From 7125ac603613d156d83a528f27b03e04282a1e37 Mon Sep 17 00:00:00 2001 From: Mike Klein Date: Wed, 5 Apr 2017 15:27:22 -0400 Subject: jumper, to_2dot2 and from_2dot2 Nothing too tricky here. Change-Id: I2a10548efc75a6fd875fcb242790880d9b9a28fd Reviewed-on: https://skia-review.googlesource.com/11388 Commit-Queue: Mike Klein Reviewed-by: Matt Sarett --- src/jumper/SkJumper_generated_win.S | 728 ++++++++++++++++++++++++++---------- 1 file changed, 531 insertions(+), 197 deletions(-) (limited to 'src/jumper/SkJumper_generated_win.S') diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 627bec906e..c4872ae7e6 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -936,6 +936,87 @@ _sk_to_srgb_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_from_2dot2_hsw +_sk_from_2dot2_hsw LABEL PROC + DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,200 ; vrsqrtps %ymm8,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 197,252,89,192 ; vmulps %ymm0,%ymm0,%ymm0 + DB 196,65,60,89,208 ; vmulps %ymm8,%ymm8,%ymm10 + DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 + DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 196,65,124,82,210 ; vrsqrtps %ymm10,%ymm10 + DB 197,244,89,201 ; vmulps %ymm1,%ymm1,%ymm1 + DB 196,65,52,89,217 ; vmulps %ymm9,%ymm9,%ymm11 + DB 196,65,52,89,203 ; vmulps %ymm11,%ymm9,%ymm9 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 + DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 + DB 197,124,82,202 ; vrsqrtps %ymm2,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 196,65,124,82,210 ; vrsqrtps %ymm10,%ymm10 + DB 197,236,89,210 ; vmulps %ymm2,%ymm2,%ymm2 + DB 196,65,52,89,217 ; vmulps %ymm9,%ymm9,%ymm11 + DB 196,65,52,89,203 ; vmulps %ymm11,%ymm9,%ymm9 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 + DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_to_2dot2_hsw +_sk_to_2dot2_hsw LABEL PROC + DB 197,252,82,192 ; vrsqrtps %ymm0,%ymm0 + DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,200 ; vrsqrtps %ymm8,%ymm9 + DB 197,252,83,192 ; vrcpps %ymm0,%ymm0 + DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 + DB 196,65,124,83,193 ; vrcpps %ymm9,%ymm8 + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 + DB 197,252,82,201 ; vrsqrtps %ymm1,%ymm1 + DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 197,252,83,201 ; vrcpps %ymm1,%ymm1 + DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 + DB 196,65,124,83,202 ; vrcpps %ymm10,%ymm9 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 + DB 197,252,82,210 ; vrsqrtps %ymm2,%ymm2 + DB 197,124,82,202 ; vrsqrtps %ymm2,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 197,252,83,210 ; vrcpps %ymm2,%ymm2 + DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2 + DB 196,65,124,83,202 ; vrcpps %ymm10,%ymm9 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_scale_1_float_hsw _sk_scale_1_float_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -954,7 +1035,7 @@ _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 de4 <_sk_scale_u8_hsw+0x48> + DB 117,56 ; jne f3a <_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 @@ -978,9 +1059,9 @@ _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 dec <_sk_scale_u8_hsw+0x50> + DB 117,234 ; jne f42 <_sk_scale_u8_hsw+0x50> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,167 ; jmp db0 <_sk_scale_u8_hsw+0x14> + DB 235,167 ; jmp f06 <_sk_scale_u8_hsw+0x14> PUBLIC _sk_lerp_1_float_hsw _sk_lerp_1_float_hsw LABEL PROC @@ -1004,7 +1085,7 @@ _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 e94 <_sk_lerp_u8_hsw+0x5c> + DB 117,76 ; jne fea <_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 @@ -1032,16 +1113,16 @@ _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 e9c <_sk_lerp_u8_hsw+0x64> + DB 117,234 ; jne ff2 <_sk_lerp_u8_hsw+0x64> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,147 ; jmp e4c <_sk_lerp_u8_hsw+0x14> + DB 235,147 ; jmp fa2 <_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 f7a <_sk_lerp_565_hsw+0xc1> + DB 15,133,179,0,0,0 ; jne 10d0 <_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 @@ -1087,9 +1168,9 @@ _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 ecd <_sk_lerp_565_hsw+0x14> + DB 15,135,59,255,255,255 ; ja 1023 <_sk_lerp_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # fe8 <_sk_lerp_565_hsw+0x12f> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 113c <_sk_lerp_565_hsw+0x12d> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -1101,28 +1182,27 @@ _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 ecd <_sk_lerp_565_hsw+0x14> - DB 102,144 ; xchg %ax,%ax - DB 242,255 ; repnz (bad) + DB 233,231,254,255,255 ; jmpq 1023 <_sk_lerp_565_hsw+0x14> + DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) + DB 236 ; in (%dx),%al DB 255 ; (bad) - DB 255,226 ; jmpq *%rdx DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) - DB 218,255 ; (bad) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) + DB 255,212 ; callq *%rsp DB 255 ; (bad) - DB 255,202 ; dec %edx DB 255 ; (bad) + DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 190 ; .byte 0xbe + DB 255,192 ; inc %eax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -1134,7 +1214,7 @@ _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 1092 <_sk_load_tables_hsw+0x8e> + DB 117,121 ; jne 11e6 <_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 @@ -1170,7 +1250,7 @@ _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 101e <_sk_load_tables_hsw+0x1a> + DB 233,99,255,255,255 ; jmpq 1172 <_sk_load_tables_hsw+0x1a> PUBLIC _sk_load_a8_hsw _sk_load_a8_hsw LABEL PROC @@ -1179,7 +1259,7 @@ _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 10fd <_sk_load_a8_hsw+0x42> + DB 117,50 ; jne 1251 <_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 @@ -1202,9 +1282,9 @@ _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 1105 <_sk_load_a8_hsw+0x4a> + DB 117,234 ; jne 1259 <_sk_load_a8_hsw+0x4a> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,173 ; jmp 10cf <_sk_load_a8_hsw+0x14> + DB 235,173 ; jmp 1223 <_sk_load_a8_hsw+0x14> PUBLIC _sk_store_a8_hsw _sk_store_a8_hsw LABEL PROC @@ -1219,7 +1299,7 @@ _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 115d <_sk_store_a8_hsw+0x3b> + DB 117,10 ; jne 12b1 <_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 @@ -1227,10 +1307,10 @@ _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 1159 <_sk_store_a8_hsw+0x37> + DB 119,236 ; ja 12ad <_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 # 11c0 <_sk_store_a8_hsw+0x9e> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1314 <_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 @@ -1241,7 +1321,7 @@ _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 1159 <_sk_store_a8_hsw+0x37> + DB 235,154 ; jmp 12ad <_sk_store_a8_hsw+0x37> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -1273,7 +1353,7 @@ _sk_load_g8_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 117,60 ; jne 1228 <_sk_load_g8_hsw+0x4c> + DB 117,60 ; jne 137c <_sk_load_g8_hsw+0x4c> 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 @@ -1298,16 +1378,16 @@ _sk_load_g8_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 1230 <_sk_load_g8_hsw+0x54> + DB 117,234 ; jne 1384 <_sk_load_g8_hsw+0x54> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,163 ; jmp 11f0 <_sk_load_g8_hsw+0x14> + DB 235,163 ; jmp 1344 <_sk_load_g8_hsw+0x14> PUBLIC _sk_load_565_hsw _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 12f0 <_sk_load_565_hsw+0xa3> + DB 15,133,149,0,0,0 ; jne 1444 <_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 @@ -1347,9 +1427,9 @@ _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 1261 <_sk_load_565_hsw+0x14> + DB 15,135,89,255,255,255 ; ja 13b5 <_sk_load_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 135c <_sk_load_565_hsw+0x10f> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 14b0 <_sk_load_565_hsw+0x10f> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -1361,7 +1441,7 @@ _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 1261 <_sk_load_565_hsw+0x14> + DB 233,5,255,255,255 ; jmpq 13b5 <_sk_load_565_hsw+0x14> DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) @@ -1409,7 +1489,7 @@ _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 13e4 <_sk_store_565_hsw+0x6c> + DB 117,10 ; jne 1538 <_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 @@ -1417,9 +1497,9 @@ _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 13e0 <_sk_store_565_hsw+0x68> + DB 119,236 ; ja 1534 <_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 # 1444 <_sk_store_565_hsw+0xcc> + DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 1598 <_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 @@ -1430,7 +1510,7 @@ _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 13e0 <_sk_store_565_hsw+0x68> + DB 235,159 ; jmp 1534 <_sk_store_565_hsw+0x68> DB 15,31,0 ; nopl (%rax) DB 244 ; hlt DB 255 ; (bad) @@ -1461,7 +1541,7 @@ _sk_load_4444_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 1521 <_sk_load_4444_hsw+0xc1> + DB 15,133,179,0,0,0 ; jne 1675 <_sk_load_4444_hsw+0xc1> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,98,125,51,200 ; vpmovzxwd %xmm0,%ymm9 DB 184,0,240,0,0 ; mov $0xf000,%eax @@ -1507,9 +1587,9 @@ _sk_load_4444_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 15,135,59,255,255,255 ; ja 1474 <_sk_load_4444_hsw+0x14> + DB 15,135,59,255,255,255 ; ja 15c8 <_sk_load_4444_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1590 <_sk_load_4444_hsw+0x130> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 16e4 <_sk_load_4444_hsw+0x130> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -1521,13 +1601,13 @@ _sk_load_4444_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 233,231,254,255,255 ; jmpq 1474 <_sk_load_4444_hsw+0x14> + DB 233,231,254,255,255 ; jmpq 15c8 <_sk_load_4444_hsw+0x14> DB 15,31,0 ; nopl (%rax) DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe2001598 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff563> + DB 233,255,255,255,225 ; jmpq ffffffffe20016ec <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff563> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -1569,7 +1649,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 161e <_sk_store_4444_hsw+0x72> + DB 117,10 ; jne 1772 <_sk_store_4444_hsw+0x72> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1577,9 +1657,9 @@ _sk_store_4444_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 161a <_sk_store_4444_hsw+0x6e> + DB 119,236 ; ja 176e <_sk_store_4444_hsw+0x6e> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 167c <_sk_store_4444_hsw+0xd0> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 17d0 <_sk_store_4444_hsw+0xd0> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -1590,7 +1670,7 @@ _sk_store_4444_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 161a <_sk_store_4444_hsw+0x6e> + DB 235,159 ; jmp 176e <_sk_store_4444_hsw+0x6e> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -1622,7 +1702,7 @@ _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 1715 <_sk_load_8888_hsw+0x7d> + DB 117,104 ; jne 1869 <_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 @@ -1655,7 +1735,7 @@ _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 16b2 <_sk_load_8888_hsw+0x1a> + DB 233,116,255,255,255 ; jmpq 1806 <_sk_load_8888_hsw+0x1a> PUBLIC _sk_store_8888_hsw _sk_store_8888_hsw LABEL PROC @@ -1681,7 +1761,7 @@ _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 17b2 <_sk_store_8888_hsw+0x74> + DB 117,12 ; jne 1906 <_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 @@ -1694,14 +1774,14 @@ _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 17ab <_sk_store_8888_hsw+0x6d> + DB 235,211 ; jmp 18ff <_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 1843 <_sk_load_f16_hsw+0x6b> + DB 117,97 ; jne 1997 <_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 @@ -1727,29 +1807,29 @@ _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 18a2 <_sk_load_f16_hsw+0xca> + DB 116,79 ; je 19f6 <_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 18a2 <_sk_load_f16_hsw+0xca> + DB 114,67 ; jb 19f6 <_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 18af <_sk_load_f16_hsw+0xd7> + DB 116,68 ; je 1a03 <_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 18af <_sk_load_f16_hsw+0xd7> + DB 114,56 ; jb 1a03 <_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 17f9 <_sk_load_f16_hsw+0x21> + DB 15,132,114,255,255,255 ; je 194d <_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 17f9 <_sk_load_f16_hsw+0x21> + DB 15,130,98,255,255,255 ; jb 194d <_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 17f9 <_sk_load_f16_hsw+0x21> + DB 233,87,255,255,255 ; jmpq 194d <_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 17f9 <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq 194d <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq 17f9 <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 194d <_sk_load_f16_hsw+0x21> PUBLIC _sk_store_f16_hsw _sk_store_f16_hsw LABEL PROC @@ -1768,7 +1848,7 @@ _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 191d <_sk_store_f16_hsw+0x65> + DB 117,27 ; jne 1a71 <_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) @@ -1777,29 +1857,29 @@ _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 1919 <_sk_store_f16_hsw+0x61> + DB 116,241 ; je 1a6d <_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 1919 <_sk_store_f16_hsw+0x61> + DB 114,229 ; jb 1a6d <_sk_store_f16_hsw+0x61> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je 1919 <_sk_store_f16_hsw+0x61> + DB 116,221 ; je 1a6d <_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 1919 <_sk_store_f16_hsw+0x61> + DB 114,209 ; jb 1a6d <_sk_store_f16_hsw+0x61> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je 1919 <_sk_store_f16_hsw+0x61> + DB 116,201 ; je 1a6d <_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 1919 <_sk_store_f16_hsw+0x61> + DB 114,189 ; jb 1a6d <_sk_store_f16_hsw+0x61> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp 1919 <_sk_store_f16_hsw+0x61> + DB 235,181 ; jmp 1a6d <_sk_store_f16_hsw+0x61> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_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 15,133,201,0,0,0 ; jne 1a3b <_sk_load_u16_be_hsw+0xd7> + DB 15,133,201,0,0,0 ; jne 1b8f <_sk_load_u16_be_hsw+0xd7> 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 @@ -1848,29 +1928,29 @@ _sk_load_u16_be_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 1a9a <_sk_load_u16_be_hsw+0x136> + DB 116,79 ; je 1bee <_sk_load_u16_be_hsw+0x136> 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 1a9a <_sk_load_u16_be_hsw+0x136> + DB 114,67 ; jb 1bee <_sk_load_u16_be_hsw+0x136> 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 1aa7 <_sk_load_u16_be_hsw+0x143> + DB 116,68 ; je 1bfb <_sk_load_u16_be_hsw+0x143> 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 1aa7 <_sk_load_u16_be_hsw+0x143> + DB 114,56 ; jb 1bfb <_sk_load_u16_be_hsw+0x143> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,10,255,255,255 ; je 1989 <_sk_load_u16_be_hsw+0x25> + DB 15,132,10,255,255,255 ; je 1add <_sk_load_u16_be_hsw+0x25> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,250,254,255,255 ; jb 1989 <_sk_load_u16_be_hsw+0x25> + DB 15,130,250,254,255,255 ; jb 1add <_sk_load_u16_be_hsw+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,239,254,255,255 ; jmpq 1989 <_sk_load_u16_be_hsw+0x25> + DB 233,239,254,255,255 ; jmpq 1add <_sk_load_u16_be_hsw+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,226,254,255,255 ; jmpq 1989 <_sk_load_u16_be_hsw+0x25> + DB 233,226,254,255,255 ; jmpq 1add <_sk_load_u16_be_hsw+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,217,254,255,255 ; jmpq 1989 <_sk_load_u16_be_hsw+0x25> + DB 233,217,254,255,255 ; jmpq 1add <_sk_load_u16_be_hsw+0x25> PUBLIC _sk_store_u16_be_hsw _sk_store_u16_be_hsw LABEL PROC @@ -1916,7 +1996,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 1ba3 <_sk_store_u16_be_hsw+0xf3> + DB 117,31 ; jne 1cf7 <_sk_store_u16_be_hsw+0xf3> 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) @@ -1925,22 +2005,22 @@ _sk_store_u16_be_hsw 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 1b9f <_sk_store_u16_be_hsw+0xef> + DB 116,240 ; je 1cf3 <_sk_store_u16_be_hsw+0xef> 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 1b9f <_sk_store_u16_be_hsw+0xef> + DB 114,227 ; jb 1cf3 <_sk_store_u16_be_hsw+0xef> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 1b9f <_sk_store_u16_be_hsw+0xef> + DB 116,218 ; je 1cf3 <_sk_store_u16_be_hsw+0xef> 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 1b9f <_sk_store_u16_be_hsw+0xef> + DB 114,205 ; jb 1cf3 <_sk_store_u16_be_hsw+0xef> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 1b9f <_sk_store_u16_be_hsw+0xef> + DB 116,196 ; je 1cf3 <_sk_store_u16_be_hsw+0xef> 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 1b9f <_sk_store_u16_be_hsw+0xef> + DB 114,183 ; jb 1cf3 <_sk_store_u16_be_hsw+0xef> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 1b9f <_sk_store_u16_be_hsw+0xef> + DB 235,174 ; jmp 1cf3 <_sk_store_u16_be_hsw+0xef> PUBLIC _sk_store_f32_hsw _sk_store_f32_hsw LABEL PROC @@ -1956,7 +2036,7 @@ _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 1c5e <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 1db2 <_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 @@ -1969,22 +2049,22 @@ _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 1c5a <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 1dae <_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 1c5a <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 1dae <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 1c5a <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 1dae <_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 1c5a <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 1dae <_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 1c5a <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 1dae <_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 1c5a <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 1dae <_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 1c5a <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 1dae <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -3275,6 +3355,87 @@ _sk_to_srgb_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_from_2dot2_avx +_sk_from_2dot2_avx LABEL PROC + DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,200 ; vrsqrtps %ymm8,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 197,252,89,192 ; vmulps %ymm0,%ymm0,%ymm0 + DB 196,65,60,89,208 ; vmulps %ymm8,%ymm8,%ymm10 + DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 + DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 196,65,124,82,210 ; vrsqrtps %ymm10,%ymm10 + DB 197,244,89,201 ; vmulps %ymm1,%ymm1,%ymm1 + DB 196,65,52,89,217 ; vmulps %ymm9,%ymm9,%ymm11 + DB 196,65,52,89,203 ; vmulps %ymm11,%ymm9,%ymm9 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 + DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 + DB 197,124,82,202 ; vrsqrtps %ymm2,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 196,65,124,82,210 ; vrsqrtps %ymm10,%ymm10 + DB 197,236,89,210 ; vmulps %ymm2,%ymm2,%ymm2 + DB 196,65,52,89,217 ; vmulps %ymm9,%ymm9,%ymm11 + DB 196,65,52,89,203 ; vmulps %ymm11,%ymm9,%ymm9 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 + DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_to_2dot2_avx +_sk_to_2dot2_avx LABEL PROC + DB 197,252,82,192 ; vrsqrtps %ymm0,%ymm0 + DB 197,124,82,192 ; vrsqrtps %ymm0,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,192 ; vrsqrtps %ymm8,%ymm8 + DB 196,65,124,82,200 ; vrsqrtps %ymm8,%ymm9 + DB 197,252,83,192 ; vrcpps %ymm0,%ymm0 + DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 + DB 196,65,124,83,193 ; vrcpps %ymm9,%ymm8 + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 + DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 + DB 197,252,82,201 ; vrsqrtps %ymm1,%ymm1 + DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 197,252,83,201 ; vrcpps %ymm1,%ymm1 + DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 + DB 196,65,124,83,202 ; vrcpps %ymm10,%ymm9 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 + DB 197,252,82,210 ; vrsqrtps %ymm2,%ymm2 + DB 197,124,82,202 ; vrsqrtps %ymm2,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,201 ; vrsqrtps %ymm9,%ymm9 + DB 196,65,124,82,209 ; vrsqrtps %ymm9,%ymm10 + DB 197,252,83,210 ; vrcpps %ymm2,%ymm2 + DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2 + DB 196,65,124,83,202 ; vrcpps %ymm10,%ymm9 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_scale_1_float_avx _sk_scale_1_float_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -3293,7 +3454,7 @@ _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 1015 <_sk_scale_u8_avx+0x60> + DB 117,80 ; jne 116b <_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 @@ -3321,9 +3482,9 @@ _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 101d <_sk_scale_u8_avx+0x68> + DB 117,234 ; jne 1173 <_sk_scale_u8_avx+0x68> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,143 ; jmp fc9 <_sk_scale_u8_avx+0x14> + DB 235,143 ; jmp 111f <_sk_scale_u8_avx+0x14> PUBLIC _sk_lerp_1_float_avx _sk_lerp_1_float_avx LABEL PROC @@ -3351,7 +3512,7 @@ _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 10fd <_sk_lerp_u8_avx+0x84> + DB 117,116 ; jne 1253 <_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 @@ -3387,16 +3548,16 @@ _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 1105 <_sk_lerp_u8_avx+0x8c> + DB 117,234 ; jne 125b <_sk_lerp_u8_avx+0x8c> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 233,104,255,255,255 ; jmpq 108d <_sk_lerp_u8_avx+0x14> + DB 233,104,255,255,255 ; jmpq 11e3 <_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 122d <_sk_lerp_565_avx+0x108> + DB 15,133,250,0,0,0 ; jne 1383 <_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 @@ -3455,9 +3616,9 @@ _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 1139 <_sk_lerp_565_avx+0x14> + DB 15,135,243,254,255,255 ; ja 128f <_sk_lerp_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 129c <_sk_lerp_565_avx+0x177> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 13f0 <_sk_lerp_565_avx+0x175> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -3469,28 +3630,27 @@ _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 1139 <_sk_lerp_565_avx+0x14> - DB 102,144 ; xchg %ax,%ax - DB 242,255 ; repnz (bad) + DB 233,159,254,255,255 ; jmpq 128f <_sk_lerp_565_avx+0x14> + DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) + DB 236 ; in (%dx),%al DB 255 ; (bad) - DB 255,226 ; jmpq *%rdx DB 255 ; (bad) + DB 255,228 ; jmpq *%rsp DB 255 ; (bad) DB 255 ; (bad) - DB 218,255 ; (bad) DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 220,255 ; fdivr %st,%st(7) DB 255 ; (bad) + DB 255,212 ; callq *%rsp DB 255 ; (bad) - DB 255,202 ; dec %edx DB 255 ; (bad) + DB 255,204 ; dec %esp DB 255 ; (bad) DB 255 ; (bad) - DB 190 ; .byte 0xbe + DB 255,192 ; inc %eax DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -3506,7 +3666,7 @@ _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 1508 <_sk_load_tables_avx+0x250> + DB 15,133,56,2,0,0 ; jne 165c <_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 @@ -3625,9 +3785,9 @@ _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 12d6 <_sk_load_tables_avx+0x1e> + DB 15,135,185,253,255,255 ; ja 142a <_sk_load_tables_avx+0x1e> DB 15,182,219 ; movzbl %bl,%ebx - DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 15b0 <_sk_load_tables_avx+0x2f8> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 1704 <_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 @@ -3650,7 +3810,7 @@ _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 12d6 <_sk_load_tables_avx+0x1e> + DB 233,38,253,255,255 ; jmpq 142a <_sk_load_tables_avx+0x1e> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -3677,7 +3837,7 @@ _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 1626 <_sk_load_a8_avx+0x5a> + DB 117,74 ; jne 177a <_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 @@ -3704,9 +3864,9 @@ _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 162e <_sk_load_a8_avx+0x62> + DB 117,234 ; jne 1782 <_sk_load_a8_avx+0x62> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,149 ; jmp 15e0 <_sk_load_a8_avx+0x14> + DB 235,149 ; jmp 1734 <_sk_load_a8_avx+0x14> PUBLIC _sk_store_a8_avx _sk_store_a8_avx LABEL PROC @@ -3722,7 +3882,7 @@ _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 168d <_sk_store_a8_avx+0x42> + DB 117,10 ; jne 17e1 <_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 @@ -3730,10 +3890,10 @@ _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 1689 <_sk_store_a8_avx+0x3e> + DB 119,236 ; ja 17dd <_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 # 16f0 <_sk_store_a8_avx+0xa5> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1844 <_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 @@ -3744,7 +3904,7 @@ _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 1689 <_sk_store_a8_avx+0x3e> + DB 235,154 ; jmp 17dd <_sk_store_a8_avx+0x3e> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -3776,7 +3936,7 @@ _sk_load_g8_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 117,91 ; jne 1777 <_sk_load_g8_avx+0x6b> + DB 117,91 ; jne 18cb <_sk_load_g8_avx+0x6b> 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 @@ -3806,16 +3966,16 @@ _sk_load_g8_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 177f <_sk_load_g8_avx+0x73> + DB 117,234 ; jne 18d3 <_sk_load_g8_avx+0x73> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,132 ; jmp 1720 <_sk_load_g8_avx+0x14> + DB 235,132 ; jmp 1874 <_sk_load_g8_avx+0x14> PUBLIC _sk_load_565_avx _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 187b <_sk_load_565_avx+0xdf> + DB 15,133,209,0,0,0 ; jne 19cf <_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 @@ -3865,9 +4025,9 @@ _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 17b0 <_sk_load_565_avx+0x14> + DB 15,135,29,255,255,255 ; ja 1904 <_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 # 18e8 <_sk_load_565_avx+0x14c> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1a3c <_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 @@ -3879,12 +4039,12 @@ _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 17b0 <_sk_load_565_avx+0x14> + DB 233,201,254,255,255 ; jmpq 1904 <_sk_load_565_avx+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 18ed <_sk_load_565_avx+0x151> + DB 235,255 ; jmp 1a41 <_sk_load_565_avx+0x151> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -3935,7 +4095,7 @@ _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 19a2 <_sk_store_565_avx+0x9e> + DB 117,10 ; jne 1af6 <_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 @@ -3943,9 +4103,9 @@ _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 199e <_sk_store_565_avx+0x9a> + DB 119,236 ; ja 1af2 <_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 # 1a00 <_sk_store_565_avx+0xfc> + DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1b54 <_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 @@ -3956,7 +4116,7 @@ _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 199e <_sk_store_565_avx+0x9a> + DB 235,159 ; jmp 1af2 <_sk_store_565_avx+0x9a> DB 144 ; nop DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -3986,7 +4146,7 @@ _sk_load_4444_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 15,133,245,0,0,0 ; jne 1b1f <_sk_load_4444_avx+0x103> + DB 15,133,245,0,0,0 ; jne 1c73 <_sk_load_4444_avx+0x103> 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 @@ -4043,9 +4203,9 @@ _sk_load_4444_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 15,135,249,254,255,255 ; ja 1a30 <_sk_load_4444_avx+0x14> + DB 15,135,249,254,255,255 ; ja 1b84 <_sk_load_4444_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1b8c <_sk_load_4444_avx+0x170> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1ce0 <_sk_load_4444_avx+0x170> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -4057,12 +4217,12 @@ _sk_load_4444_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 233,165,254,255,255 ; jmpq 1a30 <_sk_load_4444_avx+0x14> + DB 233,165,254,255,255 ; jmpq 1b84 <_sk_load_4444_avx+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 1b91 <_sk_load_4444_avx+0x175> + DB 235,255 ; jmp 1ce5 <_sk_load_4444_avx+0x175> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -4116,7 +4276,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 1c57 <_sk_store_4444_avx+0xaf> + DB 117,10 ; jne 1dab <_sk_store_4444_avx+0xaf> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -4124,9 +4284,9 @@ _sk_store_4444_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 1c53 <_sk_store_4444_avx+0xab> + DB 119,236 ; ja 1da7 <_sk_store_4444_avx+0xab> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1cb4 <_sk_store_4444_avx+0x10c> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1e08 <_sk_store_4444_avx+0x10c> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -4137,7 +4297,7 @@ _sk_store_4444_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 1c53 <_sk_store_4444_avx+0xab> + DB 235,159 ; jmp 1da7 <_sk_store_4444_avx+0xab> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -4166,7 +4326,7 @@ _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 1d7b <_sk_load_8888_avx+0xab> + DB 15,133,157,0,0,0 ; jne 1ecf <_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 @@ -4204,9 +4364,9 @@ _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 1ce4 <_sk_load_8888_avx+0x14> + DB 15,135,80,255,255,255 ; ja 1e38 <_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 # 1e28 <_sk_load_8888_avx+0x158> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 1f7c <_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 @@ -4229,7 +4389,7 @@ _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 1ce4 <_sk_load_8888_avx+0x14> + DB 233,188,254,255,255 ; jmpq 1e38 <_sk_load_8888_avx+0x14> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -4281,7 +4441,7 @@ _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 1ee8 <_sk_store_8888_avx+0xa4> + DB 117,10 ; jne 203c <_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 @@ -4289,9 +4449,9 @@ _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 1ee4 <_sk_store_8888_avx+0xa0> + DB 119,236 ; ja 2038 <_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 # 1f58 <_sk_store_8888_avx+0x114> + DB 76,141,5,85,0,0,0 ; lea 0x55(%rip),%r8 # 20ac <_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 @@ -4305,7 +4465,7 @@ _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 1ee4 <_sk_store_8888_avx+0xa0> + DB 235,143 ; jmp 2038 <_sk_store_8888_avx+0xa0> DB 15,31,0 ; nopl (%rax) DB 245 ; cmc DB 255 ; (bad) @@ -4336,7 +4496,7 @@ _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 15,133,17,1,0,0 ; jne 2093 <_sk_load_f16_avx+0x11f> + DB 15,133,17,1,0,0 ; jne 21e7 <_sk_load_f16_avx+0x11f> 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 @@ -4398,29 +4558,29 @@ _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 20f2 <_sk_load_f16_avx+0x17e> + DB 116,79 ; je 2246 <_sk_load_f16_avx+0x17e> 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 20f2 <_sk_load_f16_avx+0x17e> + DB 114,67 ; jb 2246 <_sk_load_f16_avx+0x17e> 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 20ff <_sk_load_f16_avx+0x18b> + DB 116,68 ; je 2253 <_sk_load_f16_avx+0x18b> 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 20ff <_sk_load_f16_avx+0x18b> + DB 114,56 ; jb 2253 <_sk_load_f16_avx+0x18b> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,194,254,255,255 ; je 1f99 <_sk_load_f16_avx+0x25> + DB 15,132,194,254,255,255 ; je 20ed <_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 15,130,178,254,255,255 ; jb 1f99 <_sk_load_f16_avx+0x25> + DB 15,130,178,254,255,255 ; jb 20ed <_sk_load_f16_avx+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,167,254,255,255 ; jmpq 1f99 <_sk_load_f16_avx+0x25> + DB 233,167,254,255,255 ; jmpq 20ed <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,154,254,255,255 ; jmpq 1f99 <_sk_load_f16_avx+0x25> + DB 233,154,254,255,255 ; jmpq 20ed <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,145,254,255,255 ; jmpq 1f99 <_sk_load_f16_avx+0x25> + DB 233,145,254,255,255 ; jmpq 20ed <_sk_load_f16_avx+0x25> PUBLIC _sk_store_f16_avx _sk_store_f16_avx LABEL PROC @@ -4459,7 +4619,7 @@ _sk_store_f16_avx LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 21da <_sk_store_f16_avx+0xd2> + DB 117,31 ; jne 232e <_sk_store_f16_avx+0xd2> 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) @@ -4468,29 +4628,29 @@ _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 21d6 <_sk_store_f16_avx+0xce> + DB 116,240 ; je 232a <_sk_store_f16_avx+0xce> 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 21d6 <_sk_store_f16_avx+0xce> + DB 114,227 ; jb 232a <_sk_store_f16_avx+0xce> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 21d6 <_sk_store_f16_avx+0xce> + DB 116,218 ; je 232a <_sk_store_f16_avx+0xce> 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 21d6 <_sk_store_f16_avx+0xce> + DB 114,205 ; jb 232a <_sk_store_f16_avx+0xce> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 21d6 <_sk_store_f16_avx+0xce> + DB 116,196 ; je 232a <_sk_store_f16_avx+0xce> 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 21d6 <_sk_store_f16_avx+0xce> + DB 114,183 ; jb 232a <_sk_store_f16_avx+0xce> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 21d6 <_sk_store_f16_avx+0xce> + DB 235,174 ; jmp 232a <_sk_store_f16_avx+0xce> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_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 15,133,1,1,0,0 ; jne 2337 <_sk_load_u16_be_avx+0x10f> + DB 15,133,1,1,0,0 ; jne 248b <_sk_load_u16_be_avx+0x10f> 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 @@ -4549,29 +4709,29 @@ _sk_load_u16_be_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 2396 <_sk_load_u16_be_avx+0x16e> + DB 116,79 ; je 24ea <_sk_load_u16_be_avx+0x16e> 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 2396 <_sk_load_u16_be_avx+0x16e> + DB 114,67 ; jb 24ea <_sk_load_u16_be_avx+0x16e> 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 23a3 <_sk_load_u16_be_avx+0x17b> + DB 116,68 ; je 24f7 <_sk_load_u16_be_avx+0x17b> 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 23a3 <_sk_load_u16_be_avx+0x17b> + DB 114,56 ; jb 24f7 <_sk_load_u16_be_avx+0x17b> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,210,254,255,255 ; je 224d <_sk_load_u16_be_avx+0x25> + DB 15,132,210,254,255,255 ; je 23a1 <_sk_load_u16_be_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 15,130,194,254,255,255 ; jb 224d <_sk_load_u16_be_avx+0x25> + DB 15,130,194,254,255,255 ; jb 23a1 <_sk_load_u16_be_avx+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,183,254,255,255 ; jmpq 224d <_sk_load_u16_be_avx+0x25> + DB 233,183,254,255,255 ; jmpq 23a1 <_sk_load_u16_be_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,170,254,255,255 ; jmpq 224d <_sk_load_u16_be_avx+0x25> + DB 233,170,254,255,255 ; jmpq 23a1 <_sk_load_u16_be_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,161,254,255,255 ; jmpq 224d <_sk_load_u16_be_avx+0x25> + DB 233,161,254,255,255 ; jmpq 23a1 <_sk_load_u16_be_avx+0x25> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC @@ -4618,7 +4778,7 @@ _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 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 24a6 <_sk_store_u16_be_avx+0xfa> + DB 117,31 ; jne 25fa <_sk_store_u16_be_avx+0xfa> 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) @@ -4627,22 +4787,22 @@ _sk_store_u16_be_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 24a2 <_sk_store_u16_be_avx+0xf6> + DB 116,240 ; je 25f6 <_sk_store_u16_be_avx+0xf6> 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 24a2 <_sk_store_u16_be_avx+0xf6> + DB 114,227 ; jb 25f6 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 24a2 <_sk_store_u16_be_avx+0xf6> + DB 116,218 ; je 25f6 <_sk_store_u16_be_avx+0xf6> 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 24a2 <_sk_store_u16_be_avx+0xf6> + DB 114,205 ; jb 25f6 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 24a2 <_sk_store_u16_be_avx+0xf6> + DB 116,196 ; je 25f6 <_sk_store_u16_be_avx+0xf6> 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 24a2 <_sk_store_u16_be_avx+0xf6> + DB 114,183 ; jb 25f6 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 24a2 <_sk_store_u16_be_avx+0xf6> + DB 235,174 ; jmp 25f6 <_sk_store_u16_be_avx+0xf6> PUBLIC _sk_store_f32_avx _sk_store_f32_avx LABEL PROC @@ -4658,7 +4818,7 @@ _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 2561 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 26b5 <_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 @@ -4671,22 +4831,22 @@ _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 255d <_sk_store_f32_avx+0x69> + DB 116,240 ; je 26b1 <_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 255d <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 26b1 <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 255d <_sk_store_f32_avx+0x69> + DB 116,218 ; je 26b1 <_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 255d <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 26b1 <_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 255d <_sk_store_f32_avx+0x69> + DB 116,195 ; je 26b1 <_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 255d <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 26b1 <_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 255d <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 26b1 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -6239,6 +6399,93 @@ _sk_to_srgb_sse41 LABEL PROC DB 72,131,196,24 ; add $0x18,%rsp DB 255,224 ; jmpq *%rax +PUBLIC _sk_from_2dot2_sse41 +_sk_from_2dot2_sse41 LABEL PROC + DB 68,15,40,192 ; movaps %xmm0,%xmm8 + DB 65,15,82,192 ; rsqrtps %xmm8,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 68,15,82,200 ; rsqrtps %xmm0,%xmm9 + DB 65,15,82,193 ; rsqrtps %xmm9,%xmm0 + DB 68,15,82,208 ; rsqrtps %xmm0,%xmm10 + DB 69,15,89,192 ; mulps %xmm8,%xmm8 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 15,89,192 ; mulps %xmm0,%xmm0 + DB 65,15,89,193 ; mulps %xmm9,%xmm0 + DB 65,15,89,192 ; mulps %xmm8,%xmm0 + DB 65,15,89,194 ; mulps %xmm10,%xmm0 + DB 69,15,87,210 ; xorps %xmm10,%xmm10 + DB 65,15,95,194 ; maxps %xmm10,%xmm0 + DB 68,15,82,193 ; rsqrtps %xmm1,%xmm8 + DB 69,15,82,192 ; rsqrtps %xmm8,%xmm8 + DB 69,15,82,192 ; rsqrtps %xmm8,%xmm8 + DB 69,15,82,200 ; rsqrtps %xmm8,%xmm9 + DB 69,15,82,193 ; rsqrtps %xmm9,%xmm8 + DB 69,15,82,216 ; rsqrtps %xmm8,%xmm11 + DB 15,89,201 ; mulps %xmm1,%xmm1 + DB 69,15,40,193 ; movaps %xmm9,%xmm8 + DB 69,15,89,192 ; mulps %xmm8,%xmm8 + DB 69,15,89,193 ; mulps %xmm9,%xmm8 + DB 68,15,89,193 ; mulps %xmm1,%xmm8 + DB 69,15,89,195 ; mulps %xmm11,%xmm8 + DB 69,15,95,194 ; maxps %xmm10,%xmm8 + DB 15,82,202 ; rsqrtps %xmm2,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 68,15,82,217 ; rsqrtps %xmm1,%xmm11 + DB 65,15,82,203 ; rsqrtps %xmm11,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,89,210 ; mulps %xmm2,%xmm2 + DB 69,15,40,203 ; movaps %xmm11,%xmm9 + DB 69,15,89,201 ; mulps %xmm9,%xmm9 + DB 69,15,89,203 ; mulps %xmm11,%xmm9 + DB 68,15,89,202 ; mulps %xmm2,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 69,15,95,202 ; maxps %xmm10,%xmm9 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 65,15,40,209 ; movaps %xmm9,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_to_2dot2_sse41 +_sk_to_2dot2_sse41 LABEL PROC + DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 + DB 65,15,82,192 ; rsqrtps %xmm8,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 68,15,82,200 ; rsqrtps %xmm0,%xmm9 + DB 69,15,83,192 ; rcpps %xmm8,%xmm8 + DB 68,15,89,192 ; mulps %xmm0,%xmm8 + DB 65,15,83,193 ; rcpps %xmm9,%xmm0 + DB 65,15,89,192 ; mulps %xmm8,%xmm0 + DB 69,15,87,192 ; xorps %xmm8,%xmm8 + DB 65,15,95,192 ; maxps %xmm8,%xmm0 + DB 68,15,82,201 ; rsqrtps %xmm1,%xmm9 + DB 65,15,82,201 ; rsqrtps %xmm9,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 68,15,82,209 ; rsqrtps %xmm1,%xmm10 + DB 69,15,83,201 ; rcpps %xmm9,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 65,15,83,202 ; rcpps %xmm10,%xmm1 + DB 65,15,89,201 ; mulps %xmm9,%xmm1 + DB 65,15,95,200 ; maxps %xmm8,%xmm1 + DB 68,15,82,202 ; rsqrtps %xmm2,%xmm9 + DB 65,15,82,209 ; rsqrtps %xmm9,%xmm2 + DB 15,82,210 ; rsqrtps %xmm2,%xmm2 + DB 15,82,210 ; rsqrtps %xmm2,%xmm2 + DB 15,82,210 ; rsqrtps %xmm2,%xmm2 + DB 68,15,82,210 ; rsqrtps %xmm2,%xmm10 + DB 69,15,83,201 ; rcpps %xmm9,%xmm9 + DB 68,15,89,202 ; mulps %xmm2,%xmm9 + DB 65,15,83,210 ; rcpps %xmm10,%xmm2 + DB 65,15,89,209 ; mulps %xmm9,%xmm2 + DB 65,15,95,208 ; maxps %xmm8,%xmm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_scale_1_float_sse41 _sk_scale_1_float_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -8491,6 +8738,93 @@ _sk_to_srgb_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_from_2dot2_sse2 +_sk_from_2dot2_sse2 LABEL PROC + DB 68,15,40,192 ; movaps %xmm0,%xmm8 + DB 65,15,82,192 ; rsqrtps %xmm8,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 68,15,82,200 ; rsqrtps %xmm0,%xmm9 + DB 65,15,82,193 ; rsqrtps %xmm9,%xmm0 + DB 68,15,82,208 ; rsqrtps %xmm0,%xmm10 + DB 69,15,89,192 ; mulps %xmm8,%xmm8 + DB 65,15,40,193 ; movaps %xmm9,%xmm0 + DB 15,89,192 ; mulps %xmm0,%xmm0 + DB 65,15,89,193 ; mulps %xmm9,%xmm0 + DB 65,15,89,192 ; mulps %xmm8,%xmm0 + DB 65,15,89,194 ; mulps %xmm10,%xmm0 + DB 69,15,87,210 ; xorps %xmm10,%xmm10 + DB 65,15,95,194 ; maxps %xmm10,%xmm0 + DB 68,15,82,193 ; rsqrtps %xmm1,%xmm8 + DB 69,15,82,192 ; rsqrtps %xmm8,%xmm8 + DB 69,15,82,192 ; rsqrtps %xmm8,%xmm8 + DB 69,15,82,200 ; rsqrtps %xmm8,%xmm9 + DB 69,15,82,193 ; rsqrtps %xmm9,%xmm8 + DB 69,15,82,216 ; rsqrtps %xmm8,%xmm11 + DB 15,89,201 ; mulps %xmm1,%xmm1 + DB 69,15,40,193 ; movaps %xmm9,%xmm8 + DB 69,15,89,192 ; mulps %xmm8,%xmm8 + DB 69,15,89,193 ; mulps %xmm9,%xmm8 + DB 68,15,89,193 ; mulps %xmm1,%xmm8 + DB 69,15,89,195 ; mulps %xmm11,%xmm8 + DB 69,15,95,194 ; maxps %xmm10,%xmm8 + DB 15,82,202 ; rsqrtps %xmm2,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 68,15,82,217 ; rsqrtps %xmm1,%xmm11 + DB 65,15,82,203 ; rsqrtps %xmm11,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,89,210 ; mulps %xmm2,%xmm2 + DB 69,15,40,203 ; movaps %xmm11,%xmm9 + DB 69,15,89,201 ; mulps %xmm9,%xmm9 + DB 69,15,89,203 ; mulps %xmm11,%xmm9 + DB 68,15,89,202 ; mulps %xmm2,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 69,15,95,202 ; maxps %xmm10,%xmm9 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 65,15,40,200 ; movaps %xmm8,%xmm1 + DB 65,15,40,209 ; movaps %xmm9,%xmm2 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_to_2dot2_sse2 +_sk_to_2dot2_sse2 LABEL PROC + DB 68,15,82,192 ; rsqrtps %xmm0,%xmm8 + DB 65,15,82,192 ; rsqrtps %xmm8,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 15,82,192 ; rsqrtps %xmm0,%xmm0 + DB 68,15,82,200 ; rsqrtps %xmm0,%xmm9 + DB 69,15,83,192 ; rcpps %xmm8,%xmm8 + DB 68,15,89,192 ; mulps %xmm0,%xmm8 + DB 65,15,83,193 ; rcpps %xmm9,%xmm0 + DB 65,15,89,192 ; mulps %xmm8,%xmm0 + DB 69,15,87,192 ; xorps %xmm8,%xmm8 + DB 65,15,95,192 ; maxps %xmm8,%xmm0 + DB 68,15,82,201 ; rsqrtps %xmm1,%xmm9 + DB 65,15,82,201 ; rsqrtps %xmm9,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 15,82,201 ; rsqrtps %xmm1,%xmm1 + DB 68,15,82,209 ; rsqrtps %xmm1,%xmm10 + DB 69,15,83,201 ; rcpps %xmm9,%xmm9 + DB 68,15,89,201 ; mulps %xmm1,%xmm9 + DB 65,15,83,202 ; rcpps %xmm10,%xmm1 + DB 65,15,89,201 ; mulps %xmm9,%xmm1 + DB 65,15,95,200 ; maxps %xmm8,%xmm1 + DB 68,15,82,202 ; rsqrtps %xmm2,%xmm9 + DB 65,15,82,209 ; rsqrtps %xmm9,%xmm2 + DB 15,82,210 ; rsqrtps %xmm2,%xmm2 + DB 15,82,210 ; rsqrtps %xmm2,%xmm2 + DB 15,82,210 ; rsqrtps %xmm2,%xmm2 + DB 68,15,82,210 ; rsqrtps %xmm2,%xmm10 + DB 69,15,83,201 ; rcpps %xmm9,%xmm9 + DB 68,15,89,202 ; mulps %xmm2,%xmm9 + DB 65,15,83,210 ; rcpps %xmm10,%xmm2 + DB 65,15,89,209 ; mulps %xmm9,%xmm2 + DB 65,15,95,208 ; maxps %xmm8,%xmm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_scale_1_float_sse2 _sk_scale_1_float_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax -- cgit v1.2.3