diff options
Diffstat (limited to 'src/jumper/SkJumper_generated_win.S')
-rw-r--r-- | src/jumper/SkJumper_generated_win.S | 748 |
1 files changed, 610 insertions, 138 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index c4d144d474..25da2f0884 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -319,6 +319,106 @@ _sk_xor__hsw LABEL PROC DB 197,124,41,195 ; vmovaps %ymm8,%ymm3 DB 255,224 ; jmpq *%rax +PUBLIC _sk_darken_hsw +_sk_darken_hsw LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 + DB 197,100,89,204 ; vmulps %ymm4,%ymm3,%ymm9 + DB 196,193,124,95,193 ; vmaxps %ymm9,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 + DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 + DB 196,193,116,95,201 ; vmaxps %ymm9,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + DB 197,236,89,215 ; vmulps %ymm7,%ymm2,%ymm2 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 + 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 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_lighten_hsw +_sk_lighten_hsw LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 + DB 197,100,89,204 ; vmulps %ymm4,%ymm3,%ymm9 + DB 196,193,124,93,193 ; vminps %ymm9,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 + DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 + DB 196,193,116,93,201 ; vminps %ymm9,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + DB 197,236,89,215 ; vmulps %ymm7,%ymm2,%ymm2 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 + 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 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_difference_hsw +_sk_difference_hsw LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 + DB 197,100,89,204 ; vmulps %ymm4,%ymm3,%ymm9 + DB 196,193,124,93,193 ; vminps %ymm9,%ymm0,%ymm0 + DB 197,252,88,192 ; vaddps %ymm0,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 + DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 + DB 196,193,116,93,201 ; vminps %ymm9,%ymm1,%ymm1 + DB 197,244,88,201 ; vaddps %ymm1,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + DB 197,236,89,215 ; vmulps %ymm7,%ymm2,%ymm2 + DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 + 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 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_hsw +_sk_exclusion_hsw LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,196 ; vmulps %ymm4,%ymm0,%ymm0 + DB 197,252,88,192 ; vaddps %ymm0,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,205 ; vmulps %ymm5,%ymm1,%ymm1 + DB 197,244,88,201 ; vaddps %ymm1,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8 + 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 + DB 255,224 ; jmpq *%rax + PUBLIC _sk_clamp_0_hsw _sk_clamp_0_hsw LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 @@ -537,7 +637,7 @@ _sk_scale_u8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,56 ; jne 70e <_sk_scale_u8_hsw+0x48> + DB 117,56 ; jne 873 <_sk_scale_u8_hsw+0x48> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 @@ -561,9 +661,9 @@ _sk_scale_u8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 716 <_sk_scale_u8_hsw+0x50> + DB 117,234 ; jne 87b <_sk_scale_u8_hsw+0x50> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,167 ; jmp 6da <_sk_scale_u8_hsw+0x14> + DB 235,167 ; jmp 83f <_sk_scale_u8_hsw+0x14> PUBLIC _sk_lerp_1_float_hsw _sk_lerp_1_float_hsw LABEL PROC @@ -587,7 +687,7 @@ _sk_lerp_u8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,76 ; jne 7be <_sk_lerp_u8_hsw+0x5c> + DB 117,76 ; jne 923 <_sk_lerp_u8_hsw+0x5c> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,125,49,192 ; vpmovzxbd %xmm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 @@ -615,16 +715,16 @@ _sk_lerp_u8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 7c6 <_sk_lerp_u8_hsw+0x64> + DB 117,234 ; jne 92b <_sk_lerp_u8_hsw+0x64> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,147 ; jmp 776 <_sk_lerp_u8_hsw+0x14> + DB 235,147 ; jmp 8db <_sk_lerp_u8_hsw+0x14> PUBLIC _sk_lerp_565_hsw _sk_lerp_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,179,0,0,0 ; jne 8a4 <_sk_lerp_565_hsw+0xc1> + DB 15,133,179,0,0,0 ; jne a09 <_sk_lerp_565_hsw+0xc1> DB 196,193,122,111,28,122 ; vmovdqu (%r10,%rdi,2),%xmm3 DB 196,98,125,51,195 ; vpmovzxwd %xmm3,%ymm8 DB 184,0,248,0,0 ; mov $0xf800,%eax @@ -670,9 +770,9 @@ _sk_lerp_565_hsw LABEL PROC DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,59,255,255,255 ; ja 7f7 <_sk_lerp_565_hsw+0x14> + DB 15,135,59,255,255,255 ; ja 95c <_sk_lerp_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 910 <_sk_lerp_565_hsw+0x12d> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # a78 <_sk_lerp_565_hsw+0x130> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -684,27 +784,26 @@ _sk_lerp_565_hsw LABEL PROC DB 196,193,97,196,92,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3 DB 196,193,97,196,92,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3 DB 196,193,97,196,28,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3 - DB 233,231,254,255,255 ; jmpq 7f7 <_sk_lerp_565_hsw+0x14> - DB 244 ; hlt - DB 255 ; (bad) + DB 233,231,254,255,255 ; jmpq 95c <_sk_lerp_565_hsw+0x14> + DB 15,31,0 ; nopl (%rax) + DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) - DB 236 ; in (%dx),%al DB 255 ; (bad) + DB 233,255,255,255,225 ; jmpq ffffffffe2000a80 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff4f0> DB 255 ; (bad) - DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) + DB 217,255 ; fcos DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) + DB 255,209 ; callq *%rcx DB 255 ; (bad) - DB 255,212 ; callq *%rsp DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) - DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax + DB 189 ; .byte 0xbd DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -716,7 +815,7 @@ _sk_load_tables_hsw LABEL PROC DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 - DB 117,121 ; jne 9ba <_sk_load_tables_hsw+0x8e> + DB 117,121 ; jne b22 <_sk_load_tables_hsw+0x8e> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 DB 185,255,0,0,0 ; mov $0xff,%ecx DB 197,249,110,193 ; vmovd %ecx,%xmm0 @@ -752,7 +851,7 @@ _sk_load_tables_hsw LABEL PROC DB 196,193,249,110,194 ; vmovq %r10,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3 - DB 233,99,255,255,255 ; jmpq 946 <_sk_load_tables_hsw+0x1a> + DB 233,99,255,255,255 ; jmpq aae <_sk_load_tables_hsw+0x1a> PUBLIC _sk_load_a8_hsw _sk_load_a8_hsw LABEL PROC @@ -761,7 +860,7 @@ _sk_load_a8_hsw LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,50 ; jne a25 <_sk_load_a8_hsw+0x42> + DB 117,50 ; jne b8d <_sk_load_a8_hsw+0x42> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 @@ -784,9 +883,9 @@ _sk_load_a8_hsw LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne a2d <_sk_load_a8_hsw+0x4a> + DB 117,234 ; jne b95 <_sk_load_a8_hsw+0x4a> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,173 ; jmp 9f7 <_sk_load_a8_hsw+0x14> + DB 235,173 ; jmp b5f <_sk_load_a8_hsw+0x14> PUBLIC _sk_store_a8_hsw _sk_store_a8_hsw LABEL PROC @@ -801,7 +900,7 @@ _sk_store_a8_hsw LABEL PROC DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne a85 <_sk_store_a8_hsw+0x3b> + DB 117,10 ; jne bed <_sk_store_a8_hsw+0x3b> DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -809,10 +908,10 @@ _sk_store_a8_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja a81 <_sk_store_a8_hsw+0x37> + DB 119,236 ; ja be9 <_sk_store_a8_hsw+0x37> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # ae8 <_sk_store_a8_hsw+0x9e> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # c50 <_sk_store_a8_hsw+0x9e> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -823,7 +922,7 @@ _sk_store_a8_hsw LABEL PROC DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,154 ; jmp a81 <_sk_store_a8_hsw+0x37> + DB 235,154 ; jmp be9 <_sk_store_a8_hsw+0x37> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -853,7 +952,7 @@ _sk_load_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,149,0,0,0 ; jne ba7 <_sk_load_565_hsw+0xa3> + DB 15,133,149,0,0,0 ; jne d0f <_sk_load_565_hsw+0xa3> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 DB 184,0,248,0,0 ; mov $0xf800,%eax @@ -893,9 +992,9 @@ _sk_load_565_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,89,255,255,255 ; ja b18 <_sk_load_565_hsw+0x14> + DB 15,135,89,255,255,255 ; ja c80 <_sk_load_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # c14 <_sk_load_565_hsw+0x110> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # d7c <_sk_load_565_hsw+0x110> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -907,12 +1006,12 @@ _sk_load_565_hsw LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,5,255,255,255 ; jmpq b18 <_sk_load_565_hsw+0x14> + DB 233,5,255,255,255 ; jmpq c80 <_sk_load_565_hsw+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp c19 <_sk_load_565_hsw+0x115> + DB 235,255 ; jmp d81 <_sk_load_565_hsw+0x115> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -955,7 +1054,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne c9c <_sk_store_565_hsw+0x6c> + DB 117,10 ; jne e04 <_sk_store_565_hsw+0x6c> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -963,9 +1062,9 @@ _sk_store_565_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja c98 <_sk_store_565_hsw+0x68> + DB 119,236 ; ja e00 <_sk_store_565_hsw+0x68> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # cfc <_sk_store_565_hsw+0xcc> + DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # e64 <_sk_store_565_hsw+0xcc> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -976,7 +1075,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp c98 <_sk_store_565_hsw+0x68> + DB 235,159 ; jmp e00 <_sk_store_565_hsw+0x68> DB 15,31,0 ; nopl (%rax) DB 244 ; hlt DB 255 ; (bad) @@ -1009,7 +1108,7 @@ _sk_load_8888_hsw LABEL PROC DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 - DB 117,104 ; jne d95 <_sk_load_8888_hsw+0x7d> + DB 117,104 ; jne efd <_sk_load_8888_hsw+0x7d> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -1042,7 +1141,7 @@ _sk_load_8888_hsw LABEL PROC DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3 - DB 233,116,255,255,255 ; jmpq d32 <_sk_load_8888_hsw+0x1a> + DB 233,116,255,255,255 ; jmpq e9a <_sk_load_8888_hsw+0x1a> PUBLIC _sk_store_8888_hsw _sk_store_8888_hsw LABEL PROC @@ -1068,7 +1167,7 @@ _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 77,133,192 ; test %r8,%r8 - DB 117,12 ; jne e32 <_sk_store_8888_hsw+0x74> + DB 117,12 ; jne f9a <_sk_store_8888_hsw+0x74> DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx @@ -1081,14 +1180,14 @@ _sk_store_8888_hsw LABEL PROC DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9 DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9) - DB 235,211 ; jmp e2b <_sk_store_8888_hsw+0x6d> + DB 235,211 ; jmp f93 <_sk_store_8888_hsw+0x6d> PUBLIC _sk_load_f16_hsw _sk_load_f16_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 117,97 ; jne ec3 <_sk_load_f16_hsw+0x6b> + DB 117,97 ; jne 102b <_sk_load_f16_hsw+0x6b> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -1114,29 +1213,29 @@ _sk_load_f16_hsw LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je f22 <_sk_load_f16_hsw+0xca> + DB 116,79 ; je 108a <_sk_load_f16_hsw+0xca> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb f22 <_sk_load_f16_hsw+0xca> + DB 114,67 ; jb 108a <_sk_load_f16_hsw+0xca> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je f2f <_sk_load_f16_hsw+0xd7> + DB 116,68 ; je 1097 <_sk_load_f16_hsw+0xd7> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb f2f <_sk_load_f16_hsw+0xd7> + DB 114,56 ; jb 1097 <_sk_load_f16_hsw+0xd7> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,114,255,255,255 ; je e79 <_sk_load_f16_hsw+0x21> + DB 15,132,114,255,255,255 ; je fe1 <_sk_load_f16_hsw+0x21> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,98,255,255,255 ; jb e79 <_sk_load_f16_hsw+0x21> + DB 15,130,98,255,255,255 ; jb fe1 <_sk_load_f16_hsw+0x21> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,87,255,255,255 ; jmpq e79 <_sk_load_f16_hsw+0x21> + DB 233,87,255,255,255 ; jmpq fe1 <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,74,255,255,255 ; jmpq e79 <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq fe1 <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq e79 <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq fe1 <_sk_load_f16_hsw+0x21> PUBLIC _sk_store_f16_hsw _sk_store_f16_hsw LABEL PROC @@ -1155,7 +1254,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,27 ; jne f9d <_sk_store_f16_hsw+0x65> + DB 117,27 ; jne 1105 <_sk_store_f16_hsw+0x65> DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8) DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8) DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8) @@ -1164,22 +1263,22 @@ _sk_store_f16_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,241 ; je f99 <_sk_store_f16_hsw+0x61> + DB 116,241 ; je 1101 <_sk_store_f16_hsw+0x61> DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,229 ; jb f99 <_sk_store_f16_hsw+0x61> + DB 114,229 ; jb 1101 <_sk_store_f16_hsw+0x61> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je f99 <_sk_store_f16_hsw+0x61> + DB 116,221 ; je 1101 <_sk_store_f16_hsw+0x61> DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,209 ; jb f99 <_sk_store_f16_hsw+0x61> + DB 114,209 ; jb 1101 <_sk_store_f16_hsw+0x61> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je f99 <_sk_store_f16_hsw+0x61> + DB 116,201 ; je 1101 <_sk_store_f16_hsw+0x61> DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,189 ; jb f99 <_sk_store_f16_hsw+0x61> + DB 114,189 ; jb 1101 <_sk_store_f16_hsw+0x61> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp f99 <_sk_store_f16_hsw+0x61> + DB 235,181 ; jmp 1101 <_sk_store_f16_hsw+0x61> PUBLIC _sk_store_f32_hsw _sk_store_f32_hsw LABEL PROC @@ -1195,7 +1294,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 1051 <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 11b9 <_sk_store_f32_hsw+0x6d> 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 @@ -1208,22 +1307,22 @@ _sk_store_f32_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 104d <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 11b5 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 104d <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 11b5 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 104d <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 11b5 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 104d <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 11b5 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 104d <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 11b5 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 104d <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 11b5 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 104d <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 11b5 <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -1826,6 +1925,114 @@ _sk_xor__avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_darken_avx +_sk_darken_avx LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 + DB 197,100,89,204 ; vmulps %ymm4,%ymm3,%ymm9 + DB 196,193,124,95,193 ; vmaxps %ymm9,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 + DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 + DB 196,193,116,95,201 ; vmaxps %ymm9,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + DB 197,236,89,215 ; vmulps %ymm7,%ymm2,%ymm2 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,67,121,4,192,0 ; vpermilps $0x0,%xmm8,%xmm8 + DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 + 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 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_lighten_avx +_sk_lighten_avx LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 + DB 197,100,89,204 ; vmulps %ymm4,%ymm3,%ymm9 + DB 196,193,124,93,193 ; vminps %ymm9,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 + DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 + DB 196,193,116,93,201 ; vminps %ymm9,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + DB 197,236,89,215 ; vmulps %ymm7,%ymm2,%ymm2 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,67,121,4,192,0 ; vpermilps $0x0,%xmm8,%xmm8 + DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 + 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 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_difference_avx +_sk_difference_avx LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 + DB 197,100,89,204 ; vmulps %ymm4,%ymm3,%ymm9 + DB 196,193,124,93,193 ; vminps %ymm9,%ymm0,%ymm0 + DB 197,252,88,192 ; vaddps %ymm0,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,207 ; vmulps %ymm7,%ymm1,%ymm1 + DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 + DB 196,193,116,93,201 ; vminps %ymm9,%ymm1,%ymm1 + DB 197,244,88,201 ; vaddps %ymm1,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + DB 197,236,89,215 ; vmulps %ymm7,%ymm2,%ymm2 + DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,67,121,4,192,0 ; vpermilps $0x0,%xmm8,%xmm8 + DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 + 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 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_avx +_sk_exclusion_avx LABEL PROC + DB 197,124,88,196 ; vaddps %ymm4,%ymm0,%ymm8 + DB 197,252,89,196 ; vmulps %ymm4,%ymm0,%ymm0 + DB 197,252,88,192 ; vaddps %ymm0,%ymm0,%ymm0 + DB 197,188,92,192 ; vsubps %ymm0,%ymm8,%ymm0 + DB 197,116,88,197 ; vaddps %ymm5,%ymm1,%ymm8 + DB 197,244,89,205 ; vmulps %ymm5,%ymm1,%ymm1 + DB 197,244,88,201 ; vaddps %ymm1,%ymm1,%ymm1 + DB 197,188,92,201 ; vsubps %ymm1,%ymm8,%ymm1 + DB 197,108,88,198 ; vaddps %ymm6,%ymm2,%ymm8 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 197,121,110,192 ; vmovd %eax,%xmm8 + DB 196,67,121,4,192,0 ; vpermilps $0x0,%xmm8,%xmm8 + DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 + 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 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_clamp_0_avx _sk_clamp_0_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 @@ -2068,7 +2275,7 @@ _sk_scale_u8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,80 ; jne 86b <_sk_scale_u8_avx+0x60> + DB 117,80 ; jne 9f8 <_sk_scale_u8_avx+0x60> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9 DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8 @@ -2096,9 +2303,9 @@ _sk_scale_u8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 873 <_sk_scale_u8_avx+0x68> + DB 117,234 ; jne a00 <_sk_scale_u8_avx+0x68> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,143 ; jmp 81f <_sk_scale_u8_avx+0x14> + DB 235,143 ; jmp 9ac <_sk_scale_u8_avx+0x14> PUBLIC _sk_lerp_1_float_avx _sk_lerp_1_float_avx LABEL PROC @@ -2126,7 +2333,7 @@ _sk_lerp_u8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,116 ; jne 953 <_sk_lerp_u8_avx+0x84> + DB 117,116 ; jne ae0 <_sk_lerp_u8_avx+0x84> DB 197,122,126,0 ; vmovq (%rax),%xmm8 DB 196,66,121,49,200 ; vpmovzxbd %xmm8,%xmm9 DB 196,67,121,4,192,229 ; vpermilps $0xe5,%xmm8,%xmm8 @@ -2162,16 +2369,16 @@ _sk_lerp_u8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne 95b <_sk_lerp_u8_avx+0x8c> + DB 117,234 ; jne ae8 <_sk_lerp_u8_avx+0x8c> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 233,104,255,255,255 ; jmpq 8e3 <_sk_lerp_u8_avx+0x14> + DB 233,104,255,255,255 ; jmpq a70 <_sk_lerp_u8_avx+0x14> PUBLIC _sk_lerp_565_avx _sk_lerp_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,250,0,0,0 ; jne a83 <_sk_lerp_565_avx+0x108> + DB 15,133,250,0,0,0 ; jne c10 <_sk_lerp_565_avx+0x108> DB 196,65,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm8 DB 197,225,239,219 ; vpxor %xmm3,%xmm3,%xmm3 DB 197,185,105,219 ; vpunpckhwd %xmm3,%xmm8,%xmm3 @@ -2230,9 +2437,9 @@ _sk_lerp_565_avx LABEL PROC DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,243,254,255,255 ; ja 98f <_sk_lerp_565_avx+0x14> + DB 15,135,243,254,255,255 ; ja b1c <_sk_lerp_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # af0 <_sk_lerp_565_avx+0x175> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # c80 <_sk_lerp_565_avx+0x178> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2244,27 +2451,26 @@ _sk_lerp_565_avx LABEL PROC DB 196,65,57,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 DB 196,65,57,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 DB 196,65,57,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 - DB 233,159,254,255,255 ; jmpq 98f <_sk_lerp_565_avx+0x14> - DB 244 ; hlt - DB 255 ; (bad) + DB 233,159,254,255,255 ; jmpq b1c <_sk_lerp_565_avx+0x14> + DB 15,31,0 ; nopl (%rax) + DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) - DB 236 ; in (%dx),%al DB 255 ; (bad) + DB 233,255,255,255,225 ; jmpq ffffffffe2000c88 <_sk_linear_gradient_2stops_avx+0xffffffffe1ffee62> DB 255 ; (bad) - DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) + DB 217,255 ; fcos DB 255 ; (bad) - DB 220,255 ; fdivr %st,%st(7) + DB 255,209 ; callq *%rcx DB 255 ; (bad) - DB 255,212 ; callq *%rsp DB 255 ; (bad) + DB 255,201 ; dec %ecx DB 255 ; (bad) - DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 255,192 ; inc %eax + DB 189 ; .byte 0xbd DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -2280,7 +2486,7 @@ _sk_load_tables_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,56,2,0,0 ; jne d5c <_sk_load_tables_avx+0x250> + DB 15,133,56,2,0,0 ; jne eec <_sk_load_tables_avx+0x250> DB 196,65,124,16,4,184 ; vmovups (%r8,%rdi,4),%ymm8 DB 187,255,0,0,0 ; mov $0xff,%ebx DB 197,249,110,195 ; vmovd %ebx,%xmm0 @@ -2399,9 +2605,9 @@ _sk_load_tables_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 254,203 ; dec %bl DB 128,251,6 ; cmp $0x6,%bl - DB 15,135,185,253,255,255 ; ja b2a <_sk_load_tables_avx+0x1e> + DB 15,135,185,253,255,255 ; ja cba <_sk_load_tables_avx+0x1e> DB 15,182,219 ; movzbl %bl,%ebx - DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # e04 <_sk_load_tables_avx+0x2f8> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # f94 <_sk_load_tables_avx+0x2f8> DB 73,99,28,153 ; movslq (%r9,%rbx,4),%rbx DB 76,1,203 ; add %r9,%rbx DB 255,227 ; jmpq *%rbx @@ -2424,7 +2630,7 @@ _sk_load_tables_avx LABEL PROC DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8 DB 196,195,57,34,4,184,0 ; vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0 DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8 - DB 233,38,253,255,255 ; jmpq b2a <_sk_load_tables_avx+0x1e> + DB 233,38,253,255,255 ; jmpq cba <_sk_load_tables_avx+0x1e> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -2451,7 +2657,7 @@ _sk_load_a8_avx LABEL PROC DB 72,139,0 ; mov (%rax),%rax DB 72,1,248 ; add %rdi,%rax DB 77,133,192 ; test %r8,%r8 - DB 117,74 ; jne e7a <_sk_load_a8_avx+0x5a> + DB 117,74 ; jne 100a <_sk_load_a8_avx+0x5a> DB 197,250,126,0 ; vmovq (%rax),%xmm0 DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1 DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0 @@ -2478,9 +2684,9 @@ _sk_load_a8_avx LABEL PROC DB 77,9,217 ; or %r11,%r9 DB 72,131,193,8 ; add $0x8,%rcx DB 73,255,202 ; dec %r10 - DB 117,234 ; jne e82 <_sk_load_a8_avx+0x62> + DB 117,234 ; jne 1012 <_sk_load_a8_avx+0x62> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,149 ; jmp e34 <_sk_load_a8_avx+0x14> + DB 235,149 ; jmp fc4 <_sk_load_a8_avx+0x14> PUBLIC _sk_store_a8_avx _sk_store_a8_avx LABEL PROC @@ -2496,7 +2702,7 @@ _sk_store_a8_avx LABEL PROC DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne ee1 <_sk_store_a8_avx+0x42> + DB 117,10 ; jne 1071 <_sk_store_a8_avx+0x42> DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2504,10 +2710,10 @@ _sk_store_a8_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja edd <_sk_store_a8_avx+0x3e> + DB 119,236 ; ja 106d <_sk_store_a8_avx+0x3e> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # f44 <_sk_store_a8_avx+0xa5> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 10d4 <_sk_store_a8_avx+0xa5> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2518,7 +2724,7 @@ _sk_store_a8_avx LABEL PROC DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1) - DB 235,154 ; jmp edd <_sk_store_a8_avx+0x3e> + DB 235,154 ; jmp 106d <_sk_store_a8_avx+0x3e> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -2548,7 +2754,7 @@ _sk_load_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,209,0,0,0 ; jne 103f <_sk_load_565_avx+0xdf> + DB 15,133,209,0,0,0 ; jne 11cf <_sk_load_565_avx+0xdf> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -2598,9 +2804,9 @@ _sk_load_565_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,29,255,255,255 ; ja f74 <_sk_load_565_avx+0x14> + DB 15,135,29,255,255,255 ; ja 1104 <_sk_load_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 10ac <_sk_load_565_avx+0x14c> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 123c <_sk_load_565_avx+0x14c> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2612,12 +2818,12 @@ _sk_load_565_avx LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,201,254,255,255 ; jmpq f74 <_sk_load_565_avx+0x14> + DB 233,201,254,255,255 ; jmpq 1104 <_sk_load_565_avx+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 10b1 <_sk_load_565_avx+0x151> + DB 235,255 ; jmp 1241 <_sk_load_565_avx+0x151> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -2668,7 +2874,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 1166 <_sk_store_565_avx+0x9e> + DB 117,10 ; jne 12f6 <_sk_store_565_avx+0x9e> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2676,9 +2882,9 @@ _sk_store_565_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 1162 <_sk_store_565_avx+0x9a> + DB 119,236 ; ja 12f2 <_sk_store_565_avx+0x9a> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 11c4 <_sk_store_565_avx+0xfc> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1354 <_sk_store_565_avx+0xfc> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2689,7 +2895,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 1162 <_sk_store_565_avx+0x9a> + DB 235,159 ; jmp 12f2 <_sk_store_565_avx+0x9a> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -2719,7 +2925,7 @@ _sk_load_8888_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,157,0,0,0 ; jne 128b <_sk_load_8888_avx+0xab> + DB 15,133,157,0,0,0 ; jne 141b <_sk_load_8888_avx+0xab> DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -2757,9 +2963,9 @@ _sk_load_8888_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,80,255,255,255 ; ja 11f4 <_sk_load_8888_avx+0x14> + DB 15,135,80,255,255,255 ; ja 1384 <_sk_load_8888_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 1338 <_sk_load_8888_avx+0x158> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 14c8 <_sk_load_8888_avx+0x158> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -2782,7 +2988,7 @@ _sk_load_8888_avx LABEL PROC DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 DB 196,195,49,34,4,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 - DB 233,188,254,255,255 ; jmpq 11f4 <_sk_load_8888_avx+0x14> + DB 233,188,254,255,255 ; jmpq 1384 <_sk_load_8888_avx+0x14> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -2834,7 +3040,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 13f8 <_sk_store_8888_avx+0xa4> + DB 117,10 ; jne 1588 <_sk_store_8888_avx+0xa4> DB 196,65,124,17,4,185 ; vmovups %ymm8,(%r9,%rdi,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -2842,9 +3048,9 @@ _sk_store_8888_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 13f4 <_sk_store_8888_avx+0xa0> + DB 119,236 ; ja 1584 <_sk_store_8888_avx+0xa0> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,85,0,0,0 ; lea 0x55(%rip),%r8 # 1468 <_sk_store_8888_avx+0x114> + DB 76,141,5,85,0,0,0 ; lea 0x55(%rip),%r8 # 15f8 <_sk_store_8888_avx+0x114> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -2858,7 +3064,7 @@ _sk_store_8888_avx LABEL PROC DB 196,67,121,22,68,185,8,2 ; vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) DB 196,67,121,22,68,185,4,1 ; vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) DB 196,65,121,126,4,185 ; vmovd %xmm8,(%r9,%rdi,4) - DB 235,143 ; jmp 13f4 <_sk_store_8888_avx+0xa0> + DB 235,143 ; jmp 1584 <_sk_store_8888_avx+0xa0> DB 15,31,0 ; nopl (%rax) DB 245 ; cmc DB 255 ; (bad) @@ -2889,7 +3095,7 @@ _sk_load_f16_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,2,1,0,0 ; jne 1594 <_sk_load_f16_avx+0x110> + DB 15,133,2,1,0,0 ; jne 1724 <_sk_load_f16_avx+0x110> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -2947,29 +3153,29 @@ _sk_load_f16_avx LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 15f3 <_sk_load_f16_avx+0x16f> + DB 116,79 ; je 1783 <_sk_load_f16_avx+0x16f> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 15f3 <_sk_load_f16_avx+0x16f> + DB 114,67 ; jb 1783 <_sk_load_f16_avx+0x16f> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 1600 <_sk_load_f16_avx+0x17c> + DB 116,68 ; je 1790 <_sk_load_f16_avx+0x17c> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 1600 <_sk_load_f16_avx+0x17c> + DB 114,56 ; jb 1790 <_sk_load_f16_avx+0x17c> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,209,254,255,255 ; je 14a9 <_sk_load_f16_avx+0x25> + DB 15,132,209,254,255,255 ; je 1639 <_sk_load_f16_avx+0x25> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,193,254,255,255 ; jb 14a9 <_sk_load_f16_avx+0x25> + DB 15,130,193,254,255,255 ; jb 1639 <_sk_load_f16_avx+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,182,254,255,255 ; jmpq 14a9 <_sk_load_f16_avx+0x25> + DB 233,182,254,255,255 ; jmpq 1639 <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,169,254,255,255 ; jmpq 14a9 <_sk_load_f16_avx+0x25> + DB 233,169,254,255,255 ; jmpq 1639 <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,160,254,255,255 ; jmpq 14a9 <_sk_load_f16_avx+0x25> + DB 233,160,254,255,255 ; jmpq 1639 <_sk_load_f16_avx+0x25> PUBLIC _sk_store_f16_avx _sk_store_f16_avx LABEL PROC @@ -3008,7 +3214,7 @@ _sk_store_f16_avx LABEL PROC DB 196,65,25,98,205 ; vpunpckldq %xmm13,%xmm12,%xmm9 DB 196,65,25,106,197 ; vpunpckhdq %xmm13,%xmm12,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 16df <_sk_store_f16_avx+0xd6> + DB 117,31 ; jne 186f <_sk_store_f16_avx+0xd6> DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8) DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8) DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8) @@ -3017,22 +3223,22 @@ _sk_store_f16_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 16db <_sk_store_f16_avx+0xd2> + DB 116,240 ; je 186b <_sk_store_f16_avx+0xd2> DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 16db <_sk_store_f16_avx+0xd2> + DB 114,227 ; jb 186b <_sk_store_f16_avx+0xd2> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 16db <_sk_store_f16_avx+0xd2> + DB 116,218 ; je 186b <_sk_store_f16_avx+0xd2> DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 16db <_sk_store_f16_avx+0xd2> + DB 114,205 ; jb 186b <_sk_store_f16_avx+0xd2> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 16db <_sk_store_f16_avx+0xd2> + DB 116,196 ; je 186b <_sk_store_f16_avx+0xd2> DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 16db <_sk_store_f16_avx+0xd2> + DB 114,183 ; jb 186b <_sk_store_f16_avx+0xd2> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 16db <_sk_store_f16_avx+0xd2> + DB 235,174 ; jmp 186b <_sk_store_f16_avx+0xd2> PUBLIC _sk_store_f32_avx _sk_store_f32_avx LABEL PROC @@ -3048,7 +3254,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 179a <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 192a <_sk_store_f32_avx+0x6d> 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 @@ -3061,22 +3267,22 @@ _sk_store_f32_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 1796 <_sk_store_f32_avx+0x69> + DB 116,240 ; je 1926 <_sk_store_f32_avx+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 1796 <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 1926 <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 1796 <_sk_store_f32_avx+0x69> + DB 116,218 ; je 1926 <_sk_store_f32_avx+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 1796 <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 1926 <_sk_store_f32_avx+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 1796 <_sk_store_f32_avx+0x69> + DB 116,195 ; je 1926 <_sk_store_f32_avx+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 1796 <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 1926 <_sk_store_f32_avx+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 1796 <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 1926 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -3760,6 +3966,139 @@ _sk_xor__sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_darken_sse41 +_sk_darken_sse41 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 68,15,40,200 ; movaps %xmm0,%xmm9 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 68,15,89,207 ; mulps %xmm7,%xmm9 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 68,15,95,201 ; maxps %xmm1,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,199 ; mulps %xmm7,%xmm8 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,205 ; mulps %xmm5,%xmm9 + DB 69,15,95,193 ; maxps %xmm9,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_lighten_sse41 +_sk_lighten_sse41 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 68,15,40,200 ; movaps %xmm0,%xmm9 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 68,15,89,207 ; mulps %xmm7,%xmm9 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 68,15,93,201 ; minps %xmm1,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,199 ; mulps %xmm7,%xmm8 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,205 ; mulps %xmm5,%xmm9 + DB 69,15,93,193 ; minps %xmm9,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_difference_sse41 +_sk_difference_sse41 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 68,15,40,200 ; movaps %xmm0,%xmm9 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 68,15,89,207 ; mulps %xmm7,%xmm9 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 68,15,93,201 ; minps %xmm1,%xmm9 + DB 69,15,88,201 ; addps %xmm9,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,199 ; mulps %xmm7,%xmm8 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,205 ; mulps %xmm5,%xmm9 + DB 69,15,93,193 ; minps %xmm9,%xmm8 + DB 69,15,88,192 ; addps %xmm8,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,206 ; mulps %xmm6,%xmm9 + DB 65,15,93,209 ; minps %xmm9,%xmm2 + DB 15,88,210 ; addps %xmm2,%xmm2 + DB 68,15,92,194 ; subps %xmm2,%xmm8 + DB 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_sse41 +_sk_exclusion_sse41 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 15,40,200 ; movaps %xmm0,%xmm1 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 15,88,201 ; addps %xmm1,%xmm1 + DB 15,92,193 ; subps %xmm1,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,197 ; mulps %xmm5,%xmm8 + DB 69,15,88,192 ; addps %xmm8,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,214 ; mulps %xmm6,%xmm2 + DB 15,88,210 ; addps %xmm2,%xmm2 + DB 68,15,92,194 ; subps %xmm2,%xmm8 + DB 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + PUBLIC _sk_clamp_0_sse41 _sk_clamp_0_sse41 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 @@ -5169,6 +5508,139 @@ _sk_xor__sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_darken_sse2 +_sk_darken_sse2 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 68,15,40,200 ; movaps %xmm0,%xmm9 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 68,15,89,207 ; mulps %xmm7,%xmm9 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 68,15,95,201 ; maxps %xmm1,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,199 ; mulps %xmm7,%xmm8 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,205 ; mulps %xmm5,%xmm9 + DB 69,15,95,193 ; maxps %xmm9,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_lighten_sse2 +_sk_lighten_sse2 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 68,15,40,200 ; movaps %xmm0,%xmm9 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 68,15,89,207 ; mulps %xmm7,%xmm9 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 68,15,93,201 ; minps %xmm1,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,199 ; mulps %xmm7,%xmm8 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,205 ; mulps %xmm5,%xmm9 + DB 69,15,93,193 ; minps %xmm9,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + 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 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_difference_sse2 +_sk_difference_sse2 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 68,15,40,200 ; movaps %xmm0,%xmm9 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 68,15,89,207 ; mulps %xmm7,%xmm9 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 68,15,93,201 ; minps %xmm1,%xmm9 + DB 69,15,88,201 ; addps %xmm9,%xmm9 + DB 65,15,92,193 ; subps %xmm9,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,199 ; mulps %xmm7,%xmm8 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,205 ; mulps %xmm5,%xmm9 + DB 69,15,93,193 ; minps %xmm9,%xmm8 + DB 69,15,88,192 ; addps %xmm8,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 68,15,40,203 ; movaps %xmm3,%xmm9 + DB 68,15,89,206 ; mulps %xmm6,%xmm9 + DB 65,15,93,209 ; minps %xmm9,%xmm2 + DB 15,88,210 ; addps %xmm2,%xmm2 + DB 68,15,92,194 ; subps %xmm2,%xmm8 + DB 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_sse2 +_sk_exclusion_sse2 LABEL PROC + DB 68,15,40,193 ; movaps %xmm1,%xmm8 + DB 15,40,200 ; movaps %xmm0,%xmm1 + DB 15,88,196 ; addps %xmm4,%xmm0 + DB 15,89,204 ; mulps %xmm4,%xmm1 + DB 15,88,201 ; addps %xmm1,%xmm1 + DB 15,92,193 ; subps %xmm1,%xmm0 + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 15,88,205 ; addps %xmm5,%xmm1 + DB 68,15,89,197 ; mulps %xmm5,%xmm8 + DB 69,15,88,192 ; addps %xmm8,%xmm8 + DB 65,15,92,200 ; subps %xmm8,%xmm1 + DB 68,15,40,194 ; movaps %xmm2,%xmm8 + DB 68,15,88,198 ; addps %xmm6,%xmm8 + DB 15,89,214 ; mulps %xmm6,%xmm2 + DB 15,88,210 ; addps %xmm2,%xmm2 + DB 68,15,92,194 ; subps %xmm2,%xmm8 + DB 184,0,0,128,63 ; mov $0x3f800000,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2 + DB 15,92,211 ; subps %xmm3,%xmm2 + DB 15,89,215 ; mulps %xmm7,%xmm2 + DB 15,88,218 ; addps %xmm2,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 255,224 ; jmpq *%rax + PUBLIC _sk_clamp_0_sse2 _sk_clamp_0_sse2 LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 |