sk_load_g8_dst_hsw_8bit+0x1e> + DB 102,144 ; xchg %ax,%ax + DB 68,255 ; rex.R (bad) DB 255 ; (bad) - DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) - DB 255,80,255 ; callq *-0x1(%rax) + DB 255,91,255 ; lcall *-0x1(%rbx) + DB 255 ; (bad) + DB 255,79,255 ; decl -0x1(%rdi) + DB 255 ; (bad) + DB 255,139,255,255,255,131 ; decl -0x7c000001(%rbx) DB 255 ; (bad) - DB 255,140,255,255,255,132,255 ; decl -0x7b0001(%rdi,%rdi,8) DB 255 ; (bad) DB 255 ; (bad) - DB 124,255 ; jl 2443d <_sk_load_g8_dst_hsw_8bit+0x14d> + DB 123,255 ; jnp 24605 <_sk_load_g8_dst_hsw_8bit+0x14d> DB 255 ; (bad) - DB 255,112,255 ; pushq -0x1(%rax) + DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 185,255,255,255,177 ; mov $0xb1ffffff,%ecx + DB 184,255,255,255,176 ; mov $0xb0ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,169,255,255,255,157 ; ljmp *-0x62000001(%rcx) + DB 255,168,255,255,255,156 ; ljmp *-0x63000001(%rax) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) + DB 221,255 ; (bad) DB 255 ; (bad) - DB 255,214 ; callq *%rsi + DB 255,213 ; callq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 255,202 ; dec %edx + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -34655,218 +34678,215 @@ _sk_load_g8_dst_hsw_8bit LABEL PROC PUBLIC _sk_srcover_rgba_8888_hsw_8bit _sk_srcover_rgba_8888_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,99,72,8 ; movslq 0x8(%rax),%r9 - DB 76,99,193 ; movslq %ecx,%r8 - DB 77,15,175,193 ; imul %r9,%r8 - DB 73,193,224,2 ; shl $0x2,%r8 - DB 76,3,0 ; add (%rax),%r8 - DB 76,99,218 ; movslq %edx,%r11 + DB 76,99,64,8 ; movslq 0x8(%rax),%r8 + DB 76,99,201 ; movslq %ecx,%r9 + DB 77,15,175,200 ; imul %r8,%r9 + DB 73,193,225,2 ; shl $0x2,%r9 + DB 76,3,8 ; add (%rax),%r9 + DB 72,99,194 ; movslq %edx,%rax + DB 77,141,20,129 ; lea (%r9,%rax,4),%r10 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,222,0,0,0 ; jne 24562 <_sk_srcover_rgba_8888_hsw_8bit+0xfe> - DB 196,1,126,111,76,152,32 ; vmovdqu 0x20(%r8,%r11,4),%ymm9 - DB 196,1,126,111,28,152 ; vmovdqu (%r8,%r11,4),%ymm11 - DB 197,253,111,53,39,139,0,0 ; vmovdqa 0x8b27(%rip),%ymm6 # 2cfc0 <_sk_overlay_sse2_8bit+0xa68> + DB 15,133,218,0,0,0 ; jne 2472a <_sk_srcover_rgba_8888_hsw_8bit+0xfe> + DB 196,65,126,111,18 ; vmovdqu (%r10),%ymm10 + DB 196,65,126,111,66,32 ; vmovdqu 0x20(%r10),%ymm8 + DB 197,253,111,53,61,136,0,0 ; vmovdqa 0x883d(%rip),%ymm6 # 2cea0 <_sk_overlay_sse2_8bit+0x9f1> DB 196,226,117,0,254 ; vpshufb %ymm6,%ymm1,%ymm7 DB 196,226,125,0,246 ; vpshufb %ymm6,%ymm0,%ymm6 - DB 196,66,125,48,195 ; vpmovzxbw %xmm11,%ymm8 - DB 196,99,125,57,220,1 ; vextracti128 $0x1,%ymm11,%xmm4 - DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,66,125,48,209 ; vpmovzxbw %xmm9,%ymm10 - DB 196,99,125,57,205,1 ; vextracti128 $0x1,%ymm9,%xmm5 + DB 196,99,125,57,212,1 ; vextracti128 $0x1,%ymm10,%xmm4 + DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 + DB 196,66,125,48,202 ; vpmovzxbw %xmm10,%ymm9 + DB 196,99,125,57,197,1 ; vextracti128 $0x1,%ymm8,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,230 ; vpmovzxbw %xmm6,%ymm12 - DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 + DB 196,66,125,48,216 ; vpmovzxbw %xmm8,%ymm11 + DB 196,227,125,57,244,1 ; vextracti128 $0x1,%ymm6,%xmm4 + DB 196,98,125,48,236 ; vpmovzxbw %xmm4,%ymm13 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,239 ; vpmovzxbw %xmm7,%ymm13 - DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 + DB 196,227,125,57,252,1 ; vextracti128 $0x1,%ymm7,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 197,213,213,255 ; vpmullw %ymm7,%ymm5,%ymm7 - DB 196,65,45,213,237 ; vpmullw %ymm13,%ymm10,%ymm13 - DB 197,221,213,246 ; vpmullw %ymm6,%ymm4,%ymm6 - DB 196,65,61,213,228 ; vpmullw %ymm12,%ymm8,%ymm12 - DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 - DB 197,205,253,228 ; vpaddw %ymm4,%ymm6,%ymm4 - DB 196,193,21,253,242 ; vpaddw %ymm10,%ymm13,%ymm6 - DB 197,197,253,237 ; vpaddw %ymm5,%ymm7,%ymm5 + DB 197,165,213,255 ; vpmullw %ymm7,%ymm11,%ymm7 + DB 197,213,213,228 ; vpmullw %ymm4,%ymm5,%ymm4 + DB 197,181,213,246 ; vpmullw %ymm6,%ymm9,%ymm6 + DB 196,65,29,213,237 ; vpmullw %ymm13,%ymm12,%ymm13 + DB 196,65,21,253,228 ; vpaddw %ymm12,%ymm13,%ymm12 + DB 196,193,77,253,241 ; vpaddw %ymm9,%ymm6,%ymm6 + DB 197,221,253,229 ; vpaddw %ymm5,%ymm4,%ymm4 + DB 196,193,69,253,235 ; vpaddw %ymm11,%ymm7,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 - DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 - DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 - DB 197,181,248,237 ; vpsubb %ymm5,%ymm9,%ymm5 - DB 197,165,248,228 ; vpsubb %ymm4,%ymm11,%ymm4 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 196,193,69,113,212,8 ; vpsrlw $0x8,%ymm12,%ymm7 + DB 196,99,77,56,207,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm9 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,181,103,246 ; vpackuswb %ymm6,%ymm9,%ymm6 + DB 196,227,85,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm7 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 + DB 197,189,248,236 ; vpsubb %ymm4,%ymm8,%ymm5 + DB 197,173,248,230 ; vpsubb %ymm6,%ymm10,%ymm4 DB 197,221,252,224 ; vpaddb %ymm0,%ymm4,%ymm4 DB 197,213,252,233 ; vpaddb %ymm1,%ymm5,%ymm5 DB 72,133,255 ; test %rdi,%rdi - DB 117,78 ; jne 2459f <_sk_srcover_rgba_8888_hsw_8bit+0x13b> - DB 196,129,126,127,36,152 ; vmovdqu %ymm4,(%r8,%r11,4) - DB 196,129,126,127,108,152,32 ; vmovdqu %ymm5,0x20(%r8,%r11,4) - DB 72,173 ; lods %ds:(%rsi),%rax - DB 255,224 ; jmpq *%rax - DB 65,137,249 ; mov %edi,%r9d - DB 65,128,225,15 ; and $0xf,%r9b - DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9 - DB 196,65,37,239,219 ; vpxor %ymm11,%ymm11,%ymm11 - DB 65,254,201 ; dec %r9b - DB 65,128,249,14 ; cmp $0xe,%r9b - DB 15,135,17,255,255,255 ; ja 24491 <_sk_srcover_rgba_8888_hsw_8bit+0x2d> - DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,237,1,0,0 ; lea 0x1ed(%rip),%r10 # 24778 <_sk_srcover_rgba_8888_hsw_8bit+0x314> - DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax - DB 76,1,208 ; add %r10,%rax - DB 255,224 ; jmpq *%rax - DB 196,1,121,110,28,152 ; vmovd (%r8,%r11,4),%xmm11 - DB 233,242,254,255,255 ; jmpq 24491 <_sk_srcover_rgba_8888_hsw_8bit+0x2d> - DB 65,137,249 ; mov %edi,%r9d - DB 65,128,225,15 ; and $0xf,%r9b - DB 65,254,201 ; dec %r9b - DB 65,128,249,14 ; cmp $0xe,%r9b - DB 119,175 ; ja 2455e <_sk_srcover_rgba_8888_hsw_8bit+0xfa> - DB 65,15,182,193 ; movzbl %r9b,%eax - DB 76,141,13,250,1,0,0 ; lea 0x1fa(%rip),%r9 # 247b4 <_sk_srcover_rgba_8888_hsw_8bit+0x350> - DB 73,99,4,129 ; movslq (%r9,%rax,4),%rax + DB 117,75 ; jne 24766 <_sk_srcover_rgba_8888_hsw_8bit+0x13a> + DB 196,193,126,127,34 ; vmovdqu %ymm4,(%r10) + DB 196,193,126,127,106,32 ; vmovdqu %ymm5,0x20(%r10) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 65,137,248 ; mov %edi,%r8d + DB 65,128,224,15 ; and $0xf,%r8b + DB 196,65,61,239,192 ; vpxor %ymm8,%ymm8,%ymm8 + DB 196,65,45,239,210 ; vpxor %ymm10,%ymm10,%ymm10 + DB 65,254,200 ; dec %r8b + DB 65,128,248,14 ; cmp $0xe,%r8b + DB 15,135,19,255,255,255 ; ja 2465b <_sk_srcover_rgba_8888_hsw_8bit+0x2f> + DB 69,15,182,192 ; movzbl %r8b,%r8d + DB 76,141,13,213,1,0,0 ; lea 0x1d5(%rip),%r9 # 24928 <_sk_srcover_rgba_8888_hsw_8bit+0x2fc> + DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax - DB 196,129,121,126,36,152 ; vmovd %xmm4,(%r8,%r11,4) - DB 235,147 ; jmp 2455e <_sk_srcover_rgba_8888_hsw_8bit+0xfa> - DB 196,129,121,110,100,152,8 ; vmovd 0x8(%r8,%r11,4),%xmm4 + DB 196,65,121,110,18 ; vmovd (%r10),%xmm10 + DB 233,245,254,255,255 ; jmpq 2465b <_sk_srcover_rgba_8888_hsw_8bit+0x2f> + DB 65,137,248 ; mov %edi,%r8d + DB 65,128,224,15 ; and $0xf,%r8b + DB 65,254,200 ; dec %r8b + DB 65,128,248,14 ; cmp $0xe,%r8b + DB 119,176 ; ja 24726 <_sk_srcover_rgba_8888_hsw_8bit+0xfa> + DB 65,15,182,192 ; movzbl %r8b,%eax + DB 76,141,5,227,1,0,0 ; lea 0x1e3(%rip),%r8 # 24964 <_sk_srcover_rgba_8888_hsw_8bit+0x338> + DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax + DB 76,1,192 ; add %r8,%rax + DB 255,224 ; jmpq *%rax + DB 196,193,121,126,34 ; vmovd %xmm4,(%r10) + DB 235,149 ; jmp 24726 <_sk_srcover_rgba_8888_hsw_8bit+0xfa> + DB 196,193,121,110,98,8 ; vmovd 0x8(%r10),%xmm4 DB 196,226,121,89,236 ; vpbroadcastq %xmm4,%xmm5 - DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9 - DB 196,99,53,2,221,4 ; vpblendd $0x4,%ymm5,%ymm9,%ymm11 - DB 196,130,121,53,52,152 ; vpmovzxdq (%r8,%r11,4),%xmm6 + DB 196,65,61,239,192 ; vpxor %ymm8,%ymm8,%ymm8 + DB 196,99,61,2,213,4 ; vpblendd $0x4,%ymm5,%ymm8,%ymm10 + DB 196,194,121,53,50 ; vpmovzxdq (%r10),%xmm6 DB 197,249,112,246,232 ; vpshufd $0xe8,%xmm6,%xmm6 - DB 196,99,37,2,222,3 ; vpblendd $0x3,%ymm6,%ymm11,%ymm11 - DB 233,153,254,255,255 ; jmpq 24491 <_sk_srcover_rgba_8888_hsw_8bit+0x2d> - DB 196,129,121,110,100,152,24 ; vmovd 0x18(%r8,%r11,4),%xmm4 + DB 196,99,45,2,214,3 ; vpblendd $0x3,%ymm6,%ymm10,%ymm10 + DB 233,159,254,255,255 ; jmpq 2465b <_sk_srcover_rgba_8888_hsw_8bit+0x2f> + DB 196,193,121,110,98,24 ; vmovd 0x18(%r10),%xmm4 DB 196,226,125,89,236 ; vpbroadcastq %xmm4,%ymm5 - DB 196,65,53,239,201 ; vpxor %ymm9,%ymm9,%ymm9 - DB 196,99,53,2,221,64 ; vpblendd $0x40,%ymm5,%ymm9,%ymm11 - DB 196,99,125,57,222,1 ; vextracti128 $0x1,%ymm11,%xmm6 - DB 196,131,73,34,116,152,20,1 ; vpinsrd $0x1,0x14(%r8,%r11,4),%xmm6,%xmm6 - DB 196,99,37,56,222,1 ; vinserti128 $0x1,%xmm6,%ymm11,%ymm11 - DB 196,99,125,57,222,1 ; vextracti128 $0x1,%ymm11,%xmm6 - DB 196,131,73,34,116,152,16,0 ; vpinsrd $0x0,0x10(%r8,%r11,4),%xmm6,%xmm6 - DB 196,99,37,56,222,1 ; vinserti128 $0x1,%xmm6,%ymm11,%ymm11 - DB 196,129,122,111,52,152 ; vmovdqu (%r8,%r11,4),%xmm6 - DB 196,67,77,2,219,240 ; vpblendd $0xf0,%ymm11,%ymm6,%ymm11 - DB 233,73,254,255,255 ; jmpq 24491 <_sk_srcover_rgba_8888_hsw_8bit+0x2d> - DB 196,129,121,110,100,152,40 ; vmovd 0x28(%r8,%r11,4),%xmm4 + DB 196,65,61,239,192 ; vpxor %ymm8,%ymm8,%ymm8 + DB 196,99,61,2,213,64 ; vpblendd $0x40,%ymm5,%ymm8,%ymm10 + DB 196,99,125,57,214,1 ; vextracti128 $0x1,%ymm10,%xmm6 + DB 196,195,73,34,114,20,1 ; vpinsrd $0x1,0x14(%r10),%xmm6,%xmm6 + DB 196,99,45,56,214,1 ; vinserti128 $0x1,%xmm6,%ymm10,%ymm10 + DB 196,99,125,57,214,1 ; vextracti128 $0x1,%ymm10,%xmm6 + DB 196,195,73,34,114,16,0 ; vpinsrd $0x0,0x10(%r10),%xmm6,%xmm6 + DB 196,99,45,56,214,1 ; vinserti128 $0x1,%xmm6,%ymm10,%ymm10 + DB 196,193,122,111,50 ; vmovdqu (%r10),%xmm6 + DB 196,67,77,2,210,240 ; vpblendd $0xf0,%ymm10,%ymm6,%ymm10 + DB 233,83,254,255,255 ; jmpq 2465b <_sk_srcover_rgba_8888_hsw_8bit+0x2f> + DB 196,193,121,110,98,40 ; vmovd 0x28(%r10),%xmm4 DB 196,226,121,89,228 ; vpbroadcastq %xmm4,%xmm4 DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 - DB 196,99,85,2,204,4 ; vpblendd $0x4,%ymm4,%ymm5,%ymm9 - DB 196,131,49,34,108,152,36,1 ; vpinsrd $0x1,0x24(%r8,%r11,4),%xmm9,%xmm5 - DB 196,99,53,2,205,15 ; vpblendd $0xf,%ymm5,%ymm9,%ymm9 - DB 196,129,121,110,108,152,32 ; vmovd 0x20(%r8,%r11,4),%xmm5 - DB 196,99,53,2,205,1 ; vpblendd $0x1,%ymm5,%ymm9,%ymm9 - DB 233,13,254,255,255 ; jmpq 2448b <_sk_srcover_rgba_8888_hsw_8bit+0x27> - DB 196,129,121,110,100,152,56 ; vmovd 0x38(%r8,%r11,4),%xmm4 + DB 196,99,85,2,196,4 ; vpblendd $0x4,%ymm4,%ymm5,%ymm8 + DB 196,195,57,34,106,36,1 ; vpinsrd $0x1,0x24(%r10),%xmm8,%xmm5 + DB 196,99,61,2,197,15 ; vpblendd $0xf,%ymm5,%ymm8,%ymm8 + DB 196,193,121,110,106,32 ; vmovd 0x20(%r10),%xmm5 + DB 196,99,61,2,197,1 ; vpblendd $0x1,%ymm5,%ymm8,%ymm8 + DB 196,65,126,111,18 ; vmovdqu (%r10),%ymm10 + DB 233,27,254,255,255 ; jmpq 2465b <_sk_srcover_rgba_8888_hsw_8bit+0x2f> + DB 196,193,121,110,98,56 ; vmovd 0x38(%r10),%xmm4 DB 196,226,125,89,228 ; vpbroadcastq %xmm4,%ymm4 DB 197,213,239,237 ; vpxor %ymm5,%ymm5,%ymm5 - DB 196,99,85,2,204,64 ; vpblendd $0x40,%ymm4,%ymm5,%ymm9 - DB 196,99,125,57,205,1 ; vextracti128 $0x1,%ymm9,%xmm5 - DB 196,131,81,34,108,152,52,1 ; vpinsrd $0x1,0x34(%r8,%r11,4),%xmm5,%xmm5 - DB 196,99,53,56,205,1 ; vinserti128 $0x1,%xmm5,%ymm9,%ymm9 - DB 196,99,125,57,205,1 ; vextracti128 $0x1,%ymm9,%xmm5 - DB 196,131,81,34,108,152,48,0 ; vpinsrd $0x0,0x30(%r8,%r11,4),%xmm5,%xmm5 - DB 196,99,53,56,205,1 ; vinserti128 $0x1,%xmm5,%ymm9,%ymm9 - DB 196,1,126,111,28,152 ; vmovdqu (%r8,%r11,4),%ymm11 - DB 196,129,122,111,116,152,32 ; vmovdqu 0x20(%r8,%r11,4),%xmm6 - DB 196,67,77,2,201,240 ; vpblendd $0xf0,%ymm9,%ymm6,%ymm9 - DB 233,189,253,255,255 ; jmpq 24491 <_sk_srcover_rgba_8888_hsw_8bit+0x2d> - DB 196,131,121,22,100,152,8,2 ; vpextrd $0x2,%xmm4,0x8(%r8,%r11,4) - DB 196,129,121,214,36,152 ; vmovq %xmm4,(%r8,%r11,4) - DB 233,119,254,255,255 ; jmpq 2455e <_sk_srcover_rgba_8888_hsw_8bit+0xfa> + DB 196,99,85,2,196,64 ; vpblendd $0x40,%ymm4,%ymm5,%ymm8 + DB 196,99,125,57,197,1 ; vextracti128 $0x1,%ymm8,%xmm5 + DB 196,195,81,34,106,52,1 ; vpinsrd $0x1,0x34(%r10),%xmm5,%xmm5 + DB 196,99,61,56,197,1 ; vinserti128 $0x1,%xmm5,%ymm8,%ymm8 + DB 196,99,125,57,197,1 ; vextracti128 $0x1,%ymm8,%xmm5 + DB 196,195,81,34,106,48,0 ; vpinsrd $0x0,0x30(%r10),%xmm5,%xmm5 + DB 196,99,61,56,197,1 ; vinserti128 $0x1,%xmm5,%ymm8,%ymm8 + DB 196,65,126,111,18 ; vmovdqu (%r10),%ymm10 + DB 196,193,122,111,114,32 ; vmovdqu 0x20(%r10),%xmm6 + DB 196,67,77,2,192,240 ; vpblendd $0xf0,%ymm8,%ymm6,%ymm8 + DB 233,202,253,255,255 ; jmpq 2465b <_sk_srcover_rgba_8888_hsw_8bit+0x2f> + DB 196,195,121,22,98,8,2 ; vpextrd $0x2,%xmm4,0x8(%r10) + DB 196,193,121,214,34 ; vmovq %xmm4,(%r10) + DB 233,132,254,255,255 ; jmpq 24726 <_sk_srcover_rgba_8888_hsw_8bit+0xfa> DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 - DB 196,131,121,22,108,152,24,2 ; vpextrd $0x2,%xmm5,0x18(%r8,%r11,4) + DB 196,195,121,22,106,24,2 ; vpextrd $0x2,%xmm5,0x18(%r10) DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 - DB 196,131,121,22,108,152,20,1 ; vpextrd $0x1,%xmm5,0x14(%r8,%r11,4) + DB 196,195,121,22,106,20,1 ; vpextrd $0x1,%xmm5,0x14(%r10) DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 - DB 196,129,121,126,108,152,16 ; vmovd %xmm5,0x10(%r8,%r11,4) - DB 196,129,122,127,36,152 ; vmovdqu %xmm4,(%r8,%r11,4) - DB 233,67,254,255,255 ; jmpq 2455e <_sk_srcover_rgba_8888_hsw_8bit+0xfa> - DB 196,131,121,22,108,152,40,2 ; vpextrd $0x2,%xmm5,0x28(%r8,%r11,4) - DB 196,131,121,22,108,152,36,1 ; vpextrd $0x1,%xmm5,0x24(%r8,%r11,4) - DB 196,129,121,126,108,152,32 ; vmovd %xmm5,0x20(%r8,%r11,4) - DB 196,129,126,127,36,152 ; vmovdqu %ymm4,(%r8,%r11,4) - DB 233,33,254,255,255 ; jmpq 2455e <_sk_srcover_rgba_8888_hsw_8bit+0xfa> + DB 196,193,121,126,106,16 ; vmovd %xmm5,0x10(%r10) + DB 196,193,122,127,34 ; vmovdqu %xmm4,(%r10) + DB 233,84,254,255,255 ; jmpq 24726 <_sk_srcover_rgba_8888_hsw_8bit+0xfa> + DB 196,195,121,22,106,40,2 ; vpextrd $0x2,%xmm5,0x28(%r10) + DB 196,195,121,22,106,36,1 ; vpextrd $0x1,%xmm5,0x24(%r10) + DB 196,193,121,126,106,32 ; vmovd %xmm5,0x20(%r10) + DB 196,193,126,127,34 ; vmovdqu %ymm4,(%r10) + DB 233,54,254,255,255 ; jmpq 24726 <_sk_srcover_rgba_8888_hsw_8bit+0xfa> DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 - DB 196,131,121,22,116,152,56,2 ; vpextrd $0x2,%xmm6,0x38(%r8,%r11,4) + DB 196,195,121,22,114,56,2 ; vpextrd $0x2,%xmm6,0x38(%r10) DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 - DB 196,131,121,22,116,152,52,1 ; vpextrd $0x1,%xmm6,0x34(%r8,%r11,4) + DB 196,195,121,22,114,52,1 ; vpextrd $0x1,%xmm6,0x34(%r10) DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 - DB 196,129,121,126,116,152,48 ; vmovd %xmm6,0x30(%r8,%r11,4) - DB 196,129,126,127,36,152 ; vmovdqu %ymm4,(%r8,%r11,4) - DB 196,129,122,127,108,152,32 ; vmovdqu %xmm5,0x20(%r8,%r11,4) - DB 233,230,253,255,255 ; jmpq 2455e <_sk_srcover_rgba_8888_hsw_8bit+0xfa> - DB 28,254 ; sbb $0xfe,%al + DB 196,193,121,126,114,48 ; vmovd %xmm6,0x30(%r10) + DB 196,193,126,127,34 ; vmovdqu %ymm4,(%r10) + DB 196,193,122,127,106,32 ; vmovdqu %xmm5,0x20(%r10) + DB 233,0,254,255,255 ; jmpq 24726 <_sk_srcover_rgba_8888_hsw_8bit+0xfa> + DB 102,144 ; xchg %ax,%ax + DB 52,254 ; xor $0xfe,%al DB 255 ; (bad) - DB 255,106,254 ; ljmp *-0x2(%rdx) DB 255 ; (bad) - DB 255,83,254 ; callq *-0x2(%rbx) + DB 127,254 ; jg 2492c <_sk_srcover_rgba_8888_hsw_8bit+0x300> DB 255 ; (bad) + DB 255,105,254 ; ljmp *-0x2(%rcx) DB 255 ; (bad) - DB 191,254,255,255,171 ; mov $0xabfffffe,%edi + DB 255,208 ; callq *%rax DB 254 ; (bad) DB 255 ; (bad) - DB 255,151,254,255,255,128 ; callq *-0x7f000002(%rdi) + DB 255 ; (bad) + DB 189,254,255,255,170 ; mov $0xaafffffe,%ebp DB 254 ; (bad) DB 255 ; (bad) - DB 255,19 ; callq *(%rbx) - DB 253 ; std + DB 255,148,254,255,255,14,255 ; callq *-0xf10001(%rsi,%rdi,8) DB 255 ; (bad) - DB 255,244 ; push %rsp - DB 254 ; (bad) + DB 255,2 ; incl (%rdx) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi - DB 254 ; (bad) DB 255 ; (bad) - DB 255,208 ; callq *%rax + DB 255,245 ; push %rbp DB 254 ; (bad) DB 255 ; (bad) - DB 255,68,255,255 ; incl -0x1(%rdi,%rdi,8) - DB 255,48 ; pushq (%rax) + DB 255,224 ; jmpq *%rax + DB 254 ; (bad) DB 255 ; (bad) + DB 255,83,255 ; callq *-0x1(%rbx) DB 255 ; (bad) - DB 255,28,255 ; lcall *(%rdi,%rdi,8) + DB 255,64,255 ; incl -0x1(%rax) DB 255 ; (bad) - DB 255,6 ; incl (%rsi) + DB 255,45,255,255,255,24 ; ljmp *0x18ffffff(%rip) # 19024960 <_sk_overlay_sse2_8bit+0x18ff84b1> DB 255 ; (bad) DB 255 ; (bad) - DB 255,15 ; decl (%rdi) + DB 255,38 ; jmpq *(%rsi) DB 254 ; (bad) DB 255 ; (bad) - DB 255,40 ; ljmp *(%rax) + DB 255,52,255 ; pushq (%rdi,%rdi,8) DB 255 ; (bad) + DB 255,45,255,255,255,100 ; ljmp *0x64ffffff(%rip) # 65024970 <_sk_overlay_sse2_8bit+0x64ff84c1> DB 255 ; (bad) - DB 255,32 ; jmpq *(%rax) DB 255 ; (bad) + DB 255,88,255 ; lcall *-0x1(%rax) DB 255 ; (bad) - DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) - DB 255,79,255 ; decl -0x1(%rdi) - DB 255 ; (bad) - DB 255,65,255 ; incl -0x1(%rcx) - DB 255 ; (bad) - DB 255,51 ; pushq (%rbx) + DB 255,75,255 ; decl -0x1(%rbx) DB 255 ; (bad) DB 255 ; (bad) + DB 62,255 ; ds (bad) DB 255 ; (bad) - DB 126,255 ; jle 247d1 <_sk_srcover_rgba_8888_hsw_8bit+0x36d> + DB 255,130,255,255,255,124 ; incl 0x7cffffff(%rdx) DB 255 ; (bad) - DB 255,119,255 ; pushq -0x1(%rdi) DB 255 ; (bad) - DB 255,111,255 ; ljmp *-0x1(%rdi) + DB 255,117,255 ; pushq -0x1(%rbp) DB 255 ; (bad) - DB 255,103,255 ; jmpq *-0x1(%rdi) + DB 255,110,255 ; ljmp *-0x1(%rsi) DB 255 ; (bad) - DB 255,178,255,255,255,165 ; pushq -0x5a000001(%rdx) + DB 255,178,255,255,255,166 ; pushq -0x59000001(%rdx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,151,255,255,255,137 ; callq *-0x76000001(%rdi) + DB 255,153,255,255,255,140 ; lcall *-0x73000001(%rcx) DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -34875,17 +34895,17 @@ PUBLIC _sk_scale_1_float_hsw_8bit _sk_scale_1_float_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,250,16,32 ; vmovss (%rax),%xmm4 - DB 197,218,89,37,218,138,0,0 ; vmulss 0x8ada(%rip),%xmm4,%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 197,218,89,37,10,136,0,0 ; vmulss 0x880a(%rip),%xmm4,%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 197,250,44,196 ; vcvttss2si %xmm4,%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 DB 196,226,125,120,228 ; vpbroadcastb %xmm4,%ymm4 - DB 196,226,125,48,232 ; vpmovzxbw %xmm0,%ymm5 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,241 ; vpmovzxbw %xmm1,%ymm6 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,221,219,37,173,135,0,0 ; vpand 0x87ad(%rip),%ymm4,%ymm4 # 2cfe0 <_sk_overlay_sse2_8bit+0xa88> + DB 197,221,219,37,221,132,0,0 ; vpand 0x84dd(%rip),%ymm4,%ymm4 # 2cec0 <_sk_overlay_sse2_8bit+0xa11> DB 197,221,213,249 ; vpmullw %ymm1,%ymm4,%ymm7 DB 197,93,213,198 ; vpmullw %ymm6,%ymm4,%ymm8 DB 197,93,213,200 ; vpmullw %ymm0,%ymm4,%ymm9 @@ -34898,11 +34918,11 @@ _sk_scale_1_float_hsw_8bit LABEL PROC DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,227,93,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm4,%ymm6 - DB 196,227,93,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm4,%ymm0 + DB 196,227,125,56,244,1 ; vinserti128 $0x1,%xmm4,%ymm0,%ymm6 + DB 196,227,125,70,196,49 ; vperm2i128 $0x31,%ymm4,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,85,56,225,1 ; vinserti128 $0x1,%xmm1,%ymm5,%ymm4 - DB 196,227,85,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm5,%ymm1 + DB 196,227,117,56,229,1 ; vinserti128 $0x1,%xmm5,%ymm1,%ymm4 + DB 196,227,117,70,205,49 ; vperm2i128 $0x31,%ymm5,%ymm1,%ymm1 DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -34916,43 +34936,43 @@ _sk_scale_u8_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,191,0,0,0 ; jne 24966 <_sk_scale_u8_hsw_8bit+0xdb> + DB 15,133,194,0,0,0 ; jne 24b19 <_sk_scale_u8_hsw_8bit+0xde> DB 196,129,122,111,36,24 ; vmovdqu (%r8,%r11,1),%xmm4 DB 196,226,125,49,236 ; vpmovzxbd %xmm4,%ymm5 DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,125,49,228 ; vpmovzxbd %xmm4,%ymm4 - DB 197,253,111,53,60,135,0,0 ; vmovdqa 0x873c(%rip),%ymm6 # 2d000 <_sk_overlay_sse2_8bit+0xaa8> + DB 197,253,111,53,108,132,0,0 ; vmovdqa 0x846c(%rip),%ymm6 # 2cee0 <_sk_overlay_sse2_8bit+0xa31> DB 196,226,93,0,230 ; vpshufb %ymm6,%ymm4,%ymm4 DB 196,226,85,0,238 ; vpshufb %ymm6,%ymm5,%ymm5 - DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,98,125,48,197 ; vpmovzxbw %xmm5,%ymm8 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 + DB 196,98,125,48,206 ; vpmovzxbw %xmm6,%ymm9 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,230,1 ; vextracti128 $0x1,%ymm4,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 197,221,213,225 ; vpmullw %ymm1,%ymm4,%ymm4 - DB 197,53,213,207 ; vpmullw %ymm7,%ymm9,%ymm9 + DB 197,205,213,247 ; vpmullw %ymm7,%ymm6,%ymm6 DB 197,213,213,232 ; vpmullw %ymm0,%ymm5,%ymm5 - DB 197,61,213,198 ; vpmullw %ymm6,%ymm8,%ymm8 - DB 197,189,253,246 ; vpaddw %ymm6,%ymm8,%ymm6 + DB 196,65,53,213,200 ; vpmullw %ymm8,%ymm9,%ymm9 + DB 196,65,53,253,192 ; vpaddw %ymm8,%ymm9,%ymm8 DB 197,213,253,192 ; vpaddw %ymm0,%ymm5,%ymm0 - DB 197,181,253,239 ; vpaddw %ymm7,%ymm9,%ymm5 + DB 197,205,253,239 ; vpaddw %ymm7,%ymm6,%ymm5 DB 197,221,253,201 ; vpaddw %ymm1,%ymm4,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,213,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm5 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,93,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm5 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -34961,35 +34981,35 @@ _sk_scale_u8_hsw_8bit LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 15,135,47,255,255,255 ; ja 248ad <_sk_scale_u8_hsw_8bit+0x22> + DB 15,135,44,255,255,255 ; ja 24a5d <_sk_scale_u8_hsw_8bit+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,199,0,0,0 ; lea 0xc7(%rip),%r10 # 24a50 <_sk_scale_u8_hsw_8bit+0x1c5> + DB 76,141,21,200,0,0,0 ; lea 0xc8(%rip),%r10 # 24c04 <_sk_scale_u8_hsw_8bit+0x1c9> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 233,13,255,255,255 ; jmpq 248ad <_sk_scale_u8_hsw_8bit+0x22> + DB 233,10,255,255,255 ; jmpq 24a5d <_sk_scale_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,2,2 ; vpinsrb $0x2,0x2(%r8,%r11,1),%xmm4,%xmm4 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax DB 197,249,110,232 ; vmovd %eax,%xmm5 DB 196,227,89,14,229,1 ; vpblendw $0x1,%xmm5,%xmm4,%xmm4 - DB 233,237,254,255,255 ; jmpq 248ad <_sk_scale_u8_hsw_8bit+0x22> + DB 233,234,254,255,255 ; jmpq 24a5d <_sk_scale_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,6,6 ; vpinsrb $0x6,0x6(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,5,5 ; vpinsrb $0x5,0x5(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,4,4 ; vpinsrb $0x4,0x4(%r8,%r11,1),%xmm4,%xmm4 DB 196,129,121,110,44,24 ; vmovd (%r8,%r11,1),%xmm5 DB 196,227,89,2,229,1 ; vpblendd $0x1,%xmm5,%xmm4,%xmm4 - DB 233,192,254,255,255 ; jmpq 248ad <_sk_scale_u8_hsw_8bit+0x22> + DB 233,189,254,255,255 ; jmpq 24a5d <_sk_scale_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,10,10 ; vpinsrb $0xa,0xa(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,9,9 ; vpinsrb $0x9,0x9(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,8,8 ; vpinsrb $0x8,0x8(%r8,%r11,1),%xmm4,%xmm4 DB 196,129,122,126,44,24 ; vmovq (%r8,%r11,1),%xmm5 DB 196,227,81,2,228,12 ; vpblendd $0xc,%xmm4,%xmm5,%xmm4 - DB 233,147,254,255,255 ; jmpq 248ad <_sk_scale_u8_hsw_8bit+0x22> + DB 233,144,254,255,255 ; jmpq 24a5d <_sk_scale_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,14,14 ; vpinsrb $0xe,0xe(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,13,13 ; vpinsrb $0xd,0xd(%r8,%r11,1),%xmm4,%xmm4 @@ -34997,37 +35017,39 @@ _sk_scale_u8_hsw_8bit LABEL PROC DB 196,129,122,126,44,24 ; vmovq (%r8,%r11,1),%xmm5 DB 196,131,81,34,108,24,8,2 ; vpinsrd $0x2,0x8(%r8,%r11,1),%xmm5,%xmm5 DB 196,227,81,2,228,8 ; vpblendd $0x8,%xmm4,%xmm5,%xmm4 - DB 233,94,254,255,255 ; jmpq 248ad <_sk_scale_u8_hsw_8bit+0x22> - DB 144 ; nop - DB 66,255 ; rex.X (bad) + DB 233,91,254,255,255 ; jmpq 24a5d <_sk_scale_u8_hsw_8bit+0x22> + DB 102,144 ; xchg %ax,%ax + DB 65,255 ; rex.B (bad) DB 255 ; (bad) - DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) - DB 255,80,255 ; callq *-0x1(%rax) + DB 255,91,255 ; lcall *-0x1(%rbx) DB 255 ; (bad) - DB 255,140,255,255,255,132,255 ; decl -0x7b0001(%rdi,%rdi,8) + DB 255,79,255 ; decl -0x1(%rdi) DB 255 ; (bad) + DB 255,139,255,255,255,131 ; decl -0x7c000001(%rbx) DB 255 ; (bad) - DB 124,255 ; jl 24a65 <_sk_scale_u8_hsw_8bit+0x1da> DB 255 ; (bad) - DB 255,112,255 ; pushq -0x1(%rax) DB 255 ; (bad) + DB 123,255 ; jnp 24c19 <_sk_scale_u8_hsw_8bit+0x1de> DB 255 ; (bad) - DB 185,255,255,255,177 ; mov $0xb1ffffff,%ecx + DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,169,255,255,255,157 ; ljmp *-0x62000001(%rcx) + DB 184,255,255,255,176 ; mov $0xb0ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,168,255,255,255,156 ; ljmp *-0x63000001(%rax) DB 255 ; (bad) DB 255 ; (bad) + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,214 ; callq *%rsi DB 255 ; (bad) + DB 221,255 ; (bad) DB 255 ; (bad) - DB 255,202 ; dec %edx + DB 255,213 ; callq *%rbp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -35036,17 +35058,17 @@ PUBLIC _sk_lerp_1_float_hsw_8bit _sk_lerp_1_float_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,250,16,32 ; vmovss (%rax),%xmm4 - DB 197,218,89,37,62,136,0,0 ; vmulss 0x883e(%rip),%xmm4,%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 197,218,89,37,106,133,0,0 ; vmulss 0x856a(%rip),%xmm4,%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 197,250,44,196 ; vcvttss2si %xmm4,%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 DB 196,226,125,120,228 ; vpbroadcastb %xmm4,%ymm4 - DB 196,226,125,48,232 ; vpmovzxbw %xmm0,%ymm5 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,241 ; vpmovzxbw %xmm1,%ymm6 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,221,219,61,81,133,0,0 ; vpand 0x8551(%rip),%ymm4,%ymm7 # 2d020 <_sk_overlay_sse2_8bit+0xac8> + DB 197,221,219,61,125,130,0,0 ; vpand 0x827d(%rip),%ymm4,%ymm7 # 2cf00 <_sk_overlay_sse2_8bit+0xa51> DB 197,69,213,193 ; vpmullw %ymm1,%ymm7,%ymm8 DB 197,69,213,206 ; vpmullw %ymm6,%ymm7,%ymm9 DB 197,69,213,208 ; vpmullw %ymm0,%ymm7,%ymm10 @@ -35059,43 +35081,43 @@ _sk_lerp_1_float_hsw_8bit LABEL PROC DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 196,227,85,56,248,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm7 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 - DB 197,69,103,192 ; vpackuswb %ymm0,%ymm7,%ymm8 - DB 196,227,77,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm6,%ymm5 - DB 196,227,77,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm6,%ymm1 - DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 + DB 196,227,125,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm7 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 + DB 197,197,103,192 ; vpackuswb %ymm0,%ymm7,%ymm0 + DB 196,227,117,56,238,1 ; vinserti128 $0x1,%xmm6,%ymm1,%ymm5 + DB 196,227,117,70,206,49 ; vperm2i128 $0x31,%ymm6,%ymm1,%ymm1 + DB 197,85,103,201 ; vpackuswb %ymm1,%ymm5,%ymm9 DB 197,213,118,237 ; vpcmpeqd %ymm5,%ymm5,%ymm5 DB 197,221,239,229 ; vpxor %ymm5,%ymm4,%ymm4 - DB 196,226,125,48,234 ; vpmovzxbw %xmm2,%ymm5 - DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 - DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,226,125,48,251 ; vpmovzxbw %xmm3,%ymm7 - DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0 - DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,213,1 ; vextracti128 $0x1,%ymm2,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,226,125,48,242 ; vpmovzxbw %xmm2,%ymm6 + DB 196,227,125,57,223,1 ; vextracti128 $0x1,%ymm3,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,98,125,48,195 ; vpmovzxbw %xmm3,%ymm8 + DB 196,227,125,57,225,1 ; vextracti128 $0x1,%ymm4,%xmm1 + DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,93,213,208 ; vpmullw %ymm0,%ymm4,%ymm10 - DB 197,53,213,223 ; vpmullw %ymm7,%ymm9,%ymm11 + DB 196,65,93,213,208 ; vpmullw %ymm8,%ymm4,%ymm10 + DB 197,117,213,223 ; vpmullw %ymm7,%ymm1,%ymm11 DB 197,221,213,230 ; vpmullw %ymm6,%ymm4,%ymm4 - DB 197,53,213,205 ; vpmullw %ymm5,%ymm9,%ymm9 - DB 197,181,253,237 ; vpaddw %ymm5,%ymm9,%ymm5 + DB 197,245,213,205 ; vpmullw %ymm5,%ymm1,%ymm1 + DB 197,245,253,205 ; vpaddw %ymm5,%ymm1,%ymm1 DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 - DB 197,165,253,247 ; vpaddw %ymm7,%ymm11,%ymm6 - DB 197,173,253,192 ; vpaddw %ymm0,%ymm10,%ymm0 - DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 + DB 197,165,253,239 ; vpaddw %ymm7,%ymm11,%ymm5 + DB 196,193,45,253,240 ; vpaddw %ymm8,%ymm10,%ymm6 DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 196,227,85,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm7 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 - DB 196,227,77,56,232,1 ; vinserti128 $0x1,%xmm0,%ymm6,%ymm5 - DB 196,227,77,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm6,%ymm0 - DB 197,213,103,232 ; vpackuswb %ymm0,%ymm5,%ymm5 - DB 196,193,93,252,192 ; vpaddb %ymm8,%ymm4,%ymm0 - DB 197,213,252,201 ; vpaddb %ymm1,%ymm5,%ymm1 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 + DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 + DB 196,227,93,56,249,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm7 + DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 + DB 197,197,103,201 ; vpackuswb %ymm1,%ymm7,%ymm1 + DB 197,245,252,192 ; vpaddb %ymm0,%ymm1,%ymm0 + DB 196,227,77,56,205,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm1 + DB 196,227,77,70,229,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm4 + DB 197,245,103,204 ; vpackuswb %ymm4,%ymm1,%ymm1 + DB 196,193,117,252,201 ; vpaddb %ymm9,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -35108,79 +35130,79 @@ _sk_lerp_u8_hsw_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,106,1,0,0 ; jne 24d42 <_sk_lerp_u8_hsw_8bit+0x186> + DB 15,133,107,1,0,0 ; jne 24ef9 <_sk_lerp_u8_hsw_8bit+0x187> DB 196,129,122,111,36,24 ; vmovdqu (%r8,%r11,1),%xmm4 DB 196,226,125,49,236 ; vpmovzxbd %xmm4,%ymm5 DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,125,49,228 ; vpmovzxbd %xmm4,%ymm4 - DB 197,253,111,53,75,132,0,0 ; vmovdqa 0x844b(%rip),%ymm6 # 2d040 <_sk_overlay_sse2_8bit+0xae8> - DB 196,98,93,0,206 ; vpshufb %ymm6,%ymm4,%ymm9 - DB 196,98,85,0,222 ; vpshufb %ymm6,%ymm5,%ymm11 - DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 197,253,111,53,117,129,0,0 ; vmovdqa 0x8175(%rip),%ymm6 # 2cf20 <_sk_overlay_sse2_8bit+0xa71> + DB 196,98,93,0,198 ; vpshufb %ymm6,%ymm4,%ymm8 + DB 196,226,85,0,238 ; vpshufb %ymm6,%ymm5,%ymm5 + DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,66,125,48,195 ; vpmovzxbw %xmm11,%ymm8 - DB 196,99,125,57,220,1 ; vextracti128 $0x1,%ymm11,%xmm4 + DB 196,227,125,57,236,1 ; vextracti128 $0x1,%ymm5,%xmm4 + DB 196,98,125,48,212 ; vpmovzxbw %xmm4,%ymm10 + DB 196,98,125,48,205 ; vpmovzxbw %xmm5,%ymm9 + DB 196,99,125,57,196,1 ; vextracti128 $0x1,%ymm8,%xmm4 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,66,125,48,209 ; vpmovzxbw %xmm9,%ymm10 - DB 196,99,125,57,205,1 ; vextracti128 $0x1,%ymm9,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,213,213,233 ; vpmullw %ymm1,%ymm5,%ymm5 - DB 197,45,213,215 ; vpmullw %ymm7,%ymm10,%ymm10 - DB 197,221,213,224 ; vpmullw %ymm0,%ymm4,%ymm4 - DB 197,61,213,198 ; vpmullw %ymm6,%ymm8,%ymm8 - DB 197,189,253,246 ; vpaddw %ymm6,%ymm8,%ymm6 - DB 197,221,253,192 ; vpaddw %ymm0,%ymm4,%ymm0 - DB 197,173,253,231 ; vpaddw %ymm7,%ymm10,%ymm4 - DB 197,213,253,201 ; vpaddw %ymm1,%ymm5,%ymm1 + DB 196,66,125,48,216 ; vpmovzxbw %xmm8,%ymm11 + DB 197,37,213,217 ; vpmullw %ymm1,%ymm11,%ymm11 + DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 + DB 197,53,213,200 ; vpmullw %ymm0,%ymm9,%ymm9 + DB 197,45,213,214 ; vpmullw %ymm6,%ymm10,%ymm10 + DB 197,173,253,246 ; vpaddw %ymm6,%ymm10,%ymm6 + DB 197,181,253,192 ; vpaddw %ymm0,%ymm9,%ymm0 + DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 + DB 197,165,253,201 ; vpaddw %ymm1,%ymm11,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,213,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm5 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 - DB 197,77,103,208 ; vpackuswb %ymm0,%ymm6,%ymm10 - DB 196,227,93,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm5 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 - DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 196,227,125,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm0,%ymm7 + DB 196,227,125,70,198,49 ; vperm2i128 $0x31,%ymm6,%ymm0,%ymm0 + DB 197,69,103,208 ; vpackuswb %ymm0,%ymm7,%ymm10 + DB 196,227,117,56,244,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm6 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 + DB 197,77,103,201 ; vpackuswb %ymm1,%ymm6,%ymm9 DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 - DB 197,181,239,236 ; vpxor %ymm4,%ymm9,%ymm5 - DB 197,165,239,228 ; vpxor %ymm4,%ymm11,%ymm4 - DB 196,226,125,48,242 ; vpmovzxbw %xmm2,%ymm6 - DB 196,227,125,57,215,1 ; vextracti128 $0x1,%ymm2,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 197,189,239,244 ; vpxor %ymm4,%ymm8,%ymm6 + DB 197,213,239,228 ; vpxor %ymm4,%ymm5,%ymm4 + DB 196,227,125,57,213,1 ; vextracti128 $0x1,%ymm2,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,226,125,48,250 ; vpmovzxbw %xmm2,%ymm7 + DB 196,227,125,57,217,1 ; vextracti128 $0x1,%ymm3,%xmm1 + DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 DB 196,98,125,48,195 ; vpmovzxbw %xmm3,%ymm8 - DB 196,227,125,57,216,1 ; vextracti128 $0x1,%ymm3,%xmm0 - DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,224,1 ; vextracti128 $0x1,%ymm4,%xmm0 + DB 196,98,125,48,216 ; vpmovzxbw %xmm0,%ymm11 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,221 ; vpmovzxbw %xmm5,%ymm11 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,213,213,232 ; vpmullw %ymm0,%ymm5,%ymm5 - DB 196,65,37,213,216 ; vpmullw %ymm8,%ymm11,%ymm11 + DB 196,227,125,57,240,1 ; vextracti128 $0x1,%ymm6,%xmm0 + DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,193,77,213,240 ; vpmullw %ymm8,%ymm6,%ymm6 + DB 197,253,213,193 ; vpmullw %ymm1,%ymm0,%ymm0 DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 197,53,213,206 ; vpmullw %ymm6,%ymm9,%ymm9 - DB 197,181,253,246 ; vpaddw %ymm6,%ymm9,%ymm6 + DB 197,37,213,221 ; vpmullw %ymm5,%ymm11,%ymm11 + DB 197,165,253,237 ; vpaddw %ymm5,%ymm11,%ymm5 DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 196,193,37,253,248 ; vpaddw %ymm8,%ymm11,%ymm7 - DB 197,213,253,192 ; vpaddw %ymm0,%ymm5,%ymm0 - DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,213,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm5 - DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 196,227,77,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm6,%ymm7 - DB 196,227,77,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm6,%ymm4 - DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 - DB 197,205,103,232 ; vpackuswb %ymm0,%ymm6,%ymm5 - DB 196,193,93,252,194 ; vpaddb %ymm10,%ymm4,%ymm0 - DB 197,213,252,201 ; vpaddb %ymm1,%ymm5,%ymm1 + DB 197,253,253,193 ; vpaddw %ymm1,%ymm0,%ymm0 + DB 196,193,77,253,200 ; vpaddw %ymm8,%ymm6,%ymm1 + DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 + DB 197,205,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm6 + DB 197,253,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm0 + DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 + DB 196,227,125,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm0,%ymm5 + DB 196,227,125,70,196,49 ; vperm2i128 $0x31,%ymm4,%ymm0,%ymm0 + DB 197,213,103,192 ; vpackuswb %ymm0,%ymm5,%ymm0 + DB 196,193,125,252,194 ; vpaddb %ymm10,%ymm0,%ymm0 + DB 196,227,117,56,230,1 ; vinserti128 $0x1,%xmm6,%ymm1,%ymm4 + DB 196,227,117,70,206,49 ; vperm2i128 $0x31,%ymm6,%ymm1,%ymm1 + DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 + DB 196,193,117,252,201 ; vpaddb %ymm9,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -35188,35 +35210,35 @@ _sk_lerp_u8_hsw_8bit LABEL PROC DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,14 ; cmp $0xe,%r9b - DB 15,135,132,254,255,255 ; ja 24bde <_sk_lerp_u8_hsw_8bit+0x22> + DB 15,135,131,254,255,255 ; ja 24d94 <_sk_lerp_u8_hsw_8bit+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,199,0,0,0 ; lea 0xc7(%rip),%r10 # 24e2c <_sk_lerp_u8_hsw_8bit+0x270> + DB 76,141,21,200,0,0,0 ; lea 0xc8(%rip),%r10 # 24fe4 <_sk_lerp_u8_hsw_8bit+0x272> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 - DB 233,98,254,255,255 ; jmpq 24bde <_sk_lerp_u8_hsw_8bit+0x22> + DB 233,97,254,255,255 ; jmpq 24d94 <_sk_lerp_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,2,2 ; vpinsrb $0x2,0x2(%r8,%r11,1),%xmm4,%xmm4 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax DB 197,249,110,232 ; vmovd %eax,%xmm5 DB 196,227,89,14,229,1 ; vpblendw $0x1,%xmm5,%xmm4,%xmm4 - DB 233,66,254,255,255 ; jmpq 24bde <_sk_lerp_u8_hsw_8bit+0x22> + DB 233,65,254,255,255 ; jmpq 24d94 <_sk_lerp_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,6,6 ; vpinsrb $0x6,0x6(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,5,5 ; vpinsrb $0x5,0x5(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,4,4 ; vpinsrb $0x4,0x4(%r8,%r11,1),%xmm4,%xmm4 DB 196,129,121,110,44,24 ; vmovd (%r8,%r11,1),%xmm5 DB 196,227,89,2,229,1 ; vpblendd $0x1,%xmm5,%xmm4,%xmm4 - DB 233,21,254,255,255 ; jmpq 24bde <_sk_lerp_u8_hsw_8bit+0x22> + DB 233,20,254,255,255 ; jmpq 24d94 <_sk_lerp_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,10,10 ; vpinsrb $0xa,0xa(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,9,9 ; vpinsrb $0x9,0x9(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,8,8 ; vpinsrb $0x8,0x8(%r8,%r11,1),%xmm4,%xmm4 DB 196,129,122,126,44,24 ; vmovq (%r8,%r11,1),%xmm5 DB 196,227,81,2,228,12 ; vpblendd $0xc,%xmm4,%xmm5,%xmm4 - DB 233,232,253,255,255 ; jmpq 24bde <_sk_lerp_u8_hsw_8bit+0x22> + DB 233,231,253,255,255 ; jmpq 24d94 <_sk_lerp_u8_hsw_8bit+0x22> DB 197,217,239,228 ; vpxor %xmm4,%xmm4,%xmm4 DB 196,131,89,32,100,24,14,14 ; vpinsrb $0xe,0xe(%r8,%r11,1),%xmm4,%xmm4 DB 196,131,89,32,100,24,13,13 ; vpinsrb $0xd,0xd(%r8,%r11,1),%xmm4,%xmm4 @@ -35224,37 +35246,39 @@ _sk_lerp_u8_hsw_8bit LABEL PROC DB 196,129,122,126,44,24 ; vmovq (%r8,%r11,1),%xmm5 DB 196,131,81,34,108,24,8,2 ; vpinsrd $0x2,0x8(%r8,%r11,1),%xmm5,%xmm5 DB 196,227,81,2,228,8 ; vpblendd $0x8,%xmm4,%xmm5,%xmm4 - DB 233,179,253,255,255 ; jmpq 24bde <_sk_lerp_u8_hsw_8bit+0x22> - DB 144 ; nop - DB 66,255 ; rex.X (bad) + DB 233,178,253,255,255 ; jmpq 24d94 <_sk_lerp_u8_hsw_8bit+0x22> + DB 102,144 ; xchg %ax,%ax + DB 65,255 ; rex.B (bad) DB 255 ; (bad) - DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) - DB 255,80,255 ; callq *-0x1(%rax) + DB 255,91,255 ; lcall *-0x1(%rbx) DB 255 ; (bad) - DB 255,140,255,255,255,132,255 ; decl -0x7b0001(%rdi,%rdi,8) + DB 255,79,255 ; decl -0x1(%rdi) DB 255 ; (bad) + DB 255,139,255,255,255,131 ; decl -0x7c000001(%rbx) DB 255 ; (bad) - DB 124,255 ; jl 24e41 <_sk_lerp_u8_hsw_8bit+0x285> DB 255 ; (bad) - DB 255,112,255 ; pushq -0x1(%rax) DB 255 ; (bad) + DB 123,255 ; jnp 24ff9 <_sk_lerp_u8_hsw_8bit+0x287> DB 255 ; (bad) - DB 185,255,255,255,177 ; mov $0xb1ffffff,%ecx + DB 255,111,255 ; ljmp *-0x1(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,169,255,255,255,157 ; ljmp *-0x62000001(%rcx) + DB 184,255,255,255,176 ; mov $0xb0ffffff,%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 255,168,255,255,255,156 ; ljmp *-0x63000001(%rax) DB 255 ; (bad) DB 255 ; (bad) + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,214 ; callq *%rsi DB 255 ; (bad) + DB 221,255 ; (bad) DB 255 ; (bad) - DB 255,202 ; dec %edx + DB 255,213 ; callq *%rbp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -35276,7 +35300,7 @@ _sk_move_dst_src_hsw_8bit LABEL PROC PUBLIC _sk_black_color_hsw_8bit _sk_black_color_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,185,133,0,0 ; vbroadcastss 0x85b9(%rip),%ymm0 # 2d444 <_sk_overlay_sse2_8bit+0xeec> + DB 196,226,125,24,5,237,130,0,0 ; vbroadcastss 0x82ed(%rip),%ymm0 # 2d330 <_sk_overlay_sse2_8bit+0xe81> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 255,224 ; jmpq *%rax @@ -35296,485 +35320,480 @@ _sk_clear_hsw_8bit LABEL PROC PUBLIC _sk_srcatop_hsw_8bit _sk_srcatop_hsw_8bit LABEL PROC - DB 197,125,111,5,175,129,0,0 ; vmovdqa 0x81af(%rip),%ymm8 # 2d060 <_sk_overlay_sse2_8bit+0xb08> - DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 - DB 196,194,109,0,232 ; vpshufb %ymm8,%ymm2,%ymm5 - DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 - DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 197,125,111,13,215,126,0,0 ; vmovdqa 0x7ed7(%rip),%ymm9 # 2cf40 <_sk_overlay_sse2_8bit+0xa91> + DB 196,194,101,0,225 ; vpshufb %ymm9,%ymm3,%ymm4 + DB 196,194,109,0,233 ; vpshufb %ymm9,%ymm2,%ymm5 + DB 196,227,125,57,199,1 ; vextracti128 $0x1,%ymm0,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 + DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 + DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,221 ; vpmovzxbw %xmm5,%ymm11 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,98,125,48,209 ; vpmovzxbw %xmm1,%ymm10 + DB 196,227,125,57,239,1 ; vextracti128 $0x1,%ymm5,%xmm7 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,231,1 ; vextracti128 $0x1,%ymm4,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 196,65,29,213,225 ; vpmullw %ymm9,%ymm12,%ymm12 - DB 197,213,213,238 ; vpmullw %ymm6,%ymm5,%ymm5 - DB 196,65,37,213,218 ; vpmullw %ymm10,%ymm11,%ymm11 - DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 - DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 - DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 + DB 196,193,93,213,226 ; vpmullw %ymm10,%ymm4,%ymm4 + DB 197,197,213,254 ; vpmullw %ymm6,%ymm7,%ymm7 + DB 196,193,85,213,232 ; vpmullw %ymm8,%ymm5,%ymm5 + DB 196,65,29,213,227 ; vpmullw %ymm11,%ymm12,%ymm12 + DB 196,65,29,253,219 ; vpaddw %ymm11,%ymm12,%ymm11 + DB 196,193,85,253,232 ; vpaddw %ymm8,%ymm5,%ymm5 + DB 197,197,253,246 ; vpaddw %ymm6,%ymm7,%ymm6 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 - DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 - DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 197,53,103,204 ; vpackuswb %ymm4,%ymm9,%ymm9 - DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 - DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 - DB 197,213,103,238 ; vpackuswb %ymm6,%ymm5,%ymm5 - DB 196,194,125,0,192 ; vpshufb %ymm8,%ymm0,%ymm0 - DB 196,194,117,0,200 ; vpshufb %ymm8,%ymm1,%ymm1 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 196,193,69,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm7 + DB 196,99,85,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm5,%ymm8 + DB 196,227,85,70,239,49 ; vperm2i128 $0x31,%ymm7,%ymm5,%ymm5 + DB 197,61,103,213 ; vpackuswb %ymm5,%ymm8,%ymm10 + DB 196,227,93,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm4,%ymm7 + DB 196,227,93,70,230,49 ; vperm2i128 $0x31,%ymm6,%ymm4,%ymm4 + DB 197,69,103,196 ; vpackuswb %ymm4,%ymm7,%ymm8 + DB 196,194,125,0,193 ; vpshufb %ymm9,%ymm0,%ymm0 + DB 196,194,117,0,201 ; vpshufb %ymm9,%ymm1,%ymm1 DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 DB 197,245,239,206 ; vpxor %ymm6,%ymm1,%ymm1 DB 197,253,239,198 ; vpxor %ymm6,%ymm0,%ymm0 - DB 196,226,125,48,242 ; vpmovzxbw %xmm2,%ymm6 - DB 196,227,125,57,215,1 ; vextracti128 $0x1,%ymm2,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,195 ; vpmovzxbw %xmm3,%ymm8 + DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,226,125,48,250 ; vpmovzxbw %xmm2,%ymm7 DB 196,227,125,57,220,1 ; vextracti128 $0x1,%ymm3,%xmm4 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,98,125,48,203 ; vpmovzxbw %xmm3,%ymm9 + DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 196,98,125,48,221 ; vpmovzxbw %xmm5,%ymm11 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,217 ; vpmovzxbw %xmm1,%ymm11 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,205,1 ; vextracti128 $0x1,%ymm1,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,221,213,201 ; vpmullw %ymm1,%ymm4,%ymm1 - DB 196,65,61,213,219 ; vpmullw %ymm11,%ymm8,%ymm11 + DB 197,181,213,201 ; vpmullw %ymm1,%ymm9,%ymm1 + DB 197,221,213,237 ; vpmullw %ymm5,%ymm4,%ymm5 DB 197,197,213,192 ; vpmullw %ymm0,%ymm7,%ymm0 - DB 196,65,77,213,210 ; vpmullw %ymm10,%ymm6,%ymm10 - DB 197,173,253,246 ; vpaddw %ymm6,%ymm10,%ymm6 + DB 196,65,77,213,219 ; vpmullw %ymm11,%ymm6,%ymm11 + DB 197,165,253,246 ; vpaddw %ymm6,%ymm11,%ymm6 DB 197,253,253,199 ; vpaddw %ymm7,%ymm0,%ymm0 - DB 196,193,37,253,248 ; vpaddw %ymm8,%ymm11,%ymm7 - DB 197,245,253,204 ; vpaddw %ymm4,%ymm1,%ymm1 + DB 197,213,253,228 ; vpaddw %ymm4,%ymm5,%ymm4 + DB 196,193,117,253,201 ; vpaddw %ymm9,%ymm1,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 - DB 197,221,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm4 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 196,227,77,56,248,1 ; vinserti128 $0x1,%xmm0,%ymm6,%ymm7 - DB 196,227,77,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm6,%ymm0 - DB 197,197,103,192 ; vpackuswb %ymm0,%ymm7,%ymm0 - DB 196,227,93,56,241,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm6 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 - DB 197,205,103,201 ; vpackuswb %ymm1,%ymm6,%ymm1 - DB 196,193,125,252,193 ; vpaddb %ymm9,%ymm0,%ymm0 - DB 197,245,252,205 ; vpaddb %ymm5,%ymm1,%ymm1 + DB 197,213,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 + DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 + DB 196,193,125,252,194 ; vpaddb %ymm10,%ymm0,%ymm0 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 + DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 + DB 196,193,117,252,200 ; vpaddb %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstatop_hsw_8bit _sk_dstatop_hsw_8bit LABEL PROC - DB 197,125,111,5,106,128,0,0 ; vmovdqa 0x806a(%rip),%ymm8 # 2d080 <_sk_overlay_sse2_8bit+0xb28> - DB 196,194,117,0,224 ; vpshufb %ymm8,%ymm1,%ymm4 - DB 196,194,125,0,232 ; vpshufb %ymm8,%ymm0,%ymm5 - DB 196,98,125,48,210 ; vpmovzxbw %xmm2,%ymm10 - DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 + DB 197,125,111,13,144,125,0,0 ; vmovdqa 0x7d90(%rip),%ymm9 # 2cf60 <_sk_overlay_sse2_8bit+0xab1> + DB 196,194,117,0,225 ; vpshufb %ymm9,%ymm1,%ymm4 + DB 196,194,125,0,233 ; vpshufb %ymm9,%ymm0,%ymm5 + DB 196,227,125,57,215,1 ; vextracti128 $0x1,%ymm2,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 + DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,203 ; vpmovzxbw %xmm3,%ymm9 - DB 196,227,125,57,223,1 ; vextracti128 $0x1,%ymm3,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,221 ; vpmovzxbw %xmm5,%ymm11 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,227,125,57,239,1 ; vextracti128 $0x1,%ymm5,%xmm7 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,231,1 ; vextracti128 $0x1,%ymm4,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,197,213,228 ; vpmullw %ymm4,%ymm7,%ymm4 - DB 196,65,53,213,228 ; vpmullw %ymm12,%ymm9,%ymm12 - DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 - DB 196,65,45,213,219 ; vpmullw %ymm11,%ymm10,%ymm11 - DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 - DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 - DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 + DB 197,173,213,228 ; vpmullw %ymm4,%ymm10,%ymm4 + DB 197,205,213,255 ; vpmullw %ymm7,%ymm6,%ymm7 + DB 197,189,213,237 ; vpmullw %ymm5,%ymm8,%ymm5 + DB 196,65,37,213,228 ; vpmullw %ymm12,%ymm11,%ymm12 + DB 196,65,29,253,219 ; vpaddw %ymm11,%ymm12,%ymm11 + DB 196,193,85,253,232 ; vpaddw %ymm8,%ymm5,%ymm5 + DB 197,197,253,246 ; vpaddw %ymm6,%ymm7,%ymm6 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 - DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 - DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 197,181,103,228 ; vpackuswb %ymm4,%ymm9,%ymm4 - DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 - DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 - DB 197,213,103,238 ; vpackuswb %ymm6,%ymm5,%ymm5 - DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 - DB 196,194,101,0,248 ; vpshufb %ymm8,%ymm3,%ymm7 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 196,193,69,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm7 + DB 196,99,85,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm5,%ymm8 + DB 196,227,85,70,239,49 ; vperm2i128 $0x31,%ymm7,%ymm5,%ymm5 + DB 197,61,103,221 ; vpackuswb %ymm5,%ymm8,%ymm11 + DB 196,227,93,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm4,%ymm7 + DB 196,227,93,70,230,49 ; vperm2i128 $0x31,%ymm6,%ymm4,%ymm4 + DB 197,69,103,212 ; vpackuswb %ymm4,%ymm7,%ymm10 + DB 196,194,109,0,241 ; vpshufb %ymm9,%ymm2,%ymm6 + DB 196,194,101,0,249 ; vpshufb %ymm9,%ymm3,%ymm7 DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 DB 196,193,69,239,248 ; vpxor %ymm8,%ymm7,%ymm7 DB 196,193,77,239,240 ; vpxor %ymm8,%ymm6,%ymm6 - DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,196 ; vpmovzxbw %xmm4,%ymm8 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,205,1 ; vextracti128 $0x1,%ymm1,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,98,125,48,214 ; vpmovzxbw %xmm6,%ymm10 - DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 + DB 196,227,125,57,244,1 ; vextracti128 $0x1,%ymm6,%xmm4 + DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 - DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 + DB 196,227,125,57,252,1 ; vextracti128 $0x1,%ymm7,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 197,197,213,249 ; vpmullw %ymm1,%ymm7,%ymm7 - DB 196,65,37,213,217 ; vpmullw %ymm9,%ymm11,%ymm11 + DB 197,221,213,229 ; vpmullw %ymm5,%ymm4,%ymm4 DB 197,205,213,240 ; vpmullw %ymm0,%ymm6,%ymm6 - DB 196,65,45,213,208 ; vpmullw %ymm8,%ymm10,%ymm10 - DB 196,65,45,253,192 ; vpaddw %ymm8,%ymm10,%ymm8 + DB 196,65,53,213,200 ; vpmullw %ymm8,%ymm9,%ymm9 + DB 196,65,53,253,192 ; vpaddw %ymm8,%ymm9,%ymm8 DB 197,205,253,192 ; vpaddw %ymm0,%ymm6,%ymm0 - DB 196,193,37,253,241 ; vpaddw %ymm9,%ymm11,%ymm6 + DB 197,221,253,229 ; vpaddw %ymm5,%ymm4,%ymm4 DB 197,197,253,201 ; vpaddw %ymm1,%ymm7,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,192,1 ; vinserti128 $0x1,%xmm0,%ymm7,%ymm8 - DB 196,227,69,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm7,%ymm0 - DB 197,189,103,192 ; vpackuswb %ymm0,%ymm8,%ymm0 - DB 196,227,77,56,249,1 ; vinserti128 $0x1,%xmm1,%ymm6,%ymm7 - DB 196,227,77,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm6,%ymm1 - DB 197,197,103,201 ; vpackuswb %ymm1,%ymm7,%ymm1 - DB 197,253,252,196 ; vpaddb %ymm4,%ymm0,%ymm0 - DB 197,245,252,205 ; vpaddb %ymm5,%ymm1,%ymm1 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 + DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 + DB 196,193,125,252,195 ; vpaddb %ymm11,%ymm0,%ymm0 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 + DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 + DB 196,193,117,252,202 ; vpaddb %ymm10,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_srcin_hsw_8bit _sk_srcin_hsw_8bit LABEL PROC - DB 197,253,111,37,33,127,0,0 ; vmovdqa 0x7f21(%rip),%ymm4 # 2d0a0 <_sk_overlay_sse2_8bit+0xb48> + DB 197,253,111,37,71,124,0,0 ; vmovdqa 0x7c47(%rip),%ymm4 # 2cf80 <_sk_overlay_sse2_8bit+0xad1> DB 196,226,101,0,236 ; vpshufb %ymm4,%ymm3,%ymm5 DB 196,226,109,0,228 ; vpshufb %ymm4,%ymm2,%ymm4 - DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,98,125,48,196 ; vpmovzxbw %xmm4,%ymm8 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,230,1 ; vextracti128 $0x1,%ymm4,%xmm6 + DB 196,98,125,48,206 ; vpmovzxbw %xmm6,%ymm9 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,205 ; vpmovzxbw %xmm5,%ymm9 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 197,213,213,233 ; vpmullw %ymm1,%ymm5,%ymm5 - DB 197,53,213,207 ; vpmullw %ymm7,%ymm9,%ymm9 + DB 197,205,213,247 ; vpmullw %ymm7,%ymm6,%ymm6 DB 197,221,213,224 ; vpmullw %ymm0,%ymm4,%ymm4 - DB 197,61,213,198 ; vpmullw %ymm6,%ymm8,%ymm8 - DB 197,189,253,246 ; vpaddw %ymm6,%ymm8,%ymm6 + DB 196,65,53,213,200 ; vpmullw %ymm8,%ymm9,%ymm9 + DB 196,65,53,253,192 ; vpaddw %ymm8,%ymm9,%ymm8 DB 197,221,253,192 ; vpaddw %ymm0,%ymm4,%ymm0 - DB 197,181,253,231 ; vpaddw %ymm7,%ymm9,%ymm4 + DB 197,205,253,231 ; vpaddw %ymm7,%ymm6,%ymm4 DB 197,213,253,201 ; vpaddw %ymm1,%ymm5,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,213,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm5 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,93,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm5 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstin_hsw_8bit _sk_dstin_hsw_8bit LABEL PROC - DB 197,253,111,37,151,126,0,0 ; vmovdqa 0x7e97(%rip),%ymm4 # 2d0c0 <_sk_overlay_sse2_8bit+0xb68> + DB 197,253,111,37,186,123,0,0 ; vmovdqa 0x7bba(%rip),%ymm4 # 2cfa0 <_sk_overlay_sse2_8bit+0xaf1> DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 - DB 196,226,125,48,226 ; vpmovzxbw %xmm2,%ymm4 - DB 196,227,125,57,213,1 ; vextracti128 $0x1,%ymm2,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,226,125,48,243 ; vpmovzxbw %xmm3,%ymm6 - DB 196,227,125,57,223,1 ; vextracti128 $0x1,%ymm3,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 + DB 196,98,125,48,196 ; vpmovzxbw %xmm4,%ymm8 + DB 196,226,125,48,234 ; vpmovzxbw %xmm2,%ymm5 + DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,226,125,48,251 ; vpmovzxbw %xmm3,%ymm7 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 DB 197,197,213,201 ; vpmullw %ymm1,%ymm7,%ymm1 - DB 196,65,77,213,201 ; vpmullw %ymm9,%ymm6,%ymm9 + DB 197,205,213,228 ; vpmullw %ymm4,%ymm6,%ymm4 DB 197,213,213,192 ; vpmullw %ymm0,%ymm5,%ymm0 - DB 196,65,93,213,192 ; vpmullw %ymm8,%ymm4,%ymm8 - DB 197,189,253,228 ; vpaddw %ymm4,%ymm8,%ymm4 + DB 196,65,61,213,201 ; vpmullw %ymm9,%ymm8,%ymm9 + DB 196,65,53,253,192 ; vpaddw %ymm8,%ymm9,%ymm8 DB 197,253,253,197 ; vpaddw %ymm5,%ymm0,%ymm0 - DB 197,181,253,238 ; vpaddw %ymm6,%ymm9,%ymm5 + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 DB 197,245,253,207 ; vpaddw %ymm7,%ymm1,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 - DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,227,93,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm4,%ymm6 - DB 196,227,93,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm4,%ymm0 + DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,85,56,225,1 ; vinserti128 $0x1,%xmm1,%ymm5,%ymm4 - DB 196,227,85,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm5,%ymm1 - DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 + DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_srcout_hsw_8bit _sk_srcout_hsw_8bit LABEL PROC - DB 197,253,111,37,11,126,0,0 ; vmovdqa 0x7e0b(%rip),%ymm4 # 2d0e0 <_sk_overlay_sse2_8bit+0xb88> + DB 197,253,111,37,45,123,0,0 ; vmovdqa 0x7b2d(%rip),%ymm4 # 2cfc0 <_sk_overlay_sse2_8bit+0xb11> DB 196,226,109,0,236 ; vpshufb %ymm4,%ymm2,%ymm5 DB 196,226,101,0,228 ; vpshufb %ymm4,%ymm3,%ymm4 DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 DB 197,221,239,230 ; vpxor %ymm6,%ymm4,%ymm4 DB 197,213,239,238 ; vpxor %ymm6,%ymm5,%ymm5 - DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,98,125,48,197 ; vpmovzxbw %xmm5,%ymm8 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,238,1 ; vextracti128 $0x1,%ymm5,%xmm6 + DB 196,98,125,48,206 ; vpmovzxbw %xmm6,%ymm9 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,230,1 ; vextracti128 $0x1,%ymm4,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 197,221,213,225 ; vpmullw %ymm1,%ymm4,%ymm4 - DB 197,53,213,207 ; vpmullw %ymm7,%ymm9,%ymm9 + DB 197,205,213,247 ; vpmullw %ymm7,%ymm6,%ymm6 DB 197,213,213,232 ; vpmullw %ymm0,%ymm5,%ymm5 - DB 197,61,213,198 ; vpmullw %ymm6,%ymm8,%ymm8 - DB 197,189,253,246 ; vpaddw %ymm6,%ymm8,%ymm6 + DB 196,65,53,213,200 ; vpmullw %ymm8,%ymm9,%ymm9 + DB 196,65,53,253,192 ; vpaddw %ymm8,%ymm9,%ymm8 DB 197,213,253,192 ; vpaddw %ymm0,%ymm5,%ymm0 - DB 197,181,253,239 ; vpaddw %ymm7,%ymm9,%ymm5 + DB 197,205,253,239 ; vpaddw %ymm7,%ymm6,%ymm5 DB 197,221,253,201 ; vpaddw %ymm1,%ymm4,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 197,213,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm5 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,93,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm5 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstout_hsw_8bit _sk_dstout_hsw_8bit LABEL PROC - DB 197,253,111,37,117,125,0,0 ; vmovdqa 0x7d75(%rip),%ymm4 # 2d100 <_sk_overlay_sse2_8bit+0xba8> + DB 197,253,111,37,148,122,0,0 ; vmovdqa 0x7a94(%rip),%ymm4 # 2cfe0 <_sk_overlay_sse2_8bit+0xb31> DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 DB 197,245,239,204 ; vpxor %ymm4,%ymm1,%ymm1 DB 197,253,239,196 ; vpxor %ymm4,%ymm0,%ymm0 - DB 196,226,125,48,226 ; vpmovzxbw %xmm2,%ymm4 - DB 196,227,125,57,213,1 ; vextracti128 $0x1,%ymm2,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,226,125,48,243 ; vpmovzxbw %xmm3,%ymm6 - DB 196,227,125,57,223,1 ; vextracti128 $0x1,%ymm3,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 + DB 196,98,125,48,196 ; vpmovzxbw %xmm4,%ymm8 + DB 196,226,125,48,234 ; vpmovzxbw %xmm2,%ymm5 + DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,226,125,48,251 ; vpmovzxbw %xmm3,%ymm7 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 DB 197,197,213,201 ; vpmullw %ymm1,%ymm7,%ymm1 - DB 196,65,77,213,201 ; vpmullw %ymm9,%ymm6,%ymm9 + DB 197,205,213,228 ; vpmullw %ymm4,%ymm6,%ymm4 DB 197,213,213,192 ; vpmullw %ymm0,%ymm5,%ymm0 - DB 196,65,93,213,192 ; vpmullw %ymm8,%ymm4,%ymm8 - DB 197,189,253,228 ; vpaddw %ymm4,%ymm8,%ymm4 + DB 196,65,61,213,201 ; vpmullw %ymm9,%ymm8,%ymm9 + DB 196,65,53,253,192 ; vpaddw %ymm8,%ymm9,%ymm8 DB 197,253,253,197 ; vpaddw %ymm5,%ymm0,%ymm0 - DB 197,181,253,238 ; vpaddw %ymm6,%ymm9,%ymm5 + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 DB 197,245,253,207 ; vpaddw %ymm7,%ymm1,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 - DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,227,93,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm4,%ymm6 - DB 196,227,93,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm4,%ymm0 + DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 + DB 196,227,125,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm0,%ymm6 + DB 196,227,125,70,197,49 ; vperm2i128 $0x31,%ymm5,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,85,56,225,1 ; vinserti128 $0x1,%xmm1,%ymm5,%ymm4 - DB 196,227,85,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm5,%ymm1 - DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 + DB 196,227,117,56,236,1 ; vinserti128 $0x1,%xmm4,%ymm1,%ymm5 + DB 196,227,117,70,204,49 ; vperm2i128 $0x31,%ymm4,%ymm1,%ymm1 + DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_srcover_hsw_8bit _sk_srcover_hsw_8bit LABEL PROC - DB 197,253,111,37,221,124,0,0 ; vmovdqa 0x7cdd(%rip),%ymm4 # 2d120 <_sk_overlay_sse2_8bit+0xbc8> + DB 197,253,111,37,251,121,0,0 ; vmovdqa 0x79fb(%rip),%ymm4 # 2d000 <_sk_overlay_sse2_8bit+0xb51> DB 196,226,117,0,236 ; vpshufb %ymm4,%ymm1,%ymm5 DB 196,226,125,0,228 ; vpshufb %ymm4,%ymm0,%ymm4 - DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 - DB 196,227,125,57,215,1 ; vextracti128 $0x1,%ymm2,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,195 ; vpmovzxbw %xmm3,%ymm8 + DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 + DB 196,98,125,48,210 ; vpmovzxbw %xmm2,%ymm10 DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,212 ; vpmovzxbw %xmm4,%ymm10 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,98,125,48,203 ; vpmovzxbw %xmm3,%ymm9 + DB 196,227,125,57,231,1 ; vextracti128 $0x1,%ymm4,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,221 ; vpmovzxbw %xmm5,%ymm11 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,239,1 ; vextracti128 $0x1,%ymm5,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 + DB 197,181,213,237 ; vpmullw %ymm5,%ymm9,%ymm5 + DB 197,205,213,255 ; vpmullw %ymm7,%ymm6,%ymm7 + DB 197,173,213,228 ; vpmullw %ymm4,%ymm10,%ymm4 DB 196,65,61,213,219 ; vpmullw %ymm11,%ymm8,%ymm11 - DB 197,197,213,228 ; vpmullw %ymm4,%ymm7,%ymm4 - DB 196,65,53,213,210 ; vpmullw %ymm10,%ymm9,%ymm10 - DB 196,65,45,253,201 ; vpaddw %ymm9,%ymm10,%ymm9 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 196,193,37,253,248 ; vpaddw %ymm8,%ymm11,%ymm7 - DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 196,65,37,253,192 ; vpaddw %ymm8,%ymm11,%ymm8 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 + DB 197,197,253,246 ; vpaddw %ymm6,%ymm7,%ymm6 + DB 196,193,85,253,233 ; vpaddw %ymm9,%ymm5,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,209,8 ; vpsrlw $0x8,%ymm9,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 + DB 196,99,93,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm4,%ymm8 + DB 196,227,93,70,231,49 ; vperm2i128 $0x31,%ymm7,%ymm4,%ymm4 DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + DB 196,227,85,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm5,%ymm7 + DB 196,227,85,70,238,49 ; vperm2i128 $0x31,%ymm6,%ymm5,%ymm5 DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 + DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 - DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstover_hsw_8bit _sk_dstover_hsw_8bit LABEL PROC - DB 197,253,111,37,62,124,0,0 ; vmovdqa 0x7c3e(%rip),%ymm4 # 2d140 <_sk_overlay_sse2_8bit+0xbe8> + DB 197,253,111,37,92,121,0,0 ; vmovdqa 0x795c(%rip),%ymm4 # 2d020 <_sk_overlay_sse2_8bit+0xb71> DB 196,226,101,0,236 ; vpshufb %ymm4,%ymm3,%ymm5 DB 196,226,109,0,228 ; vpshufb %ymm4,%ymm2,%ymm4 - DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 - DB 196,227,125,57,199,1 ; vextracti128 $0x1,%ymm0,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,193 ; vpmovzxbw %xmm1,%ymm8 + DB 196,227,125,57,198,1 ; vextracti128 $0x1,%ymm0,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 + DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,212 ; vpmovzxbw %xmm4,%ymm10 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 + DB 196,227,125,57,231,1 ; vextracti128 $0x1,%ymm4,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,221 ; vpmovzxbw %xmm5,%ymm11 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,239,1 ; vextracti128 $0x1,%ymm5,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,213,213,238 ; vpmullw %ymm6,%ymm5,%ymm5 + DB 196,193,85,213,233 ; vpmullw %ymm9,%ymm5,%ymm5 + DB 197,197,213,254 ; vpmullw %ymm6,%ymm7,%ymm7 + DB 196,193,93,213,226 ; vpmullw %ymm10,%ymm4,%ymm4 DB 196,65,37,213,216 ; vpmullw %ymm8,%ymm11,%ymm11 - DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 196,65,45,213,209 ; vpmullw %ymm9,%ymm10,%ymm10 - DB 196,65,45,253,201 ; vpaddw %ymm9,%ymm10,%ymm9 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 196,193,37,253,248 ; vpaddw %ymm8,%ymm11,%ymm7 - DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 196,65,37,253,192 ; vpaddw %ymm8,%ymm11,%ymm8 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 + DB 197,197,253,246 ; vpaddw %ymm6,%ymm7,%ymm6 + DB 196,193,85,253,233 ; vpaddw %ymm9,%ymm5,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,209,8 ; vpsrlw $0x8,%ymm9,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 + DB 196,99,93,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm4,%ymm8 + DB 196,227,93,70,231,49 ; vperm2i128 $0x31,%ymm7,%ymm4,%ymm4 DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + DB 196,227,85,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm5,%ymm7 + DB 196,227,85,70,238,49 ; vperm2i128 $0x31,%ymm6,%ymm5,%ymm5 DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 + DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 - DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_modulate_hsw_8bit _sk_modulate_hsw_8bit LABEL PROC - DB 196,226,125,48,224 ; vpmovzxbw %xmm0,%ymm4 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,226,125,48,233 ; vpmovzxbw %xmm1,%ymm5 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,205,1 ; vextracti128 $0x1,%ymm1,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 - DB 196,227,125,57,215,1 ; vextracti128 $0x1,%ymm2,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,195 ; vpmovzxbw %xmm3,%ymm8 + DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 + DB 196,226,125,48,250 ; vpmovzxbw %xmm2,%ymm7 DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 197,205,213,241 ; vpmullw %ymm1,%ymm6,%ymm6 - DB 197,61,213,197 ; vpmullw %ymm5,%ymm8,%ymm8 + DB 196,98,125,48,203 ; vpmovzxbw %xmm3,%ymm9 + DB 197,53,213,201 ; vpmullw %ymm1,%ymm9,%ymm9 + DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 DB 197,197,213,248 ; vpmullw %ymm0,%ymm7,%ymm7 - DB 197,53,213,204 ; vpmullw %ymm4,%ymm9,%ymm9 - DB 197,181,253,228 ; vpaddw %ymm4,%ymm9,%ymm4 + DB 197,61,213,196 ; vpmullw %ymm4,%ymm8,%ymm8 + DB 197,189,253,228 ; vpaddw %ymm4,%ymm8,%ymm4 DB 197,197,253,192 ; vpaddw %ymm0,%ymm7,%ymm0 - DB 197,189,253,237 ; vpaddw %ymm5,%ymm8,%ymm5 - DB 197,205,253,201 ; vpaddw %ymm1,%ymm6,%ymm1 + DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 + DB 197,181,253,201 ; vpaddw %ymm1,%ymm9,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,227,93,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm4,%ymm6 - DB 196,227,93,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm4,%ymm0 + DB 196,227,125,56,244,1 ; vinserti128 $0x1,%xmm4,%ymm0,%ymm6 + DB 196,227,125,70,196,49 ; vperm2i128 $0x31,%ymm4,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,85,56,225,1 ; vinserti128 $0x1,%xmm1,%ymm5,%ymm4 - DB 196,227,85,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm5,%ymm1 + DB 196,227,117,56,229,1 ; vinserti128 $0x1,%xmm5,%ymm1,%ymm4 + DB 196,227,117,70,205,49 ; vperm2i128 $0x31,%ymm5,%ymm1,%ymm1 DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_multiply_hsw_8bit _sk_multiply_hsw_8bit LABEL PROC - DB 72,131,236,56 ; sub $0x38,%rsp - DB 197,253,111,243 ; vmovdqa %ymm3,%ymm6 - DB 197,253,111,218 ; vmovdqa %ymm2,%ymm3 - DB 197,125,111,13,251,122,0,0 ; vmovdqa 0x7afb(%rip),%ymm9 # 2d160 <_sk_overlay_sse2_8bit+0xc08> - DB 196,194,101,0,225 ; vpshufb %ymm9,%ymm3,%ymm4 - DB 196,194,77,0,233 ; vpshufb %ymm9,%ymm6,%ymm5 - DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 - DB 196,65,85,239,194 ; vpxor %ymm10,%ymm5,%ymm8 - DB 196,65,93,239,218 ; vpxor %ymm10,%ymm4,%ymm11 - DB 196,98,125,48,232 ; vpmovzxbw %xmm0,%ymm13 - DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 - DB 196,98,125,48,245 ; vpmovzxbw %xmm5,%ymm14 - DB 196,226,125,48,209 ; vpmovzxbw %xmm1,%ymm2 - DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,66,125,48,227 ; vpmovzxbw %xmm11,%ymm12 - DB 196,99,125,57,220,1 ; vextracti128 $0x1,%ymm11,%xmm4 + DB 197,125,111,21,35,120,0,0 ; vmovdqa 0x7823(%rip),%ymm10 # 2d040 <_sk_overlay_sse2_8bit+0xb91> + DB 196,194,109,0,226 ; vpshufb %ymm10,%ymm2,%ymm4 + DB 196,194,101,0,234 ; vpshufb %ymm10,%ymm3,%ymm5 + DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 + DB 196,65,85,239,195 ; vpxor %ymm11,%ymm5,%ymm8 + DB 196,65,93,239,203 ; vpxor %ymm11,%ymm4,%ymm9 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,236 ; vpmovzxbw %xmm4,%ymm13 + DB 196,98,125,48,240 ; vpmovzxbw %xmm0,%ymm14 + DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 + DB 196,98,125,48,254 ; vpmovzxbw %xmm6,%ymm15 + DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 + DB 196,99,125,57,204,1 ; vextracti128 $0x1,%ymm9,%xmm4 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,66,125,48,216 ; vpmovzxbw %xmm8,%ymm11 + DB 196,66,125,48,201 ; vpmovzxbw %xmm9,%ymm9 DB 196,99,125,57,197,1 ; vextracti128 $0x1,%ymm8,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,213,213,239 ; vpmullw %ymm7,%ymm5,%ymm5 - DB 197,37,213,194 ; vpmullw %ymm2,%ymm11,%ymm8 - DB 196,193,93,213,230 ; vpmullw %ymm14,%ymm4,%ymm4 - DB 196,65,29,213,221 ; vpmullw %ymm13,%ymm12,%ymm11 - DB 196,65,37,253,221 ; vpaddw %ymm13,%ymm11,%ymm11 - DB 196,193,93,253,230 ; vpaddw %ymm14,%ymm4,%ymm4 - DB 197,61,253,194 ; vpaddw %ymm2,%ymm8,%ymm8 - DB 197,213,253,239 ; vpaddw %ymm7,%ymm5,%ymm5 - DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 196,66,125,48,192 ; vpmovzxbw %xmm8,%ymm8 + DB 197,61,213,199 ; vpmullw %ymm7,%ymm8,%ymm8 + DB 196,193,85,213,239 ; vpmullw %ymm15,%ymm5,%ymm5 + DB 196,65,53,213,206 ; vpmullw %ymm14,%ymm9,%ymm9 + DB 196,193,93,213,229 ; vpmullw %ymm13,%ymm4,%ymm4 + DB 196,193,93,253,229 ; vpaddw %ymm13,%ymm4,%ymm4 + DB 196,65,53,253,206 ; vpaddw %ymm14,%ymm9,%ymm9 + DB 196,193,85,253,239 ; vpaddw %ymm15,%ymm5,%ymm5 + DB 197,61,253,199 ; vpaddw %ymm7,%ymm8,%ymm8 DB 196,193,29,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm12 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 196,193,61,113,209,8 ; vpsrlw $0x8,%ymm9,%ymm8 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,61,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm8 - DB 196,99,61,56,220,1 ; vinserti128 $0x1,%xmm4,%ymm8,%ymm11 + DB 196,99,61,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm8,%ymm9 DB 196,227,61,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm8,%ymm4 - DB 197,37,103,196 ; vpackuswb %ymm4,%ymm11,%ymm8 + DB 197,53,103,196 ; vpackuswb %ymm4,%ymm9,%ymm8 DB 196,227,29,56,229,1 ; vinserti128 $0x1,%xmm5,%ymm12,%ymm4 DB 196,227,29,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm12,%ymm5 - DB 197,93,103,221 ; vpackuswb %ymm5,%ymm4,%ymm11 - DB 196,194,125,0,193 ; vpshufb %ymm9,%ymm0,%ymm0 - DB 196,194,117,0,201 ; vpshufb %ymm9,%ymm1,%ymm1 - DB 196,65,117,239,226 ; vpxor %ymm10,%ymm1,%ymm12 - DB 196,193,125,239,226 ; vpxor %ymm10,%ymm0,%ymm4 - DB 196,226,125,48,195 ; vpmovzxbw %xmm3,%ymm0 - DB 196,227,125,57,217,1 ; vextracti128 $0x1,%ymm3,%xmm1 - DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,254,127,52,36 ; vmovdqu %ymm6,(%rsp) - DB 196,98,125,48,206 ; vpmovzxbw %xmm6,%ymm9 - DB 196,227,125,57,245,1 ; vextracti128 $0x1,%ymm6,%xmm5 + DB 197,93,103,205 ; vpackuswb %ymm5,%ymm4,%ymm9 + DB 196,194,125,0,194 ; vpshufb %ymm10,%ymm0,%ymm0 + DB 196,194,117,0,202 ; vpshufb %ymm10,%ymm1,%ymm1 + DB 196,65,117,239,227 ; vpxor %ymm11,%ymm1,%ymm12 + DB 196,193,125,239,227 ; vpxor %ymm11,%ymm0,%ymm4 + DB 196,227,125,57,208,1 ; vextracti128 $0x1,%ymm2,%xmm0 + DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 + DB 196,226,125,48,202 ; vpmovzxbw %xmm2,%ymm1 + DB 196,227,125,57,221,1 ; vextracti128 $0x1,%ymm3,%xmm5 DB 196,98,125,48,213 ; vpmovzxbw %xmm5,%ymm10 - DB 196,226,125,48,236 ; vpmovzxbw %xmm4,%ymm5 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 197,245,213,228 ; vpmullw %ymm4,%ymm1,%ymm4 DB 197,253,213,237 ; vpmullw %ymm5,%ymm0,%ymm5 @@ -35782,47 +35801,45 @@ _sk_multiply_hsw_8bit LABEL PROC DB 197,221,253,225 ; vpaddw %ymm1,%ymm4,%ymm4 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 196,99,85,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm15 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 196,194,125,48,236 ; vpmovzxbw %xmm12,%ymm5 + DB 196,99,93,56,221,1 ; vinserti128 $0x1,%xmm5,%ymm4,%ymm11 + DB 196,227,93,70,229,49 ; vperm2i128 $0x31,%ymm5,%ymm4,%ymm4 + DB 196,226,125,48,235 ; vpmovzxbw %xmm3,%ymm5 + DB 197,165,103,228 ; vpackuswb %ymm4,%ymm11,%ymm4 DB 196,99,125,57,230,1 ; vextracti128 $0x1,%ymm12,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,66,125,48,220 ; vpmovzxbw %xmm12,%ymm11 + DB 196,65,85,213,219 ; vpmullw %ymm11,%ymm5,%ymm11 DB 197,173,213,246 ; vpmullw %ymm6,%ymm10,%ymm6 - DB 197,181,213,237 ; vpmullw %ymm5,%ymm9,%ymm5 - DB 196,193,85,253,233 ; vpaddw %ymm9,%ymm5,%ymm5 DB 196,193,77,253,242 ; vpaddw %ymm10,%ymm6,%ymm6 + DB 197,37,253,221 ; vpaddw %ymm5,%ymm11,%ymm11 + DB 196,193,37,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm11 DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,133,103,228 ; vpackuswb %ymm4,%ymm15,%ymm4 - DB 196,99,85,56,230,1 ; vinserti128 $0x1,%xmm6,%ymm5,%ymm12 - DB 196,227,85,70,238,49 ; vperm2i128 $0x31,%ymm6,%ymm5,%ymm5 - DB 197,157,103,237 ; vpackuswb %ymm5,%ymm12,%ymm5 - DB 196,193,85,252,235 ; vpaddb %ymm11,%ymm5,%ymm5 DB 196,193,93,252,224 ; vpaddb %ymm8,%ymm4,%ymm4 + DB 196,99,37,56,198,1 ; vinserti128 $0x1,%xmm6,%ymm11,%ymm8 + DB 196,227,37,70,246,49 ; vperm2i128 $0x31,%ymm6,%ymm11,%ymm6 + DB 197,189,103,246 ; vpackuswb %ymm6,%ymm8,%ymm6 + DB 196,193,77,252,241 ; vpaddb %ymm9,%ymm6,%ymm6 DB 196,193,125,213,197 ; vpmullw %ymm13,%ymm0,%ymm0 DB 196,193,125,253,197 ; vpaddw %ymm13,%ymm0,%ymm0 DB 196,193,117,213,206 ; vpmullw %ymm14,%ymm1,%ymm1 DB 196,193,117,253,206 ; vpaddw %ymm14,%ymm1,%ymm1 - DB 197,181,213,242 ; vpmullw %ymm2,%ymm9,%ymm6 - DB 197,205,253,210 ; vpaddw %ymm2,%ymm6,%ymm2 - DB 197,173,213,247 ; vpmullw %ymm7,%ymm10,%ymm6 - DB 197,205,253,247 ; vpaddw %ymm7,%ymm6,%ymm6 + DB 196,65,45,213,199 ; vpmullw %ymm15,%ymm10,%ymm8 + DB 196,65,61,253,199 ; vpaddw %ymm15,%ymm8,%ymm8 + DB 197,213,213,239 ; vpmullw %ymm7,%ymm5,%ymm5 + DB 197,213,253,239 ; vpaddw %ymm7,%ymm5,%ymm5 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 - DB 196,227,125,56,249,1 ; vinserti128 $0x1,%xmm1,%ymm0,%ymm7 - DB 196,227,125,70,193,49 ; vperm2i128 $0x31,%ymm1,%ymm0,%ymm0 - DB 197,245,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm1 - DB 197,237,113,210,8 ; vpsrlw $0x8,%ymm2,%ymm2 + DB 196,227,117,56,248,1 ; vinserti128 $0x1,%xmm0,%ymm1,%ymm7 + DB 196,227,117,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm1,%ymm0 DB 197,197,103,192 ; vpackuswb %ymm0,%ymm7,%ymm0 - DB 196,227,109,56,241,1 ; vinserti128 $0x1,%xmm1,%ymm2,%ymm6 - DB 196,227,109,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm2,%ymm1 - DB 197,205,103,201 ; vpackuswb %ymm1,%ymm6,%ymm1 + DB 197,245,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm1 + DB 196,193,85,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm5 DB 197,221,252,192 ; vpaddb %ymm0,%ymm4,%ymm0 - DB 197,213,252,201 ; vpaddb %ymm1,%ymm5,%ymm1 + DB 196,227,117,56,229,1 ; vinserti128 $0x1,%xmm5,%ymm1,%ymm4 + DB 196,227,117,70,205,49 ; vperm2i128 $0x31,%ymm5,%ymm1,%ymm1 + DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 + DB 197,205,252,201 ; vpaddb %ymm1,%ymm6,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,253,111,211 ; vmovdqa %ymm3,%ymm2 - DB 197,252,16,28,36 ; vmovups (%rsp),%ymm3 - DB 72,131,196,56 ; add $0x38,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_screen_hsw_8bit @@ -35830,35 +35847,35 @@ _sk_screen_hsw_8bit LABEL PROC DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 DB 197,245,239,236 ; vpxor %ymm4,%ymm1,%ymm5 DB 197,253,239,228 ; vpxor %ymm4,%ymm0,%ymm4 - DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,230,1 ; vextracti128 $0x1,%ymm4,%xmm6 + DB 196,98,125,48,198 ; vpmovzxbw %xmm6,%ymm8 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,226,125,48,253 ; vpmovzxbw %xmm5,%ymm7 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,239,1 ; vextracti128 $0x1,%ymm5,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 - DB 196,98,125,48,222 ; vpmovzxbw %xmm6,%ymm11 - DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,98,125,48,214 ; vpmovzxbw %xmm6,%ymm10 + DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 - DB 197,45,213,215 ; vpmullw %ymm7,%ymm10,%ymm10 - DB 197,37,213,220 ; vpmullw %ymm4,%ymm11,%ymm11 - DB 196,65,61,213,193 ; vpmullw %ymm9,%ymm8,%ymm8 - DB 196,65,61,253,193 ; vpaddw %ymm9,%ymm8,%ymm8 - DB 197,165,253,228 ; vpaddw %ymm4,%ymm11,%ymm4 - DB 197,173,253,255 ; vpaddw %ymm7,%ymm10,%ymm7 - DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 + DB 196,98,125,48,219 ; vpmovzxbw %xmm3,%ymm11 + DB 197,37,213,221 ; vpmullw %ymm5,%ymm11,%ymm11 + DB 197,205,213,247 ; vpmullw %ymm7,%ymm6,%ymm6 + DB 197,53,213,204 ; vpmullw %ymm4,%ymm9,%ymm9 + DB 196,65,45,213,208 ; vpmullw %ymm8,%ymm10,%ymm10 + DB 196,65,45,253,192 ; vpaddw %ymm8,%ymm10,%ymm8 + DB 197,181,253,228 ; vpaddw %ymm4,%ymm9,%ymm4 + DB 197,205,253,247 ; vpaddw %ymm7,%ymm6,%ymm6 + DB 197,165,253,237 ; vpaddw %ymm5,%ymm11,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + DB 196,99,93,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm4,%ymm8 + DB 196,227,93,70,231,49 ; vperm2i128 $0x31,%ymm7,%ymm4,%ymm4 DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + DB 196,227,85,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm5,%ymm7 + DB 196,227,85,70,238,49 ; vperm2i128 $0x31,%ymm6,%ymm5,%ymm5 DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 DB 197,221,252,192 ; vpaddb %ymm0,%ymm4,%ymm0 DB 197,213,252,201 ; vpaddb %ymm1,%ymm5,%ymm1 @@ -35867,78 +35884,78 @@ _sk_screen_hsw_8bit LABEL PROC PUBLIC _sk_xor__hsw_8bit _sk_xor__hsw_8bit LABEL PROC - DB 197,125,111,13,135,120,0,0 ; vmovdqa 0x7887(%rip),%ymm9 # 2d180 <_sk_overlay_sse2_8bit+0xc28> - DB 196,194,109,0,225 ; vpshufb %ymm9,%ymm2,%ymm4 - DB 196,194,101,0,249 ; vpshufb %ymm9,%ymm3,%ymm7 - DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 - DB 196,193,69,239,251 ; vpxor %ymm11,%ymm7,%ymm7 - DB 196,193,93,239,227 ; vpxor %ymm11,%ymm4,%ymm4 - DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 - DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 197,125,111,5,187,117,0,0 ; vmovdqa 0x75bb(%rip),%ymm8 # 2d060 <_sk_overlay_sse2_8bit+0xbb1> + DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 + DB 196,194,101,0,248 ; vpshufb %ymm8,%ymm3,%ymm7 + DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 + DB 196,193,69,239,250 ; vpxor %ymm10,%ymm7,%ymm7 + DB 196,193,77,239,242 ; vpxor %ymm10,%ymm6,%ymm6 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 + DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 + DB 196,227,125,57,205,1 ; vextracti128 $0x1,%ymm1,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,209 ; vpmovzxbw %xmm1,%ymm10 - DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 + DB 196,98,125,48,217 ; vpmovzxbw %xmm1,%ymm11 + DB 196,227,125,57,244,1 ; vextracti128 $0x1,%ymm6,%xmm4 + DB 196,98,125,48,236 ; vpmovzxbw %xmm4,%ymm13 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,252,1 ; vextracti128 $0x1,%ymm7,%xmm4 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,239 ; vpmovzxbw %xmm7,%ymm13 - DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 197,197,213,254 ; vpmullw %ymm6,%ymm7,%ymm7 - DB 196,65,21,213,234 ; vpmullw %ymm10,%ymm13,%ymm13 + DB 196,193,69,213,251 ; vpmullw %ymm11,%ymm7,%ymm7 DB 197,221,213,229 ; vpmullw %ymm5,%ymm4,%ymm4 - DB 196,65,29,213,224 ; vpmullw %ymm8,%ymm12,%ymm12 - DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 + DB 196,193,77,213,241 ; vpmullw %ymm9,%ymm6,%ymm6 + DB 196,65,21,213,236 ; vpmullw %ymm12,%ymm13,%ymm13 + DB 196,65,21,253,228 ; vpaddw %ymm12,%ymm13,%ymm12 + DB 196,193,77,253,241 ; vpaddw %ymm9,%ymm6,%ymm6 DB 197,221,253,229 ; vpaddw %ymm5,%ymm4,%ymm4 - DB 196,193,21,253,234 ; vpaddw %ymm10,%ymm13,%ymm5 - DB 197,197,253,246 ; vpaddw %ymm6,%ymm7,%ymm6 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 196,193,69,253,235 ; vpaddw %ymm11,%ymm7,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 - DB 197,61,103,212 ; vpackuswb %ymm4,%ymm8,%ymm10 - DB 196,227,85,56,254,1 ; vinserti128 $0x1,%xmm6,%ymm5,%ymm7 - DB 196,227,85,70,238,49 ; vperm2i128 $0x31,%ymm6,%ymm5,%ymm5 - DB 197,197,103,253 ; vpackuswb %ymm5,%ymm7,%ymm7 - DB 196,194,125,0,193 ; vpshufb %ymm9,%ymm0,%ymm0 - DB 196,194,117,0,201 ; vpshufb %ymm9,%ymm1,%ymm1 - DB 196,193,117,239,203 ; vpxor %ymm11,%ymm1,%ymm1 - DB 196,193,125,239,195 ; vpxor %ymm11,%ymm0,%ymm0 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 196,193,69,113,212,8 ; vpsrlw $0x8,%ymm12,%ymm7 + DB 196,99,77,56,207,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm9 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,53,103,222 ; vpackuswb %ymm6,%ymm9,%ymm11 + DB 196,227,85,56,244,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm6 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,77,103,204 ; vpackuswb %ymm4,%ymm6,%ymm9 + DB 196,194,125,0,192 ; vpshufb %ymm8,%ymm0,%ymm0 + DB 196,194,117,0,200 ; vpshufb %ymm8,%ymm1,%ymm1 + DB 196,193,117,239,202 ; vpxor %ymm10,%ymm1,%ymm1 + DB 196,193,125,239,194 ; vpxor %ymm10,%ymm0,%ymm0 + DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,234 ; vpmovzxbw %xmm2,%ymm5 - DB 196,227,125,57,214,1 ; vextracti128 $0x1,%ymm2,%xmm6 + DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 DB 196,98,125,48,195 ; vpmovzxbw %xmm3,%ymm8 - DB 196,227,125,57,220,1 ; vextracti128 $0x1,%ymm3,%xmm4 - DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 - DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,125,57,199,1 ; vextracti128 $0x1,%ymm0,%xmm7 + DB 196,98,125,48,215 ; vpmovzxbw %xmm7,%ymm10 DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 - DB 196,98,125,48,217 ; vpmovzxbw %xmm1,%ymm11 - DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,221,213,201 ; vpmullw %ymm1,%ymm4,%ymm1 - DB 196,65,61,213,219 ; vpmullw %ymm11,%ymm8,%ymm11 - DB 197,205,213,192 ; vpmullw %ymm0,%ymm6,%ymm0 - DB 196,65,85,213,201 ; vpmullw %ymm9,%ymm5,%ymm9 - DB 197,181,253,237 ; vpaddw %ymm5,%ymm9,%ymm5 - DB 197,253,253,198 ; vpaddw %ymm6,%ymm0,%ymm0 - DB 196,193,37,253,240 ; vpaddw %ymm8,%ymm11,%ymm6 - DB 197,245,253,204 ; vpaddw %ymm4,%ymm1,%ymm1 + DB 197,189,213,201 ; vpmullw %ymm1,%ymm8,%ymm1 + DB 197,205,213,255 ; vpmullw %ymm7,%ymm6,%ymm7 + DB 197,213,213,192 ; vpmullw %ymm0,%ymm5,%ymm0 + DB 196,65,93,213,210 ; vpmullw %ymm10,%ymm4,%ymm10 + DB 197,173,253,228 ; vpaddw %ymm4,%ymm10,%ymm4 + DB 197,253,253,197 ; vpaddw %ymm5,%ymm0,%ymm0 + DB 197,197,253,238 ; vpaddw %ymm6,%ymm7,%ymm5 + DB 196,193,117,253,200 ; vpaddw %ymm8,%ymm1,%ymm1 DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 - DB 197,221,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm4 - DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 196,227,85,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm5,%ymm6 - DB 196,227,85,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm5,%ymm0 + DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 + DB 196,227,125,56,244,1 ; vinserti128 $0x1,%xmm4,%ymm0,%ymm6 + DB 196,227,125,70,196,49 ; vperm2i128 $0x31,%ymm4,%ymm0,%ymm0 DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 - DB 196,227,93,56,233,1 ; vinserti128 $0x1,%xmm1,%ymm4,%ymm5 - DB 196,227,93,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm4,%ymm1 - DB 197,213,103,201 ; vpackuswb %ymm1,%ymm5,%ymm1 - DB 196,193,125,252,194 ; vpaddb %ymm10,%ymm0,%ymm0 - DB 197,245,252,207 ; vpaddb %ymm7,%ymm1,%ymm1 + DB 196,193,125,252,195 ; vpaddb %ymm11,%ymm0,%ymm0 + DB 196,227,117,56,229,1 ; vinserti128 $0x1,%xmm5,%ymm1,%ymm4 + DB 196,227,117,70,205,49 ; vperm2i128 $0x31,%ymm5,%ymm1,%ymm1 + DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 + DB 196,193,117,252,201 ; vpaddb %ymm9,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -35951,633 +35968,572 @@ _sk_plus__hsw_8bit LABEL PROC PUBLIC _sk_darken_hsw_8bit _sk_darken_hsw_8bit LABEL PROC - DB 197,125,111,5,41,119,0,0 ; vmovdqa 0x7729(%rip),%ymm8 # 2d1a0 <_sk_overlay_sse2_8bit+0xc48> - DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 - DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 - DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 - DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 197,125,111,13,91,116,0,0 ; vmovdqa 0x745b(%rip),%ymm9 # 2d080 <_sk_overlay_sse2_8bit+0xbd1> + DB 196,194,101,0,225 ; vpshufb %ymm9,%ymm3,%ymm4 + DB 196,194,109,0,241 ; vpshufb %ymm9,%ymm2,%ymm6 + DB 196,227,125,57,199,1 ; vextracti128 $0x1,%ymm0,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 + DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 + DB 196,227,125,57,205,1 ; vextracti128 $0x1,%ymm1,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,222 ; vpmovzxbw %xmm6,%ymm11 - DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 + DB 196,98,125,48,209 ; vpmovzxbw %xmm1,%ymm10 + DB 196,227,125,57,247,1 ; vextracti128 $0x1,%ymm6,%xmm7 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,231,1 ; vextracti128 $0x1,%ymm4,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 196,65,29,213,225 ; vpmullw %ymm9,%ymm12,%ymm12 - DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 - DB 196,65,37,213,218 ; vpmullw %ymm10,%ymm11,%ymm11 - DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 - DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 - DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 + DB 196,193,93,213,226 ; vpmullw %ymm10,%ymm4,%ymm4 + DB 197,197,213,253 ; vpmullw %ymm5,%ymm7,%ymm7 + DB 196,193,77,213,240 ; vpmullw %ymm8,%ymm6,%ymm6 + DB 196,65,29,213,227 ; vpmullw %ymm11,%ymm12,%ymm12 + DB 196,65,29,253,219 ; vpaddw %ymm11,%ymm12,%ymm11 + DB 196,193,77,253,240 ; vpaddw %ymm8,%ymm6,%ymm6 + DB 197,197,253,237 ; vpaddw %ymm5,%ymm7,%ymm5 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 - DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 - DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 197,53,103,204 ; vpackuswb %ymm4,%ymm9,%ymm9 - DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 - DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 - DB 197,85,103,222 ; vpackuswb %ymm6,%ymm5,%ymm11 - DB 196,194,117,0,232 ; vpshufb %ymm8,%ymm1,%ymm5 - DB 196,194,125,0,248 ; vpshufb %ymm8,%ymm0,%ymm7 - DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 197,221,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm4 + DB 196,193,77,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm6 + DB 196,99,93,56,198,1 ; vinserti128 $0x1,%xmm6,%ymm4,%ymm8 + DB 196,227,93,70,230,49 ; vperm2i128 $0x31,%ymm6,%ymm4,%ymm4 + DB 197,61,103,196 ; vpackuswb %ymm4,%ymm8,%ymm8 + DB 196,227,69,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm7,%ymm6 + DB 196,227,69,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm7,%ymm5 + DB 197,77,103,213 ; vpackuswb %ymm5,%ymm6,%ymm10 + DB 196,194,117,0,233 ; vpshufb %ymm9,%ymm1,%ymm5 + DB 196,194,125,0,249 ; vpshufb %ymm9,%ymm0,%ymm7 DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 - DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 + DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 - DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 + DB 196,98,125,48,219 ; vpmovzxbw %xmm3,%ymm11 + DB 196,227,125,57,252,1 ; vextracti128 $0x1,%ymm7,%xmm4 + DB 196,98,125,48,236 ; vpmovzxbw %xmm4,%ymm13 DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,237 ; vpmovzxbw %xmm5,%ymm13 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,236,1 ; vextracti128 $0x1,%ymm5,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 - DB 196,65,45,213,237 ; vpmullw %ymm13,%ymm10,%ymm13 - DB 197,221,213,255 ; vpmullw %ymm7,%ymm4,%ymm7 - DB 196,65,61,213,228 ; vpmullw %ymm12,%ymm8,%ymm12 - DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 - DB 197,197,253,228 ; vpaddw %ymm4,%ymm7,%ymm4 - DB 196,193,21,253,250 ; vpaddw %ymm10,%ymm13,%ymm7 - DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 197,165,213,237 ; vpmullw %ymm5,%ymm11,%ymm5 + DB 197,205,213,228 ; vpmullw %ymm4,%ymm6,%ymm4 + DB 197,181,213,255 ; vpmullw %ymm7,%ymm9,%ymm7 + DB 196,65,29,213,237 ; vpmullw %ymm13,%ymm12,%ymm13 + DB 196,65,21,253,228 ; vpaddw %ymm12,%ymm13,%ymm12 + DB 196,193,69,253,249 ; vpaddw %ymm9,%ymm7,%ymm7 + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 + DB 196,193,85,253,235 ; vpaddw %ymm11,%ymm5,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 - DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 - DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 - DB 197,181,222,244 ; vpmaxub %ymm4,%ymm9,%ymm6 - DB 197,165,222,253 ; vpmaxub %ymm5,%ymm11,%ymm7 + DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 196,193,69,113,212,8 ; vpsrlw $0x8,%ymm12,%ymm7 + DB 196,99,77,56,207,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm9 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,181,103,246 ; vpackuswb %ymm6,%ymm9,%ymm6 + DB 196,227,85,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm7 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 + DB 197,189,222,238 ; vpmaxub %ymm6,%ymm8,%ymm5 + DB 197,173,222,252 ; vpmaxub %ymm4,%ymm10,%ymm7 DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 DB 197,245,248,255 ; vpsubb %ymm7,%ymm1,%ymm7 - DB 197,253,248,246 ; vpsubb %ymm6,%ymm0,%ymm6 - DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 - DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 - DB 196,226,125,88,37,102,120,0,0 ; vpbroadcastd 0x7866(%rip),%ymm4 # 2d44c <_sk_overlay_sse2_8bit+0xef4> - DB 196,227,125,76,198,64 ; vpblendvb %ymm4,%ymm6,%ymm0,%ymm0 + DB 197,253,248,237 ; vpsubb %ymm5,%ymm0,%ymm5 + DB 197,245,248,204 ; vpsubb %ymm4,%ymm1,%ymm1 + DB 197,253,248,198 ; vpsubb %ymm6,%ymm0,%ymm0 + DB 196,226,125,88,37,162,117,0,0 ; vpbroadcastd 0x75a2(%rip),%ymm4 # 2d338 <_sk_overlay_sse2_8bit+0xe89> + DB 196,227,125,76,197,64 ; vpblendvb %ymm4,%ymm5,%ymm0,%ymm0 DB 196,227,117,76,207,64 ; vpblendvb %ymm4,%ymm7,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_lighten_hsw_8bit _sk_lighten_hsw_8bit LABEL PROC - DB 197,125,111,5,194,117,0,0 ; vmovdqa 0x75c2(%rip),%ymm8 # 2d1c0 <_sk_overlay_sse2_8bit+0xc68> - DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 - DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 - DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 - DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 197,125,111,13,242,114,0,0 ; vmovdqa 0x72f2(%rip),%ymm9 # 2d0a0 <_sk_overlay_sse2_8bit+0xbf1> + DB 196,194,101,0,225 ; vpshufb %ymm9,%ymm3,%ymm4 + DB 196,194,109,0,241 ; vpshufb %ymm9,%ymm2,%ymm6 + DB 196,227,125,57,199,1 ; vextracti128 $0x1,%ymm0,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 + DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 + DB 196,227,125,57,205,1 ; vextracti128 $0x1,%ymm1,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,222 ; vpmovzxbw %xmm6,%ymm11 - DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 + DB 196,98,125,48,209 ; vpmovzxbw %xmm1,%ymm10 + DB 196,227,125,57,247,1 ; vextracti128 $0x1,%ymm6,%xmm7 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,227,125,57,231,1 ; vextracti128 $0x1,%ymm4,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 196,65,29,213,225 ; vpmullw %ymm9,%ymm12,%ymm12 - DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 - DB 196,65,37,213,218 ; vpmullw %ymm10,%ymm11,%ymm11 - DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 - DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 - DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 + DB 196,193,93,213,226 ; vpmullw %ymm10,%ymm4,%ymm4 + DB 197,197,213,253 ; vpmullw %ymm5,%ymm7,%ymm7 + DB 196,193,77,213,240 ; vpmullw %ymm8,%ymm6,%ymm6 + DB 196,65,29,213,227 ; vpmullw %ymm11,%ymm12,%ymm12 + DB 196,65,29,253,219 ; vpaddw %ymm11,%ymm12,%ymm11 + DB 196,193,77,253,240 ; vpaddw %ymm8,%ymm6,%ymm6 + DB 197,197,253,237 ; vpaddw %ymm5,%ymm7,%ymm5 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 - DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 - DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 - DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 197,53,103,204 ; vpackuswb %ymm4,%ymm9,%ymm9 - DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 - DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 - DB 197,85,103,222 ; vpackuswb %ymm6,%ymm5,%ymm11 - DB 196,194,117,0,232 ; vpshufb %ymm8,%ymm1,%ymm5 - DB 196,194,125,0,248 ; vpshufb %ymm8,%ymm0,%ymm7 - DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 197,221,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm4 + DB 196,193,77,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm6 + DB 196,99,93,56,198,1 ; vinserti128 $0x1,%xmm6,%ymm4,%ymm8 + DB 196,227,93,70,230,49 ; vperm2i128 $0x31,%ymm6,%ymm4,%ymm4 + DB 197,61,103,196 ; vpackuswb %ymm4,%ymm8,%ymm8 + DB 196,227,69,56,245,1 ; vinserti128 $0x1,%xmm5,%ymm7,%ymm6 + DB 196,227,69,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm7,%ymm5 + DB 197,77,103,213 ; vpackuswb %ymm5,%ymm6,%ymm10 + DB 196,194,117,0,233 ; vpshufb %ymm9,%ymm1,%ymm5 + DB 196,194,125,0,249 ; vpshufb %ymm9,%ymm0,%ymm7 DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 - DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 + DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 - DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 + DB 196,98,125,48,219 ; vpmovzxbw %xmm3,%ymm11 + DB 196,227,125,57,252,1 ; vextracti128 $0x1,%ymm7,%xmm4 + DB 196,98,125,48,236 ; vpmovzxbw %xmm4,%ymm13 DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,237 ; vpmovzxbw %xmm5,%ymm13 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,227,125,57,236,1 ; vextracti128 $0x1,%ymm5,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 - DB 196,65,45,213,237 ; vpmullw %ymm13,%ymm10,%ymm13 - DB 197,221,213,255 ; vpmullw %ymm7,%ymm4,%ymm7 - DB 196,65,61,213,228 ; vpmullw %ymm12,%ymm8,%ymm12 - DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 - DB 197,197,253,228 ; vpaddw %ymm4,%ymm7,%ymm4 - DB 196,193,21,253,250 ; vpaddw %ymm10,%ymm13,%ymm7 - DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 197,165,213,237 ; vpmullw %ymm5,%ymm11,%ymm5 + DB 197,205,213,228 ; vpmullw %ymm4,%ymm6,%ymm4 + DB 197,181,213,255 ; vpmullw %ymm7,%ymm9,%ymm7 + DB 196,65,29,213,237 ; vpmullw %ymm13,%ymm12,%ymm13 + DB 196,65,21,253,228 ; vpaddw %ymm12,%ymm13,%ymm12 + DB 196,193,69,253,249 ; vpaddw %ymm9,%ymm7,%ymm7 + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 + DB 196,193,85,253,235 ; vpaddw %ymm11,%ymm5,%ymm5 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 - DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 - DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 - DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 - DB 197,181,218,244 ; vpminub %ymm4,%ymm9,%ymm6 - DB 197,165,218,253 ; vpminub %ymm5,%ymm11,%ymm7 + DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 196,193,69,113,212,8 ; vpsrlw $0x8,%ymm12,%ymm7 + DB 196,99,77,56,207,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm9 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,181,103,246 ; vpackuswb %ymm6,%ymm9,%ymm6 + DB 196,227,85,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm7 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 + DB 197,189,218,238 ; vpminub %ymm6,%ymm8,%ymm5 + DB 197,173,218,252 ; vpminub %ymm4,%ymm10,%ymm7 DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 DB 197,245,248,255 ; vpsubb %ymm7,%ymm1,%ymm7 - DB 197,253,248,246 ; vpsubb %ymm6,%ymm0,%ymm6 - DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 - DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 - DB 196,226,125,88,37,223,118,0,0 ; vpbroadcastd 0x76df(%rip),%ymm4 # 2d44c <_sk_overlay_sse2_8bit+0xef4> - DB 196,227,125,76,198,64 ; vpblendvb %ymm4,%ymm6,%ymm0,%ymm0 + DB 197,253,248,237 ; vpsubb %ymm5,%ymm0,%ymm5 + DB 197,245,248,204 ; vpsubb %ymm4,%ymm1,%ymm1 + DB 197,253,248,198 ; vpsubb %ymm6,%ymm0,%ymm0 + DB 196,226,125,88,37,25,116,0,0 ; vpbroadcastd 0x7419(%rip),%ymm4 # 2d338 <_sk_overlay_sse2_8bit+0xe89> + DB 196,227,125,76,197,64 ; vpblendvb %ymm4,%ymm5,%ymm0,%ymm0 DB 196,227,117,76,207,64 ; vpblendvb %ymm4,%ymm7,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_exclusion_hsw_8bit _sk_exclusion_hsw_8bit LABEL PROC - DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 - DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,226,125,48,241 ; vpmovzxbw %xmm1,%ymm6 - DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,196 ; vpmovzxbw %xmm4,%ymm8 + DB 196,226,125,48,232 ; vpmovzxbw %xmm0,%ymm5 + DB 196,227,125,57,206,1 ; vextracti128 $0x1,%ymm1,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,226,125,48,249 ; vpmovzxbw %xmm1,%ymm7 DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 - DB 196,98,125,48,220 ; vpmovzxbw %xmm4,%ymm11 - DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,98,125,48,212 ; vpmovzxbw %xmm4,%ymm10 + DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 DB 196,227,125,57,220,1 ; vextracti128 $0x1,%ymm3,%xmm4 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 197,45,213,214 ; vpmullw %ymm6,%ymm10,%ymm10 - DB 197,37,213,221 ; vpmullw %ymm5,%ymm11,%ymm11 - DB 196,65,61,213,193 ; vpmullw %ymm9,%ymm8,%ymm8 - DB 196,65,61,253,193 ; vpaddw %ymm9,%ymm8,%ymm8 - DB 197,165,253,237 ; vpaddw %ymm5,%ymm11,%ymm5 - DB 197,173,253,246 ; vpaddw %ymm6,%ymm10,%ymm6 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 + DB 196,98,125,48,219 ; vpmovzxbw %xmm3,%ymm11 + DB 197,37,213,223 ; vpmullw %ymm7,%ymm11,%ymm11 + DB 197,221,213,230 ; vpmullw %ymm6,%ymm4,%ymm4 + DB 197,53,213,205 ; vpmullw %ymm5,%ymm9,%ymm9 + DB 196,65,45,213,208 ; vpmullw %ymm8,%ymm10,%ymm10 + DB 196,65,45,253,192 ; vpaddw %ymm8,%ymm10,%ymm8 + DB 197,181,253,237 ; vpaddw %ymm5,%ymm9,%ymm5 + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 + DB 197,165,253,247 ; vpaddw %ymm7,%ymm11,%ymm6 DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,197,1 ; vinserti128 $0x1,%xmm5,%ymm7,%ymm8 - DB 196,227,69,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm7,%ymm5 + DB 196,99,85,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm5,%ymm8 + DB 196,227,85,70,239,49 ; vperm2i128 $0x31,%ymm7,%ymm5,%ymm5 DB 197,189,103,237 ; vpackuswb %ymm5,%ymm8,%ymm5 DB 196,227,77,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm6,%ymm7 DB 196,227,77,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm6,%ymm4 DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 - DB 197,253,111,53,196,115,0,0 ; vmovdqa 0x73c4(%rip),%ymm6 # 2d1e0 <_sk_overlay_sse2_8bit+0xc88> + DB 197,253,111,53,242,112,0,0 ; vmovdqa 0x70f2(%rip),%ymm6 # 2d0c0 <_sk_overlay_sse2_8bit+0xc11> DB 197,221,219,254 ; vpand %ymm6,%ymm4,%ymm7 DB 197,213,219,246 ; vpand %ymm6,%ymm5,%ymm6 DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 DB 197,245,248,204 ; vpsubb %ymm4,%ymm1,%ymm1 + DB 197,245,248,207 ; vpsubb %ymm7,%ymm1,%ymm1 DB 197,253,248,197 ; vpsubb %ymm5,%ymm0,%ymm0 DB 197,253,248,198 ; vpsubb %ymm6,%ymm0,%ymm0 - DB 197,245,248,207 ; vpsubb %ymm7,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_difference_hsw_8bit _sk_difference_hsw_8bit LABEL PROC - DB 197,125,111,5,184,115,0,0 ; vmovdqa 0x73b8(%rip),%ymm8 # 2d200 <_sk_overlay_sse2_8bit+0xca8> - DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 - DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 - DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 - DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 - DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 196,98,125,48,222 ; vpmovzxbw %xmm6,%ymm11 - DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 - DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 - DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 197,125,111,13,230,112,0,0 ; vmovdqa 0x70e6(%rip),%ymm9 # 2d0e0 <_sk_overlay_sse2_8bit+0xc31> + DB 196,194,101,0,233 ; vpshufb %ymm9,%ymm3,%ymm5 + DB 196,194,109,0,241 ; vpshufb %ymm9,%ymm2,%ymm6 + DB 196,227,125,57,199,1 ; vextracti128 $0x1,%ymm0,%xmm7 + DB 196,98,125,48,223 ; vpmovzxbw %xmm7,%ymm11 + DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 + DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 - DB 196,65,29,213,225 ; vpmullw %ymm9,%ymm12,%ymm12 - DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 - DB 196,65,37,213,218 ; vpmullw %ymm10,%ymm11,%ymm11 - DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 - DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 - DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 - DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 - DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 + DB 196,98,125,48,209 ; vpmovzxbw %xmm1,%ymm10 + DB 196,227,125,57,247,1 ; vextracti128 $0x1,%ymm6,%xmm7 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,227,125,57,239,1 ; vextracti128 $0x1,%ymm5,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,193,85,213,234 ; vpmullw %ymm10,%ymm5,%ymm5 + DB 197,197,213,252 ; vpmullw %ymm4,%ymm7,%ymm7 + DB 196,193,77,213,240 ; vpmullw %ymm8,%ymm6,%ymm6 + DB 196,65,29,213,227 ; vpmullw %ymm11,%ymm12,%ymm12 + DB 196,65,29,253,219 ; vpaddw %ymm11,%ymm12,%ymm11 + DB 196,193,77,253,240 ; vpaddw %ymm8,%ymm6,%ymm6 + DB 197,197,253,228 ; vpaddw %ymm4,%ymm7,%ymm4 + DB 196,193,85,253,234 ; vpaddw %ymm10,%ymm5,%ymm5 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 - DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 - DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 - DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 - DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 - DB 197,53,103,204 ; vpackuswb %ymm4,%ymm9,%ymm9 - DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 + DB 196,193,69,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm7 + DB 196,99,77,56,199,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm8 DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 - DB 197,85,103,222 ; vpackuswb %ymm6,%ymm5,%ymm11 - DB 196,194,117,0,232 ; vpshufb %ymm8,%ymm1,%ymm5 - DB 196,194,125,0,248 ; vpshufb %ymm8,%ymm0,%ymm7 - DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 - DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 - DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 197,61,103,214 ; vpackuswb %ymm6,%ymm8,%ymm10 + DB 196,227,85,56,252,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm7 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,69,103,196 ; vpackuswb %ymm4,%ymm7,%ymm8 + DB 196,194,117,0,225 ; vpshufb %ymm9,%ymm1,%ymm4 + DB 196,194,125,0,249 ; vpshufb %ymm9,%ymm0,%ymm7 + DB 196,227,125,57,213,1 ; vextracti128 $0x1,%ymm2,%xmm5 + DB 196,98,125,48,229 ; vpmovzxbw %xmm5,%ymm12 + DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 - DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,98,125,48,219 ; vpmovzxbw %xmm3,%ymm11 + DB 196,227,125,57,253,1 ; vextracti128 $0x1,%ymm7,%xmm5 DB 196,98,125,48,237 ; vpmovzxbw %xmm5,%ymm13 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,227,125,57,229,1 ; vextracti128 $0x1,%ymm4,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 + DB 197,165,213,228 ; vpmullw %ymm4,%ymm11,%ymm4 DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 - DB 196,65,45,213,237 ; vpmullw %ymm13,%ymm10,%ymm13 - DB 197,221,213,255 ; vpmullw %ymm7,%ymm4,%ymm7 - DB 196,65,61,213,228 ; vpmullw %ymm12,%ymm8,%ymm12 - DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 - DB 197,197,253,228 ; vpaddw %ymm4,%ymm7,%ymm4 - DB 196,193,21,253,250 ; vpaddw %ymm10,%ymm13,%ymm7 + DB 197,181,213,255 ; vpmullw %ymm7,%ymm9,%ymm7 + DB 196,65,29,213,237 ; vpmullw %ymm13,%ymm12,%ymm13 + DB 196,65,21,253,228 ; vpaddw %ymm12,%ymm13,%ymm12 + DB 196,193,69,253,249 ; vpaddw %ymm9,%ymm7,%ymm7 DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 196,193,93,253,227 ; vpaddw %ymm11,%ymm4,%ymm4 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 - DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 - DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 - DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 - DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 - DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 - DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 - DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 - DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 - DB 197,181,218,228 ; vpminub %ymm4,%ymm9,%ymm4 - DB 197,165,218,237 ; vpminub %ymm5,%ymm11,%ymm5 - DB 197,253,111,53,130,114,0,0 ; vmovdqa 0x7282(%rip),%ymm6 # 2d220 <_sk_overlay_sse2_8bit+0xcc8> - DB 197,213,219,254 ; vpand %ymm6,%ymm5,%ymm7 - DB 197,221,219,246 ; vpand %ymm6,%ymm4,%ymm6 + DB 196,193,69,113,212,8 ; vpsrlw $0x8,%ymm12,%ymm7 + DB 196,99,77,56,207,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm9 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,181,103,246 ; vpackuswb %ymm6,%ymm9,%ymm6 + DB 197,173,218,246 ; vpminub %ymm6,%ymm10,%ymm6 + DB 196,227,93,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm4,%ymm7 + DB 196,227,93,70,229,49 ; vperm2i128 $0x31,%ymm5,%ymm4,%ymm4 + DB 197,197,103,228 ; vpackuswb %ymm4,%ymm7,%ymm4 + DB 197,189,218,228 ; vpminub %ymm4,%ymm8,%ymm4 + DB 197,253,111,45,174,111,0,0 ; vmovdqa 0x6fae(%rip),%ymm5 # 2d100 <_sk_overlay_sse2_8bit+0xc51> + DB 197,221,219,253 ; vpand %ymm5,%ymm4,%ymm7 + DB 197,205,219,237 ; vpand %ymm5,%ymm6,%ymm5 DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 - DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 - DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 - DB 197,253,248,198 ; vpsubb %ymm6,%ymm0,%ymm0 + DB 197,245,248,204 ; vpsubb %ymm4,%ymm1,%ymm1 DB 197,245,248,207 ; vpsubb %ymm7,%ymm1,%ymm1 + DB 197,253,248,198 ; vpsubb %ymm6,%ymm0,%ymm0 + DB 197,253,248,197 ; vpsubb %ymm5,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_hardlight_hsw_8bit _sk_hardlight_hsw_8bit LABEL PROC - DB 72,129,236,24,2,0,0 ; sub $0x218,%rsp - DB 197,125,111,251 ; vmovdqa %ymm3,%ymm15 - DB 196,226,125,48,248 ; vpmovzxbw %xmm0,%ymm7 + DB 72,129,236,24,1,0,0 ; sub $0x118,%rsp + DB 197,253,111,235 ; vmovdqa %ymm3,%ymm5 + DB 197,125,111,193 ; vmovdqa %ymm1,%ymm8 DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 - DB 196,226,125,48,244 ; vpmovzxbw %xmm4,%ymm6 - DB 196,226,125,48,217 ; vpmovzxbw %xmm1,%ymm3 - DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 - DB 196,98,125,48,196 ; vpmovzxbw %xmm4,%ymm8 - DB 197,126,127,132,36,224,0,0,0 ; vmovdqu %ymm8,0xe0(%rsp) - DB 197,253,111,45,66,114,0,0 ; vmovdqa 0x7242(%rip),%ymm5 # 2d240 <_sk_overlay_sse2_8bit+0xce8> - DB 196,226,125,0,229 ; vpshufb %ymm5,%ymm0,%ymm4 - DB 196,226,117,0,205 ; vpshufb %ymm5,%ymm1,%ymm1 - DB 196,227,125,57,200,1 ; vextracti128 $0x1,%ymm1,%xmm0 - DB 196,98,125,48,240 ; vpmovzxbw %xmm0,%ymm14 - DB 196,226,125,48,233 ; vpmovzxbw %xmm1,%ymm5 - DB 196,227,125,57,224,1 ; vextracti128 $0x1,%ymm4,%xmm0 - DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 - DB 196,98,125,48,212 ; vpmovzxbw %xmm4,%ymm10 - DB 197,197,253,199 ; vpaddw %ymm7,%ymm7,%ymm0 - DB 197,205,253,206 ; vpaddw %ymm6,%ymm6,%ymm1 - DB 197,101,253,227 ; vpaddw %ymm3,%ymm3,%ymm12 - DB 196,65,61,253,216 ; vpaddw %ymm8,%ymm8,%ymm11 - DB 196,66,37,58,238 ; vpminuw %ymm14,%ymm11,%ymm13 - DB 196,193,37,117,229 ; vpcmpeqw %ymm13,%ymm11,%ymm4 - DB 197,254,127,164,36,160,1,0,0 ; vmovdqu %ymm4,0x1a0(%rsp) - DB 196,98,29,58,237 ; vpminuw %ymm5,%ymm12,%ymm13 - DB 196,193,29,117,229 ; vpcmpeqw %ymm13,%ymm12,%ymm4 - DB 197,254,127,164,36,128,1,0,0 ; vmovdqu %ymm4,0x180(%rsp) - DB 196,66,117,58,233 ; vpminuw %ymm9,%ymm1,%ymm13 - DB 196,193,117,117,205 ; vpcmpeqw %ymm13,%ymm1,%ymm1 - DB 197,254,127,140,36,64,1,0,0 ; vmovdqu %ymm1,0x140(%rsp) - DB 196,194,125,58,202 ; vpminuw %ymm10,%ymm0,%ymm1 - DB 197,253,117,193 ; vpcmpeqw %ymm1,%ymm0,%ymm0 - DB 197,254,127,132,36,0,1,0,0 ; vmovdqu %ymm0,0x100(%rsp) - DB 197,253,111,226 ; vmovdqa %ymm2,%ymm4 - DB 197,125,111,37,176,113,0,0 ; vmovdqa 0x71b0(%rip),%ymm12 # 2d240 <_sk_overlay_sse2_8bit+0xce8> - DB 196,66,93,0,196 ; vpshufb %ymm12,%ymm4,%ymm8 - DB 196,194,125,48,208 ; vpmovzxbw %xmm8,%ymm2 - DB 196,226,125,121,5,189,113,0,0 ; vpbroadcastw 0x71bd(%rip),%ymm0 # 2d260 <_sk_overlay_sse2_8bit+0xd08> - DB 197,109,239,232 ; vpxor %ymm0,%ymm2,%ymm13 - DB 197,149,213,207 ; vpmullw %ymm7,%ymm13,%ymm1 - DB 197,254,127,140,36,96,1,0,0 ; vmovdqu %ymm1,0x160(%rsp) DB 196,226,125,48,204 ; vpmovzxbw %xmm4,%ymm1 - DB 197,254,127,164,36,224,1,0,0 ; vmovdqu %ymm4,0x1e0(%rsp) - DB 197,45,239,232 ; vpxor %ymm0,%ymm10,%ymm13 - DB 197,21,213,217 ; vpmullw %ymm1,%ymm13,%ymm11 - DB 197,126,127,156,36,32,1,0,0 ; vmovdqu %ymm11,0x120(%rsp) - DB 197,117,213,223 ; vpmullw %ymm7,%ymm1,%ymm11 - DB 197,126,127,156,36,192,0,0,0 ; vmovdqu %ymm11,0xc0(%rsp) - DB 197,237,249,201 ; vpsubw %ymm1,%ymm2,%ymm1 - DB 197,254,127,12,36 ; vmovdqu %ymm1,(%rsp) - DB 196,193,109,213,202 ; vpmullw %ymm10,%ymm2,%ymm1 - DB 197,254,127,76,36,96 ; vmovdqu %ymm1,0x60(%rsp) - DB 197,45,249,223 ; vpsubw %ymm7,%ymm10,%ymm11 - DB 196,227,125,57,226,1 ; vextracti128 $0x1,%ymm4,%xmm2 + DB 196,226,125,48,216 ; vpmovzxbw %xmm0,%ymm3 + DB 197,125,111,45,131,111,0,0 ; vmovdqa 0x6f83(%rip),%ymm13 # 2d120 <_sk_overlay_sse2_8bit+0xc71> + DB 196,194,125,0,197 ; vpshufb %ymm13,%ymm0,%ymm0 + DB 196,226,125,48,224 ; vpmovzxbw %xmm0,%ymm4 + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,226,125,48,248 ; vpmovzxbw %xmm0,%ymm7 + DB 197,254,127,148,36,224,0,0,0 ; vmovdqu %ymm2,0xe0(%rsp) + DB 196,227,125,57,208,1 ; vextracti128 $0x1,%ymm2,%xmm0 + DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 + DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 + DB 196,66,109,0,221 ; vpshufb %ymm13,%ymm2,%ymm11 + DB 196,194,125,48,211 ; vpmovzxbw %xmm11,%ymm2 + DB 196,99,125,57,216,1 ; vextracti128 $0x1,%ymm11,%xmm0 + DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 + DB 197,254,127,76,36,32 ; vmovdqu %ymm1,0x20(%rsp) + DB 197,245,253,193 ; vpaddw %ymm1,%ymm1,%ymm0 + DB 197,101,253,227 ; vpaddw %ymm3,%ymm3,%ymm12 + DB 197,29,101,244 ; vpcmpgtw %ymm4,%ymm12,%ymm14 + DB 197,254,127,188,36,192,0,0,0 ; vmovdqu %ymm7,0xc0(%rsp) + DB 197,253,101,199 ; vpcmpgtw %ymm7,%ymm0,%ymm0 + DB 197,254,127,180,36,160,0,0,0 ; vmovdqu %ymm6,0xa0(%rsp) + DB 197,45,249,230 ; vpsubw %ymm6,%ymm10,%ymm12 + DB 197,69,249,249 ; vpsubw %ymm1,%ymm7,%ymm15 + DB 196,65,29,213,231 ; vpmullw %ymm15,%ymm12,%ymm12 + DB 197,45,213,255 ; vpmullw %ymm7,%ymm10,%ymm15 + DB 196,65,29,253,228 ; vpaddw %ymm12,%ymm12,%ymm12 + DB 196,65,5,249,228 ; vpsubw %ymm12,%ymm15,%ymm12 + DB 197,77,213,249 ; vpmullw %ymm1,%ymm6,%ymm15 + DB 196,65,5,253,255 ; vpaddw %ymm15,%ymm15,%ymm15 + DB 196,195,5,76,196,0 ; vpblendvb %ymm0,%ymm12,%ymm15,%ymm0 + DB 197,254,127,132,36,128,0,0,0 ; vmovdqu %ymm0,0x80(%rsp) + DB 197,126,127,76,36,64 ; vmovdqu %ymm9,0x40(%rsp) + DB 197,253,111,250 ; vmovdqa %ymm2,%ymm7 + DB 196,193,69,249,193 ; vpsubw %ymm9,%ymm7,%ymm0 + DB 197,253,111,243 ; vmovdqa %ymm3,%ymm6 + DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) + DB 197,93,249,254 ; vpsubw %ymm6,%ymm4,%ymm15 + DB 196,193,125,213,199 ; vpmullw %ymm15,%ymm0,%ymm0 + DB 197,69,213,252 ; vpmullw %ymm4,%ymm7,%ymm15 + DB 197,253,253,192 ; vpaddw %ymm0,%ymm0,%ymm0 + DB 197,133,249,192 ; vpsubw %ymm0,%ymm15,%ymm0 + DB 197,53,213,254 ; vpmullw %ymm6,%ymm9,%ymm15 + DB 196,65,5,253,255 ; vpaddw %ymm15,%ymm15,%ymm15 + DB 196,227,5,76,192,224 ; vpblendvb %ymm14,%ymm0,%ymm15,%ymm0 + DB 197,254,127,68,36,96 ; vmovdqu %ymm0,0x60(%rsp) + DB 196,99,125,57,196,1 ; vextracti128 $0x1,%ymm8,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 + DB 196,66,61,0,245 ; vpshufb %ymm13,%ymm8,%ymm14 + DB 196,99,125,57,243,1 ; vextracti128 $0x1,%ymm14,%xmm3 + DB 196,226,125,48,219 ; vpmovzxbw %xmm3,%ymm3 + DB 197,125,111,221 ; vmovdqa %ymm5,%ymm11 + DB 196,99,125,57,218,1 ; vextracti128 $0x1,%ymm11,%xmm2 DB 196,226,125,48,210 ; vpmovzxbw %xmm2,%ymm2 - DB 196,99,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm1 - DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,117,239,208 ; vpxor %ymm0,%ymm1,%ymm10 - DB 197,173,213,230 ; vpmullw %ymm6,%ymm10,%ymm4 - DB 197,254,127,164,36,160,0,0,0 ; vmovdqu %ymm4,0xa0(%rsp) - DB 197,53,239,208 ; vpxor %ymm0,%ymm9,%ymm10 - DB 197,173,213,226 ; vpmullw %ymm2,%ymm10,%ymm4 - DB 197,254,127,164,36,128,0,0,0 ; vmovdqu %ymm4,0x80(%rsp) - DB 197,109,213,238 ; vpmullw %ymm6,%ymm2,%ymm13 - DB 197,245,249,226 ; vpsubw %ymm2,%ymm1,%ymm4 - DB 196,65,117,213,193 ; vpmullw %ymm9,%ymm1,%ymm8 - DB 197,181,249,246 ; vpsubw %ymm6,%ymm9,%ymm6 - DB 197,125,127,255 ; vmovdqa %ymm15,%ymm7 - DB 197,254,127,188,36,192,1,0,0 ; vmovdqu %ymm7,0x1c0(%rsp) - DB 196,66,69,0,204 ; vpshufb %ymm12,%ymm7,%ymm9 - DB 196,194,125,48,209 ; vpmovzxbw %xmm9,%ymm2 - DB 197,109,239,248 ; vpxor %ymm0,%ymm2,%ymm15 - DB 197,125,111,227 ; vmovdqa %ymm3,%ymm12 - DB 196,193,5,213,204 ; vpmullw %ymm12,%ymm15,%ymm1 - DB 197,254,127,76,36,64 ; vmovdqu %ymm1,0x40(%rsp) - DB 196,226,125,48,207 ; vpmovzxbw %xmm7,%ymm1 - DB 197,85,239,248 ; vpxor %ymm0,%ymm5,%ymm15 - DB 197,133,213,217 ; vpmullw %ymm1,%ymm15,%ymm3 - DB 197,254,127,92,36,32 ; vmovdqu %ymm3,0x20(%rsp) - DB 196,65,117,213,212 ; vpmullw %ymm12,%ymm1,%ymm10 - DB 197,237,249,217 ; vpsubw %ymm1,%ymm2,%ymm3 - DB 197,237,213,213 ; vpmullw %ymm5,%ymm2,%ymm2 - DB 196,193,85,249,236 ; vpsubw %ymm12,%ymm5,%ymm5 - DB 196,195,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm15 - DB 196,66,125,48,255 ; vpmovzxbw %xmm15,%ymm15 - DB 196,99,125,57,207,1 ; vextracti128 $0x1,%ymm9,%xmm7 - DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 - DB 197,69,239,200 ; vpxor %ymm0,%ymm7,%ymm9 - DB 197,254,111,140,36,224,0,0,0 ; vmovdqu 0xe0(%rsp),%ymm1 - DB 197,53,213,225 ; vpmullw %ymm1,%ymm9,%ymm12 - DB 197,141,239,192 ; vpxor %ymm0,%ymm14,%ymm0 - DB 196,65,125,213,207 ; vpmullw %ymm15,%ymm0,%ymm9 - DB 197,133,213,193 ; vpmullw %ymm1,%ymm15,%ymm0 - DB 196,65,69,249,255 ; vpsubw %ymm15,%ymm7,%ymm15 - DB 196,193,69,213,254 ; vpmullw %ymm14,%ymm7,%ymm7 - DB 197,13,249,241 ; vpsubw %ymm1,%ymm14,%ymm14 - DB 196,65,5,213,246 ; vpmullw %ymm14,%ymm15,%ymm14 - DB 197,229,213,205 ; vpmullw %ymm5,%ymm3,%ymm1 - DB 197,221,213,230 ; vpmullw %ymm6,%ymm4,%ymm4 - DB 197,165,213,44,36 ; vpmullw (%rsp),%ymm11,%ymm5 - DB 196,193,13,253,246 ; vpaddw %ymm14,%ymm14,%ymm6 - DB 197,197,249,246 ; vpsubw %ymm6,%ymm7,%ymm6 - DB 197,245,253,201 ; vpaddw %ymm1,%ymm1,%ymm1 - DB 197,237,249,201 ; vpsubw %ymm1,%ymm2,%ymm1 - DB 197,221,253,212 ; vpaddw %ymm4,%ymm4,%ymm2 - DB 197,189,249,210 ; vpsubw %ymm2,%ymm8,%ymm2 - DB 197,213,253,221 ; vpaddw %ymm5,%ymm5,%ymm3 - DB 197,254,111,100,36,96 ; vmovdqu 0x60(%rsp),%ymm4 - DB 197,221,249,219 ; vpsubw %ymm3,%ymm4,%ymm3 - DB 197,254,111,164,36,192,0,0,0 ; vmovdqu 0xc0(%rsp),%ymm4 - DB 197,221,253,228 ; vpaddw %ymm4,%ymm4,%ymm4 - DB 197,254,111,172,36,0,1,0,0 ; vmovdqu 0x100(%rsp),%ymm5 - DB 197,213,223,219 ; vpandn %ymm3,%ymm5,%ymm3 - DB 197,221,219,229 ; vpand %ymm5,%ymm4,%ymm4 - DB 197,221,235,219 ; vpor %ymm3,%ymm4,%ymm3 - DB 196,193,21,253,229 ; vpaddw %ymm13,%ymm13,%ymm4 - DB 197,254,111,172,36,64,1,0,0 ; vmovdqu 0x140(%rsp),%ymm5 - DB 197,213,223,210 ; vpandn %ymm2,%ymm5,%ymm2 - DB 197,221,219,229 ; vpand %ymm5,%ymm4,%ymm4 - DB 197,221,235,210 ; vpor %ymm2,%ymm4,%ymm2 - DB 196,193,45,253,226 ; vpaddw %ymm10,%ymm10,%ymm4 - DB 197,254,111,172,36,128,1,0,0 ; vmovdqu 0x180(%rsp),%ymm5 - DB 197,213,223,201 ; vpandn %ymm1,%ymm5,%ymm1 - DB 197,221,219,229 ; vpand %ymm5,%ymm4,%ymm4 - DB 197,221,235,201 ; vpor %ymm1,%ymm4,%ymm1 + DB 196,66,37,0,253 ; vpshufb %ymm13,%ymm11,%ymm15 + DB 196,67,125,57,249,1 ; vextracti128 $0x1,%ymm15,%xmm9 + DB 196,66,125,48,201 ; vpmovzxbw %xmm9,%ymm9 + DB 197,181,249,194 ; vpsubw %ymm2,%ymm9,%ymm0 + DB 197,101,249,236 ; vpsubw %ymm4,%ymm3,%ymm13 + DB 196,193,125,213,197 ; vpmullw %ymm13,%ymm0,%ymm0 DB 197,253,253,192 ; vpaddw %ymm0,%ymm0,%ymm0 - DB 197,254,111,172,36,160,1,0,0 ; vmovdqu 0x1a0(%rsp),%ymm5 - DB 197,213,223,230 ; vpandn %ymm6,%ymm5,%ymm4 - DB 197,253,219,197 ; vpand %ymm5,%ymm0,%ymm0 - DB 197,253,235,196 ; vpor %ymm4,%ymm0,%ymm0 - DB 197,254,111,164,36,32,1,0,0 ; vmovdqu 0x120(%rsp),%ymm4 - DB 197,221,253,164,36,96,1,0,0 ; vpaddw 0x160(%rsp),%ymm4,%ymm4 - DB 197,254,111,172,36,128,0,0,0 ; vmovdqu 0x80(%rsp),%ymm5 - DB 197,213,253,172,36,160,0,0,0 ; vpaddw 0xa0(%rsp),%ymm5,%ymm5 - DB 197,254,111,116,36,32 ; vmovdqu 0x20(%rsp),%ymm6 - DB 197,205,253,116,36,64 ; vpaddw 0x40(%rsp),%ymm6,%ymm6 - DB 196,193,53,253,252 ; vpaddw %ymm12,%ymm9,%ymm7 - DB 196,98,125,121,5,170,111,0,0 ; vpbroadcastw 0x6faa(%rip),%ymm8 # 2d262 <_sk_overlay_sse2_8bit+0xd0a> - DB 196,193,93,253,224 ; vpaddw %ymm8,%ymm4,%ymm4 - DB 197,229,253,220 ; vpaddw %ymm4,%ymm3,%ymm3 - DB 196,193,85,253,224 ; vpaddw %ymm8,%ymm5,%ymm4 + DB 197,53,213,235 ; vpmullw %ymm3,%ymm9,%ymm13 + DB 197,149,249,192 ; vpsubw %ymm0,%ymm13,%ymm0 + DB 197,93,253,236 ; vpaddw %ymm4,%ymm4,%ymm13 + DB 197,21,101,235 ; vpcmpgtw %ymm3,%ymm13,%ymm13 + DB 197,109,213,228 ; vpmullw %ymm4,%ymm2,%ymm12 + DB 196,65,29,253,228 ; vpaddw %ymm12,%ymm12,%ymm12 + DB 196,227,29,76,200,208 ; vpblendvb %ymm13,%ymm0,%ymm12,%ymm1 + DB 196,66,125,48,224 ; vpmovzxbw %xmm8,%ymm12 + DB 196,66,125,48,238 ; vpmovzxbw %xmm14,%ymm13 + DB 196,66,125,48,247 ; vpmovzxbw %xmm15,%ymm14 + DB 196,66,125,48,251 ; vpmovzxbw %xmm11,%ymm15 + DB 196,193,13,249,199 ; vpsubw %ymm15,%ymm14,%ymm0 + DB 196,65,21,249,196 ; vpsubw %ymm12,%ymm13,%ymm8 + DB 196,193,125,213,192 ; vpmullw %ymm8,%ymm0,%ymm0 + DB 197,253,253,192 ; vpaddw %ymm0,%ymm0,%ymm0 + DB 196,65,13,213,197 ; vpmullw %ymm13,%ymm14,%ymm8 + DB 197,189,249,192 ; vpsubw %ymm0,%ymm8,%ymm0 + DB 196,65,29,253,196 ; vpaddw %ymm12,%ymm12,%ymm8 + DB 196,65,61,101,197 ; vpcmpgtw %ymm13,%ymm8,%ymm8 + DB 196,193,5,213,236 ; vpmullw %ymm12,%ymm15,%ymm5 + DB 197,213,253,237 ; vpaddw %ymm5,%ymm5,%ymm5 + DB 196,227,85,76,192,128 ; vpblendvb %ymm8,%ymm0,%ymm5,%ymm0 + DB 196,226,125,121,45,7,110,0,0 ; vpbroadcastw 0x6e07(%rip),%ymm5 # 2d140 <_sk_overlay_sse2_8bit+0xc91> + DB 197,13,239,197 ; vpxor %ymm5,%ymm14,%ymm8 + DB 196,65,61,213,196 ; vpmullw %ymm12,%ymm8,%ymm8 + DB 197,53,239,205 ; vpxor %ymm5,%ymm9,%ymm9 + DB 197,181,213,228 ; vpmullw %ymm4,%ymm9,%ymm4 + DB 197,69,239,205 ; vpxor %ymm5,%ymm7,%ymm9 + DB 197,53,213,206 ; vpmullw %ymm6,%ymm9,%ymm9 + DB 197,45,239,213 ; vpxor %ymm5,%ymm10,%ymm10 + DB 197,45,213,84,36,32 ; vpmullw 0x20(%rsp),%ymm10,%ymm10 + DB 197,213,239,188,36,192,0,0,0 ; vpxor 0xc0(%rsp),%ymm5,%ymm7 + DB 197,213,239,52,36 ; vpxor (%rsp),%ymm5,%ymm6 + DB 197,229,239,221 ; vpxor %ymm5,%ymm3,%ymm3 + DB 197,149,239,237 ; vpxor %ymm5,%ymm13,%ymm5 + DB 196,193,85,213,239 ; vpmullw %ymm15,%ymm5,%ymm5 + DB 196,193,85,253,232 ; vpaddw %ymm8,%ymm5,%ymm5 + DB 197,229,213,210 ; vpmullw %ymm2,%ymm3,%ymm2 DB 197,237,253,212 ; vpaddw %ymm4,%ymm2,%ymm2 - DB 196,193,77,253,224 ; vpaddw %ymm8,%ymm6,%ymm4 - DB 197,245,253,204 ; vpaddw %ymm4,%ymm1,%ymm1 - DB 196,193,69,253,224 ; vpaddw %ymm8,%ymm7,%ymm4 - DB 197,253,253,196 ; vpaddw %ymm4,%ymm0,%ymm0 - DB 196,226,125,121,37,127,111,0,0 ; vpbroadcastw 0x6f7f(%rip),%ymm4 # 2d264 <_sk_overlay_sse2_8bit+0xd0c> - DB 197,253,228,196 ; vpmulhuw %ymm4,%ymm0,%ymm0 - DB 197,245,228,204 ; vpmulhuw %ymm4,%ymm1,%ymm1 - DB 197,237,228,212 ; vpmulhuw %ymm4,%ymm2,%ymm2 - DB 197,237,113,210,7 ; vpsrlw $0x7,%ymm2,%ymm2 - DB 197,229,228,220 ; vpmulhuw %ymm4,%ymm3,%ymm3 + DB 197,205,213,92,36,64 ; vpmullw 0x40(%rsp),%ymm6,%ymm3 + DB 196,193,101,253,217 ; vpaddw %ymm9,%ymm3,%ymm3 + DB 197,197,213,164,36,160,0,0,0 ; vpmullw 0xa0(%rsp),%ymm7,%ymm4 + DB 196,193,93,253,226 ; vpaddw %ymm10,%ymm4,%ymm4 + DB 196,226,125,121,53,156,109,0,0 ; vpbroadcastw 0x6d9c(%rip),%ymm6 # 2d142 <_sk_overlay_sse2_8bit+0xc93> + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 + DB 197,221,253,164,36,128,0,0,0 ; vpaddw 0x80(%rsp),%ymm4,%ymm4 + DB 197,229,253,222 ; vpaddw %ymm6,%ymm3,%ymm3 + DB 197,229,253,92,36,96 ; vpaddw 0x60(%rsp),%ymm3,%ymm3 + DB 197,237,253,214 ; vpaddw %ymm6,%ymm2,%ymm2 + DB 197,245,253,202 ; vpaddw %ymm2,%ymm1,%ymm1 + DB 197,213,253,214 ; vpaddw %ymm6,%ymm5,%ymm2 + DB 197,253,253,194 ; vpaddw %ymm2,%ymm0,%ymm0 + DB 196,226,125,121,21,110,109,0,0 ; vpbroadcastw 0x6d6e(%rip),%ymm2 # 2d144 <_sk_overlay_sse2_8bit+0xc95> + DB 197,253,228,194 ; vpmulhuw %ymm2,%ymm0,%ymm0 + DB 197,245,228,202 ; vpmulhuw %ymm2,%ymm1,%ymm1 + DB 197,229,228,218 ; vpmulhuw %ymm2,%ymm3,%ymm3 DB 197,229,113,211,7 ; vpsrlw $0x7,%ymm3,%ymm3 + DB 197,221,228,210 ; vpmulhuw %ymm2,%ymm4,%ymm2 + DB 197,237,113,210,7 ; vpsrlw $0x7,%ymm2,%ymm2 DB 196,227,101,56,226,1 ; vinserti128 $0x1,%xmm2,%ymm3,%ymm4 DB 196,227,101,70,210,49 ; vperm2i128 $0x31,%ymm2,%ymm3,%ymm2 DB 197,229,113,208,7 ; vpsrlw $0x7,%ymm0,%ymm3 DB 197,245,113,209,7 ; vpsrlw $0x7,%ymm1,%ymm1 DB 197,221,103,194 ; vpackuswb %ymm2,%ymm4,%ymm0 - DB 196,227,117,56,211,1 ; vinserti128 $0x1,%xmm3,%ymm1,%ymm2 - DB 196,227,117,70,203,49 ; vperm2i128 $0x31,%ymm3,%ymm1,%ymm1 + DB 196,227,101,56,209,1 ; vinserti128 $0x1,%xmm1,%ymm3,%ymm2 + DB 196,227,101,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm3,%ymm1 DB 197,237,103,201 ; vpackuswb %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,148,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm2 - DB 197,252,16,156,36,192,1,0,0 ; vmovups 0x1c0(%rsp),%ymm3 - DB 72,129,196,24,2,0,0 ; add $0x218,%rsp + DB 197,252,16,148,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm2 + DB 197,125,127,219 ; vmovdqa %ymm11,%ymm3 + DB 72,129,196,24,1,0,0 ; add $0x118,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_overlay_hsw_8bit _sk_overlay_hsw_8bit LABEL PROC - DB 72,129,236,24,2,0,0 ; sub $0x218,%rsp - DB 197,252,17,140,36,192,1,0,0 ; vmovups %ymm1,0x1c0(%rsp) - DB 197,253,111,45,34,111,0,0 ; vmovdqa 0x6f22(%rip),%ymm5 # 2d280 <_sk_overlay_sse2_8bit+0xd28> - DB 197,253,111,226 ; vmovdqa %ymm2,%ymm4 - DB 196,227,125,57,225,1 ; vextracti128 $0x1,%ymm4,%xmm1 - DB 196,98,125,48,193 ; vpmovzxbw %xmm1,%ymm8 - DB 197,254,127,156,36,224,1,0,0 ; vmovdqu %ymm3,0x1e0(%rsp) - DB 196,226,125,48,251 ; vpmovzxbw %xmm3,%ymm7 - DB 196,227,125,57,217,1 ; vextracti128 $0x1,%ymm3,%xmm1 - DB 196,98,125,48,241 ; vpmovzxbw %xmm1,%ymm14 - DB 196,226,93,0,205 ; vpshufb %ymm5,%ymm4,%ymm1 - DB 197,254,127,164,36,160,1,0,0 ; vmovdqu %ymm4,0x1a0(%rsp) - DB 196,226,101,0,221 ; vpshufb %ymm5,%ymm3,%ymm3 - DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 - DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 - DB 196,98,125,48,203 ; vpmovzxbw %xmm3,%ymm9 - DB 196,227,125,57,203,1 ; vextracti128 $0x1,%ymm1,%xmm3 - DB 196,98,125,48,219 ; vpmovzxbw %xmm3,%ymm11 - DB 196,193,61,253,216 ; vpaddw %ymm8,%ymm8,%ymm3 - DB 197,69,253,231 ; vpaddw %ymm7,%ymm7,%ymm12 - DB 196,65,13,253,214 ; vpaddw %ymm14,%ymm14,%ymm10 - DB 196,98,45,58,238 ; vpminuw %ymm6,%ymm10,%ymm13 - DB 196,193,45,117,213 ; vpcmpeqw %ymm13,%ymm10,%ymm2 - DB 197,254,127,148,36,128,1,0,0 ; vmovdqu %ymm2,0x180(%rsp) - DB 196,66,29,58,233 ; vpminuw %ymm9,%ymm12,%ymm13 - DB 196,193,29,117,213 ; vpcmpeqw %ymm13,%ymm12,%ymm2 - DB 197,254,127,148,36,96,1,0,0 ; vmovdqu %ymm2,0x160(%rsp) - DB 196,66,101,58,235 ; vpminuw %ymm11,%ymm3,%ymm13 - DB 196,193,101,117,213 ; vpcmpeqw %ymm13,%ymm3,%ymm2 - DB 197,254,127,148,36,32,1,0,0 ; vmovdqu %ymm2,0x120(%rsp) - DB 196,226,125,48,212 ; vpmovzxbw %xmm4,%ymm2 - DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,237,253,226 ; vpaddw %ymm2,%ymm2,%ymm4 - DB 196,226,93,58,217 ; vpminuw %ymm1,%ymm4,%ymm3 - DB 197,93,117,251 ; vpcmpeqw %ymm3,%ymm4,%ymm15 - DB 197,253,111,216 ; vmovdqa %ymm0,%ymm3 - DB 196,226,125,48,227 ; vpmovzxbw %xmm3,%ymm4 - DB 196,226,125,121,5,124,110,0,0 ; vpbroadcastw 0x6e7c(%rip),%ymm0 # 2d2a0 <_sk_overlay_sse2_8bit+0xd48> - DB 197,117,239,232 ; vpxor %ymm0,%ymm1,%ymm13 - DB 197,21,213,212 ; vpmullw %ymm4,%ymm13,%ymm10 - DB 197,126,127,148,36,64,1,0,0 ; vmovdqu %ymm10,0x140(%rsp) - DB 197,125,111,229 ; vmovdqa %ymm5,%ymm12 - DB 196,194,101,0,236 ; vpshufb %ymm12,%ymm3,%ymm5 - DB 197,125,111,211 ; vmovdqa %ymm3,%ymm10 - DB 196,226,125,48,221 ; vpmovzxbw %xmm5,%ymm3 - DB 197,101,239,232 ; vpxor %ymm0,%ymm3,%ymm13 - DB 197,21,213,234 ; vpmullw %ymm2,%ymm13,%ymm13 - DB 197,126,127,172,36,0,1,0,0 ; vmovdqu %ymm13,0x100(%rsp) - DB 197,109,213,236 ; vpmullw %ymm4,%ymm2,%ymm13 - DB 197,126,127,172,36,224,0,0,0 ; vmovdqu %ymm13,0xe0(%rsp) - DB 197,245,249,210 ; vpsubw %ymm2,%ymm1,%ymm2 - DB 197,254,127,84,36,32 ; vmovdqu %ymm2,0x20(%rsp) - DB 197,245,213,203 ; vpmullw %ymm3,%ymm1,%ymm1 - DB 197,254,127,140,36,128,0,0,0 ; vmovdqu %ymm1,0x80(%rsp) - DB 197,229,249,220 ; vpsubw %ymm4,%ymm3,%ymm3 - DB 196,99,125,57,212,1 ; vextracti128 $0x1,%ymm10,%xmm4 - DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 72,129,236,24,1,0,0 ; sub $0x118,%rsp + DB 197,253,111,243 ; vmovdqa %ymm3,%ymm6 + DB 197,253,111,233 ; vmovdqa %ymm1,%ymm5 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,226,125,48,204 ; vpmovzxbw %xmm4,%ymm1 + DB 196,226,125,48,216 ; vpmovzxbw %xmm0,%ymm3 + DB 197,125,111,45,7,109,0,0 ; vmovdqa 0x6d07(%rip),%ymm13 # 2d160 <_sk_overlay_sse2_8bit+0xcb1> + DB 196,194,125,0,197 ; vpshufb %ymm13,%ymm0,%ymm0 + DB 196,226,125,48,248 ; vpmovzxbw %xmm0,%ymm7 + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 + DB 197,254,127,148,36,224,0,0,0 ; vmovdqu %ymm2,0xe0(%rsp) + DB 196,227,125,57,208,1 ; vextracti128 $0x1,%ymm2,%xmm0 + DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 + DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 196,66,109,0,221 ; vpshufb %ymm13,%ymm2,%ymm11 + DB 196,66,125,48,211 ; vpmovzxbw %xmm11,%ymm10 + DB 196,99,125,57,220,1 ; vextracti128 $0x1,%ymm11,%xmm4 + DB 196,98,125,48,220 ; vpmovzxbw %xmm4,%ymm11 + DB 197,253,253,224 ; vpaddw %ymm0,%ymm0,%ymm4 + DB 196,65,61,253,224 ; vpaddw %ymm8,%ymm8,%ymm12 + DB 196,65,29,101,242 ; vpcmpgtw %ymm10,%ymm12,%ymm14 + DB 196,193,93,101,227 ; vpcmpgtw %ymm11,%ymm4,%ymm4 + DB 197,37,249,224 ; vpsubw %ymm0,%ymm11,%ymm12 + DB 197,254,127,132,36,160,0,0,0 ; vmovdqu %ymm0,0xa0(%rsp) + DB 197,254,127,76,36,32 ; vmovdqu %ymm1,0x20(%rsp) + DB 197,126,127,140,36,192,0,0,0 ; vmovdqu %ymm9,0xc0(%rsp) + DB 197,53,249,249 ; vpsubw %ymm1,%ymm9,%ymm15 + DB 196,65,29,213,231 ; vpmullw %ymm15,%ymm12,%ymm12 + DB 196,65,37,213,249 ; vpmullw %ymm9,%ymm11,%ymm15 + DB 196,65,29,253,228 ; vpaddw %ymm12,%ymm12,%ymm12 + DB 196,65,5,249,228 ; vpsubw %ymm12,%ymm15,%ymm12 + DB 197,125,213,249 ; vpmullw %ymm1,%ymm0,%ymm15 + DB 196,65,5,253,255 ; vpaddw %ymm15,%ymm15,%ymm15 + DB 196,195,5,76,196,64 ; vpblendvb %ymm4,%ymm12,%ymm15,%ymm0 + DB 197,254,127,132,36,128,0,0,0 ; vmovdqu %ymm0,0x80(%rsp) + DB 197,126,127,68,36,64 ; vmovdqu %ymm8,0x40(%rsp) + DB 196,193,45,249,224 ; vpsubw %ymm8,%ymm10,%ymm4 + DB 197,254,127,60,36 ; vmovdqu %ymm7,(%rsp) + DB 197,69,249,251 ; vpsubw %ymm3,%ymm7,%ymm15 + DB 196,193,93,213,231 ; vpmullw %ymm15,%ymm4,%ymm4 + DB 197,45,213,255 ; vpmullw %ymm7,%ymm10,%ymm15 + DB 197,221,253,228 ; vpaddw %ymm4,%ymm4,%ymm4 + DB 197,133,249,228 ; vpsubw %ymm4,%ymm15,%ymm4 + DB 197,61,213,251 ; vpmullw %ymm3,%ymm8,%ymm15 + DB 197,125,111,195 ; vmovdqa %ymm3,%ymm8 + DB 196,65,5,253,255 ; vpaddw %ymm15,%ymm15,%ymm15 + DB 196,227,5,76,196,224 ; vpblendvb %ymm14,%ymm4,%ymm15,%ymm0 + DB 197,254,127,68,36,96 ; vmovdqu %ymm0,0x60(%rsp) + DB 197,253,111,197 ; vmovdqa %ymm5,%ymm0 + DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,37,239,232 ; vpxor %ymm0,%ymm11,%ymm13 - DB 197,149,213,204 ; vpmullw %ymm4,%ymm13,%ymm1 - DB 197,254,127,140,36,192,0,0,0 ; vmovdqu %ymm1,0xc0(%rsp) - DB 197,85,239,232 ; vpxor %ymm0,%ymm5,%ymm13 - DB 196,193,21,213,200 ; vpmullw %ymm8,%ymm13,%ymm1 - DB 197,254,127,140,36,160,0,0,0 ; vmovdqu %ymm1,0xa0(%rsp) - DB 197,189,213,204 ; vpmullw %ymm4,%ymm8,%ymm1 - DB 197,254,127,12,36 ; vmovdqu %ymm1,(%rsp) - DB 196,65,37,249,192 ; vpsubw %ymm8,%ymm11,%ymm8 - DB 197,37,213,221 ; vpmullw %ymm5,%ymm11,%ymm11 - DB 197,213,249,228 ; vpsubw %ymm4,%ymm5,%ymm4 - DB 197,254,111,140,36,192,1,0,0 ; vmovdqu 0x1c0(%rsp),%ymm1 - DB 196,194,117,0,236 ; vpshufb %ymm12,%ymm1,%ymm5 - DB 196,226,125,48,209 ; vpmovzxbw %xmm1,%ymm2 - DB 197,125,111,225 ; vmovdqa %ymm1,%ymm12 - DB 197,53,239,232 ; vpxor %ymm0,%ymm9,%ymm13 - DB 197,149,213,202 ; vpmullw %ymm2,%ymm13,%ymm1 - DB 197,254,127,76,36,96 ; vmovdqu %ymm1,0x60(%rsp) - DB 196,226,125,48,205 ; vpmovzxbw %xmm5,%ymm1 - DB 197,117,239,232 ; vpxor %ymm0,%ymm1,%ymm13 - DB 197,21,213,215 ; vpmullw %ymm7,%ymm13,%ymm10 - DB 197,126,127,84,36,64 ; vmovdqu %ymm10,0x40(%rsp) - DB 197,69,213,210 ; vpmullw %ymm2,%ymm7,%ymm10 - DB 197,181,249,255 ; vpsubw %ymm7,%ymm9,%ymm7 - DB 197,53,213,201 ; vpmullw %ymm1,%ymm9,%ymm9 - DB 197,245,249,202 ; vpsubw %ymm2,%ymm1,%ymm1 - DB 196,99,125,57,226,1 ; vextracti128 $0x1,%ymm12,%xmm2 + DB 196,66,125,0,245 ; vpshufb %ymm13,%ymm0,%ymm14 + DB 197,253,111,224 ; vmovdqa %ymm0,%ymm4 + DB 196,99,125,57,243,1 ; vextracti128 $0x1,%ymm14,%xmm3 + DB 196,226,125,48,219 ; vpmovzxbw %xmm3,%ymm3 + DB 197,253,111,254 ; vmovdqa %ymm6,%ymm7 + DB 196,227,125,57,250,1 ; vextracti128 $0x1,%ymm7,%xmm2 DB 196,226,125,48,210 ; vpmovzxbw %xmm2,%ymm2 - DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 - DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 - DB 197,77,239,232 ; vpxor %ymm0,%ymm6,%ymm13 - DB 197,21,213,234 ; vpmullw %ymm2,%ymm13,%ymm13 - DB 197,213,239,192 ; vpxor %ymm0,%ymm5,%ymm0 - DB 196,65,125,213,230 ; vpmullw %ymm14,%ymm0,%ymm12 - DB 197,141,213,194 ; vpmullw %ymm2,%ymm14,%ymm0 - DB 196,65,77,249,246 ; vpsubw %ymm14,%ymm6,%ymm14 - DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 - DB 197,213,249,210 ; vpsubw %ymm2,%ymm5,%ymm2 - DB 197,141,213,210 ; vpmullw %ymm2,%ymm14,%ymm2 - DB 197,197,213,201 ; vpmullw %ymm1,%ymm7,%ymm1 - DB 197,189,213,228 ; vpmullw %ymm4,%ymm8,%ymm4 - DB 197,229,213,92,36,32 ; vpmullw 0x20(%rsp),%ymm3,%ymm3 - DB 197,237,253,210 ; vpaddw %ymm2,%ymm2,%ymm2 - DB 197,205,249,210 ; vpsubw %ymm2,%ymm6,%ymm2 - DB 197,245,253,201 ; vpaddw %ymm1,%ymm1,%ymm1 - DB 197,181,249,201 ; vpsubw %ymm1,%ymm9,%ymm1 - DB 197,221,253,228 ; vpaddw %ymm4,%ymm4,%ymm4 - DB 197,165,249,228 ; vpsubw %ymm4,%ymm11,%ymm4 - DB 197,229,253,219 ; vpaddw %ymm3,%ymm3,%ymm3 - DB 197,254,111,172,36,128,0,0,0 ; vmovdqu 0x80(%rsp),%ymm5 - DB 197,213,249,219 ; vpsubw %ymm3,%ymm5,%ymm3 - DB 197,254,111,172,36,224,0,0,0 ; vmovdqu 0xe0(%rsp),%ymm5 - DB 197,213,253,237 ; vpaddw %ymm5,%ymm5,%ymm5 - DB 197,133,223,219 ; vpandn %ymm3,%ymm15,%ymm3 - DB 196,193,85,219,239 ; vpand %ymm15,%ymm5,%ymm5 - DB 197,213,235,219 ; vpor %ymm3,%ymm5,%ymm3 - DB 197,254,111,44,36 ; vmovdqu (%rsp),%ymm5 - DB 197,213,253,237 ; vpaddw %ymm5,%ymm5,%ymm5 - DB 197,254,111,180,36,32,1,0,0 ; vmovdqu 0x120(%rsp),%ymm6 - DB 197,205,223,228 ; vpandn %ymm4,%ymm6,%ymm4 - DB 197,213,219,238 ; vpand %ymm6,%ymm5,%ymm5 - DB 197,213,235,228 ; vpor %ymm4,%ymm5,%ymm4 - DB 196,193,45,253,234 ; vpaddw %ymm10,%ymm10,%ymm5 - DB 197,254,111,180,36,96,1,0,0 ; vmovdqu 0x160(%rsp),%ymm6 - DB 197,205,223,201 ; vpandn %ymm1,%ymm6,%ymm1 - DB 197,213,219,238 ; vpand %ymm6,%ymm5,%ymm5 - DB 197,213,235,201 ; vpor %ymm1,%ymm5,%ymm1 + DB 196,66,69,0,253 ; vpshufb %ymm13,%ymm7,%ymm15 + DB 196,67,125,57,253,1 ; vextracti128 $0x1,%ymm15,%xmm13 + DB 196,66,125,48,237 ; vpmovzxbw %xmm13,%ymm13 + DB 197,149,249,194 ; vpsubw %ymm2,%ymm13,%ymm0 + DB 197,101,249,205 ; vpsubw %ymm5,%ymm3,%ymm9 + DB 196,193,125,213,193 ; vpmullw %ymm9,%ymm0,%ymm0 DB 197,253,253,192 ; vpaddw %ymm0,%ymm0,%ymm0 - DB 197,254,111,172,36,128,1,0,0 ; vmovdqu 0x180(%rsp),%ymm5 - DB 197,213,223,210 ; vpandn %ymm2,%ymm5,%ymm2 - DB 197,253,219,197 ; vpand %ymm5,%ymm0,%ymm0 - DB 197,253,235,194 ; vpor %ymm2,%ymm0,%ymm0 - DB 197,254,111,148,36,0,1,0,0 ; vmovdqu 0x100(%rsp),%ymm2 - DB 197,237,253,148,36,64,1,0,0 ; vpaddw 0x140(%rsp),%ymm2,%ymm2 - DB 197,254,111,172,36,160,0,0,0 ; vmovdqu 0xa0(%rsp),%ymm5 - DB 197,213,253,172,36,192,0,0,0 ; vpaddw 0xc0(%rsp),%ymm5,%ymm5 - DB 197,254,111,116,36,64 ; vmovdqu 0x40(%rsp),%ymm6 - DB 197,205,253,116,36,96 ; vpaddw 0x60(%rsp),%ymm6,%ymm6 - DB 196,193,29,253,253 ; vpaddw %ymm13,%ymm12,%ymm7 - DB 196,98,125,121,5,111,108,0,0 ; vpbroadcastw 0x6c6f(%rip),%ymm8 # 2d2a2 <_sk_overlay_sse2_8bit+0xd4a> - DB 196,193,109,253,208 ; vpaddw %ymm8,%ymm2,%ymm2 - DB 197,229,253,210 ; vpaddw %ymm2,%ymm3,%ymm2 - DB 196,193,85,253,216 ; vpaddw %ymm8,%ymm5,%ymm3 - DB 197,221,253,219 ; vpaddw %ymm3,%ymm4,%ymm3 - DB 196,193,77,253,224 ; vpaddw %ymm8,%ymm6,%ymm4 - DB 197,245,253,204 ; vpaddw %ymm4,%ymm1,%ymm1 - DB 196,193,69,253,224 ; vpaddw %ymm8,%ymm7,%ymm4 - DB 197,253,253,196 ; vpaddw %ymm4,%ymm0,%ymm0 - DB 196,226,125,121,37,68,108,0,0 ; vpbroadcastw 0x6c44(%rip),%ymm4 # 2d2a4 <_sk_overlay_sse2_8bit+0xd4c> - DB 197,253,228,196 ; vpmulhuw %ymm4,%ymm0,%ymm0 - DB 197,245,228,204 ; vpmulhuw %ymm4,%ymm1,%ymm1 - DB 197,229,228,220 ; vpmulhuw %ymm4,%ymm3,%ymm3 + DB 197,21,213,203 ; vpmullw %ymm3,%ymm13,%ymm9 + DB 197,181,249,192 ; vpsubw %ymm0,%ymm9,%ymm0 + DB 197,109,253,202 ; vpaddw %ymm2,%ymm2,%ymm9 + DB 196,65,53,101,205 ; vpcmpgtw %ymm13,%ymm9,%ymm9 + DB 197,237,213,245 ; vpmullw %ymm5,%ymm2,%ymm6 + DB 197,205,253,246 ; vpaddw %ymm6,%ymm6,%ymm6 + DB 196,227,77,76,200,144 ; vpblendvb %ymm9,%ymm0,%ymm6,%ymm1 + DB 196,226,125,48,244 ; vpmovzxbw %xmm4,%ymm6 + DB 196,66,125,48,206 ; vpmovzxbw %xmm14,%ymm9 + DB 196,66,125,48,247 ; vpmovzxbw %xmm15,%ymm14 + DB 196,98,125,48,255 ; vpmovzxbw %xmm7,%ymm15 + DB 196,193,13,249,199 ; vpsubw %ymm15,%ymm14,%ymm0 + DB 197,53,249,230 ; vpsubw %ymm6,%ymm9,%ymm12 + DB 196,193,125,213,196 ; vpmullw %ymm12,%ymm0,%ymm0 + DB 197,253,253,192 ; vpaddw %ymm0,%ymm0,%ymm0 + DB 196,65,13,213,225 ; vpmullw %ymm9,%ymm14,%ymm12 + DB 197,157,249,192 ; vpsubw %ymm0,%ymm12,%ymm0 + DB 196,65,5,253,231 ; vpaddw %ymm15,%ymm15,%ymm12 + DB 196,65,29,101,230 ; vpcmpgtw %ymm14,%ymm12,%ymm12 + DB 197,133,213,230 ; vpmullw %ymm6,%ymm15,%ymm4 + DB 197,221,253,228 ; vpaddw %ymm4,%ymm4,%ymm4 + DB 196,227,93,76,192,192 ; vpblendvb %ymm12,%ymm0,%ymm4,%ymm0 + DB 196,226,125,121,37,133,107,0,0 ; vpbroadcastw 0x6b85(%rip),%ymm4 # 2d180 <_sk_overlay_sse2_8bit+0xcd1> + DB 197,13,239,228 ; vpxor %ymm4,%ymm14,%ymm12 + DB 197,157,213,246 ; vpmullw %ymm6,%ymm12,%ymm6 + DB 197,21,239,228 ; vpxor %ymm4,%ymm13,%ymm12 + DB 197,157,213,237 ; vpmullw %ymm5,%ymm12,%ymm5 + DB 197,45,239,212 ; vpxor %ymm4,%ymm10,%ymm10 + DB 196,65,45,213,208 ; vpmullw %ymm8,%ymm10,%ymm10 + DB 197,37,239,220 ; vpxor %ymm4,%ymm11,%ymm11 + DB 197,37,213,92,36,32 ; vpmullw 0x20(%rsp),%ymm11,%ymm11 + DB 197,93,239,164,36,192,0,0,0 ; vpxor 0xc0(%rsp),%ymm4,%ymm12 + DB 197,93,239,44,36 ; vpxor (%rsp),%ymm4,%ymm13 + DB 197,229,239,220 ; vpxor %ymm4,%ymm3,%ymm3 + DB 197,181,239,228 ; vpxor %ymm4,%ymm9,%ymm4 + DB 196,193,93,213,231 ; vpmullw %ymm15,%ymm4,%ymm4 + DB 197,221,253,230 ; vpaddw %ymm6,%ymm4,%ymm4 + DB 197,229,213,210 ; vpmullw %ymm2,%ymm3,%ymm2 + DB 197,237,253,213 ; vpaddw %ymm5,%ymm2,%ymm2 + DB 197,149,213,92,36,64 ; vpmullw 0x40(%rsp),%ymm13,%ymm3 + DB 196,193,101,253,218 ; vpaddw %ymm10,%ymm3,%ymm3 + DB 197,157,213,172,36,160,0,0,0 ; vpmullw 0xa0(%rsp),%ymm12,%ymm5 + DB 196,193,85,253,235 ; vpaddw %ymm11,%ymm5,%ymm5 + DB 196,226,125,121,53,27,107,0,0 ; vpbroadcastw 0x6b1b(%rip),%ymm6 # 2d182 <_sk_overlay_sse2_8bit+0xcd3> + DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 197,213,253,172,36,128,0,0,0 ; vpaddw 0x80(%rsp),%ymm5,%ymm5 + DB 197,229,253,222 ; vpaddw %ymm6,%ymm3,%ymm3 + DB 197,229,253,92,36,96 ; vpaddw 0x60(%rsp),%ymm3,%ymm3 + DB 197,237,253,214 ; vpaddw %ymm6,%ymm2,%ymm2 + DB 197,245,253,202 ; vpaddw %ymm2,%ymm1,%ymm1 + DB 197,221,253,214 ; vpaddw %ymm6,%ymm4,%ymm2 + DB 197,253,253,194 ; vpaddw %ymm2,%ymm0,%ymm0 + DB 196,226,125,121,21,237,106,0,0 ; vpbroadcastw 0x6aed(%rip),%ymm2 # 2d184 <_sk_overlay_sse2_8bit+0xcd5> + DB 197,253,228,194 ; vpmulhuw %ymm2,%ymm0,%ymm0 + DB 197,245,228,202 ; vpmulhuw %ymm2,%ymm1,%ymm1 + DB 197,229,228,218 ; vpmulhuw %ymm2,%ymm3,%ymm3 DB 197,229,113,211,7 ; vpsrlw $0x7,%ymm3,%ymm3 - DB 197,237,228,212 ; vpmulhuw %ymm4,%ymm2,%ymm2 + DB 197,213,228,210 ; vpmulhuw %ymm2,%ymm5,%ymm2 DB 197,237,113,210,7 ; vpsrlw $0x7,%ymm2,%ymm2 - DB 196,227,109,56,227,1 ; vinserti128 $0x1,%xmm3,%ymm2,%ymm4 - DB 196,227,109,70,211,49 ; vperm2i128 $0x31,%ymm3,%ymm2,%ymm2 + DB 196,227,101,56,226,1 ; vinserti128 $0x1,%xmm2,%ymm3,%ymm4 + DB 196,227,101,70,210,49 ; vperm2i128 $0x31,%ymm2,%ymm3,%ymm2 DB 197,229,113,208,7 ; vpsrlw $0x7,%ymm0,%ymm3 DB 197,245,113,209,7 ; vpsrlw $0x7,%ymm1,%ymm1 DB 197,221,103,194 ; vpackuswb %ymm2,%ymm4,%ymm0 - DB 196,227,117,56,211,1 ; vinserti128 $0x1,%xmm3,%ymm1,%ymm2 - DB 196,227,117,70,203,49 ; vperm2i128 $0x31,%ymm3,%ymm1,%ymm1 + DB 196,227,101,56,209,1 ; vinserti128 $0x1,%xmm1,%ymm3,%ymm2 + DB 196,227,101,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm3,%ymm1 DB 197,237,103,201 ; vpackuswb %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,252,16,148,36,160,1,0,0 ; vmovups 0x1a0(%rsp),%ymm2 - DB 197,252,16,156,36,224,1,0,0 ; vmovups 0x1e0(%rsp),%ymm3 - DB 72,129,196,24,2,0,0 ; add $0x218,%rsp + DB 197,252,16,148,36,224,0,0,0 ; vmovups 0xe0(%rsp),%ymm2 + DB 197,253,111,223 ; vmovdqa %ymm7,%ymm3 + DB 72,129,196,24,1,0,0 ; add $0x118,%rsp DB 255,224 ; jmpq *%rax DB 144 ; nop - DB 144 ; nop - DB 144 ; nop PUBLIC _sk_start_pipeline_sse41_8bit _sk_start_pipeline_sse41_8bit LABEL PROC @@ -36610,13 +36566,13 @@ _sk_start_pipeline_sse41_8bit LABEL PROC DB 73,137,246 ; mov %rsi,%r14 DB 76,137,141,8,255,255,255 ; mov %r9,-0xf8(%rbp) DB 76,57,203 ; cmp %r9,%rbx - DB 115,101 ; jae 267a6 <_sk_start_pipeline_sse41_8bit+0xe2> + DB 115,101 ; jae 267d6 <_sk_start_pipeline_sse41_8bit+0xe2> DB 72,139,133,24,255,255,255 ; mov -0xe8(%rbp),%rax DB 72,141,64,8 ; lea 0x8(%rax),%rax DB 72,137,133,16,255,255,255 ; mov %rax,-0xf0(%rbp) DB 76,57,189,16,255,255,255 ; cmp %r15,-0xf0(%rbp) DB 72,139,149,24,255,255,255 ; mov -0xe8(%rbp),%rdx - DB 119,38 ; ja 26789 <_sk_start_pipeline_sse41_8bit+0xc5> + DB 119,38 ; ja 267b9 <_sk_start_pipeline_sse41_8bit+0xc5> DB 76,139,165,24,255,255,255 ; mov -0xe8(%rbp),%r12 DB 49,255 ; xor %edi,%edi DB 76,137,246 ; mov %r14,%rsi @@ -36627,16 +36583,16 @@ _sk_start_pipeline_sse41_8bit LABEL PROC DB 73,131,196,16 ; add $0x10,%r12 DB 77,57,252 ; cmp %r15,%r12 DB 73,137,212 ; mov %rdx,%r12 - DB 118,225 ; jbe 2676a <_sk_start_pipeline_sse41_8bit+0xa6> + DB 118,225 ; jbe 2679a <_sk_start_pipeline_sse41_8bit+0xa6> DB 76,137,255 ; mov %r15,%rdi DB 72,41,215 ; sub %rdx,%rdi - DB 116,9 ; je 2679a <_sk_start_pipeline_sse41_8bit+0xd6> + DB 116,9 ; je 267ca <_sk_start_pipeline_sse41_8bit+0xd6> DB 76,137,246 ; mov %r14,%rsi DB 72,137,217 ; mov %rbx,%rcx DB 65,255,213 ; callq *%r13 DB 72,255,195 ; inc %rbx DB 72,59,157,8,255,255,255 ; cmp -0xf8(%rbp),%rbx - DB 117,173 ; jne 26753 <_sk_start_pipeline_sse41_8bit+0x8f> + DB 117,173 ; jne 26783 <_sk_start_pipeline_sse41_8bit+0x8f> DB 15,40,181,32,255,255,255 ; movaps -0xe0(%rbp),%xmm6 DB 15,40,189,48,255,255,255 ; movaps -0xd0(%rbp),%xmm7 DB 68,15,40,133,64,255,255,255 ; movaps -0xc0(%rbp),%xmm8 @@ -36674,7 +36630,7 @@ _sk_uniform_color_sse41_8bit LABEL PROC PUBLIC _sk_set_rgb_sse41_8bit _sk_set_rgb_sse41_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,15,16,37,189,106,0,0 ; movss 0x6abd(%rip),%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 243,15,16,37,109,105,0,0 ; movss 0x696d(%rip),%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 243,15,16,40 ; movss (%rax),%xmm5 DB 243,15,89,236 ; mulss %xmm4,%xmm5 DB 243,76,15,44,197 ; cvttss2si %xmm5,%r8 @@ -36689,7 +36645,7 @@ _sk_set_rgb_sse41_8bit LABEL PROC DB 68,9,200 ; or %r9d,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 102,15,111,45,50,115,0,0 ; movdqa 0x7332(%rip),%xmm5 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,45,18,114,0,0 ; movdqa 0x7212(%rip),%xmm5 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,219,205 ; pand %xmm5,%xmm1 DB 102,15,219,197 ; pand %xmm5,%xmm0 DB 102,15,235,196 ; por %xmm4,%xmm0 @@ -36699,45 +36655,44 @@ _sk_set_rgb_sse41_8bit LABEL PROC PUBLIC _sk_premul_sse41_8bit _sk_premul_sse41_8bit LABEL PROC - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,30,115,0,0 ; movdqa 0x731e(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,15,111,253 ; movdqa %xmm5,%xmm7 - DB 102,15,56,0,248 ; pshufb %xmm0,%xmm7 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,56,0,240 ; pshufb %xmm0,%xmm6 - DB 102,15,111,5,244,114,0,0 ; movdqa 0x72f4(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x1638> - DB 102,15,235,240 ; por %xmm0,%xmm6 - DB 102,15,235,248 ; por %xmm0,%xmm7 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,56,48,205 ; pmovzxbw %xmm5,%xmm9 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,56,48,212 ; pmovzxbw %xmm4,%xmm10 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 - DB 102,15,56,48,206 ; pmovzxbw %xmm6,%xmm1 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,213,253 ; pmullw %xmm5,%xmm7 - DB 102,65,15,213,202 ; pmullw %xmm10,%xmm1 - DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 - DB 102,15,253,253 ; paddw %xmm5,%xmm7 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 - DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 - DB 102,65,15,253,202 ; paddw %xmm10,%xmm1 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,68,15,111,5,5,114,0,0 ; movdqa 0x7205(%rip),%xmm8 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,65,15,56,0,192 ; pshufb %xmm8,%xmm0 + DB 102,15,112,249,78 ; pshufd $0x4e,%xmm1,%xmm7 + DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,65,15,56,0,200 ; pshufb %xmm8,%xmm1 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,68,15,111,5,192,113,0,0 ; movdqa 0x71c0(%rip),%xmm8 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> + DB 102,65,15,235,200 ; por %xmm8,%xmm1 + DB 102,65,15,235,192 ; por %xmm8,%xmm0 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 + DB 102,15,213,247 ; pmullw %xmm7,%xmm6 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,15,213,229 ; pmullw %xmm5,%xmm4 + DB 102,15,253,229 ; paddw %xmm5,%xmm4 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 + DB 102,15,253,247 ; paddw %xmm7,%xmm6 + DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,103,206 ; packuswb %xmm6,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,196 ; packuswb %xmm4,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_swap_rb_sse41_8bit _sk_swap_rb_sse41_8bit LABEL PROC - DB 102,15,111,37,145,114,0,0 ; movdqa 0x7291(%rip),%xmm4 # 2dbb0 <_sk_overlay_sse2_8bit+0x1658> + DB 102,15,111,37,108,113,0,0 ; movdqa 0x716c(%rip),%xmm4 # 2dac0 <_sk_overlay_sse2_8bit+0x1611> DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -36761,9 +36716,9 @@ _sk_load_8888_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 2696a <_sk_load_8888_sse41_8bit+0x2d> - DB 243,67,15,111,76,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm1 + DB 117,17 ; jne 2699f <_sk_load_8888_sse41_8bit+0x2d> DB 243,67,15,111,4,152 ; movdqu (%r8,%r11,4),%xmm0 + DB 243,67,15,111,76,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -36772,40 +36727,45 @@ _sk_load_8888_sse41_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,228 ; ja 26966 <_sk_load_8888_sse41_8bit+0x29> + DB 119,228 ; ja 2699b <_sk_load_8888_sse41_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,75,0,0,0 ; lea 0x4b(%rip),%r10 # 269d8 <_sk_load_8888_sse41_8bit+0x9b> + DB 76,141,21,82,0,0,0 ; lea 0x52(%rip),%r10 # 26a14 <_sk_load_8888_sse41_8bit+0xa2> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,110,4,152 ; movd (%r8,%r11,4),%xmm0 - DB 235,200 ; jmp 26966 <_sk_load_8888_sse41_8bit+0x29> + DB 235,200 ; jmp 2699b <_sk_load_8888_sse41_8bit+0x29> DB 102,67,15,110,68,152,8 ; movd 0x8(%r8,%r11,4),%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 DB 243,67,15,126,36,152 ; movq (%r8,%r11,4),%xmm4 DB 102,15,58,14,196,15 ; pblendw $0xf,%xmm4,%xmm0 - DB 235,174 ; jmp 26966 <_sk_load_8888_sse41_8bit+0x29> + DB 235,174 ; jmp 2699b <_sk_load_8888_sse41_8bit+0x29> DB 102,67,15,110,68,152,24 ; movd 0x18(%r8,%r11,4),%xmm0 DB 102,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm1 DB 102,67,15,58,34,76,152,20,1 ; pinsrd $0x1,0x14(%r8,%r11,4),%xmm1 DB 102,67,15,58,34,76,152,16,0 ; pinsrd $0x0,0x10(%r8,%r11,4),%xmm1 - DB 235,136 ; jmp 26960 <_sk_load_8888_sse41_8bit+0x23> - DB 190,255,255,255,210 ; mov $0xd2ffffff,%esi + DB 243,67,15,111,4,152 ; movdqu (%r8,%r11,4),%xmm0 + DB 235,136 ; jmp 2699b <_sk_load_8888_sse41_8bit+0x29> + DB 144 ; nop + DB 183,255 ; mov $0xff,%bh DB 255 ; (bad) + DB 255,203 ; dec %ebx DB 255 ; (bad) - DB 255,198 ; inc %esi DB 255 ; (bad) DB 255 ; (bad) - DB 255,136,255,255,255,245 ; decl -0xa000001(%rax) + DB 191,255,255,255,247 ; mov $0xf7ffffff,%edi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 236 ; in (%dx),%al + DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,224 ; jmpq *%rax + DB 255,229 ; jmpq *%rbp + DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) + DB 217,255 ; fcos + DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_8888_dst_sse41_8bit @@ -36818,9 +36778,9 @@ _sk_load_8888_dst_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 26a21 <_sk_load_8888_dst_sse41_8bit+0x2d> - DB 243,67,15,111,92,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm3 + DB 117,17 ; jne 26a5d <_sk_load_8888_dst_sse41_8bit+0x2d> DB 243,67,15,111,20,152 ; movdqu (%r8,%r11,4),%xmm2 + DB 243,67,15,111,92,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -36829,39 +36789,44 @@ _sk_load_8888_dst_sse41_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,228 ; ja 26a1d <_sk_load_8888_dst_sse41_8bit+0x29> + DB 119,228 ; ja 26a59 <_sk_load_8888_dst_sse41_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,76,0,0,0 ; lea 0x4c(%rip),%r10 # 26a90 <_sk_load_8888_dst_sse41_8bit+0x9c> + DB 76,141,21,84,0,0,0 ; lea 0x54(%rip),%r10 # 26ad4 <_sk_load_8888_dst_sse41_8bit+0xa4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,110,20,152 ; movd (%r8,%r11,4),%xmm2 - DB 235,200 ; jmp 26a1d <_sk_load_8888_dst_sse41_8bit+0x29> + DB 235,200 ; jmp 26a59 <_sk_load_8888_dst_sse41_8bit+0x29> DB 102,67,15,110,84,152,8 ; movd 0x8(%r8,%r11,4),%xmm2 DB 102,15,112,210,69 ; pshufd $0x45,%xmm2,%xmm2 DB 243,67,15,126,36,152 ; movq (%r8,%r11,4),%xmm4 DB 102,15,58,14,212,15 ; pblendw $0xf,%xmm4,%xmm2 - DB 235,174 ; jmp 26a1d <_sk_load_8888_dst_sse41_8bit+0x29> + DB 235,174 ; jmp 26a59 <_sk_load_8888_dst_sse41_8bit+0x29> DB 102,67,15,110,84,152,24 ; movd 0x18(%r8,%r11,4),%xmm2 DB 102,15,112,218,69 ; pshufd $0x45,%xmm2,%xmm3 DB 102,67,15,58,34,92,152,20,1 ; pinsrd $0x1,0x14(%r8,%r11,4),%xmm3 DB 102,67,15,58,34,92,152,16,0 ; pinsrd $0x0,0x10(%r8,%r11,4),%xmm3 - DB 235,136 ; jmp 26a17 <_sk_load_8888_dst_sse41_8bit+0x23> - DB 144 ; nop - DB 189,255,255,255,209 ; mov $0xd1ffffff,%ebp + DB 243,67,15,111,20,152 ; movdqu (%r8,%r11,4),%xmm2 + DB 235,136 ; jmp 26a59 <_sk_load_8888_dst_sse41_8bit+0x29> + DB 15,31,0 ; nopl (%rax) + DB 181,255 ; mov $0xff,%ch + DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) - DB 255,197 ; inc %ebp + DB 255 ; (bad) + DB 189,255,255,255,245 ; mov $0xf5ffffff,%ebp + DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 255,135,255,255,255,244 ; incl -0xb000001(%rdi) + DB 236 ; in (%dx),%al DB 255 ; (bad) DB 255 ; (bad) + DB 255,227 ; jmpq *%rbx DB 255 ; (bad) - DB 235,255 ; jmp 26aa5 <_sk_load_8888_dst_sse41_8bit+0xb1> DB 255 ; (bad) + DB 255,215 ; callq *%rdi DB 255 ; (bad) - DB 223,255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -36875,7 +36840,7 @@ _sk_store_8888_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 26ad9 <_sk_store_8888_sse41_8bit+0x2d> + DB 117,17 ; jne 26b1d <_sk_store_8888_sse41_8bit+0x2d> DB 243,67,15,127,4,152 ; movdqu %xmm0,(%r8,%r11,4) DB 243,67,15,127,76,152,16 ; movdqu %xmm1,0x10(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -36884,22 +36849,22 @@ _sk_store_8888_sse41_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 26ad5 <_sk_store_8888_sse41_8bit+0x29> + DB 119,236 ; ja 26b19 <_sk_store_8888_sse41_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 26b38 <_sk_store_8888_sse41_8bit+0x8c> + DB 76,141,21,68,0,0,0 ; lea 0x44(%rip),%r10 # 26b7c <_sk_store_8888_sse41_8bit+0x8c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,126,4,152 ; movd %xmm0,(%r8,%r11,4) - DB 235,208 ; jmp 26ad5 <_sk_store_8888_sse41_8bit+0x29> + DB 235,208 ; jmp 26b19 <_sk_store_8888_sse41_8bit+0x29> DB 102,67,15,58,22,68,152,8,2 ; pextrd $0x2,%xmm0,0x8(%r8,%r11,4) DB 102,67,15,214,4,152 ; movq %xmm0,(%r8,%r11,4) - DB 235,191 ; jmp 26ad5 <_sk_store_8888_sse41_8bit+0x29> + DB 235,191 ; jmp 26b19 <_sk_store_8888_sse41_8bit+0x29> DB 102,67,15,58,22,76,152,24,2 ; pextrd $0x2,%xmm1,0x18(%r8,%r11,4) DB 102,67,15,58,22,76,152,20,1 ; pextrd $0x1,%xmm1,0x14(%r8,%r11,4) DB 102,67,15,126,76,152,16 ; movd %xmm1,0x10(%r8,%r11,4) DB 243,67,15,127,4,152 ; movdqu %xmm0,(%r8,%r11,4) - DB 235,158 ; jmp 26ad5 <_sk_store_8888_sse41_8bit+0x29> + DB 235,158 ; jmp 26b19 <_sk_store_8888_sse41_8bit+0x29> DB 144 ; nop DB 197,255,255 ; (bad) DB 255,214 ; callq *%rsi @@ -36932,10 +36897,10 @@ _sk_load_bgra_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,35 ; jne 26b93 <_sk_load_bgra_sse41_8bit+0x3f> - DB 243,67,15,111,76,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm1 + DB 117,35 ; jne 26bd7 <_sk_load_bgra_sse41_8bit+0x3f> DB 243,67,15,111,4,152 ; movdqu (%r8,%r11,4),%xmm0 - DB 102,15,111,37,43,112,0,0 ; movdqa 0x702b(%rip),%xmm4 # 2dbb0 <_sk_overlay_sse2_8bit+0x1658> + DB 243,67,15,111,76,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm1 + DB 102,15,111,37,247,110,0,0 ; movdqa 0x6ef7(%rip),%xmm4 # 2dac0 <_sk_overlay_sse2_8bit+0x1611> DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -36946,37 +36911,43 @@ _sk_load_bgra_sse41_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,210 ; ja 26b7d <_sk_load_bgra_sse41_8bit+0x29> + DB 119,210 ; ja 26bc1 <_sk_load_bgra_sse41_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,78,0,0,0 ; lea 0x4e(%rip),%r10 # 26c04 <_sk_load_bgra_sse41_8bit+0xb0> + DB 76,141,21,86,0,0,0 ; lea 0x56(%rip),%r10 # 26c50 <_sk_load_bgra_sse41_8bit+0xb8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,110,4,152 ; movd (%r8,%r11,4),%xmm0 - DB 235,182 ; jmp 26b7d <_sk_load_bgra_sse41_8bit+0x29> + DB 235,182 ; jmp 26bc1 <_sk_load_bgra_sse41_8bit+0x29> DB 102,67,15,110,68,152,8 ; movd 0x8(%r8,%r11,4),%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 DB 243,67,15,126,36,152 ; movq (%r8,%r11,4),%xmm4 DB 102,15,58,14,196,15 ; pblendw $0xf,%xmm4,%xmm0 - DB 235,156 ; jmp 26b7d <_sk_load_bgra_sse41_8bit+0x29> + DB 235,156 ; jmp 26bc1 <_sk_load_bgra_sse41_8bit+0x29> DB 102,67,15,110,68,152,24 ; movd 0x18(%r8,%r11,4),%xmm0 DB 102,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm1 DB 102,67,15,58,34,76,152,20,1 ; pinsrd $0x1,0x14(%r8,%r11,4),%xmm1 DB 102,67,15,58,34,76,152,16,0 ; pinsrd $0x0,0x10(%r8,%r11,4),%xmm1 - DB 233,115,255,255,255 ; jmpq 26b77 <_sk_load_bgra_sse41_8bit+0x23> - DB 187,255,255,255,207 ; mov $0xcfffffff,%ebx + DB 243,67,15,111,4,152 ; movdqu (%r8,%r11,4),%xmm0 + DB 233,115,255,255,255 ; jmpq 26bc1 <_sk_load_bgra_sse41_8bit+0x29> + DB 102,144 ; xchg %ax,%ax + DB 179,255 ; mov $0xff,%bl DB 255 ; (bad) + DB 255,199 ; inc %edi DB 255 ; (bad) - DB 255,195 ; inc %ebx DB 255 ; (bad) DB 255 ; (bad) - DB 255,115,255 ; pushq -0x1(%rbx) + DB 187,255,255,255,243 ; mov $0xf3ffffff,%ebx DB 255 ; (bad) - DB 255,242 ; push %rdx DB 255 ; (bad) DB 255 ; (bad) + DB 234 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,221 ; jmpq ffffffffde026c1c <_sk_overlay_sse2_8bit+0xffffffffddffa6c4> + DB 255 ; (bad) + DB 255,225 ; jmpq *%rcx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,213 ; callq *%rbp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -36991,10 +36962,10 @@ _sk_load_bgra_dst_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,35 ; jne 26c5f <_sk_load_bgra_dst_sse41_8bit+0x3f> - DB 243,67,15,111,92,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm3 + DB 117,35 ; jne 26cab <_sk_load_bgra_dst_sse41_8bit+0x3f> DB 243,67,15,111,20,152 ; movdqu (%r8,%r11,4),%xmm2 - DB 102,15,111,37,95,111,0,0 ; movdqa 0x6f5f(%rip),%xmm4 # 2dbb0 <_sk_overlay_sse2_8bit+0x1658> + DB 243,67,15,111,92,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm3 + DB 102,15,111,37,35,110,0,0 ; movdqa 0x6e23(%rip),%xmm4 # 2dac0 <_sk_overlay_sse2_8bit+0x1611> DB 102,15,56,0,212 ; pshufb %xmm4,%xmm2 DB 102,15,56,0,220 ; pshufb %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -37005,37 +36976,43 @@ _sk_load_bgra_dst_sse41_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,210 ; ja 26c49 <_sk_load_bgra_dst_sse41_8bit+0x29> + DB 119,210 ; ja 26c95 <_sk_load_bgra_dst_sse41_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,78,0,0,0 ; lea 0x4e(%rip),%r10 # 26cd0 <_sk_load_bgra_dst_sse41_8bit+0xb0> + DB 76,141,21,86,0,0,0 ; lea 0x56(%rip),%r10 # 26d24 <_sk_load_bgra_dst_sse41_8bit+0xb8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,110,20,152 ; movd (%r8,%r11,4),%xmm2 - DB 235,182 ; jmp 26c49 <_sk_load_bgra_dst_sse41_8bit+0x29> + DB 235,182 ; jmp 26c95 <_sk_load_bgra_dst_sse41_8bit+0x29> DB 102,67,15,110,84,152,8 ; movd 0x8(%r8,%r11,4),%xmm2 DB 102,15,112,210,69 ; pshufd $0x45,%xmm2,%xmm2 DB 243,67,15,126,36,152 ; movq (%r8,%r11,4),%xmm4 DB 102,15,58,14,212,15 ; pblendw $0xf,%xmm4,%xmm2 - DB 235,156 ; jmp 26c49 <_sk_load_bgra_dst_sse41_8bit+0x29> + DB 235,156 ; jmp 26c95 <_sk_load_bgra_dst_sse41_8bit+0x29> DB 102,67,15,110,84,152,24 ; movd 0x18(%r8,%r11,4),%xmm2 DB 102,15,112,218,69 ; pshufd $0x45,%xmm2,%xmm3 DB 102,67,15,58,34,92,152,20,1 ; pinsrd $0x1,0x14(%r8,%r11,4),%xmm3 DB 102,67,15,58,34,92,152,16,0 ; pinsrd $0x0,0x10(%r8,%r11,4),%xmm3 - DB 233,115,255,255,255 ; jmpq 26c43 <_sk_load_bgra_dst_sse41_8bit+0x23> - DB 187,255,255,255,207 ; mov $0xcfffffff,%ebx + DB 243,67,15,111,20,152 ; movdqu (%r8,%r11,4),%xmm2 + DB 233,115,255,255,255 ; jmpq 26c95 <_sk_load_bgra_dst_sse41_8bit+0x29> + DB 102,144 ; xchg %ax,%ax + DB 179,255 ; mov $0xff,%bl DB 255 ; (bad) + DB 255,199 ; inc %edi DB 255 ; (bad) - DB 255,195 ; inc %ebx DB 255 ; (bad) DB 255 ; (bad) - DB 255,115,255 ; pushq -0x1(%rbx) + DB 187,255,255,255,243 ; mov $0xf3ffffff,%ebx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 234 ; (bad) DB 255 ; (bad) - DB 255,242 ; push %rdx DB 255 ; (bad) + DB 255,225 ; jmpq *%rcx DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,221 ; jmpq ffffffffde026ce8 <_sk_overlay_sse2_8bit+0xffffffffddffa790> + DB 255,213 ; callq *%rbp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -37049,13 +37026,13 @@ _sk_store_bgra_sse41_8bit LABEL PROC DB 73,193,224,2 ; shl $0x2,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 102,15,111,53,165,110,0,0 ; movdqa 0x6ea5(%rip),%xmm6 # 2dbb0 <_sk_overlay_sse2_8bit+0x1658> + DB 102,15,111,53,97,109,0,0 ; movdqa 0x6d61(%rip),%xmm6 # 2dac0 <_sk_overlay_sse2_8bit+0x1611> DB 102,15,111,233 ; movdqa %xmm1,%xmm5 DB 102,15,56,0,238 ; pshufb %xmm6,%xmm5 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 DB 102,15,56,0,230 ; pshufb %xmm6,%xmm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 26d33 <_sk_store_bgra_sse41_8bit+0x47> + DB 117,17 ; jne 26d87 <_sk_store_bgra_sse41_8bit+0x47> DB 243,67,15,127,36,152 ; movdqu %xmm4,(%r8,%r11,4) DB 243,67,15,127,108,152,16 ; movdqu %xmm5,0x10(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -37064,22 +37041,22 @@ _sk_store_bgra_sse41_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 26d2f <_sk_store_bgra_sse41_8bit+0x43> + DB 119,236 ; ja 26d83 <_sk_store_bgra_sse41_8bit+0x43> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,70,0,0,0 ; lea 0x46(%rip),%r10 # 26d94 <_sk_store_bgra_sse41_8bit+0xa8> + DB 76,141,21,70,0,0,0 ; lea 0x46(%rip),%r10 # 26de8 <_sk_store_bgra_sse41_8bit+0xa8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,126,36,152 ; movd %xmm4,(%r8,%r11,4) - DB 235,208 ; jmp 26d2f <_sk_store_bgra_sse41_8bit+0x43> + DB 235,208 ; jmp 26d83 <_sk_store_bgra_sse41_8bit+0x43> DB 102,67,15,58,22,100,152,8,2 ; pextrd $0x2,%xmm4,0x8(%r8,%r11,4) DB 102,67,15,214,36,152 ; movq %xmm4,(%r8,%r11,4) - DB 235,191 ; jmp 26d2f <_sk_store_bgra_sse41_8bit+0x43> + DB 235,191 ; jmp 26d83 <_sk_store_bgra_sse41_8bit+0x43> DB 102,67,15,58,22,108,152,24,2 ; pextrd $0x2,%xmm5,0x18(%r8,%r11,4) DB 102,67,15,58,22,108,152,20,1 ; pextrd $0x1,%xmm5,0x14(%r8,%r11,4) DB 102,67,15,126,108,152,16 ; movd %xmm5,0x10(%r8,%r11,4) DB 243,67,15,127,36,152 ; movdqu %xmm4,(%r8,%r11,4) - DB 235,158 ; jmp 26d2f <_sk_store_bgra_sse41_8bit+0x43> + DB 235,158 ; jmp 26d83 <_sk_store_bgra_sse41_8bit+0x43> DB 15,31,0 ; nopl (%rax) DB 195 ; retq DB 255 ; (bad) @@ -37114,34 +37091,34 @@ _sk_store_565_sse41_8bit LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 102,68,15,111,13,241,109,0,0 ; movdqa 0x6df1(%rip),%xmm9 # 2dbc0 <_sk_overlay_sse2_8bit+0x1668> - DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 - DB 102,69,15,219,193 ; pand %xmm9,%xmm8 - DB 102,68,15,219,200 ; pand %xmm0,%xmm9 - DB 102,65,15,114,241,8 ; pslld $0x8,%xmm9 - DB 102,65,15,114,240,8 ; pslld $0x8,%xmm8 - DB 102,15,111,37,222,109,0,0 ; movdqa 0x6dde(%rip),%xmm4 # 2dbd0 <_sk_overlay_sse2_8bit+0x1678> - DB 102,15,111,249 ; movdqa %xmm1,%xmm7 - DB 102,15,219,252 ; pand %xmm4,%xmm7 + DB 102,15,111,37,174,108,0,0 ; movdqa 0x6cae(%rip),%xmm4 # 2dad0 <_sk_overlay_sse2_8bit+0x1621> + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,15,114,212,5 ; psrld $0x5,%xmm4 + DB 102,15,114,244,8 ; pslld $0x8,%xmm4 + DB 102,15,114,245,8 ; pslld $0x8,%xmm5 + DB 102,15,111,53,160,108,0,0 ; movdqa 0x6ca0(%rip),%xmm6 # 2dae0 <_sk_overlay_sse2_8bit+0x1631> + DB 102,15,111,249 ; movdqa %xmm1,%xmm7 + DB 102,15,219,254 ; pand %xmm6,%xmm7 + DB 102,15,219,240 ; pand %xmm0,%xmm6 + DB 102,15,114,214,5 ; psrld $0x5,%xmm6 + DB 102,15,235,244 ; por %xmm4,%xmm6 DB 102,15,114,215,5 ; psrld $0x5,%xmm7 - DB 102,15,111,53,208,109,0,0 ; movdqa 0x6dd0(%rip),%xmm6 # 2dbe0 <_sk_overlay_sse2_8bit+0x1688> - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,219,238 ; pand %xmm6,%xmm5 - DB 102,15,219,241 ; pand %xmm1,%xmm6 - DB 102,15,114,214,19 ; psrld $0x13,%xmm6 + DB 102,15,235,253 ; por %xmm5,%xmm7 + DB 102,15,111,45,138,108,0,0 ; movdqa 0x6c8a(%rip),%xmm5 # 2daf0 <_sk_overlay_sse2_8bit+0x1641> + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,219,229 ; pand %xmm5,%xmm4 + DB 102,15,219,233 ; pand %xmm1,%xmm5 DB 102,15,114,213,19 ; psrld $0x13,%xmm5 - DB 102,65,15,235,248 ; por %xmm8,%xmm7 - DB 102,65,15,235,225 ; por %xmm9,%xmm4 - DB 102,15,235,229 ; por %xmm5,%xmm4 - DB 102,15,235,254 ; por %xmm6,%xmm7 - DB 102,15,111,45,176,109,0,0 ; movdqa 0x6db0(%rip),%xmm5 # 2dbf0 <_sk_overlay_sse2_8bit+0x1698> - DB 102,15,56,0,253 ; pshufb %xmm5,%xmm7 - DB 102,15,56,0,229 ; pshufb %xmm5,%xmm4 - DB 102,15,108,231 ; punpcklqdq %xmm7,%xmm4 + DB 102,15,235,239 ; por %xmm7,%xmm5 + DB 102,15,114,212,19 ; psrld $0x13,%xmm4 + DB 102,15,235,230 ; por %xmm6,%xmm4 + DB 102,15,111,53,116,108,0,0 ; movdqa 0x6c74(%rip),%xmm6 # 2db00 <_sk_overlay_sse2_8bit+0x1651> + DB 102,15,56,0,238 ; pshufb %xmm6,%xmm5 + DB 102,15,56,0,230 ; pshufb %xmm6,%xmm4 + DB 102,15,108,229 ; punpcklqdq %xmm5,%xmm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 26e5d <_sk_store_565_sse41_8bit+0xad> + DB 117,10 ; jne 26ea9 <_sk_store_565_sse41_8bit+0xa5> DB 243,67,15,127,36,88 ; movdqu %xmm4,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -37149,22 +37126,22 @@ _sk_store_565_sse41_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 26e59 <_sk_store_565_sse41_8bit+0xa9> + DB 119,236 ; ja 26ea5 <_sk_store_565_sse41_8bit+0xa1> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,72,0,0,0 ; lea 0x48(%rip),%r10 # 26ec0 <_sk_store_565_sse41_8bit+0x110> + DB 76,141,21,72,0,0,0 ; lea 0x48(%rip),%r10 # 26f0c <_sk_store_565_sse41_8bit+0x108> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,58,21,36,88,0 ; pextrw $0x0,%xmm4,(%r8,%r11,2) - DB 235,206 ; jmp 26e59 <_sk_store_565_sse41_8bit+0xa9> + DB 235,206 ; jmp 26ea5 <_sk_store_565_sse41_8bit+0xa1> DB 102,67,15,58,21,100,88,4,2 ; pextrw $0x2,%xmm4,0x4(%r8,%r11,2) DB 102,67,15,126,36,88 ; movd %xmm4,(%r8,%r11,2) - DB 235,189 ; jmp 26e59 <_sk_store_565_sse41_8bit+0xa9> + DB 235,189 ; jmp 26ea5 <_sk_store_565_sse41_8bit+0xa1> DB 102,67,15,58,21,100,88,12,6 ; pextrw $0x6,%xmm4,0xc(%r8,%r11,2) DB 102,67,15,58,21,100,88,10,5 ; pextrw $0x5,%xmm4,0xa(%r8,%r11,2) DB 102,67,15,58,21,100,88,8,4 ; pextrw $0x4,%xmm4,0x8(%r8,%r11,2) DB 102,67,15,214,36,88 ; movq %xmm4,(%r8,%r11,2) - DB 235,154 ; jmp 26e59 <_sk_store_565_sse41_8bit+0xa9> + DB 235,154 ; jmp 26ea5 <_sk_store_565_sse41_8bit+0xa1> DB 144 ; nop DB 193,255,255 ; sar $0xff,%edi DB 255,212 ; callq *%rsp @@ -37198,30 +37175,30 @@ _sk_load_565_dst_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,216,0,0,0 ; jne 26fd3 <_sk_load_565_dst_sse41_8bit+0xf7> - DB 243,67,15,111,28,88 ; movdqu (%r8,%r11,2),%xmm3 - DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,15,56,51,211 ; pmovzxwd %xmm3,%xmm2 - DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 - DB 102,15,111,53,186,103,0,0 ; movdqa 0x67ba(%rip),%xmm6 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 15,133,218,0,0,0 ; jne 27021 <_sk_load_565_dst_sse41_8bit+0xf9> + DB 243,67,15,111,20,88 ; movdqu (%r8,%r11,2),%xmm2 + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,56,51,219 ; pmovzxwd %xmm3,%xmm3 + DB 102,15,56,51,210 ; pmovzxwd %xmm2,%xmm2 + DB 102,15,111,53,92,102,0,0 ; movdqa 0x665c(%rip),%xmm6 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,111,250 ; movdqa %xmm2,%xmm7 DB 102,15,219,254 ; pand %xmm6,%xmm7 DB 102,15,219,243 ; pand %xmm3,%xmm6 DB 102,15,114,214,8 ; psrld $0x8,%xmm6 DB 102,15,114,215,8 ; psrld $0x8,%xmm7 - DB 102,15,111,45,220,103,0,0 ; movdqa 0x67dc(%rip),%xmm5 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 102,15,111,45,126,102,0,0 ; movdqa 0x667e(%rip),%xmm5 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,15,219,229 ; pand %xmm5,%xmm4 DB 102,15,219,235 ; pand %xmm3,%xmm5 DB 102,15,114,245,19 ; pslld $0x13,%xmm5 + DB 102,15,235,238 ; por %xmm6,%xmm5 DB 102,15,114,244,19 ; pslld $0x13,%xmm4 DB 102,15,235,231 ; por %xmm7,%xmm4 - DB 102,15,235,238 ; por %xmm6,%xmm5 - DB 102,15,111,53,150,103,0,0 ; movdqa 0x6796(%rip),%xmm6 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 102,15,111,53,56,102,0,0 ; movdqa 0x6638(%rip),%xmm6 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,111,251 ; movdqa %xmm3,%xmm7 DB 102,15,219,254 ; pand %xmm6,%xmm7 DB 102,15,114,247,5 ; pslld $0x5,%xmm7 - DB 102,68,15,111,5,160,108,0,0 ; movdqa 0x6ca0(%rip),%xmm8 # 2dc10 <_sk_overlay_sse2_8bit+0x16b8> + DB 102,68,15,111,5,98,107,0,0 ; movdqa 0x6b62(%rip),%xmm8 # 2db20 <_sk_overlay_sse2_8bit+0x1671> DB 102,65,15,219,216 ; pand %xmm8,%xmm3 DB 102,15,114,211,1 ; psrld $0x1,%xmm3 DB 102,15,235,223 ; por %xmm7,%xmm3 @@ -37234,10 +37211,10 @@ _sk_load_565_dst_sse41_8bit LABEL PROC DB 102,15,235,214 ; por %xmm6,%xmm2 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,114,212,5 ; psrld $0x5,%xmm4 - DB 102,15,111,53,81,108,0,0 ; movdqa 0x6c51(%rip),%xmm6 # 2dc00 <_sk_overlay_sse2_8bit+0x16a8> + DB 102,15,111,53,19,107,0,0 ; movdqa 0x6b13(%rip),%xmm6 # 2db10 <_sk_overlay_sse2_8bit+0x1661> DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 102,15,219,238 ; pand %xmm6,%xmm5 - DB 102,15,111,53,209,107,0,0 ; movdqa 0x6bd1(%rip),%xmm6 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,53,147,106,0,0 ; movdqa 0x6a93(%rip),%xmm6 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,235,214 ; por %xmm6,%xmm2 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,222 ; por %xmm6,%xmm3 @@ -37246,47 +37223,49 @@ _sk_load_565_dst_sse41_8bit LABEL PROC DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b - DB 102,15,239,219 ; pxor %xmm3,%xmm3 + DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,22,255,255,255 ; ja 26f01 <_sk_load_565_dst_sse41_8bit+0x25> + DB 15,135,20,255,255,255 ; ja 26f4d <_sk_load_565_dst_sse41_8bit+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,98,0,0,0 ; lea 0x62(%rip),%r10 # 27058 <_sk_load_565_dst_sse41_8bit+0x17c> + DB 76,141,21,100,0,0,0 ; lea 0x64(%rip),%r10 # 270a8 <_sk_load_565_dst_sse41_8bit+0x180> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax - DB 102,15,110,216 ; movd %eax,%xmm3 - DB 233,244,254,255,255 ; jmpq 26f01 <_sk_load_565_dst_sse41_8bit+0x25> - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,67,15,196,92,88,4,2 ; pinsrw $0x2,0x4(%r8,%r11,2),%xmm3 - DB 102,67,15,110,20,88 ; movd (%r8,%r11,2),%xmm2 - DB 102,15,58,14,218,3 ; pblendw $0x3,%xmm2,%xmm3 - DB 233,215,254,255,255 ; jmpq 26f01 <_sk_load_565_dst_sse41_8bit+0x25> - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,67,15,196,92,88,12,6 ; pinsrw $0x6,0xc(%r8,%r11,2),%xmm3 - DB 102,67,15,196,92,88,10,5 ; pinsrw $0x5,0xa(%r8,%r11,2),%xmm3 - DB 102,67,15,196,92,88,8,4 ; pinsrw $0x4,0x8(%r8,%r11,2),%xmm3 - DB 243,67,15,126,20,88 ; movq (%r8,%r11,2),%xmm2 - DB 102,15,58,14,218,15 ; pblendw $0xf,%xmm2,%xmm3 - DB 233,170,254,255,255 ; jmpq 26f01 <_sk_load_565_dst_sse41_8bit+0x25> - DB 144 ; nop - DB 167 ; cmpsl %es:(%rdi),%ds:(%rsi) + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 233,242,254,255,255 ; jmpq 26f4d <_sk_load_565_dst_sse41_8bit+0x25> + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,67,15,196,84,88,4,2 ; pinsrw $0x2,0x4(%r8,%r11,2),%xmm2 + DB 102,67,15,110,28,88 ; movd (%r8,%r11,2),%xmm3 + DB 102,15,58,14,211,3 ; pblendw $0x3,%xmm3,%xmm2 + DB 233,213,254,255,255 ; jmpq 26f4d <_sk_load_565_dst_sse41_8bit+0x25> + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,67,15,196,84,88,12,6 ; pinsrw $0x6,0xc(%r8,%r11,2),%xmm2 + DB 102,67,15,196,84,88,10,5 ; pinsrw $0x5,0xa(%r8,%r11,2),%xmm2 + DB 102,67,15,196,84,88,8,4 ; pinsrw $0x4,0x8(%r8,%r11,2),%xmm2 + DB 243,67,15,126,28,88 ; movq (%r8,%r11,2),%xmm3 + DB 102,15,58,14,211,15 ; pblendw $0xf,%xmm3,%xmm2 + DB 233,168,254,255,255 ; jmpq 26f4d <_sk_load_565_dst_sse41_8bit+0x25> + DB 15,31,0 ; nopl (%rax) + DB 165 ; movsl %ds:(%rsi),%es:(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,193 ; inc %ecx DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) - DB 255,181,255,255,255,238 ; pushq -0x11000001(%rbp) DB 255 ; (bad) DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 236 ; in (%dx),%al DB 255 ; (bad) DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255 ; (bad) + DB 220,255 ; fdivr %st,%st(7) + DB 255 ; (bad) + DB 255,208 ; callq *%rax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -37300,13 +37279,11 @@ _sk_load_a8_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,42 ; jne 270b6 <_sk_load_a8_sse41_8bit+0x42> + DB 117,38 ; jne 27102 <_sk_load_a8_sse41_8bit+0x3e> DB 102,67,15,56,48,12,24 ; pmovzxbw (%r8,%r11,1),%xmm1 - DB 102,15,219,13,181,99,0,0 ; pand 0x63b5(%rip),%xmm1 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,219,13,85,98,0,0 ; pand 0x6255(%rip),%xmm1 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 - DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 - DB 102,15,114,241,24 ; pslld $0x18,%xmm1 + DB 102,15,56,0,13,55,106,0,0 ; pshufb 0x6a37(%rip),%xmm1 # 2db30 <_sk_overlay_sse2_8bit+0x1681> DB 102,15,114,240,24 ; pslld $0x18,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -37315,15 +37292,15 @@ _sk_load_a8_sse41_8bit LABEL PROC DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,201 ; ja 27093 <_sk_load_a8_sse41_8bit+0x1f> + DB 119,205 ; ja 270e3 <_sk_load_a8_sse41_8bit+0x1f> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 2714c <_sk_load_a8_sse41_8bit+0xd8> + DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 27198 <_sk_load_a8_sse41_8bit+0xd4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,200 ; movd %eax,%xmm1 - DB 235,170 ; jmp 27093 <_sk_load_a8_sse41_8bit+0x1f> + DB 235,174 ; jmp 270e3 <_sk_load_a8_sse41_8bit+0x1f> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,196,200,2 ; pinsrw $0x2,%eax,%xmm1 @@ -37331,7 +37308,7 @@ _sk_load_a8_sse41_8bit LABEL PROC DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 DB 102,15,58,14,200,3 ; pblendw $0x3,%xmm0,%xmm1 - DB 235,133 ; jmp 27093 <_sk_load_a8_sse41_8bit+0x1f> + DB 235,137 ; jmp 270e3 <_sk_load_a8_sse41_8bit+0x1f> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,196,200,6 ; pinsrw $0x6,%eax,%xmm1 @@ -37342,7 +37319,7 @@ _sk_load_a8_sse41_8bit LABEL PROC DB 102,67,15,110,4,24 ; movd (%r8,%r11,1),%xmm0 DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 DB 102,15,58,14,200,15 ; pblendw $0xf,%xmm0,%xmm1 - DB 233,74,255,255,255 ; jmpq 27093 <_sk_load_a8_sse41_8bit+0x1f> + DB 233,78,255,255,255 ; jmpq 270e3 <_sk_load_a8_sse41_8bit+0x1f> DB 15,31,0 ; nopl (%rax) DB 146 ; xchg %eax,%edx DB 255 ; (bad) @@ -37372,13 +37349,11 @@ _sk_load_a8_dst_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,42 ; jne 271aa <_sk_load_a8_dst_sse41_8bit+0x42> + DB 117,38 ; jne 271f2 <_sk_load_a8_dst_sse41_8bit+0x3e> DB 102,67,15,56,48,28,24 ; pmovzxbw (%r8,%r11,1),%xmm3 - DB 102,15,219,29,193,98,0,0 ; pand 0x62c1(%rip),%xmm3 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,219,29,101,97,0,0 ; pand 0x6165(%rip),%xmm3 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,56,51,211 ; pmovzxwd %xmm3,%xmm2 - DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 - DB 102,15,114,243,24 ; pslld $0x18,%xmm3 + DB 102,15,56,0,29,71,105,0,0 ; pshufb 0x6947(%rip),%xmm3 # 2db30 <_sk_overlay_sse2_8bit+0x1681> DB 102,15,114,242,24 ; pslld $0x18,%xmm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -37387,15 +37362,15 @@ _sk_load_a8_dst_sse41_8bit LABEL PROC DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,201 ; ja 27187 <_sk_load_a8_dst_sse41_8bit+0x1f> + DB 119,205 ; ja 271d3 <_sk_load_a8_dst_sse41_8bit+0x1f> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 27240 <_sk_load_a8_dst_sse41_8bit+0xd8> + DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 27288 <_sk_load_a8_dst_sse41_8bit+0xd4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,216 ; movd %eax,%xmm3 - DB 235,170 ; jmp 27187 <_sk_load_a8_dst_sse41_8bit+0x1f> + DB 235,174 ; jmp 271d3 <_sk_load_a8_dst_sse41_8bit+0x1f> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,196,216,2 ; pinsrw $0x2,%eax,%xmm3 @@ -37403,7 +37378,7 @@ _sk_load_a8_dst_sse41_8bit LABEL PROC DB 102,15,110,208 ; movd %eax,%xmm2 DB 102,15,56,48,210 ; pmovzxbw %xmm2,%xmm2 DB 102,15,58,14,218,3 ; pblendw $0x3,%xmm2,%xmm3 - DB 235,133 ; jmp 27187 <_sk_load_a8_dst_sse41_8bit+0x1f> + DB 235,137 ; jmp 271d3 <_sk_load_a8_dst_sse41_8bit+0x1f> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,196,216,6 ; pinsrw $0x6,%eax,%xmm3 @@ -37414,7 +37389,7 @@ _sk_load_a8_dst_sse41_8bit LABEL PROC DB 102,67,15,110,20,24 ; movd (%r8,%r11,1),%xmm2 DB 102,15,56,48,210 ; pmovzxbw %xmm2,%xmm2 DB 102,15,58,14,218,15 ; pblendw $0xf,%xmm2,%xmm3 - DB 233,74,255,255,255 ; jmpq 27187 <_sk_load_a8_dst_sse41_8bit+0x1f> + DB 233,78,255,255,255 ; jmpq 271d3 <_sk_load_a8_dst_sse41_8bit+0x1f> DB 15,31,0 ; nopl (%rax) DB 146 ; xchg %eax,%edx DB 255 ; (bad) @@ -37443,15 +37418,15 @@ _sk_store_a8_sse41_8bit LABEL PROC DB 77,15,175,193 ; imul %r9,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 102,15,111,45,169,105,0,0 ; movdqa 0x69a9(%rip),%xmm5 # 2dc20 <_sk_overlay_sse2_8bit+0x16c8> + DB 102,15,111,45,129,104,0,0 ; movdqa 0x6881(%rip),%xmm5 # 2db40 <_sk_overlay_sse2_8bit+0x1691> DB 102,15,111,241 ; movdqa %xmm1,%xmm6 DB 102,15,56,0,245 ; pshufb %xmm5,%xmm6 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 DB 102,15,56,0,229 ; pshufb %xmm5,%xmm4 DB 102,15,108,230 ; punpcklqdq %xmm6,%xmm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,19 ; jne 272a5 <_sk_store_a8_sse41_8bit+0x49> - DB 102,15,56,0,37,229,104,0,0 ; pshufb 0x68e5(%rip),%xmm4 # 2db80 <_sk_overlay_sse2_8bit+0x1628> + DB 117,19 ; jne 272ed <_sk_store_a8_sse41_8bit+0x49> + DB 102,15,56,0,37,173,103,0,0 ; pshufb 0x67ad(%rip),%xmm4 # 2da90 <_sk_overlay_sse2_8bit+0x15e1> DB 102,67,15,214,36,24 ; movq %xmm4,(%r8,%r11,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -37459,24 +37434,24 @@ _sk_store_a8_sse41_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 272a1 <_sk_store_a8_sse41_8bit+0x45> + DB 119,236 ; ja 272e9 <_sk_store_a8_sse41_8bit+0x45> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,92,0,0,0 ; lea 0x5c(%rip),%r10 # 2731c <_sk_store_a8_sse41_8bit+0xc0> + DB 76,141,21,92,0,0,0 ; lea 0x5c(%rip),%r10 # 27364 <_sk_store_a8_sse41_8bit+0xc0> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,58,20,36,24,0 ; pextrb $0x0,%xmm4,(%r8,%r11,1) - DB 235,206 ; jmp 272a1 <_sk_store_a8_sse41_8bit+0x45> + DB 235,206 ; jmp 272e9 <_sk_store_a8_sse41_8bit+0x45> DB 102,67,15,58,20,100,24,2,4 ; pextrb $0x4,%xmm4,0x2(%r8,%r11,1) - DB 102,15,56,0,37,123,97,0,0 ; pshufb 0x617b(%rip),%xmm4 # 2d460 <_sk_overlay_sse2_8bit+0xf08> + DB 102,15,56,0,37,35,96,0,0 ; pshufb 0x6023(%rip),%xmm4 # 2d350 <_sk_overlay_sse2_8bit+0xea1> DB 102,67,15,58,21,36,24,0 ; pextrw $0x0,%xmm4,(%r8,%r11,1) - DB 235,178 ; jmp 272a1 <_sk_store_a8_sse41_8bit+0x45> + DB 235,178 ; jmp 272e9 <_sk_store_a8_sse41_8bit+0x45> DB 102,67,15,58,20,100,24,6,12 ; pextrb $0xc,%xmm4,0x6(%r8,%r11,1) DB 102,67,15,58,20,100,24,5,10 ; pextrb $0xa,%xmm4,0x5(%r8,%r11,1) DB 102,67,15,58,20,100,24,4,8 ; pextrb $0x8,%xmm4,0x4(%r8,%r11,1) - DB 102,15,56,0,37,93,97,0,0 ; pshufb 0x615d(%rip),%xmm4 # 2d470 <_sk_overlay_sse2_8bit+0xf18> + DB 102,15,56,0,37,5,96,0,0 ; pshufb 0x6005(%rip),%xmm4 # 2d360 <_sk_overlay_sse2_8bit+0xeb1> DB 102,67,15,126,36,24 ; movd %xmm4,(%r8,%r11,1) - DB 235,134 ; jmp 272a1 <_sk_store_a8_sse41_8bit+0x45> + DB 235,134 ; jmp 272e9 <_sk_store_a8_sse41_8bit+0x45> DB 144 ; nop DB 173 ; lods %ds:(%rsi),%eax DB 255 ; (bad) @@ -37507,16 +37482,16 @@ _sk_load_g8_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 27392 <_sk_load_g8_sse41_8bit+0x5a> + DB 117,68 ; jne 273dc <_sk_load_g8_sse41_8bit+0x5c> DB 102,67,15,56,48,12,24 ; pmovzxbw (%r8,%r11,1),%xmm1 - DB 102,15,219,13,241,96,0,0 ; pand 0x60f1(%rip),%xmm1 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,219,13,153,95,0,0 ; pand 0x5f99(%rip),%xmm1 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 - DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 - DB 102,15,111,37,188,104,0,0 ; movdqa 0x68bc(%rip),%xmm4 # 2dc30 <_sk_overlay_sse2_8bit+0x16d8> + DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1 + DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1 + DB 102,15,111,37,146,103,0,0 ; movdqa 0x6792(%rip),%xmm4 # 2db50 <_sk_overlay_sse2_8bit+0x16a1> DB 102,15,56,64,204 ; pmulld %xmm4,%xmm1 DB 102,15,56,64,196 ; pmulld %xmm4,%xmm0 - DB 102,15,111,37,10,104,0,0 ; movdqa 0x680a(%rip),%xmm4 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,37,208,102,0,0 ; movdqa 0x66d0(%rip),%xmm4 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,235,196 ; por %xmm4,%xmm0 DB 102,15,235,204 ; por %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -37526,15 +37501,15 @@ _sk_load_g8_sse41_8bit LABEL PROC DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,177 ; ja 27357 <_sk_load_g8_sse41_8bit+0x1f> + DB 119,175 ; ja 2739f <_sk_load_g8_sse41_8bit+0x1f> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 27428 <_sk_load_g8_sse41_8bit+0xf0> + DB 76,141,21,121,0,0,0 ; lea 0x79(%rip),%r10 # 27474 <_sk_load_g8_sse41_8bit+0xf4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,200 ; movd %eax,%xmm1 - DB 235,146 ; jmp 27357 <_sk_load_g8_sse41_8bit+0x1f> + DB 235,144 ; jmp 2739f <_sk_load_g8_sse41_8bit+0x1f> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,196,200,2 ; pinsrw $0x2,%eax,%xmm1 @@ -37542,7 +37517,7 @@ _sk_load_g8_sse41_8bit LABEL PROC DB 102,15,110,192 ; movd %eax,%xmm0 DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 DB 102,15,58,14,200,3 ; pblendw $0x3,%xmm0,%xmm1 - DB 233,106,255,255,255 ; jmpq 27357 <_sk_load_g8_sse41_8bit+0x1f> + DB 233,104,255,255,255 ; jmpq 2739f <_sk_load_g8_sse41_8bit+0x1f> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,196,200,6 ; pinsrw $0x6,%eax,%xmm1 @@ -37553,23 +37528,22 @@ _sk_load_g8_sse41_8bit LABEL PROC DB 102,67,15,110,4,24 ; movd (%r8,%r11,1),%xmm0 DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 DB 102,15,58,14,200,15 ; pblendw $0xf,%xmm0,%xmm1 - DB 233,47,255,255,255 ; jmpq 27357 <_sk_load_g8_sse41_8bit+0x1f> - DB 146 ; xchg %eax,%edx - DB 255 ; (bad) + DB 233,45,255,255,255 ; jmpq 2739f <_sk_load_g8_sse41_8bit+0x1f> + DB 102,144 ; xchg %ax,%ax + DB 144 ; nop DB 255 ; (bad) - DB 255,172,255,255,255,157,255 ; ljmp *-0x620001(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,170,255,255,255,155 ; ljmp *-0x64000001(%rdx) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) DB 255 ; (bad) + DB 232,255,255,255,221 ; callq ffffffffde027484 <_sk_overlay_sse2_8bit+0xffffffffddffafd5> DB 255 ; (bad) - DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,212 ; callq *%rsp + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) - DB 255,197 ; inc %ebp + DB 255,195 ; inc %ebx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -37583,16 +37557,16 @@ _sk_load_g8_dst_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,66 ; jne 2749e <_sk_load_g8_dst_sse41_8bit+0x5a> + DB 117,68 ; jne 274ec <_sk_load_g8_dst_sse41_8bit+0x5c> DB 102,67,15,56,48,28,24 ; pmovzxbw (%r8,%r11,1),%xmm3 - DB 102,15,219,29,229,95,0,0 ; pand 0x5fe5(%rip),%xmm3 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,219,29,137,94,0,0 ; pand 0x5e89(%rip),%xmm3 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,56,51,211 ; pmovzxwd %xmm3,%xmm2 - DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 - DB 102,15,111,37,176,103,0,0 ; movdqa 0x67b0(%rip),%xmm4 # 2dc30 <_sk_overlay_sse2_8bit+0x16d8> + DB 102,15,112,219,78 ; pshufd $0x4e,%xmm3,%xmm3 + DB 102,15,56,51,219 ; pmovzxwd %xmm3,%xmm3 + DB 102,15,111,37,130,102,0,0 ; movdqa 0x6682(%rip),%xmm4 # 2db50 <_sk_overlay_sse2_8bit+0x16a1> DB 102,15,56,64,220 ; pmulld %xmm4,%xmm3 DB 102,15,56,64,212 ; pmulld %xmm4,%xmm2 - DB 102,15,111,37,254,102,0,0 ; movdqa 0x66fe(%rip),%xmm4 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,37,192,101,0,0 ; movdqa 0x65c0(%rip),%xmm4 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,220 ; por %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -37602,15 +37576,15 @@ _sk_load_g8_dst_sse41_8bit LABEL PROC DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,177 ; ja 27463 <_sk_load_g8_dst_sse41_8bit+0x1f> + DB 119,175 ; ja 274af <_sk_load_g8_dst_sse41_8bit+0x1f> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,119,0,0,0 ; lea 0x77(%rip),%r10 # 27534 <_sk_load_g8_dst_sse41_8bit+0xf0> + DB 76,141,21,121,0,0,0 ; lea 0x79(%rip),%r10 # 27584 <_sk_load_g8_dst_sse41_8bit+0xf4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,216 ; movd %eax,%xmm3 - DB 235,146 ; jmp 27463 <_sk_load_g8_dst_sse41_8bit+0x1f> + DB 235,144 ; jmp 274af <_sk_load_g8_dst_sse41_8bit+0x1f> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,196,216,2 ; pinsrw $0x2,%eax,%xmm3 @@ -37618,7 +37592,7 @@ _sk_load_g8_dst_sse41_8bit LABEL PROC DB 102,15,110,208 ; movd %eax,%xmm2 DB 102,15,56,48,210 ; pmovzxbw %xmm2,%xmm2 DB 102,15,58,14,218,3 ; pblendw $0x3,%xmm2,%xmm3 - DB 233,106,255,255,255 ; jmpq 27463 <_sk_load_g8_dst_sse41_8bit+0x1f> + DB 233,104,255,255,255 ; jmpq 274af <_sk_load_g8_dst_sse41_8bit+0x1f> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,196,216,6 ; pinsrw $0x6,%eax,%xmm3 @@ -37629,23 +37603,22 @@ _sk_load_g8_dst_sse41_8bit LABEL PROC DB 102,67,15,110,20,24 ; movd (%r8,%r11,1),%xmm2 DB 102,15,56,48,210 ; pmovzxbw %xmm2,%xmm2 DB 102,15,58,14,218,15 ; pblendw $0xf,%xmm2,%xmm3 - DB 233,47,255,255,255 ; jmpq 27463 <_sk_load_g8_dst_sse41_8bit+0x1f> - DB 146 ; xchg %eax,%edx - DB 255 ; (bad) + DB 233,45,255,255,255 ; jmpq 274af <_sk_load_g8_dst_sse41_8bit+0x1f> + DB 102,144 ; xchg %ax,%ax + DB 144 ; nop DB 255 ; (bad) - DB 255,172,255,255,255,157,255 ; ljmp *-0x620001(%rdi,%rdi,8) DB 255 ; (bad) + DB 255,170,255,255,255,155 ; ljmp *-0x64000001(%rdx) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) DB 255 ; (bad) + DB 232,255,255,255,221 ; callq ffffffffde027594 <_sk_overlay_sse2_8bit+0xffffffffddffb0e5> DB 255 ; (bad) - DB 223,255 ; (bad) DB 255 ; (bad) - DB 255,212 ; callq *%rsp + DB 255,210 ; callq *%rdx DB 255 ; (bad) DB 255 ; (bad) - DB 255,197 ; inc %ebp + DB 255,195 ; inc %ebx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -37660,45 +37633,46 @@ _sk_srcover_rgba_8888_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,206,0,0,0 ; jne 2763e <_sk_srcover_rgba_8888_sse41_8bit+0xee> - DB 243,71,15,111,68,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm8 + DB 15,133,215,0,0,0 ; jne 27697 <_sk_srcover_rgba_8888_sse41_8bit+0xf7> DB 243,71,15,111,12,152 ; movdqu (%r8,%r11,4),%xmm9 + DB 243,71,15,111,68,152,16 ; movdqu 0x10(%r8,%r11,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi - DB 102,15,111,37,24,102,0,0 ; movdqa 0x6618(%rip),%xmm4 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,15,111,241 ; movdqa %xmm1,%xmm6 - DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,15,56,0,252 ; pshufb %xmm4,%xmm7 - DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11 - DB 102,69,15,104,218 ; punpckhbw %xmm10,%xmm11 - DB 102,69,15,111,224 ; movdqa %xmm8,%xmm12 - DB 102,69,15,104,226 ; punpckhbw %xmm10,%xmm12 - DB 102,69,15,56,48,233 ; pmovzxbw %xmm9,%xmm13 - DB 102,69,15,56,48,240 ; pmovzxbw %xmm8,%xmm14 - DB 102,15,56,48,231 ; pmovzxbw %xmm7,%xmm4 + DB 102,15,111,53,216,100,0,0 ; movdqa 0x64d8(%rip),%xmm6 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,111,249 ; movdqa %xmm1,%xmm7 + DB 102,15,56,0,254 ; pshufb %xmm6,%xmm7 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,56,0,230 ; pshufb %xmm6,%xmm4 + DB 102,65,15,112,241,78 ; pshufd $0x4e,%xmm9,%xmm6 + DB 102,68,15,56,48,238 ; pmovzxbw %xmm6,%xmm13 + DB 102,69,15,56,48,209 ; pmovzxbw %xmm9,%xmm10 + DB 102,65,15,112,240,78 ; pshufd $0x4e,%xmm8,%xmm6 + DB 102,68,15,56,48,222 ; pmovzxbw %xmm6,%xmm11 + DB 102,69,15,56,48,224 ; pmovzxbw %xmm8,%xmm12 + DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 DB 102,15,56,48,238 ; pmovzxbw %xmm6,%xmm5 - DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 - DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 - DB 102,65,15,213,244 ; pmullw %xmm12,%xmm6 - DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 - DB 102,65,15,213,238 ; pmullw %xmm14,%xmm5 - DB 102,65,15,213,229 ; pmullw %xmm13,%xmm4 - DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 - DB 102,65,15,253,244 ; paddw %xmm12,%xmm6 - DB 102,65,15,253,229 ; paddw %xmm13,%xmm4 - DB 102,65,15,253,238 ; paddw %xmm14,%xmm5 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,56,48,244 ; pmovzxbw %xmm4,%xmm6 + DB 102,15,112,231,78 ; pshufd $0x4e,%xmm7,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,65,15,213,252 ; pmullw %xmm12,%xmm7 + DB 102,65,15,213,227 ; pmullw %xmm11,%xmm4 + DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 + DB 102,65,15,213,237 ; pmullw %xmm13,%xmm5 + DB 102,65,15,253,237 ; paddw %xmm13,%xmm5 + DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 + DB 102,65,15,253,227 ; paddw %xmm11,%xmm4 + DB 102,65,15,253,252 ; paddw %xmm12,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,231 ; packuswb %xmm7,%xmm4 - DB 102,15,103,238 ; packuswb %xmm6,%xmm5 - DB 102,68,15,248,197 ; psubb %xmm5,%xmm8 - DB 102,68,15,248,204 ; psubb %xmm4,%xmm9 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,245 ; packuswb %xmm5,%xmm6 + DB 102,68,15,248,199 ; psubb %xmm7,%xmm8 + DB 102,68,15,248,206 ; psubb %xmm6,%xmm9 DB 102,68,15,252,200 ; paddb %xmm0,%xmm9 DB 102,68,15,252,193 ; paddb %xmm1,%xmm8 - DB 117,78 ; jne 2767b <_sk_srcover_rgba_8888_sse41_8bit+0x12b> + DB 117,78 ; jne 276d4 <_sk_srcover_rgba_8888_sse41_8bit+0x134> DB 243,71,15,127,12,152 ; movdqu %xmm9,(%r8,%r11,4) DB 243,71,15,127,68,152,16 ; movdqu %xmm8,0x10(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -37709,109 +37683,111 @@ _sk_srcover_rgba_8888_sse41_8bit LABEL PROC DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,33,255,255,255 ; ja 2757d <_sk_srcover_rgba_8888_sse41_8bit+0x2d> + DB 15,135,24,255,255,255 ; ja 275cd <_sk_srcover_rgba_8888_sse41_8bit+0x2d> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,189,0,0,0 ; lea 0xbd(%rip),%r10 # 27724 <_sk_srcover_rgba_8888_sse41_8bit+0x1d4> + DB 76,141,21,196,0,0,0 ; lea 0xc4(%rip),%r10 # 27784 <_sk_srcover_rgba_8888_sse41_8bit+0x1e4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,71,15,110,12,152 ; movd (%r8,%r11,4),%xmm9 - DB 233,2,255,255,255 ; jmpq 2757d <_sk_srcover_rgba_8888_sse41_8bit+0x2d> + DB 233,249,254,255,255 ; jmpq 275cd <_sk_srcover_rgba_8888_sse41_8bit+0x2d> DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,175 ; ja 2763a <_sk_srcover_rgba_8888_sse41_8bit+0xea> + DB 119,175 ; ja 27693 <_sk_srcover_rgba_8888_sse41_8bit+0xf3> DB 65,15,182,193 ; movzbl %r9b,%eax - DB 76,141,13,170,0,0,0 ; lea 0xaa(%rip),%r9 # 27740 <_sk_srcover_rgba_8888_sse41_8bit+0x1f0> + DB 76,141,13,177,0,0,0 ; lea 0xb1(%rip),%r9 # 277a0 <_sk_srcover_rgba_8888_sse41_8bit+0x200> DB 73,99,4,129 ; movslq (%r9,%rax,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax DB 102,71,15,126,12,152 ; movd %xmm9,(%r8,%r11,4) - DB 235,147 ; jmp 2763a <_sk_srcover_rgba_8888_sse41_8bit+0xea> - DB 102,67,15,110,100,152,8 ; movd 0x8(%r8,%r11,4),%xmm4 - DB 102,68,15,112,204,69 ; pshufd $0x45,%xmm4,%xmm9 - DB 243,67,15,126,36,152 ; movq (%r8,%r11,4),%xmm4 - DB 102,68,15,58,14,204,15 ; pblendw $0xf,%xmm4,%xmm9 - DB 233,183,254,255,255 ; jmpq 2757d <_sk_srcover_rgba_8888_sse41_8bit+0x2d> + DB 235,147 ; jmp 27693 <_sk_srcover_rgba_8888_sse41_8bit+0xf3> + DB 102,67,15,110,108,152,8 ; movd 0x8(%r8,%r11,4),%xmm5 + DB 102,68,15,112,205,69 ; pshufd $0x45,%xmm5,%xmm9 + DB 243,67,15,126,52,152 ; movq (%r8,%r11,4),%xmm6 + DB 102,68,15,58,14,206,15 ; pblendw $0xf,%xmm6,%xmm9 + DB 233,174,254,255,255 ; jmpq 275cd <_sk_srcover_rgba_8888_sse41_8bit+0x2d> DB 102,67,15,110,100,152,24 ; movd 0x18(%r8,%r11,4),%xmm4 DB 102,68,15,112,196,69 ; pshufd $0x45,%xmm4,%xmm8 DB 102,71,15,58,34,68,152,20,1 ; pinsrd $0x1,0x14(%r8,%r11,4),%xmm8 DB 102,71,15,58,34,68,152,16,0 ; pinsrd $0x0,0x10(%r8,%r11,4),%xmm8 - DB 233,141,254,255,255 ; jmpq 27577 <_sk_srcover_rgba_8888_sse41_8bit+0x27> + DB 243,71,15,111,12,152 ; movdqu (%r8,%r11,4),%xmm9 + DB 233,132,254,255,255 ; jmpq 275cd <_sk_srcover_rgba_8888_sse41_8bit+0x2d> DB 102,71,15,58,22,76,152,8,2 ; pextrd $0x2,%xmm9,0x8(%r8,%r11,4) DB 102,71,15,214,12,152 ; movq %xmm9,(%r8,%r11,4) - DB 233,60,255,255,255 ; jmpq 2763a <_sk_srcover_rgba_8888_sse41_8bit+0xea> + DB 233,54,255,255,255 ; jmpq 27693 <_sk_srcover_rgba_8888_sse41_8bit+0xf3> DB 102,71,15,58,22,68,152,24,2 ; pextrd $0x2,%xmm8,0x18(%r8,%r11,4) DB 102,71,15,58,22,68,152,20,1 ; pextrd $0x1,%xmm8,0x14(%r8,%r11,4) DB 102,71,15,126,68,152,16 ; movd %xmm8,0x10(%r8,%r11,4) DB 243,71,15,127,12,152 ; movdqu %xmm9,(%r8,%r11,4) - DB 233,24,255,255,255 ; jmpq 2763a <_sk_srcover_rgba_8888_sse41_8bit+0xea> - DB 102,144 ; xchg %ax,%ax - DB 76,255 ; rex.WR (bad) + DB 233,18,255,255,255 ; jmpq 27693 <_sk_srcover_rgba_8888_sse41_8bit+0xf3> + DB 15,31,0 ; nopl (%rax) + DB 69,255 ; rex.RB (bad) DB 255 ; (bad) - DB 255,144,255,255,255,131 ; callq *-0x7c000001(%rax) + DB 255,137,255,255,255,124 ; decl 0x7cffffff(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,83,254 ; callq *-0x2(%rbx) DB 255 ; (bad) + DB 186,255,255,255,177 ; mov $0xb1ffffff,%edx DB 255 ; (bad) - DB 184,255,255,255,175 ; mov $0xafffffff,%eax DB 255 ; (bad) + DB 255,168,255,255,255,155 ; ljmp *-0x64000001(%rax) DB 255 ; (bad) - DB 255,162,255,255,255,95 ; jmpq *0x5fffffff(%rdx) DB 255 ; (bad) + DB 255,88,255 ; lcall *-0x1(%rax) DB 255 ; (bad) - DB 255,179,255,255,255,170 ; pushq -0x55000001(%rbx) + DB 255,178,255,255,255,169 ; pushq -0x56000001(%rdx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,215 ; callq *%rdi + DB 255,214 ; callq *%rsi DB 255 ; (bad) DB 255 ; (bad) - DB 255,208 ; callq *%rax + DB 255,207 ; dec %edi DB 255 ; (bad) DB 255 ; (bad) - DB 255,199 ; inc %edi + DB 255,198 ; inc %esi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 190 ; .byte 0xbe + DB 189 ; .byte 0xbd DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_scale_1_float_sse41_8bit _sk_scale_1_float_sse41_8bit LABEL PROC - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,15,16,0 ; movss (%rax),%xmm0 - DB 243,15,89,5,102,91,0,0 ; mulss 0x5b66(%rip),%xmm0 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> - DB 243,15,44,192 ; cvttss2si %xmm0,%eax - DB 15,87,192 ; xorps %xmm0,%xmm0 - DB 102,68,15,56,48,197 ; pmovzxbw %xmm5,%xmm8 - DB 102,15,104,232 ; punpckhbw %xmm0,%xmm5 - DB 102,68,15,56,48,204 ; pmovzxbw %xmm4,%xmm9 - DB 102,15,104,224 ; punpckhbw %xmm0,%xmm4 - DB 102,15,110,240 ; movd %eax,%xmm6 - DB 102,15,56,0,53,166,100,0,0 ; pshufb 0x64a6(%rip),%xmm6 # 2dc40 <_sk_overlay_sse2_8bit+0x16e8> - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,65,15,213,192 ; pmullw %xmm8,%xmm0 - DB 102,15,111,254 ; movdqa %xmm6,%xmm7 - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 - DB 102,15,213,245 ; pmullw %xmm5,%xmm6 - DB 102,15,253,245 ; paddw %xmm5,%xmm6 - DB 102,15,253,252 ; paddw %xmm4,%xmm7 - DB 102,65,15,253,192 ; paddw %xmm8,%xmm0 - DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 243,15,16,32 ; movss (%rax),%xmm4 + DB 243,15,89,37,238,89,0,0 ; mulss 0x59ee(%rip),%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> + DB 243,15,44,196 ; cvttss2si %xmm4,%eax + DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,56,48,232 ; pmovzxbw %xmm0,%xmm5 + DB 102,68,15,56,48,193 ; pmovzxbw %xmm1,%xmm8 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 + DB 102,15,110,192 ; movd %eax,%xmm0 + DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 + DB 242,15,112,192,0 ; pshuflw $0x0,%xmm0,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,65,15,213,200 ; pmullw %xmm8,%xmm1 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,15,213,196 ; pmullw %xmm4,%xmm0 + DB 102,15,253,196 ; paddw %xmm4,%xmm0 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,65,15,253,200 ; paddw %xmm8,%xmm1 + DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,206 ; packuswb %xmm6,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,198 ; packuswb %xmm6,%xmm0 - DB 102,15,103,207 ; packuswb %xmm7,%xmm1 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -37824,72 +37800,73 @@ _sk_scale_u8_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,160,0,0,0 ; jne 278a6 <_sk_scale_u8_sse41_8bit+0xbc> - DB 102,67,15,56,48,52,24 ; pmovzxbw (%r8,%r11,1),%xmm6 - DB 102,15,219,53,59,92,0,0 ; pand 0x5c3b(%rip),%xmm6 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,15,111,254 ; movdqa %xmm6,%xmm7 - DB 102,15,56,0,61,41,100,0,0 ; pshufb 0x6429(%rip),%xmm7 # 2dc50 <_sk_overlay_sse2_8bit+0x16f8> - DB 102,15,56,0,53,48,100,0,0 ; pshufb 0x6430(%rip),%xmm6 # 2dc60 <_sk_overlay_sse2_8bit+0x1708> - DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 - DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 - DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 - DB 102,65,15,104,200 ; punpckhbw %xmm8,%xmm1 - DB 102,15,56,48,230 ; pmovzxbw %xmm6,%xmm4 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,15,56,48,239 ; pmovzxbw %xmm7,%xmm5 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,15,213,249 ; pmullw %xmm1,%xmm7 - DB 102,15,213,240 ; pmullw %xmm0,%xmm6 - DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 - DB 102,65,15,213,225 ; pmullw %xmm9,%xmm4 - DB 102,15,253,240 ; paddw %xmm0,%xmm6 - DB 102,15,253,249 ; paddw %xmm1,%xmm7 - DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 - DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 15,133,170,0,0,0 ; jne 27916 <_sk_scale_u8_sse41_8bit+0xc6> + DB 102,67,15,56,48,36,24 ; pmovzxbw (%r8,%r11,1),%xmm4 + DB 102,15,219,37,197,90,0,0 ; pand 0x5ac5(%rip),%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,15,56,0,45,216,98,0,0 ; pshufb 0x62d8(%rip),%xmm5 # 2db60 <_sk_overlay_sse2_8bit+0x16b1> + DB 102,15,56,0,37,223,98,0,0 ; pshufb 0x62df(%rip),%xmm4 # 2db70 <_sk_overlay_sse2_8bit+0x16c1> + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,68,15,56,48,192 ; pmovzxbw %xmm0,%xmm8 + DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,15,112,196,78 ; pshufd $0x4e,%xmm4,%xmm0 + DB 102,15,56,48,248 ; pmovzxbw %xmm0,%xmm7 + DB 102,15,56,48,196 ; pmovzxbw %xmm4,%xmm0 + DB 102,15,112,205,78 ; pshufd $0x4e,%xmm5,%xmm1 + DB 102,15,56,48,225 ; pmovzxbw %xmm1,%xmm4 + DB 102,15,56,48,205 ; pmovzxbw %xmm5,%xmm1 + DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,15,213,254 ; pmullw %xmm6,%xmm7 + DB 102,15,253,254 ; paddw %xmm6,%xmm7 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,230 ; packuswb %xmm6,%xmm4 - DB 102,15,103,239 ; packuswb %xmm7,%xmm5 + DB 102,15,103,204 ; packuswb %xmm4,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b - DB 102,15,239,246 ; pxor %xmm6,%xmm6 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,79,255,255,255 ; ja 2780d <_sk_scale_u8_sse41_8bit+0x23> + DB 15,135,69,255,255,255 ; ja 27873 <_sk_scale_u8_sse41_8bit+0x23> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,123,0,0,0 ; lea 0x7b(%rip),%r10 # 27944 <_sk_scale_u8_sse41_8bit+0x15a> + DB 76,141,21,123,0,0,0 ; lea 0x7b(%rip),%r10 # 279b4 <_sk_scale_u8_sse41_8bit+0x164> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax - DB 102,15,110,240 ; movd %eax,%xmm6 - DB 233,45,255,255,255 ; jmpq 2780d <_sk_scale_u8_sse41_8bit+0x23> + DB 102,15,110,224 ; movd %eax,%xmm4 + DB 233,35,255,255,255 ; jmpq 27873 <_sk_scale_u8_sse41_8bit+0x23> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax - DB 102,15,239,246 ; pxor %xmm6,%xmm6 - DB 102,15,196,240,2 ; pinsrw $0x2,%eax,%xmm6 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,196,224,2 ; pinsrw $0x2,%eax,%xmm4 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax - DB 102,15,110,224 ; movd %eax,%xmm4 - DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 - DB 102,15,58,14,244,3 ; pblendw $0x3,%xmm4,%xmm6 - DB 233,5,255,255,255 ; jmpq 2780d <_sk_scale_u8_sse41_8bit+0x23> + DB 102,15,110,232 ; movd %eax,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,58,14,229,3 ; pblendw $0x3,%xmm5,%xmm4 + DB 233,251,254,255,255 ; jmpq 27873 <_sk_scale_u8_sse41_8bit+0x23> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax - DB 102,15,239,246 ; pxor %xmm6,%xmm6 - DB 102,15,196,240,6 ; pinsrw $0x6,%eax,%xmm6 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,196,224,6 ; pinsrw $0x6,%eax,%xmm4 DB 67,15,182,68,24,5 ; movzbl 0x5(%r8,%r11,1),%eax - DB 102,15,196,240,5 ; pinsrw $0x5,%eax,%xmm6 + DB 102,15,196,224,5 ; pinsrw $0x5,%eax,%xmm4 DB 67,15,182,68,24,4 ; movzbl 0x4(%r8,%r11,1),%eax - DB 102,15,196,240,4 ; pinsrw $0x4,%eax,%xmm6 - DB 102,67,15,110,36,24 ; movd (%r8,%r11,1),%xmm4 - DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 - DB 102,15,58,14,244,15 ; pblendw $0xf,%xmm4,%xmm6 - DB 233,202,254,255,255 ; jmpq 2780d <_sk_scale_u8_sse41_8bit+0x23> + DB 102,15,196,224,4 ; pinsrw $0x4,%eax,%xmm4 + DB 102,67,15,110,44,24 ; movd (%r8,%r11,1),%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,58,14,229,15 ; pblendw $0xf,%xmm5,%xmm4 + DB 233,192,254,255,255 ; jmpq 27873 <_sk_scale_u8_sse41_8bit+0x23> DB 144 ; nop DB 142,255 ; mov %edi,%? DB 255 ; (bad) @@ -37897,7 +37874,7 @@ _sk_scale_u8_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf027954 <_sk_overlay_sse2_8bit+0xffffffffdeffb3fc> + DB 233,255,255,255,222 ; jmpq ffffffffdf0279c4 <_sk_overlay_sse2_8bit+0xffffffffdeffb515> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -37912,63 +37889,65 @@ PUBLIC _sk_lerp_1_float_sse41_8bit _sk_lerp_1_float_sse41_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,32 ; movss (%rax),%xmm4 - DB 243,15,89,37,106,89,0,0 ; mulss 0x596a(%rip),%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 243,15,89,37,218,87,0,0 ; mulss 0x57da(%rip),%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 243,15,44,196 ; cvttss2si %xmm4,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 - DB 242,15,112,252,0 ; pshuflw $0x0,%xmm4,%xmm7 - DB 102,68,15,112,199,80 ; pshufd $0x50,%xmm7,%xmm8 - DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 + DB 242,15,112,228,0 ; pshuflw $0x0,%xmm4,%xmm4 + DB 102,68,15,112,196,80 ; pshufd $0x50,%xmm4,%xmm8 DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 - DB 102,65,15,104,193 ; punpckhbw %xmm9,%xmm0 - DB 102,68,15,56,48,217 ; pmovzxbw %xmm1,%xmm11 - DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 - DB 102,15,56,0,61,199,98,0,0 ; pshufb 0x62c7(%rip),%xmm7 # 2dc70 <_sk_overlay_sse2_8bit+0x1718> - DB 102,68,15,111,231 ; movdqa %xmm7,%xmm12 - DB 102,69,15,213,227 ; pmullw %xmm11,%xmm12 - DB 102,68,15,111,239 ; movdqa %xmm7,%xmm13 - DB 102,69,15,213,234 ; pmullw %xmm10,%xmm13 + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,15,56,48,252 ; pmovzxbw %xmm4,%xmm7 + DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,15,213,225 ; pmullw %xmm1,%xmm4 + DB 102,15,111,239 ; movdqa %xmm7,%xmm5 + DB 102,65,15,213,233 ; pmullw %xmm9,%xmm5 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,213,241 ; pmullw %xmm1,%xmm6 - DB 102,15,213,248 ; pmullw %xmm0,%xmm7 - DB 102,15,253,248 ; paddw %xmm0,%xmm7 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,69,15,253,234 ; paddw %xmm10,%xmm13 - DB 102,69,15,253,227 ; paddw %xmm11,%xmm12 + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 + DB 102,65,15,213,250 ; pmullw %xmm10,%xmm7 + DB 102,65,15,253,250 ; paddw %xmm10,%xmm7 + DB 102,15,253,240 ; paddw %xmm0,%xmm6 + DB 102,65,15,253,233 ; paddw %xmm9,%xmm5 + DB 102,15,253,225 ; paddw %xmm1,%xmm4 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,236 ; packuswb %xmm4,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 - DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 - DB 102,68,15,103,239 ; packuswb %xmm7,%xmm13 - DB 102,68,15,103,230 ; packuswb %xmm6,%xmm12 - DB 102,15,118,255 ; pcmpeqd %xmm7,%xmm7 - DB 102,65,15,239,248 ; pxor %xmm8,%xmm7 - DB 102,68,15,56,48,194 ; pmovzxbw %xmm2,%xmm8 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,65,15,104,241 ; punpckhbw %xmm9,%xmm6 - DB 102,68,15,56,48,211 ; pmovzxbw %xmm3,%xmm10 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,65,15,104,225 ; punpckhbw %xmm9,%xmm4 - DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 - DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 - DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,213,236 ; pmullw %xmm4,%xmm5 - DB 102,15,213,254 ; pmullw %xmm6,%xmm7 + DB 102,15,103,254 ; packuswb %xmm6,%xmm7 + DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 + DB 102,65,15,239,192 ; pxor %xmm8,%xmm0 + DB 102,15,112,202,78 ; pshufd $0x4e,%xmm2,%xmm1 + DB 102,68,15,56,48,217 ; pmovzxbw %xmm1,%xmm11 + DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 + DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 + DB 102,68,15,56,48,193 ; pmovzxbw %xmm1,%xmm8 + DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,56,48,225 ; pmovzxbw %xmm1,%xmm4 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,65,15,213,202 ; pmullw %xmm10,%xmm1 - DB 102,65,15,213,192 ; pmullw %xmm8,%xmm0 - DB 102,15,253,254 ; paddw %xmm6,%xmm7 - DB 102,15,253,236 ; paddw %xmm4,%xmm5 - DB 102,65,15,253,192 ; paddw %xmm8,%xmm0 - DB 102,65,15,253,202 ; paddw %xmm10,%xmm1 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 + DB 102,15,111,244 ; movdqa %xmm4,%xmm6 + DB 102,65,15,213,240 ; pmullw %xmm8,%xmm6 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,65,15,213,227 ; pmullw %xmm11,%xmm4 + DB 102,65,15,253,227 ; paddw %xmm11,%xmm4 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 + DB 102,65,15,253,240 ; paddw %xmm8,%xmm6 + DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,206 ; packuswb %xmm6,%xmm1 + DB 102,15,252,205 ; paddb %xmm5,%xmm1 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,205 ; packuswb %xmm5,%xmm1 - DB 102,65,15,252,197 ; paddb %xmm13,%xmm0 - DB 102,65,15,252,204 ; paddb %xmm12,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,196 ; packuswb %xmm4,%xmm0 + DB 102,15,252,199 ; paddb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -37981,117 +37960,117 @@ _sk_lerp_u8_sse41_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,46,1,0,0 ; jne 27bcc <_sk_lerp_u8_sse41_8bit+0x14a> - DB 102,67,15,56,48,60,24 ; pmovzxbw (%r8,%r11,1),%xmm7 - DB 102,15,219,61,163,89,0,0 ; pand 0x59a3(%rip),%xmm7 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,56,0,53,145,97,0,0 ; pshufb 0x6191(%rip),%xmm6 # 2dc50 <_sk_overlay_sse2_8bit+0x16f8> - DB 102,15,56,0,61,152,97,0,0 ; pshufb 0x6198(%rip),%xmm7 # 2dc60 <_sk_overlay_sse2_8bit+0x1708> - DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 - DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 + DB 15,133,77,1,0,0 ; jne 27c61 <_sk_lerp_u8_sse41_8bit+0x169> + DB 102,71,15,56,48,4,24 ; pmovzxbw (%r8,%r11,1),%xmm8 + DB 102,68,15,219,5,28,88,0,0 ; pand 0x581c(%rip),%xmm8 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 102,65,15,111,248 ; movdqa %xmm8,%xmm7 + DB 102,15,56,0,61,46,96,0,0 ; pshufb 0x602e(%rip),%xmm7 # 2db60 <_sk_overlay_sse2_8bit+0x16b1> + DB 102,68,15,56,0,5,52,96,0,0 ; pshufb 0x6034(%rip),%xmm8 # 2db70 <_sk_overlay_sse2_8bit+0x16c1> + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,56,48,216 ; pmovzxbw %xmm0,%xmm11 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,68,15,56,48,205 ; pmovzxbw %xmm5,%xmm9 DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 - DB 102,65,15,104,200 ; punpckhbw %xmm8,%xmm1 - DB 102,15,111,231 ; movdqa %xmm7,%xmm4 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,15,111,238 ; movdqa %xmm6,%xmm5 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,56,48,231 ; pmovzxbw %xmm7,%xmm12 - DB 102,68,15,56,48,222 ; pmovzxbw %xmm6,%xmm11 - DB 102,15,213,233 ; pmullw %xmm1,%xmm5 - DB 102,15,213,224 ; pmullw %xmm0,%xmm4 - DB 102,69,15,213,218 ; pmullw %xmm10,%xmm11 - DB 102,69,15,213,225 ; pmullw %xmm9,%xmm12 - DB 102,15,253,224 ; paddw %xmm0,%xmm4 - DB 102,15,253,233 ; paddw %xmm1,%xmm5 - DB 102,69,15,253,225 ; paddw %xmm9,%xmm12 - DB 102,69,15,253,218 ; paddw %xmm10,%xmm11 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 - DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 - DB 102,68,15,103,228 ; packuswb %xmm4,%xmm12 - DB 102,68,15,103,221 ; packuswb %xmm5,%xmm11 - DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 - DB 102,15,239,240 ; pxor %xmm0,%xmm6 - DB 102,15,239,248 ; pxor %xmm0,%xmm7 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,56,48,211 ; pmovzxbw %xmm3,%xmm10 - DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 - DB 102,15,56,48,206 ; pmovzxbw %xmm6,%xmm1 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,213,253 ; pmullw %xmm5,%xmm7 - DB 102,65,15,213,202 ; pmullw %xmm10,%xmm1 + DB 102,65,15,112,232,78 ; pshufd $0x4e,%xmm8,%xmm5 + DB 102,15,56,48,205 ; pmovzxbw %xmm5,%xmm1 + DB 102,69,15,56,48,232 ; pmovzxbw %xmm8,%xmm13 + DB 102,15,112,247,78 ; pshufd $0x4e,%xmm7,%xmm6 + DB 102,15,56,48,198 ; pmovzxbw %xmm6,%xmm0 + DB 102,15,56,48,247 ; pmovzxbw %xmm7,%xmm6 + DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 - DB 102,15,253,253 ; paddw %xmm5,%xmm7 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,69,15,213,235 ; pmullw %xmm11,%xmm13 + DB 102,15,213,204 ; pmullw %xmm4,%xmm1 + DB 102,15,253,204 ; paddw %xmm4,%xmm1 + DB 102,69,15,253,235 ; paddw %xmm11,%xmm13 DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 - DB 102,65,15,253,202 ; paddw %xmm10,%xmm1 + DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,240 ; packuswb %xmm0,%xmm6 + DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,68,15,103,233 ; packuswb %xmm1,%xmm13 + DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 + DB 102,15,239,248 ; pxor %xmm0,%xmm7 + DB 102,68,15,239,192 ; pxor %xmm0,%xmm8 + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 + DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 + DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 + DB 102,68,15,56,48,216 ; pmovzxbw %xmm0,%xmm11 + DB 102,68,15,56,48,227 ; pmovzxbw %xmm3,%xmm12 + DB 102,65,15,112,192,78 ; pshufd $0x4e,%xmm8,%xmm0 + DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 + DB 102,65,15,56,48,192 ; pmovzxbw %xmm8,%xmm0 + DB 102,15,112,207,78 ; pshufd $0x4e,%xmm7,%xmm1 + DB 102,15,56,48,233 ; pmovzxbw %xmm1,%xmm5 + DB 102,15,56,48,207 ; pmovzxbw %xmm7,%xmm1 + DB 102,65,15,213,204 ; pmullw %xmm12,%xmm1 + DB 102,65,15,213,235 ; pmullw %xmm11,%xmm5 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,65,15,213,225 ; pmullw %xmm9,%xmm4 + DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 + DB 102,65,15,253,235 ; paddw %xmm11,%xmm5 + DB 102,65,15,253,204 ; paddw %xmm12,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 + DB 102,15,252,206 ; paddb %xmm6,%xmm1 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 - DB 102,65,15,252,196 ; paddb %xmm12,%xmm0 - DB 102,65,15,252,203 ; paddb %xmm11,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,196 ; packuswb %xmm4,%xmm0 + DB 102,65,15,252,197 ; paddb %xmm13,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b - DB 102,15,239,255 ; pxor %xmm7,%xmm7 + DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,193,254,255,255 ; ja 27aa5 <_sk_lerp_u8_sse41_8bit+0x23> + DB 15,135,161,254,255,255 ; ja 27b1b <_sk_lerp_u8_sse41_8bit+0x23> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,125,0,0,0 ; lea 0x7d(%rip),%r10 # 27c6c <_sk_lerp_u8_sse41_8bit+0x1ea> + DB 76,141,21,131,0,0,0 ; lea 0x83(%rip),%r10 # 27d08 <_sk_lerp_u8_sse41_8bit+0x210> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax - DB 102,15,110,248 ; movd %eax,%xmm7 - DB 233,159,254,255,255 ; jmpq 27aa5 <_sk_lerp_u8_sse41_8bit+0x23> + DB 102,68,15,110,192 ; movd %eax,%xmm8 + DB 233,126,254,255,255 ; jmpq 27b1b <_sk_lerp_u8_sse41_8bit+0x23> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax - DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 102,15,196,248,2 ; pinsrw $0x2,%eax,%xmm7 + DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 + DB 102,68,15,196,192,2 ; pinsrw $0x2,%eax,%xmm8 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax - DB 102,15,110,232 ; movd %eax,%xmm5 - DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 - DB 102,15,58,14,253,3 ; pblendw $0x3,%xmm5,%xmm7 - DB 233,119,254,255,255 ; jmpq 27aa5 <_sk_lerp_u8_sse41_8bit+0x23> + DB 102,15,110,224 ; movd %eax,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,58,14,196,3 ; pblendw $0x3,%xmm4,%xmm8 + DB 233,83,254,255,255 ; jmpq 27b1b <_sk_lerp_u8_sse41_8bit+0x23> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax - DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 102,15,196,248,6 ; pinsrw $0x6,%eax,%xmm7 + DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 + DB 102,68,15,196,192,6 ; pinsrw $0x6,%eax,%xmm8 DB 67,15,182,68,24,5 ; movzbl 0x5(%r8,%r11,1),%eax - DB 102,15,196,248,5 ; pinsrw $0x5,%eax,%xmm7 + DB 102,68,15,196,192,5 ; pinsrw $0x5,%eax,%xmm8 DB 67,15,182,68,24,4 ; movzbl 0x4(%r8,%r11,1),%eax - DB 102,15,196,248,4 ; pinsrw $0x4,%eax,%xmm7 - DB 102,67,15,110,44,24 ; movd (%r8,%r11,1),%xmm5 - DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 - DB 102,15,58,14,253,15 ; pblendw $0xf,%xmm5,%xmm7 - DB 233,60,254,255,255 ; jmpq 27aa5 <_sk_lerp_u8_sse41_8bit+0x23> - DB 15,31,0 ; nopl (%rax) - DB 140,255 ; mov %?,%edi - DB 255 ; (bad) - DB 255,169,255,255,255,154 ; ljmp *-0x65000001(%rcx) + DB 102,68,15,196,192,4 ; pinsrw $0x4,%eax,%xmm8 + DB 102,67,15,110,36,24 ; movd (%r8,%r11,1),%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,58,14,196,15 ; pblendw $0xf,%xmm4,%xmm8 + DB 233,19,254,255,255 ; jmpq 27b1b <_sk_lerp_u8_sse41_8bit+0x23> + DB 134,255 ; xchg %bh,%bh DB 255 ; (bad) + DB 255,166,255,255,255,149 ; jmpq *-0x6a000001(%rsi) DB 255 ; (bad) - DB 255,231 ; jmpq *%rdi DB 255 ; (bad) DB 255 ; (bad) + DB 233,255,255,255,221 ; jmpq ffffffffde027d18 <_sk_overlay_sse2_8bit+0xffffffffddffb869> DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) - DB 255,194 ; inc %edx + DB 255,192 ; inc %eax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -38113,7 +38092,7 @@ _sk_move_dst_src_sse41_8bit LABEL PROC PUBLIC _sk_black_color_sse41_8bit _sk_black_color_sse41_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,5,235,94,0,0 ; movaps 0x5eeb(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 15,40,5,95,93,0,0 ; movaps 0x5d5f(%rip),%xmm0 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 255,224 ; jmpq *%rax @@ -38133,593 +38112,602 @@ _sk_clear_sse41_8bit LABEL PROC PUBLIC _sk_srcatop_sse41_8bit _sk_srcatop_sse41_8bit LABEL PROC - DB 102,68,15,111,21,215,94,0,0 ; movdqa 0x5ed7(%rip),%xmm10 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 - DB 102,68,15,56,48,195 ; pmovzxbw %xmm3,%xmm8 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,65,15,56,0,234 ; pshufb %xmm10,%xmm5 - DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 - DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 + DB 102,68,15,111,5,75,93,0,0 ; movdqa 0x5d4b(%rip),%xmm8 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,68,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm9 + DB 102,68,15,56,48,219 ; pmovzxbw %xmm3,%xmm11 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,65,15,56,0,248 ; pshufb %xmm8,%xmm7 + DB 102,68,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm10 + DB 102,68,15,56,48,226 ; pmovzxbw %xmm2,%xmm12 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,65,15,56,0,226 ; pshufb %xmm10,%xmm4 - DB 102,69,15,239,237 ; pxor %xmm13,%xmm13 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,65,15,104,253 ; punpckhbw %xmm13,%xmm7 - DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,69,15,104,245 ; punpckhbw %xmm13,%xmm14 - DB 102,15,56,48,244 ; pmovzxbw %xmm4,%xmm6 - DB 102,65,15,104,229 ; punpckhbw %xmm13,%xmm4 - DB 102,15,213,231 ; pmullw %xmm7,%xmm4 - DB 102,15,253,231 ; paddw %xmm7,%xmm4 - DB 102,15,56,48,253 ; pmovzxbw %xmm5,%xmm7 - DB 102,65,15,104,237 ; punpckhbw %xmm13,%xmm5 - DB 102,65,15,213,238 ; pmullw %xmm14,%xmm5 - DB 102,65,15,253,238 ; paddw %xmm14,%xmm5 - DB 102,68,15,56,48,240 ; pmovzxbw %xmm0,%xmm14 + DB 102,65,15,56,0,224 ; pshufb %xmm8,%xmm4 + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,68,15,56,48,232 ; pmovzxbw %xmm0,%xmm13 + DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,213,238 ; pmullw %xmm6,%xmm5 + DB 102,15,253,238 ; paddw %xmm6,%xmm5 + DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 + DB 102,68,15,56,48,246 ; pmovzxbw %xmm6,%xmm14 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,65,15,213,229 ; pmullw %xmm13,%xmm4 + DB 102,65,15,253,229 ; paddw %xmm13,%xmm4 + DB 102,15,112,247,78 ; pshufd $0x4e,%xmm7,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 DB 102,65,15,213,246 ; pmullw %xmm14,%xmm6 DB 102,65,15,253,246 ; paddw %xmm14,%xmm6 - DB 102,68,15,56,48,241 ; pmovzxbw %xmm1,%xmm14 - DB 102,65,15,213,254 ; pmullw %xmm14,%xmm7 - DB 102,65,15,253,254 ; paddw %xmm14,%xmm7 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,68,15,56,48,233 ; pmovzxbw %xmm1,%xmm13 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,65,15,213,253 ; pmullw %xmm13,%xmm7 + DB 102,65,15,253,253 ; paddw %xmm13,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,244 ; packuswb %xmm4,%xmm6 - DB 102,15,103,253 ; packuswb %xmm5,%xmm7 - DB 102,65,15,56,0,194 ; pshufb %xmm10,%xmm0 - DB 102,65,15,56,0,202 ; pshufb %xmm10,%xmm1 - DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 - DB 102,15,239,204 ; pxor %xmm4,%xmm1 - DB 102,15,239,196 ; pxor %xmm4,%xmm0 - DB 102,69,15,104,229 ; punpckhbw %xmm13,%xmm12 - DB 102,69,15,104,221 ; punpckhbw %xmm13,%xmm11 - DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 - DB 102,15,56,48,233 ; pmovzxbw %xmm1,%xmm5 - DB 102,65,15,104,197 ; punpckhbw %xmm13,%xmm0 - DB 102,65,15,104,205 ; punpckhbw %xmm13,%xmm1 + DB 102,15,103,254 ; packuswb %xmm6,%xmm7 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,229 ; packuswb %xmm5,%xmm4 + DB 102,65,15,56,0,192 ; pshufb %xmm8,%xmm0 + DB 102,65,15,56,0,200 ; pshufb %xmm8,%xmm1 + DB 102,15,118,237 ; pcmpeqd %xmm5,%xmm5 + DB 102,15,239,205 ; pxor %xmm5,%xmm1 + DB 102,15,239,197 ; pxor %xmm5,%xmm0 + DB 102,69,15,56,48,210 ; pmovzxbw %xmm10,%xmm10 + DB 102,69,15,56,48,193 ; pmovzxbw %xmm9,%xmm8 + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 DB 102,65,15,213,203 ; pmullw %xmm11,%xmm1 - DB 102,65,15,213,196 ; pmullw %xmm12,%xmm0 DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 - DB 102,65,15,213,225 ; pmullw %xmm9,%xmm4 + DB 102,65,15,213,196 ; pmullw %xmm12,%xmm0 + DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 + DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 DB 102,65,15,253,196 ; paddw %xmm12,%xmm0 - DB 102,65,15,253,203 ; paddw %xmm11,%xmm1 - DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 + DB 102,65,15,253,203 ; paddw %xmm11,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,224 ; packuswb %xmm0,%xmm4 - DB 102,15,103,233 ; packuswb %xmm1,%xmm5 - DB 102,15,252,230 ; paddb %xmm6,%xmm4 - DB 102,15,252,239 ; paddb %xmm7,%xmm5 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 + DB 102,15,252,207 ; paddb %xmm7,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,198 ; packuswb %xmm6,%xmm0 + DB 102,15,252,196 ; paddb %xmm4,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax PUBLIC _sk_dstatop_sse41_8bit _sk_dstatop_sse41_8bit LABEL PROC - DB 102,68,15,111,29,156,93,0,0 ; movdqa 0x5d9c(%rip),%xmm11 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 - DB 102,69,15,56,0,235 ; pshufb %xmm11,%xmm13 - DB 102,68,15,111,248 ; movdqa %xmm0,%xmm15 - DB 102,69,15,56,0,251 ; pshufb %xmm11,%xmm15 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,56,48,226 ; pmovzxbw %xmm2,%xmm12 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,56,48,243 ; pmovzxbw %xmm3,%xmm14 - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,69,15,56,48,215 ; pmovzxbw %xmm15,%xmm10 - DB 102,69,15,104,248 ; punpckhbw %xmm8,%xmm15 - DB 102,68,15,213,253 ; pmullw %xmm5,%xmm15 - DB 102,68,15,253,253 ; paddw %xmm5,%xmm15 - DB 102,69,15,56,48,205 ; pmovzxbw %xmm13,%xmm9 - DB 102,69,15,104,232 ; punpckhbw %xmm8,%xmm13 - DB 102,68,15,213,236 ; pmullw %xmm4,%xmm13 - DB 102,68,15,253,236 ; paddw %xmm4,%xmm13 - DB 102,69,15,213,206 ; pmullw %xmm14,%xmm9 - DB 102,69,15,213,212 ; pmullw %xmm12,%xmm10 - DB 102,69,15,253,212 ; paddw %xmm12,%xmm10 - DB 102,69,15,253,206 ; paddw %xmm14,%xmm9 - DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 - DB 102,65,15,113,215,8 ; psrlw $0x8,%xmm15 - DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 - DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 - DB 102,69,15,103,215 ; packuswb %xmm15,%xmm10 - DB 102,69,15,103,205 ; packuswb %xmm13,%xmm9 - DB 102,65,15,56,0,243 ; pshufb %xmm11,%xmm6 - DB 102,65,15,56,0,251 ; pshufb %xmm11,%xmm7 - DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 - DB 102,15,239,252 ; pxor %xmm4,%xmm7 - DB 102,15,239,244 ; pxor %xmm4,%xmm6 - DB 102,68,15,56,48,216 ; pmovzxbw %xmm0,%xmm11 - DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 - DB 102,68,15,56,48,225 ; pmovzxbw %xmm1,%xmm12 - DB 102,65,15,104,200 ; punpckhbw %xmm8,%xmm1 - DB 102,15,56,48,230 ; pmovzxbw %xmm6,%xmm4 - DB 102,15,56,48,239 ; pmovzxbw %xmm7,%xmm5 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,15,213,249 ; pmullw %xmm1,%xmm7 - DB 102,15,213,240 ; pmullw %xmm0,%xmm6 - DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 - DB 102,65,15,213,227 ; pmullw %xmm11,%xmm4 - DB 102,15,253,240 ; paddw %xmm0,%xmm6 - DB 102,15,253,249 ; paddw %xmm1,%xmm7 - DB 102,65,15,253,227 ; paddw %xmm11,%xmm4 - DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 + DB 102,68,15,111,5,3,92,0,0 ; movdqa 0x5c03(%rip),%xmm8 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,68,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm9 + DB 102,68,15,56,48,217 ; pmovzxbw %xmm1,%xmm11 + DB 102,65,15,56,0,200 ; pshufb %xmm8,%xmm1 + DB 102,68,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm10 + DB 102,68,15,56,48,224 ; pmovzxbw %xmm0,%xmm12 + DB 102,65,15,56,0,192 ; pshufb %xmm8,%xmm0 + DB 102,15,112,234,78 ; pshufd $0x4e,%xmm2,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,250 ; pmovzxbw %xmm2,%xmm7 + DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 + DB 102,68,15,56,48,238 ; pmovzxbw %xmm6,%xmm13 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,213,229 ; pmullw %xmm5,%xmm4 + DB 102,15,253,229 ; paddw %xmm5,%xmm4 + DB 102,15,56,48,235 ; pmovzxbw %xmm3,%xmm5 + DB 102,15,56,48,240 ; pmovzxbw %xmm0,%xmm6 + DB 102,15,213,247 ; pmullw %xmm7,%xmm6 + DB 102,15,253,247 ; paddw %xmm7,%xmm6 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,56,48,249 ; pmovzxbw %xmm1,%xmm7 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,65,15,213,197 ; pmullw %xmm13,%xmm0 + DB 102,65,15,253,197 ; paddw %xmm13,%xmm0 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,248 ; packuswb %xmm0,%xmm7 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,230 ; packuswb %xmm6,%xmm4 - DB 102,15,103,239 ; packuswb %xmm7,%xmm5 - DB 102,65,15,252,226 ; paddb %xmm10,%xmm4 - DB 102,65,15,252,233 ; paddb %xmm9,%xmm5 + DB 102,15,103,244 ; packuswb %xmm4,%xmm6 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,65,15,56,0,192 ; pshufb %xmm8,%xmm0 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,65,15,56,0,200 ; pshufb %xmm8,%xmm1 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,239,204 ; pxor %xmm4,%xmm1 + DB 102,15,239,196 ; pxor %xmm4,%xmm0 + DB 102,69,15,56,48,210 ; pmovzxbw %xmm10,%xmm10 + DB 102,69,15,56,48,193 ; pmovzxbw %xmm9,%xmm8 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,65,15,213,203 ; pmullw %xmm11,%xmm1 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 + DB 102,65,15,213,196 ; pmullw %xmm12,%xmm0 + DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 + DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 + DB 102,65,15,253,196 ; paddw %xmm12,%xmm0 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,65,15,253,203 ; paddw %xmm11,%xmm1 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,204 ; packuswb %xmm4,%xmm1 + DB 102,15,252,207 ; paddb %xmm7,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,197 ; packuswb %xmm5,%xmm0 + DB 102,15,252,198 ; paddb %xmm6,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax PUBLIC _sk_srcin_sse41_8bit _sk_srcin_sse41_8bit LABEL PROC - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,83,92,0,0 ; movdqa 0x5c53(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,15,56,0,240 ; pshufb %xmm0,%xmm6 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,56,0,248 ; pshufb %xmm0,%xmm7 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,56,48,205 ; pmovzxbw %xmm5,%xmm9 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,56,48,212 ; pmovzxbw %xmm4,%xmm10 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 - DB 102,15,56,48,206 ; pmovzxbw %xmm6,%xmm1 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,213,253 ; pmullw %xmm5,%xmm7 - DB 102,65,15,213,202 ; pmullw %xmm10,%xmm1 - DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 - DB 102,15,253,253 ; paddw %xmm5,%xmm7 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 - DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 - DB 102,65,15,253,202 ; paddw %xmm10,%xmm1 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,111,37,194,90,0,0 ; movdqa 0x5ac2(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,15,56,0,236 ; pshufb %xmm4,%xmm5 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,68,15,56,48,192 ; pmovzxbw %xmm0,%xmm8 + DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,15,112,198,78 ; pshufd $0x4e,%xmm6,%xmm0 + DB 102,15,56,48,248 ; pmovzxbw %xmm0,%xmm7 + DB 102,15,56,48,198 ; pmovzxbw %xmm6,%xmm0 + DB 102,15,112,205,78 ; pshufd $0x4e,%xmm5,%xmm1 + DB 102,15,56,48,241 ; pmovzxbw %xmm1,%xmm6 + DB 102,15,56,48,205 ; pmovzxbw %xmm5,%xmm1 + DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 + DB 102,65,15,213,240 ; pmullw %xmm8,%xmm6 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 + DB 102,65,15,253,240 ; paddw %xmm8,%xmm6 + DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,206 ; packuswb %xmm6,%xmm1 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstin_sse41_8bit _sk_dstin_sse41_8bit LABEL PROC - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,194,91,0,0 ; movdqa 0x5bc2(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,15,56,0,200 ; pshufb %xmm0,%xmm1 - DB 102,15,56,0,232 ; pshufb %xmm0,%xmm5 - DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,68,15,56,48,194 ; pmovzxbw %xmm2,%xmm8 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 - DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 - DB 102,15,56,48,197 ; pmovzxbw %xmm5,%xmm0 - DB 102,15,56,48,225 ; pmovzxbw %xmm1,%xmm4 - DB 102,65,15,104,234 ; punpckhbw %xmm10,%xmm5 - DB 102,65,15,104,202 ; punpckhbw %xmm10,%xmm1 - DB 102,15,213,206 ; pmullw %xmm6,%xmm1 - DB 102,15,213,239 ; pmullw %xmm7,%xmm5 - DB 102,65,15,213,225 ; pmullw %xmm9,%xmm4 - DB 102,65,15,213,192 ; pmullw %xmm8,%xmm0 - DB 102,15,253,239 ; paddw %xmm7,%xmm5 - DB 102,15,253,206 ; paddw %xmm6,%xmm1 - DB 102,65,15,253,192 ; paddw %xmm8,%xmm0 - DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 + DB 102,15,111,37,35,90,0,0 ; movdqa 0x5a23(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 + DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 + DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,68,15,56,48,195 ; pmovzxbw %xmm3,%xmm8 + DB 102,15,112,248,78 ; pshufd $0x4e,%xmm0,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,65,15,213,200 ; pmullw %xmm8,%xmm1 + DB 102,15,213,238 ; pmullw %xmm6,%xmm5 + DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 + DB 102,15,253,238 ; paddw %xmm6,%xmm5 + DB 102,65,15,253,200 ; paddw %xmm8,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,197 ; packuswb %xmm5,%xmm0 - DB 102,15,103,225 ; packuswb %xmm1,%xmm4 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 DB 255,224 ; jmpq *%rax PUBLIC _sk_srcout_sse41_8bit _sk_srcout_sse41_8bit LABEL PROC - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,41,91,0,0 ; movdqa 0x5b29(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,56,0,248 ; pshufb %xmm0,%xmm7 + DB 102,15,111,37,143,89,0,0 ; movdqa 0x598f(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,15,56,0,236 ; pshufb %xmm4,%xmm5 DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,15,56,0,240 ; pshufb %xmm0,%xmm6 - DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 - DB 102,15,239,240 ; pxor %xmm0,%xmm6 - DB 102,15,239,248 ; pxor %xmm0,%xmm7 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,56,48,205 ; pmovzxbw %xmm5,%xmm9 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,56,48,212 ; pmovzxbw %xmm4,%xmm10 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 + DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,239,244 ; pxor %xmm4,%xmm6 + DB 102,15,239,236 ; pxor %xmm4,%xmm5 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,68,15,56,48,192 ; pmovzxbw %xmm0,%xmm8 + DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,56,48,248 ; pmovzxbw %xmm0,%xmm7 + DB 102,15,56,48,197 ; pmovzxbw %xmm5,%xmm0 + DB 102,15,112,206,78 ; pshufd $0x4e,%xmm6,%xmm1 + DB 102,15,56,48,233 ; pmovzxbw %xmm1,%xmm5 DB 102,15,56,48,206 ; pmovzxbw %xmm6,%xmm1 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,213,253 ; pmullw %xmm5,%xmm7 - DB 102,65,15,213,202 ; pmullw %xmm10,%xmm1 - DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 - DB 102,15,253,253 ; paddw %xmm5,%xmm7 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 - DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 - DB 102,65,15,253,202 ; paddw %xmm10,%xmm1 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 + DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 + DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 + DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstout_sse41_8bit _sk_dstout_sse41_8bit LABEL PROC - DB 102,15,111,37,144,90,0,0 ; movdqa 0x5a90(%rip),%xmm4 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> + DB 102,15,111,37,228,88,0,0 ; movdqa 0x58e4(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 DB 102,15,239,204 ; pxor %xmm4,%xmm1 DB 102,15,239,196 ; pxor %xmm4,%xmm0 - DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,68,15,56,48,194 ; pmovzxbw %xmm2,%xmm8 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 - DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 - DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 - DB 102,15,56,48,233 ; pmovzxbw %xmm1,%xmm5 - DB 102,65,15,104,194 ; punpckhbw %xmm10,%xmm0 - DB 102,65,15,104,202 ; punpckhbw %xmm10,%xmm1 - DB 102,15,213,206 ; pmullw %xmm6,%xmm1 - DB 102,15,213,199 ; pmullw %xmm7,%xmm0 - DB 102,65,15,213,233 ; pmullw %xmm9,%xmm5 - DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 - DB 102,15,253,199 ; paddw %xmm7,%xmm0 - DB 102,15,253,206 ; paddw %xmm6,%xmm1 - DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 - DB 102,65,15,253,233 ; paddw %xmm9,%xmm5 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 + DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,68,15,56,48,195 ; pmovzxbw %xmm3,%xmm8 + DB 102,15,112,248,78 ; pshufd $0x4e,%xmm0,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,65,15,213,200 ; pmullw %xmm8,%xmm1 + DB 102,15,213,238 ; pmullw %xmm6,%xmm5 + DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 + DB 102,15,253,238 ; paddw %xmm6,%xmm5 + DB 102,65,15,253,200 ; paddw %xmm8,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,224 ; packuswb %xmm0,%xmm4 - DB 102,15,103,233 ; packuswb %xmm1,%xmm5 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax PUBLIC _sk_srcover_sse41_8bit _sk_srcover_sse41_8bit LABEL PROC - DB 102,15,111,53,239,89,0,0 ; movdqa 0x59ef(%rip),%xmm6 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 - DB 102,68,15,56,0,222 ; pshufb %xmm6,%xmm11 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,56,0,238 ; pshufb %xmm6,%xmm5 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 - DB 102,15,252,194 ; paddb %xmm2,%xmm0 - DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 - DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 + DB 102,15,111,37,68,88,0,0 ; movdqa 0x5844(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,15,56,0,236 ; pshufb %xmm4,%xmm5 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,68,15,56,48,220 ; pmovzxbw %xmm4,%xmm11 + DB 102,68,15,56,48,194 ; pmovzxbw %xmm2,%xmm8 + DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 + DB 102,68,15,56,48,204 ; pmovzxbw %xmm4,%xmm9 DB 102,68,15,56,48,211 ; pmovzxbw %xmm3,%xmm10 - DB 102,15,252,203 ; paddb %xmm3,%xmm1 - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,15,56,48,229 ; pmovzxbw %xmm5,%xmm4 - DB 102,65,15,56,48,243 ; pmovzxbw %xmm11,%xmm6 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,69,15,104,216 ; punpckhbw %xmm8,%xmm11 - DB 102,68,15,213,223 ; pmullw %xmm7,%xmm11 - DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 - DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 - DB 102,65,15,213,225 ; pmullw %xmm9,%xmm4 - DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 - DB 102,68,15,253,223 ; paddw %xmm7,%xmm11 - DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 - DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 - DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 + DB 102,15,112,230,78 ; pshufd $0x4e,%xmm6,%xmm4 + DB 102,15,56,48,252 ; pmovzxbw %xmm4,%xmm7 + DB 102,15,56,48,230 ; pmovzxbw %xmm6,%xmm4 + DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 + DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 + DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 + DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 + DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,229 ; packuswb %xmm5,%xmm4 - DB 102,65,15,103,243 ; packuswb %xmm11,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,231 ; packuswb %xmm7,%xmm4 + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,248,205 ; psubb %xmm5,%xmm1 + DB 102,15,252,194 ; paddb %xmm2,%xmm0 DB 102,15,248,196 ; psubb %xmm4,%xmm0 - DB 102,15,248,206 ; psubb %xmm6,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_dstover_sse41_8bit _sk_dstover_sse41_8bit LABEL PROC - DB 102,68,15,111,5,63,89,0,0 ; movdqa 0x593f(%rip),%xmm8 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 - DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,68,15,111,5,145,87,0,0 ; movdqa 0x5791(%rip),%xmm8 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,68,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm9 + DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 DB 102,15,252,203 ; paddb %xmm3,%xmm1 DB 102,15,111,251 ; movdqa %xmm3,%xmm7 DB 102,65,15,56,0,248 ; pshufb %xmm8,%xmm7 - DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 DB 102,68,15,56,48,216 ; pmovzxbw %xmm0,%xmm11 DB 102,15,252,194 ; paddb %xmm2,%xmm0 DB 102,15,111,234 ; movdqa %xmm2,%xmm5 DB 102,65,15,56,0,232 ; pshufb %xmm8,%xmm5 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 - DB 102,69,15,104,208 ; punpckhbw %xmm8,%xmm10 - DB 102,15,56,48,245 ; pmovzxbw %xmm5,%xmm6 - DB 102,15,56,48,231 ; pmovzxbw %xmm7,%xmm4 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 + DB 102,68,15,56,48,228 ; pmovzxbw %xmm4,%xmm12 + DB 102,69,15,56,48,193 ; pmovzxbw %xmm9,%xmm8 + DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,112,231,78 ; pshufd $0x4e,%xmm7,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 DB 102,65,15,213,250 ; pmullw %xmm10,%xmm7 - DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 - DB 102,65,15,213,225 ; pmullw %xmm9,%xmm4 - DB 102,65,15,213,243 ; pmullw %xmm11,%xmm6 - DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 + DB 102,65,15,213,235 ; pmullw %xmm11,%xmm5 + DB 102,65,15,213,244 ; pmullw %xmm12,%xmm6 + DB 102,65,15,253,244 ; paddw %xmm12,%xmm6 + DB 102,65,15,253,235 ; paddw %xmm11,%xmm5 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 DB 102,65,15,253,250 ; paddw %xmm10,%xmm7 - DB 102,65,15,253,243 ; paddw %xmm11,%xmm6 - DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 - DB 102,15,103,231 ; packuswb %xmm7,%xmm4 - DB 102,15,248,198 ; psubb %xmm6,%xmm0 - DB 102,15,248,204 ; psubb %xmm4,%xmm1 + DB 102,15,103,238 ; packuswb %xmm6,%xmm5 + DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,248,197 ; psubb %xmm5,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_modulate_sse41_8bit _sk_modulate_sse41_8bit LABEL PROC - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,68,15,56,48,197 ; pmovzxbw %xmm5,%xmm8 - DB 102,65,15,104,234 ; punpckhbw %xmm10,%xmm5 - DB 102,68,15,56,48,204 ; pmovzxbw %xmm4,%xmm9 - DB 102,65,15,104,226 ; punpckhbw %xmm10,%xmm4 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,56,48,240 ; pmovzxbw %xmm0,%xmm6 + DB 102,68,15,56,48,193 ; pmovzxbw %xmm1,%xmm8 + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,15,56,48,248 ; pmovzxbw %xmm0,%xmm7 DB 102,15,56,48,194 ; pmovzxbw %xmm2,%xmm0 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 + DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 + DB 102,15,56,48,233 ; pmovzxbw %xmm1,%xmm5 DB 102,15,56,48,203 ; pmovzxbw %xmm3,%xmm1 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,213,253 ; pmullw %xmm5,%xmm7 - DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 - DB 102,65,15,213,192 ; pmullw %xmm8,%xmm0 - DB 102,15,253,253 ; paddw %xmm5,%xmm7 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 - DB 102,65,15,253,192 ; paddw %xmm8,%xmm0 - DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,213,200 ; pmullw %xmm8,%xmm1 + DB 102,15,213,238 ; pmullw %xmm6,%xmm5 + DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 + DB 102,15,253,238 ; paddw %xmm6,%xmm5 + DB 102,65,15,253,200 ; paddw %xmm8,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_multiply_sse41_8bit _sk_multiply_sse41_8bit LABEL PROC - DB 102,68,15,111,211 ; movdqa %xmm3,%xmm10 - DB 102,15,111,218 ; movdqa %xmm2,%xmm3 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,68,15,111,53,253,87,0,0 ; movdqa 0x57fd(%rip),%xmm14 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,65,15,56,0,238 ; pshufb %xmm14,%xmm5 - DB 102,65,15,111,250 ; movdqa %xmm10,%xmm7 - DB 102,65,15,56,0,254 ; pshufb %xmm14,%xmm7 + DB 72,131,236,40 ; sub $0x28,%rsp + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,68,15,111,37,81,86,0,0 ; movdqa 0x5651(%rip),%xmm12 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) + DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 + DB 102,68,15,56,48,192 ; pmovzxbw %xmm0,%xmm8 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,65,15,56,0,236 ; pshufb %xmm12,%xmm5 + DB 102,68,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm14 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,15,127,28,36 ; movdqa %xmm3,(%rsp) + DB 102,65,15,56,0,244 ; pshufb %xmm12,%xmm6 DB 102,69,15,118,255 ; pcmpeqd %xmm15,%xmm15 - DB 102,65,15,239,255 ; pxor %xmm15,%xmm7 + DB 102,65,15,239,247 ; pxor %xmm15,%xmm6 DB 102,65,15,239,239 ; pxor %xmm15,%xmm5 - DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 - DB 102,68,15,104,236 ; punpckhbw %xmm4,%xmm13 - DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 - DB 102,68,15,104,228 ; punpckhbw %xmm4,%xmm12 - DB 102,68,15,56,48,217 ; pmovzxbw %xmm1,%xmm11 - DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 - DB 102,15,56,48,245 ; pmovzxbw %xmm5,%xmm6 - DB 102,15,104,236 ; punpckhbw %xmm4,%xmm5 - DB 102,65,15,213,237 ; pmullw %xmm13,%xmm5 - DB 102,65,15,213,243 ; pmullw %xmm11,%xmm6 - DB 102,65,15,253,237 ; paddw %xmm13,%xmm5 - DB 102,65,15,253,243 ; paddw %xmm11,%xmm6 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,112,252,78 ; pshufd $0x4e,%xmm4,%xmm7 + DB 102,68,15,56,48,223 ; pmovzxbw %xmm7,%xmm11 + DB 102,68,15,56,48,212 ; pmovzxbw %xmm4,%xmm10 + DB 102,15,112,249,78 ; pshufd $0x4e,%xmm1,%xmm7 + DB 102,68,15,56,48,207 ; pmovzxbw %xmm7,%xmm9 + DB 102,68,15,56,48,233 ; pmovzxbw %xmm1,%xmm13 + DB 102,15,112,254,78 ; pshufd $0x4e,%xmm6,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,65,15,213,245 ; pmullw %xmm13,%xmm6 + DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 + DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 + DB 102,65,15,253,245 ; paddw %xmm13,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 - DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 - DB 102,15,104,252 ; punpckhbw %xmm4,%xmm7 - DB 102,65,15,213,252 ; pmullw %xmm12,%xmm7 - DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 - DB 102,65,15,253,252 ; paddw %xmm12,%xmm7 - DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,247 ; packuswb %xmm7,%xmm6 + DB 102,15,112,253,78 ; pshufd $0x4e,%xmm5,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,56,48,197 ; pmovzxbw %xmm5,%xmm0 + DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 + DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 + DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 + DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,65,15,111,234 ; movdqa %xmm10,%xmm5 - DB 102,65,15,56,0,206 ; pshufb %xmm14,%xmm1 - DB 102,65,15,56,0,214 ; pshufb %xmm14,%xmm2 - DB 102,65,15,239,215 ; pxor %xmm15,%xmm2 + DB 102,15,56,48,235 ; pmovzxbw %xmm3,%xmm5 + DB 102,65,15,56,0,228 ; pshufb %xmm12,%xmm4 + DB 102,65,15,56,0,204 ; pshufb %xmm12,%xmm1 DB 102,65,15,239,207 ; pxor %xmm15,%xmm1 - DB 102,68,15,104,196 ; punpckhbw %xmm4,%xmm8 - DB 102,68,15,104,212 ; punpckhbw %xmm4,%xmm10 - DB 102,15,56,48,249 ; pmovzxbw %xmm1,%xmm7 - DB 102,68,15,56,48,242 ; pmovzxbw %xmm2,%xmm14 - DB 102,15,104,204 ; punpckhbw %xmm4,%xmm1 - DB 102,15,104,212 ; punpckhbw %xmm4,%xmm2 - DB 102,68,15,111,251 ; movdqa %xmm3,%xmm15 - DB 102,65,15,56,48,231 ; pmovzxbw %xmm15,%xmm4 - DB 102,65,15,213,200 ; pmullw %xmm8,%xmm1 - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 - DB 102,65,15,253,200 ; paddw %xmm8,%xmm1 - DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,239,231 ; pxor %xmm15,%xmm4 + DB 102,68,15,56,48,226 ; pmovzxbw %xmm2,%xmm12 + DB 102,65,15,56,48,254 ; pmovzxbw %xmm14,%xmm7 + DB 102,15,112,220,78 ; pshufd $0x4e,%xmm4,%xmm3 + DB 102,15,56,48,219 ; pmovzxbw %xmm3,%xmm3 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 + DB 102,15,56,48,210 ; pmovzxbw %xmm2,%xmm2 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,15,213,205 ; pmullw %xmm5,%xmm1 + DB 102,15,213,215 ; pmullw %xmm7,%xmm2 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 + DB 102,65,15,213,220 ; pmullw %xmm12,%xmm3 + DB 102,65,15,253,220 ; paddw %xmm12,%xmm3 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,15,253,215 ; paddw %xmm7,%xmm2 + DB 102,15,253,205 ; paddw %xmm5,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,249 ; packuswb %xmm1,%xmm7 - DB 102,15,111,221 ; movdqa %xmm5,%xmm3 - DB 102,15,56,48,235 ; pmovzxbw %xmm3,%xmm5 - DB 102,65,15,213,210 ; pmullw %xmm10,%xmm2 - DB 102,68,15,213,245 ; pmullw %xmm5,%xmm14 - DB 102,65,15,253,210 ; paddw %xmm10,%xmm2 - DB 102,68,15,253,245 ; paddw %xmm5,%xmm14 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,65,15,113,214,8 ; psrlw $0x8,%xmm14 - DB 102,68,15,103,242 ; packuswb %xmm2,%xmm14 - DB 102,68,15,252,240 ; paddb %xmm0,%xmm14 - DB 102,15,252,254 ; paddb %xmm6,%xmm7 - DB 102,69,15,213,197 ; pmullw %xmm13,%xmm8 - DB 102,69,15,253,197 ; paddw %xmm13,%xmm8 - DB 102,69,15,213,212 ; pmullw %xmm12,%xmm10 - DB 102,69,15,253,212 ; paddw %xmm12,%xmm10 - DB 102,65,15,213,227 ; pmullw %xmm11,%xmm4 - DB 102,65,15,253,227 ; paddw %xmm11,%xmm4 - DB 102,65,15,213,233 ; pmullw %xmm9,%xmm5 - DB 102,65,15,253,233 ; paddw %xmm9,%xmm5 - DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 + DB 102,15,103,202 ; packuswb %xmm2,%xmm1 + DB 102,15,252,206 ; paddb %xmm6,%xmm1 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,65,15,103,224 ; packuswb %xmm8,%xmm4 - DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,227 ; packuswb %xmm3,%xmm4 + DB 102,15,252,224 ; paddb %xmm0,%xmm4 + DB 102,65,15,213,237 ; pmullw %xmm13,%xmm5 + DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 + DB 102,69,15,213,194 ; pmullw %xmm10,%xmm8 + DB 102,69,15,213,227 ; pmullw %xmm11,%xmm12 + DB 102,69,15,253,227 ; paddw %xmm11,%xmm12 + DB 102,69,15,253,194 ; paddw %xmm10,%xmm8 + DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 + DB 102,65,15,253,237 ; paddw %xmm13,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,65,15,103,234 ; packuswb %xmm10,%xmm5 - DB 102,15,252,231 ; paddb %xmm7,%xmm4 - DB 102,65,15,252,238 ; paddb %xmm14,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,239 ; packuswb %xmm7,%xmm5 + DB 102,15,252,233 ; paddb %xmm1,%xmm5 + DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 + DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 + DB 102,69,15,103,196 ; packuswb %xmm12,%xmm8 + DB 102,68,15,252,196 ; paddb %xmm4,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,65,15,111,215 ; movdqa %xmm15,%xmm2 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 15,40,84,36,16 ; movaps 0x10(%rsp),%xmm2 + DB 15,40,28,36 ; movaps (%rsp),%xmm3 + DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 72,131,196,40 ; add $0x28,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_screen_sse41_8bit _sk_screen_sse41_8bit LABEL PROC - DB 102,69,15,118,228 ; pcmpeqd %xmm12,%xmm12 - DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 - DB 102,69,15,239,220 ; pxor %xmm12,%xmm11 - DB 102,68,15,239,224 ; pxor %xmm0,%xmm12 - DB 102,69,15,56,48,204 ; pmovzxbw %xmm12,%xmm9 - DB 102,69,15,56,48,195 ; pmovzxbw %xmm11,%xmm8 - DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 - DB 102,69,15,104,226 ; punpckhbw %xmm10,%xmm12 - DB 102,69,15,104,218 ; punpckhbw %xmm10,%xmm11 - DB 102,15,56,48,242 ; pmovzxbw %xmm2,%xmm6 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 - DB 102,15,56,48,235 ; pmovzxbw %xmm3,%xmm5 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,65,15,104,226 ; punpckhbw %xmm10,%xmm4 - DB 102,65,15,213,227 ; pmullw %xmm11,%xmm4 - DB 102,65,15,213,252 ; pmullw %xmm12,%xmm7 - DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,15,239,236 ; pxor %xmm4,%xmm5 + DB 102,15,239,224 ; pxor %xmm0,%xmm4 + DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 + DB 102,68,15,56,48,222 ; pmovzxbw %xmm6,%xmm11 + DB 102,68,15,56,48,212 ; pmovzxbw %xmm4,%xmm10 + DB 102,15,112,229,78 ; pshufd $0x4e,%xmm5,%xmm4 + DB 102,68,15,56,48,196 ; pmovzxbw %xmm4,%xmm8 + DB 102,68,15,56,48,205 ; pmovzxbw %xmm5,%xmm9 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,15,56,48,236 ; pmovzxbw %xmm4,%xmm5 + DB 102,15,56,48,226 ; pmovzxbw %xmm2,%xmm4 + DB 102,15,112,251,78 ; pshufd $0x4e,%xmm3,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,56,48,243 ; pmovzxbw %xmm3,%xmm6 DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 + DB 102,65,15,213,248 ; pmullw %xmm8,%xmm7 + DB 102,65,15,213,226 ; pmullw %xmm10,%xmm4 + DB 102,65,15,213,235 ; pmullw %xmm11,%xmm5 + DB 102,65,15,253,235 ; paddw %xmm11,%xmm5 + DB 102,65,15,253,226 ; paddw %xmm10,%xmm4 + DB 102,65,15,253,248 ; paddw %xmm8,%xmm7 DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 - DB 102,65,15,253,252 ; paddw %xmm12,%xmm7 - DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 - DB 102,65,15,253,227 ; paddw %xmm11,%xmm4 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,247 ; packuswb %xmm7,%xmm6 - DB 102,15,103,236 ; packuswb %xmm4,%xmm5 - DB 102,15,252,198 ; paddb %xmm6,%xmm0 - DB 102,15,252,205 ; paddb %xmm5,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,229 ; packuswb %xmm5,%xmm4 + DB 102,15,252,196 ; paddb %xmm4,%xmm0 + DB 102,15,252,206 ; paddb %xmm6,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_xor__sse41_8bit _sk_xor__sse41_8bit LABEL PROC - DB 102,68,15,111,21,187,85,0,0 ; movdqa 0x55bb(%rip),%xmm10 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 - DB 102,68,15,56,48,194 ; pmovzxbw %xmm2,%xmm8 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,56,0,234 ; pshufb %xmm10,%xmm5 - DB 102,68,15,111,235 ; movdqa %xmm3,%xmm13 - DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 + DB 102,68,15,111,5,234,83,0,0 ; movdqa 0x53ea(%rip),%xmm8 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,68,15,112,202,78 ; pshufd $0x4e,%xmm2,%xmm9 + DB 102,68,15,56,48,218 ; pmovzxbw %xmm2,%xmm11 + DB 102,15,111,250 ; movdqa %xmm2,%xmm7 + DB 102,65,15,56,0,248 ; pshufb %xmm8,%xmm7 + DB 102,68,15,112,211,78 ; pshufd $0x4e,%xmm3,%xmm10 + DB 102,68,15,56,48,227 ; pmovzxbw %xmm3,%xmm12 DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,65,15,56,0,226 ; pshufb %xmm10,%xmm4 - DB 102,69,15,118,219 ; pcmpeqd %xmm11,%xmm11 - DB 102,65,15,239,227 ; pxor %xmm11,%xmm4 - DB 102,65,15,239,235 ; pxor %xmm11,%xmm5 - DB 102,69,15,239,246 ; pxor %xmm14,%xmm14 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,65,15,104,254 ; punpckhbw %xmm14,%xmm7 - DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 - DB 102,69,15,104,254 ; punpckhbw %xmm14,%xmm15 - DB 102,15,56,48,245 ; pmovzxbw %xmm5,%xmm6 - DB 102,65,15,104,238 ; punpckhbw %xmm14,%xmm5 - DB 102,15,213,239 ; pmullw %xmm7,%xmm5 - DB 102,15,253,239 ; paddw %xmm7,%xmm5 - DB 102,15,56,48,252 ; pmovzxbw %xmm4,%xmm7 - DB 102,65,15,104,230 ; punpckhbw %xmm14,%xmm4 - DB 102,65,15,213,231 ; pmullw %xmm15,%xmm4 - DB 102,65,15,253,231 ; paddw %xmm15,%xmm4 - DB 102,68,15,56,48,248 ; pmovzxbw %xmm0,%xmm15 - DB 102,65,15,213,247 ; pmullw %xmm15,%xmm6 - DB 102,65,15,253,247 ; paddw %xmm15,%xmm6 - DB 102,68,15,56,48,249 ; pmovzxbw %xmm1,%xmm15 - DB 102,65,15,213,255 ; pmullw %xmm15,%xmm7 - DB 102,65,15,253,255 ; paddw %xmm15,%xmm7 + DB 102,65,15,56,0,224 ; pshufb %xmm8,%xmm4 + DB 102,69,15,118,237 ; pcmpeqd %xmm13,%xmm13 + DB 102,65,15,239,229 ; pxor %xmm13,%xmm4 + DB 102,65,15,239,253 ; pxor %xmm13,%xmm7 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,68,15,56,48,240 ; pmovzxbw %xmm0,%xmm14 + DB 102,15,112,247,78 ; pshufd $0x4e,%xmm7,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,15,213,245 ; pmullw %xmm5,%xmm6 + DB 102,15,253,245 ; paddw %xmm5,%xmm6 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,68,15,56,48,253 ; pmovzxbw %xmm5,%xmm15 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,65,15,213,254 ; pmullw %xmm14,%xmm7 + DB 102,65,15,253,254 ; paddw %xmm14,%xmm7 + DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,65,15,213,239 ; pmullw %xmm15,%xmm5 + DB 102,65,15,253,239 ; paddw %xmm15,%xmm5 + DB 102,68,15,56,48,241 ; pmovzxbw %xmm1,%xmm14 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,65,15,213,230 ; pmullw %xmm14,%xmm4 + DB 102,65,15,253,230 ; paddw %xmm14,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,229 ; packuswb %xmm5,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 - DB 102,15,103,252 ; packuswb %xmm4,%xmm7 - DB 102,65,15,56,0,194 ; pshufb %xmm10,%xmm0 - DB 102,65,15,56,0,202 ; pshufb %xmm10,%xmm1 - DB 102,65,15,239,203 ; pxor %xmm11,%xmm1 - DB 102,65,15,239,195 ; pxor %xmm11,%xmm0 - DB 102,69,15,104,230 ; punpckhbw %xmm14,%xmm12 - DB 102,69,15,104,238 ; punpckhbw %xmm14,%xmm13 - DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 - DB 102,15,56,48,233 ; pmovzxbw %xmm1,%xmm5 - DB 102,65,15,104,198 ; punpckhbw %xmm14,%xmm0 - DB 102,65,15,104,206 ; punpckhbw %xmm14,%xmm1 - DB 102,65,15,213,205 ; pmullw %xmm13,%xmm1 - DB 102,65,15,213,196 ; pmullw %xmm12,%xmm0 - DB 102,65,15,213,233 ; pmullw %xmm9,%xmm5 - DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 - DB 102,65,15,253,196 ; paddw %xmm12,%xmm0 - DB 102,65,15,253,205 ; paddw %xmm13,%xmm1 - DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 - DB 102,65,15,253,233 ; paddw %xmm9,%xmm5 + DB 102,15,103,254 ; packuswb %xmm6,%xmm7 + DB 102,65,15,56,0,192 ; pshufb %xmm8,%xmm0 + DB 102,65,15,56,0,200 ; pshufb %xmm8,%xmm1 + DB 102,65,15,239,205 ; pxor %xmm13,%xmm1 + DB 102,65,15,239,197 ; pxor %xmm13,%xmm0 + DB 102,69,15,56,48,201 ; pmovzxbw %xmm9,%xmm9 + DB 102,69,15,56,48,194 ; pmovzxbw %xmm10,%xmm8 + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,65,15,213,204 ; pmullw %xmm12,%xmm1 + DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 + DB 102,65,15,213,195 ; pmullw %xmm11,%xmm0 + DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 + DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 + DB 102,65,15,253,195 ; paddw %xmm11,%xmm0 + DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 + DB 102,65,15,253,204 ; paddw %xmm12,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,224 ; packuswb %xmm0,%xmm4 - DB 102,15,103,233 ; packuswb %xmm1,%xmm5 - DB 102,15,252,230 ; paddb %xmm6,%xmm4 - DB 102,15,252,239 ; paddb %xmm7,%xmm5 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 + DB 102,15,252,204 ; paddb %xmm4,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,198 ; packuswb %xmm6,%xmm0 + DB 102,15,252,199 ; paddb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax PUBLIC _sk_plus__sse41_8bit @@ -38731,682 +38719,624 @@ _sk_plus__sse41_8bit LABEL PROC PUBLIC _sk_darken_sse41_8bit _sk_darken_sse41_8bit LABEL PROC - DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 - DB 102,68,15,111,5,98,84,0,0 ; movdqa 0x5462(%rip),%xmm8 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 - DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,56,0,240 ; pshufb %xmm8,%xmm6 - DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 - DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,56,0,232 ; pshufb %xmm8,%xmm5 - DB 102,69,15,239,228 ; pxor %xmm12,%xmm12 - DB 102,65,15,104,196 ; punpckhbw %xmm12,%xmm0 - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 - DB 102,69,15,56,48,254 ; pmovzxbw %xmm14,%xmm15 - DB 102,15,56,48,253 ; pmovzxbw %xmm5,%xmm7 - DB 102,65,15,104,236 ; punpckhbw %xmm12,%xmm5 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,15,56,48,198 ; pmovzxbw %xmm6,%xmm0 - DB 102,65,15,104,244 ; punpckhbw %xmm12,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 - DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 - DB 102,68,15,252,242 ; paddb %xmm2,%xmm14 - DB 102,65,15,213,255 ; pmullw %xmm15,%xmm7 - DB 102,65,15,253,255 ; paddw %xmm15,%xmm7 - DB 102,68,15,56,48,249 ; pmovzxbw %xmm1,%xmm15 - DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 - DB 102,65,15,253,199 ; paddw %xmm15,%xmm0 - DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 - DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 72,131,236,40 ; sub $0x28,%rsp + DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9 + DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10 + DB 102,15,111,37,124,82,0,0 ; movdqa 0x527c(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3 + DB 102,15,127,92,36,16 ; movdqa %xmm3,0x10(%rsp) + DB 102,69,15,56,48,217 ; pmovzxbw %xmm9,%xmm11 + DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 + DB 102,68,15,56,0,196 ; pshufb %xmm4,%xmm8 + DB 102,69,15,112,226,78 ; pshufd $0x4e,%xmm10,%xmm12 + DB 102,69,15,56,48,242 ; pmovzxbw %xmm10,%xmm14 + DB 102,69,15,111,250 ; movdqa %xmm10,%xmm15 + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,127,28,36 ; movdqa %xmm3,(%rsp) + DB 102,68,15,56,48,234 ; pmovzxbw %xmm2,%xmm13 + DB 102,68,15,252,210 ; paddb %xmm2,%xmm10 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 + DB 102,65,15,56,48,236 ; pmovzxbw %xmm12,%xmm5 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,254,78 ; pshufd $0x4e,%xmm6,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,68,15,56,48,225 ; pmovzxbw %xmm1,%xmm12 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,65,15,213,246 ; pmullw %xmm14,%xmm6 + DB 102,65,15,253,246 ; paddw %xmm14,%xmm6 + DB 102,65,15,112,216,78 ; pshufd $0x4e,%xmm8,%xmm3 + DB 102,15,56,48,219 ; pmovzxbw %xmm3,%xmm3 + DB 102,15,213,216 ; pmullw %xmm0,%xmm3 + DB 102,15,253,216 ; paddw %xmm0,%xmm3 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,65,15,252,201 ; paddb %xmm9,%xmm1 + DB 102,65,15,56,48,232 ; pmovzxbw %xmm8,%xmm5 + DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 + DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,253 ; packuswb %xmm5,%xmm7 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,235 ; packuswb %xmm3,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,247 ; packuswb %xmm7,%xmm6 + DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 + DB 102,68,15,56,0,252 ; pshufb %xmm4,%xmm15 + DB 102,68,15,56,48,36,36 ; pmovzxbw (%rsp),%xmm12 + DB 102,68,15,56,48,68,36,16 ; pmovzxbw 0x10(%rsp),%xmm8 + DB 102,65,15,112,231,78 ; pshufd $0x4e,%xmm15,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,65,15,56,48,255 ; pmovzxbw %xmm15,%xmm7 + DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 + DB 102,15,56,48,219 ; pmovzxbw %xmm3,%xmm3 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,65,15,213,195 ; pmullw %xmm11,%xmm0 + DB 102,65,15,213,216 ; pmullw %xmm8,%xmm3 + DB 102,65,15,213,253 ; pmullw %xmm13,%xmm7 + DB 102,65,15,213,228 ; pmullw %xmm12,%xmm4 + DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 + DB 102,65,15,253,253 ; paddw %xmm13,%xmm7 + DB 102,65,15,253,216 ; paddw %xmm8,%xmm3 + DB 102,65,15,253,195 ; paddw %xmm11,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,198 ; packuswb %xmm6,%xmm0 - DB 102,69,15,56,0,248 ; pshufb %xmm8,%xmm15 - DB 102,65,15,56,0,224 ; pshufb %xmm8,%xmm4 - DB 102,69,15,104,236 ; punpckhbw %xmm12,%xmm13 - DB 102,69,15,104,220 ; punpckhbw %xmm12,%xmm11 - DB 102,15,56,48,236 ; pmovzxbw %xmm4,%xmm5 - DB 102,65,15,56,48,247 ; pmovzxbw %xmm15,%xmm6 - DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 - DB 102,69,15,104,252 ; punpckhbw %xmm12,%xmm15 - DB 102,69,15,213,251 ; pmullw %xmm11,%xmm15 - DB 102,65,15,213,229 ; pmullw %xmm13,%xmm4 - DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 - DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 - DB 102,65,15,253,229 ; paddw %xmm13,%xmm4 - DB 102,69,15,253,251 ; paddw %xmm11,%xmm15 - DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 - DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 - DB 102,65,15,113,215,8 ; psrlw $0x8,%xmm15 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,195 ; packuswb %xmm3,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,236 ; packuswb %xmm4,%xmm5 - DB 102,65,15,103,247 ; packuswb %xmm15,%xmm6 - DB 102,15,222,253 ; pmaxub %xmm5,%xmm7 - DB 102,15,222,198 ; pmaxub %xmm6,%xmm0 - DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 - DB 102,68,15,248,192 ; psubb %xmm0,%xmm8 - DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 - DB 102,15,248,231 ; psubb %xmm7,%xmm4 - DB 102,15,248,206 ; psubb %xmm6,%xmm1 - DB 102,68,15,248,245 ; psubb %xmm5,%xmm14 - DB 15,40,5,246,83,0,0 ; movaps 0x53f6(%rip),%xmm0 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> - DB 102,68,15,56,16,244 ; pblendvb %xmm0,%xmm4,%xmm14 - DB 102,65,15,56,16,200 ; pblendvb %xmm0,%xmm8,%xmm1 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,222,247 ; pmaxub %xmm7,%xmm6 + DB 102,15,222,232 ; pmaxub %xmm0,%xmm5 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,248,221 ; psubb %xmm5,%xmm3 + DB 102,65,15,111,226 ; movdqa %xmm10,%xmm4 + DB 102,15,248,230 ; psubb %xmm6,%xmm4 + DB 102,15,248,200 ; psubb %xmm0,%xmm1 + DB 102,68,15,248,215 ; psubb %xmm7,%xmm10 + DB 15,40,5,218,81,0,0 ; movaps 0x51da(%rip),%xmm0 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> + DB 102,68,15,56,16,212 ; pblendvb %xmm0,%xmm4,%xmm10 + DB 102,15,56,16,203 ; pblendvb %xmm0,%xmm3,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 + DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 + DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 + DB 72,131,196,40 ; add $0x28,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_lighten_sse41_8bit _sk_lighten_sse41_8bit LABEL PROC - DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 - DB 102,68,15,111,5,243,82,0,0 ; movdqa 0x52f3(%rip),%xmm8 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 - DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,56,0,240 ; pshufb %xmm8,%xmm6 - DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 - DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,56,0,232 ; pshufb %xmm8,%xmm5 - DB 102,69,15,239,228 ; pxor %xmm12,%xmm12 - DB 102,65,15,104,196 ; punpckhbw %xmm12,%xmm0 - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 - DB 102,69,15,56,48,254 ; pmovzxbw %xmm14,%xmm15 - DB 102,15,56,48,253 ; pmovzxbw %xmm5,%xmm7 - DB 102,65,15,104,236 ; punpckhbw %xmm12,%xmm5 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,15,56,48,198 ; pmovzxbw %xmm6,%xmm0 - DB 102,65,15,104,244 ; punpckhbw %xmm12,%xmm6 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,15,253,244 ; paddw %xmm4,%xmm6 - DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 - DB 102,68,15,252,242 ; paddb %xmm2,%xmm14 - DB 102,65,15,213,255 ; pmullw %xmm15,%xmm7 - DB 102,65,15,253,255 ; paddw %xmm15,%xmm7 - DB 102,68,15,56,48,249 ; pmovzxbw %xmm1,%xmm15 - DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 - DB 102,65,15,253,199 ; paddw %xmm15,%xmm0 - DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 - DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 72,131,236,40 ; sub $0x28,%rsp + DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9 + DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10 + DB 102,15,111,37,215,80,0,0 ; movdqa 0x50d7(%rip),%xmm4 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3 + DB 102,15,127,92,36,16 ; movdqa %xmm3,0x10(%rsp) + DB 102,69,15,56,48,217 ; pmovzxbw %xmm9,%xmm11 + DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 + DB 102,68,15,56,0,196 ; pshufb %xmm4,%xmm8 + DB 102,69,15,112,226,78 ; pshufd $0x4e,%xmm10,%xmm12 + DB 102,69,15,56,48,242 ; pmovzxbw %xmm10,%xmm14 + DB 102,69,15,111,250 ; movdqa %xmm10,%xmm15 + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,127,28,36 ; movdqa %xmm3,(%rsp) + DB 102,68,15,56,48,234 ; pmovzxbw %xmm2,%xmm13 + DB 102,68,15,252,210 ; paddb %xmm2,%xmm10 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 + DB 102,65,15,56,48,236 ; pmovzxbw %xmm12,%xmm5 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,15,112,254,78 ; pshufd $0x4e,%xmm6,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,68,15,56,48,225 ; pmovzxbw %xmm1,%xmm12 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,65,15,213,246 ; pmullw %xmm14,%xmm6 + DB 102,65,15,253,246 ; paddw %xmm14,%xmm6 + DB 102,65,15,112,216,78 ; pshufd $0x4e,%xmm8,%xmm3 + DB 102,15,56,48,219 ; pmovzxbw %xmm3,%xmm3 + DB 102,15,213,216 ; pmullw %xmm0,%xmm3 + DB 102,15,253,216 ; paddw %xmm0,%xmm3 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,65,15,252,201 ; paddb %xmm9,%xmm1 + DB 102,65,15,56,48,232 ; pmovzxbw %xmm8,%xmm5 + DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 + DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,253 ; packuswb %xmm5,%xmm7 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,235 ; packuswb %xmm3,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,247 ; packuswb %xmm7,%xmm6 + DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 + DB 102,68,15,56,0,252 ; pshufb %xmm4,%xmm15 + DB 102,68,15,56,48,36,36 ; pmovzxbw (%rsp),%xmm12 + DB 102,68,15,56,48,68,36,16 ; pmovzxbw 0x10(%rsp),%xmm8 + DB 102,65,15,112,231,78 ; pshufd $0x4e,%xmm15,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,65,15,56,48,255 ; pmovzxbw %xmm15,%xmm7 + DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 + DB 102,15,56,48,219 ; pmovzxbw %xmm3,%xmm3 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,65,15,213,195 ; pmullw %xmm11,%xmm0 + DB 102,65,15,213,216 ; pmullw %xmm8,%xmm3 + DB 102,65,15,213,253 ; pmullw %xmm13,%xmm7 + DB 102,65,15,213,228 ; pmullw %xmm12,%xmm4 + DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 + DB 102,65,15,253,253 ; paddw %xmm13,%xmm7 + DB 102,65,15,253,216 ; paddw %xmm8,%xmm3 + DB 102,65,15,253,195 ; paddw %xmm11,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,198 ; packuswb %xmm6,%xmm0 - DB 102,69,15,56,0,248 ; pshufb %xmm8,%xmm15 - DB 102,65,15,56,0,224 ; pshufb %xmm8,%xmm4 - DB 102,69,15,104,236 ; punpckhbw %xmm12,%xmm13 - DB 102,69,15,104,220 ; punpckhbw %xmm12,%xmm11 - DB 102,15,56,48,236 ; pmovzxbw %xmm4,%xmm5 - DB 102,65,15,56,48,247 ; pmovzxbw %xmm15,%xmm6 - DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 - DB 102,69,15,104,252 ; punpckhbw %xmm12,%xmm15 - DB 102,69,15,213,251 ; pmullw %xmm11,%xmm15 - DB 102,65,15,213,229 ; pmullw %xmm13,%xmm4 - DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 - DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 - DB 102,65,15,253,229 ; paddw %xmm13,%xmm4 - DB 102,69,15,253,251 ; paddw %xmm11,%xmm15 - DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 - DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 - DB 102,65,15,113,215,8 ; psrlw $0x8,%xmm15 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,195 ; packuswb %xmm3,%xmm0 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,236 ; packuswb %xmm4,%xmm5 - DB 102,65,15,103,247 ; packuswb %xmm15,%xmm6 - DB 102,15,218,253 ; pminub %xmm5,%xmm7 - DB 102,15,218,198 ; pminub %xmm6,%xmm0 - DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 - DB 102,68,15,248,192 ; psubb %xmm0,%xmm8 - DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 - DB 102,15,248,231 ; psubb %xmm7,%xmm4 - DB 102,15,248,206 ; psubb %xmm6,%xmm1 - DB 102,68,15,248,245 ; psubb %xmm5,%xmm14 - DB 15,40,5,135,82,0,0 ; movaps 0x5287(%rip),%xmm0 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> - DB 102,68,15,56,16,244 ; pblendvb %xmm0,%xmm4,%xmm14 - DB 102,65,15,56,16,200 ; pblendvb %xmm0,%xmm8,%xmm1 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,218,247 ; pminub %xmm7,%xmm6 + DB 102,15,218,232 ; pminub %xmm0,%xmm5 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,248,221 ; psubb %xmm5,%xmm3 + DB 102,65,15,111,226 ; movdqa %xmm10,%xmm4 + DB 102,15,248,230 ; psubb %xmm6,%xmm4 + DB 102,15,248,200 ; psubb %xmm0,%xmm1 + DB 102,68,15,248,215 ; psubb %xmm7,%xmm10 + DB 15,40,5,53,80,0,0 ; movaps 0x5035(%rip),%xmm0 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> + DB 102,68,15,56,16,212 ; pblendvb %xmm0,%xmm4,%xmm10 + DB 102,15,56,16,203 ; pblendvb %xmm0,%xmm3,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 + DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 + DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 + DB 72,131,196,40 ; add $0x28,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_exclusion_sse41_8bit _sk_exclusion_sse41_8bit LABEL PROC - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,111,216 ; movdqa %xmm0,%xmm11 - DB 102,69,15,104,216 ; punpckhbw %xmm8,%xmm11 - DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 - DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 - DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 - DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 - DB 102,15,56,48,250 ; pmovzxbw %xmm2,%xmm7 - DB 102,15,252,194 ; paddb %xmm2,%xmm0 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,15,56,48,227 ; pmovzxbw %xmm3,%xmm4 - DB 102,15,252,203 ; paddb %xmm3,%xmm1 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,65,15,213,244 ; pmullw %xmm12,%xmm6 - DB 102,65,15,213,235 ; pmullw %xmm11,%xmm5 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,68,15,56,48,220 ; pmovzxbw %xmm4,%xmm11 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 + DB 102,68,15,56,48,196 ; pmovzxbw %xmm4,%xmm8 + DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,15,56,48,252 ; pmovzxbw %xmm4,%xmm7 + DB 102,15,56,48,226 ; pmovzxbw %xmm2,%xmm4 + DB 102,15,112,235,78 ; pshufd $0x4e,%xmm3,%xmm5 + DB 102,15,56,48,245 ; pmovzxbw %xmm5,%xmm6 + DB 102,15,56,48,235 ; pmovzxbw %xmm3,%xmm5 + DB 102,65,15,213,233 ; pmullw %xmm9,%xmm5 + DB 102,65,15,213,240 ; pmullw %xmm8,%xmm6 DB 102,65,15,213,226 ; pmullw %xmm10,%xmm4 - DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 - DB 102,65,15,253,235 ; paddw %xmm11,%xmm5 - DB 102,65,15,253,244 ; paddw %xmm12,%xmm6 - DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 + DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 + DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 DB 102,65,15,253,226 ; paddw %xmm10,%xmm4 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,65,15,253,240 ; paddw %xmm8,%xmm6 + DB 102,65,15,253,233 ; paddw %xmm9,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,253 ; packuswb %xmm5,%xmm7 - DB 102,15,103,230 ; packuswb %xmm6,%xmm4 - DB 102,15,111,45,221,81,0,0 ; movdqa 0x51dd(%rip),%xmm5 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> - DB 102,15,248,204 ; psubb %xmm4,%xmm1 - DB 102,15,219,229 ; pand %xmm5,%xmm4 - DB 102,15,219,239 ; pand %xmm7,%xmm5 - DB 102,15,248,199 ; psubb %xmm7,%xmm0 - DB 102,15,248,197 ; psubb %xmm5,%xmm0 - DB 102,15,248,204 ; psubb %xmm4,%xmm1 + DB 102,15,103,231 ; packuswb %xmm7,%xmm4 + DB 102,15,111,53,140,79,0,0 ; movdqa 0x4f8c(%rip),%xmm6 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,248,205 ; psubb %xmm5,%xmm1 + DB 102,15,219,238 ; pand %xmm6,%xmm5 + DB 102,15,219,244 ; pand %xmm4,%xmm6 + DB 102,15,252,194 ; paddb %xmm2,%xmm0 + DB 102,15,248,205 ; psubb %xmm5,%xmm1 + DB 102,15,248,196 ; psubb %xmm4,%xmm0 + DB 102,15,248,198 ; psubb %xmm6,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_difference_sse41_8bit _sk_difference_sse41_8bit LABEL PROC - DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 - DB 102,15,111,13,212,80,0,0 ; movdqa 0x50d4(%rip),%xmm1 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 - DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,15,56,0,241 ; pshufb %xmm1,%xmm6 - DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 - DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 - DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,15,56,0,225 ; pshufb %xmm1,%xmm4 - DB 102,69,15,239,228 ; pxor %xmm12,%xmm12 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,65,15,104,236 ; punpckhbw %xmm12,%xmm5 - DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 - DB 102,65,15,104,204 ; punpckhbw %xmm12,%xmm1 - DB 102,68,15,56,48,248 ; pmovzxbw %xmm0,%xmm15 - DB 102,68,15,56,48,244 ; pmovzxbw %xmm4,%xmm14 - DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 - DB 102,15,213,229 ; pmullw %xmm5,%xmm4 - DB 102,15,253,229 ; paddw %xmm5,%xmm4 - DB 102,15,56,48,254 ; pmovzxbw %xmm6,%xmm7 - DB 102,65,15,104,244 ; punpckhbw %xmm12,%xmm6 - DB 102,15,213,241 ; pmullw %xmm1,%xmm6 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 72,131,236,40 ; sub $0x28,%rsp + DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9 + DB 102,68,15,111,21,134,78,0,0 ; movdqa 0x4e86(%rip),%xmm10 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,65,15,112,225,78 ; pshufd $0x4e,%xmm9,%xmm4 + DB 102,15,127,100,36,16 ; movdqa %xmm4,0x10(%rsp) + DB 102,69,15,56,48,217 ; pmovzxbw %xmm9,%xmm11 + DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8 + DB 102,69,15,56,0,194 ; pshufb %xmm10,%xmm8 + DB 102,68,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm12 + DB 102,68,15,56,48,240 ; pmovzxbw %xmm0,%xmm14 + DB 102,68,15,111,248 ; movdqa %xmm0,%xmm15 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 + DB 102,15,127,36,36 ; movdqa %xmm4,(%rsp) + DB 102,68,15,56,48,234 ; pmovzxbw %xmm2,%xmm13 DB 102,15,252,194 ; paddb %xmm2,%xmm0 - DB 102,69,15,213,247 ; pmullw %xmm15,%xmm14 - DB 102,69,15,253,247 ; paddw %xmm15,%xmm14 - DB 102,65,15,56,48,200 ; pmovzxbw %xmm8,%xmm1 - DB 102,15,213,249 ; pmullw %xmm1,%xmm7 - DB 102,15,253,249 ; paddw %xmm1,%xmm7 - DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 - DB 102,68,15,252,195 ; paddb %xmm3,%xmm8 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,65,15,113,214,8 ; psrlw $0x8,%xmm14 - DB 102,68,15,103,244 ; packuswb %xmm4,%xmm14 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,254 ; packuswb %xmm6,%xmm7 - DB 102,15,111,37,25,80,0,0 ; movdqa 0x5019(%rip),%xmm4 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 - DB 102,15,56,0,236 ; pshufb %xmm4,%xmm5 - DB 102,69,15,104,236 ; punpckhbw %xmm12,%xmm13 - DB 102,69,15,104,220 ; punpckhbw %xmm12,%xmm11 - DB 102,15,56,48,229 ; pmovzxbw %xmm5,%xmm4 - DB 102,15,56,48,241 ; pmovzxbw %xmm1,%xmm6 - DB 102,65,15,104,236 ; punpckhbw %xmm12,%xmm5 - DB 102,65,15,104,204 ; punpckhbw %xmm12,%xmm1 - DB 102,65,15,213,203 ; pmullw %xmm11,%xmm1 - DB 102,65,15,213,237 ; pmullw %xmm13,%xmm5 - DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 - DB 102,65,15,213,226 ; pmullw %xmm10,%xmm4 - DB 102,65,15,253,237 ; paddw %xmm13,%xmm5 - DB 102,65,15,253,203 ; paddw %xmm11,%xmm1 - DB 102,65,15,253,226 ; paddw %xmm10,%xmm4 - DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,65,15,56,0,242 ; pshufb %xmm10,%xmm6 + DB 102,65,15,56,48,236 ; pmovzxbw %xmm12,%xmm5 + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,15,112,254,78 ; pshufd $0x4e,%xmm6,%xmm7 + DB 102,15,56,48,255 ; pmovzxbw %xmm7,%xmm7 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,68,15,56,48,225 ; pmovzxbw %xmm1,%xmm12 + DB 102,15,56,48,246 ; pmovzxbw %xmm6,%xmm6 + DB 102,65,15,213,246 ; pmullw %xmm14,%xmm6 + DB 102,65,15,253,246 ; paddw %xmm14,%xmm6 + DB 102,65,15,112,216,78 ; pshufd $0x4e,%xmm8,%xmm3 + DB 102,15,56,48,219 ; pmovzxbw %xmm3,%xmm3 + DB 102,15,213,220 ; pmullw %xmm4,%xmm3 + DB 102,15,253,220 ; paddw %xmm4,%xmm3 + DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 + DB 102,65,15,252,201 ; paddb %xmm9,%xmm1 + DB 102,65,15,56,48,232 ; pmovzxbw %xmm8,%xmm5 + DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 + DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,235 ; packuswb %xmm3,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,247 ; packuswb %xmm7,%xmm6 + DB 102,69,15,56,0,242 ; pshufb %xmm10,%xmm14 + DB 102,69,15,56,0,250 ; pshufb %xmm10,%xmm15 + DB 102,68,15,56,48,20,36 ; pmovzxbw (%rsp),%xmm10 + DB 102,68,15,56,48,68,36,16 ; pmovzxbw 0x10(%rsp),%xmm8 + DB 102,65,15,112,255,78 ; pshufd $0x4e,%xmm15,%xmm7 + DB 102,68,15,56,48,231 ; pmovzxbw %xmm7,%xmm12 + DB 102,65,15,56,48,223 ; pmovzxbw %xmm15,%xmm3 + DB 102,65,15,112,230,78 ; pshufd $0x4e,%xmm14,%xmm4 + DB 102,15,56,48,228 ; pmovzxbw %xmm4,%xmm4 + DB 102,65,15,56,48,254 ; pmovzxbw %xmm14,%xmm7 + DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 + DB 102,65,15,213,221 ; pmullw %xmm13,%xmm3 + DB 102,69,15,213,226 ; pmullw %xmm10,%xmm12 + DB 102,69,15,253,226 ; paddw %xmm10,%xmm12 + DB 102,65,15,253,221 ; paddw %xmm13,%xmm3 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,229 ; packuswb %xmm5,%xmm4 - DB 102,15,103,241 ; packuswb %xmm1,%xmm6 - DB 102,65,15,218,230 ; pminub %xmm14,%xmm4 - DB 102,15,218,247 ; pminub %xmm7,%xmm6 - DB 102,15,111,13,124,80,0,0 ; movdqa 0x507c(%rip),%xmm1 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> - DB 102,68,15,248,198 ; psubb %xmm6,%xmm8 - DB 102,15,219,241 ; pand %xmm1,%xmm6 - DB 102,15,219,204 ; pand %xmm4,%xmm1 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,218,253 ; pminub %xmm5,%xmm7 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 + DB 102,65,15,103,220 ; packuswb %xmm12,%xmm3 + DB 102,15,218,222 ; pminub %xmm6,%xmm3 + DB 102,15,111,37,246,77,0,0 ; movdqa 0x4df6(%rip),%xmm4 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> + DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,219,252 ; pand %xmm4,%xmm7 + DB 102,15,219,227 ; pand %xmm3,%xmm4 + DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,248,195 ; psubb %xmm3,%xmm0 DB 102,15,248,196 ; psubb %xmm4,%xmm0 - DB 102,15,248,193 ; psubb %xmm1,%xmm0 - DB 102,68,15,248,198 ; psubb %xmm6,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 + DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 + DB 72,131,196,40 ; add $0x28,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_hardlight_sse41_8bit _sk_hardlight_sse41_8bit LABEL PROC - DB 72,129,236,24,1,0,0 ; sub $0x118,%rsp - DB 102,68,15,111,211 ; movdqa %xmm3,%xmm10 + DB 72,129,236,152,0,0,0 ; sub $0x98,%rsp DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 - DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 - DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 - DB 102,15,104,224 ; punpckhbw %xmm0,%xmm4 - DB 102,65,15,111,253 ; movdqa %xmm13,%xmm7 - DB 102,15,104,248 ; punpckhbw %xmm0,%xmm7 - DB 102,65,15,56,48,206 ; pmovzxbw %xmm14,%xmm1 - DB 102,15,127,12,36 ; movdqa %xmm1,(%rsp) - DB 102,69,15,56,48,221 ; pmovzxbw %xmm13,%xmm11 - DB 102,68,15,127,156,36,176,0,0,0 ; movdqa %xmm11,0xb0(%rsp) - DB 102,15,111,13,42,79,0,0 ; movdqa 0x4f2a(%rip),%xmm1 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,56,0,241 ; pshufb %xmm1,%xmm14 - DB 102,68,15,56,0,233 ; pshufb %xmm1,%xmm13 - DB 102,69,15,56,48,197 ; pmovzxbw %xmm13,%xmm8 - DB 102,68,15,104,232 ; punpckhbw %xmm0,%xmm13 - DB 102,69,15,56,48,254 ; pmovzxbw %xmm14,%xmm15 - DB 102,68,15,104,240 ; punpckhbw %xmm0,%xmm14 - DB 102,15,127,84,36,80 ; movdqa %xmm2,0x50(%rsp) - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,15,104,232 ; punpckhbw %xmm0,%xmm5 - DB 102,68,15,127,84,36,96 ; movdqa %xmm10,0x60(%rsp) - DB 102,65,15,111,210 ; movdqa %xmm10,%xmm2 - DB 102,68,15,104,208 ; punpckhbw %xmm0,%xmm10 - DB 102,15,56,0,241 ; pshufb %xmm1,%xmm6 - DB 102,15,56,0,209 ; pshufb %xmm1,%xmm2 - DB 102,15,56,48,202 ; pmovzxbw %xmm2,%xmm1 - DB 102,15,127,76,36,16 ; movdqa %xmm1,0x10(%rsp) - DB 102,15,104,208 ; punpckhbw %xmm0,%xmm2 - DB 102,15,127,84,36,64 ; movdqa %xmm2,0x40(%rsp) - DB 102,15,56,48,214 ; pmovzxbw %xmm6,%xmm2 - DB 102,15,127,84,36,48 ; movdqa %xmm2,0x30(%rsp) - DB 102,15,104,240 ; punpckhbw %xmm0,%xmm6 - DB 102,68,15,111,206 ; movdqa %xmm6,%xmm9 - DB 102,15,111,199 ; movdqa %xmm7,%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,65,15,56,58,213 ; pminuw %xmm13,%xmm2 - DB 102,15,117,208 ; pcmpeqw %xmm0,%xmm2 - DB 102,15,127,148,36,240,0,0,0 ; movdqa %xmm2,0xf0(%rsp) - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,65,15,56,58,214 ; pminuw %xmm14,%xmm2 - DB 102,15,117,208 ; pcmpeqw %xmm0,%xmm2 - DB 102,15,127,148,36,208,0,0,0 ; movdqa %xmm2,0xd0(%rsp) - DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,56,48,201 ; pmovzxbw %xmm1,%xmm1 + DB 102,15,56,48,248 ; pmovzxbw %xmm0,%xmm7 + DB 102,68,15,111,37,221,76,0,0 ; movdqa 0x4cdd(%rip),%xmm12 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,65,15,56,0,196 ; pshufb %xmm12,%xmm0 + DB 102,15,56,48,240 ; pmovzxbw %xmm0,%xmm6 + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 + DB 102,15,127,148,36,128,0,0,0 ; movdqa %xmm2,0x80(%rsp) + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 + DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,65,15,56,0,196 ; pshufb %xmm12,%xmm0 + DB 102,68,15,56,48,240 ; pmovzxbw %xmm0,%xmm14 + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,56,48,232 ; pmovzxbw %xmm0,%xmm5 + DB 102,15,127,108,36,16 ; movdqa %xmm5,0x10(%rsp) + DB 102,68,15,111,252 ; movdqa %xmm4,%xmm15 + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,15,127,84,36,96 ; movdqa %xmm2,0x60(%rsp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,65,15,111,216 ; movdqa %xmm8,%xmm3 - DB 102,15,127,92,36,112 ; movdqa %xmm3,0x70(%rsp) - DB 102,15,56,58,211 ; pminuw %xmm3,%xmm2 - DB 102,15,117,208 ; pcmpeqw %xmm0,%xmm2 - DB 102,15,127,148,36,0,1,0,0 ; movdqa %xmm2,0x100(%rsp) - DB 102,15,111,4,36 ; movdqa (%rsp),%xmm0 + DB 102,15,101,196 ; pcmpgtw %xmm4,%xmm0 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,15,213,204 ; pmullw %xmm4,%xmm1 + DB 102,68,15,127,76,36,112 ; movdqa %xmm9,0x70(%rsp) + DB 102,65,15,249,233 ; psubw %xmm9,%xmm5 + DB 102,15,249,226 ; psubw %xmm2,%xmm4 + DB 102,15,213,229 ; pmullw %xmm5,%xmm4 + DB 102,68,15,127,52,36 ; movdqa %xmm14,(%rsp) + DB 102,15,253,228 ; paddw %xmm4,%xmm4 + DB 102,15,249,204 ; psubw %xmm4,%xmm1 + DB 102,68,15,213,202 ; pmullw %xmm2,%xmm9 + DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 + DB 102,68,15,56,16,201 ; pblendvb %xmm0,%xmm1,%xmm9 + DB 102,15,127,124,36,80 ; movdqa %xmm7,0x50(%rsp) + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,65,15,56,58,247 ; pminuw %xmm15,%xmm6 - DB 102,15,117,240 ; pcmpeqw %xmm0,%xmm6 - DB 102,15,127,180,36,192,0,0,0 ; movdqa %xmm6,0xc0(%rsp) - DB 102,15,111,53,210,70,0,0 ; movdqa 0x46d2(%rip),%xmm6 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,213,196 ; pmullw %xmm4,%xmm0 - DB 102,15,127,132,36,224,0,0,0 ; movdqa %xmm0,0xe0(%rsp) - DB 102,69,15,111,222 ; movdqa %xmm14,%xmm11 - DB 102,68,15,239,222 ; pxor %xmm6,%xmm11 - DB 102,68,15,213,221 ; pmullw %xmm5,%xmm11 - DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 - DB 102,68,15,249,205 ; psubw %xmm5,%xmm9 - DB 102,68,15,127,140,36,128,0,0,0 ; movdqa %xmm9,0x80(%rsp) - DB 102,68,15,111,205 ; movdqa %xmm5,%xmm9 - DB 102,68,15,213,204 ; pmullw %xmm4,%xmm9 - DB 102,65,15,213,198 ; pmullw %xmm14,%xmm0 - DB 102,15,127,132,36,144,0,0,0 ; movdqa %xmm0,0x90(%rsp) - DB 102,68,15,249,244 ; psubw %xmm4,%xmm14 - DB 102,15,111,108,36,48 ; movdqa 0x30(%rsp),%xmm5 - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,127,68,36,32 ; movdqa %xmm0,0x20(%rsp) - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,239,214 ; pxor %xmm6,%xmm2 - DB 102,15,111,100,36,64 ; movdqa 0x40(%rsp),%xmm4 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,213,193 ; pmullw %xmm1,%xmm0 - DB 102,15,127,132,36,160,0,0,0 ; movdqa %xmm0,0xa0(%rsp) - DB 102,65,15,111,199 ; movdqa %xmm15,%xmm0 - DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 - DB 102,68,15,239,198 ; pxor %xmm6,%xmm8 - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,15,239,254 ; pxor %xmm6,%xmm7 - DB 102,65,15,239,245 ; pxor %xmm13,%xmm6 - DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 - DB 102,68,15,111,252 ; movdqa %xmm4,%xmm15 - DB 102,65,15,249,226 ; psubw %xmm10,%xmm4 - DB 102,15,127,100,36,64 ; movdqa %xmm4,0x40(%rsp) - DB 102,65,15,111,226 ; movdqa %xmm10,%xmm4 - DB 102,15,213,225 ; pmullw %xmm1,%xmm4 - DB 102,69,15,213,253 ; pmullw %xmm13,%xmm15 - DB 102,68,15,249,233 ; psubw %xmm1,%xmm13 - DB 102,15,111,28,36 ; movdqa (%rsp),%xmm3 - DB 102,15,111,76,36,32 ; movdqa 0x20(%rsp),%xmm1 - DB 102,15,213,203 ; pmullw %xmm3,%xmm1 - DB 102,15,127,76,36,32 ; movdqa %xmm1,0x20(%rsp) - DB 102,15,56,48,76,36,80 ; pmovzxbw 0x50(%rsp),%xmm1 - DB 102,68,15,213,193 ; pmullw %xmm1,%xmm8 - DB 102,68,15,111,229 ; movdqa %xmm5,%xmm12 - DB 102,15,249,233 ; psubw %xmm1,%xmm5 - DB 102,15,127,108,36,48 ; movdqa %xmm5,0x30(%rsp) - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,68,15,111,211 ; movdqa %xmm3,%xmm10 - DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 - DB 102,68,15,213,224 ; pmullw %xmm0,%xmm12 - DB 102,65,15,249,194 ; psubw %xmm10,%xmm0 + DB 102,15,101,198 ; pcmpgtw %xmm6,%xmm0 + DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 + DB 102,15,213,206 ; pmullw %xmm6,%xmm1 + DB 102,68,15,127,84,36,64 ; movdqa %xmm10,0x40(%rsp) + DB 102,69,15,249,242 ; psubw %xmm10,%xmm14 + DB 102,15,111,230 ; movdqa %xmm6,%xmm4 + DB 102,15,249,247 ; psubw %xmm7,%xmm6 + DB 102,65,15,213,246 ; pmullw %xmm14,%xmm6 + DB 102,15,253,246 ; paddw %xmm6,%xmm6 + DB 102,15,249,206 ; psubw %xmm6,%xmm1 + DB 102,69,15,111,242 ; movdqa %xmm10,%xmm14 + DB 102,68,15,213,247 ; pmullw %xmm7,%xmm14 + DB 102,69,15,253,246 ; paddw %xmm14,%xmm14 + DB 102,68,15,56,16,241 ; pblendvb %xmm0,%xmm1,%xmm14 + DB 102,65,15,112,197,78 ; pshufd $0x4e,%xmm13,%xmm0 + DB 102,69,15,56,48,221 ; pmovzxbw %xmm13,%xmm11 + DB 102,69,15,56,0,236 ; pshufb %xmm12,%xmm13 + DB 102,15,127,92,36,32 ; movdqa %xmm3,0x20(%rsp) + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,65,15,56,0,236 ; pshufb %xmm12,%xmm5 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,65,15,112,197,78 ; pshufd $0x4e,%xmm13,%xmm0 + DB 102,15,56,48,240 ; pmovzxbw %xmm0,%xmm6 + DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 + DB 102,68,15,56,48,224 ; pmovzxbw %xmm0,%xmm12 + DB 102,68,15,127,100,36,48 ; movdqa %xmm12,0x30(%rsp) + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,65,15,111,218 ; movdqa %xmm10,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,101,222 ; pcmpgtw %xmm6,%xmm3 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,156,36,176,0,0,0 ; movdqa 0xb0(%rsp),%xmm3 - DB 102,15,213,211 ; pmullw %xmm3,%xmm2 - DB 102,15,127,20,36 ; movdqa %xmm2,(%rsp) - DB 102,15,56,48,84,36,96 ; pmovzxbw 0x60(%rsp),%xmm2 - DB 102,15,213,250 ; pmullw %xmm2,%xmm7 + DB 102,15,213,206 ; pmullw %xmm6,%xmm1 + DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 + DB 102,65,15,249,196 ; psubw %xmm12,%xmm0 + DB 102,15,111,214 ; movdqa %xmm6,%xmm2 + DB 102,65,15,249,242 ; psubw %xmm10,%xmm6 + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 + DB 102,15,253,246 ; paddw %xmm6,%xmm6 + DB 102,15,249,206 ; psubw %xmm6,%xmm1 + DB 102,69,15,213,226 ; pmullw %xmm10,%xmm12 + DB 102,69,15,253,228 ; paddw %xmm12,%xmm12 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,68,15,56,16,225 ; pblendvb %xmm0,%xmm1,%xmm12 + DB 102,65,15,56,48,221 ; pmovzxbw %xmm13,%xmm3 + DB 102,68,15,56,48,237 ; pmovzxbw %xmm5,%xmm13 + DB 102,15,111,13,205,67,0,0 ; movdqa 0x43cd(%rip),%xmm1 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 - DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10 - DB 102,15,249,194 ; psubw %xmm2,%xmm0 + DB 102,15,239,193 ; pxor %xmm1,%xmm0 DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) - DB 102,15,213,211 ; pmullw %xmm3,%xmm2 - DB 102,15,111,68,36,112 ; movdqa 0x70(%rsp),%xmm0 - DB 102,68,15,213,208 ; pmullw %xmm0,%xmm10 - DB 102,15,249,195 ; psubw %xmm3,%xmm0 - DB 102,68,15,213,108,36,64 ; pmullw 0x40(%rsp),%xmm13 - DB 102,68,15,213,180,36,128,0,0,0 ; pmullw 0x80(%rsp),%xmm14 - DB 102,15,213,68,36,16 ; pmullw 0x10(%rsp),%xmm0 - DB 102,15,213,76,36,48 ; pmullw 0x30(%rsp),%xmm1 + DB 102,15,111,4,36 ; movdqa (%rsp),%xmm0 + DB 102,15,239,193 ; pxor %xmm1,%xmm0 + DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) + DB 102,68,15,239,193 ; pxor %xmm1,%xmm8 + DB 102,65,15,111,245 ; movdqa %xmm13,%xmm6 + DB 102,15,239,241 ; pxor %xmm1,%xmm6 + DB 102,68,15,239,249 ; pxor %xmm1,%xmm15 + DB 102,15,239,225 ; pxor %xmm1,%xmm4 + DB 102,15,239,209 ; pxor %xmm1,%xmm2 + DB 102,15,239,203 ; pxor %xmm3,%xmm1 + DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,68,15,249,208 ; psubw %xmm0,%xmm10 - DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,68,15,249,225 ; psubw %xmm1,%xmm12 - DB 102,69,15,253,237 ; paddw %xmm13,%xmm13 - DB 102,69,15,249,253 ; psubw %xmm13,%xmm15 - DB 102,69,15,253,246 ; paddw %xmm14,%xmm14 - DB 102,15,111,140,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm1 - DB 102,65,15,249,206 ; psubw %xmm14,%xmm1 - DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 - DB 102,15,111,132,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm0 - DB 102,68,15,219,200 ; pand %xmm0,%xmm9 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,65,15,235,193 ; por %xmm9,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,15,111,132,36,240,0,0,0 ; movdqa 0xf0(%rsp),%xmm0 - DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,65,15,223,199 ; pandn %xmm15,%xmm0 - DB 102,15,235,196 ; por %xmm4,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,253,237 ; paddw %xmm5,%xmm5 - DB 102,15,111,132,36,192,0,0,0 ; movdqa 0xc0(%rsp),%xmm0 - DB 102,15,219,232 ; pand %xmm0,%xmm5 - DB 102,65,15,223,196 ; pandn %xmm12,%xmm0 - DB 102,15,235,197 ; por %xmm5,%xmm0 - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,111,132,36,0,1,0,0 ; movdqa 0x100(%rsp),%xmm0 - DB 102,15,219,208 ; pand %xmm0,%xmm2 - DB 102,65,15,223,194 ; pandn %xmm10,%xmm0 - DB 102,15,235,194 ; por %xmm2,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,68,15,253,156,36,224,0,0,0 ; paddw 0xe0(%rsp),%xmm11 - DB 102,15,253,180,36,160,0,0,0 ; paddw 0xa0(%rsp),%xmm6 - DB 102,68,15,253,68,36,32 ; paddw 0x20(%rsp),%xmm8 - DB 102,15,253,60,36 ; paddw (%rsp),%xmm7 - DB 102,15,111,5,191,76,0,0 ; movdqa 0x4cbf(%rip),%xmm0 # 2dc90 <_sk_overlay_sse2_8bit+0x1738> - DB 102,68,15,253,216 ; paddw %xmm0,%xmm11 - DB 102,68,15,253,219 ; paddw %xmm3,%xmm11 - DB 102,15,253,240 ; paddw %xmm0,%xmm6 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,68,15,253,192 ; paddw %xmm0,%xmm8 - DB 102,68,15,253,196 ; paddw %xmm4,%xmm8 - DB 102,15,253,248 ; paddw %xmm0,%xmm7 - DB 102,15,253,250 ; paddw %xmm2,%xmm7 - DB 102,15,111,5,163,76,0,0 ; movdqa 0x4ca3(%rip),%xmm0 # 2dca0 <_sk_overlay_sse2_8bit+0x1748> - DB 102,15,228,240 ; pmulhuw %xmm0,%xmm6 - DB 102,68,15,228,216 ; pmulhuw %xmm0,%xmm11 - DB 102,15,228,248 ; pmulhuw %xmm0,%xmm7 - DB 102,68,15,228,192 ; pmulhuw %xmm0,%xmm8 - DB 102,65,15,113,211,7 ; psrlw $0x7,%xmm11 - DB 102,65,15,113,208,7 ; psrlw $0x7,%xmm8 - DB 102,69,15,103,195 ; packuswb %xmm11,%xmm8 - DB 102,15,113,214,7 ; psrlw $0x7,%xmm6 - DB 102,15,113,215,7 ; psrlw $0x7,%xmm7 - DB 102,15,103,254 ; packuswb %xmm6,%xmm7 + DB 102,15,101,195 ; pcmpgtw %xmm3,%xmm0 + DB 102,65,15,111,237 ; movdqa %xmm13,%xmm5 + DB 102,15,213,235 ; pmullw %xmm3,%xmm5 + DB 102,15,56,48,124,36,32 ; pmovzxbw 0x20(%rsp),%xmm7 + DB 102,68,15,249,239 ; psubw %xmm7,%xmm13 + DB 102,65,15,249,219 ; psubw %xmm11,%xmm3 + DB 102,65,15,213,221 ; pmullw %xmm13,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,249,235 ; psubw %xmm3,%xmm5 + DB 102,15,111,223 ; movdqa %xmm7,%xmm3 + DB 102,65,15,213,219 ; pmullw %xmm11,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,56,16,221 ; pblendvb %xmm0,%xmm5,%xmm3 + DB 102,15,213,207 ; pmullw %xmm7,%xmm1 + DB 102,15,213,84,36,48 ; pmullw 0x30(%rsp),%xmm2 + DB 102,15,213,100,36,64 ; pmullw 0x40(%rsp),%xmm4 + DB 102,68,15,213,124,36,112 ; pmullw 0x70(%rsp),%xmm15 + DB 102,65,15,213,243 ; pmullw %xmm11,%xmm6 + DB 102,69,15,213,194 ; pmullw %xmm10,%xmm8 + DB 102,15,111,44,36 ; movdqa (%rsp),%xmm5 + DB 102,15,213,108,36,80 ; pmullw 0x50(%rsp),%xmm5 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,213,68,36,96 ; pmullw 0x60(%rsp),%xmm0 + DB 102,15,253,206 ; paddw %xmm6,%xmm1 + DB 102,65,15,253,208 ; paddw %xmm8,%xmm2 + DB 102,15,253,229 ; paddw %xmm5,%xmm4 + DB 102,68,15,253,248 ; paddw %xmm0,%xmm15 + DB 102,15,111,5,72,75,0,0 ; movdqa 0x4b48(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x16e1> + DB 102,68,15,253,248 ; paddw %xmm0,%xmm15 + DB 102,69,15,253,249 ; paddw %xmm9,%xmm15 + DB 102,15,253,224 ; paddw %xmm0,%xmm4 + DB 102,65,15,253,230 ; paddw %xmm14,%xmm4 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,65,15,253,212 ; paddw %xmm12,%xmm2 + DB 102,15,253,200 ; paddw %xmm0,%xmm1 + DB 102,15,253,203 ; paddw %xmm3,%xmm1 + DB 102,15,111,5,44,75,0,0 ; movdqa 0x4b2c(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x16f1> + DB 102,15,228,200 ; pmulhuw %xmm0,%xmm1 + DB 102,15,113,209,7 ; psrlw $0x7,%xmm1 + DB 102,15,228,208 ; pmulhuw %xmm0,%xmm2 + DB 102,15,113,210,7 ; psrlw $0x7,%xmm2 + DB 102,15,103,202 ; packuswb %xmm2,%xmm1 + DB 102,15,228,224 ; pmulhuw %xmm0,%xmm4 + DB 102,68,15,228,248 ; pmulhuw %xmm0,%xmm15 + DB 102,15,113,212,7 ; psrlw $0x7,%xmm4 + DB 102,65,15,113,215,7 ; psrlw $0x7,%xmm15 + DB 102,65,15,103,231 ; packuswb %xmm15,%xmm4 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,84,36,80 ; movaps 0x50(%rsp),%xmm2 - DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 - DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 72,129,196,24,1,0,0 ; add $0x118,%rsp + DB 15,40,148,36,128,0,0,0 ; movaps 0x80(%rsp),%xmm2 + DB 15,40,92,36,32 ; movaps 0x20(%rsp),%xmm3 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 72,129,196,152,0,0,0 ; add $0x98,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_overlay_sse41_8bit _sk_overlay_sse41_8bit LABEL PROC - DB 72,129,236,72,1,0,0 ; sub $0x148,%rsp - DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,68,15,111,248 ; movdqa %xmm0,%xmm15 - DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,65,15,111,231 ; movdqa %xmm15,%xmm4 - DB 102,15,104,224 ; punpckhbw %xmm0,%xmm4 - DB 102,69,15,111,230 ; movdqa %xmm14,%xmm12 - DB 102,68,15,104,224 ; punpckhbw %xmm0,%xmm12 - DB 102,65,15,56,48,207 ; pmovzxbw %xmm15,%xmm1 - DB 102,15,127,12,36 ; movdqa %xmm1,(%rsp) - DB 102,65,15,56,48,206 ; pmovzxbw %xmm14,%xmm1 - DB 102,15,127,140,36,16,1,0,0 ; movdqa %xmm1,0x110(%rsp) - DB 102,15,111,13,10,75,0,0 ; movdqa 0x4b0a(%rip),%xmm1 # 2dba0 <_sk_overlay_sse2_8bit+0x1648> - DB 102,68,15,56,0,249 ; pshufb %xmm1,%xmm15 - DB 102,68,15,56,0,241 ; pshufb %xmm1,%xmm14 - DB 102,65,15,56,48,238 ; pmovzxbw %xmm14,%xmm5 - DB 102,15,127,108,36,32 ; movdqa %xmm5,0x20(%rsp) - DB 102,68,15,104,240 ; punpckhbw %xmm0,%xmm14 - DB 102,65,15,56,48,239 ; pmovzxbw %xmm15,%xmm5 - DB 102,15,127,172,36,0,1,0,0 ; movdqa %xmm5,0x100(%rsp) - DB 102,68,15,104,248 ; punpckhbw %xmm0,%xmm15 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,68,15,111,218 ; movdqa %xmm2,%xmm11 - DB 102,68,15,127,156,36,32,1,0,0 ; movdqa %xmm11,0x120(%rsp) - DB 102,15,104,232 ; punpckhbw %xmm0,%xmm5 - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,127,156,36,48,1,0,0 ; movdqa %xmm3,0x130(%rsp) - DB 102,15,104,208 ; punpckhbw %xmm0,%xmm2 - DB 102,15,56,0,241 ; pshufb %xmm1,%xmm6 - DB 102,15,56,0,249 ; pshufb %xmm1,%xmm7 - DB 102,15,56,48,207 ; pmovzxbw %xmm7,%xmm1 - DB 102,15,104,248 ; punpckhbw %xmm0,%xmm7 - DB 102,68,15,56,48,238 ; pmovzxbw %xmm6,%xmm13 - DB 102,15,104,240 ; punpckhbw %xmm0,%xmm6 - DB 102,68,15,111,206 ; movdqa %xmm6,%xmm9 + DB 72,129,236,168,0,0,0 ; sub $0xa8,%rsp + DB 102,68,15,111,211 ; movdqa %xmm3,%xmm10 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,56,48,244 ; pmovzxbw %xmm4,%xmm6 + DB 102,68,15,56,48,192 ; pmovzxbw %xmm0,%xmm8 + DB 102,68,15,111,45,204,73,0,0 ; movdqa 0x49cc(%rip),%xmm13 # 2dab0 <_sk_overlay_sse2_8bit+0x1601> + DB 102,65,15,56,0,197 ; pshufb %xmm13,%xmm0 + DB 102,68,15,56,48,248 ; pmovzxbw %xmm0,%xmm15 + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 + DB 102,15,127,148,36,144,0,0,0 ; movdqa %xmm2,0x90(%rsp) + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,68,15,56,48,224 ; pmovzxbw %xmm0,%xmm12 + DB 102,68,15,56,48,202 ; pmovzxbw %xmm2,%xmm9 DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,65,15,56,0,197 ; pshufb %xmm13,%xmm0 + DB 102,68,15,56,48,216 ; pmovzxbw %xmm0,%xmm11 + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,56,48,232 ; pmovzxbw %xmm0,%xmm5 + DB 102,15,127,108,36,16 ; movdqa %xmm5,0x10(%rsp) + DB 102,68,15,111,244 ; movdqa %xmm4,%xmm14 + DB 102,65,15,111,196 ; movdqa %xmm12,%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,15,56,58,247 ; pminuw %xmm7,%xmm6 - DB 102,15,117,240 ; pcmpeqw %xmm0,%xmm6 - DB 102,15,127,180,36,208,0,0,0 ; movdqa %xmm6,0xd0(%rsp) - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,65,15,56,58,241 ; pminuw %xmm9,%xmm6 - DB 102,15,117,240 ; pcmpeqw %xmm0,%xmm6 - DB 102,15,127,180,36,192,0,0,0 ; movdqa %xmm6,0xc0(%rsp) - DB 102,15,56,48,195 ; pmovzxbw %xmm3,%xmm0 - DB 102,15,127,132,36,144,0,0,0 ; movdqa %xmm0,0x90(%rsp) - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,56,58,217 ; pminuw %xmm1,%xmm3 - DB 102,15,127,76,36,16 ; movdqa %xmm1,0x10(%rsp) - DB 102,15,117,216 ; pcmpeqw %xmm0,%xmm3 - DB 102,15,127,156,36,224,0,0,0 ; movdqa %xmm3,0xe0(%rsp) - DB 102,65,15,56,48,195 ; pmovzxbw %xmm11,%xmm0 - DB 102,15,127,68,36,96 ; movdqa %xmm0,0x60(%rsp) - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,65,15,56,58,221 ; pminuw %xmm13,%xmm3 - DB 102,15,117,216 ; pcmpeqw %xmm0,%xmm3 - DB 102,15,127,156,36,176,0,0,0 ; movdqa %xmm3,0xb0(%rsp) - DB 102,15,111,53,157,66,0,0 ; movdqa 0x429d(%rip),%xmm6 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,213,196 ; pmullw %xmm4,%xmm0 - DB 102,15,127,132,36,240,0,0,0 ; movdqa %xmm0,0xf0(%rsp) - DB 102,69,15,111,223 ; movdqa %xmm15,%xmm11 - DB 102,68,15,239,222 ; pxor %xmm6,%xmm11 - DB 102,68,15,213,221 ; pmullw %xmm5,%xmm11 + DB 102,15,101,197 ; pcmpgtw %xmm5,%xmm0 + DB 102,15,111,253 ; movdqa %xmm5,%xmm7 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,65,15,249,236 ; psubw %xmm12,%xmm5 + DB 102,68,15,127,164,36,128,0,0,0 ; movdqa %xmm12,0x80(%rsp) + DB 102,15,127,116,36,112 ; movdqa %xmm6,0x70(%rsp) + DB 102,15,249,230 ; psubw %xmm6,%xmm4 + DB 102,15,213,229 ; pmullw %xmm5,%xmm4 + DB 102,68,15,127,28,36 ; movdqa %xmm11,(%rsp) + DB 102,15,253,228 ; paddw %xmm4,%xmm4 + DB 102,15,249,252 ; psubw %xmm4,%xmm7 + DB 102,68,15,213,230 ; pmullw %xmm6,%xmm12 + DB 102,69,15,253,228 ; paddw %xmm12,%xmm12 + DB 102,68,15,56,16,231 ; pblendvb %xmm0,%xmm7,%xmm12 + DB 102,68,15,127,76,36,64 ; movdqa %xmm9,0x40(%rsp) DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 - DB 102,68,15,249,205 ; psubw %xmm5,%xmm9 - DB 102,68,15,127,76,36,80 ; movdqa %xmm9,0x50(%rsp) - DB 102,68,15,111,205 ; movdqa %xmm5,%xmm9 - DB 102,68,15,213,204 ; pmullw %xmm4,%xmm9 - DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 - DB 102,15,127,68,36,112 ; movdqa %xmm0,0x70(%rsp) - DB 102,68,15,249,252 ; psubw %xmm4,%xmm15 - DB 102,65,15,111,197 ; movdqa %xmm13,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,239,206 ; pxor %xmm6,%xmm1 - DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,65,15,111,252 ; movdqa %xmm12,%xmm7 - DB 102,15,213,199 ; pmullw %xmm7,%xmm0 - DB 102,15,127,132,36,160,0,0,0 ; movdqa %xmm0,0xa0(%rsp) - DB 102,15,111,132,36,0,1,0,0 ; movdqa 0x100(%rsp),%xmm0 - DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 - DB 102,68,15,239,198 ; pxor %xmm6,%xmm8 - DB 102,15,111,108,36,32 ; movdqa 0x20(%rsp),%xmm5 - DB 102,15,239,238 ; pxor %xmm6,%xmm5 - DB 102,65,15,239,246 ; pxor %xmm14,%xmm6 - DB 102,15,213,242 ; pmullw %xmm2,%xmm6 - DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 - DB 102,15,249,202 ; psubw %xmm2,%xmm1 - DB 102,15,127,76,36,64 ; movdqa %xmm1,0x40(%rsp) - DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,15,213,231 ; pmullw %xmm7,%xmm4 - DB 102,69,15,213,230 ; pmullw %xmm14,%xmm12 - DB 102,68,15,249,247 ; psubw %xmm7,%xmm14 - DB 102,15,111,12,36 ; movdqa (%rsp),%xmm1 - DB 102,15,213,217 ; pmullw %xmm1,%xmm3 - DB 102,15,127,156,36,128,0,0,0 ; movdqa %xmm3,0x80(%rsp) - DB 102,15,111,84,36,96 ; movdqa 0x60(%rsp),%xmm2 - DB 102,68,15,213,194 ; pmullw %xmm2,%xmm8 - DB 102,65,15,111,221 ; movdqa %xmm13,%xmm3 - DB 102,15,249,218 ; psubw %xmm2,%xmm3 - DB 102,15,127,92,36,48 ; movdqa %xmm3,0x30(%rsp) - DB 102,15,111,218 ; movdqa %xmm2,%xmm3 - DB 102,15,213,217 ; pmullw %xmm1,%xmm3 - DB 102,68,15,213,232 ; pmullw %xmm0,%xmm13 - DB 102,15,249,193 ; psubw %xmm1,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,148,36,16,1,0,0 ; movdqa 0x110(%rsp),%xmm2 - DB 102,68,15,213,210 ; pmullw %xmm2,%xmm10 - DB 102,68,15,127,20,36 ; movdqa %xmm10,(%rsp) - DB 102,15,111,132,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm0 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,15,111,124,36,16 ; movdqa 0x10(%rsp),%xmm7 - DB 102,68,15,111,215 ; movdqa %xmm7,%xmm10 - DB 102,15,249,248 ; psubw %xmm0,%xmm7 - DB 102,15,127,124,36,16 ; movdqa %xmm7,0x10(%rsp) - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,15,213,250 ; pmullw %xmm2,%xmm7 - DB 102,15,111,68,36,32 ; movdqa 0x20(%rsp),%xmm0 - DB 102,68,15,213,208 ; pmullw %xmm0,%xmm10 - DB 102,15,249,194 ; psubw %xmm2,%xmm0 - DB 102,68,15,213,116,36,64 ; pmullw 0x40(%rsp),%xmm14 - DB 102,68,15,213,124,36,80 ; pmullw 0x50(%rsp),%xmm15 - DB 102,15,213,68,36,16 ; pmullw 0x10(%rsp),%xmm0 - DB 102,15,213,76,36,48 ; pmullw 0x30(%rsp),%xmm1 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,68,15,249,208 ; psubw %xmm0,%xmm10 - DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,68,15,249,233 ; psubw %xmm1,%xmm13 - DB 102,69,15,253,246 ; paddw %xmm14,%xmm14 - DB 102,69,15,249,230 ; psubw %xmm14,%xmm12 + DB 102,65,15,101,195 ; pcmpgtw %xmm11,%xmm0 + DB 102,65,15,111,251 ; movdqa %xmm11,%xmm7 + DB 102,65,15,213,255 ; pmullw %xmm15,%xmm7 + DB 102,69,15,249,217 ; psubw %xmm9,%xmm11 + DB 102,65,15,111,247 ; movdqa %xmm15,%xmm6 + DB 102,68,15,127,68,36,80 ; movdqa %xmm8,0x50(%rsp) + DB 102,69,15,249,248 ; psubw %xmm8,%xmm15 + DB 102,69,15,213,251 ; pmullw %xmm11,%xmm15 DB 102,69,15,253,255 ; paddw %xmm15,%xmm15 - DB 102,15,111,76,36,112 ; movdqa 0x70(%rsp),%xmm1 - DB 102,65,15,249,207 ; psubw %xmm15,%xmm1 - DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 - DB 102,15,111,132,36,192,0,0,0 ; movdqa 0xc0(%rsp),%xmm0 - DB 102,68,15,219,200 ; pand %xmm0,%xmm9 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,65,15,235,193 ; por %xmm9,%xmm0 - DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,15,111,132,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm0 - DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,65,15,223,196 ; pandn %xmm12,%xmm0 - DB 102,15,235,196 ; por %xmm4,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,65,15,249,255 ; psubw %xmm15,%xmm7 + DB 102,69,15,111,249 ; movdqa %xmm9,%xmm15 + DB 102,69,15,213,248 ; pmullw %xmm8,%xmm15 + DB 102,69,15,253,255 ; paddw %xmm15,%xmm15 + DB 102,68,15,56,16,255 ; pblendvb %xmm0,%xmm7,%xmm15 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,56,48,209 ; pmovzxbw %xmm1,%xmm2 + DB 102,15,127,84,36,48 ; movdqa %xmm2,0x30(%rsp) + DB 102,65,15,56,0,205 ; pshufb %xmm13,%xmm1 + DB 102,65,15,111,234 ; movdqa %xmm10,%xmm5 + DB 102,65,15,56,0,237 ; pshufb %xmm13,%xmm5 + DB 102,15,56,48,224 ; pmovzxbw %xmm0,%xmm4 + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 + DB 102,15,56,48,248 ; pmovzxbw %xmm0,%xmm7 + DB 102,65,15,112,194,78 ; pshufd $0x4e,%xmm10,%xmm0 + DB 102,68,15,127,84,36,96 ; movdqa %xmm10,0x60(%rsp) + DB 102,68,15,56,48,216 ; pmovzxbw %xmm0,%xmm11 + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,56,48,192 ; pmovzxbw %xmm0,%xmm0 + DB 102,65,15,111,219 ; movdqa %xmm11,%xmm3 DB 102,15,253,219 ; paddw %xmm3,%xmm3 - DB 102,15,111,132,36,176,0,0,0 ; movdqa 0xb0(%rsp),%xmm0 - DB 102,15,219,216 ; pand %xmm0,%xmm3 - DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 - DB 102,15,235,195 ; por %xmm3,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,101,216 ; pcmpgtw %xmm0,%xmm3 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,15,213,207 ; pmullw %xmm7,%xmm1 + DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 + DB 102,65,15,249,195 ; psubw %xmm11,%xmm0 + DB 102,15,111,215 ; movdqa %xmm7,%xmm2 + DB 102,15,249,252 ; psubw %xmm4,%xmm7 + DB 102,15,127,100,36,32 ; movdqa %xmm4,0x20(%rsp) + DB 102,15,213,248 ; pmullw %xmm0,%xmm7 DB 102,15,253,255 ; paddw %xmm7,%xmm7 - DB 102,15,111,132,36,224,0,0,0 ; movdqa 0xe0(%rsp),%xmm0 - DB 102,15,219,248 ; pand %xmm0,%xmm7 - DB 102,65,15,223,194 ; pandn %xmm10,%xmm0 - DB 102,15,235,199 ; por %xmm7,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,68,15,253,156,36,240,0,0,0 ; paddw 0xf0(%rsp),%xmm11 - DB 102,15,253,180,36,160,0,0,0 ; paddw 0xa0(%rsp),%xmm6 - DB 102,68,15,253,132,36,128,0,0,0 ; paddw 0x80(%rsp),%xmm8 - DB 102,15,253,44,36 ; paddw (%rsp),%xmm5 - DB 102,15,111,5,153,72,0,0 ; movdqa 0x4899(%rip),%xmm0 # 2dc90 <_sk_overlay_sse2_8bit+0x1738> - DB 102,68,15,253,216 ; paddw %xmm0,%xmm11 - DB 102,69,15,253,217 ; paddw %xmm9,%xmm11 + DB 102,15,249,207 ; psubw %xmm7,%xmm1 + DB 102,69,15,111,203 ; movdqa %xmm11,%xmm9 + DB 102,68,15,213,204 ; pmullw %xmm4,%xmm9 + DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,68,15,56,16,201 ; pblendvb %xmm0,%xmm1,%xmm9 + DB 102,65,15,56,48,221 ; pmovzxbw %xmm13,%xmm3 + DB 102,15,56,48,237 ; pmovzxbw %xmm5,%xmm5 + DB 102,15,111,13,167,64,0,0 ; movdqa 0x40a7(%rip),%xmm1 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,239,193 ; pxor %xmm1,%xmm0 + DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) + DB 102,15,111,4,36 ; movdqa (%rsp),%xmm0 + DB 102,15,239,193 ; pxor %xmm1,%xmm0 + DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) + DB 102,68,15,239,193 ; pxor %xmm1,%xmm8 + DB 102,68,15,111,237 ; movdqa %xmm5,%xmm13 + DB 102,68,15,239,233 ; pxor %xmm1,%xmm13 + DB 102,68,15,239,241 ; pxor %xmm1,%xmm14 + DB 102,15,239,241 ; pxor %xmm1,%xmm6 + DB 102,15,239,209 ; pxor %xmm1,%xmm2 + DB 102,15,239,203 ; pxor %xmm3,%xmm1 + DB 102,65,15,56,48,250 ; pmovzxbw %xmm10,%xmm7 + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,101,197 ; pcmpgtw %xmm5,%xmm0 + DB 102,15,111,229 ; movdqa %xmm5,%xmm4 + DB 102,15,213,227 ; pmullw %xmm3,%xmm4 + DB 102,15,249,239 ; psubw %xmm7,%xmm5 + DB 102,68,15,111,84,36,48 ; movdqa 0x30(%rsp),%xmm10 + DB 102,65,15,249,218 ; psubw %xmm10,%xmm3 + DB 102,15,213,221 ; pmullw %xmm5,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,249,227 ; psubw %xmm3,%xmm4 + DB 102,15,111,223 ; movdqa %xmm7,%xmm3 + DB 102,65,15,213,218 ; pmullw %xmm10,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,56,16,220 ; pblendvb %xmm0,%xmm4,%xmm3 + DB 102,15,213,207 ; pmullw %xmm7,%xmm1 + DB 102,65,15,213,211 ; pmullw %xmm11,%xmm2 + DB 102,15,213,116,36,64 ; pmullw 0x40(%rsp),%xmm6 + DB 102,68,15,213,180,36,128,0,0,0 ; pmullw 0x80(%rsp),%xmm14 + DB 102,69,15,213,234 ; pmullw %xmm10,%xmm13 + DB 102,68,15,213,68,36,32 ; pmullw 0x20(%rsp),%xmm8 + DB 102,15,111,36,36 ; movdqa (%rsp),%xmm4 + DB 102,15,213,100,36,80 ; pmullw 0x50(%rsp),%xmm4 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,213,68,36,112 ; pmullw 0x70(%rsp),%xmm0 + DB 102,65,15,253,205 ; paddw %xmm13,%xmm1 + DB 102,65,15,253,208 ; paddw %xmm8,%xmm2 + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,68,15,253,240 ; paddw %xmm0,%xmm14 + DB 102,15,111,5,26,72,0,0 ; movdqa 0x481a(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x16e1> + DB 102,68,15,253,240 ; paddw %xmm0,%xmm14 + DB 102,69,15,253,244 ; paddw %xmm12,%xmm14 DB 102,15,253,240 ; paddw %xmm0,%xmm6 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,68,15,253,192 ; paddw %xmm0,%xmm8 - DB 102,68,15,253,195 ; paddw %xmm3,%xmm8 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,15,253,234 ; paddw %xmm2,%xmm5 - DB 102,15,111,5,125,72,0,0 ; movdqa 0x487d(%rip),%xmm0 # 2dca0 <_sk_overlay_sse2_8bit+0x1748> + DB 102,65,15,253,247 ; paddw %xmm15,%xmm6 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,65,15,253,209 ; paddw %xmm9,%xmm2 + DB 102,15,253,200 ; paddw %xmm0,%xmm1 + DB 102,15,253,203 ; paddw %xmm3,%xmm1 + DB 102,15,111,5,254,71,0,0 ; movdqa 0x47fe(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x16f1> + DB 102,15,228,200 ; pmulhuw %xmm0,%xmm1 + DB 102,15,113,209,7 ; psrlw $0x7,%xmm1 + DB 102,15,228,208 ; pmulhuw %xmm0,%xmm2 + DB 102,15,113,210,7 ; psrlw $0x7,%xmm2 + DB 102,15,103,202 ; packuswb %xmm2,%xmm1 DB 102,15,228,240 ; pmulhuw %xmm0,%xmm6 - DB 102,68,15,228,216 ; pmulhuw %xmm0,%xmm11 - DB 102,15,228,232 ; pmulhuw %xmm0,%xmm5 - DB 102,68,15,228,192 ; pmulhuw %xmm0,%xmm8 - DB 102,65,15,113,211,7 ; psrlw $0x7,%xmm11 - DB 102,65,15,113,208,7 ; psrlw $0x7,%xmm8 - DB 102,69,15,103,195 ; packuswb %xmm11,%xmm8 + DB 102,68,15,228,240 ; pmulhuw %xmm0,%xmm14 DB 102,15,113,214,7 ; psrlw $0x7,%xmm6 - DB 102,15,113,213,7 ; psrlw $0x7,%xmm5 - DB 102,15,103,238 ; packuswb %xmm6,%xmm5 + DB 102,65,15,113,214,7 ; psrlw $0x7,%xmm14 + DB 102,65,15,103,246 ; packuswb %xmm14,%xmm6 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,148,36,32,1,0,0 ; movaps 0x120(%rsp),%xmm2 - DB 15,40,156,36,48,1,0,0 ; movaps 0x130(%rsp),%xmm3 - DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 72,129,196,72,1,0,0 ; add $0x148,%rsp + DB 15,40,148,36,144,0,0,0 ; movaps 0x90(%rsp),%xmm2 + DB 15,40,92,36,96 ; movaps 0x60(%rsp),%xmm3 + DB 102,15,111,198 ; movdqa %xmm6,%xmm0 + DB 72,129,196,168,0,0,0 ; add $0xa8,%rsp DB 255,224 ; jmpq *%rax + DB 144 ; nop + DB 144 ; nop + DB 144 ; nop PUBLIC _sk_start_pipeline_sse2_8bit _sk_start_pipeline_sse2_8bit LABEL PROC @@ -39439,13 +39369,13 @@ _sk_start_pipeline_sse2_8bit LABEL PROC DB 73,137,246 ; mov %rsi,%r14 DB 76,137,141,8,255,255,255 ; mov %r9,-0xf8(%rbp) DB 76,57,203 ; cmp %r9,%rbx - DB 115,101 ; jae 2955a <_sk_start_pipeline_sse2_8bit+0xe2> + DB 115,101 ; jae 294d2 <_sk_start_pipeline_sse2_8bit+0xe2> DB 72,139,133,24,255,255,255 ; mov -0xe8(%rbp),%rax DB 72,141,64,8 ; lea 0x8(%rax),%rax DB 72,137,133,16,255,255,255 ; mov %rax,-0xf0(%rbp) DB 76,57,189,16,255,255,255 ; cmp %r15,-0xf0(%rbp) DB 72,139,149,24,255,255,255 ; mov -0xe8(%rbp),%rdx - DB 119,38 ; ja 2953d <_sk_start_pipeline_sse2_8bit+0xc5> + DB 119,38 ; ja 294b5 <_sk_start_pipeline_sse2_8bit+0xc5> DB 76,139,165,24,255,255,255 ; mov -0xe8(%rbp),%r12 DB 49,255 ; xor %edi,%edi DB 76,137,246 ; mov %r14,%rsi @@ -39456,16 +39386,16 @@ _sk_start_pipeline_sse2_8bit LABEL PROC DB 73,131,196,16 ; add $0x10,%r12 DB 77,57,252 ; cmp %r15,%r12 DB 73,137,212 ; mov %rdx,%r12 - DB 118,225 ; jbe 2951e <_sk_start_pipeline_sse2_8bit+0xa6> + DB 118,225 ; jbe 29496 <_sk_start_pipeline_sse2_8bit+0xa6> DB 76,137,255 ; mov %r15,%rdi DB 72,41,215 ; sub %rdx,%rdi - DB 116,9 ; je 2954e <_sk_start_pipeline_sse2_8bit+0xd6> + DB 116,9 ; je 294c6 <_sk_start_pipeline_sse2_8bit+0xd6> DB 76,137,246 ; mov %r14,%rsi DB 72,137,217 ; mov %rbx,%rcx DB 65,255,213 ; callq *%r13 DB 72,255,195 ; inc %rbx DB 72,59,157,8,255,255,255 ; cmp -0xf8(%rbp),%rbx - DB 117,173 ; jne 29507 <_sk_start_pipeline_sse2_8bit+0x8f> + DB 117,173 ; jne 2947f <_sk_start_pipeline_sse2_8bit+0x8f> DB 15,40,181,32,255,255,255 ; movaps -0xe0(%rbp),%xmm6 DB 15,40,189,48,255,255,255 ; movaps -0xd0(%rbp),%xmm7 DB 68,15,40,133,64,255,255,255 ; movaps -0xc0(%rbp),%xmm8 @@ -39503,7 +39433,7 @@ _sk_uniform_color_sse2_8bit LABEL PROC PUBLIC _sk_set_rgb_sse2_8bit _sk_set_rgb_sse2_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,15,16,37,9,61,0,0 ; movss 0x3d09(%rip),%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 243,15,16,37,113,60,0,0 ; movss 0x3c71(%rip),%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 243,15,16,40 ; movss (%rax),%xmm5 DB 243,15,89,236 ; mulss %xmm4,%xmm5 DB 243,76,15,44,197 ; cvttss2si %xmm5,%r8 @@ -39518,7 +39448,7 @@ _sk_set_rgb_sse2_8bit LABEL PROC DB 68,9,200 ; or %r9d,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 102,15,111,45,126,69,0,0 ; movdqa 0x457e(%rip),%xmm5 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,45,22,69,0,0 ; movdqa 0x4516(%rip),%xmm5 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,219,205 ; pand %xmm5,%xmm1 DB 102,15,219,197 ; pand %xmm5,%xmm0 DB 102,15,235,196 ; por %xmm4,%xmm0 @@ -39541,38 +39471,38 @@ _sk_premul_sse2_8bit LABEL PROC DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 DB 243,15,112,248,95 ; pshufhw $0x5f,%xmm0,%xmm7 - DB 102,15,111,5,34,69,0,0 ; movdqa 0x4522(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x1638> - DB 102,15,235,248 ; por %xmm0,%xmm7 - DB 102,15,235,240 ; por %xmm0,%xmm6 - DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 + DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 - DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 - DB 102,69,15,104,193 ; punpckhbw %xmm9,%xmm8 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,65,15,96,233 ; punpcklbw %xmm9,%xmm5 - DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 - DB 102,15,111,230 ; movdqa %xmm6,%xmm4 - DB 102,65,15,96,225 ; punpcklbw %xmm9,%xmm4 - DB 102,65,15,104,241 ; punpckhbw %xmm9,%xmm6 - DB 102,68,15,111,215 ; movdqa %xmm7,%xmm10 - DB 102,69,15,96,209 ; punpcklbw %xmm9,%xmm10 - DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 + DB 102,65,15,96,194 ; punpcklbw %xmm10,%xmm0 + DB 102,69,15,104,194 ; punpckhbw %xmm10,%xmm8 + DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 + DB 102,69,15,96,218 ; punpcklbw %xmm10,%xmm11 + DB 102,65,15,104,202 ; punpckhbw %xmm10,%xmm1 + DB 102,68,15,111,13,150,68,0,0 ; movdqa 0x4496(%rip),%xmm9 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> + DB 102,65,15,235,249 ; por %xmm9,%xmm7 + DB 102,65,15,235,241 ; por %xmm9,%xmm6 + DB 102,15,111,238 ; movdqa %xmm6,%xmm5 + DB 102,65,15,96,234 ; punpcklbw %xmm10,%xmm5 + DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 + DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,65,15,96,226 ; punpcklbw %xmm10,%xmm4 + DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 DB 102,15,213,249 ; pmullw %xmm1,%xmm7 - DB 102,68,15,213,213 ; pmullw %xmm5,%xmm10 + DB 102,65,15,213,227 ; pmullw %xmm11,%xmm4 DB 102,65,15,213,240 ; pmullw %xmm8,%xmm6 - DB 102,15,213,224 ; pmullw %xmm0,%xmm4 - DB 102,15,253,196 ; paddw %xmm4,%xmm0 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,253,197 ; paddw %xmm5,%xmm0 DB 102,65,15,253,240 ; paddw %xmm8,%xmm6 - DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 + DB 102,68,15,253,220 ; paddw %xmm4,%xmm11 DB 102,15,253,249 ; paddw %xmm1,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 + DB 102,68,15,103,223 ; packuswb %xmm7,%xmm11 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,198 ; packuswb %xmm6,%xmm0 - DB 102,15,103,239 ; packuswb %xmm7,%xmm5 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,65,15,111,203 ; movdqa %xmm11,%xmm1 DB 255,224 ; jmpq *%rax PUBLIC _sk_swap_rb_sse2_8bit @@ -39615,9 +39545,9 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,16 ; jne 29789 <_sk_load_8888_sse2_8bit+0x2c> - DB 67,15,16,76,152,16 ; movups 0x10(%r8,%r11,4),%xmm1 + DB 117,16 ; jne 29707 <_sk_load_8888_sse2_8bit+0x2c> DB 102,67,15,16,4,152 ; movupd (%r8,%r11,4),%xmm0 + DB 67,15,16,76,152,16 ; movups 0x10(%r8,%r11,4),%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -39626,18 +39556,18 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,228 ; ja 29785 <_sk_load_8888_sse2_8bit+0x28> + DB 119,228 ; ja 29703 <_sk_load_8888_sse2_8bit+0x28> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,80,0,0,0 ; lea 0x50(%rip),%r10 # 297fc <_sk_load_8888_sse2_8bit+0x9f> + DB 76,141,21,86,0,0,0 ; lea 0x56(%rip),%r10 # 29780 <_sk_load_8888_sse2_8bit+0xa5> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0 - DB 235,200 ; jmp 29785 <_sk_load_8888_sse2_8bit+0x28> + DB 235,200 ; jmp 29703 <_sk_load_8888_sse2_8bit+0x28> DB 102,67,15,110,68,152,8 ; movd 0x8(%r8,%r11,4),%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 DB 102,67,15,18,4,152 ; movlpd (%r8,%r11,4),%xmm0 - DB 235,180 ; jmp 29785 <_sk_load_8888_sse2_8bit+0x28> + DB 235,180 ; jmp 29703 <_sk_load_8888_sse2_8bit+0x28> DB 102,67,15,110,68,152,24 ; movd 0x18(%r8,%r11,4),%xmm0 DB 102,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm1 DB 243,67,15,16,68,152,20 ; movss 0x14(%r8,%r11,4),%xmm0 @@ -39646,20 +39576,25 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 15,40,200 ; movaps %xmm0,%xmm1 DB 243,67,15,16,68,152,16 ; movss 0x10(%r8,%r11,4),%xmm0 DB 243,15,16,200 ; movss %xmm0,%xmm1 - DB 235,131 ; jmp 2977f <_sk_load_8888_sse2_8bit+0x22> - DB 185,255,255,255,205 ; mov $0xcdffffff,%ecx + DB 102,67,15,16,4,152 ; movupd (%r8,%r11,4),%xmm0 + DB 235,131 ; jmp 29703 <_sk_load_8888_sse2_8bit+0x28> + DB 179,255 ; mov $0xff,%bl + DB 255 ; (bad) + DB 255,199 ; inc %edi DB 255 ; (bad) DB 255 ; (bad) - DB 255,193 ; inc %ecx DB 255 ; (bad) + DB 187,255,255,255,248 ; mov $0xf8ffffff,%ebx DB 255 ; (bad) - DB 255,131,255,255,255,243 ; incl -0xc000001(%rbx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,225 ; jmpq *%rcx + DB 237 ; in (%dx),%eax DB 255 ; (bad) DB 255 ; (bad) - DB 255,213 ; callq *%rbp + DB 255 ; (bad) + DB 219,255 ; (bad) + DB 255 ; (bad) + DB 255,207 ; dec %edi DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -39674,9 +39609,9 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,16 ; jne 29844 <_sk_load_8888_dst_sse2_8bit+0x2c> - DB 67,15,16,92,152,16 ; movups 0x10(%r8,%r11,4),%xmm3 + DB 117,16 ; jne 297c8 <_sk_load_8888_dst_sse2_8bit+0x2c> DB 102,67,15,16,20,152 ; movupd (%r8,%r11,4),%xmm2 + DB 67,15,16,92,152,16 ; movups 0x10(%r8,%r11,4),%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d @@ -39685,18 +39620,18 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,228 ; ja 29840 <_sk_load_8888_dst_sse2_8bit+0x28> + DB 119,228 ; ja 297c4 <_sk_load_8888_dst_sse2_8bit+0x28> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,81,0,0,0 ; lea 0x51(%rip),%r10 # 298b8 <_sk_load_8888_dst_sse2_8bit+0xa0> + DB 76,141,21,89,0,0,0 ; lea 0x59(%rip),%r10 # 29844 <_sk_load_8888_dst_sse2_8bit+0xa8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2 - DB 235,200 ; jmp 29840 <_sk_load_8888_dst_sse2_8bit+0x28> + DB 235,200 ; jmp 297c4 <_sk_load_8888_dst_sse2_8bit+0x28> DB 102,67,15,110,84,152,8 ; movd 0x8(%r8,%r11,4),%xmm2 DB 102,15,112,210,69 ; pshufd $0x45,%xmm2,%xmm2 DB 102,67,15,18,20,152 ; movlpd (%r8,%r11,4),%xmm2 - DB 235,180 ; jmp 29840 <_sk_load_8888_dst_sse2_8bit+0x28> + DB 235,180 ; jmp 297c4 <_sk_load_8888_dst_sse2_8bit+0x28> DB 102,67,15,110,84,152,24 ; movd 0x18(%r8,%r11,4),%xmm2 DB 102,15,112,218,69 ; pshufd $0x45,%xmm2,%xmm3 DB 243,67,15,16,84,152,20 ; movss 0x14(%r8,%r11,4),%xmm2 @@ -39705,21 +39640,26 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 243,67,15,16,84,152,16 ; movss 0x10(%r8,%r11,4),%xmm2 DB 243,15,16,218 ; movss %xmm2,%xmm3 - DB 235,131 ; jmp 2983a <_sk_load_8888_dst_sse2_8bit+0x22> - DB 144 ; nop - DB 184,255,255,255,204 ; mov $0xccffffff,%eax + DB 102,67,15,16,20,152 ; movupd (%r8,%r11,4),%xmm2 + DB 235,131 ; jmp 297c4 <_sk_load_8888_dst_sse2_8bit+0x28> + DB 15,31,0 ; nopl (%rax) + DB 176,255 ; mov $0xff,%al + DB 255 ; (bad) + DB 255,196 ; inc %esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax DB 255 ; (bad) + DB 184,255,255,255,245 ; mov $0xf5ffffff,%eax DB 255 ; (bad) - DB 255,130,255,255,255,242 ; incl -0xd000001(%rdx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,224 ; jmpq *%rax + DB 234 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 255,212 ; callq *%rsp + DB 255 ; (bad) + DB 216,255 ; fdivr %st(7),%st + DB 255 ; (bad) + DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -39734,7 +39674,7 @@ _sk_store_8888_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 29901 <_sk_store_8888_sse2_8bit+0x2d> + DB 117,17 ; jne 2988d <_sk_store_8888_sse2_8bit+0x2d> DB 243,67,15,127,4,152 ; movdqu %xmm0,(%r8,%r11,4) DB 243,67,15,127,76,152,16 ; movdqu %xmm1,0x10(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -39743,25 +39683,25 @@ _sk_store_8888_sse2_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 298fd <_sk_store_8888_sse2_8bit+0x29> + DB 119,236 ; ja 29889 <_sk_store_8888_sse2_8bit+0x29> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,76,0,0,0 ; lea 0x4c(%rip),%r10 # 29968 <_sk_store_8888_sse2_8bit+0x94> + DB 76,141,21,76,0,0,0 ; lea 0x4c(%rip),%r10 # 298f4 <_sk_store_8888_sse2_8bit+0x94> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,126,4,152 ; movd %xmm0,(%r8,%r11,4) - DB 235,208 ; jmp 298fd <_sk_store_8888_sse2_8bit+0x29> + DB 235,208 ; jmp 29889 <_sk_store_8888_sse2_8bit+0x29> DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 DB 102,67,15,126,100,152,8 ; movd %xmm4,0x8(%r8,%r11,4) DB 102,67,15,214,4,152 ; movq %xmm0,(%r8,%r11,4) - DB 235,188 ; jmp 298fd <_sk_store_8888_sse2_8bit+0x29> + DB 235,188 ; jmp 29889 <_sk_store_8888_sse2_8bit+0x29> DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 DB 102,67,15,126,100,152,24 ; movd %xmm4,0x18(%r8,%r11,4) DB 102,15,112,225,229 ; pshufd $0xe5,%xmm1,%xmm4 DB 102,67,15,126,100,152,20 ; movd %xmm4,0x14(%r8,%r11,4) DB 102,67,15,126,76,152,16 ; movd %xmm1,0x10(%r8,%r11,4) DB 243,67,15,127,4,152 ; movdqu %xmm0,(%r8,%r11,4) - DB 235,149 ; jmp 298fd <_sk_store_8888_sse2_8bit+0x29> + DB 235,149 ; jmp 29889 <_sk_store_8888_sse2_8bit+0x29> DB 189,255,255,255,209 ; mov $0xd1ffffff,%ebp DB 255 ; (bad) DB 255 ; (bad) @@ -39793,9 +39733,9 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,92 ; jne 299fc <_sk_load_bgra_sse2_8bit+0x78> - DB 67,15,16,76,152,16 ; movups 0x10(%r8,%r11,4),%xmm1 + DB 117,92 ; jne 29988 <_sk_load_bgra_sse2_8bit+0x78> DB 102,67,15,16,4,152 ; movupd (%r8,%r11,4),%xmm0 + DB 67,15,16,76,152,16 ; movups 0x10(%r8,%r11,4),%xmm1 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,40,232 ; movapd %xmm0,%xmm5 DB 102,15,104,236 ; punpckhbw %xmm4,%xmm5 @@ -39821,18 +39761,18 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,152 ; ja 299ac <_sk_load_bgra_sse2_8bit+0x28> + DB 119,152 ; ja 29938 <_sk_load_bgra_sse2_8bit+0x28> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,89,0,0,0 ; lea 0x59(%rip),%r10 # 29a78 <_sk_load_bgra_sse2_8bit+0xf4> + DB 76,141,21,97,0,0,0 ; lea 0x61(%rip),%r10 # 29a0c <_sk_load_bgra_sse2_8bit+0xfc> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0 - DB 233,121,255,255,255 ; jmpq 299ac <_sk_load_bgra_sse2_8bit+0x28> + DB 233,121,255,255,255 ; jmpq 29938 <_sk_load_bgra_sse2_8bit+0x28> DB 102,67,15,110,68,152,8 ; movd 0x8(%r8,%r11,4),%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 DB 102,67,15,18,4,152 ; movlpd (%r8,%r11,4),%xmm0 - DB 233,98,255,255,255 ; jmpq 299ac <_sk_load_bgra_sse2_8bit+0x28> + DB 233,98,255,255,255 ; jmpq 29938 <_sk_load_bgra_sse2_8bit+0x28> DB 102,67,15,110,68,152,24 ; movd 0x18(%r8,%r11,4),%xmm0 DB 102,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm1 DB 243,67,15,16,68,152,20 ; movss 0x14(%r8,%r11,4),%xmm0 @@ -39841,23 +39781,23 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 15,40,200 ; movaps %xmm0,%xmm1 DB 243,67,15,16,68,152,16 ; movss 0x10(%r8,%r11,4),%xmm0 DB 243,15,16,200 ; movss %xmm0,%xmm1 - DB 233,46,255,255,255 ; jmpq 299a6 <_sk_load_bgra_sse2_8bit+0x22> - DB 176,255 ; mov $0xff,%al + DB 102,67,15,16,4,152 ; movupd (%r8,%r11,4),%xmm0 + DB 233,46,255,255,255 ; jmpq 29938 <_sk_load_bgra_sse2_8bit+0x28> + DB 102,144 ; xchg %ax,%ax + DB 168,255 ; test $0xff,%al DB 255 ; (bad) - DB 255,199 ; inc %edi DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) DB 255 ; (bad) - DB 187,255,255,255,46 ; mov $0x2effffff,%ebx + DB 255,243 ; push %rbx DB 255 ; (bad) DB 255 ; (bad) - DB 255,240 ; push %rax DB 255 ; (bad) + DB 232,255,255,255,214 ; callq ffffffffd7029a20 <_sk_overlay_sse2_8bit+0xffffffffd6ffd571> DB 255 ; (bad) DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) - DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255,202 ; dec %edx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -39872,9 +39812,9 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,92 ; jne 29b0c <_sk_load_bgra_dst_sse2_8bit+0x78> - DB 67,15,16,92,152,16 ; movups 0x10(%r8,%r11,4),%xmm3 + DB 117,92 ; jne 29aa0 <_sk_load_bgra_dst_sse2_8bit+0x78> DB 102,67,15,16,20,152 ; movupd (%r8,%r11,4),%xmm2 + DB 67,15,16,92,152,16 ; movups 0x10(%r8,%r11,4),%xmm3 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,40,234 ; movapd %xmm2,%xmm5 DB 102,15,104,236 ; punpckhbw %xmm4,%xmm5 @@ -39900,18 +39840,18 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,152 ; ja 29abc <_sk_load_bgra_dst_sse2_8bit+0x28> + DB 119,152 ; ja 29a50 <_sk_load_bgra_dst_sse2_8bit+0x28> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,89,0,0,0 ; lea 0x59(%rip),%r10 # 29b88 <_sk_load_bgra_dst_sse2_8bit+0xf4> + DB 76,141,21,97,0,0,0 ; lea 0x61(%rip),%r10 # 29b24 <_sk_load_bgra_dst_sse2_8bit+0xfc> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2 - DB 233,121,255,255,255 ; jmpq 29abc <_sk_load_bgra_dst_sse2_8bit+0x28> + DB 233,121,255,255,255 ; jmpq 29a50 <_sk_load_bgra_dst_sse2_8bit+0x28> DB 102,67,15,110,84,152,8 ; movd 0x8(%r8,%r11,4),%xmm2 DB 102,15,112,210,69 ; pshufd $0x45,%xmm2,%xmm2 DB 102,67,15,18,20,152 ; movlpd (%r8,%r11,4),%xmm2 - DB 233,98,255,255,255 ; jmpq 29abc <_sk_load_bgra_dst_sse2_8bit+0x28> + DB 233,98,255,255,255 ; jmpq 29a50 <_sk_load_bgra_dst_sse2_8bit+0x28> DB 102,67,15,110,84,152,24 ; movd 0x18(%r8,%r11,4),%xmm2 DB 102,15,112,218,69 ; pshufd $0x45,%xmm2,%xmm3 DB 243,67,15,16,84,152,20 ; movss 0x14(%r8,%r11,4),%xmm2 @@ -39920,23 +39860,23 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 243,67,15,16,84,152,16 ; movss 0x10(%r8,%r11,4),%xmm2 DB 243,15,16,218 ; movss %xmm2,%xmm3 - DB 233,46,255,255,255 ; jmpq 29ab6 <_sk_load_bgra_dst_sse2_8bit+0x22> - DB 176,255 ; mov $0xff,%al - DB 255 ; (bad) - DB 255,199 ; inc %edi + DB 102,67,15,16,20,152 ; movupd (%r8,%r11,4),%xmm2 + DB 233,46,255,255,255 ; jmpq 29a50 <_sk_load_bgra_dst_sse2_8bit+0x28> + DB 102,144 ; xchg %ax,%ax + DB 168,255 ; test $0xff,%al DB 255 ; (bad) DB 255 ; (bad) + DB 191,255,255,255,179 ; mov $0xb3ffffff,%edi DB 255 ; (bad) - DB 187,255,255,255,46 ; mov $0x2effffff,%ebx DB 255 ; (bad) + DB 255,243 ; push %rbx DB 255 ; (bad) - DB 255,240 ; push %rax DB 255 ; (bad) DB 255 ; (bad) + DB 232,255,255,255,214 ; callq ffffffffd7029b38 <_sk_overlay_sse2_8bit+0xffffffffd6ffd689> DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255,202 ; dec %edx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -39970,7 +39910,7 @@ _sk_store_bgra_sse2_8bit LABEL PROC DB 243,15,112,237,198 ; pshufhw $0xc6,%xmm5,%xmm5 DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 72,133,255 ; test %rdi,%rdi - DB 117,17 ; jne 29c25 <_sk_store_bgra_sse2_8bit+0x81> + DB 117,17 ; jne 29bc1 <_sk_store_bgra_sse2_8bit+0x81> DB 243,67,15,127,44,152 ; movdqu %xmm5,(%r8,%r11,4) DB 243,67,15,127,100,152,16 ; movdqu %xmm4,0x10(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -39979,25 +39919,25 @@ _sk_store_bgra_sse2_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 29c21 <_sk_store_bgra_sse2_8bit+0x7d> + DB 119,236 ; ja 29bbd <_sk_store_bgra_sse2_8bit+0x7d> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,76,0,0,0 ; lea 0x4c(%rip),%r10 # 29c8c <_sk_store_bgra_sse2_8bit+0xe8> + DB 76,141,21,76,0,0,0 ; lea 0x4c(%rip),%r10 # 29c28 <_sk_store_bgra_sse2_8bit+0xe8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,67,15,126,44,152 ; movd %xmm5,(%r8,%r11,4) - DB 235,208 ; jmp 29c21 <_sk_store_bgra_sse2_8bit+0x7d> + DB 235,208 ; jmp 29bbd <_sk_store_bgra_sse2_8bit+0x7d> DB 102,15,112,229,78 ; pshufd $0x4e,%xmm5,%xmm4 DB 102,67,15,126,100,152,8 ; movd %xmm4,0x8(%r8,%r11,4) DB 102,67,15,214,44,152 ; movq %xmm5,(%r8,%r11,4) - DB 235,188 ; jmp 29c21 <_sk_store_bgra_sse2_8bit+0x7d> + DB 235,188 ; jmp 29bbd <_sk_store_bgra_sse2_8bit+0x7d> DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 DB 102,67,15,126,116,152,24 ; movd %xmm6,0x18(%r8,%r11,4) DB 102,15,112,244,229 ; pshufd $0xe5,%xmm4,%xmm6 DB 102,67,15,126,116,152,20 ; movd %xmm6,0x14(%r8,%r11,4) DB 102,67,15,126,100,152,16 ; movd %xmm4,0x10(%r8,%r11,4) DB 243,67,15,127,44,152 ; movdqu %xmm5,(%r8,%r11,4) - DB 235,149 ; jmp 29c21 <_sk_store_bgra_sse2_8bit+0x7d> + DB 235,149 ; jmp 29bbd <_sk_store_bgra_sse2_8bit+0x7d> DB 189,255,255,255,209 ; mov $0xd1ffffff,%ebp DB 255 ; (bad) DB 255 ; (bad) @@ -40028,35 +39968,35 @@ _sk_store_565_sse2_8bit LABEL PROC DB 77,1,192 ; add %r8,%r8 DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 - DB 102,68,15,111,13,249,62,0,0 ; movdqa 0x3ef9(%rip),%xmm9 # 2dbc0 <_sk_overlay_sse2_8bit+0x1668> - DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 - DB 102,69,15,219,193 ; pand %xmm9,%xmm8 - DB 102,68,15,219,200 ; pand %xmm0,%xmm9 - DB 102,65,15,114,241,8 ; pslld $0x8,%xmm9 - DB 102,65,15,114,240,8 ; pslld $0x8,%xmm8 - DB 102,15,111,37,230,62,0,0 ; movdqa 0x3ee6(%rip),%xmm4 # 2dbd0 <_sk_overlay_sse2_8bit+0x1678> + DB 102,15,111,37,110,62,0,0 ; movdqa 0x3e6e(%rip),%xmm4 # 2dad0 <_sk_overlay_sse2_8bit+0x1621> DB 102,15,111,233 ; movdqa %xmm1,%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,15,114,212,5 ; psrld $0x5,%xmm4 - DB 102,15,114,213,5 ; psrld $0x5,%xmm5 - DB 102,15,111,61,216,62,0,0 ; movdqa 0x3ed8(%rip),%xmm7 # 2dbe0 <_sk_overlay_sse2_8bit+0x1688> - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,15,219,247 ; pand %xmm7,%xmm6 - DB 102,15,219,249 ; pand %xmm1,%xmm7 - DB 102,15,114,215,19 ; psrld $0x13,%xmm7 - DB 102,15,114,214,19 ; psrld $0x13,%xmm6 - DB 102,65,15,235,232 ; por %xmm8,%xmm5 - DB 102,65,15,235,225 ; por %xmm9,%xmm4 - DB 102,15,235,230 ; por %xmm6,%xmm4 + DB 102,15,114,244,8 ; pslld $0x8,%xmm4 + DB 102,15,114,245,8 ; pslld $0x8,%xmm5 + DB 102,15,111,53,96,62,0,0 ; movdqa 0x3e60(%rip),%xmm6 # 2dae0 <_sk_overlay_sse2_8bit+0x1631> + DB 102,15,111,249 ; movdqa %xmm1,%xmm7 + DB 102,15,219,254 ; pand %xmm6,%xmm7 + DB 102,15,219,240 ; pand %xmm0,%xmm6 + DB 102,15,114,214,5 ; psrld $0x5,%xmm6 + DB 102,15,235,244 ; por %xmm4,%xmm6 + DB 102,15,114,215,5 ; psrld $0x5,%xmm7 + DB 102,15,235,253 ; por %xmm5,%xmm7 + DB 102,15,111,45,74,62,0,0 ; movdqa 0x3e4a(%rip),%xmm5 # 2daf0 <_sk_overlay_sse2_8bit+0x1641> + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,219,229 ; pand %xmm5,%xmm4 + DB 102,15,219,233 ; pand %xmm1,%xmm5 + DB 102,15,114,213,19 ; psrld $0x13,%xmm5 DB 102,15,235,239 ; por %xmm7,%xmm5 + DB 102,15,114,212,19 ; psrld $0x13,%xmm4 + DB 102,15,235,230 ; por %xmm6,%xmm4 DB 102,15,114,245,16 ; pslld $0x10,%xmm5 DB 102,15,114,229,16 ; psrad $0x10,%xmm5 DB 102,15,114,244,16 ; pslld $0x10,%xmm4 DB 102,15,114,228,16 ; psrad $0x10,%xmm4 DB 102,15,107,229 ; packssdw %xmm5,%xmm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,10 ; jne 29d57 <_sk_store_565_sse2_8bit+0xaf> + DB 117,10 ; jne 29ceb <_sk_store_565_sse2_8bit+0xa7> DB 243,67,15,127,36,88 ; movdqu %xmm4,(%r8,%r11,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -40064,19 +40004,19 @@ _sk_store_565_sse2_8bit LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 29d53 <_sk_store_565_sse2_8bit+0xab> + DB 119,236 ; ja 29ce7 <_sk_store_565_sse2_8bit+0xa3> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,82,0,0,0 ; lea 0x52(%rip),%r10 # 29dc4 <_sk_store_565_sse2_8bit+0x11c> + DB 76,141,21,82,0,0,0 ; lea 0x52(%rip),%r10 # 29d58 <_sk_store_565_sse2_8bit+0x114> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 102,15,126,224 ; movd %xmm4,%eax DB 102,67,137,4,88 ; mov %ax,(%r8,%r11,2) - DB 235,205 ; jmp 29d53 <_sk_store_565_sse2_8bit+0xab> + DB 235,205 ; jmp 29ce7 <_sk_store_565_sse2_8bit+0xa3> DB 102,15,197,196,2 ; pextrw $0x2,%xmm4,%eax DB 102,67,137,68,88,4 ; mov %ax,0x4(%r8,%r11,2) DB 102,67,15,126,36,88 ; movd %xmm4,(%r8,%r11,2) - DB 235,186 ; jmp 29d53 <_sk_store_565_sse2_8bit+0xab> + DB 235,186 ; jmp 29ce7 <_sk_store_565_sse2_8bit+0xa3> DB 102,15,197,196,6 ; pextrw $0x6,%xmm4,%eax DB 102,67,137,68,88,12 ; mov %ax,0xc(%r8,%r11,2) DB 102,15,197,196,5 ; pextrw $0x5,%xmm4,%eax @@ -40084,7 +40024,7 @@ _sk_store_565_sse2_8bit LABEL PROC DB 102,15,197,196,4 ; pextrw $0x4,%xmm4,%eax DB 102,67,137,68,88,8 ; mov %ax,0x8(%r8,%r11,2) DB 102,67,15,214,36,88 ; movq %xmm4,(%r8,%r11,2) - DB 235,145 ; jmp 29d53 <_sk_store_565_sse2_8bit+0xab> + DB 235,145 ; jmp 29ce7 <_sk_store_565_sse2_8bit+0xa3> DB 102,144 ; xchg %ax,%ax DB 183,255 ; mov $0xff,%bh DB 255 ; (bad) @@ -40098,7 +40038,7 @@ _sk_store_565_sse2_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 29dd5 <_sk_store_565_sse2_8bit+0x12d> + DB 235,255 ; jmp 29d69 <_sk_store_565_sse2_8bit+0x125> DB 255 ; (bad) DB 255,224 ; jmpq *%rax DB 255 ; (bad) @@ -40118,31 +40058,31 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,219,0,0,0 ; jne 29eda <_sk_load_565_dst_sse2_8bit+0xfa> + DB 15,133,219,0,0,0 ; jne 29e6e <_sk_load_565_dst_sse2_8bit+0xfa> DB 243,67,15,111,20,88 ; movdqu (%r8,%r11,2),%xmm2 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,111,218 ; movdqa %xmm2,%xmm3 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 DB 102,15,97,212 ; punpcklwd %xmm4,%xmm2 - DB 102,15,111,53,179,56,0,0 ; movdqa 0x38b3(%rip),%xmm6 # 2d6d0 <_sk_overlay_sse2_8bit+0x1178> + DB 102,15,111,53,15,56,0,0 ; movdqa 0x380f(%rip),%xmm6 # 2d5c0 <_sk_overlay_sse2_8bit+0x1111> DB 102,15,111,250 ; movdqa %xmm2,%xmm7 DB 102,15,219,254 ; pand %xmm6,%xmm7 DB 102,15,219,243 ; pand %xmm3,%xmm6 DB 102,15,114,214,8 ; psrld $0x8,%xmm6 DB 102,15,114,215,8 ; psrld $0x8,%xmm7 - DB 102,15,111,45,213,56,0,0 ; movdqa 0x38d5(%rip),%xmm5 # 2d710 <_sk_overlay_sse2_8bit+0x11b8> + DB 102,15,111,45,49,56,0,0 ; movdqa 0x3831(%rip),%xmm5 # 2d600 <_sk_overlay_sse2_8bit+0x1151> DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,15,219,229 ; pand %xmm5,%xmm4 DB 102,15,219,235 ; pand %xmm3,%xmm5 DB 102,15,114,245,19 ; pslld $0x13,%xmm5 + DB 102,15,235,238 ; por %xmm6,%xmm5 DB 102,15,114,244,19 ; pslld $0x13,%xmm4 DB 102,15,235,231 ; por %xmm7,%xmm4 - DB 102,15,235,238 ; por %xmm6,%xmm5 - DB 102,15,111,53,143,56,0,0 ; movdqa 0x388f(%rip),%xmm6 # 2d6f0 <_sk_overlay_sse2_8bit+0x1198> + DB 102,15,111,53,235,55,0,0 ; movdqa 0x37eb(%rip),%xmm6 # 2d5e0 <_sk_overlay_sse2_8bit+0x1131> DB 102,15,111,251 ; movdqa %xmm3,%xmm7 DB 102,15,219,254 ; pand %xmm6,%xmm7 DB 102,15,114,247,5 ; pslld $0x5,%xmm7 - DB 102,68,15,111,5,153,61,0,0 ; movdqa 0x3d99(%rip),%xmm8 # 2dc10 <_sk_overlay_sse2_8bit+0x16b8> + DB 102,68,15,111,5,21,61,0,0 ; movdqa 0x3d15(%rip),%xmm8 # 2db20 <_sk_overlay_sse2_8bit+0x1671> DB 102,65,15,219,216 ; pand %xmm8,%xmm3 DB 102,15,114,211,1 ; psrld $0x1,%xmm3 DB 102,15,235,223 ; por %xmm7,%xmm3 @@ -40155,10 +40095,10 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 102,15,235,214 ; por %xmm6,%xmm2 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,114,212,5 ; psrld $0x5,%xmm4 - DB 102,15,111,53,74,61,0,0 ; movdqa 0x3d4a(%rip),%xmm6 # 2dc00 <_sk_overlay_sse2_8bit+0x16a8> + DB 102,15,111,53,198,60,0,0 ; movdqa 0x3cc6(%rip),%xmm6 # 2db10 <_sk_overlay_sse2_8bit+0x1661> DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 102,15,219,238 ; pand %xmm6,%xmm5 - DB 102,15,111,53,202,60,0,0 ; movdqa 0x3cca(%rip),%xmm6 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,53,70,60,0,0 ; movdqa 0x3c46(%rip),%xmm6 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,235,214 ; por %xmm6,%xmm2 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,222 ; por %xmm6,%xmm3 @@ -40170,26 +40110,26 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,19,255,255,255 ; ja 29e05 <_sk_load_565_dst_sse2_8bit+0x25> + DB 15,135,19,255,255,255 ; ja 29d99 <_sk_load_565_dst_sse2_8bit+0x25> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,91,0,0,0 ; lea 0x5b(%rip),%r10 # 29f58 <_sk_load_565_dst_sse2_8bit+0x178> + DB 76,141,21,91,0,0,0 ; lea 0x5b(%rip),%r10 # 29eec <_sk_load_565_dst_sse2_8bit+0x178> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,183,4,88 ; movzwl (%r8,%r11,2),%eax DB 102,15,110,208 ; movd %eax,%xmm2 - DB 233,241,254,255,255 ; jmpq 29e05 <_sk_load_565_dst_sse2_8bit+0x25> + DB 233,241,254,255,255 ; jmpq 29d99 <_sk_load_565_dst_sse2_8bit+0x25> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,67,15,196,84,88,4,2 ; pinsrw $0x2,0x4(%r8,%r11,2),%xmm2 DB 243,67,15,16,28,88 ; movss (%r8,%r11,2),%xmm3 DB 243,15,16,211 ; movss %xmm3,%xmm2 - DB 233,214,254,255,255 ; jmpq 29e05 <_sk_load_565_dst_sse2_8bit+0x25> + DB 233,214,254,255,255 ; jmpq 29d99 <_sk_load_565_dst_sse2_8bit+0x25> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,67,15,196,84,88,12,6 ; pinsrw $0x6,0xc(%r8,%r11,2),%xmm2 DB 102,67,15,196,84,88,10,5 ; pinsrw $0x5,0xa(%r8,%r11,2),%xmm2 DB 102,67,15,196,84,88,8,4 ; pinsrw $0x4,0x8(%r8,%r11,2),%xmm2 DB 102,67,15,18,20,88 ; movlpd (%r8,%r11,2),%xmm2 - DB 233,175,254,255,255 ; jmpq 29e05 <_sk_load_565_dst_sse2_8bit+0x25> + DB 233,175,254,255,255 ; jmpq 29d99 <_sk_load_565_dst_sse2_8bit+0x25> DB 102,144 ; xchg %ax,%ax DB 174 ; scas %es:(%rdi),%al DB 255 ; (bad) @@ -40202,7 +40142,7 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 29f69 <_sk_load_565_dst_sse2_8bit+0x189> + DB 235,255 ; jmp 29efd <_sk_load_565_dst_sse2_8bit+0x189> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -40221,10 +40161,10 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,48 ; jne 29fbc <_sk_load_a8_sse2_8bit+0x48> + DB 117,48 ; jne 29f50 <_sk_load_a8_sse2_8bit+0x48> DB 243,67,15,126,4,24 ; movq (%r8,%r11,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 102,15,219,5,178,52,0,0 ; pand 0x34b2(%rip),%xmm0 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,5,14,52,0,0 ; pand 0x340e(%rip),%xmm0 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 @@ -40238,15 +40178,15 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,198 ; ja 29f96 <_sk_load_a8_sse2_8bit+0x22> + DB 119,198 ; ja 29f2a <_sk_load_a8_sse2_8bit+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,113,0,0,0 ; lea 0x71(%rip),%r10 # 2a04c <_sk_load_a8_sse2_8bit+0xd8> + DB 76,141,21,113,0,0,0 ; lea 0x71(%rip),%r10 # 29fe0 <_sk_load_a8_sse2_8bit+0xd8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 - DB 235,167 ; jmp 29f96 <_sk_load_a8_sse2_8bit+0x22> + DB 235,167 ; jmp 29f2a <_sk_load_a8_sse2_8bit+0x22> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,192,2 ; pinsrw $0x2,%eax,%xmm0 @@ -40254,7 +40194,7 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 102,15,110,200 ; movd %eax,%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 243,15,16,193 ; movss %xmm1,%xmm0 - DB 235,133 ; jmp 29f96 <_sk_load_a8_sse2_8bit+0x22> + DB 235,133 ; jmp 29f2a <_sk_load_a8_sse2_8bit+0x22> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,192,6 ; pinsrw $0x6,%eax,%xmm0 @@ -40265,7 +40205,7 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 102,67,15,110,12,24 ; movd (%r8,%r11,1),%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 233,77,255,255,255 ; jmpq 29f96 <_sk_load_a8_sse2_8bit+0x22> + DB 233,77,255,255,255 ; jmpq 29f2a <_sk_load_a8_sse2_8bit+0x22> DB 15,31,0 ; nopl (%rax) DB 152 ; cwtl DB 255 ; (bad) @@ -40297,10 +40237,10 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,48 ; jne 2a0b0 <_sk_load_a8_dst_sse2_8bit+0x48> + DB 117,48 ; jne 2a044 <_sk_load_a8_dst_sse2_8bit+0x48> DB 243,67,15,126,20,24 ; movq (%r8,%r11,1),%xmm2 DB 102,15,96,208 ; punpcklbw %xmm0,%xmm2 - DB 102,15,219,21,190,51,0,0 ; pand 0x33be(%rip),%xmm2 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,21,26,51,0,0 ; pand 0x331a(%rip),%xmm2 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,111,218 ; movdqa %xmm2,%xmm3 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 @@ -40314,15 +40254,15 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,198 ; ja 2a08a <_sk_load_a8_dst_sse2_8bit+0x22> + DB 119,198 ; ja 2a01e <_sk_load_a8_dst_sse2_8bit+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,113,0,0,0 ; lea 0x71(%rip),%r10 # 2a140 <_sk_load_a8_dst_sse2_8bit+0xd8> + DB 76,141,21,113,0,0,0 ; lea 0x71(%rip),%r10 # 2a0d4 <_sk_load_a8_dst_sse2_8bit+0xd8> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,208 ; movd %eax,%xmm2 - DB 235,167 ; jmp 2a08a <_sk_load_a8_dst_sse2_8bit+0x22> + DB 235,167 ; jmp 2a01e <_sk_load_a8_dst_sse2_8bit+0x22> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,208,2 ; pinsrw $0x2,%eax,%xmm2 @@ -40330,7 +40270,7 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 102,15,110,216 ; movd %eax,%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 243,15,16,211 ; movss %xmm3,%xmm2 - DB 235,133 ; jmp 2a08a <_sk_load_a8_dst_sse2_8bit+0x22> + DB 235,133 ; jmp 2a01e <_sk_load_a8_dst_sse2_8bit+0x22> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,208,6 ; pinsrw $0x6,%eax,%xmm2 @@ -40341,7 +40281,7 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 102,67,15,110,28,24 ; movd (%r8,%r11,1),%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 242,15,16,211 ; movsd %xmm3,%xmm2 - DB 233,77,255,255,255 ; jmpq 2a08a <_sk_load_a8_dst_sse2_8bit+0x22> + DB 233,77,255,255,255 ; jmpq 2a01e <_sk_load_a8_dst_sse2_8bit+0x22> DB 15,31,0 ; nopl (%rax) DB 152 ; cwtl DB 255 ; (bad) @@ -40366,7 +40306,6 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC PUBLIC _sk_store_a8_sse2_8bit _sk_store_a8_sse2_8bit LABEL PROC - DB 72,131,236,88 ; sub $0x58,%rsp DB 72,173 ; lods %ds:(%rsi),%rax DB 76,99,72,8 ; movslq 0x8(%rax),%r9 DB 76,99,193 ; movslq %ecx,%r8 @@ -40383,65 +40322,53 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 102,15,114,228,16 ; psrad $0x10,%xmm4 DB 102,15,107,229 ; packssdw %xmm5,%xmm4 DB 72,133,255 ; test %rdi,%rdi - DB 117,26 ; jne 2a1bc <_sk_store_a8_sse2_8bit+0x60> - DB 102,15,219,37,166,50,0,0 ; pand 0x32a6(%rip),%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 117,22 ; jne 2a148 <_sk_store_a8_sse2_8bit+0x58> + DB 102,15,219,37,6,50,0,0 ; pand 0x3206(%rip),%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,67,15,214,36,24 ; movq %xmm4,(%r8,%r11,1) DB 72,173 ; lods %ds:(%rsi),%rax - DB 72,131,196,88 ; add $0x58,%rsp DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,232 ; ja 2a1b4 <_sk_store_a8_sse2_8bit+0x58> + DB 119,236 ; ja 2a144 <_sk_store_a8_sse2_8bit+0x54> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 2a25c <_sk_store_a8_sse2_8bit+0x100> + DB 76,141,21,105,0,0,0 ; lea 0x69(%rip),%r10 # 2a1cc <_sk_store_a8_sse2_8bit+0xdc> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax - DB 102,15,127,100,36,64 ; movdqa %xmm4,0x40(%rsp) - DB 138,68,36,64 ; mov 0x40(%rsp),%al + DB 102,15,126,224 ; movd %xmm4,%eax DB 67,136,4,24 ; mov %al,(%r8,%r11,1) - DB 235,196 ; jmp 2a1b4 <_sk_store_a8_sse2_8bit+0x58> - DB 102,15,127,100,36,48 ; movdqa %xmm4,0x30(%rsp) - DB 138,68,36,52 ; mov 0x34(%rsp),%al + DB 235,206 ; jmp 2a144 <_sk_store_a8_sse2_8bit+0x54> + DB 102,15,197,196,2 ; pextrw $0x2,%xmm4,%eax DB 67,136,68,24,2 ; mov %al,0x2(%r8,%r11,1) - DB 102,15,219,37,73,50,0,0 ; pand 0x3249(%rip),%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,37,184,49,0,0 ; pand 0x31b8(%rip),%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax DB 102,67,137,4,24 ; mov %ax,(%r8,%r11,1) - DB 235,158 ; jmp 2a1b4 <_sk_store_a8_sse2_8bit+0x58> - DB 102,15,127,100,36,32 ; movdqa %xmm4,0x20(%rsp) - DB 138,68,36,44 ; mov 0x2c(%rsp),%al + DB 235,173 ; jmp 2a144 <_sk_store_a8_sse2_8bit+0x54> + DB 102,15,197,196,6 ; pextrw $0x6,%xmm4,%eax DB 67,136,68,24,6 ; mov %al,0x6(%r8,%r11,1) - DB 102,15,127,100,36,16 ; movdqa %xmm4,0x10(%rsp) - DB 138,68,36,26 ; mov 0x1a(%rsp),%al + DB 102,15,197,196,5 ; pextrw $0x5,%xmm4,%eax DB 67,136,68,24,5 ; mov %al,0x5(%r8,%r11,1) - DB 102,15,127,36,36 ; movdqa %xmm4,(%rsp) - DB 138,68,36,8 ; mov 0x8(%rsp),%al + DB 102,15,197,196,4 ; pextrw $0x4,%xmm4,%eax DB 67,136,68,24,4 ; mov %al,0x4(%r8,%r11,1) - DB 102,15,219,37,6,50,0,0 ; pand 0x3206(%rip),%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,37,131,49,0,0 ; pand 0x3183(%rip),%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,67,15,126,36,24 ; movd %xmm4,(%r8,%r11,1) - DB 233,91,255,255,255 ; jmpq 2a1b4 <_sk_store_a8_sse2_8bit+0x58> - DB 15,31,0 ; nopl (%rax) - DB 132,255 ; test %bh,%bh - DB 255 ; (bad) - DB 255,163,255,255,255,148 ; jmpq *-0x6b000001(%rbx) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi + DB 233,120,255,255,255 ; jmpq 2a144 <_sk_store_a8_sse2_8bit+0x54> + DB 160,255,255,255,180,255,255,255,170 ; movabs 0xaaffffffb4ffffff,%al DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 216,255 ; fdivr %st(7),%st + DB 233,255,255,255,223 ; jmpq ffffffffe002a1dc <_sk_overlay_sse2_8bit+0xffffffffdfffdd2d> DB 255 ; (bad) - DB 255,201 ; dec %ecx DB 255 ; (bad) + DB 255,213 ; callq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 186 ; .byte 0xba + DB 255,203 ; dec %ebx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -40455,15 +40382,15 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,116 ; jne 2a304 <_sk_load_g8_sse2_8bit+0x8c> + DB 117,116 ; jne 2a274 <_sk_load_g8_sse2_8bit+0x8c> DB 243,67,15,126,4,24 ; movq (%r8,%r11,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 102,15,219,5,174,49,0,0 ; pand 0x31ae(%rip),%xmm0 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,5,46,49,0,0 ; pand 0x312e(%rip),%xmm0 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 DB 102,15,97,225 ; punpcklwd %xmm1,%xmm4 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,15,111,45,118,57,0,0 ; movdqa 0x3976(%rip),%xmm5 # 2dc30 <_sk_overlay_sse2_8bit+0x16d8> + DB 102,15,111,45,38,57,0,0 ; movdqa 0x3926(%rip),%xmm5 # 2db50 <_sk_overlay_sse2_8bit+0x16a1> DB 102,15,112,240,245 ; pshufd $0xf5,%xmm0,%xmm6 DB 102,15,244,197 ; pmuludq %xmm5,%xmm0 DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 @@ -40476,7 +40403,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,244,245 ; pmuludq %xmm5,%xmm6 DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 DB 102,15,98,196 ; punpckldq %xmm4,%xmm0 - DB 102,15,111,37,152,56,0,0 ; movdqa 0x3898(%rip),%xmm4 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,37,56,56,0,0 ; movdqa 0x3838(%rip),%xmm4 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,235,196 ; por %xmm4,%xmm0 DB 102,15,235,204 ; por %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -40486,15 +40413,15 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,130 ; ja 2a29a <_sk_load_g8_sse2_8bit+0x22> + DB 119,130 ; ja 2a20a <_sk_load_g8_sse2_8bit+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,117,0,0,0 ; lea 0x75(%rip),%r10 # 2a398 <_sk_load_g8_sse2_8bit+0x120> + DB 76,141,21,117,0,0,0 ; lea 0x75(%rip),%r10 # 2a308 <_sk_load_g8_sse2_8bit+0x120> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,192 ; movd %eax,%xmm0 - DB 233,96,255,255,255 ; jmpq 2a29a <_sk_load_g8_sse2_8bit+0x22> + DB 233,96,255,255,255 ; jmpq 2a20a <_sk_load_g8_sse2_8bit+0x22> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,192,2 ; pinsrw $0x2,%eax,%xmm0 @@ -40502,7 +40429,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,110,200 ; movd %eax,%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 243,15,16,193 ; movss %xmm1,%xmm0 - DB 233,59,255,255,255 ; jmpq 2a29a <_sk_load_g8_sse2_8bit+0x22> + DB 233,59,255,255,255 ; jmpq 2a20a <_sk_load_g8_sse2_8bit+0x22> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,192,6 ; pinsrw $0x6,%eax,%xmm0 @@ -40513,7 +40440,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,67,15,110,12,24 ; movd (%r8,%r11,1),%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 233,3,255,255,255 ; jmpq 2a29a <_sk_load_g8_sse2_8bit+0x22> + DB 233,3,255,255,255 ; jmpq 2a20a <_sk_load_g8_sse2_8bit+0x22> DB 144 ; nop DB 148 ; xchg %eax,%esp DB 255 ; (bad) @@ -40545,15 +40472,15 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 117,116 ; jne 2a440 <_sk_load_g8_dst_sse2_8bit+0x8c> + DB 117,116 ; jne 2a3b0 <_sk_load_g8_dst_sse2_8bit+0x8c> DB 243,67,15,126,20,24 ; movq (%r8,%r11,1),%xmm2 DB 102,15,96,208 ; punpcklbw %xmm0,%xmm2 - DB 102,15,219,21,114,48,0,0 ; pand 0x3072(%rip),%xmm2 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,21,242,47,0,0 ; pand 0x2ff2(%rip),%xmm2 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,15,97,227 ; punpcklwd %xmm3,%xmm4 DB 102,15,105,211 ; punpckhwd %xmm3,%xmm2 - DB 102,15,111,45,58,56,0,0 ; movdqa 0x383a(%rip),%xmm5 # 2dc30 <_sk_overlay_sse2_8bit+0x16d8> + DB 102,15,111,45,234,55,0,0 ; movdqa 0x37ea(%rip),%xmm5 # 2db50 <_sk_overlay_sse2_8bit+0x16a1> DB 102,15,112,242,245 ; pshufd $0xf5,%xmm2,%xmm6 DB 102,15,244,213 ; pmuludq %xmm5,%xmm2 DB 102,15,112,218,232 ; pshufd $0xe8,%xmm2,%xmm3 @@ -40566,7 +40493,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,244,245 ; pmuludq %xmm5,%xmm6 DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 DB 102,15,98,212 ; punpckldq %xmm4,%xmm2 - DB 102,15,111,37,92,55,0,0 ; movdqa 0x375c(%rip),%xmm4 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 102,15,111,37,252,54,0,0 ; movdqa 0x36fc(%rip),%xmm4 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,220 ; por %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -40576,15 +40503,15 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,130 ; ja 2a3d6 <_sk_load_g8_dst_sse2_8bit+0x22> + DB 119,130 ; ja 2a346 <_sk_load_g8_dst_sse2_8bit+0x22> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,117,0,0,0 ; lea 0x75(%rip),%r10 # 2a4d4 <_sk_load_g8_dst_sse2_8bit+0x120> + DB 76,141,21,117,0,0,0 ; lea 0x75(%rip),%r10 # 2a444 <_sk_load_g8_dst_sse2_8bit+0x120> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,208 ; movd %eax,%xmm2 - DB 233,96,255,255,255 ; jmpq 2a3d6 <_sk_load_g8_dst_sse2_8bit+0x22> + DB 233,96,255,255,255 ; jmpq 2a346 <_sk_load_g8_dst_sse2_8bit+0x22> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,208,2 ; pinsrw $0x2,%eax,%xmm2 @@ -40592,7 +40519,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,110,216 ; movd %eax,%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 243,15,16,211 ; movss %xmm3,%xmm2 - DB 233,59,255,255,255 ; jmpq 2a3d6 <_sk_load_g8_dst_sse2_8bit+0x22> + DB 233,59,255,255,255 ; jmpq 2a346 <_sk_load_g8_dst_sse2_8bit+0x22> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,208,6 ; pinsrw $0x6,%eax,%xmm2 @@ -40603,7 +40530,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,67,15,110,28,24 ; movd (%r8,%r11,1),%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 242,15,16,211 ; movsd %xmm3,%xmm2 - DB 233,3,255,255,255 ; jmpq 2a3d6 <_sk_load_g8_dst_sse2_8bit+0x22> + DB 233,3,255,255,255 ; jmpq 2a346 <_sk_load_g8_dst_sse2_8bit+0x22> DB 144 ; nop DB 148 ; xchg %eax,%esp DB 255 ; (bad) @@ -40636,9 +40563,9 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,253,0,0,0 ; jne 2a60d <_sk_srcover_rgba_8888_sse2_8bit+0x11d> - DB 71,15,16,68,152,16 ; movups 0x10(%r8,%r11,4),%xmm8 + DB 15,133,253,0,0,0 ; jne 2a57d <_sk_srcover_rgba_8888_sse2_8bit+0x11d> DB 102,71,15,16,12,152 ; movupd (%r8,%r11,4),%xmm9 + DB 71,15,16,68,152,16 ; movups 0x10(%r8,%r11,4),%xmm8 DB 72,133,255 ; test %rdi,%rdi DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 @@ -40677,15 +40604,15 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 102,65,15,253,238 ; paddw %xmm14,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,245 ; packuswb %xmm5,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,252 ; packuswb %xmm4,%xmm7 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 DB 102,68,15,248,198 ; psubb %xmm6,%xmm8 DB 102,68,15,248,207 ; psubb %xmm7,%xmm9 DB 102,68,15,252,200 ; paddb %xmm0,%xmm9 DB 102,68,15,252,193 ; paddb %xmm1,%xmm8 - DB 117,78 ; jne 2a64a <_sk_srcover_rgba_8888_sse2_8bit+0x15a> + DB 117,78 ; jne 2a5ba <_sk_srcover_rgba_8888_sse2_8bit+0x15a> DB 243,71,15,127,12,152 ; movdqu %xmm9,(%r8,%r11,4) DB 243,71,15,127,68,152,16 ; movdqu %xmm8,0x10(%r8,%r11,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -40696,30 +40623,30 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,241,254,255,255 ; ja 2a51c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> + DB 15,135,241,254,255,255 ; ja 2a48c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,210,0,0,0 ; lea 0xd2(%rip),%r10 # 2a708 <_sk_srcover_rgba_8888_sse2_8bit+0x218> + DB 76,141,21,214,0,0,0 ; lea 0xd6(%rip),%r10 # 2a67c <_sk_srcover_rgba_8888_sse2_8bit+0x21c> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 243,71,15,16,12,152 ; movss (%r8,%r11,4),%xmm9 - DB 233,210,254,255,255 ; jmpq 2a51c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> + DB 233,210,254,255,255 ; jmpq 2a48c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,175 ; ja 2a609 <_sk_srcover_rgba_8888_sse2_8bit+0x119> + DB 119,175 ; ja 2a579 <_sk_srcover_rgba_8888_sse2_8bit+0x119> DB 65,15,182,193 ; movzbl %r9b,%eax - DB 76,141,13,191,0,0,0 ; lea 0xbf(%rip),%r9 # 2a724 <_sk_srcover_rgba_8888_sse2_8bit+0x234> + DB 76,141,13,195,0,0,0 ; lea 0xc3(%rip),%r9 # 2a698 <_sk_srcover_rgba_8888_sse2_8bit+0x238> DB 73,99,4,129 ; movslq (%r9,%rax,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax DB 102,71,15,126,12,152 ; movd %xmm9,(%r8,%r11,4) - DB 235,147 ; jmp 2a609 <_sk_srcover_rgba_8888_sse2_8bit+0x119> + DB 235,147 ; jmp 2a579 <_sk_srcover_rgba_8888_sse2_8bit+0x119> DB 102,67,15,110,100,152,8 ; movd 0x8(%r8,%r11,4),%xmm4 DB 102,68,15,112,204,69 ; pshufd $0x45,%xmm4,%xmm9 DB 102,71,15,18,12,152 ; movlpd (%r8,%r11,4),%xmm9 - DB 233,142,254,255,255 ; jmpq 2a51c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> + DB 233,142,254,255,255 ; jmpq 2a48c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> DB 102,67,15,110,100,152,24 ; movd 0x18(%r8,%r11,4),%xmm4 DB 102,68,15,112,196,69 ; pshufd $0x45,%xmm4,%xmm8 DB 243,67,15,16,100,152,20 ; movss 0x14(%r8,%r11,4),%xmm4 @@ -40728,50 +40655,47 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 68,15,40,196 ; movaps %xmm4,%xmm8 DB 243,67,15,16,100,152,16 ; movss 0x10(%r8,%r11,4),%xmm4 DB 243,68,15,16,196 ; movss %xmm4,%xmm8 - DB 233,85,254,255,255 ; jmpq 2a516 <_sk_srcover_rgba_8888_sse2_8bit+0x26> + DB 102,71,15,16,12,152 ; movupd (%r8,%r11,4),%xmm9 + DB 233,85,254,255,255 ; jmpq 2a48c <_sk_srcover_rgba_8888_sse2_8bit+0x2c> DB 102,65,15,112,225,78 ; pshufd $0x4e,%xmm9,%xmm4 DB 102,67,15,126,100,152,8 ; movd %xmm4,0x8(%r8,%r11,4) DB 102,71,15,214,12,152 ; movq %xmm9,(%r8,%r11,4) - DB 233,48,255,255,255 ; jmpq 2a609 <_sk_srcover_rgba_8888_sse2_8bit+0x119> + DB 233,42,255,255,255 ; jmpq 2a579 <_sk_srcover_rgba_8888_sse2_8bit+0x119> DB 102,65,15,112,224,78 ; pshufd $0x4e,%xmm8,%xmm4 DB 102,67,15,126,100,152,24 ; movd %xmm4,0x18(%r8,%r11,4) DB 102,65,15,112,224,229 ; pshufd $0xe5,%xmm8,%xmm4 DB 102,67,15,126,100,152,20 ; movd %xmm4,0x14(%r8,%r11,4) DB 102,71,15,126,68,152,16 ; movd %xmm8,0x10(%r8,%r11,4) DB 243,71,15,127,12,152 ; movdqu %xmm9,(%r8,%r11,4) - DB 233,4,255,255,255 ; jmpq 2a609 <_sk_srcover_rgba_8888_sse2_8bit+0x119> - DB 15,31,0 ; nopl (%rax) - DB 55 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 123,255 ; jnp 2a70d <_sk_srcover_rgba_8888_sse2_8bit+0x21d> + DB 233,254,254,255,255 ; jmpq 2a579 <_sk_srcover_rgba_8888_sse2_8bit+0x119> + DB 144 ; nop + DB 51,255 ; xor %edi,%edi DB 255 ; (bad) - DB 255,110,255 ; ljmp *-0x1(%rsi) + DB 255,119,255 ; pushq -0x1(%rdi) DB 255 ; (bad) - DB 255,14 ; decl (%rsi) - DB 254 ; (bad) + DB 255,106,255 ; ljmp *-0x1(%rdx) DB 255 ; (bad) - DB 255,168,255,255,255,147 ; ljmp *-0x6c000001(%rax) + DB 255,176,255,255,255,164 ; pushq -0x5b000001(%rax) DB 255 ; (bad) DB 255 ; (bad) - DB 255,134,255,255,255,74 ; incl 0x4affffff(%rsi) + DB 255,143,255,255,255,130 ; decl -0x7d000001(%rdi) DB 255 ; (bad) DB 255 ; (bad) - DB 255,170,255,255,255,157 ; ljmp *-0x62000001(%rdx) + DB 255,70,255 ; incl -0x1(%rsi) DB 255 ; (bad) + DB 255,172,255,255,255,159,255 ; ljmp *-0x600001(%rdi,%rdi,8) DB 255 ; (bad) - DB 255,214 ; callq *%rsi DB 255 ; (bad) + DB 216,255 ; fdivr %st(7),%st DB 255 ; (bad) - DB 255,207 ; dec %edi + DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) - DB 255,194 ; inc %edx + DB 255,196 ; inc %esp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff - DB 181,255 ; mov $0xff,%ch + DB 183,255 ; mov $0xff,%bh DB 255 ; (bad) DB 255 ; .byte 0xff @@ -40781,7 +40705,7 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,0 ; movss (%rax),%xmm0 - DB 243,15,89,5,128,43,0,0 ; mulss 0x2b80(%rip),%xmm0 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 243,15,89,5,236,42,0,0 ; mulss 0x2aec(%rip),%xmm0 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 243,15,44,192 ; cvttss2si %xmm0,%eax DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 @@ -40793,7 +40717,7 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,15,96,246 ; punpcklbw %xmm6,%xmm6 DB 242,15,112,246,0 ; pshuflw $0x0,%xmm6,%xmm6 DB 102,15,112,246,80 ; pshufd $0x50,%xmm6,%xmm6 - DB 102,15,219,53,191,44,0,0 ; pand 0x2cbf(%rip),%xmm6 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,53,59,44,0,0 ; pand 0x2c3b(%rip),%xmm6 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,15,111,254 ; movdqa %xmm6,%xmm7 DB 102,65,15,213,248 ; pmullw %xmm8,%xmm7 DB 102,15,111,230 ; movdqa %xmm6,%xmm4 @@ -40807,10 +40731,10 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,65,15,253,248 ; paddw %xmm8,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,207 ; packuswb %xmm7,%xmm1 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,197 ; packuswb %xmm5,%xmm0 - DB 102,15,103,207 ; packuswb %xmm7,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -40823,10 +40747,10 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,239,0,0,0 ; jne 2a8ec <_sk_scale_u8_sse2_8bit+0x10b> + DB 15,133,239,0,0,0 ; jne 2a860 <_sk_scale_u8_sse2_8bit+0x10b> DB 243,67,15,126,36,24 ; movq (%r8,%r11,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 - DB 102,15,219,37,65,44,0,0 ; pand 0x2c41(%rip),%xmm4 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 102,15,219,37,189,43,0,0 ; pand 0x2bbd(%rip),%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,15,111,236 ; movdqa %xmm4,%xmm5 DB 102,65,15,105,232 ; punpckhwd %xmm8,%xmm5 @@ -40867,10 +40791,10 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,253,241 ; paddw %xmm1,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,103,231 ; packuswb %xmm7,%xmm4 - DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 72,173 ; lods %ds:(%rsi),%rax DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,15,111,205 ; movdqa %xmm5,%xmm1 @@ -40880,15 +40804,15 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,3,255,255,255 ; ja 2a807 <_sk_scale_u8_sse2_8bit+0x26> + DB 15,135,3,255,255,255 ; ja 2a77b <_sk_scale_u8_sse2_8bit+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,117,0,0,0 ; lea 0x75(%rip),%r10 # 2a984 <_sk_scale_u8_sse2_8bit+0x1a3> + DB 76,141,21,117,0,0,0 ; lea 0x75(%rip),%r10 # 2a8f8 <_sk_scale_u8_sse2_8bit+0x1a3> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax DB 102,15,110,224 ; movd %eax,%xmm4 - DB 233,225,254,255,255 ; jmpq 2a807 <_sk_scale_u8_sse2_8bit+0x26> + DB 233,225,254,255,255 ; jmpq 2a77b <_sk_scale_u8_sse2_8bit+0x26> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,196,224,2 ; pinsrw $0x2,%eax,%xmm4 @@ -40896,7 +40820,7 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,110,232 ; movd %eax,%xmm5 DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 243,15,16,229 ; movss %xmm5,%xmm4 - DB 233,188,254,255,255 ; jmpq 2a807 <_sk_scale_u8_sse2_8bit+0x26> + DB 233,188,254,255,255 ; jmpq 2a77b <_sk_scale_u8_sse2_8bit+0x26> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,196,224,6 ; pinsrw $0x6,%eax,%xmm4 @@ -40907,7 +40831,7 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,67,15,110,44,24 ; movd (%r8,%r11,1),%xmm5 DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 233,132,254,255,255 ; jmpq 2a807 <_sk_scale_u8_sse2_8bit+0x26> + DB 233,132,254,255,255 ; jmpq 2a77b <_sk_scale_u8_sse2_8bit+0x26> DB 144 ; nop DB 148 ; xchg %eax,%esp DB 255 ; (bad) @@ -40934,7 +40858,7 @@ PUBLIC _sk_lerp_1_float_sse2_8bit _sk_lerp_1_float_sse2_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,32 ; movss (%rax),%xmm4 - DB 243,15,89,37,42,41,0,0 ; mulss 0x292a(%rip),%xmm4 # 2d2d8 <_sk_overlay_sse2_8bit+0xd80> + DB 243,15,89,37,150,40,0,0 ; mulss 0x2896(%rip),%xmm4 # 2d1b8 <_sk_overlay_sse2_8bit+0xd09> DB 243,15,44,196 ; cvttss2si %xmm4,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 @@ -40944,34 +40868,34 @@ _sk_lerp_1_float_sse2_8bit LABEL PROC DB 102,15,111,248 ; movdqa %xmm0,%xmm7 DB 102,65,15,96,249 ; punpcklbw %xmm9,%xmm7 DB 102,65,15,104,193 ; punpckhbw %xmm9,%xmm0 - DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 - DB 102,69,15,96,217 ; punpcklbw %xmm9,%xmm11 + DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 + DB 102,69,15,96,209 ; punpcklbw %xmm9,%xmm10 DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 - DB 102,15,111,53,97,42,0,0 ; movdqa 0x2a61(%rip),%xmm6 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,65,15,219,240 ; pand %xmm8,%xmm6 - DB 102,15,111,230 ; movdqa %xmm6,%xmm4 + DB 102,15,111,45,221,41,0,0 ; movdqa 0x29dd(%rip),%xmm5 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 102,65,15,219,232 ; pand %xmm8,%xmm5 + DB 102,15,111,229 ; movdqa %xmm5,%xmm4 DB 102,15,213,225 ; pmullw %xmm1,%xmm4 - DB 102,68,15,111,214 ; movdqa %xmm6,%xmm10 - DB 102,69,15,213,211 ; pmullw %xmm11,%xmm10 - DB 102,15,111,238 ; movdqa %xmm6,%xmm5 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,15,213,247 ; pmullw %xmm7,%xmm6 - DB 102,15,253,247 ; paddw %xmm7,%xmm6 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,69,15,253,211 ; paddw %xmm11,%xmm10 + DB 102,68,15,111,221 ; movdqa %xmm5,%xmm11 + DB 102,69,15,213,218 ; pmullw %xmm10,%xmm11 + DB 102,15,111,245 ; movdqa %xmm5,%xmm6 + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 + DB 102,15,213,239 ; pmullw %xmm7,%xmm5 + DB 102,15,253,239 ; paddw %xmm7,%xmm5 + DB 102,15,253,240 ; paddw %xmm0,%xmm6 + DB 102,69,15,253,218 ; paddw %xmm10,%xmm11 DB 102,15,253,225 ; paddw %xmm1,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 + DB 102,68,15,103,220 ; packuswb %xmm4,%xmm11 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 - DB 102,68,15,103,212 ; packuswb %xmm4,%xmm10 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 102,15,118,255 ; pcmpeqd %xmm7,%xmm7 DB 102,65,15,239,248 ; pxor %xmm8,%xmm7 - DB 102,68,15,111,218 ; movdqa %xmm2,%xmm11 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,65,15,96,233 ; punpcklbw %xmm9,%xmm5 - DB 102,69,15,104,217 ; punpckhbw %xmm9,%xmm11 + DB 102,68,15,111,210 ; movdqa %xmm2,%xmm10 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,65,15,96,241 ; punpcklbw %xmm9,%xmm6 + DB 102,69,15,104,209 ; punpckhbw %xmm9,%xmm10 DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 DB 102,68,15,111,227 ; movdqa %xmm3,%xmm12 DB 102,69,15,96,225 ; punpcklbw %xmm9,%xmm12 @@ -40983,20 +40907,20 @@ _sk_lerp_1_float_sse2_8bit LABEL PROC DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,65,15,213,204 ; pmullw %xmm12,%xmm1 - DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 - DB 102,15,213,197 ; pmullw %xmm5,%xmm0 - DB 102,15,253,197 ; paddw %xmm5,%xmm0 - DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 + DB 102,65,15,213,250 ; pmullw %xmm10,%xmm7 + DB 102,15,213,198 ; pmullw %xmm6,%xmm0 + DB 102,15,253,198 ; paddw %xmm6,%xmm0 + DB 102,65,15,253,250 ; paddw %xmm10,%xmm7 DB 102,65,15,253,204 ; paddw %xmm12,%xmm1 DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,204 ; packuswb %xmm4,%xmm1 + DB 102,65,15,252,203 ; paddb %xmm11,%xmm1 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,204 ; packuswb %xmm4,%xmm1 - DB 102,15,252,198 ; paddb %xmm6,%xmm0 - DB 102,65,15,252,202 ; paddb %xmm10,%xmm1 + DB 102,15,252,197 ; paddb %xmm5,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41009,124 +40933,124 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 76,3,0 ; add (%rax),%r8 DB 76,99,218 ; movslq %edx,%r11 DB 72,133,255 ; test %rdi,%rdi - DB 15,133,141,1,0,0 ; jne 2ac7f <_sk_lerp_u8_sse2_8bit+0x1a9> - DB 243,67,15,126,44,24 ; movq (%r8,%r11,1),%xmm5 - DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 - DB 102,15,219,45,76,41,0,0 ; pand 0x294c(%rip),%xmm5 # 2d450 <_sk_overlay_sse2_8bit+0xef8> + DB 15,133,141,1,0,0 ; jne 2abf3 <_sk_lerp_u8_sse2_8bit+0x1a9> + DB 243,67,15,126,36,24 ; movq (%r8,%r11,1),%xmm4 + DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 + DB 102,15,219,37,200,40,0,0 ; pand 0x28c8(%rip),%xmm4 # 2d340 <_sk_overlay_sse2_8bit+0xe91> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,15,111,229 ; movdqa %xmm5,%xmm4 - DB 102,65,15,105,224 ; punpckhwd %xmm8,%xmm4 - DB 102,65,15,97,232 ; punpcklwd %xmm8,%xmm5 - DB 102,15,114,245,24 ; pslld $0x18,%xmm5 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,65,15,105,232 ; punpckhwd %xmm8,%xmm5 + DB 102,65,15,97,224 ; punpcklwd %xmm8,%xmm4 DB 102,15,114,244,24 ; pslld $0x18,%xmm4 - DB 242,15,112,228,231 ; pshuflw $0xe7,%xmm4,%xmm4 - DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 - DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 - DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,244,95 ; pshufhw $0x5f,%xmm4,%xmm6 + DB 102,15,114,245,24 ; pslld $0x18,%xmm5 DB 242,15,112,237,231 ; pshuflw $0xe7,%xmm5,%xmm5 DB 243,15,112,237,231 ; pshufhw $0xe7,%xmm5,%xmm5 DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 DB 243,15,112,253,95 ; pshufhw $0x5f,%xmm5,%xmm7 - DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 - DB 102,69,15,96,200 ; punpcklbw %xmm8,%xmm9 - DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 - DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 + DB 242,15,112,228,231 ; pshuflw $0xe7,%xmm4,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,244,95 ; pshufhw $0x5f,%xmm4,%xmm6 + DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10 DB 102,69,15,96,208 ; punpcklbw %xmm8,%xmm10 - DB 102,65,15,104,200 ; punpckhbw %xmm8,%xmm1 - DB 102,68,15,111,223 ; movdqa %xmm7,%xmm11 + DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 + DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 DB 102,69,15,96,216 ; punpcklbw %xmm8,%xmm11 - DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,111,230 ; movdqa %xmm6,%xmm12 - DB 102,69,15,96,224 ; punpcklbw %xmm8,%xmm12 + DB 102,65,15,104,200 ; punpckhbw %xmm8,%xmm1 + DB 102,68,15,111,206 ; movdqa %xmm6,%xmm9 + DB 102,69,15,96,200 ; punpcklbw %xmm8,%xmm9 DB 102,15,111,230 ; movdqa %xmm6,%xmm4 DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 - DB 102,15,213,225 ; pmullw %xmm1,%xmm4 - DB 102,69,15,213,226 ; pmullw %xmm10,%xmm12 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,69,15,213,217 ; pmullw %xmm9,%xmm11 - DB 102,69,15,253,217 ; paddw %xmm9,%xmm11 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,69,15,253,226 ; paddw %xmm10,%xmm12 - DB 102,15,253,225 ; paddw %xmm1,%xmm4 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 + DB 102,68,15,111,231 ; movdqa %xmm7,%xmm12 + DB 102,69,15,96,224 ; punpcklbw %xmm8,%xmm12 + DB 102,15,111,239 ; movdqa %xmm7,%xmm5 + DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 + DB 102,15,213,233 ; pmullw %xmm1,%xmm5 + DB 102,69,15,213,227 ; pmullw %xmm11,%xmm12 + DB 102,15,213,224 ; pmullw %xmm0,%xmm4 + DB 102,69,15,213,202 ; pmullw %xmm10,%xmm9 + DB 102,69,15,253,202 ; paddw %xmm10,%xmm9 + DB 102,15,253,224 ; paddw %xmm0,%xmm4 + DB 102,69,15,253,227 ; paddw %xmm11,%xmm12 + DB 102,15,253,233 ; paddw %xmm1,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 - DB 102,68,15,103,221 ; packuswb %xmm5,%xmm11 - DB 102,68,15,103,228 ; packuswb %xmm4,%xmm12 + DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 + DB 102,68,15,103,229 ; packuswb %xmm5,%xmm12 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 + DB 102,68,15,103,204 ; packuswb %xmm4,%xmm9 DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 - DB 102,15,239,240 ; pxor %xmm0,%xmm6 DB 102,15,239,248 ; pxor %xmm0,%xmm7 + DB 102,15,239,240 ; pxor %xmm0,%xmm6 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,65,15,96,224 ; punpcklbw %xmm8,%xmm4 DB 102,15,111,234 ; movdqa %xmm2,%xmm5 DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9 - DB 102,69,15,96,200 ; punpcklbw %xmm8,%xmm9 DB 102,68,15,111,211 ; movdqa %xmm3,%xmm10 - DB 102,69,15,104,208 ; punpckhbw %xmm8,%xmm10 - DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,69,15,96,208 ; punpcklbw %xmm8,%xmm10 + DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 + DB 102,69,15,104,216 ; punpckhbw %xmm8,%xmm11 + DB 102,15,111,198 ; movdqa %xmm6,%xmm0 DB 102,65,15,96,192 ; punpcklbw %xmm8,%xmm0 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,65,15,96,200 ; punpcklbw %xmm8,%xmm1 DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 - DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 - DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,15,111,207 ; movdqa %xmm7,%xmm1 + DB 102,65,15,96,200 ; punpcklbw %xmm8,%xmm1 + DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 + DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 + DB 102,65,15,213,202 ; pmullw %xmm10,%xmm1 + DB 102,15,213,245 ; pmullw %xmm5,%xmm6 DB 102,15,213,196 ; pmullw %xmm4,%xmm0 DB 102,15,253,196 ; paddw %xmm4,%xmm0 - DB 102,15,253,253 ; paddw %xmm5,%xmm7 - DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 - DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,253,245 ; paddw %xmm5,%xmm6 + DB 102,65,15,253,202 ; paddw %xmm10,%xmm1 + DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 - DB 102,65,15,252,195 ; paddb %xmm11,%xmm0 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,207 ; packuswb %xmm7,%xmm1 DB 102,65,15,252,204 ; paddb %xmm12,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,198 ; packuswb %xmm6,%xmm0 + DB 102,65,15,252,193 ; paddb %xmm9,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax DB 65,137,249 ; mov %edi,%r9d DB 65,128,225,7 ; and $0x7,%r9b - DB 102,15,239,237 ; pxor %xmm5,%xmm5 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,101,254,255,255 ; ja 2aafc <_sk_lerp_u8_sse2_8bit+0x26> + DB 15,135,101,254,255,255 ; ja 2aa70 <_sk_lerp_u8_sse2_8bit+0x26> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,118,0,0,0 ; lea 0x76(%rip),%r10 # 2ad18 <_sk_lerp_u8_sse2_8bit+0x242> + DB 76,141,21,118,0,0,0 ; lea 0x76(%rip),%r10 # 2ac8c <_sk_lerp_u8_sse2_8bit+0x242> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax DB 67,15,182,4,24 ; movzbl (%r8,%r11,1),%eax - DB 102,15,110,232 ; movd %eax,%xmm5 - DB 233,67,254,255,255 ; jmpq 2aafc <_sk_lerp_u8_sse2_8bit+0x26> + DB 102,15,110,224 ; movd %eax,%xmm4 + DB 233,67,254,255,255 ; jmpq 2aa70 <_sk_lerp_u8_sse2_8bit+0x26> DB 67,15,182,68,24,2 ; movzbl 0x2(%r8,%r11,1),%eax - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,196,232,2 ; pinsrw $0x2,%eax,%xmm5 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,196,224,2 ; pinsrw $0x2,%eax,%xmm4 DB 67,15,183,4,24 ; movzwl (%r8,%r11,1),%eax - DB 102,15,110,224 ; movd %eax,%xmm4 - DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 - DB 243,15,16,236 ; movss %xmm4,%xmm5 - DB 233,30,254,255,255 ; jmpq 2aafc <_sk_lerp_u8_sse2_8bit+0x26> + DB 102,15,110,232 ; movd %eax,%xmm5 + DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 + DB 243,15,16,229 ; movss %xmm5,%xmm4 + DB 233,30,254,255,255 ; jmpq 2aa70 <_sk_lerp_u8_sse2_8bit+0x26> DB 67,15,182,68,24,6 ; movzbl 0x6(%r8,%r11,1),%eax - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,196,232,6 ; pinsrw $0x6,%eax,%xmm5 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,196,224,6 ; pinsrw $0x6,%eax,%xmm4 DB 67,15,182,68,24,5 ; movzbl 0x5(%r8,%r11,1),%eax - DB 102,15,196,232,5 ; pinsrw $0x5,%eax,%xmm5 + DB 102,15,196,224,5 ; pinsrw $0x5,%eax,%xmm4 DB 67,15,182,68,24,4 ; movzbl 0x4(%r8,%r11,1),%eax - DB 102,15,196,232,4 ; pinsrw $0x4,%eax,%xmm5 - DB 102,67,15,110,36,24 ; movd (%r8,%r11,1),%xmm4 - DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 - DB 242,15,16,236 ; movsd %xmm4,%xmm5 - DB 233,230,253,255,255 ; jmpq 2aafc <_sk_lerp_u8_sse2_8bit+0x26> + DB 102,15,196,224,4 ; pinsrw $0x4,%eax,%xmm4 + DB 102,67,15,110,44,24 ; movd (%r8,%r11,1),%xmm5 + DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 + DB 242,15,16,229 ; movsd %xmm5,%xmm4 + DB 233,230,253,255,255 ; jmpq 2aa70 <_sk_lerp_u8_sse2_8bit+0x26> DB 102,144 ; xchg %ax,%ax DB 147 ; xchg %eax,%ebx DB 255 ; (bad) @@ -41135,7 +41059,7 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 2ad25 <_sk_lerp_u8_sse2_8bit+0x24f> + DB 235,255 ; jmp 2ac99 <_sk_lerp_u8_sse2_8bit+0x24f> DB 255 ; (bad) DB 255,224 ; jmpq *%rax DB 255 ; (bad) @@ -41165,7 +41089,7 @@ _sk_move_dst_src_sse2_8bit LABEL PROC PUBLIC _sk_black_color_sse2_8bit _sk_black_color_sse2_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,5,63,46,0,0 ; movaps 0x2e3f(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x1638> + DB 15,40,5,219,45,0,0 ; movaps 0x2ddb(%rip),%xmm0 # 2daa0 <_sk_overlay_sse2_8bit+0x15f1> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 255,224 ; jmpq *%rax @@ -41191,41 +41115,41 @@ _sk_srcatop_sse2_8bit LABEL PROC DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 DB 243,68,15,112,220,95 ; pshufhw $0x5f,%xmm4,%xmm11 - DB 242,15,112,226,231 ; pshuflw $0xe7,%xmm2,%xmm4 - DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 - DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 - DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 + DB 242,15,112,234,231 ; pshuflw $0xe7,%xmm2,%xmm5 + DB 243,15,112,237,231 ; pshufhw $0xe7,%xmm5,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 + DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 + DB 243,15,112,245,95 ; pshufhw $0x5f,%xmm5,%xmm6 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 DB 242,68,15,112,208,231 ; pshuflw $0xe7,%xmm0,%xmm10 DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,69,15,104,193 ; punpckhbw %xmm9,%xmm8 - DB 102,15,111,249 ; movdqa %xmm1,%xmm7 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 242,68,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm12 DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1 - DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,65,15,96,241 ; punpcklbw %xmm9,%xmm6 DB 102,65,15,104,225 ; punpckhbw %xmm9,%xmm4 + DB 102,15,111,254 ; movdqa %xmm6,%xmm7 + DB 102,65,15,96,249 ; punpcklbw %xmm9,%xmm7 + DB 102,65,15,104,241 ; punpckhbw %xmm9,%xmm6 DB 102,65,15,111,235 ; movdqa %xmm11,%xmm5 DB 102,65,15,96,233 ; punpcklbw %xmm9,%xmm5 DB 102,69,15,104,217 ; punpckhbw %xmm9,%xmm11 - DB 102,68,15,213,223 ; pmullw %xmm7,%xmm11 + DB 102,68,15,213,220 ; pmullw %xmm4,%xmm11 DB 102,15,213,233 ; pmullw %xmm1,%xmm5 - DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 - DB 102,15,213,240 ; pmullw %xmm0,%xmm6 - DB 102,15,253,240 ; paddw %xmm0,%xmm6 - DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,65,15,213,240 ; pmullw %xmm8,%xmm6 + DB 102,15,213,248 ; pmullw %xmm0,%xmm7 + DB 102,15,253,248 ; paddw %xmm0,%xmm7 + DB 102,65,15,253,240 ; paddw %xmm8,%xmm6 DB 102,15,253,233 ; paddw %xmm1,%xmm5 - DB 102,68,15,253,223 ; paddw %xmm7,%xmm11 + DB 102,68,15,253,220 ; paddw %xmm4,%xmm11 DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,244 ; packuswb %xmm4,%xmm6 DB 102,65,15,103,235 ; packuswb %xmm11,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,254 ; packuswb %xmm6,%xmm7 DB 243,65,15,112,194,231 ; pshufhw $0xe7,%xmm10,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 @@ -41235,9 +41159,9 @@ _sk_srcatop_sse2_8bit LABEL PROC DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,248,95 ; pshufhw $0x5f,%xmm0,%xmm7 + DB 243,15,112,240,95 ; pshufhw $0x5f,%xmm0,%xmm6 DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 - DB 102,15,239,248 ; pxor %xmm0,%xmm7 + DB 102,15,239,240 ; pxor %xmm0,%xmm6 DB 102,15,239,224 ; pxor %xmm0,%xmm4 DB 102,68,15,111,194 ; movdqa %xmm2,%xmm8 DB 102,68,15,111,210 ; movdqa %xmm2,%xmm10 @@ -41250,25 +41174,25 @@ _sk_srcatop_sse2_8bit LABEL PROC DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 DB 102,65,15,104,225 ; punpckhbw %xmm9,%xmm4 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1 - DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 - DB 102,65,15,213,251 ; pmullw %xmm11,%xmm7 + DB 102,65,15,104,241 ; punpckhbw %xmm9,%xmm6 + DB 102,65,15,213,243 ; pmullw %xmm11,%xmm6 DB 102,65,15,213,204 ; pmullw %xmm12,%xmm1 DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 DB 102,65,15,213,194 ; pmullw %xmm10,%xmm0 DB 102,65,15,253,194 ; paddw %xmm10,%xmm0 DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 DB 102,65,15,253,204 ; paddw %xmm12,%xmm1 - DB 102,65,15,253,251 ; paddw %xmm11,%xmm7 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,253,243 ; paddw %xmm11,%xmm6 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,206 ; packuswb %xmm6,%xmm1 + DB 102,15,252,205 ; paddb %xmm5,%xmm1 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,196 ; packuswb %xmm4,%xmm0 - DB 102,15,103,207 ; packuswb %xmm7,%xmm1 - DB 102,15,252,198 ; paddb %xmm6,%xmm0 - DB 102,15,252,205 ; paddb %xmm5,%xmm1 + DB 102,15,252,199 ; paddb %xmm7,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41279,87 +41203,87 @@ _sk_dstatop_sse2_8bit LABEL PROC DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 - DB 242,15,112,232,231 ; pshuflw $0xe7,%xmm0,%xmm5 - DB 243,15,112,237,231 ; pshufhw $0xe7,%xmm5,%xmm5 - DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 - DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 - DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 - DB 243,15,112,237,95 ; pshufhw $0x5f,%xmm5,%xmm5 - DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 - DB 102,68,15,111,210 ; movdqa %xmm2,%xmm10 + DB 243,68,15,112,236,95 ; pshufhw $0x5f,%xmm4,%xmm13 + DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,236,95 ; pshufhw $0x5f,%xmm4,%xmm5 + DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 + DB 102,68,15,111,194 ; movdqa %xmm2,%xmm8 DB 242,68,15,112,218,231 ; pshuflw $0xe7,%xmm2,%xmm11 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,65,15,96,248 ; punpcklbw %xmm8,%xmm7 - DB 102,69,15,104,208 ; punpckhbw %xmm8,%xmm10 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,68,15,111,205 ; movdqa %xmm5,%xmm9 - DB 102,69,15,96,200 ; punpcklbw %xmm8,%xmm9 - DB 102,68,15,213,207 ; pmullw %xmm7,%xmm9 - DB 102,68,15,253,207 ; paddw %xmm7,%xmm9 - DB 242,68,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm12 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,65,15,96,225 ; punpcklbw %xmm9,%xmm4 + DB 102,69,15,104,193 ; punpckhbw %xmm9,%xmm8 DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,65,15,96,248 ; punpcklbw %xmm8,%xmm7 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 - DB 102,65,15,104,232 ; punpckhbw %xmm8,%xmm5 - DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 - DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 - DB 102,68,15,111,212 ; movdqa %xmm4,%xmm10 - DB 102,69,15,96,208 ; punpcklbw %xmm8,%xmm10 - DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 + DB 102,68,15,111,213 ; movdqa %xmm5,%xmm10 + DB 102,69,15,96,209 ; punpcklbw %xmm9,%xmm10 + DB 102,68,15,213,212 ; pmullw %xmm4,%xmm10 + DB 102,68,15,253,212 ; paddw %xmm4,%xmm10 + DB 242,68,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm12 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,65,15,96,241 ; punpcklbw %xmm9,%xmm6 + DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 + DB 102,65,15,104,233 ; punpckhbw %xmm9,%xmm5 + DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 + DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 + DB 102,65,15,111,229 ; movdqa %xmm13,%xmm4 + DB 102,65,15,96,225 ; punpcklbw %xmm9,%xmm4 + DB 102,69,15,104,233 ; punpckhbw %xmm9,%xmm13 + DB 102,68,15,213,239 ; pmullw %xmm7,%xmm13 DB 102,15,213,230 ; pmullw %xmm6,%xmm4 - DB 102,68,15,213,215 ; pmullw %xmm7,%xmm10 - DB 102,68,15,253,215 ; paddw %xmm7,%xmm10 DB 102,15,253,230 ; paddw %xmm6,%xmm4 + DB 102,68,15,253,239 ; paddw %xmm7,%xmm13 + DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 + DB 102,65,15,103,229 ; packuswb %xmm13,%xmm4 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,65,15,113,209,8 ; psrlw $0x8,%xmm9 - DB 102,68,15,103,205 ; packuswb %xmm5,%xmm9 - DB 102,68,15,103,212 ; packuswb %xmm4,%xmm10 - DB 243,65,15,112,227,231 ; pshufhw $0xe7,%xmm11,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 - DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 - DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,252,95 ; pshufhw $0x5f,%xmm4,%xmm7 - DB 243,65,15,112,228,231 ; pshufhw $0xe7,%xmm12,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 - DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 - DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,244,95 ; pshufhw $0x5f,%xmm4,%xmm6 - DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 - DB 102,15,239,244 ; pxor %xmm4,%xmm6 - DB 102,15,239,252 ; pxor %xmm4,%xmm7 - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,65,15,96,224 ; punpcklbw %xmm8,%xmm4 - DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 + DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 + DB 102,68,15,103,213 ; packuswb %xmm5,%xmm10 + DB 243,65,15,112,235,231 ; pshufhw $0xe7,%xmm11,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 + DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 + DB 243,15,112,253,95 ; pshufhw $0x5f,%xmm5,%xmm7 + DB 243,65,15,112,236,231 ; pshufhw $0xe7,%xmm12,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 + DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 + DB 243,15,112,245,95 ; pshufhw $0x5f,%xmm5,%xmm6 + DB 102,15,118,237 ; pcmpeqd %xmm5,%xmm5 + DB 102,15,239,245 ; pxor %xmm5,%xmm6 + DB 102,15,239,253 ; pxor %xmm5,%xmm7 + DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 + DB 102,69,15,96,193 ; punpcklbw %xmm9,%xmm8 + DB 102,65,15,104,193 ; punpckhbw %xmm9,%xmm0 DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,65,15,96,232 ; punpcklbw %xmm8,%xmm5 - DB 102,65,15,104,200 ; punpckhbw %xmm8,%xmm1 + DB 102,65,15,96,233 ; punpcklbw %xmm9,%xmm5 + DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 DB 102,68,15,111,223 ; movdqa %xmm7,%xmm11 - DB 102,69,15,96,216 ; punpcklbw %xmm8,%xmm11 - DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 + DB 102,69,15,96,217 ; punpcklbw %xmm9,%xmm11 + DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 DB 102,68,15,111,230 ; movdqa %xmm6,%xmm12 - DB 102,69,15,96,224 ; punpcklbw %xmm8,%xmm12 - DB 102,65,15,104,240 ; punpckhbw %xmm8,%xmm6 + DB 102,69,15,96,225 ; punpcklbw %xmm9,%xmm12 + DB 102,65,15,104,241 ; punpckhbw %xmm9,%xmm6 DB 102,15,213,241 ; pmullw %xmm1,%xmm6 DB 102,68,15,213,229 ; pmullw %xmm5,%xmm12 DB 102,15,213,248 ; pmullw %xmm0,%xmm7 - DB 102,68,15,213,220 ; pmullw %xmm4,%xmm11 - DB 102,65,15,253,227 ; paddw %xmm11,%xmm4 + DB 102,69,15,213,216 ; pmullw %xmm8,%xmm11 + DB 102,69,15,253,195 ; paddw %xmm11,%xmm8 DB 102,15,253,248 ; paddw %xmm0,%xmm7 DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 DB 102,15,253,241 ; paddw %xmm1,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,231 ; packuswb %xmm7,%xmm4 DB 102,15,103,238 ; packuswb %xmm6,%xmm5 - DB 102,65,15,252,225 ; paddb %xmm9,%xmm4 - DB 102,65,15,252,234 ; paddb %xmm10,%xmm5 + DB 102,15,252,236 ; paddb %xmm4,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 + DB 102,68,15,103,199 ; packuswb %xmm7,%xmm8 + DB 102,69,15,252,194 ; paddb %xmm10,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax @@ -41401,10 +41325,10 @@ _sk_srcin_sse2_8bit LABEL PROC DB 102,15,253,241 ; paddw %xmm1,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,230 ; packuswb %xmm6,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,199 ; packuswb %xmm7,%xmm0 - DB 102,15,103,230 ; packuswb %xmm6,%xmm4 DB 72,173 ; lods %ds:(%rsi),%rax DB 102,15,111,204 ; movdqa %xmm4,%xmm1 DB 255,224 ; jmpq *%rax @@ -41448,10 +41372,10 @@ _sk_dstin_sse2_8bit LABEL PROC DB 102,65,15,253,225 ; paddw %xmm9,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,204 ; packuswb %xmm4,%xmm1 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,197 ; packuswb %xmm5,%xmm0 - DB 102,15,103,204 ; packuswb %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41496,10 +41420,10 @@ _sk_srcout_sse2_8bit LABEL PROC DB 102,15,253,249 ; paddw %xmm1,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,239 ; packuswb %xmm7,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,198 ; packuswb %xmm6,%xmm0 - DB 102,15,103,239 ; packuswb %xmm7,%xmm5 DB 72,173 ; lods %ds:(%rsi),%rax DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 255,224 ; jmpq *%rax @@ -41546,10 +41470,10 @@ _sk_dstout_sse2_8bit LABEL PROC DB 102,65,15,253,233 ; paddw %xmm9,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,205 ; packuswb %xmm5,%xmm1 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,196 ; packuswb %xmm4,%xmm0 - DB 102,15,103,205 ; packuswb %xmm5,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41570,36 +41494,36 @@ _sk_srcover_sse2_8bit LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 DB 102,15,252,194 ; paddb %xmm2,%xmm0 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,65,15,96,240 ; punpcklbw %xmm8,%xmm6 + DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 + DB 102,69,15,96,224 ; punpcklbw %xmm8,%xmm12 DB 102,69,15,104,200 ; punpckhbw %xmm8,%xmm9 - DB 102,68,15,111,227 ; movdqa %xmm3,%xmm12 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 DB 102,15,252,203 ; paddb %xmm3,%xmm1 DB 102,15,111,227 ; movdqa %xmm3,%xmm4 DB 102,65,15,96,224 ; punpcklbw %xmm8,%xmm4 - DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 + DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 DB 102,65,15,111,235 ; movdqa %xmm11,%xmm5 DB 102,65,15,96,232 ; punpcklbw %xmm8,%xmm5 DB 102,69,15,104,216 ; punpckhbw %xmm8,%xmm11 - DB 102,65,15,111,250 ; movdqa %xmm10,%xmm7 - DB 102,65,15,96,248 ; punpcklbw %xmm8,%xmm7 + DB 102,65,15,111,242 ; movdqa %xmm10,%xmm6 + DB 102,65,15,96,240 ; punpcklbw %xmm8,%xmm6 DB 102,69,15,104,208 ; punpckhbw %xmm8,%xmm10 - DB 102,69,15,213,212 ; pmullw %xmm12,%xmm10 - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,68,15,213,215 ; pmullw %xmm7,%xmm10 + DB 102,15,213,244 ; pmullw %xmm4,%xmm6 DB 102,69,15,213,217 ; pmullw %xmm9,%xmm11 - DB 102,15,213,238 ; pmullw %xmm6,%xmm5 - DB 102,15,253,238 ; paddw %xmm6,%xmm5 + DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 + DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 DB 102,69,15,253,217 ; paddw %xmm9,%xmm11 - DB 102,15,253,252 ; paddw %xmm4,%xmm7 - DB 102,69,15,253,212 ; paddw %xmm12,%xmm10 + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,68,15,253,215 ; paddw %xmm7,%xmm10 DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,65,15,103,242 ; packuswb %xmm10,%xmm6 DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,65,15,103,235 ; packuswb %xmm11,%xmm5 - DB 102,65,15,103,250 ; packuswb %xmm10,%xmm7 + DB 102,15,248,206 ; psubb %xmm6,%xmm1 DB 102,15,248,197 ; psubb %xmm5,%xmm0 - DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41642,14 +41566,14 @@ _sk_dstover_sse2_8bit LABEL PROC DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,103,245 ; packuswb %xmm5,%xmm6 - DB 102,15,103,252 ; packuswb %xmm4,%xmm7 DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 102,15,252,194 ; paddb %xmm2,%xmm0 DB 102,15,248,198 ; psubb %xmm6,%xmm0 - DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41662,28 +41586,28 @@ _sk_modulate_sse2_8bit LABEL PROC DB 102,69,15,104,202 ; punpckhbw %xmm10,%xmm9 DB 102,65,15,96,202 ; punpcklbw %xmm10,%xmm1 DB 102,69,15,104,194 ; punpckhbw %xmm10,%xmm8 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,65,15,96,226 ; punpcklbw %xmm10,%xmm4 - DB 102,65,15,104,250 ; punpckhbw %xmm10,%xmm7 + DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,65,15,96,242 ; punpcklbw %xmm10,%xmm6 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,65,15,96,250 ; punpcklbw %xmm10,%xmm7 DB 102,65,15,104,234 ; punpckhbw %xmm10,%xmm5 DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 - DB 102,15,213,241 ; pmullw %xmm1,%xmm6 - DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 + DB 102,15,213,249 ; pmullw %xmm1,%xmm7 + DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 DB 102,15,213,224 ; pmullw %xmm0,%xmm4 DB 102,15,253,196 ; paddw %xmm4,%xmm0 - DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 - DB 102,15,253,206 ; paddw %xmm6,%xmm1 + DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 + DB 102,15,253,207 ; paddw %xmm7,%xmm1 DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 102,15,103,205 ; packuswb %xmm5,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,198 ; packuswb %xmm6,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41692,115 +41616,114 @@ _sk_multiply_sse2_8bit LABEL PROC DB 72,131,236,24 ; sub $0x18,%rsp DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 DB 102,15,111,218 ; movdqa %xmm2,%xmm3 - DB 242,15,112,211,231 ; pshuflw $0xe7,%xmm3,%xmm2 - DB 243,15,112,210,231 ; pshufhw $0xe7,%xmm2,%xmm2 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 - DB 242,15,112,210,95 ; pshuflw $0x5f,%xmm2,%xmm2 - DB 243,15,112,250,95 ; pshufhw $0x5f,%xmm2,%xmm7 - DB 242,65,15,112,214,231 ; pshuflw $0xe7,%xmm14,%xmm2 - DB 243,15,112,210,231 ; pshufhw $0xe7,%xmm2,%xmm2 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 - DB 242,15,112,210,95 ; pshuflw $0x5f,%xmm2,%xmm2 - DB 243,15,112,234,95 ; pshufhw $0x5f,%xmm2,%xmm5 + DB 242,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,252,95 ; pshufhw $0x5f,%xmm4,%xmm7 + DB 242,65,15,112,230,231 ; pshuflw $0xe7,%xmm14,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,244,95 ; pshufhw $0x5f,%xmm4,%xmm6 DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 - DB 102,65,15,239,232 ; pxor %xmm8,%xmm5 + DB 102,65,15,239,240 ; pxor %xmm8,%xmm6 DB 102,65,15,239,248 ; pxor %xmm8,%xmm7 DB 102,69,15,239,237 ; pxor %xmm13,%xmm13 DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 - DB 242,15,112,208,231 ; pshuflw $0xe7,%xmm0,%xmm2 + DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 DB 102,68,15,111,216 ; movdqa %xmm0,%xmm11 DB 102,69,15,96,221 ; punpcklbw %xmm13,%xmm11 DB 102,69,15,104,205 ; punpckhbw %xmm13,%xmm9 DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 - DB 242,15,112,241,231 ; pshuflw $0xe7,%xmm1,%xmm6 + DB 242,15,112,193,231 ; pshuflw $0xe7,%xmm1,%xmm0 DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 DB 102,69,15,96,229 ; punpcklbw %xmm13,%xmm12 DB 102,69,15,104,213 ; punpckhbw %xmm13,%xmm10 DB 102,68,15,111,255 ; movdqa %xmm7,%xmm15 DB 102,69,15,96,253 ; punpcklbw %xmm13,%xmm15 DB 102,65,15,104,253 ; punpckhbw %xmm13,%xmm7 - DB 102,15,111,229 ; movdqa %xmm5,%xmm4 - DB 102,65,15,96,229 ; punpcklbw %xmm13,%xmm4 - DB 102,65,15,104,237 ; punpckhbw %xmm13,%xmm5 - DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 - DB 102,65,15,213,228 ; pmullw %xmm12,%xmm4 + DB 102,15,111,238 ; movdqa %xmm6,%xmm5 + DB 102,65,15,96,237 ; punpcklbw %xmm13,%xmm5 + DB 102,65,15,104,245 ; punpckhbw %xmm13,%xmm6 + DB 102,65,15,213,242 ; pmullw %xmm10,%xmm6 + DB 102,65,15,213,236 ; pmullw %xmm12,%xmm5 DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 DB 102,69,15,213,251 ; pmullw %xmm11,%xmm15 DB 102,69,15,253,251 ; paddw %xmm11,%xmm15 DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 - DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 - DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 + DB 102,65,15,253,236 ; paddw %xmm12,%xmm5 + DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,65,15,113,215,8 ; psrlw $0x8,%xmm15 DB 102,68,15,103,255 ; packuswb %xmm7,%xmm15 - DB 102,15,103,229 ; packuswb %xmm5,%xmm4 - DB 243,15,112,194,231 ; pshufhw $0xe7,%xmm2,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,192,95 ; pshufhw $0x5f,%xmm0,%xmm0 - DB 243,15,112,206,231 ; pshufhw $0xe7,%xmm6,%xmm1 + DB 243,15,112,204,231 ; pshufhw $0xe7,%xmm4,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,241,95 ; pshufhw $0x5f,%xmm1,%xmm6 - DB 102,65,15,239,240 ; pxor %xmm8,%xmm6 - DB 102,65,15,239,192 ; pxor %xmm8,%xmm0 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,65,15,96,237 ; punpcklbw %xmm13,%xmm5 - DB 102,65,15,104,213 ; punpckhbw %xmm13,%xmm2 - DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1 - DB 102,15,127,12,36 ; movdqa %xmm1,(%rsp) + DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 + DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 + DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 + DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 + DB 102,65,15,239,200 ; pxor %xmm8,%xmm1 + DB 102,65,15,239,224 ; pxor %xmm8,%xmm4 + DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 + DB 102,65,15,96,213 ; punpcklbw %xmm13,%xmm2 + DB 102,69,15,104,197 ; punpckhbw %xmm13,%xmm8 + DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 + DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,65,15,96,245 ; punpcklbw %xmm13,%xmm6 + DB 102,69,15,104,245 ; punpckhbw %xmm13,%xmm14 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,65,15,96,197 ; punpcklbw %xmm13,%xmm0 + DB 102,65,15,104,229 ; punpckhbw %xmm13,%xmm4 DB 102,15,111,249 ; movdqa %xmm1,%xmm7 DB 102,65,15,96,253 ; punpcklbw %xmm13,%xmm7 - DB 102,69,15,104,245 ; punpckhbw %xmm13,%xmm14 - DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 - DB 102,69,15,96,197 ; punpcklbw %xmm13,%xmm8 - DB 102,65,15,104,197 ; punpckhbw %xmm13,%xmm0 - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,65,15,96,205 ; punpcklbw %xmm13,%xmm1 - DB 102,65,15,104,245 ; punpckhbw %xmm13,%xmm6 + DB 102,65,15,104,205 ; punpckhbw %xmm13,%xmm1 + DB 102,65,15,213,206 ; pmullw %xmm14,%xmm1 + DB 102,15,213,254 ; pmullw %xmm6,%xmm7 + DB 102,15,253,254 ; paddw %xmm6,%xmm7 + DB 102,65,15,253,206 ; paddw %xmm14,%xmm1 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,249 ; packuswb %xmm1,%xmm7 + DB 102,15,252,253 ; paddb %xmm5,%xmm7 + DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 DB 102,15,213,194 ; pmullw %xmm2,%xmm0 - DB 102,68,15,213,197 ; pmullw %xmm5,%xmm8 - DB 102,68,15,253,197 ; paddw %xmm5,%xmm8 DB 102,15,253,194 ; paddw %xmm2,%xmm0 + DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 - DB 102,68,15,103,192 ; packuswb %xmm0,%xmm8 - DB 102,65,15,213,246 ; pmullw %xmm14,%xmm6 - DB 102,15,213,207 ; pmullw %xmm7,%xmm1 - DB 102,15,253,207 ; paddw %xmm7,%xmm1 - DB 102,65,15,253,246 ; paddw %xmm14,%xmm6 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 - DB 102,15,252,204 ; paddb %xmm4,%xmm1 - DB 102,69,15,252,199 ; paddb %xmm15,%xmm8 - DB 102,65,15,213,235 ; pmullw %xmm11,%xmm5 - DB 102,65,15,253,235 ; paddw %xmm11,%xmm5 - DB 102,65,15,213,209 ; pmullw %xmm9,%xmm2 - DB 102,65,15,253,209 ; paddw %xmm9,%xmm2 - DB 102,65,15,213,252 ; pmullw %xmm12,%xmm7 - DB 102,65,15,253,252 ; paddw %xmm12,%xmm7 + DB 102,15,103,196 ; packuswb %xmm4,%xmm0 + DB 102,65,15,252,199 ; paddb %xmm15,%xmm0 + DB 102,65,15,213,211 ; pmullw %xmm11,%xmm2 + DB 102,65,15,253,211 ; paddw %xmm11,%xmm2 + DB 102,69,15,213,193 ; pmullw %xmm9,%xmm8 + DB 102,69,15,253,193 ; paddw %xmm9,%xmm8 + DB 102,65,15,213,244 ; pmullw %xmm12,%xmm6 + DB 102,65,15,253,244 ; paddw %xmm12,%xmm6 DB 102,69,15,213,242 ; pmullw %xmm10,%xmm14 DB 102,69,15,253,242 ; paddw %xmm10,%xmm14 - DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,234 ; packuswb %xmm2,%xmm5 DB 102,65,15,113,214,8 ; psrlw $0x8,%xmm14 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,65,15,103,254 ; packuswb %xmm14,%xmm7 - DB 102,68,15,252,197 ; paddb %xmm5,%xmm8 - DB 102,15,252,207 ; paddb %xmm7,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,65,15,103,246 ; packuswb %xmm14,%xmm6 + DB 102,15,252,254 ; paddb %xmm6,%xmm7 + DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,65,15,103,208 ; packuswb %xmm8,%xmm2 + DB 102,15,252,194 ; paddb %xmm2,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 102,15,111,211 ; movdqa %xmm3,%xmm2 DB 15,40,28,36 ; movaps (%rsp),%xmm3 - DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 + DB 102,15,111,207 ; movdqa %xmm7,%xmm1 DB 72,131,196,24 ; add $0x18,%rsp DB 255,224 ; jmpq *%rax @@ -41836,10 +41759,10 @@ _sk_screen_sse2_8bit LABEL PROC DB 102,65,15,253,242 ; paddw %xmm10,%xmm6 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,206 ; packuswb %xmm6,%xmm1 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,236 ; packuswb %xmm4,%xmm5 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 DB 102,15,252,197 ; paddb %xmm5,%xmm0 DB 102,65,15,252,200 ; paddb %xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -41853,44 +41776,44 @@ _sk_xor__sse2_8bit LABEL PROC DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 DB 243,68,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm12 - DB 242,15,112,235,231 ; pshuflw $0xe7,%xmm3,%xmm5 - DB 243,15,112,237,231 ; pshufhw $0xe7,%xmm5,%xmm5 - DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 - DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 - DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 - DB 243,15,112,237,95 ; pshufhw $0x5f,%xmm5,%xmm5 + DB 242,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 DB 102,69,15,118,192 ; pcmpeqd %xmm8,%xmm8 - DB 102,65,15,239,232 ; pxor %xmm8,%xmm5 + DB 102,65,15,239,224 ; pxor %xmm8,%xmm4 DB 102,69,15,239,224 ; pxor %xmm8,%xmm12 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 DB 242,68,15,112,216,231 ; pshuflw $0xe7,%xmm0,%xmm11 DB 102,65,15,96,194 ; punpcklbw %xmm10,%xmm0 DB 102,69,15,104,202 ; punpckhbw %xmm10,%xmm9 - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 DB 242,68,15,112,233,231 ; pshuflw $0xe7,%xmm1,%xmm13 DB 102,65,15,96,202 ; punpcklbw %xmm10,%xmm1 - DB 102,65,15,104,226 ; punpckhbw %xmm10,%xmm4 - DB 102,65,15,111,244 ; movdqa %xmm12,%xmm6 - DB 102,65,15,96,242 ; punpcklbw %xmm10,%xmm6 + DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 + DB 102,65,15,111,236 ; movdqa %xmm12,%xmm5 + DB 102,65,15,96,234 ; punpcklbw %xmm10,%xmm5 DB 102,69,15,104,226 ; punpckhbw %xmm10,%xmm12 - DB 102,15,111,253 ; movdqa %xmm5,%xmm7 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 DB 102,65,15,96,250 ; punpcklbw %xmm10,%xmm7 - DB 102,65,15,104,234 ; punpckhbw %xmm10,%xmm5 - DB 102,15,213,236 ; pmullw %xmm4,%xmm5 + DB 102,65,15,104,226 ; punpckhbw %xmm10,%xmm4 + DB 102,15,213,230 ; pmullw %xmm6,%xmm4 DB 102,15,213,249 ; pmullw %xmm1,%xmm7 DB 102,69,15,213,225 ; pmullw %xmm9,%xmm12 - DB 102,15,213,240 ; pmullw %xmm0,%xmm6 - DB 102,15,253,240 ; paddw %xmm0,%xmm6 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,253,232 ; paddw %xmm0,%xmm5 DB 102,69,15,253,225 ; paddw %xmm9,%xmm12 DB 102,15,253,249 ; paddw %xmm1,%xmm7 - DB 102,15,253,236 ; paddw %xmm4,%xmm5 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,253,230 ; paddw %xmm6,%xmm4 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,65,15,103,244 ; packuswb %xmm12,%xmm6 - DB 102,15,103,253 ; packuswb %xmm5,%xmm7 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,65,15,103,236 ; packuswb %xmm12,%xmm5 DB 243,65,15,112,195,231 ; pshufhw $0xe7,%xmm11,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 @@ -41900,8 +41823,8 @@ _sk_xor__sse2_8bit LABEL PROC DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,232,95 ; pshufhw $0x5f,%xmm0,%xmm5 - DB 102,65,15,239,232 ; pxor %xmm8,%xmm5 + DB 243,15,112,240,95 ; pshufhw $0x5f,%xmm0,%xmm6 + DB 102,65,15,239,240 ; pxor %xmm8,%xmm6 DB 102,65,15,239,224 ; pxor %xmm8,%xmm4 DB 102,68,15,111,194 ; movdqa %xmm2,%xmm8 DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 @@ -41914,25 +41837,25 @@ _sk_xor__sse2_8bit LABEL PROC DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,65,15,96,194 ; punpcklbw %xmm10,%xmm0 DB 102,65,15,104,226 ; punpckhbw %xmm10,%xmm4 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 DB 102,65,15,96,202 ; punpcklbw %xmm10,%xmm1 - DB 102,65,15,104,234 ; punpckhbw %xmm10,%xmm5 - DB 102,65,15,213,235 ; pmullw %xmm11,%xmm5 + DB 102,65,15,104,242 ; punpckhbw %xmm10,%xmm6 + DB 102,65,15,213,243 ; pmullw %xmm11,%xmm6 DB 102,65,15,213,204 ; pmullw %xmm12,%xmm1 DB 102,65,15,213,224 ; pmullw %xmm8,%xmm4 DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 DB 102,65,15,253,193 ; paddw %xmm9,%xmm0 DB 102,65,15,253,224 ; paddw %xmm8,%xmm4 DB 102,65,15,253,204 ; paddw %xmm12,%xmm1 - DB 102,65,15,253,235 ; paddw %xmm11,%xmm5 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,65,15,253,243 ; paddw %xmm11,%xmm6 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,206 ; packuswb %xmm6,%xmm1 + DB 102,15,252,207 ; paddb %xmm7,%xmm1 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,196 ; packuswb %xmm4,%xmm0 - DB 102,15,103,205 ; packuswb %xmm5,%xmm1 - DB 102,15,252,198 ; paddb %xmm6,%xmm0 - DB 102,15,252,207 ; paddb %xmm7,%xmm1 + DB 102,15,252,197 ; paddb %xmm5,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -41982,10 +41905,10 @@ _sk_darken_sse2_8bit LABEL PROC DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 + DB 102,68,15,103,236 ; packuswb %xmm4,%xmm13 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 DB 102,68,15,103,215 ; packuswb %xmm7,%xmm10 - DB 102,68,15,103,236 ; packuswb %xmm4,%xmm13 DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 @@ -42024,10 +41947,10 @@ _sk_darken_sse2_8bit LABEL PROC DB 102,69,15,253,222 ; paddw %xmm14,%xmm11 DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,103,251 ; packuswb %xmm11,%xmm7 DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,65,15,103,228 ; packuswb %xmm12,%xmm4 - DB 102,65,15,103,251 ; packuswb %xmm11,%xmm7 DB 102,68,15,222,212 ; pmaxub %xmm4,%xmm10 DB 102,68,15,222,239 ; pmaxub %xmm7,%xmm13 DB 102,15,111,241 ; movdqa %xmm1,%xmm6 @@ -42036,7 +41959,7 @@ _sk_darken_sse2_8bit LABEL PROC DB 102,65,15,248,234 ; psubb %xmm10,%xmm5 DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 102,15,248,196 ; psubb %xmm4,%xmm0 - DB 102,15,111,37,41,32,0,0 ; movdqa 0x2029(%rip),%xmm4 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> + DB 102,15,111,37,181,31,0,0 ; movdqa 0x1fb5(%rip),%xmm4 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,111,252 ; movdqa %xmm4,%xmm7 DB 102,15,223,248 ; pandn %xmm0,%xmm7 @@ -42088,10 +42011,10 @@ _sk_lighten_sse2_8bit LABEL PROC DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 + DB 102,68,15,103,236 ; packuswb %xmm4,%xmm13 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 DB 102,68,15,103,215 ; packuswb %xmm7,%xmm10 - DB 102,68,15,103,236 ; packuswb %xmm4,%xmm13 DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 @@ -42130,10 +42053,10 @@ _sk_lighten_sse2_8bit LABEL PROC DB 102,69,15,253,222 ; paddw %xmm14,%xmm11 DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,103,251 ; packuswb %xmm11,%xmm7 DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,65,15,103,228 ; packuswb %xmm12,%xmm4 - DB 102,65,15,103,251 ; packuswb %xmm11,%xmm7 DB 102,68,15,218,212 ; pminub %xmm4,%xmm10 DB 102,68,15,218,239 ; pminub %xmm7,%xmm13 DB 102,15,111,241 ; movdqa %xmm1,%xmm6 @@ -42142,7 +42065,7 @@ _sk_lighten_sse2_8bit LABEL PROC DB 102,65,15,248,234 ; psubb %xmm10,%xmm5 DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 102,15,248,196 ; psubb %xmm4,%xmm0 - DB 102,15,111,37,61,30,0,0 ; movdqa 0x1e3d(%rip),%xmm4 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> + DB 102,15,111,37,201,29,0,0 ; movdqa 0x1dc9(%rip),%xmm4 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,111,252 ; movdqa %xmm4,%xmm7 DB 102,15,223,248 ; pandn %xmm0,%xmm7 @@ -42186,17 +42109,17 @@ _sk_exclusion_sse2_8bit LABEL PROC DB 102,65,15,253,252 ; paddw %xmm12,%xmm7 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,239 ; packuswb %xmm7,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,103,244 ; packuswb %xmm4,%xmm6 - DB 102,15,103,239 ; packuswb %xmm7,%xmm5 - DB 102,15,111,37,112,29,0,0 ; movdqa 0x1d70(%rip),%xmm4 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> + DB 102,15,111,37,252,28,0,0 ; movdqa 0x1cfc(%rip),%xmm4 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> DB 102,15,248,205 ; psubb %xmm5,%xmm1 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,219,230 ; pand %xmm6,%xmm4 + DB 102,15,248,205 ; psubb %xmm5,%xmm1 DB 102,15,248,198 ; psubb %xmm6,%xmm0 DB 102,15,248,196 ; psubb %xmm4,%xmm0 - DB 102,15,248,205 ; psubb %xmm5,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -42239,10 +42162,10 @@ _sk_difference_sse2_8bit LABEL PROC DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 + DB 102,68,15,103,220 ; packuswb %xmm4,%xmm11 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 DB 102,68,15,103,215 ; packuswb %xmm7,%xmm10 - DB 102,68,15,103,220 ; packuswb %xmm4,%xmm11 DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 @@ -42281,457 +42204,440 @@ _sk_difference_sse2_8bit LABEL PROC DB 102,69,15,253,230 ; paddw %xmm14,%xmm12 DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,65,15,103,244 ; packuswb %xmm12,%xmm6 + DB 102,65,15,218,243 ; pminub %xmm11,%xmm6 DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,65,15,103,229 ; packuswb %xmm13,%xmm4 - DB 102,65,15,103,244 ; packuswb %xmm12,%xmm6 DB 102,65,15,218,226 ; pminub %xmm10,%xmm4 - DB 102,65,15,218,243 ; pminub %xmm11,%xmm6 - DB 102,15,111,45,173,27,0,0 ; movdqa 0x1bad(%rip),%xmm5 # 2dc80 <_sk_overlay_sse2_8bit+0x1728> + DB 102,15,111,45,57,27,0,0 ; movdqa 0x1b39(%rip),%xmm5 # 2db80 <_sk_overlay_sse2_8bit+0x16d1> DB 102,15,248,206 ; psubb %xmm6,%xmm1 DB 102,15,219,245 ; pand %xmm5,%xmm6 DB 102,15,219,236 ; pand %xmm4,%xmm5 + DB 102,15,248,206 ; psubb %xmm6,%xmm1 DB 102,15,248,196 ; psubb %xmm4,%xmm0 DB 102,15,248,197 ; psubb %xmm5,%xmm0 - DB 102,15,248,206 ; psubb %xmm6,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_hardlight_sse2_8bit _sk_hardlight_sse2_8bit LABEL PROC - DB 72,129,236,56,1,0,0 ; sub $0x138,%rsp + DB 72,129,236,24,1,0,0 ; sub $0x118,%rsp DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 - DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,69,15,111,240 ; movdqa %xmm8,%xmm14 - DB 102,68,15,96,245 ; punpcklbw %xmm5,%xmm14 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,96,196 ; punpcklbw %xmm4,%xmm0 + DB 102,15,127,132,36,176,0,0,0 ; movdqa %xmm0,0xb0(%rsp) DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,96,197 ; punpcklbw %xmm5,%xmm0 - DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 - DB 242,65,15,112,192,231 ; pshuflw $0xe7,%xmm8,%xmm0 - DB 102,68,15,104,197 ; punpckhbw %xmm5,%xmm8 + DB 102,15,96,196 ; punpcklbw %xmm4,%xmm0 + DB 102,15,127,68,36,64 ; movdqa %xmm0,0x40(%rsp) + DB 242,15,112,198,231 ; pshuflw $0xe7,%xmm6,%xmm0 + DB 102,15,104,244 ; punpckhbw %xmm4,%xmm6 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,224,95 ; pshufhw $0x5f,%xmm0,%xmm4 + DB 243,68,15,112,192,95 ; pshufhw $0x5f,%xmm0,%xmm8 DB 242,15,112,193,231 ; pshuflw $0xe7,%xmm1,%xmm0 - DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 - DB 102,15,127,76,36,64 ; movdqa %xmm1,0x40(%rsp) + DB 102,15,104,204 ; punpckhbw %xmm4,%xmm1 + DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 + DB 102,68,15,127,180,36,160,0,0,0 ; movdqa %xmm14,0xa0(%rsp) DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,68,15,112,224,95 ; pshufhw $0x5f,%xmm0,%xmm12 - DB 102,65,15,111,252 ; movdqa %xmm12,%xmm7 - DB 102,15,104,253 ; punpckhbw %xmm5,%xmm7 - DB 102,15,127,188,36,176,0,0,0 ; movdqa %xmm7,0xb0(%rsp) - DB 102,68,15,96,229 ; punpcklbw %xmm5,%xmm12 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 - DB 102,15,127,68,36,32 ; movdqa %xmm0,0x20(%rsp) - DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 - DB 102,68,15,127,188,36,32,1,0,0 ; movdqa %xmm15,0x120(%rsp) + DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,104,196 ; punpckhbw %xmm4,%xmm0 + DB 102,15,127,68,36,80 ; movdqa %xmm0,0x50(%rsp) + DB 102,15,96,204 ; punpcklbw %xmm4,%xmm1 + DB 102,15,127,76,36,48 ; movdqa %xmm1,0x30(%rsp) + DB 102,69,15,111,232 ; movdqa %xmm8,%xmm13 + DB 102,68,15,104,236 ; punpckhbw %xmm4,%xmm13 + DB 102,68,15,127,108,36,16 ; movdqa %xmm13,0x10(%rsp) + DB 102,68,15,96,196 ; punpcklbw %xmm4,%xmm8 + DB 102,68,15,127,4,36 ; movdqa %xmm8,(%rsp) + DB 102,68,15,127,188,36,0,1,0,0 ; movdqa %xmm15,0x100(%rsp) DB 102,65,15,111,207 ; movdqa %xmm15,%xmm1 DB 242,65,15,112,199,231 ; pshuflw $0xe7,%xmm15,%xmm0 - DB 102,68,15,96,253 ; punpcklbw %xmm5,%xmm15 - DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 - DB 102,15,127,76,36,16 ; movdqa %xmm1,0x10(%rsp) - DB 102,15,127,156,36,16,1,0,0 ; movdqa %xmm3,0x110(%rsp) - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,96,205 ; punpcklbw %xmm5,%xmm1 - DB 102,15,127,12,36 ; movdqa %xmm1,(%rsp) - DB 102,15,104,213 ; punpckhbw %xmm5,%xmm2 - DB 102,15,127,148,36,0,1,0,0 ; movdqa %xmm2,0x100(%rsp) - DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,240,95 ; pshufhw $0x5f,%xmm0,%xmm6 - DB 242,15,112,195,231 ; pshuflw $0xe7,%xmm3,%xmm0 + DB 102,68,15,96,252 ; punpcklbw %xmm4,%xmm15 + DB 102,15,104,204 ; punpckhbw %xmm4,%xmm1 + DB 102,15,127,76,36,112 ; movdqa %xmm1,0x70(%rsp) + DB 102,15,127,156,36,240,0,0,0 ; movdqa %xmm3,0xf0(%rsp) + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,68,15,111,227 ; movdqa %xmm3,%xmm12 + DB 102,68,15,96,228 ; punpcklbw %xmm4,%xmm12 + DB 102,68,15,127,164,36,128,0,0,0 ; movdqa %xmm12,0x80(%rsp) + DB 102,15,104,252 ; punpckhbw %xmm4,%xmm7 + DB 102,15,127,188,36,144,0,0,0 ; movdqa %xmm7,0x90(%rsp) DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,68,15,112,232,95 ; pshufhw $0x5f,%xmm0,%xmm13 - DB 102,65,15,111,205 ; movdqa %xmm13,%xmm1 - DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 - DB 102,15,127,76,36,48 ; movdqa %xmm1,0x30(%rsp) - DB 102,68,15,96,237 ; punpcklbw %xmm5,%xmm13 - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 - DB 102,15,96,245 ; punpcklbw %xmm5,%xmm6 - DB 102,69,15,111,214 ; movdqa %xmm14,%xmm10 - DB 102,69,15,253,210 ; paddw %xmm10,%xmm10 - DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 - DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 - DB 102,15,111,84,36,64 ; movdqa 0x40(%rsp),%xmm2 - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,217,215 ; psubusw %xmm7,%xmm2 - DB 102,15,117,213 ; pcmpeqw %xmm5,%xmm2 - DB 102,15,127,148,36,240,0,0,0 ; movdqa %xmm2,0xf0(%rsp) - DB 102,69,15,217,204 ; psubusw %xmm12,%xmm9 - DB 102,68,15,117,205 ; pcmpeqw %xmm5,%xmm9 - DB 102,15,111,84,36,32 ; movdqa 0x20(%rsp),%xmm2 - DB 102,15,217,194 ; psubusw %xmm2,%xmm0 - DB 102,15,117,197 ; pcmpeqw %xmm5,%xmm0 - DB 102,15,127,132,36,192,0,0,0 ; movdqa %xmm0,0xc0(%rsp) - DB 102,68,15,111,220 ; movdqa %xmm4,%xmm11 - DB 102,69,15,217,211 ; psubusw %xmm11,%xmm10 - DB 102,68,15,117,213 ; pcmpeqw %xmm5,%xmm10 - DB 102,15,111,45,176,17,0,0 ; movdqa 0x11b0(%rip),%xmm5 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,15,239,197 ; pxor %xmm5,%xmm0 - DB 102,65,15,213,198 ; pmullw %xmm14,%xmm0 - DB 102,15,127,132,36,224,0,0,0 ; movdqa %xmm0,0xe0(%rsp) + DB 243,68,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm9 + DB 242,15,112,203,231 ; pshuflw $0xe7,%xmm3,%xmm1 + DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 + DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 + DB 243,68,15,112,217,95 ; pshufhw $0x5f,%xmm1,%xmm11 DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 - DB 102,15,239,197 ; pxor %xmm5,%xmm0 - DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 - DB 102,15,127,132,36,208,0,0,0 ; movdqa %xmm0,0xd0(%rsp) - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,65,15,249,247 ; psubw %xmm15,%xmm6 - DB 102,15,127,180,36,128,0,0,0 ; movdqa %xmm6,0x80(%rsp) - DB 102,65,15,111,255 ; movdqa %xmm15,%xmm7 - DB 102,65,15,213,254 ; pmullw %xmm14,%xmm7 - DB 102,65,15,213,195 ; pmullw %xmm11,%xmm0 - DB 102,15,127,132,36,144,0,0,0 ; movdqa %xmm0,0x90(%rsp) - DB 102,69,15,249,222 ; psubw %xmm14,%xmm11 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,239,197 ; pxor %xmm5,%xmm0 - DB 102,65,15,213,192 ; pmullw %xmm8,%xmm0 - DB 102,15,127,132,36,160,0,0,0 ; movdqa %xmm0,0xa0(%rsp) - DB 102,68,15,111,250 ; movdqa %xmm2,%xmm15 - DB 102,68,15,239,253 ; pxor %xmm5,%xmm15 - DB 102,15,111,116,36,16 ; movdqa 0x10(%rsp),%xmm6 - DB 102,68,15,213,254 ; pmullw %xmm6,%xmm15 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,249,206 ; psubw %xmm6,%xmm1 - DB 102,15,127,76,36,96 ; movdqa %xmm1,0x60(%rsp) - DB 102,65,15,213,240 ; pmullw %xmm8,%xmm6 - DB 102,15,213,194 ; pmullw %xmm2,%xmm0 - DB 102,15,127,68,36,112 ; movdqa %xmm0,0x70(%rsp) - DB 102,65,15,249,208 ; psubw %xmm8,%xmm2 - DB 102,15,127,84,36,32 ; movdqa %xmm2,0x20(%rsp) - DB 102,65,15,111,197 ; movdqa %xmm13,%xmm0 + DB 102,15,104,196 ; punpckhbw %xmm4,%xmm0 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,239,205 ; pxor %xmm5,%xmm1 - DB 102,15,213,203 ; pmullw %xmm3,%xmm1 - DB 102,15,127,76,36,16 ; movdqa %xmm1,0x10(%rsp) + DB 102,15,127,76,36,32 ; movdqa %xmm1,0x20(%rsp) + DB 102,68,15,96,220 ; punpcklbw %xmm4,%xmm11 + DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3 + DB 102,15,104,220 ; punpckhbw %xmm4,%xmm3 + DB 102,68,15,96,204 ; punpcklbw %xmm4,%xmm9 + DB 102,68,15,111,21,112,17,0,0 ; movdqa 0x1170(%rip),%xmm10 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 102,65,15,111,209 ; movdqa %xmm9,%xmm2 + DB 102,65,15,239,210 ; pxor %xmm10,%xmm2 + DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,65,15,239,226 ; pxor %xmm10,%xmm4 + DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 + DB 102,65,15,239,194 ; pxor %xmm10,%xmm0 + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,65,15,239,234 ; pxor %xmm10,%xmm5 + DB 102,65,15,213,238 ; pmullw %xmm14,%xmm5 + DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1 + DB 102,65,15,239,202 ; pxor %xmm10,%xmm1 + DB 102,69,15,239,234 ; pxor %xmm10,%xmm13 + DB 102,68,15,111,68,36,48 ; movdqa 0x30(%rsp),%xmm8 + DB 102,69,15,239,194 ; pxor %xmm10,%xmm8 + DB 102,68,15,239,84,36,80 ; pxor 0x50(%rsp),%xmm10 + DB 102,68,15,213,215 ; pmullw %xmm7,%xmm10 + DB 102,68,15,253,213 ; paddw %xmm5,%xmm10 + DB 102,15,213,68,36,64 ; pmullw 0x40(%rsp),%xmm0 + DB 102,69,15,213,196 ; pmullw %xmm12,%xmm8 + DB 102,68,15,253,192 ; paddw %xmm0,%xmm8 + DB 102,68,15,127,132,36,224,0,0,0 ; movdqa %xmm8,0xe0(%rsp) + DB 102,15,213,230 ; pmullw %xmm6,%xmm4 + DB 102,15,111,108,36,112 ; movdqa 0x70(%rsp),%xmm5 + DB 102,68,15,213,237 ; pmullw %xmm5,%xmm13 + DB 102,68,15,253,236 ; paddw %xmm4,%xmm13 + DB 102,68,15,127,172,36,208,0,0,0 ; movdqa %xmm13,0xd0(%rsp) + DB 102,68,15,111,164,36,176,0,0,0 ; movdqa 0xb0(%rsp),%xmm12 + DB 102,65,15,213,212 ; pmullw %xmm12,%xmm2 + DB 102,65,15,213,207 ; pmullw %xmm15,%xmm1 + DB 102,15,253,202 ; paddw %xmm2,%xmm1 + DB 102,15,127,140,36,192,0,0,0 ; movdqa %xmm1,0xc0(%rsp) DB 102,69,15,111,196 ; movdqa %xmm12,%xmm8 - DB 102,68,15,239,197 ; pxor %xmm5,%xmm8 - DB 102,15,111,36,36 ; movdqa (%rsp),%xmm4 - DB 102,68,15,213,196 ; pmullw %xmm4,%xmm8 - DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13 - DB 102,15,249,196 ; psubw %xmm4,%xmm0 - DB 102,15,127,68,36,80 ; movdqa %xmm0,0x50(%rsp) - DB 102,15,213,227 ; pmullw %xmm3,%xmm4 - DB 102,69,15,213,236 ; pmullw %xmm12,%xmm13 - DB 102,68,15,249,227 ; psubw %xmm3,%xmm12 - DB 102,15,111,76,36,48 ; movdqa 0x30(%rsp),%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,239,197 ; pxor %xmm5,%xmm0 - DB 102,15,111,92,36,64 ; movdqa 0x40(%rsp),%xmm3 - DB 102,15,213,195 ; pmullw %xmm3,%xmm0 + DB 102,65,15,111,204 ; movdqa %xmm12,%xmm1 + DB 102,69,15,253,192 ; paddw %xmm8,%xmm8 + DB 102,15,111,4,36 ; movdqa (%rsp),%xmm0 + DB 102,68,15,101,192 ; pcmpgtw %xmm0,%xmm8 + DB 102,69,15,111,233 ; movdqa %xmm9,%xmm13 + DB 102,69,15,249,207 ; psubw %xmm15,%xmm9 + DB 102,68,15,213,249 ; pmullw %xmm1,%xmm15 + DB 102,68,15,127,124,36,96 ; movdqa %xmm15,0x60(%rsp) + DB 102,68,15,213,232 ; pmullw %xmm0,%xmm13 + DB 102,15,249,193 ; psubw %xmm1,%xmm0 + DB 102,65,15,213,193 ; pmullw %xmm9,%xmm0 DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) - DB 102,15,111,132,36,176,0,0,0 ; movdqa 0xb0(%rsp),%xmm0 - DB 102,15,239,232 ; pxor %xmm0,%xmm5 - DB 102,15,111,148,36,0,1,0,0 ; movdqa 0x100(%rsp),%xmm2 - DB 102,15,213,234 ; pmullw %xmm2,%xmm5 - DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,15,249,202 ; psubw %xmm2,%xmm1 - DB 102,15,127,76,36,48 ; movdqa %xmm1,0x30(%rsp) - DB 102,15,213,211 ; pmullw %xmm3,%xmm2 - DB 102,68,15,213,240 ; pmullw %xmm0,%xmm14 + DB 102,68,15,111,206 ; movdqa %xmm6,%xmm9 + DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,68,15,101,200 ; pcmpgtw %xmm0,%xmm9 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,15,249,221 ; psubw %xmm5,%xmm3 + DB 102,68,15,111,245 ; movdqa %xmm5,%xmm14 + DB 102,68,15,213,246 ; pmullw %xmm6,%xmm14 + DB 102,15,213,248 ; pmullw %xmm0,%xmm7 + DB 102,15,249,198 ; psubw %xmm6,%xmm0 + DB 102,15,213,195 ; pmullw %xmm3,%xmm0 + DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) + DB 102,15,111,92,36,64 ; movdqa 0x40(%rsp),%xmm3 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,15,253,237 ; paddw %xmm5,%xmm5 + DB 102,15,111,68,36,48 ; movdqa 0x30(%rsp),%xmm0 + DB 102,15,101,232 ; pcmpgtw %xmm0,%xmm5 + DB 102,65,15,111,243 ; movdqa %xmm11,%xmm6 + DB 102,68,15,111,188,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm15 + DB 102,69,15,249,223 ; psubw %xmm15,%xmm11 + DB 102,68,15,213,251 ; pmullw %xmm3,%xmm15 + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 DB 102,15,249,195 ; psubw %xmm3,%xmm0 - DB 102,15,213,68,36,48 ; pmullw 0x30(%rsp),%xmm0 - DB 102,68,15,213,100,36,80 ; pmullw 0x50(%rsp),%xmm12 + DB 102,65,15,213,195 ; pmullw %xmm11,%xmm0 + DB 102,68,15,111,216 ; movdqa %xmm0,%xmm11 + DB 102,15,111,148,36,160,0,0,0 ; movdqa 0xa0(%rsp),%xmm2 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,15,253,228 ; paddw %xmm4,%xmm4 + DB 102,15,111,68,36,80 ; movdqa 0x50(%rsp),%xmm0 + DB 102,15,101,224 ; pcmpgtw %xmm0,%xmm4 DB 102,15,111,76,36,32 ; movdqa 0x20(%rsp),%xmm1 - DB 102,15,213,76,36,96 ; pmullw 0x60(%rsp),%xmm1 - DB 102,68,15,213,156,36,128,0,0,0 ; pmullw 0x80(%rsp),%xmm11 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,68,15,249,240 ; psubw %xmm0,%xmm14 - DB 102,69,15,253,228 ; paddw %xmm12,%xmm12 - DB 102,69,15,249,236 ; psubw %xmm12,%xmm13 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,68,15,111,164,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm12 + DB 102,65,15,249,204 ; psubw %xmm12,%xmm1 + DB 102,15,127,76,36,32 ; movdqa %xmm1,0x20(%rsp) + DB 102,68,15,213,226 ; pmullw %xmm2,%xmm12 + DB 102,15,213,216 ; pmullw %xmm0,%xmm3 + DB 102,15,249,194 ; psubw %xmm2,%xmm0 + DB 102,15,213,68,36,32 ; pmullw 0x20(%rsp),%xmm0 + DB 102,15,111,12,36 ; movdqa (%rsp),%xmm1 DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,15,111,92,36,112 ; movdqa 0x70(%rsp),%xmm3 - DB 102,15,249,217 ; psubw %xmm1,%xmm3 + DB 102,68,15,249,233 ; psubw %xmm1,%xmm13 + DB 102,15,111,76,36,16 ; movdqa 0x10(%rsp),%xmm1 + DB 102,15,253,201 ; paddw %xmm1,%xmm1 + DB 102,15,249,249 ; psubw %xmm1,%xmm7 DB 102,69,15,253,219 ; paddw %xmm11,%xmm11 - DB 102,15,111,132,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm0 - DB 102,65,15,249,195 ; psubw %xmm11,%xmm0 - DB 102,15,253,255 ; paddw %xmm7,%xmm7 - DB 102,65,15,219,250 ; pand %xmm10,%xmm7 - DB 102,68,15,223,208 ; pandn %xmm0,%xmm10 - DB 102,68,15,235,215 ; por %xmm7,%xmm10 - DB 102,15,253,246 ; paddw %xmm6,%xmm6 - DB 102,15,111,132,36,192,0,0,0 ; movdqa 0xc0(%rsp),%xmm0 - DB 102,15,219,240 ; pand %xmm0,%xmm6 - DB 102,15,223,195 ; pandn %xmm3,%xmm0 - DB 102,15,235,198 ; por %xmm6,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,65,15,219,225 ; pand %xmm9,%xmm4 - DB 102,69,15,223,205 ; pandn %xmm13,%xmm9 - DB 102,68,15,235,204 ; por %xmm4,%xmm9 - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,111,132,36,240,0,0,0 ; movdqa 0xf0(%rsp),%xmm0 - DB 102,15,219,208 ; pand %xmm0,%xmm2 - DB 102,65,15,223,198 ; pandn %xmm14,%xmm0 - DB 102,15,235,194 ; por %xmm2,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,148,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm2 - DB 102,15,253,148,36,224,0,0,0 ; paddw 0xe0(%rsp),%xmm2 - DB 102,68,15,253,188,36,160,0,0,0 ; paddw 0xa0(%rsp),%xmm15 - DB 102,68,15,253,68,36,16 ; paddw 0x10(%rsp),%xmm8 - DB 102,15,253,44,36 ; paddw (%rsp),%xmm5 - DB 102,15,111,5,190,23,0,0 ; movdqa 0x17be(%rip),%xmm0 # 2dc90 <_sk_overlay_sse2_8bit+0x1738> - DB 102,15,253,208 ; paddw %xmm0,%xmm2 - DB 102,68,15,253,210 ; paddw %xmm2,%xmm10 - DB 102,68,15,253,248 ; paddw %xmm0,%xmm15 - DB 102,68,15,253,251 ; paddw %xmm3,%xmm15 - DB 102,68,15,253,192 ; paddw %xmm0,%xmm8 - DB 102,69,15,253,200 ; paddw %xmm8,%xmm9 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,15,253,233 ; paddw %xmm1,%xmm5 - DB 102,15,111,5,161,23,0,0 ; movdqa 0x17a1(%rip),%xmm0 # 2dca0 <_sk_overlay_sse2_8bit+0x1748> - DB 102,15,228,232 ; pmulhuw %xmm0,%xmm5 - DB 102,68,15,228,200 ; pmulhuw %xmm0,%xmm9 - DB 102,68,15,228,248 ; pmulhuw %xmm0,%xmm15 + DB 102,65,15,249,243 ; psubw %xmm11,%xmm6 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,249,216 ; psubw %xmm0,%xmm3 + DB 102,15,111,68,36,96 ; movdqa 0x60(%rsp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,69,15,219,232 ; pand %xmm8,%xmm13 + DB 102,68,15,223,192 ; pandn %xmm0,%xmm8 + DB 102,69,15,235,197 ; por %xmm13,%xmm8 + DB 102,69,15,253,246 ; paddw %xmm14,%xmm14 + DB 102,65,15,219,249 ; pand %xmm9,%xmm7 + DB 102,69,15,223,206 ; pandn %xmm14,%xmm9 + DB 102,68,15,235,207 ; por %xmm7,%xmm9 + DB 102,69,15,253,255 ; paddw %xmm15,%xmm15 + DB 102,15,219,245 ; pand %xmm5,%xmm6 + DB 102,65,15,223,239 ; pandn %xmm15,%xmm5 + DB 102,15,235,238 ; por %xmm6,%xmm5 + DB 102,69,15,253,228 ; paddw %xmm12,%xmm12 + DB 102,15,219,220 ; pand %xmm4,%xmm3 + DB 102,65,15,223,228 ; pandn %xmm12,%xmm4 + DB 102,15,235,227 ; por %xmm3,%xmm4 + DB 102,15,111,5,126,23,0,0 ; movdqa 0x177e(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x16e1> + DB 102,15,111,140,36,192,0,0,0 ; movdqa 0xc0(%rsp),%xmm1 + DB 102,15,253,200 ; paddw %xmm0,%xmm1 + DB 102,68,15,253,193 ; paddw %xmm1,%xmm8 + DB 102,15,111,140,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm1 + DB 102,15,253,200 ; paddw %xmm0,%xmm1 + DB 102,65,15,253,201 ; paddw %xmm9,%xmm1 + DB 102,15,111,156,36,224,0,0,0 ; movdqa 0xe0(%rsp),%xmm3 + DB 102,15,253,216 ; paddw %xmm0,%xmm3 + DB 102,15,253,235 ; paddw %xmm3,%xmm5 + DB 102,68,15,253,208 ; paddw %xmm0,%xmm10 + DB 102,68,15,253,212 ; paddw %xmm4,%xmm10 + DB 102,15,111,5,71,23,0,0 ; movdqa 0x1747(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x16f1> DB 102,68,15,228,208 ; pmulhuw %xmm0,%xmm10 - DB 102,65,15,113,215,7 ; psrlw $0x7,%xmm15 DB 102,65,15,113,210,7 ; psrlw $0x7,%xmm10 - DB 102,69,15,103,215 ; packuswb %xmm15,%xmm10 + DB 102,15,228,232 ; pmulhuw %xmm0,%xmm5 DB 102,15,113,213,7 ; psrlw $0x7,%xmm5 - DB 102,65,15,113,209,7 ; psrlw $0x7,%xmm9 - DB 102,68,15,103,205 ; packuswb %xmm5,%xmm9 + DB 102,65,15,103,234 ; packuswb %xmm10,%xmm5 + DB 102,15,228,200 ; pmulhuw %xmm0,%xmm1 + DB 102,68,15,228,192 ; pmulhuw %xmm0,%xmm8 + DB 102,15,113,209,7 ; psrlw $0x7,%xmm1 + DB 102,65,15,113,208,7 ; psrlw $0x7,%xmm8 + DB 102,68,15,103,193 ; packuswb %xmm1,%xmm8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,148,36,32,1,0,0 ; movaps 0x120(%rsp),%xmm2 - DB 15,40,156,36,16,1,0,0 ; movaps 0x110(%rsp),%xmm3 - DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 - DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1 - DB 72,129,196,56,1,0,0 ; add $0x138,%rsp + DB 15,40,148,36,0,1,0,0 ; movaps 0x100(%rsp),%xmm2 + DB 15,40,156,36,240,0,0,0 ; movaps 0xf0(%rsp),%xmm3 + DB 102,65,15,111,192 ; movdqa %xmm8,%xmm0 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 72,129,196,24,1,0,0 ; add $0x118,%rsp DB 255,224 ; jmpq *%rax PUBLIC _sk_overlay_sse2_8bit _sk_overlay_sse2_8bit LABEL PROC - DB 72,129,236,72,1,0,0 ; sub $0x148,%rsp - DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,68,15,111,248 ; movdqa %xmm0,%xmm15 - DB 102,68,15,96,252 ; punpcklbw %xmm4,%xmm15 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,96,236 ; punpcklbw %xmm4,%xmm5 - DB 102,15,127,108,36,16 ; movdqa %xmm5,0x10(%rsp) - DB 242,15,112,232,231 ; pshuflw $0xe7,%xmm0,%xmm5 - DB 102,15,104,196 ; punpckhbw %xmm4,%xmm0 - DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) - DB 243,15,112,197,231 ; pshufhw $0xe7,%xmm5,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,232,95 ; pshufhw $0x5f,%xmm0,%xmm5 - DB 242,15,112,193,231 ; pshuflw $0xe7,%xmm1,%xmm0 - DB 102,15,104,204 ; punpckhbw %xmm4,%xmm1 - DB 102,15,127,140,36,16,1,0,0 ; movdqa %xmm1,0x110(%rsp) - DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,104,196 ; punpckhbw %xmm4,%xmm0 - DB 102,15,127,132,36,0,1,0,0 ; movdqa %xmm0,0x100(%rsp) - DB 102,15,96,204 ; punpcklbw %xmm4,%xmm1 - DB 102,15,127,140,36,224,0,0,0 ; movdqa %xmm1,0xe0(%rsp) - DB 102,68,15,111,221 ; movdqa %xmm5,%xmm11 - DB 102,68,15,104,220 ; punpckhbw %xmm4,%xmm11 - DB 102,15,96,236 ; punpcklbw %xmm4,%xmm5 - DB 102,68,15,111,213 ; movdqa %xmm5,%xmm10 - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,15,127,140,36,48,1,0,0 ; movdqa %xmm1,0x130(%rsp) - DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 - DB 242,15,112,193,231 ; pshuflw $0xe7,%xmm1,%xmm0 - DB 102,15,96,204 ; punpcklbw %xmm4,%xmm1 - DB 102,68,15,104,196 ; punpckhbw %xmm4,%xmm8 - DB 102,15,127,156,36,32,1,0,0 ; movdqa %xmm3,0x120(%rsp) - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 - DB 102,68,15,96,244 ; punpcklbw %xmm4,%xmm14 - DB 102,15,104,212 ; punpckhbw %xmm4,%xmm2 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,15,127,180,36,144,0,0,0 ; movdqa %xmm6,0x90(%rsp) - DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,208,95 ; pshufhw $0x5f,%xmm0,%xmm2 - DB 242,15,112,195,231 ; pshuflw $0xe7,%xmm3,%xmm0 - DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,216,95 ; pshufhw $0x5f,%xmm0,%xmm3 - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,104,196 ; punpckhbw %xmm4,%xmm0 - DB 102,15,127,132,36,128,0,0,0 ; movdqa %xmm0,0x80(%rsp) - DB 102,15,96,220 ; punpcklbw %xmm4,%xmm3 - DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 - DB 102,68,15,104,204 ; punpckhbw %xmm4,%xmm9 - DB 102,15,96,212 ; punpcklbw %xmm4,%xmm2 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,253,237 ; paddw %xmm5,%xmm5 - DB 102,65,15,111,248 ; movdqa %xmm8,%xmm7 - DB 102,15,253,255 ; paddw %xmm7,%xmm7 - DB 102,69,15,111,238 ; movdqa %xmm14,%xmm13 - DB 102,69,15,253,237 ; paddw %xmm13,%xmm13 - DB 102,15,253,246 ; paddw %xmm6,%xmm6 - DB 102,15,217,240 ; psubusw %xmm0,%xmm6 - DB 102,15,117,244 ; pcmpeqw %xmm4,%xmm6 - DB 102,15,127,180,36,240,0,0,0 ; movdqa %xmm6,0xf0(%rsp) - DB 102,68,15,217,235 ; psubusw %xmm3,%xmm13 - DB 102,68,15,117,236 ; pcmpeqw %xmm4,%xmm13 - DB 102,68,15,127,172,36,208,0,0,0 ; movdqa %xmm13,0xd0(%rsp) - DB 102,65,15,217,249 ; psubusw %xmm9,%xmm7 - DB 102,15,117,252 ; pcmpeqw %xmm4,%xmm7 - DB 102,15,127,188,36,176,0,0,0 ; movdqa %xmm7,0xb0(%rsp) - DB 102,15,217,234 ; psubusw %xmm2,%xmm5 - DB 102,15,117,236 ; pcmpeqw %xmm4,%xmm5 - DB 102,15,127,172,36,160,0,0,0 ; movdqa %xmm5,0xa0(%rsp) - DB 102,15,111,53,69,13,0,0 ; movdqa 0xd45(%rip),%xmm6 # 2d450 <_sk_overlay_sse2_8bit+0xef8> - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 - DB 102,15,127,132,36,192,0,0,0 ; movdqa %xmm0,0xc0(%rsp) - DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 - DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 - DB 102,68,15,239,230 ; pxor %xmm6,%xmm12 - DB 102,68,15,213,225 ; pmullw %xmm1,%xmm12 - DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,15,249,209 ; psubw %xmm1,%xmm2 - DB 102,15,127,84,36,80 ; movdqa %xmm2,0x50(%rsp) + DB 72,129,236,8,1,0,0 ; sub $0x108,%rsp + DB 102,68,15,111,242 ; movdqa %xmm2,%xmm14 + DB 102,15,239,237 ; pxor %xmm5,%xmm5 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,15,127,36,36 ; movdqa %xmm4,(%rsp) + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,15,127,100,36,48 ; movdqa %xmm4,0x30(%rsp) + DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,132,36,144,0,0,0 ; movdqa %xmm0,0x90(%rsp) + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,68,15,112,252,95 ; pshufhw $0x5f,%xmm4,%xmm15 + DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 + DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 - DB 102,69,15,213,215 ; pmullw %xmm15,%xmm10 - DB 102,15,213,224 ; pmullw %xmm0,%xmm4 - DB 102,15,127,100,36,96 ; movdqa %xmm4,0x60(%rsp) - DB 102,65,15,249,199 ; psubw %xmm15,%xmm0 - DB 102,15,127,68,36,48 ; movdqa %xmm0,0x30(%rsp) - DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,111,20,36 ; movdqa (%rsp),%xmm2 - DB 102,15,213,194 ; pmullw %xmm2,%xmm0 - DB 102,15,127,68,36,112 ; movdqa %xmm0,0x70(%rsp) + DB 102,68,15,127,148,36,160,0,0,0 ; movdqa %xmm10,0xa0(%rsp) + DB 243,15,112,204,231 ; pshufhw $0xe7,%xmm4,%xmm1 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 + DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 + DB 243,68,15,112,217,95 ; pshufhw $0x5f,%xmm1,%xmm11 DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 - DB 102,68,15,239,222 ; pxor %xmm6,%xmm11 - DB 102,69,15,213,216 ; pmullw %xmm8,%xmm11 - DB 102,69,15,111,249 ; movdqa %xmm9,%xmm15 - DB 102,69,15,249,200 ; psubw %xmm8,%xmm9 - DB 102,68,15,127,76,36,64 ; movdqa %xmm9,0x40(%rsp) - DB 102,68,15,213,194 ; pmullw %xmm2,%xmm8 - DB 102,68,15,213,248 ; pmullw %xmm0,%xmm15 - DB 102,15,249,194 ; psubw %xmm2,%xmm0 - DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,68,36,32 ; movdqa %xmm0,0x20(%rsp) + DB 102,68,15,96,221 ; punpcklbw %xmm5,%xmm11 + DB 102,68,15,127,156,36,128,0,0,0 ; movdqa %xmm11,0x80(%rsp) + DB 102,65,15,111,207 ; movdqa %xmm15,%xmm1 + DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 + DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 + DB 102,68,15,127,108,36,96 ; movdqa %xmm13,0x60(%rsp) + DB 102,68,15,96,253 ; punpcklbw %xmm5,%xmm15 + DB 102,68,15,127,180,36,240,0,0,0 ; movdqa %xmm14,0xf0(%rsp) + DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 + DB 242,65,15,112,206,231 ; pshuflw $0xe7,%xmm14,%xmm1 + DB 102,68,15,96,245 ; punpcklbw %xmm5,%xmm14 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,68,36,64 ; movdqa %xmm0,0x40(%rsp) + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,127,148,36,224,0,0,0 ; movdqa %xmm2,0xe0(%rsp) + DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 + DB 102,68,15,96,205 ; punpcklbw %xmm5,%xmm9 + DB 102,15,104,221 ; punpckhbw %xmm5,%xmm3 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,111,84,36,16 ; movdqa 0x10(%rsp),%xmm2 - DB 102,15,213,194 ; pmullw %xmm2,%xmm0 - DB 102,15,127,4,36 ; movdqa %xmm0,(%rsp) - DB 102,15,111,132,36,224,0,0,0 ; movdqa 0xe0(%rsp),%xmm0 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,127,68,36,112 ; movdqa %xmm0,0x70(%rsp) + DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 + DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 + DB 243,15,112,217,95 ; pshufhw $0x5f,%xmm1,%xmm3 + DB 242,15,112,202,231 ; pshuflw $0xe7,%xmm2,%xmm1 + DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 + DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 + DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 + DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 + DB 102,68,15,127,68,36,16 ; movdqa %xmm8,0x10(%rsp) + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,68,15,111,227 ; movdqa %xmm3,%xmm12 + DB 102,68,15,104,229 ; punpckhbw %xmm5,%xmm12 + DB 102,15,96,221 ; punpcklbw %xmm5,%xmm3 + DB 102,15,111,53,38,13,0,0 ; movdqa 0xd26(%rip),%xmm6 # 2d340 <_sk_overlay_sse2_8bit+0xe91> + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,239,214 ; pxor %xmm6,%xmm2 + DB 102,65,15,111,252 ; movdqa %xmm12,%xmm7 + DB 102,15,239,254 ; pxor %xmm6,%xmm7 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,239,206 ; pxor %xmm6,%xmm1 + DB 102,65,15,111,232 ; movdqa %xmm8,%xmm5 DB 102,15,239,238 ; pxor %xmm6,%xmm5 - DB 102,65,15,213,238 ; pmullw %xmm14,%xmm5 - DB 102,68,15,111,235 ; movdqa %xmm3,%xmm13 + DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 + DB 102,69,15,111,199 ; movdqa %xmm15,%xmm8 + DB 102,68,15,239,198 ; pxor %xmm6,%xmm8 + DB 102,69,15,111,213 ; movdqa %xmm13,%xmm10 + DB 102,68,15,239,214 ; pxor %xmm6,%xmm10 + DB 102,68,15,239,222 ; pxor %xmm6,%xmm11 + DB 102,15,239,116,36,32 ; pxor 0x20(%rsp),%xmm6 + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 + DB 102,15,253,245 ; paddw %xmm5,%xmm6 + DB 102,15,213,76,36,48 ; pmullw 0x30(%rsp),%xmm1 + DB 102,69,15,213,217 ; pmullw %xmm9,%xmm11 + DB 102,68,15,253,217 ; paddw %xmm1,%xmm11 + DB 102,68,15,127,156,36,176,0,0,0 ; movdqa %xmm11,0xb0(%rsp) + DB 102,15,111,132,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm0 + DB 102,15,213,248 ; pmullw %xmm0,%xmm7 + DB 102,15,111,108,36,64 ; movdqa 0x40(%rsp),%xmm5 + DB 102,68,15,213,213 ; pmullw %xmm5,%xmm10 + DB 102,68,15,253,215 ; paddw %xmm7,%xmm10 + DB 102,68,15,127,148,36,208,0,0,0 ; movdqa %xmm10,0xd0(%rsp) + DB 102,15,111,12,36 ; movdqa (%rsp),%xmm1 + DB 102,15,213,209 ; pmullw %xmm1,%xmm2 + DB 102,69,15,213,198 ; pmullw %xmm14,%xmm8 + DB 102,68,15,253,194 ; paddw %xmm2,%xmm8 + DB 102,68,15,127,132,36,192,0,0,0 ; movdqa %xmm8,0xc0(%rsp) + DB 102,65,15,111,214 ; movdqa %xmm14,%xmm2 + DB 102,15,253,210 ; paddw %xmm2,%xmm2 + DB 102,15,101,211 ; pcmpgtw %xmm3,%xmm2 + DB 102,15,127,84,36,80 ; movdqa %xmm2,0x50(%rsp) + DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 DB 102,65,15,249,222 ; psubw %xmm14,%xmm3 - DB 102,15,127,92,36,32 ; movdqa %xmm3,0x20(%rsp) - DB 102,65,15,111,254 ; movdqa %xmm14,%xmm7 - DB 102,15,213,250 ; pmullw %xmm2,%xmm7 - DB 102,68,15,213,232 ; pmullw %xmm0,%xmm13 - DB 102,15,249,194 ; psubw %xmm2,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,111,140,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,111,156,36,16,1,0,0 ; movdqa 0x110(%rsp),%xmm3 - DB 102,15,213,195 ; pmullw %xmm3,%xmm0 + DB 102,68,15,213,241 ; pmullw %xmm1,%xmm14 + DB 102,69,15,213,223 ; pmullw %xmm15,%xmm11 + DB 102,68,15,249,249 ; psubw %xmm1,%xmm15 + DB 102,68,15,213,251 ; pmullw %xmm3,%xmm15 + DB 102,68,15,127,60,36 ; movdqa %xmm15,(%rsp) + DB 102,15,111,221 ; movdqa %xmm5,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,65,15,101,220 ; pcmpgtw %xmm12,%xmm3 + DB 102,69,15,111,196 ; movdqa %xmm12,%xmm8 + DB 102,68,15,249,229 ; psubw %xmm5,%xmm12 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,68,15,111,108,36,96 ; movdqa 0x60(%rsp),%xmm13 + DB 102,69,15,213,197 ; pmullw %xmm13,%xmm8 + DB 102,68,15,249,232 ; psubw %xmm0,%xmm13 + DB 102,69,15,213,236 ; pmullw %xmm12,%xmm13 + DB 102,69,15,111,249 ; movdqa %xmm9,%xmm15 + DB 102,69,15,253,255 ; paddw %xmm15,%xmm15 + DB 102,68,15,101,252 ; pcmpgtw %xmm4,%xmm15 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,65,15,249,225 ; psubw %xmm9,%xmm4 + DB 102,69,15,111,209 ; movdqa %xmm9,%xmm10 + DB 102,15,111,68,36,48 ; movdqa 0x30(%rsp),%xmm0 + DB 102,68,15,213,208 ; pmullw %xmm0,%xmm10 + DB 102,15,111,188,36,128,0,0,0 ; movdqa 0x80(%rsp),%xmm7 + DB 102,15,213,207 ; pmullw %xmm7,%xmm1 + DB 102,15,249,248 ; psubw %xmm0,%xmm7 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,68,15,111,76,36,112 ; movdqa 0x70(%rsp),%xmm9 + DB 102,65,15,111,225 ; movdqa %xmm9,%xmm4 + DB 102,15,253,228 ; paddw %xmm4,%xmm4 + DB 102,15,111,68,36,16 ; movdqa 0x10(%rsp),%xmm0 + DB 102,15,101,224 ; pcmpgtw %xmm0,%xmm4 + DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12 + DB 102,65,15,249,193 ; psubw %xmm9,%xmm0 DB 102,15,127,68,36,16 ; movdqa %xmm0,0x10(%rsp) - DB 102,15,111,132,36,0,1,0,0 ; movdqa 0x100(%rsp),%xmm0 - DB 102,15,239,240 ; pxor %xmm0,%xmm6 - DB 102,15,111,164,36,144,0,0,0 ; movdqa 0x90(%rsp),%xmm4 - DB 102,15,213,244 ; pmullw %xmm4,%xmm6 - DB 102,68,15,111,241 ; movdqa %xmm1,%xmm14 - DB 102,15,249,204 ; psubw %xmm4,%xmm1 - DB 102,15,213,227 ; pmullw %xmm3,%xmm4 - DB 102,68,15,213,240 ; pmullw %xmm0,%xmm14 - DB 102,15,249,195 ; psubw %xmm3,%xmm0 - DB 102,15,213,193 ; pmullw %xmm1,%xmm0 - DB 102,15,213,84,36,32 ; pmullw 0x20(%rsp),%xmm2 - DB 102,68,15,213,76,36,64 ; pmullw 0x40(%rsp),%xmm9 - DB 102,15,111,76,36,48 ; movdqa 0x30(%rsp),%xmm1 - DB 102,15,213,76,36,80 ; pmullw 0x50(%rsp),%xmm1 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,68,15,249,240 ; psubw %xmm0,%xmm14 + DB 102,15,111,148,36,160,0,0,0 ; movdqa 0xa0(%rsp),%xmm2 + DB 102,68,15,213,202 ; pmullw %xmm2,%xmm9 + DB 102,15,111,68,36,32 ; movdqa 0x20(%rsp),%xmm0 + DB 102,68,15,213,224 ; pmullw %xmm0,%xmm12 + DB 102,15,249,194 ; psubw %xmm2,%xmm0 + DB 102,15,213,68,36,16 ; pmullw 0x10(%rsp),%xmm0 + DB 102,15,111,20,36 ; movdqa (%rsp),%xmm2 DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,68,15,249,234 ; psubw %xmm2,%xmm13 - DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 - DB 102,69,15,249,249 ; psubw %xmm9,%xmm15 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,76,36,96 ; movdqa 0x60(%rsp),%xmm1 - DB 102,15,249,200 ; psubw %xmm0,%xmm1 - DB 102,69,15,253,210 ; paddw %xmm10,%xmm10 - DB 102,15,111,132,36,160,0,0,0 ; movdqa 0xa0(%rsp),%xmm0 - DB 102,68,15,219,208 ; pand %xmm0,%xmm10 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,65,15,235,194 ; por %xmm10,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,69,15,253,192 ; paddw %xmm8,%xmm8 - DB 102,15,111,132,36,176,0,0,0 ; movdqa 0xb0(%rsp),%xmm0 - DB 102,68,15,219,192 ; pand %xmm0,%xmm8 - DB 102,65,15,223,199 ; pandn %xmm15,%xmm0 - DB 102,65,15,235,192 ; por %xmm8,%xmm0 - DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 + DB 102,68,15,249,218 ; psubw %xmm2,%xmm11 + DB 102,69,15,253,237 ; paddw %xmm13,%xmm13 + DB 102,69,15,249,197 ; psubw %xmm13,%xmm8 DB 102,15,253,255 ; paddw %xmm7,%xmm7 - DB 102,15,111,132,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm0 - DB 102,15,219,248 ; pand %xmm0,%xmm7 - DB 102,65,15,223,197 ; pandn %xmm13,%xmm0 - DB 102,15,235,199 ; por %xmm7,%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,15,111,132,36,240,0,0,0 ; movdqa 0xf0(%rsp),%xmm0 - DB 102,15,219,224 ; pand %xmm0,%xmm4 + DB 102,15,249,207 ; psubw %xmm7,%xmm1 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,68,15,249,224 ; psubw %xmm0,%xmm12 + DB 102,69,15,253,246 ; paddw %xmm14,%xmm14 + DB 102,15,111,68,36,80 ; movdqa 0x50(%rsp),%xmm0 + DB 102,68,15,219,216 ; pand %xmm0,%xmm11 DB 102,65,15,223,198 ; pandn %xmm14,%xmm0 - DB 102,15,235,196 ; por %xmm4,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,68,15,253,164,36,192,0,0,0 ; paddw 0xc0(%rsp),%xmm12 - DB 102,68,15,253,92,36,112 ; paddw 0x70(%rsp),%xmm11 - DB 102,15,253,44,36 ; paddw (%rsp),%xmm5 - DB 102,15,253,116,36,16 ; paddw 0x10(%rsp),%xmm6 - DB 102,15,111,5,79,19,0,0 ; movdqa 0x134f(%rip),%xmm0 # 2dc90 <_sk_overlay_sse2_8bit+0x1738> - DB 102,68,15,253,224 ; paddw %xmm0,%xmm12 - DB 102,68,15,253,227 ; paddw %xmm3,%xmm12 - DB 102,68,15,253,216 ; paddw %xmm0,%xmm11 - DB 102,69,15,253,216 ; paddw %xmm8,%xmm11 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,15,253,234 ; paddw %xmm2,%xmm5 + DB 102,68,15,235,216 ; por %xmm0,%xmm11 + DB 102,15,253,237 ; paddw %xmm5,%xmm5 + DB 102,68,15,219,195 ; pand %xmm3,%xmm8 + DB 102,15,223,221 ; pandn %xmm5,%xmm3 + DB 102,65,15,235,216 ; por %xmm8,%xmm3 + DB 102,69,15,253,210 ; paddw %xmm10,%xmm10 + DB 102,65,15,219,207 ; pand %xmm15,%xmm1 + DB 102,69,15,223,250 ; pandn %xmm10,%xmm15 + DB 102,65,15,235,207 ; por %xmm15,%xmm1 + DB 102,69,15,253,201 ; paddw %xmm9,%xmm9 + DB 102,68,15,219,228 ; pand %xmm4,%xmm12 + DB 102,65,15,223,225 ; pandn %xmm9,%xmm4 + DB 102,65,15,235,228 ; por %xmm12,%xmm4 + DB 102,15,111,5,83,19,0,0 ; movdqa 0x1353(%rip),%xmm0 # 2db90 <_sk_overlay_sse2_8bit+0x16e1> + DB 102,15,111,148,36,192,0,0,0 ; movdqa 0xc0(%rsp),%xmm2 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,65,15,253,211 ; paddw %xmm11,%xmm2 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,15,111,148,36,208,0,0,0 ; movdqa 0xd0(%rsp),%xmm2 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,15,253,211 ; paddw %xmm3,%xmm2 + DB 102,15,111,156,36,176,0,0,0 ; movdqa 0xb0(%rsp),%xmm3 + DB 102,15,253,216 ; paddw %xmm0,%xmm3 + DB 102,15,253,217 ; paddw %xmm1,%xmm3 DB 102,15,253,240 ; paddw %xmm0,%xmm6 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,15,111,5,51,19,0,0 ; movdqa 0x1333(%rip),%xmm0 # 2dca0 <_sk_overlay_sse2_8bit+0x1748> + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,15,111,5,27,19,0,0 ; movdqa 0x131b(%rip),%xmm0 # 2dba0 <_sk_overlay_sse2_8bit+0x16f1> DB 102,15,228,240 ; pmulhuw %xmm0,%xmm6 - DB 102,15,228,232 ; pmulhuw %xmm0,%xmm5 - DB 102,68,15,228,216 ; pmulhuw %xmm0,%xmm11 - DB 102,68,15,228,224 ; pmulhuw %xmm0,%xmm12 - DB 102,65,15,113,211,7 ; psrlw $0x7,%xmm11 - DB 102,65,15,113,212,7 ; psrlw $0x7,%xmm12 - DB 102,69,15,103,227 ; packuswb %xmm11,%xmm12 DB 102,15,113,214,7 ; psrlw $0x7,%xmm6 - DB 102,15,113,213,7 ; psrlw $0x7,%xmm5 - DB 102,15,103,238 ; packuswb %xmm6,%xmm5 + DB 102,15,228,216 ; pmulhuw %xmm0,%xmm3 + DB 102,15,113,211,7 ; psrlw $0x7,%xmm3 + DB 102,15,103,222 ; packuswb %xmm6,%xmm3 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,228,208 ; pmulhuw %xmm0,%xmm2 + DB 102,15,111,221 ; movdqa %xmm5,%xmm3 + DB 102,15,228,216 ; pmulhuw %xmm0,%xmm3 + DB 102,15,113,210,7 ; psrlw $0x7,%xmm2 + DB 102,15,113,211,7 ; psrlw $0x7,%xmm3 + DB 102,15,103,218 ; packuswb %xmm2,%xmm3 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,148,36,48,1,0,0 ; movaps 0x130(%rsp),%xmm2 - DB 15,40,156,36,32,1,0,0 ; movaps 0x120(%rsp),%xmm3 - DB 102,65,15,111,196 ; movdqa %xmm12,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 72,129,196,72,1,0,0 ; add $0x148,%rsp + DB 15,40,148,36,240,0,0,0 ; movaps 0xf0(%rsp),%xmm2 + DB 15,40,156,36,224,0,0,0 ; movaps 0xe0(%rsp),%xmm3 + DB 72,129,196,8,1,0,0 ; add $0x108,%rsp DB 255,224 ; jmpq *%rax ALIGN 32 @@ -42782,16 +42688,16 @@ ALIGN 32 DB 0,1 ; add %al,(%rcx) DB 255 ; (bad) DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02ca48 <_sk_overlay_sse2_8bit+0xa0004f0> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02c948 <_sk_overlay_sse2_8bit+0xa000499> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202ca50 <_sk_overlay_sse2_8bit+0x120004f8> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202c950 <_sk_overlay_sse2_8bit+0x120004a1> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02ca58 <_sk_overlay_sse2_8bit+0x1a000500> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02c958 <_sk_overlay_sse2_8bit+0x1a0004a9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302ca60 <_sk_overlay_sse2_8bit+0x3000508> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302c960 <_sk_overlay_sse2_8bit+0x30004b1> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -42834,16 +42740,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02caa8 <_sk_overlay_sse2_8bit+0xa000550> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02c9a8 <_sk_overlay_sse2_8bit+0xa0004f9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cab0 <_sk_overlay_sse2_8bit+0x12000558> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202c9b0 <_sk_overlay_sse2_8bit+0x12000501> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cab8 <_sk_overlay_sse2_8bit+0x1a000560> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02c9b8 <_sk_overlay_sse2_8bit+0x1a000509> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cac0 <_sk_overlay_sse2_8bit+0x3000568> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302c9c0 <_sk_overlay_sse2_8bit+0x3000511> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -42886,16 +42792,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cb08 <_sk_overlay_sse2_8bit+0xa0005b0> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02ca08 <_sk_overlay_sse2_8bit+0xa000559> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cb10 <_sk_overlay_sse2_8bit+0x120005b8> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202ca10 <_sk_overlay_sse2_8bit+0x12000561> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cb18 <_sk_overlay_sse2_8bit+0x1a0005c0> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02ca18 <_sk_overlay_sse2_8bit+0x1a000569> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cb20 <_sk_overlay_sse2_8bit+0x30005c8> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302ca20 <_sk_overlay_sse2_8bit+0x3000571> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -42938,16 +42844,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cb68 <_sk_overlay_sse2_8bit+0xa000610> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02ca68 <_sk_overlay_sse2_8bit+0xa0005b9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cb70 <_sk_overlay_sse2_8bit+0x12000618> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202ca70 <_sk_overlay_sse2_8bit+0x120005c1> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cb78 <_sk_overlay_sse2_8bit+0x1a000620> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02ca78 <_sk_overlay_sse2_8bit+0x1a0005c9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cb80 <_sk_overlay_sse2_8bit+0x3000628> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302ca80 <_sk_overlay_sse2_8bit+0x30005d1> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -42990,16 +42896,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cbc8 <_sk_overlay_sse2_8bit+0xa000670> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cac8 <_sk_overlay_sse2_8bit+0xa000619> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cbd0 <_sk_overlay_sse2_8bit+0x12000678> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cad0 <_sk_overlay_sse2_8bit+0x12000621> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cbd8 <_sk_overlay_sse2_8bit+0x1a000680> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cad8 <_sk_overlay_sse2_8bit+0x1a000629> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cbe0 <_sk_overlay_sse2_8bit+0x3000688> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cae0 <_sk_overlay_sse2_8bit+0x3000631> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -43042,16 +42948,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cc28 <_sk_overlay_sse2_8bit+0xa0006d0> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cb28 <_sk_overlay_sse2_8bit+0xa000679> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cc30 <_sk_overlay_sse2_8bit+0x120006d8> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cb30 <_sk_overlay_sse2_8bit+0x12000681> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cc38 <_sk_overlay_sse2_8bit+0x1a0006e0> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cb38 <_sk_overlay_sse2_8bit+0x1a000689> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cc40 <_sk_overlay_sse2_8bit+0x30006e8> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cb40 <_sk_overlay_sse2_8bit+0x3000691> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -43094,16 +43000,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cc88 <_sk_overlay_sse2_8bit+0xa000730> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cb88 <_sk_overlay_sse2_8bit+0xa0006d9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cc90 <_sk_overlay_sse2_8bit+0x12000738> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cb90 <_sk_overlay_sse2_8bit+0x120006e1> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cc98 <_sk_overlay_sse2_8bit+0x1a000740> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cb98 <_sk_overlay_sse2_8bit+0x1a0006e9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cca0 <_sk_overlay_sse2_8bit+0x3000748> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cba0 <_sk_overlay_sse2_8bit+0x30006f1> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -43146,16 +43052,16 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cce8 <_sk_overlay_sse2_8bit+0xa000790> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02cbe8 <_sk_overlay_sse2_8bit+0xa000739> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202ccf0 <_sk_overlay_sse2_8bit+0x12000798> + DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 1202cbf0 <_sk_overlay_sse2_8bit+0x12000741> DB 255 ; (bad) DB 255 ; (bad) - DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02ccf8 <_sk_overlay_sse2_8bit+0x1a0007a0> + DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a02cbf8 <_sk_overlay_sse2_8bit+0x1a000749> DB 255 ; (bad) DB 255 ; (bad) - DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cd00 <_sk_overlay_sse2_8bit+0x30007a8> + DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 302cc00 <_sk_overlay_sse2_8bit+0x3000751> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -43424,87 +43330,73 @@ ALIGN 32 DB 5,4,7,10,9 ; add $0x90a0704,%eax DB 8,11 ; or %cl,(%rbx) DB 14 ; (bad) - DB 13,12,15,248,0 ; or $0xf80f0c,%eax - DB 0,0 ; add %al,(%rax) - DB 248 ; clc - DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al - DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al - DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al - DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al - DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al - DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al - DB 0,0 ; add %al,(%rax) - DB 0,0 ; add %al,(%rax) - DB 252 ; cld - DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 13,12,15,0,1 ; or $0x1000f0c,%eax + DB 4,5 ; add $0x5,%al + DB 8,9 ; or %cl,(%rcx) + DB 12,13 ; or $0xd,%al + DB 8,9 ; or %cl,(%rcx) + DB 12,13 ; or $0xd,%al + DB 12,13 ; or $0xd,%al + DB 14 ; (bad) + DB 15,16,17 ; movups (%rcx),%xmm2 + DB 20,21 ; adc $0x15,%al + DB 24,25 ; sbb %bl,(%rcx) + DB 28,29 ; sbb $0x1d,%al + DB 24,25 ; sbb %bl,(%rcx) + DB 28,29 ; sbb $0x1d,%al + DB 28,29 ; sbb $0x1d,%al + DB 30 ; (bad) + DB 31 ; (bad) + DB 3,255 ; add %edi,%edi + DB 7 ; (bad) + DB 255,11 ; decl (%rbx) + DB 255,15 ; decl (%rdi) + DB 255,11 ; decl (%rbx) + DB 255,15 ; decl (%rdi) + DB 255,15 ; decl (%rdi) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,19 ; callq *(%rbx) + DB 255,23 ; callq *(%rdi) + DB 255,27 ; lcall *(%rbx) + DB 255,31 ; lcall *(%rdi) + DB 255,27 ; lcall *(%rbx) + DB 255,31 ; lcall *(%rdi) + DB 255,31 ; lcall *(%rdi) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,252 ; add %bh,%ah + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 248 ; clc + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 0,248 ; add %bh,%al + DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 1,4,5,8,9,12,13 ; add %eax,0xd0c0908(,%rax,1) - DB 128,128,128,128,128,128,128 ; addb $0x80,-0x7f7f7f80(%rax) - DB 128,0,1 ; addb $0x1,(%rax) - DB 4,5 ; add $0x5,%al - DB 8,9 ; or %cl,(%rcx) - DB 12,13 ; or $0xd,%al - DB 128,128,128,128,128,128,128 ; addb $0x80,-0x7f7f7f80(%rax) - DB 128,3,255 ; addb $0xff,(%rbx) - DB 7 ; (bad) - DB 255,11 ; decl (%rbx) - DB 255,15 ; decl (%rdi) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,19 ; callq *(%rbx) - DB 255,23 ; callq *(%rdi) - DB 255,27 ; lcall *(%rbx) - DB 255,31 ; lcall *(%rdi) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) DB 255,3 ; incl (%rbx) DB 3,3 ; add (%rbx),%eax DB 3,7 ; add (%rdi),%eax @@ -44001,14 +43893,14 @@ ALIGN 4 DB 252 ; cld DB 190,0,0,128,191 ; mov $0xbf800000,%esi DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 2d30c <.literal4+0x64> + DB 224,64 ; loopne 2d1ec <.literal4+0x64> DB 154 ; (bad) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 2d315 <.literal4+0x6d> + DB 71,225,61 ; rex.RXB loope 2d1f5 <.literal4+0x6d> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2d31f <.literal4+0x77> + DB 127,67 ; jg 2d1ff <.literal4+0x77> DB 145 ; xchg %eax,%ecx DB 131,158,61,92,143,50,63 ; sbbl $0x3f,0x328f5c3d(%rsi) DB 10,215 ; or %bh,%dl @@ -44041,7 +43933,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 2d335 <.literal4+0x8d> + DB 224,7 ; loopne 2d215 <.literal4+0x8d> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -44052,8 +43944,8 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 0,52,255 ; add %dh,(%rdi,%rdi,8) DB 255 ; (bad) - DB 127,0 ; jg 2d348 <.literal4+0xa0> - DB 119,115 ; ja 2d3bd <.literal4+0x115> + DB 127,0 ; jg 2d228 <.literal4+0xa0> + DB 119,115 ; ja 2d29d <.literal4+0x115> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -44076,12 +43968,12 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 2d3d3 <.literal4+0x12b> + DB 118,63 ; jbe 2d2b3 <.literal4+0x12b> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 2d3e2 <.literal4+0x13a> + DB 124,66 ; jl 2d2c2 <.literal4+0x13a> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -44091,9 +43983,9 @@ ALIGN 4 DB 137,136,136,59,15,0 ; mov %ecx,0xf3b88(%rax) DB 0,0 ; add %al,(%rax) DB 137,136,136,61,0,0 ; mov %ecx,0x3d88(%rax) - DB 112,65 ; jo 2d405 <.literal4+0x15d> + DB 112,65 ; jo 2d2e5 <.literal4+0x15d> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 2d40f <.literal4+0x167> + DB 127,71 ; jg 2d2ef <.literal4+0x167> DB 208 ; (bad) DB 179,89 ; mov $0x59,%bl DB 62,89 ; ds pop %rcx @@ -44120,23 +44012,34 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 128,0,0 ; addb $0x0,(%rax) DB 0,191,0,0,192,191 ; add %bh,-0x40400000(%rdi) - DB 114,28 ; jb 2d42e <.literal4+0x186> + DB 114,28 ; jb 2d30e <.literal4+0x186> DB 199 ; (bad) DB 62,0,0 ; add %al,%ds:(%rax) DB 192,63,85 ; sarb $0x55,(%rdi) DB 85 ; push %rbp DB 149 ; xchg %eax,%ebp DB 191,57,142,99,61 ; mov $0x3d638e39,%edi - DB 114,249 ; jb 2d41b <.literal4+0x173> - DB 127,63 ; jg 2d463 <_sk_overlay_sse2_8bit+0xf0b> + DB 114,249 ; jb 2d2fb <.literal4+0x173> + DB 127,63 ; jg 2d343 <_sk_overlay_sse2_8bit+0xe94> DB 3,0 ; add (%rax),%eax DB 0,0 ; add %al,(%rax) - DB 0,128,0,0,0,4 ; add %al,0x4000000(%rax) - DB 0,128,0,0,0,56 ; add %al,0x38000000(%rax) + DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 0,0 ; add %al,(%rax) DB 0,0 ; add %al,(%rax) - DB 128,56,0 ; cmpb $0x0,(%rax) - DB 64,254 ; rex (bad) - DB 255,7 ; incl (%rdi) + DB 56,255 ; cmp %bh,%bh + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 2d315 <.literal4+0x18d> + DB 0,128,56,0,64,254 ; add %al,-0x1bfffc8(%rax) + DB 255 ; (bad) + DB 248 ; clc + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 252 ; cld + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 248 ; clc + DB 0,7 ; add %al,(%rdi) DB 0,7 ; add %al,(%rdi) DB 0,0 ; add %al,(%rax) DB 12,0 ; or $0x0,%al @@ -44247,11 +44150,11 @@ ALIGN 16 DB 0,128,191,0,0,128 ; add %al,-0x7fffff41(%rax) DB 191,0,0,224,64 ; mov $0x40e00000,%edi DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 2d578 <.literal16+0x128> + DB 224,64 ; loopne 2d468 <.literal16+0x128> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 2d57c <.literal16+0x12c> + DB 224,64 ; loopne 2d46c <.literal16+0x12c> DB 0,0 ; add %al,(%rax) - DB 224,64 ; loopne 2d580 <.literal16+0x130> + DB 224,64 ; loopne 2d470 <.literal16+0x130> DB 154 ; (bad) DB 153 ; cltd DB 153 ; cltd @@ -44271,13 +44174,13 @@ ALIGN 16 DB 10,23 ; or (%rdi),%dl DB 63 ; (bad) DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 2d5a1 <.literal16+0x151> + DB 71,225,61 ; rex.RXB loope 2d491 <.literal16+0x151> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 2d5a5 <.literal16+0x155> + DB 71,225,61 ; rex.RXB loope 2d495 <.literal16+0x155> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 2d5a9 <.literal16+0x159> + DB 71,225,61 ; rex.RXB loope 2d499 <.literal16+0x159> DB 174 ; scas %es:(%rdi),%al - DB 71,225,61 ; rex.RXB loope 2d5ad <.literal16+0x15d> + DB 71,225,61 ; rex.RXB loope 2d49d <.literal16+0x15d> DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -44288,10 +44191,10 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 1,255 ; add %edi,%edi DB 255 ; (bad) - DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02d588 <_sk_overlay_sse2_8bit+0xa001030> + DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a02d478 <_sk_overlay_sse2_8bit+0xa000fc9> DB 255 ; (bad) DB 255 ; (bad) - DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 302d590 <_sk_overlay_sse2_8bit+0x3001038> + DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 302d480 <_sk_overlay_sse2_8bit+0x3000fd1> DB 255 ; (bad) DB 255 ; (bad) DB 255,6 ; incl (%rsi) @@ -44306,11 +44209,11 @@ ALIGN 16 DB 255,0 ; incl (%rax) DB 0,127,67 ; add %bh,0x43(%rdi) DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2d5eb <.literal16+0x19b> + DB 127,67 ; jg 2d4db <.literal16+0x19b> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2d5ef <.literal16+0x19f> + DB 127,67 ; jg 2d4df <.literal16+0x19f> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2d5f3 <.literal16+0x1a3> + DB 127,67 ; jg 2d4e3 <.literal16+0x1a3> DB 145 ; xchg %eax,%ecx DB 131,158,61,145,131,158,61 ; sbbl $0x3d,-0x617c6ec3(%rsi) DB 145 ; xchg %eax,%ecx @@ -44461,13 +44364,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%rdi) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 2d6f9 <.literal16+0x2a9> + DB 224,7 ; loopne 2d5e9 <.literal16+0x2a9> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 2d6fd <.literal16+0x2ad> + DB 224,7 ; loopne 2d5ed <.literal16+0x2ad> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 2d701 <.literal16+0x2b1> + DB 224,7 ; loopne 2d5f1 <.literal16+0x2b1> DB 0,0 ; add %al,(%rax) - DB 224,7 ; loopne 2d705 <.literal16+0x2b5> + DB 224,7 ; loopne 2d5f5 <.literal16+0x2b5> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -44512,12 +44415,6 @@ ALIGN 16 DB 55 ; (bad) DB 128,0,128 ; addb $0x80,(%rax) DB 55 ; (bad) - DB 8,128,10,128,12,128 ; or %al,-0x7ff37ff6(%rax) - DB 14 ; (bad) - DB 128,0,0 ; addb $0x0,(%rax) - DB 0,0 ; add %al,(%rax) - DB 0,0 ; add %al,(%rax) - DB 0,0 ; add %al,(%rax) DB 0,0 ; add %al,(%rax) DB 0,52,0 ; add %dh,(%rax,%rax,1) DB 0,0 ; add %al,(%rax) @@ -44527,17 +44424,17 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 2d784 <.literal16+0x334> + DB 127,0 ; jg 2d664 <.literal16+0x324> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 2d788 <.literal16+0x338> + DB 127,0 ; jg 2d668 <.literal16+0x328> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 2d78c <.literal16+0x33c> + DB 127,0 ; jg 2d66c <.literal16+0x32c> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg 2d790 <.literal16+0x340> - DB 119,115 ; ja 2d805 <.literal16+0x3b5> + DB 127,0 ; jg 2d670 <.literal16+0x330> + DB 119,115 ; ja 2d6e5 <.literal16+0x3a5> DB 248 ; clc DB 194,119,115 ; retq $0x7377 DB 248 ; clc @@ -44548,7 +44445,7 @@ ALIGN 16 DB 194,117,191 ; retq $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) - DB 117,191 ; jne 2d769 <.literal16+0x319> + DB 117,191 ; jne 2d649 <.literal16+0x309> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; (bad) DB 249 ; stc @@ -44560,7 +44457,7 @@ ALIGN 16 DB 249 ; stc DB 68,180,62 ; rex.R mov $0x3e,%spl DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9 - DB 233,220,63,163,233 ; jmpq ffffffffe9a617aa <_sk_overlay_sse2_8bit+0xffffffffe9a35252> + DB 233,220,63,163,233 ; jmpq ffffffffe9a6168a <_sk_overlay_sse2_8bit+0xffffffffe9a351db> DB 220,63 ; fdivrl (%rdi) DB 81 ; push %rcx DB 140,242 ; mov %?,%edx @@ -44650,16 +44547,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%rbx) DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 2d8f3 <.literal16+0x4a3> + DB 118,63 ; jbe 2d7d3 <.literal16+0x493> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 2d8f7 <.literal16+0x4a7> + DB 118,63 ; jbe 2d7d7 <.literal16+0x497> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 2d8fb <.literal16+0x4ab> + DB 118,63 ; jbe 2d7db <.literal16+0x49b> DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 2d8ff <.literal16+0x4af> + DB 118,63 ; jbe 2d7df <.literal16+0x49f> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 246,64,83,63 ; testb $0x3f,0x53(%rax) @@ -44680,13 +44577,13 @@ ALIGN 16 DB 65,0,0 ; add %al,(%r8) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 2d936 <.literal16+0x4e6> + DB 124,66 ; jl 2d816 <.literal16+0x4d6> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 2d93a <.literal16+0x4ea> + DB 124,66 ; jl 2d81a <.literal16+0x4da> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 2d93e <.literal16+0x4ee> + DB 124,66 ; jl 2d81e <.literal16+0x4de> DB 0,0 ; add %al,(%rax) - DB 124,66 ; jl 2d942 <.literal16+0x4f2> + DB 124,66 ; jl 2d822 <.literal16+0x4e2> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 0,240 ; add %dh,%al @@ -44732,18 +44629,27 @@ ALIGN 16 DB 136,136,61,137,136,136 ; mov %cl,-0x777776c3(%rax) DB 61,137,136,136,61 ; cmp $0x3d888889,%eax DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 2d9c5 <.literal16+0x575> + DB 112,65 ; jo 2d8a5 <.literal16+0x565> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 2d9c9 <.literal16+0x579> + DB 112,65 ; jo 2d8a9 <.literal16+0x569> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 2d9cd <.literal16+0x57d> + DB 112,65 ; jo 2d8ad <.literal16+0x56d> DB 0,0 ; add %al,(%rax) - DB 112,65 ; jo 2d9d1 <.literal16+0x581> + DB 112,65 ; jo 2d8b1 <.literal16+0x571> DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) - DB 0,4,0 ; add %al,(%rax,%rax,1) + DB 255 ; (bad) + DB 127,0 ; jg 2d883 <.literal16+0x543> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg 2d887 <.literal16+0x547> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg 2d88b <.literal16+0x54b> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg 2d88f <.literal16+0x54f> + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al DB 0,0 ; add %al,(%rax) DB 4,0 ; add $0x0,%al DB 0,0 ; add %al,(%rax) @@ -44763,9 +44669,19 @@ ALIGN 16 DB 128,0,0 ; addb $0x0,(%rax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax) DB 0,0 ; add %al,(%rax) - DB 0,128,0,0,128,56 ; add %al,0x38800000(%rax) - DB 0,0 ; add %al,(%rax) - DB 128,56,0 ; cmpb $0x0,(%rax) + DB 0,128,255,255,255,127 ; add %al,0x7fffffff(%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,255 ; jg 2d8c8 <.literal16+0x588> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,255 ; jg 2d8cc <.literal16+0x58c> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg 2d8d1 <.literal16+0x591> + DB 0,128,56,0,0,128 ; add %al,-0x7fffffc8(%rax) + DB 56,0 ; cmp %al,(%rax) DB 0,128,56,0,0,128 ; add %al,-0x7fffffc8(%rax) DB 56,0 ; cmp %al,(%rax) DB 64,254 ; rex (bad) @@ -44777,13 +44693,13 @@ ALIGN 16 DB 64,254 ; rex (bad) DB 255,0 ; incl (%rax) DB 255 ; (bad) - DB 127,71 ; jg 2da3b <.literal16+0x5eb> + DB 127,71 ; jg 2d93b <.literal16+0x5fb> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 2da3f <.literal16+0x5ef> + DB 127,71 ; jg 2d93f <.literal16+0x5ff> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 2da43 <.literal16+0x5f3> + DB 127,71 ; jg 2d943 <.literal16+0x603> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 2da47 <.literal16+0x5f7> + DB 127,71 ; jg 2d947 <.literal16+0x607> DB 208 ; (bad) DB 179,89 ; mov $0x59,%bl DB 62,208 ; ds (bad) @@ -44813,10 +44729,10 @@ ALIGN 16 DB 61,152,221,147,61 ; cmp $0x3d93dd98,%eax DB 152 ; cwtl DB 221,147,61,45,16,17 ; fstl 0x11102d3d(%rbx) - DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc2eb4a <_sk_overlay_sse2_8bit+0x2dc025f2> + DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110(%rip) # 2dc2ea4a <_sk_overlay_sse2_8bit+0x2dc0259b> DB 17,192 ; adc %eax,%eax DB 45,16,17,192,18 ; sub $0x12c01110,%eax - DB 120,57 ; js 2da7c <.literal16+0x62c> + DB 120,57 ; js 2d97c <.literal16+0x63c> DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil DB 64,18,120,57 ; adc 0x39(%rax),%dil @@ -44881,13 +44797,13 @@ ALIGN 16 DB 192,191,0,0,192,191,0 ; sarb $0x0,-0x40400000(%rdi) DB 0,192 ; add %al,%al DB 191,0,0,192,191 ; mov $0xbfc00000,%edi - DB 114,28 ; jb 2db2e <.literal16+0x6de> + DB 114,28 ; jb 2da2e <.literal16+0x6ee> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 2db32 <.literal16+0x6e2> + DB 62,114,28 ; jb,pt 2da32 <.literal16+0x6f2> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 2db36 <.literal16+0x6e6> + DB 62,114,28 ; jb,pt 2da36 <.literal16+0x6f6> DB 199 ; (bad) - DB 62,114,28 ; jb,pt 2db3a <.literal16+0x6ea> + DB 62,114,28 ; jb,pt 2da3a <.literal16+0x6fa> DB 199 ; (bad) DB 62,85 ; ds push %rbp DB 85 ; push %rbp @@ -44906,15 +44822,15 @@ ALIGN 16 DB 0,192 ; add %al,%al DB 63 ; (bad) DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%rsi) - DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d666985 <_sk_overlay_sse2_8bit+0x3d63a42d> + DB 99,61,57,142,99,61 ; movslq 0x3d638e39(%rip),%edi # 3d666885 <_sk_overlay_sse2_8bit+0x3d63a3d6> DB 57,142,99,61,114,249 ; cmp %ecx,-0x68dc29d(%rsi) - DB 127,63 ; jg 2db93 <.literal16+0x743> - DB 114,249 ; jb 2db4f <.literal16+0x6ff> - DB 127,63 ; jg 2db97 <.literal16+0x747> - DB 114,249 ; jb 2db53 <.literal16+0x703> - DB 127,63 ; jg 2db9b <.literal16+0x74b> - DB 114,249 ; jb 2db57 <.literal16+0x707> - DB 127,63 ; jg 2db9f <.literal16+0x74f> + DB 127,63 ; jg 2da93 <.literal16+0x753> + DB 114,249 ; jb 2da4f <.literal16+0x70f> + DB 127,63 ; jg 2da97 <.literal16+0x757> + DB 114,249 ; jb 2da53 <.literal16+0x713> + DB 127,63 ; jg 2da9b <.literal16+0x75b> + DB 114,249 ; jb 2da57 <.literal16+0x717> + DB 127,63 ; jg 2da9f <.literal16+0x75f> DB 3,0 ; add (%rax),%eax DB 0,0 ; add %al,(%rax) DB 3,0 ; add (%rax),%eax @@ -44923,12 +44839,31 @@ ALIGN 16 DB 0,0 ; add %al,(%rax) DB 3,0 ; add (%rax),%eax DB 0,0 ; add %al,(%rax) - DB 0,4,0 ; add %al,(%rax,%rax,1) - DB 128,0,4 ; addb $0x4,(%rax) - DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) - DB 0,4,0 ; add %al,(%rax,%rax,1) - DB 128,0,2 ; addb $0x2,(%rax) - DB 4,6 ; add $0x6,%al + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 0,255 ; add %bh,%bh + DB 255,0 ; incl (%rax) + DB 0,255 ; add %bh,%bh + DB 255,0 ; incl (%rax) + DB 0,255 ; add %bh,%bh + DB 255,0 ; incl (%rax) + DB 0,255 ; add %bh,%bh + DB 255,0 ; incl (%rax) + DB 2,4,6 ; add (%rsi,%rax,1),%al DB 8,10 ; or %cl,(%rdx) DB 12,14 ; or $0xe,%al DB 0,0 ; add %al,(%rax) @@ -45007,8 +44942,19 @@ ALIGN 16 DB 12,0 ; or $0x0,%al DB 0,0 ; add %al,(%rax) DB 12,0 ; or $0x0,%al - DB 0,3 ; add %al,(%rbx) - DB 255,7 ; incl (%rdi) + DB 0,255 ; add %bh,%bh + DB 255 ; (bad) + DB 255,8 ; decl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,10 ; decl (%rdx) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,12,255 ; decl (%rdi,%rdi,8) + DB 255 ; (bad) + DB 255,14 ; decl (%rsi) + DB 3,255 ; add %edi,%edi + DB 7 ; (bad) DB 255,11 ; decl (%rbx) DB 255,15 ; decl (%rdi) DB 255,11 ; decl (%rbx) @@ -45024,15 +44970,7 @@ ALIGN 16 DB 1,1 ; add %eax,(%rcx) DB 0,1 ; add %al,(%rcx) DB 1,1 ; add %eax,(%rcx) - DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) - DB 255,0 ; incl (%rax) - DB 255,0 ; incl (%rax) - DB 255,0 ; incl (%rax) - DB 255,0 ; incl (%rax) - DB 255,0 ; incl (%rax) - DB 255,0 ; incl (%rax) - DB 255,8 ; decl (%rax) + DB 0,8 ; add %cl,(%rax) DB 8,8 ; or %cl,(%rax) DB 8,10 ; or %cl,(%rdx) DB 10,10 ; or (%rdx),%cl @@ -45052,12 +44990,8 @@ ALIGN 16 DB 6 ; (bad) DB 6 ; (bad) DB 6 ; (bad) - DB 0,128,2,128,4,128 ; add %al,-0x7ffb7ffe(%rax) - DB 6 ; (bad) - DB 128,4,128,5 ; addb $0x5,(%rax,%rax,4) - DB 128,6,128 ; addb $0x80,(%rsi) - DB 7 ; (bad) - DB 128,255,255 ; cmp $0xff,%bh + DB 255 ; (bad) + DB 255 ; (bad) DB 255,0 ; incl (%rax) DB 255 ; (bad) DB 255 ; (bad) @@ -45068,14 +45002,14 @@ ALIGN 16 DB 255 ; (bad) DB 255 ; (bad) DB 255,0 ; incl (%rax) - DB 127,0 ; jg 2dc92 <.literal16+0x842> - DB 127,0 ; jg 2dc94 <.literal16+0x844> - DB 127,0 ; jg 2dc96 <.literal16+0x846> - DB 127,0 ; jg 2dc98 <.literal16+0x848> - DB 127,0 ; jg 2dc9a <.literal16+0x84a> - DB 127,0 ; jg 2dc9c <.literal16+0x84c> - DB 127,0 ; jg 2dc9e <.literal16+0x84e> - DB 127,0 ; jg 2dca0 <.literal16+0x850> + DB 127,0 ; jg 2db92 <.literal16+0x852> + DB 127,0 ; jg 2db94 <.literal16+0x854> + DB 127,0 ; jg 2db96 <.literal16+0x856> + DB 127,0 ; jg 2db98 <.literal16+0x858> + DB 127,0 ; jg 2db9a <.literal16+0x85a> + DB 127,0 ; jg 2db9c <.literal16+0x85c> + DB 127,0 ; jg 2db9e <.literal16+0x85e> + DB 127,0 ; jg 2dba0 <.literal16+0x860> DB 129,128,129,128,129,128,129,128,129,128; addl $0x80818081,-0x7f7e7f7f(%rax) DB 129 ; .byte 0x81 DB 128 ; .byte 0x80 @@ -45083,6 +45017,28 @@ ALIGN 16 DB 128 ; .byte 0x80 DB 129 ; .byte 0x81 DB 128 ; .byte 0x80 + +ALIGN 8 + DB 1,0 ; add %eax,(%rax) + DB 0,0 ; add %al,(%rax) + DB 1,0 ; add %eax,(%rax) + DB 0,0 ; add %al,(%rax) + DB 2,0 ; add (%rax),%al + DB 0,0 ; add %al,(%rax) + DB 2,0 ; add (%rax),%al + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 4,0 ; add $0x0,%al + DB 0,0 ; add %al,(%rax) + DB 255 ; (bad) + DB 127,0 ; jg 2dbcb <.literal8+0x1b> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg 2dbcf <.literal8+0x1f> + DB 0,0 ; add %al,(%rax) + DB 128,0,0 ; addb $0x0,(%rax) + DB 0 ; .byte 0x0 + DB 128,0,0 ; addb $0x0,(%rax) ELSE .MODEL FLAT,C _text32 SEGMENT ALIGN(32) 'CODE' @@ -45167,11 +45123,11 @@ _sk_seed_shader_sse2 LABEL PROC DB 102,15,110,2 ; movd (%edx),%xmm0 DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,88,128,141,207,0,0 ; addps 0xcf8d(%eax),%xmm0 + DB 15,88,128,125,206,0,0 ; addps 0xce7d(%eax),%xmm0 DB 102,15,110,74,4 ; movd 0x4(%edx),%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,88,136,157,207,0,0 ; addps 0xcf9d(%eax),%xmm1 + DB 15,88,136,141,206,0,0 ; addps 0xce8d(%eax),%xmm1 DB 141,113,4 ; lea 0x4(%ecx),%esi DB 15,87,210 ; xorps %xmm2,%xmm2 DB 15,41,82,64 ; movaps %xmm2,0x40(%edx) @@ -45179,7 +45135,7 @@ _sk_seed_shader_sse2 LABEL PROC DB 15,41,82,32 ; movaps %xmm2,0x20(%edx) DB 15,41,82,16 ; movaps %xmm2,0x10(%edx) DB 131,236,8 ; sub $0x8,%esp - DB 15,40,144,173,207,0,0 ; movaps 0xcfad(%eax),%xmm2 + DB 15,40,144,157,206,0,0 ; movaps 0xce9d(%eax),%xmm2 DB 15,87,219 ; xorps %xmm3,%xmm3 DB 86 ; push %esi DB 82 ; push %edx @@ -45203,37 +45159,38 @@ _sk_dither_sse2 LABEL PROC DB 89 ; pop %ecx DB 139,69,8 ; mov 0x8(%ebp),%eax DB 102,15,110,32 ; movd (%eax),%xmm4 - DB 102,15,112,236,0 ; pshufd $0x0,%xmm4,%xmm5 - DB 102,15,254,169,72,207,0,0 ; paddd 0xcf48(%ecx),%xmm5 - DB 102,15,110,96,4 ; movd 0x4(%eax),%xmm4 - DB 102,15,112,252,0 ; pshufd $0x0,%xmm4,%xmm7 - DB 102,15,239,253 ; pxor %xmm5,%xmm7 - DB 102,15,111,177,88,207,0,0 ; movdqa 0xcf58(%ecx),%xmm6 + DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 + DB 102,15,254,161,56,206,0,0 ; paddd 0xce38(%ecx),%xmm4 + DB 102,15,110,104,4 ; movd 0x4(%eax),%xmm5 + DB 102,15,112,253,0 ; pshufd $0x0,%xmm5,%xmm7 + DB 102,15,239,252 ; pxor %xmm4,%xmm7 + DB 102,15,111,177,72,206,0,0 ; movdqa 0xce48(%ecx),%xmm6 DB 102,15,111,207 ; movdqa %xmm7,%xmm1 DB 102,15,219,206 ; pand %xmm6,%xmm1 - DB 102,15,219,245 ; pand %xmm5,%xmm6 - DB 102,15,111,129,120,207,0,0 ; movdqa 0xcf78(%ecx),%xmm0 - DB 102,15,111,231 ; movdqa %xmm7,%xmm4 - DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,15,219,197 ; pand %xmm5,%xmm0 - DB 102,15,219,185,104,207,0,0 ; pand 0xcf68(%ecx),%xmm7 - DB 102,15,219,169,104,207,0,0 ; pand 0xcf68(%ecx),%xmm5 + DB 102,15,114,241,5 ; pslld $0x5,%xmm1 + DB 102,15,111,129,104,206,0,0 ; movdqa 0xce68(%ecx),%xmm0 + DB 102,15,111,239 ; movdqa %xmm7,%xmm5 + DB 102,15,219,232 ; pand %xmm0,%xmm5 + DB 102,15,114,245,2 ; pslld $0x2,%xmm5 + DB 102,15,235,233 ; por %xmm1,%xmm5 + DB 102,15,219,244 ; pand %xmm4,%xmm6 DB 102,15,114,246,4 ; pslld $0x4,%xmm6 + DB 102,15,219,196 ; pand %xmm4,%xmm0 DB 102,15,254,192 ; paddd %xmm0,%xmm0 DB 102,15,235,198 ; por %xmm6,%xmm0 - DB 102,15,114,213,2 ; psrld $0x2,%xmm5 - DB 102,15,235,197 ; por %xmm5,%xmm0 - DB 102,15,114,241,5 ; pslld $0x5,%xmm1 - DB 102,15,114,244,2 ; pslld $0x2,%xmm4 - DB 102,15,235,225 ; por %xmm1,%xmm4 + DB 102,15,111,137,88,206,0,0 ; movdqa 0xce58(%ecx),%xmm1 + DB 102,15,219,249 ; pand %xmm1,%xmm7 + DB 102,15,219,225 ; pand %xmm1,%xmm4 + DB 102,15,114,212,2 ; psrld $0x2,%xmm4 DB 102,15,235,224 ; por %xmm0,%xmm4 + DB 102,15,235,229 ; por %xmm5,%xmm4 DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 102,15,114,215,1 ; psrld $0x1,%xmm7 DB 102,15,235,231 ; por %xmm7,%xmm4 DB 15,91,196 ; cvtdq2ps %xmm4,%xmm0 - DB 15,89,129,136,207,0,0 ; mulps 0xcf88(%ecx),%xmm0 - DB 15,88,129,152,207,0,0 ; addps 0xcf98(%ecx),%xmm0 + DB 15,89,129,120,206,0,0 ; mulps 0xce78(%ecx),%xmm0 + DB 15,88,129,136,206,0,0 ; addps 0xce88(%ecx),%xmm0 DB 243,15,16,14 ; movss (%esi),%xmm1 DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1 DB 15,89,200 ; mulps %xmm0,%xmm1 @@ -45297,7 +45254,7 @@ _sk_black_color_sse2 LABEL PROC DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 141,81,4 ; lea 0x4(%ecx),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,152,252,205,0,0 ; movaps 0xcdfc(%eax),%xmm3 + DB 15,40,152,236,204,0,0 ; movaps 0xccec(%eax),%xmm3 DB 15,87,192 ; xorps %xmm0,%xmm0 DB 15,87,201 ; xorps %xmm1,%xmm1 DB 15,87,210 ; xorps %xmm2,%xmm2 @@ -45318,7 +45275,7 @@ _sk_white_color_sse2 LABEL PROC DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 141,81,4 ; lea 0x4(%ecx),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,128,204,205,0,0 ; movaps 0xcdcc(%eax),%xmm0 + DB 15,40,128,188,204,0,0 ; movaps 0xccbc(%eax),%xmm0 DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,40,216 ; movaps %xmm0,%xmm3 @@ -45400,7 +45357,7 @@ _sk_srcatop_sse2 LABEL PROC DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 15,40,97,64 ; movaps 0x40(%ecx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 - DB 15,40,170,29,205,0,0 ; movaps 0xcd1d(%edx),%xmm5 + DB 15,40,170,13,204,0,0 ; movaps 0xcc0d(%edx),%xmm5 DB 15,92,235 ; subps %xmm3,%xmm5 DB 15,40,113,16 ; movaps 0x10(%ecx),%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 @@ -45437,7 +45394,7 @@ _sk_dstatop_sse2 LABEL PROC DB 15,40,105,64 ; movaps 0x40(%ecx),%xmm5 DB 15,40,113,16 ; movaps 0x10(%ecx),%xmm6 DB 15,89,243 ; mulps %xmm3,%xmm6 - DB 15,40,162,190,204,0,0 ; movaps 0xccbe(%edx),%xmm4 + DB 15,40,162,174,203,0,0 ; movaps 0xcbae(%edx),%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,88,198 ; addps %xmm6,%xmm0 @@ -45515,7 +45472,7 @@ _sk_srcout_sse2 LABEL PROC DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 139,85,8 ; mov 0x8(%ebp),%edx - DB 15,40,160,253,203,0,0 ; movaps 0xcbfd(%eax),%xmm4 + DB 15,40,160,237,202,0,0 ; movaps 0xcaed(%eax),%xmm4 DB 15,92,98,64 ; subps 0x40(%edx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,89,204 ; mulps %xmm4,%xmm1 @@ -45539,7 +45496,7 @@ _sk_dstout_sse2 LABEL PROC DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 139,85,8 ; mov 0x8(%ebp),%edx - DB 15,40,160,197,203,0,0 ; movaps 0xcbc5(%eax),%xmm4 + DB 15,40,160,181,202,0,0 ; movaps 0xcab5(%eax),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 DB 15,40,66,16 ; movaps 0x10(%edx),%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 @@ -45567,7 +45524,7 @@ _sk_srcover_sse2 LABEL PROC DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 139,85,8 ; mov 0x8(%ebp),%edx - DB 15,40,160,126,203,0,0 ; movaps 0xcb7e(%eax),%xmm4 + DB 15,40,160,110,202,0,0 ; movaps 0xca6e(%eax),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 DB 15,40,106,16 ; movaps 0x10(%edx),%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 @@ -45599,7 +45556,7 @@ _sk_dstover_sse2 LABEL PROC DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 139,85,8 ; mov 0x8(%ebp),%edx DB 15,40,106,64 ; movaps 0x40(%edx),%xmm5 - DB 15,40,160,46,203,0,0 ; movaps 0xcb2e(%eax),%xmm4 + DB 15,40,160,30,202,0,0 ; movaps 0xca1e(%eax),%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,88,66,16 ; addps 0x10(%edx),%xmm0 @@ -45651,7 +45608,7 @@ _sk_multiply_sse2 LABEL PROC DB 139,69,8 ; mov 0x8(%ebp),%eax DB 15,40,80,64 ; movaps 0x40(%eax),%xmm2 DB 15,41,85,200 ; movaps %xmm2,-0x38(%ebp) - DB 15,40,129,175,202,0,0 ; movaps 0xcaaf(%ecx),%xmm0 + DB 15,40,129,159,201,0,0 ; movaps 0xc99f(%ecx),%xmm0 DB 15,40,240 ; movaps %xmm0,%xmm6 DB 15,92,242 ; subps %xmm2,%xmm6 DB 15,40,254 ; movaps %xmm6,%xmm7 @@ -45709,7 +45666,7 @@ _sk_plus__sse2 LABEL PROC DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 139,85,8 ; mov 0x8(%ebp),%edx DB 15,88,66,16 ; addps 0x10(%edx),%xmm0 - DB 15,40,160,9,202,0,0 ; movaps 0xca09(%eax),%xmm4 + DB 15,40,160,249,200,0,0 ; movaps 0xc8f9(%eax),%xmm4 DB 15,93,196 ; minps %xmm4,%xmm0 DB 15,88,74,32 ; addps 0x20(%edx),%xmm1 DB 15,93,204 ; minps %xmm4,%xmm1 @@ -45777,7 +45734,7 @@ _sk_xor__sse2 LABEL PROC DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 15,40,105,64 ; movaps 0x40(%ecx),%xmm5 - DB 15,40,154,91,201,0,0 ; movaps 0xc95b(%edx),%xmm3 + DB 15,40,154,75,200,0,0 ; movaps 0xc84b(%edx),%xmm3 DB 15,40,243 ; movaps %xmm3,%xmm6 DB 15,92,245 ; subps %xmm5,%xmm6 DB 15,89,198 ; mulps %xmm6,%xmm0 @@ -45837,7 +45794,7 @@ _sk_darken_sse2 LABEL PROC DB 15,89,203 ; mulps %xmm3,%xmm1 DB 15,95,209 ; maxps %xmm1,%xmm2 DB 15,92,234 ; subps %xmm2,%xmm5 - DB 15,40,138,243,200,0,0 ; movaps 0xc8f3(%edx),%xmm1 + DB 15,40,138,227,199,0,0 ; movaps 0xc7e3(%edx),%xmm1 DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,206 ; mulps %xmm6,%xmm1 DB 15,88,217 ; addps %xmm1,%xmm3 @@ -45884,7 +45841,7 @@ _sk_lighten_sse2 LABEL PROC DB 15,89,203 ; mulps %xmm3,%xmm1 DB 15,93,209 ; minps %xmm1,%xmm2 DB 15,92,234 ; subps %xmm2,%xmm5 - DB 15,40,138,115,200,0,0 ; movaps 0xc873(%edx),%xmm1 + DB 15,40,138,99,199,0,0 ; movaps 0xc763(%edx),%xmm1 DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,206 ; mulps %xmm6,%xmm1 DB 15,88,217 ; addps %xmm1,%xmm3 @@ -45935,7 +45892,7 @@ _sk_difference_sse2 LABEL PROC DB 15,93,214 ; minps %xmm6,%xmm2 DB 15,88,210 ; addps %xmm2,%xmm2 DB 15,92,226 ; subps %xmm2,%xmm4 - DB 15,40,146,240,199,0,0 ; movaps 0xc7f0(%edx),%xmm2 + DB 15,40,146,224,198,0,0 ; movaps 0xc6e0(%edx),%xmm2 DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,213 ; mulps %xmm5,%xmm2 DB 15,88,218 ; addps %xmm2,%xmm3 @@ -45977,7 +45934,7 @@ _sk_exclusion_sse2 LABEL PROC DB 15,89,202 ; mulps %xmm2,%xmm1 DB 15,88,201 ; addps %xmm1,%xmm1 DB 15,92,233 ; subps %xmm1,%xmm5 - DB 15,40,138,106,199,0,0 ; movaps 0xc76a(%edx),%xmm1 + DB 15,40,138,90,198,0,0 ; movaps 0xc65a(%edx),%xmm1 DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,73,64 ; mulps 0x40(%ecx),%xmm1 DB 15,88,217 ; addps %xmm1,%xmm3 @@ -46016,7 +45973,7 @@ _sk_colorburn_sse2 LABEL PROC DB 15,93,198 ; minps %xmm6,%xmm0 DB 15,40,233 ; movaps %xmm1,%xmm5 DB 15,92,232 ; subps %xmm0,%xmm5 - DB 15,40,177,235,198,0,0 ; movaps 0xc6eb(%ecx),%xmm6 + DB 15,40,177,219,197,0,0 ; movaps 0xc5db(%ecx),%xmm6 DB 15,40,198 ; movaps %xmm6,%xmm0 DB 15,92,193 ; subps %xmm1,%xmm0 DB 15,41,69,200 ; movaps %xmm0,-0x38(%ebp) @@ -46131,7 +46088,7 @@ _sk_colordodge_sse2 LABEL PROC DB 89 ; pop %ecx DB 139,69,8 ; mov 0x8(%ebp),%eax DB 15,40,72,16 ; movaps 0x10(%eax),%xmm1 - DB 15,40,145,98,197,0,0 ; movaps 0xc562(%ecx),%xmm2 + DB 15,40,145,82,196,0,0 ; movaps 0xc452(%ecx),%xmm2 DB 15,40,226 ; movaps %xmm2,%xmm4 DB 15,92,214 ; subps %xmm6,%xmm2 DB 15,40,250 ; movaps %xmm2,%xmm7 @@ -46246,7 +46203,7 @@ _sk_hardlight_sse2 LABEL PROC DB 89 ; pop %ecx DB 139,69,8 ; mov 0x8(%ebp),%eax DB 15,40,72,64 ; movaps 0x40(%eax),%xmm1 - DB 15,40,145,1,196,0,0 ; movaps 0xc401(%ecx),%xmm2 + DB 15,40,145,241,194,0,0 ; movaps 0xc2f1(%ecx),%xmm2 DB 15,40,234 ; movaps %xmm2,%xmm5 DB 15,92,233 ; subps %xmm1,%xmm5 DB 15,40,225 ; movaps %xmm1,%xmm4 @@ -46354,7 +46311,7 @@ _sk_overlay_sse2 LABEL PROC DB 89 ; pop %ecx DB 139,69,8 ; mov 0x8(%ebp),%eax DB 15,40,80,64 ; movaps 0x40(%eax),%xmm2 - DB 15,40,161,182,194,0,0 ; movaps 0xc2b6(%ecx),%xmm4 + DB 15,40,161,166,193,0,0 ; movaps 0xc1a6(%ecx),%xmm4 DB 15,40,244 ; movaps %xmm4,%xmm6 DB 15,92,242 ; subps %xmm2,%xmm6 DB 15,40,206 ; movaps %xmm6,%xmm1 @@ -46480,17 +46437,17 @@ _sk_softlight_sse2 LABEL PROC DB 232,0,0,0,0 ; call f2a <_sk_softlight_sse2+0x52> DB 89 ; pop %ecx DB 15,40,194 ; movaps %xmm2,%xmm0 - DB 15,40,137,150,193,0,0 ; movaps 0xc196(%ecx),%xmm1 + DB 15,40,137,134,192,0,0 ; movaps 0xc086(%ecx),%xmm1 DB 15,41,141,104,255,255,255 ; movaps %xmm1,-0x98(%ebp) DB 15,88,193 ; addps %xmm1,%xmm0 DB 15,89,198 ; mulps %xmm6,%xmm0 - DB 15,40,185,38,193,0,0 ; movaps 0xc126(%ecx),%xmm7 + DB 15,40,185,22,192,0,0 ; movaps 0xc016(%ecx),%xmm7 DB 15,41,125,232 ; movaps %xmm7,-0x18(%ebp) DB 15,92,250 ; subps %xmm2,%xmm7 DB 15,82,242 ; rsqrtps %xmm2,%xmm6 DB 15,83,206 ; rcpps %xmm6,%xmm1 DB 15,92,202 ; subps %xmm2,%xmm1 - DB 15,40,161,166,193,0,0 ; movaps 0xc1a6(%ecx),%xmm4 + DB 15,40,161,150,192,0,0 ; movaps 0xc096(%ecx),%xmm4 DB 15,41,101,152 ; movaps %xmm4,-0x68(%ebp) DB 15,89,212 ; mulps %xmm4,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 @@ -46694,7 +46651,7 @@ _sk_hue_sse2 LABEL PROC DB 15,40,226 ; movaps %xmm2,%xmm4 DB 15,40,72,64 ; movaps 0x40(%eax),%xmm1 DB 15,41,77,216 ; movaps %xmm1,-0x28(%ebp) - DB 15,40,169,101,190,0,0 ; movaps 0xbe65(%ecx),%xmm5 + DB 15,40,169,85,189,0,0 ; movaps 0xbd55(%ecx),%xmm5 DB 15,92,233 ; subps %xmm1,%xmm5 DB 15,89,197 ; mulps %xmm5,%xmm0 DB 15,41,133,72,255,255,255 ; movaps %xmm0,-0xb8(%ebp) @@ -46727,17 +46684,17 @@ _sk_hue_sse2 LABEL PROC DB 15,89,233 ; mulps %xmm1,%xmm5 DB 15,41,173,120,255,255,255 ; movaps %xmm5,-0x88(%ebp) DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,40,153,245,190,0,0 ; movaps 0xbef5(%ecx),%xmm3 + DB 15,40,153,229,189,0,0 ; movaps 0xbde5(%ecx),%xmm3 DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) DB 15,40,117,152 ; movaps -0x68(%ebp),%xmm6 DB 15,40,206 ; movaps %xmm6,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 15,40,153,5,191,0,0 ; movaps 0xbf05(%ecx),%xmm3 + DB 15,40,153,245,189,0,0 ; movaps 0xbdf5(%ecx),%xmm3 DB 15,40,109,136 ; movaps -0x78(%ebp),%xmm5 DB 15,40,253 ; movaps %xmm5,%xmm7 DB 15,89,251 ; mulps %xmm3,%xmm7 DB 15,88,249 ; addps %xmm1,%xmm7 - DB 15,40,137,101,190,0,0 ; movaps 0xbe65(%ecx),%xmm1 + DB 15,40,137,85,189,0,0 ; movaps 0xbd55(%ecx),%xmm1 DB 15,92,77,184 ; subps -0x48(%ebp),%xmm1 DB 15,89,241 ; mulps %xmm1,%xmm6 DB 15,41,117,152 ; movaps %xmm6,-0x68(%ebp) @@ -46747,7 +46704,7 @@ _sk_hue_sse2 LABEL PROC DB 15,40,77,200 ; movaps -0x38(%ebp),%xmm1 DB 15,89,241 ; mulps %xmm1,%xmm6 DB 15,41,117,168 ; movaps %xmm6,-0x58(%ebp) - DB 15,40,169,21,191,0,0 ; movaps 0xbf15(%ecx),%xmm5 + DB 15,40,169,5,190,0,0 ; movaps 0xbe05(%ecx),%xmm5 DB 15,89,205 ; mulps %xmm5,%xmm1 DB 15,88,207 ; addps %xmm7,%xmm1 DB 15,94,212 ; divps %xmm4,%xmm2 @@ -46938,17 +46895,17 @@ _sk_saturation_sse2 LABEL PROC DB 15,89,253 ; mulps %xmm5,%xmm7 DB 232,0,0,0,0 ; call 153f <_sk_saturation_sse2+0xb0> DB 89 ; pop %ecx - DB 15,40,153,161,187,0,0 ; movaps 0xbba1(%ecx),%xmm3 + DB 15,40,153,145,186,0,0 ; movaps 0xba91(%ecx),%xmm3 DB 15,41,157,88,255,255,255 ; movaps %xmm3,-0xa8(%ebp) DB 15,40,77,184 ; movaps -0x48(%ebp),%xmm1 DB 15,40,193 ; movaps %xmm1,%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,40,153,177,187,0,0 ; movaps 0xbbb1(%ecx),%xmm3 + DB 15,40,153,161,186,0,0 ; movaps 0xbaa1(%ecx),%xmm3 DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) DB 15,40,238 ; movaps %xmm6,%xmm5 DB 15,89,235 ; mulps %xmm3,%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 - DB 15,40,129,17,187,0,0 ; movaps 0xbb11(%ecx),%xmm0 + DB 15,40,129,1,186,0,0 ; movaps 0xba01(%ecx),%xmm0 DB 15,41,133,104,255,255,255 ; movaps %xmm0,-0x98(%ebp) DB 15,92,69,216 ; subps -0x28(%ebp),%xmm0 DB 15,89,200 ; mulps %xmm0,%xmm1 @@ -46960,7 +46917,7 @@ _sk_saturation_sse2 LABEL PROC DB 15,89,200 ; mulps %xmm0,%xmm1 DB 15,41,141,120,255,255,255 ; movaps %xmm1,-0x88(%ebp) DB 15,40,200 ; movaps %xmm0,%xmm1 - DB 15,40,129,193,187,0,0 ; movaps 0xbbc1(%ecx),%xmm0 + DB 15,40,129,177,186,0,0 ; movaps 0xbab1(%ecx),%xmm0 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 15,88,205 ; addps %xmm5,%xmm1 DB 15,40,109,168 ; movaps -0x58(%ebp),%xmm5 @@ -47117,16 +47074,16 @@ _sk_color_sse2 LABEL PROC DB 139,69,8 ; mov 0x8(%ebp),%eax DB 15,40,80,16 ; movaps 0x10(%eax),%xmm2 DB 15,40,72,32 ; movaps 0x20(%eax),%xmm1 - DB 15,40,177,70,185,0,0 ; movaps 0xb946(%ecx),%xmm6 + DB 15,40,177,54,184,0,0 ; movaps 0xb836(%ecx),%xmm6 DB 15,41,117,216 ; movaps %xmm6,-0x28(%ebp) DB 15,40,194 ; movaps %xmm2,%xmm0 DB 15,89,198 ; mulps %xmm6,%xmm0 - DB 15,40,153,86,185,0,0 ; movaps 0xb956(%ecx),%xmm3 + DB 15,40,153,70,184,0,0 ; movaps 0xb846(%ecx),%xmm3 DB 15,41,93,184 ; movaps %xmm3,-0x48(%ebp) DB 15,40,241 ; movaps %xmm1,%xmm6 DB 15,89,243 ; mulps %xmm3,%xmm6 DB 15,88,240 ; addps %xmm0,%xmm6 - DB 15,40,129,182,184,0,0 ; movaps 0xb8b6(%ecx),%xmm0 + DB 15,40,129,166,183,0,0 ; movaps 0xb7a6(%ecx),%xmm0 DB 15,41,69,168 ; movaps %xmm0,-0x58(%ebp) DB 15,92,69,200 ; subps -0x38(%ebp),%xmm0 DB 15,89,208 ; mulps %xmm0,%xmm2 @@ -47136,7 +47093,7 @@ _sk_color_sse2 LABEL PROC DB 15,40,80,48 ; movaps 0x30(%eax),%xmm2 DB 15,89,194 ; mulps %xmm2,%xmm0 DB 15,41,133,72,255,255,255 ; movaps %xmm0,-0xb8(%ebp) - DB 15,40,137,102,185,0,0 ; movaps 0xb966(%ecx),%xmm1 + DB 15,40,137,86,184,0,0 ; movaps 0xb856(%ecx),%xmm1 DB 15,89,209 ; mulps %xmm1,%xmm2 DB 15,88,214 ; addps %xmm6,%xmm2 DB 15,40,88,64 ; movaps 0x40(%eax),%xmm3 @@ -47293,18 +47250,18 @@ _sk_luminosity_sse2 LABEL PROC DB 232,0,0,0,0 ; call 19fd <_sk_luminosity_sse2+0x11> DB 89 ; pop %ecx DB 139,69,8 ; mov 0x8(%ebp),%eax - DB 15,40,161,227,182,0,0 ; movaps 0xb6e3(%ecx),%xmm4 + DB 15,40,161,211,181,0,0 ; movaps 0xb5d3(%ecx),%xmm4 DB 15,41,101,232 ; movaps %xmm4,-0x18(%ebp) DB 15,40,208 ; movaps %xmm0,%xmm2 DB 15,89,212 ; mulps %xmm4,%xmm2 - DB 15,40,161,243,182,0,0 ; movaps 0xb6f3(%ecx),%xmm4 + DB 15,40,161,227,181,0,0 ; movaps 0xb5e3(%ecx),%xmm4 DB 15,41,101,216 ; movaps %xmm4,-0x28(%ebp) DB 15,40,233 ; movaps %xmm1,%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 DB 15,88,234 ; addps %xmm2,%xmm5 DB 15,40,80,64 ; movaps 0x40(%eax),%xmm2 DB 15,41,85,184 ; movaps %xmm2,-0x48(%ebp) - DB 15,40,161,83,182,0,0 ; movaps 0xb653(%ecx),%xmm4 + DB 15,40,161,67,181,0,0 ; movaps 0xb543(%ecx),%xmm4 DB 15,40,252 ; movaps %xmm4,%xmm7 DB 15,92,250 ; subps %xmm2,%xmm7 DB 15,89,199 ; mulps %xmm7,%xmm0 @@ -47314,7 +47271,7 @@ _sk_luminosity_sse2 LABEL PROC DB 15,89,254 ; mulps %xmm6,%xmm7 DB 15,41,189,104,255,255,255 ; movaps %xmm7,-0x98(%ebp) DB 15,40,214 ; movaps %xmm6,%xmm2 - DB 15,40,137,3,183,0,0 ; movaps 0xb703(%ecx),%xmm1 + DB 15,40,137,243,181,0,0 ; movaps 0xb5f3(%ecx),%xmm1 DB 15,89,209 ; mulps %xmm1,%xmm2 DB 15,88,213 ; addps %xmm5,%xmm2 DB 15,40,104,16 ; movaps 0x10(%eax),%xmm5 @@ -47488,7 +47445,7 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 139,49 ; mov (%ecx),%esi DB 15,133,224,0,0,0 ; jne 1d6a <_sk_srcover_rgba_8888_sse2+0x119> DB 243,15,111,20,178 ; movdqu (%edx,%esi,4),%xmm2 - DB 102,15,111,167,168,180,0,0 ; movdqa 0xb4a8(%edi),%xmm4 + DB 102,15,111,167,152,179,0,0 ; movdqa 0xb398(%edi),%xmm4 DB 102,15,111,234 ; movdqa %xmm2,%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,111,242 ; movdqa %xmm2,%xmm6 @@ -47499,9 +47456,9 @@ _sk_srcover_rgba_8888_sse2 LABEL PROC DB 102,15,219,252 ; pand %xmm4,%xmm7 DB 15,91,229 ; cvtdq2ps %xmm5,%xmm4 DB 15,41,97,16 ; movaps %xmm4,0x10(%ecx) - DB 15,40,175,232,179,0,0 ; movaps 0xb3e8(%edi),%xmm5 + DB 15,40,175,216,178,0,0 ; movaps 0xb2d8(%edi),%xmm5 DB 15,92,235 ; subps %xmm3,%xmm5 - DB 15,40,143,184,180,0,0 ; movaps 0xb4b8(%edi),%xmm1 + DB 15,40,143,168,179,0,0 ; movaps 0xb3a8(%edi),%xmm1 DB 15,89,193 ; mulps %xmm1,%xmm0 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,88,224 ; addps %xmm0,%xmm4 @@ -47611,7 +47568,7 @@ _sk_clamp_1_sse2 LABEL PROC DB 232,0,0,0,0 ; call 1e14 <_sk_clamp_1_sse2+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 15,40,160,60,178,0,0 ; movaps 0xb23c(%eax),%xmm4 + DB 15,40,160,44,177,0,0 ; movaps 0xb12c(%eax),%xmm4 DB 15,93,196 ; minps %xmm4,%xmm0 DB 15,93,204 ; minps %xmm4,%xmm1 DB 15,93,212 ; minps %xmm4,%xmm2 @@ -47633,7 +47590,7 @@ _sk_clamp_a_sse2 LABEL PROC DB 232,0,0,0,0 ; call 1e47 <_sk_clamp_a_sse2+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 15,93,152,9,178,0,0 ; minps 0xb209(%eax),%xmm3 + DB 15,93,152,249,176,0,0 ; minps 0xb0f9(%eax),%xmm3 DB 15,93,195 ; minps %xmm3,%xmm0 DB 15,93,203 ; minps %xmm3,%xmm1 DB 15,93,211 ; minps %xmm3,%xmm2 @@ -47656,14 +47613,14 @@ _sk_clamp_a_dst_sse2 LABEL PROC DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 139,85,8 ; mov 0x8(%ebp),%edx DB 15,40,98,64 ; movaps 0x40(%edx),%xmm4 - DB 15,93,160,217,177,0,0 ; minps 0xb1d9(%eax),%xmm4 + DB 15,93,160,201,176,0,0 ; minps 0xb0c9(%eax),%xmm4 DB 15,41,98,64 ; movaps %xmm4,0x40(%edx) DB 15,40,106,16 ; movaps 0x10(%edx),%xmm5 + DB 15,40,114,32 ; movaps 0x20(%edx),%xmm6 DB 15,93,236 ; minps %xmm4,%xmm5 DB 15,41,106,16 ; movaps %xmm5,0x10(%edx) - DB 15,40,106,32 ; movaps 0x20(%edx),%xmm5 - DB 15,93,236 ; minps %xmm4,%xmm5 - DB 15,41,106,32 ; movaps %xmm5,0x20(%edx) + DB 15,93,244 ; minps %xmm4,%xmm6 + DB 15,41,114,32 ; movaps %xmm6,0x20(%edx) DB 15,40,106,48 ; movaps 0x30(%edx),%xmm5 DB 15,93,236 ; minps %xmm4,%xmm5 DB 15,41,106,48 ; movaps %xmm5,0x30(%edx) @@ -47724,7 +47681,7 @@ _sk_invert_sse2 LABEL PROC DB 232,0,0,0,0 ; call 1f22 <_sk_invert_sse2+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 15,40,160,46,177,0,0 ; movaps 0xb12e(%eax),%xmm4 + DB 15,40,160,30,176,0,0 ; movaps 0xb01e(%eax),%xmm4 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,92,232 ; subps %xmm0,%xmm5 DB 15,40,244 ; movaps %xmm4,%xmm6 @@ -47837,7 +47794,7 @@ _sk_unpremul_sse2 LABEL PROC DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 15,87,228 ; xorps %xmm4,%xmm4 - DB 15,40,168,48,176,0,0 ; movaps 0xb030(%eax),%xmm5 + DB 15,40,168,32,175,0,0 ; movaps 0xaf20(%eax),%xmm5 DB 15,94,235 ; divps %xmm3,%xmm5 DB 15,194,227,4 ; cmpneqps %xmm3,%xmm4 DB 15,84,229 ; andps %xmm5,%xmm4 @@ -47866,19 +47823,19 @@ _sk_from_srgb_sse2 LABEL PROC DB 88 ; pop %eax DB 15,40,202 ; movaps %xmm2,%xmm1 DB 15,89,201 ; mulps %xmm1,%xmm1 - DB 15,40,160,118,176,0,0 ; movaps 0xb076(%eax),%xmm4 + DB 15,40,160,102,175,0,0 ; movaps 0xaf66(%eax),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 - DB 15,40,176,214,176,0,0 ; movaps 0xb0d6(%eax),%xmm6 + DB 15,40,176,198,175,0,0 ; movaps 0xafc6(%eax),%xmm6 DB 15,41,117,216 ; movaps %xmm6,-0x28(%ebp) DB 15,88,198 ; addps %xmm6,%xmm0 DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,40,176,198,176,0,0 ; movaps 0xb0c6(%eax),%xmm6 + DB 15,40,176,182,175,0,0 ; movaps 0xafb6(%eax),%xmm6 DB 15,40,202 ; movaps %xmm2,%xmm1 DB 15,89,206 ; mulps %xmm6,%xmm1 - DB 15,40,184,230,176,0,0 ; movaps 0xb0e6(%eax),%xmm7 + DB 15,40,184,214,175,0,0 ; movaps 0xafd6(%eax),%xmm7 DB 15,41,125,232 ; movaps %xmm7,-0x18(%ebp) DB 15,88,199 ; addps %xmm7,%xmm0 - DB 15,40,184,246,176,0,0 ; movaps 0xb0f6(%eax),%xmm7 + DB 15,40,184,230,175,0,0 ; movaps 0xafe6(%eax),%xmm7 DB 15,194,215,1 ; cmpltps %xmm7,%xmm2 DB 15,84,202 ; andps %xmm2,%xmm1 DB 15,85,208 ; andnps %xmm0,%xmm2 @@ -47936,24 +47893,24 @@ _sk_from_srgb_dst_sse2 LABEL PROC DB 15,40,72,16 ; movaps 0x10(%eax),%xmm1 DB 15,40,233 ; movaps %xmm1,%xmm5 DB 15,89,237 ; mulps %xmm5,%xmm5 - DB 15,40,161,155,175,0,0 ; movaps 0xaf9b(%ecx),%xmm4 + DB 15,40,161,139,174,0,0 ; movaps 0xae8b(%ecx),%xmm4 DB 15,40,193 ; movaps %xmm1,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 - DB 15,40,177,251,175,0,0 ; movaps 0xaffb(%ecx),%xmm6 + DB 15,40,177,235,174,0,0 ; movaps 0xaeeb(%ecx),%xmm6 DB 15,88,198 ; addps %xmm6,%xmm0 DB 15,89,197 ; mulps %xmm5,%xmm0 - DB 15,40,169,235,175,0,0 ; movaps 0xafeb(%ecx),%xmm5 + DB 15,40,169,219,174,0,0 ; movaps 0xaedb(%ecx),%xmm5 DB 15,40,209 ; movaps %xmm1,%xmm2 DB 15,89,213 ; mulps %xmm5,%xmm2 - DB 15,40,185,11,176,0,0 ; movaps 0xb00b(%ecx),%xmm7 + DB 15,40,185,251,174,0,0 ; movaps 0xaefb(%ecx),%xmm7 DB 15,88,199 ; addps %xmm7,%xmm0 - DB 15,40,153,27,176,0,0 ; movaps 0xb01b(%ecx),%xmm3 + DB 15,40,153,11,175,0,0 ; movaps 0xaf0b(%ecx),%xmm3 DB 15,194,203,1 ; cmpltps %xmm3,%xmm1 DB 15,84,209 ; andps %xmm1,%xmm2 DB 15,85,200 ; andnps %xmm0,%xmm1 DB 15,86,202 ; orps %xmm2,%xmm1 - DB 15,41,72,16 ; movaps %xmm1,0x10(%eax) DB 15,40,64,32 ; movaps 0x20(%eax),%xmm0 + DB 15,41,72,16 ; movaps %xmm1,0x10(%eax) DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,89,201 ; mulps %xmm1,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 @@ -48007,25 +47964,25 @@ _sk_to_srgb_sse2 LABEL PROC DB 232,0,0,0,0 ; call 2230 <_sk_to_srgb_sse2+0x18> DB 88 ; pop %eax DB 15,82,218 ; rsqrtps %xmm2,%xmm3 - DB 15,40,160,80,175,0,0 ; movaps 0xaf50(%eax),%xmm4 + DB 15,40,160,64,174,0,0 ; movaps 0xae40(%eax),%xmm4 DB 15,40,235 ; movaps %xmm3,%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 - DB 15,40,128,96,175,0,0 ; movaps 0xaf60(%eax),%xmm0 + DB 15,40,128,80,174,0,0 ; movaps 0xae50(%eax),%xmm0 DB 15,41,69,232 ; movaps %xmm0,-0x18(%ebp) DB 15,88,232 ; addps %xmm0,%xmm5 DB 15,89,235 ; mulps %xmm3,%xmm5 - DB 15,40,128,112,175,0,0 ; movaps 0xaf70(%eax),%xmm0 + DB 15,40,128,96,174,0,0 ; movaps 0xae60(%eax),%xmm0 DB 15,41,69,200 ; movaps %xmm0,-0x38(%ebp) DB 15,88,232 ; addps %xmm0,%xmm5 - DB 15,40,128,128,175,0,0 ; movaps 0xaf80(%eax),%xmm0 + DB 15,40,128,112,174,0,0 ; movaps 0xae70(%eax),%xmm0 DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) DB 15,88,216 ; addps %xmm0,%xmm3 DB 15,83,195 ; rcpps %xmm3,%xmm0 DB 15,89,197 ; mulps %xmm5,%xmm0 - DB 15,40,168,64,175,0,0 ; movaps 0xaf40(%eax),%xmm5 + DB 15,40,168,48,174,0,0 ; movaps 0xae30(%eax),%xmm5 DB 15,40,202 ; movaps %xmm2,%xmm1 DB 15,89,205 ; mulps %xmm5,%xmm1 - DB 15,40,152,144,175,0,0 ; movaps 0xaf90(%eax),%xmm3 + DB 15,40,152,128,174,0,0 ; movaps 0xae80(%eax),%xmm3 DB 15,194,211,1 ; cmpltps %xmm3,%xmm2 DB 15,84,202 ; andps %xmm2,%xmm1 DB 15,85,208 ; andnps %xmm0,%xmm2 @@ -48094,14 +48051,14 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 15,40,195 ; movaps %xmm3,%xmm0 DB 15,92,193 ; subps %xmm1,%xmm0 DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) - DB 15,40,136,41,173,0,0 ; movaps 0xad29(%eax),%xmm1 + DB 15,40,136,25,172,0,0 ; movaps 0xac19(%eax),%xmm1 DB 15,94,200 ; divps %xmm0,%xmm1 DB 15,40,199 ; movaps %xmm7,%xmm0 DB 15,92,194 ; subps %xmm2,%xmm0 DB 15,89,193 ; mulps %xmm1,%xmm0 DB 15,40,239 ; movaps %xmm7,%xmm5 DB 15,194,234,1 ; cmpltps %xmm2,%xmm5 - DB 15,84,168,169,174,0,0 ; andps 0xaea9(%eax),%xmm5 + DB 15,84,168,153,173,0,0 ; andps 0xad99(%eax),%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 DB 15,40,195 ; movaps %xmm3,%xmm0 DB 15,194,196,0 ; cmpeqps %xmm4,%xmm0 @@ -48111,9 +48068,9 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 15,92,231 ; subps %xmm7,%xmm4 DB 15,89,209 ; mulps %xmm1,%xmm2 DB 15,89,225 ; mulps %xmm1,%xmm4 - DB 15,40,136,185,174,0,0 ; movaps 0xaeb9(%eax),%xmm1 + DB 15,40,136,169,173,0,0 ; movaps 0xada9(%eax),%xmm1 DB 15,88,209 ; addps %xmm1,%xmm2 - DB 15,88,160,201,174,0,0 ; addps 0xaec9(%eax),%xmm4 + DB 15,88,160,185,173,0,0 ; addps 0xadb9(%eax),%xmm4 DB 15,84,214 ; andps %xmm6,%xmm2 DB 15,85,244 ; andnps %xmm4,%xmm6 DB 15,86,242 ; orps %xmm2,%xmm6 @@ -48125,7 +48082,7 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 15,194,226,4 ; cmpneqps %xmm2,%xmm4 DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,88,218 ; addps %xmm2,%xmm3 - DB 15,40,168,25,173,0,0 ; movaps 0xad19(%eax),%xmm5 + DB 15,40,168,9,172,0,0 ; movaps 0xac09(%eax),%xmm5 DB 15,92,202 ; subps %xmm2,%xmm1 DB 15,40,211 ; movaps %xmm3,%xmm2 DB 15,89,213 ; mulps %xmm5,%xmm2 @@ -48135,7 +48092,7 @@ _sk_rgb_to_hsl_sse2 LABEL PROC DB 15,86,233 ; orps %xmm1,%xmm5 DB 15,40,77,216 ; movaps -0x28(%ebp),%xmm1 DB 15,94,205 ; divps %xmm5,%xmm1 - DB 15,89,128,217,174,0,0 ; mulps 0xaed9(%eax),%xmm0 + DB 15,89,128,201,173,0,0 ; mulps 0xadc9(%eax),%xmm0 DB 15,84,196 ; andps %xmm4,%xmm0 DB 15,84,204 ; andps %xmm4,%xmm1 DB 139,69,12 ; mov 0xc(%ebp),%eax @@ -48158,7 +48115,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,41,69,184 ; movaps %xmm0,-0x48(%ebp) DB 232,0,0,0,0 ; call 2422 <_sk_hsl_to_rgb_sse2+0x19> DB 88 ; pop %eax - DB 15,40,160,30,172,0,0 ; movaps 0xac1e(%eax),%xmm4 + DB 15,40,160,14,171,0,0 ; movaps 0xab0e(%eax),%xmm4 DB 15,40,220 ; movaps %xmm4,%xmm3 DB 15,41,165,104,255,255,255 ; movaps %xmm4,-0x98(%ebp) DB 15,194,218,2 ; cmpleps %xmm2,%xmm3 @@ -48171,13 +48128,13 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,84,243 ; andps %xmm3,%xmm6 DB 15,85,217 ; andnps %xmm1,%xmm3 DB 15,86,222 ; orps %xmm6,%xmm3 - DB 15,40,168,238,173,0,0 ; movaps 0xadee(%eax),%xmm5 + DB 15,40,168,222,172,0,0 ; movaps 0xacde(%eax),%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 DB 243,15,91,205 ; cvttps2dq %xmm5,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,194,193,1 ; cmpltps %xmm1,%xmm0 - DB 15,40,176,46,172,0,0 ; movaps 0xac2e(%eax),%xmm6 + DB 15,40,176,30,171,0,0 ; movaps 0xab1e(%eax),%xmm6 DB 15,41,117,136 ; movaps %xmm6,-0x78(%ebp) DB 15,84,198 ; andps %xmm6,%xmm0 DB 15,92,200 ; subps %xmm0,%xmm1 @@ -48187,19 +48144,19 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) DB 15,88,210 ; addps %xmm2,%xmm2 DB 15,92,211 ; subps %xmm3,%xmm2 - DB 15,40,184,254,173,0,0 ; movaps 0xadfe(%eax),%xmm7 + DB 15,40,184,238,172,0,0 ; movaps 0xacee(%eax),%xmm7 DB 15,41,189,88,255,255,255 ; movaps %xmm7,-0xa8(%ebp) DB 15,194,253,2 ; cmpleps %xmm5,%xmm7 DB 15,92,218 ; subps %xmm2,%xmm3 DB 15,41,93,200 ; movaps %xmm3,-0x38(%ebp) DB 15,40,204 ; movaps %xmm4,%xmm1 DB 15,194,205,2 ; cmpleps %xmm5,%xmm1 - DB 15,40,160,222,173,0,0 ; movaps 0xadde(%eax),%xmm4 + DB 15,40,160,206,172,0,0 ; movaps 0xacce(%eax),%xmm4 DB 15,41,165,120,255,255,255 ; movaps %xmm4,-0x88(%ebp) DB 15,194,229,2 ; cmpleps %xmm5,%xmm4 - DB 15,40,176,174,173,0,0 ; movaps 0xadae(%eax),%xmm6 + DB 15,40,176,158,172,0,0 ; movaps 0xac9e(%eax),%xmm6 DB 15,89,238 ; mulps %xmm6,%xmm5 - DB 15,40,128,206,173,0,0 ; movaps 0xadce(%eax),%xmm0 + DB 15,40,128,190,172,0,0 ; movaps 0xacbe(%eax),%xmm0 DB 15,41,69,168 ; movaps %xmm0,-0x58(%ebp) DB 15,92,197 ; subps %xmm5,%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 @@ -48225,18 +48182,18 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,203 ; movaps %xmm3,%xmm1 DB 15,194,200,1 ; cmpltps %xmm0,%xmm1 - DB 15,84,136,46,172,0,0 ; andps 0xac2e(%eax),%xmm1 + DB 15,84,136,30,171,0,0 ; andps 0xab1e(%eax),%xmm1 DB 15,92,193 ; subps %xmm1,%xmm0 DB 15,40,203 ; movaps %xmm3,%xmm1 DB 15,92,200 ; subps %xmm0,%xmm1 - DB 15,40,128,254,173,0,0 ; movaps 0xadfe(%eax),%xmm0 + DB 15,40,128,238,172,0,0 ; movaps 0xacee(%eax),%xmm0 DB 15,194,193,2 ; cmpleps %xmm1,%xmm0 - DB 15,40,152,30,172,0,0 ; movaps 0xac1e(%eax),%xmm3 + DB 15,40,152,14,171,0,0 ; movaps 0xab0e(%eax),%xmm3 DB 15,194,217,2 ; cmpleps %xmm1,%xmm3 - DB 15,40,160,222,173,0,0 ; movaps 0xadde(%eax),%xmm4 + DB 15,40,160,206,172,0,0 ; movaps 0xacce(%eax),%xmm4 DB 15,194,225,2 ; cmpleps %xmm1,%xmm4 DB 15,89,206 ; mulps %xmm6,%xmm1 - DB 15,40,168,206,173,0,0 ; movaps 0xadce(%eax),%xmm5 + DB 15,40,168,190,172,0,0 ; movaps 0xacbe(%eax),%xmm5 DB 15,92,233 ; subps %xmm1,%xmm5 DB 15,89,239 ; mulps %xmm7,%xmm5 DB 15,88,234 ; addps %xmm2,%xmm5 @@ -48256,7 +48213,7 @@ _sk_hsl_to_rgb_sse2 LABEL PROC DB 15,40,125,232 ; movaps -0x18(%ebp),%xmm7 DB 15,85,252 ; andnps %xmm4,%xmm7 DB 15,40,93,184 ; movaps -0x48(%ebp),%xmm3 - DB 15,88,152,14,174,0,0 ; addps 0xae0e(%eax),%xmm3 + DB 15,88,152,254,172,0,0 ; addps 0xacfe(%eax),%xmm3 DB 243,15,91,195 ; cvttps2dq %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,40,203 ; movaps %xmm3,%xmm1 @@ -48354,9 +48311,9 @@ _sk_scale_u8_sse2 LABEL PROC DB 102,15,110,36,62 ; movd (%esi,%edi,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,219,162,116,170,0,0 ; pand 0xaa74(%edx),%xmm4 + DB 102,15,219,162,100,169,0,0 ; pand 0xa964(%edx),%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,162,164,171,0,0 ; mulps 0xaba4(%edx),%xmm4 + DB 15,89,162,148,170,0,0 ; mulps 0xaa94(%edx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,89,204 ; mulps %xmm4,%xmm1 DB 15,89,212 ; mulps %xmm4,%xmm2 @@ -48420,17 +48377,17 @@ _sk_scale_565_sse2 LABEL PROC DB 15,133,157,0,0,0 ; jne 2814 <_sk_scale_565_sse2+0xd8> DB 243,15,126,52,126 ; movq (%esi,%edi,2),%xmm6 DB 102,15,97,240 ; punpcklwd %xmm0,%xmm6 - DB 102,15,111,162,250,170,0,0 ; movdqa 0xaafa(%edx),%xmm4 + DB 102,15,111,162,234,169,0,0 ; movdqa 0xa9ea(%edx),%xmm4 DB 102,15,219,230 ; pand %xmm6,%xmm4 DB 15,91,212 ; cvtdq2ps %xmm4,%xmm2 - DB 15,89,146,10,171,0,0 ; mulps 0xab0a(%edx),%xmm2 - DB 102,15,111,170,26,171,0,0 ; movdqa 0xab1a(%edx),%xmm5 + DB 15,89,146,250,169,0,0 ; mulps 0xa9fa(%edx),%xmm2 + DB 102,15,111,170,10,170,0,0 ; movdqa 0xaa0a(%edx),%xmm5 DB 102,15,219,238 ; pand %xmm6,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,170,42,171,0,0 ; mulps 0xab2a(%edx),%xmm5 - DB 102,15,219,178,58,171,0,0 ; pand 0xab3a(%edx),%xmm6 + DB 15,89,170,26,170,0,0 ; mulps 0xaa1a(%edx),%xmm5 + DB 102,15,219,178,42,170,0,0 ; pand 0xaa2a(%edx),%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,89,178,74,171,0,0 ; mulps 0xab4a(%edx),%xmm6 + DB 15,89,178,58,170,0,0 ; mulps 0xaa3a(%edx),%xmm6 DB 15,40,253 ; movaps %xmm5,%xmm7 DB 15,93,254 ; minps %xmm6,%xmm7 DB 15,40,226 ; movaps %xmm2,%xmm4 @@ -48465,7 +48422,7 @@ _sk_scale_565_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,50 ; je 284e <_sk_scale_565_sse2+0x112> + DB 116,51 ; je 284f <_sk_scale_565_sse2+0x113> DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 128,251,2 ; cmp $0x2,%bl DB 116,23 ; je 283c <_sk_scale_565_sse2+0x100> @@ -48475,12 +48432,12 @@ _sk_scale_565_sse2 LABEL PROC DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,244,69 ; pshufd $0x45,%xmm4,%xmm6 DB 102,15,110,36,126 ; movd (%esi,%edi,2),%xmm4 - DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 + DB 242,15,112,228,212 ; pshuflw $0xd4,%xmm4,%xmm4 DB 242,15,16,244 ; movsd %xmm4,%xmm6 - DB 233,50,255,255,255 ; jmp 2780 <_sk_scale_565_sse2+0x44> + DB 233,49,255,255,255 ; jmp 2780 <_sk_scale_565_sse2+0x44> DB 15,183,52,126 ; movzwl (%esi,%edi,2),%esi DB 102,15,110,246 ; movd %esi,%xmm6 - DB 233,37,255,255,255 ; jmp 2780 <_sk_scale_565_sse2+0x44> + DB 233,36,255,255,255 ; jmp 2780 <_sk_scale_565_sse2+0x44> PUBLIC _sk_lerp_1_float_sse2 _sk_lerp_1_float_sse2 LABEL PROC @@ -48525,7 +48482,7 @@ _sk_lerp_u8_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 28c3 <_sk_lerp_u8_sse2+0xe> + DB 232,0,0,0,0 ; call 28c4 <_sk_lerp_u8_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -48536,13 +48493,13 @@ _sk_lerp_u8_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,102 ; jne 2944 <_sk_lerp_u8_sse2+0x8f> + DB 117,102 ; jne 2945 <_sk_lerp_u8_sse2+0x8f> DB 102,15,110,36,62 ; movd (%esi,%edi,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,219,162,77,168,0,0 ; pand 0xa84d(%edx),%xmm4 + DB 102,15,219,162,60,167,0,0 ; pand 0xa73c(%edx),%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,162,125,169,0,0 ; mulps 0xa97d(%edx),%xmm4 + DB 15,89,162,108,168,0,0 ; mulps 0xa86c(%edx),%xmm4 DB 15,40,105,16 ; movaps 0x10(%ecx),%xmm5 DB 15,40,113,32 ; movaps 0x20(%ecx),%xmm6 DB 15,92,197 ; subps %xmm5,%xmm0 @@ -48572,12 +48529,12 @@ _sk_lerp_u8_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,53 ; je 2981 <_sk_lerp_u8_sse2+0xcc> + DB 116,53 ; je 2982 <_sk_lerp_u8_sse2+0xcc> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 2968 <_sk_lerp_u8_sse2+0xb3> + DB 116,19 ; je 2969 <_sk_lerp_u8_sse2+0xb3> DB 128,251,3 ; cmp $0x3,%bl - DB 117,145 ; jne 28eb <_sk_lerp_u8_sse2+0x36> + DB 117,145 ; jne 28ec <_sk_lerp_u8_sse2+0x36> DB 15,182,92,62,2 ; movzbl 0x2(%esi,%edi,1),%ebx DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -48586,10 +48543,10 @@ _sk_lerp_u8_sse2 LABEL PROC DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 233,106,255,255,255 ; jmp 28eb <_sk_lerp_u8_sse2+0x36> + DB 233,106,255,255,255 ; jmp 28ec <_sk_lerp_u8_sse2+0x36> DB 15,182,52,62 ; movzbl (%esi,%edi,1),%esi DB 102,15,110,230 ; movd %esi,%xmm4 - DB 233,93,255,255,255 ; jmp 28eb <_sk_lerp_u8_sse2+0x36> + DB 233,93,255,255,255 ; jmp 28ec <_sk_lerp_u8_sse2+0x36> PUBLIC _sk_lerp_565_sse2 _sk_lerp_565_sse2 LABEL PROC @@ -48603,7 +48560,7 @@ _sk_lerp_565_sse2 LABEL PROC DB 15,41,85,136 ; movaps %xmm2,-0x78(%ebp) DB 15,41,77,152 ; movaps %xmm1,-0x68(%ebp) DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) - DB 232,0,0,0,0 ; call 29ad <_sk_lerp_565_sse2+0x1f> + DB 232,0,0,0,0 ; call 29ae <_sk_lerp_565_sse2+0x1f> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -48615,20 +48572,20 @@ _sk_lerp_565_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 15,133,211,0,0,0 ; jne 2aa1 <_sk_lerp_565_sse2+0x113> + DB 15,133,211,0,0,0 ; jne 2aa2 <_sk_lerp_565_sse2+0x113> DB 243,15,126,44,126 ; movq (%esi,%edi,2),%xmm5 DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 - DB 102,15,111,162,163,168,0,0 ; movdqa 0xa8a3(%edx),%xmm4 + DB 102,15,111,162,146,167,0,0 ; movdqa 0xa792(%edx),%xmm4 DB 102,15,219,229 ; pand %xmm5,%xmm4 DB 15,91,204 ; cvtdq2ps %xmm4,%xmm1 - DB 15,89,138,179,168,0,0 ; mulps 0xa8b3(%edx),%xmm1 - DB 102,15,111,178,195,168,0,0 ; movdqa 0xa8c3(%edx),%xmm6 + DB 15,89,138,162,167,0,0 ; mulps 0xa7a2(%edx),%xmm1 + DB 102,15,111,178,178,167,0,0 ; movdqa 0xa7b2(%edx),%xmm6 DB 102,15,219,245 ; pand %xmm5,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,89,178,211,168,0,0 ; mulps 0xa8d3(%edx),%xmm6 - DB 102,15,219,170,227,168,0,0 ; pand 0xa8e3(%edx),%xmm5 + DB 15,89,178,194,167,0,0 ; mulps 0xa7c2(%edx),%xmm6 + DB 102,15,219,170,210,167,0,0 ; pand 0xa7d2(%edx),%xmm5 DB 15,91,213 ; cvtdq2ps %xmm5,%xmm2 - DB 15,89,146,243,168,0,0 ; mulps 0xa8f3(%edx),%xmm2 + DB 15,89,146,226,167,0,0 ; mulps 0xa7e2(%edx),%xmm2 DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,41,93,200 ; movaps %xmm3,-0x38(%ebp) DB 15,40,254 ; movaps %xmm6,%xmm7 @@ -48678,22 +48635,22 @@ _sk_lerp_565_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,50 ; je 2adb <_sk_lerp_565_sse2+0x14d> + DB 116,51 ; je 2add <_sk_lerp_565_sse2+0x14e> DB 102,15,239,237 ; pxor %xmm5,%xmm5 DB 128,251,2 ; cmp $0x2,%bl - DB 116,23 ; je 2ac9 <_sk_lerp_565_sse2+0x13b> + DB 116,23 ; je 2aca <_sk_lerp_565_sse2+0x13b> DB 128,251,3 ; cmp $0x3,%bl - DB 15,133,28,255,255,255 ; jne 29d7 <_sk_lerp_565_sse2+0x49> + DB 15,133,28,255,255,255 ; jne 29d8 <_sk_lerp_565_sse2+0x49> DB 15,183,92,126,4 ; movzwl 0x4(%esi,%edi,2),%ebx DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,236,69 ; pshufd $0x45,%xmm4,%xmm5 DB 102,15,110,36,126 ; movd (%esi,%edi,2),%xmm4 - DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 + DB 242,15,112,228,212 ; pshuflw $0xd4,%xmm4,%xmm4 DB 242,15,16,236 ; movsd %xmm4,%xmm5 - DB 233,252,254,255,255 ; jmp 29d7 <_sk_lerp_565_sse2+0x49> + DB 233,251,254,255,255 ; jmp 29d8 <_sk_lerp_565_sse2+0x49> DB 15,183,52,126 ; movzwl (%esi,%edi,2),%esi DB 102,15,110,238 ; movd %esi,%xmm5 - DB 233,239,254,255,255 ; jmp 29d7 <_sk_lerp_565_sse2+0x49> + DB 233,238,254,255,255 ; jmp 29d8 <_sk_lerp_565_sse2+0x49> PUBLIC _sk_load_tables_sse2 _sk_load_tables_sse2 LABEL PROC @@ -48703,7 +48660,7 @@ _sk_load_tables_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 2af6 <_sk_load_tables_sse2+0xe> + DB 232,0,0,0,0 ; call 2af8 <_sk_load_tables_sse2+0xe> DB 90 ; pop %edx DB 139,117,12 ; mov 0xc(%ebp),%esi DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -48712,12 +48669,12 @@ _sk_load_tables_sse2 LABEL PROC DB 133,192 ; test %eax,%eax DB 139,54 ; mov (%esi),%esi DB 139,30 ; mov (%esi),%ebx - DB 15,133,19,1,0,0 ; jne 2c21 <_sk_load_tables_sse2+0x139> + DB 15,133,22,1,0,0 ; jne 2c26 <_sk_load_tables_sse2+0x13c> DB 243,15,111,28,187 ; movdqu (%ebx,%edi,4),%xmm3 - DB 102,15,111,146,26,166,0,0 ; movdqa 0xa61a(%edx),%xmm2 + DB 102,15,111,146,8,165,0,0 ; movdqa 0xa508(%edx),%xmm2 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,112,224,231 ; pshufd $0xe7,%xmm0,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax DB 139,126,4 ; mov 0x4(%esi),%edi @@ -48725,7 +48682,7 @@ _sk_load_tables_sse2 LABEL PROC DB 102,15,126,200 ; movd %xmm1,%eax DB 243,15,16,44,135 ; movss (%edi,%eax,4),%xmm5 DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 243,15,16,4,135 ; movss (%edi,%eax,4),%xmm0 DB 102,15,126,200 ; movd %xmm1,%eax DB 243,15,16,12,135 ; movss (%edi,%eax,4),%xmm1 @@ -48735,42 +48692,42 @@ _sk_load_tables_sse2 LABEL PROC DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 DB 102,15,219,202 ; pand %xmm2,%xmm1 - DB 102,15,112,225,229 ; pshufd $0xe5,%xmm1,%xmm4 + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 DB 102,15,112,241,231 ; pshufd $0xe7,%xmm1,%xmm6 DB 102,15,126,247 ; movd %xmm6,%edi DB 243,15,16,52,184 ; movss (%eax,%edi,4),%xmm6 DB 102,15,126,231 ; movd %xmm4,%edi DB 243,15,16,36,184 ; movss (%eax,%edi,4),%xmm4 DB 102,15,126,207 ; movd %xmm1,%edi - DB 102,15,112,249,78 ; pshufd $0x4e,%xmm1,%xmm7 + DB 102,15,112,249,229 ; pshufd $0xe5,%xmm1,%xmm7 DB 243,15,16,12,184 ; movss (%eax,%edi,4),%xmm1 DB 102,15,126,255 ; movd %xmm7,%edi DB 243,15,16,60,184 ; movss (%eax,%edi,4),%xmm7 - DB 15,20,197 ; unpcklps %xmm5,%xmm0 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 DB 15,20,230 ; unpcklps %xmm6,%xmm4 DB 15,20,207 ; unpcklps %xmm7,%xmm1 DB 139,70,12 ; mov 0xc(%esi),%eax DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,114,213,16 ; psrld $0x10,%xmm5 DB 102,15,219,234 ; pand %xmm2,%xmm5 - DB 102,15,112,213,229 ; pshufd $0xe5,%xmm5,%xmm2 + DB 102,15,112,213,78 ; pshufd $0x4e,%xmm5,%xmm2 DB 102,15,112,245,231 ; pshufd $0xe7,%xmm5,%xmm6 DB 102,15,126,246 ; movd %xmm6,%esi DB 243,15,16,52,176 ; movss (%eax,%esi,4),%xmm6 DB 102,15,126,214 ; movd %xmm2,%esi DB 243,15,16,60,176 ; movss (%eax,%esi,4),%xmm7 DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,112,237,78 ; pshufd $0x4e,%xmm5,%xmm5 + DB 102,15,112,237,229 ; pshufd $0xe5,%xmm5,%xmm5 DB 243,15,16,20,176 ; movss (%eax,%esi,4),%xmm2 DB 102,15,126,238 ; movd %xmm5,%esi DB 243,15,16,44,176 ; movss (%eax,%esi,4),%xmm5 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 + DB 102,15,20,204 ; unpcklpd %xmm4,%xmm1 DB 15,20,254 ; unpcklps %xmm6,%xmm7 DB 15,20,213 ; unpcklps %xmm5,%xmm2 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 + DB 102,15,20,215 ; unpcklpd %xmm7,%xmm2 DB 102,15,114,211,24 ; psrld $0x18,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,154,74,167,0,0 ; mulps 0xa74a(%edx),%xmm3 + DB 15,89,154,56,166,0,0 ; mulps 0xa638(%edx),%xmm3 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 137,194 ; mov %eax,%edx DB 141,66,8 ; lea 0x8(%edx),%eax @@ -48786,18 +48743,18 @@ _sk_load_tables_sse2 LABEL PROC DB 195 ; ret DB 36,3 ; and $0x3,%al DB 60,1 ; cmp $0x1,%al - DB 116,37 ; je 2c4c <_sk_load_tables_sse2+0x164> + DB 116,37 ; je 2c51 <_sk_load_tables_sse2+0x167> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 60,2 ; cmp $0x2,%al - DB 116,19 ; je 2c42 <_sk_load_tables_sse2+0x15a> + DB 116,19 ; je 2c47 <_sk_load_tables_sse2+0x15d> DB 60,3 ; cmp $0x3,%al - DB 15,133,220,254,255,255 ; jne 2b13 <_sk_load_tables_sse2+0x2b> + DB 15,133,217,254,255,255 ; jne 2b15 <_sk_load_tables_sse2+0x2b> DB 102,15,110,68,187,8 ; movd 0x8(%ebx,%edi,4),%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 102,15,18,28,187 ; movlpd (%ebx,%edi,4),%xmm3 - DB 233,199,254,255,255 ; jmp 2b13 <_sk_load_tables_sse2+0x2b> + DB 233,196,254,255,255 ; jmp 2b15 <_sk_load_tables_sse2+0x2b> DB 102,15,110,28,187 ; movd (%ebx,%edi,4),%xmm3 - DB 233,189,254,255,255 ; jmp 2b13 <_sk_load_tables_sse2+0x2b> + DB 233,186,254,255,255 ; jmp 2b15 <_sk_load_tables_sse2+0x2b> PUBLIC _sk_load_tables_u16_be_sse2 _sk_load_tables_u16_be_sse2 LABEL PROC @@ -48807,137 +48764,137 @@ _sk_load_tables_u16_be_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,44 ; sub $0x2c,%esp - DB 232,0,0,0,0 ; call 2c64 <_sk_load_tables_u16_be_sse2+0xe> + DB 232,0,0,0,0 ; call 2c69 <_sk_load_tables_u16_be_sse2+0xe> DB 90 ; pop %edx - DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 139,69,8 ; mov 0x8(%ebp),%eax - DB 139,24 ; mov (%eax),%ebx - DB 139,64,8 ; mov 0x8(%eax),%eax - DB 193,227,2 ; shl $0x2,%ebx - DB 133,192 ; test %eax,%eax + DB 139,69,12 ; mov 0xc(%ebp),%eax + DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,57 ; mov (%ecx),%edi - DB 139,15 ; mov (%edi),%ecx - DB 190,255,0,255,0 ; mov $0xff00ff,%esi - DB 15,133,166,1,0,0 ; jne 2e2a <_sk_load_tables_u16_be_sse2+0x1d4> - DB 102,15,16,4,89 ; movupd (%ecx,%ebx,2),%xmm0 - DB 243,15,111,76,89,16 ; movdqu 0x10(%ecx,%ebx,2),%xmm1 + DB 139,89,8 ; mov 0x8(%ecx),%ebx + DB 193,231,2 ; shl $0x2,%edi + DB 133,219 ; test %ebx,%ebx + DB 139,48 ; mov (%eax),%esi + DB 139,6 ; mov (%esi),%eax + DB 15,133,210,1,0,0 ; jne 2e56 <_sk_load_tables_u16_be_sse2+0x1fb> + DB 102,15,16,4,120 ; movupd (%eax,%edi,2),%xmm0 + DB 243,15,111,76,120,16 ; movdqu 0x10(%eax,%edi,2),%xmm1 DB 102,15,40,216 ; movapd %xmm0,%xmm3 DB 102,15,97,217 ; punpcklwd %xmm1,%xmm3 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,15,105,216 ; punpckhwd %xmm0,%xmm3 - DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,126,209 ; movd %xmm2,%ecx - DB 102,15,112,202,78 ; pshufd $0x4e,%xmm2,%xmm1 - DB 102,15,112,234,231 ; pshufd $0xe7,%xmm2,%xmm5 - DB 33,241 ; and %esi,%ecx - DB 137,77,216 ; mov %ecx,-0x28(%ebp) - DB 33,240 ; and %esi,%eax - DB 137,69,220 ; mov %eax,-0x24(%ebp) - DB 243,15,126,69,216 ; movq -0x28(%ebp),%xmm0 - DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 102,15,97,194 ; punpcklwd %xmm2,%xmm0 + DB 102,15,126,77,208 ; movd %xmm1,-0x30(%ebp) + DB 102,15,112,193,229 ; pshufd $0xe5,%xmm1,%xmm0 + DB 102,15,126,69,212 ; movd %xmm0,-0x2c(%ebp) + DB 243,15,126,69,208 ; movq -0x30(%ebp),%xmm0 + DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 + DB 102,15,112,209,231 ; pshufd $0xe7,%xmm1,%xmm2 + DB 102,15,126,85,220 ; movd %xmm2,-0x24(%ebp) + DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1 + DB 102,15,126,77,216 ; movd %xmm1,-0x28(%ebp) + DB 243,15,126,77,216 ; movq -0x28(%ebp),%xmm1 + DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 + DB 102,15,126,93,224 ; movd %xmm3,-0x20(%ebp) + DB 102,15,112,211,229 ; pshufd $0xe5,%xmm3,%xmm2 + DB 102,15,126,85,228 ; movd %xmm2,-0x1c(%ebp) + DB 242,15,112,192,232 ; pshuflw $0xe8,%xmm0,%xmm0 + DB 243,15,112,192,232 ; pshufhw $0xe8,%xmm0,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,111,146,55,165,0,0 ; movdqa 0xa537(%edx),%xmm2 + DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,15,112,208,229 ; pshufd $0xe5,%xmm0,%xmm2 + DB 102,15,97,196 ; punpcklwd %xmm4,%xmm0 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 DB 102,15,112,240,231 ; pshufd $0xe7,%xmm0,%xmm6 DB 102,15,126,240 ; movd %xmm6,%eax - DB 139,79,4 ; mov 0x4(%edi),%ecx - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 + DB 139,126,4 ; mov 0x4(%esi),%edi + DB 243,15,16,52,135 ; movss (%edi,%eax,4),%xmm6 + DB 102,15,126,232 ; movd %xmm5,%eax + DB 243,15,16,44,135 ; movss (%edi,%eax,4),%xmm5 DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,248,78 ; pshufd $0x4e,%xmm0,%xmm7 - DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 102,15,112,248,229 ; pshufd $0xe5,%xmm0,%xmm7 + DB 243,15,16,4,135 ; movss (%edi,%eax,4),%xmm0 DB 102,15,126,248 ; movd %xmm7,%eax - DB 243,15,16,60,129 ; movss (%ecx,%eax,4),%xmm7 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 139,79,8 ; mov 0x8(%edi),%ecx - DB 33,240 ; and %esi,%eax - DB 137,69,228 ; mov %eax,-0x1c(%ebp) - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,203,229 ; pshufd $0xe5,%xmm3,%xmm1 - DB 33,240 ; and %esi,%eax - DB 137,69,224 ; mov %eax,-0x20(%ebp) - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,15,126,109,224 ; movq -0x20(%ebp),%xmm5 - DB 102,15,97,236 ; punpcklwd %xmm4,%xmm5 - DB 102,15,112,205,231 ; pshufd $0xe7,%xmm5,%xmm1 - DB 102,15,126,203 ; movd %xmm1,%ebx - DB 243,15,16,36,153 ; movss (%ecx,%ebx,4),%xmm4 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 102,15,112,205,229 ; pshufd $0xe5,%xmm5,%xmm1 - DB 102,15,126,203 ; movd %xmm1,%ebx - DB 243,15,16,52,153 ; movss (%ecx,%ebx,4),%xmm6 + DB 243,15,16,60,135 ; movss (%edi,%eax,4),%xmm7 + DB 139,70,8 ; mov 0x8(%esi),%eax + DB 15,20,238 ; unpcklps %xmm6,%xmm5 DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 102,15,126,235 ; movd %xmm5,%ebx - DB 243,15,16,12,153 ; movss (%ecx,%ebx,4),%xmm1 - DB 102,15,112,237,78 ; pshufd $0x4e,%xmm5,%xmm5 - DB 102,15,126,235 ; movd %xmm5,%ebx - DB 243,15,16,44,153 ; movss (%ecx,%ebx,4),%xmm5 - DB 102,15,126,217 ; movd %xmm3,%ecx - DB 139,127,12 ; mov 0xc(%edi),%edi - DB 33,241 ; and %esi,%ecx - DB 137,77,232 ; mov %ecx,-0x18(%ebp) - DB 33,240 ; and %esi,%eax - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 137,69,236 ; mov %eax,-0x14(%ebp) - DB 243,15,126,85,232 ; movq -0x18(%ebp),%xmm2 - DB 15,87,228 ; xorps %xmm4,%xmm4 - DB 102,15,97,212 ; punpcklwd %xmm4,%xmm2 - DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 102,15,112,226,229 ; pshufd $0xe5,%xmm2,%xmm4 - DB 102,15,112,234,231 ; pshufd $0xe7,%xmm2,%xmm5 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 243,15,16,44,135 ; movss (%edi,%eax,4),%xmm5 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,15,16,36,135 ; movss (%edi,%eax,4),%xmm4 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 - DB 243,15,16,20,135 ; movss (%edi,%eax,4),%xmm2 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 243,15,16,52,135 ; movss (%edi,%eax,4),%xmm6 - DB 15,20,229 ; unpcklps %xmm5,%xmm4 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 + DB 242,15,112,201,232 ; pshuflw $0xe8,%xmm1,%xmm1 + DB 243,15,112,201,232 ; pshufhw $0xe8,%xmm1,%xmm1 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,219,202 ; pand %xmm2,%xmm1 + DB 102,15,97,204 ; punpcklwd %xmm4,%xmm1 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,112,241,231 ; pshufd $0xe7,%xmm1,%xmm6 + DB 102,15,126,247 ; movd %xmm6,%edi + DB 243,15,16,52,184 ; movss (%eax,%edi,4),%xmm6 + DB 102,15,126,239 ; movd %xmm5,%edi + DB 243,15,16,44,184 ; movss (%eax,%edi,4),%xmm5 + DB 102,15,126,207 ; movd %xmm1,%edi + DB 102,15,112,249,229 ; pshufd $0xe5,%xmm1,%xmm7 + DB 243,15,16,12,184 ; movss (%eax,%edi,4),%xmm1 + DB 102,15,126,255 ; movd %xmm7,%edi + DB 243,15,16,60,184 ; movss (%eax,%edi,4),%xmm7 + DB 15,20,238 ; unpcklps %xmm6,%xmm5 + DB 243,15,126,117,224 ; movq -0x20(%ebp),%xmm6 + DB 102,15,97,240 ; punpcklwd %xmm0,%xmm6 + DB 15,20,207 ; unpcklps %xmm7,%xmm1 + DB 102,15,20,205 ; unpcklpd %xmm5,%xmm1 + DB 139,70,12 ; mov 0xc(%esi),%eax + DB 242,15,112,238,232 ; pshuflw $0xe8,%xmm6,%xmm5 + DB 243,15,112,237,232 ; pshufhw $0xe8,%xmm5,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,15,219,234 ; pand %xmm2,%xmm5 + DB 102,15,97,236 ; punpcklwd %xmm4,%xmm5 + DB 102,15,112,213,78 ; pshufd $0x4e,%xmm5,%xmm2 + DB 102,15,112,245,231 ; pshufd $0xe7,%xmm5,%xmm6 + DB 102,15,126,246 ; movd %xmm6,%esi + DB 243,15,16,52,176 ; movss (%eax,%esi,4),%xmm6 + DB 102,15,126,214 ; movd %xmm2,%esi + DB 243,15,16,60,176 ; movss (%eax,%esi,4),%xmm7 + DB 102,15,126,238 ; movd %xmm5,%esi + DB 102,15,112,237,229 ; pshufd $0xe5,%xmm5,%xmm5 + DB 243,15,16,20,176 ; movss (%eax,%esi,4),%xmm2 + DB 102,15,126,238 ; movd %xmm5,%esi + DB 243,15,16,44,176 ; movss (%eax,%esi,4),%xmm5 + DB 15,20,254 ; unpcklps %xmm6,%xmm7 + DB 15,20,213 ; unpcklps %xmm5,%xmm2 + DB 102,15,20,215 ; unpcklpd %xmm7,%xmm2 DB 102,15,112,219,78 ; pshufd $0x4e,%xmm3,%xmm3 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,113,244,8 ; psllw $0x8,%xmm4 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,15,113,245,8 ; psllw $0x8,%xmm5 DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,235,220 ; por %xmm4,%xmm3 - DB 102,15,126,93,208 ; movd %xmm3,-0x30(%ebp) + DB 102,15,235,221 ; por %xmm5,%xmm3 + DB 102,15,126,93,232 ; movd %xmm3,-0x18(%ebp) DB 102,15,112,219,229 ; pshufd $0xe5,%xmm3,%xmm3 - DB 102,15,126,93,212 ; movd %xmm3,-0x2c(%ebp) - DB 243,15,126,93,208 ; movq -0x30(%ebp),%xmm3 - DB 102,15,97,223 ; punpcklwd %xmm7,%xmm3 + DB 102,15,126,93,236 ; movd %xmm3,-0x14(%ebp) + DB 243,15,126,93,232 ; movq -0x18(%ebp),%xmm3 + DB 102,15,97,220 ; punpcklwd %xmm4,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,154,76,166,0,0 ; mulps 0xa64c(%edx),%xmm3 + DB 15,89,154,71,165,0,0 ; mulps 0xa547(%edx),%xmm3 DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 137,193 ; mov %eax,%ecx - DB 141,65,8 ; lea 0x8(%ecx),%eax + DB 137,194 ; mov %eax,%edx + DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp DB 80 ; push %eax - DB 255,117,8 ; pushl 0x8(%ebp) - DB 255,81,4 ; call *0x4(%ecx) + DB 81 ; push %ecx + DB 255,82,4 ; call *0x4(%edx) DB 131,196,60 ; add $0x3c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx DB 93 ; pop %ebp DB 195 ; ret - DB 242,15,16,4,89 ; movsd (%ecx,%ebx,2),%xmm0 + DB 242,15,16,4,120 ; movsd (%eax,%edi,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 131,248,1 ; cmp $0x1,%eax - DB 15,132,83,254,255,255 ; je 2c8f <_sk_load_tables_u16_be_sse2+0x39> - DB 102,15,22,68,89,8 ; movhpd 0x8(%ecx,%ebx,2),%xmm0 - DB 131,248,3 ; cmp $0x3,%eax - DB 15,130,68,254,255,255 ; jb 2c8f <_sk_load_tables_u16_be_sse2+0x39> - DB 243,15,126,76,89,16 ; movq 0x10(%ecx,%ebx,2),%xmm1 - DB 233,57,254,255,255 ; jmp 2c8f <_sk_load_tables_u16_be_sse2+0x39> + DB 131,251,1 ; cmp $0x1,%ebx + DB 15,132,39,254,255,255 ; je 2c8f <_sk_load_tables_u16_be_sse2+0x34> + DB 102,15,22,68,120,8 ; movhpd 0x8(%eax,%edi,2),%xmm0 + DB 131,251,3 ; cmp $0x3,%ebx + DB 15,130,24,254,255,255 ; jb 2c8f <_sk_load_tables_u16_be_sse2+0x34> + DB 243,15,126,76,120,16 ; movq 0x10(%eax,%edi,2),%xmm1 + DB 233,13,254,255,255 ; jmp 2c8f <_sk_load_tables_u16_be_sse2+0x34> PUBLIC _sk_load_tables_rgb_u16_be_sse2 _sk_load_tables_rgb_u16_be_sse2 LABEL PROC @@ -48947,141 +48904,138 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,28 ; sub $0x1c,%esp - DB 232,0,0,0,0 ; call 2e64 <_sk_load_tables_rgb_u16_be_sse2+0xe> - DB 88 ; pop %eax - DB 137,69,240 ; mov %eax,-0x10(%ebp) - DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 139,85,8 ; mov 0x8(%ebp),%edx - DB 139,74,8 ; mov 0x8(%edx),%ecx - DB 107,26,3 ; imul $0x3,(%edx),%ebx - DB 133,201 ; test %ecx,%ecx - DB 139,56 ; mov (%eax),%edi - DB 139,7 ; mov (%edi),%eax - DB 190,255,0,255,0 ; mov $0xff00ff,%esi - DB 15,133,141,1,0,0 ; jne 3012 <_sk_load_tables_rgb_u16_be_sse2+0x1bc> - DB 243,15,111,12,88 ; movdqu (%eax,%ebx,2),%xmm1 - DB 243,15,111,92,88,8 ; movdqu 0x8(%eax,%ebx,2),%xmm3 + DB 232,0,0,0,0 ; call 2e90 <_sk_load_tables_rgb_u16_be_sse2+0xe> + DB 90 ; pop %edx + DB 139,117,12 ; mov 0xc(%ebp),%esi + DB 139,77,8 ; mov 0x8(%ebp),%ecx + DB 139,65,8 ; mov 0x8(%ecx),%eax + DB 107,57,3 ; imul $0x3,(%ecx),%edi + DB 133,192 ; test %eax,%eax + DB 139,54 ; mov (%esi),%esi + DB 139,30 ; mov (%esi),%ebx + DB 15,133,179,1,0,0 ; jne 305c <_sk_load_tables_rgb_u16_be_sse2+0x1da> + DB 243,15,111,4,123 ; movdqu (%ebx,%edi,2),%xmm0 + DB 243,15,111,92,123,8 ; movdqu 0x8(%ebx,%edi,2),%xmm3 DB 102,15,115,219,4 ; psrldq $0x4,%xmm3 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,15,115,217,6 ; psrldq $0x6,%xmm1 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 DB 102,15,115,218,6 ; psrldq $0x6,%xmm2 - DB 102,15,97,203 ; punpcklwd %xmm3,%xmm1 - DB 102,15,97,194 ; punpcklwd %xmm2,%xmm0 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,15,97,202 ; punpcklwd %xmm2,%xmm1 + DB 102,15,97,195 ; punpcklwd %xmm3,%xmm0 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,97,209 ; punpcklwd %xmm1,%xmm2 + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 + DB 102,15,126,85,216 ; movd %xmm2,-0x28(%ebp) + DB 102,15,112,202,229 ; pshufd $0xe5,%xmm2,%xmm1 + DB 102,15,126,77,220 ; movd %xmm1,-0x24(%ebp) + DB 243,15,126,85,216 ; movq -0x28(%ebp),%xmm2 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 - DB 102,15,105,200 ; punpckhwd %xmm0,%xmm1 - DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,126,209 ; movd %xmm2,%ecx - DB 102,15,112,196,229 ; pshufd $0xe5,%xmm4,%xmm0 - DB 102,15,126,195 ; movd %xmm0,%ebx - DB 33,241 ; and %esi,%ecx - DB 137,77,216 ; mov %ecx,-0x28(%ebp) - DB 33,240 ; and %esi,%eax - DB 137,69,220 ; mov %eax,-0x24(%ebp) - DB 243,15,126,69,216 ; movq -0x28(%ebp),%xmm0 - DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 102,15,97,194 ; punpcklwd %xmm2,%xmm0 - DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 102,15,112,216,229 ; pshufd $0xe5,%xmm0,%xmm3 + DB 102,15,126,93,224 ; movd %xmm3,-0x20(%ebp) + DB 102,15,112,203,229 ; pshufd $0xe5,%xmm3,%xmm1 + DB 102,15,126,77,228 ; movd %xmm1,-0x1c(%ebp) + DB 243,15,126,77,224 ; movq -0x20(%ebp),%xmm1 + DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 + DB 102,15,126,69,232 ; movd %xmm0,-0x18(%ebp) + DB 102,15,112,192,229 ; pshufd $0xe5,%xmm0,%xmm0 + DB 102,15,126,69,236 ; movd %xmm0,-0x14(%ebp) + DB 242,15,112,194,232 ; pshuflw $0xe8,%xmm2,%xmm0 + DB 243,15,112,192,232 ; pshufhw $0xe8,%xmm0,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,111,146,16,163,0,0 ; movdqa 0xa310(%edx),%xmm2 + DB 102,15,219,194 ; pand %xmm2,%xmm0 + DB 102,15,239,219 ; pxor %xmm3,%xmm3 + DB 102,15,97,195 ; punpcklwd %xmm3,%xmm0 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 DB 102,15,112,232,231 ; pshufd $0xe7,%xmm0,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax - DB 139,79,4 ; mov 0x4(%edi),%ecx - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 - DB 243,15,16,28,129 ; movss (%ecx,%eax,4),%xmm3 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 102,15,112,225,229 ; pshufd $0xe5,%xmm1,%xmm4 - DB 139,79,8 ; mov 0x8(%edi),%ecx - DB 33,240 ; and %esi,%eax - DB 137,69,224 ; mov %eax,-0x20(%ebp) + DB 139,126,4 ; mov 0x4(%esi),%edi + DB 243,15,16,44,135 ; movss (%edi,%eax,4),%xmm5 DB 102,15,126,224 ; movd %xmm4,%eax - DB 33,243 ; and %esi,%ebx - DB 137,93,228 ; mov %ebx,-0x1c(%ebp) - DB 243,15,126,101,224 ; movq -0x20(%ebp),%xmm4 - DB 102,15,97,231 ; punpcklwd %xmm7,%xmm4 - DB 102,15,112,252,231 ; pshufd $0xe7,%xmm4,%xmm7 - DB 102,15,126,251 ; movd %xmm7,%ebx - DB 243,15,16,60,153 ; movss (%ecx,%ebx,4),%xmm7 - DB 102,15,126,203 ; movd %xmm1,%ebx - DB 102,15,112,204,229 ; pshufd $0xe5,%xmm4,%xmm1 - DB 102,15,126,202 ; movd %xmm1,%edx - DB 243,15,16,4,145 ; movss (%ecx,%edx,4),%xmm0 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 102,15,126,226 ; movd %xmm4,%edx - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 102,15,112,228,78 ; pshufd $0x4e,%xmm4,%xmm4 - DB 102,15,126,226 ; movd %xmm4,%edx - DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 - DB 139,79,12 ; mov 0xc(%edi),%ecx - DB 33,243 ; and %esi,%ebx - DB 137,93,232 ; mov %ebx,-0x18(%ebp) - DB 33,240 ; and %esi,%eax - DB 15,20,221 ; unpcklps %xmm5,%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 137,69,236 ; mov %eax,-0x14(%ebp) - DB 243,15,126,85,232 ; movq -0x18(%ebp),%xmm2 - DB 15,87,237 ; xorps %xmm5,%xmm5 - DB 102,15,97,213 ; punpcklwd %xmm5,%xmm2 - DB 102,15,112,234,229 ; pshufd $0xe5,%xmm2,%xmm5 - DB 102,15,112,242,231 ; pshufd $0xe7,%xmm2,%xmm6 + DB 243,15,16,36,135 ; movss (%edi,%eax,4),%xmm4 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 + DB 243,15,16,4,135 ; movss (%edi,%eax,4),%xmm0 DB 102,15,126,240 ; movd %xmm6,%eax - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,250,78 ; pshufd $0x4e,%xmm2,%xmm7 - DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 - DB 102,15,126,248 ; movd %xmm7,%eax - DB 243,15,16,60,129 ; movss (%ecx,%eax,4),%xmm7 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 243,15,16,52,135 ; movss (%edi,%eax,4),%xmm6 + DB 139,70,8 ; mov 0x8(%esi),%eax + DB 15,20,229 ; unpcklps %xmm5,%xmm4 + DB 15,20,198 ; unpcklps %xmm6,%xmm0 + DB 242,15,112,201,232 ; pshuflw $0xe8,%xmm1,%xmm1 + DB 243,15,112,201,232 ; pshufhw $0xe8,%xmm1,%xmm1 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,219,202 ; pand %xmm2,%xmm1 + DB 102,15,97,203 ; punpcklwd %xmm3,%xmm1 + DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,112,241,231 ; pshufd $0xe7,%xmm1,%xmm6 + DB 102,15,126,247 ; movd %xmm6,%edi + DB 243,15,16,52,184 ; movss (%eax,%edi,4),%xmm6 + DB 102,15,126,239 ; movd %xmm5,%edi + DB 243,15,16,44,184 ; movss (%eax,%edi,4),%xmm5 + DB 102,15,126,207 ; movd %xmm1,%edi + DB 102,15,112,249,229 ; pshufd $0xe5,%xmm1,%xmm7 + DB 243,15,16,12,184 ; movss (%eax,%edi,4),%xmm1 + DB 102,15,126,255 ; movd %xmm7,%edi + DB 243,15,16,60,184 ; movss (%eax,%edi,4),%xmm7 + DB 102,15,20,196 ; unpcklpd %xmm4,%xmm0 + DB 243,15,126,101,232 ; movq -0x18(%ebp),%xmm4 + DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 15,20,213 ; unpcklps %xmm5,%xmm2 + DB 139,70,12 ; mov 0xc(%esi),%eax + DB 242,15,112,228,232 ; pshuflw $0xe8,%xmm4,%xmm4 + DB 243,15,112,228,232 ; pshufhw $0xe8,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,219,226 ; pand %xmm2,%xmm4 + DB 102,15,97,227 ; punpcklwd %xmm3,%xmm4 + DB 102,15,112,212,78 ; pshufd $0x4e,%xmm4,%xmm2 + DB 102,15,112,220,231 ; pshufd $0xe7,%xmm4,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 243,15,16,28,176 ; movss (%eax,%esi,4),%xmm3 + DB 102,15,126,214 ; movd %xmm2,%esi + DB 243,15,16,52,176 ; movss (%eax,%esi,4),%xmm6 + DB 102,15,126,230 ; movd %xmm4,%esi + DB 102,15,112,228,229 ; pshufd $0xe5,%xmm4,%xmm4 + DB 243,15,16,20,176 ; movss (%eax,%esi,4),%xmm2 + DB 102,15,126,230 ; movd %xmm4,%esi + DB 243,15,16,36,176 ; movss (%eax,%esi,4),%xmm4 + DB 15,20,207 ; unpcklps %xmm7,%xmm1 + DB 102,15,20,205 ; unpcklpd %xmm5,%xmm1 + DB 15,20,243 ; unpcklps %xmm3,%xmm6 + DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 102,15,20,214 ; unpcklpd %xmm6,%xmm2 DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 137,193 ; mov %eax,%ecx - DB 141,65,8 ; lea 0x8(%ecx),%eax + DB 137,198 ; mov %eax,%esi + DB 141,70,8 ; lea 0x8(%esi),%eax DB 131,236,8 ; sub $0x8,%esp - DB 139,85,240 ; mov -0x10(%ebp),%edx - DB 15,40,162,236,161,0,0 ; movaps 0xa1ec(%edx),%xmm4 - DB 15,40,195 ; movaps %xmm3,%xmm0 - DB 15,40,220 ; movaps %xmm4,%xmm3 + DB 15,40,154,176,160,0,0 ; movaps 0xa0b0(%edx),%xmm3 DB 80 ; push %eax - DB 255,117,8 ; pushl 0x8(%ebp) - DB 255,81,4 ; call *0x4(%ecx) + DB 81 ; push %ecx + DB 255,86,4 ; call *0x4(%esi) DB 131,196,44 ; add $0x2c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx DB 93 ; pop %ebp DB 195 ; ret - DB 102,15,110,12,88 ; movd (%eax,%ebx,2),%xmm1 - DB 102,15,196,76,88,4,2 ; pinsrw $0x2,0x4(%eax,%ebx,2),%xmm1 + DB 102,15,110,4,123 ; movd (%ebx,%edi,2),%xmm0 + DB 102,15,196,68,123,4,2 ; pinsrw $0x2,0x4(%ebx,%edi,2),%xmm0 DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 131,249,1 ; cmp $0x1,%ecx - DB 117,13 ; jne 3034 <_sk_load_tables_rgb_u16_be_sse2+0x1de> + DB 131,248,1 ; cmp $0x1,%eax + DB 117,13 ; jne 307e <_sk_load_tables_rgb_u16_be_sse2+0x1fc> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,115,254,255,255 ; jmp 2ea7 <_sk_load_tables_rgb_u16_be_sse2+0x51> - DB 102,15,110,68,88,6 ; movd 0x6(%eax,%ebx,2),%xmm0 - DB 102,15,196,68,88,10,2 ; pinsrw $0x2,0xa(%eax,%ebx,2),%xmm0 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 233,77,254,255,255 ; jmp 2ecb <_sk_load_tables_rgb_u16_be_sse2+0x49> + DB 102,15,110,76,123,6 ; movd 0x6(%ebx,%edi,2),%xmm1 + DB 102,15,196,76,123,10,2 ; pinsrw $0x2,0xa(%ebx,%edi,2),%xmm1 DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 131,249,3 ; cmp $0x3,%ecx - DB 114,18 ; jb 305c <_sk_load_tables_rgb_u16_be_sse2+0x206> - DB 102,15,110,92,88,12 ; movd 0xc(%eax,%ebx,2),%xmm3 - DB 102,15,196,92,88,16,2 ; pinsrw $0x2,0x10(%eax,%ebx,2),%xmm3 - DB 233,75,254,255,255 ; jmp 2ea7 <_sk_load_tables_rgb_u16_be_sse2+0x51> + DB 131,248,3 ; cmp $0x3,%eax + DB 114,18 ; jb 30a6 <_sk_load_tables_rgb_u16_be_sse2+0x224> + DB 102,15,110,92,123,12 ; movd 0xc(%ebx,%edi,2),%xmm3 + DB 102,15,196,92,123,16,2 ; pinsrw $0x2,0x10(%ebx,%edi,2),%xmm3 + DB 233,37,254,255,255 ; jmp 2ecb <_sk_load_tables_rgb_u16_be_sse2+0x49> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 233,66,254,255,255 ; jmp 2ea7 <_sk_load_tables_rgb_u16_be_sse2+0x51> + DB 233,28,254,255,255 ; jmp 2ecb <_sk_load_tables_rgb_u16_be_sse2+0x49> PUBLIC _sk_byte_tables_sse2 _sk_byte_tables_sse2 LABEL PROC @@ -49091,11 +49045,11 @@ _sk_byte_tables_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 3073 <_sk_byte_tables_sse2+0xe> + DB 232,0,0,0,0 ; call 30bd <_sk_byte_tables_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,8 ; mov (%eax),%ecx - DB 15,40,162,173,160,0,0 ; movaps 0xa0ad(%edx),%xmm4 + DB 15,40,162,83,159,0,0 ; movaps 0x9f53(%edx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 @@ -49114,10 +49068,10 @@ _sk_byte_tables_sse2 LABEL PROC DB 15,182,4,7 ; movzbl (%edi,%eax,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 9,240 ; or %esi,%eax - DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 + DB 102,15,110,192 ; movd %eax,%xmm0 DB 139,65,4 ; mov 0x4(%ecx),%eax DB 102,15,196,195,1 ; pinsrw $0x1,%ebx,%xmm0 - DB 15,40,170,205,161,0,0 ; movaps 0xa1cd(%edx),%xmm5 + DB 15,40,170,115,160,0,0 ; movaps 0xa073(%edx),%xmm5 DB 15,89,204 ; mulps %xmm4,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 @@ -49135,7 +49089,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 15,182,4,56 ; movzbl (%eax,%edi,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 9,208 ; or %edx,%eax - DB 102,15,196,200,0 ; pinsrw $0x0,%eax,%xmm1 + DB 102,15,110,200 ; movd %eax,%xmm1 DB 102,15,196,206,1 ; pinsrw $0x1,%esi,%xmm1 DB 139,65,8 ; mov 0x8(%ecx),%eax DB 15,89,212 ; mulps %xmm4,%xmm2 @@ -49155,7 +49109,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 15,182,4,56 ; movzbl (%eax,%edi,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 9,208 ; or %edx,%eax - DB 102,15,196,208,0 ; pinsrw $0x0,%eax,%xmm2 + DB 102,15,110,208 ; movd %eax,%xmm2 DB 102,15,196,214,1 ; pinsrw $0x1,%esi,%xmm2 DB 139,65,12 ; mov 0xc(%ecx),%eax DB 15,89,220 ; mulps %xmm4,%xmm3 @@ -49175,7 +49129,7 @@ _sk_byte_tables_sse2 LABEL PROC DB 15,182,4,16 ; movzbl (%eax,%edx,1),%eax DB 193,224,8 ; shl $0x8,%eax DB 9,200 ; or %ecx,%eax - DB 102,15,196,216,0 ; pinsrw $0x0,%eax,%xmm3 + DB 102,15,110,216 ; movd %eax,%xmm3 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,96,196 ; punpcklbw %xmm4,%xmm0 DB 102,15,97,196 ; punpcklwd %xmm4,%xmm0 @@ -49240,12 +49194,12 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 15,182,60,31 ; movzbl (%edi,%ebx,1),%edi DB 193,231,8 ; shl $0x8,%edi DB 9,215 ; or %edx,%edi - DB 232,0,0,0,0 ; call 3273 <_sk_byte_tables_rgb_sse2+0x65> + DB 232,0,0,0,0 ; call 32b9 <_sk_byte_tables_rgb_sse2+0x65> DB 90 ; pop %edx - DB 102,15,196,199,0 ; pinsrw $0x0,%edi,%xmm0 + DB 102,15,110,199 ; movd %edi,%xmm0 DB 139,121,4 ; mov 0x4(%ecx),%edi DB 102,15,196,198,1 ; pinsrw $0x1,%esi,%xmm0 - DB 15,40,170,205,159,0,0 ; movaps 0x9fcd(%edx),%xmm5 + DB 15,40,170,119,158,0,0 ; movaps 0x9e77(%edx),%xmm5 DB 15,89,204 ; mulps %xmm4,%xmm1 DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1 DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 @@ -49263,7 +49217,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 15,182,60,31 ; movzbl (%edi,%ebx,1),%edi DB 193,231,8 ; shl $0x8,%edi DB 9,215 ; or %edx,%edi - DB 102,15,196,207,0 ; pinsrw $0x0,%edi,%xmm1 + DB 102,15,110,207 ; movd %edi,%xmm1 DB 102,15,196,206,1 ; pinsrw $0x1,%esi,%xmm1 DB 139,73,8 ; mov 0x8(%ecx),%ecx DB 15,89,212 ; mulps %xmm4,%xmm2 @@ -49283,7 +49237,7 @@ _sk_byte_tables_rgb_sse2 LABEL PROC DB 15,182,12,49 ; movzbl (%ecx,%esi,1),%ecx DB 193,225,8 ; shl $0x8,%ecx DB 9,209 ; or %edx,%ecx - DB 102,15,196,209,0 ; pinsrw $0x0,%ecx,%xmm2 + DB 102,15,110,209 ; movd %ecx,%xmm2 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,96,196 ; punpcklbw %xmm4,%xmm0 DB 102,15,97,196 ; punpcklwd %xmm4,%xmm0 @@ -49325,20 +49279,20 @@ _sk_table_r_sse2 LABEL PROC DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,89,224 ; mulps %xmm0,%xmm4 DB 102,15,91,196 ; cvtps2dq %xmm4,%xmm0 - DB 102,15,112,224,229 ; pshufd $0xe5,%xmm0,%xmm4 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 DB 102,15,112,232,231 ; pshufd $0xe7,%xmm0,%xmm5 DB 102,15,126,233 ; movd %xmm5,%ecx DB 243,15,16,44,138 ; movss (%edx,%ecx,4),%xmm5 DB 102,15,126,225 ; movd %xmm4,%ecx DB 243,15,16,36,138 ; movss (%edx,%ecx,4),%xmm4 DB 102,15,126,193 ; movd %xmm0,%ecx - DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 DB 243,15,16,4,138 ; movss (%edx,%ecx,4),%xmm0 DB 102,15,126,241 ; movd %xmm6,%ecx DB 243,15,16,52,138 ; movss (%edx,%ecx,4),%xmm6 DB 15,20,229 ; unpcklps %xmm5,%xmm4 DB 15,20,198 ; unpcklps %xmm6,%xmm0 - DB 15,20,196 ; unpcklps %xmm4,%xmm0 + DB 102,15,20,196 ; unpcklpd %xmm4,%xmm0 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp DB 81 ; push %ecx @@ -49363,20 +49317,20 @@ _sk_table_g_sse2 LABEL PROC DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,89,225 ; mulps %xmm1,%xmm4 DB 102,15,91,204 ; cvtps2dq %xmm4,%xmm1 - DB 102,15,112,225,229 ; pshufd $0xe5,%xmm1,%xmm4 + DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 DB 102,15,112,233,231 ; pshufd $0xe7,%xmm1,%xmm5 DB 102,15,126,233 ; movd %xmm5,%ecx DB 243,15,16,44,138 ; movss (%edx,%ecx,4),%xmm5 DB 102,15,126,225 ; movd %xmm4,%ecx DB 243,15,16,36,138 ; movss (%edx,%ecx,4),%xmm4 DB 102,15,126,201 ; movd %xmm1,%ecx - DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 + DB 102,15,112,241,229 ; pshufd $0xe5,%xmm1,%xmm6 DB 243,15,16,12,138 ; movss (%edx,%ecx,4),%xmm1 DB 102,15,126,241 ; movd %xmm6,%ecx DB 243,15,16,52,138 ; movss (%edx,%ecx,4),%xmm6 DB 15,20,229 ; unpcklps %xmm5,%xmm4 DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 + DB 102,15,20,204 ; unpcklpd %xmm4,%xmm1 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp DB 81 ; push %ecx @@ -49401,20 +49355,20 @@ _sk_table_b_sse2 LABEL PROC DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,89,226 ; mulps %xmm2,%xmm4 DB 102,15,91,212 ; cvtps2dq %xmm4,%xmm2 - DB 102,15,112,226,229 ; pshufd $0xe5,%xmm2,%xmm4 + DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 DB 102,15,112,234,231 ; pshufd $0xe7,%xmm2,%xmm5 DB 102,15,126,233 ; movd %xmm5,%ecx DB 243,15,16,44,138 ; movss (%edx,%ecx,4),%xmm5 DB 102,15,126,225 ; movd %xmm4,%ecx DB 243,15,16,36,138 ; movss (%edx,%ecx,4),%xmm4 DB 102,15,126,209 ; movd %xmm2,%ecx - DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 + DB 102,15,112,242,229 ; pshufd $0xe5,%xmm2,%xmm6 DB 243,15,16,20,138 ; movss (%edx,%ecx,4),%xmm2 DB 102,15,126,241 ; movd %xmm6,%ecx DB 243,15,16,52,138 ; movss (%edx,%ecx,4),%xmm6 DB 15,20,229 ; unpcklps %xmm5,%xmm4 DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 102,15,20,212 ; unpcklpd %xmm4,%xmm2 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp DB 81 ; push %ecx @@ -49439,20 +49393,20 @@ _sk_table_a_sse2 LABEL PROC DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,89,227 ; mulps %xmm3,%xmm4 DB 102,15,91,220 ; cvtps2dq %xmm4,%xmm3 - DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 + DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 DB 102,15,112,235,231 ; pshufd $0xe7,%xmm3,%xmm5 DB 102,15,126,233 ; movd %xmm5,%ecx DB 243,15,16,44,138 ; movss (%edx,%ecx,4),%xmm5 DB 102,15,126,225 ; movd %xmm4,%ecx DB 243,15,16,36,138 ; movss (%edx,%ecx,4),%xmm4 DB 102,15,126,217 ; movd %xmm3,%ecx - DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 + DB 102,15,112,243,229 ; pshufd $0xe5,%xmm3,%xmm6 DB 243,15,16,28,138 ; movss (%edx,%ecx,4),%xmm3 DB 102,15,126,241 ; movd %xmm6,%ecx DB 243,15,16,52,138 ; movss (%edx,%ecx,4),%xmm6 DB 15,20,229 ; unpcklps %xmm5,%xmm4 DB 15,20,222 ; unpcklps %xmm6,%xmm3 - DB 15,20,220 ; unpcklps %xmm4,%xmm3 + DB 102,15,20,220 ; unpcklpd %xmm4,%xmm3 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp DB 81 ; push %ecx @@ -49471,7 +49425,7 @@ _sk_parametric_r_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,209 ; movaps %xmm1,%xmm2 DB 15,40,200 ; movaps %xmm0,%xmm1 - DB 232,0,0,0,0 ; call 3542 <_sk_parametric_r_sse2+0x18> + DB 232,0,0,0,0 ; call 3589 <_sk_parametric_r_sse2+0x18> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,8 ; mov (%eax),%ecx @@ -49491,15 +49445,15 @@ _sk_parametric_r_sse2 LABEL PROC DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,88,245 ; addps %xmm5,%xmm6 DB 15,91,238 ; cvtdq2ps %xmm6,%xmm5 - DB 15,89,170,126,157,0,0 ; mulps 0x9d7e(%edx),%xmm5 - DB 15,84,178,142,157,0,0 ; andps 0x9d8e(%edx),%xmm6 - DB 15,86,178,254,154,0,0 ; orps 0x9afe(%edx),%xmm6 - DB 15,88,170,158,157,0,0 ; addps 0x9d9e(%edx),%xmm5 - DB 15,40,186,174,157,0,0 ; movaps 0x9dae(%edx),%xmm7 + DB 15,89,170,55,156,0,0 ; mulps 0x9c37(%edx),%xmm5 + DB 15,84,178,71,156,0,0 ; andps 0x9c47(%edx),%xmm6 + DB 15,86,178,167,153,0,0 ; orps 0x99a7(%edx),%xmm6 + DB 15,88,170,87,156,0,0 ; addps 0x9c57(%edx),%xmm5 + DB 15,40,186,103,156,0,0 ; movaps 0x9c67(%edx),%xmm7 DB 15,89,254 ; mulps %xmm6,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,88,178,190,157,0,0 ; addps 0x9dbe(%edx),%xmm6 - DB 15,40,186,206,157,0,0 ; movaps 0x9dce(%edx),%xmm7 + DB 15,88,178,119,156,0,0 ; addps 0x9c77(%edx),%xmm6 + DB 15,40,186,135,156,0,0 ; movaps 0x9c87(%edx),%xmm7 DB 15,94,254 ; divps %xmm6,%xmm7 DB 243,15,16,49 ; movss (%ecx),%xmm6 DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 @@ -49509,21 +49463,21 @@ _sk_parametric_r_sse2 LABEL PROC DB 15,91,254 ; cvtdq2ps %xmm6,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 - DB 15,40,178,14,155,0,0 ; movaps 0x9b0e(%edx),%xmm6 + DB 15,40,178,183,153,0,0 ; movaps 0x99b7(%edx),%xmm6 DB 15,84,198 ; andps %xmm6,%xmm0 DB 15,92,248 ; subps %xmm0,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,92,199 ; subps %xmm7,%xmm0 - DB 15,88,170,222,157,0,0 ; addps 0x9dde(%edx),%xmm5 - DB 15,40,186,238,157,0,0 ; movaps 0x9dee(%edx),%xmm7 + DB 15,88,170,151,156,0,0 ; addps 0x9c97(%edx),%xmm5 + DB 15,40,186,167,156,0,0 ; movaps 0x9ca7(%edx),%xmm7 DB 15,89,248 ; mulps %xmm0,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,40,186,254,157,0,0 ; movaps 0x9dfe(%edx),%xmm7 + DB 15,40,186,183,156,0,0 ; movaps 0x9cb7(%edx),%xmm7 DB 15,92,248 ; subps %xmm0,%xmm7 - DB 15,40,130,14,158,0,0 ; movaps 0x9e0e(%edx),%xmm0 + DB 15,40,130,199,156,0,0 ; movaps 0x9cc7(%edx),%xmm0 DB 15,94,199 ; divps %xmm7,%xmm0 DB 15,88,197 ; addps %xmm5,%xmm0 - DB 15,89,130,30,158,0,0 ; mulps 0x9e1e(%edx),%xmm0 + DB 15,89,130,215,156,0,0 ; mulps 0x9cd7(%edx),%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 243,15,16,105,20 ; movss 0x14(%ecx),%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -49555,7 +49509,7 @@ _sk_parametric_g_sse2 LABEL PROC DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,208 ; movaps %xmm0,%xmm2 - DB 232,0,0,0,0 ; call 3679 <_sk_parametric_g_sse2+0x15> + DB 232,0,0,0,0 ; call 36c0 <_sk_parametric_g_sse2+0x15> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,8 ; mov (%eax),%ecx @@ -49575,15 +49529,15 @@ _sk_parametric_g_sse2 LABEL PROC DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,88,245 ; addps %xmm5,%xmm6 DB 15,91,238 ; cvtdq2ps %xmm6,%xmm5 - DB 15,89,170,71,156,0,0 ; mulps 0x9c47(%edx),%xmm5 - DB 15,84,178,87,156,0,0 ; andps 0x9c57(%edx),%xmm6 - DB 15,86,178,199,153,0,0 ; orps 0x99c7(%edx),%xmm6 - DB 15,88,170,103,156,0,0 ; addps 0x9c67(%edx),%xmm5 - DB 15,40,186,119,156,0,0 ; movaps 0x9c77(%edx),%xmm7 + DB 15,89,170,0,155,0,0 ; mulps 0x9b00(%edx),%xmm5 + DB 15,84,178,16,155,0,0 ; andps 0x9b10(%edx),%xmm6 + DB 15,86,178,112,152,0,0 ; orps 0x9870(%edx),%xmm6 + DB 15,88,170,32,155,0,0 ; addps 0x9b20(%edx),%xmm5 + DB 15,40,186,48,155,0,0 ; movaps 0x9b30(%edx),%xmm7 DB 15,89,254 ; mulps %xmm6,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,88,178,135,156,0,0 ; addps 0x9c87(%edx),%xmm6 - DB 15,40,186,151,156,0,0 ; movaps 0x9c97(%edx),%xmm7 + DB 15,88,178,64,155,0,0 ; addps 0x9b40(%edx),%xmm6 + DB 15,40,186,80,155,0,0 ; movaps 0x9b50(%edx),%xmm7 DB 15,94,254 ; divps %xmm6,%xmm7 DB 243,15,16,49 ; movss (%ecx),%xmm6 DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 @@ -49593,21 +49547,21 @@ _sk_parametric_g_sse2 LABEL PROC DB 15,91,254 ; cvtdq2ps %xmm6,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 - DB 15,40,178,215,153,0,0 ; movaps 0x99d7(%edx),%xmm6 + DB 15,40,178,128,152,0,0 ; movaps 0x9880(%edx),%xmm6 DB 15,84,198 ; andps %xmm6,%xmm0 DB 15,92,248 ; subps %xmm0,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,92,199 ; subps %xmm7,%xmm0 - DB 15,88,170,167,156,0,0 ; addps 0x9ca7(%edx),%xmm5 - DB 15,40,186,183,156,0,0 ; movaps 0x9cb7(%edx),%xmm7 + DB 15,88,170,96,155,0,0 ; addps 0x9b60(%edx),%xmm5 + DB 15,40,186,112,155,0,0 ; movaps 0x9b70(%edx),%xmm7 DB 15,89,248 ; mulps %xmm0,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,40,186,199,156,0,0 ; movaps 0x9cc7(%edx),%xmm7 + DB 15,40,186,128,155,0,0 ; movaps 0x9b80(%edx),%xmm7 DB 15,92,248 ; subps %xmm0,%xmm7 - DB 15,40,130,215,156,0,0 ; movaps 0x9cd7(%edx),%xmm0 + DB 15,40,130,144,155,0,0 ; movaps 0x9b90(%edx),%xmm0 DB 15,94,199 ; divps %xmm7,%xmm0 DB 15,88,197 ; addps %xmm5,%xmm0 - DB 15,89,130,231,156,0,0 ; mulps 0x9ce7(%edx),%xmm0 + DB 15,89,130,160,155,0,0 ; mulps 0x9ba0(%edx),%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 243,15,16,105,20 ; movss 0x14(%ecx),%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -49638,7 +49592,7 @@ _sk_parametric_b_sse2 LABEL PROC DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) DB 15,40,217 ; movaps %xmm1,%xmm3 DB 15,40,200 ; movaps %xmm0,%xmm1 - DB 232,0,0,0,0 ; call 37ad <_sk_parametric_b_sse2+0x15> + DB 232,0,0,0,0 ; call 37f4 <_sk_parametric_b_sse2+0x15> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,8 ; mov (%eax),%ecx @@ -49658,15 +49612,15 @@ _sk_parametric_b_sse2 LABEL PROC DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,88,245 ; addps %xmm5,%xmm6 DB 15,91,238 ; cvtdq2ps %xmm6,%xmm5 - DB 15,89,170,19,155,0,0 ; mulps 0x9b13(%edx),%xmm5 - DB 15,84,178,35,155,0,0 ; andps 0x9b23(%edx),%xmm6 - DB 15,86,178,147,152,0,0 ; orps 0x9893(%edx),%xmm6 - DB 15,88,170,51,155,0,0 ; addps 0x9b33(%edx),%xmm5 - DB 15,40,186,67,155,0,0 ; movaps 0x9b43(%edx),%xmm7 + DB 15,89,170,204,153,0,0 ; mulps 0x99cc(%edx),%xmm5 + DB 15,84,178,220,153,0,0 ; andps 0x99dc(%edx),%xmm6 + DB 15,86,178,60,151,0,0 ; orps 0x973c(%edx),%xmm6 + DB 15,88,170,236,153,0,0 ; addps 0x99ec(%edx),%xmm5 + DB 15,40,186,252,153,0,0 ; movaps 0x99fc(%edx),%xmm7 DB 15,89,254 ; mulps %xmm6,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,88,178,83,155,0,0 ; addps 0x9b53(%edx),%xmm6 - DB 15,40,186,99,155,0,0 ; movaps 0x9b63(%edx),%xmm7 + DB 15,88,178,12,154,0,0 ; addps 0x9a0c(%edx),%xmm6 + DB 15,40,186,28,154,0,0 ; movaps 0x9a1c(%edx),%xmm7 DB 15,94,254 ; divps %xmm6,%xmm7 DB 243,15,16,49 ; movss (%ecx),%xmm6 DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 @@ -49676,21 +49630,21 @@ _sk_parametric_b_sse2 LABEL PROC DB 15,91,254 ; cvtdq2ps %xmm6,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 - DB 15,40,178,163,152,0,0 ; movaps 0x98a3(%edx),%xmm6 + DB 15,40,178,76,151,0,0 ; movaps 0x974c(%edx),%xmm6 DB 15,84,198 ; andps %xmm6,%xmm0 DB 15,92,248 ; subps %xmm0,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,92,199 ; subps %xmm7,%xmm0 - DB 15,88,170,115,155,0,0 ; addps 0x9b73(%edx),%xmm5 - DB 15,40,186,131,155,0,0 ; movaps 0x9b83(%edx),%xmm7 + DB 15,88,170,44,154,0,0 ; addps 0x9a2c(%edx),%xmm5 + DB 15,40,186,60,154,0,0 ; movaps 0x9a3c(%edx),%xmm7 DB 15,89,248 ; mulps %xmm0,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,40,186,147,155,0,0 ; movaps 0x9b93(%edx),%xmm7 + DB 15,40,186,76,154,0,0 ; movaps 0x9a4c(%edx),%xmm7 DB 15,92,248 ; subps %xmm0,%xmm7 - DB 15,40,130,163,155,0,0 ; movaps 0x9ba3(%edx),%xmm0 + DB 15,40,130,92,154,0,0 ; movaps 0x9a5c(%edx),%xmm0 DB 15,94,199 ; divps %xmm7,%xmm0 DB 15,88,197 ; addps %xmm5,%xmm0 - DB 15,89,130,179,155,0,0 ; mulps 0x9bb3(%edx),%xmm0 + DB 15,89,130,108,154,0,0 ; mulps 0x9a6c(%edx),%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 243,15,16,105,20 ; movss 0x14(%ecx),%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -49721,7 +49675,7 @@ _sk_parametric_a_sse2 LABEL PROC DB 15,41,85,232 ; movaps %xmm2,-0x18(%ebp) DB 15,40,209 ; movaps %xmm1,%xmm2 DB 15,40,200 ; movaps %xmm0,%xmm1 - DB 232,0,0,0,0 ; call 38e1 <_sk_parametric_a_sse2+0x15> + DB 232,0,0,0,0 ; call 3928 <_sk_parametric_a_sse2+0x15> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,8 ; mov (%eax),%ecx @@ -49741,15 +49695,15 @@ _sk_parametric_a_sse2 LABEL PROC DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,88,245 ; addps %xmm5,%xmm6 DB 15,91,238 ; cvtdq2ps %xmm6,%xmm5 - DB 15,89,170,223,153,0,0 ; mulps 0x99df(%edx),%xmm5 - DB 15,84,178,239,153,0,0 ; andps 0x99ef(%edx),%xmm6 - DB 15,86,178,95,151,0,0 ; orps 0x975f(%edx),%xmm6 - DB 15,88,170,255,153,0,0 ; addps 0x99ff(%edx),%xmm5 - DB 15,40,186,15,154,0,0 ; movaps 0x9a0f(%edx),%xmm7 + DB 15,89,170,152,152,0,0 ; mulps 0x9898(%edx),%xmm5 + DB 15,84,178,168,152,0,0 ; andps 0x98a8(%edx),%xmm6 + DB 15,86,178,8,150,0,0 ; orps 0x9608(%edx),%xmm6 + DB 15,88,170,184,152,0,0 ; addps 0x98b8(%edx),%xmm5 + DB 15,40,186,200,152,0,0 ; movaps 0x98c8(%edx),%xmm7 DB 15,89,254 ; mulps %xmm6,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,88,178,31,154,0,0 ; addps 0x9a1f(%edx),%xmm6 - DB 15,40,186,47,154,0,0 ; movaps 0x9a2f(%edx),%xmm7 + DB 15,88,178,216,152,0,0 ; addps 0x98d8(%edx),%xmm6 + DB 15,40,186,232,152,0,0 ; movaps 0x98e8(%edx),%xmm7 DB 15,94,254 ; divps %xmm6,%xmm7 DB 243,15,16,49 ; movss (%ecx),%xmm6 DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 @@ -49759,21 +49713,21 @@ _sk_parametric_a_sse2 LABEL PROC DB 15,91,254 ; cvtdq2ps %xmm6,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,194,199,1 ; cmpltps %xmm7,%xmm0 - DB 15,40,178,111,151,0,0 ; movaps 0x976f(%edx),%xmm6 + DB 15,40,178,24,150,0,0 ; movaps 0x9618(%edx),%xmm6 DB 15,84,198 ; andps %xmm6,%xmm0 DB 15,92,248 ; subps %xmm0,%xmm7 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,92,199 ; subps %xmm7,%xmm0 - DB 15,88,170,63,154,0,0 ; addps 0x9a3f(%edx),%xmm5 - DB 15,40,186,79,154,0,0 ; movaps 0x9a4f(%edx),%xmm7 + DB 15,88,170,248,152,0,0 ; addps 0x98f8(%edx),%xmm5 + DB 15,40,186,8,153,0,0 ; movaps 0x9908(%edx),%xmm7 DB 15,89,248 ; mulps %xmm0,%xmm7 DB 15,92,239 ; subps %xmm7,%xmm5 - DB 15,40,186,95,154,0,0 ; movaps 0x9a5f(%edx),%xmm7 + DB 15,40,186,24,153,0,0 ; movaps 0x9918(%edx),%xmm7 DB 15,92,248 ; subps %xmm0,%xmm7 - DB 15,40,130,111,154,0,0 ; movaps 0x9a6f(%edx),%xmm0 + DB 15,40,130,40,153,0,0 ; movaps 0x9928(%edx),%xmm0 DB 15,94,199 ; divps %xmm7,%xmm0 DB 15,88,197 ; addps %xmm5,%xmm0 - DB 15,89,130,127,154,0,0 ; mulps 0x9a7f(%edx),%xmm0 + DB 15,89,130,56,153,0,0 ; mulps 0x9938(%edx),%xmm0 DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0 DB 243,15,16,105,20 ; movss 0x14(%ecx),%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -49803,18 +49757,18 @@ _sk_gamma_sse2 LABEL PROC DB 131,236,40 ; sub $0x28,%esp DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) DB 15,40,216 ; movaps %xmm0,%xmm3 - DB 232,0,0,0,0 ; call 3a12 <_sk_gamma_sse2+0x12> + DB 232,0,0,0,0 ; call 3a59 <_sk_gamma_sse2+0x12> DB 88 ; pop %eax DB 15,91,195 ; cvtdq2ps %xmm3,%xmm0 - DB 15,89,128,174,152,0,0 ; mulps 0x98ae(%eax),%xmm0 - DB 15,84,152,190,152,0,0 ; andps 0x98be(%eax),%xmm3 - DB 15,86,152,46,150,0,0 ; orps 0x962e(%eax),%xmm3 - DB 15,88,128,206,152,0,0 ; addps 0x98ce(%eax),%xmm0 + DB 15,89,128,103,151,0,0 ; mulps 0x9767(%eax),%xmm0 + DB 15,84,152,119,151,0,0 ; andps 0x9777(%eax),%xmm3 + DB 15,86,152,215,148,0,0 ; orps 0x94d7(%eax),%xmm3 + DB 15,88,128,135,151,0,0 ; addps 0x9787(%eax),%xmm0 DB 15,40,227 ; movaps %xmm3,%xmm4 - DB 15,89,160,222,152,0,0 ; mulps 0x98de(%eax),%xmm4 + DB 15,89,160,151,151,0,0 ; mulps 0x9797(%eax),%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,88,152,238,152,0,0 ; addps 0x98ee(%eax),%xmm3 - DB 15,40,160,254,152,0,0 ; movaps 0x98fe(%eax),%xmm4 + DB 15,88,152,167,151,0,0 ; addps 0x97a7(%eax),%xmm3 + DB 15,40,160,183,151,0,0 ; movaps 0x97b7(%eax),%xmm4 DB 15,94,227 ; divps %xmm3,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 DB 139,77,12 ; mov 0xc(%ebp),%ecx @@ -49827,30 +49781,30 @@ _sk_gamma_sse2 LABEL PROC DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 DB 15,40,224 ; movaps %xmm0,%xmm4 DB 15,194,227,1 ; cmpltps %xmm3,%xmm4 - DB 15,84,160,62,150,0,0 ; andps 0x963e(%eax),%xmm4 + DB 15,84,160,231,148,0,0 ; andps 0x94e7(%eax),%xmm4 DB 15,92,220 ; subps %xmm4,%xmm3 DB 15,40,224 ; movaps %xmm0,%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 - DB 15,88,128,14,153,0,0 ; addps 0x990e(%eax),%xmm0 - DB 15,40,176,46,153,0,0 ; movaps 0x992e(%eax),%xmm6 + DB 15,88,128,199,151,0,0 ; addps 0x97c7(%eax),%xmm0 + DB 15,40,176,231,151,0,0 ; movaps 0x97e7(%eax),%xmm6 DB 15,92,244 ; subps %xmm4,%xmm6 - DB 15,40,168,30,153,0,0 ; movaps 0x991e(%eax),%xmm5 + DB 15,40,168,215,151,0,0 ; movaps 0x97d7(%eax),%xmm5 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,40,152,62,153,0,0 ; movaps 0x993e(%eax),%xmm3 + DB 15,40,152,247,151,0,0 ; movaps 0x97f7(%eax),%xmm3 DB 15,40,227 ; movaps %xmm3,%xmm4 DB 15,94,230 ; divps %xmm6,%xmm4 DB 15,88,224 ; addps %xmm0,%xmm4 DB 15,91,193 ; cvtdq2ps %xmm1,%xmm0 - DB 15,89,128,174,152,0,0 ; mulps 0x98ae(%eax),%xmm0 - DB 15,84,136,190,152,0,0 ; andps 0x98be(%eax),%xmm1 - DB 15,86,136,46,150,0,0 ; orps 0x962e(%eax),%xmm1 - DB 15,88,128,206,152,0,0 ; addps 0x98ce(%eax),%xmm0 + DB 15,89,128,103,151,0,0 ; mulps 0x9767(%eax),%xmm0 + DB 15,84,136,119,151,0,0 ; andps 0x9777(%eax),%xmm1 + DB 15,86,136,215,148,0,0 ; orps 0x94d7(%eax),%xmm1 + DB 15,88,128,135,151,0,0 ; addps 0x9787(%eax),%xmm0 DB 15,40,241 ; movaps %xmm1,%xmm6 - DB 15,89,176,222,152,0,0 ; mulps 0x98de(%eax),%xmm6 + DB 15,89,176,151,151,0,0 ; mulps 0x9797(%eax),%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 - DB 15,88,136,238,152,0,0 ; addps 0x98ee(%eax),%xmm1 - DB 15,40,176,254,152,0,0 ; movaps 0x98fe(%eax),%xmm6 + DB 15,88,136,167,151,0,0 ; addps 0x97a7(%eax),%xmm1 + DB 15,40,176,183,151,0,0 ; movaps 0x97b7(%eax),%xmm6 DB 15,94,241 ; divps %xmm1,%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 DB 15,89,199 ; mulps %xmm7,%xmm0 @@ -49858,12 +49812,12 @@ _sk_gamma_sse2 LABEL PROC DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 DB 15,40,240 ; movaps %xmm0,%xmm6 DB 15,194,241,1 ; cmpltps %xmm1,%xmm6 - DB 15,84,176,62,150,0,0 ; andps 0x963e(%eax),%xmm6 + DB 15,84,176,231,148,0,0 ; andps 0x94e7(%eax),%xmm6 DB 15,92,206 ; subps %xmm6,%xmm1 DB 15,40,240 ; movaps %xmm0,%xmm6 DB 15,92,241 ; subps %xmm1,%xmm6 - DB 15,88,128,14,153,0,0 ; addps 0x990e(%eax),%xmm0 - DB 15,40,184,46,153,0,0 ; movaps 0x992e(%eax),%xmm7 + DB 15,88,128,199,151,0,0 ; addps 0x97c7(%eax),%xmm0 + DB 15,40,184,231,151,0,0 ; movaps 0x97e7(%eax),%xmm7 DB 15,92,254 ; subps %xmm6,%xmm7 DB 15,89,245 ; mulps %xmm5,%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 @@ -49871,15 +49825,15 @@ _sk_gamma_sse2 LABEL PROC DB 15,94,207 ; divps %xmm7,%xmm1 DB 15,88,200 ; addps %xmm0,%xmm1 DB 15,91,194 ; cvtdq2ps %xmm2,%xmm0 - DB 15,89,128,174,152,0,0 ; mulps 0x98ae(%eax),%xmm0 - DB 15,84,144,190,152,0,0 ; andps 0x98be(%eax),%xmm2 - DB 15,86,144,46,150,0,0 ; orps 0x962e(%eax),%xmm2 - DB 15,88,128,206,152,0,0 ; addps 0x98ce(%eax),%xmm0 - DB 15,40,176,222,152,0,0 ; movaps 0x98de(%eax),%xmm6 + DB 15,89,128,103,151,0,0 ; mulps 0x9767(%eax),%xmm0 + DB 15,84,144,119,151,0,0 ; andps 0x9777(%eax),%xmm2 + DB 15,86,144,215,148,0,0 ; orps 0x94d7(%eax),%xmm2 + DB 15,88,128,135,151,0,0 ; addps 0x9787(%eax),%xmm0 + DB 15,40,176,151,151,0,0 ; movaps 0x9797(%eax),%xmm6 DB 15,89,242 ; mulps %xmm2,%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 - DB 15,88,144,238,152,0,0 ; addps 0x98ee(%eax),%xmm2 - DB 15,40,176,254,152,0,0 ; movaps 0x98fe(%eax),%xmm6 + DB 15,88,144,167,151,0,0 ; addps 0x97a7(%eax),%xmm2 + DB 15,40,176,183,151,0,0 ; movaps 0x97b7(%eax),%xmm6 DB 15,94,242 ; divps %xmm2,%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 DB 15,89,69,232 ; mulps -0x18(%ebp),%xmm0 @@ -49887,18 +49841,18 @@ _sk_gamma_sse2 LABEL PROC DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 15,40,240 ; movaps %xmm0,%xmm6 DB 15,194,242,1 ; cmpltps %xmm2,%xmm6 - DB 15,84,176,62,150,0,0 ; andps 0x963e(%eax),%xmm6 + DB 15,84,176,231,148,0,0 ; andps 0x94e7(%eax),%xmm6 DB 15,92,214 ; subps %xmm6,%xmm2 DB 15,40,240 ; movaps %xmm0,%xmm6 DB 15,92,242 ; subps %xmm2,%xmm6 - DB 15,88,128,14,153,0,0 ; addps 0x990e(%eax),%xmm0 + DB 15,88,128,199,151,0,0 ; addps 0x97c7(%eax),%xmm0 DB 15,89,238 ; mulps %xmm6,%xmm5 DB 15,92,197 ; subps %xmm5,%xmm0 - DB 15,40,144,46,153,0,0 ; movaps 0x992e(%eax),%xmm2 + DB 15,40,144,231,151,0,0 ; movaps 0x97e7(%eax),%xmm2 DB 15,92,214 ; subps %xmm6,%xmm2 DB 15,94,218 ; divps %xmm2,%xmm3 DB 15,88,216 ; addps %xmm0,%xmm3 - DB 15,40,128,78,153,0,0 ; movaps 0x994e(%eax),%xmm0 + DB 15,40,128,7,152,0,0 ; movaps 0x9807(%eax),%xmm0 DB 15,89,224 ; mulps %xmm0,%xmm4 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 15,89,216 ; mulps %xmm0,%xmm3 @@ -49923,31 +49877,31 @@ _sk_lab_to_xyz_sse2 LABEL PROC DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,208 ; movaps %xmm0,%xmm2 - DB 232,0,0,0,0 ; call 3bfa <_sk_lab_to_xyz_sse2+0x15> + DB 232,0,0,0,0 ; call 3c41 <_sk_lab_to_xyz_sse2+0x15> DB 88 ; pop %eax - DB 15,89,144,118,151,0,0 ; mulps 0x9776(%eax),%xmm2 - DB 15,40,160,38,149,0,0 ; movaps 0x9526(%eax),%xmm4 + DB 15,89,144,47,150,0,0 ; mulps 0x962f(%eax),%xmm2 + DB 15,40,160,207,147,0,0 ; movaps 0x93cf(%eax),%xmm4 DB 15,89,204 ; mulps %xmm4,%xmm1 - DB 15,40,168,134,151,0,0 ; movaps 0x9786(%eax),%xmm5 + DB 15,40,168,63,150,0,0 ; movaps 0x963f(%eax),%xmm5 DB 15,88,205 ; addps %xmm5,%xmm1 DB 15,89,220 ; mulps %xmm4,%xmm3 DB 15,88,221 ; addps %xmm5,%xmm3 - DB 15,88,144,150,151,0,0 ; addps 0x9796(%eax),%xmm2 - DB 15,89,144,166,151,0,0 ; mulps 0x97a6(%eax),%xmm2 - DB 15,89,136,182,151,0,0 ; mulps 0x97b6(%eax),%xmm1 + DB 15,88,144,79,150,0,0 ; addps 0x964f(%eax),%xmm2 + DB 15,89,144,95,150,0,0 ; mulps 0x965f(%eax),%xmm2 + DB 15,89,136,111,150,0,0 ; mulps 0x966f(%eax),%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 - DB 15,89,152,198,151,0,0 ; mulps 0x97c6(%eax),%xmm3 + DB 15,89,152,127,150,0,0 ; mulps 0x967f(%eax),%xmm3 DB 15,40,234 ; movaps %xmm2,%xmm5 DB 15,92,235 ; subps %xmm3,%xmm5 DB 15,40,193 ; movaps %xmm1,%xmm0 DB 15,89,192 ; mulps %xmm0,%xmm0 DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,40,152,214,151,0,0 ; movaps 0x97d6(%eax),%xmm3 + DB 15,40,152,143,150,0,0 ; movaps 0x968f(%eax),%xmm3 DB 15,40,227 ; movaps %xmm3,%xmm4 DB 15,194,224,1 ; cmpltps %xmm0,%xmm4 - DB 15,40,176,230,151,0,0 ; movaps 0x97e6(%eax),%xmm6 + DB 15,40,176,159,150,0,0 ; movaps 0x969f(%eax),%xmm6 DB 15,88,206 ; addps %xmm6,%xmm1 - DB 15,40,184,246,151,0,0 ; movaps 0x97f6(%eax),%xmm7 + DB 15,40,184,175,150,0,0 ; movaps 0x96af(%eax),%xmm7 DB 15,89,207 ; mulps %xmm7,%xmm1 DB 15,84,196 ; andps %xmm4,%xmm0 DB 15,85,225 ; andnps %xmm1,%xmm4 @@ -49972,8 +49926,8 @@ _sk_lab_to_xyz_sse2 LABEL PROC DB 15,85,221 ; andnps %xmm5,%xmm3 DB 15,86,216 ; orps %xmm0,%xmm3 DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 15,89,160,6,152,0,0 ; mulps 0x9806(%eax),%xmm4 - DB 15,89,152,22,152,0,0 ; mulps 0x9816(%eax),%xmm3 + DB 15,89,160,191,150,0,0 ; mulps 0x96bf(%eax),%xmm4 + DB 15,89,152,207,150,0,0 ; mulps 0x96cf(%eax),%xmm3 DB 141,65,4 ; lea 0x4(%ecx),%eax DB 131,236,8 ; sub $0x8,%esp DB 15,40,196 ; movaps %xmm4,%xmm0 @@ -49994,7 +49948,7 @@ _sk_load_a8_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 3cea <_sk_load_a8_sse2+0xe> + DB 232,0,0,0,0 ; call 3d31 <_sk_load_a8_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50005,13 +49959,13 @@ _sk_load_a8_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,60 ; jne 3d41 <_sk_load_a8_sse2+0x65> + DB 117,60 ; jne 3d88 <_sk_load_a8_sse2+0x65> DB 102,15,110,4,62 ; movd (%esi,%edi,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,130,38,148,0,0 ; pand 0x9426(%edx),%xmm0 + DB 102,15,219,130,207,146,0,0 ; pand 0x92cf(%edx),%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,154,86,149,0,0 ; mulps 0x9556(%edx),%xmm3 + DB 15,89,154,255,147,0,0 ; mulps 0x93ff(%edx),%xmm3 DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -50028,12 +49982,12 @@ _sk_load_a8_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,50 ; je 3d7b <_sk_load_a8_sse2+0x9f> + DB 116,50 ; je 3dc2 <_sk_load_a8_sse2+0x9f> DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 3d65 <_sk_load_a8_sse2+0x89> + DB 116,19 ; je 3dac <_sk_load_a8_sse2+0x89> DB 128,251,3 ; cmp $0x3,%bl - DB 117,187 ; jne 3d12 <_sk_load_a8_sse2+0x36> + DB 117,187 ; jne 3d59 <_sk_load_a8_sse2+0x36> DB 15,182,92,62,2 ; movzbl 0x2(%esi,%edi,1),%ebx DB 102,15,110,195 ; movd %ebx,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 @@ -50042,10 +49996,10 @@ _sk_load_a8_sse2 LABEL PROC DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 235,151 ; jmp 3d12 <_sk_load_a8_sse2+0x36> + DB 235,151 ; jmp 3d59 <_sk_load_a8_sse2+0x36> DB 15,182,52,62 ; movzbl (%esi,%edi,1),%esi DB 102,15,110,198 ; movd %esi,%xmm0 - DB 235,141 ; jmp 3d12 <_sk_load_a8_sse2+0x36> + DB 235,141 ; jmp 3d59 <_sk_load_a8_sse2+0x36> PUBLIC _sk_load_a8_dst_sse2 _sk_load_a8_dst_sse2 LABEL PROC @@ -50068,15 +50022,15 @@ _sk_load_a8_dst_sse2 LABEL PROC DB 102,15,127,97,32 ; movdqa %xmm4,0x20(%ecx) DB 102,15,127,97,16 ; movdqa %xmm4,0x10(%ecx) DB 133,219 ; test %ebx,%ebx - DB 232,0,0,0,0 ; call 3dbe <_sk_load_a8_dst_sse2+0x39> + DB 232,0,0,0,0 ; call 3e05 <_sk_load_a8_dst_sse2+0x39> DB 90 ; pop %edx - DB 117,54 ; jne 3df7 <_sk_load_a8_dst_sse2+0x72> + DB 117,54 ; jne 3e3e <_sk_load_a8_dst_sse2+0x72> DB 102,15,110,36,55 ; movd (%edi,%esi,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,219,162,82,147,0,0 ; pand 0x9352(%edx),%xmm4 + DB 102,15,219,162,251,145,0,0 ; pand 0x91fb(%edx),%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,162,130,148,0,0 ; mulps 0x9482(%edx),%xmm4 + DB 15,89,162,43,147,0,0 ; mulps 0x932b(%edx),%xmm4 DB 15,41,97,64 ; movaps %xmm4,0x40(%ecx) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp @@ -50091,11 +50045,11 @@ _sk_load_a8_dst_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,46 ; je 3e2d <_sk_load_a8_dst_sse2+0xa8> + DB 116,46 ; je 3e74 <_sk_load_a8_dst_sse2+0xa8> DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 3e17 <_sk_load_a8_dst_sse2+0x92> + DB 116,19 ; je 3e5e <_sk_load_a8_dst_sse2+0x92> DB 128,251,3 ; cmp $0x3,%bl - DB 117,197 ; jne 3dce <_sk_load_a8_dst_sse2+0x49> + DB 117,197 ; jne 3e15 <_sk_load_a8_dst_sse2+0x49> DB 15,182,92,55,2 ; movzbl 0x2(%edi,%esi,1),%ebx DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -50104,10 +50058,10 @@ _sk_load_a8_dst_sse2 LABEL PROC DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 235,161 ; jmp 3dce <_sk_load_a8_dst_sse2+0x49> + DB 235,161 ; jmp 3e15 <_sk_load_a8_dst_sse2+0x49> DB 15,182,52,55 ; movzbl (%edi,%esi,1),%esi DB 102,15,110,230 ; movd %esi,%xmm4 - DB 235,151 ; jmp 3dce <_sk_load_a8_dst_sse2+0x49> + DB 235,151 ; jmp 3e15 <_sk_load_a8_dst_sse2+0x49> PUBLIC _sk_gather_a8_sse2 _sk_gather_a8_sse2 LABEL PROC @@ -50144,15 +50098,15 @@ _sk_gather_a8_sse2 LABEL PROC DB 15,182,20,50 ; movzbl (%edx,%esi,1),%edx DB 193,226,8 ; shl $0x8,%edx DB 9,202 ; or %ecx,%edx - DB 102,15,196,194,0 ; pinsrw $0x0,%edx,%xmm0 + DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,196,199,1 ; pinsrw $0x1,%edi,%xmm0 - DB 232,0,0,0,0 ; call 3ebc <_sk_gather_a8_sse2+0x85> + DB 232,0,0,0,0 ; call 3f02 <_sk_gather_a8_sse2+0x84> DB 89 ; pop %ecx DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,216 ; cvtdq2ps %xmm0,%xmm3 - DB 15,89,153,132,147,0,0 ; mulps 0x9384(%ecx),%xmm3 + DB 15,89,153,46,146,0,0 ; mulps 0x922e(%ecx),%xmm3 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp DB 15,87,192 ; xorps %xmm0,%xmm0 @@ -50174,8 +50128,8 @@ _sk_store_a8_sse2 LABEL PROC DB 83 ; push %ebx DB 87 ; push %edi DB 86 ; push %esi - DB 131,236,60 ; sub $0x3c,%esp - DB 232,0,0,0,0 ; call 3f00 <_sk_store_a8_sse2+0xe> + DB 131,236,12 ; sub $0xc,%esp + DB 232,0,0,0,0 ; call 3f46 <_sk_store_a8_sse2+0xe> DB 94 ; pop %esi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50184,16 +50138,16 @@ _sk_store_a8_sse2 LABEL PROC DB 15,175,81,4 ; imul 0x4(%ecx),%edx DB 3,23 ; add (%edi),%edx DB 139,89,8 ; mov 0x8(%ecx),%ebx - DB 15,40,166,32,146,0,0 ; movaps 0x9220(%esi),%xmm4 + DB 15,40,166,202,144,0,0 ; movaps 0x90ca(%esi),%xmm4 DB 15,89,227 ; mulps %xmm3,%xmm4 DB 102,15,91,228 ; cvtps2dq %xmm4,%xmm4 DB 102,15,114,244,16 ; pslld $0x10,%xmm4 DB 102,15,114,228,16 ; psrad $0x10,%xmm4 - DB 102,15,107,228 ; packssdw %xmm4,%xmm4 - DB 102,15,103,228 ; packuswb %xmm4,%xmm4 + DB 102,15,107,224 ; packssdw %xmm0,%xmm4 + DB 102,15,103,224 ; packuswb %xmm0,%xmm4 DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,26 ; jne 3f55 <_sk_store_a8_sse2+0x63> + DB 117,26 ; jne 3f9b <_sk_store_a8_sse2+0x63> DB 102,15,126,230 ; movd %xmm4,%esi DB 137,52,58 ; mov %esi,(%edx,%edi,1) DB 141,80,8 ; lea 0x8(%eax),%edx @@ -50201,7 +50155,7 @@ _sk_store_a8_sse2 LABEL PROC DB 82 ; push %edx DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) - DB 131,196,76 ; add $0x4c,%esp + DB 131,196,28 ; add $0x1c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx @@ -50211,26 +50165,24 @@ _sk_store_a8_sse2 LABEL PROC DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,57 ; je 3f9e <_sk_store_a8_sse2+0xac> + DB 116,54 ; je 3fe1 <_sk_store_a8_sse2+0xa9> DB 128,251,2 ; cmp $0x2,%bl - DB 116,17 ; je 3f7b <_sk_store_a8_sse2+0x89> + DB 116,14 ; je 3fbe <_sk_store_a8_sse2+0x86> DB 128,251,3 ; cmp $0x3,%bl - DB 117,211 ; jne 3f42 <_sk_store_a8_sse2+0x50> - DB 102,15,127,101,200 ; movdqa %xmm4,-0x38(%ebp) - DB 138,93,208 ; mov -0x30(%ebp),%bl + DB 117,211 ; jne 3f88 <_sk_store_a8_sse2+0x50> + DB 102,15,197,220,4 ; pextrw $0x4,%xmm4,%ebx DB 136,92,58,2 ; mov %bl,0x2(%edx,%edi,1) DB 102,15,112,228,212 ; pshufd $0xd4,%xmm4,%xmm4 - DB 102,15,219,166,32,149,0,0 ; pand 0x9520(%esi),%xmm4 + DB 102,15,219,166,218,147,0,0 ; pand 0x93da(%esi),%xmm4 DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,126,230 ; movd %xmm4,%esi DB 102,137,52,58 ; mov %si,(%edx,%edi,1) - DB 235,164 ; jmp 3f42 <_sk_store_a8_sse2+0x50> - DB 102,15,127,101,216 ; movdqa %xmm4,-0x28(%ebp) - DB 138,93,216 ; mov -0x28(%ebp),%bl + DB 235,167 ; jmp 3f88 <_sk_store_a8_sse2+0x50> + DB 102,15,126,227 ; movd %xmm4,%ebx DB 136,28,58 ; mov %bl,(%edx,%edi,1) - DB 235,151 ; jmp 3f42 <_sk_store_a8_sse2+0x50> + DB 235,158 ; jmp 3f88 <_sk_store_a8_sse2+0x50> PUBLIC _sk_load_g8_sse2 _sk_load_g8_sse2 LABEL PROC @@ -50240,7 +50192,7 @@ _sk_load_g8_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 3fb9 <_sk_load_g8_sse2+0xe> + DB 232,0,0,0,0 ; call 3ff8 <_sk_load_g8_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50251,16 +50203,16 @@ _sk_load_g8_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,63 ; jne 4013 <_sk_load_g8_sse2+0x68> + DB 117,63 ; jne 4052 <_sk_load_g8_sse2+0x68> DB 102,15,110,4,62 ; movd (%esi,%edi,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 - DB 102,15,219,130,87,145,0,0 ; pand 0x9157(%edx),%xmm0 + DB 102,15,219,130,8,144,0,0 ; pand 0x9008(%edx),%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,130,135,146,0,0 ; mulps 0x9287(%edx),%xmm0 + DB 15,89,130,56,145,0,0 ; mulps 0x9138(%edx),%xmm0 DB 141,112,8 ; lea 0x8(%eax),%esi DB 131,236,8 ; sub $0x8,%esp - DB 15,40,154,151,144,0,0 ; movaps 0x9097(%edx),%xmm3 + DB 15,40,154,72,143,0,0 ; movaps 0x8f48(%edx),%xmm3 DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 86 ; push %esi @@ -50274,12 +50226,12 @@ _sk_load_g8_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,50 ; je 404d <_sk_load_g8_sse2+0xa2> + DB 116,50 ; je 408c <_sk_load_g8_sse2+0xa2> DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 4037 <_sk_load_g8_sse2+0x8c> + DB 116,19 ; je 4076 <_sk_load_g8_sse2+0x8c> DB 128,251,3 ; cmp $0x3,%bl - DB 117,184 ; jne 3fe1 <_sk_load_g8_sse2+0x36> + DB 117,184 ; jne 4020 <_sk_load_g8_sse2+0x36> DB 15,182,92,62,2 ; movzbl 0x2(%esi,%edi,1),%ebx DB 102,15,110,195 ; movd %ebx,%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 @@ -50288,10 +50240,10 @@ _sk_load_g8_sse2 LABEL PROC DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 235,148 ; jmp 3fe1 <_sk_load_g8_sse2+0x36> + DB 235,148 ; jmp 4020 <_sk_load_g8_sse2+0x36> DB 15,182,52,62 ; movzbl (%esi,%edi,1),%esi DB 102,15,110,198 ; movd %esi,%xmm0 - DB 235,138 ; jmp 3fe1 <_sk_load_g8_sse2+0x36> + DB 235,138 ; jmp 4020 <_sk_load_g8_sse2+0x36> PUBLIC _sk_load_g8_dst_sse2 _sk_load_g8_dst_sse2 LABEL PROC @@ -50301,7 +50253,7 @@ _sk_load_g8_dst_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4065 <_sk_load_g8_dst_sse2+0xe> + DB 232,0,0,0,0 ; call 40a4 <_sk_load_g8_dst_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50312,17 +50264,17 @@ _sk_load_g8_dst_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,73 ; jne 40c9 <_sk_load_g8_dst_sse2+0x72> + DB 117,73 ; jne 4108 <_sk_load_g8_dst_sse2+0x72> DB 102,15,110,36,62 ; movd (%esi,%edi,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,219,162,171,144,0,0 ; pand 0x90ab(%edx),%xmm4 + DB 102,15,219,162,92,143,0,0 ; pand 0x8f5c(%edx),%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,162,219,145,0,0 ; mulps 0x91db(%edx),%xmm4 + DB 15,89,162,140,144,0,0 ; mulps 0x908c(%edx),%xmm4 DB 15,41,97,48 ; movaps %xmm4,0x30(%ecx) DB 15,41,97,32 ; movaps %xmm4,0x20(%ecx) DB 15,41,97,16 ; movaps %xmm4,0x10(%ecx) - DB 15,40,162,235,143,0,0 ; movaps 0x8feb(%edx),%xmm4 + DB 15,40,162,156,142,0,0 ; movaps 0x8e9c(%edx),%xmm4 DB 15,41,97,64 ; movaps %xmm4,0x40(%ecx) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp @@ -50337,12 +50289,12 @@ _sk_load_g8_dst_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,50 ; je 4103 <_sk_load_g8_dst_sse2+0xac> + DB 116,50 ; je 4142 <_sk_load_g8_dst_sse2+0xac> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 40ed <_sk_load_g8_dst_sse2+0x96> + DB 116,19 ; je 412c <_sk_load_g8_dst_sse2+0x96> DB 128,251,3 ; cmp $0x3,%bl - DB 117,174 ; jne 408d <_sk_load_g8_dst_sse2+0x36> + DB 117,174 ; jne 40cc <_sk_load_g8_dst_sse2+0x36> DB 15,182,92,62,2 ; movzbl 0x2(%esi,%edi,1),%ebx DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 @@ -50351,10 +50303,10 @@ _sk_load_g8_dst_sse2 LABEL PROC DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 235,138 ; jmp 408d <_sk_load_g8_dst_sse2+0x36> + DB 235,138 ; jmp 40cc <_sk_load_g8_dst_sse2+0x36> DB 15,182,52,62 ; movzbl (%esi,%edi,1),%esi DB 102,15,110,230 ; movd %esi,%xmm4 - DB 235,128 ; jmp 408d <_sk_load_g8_dst_sse2+0x36> + DB 235,128 ; jmp 40cc <_sk_load_g8_dst_sse2+0x36> PUBLIC _sk_gather_g8_sse2 _sk_gather_g8_sse2 LABEL PROC @@ -50391,18 +50343,18 @@ _sk_gather_g8_sse2 LABEL PROC DB 15,182,20,50 ; movzbl (%edx,%esi,1),%edx DB 193,226,8 ; shl $0x8,%edx DB 9,202 ; or %ecx,%edx - DB 102,15,196,194,0 ; pinsrw $0x0,%edx,%xmm0 - DB 232,0,0,0,0 ; call 418d <_sk_gather_g8_sse2+0x80> + DB 102,15,110,194 ; movd %edx,%xmm0 + DB 232,0,0,0,0 ; call 41cb <_sk_gather_g8_sse2+0x7f> DB 89 ; pop %ecx DB 102,15,196,199,1 ; pinsrw $0x1,%edi,%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,129,179,144,0,0 ; mulps 0x90b3(%ecx),%xmm0 + DB 15,89,129,101,143,0,0 ; mulps 0x8f65(%ecx),%xmm0 DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,153,195,142,0,0 ; movaps 0x8ec3(%ecx),%xmm3 + DB 15,40,153,117,141,0,0 ; movaps 0x8d75(%ecx),%xmm3 DB 15,40,200 ; movaps %xmm0,%xmm1 DB 15,40,208 ; movaps %xmm0,%xmm2 DB 82 ; push %edx @@ -50422,7 +50374,7 @@ _sk_load_565_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 41d8 <_sk_load_565_sse2+0xe> + DB 232,0,0,0,0 ; call 4216 <_sk_load_565_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50434,23 +50386,23 @@ _sk_load_565_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,97 ; jne 4256 <_sk_load_565_sse2+0x8c> + DB 117,97 ; jne 4294 <_sk_load_565_sse2+0x8c> DB 243,15,126,20,126 ; movq (%esi,%edi,2),%xmm2 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,130,120,144,0,0 ; movdqa 0x9078(%edx),%xmm0 + DB 102,15,111,130,42,143,0,0 ; movdqa 0x8f2a(%edx),%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,130,136,144,0,0 ; mulps 0x9088(%edx),%xmm0 - DB 102,15,111,138,152,144,0,0 ; movdqa 0x9098(%edx),%xmm1 + DB 15,89,130,58,143,0,0 ; mulps 0x8f3a(%edx),%xmm0 + DB 102,15,111,138,74,143,0,0 ; movdqa 0x8f4a(%edx),%xmm1 DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,138,168,144,0,0 ; mulps 0x90a8(%edx),%xmm1 - DB 102,15,219,146,184,144,0,0 ; pand 0x90b8(%edx),%xmm2 + DB 15,89,138,90,143,0,0 ; mulps 0x8f5a(%edx),%xmm1 + DB 102,15,219,146,106,143,0,0 ; pand 0x8f6a(%edx),%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,146,200,144,0,0 ; mulps 0x90c8(%edx),%xmm2 + DB 15,89,146,122,143,0,0 ; mulps 0x8f7a(%edx),%xmm2 DB 141,112,8 ; lea 0x8(%eax),%esi DB 131,236,8 ; sub $0x8,%esp - DB 15,40,154,120,142,0,0 ; movaps 0x8e78(%edx),%xmm3 + DB 15,40,154,42,141,0,0 ; movaps 0x8d2a(%edx),%xmm3 DB 86 ; push %esi DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) @@ -50462,22 +50414,22 @@ _sk_load_565_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,46 ; je 428c <_sk_load_565_sse2+0xc2> + DB 116,47 ; je 42cb <_sk_load_565_sse2+0xc3> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 427a <_sk_load_565_sse2+0xb0> + DB 116,19 ; je 42b8 <_sk_load_565_sse2+0xb0> DB 128,251,3 ; cmp $0x3,%bl - DB 117,146 ; jne 41fe <_sk_load_565_sse2+0x34> + DB 117,146 ; jne 423c <_sk_load_565_sse2+0x34> DB 15,183,92,126,4 ; movzwl 0x4(%esi,%edi,2),%ebx DB 102,15,110,195 ; movd %ebx,%xmm0 DB 102,15,112,208,69 ; pshufd $0x45,%xmm0,%xmm2 DB 102,15,110,4,126 ; movd (%esi,%edi,2),%xmm0 - DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 + DB 242,15,112,192,212 ; pshuflw $0xd4,%xmm0,%xmm0 DB 242,15,16,208 ; movsd %xmm0,%xmm2 - DB 233,114,255,255,255 ; jmp 41fe <_sk_load_565_sse2+0x34> + DB 233,113,255,255,255 ; jmp 423c <_sk_load_565_sse2+0x34> DB 15,183,52,126 ; movzwl (%esi,%edi,2),%esi DB 102,15,110,214 ; movd %esi,%xmm2 - DB 233,101,255,255,255 ; jmp 41fe <_sk_load_565_sse2+0x34> + DB 233,100,255,255,255 ; jmp 423c <_sk_load_565_sse2+0x34> PUBLIC _sk_load_565_dst_sse2 _sk_load_565_dst_sse2 LABEL PROC @@ -50487,7 +50439,7 @@ _sk_load_565_dst_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 42a7 <_sk_load_565_dst_sse2+0xe> + DB 232,0,0,0,0 ; call 42e6 <_sk_load_565_dst_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50499,24 +50451,24 @@ _sk_load_565_dst_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,113 ; jne 4335 <_sk_load_565_dst_sse2+0x9c> + DB 117,113 ; jne 4374 <_sk_load_565_dst_sse2+0x9c> DB 243,15,126,36,126 ; movq (%esi,%edi,2),%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,111,170,169,143,0,0 ; movdqa 0x8fa9(%edx),%xmm5 + DB 102,15,111,170,90,142,0,0 ; movdqa 0x8e5a(%edx),%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,170,185,143,0,0 ; mulps 0x8fb9(%edx),%xmm5 + DB 15,89,170,106,142,0,0 ; mulps 0x8e6a(%edx),%xmm5 DB 15,41,105,16 ; movaps %xmm5,0x10(%ecx) - DB 102,15,111,170,201,143,0,0 ; movdqa 0x8fc9(%edx),%xmm5 + DB 102,15,111,170,122,142,0,0 ; movdqa 0x8e7a(%edx),%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,170,217,143,0,0 ; mulps 0x8fd9(%edx),%xmm5 + DB 15,89,170,138,142,0,0 ; mulps 0x8e8a(%edx),%xmm5 DB 15,41,105,32 ; movaps %xmm5,0x20(%ecx) - DB 102,15,219,162,233,143,0,0 ; pand 0x8fe9(%edx),%xmm4 + DB 102,15,219,162,154,142,0,0 ; pand 0x8e9a(%edx),%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,162,249,143,0,0 ; mulps 0x8ff9(%edx),%xmm4 + DB 15,89,162,170,142,0,0 ; mulps 0x8eaa(%edx),%xmm4 DB 15,41,97,48 ; movaps %xmm4,0x30(%ecx) - DB 15,40,162,169,141,0,0 ; movaps 0x8da9(%edx),%xmm4 + DB 15,40,162,90,140,0,0 ; movaps 0x8c5a(%edx),%xmm4 DB 15,41,97,64 ; movaps %xmm4,0x40(%ecx) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp @@ -50531,22 +50483,22 @@ _sk_load_565_dst_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,46 ; je 436b <_sk_load_565_dst_sse2+0xd2> + DB 116,47 ; je 43ab <_sk_load_565_dst_sse2+0xd3> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 4359 <_sk_load_565_dst_sse2+0xc0> + DB 116,19 ; je 4398 <_sk_load_565_dst_sse2+0xc0> DB 128,251,3 ; cmp $0x3,%bl - DB 117,130 ; jne 42cd <_sk_load_565_dst_sse2+0x34> + DB 117,130 ; jne 430c <_sk_load_565_dst_sse2+0x34> DB 15,183,92,126,4 ; movzwl 0x4(%esi,%edi,2),%ebx DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 DB 102,15,110,44,126 ; movd (%esi,%edi,2),%xmm5 - DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 + DB 242,15,112,237,212 ; pshuflw $0xd4,%xmm5,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 233,98,255,255,255 ; jmp 42cd <_sk_load_565_dst_sse2+0x34> + DB 233,97,255,255,255 ; jmp 430c <_sk_load_565_dst_sse2+0x34> DB 15,183,52,126 ; movzwl (%esi,%edi,2),%esi DB 102,15,110,230 ; movd %esi,%xmm4 - DB 233,85,255,255,255 ; jmp 42cd <_sk_load_565_dst_sse2+0x34> + DB 233,84,255,255,255 ; jmp 430c <_sk_load_565_dst_sse2+0x34> PUBLIC _sk_gather_565_sse2 _sk_gather_565_sse2 LABEL PROC @@ -50569,35 +50521,37 @@ _sk_gather_565_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,126,193 ; movd %xmm0,%ecx DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,15,196,20,74,0 ; pinsrw $0x0,(%edx,%ecx,2),%xmm2 + DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx + DB 102,15,110,209 ; movd %ecx,%xmm2 DB 102,15,126,201 ; movd %xmm1,%ecx DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,15,196,20,74,1 ; pinsrw $0x1,(%edx,%ecx,2),%xmm2 + DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx + DB 102,15,196,209,1 ; pinsrw $0x1,%ecx,%xmm2 DB 102,15,126,201 ; movd %xmm1,%ecx DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx DB 102,15,196,209,2 ; pinsrw $0x2,%ecx,%xmm2 DB 102,15,126,193 ; movd %xmm0,%ecx DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx - DB 232,0,0,0,0 ; call 43f3 <_sk_gather_565_sse2+0x7b> + DB 232,0,0,0,0 ; call 4438 <_sk_gather_565_sse2+0x80> DB 90 ; pop %edx DB 102,15,196,209,3 ; pinsrw $0x3,%ecx,%xmm2 DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2 - DB 102,15,111,130,93,142,0,0 ; movdqa 0x8e5d(%edx),%xmm0 + DB 102,15,111,130,8,141,0,0 ; movdqa 0x8d08(%edx),%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,130,109,142,0,0 ; mulps 0x8e6d(%edx),%xmm0 - DB 102,15,111,138,125,142,0,0 ; movdqa 0x8e7d(%edx),%xmm1 + DB 15,89,130,24,141,0,0 ; mulps 0x8d18(%edx),%xmm0 + DB 102,15,111,138,40,141,0,0 ; movdqa 0x8d28(%edx),%xmm1 DB 102,15,219,202 ; pand %xmm2,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,138,141,142,0,0 ; mulps 0x8e8d(%edx),%xmm1 - DB 102,15,219,146,157,142,0,0 ; pand 0x8e9d(%edx),%xmm2 + DB 15,89,138,56,141,0,0 ; mulps 0x8d38(%edx),%xmm1 + DB 102,15,219,146,72,141,0,0 ; pand 0x8d48(%edx),%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,146,173,142,0,0 ; mulps 0x8ead(%edx),%xmm2 + DB 15,89,146,88,141,0,0 ; mulps 0x8d58(%edx),%xmm2 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,154,93,140,0,0 ; movaps 0x8c5d(%edx),%xmm3 + DB 15,40,154,8,139,0,0 ; movaps 0x8b08(%edx),%xmm3 DB 81 ; push %ecx DB 255,117,8 ; pushl 0x8(%ebp) DB 255,80,4 ; call *0x4(%eax) @@ -50613,7 +50567,7 @@ _sk_store_565_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4466 <_sk_store_565_sse2+0xe> + DB 232,0,0,0,0 ; call 44ab <_sk_store_565_sse2+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50623,8 +50577,8 @@ _sk_store_565_sse2 LABEL PROC DB 1,210 ; add %edx,%edx DB 3,22 ; add (%esi),%edx DB 139,49 ; mov (%ecx),%esi - DB 15,40,167,202,143,0,0 ; movaps 0x8fca(%edi),%xmm4 - DB 15,40,175,218,143,0,0 ; movaps 0x8fda(%edi),%xmm5 + DB 15,40,167,133,142,0,0 ; movaps 0x8e85(%edi),%xmm4 + DB 15,40,175,149,142,0,0 ; movaps 0x8e95(%edi),%xmm5 DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 15,40,240 ; movaps %xmm0,%xmm6 DB 15,89,244 ; mulps %xmm4,%xmm6 @@ -50639,12 +50593,12 @@ _sk_store_565_sse2 LABEL PROC DB 102,15,86,229 ; orpd %xmm5,%xmm4 DB 102,15,114,244,16 ; pslld $0x10,%xmm4 DB 102,15,114,228,16 ; psrad $0x10,%xmm4 - DB 102,15,107,228 ; packssdw %xmm4,%xmm4 + DB 102,15,107,224 ; packssdw %xmm0,%xmm4 DB 102,15,126,101,232 ; movd %xmm4,-0x18(%ebp) DB 102,15,112,236,229 ; pshufd $0xe5,%xmm4,%xmm5 DB 102,15,126,109,236 ; movd %xmm5,-0x14(%ebp) DB 133,219 ; test %ebx,%ebx - DB 117,34 ; jne 44fa <_sk_store_565_sse2+0xa2> + DB 117,34 ; jne 453f <_sk_store_565_sse2+0xa2> DB 102,15,126,231 ; movd %xmm4,%edi DB 102,15,126,235 ; movd %xmm5,%ebx DB 137,60,114 ; mov %edi,(%edx,%esi,2) @@ -50664,20 +50618,20 @@ _sk_store_565_sse2 LABEL PROC DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,37 ; je 4530 <_sk_store_565_sse2+0xd8> + DB 116,37 ; je 4575 <_sk_store_565_sse2+0xd8> DB 128,251,2 ; cmp $0x2,%bl - DB 116,15 ; je 451f <_sk_store_565_sse2+0xc7> + DB 116,15 ; je 4564 <_sk_store_565_sse2+0xc7> DB 128,251,3 ; cmp $0x3,%bl - DB 117,210 ; jne 44e7 <_sk_store_565_sse2+0x8f> + DB 117,210 ; jne 452c <_sk_store_565_sse2+0x8f> DB 102,15,197,252,4 ; pextrw $0x4,%xmm4,%edi DB 102,137,124,114,4 ; mov %di,0x4(%edx,%esi,2) DB 102,15,112,228,212 ; pshufd $0xd4,%xmm4,%xmm4 DB 242,15,112,228,232 ; pshuflw $0xe8,%xmm4,%xmm4 DB 102,15,126,36,114 ; movd %xmm4,(%edx,%esi,2) - DB 235,183 ; jmp 44e7 <_sk_store_565_sse2+0x8f> + DB 235,183 ; jmp 452c <_sk_store_565_sse2+0x8f> DB 102,15,126,231 ; movd %xmm4,%edi DB 102,137,60,114 ; mov %di,(%edx,%esi,2) - DB 235,173 ; jmp 44e7 <_sk_store_565_sse2+0x8f> + DB 235,173 ; jmp 452c <_sk_store_565_sse2+0x8f> PUBLIC _sk_load_4444_sse2 _sk_load_4444_sse2 LABEL PROC @@ -50687,7 +50641,7 @@ _sk_load_4444_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4548 <_sk_load_4444_sse2+0xe> + DB 232,0,0,0,0 ; call 458d <_sk_load_4444_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50699,24 +50653,24 @@ _sk_load_4444_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,112 ; jne 45d5 <_sk_load_4444_sse2+0x9b> + DB 117,112 ; jne 461a <_sk_load_4444_sse2+0x9b> DB 243,15,126,28,126 ; movq (%esi,%edi,2),%xmm3 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,130,8,143,0,0 ; movdqa 0x8f08(%edx),%xmm0 + DB 102,15,111,130,195,141,0,0 ; movdqa 0x8dc3(%edx),%xmm0 DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,130,24,143,0,0 ; mulps 0x8f18(%edx),%xmm0 - DB 102,15,111,138,40,143,0,0 ; movdqa 0x8f28(%edx),%xmm1 + DB 15,89,130,211,141,0,0 ; mulps 0x8dd3(%edx),%xmm0 + DB 102,15,111,138,227,141,0,0 ; movdqa 0x8de3(%edx),%xmm1 DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,138,56,143,0,0 ; mulps 0x8f38(%edx),%xmm1 - DB 102,15,111,146,72,143,0,0 ; movdqa 0x8f48(%edx),%xmm2 + DB 15,89,138,243,141,0,0 ; mulps 0x8df3(%edx),%xmm1 + DB 102,15,111,146,3,142,0,0 ; movdqa 0x8e03(%edx),%xmm2 DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,146,88,143,0,0 ; mulps 0x8f58(%edx),%xmm2 - DB 102,15,219,154,104,143,0,0 ; pand 0x8f68(%edx),%xmm3 + DB 15,89,146,19,142,0,0 ; mulps 0x8e13(%edx),%xmm2 + DB 102,15,219,154,35,142,0,0 ; pand 0x8e23(%edx),%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,154,120,143,0,0 ; mulps 0x8f78(%edx),%xmm3 + DB 15,89,154,51,142,0,0 ; mulps 0x8e33(%edx),%xmm3 DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp DB 82 ; push %edx @@ -50730,22 +50684,22 @@ _sk_load_4444_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,46 ; je 460b <_sk_load_4444_sse2+0xd1> + DB 116,47 ; je 4651 <_sk_load_4444_sse2+0xd2> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 128,251,2 ; cmp $0x2,%bl - DB 116,19 ; je 45f9 <_sk_load_4444_sse2+0xbf> + DB 116,19 ; je 463e <_sk_load_4444_sse2+0xbf> DB 128,251,3 ; cmp $0x3,%bl - DB 117,131 ; jne 456e <_sk_load_4444_sse2+0x34> + DB 117,131 ; jne 45b3 <_sk_load_4444_sse2+0x34> DB 15,183,92,126,4 ; movzwl 0x4(%esi,%edi,2),%ebx DB 102,15,110,195 ; movd %ebx,%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 102,15,110,4,126 ; movd (%esi,%edi,2),%xmm0 - DB 102,15,97,192 ; punpcklwd %xmm0,%xmm0 + DB 242,15,112,192,212 ; pshuflw $0xd4,%xmm0,%xmm0 DB 242,15,16,216 ; movsd %xmm0,%xmm3 - DB 233,99,255,255,255 ; jmp 456e <_sk_load_4444_sse2+0x34> + DB 233,98,255,255,255 ; jmp 45b3 <_sk_load_4444_sse2+0x34> DB 15,183,52,126 ; movzwl (%esi,%edi,2),%esi DB 102,15,110,222 ; movd %esi,%xmm3 - DB 233,86,255,255,255 ; jmp 456e <_sk_load_4444_sse2+0x34> + DB 233,85,255,255,255 ; jmp 45b3 <_sk_load_4444_sse2+0x34> PUBLIC _sk_load_4444_dst_sse2 _sk_load_4444_dst_sse2 LABEL PROC @@ -50755,7 +50709,7 @@ _sk_load_4444_dst_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4626 <_sk_load_4444_dst_sse2+0xe> + DB 232,0,0,0,0 ; call 466c <_sk_load_4444_dst_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50767,27 +50721,27 @@ _sk_load_4444_dst_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 15,133,128,0,0,0 ; jne 46c7 <_sk_load_4444_dst_sse2+0xaf> + DB 15,133,128,0,0,0 ; jne 470d <_sk_load_4444_dst_sse2+0xaf> DB 243,15,126,36,126 ; movq (%esi,%edi,2),%xmm4 DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 - DB 102,15,111,170,42,142,0,0 ; movdqa 0x8e2a(%edx),%xmm5 + DB 102,15,111,170,228,140,0,0 ; movdqa 0x8ce4(%edx),%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,170,58,142,0,0 ; mulps 0x8e3a(%edx),%xmm5 + DB 15,89,170,244,140,0,0 ; mulps 0x8cf4(%edx),%xmm5 DB 15,41,105,16 ; movaps %xmm5,0x10(%ecx) - DB 102,15,111,170,74,142,0,0 ; movdqa 0x8e4a(%edx),%xmm5 + DB 102,15,111,170,4,141,0,0 ; movdqa 0x8d04(%edx),%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,170,90,142,0,0 ; mulps 0x8e5a(%edx),%xmm5 + DB 15,89,170,20,141,0,0 ; mulps 0x8d14(%edx),%xmm5 DB 15,41,105,32 ; movaps %xmm5,0x20(%ecx) - DB 102,15,111,170,106,142,0,0 ; movdqa 0x8e6a(%edx),%xmm5 + DB 102,15,111,170,36,141,0,0 ; movdqa 0x8d24(%edx),%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 15,91,237 ; cvtdq2ps %xmm5,%xmm5 - DB 15,89,170,122,142,0,0 ; mulps 0x8e7a(%edx),%xmm5 + DB 15,89,170,52,141,0,0 ; mulps 0x8d34(%edx),%xmm5 DB 15,41,105,48 ; movaps %xmm5,0x30(%ecx) - DB 102,15,219,162,138,142,0,0 ; pand 0x8e8a(%edx),%xmm4 + DB 102,15,219,162,68,141,0,0 ; pand 0x8d44(%edx),%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 - DB 15,89,162,154,142,0,0 ; mulps 0x8e9a(%edx),%xmm4 + DB 15,89,162,84,141,0,0 ; mulps 0x8d54(%edx),%xmm4 DB 15,41,97,64 ; movaps %xmm4,0x40(%ecx) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp @@ -50802,22 +50756,22 @@ _sk_load_4444_dst_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,50 ; je 4701 <_sk_load_4444_dst_sse2+0xe9> + DB 116,51 ; je 4748 <_sk_load_4444_dst_sse2+0xea> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 128,251,2 ; cmp $0x2,%bl - DB 116,23 ; je 46ef <_sk_load_4444_dst_sse2+0xd7> + DB 116,23 ; je 4735 <_sk_load_4444_dst_sse2+0xd7> DB 128,251,3 ; cmp $0x3,%bl - DB 15,133,111,255,255,255 ; jne 4650 <_sk_load_4444_dst_sse2+0x38> + DB 15,133,111,255,255,255 ; jne 4696 <_sk_load_4444_dst_sse2+0x38> DB 15,183,92,126,4 ; movzwl 0x4(%esi,%edi,2),%ebx DB 102,15,110,227 ; movd %ebx,%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 DB 102,15,110,44,126 ; movd (%esi,%edi,2),%xmm5 - DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 + DB 242,15,112,237,212 ; pshuflw $0xd4,%xmm5,%xmm5 DB 242,15,16,229 ; movsd %xmm5,%xmm4 - DB 233,79,255,255,255 ; jmp 4650 <_sk_load_4444_dst_sse2+0x38> + DB 233,78,255,255,255 ; jmp 4696 <_sk_load_4444_dst_sse2+0x38> DB 15,183,52,126 ; movzwl (%esi,%edi,2),%esi DB 102,15,110,230 ; movd %esi,%xmm4 - DB 233,66,255,255,255 ; jmp 4650 <_sk_load_4444_dst_sse2+0x38> + DB 233,65,255,255,255 ; jmp 4696 <_sk_load_4444_dst_sse2+0x38> PUBLIC _sk_gather_4444_sse2 _sk_gather_4444_sse2 LABEL PROC @@ -50840,10 +50794,12 @@ _sk_gather_4444_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,126,193 ; movd %xmm0,%ecx DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,15,196,28,74,0 ; pinsrw $0x0,(%edx,%ecx,2),%xmm3 + DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx + DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,15,126,201 ; movd %xmm1,%ecx DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 - DB 102,15,196,28,74,1 ; pinsrw $0x1,(%edx,%ecx,2),%xmm3 + DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx + DB 102,15,196,217,1 ; pinsrw $0x1,%ecx,%xmm3 DB 102,15,126,201 ; movd %xmm1,%ecx DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx @@ -50851,25 +50807,25 @@ _sk_gather_4444_sse2 LABEL PROC DB 102,15,126,193 ; movd %xmm0,%ecx DB 15,183,12,74 ; movzwl (%edx,%ecx,2),%ecx DB 102,15,196,217,3 ; pinsrw $0x3,%ecx,%xmm3 - DB 232,0,0,0,0 ; call 478e <_sk_gather_4444_sse2+0x80> + DB 232,0,0,0,0 ; call 47da <_sk_gather_4444_sse2+0x85> DB 89 ; pop %ecx DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3 - DB 102,15,111,129,194,140,0,0 ; movdqa 0x8cc2(%ecx),%xmm0 + DB 102,15,111,129,118,139,0,0 ; movdqa 0x8b76(%ecx),%xmm0 DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,89,129,210,140,0,0 ; mulps 0x8cd2(%ecx),%xmm0 - DB 102,15,111,137,226,140,0,0 ; movdqa 0x8ce2(%ecx),%xmm1 + DB 15,89,129,134,139,0,0 ; mulps 0x8b86(%ecx),%xmm0 + DB 102,15,111,137,150,139,0,0 ; movdqa 0x8b96(%ecx),%xmm1 DB 102,15,219,203 ; pand %xmm3,%xmm1 DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 - DB 15,89,137,242,140,0,0 ; mulps 0x8cf2(%ecx),%xmm1 - DB 102,15,111,145,2,141,0,0 ; movdqa 0x8d02(%ecx),%xmm2 + DB 15,89,137,166,139,0,0 ; mulps 0x8ba6(%ecx),%xmm1 + DB 102,15,111,145,182,139,0,0 ; movdqa 0x8bb6(%ecx),%xmm2 DB 102,15,219,211 ; pand %xmm3,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 - DB 15,89,145,18,141,0,0 ; mulps 0x8d12(%ecx),%xmm2 - DB 102,15,219,153,34,141,0,0 ; pand 0x8d22(%ecx),%xmm3 + DB 15,89,145,198,139,0,0 ; mulps 0x8bc6(%ecx),%xmm2 + DB 102,15,219,153,214,139,0,0 ; pand 0x8bd6(%ecx),%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 - DB 15,89,153,50,141,0,0 ; mulps 0x8d32(%ecx),%xmm3 + DB 15,89,153,230,139,0,0 ; mulps 0x8be6(%ecx),%xmm3 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp DB 81 ; push %ecx @@ -50887,7 +50843,7 @@ _sk_store_4444_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 480b <_sk_store_4444_sse2+0xe> + DB 232,0,0,0,0 ; call 4857 <_sk_store_4444_sse2+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50897,7 +50853,7 @@ _sk_store_4444_sse2 LABEL PROC DB 1,210 ; add %edx,%edx DB 3,22 ; add (%esi),%edx DB 139,49 ; mov (%ecx),%esi - DB 15,40,167,197,140,0,0 ; movaps 0x8cc5(%edi),%xmm4 + DB 15,40,167,121,139,0,0 ; movaps 0x8b79(%edi),%xmm4 DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 15,40,232 ; movaps %xmm0,%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 @@ -50918,12 +50874,12 @@ _sk_store_4444_sse2 LABEL PROC DB 102,15,86,230 ; orpd %xmm6,%xmm4 DB 102,15,114,244,16 ; pslld $0x10,%xmm4 DB 102,15,114,228,16 ; psrad $0x10,%xmm4 - DB 102,15,107,228 ; packssdw %xmm4,%xmm4 + DB 102,15,107,224 ; packssdw %xmm0,%xmm4 DB 102,15,126,101,232 ; movd %xmm4,-0x18(%ebp) DB 102,15,112,236,229 ; pshufd $0xe5,%xmm4,%xmm5 DB 102,15,126,109,236 ; movd %xmm5,-0x14(%ebp) DB 133,219 ; test %ebx,%ebx - DB 117,34 ; jne 48ae <_sk_store_4444_sse2+0xb1> + DB 117,34 ; jne 48fa <_sk_store_4444_sse2+0xb1> DB 102,15,126,231 ; movd %xmm4,%edi DB 102,15,126,235 ; movd %xmm5,%ebx DB 137,60,114 ; mov %edi,(%edx,%esi,2) @@ -50943,20 +50899,20 @@ _sk_store_4444_sse2 LABEL PROC DB 102,15,97,224 ; punpcklwd %xmm0,%xmm4 DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,37 ; je 48e4 <_sk_store_4444_sse2+0xe7> + DB 116,37 ; je 4930 <_sk_store_4444_sse2+0xe7> DB 128,251,2 ; cmp $0x2,%bl - DB 116,15 ; je 48d3 <_sk_store_4444_sse2+0xd6> + DB 116,15 ; je 491f <_sk_store_4444_sse2+0xd6> DB 128,251,3 ; cmp $0x3,%bl - DB 117,210 ; jne 489b <_sk_store_4444_sse2+0x9e> + DB 117,210 ; jne 48e7 <_sk_store_4444_sse2+0x9e> DB 102,15,197,252,4 ; pextrw $0x4,%xmm4,%edi DB 102,137,124,114,4 ; mov %di,0x4(%edx,%esi,2) DB 102,15,112,228,212 ; pshufd $0xd4,%xmm4,%xmm4 DB 242,15,112,228,232 ; pshuflw $0xe8,%xmm4,%xmm4 DB 102,15,126,36,114 ; movd %xmm4,(%edx,%esi,2) - DB 235,183 ; jmp 489b <_sk_store_4444_sse2+0x9e> + DB 235,183 ; jmp 48e7 <_sk_store_4444_sse2+0x9e> DB 102,15,126,231 ; movd %xmm4,%edi DB 102,137,60,114 ; mov %di,(%edx,%esi,2) - DB 235,173 ; jmp 489b <_sk_store_4444_sse2+0x9e> + DB 235,173 ; jmp 48e7 <_sk_store_4444_sse2+0x9e> PUBLIC _sk_load_8888_sse2 _sk_load_8888_sse2 LABEL PROC @@ -50966,7 +50922,7 @@ _sk_load_8888_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 48fc <_sk_load_8888_sse2+0xe> + DB 232,0,0,0,0 ; call 4948 <_sk_load_8888_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -50978,13 +50934,13 @@ _sk_load_8888_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,102 ; jne 4980 <_sk_load_8888_sse2+0x92> + DB 117,102 ; jne 49cc <_sk_load_8888_sse2+0x92> DB 243,15,111,28,190 ; movdqu (%esi,%edi,4),%xmm3 - DB 102,15,111,146,20,136,0,0 ; movdqa 0x8814(%edx),%xmm2 + DB 102,15,111,146,184,134,0,0 ; movdqa 0x86b8(%edx),%xmm2 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,40,162,68,137,0,0 ; movaps 0x8944(%edx),%xmm4 + DB 15,40,162,232,135,0,0 ; movaps 0x87e8(%edx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -51012,18 +50968,18 @@ _sk_load_8888_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,35 ; je 49ab <_sk_load_8888_sse2+0xbd> + DB 116,35 ; je 49f7 <_sk_load_8888_sse2+0xbd> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 128,251,2 ; cmp $0x2,%bl - DB 116,16 ; je 49a1 <_sk_load_8888_sse2+0xb3> + DB 116,16 ; je 49ed <_sk_load_8888_sse2+0xb3> DB 128,251,3 ; cmp $0x3,%bl - DB 117,137 ; jne 491f <_sk_load_8888_sse2+0x31> + DB 117,137 ; jne 496b <_sk_load_8888_sse2+0x31> DB 102,15,110,68,190,8 ; movd 0x8(%esi,%edi,4),%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 102,15,18,28,190 ; movlpd (%esi,%edi,4),%xmm3 - DB 233,116,255,255,255 ; jmp 491f <_sk_load_8888_sse2+0x31> + DB 233,116,255,255,255 ; jmp 496b <_sk_load_8888_sse2+0x31> DB 102,15,110,28,190 ; movd (%esi,%edi,4),%xmm3 - DB 233,106,255,255,255 ; jmp 491f <_sk_load_8888_sse2+0x31> + DB 233,106,255,255,255 ; jmp 496b <_sk_load_8888_sse2+0x31> PUBLIC _sk_load_8888_dst_sse2 _sk_load_8888_dst_sse2 LABEL PROC @@ -51033,7 +50989,7 @@ _sk_load_8888_dst_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 49c3 <_sk_load_8888_dst_sse2+0xe> + DB 232,0,0,0,0 ; call 4a0f <_sk_load_8888_dst_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51045,13 +51001,13 @@ _sk_load_8888_dst_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,118 ; jne 4a57 <_sk_load_8888_dst_sse2+0xa2> + DB 117,118 ; jne 4aa3 <_sk_load_8888_dst_sse2+0xa2> DB 243,15,111,36,190 ; movdqu (%esi,%edi,4),%xmm4 - DB 102,15,111,170,77,135,0,0 ; movdqa 0x874d(%edx),%xmm5 + DB 102,15,111,170,241,133,0,0 ; movdqa 0x85f1(%edx),%xmm5 DB 102,15,111,244 ; movdqa %xmm4,%xmm6 DB 102,15,219,245 ; pand %xmm5,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,40,186,125,136,0,0 ; movaps 0x887d(%edx),%xmm7 + DB 15,40,186,33,135,0,0 ; movaps 0x8721(%edx),%xmm7 DB 15,89,247 ; mulps %xmm7,%xmm6 DB 15,41,113,16 ; movaps %xmm6,0x10(%ecx) DB 102,15,111,244 ; movdqa %xmm4,%xmm6 @@ -51083,18 +51039,18 @@ _sk_load_8888_dst_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,39 ; je 4a86 <_sk_load_8888_dst_sse2+0xd1> + DB 116,39 ; je 4ad2 <_sk_load_8888_dst_sse2+0xd1> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 128,251,2 ; cmp $0x2,%bl - DB 116,20 ; je 4a7c <_sk_load_8888_dst_sse2+0xc7> + DB 116,20 ; je 4ac8 <_sk_load_8888_dst_sse2+0xc7> DB 128,251,3 ; cmp $0x3,%bl - DB 15,133,117,255,255,255 ; jne 49e6 <_sk_load_8888_dst_sse2+0x31> + DB 15,133,117,255,255,255 ; jne 4a32 <_sk_load_8888_dst_sse2+0x31> DB 102,15,110,100,190,8 ; movd 0x8(%esi,%edi,4),%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 DB 102,15,18,36,190 ; movlpd (%esi,%edi,4),%xmm4 - DB 233,96,255,255,255 ; jmp 49e6 <_sk_load_8888_dst_sse2+0x31> + DB 233,96,255,255,255 ; jmp 4a32 <_sk_load_8888_dst_sse2+0x31> DB 102,15,110,36,190 ; movd (%esi,%edi,4),%xmm4 - DB 233,86,255,255,255 ; jmp 49e6 <_sk_load_8888_dst_sse2+0x31> + DB 233,86,255,255,255 ; jmp 4a32 <_sk_load_8888_dst_sse2+0x31> PUBLIC _sk_gather_8888_sse2 _sk_gather_8888_sse2 LABEL PROC @@ -51115,27 +51071,27 @@ _sk_gather_8888_sse2 LABEL PROC DB 102,15,98,203 ; punpckldq %xmm3,%xmm1 DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 DB 102,15,126,209 ; movd %xmm2,%ecx DB 102,15,110,20,138 ; movd (%edx,%ecx,4),%xmm2 DB 102,15,126,201 ; movd %xmm1,%ecx DB 102,15,110,12,138 ; movd (%edx,%ecx,4),%xmm1 DB 102,15,126,193 ; movd %xmm0,%ecx - DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,112,192,229 ; pshufd $0xe5,%xmm0,%xmm0 DB 102,15,110,28,138 ; movd (%edx,%ecx,4),%xmm3 DB 102,15,126,193 ; movd %xmm0,%ecx DB 102,15,110,4,138 ; movd (%edx,%ecx,4),%xmm0 - DB 232,0,0,0,0 ; call 4b06 <_sk_gather_8888_sse2+0x76> + DB 232,0,0,0,0 ; call 4b52 <_sk_gather_8888_sse2+0x76> DB 89 ; pop %ecx DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 DB 102,15,98,216 ; punpckldq %xmm0,%xmm3 - DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 - DB 102,15,111,145,10,134,0,0 ; movdqa 0x860a(%ecx),%xmm2 + DB 102,15,108,217 ; punpcklqdq %xmm1,%xmm3 + DB 102,15,111,145,174,132,0,0 ; movdqa 0x84ae(%ecx),%xmm2 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,219,194 ; pand %xmm2,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,40,161,58,135,0,0 ; movaps 0x873a(%ecx),%xmm4 + DB 15,40,161,222,133,0,0 ; movaps 0x85de(%ecx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -51167,7 +51123,7 @@ _sk_store_8888_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4b81 <_sk_store_8888_sse2+0xe> + DB 232,0,0,0,0 ; call 4bcd <_sk_store_8888_sse2+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51177,7 +51133,7 @@ _sk_store_8888_sse2 LABEL PROC DB 193,226,2 ; shl $0x2,%edx DB 3,22 ; add (%esi),%edx DB 139,49 ; mov (%ecx),%esi - DB 15,40,167,159,133,0,0 ; movaps 0x859f(%edi),%xmm4 + DB 15,40,167,67,132,0,0 ; movaps 0x8443(%edi),%xmm4 DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 15,40,232 ; movaps %xmm0,%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 @@ -51197,7 +51153,7 @@ _sk_store_8888_sse2 LABEL PROC DB 102,15,235,229 ; por %xmm5,%xmm4 DB 102,15,235,230 ; por %xmm6,%xmm4 DB 133,219 ; test %ebx,%ebx - DB 117,24 ; jne 4bfe <_sk_store_8888_sse2+0x8b> + DB 117,24 ; jne 4c4a <_sk_store_8888_sse2+0x8b> DB 243,15,127,36,178 ; movdqu %xmm4,(%edx,%esi,4) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp @@ -51212,17 +51168,17 @@ _sk_store_8888_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,28 ; je 4c22 <_sk_store_8888_sse2+0xaf> + DB 116,28 ; je 4c6e <_sk_store_8888_sse2+0xaf> DB 128,251,2 ; cmp $0x2,%bl - DB 116,16 ; je 4c1b <_sk_store_8888_sse2+0xa8> + DB 116,16 ; je 4c67 <_sk_store_8888_sse2+0xa8> DB 128,251,3 ; cmp $0x3,%bl - DB 117,219 ; jne 4beb <_sk_store_8888_sse2+0x78> + DB 117,219 ; jne 4c37 <_sk_store_8888_sse2+0x78> DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 DB 102,15,126,108,178,8 ; movd %xmm5,0x8(%edx,%esi,4) DB 102,15,214,36,178 ; movq %xmm4,(%edx,%esi,4) - DB 235,201 ; jmp 4beb <_sk_store_8888_sse2+0x78> + DB 235,201 ; jmp 4c37 <_sk_store_8888_sse2+0x78> DB 102,15,126,36,178 ; movd %xmm4,(%edx,%esi,4) - DB 235,194 ; jmp 4beb <_sk_store_8888_sse2+0x78> + DB 235,194 ; jmp 4c37 <_sk_store_8888_sse2+0x78> PUBLIC _sk_load_bgra_sse2 _sk_load_bgra_sse2 LABEL PROC @@ -51232,7 +51188,7 @@ _sk_load_bgra_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4c37 <_sk_load_bgra_sse2+0xe> + DB 232,0,0,0,0 ; call 4c83 <_sk_load_bgra_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51244,13 +51200,13 @@ _sk_load_bgra_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,102 ; jne 4cbb <_sk_load_bgra_sse2+0x92> + DB 117,102 ; jne 4d07 <_sk_load_bgra_sse2+0x92> DB 243,15,111,28,190 ; movdqu (%esi,%edi,4),%xmm3 - DB 102,15,111,130,217,132,0,0 ; movdqa 0x84d9(%edx),%xmm0 + DB 102,15,111,130,125,131,0,0 ; movdqa 0x837d(%edx),%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,219,200 ; pand %xmm0,%xmm1 DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2 - DB 15,40,162,9,134,0,0 ; movaps 0x8609(%edx),%xmm4 + DB 15,40,162,173,132,0,0 ; movaps 0x84ad(%edx),%xmm4 DB 15,89,212 ; mulps %xmm4,%xmm2 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -51278,18 +51234,18 @@ _sk_load_bgra_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,35 ; je 4ce6 <_sk_load_bgra_sse2+0xbd> + DB 116,35 ; je 4d32 <_sk_load_bgra_sse2+0xbd> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 128,251,2 ; cmp $0x2,%bl - DB 116,16 ; je 4cdc <_sk_load_bgra_sse2+0xb3> + DB 116,16 ; je 4d28 <_sk_load_bgra_sse2+0xb3> DB 128,251,3 ; cmp $0x3,%bl - DB 117,137 ; jne 4c5a <_sk_load_bgra_sse2+0x31> + DB 117,137 ; jne 4ca6 <_sk_load_bgra_sse2+0x31> DB 102,15,110,68,190,8 ; movd 0x8(%esi,%edi,4),%xmm0 DB 102,15,112,216,69 ; pshufd $0x45,%xmm0,%xmm3 DB 102,15,18,28,190 ; movlpd (%esi,%edi,4),%xmm3 - DB 233,116,255,255,255 ; jmp 4c5a <_sk_load_bgra_sse2+0x31> + DB 233,116,255,255,255 ; jmp 4ca6 <_sk_load_bgra_sse2+0x31> DB 102,15,110,28,190 ; movd (%esi,%edi,4),%xmm3 - DB 233,106,255,255,255 ; jmp 4c5a <_sk_load_bgra_sse2+0x31> + DB 233,106,255,255,255 ; jmp 4ca6 <_sk_load_bgra_sse2+0x31> PUBLIC _sk_load_bgra_dst_sse2 _sk_load_bgra_dst_sse2 LABEL PROC @@ -51299,7 +51255,7 @@ _sk_load_bgra_dst_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4cfe <_sk_load_bgra_dst_sse2+0xe> + DB 232,0,0,0,0 ; call 4d4a <_sk_load_bgra_dst_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51311,13 +51267,13 @@ _sk_load_bgra_dst_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 117,118 ; jne 4d92 <_sk_load_bgra_dst_sse2+0xa2> + DB 117,118 ; jne 4dde <_sk_load_bgra_dst_sse2+0xa2> DB 243,15,111,36,190 ; movdqu (%esi,%edi,4),%xmm4 - DB 102,15,111,170,18,132,0,0 ; movdqa 0x8412(%edx),%xmm5 + DB 102,15,111,170,182,130,0,0 ; movdqa 0x82b6(%edx),%xmm5 DB 102,15,111,244 ; movdqa %xmm4,%xmm6 DB 102,15,219,245 ; pand %xmm5,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 - DB 15,40,186,66,133,0,0 ; movaps 0x8542(%edx),%xmm7 + DB 15,40,186,230,131,0,0 ; movaps 0x83e6(%edx),%xmm7 DB 15,89,247 ; mulps %xmm7,%xmm6 DB 15,41,113,48 ; movaps %xmm6,0x30(%ecx) DB 102,15,111,244 ; movdqa %xmm4,%xmm6 @@ -51349,18 +51305,18 @@ _sk_load_bgra_dst_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,39 ; je 4dc1 <_sk_load_bgra_dst_sse2+0xd1> + DB 116,39 ; je 4e0d <_sk_load_bgra_dst_sse2+0xd1> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 128,251,2 ; cmp $0x2,%bl - DB 116,20 ; je 4db7 <_sk_load_bgra_dst_sse2+0xc7> + DB 116,20 ; je 4e03 <_sk_load_bgra_dst_sse2+0xc7> DB 128,251,3 ; cmp $0x3,%bl - DB 15,133,117,255,255,255 ; jne 4d21 <_sk_load_bgra_dst_sse2+0x31> + DB 15,133,117,255,255,255 ; jne 4d6d <_sk_load_bgra_dst_sse2+0x31> DB 102,15,110,100,190,8 ; movd 0x8(%esi,%edi,4),%xmm4 DB 102,15,112,228,69 ; pshufd $0x45,%xmm4,%xmm4 DB 102,15,18,36,190 ; movlpd (%esi,%edi,4),%xmm4 - DB 233,96,255,255,255 ; jmp 4d21 <_sk_load_bgra_dst_sse2+0x31> + DB 233,96,255,255,255 ; jmp 4d6d <_sk_load_bgra_dst_sse2+0x31> DB 102,15,110,36,190 ; movd (%esi,%edi,4),%xmm4 - DB 233,86,255,255,255 ; jmp 4d21 <_sk_load_bgra_dst_sse2+0x31> + DB 233,86,255,255,255 ; jmp 4d6d <_sk_load_bgra_dst_sse2+0x31> PUBLIC _sk_gather_bgra_sse2 _sk_gather_bgra_sse2 LABEL PROC @@ -51381,27 +51337,27 @@ _sk_gather_bgra_sse2 LABEL PROC DB 102,15,98,203 ; punpckldq %xmm3,%xmm1 DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 DB 102,15,126,209 ; movd %xmm2,%ecx DB 102,15,110,20,138 ; movd (%edx,%ecx,4),%xmm2 DB 102,15,126,201 ; movd %xmm1,%ecx DB 102,15,110,12,138 ; movd (%edx,%ecx,4),%xmm1 DB 102,15,126,193 ; movd %xmm0,%ecx - DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,15,112,192,229 ; pshufd $0xe5,%xmm0,%xmm0 DB 102,15,110,28,138 ; movd (%edx,%ecx,4),%xmm3 DB 102,15,126,193 ; movd %xmm0,%ecx DB 102,15,110,4,138 ; movd (%edx,%ecx,4),%xmm0 - DB 232,0,0,0,0 ; call 4e41 <_sk_gather_bgra_sse2+0x76> + DB 232,0,0,0,0 ; call 4e8d <_sk_gather_bgra_sse2+0x76> DB 89 ; pop %ecx DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 DB 102,15,98,216 ; punpckldq %xmm0,%xmm3 - DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 - DB 102,15,111,129,207,130,0,0 ; movdqa 0x82cf(%ecx),%xmm0 + DB 102,15,108,217 ; punpcklqdq %xmm1,%xmm3 + DB 102,15,111,129,115,129,0,0 ; movdqa 0x8173(%ecx),%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,219,200 ; pand %xmm0,%xmm1 DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2 - DB 15,40,161,255,131,0,0 ; movaps 0x83ff(%ecx),%xmm4 + DB 15,40,161,163,130,0,0 ; movaps 0x82a3(%ecx),%xmm4 DB 15,89,212 ; mulps %xmm4,%xmm2 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,114,209,8 ; psrld $0x8,%xmm1 @@ -51433,7 +51389,7 @@ _sk_store_bgra_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 4ebc <_sk_store_bgra_sse2+0xe> + DB 232,0,0,0,0 ; call 4f08 <_sk_store_bgra_sse2+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51443,7 +51399,7 @@ _sk_store_bgra_sse2 LABEL PROC DB 193,226,2 ; shl $0x2,%edx DB 3,22 ; add (%esi),%edx DB 139,49 ; mov (%ecx),%esi - DB 15,40,167,100,130,0,0 ; movaps 0x8264(%edi),%xmm4 + DB 15,40,167,8,129,0,0 ; movaps 0x8108(%edi),%xmm4 DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 15,40,234 ; movaps %xmm2,%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 @@ -51463,7 +51419,7 @@ _sk_store_bgra_sse2 LABEL PROC DB 102,15,235,229 ; por %xmm5,%xmm4 DB 102,15,235,230 ; por %xmm6,%xmm4 DB 133,219 ; test %ebx,%ebx - DB 117,24 ; jne 4f39 <_sk_store_bgra_sse2+0x8b> + DB 117,24 ; jne 4f85 <_sk_store_bgra_sse2+0x8b> DB 243,15,127,36,178 ; movdqu %xmm4,(%edx,%esi,4) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp @@ -51478,17 +51434,17 @@ _sk_store_bgra_sse2 LABEL PROC DB 195 ; ret DB 128,227,3 ; and $0x3,%bl DB 128,251,1 ; cmp $0x1,%bl - DB 116,28 ; je 4f5d <_sk_store_bgra_sse2+0xaf> + DB 116,28 ; je 4fa9 <_sk_store_bgra_sse2+0xaf> DB 128,251,2 ; cmp $0x2,%bl - DB 116,16 ; je 4f56 <_sk_store_bgra_sse2+0xa8> + DB 116,16 ; je 4fa2 <_sk_store_bgra_sse2+0xa8> DB 128,251,3 ; cmp $0x3,%bl - DB 117,219 ; jne 4f26 <_sk_store_bgra_sse2+0x78> + DB 117,219 ; jne 4f72 <_sk_store_bgra_sse2+0x78> DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 DB 102,15,126,108,178,8 ; movd %xmm5,0x8(%edx,%esi,4) DB 102,15,214,36,178 ; movq %xmm4,(%edx,%esi,4) - DB 235,201 ; jmp 4f26 <_sk_store_bgra_sse2+0x78> + DB 235,201 ; jmp 4f72 <_sk_store_bgra_sse2+0x78> DB 102,15,126,36,178 ; movd %xmm4,(%edx,%esi,4) - DB 235,194 ; jmp 4f26 <_sk_store_bgra_sse2+0x78> + DB 235,194 ; jmp 4f72 <_sk_store_bgra_sse2+0x78> PUBLIC _sk_load_f16_sse2 _sk_load_f16_sse2 LABEL PROC @@ -51497,8 +51453,8 @@ _sk_load_f16_sse2 LABEL PROC DB 83 ; push %ebx DB 87 ; push %edi DB 86 ; push %esi - DB 131,236,92 ; sub $0x5c,%esp - DB 232,0,0,0,0 ; call 4f72 <_sk_load_f16_sse2+0xe> + DB 131,236,108 ; sub $0x6c,%esp + DB 232,0,0,0,0 ; call 4fbe <_sk_load_f16_sse2+0xe> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51510,7 +51466,7 @@ _sk_load_f16_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 15,133,160,1,0,0 ; jne 5134 <_sk_load_f16_sse2+0x1d0> + DB 15,133,173,1,0,0 ; jne 518d <_sk_load_f16_sse2+0x1dd> DB 102,15,16,4,254 ; movupd (%esi,%edi,8),%xmm0 DB 243,15,111,76,254,16 ; movdqu 0x10(%esi,%edi,8),%xmm1 DB 102,15,40,216 ; movapd %xmm0,%xmm3 @@ -51527,82 +51483,83 @@ _sk_load_f16_sse2 LABEL PROC DB 102,15,126,69,220 ; movd %xmm0,-0x24(%ebp) DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,69,216 ; movd %xmm0,-0x28(%ebp) - DB 243,15,126,125,216 ; movq -0x28(%ebp),%xmm7 + DB 243,15,126,101,216 ; movq -0x28(%ebp),%xmm4 DB 102,15,126,93,224 ; movd %xmm3,-0x20(%ebp) DB 102,15,112,195,229 ; pshufd $0xe5,%xmm3,%xmm0 DB 102,15,126,69,228 ; movd %xmm0,-0x1c(%ebp) - DB 243,15,126,109,224 ; movq -0x20(%ebp),%xmm5 + DB 242,15,16,69,224 ; movsd -0x20(%ebp),%xmm0 + DB 15,41,69,152 ; movaps %xmm0,-0x68(%ebp) DB 102,15,112,195,231 ; pshufd $0xe7,%xmm3,%xmm0 DB 102,15,126,69,236 ; movd %xmm0,-0x14(%ebp) DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 DB 102,15,126,69,232 ; movd %xmm0,-0x18(%ebp) DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 - DB 102,15,111,154,110,133,0,0 ; movdqa 0x856e(%edx),%xmm3 - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,219,227 ; pand %xmm3,%xmm4 - DB 102,15,239,204 ; pxor %xmm4,%xmm1 - DB 102,15,114,244,16 ; pslld $0x10,%xmm4 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,114,242,13 ; pslld $0xd,%xmm2 - DB 102,15,235,212 ; por %xmm4,%xmm2 - DB 102,15,97,248 ; punpcklwd %xmm0,%xmm7 - DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,15,111,170,34,132,0,0 ; movdqa 0x8422(%edx),%xmm5 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,15,219,245 ; pand %xmm5,%xmm6 + DB 102,15,219,138,50,132,0,0 ; pand 0x8432(%edx),%xmm1 + DB 102,15,111,130,66,132,0,0 ; movdqa 0x8442(%edx),%xmm0 + DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) + DB 102,15,114,246,16 ; pslld $0x10,%xmm6 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,102,217 ; pcmpgtd %xmm1,%xmm3 + DB 102,15,127,93,136 ; movdqa %xmm3,-0x78(%ebp) + DB 102,15,114,241,13 ; pslld $0xd,%xmm1 + DB 102,15,235,206 ; por %xmm6,%xmm1 + DB 102,15,97,226 ; punpcklwd %xmm2,%xmm4 + DB 102,15,239,255 ; pxor %xmm7,%xmm7 + DB 102,15,111,212 ; movdqa %xmm4,%xmm2 + DB 102,15,219,213 ; pand %xmm5,%xmm2 + DB 102,15,111,154,50,132,0,0 ; movdqa 0x8432(%edx),%xmm3 DB 102,15,219,227 ; pand %xmm3,%xmm4 - DB 102,15,239,252 ; pxor %xmm4,%xmm7 - DB 102,15,127,125,152 ; movdqa %xmm7,-0x68(%ebp) - DB 102,15,114,244,16 ; pslld $0x10,%xmm4 - DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,114,246,13 ; pslld $0xd,%xmm6 - DB 102,15,235,244 ; por %xmm4,%xmm6 - DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 - DB 102,15,111,253 ; movdqa %xmm5,%xmm7 - DB 102,15,219,251 ; pand %xmm3,%xmm7 - DB 102,15,239,239 ; pxor %xmm7,%xmm5 - DB 102,15,127,109,168 ; movdqa %xmm5,-0x58(%ebp) + DB 102,15,114,242,16 ; pslld $0x10,%xmm2 + DB 102,15,102,196 ; pcmpgtd %xmm4,%xmm0 + DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) + DB 102,15,114,244,13 ; pslld $0xd,%xmm4 + DB 102,15,235,226 ; por %xmm2,%xmm4 + DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 + DB 102,15,97,199 ; punpcklwd %xmm7,%xmm0 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,219,253 ; pand %xmm5,%xmm7 + DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 102,15,114,247,16 ; pslld $0x10,%xmm7 - DB 102,15,114,245,13 ; pslld $0xd,%xmm5 - DB 102,15,235,239 ; por %xmm7,%xmm5 - DB 243,15,126,125,232 ; movq -0x18(%ebp),%xmm7 - DB 102,15,97,248 ; punpcklwd %xmm0,%xmm7 - DB 102,15,219,223 ; pand %xmm7,%xmm3 - DB 102,15,239,251 ; pxor %xmm3,%xmm7 - DB 102,15,114,243,16 ; pslld $0x10,%xmm3 - DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,15,111,146,66,132,0,0 ; movdqa 0x8442(%edx),%xmm2 + DB 102,15,102,208 ; pcmpgtd %xmm0,%xmm2 DB 102,15,114,240,13 ; pslld $0xd,%xmm0 - DB 102,15,235,195 ; por %xmm3,%xmm0 - DB 102,15,111,154,126,133,0,0 ; movdqa 0x857e(%edx),%xmm3 - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,254,243 ; paddd %xmm3,%xmm6 - DB 102,15,254,235 ; paddd %xmm3,%xmm5 - DB 102,15,254,195 ; paddd %xmm3,%xmm0 - DB 102,15,111,154,142,133,0,0 ; movdqa 0x858e(%edx),%xmm3 - DB 102,15,127,93,184 ; movdqa %xmm3,-0x48(%ebp) - DB 102,15,239,203 ; pxor %xmm3,%xmm1 - DB 102,15,111,154,158,133,0,0 ; movdqa 0x859e(%edx),%xmm3 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,102,225 ; pcmpgtd %xmm1,%xmm4 - DB 102,15,223,226 ; pandn %xmm2,%xmm4 - DB 102,15,111,85,152 ; movdqa -0x68(%ebp),%xmm2 - DB 102,15,239,85,184 ; pxor -0x48(%ebp),%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,102,202 ; pcmpgtd %xmm2,%xmm1 - DB 102,15,223,206 ; pandn %xmm6,%xmm1 - DB 102,15,111,117,168 ; movdqa -0x58(%ebp),%xmm6 - DB 102,15,239,117,184 ; pxor -0x48(%ebp),%xmm6 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,102,214 ; pcmpgtd %xmm6,%xmm2 - DB 102,15,223,213 ; pandn %xmm5,%xmm2 - DB 102,15,239,125,184 ; pxor -0x48(%ebp),%xmm7 + DB 102,15,235,199 ; por %xmm7,%xmm0 + DB 243,15,126,125,232 ; movq -0x18(%ebp),%xmm7 + DB 102,15,239,246 ; pxor %xmm6,%xmm6 + DB 102,15,97,254 ; punpcklwd %xmm6,%xmm7 + DB 102,15,219,239 ; pand %xmm7,%xmm5 + DB 102,15,219,251 ; pand %xmm3,%xmm7 + DB 102,15,114,245,16 ; pslld $0x10,%xmm5 + DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 DB 102,15,102,223 ; pcmpgtd %xmm7,%xmm3 - DB 102,15,223,216 ; pandn %xmm0,%xmm3 + DB 102,15,127,93,184 ; movdqa %xmm3,-0x48(%ebp) + DB 102,15,114,247,13 ; pslld $0xd,%xmm7 + DB 102,15,235,253 ; por %xmm5,%xmm7 + DB 102,15,111,154,82,132,0,0 ; movdqa 0x8452(%edx),%xmm3 + DB 102,15,254,203 ; paddd %xmm3,%xmm1 + DB 102,15,254,227 ; paddd %xmm3,%xmm4 + DB 102,15,254,195 ; paddd %xmm3,%xmm0 + DB 102,15,254,251 ; paddd %xmm3,%xmm7 + DB 102,15,111,93,136 ; movdqa -0x78(%ebp),%xmm3 + DB 102,15,223,217 ; pandn %xmm1,%xmm3 + DB 102,15,111,77,168 ; movdqa -0x58(%ebp),%xmm1 + DB 102,15,223,204 ; pandn %xmm4,%xmm1 + DB 102,15,223,208 ; pandn %xmm0,%xmm2 + DB 102,15,111,101,184 ; movdqa -0x48(%ebp),%xmm4 + DB 102,15,223,231 ; pandn %xmm7,%xmm4 DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 82 ; push %edx DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) - DB 131,196,108 ; add $0x6c,%esp + DB 131,196,124 ; add $0x7c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx @@ -51611,12 +51568,12 @@ _sk_load_f16_sse2 LABEL PROC DB 242,15,16,4,254 ; movsd (%esi,%edi,8),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 131,251,1 ; cmp $0x1,%ebx - DB 15,132,89,254,255,255 ; je 4f9f <_sk_load_f16_sse2+0x3b> + DB 15,132,76,254,255,255 ; je 4feb <_sk_load_f16_sse2+0x3b> DB 102,15,22,68,254,8 ; movhpd 0x8(%esi,%edi,8),%xmm0 DB 131,251,3 ; cmp $0x3,%ebx - DB 15,130,74,254,255,255 ; jb 4f9f <_sk_load_f16_sse2+0x3b> + DB 15,130,61,254,255,255 ; jb 4feb <_sk_load_f16_sse2+0x3b> DB 243,15,126,76,254,16 ; movq 0x10(%esi,%edi,8),%xmm1 - DB 233,63,254,255,255 ; jmp 4f9f <_sk_load_f16_sse2+0x3b> + DB 233,50,254,255,255 ; jmp 4feb <_sk_load_f16_sse2+0x3b> PUBLIC _sk_load_f16_dst_sse2 _sk_load_f16_dst_sse2 LABEL PROC @@ -51625,12 +51582,12 @@ _sk_load_f16_dst_sse2 LABEL PROC DB 83 ; push %ebx DB 87 ; push %edi DB 86 ; push %esi - DB 129,236,140,0,0,0 ; sub $0x8c,%esp - DB 15,41,157,104,255,255,255 ; movaps %xmm3,-0x98(%ebp) - DB 15,41,149,120,255,255,255 ; movaps %xmm2,-0x88(%ebp) - DB 102,15,127,77,136 ; movdqa %xmm1,-0x78(%ebp) - DB 15,41,69,152 ; movaps %xmm0,-0x68(%ebp) - DB 232,0,0,0,0 ; call 5188 <_sk_load_f16_dst_sse2+0x28> + DB 131,236,124 ; sub $0x7c,%esp + DB 15,41,157,120,255,255,255 ; movaps %xmm3,-0x88(%ebp) + DB 15,41,85,136 ; movaps %xmm2,-0x78(%ebp) + DB 102,15,127,77,152 ; movdqa %xmm1,-0x68(%ebp) + DB 15,41,69,168 ; movaps %xmm0,-0x58(%ebp) + DB 232,0,0,0,0 ; call 51db <_sk_load_f16_dst_sse2+0x22> DB 90 ; pop %edx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx @@ -51642,7 +51599,7 @@ _sk_load_f16_dst_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 133,219 ; test %ebx,%ebx DB 139,57 ; mov (%ecx),%edi - DB 15,133,206,1,0,0 ; jne 5378 <_sk_load_f16_dst_sse2+0x218> + DB 15,133,168,1,0,0 ; jne 53a5 <_sk_load_f16_dst_sse2+0x1ec> DB 102,15,16,4,254 ; movupd (%esi,%edi,8),%xmm0 DB 243,15,111,76,254,16 ; movdqu 0x10(%esi,%edi,8),%xmm1 DB 102,15,40,208 ; movapd %xmm0,%xmm2 @@ -51654,7 +51611,7 @@ _sk_load_f16_dst_sse2 LABEL PROC DB 102,15,126,93,208 ; movd %xmm3,-0x30(%ebp) DB 102,15,112,195,229 ; pshufd $0xe5,%xmm3,%xmm0 DB 102,15,126,69,212 ; movd %xmm0,-0x2c(%ebp) - DB 243,15,126,101,208 ; movq -0x30(%ebp),%xmm4 + DB 243,15,126,77,208 ; movq -0x30(%ebp),%xmm1 DB 102,15,112,195,231 ; pshufd $0xe7,%xmm3,%xmm0 DB 102,15,126,69,220 ; movd %xmm0,-0x24(%ebp) DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 @@ -51663,87 +51620,79 @@ _sk_load_f16_dst_sse2 LABEL PROC DB 102,15,126,85,224 ; movd %xmm2,-0x20(%ebp) DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 DB 102,15,126,93,228 ; movd %xmm3,-0x1c(%ebp) - DB 243,15,126,109,224 ; movq -0x20(%ebp),%xmm5 + DB 243,15,126,117,224 ; movq -0x20(%ebp),%xmm6 DB 102,15,112,218,231 ; pshufd $0xe7,%xmm2,%xmm3 DB 102,15,126,93,236 ; movd %xmm3,-0x14(%ebp) DB 102,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm2 DB 102,15,126,85,232 ; movd %xmm2,-0x18(%ebp) - DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 102,15,97,225 ; punpcklwd %xmm1,%xmm4 - DB 102,15,111,154,88,131,0,0 ; movdqa 0x8358(%edx),%xmm3 - DB 102,15,111,212 ; movdqa %xmm4,%xmm2 - DB 102,15,219,211 ; pand %xmm3,%xmm2 - DB 102,15,239,226 ; pxor %xmm2,%xmm4 + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,15,97,202 ; punpcklwd %xmm2,%xmm1 + DB 102,15,111,186,5,130,0,0 ; movdqa 0x8205(%edx),%xmm7 + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,15,219,215 ; pand %xmm7,%xmm2 + DB 102,15,111,154,21,130,0,0 ; movdqa 0x8215(%edx),%xmm3 + DB 102,15,219,203 ; pand %xmm3,%xmm1 + DB 102,15,111,162,37,130,0,0 ; movdqa 0x8225(%edx),%xmm4 DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,15,102,233 ; pcmpgtd %xmm1,%xmm5 DB 102,15,114,241,13 ; pslld $0xd,%xmm1 DB 102,15,235,202 ; por %xmm2,%xmm1 - DB 102,15,111,146,120,131,0,0 ; movdqa 0x8378(%edx),%xmm2 - DB 102,15,127,85,168 ; movdqa %xmm2,-0x58(%ebp) - DB 102,15,239,226 ; pxor %xmm2,%xmm4 - DB 102,15,111,178,136,131,0,0 ; movdqa 0x8388(%edx),%xmm6 - DB 102,15,111,254 ; movdqa %xmm6,%xmm7 - DB 102,15,102,252 ; pcmpgtd %xmm4,%xmm7 - DB 102,15,111,146,104,131,0,0 ; movdqa 0x8368(%edx),%xmm2 + DB 102,15,111,146,53,130,0,0 ; movdqa 0x8235(%edx),%xmm2 DB 102,15,127,85,184 ; movdqa %xmm2,-0x48(%ebp) DB 102,15,254,202 ; paddd %xmm2,%xmm1 - DB 102,15,223,249 ; pandn %xmm1,%xmm7 - DB 243,15,126,85,232 ; movq -0x18(%ebp),%xmm2 - DB 102,15,127,121,16 ; movdqa %xmm7,0x10(%ecx) - DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,219,203 ; pand %xmm3,%xmm1 - DB 102,15,239,193 ; pxor %xmm1,%xmm0 - DB 102,15,114,241,16 ; pslld $0x10,%xmm1 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,15,114,247,13 ; pslld $0xd,%xmm7 - DB 102,15,235,249 ; por %xmm1,%xmm7 - DB 102,15,111,101,168 ; movdqa -0x58(%ebp),%xmm4 - DB 102,15,239,196 ; pxor %xmm4,%xmm0 - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,15,102,200 ; pcmpgtd %xmm0,%xmm1 - DB 102,15,254,125,184 ; paddd -0x48(%ebp),%xmm7 - DB 102,15,223,207 ; pandn %xmm7,%xmm1 - DB 102,15,127,73,32 ; movdqa %xmm1,0x20(%ecx) - DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 102,15,97,239 ; punpcklwd %xmm7,%xmm5 - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,223,233 ; pandn %xmm1,%xmm5 + DB 243,15,126,77,232 ; movq -0x18(%ebp),%xmm1 + DB 102,15,127,105,16 ; movdqa %xmm5,0x10(%ecx) + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,15,97,194 ; punpcklwd %xmm2,%xmm0 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,219,239 ; pand %xmm7,%xmm5 DB 102,15,219,195 ; pand %xmm3,%xmm0 - DB 102,15,239,232 ; pxor %xmm0,%xmm5 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,114,245,16 ; pslld $0x10,%xmm5 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 + DB 102,15,102,216 ; pcmpgtd %xmm0,%xmm3 + DB 102,15,114,240,13 ; pslld $0xd,%xmm0 + DB 102,15,235,197 ; por %xmm5,%xmm0 + DB 102,15,111,109,184 ; movdqa -0x48(%ebp),%xmm5 + DB 102,15,254,197 ; paddd %xmm5,%xmm0 + DB 102,15,223,216 ; pandn %xmm0,%xmm3 + DB 102,15,127,89,32 ; movdqa %xmm3,0x20(%ecx) + DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 102,15,97,240 ; punpcklwd %xmm0,%xmm6 + DB 102,15,111,198 ; movdqa %xmm6,%xmm0 + DB 102,15,219,199 ; pand %xmm7,%xmm0 + DB 102,15,219,242 ; pand %xmm2,%xmm6 DB 102,15,114,240,16 ; pslld $0x10,%xmm0 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 + DB 102,15,102,222 ; pcmpgtd %xmm6,%xmm3 + DB 102,15,114,246,13 ; pslld $0xd,%xmm6 + DB 102,15,235,240 ; por %xmm0,%xmm6 + DB 102,15,254,245 ; paddd %xmm5,%xmm6 + DB 102,15,223,222 ; pandn %xmm6,%xmm3 + DB 102,15,127,89,48 ; movdqa %xmm3,0x30(%ecx) + DB 102,15,239,219 ; pxor %xmm3,%xmm3 + DB 102,15,97,203 ; punpcklwd %xmm3,%xmm1 + DB 102,15,219,249 ; pand %xmm1,%xmm7 + DB 102,15,219,202 ; pand %xmm2,%xmm1 + DB 102,15,114,247,16 ; pslld $0x10,%xmm7 + DB 102,15,102,225 ; pcmpgtd %xmm1,%xmm4 DB 102,15,114,241,13 ; pslld $0xd,%xmm1 - DB 102,15,235,200 ; por %xmm0,%xmm1 - DB 102,15,239,236 ; pxor %xmm4,%xmm5 - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,15,102,197 ; pcmpgtd %xmm5,%xmm0 - DB 102,15,111,109,184 ; movdqa -0x48(%ebp),%xmm5 + DB 102,15,235,207 ; por %xmm7,%xmm1 DB 102,15,254,205 ; paddd %xmm5,%xmm1 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,15,127,65,48 ; movdqa %xmm0,0x30(%ecx) - DB 102,15,97,215 ; punpcklwd %xmm7,%xmm2 - DB 102,15,219,218 ; pand %xmm2,%xmm3 - DB 102,15,239,211 ; pxor %xmm3,%xmm2 - DB 102,15,114,243,16 ; pslld $0x10,%xmm3 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,114,240,13 ; pslld $0xd,%xmm0 - DB 102,15,235,195 ; por %xmm3,%xmm0 - DB 102,15,254,197 ; paddd %xmm5,%xmm0 - DB 102,15,239,212 ; pxor %xmm4,%xmm2 - DB 102,15,102,242 ; pcmpgtd %xmm2,%xmm6 - DB 102,15,223,240 ; pandn %xmm0,%xmm6 - DB 102,15,127,113,64 ; movdqa %xmm6,0x40(%ecx) + DB 102,15,223,225 ; pandn %xmm1,%xmm4 + DB 102,15,127,97,64 ; movdqa %xmm4,0x40(%ecx) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,69,152 ; movaps -0x68(%ebp),%xmm0 - DB 15,40,77,136 ; movaps -0x78(%ebp),%xmm1 - DB 15,40,149,120,255,255,255 ; movaps -0x88(%ebp),%xmm2 - DB 15,40,157,104,255,255,255 ; movaps -0x98(%ebp),%xmm3 + DB 15,40,69,168 ; movaps -0x58(%ebp),%xmm0 + DB 15,40,77,152 ; movaps -0x68(%ebp),%xmm1 + DB 15,40,85,136 ; movaps -0x78(%ebp),%xmm2 + DB 15,40,157,120,255,255,255 ; movaps -0x88(%ebp),%xmm3 DB 82 ; push %edx DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) - DB 129,196,156,0,0,0 ; add $0x9c,%esp + DB 129,196,140,0,0,0 ; add $0x8c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx @@ -51752,12 +51701,12 @@ _sk_load_f16_dst_sse2 LABEL PROC DB 242,15,16,4,254 ; movsd (%esi,%edi,8),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 131,251,1 ; cmp $0x1,%ebx - DB 15,132,43,254,255,255 ; je 51b5 <_sk_load_f16_dst_sse2+0x55> + DB 15,132,81,254,255,255 ; je 5208 <_sk_load_f16_dst_sse2+0x4f> DB 102,15,22,68,254,8 ; movhpd 0x8(%esi,%edi,8),%xmm0 DB 131,251,3 ; cmp $0x3,%ebx - DB 15,130,28,254,255,255 ; jb 51b5 <_sk_load_f16_dst_sse2+0x55> + DB 15,130,66,254,255,255 ; jb 5208 <_sk_load_f16_dst_sse2+0x4f> DB 243,15,126,76,254,16 ; movq 0x10(%esi,%edi,8),%xmm1 - DB 233,17,254,255,255 ; jmp 51b5 <_sk_load_f16_dst_sse2+0x55> + DB 233,55,254,255,255 ; jmp 5208 <_sk_load_f16_dst_sse2+0x4f> PUBLIC _sk_gather_f16_sse2 _sk_gather_f16_sse2 LABEL PROC @@ -51765,10 +51714,10 @@ _sk_gather_f16_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 131,236,88 ; sub $0x58,%esp DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 139,16 ; mov (%eax),%edx - DB 139,10 ; mov (%edx),%ecx + DB 139,8 ; mov (%eax),%ecx + DB 139,17 ; mov (%ecx),%edx DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 - DB 102,15,110,82,4 ; movd 0x4(%edx),%xmm2 + DB 102,15,110,81,4 ; movd 0x4(%ecx),%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 DB 102,15,112,217,245 ; pshufd $0xf5,%xmm1,%xmm3 DB 102,15,244,218 ; pmuludq %xmm2,%xmm3 @@ -51780,33 +51729,25 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,15,254,193 ; paddd %xmm1,%xmm0 DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 - DB 102,15,126,210 ; movd %xmm2,%edx - DB 102,15,110,84,209,4 ; movd 0x4(%ecx,%edx,8),%xmm2 - DB 102,15,110,28,209 ; movd (%ecx,%edx,8),%xmm3 - DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,110,76,209,4 ; movd 0x4(%ecx,%edx,8),%xmm1 - DB 102,15,110,36,209 ; movd (%ecx,%edx,8),%xmm4 - DB 102,15,126,194 ; movd %xmm0,%edx + DB 102,15,126,209 ; movd %xmm2,%ecx + DB 243,15,126,20,202 ; movq (%edx,%ecx,8),%xmm2 + DB 102,15,126,201 ; movd %xmm1,%ecx + DB 243,15,126,12,202 ; movq (%edx,%ecx,8),%xmm1 + DB 102,15,126,193 ; movd %xmm0,%ecx DB 102,15,112,192,229 ; pshufd $0xe5,%xmm0,%xmm0 - DB 102,15,110,108,209,4 ; movd 0x4(%ecx,%edx,8),%xmm5 - DB 102,15,110,52,209 ; movd (%ecx,%edx,8),%xmm6 - DB 102,15,126,194 ; movd %xmm0,%edx - DB 102,15,110,68,209,4 ; movd 0x4(%ecx,%edx,8),%xmm0 - DB 102,15,110,60,209 ; movd (%ecx,%edx,8),%xmm7 - DB 232,0,0,0,0 ; call 5432 <_sk_gather_f16_sse2+0x8e> + DB 243,15,126,28,202 ; movq (%edx,%ecx,8),%xmm3 + DB 102,15,126,193 ; movd %xmm0,%ecx + DB 243,15,126,4,202 ; movq (%edx,%ecx,8),%xmm0 + DB 232,0,0,0,0 ; call 5447 <_sk_gather_f16_sse2+0x76> DB 89 ; pop %ecx - DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 - DB 102,15,98,227 ; punpckldq %xmm3,%xmm4 - DB 102,15,98,225 ; punpckldq %xmm1,%xmm4 - DB 102,15,98,232 ; punpckldq %xmm0,%xmm5 - DB 102,15,98,247 ; punpckldq %xmm7,%xmm6 - DB 102,15,98,245 ; punpckldq %xmm5,%xmm6 - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,15,97,196 ; punpcklwd %xmm4,%xmm0 - DB 102,15,105,244 ; punpckhwd %xmm4,%xmm6 + DB 102,15,108,202 ; punpcklqdq %xmm2,%xmm1 + DB 102,15,108,216 ; punpcklqdq %xmm0,%xmm3 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 + DB 102,15,105,217 ; punpckhwd %xmm1,%xmm3 DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,97,214 ; punpcklwd %xmm6,%xmm2 - DB 102,15,105,198 ; punpckhwd %xmm6,%xmm0 + DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 + DB 102,15,105,195 ; punpckhwd %xmm3,%xmm0 DB 102,15,126,85,224 ; movd %xmm2,-0x20(%ebp) DB 102,15,112,202,229 ; pshufd $0xe5,%xmm2,%xmm1 DB 102,15,126,77,228 ; movd %xmm1,-0x1c(%ebp) @@ -51815,79 +51756,80 @@ _sk_gather_f16_sse2 LABEL PROC DB 102,15,126,93,236 ; movd %xmm3,-0x14(%ebp) DB 102,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm2 DB 102,15,126,85,232 ; movd %xmm2,-0x18(%ebp) - DB 243,15,126,125,232 ; movq -0x18(%ebp),%xmm7 + DB 243,15,126,101,232 ; movq -0x18(%ebp),%xmm4 DB 102,15,126,69,240 ; movd %xmm0,-0x10(%ebp) DB 102,15,112,216,229 ; pshufd $0xe5,%xmm0,%xmm3 DB 102,15,126,93,244 ; movd %xmm3,-0xc(%ebp) - DB 243,15,126,109,240 ; movq -0x10(%ebp),%xmm5 + DB 243,15,126,85,240 ; movq -0x10(%ebp),%xmm2 + DB 102,15,127,85,184 ; movdqa %xmm2,-0x48(%ebp) DB 102,15,112,216,231 ; pshufd $0xe7,%xmm0,%xmm3 DB 102,15,126,93,252 ; movd %xmm3,-0x4(%ebp) DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 DB 102,15,126,69,248 ; movd %xmm0,-0x8(%ebp) DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 - DB 102,15,111,153,174,128,0,0 ; movdqa 0x80ae(%ecx),%xmm3 - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,219,227 ; pand %xmm3,%xmm4 - DB 102,15,239,204 ; pxor %xmm4,%xmm1 - DB 102,15,114,244,16 ; pslld $0x10,%xmm4 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,114,242,13 ; pslld $0xd,%xmm2 - DB 102,15,235,212 ; por %xmm4,%xmm2 - DB 102,15,97,248 ; punpcklwd %xmm0,%xmm7 - DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,15,111,169,153,127,0,0 ; movdqa 0x7f99(%ecx),%xmm5 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,15,219,245 ; pand %xmm5,%xmm6 + DB 102,15,219,137,169,127,0,0 ; pand 0x7fa9(%ecx),%xmm1 + DB 102,15,111,129,185,127,0,0 ; movdqa 0x7fb9(%ecx),%xmm0 + DB 102,15,114,246,16 ; pslld $0x10,%xmm6 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,102,217 ; pcmpgtd %xmm1,%xmm3 + DB 102,15,127,93,168 ; movdqa %xmm3,-0x58(%ebp) + DB 102,15,114,241,13 ; pslld $0xd,%xmm1 + DB 102,15,235,206 ; por %xmm6,%xmm1 + DB 102,15,97,226 ; punpcklwd %xmm2,%xmm4 + DB 102,15,239,255 ; pxor %xmm7,%xmm7 + DB 102,15,111,212 ; movdqa %xmm4,%xmm2 + DB 102,15,219,213 ; pand %xmm5,%xmm2 + DB 102,15,111,153,169,127,0,0 ; movdqa 0x7fa9(%ecx),%xmm3 DB 102,15,219,227 ; pand %xmm3,%xmm4 - DB 102,15,239,252 ; pxor %xmm4,%xmm7 - DB 102,15,127,125,168 ; movdqa %xmm7,-0x58(%ebp) - DB 102,15,114,244,16 ; pslld $0x10,%xmm4 - DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,114,246,13 ; pslld $0xd,%xmm6 - DB 102,15,235,244 ; por %xmm4,%xmm6 - DB 102,15,97,232 ; punpcklwd %xmm0,%xmm5 - DB 102,15,111,253 ; movdqa %xmm5,%xmm7 - DB 102,15,219,251 ; pand %xmm3,%xmm7 - DB 102,15,239,239 ; pxor %xmm7,%xmm5 - DB 102,15,127,109,184 ; movdqa %xmm5,-0x48(%ebp) + DB 102,15,114,242,16 ; pslld $0x10,%xmm2 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,102,244 ; pcmpgtd %xmm4,%xmm6 + DB 102,15,127,117,200 ; movdqa %xmm6,-0x38(%ebp) + DB 102,15,114,244,13 ; pslld $0xd,%xmm4 + DB 102,15,235,226 ; por %xmm2,%xmm4 + DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 + DB 102,15,97,199 ; punpcklwd %xmm7,%xmm0 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,219,253 ; pand %xmm5,%xmm7 + DB 102,15,219,195 ; pand %xmm3,%xmm0 DB 102,15,114,247,16 ; pslld $0x10,%xmm7 - DB 102,15,114,245,13 ; pslld $0xd,%xmm5 - DB 102,15,235,239 ; por %xmm7,%xmm5 - DB 243,15,126,125,248 ; movq -0x8(%ebp),%xmm7 - DB 102,15,97,248 ; punpcklwd %xmm0,%xmm7 - DB 102,15,219,223 ; pand %xmm7,%xmm3 - DB 102,15,239,251 ; pxor %xmm3,%xmm7 - DB 102,15,114,243,16 ; pslld $0x10,%xmm3 - DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,15,111,145,185,127,0,0 ; movdqa 0x7fb9(%ecx),%xmm2 + DB 102,15,111,218 ; movdqa %xmm2,%xmm3 + DB 102,15,102,216 ; pcmpgtd %xmm0,%xmm3 DB 102,15,114,240,13 ; pslld $0xd,%xmm0 - DB 102,15,235,195 ; por %xmm3,%xmm0 - DB 102,15,111,153,190,128,0,0 ; movdqa 0x80be(%ecx),%xmm3 - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,254,243 ; paddd %xmm3,%xmm6 - DB 102,15,254,235 ; paddd %xmm3,%xmm5 - DB 102,15,254,195 ; paddd %xmm3,%xmm0 - DB 102,15,111,153,206,128,0,0 ; movdqa 0x80ce(%ecx),%xmm3 - DB 102,15,127,93,200 ; movdqa %xmm3,-0x38(%ebp) - DB 102,15,239,203 ; pxor %xmm3,%xmm1 - DB 102,15,111,153,222,128,0,0 ; movdqa 0x80de(%ecx),%xmm3 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,102,225 ; pcmpgtd %xmm1,%xmm4 - DB 102,15,223,226 ; pandn %xmm2,%xmm4 - DB 102,15,111,85,168 ; movdqa -0x58(%ebp),%xmm2 - DB 102,15,239,85,200 ; pxor -0x38(%ebp),%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,102,202 ; pcmpgtd %xmm2,%xmm1 - DB 102,15,223,206 ; pandn %xmm6,%xmm1 - DB 102,15,111,117,184 ; movdqa -0x48(%ebp),%xmm6 - DB 102,15,239,117,200 ; pxor -0x38(%ebp),%xmm6 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,102,214 ; pcmpgtd %xmm6,%xmm2 - DB 102,15,223,213 ; pandn %xmm5,%xmm2 - DB 102,15,239,125,200 ; pxor -0x38(%ebp),%xmm7 - DB 102,15,102,223 ; pcmpgtd %xmm7,%xmm3 + DB 102,15,235,199 ; por %xmm7,%xmm0 + DB 243,15,126,125,248 ; movq -0x8(%ebp),%xmm7 + DB 102,15,239,246 ; pxor %xmm6,%xmm6 + DB 102,15,97,254 ; punpcklwd %xmm6,%xmm7 + DB 102,15,219,239 ; pand %xmm7,%xmm5 + DB 102,15,219,185,169,127,0,0 ; pand 0x7fa9(%ecx),%xmm7 + DB 102,15,114,245,16 ; pslld $0x10,%xmm5 + DB 102,15,102,215 ; pcmpgtd %xmm7,%xmm2 + DB 102,15,114,247,13 ; pslld $0xd,%xmm7 + DB 102,15,235,253 ; por %xmm5,%xmm7 + DB 102,15,111,169,201,127,0,0 ; movdqa 0x7fc9(%ecx),%xmm5 + DB 102,15,254,205 ; paddd %xmm5,%xmm1 + DB 102,15,254,229 ; paddd %xmm5,%xmm4 + DB 102,15,254,197 ; paddd %xmm5,%xmm0 + DB 102,15,254,253 ; paddd %xmm5,%xmm7 + DB 102,15,111,109,168 ; movdqa -0x58(%ebp),%xmm5 + DB 102,15,223,233 ; pandn %xmm1,%xmm5 + DB 102,15,111,77,200 ; movdqa -0x38(%ebp),%xmm1 + DB 102,15,223,204 ; pandn %xmm4,%xmm1 DB 102,15,223,216 ; pandn %xmm0,%xmm3 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,15,223,231 ; pandn %xmm7,%xmm4 DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 82 ; push %edx DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) @@ -51901,74 +51843,77 @@ _sk_store_f16_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 87 ; push %edi DB 86 ; push %esi - DB 131,236,80 ; sub $0x50,%esp - DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) - DB 102,15,111,218 ; movdqa %xmm2,%xmm3 - DB 232,0,0,0,0 ; call 55f5 <_sk_store_f16_sse2+0x15> + DB 131,236,96 ; sub $0x60,%esp + DB 15,41,93,184 ; movaps %xmm3,-0x48(%ebp) + DB 15,41,85,200 ; movaps %xmm2,-0x38(%ebp) + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 232,0,0,0,0 ; call 560c <_sk_store_f16_sse2+0x19> DB 88 ; pop %eax - DB 102,15,111,176,11,127,0,0 ; movdqa 0x7f0b(%eax),%xmm6 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,219,214 ; pand %xmm6,%xmm2 - DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,239,234 ; pxor %xmm2,%xmm5 - DB 102,15,111,160,43,127,0,0 ; movdqa 0x7f2b(%eax),%xmm4 - DB 102,15,114,210,16 ; psrld $0x10,%xmm2 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,102,197 ; pcmpgtd %xmm5,%xmm0 - DB 102,15,114,213,13 ; psrld $0xd,%xmm5 - DB 102,15,111,184,59,127,0,0 ; movdqa 0x7f3b(%eax),%xmm7 - DB 102,15,254,215 ; paddd %xmm7,%xmm2 - DB 102,15,127,125,232 ; movdqa %xmm7,-0x18(%ebp) - DB 102,15,254,213 ; paddd %xmm5,%xmm2 - DB 102,15,114,242,16 ; pslld $0x10,%xmm2 - DB 102,15,114,226,16 ; psrad $0x10,%xmm2 - DB 102,15,223,194 ; pandn %xmm2,%xmm0 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,219,238 ; pand %xmm6,%xmm5 - DB 102,15,127,77,184 ; movdqa %xmm1,-0x48(%ebp) - DB 102,15,239,205 ; pxor %xmm5,%xmm1 - DB 102,15,114,213,16 ; psrld $0x10,%xmm5 - DB 102,15,111,212 ; movdqa %xmm4,%xmm2 - DB 102,15,102,209 ; pcmpgtd %xmm1,%xmm2 - DB 102,15,114,209,13 ; psrld $0xd,%xmm1 - DB 102,15,254,239 ; paddd %xmm7,%xmm5 - DB 102,15,254,233 ; paddd %xmm1,%xmm5 - DB 102,15,114,245,16 ; pslld $0x10,%xmm5 - DB 102,15,114,229,16 ; psrad $0x10,%xmm5 - DB 102,15,223,213 ; pandn %xmm5,%xmm2 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,219,206 ; pand %xmm6,%xmm1 - DB 102,15,127,93,168 ; movdqa %xmm3,-0x58(%ebp) - DB 102,15,239,217 ; pxor %xmm1,%xmm3 - DB 102,15,114,209,16 ; psrld $0x10,%xmm1 - DB 102,15,111,236 ; movdqa %xmm4,%xmm5 - DB 102,15,102,235 ; pcmpgtd %xmm3,%xmm5 - DB 102,15,114,211,13 ; psrld $0xd,%xmm3 - DB 102,15,254,77,232 ; paddd -0x18(%ebp),%xmm1 - DB 102,15,254,203 ; paddd %xmm3,%xmm1 - DB 102,15,111,93,216 ; movdqa -0x28(%ebp),%xmm3 - DB 102,15,114,241,16 ; pslld $0x10,%xmm1 - DB 102,15,114,225,16 ; psrad $0x10,%xmm1 - DB 102,15,223,233 ; pandn %xmm1,%xmm5 - DB 102,15,219,243 ; pand %xmm3,%xmm6 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,239,206 ; pxor %xmm6,%xmm1 - DB 102,15,114,214,16 ; psrld $0x10,%xmm6 - DB 102,15,254,117,232 ; paddd -0x18(%ebp),%xmm6 - DB 102,15,102,225 ; pcmpgtd %xmm1,%xmm4 - DB 102,15,114,209,13 ; psrld $0xd,%xmm1 - DB 102,15,254,241 ; paddd %xmm1,%xmm6 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,114,246,16 ; pslld $0x10,%xmm6 + DB 102,15,111,152,20,126,0,0 ; movdqa 0x7e14(%eax),%xmm3 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,219,227 ; pand %xmm3,%xmm4 + DB 102,15,111,136,68,126,0,0 ; movdqa 0x7e44(%eax),%xmm1 + DB 102,15,127,77,232 ; movdqa %xmm1,-0x18(%ebp) + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,114,246,3 ; pslld $0x3,%xmm6 + DB 102,15,219,241 ; pand %xmm1,%xmm6 + DB 102,15,254,244 ; paddd %xmm4,%xmm6 + DB 102,15,111,184,36,126,0,0 ; movdqa 0x7e24(%eax),%xmm7 + DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) + DB 102,15,219,199 ; pand %xmm7,%xmm0 + DB 102,15,111,168,52,126,0,0 ; movdqa 0x7e34(%eax),%xmm5 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,15,102,200 ; pcmpgtd %xmm0,%xmm1 + DB 102,15,111,128,196,122,0,0 ; movdqa 0x7ac4(%eax),%xmm0 + DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) + DB 102,15,254,240 ; paddd %xmm0,%xmm6 + DB 102,15,114,230,16 ; psrad $0x10,%xmm6 + DB 102,15,223,206 ; pandn %xmm6,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,219,195 ; pand %xmm3,%xmm0 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,114,246,3 ; pslld $0x3,%xmm6 + DB 102,15,219,117,232 ; pand -0x18(%ebp),%xmm6 + DB 102,15,254,240 ; paddd %xmm0,%xmm6 + DB 102,15,127,85,152 ; movdqa %xmm2,-0x68(%ebp) + DB 102,15,219,215 ; pand %xmm7,%xmm2 + DB 102,15,111,229 ; movdqa %xmm5,%xmm4 + DB 102,15,102,226 ; pcmpgtd %xmm2,%xmm4 + DB 102,15,254,117,216 ; paddd -0x28(%ebp),%xmm6 DB 102,15,114,230,16 ; psrad $0x10,%xmm6 DB 102,15,223,230 ; pandn %xmm6,%xmm4 - DB 102,15,107,201 ; packssdw %xmm1,%xmm1 - DB 102,15,107,210 ; packssdw %xmm2,%xmm2 - DB 102,15,97,202 ; punpcklwd %xmm2,%xmm1 - DB 102,15,107,237 ; packssdw %xmm5,%xmm5 - DB 102,15,107,228 ; packssdw %xmm4,%xmm4 - DB 102,15,97,236 ; punpcklwd %xmm4,%xmm5 + DB 102,15,107,200 ; packssdw %xmm0,%xmm1 + DB 102,15,107,224 ; packssdw %xmm0,%xmm4 + DB 102,15,97,204 ; punpcklwd %xmm4,%xmm1 + DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,219,195 ; pand %xmm3,%xmm0 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,15,114,244,3 ; pslld $0x3,%xmm4 + DB 102,15,219,101,232 ; pand -0x18(%ebp),%xmm4 + DB 102,15,254,224 ; paddd %xmm0,%xmm4 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,219,199 ; pand %xmm7,%xmm0 + DB 102,15,111,245 ; movdqa %xmm5,%xmm6 + DB 102,15,102,240 ; pcmpgtd %xmm0,%xmm6 + DB 102,15,111,85,216 ; movdqa -0x28(%ebp),%xmm2 + DB 102,15,254,226 ; paddd %xmm2,%xmm4 + DB 102,15,114,228,16 ; psrad $0x10,%xmm4 + DB 102,15,223,244 ; pandn %xmm4,%xmm6 + DB 102,15,111,101,184 ; movdqa -0x48(%ebp),%xmm4 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,114,240,3 ; pslld $0x3,%xmm0 + DB 102,15,219,69,232 ; pand -0x18(%ebp),%xmm0 + DB 102,15,219,220 ; pand %xmm4,%xmm3 + DB 102,15,254,195 ; paddd %xmm3,%xmm0 + DB 102,15,254,194 ; paddd %xmm2,%xmm0 + DB 102,15,219,252 ; pand %xmm4,%xmm7 + DB 102,15,102,239 ; pcmpgtd %xmm7,%xmm5 + DB 102,15,114,224,16 ; psrad $0x10,%xmm0 + DB 102,15,223,232 ; pandn %xmm0,%xmm5 + DB 102,15,107,240 ; packssdw %xmm0,%xmm6 + DB 102,15,107,232 ; packssdw %xmm0,%xmm5 + DB 102,15,97,245 ; punpcklwd %xmm5,%xmm6 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,17 ; mov (%ecx),%edx @@ -51978,37 +51923,36 @@ _sk_store_f16_sse2 LABEL PROC DB 193,230,3 ; shl $0x3,%esi DB 3,55 ; add (%edi),%esi DB 139,121,8 ; mov 0x8(%ecx),%edi - DB 133,255 ; test %edi,%edi - DB 117,53 ; jne 5767 <_sk_store_f16_sse2+0x187> DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,98,197 ; punpckldq %xmm5,%xmm0 - DB 243,15,127,4,214 ; movdqu %xmm0,(%esi,%edx,8) - DB 102,15,106,205 ; punpckhdq %xmm5,%xmm1 + DB 102,15,98,198 ; punpckldq %xmm6,%xmm0 + DB 133,255 ; test %edi,%edi + DB 117,48 ; jne 578d <_sk_store_f16_sse2+0x19a> + DB 15,17,4,214 ; movups %xmm0,(%esi,%edx,8) + DB 102,15,106,206 ; punpckhdq %xmm6,%xmm1 DB 243,15,127,76,214,16 ; movdqu %xmm1,0x10(%esi,%edx,8) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,69,200 ; movaps -0x38(%ebp),%xmm0 - DB 15,40,77,184 ; movaps -0x48(%ebp),%xmm1 - DB 15,40,85,168 ; movaps -0x58(%ebp),%xmm2 + DB 15,40,69,168 ; movaps -0x58(%ebp),%xmm0 + DB 15,40,77,152 ; movaps -0x68(%ebp),%xmm1 + DB 15,40,85,200 ; movaps -0x38(%ebp),%xmm2 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 82 ; push %edx DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) - DB 131,196,96 ; add $0x60,%esp + DB 131,196,112 ; add $0x70,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 93 ; pop %ebp DB 195 ; ret - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,98,213 ; punpckldq %xmm5,%xmm2 - DB 102,15,214,20,214 ; movq %xmm2,(%esi,%edx,8) + DB 102,15,214,4,214 ; movq %xmm0,(%esi,%edx,8) DB 131,255,1 ; cmp $0x1,%edi - DB 116,208 ; je 5749 <_sk_store_f16_sse2+0x169> - DB 102,15,23,84,214,8 ; movhpd %xmm2,0x8(%esi,%edx,8) + DB 116,212 ; je 576b <_sk_store_f16_sse2+0x178> + DB 102,15,23,68,214,8 ; movhpd %xmm0,0x8(%esi,%edx,8) DB 131,255,3 ; cmp $0x3,%edi - DB 114,197 ; jb 5749 <_sk_store_f16_sse2+0x169> - DB 102,15,106,205 ; punpckhdq %xmm5,%xmm1 + DB 114,201 ; jb 576b <_sk_store_f16_sse2+0x178> + DB 102,15,106,206 ; punpckhdq %xmm6,%xmm1 DB 102,15,214,76,214,16 ; movq %xmm1,0x10(%esi,%edx,8) - DB 235,185 ; jmp 5749 <_sk_store_f16_sse2+0x169> + DB 235,189 ; jmp 576b <_sk_store_f16_sse2+0x178> PUBLIC _sk_load_u16_be_sse2 _sk_load_u16_be_sse2 LABEL PROC @@ -52029,9 +51973,9 @@ _sk_load_u16_be_sse2 LABEL PROC DB 139,89,8 ; mov 0x8(%ecx),%ebx DB 193,230,2 ; shl $0x2,%esi DB 133,219 ; test %ebx,%ebx - DB 232,0,0,0,0 ; call 57bb <_sk_load_u16_be_sse2+0x2b> + DB 232,0,0,0,0 ; call 57d9 <_sk_load_u16_be_sse2+0x2b> DB 90 ; pop %edx - DB 15,133,11,1,0,0 ; jne 58cd <_sk_load_u16_be_sse2+0x13d> + DB 15,133,11,1,0,0 ; jne 58eb <_sk_load_u16_be_sse2+0x13d> DB 102,15,16,4,119 ; movupd (%edi,%esi,2),%xmm0 DB 243,15,111,76,119,16 ; movdqu 0x10(%edi,%esi,2),%xmm1 DB 102,15,40,208 ; movapd %xmm0,%xmm2 @@ -52052,7 +51996,7 @@ _sk_load_u16_be_sse2 LABEL PROC DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,97,195 ; punpcklwd %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,40,162,245,122,0,0 ; movaps 0x7af5(%edx),%xmm4 + DB 15,40,162,215,121,0,0 ; movaps 0x79d7(%edx),%xmm4 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -52102,12 +52046,12 @@ _sk_load_u16_be_sse2 LABEL PROC DB 242,15,16,4,119 ; movsd (%edi,%esi,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 131,251,1 ; cmp $0x1,%ebx - DB 15,132,238,254,255,255 ; je 57cd <_sk_load_u16_be_sse2+0x3d> + DB 15,132,238,254,255,255 ; je 57eb <_sk_load_u16_be_sse2+0x3d> DB 102,15,22,68,119,8 ; movhpd 0x8(%edi,%esi,2),%xmm0 DB 131,251,3 ; cmp $0x3,%ebx - DB 15,130,223,254,255,255 ; jb 57cd <_sk_load_u16_be_sse2+0x3d> + DB 15,130,223,254,255,255 ; jb 57eb <_sk_load_u16_be_sse2+0x3d> DB 243,15,126,76,119,16 ; movq 0x10(%edi,%esi,2),%xmm1 - DB 233,212,254,255,255 ; jmp 57cd <_sk_load_u16_be_sse2+0x3d> + DB 233,212,254,255,255 ; jmp 57eb <_sk_load_u16_be_sse2+0x3d> PUBLIC _sk_load_rgb_u16_be_sse2 _sk_load_rgb_u16_be_sse2 LABEL PROC @@ -52127,9 +52071,9 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 3,50 ; add (%edx),%esi DB 107,25,3 ; imul $0x3,(%ecx),%ebx DB 133,255 ; test %edi,%edi - DB 232,0,0,0,0 ; call 5922 <_sk_load_rgb_u16_be_sse2+0x29> + DB 232,0,0,0,0 ; call 5940 <_sk_load_rgb_u16_be_sse2+0x29> DB 90 ; pop %edx - DB 15,133,240,0,0,0 ; jne 5a19 <_sk_load_rgb_u16_be_sse2+0x120> + DB 15,133,240,0,0,0 ; jne 5a37 <_sk_load_rgb_u16_be_sse2+0x120> DB 243,15,111,20,94 ; movdqu (%esi,%ebx,2),%xmm2 DB 243,15,111,92,94,8 ; movdqu 0x8(%esi,%ebx,2),%xmm3 DB 102,15,115,219,4 ; psrldq $0x4,%xmm3 @@ -52137,8 +52081,8 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,15,115,216,6 ; psrldq $0x6,%xmm0 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,115,217,6 ; psrldq $0x6,%xmm1 - DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 + DB 102,15,97,211 ; punpcklwd %xmm3,%xmm2 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1 DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 @@ -52154,7 +52098,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,97,195 ; punpcklwd %xmm3,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 - DB 15,40,170,142,121,0,0 ; movaps 0x798e(%edx),%xmm5 + DB 15,40,170,112,120,0,0 ; movaps 0x7870(%edx),%xmm5 DB 15,89,197 ; mulps %xmm5,%xmm0 DB 102,15,111,204 ; movdqa %xmm4,%xmm1 DB 102,15,113,241,8 ; psllw $0x8,%xmm1 @@ -52180,7 +52124,7 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 15,89,213 ; mulps %xmm5,%xmm2 DB 141,112,8 ; lea 0x8(%eax),%esi DB 131,236,8 ; sub $0x8,%esp - DB 15,40,154,46,119,0,0 ; movaps 0x772e(%edx),%xmm3 + DB 15,40,154,0,118,0,0 ; movaps 0x7600(%edx),%xmm3 DB 86 ; push %esi DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) @@ -52194,20 +52138,20 @@ _sk_load_rgb_u16_be_sse2 LABEL PROC DB 102,15,196,84,94,4,2 ; pinsrw $0x2,0x4(%esi,%ebx,2),%xmm2 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 131,255,1 ; cmp $0x1,%edi - DB 117,13 ; jne 5a3b <_sk_load_rgb_u16_be_sse2+0x142> + DB 117,13 ; jne 5a59 <_sk_load_rgb_u16_be_sse2+0x142> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 233,16,255,255,255 ; jmp 594b <_sk_load_rgb_u16_be_sse2+0x52> + DB 233,16,255,255,255 ; jmp 5969 <_sk_load_rgb_u16_be_sse2+0x52> DB 102,15,110,68,94,6 ; movd 0x6(%esi,%ebx,2),%xmm0 DB 102,15,196,68,94,10,2 ; pinsrw $0x2,0xa(%esi,%ebx,2),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 131,255,3 ; cmp $0x3,%edi - DB 114,18 ; jb 5a63 <_sk_load_rgb_u16_be_sse2+0x16a> + DB 114,18 ; jb 5a81 <_sk_load_rgb_u16_be_sse2+0x16a> DB 102,15,110,92,94,12 ; movd 0xc(%esi,%ebx,2),%xmm3 DB 102,15,196,92,94,16,2 ; pinsrw $0x2,0x10(%esi,%ebx,2),%xmm3 - DB 233,232,254,255,255 ; jmp 594b <_sk_load_rgb_u16_be_sse2+0x52> + DB 233,232,254,255,255 ; jmp 5969 <_sk_load_rgb_u16_be_sse2+0x52> DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 233,223,254,255,255 ; jmp 594b <_sk_load_rgb_u16_be_sse2+0x52> + DB 233,223,254,255,255 ; jmp 5969 <_sk_load_rgb_u16_be_sse2+0x52> PUBLIC _sk_store_u16_be_sse2 _sk_store_u16_be_sse2 LABEL PROC @@ -52215,91 +52159,86 @@ _sk_store_u16_be_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 87 ; push %edi DB 86 ; push %esi - DB 131,236,16 ; sub $0x10,%esp - DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) - DB 232,0,0,0,0 ; call 5a7d <_sk_store_u16_be_sse2+0x11> - DB 88 ; pop %eax - DB 15,40,176,195,122,0,0 ; movaps 0x7ac3(%eax),%xmm6 + DB 232,0,0,0,0 ; call 5a94 <_sk_store_u16_be_sse2+0xa> + DB 95 ; pop %edi + DB 139,69,12 ; mov 0xc(%ebp),%eax + DB 139,77,8 ; mov 0x8(%ebp),%ecx + DB 139,48 ; mov (%eax),%esi + DB 139,86,4 ; mov 0x4(%esi),%edx + DB 15,175,81,4 ; imul 0x4(%ecx),%edx + DB 1,210 ; add %edx,%edx + DB 3,22 ; add (%esi),%edx + DB 139,49 ; mov (%ecx),%esi + DB 15,40,183,204,121,0,0 ; movaps 0x79cc(%edi),%xmm6 + DB 139,121,8 ; mov 0x8(%ecx),%edi + DB 193,230,2 ; shl $0x2,%esi DB 15,40,224 ; movaps %xmm0,%xmm4 DB 15,89,230 ; mulps %xmm6,%xmm4 DB 102,15,91,228 ; cvtps2dq %xmm4,%xmm4 DB 102,15,114,244,16 ; pslld $0x10,%xmm4 DB 102,15,114,228,16 ; psrad $0x10,%xmm4 - DB 102,15,107,228 ; packssdw %xmm4,%xmm4 + DB 102,15,107,224 ; packssdw %xmm0,%xmm4 DB 102,15,111,236 ; movdqa %xmm4,%xmm5 DB 102,15,113,245,8 ; psllw $0x8,%xmm5 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,235,229 ; por %xmm5,%xmm4 DB 15,40,233 ; movaps %xmm1,%xmm5 DB 15,89,238 ; mulps %xmm6,%xmm5 - DB 102,15,91,253 ; cvtps2dq %xmm5,%xmm7 - DB 102,15,114,247,16 ; pslld $0x10,%xmm7 - DB 102,15,114,231,16 ; psrad $0x10,%xmm7 - DB 102,15,107,255 ; packssdw %xmm7,%xmm7 - DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,113,245,8 ; psllw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,235,253 ; por %xmm5,%xmm7 + DB 102,15,91,237 ; cvtps2dq %xmm5,%xmm5 + DB 102,15,114,245,16 ; pslld $0x10,%xmm5 + DB 102,15,114,229,16 ; psrad $0x10,%xmm5 + DB 102,15,107,232 ; packssdw %xmm0,%xmm5 + DB 102,15,111,253 ; movdqa %xmm5,%xmm7 + DB 102,15,113,247,8 ; psllw $0x8,%xmm7 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,235,239 ; por %xmm7,%xmm5 + DB 102,15,97,229 ; punpcklwd %xmm5,%xmm4 DB 15,40,234 ; movaps %xmm2,%xmm5 DB 15,89,238 ; mulps %xmm6,%xmm5 DB 102,15,91,237 ; cvtps2dq %xmm5,%xmm5 DB 102,15,114,245,16 ; pslld $0x10,%xmm5 DB 102,15,114,229,16 ; psrad $0x10,%xmm5 - DB 102,15,107,237 ; packssdw %xmm5,%xmm5 - DB 102,15,111,221 ; movdqa %xmm5,%xmm3 - DB 102,15,113,243,8 ; psllw $0x8,%xmm3 + DB 102,15,107,232 ; packssdw %xmm0,%xmm5 + DB 102,15,111,253 ; movdqa %xmm5,%xmm7 + DB 102,15,113,247,8 ; psllw $0x8,%xmm7 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,235,235 ; por %xmm3,%xmm5 - DB 15,89,117,232 ; mulps -0x18(%ebp),%xmm6 - DB 102,15,91,222 ; cvtps2dq %xmm6,%xmm3 - DB 102,15,114,243,16 ; pslld $0x10,%xmm3 - DB 102,15,114,227,16 ; psrad $0x10,%xmm3 - DB 102,15,107,219 ; packssdw %xmm3,%xmm3 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,15,113,246,8 ; psllw $0x8,%xmm6 - DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,235,222 ; por %xmm6,%xmm3 - DB 102,15,97,231 ; punpcklwd %xmm7,%xmm4 - DB 102,15,97,235 ; punpcklwd %xmm3,%xmm5 - DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 139,77,8 ; mov 0x8(%ebp),%ecx - DB 139,17 ; mov (%ecx),%edx - DB 139,56 ; mov (%eax),%edi - DB 139,119,4 ; mov 0x4(%edi),%esi - DB 15,175,113,4 ; imul 0x4(%ecx),%esi - DB 1,246 ; add %esi,%esi - DB 3,55 ; add (%edi),%esi - DB 139,121,8 ; mov 0x8(%ecx),%edi - DB 193,226,2 ; shl $0x2,%edx + DB 102,15,235,239 ; por %xmm7,%xmm5 + DB 15,89,243 ; mulps %xmm3,%xmm6 + DB 102,15,91,246 ; cvtps2dq %xmm6,%xmm6 + DB 102,15,114,246,16 ; pslld $0x10,%xmm6 + DB 102,15,114,230,16 ; psrad $0x10,%xmm6 + DB 102,15,107,240 ; packssdw %xmm0,%xmm6 + DB 102,15,111,254 ; movdqa %xmm6,%xmm7 + DB 102,15,113,247,8 ; psllw $0x8,%xmm7 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,235,247 ; por %xmm7,%xmm6 + DB 102,15,97,238 ; punpcklwd %xmm6,%xmm5 + DB 102,15,111,244 ; movdqa %xmm4,%xmm6 + DB 102,15,98,245 ; punpckldq %xmm5,%xmm6 DB 133,255 ; test %edi,%edi - DB 117,45 ; jne 5b7f <_sk_store_u16_be_sse2+0x113> - DB 102,15,111,220 ; movdqa %xmm4,%xmm3 - DB 102,15,98,221 ; punpckldq %xmm5,%xmm3 - DB 243,15,127,28,86 ; movdqu %xmm3,(%esi,%edx,2) + DB 117,32 ; jne 5b90 <_sk_store_u16_be_sse2+0x106> + DB 15,17,52,114 ; movups %xmm6,(%edx,%esi,2) DB 102,15,106,229 ; punpckhdq %xmm5,%xmm4 - DB 243,15,127,100,86,16 ; movdqu %xmm4,0x10(%esi,%edx,2) + DB 243,15,127,100,114,16 ; movdqu %xmm4,0x10(%edx,%esi,2) DB 141,80,8 ; lea 0x8(%eax),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,93,232 ; movaps -0x18(%ebp),%xmm3 DB 82 ; push %edx DB 81 ; push %ecx DB 255,80,4 ; call *0x4(%eax) - DB 131,196,32 ; add $0x20,%esp + DB 131,196,16 ; add $0x10,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 93 ; pop %ebp DB 195 ; ret - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,98,245 ; punpckldq %xmm5,%xmm6 - DB 102,15,214,52,86 ; movq %xmm6,(%esi,%edx,2) + DB 102,15,214,52,114 ; movq %xmm6,(%edx,%esi,2) DB 131,255,1 ; cmp $0x1,%edi - DB 116,216 ; je 5b69 <_sk_store_u16_be_sse2+0xfd> - DB 102,15,23,116,86,8 ; movhpd %xmm6,0x8(%esi,%edx,2) + DB 116,228 ; je 5b7e <_sk_store_u16_be_sse2+0xf4> + DB 102,15,23,116,114,8 ; movhpd %xmm6,0x8(%edx,%esi,2) DB 131,255,3 ; cmp $0x3,%edi - DB 114,205 ; jb 5b69 <_sk_store_u16_be_sse2+0xfd> + DB 114,217 ; jb 5b7e <_sk_store_u16_be_sse2+0xf4> DB 102,15,106,229 ; punpckhdq %xmm5,%xmm4 - DB 102,15,214,100,86,16 ; movq %xmm4,0x10(%esi,%edx,2) - DB 235,193 ; jmp 5b69 <_sk_store_u16_be_sse2+0xfd> + DB 102,15,214,100,114,16 ; movq %xmm4,0x10(%edx,%esi,2) + DB 235,205 ; jmp 5b7e <_sk_store_u16_be_sse2+0xf4> PUBLIC _sk_load_f32_sse2 _sk_load_f32_sse2 LABEL PROC @@ -52320,7 +52259,7 @@ _sk_load_f32_sse2 LABEL PROC DB 15,16,36,23 ; movups (%edi,%edx,1),%xmm4 DB 139,121,8 ; mov 0x8(%ecx),%edi DB 133,255 ; test %edi,%edi - DB 117,71 ; jne 5c1f <_sk_load_f32_sse2+0x77> + DB 117,71 ; jne 5c28 <_sk_load_f32_sse2+0x77> DB 15,16,84,178,16 ; movups 0x10(%edx,%esi,4),%xmm2 DB 15,16,92,178,32 ; movups 0x20(%edx,%esi,4),%xmm3 DB 15,16,68,178,48 ; movups 0x30(%edx,%esi,4),%xmm0 @@ -52348,17 +52287,17 @@ _sk_load_f32_sse2 LABEL PROC DB 195 ; ret DB 15,87,192 ; xorps %xmm0,%xmm0 DB 131,255,1 ; cmp $0x1,%edi - DB 117,8 ; jne 5c2f <_sk_load_f32_sse2+0x87> + DB 117,8 ; jne 5c38 <_sk_load_f32_sse2+0x87> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,210 ; xorps %xmm2,%xmm2 - DB 235,184 ; jmp 5be7 <_sk_load_f32_sse2+0x3f> + DB 235,184 ; jmp 5bf0 <_sk_load_f32_sse2+0x3f> DB 15,16,84,178,16 ; movups 0x10(%edx,%esi,4),%xmm2 DB 131,255,3 ; cmp $0x3,%edi - DB 114,7 ; jb 5c40 <_sk_load_f32_sse2+0x98> + DB 114,7 ; jb 5c49 <_sk_load_f32_sse2+0x98> DB 15,16,92,178,32 ; movups 0x20(%edx,%esi,4),%xmm3 - DB 235,167 ; jmp 5be7 <_sk_load_f32_sse2+0x3f> + DB 235,167 ; jmp 5bf0 <_sk_load_f32_sse2+0x3f> DB 15,87,219 ; xorps %xmm3,%xmm3 - DB 235,162 ; jmp 5be7 <_sk_load_f32_sse2+0x3f> + DB 235,162 ; jmp 5bf0 <_sk_load_f32_sse2+0x3f> PUBLIC _sk_load_f32_dst_sse2 _sk_load_f32_dst_sse2 LABEL PROC @@ -52382,7 +52321,7 @@ _sk_load_f32_dst_sse2 LABEL PROC DB 15,16,52,23 ; movups (%edi,%edx,1),%xmm6 DB 139,121,8 ; mov 0x8(%ecx),%edi DB 133,255 ; test %edi,%edi - DB 117,97 ; jne 5ce1 <_sk_load_f32_dst_sse2+0x9c> + DB 117,97 ; jne 5cea <_sk_load_f32_dst_sse2+0x9c> DB 15,16,124,178,16 ; movups 0x10(%edx,%esi,4),%xmm7 DB 15,16,92,178,32 ; movups 0x20(%edx,%esi,4),%xmm3 DB 15,16,84,178,48 ; movups 0x30(%edx,%esi,4),%xmm2 @@ -52416,17 +52355,17 @@ _sk_load_f32_dst_sse2 LABEL PROC DB 195 ; ret DB 15,87,210 ; xorps %xmm2,%xmm2 DB 131,255,1 ; cmp $0x1,%edi - DB 117,8 ; jne 5cf1 <_sk_load_f32_dst_sse2+0xac> + DB 117,8 ; jne 5cfa <_sk_load_f32_dst_sse2+0xac> DB 15,87,219 ; xorps %xmm3,%xmm3 DB 15,87,255 ; xorps %xmm7,%xmm7 - DB 235,158 ; jmp 5c8f <_sk_load_f32_dst_sse2+0x4a> + DB 235,158 ; jmp 5c98 <_sk_load_f32_dst_sse2+0x4a> DB 15,16,124,178,16 ; movups 0x10(%edx,%esi,4),%xmm7 DB 131,255,3 ; cmp $0x3,%edi - DB 114,7 ; jb 5d02 <_sk_load_f32_dst_sse2+0xbd> + DB 114,7 ; jb 5d0b <_sk_load_f32_dst_sse2+0xbd> DB 15,16,92,178,32 ; movups 0x20(%edx,%esi,4),%xmm3 - DB 235,141 ; jmp 5c8f <_sk_load_f32_dst_sse2+0x4a> + DB 235,141 ; jmp 5c98 <_sk_load_f32_dst_sse2+0x4a> DB 15,87,219 ; xorps %xmm3,%xmm3 - DB 235,136 ; jmp 5c8f <_sk_load_f32_dst_sse2+0x4a> + DB 235,136 ; jmp 5c98 <_sk_load_f32_dst_sse2+0x4a> PUBLIC _sk_store_f32_sse2 _sk_store_f32_sse2 LABEL PROC @@ -52462,7 +52401,7 @@ _sk_store_f32_sse2 LABEL PROC DB 15,40,238 ; movaps %xmm6,%xmm5 DB 102,15,20,239 ; unpcklpd %xmm7,%xmm5 DB 133,255 ; test %edi,%edi - DB 117,40 ; jne 5d8d <_sk_store_f32_sse2+0x86> + DB 117,40 ; jne 5d96 <_sk_store_f32_sse2+0x86> DB 102,15,21,247 ; unpckhpd %xmm7,%xmm6 DB 15,17,100,178,16 ; movups %xmm4,0x10(%edx,%esi,4) DB 102,15,17,108,178,32 ; movupd %xmm5,0x20(%edx,%esi,4) @@ -52479,12 +52418,12 @@ _sk_store_f32_sse2 LABEL PROC DB 93 ; pop %ebp DB 195 ; ret DB 131,255,1 ; cmp $0x1,%edi - DB 116,232 ; je 5d7a <_sk_store_f32_sse2+0x73> + DB 116,232 ; je 5d83 <_sk_store_f32_sse2+0x73> DB 15,17,100,178,16 ; movups %xmm4,0x10(%edx,%esi,4) DB 131,255,3 ; cmp $0x3,%edi - DB 114,222 ; jb 5d7a <_sk_store_f32_sse2+0x73> + DB 114,222 ; jb 5d83 <_sk_store_f32_sse2+0x73> DB 102,15,17,108,178,32 ; movupd %xmm5,0x20(%edx,%esi,4) - DB 235,214 ; jmp 5d7a <_sk_store_f32_sse2+0x73> + DB 235,214 ; jmp 5d83 <_sk_store_f32_sse2+0x73> PUBLIC _sk_clamp_x_sse2 _sk_clamp_x_sse2 LABEL PROC @@ -52539,7 +52478,7 @@ _sk_repeat_x_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 5e21 <_sk_repeat_x_sse2+0xb> + DB 232,0,0,0,0 ; call 5e2a <_sk_repeat_x_sse2+0xb> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -52550,7 +52489,7 @@ _sk_repeat_x_sse2 LABEL PROC DB 243,15,91,230 ; cvttps2dq %xmm6,%xmm4 DB 15,91,252 ; cvtdq2ps %xmm4,%xmm7 DB 15,194,247,1 ; cmpltps %xmm7,%xmm6 - DB 15,84,177,47,114,0,0 ; andps 0x722f(%ecx),%xmm6 + DB 15,84,177,22,113,0,0 ; andps 0x7116(%ecx),%xmm6 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,92,254 ; subps %xmm6,%xmm7 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -52575,7 +52514,7 @@ _sk_repeat_y_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 5e87 <_sk_repeat_y_sse2+0xb> + DB 232,0,0,0,0 ; call 5e90 <_sk_repeat_y_sse2+0xb> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -52586,7 +52525,7 @@ _sk_repeat_y_sse2 LABEL PROC DB 243,15,91,230 ; cvttps2dq %xmm6,%xmm4 DB 15,91,252 ; cvtdq2ps %xmm4,%xmm7 DB 15,194,247,1 ; cmpltps %xmm7,%xmm6 - DB 15,84,177,201,113,0,0 ; andps 0x71c9(%ecx),%xmm6 + DB 15,84,177,176,112,0,0 ; andps 0x70b0(%ecx),%xmm6 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,92,254 ; subps %xmm6,%xmm7 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -52611,7 +52550,7 @@ _sk_mirror_x_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 5eed <_sk_mirror_x_sse2+0xb> + DB 232,0,0,0,0 ; call 5ef6 <_sk_mirror_x_sse2+0xb> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -52620,13 +52559,13 @@ _sk_mirror_x_sse2 LABEL PROC DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,198,228,0 ; shufps $0x0,%xmm4,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 243,15,89,177,179,120,0,0 ; mulss 0x78b3(%ecx),%xmm6 + DB 243,15,89,177,202,119,0,0 ; mulss 0x77ca(%ecx),%xmm6 DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 DB 15,89,240 ; mulps %xmm0,%xmm6 DB 243,15,91,254 ; cvttps2dq %xmm6,%xmm7 DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 DB 15,194,247,1 ; cmpltps %xmm7,%xmm6 - DB 15,84,177,99,113,0,0 ; andps 0x7163(%ecx),%xmm6 + DB 15,84,177,74,112,0,0 ; andps 0x704a(%ecx),%xmm6 DB 15,92,254 ; subps %xmm6,%xmm7 DB 243,15,88,237 ; addss %xmm5,%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -52655,7 +52594,7 @@ _sk_mirror_y_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 5f72 <_sk_mirror_y_sse2+0xb> + DB 232,0,0,0,0 ; call 5f7b <_sk_mirror_y_sse2+0xb> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -52664,13 +52603,13 @@ _sk_mirror_y_sse2 LABEL PROC DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,198,228,0 ; shufps $0x0,%xmm4,%xmm4 DB 15,92,204 ; subps %xmm4,%xmm1 - DB 243,15,89,177,46,120,0,0 ; mulss 0x782e(%ecx),%xmm6 + DB 243,15,89,177,69,119,0,0 ; mulss 0x7745(%ecx),%xmm6 DB 15,198,246,0 ; shufps $0x0,%xmm6,%xmm6 DB 15,89,241 ; mulps %xmm1,%xmm6 DB 243,15,91,254 ; cvttps2dq %xmm6,%xmm7 DB 15,91,255 ; cvtdq2ps %xmm7,%xmm7 DB 15,194,247,1 ; cmpltps %xmm7,%xmm6 - DB 15,84,177,222,112,0,0 ; andps 0x70de(%ecx),%xmm6 + DB 15,84,177,197,111,0,0 ; andps 0x6fc5(%ecx),%xmm6 DB 15,92,254 ; subps %xmm6,%xmm7 DB 243,15,88,237 ; addss %xmm5,%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 @@ -52699,12 +52638,12 @@ _sk_clamp_x_1_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 5ff7 <_sk_clamp_x_1_sse2+0xb> + DB 232,0,0,0,0 ; call 6000 <_sk_clamp_x_1_sse2+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,95,224 ; maxps %xmm0,%xmm4 - DB 15,93,160,89,112,0,0 ; minps 0x7059(%eax),%xmm4 + DB 15,93,160,64,111,0,0 ; minps 0x6f40(%eax),%xmm4 DB 141,65,4 ; lea 0x4(%ecx),%eax DB 131,236,8 ; sub $0x8,%esp DB 15,40,196 ; movaps %xmm4,%xmm0 @@ -52720,14 +52659,14 @@ _sk_repeat_x_1_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 6027 <_sk_repeat_x_1_sse2+0xb> + DB 232,0,0,0,0 ; call 6030 <_sk_repeat_x_1_sse2+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 243,15,91,224 ; cvttps2dq %xmm0,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,40,232 ; movaps %xmm0,%xmm5 DB 15,194,236,1 ; cmpltps %xmm4,%xmm5 - DB 15,84,168,41,112,0,0 ; andps 0x7029(%eax),%xmm5 + DB 15,84,168,16,111,0,0 ; andps 0x6f10(%eax),%xmm5 DB 15,92,229 ; subps %xmm5,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 DB 141,65,4 ; lea 0x4(%ecx),%eax @@ -52744,17 +52683,17 @@ _sk_mirror_x_1_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 6062 <_sk_mirror_x_1_sse2+0xb> + DB 232,0,0,0,0 ; call 606b <_sk_mirror_x_1_sse2+0xb> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 15,40,161,94,112,0,0 ; movaps 0x705e(%ecx),%xmm4 + DB 15,40,161,69,111,0,0 ; movaps 0x6f45(%ecx),%xmm4 DB 15,88,196 ; addps %xmm4,%xmm0 - DB 15,40,169,222,111,0,0 ; movaps 0x6fde(%ecx),%xmm5 + DB 15,40,169,197,110,0,0 ; movaps 0x6ec5(%ecx),%xmm5 DB 15,89,232 ; mulps %xmm0,%xmm5 DB 243,15,91,245 ; cvttps2dq %xmm5,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 DB 15,194,238,1 ; cmpltps %xmm6,%xmm5 - DB 15,84,169,238,111,0,0 ; andps 0x6fee(%ecx),%xmm5 + DB 15,84,169,213,110,0,0 ; andps 0x6ed5(%ecx),%xmm5 DB 15,87,255 ; xorps %xmm7,%xmm7 DB 15,92,245 ; subps %xmm5,%xmm6 DB 15,88,246 ; addps %xmm6,%xmm6 @@ -52777,13 +52716,13 @@ _sk_luminance_to_alpha_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 232,0,0,0,0 ; call 60c0 <_sk_luminance_to_alpha_sse2+0xe> + DB 232,0,0,0,0 ; call 60c9 <_sk_luminance_to_alpha_sse2+0xe> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 15,89,128,144,116,0,0 ; mulps 0x7490(%eax),%xmm0 - DB 15,89,136,160,116,0,0 ; mulps 0x74a0(%eax),%xmm1 + DB 15,89,128,167,115,0,0 ; mulps 0x73a7(%eax),%xmm0 + DB 15,89,136,183,115,0,0 ; mulps 0x73b7(%eax),%xmm1 DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,89,152,176,116,0,0 ; mulps 0x74b0(%eax),%xmm3 + DB 15,89,152,199,115,0,0 ; mulps 0x73c7(%eax),%xmm3 DB 15,88,217 ; addps %xmm1,%xmm3 DB 141,65,4 ; lea 0x4(%ecx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -53157,7 +53096,7 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 86 ; push %esi DB 131,236,60 ; sub $0x3c,%esp DB 15,41,69,184 ; movaps %xmm0,-0x48(%ebp) - DB 232,0,0,0,0 ; call 655d <_sk_evenly_spaced_gradient_sse2+0x12> + DB 232,0,0,0,0 ; call 6566 <_sk_evenly_spaced_gradient_sse2+0x12> DB 90 ; pop %edx DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 141,65,4 ; lea 0x4(%ecx),%eax @@ -53168,18 +53107,18 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 78 ; dec %esi DB 102,15,110,206 ; movd %esi,%xmm1 DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1 - DB 102,15,111,146,35,112,0,0 ; movdqa 0x7023(%edx),%xmm2 + DB 102,15,111,146,58,111,0,0 ; movdqa 0x6f3a(%edx),%xmm2 DB 102,15,219,209 ; pand %xmm1,%xmm2 - DB 102,15,235,146,3,110,0,0 ; por 0x6e03(%edx),%xmm2 + DB 102,15,235,146,250,108,0,0 ; por 0x6cfa(%edx),%xmm2 DB 102,15,114,209,16 ; psrld $0x10,%xmm1 - DB 102,15,235,138,51,112,0,0 ; por 0x7033(%edx),%xmm1 - DB 15,88,138,67,112,0,0 ; addps 0x7043(%edx),%xmm1 + DB 102,15,235,138,74,111,0,0 ; por 0x6f4a(%edx),%xmm1 + DB 15,88,138,90,111,0,0 ; addps 0x6f5a(%edx),%xmm1 DB 15,88,202 ; addps %xmm2,%xmm1 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 - DB 102,15,126,210 ; movd %xmm2,%edx - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,218 ; movd %xmm3,%edx DB 102,15,112,217,231 ; pshufd $0xe7,%xmm1,%xmm3 DB 102,15,126,223 ; movd %xmm3,%edi DB 243,15,16,28,184 ; movss (%eax,%edi,4),%xmm3 @@ -53195,7 +53134,7 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,4,176 ; movss (%eax,%esi,4),%xmm0 DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 DB 15,20,225 ; unpcklps %xmm1,%xmm4 - DB 15,20,230 ; unpcklps %xmm6,%xmm4 + DB 102,15,20,230 ; unpcklpd %xmm6,%xmm4 DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 15,20,199 ; unpcklps %xmm7,%xmm0 DB 139,65,8 ; mov 0x8(%ecx),%eax @@ -53203,11 +53142,11 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,52,144 ; movss (%eax,%edx,4),%xmm6 DB 243,15,16,12,176 ; movss (%eax,%esi,4),%xmm1 DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,41,69,200 ; movaps %xmm0,-0x38(%ebp) + DB 102,15,20,195 ; unpcklpd %xmm3,%xmm0 + DB 102,15,41,69,200 ; movapd %xmm0,-0x38(%ebp) DB 15,20,242 ; unpcklps %xmm2,%xmm6 DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 DB 139,65,24 ; mov 0x18(%ecx),%eax DB 243,15,16,20,184 ; movss (%eax,%edi,4),%xmm2 DB 243,15,16,28,144 ; movss (%eax,%edx,4),%xmm3 @@ -53215,8 +53154,8 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) + DB 102,15,20,195 ; unpcklpd %xmm3,%xmm0 + DB 102,15,41,69,216 ; movapd %xmm0,-0x28(%ebp) DB 139,65,12 ; mov 0xc(%ecx),%eax DB 243,15,16,20,184 ; movss (%eax,%edi,4),%xmm2 DB 243,15,16,28,144 ; movss (%eax,%edx,4),%xmm3 @@ -53224,7 +53163,7 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,20,176 ; movss (%eax,%esi,4),%xmm2 DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 139,65,28 ; mov 0x1c(%ecx),%eax DB 243,15,16,60,184 ; movss (%eax,%edi,4),%xmm7 DB 243,15,16,28,144 ; movss (%eax,%edx,4),%xmm3 @@ -53232,7 +53171,7 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,60,176 ; movss (%eax,%esi,4),%xmm7 DB 243,15,16,44,152 ; movss (%eax,%ebx,4),%xmm5 DB 15,20,253 ; unpcklps %xmm5,%xmm7 - DB 15,20,251 ; unpcklps %xmm3,%xmm7 + DB 102,15,20,251 ; unpcklpd %xmm3,%xmm7 DB 139,65,16 ; mov 0x10(%ecx),%eax DB 243,15,16,28,184 ; movss (%eax,%edi,4),%xmm3 DB 243,15,16,44,144 ; movss (%eax,%edx,4),%xmm5 @@ -53240,7 +53179,7 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,28,176 ; movss (%eax,%esi,4),%xmm3 DB 243,15,16,4,152 ; movss (%eax,%ebx,4),%xmm0 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 + DB 102,15,20,221 ; unpcklpd %xmm5,%xmm3 DB 139,65,32 ; mov 0x20(%ecx),%eax DB 243,15,16,4,184 ; movss (%eax,%edi,4),%xmm0 DB 243,15,16,44,144 ; movss (%eax,%edx,4),%xmm5 @@ -53248,7 +53187,7 @@ _sk_evenly_spaced_gradient_sse2 LABEL PROC DB 243,15,16,4,176 ; movss (%eax,%esi,4),%xmm0 DB 243,15,16,52,152 ; movss (%eax,%ebx,4),%xmm6 DB 15,20,198 ; unpcklps %xmm6,%xmm0 - DB 15,20,197 ; unpcklps %xmm5,%xmm0 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 DB 15,40,109,184 ; movaps -0x48(%ebp),%xmm5 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,88,101,200 ; addps -0x38(%ebp),%xmm4 @@ -53278,18 +53217,18 @@ _sk_gauss_a_to_rgba_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 6713 <_sk_gauss_a_to_rgba_sse2+0xb> + DB 232,0,0,0,0 ; call 6726 <_sk_gauss_a_to_rgba_sse2+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 15,40,128,157,110,0,0 ; movaps 0x6e9d(%eax),%xmm0 + DB 15,40,128,170,109,0,0 ; movaps 0x6daa(%eax),%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,128,173,110,0,0 ; addps 0x6ead(%eax),%xmm0 + DB 15,88,128,186,109,0,0 ; addps 0x6dba(%eax),%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,128,189,110,0,0 ; addps 0x6ebd(%eax),%xmm0 + DB 15,88,128,202,109,0,0 ; addps 0x6dca(%eax),%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,128,205,110,0,0 ; addps 0x6ecd(%eax),%xmm0 + DB 15,88,128,218,109,0,0 ; addps 0x6dda(%eax),%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,128,221,110,0,0 ; addps 0x6edd(%eax),%xmm0 + DB 15,88,128,234,109,0,0 ; addps 0x6dea(%eax),%xmm0 DB 141,65,4 ; lea 0x4(%ecx),%eax DB 131,236,8 ; sub $0x8,%esp DB 15,40,200 ; movaps %xmm0,%xmm1 @@ -53310,8 +53249,6 @@ _sk_gradient_sse2 LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,60 ; sub $0x3c,%esp - DB 232,0,0,0,0 ; call 676e <_sk_gradient_sse2+0xe> - DB 95 ; pop %edi DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 141,81,4 ; lea 0x4(%ecx),%edx DB 137,85,240 ; mov %edx,-0x10(%ebp) @@ -53319,24 +53256,22 @@ _sk_gradient_sse2 LABEL PROC DB 139,17 ; mov (%ecx),%edx DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 131,250,2 ; cmp $0x2,%edx - DB 114,43 ; jb 67b0 <_sk_gradient_sse2+0x50> + DB 114,33 ; jb 67b3 <_sk_gradient_sse2+0x40> DB 139,113,36 ; mov 0x24(%ecx),%esi DB 74 ; dec %edx DB 131,198,4 ; add $0x4,%esi DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 15,40,151,2,105,0,0 ; movaps 0x6902(%edi),%xmm2 - DB 243,15,16,30 ; movss (%esi),%xmm3 - DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3 - DB 15,194,216,2 ; cmpleps %xmm0,%xmm3 - DB 15,84,218 ; andps %xmm2,%xmm3 - DB 102,15,254,203 ; paddd %xmm3,%xmm1 + DB 243,15,16,22 ; movss (%esi),%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,194,208,2 ; cmpleps %xmm0,%xmm2 + DB 102,15,250,202 ; psubd %xmm2,%xmm1 DB 131,198,4 ; add $0x4,%esi DB 74 ; dec %edx - DB 117,231 ; jne 6797 <_sk_gradient_sse2+0x37> + DB 117,234 ; jne 679d <_sk_gradient_sse2+0x2a> DB 102,15,126,202 ; movd %xmm1,%edx DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,207 ; movd %xmm1,%edi DB 139,65,4 ; mov 0x4(%ecx),%eax @@ -53352,8 +53287,8 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,44,144 ; movss (%eax,%edx,4),%xmm5 DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 DB 15,20,226 ; unpcklps %xmm2,%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 15,41,101,184 ; movaps %xmm4,-0x48(%ebp) + DB 102,15,20,227 ; unpcklpd %xmm3,%xmm4 + DB 102,15,41,101,184 ; movapd %xmm4,-0x48(%ebp) DB 15,20,241 ; unpcklps %xmm1,%xmm6 DB 15,20,239 ; unpcklps %xmm7,%xmm5 DB 139,65,8 ; mov 0x8(%ecx),%eax @@ -53361,11 +53296,11 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,28,176 ; movss (%eax,%esi,4),%xmm3 DB 243,15,16,12,144 ; movss (%eax,%edx,4),%xmm1 DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 15,41,109,200 ; movaps %xmm5,-0x38(%ebp) + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 + DB 102,15,41,109,200 ; movapd %xmm5,-0x38(%ebp) DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 15,20,203 ; unpcklps %xmm3,%xmm1 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 DB 139,65,24 ; mov 0x18(%ecx),%eax DB 243,15,16,20,184 ; movss (%eax,%edi,4),%xmm2 DB 243,15,16,28,176 ; movss (%eax,%esi,4),%xmm3 @@ -53373,8 +53308,8 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 15,20,231 ; unpcklps %xmm7,%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 15,41,101,216 ; movaps %xmm4,-0x28(%ebp) + DB 102,15,20,227 ; unpcklpd %xmm3,%xmm4 + DB 102,15,41,101,216 ; movapd %xmm4,-0x28(%ebp) DB 139,65,12 ; mov 0xc(%ecx),%eax DB 243,15,16,20,184 ; movss (%eax,%edi,4),%xmm2 DB 243,15,16,28,176 ; movss (%eax,%esi,4),%xmm3 @@ -53382,7 +53317,7 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,20,144 ; movss (%eax,%edx,4),%xmm2 DB 243,15,16,60,152 ; movss (%eax,%ebx,4),%xmm7 DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 139,65,28 ; mov 0x1c(%ecx),%eax DB 243,15,16,60,184 ; movss (%eax,%edi,4),%xmm7 DB 243,15,16,28,176 ; movss (%eax,%esi,4),%xmm3 @@ -53390,7 +53325,7 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,60,144 ; movss (%eax,%edx,4),%xmm7 DB 243,15,16,44,152 ; movss (%eax,%ebx,4),%xmm5 DB 15,20,253 ; unpcklps %xmm5,%xmm7 - DB 15,20,251 ; unpcklps %xmm3,%xmm7 + DB 102,15,20,251 ; unpcklpd %xmm3,%xmm7 DB 139,65,16 ; mov 0x10(%ecx),%eax DB 243,15,16,28,184 ; movss (%eax,%edi,4),%xmm3 DB 243,15,16,44,176 ; movss (%eax,%esi,4),%xmm5 @@ -53398,7 +53333,7 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,28,144 ; movss (%eax,%edx,4),%xmm3 DB 243,15,16,52,152 ; movss (%eax,%ebx,4),%xmm6 DB 15,20,222 ; unpcklps %xmm6,%xmm3 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 + DB 102,15,20,221 ; unpcklpd %xmm5,%xmm3 DB 139,65,32 ; mov 0x20(%ecx),%eax DB 243,15,16,44,184 ; movss (%eax,%edi,4),%xmm5 DB 243,15,16,52,176 ; movss (%eax,%esi,4),%xmm6 @@ -53406,7 +53341,7 @@ _sk_gradient_sse2 LABEL PROC DB 243,15,16,44,144 ; movss (%eax,%edx,4),%xmm5 DB 243,15,16,36,152 ; movss (%eax,%ebx,4),%xmm4 DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 DB 15,40,101,184 ; movaps -0x48(%ebp),%xmm4 DB 15,89,224 ; mulps %xmm0,%xmm4 DB 15,88,101,200 ; addps -0x38(%ebp),%xmm4 @@ -53480,7 +53415,7 @@ _sk_xy_to_unit_angle_sse2 LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 15,40,209 ; movaps %xmm1,%xmm2 DB 15,40,200 ; movaps %xmm0,%xmm1 - DB 232,0,0,0,0 ; call 69aa <_sk_xy_to_unit_angle_sse2+0x18> + DB 232,0,0,0,0 ; call 69b8 <_sk_xy_to_unit_angle_sse2+0x18> DB 88 ; pop %eax DB 15,87,237 ; xorps %xmm5,%xmm5 DB 15,92,233 ; subps %xmm1,%xmm5 @@ -53495,30 +53430,30 @@ _sk_xy_to_unit_angle_sse2 LABEL PROC DB 15,94,247 ; divps %xmm7,%xmm6 DB 15,40,254 ; movaps %xmm6,%xmm7 DB 15,89,255 ; mulps %xmm7,%xmm7 - DB 15,40,128,86,108,0,0 ; movaps 0x6c56(%eax),%xmm0 + DB 15,40,128,104,107,0,0 ; movaps 0x6b68(%eax),%xmm0 DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 15,88,128,102,108,0,0 ; addps 0x6c66(%eax),%xmm0 + DB 15,88,128,120,107,0,0 ; addps 0x6b78(%eax),%xmm0 DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 15,88,128,118,108,0,0 ; addps 0x6c76(%eax),%xmm0 + DB 15,88,128,136,107,0,0 ; addps 0x6b88(%eax),%xmm0 DB 15,89,199 ; mulps %xmm7,%xmm0 - DB 15,88,128,134,108,0,0 ; addps 0x6c86(%eax),%xmm0 + DB 15,88,128,152,107,0,0 ; addps 0x6b98(%eax),%xmm0 DB 15,89,198 ; mulps %xmm6,%xmm0 DB 15,194,236,1 ; cmpltps %xmm4,%xmm5 - DB 15,40,176,150,108,0,0 ; movaps 0x6c96(%eax),%xmm6 + DB 15,40,176,168,107,0,0 ; movaps 0x6ba8(%eax),%xmm6 DB 15,92,240 ; subps %xmm0,%xmm6 DB 15,84,245 ; andps %xmm5,%xmm6 DB 15,85,232 ; andnps %xmm0,%xmm5 DB 15,87,228 ; xorps %xmm4,%xmm4 DB 15,86,238 ; orps %xmm6,%xmm5 DB 15,194,204,1 ; cmpltps %xmm4,%xmm1 - DB 15,40,128,150,102,0,0 ; movaps 0x6696(%eax),%xmm0 + DB 15,40,128,120,101,0,0 ; movaps 0x6578(%eax),%xmm0 DB 15,92,197 ; subps %xmm5,%xmm0 DB 15,84,193 ; andps %xmm1,%xmm0 DB 15,85,205 ; andnps %xmm5,%xmm1 DB 15,86,200 ; orps %xmm0,%xmm1 DB 15,40,194 ; movaps %xmm2,%xmm0 DB 15,194,196,1 ; cmpltps %xmm4,%xmm0 - DB 15,40,168,166,102,0,0 ; movaps 0x66a6(%eax),%xmm5 + DB 15,40,168,136,101,0,0 ; movaps 0x6588(%eax),%xmm5 DB 15,92,233 ; subps %xmm1,%xmm5 DB 15,84,232 ; andps %xmm0,%xmm5 DB 15,85,193 ; andnps %xmm1,%xmm0 @@ -53575,15 +53510,15 @@ _sk_xy_to_2pt_conical_quadratic_max_sse2 LABEL PROC DB 15,40,241 ; movaps %xmm1,%xmm6 DB 15,89,246 ; mulps %xmm6,%xmm6 DB 15,88,240 ; addps %xmm0,%xmm6 - DB 232,0,0,0,0 ; call 6ac7 <_sk_xy_to_2pt_conical_quadratic_max_sse2+0x31> + DB 232,0,0,0,0 ; call 6ad5 <_sk_xy_to_2pt_conical_quadratic_max_sse2+0x31> DB 90 ; pop %edx DB 243,15,89,237 ; mulss %xmm5,%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,92,245 ; subps %xmm5,%xmm6 DB 243,15,16,65,32 ; movss 0x20(%ecx),%xmm0 DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0 - DB 15,89,162,137,107,0,0 ; mulps 0x6b89(%edx),%xmm4 - DB 15,89,130,153,107,0,0 ; mulps 0x6b99(%edx),%xmm0 + DB 15,89,162,155,106,0,0 ; mulps 0x6a9b(%edx),%xmm4 + DB 15,89,130,171,106,0,0 ; mulps 0x6aab(%edx),%xmm0 DB 15,89,198 ; mulps %xmm6,%xmm0 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,237 ; mulps %xmm5,%xmm5 @@ -53591,11 +53526,11 @@ _sk_xy_to_2pt_conical_quadratic_max_sse2 LABEL PROC DB 15,81,237 ; sqrtps %xmm5,%xmm5 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,87,162,57,106,0,0 ; xorps 0x6a39(%edx),%xmm4 + DB 15,87,162,75,105,0,0 ; xorps 0x694b(%edx),%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 243,15,16,105,36 ; movss 0x24(%ecx),%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 - DB 15,89,170,121,101,0,0 ; mulps 0x6579(%edx),%xmm5 + DB 15,89,170,91,100,0,0 ; mulps 0x645b(%edx),%xmm5 DB 15,89,197 ; mulps %xmm5,%xmm0 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,95,196 ; maxps %xmm4,%xmm0 @@ -53624,15 +53559,15 @@ _sk_xy_to_2pt_conical_quadratic_min_sse2 LABEL PROC DB 15,40,241 ; movaps %xmm1,%xmm6 DB 15,89,246 ; mulps %xmm6,%xmm6 DB 15,88,240 ; addps %xmm0,%xmm6 - DB 232,0,0,0,0 ; call 6b65 <_sk_xy_to_2pt_conical_quadratic_min_sse2+0x31> + DB 232,0,0,0,0 ; call 6b73 <_sk_xy_to_2pt_conical_quadratic_min_sse2+0x31> DB 90 ; pop %edx DB 243,15,89,237 ; mulss %xmm5,%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,92,245 ; subps %xmm5,%xmm6 DB 243,15,16,65,32 ; movss 0x20(%ecx),%xmm0 DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0 - DB 15,89,162,235,106,0,0 ; mulps 0x6aeb(%edx),%xmm4 - DB 15,89,130,251,106,0,0 ; mulps 0x6afb(%edx),%xmm0 + DB 15,89,162,253,105,0,0 ; mulps 0x69fd(%edx),%xmm4 + DB 15,89,130,13,106,0,0 ; mulps 0x6a0d(%edx),%xmm0 DB 15,89,198 ; mulps %xmm6,%xmm0 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,237 ; mulps %xmm5,%xmm5 @@ -53640,11 +53575,11 @@ _sk_xy_to_2pt_conical_quadratic_min_sse2 LABEL PROC DB 15,81,237 ; sqrtps %xmm5,%xmm5 DB 15,40,197 ; movaps %xmm5,%xmm0 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,87,162,155,105,0,0 ; xorps 0x699b(%edx),%xmm4 + DB 15,87,162,173,104,0,0 ; xorps 0x68ad(%edx),%xmm4 DB 15,92,229 ; subps %xmm5,%xmm4 DB 243,15,16,105,36 ; movss 0x24(%ecx),%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 - DB 15,89,170,219,100,0,0 ; mulps 0x64db(%edx),%xmm5 + DB 15,89,170,189,99,0,0 ; mulps 0x63bd(%edx),%xmm5 DB 15,89,197 ; mulps %xmm5,%xmm0 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,93,196 ; minps %xmm4,%xmm0 @@ -53662,7 +53597,7 @@ _sk_xy_to_2pt_conical_linear_sse2 LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call 6bdd <_sk_xy_to_2pt_conical_linear_sse2+0xb> + DB 232,0,0,0,0 ; call 6beb <_sk_xy_to_2pt_conical_linear_sse2+0xb> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -53671,7 +53606,7 @@ _sk_xy_to_2pt_conical_linear_sse2 LABEL PROC DB 243,15,89,236 ; mulss %xmm4,%xmm5 DB 15,198,237,0 ; shufps $0x0,%xmm5,%xmm5 DB 15,88,232 ; addps %xmm0,%xmm5 - DB 15,89,169,115,106,0,0 ; mulps 0x6a73(%ecx),%xmm5 + DB 15,89,169,133,105,0,0 ; mulps 0x6985(%ecx),%xmm5 DB 15,89,192 ; mulps %xmm0,%xmm0 DB 15,40,241 ; movaps %xmm1,%xmm6 DB 15,89,246 ; mulps %xmm6,%xmm6 @@ -53679,7 +53614,7 @@ _sk_xy_to_2pt_conical_linear_sse2 LABEL PROC DB 243,15,89,228 ; mulss %xmm4,%xmm4 DB 15,198,228,0 ; shufps $0x0,%xmm4,%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,87,129,35,105,0,0 ; xorps 0x6923(%ecx),%xmm0 + DB 15,87,129,53,104,0,0 ; xorps 0x6835(%ecx),%xmm0 DB 15,94,197 ; divps %xmm5,%xmm0 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 131,236,8 ; sub $0x8,%esp @@ -53747,16 +53682,16 @@ _sk_save_xy_sse2 LABEL PROC DB 131,236,24 ; sub $0x18,%esp DB 15,41,93,232 ; movaps %xmm3,-0x18(%ebp) DB 15,40,218 ; movaps %xmm2,%xmm3 - DB 232,0,0,0,0 ; call 6cba <_sk_save_xy_sse2+0x12> + DB 232,0,0,0,0 ; call 6cc8 <_sk_save_xy_sse2+0x12> DB 88 ; pop %eax - DB 15,40,160,134,99,0,0 ; movaps 0x6386(%eax),%xmm4 + DB 15,40,160,104,98,0,0 ; movaps 0x6268(%eax),%xmm4 DB 15,40,232 ; movaps %xmm0,%xmm5 DB 15,88,236 ; addps %xmm4,%xmm5 DB 243,15,91,245 ; cvttps2dq %xmm5,%xmm6 DB 15,91,246 ; cvtdq2ps %xmm6,%xmm6 DB 15,40,253 ; movaps %xmm5,%xmm7 DB 15,194,254,1 ; cmpltps %xmm6,%xmm7 - DB 15,40,144,150,99,0,0 ; movaps 0x6396(%eax),%xmm2 + DB 15,40,144,120,98,0,0 ; movaps 0x6278(%eax),%xmm2 DB 15,84,250 ; andps %xmm2,%xmm7 DB 15,92,247 ; subps %xmm7,%xmm6 DB 15,92,238 ; subps %xmm6,%xmm5 @@ -53827,15 +53762,15 @@ _sk_bilinear_nx_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6d9c <_sk_bilinear_nx_sse2+0xa> + DB 232,0,0,0,0 ; call 6daa <_sk_bilinear_nx_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,6 ; movups (%esi),%xmm0 DB 15,16,102,64 ; movups 0x40(%esi),%xmm4 - DB 15,88,128,212,104,0,0 ; addps 0x68d4(%eax),%xmm0 - DB 15,40,168,180,98,0,0 ; movaps 0x62b4(%eax),%xmm5 + DB 15,88,128,230,103,0,0 ; addps 0x67e6(%eax),%xmm0 + DB 15,40,168,150,97,0,0 ; movaps 0x6196(%eax),%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 DB 15,17,174,128,0,0,0 ; movups %xmm5,0x80(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax @@ -53854,14 +53789,14 @@ _sk_bilinear_px_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6ddf <_sk_bilinear_px_sse2+0xa> + DB 232,0,0,0,0 ; call 6ded <_sk_bilinear_px_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,6 ; movups (%esi),%xmm0 DB 15,16,102,64 ; movups 0x40(%esi),%xmm4 - DB 15,88,128,97,98,0,0 ; addps 0x6261(%eax),%xmm0 + DB 15,88,128,67,97,0,0 ; addps 0x6143(%eax),%xmm0 DB 15,17,166,128,0,0,0 ; movups %xmm4,0x80(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -53879,15 +53814,15 @@ _sk_bilinear_ny_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6e18 <_sk_bilinear_ny_sse2+0xa> + DB 232,0,0,0,0 ; call 6e26 <_sk_bilinear_ny_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,78,32 ; movups 0x20(%esi),%xmm1 DB 15,16,102,96 ; movups 0x60(%esi),%xmm4 - DB 15,88,136,88,104,0,0 ; addps 0x6858(%eax),%xmm1 - DB 15,40,168,56,98,0,0 ; movaps 0x6238(%eax),%xmm5 + DB 15,88,136,106,103,0,0 ; addps 0x676a(%eax),%xmm1 + DB 15,40,168,26,97,0,0 ; movaps 0x611a(%eax),%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 DB 15,17,174,160,0,0,0 ; movups %xmm5,0xa0(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax @@ -53906,14 +53841,14 @@ _sk_bilinear_py_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6e5c <_sk_bilinear_py_sse2+0xa> + DB 232,0,0,0,0 ; call 6e6a <_sk_bilinear_py_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,78,32 ; movups 0x20(%esi),%xmm1 DB 15,16,102,96 ; movups 0x60(%esi),%xmm4 - DB 15,88,136,228,97,0,0 ; addps 0x61e4(%eax),%xmm1 + DB 15,88,136,198,96,0,0 ; addps 0x60c6(%eax),%xmm1 DB 15,17,166,160,0,0,0 ; movups %xmm4,0xa0(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -53931,20 +53866,20 @@ _sk_bicubic_n3x_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6e96 <_sk_bicubic_n3x_sse2+0xa> + DB 232,0,0,0,0 ; call 6ea4 <_sk_bicubic_n3x_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,6 ; movups (%esi),%xmm0 DB 15,16,102,64 ; movups 0x40(%esi),%xmm4 - DB 15,88,128,234,103,0,0 ; addps 0x67ea(%eax),%xmm0 - DB 15,40,168,186,97,0,0 ; movaps 0x61ba(%eax),%xmm5 + DB 15,88,128,252,102,0,0 ; addps 0x66fc(%eax),%xmm0 + DB 15,40,168,156,96,0,0 ; movaps 0x609c(%eax),%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,89,228 ; mulps %xmm4,%xmm4 - DB 15,89,168,250,103,0,0 ; mulps 0x67fa(%eax),%xmm5 - DB 15,88,168,154,99,0,0 ; addps 0x639a(%eax),%xmm5 + DB 15,89,168,12,103,0,0 ; mulps 0x670c(%eax),%xmm5 + DB 15,88,168,124,98,0,0 ; addps 0x627c(%eax),%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 DB 15,17,174,128,0,0,0 ; movups %xmm5,0x80(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax @@ -53963,23 +53898,23 @@ _sk_bicubic_n1x_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6ef0 <_sk_bicubic_n1x_sse2+0xa> + DB 232,0,0,0,0 ; call 6efe <_sk_bicubic_n1x_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,6 ; movups (%esi),%xmm0 DB 15,16,102,64 ; movups 0x40(%esi),%xmm4 - DB 15,88,128,128,103,0,0 ; addps 0x6780(%eax),%xmm0 - DB 15,40,168,96,97,0,0 ; movaps 0x6160(%eax),%xmm5 + DB 15,88,128,146,102,0,0 ; addps 0x6692(%eax),%xmm0 + DB 15,40,168,66,96,0,0 ; movaps 0x6042(%eax),%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 - DB 15,40,160,176,103,0,0 ; movaps 0x67b0(%eax),%xmm4 + DB 15,40,160,194,102,0,0 ; movaps 0x66c2(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 - DB 15,88,160,192,103,0,0 ; addps 0x67c0(%eax),%xmm4 + DB 15,88,160,210,102,0,0 ; addps 0x66d2(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 - DB 15,88,160,80,97,0,0 ; addps 0x6150(%eax),%xmm4 + DB 15,88,160,50,96,0,0 ; addps 0x6032(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 - DB 15,88,160,208,103,0,0 ; addps 0x67d0(%eax),%xmm4 + DB 15,88,160,226,102,0,0 ; addps 0x66e2(%eax),%xmm4 DB 15,17,166,128,0,0,0 ; movups %xmm4,0x80(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -53997,22 +53932,22 @@ _sk_bicubic_p1x_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6f58 <_sk_bicubic_p1x_sse2+0xa> + DB 232,0,0,0,0 ; call 6f66 <_sk_bicubic_p1x_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi - DB 15,40,160,232,96,0,0 ; movaps 0x60e8(%eax),%xmm4 + DB 15,40,160,202,95,0,0 ; movaps 0x5fca(%eax),%xmm4 DB 15,16,6 ; movups (%esi),%xmm0 DB 15,16,110,64 ; movups 0x40(%esi),%xmm5 DB 15,88,196 ; addps %xmm4,%xmm0 - DB 15,40,176,72,103,0,0 ; movaps 0x6748(%eax),%xmm6 + DB 15,40,176,90,102,0,0 ; movaps 0x665a(%eax),%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 - DB 15,88,176,88,103,0,0 ; addps 0x6758(%eax),%xmm6 + DB 15,88,176,106,102,0,0 ; addps 0x666a(%eax),%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 DB 15,88,244 ; addps %xmm4,%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 - DB 15,88,176,104,103,0,0 ; addps 0x6768(%eax),%xmm6 + DB 15,88,176,122,102,0,0 ; addps 0x667a(%eax),%xmm6 DB 15,17,182,128,0,0,0 ; movups %xmm6,0x80(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -54030,18 +53965,18 @@ _sk_bicubic_p3x_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 6fb5 <_sk_bicubic_p3x_sse2+0xa> + DB 232,0,0,0,0 ; call 6fc3 <_sk_bicubic_p3x_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,6 ; movups (%esi),%xmm0 DB 15,16,102,64 ; movups 0x40(%esi),%xmm4 - DB 15,88,128,251,102,0,0 ; addps 0x66fb(%eax),%xmm0 + DB 15,88,128,13,102,0,0 ; addps 0x660d(%eax),%xmm0 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,237 ; mulps %xmm5,%xmm5 - DB 15,89,160,219,102,0,0 ; mulps 0x66db(%eax),%xmm4 - DB 15,88,160,123,98,0,0 ; addps 0x627b(%eax),%xmm4 + DB 15,89,160,237,101,0,0 ; mulps 0x65ed(%eax),%xmm4 + DB 15,88,160,93,97,0,0 ; addps 0x615d(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,17,166,128,0,0,0 ; movups %xmm4,0x80(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax @@ -54060,20 +53995,20 @@ _sk_bicubic_n3y_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 7005 <_sk_bicubic_n3y_sse2+0xa> + DB 232,0,0,0,0 ; call 7013 <_sk_bicubic_n3y_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,78,32 ; movups 0x20(%esi),%xmm1 DB 15,16,102,96 ; movups 0x60(%esi),%xmm4 - DB 15,88,136,123,102,0,0 ; addps 0x667b(%eax),%xmm1 - DB 15,40,168,75,96,0,0 ; movaps 0x604b(%eax),%xmm5 + DB 15,88,136,141,101,0,0 ; addps 0x658d(%eax),%xmm1 + DB 15,40,168,45,95,0,0 ; movaps 0x5f2d(%eax),%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 DB 15,40,229 ; movaps %xmm5,%xmm4 DB 15,89,228 ; mulps %xmm4,%xmm4 - DB 15,89,168,139,102,0,0 ; mulps 0x668b(%eax),%xmm5 - DB 15,88,168,43,98,0,0 ; addps 0x622b(%eax),%xmm5 + DB 15,89,168,157,101,0,0 ; mulps 0x659d(%eax),%xmm5 + DB 15,88,168,13,97,0,0 ; addps 0x610d(%eax),%xmm5 DB 15,89,236 ; mulps %xmm4,%xmm5 DB 15,17,174,160,0,0,0 ; movups %xmm5,0xa0(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax @@ -54092,23 +54027,23 @@ _sk_bicubic_n1y_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 7060 <_sk_bicubic_n1y_sse2+0xa> + DB 232,0,0,0,0 ; call 706e <_sk_bicubic_n1y_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,78,32 ; movups 0x20(%esi),%xmm1 DB 15,16,102,96 ; movups 0x60(%esi),%xmm4 - DB 15,88,136,16,102,0,0 ; addps 0x6610(%eax),%xmm1 - DB 15,40,168,240,95,0,0 ; movaps 0x5ff0(%eax),%xmm5 + DB 15,88,136,34,101,0,0 ; addps 0x6522(%eax),%xmm1 + DB 15,40,168,210,94,0,0 ; movaps 0x5ed2(%eax),%xmm5 DB 15,92,236 ; subps %xmm4,%xmm5 - DB 15,40,160,64,102,0,0 ; movaps 0x6640(%eax),%xmm4 + DB 15,40,160,82,101,0,0 ; movaps 0x6552(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 - DB 15,88,160,80,102,0,0 ; addps 0x6650(%eax),%xmm4 + DB 15,88,160,98,101,0,0 ; addps 0x6562(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 - DB 15,88,160,224,95,0,0 ; addps 0x5fe0(%eax),%xmm4 + DB 15,88,160,194,94,0,0 ; addps 0x5ec2(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 - DB 15,88,160,96,102,0,0 ; addps 0x6660(%eax),%xmm4 + DB 15,88,160,114,101,0,0 ; addps 0x6572(%eax),%xmm4 DB 15,17,166,160,0,0,0 ; movups %xmm4,0xa0(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -54126,22 +54061,22 @@ _sk_bicubic_p1y_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 70c9 <_sk_bicubic_p1y_sse2+0xa> + DB 232,0,0,0,0 ; call 70d7 <_sk_bicubic_p1y_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi - DB 15,40,160,119,95,0,0 ; movaps 0x5f77(%eax),%xmm4 + DB 15,40,160,89,94,0,0 ; movaps 0x5e59(%eax),%xmm4 DB 15,16,78,32 ; movups 0x20(%esi),%xmm1 DB 15,16,110,96 ; movups 0x60(%esi),%xmm5 DB 15,88,204 ; addps %xmm4,%xmm1 - DB 15,40,176,215,101,0,0 ; movaps 0x65d7(%eax),%xmm6 + DB 15,40,176,233,100,0,0 ; movaps 0x64e9(%eax),%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 - DB 15,88,176,231,101,0,0 ; addps 0x65e7(%eax),%xmm6 + DB 15,88,176,249,100,0,0 ; addps 0x64f9(%eax),%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 DB 15,88,244 ; addps %xmm4,%xmm6 DB 15,89,245 ; mulps %xmm5,%xmm6 - DB 15,88,176,247,101,0,0 ; addps 0x65f7(%eax),%xmm6 + DB 15,88,176,9,101,0,0 ; addps 0x6509(%eax),%xmm6 DB 15,17,182,160,0,0,0 ; movups %xmm6,0xa0(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax DB 131,236,8 ; sub $0x8,%esp @@ -54159,18 +54094,18 @@ _sk_bicubic_p3y_sse2 LABEL PROC DB 137,229 ; mov %esp,%ebp DB 86 ; push %esi DB 80 ; push %eax - DB 232,0,0,0,0 ; call 7127 <_sk_bicubic_p3y_sse2+0xa> + DB 232,0,0,0,0 ; call 7135 <_sk_bicubic_p3y_sse2+0xa> DB 88 ; pop %eax DB 139,77,8 ; mov 0x8(%ebp),%ecx DB 139,85,12 ; mov 0xc(%ebp),%edx DB 139,50 ; mov (%edx),%esi DB 15,16,78,32 ; movups 0x20(%esi),%xmm1 DB 15,16,102,96 ; movups 0x60(%esi),%xmm4 - DB 15,88,136,137,101,0,0 ; addps 0x6589(%eax),%xmm1 + DB 15,88,136,155,100,0,0 ; addps 0x649b(%eax),%xmm1 DB 15,40,236 ; movaps %xmm4,%xmm5 DB 15,89,237 ; mulps %xmm5,%xmm5 - DB 15,89,160,105,101,0,0 ; mulps 0x6569(%eax),%xmm4 - DB 15,88,160,9,97,0,0 ; addps 0x6109(%eax),%xmm4 + DB 15,89,160,123,100,0,0 ; mulps 0x647b(%eax),%xmm4 + DB 15,88,160,235,95,0,0 ; addps 0x5feb(%eax),%xmm4 DB 15,89,229 ; mulps %xmm5,%xmm4 DB 15,17,166,160,0,0,0 ; movups %xmm4,0xa0(%esi) DB 141,66,8 ; lea 0x8(%edx),%eax @@ -54260,7 +54195,7 @@ _sk_clut_3D_sse2 LABEL PROC DB 15,40,216 ; movaps %xmm0,%xmm3 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx - DB 137,141,4,255,255,255 ; mov %ecx,-0xfc(%ebp) + DB 137,141,20,255,255,255 ; mov %ecx,-0xec(%ebp) DB 139,8 ; mov (%eax),%ecx DB 139,81,12 ; mov 0xc(%ecx),%edx DB 141,114,255 ; lea -0x1(%edx),%esi @@ -54269,26 +54204,28 @@ _sk_clut_3D_sse2 LABEL PROC DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,89,194 ; mulps %xmm2,%xmm0 DB 15,40,224 ; movaps %xmm0,%xmm4 + DB 15,41,165,232,254,255,255 ; movaps %xmm4,-0x118(%ebp) DB 102,15,110,194 ; movd %edx,%xmm0 DB 102,15,112,240,0 ; pshufd $0x0,%xmm0,%xmm6 - DB 102,15,127,117,136 ; movdqa %xmm6,-0x78(%ebp) + DB 102,15,127,117,152 ; movdqa %xmm6,-0x68(%ebp) DB 139,81,8 ; mov 0x8(%ecx),%edx DB 141,114,255 ; lea -0x1(%edx),%esi DB 102,15,110,214 ; movd %esi,%xmm2 DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 15,89,209 ; mulps %xmm1,%xmm2 - DB 15,41,149,56,255,255,255 ; movaps %xmm2,-0xc8(%ebp) + DB 15,41,149,88,255,255,255 ; movaps %xmm2,-0xa8(%ebp) DB 243,15,91,234 ; cvttps2dq %xmm2,%xmm5 - DB 102,15,127,173,8,255,255,255 ; movdqa %xmm5,-0xf8(%ebp) - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 102,15,244,206 ; pmuludq %xmm6,%xmm1 - DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 + DB 102,15,127,173,40,255,255,255 ; movdqa %xmm5,-0xd8(%ebp) DB 102,15,112,205,245 ; pshufd $0xf5,%xmm5,%xmm1 DB 102,15,244,206 ; pmuludq %xmm6,%xmm1 + DB 102,15,111,214 ; movdqa %xmm6,%xmm2 + DB 102,15,244,213 ; pmuludq %xmm5,%xmm2 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,98,209 ; punpckldq %xmm1,%xmm2 - DB 102,15,127,149,24,255,255,255 ; movdqa %xmm2,-0xe8(%ebp) + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,15,127,109,168 ; movdqa %xmm5,-0x58(%ebp) DB 102,15,110,202 ; movd %edx,%xmm1 DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 DB 139,81,4 ; mov 0x4(%ecx),%edx @@ -54297,73 +54234,68 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,41,69,200 ; movaps %xmm0,-0x38(%ebp) - DB 102,15,112,217,0 ; pshufd $0x0,%xmm1,%xmm3 - DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 - DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 - DB 102,15,112,233,232 ; pshufd $0xe8,%xmm1,%xmm5 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 + DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) + DB 102,15,112,209,0 ; pshufd $0x0,%xmm1,%xmm2 + DB 243,15,91,216 ; cvttps2dq %xmm0,%xmm3 + DB 102,15,127,93,184 ; movdqa %xmm3,-0x48(%ebp) DB 102,15,112,195,245 ; pshufd $0xf5,%xmm3,%xmm0 - DB 102,15,127,133,120,255,255,255 ; movdqa %xmm0,-0x88(%ebp) - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 + DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,233 ; punpckldq %xmm1,%xmm5 - DB 102,15,127,173,104,255,255,255 ; movdqa %xmm5,-0x98(%ebp) - DB 232,0,0,0,0 ; call 731b <_sk_clut_3D_sse2+0xff> + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,98,200 ; punpckldq %xmm0,%xmm1 + DB 102,15,127,77,136 ; movdqa %xmm1,-0x78(%ebp) + DB 232,0,0,0,0 ; call 7321 <_sk_clut_3D_sse2+0xf7> DB 90 ; pop %edx DB 243,15,91,196 ; cvttps2dq %xmm4,%xmm0 - DB 102,15,127,133,232,254,255,255 ; movdqa %xmm0,-0x118(%ebp) - DB 15,40,244 ; movaps %xmm4,%xmm6 - DB 15,41,181,216,254,255,255 ; movaps %xmm6,-0x128(%ebp) - DB 102,15,254,208 ; paddd %xmm0,%xmm2 - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 102,15,254,202 ; paddd %xmm2,%xmm1 - DB 102,15,111,130,197,99,0,0 ; movdqa 0x63c5(%edx),%xmm0 - DB 102,15,127,133,88,255,255,255 ; movdqa %xmm0,-0xa8(%ebp) - DB 102,15,112,225,245 ; pshufd $0xf5,%xmm1,%xmm4 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,244,224 ; pmuludq %xmm0,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,127,133,248,254,255,255 ; movdqa %xmm0,-0x108(%ebp) + DB 102,15,111,221 ; movdqa %xmm5,%xmm3 + DB 102,15,254,216 ; paddd %xmm0,%xmm3 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,254,195 ; paddd %xmm3,%xmm0 + DB 102,15,111,162,223,98,0,0 ; movdqa 0x62df(%edx),%xmm4 + DB 102,15,127,165,120,255,255,255 ; movdqa %xmm4,-0x88(%ebp) + DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 + DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 + DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 + DB 102,15,112,225,232 ; pshufd $0xe8,%xmm1,%xmm4 + DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 DB 102,15,98,204 ; punpckldq %xmm4,%xmm1 - DB 102,15,112,225,229 ; pshufd $0xe5,%xmm1,%xmm4 - DB 102,15,126,230 ; movd %xmm4,%esi - DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 - DB 102,15,126,231 ; movd %xmm4,%edi - DB 102,15,112,225,231 ; pshufd $0xe7,%xmm1,%xmm4 - DB 102,15,126,227 ; movd %xmm4,%ebx + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,112,193,229 ; pshufd $0xe5,%xmm1,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 102,15,112,193,231 ; pshufd $0xe7,%xmm1,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax DB 139,9 ; mov (%ecx),%ecx + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 DB 243,15,16,36,153 ; movss (%ecx,%ebx,4),%xmm4 - DB 243,15,16,60,177 ; movss (%ecx,%esi,4),%xmm7 - DB 15,20,252 ; unpcklps %xmm4,%xmm7 - DB 102,15,126,206 ; movd %xmm1,%esi - DB 243,15,16,4,177 ; movss (%ecx,%esi,4),%xmm0 - DB 243,15,16,36,185 ; movss (%ecx,%edi,4),%xmm4 - DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,41,69,168 ; movaps %xmm0,-0x58(%ebp) - DB 102,15,111,130,85,93,0,0 ; movdqa 0x5d55(%edx),%xmm0 - DB 102,15,127,133,72,255,255,255 ; movdqa %xmm0,-0xb8(%ebp) - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,254,224 ; paddd %xmm0,%xmm4 - DB 102,15,112,252,229 ; pshufd $0xe5,%xmm4,%xmm7 - DB 102,15,126,254 ; movd %xmm7,%esi - DB 102,15,112,252,78 ; pshufd $0x4e,%xmm4,%xmm7 - DB 102,15,126,255 ; movd %xmm7,%edi - DB 102,15,126,227 ; movd %xmm4,%ebx - DB 102,15,112,228,231 ; pshufd $0xe7,%xmm4,%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 + DB 243,15,16,44,177 ; movss (%ecx,%esi,4),%xmm5 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 102,15,20,236 ; unpcklpd %xmm4,%xmm5 + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,250,196 ; psubd %xmm4,%xmm0 + DB 102,15,112,224,229 ; pshufd $0xe5,%xmm0,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 - DB 243,15,16,4,177 ; movss (%ecx,%esi,4),%xmm0 - DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 243,15,16,60,153 ; movss (%ecx,%ebx,4),%xmm7 - DB 243,15,16,36,185 ; movss (%ecx,%edi,4),%xmm4 - DB 15,20,252 ; unpcklps %xmm4,%xmm7 + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,126,230 ; movd %xmm4,%esi + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,153 ; movss (%ecx,%ebx,4),%xmm0 + DB 243,15,16,36,177 ; movss (%ecx,%esi,4),%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 243,15,16,60,185 ; movss (%ecx,%edi,4),%xmm7 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 DB 15,20,248 ; unpcklps %xmm0,%xmm7 - DB 102,15,111,130,117,93,0,0 ; movdqa 0x5d75(%edx),%xmm0 - DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) + DB 102,15,20,252 ; unpcklpd %xmm4,%xmm7 + DB 102,15,111,130,95,92,0,0 ; movdqa 0x5c5f(%edx),%xmm0 + DB 102,15,127,133,104,255,255,255 ; movdqa %xmm0,-0x98(%ebp) DB 102,15,254,200 ; paddd %xmm0,%xmm1 DB 102,15,112,193,229 ; pshufd $0xe5,%xmm1,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax @@ -54373,275 +54305,282 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,193,231 ; pshufd $0xe7,%xmm1,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,153 ; movss (%ecx,%ebx,4),%xmm0 - DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 - DB 15,20,224 ; unpcklps %xmm0,%xmm4 - DB 243,15,16,44,185 ; movss (%ecx,%edi,4),%xmm5 - DB 243,15,16,4,177 ; movss (%ecx,%esi,4),%xmm0 - DB 15,20,232 ; unpcklps %xmm0,%xmm5 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 15,40,130,181,99,0,0 ; movaps 0x63b5(%edx),%xmm0 - DB 15,88,240 ; addps %xmm0,%xmm6 - DB 15,41,117,184 ; movaps %xmm6,-0x48(%ebp) - DB 15,40,141,56,255,255,255 ; movaps -0xc8(%ebp),%xmm1 + DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 243,15,16,52,185 ; movss (%ecx,%edi,4),%xmm6 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 102,15,20,241 ; unpcklpd %xmm1,%xmm6 + DB 15,40,130,207,98,0,0 ; movaps 0x62cf(%edx),%xmm0 + DB 15,40,141,232,254,255,255 ; movaps -0x118(%ebp),%xmm1 DB 15,88,200 ; addps %xmm0,%xmm1 - DB 15,41,141,40,255,255,255 ; movaps %xmm1,-0xd8(%ebp) - DB 15,88,69,200 ; addps -0x38(%ebp),%xmm0 + DB 15,41,77,200 ; movaps %xmm1,-0x38(%ebp) + DB 15,40,141,88,255,255,255 ; movaps -0xa8(%ebp),%xmm1 + DB 15,88,200 ; addps %xmm0,%xmm1 + DB 15,41,141,72,255,255,255 ; movaps %xmm1,-0xb8(%ebp) + DB 15,88,69,216 ; addps -0x28(%ebp),%xmm0 DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 - DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 - DB 102,15,112,192,245 ; pshufd $0xf5,%xmm0,%xmm0 - DB 102,15,244,133,120,255,255,255 ; pmuludq -0x88(%ebp),%xmm0 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,98,216 ; punpckldq %xmm0,%xmm3 - DB 102,15,127,157,120,255,255,255 ; movdqa %xmm3,-0x88(%ebp) - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,112,218,245 ; pshufd $0xf5,%xmm2,%xmm3 - DB 102,15,111,141,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm1 - DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 - DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 + DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 + DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 + DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 + DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 + DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 + DB 102,15,127,149,24,255,255,255 ; movdqa %xmm2,-0xe8(%ebp) + DB 102,15,254,218 ; paddd %xmm2,%xmm3 + DB 102,15,112,195,245 ; pshufd $0xf5,%xmm3,%xmm0 + DB 102,15,111,141,120,255,255,255 ; movdqa -0x88(%ebp),%xmm1 DB 102,15,244,217 ; pmuludq %xmm1,%xmm3 - DB 102,15,112,211,232 ; pshufd $0xe8,%xmm3,%xmm2 - DB 102,15,98,194 ; punpckldq %xmm2,%xmm0 - DB 102,15,112,208,229 ; pshufd $0xe5,%xmm0,%xmm2 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 + DB 102,15,112,195,232 ; pshufd $0xe8,%xmm3,%xmm0 + DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,126,202 ; movd %xmm1,%edx + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,126,206 ; movd %xmm1,%esi + DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,126,207 ; movd %xmm1,%edi + DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 + DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 + DB 243,15,16,28,129 ; movss (%ecx,%eax,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 + DB 102,15,20,218 ; unpcklpd %xmm2,%xmm3 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 + DB 102,15,250,202 ; psubd %xmm2,%xmm1 + DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 + DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 DB 102,15,126,210 ; movd %xmm2,%edx - DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi - DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 - DB 243,15,16,20,145 ; movss (%ecx,%edx,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,254,149,72,255,255,255 ; paddd -0xb8(%ebp),%xmm2 - DB 102,15,112,242,229 ; pshufd $0xe5,%xmm2,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 - DB 102,15,126,242 ; movd %xmm6,%edx - DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 - DB 102,15,126,215 ; movd %xmm2,%edi + DB 102,15,126,206 ; movd %xmm1,%esi + DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 + DB 102,15,126,207 ; movd %xmm1,%edi DB 243,15,16,20,185 ; movss (%ecx,%edi,4),%xmm2 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 + DB 15,20,202 ; unpcklps %xmm2,%xmm1 DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 - DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 + DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 DB 15,20,212 ; unpcklps %xmm4,%xmm2 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 102,15,254,69,216 ; paddd -0x28(%ebp),%xmm0 - DB 102,15,112,224,229 ; pshufd $0xe5,%xmm0,%xmm4 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 - DB 102,15,126,226 ; movd %xmm4,%edx + DB 102,15,20,209 ; unpcklpd %xmm1,%xmm2 + DB 102,15,254,133,104,255,255,255 ; paddd -0x98(%ebp),%xmm0 + DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,126,200 ; movd %xmm1,%eax + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,126,202 ; movd %xmm1,%edx DB 102,15,126,198 ; movd %xmm0,%esi DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 - DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 + DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 DB 15,20,224 ; unpcklps %xmm0,%xmm4 - DB 243,15,16,28,177 ; movss (%ecx,%esi,4),%xmm3 - DB 243,15,16,52,145 ; movss (%ecx,%edx,4),%xmm6 - DB 15,20,222 ; unpcklps %xmm6,%xmm3 - DB 15,20,220 ; unpcklps %xmm4,%xmm3 - DB 15,91,101,152 ; cvtdq2ps -0x68(%ebp),%xmm4 - DB 15,40,69,200 ; movaps -0x38(%ebp),%xmm0 - DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,41,69,200 ; movaps %xmm0,-0x38(%ebp) - DB 15,40,101,168 ; movaps -0x58(%ebp),%xmm4 - DB 15,92,204 ; subps %xmm4,%xmm1 - DB 15,89,200 ; mulps %xmm0,%xmm1 - DB 15,88,204 ; addps %xmm4,%xmm1 - DB 15,41,77,152 ; movaps %xmm1,-0x68(%ebp) - DB 15,92,215 ; subps %xmm7,%xmm2 - DB 15,89,208 ; mulps %xmm0,%xmm2 - DB 15,88,215 ; addps %xmm7,%xmm2 - DB 15,41,85,168 ; movaps %xmm2,-0x58(%ebp) + DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 102,15,20,204 ; unpcklpd %xmm4,%xmm1 + DB 15,91,69,184 ; cvtdq2ps -0x48(%ebp),%xmm0 + DB 15,40,101,216 ; movaps -0x28(%ebp),%xmm4 + DB 15,92,224 ; subps %xmm0,%xmm4 + DB 15,41,101,216 ; movaps %xmm4,-0x28(%ebp) DB 15,92,221 ; subps %xmm5,%xmm3 - DB 15,89,216 ; mulps %xmm0,%xmm3 + DB 15,89,220 ; mulps %xmm4,%xmm3 DB 15,88,221 ; addps %xmm5,%xmm3 - DB 243,15,91,141,40,255,255,255 ; cvttps2dq -0xd8(%ebp),%xmm1 - DB 102,15,112,225,245 ; pshufd $0xf5,%xmm1,%xmm4 - DB 102,15,111,69,136 ; movdqa -0x78(%ebp),%xmm0 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,244,224 ; pmuludq %xmm0,%xmm4 - DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 - DB 102,15,112,204,232 ; pshufd $0xe8,%xmm4,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,15,127,69,136 ; movdqa %xmm0,-0x78(%ebp) - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,15,254,189,232,254,255,255 ; paddd -0x118(%ebp),%xmm7 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,254,141,104,255,255,255 ; paddd -0x98(%ebp),%xmm1 - DB 102,15,112,225,245 ; pshufd $0xf5,%xmm1,%xmm4 - DB 102,15,111,133,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm0 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,233,232 ; pshufd $0xe8,%xmm1,%xmm5 - DB 102,15,244,224 ; pmuludq %xmm0,%xmm4 - DB 102,15,112,204,232 ; pshufd $0xe8,%xmm4,%xmm1 - DB 102,15,98,233 ; punpckldq %xmm1,%xmm5 - DB 102,15,112,205,229 ; pshufd $0xe5,%xmm5,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,205,78 ; pshufd $0x4e,%xmm5,%xmm1 - DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,112,205,231 ; pshufd $0xe7,%xmm5,%xmm1 - DB 102,15,126,206 ; movd %xmm1,%esi - DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 - DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 - DB 15,20,225 ; unpcklps %xmm1,%xmm4 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 - DB 243,15,16,52,145 ; movss (%ecx,%edx,4),%xmm6 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 15,41,141,40,255,255,255 ; movaps %xmm1,-0xd8(%ebp) - DB 102,15,111,229 ; movdqa %xmm5,%xmm4 - DB 102,15,111,149,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm2 - DB 102,15,254,226 ; paddd %xmm2,%xmm4 - DB 102,15,112,244,229 ; pshufd $0xe5,%xmm4,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 - DB 102,15,126,242 ; movd %xmm6,%edx - DB 102,15,126,230 ; movd %xmm4,%esi - DB 102,15,112,228,231 ; pshufd $0xe7,%xmm4,%xmm4 - DB 102,15,126,231 ; movd %xmm4,%edi + DB 15,41,93,184 ; movaps %xmm3,-0x48(%ebp) + DB 15,92,215 ; subps %xmm7,%xmm2 + DB 15,89,212 ; mulps %xmm4,%xmm2 + DB 15,88,215 ; addps %xmm7,%xmm2 + DB 15,41,149,56,255,255,255 ; movaps %xmm2,-0xc8(%ebp) + DB 15,92,206 ; subps %xmm6,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 15,88,206 ; addps %xmm6,%xmm1 + DB 15,41,141,216,254,255,255 ; movaps %xmm1,-0x128(%ebp) + DB 243,15,91,133,72,255,255,255 ; cvttps2dq -0xb8(%ebp),%xmm0 + DB 102,15,112,224,245 ; pshufd $0xf5,%xmm0,%xmm4 + DB 102,15,111,77,152 ; movdqa -0x68(%ebp),%xmm1 + DB 102,15,244,225 ; pmuludq %xmm1,%xmm4 + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,15,244,232 ; pmuludq %xmm0,%xmm5 + DB 102,15,112,253,232 ; pshufd $0xe8,%xmm5,%xmm7 + DB 102,15,112,196,232 ; pshufd $0xe8,%xmm4,%xmm0 + DB 102,15,98,248 ; punpckldq %xmm0,%xmm7 + DB 102,15,127,125,152 ; movdqa %xmm7,-0x68(%ebp) + DB 102,15,254,189,248,254,255,255 ; paddd -0x108(%ebp),%xmm7 + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,15,254,69,136 ; paddd -0x78(%ebp),%xmm0 + DB 102,15,112,224,245 ; pshufd $0xf5,%xmm0,%xmm4 + DB 102,15,111,141,120,255,255,255 ; movdqa -0x88(%ebp),%xmm1 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,244,225 ; pmuludq %xmm1,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,112,232,232 ; pshufd $0xe8,%xmm0,%xmm5 + DB 102,15,98,236 ; punpckldq %xmm4,%xmm5 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,197,229 ; pshufd $0xe5,%xmm5,%xmm0 + DB 102,15,126,194 ; movd %xmm0,%edx + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,112,197,231 ; pshufd $0xe7,%xmm5,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 + DB 243,15,16,36,177 ; movss (%ecx,%esi,4),%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 243,15,16,4,145 ; movss (%ecx,%edx,4),%xmm0 + DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 + DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 102,15,20,212 ; unpcklpd %xmm4,%xmm2 + DB 102,15,41,149,72,255,255,255 ; movapd %xmm2,-0xb8(%ebp) + DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,250,195 ; psubd %xmm3,%xmm0 + DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 + DB 102,15,112,224,229 ; pshufd $0xe5,%xmm0,%xmm4 + DB 102,15,126,224 ; movd %xmm4,%eax + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,126,226 ; movd %xmm4,%edx + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi DB 243,15,16,36,185 ; movss (%ecx,%edi,4),%xmm4 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 243,15,16,4,145 ; movss (%ecx,%edx,4),%xmm0 + DB 15,20,196 ; unpcklps %xmm4,%xmm0 DB 243,15,16,36,177 ; movss (%ecx,%esi,4),%xmm4 - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 DB 15,20,230 ; unpcklps %xmm6,%xmm4 - DB 102,15,254,109,216 ; paddd -0x28(%ebp),%xmm5 - DB 102,15,112,205,229 ; pshufd $0xe5,%xmm5,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,205,78 ; pshufd $0x4e,%xmm5,%xmm1 - DB 102,15,126,202 ; movd %xmm1,%edx + DB 102,15,20,224 ; unpcklpd %xmm0,%xmm4 + DB 102,15,111,157,104,255,255,255 ; movdqa -0x98(%ebp),%xmm3 + DB 102,15,254,235 ; paddd %xmm3,%xmm5 + DB 102,15,112,197,229 ; pshufd $0xe5,%xmm5,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,126,194 ; movd %xmm0,%edx DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,112,205,231 ; pshufd $0xe7,%xmm5,%xmm1 - DB 102,15,126,207 ; movd %xmm1,%edi - DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 + DB 102,15,112,197,231 ; pshufd $0xe7,%xmm5,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 + DB 243,15,16,52,145 ; movss (%ecx,%edx,4),%xmm6 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 DB 243,15,16,44,177 ; movss (%ecx,%esi,4),%xmm5 - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,233 ; unpcklps %xmm1,%xmm5 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 102,15,254,189,120,255,255,255 ; paddd -0x88(%ebp),%xmm7 - DB 102,15,112,207,245 ; pshufd $0xf5,%xmm7,%xmm1 - DB 102,15,244,248 ; pmuludq %xmm0,%xmm7 - DB 102,15,112,255,232 ; pshufd $0xe8,%xmm7,%xmm7 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,249 ; punpckldq %xmm1,%xmm7 - DB 102,15,112,207,229 ; pshufd $0xe5,%xmm7,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,207,78 ; pshufd $0x4e,%xmm7,%xmm1 - DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,112,207,231 ; pshufd $0xe7,%xmm7,%xmm1 - DB 102,15,126,206 ; movd %xmm1,%esi - DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 102,15,126,248 ; movd %xmm7,%eax DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 15,20,198 ; unpcklps %xmm6,%xmm0 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,254,202 ; paddd %xmm2,%xmm1 - DB 102,15,112,241,229 ; pshufd $0xe5,%xmm1,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 + DB 102,15,254,189,24,255,255,255 ; paddd -0xe8(%ebp),%xmm7 + DB 102,15,112,199,245 ; pshufd $0xf5,%xmm7,%xmm0 + DB 102,15,244,249 ; pmuludq %xmm1,%xmm7 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 102,15,112,240,232 ; pshufd $0xe8,%xmm0,%xmm6 + DB 102,15,112,199,232 ; pshufd $0xe8,%xmm7,%xmm0 + DB 102,15,98,198 ; punpckldq %xmm6,%xmm0 + DB 102,15,126,248 ; movd %xmm7,%eax + DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 DB 102,15,126,242 ; movd %xmm6,%edx - DB 102,15,126,206 ; movd %xmm1,%esi - DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 - DB 102,15,126,207 ; movd %xmm1,%edi - DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 102,15,254,125,216 ; paddd -0x28(%ebp),%xmm7 - DB 102,15,112,207,229 ; pshufd $0xe5,%xmm7,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,207,78 ; pshufd $0x4e,%xmm7,%xmm1 - DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,126,254 ; movd %xmm7,%esi - DB 102,15,112,207,231 ; pshufd $0xe7,%xmm7,%xmm1 - DB 102,15,126,207 ; movd %xmm1,%edi - DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,112,240,231 ; pshufd $0xe7,%xmm0,%xmm6 + DB 102,15,126,247 ; movd %xmm6,%edi + DB 243,15,16,52,185 ; movss (%ecx,%edi,4),%xmm6 DB 243,15,16,60,177 ; movss (%ecx,%esi,4),%xmm7 - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 DB 15,20,254 ; unpcklps %xmm6,%xmm7 - DB 15,40,181,40,255,255,255 ; movaps -0xd8(%ebp),%xmm6 - DB 15,92,198 ; subps %xmm6,%xmm0 - DB 15,40,77,200 ; movaps -0x38(%ebp),%xmm1 - DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,88,198 ; addps %xmm6,%xmm0 + DB 243,15,16,52,145 ; movss (%ecx,%edx,4),%xmm6 + DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 + DB 15,20,206 ; unpcklps %xmm6,%xmm1 + DB 102,15,20,207 ; unpcklpd %xmm7,%xmm1 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,250,242 ; psubd %xmm2,%xmm6 + DB 102,15,112,254,229 ; pshufd $0xe5,%xmm6,%xmm7 + DB 102,15,126,248 ; movd %xmm7,%eax + DB 102,15,112,254,78 ; pshufd $0x4e,%xmm6,%xmm7 + DB 102,15,126,250 ; movd %xmm7,%edx + DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,112,246,231 ; pshufd $0xe7,%xmm6,%xmm6 + DB 102,15,126,247 ; movd %xmm6,%edi + DB 243,15,16,52,185 ; movss (%ecx,%edi,4),%xmm6 + DB 243,15,16,60,145 ; movss (%ecx,%edx,4),%xmm7 + DB 15,20,254 ; unpcklps %xmm6,%xmm7 + DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 + DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 + DB 15,20,214 ; unpcklps %xmm6,%xmm2 + DB 102,15,20,215 ; unpcklpd %xmm7,%xmm2 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,15,254,199 ; paddd %xmm7,%xmm0 + DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 + DB 102,15,126,240 ; movd %xmm6,%eax + DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,126,242 ; movd %xmm6,%edx + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 + DB 243,15,16,52,145 ; movss (%ecx,%edx,4),%xmm6 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 243,15,16,28,177 ; movss (%ecx,%esi,4),%xmm3 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 15,20,216 ; unpcklps %xmm0,%xmm3 + DB 102,15,20,222 ; unpcklpd %xmm6,%xmm3 + DB 15,40,181,72,255,255,255 ; movaps -0xb8(%ebp),%xmm6 + DB 15,92,206 ; subps %xmm6,%xmm1 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 + DB 15,89,200 ; mulps %xmm0,%xmm1 + DB 15,88,206 ; addps %xmm6,%xmm1 DB 15,92,212 ; subps %xmm4,%xmm2 - DB 15,89,209 ; mulps %xmm1,%xmm2 + DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,88,212 ; addps %xmm4,%xmm2 - DB 15,92,253 ; subps %xmm5,%xmm7 - DB 15,89,249 ; mulps %xmm1,%xmm7 - DB 15,88,253 ; addps %xmm5,%xmm7 - DB 15,91,141,8,255,255,255 ; cvtdq2ps -0xf8(%ebp),%xmm1 - DB 15,40,165,56,255,255,255 ; movaps -0xc8(%ebp),%xmm4 - DB 15,92,225 ; subps %xmm1,%xmm4 - DB 15,41,165,56,255,255,255 ; movaps %xmm4,-0xc8(%ebp) - DB 15,40,77,152 ; movaps -0x68(%ebp),%xmm1 - DB 15,92,193 ; subps %xmm1,%xmm0 - DB 15,89,196 ; mulps %xmm4,%xmm0 - DB 15,88,193 ; addps %xmm1,%xmm0 - DB 15,41,69,152 ; movaps %xmm0,-0x68(%ebp) - DB 15,40,69,168 ; movaps -0x58(%ebp),%xmm0 + DB 15,92,221 ; subps %xmm5,%xmm3 + DB 15,89,216 ; mulps %xmm0,%xmm3 + DB 15,88,221 ; addps %xmm5,%xmm3 + DB 15,91,133,40,255,255,255 ; cvtdq2ps -0xd8(%ebp),%xmm0 + DB 15,40,165,88,255,255,255 ; movaps -0xa8(%ebp),%xmm4 + DB 15,92,224 ; subps %xmm0,%xmm4 + DB 15,41,165,88,255,255,255 ; movaps %xmm4,-0xa8(%ebp) + DB 15,40,69,184 ; movaps -0x48(%ebp),%xmm0 + DB 15,92,200 ; subps %xmm0,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 15,88,200 ; addps %xmm0,%xmm1 + DB 15,41,77,184 ; movaps %xmm1,-0x48(%ebp) + DB 15,40,133,56,255,255,255 ; movaps -0xc8(%ebp),%xmm0 DB 15,92,208 ; subps %xmm0,%xmm2 DB 15,89,212 ; mulps %xmm4,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 - DB 15,41,85,168 ; movaps %xmm2,-0x58(%ebp) - DB 15,92,251 ; subps %xmm3,%xmm7 - DB 15,89,252 ; mulps %xmm4,%xmm7 - DB 15,88,251 ; addps %xmm3,%xmm7 - DB 15,41,189,8,255,255,255 ; movaps %xmm7,-0xf8(%ebp) - DB 243,15,91,77,184 ; cvttps2dq -0x48(%ebp),%xmm1 - DB 102,15,127,77,184 ; movdqa %xmm1,-0x48(%ebp) - DB 102,15,111,133,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm0 + DB 15,41,149,56,255,255,255 ; movaps %xmm2,-0xc8(%ebp) + DB 15,40,133,216,254,255,255 ; movaps -0x128(%ebp),%xmm0 + DB 15,92,216 ; subps %xmm0,%xmm3 + DB 15,89,220 ; mulps %xmm4,%xmm3 + DB 15,88,216 ; addps %xmm0,%xmm3 + DB 15,41,157,40,255,255,255 ; movaps %xmm3,-0xd8(%ebp) + DB 243,15,91,77,200 ; cvttps2dq -0x38(%ebp),%xmm1 + DB 102,15,127,77,200 ; movdqa %xmm1,-0x38(%ebp) + DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,111,141,104,255,255,255 ; movdqa -0x98(%ebp),%xmm1 + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 DB 102,15,254,200 ; paddd %xmm0,%xmm1 DB 102,15,111,240 ; movdqa %xmm0,%xmm6 DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 - DB 102,15,111,133,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm0 + DB 102,15,111,133,120,255,255,255 ; movdqa -0x88(%ebp),%xmm0 DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 - DB 102,15,112,202,232 ; pshufd $0xe8,%xmm2,%xmm1 - DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 - DB 102,15,112,203,229 ; pshufd $0xe5,%xmm3,%xmm1 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 + DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 + DB 102,15,98,218 ; punpckldq %xmm2,%xmm3 DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 + DB 102,15,112,203,229 ; pshufd $0xe5,%xmm3,%xmm1 DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,112,203,231 ; pshufd $0xe7,%xmm3,%xmm1 + DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 DB 102,15,126,206 ; movd %xmm1,%esi - DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 - DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 + DB 102,15,112,203,231 ; pshufd $0xe7,%xmm3,%xmm1 + DB 102,15,126,207 ; movd %xmm1,%edi + DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 + DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,15,16,60,129 ; movss (%ecx,%eax,4),%xmm7 DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 - DB 15,20,252 ; unpcklps %xmm4,%xmm7 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 15,20,196 ; unpcklps %xmm4,%xmm0 + DB 102,15,20,194 ; unpcklpd %xmm2,%xmm0 + DB 102,15,41,69,168 ; movapd %xmm0,-0x58(%ebp) DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,111,133,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm0 - DB 102,15,254,208 ; paddd %xmm0,%xmm2 + DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 + DB 102,15,250,208 ; psubd %xmm0,%xmm2 DB 102,15,112,226,229 ; pshufd $0xe5,%xmm2,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax DB 102,15,112,226,78 ; pshufd $0x4e,%xmm2,%xmm4 @@ -54650,13 +54589,13 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 DB 102,15,126,215 ; movd %xmm2,%edi DB 243,15,16,20,185 ; movss (%ecx,%edi,4),%xmm2 - DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 + DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 DB 15,20,226 ; unpcklps %xmm2,%xmm4 DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 - DB 243,15,16,44,145 ; movss (%ecx,%edx,4),%xmm5 + DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 DB 15,20,213 ; unpcklps %xmm5,%xmm2 - DB 15,20,212 ; unpcklps %xmm4,%xmm2 - DB 102,15,254,93,216 ; paddd -0x28(%ebp),%xmm3 + DB 102,15,20,212 ; unpcklpd %xmm4,%xmm2 + DB 102,15,254,223 ; paddd %xmm7,%xmm3 DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 @@ -54665,37 +54604,38 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,219,231 ; pshufd $0xe7,%xmm3,%xmm3 DB 102,15,126,223 ; movd %xmm3,%edi DB 243,15,16,28,185 ; movss (%ecx,%edi,4),%xmm3 - DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 + DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 DB 15,20,227 ; unpcklps %xmm3,%xmm4 DB 243,15,16,44,177 ; movss (%ecx,%esi,4),%xmm5 - DB 243,15,16,28,145 ; movss (%ecx,%edx,4),%xmm3 + DB 243,15,16,28,129 ; movss (%ecx,%eax,4),%xmm3 DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 102,15,111,222 ; movdqa %xmm6,%xmm3 - DB 102,15,254,157,120,255,255,255 ; paddd -0x88(%ebp),%xmm3 - DB 102,15,112,227,245 ; pshufd $0xf5,%xmm3,%xmm4 - DB 102,15,111,141,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm1 - DB 102,15,244,217 ; pmuludq %xmm1,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 + DB 102,15,20,236 ; unpcklpd %xmm4,%xmm5 + DB 102,15,111,230 ; movdqa %xmm6,%xmm4 + DB 102,15,254,165,24,255,255,255 ; paddd -0xe8(%ebp),%xmm4 + DB 102,15,112,220,245 ; pshufd $0xf5,%xmm4,%xmm3 + DB 102,15,111,141,120,255,255,255 ; movdqa -0x88(%ebp),%xmm1 DB 102,15,244,225 ; pmuludq %xmm1,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,111,244 ; movdqa %xmm4,%xmm6 + DB 102,15,244,217 ; pmuludq %xmm1,%xmm3 + DB 102,15,112,227,232 ; pshufd $0xe8,%xmm3,%xmm4 + DB 102,15,112,222,232 ; pshufd $0xe8,%xmm6,%xmm3 DB 102,15,98,220 ; punpckldq %xmm4,%xmm3 + DB 102,15,126,240 ; movd %xmm6,%eax DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 DB 102,15,126,226 ; movd %xmm4,%edx - DB 102,15,112,227,231 ; pshufd $0xe7,%xmm3,%xmm4 + DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 DB 102,15,126,230 ; movd %xmm4,%esi - DB 243,15,16,36,177 ; movss (%ecx,%esi,4),%xmm4 - DB 243,15,16,52,129 ; movss (%ecx,%eax,4),%xmm6 + DB 102,15,112,227,231 ; pshufd $0xe7,%xmm3,%xmm4 + DB 102,15,126,231 ; movd %xmm4,%edi + DB 243,15,16,36,185 ; movss (%ecx,%edi,4),%xmm4 + DB 243,15,16,52,177 ; movss (%ecx,%esi,4),%xmm6 DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 + DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,254,224 ; paddd %xmm0,%xmm4 + DB 102,15,250,224 ; psubd %xmm0,%xmm4 DB 102,15,112,244,229 ; pshufd $0xe5,%xmm4,%xmm6 DB 102,15,126,240 ; movd %xmm6,%eax DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 @@ -54704,13 +54644,14 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,228,231 ; pshufd $0xe7,%xmm4,%xmm4 DB 102,15,126,231 ; movd %xmm4,%edi DB 243,15,16,36,185 ; movss (%ecx,%edi,4),%xmm4 - DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 243,15,16,4,145 ; movss (%ecx,%edx,4),%xmm0 DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 243,15,16,52,177 ; movss (%ecx,%esi,4),%xmm6 - DB 243,15,16,36,145 ; movss (%ecx,%edx,4),%xmm4 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 102,15,254,93,216 ; paddd -0x28(%ebp),%xmm3 + DB 243,15,16,60,177 ; movss (%ecx,%esi,4),%xmm7 + DB 243,15,16,36,129 ; movss (%ecx,%eax,4),%xmm4 + DB 15,20,252 ; unpcklps %xmm4,%xmm7 + DB 102,15,20,248 ; unpcklpd %xmm0,%xmm7 + DB 102,15,111,165,104,255,255,255 ; movdqa -0x98(%ebp),%xmm4 + DB 102,15,254,220 ; paddd %xmm4,%xmm3 DB 102,15,112,195,229 ; pshufd $0xe5,%xmm3,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 @@ -54719,53 +54660,54 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,195,231 ; pshufd $0xe7,%xmm3,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 - DB 243,15,16,28,129 ; movss (%ecx,%eax,4),%xmm3 + DB 243,15,16,28,145 ; movss (%ecx,%edx,4),%xmm3 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 243,15,16,36,177 ; movss (%ecx,%esi,4),%xmm4 - DB 243,15,16,4,145 ; movss (%ecx,%edx,4),%xmm0 - DB 15,20,224 ; unpcklps %xmm0,%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 15,92,207 ; subps %xmm7,%xmm1 - DB 15,40,69,200 ; movaps -0x38(%ebp),%xmm0 + DB 243,15,16,52,177 ; movss (%ecx,%esi,4),%xmm6 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 102,15,20,243 ; unpcklpd %xmm3,%xmm6 + DB 15,40,93,168 ; movaps -0x58(%ebp),%xmm3 + DB 15,92,203 ; subps %xmm3,%xmm1 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 DB 15,89,200 ; mulps %xmm0,%xmm1 - DB 15,88,207 ; addps %xmm7,%xmm1 - DB 15,41,141,24,255,255,255 ; movaps %xmm1,-0xe8(%ebp) - DB 15,92,242 ; subps %xmm2,%xmm6 + DB 15,88,203 ; addps %xmm3,%xmm1 + DB 15,41,77,168 ; movaps %xmm1,-0x58(%ebp) + DB 15,92,250 ; subps %xmm2,%xmm7 + DB 15,89,248 ; mulps %xmm0,%xmm7 + DB 15,88,250 ; addps %xmm2,%xmm7 + DB 15,92,245 ; subps %xmm5,%xmm6 DB 15,89,240 ; mulps %xmm0,%xmm6 - DB 15,88,242 ; addps %xmm2,%xmm6 - DB 15,92,229 ; subps %xmm5,%xmm4 - DB 15,89,224 ; mulps %xmm0,%xmm4 - DB 15,88,229 ; addps %xmm5,%xmm4 - DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 - DB 102,15,254,77,184 ; paddd -0x48(%ebp),%xmm1 - DB 102,15,127,77,136 ; movdqa %xmm1,-0x78(%ebp) - DB 102,15,111,133,104,255,255,255 ; movdqa -0x98(%ebp),%xmm0 - DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,189,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm7 - DB 102,15,244,199 ; pmuludq %xmm7,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 15,88,245 ; addps %xmm5,%xmm6 + DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 + DB 102,15,254,69,200 ; paddd -0x38(%ebp),%xmm0 + DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) + DB 102,15,111,85,136 ; movdqa -0x78(%ebp),%xmm2 + DB 102,15,254,208 ; paddd %xmm0,%xmm2 + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,157,120,255,255,255 ; movdqa -0x88(%ebp),%xmm3 + DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 + DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 + DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,15,126,208 ; movd %xmm2,%eax DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,126,206 ; movd %xmm1,%esi - DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 - DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 + DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,126,207 ; movd %xmm1,%edi + DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 + DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,15,16,28,129 ; movss (%ecx,%eax,4),%xmm3 DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,41,157,104,255,255,255 ; movaps %xmm3,-0x98(%ebp) + DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 + DB 15,20,233 ; unpcklps %xmm1,%xmm5 + DB 102,15,20,234 ; unpcklpd %xmm2,%xmm5 + DB 102,15,41,109,136 ; movapd %xmm5,-0x78(%ebp) DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,173,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm5 - DB 102,15,254,205 ; paddd %xmm5,%xmm1 + DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 + DB 102,15,250,202 ; psubd %xmm2,%xmm1 DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 DB 102,15,126,208 ; movd %xmm2,%eax DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 @@ -54774,14 +54716,14 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,207 ; movd %xmm1,%edi DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 - DB 243,15,16,20,129 ; movss (%ecx,%eax,4),%xmm2 + DB 243,15,16,20,145 ; movss (%ecx,%edx,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,15,16,28,177 ; movss (%ecx,%esi,4),%xmm3 - DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,41,93,184 ; movaps %xmm3,-0x48(%ebp) - DB 102,15,254,69,216 ; paddd -0x28(%ebp),%xmm0 + DB 243,15,16,44,177 ; movss (%ecx,%esi,4),%xmm5 + DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 + DB 15,20,233 ; unpcklps %xmm1,%xmm5 + DB 102,15,20,234 ; unpcklpd %xmm2,%xmm5 + DB 102,15,41,109,200 ; movapd %xmm5,-0x38(%ebp) + DB 102,15,254,196 ; paddd %xmm4,%xmm0 DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 @@ -54790,99 +54732,101 @@ _sk_clut_3D_sse2 LABEL PROC DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 - DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 + DB 243,15,16,12,145 ; movss (%ecx,%edx,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 243,15,16,28,177 ; movss (%ecx,%esi,4),%xmm3 - DB 243,15,16,4,145 ; movss (%ecx,%edx,4),%xmm0 - DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 - DB 102,15,254,141,120,255,255,255 ; paddd -0x88(%ebp),%xmm1 + DB 243,15,16,36,177 ; movss (%ecx,%esi,4),%xmm4 + DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,20,225 ; unpcklpd %xmm1,%xmm4 + DB 102,15,111,77,152 ; movdqa -0x68(%ebp),%xmm1 + DB 102,15,254,141,24,255,255,255 ; paddd -0xe8(%ebp),%xmm1 DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 - DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 - DB 102,15,244,199 ; pmuludq %xmm7,%xmm0 - DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 + DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 + DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 + DB 102,15,126,200 ; movd %xmm1,%eax DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,194 ; movd %xmm0,%edx - DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,198 ; movd %xmm0,%esi - DB 243,15,16,4,177 ; movss (%ecx,%esi,4),%xmm0 - DB 243,15,16,12,129 ; movss (%ecx,%eax,4),%xmm1 + DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 243,15,16,4,185 ; movss (%ecx,%edi,4),%xmm0 + DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax + DB 243,15,16,28,145 ; movss (%ecx,%edx,4),%xmm3 DB 243,15,16,4,129 ; movss (%ecx,%eax,4),%xmm0 - DB 243,15,16,60,145 ; movss (%ecx,%edx,4),%xmm7 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 102,15,254,234 ; paddd %xmm2,%xmm5 - DB 102,15,112,205,229 ; pshufd $0xe5,%xmm5,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,205,78 ; pshufd $0x4e,%xmm5,%xmm1 - DB 102,15,126,202 ; movd %xmm1,%edx - DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,112,205,231 ; pshufd $0xe7,%xmm5,%xmm1 + DB 15,20,195 ; unpcklps %xmm3,%xmm0 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,250,203 ; psubd %xmm3,%xmm1 + DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,218 ; movd %xmm3,%edx + DB 102,15,126,206 ; movd %xmm1,%esi + DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,207 ; movd %xmm1,%edi DB 243,15,16,12,185 ; movss (%ecx,%edi,4),%xmm1 - DB 243,15,16,60,129 ; movss (%ecx,%eax,4),%xmm7 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 + DB 243,15,16,28,145 ; movss (%ecx,%edx,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 DB 243,15,16,12,177 ; movss (%ecx,%esi,4),%xmm1 - DB 243,15,16,44,145 ; movss (%ecx,%edx,4),%xmm5 + DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 102,15,254,85,216 ; paddd -0x28(%ebp),%xmm2 - DB 102,15,112,234,229 ; pshufd $0xe5,%xmm2,%xmm5 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 102,15,112,234,78 ; pshufd $0x4e,%xmm2,%xmm5 - DB 102,15,126,234 ; movd %xmm5,%edx + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 + DB 102,15,254,149,104,255,255,255 ; paddd -0x98(%ebp),%xmm2 + DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,126,218 ; movd %xmm3,%edx DB 102,15,126,214 ; movd %xmm2,%esi DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 DB 102,15,126,215 ; movd %xmm2,%edi DB 243,15,16,20,185 ; movss (%ecx,%edi,4),%xmm2 - DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 + DB 243,15,16,28,145 ; movss (%ecx,%edx,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 243,15,16,20,177 ; movss (%ecx,%esi,4),%xmm2 - DB 243,15,16,60,145 ; movss (%ecx,%edx,4),%xmm7 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 + DB 243,15,16,44,129 ; movss (%ecx,%eax,4),%xmm5 DB 15,20,213 ; unpcklps %xmm5,%xmm2 - DB 15,40,189,104,255,255,255 ; movaps -0x98(%ebp),%xmm7 - DB 15,92,199 ; subps %xmm7,%xmm0 - DB 15,40,109,200 ; movaps -0x38(%ebp),%xmm5 - DB 15,89,197 ; mulps %xmm5,%xmm0 - DB 15,88,199 ; addps %xmm7,%xmm0 - DB 15,40,125,184 ; movaps -0x48(%ebp),%xmm7 - DB 15,92,207 ; subps %xmm7,%xmm1 - DB 15,89,205 ; mulps %xmm5,%xmm1 - DB 15,88,207 ; addps %xmm7,%xmm1 - DB 15,92,211 ; subps %xmm3,%xmm2 - DB 15,89,213 ; mulps %xmm5,%xmm2 - DB 15,88,211 ; addps %xmm3,%xmm2 - DB 15,40,173,24,255,255,255 ; movaps -0xe8(%ebp),%xmm5 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 15,40,109,136 ; movaps -0x78(%ebp),%xmm5 DB 15,92,197 ; subps %xmm5,%xmm0 - DB 15,40,157,56,255,255,255 ; movaps -0xc8(%ebp),%xmm3 + DB 15,40,93,216 ; movaps -0x28(%ebp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 DB 15,88,197 ; addps %xmm5,%xmm0 - DB 15,92,206 ; subps %xmm6,%xmm1 + DB 15,40,109,200 ; movaps -0x38(%ebp),%xmm5 + DB 15,92,205 ; subps %xmm5,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 15,88,206 ; addps %xmm6,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 DB 15,92,212 ; subps %xmm4,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 DB 15,88,212 ; addps %xmm4,%xmm2 - DB 15,91,157,232,254,255,255 ; cvtdq2ps -0x118(%ebp),%xmm3 - DB 15,40,165,216,254,255,255 ; movaps -0x128(%ebp),%xmm4 + DB 15,40,101,168 ; movaps -0x58(%ebp),%xmm4 + DB 15,92,196 ; subps %xmm4,%xmm0 + DB 15,40,157,88,255,255,255 ; movaps -0xa8(%ebp),%xmm3 + DB 15,89,195 ; mulps %xmm3,%xmm0 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 15,92,207 ; subps %xmm7,%xmm1 + DB 15,89,203 ; mulps %xmm3,%xmm1 + DB 15,88,207 ; addps %xmm7,%xmm1 + DB 15,92,214 ; subps %xmm6,%xmm2 + DB 15,89,211 ; mulps %xmm3,%xmm2 + DB 15,88,214 ; addps %xmm6,%xmm2 + DB 15,91,157,248,254,255,255 ; cvtdq2ps -0x108(%ebp),%xmm3 + DB 15,40,165,232,254,255,255 ; movaps -0x118(%ebp),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 - DB 15,40,93,152 ; movaps -0x68(%ebp),%xmm3 + DB 15,40,93,184 ; movaps -0x48(%ebp),%xmm3 DB 15,92,195 ; subps %xmm3,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,88,195 ; addps %xmm3,%xmm0 - DB 15,40,93,168 ; movaps -0x58(%ebp),%xmm3 + DB 15,40,157,56,255,255,255 ; movaps -0xc8(%ebp),%xmm3 DB 15,92,203 ; subps %xmm3,%xmm1 DB 15,89,204 ; mulps %xmm4,%xmm1 DB 15,88,203 ; addps %xmm3,%xmm1 - DB 15,40,157,8,255,255,255 ; movaps -0xf8(%ebp),%xmm3 + DB 15,40,157,40,255,255,255 ; movaps -0xd8(%ebp),%xmm3 DB 15,92,211 ; subps %xmm3,%xmm2 DB 15,89,212 ; mulps %xmm4,%xmm2 DB 15,88,211 ; addps %xmm3,%xmm2 @@ -54908,7 +54852,7 @@ _sk_clut_4D_sse2 LABEL PROC DB 83 ; push %ebx DB 87 ; push %edi DB 86 ; push %esi - DB 129,236,124,1,0,0 ; sub $0x17c,%esp + DB 129,236,92,1,0,0 ; sub $0x15c,%esp DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 137,141,212,254,255,255 ; mov %ecx,-0x12c(%ebp) @@ -54922,25 +54866,25 @@ _sk_clut_4D_sse2 LABEL PROC DB 15,40,236 ; movaps %xmm4,%xmm5 DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,15,112,251,0 ; pshufd $0x0,%xmm3,%xmm7 - DB 102,15,127,189,248,254,255,255 ; movdqa %xmm7,-0x108(%ebp) + DB 102,15,127,189,24,255,255,255 ; movdqa %xmm7,-0xe8(%ebp) DB 139,74,12 ; mov 0xc(%edx),%ecx DB 141,113,255 ; lea -0x1(%ecx),%esi DB 102,15,110,230 ; movd %esi,%xmm4 DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 DB 15,91,228 ; cvtdq2ps %xmm4,%xmm4 DB 15,89,226 ; mulps %xmm2,%xmm4 - DB 15,41,165,232,254,255,255 ; movaps %xmm4,-0x118(%ebp) + DB 15,41,165,248,254,255,255 ; movaps %xmm4,-0x108(%ebp) DB 243,15,91,244 ; cvttps2dq %xmm4,%xmm6 DB 102,15,127,181,216,254,255,255 ; movdqa %xmm6,-0x128(%ebp) - DB 102,15,111,214 ; movdqa %xmm6,%xmm2 - DB 102,15,244,215 ; pmuludq %xmm7,%xmm2 - DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 DB 102,15,112,214,245 ; pshufd $0xf5,%xmm6,%xmm2 DB 102,15,244,215 ; pmuludq %xmm7,%xmm2 + DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,15,244,230 ; pmuludq %xmm6,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 DB 102,15,98,226 ; punpckldq %xmm2,%xmm4 DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,127,181,56,255,255,255 ; movdqa %xmm6,-0xc8(%ebp) + DB 102,15,127,181,8,255,255,255 ; movdqa %xmm6,-0xf8(%ebp) DB 102,15,110,209 ; movd %ecx,%xmm2 DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 @@ -54953,21 +54897,20 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3 DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 DB 15,89,217 ; mulps %xmm1,%xmm3 - DB 15,41,157,40,255,255,255 ; movaps %xmm3,-0xd8(%ebp) + DB 15,41,157,56,255,255,255 ; movaps %xmm3,-0xc8(%ebp) DB 102,15,112,250,0 ; pshufd $0x0,%xmm2,%xmm7 - DB 102,15,127,189,104,255,255,255 ; movdqa %xmm7,-0x98(%ebp) + DB 102,15,127,189,88,255,255,255 ; movdqa %xmm7,-0xa8(%ebp) DB 243,15,91,227 ; cvttps2dq %xmm3,%xmm4 - DB 102,15,127,165,8,255,255,255 ; movdqa %xmm4,-0xf8(%ebp) - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 - DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 - DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 + DB 102,15,127,165,40,255,255,255 ; movdqa %xmm4,-0xd8(%ebp) DB 102,15,112,204,245 ; pshufd $0xf5,%xmm4,%xmm1 - DB 102,15,112,231,245 ; pshufd $0xf5,%xmm7,%xmm4 - DB 102,15,127,165,88,255,255,255 ; movdqa %xmm4,-0xa8(%ebp) - DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 + DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 + DB 102,15,111,223 ; movdqa %xmm7,%xmm3 + DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 + DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 - DB 102,15,127,157,72,255,255,255 ; movdqa %xmm3,-0xb8(%ebp) + DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,15,127,165,72,255,255,255 ; movdqa %xmm4,-0xb8(%ebp) DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 DB 139,74,4 ; mov 0x4(%edx),%ecx @@ -54976,150 +54919,268 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 DB 15,89,208 ; mulps %xmm0,%xmm2 - DB 15,41,85,152 ; movaps %xmm2,-0x68(%ebp) - DB 102,15,112,193,0 ; pshufd $0x0,%xmm1,%xmm0 - DB 243,15,91,226 ; cvttps2dq %xmm2,%xmm4 - DB 102,15,127,101,200 ; movdqa %xmm4,-0x38(%ebp) - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 - DB 102,15,112,204,245 ; pshufd $0xf5,%xmm4,%xmm1 - DB 102,15,112,224,245 ; pshufd $0xf5,%xmm0,%xmm4 - DB 102,15,127,101,184 ; movdqa %xmm4,-0x48(%ebp) - DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 + DB 15,41,85,216 ; movaps %xmm2,-0x28(%ebp) + DB 102,15,112,217,0 ; pshufd $0x0,%xmm1,%xmm3 + DB 102,15,127,93,168 ; movdqa %xmm3,-0x58(%ebp) + DB 243,15,91,194 ; cvttps2dq %xmm2,%xmm0 + DB 102,15,127,133,104,255,255,255 ; movdqa %xmm0,-0x98(%ebp) + DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 + DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 + DB 102,15,112,218,232 ; pshufd $0xe8,%xmm2,%xmm3 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,209 ; punpckldq %xmm1,%xmm2 - DB 102,15,127,85,168 ; movdqa %xmm2,-0x58(%ebp) - DB 232,0,0,0,0 ; call 7e91 <_sk_clut_4D_sse2+0x179> + DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 + DB 102,15,127,93,152 ; movdqa %xmm3,-0x68(%ebp) + DB 232,0,0,0,0 ; call 7eb7 <_sk_clut_4D_sse2+0x16e> DB 89 ; pop %ecx - DB 15,40,229 ; movaps %xmm5,%xmm4 - DB 15,41,165,120,254,255,255 ; movaps %xmm4,-0x188(%ebp) - DB 243,15,91,204 ; cvttps2dq %xmm4,%xmm1 + DB 15,40,197 ; movaps %xmm5,%xmm0 + DB 15,41,133,152,254,255,255 ; movaps %xmm0,-0x168(%ebp) + DB 243,15,91,200 ; cvttps2dq %xmm0,%xmm1 DB 102,15,127,141,184,254,255,255 ; movdqa %xmm1,-0x148(%ebp) - DB 102,15,111,238 ; movdqa %xmm6,%xmm5 - DB 102,15,254,233 ; paddd %xmm1,%xmm5 - DB 102,15,127,173,24,255,255,255 ; movdqa %xmm5,-0xe8(%ebp) - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,15,254,245 ; paddd %xmm5,%xmm6 - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,15,254,206 ; paddd %xmm6,%xmm1 - DB 102,15,111,153,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm3 - DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 - DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 - DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 + DB 102,15,254,241 ; paddd %xmm1,%xmm6 + DB 102,15,127,117,200 ; movdqa %xmm6,-0x38(%ebp) + DB 102,15,111,212 ; movdqa %xmm4,%xmm2 + DB 102,15,254,214 ; paddd %xmm6,%xmm2 + DB 102,15,254,218 ; paddd %xmm2,%xmm3 + DB 102,15,111,161,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm4 + DB 102,15,112,203,245 ; pshufd $0xf5,%xmm3,%xmm1 + DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 + DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 + DB 102,15,112,233,232 ; pshufd $0xe8,%xmm1,%xmm5 + DB 102,15,112,203,232 ; pshufd $0xe8,%xmm3,%xmm1 + DB 102,15,98,205 ; punpckldq %xmm5,%xmm1 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 + DB 102,15,126,223 ; movd %xmm3,%edi + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,219 ; movd %xmm3,%ebx + DB 102,15,112,217,231 ; pshufd $0xe7,%xmm1,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 139,18 ; mov (%edx),%edx + DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 243,15,16,52,154 ; movss (%edx,%ebx,4),%xmm6 + DB 15,20,243 ; unpcklps %xmm3,%xmm6 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 + DB 15,20,227 ; unpcklps %xmm3,%xmm4 + DB 102,15,20,230 ; unpcklpd %xmm6,%xmm4 + DB 102,15,41,101,184 ; movapd %xmm4,-0x48(%ebp) + DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,250,220 ; psubd %xmm4,%xmm3 + DB 102,15,118,237 ; pcmpeqd %xmm5,%xmm5 + DB 102,15,112,243,229 ; pshufd $0xe5,%xmm3,%xmm6 + DB 102,15,126,240 ; movd %xmm6,%eax + DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 + DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,126,223 ; movd %xmm3,%edi + DB 102,15,112,219,231 ; pshufd $0xe7,%xmm3,%xmm3 + DB 102,15,126,219 ; movd %xmm3,%ebx + DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 + DB 15,20,243 ; unpcklps %xmm3,%xmm6 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 + DB 15,20,223 ; unpcklps %xmm7,%xmm3 + DB 102,15,20,222 ; unpcklpd %xmm6,%xmm3 + DB 102,15,41,157,232,254,255,255 ; movapd %xmm3,-0x118(%ebp) + DB 102,15,254,137,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm1 + DB 102,15,112,241,229 ; pshufd $0xe5,%xmm1,%xmm6 + DB 102,15,126,240 ; movd %xmm6,%eax + DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 + DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,126,207 ; movd %xmm1,%edi + DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 + DB 102,15,126,203 ; movd %xmm1,%ebx + DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 + DB 15,20,241 ; unpcklps %xmm1,%xmm6 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 + DB 15,20,223 ; unpcklps %xmm7,%xmm3 + DB 102,15,20,222 ; unpcklpd %xmm6,%xmm3 + DB 15,40,185,57,87,0,0 ; movaps 0x5739(%ecx),%xmm7 + DB 15,88,199 ; addps %xmm7,%xmm0 + DB 15,41,133,120,255,255,255 ; movaps %xmm0,-0x88(%ebp) + DB 15,40,133,248,254,255,255 ; movaps -0x108(%ebp),%xmm0 + DB 15,88,199 ; addps %xmm7,%xmm0 + DB 15,41,69,136 ; movaps %xmm0,-0x78(%ebp) + DB 15,40,133,56,255,255,255 ; movaps -0xc8(%ebp),%xmm0 + DB 15,88,199 ; addps %xmm7,%xmm0 + DB 15,41,133,168,254,255,255 ; movaps %xmm0,-0x158(%ebp) + DB 15,88,125,216 ; addps -0x28(%ebp),%xmm7 + DB 243,15,91,255 ; cvttps2dq %xmm7,%xmm7 + DB 102,15,112,199,245 ; pshufd $0xf5,%xmm7,%xmm0 + DB 102,15,111,117,168 ; movdqa -0x58(%ebp),%xmm6 + DB 102,15,244,198 ; pmuludq %xmm6,%xmm0 + DB 102,15,244,247 ; pmuludq %xmm7,%xmm6 + DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,98,224 ; punpckldq %xmm0,%xmm4 + DB 102,15,127,101,168 ; movdqa %xmm4,-0x58(%ebp) + DB 102,15,254,212 ; paddd %xmm4,%xmm2 + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,161,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm4 + DB 102,15,244,212 ; pmuludq %xmm4,%xmm2 + DB 102,15,244,196 ; pmuludq %xmm4,%xmm0 + DB 102,15,112,240,232 ; pshufd $0xe8,%xmm0,%xmm6 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 + DB 102,15,98,198 ; punpckldq %xmm6,%xmm0 + DB 102,15,126,208 ; movd %xmm2,%eax + DB 102,15,112,208,229 ; pshufd $0xe5,%xmm0,%xmm2 DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 + DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 DB 102,15,126,215 ; movd %xmm2,%edi - DB 102,15,112,209,231 ; pshufd $0xe7,%xmm1,%xmm2 + DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 DB 102,15,126,211 ; movd %xmm2,%ebx - DB 139,18 ; mov (%edx),%edx DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 - DB 102,15,126,206 ; movd %xmm1,%esi - DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 - DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 - DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) - DB 102,15,111,153,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm3 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,112,234,229 ; pshufd $0xe5,%xmm2,%xmm5 - DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,112,234,78 ; pshufd $0x4e,%xmm2,%xmm5 - DB 102,15,126,239 ; movd %xmm5,%edi - DB 102,15,126,211 ; movd %xmm2,%ebx + DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 + DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 + DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 + DB 15,20,202 ; unpcklps %xmm2,%xmm1 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,250,213 ; psubd %xmm5,%xmm2 + DB 102,15,112,242,229 ; pshufd $0xe5,%xmm2,%xmm6 + DB 102,15,126,240 ; movd %xmm6,%eax + DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 + DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,126,215 ; movd %xmm2,%edi DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 - DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 + DB 102,15,126,211 ; movd %xmm2,%ebx DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,60,186 ; movss (%edx,%edi,4),%xmm7 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 + DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 + DB 243,15,16,44,130 ; movss (%edx,%eax,4),%xmm5 DB 15,20,213 ; unpcklps %xmm5,%xmm2 - DB 15,41,149,136,254,255,255 ; movaps %xmm2,-0x178(%ebp) - DB 102,15,254,137,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm1 - DB 102,15,112,233,229 ; pshufd $0xe5,%xmm1,%xmm5 + DB 102,15,20,214 ; unpcklpd %xmm6,%xmm2 + DB 102,15,254,129,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm0 + DB 102,15,112,232,229 ; pshufd $0xe5,%xmm0,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax - DB 102,15,112,233,78 ; pshufd $0x4e,%xmm1,%xmm5 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,126,207 ; movd %xmm1,%edi - DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 - DB 102,15,126,203 ; movd %xmm1,%ebx - DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 - DB 243,15,16,44,130 ; movss (%edx,%eax,4),%xmm5 - DB 15,20,233 ; unpcklps %xmm1,%xmm5 - DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 15,41,141,152,254,255,255 ; movaps %xmm1,-0x168(%ebp) - DB 15,40,185,63,88,0,0 ; movaps 0x583f(%ecx),%xmm7 - DB 15,88,231 ; addps %xmm7,%xmm4 - DB 15,41,165,120,255,255,255 ; movaps %xmm4,-0x88(%ebp) + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 243,15,16,60,186 ; movss (%edx,%edi,4),%xmm7 + DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 + DB 15,20,254 ; unpcklps %xmm6,%xmm7 + DB 102,15,20,253 ; unpcklpd %xmm5,%xmm7 + DB 15,91,173,104,255,255,255 ; cvtdq2ps -0x98(%ebp),%xmm5 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 + DB 15,92,197 ; subps %xmm5,%xmm0 + DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) + DB 15,40,109,184 ; movaps -0x48(%ebp),%xmm5 + DB 15,92,205 ; subps %xmm5,%xmm1 + DB 15,89,200 ; mulps %xmm0,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 15,41,141,104,255,255,255 ; movaps %xmm1,-0x98(%ebp) DB 15,40,141,232,254,255,255 ; movaps -0x118(%ebp),%xmm1 - DB 15,88,207 ; addps %xmm7,%xmm1 - DB 15,41,77,136 ; movaps %xmm1,-0x78(%ebp) - DB 15,40,141,40,255,255,255 ; movaps -0xd8(%ebp),%xmm1 - DB 15,88,207 ; addps %xmm7,%xmm1 - DB 15,41,141,168,254,255,255 ; movaps %xmm1,-0x158(%ebp) - DB 15,40,101,152 ; movaps -0x68(%ebp),%xmm4 - DB 15,88,252 ; addps %xmm4,%xmm7 - DB 243,15,91,255 ; cvttps2dq %xmm7,%xmm7 - DB 102,15,244,199 ; pmuludq %xmm7,%xmm0 - DB 102,15,112,255,245 ; pshufd $0xf5,%xmm7,%xmm7 - DB 102,15,244,125,184 ; pmuludq -0x48(%ebp),%xmm7 - DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 - DB 102,15,112,199,232 ; pshufd $0xe8,%xmm7,%xmm0 - DB 102,15,98,200 ; punpckldq %xmm0,%xmm1 - DB 102,15,127,77,184 ; movdqa %xmm1,-0x48(%ebp) - DB 102,15,254,241 ; paddd %xmm1,%xmm6 - DB 102,15,112,254,245 ; pshufd $0xf5,%xmm6,%xmm7 - DB 102,15,111,137,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm1 - DB 102,15,244,241 ; pmuludq %xmm1,%xmm6 - DB 102,15,112,198,232 ; pshufd $0xe8,%xmm6,%xmm0 - DB 102,15,244,249 ; pmuludq %xmm1,%xmm7 - DB 102,15,112,247,232 ; pshufd $0xe8,%xmm7,%xmm6 - DB 102,15,98,198 ; punpckldq %xmm6,%xmm0 - DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 + DB 15,92,209 ; subps %xmm1,%xmm2 + DB 15,89,208 ; mulps %xmm0,%xmm2 + DB 15,88,209 ; addps %xmm1,%xmm2 + DB 15,41,85,184 ; movaps %xmm2,-0x48(%ebp) + DB 15,92,251 ; subps %xmm3,%xmm7 + DB 15,89,248 ; mulps %xmm0,%xmm7 + DB 15,88,251 ; addps %xmm3,%xmm7 + DB 243,15,91,141,168,254,255,255 ; cvttps2dq -0x158(%ebp),%xmm1 + DB 102,15,112,217,245 ; pshufd $0xf5,%xmm1,%xmm3 + DB 102,15,111,133,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm0 + DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,244,233 ; pmuludq %xmm1,%xmm5 + DB 102,15,112,197,232 ; pshufd $0xe8,%xmm5,%xmm0 + DB 102,15,112,203,232 ; pshufd $0xe8,%xmm3,%xmm1 + DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,15,127,133,88,255,255,255 ; movdqa %xmm0,-0xa8(%ebp) + DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 + DB 102,15,254,208 ; paddd %xmm0,%xmm2 + DB 102,15,111,218 ; movdqa %xmm2,%xmm3 + DB 102,15,254,93,152 ; paddd -0x68(%ebp),%xmm3 + DB 102,15,112,203,245 ; pshufd $0xf5,%xmm3,%xmm1 + DB 102,15,244,220 ; pmuludq %xmm4,%xmm3 + DB 102,15,244,204 ; pmuludq %xmm4,%xmm1 + DB 102,15,112,233,232 ; pshufd $0xe8,%xmm1,%xmm5 + DB 102,15,112,203,232 ; pshufd $0xe8,%xmm3,%xmm1 + DB 102,15,98,205 ; punpckldq %xmm5,%xmm1 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,223 ; movd %xmm3,%edi + DB 102,15,112,217,231 ; pshufd $0xe7,%xmm1,%xmm3 + DB 102,15,126,219 ; movd %xmm3,%ebx + DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,198 ; unpcklps %xmm6,%xmm0 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 + DB 102,15,41,69,200 ; movapd %xmm0,-0x38(%ebp) + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,250,235 ; psubd %xmm3,%xmm5 + DB 102,15,112,245,229 ; pshufd $0xe5,%xmm5,%xmm6 DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,112,245,78 ; pshufd $0x4e,%xmm5,%xmm6 DB 102,15,126,246 ; movd %xmm6,%esi - DB 102,15,112,240,231 ; pshufd $0xe7,%xmm0,%xmm6 - DB 102,15,126,247 ; movd %xmm6,%edi - DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 - DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 - DB 15,20,254 ; unpcklps %xmm6,%xmm7 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 - DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,213 ; unpcklps %xmm5,%xmm2 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,254,233 ; paddd %xmm1,%xmm5 - DB 102,15,112,253,229 ; pshufd $0xe5,%xmm5,%xmm7 - DB 102,15,126,248 ; movd %xmm7,%eax - DB 102,15,112,253,78 ; pshufd $0x4e,%xmm5,%xmm7 - DB 102,15,126,254 ; movd %xmm7,%esi DB 102,15,126,239 ; movd %xmm5,%edi DB 102,15,112,237,231 ; pshufd $0xe7,%xmm5,%xmm5 DB 102,15,126,235 ; movd %xmm5,%ebx DB 243,15,16,44,154 ; movss (%edx,%ebx,4),%xmm5 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 - DB 15,20,205 ; unpcklps %xmm5,%xmm1 - DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 + DB 15,20,197 ; unpcklps %xmm5,%xmm0 + DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 + DB 243,15,16,44,130 ; movss (%edx,%eax,4),%xmm5 + DB 15,20,245 ; unpcklps %xmm5,%xmm6 + DB 102,15,20,240 ; unpcklpd %xmm0,%xmm6 + DB 102,15,254,137,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm1 + DB 102,15,112,193,229 ; pshufd $0xe5,%xmm1,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,193,78 ; pshufd $0x4e,%xmm1,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,126,207 ; movd %xmm1,%edi + DB 102,15,112,193,231 ; pshufd $0xe7,%xmm1,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,221 ; unpcklps %xmm5,%xmm3 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 102,15,111,177,255,81,0,0 ; movdqa 0x51ff(%ecx),%xmm6 - DB 102,15,254,198 ; paddd %xmm6,%xmm0 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,20,229 ; unpcklpd %xmm5,%xmm4 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,254,77,168 ; paddd -0x58(%ebp),%xmm1 + DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 + DB 102,15,111,145,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm2 + DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 + DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,112,233,232 ; pshufd $0xe8,%xmm1,%xmm5 + DB 102,15,98,232 ; punpckldq %xmm0,%xmm5 + DB 102,15,126,200 ; movd %xmm1,%eax + DB 102,15,112,197,229 ; pshufd $0xe5,%xmm5,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,197,231 ; pshufd $0xe7,%xmm5,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 + DB 15,20,193 ; unpcklps %xmm1,%xmm0 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 + DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,15,20,208 ; unpcklpd %xmm0,%xmm2 + DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,250,195 ; psubd %xmm3,%xmm0 DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 @@ -55128,69 +55189,97 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 243,15,16,60,186 ; movss (%edx,%edi,4),%xmm7 - DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,253 ; unpcklps %xmm5,%xmm7 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 - DB 15,91,77,200 ; cvtdq2ps -0x38(%ebp),%xmm1 - DB 15,92,225 ; subps %xmm1,%xmm4 - DB 15,41,101,152 ; movaps %xmm4,-0x68(%ebp) - DB 15,40,77,216 ; movaps -0x28(%ebp),%xmm1 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,216 ; unpcklps %xmm0,%xmm3 + DB 102,15,20,217 ; unpcklpd %xmm1,%xmm3 + DB 102,15,254,169,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm5 + DB 102,15,112,197,229 ; pshufd $0xe5,%xmm5,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,197,78 ; pshufd $0x4e,%xmm5,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,126,239 ; movd %xmm5,%edi + DB 102,15,112,197,231 ; pshufd $0xe7,%xmm5,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 102,15,20,233 ; unpcklpd %xmm1,%xmm5 + DB 15,40,77,200 ; movaps -0x38(%ebp),%xmm1 DB 15,92,209 ; subps %xmm1,%xmm2 - DB 15,89,212 ; mulps %xmm4,%xmm2 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 + DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,88,209 ; addps %xmm1,%xmm2 - DB 15,41,85,200 ; movaps %xmm2,-0x38(%ebp) - DB 15,40,141,136,254,255,255 ; movaps -0x178(%ebp),%xmm1 - DB 15,92,217 ; subps %xmm1,%xmm3 - DB 15,89,220 ; mulps %xmm4,%xmm3 - DB 15,88,217 ; addps %xmm1,%xmm3 - DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) - DB 15,40,141,152,254,255,255 ; movaps -0x168(%ebp),%xmm1 - DB 15,92,249 ; subps %xmm1,%xmm7 - DB 15,89,252 ; mulps %xmm4,%xmm7 - DB 15,88,249 ; addps %xmm1,%xmm7 - DB 243,15,91,141,168,254,255,255 ; cvttps2dq -0x158(%ebp),%xmm1 - DB 102,15,111,133,104,255,255,255 ; movdqa -0x98(%ebp),%xmm0 - DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 - DB 102,15,112,201,245 ; pshufd $0xf5,%xmm1,%xmm1 - DB 102,15,244,141,88,255,255,255 ; pmuludq -0xa8(%ebp),%xmm1 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,15,127,133,88,255,255,255 ; movdqa %xmm0,-0xa8(%ebp) + DB 15,92,222 ; subps %xmm6,%xmm3 + DB 15,89,216 ; mulps %xmm0,%xmm3 + DB 15,88,222 ; addps %xmm6,%xmm3 + DB 15,92,236 ; subps %xmm4,%xmm5 + DB 15,89,232 ; mulps %xmm0,%xmm5 + DB 15,88,236 ; addps %xmm4,%xmm5 + DB 15,91,133,40,255,255,255 ; cvtdq2ps -0xd8(%ebp),%xmm0 + DB 15,40,141,56,255,255,255 ; movaps -0xc8(%ebp),%xmm1 + DB 15,92,200 ; subps %xmm0,%xmm1 + DB 15,41,141,56,255,255,255 ; movaps %xmm1,-0xc8(%ebp) + DB 15,40,133,104,255,255,255 ; movaps -0x98(%ebp),%xmm0 + DB 15,92,208 ; subps %xmm0,%xmm2 + DB 15,89,209 ; mulps %xmm1,%xmm2 + DB 15,88,208 ; addps %xmm0,%xmm2 + DB 15,41,149,40,255,255,255 ; movaps %xmm2,-0xd8(%ebp) + DB 15,40,69,184 ; movaps -0x48(%ebp),%xmm0 + DB 15,92,216 ; subps %xmm0,%xmm3 + DB 15,89,217 ; mulps %xmm1,%xmm3 + DB 15,88,216 ; addps %xmm0,%xmm3 + DB 15,41,157,104,255,255,255 ; movaps %xmm3,-0x98(%ebp) + DB 15,92,239 ; subps %xmm7,%xmm5 + DB 15,89,233 ; mulps %xmm1,%xmm5 + DB 15,88,239 ; addps %xmm7,%xmm5 + DB 15,41,109,184 ; movaps %xmm5,-0x48(%ebp) + DB 243,15,91,69,136 ; cvttps2dq -0x78(%ebp),%xmm0 + DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 DB 102,15,111,149,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm2 - DB 102,15,254,208 ; paddd %xmm0,%xmm2 - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,15,254,77,168 ; paddd -0x58(%ebp),%xmm1 - DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 - DB 102,15,111,129,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm0 - DB 102,15,244,200 ; pmuludq %xmm0,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,98,202 ; punpckldq %xmm2,%xmm1 - DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 + DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 + DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 + DB 102,15,98,224 ; punpckldq %xmm0,%xmm4 + DB 102,15,127,165,24,255,255,255 ; movdqa %xmm4,-0xe8(%ebp) + DB 102,15,254,165,184,254,255,255 ; paddd -0x148(%ebp),%xmm4 + DB 102,15,127,101,200 ; movdqa %xmm4,-0x38(%ebp) + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,254,141,72,255,255,255 ; paddd -0xb8(%ebp),%xmm1 + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,15,254,85,152 ; paddd -0x68(%ebp),%xmm2 + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,153,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm3 + DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 + DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,15,112,216,232 ; pshufd $0xe8,%xmm0,%xmm3 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 + DB 102,15,98,195 ; punpckldq %xmm3,%xmm0 DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 + DB 102,15,112,208,229 ; pshufd $0xe5,%xmm0,%xmm2 DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,112,209,231 ; pshufd $0xe7,%xmm1,%xmm2 + DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 DB 102,15,126,215 ; movd %xmm2,%edi - DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 102,15,112,208,231 ; pshufd $0xe7,%xmm0,%xmm2 + DB 102,15,126,211 ; movd %xmm2,%ebx + DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,41,133,104,255,255,255 ; movaps %xmm0,-0x98(%ebp) - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,111,129,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm0 - DB 102,15,254,208 ; paddd %xmm0,%xmm2 - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 15,20,226 ; unpcklps %xmm2,%xmm4 + DB 102,15,20,227 ; unpcklpd %xmm3,%xmm4 + DB 102,15,41,101,136 ; movapd %xmm4,-0x78(%ebp) + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,118,246 ; pcmpeqd %xmm6,%xmm6 + DB 102,15,250,214 ; psubd %xmm6,%xmm2 DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 DB 102,15,126,216 ; movd %xmm3,%eax DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 @@ -55199,53 +55288,13 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 DB 102,15,126,211 ; movd %xmm2,%ebx DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,41,133,24,255,255,255 ; movaps %xmm0,-0xe8(%ebp) - DB 102,15,254,206 ; paddd %xmm6,%xmm1 - DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,126,207 ; movd %xmm1,%edi - DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 - DB 102,15,126,203 ; movd %xmm1,%ebx - DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 - DB 15,20,243 ; unpcklps %xmm3,%xmm6 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 - DB 102,15,254,69,184 ; paddd -0x48(%ebp),%xmm0 - DB 102,15,112,216,245 ; pshufd $0xf5,%xmm0,%xmm3 - DB 102,15,111,137,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm1 - DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 - DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 - DB 102,15,244,217 ; pmuludq %xmm1,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,15,98,211 ; punpckldq %xmm3,%xmm2 - DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 - DB 102,15,126,222 ; movd %xmm3,%esi - DB 102,15,112,218,231 ; pshufd $0xe7,%xmm2,%xmm3 - DB 102,15,126,223 ; movd %xmm3,%edi - DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 102,15,126,208 ; movd %xmm2,%eax + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 DB 243,15,16,44,130 ; movss (%edx,%eax,4),%xmm5 - DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 - DB 15,20,235 ; unpcklps %xmm3,%xmm5 - DB 15,20,232 ; unpcklps %xmm0,%xmm5 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,254,196 ; paddd %xmm4,%xmm0 + DB 15,20,229 ; unpcklps %xmm5,%xmm4 + DB 102,15,20,227 ; unpcklpd %xmm3,%xmm4 + DB 102,15,254,129,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm0 DB 102,15,112,216,229 ; pshufd $0xe5,%xmm0,%xmm3 DB 102,15,126,216 ; movd %xmm3,%eax DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 @@ -55254,240 +55303,123 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 - DB 15,20,224 ; unpcklps %xmm0,%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 102,15,111,153,255,81,0,0 ; movdqa 0x51ff(%ecx),%xmm3 - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 - DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,126,215 ; movd %xmm2,%edi - DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 - DB 102,15,126,195 ; movd %xmm0,%ebx - DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 - DB 15,20,208 ; unpcklps %xmm0,%xmm2 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 15,40,205 ; movaps %xmm5,%xmm1 - DB 15,40,173,104,255,255,255 ; movaps -0x98(%ebp),%xmm5 - DB 15,92,205 ; subps %xmm5,%xmm1 - DB 15,40,69,152 ; movaps -0x68(%ebp),%xmm0 - DB 15,89,200 ; mulps %xmm0,%xmm1 - DB 15,88,205 ; addps %xmm5,%xmm1 - DB 15,40,233 ; movaps %xmm1,%xmm5 - DB 15,40,141,24,255,255,255 ; movaps -0xe8(%ebp),%xmm1 - DB 15,92,225 ; subps %xmm1,%xmm4 - DB 15,89,224 ; mulps %xmm0,%xmm4 - DB 15,88,225 ; addps %xmm1,%xmm4 - DB 15,92,214 ; subps %xmm6,%xmm2 - DB 15,89,208 ; mulps %xmm0,%xmm2 - DB 15,88,214 ; addps %xmm6,%xmm2 - DB 15,91,133,8,255,255,255 ; cvtdq2ps -0xf8(%ebp),%xmm0 - DB 15,40,141,40,255,255,255 ; movaps -0xd8(%ebp),%xmm1 - DB 15,92,200 ; subps %xmm0,%xmm1 - DB 15,41,141,40,255,255,255 ; movaps %xmm1,-0xd8(%ebp) - DB 15,40,69,200 ; movaps -0x38(%ebp),%xmm0 - DB 15,92,232 ; subps %xmm0,%xmm5 - DB 15,89,233 ; mulps %xmm1,%xmm5 - DB 15,88,232 ; addps %xmm0,%xmm5 - DB 15,41,173,8,255,255,255 ; movaps %xmm5,-0xf8(%ebp) - DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 - DB 15,92,224 ; subps %xmm0,%xmm4 - DB 15,89,225 ; mulps %xmm1,%xmm4 - DB 15,88,224 ; addps %xmm0,%xmm4 - DB 15,41,165,104,255,255,255 ; movaps %xmm4,-0x98(%ebp) - DB 15,92,215 ; subps %xmm7,%xmm2 - DB 15,89,209 ; mulps %xmm1,%xmm2 - DB 15,88,215 ; addps %xmm7,%xmm2 - DB 15,41,85,200 ; movaps %xmm2,-0x38(%ebp) - DB 243,15,91,69,136 ; cvttps2dq -0x78(%ebp),%xmm0 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,149,248,254,255,255 ; movdqa -0x108(%ebp),%xmm2 - DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 - DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 - DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 - DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 - DB 102,15,127,149,248,254,255,255 ; movdqa %xmm2,-0x108(%ebp) - DB 102,15,254,149,184,254,255,255 ; paddd -0x148(%ebp),%xmm2 - DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) - DB 102,15,111,234 ; movdqa %xmm2,%xmm5 - DB 102,15,254,173,72,255,255,255 ; paddd -0xb8(%ebp),%xmm5 - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 - DB 102,15,254,69,168 ; paddd -0x58(%ebp),%xmm0 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,145,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm2 - DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 - DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 243,15,16,44,130 ; movss (%edx,%eax,4),%xmm5 + DB 15,20,213 ; unpcklps %xmm5,%xmm2 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 102,15,254,77,168 ; paddd -0x58(%ebp),%xmm1 + DB 102,15,112,217,245 ; pshufd $0xf5,%xmm1,%xmm3 + DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 + DB 102,15,244,223 ; pmuludq %xmm7,%xmm3 + DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 + DB 102,15,112,249,232 ; pshufd $0xe8,%xmm1,%xmm7 + DB 102,15,98,251 ; punpckldq %xmm3,%xmm7 DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,112,207,229 ; pshufd $0xe5,%xmm7,%xmm1 DB 102,15,126,206 ; movd %xmm1,%esi - DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,112,207,78 ; pshufd $0x4e,%xmm7,%xmm1 DB 102,15,126,207 ; movd %xmm1,%edi - DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 + DB 102,15,112,207,231 ; pshufd $0xe7,%xmm7,%xmm1 + DB 102,15,126,203 ; movd %xmm1,%ebx + DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 - DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 15,41,85,136 ; movaps %xmm2,-0x78(%ebp) - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,111,145,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm2 - DB 102,15,254,202 ; paddd %xmm2,%xmm1 - DB 102,15,112,241,229 ; pshufd $0xe5,%xmm1,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 - DB 102,15,126,246 ; movd %xmm6,%esi + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,193 ; unpcklps %xmm1,%xmm0 + DB 102,15,20,195 ; unpcklpd %xmm3,%xmm0 + DB 102,15,111,207 ; movdqa %xmm7,%xmm1 + DB 102,15,250,206 ; psubd %xmm6,%xmm1 + DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi DB 102,15,126,207 ; movd %xmm1,%edi DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,203 ; movd %xmm1,%ebx DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 - DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 - DB 15,20,231 ; unpcklps %xmm7,%xmm4 - DB 15,20,230 ; unpcklps %xmm6,%xmm4 - DB 102,15,254,195 ; paddd %xmm3,%xmm0 - DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 - DB 102,15,126,246 ; movd %xmm6,%esi - DB 102,15,126,199 ; movd %xmm0,%edi - DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 - DB 102,15,126,195 ; movd %xmm0,%ebx - DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 102,15,254,109,184 ; paddd -0x48(%ebp),%xmm5 - DB 102,15,112,245,245 ; pshufd $0xf5,%xmm5,%xmm6 - DB 102,15,111,129,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm0 - DB 102,15,244,232 ; pmuludq %xmm0,%xmm5 - DB 102,15,112,253,232 ; pshufd $0xe8,%xmm5,%xmm7 - DB 102,15,244,240 ; pmuludq %xmm0,%xmm6 - DB 102,15,112,238,232 ; pshufd $0xe8,%xmm6,%xmm5 - DB 102,15,98,253 ; punpckldq %xmm5,%xmm7 - DB 102,15,112,239,229 ; pshufd $0xe5,%xmm7,%xmm5 - DB 102,15,126,232 ; movd %xmm5,%eax - DB 102,15,112,239,78 ; pshufd $0x4e,%xmm7,%xmm5 - DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,112,239,231 ; pshufd $0xe7,%xmm7,%xmm5 - DB 102,15,126,239 ; movd %xmm5,%edi + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,245 ; unpcklps %xmm5,%xmm6 - DB 102,15,126,248 ; movd %xmm7,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,20,198 ; unpcklps %xmm6,%xmm0 - DB 102,15,111,223 ; movdqa %xmm7,%xmm3 - DB 102,15,254,218 ; paddd %xmm2,%xmm3 - DB 102,15,112,243,229 ; pshufd $0xe5,%xmm3,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,243,78 ; pshufd $0x4e,%xmm3,%xmm6 - DB 102,15,126,246 ; movd %xmm6,%esi - DB 102,15,126,223 ; movd %xmm3,%edi - DB 102,15,112,219,231 ; pshufd $0xe7,%xmm3,%xmm3 + DB 15,20,238 ; unpcklps %xmm6,%xmm5 + DB 102,15,20,235 ; unpcklpd %xmm3,%xmm5 + DB 102,15,111,137,201,80,0,0 ; movdqa 0x50c9(%ecx),%xmm1 + DB 102,15,254,249 ; paddd %xmm1,%xmm7 + DB 102,15,112,223,229 ; pshufd $0xe5,%xmm7,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,223,78 ; pshufd $0x4e,%xmm7,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,126,255 ; movd %xmm7,%edi + DB 102,15,112,223,231 ; pshufd $0xe7,%xmm7,%xmm3 DB 102,15,126,219 ; movd %xmm3,%ebx DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 DB 15,20,243 ; unpcklps %xmm3,%xmm6 - DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 102,15,111,153,255,81,0,0 ; movdqa 0x51ff(%ecx),%xmm3 - DB 102,15,254,251 ; paddd %xmm3,%xmm7 - DB 102,15,112,215,229 ; pshufd $0xe5,%xmm7,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,215,78 ; pshufd $0x4e,%xmm7,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,126,255 ; movd %xmm7,%edi - DB 102,15,112,215,231 ; pshufd $0xe7,%xmm7,%xmm2 - DB 102,15,126,211 ; movd %xmm2,%ebx - DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 DB 243,15,16,60,186 ; movss (%edx,%edi,4),%xmm7 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 - DB 15,20,254 ; unpcklps %xmm6,%xmm7 + DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 15,20,251 ; unpcklps %xmm3,%xmm7 + DB 102,15,20,254 ; unpcklpd %xmm6,%xmm7 DB 15,40,117,136 ; movaps -0x78(%ebp),%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 - DB 15,40,85,152 ; movaps -0x68(%ebp),%xmm2 - DB 15,89,194 ; mulps %xmm2,%xmm0 + DB 15,40,93,216 ; movaps -0x28(%ebp),%xmm3 + DB 15,89,195 ; mulps %xmm3,%xmm0 DB 15,88,198 ; addps %xmm6,%xmm0 DB 15,41,69,136 ; movaps %xmm0,-0x78(%ebp) DB 15,92,236 ; subps %xmm4,%xmm5 - DB 15,89,234 ; mulps %xmm2,%xmm5 + DB 15,89,235 ; mulps %xmm3,%xmm5 DB 15,88,236 ; addps %xmm4,%xmm5 - DB 15,92,249 ; subps %xmm1,%xmm7 - DB 15,89,250 ; mulps %xmm2,%xmm7 - DB 15,88,249 ; addps %xmm1,%xmm7 - DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 - DB 102,15,254,141,88,255,255,255 ; paddd -0xa8(%ebp),%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,254,69,168 ; paddd -0x58(%ebp),%xmm0 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,177,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm6 - DB 102,15,244,198 ; pmuludq %xmm6,%xmm0 + DB 15,92,250 ; subps %xmm2,%xmm7 + DB 15,89,251 ; mulps %xmm3,%xmm7 + DB 15,88,250 ; addps %xmm2,%xmm7 + DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 + DB 102,15,254,149,88,255,255,255 ; paddd -0xa8(%ebp),%xmm2 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,15,254,69,152 ; paddd -0x68(%ebp),%xmm0 + DB 102,15,112,208,245 ; pshufd $0xf5,%xmm0,%xmm2 + DB 102,15,111,153,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm3 + DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 + DB 102,15,244,211 ; pmuludq %xmm3,%xmm2 + DB 102,15,112,218,232 ; pshufd $0xe8,%xmm2,%xmm3 DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 - DB 102,15,244,206 ; pmuludq %xmm6,%xmm1 - DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 - DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 - DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 + DB 102,15,98,211 ; punpckldq %xmm3,%xmm2 DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 - DB 15,20,198 ; unpcklps %xmm6,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,15,254,137,223,81,0,0 ; paddd 0x51df(%ecx),%xmm1 - DB 102,15,112,241,229 ; pshufd $0xe5,%xmm1,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,241,78 ; pshufd $0x4e,%xmm1,%xmm6 - DB 102,15,126,246 ; movd %xmm6,%esi - DB 102,15,126,207 ; movd %xmm1,%edi - DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 - DB 102,15,126,203 ; movd %xmm1,%ebx - DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 15,20,216 ; unpcklps %xmm0,%xmm3 + DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,241 ; unpcklps %xmm1,%xmm6 - DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 102,15,20,243 ; unpcklpd %xmm3,%xmm6 + DB 102,15,41,117,200 ; movapd %xmm6,-0x38(%ebp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,250,195 ; psubd %xmm3,%xmm0 + DB 102,15,112,216,229 ; pshufd $0xe5,%xmm0,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 15,41,141,24,255,255,255 ; movaps %xmm1,-0xe8(%ebp) - DB 102,15,254,211 ; paddd %xmm3,%xmm2 + DB 15,20,195 ; unpcklps %xmm3,%xmm0 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 + DB 15,20,222 ; unpcklps %xmm6,%xmm3 + DB 102,15,20,216 ; unpcklpd %xmm0,%xmm3 + DB 102,15,41,157,232,254,255,255 ; movapd %xmm3,-0x118(%ebp) + DB 102,15,254,209 ; paddd %xmm1,%xmm2 DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 @@ -55496,72 +55428,73 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 DB 15,20,240 ; unpcklps %xmm0,%xmm6 DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 15,20,222 ; unpcklps %xmm6,%xmm3 + DB 102,15,20,222 ; unpcklpd %xmm6,%xmm3 DB 102,15,111,212 ; movdqa %xmm4,%xmm2 - DB 102,15,254,85,184 ; paddd -0x48(%ebp),%xmm2 + DB 102,15,254,85,168 ; paddd -0x58(%ebp),%xmm2 DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 - DB 102,15,111,137,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm1 + DB 102,15,111,137,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm1 DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 - DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,98,224 ; punpckldq %xmm0,%xmm4 - DB 102,15,112,196,229 ; pshufd $0xe5,%xmm4,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,196,78 ; pshufd $0x4e,%xmm4,%xmm0 - DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,112,196,231 ; pshufd $0xe7,%xmm4,%xmm0 - DB 102,15,126,199 ; movd %xmm0,%edi - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 102,15,126,224 ; movd %xmm4,%eax + DB 102,15,112,240,232 ; pshufd $0xe8,%xmm0,%xmm6 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 + DB 102,15,98,198 ; punpckldq %xmm6,%xmm0 + DB 102,15,126,208 ; movd %xmm2,%eax + DB 102,15,112,224,229 ; pshufd $0xe5,%xmm0,%xmm4 + DB 102,15,126,230 ; movd %xmm4,%esi + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,126,231 ; movd %xmm4,%edi + DB 102,15,112,224,231 ; pshufd $0xe7,%xmm0,%xmm4 + DB 102,15,126,227 ; movd %xmm4,%ebx + DB 243,15,16,36,154 ; movss (%edx,%ebx,4),%xmm4 + DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,254,129,223,81,0,0 ; paddd 0x51df(%ecx),%xmm0 - DB 102,15,112,240,229 ; pshufd $0xe5,%xmm0,%xmm6 + DB 15,20,204 ; unpcklps %xmm4,%xmm1 + DB 102,15,20,206 ; unpcklpd %xmm6,%xmm1 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 + DB 102,15,250,226 ; psubd %xmm2,%xmm4 + DB 102,15,112,244,229 ; pshufd $0xe5,%xmm4,%xmm6 DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,240,78 ; pshufd $0x4e,%xmm0,%xmm6 + DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,126,231 ; movd %xmm4,%edi + DB 102,15,112,228,231 ; pshufd $0xe7,%xmm4,%xmm4 + DB 102,15,126,227 ; movd %xmm4,%ebx + DB 243,15,16,36,154 ; movss (%edx,%ebx,4),%xmm4 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 + DB 15,20,244 ; unpcklps %xmm4,%xmm6 + DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 15,20,212 ; unpcklps %xmm4,%xmm2 + DB 102,15,20,214 ; unpcklpd %xmm6,%xmm2 + DB 102,15,254,129,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm0 + DB 102,15,112,224,229 ; pshufd $0xe5,%xmm0,%xmm4 + DB 102,15,126,224 ; movd %xmm4,%eax + DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 + DB 102,15,126,230 ; movd %xmm4,%esi DB 102,15,126,199 ; movd %xmm0,%edi DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 - DB 15,20,208 ; unpcklps %xmm0,%xmm2 - DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 102,15,254,161,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm4 - DB 102,15,112,196,229 ; pshufd $0xe5,%xmm4,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,196,78 ; pshufd $0x4e,%xmm4,%xmm0 - DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,126,231 ; movd %xmm4,%edi - DB 102,15,112,196,231 ; pshufd $0xe7,%xmm4,%xmm0 - DB 102,15,126,195 ; movd %xmm0,%ebx - DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 DB 15,20,224 ; unpcklps %xmm0,%xmm4 DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 15,20,240 ; unpcklps %xmm0,%xmm6 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 15,40,101,216 ; movaps -0x28(%ebp),%xmm4 + DB 102,15,20,244 ; unpcklpd %xmm4,%xmm6 + DB 15,40,101,200 ; movaps -0x38(%ebp),%xmm4 DB 15,92,204 ; subps %xmm4,%xmm1 - DB 15,40,69,152 ; movaps -0x68(%ebp),%xmm0 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 15,88,204 ; addps %xmm4,%xmm1 - DB 15,40,165,24,255,255,255 ; movaps -0xe8(%ebp),%xmm4 + DB 15,40,165,232,254,255,255 ; movaps -0x118(%ebp),%xmm4 DB 15,92,212 ; subps %xmm4,%xmm2 DB 15,89,208 ; mulps %xmm0,%xmm2 DB 15,88,212 ; addps %xmm4,%xmm2 @@ -55570,7 +55503,7 @@ _sk_clut_4D_sse2 LABEL PROC DB 15,88,243 ; addps %xmm3,%xmm6 DB 15,40,93,136 ; movaps -0x78(%ebp),%xmm3 DB 15,92,203 ; subps %xmm3,%xmm1 - DB 15,40,133,40,255,255,255 ; movaps -0xd8(%ebp),%xmm0 + DB 15,40,133,56,255,255,255 ; movaps -0xc8(%ebp),%xmm0 DB 15,89,200 ; mulps %xmm0,%xmm1 DB 15,88,203 ; addps %xmm3,%xmm1 DB 15,40,217 ; movaps %xmm1,%xmm3 @@ -55581,250 +55514,253 @@ _sk_clut_4D_sse2 LABEL PROC DB 15,89,240 ; mulps %xmm0,%xmm6 DB 15,88,247 ; addps %xmm7,%xmm6 DB 15,91,133,216,254,255,255 ; cvtdq2ps -0x128(%ebp),%xmm0 - DB 15,40,141,232,254,255,255 ; movaps -0x118(%ebp),%xmm1 + DB 15,40,141,248,254,255,255 ; movaps -0x108(%ebp),%xmm1 DB 15,92,200 ; subps %xmm0,%xmm1 - DB 15,41,141,232,254,255,255 ; movaps %xmm1,-0x118(%ebp) - DB 15,40,133,8,255,255,255 ; movaps -0xf8(%ebp),%xmm0 + DB 15,41,141,248,254,255,255 ; movaps %xmm1,-0x108(%ebp) + DB 15,40,133,40,255,255,255 ; movaps -0xd8(%ebp),%xmm0 DB 15,92,216 ; subps %xmm0,%xmm3 DB 15,89,217 ; mulps %xmm1,%xmm3 DB 15,88,216 ; addps %xmm0,%xmm3 - DB 15,41,157,8,255,255,255 ; movaps %xmm3,-0xf8(%ebp) + DB 15,41,157,40,255,255,255 ; movaps %xmm3,-0xd8(%ebp) DB 15,40,133,104,255,255,255 ; movaps -0x98(%ebp),%xmm0 DB 15,92,208 ; subps %xmm0,%xmm2 DB 15,89,209 ; mulps %xmm1,%xmm2 DB 15,88,208 ; addps %xmm0,%xmm2 DB 15,41,149,104,255,255,255 ; movaps %xmm2,-0x98(%ebp) - DB 15,40,69,200 ; movaps -0x38(%ebp),%xmm0 + DB 15,40,69,184 ; movaps -0x48(%ebp),%xmm0 DB 15,92,240 ; subps %xmm0,%xmm6 DB 15,89,241 ; mulps %xmm1,%xmm6 DB 15,88,240 ; addps %xmm0,%xmm6 DB 15,41,181,216,254,255,255 ; movaps %xmm6,-0x128(%ebp) DB 243,15,91,141,120,255,255,255 ; cvttps2dq -0x88(%ebp),%xmm1 DB 102,15,127,77,136 ; movdqa %xmm1,-0x78(%ebp) - DB 102,15,111,133,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm0 + DB 102,15,111,133,8,255,255,255 ; movdqa -0xf8(%ebp),%xmm0 DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,127,133,56,255,255,255 ; movdqa %xmm0,-0xc8(%ebp) + DB 102,15,127,133,8,255,255,255 ; movdqa %xmm0,-0xf8(%ebp) DB 102,15,111,141,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm1 DB 102,15,254,200 ; paddd %xmm0,%xmm1 - DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 - DB 102,15,254,193 ; paddd %xmm1,%xmm0 - DB 102,15,112,224,245 ; pshufd $0xf5,%xmm0,%xmm4 - DB 102,15,111,145,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm2 - DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 - DB 102,15,112,216,232 ; pshufd $0xe8,%xmm0,%xmm3 - DB 102,15,244,226 ; pmuludq %xmm2,%xmm4 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,112,196,232 ; pshufd $0xe8,%xmm4,%xmm0 - DB 102,15,98,216 ; punpckldq %xmm0,%xmm3 - DB 102,15,112,195,229 ; pshufd $0xe5,%xmm3,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 - DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,112,195,231 ; pshufd $0xe7,%xmm3,%xmm0 - DB 102,15,126,199 ; movd %xmm0,%edi - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 - DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,111,93,152 ; movdqa -0x68(%ebp),%xmm3 + DB 102,15,254,217 ; paddd %xmm1,%xmm3 + DB 102,15,112,211,245 ; pshufd $0xf5,%xmm3,%xmm2 + DB 102,15,111,129,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm0 + DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 + DB 102,15,244,208 ; pmuludq %xmm0,%xmm2 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,112,226,232 ; pshufd $0xe8,%xmm2,%xmm4 + DB 102,15,112,211,232 ; pshufd $0xe8,%xmm3,%xmm2 + DB 102,15,98,212 ; punpckldq %xmm4,%xmm2 DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,126,223 ; movd %xmm3,%edi + DB 102,15,112,218,231 ; pshufd $0xe7,%xmm2,%xmm3 + DB 102,15,126,219 ; movd %xmm3,%ebx + DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,197 ; unpcklps %xmm5,%xmm0 - DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 15,41,133,120,255,255,255 ; movaps %xmm0,-0x88(%ebp) - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,111,145,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm2 - DB 102,15,254,226 ; paddd %xmm2,%xmm4 - DB 102,15,112,236,229 ; pshufd $0xe5,%xmm4,%xmm5 + DB 15,20,195 ; unpcklps %xmm3,%xmm0 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 + DB 102,15,41,133,120,255,255,255 ; movapd %xmm0,-0x88(%ebp) + DB 102,15,111,218 ; movdqa %xmm2,%xmm3 + DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 + DB 102,15,250,216 ; psubd %xmm0,%xmm3 + DB 102,15,112,235,229 ; pshufd $0xe5,%xmm3,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax - DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 + DB 102,15,112,235,78 ; pshufd $0x4e,%xmm3,%xmm5 DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,126,231 ; movd %xmm4,%edi - DB 102,15,112,228,231 ; pshufd $0xe7,%xmm4,%xmm4 - DB 102,15,126,227 ; movd %xmm4,%ebx - DB 243,15,16,36,154 ; movss (%edx,%ebx,4),%xmm4 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 - DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 102,15,254,153,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm3 - DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 - DB 102,15,126,230 ; movd %xmm4,%esi DB 102,15,126,223 ; movd %xmm3,%edi DB 102,15,112,219,231 ; pshufd $0xe7,%xmm3,%xmm3 DB 102,15,126,219 ; movd %xmm3,%ebx DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 - DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 - DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 DB 15,20,243 ; unpcklps %xmm3,%xmm6 - DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 102,15,254,77,184 ; paddd -0x48(%ebp),%xmm1 - DB 102,15,112,217,245 ; pshufd $0xf5,%xmm1,%xmm3 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 + DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 + DB 102,15,254,145,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm2 + DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,126,215 ; movd %xmm2,%edi + DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 + DB 102,15,126,211 ; movd %xmm2,%ebx + DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 + DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 + DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 + DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 102,15,20,243 ; unpcklpd %xmm3,%xmm6 + DB 102,15,254,77,168 ; paddd -0x58(%ebp),%xmm1 + DB 102,15,112,209,245 ; pshufd $0xf5,%xmm1,%xmm2 DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 - DB 102,15,112,225,232 ; pshufd $0xe8,%xmm1,%xmm4 - DB 102,15,244,223 ; pmuludq %xmm7,%xmm3 - DB 102,15,112,203,232 ; pshufd $0xe8,%xmm3,%xmm1 - DB 102,15,98,225 ; punpckldq %xmm1,%xmm4 - DB 102,15,112,204,229 ; pshufd $0xe5,%xmm4,%xmm1 + DB 102,15,244,215 ; pmuludq %xmm7,%xmm2 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 + DB 102,15,112,217,232 ; pshufd $0xe8,%xmm1,%xmm3 + DB 102,15,98,218 ; punpckldq %xmm2,%xmm3 DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,204,78 ; pshufd $0x4e,%xmm4,%xmm1 + DB 102,15,112,203,229 ; pshufd $0xe5,%xmm3,%xmm1 DB 102,15,126,206 ; movd %xmm1,%esi - DB 102,15,112,204,231 ; pshufd $0xe7,%xmm4,%xmm1 + DB 102,15,112,203,78 ; pshufd $0x4e,%xmm3,%xmm1 DB 102,15,126,207 ; movd %xmm1,%edi - DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 102,15,112,203,231 ; pshufd $0xe7,%xmm3,%xmm1 + DB 102,15,126,203 ; movd %xmm1,%ebx + DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 102,15,111,220 ; movdqa %xmm4,%xmm3 - DB 102,15,254,218 ; paddd %xmm2,%xmm3 - DB 102,15,112,251,229 ; pshufd $0xe5,%xmm3,%xmm7 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 15,20,231 ; unpcklps %xmm7,%xmm4 + DB 102,15,20,226 ; unpcklpd %xmm2,%xmm4 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,250,208 ; psubd %xmm0,%xmm2 + DB 102,15,112,250,229 ; pshufd $0xe5,%xmm2,%xmm7 DB 102,15,126,248 ; movd %xmm7,%eax - DB 102,15,112,251,78 ; pshufd $0x4e,%xmm3,%xmm7 + DB 102,15,112,250,78 ; pshufd $0x4e,%xmm2,%xmm7 DB 102,15,126,254 ; movd %xmm7,%esi - DB 102,15,126,223 ; movd %xmm3,%edi - DB 102,15,112,219,231 ; pshufd $0xe7,%xmm3,%xmm3 - DB 102,15,126,219 ; movd %xmm3,%ebx - DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 - DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 - DB 15,20,251 ; unpcklps %xmm3,%xmm7 - DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,20,223 ; unpcklps %xmm7,%xmm3 - DB 102,15,254,161,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm4 - DB 102,15,112,212,229 ; pshufd $0xe5,%xmm4,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,212,78 ; pshufd $0x4e,%xmm4,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,126,231 ; movd %xmm4,%edi - DB 102,15,112,212,231 ; pshufd $0xe7,%xmm4,%xmm2 + DB 102,15,126,215 ; movd %xmm2,%edi + DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 DB 102,15,126,211 ; movd %xmm2,%ebx DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 + DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 DB 15,20,250 ; unpcklps %xmm2,%xmm7 - DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,226 ; unpcklps %xmm2,%xmm4 - DB 15,20,231 ; unpcklps %xmm7,%xmm4 - DB 15,40,149,120,255,255,255 ; movaps -0x88(%ebp),%xmm2 - DB 15,92,194 ; subps %xmm2,%xmm0 - DB 15,40,77,152 ; movaps -0x68(%ebp),%xmm1 - DB 15,89,193 ; mulps %xmm1,%xmm0 - DB 15,88,194 ; addps %xmm2,%xmm0 - DB 15,41,133,120,255,255,255 ; movaps %xmm0,-0x88(%ebp) - DB 15,92,221 ; subps %xmm5,%xmm3 - DB 15,89,217 ; mulps %xmm1,%xmm3 - DB 15,88,221 ; addps %xmm5,%xmm3 - DB 15,41,93,200 ; movaps %xmm3,-0x38(%ebp) - DB 15,92,230 ; subps %xmm6,%xmm4 - DB 15,89,225 ; mulps %xmm1,%xmm4 - DB 15,88,230 ; addps %xmm6,%xmm4 - DB 102,15,111,133,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm0 - DB 102,15,254,133,88,255,255,255 ; paddd -0xa8(%ebp),%xmm0 - DB 102,15,127,133,56,255,255,255 ; movdqa %xmm0,-0xc8(%ebp) - DB 102,15,254,69,168 ; paddd -0x58(%ebp),%xmm0 - DB 102,15,112,208,245 ; pshufd $0xf5,%xmm0,%xmm2 - DB 102,15,111,137,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm1 - DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 + DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 102,15,20,215 ; unpcklpd %xmm7,%xmm2 + DB 102,15,111,137,201,80,0,0 ; movdqa 0x50c9(%ecx),%xmm1 + DB 102,15,254,217 ; paddd %xmm1,%xmm3 + DB 102,15,112,195,229 ; pshufd $0xe5,%xmm3,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,195,78 ; pshufd $0x4e,%xmm3,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,126,223 ; movd %xmm3,%edi + DB 102,15,112,195,231 ; pshufd $0xe7,%xmm3,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 + DB 15,20,248 ; unpcklps %xmm0,%xmm7 + DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,216 ; unpcklps %xmm0,%xmm3 + DB 102,15,20,223 ; unpcklpd %xmm7,%xmm3 + DB 15,40,189,120,255,255,255 ; movaps -0x88(%ebp),%xmm7 + DB 15,92,231 ; subps %xmm7,%xmm4 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 + DB 15,89,224 ; mulps %xmm0,%xmm4 + DB 15,88,231 ; addps %xmm7,%xmm4 + DB 15,41,165,120,255,255,255 ; movaps %xmm4,-0x88(%ebp) + DB 15,92,213 ; subps %xmm5,%xmm2 + DB 15,89,208 ; mulps %xmm0,%xmm2 + DB 15,88,213 ; addps %xmm5,%xmm2 + DB 15,41,85,200 ; movaps %xmm2,-0x38(%ebp) + DB 15,92,222 ; subps %xmm6,%xmm3 + DB 15,89,216 ; mulps %xmm0,%xmm3 + DB 15,88,222 ; addps %xmm6,%xmm3 + DB 15,41,93,184 ; movaps %xmm3,-0x48(%ebp) + DB 102,15,111,149,8,255,255,255 ; movdqa -0xf8(%ebp),%xmm2 + DB 102,15,254,149,88,255,255,255 ; paddd -0xa8(%ebp),%xmm2 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,254,69,152 ; paddd -0x68(%ebp),%xmm0 + DB 102,15,112,224,245 ; pshufd $0xf5,%xmm0,%xmm4 + DB 102,15,111,169,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm5 + DB 102,15,111,221 ; movdqa %xmm5,%xmm3 + DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 + DB 102,15,244,227 ; pmuludq %xmm3,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,112,240,232 ; pshufd $0xe8,%xmm0,%xmm6 - DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 - DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 - DB 102,15,98,240 ; punpckldq %xmm0,%xmm6 - DB 102,15,112,198,229 ; pshufd $0xe5,%xmm6,%xmm0 + DB 102,15,98,244 ; punpckldq %xmm4,%xmm6 DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,198,78 ; pshufd $0x4e,%xmm6,%xmm0 + DB 102,15,112,198,229 ; pshufd $0xe5,%xmm6,%xmm0 DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,112,198,231 ; pshufd $0xe7,%xmm6,%xmm0 + DB 102,15,112,198,78 ; pshufd $0x4e,%xmm6,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 - DB 15,20,208 ; unpcklps %xmm0,%xmm2 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 243,15,16,44,178 ; movss (%edx,%esi,4),%xmm5 - DB 15,20,197 ; unpcklps %xmm5,%xmm0 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 15,41,69,216 ; movaps %xmm0,-0x28(%ebp) - DB 102,15,111,214 ; movdqa %xmm6,%xmm2 - DB 102,15,111,153,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm3 - DB 102,15,254,211 ; paddd %xmm3,%xmm2 - DB 102,15,112,234,229 ; pshufd $0xe5,%xmm2,%xmm5 + DB 102,15,112,198,231 ; pshufd $0xe7,%xmm6,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,20,229 ; unpcklpd %xmm5,%xmm4 + DB 102,15,111,198 ; movdqa %xmm6,%xmm0 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,250,195 ; psubd %xmm3,%xmm0 + DB 102,15,112,232,229 ; pshufd $0xe5,%xmm0,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax - DB 102,15,112,234,78 ; pshufd $0x4e,%xmm2,%xmm5 + DB 102,15,112,232,78 ; pshufd $0x4e,%xmm0,%xmm5 DB 102,15,126,238 ; movd %xmm5,%esi - DB 102,15,126,215 ; movd %xmm2,%edi - DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 - DB 102,15,126,211 ; movd %xmm2,%ebx - DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 + DB 15,20,248 ; unpcklps %xmm0,%xmm7 DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,234 ; unpcklps %xmm2,%xmm5 - DB 15,20,239 ; unpcklps %xmm7,%xmm5 - DB 102,15,254,177,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm6 - DB 102,15,112,214,229 ; pshufd $0xe5,%xmm6,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,214,78 ; pshufd $0x4e,%xmm6,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,232 ; unpcklps %xmm0,%xmm5 + DB 102,15,20,239 ; unpcklpd %xmm7,%xmm5 + DB 102,15,254,241 ; paddd %xmm1,%xmm6 + DB 102,15,112,198,229 ; pshufd $0xe5,%xmm6,%xmm0 + DB 102,15,126,192 ; movd %xmm0,%eax + DB 102,15,112,198,78 ; pshufd $0x4e,%xmm6,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi DB 102,15,126,247 ; movd %xmm6,%edi - DB 102,15,112,214,231 ; pshufd $0xe7,%xmm6,%xmm2 - DB 102,15,126,211 ; movd %xmm2,%ebx - DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 + DB 102,15,112,198,231 ; pshufd $0xe7,%xmm6,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 + DB 15,20,248 ; unpcklps %xmm0,%xmm7 DB 243,15,16,52,186 ; movss (%edx,%edi,4),%xmm6 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 - DB 15,20,247 ; unpcklps %xmm7,%xmm6 - DB 102,15,111,133,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm0 - DB 102,15,254,69,184 ; paddd -0x48(%ebp),%xmm0 - DB 102,15,112,208,245 ; pshufd $0xf5,%xmm0,%xmm2 - DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 - DB 102,15,112,248,232 ; pshufd $0xe8,%xmm0,%xmm7 - DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 - DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 - DB 102,15,98,250 ; punpckldq %xmm2,%xmm7 - DB 102,15,112,215,229 ; pshufd $0xe5,%xmm7,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,215,78 ; pshufd $0x4e,%xmm7,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi - DB 102,15,112,215,231 ; pshufd $0xe7,%xmm7,%xmm2 - DB 102,15,126,215 ; movd %xmm2,%edi - DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 15,20,194 ; unpcklps %xmm2,%xmm0 - DB 102,15,126,248 ; movd %xmm7,%eax - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 - DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 - DB 15,20,202 ; unpcklps %xmm2,%xmm1 - DB 15,20,200 ; unpcklps %xmm0,%xmm1 + DB 15,20,240 ; unpcklps %xmm0,%xmm6 + DB 102,15,20,247 ; unpcklpd %xmm7,%xmm6 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,254,77,168 ; paddd -0x58(%ebp),%xmm1 + DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 + DB 102,15,111,145,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm2 + DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 + DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,112,249,232 ; pshufd $0xe8,%xmm1,%xmm7 + DB 102,15,98,248 ; punpckldq %xmm0,%xmm7 + DB 102,15,126,200 ; movd %xmm1,%eax + DB 102,15,112,199,229 ; pshufd $0xe5,%xmm7,%xmm0 + DB 102,15,126,198 ; movd %xmm0,%esi + DB 102,15,112,199,78 ; pshufd $0x4e,%xmm7,%xmm0 + DB 102,15,126,199 ; movd %xmm0,%edi + DB 102,15,112,199,231 ; pshufd $0xe7,%xmm7,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 + DB 15,20,193 ; unpcklps %xmm1,%xmm0 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 + DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 + DB 15,20,209 ; unpcklps %xmm1,%xmm2 + DB 102,15,20,208 ; unpcklpd %xmm0,%xmm2 DB 102,15,111,199 ; movdqa %xmm7,%xmm0 - DB 102,15,254,195 ; paddd %xmm3,%xmm0 - DB 102,15,112,208,229 ; pshufd $0xe5,%xmm0,%xmm2 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 102,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm2 - DB 102,15,126,214 ; movd %xmm2,%esi + DB 102,15,250,195 ; psubd %xmm3,%xmm0 + DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 + DB 102,15,126,200 ; movd %xmm1,%eax + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,15,126,206 ; movd %xmm1,%esi DB 102,15,126,199 ; movd %xmm0,%edi DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 - DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 102,15,254,185,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm7 + DB 102,15,20,217 ; unpcklpd %xmm1,%xmm3 + DB 102,15,254,185,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm7 DB 102,15,112,199,229 ; pshufd $0xe5,%xmm7,%xmm0 DB 102,15,126,192 ; movd %xmm0,%eax DB 102,15,112,199,78 ; pshufd $0x4e,%xmm7,%xmm0 @@ -55833,71 +55769,72 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,199,231 ; pshufd $0xe7,%xmm7,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 - DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 + DB 15,20,200 ; unpcklps %xmm0,%xmm1 DB 243,15,16,60,186 ; movss (%edx,%edi,4),%xmm7 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 15,20,248 ; unpcklps %xmm0,%xmm7 - DB 15,20,250 ; unpcklps %xmm2,%xmm7 - DB 15,40,85,216 ; movaps -0x28(%ebp),%xmm2 - DB 15,92,202 ; subps %xmm2,%xmm1 - DB 15,40,69,152 ; movaps -0x68(%ebp),%xmm0 - DB 15,89,200 ; mulps %xmm0,%xmm1 - DB 15,88,202 ; addps %xmm2,%xmm1 + DB 102,15,20,249 ; unpcklpd %xmm1,%xmm7 + DB 15,92,212 ; subps %xmm4,%xmm2 + DB 15,40,69,216 ; movaps -0x28(%ebp),%xmm0 + DB 15,89,208 ; mulps %xmm0,%xmm2 + DB 15,88,212 ; addps %xmm4,%xmm2 DB 15,92,221 ; subps %xmm5,%xmm3 DB 15,89,216 ; mulps %xmm0,%xmm3 DB 15,88,221 ; addps %xmm5,%xmm3 DB 15,92,254 ; subps %xmm6,%xmm7 DB 15,89,248 ; mulps %xmm0,%xmm7 DB 15,88,254 ; addps %xmm6,%xmm7 - DB 15,40,149,120,255,255,255 ; movaps -0x88(%ebp),%xmm2 - DB 15,92,202 ; subps %xmm2,%xmm1 - DB 15,40,133,40,255,255,255 ; movaps -0xd8(%ebp),%xmm0 - DB 15,89,200 ; mulps %xmm0,%xmm1 - DB 15,88,202 ; addps %xmm2,%xmm1 - DB 15,41,141,56,255,255,255 ; movaps %xmm1,-0xc8(%ebp) + DB 15,40,141,120,255,255,255 ; movaps -0x88(%ebp),%xmm1 + DB 15,92,209 ; subps %xmm1,%xmm2 + DB 15,40,133,56,255,255,255 ; movaps -0xc8(%ebp),%xmm0 + DB 15,89,208 ; mulps %xmm0,%xmm2 + DB 15,88,209 ; addps %xmm1,%xmm2 + DB 15,41,149,8,255,255,255 ; movaps %xmm2,-0xf8(%ebp) DB 15,40,77,200 ; movaps -0x38(%ebp),%xmm1 DB 15,92,217 ; subps %xmm1,%xmm3 DB 15,89,216 ; mulps %xmm0,%xmm3 DB 15,88,217 ; addps %xmm1,%xmm3 DB 15,41,157,120,255,255,255 ; movaps %xmm3,-0x88(%ebp) - DB 15,92,252 ; subps %xmm4,%xmm7 + DB 15,40,77,184 ; movaps -0x48(%ebp),%xmm1 + DB 15,92,249 ; subps %xmm1,%xmm7 DB 15,89,248 ; mulps %xmm0,%xmm7 - DB 15,88,252 ; addps %xmm4,%xmm7 - DB 15,41,125,200 ; movaps %xmm7,-0x38(%ebp) - DB 102,15,111,133,248,254,255,255 ; movdqa -0x108(%ebp),%xmm0 + DB 15,88,249 ; addps %xmm1,%xmm7 + DB 15,41,125,184 ; movaps %xmm7,-0x48(%ebp) + DB 102,15,111,133,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm0 DB 102,15,254,69,136 ; paddd -0x78(%ebp),%xmm0 - DB 102,15,127,133,248,254,255,255 ; movdqa %xmm0,-0x108(%ebp) + DB 102,15,127,133,24,255,255,255 ; movdqa %xmm0,-0xe8(%ebp) DB 102,15,111,141,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm1 DB 102,15,254,200 ; paddd %xmm0,%xmm1 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,254,69,168 ; paddd -0x58(%ebp),%xmm0 + DB 102,15,254,69,152 ; paddd -0x68(%ebp),%xmm0 DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,153,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm3 - DB 102,15,244,195 ; pmuludq %xmm3,%xmm0 + DB 102,15,111,145,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm2 + DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 + DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 + DB 102,15,111,250 ; movdqa %xmm2,%xmm7 + DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,112,208,232 ; pshufd $0xe8,%xmm0,%xmm2 - DB 102,15,244,203 ; pmuludq %xmm3,%xmm1 - DB 102,15,112,193,232 ; pshufd $0xe8,%xmm1,%xmm0 - DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 - DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 + DB 102,15,98,209 ; punpckldq %xmm1,%xmm2 DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 + DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 + DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 15,20,195 ; unpcklps %xmm3,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 15,41,133,72,255,255,255 ; movaps %xmm0,-0xb8(%ebp) + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 102,15,41,133,72,255,255,255 ; movapd %xmm0,-0xb8(%ebp) DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,15,111,177,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm6 - DB 102,15,254,206 ; paddd %xmm6,%xmm1 + DB 102,15,118,246 ; pcmpeqd %xmm6,%xmm6 + DB 102,15,250,206 ; psubd %xmm6,%xmm1 DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 DB 102,15,126,216 ; movd %xmm3,%eax DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 @@ -55906,14 +55843,13 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,203 ; movd %xmm1,%ebx DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 DB 15,20,217 ; unpcklps %xmm1,%xmm3 DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 DB 15,20,204 ; unpcklps %xmm4,%xmm1 - DB 15,20,203 ; unpcklps %xmm3,%xmm1 - DB 102,15,111,185,255,81,0,0 ; movdqa 0x51ff(%ecx),%xmm7 - DB 102,15,254,215 ; paddd %xmm7,%xmm2 + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 + DB 102,15,254,145,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm2 DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 DB 102,15,126,216 ; movd %xmm3,%eax DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 @@ -55922,37 +55858,37 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 DB 102,15,126,211 ; movd %xmm2,%ebx DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 - DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 DB 15,20,212 ; unpcklps %xmm4,%xmm2 - DB 15,20,211 ; unpcklps %xmm3,%xmm2 - DB 102,15,111,221 ; movdqa %xmm5,%xmm3 - DB 102,15,254,93,184 ; paddd -0x48(%ebp),%xmm3 - DB 102,15,112,227,245 ; pshufd $0xf5,%xmm3,%xmm4 - DB 102,15,111,129,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm0 - DB 102,15,244,216 ; pmuludq %xmm0,%xmm3 - DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 - DB 102,15,244,224 ; pmuludq %xmm0,%xmm4 - DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 102,15,111,229 ; movdqa %xmm5,%xmm4 + DB 102,15,254,101,168 ; paddd -0x58(%ebp),%xmm4 + DB 102,15,112,220,245 ; pshufd $0xf5,%xmm4,%xmm3 + DB 102,15,244,231 ; pmuludq %xmm7,%xmm4 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,15,244,223 ; pmuludq %xmm7,%xmm3 + DB 102,15,112,227,232 ; pshufd $0xe8,%xmm3,%xmm4 + DB 102,15,112,221,232 ; pshufd $0xe8,%xmm5,%xmm3 DB 102,15,98,220 ; punpckldq %xmm4,%xmm3 + DB 102,15,126,232 ; movd %xmm5,%eax DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 - DB 102,15,126,224 ; movd %xmm4,%eax - DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 DB 102,15,126,230 ; movd %xmm4,%esi - DB 102,15,112,227,231 ; pshufd $0xe7,%xmm3,%xmm4 + DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 DB 102,15,126,231 ; movd %xmm4,%edi - DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 - DB 243,15,16,44,130 ; movss (%edx,%eax,4),%xmm5 + DB 102,15,112,227,231 ; pshufd $0xe7,%xmm3,%xmm4 + DB 102,15,126,227 ; movd %xmm4,%ebx + DB 243,15,16,36,154 ; movss (%edx,%ebx,4),%xmm4 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 102,15,126,216 ; movd %xmm3,%eax - DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 DB 15,20,196 ; unpcklps %xmm4,%xmm0 - DB 15,20,197 ; unpcklps %xmm5,%xmm0 + DB 102,15,20,197 ; unpcklpd %xmm5,%xmm0 DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,254,230 ; paddd %xmm6,%xmm4 + DB 102,15,250,230 ; psubd %xmm6,%xmm4 DB 102,15,112,236,229 ; pshufd $0xe5,%xmm4,%xmm5 DB 102,15,126,232 ; movd %xmm5,%eax DB 102,15,112,236,78 ; pshufd $0x4e,%xmm4,%xmm5 @@ -55961,13 +55897,13 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,228,231 ; pshufd $0xe7,%xmm4,%xmm4 DB 102,15,126,227 ; movd %xmm4,%ebx DB 243,15,16,36,154 ; movss (%edx,%ebx,4),%xmm4 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 DB 15,20,244 ; unpcklps %xmm4,%xmm6 - DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 - DB 243,15,16,36,178 ; movss (%edx,%esi,4),%xmm4 - DB 15,20,236 ; unpcklps %xmm4,%xmm5 - DB 15,20,238 ; unpcklps %xmm6,%xmm5 - DB 102,15,254,223 ; paddd %xmm7,%xmm3 + DB 243,15,16,60,186 ; movss (%edx,%edi,4),%xmm7 + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 15,20,252 ; unpcklps %xmm4,%xmm7 + DB 102,15,20,254 ; unpcklpd %xmm6,%xmm7 + DB 102,15,254,153,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm3 DB 102,15,112,227,229 ; pshufd $0xe5,%xmm3,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax DB 102,15,112,227,78 ; pshufd $0x4e,%xmm3,%xmm4 @@ -55976,54 +55912,55 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,219,231 ; pshufd $0xe7,%xmm3,%xmm3 DB 102,15,126,219 ; movd %xmm3,%ebx DB 243,15,16,28,154 ; movss (%edx,%ebx,4),%xmm3 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 + DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 DB 15,20,243 ; unpcklps %xmm3,%xmm6 - DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 - DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 - DB 15,20,227 ; unpcklps %xmm3,%xmm4 - DB 15,20,230 ; unpcklps %xmm6,%xmm4 - DB 15,40,181,72,255,255,255 ; movaps -0xb8(%ebp),%xmm6 - DB 15,92,198 ; subps %xmm6,%xmm0 - DB 15,40,93,152 ; movaps -0x68(%ebp),%xmm3 + DB 243,15,16,44,186 ; movss (%edx,%edi,4),%xmm5 + DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 + DB 15,20,235 ; unpcklps %xmm3,%xmm5 + DB 102,15,20,238 ; unpcklpd %xmm6,%xmm5 + DB 15,40,165,72,255,255,255 ; movaps -0xb8(%ebp),%xmm4 + DB 15,92,196 ; subps %xmm4,%xmm0 + DB 15,40,93,216 ; movaps -0x28(%ebp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 - DB 15,88,198 ; addps %xmm6,%xmm0 + DB 15,88,196 ; addps %xmm4,%xmm0 DB 15,41,133,72,255,255,255 ; movaps %xmm0,-0xb8(%ebp) - DB 15,92,233 ; subps %xmm1,%xmm5 + DB 15,92,249 ; subps %xmm1,%xmm7 + DB 15,89,251 ; mulps %xmm3,%xmm7 + DB 15,88,249 ; addps %xmm1,%xmm7 + DB 15,92,234 ; subps %xmm2,%xmm5 DB 15,89,235 ; mulps %xmm3,%xmm5 - DB 15,88,233 ; addps %xmm1,%xmm5 - DB 15,92,226 ; subps %xmm2,%xmm4 - DB 15,89,227 ; mulps %xmm3,%xmm4 - DB 15,88,226 ; addps %xmm2,%xmm4 - DB 102,15,111,141,248,254,255,255 ; movdqa -0x108(%ebp),%xmm1 - DB 102,15,254,141,88,255,255,255 ; paddd -0xa8(%ebp),%xmm1 - DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 - DB 102,15,254,193 ; paddd %xmm1,%xmm0 + DB 15,88,234 ; addps %xmm2,%xmm5 + DB 102,15,111,133,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm0 + DB 102,15,254,133,88,255,255,255 ; paddd -0xa8(%ebp),%xmm0 + DB 102,15,111,85,152 ; movdqa -0x68(%ebp),%xmm2 + DB 102,15,254,208 ; paddd %xmm0,%xmm2 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,112,194,245 ; pshufd $0xf5,%xmm2,%xmm0 + DB 102,15,111,137,73,87,0,0 ; movdqa 0x5749(%ecx),%xmm1 + DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 + DB 102,15,244,193 ; pmuludq %xmm1,%xmm0 DB 102,15,111,241 ; movdqa %xmm1,%xmm6 - DB 102,15,112,200,245 ; pshufd $0xf5,%xmm0,%xmm1 - DB 102,15,111,145,79,88,0,0 ; movdqa 0x584f(%ecx),%xmm2 - DB 102,15,244,194 ; pmuludq %xmm2,%xmm0 - DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 - DB 102,15,244,202 ; pmuludq %xmm2,%xmm1 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 + DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 + DB 102,15,112,194,232 ; pshufd $0xe8,%xmm2,%xmm0 DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,15,126,208 ; movd %xmm2,%eax DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,126,206 ; movd %xmm1,%esi - DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 DB 102,15,126,207 ; movd %xmm1,%edi - DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 + DB 102,15,112,200,231 ; pshufd $0xe7,%xmm0,%xmm1 + DB 102,15,126,203 ; movd %xmm1,%ebx + DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 + DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 243,15,16,28,130 ; movss (%edx,%eax,4),%xmm3 DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,41,93,168 ; movaps %xmm3,-0x58(%ebp) + DB 243,15,16,36,130 ; movss (%edx,%eax,4),%xmm4 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 102,15,20,226 ; unpcklpd %xmm2,%xmm4 + DB 102,15,41,101,152 ; movapd %xmm4,-0x68(%ebp) DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,254,137,223,81,0,0 ; paddd 0x51df(%ecx),%xmm1 + DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 + DB 102,15,250,202 ; psubd %xmm2,%xmm1 DB 102,15,112,209,229 ; pshufd $0xe5,%xmm1,%xmm2 DB 102,15,126,208 ; movd %xmm2,%eax DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2 @@ -56032,14 +55969,14 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,203 ; movd %xmm1,%ebx DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 - DB 243,15,16,20,130 ; movss (%edx,%eax,4),%xmm2 + DB 243,15,16,20,178 ; movss (%edx,%esi,4),%xmm2 DB 15,20,209 ; unpcklps %xmm1,%xmm2 - DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 - DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 15,20,218 ; unpcklps %xmm2,%xmm3 - DB 15,41,157,88,255,255,255 ; movaps %xmm3,-0xa8(%ebp) - DB 102,15,254,129,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm0 + DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 + DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 + DB 15,20,225 ; unpcklps %xmm1,%xmm4 + DB 102,15,20,226 ; unpcklpd %xmm2,%xmm4 + DB 102,15,41,165,88,255,255,255 ; movapd %xmm4,-0xa8(%ebp) + DB 102,15,254,129,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm0 DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,15,126,200 ; movd %xmm1,%eax DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 @@ -56048,105 +55985,106 @@ _sk_clut_4D_sse2 LABEL PROC DB 102,15,112,192,231 ; pshufd $0xe7,%xmm0,%xmm0 DB 102,15,126,195 ; movd %xmm0,%ebx DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 + DB 243,15,16,12,178 ; movss (%edx,%esi,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 243,15,16,28,186 ; movss (%edx,%edi,4),%xmm3 - DB 243,15,16,4,178 ; movss (%edx,%esi,4),%xmm0 - DB 15,20,216 ; unpcklps %xmm0,%xmm3 - DB 15,20,217 ; unpcklps %xmm1,%xmm3 - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,15,254,77,184 ; paddd -0x48(%ebp),%xmm1 + DB 243,15,16,36,186 ; movss (%edx,%edi,4),%xmm4 + DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 + DB 15,20,224 ; unpcklps %xmm0,%xmm4 + DB 102,15,20,225 ; unpcklpd %xmm1,%xmm4 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,254,77,168 ; paddd -0x58(%ebp),%xmm1 DB 102,15,112,193,245 ; pshufd $0xf5,%xmm1,%xmm0 - DB 102,15,244,207 ; pmuludq %xmm7,%xmm1 - DB 102,15,244,199 ; pmuludq %xmm7,%xmm0 - DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 + DB 102,15,244,206 ; pmuludq %xmm6,%xmm1 + DB 102,15,244,198 ; pmuludq %xmm6,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,112,209,232 ; pshufd $0xe8,%xmm1,%xmm2 DB 102,15,98,208 ; punpckldq %xmm0,%xmm2 + DB 102,15,126,200 ; movd %xmm1,%eax DB 102,15,112,194,229 ; pshufd $0xe5,%xmm2,%xmm0 - DB 102,15,126,192 ; movd %xmm0,%eax - DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,198 ; movd %xmm0,%esi - DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,112,194,78 ; pshufd $0x4e,%xmm2,%xmm0 DB 102,15,126,199 ; movd %xmm0,%edi - DB 243,15,16,4,186 ; movss (%edx,%edi,4),%xmm0 - DB 243,15,16,12,130 ; movss (%edx,%eax,4),%xmm1 + DB 102,15,112,194,231 ; pshufd $0xe7,%xmm2,%xmm0 + DB 102,15,126,195 ; movd %xmm0,%ebx + DB 243,15,16,4,154 ; movss (%edx,%ebx,4),%xmm0 + DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 DB 15,20,200 ; unpcklps %xmm0,%xmm1 - DB 102,15,126,208 ; movd %xmm2,%eax + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 DB 243,15,16,4,130 ; movss (%edx,%eax,4),%xmm0 - DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 - DB 15,20,199 ; unpcklps %xmm7,%xmm0 - DB 15,20,193 ; unpcklps %xmm1,%xmm0 - DB 102,15,111,177,223,81,0,0 ; movdqa 0x51df(%ecx),%xmm6 - DB 102,15,254,242 ; paddd %xmm2,%xmm6 - DB 102,15,112,206,229 ; pshufd $0xe5,%xmm6,%xmm1 - DB 102,15,126,200 ; movd %xmm1,%eax - DB 102,15,112,206,78 ; pshufd $0x4e,%xmm6,%xmm1 - DB 102,15,126,206 ; movd %xmm1,%esi - DB 102,15,126,247 ; movd %xmm6,%edi - DB 102,15,112,206,231 ; pshufd $0xe7,%xmm6,%xmm1 + DB 15,20,195 ; unpcklps %xmm3,%xmm0 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,250,203 ; psubd %xmm3,%xmm1 + DB 102,15,112,217,229 ; pshufd $0xe5,%xmm1,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,217,78 ; pshufd $0x4e,%xmm1,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi + DB 102,15,126,207 ; movd %xmm1,%edi + DB 102,15,112,201,231 ; pshufd $0xe7,%xmm1,%xmm1 DB 102,15,126,203 ; movd %xmm1,%ebx DB 243,15,16,12,154 ; movss (%edx,%ebx,4),%xmm1 - DB 243,15,16,60,130 ; movss (%edx,%eax,4),%xmm7 - DB 15,20,249 ; unpcklps %xmm1,%xmm7 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 + DB 15,20,217 ; unpcklps %xmm1,%xmm3 DB 243,15,16,12,186 ; movss (%edx,%edi,4),%xmm1 - DB 243,15,16,52,178 ; movss (%edx,%esi,4),%xmm6 + DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 DB 15,20,206 ; unpcklps %xmm6,%xmm1 - DB 15,20,207 ; unpcklps %xmm7,%xmm1 - DB 102,15,254,145,255,81,0,0 ; paddd 0x51ff(%ecx),%xmm2 - DB 102,15,112,242,229 ; pshufd $0xe5,%xmm2,%xmm6 - DB 102,15,126,240 ; movd %xmm6,%eax - DB 102,15,112,242,78 ; pshufd $0x4e,%xmm2,%xmm6 - DB 102,15,126,246 ; movd %xmm6,%esi + DB 102,15,20,203 ; unpcklpd %xmm3,%xmm1 + DB 102,15,254,145,201,80,0,0 ; paddd 0x50c9(%ecx),%xmm2 + DB 102,15,112,218,229 ; pshufd $0xe5,%xmm2,%xmm3 + DB 102,15,126,216 ; movd %xmm3,%eax + DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3 + DB 102,15,126,222 ; movd %xmm3,%esi DB 102,15,126,215 ; movd %xmm2,%edi DB 102,15,112,210,231 ; pshufd $0xe7,%xmm2,%xmm2 DB 102,15,126,211 ; movd %xmm2,%ebx DB 243,15,16,20,154 ; movss (%edx,%ebx,4),%xmm2 - DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 - DB 15,20,242 ; unpcklps %xmm2,%xmm6 + DB 243,15,16,28,178 ; movss (%edx,%esi,4),%xmm3 + DB 15,20,218 ; unpcklps %xmm2,%xmm3 DB 243,15,16,20,186 ; movss (%edx,%edi,4),%xmm2 - DB 243,15,16,60,178 ; movss (%edx,%esi,4),%xmm7 - DB 15,20,215 ; unpcklps %xmm7,%xmm2 + DB 243,15,16,52,130 ; movss (%edx,%eax,4),%xmm6 DB 15,20,214 ; unpcklps %xmm6,%xmm2 - DB 15,40,125,168 ; movaps -0x58(%ebp),%xmm7 - DB 15,92,199 ; subps %xmm7,%xmm0 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 DB 15,40,117,152 ; movaps -0x68(%ebp),%xmm6 - DB 15,89,198 ; mulps %xmm6,%xmm0 - DB 15,88,199 ; addps %xmm7,%xmm0 - DB 15,40,189,88,255,255,255 ; movaps -0xa8(%ebp),%xmm7 - DB 15,92,207 ; subps %xmm7,%xmm1 - DB 15,89,206 ; mulps %xmm6,%xmm1 - DB 15,88,207 ; addps %xmm7,%xmm1 - DB 15,92,211 ; subps %xmm3,%xmm2 - DB 15,89,214 ; mulps %xmm6,%xmm2 - DB 15,88,211 ; addps %xmm3,%xmm2 - DB 15,40,181,72,255,255,255 ; movaps -0xb8(%ebp),%xmm6 DB 15,92,198 ; subps %xmm6,%xmm0 - DB 15,40,157,40,255,255,255 ; movaps -0xd8(%ebp),%xmm3 + DB 15,40,93,216 ; movaps -0x28(%ebp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 DB 15,88,198 ; addps %xmm6,%xmm0 - DB 15,92,205 ; subps %xmm5,%xmm1 + DB 15,40,181,88,255,255,255 ; movaps -0xa8(%ebp),%xmm6 + DB 15,92,206 ; subps %xmm6,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 - DB 15,88,205 ; addps %xmm5,%xmm1 + DB 15,88,206 ; addps %xmm6,%xmm1 DB 15,92,212 ; subps %xmm4,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 DB 15,88,212 ; addps %xmm4,%xmm2 - DB 15,40,165,56,255,255,255 ; movaps -0xc8(%ebp),%xmm4 + DB 15,40,165,72,255,255,255 ; movaps -0xb8(%ebp),%xmm4 + DB 15,92,196 ; subps %xmm4,%xmm0 + DB 15,40,157,56,255,255,255 ; movaps -0xc8(%ebp),%xmm3 + DB 15,89,195 ; mulps %xmm3,%xmm0 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 15,92,207 ; subps %xmm7,%xmm1 + DB 15,89,203 ; mulps %xmm3,%xmm1 + DB 15,88,207 ; addps %xmm7,%xmm1 + DB 15,92,213 ; subps %xmm5,%xmm2 + DB 15,89,211 ; mulps %xmm3,%xmm2 + DB 15,88,213 ; addps %xmm5,%xmm2 + DB 15,40,165,8,255,255,255 ; movaps -0xf8(%ebp),%xmm4 DB 15,92,196 ; subps %xmm4,%xmm0 - DB 15,40,157,232,254,255,255 ; movaps -0x118(%ebp),%xmm3 + DB 15,40,157,248,254,255,255 ; movaps -0x108(%ebp),%xmm3 DB 15,89,195 ; mulps %xmm3,%xmm0 DB 15,88,196 ; addps %xmm4,%xmm0 DB 15,40,165,120,255,255,255 ; movaps -0x88(%ebp),%xmm4 DB 15,92,204 ; subps %xmm4,%xmm1 DB 15,89,203 ; mulps %xmm3,%xmm1 DB 15,88,204 ; addps %xmm4,%xmm1 - DB 15,40,101,200 ; movaps -0x38(%ebp),%xmm4 + DB 15,40,101,184 ; movaps -0x48(%ebp),%xmm4 DB 15,92,212 ; subps %xmm4,%xmm2 DB 15,89,211 ; mulps %xmm3,%xmm2 DB 15,88,212 ; addps %xmm4,%xmm2 DB 15,91,157,184,254,255,255 ; cvtdq2ps -0x148(%ebp),%xmm3 - DB 15,40,165,120,254,255,255 ; movaps -0x188(%ebp),%xmm4 + DB 15,40,165,152,254,255,255 ; movaps -0x168(%ebp),%xmm4 DB 15,92,227 ; subps %xmm3,%xmm4 - DB 15,40,157,8,255,255,255 ; movaps -0xf8(%ebp),%xmm3 + DB 15,40,157,40,255,255,255 ; movaps -0xd8(%ebp),%xmm3 DB 15,92,195 ; subps %xmm3,%xmm0 DB 15,89,196 ; mulps %xmm4,%xmm0 DB 15,88,195 ; addps %xmm3,%xmm0 @@ -56162,11 +56100,11 @@ _sk_clut_4D_sse2 LABEL PROC DB 139,117,12 ; mov 0xc(%ebp),%esi DB 141,86,8 ; lea 0x8(%esi),%edx DB 131,236,8 ; sub $0x8,%esp - DB 15,40,153,191,81,0,0 ; movaps 0x51bf(%ecx),%xmm3 + DB 15,40,153,137,80,0,0 ; movaps 0x5089(%ecx),%xmm3 DB 82 ; push %edx DB 80 ; push %eax DB 255,86,4 ; call *0x4(%esi) - DB 129,196,140,1,0,0 ; add $0x18c,%esp + DB 129,196,108,1,0,0 ; add $0x16c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx @@ -56186,7 +56124,7 @@ _sk_start_pipeline_sse2_8bit LABEL PROC DB 131,236,16 ; sub $0x10,%esp DB 139,93,12 ; mov 0xc(%ebp),%ebx DB 59,93,20 ; cmp 0x14(%ebp),%ebx - DB 115,97 ; jae 9369 <_sk_start_pipeline_sse2_8bit+0x75> + DB 115,97 ; jae 9379 <_sk_start_pipeline_sse2_8bit+0x75> DB 139,85,16 ; mov 0x10(%ebp),%edx DB 139,69,8 ; mov 0x8(%ebp),%eax DB 139,117,24 ; mov 0x18(%ebp),%esi @@ -56198,7 +56136,7 @@ _sk_start_pipeline_sse2_8bit LABEL PROC DB 57,84,36,8 ; cmp %edx,0x8(%esp) DB 139,69,8 ; mov 0x8(%ebp),%eax DB 137,243 ; mov %esi,%ebx - DB 119,31 ; ja 934b <_sk_start_pipeline_sse2_8bit+0x57> + DB 119,31 ; ja 935b <_sk_start_pipeline_sse2_8bit+0x57> DB 139,117,8 ; mov 0x8(%ebp),%esi DB 255,116,36,4 ; pushl 0x4(%esp) DB 86 ; push %esi @@ -56211,12 +56149,12 @@ _sk_start_pipeline_sse2_8bit LABEL PROC DB 131,198,16 ; add $0x10,%esi DB 57,214 ; cmp %edx,%esi DB 137,198 ; mov %eax,%esi - DB 118,228 ; jbe 932f <_sk_start_pipeline_sse2_8bit+0x3b> + DB 118,228 ; jbe 933f <_sk_start_pipeline_sse2_8bit+0x3b> DB 137,209 ; mov %edx,%ecx DB 41,193 ; sub %eax,%ecx DB 137,222 ; mov %ebx,%esi DB 139,92,36,4 ; mov 0x4(%esp),%ebx - DB 116,12 ; je 9363 <_sk_start_pipeline_sse2_8bit+0x6f> + DB 116,12 ; je 9373 <_sk_start_pipeline_sse2_8bit+0x6f> DB 83 ; push %ebx DB 80 ; push %eax DB 86 ; push %esi @@ -56226,7 +56164,7 @@ _sk_start_pipeline_sse2_8bit LABEL PROC DB 131,196,16 ; add $0x10,%esp DB 67 ; inc %ebx DB 59,93,20 ; cmp 0x14(%ebp),%ebx - DB 117,180 ; jne 931d <_sk_start_pipeline_sse2_8bit+0x29> + DB 117,180 ; jne 932d <_sk_start_pipeline_sse2_8bit+0x29> DB 141,101,244 ; lea -0xc(%ebp),%esp DB 94 ; pop %esi DB 95 ; pop %edi @@ -56266,15 +56204,15 @@ _sk_set_rgb_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 93ae <_sk_set_rgb_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 93be <_sk_set_rgb_sse2_8bit+0xe> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx - DB 243,15,16,161,246,67,0,0 ; movss 0x43f6(%ecx),%xmm4 + DB 243,15,16,161,6,67,0,0 ; movss 0x4306(%ecx),%xmm4 DB 243,15,16,50 ; movss (%edx),%xmm6 DB 243,15,89,244 ; mulss %xmm4,%xmm6 DB 243,15,44,246 ; cvttss2si %xmm6,%esi - DB 243,15,16,169,250,67,0,0 ; movss 0x43fa(%ecx),%xmm5 + DB 243,15,16,169,10,67,0,0 ; movss 0x430a(%ecx),%xmm5 DB 15,40,254 ; movaps %xmm6,%xmm7 DB 243,15,92,253 ; subss %xmm5,%xmm7 DB 243,15,44,255 ; cvttss2si %xmm7,%edi @@ -56304,7 +56242,7 @@ _sk_set_rgb_sse2_8bit LABEL PROC DB 9,247 ; or %esi,%edi DB 102,15,110,231 ; movd %edi,%xmm4 DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 102,15,111,169,66,67,0,0 ; movdqa 0x4342(%ecx),%xmm5 + DB 102,15,111,169,82,66,0,0 ; movdqa 0x4252(%ecx),%xmm5 DB 102,15,219,205 ; pand %xmm5,%xmm1 DB 102,15,219,197 ; pand %xmm5,%xmm0 DB 102,15,235,196 ; por %xmm4,%xmm0 @@ -56330,7 +56268,7 @@ _sk_premul_sse2_8bit LABEL PROC DB 15,41,93,200 ; movaps %xmm3,-0x38(%ebp) DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 232,0,0,0,0 ; call 9486 <_sk_premul_sse2_8bit+0x18> + DB 232,0,0,0,0 ; call 9496 <_sk_premul_sse2_8bit+0x18> DB 88 ; pop %eax DB 242,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 @@ -56344,7 +56282,7 @@ _sk_premul_sse2_8bit LABEL PROC DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 DB 243,15,112,252,95 ; pshufhw $0x5f,%xmm4,%xmm7 - DB 102,15,111,160,106,66,0,0 ; movdqa 0x426a(%eax),%xmm4 + DB 102,15,111,160,122,65,0,0 ; movdqa 0x417a(%eax),%xmm4 DB 102,15,235,252 ; por %xmm4,%xmm7 DB 102,15,235,244 ; por %xmm4,%xmm6 DB 102,15,239,210 ; pxor %xmm2,%xmm2 @@ -56370,12 +56308,12 @@ _sk_premul_sse2_8bit LABEL PROC DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 DB 102,15,213,248 ; pmullw %xmm0,%xmm7 DB 102,15,253,248 ; paddw %xmm0,%xmm7 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,230 ; packuswb %xmm6,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,239 ; packuswb %xmm7,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,230 ; packuswb %xmm6,%xmm4 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 102,15,111,196 ; movdqa %xmm4,%xmm0 @@ -56451,7 +56389,7 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 9612 <_sk_load_8888_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 9622 <_sk_load_8888_sse2_8bit+0xe> DB 91 ; pop %ebx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -56462,9 +56400,9 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,30 ; jne 964c <_sk_load_8888_sse2_8bit+0x48> - DB 15,16,76,183,16 ; movups 0x10(%edi,%esi,4),%xmm1 + DB 117,30 ; jne 965c <_sk_load_8888_sse2_8bit+0x48> DB 102,15,16,4,183 ; movupd (%edi,%esi,4),%xmm0 + DB 15,16,76,183,16 ; movups 0x10(%edi,%esi,4),%xmm1 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 255,117,20 ; pushl 0x14(%ebp) DB 86 ; push %esi @@ -56483,16 +56421,16 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,216 ; ja 9638 <_sk_load_8888_sse2_8bit+0x34> + DB 119,216 ; ja 9648 <_sk_load_8888_sse2_8bit+0x34> DB 15,182,201 ; movzbl %cl,%ecx - DB 3,156,139,158,0,0,0 ; add 0x9e(%ebx,%ecx,4),%ebx + DB 3,156,139,162,0,0,0 ; add 0xa2(%ebx,%ecx,4),%ebx DB 255,227 ; jmp *%ebx DB 243,15,16,4,183 ; movss (%edi,%esi,4),%xmm0 - DB 235,197 ; jmp 9638 <_sk_load_8888_sse2_8bit+0x34> + DB 235,197 ; jmp 9648 <_sk_load_8888_sse2_8bit+0x34> DB 102,15,110,68,183,8 ; movd 0x8(%edi,%esi,4),%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 DB 102,15,18,4,183 ; movlpd (%edi,%esi,4),%xmm0 - DB 235,179 ; jmp 9638 <_sk_load_8888_sse2_8bit+0x34> + DB 235,179 ; jmp 9648 <_sk_load_8888_sse2_8bit+0x34> DB 102,15,110,68,183,24 ; movd 0x18(%edi,%esi,4),%xmm0 DB 102,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm1 DB 243,15,16,68,183,20 ; movss 0x14(%edi,%esi,4),%xmm0 @@ -56501,20 +56439,19 @@ _sk_load_8888_sse2_8bit LABEL PROC DB 15,40,200 ; movaps %xmm0,%xmm1 DB 243,15,16,68,183,16 ; movss 0x10(%edi,%esi,4),%xmm0 DB 243,15,16,200 ; movss %xmm0,%xmm1 - DB 235,134 ; jmp 9633 <_sk_load_8888_sse2_8bit+0x2f> - DB 15,31,0 ; nopl (%eax) + DB 102,15,16,4,183 ; movupd (%edi,%esi,4),%xmm0 + DB 235,134 ; jmp 9648 <_sk_load_8888_sse2_8bit+0x34> + DB 102,144 ; xchg %ax,%ax DB 90 ; pop %edx DB 0,0 ; add %al,(%eax) DB 0,108,0,0 ; add %ch,0x0(%eax,%eax,1) DB 0,97,0 ; add %ah,0x0(%ecx) DB 0,0 ; add %al,(%eax) - DB 33,0 ; and %eax,(%eax) - DB 0,0 ; add %al,(%eax) - DB 143,0 ; popl (%eax) + DB 153 ; cltd DB 0,0 ; add %al,(%eax) - DB 126,0 ; jle 96c6 <_sk_load_8888_sse2_8bit+0xc2> + DB 0,143,0,0,0,126 ; add %cl,0x7e000000(%edi) DB 0,0 ; add %al,(%eax) - DB 115,0 ; jae 96ca <_sk_load_8888_sse2_8bit+0xc6> + DB 0,115,0 ; add %dh,0x0(%ebx) DB 0,0 ; add %al,(%eax) PUBLIC _sk_load_8888_dst_sse2_8bit @@ -56525,7 +56462,7 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 96da <_sk_load_8888_dst_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 96ee <_sk_load_8888_dst_sse2_8bit+0xe> DB 91 ; pop %ebx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -56536,9 +56473,9 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,30 ; jne 9714 <_sk_load_8888_dst_sse2_8bit+0x48> - DB 15,16,92,183,16 ; movups 0x10(%edi,%esi,4),%xmm3 + DB 117,30 ; jne 9728 <_sk_load_8888_dst_sse2_8bit+0x48> DB 102,15,16,20,183 ; movupd (%edi,%esi,4),%xmm2 + DB 15,16,92,183,16 ; movups 0x10(%edi,%esi,4),%xmm3 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 255,117,20 ; pushl 0x14(%ebp) DB 86 ; push %esi @@ -56557,16 +56494,16 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,216 ; ja 9700 <_sk_load_8888_dst_sse2_8bit+0x34> + DB 119,216 ; ja 9714 <_sk_load_8888_dst_sse2_8bit+0x34> DB 15,182,201 ; movzbl %cl,%ecx - DB 3,156,139,158,0,0,0 ; add 0x9e(%ebx,%ecx,4),%ebx + DB 3,156,139,162,0,0,0 ; add 0xa2(%ebx,%ecx,4),%ebx DB 255,227 ; jmp *%ebx DB 243,15,16,20,183 ; movss (%edi,%esi,4),%xmm2 - DB 235,197 ; jmp 9700 <_sk_load_8888_dst_sse2_8bit+0x34> + DB 235,197 ; jmp 9714 <_sk_load_8888_dst_sse2_8bit+0x34> DB 102,15,110,84,183,8 ; movd 0x8(%edi,%esi,4),%xmm2 DB 102,15,112,210,69 ; pshufd $0x45,%xmm2,%xmm2 DB 102,15,18,20,183 ; movlpd (%edi,%esi,4),%xmm2 - DB 235,179 ; jmp 9700 <_sk_load_8888_dst_sse2_8bit+0x34> + DB 235,179 ; jmp 9714 <_sk_load_8888_dst_sse2_8bit+0x34> DB 102,15,110,84,183,24 ; movd 0x18(%edi,%esi,4),%xmm2 DB 102,15,112,218,69 ; pshufd $0x45,%xmm2,%xmm3 DB 243,15,16,84,183,20 ; movss 0x14(%edi,%esi,4),%xmm2 @@ -56575,20 +56512,19 @@ _sk_load_8888_dst_sse2_8bit LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 243,15,16,84,183,16 ; movss 0x10(%edi,%esi,4),%xmm2 DB 243,15,16,218 ; movss %xmm2,%xmm3 - DB 235,134 ; jmp 96fb <_sk_load_8888_dst_sse2_8bit+0x2f> - DB 15,31,0 ; nopl (%eax) + DB 102,15,16,20,183 ; movupd (%edi,%esi,4),%xmm2 + DB 235,134 ; jmp 9714 <_sk_load_8888_dst_sse2_8bit+0x34> + DB 102,144 ; xchg %ax,%ax DB 90 ; pop %edx DB 0,0 ; add %al,(%eax) DB 0,108,0,0 ; add %ch,0x0(%eax,%eax,1) DB 0,97,0 ; add %ah,0x0(%ecx) DB 0,0 ; add %al,(%eax) - DB 33,0 ; and %eax,(%eax) - DB 0,0 ; add %al,(%eax) - DB 143,0 ; popl (%eax) + DB 153 ; cltd DB 0,0 ; add %al,(%eax) - DB 126,0 ; jle 978e <_sk_load_8888_dst_sse2_8bit+0xc2> + DB 0,143,0,0,0,126 ; add %cl,0x7e000000(%edi) DB 0,0 ; add %al,(%eax) - DB 115,0 ; jae 9792 <_sk_load_8888_dst_sse2_8bit+0xc6> + DB 0,115,0 ; add %dh,0x0(%ebx) DB 0,0 ; add %al,(%eax) PUBLIC _sk_store_8888_sse2_8bit @@ -56599,7 +56535,7 @@ _sk_store_8888_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 97a2 <_sk_store_8888_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 97ba <_sk_store_8888_sse2_8bit+0xe> DB 91 ; pop %ebx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -56610,7 +56546,7 @@ _sk_store_8888_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,31 ; jne 97dd <_sk_store_8888_sse2_8bit+0x49> + DB 117,31 ; jne 97f5 <_sk_store_8888_sse2_8bit+0x49> DB 243,15,127,4,183 ; movdqu %xmm0,(%edi,%esi,4) DB 243,15,127,76,183,16 ; movdqu %xmm1,0x10(%edi,%esi,4) DB 141,72,8 ; lea 0x8(%eax),%ecx @@ -56629,23 +56565,23 @@ _sk_store_8888_sse2_8bit LABEL PROC DB 128,225,7 ; and $0x7,%cl DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,224 ; ja 97c9 <_sk_store_8888_sse2_8bit+0x35> + DB 119,224 ; ja 97e1 <_sk_store_8888_sse2_8bit+0x35> DB 15,182,201 ; movzbl %cl,%ecx DB 3,156,139,146,0,0,0 ; add 0x92(%ebx,%ecx,4),%ebx DB 255,227 ; jmp *%ebx DB 102,15,126,4,183 ; movd %xmm0,(%edi,%esi,4) - DB 235,205 ; jmp 97c9 <_sk_store_8888_sse2_8bit+0x35> + DB 235,205 ; jmp 97e1 <_sk_store_8888_sse2_8bit+0x35> DB 102,15,112,224,78 ; pshufd $0x4e,%xmm0,%xmm4 DB 102,15,126,100,183,8 ; movd %xmm4,0x8(%edi,%esi,4) DB 102,15,214,4,183 ; movq %xmm0,(%edi,%esi,4) - DB 235,187 ; jmp 97c9 <_sk_store_8888_sse2_8bit+0x35> + DB 235,187 ; jmp 97e1 <_sk_store_8888_sse2_8bit+0x35> DB 102,15,112,225,78 ; pshufd $0x4e,%xmm1,%xmm4 DB 102,15,126,100,183,24 ; movd %xmm4,0x18(%edi,%esi,4) DB 102,15,112,225,229 ; pshufd $0xe5,%xmm1,%xmm4 DB 102,15,126,100,183,20 ; movd %xmm4,0x14(%edi,%esi,4) DB 102,15,126,76,183,16 ; movd %xmm1,0x10(%edi,%esi,4) DB 243,15,127,4,183 ; movdqu %xmm0,(%edi,%esi,4) - DB 235,152 ; jmp 97c9 <_sk_store_8888_sse2_8bit+0x35> + DB 235,152 ; jmp 97e1 <_sk_store_8888_sse2_8bit+0x35> DB 15,31,0 ; nopl (%eax) DB 83 ; push %ebx DB 0,0 ; add %al,(%eax) @@ -56669,7 +56605,7 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 985e <_sk_load_bgra_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 9876 <_sk_load_bgra_sse2_8bit+0xe> DB 91 ; pop %ebx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -56680,9 +56616,9 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,106 ; jne 98e4 <_sk_load_bgra_sse2_8bit+0x94> - DB 15,16,76,183,16 ; movups 0x10(%edi,%esi,4),%xmm1 + DB 117,106 ; jne 98fc <_sk_load_bgra_sse2_8bit+0x94> DB 102,15,16,4,183 ; movupd (%edi,%esi,4),%xmm0 + DB 15,16,76,183,16 ; movups 0x10(%edi,%esi,4),%xmm1 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,40,232 ; movapd %xmm0,%xmm5 DB 102,15,104,236 ; punpckhbw %xmm4,%xmm5 @@ -56718,16 +56654,16 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,140 ; ja 9884 <_sk_load_bgra_sse2_8bit+0x34> + DB 119,140 ; ja 989c <_sk_load_bgra_sse2_8bit+0x34> DB 15,182,201 ; movzbl %cl,%ecx - DB 3,156,139,242,0,0,0 ; add 0xf2(%ebx,%ecx,4),%ebx + DB 3,156,139,246,0,0,0 ; add 0xf6(%ebx,%ecx,4),%ebx DB 255,227 ; jmp *%ebx DB 243,15,16,4,183 ; movss (%edi,%esi,4),%xmm0 - DB 233,118,255,255,255 ; jmp 9884 <_sk_load_bgra_sse2_8bit+0x34> + DB 233,118,255,255,255 ; jmp 989c <_sk_load_bgra_sse2_8bit+0x34> DB 102,15,110,68,183,8 ; movd 0x8(%edi,%esi,4),%xmm0 DB 102,15,112,192,69 ; pshufd $0x45,%xmm0,%xmm0 DB 102,15,18,4,183 ; movlpd (%edi,%esi,4),%xmm0 - DB 233,97,255,255,255 ; jmp 9884 <_sk_load_bgra_sse2_8bit+0x34> + DB 233,97,255,255,255 ; jmp 989c <_sk_load_bgra_sse2_8bit+0x34> DB 102,15,110,68,183,24 ; movd 0x18(%edi,%esi,4),%xmm0 DB 102,15,112,200,69 ; pshufd $0x45,%xmm0,%xmm1 DB 243,15,16,68,183,20 ; movss 0x14(%edi,%esi,4),%xmm0 @@ -56736,13 +56672,14 @@ _sk_load_bgra_sse2_8bit LABEL PROC DB 15,40,200 ; movaps %xmm0,%xmm1 DB 243,15,16,68,183,16 ; movss 0x10(%edi,%esi,4),%xmm0 DB 243,15,16,200 ; movss %xmm0,%xmm1 - DB 233,49,255,255,255 ; jmp 987f <_sk_load_bgra_sse2_8bit+0x2f> - DB 102,144 ; xchg %ax,%ax + DB 102,15,16,4,183 ; movupd (%edi,%esi,4),%xmm0 + DB 233,49,255,255,255 ; jmp 989c <_sk_load_bgra_sse2_8bit+0x34> + DB 144 ; nop DB 166 ; cmpsb %es:(%edi),%ds:(%esi) DB 0,0 ; add %al,(%eax) DB 0,187,0,0,0,176 ; add %bh,-0x50000000(%ebx) DB 0,0 ; add %al,(%eax) - DB 0,33 ; add %ah,(%ecx) + DB 0,235 ; add %ch,%bl DB 0,0 ; add %al,(%eax) DB 0,225 ; add %ah,%cl DB 0,0 ; add %al,(%eax) @@ -56760,7 +56697,7 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 997a <_sk_load_bgra_dst_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 9996 <_sk_load_bgra_dst_sse2_8bit+0xe> DB 91 ; pop %ebx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -56771,9 +56708,9 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,106 ; jne 9a00 <_sk_load_bgra_dst_sse2_8bit+0x94> - DB 15,16,92,183,16 ; movups 0x10(%edi,%esi,4),%xmm3 + DB 117,106 ; jne 9a1c <_sk_load_bgra_dst_sse2_8bit+0x94> DB 102,15,16,20,183 ; movupd (%edi,%esi,4),%xmm2 + DB 15,16,92,183,16 ; movups 0x10(%edi,%esi,4),%xmm3 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,40,234 ; movapd %xmm2,%xmm5 DB 102,15,104,236 ; punpckhbw %xmm4,%xmm5 @@ -56809,16 +56746,16 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,140 ; ja 99a0 <_sk_load_bgra_dst_sse2_8bit+0x34> + DB 119,140 ; ja 99bc <_sk_load_bgra_dst_sse2_8bit+0x34> DB 15,182,201 ; movzbl %cl,%ecx - DB 3,156,139,242,0,0,0 ; add 0xf2(%ebx,%ecx,4),%ebx + DB 3,156,139,246,0,0,0 ; add 0xf6(%ebx,%ecx,4),%ebx DB 255,227 ; jmp *%ebx DB 243,15,16,20,183 ; movss (%edi,%esi,4),%xmm2 - DB 233,118,255,255,255 ; jmp 99a0 <_sk_load_bgra_dst_sse2_8bit+0x34> + DB 233,118,255,255,255 ; jmp 99bc <_sk_load_bgra_dst_sse2_8bit+0x34> DB 102,15,110,84,183,8 ; movd 0x8(%edi,%esi,4),%xmm2 DB 102,15,112,210,69 ; pshufd $0x45,%xmm2,%xmm2 DB 102,15,18,20,183 ; movlpd (%edi,%esi,4),%xmm2 - DB 233,97,255,255,255 ; jmp 99a0 <_sk_load_bgra_dst_sse2_8bit+0x34> + DB 233,97,255,255,255 ; jmp 99bc <_sk_load_bgra_dst_sse2_8bit+0x34> DB 102,15,110,84,183,24 ; movd 0x18(%edi,%esi,4),%xmm2 DB 102,15,112,218,69 ; pshufd $0x45,%xmm2,%xmm3 DB 243,15,16,84,183,20 ; movss 0x14(%edi,%esi,4),%xmm2 @@ -56827,13 +56764,14 @@ _sk_load_bgra_dst_sse2_8bit LABEL PROC DB 15,40,218 ; movaps %xmm2,%xmm3 DB 243,15,16,84,183,16 ; movss 0x10(%edi,%esi,4),%xmm2 DB 243,15,16,218 ; movss %xmm2,%xmm3 - DB 233,49,255,255,255 ; jmp 999b <_sk_load_bgra_dst_sse2_8bit+0x2f> - DB 102,144 ; xchg %ax,%ax + DB 102,15,16,20,183 ; movupd (%edi,%esi,4),%xmm2 + DB 233,49,255,255,255 ; jmp 99bc <_sk_load_bgra_dst_sse2_8bit+0x34> + DB 144 ; nop DB 166 ; cmpsb %es:(%edi),%ds:(%esi) DB 0,0 ; add %al,(%eax) DB 0,187,0,0,0,176 ; add %bh,-0x50000000(%ebx) DB 0,0 ; add %al,(%eax) - DB 0,33 ; add %ah,(%ecx) + DB 0,235 ; add %ch,%bl DB 0,0 ; add %al,(%eax) DB 0,225 ; add %ah,%cl DB 0,0 ; add %al,(%eax) @@ -56851,8 +56789,8 @@ _sk_store_bgra_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 9a96 <_sk_store_bgra_sse2_8bit+0xe> - DB 95 ; pop %edi + DB 232,0,0,0,0 ; call 9ab6 <_sk_store_bgra_sse2_8bit+0xe> + DB 91 ; pop %ebx DB 139,77,20 ; mov 0x14(%ebp),%ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -56881,13 +56819,13 @@ _sk_store_bgra_sse2_8bit LABEL PROC DB 243,15,112,237,198 ; pshufhw $0xc6,%xmm5,%xmm5 DB 102,15,103,238 ; packuswb %xmm6,%xmm5 DB 133,210 ; test %edx,%edx - DB 139,93,16 ; mov 0x10(%ebp),%ebx - DB 117,29 ; jne 9b25 <_sk_store_bgra_sse2_8bit+0x9d> - DB 243,15,127,44,158 ; movdqu %xmm5,(%esi,%ebx,4) - DB 243,15,127,100,158,16 ; movdqu %xmm4,0x10(%esi,%ebx,4) + DB 139,125,16 ; mov 0x10(%ebp),%edi + DB 117,29 ; jne 9b45 <_sk_store_bgra_sse2_8bit+0x9d> + DB 243,15,127,44,190 ; movdqu %xmm5,(%esi,%edi,4) + DB 243,15,127,100,190,16 ; movdqu %xmm4,0x10(%esi,%edi,4) DB 141,112,8 ; lea 0x8(%eax),%esi DB 81 ; push %ecx - DB 83 ; push %ebx + DB 87 ; push %edi DB 86 ; push %esi DB 82 ; push %edx DB 255,80,4 ; call *0x4(%eax) @@ -56897,46 +56835,42 @@ _sk_store_bgra_sse2_8bit LABEL PROC DB 91 ; pop %ebx DB 93 ; pop %ebp DB 195 ; ret - DB 137,125,236 ; mov %edi,-0x14(%ebp) - DB 139,85,8 ; mov 0x8(%ebp),%edx - DB 136,85,243 ; mov %dl,-0xd(%ebp) - DB 139,85,8 ; mov 0x8(%ebp),%edx - DB 128,101,243,7 ; andb $0x7,-0xd(%ebp) - DB 254,77,243 ; decb -0xd(%ebp) - DB 128,125,243,6 ; cmpb $0x6,-0xd(%ebp) - DB 119,213 ; ja 9b13 <_sk_store_bgra_sse2_8bit+0x8b> - DB 15,182,125,243 ; movzbl -0xd(%ebp),%edi - DB 139,77,236 ; mov -0x14(%ebp),%ecx - DB 3,140,185,254,0,0,0 ; add 0xfe(%ecx,%edi,4),%ecx + DB 137,93,240 ; mov %ebx,-0x10(%ebp) + DB 137,211 ; mov %edx,%ebx + DB 128,227,7 ; and $0x7,%bl + DB 254,203 ; dec %bl + DB 128,251,6 ; cmp $0x6,%bl + DB 119,223 ; ja 9b33 <_sk_store_bgra_sse2_8bit+0x8b> + DB 15,182,219 ; movzbl %bl,%ebx + DB 139,77,240 ; mov -0x10(%ebp),%ecx + DB 3,140,153,238,0,0,0 ; add 0xee(%ecx,%ebx,4),%ecx DB 255,225 ; jmp *%ecx - DB 102,15,126,44,158 ; movd %xmm5,(%esi,%ebx,4) - DB 139,77,20 ; mov 0x14(%ebp),%ecx - DB 235,187 ; jmp 9b13 <_sk_store_bgra_sse2_8bit+0x8b> + DB 102,15,126,44,190 ; movd %xmm5,(%esi,%edi,4) + DB 235,51 ; jmp 9b9d <_sk_store_bgra_sse2_8bit+0xf5> DB 102,15,112,229,78 ; pshufd $0x4e,%xmm5,%xmm4 - DB 102,15,126,100,158,8 ; movd %xmm4,0x8(%esi,%ebx,4) - DB 102,15,214,44,158 ; movq %xmm5,(%esi,%ebx,4) - DB 139,77,20 ; mov 0x14(%ebp),%ecx - DB 235,166 ; jmp 9b13 <_sk_store_bgra_sse2_8bit+0x8b> + DB 102,15,126,100,190,8 ; movd %xmm4,0x8(%esi,%edi,4) + DB 102,15,214,44,190 ; movq %xmm5,(%esi,%edi,4) + DB 235,33 ; jmp 9b9d <_sk_store_bgra_sse2_8bit+0xf5> DB 102,15,112,244,78 ; pshufd $0x4e,%xmm4,%xmm6 - DB 102,15,126,116,158,24 ; movd %xmm6,0x18(%esi,%ebx,4) + DB 102,15,126,116,190,24 ; movd %xmm6,0x18(%esi,%edi,4) DB 102,15,112,244,229 ; pshufd $0xe5,%xmm4,%xmm6 - DB 102,15,126,116,158,20 ; movd %xmm6,0x14(%esi,%ebx,4) - DB 102,15,126,100,158,16 ; movd %xmm4,0x10(%esi,%ebx,4) - DB 243,15,127,44,158 ; movdqu %xmm5,(%esi,%ebx,4) + DB 102,15,126,116,190,20 ; movd %xmm6,0x14(%esi,%edi,4) + DB 102,15,126,100,190,16 ; movd %xmm4,0x10(%esi,%edi,4) + DB 243,15,127,44,190 ; movdqu %xmm5,(%esi,%edi,4) DB 139,77,20 ; mov 0x14(%ebp),%ecx - DB 235,128 ; jmp 9b13 <_sk_store_bgra_sse2_8bit+0x8b> - DB 144 ; nop - DB 184,0,0,0,205 ; mov $0xcd000000,%eax + DB 235,145 ; jmp 9b33 <_sk_store_bgra_sse2_8bit+0x8b> + DB 102,144 ; xchg %ax,%ax + DB 173 ; lods %ds:(%esi),%eax DB 0,0 ; add %al,(%eax) - DB 0,194 ; add %al,%dl + DB 0,191,0,0,0,180 ; add %bh,-0x4c000000(%edi) DB 0,0 ; add %al,(%eax) - DB 0,243 ; add %dh,%bl + DB 0,226 ; add %ah,%dl DB 0,0 ; add %al,(%eax) - DB 0,237 ; add %ch,%ch + DB 0,220 ; add %bl,%ah DB 0,0 ; add %al,(%eax) - DB 0,226 ; add %ah,%dl + DB 0,209 ; add %dl,%cl DB 0,0 ; add %al,(%eax) - DB 0,215 ; add %dl,%bh + DB 0,198 ; add %al,%dh DB 0,0 ; add %al,(%eax) DB 0 ; .byte 0x0 @@ -56948,99 +56882,94 @@ _sk_store_565_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 9bbe <_sk_store_565_sse2_8bit+0xe> - DB 88 ; pop %eax - DB 102,15,111,176,66,59,0,0 ; movdqa 0x3b42(%eax),%xmm6 - DB 102,15,111,249 ; movdqa %xmm1,%xmm7 - DB 102,15,219,254 ; pand %xmm6,%xmm7 - DB 102,15,219,240 ; pand %xmm0,%xmm6 - DB 102,15,114,246,8 ; pslld $0x8,%xmm6 - DB 102,15,114,247,8 ; pslld $0x8,%xmm7 - DB 102,15,111,160,82,59,0,0 ; movdqa 0x3b52(%eax),%xmm4 + DB 232,0,0,0,0 ; call 9bce <_sk_store_565_sse2_8bit+0xe> + DB 95 ; pop %edi + DB 139,69,12 ; mov 0xc(%ebp),%eax + DB 139,16 ; mov (%eax),%edx + DB 139,114,4 ; mov 0x4(%edx),%esi + DB 15,175,117,20 ; imul 0x14(%ebp),%esi + DB 1,246 ; add %esi,%esi + DB 3,50 ; add (%edx),%esi + DB 139,85,8 ; mov 0x8(%ebp),%edx + DB 102,15,111,167,82,58,0,0 ; movdqa 0x3a52(%edi),%xmm4 DB 102,15,111,233 ; movdqa %xmm1,%xmm5 DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,15,114,212,5 ; psrld $0x5,%xmm4 - DB 102,15,114,213,5 ; psrld $0x5,%xmm5 + DB 102,15,114,244,8 ; pslld $0x8,%xmm4 + DB 102,15,114,245,8 ; pslld $0x8,%xmm5 + DB 102,15,111,183,98,58,0,0 ; movdqa 0x3a62(%edi),%xmm6 + DB 102,15,111,249 ; movdqa %xmm1,%xmm7 + DB 102,15,219,254 ; pand %xmm6,%xmm7 + DB 102,15,219,240 ; pand %xmm0,%xmm6 + DB 102,15,114,214,5 ; psrld $0x5,%xmm6 + DB 102,15,235,244 ; por %xmm4,%xmm6 + DB 102,15,114,215,5 ; psrld $0x5,%xmm7 + DB 102,15,235,253 ; por %xmm5,%xmm7 + DB 102,15,111,175,114,58,0,0 ; movdqa 0x3a72(%edi),%xmm5 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,219,229 ; pand %xmm5,%xmm4 + DB 102,15,219,233 ; pand %xmm1,%xmm5 + DB 102,15,114,213,19 ; psrld $0x13,%xmm5 DB 102,15,235,239 ; por %xmm7,%xmm5 - DB 102,15,111,184,98,59,0,0 ; movdqa 0x3b62(%eax),%xmm7 - DB 102,15,235,230 ; por %xmm6,%xmm4 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,15,219,247 ; pand %xmm7,%xmm6 - DB 102,15,114,214,19 ; psrld $0x13,%xmm6 + DB 102,15,114,212,19 ; psrld $0x13,%xmm4 DB 102,15,235,230 ; por %xmm6,%xmm4 - DB 102,15,219,249 ; pand %xmm1,%xmm7 - DB 102,15,114,215,19 ; psrld $0x13,%xmm7 - DB 102,15,235,239 ; por %xmm7,%xmm5 DB 102,15,114,245,16 ; pslld $0x10,%xmm5 DB 102,15,114,229,16 ; psrad $0x10,%xmm5 DB 102,15,114,244,16 ; pslld $0x10,%xmm4 DB 102,15,114,228,16 ; psrad $0x10,%xmm4 DB 102,15,107,229 ; packssdw %xmm5,%xmm4 - DB 139,85,20 ; mov 0x14(%ebp),%edx - DB 139,77,12 ; mov 0xc(%ebp),%ecx - DB 139,49 ; mov (%ecx),%esi - DB 139,126,4 ; mov 0x4(%esi),%edi - DB 15,175,250 ; imul %edx,%edi - DB 1,255 ; add %edi,%edi - DB 3,62 ; add (%esi),%edi - DB 139,93,8 ; mov 0x8(%ebp),%ebx - DB 133,219 ; test %ebx,%ebx - DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,23 ; jne 9c74 <_sk_store_565_sse2_8bit+0xc4> - DB 243,15,127,36,119 ; movdqu %xmm4,(%edi,%esi,2) - DB 141,65,8 ; lea 0x8(%ecx),%eax - DB 82 ; push %edx - DB 86 ; push %esi - DB 80 ; push %eax + DB 133,210 ; test %edx,%edx + DB 139,93,16 ; mov 0x10(%ebp),%ebx + DB 117,25 ; jne 9c84 <_sk_store_565_sse2_8bit+0xc4> + DB 243,15,127,36,94 ; movdqu %xmm4,(%esi,%ebx,2) + DB 141,72,8 ; lea 0x8(%eax),%ecx + DB 255,117,20 ; pushl 0x14(%ebp) DB 83 ; push %ebx - DB 255,81,4 ; call *0x4(%ecx) + DB 81 ; push %ecx + DB 82 ; push %edx + DB 255,80,4 ; call *0x4(%eax) DB 131,196,28 ; add $0x1c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx DB 93 ; pop %ebp DB 195 ; ret - DB 139,93,8 ; mov 0x8(%ebp),%ebx - DB 136,93,243 ; mov %bl,-0xd(%ebp) - DB 139,93,8 ; mov 0x8(%ebp),%ebx - DB 128,101,243,7 ; andb $0x7,-0xd(%ebp) - DB 254,77,243 ; decb -0xd(%ebp) - DB 128,125,243,6 ; cmpb $0x6,-0xd(%ebp) - DB 119,216 ; ja 9c62 <_sk_store_565_sse2_8bit+0xb2> - DB 15,182,85,243 ; movzbl -0xd(%ebp),%edx - DB 3,132,144,34,1,0,0 ; add 0x122(%eax,%edx,4),%eax - DB 255,224 ; jmp *%eax - DB 102,15,126,224 ; movd %xmm4,%eax - DB 102,137,4,119 ; mov %ax,(%edi,%esi,2) - DB 139,85,20 ; mov 0x14(%ebp),%edx - DB 235,190 ; jmp 9c62 <_sk_store_565_sse2_8bit+0xb2> - DB 102,15,197,196,2 ; pextrw $0x2,%xmm4,%eax - DB 102,137,68,119,4 ; mov %ax,0x4(%edi,%esi,2) - DB 102,15,126,36,119 ; movd %xmm4,(%edi,%esi,2) - DB 139,85,20 ; mov 0x14(%ebp),%edx - DB 235,170 ; jmp 9c62 <_sk_store_565_sse2_8bit+0xb2> - DB 102,15,197,196,6 ; pextrw $0x6,%xmm4,%eax - DB 102,137,68,119,12 ; mov %ax,0xc(%edi,%esi,2) - DB 102,15,197,196,5 ; pextrw $0x5,%xmm4,%eax - DB 102,137,68,119,10 ; mov %ax,0xa(%edi,%esi,2) - DB 102,15,197,196,4 ; pextrw $0x4,%xmm4,%eax - DB 102,137,68,119,8 ; mov %ax,0x8(%edi,%esi,2) - DB 102,15,214,36,119 ; movq %xmm4,(%edi,%esi,2) - DB 139,85,20 ; mov 0x14(%ebp),%edx - DB 235,130 ; jmp 9c62 <_sk_store_565_sse2_8bit+0xb2> - DB 217,0 ; flds (%eax) + DB 137,209 ; mov %edx,%ecx + DB 128,225,7 ; and $0x7,%cl + DB 254,201 ; dec %cl + DB 128,249,6 ; cmp $0x6,%cl + DB 119,224 ; ja 9c70 <_sk_store_565_sse2_8bit+0xb0> + DB 15,182,201 ; movzbl %cl,%ecx + DB 3,188,143,14,1,0,0 ; add 0x10e(%edi,%ecx,4),%edi + DB 255,231 ; jmp *%edi + DB 102,15,126,225 ; movd %xmm4,%ecx + DB 102,137,12,94 ; mov %cx,(%esi,%ebx,2) + DB 235,202 ; jmp 9c70 <_sk_store_565_sse2_8bit+0xb0> + DB 102,15,197,204,2 ; pextrw $0x2,%xmm4,%ecx + DB 102,137,76,94,4 ; mov %cx,0x4(%esi,%ebx,2) + DB 102,15,126,36,94 ; movd %xmm4,(%esi,%ebx,2) + DB 235,185 ; jmp 9c70 <_sk_store_565_sse2_8bit+0xb0> + DB 102,15,197,204,6 ; pextrw $0x6,%xmm4,%ecx + DB 102,137,76,94,12 ; mov %cx,0xc(%esi,%ebx,2) + DB 102,15,197,204,5 ; pextrw $0x5,%xmm4,%ecx + DB 102,137,76,94,10 ; mov %cx,0xa(%esi,%ebx,2) + DB 102,15,197,204,4 ; pextrw $0x4,%xmm4,%ecx + DB 102,137,76,94,8 ; mov %cx,0x8(%esi,%ebx,2) + DB 102,15,214,36,94 ; movq %xmm4,(%esi,%ebx,2) + DB 235,148 ; jmp 9c70 <_sk_store_565_sse2_8bit+0xb0> + DB 206 ; into DB 0,0 ; add %al,(%eax) - DB 240,0,0 ; lock add %al,(%eax) - DB 0,230 ; add %ah,%dh + DB 0,226 ; add %ah,%dl DB 0,0 ; add %al,(%eax) - DB 0,24 ; add %bl,(%eax) - DB 1,0 ; add %eax,(%eax) - DB 0,14 ; add %cl,(%esi) + DB 0,216 ; add %bl,%al + DB 0,0 ; add %al,(%eax) + DB 0,7 ; add %al,(%edi) DB 1,0 ; add %eax,(%eax) - DB 0,4,1 ; add %al,(%ecx,%eax,1) + DB 0,253 ; add %bh,%ch + DB 0,0 ; add %al,(%eax) + DB 0,243 ; add %dh,%bl DB 0,0 ; add %al,(%eax) - DB 250 ; cli + DB 0,233 ; add %ch,%cl DB 0,0 ; add %al,(%eax) DB 0 ; .byte 0x0 @@ -57053,7 +56982,7 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 86 ; push %esi DB 131,236,28 ; sub $0x1c,%esp DB 15,41,77,216 ; movaps %xmm1,-0x28(%ebp) - DB 232,0,0,0,0 ; call 9d0e <_sk_load_565_dst_sse2_8bit+0x12> + DB 232,0,0,0,0 ; call 9d0a <_sk_load_565_dst_sse2_8bit+0x12> DB 94 ; pop %esi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57064,31 +56993,31 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,125,16 ; mov 0x10(%ebp),%edi - DB 15,133,235,0,0,0 ; jne 9e18 <_sk_load_565_dst_sse2_8bit+0x11c> + DB 15,133,235,0,0,0 ; jne 9e14 <_sk_load_565_dst_sse2_8bit+0x11c> DB 243,15,111,20,123 ; movdqu (%ebx,%edi,2),%xmm2 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,111,218 ; movdqa %xmm2,%xmm3 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 DB 102,15,97,212 ; punpcklwd %xmm4,%xmm2 - DB 102,15,111,182,66,53,0,0 ; movdqa 0x3542(%esi),%xmm6 + DB 102,15,111,182,54,52,0,0 ; movdqa 0x3436(%esi),%xmm6 DB 102,15,111,250 ; movdqa %xmm2,%xmm7 DB 102,15,219,254 ; pand %xmm6,%xmm7 DB 102,15,219,243 ; pand %xmm3,%xmm6 DB 102,15,114,214,8 ; psrld $0x8,%xmm6 DB 102,15,114,215,8 ; psrld $0x8,%xmm7 - DB 102,15,111,174,130,53,0,0 ; movdqa 0x3582(%esi),%xmm5 + DB 102,15,111,174,118,52,0,0 ; movdqa 0x3476(%esi),%xmm5 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,15,219,229 ; pand %xmm5,%xmm4 DB 102,15,219,235 ; pand %xmm3,%xmm5 DB 102,15,114,245,19 ; pslld $0x13,%xmm5 + DB 102,15,235,238 ; por %xmm6,%xmm5 DB 102,15,114,244,19 ; pslld $0x13,%xmm4 DB 102,15,235,231 ; por %xmm7,%xmm4 - DB 102,15,235,238 ; por %xmm6,%xmm5 - DB 102,15,111,182,98,53,0,0 ; movdqa 0x3562(%esi),%xmm6 + DB 102,15,111,182,86,52,0,0 ; movdqa 0x3456(%esi),%xmm6 DB 102,15,111,251 ; movdqa %xmm3,%xmm7 DB 102,15,219,254 ; pand %xmm6,%xmm7 DB 102,15,219,242 ; pand %xmm2,%xmm6 - DB 102,15,111,142,50,58,0,0 ; movdqa 0x3a32(%esi),%xmm1 + DB 102,15,111,142,86,57,0,0 ; movdqa 0x3956(%esi),%xmm1 DB 102,15,219,217 ; pand %xmm1,%xmm3 DB 102,15,219,209 ; pand %xmm1,%xmm2 DB 102,15,114,247,5 ; pslld $0x5,%xmm7 @@ -57101,10 +57030,10 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 102,15,235,214 ; por %xmm6,%xmm2 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,114,212,5 ; psrld $0x5,%xmm4 - DB 102,15,111,142,34,58,0,0 ; movdqa 0x3a22(%esi),%xmm1 + DB 102,15,111,142,70,57,0,0 ; movdqa 0x3946(%esi),%xmm1 DB 102,15,219,225 ; pand %xmm1,%xmm4 DB 102,15,219,233 ; pand %xmm1,%xmm5 - DB 102,15,111,142,226,57,0,0 ; movdqa 0x39e2(%esi),%xmm1 + DB 102,15,111,142,6,57,0,0 ; movdqa 0x3906(%esi),%xmm1 DB 102,15,235,209 ; por %xmm1,%xmm2 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,217 ; por %xmm1,%xmm3 @@ -57127,25 +57056,25 @@ _sk_load_565_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 15,135,6,255,255,255 ; ja 9d32 <_sk_load_565_dst_sse2_8bit+0x36> + DB 15,135,6,255,255,255 ; ja 9d2e <_sk_load_565_dst_sse2_8bit+0x36> DB 15,182,201 ; movzbl %cl,%ecx DB 139,140,142,118,1,0,0 ; mov 0x176(%esi,%ecx,4),%ecx DB 1,241 ; add %esi,%ecx DB 255,225 ; jmp *%ecx DB 15,183,12,123 ; movzwl (%ebx,%edi,2),%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 233,235,254,255,255 ; jmp 9d32 <_sk_load_565_dst_sse2_8bit+0x36> + DB 233,235,254,255,255 ; jmp 9d2e <_sk_load_565_dst_sse2_8bit+0x36> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,84,123,4,2 ; pinsrw $0x2,0x4(%ebx,%edi,2),%xmm2 DB 243,15,16,28,123 ; movss (%ebx,%edi,2),%xmm3 DB 243,15,16,211 ; movss %xmm3,%xmm2 - DB 233,210,254,255,255 ; jmp 9d32 <_sk_load_565_dst_sse2_8bit+0x36> + DB 233,210,254,255,255 ; jmp 9d2e <_sk_load_565_dst_sse2_8bit+0x36> DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,84,123,12,6 ; pinsrw $0x6,0xc(%ebx,%edi,2),%xmm2 DB 102,15,196,84,123,10,5 ; pinsrw $0x5,0xa(%ebx,%edi,2),%xmm2 DB 102,15,196,84,123,8,4 ; pinsrw $0x4,0x8(%ebx,%edi,2),%xmm2 DB 102,15,18,20,123 ; movlpd (%ebx,%edi,2),%xmm2 - DB 233,175,254,255,255 ; jmp 9d32 <_sk_load_565_dst_sse2_8bit+0x36> + DB 233,175,254,255,255 ; jmp 9d2e <_sk_load_565_dst_sse2_8bit+0x36> DB 144 ; nop DB 44,1 ; sub $0x1,%al DB 0,0 ; add %al,(%eax) @@ -57170,7 +57099,7 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 9eae <_sk_load_a8_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 9eaa <_sk_load_a8_sse2_8bit+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57180,10 +57109,10 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,63 ; jne 9f06 <_sk_load_a8_sse2_8bit+0x66> + DB 117,63 ; jne 9f02 <_sk_load_a8_sse2_8bit+0x66> DB 243,15,126,4,51 ; movq (%ebx,%esi,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 102,15,219,135,162,56,0,0 ; pand 0x38a2(%edi),%xmm0 + DB 102,15,219,135,198,55,0,0 ; pand 0x37c6(%edi),%xmm0 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 @@ -57207,14 +57136,14 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,186 ; ja 9ed0 <_sk_load_a8_sse2_8bit+0x30> + DB 119,186 ; ja 9ecc <_sk_load_a8_sse2_8bit+0x30> DB 15,182,201 ; movzbl %cl,%ecx DB 139,140,143,214,0,0,0 ; mov 0xd6(%edi,%ecx,4),%ecx DB 1,249 ; add %edi,%ecx DB 255,225 ; jmp *%ecx DB 15,182,12,51 ; movzbl (%ebx,%esi,1),%ecx DB 102,15,110,193 ; movd %ecx,%xmm0 - DB 235,162 ; jmp 9ed0 <_sk_load_a8_sse2_8bit+0x30> + DB 235,162 ; jmp 9ecc <_sk_load_a8_sse2_8bit+0x30> DB 15,182,76,51,2 ; movzbl 0x2(%ebx,%esi,1),%ecx DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,193,2 ; pinsrw $0x2,%ecx,%xmm0 @@ -57222,7 +57151,7 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 243,15,16,193 ; movss %xmm1,%xmm0 - DB 235,130 ; jmp 9ed0 <_sk_load_a8_sse2_8bit+0x30> + DB 235,130 ; jmp 9ecc <_sk_load_a8_sse2_8bit+0x30> DB 15,182,76,51,6 ; movzbl 0x6(%ebx,%esi,1),%ecx DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,193,6 ; pinsrw $0x6,%ecx,%xmm0 @@ -57233,9 +57162,9 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 102,15,110,12,51 ; movd (%ebx,%esi,1),%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 233,78,255,255,255 ; jmp 9ed0 <_sk_load_a8_sse2_8bit+0x30> + DB 233,78,255,255,255 ; jmp 9ecc <_sk_load_a8_sse2_8bit+0x30> DB 102,144 ; xchg %ax,%ax - DB 118,0 ; jbe 9f86 <_sk_load_a8_sse2_8bit+0xe6> + DB 118,0 ; jbe 9f82 <_sk_load_a8_sse2_8bit+0xe6> DB 0,0 ; add %al,(%eax) DB 142,0 ; mov (%eax),%es DB 0,0 ; add %al,(%eax) @@ -57257,7 +57186,7 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call 9fae <_sk_load_a8_dst_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call 9faa <_sk_load_a8_dst_sse2_8bit+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57267,10 +57196,10 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 117,63 ; jne a006 <_sk_load_a8_dst_sse2_8bit+0x66> + DB 117,63 ; jne a002 <_sk_load_a8_dst_sse2_8bit+0x66> DB 243,15,126,20,51 ; movq (%ebx,%esi,1),%xmm2 DB 102,15,96,208 ; punpcklbw %xmm0,%xmm2 - DB 102,15,219,151,162,55,0,0 ; pand 0x37a2(%edi),%xmm2 + DB 102,15,219,151,198,54,0,0 ; pand 0x36c6(%edi),%xmm2 DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,111,218 ; movdqa %xmm2,%xmm3 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 @@ -57294,14 +57223,14 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,186 ; ja 9fd0 <_sk_load_a8_dst_sse2_8bit+0x30> + DB 119,186 ; ja 9fcc <_sk_load_a8_dst_sse2_8bit+0x30> DB 15,182,201 ; movzbl %cl,%ecx DB 139,140,143,214,0,0,0 ; mov 0xd6(%edi,%ecx,4),%ecx DB 1,249 ; add %edi,%ecx DB 255,225 ; jmp *%ecx DB 15,182,12,51 ; movzbl (%ebx,%esi,1),%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 235,162 ; jmp 9fd0 <_sk_load_a8_dst_sse2_8bit+0x30> + DB 235,162 ; jmp 9fcc <_sk_load_a8_dst_sse2_8bit+0x30> DB 15,182,76,51,2 ; movzbl 0x2(%ebx,%esi,1),%ecx DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,209,2 ; pinsrw $0x2,%ecx,%xmm2 @@ -57309,7 +57238,7 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 243,15,16,211 ; movss %xmm3,%xmm2 - DB 235,130 ; jmp 9fd0 <_sk_load_a8_dst_sse2_8bit+0x30> + DB 235,130 ; jmp 9fcc <_sk_load_a8_dst_sse2_8bit+0x30> DB 15,182,76,51,6 ; movzbl 0x6(%ebx,%esi,1),%ecx DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,209,6 ; pinsrw $0x6,%ecx,%xmm2 @@ -57320,9 +57249,9 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 102,15,110,28,51 ; movd (%ebx,%esi,1),%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 242,15,16,211 ; movsd %xmm3,%xmm2 - DB 233,78,255,255,255 ; jmp 9fd0 <_sk_load_a8_dst_sse2_8bit+0x30> + DB 233,78,255,255,255 ; jmp 9fcc <_sk_load_a8_dst_sse2_8bit+0x30> DB 102,144 ; xchg %ax,%ax - DB 118,0 ; jbe a086 <_sk_load_a8_dst_sse2_8bit+0xe6> + DB 118,0 ; jbe a082 <_sk_load_a8_dst_sse2_8bit+0xe6> DB 0,0 ; add %al,(%eax) DB 142,0 ; mov (%eax),%es DB 0,0 ; add %al,(%eax) @@ -57343,8 +57272,8 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 83 ; push %ebx DB 87 ; push %edi DB 86 ; push %esi - DB 131,236,92 ; sub $0x5c,%esp - DB 232,0,0,0,0 ; call a0ae <_sk_store_a8_sse2_8bit+0xe> + DB 131,236,12 ; sub $0xc,%esp + DB 232,0,0,0,0 ; call a0aa <_sk_store_a8_sse2_8bit+0xe> DB 91 ; pop %ebx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57363,8 +57292,8 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 102,15,107,229 ; packssdw %xmm5,%xmm4 DB 133,210 ; test %edx,%edx DB 139,125,16 ; mov 0x10(%ebp),%edi - DB 117,37 ; jne a116 <_sk_store_a8_sse2_8bit+0x76> - DB 102,15,219,163,162,54,0,0 ; pand 0x36a2(%ebx),%xmm4 + DB 117,37 ; jne a112 <_sk_store_a8_sse2_8bit+0x76> + DB 102,15,219,163,198,53,0,0 ; pand 0x35c6(%ebx),%xmm4 DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,214,36,62 ; movq %xmm4,(%esi,%edi,1) DB 141,72,8 ; lea 0x8(%eax),%ecx @@ -57373,7 +57302,7 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 81 ; push %ecx DB 82 ; push %edx DB 255,80,4 ; call *0x4(%eax) - DB 131,196,108 ; add $0x6c,%esp + DB 131,196,28 ; add $0x1c,%esp DB 94 ; pop %esi DB 95 ; pop %edi DB 91 ; pop %ebx @@ -57383,45 +57312,41 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 128,225,7 ; and $0x7,%cl DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 119,224 ; ja a102 <_sk_store_a8_sse2_8bit+0x62> + DB 119,224 ; ja a0fe <_sk_store_a8_sse2_8bit+0x62> DB 15,182,201 ; movzbl %cl,%ecx - DB 139,140,139,238,0,0,0 ; mov 0xee(%ebx,%ecx,4),%ecx + DB 139,140,139,222,0,0,0 ; mov 0xde(%ebx,%ecx,4),%ecx DB 1,217 ; add %ebx,%ecx DB 255,225 ; jmp *%ecx - DB 102,15,127,101,152 ; movdqa %xmm4,-0x68(%ebp) - DB 138,77,152 ; mov -0x68(%ebp),%cl + DB 102,15,126,225 ; movd %xmm4,%ecx DB 136,12,62 ; mov %cl,(%esi,%edi,1) - DB 235,197 ; jmp a102 <_sk_store_a8_sse2_8bit+0x62> - DB 102,15,127,101,168 ; movdqa %xmm4,-0x58(%ebp) - DB 138,77,172 ; mov -0x54(%ebp),%cl + DB 235,201 ; jmp a0fe <_sk_store_a8_sse2_8bit+0x62> + DB 102,15,197,204,2 ; pextrw $0x2,%xmm4,%ecx DB 136,76,62,2 ; mov %cl,0x2(%esi,%edi,1) - DB 102,15,219,163,162,54,0,0 ; pand 0x36a2(%ebx),%xmm4 + DB 102,15,219,163,198,53,0,0 ; pand 0x35c6(%ebx),%xmm4 DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,126,225 ; movd %xmm4,%ecx DB 102,137,12,62 ; mov %cx,(%esi,%edi,1) - DB 235,163 ; jmp a102 <_sk_store_a8_sse2_8bit+0x62> - DB 102,15,127,101,184 ; movdqa %xmm4,-0x48(%ebp) - DB 138,77,196 ; mov -0x3c(%ebp),%cl + DB 235,170 ; jmp a0fe <_sk_store_a8_sse2_8bit+0x62> + DB 102,15,197,204,6 ; pextrw $0x6,%xmm4,%ecx DB 136,76,62,6 ; mov %cl,0x6(%esi,%edi,1) - DB 102,15,127,101,200 ; movdqa %xmm4,-0x38(%ebp) - DB 138,77,210 ; mov -0x2e(%ebp),%cl + DB 102,15,197,204,5 ; pextrw $0x5,%xmm4,%ecx DB 136,76,62,5 ; mov %cl,0x5(%esi,%edi,1) - DB 102,15,127,101,216 ; movdqa %xmm4,-0x28(%ebp) - DB 138,77,224 ; mov -0x20(%ebp),%cl + DB 102,15,197,204,4 ; pextrw $0x4,%xmm4,%ecx DB 136,76,62,4 ; mov %cl,0x4(%esi,%edi,1) - DB 102,15,219,163,162,54,0,0 ; pand 0x36a2(%ebx),%xmm4 + DB 102,15,219,163,198,53,0,0 ; pand 0x35c6(%ebx),%xmm4 DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,126,36,62 ; movd %xmm4,(%esi,%edi,1) - DB 233,105,255,255,255 ; jmp a102 <_sk_store_a8_sse2_8bit+0x62> + DB 233,121,255,255,255 ; jmp a0fe <_sk_store_a8_sse2_8bit+0x62> DB 15,31,0 ; nopl (%eax) DB 130,0,0 ; addb $0x0,(%eax) - DB 0,155,0,0,0,143 ; add %bl,-0x71000000(%ebx) + DB 0,148,0,0,0,139,0 ; add %dl,0x8b0000(%eax,%eax,1) DB 0,0 ; add %al,(%eax) - DB 0,213 ; add %dl,%ch + DB 197,0 ; lds (%eax),%eax DB 0,0 ; add %al,(%eax) - DB 0,201 ; add %cl,%cl + DB 188,0,0,0,179 ; mov $0xb3000000,%esp DB 0,0 ; add %al,(%eax) - DB 0,189,0,0,0,177 ; add %bh,-0x4f000000(%ebp) + DB 0 ; .byte 0x0 + DB 170 ; stos %al,%es:(%edi) DB 0,0 ; add %al,(%eax) DB 0 ; .byte 0x0 @@ -57433,7 +57358,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call a1c6 <_sk_load_g8_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call a1b2 <_sk_load_g8_sse2_8bit+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57443,15 +57368,15 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 15,133,131,0,0,0 ; jne a266 <_sk_load_g8_sse2_8bit+0xae> + DB 15,133,131,0,0,0 ; jne a252 <_sk_load_g8_sse2_8bit+0xae> DB 243,15,126,4,51 ; movq (%ebx,%esi,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 102,15,219,135,138,53,0,0 ; pand 0x358a(%edi),%xmm0 + DB 102,15,219,135,190,52,0,0 ; pand 0x34be(%edi),%xmm0 DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 DB 102,15,97,225 ; punpcklwd %xmm1,%xmm4 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,15,111,175,154,53,0,0 ; movdqa 0x359a(%edi),%xmm5 + DB 102,15,111,175,206,52,0,0 ; movdqa 0x34ce(%edi),%xmm5 DB 102,15,112,240,245 ; pshufd $0xf5,%xmm0,%xmm6 DB 102,15,244,197 ; pmuludq %xmm5,%xmm0 DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 @@ -57464,7 +57389,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,244,245 ; pmuludq %xmm5,%xmm6 DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 DB 102,15,98,196 ; punpckldq %xmm4,%xmm0 - DB 102,15,111,167,42,53,0,0 ; movdqa 0x352a(%edi),%xmm4 + DB 102,15,111,167,94,52,0,0 ; movdqa 0x345e(%edi),%xmm4 DB 102,15,235,196 ; por %xmm4,%xmm0 DB 102,15,235,204 ; por %xmm4,%xmm1 DB 141,72,8 ; lea 0x8(%eax),%ecx @@ -57484,14 +57409,14 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 15,135,114,255,255,255 ; ja a1ec <_sk_load_g8_sse2_8bit+0x34> + DB 15,135,114,255,255,255 ; ja a1d8 <_sk_load_g8_sse2_8bit+0x34> DB 15,182,201 ; movzbl %cl,%ecx DB 139,140,143,38,1,0,0 ; mov 0x126(%edi,%ecx,4),%ecx DB 1,249 ; add %edi,%ecx DB 255,225 ; jmp *%ecx DB 15,182,12,51 ; movzbl (%ebx,%esi,1),%ecx DB 102,15,110,193 ; movd %ecx,%xmm0 - DB 233,87,255,255,255 ; jmp a1ec <_sk_load_g8_sse2_8bit+0x34> + DB 233,87,255,255,255 ; jmp a1d8 <_sk_load_g8_sse2_8bit+0x34> DB 15,182,76,51,2 ; movzbl 0x2(%ebx,%esi,1),%ecx DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,193,2 ; pinsrw $0x2,%ecx,%xmm0 @@ -57499,7 +57424,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,110,201 ; movd %ecx,%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 243,15,16,193 ; movss %xmm1,%xmm0 - DB 233,52,255,255,255 ; jmp a1ec <_sk_load_g8_sse2_8bit+0x34> + DB 233,52,255,255,255 ; jmp a1d8 <_sk_load_g8_sse2_8bit+0x34> DB 15,182,76,51,6 ; movzbl 0x6(%ebx,%esi,1),%ecx DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,196,193,6 ; pinsrw $0x6,%ecx,%xmm0 @@ -57510,7 +57435,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,110,12,51 ; movd (%ebx,%esi,1),%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 242,15,16,193 ; movsd %xmm1,%xmm0 - DB 233,0,255,255,255 ; jmp a1ec <_sk_load_g8_sse2_8bit+0x34> + DB 233,0,255,255,255 ; jmp a1d8 <_sk_load_g8_sse2_8bit+0x34> DB 194,0,0 ; ret $0x0 DB 0,221 ; add %bl,%ch DB 0,0 ; add %al,(%eax) @@ -57533,7 +57458,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 87 ; push %edi DB 86 ; push %esi DB 131,236,12 ; sub $0xc,%esp - DB 232,0,0,0,0 ; call a316 <_sk_load_g8_dst_sse2_8bit+0xe> + DB 232,0,0,0,0 ; call a302 <_sk_load_g8_dst_sse2_8bit+0xe> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57543,15 +57468,15 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 15,133,131,0,0,0 ; jne a3b6 <_sk_load_g8_dst_sse2_8bit+0xae> + DB 15,133,131,0,0,0 ; jne a3a2 <_sk_load_g8_dst_sse2_8bit+0xae> DB 243,15,126,20,51 ; movq (%ebx,%esi,1),%xmm2 DB 102,15,96,208 ; punpcklbw %xmm0,%xmm2 - DB 102,15,219,151,58,52,0,0 ; pand 0x343a(%edi),%xmm2 + DB 102,15,219,151,110,51,0,0 ; pand 0x336e(%edi),%xmm2 DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 DB 102,15,97,227 ; punpcklwd %xmm3,%xmm4 DB 102,15,105,211 ; punpckhwd %xmm3,%xmm2 - DB 102,15,111,175,74,52,0,0 ; movdqa 0x344a(%edi),%xmm5 + DB 102,15,111,175,126,51,0,0 ; movdqa 0x337e(%edi),%xmm5 DB 102,15,112,242,245 ; pshufd $0xf5,%xmm2,%xmm6 DB 102,15,244,213 ; pmuludq %xmm5,%xmm2 DB 102,15,112,218,232 ; pshufd $0xe8,%xmm2,%xmm3 @@ -57564,7 +57489,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,244,245 ; pmuludq %xmm5,%xmm6 DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 DB 102,15,98,212 ; punpckldq %xmm4,%xmm2 - DB 102,15,111,167,218,51,0,0 ; movdqa 0x33da(%edi),%xmm4 + DB 102,15,111,167,14,51,0,0 ; movdqa 0x330e(%edi),%xmm4 DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,220 ; por %xmm4,%xmm3 DB 141,72,8 ; lea 0x8(%eax),%ecx @@ -57584,14 +57509,14 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 15,135,114,255,255,255 ; ja a33c <_sk_load_g8_dst_sse2_8bit+0x34> + DB 15,135,114,255,255,255 ; ja a328 <_sk_load_g8_dst_sse2_8bit+0x34> DB 15,182,201 ; movzbl %cl,%ecx DB 139,140,143,38,1,0,0 ; mov 0x126(%edi,%ecx,4),%ecx DB 1,249 ; add %edi,%ecx DB 255,225 ; jmp *%ecx DB 15,182,12,51 ; movzbl (%ebx,%esi,1),%ecx DB 102,15,110,209 ; movd %ecx,%xmm2 - DB 233,87,255,255,255 ; jmp a33c <_sk_load_g8_dst_sse2_8bit+0x34> + DB 233,87,255,255,255 ; jmp a328 <_sk_load_g8_dst_sse2_8bit+0x34> DB 15,182,76,51,2 ; movzbl 0x2(%ebx,%esi,1),%ecx DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,209,2 ; pinsrw $0x2,%ecx,%xmm2 @@ -57599,7 +57524,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 243,15,16,211 ; movss %xmm3,%xmm2 - DB 233,52,255,255,255 ; jmp a33c <_sk_load_g8_dst_sse2_8bit+0x34> + DB 233,52,255,255,255 ; jmp a328 <_sk_load_g8_dst_sse2_8bit+0x34> DB 15,182,76,51,6 ; movzbl 0x6(%ebx,%esi,1),%ecx DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,196,209,6 ; pinsrw $0x6,%ecx,%xmm2 @@ -57610,7 +57535,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,110,28,51 ; movd (%ebx,%esi,1),%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 242,15,16,211 ; movsd %xmm3,%xmm2 - DB 233,0,255,255,255 ; jmp a33c <_sk_load_g8_dst_sse2_8bit+0x34> + DB 233,0,255,255,255 ; jmp a328 <_sk_load_g8_dst_sse2_8bit+0x34> DB 194,0,0 ; ret $0x0 DB 0,221 ; add %bl,%ch DB 0,0 ; add %al,(%eax) @@ -57635,7 +57560,7 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 131,236,124 ; sub $0x7c,%esp DB 15,41,157,120,255,255,255 ; movaps %xmm3,-0x88(%ebp) DB 102,15,127,85,136 ; movdqa %xmm2,-0x78(%ebp) - DB 232,0,0,0,0 ; call a472 <_sk_srcover_rgba_8888_sse2_8bit+0x1a> + DB 232,0,0,0,0 ; call a45e <_sk_srcover_rgba_8888_sse2_8bit+0x1a> DB 91 ; pop %ebx DB 139,77,20 ; mov 0x14(%ebp),%ecx DB 139,69,12 ; mov 0xc(%ebp),%eax @@ -57647,9 +57572,9 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 15,133,51,1,0,0 ; jne a5c7 <_sk_srcover_rgba_8888_sse2_8bit+0x16f> - DB 15,16,116,183,16 ; movups 0x10(%edi,%esi,4),%xmm6 + DB 15,133,51,1,0,0 ; jne a5b3 <_sk_srcover_rgba_8888_sse2_8bit+0x16f> DB 102,15,16,20,183 ; movupd (%edi,%esi,4),%xmm2 + DB 15,16,116,183,16 ; movups 0x10(%edi,%esi,4),%xmm6 DB 102,15,41,85,200 ; movapd %xmm2,-0x38(%ebp) DB 15,41,117,216 ; movaps %xmm6,-0x28(%ebp) DB 242,15,112,217,231 ; pshuflw $0xe7,%xmm1,%xmm3 @@ -57692,13 +57617,13 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 DB 102,15,213,216 ; pmullw %xmm0,%xmm3 DB 102,15,253,216 ; paddw %xmm0,%xmm3 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,243 ; packuswb %xmm3,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,252 ; packuswb %xmm4,%xmm7 DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 - DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,243 ; packuswb %xmm3,%xmm6 DB 102,15,111,85,216 ; movdqa -0x28(%ebp),%xmm2 DB 102,15,248,214 ; psubb %xmm6,%xmm2 DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) @@ -57707,7 +57632,7 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 133,210 ; test %edx,%edx DB 102,15,252,232 ; paddb %xmm0,%xmm5 DB 102,15,252,217 ; paddb %xmm1,%xmm3 - DB 117,98 ; jne a5fe <_sk_srcover_rgba_8888_sse2_8bit+0x1a6> + DB 117,98 ; jne a5ea <_sk_srcover_rgba_8888_sse2_8bit+0x1a6> DB 243,15,127,44,183 ; movdqu %xmm5,(%edi,%esi,4) DB 243,15,127,92,183,16 ; movdqu %xmm3,0x10(%edi,%esi,4) DB 141,120,8 ; lea 0x8(%eax),%edi @@ -57732,33 +57657,32 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 254,77,200 ; decb -0x38(%ebp) DB 128,125,200,6 ; cmpb $0x6,-0x38(%ebp) - DB 119,66 ; ja a626 <_sk_srcover_rgba_8888_sse2_8bit+0x1ce> + DB 119,66 ; ja a612 <_sk_srcover_rgba_8888_sse2_8bit+0x1ce> DB 15,182,93,200 ; movzbl -0x38(%ebp),%ebx DB 139,69,216 ; mov -0x28(%ebp),%eax - DB 139,156,152,70,2,0,0 ; mov 0x246(%eax,%ebx,4),%ebx + DB 139,156,152,66,2,0,0 ; mov 0x242(%eax,%ebx,4),%ebx DB 3,93,216 ; add -0x28(%ebp),%ebx DB 255,227 ; jmp *%ebx DB 243,15,16,20,183 ; movss (%edi,%esi,4),%xmm2 - DB 235,109 ; jmp a66b <_sk_srcover_rgba_8888_sse2_8bit+0x213> + DB 235,109 ; jmp a657 <_sk_srcover_rgba_8888_sse2_8bit+0x213> DB 137,93,216 ; mov %ebx,-0x28(%ebp) DB 137,211 ; mov %edx,%ebx DB 128,227,7 ; and $0x7,%bl DB 254,203 ; dec %bl DB 128,251,6 ; cmp $0x6,%bl - DB 119,154 ; ja a5a7 <_sk_srcover_rgba_8888_sse2_8bit+0x14f> + DB 119,154 ; ja a593 <_sk_srcover_rgba_8888_sse2_8bit+0x14f> DB 15,182,219 ; movzbl %bl,%ebx DB 139,77,216 ; mov -0x28(%ebp),%ecx - DB 3,140,153,98,2,0,0 ; add 0x262(%ecx,%ebx,4),%ecx + DB 3,140,153,94,2,0,0 ; add 0x25e(%ecx,%ebx,4),%ecx DB 255,225 ; jmp *%ecx DB 102,15,126,44,183 ; movd %xmm5,(%edi,%esi,4) - DB 139,77,20 ; mov 0x14(%ebp),%ecx - DB 235,129 ; jmp a5a7 <_sk_srcover_rgba_8888_sse2_8bit+0x14f> + DB 233,131,0,0,0 ; jmp a695 <_sk_srcover_rgba_8888_sse2_8bit+0x251> DB 139,93,216 ; mov -0x28(%ebp),%ebx - DB 233,112,254,255,255 ; jmp a49e <_sk_srcover_rgba_8888_sse2_8bit+0x46> + DB 233,112,254,255,255 ; jmp a48a <_sk_srcover_rgba_8888_sse2_8bit+0x46> DB 102,15,110,92,183,8 ; movd 0x8(%edi,%esi,4),%xmm3 DB 102,15,112,211,69 ; pshufd $0x45,%xmm3,%xmm2 DB 102,15,18,20,183 ; movlpd (%edi,%esi,4),%xmm2 - DB 235,43 ; jmp a66b <_sk_srcover_rgba_8888_sse2_8bit+0x213> + DB 235,43 ; jmp a657 <_sk_srcover_rgba_8888_sse2_8bit+0x213> DB 102,15,110,84,183,24 ; movd 0x18(%edi,%esi,4),%xmm2 DB 102,15,112,242,69 ; pshufd $0x45,%xmm2,%xmm6 DB 243,15,16,92,183,20 ; movss 0x14(%edi,%esi,4),%xmm3 @@ -57770,12 +57694,11 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 102,15,16,20,183 ; movupd (%edi,%esi,4),%xmm2 DB 139,93,216 ; mov -0x28(%ebp),%ebx DB 139,69,12 ; mov 0xc(%ebp),%eax - DB 233,40,254,255,255 ; jmp a49e <_sk_srcover_rgba_8888_sse2_8bit+0x46> + DB 233,40,254,255,255 ; jmp a48a <_sk_srcover_rgba_8888_sse2_8bit+0x46> DB 102,15,112,213,78 ; pshufd $0x4e,%xmm5,%xmm2 DB 102,15,126,84,183,8 ; movd %xmm2,0x8(%edi,%esi,4) DB 102,15,214,44,183 ; movq %xmm5,(%edi,%esi,4) - DB 139,77,20 ; mov 0x14(%ebp),%ecx - DB 233,25,255,255,255 ; jmp a5a7 <_sk_srcover_rgba_8888_sse2_8bit+0x14f> + DB 235,33 ; jmp a695 <_sk_srcover_rgba_8888_sse2_8bit+0x251> DB 102,15,112,211,78 ; pshufd $0x4e,%xmm3,%xmm2 DB 102,15,126,84,183,24 ; movd %xmm2,0x18(%edi,%esi,4) DB 102,15,112,211,229 ; pshufd $0xe5,%xmm3,%xmm2 @@ -57783,8 +57706,8 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 102,15,126,92,183,16 ; movd %xmm3,0x10(%edi,%esi,4) DB 243,15,127,44,183 ; movdqu %xmm5,(%edi,%esi,4) DB 139,77,20 ; mov 0x14(%ebp),%ecx - DB 233,240,254,255,255 ; jmp a5a7 <_sk_srcover_rgba_8888_sse2_8bit+0x14f> - DB 144 ; nop + DB 233,246,254,255,255 ; jmp a593 <_sk_srcover_rgba_8888_sse2_8bit+0x14f> + DB 15,31,0 ; nopl (%eax) DB 133,1 ; test %eax,(%ecx) DB 0,0 ; add %al,(%eax) DB 199,1,0,0,188,1 ; movl $0x1bc0000,(%ecx) @@ -57801,14 +57724,15 @@ _sk_srcover_rgba_8888_sse2_8bit LABEL PROC DB 2,0 ; add (%eax),%al DB 0,4,2 ; add %al,(%edx,%eax,1) DB 0,0 ; add %al,(%eax) - DB 56,2 ; cmp %al,(%edx) - DB 0,0 ; add %al,(%eax) DB 50,2 ; xor (%edx),%al DB 0,0 ; add %al,(%eax) - DB 39 ; daa - DB 2,0 ; add (%eax),%al - DB 0,28,2 ; add %bl,(%edx,%eax,1) + DB 44,2 ; sub $0x2,%al DB 0,0 ; add %al,(%eax) + DB 33,2 ; and %eax,(%edx) + DB 0,0 ; add %al,(%eax) + DB 22 ; push %ss + DB 2,0 ; add (%eax),%al + DB 0 ; .byte 0x0 PUBLIC _sk_scale_1_float_sse2_8bit _sk_scale_1_float_sse2_8bit LABEL PROC @@ -57819,12 +57743,12 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 15,41,85,232 ; movaps %xmm2,-0x18(%ebp) DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 232,0,0,0,0 ; call a70b <_sk_scale_1_float_sse2_8bit+0x1b> + DB 232,0,0,0,0 ; call a6f3 <_sk_scale_1_float_sse2_8bit+0x1b> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx DB 243,15,16,2 ; movss (%edx),%xmm0 - DB 243,15,89,129,153,48,0,0 ; mulss 0x3099(%ecx),%xmm0 + DB 243,15,89,129,209,47,0,0 ; mulss 0x2fd1(%ecx),%xmm0 DB 243,15,44,208 ; cvttss2si %xmm0,%edx DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 102,15,111,213 ; movdqa %xmm5,%xmm2 @@ -57836,7 +57760,7 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,15,96,246 ; punpcklbw %xmm6,%xmm6 DB 242,15,112,246,0 ; pshuflw $0x0,%xmm6,%xmm6 DB 102,15,112,246,80 ; pshufd $0x50,%xmm6,%xmm6 - DB 102,15,219,177,69,48,0,0 ; pand 0x3045(%ecx),%xmm6 + DB 102,15,219,177,125,47,0,0 ; pand 0x2f7d(%ecx),%xmm6 DB 102,15,111,254 ; movdqa %xmm6,%xmm7 DB 102,15,111,222 ; movdqa %xmm6,%xmm3 DB 102,15,111,198 ; movdqa %xmm6,%xmm0 @@ -57848,12 +57772,12 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,15,253,203 ; paddw %xmm3,%xmm1 DB 102,15,213,252 ; pmullw %xmm4,%xmm7 DB 102,15,253,252 ; paddw %xmm4,%xmm7 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,103,208 ; packuswb %xmm0,%xmm2 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,207 ; packuswb %xmm7,%xmm1 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,103,208 ; packuswb %xmm0,%xmm2 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 15,40,85,232 ; movaps -0x18(%ebp),%xmm2 @@ -57877,7 +57801,7 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 131,236,60 ; sub $0x3c,%esp DB 15,41,93,184 ; movaps %xmm3,-0x48(%ebp) DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) - DB 232,0,0,0,0 ; call a7d3 <_sk_scale_u8_sse2_8bit+0x17> + DB 232,0,0,0,0 ; call a7bb <_sk_scale_u8_sse2_8bit+0x17> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -57887,10 +57811,10 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 15,133,11,1,0,0 ; jne a8fb <_sk_scale_u8_sse2_8bit+0x13f> + DB 15,133,11,1,0,0 ; jne a8e3 <_sk_scale_u8_sse2_8bit+0x13f> DB 243,15,126,36,51 ; movq (%ebx,%esi,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 - DB 102,15,219,167,125,47,0,0 ; pand 0x2f7d(%edi),%xmm4 + DB 102,15,219,167,181,46,0,0 ; pand 0x2eb5(%edi),%xmm4 DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,111,236 ; movdqa %xmm4,%xmm5 DB 102,15,105,234 ; punpckhwd %xmm2,%xmm5 @@ -57934,12 +57858,12 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,253,232 ; paddw %xmm0,%xmm5 DB 102,15,213,241 ; pmullw %xmm1,%xmm6 DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,231 ; packuswb %xmm7,%xmm4 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,238 ; packuswb %xmm6,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,231 ; packuswb %xmm7,%xmm4 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,15,111,205 ; movdqa %xmm5,%xmm1 @@ -57961,14 +57885,14 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 15,135,234,254,255,255 ; ja a7f9 <_sk_scale_u8_sse2_8bit+0x3d> + DB 15,135,234,254,255,255 ; ja a7e1 <_sk_scale_u8_sse2_8bit+0x3d> DB 15,182,201 ; movzbl %cl,%ecx DB 139,140,143,177,1,0,0 ; mov 0x1b1(%edi,%ecx,4),%ecx DB 1,249 ; add %edi,%ecx DB 255,225 ; jmp *%ecx DB 15,182,12,51 ; movzbl (%ebx,%esi,1),%ecx DB 102,15,110,225 ; movd %ecx,%xmm4 - DB 233,207,254,255,255 ; jmp a7f9 <_sk_scale_u8_sse2_8bit+0x3d> + DB 233,207,254,255,255 ; jmp a7e1 <_sk_scale_u8_sse2_8bit+0x3d> DB 15,182,76,51,2 ; movzbl 0x2(%ebx,%esi,1),%ecx DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,196,225,2 ; pinsrw $0x2,%ecx,%xmm4 @@ -57976,7 +57900,7 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 243,15,16,227 ; movss %xmm3,%xmm4 - DB 233,172,254,255,255 ; jmp a7f9 <_sk_scale_u8_sse2_8bit+0x3d> + DB 233,172,254,255,255 ; jmp a7e1 <_sk_scale_u8_sse2_8bit+0x3d> DB 15,182,76,51,6 ; movzbl 0x6(%ebx,%esi,1),%ecx DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,196,225,6 ; pinsrw $0x6,%ecx,%xmm4 @@ -57987,7 +57911,7 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 102,15,110,28,51 ; movd (%ebx,%esi,1),%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 242,15,16,227 ; movsd %xmm3,%xmm4 - DB 233,120,254,255,255 ; jmp a7f9 <_sk_scale_u8_sse2_8bit+0x3d> + DB 233,120,254,255,255 ; jmp a7e1 <_sk_scale_u8_sse2_8bit+0x3d> DB 15,31,0 ; nopl (%eax) DB 74 ; dec %edx DB 1,0 ; add %eax,(%eax) @@ -57999,92 +57923,91 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 0,0 ; add %al,(%eax) DB 136,1 ; mov %al,(%ecx) DB 0,0 ; add %al,(%eax) - DB 122,1 ; jp a99f <_sk_scale_u8_sse2_8bit+0x1e3> + DB 122,1 ; jp a987 <_sk_scale_u8_sse2_8bit+0x1e3> DB 0,0 ; add %al,(%eax) PUBLIC _sk_lerp_1_float_sse2_8bit _sk_lerp_1_float_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp - DB 131,236,72 ; sub $0x48,%esp + DB 131,236,56 ; sub $0x38,%esp DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) DB 15,41,85,232 ; movaps %xmm2,-0x18(%ebp) - DB 102,15,111,241 ; movdqa %xmm1,%xmm6 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 232,0,0,0,0 ; call a9bb <_sk_lerp_1_float_sse2_8bit+0x1b> + DB 232,0,0,0,0 ; call a99b <_sk_lerp_1_float_sse2_8bit+0x13> DB 89 ; pop %ecx DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx - DB 243,15,16,34 ; movss (%edx),%xmm4 - DB 243,15,89,161,233,45,0,0 ; mulss 0x2de9(%ecx),%xmm4 - DB 243,15,44,212 ; cvttss2si %xmm4,%edx - DB 102,15,110,226 ; movd %edx,%xmm4 - DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 - DB 242,15,112,228,0 ; pshuflw $0x0,%xmm4,%xmm4 - DB 102,15,112,212,80 ; pshufd $0x50,%xmm4,%xmm2 - DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,96,197 ; punpcklbw %xmm5,%xmm0 - DB 102,15,111,153,149,45,0,0 ; movdqa 0x2d95(%ecx),%xmm3 - DB 102,15,219,218 ; pand %xmm2,%xmm3 + DB 243,15,16,18 ; movss (%edx),%xmm2 + DB 243,15,89,145,41,45,0,0 ; mulss 0x2d29(%ecx),%xmm2 + DB 243,15,44,210 ; cvttss2si %xmm2,%edx + DB 102,15,110,210 ; movd %edx,%xmm2 + DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 + DB 242,15,112,210,0 ; pshuflw $0x0,%xmm2,%xmm2 + DB 102,15,112,242,80 ; pshufd $0x50,%xmm2,%xmm6 + DB 102,15,239,210 ; pxor %xmm2,%xmm2 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,96,250 ; punpcklbw %xmm2,%xmm7 + DB 102,15,111,153,213,44,0,0 ; movdqa 0x2cd5(%ecx),%xmm3 + DB 102,15,219,222 ; pand %xmm6,%xmm3 DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,111,211 ; movdqa %xmm3,%xmm2 - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,15,213,216 ; pmullw %xmm0,%xmm3 - DB 102,15,253,216 ; paddw %xmm0,%xmm3 - DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 - DB 102,15,213,249 ; pmullw %xmm1,%xmm7 - DB 102,15,253,249 ; paddw %xmm1,%xmm7 - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,15,96,197 ; punpcklbw %xmm5,%xmm0 + DB 102,15,213,223 ; pmullw %xmm7,%xmm3 + DB 102,15,253,223 ; paddw %xmm7,%xmm3 + DB 102,15,239,255 ; pxor %xmm7,%xmm7 + DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 DB 102,15,213,208 ; pmullw %xmm0,%xmm2 DB 102,15,253,208 ; paddw %xmm0,%xmm2 - DB 102,15,104,245 ; punpckhbw %xmm5,%xmm6 - DB 102,15,213,230 ; pmullw %xmm6,%xmm4 - DB 102,15,253,230 ; paddw %xmm6,%xmm4 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,253,232 ; paddw %xmm0,%xmm5 + DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 + DB 102,15,213,225 ; pmullw %xmm1,%xmm4 + DB 102,15,253,225 ; paddw %xmm1,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,236 ; packuswb %xmm4,%xmm5 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,103,223 ; packuswb %xmm7,%xmm3 - DB 102,15,103,212 ; packuswb %xmm4,%xmm2 - DB 102,15,127,85,184 ; movdqa %xmm2,-0x48(%ebp) - DB 102,15,118,201 ; pcmpeqd %xmm1,%xmm1 - DB 102,15,239,77,200 ; pxor -0x38(%ebp),%xmm1 - DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,239,246 ; pxor %xmm6,%xmm6 - DB 102,15,96,214 ; punpcklbw %xmm6,%xmm2 - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,104,230 ; punpckhbw %xmm6,%xmm4 - DB 102,15,111,125,216 ; movdqa -0x28(%ebp),%xmm7 - DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,96,238 ; punpcklbw %xmm6,%xmm5 - DB 102,15,239,246 ; pxor %xmm6,%xmm6 - DB 102,15,104,254 ; punpckhbw %xmm6,%xmm7 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 - DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 + DB 102,15,103,218 ; packuswb %xmm2,%xmm3 + DB 102,15,118,255 ; pcmpeqd %xmm7,%xmm7 + DB 102,15,239,254 ; pxor %xmm6,%xmm7 + DB 102,15,111,117,232 ; movdqa -0x18(%ebp),%xmm6 + DB 102,15,111,214 ; movdqa %xmm6,%xmm2 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,15,96,209 ; punpcklbw %xmm1,%xmm2 + DB 102,15,111,230 ; movdqa %xmm6,%xmm4 + DB 102,15,104,225 ; punpckhbw %xmm1,%xmm4 + DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,96,241 ; punpcklbw %xmm1,%xmm6 + DB 102,15,104,193 ; punpckhbw %xmm1,%xmm0 + DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 + DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,15,104,249 ; punpckhbw %xmm1,%xmm7 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,213,194 ; pmullw %xmm2,%xmm0 DB 102,15,253,194 ; paddw %xmm2,%xmm0 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,213,204 ; pmullw %xmm4,%xmm1 - DB 102,15,253,204 ; paddw %xmm4,%xmm1 - DB 102,15,213,245 ; pmullw %xmm5,%xmm6 - DB 102,15,253,245 ; paddw %xmm5,%xmm6 - DB 102,15,213,215 ; pmullw %xmm7,%xmm2 - DB 102,15,253,215 ; paddw %xmm7,%xmm2 + DB 102,15,111,215 ; movdqa %xmm7,%xmm2 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,15,213,206 ; pmullw %xmm6,%xmm1 + DB 102,15,253,206 ; paddw %xmm6,%xmm1 + DB 102,15,111,101,200 ; movdqa -0x38(%ebp),%xmm4 + DB 102,15,213,212 ; pmullw %xmm4,%xmm2 + DB 102,15,253,212 ; paddw %xmm4,%xmm2 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,103,202 ; packuswb %xmm2,%xmm1 + DB 102,15,252,205 ; paddb %xmm5,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,193 ; packuswb %xmm1,%xmm0 - DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,242 ; packuswb %xmm2,%xmm6 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 102,15,252,195 ; paddb %xmm3,%xmm0 - DB 102,15,252,117,184 ; paddb -0x48(%ebp),%xmm6 DB 141,72,8 ; lea 0x8(%eax),%ecx - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 DB 15,40,85,232 ; movaps -0x18(%ebp),%xmm2 DB 15,40,93,216 ; movaps -0x28(%ebp),%xmm3 DB 255,117,20 ; pushl 0x14(%ebp) @@ -58092,7 +58015,7 @@ _sk_lerp_1_float_sse2_8bit LABEL PROC DB 81 ; push %ecx DB 255,117,8 ; pushl 0x8(%ebp) DB 255,80,4 ; call *0x4(%eax) - DB 131,196,88 ; add $0x58,%esp + DB 131,196,72 ; add $0x48,%esp DB 93 ; pop %ebp DB 195 ; ret @@ -58106,7 +58029,7 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 131,236,92 ; sub $0x5c,%esp DB 15,41,93,200 ; movaps %xmm3,-0x38(%ebp) DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) - DB 232,0,0,0,0 ; call ab21 <_sk_lerp_u8_sse2_8bit+0x17> + DB 232,0,0,0,0 ; call ab03 <_sk_lerp_u8_sse2_8bit+0x17> DB 95 ; pop %edi DB 139,69,12 ; mov 0xc(%ebp),%eax DB 139,16 ; mov (%eax),%edx @@ -58116,15 +58039,15 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 139,85,8 ; mov 0x8(%ebp),%edx DB 133,210 ; test %edx,%edx DB 139,117,16 ; mov 0x10(%ebp),%esi - DB 15,133,181,1,0,0 ; jne acf3 <_sk_lerp_u8_sse2_8bit+0x1e9> - DB 243,15,126,44,51 ; movq (%ebx,%esi,1),%xmm5 - DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 - DB 102,15,219,175,47,44,0,0 ; pand 0x2c2f(%edi),%xmm5 + DB 15,133,181,1,0,0 ; jne acd5 <_sk_lerp_u8_sse2_8bit+0x1e9> + DB 243,15,126,36,51 ; movq (%ebx,%esi,1),%xmm4 + DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 + DB 102,15,219,167,109,43,0,0 ; pand 0x2b6d(%edi),%xmm4 DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 102,15,111,221 ; movdqa %xmm5,%xmm3 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 102,15,105,218 ; punpckhwd %xmm2,%xmm3 - DB 102,15,97,234 ; punpcklwd %xmm2,%xmm5 - DB 102,15,114,245,24 ; pslld $0x18,%xmm5 + DB 102,15,97,226 ; punpcklwd %xmm2,%xmm4 + DB 102,15,114,244,24 ; pslld $0x18,%xmm4 DB 102,15,114,243,24 ; pslld $0x18,%xmm3 DB 242,15,112,219,231 ; pshuflw $0xe7,%xmm3,%xmm3 DB 243,15,112,219,231 ; pshufhw $0xe7,%xmm3,%xmm3 @@ -58132,51 +58055,51 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 102,15,96,219 ; punpcklbw %xmm3,%xmm3 DB 242,15,112,219,95 ; pshuflw $0x5f,%xmm3,%xmm3 DB 243,15,112,219,95 ; pshufhw $0x5f,%xmm3,%xmm3 - DB 242,15,112,229,231 ; pshuflw $0xe7,%xmm5,%xmm4 + DB 242,15,112,228,231 ; pshuflw $0xe7,%xmm4,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,244,95 ; pshufhw $0x5f,%xmm4,%xmm6 - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 - DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 - DB 102,15,111,209 ; movdqa %xmm1,%xmm2 - DB 102,15,96,213 ; punpcklbw %xmm5,%xmm2 - DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 - DB 102,15,111,254 ; movdqa %xmm6,%xmm7 - DB 102,15,96,253 ; punpcklbw %xmm5,%xmm7 - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 - DB 102,15,253,252 ; paddw %xmm4,%xmm7 - DB 102,15,111,238 ; movdqa %xmm6,%xmm5 + DB 243,15,112,252,95 ; pshufhw $0x5f,%xmm4,%xmm7 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,15,104,236 ; punpckhbw %xmm4,%xmm5 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,15,253,232 ; paddw %xmm0,%xmm5 - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,15,96,236 ; punpcklbw %xmm4,%xmm5 + DB 102,15,104,196 ; punpckhbw %xmm4,%xmm0 + DB 102,15,111,209 ; movdqa %xmm1,%xmm2 + DB 102,15,96,212 ; punpcklbw %xmm4,%xmm2 + DB 102,15,104,204 ; punpckhbw %xmm4,%xmm1 + DB 102,15,111,247 ; movdqa %xmm7,%xmm6 + DB 102,15,96,244 ; punpcklbw %xmm4,%xmm6 + DB 102,15,213,245 ; pmullw %xmm5,%xmm6 + DB 102,15,253,245 ; paddw %xmm5,%xmm6 + DB 102,15,111,231 ; movdqa %xmm7,%xmm4 + DB 102,15,239,237 ; pxor %xmm5,%xmm5 + DB 102,15,104,229 ; punpckhbw %xmm5,%xmm4 + DB 102,15,213,224 ; pmullw %xmm0,%xmm4 + DB 102,15,253,224 ; paddw %xmm0,%xmm4 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 - DB 102,15,213,226 ; pmullw %xmm2,%xmm4 - DB 102,15,253,226 ; paddw %xmm2,%xmm4 + DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 + DB 102,15,213,234 ; pmullw %xmm2,%xmm5 + DB 102,15,253,234 ; paddw %xmm2,%xmm5 DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,104,194 ; punpckhbw %xmm2,%xmm0 DB 102,15,213,193 ; pmullw %xmm1,%xmm0 DB 102,15,253,193 ; paddw %xmm1,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,253 ; packuswb %xmm5,%xmm7 - DB 102,15,103,224 ; packuswb %xmm0,%xmm4 + DB 102,15,103,232 ; packuswb %xmm0,%xmm5 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,244 ; packuswb %xmm4,%xmm6 DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 DB 102,15,239,216 ; pxor %xmm0,%xmm3 - DB 102,15,239,240 ; pxor %xmm0,%xmm6 + DB 102,15,239,248 ; pxor %xmm0,%xmm7 DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 102,15,96,234 ; punpcklbw %xmm2,%xmm5 + DB 102,15,96,226 ; punpcklbw %xmm2,%xmm4 DB 102,15,104,194 ; punpckhbw %xmm2,%xmm0 DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 @@ -58185,32 +58108,32 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 102,15,127,77,168 ; movdqa %xmm1,-0x58(%ebp) DB 102,15,104,194 ; punpckhbw %xmm2,%xmm0 DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 DB 102,15,96,194 ; punpcklbw %xmm2,%xmm0 - DB 102,15,104,242 ; punpckhbw %xmm2,%xmm6 + DB 102,15,104,250 ; punpckhbw %xmm2,%xmm7 DB 102,15,111,203 ; movdqa %xmm3,%xmm1 DB 102,15,96,202 ; punpcklbw %xmm2,%xmm1 DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,104,218 ; punpckhbw %xmm2,%xmm3 - DB 102,15,213,197 ; pmullw %xmm5,%xmm0 - DB 102,15,253,197 ; paddw %xmm5,%xmm0 + DB 102,15,213,196 ; pmullw %xmm4,%xmm0 + DB 102,15,253,196 ; paddw %xmm4,%xmm0 DB 102,15,111,85,152 ; movdqa -0x68(%ebp),%xmm2 - DB 102,15,213,242 ; pmullw %xmm2,%xmm6 - DB 102,15,253,242 ; paddw %xmm2,%xmm6 + DB 102,15,213,250 ; pmullw %xmm2,%xmm7 + DB 102,15,253,250 ; paddw %xmm2,%xmm7 DB 102,15,111,85,168 ; movdqa -0x58(%ebp),%xmm2 DB 102,15,213,202 ; pmullw %xmm2,%xmm1 DB 102,15,253,202 ; paddw %xmm2,%xmm1 DB 102,15,111,85,184 ; movdqa -0x48(%ebp),%xmm2 DB 102,15,213,218 ; pmullw %xmm2,%xmm3 DB 102,15,253,218 ; paddw %xmm2,%xmm3 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,198 ; packuswb %xmm6,%xmm0 DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,203 ; packuswb %xmm3,%xmm1 - DB 102,15,252,199 ; paddb %xmm7,%xmm0 - DB 102,15,252,204 ; paddb %xmm4,%xmm1 + DB 102,15,252,205 ; paddb %xmm5,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 + DB 102,15,252,198 ; paddb %xmm6,%xmm0 DB 141,72,8 ; lea 0x8(%eax),%ecx DB 15,40,85,216 ; movaps -0x28(%ebp),%xmm2 DB 15,40,93,200 ; movaps -0x38(%ebp),%xmm3 @@ -58227,37 +58150,37 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 195 ; ret DB 137,209 ; mov %edx,%ecx DB 128,225,7 ; and $0x7,%cl - DB 102,15,239,237 ; pxor %xmm5,%xmm5 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 254,201 ; dec %cl DB 128,249,6 ; cmp $0x6,%cl - DB 15,135,64,254,255,255 ; ja ab47 <_sk_lerp_u8_sse2_8bit+0x3d> + DB 15,135,64,254,255,255 ; ja ab29 <_sk_lerp_u8_sse2_8bit+0x3d> DB 15,182,201 ; movzbl %cl,%ecx - DB 139,140,143,91,2,0,0 ; mov 0x25b(%edi,%ecx,4),%ecx + DB 139,140,143,89,2,0,0 ; mov 0x259(%edi,%ecx,4),%ecx DB 1,249 ; add %edi,%ecx DB 255,225 ; jmp *%ecx DB 15,182,12,51 ; movzbl (%ebx,%esi,1),%ecx - DB 102,15,110,233 ; movd %ecx,%xmm5 - DB 233,37,254,255,255 ; jmp ab47 <_sk_lerp_u8_sse2_8bit+0x3d> + DB 102,15,110,225 ; movd %ecx,%xmm4 + DB 233,37,254,255,255 ; jmp ab29 <_sk_lerp_u8_sse2_8bit+0x3d> DB 15,182,76,51,2 ; movzbl 0x2(%ebx,%esi,1),%ecx - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,196,233,2 ; pinsrw $0x2,%ecx,%xmm5 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,196,225,2 ; pinsrw $0x2,%ecx,%xmm4 DB 15,183,12,51 ; movzwl (%ebx,%esi,1),%ecx DB 102,15,110,217 ; movd %ecx,%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 - DB 243,15,16,235 ; movss %xmm3,%xmm5 - DB 233,2,254,255,255 ; jmp ab47 <_sk_lerp_u8_sse2_8bit+0x3d> + DB 243,15,16,227 ; movss %xmm3,%xmm4 + DB 233,2,254,255,255 ; jmp ab29 <_sk_lerp_u8_sse2_8bit+0x3d> DB 15,182,76,51,6 ; movzbl 0x6(%ebx,%esi,1),%ecx - DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,196,233,6 ; pinsrw $0x6,%ecx,%xmm5 + DB 102,15,239,228 ; pxor %xmm4,%xmm4 + DB 102,15,196,225,6 ; pinsrw $0x6,%ecx,%xmm4 DB 15,182,76,51,5 ; movzbl 0x5(%ebx,%esi,1),%ecx - DB 102,15,196,233,5 ; pinsrw $0x5,%ecx,%xmm5 + DB 102,15,196,225,5 ; pinsrw $0x5,%ecx,%xmm4 DB 15,182,76,51,4 ; movzbl 0x4(%ebx,%esi,1),%ecx - DB 102,15,196,233,4 ; pinsrw $0x4,%ecx,%xmm5 + DB 102,15,196,225,4 ; pinsrw $0x4,%ecx,%xmm4 DB 102,15,110,28,51 ; movd (%ebx,%esi,1),%xmm3 DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 - DB 242,15,16,235 ; movsd %xmm3,%xmm5 - DB 233,206,253,255,255 ; jmp ab47 <_sk_lerp_u8_sse2_8bit+0x3d> - DB 15,31,0 ; nopl (%eax) + DB 242,15,16,227 ; movsd %xmm3,%xmm4 + DB 233,206,253,255,255 ; jmp ab29 <_sk_lerp_u8_sse2_8bit+0x3d> + DB 144 ; nop DB 244 ; hlt DB 1,0 ; add %eax,(%eax) DB 0,15 ; add %cl,(%edi) @@ -58314,11 +58237,11 @@ _sk_black_color_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,8 ; sub $0x8,%esp - DB 232,0,0,0,0 ; call ade9 <_sk_black_color_sse2_8bit+0xb> + DB 232,0,0,0,0 ; call adc9 <_sk_black_color_sse2_8bit+0xb> DB 88 ; pop %eax DB 139,77,12 ; mov 0xc(%ebp),%ecx DB 141,81,4 ; lea 0x4(%ecx),%edx - DB 15,40,128,7,41,0,0 ; movaps 0x2907(%eax),%xmm0 + DB 15,40,128,71,40,0,0 ; movaps 0x2847(%eax),%xmm0 DB 15,40,200 ; movaps %xmm0,%xmm1 DB 255,117,20 ; pushl 0x14(%ebp) DB 255,117,16 ; pushl 0x10(%ebp) @@ -58370,8 +58293,8 @@ _sk_srcatop_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,88 ; sub $0x58,%esp - DB 102,15,111,251 ; movdqa %xmm3,%xmm7 - DB 102,15,127,125,232 ; movdqa %xmm7,-0x18(%ebp) + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,15,127,117,232 ; movdqa %xmm6,-0x18(%ebp) DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) DB 102,15,111,233 ; movdqa %xmm1,%xmm5 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 @@ -58380,45 +58303,45 @@ _sk_srcatop_sse2_8bit LABEL PROC DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 DB 242,15,112,210,95 ; pshuflw $0x5f,%xmm2,%xmm2 - DB 243,15,112,210,95 ; pshufhw $0x5f,%xmm2,%xmm2 + DB 243,15,112,250,95 ; pshufhw $0x5f,%xmm2,%xmm7 DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,96,243 ; punpcklbw %xmm3,%xmm6 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,111,212 ; movdqa %xmm4,%xmm2 + DB 102,15,96,211 ; punpcklbw %xmm3,%xmm2 + DB 102,15,111,199 ; movdqa %xmm7,%xmm0 DB 102,15,96,195 ; punpcklbw %xmm3,%xmm0 - DB 102,15,213,198 ; pmullw %xmm6,%xmm0 - DB 102,15,253,198 ; paddw %xmm6,%xmm0 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,104,243 ; punpckhbw %xmm3,%xmm6 + DB 102,15,213,194 ; pmullw %xmm2,%xmm0 + DB 102,15,253,194 ; paddw %xmm2,%xmm0 + DB 102,15,111,212 ; movdqa %xmm4,%xmm2 DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 - DB 102,15,213,214 ; pmullw %xmm6,%xmm2 - DB 102,15,253,214 ; paddw %xmm6,%xmm2 - DB 242,15,112,247,231 ; pshuflw $0xe7,%xmm7,%xmm6 - DB 243,15,112,246,231 ; pshufhw $0xe7,%xmm6,%xmm6 - DB 102,15,112,246,232 ; pshufd $0xe8,%xmm6,%xmm6 - DB 102,15,96,246 ; punpcklbw %xmm6,%xmm6 - DB 242,15,112,246,95 ; pshuflw $0x5f,%xmm6,%xmm6 - DB 243,15,112,254,95 ; pshufhw $0x5f,%xmm6,%xmm7 + DB 102,15,104,251 ; punpckhbw %xmm3,%xmm7 + DB 102,15,213,250 ; pmullw %xmm2,%xmm7 + DB 102,15,253,250 ; paddw %xmm2,%xmm7 + DB 242,15,112,214,231 ; pshuflw $0xe7,%xmm6,%xmm2 + DB 243,15,112,210,231 ; pshufhw $0xe7,%xmm2,%xmm2 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 + DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 + DB 242,15,112,210,95 ; pshuflw $0x5f,%xmm2,%xmm2 + DB 243,15,112,210,95 ; pshufhw $0x5f,%xmm2,%xmm2 DB 102,15,96,203 ; punpcklbw %xmm3,%xmm1 - DB 102,15,111,247 ; movdqa %xmm7,%xmm6 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 DB 102,15,96,243 ; punpcklbw %xmm3,%xmm6 DB 102,15,213,241 ; pmullw %xmm1,%xmm6 DB 102,15,253,241 ; paddw %xmm1,%xmm6 DB 102,15,111,205 ; movdqa %xmm5,%xmm1 DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 - DB 102,15,104,251 ; punpckhbw %xmm3,%xmm7 - DB 102,15,213,249 ; pmullw %xmm1,%xmm7 - DB 102,15,253,249 ; paddw %xmm1,%xmm7 + DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 + DB 102,15,213,209 ; pmullw %xmm1,%xmm2 + DB 102,15,253,209 ; paddw %xmm1,%xmm2 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,194 ; packuswb %xmm2,%xmm0 - DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) - DB 242,15,112,196,231 ; pshuflw $0xe7,%xmm4,%xmm0 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,242 ; packuswb %xmm2,%xmm6 + DB 242,15,112,212,231 ; pshuflw $0xe7,%xmm4,%xmm2 DB 242,15,112,205,231 ; pshuflw $0xe7,%xmm5,%xmm1 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,247 ; packuswb %xmm7,%xmm6 - DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 + DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) + DB 243,15,112,194,231 ; pshufhw $0xe7,%xmm2,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 @@ -58460,14 +58383,14 @@ _sk_srcatop_sse2_8bit LABEL PROC DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 DB 102,15,213,211 ; pmullw %xmm3,%xmm2 DB 102,15,253,211 ; paddw %xmm3,%xmm2 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,202 ; packuswb %xmm2,%xmm1 - DB 102,15,252,69,168 ; paddb -0x58(%ebp),%xmm0 DB 102,15,252,206 ; paddb %xmm6,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 + DB 102,15,252,69,168 ; paddb -0x58(%ebp),%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 15,40,85,216 ; movaps -0x28(%ebp),%xmm2 @@ -58526,13 +58449,13 @@ _sk_dstatop_sse2_8bit LABEL PROC DB 102,15,104,202 ; punpckhbw %xmm2,%xmm1 DB 102,15,213,204 ; pmullw %xmm4,%xmm1 DB 102,15,253,204 ; paddw %xmm4,%xmm1 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,241 ; packuswb %xmm1,%xmm6 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,248 ; packuswb %xmm0,%xmm7 DB 102,15,127,125,200 ; movdqa %xmm7,-0x38(%ebp) - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,241 ; packuswb %xmm1,%xmm6 DB 242,15,112,197,231 ; pshuflw $0xe7,%xmm5,%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 @@ -58577,14 +58500,14 @@ _sk_dstatop_sse2_8bit LABEL PROC DB 102,15,111,85,232 ; movdqa -0x18(%ebp),%xmm2 DB 102,15,213,194 ; pmullw %xmm2,%xmm0 DB 102,15,253,194 ; paddw %xmm2,%xmm0 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,225 ; packuswb %xmm1,%xmm4 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,232 ; packuswb %xmm0,%xmm5 - DB 102,15,252,101,200 ; paddb -0x38(%ebp),%xmm4 DB 102,15,252,238 ; paddb %xmm6,%xmm5 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,225 ; packuswb %xmm1,%xmm4 + DB 102,15,252,101,200 ; paddb -0x38(%ebp),%xmm4 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 102,15,111,196 ; movdqa %xmm4,%xmm0 @@ -58643,12 +58566,12 @@ _sk_srcin_sse2_8bit LABEL PROC DB 102,15,111,77,232 ; movdqa -0x18(%ebp),%xmm1 DB 102,15,213,241 ; pmullw %xmm1,%xmm6 DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,238 ; packuswb %xmm6,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 102,15,111,205 ; movdqa %xmm5,%xmm1 @@ -58705,12 +58628,12 @@ _sk_dstin_sse2_8bit LABEL PROC DB 102,15,111,85,232 ; movdqa -0x18(%ebp),%xmm2 DB 102,15,213,226 ; pmullw %xmm2,%xmm4 DB 102,15,253,226 ; paddw %xmm2,%xmm4 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,197 ; packuswb %xmm5,%xmm0 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,204 ; packuswb %xmm4,%xmm1 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,197 ; packuswb %xmm5,%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 15,40,85,200 ; movaps -0x38(%ebp),%xmm2 @@ -58771,12 +58694,12 @@ _sk_srcout_sse2_8bit LABEL PROC DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 DB 102,15,213,248 ; pmullw %xmm0,%xmm7 DB 102,15,253,248 ; paddw %xmm0,%xmm7 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,230 ; packuswb %xmm6,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,239 ; packuswb %xmm7,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,230 ; packuswb %xmm6,%xmm4 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 102,15,111,196 ; movdqa %xmm4,%xmm0 @@ -58837,12 +58760,12 @@ _sk_dstout_sse2_8bit LABEL PROC DB 102,15,111,85,232 ; movdqa -0x18(%ebp),%xmm2 DB 102,15,213,234 ; pmullw %xmm2,%xmm5 DB 102,15,253,234 ; paddw %xmm2,%xmm5 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,196 ; packuswb %xmm4,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,205 ; packuswb %xmm5,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,196 ; packuswb %xmm4,%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 15,40,85,200 ; movaps -0x38(%ebp),%xmm2 @@ -58902,20 +58825,21 @@ _sk_srcover_sse2_8bit LABEL PROC DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 DB 102,15,213,224 ; pmullw %xmm0,%xmm4 DB 102,15,253,224 ; paddw %xmm0,%xmm4 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,252 ; packuswb %xmm4,%xmm7 - DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,245 ; packuswb %xmm5,%xmm6 + DB 102,15,111,93,216 ; movdqa -0x28(%ebp),%xmm3 + DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 + DB 102,15,252,195 ; paddb %xmm3,%xmm0 + DB 102,15,248,199 ; psubb %xmm7,%xmm0 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 + DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 DB 102,15,252,194 ; paddb %xmm2,%xmm0 DB 102,15,248,198 ; psubb %xmm6,%xmm0 - DB 102,15,111,93,216 ; movdqa -0x28(%ebp),%xmm3 - DB 102,15,111,77,168 ; movdqa -0x58(%ebp),%xmm1 - DB 102,15,252,203 ; paddb %xmm3,%xmm1 - DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 255,117,20 ; pushl 0x14(%ebp) @@ -58973,20 +58897,21 @@ _sk_dstover_sse2_8bit LABEL PROC DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 DB 102,15,213,224 ; pmullw %xmm0,%xmm4 DB 102,15,253,224 ; paddw %xmm0,%xmm4 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,245 ; packuswb %xmm5,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,245 ; packuswb %xmm5,%xmm6 + DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 + DB 102,15,111,93,168 ; movdqa -0x58(%ebp),%xmm3 + DB 102,15,252,195 ; paddb %xmm3,%xmm0 + DB 102,15,248,199 ; psubb %xmm7,%xmm0 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,111,85,184 ; movdqa -0x48(%ebp),%xmm2 DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 DB 102,15,252,194 ; paddb %xmm2,%xmm0 DB 102,15,248,198 ; psubb %xmm6,%xmm0 - DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 - DB 102,15,111,93,168 ; movdqa -0x58(%ebp),%xmm3 - DB 102,15,252,203 ; paddb %xmm3,%xmm1 - DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 255,117,20 ; pushl 0x14(%ebp) @@ -59003,45 +58928,46 @@ _sk_modulate_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 131,236,56 ; sub $0x38,%esp - DB 102,15,111,227 ; movdqa %xmm3,%xmm4 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 DB 102,15,239,237 ; pxor %xmm5,%xmm5 - DB 102,15,111,222 ; movdqa %xmm6,%xmm3 - DB 102,15,96,221 ; punpcklbw %xmm5,%xmm3 - DB 102,15,104,245 ; punpckhbw %xmm5,%xmm6 + DB 102,15,111,215 ; movdqa %xmm7,%xmm2 + DB 102,15,96,213 ; punpcklbw %xmm5,%xmm2 + DB 102,15,104,253 ; punpckhbw %xmm5,%xmm7 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,96,205 ; punpcklbw %xmm5,%xmm1 DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) - DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,96,253 ; punpcklbw %xmm5,%xmm7 - DB 102,15,104,213 ; punpckhbw %xmm5,%xmm2 + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,104,245 ; punpckhbw %xmm5,%xmm6 DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,127,109,232 ; movdqa %xmm5,-0x18(%ebp) - DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 + DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 DB 102,15,111,197 ; movdqa %xmm5,%xmm0 DB 102,15,239,237 ; pxor %xmm5,%xmm5 DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 - DB 102,15,213,251 ; pmullw %xmm3,%xmm7 - DB 102,15,253,223 ; paddw %xmm7,%xmm3 - DB 102,15,213,214 ; pmullw %xmm6,%xmm2 - DB 102,15,253,214 ; paddw %xmm6,%xmm2 - DB 102,15,213,225 ; pmullw %xmm1,%xmm4 - DB 102,15,253,204 ; paddw %xmm4,%xmm1 - DB 102,15,111,101,216 ; movdqa -0x28(%ebp),%xmm4 - DB 102,15,213,196 ; pmullw %xmm4,%xmm0 - DB 102,15,253,196 ; paddw %xmm4,%xmm0 - DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,103,218 ; packuswb %xmm2,%xmm3 + DB 102,15,213,226 ; pmullw %xmm2,%xmm4 + DB 102,15,253,212 ; paddw %xmm4,%xmm2 + DB 102,15,213,247 ; pmullw %xmm7,%xmm6 + DB 102,15,253,247 ; paddw %xmm7,%xmm6 + DB 102,15,213,217 ; pmullw %xmm1,%xmm3 + DB 102,15,253,203 ; paddw %xmm3,%xmm1 + DB 102,15,111,93,216 ; movdqa -0x28(%ebp),%xmm3 + DB 102,15,213,195 ; pmullw %xmm3,%xmm0 + DB 102,15,253,195 ; paddw %xmm3,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,200 ; packuswb %xmm0,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,103,214 ; packuswb %xmm6,%xmm2 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 15,40,85,200 ; movaps -0x38(%ebp),%xmm2 DB 15,40,93,232 ; movaps -0x18(%ebp),%xmm3 DB 255,117,20 ; pushl 0x14(%ebp) @@ -59058,64 +58984,66 @@ _sk_multiply_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp DB 129,236,152,0,0,0 ; sub $0x98,%esp - DB 15,41,93,216 ; movaps %xmm3,-0x28(%ebp) - DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) - DB 102,15,111,241 ; movdqa %xmm1,%xmm6 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 242,15,112,226,231 ; pshuflw $0xe7,%xmm2,%xmm4 + DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,15,127,101,232 ; movdqa %xmm4,-0x18(%ebp) + DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 242,15,112,228,231 ; pshuflw $0xe7,%xmm4,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 + DB 243,15,112,196,95 ; pshufhw $0x5f,%xmm4,%xmm0 DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 - DB 102,15,239,226 ; pxor %xmm2,%xmm4 - DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,15,96,233 ; punpcklbw %xmm1,%xmm5 - DB 102,15,111,253 ; movdqa %xmm5,%xmm7 - DB 102,15,111,235 ; movdqa %xmm3,%xmm5 - DB 102,15,104,233 ; punpckhbw %xmm1,%xmm5 - DB 102,15,127,173,104,255,255,255 ; movdqa %xmm5,-0x98(%ebp) - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 - DB 102,15,104,225 ; punpckhbw %xmm1,%xmm4 - DB 102,15,213,229 ; pmullw %xmm5,%xmm4 - DB 102,15,127,189,120,255,255,255 ; movdqa %xmm7,-0x88(%ebp) - DB 102,15,213,199 ; pmullw %xmm7,%xmm0 - DB 102,15,253,199 ; paddw %xmm7,%xmm0 - DB 102,15,253,229 ; paddw %xmm5,%xmm4 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,239,194 ; pxor %xmm2,%xmm0 + DB 102,15,239,237 ; pxor %xmm5,%xmm5 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,15,96,245 ; punpcklbw %xmm5,%xmm6 + DB 102,15,111,254 ; movdqa %xmm6,%xmm7 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,15,104,245 ; punpckhbw %xmm5,%xmm6 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,96,213 ; punpcklbw %xmm5,%xmm2 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,213,198 ; pmullw %xmm6,%xmm0 + DB 102,15,127,181,120,255,255,255 ; movdqa %xmm6,-0x88(%ebp) + DB 102,15,127,189,104,255,255,255 ; movdqa %xmm7,-0x98(%ebp) + DB 102,15,213,215 ; pmullw %xmm7,%xmm2 + DB 102,15,253,215 ; paddw %xmm7,%xmm2 + DB 102,15,253,198 ; paddw %xmm6,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,196 ; packuswb %xmm4,%xmm0 - DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) - DB 102,15,111,125,216 ; movdqa -0x28(%ebp),%xmm7 - DB 242,15,112,231,231 ; pshuflw $0xe7,%xmm7,%xmm4 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,103,208 ; packuswb %xmm0,%xmm2 + DB 102,15,127,85,168 ; movdqa %xmm2,-0x58(%ebp) + DB 102,15,111,85,216 ; movdqa -0x28(%ebp),%xmm2 + DB 242,15,112,226,231 ; pshuflw $0xe7,%xmm2,%xmm4 DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 - DB 102,15,239,226 ; pxor %xmm2,%xmm4 - DB 102,15,111,214 ; movdqa %xmm6,%xmm2 - DB 102,15,239,201 ; pxor %xmm1,%xmm1 - DB 102,15,96,209 ; punpcklbw %xmm1,%xmm2 - DB 102,15,127,85,152 ; movdqa %xmm2,-0x68(%ebp) - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,15,104,193 ; punpckhbw %xmm1,%xmm0 + DB 102,15,118,192 ; pcmpeqd %xmm0,%xmm0 + DB 102,15,239,224 ; pxor %xmm0,%xmm4 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,15,96,245 ; punpcklbw %xmm5,%xmm6 + DB 102,15,127,117,152 ; movdqa %xmm6,-0x68(%ebp) + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 DB 102,15,127,69,136 ; movdqa %xmm0,-0x78(%ebp) - DB 102,15,111,236 ; movdqa %xmm4,%xmm5 - DB 102,15,96,233 ; punpcklbw %xmm1,%xmm5 - DB 102,15,104,225 ; punpckhbw %xmm1,%xmm4 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 + DB 102,15,96,253 ; punpcklbw %xmm5,%xmm7 + DB 102,15,104,229 ; punpckhbw %xmm5,%xmm4 DB 102,15,213,224 ; pmullw %xmm0,%xmm4 - DB 102,15,213,234 ; pmullw %xmm2,%xmm5 - DB 102,15,253,234 ; paddw %xmm2,%xmm5 + DB 102,15,213,254 ; pmullw %xmm6,%xmm7 + DB 102,15,253,254 ; paddw %xmm6,%xmm7 DB 102,15,253,224 ; paddw %xmm0,%xmm4 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,236 ; packuswb %xmm4,%xmm5 - DB 242,15,112,195,231 ; pshuflw $0xe7,%xmm3,%xmm0 - DB 242,15,112,206,231 ; pshuflw $0xe7,%xmm6,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,252 ; packuswb %xmm4,%xmm7 + DB 102,15,127,125,200 ; movdqa %xmm7,-0x38(%ebp) + DB 242,15,112,193,231 ; pshuflw $0xe7,%xmm1,%xmm0 + DB 242,15,112,203,231 ; pshuflw $0xe7,%xmm3,%xmm1 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 @@ -59126,22 +59054,18 @@ _sk_multiply_sse2_8bit LABEL PROC DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 DB 243,15,112,201,95 ; pshufhw $0x5f,%xmm1,%xmm1 - DB 102,15,118,210 ; pcmpeqd %xmm2,%xmm2 - DB 102,15,239,202 ; pxor %xmm2,%xmm1 - DB 102,15,239,194 ; pxor %xmm2,%xmm0 - DB 102,15,111,93,200 ; movdqa -0x38(%ebp),%xmm3 - DB 102,15,111,243 ; movdqa %xmm3,%xmm6 - DB 102,15,239,210 ; pxor %xmm2,%xmm2 - DB 102,15,96,242 ; punpcklbw %xmm2,%xmm6 - DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,118,219 ; pcmpeqd %xmm3,%xmm3 + DB 102,15,239,203 ; pxor %xmm3,%xmm1 + DB 102,15,239,195 ; pxor %xmm3,%xmm0 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,96,245 ; punpcklbw %xmm5,%xmm6 DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 - DB 102,15,127,85,232 ; movdqa %xmm2,-0x18(%ebp) - DB 102,15,111,215 ; movdqa %xmm7,%xmm2 - DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,15,96,227 ; punpcklbw %xmm3,%xmm4 - DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 DB 102,15,127,85,184 ; movdqa %xmm2,-0x48(%ebp) + DB 102,15,111,109,232 ; movdqa -0x18(%ebp),%xmm5 + DB 102,15,111,229 ; movdqa %xmm5,%xmm4 + DB 102,15,96,227 ; punpcklbw %xmm3,%xmm4 + DB 102,15,104,235 ; punpckhbw %xmm3,%xmm5 DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,15,111,216 ; movdqa %xmm0,%xmm3 DB 102,15,96,218 ; punpcklbw %xmm2,%xmm3 @@ -59151,51 +59075,51 @@ _sk_multiply_sse2_8bit LABEL PROC DB 102,15,96,215 ; punpcklbw %xmm7,%xmm2 DB 102,15,239,255 ; pxor %xmm7,%xmm7 DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 - DB 102,15,213,69,232 ; pmullw -0x18(%ebp),%xmm0 - DB 102,15,213,222 ; pmullw %xmm6,%xmm3 - DB 102,15,253,222 ; paddw %xmm6,%xmm3 - DB 102,15,253,69,232 ; paddw -0x18(%ebp),%xmm0 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,103,216 ; packuswb %xmm0,%xmm3 - DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 - DB 102,15,213,200 ; pmullw %xmm0,%xmm1 + DB 102,15,213,205 ; pmullw %xmm5,%xmm1 DB 102,15,213,212 ; pmullw %xmm4,%xmm2 DB 102,15,253,212 ; paddw %xmm4,%xmm2 - DB 102,15,253,200 ; paddw %xmm0,%xmm1 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,253,205 ; paddw %xmm5,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 DB 102,15,103,209 ; packuswb %xmm1,%xmm2 - DB 102,15,252,213 ; paddb %xmm5,%xmm2 - DB 102,15,252,93,168 ; paddb -0x58(%ebp),%xmm3 - DB 102,15,111,133,120,255,255,255 ; movdqa -0x88(%ebp),%xmm0 + DB 102,15,252,85,168 ; paddb -0x58(%ebp),%xmm2 + DB 102,15,111,77,184 ; movdqa -0x48(%ebp),%xmm1 + DB 102,15,213,193 ; pmullw %xmm1,%xmm0 + DB 102,15,213,222 ; pmullw %xmm6,%xmm3 + DB 102,15,253,222 ; paddw %xmm6,%xmm3 + DB 102,15,253,193 ; paddw %xmm1,%xmm0 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 + DB 102,15,103,216 ; packuswb %xmm0,%xmm3 + DB 102,15,252,93,200 ; paddb -0x38(%ebp),%xmm3 + DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 DB 102,15,213,240 ; pmullw %xmm0,%xmm6 DB 102,15,253,240 ; paddw %xmm0,%xmm6 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 + DB 102,15,213,193 ; pmullw %xmm1,%xmm0 + DB 102,15,253,193 ; paddw %xmm1,%xmm0 + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,111,133,104,255,255,255 ; movdqa -0x98(%ebp),%xmm0 - DB 102,15,111,77,232 ; movdqa -0x18(%ebp),%xmm1 - DB 102,15,213,200 ; pmullw %xmm0,%xmm1 - DB 102,15,253,200 ; paddw %xmm0,%xmm1 - DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 DB 102,15,213,224 ; pmullw %xmm0,%xmm4 DB 102,15,253,224 ; paddw %xmm0,%xmm4 - DB 102,15,111,69,136 ; movdqa -0x78(%ebp),%xmm0 - DB 102,15,213,248 ; pmullw %xmm0,%xmm7 - DB 102,15,253,248 ; paddw %xmm0,%xmm7 + DB 102,15,111,133,120,255,255,255 ; movdqa -0x88(%ebp),%xmm0 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,253,232 ; paddw %xmm0,%xmm5 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,229 ; packuswb %xmm5,%xmm4 + DB 102,15,252,212 ; paddb %xmm4,%xmm2 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 DB 102,15,103,241 ; packuswb %xmm1,%xmm6 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,231 ; packuswb %xmm7,%xmm4 DB 102,15,252,222 ; paddb %xmm6,%xmm3 - DB 102,15,252,212 ; paddb %xmm4,%xmm2 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 15,40,85,200 ; movaps -0x38(%ebp),%xmm2 - DB 15,40,93,216 ; movaps -0x28(%ebp),%xmm3 + DB 15,40,85,216 ; movaps -0x28(%ebp),%xmm2 + DB 15,40,93,232 ; movaps -0x18(%ebp),%xmm3 DB 255,117,20 ; pushl 0x14(%ebp) DB 255,117,16 ; pushl 0x10(%ebp) DB 81 ; push %ecx @@ -59216,41 +59140,39 @@ _sk_screen_sse2_8bit LABEL PROC DB 102,15,118,255 ; pcmpeqd %xmm7,%xmm7 DB 102,15,239,207 ; pxor %xmm7,%xmm1 DB 102,15,239,248 ; pxor %xmm0,%xmm7 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 + DB 102,15,239,192 ; pxor %xmm0,%xmm0 DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,96,235 ; punpcklbw %xmm3,%xmm5 - DB 102,15,104,251 ; punpckhbw %xmm3,%xmm7 + DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 + DB 102,15,104,248 ; punpckhbw %xmm0,%xmm7 DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,96,227 ; punpcklbw %xmm3,%xmm4 - DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 + DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 + DB 102,15,104,200 ; punpckhbw %xmm0,%xmm1 DB 102,15,127,77,216 ; movdqa %xmm1,-0x28(%ebp) - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) - DB 102,15,96,211 ; punpcklbw %xmm3,%xmm2 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,15,104,216 ; punpckhbw %xmm0,%xmm3 + DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) + DB 102,15,111,218 ; movdqa %xmm2,%xmm3 + DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 + DB 102,15,104,208 ; punpckhbw %xmm0,%xmm2 DB 102,15,127,117,232 ; movdqa %xmm6,-0x18(%ebp) DB 102,15,111,206 ; movdqa %xmm6,%xmm1 DB 102,15,96,200 ; punpcklbw %xmm0,%xmm1 DB 102,15,111,198 ; movdqa %xmm6,%xmm0 DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 - DB 102,15,213,213 ; pmullw %xmm5,%xmm2 - DB 102,15,253,234 ; paddw %xmm2,%xmm5 - DB 102,15,213,223 ; pmullw %xmm7,%xmm3 - DB 102,15,253,223 ; paddw %xmm7,%xmm3 + DB 102,15,213,221 ; pmullw %xmm5,%xmm3 + DB 102,15,253,235 ; paddw %xmm3,%xmm5 + DB 102,15,213,215 ; pmullw %xmm7,%xmm2 + DB 102,15,253,215 ; paddw %xmm7,%xmm2 DB 102,15,213,204 ; pmullw %xmm4,%xmm1 DB 102,15,253,225 ; paddw %xmm1,%xmm4 DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 DB 102,15,213,193 ; pmullw %xmm1,%xmm0 DB 102,15,253,193 ; paddw %xmm1,%xmm0 - DB 102,15,113,211,8 ; psrlw $0x8,%xmm3 - DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,235 ; packuswb %xmm3,%xmm5 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,103,224 ; packuswb %xmm0,%xmm4 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,234 ; packuswb %xmm2,%xmm5 DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 DB 102,15,252,197 ; paddb %xmm5,%xmm0 DB 102,15,252,101,168 ; paddb -0x58(%ebp),%xmm4 @@ -59316,13 +59238,13 @@ _sk_xor__sse2_8bit LABEL PROC DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 DB 102,15,213,209 ; pmullw %xmm1,%xmm2 DB 102,15,253,209 ; paddw %xmm1,%xmm2 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,242 ; packuswb %xmm2,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,103,196 ; packuswb %xmm4,%xmm0 DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) - DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,242 ; packuswb %xmm2,%xmm6 DB 242,15,112,197,231 ; pshuflw $0xe7,%xmm5,%xmm0 DB 242,15,112,207,231 ; pshuflw $0xe7,%xmm7,%xmm1 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 @@ -59367,14 +59289,14 @@ _sk_xor__sse2_8bit LABEL PROC DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 DB 102,15,213,211 ; pmullw %xmm3,%xmm2 DB 102,15,253,211 ; paddw %xmm3,%xmm2 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,196 ; packuswb %xmm4,%xmm0 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,103,202 ; packuswb %xmm2,%xmm1 - DB 102,15,252,69,168 ; paddb -0x58(%ebp),%xmm0 DB 102,15,252,206 ; paddb %xmm6,%xmm1 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,196 ; packuswb %xmm4,%xmm0 + DB 102,15,252,69,168 ; paddb -0x58(%ebp),%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx DB 15,40,85,216 ; movaps -0x28(%ebp),%xmm2 @@ -59413,122 +59335,116 @@ _sk_darken_sse2_8bit LABEL PROC DB 131,236,120 ; sub $0x78,%esp DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,127,109,232 ; movdqa %xmm5,-0x18(%ebp) - DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) - DB 15,41,77,216 ; movaps %xmm1,-0x28(%ebp) - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,127,93,136 ; movdqa %xmm3,-0x78(%ebp) DB 242,15,112,202,231 ; pshuflw $0xe7,%xmm2,%xmm1 - DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 - DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,96,203 ; punpcklbw %xmm3,%xmm1 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,96,243 ; punpcklbw %xmm3,%xmm6 - DB 102,15,213,241 ; pmullw %xmm1,%xmm6 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,127,125,136 ; movdqa %xmm7,-0x78(%ebp) - DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 - DB 102,15,104,227 ; punpckhbw %xmm3,%xmm4 - DB 102,15,213,225 ; pmullw %xmm1,%xmm4 - DB 102,15,253,225 ; paddw %xmm1,%xmm4 - DB 242,15,112,205,231 ; pshuflw $0xe7,%xmm5,%xmm1 - DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 - DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,201,95 ; pshufhw $0x5f,%xmm1,%xmm1 - DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 + DB 243,15,112,209,231 ; pshufhw $0xe7,%xmm1,%xmm2 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 + DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 + DB 242,15,112,210,95 ; pshuflw $0x5f,%xmm2,%xmm2 + DB 243,15,112,250,95 ; pshufhw $0x5f,%xmm2,%xmm7 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,96,211 ; punpcklbw %xmm3,%xmm2 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,96,235 ; punpcklbw %xmm3,%xmm5 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,96,209 ; punpcklbw %xmm1,%xmm2 + DB 102,15,111,247 ; movdqa %xmm7,%xmm6 + DB 102,15,96,241 ; punpcklbw %xmm1,%xmm6 + DB 102,15,213,242 ; pmullw %xmm2,%xmm6 + DB 102,15,253,242 ; paddw %xmm2,%xmm6 + DB 102,15,127,101,152 ; movdqa %xmm4,-0x68(%ebp) + DB 102,15,104,193 ; punpckhbw %xmm1,%xmm0 + DB 102,15,104,249 ; punpckhbw %xmm1,%xmm7 + DB 102,15,213,248 ; pmullw %xmm0,%xmm7 + DB 102,15,253,248 ; paddw %xmm0,%xmm7 + DB 242,15,112,197,231 ; pshuflw $0xe7,%xmm5,%xmm0 + DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 + DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 + DB 243,15,112,192,95 ; pshufhw $0x5f,%xmm0,%xmm0 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,96,209 ; punpcklbw %xmm1,%xmm2 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,96,233 ; punpcklbw %xmm1,%xmm5 DB 102,15,213,234 ; pmullw %xmm2,%xmm5 DB 102,15,253,234 ; paddw %xmm2,%xmm5 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 - DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 - DB 102,15,213,202 ; pmullw %xmm2,%xmm1 - DB 102,15,253,202 ; paddw %xmm2,%xmm1 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,244 ; packuswb %xmm4,%xmm6 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,104,209 ; punpckhbw %xmm1,%xmm2 + DB 102,15,104,193 ; punpckhbw %xmm1,%xmm0 + DB 102,15,213,194 ; pmullw %xmm2,%xmm0 + DB 102,15,253,194 ; paddw %xmm2,%xmm0 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,233 ; packuswb %xmm1,%xmm5 - DB 242,15,112,192,231 ; pshuflw $0xe7,%xmm0,%xmm0 + DB 102,15,103,232 ; packuswb %xmm0,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,247 ; packuswb %xmm7,%xmm6 + DB 242,15,112,195,231 ; pshuflw $0xe7,%xmm3,%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 DB 243,15,112,208,95 ; pshufhw $0x5f,%xmm0,%xmm2 - DB 242,15,112,199,231 ; pshuflw $0xe7,%xmm7,%xmm0 + DB 242,15,112,196,231 ; pshuflw $0xe7,%xmm4,%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 - DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,15,96,220 ; punpcklbw %xmm4,%xmm3 - DB 102,15,127,93,184 ; movdqa %xmm3,-0x48(%ebp) - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,15,104,224 ; punpckhbw %xmm0,%xmm4 - DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,96,251 ; punpcklbw %xmm3,%xmm7 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,127,125,168 ; movdqa %xmm7,-0x58(%ebp) - DB 102,15,104,195 ; punpckhbw %xmm3,%xmm0 - DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) + DB 102,15,111,101,216 ; movdqa -0x28(%ebp),%xmm4 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,96,223 ; punpcklbw %xmm7,%xmm3 + DB 102,15,127,93,200 ; movdqa %xmm3,-0x38(%ebp) + DB 102,15,104,231 ; punpckhbw %xmm7,%xmm4 + DB 102,15,111,93,232 ; movdqa -0x18(%ebp),%xmm3 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 - DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 + DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) + DB 102,15,104,223 ; punpckhbw %xmm7,%xmm3 + DB 102,15,127,93,168 ; movdqa %xmm3,-0x58(%ebp) DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 + DB 102,15,111,249 ; movdqa %xmm1,%xmm7 DB 102,15,96,251 ; punpcklbw %xmm3,%xmm7 + DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,96,195 ; punpcklbw %xmm3,%xmm0 DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 + DB 102,15,111,93,200 ; movdqa -0x38(%ebp),%xmm3 + DB 102,15,213,251 ; pmullw %xmm3,%xmm7 + DB 102,15,253,251 ; paddw %xmm3,%xmm7 + DB 102,15,213,204 ; pmullw %xmm4,%xmm1 + DB 102,15,253,204 ; paddw %xmm4,%xmm1 DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 DB 102,15,213,195 ; pmullw %xmm3,%xmm0 DB 102,15,253,195 ; paddw %xmm3,%xmm0 - DB 102,15,213,204 ; pmullw %xmm4,%xmm1 - DB 102,15,253,204 ; paddw %xmm4,%xmm1 DB 102,15,111,93,168 ; movdqa -0x58(%ebp),%xmm3 - DB 102,15,213,251 ; pmullw %xmm3,%xmm7 - DB 102,15,253,251 ; paddw %xmm3,%xmm7 - DB 102,15,111,93,152 ; movdqa -0x68(%ebp),%xmm3 DB 102,15,213,211 ; pmullw %xmm3,%xmm2 DB 102,15,253,211 ; paddw %xmm3,%xmm2 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,193 ; packuswb %xmm1,%xmm0 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,194 ; packuswb %xmm2,%xmm0 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,250 ; packuswb %xmm2,%xmm7 - DB 102,15,222,239 ; pmaxub %xmm7,%xmm5 - DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 + DB 102,15,103,249 ; packuswb %xmm1,%xmm7 + DB 102,15,222,232 ; pmaxub %xmm0,%xmm5 + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 DB 102,15,252,77,232 ; paddb -0x18(%ebp),%xmm1 DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 102,15,248,229 ; psubb %xmm5,%xmm4 - DB 102,15,222,240 ; pmaxub %xmm0,%xmm6 - DB 102,15,111,93,136 ; movdqa -0x78(%ebp),%xmm3 - DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 + DB 102,15,222,247 ; pmaxub %xmm7,%xmm6 + DB 102,15,111,93,152 ; movdqa -0x68(%ebp),%xmm3 + DB 102,15,111,85,216 ; movdqa -0x28(%ebp),%xmm2 DB 102,15,252,218 ; paddb %xmm2,%xmm3 DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,248,238 ; psubb %xmm6,%xmm5 - DB 232,0,0,0,0 ; call c00c <_sk_darken_sse2_8bit+0x1f7> + DB 232,0,0,0,0 ; call bfcf <_sk_darken_sse2_8bit+0x1df> DB 88 ; pop %eax - DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,248,200 ; psubb %xmm0,%xmm1 DB 102,15,111,241 ; movdqa %xmm1,%xmm6 - DB 102,15,248,216 ; psubb %xmm0,%xmm3 - DB 102,15,111,128,100,23,0,0 ; movdqa 0x1764(%eax),%xmm0 + DB 102,15,248,223 ; psubb %xmm7,%xmm3 + DB 102,15,111,128,193,22,0,0 ; movdqa 0x16c1(%eax),%xmm0 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,223,203 ; pandn %xmm3,%xmm1 DB 102,15,219,232 ; pand %xmm0,%xmm5 @@ -59557,122 +59473,116 @@ _sk_lighten_sse2_8bit LABEL PROC DB 131,236,120 ; sub $0x78,%esp DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,127,109,232 ; movdqa %xmm5,-0x18(%ebp) - DB 102,15,127,85,200 ; movdqa %xmm2,-0x38(%ebp) - DB 15,41,77,216 ; movaps %xmm1,-0x28(%ebp) - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,127,93,136 ; movdqa %xmm3,-0x78(%ebp) DB 242,15,112,202,231 ; pshuflw $0xe7,%xmm2,%xmm1 - DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 - DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,96,203 ; punpcklbw %xmm3,%xmm1 - DB 102,15,111,244 ; movdqa %xmm4,%xmm6 - DB 102,15,96,243 ; punpcklbw %xmm3,%xmm6 - DB 102,15,213,241 ; pmullw %xmm1,%xmm6 - DB 102,15,253,241 ; paddw %xmm1,%xmm6 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 - DB 102,15,127,125,136 ; movdqa %xmm7,-0x78(%ebp) - DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 - DB 102,15,104,227 ; punpckhbw %xmm3,%xmm4 - DB 102,15,213,225 ; pmullw %xmm1,%xmm4 - DB 102,15,253,225 ; paddw %xmm1,%xmm4 - DB 242,15,112,205,231 ; pshuflw $0xe7,%xmm5,%xmm1 - DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 - DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 - DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 - DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,201,95 ; pshufhw $0x5f,%xmm1,%xmm1 - DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 + DB 243,15,112,209,231 ; pshufhw $0xe7,%xmm1,%xmm2 + DB 102,15,112,210,232 ; pshufd $0xe8,%xmm2,%xmm2 + DB 102,15,96,210 ; punpcklbw %xmm2,%xmm2 + DB 242,15,112,210,95 ; pshuflw $0x5f,%xmm2,%xmm2 + DB 243,15,112,250,95 ; pshufhw $0x5f,%xmm2,%xmm7 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,96,211 ; punpcklbw %xmm3,%xmm2 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,96,235 ; punpcklbw %xmm3,%xmm5 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,96,209 ; punpcklbw %xmm1,%xmm2 + DB 102,15,111,247 ; movdqa %xmm7,%xmm6 + DB 102,15,96,241 ; punpcklbw %xmm1,%xmm6 + DB 102,15,213,242 ; pmullw %xmm2,%xmm6 + DB 102,15,253,242 ; paddw %xmm2,%xmm6 + DB 102,15,127,101,152 ; movdqa %xmm4,-0x68(%ebp) + DB 102,15,104,193 ; punpckhbw %xmm1,%xmm0 + DB 102,15,104,249 ; punpckhbw %xmm1,%xmm7 + DB 102,15,213,248 ; pmullw %xmm0,%xmm7 + DB 102,15,253,248 ; paddw %xmm0,%xmm7 + DB 242,15,112,197,231 ; pshuflw $0xe7,%xmm5,%xmm0 + DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 + DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 + DB 243,15,112,192,95 ; pshufhw $0x5f,%xmm0,%xmm0 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,96,209 ; punpcklbw %xmm1,%xmm2 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,96,233 ; punpcklbw %xmm1,%xmm5 DB 102,15,213,234 ; pmullw %xmm2,%xmm5 DB 102,15,253,234 ; paddw %xmm2,%xmm5 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 - DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 - DB 102,15,213,202 ; pmullw %xmm2,%xmm1 - DB 102,15,253,202 ; paddw %xmm2,%xmm1 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,244 ; packuswb %xmm4,%xmm6 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,104,209 ; punpckhbw %xmm1,%xmm2 + DB 102,15,104,193 ; punpckhbw %xmm1,%xmm0 + DB 102,15,213,194 ; pmullw %xmm2,%xmm0 + DB 102,15,253,194 ; paddw %xmm2,%xmm0 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 - DB 102,15,103,233 ; packuswb %xmm1,%xmm5 - DB 242,15,112,192,231 ; pshuflw $0xe7,%xmm0,%xmm0 + DB 102,15,103,232 ; packuswb %xmm0,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,247 ; packuswb %xmm7,%xmm6 + DB 242,15,112,195,231 ; pshuflw $0xe7,%xmm3,%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 DB 243,15,112,208,95 ; pshufhw $0x5f,%xmm0,%xmm2 - DB 242,15,112,199,231 ; pshuflw $0xe7,%xmm7,%xmm0 + DB 242,15,112,196,231 ; pshuflw $0xe7,%xmm4,%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 - DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,239,228 ; pxor %xmm4,%xmm4 - DB 102,15,96,220 ; punpcklbw %xmm4,%xmm3 - DB 102,15,127,93,184 ; movdqa %xmm3,-0x48(%ebp) - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,15,104,224 ; punpckhbw %xmm0,%xmm4 - DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 - DB 102,15,111,248 ; movdqa %xmm0,%xmm7 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,96,251 ; punpcklbw %xmm3,%xmm7 - DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,127,125,168 ; movdqa %xmm7,-0x58(%ebp) - DB 102,15,104,195 ; punpckhbw %xmm3,%xmm0 - DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) + DB 102,15,111,101,216 ; movdqa -0x28(%ebp),%xmm4 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 102,15,239,255 ; pxor %xmm7,%xmm7 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,96,223 ; punpcklbw %xmm7,%xmm3 + DB 102,15,127,93,200 ; movdqa %xmm3,-0x38(%ebp) + DB 102,15,104,231 ; punpckhbw %xmm7,%xmm4 + DB 102,15,111,93,232 ; movdqa -0x18(%ebp),%xmm3 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 - DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 + DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) + DB 102,15,104,223 ; punpckhbw %xmm7,%xmm3 + DB 102,15,127,93,168 ; movdqa %xmm3,-0x58(%ebp) DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 + DB 102,15,111,249 ; movdqa %xmm1,%xmm7 DB 102,15,96,251 ; punpcklbw %xmm3,%xmm7 + DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,96,195 ; punpcklbw %xmm3,%xmm0 DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 + DB 102,15,111,93,200 ; movdqa -0x38(%ebp),%xmm3 + DB 102,15,213,251 ; pmullw %xmm3,%xmm7 + DB 102,15,253,251 ; paddw %xmm3,%xmm7 + DB 102,15,213,204 ; pmullw %xmm4,%xmm1 + DB 102,15,253,204 ; paddw %xmm4,%xmm1 DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 DB 102,15,213,195 ; pmullw %xmm3,%xmm0 DB 102,15,253,195 ; paddw %xmm3,%xmm0 - DB 102,15,213,204 ; pmullw %xmm4,%xmm1 - DB 102,15,253,204 ; paddw %xmm4,%xmm1 DB 102,15,111,93,168 ; movdqa -0x58(%ebp),%xmm3 - DB 102,15,213,251 ; pmullw %xmm3,%xmm7 - DB 102,15,253,251 ; paddw %xmm3,%xmm7 - DB 102,15,111,93,152 ; movdqa -0x68(%ebp),%xmm3 DB 102,15,213,211 ; pmullw %xmm3,%xmm2 DB 102,15,253,211 ; paddw %xmm3,%xmm2 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,193 ; packuswb %xmm1,%xmm0 DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,194 ; packuswb %xmm2,%xmm0 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,250 ; packuswb %xmm2,%xmm7 - DB 102,15,218,239 ; pminub %xmm7,%xmm5 - DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 + DB 102,15,103,249 ; packuswb %xmm1,%xmm7 + DB 102,15,218,232 ; pminub %xmm0,%xmm5 + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 DB 102,15,252,77,232 ; paddb -0x18(%ebp),%xmm1 DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 102,15,248,229 ; psubb %xmm5,%xmm4 - DB 102,15,218,240 ; pminub %xmm0,%xmm6 - DB 102,15,111,93,136 ; movdqa -0x78(%ebp),%xmm3 - DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 + DB 102,15,218,247 ; pminub %xmm7,%xmm6 + DB 102,15,111,93,152 ; movdqa -0x68(%ebp),%xmm3 + DB 102,15,111,85,216 ; movdqa -0x28(%ebp),%xmm2 DB 102,15,252,218 ; paddb %xmm2,%xmm3 DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,248,238 ; psubb %xmm6,%xmm5 - DB 232,0,0,0,0 ; call c25a <_sk_lighten_sse2_8bit+0x1f7> + DB 232,0,0,0,0 ; call c205 <_sk_lighten_sse2_8bit+0x1df> DB 88 ; pop %eax - DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,248,200 ; psubb %xmm0,%xmm1 DB 102,15,111,241 ; movdqa %xmm1,%xmm6 - DB 102,15,248,216 ; psubb %xmm0,%xmm3 - DB 102,15,111,128,22,21,0,0 ; movdqa 0x1516(%eax),%xmm0 + DB 102,15,248,223 ; psubb %xmm7,%xmm3 + DB 102,15,111,128,139,20,0,0 ; movdqa 0x148b(%eax),%xmm0 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 DB 102,15,223,203 ; pandn %xmm3,%xmm1 DB 102,15,219,232 ; pand %xmm0,%xmm5 @@ -59701,63 +59611,65 @@ _sk_exclusion_sse2_8bit LABEL PROC DB 131,236,88 ; sub $0x58,%esp DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,15,111,226 ; movdqa %xmm2,%xmm4 - DB 102,15,111,249 ; movdqa %xmm1,%xmm7 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 - DB 102,15,239,246 ; pxor %xmm6,%xmm6 - DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,239,255 ; pxor %xmm7,%xmm7 + DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 DB 102,15,127,69,232 ; movdqa %xmm0,-0x18(%ebp) - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) - DB 102,15,111,199 ; movdqa %xmm7,%xmm0 - DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) - DB 102,15,111,199 ; movdqa %xmm7,%xmm0 - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 - DB 102,15,127,77,168 ; movdqa %xmm1,-0x58(%ebp) - DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 - DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 - DB 102,15,111,213 ; movdqa %xmm5,%xmm2 - DB 102,15,96,238 ; punpcklbw %xmm6,%xmm5 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 - DB 102,15,111,117,232 ; movdqa -0x18(%ebp),%xmm6 - DB 102,15,213,230 ; pmullw %xmm6,%xmm4 - DB 102,15,253,230 ; paddw %xmm6,%xmm4 - DB 102,15,111,117,200 ; movdqa -0x38(%ebp),%xmm6 - DB 102,15,213,206 ; pmullw %xmm6,%xmm1 - DB 102,15,253,206 ; paddw %xmm6,%xmm1 - DB 102,15,111,117,216 ; movdqa -0x28(%ebp),%xmm6 - DB 102,15,213,238 ; pmullw %xmm6,%xmm5 - DB 102,15,253,238 ; paddw %xmm6,%xmm5 - DB 102,15,111,117,184 ; movdqa -0x48(%ebp),%xmm6 - DB 102,15,213,198 ; pmullw %xmm6,%xmm0 - DB 102,15,253,198 ; paddw %xmm6,%xmm0 - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 - DB 102,15,103,225 ; packuswb %xmm1,%xmm4 - DB 232,0,0,0,0 ; call c363 <_sk_exclusion_sse2_8bit+0xb2> - DB 88 ; pop %eax + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) + DB 102,15,96,231 ; punpcklbw %xmm7,%xmm4 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,15,104,247 ; punpckhbw %xmm7,%xmm6 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 102,15,96,239 ; punpcklbw %xmm7,%xmm5 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 + DB 102,15,111,125,232 ; movdqa -0x18(%ebp),%xmm7 + DB 102,15,213,231 ; pmullw %xmm7,%xmm4 + DB 102,15,253,231 ; paddw %xmm7,%xmm4 + DB 102,15,111,125,200 ; movdqa -0x38(%ebp),%xmm7 + DB 102,15,213,247 ; pmullw %xmm7,%xmm6 + DB 102,15,253,247 ; paddw %xmm7,%xmm6 + DB 102,15,111,125,216 ; movdqa -0x28(%ebp),%xmm7 + DB 102,15,213,239 ; pmullw %xmm7,%xmm5 + DB 102,15,253,239 ; paddw %xmm7,%xmm5 + DB 102,15,111,125,184 ; movdqa -0x48(%ebp),%xmm7 + DB 102,15,213,199 ; pmullw %xmm7,%xmm0 + DB 102,15,253,199 ; paddw %xmm7,%xmm0 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 DB 102,15,103,232 ; packuswb %xmm0,%xmm5 - DB 102,15,111,128,13,20,0,0 ; movdqa 0x140d(%eax),%xmm0 - DB 102,15,111,242 ; movdqa %xmm2,%xmm6 - DB 102,15,252,254 ; paddb %xmm6,%xmm7 - DB 102,15,248,253 ; psubb %xmm5,%xmm7 + DB 232,0,0,0,0 ; call c312 <_sk_exclusion_sse2_8bit+0xb6> + DB 88 ; pop %eax + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,230 ; packuswb %xmm6,%xmm4 + DB 102,15,111,128,126,19,0,0 ; movdqa 0x137e(%eax),%xmm0 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,15,252,222 ; paddb %xmm6,%xmm3 + DB 102,15,248,221 ; psubb %xmm5,%xmm3 DB 102,15,219,232 ; pand %xmm0,%xmm5 + DB 102,15,248,221 ; psubb %xmm5,%xmm3 DB 102,15,219,196 ; pand %xmm4,%xmm0 - DB 102,15,111,85,168 ; movdqa -0x58(%ebp),%xmm2 - DB 102,15,252,218 ; paddb %xmm2,%xmm3 - DB 102,15,248,220 ; psubb %xmm4,%xmm3 - DB 102,15,248,216 ; psubb %xmm0,%xmm3 - DB 102,15,248,253 ; psubb %xmm5,%xmm7 + DB 102,15,111,109,168 ; movdqa -0x58(%ebp),%xmm5 + DB 102,15,252,213 ; paddb %xmm5,%xmm2 + DB 102,15,248,212 ; psubb %xmm4,%xmm2 + DB 102,15,248,208 ; psubb %xmm0,%xmm2 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx - DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,111,207 ; movdqa %xmm7,%xmm1 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,111,213 ; movdqa %xmm5,%xmm2 DB 102,15,111,222 ; movdqa %xmm6,%xmm3 DB 255,117,20 ; pushl 0x14(%ebp) DB 255,117,16 ; pushl 0x10(%ebp) @@ -59814,25 +59726,25 @@ _sk_difference_sse2_8bit LABEL PROC DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 DB 102,15,213,202 ; pmullw %xmm2,%xmm1 DB 102,15,253,202 ; paddw %xmm2,%xmm1 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,241 ; packuswb %xmm1,%xmm6 DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 DB 102,15,103,252 ; packuswb %xmm4,%xmm7 DB 102,15,127,125,152 ; movdqa %xmm7,-0x68(%ebp) - DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 - DB 102,15,103,241 ; packuswb %xmm1,%xmm6 DB 242,15,112,192,231 ; pshuflw $0xe7,%xmm0,%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,208,95 ; pshufhw $0x5f,%xmm0,%xmm2 + DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 DB 242,15,112,69,200,231 ; pshuflw $0xe7,-0x38(%ebp),%xmm0 DB 243,15,112,192,231 ; pshufhw $0xe7,%xmm0,%xmm0 DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 - DB 243,15,112,200,95 ; pshufhw $0x5f,%xmm0,%xmm1 + DB 243,15,112,208,95 ; pshufhw $0x5f,%xmm0,%xmm2 DB 102,15,111,101,216 ; movdqa -0x28(%ebp),%xmm4 DB 102,15,111,220 ; movdqa %xmm4,%xmm3 DB 102,15,239,255 ; pxor %xmm7,%xmm7 @@ -59845,51 +59757,49 @@ _sk_difference_sse2_8bit LABEL PROC DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) DB 102,15,239,192 ; pxor %xmm0,%xmm0 - DB 102,15,111,249 ; movdqa %xmm1,%xmm7 + DB 102,15,111,250 ; movdqa %xmm2,%xmm7 DB 102,15,96,248 ; punpcklbw %xmm0,%xmm7 - DB 102,15,104,200 ; punpckhbw %xmm0,%xmm1 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,104,208 ; punpckhbw %xmm0,%xmm2 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,96,195 ; punpcklbw %xmm3,%xmm0 DB 102,15,239,219 ; pxor %xmm3,%xmm3 - DB 102,15,104,211 ; punpckhbw %xmm3,%xmm2 + DB 102,15,104,203 ; punpckhbw %xmm3,%xmm1 DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 DB 102,15,213,251 ; pmullw %xmm3,%xmm7 DB 102,15,253,251 ; paddw %xmm3,%xmm7 - DB 102,15,213,204 ; pmullw %xmm4,%xmm1 - DB 102,15,253,204 ; paddw %xmm4,%xmm1 + DB 102,15,213,212 ; pmullw %xmm4,%xmm2 + DB 102,15,253,212 ; paddw %xmm4,%xmm2 DB 102,15,213,197 ; pmullw %xmm5,%xmm0 DB 102,15,253,197 ; paddw %xmm5,%xmm0 DB 102,15,111,93,168 ; movdqa -0x58(%ebp),%xmm3 - DB 102,15,213,211 ; pmullw %xmm3,%xmm2 - DB 102,15,253,211 ; paddw %xmm3,%xmm2 + DB 102,15,213,203 ; pmullw %xmm3,%xmm1 + DB 102,15,253,203 ; paddw %xmm3,%xmm1 DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 - DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 - DB 102,15,103,249 ; packuswb %xmm1,%xmm7 - DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 - DB 102,15,103,194 ; packuswb %xmm2,%xmm0 - DB 102,15,218,125,152 ; pminub -0x68(%ebp),%xmm7 - DB 232,0,0,0,0 ; call c57e <_sk_difference_sse2_8bit+0x1b8> - DB 88 ; pop %eax + DB 102,15,103,193 ; packuswb %xmm1,%xmm0 DB 102,15,218,198 ; pminub %xmm6,%xmm0 - DB 102,15,111,136,242,17,0,0 ; movdqa 0x11f2(%eax),%xmm1 - DB 102,15,111,101,136 ; movdqa -0x78(%ebp),%xmm4 + DB 102,15,113,210,8 ; psrlw $0x8,%xmm2 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,250 ; packuswb %xmm2,%xmm7 + DB 232,0,0,0,0 ; call c530 <_sk_difference_sse2_8bit+0x1b7> + DB 88 ; pop %eax + DB 102,15,218,125,152 ; pminub -0x68(%ebp),%xmm7 + DB 102,15,111,160,96,17,0,0 ; movdqa 0x1160(%eax),%xmm4 + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 DB 102,15,111,93,232 ; movdqa -0x18(%ebp),%xmm3 - DB 102,15,252,227 ; paddb %xmm3,%xmm4 - DB 102,15,248,224 ; psubb %xmm0,%xmm4 - DB 102,15,219,193 ; pand %xmm1,%xmm0 - DB 102,15,219,207 ; pand %xmm7,%xmm1 - DB 102,15,111,109,200 ; movdqa -0x38(%ebp),%xmm5 + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,248,200 ; psubb %xmm0,%xmm1 + DB 102,15,219,196 ; pand %xmm4,%xmm0 + DB 102,15,248,200 ; psubb %xmm0,%xmm1 + DB 102,15,219,231 ; pand %xmm7,%xmm4 + DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 DB 102,15,111,85,216 ; movdqa -0x28(%ebp),%xmm2 - DB 102,15,252,234 ; paddb %xmm2,%xmm5 - DB 102,15,248,239 ; psubb %xmm7,%xmm5 - DB 102,15,248,233 ; psubb %xmm1,%xmm5 - DB 102,15,248,224 ; psubb %xmm0,%xmm4 - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,252,194 ; paddb %xmm2,%xmm0 + DB 102,15,248,199 ; psubb %xmm7,%xmm0 + DB 102,15,248,196 ; psubb %xmm4,%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 DB 255,117,20 ; pushl 0x14(%ebp) DB 255,117,16 ; pushl 0x10(%ebp) DB 81 ; push %ecx @@ -59903,53 +59813,55 @@ PUBLIC _sk_hardlight_sse2_8bit _sk_hardlight_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp - DB 129,236,168,1,0,0 ; sub $0x1a8,%esp - DB 102,15,239,255 ; pxor %xmm7,%xmm7 + DB 129,236,120,1,0,0 ; sub $0x178,%esp + DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,96,231 ; punpcklbw %xmm7,%xmm4 - DB 102,15,127,101,232 ; movdqa %xmm4,-0x18(%ebp) + DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 + DB 102,15,127,165,120,255,255,255 ; movdqa %xmm4,-0x88(%ebp) DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,96,231 ; punpcklbw %xmm7,%xmm4 - DB 102,15,127,101,136 ; movdqa %xmm4,-0x78(%ebp) + DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 + DB 102,15,127,165,40,255,255,255 ; movdqa %xmm4,-0xd8(%ebp) DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 - DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 - DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) + DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,127,133,248,254,255,255 ; movdqa %xmm0,-0x108(%ebp) DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,236,95 ; pshufhw $0x5f,%xmm4,%xmm5 + DB 243,15,112,196,95 ; pshufhw $0x5f,%xmm4,%xmm0 DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 - DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 - DB 102,15,127,141,40,255,255,255 ; movdqa %xmm1,-0xd8(%ebp) + DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 DB 243,15,112,204,231 ; pshufhw $0xe7,%xmm4,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,193,95 ; pshufhw $0x5f,%xmm1,%xmm0 - DB 102,15,111,240 ; movdqa %xmm0,%xmm6 - DB 102,15,104,247 ; punpckhbw %xmm7,%xmm6 - DB 102,15,127,181,200,254,255,255 ; movdqa %xmm6,-0x138(%ebp) - DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 - DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 - DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 - DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) - DB 102,15,96,239 ; punpcklbw %xmm7,%xmm5 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,96,199 ; punpcklbw %xmm7,%xmm0 - DB 102,15,127,133,120,255,255,255 ; movdqa %xmm0,-0x88(%ebp) - DB 102,15,111,202 ; movdqa %xmm2,%xmm1 - DB 102,15,127,149,88,254,255,255 ; movdqa %xmm2,-0x1a8(%ebp) - DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 - DB 102,15,127,141,104,255,255,255 ; movdqa %xmm1,-0x98(%ebp) - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,96,207 ; punpcklbw %xmm7,%xmm1 + DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 + DB 102,15,127,77,152 ; movdqa %xmm1,-0x68(%ebp) + DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 + DB 102,15,127,101,168 ; movdqa %xmm4,-0x58(%ebp) + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 DB 102,15,127,77,184 ; movdqa %xmm1,-0x48(%ebp) - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,127,157,104,254,255,255 ; movdqa %xmm3,-0x198(%ebp) - DB 102,15,104,207 ; punpckhbw %xmm7,%xmm1 - DB 102,15,127,141,120,254,255,255 ; movdqa %xmm1,-0x188(%ebp) + DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 + DB 102,15,127,165,88,255,255,255 ; movdqa %xmm4,-0xa8(%ebp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 + DB 102,15,127,69,136 ; movdqa %xmm0,-0x78(%ebp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,127,149,136,254,255,255 ; movdqa %xmm2,-0x178(%ebp) + DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,127,133,24,255,255,255 ; movdqa %xmm0,-0xe8(%ebp) + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 + DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) + DB 102,15,111,195 ; movdqa %xmm3,%xmm0 + DB 102,15,127,157,152,254,255,255 ; movdqa %xmm3,-0x168(%ebp) + DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) DB 242,15,112,202,231 ; pshuflw $0xe7,%xmm2,%xmm1 DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 @@ -59963,205 +59875,187 @@ _sk_hardlight_sse2_8bit LABEL PROC DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 DB 243,15,112,201,95 ; pshufhw $0x5f,%xmm1,%xmm1 DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 - DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) - DB 102,15,96,207 ; punpcklbw %xmm7,%xmm1 - DB 102,15,127,141,8,255,255,255 ; movdqa %xmm1,-0xf8(%ebp) - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,104,199 ; punpckhbw %xmm7,%xmm0 - DB 102,15,127,133,24,255,255,255 ; movdqa %xmm0,-0xe8(%ebp) - DB 102,15,96,215 ; punpcklbw %xmm7,%xmm2 - DB 102,15,127,149,88,255,255,255 ; movdqa %xmm2,-0xa8(%ebp) - DB 102,15,111,101,232 ; movdqa -0x18(%ebp),%xmm4 - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 - DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,15,111,93,136 ; movdqa -0x78(%ebp),%xmm3 - DB 102,15,253,219 ; paddw %xmm3,%xmm3 - DB 102,15,111,149,40,255,255,255 ; movdqa -0xd8(%ebp),%xmm2 - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,217,214 ; psubusw %xmm6,%xmm2 - DB 102,15,117,215 ; pcmpeqw %xmm7,%xmm2 - DB 102,15,127,149,136,254,255,255 ; movdqa %xmm2,-0x178(%ebp) - DB 102,15,217,93,200 ; psubusw -0x38(%ebp),%xmm3 - DB 102,15,117,223 ; pcmpeqw %xmm7,%xmm3 - DB 102,15,127,157,152,254,255,255 ; movdqa %xmm3,-0x168(%ebp) - DB 102,15,111,85,152 ; movdqa -0x68(%ebp),%xmm2 - DB 102,15,217,202 ; psubusw %xmm2,%xmm1 - DB 102,15,117,207 ; pcmpeqw %xmm7,%xmm1 - DB 102,15,127,141,56,255,255,255 ; movdqa %xmm1,-0xc8(%ebp) - DB 102,15,217,229 ; psubusw %xmm5,%xmm4 - DB 102,15,117,231 ; pcmpeqw %xmm7,%xmm4 - DB 102,15,127,165,72,255,255,255 ; movdqa %xmm4,-0xb8(%ebp) - DB 232,0,0,0,0 ; call c7af <_sk_hardlight_sse2_8bit+0x1ce> + DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,127,133,72,255,255,255 ; movdqa %xmm0,-0xb8(%ebp) + DB 102,15,96,206 ; punpcklbw %xmm6,%xmm1 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 + DB 102,15,127,157,104,255,255,255 ; movdqa %xmm3,-0x98(%ebp) + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 + DB 102,15,127,141,232,254,255,255 ; movdqa %xmm1,-0x118(%ebp) + DB 102,15,96,214 ; punpcklbw %xmm6,%xmm2 + DB 102,15,127,85,232 ; movdqa %xmm2,-0x18(%ebp) + DB 232,0,0,0,0 ; call c6f9 <_sk_hardlight_sse2_8bit+0x16d> DB 88 ; pop %eax - DB 102,15,111,176,161,15,0,0 ; movdqa 0xfa1(%eax),%xmm6 - DB 102,15,111,157,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm3 - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,239,206 ; pxor %xmm6,%xmm1 - DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 - DB 102,15,213,200 ; pmullw %xmm0,%xmm1 - DB 102,15,127,141,168,254,255,255 ; movdqa %xmm1,-0x158(%ebp) - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 102,15,111,229 ; movdqa %xmm5,%xmm4 - DB 102,15,239,206 ; pxor %xmm6,%xmm1 - DB 102,15,111,173,120,255,255,255 ; movdqa -0x88(%ebp),%xmm5 - DB 102,15,213,205 ; pmullw %xmm5,%xmm1 - DB 102,15,127,141,184,254,255,255 ; movdqa %xmm1,-0x148(%ebp) - DB 102,15,111,203 ; movdqa %xmm3,%xmm1 - DB 102,15,111,249 ; movdqa %xmm1,%xmm7 - DB 102,15,111,221 ; movdqa %xmm5,%xmm3 - DB 102,15,249,203 ; psubw %xmm3,%xmm1 - DB 102,15,127,141,88,255,255,255 ; movdqa %xmm1,-0xa8(%ebp) + DB 102,15,111,184,119,15,0,0 ; movdqa 0xf77(%eax),%xmm7 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,239,247 ; pxor %xmm7,%xmm6 + DB 102,15,239,207 ; pxor %xmm7,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,239,215 ; pxor %xmm7,%xmm2 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,239,223 ; pxor %xmm7,%xmm3 + DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,127,133,56,255,255,255 ; movdqa %xmm0,-0xc8(%ebp) DB 102,15,213,216 ; pmullw %xmm0,%xmm3 - DB 102,15,127,157,232,254,255,255 ; movdqa %xmm3,-0x118(%ebp) - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,111,236 ; movdqa %xmm4,%xmm5 + DB 102,15,239,239 ; pxor %xmm7,%xmm5 + DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 + DB 102,15,239,199 ; pxor %xmm7,%xmm0 + DB 102,15,111,101,168 ; movdqa -0x58(%ebp),%xmm4 + DB 102,15,239,231 ; pxor %xmm7,%xmm4 + DB 102,15,239,125,152 ; pxor -0x68(%ebp),%xmm7 + DB 102,15,213,125,216 ; pmullw -0x28(%ebp),%xmm7 + DB 102,15,253,251 ; paddw %xmm3,%xmm7 DB 102,15,127,189,216,254,255,255 ; movdqa %xmm7,-0x128(%ebp) - DB 102,15,249,224 ; psubw %xmm0,%xmm4 - DB 102,15,127,165,248,254,255,255 ; movdqa %xmm4,-0x108(%ebp) - DB 102,15,111,141,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm1 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,15,239,222 ; pxor %xmm6,%xmm3 - DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 - DB 102,15,213,216 ; pmullw %xmm0,%xmm3 - DB 102,15,127,157,120,255,255,255 ; movdqa %xmm3,-0x88(%ebp) - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,239,254 ; pxor %xmm6,%xmm7 - DB 102,15,111,165,104,255,255,255 ; movdqa -0x98(%ebp),%xmm4 - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,15,249,204 ; psubw %xmm4,%xmm1 - DB 102,15,127,141,24,255,255,255 ; movdqa %xmm1,-0xe8(%ebp) - DB 102,15,213,224 ; pmullw %xmm0,%xmm4 - DB 102,15,213,218 ; pmullw %xmm2,%xmm3 - DB 102,15,127,93,232 ; movdqa %xmm3,-0x18(%ebp) - DB 102,15,249,208 ; psubw %xmm0,%xmm2 - DB 102,15,127,85,152 ; movdqa %xmm2,-0x68(%ebp) - DB 102,15,111,141,8,255,255,255 ; movdqa -0xf8(%ebp),%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,239,198 ; pxor %xmm6,%xmm0 - DB 102,15,111,93,136 ; movdqa -0x78(%ebp),%xmm3 + DB 102,15,213,149,40,255,255,255 ; pmullw -0xd8(%ebp),%xmm2 + DB 102,15,213,101,200 ; pmullw -0x38(%ebp),%xmm4 + DB 102,15,253,226 ; paddw %xmm2,%xmm4 + DB 102,15,127,165,168,254,255,255 ; movdqa %xmm4,-0x158(%ebp) + DB 102,15,111,149,248,254,255,255 ; movdqa -0x108(%ebp),%xmm2 + DB 102,15,213,202 ; pmullw %xmm2,%xmm1 + DB 102,15,111,157,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm3 DB 102,15,213,195 ; pmullw %xmm3,%xmm0 - DB 102,15,127,133,104,255,255,255 ; movdqa %xmm0,-0x98(%ebp) - DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,239,214 ; pxor %xmm6,%xmm2 - DB 102,15,127,85,216 ; movdqa %xmm2,-0x28(%ebp) - DB 102,15,111,85,184 ; movdqa -0x48(%ebp),%xmm2 - DB 102,15,111,109,216 ; movdqa -0x28(%ebp),%xmm5 - DB 102,15,213,234 ; pmullw %xmm2,%xmm5 - DB 102,15,127,109,216 ; movdqa %xmm5,-0x28(%ebp) + DB 102,15,253,193 ; paddw %xmm1,%xmm0 + DB 102,15,127,133,184,254,255,255 ; movdqa %xmm0,-0x148(%ebp) + DB 102,15,111,141,120,255,255,255 ; movdqa -0x88(%ebp),%xmm1 + DB 102,15,213,241 ; pmullw %xmm1,%xmm6 + DB 102,15,111,101,136 ; movdqa -0x78(%ebp),%xmm4 + DB 102,15,213,236 ; pmullw %xmm4,%xmm5 + DB 102,15,253,238 ; paddw %xmm6,%xmm5 + DB 102,15,127,173,200,254,255,255 ; movdqa %xmm5,-0x138(%ebp) DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,249,202 ; psubw %xmm2,%xmm1 - DB 102,15,127,141,8,255,255,255 ; movdqa %xmm1,-0xf8(%ebp) - DB 102,15,213,211 ; pmullw %xmm3,%xmm2 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 - DB 102,15,127,109,184 ; movdqa %xmm5,-0x48(%ebp) - DB 102,15,249,195 ; psubw %xmm3,%xmm0 - DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,253,237 ; paddw %xmm5,%xmm5 + DB 102,15,111,141,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm1 + DB 102,15,101,233 ; pcmpgtw %xmm1,%xmm5 + DB 102,15,127,173,8,255,255,255 ; movdqa %xmm5,-0xf8(%ebp) + DB 102,15,111,109,232 ; movdqa -0x18(%ebp),%xmm5 + DB 102,15,111,245 ; movdqa %xmm5,%xmm6 + DB 102,15,249,236 ; psubw %xmm4,%xmm5 + DB 102,15,213,224 ; pmullw %xmm0,%xmm4 + DB 102,15,127,101,232 ; movdqa %xmm4,-0x18(%ebp) + DB 102,15,213,241 ; pmullw %xmm1,%xmm6 + DB 102,15,127,117,136 ; movdqa %xmm6,-0x78(%ebp) + DB 102,15,249,200 ; psubw %xmm0,%xmm1 + DB 102,15,213,205 ; pmullw %xmm5,%xmm1 + DB 102,15,127,141,88,255,255,255 ; movdqa %xmm1,-0xa8(%ebp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,253,210 ; paddw %xmm2,%xmm2 + DB 102,15,111,77,184 ; movdqa -0x48(%ebp),%xmm1 + DB 102,15,101,209 ; pcmpgtw %xmm1,%xmm2 + DB 102,15,111,165,232,254,255,255 ; movdqa -0x118(%ebp),%xmm4 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 + DB 102,15,249,227 ; psubw %xmm3,%xmm4 + DB 102,15,213,216 ; pmullw %xmm0,%xmm3 + DB 102,15,127,157,120,255,255,255 ; movdqa %xmm3,-0x88(%ebp) + DB 102,15,213,249 ; pmullw %xmm1,%xmm7 + DB 102,15,249,200 ; psubw %xmm0,%xmm1 + DB 102,15,213,204 ; pmullw %xmm4,%xmm1 + DB 102,15,127,77,184 ; movdqa %xmm1,-0x48(%ebp) + DB 102,15,111,133,40,255,255,255 ; movdqa -0xd8(%ebp),%xmm0 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,15,253,237 ; paddw %xmm5,%xmm5 DB 102,15,111,77,168 ; movdqa -0x58(%ebp),%xmm1 - DB 102,15,111,217 ; movdqa %xmm1,%xmm3 - DB 102,15,239,222 ; pxor %xmm6,%xmm3 - DB 102,15,213,157,40,255,255,255 ; pmullw -0xd8(%ebp),%xmm3 - DB 102,15,127,93,136 ; movdqa %xmm3,-0x78(%ebp) - DB 102,15,111,133,200,254,255,255 ; movdqa -0x138(%ebp),%xmm0 - DB 102,15,239,240 ; pxor %xmm0,%xmm6 - DB 102,15,111,157,120,254,255,255 ; movdqa -0x188(%ebp),%xmm3 - DB 102,15,213,243 ; pmullw %xmm3,%xmm6 - DB 102,15,111,233 ; movdqa %xmm1,%xmm5 - DB 102,15,249,203 ; psubw %xmm3,%xmm1 + DB 102,15,101,233 ; pcmpgtw %xmm1,%xmm5 + DB 102,15,111,165,104,255,255,255 ; movdqa -0x98(%ebp),%xmm4 + DB 102,15,111,220 ; movdqa %xmm4,%xmm3 + DB 102,15,111,117,200 ; movdqa -0x38(%ebp),%xmm6 + DB 102,15,249,230 ; psubw %xmm6,%xmm4 + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 + DB 102,15,127,117,200 ; movdqa %xmm6,-0x38(%ebp) + DB 102,15,213,217 ; pmullw %xmm1,%xmm3 + DB 102,15,127,157,104,255,255,255 ; movdqa %xmm3,-0x98(%ebp) + DB 102,15,249,200 ; psubw %xmm0,%xmm1 + DB 102,15,213,204 ; pmullw %xmm4,%xmm1 DB 102,15,127,77,168 ; movdqa %xmm1,-0x58(%ebp) - DB 102,15,111,141,40,255,255,255 ; movdqa -0xd8(%ebp),%xmm1 + DB 102,15,111,165,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm4 + DB 102,15,253,228 ; paddw %xmm4,%xmm4 + DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 + DB 102,15,101,224 ; pcmpgtw %xmm0,%xmm4 + DB 102,15,111,141,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm1 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,15,249,77,216 ; psubw -0x28(%ebp),%xmm1 + DB 102,15,127,141,72,255,255,255 ; movdqa %xmm1,-0xb8(%ebp) + DB 102,15,111,93,216 ; movdqa -0x28(%ebp),%xmm3 + DB 102,15,111,141,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm1 DB 102,15,213,217 ; pmullw %xmm1,%xmm3 - DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,127,93,216 ; movdqa %xmm3,-0x28(%ebp) + DB 102,15,213,240 ; pmullw %xmm0,%xmm6 DB 102,15,249,193 ; psubw %xmm1,%xmm0 - DB 102,15,213,69,168 ; pmullw -0x58(%ebp),%xmm0 - DB 102,15,111,77,200 ; movdqa -0x38(%ebp),%xmm1 - DB 102,15,213,141,8,255,255,255 ; pmullw -0xf8(%ebp),%xmm1 - DB 102,15,127,77,200 ; movdqa %xmm1,-0x38(%ebp) - DB 102,15,111,77,152 ; movdqa -0x68(%ebp),%xmm1 - DB 102,15,213,141,24,255,255,255 ; pmullw -0xe8(%ebp),%xmm1 - DB 102,15,127,77,152 ; movdqa %xmm1,-0x68(%ebp) - DB 102,15,111,141,248,254,255,255 ; movdqa -0x108(%ebp),%xmm1 - DB 102,15,213,141,88,255,255,255 ; pmullw -0xa8(%ebp),%xmm1 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,249,232 ; psubw %xmm0,%xmm5 - DB 102,15,127,109,168 ; movdqa %xmm5,-0x58(%ebp) - DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 + DB 102,15,213,133,72,255,255,255 ; pmullw -0xb8(%ebp),%xmm0 + DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) + DB 102,15,111,157,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 + DB 102,15,249,203 ; psubw %xmm3,%xmm1 + DB 102,15,111,93,184 ; movdqa -0x48(%ebp),%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,249,251 ; psubw %xmm3,%xmm7 + DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,109,184 ; movdqa -0x48(%ebp),%xmm5 - DB 102,15,249,232 ; psubw %xmm0,%xmm5 - DB 102,15,127,109,184 ; movdqa %xmm5,-0x48(%ebp) + DB 102,15,111,157,104,255,255,255 ; movdqa -0x98(%ebp),%xmm3 + DB 102,15,249,216 ; psubw %xmm0,%xmm3 DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,109,232 ; movdqa -0x18(%ebp),%xmm5 - DB 102,15,249,232 ; psubw %xmm0,%xmm5 - DB 102,15,127,109,232 ; movdqa %xmm5,-0x18(%ebp) - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,249,240 ; psubw %xmm0,%xmm6 + DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,141,216,254,255,255 ; movdqa -0x128(%ebp),%xmm1 - DB 102,15,249,200 ; psubw %xmm0,%xmm1 - DB 102,15,111,173,232,254,255,255 ; movdqa -0x118(%ebp),%xmm5 - DB 102,15,253,237 ; paddw %xmm5,%xmm5 - DB 102,15,111,133,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm0 - DB 102,15,219,232 ; pand %xmm0,%xmm5 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,15,235,197 ; por %xmm5,%xmm0 - DB 102,15,127,133,72,255,255,255 ; movdqa %xmm0,-0xb8(%ebp) - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,15,111,133,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm0 - DB 102,15,219,224 ; pand %xmm0,%xmm4 + DB 102,15,127,69,232 ; movdqa %xmm0,-0x18(%ebp) + DB 102,15,111,133,8,255,255,255 ; movdqa -0xf8(%ebp),%xmm0 + DB 102,15,219,200 ; pand %xmm0,%xmm1 DB 102,15,223,69,232 ; pandn -0x18(%ebp),%xmm0 - DB 102,15,235,196 ; por %xmm4,%xmm0 - DB 102,15,127,133,56,255,255,255 ; movdqa %xmm0,-0xc8(%ebp) - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,111,141,152,254,255,255 ; movdqa -0x168(%ebp),%xmm1 - DB 102,15,219,209 ; pand %xmm1,%xmm2 - DB 102,15,223,77,184 ; pandn -0x48(%ebp),%xmm1 - DB 102,15,235,202 ; por %xmm2,%xmm1 - DB 102,15,253,219 ; paddw %xmm3,%xmm3 - DB 102,15,111,173,136,254,255,255 ; movdqa -0x178(%ebp),%xmm5 + DB 102,15,235,193 ; por %xmm1,%xmm0 + DB 102,15,111,141,120,255,255,255 ; movdqa -0x88(%ebp),%xmm1 + DB 102,15,253,201 ; paddw %xmm1,%xmm1 + DB 102,15,219,250 ; pand %xmm2,%xmm7 + DB 102,15,223,209 ; pandn %xmm1,%xmm2 + DB 102,15,235,215 ; por %xmm7,%xmm2 + DB 102,15,111,77,200 ; movdqa -0x38(%ebp),%xmm1 + DB 102,15,253,201 ; paddw %xmm1,%xmm1 DB 102,15,219,221 ; pand %xmm5,%xmm3 - DB 102,15,223,109,168 ; pandn -0x58(%ebp),%xmm5 + DB 102,15,223,233 ; pandn %xmm1,%xmm5 DB 102,15,235,235 ; por %xmm3,%xmm5 - DB 102,15,111,165,184,254,255,255 ; movdqa -0x148(%ebp),%xmm4 - DB 102,15,253,165,168,254,255,255 ; paddw -0x158(%ebp),%xmm4 - DB 102,15,253,189,120,255,255,255 ; paddw -0x88(%ebp),%xmm7 - DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 - DB 102,15,253,133,104,255,255,255 ; paddw -0x98(%ebp),%xmm0 - DB 102,15,253,117,136 ; paddw -0x78(%ebp),%xmm6 - DB 102,15,111,152,209,15,0,0 ; movdqa 0xfd1(%eax),%xmm3 - DB 102,15,253,227 ; paddw %xmm3,%xmm4 - DB 102,15,111,149,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm2 - DB 102,15,253,212 ; paddw %xmm4,%xmm2 - DB 102,15,253,251 ; paddw %xmm3,%xmm7 - DB 102,15,253,189,56,255,255,255 ; paddw -0xc8(%ebp),%xmm7 + DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 + DB 102,15,253,201 ; paddw %xmm1,%xmm1 + DB 102,15,219,244 ; pand %xmm4,%xmm6 + DB 102,15,223,225 ; pandn %xmm1,%xmm4 + DB 102,15,235,230 ; por %xmm6,%xmm4 + DB 102,15,111,136,167,15,0,0 ; movdqa 0xfa7(%eax),%xmm1 + DB 102,15,111,157,200,254,255,255 ; movdqa -0x138(%ebp),%xmm3 + DB 102,15,253,217 ; paddw %xmm1,%xmm3 DB 102,15,253,195 ; paddw %xmm3,%xmm0 - DB 102,15,253,200 ; paddw %xmm0,%xmm1 - DB 102,15,253,243 ; paddw %xmm3,%xmm6 - DB 102,15,253,245 ; paddw %xmm5,%xmm6 - DB 102,15,111,152,225,15,0,0 ; movdqa 0xfe1(%eax),%xmm3 - DB 102,15,228,243 ; pmulhuw %xmm3,%xmm6 - DB 102,15,228,203 ; pmulhuw %xmm3,%xmm1 - DB 102,15,228,251 ; pmulhuw %xmm3,%xmm7 - DB 102,15,228,211 ; pmulhuw %xmm3,%xmm2 - DB 102,15,113,215,7 ; psrlw $0x7,%xmm7 + DB 102,15,111,157,184,254,255,255 ; movdqa -0x148(%ebp),%xmm3 + DB 102,15,253,217 ; paddw %xmm1,%xmm3 + DB 102,15,253,218 ; paddw %xmm2,%xmm3 + DB 102,15,111,149,168,254,255,255 ; movdqa -0x158(%ebp),%xmm2 + DB 102,15,253,209 ; paddw %xmm1,%xmm2 + DB 102,15,253,234 ; paddw %xmm2,%xmm5 + DB 102,15,111,149,216,254,255,255 ; movdqa -0x128(%ebp),%xmm2 + DB 102,15,253,209 ; paddw %xmm1,%xmm2 + DB 102,15,253,212 ; paddw %xmm4,%xmm2 + DB 102,15,111,136,183,15,0,0 ; movdqa 0xfb7(%eax),%xmm1 + DB 102,15,228,209 ; pmulhuw %xmm1,%xmm2 DB 102,15,113,210,7 ; psrlw $0x7,%xmm2 - DB 102,15,103,215 ; packuswb %xmm7,%xmm2 - DB 102,15,113,214,7 ; psrlw $0x7,%xmm6 - DB 102,15,113,209,7 ; psrlw $0x7,%xmm1 - DB 102,15,103,206 ; packuswb %xmm6,%xmm1 + DB 102,15,228,233 ; pmulhuw %xmm1,%xmm5 + DB 102,15,113,213,7 ; psrlw $0x7,%xmm5 + DB 102,15,103,234 ; packuswb %xmm2,%xmm5 + DB 102,15,228,217 ; pmulhuw %xmm1,%xmm3 + DB 102,15,228,193 ; pmulhuw %xmm1,%xmm0 + DB 102,15,113,211,7 ; psrlw $0x7,%xmm3 + DB 102,15,113,208,7 ; psrlw $0x7,%xmm0 + DB 102,15,103,195 ; packuswb %xmm3,%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 15,40,149,88,254,255,255 ; movaps -0x1a8(%ebp),%xmm2 - DB 15,40,157,104,254,255,255 ; movaps -0x198(%ebp),%xmm3 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 15,40,149,136,254,255,255 ; movaps -0x178(%ebp),%xmm2 + DB 15,40,157,152,254,255,255 ; movaps -0x168(%ebp),%xmm3 DB 255,117,20 ; pushl 0x14(%ebp) DB 255,117,16 ; pushl 0x10(%ebp) DB 81 ; push %ecx DB 255,117,8 ; pushl 0x8(%ebp) DB 255,16 ; call *(%eax) - DB 129,196,184,1,0,0 ; add $0x1b8,%esp + DB 129,196,136,1,0,0 ; add $0x188,%esp DB 93 ; pop %ebp DB 195 ; ret @@ -60169,261 +60063,258 @@ PUBLIC _sk_overlay_sse2_8bit _sk_overlay_sse2_8bit LABEL PROC DB 85 ; push %ebp DB 137,229 ; mov %esp,%ebp - DB 129,236,184,1,0,0 ; sub $0x1b8,%esp - DB 102,15,239,246 ; pxor %xmm6,%xmm6 + DB 129,236,120,1,0,0 ; sub $0x178,%esp + DB 102,15,239,237 ; pxor %xmm5,%xmm5 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 - DB 102,15,127,165,120,255,255,255 ; movdqa %xmm4,-0x88(%ebp) + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,15,127,165,104,255,255,255 ; movdqa %xmm4,-0x98(%ebp) DB 102,15,111,225 ; movdqa %xmm1,%xmm4 - DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 - DB 102,15,127,101,200 ; movdqa %xmm4,-0x38(%ebp) + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,15,127,101,184 ; movdqa %xmm4,-0x48(%ebp) DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 - DB 102,15,127,133,104,255,255,255 ; movdqa %xmm0,-0x98(%ebp) + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,133,248,254,255,255 ; movdqa %xmm0,-0x108(%ebp) DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 - DB 243,15,112,236,95 ; pshufhw $0x5f,%xmm4,%xmm5 + DB 243,15,112,196,95 ; pshufhw $0x5f,%xmm4,%xmm0 DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 - DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 - DB 102,15,127,141,104,254,255,255 ; movdqa %xmm1,-0x198(%ebp) + DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 + DB 102,15,127,141,72,255,255,255 ; movdqa %xmm1,-0xb8(%ebp) DB 243,15,112,204,231 ; pshufhw $0xe7,%xmm4,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,193,95 ; pshufhw $0x5f,%xmm1,%xmm0 + DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 + DB 102,15,111,204 ; movdqa %xmm4,%xmm1 + DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 + DB 102,15,127,77,152 ; movdqa %xmm1,-0x68(%ebp) + DB 102,15,96,229 ; punpcklbw %xmm5,%xmm4 + DB 102,15,127,101,168 ; movdqa %xmm4,-0x58(%ebp) DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 - DB 102,15,127,141,120,254,255,255 ; movdqa %xmm1,-0x188(%ebp) - DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 - DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) - DB 102,15,111,197 ; movdqa %xmm5,%xmm0 - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 - DB 102,15,127,69,184 ; movdqa %xmm0,-0x48(%ebp) - DB 102,15,96,238 ; punpcklbw %xmm6,%xmm5 - DB 102,15,127,173,40,255,255,255 ; movdqa %xmm5,-0xd8(%ebp) - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,96,254 ; punpcklbw %xmm6,%xmm7 - DB 102,15,111,194 ; movdqa %xmm2,%xmm0 - DB 102,15,127,149,72,254,255,255 ; movdqa %xmm2,-0x1b8(%ebp) - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) + DB 102,15,96,205 ; punpcklbw %xmm5,%xmm1 + DB 102,15,127,77,232 ; movdqa %xmm1,-0x18(%ebp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,96,197 ; punpcklbw %xmm5,%xmm0 + DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 + DB 102,15,127,149,136,254,255,255 ; movdqa %xmm2,-0x178(%ebp) + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,133,24,255,255,255 ; movdqa %xmm0,-0xe8(%ebp) DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 - DB 102,15,127,69,232 ; movdqa %xmm0,-0x18(%ebp) + DB 102,15,96,197 ; punpcklbw %xmm5,%xmm0 + DB 102,15,127,133,40,255,255,255 ; movdqa %xmm0,-0xd8(%ebp) DB 102,15,111,195 ; movdqa %xmm3,%xmm0 - DB 102,15,127,157,88,254,255,255 ; movdqa %xmm3,-0x1a8(%ebp) - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 - DB 102,15,127,133,24,255,255,255 ; movdqa %xmm0,-0xe8(%ebp) + DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,15,127,165,152,254,255,255 ; movdqa %xmm4,-0x168(%ebp) + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,133,56,255,255,255 ; movdqa %xmm0,-0xc8(%ebp) DB 242,15,112,202,231 ; pshuflw $0xe7,%xmm2,%xmm1 DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,225,95 ; pshufhw $0x5f,%xmm1,%xmm4 - DB 242,15,112,203,231 ; pshuflw $0xe7,%xmm3,%xmm1 + DB 243,15,112,209,95 ; pshufhw $0x5f,%xmm1,%xmm2 + DB 242,15,112,204,231 ; pshuflw $0xe7,%xmm4,%xmm1 DB 243,15,112,201,231 ; pshufhw $0xe7,%xmm1,%xmm1 DB 102,15,112,201,232 ; pshufd $0xe8,%xmm1,%xmm1 DB 102,15,96,201 ; punpcklbw %xmm1,%xmm1 DB 242,15,112,201,95 ; pshuflw $0x5f,%xmm1,%xmm1 - DB 243,15,112,193,95 ; pshufhw $0x5f,%xmm1,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,104,206 ; punpckhbw %xmm6,%xmm1 - DB 102,15,127,77,152 ; movdqa %xmm1,-0x68(%ebp) - DB 102,15,96,198 ; punpcklbw %xmm6,%xmm0 - DB 102,15,127,69,136 ; movdqa %xmm0,-0x78(%ebp) - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,104,198 ; punpckhbw %xmm6,%xmm0 - DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 243,15,112,201,95 ; pshufhw $0x5f,%xmm1,%xmm1 + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,15,104,197 ; punpckhbw %xmm5,%xmm0 + DB 102,15,127,133,232,254,255,255 ; movdqa %xmm0,-0x118(%ebp) + DB 102,15,96,205 ; punpcklbw %xmm5,%xmm1 + DB 102,15,111,217 ; movdqa %xmm1,%xmm3 DB 102,15,127,157,88,255,255,255 ; movdqa %xmm3,-0xa8(%ebp) - DB 102,15,96,230 ; punpcklbw %xmm6,%xmm4 - DB 102,15,111,239 ; movdqa %xmm7,%xmm5 - DB 102,15,253,237 ; paddw %xmm5,%xmm5 - DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 - DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,15,111,69,232 ; movdqa -0x18(%ebp),%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,149,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm2 - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,217,85,152 ; psubusw -0x68(%ebp),%xmm2 - DB 102,15,117,214 ; pcmpeqw %xmm6,%xmm2 - DB 102,15,127,149,136,254,255,255 ; movdqa %xmm2,-0x178(%ebp) - DB 102,15,217,69,136 ; psubusw -0x78(%ebp),%xmm0 - DB 102,15,117,198 ; pcmpeqw %xmm6,%xmm0 - DB 102,15,127,133,152,254,255,255 ; movdqa %xmm0,-0x168(%ebp) - DB 102,15,217,203 ; psubusw %xmm3,%xmm1 - DB 102,15,117,206 ; pcmpeqw %xmm6,%xmm1 - DB 102,15,127,141,56,255,255,255 ; movdqa %xmm1,-0xc8(%ebp) - DB 102,15,217,236 ; psubusw %xmm4,%xmm5 - DB 102,15,117,238 ; pcmpeqw %xmm6,%xmm5 - DB 102,15,127,173,72,255,255,255 ; movdqa %xmm5,-0xb8(%ebp) - DB 232,0,0,0,0 ; call ccd5 <_sk_overlay_sse2_8bit+0x1c5> - DB 88 ; pop %eax - DB 102,15,111,152,123,10,0,0 ; movdqa 0xa7b(%eax),%xmm3 - DB 102,15,111,196 ; movdqa %xmm4,%xmm0 - DB 102,15,239,195 ; pxor %xmm3,%xmm0 - DB 102,15,111,181,120,255,255,255 ; movdqa -0x88(%ebp),%xmm6 - DB 102,15,213,198 ; pmullw %xmm6,%xmm0 - DB 102,15,127,133,168,254,255,255 ; movdqa %xmm0,-0x158(%ebp) - DB 102,15,111,133,40,255,255,255 ; movdqa -0xd8(%ebp),%xmm0 - DB 102,15,111,208 ; movdqa %xmm0,%xmm2 - DB 102,15,239,211 ; pxor %xmm3,%xmm2 - DB 102,15,213,215 ; pmullw %xmm7,%xmm2 - DB 102,15,127,149,184,254,255,255 ; movdqa %xmm2,-0x148(%ebp) - DB 102,15,111,204 ; movdqa %xmm4,%xmm1 - DB 102,15,249,231 ; psubw %xmm7,%xmm4 - DB 102,15,127,165,8,255,255,255 ; movdqa %xmm4,-0xf8(%ebp) - DB 102,15,111,214 ; movdqa %xmm6,%xmm2 - DB 102,15,213,250 ; pmullw %xmm2,%xmm7 - DB 102,15,127,189,232,254,255,255 ; movdqa %xmm7,-0x118(%ebp) - DB 102,15,213,200 ; pmullw %xmm0,%xmm1 - DB 102,15,127,141,200,254,255,255 ; movdqa %xmm1,-0x138(%ebp) - DB 102,15,249,194 ; psubw %xmm2,%xmm0 - DB 102,15,127,133,40,255,255,255 ; movdqa %xmm0,-0xd8(%ebp) - DB 102,15,111,133,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,239,203 ; pxor %xmm3,%xmm1 - DB 102,15,111,173,104,255,255,255 ; movdqa -0x98(%ebp),%xmm5 - DB 102,15,213,205 ; pmullw %xmm5,%xmm1 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,104,205 ; punpckhbw %xmm5,%xmm1 DB 102,15,127,141,120,255,255,255 ; movdqa %xmm1,-0x88(%ebp) - DB 102,15,111,85,184 ; movdqa -0x48(%ebp),%xmm2 - DB 102,15,111,250 ; movdqa %xmm2,%xmm7 - DB 102,15,239,251 ; pxor %xmm3,%xmm7 - DB 102,15,111,101,216 ; movdqa -0x28(%ebp),%xmm4 - DB 102,15,213,252 ; pmullw %xmm4,%xmm7 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,249,196 ; psubw %xmm4,%xmm0 - DB 102,15,127,133,88,255,255,255 ; movdqa %xmm0,-0xa8(%ebp) - DB 102,15,213,229 ; pmullw %xmm5,%xmm4 - DB 102,15,127,165,248,254,255,255 ; movdqa %xmm4,-0x108(%ebp) - DB 102,15,213,202 ; pmullw %xmm2,%xmm1 - DB 102,15,127,77,216 ; movdqa %xmm1,-0x28(%ebp) - DB 102,15,249,213 ; psubw %xmm5,%xmm2 - DB 102,15,127,85,184 ; movdqa %xmm2,-0x48(%ebp) - DB 102,15,111,117,136 ; movdqa -0x78(%ebp),%xmm6 - DB 102,15,111,198 ; movdqa %xmm6,%xmm0 - DB 102,15,239,195 ; pxor %xmm3,%xmm0 - DB 102,15,111,85,200 ; movdqa -0x38(%ebp),%xmm2 - DB 102,15,213,194 ; pmullw %xmm2,%xmm0 - DB 102,15,127,133,104,255,255,255 ; movdqa %xmm0,-0x98(%ebp) + DB 102,15,96,213 ; punpcklbw %xmm5,%xmm2 + DB 102,15,127,85,136 ; movdqa %xmm2,-0x78(%ebp) + DB 232,0,0,0,0 ; call cbb4 <_sk_overlay_sse2_8bit+0x171> + DB 88 ; pop %eax + DB 102,15,111,176,188,10,0,0 ; movdqa 0xabc(%eax),%xmm6 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,15,239,238 ; pxor %xmm6,%xmm5 + DB 102,15,239,206 ; pxor %xmm6,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,239,214 ; pxor %xmm6,%xmm2 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,15,239,230 ; pxor %xmm6,%xmm4 + DB 102,15,213,165,72,255,255,255 ; pmullw -0xb8(%ebp),%xmm4 + DB 102,15,111,125,232 ; movdqa -0x18(%ebp),%xmm7 + DB 102,15,239,254 ; pxor %xmm6,%xmm7 + DB 102,15,111,93,216 ; movdqa -0x28(%ebp),%xmm3 + DB 102,15,239,222 ; pxor %xmm6,%xmm3 DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 - DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,239,235 ; pxor %xmm3,%xmm5 - DB 102,15,111,101,232 ; movdqa -0x18(%ebp),%xmm4 - DB 102,15,213,236 ; pmullw %xmm4,%xmm5 - DB 102,15,111,206 ; movdqa %xmm6,%xmm1 - DB 102,15,249,244 ; psubw %xmm4,%xmm6 - DB 102,15,127,117,136 ; movdqa %xmm6,-0x78(%ebp) - DB 102,15,213,226 ; pmullw %xmm2,%xmm4 + DB 102,15,239,198 ; pxor %xmm6,%xmm0 + DB 102,15,239,117,152 ; pxor -0x68(%ebp),%xmm6 + DB 102,15,213,181,56,255,255,255 ; pmullw -0xc8(%ebp),%xmm6 + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,15,127,181,168,254,255,255 ; movdqa %xmm6,-0x158(%ebp) + DB 102,15,213,85,184 ; pmullw -0x48(%ebp),%xmm2 + DB 102,15,111,165,40,255,255,255 ; movdqa -0xd8(%ebp),%xmm4 + DB 102,15,213,196 ; pmullw %xmm4,%xmm0 + DB 102,15,253,194 ; paddw %xmm2,%xmm0 + DB 102,15,127,133,184,254,255,255 ; movdqa %xmm0,-0x148(%ebp) + DB 102,15,111,133,248,254,255,255 ; movdqa -0x108(%ebp),%xmm0 DB 102,15,213,200 ; pmullw %xmm0,%xmm1 - DB 102,15,127,77,232 ; movdqa %xmm1,-0x18(%ebp) - DB 102,15,249,194 ; psubw %xmm2,%xmm0 - DB 102,15,127,69,168 ; movdqa %xmm0,-0x58(%ebp) - DB 102,15,111,77,152 ; movdqa -0x68(%ebp),%xmm1 - DB 102,15,111,193 ; movdqa %xmm1,%xmm0 - DB 102,15,239,195 ; pxor %xmm3,%xmm0 - DB 102,15,111,149,104,254,255,255 ; movdqa -0x198(%ebp),%xmm2 - DB 102,15,213,194 ; pmullw %xmm2,%xmm0 - DB 102,15,127,133,216,254,255,255 ; movdqa %xmm0,-0x128(%ebp) - DB 102,15,111,133,120,254,255,255 ; movdqa -0x188(%ebp),%xmm0 - DB 102,15,239,216 ; pxor %xmm0,%xmm3 - DB 102,15,111,181,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm6 - DB 102,15,213,222 ; pmullw %xmm6,%xmm3 - DB 102,15,127,77,200 ; movdqa %xmm1,-0x38(%ebp) - DB 102,15,249,206 ; psubw %xmm6,%xmm1 - DB 102,15,127,77,152 ; movdqa %xmm1,-0x68(%ebp) - DB 102,15,213,242 ; pmullw %xmm2,%xmm6 - DB 102,15,111,77,200 ; movdqa -0x38(%ebp),%xmm1 + DB 102,15,111,149,24,255,255,255 ; movdqa -0xe8(%ebp),%xmm2 + DB 102,15,213,218 ; pmullw %xmm2,%xmm3 + DB 102,15,253,217 ; paddw %xmm1,%xmm3 + DB 102,15,127,157,200,254,255,255 ; movdqa %xmm3,-0x138(%ebp) + DB 102,15,111,141,104,255,255,255 ; movdqa -0x98(%ebp),%xmm1 + DB 102,15,213,233 ; pmullw %xmm1,%xmm5 + DB 102,15,111,93,200 ; movdqa -0x38(%ebp),%xmm3 + DB 102,15,213,251 ; pmullw %xmm3,%xmm7 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,15,127,189,216,254,255,255 ; movdqa %xmm7,-0x128(%ebp) + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,15,253,237 ; paddw %xmm5,%xmm5 + DB 102,15,111,245 ; movdqa %xmm5,%xmm6 + DB 102,15,111,109,136 ; movdqa -0x78(%ebp),%xmm5 + DB 102,15,101,245 ; pcmpgtw %xmm5,%xmm6 + DB 102,15,127,181,8,255,255,255 ; movdqa %xmm6,-0xf8(%ebp) + DB 102,15,111,245 ; movdqa %xmm5,%xmm6 + DB 102,15,249,235 ; psubw %xmm3,%xmm5 + DB 102,15,213,217 ; pmullw %xmm1,%xmm3 + DB 102,15,127,93,200 ; movdqa %xmm3,-0x38(%ebp) + DB 102,15,111,93,232 ; movdqa -0x18(%ebp),%xmm3 + DB 102,15,213,243 ; pmullw %xmm3,%xmm6 + DB 102,15,127,117,136 ; movdqa %xmm6,-0x78(%ebp) + DB 102,15,249,217 ; psubw %xmm1,%xmm3 + DB 102,15,213,221 ; pmullw %xmm5,%xmm3 + DB 102,15,127,93,232 ; movdqa %xmm3,-0x18(%ebp) + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,111,233 ; movdqa %xmm1,%xmm5 + DB 102,15,253,237 ; paddw %xmm5,%xmm5 + DB 102,15,111,149,120,255,255,255 ; movdqa -0x88(%ebp),%xmm2 + DB 102,15,101,234 ; pcmpgtw %xmm2,%xmm5 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,249,209 ; psubw %xmm1,%xmm2 DB 102,15,213,200 ; pmullw %xmm0,%xmm1 - DB 102,15,249,194 ; psubw %xmm2,%xmm0 - DB 102,15,213,69,152 ; pmullw -0x68(%ebp),%xmm0 + DB 102,15,127,141,104,255,255,255 ; movdqa %xmm1,-0x98(%ebp) + DB 102,15,111,77,216 ; movdqa -0x28(%ebp),%xmm1 + DB 102,15,213,241 ; pmullw %xmm1,%xmm6 + DB 102,15,249,200 ; psubw %xmm0,%xmm1 + DB 102,15,213,202 ; pmullw %xmm2,%xmm1 + DB 102,15,127,77,216 ; movdqa %xmm1,-0x28(%ebp) + DB 102,15,111,196 ; movdqa %xmm4,%xmm0 + DB 102,15,111,216 ; movdqa %xmm0,%xmm3 + DB 102,15,253,219 ; paddw %xmm3,%xmm3 + DB 102,15,111,149,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm2 + DB 102,15,101,218 ; pcmpgtw %xmm2,%xmm3 + DB 102,15,111,226 ; movdqa %xmm2,%xmm4 + DB 102,15,249,208 ; psubw %xmm0,%xmm2 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 + DB 102,15,213,208 ; pmullw %xmm0,%xmm2 + DB 102,15,127,149,88,255,255,255 ; movdqa %xmm2,-0xa8(%ebp) DB 102,15,111,85,168 ; movdqa -0x58(%ebp),%xmm2 - DB 102,15,213,85,136 ; pmullw -0x78(%ebp),%xmm2 + DB 102,15,213,226 ; pmullw %xmm2,%xmm4 + DB 102,15,127,165,120,255,255,255 ; movdqa %xmm4,-0x88(%ebp) + DB 102,15,249,208 ; psubw %xmm0,%xmm2 + DB 102,15,213,209 ; pmullw %xmm1,%xmm2 DB 102,15,127,85,168 ; movdqa %xmm2,-0x58(%ebp) - DB 102,15,111,85,184 ; movdqa -0x48(%ebp),%xmm2 - DB 102,15,213,149,88,255,255,255 ; pmullw -0xa8(%ebp),%xmm2 - DB 102,15,127,85,184 ; movdqa %xmm2,-0x48(%ebp) - DB 102,15,111,149,40,255,255,255 ; movdqa -0xd8(%ebp),%xmm2 - DB 102,15,213,149,8,255,255,255 ; pmullw -0xf8(%ebp),%xmm2 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,249,200 ; psubw %xmm0,%xmm1 - DB 102,15,127,77,200 ; movdqa %xmm1,-0x38(%ebp) - DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 - DB 102,15,253,192 ; paddw %xmm0,%xmm0 - DB 102,15,111,77,232 ; movdqa -0x18(%ebp),%xmm1 - DB 102,15,249,200 ; psubw %xmm0,%xmm1 - DB 102,15,127,77,232 ; movdqa %xmm1,-0x18(%ebp) - DB 102,15,111,77,184 ; movdqa -0x48(%ebp),%xmm1 - DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,15,111,69,216 ; movdqa -0x28(%ebp),%xmm0 + DB 102,15,111,141,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm1 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,15,253,228 ; paddw %xmm4,%xmm4 + DB 102,15,111,133,232,254,255,255 ; movdqa -0x118(%ebp),%xmm0 + DB 102,15,101,224 ; pcmpgtw %xmm0,%xmm4 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 DB 102,15,249,193 ; psubw %xmm1,%xmm0 - DB 102,15,127,69,216 ; movdqa %xmm0,-0x28(%ebp) - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,111,141,200,254,255,255 ; movdqa -0x138(%ebp),%xmm1 - DB 102,15,249,202 ; psubw %xmm2,%xmm1 - DB 102,15,111,149,232,254,255,255 ; movdqa -0x118(%ebp),%xmm2 - DB 102,15,253,210 ; paddw %xmm2,%xmm2 - DB 102,15,111,133,72,255,255,255 ; movdqa -0xb8(%ebp),%xmm0 - DB 102,15,219,208 ; pand %xmm0,%xmm2 - DB 102,15,223,193 ; pandn %xmm1,%xmm0 - DB 102,15,235,194 ; por %xmm2,%xmm0 - DB 102,15,127,133,72,255,255,255 ; movdqa %xmm0,-0xb8(%ebp) - DB 102,15,111,141,248,254,255,255 ; movdqa -0x108(%ebp),%xmm1 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,15,213,141,72,255,255,255 ; pmullw -0xb8(%ebp),%xmm1 + DB 102,15,127,77,184 ; movdqa %xmm1,-0x48(%ebp) + DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 + DB 102,15,213,208 ; pmullw %xmm0,%xmm2 + DB 102,15,249,133,72,255,255,255 ; psubw -0xb8(%ebp),%xmm0 + DB 102,15,213,199 ; pmullw %xmm7,%xmm0 + DB 102,15,127,69,152 ; movdqa %xmm0,-0x68(%ebp) + DB 102,15,111,77,232 ; movdqa -0x18(%ebp),%xmm1 DB 102,15,253,201 ; paddw %xmm1,%xmm1 - DB 102,15,111,133,56,255,255,255 ; movdqa -0xc8(%ebp),%xmm0 + DB 102,15,127,77,232 ; movdqa %xmm1,-0x18(%ebp) + DB 102,15,111,77,136 ; movdqa -0x78(%ebp),%xmm1 + DB 102,15,249,77,232 ; psubw -0x18(%ebp),%xmm1 + DB 102,15,111,125,216 ; movdqa -0x28(%ebp),%xmm7 + DB 102,15,253,255 ; paddw %xmm7,%xmm7 + DB 102,15,127,125,216 ; movdqa %xmm7,-0x28(%ebp) + DB 102,15,249,117,216 ; psubw -0x28(%ebp),%xmm6 + DB 102,15,111,69,168 ; movdqa -0x58(%ebp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,111,189,120,255,255,255 ; movdqa -0x88(%ebp),%xmm7 + DB 102,15,249,248 ; psubw %xmm0,%xmm7 + DB 102,15,111,69,152 ; movdqa -0x68(%ebp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,249,208 ; psubw %xmm0,%xmm2 + DB 102,15,111,69,200 ; movdqa -0x38(%ebp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,127,69,200 ; movdqa %xmm0,-0x38(%ebp) + DB 102,15,111,133,8,255,255,255 ; movdqa -0xf8(%ebp),%xmm0 DB 102,15,219,200 ; pand %xmm0,%xmm1 - DB 102,15,223,69,216 ; pandn -0x28(%ebp),%xmm0 - DB 102,15,235,193 ; por %xmm1,%xmm0 - DB 102,15,127,133,56,255,255,255 ; movdqa %xmm0,-0xc8(%ebp) - DB 102,15,253,228 ; paddw %xmm4,%xmm4 - DB 102,15,111,133,152,254,255,255 ; movdqa -0x168(%ebp),%xmm0 - DB 102,15,219,224 ; pand %xmm0,%xmm4 - DB 102,15,223,69,232 ; pandn -0x18(%ebp),%xmm0 - DB 102,15,235,196 ; por %xmm4,%xmm0 - DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,15,253,246 ; paddw %xmm6,%xmm6 - DB 102,15,111,133,136,254,255,255 ; movdqa -0x178(%ebp),%xmm0 - DB 102,15,219,240 ; pand %xmm0,%xmm6 DB 102,15,223,69,200 ; pandn -0x38(%ebp),%xmm0 - DB 102,15,235,198 ; por %xmm6,%xmm0 - DB 102,15,111,224 ; movdqa %xmm0,%xmm4 - DB 102,15,111,133,184,254,255,255 ; movdqa -0x148(%ebp),%xmm0 - DB 102,15,253,133,168,254,255,255 ; paddw -0x158(%ebp),%xmm0 - DB 102,15,253,189,120,255,255,255 ; paddw -0x88(%ebp),%xmm7 - DB 102,15,253,173,104,255,255,255 ; paddw -0x98(%ebp),%xmm5 - DB 102,15,253,157,216,254,255,255 ; paddw -0x128(%ebp),%xmm3 - DB 102,15,111,144,171,10,0,0 ; movdqa 0xaab(%eax),%xmm2 - DB 102,15,253,194 ; paddw %xmm2,%xmm0 - DB 102,15,253,133,72,255,255,255 ; paddw -0xb8(%ebp),%xmm0 - DB 102,15,253,250 ; paddw %xmm2,%xmm7 - DB 102,15,253,189,56,255,255,255 ; paddw -0xc8(%ebp),%xmm7 - DB 102,15,253,234 ; paddw %xmm2,%xmm5 - DB 102,15,253,233 ; paddw %xmm1,%xmm5 - DB 102,15,253,218 ; paddw %xmm2,%xmm3 - DB 102,15,253,220 ; paddw %xmm4,%xmm3 - DB 102,15,111,144,187,10,0,0 ; movdqa 0xabb(%eax),%xmm2 - DB 102,15,228,218 ; pmulhuw %xmm2,%xmm3 - DB 102,15,228,234 ; pmulhuw %xmm2,%xmm5 - DB 102,15,228,250 ; pmulhuw %xmm2,%xmm7 - DB 102,15,228,194 ; pmulhuw %xmm2,%xmm0 - DB 102,15,113,215,7 ; psrlw $0x7,%xmm7 - DB 102,15,113,208,7 ; psrlw $0x7,%xmm0 - DB 102,15,103,199 ; packuswb %xmm7,%xmm0 + DB 102,15,235,200 ; por %xmm0,%xmm1 + DB 102,15,111,133,104,255,255,255 ; movdqa -0x98(%ebp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,219,245 ; pand %xmm5,%xmm6 + DB 102,15,223,232 ; pandn %xmm0,%xmm5 + DB 102,15,235,238 ; por %xmm6,%xmm5 + DB 102,15,111,133,88,255,255,255 ; movdqa -0xa8(%ebp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,111,247 ; movdqa %xmm7,%xmm6 + DB 102,15,219,243 ; pand %xmm3,%xmm6 + DB 102,15,223,216 ; pandn %xmm0,%xmm3 + DB 102,15,235,243 ; por %xmm3,%xmm6 + DB 102,15,111,254 ; movdqa %xmm6,%xmm7 + DB 102,15,111,69,184 ; movdqa -0x48(%ebp),%xmm0 + DB 102,15,253,192 ; paddw %xmm0,%xmm0 + DB 102,15,219,212 ; pand %xmm4,%xmm2 + DB 102,15,223,224 ; pandn %xmm0,%xmm4 + DB 102,15,235,226 ; por %xmm2,%xmm4 + DB 102,15,111,128,236,10,0,0 ; movdqa 0xaec(%eax),%xmm0 + DB 102,15,111,149,216,254,255,255 ; movdqa -0x128(%ebp),%xmm2 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,15,253,209 ; paddw %xmm1,%xmm2 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,15,111,149,200,254,255,255 ; movdqa -0x138(%ebp),%xmm2 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,15,253,213 ; paddw %xmm5,%xmm2 + DB 102,15,111,218 ; movdqa %xmm2,%xmm3 + DB 102,15,111,149,184,254,255,255 ; movdqa -0x148(%ebp),%xmm2 + DB 102,15,253,208 ; paddw %xmm0,%xmm2 + DB 102,15,253,215 ; paddw %xmm7,%xmm2 + DB 102,15,111,141,168,254,255,255 ; movdqa -0x158(%ebp),%xmm1 + DB 102,15,253,200 ; paddw %xmm0,%xmm1 + DB 102,15,253,204 ; paddw %xmm4,%xmm1 + DB 102,15,111,128,252,10,0,0 ; movdqa 0xafc(%eax),%xmm0 + DB 102,15,228,200 ; pmulhuw %xmm0,%xmm1 + DB 102,15,113,209,7 ; psrlw $0x7,%xmm1 + DB 102,15,228,208 ; pmulhuw %xmm0,%xmm2 + DB 102,15,113,210,7 ; psrlw $0x7,%xmm2 + DB 102,15,103,209 ; packuswb %xmm1,%xmm2 + DB 102,15,111,202 ; movdqa %xmm2,%xmm1 + DB 102,15,228,216 ; pmulhuw %xmm0,%xmm3 + DB 102,15,111,214 ; movdqa %xmm6,%xmm2 + DB 102,15,228,208 ; pmulhuw %xmm0,%xmm2 DB 102,15,113,211,7 ; psrlw $0x7,%xmm3 - DB 102,15,113,213,7 ; psrlw $0x7,%xmm5 - DB 102,15,103,235 ; packuswb %xmm3,%xmm5 + DB 102,15,113,210,7 ; psrlw $0x7,%xmm2 + DB 102,15,103,211 ; packuswb %xmm3,%xmm2 + DB 102,15,111,194 ; movdqa %xmm2,%xmm0 DB 139,69,12 ; mov 0xc(%ebp),%eax DB 141,72,4 ; lea 0x4(%eax),%ecx - DB 102,15,111,205 ; movdqa %xmm5,%xmm1 - DB 15,40,149,72,254,255,255 ; movaps -0x1b8(%ebp),%xmm2 - DB 15,40,157,88,254,255,255 ; movaps -0x1a8(%ebp),%xmm3 + DB 15,40,149,136,254,255,255 ; movaps -0x178(%ebp),%xmm2 + DB 15,40,157,152,254,255,255 ; movaps -0x168(%ebp),%xmm3 DB 255,117,20 ; pushl 0x14(%ebp) DB 255,117,16 ; pushl 0x10(%ebp) DB 81 ; push %ecx DB 255,117,8 ; pushl 0x8(%ebp) DB 255,16 ; call *(%eax) - DB 129,196,200,1,0,0 ; add $0x1c8,%esp + DB 129,196,136,1,0,0 ; add $0x188,%esp DB 93 ; pop %ebp DB 195 ; ret @@ -60497,11 +60388,11 @@ ALIGN 16 DB 0,128,191,0,0,128 ; add %al,-0x7fffff41(%eax) DB 191,0,0,224,64 ; mov $0x40e00000,%edi DB 0,0 ; add %al,(%eax) - DB 224,64 ; loopne d118 <.literal16+0xe8> + DB 224,64 ; loopne d008 <.literal16+0xe8> DB 0,0 ; add %al,(%eax) - DB 224,64 ; loopne d11c <.literal16+0xec> + DB 224,64 ; loopne d00c <.literal16+0xec> DB 0,0 ; add %al,(%eax) - DB 224,64 ; loopne d120 <.literal16+0xf0> + DB 224,64 ; loopne d010 <.literal16+0xf0> DB 154,153,153,62,154,153,153 ; lcall $0x9999,$0x9a3e9999 DB 62,154,153,153,62,154,153,153 ; ds lcall $0x9999,$0x9a3e9999 DB 62,61,10,23,63,61 ; ds cmp $0x3d3f170a,%eax @@ -60512,16 +60403,16 @@ ALIGN 16 DB 63 ; aas DB 174 ; scas %es:(%edi),%al DB 71 ; inc %edi - DB 225,61 ; loope d141 <.literal16+0x111> + DB 225,61 ; loope d031 <.literal16+0x111> DB 174 ; scas %es:(%edi),%al DB 71 ; inc %edi - DB 225,61 ; loope d145 <.literal16+0x115> + DB 225,61 ; loope d035 <.literal16+0x115> DB 174 ; scas %es:(%edi),%al DB 71 ; inc %edi - DB 225,61 ; loope d149 <.literal16+0x119> + DB 225,61 ; loope d039 <.literal16+0x119> DB 174 ; scas %es:(%edi),%al DB 71 ; inc %edi - DB 225,61 ; loope d14d <.literal16+0x11d> + DB 225,61 ; loope d03d <.literal16+0x11d> DB 255,0 ; incl (%eax) DB 0,0 ; add %al,(%eax) DB 255,0 ; incl (%eax) @@ -60531,13 +60422,13 @@ ALIGN 16 DB 255,0 ; incl (%eax) DB 0,0 ; add %al,(%eax) DB 0,0 ; add %al,(%eax) - DB 127,67 ; jg d167 <.literal16+0x137> + DB 127,67 ; jg d057 <.literal16+0x137> DB 0,0 ; add %al,(%eax) - DB 127,67 ; jg d16b <.literal16+0x13b> + DB 127,67 ; jg d05b <.literal16+0x13b> DB 0,0 ; add %al,(%eax) - DB 127,67 ; jg d16f <.literal16+0x13f> + DB 127,67 ; jg d05f <.literal16+0x13f> DB 0,0 ; add %al,(%eax) - DB 127,67 ; jg d173 <.literal16+0x143> + DB 127,67 ; jg d063 <.literal16+0x143> DB 145 ; xchg %eax,%ecx DB 131,158,61,145,131,158,61 ; sbbl $0x3d,-0x617c6ec3(%esi) DB 145 ; xchg %eax,%ecx @@ -60688,13 +60579,13 @@ ALIGN 16 DB 132,55 ; test %dh,(%edi) DB 8,33 ; or %ah,(%ecx) DB 132,55 ; test %dh,(%edi) - DB 224,7 ; loopne d279 <.literal16+0x249> + DB 224,7 ; loopne d169 <.literal16+0x249> DB 0,0 ; add %al,(%eax) - DB 224,7 ; loopne d27d <.literal16+0x24d> + DB 224,7 ; loopne d16d <.literal16+0x24d> DB 0,0 ; add %al,(%eax) - DB 224,7 ; loopne d281 <.literal16+0x251> + DB 224,7 ; loopne d171 <.literal16+0x251> DB 0,0 ; add %al,(%eax) - DB 224,7 ; loopne d285 <.literal16+0x255> + DB 224,7 ; loopne d175 <.literal16+0x255> DB 0,0 ; add %al,(%eax) DB 33,8 ; and %ecx,(%eax) DB 2,58 ; add (%edx),%bh @@ -60718,6 +60609,14 @@ ALIGN 16 DB 4,61 ; add $0x3d,%al DB 8,33 ; or %ah,(%ecx) DB 4,61 ; add $0x3d,%al + DB 255,0 ; incl (%eax) + DB 255,0 ; incl (%eax) + DB 255,0 ; incl (%eax) + DB 255,0 ; incl (%eax) + DB 0,0 ; add %al,(%eax) + DB 0,0 ; add %al,(%eax) + DB 0,0 ; add %al,(%eax) + DB 0,0 ; add %al,(%eax) DB 128,0,128 ; addb $0x80,(%eax) DB 55 ; aaa DB 128,0,128 ; addb $0x80,(%eax) @@ -60735,17 +60634,17 @@ ALIGN 16 DB 0,0 ; add %al,(%eax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg d2d4 <.literal16+0x2a4> + DB 127,0 ; jg d1d4 <.literal16+0x2b4> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg d2d8 <.literal16+0x2a8> + DB 127,0 ; jg d1d8 <.literal16+0x2b8> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg d2dc <.literal16+0x2ac> + DB 127,0 ; jg d1dc <.literal16+0x2bc> DB 255 ; (bad) DB 255 ; (bad) - DB 127,0 ; jg d2e0 <.literal16+0x2b0> - DB 119,115 ; ja d355 <.literal16+0x325> + DB 127,0 ; jg d1e0 <.literal16+0x2c0> + DB 119,115 ; ja d255 <.literal16+0x335> DB 248 ; clc DB 194,119,115 ; ret $0x7377 DB 248 ; clc @@ -60756,7 +60655,7 @@ ALIGN 16 DB 194,117,191 ; ret $0xbf75 DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; aas - DB 117,191 ; jne d2b9 <.literal16+0x289> + DB 117,191 ; jne d1b9 <.literal16+0x299> DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi DB 63 ; aas DB 249 ; stc @@ -60772,7 +60671,7 @@ ALIGN 16 DB 68 ; inc %esp DB 180,62 ; mov $0x3e,%ah DB 163,233,220,63,163 ; mov %eax,0xa33fdce9 - DB 233,220,63,163,233 ; jmp e9a412f6 <_sk_overlay_sse2_8bit+0xe9a347e6> + DB 233,220,63,163,233 ; jmp e9a411f6 <_sk_overlay_sse2_8bit+0xe9a347b3> DB 220,63 ; fdivrl (%edi) DB 163,233,220,63,81 ; mov %eax,0x513fdce9 DB 140,242 ; mov %?,%edx @@ -60868,16 +60767,16 @@ ALIGN 16 DB 128,3,62 ; addb $0x3e,(%ebx) DB 31 ; pop %ds DB 215 ; xlat %ds:(%ebx) - DB 118,63 ; jbe d443 <.literal16+0x413> + DB 118,63 ; jbe d343 <.literal16+0x423> DB 31 ; pop %ds DB 215 ; xlat %ds:(%ebx) - DB 118,63 ; jbe d447 <.literal16+0x417> + DB 118,63 ; jbe d347 <.literal16+0x427> DB 31 ; pop %ds DB 215 ; xlat %ds:(%ebx) - DB 118,63 ; jbe d44b <.literal16+0x41b> + DB 118,63 ; jbe d34b <.literal16+0x42b> DB 31 ; pop %ds DB 215 ; xlat %ds:(%ebx) - DB 118,63 ; jbe d44f <.literal16+0x41f> + DB 118,63 ; jbe d34f <.literal16+0x42f> DB 246,64,83,63 ; testb $0x3f,0x53(%eax) DB 246,64,83,63 ; testb $0x3f,0x53(%eax) DB 246,64,83,63 ; testb $0x3f,0x53(%eax) @@ -60903,13 +60802,13 @@ ALIGN 16 DB 248 ; clc DB 65 ; inc %ecx DB 0,0 ; add %al,(%eax) - DB 124,66 ; jl d486 <.literal16+0x456> + DB 124,66 ; jl d386 <.literal16+0x466> DB 0,0 ; add %al,(%eax) - DB 124,66 ; jl d48a <.literal16+0x45a> + DB 124,66 ; jl d38a <.literal16+0x46a> DB 0,0 ; add %al,(%eax) - DB 124,66 ; jl d48e <.literal16+0x45e> + DB 124,66 ; jl d38e <.literal16+0x46e> DB 0,0 ; add %al,(%eax) - DB 124,66 ; jl d492 <.literal16+0x462> + DB 124,66 ; jl d392 <.literal16+0x472> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%eax) DB 0,240 ; add %dh,%al @@ -60955,56 +60854,77 @@ ALIGN 16 DB 136,136,61,137,136,136 ; mov %cl,-0x777776c3(%eax) DB 61,137,136,136,61 ; cmp $0x3d888889,%eax DB 0,0 ; add %al,(%eax) - DB 112,65 ; jo d515 <.literal16+0x4e5> + DB 112,65 ; jo d415 <.literal16+0x4f5> DB 0,0 ; add %al,(%eax) - DB 112,65 ; jo d519 <.literal16+0x4e9> + DB 112,65 ; jo d419 <.literal16+0x4f9> DB 0,0 ; add %al,(%eax) - DB 112,65 ; jo d51d <.literal16+0x4ed> + DB 112,65 ; jo d41d <.literal16+0x4fd> DB 0,0 ; add %al,(%eax) - DB 112,65 ; jo d521 <.literal16+0x4f1> + DB 112,65 ; jo d421 <.literal16+0x501> DB 0,128,0,0,0,128 ; add %al,-0x80000000(%eax) DB 0,0 ; add %al,(%eax) DB 0,128,0,0,0,128 ; add %al,-0x80000000(%eax) DB 0,0 ; add %al,(%eax) + DB 255 ; (bad) + DB 127,0 ; jg d3f3 <.literal16+0x4d3> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg d3f7 <.literal16+0x4d7> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg d3fb <.literal16+0x4db> + DB 0,255 ; add %bh,%bh + DB 127,0 ; jg d3ff <.literal16+0x4df> DB 0,0 ; add %al,(%eax) - DB 0,56 ; add %bh,(%eax) + DB 4,0 ; add $0x0,%al DB 0,0 ; add %al,(%eax) - DB 0,56 ; add %bh,(%eax) + DB 4,0 ; add $0x0,%al DB 0,0 ; add %al,(%eax) - DB 0,56 ; add %bh,(%eax) + DB 4,0 ; add $0x0,%al DB 0,0 ; add %al,(%eax) - DB 0,56 ; add %bh,(%eax) + DB 4,0 ; add $0x0,%al DB 0,0 ; add %al,(%eax) - DB 0,128,0,0,0,128 ; add %al,-0x80000000(%eax) DB 0,0 ; add %al,(%eax) - DB 0,128,0,0,0,128 ; add %al,-0x80000000(%eax) - DB 0,4,0 ; add %al,(%eax,%eax,1) - DB 128,0,4 ; addb $0x4,(%eax) - DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%eax) - DB 0,4,0 ; add %al,(%eax,%eax,1) + DB 56,0 ; cmp %al,(%eax) + DB 0,0 ; add %al,(%eax) + DB 56,0 ; cmp %al,(%eax) + DB 0,0 ; add %al,(%eax) + DB 56,0 ; cmp %al,(%eax) + DB 0,0 ; add %al,(%eax) + DB 56,0 ; cmp %al,(%eax) + DB 0,0 ; add %al,(%eax) DB 128,0,0 ; addb $0x0,(%eax) - DB 128,56,0 ; cmpb $0x0,(%eax) + DB 0,128,0,0,0,128 ; add %al,-0x80000000(%eax) + DB 0,0 ; add %al,(%eax) + DB 0,128,255,255,255,127 ; add %al,0x7fffffff(%eax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,255 ; jg d438 <.literal16+0x518> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,255 ; jg d43c <.literal16+0x51c> + DB 255 ; (bad) + DB 255 ; (bad) + DB 127,0 ; jg d441 <.literal16+0x521> DB 0,128,56,0,0,128 ; add %al,-0x7fffffc8(%eax) DB 56,0 ; cmp %al,(%eax) - DB 0,128,56,0,64,254 ; add %al,-0x1bfffc8(%eax) + DB 0,128,56,0,0,128 ; add %al,-0x7fffffc8(%eax) + DB 56,0 ; cmp %al,(%eax) + DB 0,255 ; add %bh,%bh DB 255,0 ; incl (%eax) - DB 64 ; inc %eax - DB 254 ; (bad) + DB 0,255 ; add %bh,%bh DB 255,0 ; incl (%eax) - DB 64 ; inc %eax - DB 254 ; (bad) + DB 0,255 ; add %bh,%bh DB 255,0 ; incl (%eax) - DB 64 ; inc %eax - DB 254 ; (bad) + DB 0,255 ; add %bh,%bh DB 255,0 ; incl (%eax) DB 255 ; (bad) - DB 127,71 ; jg d58b <.literal16+0x55b> + DB 127,71 ; jg d4ab <.literal16+0x58b> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg d58f <.literal16+0x55f> + DB 127,71 ; jg d4af <.literal16+0x58f> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg d593 <.literal16+0x563> + DB 127,71 ; jg d4b3 <.literal16+0x593> DB 0,255 ; add %bh,%bh - DB 127,71 ; jg d597 <.literal16+0x567> + DB 127,71 ; jg d4b7 <.literal16+0x597> DB 208 ; (bad) DB 179,89 ; mov $0x59,%bl DB 62,208 ; ds (bad) @@ -61057,7 +60977,7 @@ ALIGN 16 DB 192,45,16,17,192,45,16 ; shrb $0x10,0x2dc01110 DB 17,192 ; adc %eax,%eax DB 45,16,17,192,18 ; sub $0x12c01110,%eax - DB 120,57 ; js d5fc <.literal16+0x5cc> + DB 120,57 ; js d51c <.literal16+0x5fc> DB 64 ; inc %eax DB 18,120,57 ; adc 0x39(%eax),%bh DB 64 ; inc %eax @@ -61128,13 +61048,13 @@ ALIGN 16 DB 192,191,0,0,192,191,0 ; sarb $0x0,-0x40400000(%edi) DB 0,192 ; add %al,%al DB 191,0,0,192,191 ; mov $0xbfc00000,%edi - DB 114,28 ; jb d6ae <.literal16+0x67e> + DB 114,28 ; jb d5ce <.literal16+0x6ae> DB 199 ; (bad) - DB 62,114,28 ; jb,pt d6b2 <.literal16+0x682> + DB 62,114,28 ; jb,pt d5d2 <.literal16+0x6b2> DB 199 ; (bad) - DB 62,114,28 ; jb,pt d6b6 <.literal16+0x686> + DB 62,114,28 ; jb,pt d5d6 <.literal16+0x6b6> DB 199 ; (bad) - DB 62,114,28 ; jb,pt d6ba <.literal16+0x68a> + DB 62,114,28 ; jb,pt d5da <.literal16+0x6ba> DB 199 ; (bad) DB 62,85 ; ds push %ebp DB 85 ; push %ebp @@ -61155,13 +61075,13 @@ ALIGN 16 DB 57,142,99,61,57,142 ; cmp %ecx,-0x71c6c29d(%esi) DB 99,61,57,142,99,61 ; arpl %di,0x3d638e39 DB 57,142,99,61,114,249 ; cmp %ecx,-0x68dc29d(%esi) - DB 127,63 ; jg d713 <.literal16+0x6e3> - DB 114,249 ; jb d6cf <.literal16+0x69f> - DB 127,63 ; jg d717 <.literal16+0x6e7> - DB 114,249 ; jb d6d3 <.literal16+0x6a3> - DB 127,63 ; jg d71b <.literal16+0x6eb> - DB 114,249 ; jb d6d7 <.literal16+0x6a7> - DB 127,63 ; jg d71f <.literal16+0x6ef> + DB 127,63 ; jg d633 <.literal16+0x713> + DB 114,249 ; jb d5ef <.literal16+0x6cf> + DB 127,63 ; jg d637 <.literal16+0x717> + DB 114,249 ; jb d5f3 <.literal16+0x6d3> + DB 127,63 ; jg d63b <.literal16+0x71b> + DB 114,249 ; jb d5f7 <.literal16+0x6d7> + DB 127,63 ; jg d63f <.literal16+0x71f> DB 3,0 ; add (%eax),%eax DB 0,0 ; add %al,(%eax) DB 3,0 ; add (%eax),%eax @@ -61247,14 +61167,14 @@ ALIGN 16 DB 255 ; (bad) DB 255 ; (bad) DB 255,0 ; incl (%eax) - DB 127,0 ; jg d782 <.literal16+0x752> - DB 127,0 ; jg d784 <.literal16+0x754> - DB 127,0 ; jg d786 <.literal16+0x756> - DB 127,0 ; jg d788 <.literal16+0x758> - DB 127,0 ; jg d78a <.literal16+0x75a> - DB 127,0 ; jg d78c <.literal16+0x75c> - DB 127,0 ; jg d78e <.literal16+0x75e> - DB 127,0 ; jg d790 <.literal16+0x760> + DB 127,0 ; jg d6a2 <.literal16+0x782> + DB 127,0 ; jg d6a4 <.literal16+0x784> + DB 127,0 ; jg d6a6 <.literal16+0x786> + DB 127,0 ; jg d6a8 <.literal16+0x788> + DB 127,0 ; jg d6aa <.literal16+0x78a> + DB 127,0 ; jg d6ac <.literal16+0x78c> + DB 127,0 ; jg d6ae <.literal16+0x78e> + DB 127,0 ; jg d6b0 <.literal16+0x790> DB 129,128,129,128,129,128,129,128,129,128; addl $0x80818081,-0x7f7e7f7f(%eax) DB 129 ; .byte 0x81 DB 128 ; .byte 0x80 @@ -61267,7 +61187,7 @@ ALIGN 4 DB 0,0 ; add %al,(%eax) DB 0,63 ; add %bh,(%edi) DB 0,0 ; add %al,(%eax) - DB 127,67 ; jg d7eb <_sk_overlay_sse2_8bit+0xcdb> + DB 127,67 ; jg d70b <_sk_overlay_sse2_8bit+0xcc8> DB 0,0 ; add %al,(%eax) DB 0 ; .byte 0x0 DB 79 ; dec %edi |