diff options
| author |  Mike Klein <mtklein@chromium.org> | 2017-05-31 16:59:00 -0400 |
|---|---|---|
| committer |  Skia Commit-Bot <skia-commit-bot@chromium.org> | 2017-06-01 02:26:50 +0000 |
| commit | 823103384cada2e3b84726424b44e6cd80bee5fd (patch) | |
| tree | a7388fc6c0bbadb00ac311d0fa9f480555693ff3 /src/jumper | |
| parent | 8bf1f9ffcf4b0168a1a05399eb0ed4e5e04eab80 (diff) | |
reland: We can mask load and store with just AVX
Originally reviewed here: https://skia-review.googlesource.com/c/17452/
CQ_INCLUDE_TRYBOTS=skia.primary:Test-Ubuntu-GCC-ShuttleA-GPU-GTX550Ti-x86_64-Release-Valgrind
Change-Id: I2e593e897ce93147ec593c2a5de143217274ba2a
Reviewed-on: https://skia-review.googlesource.com/18267
Reviewed-by: Mike Klein <mtklein@chromium.org>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/jumper')
| -rw-r--r-- | src/jumper/SkJumper_generated.S | 1481 | ||||
| -rw-r--r-- | src/jumper/SkJumper_generated_win.S | 1475 | ||||
| -rw-r--r-- | src/jumper/SkJumper_stages.cpp | 12 |
3 files changed, 1402 insertions, 1566 deletions
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index 7fe532625a..aebbeb4754 100644 --- a/src/jumper/SkJumper_generated.S +++ b/src/jumper/SkJumper_generated.S @@ -10847,8 +10847,8 @@ _sk_srcover_rgba_8888_hsw: .byte 76,3,8 // add (%rax),%r9 .byte 77,133,192 // test %r8,%r8 .byte 15,133,180,0,0,0 // jne 1345 <_sk_srcover_rgba_8888_hsw+0xcd> - .byte 196,193,126,111,57 // vmovdqu (%r9),%ymm7 - .byte 197,197,219,37,98,59,0,0 // vpand 0x3b62(%rip),%ymm7,%ymm4 # 4e00 <_sk_callback_hsw+0x53e> + .byte 196,193,124,16,57 // vmovups (%r9),%ymm7 + .byte 197,196,84,37,98,59,0,0 // vandps 0x3b62(%rip),%ymm7,%ymm4 # 4e00 <_sk_callback_hsw+0x53e> .byte 197,252,91,228 // vcvtdq2ps %ymm4,%ymm4 .byte 196,226,69,0,45,117,59,0,0 // vpshufb 0x3b75(%rip),%ymm7,%ymm5 # 4e20 <_sk_callback_hsw+0x55e> .byte 197,252,91,237 // vcvtdq2ps %ymm5,%ymm5 @@ -10879,7 +10879,7 @@ _sk_srcover_rgba_8888_hsw: .byte 196,65,61,235,193 // vpor %ymm9,%ymm8,%ymm8 .byte 77,133,192 // test %r8,%r8 .byte 117,53 // jne 136e <_sk_srcover_rgba_8888_hsw+0xf6> - .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9) + .byte 196,65,124,17,1 // vmovups %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,193 // mov %r8,%rcx .byte 255,224 // jmpq *%rax @@ -10890,7 +10890,7 @@ _sk_srcover_rgba_8888_hsw: .byte 72,211,232 // shr %cl,%rax .byte 196,225,249,110,224 // vmovq %rax,%xmm4 .byte 196,226,125,33,228 // vpmovsxbd %xmm4,%ymm4 - .byte 196,194,93,140,57 // vpmaskmovd (%r9),%ymm4,%ymm7 + .byte 196,194,93,44,57 // vmaskmovps (%r9),%ymm4,%ymm7 .byte 233,40,255,255,255 // jmpq 1296 <_sk_srcover_rgba_8888_hsw+0x1e> .byte 185,8,0,0,0 // mov $0x8,%ecx .byte 68,41,193 // sub %r8d,%ecx @@ -10899,7 +10899,7 @@ _sk_srcover_rgba_8888_hsw: .byte 72,211,232 // shr %cl,%rax .byte 196,97,249,110,200 // vmovq %rax,%xmm9 .byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9 - .byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9) + .byte 196,66,53,46,1 // vmaskmovps %ymm8,%ymm9,(%r9) .byte 235,170 // jmp 133e <_sk_srcover_rgba_8888_hsw+0xc6> HIDDEN _sk_clamp_0_hsw @@ -11418,8 +11418,8 @@ _sk_load_tables_hsw: .byte 76,3,8 // add (%rax),%r9 .byte 77,133,192 // test %r8,%r8 .byte 117,105 // jne 1b6e <_sk_load_tables_hsw+0x7e> - .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 - .byte 197,229,219,13,78,51,0,0 // vpand 0x334e(%rip),%ymm3,%ymm1 # 4e60 <_sk_callback_hsw+0x59e> + .byte 196,193,124,16,25 // vmovups (%r9),%ymm3 + .byte 197,228,84,13,78,51,0,0 // vandps 0x334e(%rip),%ymm3,%ymm1 # 4e60 <_sk_callback_hsw+0x59e> .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 .byte 72,139,72,8 // mov 0x8(%rax),%rcx .byte 76,139,72,16 // mov 0x10(%rax),%r9 @@ -11445,7 +11445,7 @@ _sk_load_tables_hsw: .byte 73,211,234 // shr %cl,%r10 .byte 196,193,249,110,194 // vmovq %r10,%xmm0 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 - .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 + .byte 196,194,125,44,25 // vmaskmovps (%r9),%ymm0,%ymm3 .byte 233,115,255,255,255 // jmpq 1b0a <_sk_load_tables_hsw+0x1a> HIDDEN _sk_load_tables_u16_be_hsw @@ -12970,8 +12970,8 @@ _sk_load_8888_hsw: .byte 76,3,8 // add (%rax),%r9 .byte 77,133,192 // test %r8,%r8 .byte 117,88 // jne 3491 <_sk_load_8888_hsw+0x6d> - .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 - .byte 197,229,219,5,218,26,0,0 // vpand 0x1ada(%rip),%ymm3,%ymm0 # 4f20 <_sk_callback_hsw+0x65e> + .byte 196,193,124,16,25 // vmovups (%r9),%ymm3 + .byte 197,228,84,5,218,26,0,0 // vandps 0x1ada(%rip),%ymm3,%ymm0 # 4f20 <_sk_callback_hsw+0x65e> .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 .byte 196,98,125,24,5,145,24,0,0 // vbroadcastss 0x1891(%rip),%ymm8 # 4ce4 <_sk_callback_hsw+0x422> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 @@ -12994,7 +12994,7 @@ _sk_load_8888_hsw: .byte 72,211,232 // shr %cl,%rax .byte 196,225,249,110,192 // vmovq %rax,%xmm0 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 - .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 + .byte 196,194,125,44,25 // vmaskmovps (%r9),%ymm0,%ymm3 .byte 235,135 // jmp 343e <_sk_load_8888_hsw+0x1a> HIDDEN _sk_gather_8888_hsw @@ -13051,7 +13051,7 @@ _sk_store_8888_hsw: .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8 .byte 77,133,192 // test %r8,%r8 .byte 117,12 // jne 35a0 <_sk_store_8888_hsw+0x73> - .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9) + .byte 196,65,124,17,1 // vmovups %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,137,193 // mov %r8,%rcx .byte 255,224 // jmpq *%rax @@ -13062,7 +13062,7 @@ _sk_store_8888_hsw: .byte 72,211,232 // shr %cl,%rax .byte 196,97,249,110,200 // vmovq %rax,%xmm9 .byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9 - .byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9) + .byte 196,66,53,46,1 // vmaskmovps %ymm8,%ymm9,(%r9) .byte 235,211 // jmp 3599 <_sk_store_8888_hsw+0x6c> HIDDEN _sk_load_f16_hsw @@ -15093,14 +15093,14 @@ _sk_seed_shader_avx: .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,171,102,0,0 // vbroadcastss 0x66ab(%rip),%ymm1 # 677c <_sk_callback_avx+0x128> + .byte 196,226,125,24,13,211,100,0,0 // vbroadcastss 0x64d3(%rip),%ymm1 # 65a4 <_sk_callback_avx+0x128> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,21,143,102,0,0 // vbroadcastss 0x668f(%rip),%ymm2 # 6780 <_sk_callback_avx+0x12c> + .byte 196,226,125,24,21,183,100,0,0 // vbroadcastss 0x64b7(%rip),%ymm2 # 65a8 <_sk_callback_avx+0x12c> .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 @@ -15123,7 +15123,7 @@ _sk_dither_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 196,66,125,24,8 // vbroadcastss (%r8),%ymm9 .byte 196,65,60,87,209 // vxorps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,29,64,102,0,0 // vbroadcastss 0x6640(%rip),%ymm11 # 6784 <_sk_callback_avx+0x130> + .byte 196,98,125,24,29,104,100,0,0 // vbroadcastss 0x6468(%rip),%ymm11 # 65ac <_sk_callback_avx+0x130> .byte 196,65,44,84,203 // vandps %ymm11,%ymm10,%ymm9 .byte 196,193,25,114,241,5 // vpslld $0x5,%xmm9,%xmm12 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 @@ -15134,8 +15134,8 @@ _sk_dither_avx: .byte 196,67,125,25,219,1 // vextractf128 $0x1,%ymm11,%xmm11 .byte 196,193,33,114,243,4 // vpslld $0x4,%xmm11,%xmm11 .byte 196,67,29,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm12,%ymm11 - .byte 196,98,125,24,37,1,102,0,0 // vbroadcastss 0x6601(%rip),%ymm12 # 6788 <_sk_callback_avx+0x134> - .byte 196,98,125,24,45,252,101,0,0 // vbroadcastss 0x65fc(%rip),%ymm13 # 678c <_sk_callback_avx+0x138> + .byte 196,98,125,24,37,41,100,0,0 // vbroadcastss 0x6429(%rip),%ymm12 # 65b0 <_sk_callback_avx+0x134> + .byte 196,98,125,24,45,36,100,0,0 // vbroadcastss 0x6424(%rip),%ymm13 # 65b4 <_sk_callback_avx+0x138> .byte 196,65,44,84,245 // vandps %ymm13,%ymm10,%ymm14 .byte 196,193,1,114,246,2 // vpslld $0x2,%xmm14,%xmm15 .byte 196,67,125,25,246,1 // vextractf128 $0x1,%ymm14,%xmm14 @@ -15162,9 +15162,9 @@ _sk_dither_avx: .byte 196,65,12,86,202 // vorps %ymm10,%ymm14,%ymm9 .byte 196,65,60,86,193 // vorps %ymm9,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,103,101,0,0 // vbroadcastss 0x6567(%rip),%ymm9 # 6790 <_sk_callback_avx+0x13c> + .byte 196,98,125,24,13,143,99,0,0 // vbroadcastss 0x638f(%rip),%ymm9 # 65b8 <_sk_callback_avx+0x13c> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,93,101,0,0 // vbroadcastss 0x655d(%rip),%ymm9 # 6794 <_sk_callback_avx+0x140> + .byte 196,98,125,24,13,133,99,0,0 // vbroadcastss 0x6385(%rip),%ymm9 # 65bc <_sk_callback_avx+0x140> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 196,98,125,24,72,8 // vbroadcastss 0x8(%rax),%ymm9 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 @@ -15233,7 +15233,7 @@ HIDDEN _sk_srcatop_avx FUNCTION(_sk_srcatop_avx) _sk_srcatop_avx: .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 - .byte 196,98,125,24,5,180,100,0,0 // vbroadcastss 0x64b4(%rip),%ymm8 # 6798 <_sk_callback_avx+0x144> + .byte 196,98,125,24,5,220,98,0,0 // vbroadcastss 0x62dc(%rip),%ymm8 # 65c0 <_sk_callback_avx+0x144> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -15254,7 +15254,7 @@ HIDDEN _sk_dstatop_avx FUNCTION(_sk_dstatop_avx) _sk_dstatop_avx: .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 - .byte 196,98,125,24,13,118,100,0,0 // vbroadcastss 0x6476(%rip),%ymm9 # 679c <_sk_callback_avx+0x148> + .byte 196,98,125,24,13,158,98,0,0 // vbroadcastss 0x629e(%rip),%ymm9 # 65c4 <_sk_callback_avx+0x148> .byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 @@ -15296,7 +15296,7 @@ HIDDEN _sk_srcout_avx .globl _sk_srcout_avx FUNCTION(_sk_srcout_avx) _sk_srcout_avx: - .byte 196,98,125,24,5,21,100,0,0 // vbroadcastss 0x6415(%rip),%ymm8 # 67a0 <_sk_callback_avx+0x14c> + .byte 196,98,125,24,5,61,98,0,0 // vbroadcastss 0x623d(%rip),%ymm8 # 65c8 <_sk_callback_avx+0x14c> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -15309,7 +15309,7 @@ HIDDEN _sk_dstout_avx .globl _sk_dstout_avx FUNCTION(_sk_dstout_avx) _sk_dstout_avx: - .byte 196,226,125,24,5,248,99,0,0 // vbroadcastss 0x63f8(%rip),%ymm0 # 67a4 <_sk_callback_avx+0x150> + .byte 196,226,125,24,5,32,98,0,0 // vbroadcastss 0x6220(%rip),%ymm0 # 65cc <_sk_callback_avx+0x150> .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 @@ -15322,7 +15322,7 @@ HIDDEN _sk_srcover_avx .globl _sk_srcover_avx FUNCTION(_sk_srcover_avx) _sk_srcover_avx: - .byte 196,98,125,24,5,219,99,0,0 // vbroadcastss 0x63db(%rip),%ymm8 # 67a8 <_sk_callback_avx+0x154> + .byte 196,98,125,24,5,3,98,0,0 // vbroadcastss 0x6203(%rip),%ymm8 # 65d0 <_sk_callback_avx+0x154> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -15339,7 +15339,7 @@ HIDDEN _sk_dstover_avx .globl _sk_dstover_avx FUNCTION(_sk_dstover_avx) _sk_dstover_avx: - .byte 196,98,125,24,5,174,99,0,0 // vbroadcastss 0x63ae(%rip),%ymm8 # 67ac <_sk_callback_avx+0x158> + .byte 196,98,125,24,5,214,97,0,0 // vbroadcastss 0x61d6(%rip),%ymm8 # 65d4 <_sk_callback_avx+0x158> .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 @@ -15367,7 +15367,7 @@ HIDDEN _sk_multiply_avx .globl _sk_multiply_avx FUNCTION(_sk_multiply_avx) _sk_multiply_avx: - .byte 196,98,125,24,5,109,99,0,0 // vbroadcastss 0x636d(%rip),%ymm8 # 67b0 <_sk_callback_avx+0x15c> + .byte 196,98,125,24,5,149,97,0,0 // vbroadcastss 0x6195(%rip),%ymm8 # 65d8 <_sk_callback_avx+0x15c> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -15427,7 +15427,7 @@ HIDDEN _sk_xor__avx .globl _sk_xor__avx FUNCTION(_sk_xor__avx) _sk_xor__avx: - .byte 196,98,125,24,5,188,98,0,0 // vbroadcastss 0x62bc(%rip),%ymm8 # 67b4 <_sk_callback_avx+0x160> + .byte 196,98,125,24,5,228,96,0,0 // vbroadcastss 0x60e4(%rip),%ymm8 # 65dc <_sk_callback_avx+0x160> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -15464,7 +15464,7 @@ _sk_darken_avx: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,60,98,0,0 // vbroadcastss 0x623c(%rip),%ymm8 # 67b8 <_sk_callback_avx+0x164> + .byte 196,98,125,24,5,100,96,0,0 // vbroadcastss 0x6064(%rip),%ymm8 # 65e0 <_sk_callback_avx+0x164> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15490,7 +15490,7 @@ _sk_lighten_avx: .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,232,97,0,0 // vbroadcastss 0x61e8(%rip),%ymm8 # 67bc <_sk_callback_avx+0x168> + .byte 196,98,125,24,5,16,96,0,0 // vbroadcastss 0x6010(%rip),%ymm8 # 65e4 <_sk_callback_avx+0x168> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15519,7 +15519,7 @@ _sk_difference_avx: .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,136,97,0,0 // vbroadcastss 0x6188(%rip),%ymm8 # 67c0 <_sk_callback_avx+0x16c> + .byte 196,98,125,24,5,176,95,0,0 // vbroadcastss 0x5fb0(%rip),%ymm8 # 65e8 <_sk_callback_avx+0x16c> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15542,7 +15542,7 @@ _sk_exclusion_avx: .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 - .byte 196,98,125,24,5,67,97,0,0 // vbroadcastss 0x6143(%rip),%ymm8 # 67c4 <_sk_callback_avx+0x170> + .byte 196,98,125,24,5,107,95,0,0 // vbroadcastss 0x5f6b(%rip),%ymm8 # 65ec <_sk_callback_avx+0x170> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 @@ -15553,7 +15553,7 @@ HIDDEN _sk_colorburn_avx .globl _sk_colorburn_avx FUNCTION(_sk_colorburn_avx) _sk_colorburn_avx: - .byte 196,98,125,24,5,46,97,0,0 // vbroadcastss 0x612e(%rip),%ymm8 # 67c8 <_sk_callback_avx+0x174> + .byte 196,98,125,24,5,86,95,0,0 // vbroadcastss 0x5f56(%rip),%ymm8 # 65f0 <_sk_callback_avx+0x174> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 @@ -15615,7 +15615,7 @@ HIDDEN _sk_colordodge_avx FUNCTION(_sk_colordodge_avx) _sk_colordodge_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 196,98,125,24,13,42,96,0,0 // vbroadcastss 0x602a(%rip),%ymm9 # 67cc <_sk_callback_avx+0x178> + .byte 196,98,125,24,13,82,94,0,0 // vbroadcastss 0x5e52(%rip),%ymm9 # 65f4 <_sk_callback_avx+0x178> .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 @@ -15672,7 +15672,7 @@ HIDDEN _sk_hardlight_avx .globl _sk_hardlight_avx FUNCTION(_sk_hardlight_avx) _sk_hardlight_avx: - .byte 196,98,125,24,5,60,95,0,0 // vbroadcastss 0x5f3c(%rip),%ymm8 # 67d0 <_sk_callback_avx+0x17c> + .byte 196,98,125,24,5,100,93,0,0 // vbroadcastss 0x5d64(%rip),%ymm8 # 65f8 <_sk_callback_avx+0x17c> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -15727,7 +15727,7 @@ HIDDEN _sk_overlay_avx .globl _sk_overlay_avx FUNCTION(_sk_overlay_avx) _sk_overlay_avx: - .byte 196,98,125,24,5,101,94,0,0 // vbroadcastss 0x5e65(%rip),%ymm8 # 67d4 <_sk_callback_avx+0x180> + .byte 196,98,125,24,5,141,92,0,0 // vbroadcastss 0x5c8d(%rip),%ymm8 # 65fc <_sk_callback_avx+0x180> .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -15793,10 +15793,10 @@ _sk_softlight_avx: .byte 196,65,60,88,192 // vaddps %ymm8,%ymm8,%ymm8 .byte 196,65,60,89,216 // vmulps %ymm8,%ymm8,%ymm11 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 - .byte 196,98,125,24,29,92,93,0,0 // vbroadcastss 0x5d5c(%rip),%ymm11 # 67dc <_sk_callback_avx+0x188> + .byte 196,98,125,24,29,132,91,0,0 // vbroadcastss 0x5b84(%rip),%ymm11 # 6604 <_sk_callback_avx+0x188> .byte 196,65,28,88,235 // vaddps %ymm11,%ymm12,%ymm13 .byte 196,65,20,89,192 // vmulps %ymm8,%ymm13,%ymm8 - .byte 196,98,125,24,45,77,93,0,0 // vbroadcastss 0x5d4d(%rip),%ymm13 # 67e0 <_sk_callback_avx+0x18c> + .byte 196,98,125,24,45,117,91,0,0 // vbroadcastss 0x5b75(%rip),%ymm13 # 6608 <_sk_callback_avx+0x18c> .byte 196,65,28,89,245 // vmulps %ymm13,%ymm12,%ymm14 .byte 196,65,12,88,192 // vaddps %ymm8,%ymm14,%ymm8 .byte 196,65,124,82,244 // vrsqrtps %ymm12,%ymm14 @@ -15807,7 +15807,7 @@ _sk_softlight_avx: .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15 .byte 196,67,13,74,240,240 // vblendvps %ymm15,%ymm8,%ymm14,%ymm14 .byte 197,116,88,249 // vaddps %ymm1,%ymm1,%ymm15 - .byte 196,98,125,24,5,11,93,0,0 // vbroadcastss 0x5d0b(%rip),%ymm8 # 67d8 <_sk_callback_avx+0x184> + .byte 196,98,125,24,5,51,91,0,0 // vbroadcastss 0x5b33(%rip),%ymm8 # 6600 <_sk_callback_avx+0x184> .byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12 .byte 197,132,92,195 // vsubps %ymm3,%ymm15,%ymm0 .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 @@ -15934,12 +15934,12 @@ _sk_hue_avx: .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11 .byte 196,65,36,94,222 // vdivps %ymm14,%ymm11,%ymm11 .byte 196,67,37,74,224,240 // vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - .byte 196,98,125,24,53,218,90,0,0 // vbroadcastss 0x5ada(%rip),%ymm14 # 67e4 <_sk_callback_avx+0x190> + .byte 196,98,125,24,53,2,89,0,0 // vbroadcastss 0x5902(%rip),%ymm14 # 660c <_sk_callback_avx+0x190> .byte 196,65,92,89,222 // vmulps %ymm14,%ymm4,%ymm11 - .byte 196,98,125,24,61,208,90,0,0 // vbroadcastss 0x5ad0(%rip),%ymm15 # 67e8 <_sk_callback_avx+0x194> + .byte 196,98,125,24,61,248,88,0,0 // vbroadcastss 0x58f8(%rip),%ymm15 # 6610 <_sk_callback_avx+0x194> .byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 - .byte 196,226,125,24,5,193,90,0,0 // vbroadcastss 0x5ac1(%rip),%ymm0 # 67ec <_sk_callback_avx+0x198> + .byte 196,226,125,24,5,233,88,0,0 // vbroadcastss 0x58e9(%rip),%ymm0 # 6614 <_sk_callback_avx+0x198> .byte 197,76,89,232 // vmulps %ymm0,%ymm6,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 .byte 196,65,52,89,238 // vmulps %ymm14,%ymm9,%ymm13 @@ -16000,7 +16000,7 @@ _sk_hue_avx: .byte 196,65,36,95,208 // vmaxps %ymm8,%ymm11,%ymm10 .byte 196,195,109,74,209,240 // vblendvps %ymm15,%ymm9,%ymm2,%ymm2 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,154,89,0,0 // vbroadcastss 0x599a(%rip),%ymm8 # 67f0 <_sk_callback_avx+0x19c> + .byte 196,98,125,24,5,194,87,0,0 // vbroadcastss 0x57c2(%rip),%ymm8 # 6618 <_sk_callback_avx+0x19c> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -16057,12 +16057,12 @@ _sk_saturation_avx: .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11 .byte 196,65,36,94,222 // vdivps %ymm14,%ymm11,%ymm11 .byte 196,67,37,74,224,240 // vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - .byte 196,98,125,24,53,168,88,0,0 // vbroadcastss 0x58a8(%rip),%ymm14 # 67f4 <_sk_callback_avx+0x1a0> + .byte 196,98,125,24,53,208,86,0,0 // vbroadcastss 0x56d0(%rip),%ymm14 # 661c <_sk_callback_avx+0x1a0> .byte 196,65,92,89,222 // vmulps %ymm14,%ymm4,%ymm11 - .byte 196,98,125,24,61,158,88,0,0 // vbroadcastss 0x589e(%rip),%ymm15 # 67f8 <_sk_callback_avx+0x1a4> + .byte 196,98,125,24,61,198,86,0,0 // vbroadcastss 0x56c6(%rip),%ymm15 # 6620 <_sk_callback_avx+0x1a4> .byte 196,65,84,89,239 // vmulps %ymm15,%ymm5,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 - .byte 196,226,125,24,5,143,88,0,0 // vbroadcastss 0x588f(%rip),%ymm0 # 67fc <_sk_callback_avx+0x1a8> + .byte 196,226,125,24,5,183,86,0,0 // vbroadcastss 0x56b7(%rip),%ymm0 # 6624 <_sk_callback_avx+0x1a8> .byte 197,76,89,232 // vmulps %ymm0,%ymm6,%ymm13 .byte 196,65,36,88,221 // vaddps %ymm13,%ymm11,%ymm11 .byte 196,65,52,89,238 // vmulps %ymm14,%ymm9,%ymm13 @@ -16123,7 +16123,7 @@ _sk_saturation_avx: .byte 196,65,36,95,208 // vmaxps %ymm8,%ymm11,%ymm10 .byte 196,195,109,74,209,240 // vblendvps %ymm15,%ymm9,%ymm2,%ymm2 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,104,87,0,0 // vbroadcastss 0x5768(%rip),%ymm8 # 6800 <_sk_callback_avx+0x1ac> + .byte 196,98,125,24,5,144,85,0,0 // vbroadcastss 0x5590(%rip),%ymm8 # 6628 <_sk_callback_avx+0x1ac> .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 @@ -16152,12 +16152,12 @@ _sk_color_avx: .byte 197,252,17,68,36,168 // vmovups %ymm0,-0x58(%rsp) .byte 197,124,89,199 // vmulps %ymm7,%ymm0,%ymm8 .byte 197,116,89,207 // vmulps %ymm7,%ymm1,%ymm9 - .byte 196,98,125,24,45,254,86,0,0 // vbroadcastss 0x56fe(%rip),%ymm13 # 6804 <_sk_callback_avx+0x1b0> + .byte 196,98,125,24,45,38,85,0,0 // vbroadcastss 0x5526(%rip),%ymm13 # 662c <_sk_callback_avx+0x1b0> .byte 196,65,92,89,213 // vmulps %ymm13,%ymm4,%ymm10 - .byte 196,98,125,24,53,244,86,0,0 // vbroadcastss 0x56f4(%rip),%ymm14 # 6808 <_sk_callback_avx+0x1b4> + .byte 196,98,125,24,53,28,85,0,0 // vbroadcastss 0x551c(%rip),%ymm14 # 6630 <_sk_callback_avx+0x1b4> .byte 196,65,84,89,222 // vmulps %ymm14,%ymm5,%ymm11 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,24,61,229,86,0,0 // vbroadcastss 0x56e5(%rip),%ymm15 # 680c <_sk_callback_avx+0x1b8> + .byte 196,98,125,24,61,13,85,0,0 // vbroadcastss 0x550d(%rip),%ymm15 # 6634 <_sk_callback_avx+0x1b8> .byte 196,65,76,89,223 // vmulps %ymm15,%ymm6,%ymm11 .byte 196,193,44,88,195 // vaddps %ymm11,%ymm10,%ymm0 .byte 196,65,60,89,221 // vmulps %ymm13,%ymm8,%ymm11 @@ -16220,7 +16220,7 @@ _sk_color_avx: .byte 196,65,44,95,207 // vmaxps %ymm15,%ymm10,%ymm9 .byte 196,195,37,74,192,0 // vblendvps %ymm0,%ymm8,%ymm11,%ymm0 .byte 196,65,124,95,199 // vmaxps %ymm15,%ymm0,%ymm8 - .byte 196,226,125,24,5,172,85,0,0 // vbroadcastss 0x55ac(%rip),%ymm0 # 6810 <_sk_callback_avx+0x1bc> + .byte 196,226,125,24,5,212,83,0,0 // vbroadcastss 0x53d4(%rip),%ymm0 # 6638 <_sk_callback_avx+0x1bc> .byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10 .byte 197,172,89,84,36,168 // vmulps -0x58(%rsp),%ymm10,%ymm2 .byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11 @@ -16250,12 +16250,12 @@ _sk_luminosity_avx: .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 - .byte 196,98,125,24,45,62,85,0,0 // vbroadcastss 0x553e(%rip),%ymm13 # 6814 <_sk_callback_avx+0x1c0> + .byte 196,98,125,24,45,102,83,0,0 // vbroadcastss 0x5366(%rip),%ymm13 # 663c <_sk_callback_avx+0x1c0> .byte 196,65,108,89,213 // vmulps %ymm13,%ymm2,%ymm10 - .byte 196,98,125,24,53,52,85,0,0 // vbroadcastss 0x5534(%rip),%ymm14 # 6818 <_sk_callback_avx+0x1c4> + .byte 196,98,125,24,53,92,83,0,0 // vbroadcastss 0x535c(%rip),%ymm14 # 6640 <_sk_callback_avx+0x1c4> .byte 196,65,116,89,222 // vmulps %ymm14,%ymm1,%ymm11 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,24,61,37,85,0,0 // vbroadcastss 0x5525(%rip),%ymm15 # 681c <_sk_callback_avx+0x1c8> + .byte 196,98,125,24,61,77,83,0,0 // vbroadcastss 0x534d(%rip),%ymm15 # 6644 <_sk_callback_avx+0x1c8> .byte 196,65,28,89,223 // vmulps %ymm15,%ymm12,%ymm11 .byte 196,193,44,88,195 // vaddps %ymm11,%ymm10,%ymm0 .byte 196,65,60,89,221 // vmulps %ymm13,%ymm8,%ymm11 @@ -16318,7 +16318,7 @@ _sk_luminosity_avx: .byte 196,65,44,95,207 // vmaxps %ymm15,%ymm10,%ymm9 .byte 196,195,37,74,192,0 // vblendvps %ymm0,%ymm8,%ymm11,%ymm0 .byte 196,65,124,95,199 // vmaxps %ymm15,%ymm0,%ymm8 - .byte 196,226,125,24,5,236,83,0,0 // vbroadcastss 0x53ec(%rip),%ymm0 # 6820 <_sk_callback_avx+0x1cc> + .byte 196,226,125,24,5,20,82,0,0 // vbroadcastss 0x5214(%rip),%ymm0 # 6648 <_sk_callback_avx+0x1cc> .byte 197,124,92,215 // vsubps %ymm7,%ymm0,%ymm10 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 197,124,92,219 // vsubps %ymm3,%ymm0,%ymm11 @@ -16342,32 +16342,34 @@ HIDDEN _sk_srcover_rgba_8888_avx .globl _sk_srcover_rgba_8888_avx FUNCTION(_sk_srcover_rgba_8888_avx) _sk_srcover_rgba_8888_avx: + .byte 73,137,200 // mov %rcx,%r8 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 76,139,16 // mov (%rax),%r10 - .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,46,1,0,0 // jne 15bb <_sk_srcover_rgba_8888_avx+0x13c> - .byte 196,65,124,16,4,186 // vmovups (%r10,%rdi,4),%ymm8 - .byte 197,124,40,13,37,87,0,0 // vmovaps 0x5725(%rip),%ymm9 # 6bc0 <_sk_callback_avx+0x56c> - .byte 196,193,60,84,225 // vandps %ymm9,%ymm8,%ymm4 + .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 + .byte 76,3,8 // add (%rax),%r9 + .byte 77,133,192 // test %r8,%r8 + .byte 15,133,43,1,0,0 // jne 15c3 <_sk_srcover_rgba_8888_avx+0x144> + .byte 196,193,124,16,57 // vmovups (%r9),%ymm7 + .byte 197,124,40,13,219,85,0,0 // vmovaps 0x55db(%rip),%ymm9 # 6a80 <_sk_callback_avx+0x604> + .byte 196,193,68,84,225 // vandps %ymm9,%ymm7,%ymm4 .byte 197,252,91,228 // vcvtdq2ps %ymm4,%ymm4 - .byte 196,193,81,114,208,8 // vpsrld $0x8,%xmm8,%xmm5 - .byte 196,99,125,25,199,1 // vextractf128 $0x1,%ymm8,%xmm7 - .byte 197,201,114,215,8 // vpsrld $0x8,%xmm7,%xmm6 + .byte 197,209,114,215,8 // vpsrld $0x8,%xmm7,%xmm5 + .byte 196,195,125,25,248,1 // vextractf128 $0x1,%ymm7,%xmm8 + .byte 196,193,73,114,208,8 // vpsrld $0x8,%xmm8,%xmm6 .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 .byte 196,193,84,84,233 // vandps %ymm9,%ymm5,%ymm5 .byte 197,252,91,237 // vcvtdq2ps %ymm5,%ymm5 - .byte 196,193,41,114,208,16 // vpsrld $0x10,%xmm8,%xmm10 - .byte 197,201,114,215,16 // vpsrld $0x10,%xmm7,%xmm6 + .byte 197,169,114,215,16 // vpsrld $0x10,%xmm7,%xmm10 + .byte 196,193,73,114,208,16 // vpsrld $0x10,%xmm8,%xmm6 .byte 196,227,45,24,246,1 // vinsertf128 $0x1,%xmm6,%ymm10,%ymm6 .byte 196,193,76,84,241 // vandps %ymm9,%ymm6,%ymm6 .byte 197,252,91,246 // vcvtdq2ps %ymm6,%ymm6 - .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8 - .byte 197,193,114,215,24 // vpsrld $0x18,%xmm7,%xmm7 - .byte 196,227,61,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm8,%ymm7 + .byte 197,177,114,215,24 // vpsrld $0x18,%xmm7,%xmm9 + .byte 196,193,65,114,208,24 // vpsrld $0x18,%xmm8,%xmm7 + .byte 196,227,53,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm9,%ymm7 .byte 197,252,91,255 // vcvtdq2ps %ymm7,%ymm7 - .byte 196,98,125,24,5,40,83,0,0 // vbroadcastss 0x5328(%rip),%ymm8 # 6824 <_sk_callback_avx+0x1d0> + .byte 196,98,125,24,5,70,81,0,0 // vbroadcastss 0x5146(%rip),%ymm8 # 664c <_sk_callback_avx+0x1d0> .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 - .byte 196,98,125,24,13,31,83,0,0 // vbroadcastss 0x531f(%rip),%ymm9 # 6828 <_sk_callback_avx+0x1d4> + .byte 196,98,125,24,13,61,81,0,0 // vbroadcastss 0x513d(%rip),%ymm9 # 6650 <_sk_callback_avx+0x1d4> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 @@ -16399,97 +16401,40 @@ _sk_srcover_rgba_8888_avx: .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 .byte 196,65,53,86,202 // vorpd %ymm10,%ymm9,%ymm9 .byte 196,65,61,86,193 // vorpd %ymm9,%ymm8,%ymm8 - .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,183,0,0,0 // jne 1668 <_sk_srcover_rgba_8888_avx+0x1e9> - .byte 196,65,124,17,4,186 // vmovups %ymm8,(%r10,%rdi,4) + .byte 77,133,192 // test %r8,%r8 + .byte 117,87 // jne 160e <_sk_srcover_rgba_8888_avx+0x18f> + .byte 196,65,124,17,1 // vmovups %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,137,193 // mov %r8,%rcx .byte 255,224 // jmpq *%rax - .byte 65,137,200 // mov %ecx,%r8d - .byte 65,128,224,7 // and $0x7,%r8b - .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 65,254,200 // dec %r8b - .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,191,254,255,255 // ja 1493 <_sk_srcover_rgba_8888_avx+0x14> - .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,253,0,0,0 // lea 0xfd(%rip),%r9 # 16dc <_sk_srcover_rgba_8888_avx+0x25d> - .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax - .byte 76,1,200 // add %r9,%rax - .byte 255,224 // jmpq *%rax - .byte 196,193,121,110,100,186,24 // vmovd 0x18(%r10,%rdi,4),%xmm4 - .byte 197,249,112,228,68 // vpshufd $0x44,%xmm4,%xmm4 - .byte 196,227,125,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm0,%ymm4 - .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 - .byte 196,99,85,12,196,64 // vblendps $0x40,%ymm4,%ymm5,%ymm8 - .byte 196,99,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm4 - .byte 196,195,89,34,100,186,20,1 // vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm4,%xmm4 - .byte 196,99,61,24,196,1 // vinsertf128 $0x1,%xmm4,%ymm8,%ymm8 - .byte 196,99,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm4 - .byte 196,195,89,34,100,186,16,0 // vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm4,%xmm4 - .byte 196,99,61,24,196,1 // vinsertf128 $0x1,%xmm4,%ymm8,%ymm8 - .byte 196,195,57,34,100,186,12,3 // vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm8,%xmm4 - .byte 196,99,61,12,196,15 // vblendps $0xf,%ymm4,%ymm8,%ymm8 - .byte 196,195,57,34,100,186,8,2 // vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm8,%xmm4 - .byte 196,99,61,12,196,15 // vblendps $0xf,%ymm4,%ymm8,%ymm8 - .byte 196,195,57,34,100,186,4,1 // vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm8,%xmm4 - .byte 196,99,61,12,196,15 // vblendps $0xf,%ymm4,%ymm8,%ymm8 - .byte 196,195,57,34,36,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm8,%xmm4 - .byte 196,99,61,12,196,15 // vblendps $0xf,%ymm4,%ymm8,%ymm8 - .byte 233,43,254,255,255 // jmpq 1493 <_sk_srcover_rgba_8888_avx+0x14> - .byte 65,137,200 // mov %ecx,%r8d - .byte 65,128,224,7 // and $0x7,%r8b - .byte 65,254,200 // dec %r8b - .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,59,255,255,255 // ja 15b7 <_sk_srcover_rgba_8888_avx+0x138> - .byte 65,15,182,192 // movzbl %r8b,%eax - .byte 76,141,5,113,0,0,0 // lea 0x71(%rip),%r8 # 16f8 <_sk_srcover_rgba_8888_avx+0x279> - .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax - .byte 76,1,192 // add %r8,%rax - .byte 255,224 // jmpq *%rax - .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 - .byte 196,67,121,22,76,186,24,2 // vpextrd $0x2,%xmm9,0x18(%r10,%rdi,4) - .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 - .byte 196,67,121,22,76,186,20,1 // vpextrd $0x1,%xmm9,0x14(%r10,%rdi,4) - .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 - .byte 196,65,122,17,76,186,16 // vmovss %xmm9,0x10(%r10,%rdi,4) - .byte 196,67,121,22,68,186,12,3 // vpextrd $0x3,%xmm8,0xc(%r10,%rdi,4) - .byte 196,67,121,22,68,186,8,2 // vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4) - .byte 196,67,121,22,68,186,4,1 // vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4) - .byte 196,65,121,126,4,186 // vmovd %xmm8,(%r10,%rdi,4) - .byte 233,219,254,255,255 // jmpq 15b7 <_sk_srcover_rgba_8888_avx+0x138> - .byte 122,255 // jp 16dd <_sk_srcover_rgba_8888_avx+0x25e> - .byte 255 // (bad) - .byte 255,108,255,255 // ljmp *-0x1(%rdi,%rdi,8) - .byte 255,94,255 // lcall *-0x1(%rsi) - .byte 255 // (bad) - .byte 255,80,255 // callq *-0x1(%rax) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 60,255 // cmp $0xff,%al - .byte 255 // (bad) - .byte 255,40 // ljmp *(%rax) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,12,255 // decl (%rdi,%rdi,8) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 217,255 // fcos - .byte 255 // (bad) - .byte 255,209 // callq *%rcx - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,201 // dec %ecx - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,193 // inc %ecx - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,180,255,255,255,166,255 // pushq -0x590001(%rdi,%rdi,8) - .byte 255 // (bad) - .byte 255 // .byte 0xff - .byte 152 // cwtl - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // .byte 0xff + .byte 185,8,0,0,0 // mov $0x8,%ecx + .byte 68,41,193 // sub %r8d,%ecx + .byte 192,225,3 // shl $0x3,%cl + .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax + .byte 72,211,232 // shr %cl,%rax + .byte 196,225,249,110,224 // vmovq %rax,%xmm4 + .byte 196,226,121,48,228 // vpmovzxbw %xmm4,%xmm4 + .byte 196,226,89,0,45,245,83,0,0 // vpshufb 0x53f5(%rip),%xmm4,%xmm5 # 69e0 <_sk_callback_avx+0x564> + .byte 196,226,121,33,237 // vpmovsxbd %xmm5,%xmm5 + .byte 196,226,89,0,37,247,83,0,0 // vpshufb 0x53f7(%rip),%xmm4,%xmm4 # 69f0 <_sk_callback_avx+0x574> + .byte 196,226,121,33,228 // vpmovsxbd %xmm4,%xmm4 + .byte 196,227,85,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 + .byte 196,194,93,44,57 // vmaskmovps (%r9),%ymm4,%ymm7 + .byte 233,143,254,255,255 // jmpq 149d <_sk_srcover_rgba_8888_avx+0x1e> + .byte 185,8,0,0,0 // mov $0x8,%ecx + .byte 68,41,193 // sub %r8d,%ecx + .byte 192,225,3 // shl $0x3,%cl + .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax + .byte 72,211,232 // shr %cl,%rax + .byte 196,97,249,110,200 // vmovq %rax,%xmm9 + .byte 196,66,121,48,201 // vpmovzxbw %xmm9,%xmm9 + .byte 196,98,49,0,21,170,83,0,0 // vpshufb 0x53aa(%rip),%xmm9,%xmm10 # 69e0 <_sk_callback_avx+0x564> + .byte 196,66,121,33,210 // vpmovsxbd %xmm10,%xmm10 + .byte 196,98,49,0,13,172,83,0,0 // vpshufb 0x53ac(%rip),%xmm9,%xmm9 # 69f0 <_sk_callback_avx+0x574> + .byte 196,66,121,33,201 // vpmovsxbd %xmm9,%xmm9 + .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 + .byte 196,66,53,46,1 // vmaskmovps %ymm8,%ymm9,(%r9) + .byte 233,99,255,255,255 // jmpq 15bc <_sk_srcover_rgba_8888_avx+0x13d> HIDDEN _sk_clamp_0_avx .globl _sk_clamp_0_avx @@ -16507,7 +16452,7 @@ HIDDEN _sk_clamp_1_avx .globl _sk_clamp_1_avx FUNCTION(_sk_clamp_1_avx) _sk_clamp_1_avx: - .byte 196,98,125,24,5,242,80,0,0 // vbroadcastss 0x50f2(%rip),%ymm8 # 682c <_sk_callback_avx+0x1d8> + .byte 196,98,125,24,5,213,79,0,0 // vbroadcastss 0x4fd5(%rip),%ymm8 # 6654 <_sk_callback_avx+0x1d8> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 @@ -16519,7 +16464,7 @@ HIDDEN _sk_clamp_a_avx .globl _sk_clamp_a_avx FUNCTION(_sk_clamp_a_avx) _sk_clamp_a_avx: - .byte 196,98,125,24,5,213,80,0,0 // vbroadcastss 0x50d5(%rip),%ymm8 # 6830 <_sk_callback_avx+0x1dc> + .byte 196,98,125,24,5,184,79,0,0 // vbroadcastss 0x4fb8(%rip),%ymm8 # 6658 <_sk_callback_avx+0x1dc> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 @@ -16605,7 +16550,7 @@ FUNCTION(_sk_unpremul_avx) _sk_unpremul_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 - .byte 196,98,125,24,21,29,80,0,0 // vbroadcastss 0x501d(%rip),%ymm10 # 6834 <_sk_callback_avx+0x1e0> + .byte 196,98,125,24,21,0,79,0,0 // vbroadcastss 0x4f00(%rip),%ymm10 # 665c <_sk_callback_avx+0x1e0> .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 @@ -16618,17 +16563,17 @@ HIDDEN _sk_from_srgb_avx .globl _sk_from_srgb_avx FUNCTION(_sk_from_srgb_avx) _sk_from_srgb_avx: - .byte 196,98,125,24,5,254,79,0,0 // vbroadcastss 0x4ffe(%rip),%ymm8 # 6838 <_sk_callback_avx+0x1e4> + .byte 196,98,125,24,5,225,78,0,0 // vbroadcastss 0x4ee1(%rip),%ymm8 # 6660 <_sk_callback_avx+0x1e4> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 - .byte 196,98,125,24,29,240,79,0,0 // vbroadcastss 0x4ff0(%rip),%ymm11 # 683c <_sk_callback_avx+0x1e8> + .byte 196,98,125,24,29,211,78,0,0 // vbroadcastss 0x4ed3(%rip),%ymm11 # 6664 <_sk_callback_avx+0x1e8> .byte 196,65,124,89,227 // vmulps %ymm11,%ymm0,%ymm12 - .byte 196,98,125,24,45,230,79,0,0 // vbroadcastss 0x4fe6(%rip),%ymm13 # 6840 <_sk_callback_avx+0x1ec> + .byte 196,98,125,24,45,201,78,0,0 // vbroadcastss 0x4ec9(%rip),%ymm13 # 6668 <_sk_callback_avx+0x1ec> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,37,215,79,0,0 // vbroadcastss 0x4fd7(%rip),%ymm12 # 6844 <_sk_callback_avx+0x1f0> + .byte 196,98,125,24,37,186,78,0,0 // vbroadcastss 0x4eba(%rip),%ymm12 # 666c <_sk_callback_avx+0x1f0> .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10 - .byte 196,98,125,24,53,205,79,0,0 // vbroadcastss 0x4fcd(%rip),%ymm14 # 6848 <_sk_callback_avx+0x1f4> + .byte 196,98,125,24,53,176,78,0,0 // vbroadcastss 0x4eb0(%rip),%ymm14 # 6670 <_sk_callback_avx+0x1f4> .byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0 .byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0 .byte 196,65,116,89,200 // vmulps %ymm8,%ymm1,%ymm9 @@ -16655,20 +16600,20 @@ HIDDEN _sk_to_srgb_avx FUNCTION(_sk_to_srgb_avx) _sk_to_srgb_avx: .byte 197,124,82,200 // vrsqrtps %ymm0,%ymm9 - .byte 196,98,125,24,5,98,79,0,0 // vbroadcastss 0x4f62(%rip),%ymm8 # 684c <_sk_callback_avx+0x1f8> + .byte 196,98,125,24,5,69,78,0,0 // vbroadcastss 0x4e45(%rip),%ymm8 # 6674 <_sk_callback_avx+0x1f8> .byte 196,65,124,89,208 // vmulps %ymm8,%ymm0,%ymm10 - .byte 196,98,125,24,29,88,79,0,0 // vbroadcastss 0x4f58(%rip),%ymm11 # 6850 <_sk_callback_avx+0x1fc> + .byte 196,98,125,24,29,59,78,0,0 // vbroadcastss 0x4e3b(%rip),%ymm11 # 6678 <_sk_callback_avx+0x1fc> .byte 196,65,52,89,227 // vmulps %ymm11,%ymm9,%ymm12 - .byte 196,98,125,24,45,78,79,0,0 // vbroadcastss 0x4f4e(%rip),%ymm13 # 6854 <_sk_callback_avx+0x200> + .byte 196,98,125,24,45,49,78,0,0 // vbroadcastss 0x4e31(%rip),%ymm13 # 667c <_sk_callback_avx+0x200> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,52,89,228 // vmulps %ymm12,%ymm9,%ymm12 - .byte 196,98,125,24,53,63,79,0,0 // vbroadcastss 0x4f3f(%rip),%ymm14 # 6858 <_sk_callback_avx+0x204> + .byte 196,98,125,24,53,34,78,0,0 // vbroadcastss 0x4e22(%rip),%ymm14 # 6680 <_sk_callback_avx+0x204> .byte 196,65,28,88,230 // vaddps %ymm14,%ymm12,%ymm12 - .byte 196,98,125,24,61,53,79,0,0 // vbroadcastss 0x4f35(%rip),%ymm15 # 685c <_sk_callback_avx+0x208> + .byte 196,98,125,24,61,24,78,0,0 // vbroadcastss 0x4e18(%rip),%ymm15 # 6684 <_sk_callback_avx+0x208> .byte 196,65,52,88,207 // vaddps %ymm15,%ymm9,%ymm9 .byte 196,65,124,83,201 // vrcpps %ymm9,%ymm9 .byte 196,65,52,89,204 // vmulps %ymm12,%ymm9,%ymm9 - .byte 196,98,125,24,37,33,79,0,0 // vbroadcastss 0x4f21(%rip),%ymm12 # 6860 <_sk_callback_avx+0x20c> + .byte 196,98,125,24,37,4,78,0,0 // vbroadcastss 0x4e04(%rip),%ymm12 # 6688 <_sk_callback_avx+0x20c> .byte 196,193,124,194,196,1 // vcmpltps %ymm12,%ymm0,%ymm0 .byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0 .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 @@ -16705,7 +16650,7 @@ _sk_rgb_to_hsl_avx: .byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9 .byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9 .byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10 - .byte 196,98,125,24,29,135,78,0,0 // vbroadcastss 0x4e87(%rip),%ymm11 # 6864 <_sk_callback_avx+0x210> + .byte 196,98,125,24,29,106,77,0,0 // vbroadcastss 0x4d6a(%rip),%ymm11 # 668c <_sk_callback_avx+0x210> .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11 .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12 .byte 196,65,28,89,227 // vmulps %ymm11,%ymm12,%ymm12 @@ -16715,19 +16660,19 @@ _sk_rgb_to_hsl_avx: .byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2 .byte 197,252,92,201 // vsubps %ymm1,%ymm0,%ymm1 .byte 196,193,116,89,203 // vmulps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,96,78,0,0 // vbroadcastss 0x4e60(%rip),%ymm11 # 6870 <_sk_callback_avx+0x21c> + .byte 196,98,125,24,29,67,77,0,0 // vbroadcastss 0x4d43(%rip),%ymm11 # 6698 <_sk_callback_avx+0x21c> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,78,78,0,0 // vbroadcastss 0x4e4e(%rip),%ymm11 # 686c <_sk_callback_avx+0x218> + .byte 196,98,125,24,29,49,77,0,0 // vbroadcastss 0x4d31(%rip),%ymm11 # 6694 <_sk_callback_avx+0x218> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,227,117,74,202,224 // vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,54,78,0,0 // vbroadcastss 0x4e36(%rip),%ymm2 # 6868 <_sk_callback_avx+0x214> + .byte 196,226,125,24,21,25,77,0,0 // vbroadcastss 0x4d19(%rip),%ymm2 # 6690 <_sk_callback_avx+0x214> .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14 .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2 .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0 .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 .byte 196,227,117,74,194,0 // vblendvps %ymm0,%ymm2,%ymm1,%ymm0 .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1 - .byte 196,98,125,24,37,29,78,0,0 // vbroadcastss 0x4e1d(%rip),%ymm12 # 6878 <_sk_callback_avx+0x224> + .byte 196,98,125,24,37,0,77,0,0 // vbroadcastss 0x4d00(%rip),%ymm12 # 66a0 <_sk_callback_avx+0x224> .byte 196,193,116,89,212 // vmulps %ymm12,%ymm1,%ymm2 .byte 197,28,194,226,1 // vcmpltps %ymm2,%ymm12,%ymm12 .byte 196,65,36,92,216 // vsubps %ymm8,%ymm11,%ymm11 @@ -16737,7 +16682,7 @@ _sk_rgb_to_hsl_avx: .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 .byte 196,195,125,74,198,128 // vblendvps %ymm8,%ymm14,%ymm0,%ymm0 .byte 196,195,117,74,206,128 // vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - .byte 196,98,125,24,5,224,77,0,0 // vbroadcastss 0x4de0(%rip),%ymm8 # 6874 <_sk_callback_avx+0x220> + .byte 196,98,125,24,5,195,76,0,0 // vbroadcastss 0x4cc3(%rip),%ymm8 # 669c <_sk_callback_avx+0x220> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -16754,7 +16699,7 @@ _sk_hsl_to_rgb_avx: .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) .byte 197,252,40,225 // vmovaps %ymm1,%ymm4 .byte 197,252,40,216 // vmovaps %ymm0,%ymm3 - .byte 196,98,125,24,5,173,77,0,0 // vbroadcastss 0x4dad(%rip),%ymm8 # 687c <_sk_callback_avx+0x228> + .byte 196,98,125,24,5,144,76,0,0 // vbroadcastss 0x4c90(%rip),%ymm8 # 66a4 <_sk_callback_avx+0x228> .byte 197,60,194,202,2 // vcmpleps %ymm2,%ymm8,%ymm9 .byte 197,92,89,210 // vmulps %ymm2,%ymm4,%ymm10 .byte 196,65,92,92,218 // vsubps %ymm10,%ymm4,%ymm11 @@ -16762,23 +16707,23 @@ _sk_hsl_to_rgb_avx: .byte 197,52,88,210 // vaddps %ymm2,%ymm9,%ymm10 .byte 197,108,88,202 // vaddps %ymm2,%ymm2,%ymm9 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,29,135,77,0,0 // vbroadcastss 0x4d87(%rip),%ymm11 # 6880 <_sk_callback_avx+0x22c> + .byte 196,98,125,24,29,106,76,0,0 // vbroadcastss 0x4c6a(%rip),%ymm11 # 66a8 <_sk_callback_avx+0x22c> .byte 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 .byte 196,67,125,8,227,1 // vroundps $0x1,%ymm11,%ymm12 .byte 196,65,36,92,252 // vsubps %ymm12,%ymm11,%ymm15 .byte 196,65,44,92,217 // vsubps %ymm9,%ymm10,%ymm11 - .byte 196,98,125,24,37,113,77,0,0 // vbroadcastss 0x4d71(%rip),%ymm12 # 6888 <_sk_callback_avx+0x234> + .byte 196,98,125,24,37,84,76,0,0 // vbroadcastss 0x4c54(%rip),%ymm12 # 66b0 <_sk_callback_avx+0x234> .byte 196,193,4,89,196 // vmulps %ymm12,%ymm15,%ymm0 - .byte 196,98,125,24,45,103,77,0,0 // vbroadcastss 0x4d67(%rip),%ymm13 # 688c <_sk_callback_avx+0x238> + .byte 196,98,125,24,45,74,76,0,0 // vbroadcastss 0x4c4a(%rip),%ymm13 # 66b4 <_sk_callback_avx+0x238> .byte 197,20,92,240 // vsubps %ymm0,%ymm13,%ymm14 .byte 196,65,36,89,246 // vmulps %ymm14,%ymm11,%ymm14 .byte 196,65,52,88,246 // vaddps %ymm14,%ymm9,%ymm14 - .byte 196,226,125,24,13,72,77,0,0 // vbroadcastss 0x4d48(%rip),%ymm1 # 6884 <_sk_callback_avx+0x230> + .byte 196,226,125,24,13,43,76,0,0 // vbroadcastss 0x4c2b(%rip),%ymm1 # 66ac <_sk_callback_avx+0x230> .byte 196,193,116,194,255,2 // vcmpleps %ymm15,%ymm1,%ymm7 .byte 196,195,13,74,249,112 // vblendvps %ymm7,%ymm9,%ymm14,%ymm7 .byte 196,65,60,194,247,2 // vcmpleps %ymm15,%ymm8,%ymm14 .byte 196,227,45,74,255,224 // vblendvps %ymm14,%ymm7,%ymm10,%ymm7 - .byte 196,98,125,24,53,51,77,0,0 // vbroadcastss 0x4d33(%rip),%ymm14 # 6890 <_sk_callback_avx+0x23c> + .byte 196,98,125,24,53,22,76,0,0 // vbroadcastss 0x4c16(%rip),%ymm14 # 66b8 <_sk_callback_avx+0x23c> .byte 196,65,12,194,255,2 // vcmpleps %ymm15,%ymm14,%ymm15 .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 @@ -16797,7 +16742,7 @@ _sk_hsl_to_rgb_avx: .byte 197,164,89,247 // vmulps %ymm7,%ymm11,%ymm6 .byte 197,180,88,246 // vaddps %ymm6,%ymm9,%ymm6 .byte 196,227,77,74,237,0 // vblendvps %ymm0,%ymm5,%ymm6,%ymm5 - .byte 196,226,125,24,5,213,76,0,0 // vbroadcastss 0x4cd5(%rip),%ymm0 # 6894 <_sk_callback_avx+0x240> + .byte 196,226,125,24,5,184,75,0,0 // vbroadcastss 0x4bb8(%rip),%ymm0 # 66bc <_sk_callback_avx+0x240> .byte 197,228,88,192 // vaddps %ymm0,%ymm3,%ymm0 .byte 196,227,125,8,216,1 // vroundps $0x1,%ymm0,%ymm3 .byte 197,252,92,195 // vsubps %ymm3,%ymm0,%ymm0 @@ -16849,14 +16794,14 @@ _sk_scale_u8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,68 // jne 1cb6 <_sk_scale_u8_avx+0x54> + .byte 117,68 // jne 1bfb <_sk_scale_u8_avx+0x54> .byte 197,122,126,0 // vmovq (%rax),%xmm8 .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,254,75,0,0 // vbroadcastss 0x4bfe(%rip),%ymm9 # 6898 <_sk_callback_avx+0x244> + .byte 196,98,125,24,13,225,74,0,0 // vbroadcastss 0x4ae1(%rip),%ymm9 # 66c0 <_sk_callback_avx+0x244> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 @@ -16874,9 +16819,9 @@ _sk_scale_u8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 1cbe <_sk_scale_u8_avx+0x5c> + .byte 117,234 // jne 1c03 <_sk_scale_u8_avx+0x5c> .byte 196,65,249,110,193 // vmovq %r9,%xmm8 - .byte 235,155 // jmp 1c76 <_sk_scale_u8_avx+0x14> + .byte 235,155 // jmp 1bbb <_sk_scale_u8_avx+0x14> HIDDEN _sk_lerp_1_float_avx .globl _sk_lerp_1_float_avx @@ -16908,14 +16853,14 @@ _sk_lerp_u8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,104 // jne 1d92 <_sk_lerp_u8_avx+0x78> + .byte 117,104 // jne 1cd7 <_sk_lerp_u8_avx+0x78> .byte 197,122,126,0 // vmovq (%rax),%xmm8 .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,13,74,75,0,0 // vbroadcastss 0x4b4a(%rip),%ymm9 # 689c <_sk_callback_avx+0x248> + .byte 196,98,125,24,13,45,74,0,0 // vbroadcastss 0x4a2d(%rip),%ymm9 # 66c4 <_sk_callback_avx+0x248> .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 @@ -16941,9 +16886,9 @@ _sk_lerp_u8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 1d9a <_sk_lerp_u8_avx+0x80> + .byte 117,234 // jne 1cdf <_sk_lerp_u8_avx+0x80> .byte 196,65,249,110,193 // vmovq %r9,%xmm8 - .byte 233,116,255,255,255 // jmpq 1d2e <_sk_lerp_u8_avx+0x14> + .byte 233,116,255,255,255 // jmpq 1c73 <_sk_lerp_u8_avx+0x14> HIDDEN _sk_lerp_565_avx .globl _sk_lerp_565_avx @@ -16952,26 +16897,26 @@ _sk_lerp_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,208,0,0,0 // jne 1e98 <_sk_lerp_565_avx+0xde> + .byte 15,133,208,0,0,0 // jne 1ddd <_sk_lerp_565_avx+0xde> .byte 196,65,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 196,65,57,105,201 // vpunpckhwd %xmm9,%xmm8,%xmm9 .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 .byte 196,67,61,24,193,1 // vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,180,74,0,0 // vbroadcastss 0x4ab4(%rip),%ymm9 # 68a0 <_sk_callback_avx+0x24c> + .byte 196,98,125,24,13,151,73,0,0 // vbroadcastss 0x4997(%rip),%ymm9 # 66c8 <_sk_callback_avx+0x24c> .byte 196,65,60,84,201 // vandps %ymm9,%ymm8,%ymm9 .byte 196,65,124,91,201 // vcvtdq2ps %ymm9,%ymm9 - .byte 196,98,125,24,21,165,74,0,0 // vbroadcastss 0x4aa5(%rip),%ymm10 # 68a4 <_sk_callback_avx+0x250> + .byte 196,98,125,24,21,136,73,0,0 // vbroadcastss 0x4988(%rip),%ymm10 # 66cc <_sk_callback_avx+0x250> .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,21,155,74,0,0 // vbroadcastss 0x4a9b(%rip),%ymm10 # 68a8 <_sk_callback_avx+0x254> + .byte 196,98,125,24,21,126,73,0,0 // vbroadcastss 0x497e(%rip),%ymm10 # 66d0 <_sk_callback_avx+0x254> .byte 196,65,60,84,210 // vandps %ymm10,%ymm8,%ymm10 .byte 196,65,124,91,210 // vcvtdq2ps %ymm10,%ymm10 - .byte 196,98,125,24,29,140,74,0,0 // vbroadcastss 0x4a8c(%rip),%ymm11 # 68ac <_sk_callback_avx+0x258> + .byte 196,98,125,24,29,111,73,0,0 // vbroadcastss 0x496f(%rip),%ymm11 # 66d4 <_sk_callback_avx+0x258> .byte 196,65,44,89,211 // vmulps %ymm11,%ymm10,%ymm10 - .byte 196,98,125,24,29,130,74,0,0 // vbroadcastss 0x4a82(%rip),%ymm11 # 68b0 <_sk_callback_avx+0x25c> + .byte 196,98,125,24,29,101,73,0,0 // vbroadcastss 0x4965(%rip),%ymm11 # 66d8 <_sk_callback_avx+0x25c> .byte 196,65,60,84,195 // vandps %ymm11,%ymm8,%ymm8 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 - .byte 196,98,125,24,29,115,74,0,0 // vbroadcastss 0x4a73(%rip),%ymm11 # 68b4 <_sk_callback_avx+0x260> + .byte 196,98,125,24,29,86,73,0,0 // vbroadcastss 0x4956(%rip),%ymm11 # 66dc <_sk_callback_avx+0x260> .byte 196,65,60,89,195 // vmulps %ymm11,%ymm8,%ymm8 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 @@ -16998,9 +16943,9 @@ _sk_lerp_565_avx: .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,29,255,255,255 // ja 1dce <_sk_lerp_565_avx+0x14> + .byte 15,135,29,255,255,255 // ja 1d13 <_sk_lerp_565_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 1f08 <_sk_lerp_565_avx+0x14e> + .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 1e4c <_sk_lerp_565_avx+0x14d> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -17012,28 +16957,27 @@ _sk_lerp_565_avx: .byte 196,65,57,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 .byte 196,65,57,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 .byte 196,65,57,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 - .byte 233,200,254,255,255 // jmpq 1dce <_sk_lerp_565_avx+0x14> - .byte 102,144 // xchg %ax,%ax - .byte 242,255 // repnz (bad) - .byte 255 // (bad) + .byte 233,200,254,255,255 // jmpq 1d13 <_sk_lerp_565_avx+0x14> + .byte 144 // nop + .byte 243,255 // repz (bad) .byte 255 // (bad) - .byte 234 // (bad) .byte 255 // (bad) + .byte 235,255 // jmp 1e51 <_sk_lerp_565_avx+0x152> .byte 255 // (bad) - .byte 255,226 // jmpq *%rdx + .byte 255,227 // jmpq *%rbx .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 218,255 // (bad) + .byte 219,255 // (bad) .byte 255 // (bad) - .byte 255,210 // callq *%rdx + .byte 255,211 // callq *%rbx .byte 255 // (bad) .byte 255 // (bad) - .byte 255,202 // dec %edx + .byte 255,203 // dec %ebx .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 189 // .byte 0xbd + .byte 190 // .byte 0xbe .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -17042,19 +16986,21 @@ HIDDEN _sk_load_tables_avx .globl _sk_load_tables_avx FUNCTION(_sk_load_tables_avx) _sk_load_tables_avx: + .byte 73,137,200 // mov %rcx,%r8 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 76,139,0 // mov (%rax),%r8 - .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,26,2,0,0 // jne 214c <_sk_load_tables_avx+0x228> - .byte 196,65,124,16,4,184 // vmovups (%r8,%rdi,4),%ymm8 + .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 + .byte 76,3,8 // add (%rax),%r9 + .byte 77,133,192 // test %r8,%r8 + .byte 15,133,26,2,0,0 // jne 209b <_sk_load_tables_avx+0x233> + .byte 196,65,124,16,17 // vmovups (%r9),%ymm10 .byte 85 // push %rbp .byte 65,87 // push %r15 .byte 65,86 // push %r14 .byte 65,85 // push %r13 .byte 65,84 // push %r12 .byte 83 // push %rbx - .byte 197,124,40,13,150,76,0,0 // vmovaps 0x4c96(%rip),%ymm9 # 6be0 <_sk_callback_avx+0x58c> - .byte 196,193,60,84,193 // vandps %ymm9,%ymm8,%ymm0 + .byte 197,124,40,13,8,76,0,0 // vmovaps 0x4c08(%rip),%ymm9 # 6aa0 <_sk_callback_avx+0x624> + .byte 196,193,44,84,193 // vandps %ymm9,%ymm10,%ymm0 .byte 196,193,249,126,193 // vmovq %xmm0,%r9 .byte 69,137,203 // mov %r9d,%r11d .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10 @@ -17062,26 +17008,26 @@ _sk_load_tables_avx: .byte 73,193,234,32 // shr $0x20,%r10 .byte 73,193,233,32 // shr $0x20,%r9 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 - .byte 196,193,249,126,196 // vmovq %xmm0,%r12 - .byte 69,137,231 // mov %r12d,%r15d - .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx - .byte 65,137,221 // mov %ebx,%r13d + .byte 196,225,249,126,195 // vmovq %xmm0,%rbx + .byte 65,137,223 // mov %ebx,%r15d + .byte 196,227,249,22,193,1 // vpextrq $0x1,%xmm0,%rcx + .byte 65,137,205 // mov %ecx,%r13d + .byte 72,193,233,32 // shr $0x20,%rcx .byte 72,193,235,32 // shr $0x20,%rbx - .byte 73,193,236,32 // shr $0x20,%r12 .byte 72,139,104,8 // mov 0x8(%rax),%rbp - .byte 76,139,64,16 // mov 0x10(%rax),%r8 + .byte 76,139,96,16 // mov 0x10(%rax),%r12 .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0 - .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0 + .byte 196,227,121,33,68,157,0,16 // vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm0,%xmm0 .byte 196,163,121,33,68,173,0,32 // vinsertps $0x20,0x0(%rbp,%r13,4),%xmm0,%xmm0 - .byte 196,227,121,33,68,157,0,48 // vinsertps $0x30,0x0(%rbp,%rbx,4),%xmm0,%xmm0 + .byte 196,227,121,33,68,141,0,48 // vinsertps $0x30,0x0(%rbp,%rcx,4),%xmm0,%xmm0 .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1 .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 .byte 196,163,113,33,76,181,0,32 // vinsertps $0x20,0x0(%rbp,%r14,4),%xmm1,%xmm1 .byte 196,163,113,33,76,149,0,48 // vinsertps $0x30,0x0(%rbp,%r10,4),%xmm1,%xmm1 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 - .byte 196,67,125,25,194,1 // vextractf128 $0x1,%ymm8,%xmm10 - .byte 196,193,105,114,210,8 // vpsrld $0x8,%xmm10,%xmm2 + .byte 196,193,113,114,210,8 // vpsrld $0x8,%xmm10,%xmm1 + .byte 196,67,125,25,208,1 // vextractf128 $0x1,%ymm10,%xmm8 + .byte 196,193,105,114,208,8 // vpsrld $0x8,%xmm8,%xmm2 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1 .byte 196,193,249,126,201 // vmovq %xmm1,%r9 @@ -17091,36 +17037,36 @@ _sk_load_tables_avx: .byte 73,193,234,32 // shr $0x20,%r10 .byte 73,193,233,32 // shr $0x20,%r9 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 - .byte 196,225,249,126,205 // vmovq %xmm1,%rbp - .byte 65,137,239 // mov %ebp,%r15d - .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx - .byte 65,137,220 // mov %ebx,%r12d - .byte 72,193,235,32 // shr $0x20,%rbx + .byte 196,225,249,126,203 // vmovq %xmm1,%rbx + .byte 65,137,223 // mov %ebx,%r15d + .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp + .byte 137,233 // mov %ebp,%ecx .byte 72,193,237,32 // shr $0x20,%rbp - .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1 - .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1 - .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2 + .byte 72,193,235,32 // shr $0x20,%rbx + .byte 196,129,122,16,12,188 // vmovss (%r12,%r15,4),%xmm1 + .byte 196,195,113,33,12,156,16 // vinsertps $0x10,(%r12,%rbx,4),%xmm1,%xmm1 + .byte 196,193,122,16,20,140 // vmovss (%r12,%rcx,4),%xmm2 .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1 - .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2 + .byte 196,193,122,16,20,172 // vmovss (%r12,%rbp,4),%xmm2 .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1 - .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2 - .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2 - .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3 + .byte 196,129,122,16,20,156 // vmovss (%r12,%r11,4),%xmm2 + .byte 196,131,105,33,20,140,16 // vinsertps $0x10,(%r12,%r9,4),%xmm2,%xmm2 + .byte 196,129,122,16,28,180 // vmovss (%r12,%r14,4),%xmm3 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 - .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3 + .byte 196,129,122,16,28,148 // vmovss (%r12,%r10,4),%xmm3 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 .byte 72,139,64,24 // mov 0x18(%rax),%rax - .byte 196,193,105,114,208,16 // vpsrld $0x10,%xmm8,%xmm2 - .byte 196,193,97,114,210,16 // vpsrld $0x10,%xmm10,%xmm3 + .byte 196,193,105,114,210,16 // vpsrld $0x10,%xmm10,%xmm2 + .byte 196,193,97,114,208,16 // vpsrld $0x10,%xmm8,%xmm3 .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2 - .byte 196,193,249,126,208 // vmovq %xmm2,%r8 - .byte 69,137,194 // mov %r8d,%r10d - .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9 - .byte 69,137,203 // mov %r9d,%r11d + .byte 196,193,249,126,209 // vmovq %xmm2,%r9 + .byte 69,137,202 // mov %r9d,%r10d + .byte 196,227,249,22,209,1 // vpextrq $0x1,%xmm2,%rcx + .byte 65,137,203 // mov %ecx,%r11d + .byte 72,193,233,32 // shr $0x20,%rcx .byte 73,193,233,32 // shr $0x20,%r9 - .byte 73,193,232,32 // shr $0x20,%r8 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 .byte 196,225,249,126,213 // vmovq %xmm2,%rbp .byte 65,137,238 // mov %ebp,%r14d @@ -17135,19 +17081,20 @@ _sk_load_tables_avx: .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3 .byte 196,99,105,33,203,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm9 .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3 - .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + .byte 196,163,97,33,28,136,16 // vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2 .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2 - .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3 + .byte 197,250,16,28,136 // vmovss (%rax,%rcx,4),%xmm3 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 .byte 196,195,109,24,209,1 // vinsertf128 $0x1,%xmm9,%ymm2,%ymm2 - .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8 - .byte 196,193,97,114,210,24 // vpsrld $0x18,%xmm10,%xmm3 - .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 + .byte 196,193,49,114,210,24 // vpsrld $0x18,%xmm10,%xmm9 + .byte 196,193,97,114,208,24 // vpsrld $0x18,%xmm8,%xmm3 + .byte 196,227,53,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,127,71,0,0 // vbroadcastss 0x477f(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x264> + .byte 196,98,125,24,5,91,70,0,0 // vbroadcastss 0x465b(%rip),%ymm8 # 66e0 <_sk_callback_avx+0x264> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,137,193 // mov %r8,%rcx .byte 91 // pop %rbx .byte 65,92 // pop %r12 .byte 65,93 // pop %r13 @@ -17155,57 +17102,20 @@ _sk_load_tables_avx: .byte 65,95 // pop %r15 .byte 93 // pop %rbp .byte 255,224 // jmpq *%rax - .byte 65,137,201 // mov %ecx,%r9d - .byte 65,128,225,7 // and $0x7,%r9b - .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 - .byte 65,254,201 // dec %r9b - .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 15,135,211,253,255,255 // ja 1f38 <_sk_load_tables_avx+0x14> - .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,140,0,0,0 // lea 0x8c(%rip),%r10 # 21fc <_sk_load_tables_avx+0x2d8> - .byte 79,99,12,138 // movslq (%r10,%r9,4),%r9 - .byte 77,1,209 // add %r10,%r9 - .byte 65,255,225 // jmpq *%r9 - .byte 196,193,121,110,68,184,24 // vmovd 0x18(%r8,%rdi,4),%xmm0 - .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 - .byte 196,99,117,12,192,64 // vblendps $0x40,%ymm0,%ymm1,%ymm8 - .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 - .byte 196,195,121,34,68,184,20,1 // vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0 - .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 - .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 - .byte 196,195,121,34,68,184,16,0 // vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0 - .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 - .byte 196,195,57,34,68,184,12,3 // vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm8,%xmm0 - .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 - .byte 196,195,57,34,68,184,8,2 // vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm8,%xmm0 - .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 - .byte 196,195,57,34,68,184,4,1 // vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm8,%xmm0 - .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 - .byte 196,195,57,34,4,184,0 // vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0 - .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 - .byte 233,62,253,255,255 // jmpq 1f38 <_sk_load_tables_avx+0x14> - .byte 102,144 // xchg %ax,%ax - .byte 236 // in (%dx),%al - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 222,255 // fdivrp %st,%st(7) - .byte 255 // (bad) - .byte 255,208 // callq *%rax - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,194 // inc %edx - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,174,255,255,255,154 // ljmp *-0x65000001(%rsi) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 126,255 // jle 2215 <_sk_load_tables_avx+0x2f1> - .byte 255 // (bad) - .byte 255 // .byte 0xff + .byte 185,8,0,0,0 // mov $0x8,%ecx + .byte 68,41,193 // sub %r8d,%ecx + .byte 192,225,3 // shl $0x3,%cl + .byte 73,199,194,255,255,255,255 // mov $0xffffffffffffffff,%r10 + .byte 73,211,234 // shr %cl,%r10 + .byte 196,193,249,110,194 // vmovq %r10,%xmm0 + .byte 196,226,121,48,192 // vpmovzxbw %xmm0,%xmm0 + .byte 196,226,121,0,13,61,73,0,0 // vpshufb 0x493d(%rip),%xmm0,%xmm1 # 6a00 <_sk_callback_avx+0x584> + .byte 196,226,121,33,201 // vpmovsxbd %xmm1,%xmm1 + .byte 196,226,121,0,5,63,73,0,0 // vpshufb 0x493f(%rip),%xmm0,%xmm0 # 6a10 <_sk_callback_avx+0x594> + .byte 196,226,121,33,192 // vpmovsxbd %xmm0,%xmm0 + .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + .byte 196,66,125,44,17 // vmaskmovps (%r9),%ymm0,%ymm10 + .byte 233,160,253,255,255 // jmpq 1e86 <_sk_load_tables_avx+0x1e> HIDDEN _sk_load_tables_u16_be_avx .globl _sk_load_tables_u16_be_avx @@ -17215,7 +17125,7 @@ _sk_load_tables_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,113,2,0,0 // jne 249f <_sk_load_tables_u16_be_avx+0x287> + .byte 15,133,113,2,0,0 // jne 236d <_sk_load_tables_u16_be_avx+0x287> .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8 .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2 .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3 @@ -17237,7 +17147,7 @@ _sk_load_tables_u16_be_avx: .byte 197,177,108,208 // vpunpcklqdq %xmm0,%xmm9,%xmm2 .byte 197,177,109,200 // vpunpckhqdq %xmm0,%xmm9,%xmm1 .byte 196,65,57,108,212 // vpunpcklqdq %xmm12,%xmm8,%xmm10 - .byte 197,121,111,29,214,73,0,0 // vmovdqa 0x49d6(%rip),%xmm11 # 6c60 <_sk_callback_avx+0x60c> + .byte 197,121,111,29,200,72,0,0 // vmovdqa 0x48c8(%rip),%xmm11 # 6a20 <_sk_callback_avx+0x5a4> .byte 196,193,105,219,195 // vpand %xmm11,%xmm2,%xmm0 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 196,193,121,105,209 // vpunpckhwd %xmm9,%xmm0,%xmm2 @@ -17336,7 +17246,7 @@ _sk_load_tables_u16_be_avx: .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3 .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,48,68,0,0 // vbroadcastss 0x4430(%rip),%ymm8 # 68bc <_sk_callback_avx+0x268> + .byte 196,98,125,24,5,138,67,0,0 // vbroadcastss 0x438a(%rip),%ymm8 # 66e4 <_sk_callback_avx+0x268> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx @@ -17349,29 +17259,29 @@ _sk_load_tables_u16_be_avx: .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 2505 <_sk_load_tables_u16_be_avx+0x2ed> + .byte 116,85 // je 23d3 <_sk_load_tables_u16_be_avx+0x2ed> .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 2505 <_sk_load_tables_u16_be_avx+0x2ed> + .byte 114,72 // jb 23d3 <_sk_load_tables_u16_be_avx+0x2ed> .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 2512 <_sk_load_tables_u16_be_avx+0x2fa> + .byte 116,72 // je 23e0 <_sk_load_tables_u16_be_avx+0x2fa> .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 2512 <_sk_load_tables_u16_be_avx+0x2fa> + .byte 114,59 // jb 23e0 <_sk_load_tables_u16_be_avx+0x2fa> .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,97,253,255,255 // je 2249 <_sk_load_tables_u16_be_avx+0x31> + .byte 15,132,97,253,255,255 // je 2117 <_sk_load_tables_u16_be_avx+0x31> .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,80,253,255,255 // jb 2249 <_sk_load_tables_u16_be_avx+0x31> + .byte 15,130,80,253,255,255 // jb 2117 <_sk_load_tables_u16_be_avx+0x31> .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9 - .byte 233,68,253,255,255 // jmpq 2249 <_sk_load_tables_u16_be_avx+0x31> + .byte 233,68,253,255,255 // jmpq 2117 <_sk_load_tables_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,55,253,255,255 // jmpq 2249 <_sk_load_tables_u16_be_avx+0x31> + .byte 233,55,253,255,255 // jmpq 2117 <_sk_load_tables_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,46,253,255,255 // jmpq 2249 <_sk_load_tables_u16_be_avx+0x31> + .byte 233,46,253,255,255 // jmpq 2117 <_sk_load_tables_u16_be_avx+0x31> HIDDEN _sk_load_tables_rgb_u16_be_avx .globl _sk_load_tables_rgb_u16_be_avx @@ -17381,7 +17291,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,93,2,0,0 // jne 278a <_sk_load_tables_rgb_u16_be_avx+0x26f> + .byte 15,133,93,2,0,0 // jne 2658 <_sk_load_tables_rgb_u16_be_avx+0x26f> .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0 .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2 .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1 @@ -17408,7 +17318,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 197,185,108,202 // vpunpcklqdq %xmm2,%xmm8,%xmm1 .byte 197,185,109,210 // vpunpckhqdq %xmm2,%xmm8,%xmm2 .byte 197,121,108,195 // vpunpcklqdq %xmm3,%xmm0,%xmm8 - .byte 197,121,111,13,207,70,0,0 // vmovdqa 0x46cf(%rip),%xmm9 # 6c70 <_sk_callback_avx+0x61c> + .byte 197,121,111,13,193,69,0,0 // vmovdqa 0x45c1(%rip),%xmm9 # 6a30 <_sk_callback_avx+0x5b4> .byte 196,193,113,219,193 // vpand %xmm9,%xmm1,%xmm0 .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10 .byte 196,193,121,105,202 // vpunpckhwd %xmm10,%xmm0,%xmm1 @@ -17500,7 +17410,7 @@ _sk_load_tables_rgb_u16_be_avx: .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 .byte 196,195,109,24,208,1 // vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,66,65,0,0 // vbroadcastss 0x4142(%rip),%ymm3 # 68c0 <_sk_callback_avx+0x26c> + .byte 196,226,125,24,29,156,64,0,0 // vbroadcastss 0x409c(%rip),%ymm3 # 66e8 <_sk_callback_avx+0x26c> .byte 91 // pop %rbx .byte 65,92 // pop %r12 .byte 65,93 // pop %r13 @@ -17511,36 +17421,36 @@ _sk_load_tables_rgb_u16_be_avx: .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0 .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 27a3 <_sk_load_tables_rgb_u16_be_avx+0x288> - .byte 233,190,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 117,5 // jne 2671 <_sk_load_tables_rgb_u16_be_avx+0x288> + .byte 233,190,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1 .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 27d2 <_sk_load_tables_rgb_u16_be_avx+0x2b7> + .byte 114,26 // jb 26a0 <_sk_load_tables_rgb_u16_be_avx+0x2b7> .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1 .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 27d7 <_sk_load_tables_rgb_u16_be_avx+0x2bc> - .byte 233,143,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> - .byte 233,138,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 26a5 <_sk_load_tables_rgb_u16_be_avx+0x2bc> + .byte 233,143,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,138,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1 .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 2806 <_sk_load_tables_rgb_u16_be_avx+0x2eb> + .byte 114,26 // jb 26d4 <_sk_load_tables_rgb_u16_be_avx+0x2eb> .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1 .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 280b <_sk_load_tables_rgb_u16_be_avx+0x2f0> - .byte 233,91,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> - .byte 233,86,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 26d9 <_sk_load_tables_rgb_u16_be_avx+0x2f0> + .byte 233,91,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,86,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3 .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 2834 <_sk_load_tables_rgb_u16_be_avx+0x319> + .byte 114,20 // jb 2702 <_sk_load_tables_rgb_u16_be_avx+0x319> .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3 .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 - .byte 233,45,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> - .byte 233,40,253,255,255 // jmpq 2561 <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,45,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> + .byte 233,40,253,255,255 // jmpq 242f <_sk_load_tables_rgb_u16_be_avx+0x46> HIDDEN _sk_byte_tables_avx .globl _sk_byte_tables_avx @@ -17553,7 +17463,7 @@ _sk_byte_tables_avx: .byte 65,84 // push %r12 .byte 83 // push %rbx .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,118,64,0,0 // vbroadcastss 0x4076(%rip),%ymm8 # 68c4 <_sk_callback_avx+0x270> + .byte 196,98,125,24,5,208,63,0,0 // vbroadcastss 0x3fd0(%rip),%ymm8 # 66ec <_sk_callback_avx+0x270> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 @@ -17590,7 +17500,7 @@ _sk_byte_tables_avx: .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,196,63,0,0 // vbroadcastss 0x3fc4(%rip),%ymm9 # 68c8 <_sk_callback_avx+0x274> + .byte 196,98,125,24,13,30,63,0,0 // vbroadcastss 0x3f1e(%rip),%ymm9 # 66f0 <_sk_callback_avx+0x274> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 @@ -17752,7 +17662,7 @@ _sk_byte_tables_rgb_avx: .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,234,60,0,0 // vbroadcastss 0x3cea(%rip),%ymm9 # 68cc <_sk_callback_avx+0x278> + .byte 196,98,125,24,13,68,60,0,0 // vbroadcastss 0x3c44(%rip),%ymm9 # 66f4 <_sk_callback_avx+0x278> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 @@ -18049,36 +17959,36 @@ _sk_parametric_r_avx: .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,216 // vcvtdq2ps %ymm0,%ymm11 - .byte 196,98,125,24,37,72,56,0,0 // vbroadcastss 0x3848(%rip),%ymm12 # 68d0 <_sk_callback_avx+0x27c> + .byte 196,98,125,24,37,162,55,0,0 // vbroadcastss 0x37a2(%rip),%ymm12 # 66f8 <_sk_callback_avx+0x27c> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,62,56,0,0 // vbroadcastss 0x383e(%rip),%ymm12 # 68d4 <_sk_callback_avx+0x280> + .byte 196,98,125,24,37,152,55,0,0 // vbroadcastss 0x3798(%rip),%ymm12 # 66fc <_sk_callback_avx+0x280> .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,52,56,0,0 // vbroadcastss 0x3834(%rip),%ymm12 # 68d8 <_sk_callback_avx+0x284> + .byte 196,98,125,24,37,142,55,0,0 // vbroadcastss 0x378e(%rip),%ymm12 # 6700 <_sk_callback_avx+0x284> .byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,42,56,0,0 // vbroadcastss 0x382a(%rip),%ymm12 # 68dc <_sk_callback_avx+0x288> + .byte 196,98,125,24,37,132,55,0,0 // vbroadcastss 0x3784(%rip),%ymm12 # 6704 <_sk_callback_avx+0x288> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,32,56,0,0 // vbroadcastss 0x3820(%rip),%ymm12 # 68e0 <_sk_callback_avx+0x28c> + .byte 196,98,125,24,37,122,55,0,0 // vbroadcastss 0x377a(%rip),%ymm12 # 6708 <_sk_callback_avx+0x28c> .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,17,56,0,0 // vbroadcastss 0x3811(%rip),%ymm12 # 68e4 <_sk_callback_avx+0x290> + .byte 196,98,125,24,37,107,55,0,0 // vbroadcastss 0x376b(%rip),%ymm12 # 670c <_sk_callback_avx+0x290> .byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0 - .byte 196,98,125,24,37,7,56,0,0 // vbroadcastss 0x3807(%rip),%ymm12 # 68e8 <_sk_callback_avx+0x294> + .byte 196,98,125,24,37,97,55,0,0 // vbroadcastss 0x3761(%rip),%ymm12 # 6710 <_sk_callback_avx+0x294> .byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0 .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10 - .byte 196,98,125,24,29,235,55,0,0 // vbroadcastss 0x37eb(%rip),%ymm11 # 68ec <_sk_callback_avx+0x298> + .byte 196,98,125,24,29,69,55,0,0 // vbroadcastss 0x3745(%rip),%ymm11 # 6714 <_sk_callback_avx+0x298> .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,225,55,0,0 // vbroadcastss 0x37e1(%rip),%ymm11 # 68f0 <_sk_callback_avx+0x29c> + .byte 196,98,125,24,29,59,55,0,0 // vbroadcastss 0x373b(%rip),%ymm11 # 6718 <_sk_callback_avx+0x29c> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0 - .byte 196,98,125,24,29,210,55,0,0 // vbroadcastss 0x37d2(%rip),%ymm11 # 68f4 <_sk_callback_avx+0x2a0> + .byte 196,98,125,24,29,44,55,0,0 // vbroadcastss 0x372c(%rip),%ymm11 # 671c <_sk_callback_avx+0x2a0> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,200,55,0,0 // vbroadcastss 0x37c8(%rip),%ymm11 # 68f8 <_sk_callback_avx+0x2a4> + .byte 196,98,125,24,29,34,55,0,0 // vbroadcastss 0x3722(%rip),%ymm11 # 6720 <_sk_callback_avx+0x2a4> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 - .byte 196,98,125,24,21,185,55,0,0 // vbroadcastss 0x37b9(%rip),%ymm10 # 68fc <_sk_callback_avx+0x2a8> + .byte 196,98,125,24,21,19,55,0,0 // vbroadcastss 0x3713(%rip),%ymm10 # 6724 <_sk_callback_avx+0x2a8> .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -18086,7 +17996,7 @@ _sk_parametric_r_avx: .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,144,55,0,0 // vbroadcastss 0x3790(%rip),%ymm8 # 6900 <_sk_callback_avx+0x2ac> + .byte 196,98,125,24,5,234,54,0,0 // vbroadcastss 0x36ea(%rip),%ymm8 # 6728 <_sk_callback_avx+0x2ac> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18108,36 +18018,36 @@ _sk_parametric_g_avx: .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,217 // vcvtdq2ps %ymm1,%ymm11 - .byte 196,98,125,24,37,65,55,0,0 // vbroadcastss 0x3741(%rip),%ymm12 # 6904 <_sk_callback_avx+0x2b0> + .byte 196,98,125,24,37,155,54,0,0 // vbroadcastss 0x369b(%rip),%ymm12 # 672c <_sk_callback_avx+0x2b0> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,55,55,0,0 // vbroadcastss 0x3737(%rip),%ymm12 # 6908 <_sk_callback_avx+0x2b4> + .byte 196,98,125,24,37,145,54,0,0 // vbroadcastss 0x3691(%rip),%ymm12 # 6730 <_sk_callback_avx+0x2b4> .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,45,55,0,0 // vbroadcastss 0x372d(%rip),%ymm12 # 690c <_sk_callback_avx+0x2b8> + .byte 196,98,125,24,37,135,54,0,0 // vbroadcastss 0x3687(%rip),%ymm12 # 6734 <_sk_callback_avx+0x2b8> .byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,35,55,0,0 // vbroadcastss 0x3723(%rip),%ymm12 # 6910 <_sk_callback_avx+0x2bc> + .byte 196,98,125,24,37,125,54,0,0 // vbroadcastss 0x367d(%rip),%ymm12 # 6738 <_sk_callback_avx+0x2bc> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,25,55,0,0 // vbroadcastss 0x3719(%rip),%ymm12 # 6914 <_sk_callback_avx+0x2c0> + .byte 196,98,125,24,37,115,54,0,0 // vbroadcastss 0x3673(%rip),%ymm12 # 673c <_sk_callback_avx+0x2c0> .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,10,55,0,0 // vbroadcastss 0x370a(%rip),%ymm12 # 6918 <_sk_callback_avx+0x2c4> + .byte 196,98,125,24,37,100,54,0,0 // vbroadcastss 0x3664(%rip),%ymm12 # 6740 <_sk_callback_avx+0x2c4> .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 - .byte 196,98,125,24,37,0,55,0,0 // vbroadcastss 0x3700(%rip),%ymm12 # 691c <_sk_callback_avx+0x2c8> + .byte 196,98,125,24,37,90,54,0,0 // vbroadcastss 0x365a(%rip),%ymm12 # 6744 <_sk_callback_avx+0x2c8> .byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1 .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10 - .byte 196,98,125,24,29,228,54,0,0 // vbroadcastss 0x36e4(%rip),%ymm11 # 6920 <_sk_callback_avx+0x2cc> + .byte 196,98,125,24,29,62,54,0,0 // vbroadcastss 0x363e(%rip),%ymm11 # 6748 <_sk_callback_avx+0x2cc> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,218,54,0,0 // vbroadcastss 0x36da(%rip),%ymm11 # 6924 <_sk_callback_avx+0x2d0> + .byte 196,98,125,24,29,52,54,0,0 // vbroadcastss 0x3634(%rip),%ymm11 # 674c <_sk_callback_avx+0x2d0> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,29,203,54,0,0 // vbroadcastss 0x36cb(%rip),%ymm11 # 6928 <_sk_callback_avx+0x2d4> + .byte 196,98,125,24,29,37,54,0,0 // vbroadcastss 0x3625(%rip),%ymm11 # 6750 <_sk_callback_avx+0x2d4> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,193,54,0,0 // vbroadcastss 0x36c1(%rip),%ymm11 # 692c <_sk_callback_avx+0x2d8> + .byte 196,98,125,24,29,27,54,0,0 // vbroadcastss 0x361b(%rip),%ymm11 # 6754 <_sk_callback_avx+0x2d8> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 - .byte 196,98,125,24,21,178,54,0,0 // vbroadcastss 0x36b2(%rip),%ymm10 # 6930 <_sk_callback_avx+0x2dc> + .byte 196,98,125,24,21,12,54,0,0 // vbroadcastss 0x360c(%rip),%ymm10 # 6758 <_sk_callback_avx+0x2dc> .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -18145,7 +18055,7 @@ _sk_parametric_g_avx: .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,5,137,54,0,0 // vbroadcastss 0x3689(%rip),%ymm8 # 6934 <_sk_callback_avx+0x2e0> + .byte 196,98,125,24,5,227,53,0,0 // vbroadcastss 0x35e3(%rip),%ymm8 # 675c <_sk_callback_avx+0x2e0> .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18167,36 +18077,36 @@ _sk_parametric_b_avx: .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,218 // vcvtdq2ps %ymm2,%ymm11 - .byte 196,98,125,24,37,58,54,0,0 // vbroadcastss 0x363a(%rip),%ymm12 # 6938 <_sk_callback_avx+0x2e4> + .byte 196,98,125,24,37,148,53,0,0 // vbroadcastss 0x3594(%rip),%ymm12 # 6760 <_sk_callback_avx+0x2e4> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,48,54,0,0 // vbroadcastss 0x3630(%rip),%ymm12 # 693c <_sk_callback_avx+0x2e8> + .byte 196,98,125,24,37,138,53,0,0 // vbroadcastss 0x358a(%rip),%ymm12 # 6764 <_sk_callback_avx+0x2e8> .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,38,54,0,0 // vbroadcastss 0x3626(%rip),%ymm12 # 6940 <_sk_callback_avx+0x2ec> + .byte 196,98,125,24,37,128,53,0,0 // vbroadcastss 0x3580(%rip),%ymm12 # 6768 <_sk_callback_avx+0x2ec> .byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,28,54,0,0 // vbroadcastss 0x361c(%rip),%ymm12 # 6944 <_sk_callback_avx+0x2f0> + .byte 196,98,125,24,37,118,53,0,0 // vbroadcastss 0x3576(%rip),%ymm12 # 676c <_sk_callback_avx+0x2f0> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,18,54,0,0 // vbroadcastss 0x3612(%rip),%ymm12 # 6948 <_sk_callback_avx+0x2f4> + .byte 196,98,125,24,37,108,53,0,0 // vbroadcastss 0x356c(%rip),%ymm12 # 6770 <_sk_callback_avx+0x2f4> .byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,3,54,0,0 // vbroadcastss 0x3603(%rip),%ymm12 # 694c <_sk_callback_avx+0x2f8> + .byte 196,98,125,24,37,93,53,0,0 // vbroadcastss 0x355d(%rip),%ymm12 # 6774 <_sk_callback_avx+0x2f8> .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2 - .byte 196,98,125,24,37,249,53,0,0 // vbroadcastss 0x35f9(%rip),%ymm12 # 6950 <_sk_callback_avx+0x2fc> + .byte 196,98,125,24,37,83,53,0,0 // vbroadcastss 0x3553(%rip),%ymm12 # 6778 <_sk_callback_avx+0x2fc> .byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2 .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10 - .byte 196,98,125,24,29,221,53,0,0 // vbroadcastss 0x35dd(%rip),%ymm11 # 6954 <_sk_callback_avx+0x300> + .byte 196,98,125,24,29,55,53,0,0 // vbroadcastss 0x3537(%rip),%ymm11 # 677c <_sk_callback_avx+0x300> .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,211,53,0,0 // vbroadcastss 0x35d3(%rip),%ymm11 # 6958 <_sk_callback_avx+0x304> + .byte 196,98,125,24,29,45,53,0,0 // vbroadcastss 0x352d(%rip),%ymm11 # 6780 <_sk_callback_avx+0x304> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2 - .byte 196,98,125,24,29,196,53,0,0 // vbroadcastss 0x35c4(%rip),%ymm11 # 695c <_sk_callback_avx+0x308> + .byte 196,98,125,24,29,30,53,0,0 // vbroadcastss 0x351e(%rip),%ymm11 # 6784 <_sk_callback_avx+0x308> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,186,53,0,0 // vbroadcastss 0x35ba(%rip),%ymm11 # 6960 <_sk_callback_avx+0x30c> + .byte 196,98,125,24,29,20,53,0,0 // vbroadcastss 0x3514(%rip),%ymm11 # 6788 <_sk_callback_avx+0x30c> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 - .byte 196,98,125,24,21,171,53,0,0 // vbroadcastss 0x35ab(%rip),%ymm10 # 6964 <_sk_callback_avx+0x310> + .byte 196,98,125,24,21,5,53,0,0 // vbroadcastss 0x3505(%rip),%ymm10 # 678c <_sk_callback_avx+0x310> .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -18204,7 +18114,7 @@ _sk_parametric_b_avx: .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,130,53,0,0 // vbroadcastss 0x3582(%rip),%ymm8 # 6968 <_sk_callback_avx+0x314> + .byte 196,98,125,24,5,220,52,0,0 // vbroadcastss 0x34dc(%rip),%ymm8 # 6790 <_sk_callback_avx+0x314> .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18226,36 +18136,36 @@ _sk_parametric_a_avx: .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10 .byte 197,124,91,219 // vcvtdq2ps %ymm3,%ymm11 - .byte 196,98,125,24,37,51,53,0,0 // vbroadcastss 0x3533(%rip),%ymm12 # 696c <_sk_callback_avx+0x318> + .byte 196,98,125,24,37,141,52,0,0 // vbroadcastss 0x348d(%rip),%ymm12 # 6794 <_sk_callback_avx+0x318> .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,41,53,0,0 // vbroadcastss 0x3529(%rip),%ymm12 # 6970 <_sk_callback_avx+0x31c> + .byte 196,98,125,24,37,131,52,0,0 // vbroadcastss 0x3483(%rip),%ymm12 # 6798 <_sk_callback_avx+0x31c> .byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,31,53,0,0 // vbroadcastss 0x351f(%rip),%ymm12 # 6974 <_sk_callback_avx+0x320> + .byte 196,98,125,24,37,121,52,0,0 // vbroadcastss 0x3479(%rip),%ymm12 # 679c <_sk_callback_avx+0x320> .byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,21,53,0,0 // vbroadcastss 0x3515(%rip),%ymm12 # 6978 <_sk_callback_avx+0x324> + .byte 196,98,125,24,37,111,52,0,0 // vbroadcastss 0x346f(%rip),%ymm12 # 67a0 <_sk_callback_avx+0x324> .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,11,53,0,0 // vbroadcastss 0x350b(%rip),%ymm12 # 697c <_sk_callback_avx+0x328> + .byte 196,98,125,24,37,101,52,0,0 // vbroadcastss 0x3465(%rip),%ymm12 # 67a4 <_sk_callback_avx+0x328> .byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11 - .byte 196,98,125,24,37,252,52,0,0 // vbroadcastss 0x34fc(%rip),%ymm12 # 6980 <_sk_callback_avx+0x32c> + .byte 196,98,125,24,37,86,52,0,0 // vbroadcastss 0x3456(%rip),%ymm12 # 67a8 <_sk_callback_avx+0x32c> .byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3 - .byte 196,98,125,24,37,242,52,0,0 // vbroadcastss 0x34f2(%rip),%ymm12 # 6984 <_sk_callback_avx+0x330> + .byte 196,98,125,24,37,76,52,0,0 // vbroadcastss 0x344c(%rip),%ymm12 # 67ac <_sk_callback_avx+0x330> .byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3 .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10 - .byte 196,98,125,24,29,214,52,0,0 // vbroadcastss 0x34d6(%rip),%ymm11 # 6988 <_sk_callback_avx+0x334> + .byte 196,98,125,24,29,48,52,0,0 // vbroadcastss 0x3430(%rip),%ymm11 # 67b0 <_sk_callback_avx+0x334> .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,204,52,0,0 // vbroadcastss 0x34cc(%rip),%ymm11 # 698c <_sk_callback_avx+0x338> + .byte 196,98,125,24,29,38,52,0,0 // vbroadcastss 0x3426(%rip),%ymm11 # 67b4 <_sk_callback_avx+0x338> .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11 .byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3 - .byte 196,98,125,24,29,189,52,0,0 // vbroadcastss 0x34bd(%rip),%ymm11 # 6990 <_sk_callback_avx+0x33c> + .byte 196,98,125,24,29,23,52,0,0 // vbroadcastss 0x3417(%rip),%ymm11 # 67b8 <_sk_callback_avx+0x33c> .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 - .byte 196,98,125,24,29,179,52,0,0 // vbroadcastss 0x34b3(%rip),%ymm11 # 6994 <_sk_callback_avx+0x340> + .byte 196,98,125,24,29,13,52,0,0 // vbroadcastss 0x340d(%rip),%ymm11 # 67bc <_sk_callback_avx+0x340> .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 - .byte 196,98,125,24,21,164,52,0,0 // vbroadcastss 0x34a4(%rip),%ymm10 # 6998 <_sk_callback_avx+0x344> + .byte 196,98,125,24,21,254,51,0,0 // vbroadcastss 0x33fe(%rip),%ymm10 # 67c0 <_sk_callback_avx+0x344> .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 @@ -18263,7 +18173,7 @@ _sk_parametric_a_avx: .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 - .byte 196,98,125,24,5,123,52,0,0 // vbroadcastss 0x347b(%rip),%ymm8 # 699c <_sk_callback_avx+0x348> + .byte 196,98,125,24,5,213,51,0,0 // vbroadcastss 0x33d5(%rip),%ymm8 # 67c4 <_sk_callback_avx+0x348> .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18272,31 +18182,31 @@ HIDDEN _sk_lab_to_xyz_avx .globl _sk_lab_to_xyz_avx FUNCTION(_sk_lab_to_xyz_avx) _sk_lab_to_xyz_avx: - .byte 196,98,125,24,5,109,52,0,0 // vbroadcastss 0x346d(%rip),%ymm8 # 69a0 <_sk_callback_avx+0x34c> + .byte 196,98,125,24,5,199,51,0,0 // vbroadcastss 0x33c7(%rip),%ymm8 # 67c8 <_sk_callback_avx+0x34c> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,99,52,0,0 // vbroadcastss 0x3463(%rip),%ymm8 # 69a4 <_sk_callback_avx+0x350> + .byte 196,98,125,24,5,189,51,0,0 // vbroadcastss 0x33bd(%rip),%ymm8 # 67cc <_sk_callback_avx+0x350> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,98,125,24,13,89,52,0,0 // vbroadcastss 0x3459(%rip),%ymm9 # 69a8 <_sk_callback_avx+0x354> + .byte 196,98,125,24,13,179,51,0,0 // vbroadcastss 0x33b3(%rip),%ymm9 # 67d0 <_sk_callback_avx+0x354> .byte 196,193,116,88,201 // vaddps %ymm9,%ymm1,%ymm1 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 - .byte 196,98,125,24,5,69,52,0,0 // vbroadcastss 0x3445(%rip),%ymm8 # 69ac <_sk_callback_avx+0x358> + .byte 196,98,125,24,5,159,51,0,0 // vbroadcastss 0x339f(%rip),%ymm8 # 67d4 <_sk_callback_avx+0x358> .byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,59,52,0,0 // vbroadcastss 0x343b(%rip),%ymm8 # 69b0 <_sk_callback_avx+0x35c> + .byte 196,98,125,24,5,149,51,0,0 // vbroadcastss 0x3395(%rip),%ymm8 # 67d8 <_sk_callback_avx+0x35c> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,5,49,52,0,0 // vbroadcastss 0x3431(%rip),%ymm8 # 69b4 <_sk_callback_avx+0x360> + .byte 196,98,125,24,5,139,51,0,0 // vbroadcastss 0x338b(%rip),%ymm8 # 67dc <_sk_callback_avx+0x360> .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1 - .byte 196,98,125,24,5,35,52,0,0 // vbroadcastss 0x3423(%rip),%ymm8 # 69b8 <_sk_callback_avx+0x364> + .byte 196,98,125,24,5,125,51,0,0 // vbroadcastss 0x337d(%rip),%ymm8 # 67e0 <_sk_callback_avx+0x364> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 197,252,92,210 // vsubps %ymm2,%ymm0,%ymm2 .byte 197,116,89,193 // vmulps %ymm1,%ymm1,%ymm8 .byte 196,65,116,89,192 // vmulps %ymm8,%ymm1,%ymm8 - .byte 196,98,125,24,13,12,52,0,0 // vbroadcastss 0x340c(%rip),%ymm9 # 69bc <_sk_callback_avx+0x368> + .byte 196,98,125,24,13,102,51,0,0 // vbroadcastss 0x3366(%rip),%ymm9 # 67e4 <_sk_callback_avx+0x368> .byte 196,65,52,194,208,1 // vcmpltps %ymm8,%ymm9,%ymm10 - .byte 196,98,125,24,29,1,52,0,0 // vbroadcastss 0x3401(%rip),%ymm11 # 69c0 <_sk_callback_avx+0x36c> + .byte 196,98,125,24,29,91,51,0,0 // vbroadcastss 0x335b(%rip),%ymm11 # 67e8 <_sk_callback_avx+0x36c> .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 - .byte 196,98,125,24,37,247,51,0,0 // vbroadcastss 0x33f7(%rip),%ymm12 # 69c4 <_sk_callback_avx+0x370> + .byte 196,98,125,24,37,81,51,0,0 // vbroadcastss 0x3351(%rip),%ymm12 # 67ec <_sk_callback_avx+0x370> .byte 196,193,116,89,204 // vmulps %ymm12,%ymm1,%ymm1 .byte 196,67,117,74,192,160 // vblendvps %ymm10,%ymm8,%ymm1,%ymm8 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1 @@ -18311,9 +18221,9 @@ _sk_lab_to_xyz_avx: .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2 .byte 196,193,108,89,212 // vmulps %ymm12,%ymm2,%ymm2 .byte 196,227,109,74,208,144 // vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - .byte 196,226,125,24,5,173,51,0,0 // vbroadcastss 0x33ad(%rip),%ymm0 # 69c8 <_sk_callback_avx+0x374> + .byte 196,226,125,24,5,7,51,0,0 // vbroadcastss 0x3307(%rip),%ymm0 # 67f0 <_sk_callback_avx+0x374> .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,164,51,0,0 // vbroadcastss 0x33a4(%rip),%ymm8 # 69cc <_sk_callback_avx+0x378> + .byte 196,98,125,24,5,254,50,0,0 // vbroadcastss 0x32fe(%rip),%ymm8 # 67f4 <_sk_callback_avx+0x378> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18327,14 +18237,14 @@ _sk_load_a8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,62 // jne 367f <_sk_load_a8_avx+0x4e> + .byte 117,62 // jne 354d <_sk_load_a8_avx+0x4e> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,104,51,0,0 // vbroadcastss 0x3368(%rip),%ymm1 # 69d0 <_sk_callback_avx+0x37c> + .byte 196,226,125,24,13,194,50,0,0 // vbroadcastss 0x32c2(%rip),%ymm1 # 67f8 <_sk_callback_avx+0x37c> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -18351,9 +18261,9 @@ _sk_load_a8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 3687 <_sk_load_a8_avx+0x56> + .byte 117,234 // jne 3555 <_sk_load_a8_avx+0x56> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,161 // jmp 3645 <_sk_load_a8_avx+0x14> + .byte 235,161 // jmp 3513 <_sk_load_a8_avx+0x14> HIDDEN _sk_gather_a8_avx .globl _sk_gather_a8_avx @@ -18403,7 +18313,7 @@ _sk_gather_a8_avx: .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,93,50,0,0 // vbroadcastss 0x325d(%rip),%ymm1 # 69d4 <_sk_callback_avx+0x380> + .byte 196,226,125,24,13,183,49,0,0 // vbroadcastss 0x31b7(%rip),%ymm1 # 67fc <_sk_callback_avx+0x380> .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 @@ -18421,14 +18331,14 @@ FUNCTION(_sk_store_a8_avx) _sk_store_a8_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,56,50,0,0 // vbroadcastss 0x3238(%rip),%ymm8 # 69d8 <_sk_callback_avx+0x384> + .byte 196,98,125,24,5,146,49,0,0 // vbroadcastss 0x3192(%rip),%ymm8 # 6800 <_sk_callback_avx+0x384> .byte 196,65,100,89,192 // vmulps %ymm8,%ymm3,%ymm8 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 37c9 <_sk_store_a8_avx+0x37> + .byte 117,10 // jne 3697 <_sk_store_a8_avx+0x37> .byte 196,65,123,17,4,58 // vmovsd %xmm8,(%r10,%rdi,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18436,10 +18346,10 @@ _sk_store_a8_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 37c5 <_sk_store_a8_avx+0x33> + .byte 119,236 // ja 3693 <_sk_store_a8_avx+0x33> .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,67,0,0,0 // lea 0x43(%rip),%r9 # 382c <_sk_store_a8_avx+0x9a> + .byte 76,141,13,69,0,0,0 // lea 0x45(%rip),%r9 # 36fc <_sk_store_a8_avx+0x9c> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -18450,27 +18360,28 @@ _sk_store_a8_avx: .byte 196,67,121,20,68,58,2,4 // vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1) .byte 196,67,121,20,68,58,1,2 // vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1) .byte 196,67,121,20,4,58,0 // vpextrb $0x0,%xmm8,(%r10,%rdi,1) - .byte 235,154 // jmp 37c5 <_sk_store_a8_avx+0x33> - .byte 144 // nop - .byte 246,255 // idiv %bh + .byte 235,154 // jmp 3693 <_sk_store_a8_avx+0x33> + .byte 15,31,0 // nopl (%rax) + .byte 244 // hlt .byte 255 // (bad) .byte 255 // (bad) - .byte 238 // out %al,(%dx) .byte 255 // (bad) + .byte 236 // in (%dx),%al .byte 255 // (bad) - .byte 255,230 // jmpq *%rsi .byte 255 // (bad) + .byte 255,228 // jmpq *%rsp .byte 255 // (bad) .byte 255 // (bad) - .byte 222,255 // fdivrp %st,%st(7) .byte 255 // (bad) - .byte 255,214 // callq *%rsi + .byte 220,255 // fdivr %st,%st(7) .byte 255 // (bad) + .byte 255,212 // callq *%rsp .byte 255 // (bad) - .byte 255,206 // dec %esi .byte 255 // (bad) + .byte 255,204 // dec %esp .byte 255 // (bad) - .byte 255,198 // inc %esi + .byte 255 // (bad) + .byte 255,196 // inc %esp .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -18484,17 +18395,17 @@ _sk_load_g8_avx: .byte 72,139,0 // mov (%rax),%rax .byte 72,1,248 // add %rdi,%rax .byte 77,133,192 // test %r8,%r8 - .byte 117,67 // jne 389b <_sk_load_g8_avx+0x53> + .byte 117,67 // jne 376b <_sk_load_g8_avx+0x53> .byte 197,250,126,0 // vmovq (%rax),%xmm0 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,93,49,0,0 // vbroadcastss 0x315d(%rip),%ymm1 # 69dc <_sk_callback_avx+0x388> + .byte 196,226,125,24,13,181,48,0,0 // vbroadcastss 0x30b5(%rip),%ymm1 # 6804 <_sk_callback_avx+0x388> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,82,49,0,0 // vbroadcastss 0x3152(%rip),%ymm3 # 69e0 <_sk_callback_avx+0x38c> + .byte 196,226,125,24,29,170,48,0,0 // vbroadcastss 0x30aa(%rip),%ymm3 # 6808 <_sk_callback_avx+0x38c> .byte 76,137,193 // mov %r8,%rcx .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 @@ -18508,9 +18419,9 @@ _sk_load_g8_avx: .byte 77,9,217 // or %r11,%r9 .byte 72,131,193,8 // add $0x8,%rcx .byte 73,255,202 // dec %r10 - .byte 117,234 // jne 38a3 <_sk_load_g8_avx+0x5b> + .byte 117,234 // jne 3773 <_sk_load_g8_avx+0x5b> .byte 196,193,249,110,193 // vmovq %r9,%xmm0 - .byte 235,156 // jmp 385c <_sk_load_g8_avx+0x14> + .byte 235,156 // jmp 372c <_sk_load_g8_avx+0x14> HIDDEN _sk_gather_g8_avx .globl _sk_gather_g8_avx @@ -18560,10 +18471,10 @@ _sk_gather_g8_avx: .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,81,48,0,0 // vbroadcastss 0x3051(%rip),%ymm1 # 69e4 <_sk_callback_avx+0x390> + .byte 196,226,125,24,13,169,47,0,0 // vbroadcastss 0x2fa9(%rip),%ymm1 # 680c <_sk_callback_avx+0x390> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,70,48,0,0 // vbroadcastss 0x3046(%rip),%ymm3 # 69e8 <_sk_callback_avx+0x394> + .byte 196,226,125,24,29,158,47,0,0 // vbroadcastss 0x2f9e(%rip),%ymm3 # 6810 <_sk_callback_avx+0x394> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 .byte 91 // pop %rbx @@ -18579,9 +18490,9 @@ _sk_gather_i8_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 73,137,192 // mov %rax,%r8 .byte 77,133,192 // test %r8,%r8 - .byte 116,5 // je 39c2 <_sk_gather_i8_avx+0xf> + .byte 116,5 // je 3892 <_sk_gather_i8_avx+0xf> .byte 76,137,192 // mov %r8,%rax - .byte 235,2 // jmp 39c4 <_sk_gather_i8_avx+0x11> + .byte 235,2 // jmp 3894 <_sk_gather_i8_avx+0x11> .byte 72,173 // lods %ds:(%rsi),%rax .byte 65,87 // push %r15 .byte 65,86 // push %r14 @@ -18643,10 +18554,10 @@ _sk_gather_i8_avx: .byte 196,163,121,34,4,163,2 // vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0 .byte 196,163,121,34,28,19,3 // vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - .byte 197,124,40,21,18,49,0,0 // vmovaps 0x3112(%rip),%ymm10 # 6c00 <_sk_callback_avx+0x5ac> + .byte 197,124,40,21,2,49,0,0 // vmovaps 0x3102(%rip),%ymm10 # 6ac0 <_sk_callback_avx+0x644> .byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,236,46,0,0 // vbroadcastss 0x2eec(%rip),%ymm9 # 69ec <_sk_callback_avx+0x398> + .byte 196,98,125,24,13,68,46,0,0 // vbroadcastss 0x2e44(%rip),%ymm9 # 6814 <_sk_callback_avx+0x398> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 @@ -18680,38 +18591,38 @@ _sk_load_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,128,0,0,0 // jne 3bf8 <_sk_load_565_avx+0x8e> + .byte 15,133,128,0,0,0 // jne 3ac8 <_sk_load_565_avx+0x8e> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - .byte 196,226,125,24,5,86,46,0,0 // vbroadcastss 0x2e56(%rip),%ymm0 # 69f0 <_sk_callback_avx+0x39c> + .byte 196,226,125,24,5,174,45,0,0 // vbroadcastss 0x2dae(%rip),%ymm0 # 6818 <_sk_callback_avx+0x39c> .byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,73,46,0,0 // vbroadcastss 0x2e49(%rip),%ymm1 # 69f4 <_sk_callback_avx+0x3a0> + .byte 196,226,125,24,13,161,45,0,0 // vbroadcastss 0x2da1(%rip),%ymm1 # 681c <_sk_callback_avx+0x3a0> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,64,46,0,0 // vbroadcastss 0x2e40(%rip),%ymm1 # 69f8 <_sk_callback_avx+0x3a4> + .byte 196,226,125,24,13,152,45,0,0 // vbroadcastss 0x2d98(%rip),%ymm1 # 6820 <_sk_callback_avx+0x3a4> .byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,51,46,0,0 // vbroadcastss 0x2e33(%rip),%ymm3 # 69fc <_sk_callback_avx+0x3a8> + .byte 196,226,125,24,29,139,45,0,0 // vbroadcastss 0x2d8b(%rip),%ymm3 # 6824 <_sk_callback_avx+0x3a8> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,24,29,42,46,0,0 // vbroadcastss 0x2e2a(%rip),%ymm3 # 6a00 <_sk_callback_avx+0x3ac> + .byte 196,226,125,24,29,130,45,0,0 // vbroadcastss 0x2d82(%rip),%ymm3 # 6828 <_sk_callback_avx+0x3ac> .byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,29,46,0,0 // vbroadcastss 0x2e1d(%rip),%ymm3 # 6a04 <_sk_callback_avx+0x3b0> + .byte 196,226,125,24,29,117,45,0,0 // vbroadcastss 0x2d75(%rip),%ymm3 # 682c <_sk_callback_avx+0x3b0> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,18,46,0,0 // vbroadcastss 0x2e12(%rip),%ymm3 # 6a08 <_sk_callback_avx+0x3b4> + .byte 196,226,125,24,29,106,45,0,0 // vbroadcastss 0x2d6a(%rip),%ymm3 # 6830 <_sk_callback_avx+0x3b4> .byte 255,224 // jmpq *%rax .byte 65,137,200 // mov %ecx,%r8d .byte 65,128,224,7 // and $0x7,%r8b .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,110,255,255,255 // ja 3b7e <_sk_load_565_avx+0x14> + .byte 15,135,110,255,255,255 // ja 3a4e <_sk_load_565_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 3c64 <_sk_load_565_avx+0xfa> + .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 3b34 <_sk_load_565_avx+0xfa> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -18723,7 +18634,7 @@ _sk_load_565_avx: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,26,255,255,255 // jmpq 3b7e <_sk_load_565_avx+0x14> + .byte 233,26,255,255,255 // jmpq 3a4e <_sk_load_565_avx+0x14> .byte 244 // hlt .byte 255 // (bad) .byte 255 // (bad) @@ -18801,23 +18712,23 @@ _sk_gather_565_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - .byte 196,226,125,24,5,178,44,0,0 // vbroadcastss 0x2cb2(%rip),%ymm0 # 6a0c <_sk_callback_avx+0x3b8> + .byte 196,226,125,24,5,10,44,0,0 // vbroadcastss 0x2c0a(%rip),%ymm0 # 6834 <_sk_callback_avx+0x3b8> .byte 197,236,84,192 // vandps %ymm0,%ymm2,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,165,44,0,0 // vbroadcastss 0x2ca5(%rip),%ymm1 # 6a10 <_sk_callback_avx+0x3bc> + .byte 196,226,125,24,13,253,43,0,0 // vbroadcastss 0x2bfd(%rip),%ymm1 # 6838 <_sk_callback_avx+0x3bc> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,156,44,0,0 // vbroadcastss 0x2c9c(%rip),%ymm1 # 6a14 <_sk_callback_avx+0x3c0> + .byte 196,226,125,24,13,244,43,0,0 // vbroadcastss 0x2bf4(%rip),%ymm1 # 683c <_sk_callback_avx+0x3c0> .byte 197,236,84,201 // vandps %ymm1,%ymm2,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,29,143,44,0,0 // vbroadcastss 0x2c8f(%rip),%ymm3 # 6a18 <_sk_callback_avx+0x3c4> + .byte 196,226,125,24,29,231,43,0,0 // vbroadcastss 0x2be7(%rip),%ymm3 # 6840 <_sk_callback_avx+0x3c4> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 - .byte 196,226,125,24,29,134,44,0,0 // vbroadcastss 0x2c86(%rip),%ymm3 # 6a1c <_sk_callback_avx+0x3c8> + .byte 196,226,125,24,29,222,43,0,0 // vbroadcastss 0x2bde(%rip),%ymm3 # 6844 <_sk_callback_avx+0x3c8> .byte 197,236,84,211 // vandps %ymm3,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,226,125,24,29,121,44,0,0 // vbroadcastss 0x2c79(%rip),%ymm3 # 6a20 <_sk_callback_avx+0x3cc> + .byte 196,226,125,24,29,209,43,0,0 // vbroadcastss 0x2bd1(%rip),%ymm3 # 6848 <_sk_callback_avx+0x3cc> .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,110,44,0,0 // vbroadcastss 0x2c6e(%rip),%ymm3 # 6a24 <_sk_callback_avx+0x3d0> + .byte 196,226,125,24,29,198,43,0,0 // vbroadcastss 0x2bc6(%rip),%ymm3 # 684c <_sk_callback_avx+0x3d0> .byte 91 // pop %rbx .byte 65,92 // pop %r12 .byte 65,94 // pop %r14 @@ -18831,14 +18742,14 @@ FUNCTION(_sk_store_565_avx) _sk_store_565_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,90,44,0,0 // vbroadcastss 0x2c5a(%rip),%ymm8 # 6a28 <_sk_callback_avx+0x3d4> + .byte 196,98,125,24,5,178,43,0,0 // vbroadcastss 0x2bb2(%rip),%ymm8 # 6850 <_sk_callback_avx+0x3d4> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 .byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - .byte 196,98,125,24,21,51,44,0,0 // vbroadcastss 0x2c33(%rip),%ymm10 # 6a2c <_sk_callback_avx+0x3d8> + .byte 196,98,125,24,21,139,43,0,0 // vbroadcastss 0x2b8b(%rip),%ymm10 # 6854 <_sk_callback_avx+0x3d8> .byte 196,65,116,89,210 // vmulps %ymm10,%ymm1,%ymm10 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 .byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11 @@ -18852,7 +18763,7 @@ _sk_store_565_avx: .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 3e49 <_sk_store_565_avx+0x89> + .byte 117,10 // jne 3d19 <_sk_store_565_avx+0x89> .byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18860,9 +18771,9 @@ _sk_store_565_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 3e45 <_sk_store_565_avx+0x85> + .byte 119,236 // ja 3d15 <_sk_store_565_avx+0x85> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 3ea8 <_sk_store_565_avx+0xe8> + .byte 76,141,13,68,0,0,0 // lea 0x44(%rip),%r9 # 3d78 <_sk_store_565_avx+0xe8> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -18873,7 +18784,7 @@ _sk_store_565_avx: .byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) .byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) .byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2) - .byte 235,159 // jmp 3e45 <_sk_store_565_avx+0x85> + .byte 235,159 // jmp 3d15 <_sk_store_565_avx+0x85> .byte 102,144 // xchg %ax,%ax .byte 245 // cmc .byte 255 // (bad) @@ -18906,31 +18817,31 @@ _sk_load_4444_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,152,0,0,0 // jne 3f6a <_sk_load_4444_avx+0xa6> + .byte 15,133,152,0,0,0 // jne 3e3a <_sk_load_4444_avx+0xa6> .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - .byte 196,226,125,24,5,60,43,0,0 // vbroadcastss 0x2b3c(%rip),%ymm0 # 6a30 <_sk_callback_avx+0x3dc> + .byte 196,226,125,24,5,148,42,0,0 // vbroadcastss 0x2a94(%rip),%ymm0 # 6858 <_sk_callback_avx+0x3dc> .byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,47,43,0,0 // vbroadcastss 0x2b2f(%rip),%ymm1 # 6a34 <_sk_callback_avx+0x3e0> + .byte 196,226,125,24,13,135,42,0,0 // vbroadcastss 0x2a87(%rip),%ymm1 # 685c <_sk_callback_avx+0x3e0> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,38,43,0,0 // vbroadcastss 0x2b26(%rip),%ymm1 # 6a38 <_sk_callback_avx+0x3e4> + .byte 196,226,125,24,13,126,42,0,0 // vbroadcastss 0x2a7e(%rip),%ymm1 # 6860 <_sk_callback_avx+0x3e4> .byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,25,43,0,0 // vbroadcastss 0x2b19(%rip),%ymm2 # 6a3c <_sk_callback_avx+0x3e8> + .byte 196,226,125,24,21,113,42,0,0 // vbroadcastss 0x2a71(%rip),%ymm2 # 6864 <_sk_callback_avx+0x3e8> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,16,43,0,0 // vbroadcastss 0x2b10(%rip),%ymm2 # 6a40 <_sk_callback_avx+0x3ec> + .byte 196,226,125,24,21,104,42,0,0 // vbroadcastss 0x2a68(%rip),%ymm2 # 6868 <_sk_callback_avx+0x3ec> .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,3,43,0,0 // vbroadcastss 0x2b03(%rip),%ymm8 # 6a44 <_sk_callback_avx+0x3f0> + .byte 196,98,125,24,5,91,42,0,0 // vbroadcastss 0x2a5b(%rip),%ymm8 # 686c <_sk_callback_avx+0x3f0> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,249,42,0,0 // vbroadcastss 0x2af9(%rip),%ymm8 # 6a48 <_sk_callback_avx+0x3f4> + .byte 196,98,125,24,5,81,42,0,0 // vbroadcastss 0x2a51(%rip),%ymm8 # 6870 <_sk_callback_avx+0x3f4> .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,235,42,0,0 // vbroadcastss 0x2aeb(%rip),%ymm8 # 6a4c <_sk_callback_avx+0x3f8> + .byte 196,98,125,24,5,67,42,0,0 // vbroadcastss 0x2a43(%rip),%ymm8 # 6874 <_sk_callback_avx+0x3f8> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -18939,9 +18850,9 @@ _sk_load_4444_avx: .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,86,255,255,255 // ja 3ed8 <_sk_load_4444_avx+0x14> + .byte 15,135,86,255,255,255 // ja 3da8 <_sk_load_4444_avx+0x14> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3fd8 <_sk_load_4444_avx+0x114> + .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3ea8 <_sk_load_4444_avx+0x114> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -18953,7 +18864,7 @@ _sk_load_4444_avx: .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - .byte 233,2,255,255,255 // jmpq 3ed8 <_sk_load_4444_avx+0x14> + .byte 233,2,255,255,255 // jmpq 3da8 <_sk_load_4444_avx+0x14> .byte 102,144 // xchg %ax,%ax .byte 242,255 // repnz (bad) .byte 255 // (bad) @@ -19032,25 +18943,25 @@ _sk_gather_4444_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - .byte 196,226,125,24,5,130,41,0,0 // vbroadcastss 0x2982(%rip),%ymm0 # 6a50 <_sk_callback_avx+0x3fc> + .byte 196,226,125,24,5,218,40,0,0 // vbroadcastss 0x28da(%rip),%ymm0 # 6878 <_sk_callback_avx+0x3fc> .byte 197,228,84,192 // vandps %ymm0,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,226,125,24,13,117,41,0,0 // vbroadcastss 0x2975(%rip),%ymm1 # 6a54 <_sk_callback_avx+0x400> + .byte 196,226,125,24,13,205,40,0,0 // vbroadcastss 0x28cd(%rip),%ymm1 # 687c <_sk_callback_avx+0x400> .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,108,41,0,0 // vbroadcastss 0x296c(%rip),%ymm1 # 6a58 <_sk_callback_avx+0x404> + .byte 196,226,125,24,13,196,40,0,0 // vbroadcastss 0x28c4(%rip),%ymm1 # 6880 <_sk_callback_avx+0x404> .byte 197,228,84,201 // vandps %ymm1,%ymm3,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 - .byte 196,226,125,24,21,95,41,0,0 // vbroadcastss 0x295f(%rip),%ymm2 # 6a5c <_sk_callback_avx+0x408> + .byte 196,226,125,24,21,183,40,0,0 // vbroadcastss 0x28b7(%rip),%ymm2 # 6884 <_sk_callback_avx+0x408> .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 - .byte 196,226,125,24,21,86,41,0,0 // vbroadcastss 0x2956(%rip),%ymm2 # 6a60 <_sk_callback_avx+0x40c> + .byte 196,226,125,24,21,174,40,0,0 // vbroadcastss 0x28ae(%rip),%ymm2 # 6888 <_sk_callback_avx+0x40c> .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 - .byte 196,98,125,24,5,73,41,0,0 // vbroadcastss 0x2949(%rip),%ymm8 # 6a64 <_sk_callback_avx+0x410> + .byte 196,98,125,24,5,161,40,0,0 // vbroadcastss 0x28a1(%rip),%ymm8 # 688c <_sk_callback_avx+0x410> .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,98,125,24,5,63,41,0,0 // vbroadcastss 0x293f(%rip),%ymm8 # 6a68 <_sk_callback_avx+0x414> + .byte 196,98,125,24,5,151,40,0,0 // vbroadcastss 0x2897(%rip),%ymm8 # 6890 <_sk_callback_avx+0x414> .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 - .byte 196,98,125,24,5,49,41,0,0 // vbroadcastss 0x2931(%rip),%ymm8 # 6a6c <_sk_callback_avx+0x418> + .byte 196,98,125,24,5,137,40,0,0 // vbroadcastss 0x2889(%rip),%ymm8 # 6894 <_sk_callback_avx+0x418> .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 91 // pop %rbx @@ -19066,7 +18977,7 @@ FUNCTION(_sk_store_4444_avx) _sk_store_4444_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,22,41,0,0 // vbroadcastss 0x2916(%rip),%ymm8 # 6a70 <_sk_callback_avx+0x41c> + .byte 196,98,125,24,5,110,40,0,0 // vbroadcastss 0x286e(%rip),%ymm8 # 6898 <_sk_callback_avx+0x41c> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10 @@ -19093,7 +19004,7 @@ _sk_store_4444_avx: .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 41f3 <_sk_store_4444_avx+0xa7> + .byte 117,10 // jne 40c3 <_sk_store_4444_avx+0xa7> .byte 196,65,122,127,4,122 // vmovdqu %xmm8,(%r10,%rdi,2) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -19101,9 +19012,9 @@ _sk_store_4444_avx: .byte 65,128,224,7 // and $0x7,%r8b .byte 65,254,200 // dec %r8b .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 41ef <_sk_store_4444_avx+0xa3> + .byte 119,236 // ja 40bf <_sk_store_4444_avx+0xa3> .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,66,0,0,0 // lea 0x42(%rip),%r9 # 4250 <_sk_store_4444_avx+0x104> + .byte 76,141,13,66,0,0,0 // lea 0x42(%rip),%r9 # 4120 <_sk_store_4444_avx+0x104> .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -19114,7 +19025,7 @@ _sk_store_4444_avx: .byte 196,67,121,21,68,122,4,2 // vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) .byte 196,67,121,21,68,122,2,1 // vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) .byte 196,67,121,21,4,122,0 // vpextrw $0x0,%xmm8,(%r10,%rdi,2) - .byte 235,159 // jmp 41ef <_sk_store_4444_avx+0xa3> + .byte 235,159 // jmp 40bf <_sk_store_4444_avx+0xa3> .byte 247,255 // idiv %edi .byte 255 // (bad) .byte 255 // (bad) @@ -19142,87 +19053,53 @@ HIDDEN _sk_load_8888_avx .globl _sk_load_8888_avx FUNCTION(_sk_load_8888_avx) _sk_load_8888_avx: + .byte 73,137,200 // mov %rcx,%r8 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 76,139,16 // mov (%rax),%r10 - .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,135,0,0,0 // jne 4301 <_sk_load_8888_avx+0x95> - .byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9 - .byte 197,124,40,21,152,41,0,0 // vmovaps 0x2998(%rip),%ymm10 # 6c20 <_sk_callback_avx+0x5cc> - .byte 196,193,52,84,194 // vandps %ymm10,%ymm9,%ymm0 + .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 + .byte 76,3,8 // add (%rax),%r9 + .byte 77,133,192 // test %r8,%r8 + .byte 15,133,137,0,0,0 // jne 41de <_sk_load_8888_avx+0xa2> + .byte 196,193,124,16,25 // vmovups (%r9),%ymm3 + .byte 197,124,40,21,126,41,0,0 // vmovaps 0x297e(%rip),%ymm10 # 6ae0 <_sk_callback_avx+0x664> + .byte 196,193,100,84,194 // vandps %ymm10,%ymm3,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,5,218,39,0,0 // vbroadcastss 0x27da(%rip),%ymm8 # 6a74 <_sk_callback_avx+0x420> + .byte 196,98,125,24,5,40,39,0,0 // vbroadcastss 0x2728(%rip),%ymm8 # 689c <_sk_callback_avx+0x420> .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 - .byte 196,193,113,114,209,8 // vpsrld $0x8,%xmm9,%xmm1 - .byte 196,99,125,25,203,1 // vextractf128 $0x1,%ymm9,%xmm3 - .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 + .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1 + .byte 196,195,125,25,217,1 // vextractf128 $0x1,%ymm3,%xmm9 + .byte 196,193,105,114,209,8 // vpsrld $0x8,%xmm9,%xmm2 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 .byte 196,193,116,84,202 // vandps %ymm10,%ymm1,%ymm1 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 - .byte 196,193,33,114,209,16 // vpsrld $0x10,%xmm9,%xmm11 - .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2 + .byte 197,161,114,211,16 // vpsrld $0x10,%xmm3,%xmm11 + .byte 196,193,105,114,209,16 // vpsrld $0x10,%xmm9,%xmm2 .byte 196,227,37,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm11,%ymm2 .byte 196,193,108,84,210 // vandps %ymm10,%ymm2,%ymm2 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 - .byte 196,193,49,114,209,24 // vpsrld $0x18,%xmm9,%xmm9 - .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3 - .byte 196,227,53,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 + .byte 197,169,114,211,24 // vpsrld $0x18,%xmm3,%xmm10 + .byte 196,193,97,114,209,24 // vpsrld $0x18,%xmm9,%xmm3 + .byte 196,227,45,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm10,%ymm3 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,137,193 // mov %r8,%rcx .byte 255,224 // jmpq *%rax - .byte 65,137,200 // mov %ecx,%r8d - .byte 65,128,224,7 // and $0x7,%r8b - .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 - .byte 65,254,200 // dec %r8b - .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 15,135,102,255,255,255 // ja 4280 <_sk_load_8888_avx+0x14> - .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,139,0,0,0 // lea 0x8b(%rip),%r9 # 43b0 <_sk_load_8888_avx+0x144> - .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax - .byte 76,1,200 // add %r9,%rax - .byte 255,224 // jmpq *%rax - .byte 196,193,121,110,68,186,24 // vmovd 0x18(%r10,%rdi,4),%xmm0 - .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0 - .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 - .byte 196,99,117,12,200,64 // vblendps $0x40,%ymm0,%ymm1,%ymm9 - .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 - .byte 196,195,121,34,68,186,20,1 // vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0 - .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 - .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 - .byte 196,195,121,34,68,186,16,0 // vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0 - .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 - .byte 196,195,49,34,68,186,12,3 // vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm9,%xmm0 - .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 - .byte 196,195,49,34,68,186,8,2 // vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm9,%xmm0 - .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 - .byte 196,195,49,34,68,186,4,1 // vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm9,%xmm0 - .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 - .byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 - .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 - .byte 233,210,254,255,255 // jmpq 4280 <_sk_load_8888_avx+0x14> - .byte 102,144 // xchg %ax,%ax - .byte 236 // in (%dx),%al - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 222,255 // fdivrp %st,%st(7) - .byte 255 // (bad) - .byte 255,208 // callq *%rax - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,194 // inc %edx - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,174,255,255,255,154 // ljmp *-0x65000001(%rsi) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 126,255 // jle 43c9 <_sk_load_8888_avx+0x15d> - .byte 255 // (bad) - .byte 255 // .byte 0xff + .byte 185,8,0,0,0 // mov $0x8,%ecx + .byte 68,41,193 // sub %r8d,%ecx + .byte 192,225,3 // shl $0x3,%cl + .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax + .byte 72,211,232 // shr %cl,%rax + .byte 196,225,249,110,192 // vmovq %rax,%xmm0 + .byte 196,226,121,48,192 // vpmovzxbw %xmm0,%xmm0 + .byte 196,226,121,0,13,58,40,0,0 // vpshufb 0x283a(%rip),%xmm0,%xmm1 # 6a40 <_sk_callback_avx+0x5c4> + .byte 196,226,121,33,201 // vpmovsxbd %xmm1,%xmm1 + .byte 196,226,121,0,5,60,40,0,0 // vpshufb 0x283c(%rip),%xmm0,%xmm0 # 6a50 <_sk_callback_avx+0x5d4> + .byte 196,226,121,33,192 // vpmovsxbd %xmm0,%xmm0 + .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + .byte 196,194,125,44,25 // vmaskmovps (%r9),%ymm0,%ymm3 + .byte 233,49,255,255,255 // jmpq 415a <_sk_load_8888_avx+0x1e> HIDDEN _sk_gather_8888_avx .globl _sk_gather_8888_avx @@ -19265,10 +19142,10 @@ _sk_gather_8888_avx: .byte 196,131,121,34,4,152,2 // vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 .byte 196,131,121,34,28,144,3 // vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - .byte 197,124,40,21,194,39,0,0 // vmovaps 0x27c2(%rip),%ymm10 # 6c40 <_sk_callback_avx+0x5ec> + .byte 197,124,40,21,37,40,0,0 // vmovaps 0x2825(%rip),%ymm10 # 6b00 <_sk_callback_avx+0x684> .byte 196,193,124,84,194 // vandps %ymm10,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,13,232,37,0,0 // vbroadcastss 0x25e8(%rip),%ymm9 # 6a78 <_sk_callback_avx+0x424> + .byte 196,98,125,24,13,179,37,0,0 // vbroadcastss 0x25b3(%rip),%ymm9 # 68a0 <_sk_callback_avx+0x424> .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 .byte 197,233,114,211,8 // vpsrld $0x8,%xmm3,%xmm2 @@ -19298,9 +19175,11 @@ HIDDEN _sk_store_8888_avx .globl _sk_store_8888_avx FUNCTION(_sk_store_8888_avx) _sk_store_8888_avx: + .byte 73,137,200 // mov %rcx,%r8 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 76,139,16 // mov (%rax),%r10 - .byte 196,98,125,24,5,118,37,0,0 // vbroadcastss 0x2576(%rip),%ymm8 # 6a7c <_sk_callback_avx+0x428> + .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 + .byte 76,3,8 // add (%rax),%r9 + .byte 196,98,125,24,5,54,37,0,0 // vbroadcastss 0x2536(%rip),%ymm8 # 68a4 <_sk_callback_avx+0x428> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,65,116,89,208 // vmulps %ymm8,%ymm1,%ymm10 @@ -19324,56 +19203,26 @@ _sk_store_8888_avx: .byte 196,67,37,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm11,%ymm8 .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8 .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8 - .byte 72,133,201 // test %rcx,%rcx - .byte 117,10 // jne 4594 <_sk_store_8888_avx+0x9c> - .byte 196,65,124,17,4,186 // vmovups %ymm8,(%r10,%rdi,4) + .byte 77,133,192 // test %r8,%r8 + .byte 117,12 // jne 43fe <_sk_store_8888_avx+0xa9> + .byte 196,65,124,17,1 // vmovups %ymm8,(%r9) .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,137,193 // mov %r8,%rcx .byte 255,224 // jmpq *%rax - .byte 65,137,200 // mov %ecx,%r8d - .byte 65,128,224,7 // and $0x7,%r8b - .byte 65,254,200 // dec %r8b - .byte 65,128,248,6 // cmp $0x6,%r8b - .byte 119,236 // ja 4590 <_sk_store_8888_avx+0x98> - .byte 69,15,182,192 // movzbl %r8b,%r8d - .byte 76,141,13,85,0,0,0 // lea 0x55(%rip),%r9 # 4604 <_sk_store_8888_avx+0x10c> - .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax - .byte 76,1,200 // add %r9,%rax - .byte 255,224 // jmpq *%rax - .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 - .byte 196,67,121,22,76,186,24,2 // vpextrd $0x2,%xmm9,0x18(%r10,%rdi,4) - .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 - .byte 196,67,121,22,76,186,20,1 // vpextrd $0x1,%xmm9,0x14(%r10,%rdi,4) - .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 - .byte 196,65,122,17,76,186,16 // vmovss %xmm9,0x10(%r10,%rdi,4) - .byte 196,67,121,22,68,186,12,3 // vpextrd $0x3,%xmm8,0xc(%r10,%rdi,4) - .byte 196,67,121,22,68,186,8,2 // vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4) - .byte 196,67,121,22,68,186,4,1 // vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4) - .byte 196,65,121,126,4,186 // vmovd %xmm8,(%r10,%rdi,4) - .byte 235,143 // jmp 4590 <_sk_store_8888_avx+0x98> - .byte 15,31,0 // nopl (%rax) - .byte 245 // cmc - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 237 // in (%dx),%eax - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,229 // jmpq *%rbp - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // (bad) - .byte 221,255 // (bad) - .byte 255 // (bad) - .byte 255,208 // callq *%rax - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255,194 // inc %edx - .byte 255 // (bad) - .byte 255 // (bad) - .byte 255 // .byte 0xff - .byte 180,255 // mov $0xff,%ah - .byte 255 // (bad) - .byte 255 // .byte 0xff + .byte 185,8,0,0,0 // mov $0x8,%ecx + .byte 68,41,193 // sub %r8d,%ecx + .byte 192,225,3 // shl $0x3,%cl + .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax + .byte 72,211,232 // shr %cl,%rax + .byte 196,97,249,110,200 // vmovq %rax,%xmm9 + .byte 196,66,121,48,201 // vpmovzxbw %xmm9,%xmm9 + .byte 196,98,49,0,21,58,38,0,0 // vpshufb 0x263a(%rip),%xmm9,%xmm10 # 6a60 <_sk_callback_avx+0x5e4> + .byte 196,66,121,33,210 // vpmovsxbd %xmm10,%xmm10 + .byte 196,98,49,0,13,60,38,0,0 // vpshufb 0x263c(%rip),%xmm9,%xmm9 # 6a70 <_sk_callback_avx+0x5f4> + .byte 196,66,121,33,201 // vpmovsxbd %xmm9,%xmm9 + .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 + .byte 196,66,53,46,1 // vmaskmovps %ymm8,%ymm9,(%r9) + .byte 235,177 // jmp 43f7 <_sk_store_8888_avx+0xa2> HIDDEN _sk_load_f16_avx .globl _sk_load_f16_avx @@ -19387,7 +19236,7 @@ _sk_load_f16_avx: .byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp) .byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp) .byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp) - .byte 15,133,141,2,0,0 // jne 48d7 <_sk_load_f16_avx+0x2b7> + .byte 15,133,141,2,0,0 // jne 46fd <_sk_load_f16_avx+0x2b7> .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 .byte 197,249,16,76,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm1 @@ -19405,13 +19254,13 @@ _sk_load_f16_avx: .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - .byte 196,98,125,24,37,221,35,0,0 // vbroadcastss 0x23dd(%rip),%ymm12 # 6a80 <_sk_callback_avx+0x42c> + .byte 196,98,125,24,37,223,35,0,0 // vbroadcastss 0x23df(%rip),%ymm12 # 68a8 <_sk_callback_avx+0x42c> .byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1 .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 - .byte 196,98,121,24,29,201,35,0,0 // vbroadcastss 0x23c9(%rip),%xmm11 # 6a84 <_sk_callback_avx+0x430> + .byte 196,98,121,24,29,203,35,0,0 // vbroadcastss 0x23cb(%rip),%xmm11 # 68ac <_sk_callback_avx+0x430> .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 - .byte 196,98,121,24,45,191,35,0,0 // vbroadcastss 0x23bf(%rip),%xmm13 # 6a88 <_sk_callback_avx+0x434> + .byte 196,98,121,24,45,193,35,0,0 // vbroadcastss 0x23c1(%rip),%xmm13 # 68b0 <_sk_callback_avx+0x434> .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -19425,7 +19274,7 @@ _sk_load_f16_avx: .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 .byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0 .byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1 - .byte 196,226,121,24,29,117,35,0,0 // vbroadcastss 0x2375(%rip),%xmm3 # 6a8c <_sk_callback_avx+0x438> + .byte 196,226,121,24,29,119,35,0,0 // vbroadcastss 0x2377(%rip),%xmm3 # 68b4 <_sk_callback_avx+0x438> .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -19518,29 +19367,29 @@ _sk_load_f16_avx: .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,79 // je 4936 <_sk_load_f16_avx+0x316> + .byte 116,79 // je 475c <_sk_load_f16_avx+0x316> .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,67 // jb 4936 <_sk_load_f16_avx+0x316> + .byte 114,67 // jb 475c <_sk_load_f16_avx+0x316> .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,68 // je 4943 <_sk_load_f16_avx+0x323> + .byte 116,68 // je 4769 <_sk_load_f16_avx+0x323> .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,56 // jb 4943 <_sk_load_f16_avx+0x323> + .byte 114,56 // jb 4769 <_sk_load_f16_avx+0x323> .byte 197,251,16,76,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,70,253,255,255 // je 4661 <_sk_load_f16_avx+0x41> + .byte 15,132,70,253,255,255 // je 4487 <_sk_load_f16_avx+0x41> .byte 197,241,22,76,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,54,253,255,255 // jb 4661 <_sk_load_f16_avx+0x41> + .byte 15,130,54,253,255,255 // jb 4487 <_sk_load_f16_avx+0x41> .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 - .byte 233,43,253,255,255 // jmpq 4661 <_sk_load_f16_avx+0x41> + .byte 233,43,253,255,255 // jmpq 4487 <_sk_load_f16_avx+0x41> .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,30,253,255,255 // jmpq 4661 <_sk_load_f16_avx+0x41> + .byte 233,30,253,255,255 // jmpq 4487 <_sk_load_f16_avx+0x41> .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 - .byte 233,21,253,255,255 // jmpq 4661 <_sk_load_f16_avx+0x41> + .byte 233,21,253,255,255 // jmpq 4487 <_sk_load_f16_avx+0x41> HIDDEN _sk_gather_f16_avx .globl _sk_gather_f16_avx @@ -19604,13 +19453,13 @@ _sk_gather_f16_avx: .byte 197,249,105,210 // vpunpckhwd %xmm2,%xmm0,%xmm2 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - .byte 196,98,125,24,37,57,32,0,0 // vbroadcastss 0x2039(%rip),%ymm12 # 6a90 <_sk_callback_avx+0x43c> + .byte 196,98,125,24,37,59,32,0,0 // vbroadcastss 0x203b(%rip),%ymm12 # 68b8 <_sk_callback_avx+0x43c> .byte 196,193,124,84,212 // vandps %ymm12,%ymm0,%ymm2 .byte 197,252,87,194 // vxorps %ymm2,%ymm0,%ymm0 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14 - .byte 196,98,121,24,29,37,32,0,0 // vbroadcastss 0x2025(%rip),%xmm11 # 6a94 <_sk_callback_avx+0x440> + .byte 196,98,121,24,29,39,32,0,0 // vbroadcastss 0x2027(%rip),%xmm11 # 68bc <_sk_callback_avx+0x440> .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3 - .byte 196,98,121,24,45,27,32,0,0 // vbroadcastss 0x201b(%rip),%xmm13 # 6a98 <_sk_callback_avx+0x444> + .byte 196,98,121,24,45,29,32,0,0 // vbroadcastss 0x201d(%rip),%xmm13 # 68c0 <_sk_callback_avx+0x444> .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -19624,7 +19473,7 @@ _sk_gather_f16_avx: .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 .byte 197,252,86,194 // vorps %ymm2,%ymm0,%ymm0 .byte 196,227,125,25,194,1 // vextractf128 $0x1,%ymm0,%xmm2 - .byte 196,226,121,24,29,209,31,0,0 // vbroadcastss 0x1fd1(%rip),%xmm3 # 6a9c <_sk_callback_avx+0x448> + .byte 196,226,121,24,29,211,31,0,0 // vbroadcastss 0x1fd3(%rip),%xmm3 # 68c4 <_sk_callback_avx+0x448> .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0 .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 @@ -19728,12 +19577,12 @@ _sk_store_f16_avx: .byte 197,252,17,52,36 // vmovups %ymm6,(%rsp) .byte 197,252,17,108,36,224 // vmovups %ymm5,-0x20(%rsp) .byte 197,252,17,100,36,192 // vmovups %ymm4,-0x40(%rsp) - .byte 196,98,125,24,13,234,29,0,0 // vbroadcastss 0x1dea(%rip),%ymm9 # 6aa0 <_sk_callback_avx+0x44c> + .byte 196,98,125,24,13,236,29,0,0 // vbroadcastss 0x1dec(%rip),%ymm9 # 68c8 <_sk_callback_avx+0x44c> .byte 196,65,124,84,209 // vandps %ymm9,%ymm0,%ymm10 .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp) .byte 196,65,124,87,218 // vxorps %ymm10,%ymm0,%ymm11 .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 - .byte 196,98,121,24,5,207,29,0,0 // vbroadcastss 0x1dcf(%rip),%xmm8 # 6aa4 <_sk_callback_avx+0x450> + .byte 196,98,121,24,5,209,29,0,0 // vbroadcastss 0x1dd1(%rip),%xmm8 # 68cc <_sk_callback_avx+0x450> .byte 196,65,57,102,236 // vpcmpgtd %xmm12,%xmm8,%xmm13 .byte 196,65,57,102,243 // vpcmpgtd %xmm11,%xmm8,%xmm14 .byte 196,67,13,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 @@ -19743,7 +19592,7 @@ _sk_store_f16_avx: .byte 196,67,13,24,242,1 // vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11 .byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12 - .byte 196,98,125,24,21,150,29,0,0 // vbroadcastss 0x1d96(%rip),%ymm10 # 6aa8 <_sk_callback_avx+0x454> + .byte 196,98,125,24,21,152,29,0,0 // vbroadcastss 0x1d98(%rip),%ymm10 # 68d0 <_sk_callback_avx+0x454> .byte 196,65,12,86,242 // vorps %ymm10,%ymm14,%ymm14 .byte 196,67,125,25,247,1 // vextractf128 $0x1,%ymm14,%xmm15 .byte 196,65,1,254,228 // vpaddd %xmm12,%xmm15,%xmm12 @@ -19825,7 +19674,7 @@ _sk_store_f16_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,139,0 // mov (%rax),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 117,66 // jne 4ef0 <_sk_store_f16_avx+0x25e> + .byte 117,66 // jne 4d16 <_sk_store_f16_avx+0x25e> .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8) .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8) .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8) @@ -19841,22 +19690,22 @@ _sk_store_f16_avx: .byte 255,224 // jmpq *%rax .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,202 // je 4ec5 <_sk_store_f16_avx+0x233> + .byte 116,202 // je 4ceb <_sk_store_f16_avx+0x233> .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,190 // jb 4ec5 <_sk_store_f16_avx+0x233> + .byte 114,190 // jb 4ceb <_sk_store_f16_avx+0x233> .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) - .byte 116,182 // je 4ec5 <_sk_store_f16_avx+0x233> + .byte 116,182 // je 4ceb <_sk_store_f16_avx+0x233> .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,170 // jb 4ec5 <_sk_store_f16_avx+0x233> + .byte 114,170 // jb 4ceb <_sk_store_f16_avx+0x233> .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) - .byte 116,162 // je 4ec5 <_sk_store_f16_avx+0x233> + .byte 116,162 // je 4ceb <_sk_store_f16_avx+0x233> .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,150 // jb 4ec5 <_sk_store_f16_avx+0x233> + .byte 114,150 // jb 4ceb <_sk_store_f16_avx+0x233> .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) - .byte 235,142 // jmp 4ec5 <_sk_store_f16_avx+0x233> + .byte 235,142 // jmp 4ceb <_sk_store_f16_avx+0x233> HIDDEN _sk_load_u16_be_avx .globl _sk_load_u16_be_avx @@ -19866,7 +19715,7 @@ _sk_load_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,253,0,0,0 // jne 504a <_sk_load_u16_be_avx+0x113> + .byte 15,133,253,0,0,0 // jne 4e70 <_sk_load_u16_be_avx+0x113> .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -19888,7 +19737,7 @@ _sk_load_u16_be_avx: .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,29,238,26,0,0 // vbroadcastss 0x1aee(%rip),%ymm11 # 6aac <_sk_callback_avx+0x458> + .byte 196,98,125,24,29,240,26,0,0 // vbroadcastss 0x1af0(%rip),%ymm11 # 68d4 <_sk_callback_avx+0x458> .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,177,109,202 // vpunpckhqdq %xmm2,%xmm9,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -19922,29 +19771,29 @@ _sk_load_u16_be_avx: .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,85 // je 50b0 <_sk_load_u16_be_avx+0x179> + .byte 116,85 // je 4ed6 <_sk_load_u16_be_avx+0x179> .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,72 // jb 50b0 <_sk_load_u16_be_avx+0x179> + .byte 114,72 // jb 4ed6 <_sk_load_u16_be_avx+0x179> .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 116,72 // je 50bd <_sk_load_u16_be_avx+0x186> + .byte 116,72 // je 4ee3 <_sk_load_u16_be_avx+0x186> .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,59 // jb 50bd <_sk_load_u16_be_avx+0x186> + .byte 114,59 // jb 4ee3 <_sk_load_u16_be_avx+0x186> .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 15,132,213,254,255,255 // je 4f68 <_sk_load_u16_be_avx+0x31> + .byte 15,132,213,254,255,255 // je 4d8e <_sk_load_u16_be_avx+0x31> .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 15,130,196,254,255,255 // jb 4f68 <_sk_load_u16_be_avx+0x31> + .byte 15,130,196,254,255,255 // jb 4d8e <_sk_load_u16_be_avx+0x31> .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 - .byte 233,184,254,255,255 // jmpq 4f68 <_sk_load_u16_be_avx+0x31> + .byte 233,184,254,255,255 // jmpq 4d8e <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 - .byte 233,171,254,255,255 // jmpq 4f68 <_sk_load_u16_be_avx+0x31> + .byte 233,171,254,255,255 // jmpq 4d8e <_sk_load_u16_be_avx+0x31> .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 - .byte 233,162,254,255,255 // jmpq 4f68 <_sk_load_u16_be_avx+0x31> + .byte 233,162,254,255,255 // jmpq 4d8e <_sk_load_u16_be_avx+0x31> HIDDEN _sk_load_rgb_u16_be_avx .globl _sk_load_rgb_u16_be_avx @@ -19954,7 +19803,7 @@ _sk_load_rgb_u16_be_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax .byte 72,133,201 // test %rcx,%rcx - .byte 15,133,243,0,0,0 // jne 51cb <_sk_load_rgb_u16_be_avx+0x105> + .byte 15,133,243,0,0,0 // jne 4ff1 <_sk_load_rgb_u16_be_avx+0x105> .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -19981,7 +19830,7 @@ _sk_load_rgb_u16_be_avx: .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 - .byte 196,98,125,24,29,78,25,0,0 // vbroadcastss 0x194e(%rip),%ymm11 # 6ab0 <_sk_callback_avx+0x45c> + .byte 196,98,125,24,29,80,25,0,0 // vbroadcastss 0x1950(%rip),%ymm11 # 68d8 <_sk_callback_avx+0x45c> .byte 196,193,124,89,195 // vmulps %ymm11,%ymm0,%ymm0 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 @@ -20002,41 +19851,41 @@ _sk_load_rgb_u16_be_avx: .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 .byte 196,193,108,89,211 // vmulps %ymm11,%ymm2,%ymm2 .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,29,235,24,0,0 // vbroadcastss 0x18eb(%rip),%ymm3 # 6ab4 <_sk_callback_avx+0x460> + .byte 196,226,125,24,29,237,24,0,0 // vbroadcastss 0x18ed(%rip),%ymm3 # 68dc <_sk_callback_avx+0x460> .byte 255,224 // jmpq *%rax .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 117,5 // jne 51e4 <_sk_load_rgb_u16_be_avx+0x11e> - .byte 233,40,255,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,5 // jne 500a <_sk_load_rgb_u16_be_avx+0x11e> + .byte 233,40,255,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,26 // jb 5213 <_sk_load_rgb_u16_be_avx+0x14d> + .byte 114,26 // jb 5039 <_sk_load_rgb_u16_be_avx+0x14d> .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 117,10 // jne 5218 <_sk_load_rgb_u16_be_avx+0x152> - .byte 233,249,254,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,244,254,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 503e <_sk_load_rgb_u16_be_avx+0x152> + .byte 233,249,254,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,244,254,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,26 // jb 5247 <_sk_load_rgb_u16_be_avx+0x181> + .byte 114,26 // jb 506d <_sk_load_rgb_u16_be_avx+0x181> .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 117,10 // jne 524c <_sk_load_rgb_u16_be_avx+0x186> - .byte 233,197,254,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,192,254,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> + .byte 117,10 // jne 5072 <_sk_load_rgb_u16_be_avx+0x186> + .byte 233,197,254,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,192,254,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,20 // jb 5275 <_sk_load_rgb_u16_be_avx+0x1af> + .byte 114,20 // jb 509b <_sk_load_rgb_u16_be_avx+0x1af> .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - .byte 233,151,254,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> - .byte 233,146,254,255,255 // jmpq 510c <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,151,254,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> + .byte 233,146,254,255,255 // jmpq 4f32 <_sk_load_rgb_u16_be_avx+0x46> HIDDEN _sk_store_u16_be_avx .globl _sk_store_u16_be_avx @@ -20045,7 +19894,7 @@ _sk_store_u16_be_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,0 // mov (%rax),%r8 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax - .byte 196,98,125,24,5,40,24,0,0 // vbroadcastss 0x1828(%rip),%ymm8 # 6ab8 <_sk_callback_avx+0x464> + .byte 196,98,125,24,5,42,24,0,0 // vbroadcastss 0x182a(%rip),%ymm8 # 68e0 <_sk_callback_avx+0x464> .byte 196,65,124,89,200 // vmulps %ymm8,%ymm0,%ymm9 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 @@ -20083,7 +19932,7 @@ _sk_store_u16_be_avx: .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 .byte 72,133,201 // test %rcx,%rcx - .byte 117,31 // jne 5374 <_sk_store_u16_be_avx+0xfa> + .byte 117,31 // jne 519a <_sk_store_u16_be_avx+0xfa> .byte 196,65,120,17,28,64 // vmovups %xmm11,(%r8,%rax,2) .byte 196,65,120,17,84,64,16 // vmovups %xmm10,0x10(%r8,%rax,2) .byte 196,65,120,17,76,64,32 // vmovups %xmm9,0x20(%r8,%rax,2) @@ -20092,22 +19941,22 @@ _sk_store_u16_be_avx: .byte 255,224 // jmpq *%rax .byte 196,65,121,214,28,64 // vmovq %xmm11,(%r8,%rax,2) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 5370 <_sk_store_u16_be_avx+0xf6> + .byte 116,240 // je 5196 <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,23,92,64,8 // vmovhpd %xmm11,0x8(%r8,%rax,2) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 5370 <_sk_store_u16_be_avx+0xf6> + .byte 114,227 // jb 5196 <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,214,84,64,16 // vmovq %xmm10,0x10(%r8,%rax,2) - .byte 116,218 // je 5370 <_sk_store_u16_be_avx+0xf6> + .byte 116,218 // je 5196 <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,23,84,64,24 // vmovhpd %xmm10,0x18(%r8,%rax,2) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 5370 <_sk_store_u16_be_avx+0xf6> + .byte 114,205 // jb 5196 <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,214,76,64,32 // vmovq %xmm9,0x20(%r8,%rax,2) - .byte 116,196 // je 5370 <_sk_store_u16_be_avx+0xf6> + .byte 116,196 // je 5196 <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,23,76,64,40 // vmovhpd %xmm9,0x28(%r8,%rax,2) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,183 // jb 5370 <_sk_store_u16_be_avx+0xf6> + .byte 114,183 // jb 5196 <_sk_store_u16_be_avx+0xf6> .byte 196,65,121,214,68,64,48 // vmovq %xmm8,0x30(%r8,%rax,2) - .byte 235,174 // jmp 5370 <_sk_store_u16_be_avx+0xf6> + .byte 235,174 // jmp 5196 <_sk_store_u16_be_avx+0xf6> HIDDEN _sk_load_f32_avx .globl _sk_load_f32_avx @@ -20115,10 +19964,10 @@ FUNCTION(_sk_load_f32_avx) _sk_load_f32_avx: .byte 72,173 // lods %ds:(%rsi),%rax .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 119,110 // ja 5438 <_sk_load_f32_avx+0x76> + .byte 119,110 // ja 525e <_sk_load_f32_avx+0x76> .byte 76,139,0 // mov (%rax),%r8 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 - .byte 76,141,21,132,0,0,0 // lea 0x84(%rip),%r10 # 5460 <_sk_load_f32_avx+0x9e> + .byte 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 5288 <_sk_load_f32_avx+0xa0> .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -20144,19 +19993,19 @@ _sk_load_f32_avx: .byte 196,193,101,21,216 // vunpckhpd %ymm8,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax - .byte 133,255 // test %edi,%edi - .byte 255 // (bad) - .byte 255,204 // dec %esp + .byte 102,144 // xchg %ax,%ax + .byte 131,255,255 // cmp $0xffffffff,%edi + .byte 255,202 // dec %edx .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 191,255,255,255,178 // mov $0xb2ffffff,%edi + .byte 189,255,255,255,176 // mov $0xb0ffffff,%ebp .byte 255 // (bad) .byte 255 // (bad) - .byte 255,165,255,255,255,157 // jmpq *-0x62000001(%rbp) + .byte 255,163,255,255,255,155 // jmpq *-0x64000001(%rbx) .byte 255 // (bad) .byte 255 // (bad) - .byte 255,149,255,255,255,141 // callq *-0x72000001(%rbp) + .byte 255,147,255,255,255,139 // callq *-0x74000001(%rbx) .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -20177,7 +20026,7 @@ _sk_store_f32_avx: .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 .byte 72,133,201 // test %rcx,%rcx - .byte 117,55 // jne 54ed <_sk_store_f32_avx+0x6d> + .byte 117,55 // jne 5315 <_sk_store_f32_avx+0x6d> .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -20190,22 +20039,22 @@ _sk_store_f32_avx: .byte 255,224 // jmpq *%rax .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) .byte 72,131,249,1 // cmp $0x1,%rcx - .byte 116,240 // je 54e9 <_sk_store_f32_avx+0x69> + .byte 116,240 // je 5311 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) .byte 72,131,249,3 // cmp $0x3,%rcx - .byte 114,227 // jb 54e9 <_sk_store_f32_avx+0x69> + .byte 114,227 // jb 5311 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) - .byte 116,218 // je 54e9 <_sk_store_f32_avx+0x69> + .byte 116,218 // je 5311 <_sk_store_f32_avx+0x69> .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) .byte 72,131,249,5 // cmp $0x5,%rcx - .byte 114,205 // jb 54e9 <_sk_store_f32_avx+0x69> + .byte 114,205 // jb 5311 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - .byte 116,195 // je 54e9 <_sk_store_f32_avx+0x69> + .byte 116,195 // je 5311 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) .byte 72,131,249,7 // cmp $0x7,%rcx - .byte 114,181 // jb 54e9 <_sk_store_f32_avx+0x69> + .byte 114,181 // jb 5311 <_sk_store_f32_avx+0x69> .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - .byte 235,171 // jmp 54e9 <_sk_store_f32_avx+0x69> + .byte 235,171 // jmp 5311 <_sk_store_f32_avx+0x69> HIDDEN _sk_clamp_x_avx .globl _sk_clamp_x_avx @@ -20309,7 +20158,7 @@ FUNCTION(_sk_clamp_x_1_avx) _sk_clamp_x_1_avx: .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 .byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0 - .byte 196,98,125,24,5,90,20,0,0 // vbroadcastss 0x145a(%rip),%ymm8 # 6abc <_sk_callback_avx+0x468> + .byte 196,98,125,24,5,90,20,0,0 // vbroadcastss 0x145a(%rip),%ymm8 # 68e4 <_sk_callback_avx+0x468> .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -20327,9 +20176,9 @@ HIDDEN _sk_mirror_x_1_avx .globl _sk_mirror_x_1_avx FUNCTION(_sk_mirror_x_1_avx) _sk_mirror_x_1_avx: - .byte 196,98,125,24,5,61,20,0,0 // vbroadcastss 0x143d(%rip),%ymm8 # 6ac0 <_sk_callback_avx+0x46c> + .byte 196,98,125,24,5,61,20,0,0 // vbroadcastss 0x143d(%rip),%ymm8 # 68e8 <_sk_callback_avx+0x46c> .byte 196,193,124,88,192 // vaddps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,13,51,20,0,0 // vbroadcastss 0x1433(%rip),%ymm9 # 6ac4 <_sk_callback_avx+0x470> + .byte 196,98,125,24,13,51,20,0,0 // vbroadcastss 0x1433(%rip),%ymm9 # 68ec <_sk_callback_avx+0x470> .byte 196,65,124,89,201 // vmulps %ymm9,%ymm0,%ymm9 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 .byte 196,65,52,88,201 // vaddps %ymm9,%ymm9,%ymm9 @@ -20345,12 +20194,12 @@ HIDDEN _sk_luminance_to_alpha_avx .globl _sk_luminance_to_alpha_avx FUNCTION(_sk_luminance_to_alpha_avx) _sk_luminance_to_alpha_avx: - .byte 196,226,125,24,29,3,20,0,0 // vbroadcastss 0x1403(%rip),%ymm3 # 6ac8 <_sk_callback_avx+0x474> + .byte 196,226,125,24,29,3,20,0,0 // vbroadcastss 0x1403(%rip),%ymm3 # 68f0 <_sk_callback_avx+0x474> .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,29,250,19,0,0 // vbroadcastss 0x13fa(%rip),%ymm3 # 6acc <_sk_callback_avx+0x478> + .byte 196,226,125,24,29,250,19,0,0 // vbroadcastss 0x13fa(%rip),%ymm3 # 68f4 <_sk_callback_avx+0x478> .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 - .byte 196,226,125,24,13,237,19,0,0 // vbroadcastss 0x13ed(%rip),%ymm1 # 6ad0 <_sk_callback_avx+0x47c> + .byte 196,226,125,24,13,237,19,0,0 // vbroadcastss 0x13ed(%rip),%ymm1 # 68f8 <_sk_callback_avx+0x47c> .byte 197,236,89,201 // vmulps %ymm1,%ymm2,%ymm1 .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -20569,9 +20418,9 @@ _sk_evenly_spaced_gradient_avx: .byte 72,139,24 // mov (%rax),%rbx .byte 72,139,104,8 // mov 0x8(%rax),%rbp .byte 72,255,203 // dec %rbx - .byte 120,7 // js 5a39 <_sk_evenly_spaced_gradient_avx+0x1f> + .byte 120,7 // js 5861 <_sk_evenly_spaced_gradient_avx+0x1f> .byte 196,225,242,42,203 // vcvtsi2ss %rbx,%xmm1,%xmm1 - .byte 235,21 // jmp 5a4e <_sk_evenly_spaced_gradient_avx+0x34> + .byte 235,21 // jmp 5876 <_sk_evenly_spaced_gradient_avx+0x34> .byte 73,137,216 // mov %rbx,%r8 .byte 73,209,232 // shr %r8 .byte 131,227,1 // and $0x1,%ebx @@ -20728,18 +20577,18 @@ HIDDEN _sk_gauss_a_to_rgba_avx .globl _sk_gauss_a_to_rgba_avx FUNCTION(_sk_gauss_a_to_rgba_avx) _sk_gauss_a_to_rgba_avx: - .byte 196,226,125,24,5,88,13,0,0 // vbroadcastss 0xd58(%rip),%ymm0 # 6ad4 <_sk_callback_avx+0x480> + .byte 196,226,125,24,5,88,13,0,0 // vbroadcastss 0xd58(%rip),%ymm0 # 68fc <_sk_callback_avx+0x480> .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 - .byte 196,226,125,24,13,79,13,0,0 // vbroadcastss 0xd4f(%rip),%ymm1 # 6ad8 <_sk_callback_avx+0x484> + .byte 196,226,125,24,13,79,13,0,0 // vbroadcastss 0xd4f(%rip),%ymm1 # 6900 <_sk_callback_avx+0x484> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,13,66,13,0,0 // vbroadcastss 0xd42(%rip),%ymm1 # 6adc <_sk_callback_avx+0x488> + .byte 196,226,125,24,13,66,13,0,0 // vbroadcastss 0xd42(%rip),%ymm1 # 6904 <_sk_callback_avx+0x488> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,13,53,13,0,0 // vbroadcastss 0xd35(%rip),%ymm1 # 6ae0 <_sk_callback_avx+0x48c> + .byte 196,226,125,24,13,53,13,0,0 // vbroadcastss 0xd35(%rip),%ymm1 # 6908 <_sk_callback_avx+0x48c> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 - .byte 196,226,125,24,13,40,13,0,0 // vbroadcastss 0xd28(%rip),%ymm1 # 6ae4 <_sk_callback_avx+0x490> + .byte 196,226,125,24,13,40,13,0,0 // vbroadcastss 0xd28(%rip),%ymm1 # 690c <_sk_callback_avx+0x490> .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 @@ -20761,12 +20610,12 @@ _sk_gradient_avx: .byte 76,139,0 // mov (%rax),%r8 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 73,131,248,2 // cmp $0x2,%r8 - .byte 114,80 // jb 5e39 <_sk_gradient_avx+0x69> + .byte 114,80 // jb 5c61 <_sk_gradient_avx+0x69> .byte 72,139,88,72 // mov 0x48(%rax),%rbx .byte 73,255,200 // dec %r8 .byte 72,131,195,4 // add $0x4,%rbx .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 - .byte 196,98,125,24,21,230,12,0,0 // vbroadcastss 0xce6(%rip),%ymm10 # 6ae8 <_sk_callback_avx+0x494> + .byte 196,98,125,24,21,230,12,0,0 // vbroadcastss 0xce6(%rip),%ymm10 # 6910 <_sk_callback_avx+0x494> .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 .byte 196,98,125,24,3 // vbroadcastss (%rbx),%ymm8 .byte 197,60,194,192,2 // vcmpleps %ymm0,%ymm8,%ymm8 @@ -20778,7 +20627,7 @@ _sk_gradient_avx: .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 .byte 72,131,195,4 // add $0x4,%rbx .byte 73,255,200 // dec %r8 - .byte 117,205 // jne 5e06 <_sk_gradient_avx+0x36> + .byte 117,205 // jne 5c2e <_sk_gradient_avx+0x36> .byte 196,195,249,22,200,1 // vpextrq $0x1,%xmm1,%r8 .byte 69,137,193 // mov %r8d,%r9d .byte 73,193,232,32 // shr $0x20,%r8 @@ -20960,27 +20809,27 @@ _sk_xy_to_unit_angle_avx: .byte 196,65,52,95,226 // vmaxps %ymm10,%ymm9,%ymm12 .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 .byte 196,65,36,89,227 // vmulps %ymm11,%ymm11,%ymm12 - .byte 196,98,125,24,45,10,9,0,0 // vbroadcastss 0x90a(%rip),%ymm13 # 6aec <_sk_callback_avx+0x498> + .byte 196,98,125,24,45,10,9,0,0 // vbroadcastss 0x90a(%rip),%ymm13 # 6914 <_sk_callback_avx+0x498> .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13 - .byte 196,98,125,24,53,0,9,0,0 // vbroadcastss 0x900(%rip),%ymm14 # 6af0 <_sk_callback_avx+0x49c> + .byte 196,98,125,24,53,0,9,0,0 // vbroadcastss 0x900(%rip),%ymm14 # 6918 <_sk_callback_avx+0x49c> .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13 - .byte 196,98,125,24,53,241,8,0,0 // vbroadcastss 0x8f1(%rip),%ymm14 # 6af4 <_sk_callback_avx+0x4a0> + .byte 196,98,125,24,53,241,8,0,0 // vbroadcastss 0x8f1(%rip),%ymm14 # 691c <_sk_callback_avx+0x4a0> .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 .byte 196,65,28,89,229 // vmulps %ymm13,%ymm12,%ymm12 - .byte 196,98,125,24,45,226,8,0,0 // vbroadcastss 0x8e2(%rip),%ymm13 # 6af8 <_sk_callback_avx+0x4a4> + .byte 196,98,125,24,45,226,8,0,0 // vbroadcastss 0x8e2(%rip),%ymm13 # 6920 <_sk_callback_avx+0x4a4> .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11 .byte 196,65,52,194,202,1 // vcmpltps %ymm10,%ymm9,%ymm9 - .byte 196,98,125,24,21,205,8,0,0 // vbroadcastss 0x8cd(%rip),%ymm10 # 6afc <_sk_callback_avx+0x4a8> + .byte 196,98,125,24,21,205,8,0,0 // vbroadcastss 0x8cd(%rip),%ymm10 # 6924 <_sk_callback_avx+0x4a8> .byte 196,65,44,92,211 // vsubps %ymm11,%ymm10,%ymm10 .byte 196,67,37,74,202,144 // vblendvps %ymm9,%ymm10,%ymm11,%ymm9 .byte 196,193,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm0 - .byte 196,98,125,24,21,183,8,0,0 // vbroadcastss 0x8b7(%rip),%ymm10 # 6b00 <_sk_callback_avx+0x4ac> + .byte 196,98,125,24,21,183,8,0,0 // vbroadcastss 0x8b7(%rip),%ymm10 # 6928 <_sk_callback_avx+0x4ac> .byte 196,65,44,92,209 // vsubps %ymm9,%ymm10,%ymm10 .byte 196,195,53,74,194,0 // vblendvps %ymm0,%ymm10,%ymm9,%ymm0 .byte 196,65,116,194,200,1 // vcmpltps %ymm8,%ymm1,%ymm9 - .byte 196,98,125,24,21,161,8,0,0 // vbroadcastss 0x8a1(%rip),%ymm10 # 6b04 <_sk_callback_avx+0x4b0> + .byte 196,98,125,24,21,161,8,0,0 // vbroadcastss 0x8a1(%rip),%ymm10 # 692c <_sk_callback_avx+0x4b0> .byte 197,44,92,208 // vsubps %ymm0,%ymm10,%ymm10 .byte 196,195,125,74,194,144 // vblendvps %ymm9,%ymm10,%ymm0,%ymm0 .byte 196,65,124,194,200,3 // vcmpunordps %ymm8,%ymm0,%ymm9 @@ -21004,7 +20853,7 @@ HIDDEN _sk_save_xy_avx FUNCTION(_sk_save_xy_avx) _sk_save_xy_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,107,8,0,0 // vbroadcastss 0x86b(%rip),%ymm8 # 6b08 <_sk_callback_avx+0x4b4> + .byte 196,98,125,24,5,107,8,0,0 // vbroadcastss 0x86b(%rip),%ymm8 # 6930 <_sk_callback_avx+0x4b4> .byte 196,65,124,88,200 // vaddps %ymm8,%ymm0,%ymm9 .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 @@ -21041,9 +20890,9 @@ HIDDEN _sk_bilinear_nx_avx FUNCTION(_sk_bilinear_nx_avx) _sk_bilinear_nx_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,247,7,0,0 // vbroadcastss 0x7f7(%rip),%ymm0 # 6b0c <_sk_callback_avx+0x4b8> + .byte 196,226,125,24,5,247,7,0,0 // vbroadcastss 0x7f7(%rip),%ymm0 # 6934 <_sk_callback_avx+0x4b8> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,238,7,0,0 // vbroadcastss 0x7ee(%rip),%ymm8 # 6b10 <_sk_callback_avx+0x4bc> + .byte 196,98,125,24,5,238,7,0,0 // vbroadcastss 0x7ee(%rip),%ymm8 # 6938 <_sk_callback_avx+0x4bc> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21054,7 +20903,7 @@ HIDDEN _sk_bilinear_px_avx FUNCTION(_sk_bilinear_px_avx) _sk_bilinear_px_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,214,7,0,0 // vbroadcastss 0x7d6(%rip),%ymm0 # 6b14 <_sk_callback_avx+0x4c0> + .byte 196,226,125,24,5,214,7,0,0 // vbroadcastss 0x7d6(%rip),%ymm0 # 693c <_sk_callback_avx+0x4c0> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -21066,9 +20915,9 @@ HIDDEN _sk_bilinear_ny_avx FUNCTION(_sk_bilinear_ny_avx) _sk_bilinear_ny_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,186,7,0,0 // vbroadcastss 0x7ba(%rip),%ymm1 # 6b18 <_sk_callback_avx+0x4c4> + .byte 196,226,125,24,13,186,7,0,0 // vbroadcastss 0x7ba(%rip),%ymm1 # 6940 <_sk_callback_avx+0x4c4> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,176,7,0,0 // vbroadcastss 0x7b0(%rip),%ymm8 # 6b1c <_sk_callback_avx+0x4c8> + .byte 196,98,125,24,5,176,7,0,0 // vbroadcastss 0x7b0(%rip),%ymm8 # 6944 <_sk_callback_avx+0x4c8> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21079,7 +20928,7 @@ HIDDEN _sk_bilinear_py_avx FUNCTION(_sk_bilinear_py_avx) _sk_bilinear_py_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,152,7,0,0 // vbroadcastss 0x798(%rip),%ymm1 # 6b20 <_sk_callback_avx+0x4cc> + .byte 196,226,125,24,13,152,7,0,0 // vbroadcastss 0x798(%rip),%ymm1 # 6948 <_sk_callback_avx+0x4cc> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -21091,14 +20940,14 @@ HIDDEN _sk_bicubic_n3x_avx FUNCTION(_sk_bicubic_n3x_avx) _sk_bicubic_n3x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,123,7,0,0 // vbroadcastss 0x77b(%rip),%ymm0 # 6b24 <_sk_callback_avx+0x4d0> + .byte 196,226,125,24,5,123,7,0,0 // vbroadcastss 0x77b(%rip),%ymm0 # 694c <_sk_callback_avx+0x4d0> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,114,7,0,0 // vbroadcastss 0x772(%rip),%ymm8 # 6b28 <_sk_callback_avx+0x4d4> + .byte 196,98,125,24,5,114,7,0,0 // vbroadcastss 0x772(%rip),%ymm8 # 6950 <_sk_callback_avx+0x4d4> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,99,7,0,0 // vbroadcastss 0x763(%rip),%ymm10 # 6b2c <_sk_callback_avx+0x4d8> + .byte 196,98,125,24,21,99,7,0,0 // vbroadcastss 0x763(%rip),%ymm10 # 6954 <_sk_callback_avx+0x4d8> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,89,7,0,0 // vbroadcastss 0x759(%rip),%ymm10 # 6b30 <_sk_callback_avx+0x4dc> + .byte 196,98,125,24,21,89,7,0,0 // vbroadcastss 0x759(%rip),%ymm10 # 6958 <_sk_callback_avx+0x4dc> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -21110,19 +20959,19 @@ HIDDEN _sk_bicubic_n1x_avx FUNCTION(_sk_bicubic_n1x_avx) _sk_bicubic_n1x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,60,7,0,0 // vbroadcastss 0x73c(%rip),%ymm0 # 6b34 <_sk_callback_avx+0x4e0> + .byte 196,226,125,24,5,60,7,0,0 // vbroadcastss 0x73c(%rip),%ymm0 # 695c <_sk_callback_avx+0x4e0> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 - .byte 196,98,125,24,5,51,7,0,0 // vbroadcastss 0x733(%rip),%ymm8 # 6b38 <_sk_callback_avx+0x4e4> + .byte 196,98,125,24,5,51,7,0,0 // vbroadcastss 0x733(%rip),%ymm8 # 6960 <_sk_callback_avx+0x4e4> .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,41,7,0,0 // vbroadcastss 0x729(%rip),%ymm9 # 6b3c <_sk_callback_avx+0x4e8> + .byte 196,98,125,24,13,41,7,0,0 // vbroadcastss 0x729(%rip),%ymm9 # 6964 <_sk_callback_avx+0x4e8> .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,31,7,0,0 // vbroadcastss 0x71f(%rip),%ymm10 # 6b40 <_sk_callback_avx+0x4ec> + .byte 196,98,125,24,21,31,7,0,0 // vbroadcastss 0x71f(%rip),%ymm10 # 6968 <_sk_callback_avx+0x4ec> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,16,7,0,0 // vbroadcastss 0x710(%rip),%ymm10 # 6b44 <_sk_callback_avx+0x4f0> + .byte 196,98,125,24,21,16,7,0,0 // vbroadcastss 0x710(%rip),%ymm10 # 696c <_sk_callback_avx+0x4f0> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,1,7,0,0 // vbroadcastss 0x701(%rip),%ymm9 # 6b48 <_sk_callback_avx+0x4f4> + .byte 196,98,125,24,13,1,7,0,0 // vbroadcastss 0x701(%rip),%ymm9 # 6970 <_sk_callback_avx+0x4f4> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21133,17 +20982,17 @@ HIDDEN _sk_bicubic_p1x_avx FUNCTION(_sk_bicubic_p1x_avx) _sk_bicubic_p1x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,233,6,0,0 // vbroadcastss 0x6e9(%rip),%ymm8 # 6b4c <_sk_callback_avx+0x4f8> + .byte 196,98,125,24,5,233,6,0,0 // vbroadcastss 0x6e9(%rip),%ymm8 # 6974 <_sk_callback_avx+0x4f8> .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 - .byte 196,98,125,24,21,219,6,0,0 // vbroadcastss 0x6db(%rip),%ymm10 # 6b50 <_sk_callback_avx+0x4fc> + .byte 196,98,125,24,21,219,6,0,0 // vbroadcastss 0x6db(%rip),%ymm10 # 6978 <_sk_callback_avx+0x4fc> .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 - .byte 196,98,125,24,29,209,6,0,0 // vbroadcastss 0x6d1(%rip),%ymm11 # 6b54 <_sk_callback_avx+0x500> + .byte 196,98,125,24,29,209,6,0,0 // vbroadcastss 0x6d1(%rip),%ymm11 # 697c <_sk_callback_avx+0x500> .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 - .byte 196,98,125,24,13,184,6,0,0 // vbroadcastss 0x6b8(%rip),%ymm9 # 6b58 <_sk_callback_avx+0x504> + .byte 196,98,125,24,13,184,6,0,0 // vbroadcastss 0x6b8(%rip),%ymm9 # 6980 <_sk_callback_avx+0x504> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21154,13 +21003,13 @@ HIDDEN _sk_bicubic_p3x_avx FUNCTION(_sk_bicubic_p3x_avx) _sk_bicubic_p3x_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,160,6,0,0 // vbroadcastss 0x6a0(%rip),%ymm0 # 6b5c <_sk_callback_avx+0x508> + .byte 196,226,125,24,5,160,6,0,0 // vbroadcastss 0x6a0(%rip),%ymm0 # 6984 <_sk_callback_avx+0x508> .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,141,6,0,0 // vbroadcastss 0x68d(%rip),%ymm10 # 6b60 <_sk_callback_avx+0x50c> + .byte 196,98,125,24,21,141,6,0,0 // vbroadcastss 0x68d(%rip),%ymm10 # 6988 <_sk_callback_avx+0x50c> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,131,6,0,0 // vbroadcastss 0x683(%rip),%ymm10 # 6b64 <_sk_callback_avx+0x510> + .byte 196,98,125,24,21,131,6,0,0 // vbroadcastss 0x683(%rip),%ymm10 # 698c <_sk_callback_avx+0x510> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) @@ -21172,14 +21021,14 @@ HIDDEN _sk_bicubic_n3y_avx FUNCTION(_sk_bicubic_n3y_avx) _sk_bicubic_n3y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,102,6,0,0 // vbroadcastss 0x666(%rip),%ymm1 # 6b68 <_sk_callback_avx+0x514> + .byte 196,226,125,24,13,102,6,0,0 // vbroadcastss 0x666(%rip),%ymm1 # 6990 <_sk_callback_avx+0x514> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,92,6,0,0 // vbroadcastss 0x65c(%rip),%ymm8 # 6b6c <_sk_callback_avx+0x518> + .byte 196,98,125,24,5,92,6,0,0 // vbroadcastss 0x65c(%rip),%ymm8 # 6994 <_sk_callback_avx+0x518> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,77,6,0,0 // vbroadcastss 0x64d(%rip),%ymm10 # 6b70 <_sk_callback_avx+0x51c> + .byte 196,98,125,24,21,77,6,0,0 // vbroadcastss 0x64d(%rip),%ymm10 # 6998 <_sk_callback_avx+0x51c> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,67,6,0,0 // vbroadcastss 0x643(%rip),%ymm10 # 6b74 <_sk_callback_avx+0x520> + .byte 196,98,125,24,21,67,6,0,0 // vbroadcastss 0x643(%rip),%ymm10 # 699c <_sk_callback_avx+0x520> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -21191,19 +21040,19 @@ HIDDEN _sk_bicubic_n1y_avx FUNCTION(_sk_bicubic_n1y_avx) _sk_bicubic_n1y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,38,6,0,0 // vbroadcastss 0x626(%rip),%ymm1 # 6b78 <_sk_callback_avx+0x524> + .byte 196,226,125,24,13,38,6,0,0 // vbroadcastss 0x626(%rip),%ymm1 # 69a0 <_sk_callback_avx+0x524> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 - .byte 196,98,125,24,5,28,6,0,0 // vbroadcastss 0x61c(%rip),%ymm8 # 6b7c <_sk_callback_avx+0x528> + .byte 196,98,125,24,5,28,6,0,0 // vbroadcastss 0x61c(%rip),%ymm8 # 69a4 <_sk_callback_avx+0x528> .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 - .byte 196,98,125,24,13,18,6,0,0 // vbroadcastss 0x612(%rip),%ymm9 # 6b80 <_sk_callback_avx+0x52c> + .byte 196,98,125,24,13,18,6,0,0 // vbroadcastss 0x612(%rip),%ymm9 # 69a8 <_sk_callback_avx+0x52c> .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,8,6,0,0 // vbroadcastss 0x608(%rip),%ymm10 # 6b84 <_sk_callback_avx+0x530> + .byte 196,98,125,24,21,8,6,0,0 // vbroadcastss 0x608(%rip),%ymm10 # 69ac <_sk_callback_avx+0x530> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 - .byte 196,98,125,24,21,249,5,0,0 // vbroadcastss 0x5f9(%rip),%ymm10 # 6b88 <_sk_callback_avx+0x534> + .byte 196,98,125,24,21,249,5,0,0 // vbroadcastss 0x5f9(%rip),%ymm10 # 69b0 <_sk_callback_avx+0x534> .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 - .byte 196,98,125,24,13,234,5,0,0 // vbroadcastss 0x5ea(%rip),%ymm9 # 6b8c <_sk_callback_avx+0x538> + .byte 196,98,125,24,13,234,5,0,0 // vbroadcastss 0x5ea(%rip),%ymm9 # 69b4 <_sk_callback_avx+0x538> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21214,17 +21063,17 @@ HIDDEN _sk_bicubic_p1y_avx FUNCTION(_sk_bicubic_p1y_avx) _sk_bicubic_p1y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,98,125,24,5,210,5,0,0 // vbroadcastss 0x5d2(%rip),%ymm8 # 6b90 <_sk_callback_avx+0x53c> + .byte 196,98,125,24,5,210,5,0,0 // vbroadcastss 0x5d2(%rip),%ymm8 # 69b8 <_sk_callback_avx+0x53c> .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 - .byte 196,98,125,24,21,195,5,0,0 // vbroadcastss 0x5c3(%rip),%ymm10 # 6b94 <_sk_callback_avx+0x540> + .byte 196,98,125,24,21,195,5,0,0 // vbroadcastss 0x5c3(%rip),%ymm10 # 69bc <_sk_callback_avx+0x540> .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 - .byte 196,98,125,24,29,185,5,0,0 // vbroadcastss 0x5b9(%rip),%ymm11 # 6b98 <_sk_callback_avx+0x544> + .byte 196,98,125,24,29,185,5,0,0 // vbroadcastss 0x5b9(%rip),%ymm11 # 69c0 <_sk_callback_avx+0x544> .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 - .byte 196,98,125,24,13,160,5,0,0 // vbroadcastss 0x5a0(%rip),%ymm9 # 6b9c <_sk_callback_avx+0x548> + .byte 196,98,125,24,13,160,5,0,0 // vbroadcastss 0x5a0(%rip),%ymm9 # 69c4 <_sk_callback_avx+0x548> .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) .byte 72,173 // lods %ds:(%rsi),%rax @@ -21235,13 +21084,13 @@ HIDDEN _sk_bicubic_p3y_avx FUNCTION(_sk_bicubic_p3y_avx) _sk_bicubic_p3y_avx: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,13,136,5,0,0 // vbroadcastss 0x588(%rip),%ymm1 # 6ba0 <_sk_callback_avx+0x54c> + .byte 196,226,125,24,13,136,5,0,0 // vbroadcastss 0x588(%rip),%ymm1 # 69c8 <_sk_callback_avx+0x54c> .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 - .byte 196,98,125,24,21,116,5,0,0 // vbroadcastss 0x574(%rip),%ymm10 # 6ba4 <_sk_callback_avx+0x550> + .byte 196,98,125,24,21,116,5,0,0 // vbroadcastss 0x574(%rip),%ymm10 # 69cc <_sk_callback_avx+0x550> .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 - .byte 196,98,125,24,21,106,5,0,0 // vbroadcastss 0x56a(%rip),%ymm10 # 6ba8 <_sk_callback_avx+0x554> + .byte 196,98,125,24,21,106,5,0,0 // vbroadcastss 0x56a(%rip),%ymm10 # 69d0 <_sk_callback_avx+0x554> .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) @@ -21365,25 +21214,25 @@ BALIGN4 .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 682d <.literal4+0xb1> + .byte 71,225,61 // rex.RXB loope 6655 <.literal4+0xb1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 683d <.literal4+0xc1> + .byte 71,225,61 // rex.RXB loope 6665 <.literal4+0xc1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 684d <.literal4+0xd1> + .byte 71,225,61 // rex.RXB loope 6675 <.literal4+0xd1> .byte 0,0 // add %al,(%rax) .byte 128,63,154 // cmpb $0x9a,(%rdi) .byte 153 // cltd .byte 153 // cltd .byte 62,61,10,23,63,174 // ds cmp $0xae3f170a,%eax - .byte 71,225,61 // rex.RXB loope 685d <.literal4+0xe1> + .byte 71,225,61 // rex.RXB loope 6685 <.literal4+0xe1> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,63,0,0,127 // add %al,0x7f00003f(%rax) @@ -21435,7 +21284,7 @@ BALIGN4 .byte 190,129,128,128,59 // mov $0x3b808081,%esi .byte 129,128,128,59,0,248,0,0,8,33 // addl $0x21080000,-0x7ffc480(%rax) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 68b1 <.literal4+0x135> + .byte 224,7 // loopne 66d9 <.literal4+0x135> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21451,10 +21300,10 @@ BALIGN4 .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) .byte 0,52,255 // add %dh,(%rdi,%rdi,8) .byte 255 // (bad) - .byte 127,0 // jg 68d8 <.literal4+0x15c> + .byte 127,0 // jg 6700 <.literal4+0x15c> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 6951 <.literal4+0x1d5> + .byte 119,115 // ja 6779 <.literal4+0x1d5> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21468,10 +21317,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 690c <.literal4+0x190> + .byte 127,0 // jg 6734 <.literal4+0x190> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 6985 <.literal4+0x209> + .byte 119,115 // ja 67ad <.literal4+0x209> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21485,10 +21334,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 6940 <.literal4+0x1c4> + .byte 127,0 // jg 6768 <.literal4+0x1c4> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 69b9 <.literal4+0x23d> + .byte 119,115 // ja 67e1 <.literal4+0x23d> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21502,10 +21351,10 @@ BALIGN4 .byte 0,128,63,0,0,0 // add %al,0x3f(%rax) .byte 52,255 // xor $0xff,%al .byte 255 // (bad) - .byte 127,0 // jg 6974 <.literal4+0x1f8> + .byte 127,0 // jg 679c <.literal4+0x1f8> .byte 0,0 // add %al,(%rax) .byte 0,63 // add %bh,(%rdi) - .byte 119,115 // ja 69ed <.literal4+0x271> + .byte 119,115 // ja 6815 <.literal4+0x271> .byte 248 // clc .byte 194,117,191 // retq $0xbf75 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi @@ -21518,7 +21367,7 @@ BALIGN4 .byte 0,75,0 // add %cl,0x0(%rbx) .byte 0,128,63,0,0,200 // add %al,-0x37ffffc1(%rax) .byte 66,0,0 // rex.X add %al,(%rax) - .byte 127,67 // jg 69eb <.literal4+0x26f> + .byte 127,67 // jg 6813 <.literal4+0x26f> .byte 0,0 // add %al,(%rax) .byte 0,195 // add %al,%bl .byte 0,0 // add %al,(%rax) @@ -21530,10 +21379,10 @@ BALIGN4 .byte 190,80,128,3,62 // mov $0x3e038050,%esi .byte 31 // (bad) .byte 215 // xlat %ds:(%rbx) - .byte 118,63 // jbe 6a0b <.literal4+0x28f> + .byte 118,63 // jbe 6833 <.literal4+0x28f> .byte 246,64,83,63 // testb $0x3f,0x53(%rax) .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 6a1f <.literal4+0x2a3> + .byte 127,67 // jg 6847 <.literal4+0x2a3> .byte 129,128,128,59,0,0,128,63,129,128 // addl $0x80813f80,0x3b80(%rax) .byte 128,59,0 // cmpb $0x0,(%rbx) .byte 0,128,63,129,128,128 // add %al,-0x7f7f7ec1(%rax) @@ -21542,7 +21391,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 6a01 <.literal4+0x285> + .byte 224,7 // loopne 6829 <.literal4+0x285> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21554,7 +21403,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 8,33 // or %ah,(%rcx) .byte 132,55 // test %dh,(%rdi) - .byte 224,7 // loopne 6a1d <.literal4+0x2a1> + .byte 224,7 // loopne 6845 <.literal4+0x2a1> .byte 0,0 // add %al,(%rax) .byte 33,8 // and %ecx,(%rax) .byte 2,58 // add (%rdx),%bh @@ -21565,7 +21414,7 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 248 // clc .byte 65,0,0 // add %al,(%r8) - .byte 124,66 // jl 6a72 <.literal4+0x2f6> + .byte 124,66 // jl 689a <.literal4+0x2f6> .byte 0,240 // add %dh,%al .byte 0,0 // add %al,(%rax) .byte 137,136,136,55,0,15 // mov %ecx,0xf003788(%rax) @@ -21583,9 +21432,9 @@ BALIGN4 .byte 137,136,136,59,15,0 // mov %ecx,0xf3b88(%rax) .byte 0,0 // add %al,(%rax) .byte 137,136,136,61,0,0 // mov %ecx,0x3d88(%rax) - .byte 112,65 // jo 6ab5 <.literal4+0x339> + .byte 112,65 // jo 68dd <.literal4+0x339> .byte 129,128,128,59,129,128,128,59,0,0 // addl $0x3b80,-0x7f7ec480(%rax) - .byte 127,67 // jg 6ac3 <.literal4+0x347> + .byte 127,67 // jg 68eb <.literal4+0x347> .byte 0,128,0,0,0,0 // add %al,0x0(%rax) .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax) .byte 0,0 // add %al,(%rax) @@ -21601,7 +21450,7 @@ BALIGN4 .byte 0,128,55,0,0,128 // add %al,-0x7fffffc9(%rax) .byte 63 // (bad) .byte 0,255 // add %bh,%bh - .byte 127,71 // jg 6b03 <.literal4+0x387> + .byte 127,71 // jg 692b <.literal4+0x387> .byte 0,0 // add %al,(%rax) .byte 128,63,0 // cmpb $0x0,(%rdi) .byte 0,128,191,0,0,0 // add %al,0xbf(%rax) @@ -21697,39 +21546,89 @@ BALIGN4 .byte 170 // stos %al,%es:(%rdi) .byte 190 // .byte 0xbe -BALIGN32 - .byte 255,0 // incl (%rax) +BALIGN16 + .byte 0,2 // add %al,(%rdx) + .byte 4,6 // add $0x6,%al .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) + .byte 8,10 // or %cl,(%rdx) + .byte 12,14 // or $0xe,%al .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) + .byte 0,2 // add %al,(%rdx) + .byte 4,6 // add $0x6,%al .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 8,10 // or %cl,(%rdx) + .byte 12,14 // or $0xe,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 255,0 // incl (%rax) + .byte 0,2 // add %al,(%rdx) + .byte 4,6 // add $0x6,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 8,10 // or %cl,(%rdx) + .byte 12,14 // or $0xe,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,2 // add %al,(%rdx) + .byte 4,6 // add $0x6,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 8,10 // or %cl,(%rdx) + .byte 12,14 // or $0xe,%al + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + .byte 0,0 // add %al,(%rax) + +BALIGN32 .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) @@ -21778,24 +21677,38 @@ BALIGN32 .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) .byte 0,0 // add %al,(%rax) - -BALIGN16 .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) .byte 255,0 // incl (%rax) + .byte 0,0 // add %al,(%rax) BALIGN32 HIDDEN _sk_start_pipeline_sse41 diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 2baf791931..0543822f9e 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -1177,8 +1177,8 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 DB 15,133,180,0,0,0 ; jne 13ef <_sk_srcover_rgba_8888_hsw+0xcd> - DB 196,193,126,111,57 ; vmovdqu (%r9),%ymm7 - DB 197,197,219,37,88,59,0,0 ; vpand 0x3b58(%rip),%ymm7,%ymm4 # 4ea0 <_sk_callback_hsw+0x526> + DB 196,193,124,16,57 ; vmovups (%r9),%ymm7 + DB 197,196,84,37,88,59,0,0 ; vandps 0x3b58(%rip),%ymm7,%ymm4 # 4ea0 <_sk_callback_hsw+0x526> DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 DB 196,226,69,0,45,107,59,0,0 ; vpshufb 0x3b6b(%rip),%ymm7,%ymm5 # 4ec0 <_sk_callback_hsw+0x546> DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 @@ -1209,7 +1209,7 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 196,65,61,235,193 ; vpor %ymm9,%ymm8,%ymm8 DB 77,133,192 ; test %r8,%r8 DB 117,53 ; jne 1418 <_sk_srcover_rgba_8888_hsw+0xf6> - DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) + DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx DB 255,224 ; jmpq *%rax @@ -1220,7 +1220,7 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax DB 196,225,249,110,224 ; vmovq %rax,%xmm4 DB 196,226,125,33,228 ; vpmovsxbd %xmm4,%ymm4 - DB 196,194,93,140,57 ; vpmaskmovd (%r9),%ymm4,%ymm7 + DB 196,194,93,44,57 ; vmaskmovps (%r9),%ymm4,%ymm7 DB 233,40,255,255,255 ; jmpq 1340 <_sk_srcover_rgba_8888_hsw+0x1e> DB 185,8,0,0,0 ; mov $0x8,%ecx DB 68,41,193 ; sub %r8d,%ecx @@ -1229,7 +1229,7 @@ _sk_srcover_rgba_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax 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 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) DB 235,170 ; jmp 13e8 <_sk_srcover_rgba_8888_hsw+0xc6> PUBLIC _sk_clamp_0_hsw @@ -1707,8 +1707,8 @@ _sk_load_tables_hsw LABEL PROC DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 DB 117,105 ; jne 1c26 <_sk_load_tables_hsw+0x7e> - DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 - DB 197,229,219,13,54,51,0,0 ; vpand 0x3336(%rip),%ymm3,%ymm1 # 4f00 <_sk_callback_hsw+0x586> + DB 196,193,124,16,25 ; vmovups (%r9),%ymm3 + DB 197,228,84,13,54,51,0,0 ; vandps 0x3336(%rip),%ymm3,%ymm1 # 4f00 <_sk_callback_hsw+0x586> DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8 DB 72,139,72,8 ; mov 0x8(%rax),%rcx DB 76,139,72,16 ; mov 0x10(%rax),%r9 @@ -1734,7 +1734,7 @@ _sk_load_tables_hsw LABEL PROC DB 73,211,234 ; shr %cl,%r10 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 196,194,125,44,25 ; vmaskmovps (%r9),%ymm0,%ymm3 DB 233,115,255,255,255 ; jmpq 1bc2 <_sk_load_tables_hsw+0x1a> PUBLIC _sk_load_tables_u16_be_hsw @@ -3207,8 +3207,8 @@ _sk_load_8888_hsw LABEL PROC DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 DB 117,88 ; jne 3549 <_sk_load_8888_hsw+0x6d> - DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 - DB 197,229,219,5,194,26,0,0 ; vpand 0x1ac2(%rip),%ymm3,%ymm0 # 4fc0 <_sk_callback_hsw+0x646> + DB 196,193,124,16,25 ; vmovups (%r9),%ymm3 + DB 197,228,84,5,194,26,0,0 ; vandps 0x1ac2(%rip),%ymm3,%ymm0 # 4fc0 <_sk_callback_hsw+0x646> DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 DB 196,98,125,24,5,133,24,0,0 ; vbroadcastss 0x1885(%rip),%ymm8 # 4d90 <_sk_callback_hsw+0x416> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 @@ -3231,7 +3231,7 @@ _sk_load_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax 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 196,194,125,44,25 ; vmaskmovps (%r9),%ymm0,%ymm3 DB 235,135 ; jmp 34f6 <_sk_load_8888_hsw+0x1a> PUBLIC _sk_gather_8888_hsw @@ -3284,7 +3284,7 @@ _sk_store_8888_hsw LABEL PROC DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 77,133,192 ; test %r8,%r8 DB 117,12 ; jne 3658 <_sk_store_8888_hsw+0x73> - DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) + DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx DB 255,224 ; jmpq *%rax @@ -3295,7 +3295,7 @@ _sk_store_8888_hsw LABEL PROC DB 72,211,232 ; shr %cl,%rax 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 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) DB 235,211 ; jmp 3651 <_sk_store_8888_hsw+0x6c> PUBLIC _sk_load_f16_hsw @@ -5245,14 +5245,14 @@ _sk_seed_shader_avx LABEL PROC DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0 DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,231,102,0,0 ; vbroadcastss 0x66e7(%rip),%ymm1 # 6844 <_sk_callback_avx+0x11c> + DB 196,226,125,24,13,31,101,0,0 ; vbroadcastss 0x651f(%rip),%ymm1 # 667c <_sk_callback_avx+0x11c> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,88,2 ; vaddps (%rdx),%ymm0,%ymm0 DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,21,203,102,0,0 ; vbroadcastss 0x66cb(%rip),%ymm2 # 6848 <_sk_callback_avx+0x120> + DB 196,226,125,24,21,3,101,0,0 ; vbroadcastss 0x6503(%rip),%ymm2 # 6680 <_sk_callback_avx+0x120> DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3 DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4 DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 @@ -5273,7 +5273,7 @@ _sk_dither_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 196,66,125,24,8 ; vbroadcastss (%r8),%ymm9 DB 196,65,60,87,209 ; vxorps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,124,102,0,0 ; vbroadcastss 0x667c(%rip),%ymm11 # 684c <_sk_callback_avx+0x124> + DB 196,98,125,24,29,180,100,0,0 ; vbroadcastss 0x64b4(%rip),%ymm11 # 6684 <_sk_callback_avx+0x124> DB 196,65,44,84,203 ; vandps %ymm11,%ymm10,%ymm9 DB 196,193,25,114,241,5 ; vpslld $0x5,%xmm9,%xmm12 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 @@ -5284,8 +5284,8 @@ _sk_dither_avx LABEL PROC DB 196,67,125,25,219,1 ; vextractf128 $0x1,%ymm11,%xmm11 DB 196,193,33,114,243,4 ; vpslld $0x4,%xmm11,%xmm11 DB 196,67,29,24,219,1 ; vinsertf128 $0x1,%xmm11,%ymm12,%ymm11 - DB 196,98,125,24,37,61,102,0,0 ; vbroadcastss 0x663d(%rip),%ymm12 # 6850 <_sk_callback_avx+0x128> - DB 196,98,125,24,45,56,102,0,0 ; vbroadcastss 0x6638(%rip),%ymm13 # 6854 <_sk_callback_avx+0x12c> + DB 196,98,125,24,37,117,100,0,0 ; vbroadcastss 0x6475(%rip),%ymm12 # 6688 <_sk_callback_avx+0x128> + DB 196,98,125,24,45,112,100,0,0 ; vbroadcastss 0x6470(%rip),%ymm13 # 668c <_sk_callback_avx+0x12c> DB 196,65,44,84,245 ; vandps %ymm13,%ymm10,%ymm14 DB 196,193,1,114,246,2 ; vpslld $0x2,%xmm14,%xmm15 DB 196,67,125,25,246,1 ; vextractf128 $0x1,%ymm14,%xmm14 @@ -5312,9 +5312,9 @@ _sk_dither_avx LABEL PROC DB 196,65,12,86,202 ; vorps %ymm10,%ymm14,%ymm9 DB 196,65,60,86,193 ; vorps %ymm9,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,163,101,0,0 ; vbroadcastss 0x65a3(%rip),%ymm9 # 6858 <_sk_callback_avx+0x130> + DB 196,98,125,24,13,219,99,0,0 ; vbroadcastss 0x63db(%rip),%ymm9 # 6690 <_sk_callback_avx+0x130> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,153,101,0,0 ; vbroadcastss 0x6599(%rip),%ymm9 # 685c <_sk_callback_avx+0x134> + DB 196,98,125,24,13,209,99,0,0 ; vbroadcastss 0x63d1(%rip),%ymm9 # 6694 <_sk_callback_avx+0x134> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 196,98,125,24,72,8 ; vbroadcastss 0x8(%rax),%ymm9 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 @@ -5373,7 +5373,7 @@ _sk_clear_avx LABEL PROC PUBLIC _sk_srcatop_avx _sk_srcatop_avx LABEL PROC DB 197,252,89,199 ; vmulps %ymm7,%ymm0,%ymm0 - DB 196,98,125,24,5,240,100,0,0 ; vbroadcastss 0x64f0(%rip),%ymm8 # 6860 <_sk_callback_avx+0x138> + DB 196,98,125,24,5,40,99,0,0 ; vbroadcastss 0x6328(%rip),%ymm8 # 6698 <_sk_callback_avx+0x138> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -5392,7 +5392,7 @@ _sk_srcatop_avx LABEL PROC PUBLIC _sk_dstatop_avx _sk_dstatop_avx LABEL PROC DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 - DB 196,98,125,24,13,178,100,0,0 ; vbroadcastss 0x64b2(%rip),%ymm9 # 6864 <_sk_callback_avx+0x13c> + DB 196,98,125,24,13,234,98,0,0 ; vbroadcastss 0x62ea(%rip),%ymm9 # 669c <_sk_callback_avx+0x13c> DB 197,52,92,207 ; vsubps %ymm7,%ymm9,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,188,88,192 ; vaddps %ymm0,%ymm8,%ymm0 @@ -5428,7 +5428,7 @@ _sk_dstin_avx LABEL PROC PUBLIC _sk_srcout_avx _sk_srcout_avx LABEL PROC - DB 196,98,125,24,5,81,100,0,0 ; vbroadcastss 0x6451(%rip),%ymm8 # 6868 <_sk_callback_avx+0x140> + DB 196,98,125,24,5,137,98,0,0 ; vbroadcastss 0x6289(%rip),%ymm8 # 66a0 <_sk_callback_avx+0x140> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -5439,7 +5439,7 @@ _sk_srcout_avx LABEL PROC PUBLIC _sk_dstout_avx _sk_dstout_avx LABEL PROC - DB 196,226,125,24,5,52,100,0,0 ; vbroadcastss 0x6434(%rip),%ymm0 # 686c <_sk_callback_avx+0x144> + DB 196,226,125,24,5,108,98,0,0 ; vbroadcastss 0x626c(%rip),%ymm0 # 66a4 <_sk_callback_avx+0x144> DB 197,252,92,219 ; vsubps %ymm3,%ymm0,%ymm3 DB 197,228,89,196 ; vmulps %ymm4,%ymm3,%ymm0 DB 197,228,89,205 ; vmulps %ymm5,%ymm3,%ymm1 @@ -5450,7 +5450,7 @@ _sk_dstout_avx LABEL PROC PUBLIC _sk_srcover_avx _sk_srcover_avx LABEL PROC - DB 196,98,125,24,5,23,100,0,0 ; vbroadcastss 0x6417(%rip),%ymm8 # 6870 <_sk_callback_avx+0x148> + DB 196,98,125,24,5,79,98,0,0 ; vbroadcastss 0x624f(%rip),%ymm8 # 66a8 <_sk_callback_avx+0x148> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 DB 197,60,89,204 ; vmulps %ymm4,%ymm8,%ymm9 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -5465,7 +5465,7 @@ _sk_srcover_avx LABEL PROC PUBLIC _sk_dstover_avx _sk_dstover_avx LABEL PROC - DB 196,98,125,24,5,234,99,0,0 ; vbroadcastss 0x63ea(%rip),%ymm8 # 6874 <_sk_callback_avx+0x14c> + DB 196,98,125,24,5,34,98,0,0 ; vbroadcastss 0x6222(%rip),%ymm8 # 66ac <_sk_callback_avx+0x14c> DB 197,60,92,199 ; vsubps %ymm7,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,252,88,196 ; vaddps %ymm4,%ymm0,%ymm0 @@ -5489,7 +5489,7 @@ _sk_modulate_avx LABEL PROC PUBLIC _sk_multiply_avx _sk_multiply_avx LABEL PROC - DB 196,98,125,24,5,169,99,0,0 ; vbroadcastss 0x63a9(%rip),%ymm8 # 6878 <_sk_callback_avx+0x150> + DB 196,98,125,24,5,225,97,0,0 ; vbroadcastss 0x61e1(%rip),%ymm8 # 66b0 <_sk_callback_avx+0x150> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,208 ; vmulps %ymm0,%ymm9,%ymm10 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5543,7 +5543,7 @@ _sk_screen_avx LABEL PROC PUBLIC _sk_xor__avx _sk_xor__avx LABEL PROC - DB 196,98,125,24,5,248,98,0,0 ; vbroadcastss 0x62f8(%rip),%ymm8 # 687c <_sk_callback_avx+0x154> + DB 196,98,125,24,5,48,97,0,0 ; vbroadcastss 0x6130(%rip),%ymm8 # 66b4 <_sk_callback_avx+0x154> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,192 ; vmulps %ymm0,%ymm9,%ymm0 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5578,7 +5578,7 @@ _sk_darken_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,95,209 ; vmaxps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,120,98,0,0 ; vbroadcastss 0x6278(%rip),%ymm8 # 6880 <_sk_callback_avx+0x158> + DB 196,98,125,24,5,176,96,0,0 ; vbroadcastss 0x60b0(%rip),%ymm8 # 66b8 <_sk_callback_avx+0x158> 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 @@ -5602,7 +5602,7 @@ _sk_lighten_avx LABEL PROC DB 197,100,89,206 ; vmulps %ymm6,%ymm3,%ymm9 DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,36,98,0,0 ; vbroadcastss 0x6224(%rip),%ymm8 # 6884 <_sk_callback_avx+0x15c> + DB 196,98,125,24,5,92,96,0,0 ; vbroadcastss 0x605c(%rip),%ymm8 # 66bc <_sk_callback_avx+0x15c> 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 @@ -5629,7 +5629,7 @@ _sk_difference_avx LABEL PROC DB 196,193,108,93,209 ; vminps %ymm9,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,196,97,0,0 ; vbroadcastss 0x61c4(%rip),%ymm8 # 6888 <_sk_callback_avx+0x160> + DB 196,98,125,24,5,252,95,0,0 ; vbroadcastss 0x5ffc(%rip),%ymm8 # 66c0 <_sk_callback_avx+0x160> 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 @@ -5650,7 +5650,7 @@ _sk_exclusion_avx LABEL PROC DB 197,236,89,214 ; vmulps %ymm6,%ymm2,%ymm2 DB 197,236,88,210 ; vaddps %ymm2,%ymm2,%ymm2 DB 197,188,92,210 ; vsubps %ymm2,%ymm8,%ymm2 - DB 196,98,125,24,5,127,97,0,0 ; vbroadcastss 0x617f(%rip),%ymm8 # 688c <_sk_callback_avx+0x164> + DB 196,98,125,24,5,183,95,0,0 ; vbroadcastss 0x5fb7(%rip),%ymm8 # 66c4 <_sk_callback_avx+0x164> 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 @@ -5659,7 +5659,7 @@ _sk_exclusion_avx LABEL PROC PUBLIC _sk_colorburn_avx _sk_colorburn_avx LABEL PROC - DB 196,98,125,24,5,106,97,0,0 ; vbroadcastss 0x616a(%rip),%ymm8 # 6890 <_sk_callback_avx+0x168> + DB 196,98,125,24,5,162,95,0,0 ; vbroadcastss 0x5fa2(%rip),%ymm8 # 66c8 <_sk_callback_avx+0x168> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,52,89,216 ; vmulps %ymm0,%ymm9,%ymm11 DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10 @@ -5719,7 +5719,7 @@ _sk_colorburn_avx LABEL PROC PUBLIC _sk_colordodge_avx _sk_colordodge_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 196,98,125,24,13,102,96,0,0 ; vbroadcastss 0x6066(%rip),%ymm9 # 6894 <_sk_callback_avx+0x16c> + DB 196,98,125,24,13,158,94,0,0 ; vbroadcastss 0x5e9e(%rip),%ymm9 # 66cc <_sk_callback_avx+0x16c> DB 197,52,92,215 ; vsubps %ymm7,%ymm9,%ymm10 DB 197,44,89,216 ; vmulps %ymm0,%ymm10,%ymm11 DB 197,52,92,203 ; vsubps %ymm3,%ymm9,%ymm9 @@ -5774,7 +5774,7 @@ _sk_colordodge_avx LABEL PROC PUBLIC _sk_hardlight_avx _sk_hardlight_avx LABEL PROC - DB 196,98,125,24,5,120,95,0,0 ; vbroadcastss 0x5f78(%rip),%ymm8 # 6898 <_sk_callback_avx+0x170> + DB 196,98,125,24,5,176,93,0,0 ; vbroadcastss 0x5db0(%rip),%ymm8 # 66d0 <_sk_callback_avx+0x170> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5827,7 +5827,7 @@ _sk_hardlight_avx LABEL PROC PUBLIC _sk_overlay_avx _sk_overlay_avx LABEL PROC - DB 196,98,125,24,5,161,94,0,0 ; vbroadcastss 0x5ea1(%rip),%ymm8 # 689c <_sk_callback_avx+0x174> + DB 196,98,125,24,5,217,92,0,0 ; vbroadcastss 0x5cd9(%rip),%ymm8 # 66d4 <_sk_callback_avx+0x174> DB 197,60,92,215 ; vsubps %ymm7,%ymm8,%ymm10 DB 197,44,89,200 ; vmulps %ymm0,%ymm10,%ymm9 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -5892,10 +5892,10 @@ _sk_softlight_avx LABEL PROC DB 196,65,60,88,192 ; vaddps %ymm8,%ymm8,%ymm8 DB 196,65,60,89,216 ; vmulps %ymm8,%ymm8,%ymm11 DB 196,65,60,88,195 ; vaddps %ymm11,%ymm8,%ymm8 - DB 196,98,125,24,29,148,93,0,0 ; vbroadcastss 0x5d94(%rip),%ymm11 # 68a4 <_sk_callback_avx+0x17c> + DB 196,98,125,24,29,204,91,0,0 ; vbroadcastss 0x5bcc(%rip),%ymm11 # 66dc <_sk_callback_avx+0x17c> DB 196,65,28,88,235 ; vaddps %ymm11,%ymm12,%ymm13 DB 196,65,20,89,192 ; vmulps %ymm8,%ymm13,%ymm8 - DB 196,98,125,24,45,133,93,0,0 ; vbroadcastss 0x5d85(%rip),%ymm13 # 68a8 <_sk_callback_avx+0x180> + DB 196,98,125,24,45,189,91,0,0 ; vbroadcastss 0x5bbd(%rip),%ymm13 # 66e0 <_sk_callback_avx+0x180> DB 196,65,28,89,245 ; vmulps %ymm13,%ymm12,%ymm14 DB 196,65,12,88,192 ; vaddps %ymm8,%ymm14,%ymm8 DB 196,65,124,82,244 ; vrsqrtps %ymm12,%ymm14 @@ -5906,7 +5906,7 @@ _sk_softlight_avx LABEL PROC DB 197,4,194,255,2 ; vcmpleps %ymm7,%ymm15,%ymm15 DB 196,67,13,74,240,240 ; vblendvps %ymm15,%ymm8,%ymm14,%ymm14 DB 197,116,88,249 ; vaddps %ymm1,%ymm1,%ymm15 - DB 196,98,125,24,5,67,93,0,0 ; vbroadcastss 0x5d43(%rip),%ymm8 # 68a0 <_sk_callback_avx+0x178> + DB 196,98,125,24,5,123,91,0,0 ; vbroadcastss 0x5b7b(%rip),%ymm8 # 66d8 <_sk_callback_avx+0x178> DB 196,65,60,92,228 ; vsubps %ymm12,%ymm8,%ymm12 DB 197,132,92,195 ; vsubps %ymm3,%ymm15,%ymm0 DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 @@ -6033,12 +6033,12 @@ _sk_hue_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,13,91,0,0 ; vbroadcastss 0x5b0d(%rip),%ymm14 # 68ac <_sk_callback_avx+0x184> + DB 196,98,125,24,53,69,89,0,0 ; vbroadcastss 0x5945(%rip),%ymm14 # 66e4 <_sk_callback_avx+0x184> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,3,91,0,0 ; vbroadcastss 0x5b03(%rip),%ymm15 # 68b0 <_sk_callback_avx+0x188> + DB 196,98,125,24,61,59,89,0,0 ; vbroadcastss 0x593b(%rip),%ymm15 # 66e8 <_sk_callback_avx+0x188> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,244,90,0,0 ; vbroadcastss 0x5af4(%rip),%ymm0 # 68b4 <_sk_callback_avx+0x18c> + DB 196,226,125,24,5,44,89,0,0 ; vbroadcastss 0x592c(%rip),%ymm0 # 66ec <_sk_callback_avx+0x18c> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -6099,7 +6099,7 @@ _sk_hue_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,205,89,0,0 ; vbroadcastss 0x59cd(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x190> + DB 196,98,125,24,5,5,88,0,0 ; vbroadcastss 0x5805(%rip),%ymm8 # 66f0 <_sk_callback_avx+0x190> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6156,12 +6156,12 @@ _sk_saturation_avx LABEL PROC DB 196,65,28,89,219 ; vmulps %ymm11,%ymm12,%ymm11 DB 196,65,36,94,222 ; vdivps %ymm14,%ymm11,%ymm11 DB 196,67,37,74,224,240 ; vblendvps %ymm15,%ymm8,%ymm11,%ymm12 - DB 196,98,125,24,53,213,88,0,0 ; vbroadcastss 0x58d5(%rip),%ymm14 # 68bc <_sk_callback_avx+0x194> + DB 196,98,125,24,53,13,87,0,0 ; vbroadcastss 0x570d(%rip),%ymm14 # 66f4 <_sk_callback_avx+0x194> DB 196,65,92,89,222 ; vmulps %ymm14,%ymm4,%ymm11 - DB 196,98,125,24,61,203,88,0,0 ; vbroadcastss 0x58cb(%rip),%ymm15 # 68c0 <_sk_callback_avx+0x198> + DB 196,98,125,24,61,3,87,0,0 ; vbroadcastss 0x5703(%rip),%ymm15 # 66f8 <_sk_callback_avx+0x198> DB 196,65,84,89,239 ; vmulps %ymm15,%ymm5,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 - DB 196,226,125,24,5,188,88,0,0 ; vbroadcastss 0x58bc(%rip),%ymm0 # 68c4 <_sk_callback_avx+0x19c> + DB 196,226,125,24,5,244,86,0,0 ; vbroadcastss 0x56f4(%rip),%ymm0 # 66fc <_sk_callback_avx+0x19c> DB 197,76,89,232 ; vmulps %ymm0,%ymm6,%ymm13 DB 196,65,36,88,221 ; vaddps %ymm13,%ymm11,%ymm11 DB 196,65,52,89,238 ; vmulps %ymm14,%ymm9,%ymm13 @@ -6222,7 +6222,7 @@ _sk_saturation_avx LABEL PROC DB 196,65,36,95,208 ; vmaxps %ymm8,%ymm11,%ymm10 DB 196,195,109,74,209,240 ; vblendvps %ymm15,%ymm9,%ymm2,%ymm2 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,149,87,0,0 ; vbroadcastss 0x5795(%rip),%ymm8 # 68c8 <_sk_callback_avx+0x1a0> + DB 196,98,125,24,5,205,85,0,0 ; vbroadcastss 0x55cd(%rip),%ymm8 # 6700 <_sk_callback_avx+0x1a0> DB 197,60,92,207 ; vsubps %ymm7,%ymm8,%ymm9 DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1 DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 @@ -6251,12 +6251,12 @@ _sk_color_avx LABEL PROC DB 197,252,17,68,36,32 ; vmovups %ymm0,0x20(%rsp) DB 197,124,89,199 ; vmulps %ymm7,%ymm0,%ymm8 DB 197,116,89,207 ; vmulps %ymm7,%ymm1,%ymm9 - DB 196,98,125,24,45,37,87,0,0 ; vbroadcastss 0x5725(%rip),%ymm13 # 68cc <_sk_callback_avx+0x1a4> + DB 196,98,125,24,45,93,85,0,0 ; vbroadcastss 0x555d(%rip),%ymm13 # 6704 <_sk_callback_avx+0x1a4> DB 196,65,92,89,213 ; vmulps %ymm13,%ymm4,%ymm10 - DB 196,98,125,24,53,27,87,0,0 ; vbroadcastss 0x571b(%rip),%ymm14 # 68d0 <_sk_callback_avx+0x1a8> + DB 196,98,125,24,53,83,85,0,0 ; vbroadcastss 0x5553(%rip),%ymm14 # 6708 <_sk_callback_avx+0x1a8> DB 196,65,84,89,222 ; vmulps %ymm14,%ymm5,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,12,87,0,0 ; vbroadcastss 0x570c(%rip),%ymm15 # 68d4 <_sk_callback_avx+0x1ac> + DB 196,98,125,24,61,68,85,0,0 ; vbroadcastss 0x5544(%rip),%ymm15 # 670c <_sk_callback_avx+0x1ac> DB 196,65,76,89,223 ; vmulps %ymm15,%ymm6,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -6319,7 +6319,7 @@ _sk_color_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,211,85,0,0 ; vbroadcastss 0x55d3(%rip),%ymm0 # 68d8 <_sk_callback_avx+0x1b0> + DB 196,226,125,24,5,11,84,0,0 ; vbroadcastss 0x540b(%rip),%ymm0 # 6710 <_sk_callback_avx+0x1b0> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,84,36,32 ; vmulps 0x20(%rsp),%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -6349,12 +6349,12 @@ _sk_luminosity_avx LABEL PROC DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 197,100,89,196 ; vmulps %ymm4,%ymm3,%ymm8 DB 197,100,89,205 ; vmulps %ymm5,%ymm3,%ymm9 - DB 196,98,125,24,45,95,85,0,0 ; vbroadcastss 0x555f(%rip),%ymm13 # 68dc <_sk_callback_avx+0x1b4> + DB 196,98,125,24,45,151,83,0,0 ; vbroadcastss 0x5397(%rip),%ymm13 # 6714 <_sk_callback_avx+0x1b4> DB 196,65,108,89,213 ; vmulps %ymm13,%ymm2,%ymm10 - DB 196,98,125,24,53,85,85,0,0 ; vbroadcastss 0x5555(%rip),%ymm14 # 68e0 <_sk_callback_avx+0x1b8> + DB 196,98,125,24,53,141,83,0,0 ; vbroadcastss 0x538d(%rip),%ymm14 # 6718 <_sk_callback_avx+0x1b8> DB 196,65,116,89,222 ; vmulps %ymm14,%ymm1,%ymm11 DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,61,70,85,0,0 ; vbroadcastss 0x5546(%rip),%ymm15 # 68e4 <_sk_callback_avx+0x1bc> + DB 196,98,125,24,61,126,83,0,0 ; vbroadcastss 0x537e(%rip),%ymm15 # 671c <_sk_callback_avx+0x1bc> DB 196,65,28,89,223 ; vmulps %ymm15,%ymm12,%ymm11 DB 196,193,44,88,195 ; vaddps %ymm11,%ymm10,%ymm0 DB 196,65,60,89,221 ; vmulps %ymm13,%ymm8,%ymm11 @@ -6417,7 +6417,7 @@ _sk_luminosity_avx LABEL PROC DB 196,65,44,95,207 ; vmaxps %ymm15,%ymm10,%ymm9 DB 196,195,37,74,192,0 ; vblendvps %ymm0,%ymm8,%ymm11,%ymm0 DB 196,65,124,95,199 ; vmaxps %ymm15,%ymm0,%ymm8 - DB 196,226,125,24,5,13,84,0,0 ; vbroadcastss 0x540d(%rip),%ymm0 # 68e8 <_sk_callback_avx+0x1c0> + DB 196,226,125,24,5,69,82,0,0 ; vbroadcastss 0x5245(%rip),%ymm0 # 6720 <_sk_callback_avx+0x1c0> DB 197,124,92,215 ; vsubps %ymm7,%ymm0,%ymm10 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 197,124,92,219 ; vsubps %ymm3,%ymm0,%ymm11 @@ -6440,32 +6440,35 @@ _sk_luminosity_avx LABEL PROC PUBLIC _sk_srcover_rgba_8888_avx _sk_srcover_rgba_8888_avx LABEL PROC + DB 72,131,236,16 ; sub $0x10,%rsp + DB 73,137,200 ; mov %rcx,%r8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,46,1,0,0 ; jne 1665 <_sk_srcover_rgba_8888_avx+0x13c> - DB 196,65,124,16,4,186 ; vmovups (%r10,%rdi,4),%ymm8 - DB 197,124,40,13,59,87,0,0 ; vmovaps 0x573b(%rip),%ymm9 # 6c80 <_sk_callback_avx+0x558> - DB 196,193,60,84,225 ; vandps %ymm9,%ymm8,%ymm4 + 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 15,133,47,1,0,0 ; jne 1675 <_sk_srcover_rgba_8888_avx+0x14c> + DB 196,193,124,16,57 ; vmovups (%r9),%ymm7 + DB 197,124,40,13,13,86,0,0 ; vmovaps 0x560d(%rip),%ymm9 # 6b60 <_sk_callback_avx+0x600> + DB 196,193,68,84,225 ; vandps %ymm9,%ymm7,%ymm4 DB 197,252,91,228 ; vcvtdq2ps %ymm4,%ymm4 - DB 196,193,81,114,208,8 ; vpsrld $0x8,%xmm8,%xmm5 - DB 196,99,125,25,199,1 ; vextractf128 $0x1,%ymm8,%xmm7 - DB 197,201,114,215,8 ; vpsrld $0x8,%xmm7,%xmm6 + DB 197,209,114,215,8 ; vpsrld $0x8,%xmm7,%xmm5 + DB 196,195,125,25,248,1 ; vextractf128 $0x1,%ymm7,%xmm8 + DB 196,193,73,114,208,8 ; vpsrld $0x8,%xmm8,%xmm6 DB 196,227,85,24,238,1 ; vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 DB 196,193,84,84,233 ; vandps %ymm9,%ymm5,%ymm5 DB 197,252,91,237 ; vcvtdq2ps %ymm5,%ymm5 - DB 196,193,41,114,208,16 ; vpsrld $0x10,%xmm8,%xmm10 - DB 197,201,114,215,16 ; vpsrld $0x10,%xmm7,%xmm6 + DB 197,169,114,215,16 ; vpsrld $0x10,%xmm7,%xmm10 + DB 196,193,73,114,208,16 ; vpsrld $0x10,%xmm8,%xmm6 DB 196,227,45,24,246,1 ; vinsertf128 $0x1,%xmm6,%ymm10,%ymm6 DB 196,193,76,84,241 ; vandps %ymm9,%ymm6,%ymm6 DB 197,252,91,246 ; vcvtdq2ps %ymm6,%ymm6 - DB 196,193,57,114,208,24 ; vpsrld $0x18,%xmm8,%xmm8 - DB 197,193,114,215,24 ; vpsrld $0x18,%xmm7,%xmm7 - DB 196,227,61,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm8,%ymm7 + DB 197,177,114,215,24 ; vpsrld $0x18,%xmm7,%xmm9 + DB 196,193,65,114,208,24 ; vpsrld $0x18,%xmm8,%xmm7 + DB 196,227,53,24,255,1 ; vinsertf128 $0x1,%xmm7,%ymm9,%ymm7 DB 197,252,91,255 ; vcvtdq2ps %ymm7,%ymm7 - DB 196,98,125,24,5,70,83,0,0 ; vbroadcastss 0x5346(%rip),%ymm8 # 68ec <_sk_callback_avx+0x1c4> + DB 196,98,125,24,5,112,81,0,0 ; vbroadcastss 0x5170(%rip),%ymm8 # 6724 <_sk_callback_avx+0x1c4> DB 197,60,92,195 ; vsubps %ymm3,%ymm8,%ymm8 - DB 196,98,125,24,13,61,83,0,0 ; vbroadcastss 0x533d(%rip),%ymm9 # 68f0 <_sk_callback_avx+0x1c8> + DB 196,98,125,24,13,103,81,0,0 ; vbroadcastss 0x5167(%rip),%ymm9 # 6728 <_sk_callback_avx+0x1c8> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,60,89,212 ; vmulps %ymm4,%ymm8,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 @@ -6497,96 +6500,41 @@ _sk_srcover_rgba_8888_avx LABEL PROC DB 196,67,37,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 DB 196,65,53,86,202 ; vorpd %ymm10,%ymm9,%ymm9 DB 196,65,61,86,193 ; vorpd %ymm9,%ymm8,%ymm8 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,183,0,0,0 ; jne 1712 <_sk_srcover_rgba_8888_avx+0x1e9> - DB 196,65,124,17,4,186 ; vmovups %ymm8,(%r10,%rdi,4) + DB 77,133,192 ; test %r8,%r8 + DB 117,91 ; jne 16c0 <_sk_srcover_rgba_8888_avx+0x197> + DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,137,193 ; mov %r8,%rcx + DB 72,131,196,16 ; add $0x10,%rsp DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,191,254,255,255 ; ja 153d <_sk_srcover_rgba_8888_avx+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,255,0,0,0 ; lea 0xff(%rip),%r9 # 1788 <_sk_srcover_rgba_8888_avx+0x25f> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,193,121,110,100,186,24 ; vmovd 0x18(%r10,%rdi,4),%xmm4 - DB 197,249,112,228,68 ; vpshufd $0x44,%xmm4,%xmm4 - DB 196,227,125,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm0,%ymm4 - DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5 - DB 196,99,85,12,196,64 ; vblendps $0x40,%ymm4,%ymm5,%ymm8 - DB 196,99,125,25,196,1 ; vextractf128 $0x1,%ymm8,%xmm4 - DB 196,195,89,34,100,186,20,1 ; vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm4,%xmm4 - DB 196,99,61,24,196,1 ; vinsertf128 $0x1,%xmm4,%ymm8,%ymm8 - DB 196,99,125,25,196,1 ; vextractf128 $0x1,%ymm8,%xmm4 - DB 196,195,89,34,100,186,16,0 ; vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm4,%xmm4 - DB 196,99,61,24,196,1 ; vinsertf128 $0x1,%xmm4,%ymm8,%ymm8 - DB 196,195,57,34,100,186,12,3 ; vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm8,%xmm4 - DB 196,99,61,12,196,15 ; vblendps $0xf,%ymm4,%ymm8,%ymm8 - DB 196,195,57,34,100,186,8,2 ; vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm8,%xmm4 - DB 196,99,61,12,196,15 ; vblendps $0xf,%ymm4,%ymm8,%ymm8 - DB 196,195,57,34,100,186,4,1 ; vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm8,%xmm4 - DB 196,99,61,12,196,15 ; vblendps $0xf,%ymm4,%ymm8,%ymm8 - DB 196,195,57,34,36,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm8,%xmm4 - DB 196,99,61,12,196,15 ; vblendps $0xf,%ymm4,%ymm8,%ymm8 - DB 233,43,254,255,255 ; jmpq 153d <_sk_srcover_rgba_8888_avx+0x14> - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - DB 65,254,200 ; dec %r8b - DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,59,255,255,255 ; ja 1661 <_sk_srcover_rgba_8888_avx+0x138> - DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,115,0,0,0 ; lea 0x73(%rip),%r8 # 17a4 <_sk_srcover_rgba_8888_avx+0x27b> - DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax - DB 76,1,192 ; add %r8,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 196,67,121,22,76,186,24,2 ; vpextrd $0x2,%xmm9,0x18(%r10,%rdi,4) - DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 196,67,121,22,76,186,20,1 ; vpextrd $0x1,%xmm9,0x14(%r10,%rdi,4) - DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 196,65,122,17,76,186,16 ; vmovss %xmm9,0x10(%r10,%rdi,4) - DB 196,67,121,22,68,186,12,3 ; vpextrd $0x3,%xmm8,0xc(%r10,%rdi,4) - DB 196,67,121,22,68,186,8,2 ; vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4) - DB 196,67,121,22,68,186,4,1 ; vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4) - DB 196,65,121,126,4,186 ; vmovd %xmm8,(%r10,%rdi,4) - DB 233,219,254,255,255 ; jmpq 1661 <_sk_srcover_rgba_8888_avx+0x138> - DB 102,144 ; xchg %ax,%ax - DB 120,255 ; js 1789 <_sk_srcover_rgba_8888_avx+0x260> - DB 255 ; (bad) - DB 255,106,255 ; ljmp *-0x1(%rdx) - DB 255 ; (bad) - DB 255,92,255,255 ; lcall *-0x1(%rdi,%rdi,8) - DB 255,78,255 ; decl -0x1(%rsi) - DB 255 ; (bad) - DB 255 ; (bad) - DB 58,255 ; cmp %bh,%bh - DB 255 ; (bad) - DB 255,38 ; jmpq *(%rsi) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,10 ; decl (%rdx) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,215 ; callq *%rdi - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,207 ; dec %edi - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,199 ; inc %edi - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 191,255,255,255,178 ; mov $0xb2ffffff,%edi - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,164,255,255,255,150,255 ; jmpq *-0x690001(%rdi,%rdi,8) - DB 255 ; (bad) - DB 255 ; .byte 0xff + DB 185,8,0,0,0 ; mov $0x8,%ecx + DB 68,41,193 ; sub %r8d,%ecx + DB 192,225,3 ; shl $0x3,%cl + DB 72,199,192,255,255,255,255 ; mov $0xffffffffffffffff,%rax + DB 72,211,232 ; shr %cl,%rax + DB 196,225,249,110,224 ; vmovq %rax,%xmm4 + DB 196,226,121,48,228 ; vpmovzxbw %xmm4,%xmm4 + DB 196,226,89,0,45,19,84,0,0 ; vpshufb 0x5413(%rip),%xmm4,%xmm5 # 6ab0 <_sk_callback_avx+0x550> + DB 196,226,121,33,237 ; vpmovsxbd %xmm5,%xmm5 + DB 196,226,89,0,37,21,84,0,0 ; vpshufb 0x5415(%rip),%xmm4,%xmm4 # 6ac0 <_sk_callback_avx+0x560> + DB 196,226,121,33,228 ; vpmovsxbd %xmm4,%xmm4 + DB 196,227,85,24,228,1 ; vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 + DB 196,194,93,44,57 ; vmaskmovps (%r9),%ymm4,%ymm7 + DB 233,139,254,255,255 ; jmpq 154b <_sk_srcover_rgba_8888_avx+0x22> + DB 185,8,0,0,0 ; mov $0x8,%ecx + DB 68,41,193 ; sub %r8d,%ecx + DB 192,225,3 ; shl $0x3,%cl + DB 72,199,192,255,255,255,255 ; mov $0xffffffffffffffff,%rax + DB 72,211,232 ; shr %cl,%rax + DB 196,97,249,110,200 ; vmovq %rax,%xmm9 + DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 + DB 196,98,49,0,21,200,83,0,0 ; vpshufb 0x53c8(%rip),%xmm9,%xmm10 # 6ab0 <_sk_callback_avx+0x550> + DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10 + DB 196,98,49,0,13,202,83,0,0 ; vpshufb 0x53ca(%rip),%xmm9,%xmm9 # 6ac0 <_sk_callback_avx+0x560> + DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9 + DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 + DB 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) + DB 233,95,255,255,255 ; jmpq 166a <_sk_srcover_rgba_8888_avx+0x141> PUBLIC _sk_clamp_0_avx _sk_clamp_0_avx LABEL PROC @@ -6600,7 +6548,7 @@ _sk_clamp_0_avx LABEL PROC PUBLIC _sk_clamp_1_avx _sk_clamp_1_avx LABEL PROC - DB 196,98,125,24,5,14,81,0,0 ; vbroadcastss 0x510e(%rip),%ymm8 # 68f4 <_sk_callback_avx+0x1cc> + DB 196,98,125,24,5,251,79,0,0 ; vbroadcastss 0x4ffb(%rip),%ymm8 # 672c <_sk_callback_avx+0x1cc> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 @@ -6610,7 +6558,7 @@ _sk_clamp_1_avx LABEL PROC PUBLIC _sk_clamp_a_avx _sk_clamp_a_avx LABEL PROC - DB 196,98,125,24,5,241,80,0,0 ; vbroadcastss 0x50f1(%rip),%ymm8 # 68f8 <_sk_callback_avx+0x1d0> + DB 196,98,125,24,5,222,79,0,0 ; vbroadcastss 0x4fde(%rip),%ymm8 # 6730 <_sk_callback_avx+0x1d0> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 197,252,93,195 ; vminps %ymm3,%ymm0,%ymm0 DB 197,244,93,203 ; vminps %ymm3,%ymm1,%ymm1 @@ -6682,7 +6630,7 @@ PUBLIC _sk_unpremul_avx _sk_unpremul_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,65,100,194,200,0 ; vcmpeqps %ymm8,%ymm3,%ymm9 - DB 196,98,125,24,21,57,80,0,0 ; vbroadcastss 0x5039(%rip),%ymm10 # 68fc <_sk_callback_avx+0x1d4> + DB 196,98,125,24,21,38,79,0,0 ; vbroadcastss 0x4f26(%rip),%ymm10 # 6734 <_sk_callback_avx+0x1d4> DB 197,44,94,211 ; vdivps %ymm3,%ymm10,%ymm10 DB 196,67,45,74,192,144 ; vblendvps %ymm9,%ymm8,%ymm10,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 @@ -6693,17 +6641,17 @@ _sk_unpremul_avx LABEL PROC PUBLIC _sk_from_srgb_avx _sk_from_srgb_avx LABEL PROC - DB 196,98,125,24,5,26,80,0,0 ; vbroadcastss 0x501a(%rip),%ymm8 # 6900 <_sk_callback_avx+0x1d8> + DB 196,98,125,24,5,7,79,0,0 ; vbroadcastss 0x4f07(%rip),%ymm8 # 6738 <_sk_callback_avx+0x1d8> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 197,124,89,208 ; vmulps %ymm0,%ymm0,%ymm10 - DB 196,98,125,24,29,12,80,0,0 ; vbroadcastss 0x500c(%rip),%ymm11 # 6904 <_sk_callback_avx+0x1dc> + DB 196,98,125,24,29,249,78,0,0 ; vbroadcastss 0x4ef9(%rip),%ymm11 # 673c <_sk_callback_avx+0x1dc> DB 196,65,124,89,227 ; vmulps %ymm11,%ymm0,%ymm12 - DB 196,98,125,24,45,2,80,0,0 ; vbroadcastss 0x5002(%rip),%ymm13 # 6908 <_sk_callback_avx+0x1e0> + DB 196,98,125,24,45,239,78,0,0 ; vbroadcastss 0x4eef(%rip),%ymm13 # 6740 <_sk_callback_avx+0x1e0> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,44,89,212 ; vmulps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,37,243,79,0,0 ; vbroadcastss 0x4ff3(%rip),%ymm12 # 690c <_sk_callback_avx+0x1e4> + DB 196,98,125,24,37,224,78,0,0 ; vbroadcastss 0x4ee0(%rip),%ymm12 # 6744 <_sk_callback_avx+0x1e4> DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10 - DB 196,98,125,24,53,233,79,0,0 ; vbroadcastss 0x4fe9(%rip),%ymm14 # 6910 <_sk_callback_avx+0x1e8> + DB 196,98,125,24,53,214,78,0,0 ; vbroadcastss 0x4ed6(%rip),%ymm14 # 6748 <_sk_callback_avx+0x1e8> DB 196,193,124,194,198,1 ; vcmpltps %ymm14,%ymm0,%ymm0 DB 196,195,45,74,193,0 ; vblendvps %ymm0,%ymm9,%ymm10,%ymm0 DB 196,65,116,89,200 ; vmulps %ymm8,%ymm1,%ymm9 @@ -6728,20 +6676,20 @@ _sk_from_srgb_avx LABEL PROC PUBLIC _sk_to_srgb_avx _sk_to_srgb_avx LABEL PROC DB 197,124,82,200 ; vrsqrtps %ymm0,%ymm9 - DB 196,98,125,24,5,126,79,0,0 ; vbroadcastss 0x4f7e(%rip),%ymm8 # 6914 <_sk_callback_avx+0x1ec> + DB 196,98,125,24,5,107,78,0,0 ; vbroadcastss 0x4e6b(%rip),%ymm8 # 674c <_sk_callback_avx+0x1ec> DB 196,65,124,89,208 ; vmulps %ymm8,%ymm0,%ymm10 - DB 196,98,125,24,29,116,79,0,0 ; vbroadcastss 0x4f74(%rip),%ymm11 # 6918 <_sk_callback_avx+0x1f0> + DB 196,98,125,24,29,97,78,0,0 ; vbroadcastss 0x4e61(%rip),%ymm11 # 6750 <_sk_callback_avx+0x1f0> DB 196,65,52,89,227 ; vmulps %ymm11,%ymm9,%ymm12 - DB 196,98,125,24,45,106,79,0,0 ; vbroadcastss 0x4f6a(%rip),%ymm13 # 691c <_sk_callback_avx+0x1f4> + DB 196,98,125,24,45,87,78,0,0 ; vbroadcastss 0x4e57(%rip),%ymm13 # 6754 <_sk_callback_avx+0x1f4> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,52,89,228 ; vmulps %ymm12,%ymm9,%ymm12 - DB 196,98,125,24,53,91,79,0,0 ; vbroadcastss 0x4f5b(%rip),%ymm14 # 6920 <_sk_callback_avx+0x1f8> + DB 196,98,125,24,53,72,78,0,0 ; vbroadcastss 0x4e48(%rip),%ymm14 # 6758 <_sk_callback_avx+0x1f8> DB 196,65,28,88,230 ; vaddps %ymm14,%ymm12,%ymm12 - DB 196,98,125,24,61,81,79,0,0 ; vbroadcastss 0x4f51(%rip),%ymm15 # 6924 <_sk_callback_avx+0x1fc> + DB 196,98,125,24,61,62,78,0,0 ; vbroadcastss 0x4e3e(%rip),%ymm15 # 675c <_sk_callback_avx+0x1fc> DB 196,65,52,88,207 ; vaddps %ymm15,%ymm9,%ymm9 DB 196,65,124,83,201 ; vrcpps %ymm9,%ymm9 DB 196,65,52,89,204 ; vmulps %ymm12,%ymm9,%ymm9 - DB 196,98,125,24,37,61,79,0,0 ; vbroadcastss 0x4f3d(%rip),%ymm12 # 6928 <_sk_callback_avx+0x200> + DB 196,98,125,24,37,42,78,0,0 ; vbroadcastss 0x4e2a(%rip),%ymm12 # 6760 <_sk_callback_avx+0x200> DB 196,193,124,194,196,1 ; vcmpltps %ymm12,%ymm0,%ymm0 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 197,124,82,201 ; vrsqrtps %ymm1,%ymm9 @@ -6776,7 +6724,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,124,93,201 ; vminps %ymm1,%ymm0,%ymm9 DB 197,52,93,202 ; vminps %ymm2,%ymm9,%ymm9 DB 196,65,60,92,209 ; vsubps %ymm9,%ymm8,%ymm10 - DB 196,98,125,24,29,163,78,0,0 ; vbroadcastss 0x4ea3(%rip),%ymm11 # 692c <_sk_callback_avx+0x204> + DB 196,98,125,24,29,144,77,0,0 ; vbroadcastss 0x4d90(%rip),%ymm11 # 6764 <_sk_callback_avx+0x204> DB 196,65,36,94,218 ; vdivps %ymm10,%ymm11,%ymm11 DB 197,116,92,226 ; vsubps %ymm2,%ymm1,%ymm12 DB 196,65,28,89,227 ; vmulps %ymm11,%ymm12,%ymm12 @@ -6786,19 +6734,19 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 197,252,92,201 ; vsubps %ymm1,%ymm0,%ymm1 DB 196,193,116,89,203 ; vmulps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,124,78,0,0 ; vbroadcastss 0x4e7c(%rip),%ymm11 # 6938 <_sk_callback_avx+0x210> + DB 196,98,125,24,29,105,77,0,0 ; vbroadcastss 0x4d69(%rip),%ymm11 # 6770 <_sk_callback_avx+0x210> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,106,78,0,0 ; vbroadcastss 0x4e6a(%rip),%ymm11 # 6934 <_sk_callback_avx+0x20c> + DB 196,98,125,24,29,87,77,0,0 ; vbroadcastss 0x4d57(%rip),%ymm11 # 676c <_sk_callback_avx+0x20c> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,227,117,74,202,224 ; vblendvps %ymm14,%ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,82,78,0,0 ; vbroadcastss 0x4e52(%rip),%ymm2 # 6930 <_sk_callback_avx+0x208> + DB 196,226,125,24,21,63,77,0,0 ; vbroadcastss 0x4d3f(%rip),%ymm2 # 6768 <_sk_callback_avx+0x208> DB 196,65,12,87,246 ; vxorps %ymm14,%ymm14,%ymm14 DB 196,227,13,74,210,208 ; vblendvps %ymm13,%ymm2,%ymm14,%ymm2 DB 197,188,194,192,0 ; vcmpeqps %ymm0,%ymm8,%ymm0 DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 DB 196,227,117,74,194,0 ; vblendvps %ymm0,%ymm2,%ymm1,%ymm0 DB 196,193,60,88,201 ; vaddps %ymm9,%ymm8,%ymm1 - DB 196,98,125,24,37,57,78,0,0 ; vbroadcastss 0x4e39(%rip),%ymm12 # 6940 <_sk_callback_avx+0x218> + DB 196,98,125,24,37,38,77,0,0 ; vbroadcastss 0x4d26(%rip),%ymm12 # 6778 <_sk_callback_avx+0x218> DB 196,193,116,89,212 ; vmulps %ymm12,%ymm1,%ymm2 DB 197,28,194,226,1 ; vcmpltps %ymm2,%ymm12,%ymm12 DB 196,65,36,92,216 ; vsubps %ymm8,%ymm11,%ymm11 @@ -6808,7 +6756,7 @@ _sk_rgb_to_hsl_avx LABEL PROC DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1 DB 196,195,125,74,198,128 ; vblendvps %ymm8,%ymm14,%ymm0,%ymm0 DB 196,195,117,74,206,128 ; vblendvps %ymm8,%ymm14,%ymm1,%ymm1 - DB 196,98,125,24,5,252,77,0,0 ; vbroadcastss 0x4dfc(%rip),%ymm8 # 693c <_sk_callback_avx+0x214> + DB 196,98,125,24,5,233,76,0,0 ; vbroadcastss 0x4ce9(%rip),%ymm8 # 6774 <_sk_callback_avx+0x214> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -6823,7 +6771,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,252,17,28,36 ; vmovups %ymm3,(%rsp) DB 197,252,40,225 ; vmovaps %ymm1,%ymm4 DB 197,252,40,216 ; vmovaps %ymm0,%ymm3 - DB 196,98,125,24,5,195,77,0,0 ; vbroadcastss 0x4dc3(%rip),%ymm8 # 6944 <_sk_callback_avx+0x21c> + DB 196,98,125,24,5,176,76,0,0 ; vbroadcastss 0x4cb0(%rip),%ymm8 # 677c <_sk_callback_avx+0x21c> DB 197,60,194,202,2 ; vcmpleps %ymm2,%ymm8,%ymm9 DB 197,92,89,210 ; vmulps %ymm2,%ymm4,%ymm10 DB 196,65,92,92,218 ; vsubps %ymm10,%ymm4,%ymm11 @@ -6831,23 +6779,23 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,52,88,210 ; vaddps %ymm2,%ymm9,%ymm10 DB 197,108,88,202 ; vaddps %ymm2,%ymm2,%ymm9 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,29,157,77,0,0 ; vbroadcastss 0x4d9d(%rip),%ymm11 # 6948 <_sk_callback_avx+0x220> + DB 196,98,125,24,29,138,76,0,0 ; vbroadcastss 0x4c8a(%rip),%ymm11 # 6780 <_sk_callback_avx+0x220> DB 196,65,100,88,219 ; vaddps %ymm11,%ymm3,%ymm11 DB 196,67,125,8,227,1 ; vroundps $0x1,%ymm11,%ymm12 DB 196,65,36,92,252 ; vsubps %ymm12,%ymm11,%ymm15 DB 196,65,44,92,217 ; vsubps %ymm9,%ymm10,%ymm11 - DB 196,98,125,24,37,135,77,0,0 ; vbroadcastss 0x4d87(%rip),%ymm12 # 6950 <_sk_callback_avx+0x228> + DB 196,98,125,24,37,116,76,0,0 ; vbroadcastss 0x4c74(%rip),%ymm12 # 6788 <_sk_callback_avx+0x228> DB 196,193,4,89,196 ; vmulps %ymm12,%ymm15,%ymm0 - DB 196,98,125,24,45,125,77,0,0 ; vbroadcastss 0x4d7d(%rip),%ymm13 # 6954 <_sk_callback_avx+0x22c> + DB 196,98,125,24,45,106,76,0,0 ; vbroadcastss 0x4c6a(%rip),%ymm13 # 678c <_sk_callback_avx+0x22c> DB 197,20,92,240 ; vsubps %ymm0,%ymm13,%ymm14 DB 196,65,36,89,246 ; vmulps %ymm14,%ymm11,%ymm14 DB 196,65,52,88,246 ; vaddps %ymm14,%ymm9,%ymm14 - DB 196,226,125,24,13,94,77,0,0 ; vbroadcastss 0x4d5e(%rip),%ymm1 # 694c <_sk_callback_avx+0x224> + DB 196,226,125,24,13,75,76,0,0 ; vbroadcastss 0x4c4b(%rip),%ymm1 # 6784 <_sk_callback_avx+0x224> DB 196,193,116,194,255,2 ; vcmpleps %ymm15,%ymm1,%ymm7 DB 196,195,13,74,249,112 ; vblendvps %ymm7,%ymm9,%ymm14,%ymm7 DB 196,65,60,194,247,2 ; vcmpleps %ymm15,%ymm8,%ymm14 DB 196,227,45,74,255,224 ; vblendvps %ymm14,%ymm7,%ymm10,%ymm7 - DB 196,98,125,24,53,73,77,0,0 ; vbroadcastss 0x4d49(%rip),%ymm14 # 6958 <_sk_callback_avx+0x230> + DB 196,98,125,24,53,54,76,0,0 ; vbroadcastss 0x4c36(%rip),%ymm14 # 6790 <_sk_callback_avx+0x230> DB 196,65,12,194,255,2 ; vcmpleps %ymm15,%ymm14,%ymm15 DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,180,88,192 ; vaddps %ymm0,%ymm9,%ymm0 @@ -6866,7 +6814,7 @@ _sk_hsl_to_rgb_avx LABEL PROC DB 197,164,89,247 ; vmulps %ymm7,%ymm11,%ymm6 DB 197,180,88,246 ; vaddps %ymm6,%ymm9,%ymm6 DB 196,227,77,74,237,0 ; vblendvps %ymm0,%ymm5,%ymm6,%ymm5 - DB 196,226,125,24,5,235,76,0,0 ; vbroadcastss 0x4ceb(%rip),%ymm0 # 695c <_sk_callback_avx+0x234> + DB 196,226,125,24,5,216,75,0,0 ; vbroadcastss 0x4bd8(%rip),%ymm0 # 6794 <_sk_callback_avx+0x234> DB 197,228,88,192 ; vaddps %ymm0,%ymm3,%ymm0 DB 196,227,125,8,216,1 ; vroundps $0x1,%ymm0,%ymm3 DB 197,252,92,195 ; vsubps %ymm3,%ymm0,%ymm0 @@ -6914,14 +6862,14 @@ _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 117,68 ; jne 1d6e <_sk_scale_u8_avx+0x54> + DB 117,68 ; jne 1cb9 <_sk_scale_u8_avx+0x54> 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 DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8 DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,14,76,0,0 ; vbroadcastss 0x4c0e(%rip),%ymm9 # 6960 <_sk_callback_avx+0x238> + DB 196,98,125,24,13,251,74,0,0 ; vbroadcastss 0x4afb(%rip),%ymm9 # 6798 <_sk_callback_avx+0x238> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 @@ -6939,9 +6887,9 @@ _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 1d76 <_sk_scale_u8_avx+0x5c> + DB 117,234 ; jne 1cc1 <_sk_scale_u8_avx+0x5c> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 235,155 ; jmp 1d2e <_sk_scale_u8_avx+0x14> + DB 235,155 ; jmp 1c79 <_sk_scale_u8_avx+0x14> PUBLIC _sk_lerp_1_float_avx _sk_lerp_1_float_avx LABEL PROC @@ -6969,14 +6917,14 @@ _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 117,104 ; jne 1e4a <_sk_lerp_u8_avx+0x78> + DB 117,104 ; jne 1d95 <_sk_lerp_u8_avx+0x78> 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 DB 196,66,121,49,192 ; vpmovzxbd %xmm8,%xmm8 DB 196,67,53,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,13,90,75,0,0 ; vbroadcastss 0x4b5a(%rip),%ymm9 # 6964 <_sk_callback_avx+0x23c> + DB 196,98,125,24,13,71,74,0,0 ; vbroadcastss 0x4a47(%rip),%ymm9 # 679c <_sk_callback_avx+0x23c> DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 @@ -7002,35 +6950,35 @@ _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 1e52 <_sk_lerp_u8_avx+0x80> + DB 117,234 ; jne 1d9d <_sk_lerp_u8_avx+0x80> DB 196,65,249,110,193 ; vmovq %r9,%xmm8 - DB 233,116,255,255,255 ; jmpq 1de6 <_sk_lerp_u8_avx+0x14> + DB 233,116,255,255,255 ; jmpq 1d31 <_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 15,133,208,0,0,0 ; jne 1f50 <_sk_lerp_565_avx+0xde> + DB 15,133,208,0,0,0 ; jne 1e9b <_sk_lerp_565_avx+0xde> DB 196,65,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,65,57,105,201 ; vpunpckhwd %xmm9,%xmm8,%xmm9 DB 196,66,121,51,192 ; vpmovzxwd %xmm8,%xmm8 DB 196,67,61,24,193,1 ; vinsertf128 $0x1,%xmm9,%ymm8,%ymm8 - DB 196,98,125,24,13,196,74,0,0 ; vbroadcastss 0x4ac4(%rip),%ymm9 # 6968 <_sk_callback_avx+0x240> + DB 196,98,125,24,13,177,73,0,0 ; vbroadcastss 0x49b1(%rip),%ymm9 # 67a0 <_sk_callback_avx+0x240> DB 196,65,60,84,201 ; vandps %ymm9,%ymm8,%ymm9 DB 196,65,124,91,201 ; vcvtdq2ps %ymm9,%ymm9 - DB 196,98,125,24,21,181,74,0,0 ; vbroadcastss 0x4ab5(%rip),%ymm10 # 696c <_sk_callback_avx+0x244> + DB 196,98,125,24,21,162,73,0,0 ; vbroadcastss 0x49a2(%rip),%ymm10 # 67a4 <_sk_callback_avx+0x244> DB 196,65,52,89,202 ; vmulps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,171,74,0,0 ; vbroadcastss 0x4aab(%rip),%ymm10 # 6970 <_sk_callback_avx+0x248> + DB 196,98,125,24,21,152,73,0,0 ; vbroadcastss 0x4998(%rip),%ymm10 # 67a8 <_sk_callback_avx+0x248> DB 196,65,60,84,210 ; vandps %ymm10,%ymm8,%ymm10 DB 196,65,124,91,210 ; vcvtdq2ps %ymm10,%ymm10 - DB 196,98,125,24,29,156,74,0,0 ; vbroadcastss 0x4a9c(%rip),%ymm11 # 6974 <_sk_callback_avx+0x24c> + DB 196,98,125,24,29,137,73,0,0 ; vbroadcastss 0x4989(%rip),%ymm11 # 67ac <_sk_callback_avx+0x24c> DB 196,65,44,89,211 ; vmulps %ymm11,%ymm10,%ymm10 - DB 196,98,125,24,29,146,74,0,0 ; vbroadcastss 0x4a92(%rip),%ymm11 # 6978 <_sk_callback_avx+0x250> + DB 196,98,125,24,29,127,73,0,0 ; vbroadcastss 0x497f(%rip),%ymm11 # 67b0 <_sk_callback_avx+0x250> DB 196,65,60,84,195 ; vandps %ymm11,%ymm8,%ymm8 DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8 - DB 196,98,125,24,29,131,74,0,0 ; vbroadcastss 0x4a83(%rip),%ymm11 # 697c <_sk_callback_avx+0x254> + DB 196,98,125,24,29,112,73,0,0 ; vbroadcastss 0x4970(%rip),%ymm11 # 67b4 <_sk_callback_avx+0x254> DB 196,65,60,89,195 ; vmulps %ymm11,%ymm8,%ymm8 DB 197,252,92,196 ; vsubps %ymm4,%ymm0,%ymm0 DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 @@ -7057,9 +7005,9 @@ _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 15,135,29,255,255,255 ; ja 1e86 <_sk_lerp_565_avx+0x14> + DB 15,135,29,255,255,255 ; ja 1dd1 <_sk_lerp_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1fc0 <_sk_lerp_565_avx+0x14e> + DB 76,141,13,77,0,0,0 ; lea 0x4d(%rip),%r9 # 1f0c <_sk_lerp_565_avx+0x14f> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -7071,47 +7019,48 @@ _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 233,200,254,255,255 ; jmpq 1e86 <_sk_lerp_565_avx+0x14> - DB 102,144 ; xchg %ax,%ax - DB 242,255 ; repnz (bad) + DB 233,200,254,255,255 ; jmpq 1dd1 <_sk_lerp_565_avx+0x14> + DB 15,31,0 ; nopl (%rax) + DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) - DB 234 ; (bad) DB 255 ; (bad) - DB 255 ; (bad) - DB 255,226 ; jmpq *%rdx + DB 233,255,255,255,225 ; jmpq ffffffffe2001f14 <_sk_callback_avx+0xffffffffe1ffb9b4> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 218,255 ; (bad) + DB 217,255 ; fcos DB 255 ; (bad) - DB 255,210 ; callq *%rdx + DB 255,209 ; callq *%rcx DB 255 ; (bad) DB 255 ; (bad) - DB 255,202 ; dec %edx + DB 255,201 ; dec %ecx DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 189 ; .byte 0xbd + DB 188 ; .byte 0xbc DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff PUBLIC _sk_load_tables_avx _sk_load_tables_avx LABEL PROC + DB 73,137,200 ; mov %rcx,%r8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,0 ; mov (%rax),%r8 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,26,2,0,0 ; jne 2204 <_sk_load_tables_avx+0x228> - DB 196,65,124,16,4,184 ; vmovups (%r8,%rdi,4),%ymm8 + 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 15,133,31,2,0,0 ; jne 2160 <_sk_load_tables_avx+0x238> + DB 196,65,124,16,17 ; vmovups (%r9),%ymm10 DB 85 ; push %rbp DB 65,87 ; push %r15 DB 65,86 ; push %r14 DB 65,85 ; push %r13 DB 65,84 ; push %r12 DB 83 ; push %rbx - DB 197,124,40,13,158,76,0,0 ; vmovaps 0x4c9e(%rip),%ymm9 # 6ca0 <_sk_callback_avx+0x578> - DB 196,193,60,84,193 ; vandps %ymm9,%ymm8,%ymm0 + DB 80 ; push %rax + DB 197,124,40,13,39,76,0,0 ; vmovaps 0x4c27(%rip),%ymm9 # 6b80 <_sk_callback_avx+0x620> + DB 196,193,44,84,193 ; vandps %ymm9,%ymm10,%ymm0 DB 196,193,249,126,193 ; vmovq %xmm0,%r9 DB 69,137,203 ; mov %r9d,%r11d DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 @@ -7119,26 +7068,26 @@ _sk_load_tables_avx LABEL PROC DB 73,193,234,32 ; shr $0x20,%r10 DB 73,193,233,32 ; shr $0x20,%r9 DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0 - DB 196,193,249,126,196 ; vmovq %xmm0,%r12 - DB 69,137,231 ; mov %r12d,%r15d - DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx - DB 65,137,221 ; mov %ebx,%r13d + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,223 ; mov %ebx,%r15d + DB 196,227,249,22,193,1 ; vpextrq $0x1,%xmm0,%rcx + DB 65,137,205 ; mov %ecx,%r13d + DB 72,193,233,32 ; shr $0x20,%rcx DB 72,193,235,32 ; shr $0x20,%rbx - DB 73,193,236,32 ; shr $0x20,%r12 DB 72,139,104,8 ; mov 0x8(%rax),%rbp - DB 76,139,64,16 ; mov 0x10(%rax),%r8 + DB 76,139,96,16 ; mov 0x10(%rax),%r12 DB 196,161,122,16,68,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm0 - DB 196,163,121,33,68,165,0,16 ; vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0 + DB 196,227,121,33,68,157,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,4),%xmm0,%xmm0 DB 196,163,121,33,68,173,0,32 ; vinsertps $0x20,0x0(%rbp,%r13,4),%xmm0,%xmm0 - DB 196,227,121,33,68,157,0,48 ; vinsertps $0x30,0x0(%rbp,%rbx,4),%xmm0,%xmm0 + DB 196,227,121,33,68,141,0,48 ; vinsertps $0x30,0x0(%rbp,%rcx,4),%xmm0,%xmm0 DB 196,161,122,16,76,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm1 DB 196,163,113,33,76,141,0,16 ; vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 DB 196,163,113,33,76,181,0,32 ; vinsertps $0x20,0x0(%rbp,%r14,4),%xmm1,%xmm1 DB 196,163,113,33,76,149,0,48 ; vinsertps $0x30,0x0(%rbp,%r10,4),%xmm1,%xmm1 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 - DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 - DB 196,67,125,25,194,1 ; vextractf128 $0x1,%ymm8,%xmm10 - DB 196,193,105,114,210,8 ; vpsrld $0x8,%xmm10,%xmm2 + DB 196,193,113,114,210,8 ; vpsrld $0x8,%xmm10,%xmm1 + DB 196,67,125,25,208,1 ; vextractf128 $0x1,%ymm10,%xmm8 + DB 196,193,105,114,208,8 ; vpsrld $0x8,%xmm8,%xmm2 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 196,193,116,84,201 ; vandps %ymm9,%ymm1,%ymm1 DB 196,193,249,126,201 ; vmovq %xmm1,%r9 @@ -7148,36 +7097,36 @@ _sk_load_tables_avx LABEL PROC DB 73,193,234,32 ; shr $0x20,%r10 DB 73,193,233,32 ; shr $0x20,%r9 DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1 - DB 196,225,249,126,205 ; vmovq %xmm1,%rbp - DB 65,137,239 ; mov %ebp,%r15d - DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx - DB 65,137,220 ; mov %ebx,%r12d - DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,225,249,126,203 ; vmovq %xmm1,%rbx + DB 65,137,223 ; mov %ebx,%r15d + DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp + DB 137,233 ; mov %ebp,%ecx DB 72,193,237,32 ; shr $0x20,%rbp - DB 196,129,122,16,12,184 ; vmovss (%r8,%r15,4),%xmm1 - DB 196,195,113,33,12,168,16 ; vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1 - DB 196,129,122,16,20,160 ; vmovss (%r8,%r12,4),%xmm2 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,129,122,16,12,188 ; vmovss (%r12,%r15,4),%xmm1 + DB 196,195,113,33,12,156,16 ; vinsertps $0x10,(%r12,%rbx,4),%xmm1,%xmm1 + DB 196,193,122,16,20,140 ; vmovss (%r12,%rcx,4),%xmm2 DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1 - DB 196,193,122,16,20,152 ; vmovss (%r8,%rbx,4),%xmm2 + DB 196,193,122,16,20,172 ; vmovss (%r12,%rbp,4),%xmm2 DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1 - DB 196,129,122,16,20,152 ; vmovss (%r8,%r11,4),%xmm2 - DB 196,131,105,33,20,136,16 ; vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2 - DB 196,129,122,16,28,176 ; vmovss (%r8,%r14,4),%xmm3 + DB 196,129,122,16,20,156 ; vmovss (%r12,%r11,4),%xmm2 + DB 196,131,105,33,20,140,16 ; vinsertps $0x10,(%r12,%r9,4),%xmm2,%xmm2 + DB 196,129,122,16,28,180 ; vmovss (%r12,%r14,4),%xmm3 DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2 - DB 196,129,122,16,28,144 ; vmovss (%r8,%r10,4),%xmm3 + DB 196,129,122,16,28,148 ; vmovss (%r12,%r10,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 DB 72,139,64,24 ; mov 0x18(%rax),%rax - DB 196,193,105,114,208,16 ; vpsrld $0x10,%xmm8,%xmm2 - DB 196,193,97,114,210,16 ; vpsrld $0x10,%xmm10,%xmm3 + DB 196,193,105,114,210,16 ; vpsrld $0x10,%xmm10,%xmm2 + DB 196,193,97,114,208,16 ; vpsrld $0x10,%xmm8,%xmm3 DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 DB 196,193,108,84,209 ; vandps %ymm9,%ymm2,%ymm2 - DB 196,193,249,126,208 ; vmovq %xmm2,%r8 - DB 69,137,194 ; mov %r8d,%r10d - DB 196,195,249,22,209,1 ; vpextrq $0x1,%xmm2,%r9 - DB 69,137,203 ; mov %r9d,%r11d + DB 196,193,249,126,209 ; vmovq %xmm2,%r9 + DB 69,137,202 ; mov %r9d,%r10d + DB 196,227,249,22,209,1 ; vpextrq $0x1,%xmm2,%rcx + DB 65,137,203 ; mov %ecx,%r11d + DB 72,193,233,32 ; shr $0x20,%rcx DB 73,193,233,32 ; shr $0x20,%r9 - DB 73,193,232,32 ; shr $0x20,%r8 DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 DB 196,225,249,126,213 ; vmovq %xmm2,%rbp DB 65,137,238 ; mov %ebp,%r14d @@ -7192,19 +7141,21 @@ _sk_load_tables_avx LABEL PROC DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3 DB 196,99,105,33,203,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm9 DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3 - DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 + DB 196,163,97,33,28,136,16 ; vinsertps $0x10,(%rax,%r9,4),%xmm3,%xmm3 DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2 DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2 - DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3 + DB 197,250,16,28,136 ; vmovss (%rax,%rcx,4),%xmm3 DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,195,109,24,209,1 ; vinsertf128 $0x1,%xmm9,%ymm2,%ymm2 - DB 196,193,57,114,208,24 ; vpsrld $0x18,%xmm8,%xmm8 - DB 196,193,97,114,210,24 ; vpsrld $0x18,%xmm10,%xmm3 - DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 + DB 196,193,49,114,210,24 ; vpsrld $0x18,%xmm10,%xmm9 + DB 196,193,97,114,208,24 ; vpsrld $0x18,%xmm8,%xmm3 + DB 196,227,53,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,143,71,0,0 ; vbroadcastss 0x478f(%rip),%ymm8 # 6980 <_sk_callback_avx+0x258> + DB 196,98,125,24,5,114,70,0,0 ; vbroadcastss 0x4672(%rip),%ymm8 # 67b8 <_sk_callback_avx+0x258> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,137,193 ; mov %r8,%rcx + DB 72,131,196,8 ; add $0x8,%rsp DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,93 ; pop %r13 @@ -7212,57 +7163,20 @@ _sk_load_tables_avx LABEL PROC DB 65,95 ; pop %r15 DB 93 ; pop %rbp DB 255,224 ; jmpq *%rax - DB 65,137,201 ; mov %ecx,%r9d - DB 65,128,225,7 ; and $0x7,%r9b - DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 - DB 65,254,201 ; dec %r9b - DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,211,253,255,255 ; ja 1ff0 <_sk_load_tables_avx+0x14> - DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,140,0,0,0 ; lea 0x8c(%rip),%r10 # 22b4 <_sk_load_tables_avx+0x2d8> - DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9 - DB 77,1,209 ; add %r10,%r9 - DB 65,255,225 ; jmpq *%r9 - DB 196,193,121,110,68,184,24 ; vmovd 0x18(%r8,%rdi,4),%xmm0 - DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 - DB 196,99,117,12,192,64 ; vblendps $0x40,%ymm0,%ymm1,%ymm8 - DB 196,99,125,25,192,1 ; vextractf128 $0x1,%ymm8,%xmm0 - DB 196,195,121,34,68,184,20,1 ; vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0 - DB 196,99,61,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 - DB 196,99,125,25,192,1 ; vextractf128 $0x1,%ymm8,%xmm0 - DB 196,195,121,34,68,184,16,0 ; vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0 - DB 196,99,61,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 - DB 196,195,57,34,68,184,12,3 ; vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm8,%xmm0 - DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8 - DB 196,195,57,34,68,184,8,2 ; vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm8,%xmm0 - DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8 - DB 196,195,57,34,68,184,4,1 ; vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm8,%xmm0 - 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 233,62,253,255,255 ; jmpq 1ff0 <_sk_load_tables_avx+0x14> - DB 102,144 ; xchg %ax,%ax - DB 236 ; in (%dx),%al - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) - DB 255 ; (bad) - DB 255,208 ; callq *%rax - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,194 ; inc %edx - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,174,255,255,255,154 ; ljmp *-0x65000001(%rsi) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 126,255 ; jle 22cd <_sk_load_tables_avx+0x2f1> - DB 255 ; (bad) - DB 255 ; .byte 0xff + DB 185,8,0,0,0 ; mov $0x8,%ecx + DB 68,41,193 ; sub %r8d,%ecx + DB 192,225,3 ; shl $0x3,%cl + DB 73,199,194,255,255,255,255 ; mov $0xffffffffffffffff,%r10 + DB 73,211,234 ; shr %cl,%r10 + DB 196,193,249,110,194 ; vmovq %r10,%xmm0 + DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0 + DB 196,226,121,0,13,72,73,0,0 ; vpshufb 0x4948(%rip),%xmm0,%xmm1 # 6ad0 <_sk_callback_avx+0x570> + DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1 + DB 196,226,121,0,5,74,73,0,0 ; vpshufb 0x494a(%rip),%xmm0,%xmm0 # 6ae0 <_sk_callback_avx+0x580> + DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 196,66,125,44,17 ; vmaskmovps (%r9),%ymm0,%ymm10 + DB 233,155,253,255,255 ; jmpq 1f46 <_sk_load_tables_avx+0x1e> PUBLIC _sk_load_tables_u16_be_avx _sk_load_tables_u16_be_avx LABEL PROC @@ -7270,7 +7184,7 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,113,2,0,0 ; jne 2557 <_sk_load_tables_u16_be_avx+0x287> + DB 15,133,113,2,0,0 ; jne 2432 <_sk_load_tables_u16_be_avx+0x287> DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8 DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2 DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3 @@ -7292,7 +7206,7 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 197,177,108,208 ; vpunpcklqdq %xmm0,%xmm9,%xmm2 DB 197,177,109,200 ; vpunpckhqdq %xmm0,%xmm9,%xmm1 DB 196,65,57,108,212 ; vpunpcklqdq %xmm12,%xmm8,%xmm10 - DB 197,121,111,29,222,73,0,0 ; vmovdqa 0x49de(%rip),%xmm11 # 6d20 <_sk_callback_avx+0x5f8> + DB 197,121,111,29,211,72,0,0 ; vmovdqa 0x48d3(%rip),%xmm11 # 6af0 <_sk_callback_avx+0x590> DB 196,193,105,219,195 ; vpand %xmm11,%xmm2,%xmm0 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 196,193,121,105,209 ; vpunpckhwd %xmm9,%xmm0,%xmm2 @@ -7391,7 +7305,7 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3 DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,64,68,0,0 ; vbroadcastss 0x4440(%rip),%ymm8 # 6984 <_sk_callback_avx+0x25c> + DB 196,98,125,24,5,157,67,0,0 ; vbroadcastss 0x439d(%rip),%ymm8 # 67bc <_sk_callback_avx+0x25c> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -7404,29 +7318,29 @@ _sk_load_tables_u16_be_avx LABEL PROC DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 25bd <_sk_load_tables_u16_be_avx+0x2ed> + DB 116,85 ; je 2498 <_sk_load_tables_u16_be_avx+0x2ed> DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 25bd <_sk_load_tables_u16_be_avx+0x2ed> + DB 114,72 ; jb 2498 <_sk_load_tables_u16_be_avx+0x2ed> DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 25ca <_sk_load_tables_u16_be_avx+0x2fa> + DB 116,72 ; je 24a5 <_sk_load_tables_u16_be_avx+0x2fa> DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 25ca <_sk_load_tables_u16_be_avx+0x2fa> + DB 114,59 ; jb 24a5 <_sk_load_tables_u16_be_avx+0x2fa> DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,97,253,255,255 ; je 2301 <_sk_load_tables_u16_be_avx+0x31> + DB 15,132,97,253,255,255 ; je 21dc <_sk_load_tables_u16_be_avx+0x31> DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,80,253,255,255 ; jb 2301 <_sk_load_tables_u16_be_avx+0x31> + DB 15,130,80,253,255,255 ; jb 21dc <_sk_load_tables_u16_be_avx+0x31> DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9 - DB 233,68,253,255,255 ; jmpq 2301 <_sk_load_tables_u16_be_avx+0x31> + DB 233,68,253,255,255 ; jmpq 21dc <_sk_load_tables_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,55,253,255,255 ; jmpq 2301 <_sk_load_tables_u16_be_avx+0x31> + DB 233,55,253,255,255 ; jmpq 21dc <_sk_load_tables_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,46,253,255,255 ; jmpq 2301 <_sk_load_tables_u16_be_avx+0x31> + DB 233,46,253,255,255 ; jmpq 21dc <_sk_load_tables_u16_be_avx+0x31> PUBLIC _sk_load_tables_rgb_u16_be_avx _sk_load_tables_rgb_u16_be_avx LABEL PROC @@ -7434,7 +7348,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,93,2,0,0 ; jne 2842 <_sk_load_tables_rgb_u16_be_avx+0x26f> + DB 15,133,93,2,0,0 ; jne 271d <_sk_load_tables_rgb_u16_be_avx+0x26f> DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0 DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2 DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1 @@ -7461,7 +7375,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 197,185,108,202 ; vpunpcklqdq %xmm2,%xmm8,%xmm1 DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2 DB 197,121,108,195 ; vpunpcklqdq %xmm3,%xmm0,%xmm8 - DB 197,121,111,13,215,70,0,0 ; vmovdqa 0x46d7(%rip),%xmm9 # 6d30 <_sk_callback_avx+0x608> + DB 197,121,111,13,204,69,0,0 ; vmovdqa 0x45cc(%rip),%xmm9 # 6b00 <_sk_callback_avx+0x5a0> DB 196,193,113,219,193 ; vpand %xmm9,%xmm1,%xmm0 DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10 DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1 @@ -7553,7 +7467,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2 DB 196,195,109,24,208,1 ; vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,82,65,0,0 ; vbroadcastss 0x4152(%rip),%ymm3 # 6988 <_sk_callback_avx+0x260> + DB 196,226,125,24,29,175,64,0,0 ; vbroadcastss 0x40af(%rip),%ymm3 # 67c0 <_sk_callback_avx+0x260> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,93 ; pop %r13 @@ -7564,36 +7478,36 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0 DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 285b <_sk_load_tables_rgb_u16_be_avx+0x288> - DB 233,190,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 117,5 ; jne 2736 <_sk_load_tables_rgb_u16_be_avx+0x288> + DB 233,190,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1 DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 288a <_sk_load_tables_rgb_u16_be_avx+0x2b7> + DB 114,26 ; jb 2765 <_sk_load_tables_rgb_u16_be_avx+0x2b7> DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1 DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 288f <_sk_load_tables_rgb_u16_be_avx+0x2bc> - DB 233,143,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,138,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 276a <_sk_load_tables_rgb_u16_be_avx+0x2bc> + DB 233,143,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,138,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1 DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 28be <_sk_load_tables_rgb_u16_be_avx+0x2eb> + DB 114,26 ; jb 2799 <_sk_load_tables_rgb_u16_be_avx+0x2eb> DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1 DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 28c3 <_sk_load_tables_rgb_u16_be_avx+0x2f0> - DB 233,91,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,86,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 279e <_sk_load_tables_rgb_u16_be_avx+0x2f0> + DB 233,91,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,86,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3 DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 28ec <_sk_load_tables_rgb_u16_be_avx+0x319> + DB 114,20 ; jb 27c7 <_sk_load_tables_rgb_u16_be_avx+0x319> DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3 DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 - DB 233,45,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> - DB 233,40,253,255,255 ; jmpq 2619 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,45,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> + DB 233,40,253,255,255 ; jmpq 24f4 <_sk_load_tables_rgb_u16_be_avx+0x46> PUBLIC _sk_byte_tables_avx _sk_byte_tables_avx LABEL PROC @@ -7604,7 +7518,7 @@ _sk_byte_tables_avx LABEL PROC DB 65,84 ; push %r12 DB 83 ; push %rbx DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,134,64,0,0 ; vbroadcastss 0x4086(%rip),%ymm8 # 698c <_sk_callback_avx+0x264> + DB 196,98,125,24,5,227,63,0,0 ; vbroadcastss 0x3fe3(%rip),%ymm8 # 67c4 <_sk_callback_avx+0x264> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8 @@ -7641,7 +7555,7 @@ _sk_byte_tables_avx LABEL PROC DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,212,63,0,0 ; vbroadcastss 0x3fd4(%rip),%ymm9 # 6990 <_sk_callback_avx+0x268> + DB 196,98,125,24,13,49,63,0,0 ; vbroadcastss 0x3f31(%rip),%ymm9 # 67c8 <_sk_callback_avx+0x268> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 @@ -7801,7 +7715,7 @@ _sk_byte_tables_rgb_avx LABEL PROC DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,250,60,0,0 ; vbroadcastss 0x3cfa(%rip),%ymm9 # 6994 <_sk_callback_avx+0x26c> + DB 196,98,125,24,13,87,60,0,0 ; vbroadcastss 0x3c57(%rip),%ymm9 # 67cc <_sk_callback_avx+0x26c> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 @@ -8088,36 +8002,36 @@ _sk_parametric_r_avx LABEL PROC DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,216 ; vcvtdq2ps %ymm0,%ymm11 - DB 196,98,125,24,37,88,56,0,0 ; vbroadcastss 0x3858(%rip),%ymm12 # 6998 <_sk_callback_avx+0x270> + DB 196,98,125,24,37,181,55,0,0 ; vbroadcastss 0x37b5(%rip),%ymm12 # 67d0 <_sk_callback_avx+0x270> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,78,56,0,0 ; vbroadcastss 0x384e(%rip),%ymm12 # 699c <_sk_callback_avx+0x274> + DB 196,98,125,24,37,171,55,0,0 ; vbroadcastss 0x37ab(%rip),%ymm12 # 67d4 <_sk_callback_avx+0x274> DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,68,56,0,0 ; vbroadcastss 0x3844(%rip),%ymm12 # 69a0 <_sk_callback_avx+0x278> + DB 196,98,125,24,37,161,55,0,0 ; vbroadcastss 0x37a1(%rip),%ymm12 # 67d8 <_sk_callback_avx+0x278> DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,58,56,0,0 ; vbroadcastss 0x383a(%rip),%ymm12 # 69a4 <_sk_callback_avx+0x27c> + DB 196,98,125,24,37,151,55,0,0 ; vbroadcastss 0x3797(%rip),%ymm12 # 67dc <_sk_callback_avx+0x27c> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,48,56,0,0 ; vbroadcastss 0x3830(%rip),%ymm12 # 69a8 <_sk_callback_avx+0x280> + DB 196,98,125,24,37,141,55,0,0 ; vbroadcastss 0x378d(%rip),%ymm12 # 67e0 <_sk_callback_avx+0x280> DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,33,56,0,0 ; vbroadcastss 0x3821(%rip),%ymm12 # 69ac <_sk_callback_avx+0x284> + DB 196,98,125,24,37,126,55,0,0 ; vbroadcastss 0x377e(%rip),%ymm12 # 67e4 <_sk_callback_avx+0x284> DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0 - DB 196,98,125,24,37,23,56,0,0 ; vbroadcastss 0x3817(%rip),%ymm12 # 69b0 <_sk_callback_avx+0x288> + DB 196,98,125,24,37,116,55,0,0 ; vbroadcastss 0x3774(%rip),%ymm12 # 67e8 <_sk_callback_avx+0x288> DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0 DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0 DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0 DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10 DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10 - DB 196,98,125,24,29,251,55,0,0 ; vbroadcastss 0x37fb(%rip),%ymm11 # 69b4 <_sk_callback_avx+0x28c> + DB 196,98,125,24,29,88,55,0,0 ; vbroadcastss 0x3758(%rip),%ymm11 # 67ec <_sk_callback_avx+0x28c> DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,241,55,0,0 ; vbroadcastss 0x37f1(%rip),%ymm11 # 69b8 <_sk_callback_avx+0x290> + DB 196,98,125,24,29,78,55,0,0 ; vbroadcastss 0x374e(%rip),%ymm11 # 67f0 <_sk_callback_avx+0x290> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0 - DB 196,98,125,24,29,226,55,0,0 ; vbroadcastss 0x37e2(%rip),%ymm11 # 69bc <_sk_callback_avx+0x294> + DB 196,98,125,24,29,63,55,0,0 ; vbroadcastss 0x373f(%rip),%ymm11 # 67f4 <_sk_callback_avx+0x294> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,216,55,0,0 ; vbroadcastss 0x37d8(%rip),%ymm11 # 69c0 <_sk_callback_avx+0x298> + DB 196,98,125,24,29,53,55,0,0 ; vbroadcastss 0x3735(%rip),%ymm11 # 67f8 <_sk_callback_avx+0x298> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0 - DB 196,98,125,24,21,201,55,0,0 ; vbroadcastss 0x37c9(%rip),%ymm10 # 69c4 <_sk_callback_avx+0x29c> + DB 196,98,125,24,21,38,55,0,0 ; vbroadcastss 0x3726(%rip),%ymm10 # 67fc <_sk_callback_avx+0x29c> DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0 DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8125,7 +8039,7 @@ _sk_parametric_r_avx LABEL PROC DB 196,195,125,74,193,128 ; vblendvps %ymm8,%ymm9,%ymm0,%ymm0 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,124,95,192 ; vmaxps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,160,55,0,0 ; vbroadcastss 0x37a0(%rip),%ymm8 # 69c8 <_sk_callback_avx+0x2a0> + DB 196,98,125,24,5,253,54,0,0 ; vbroadcastss 0x36fd(%rip),%ymm8 # 6800 <_sk_callback_avx+0x2a0> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8145,36 +8059,36 @@ _sk_parametric_g_avx LABEL PROC DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,217 ; vcvtdq2ps %ymm1,%ymm11 - DB 196,98,125,24,37,81,55,0,0 ; vbroadcastss 0x3751(%rip),%ymm12 # 69cc <_sk_callback_avx+0x2a4> + DB 196,98,125,24,37,174,54,0,0 ; vbroadcastss 0x36ae(%rip),%ymm12 # 6804 <_sk_callback_avx+0x2a4> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,71,55,0,0 ; vbroadcastss 0x3747(%rip),%ymm12 # 69d0 <_sk_callback_avx+0x2a8> + DB 196,98,125,24,37,164,54,0,0 ; vbroadcastss 0x36a4(%rip),%ymm12 # 6808 <_sk_callback_avx+0x2a8> DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,61,55,0,0 ; vbroadcastss 0x373d(%rip),%ymm12 # 69d4 <_sk_callback_avx+0x2ac> + DB 196,98,125,24,37,154,54,0,0 ; vbroadcastss 0x369a(%rip),%ymm12 # 680c <_sk_callback_avx+0x2ac> DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,51,55,0,0 ; vbroadcastss 0x3733(%rip),%ymm12 # 69d8 <_sk_callback_avx+0x2b0> + DB 196,98,125,24,37,144,54,0,0 ; vbroadcastss 0x3690(%rip),%ymm12 # 6810 <_sk_callback_avx+0x2b0> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,41,55,0,0 ; vbroadcastss 0x3729(%rip),%ymm12 # 69dc <_sk_callback_avx+0x2b4> + DB 196,98,125,24,37,134,54,0,0 ; vbroadcastss 0x3686(%rip),%ymm12 # 6814 <_sk_callback_avx+0x2b4> DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,26,55,0,0 ; vbroadcastss 0x371a(%rip),%ymm12 # 69e0 <_sk_callback_avx+0x2b8> + DB 196,98,125,24,37,119,54,0,0 ; vbroadcastss 0x3677(%rip),%ymm12 # 6818 <_sk_callback_avx+0x2b8> DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1 - DB 196,98,125,24,37,16,55,0,0 ; vbroadcastss 0x3710(%rip),%ymm12 # 69e4 <_sk_callback_avx+0x2bc> + DB 196,98,125,24,37,109,54,0,0 ; vbroadcastss 0x366d(%rip),%ymm12 # 681c <_sk_callback_avx+0x2bc> DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1 DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1 DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1 DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10 DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10 - DB 196,98,125,24,29,244,54,0,0 ; vbroadcastss 0x36f4(%rip),%ymm11 # 69e8 <_sk_callback_avx+0x2c0> + DB 196,98,125,24,29,81,54,0,0 ; vbroadcastss 0x3651(%rip),%ymm11 # 6820 <_sk_callback_avx+0x2c0> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,234,54,0,0 ; vbroadcastss 0x36ea(%rip),%ymm11 # 69ec <_sk_callback_avx+0x2c4> + DB 196,98,125,24,29,71,54,0,0 ; vbroadcastss 0x3647(%rip),%ymm11 # 6824 <_sk_callback_avx+0x2c4> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,29,219,54,0,0 ; vbroadcastss 0x36db(%rip),%ymm11 # 69f0 <_sk_callback_avx+0x2c8> + DB 196,98,125,24,29,56,54,0,0 ; vbroadcastss 0x3638(%rip),%ymm11 # 6828 <_sk_callback_avx+0x2c8> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,209,54,0,0 ; vbroadcastss 0x36d1(%rip),%ymm11 # 69f4 <_sk_callback_avx+0x2cc> + DB 196,98,125,24,29,46,54,0,0 ; vbroadcastss 0x362e(%rip),%ymm11 # 682c <_sk_callback_avx+0x2cc> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1 - DB 196,98,125,24,21,194,54,0,0 ; vbroadcastss 0x36c2(%rip),%ymm10 # 69f8 <_sk_callback_avx+0x2d0> + DB 196,98,125,24,21,31,54,0,0 ; vbroadcastss 0x361f(%rip),%ymm10 # 6830 <_sk_callback_avx+0x2d0> DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1 DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8182,7 +8096,7 @@ _sk_parametric_g_avx LABEL PROC DB 196,195,117,74,201,128 ; vblendvps %ymm8,%ymm9,%ymm1,%ymm1 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,116,95,200 ; vmaxps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,5,153,54,0,0 ; vbroadcastss 0x3699(%rip),%ymm8 # 69fc <_sk_callback_avx+0x2d4> + DB 196,98,125,24,5,246,53,0,0 ; vbroadcastss 0x35f6(%rip),%ymm8 # 6834 <_sk_callback_avx+0x2d4> DB 196,193,116,93,200 ; vminps %ymm8,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8202,36 +8116,36 @@ _sk_parametric_b_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,218 ; vcvtdq2ps %ymm2,%ymm11 - DB 196,98,125,24,37,74,54,0,0 ; vbroadcastss 0x364a(%rip),%ymm12 # 6a00 <_sk_callback_avx+0x2d8> + DB 196,98,125,24,37,167,53,0,0 ; vbroadcastss 0x35a7(%rip),%ymm12 # 6838 <_sk_callback_avx+0x2d8> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,64,54,0,0 ; vbroadcastss 0x3640(%rip),%ymm12 # 6a04 <_sk_callback_avx+0x2dc> + DB 196,98,125,24,37,157,53,0,0 ; vbroadcastss 0x359d(%rip),%ymm12 # 683c <_sk_callback_avx+0x2dc> DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,54,54,0,0 ; vbroadcastss 0x3636(%rip),%ymm12 # 6a08 <_sk_callback_avx+0x2e0> + DB 196,98,125,24,37,147,53,0,0 ; vbroadcastss 0x3593(%rip),%ymm12 # 6840 <_sk_callback_avx+0x2e0> DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,44,54,0,0 ; vbroadcastss 0x362c(%rip),%ymm12 # 6a0c <_sk_callback_avx+0x2e4> + DB 196,98,125,24,37,137,53,0,0 ; vbroadcastss 0x3589(%rip),%ymm12 # 6844 <_sk_callback_avx+0x2e4> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,34,54,0,0 ; vbroadcastss 0x3622(%rip),%ymm12 # 6a10 <_sk_callback_avx+0x2e8> + DB 196,98,125,24,37,127,53,0,0 ; vbroadcastss 0x357f(%rip),%ymm12 # 6848 <_sk_callback_avx+0x2e8> DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,19,54,0,0 ; vbroadcastss 0x3613(%rip),%ymm12 # 6a14 <_sk_callback_avx+0x2ec> + DB 196,98,125,24,37,112,53,0,0 ; vbroadcastss 0x3570(%rip),%ymm12 # 684c <_sk_callback_avx+0x2ec> DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2 - DB 196,98,125,24,37,9,54,0,0 ; vbroadcastss 0x3609(%rip),%ymm12 # 6a18 <_sk_callback_avx+0x2f0> + DB 196,98,125,24,37,102,53,0,0 ; vbroadcastss 0x3566(%rip),%ymm12 # 6850 <_sk_callback_avx+0x2f0> DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2 DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2 DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2 DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10 DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10 - DB 196,98,125,24,29,237,53,0,0 ; vbroadcastss 0x35ed(%rip),%ymm11 # 6a1c <_sk_callback_avx+0x2f4> + DB 196,98,125,24,29,74,53,0,0 ; vbroadcastss 0x354a(%rip),%ymm11 # 6854 <_sk_callback_avx+0x2f4> DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,227,53,0,0 ; vbroadcastss 0x35e3(%rip),%ymm11 # 6a20 <_sk_callback_avx+0x2f8> + DB 196,98,125,24,29,64,53,0,0 ; vbroadcastss 0x3540(%rip),%ymm11 # 6858 <_sk_callback_avx+0x2f8> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2 - DB 196,98,125,24,29,212,53,0,0 ; vbroadcastss 0x35d4(%rip),%ymm11 # 6a24 <_sk_callback_avx+0x2fc> + DB 196,98,125,24,29,49,53,0,0 ; vbroadcastss 0x3531(%rip),%ymm11 # 685c <_sk_callback_avx+0x2fc> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,202,53,0,0 ; vbroadcastss 0x35ca(%rip),%ymm11 # 6a28 <_sk_callback_avx+0x300> + DB 196,98,125,24,29,39,53,0,0 ; vbroadcastss 0x3527(%rip),%ymm11 # 6860 <_sk_callback_avx+0x300> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2 - DB 196,98,125,24,21,187,53,0,0 ; vbroadcastss 0x35bb(%rip),%ymm10 # 6a2c <_sk_callback_avx+0x304> + DB 196,98,125,24,21,24,53,0,0 ; vbroadcastss 0x3518(%rip),%ymm10 # 6864 <_sk_callback_avx+0x304> DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2 DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8239,7 +8153,7 @@ _sk_parametric_b_avx LABEL PROC DB 196,195,109,74,209,128 ; vblendvps %ymm8,%ymm9,%ymm2,%ymm2 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,108,95,208 ; vmaxps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,146,53,0,0 ; vbroadcastss 0x3592(%rip),%ymm8 # 6a30 <_sk_callback_avx+0x308> + DB 196,98,125,24,5,239,52,0,0 ; vbroadcastss 0x34ef(%rip),%ymm8 # 6868 <_sk_callback_avx+0x308> DB 196,193,108,93,208 ; vminps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8259,36 +8173,36 @@ _sk_parametric_a_avx LABEL PROC DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10 DB 197,124,91,219 ; vcvtdq2ps %ymm3,%ymm11 - DB 196,98,125,24,37,67,53,0,0 ; vbroadcastss 0x3543(%rip),%ymm12 # 6a34 <_sk_callback_avx+0x30c> + DB 196,98,125,24,37,160,52,0,0 ; vbroadcastss 0x34a0(%rip),%ymm12 # 686c <_sk_callback_avx+0x30c> DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,57,53,0,0 ; vbroadcastss 0x3539(%rip),%ymm12 # 6a38 <_sk_callback_avx+0x310> + DB 196,98,125,24,37,150,52,0,0 ; vbroadcastss 0x3496(%rip),%ymm12 # 6870 <_sk_callback_avx+0x310> DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,47,53,0,0 ; vbroadcastss 0x352f(%rip),%ymm12 # 6a3c <_sk_callback_avx+0x314> + DB 196,98,125,24,37,140,52,0,0 ; vbroadcastss 0x348c(%rip),%ymm12 # 6874 <_sk_callback_avx+0x314> DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,37,53,0,0 ; vbroadcastss 0x3525(%rip),%ymm12 # 6a40 <_sk_callback_avx+0x318> + DB 196,98,125,24,37,130,52,0,0 ; vbroadcastss 0x3482(%rip),%ymm12 # 6878 <_sk_callback_avx+0x318> DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,27,53,0,0 ; vbroadcastss 0x351b(%rip),%ymm12 # 6a44 <_sk_callback_avx+0x31c> + DB 196,98,125,24,37,120,52,0,0 ; vbroadcastss 0x3478(%rip),%ymm12 # 687c <_sk_callback_avx+0x31c> DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12 DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11 - DB 196,98,125,24,37,12,53,0,0 ; vbroadcastss 0x350c(%rip),%ymm12 # 6a48 <_sk_callback_avx+0x320> + DB 196,98,125,24,37,105,52,0,0 ; vbroadcastss 0x3469(%rip),%ymm12 # 6880 <_sk_callback_avx+0x320> DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3 - DB 196,98,125,24,37,2,53,0,0 ; vbroadcastss 0x3502(%rip),%ymm12 # 6a4c <_sk_callback_avx+0x324> + DB 196,98,125,24,37,95,52,0,0 ; vbroadcastss 0x345f(%rip),%ymm12 # 6884 <_sk_callback_avx+0x324> DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3 DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3 DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3 DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10 DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10 - DB 196,98,125,24,29,230,52,0,0 ; vbroadcastss 0x34e6(%rip),%ymm11 # 6a50 <_sk_callback_avx+0x328> + DB 196,98,125,24,29,67,52,0,0 ; vbroadcastss 0x3443(%rip),%ymm11 # 6888 <_sk_callback_avx+0x328> DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,220,52,0,0 ; vbroadcastss 0x34dc(%rip),%ymm11 # 6a54 <_sk_callback_avx+0x32c> + DB 196,98,125,24,29,57,52,0,0 ; vbroadcastss 0x3439(%rip),%ymm11 # 688c <_sk_callback_avx+0x32c> DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11 DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3 - DB 196,98,125,24,29,205,52,0,0 ; vbroadcastss 0x34cd(%rip),%ymm11 # 6a58 <_sk_callback_avx+0x330> + DB 196,98,125,24,29,42,52,0,0 ; vbroadcastss 0x342a(%rip),%ymm11 # 6890 <_sk_callback_avx+0x330> DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10 - DB 196,98,125,24,29,195,52,0,0 ; vbroadcastss 0x34c3(%rip),%ymm11 # 6a5c <_sk_callback_avx+0x334> + DB 196,98,125,24,29,32,52,0,0 ; vbroadcastss 0x3420(%rip),%ymm11 # 6894 <_sk_callback_avx+0x334> DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10 DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3 - DB 196,98,125,24,21,180,52,0,0 ; vbroadcastss 0x34b4(%rip),%ymm10 # 6a60 <_sk_callback_avx+0x338> + DB 196,98,125,24,21,17,52,0,0 ; vbroadcastss 0x3411(%rip),%ymm10 # 6898 <_sk_callback_avx+0x338> DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3 DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3 DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10 @@ -8296,38 +8210,38 @@ _sk_parametric_a_avx LABEL PROC DB 196,195,101,74,217,128 ; vblendvps %ymm8,%ymm9,%ymm3,%ymm3 DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 196,193,100,95,216 ; vmaxps %ymm8,%ymm3,%ymm3 - DB 196,98,125,24,5,139,52,0,0 ; vbroadcastss 0x348b(%rip),%ymm8 # 6a64 <_sk_callback_avx+0x33c> + DB 196,98,125,24,5,232,51,0,0 ; vbroadcastss 0x33e8(%rip),%ymm8 # 689c <_sk_callback_avx+0x33c> DB 196,193,100,93,216 ; vminps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax PUBLIC _sk_lab_to_xyz_avx _sk_lab_to_xyz_avx LABEL PROC - DB 196,98,125,24,5,125,52,0,0 ; vbroadcastss 0x347d(%rip),%ymm8 # 6a68 <_sk_callback_avx+0x340> + DB 196,98,125,24,5,218,51,0,0 ; vbroadcastss 0x33da(%rip),%ymm8 # 68a0 <_sk_callback_avx+0x340> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,115,52,0,0 ; vbroadcastss 0x3473(%rip),%ymm8 # 6a6c <_sk_callback_avx+0x344> + DB 196,98,125,24,5,208,51,0,0 ; vbroadcastss 0x33d0(%rip),%ymm8 # 68a4 <_sk_callback_avx+0x344> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,98,125,24,13,105,52,0,0 ; vbroadcastss 0x3469(%rip),%ymm9 # 6a70 <_sk_callback_avx+0x348> + DB 196,98,125,24,13,198,51,0,0 ; vbroadcastss 0x33c6(%rip),%ymm9 # 68a8 <_sk_callback_avx+0x348> DB 196,193,116,88,201 ; vaddps %ymm9,%ymm1,%ymm1 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 196,193,108,88,209 ; vaddps %ymm9,%ymm2,%ymm2 - DB 196,98,125,24,5,85,52,0,0 ; vbroadcastss 0x3455(%rip),%ymm8 # 6a74 <_sk_callback_avx+0x34c> + DB 196,98,125,24,5,178,51,0,0 ; vbroadcastss 0x33b2(%rip),%ymm8 # 68ac <_sk_callback_avx+0x34c> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,75,52,0,0 ; vbroadcastss 0x344b(%rip),%ymm8 # 6a78 <_sk_callback_avx+0x350> + DB 196,98,125,24,5,168,51,0,0 ; vbroadcastss 0x33a8(%rip),%ymm8 # 68b0 <_sk_callback_avx+0x350> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,5,65,52,0,0 ; vbroadcastss 0x3441(%rip),%ymm8 # 6a7c <_sk_callback_avx+0x354> + DB 196,98,125,24,5,158,51,0,0 ; vbroadcastss 0x339e(%rip),%ymm8 # 68b4 <_sk_callback_avx+0x354> DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 DB 197,252,88,201 ; vaddps %ymm1,%ymm0,%ymm1 - DB 196,98,125,24,5,51,52,0,0 ; vbroadcastss 0x3433(%rip),%ymm8 # 6a80 <_sk_callback_avx+0x358> + DB 196,98,125,24,5,144,51,0,0 ; vbroadcastss 0x3390(%rip),%ymm8 # 68b8 <_sk_callback_avx+0x358> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 197,252,92,210 ; vsubps %ymm2,%ymm0,%ymm2 DB 197,116,89,193 ; vmulps %ymm1,%ymm1,%ymm8 DB 196,65,116,89,192 ; vmulps %ymm8,%ymm1,%ymm8 - DB 196,98,125,24,13,28,52,0,0 ; vbroadcastss 0x341c(%rip),%ymm9 # 6a84 <_sk_callback_avx+0x35c> + DB 196,98,125,24,13,121,51,0,0 ; vbroadcastss 0x3379(%rip),%ymm9 # 68bc <_sk_callback_avx+0x35c> DB 196,65,52,194,208,1 ; vcmpltps %ymm8,%ymm9,%ymm10 - DB 196,98,125,24,29,17,52,0,0 ; vbroadcastss 0x3411(%rip),%ymm11 # 6a88 <_sk_callback_avx+0x360> + DB 196,98,125,24,29,110,51,0,0 ; vbroadcastss 0x336e(%rip),%ymm11 # 68c0 <_sk_callback_avx+0x360> DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1 - DB 196,98,125,24,37,7,52,0,0 ; vbroadcastss 0x3407(%rip),%ymm12 # 6a8c <_sk_callback_avx+0x364> + DB 196,98,125,24,37,100,51,0,0 ; vbroadcastss 0x3364(%rip),%ymm12 # 68c4 <_sk_callback_avx+0x364> DB 196,193,116,89,204 ; vmulps %ymm12,%ymm1,%ymm1 DB 196,67,117,74,192,160 ; vblendvps %ymm10,%ymm8,%ymm1,%ymm8 DB 197,252,89,200 ; vmulps %ymm0,%ymm0,%ymm1 @@ -8342,9 +8256,9 @@ _sk_lab_to_xyz_avx LABEL PROC DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2 DB 196,193,108,89,212 ; vmulps %ymm12,%ymm2,%ymm2 DB 196,227,109,74,208,144 ; vblendvps %ymm9,%ymm0,%ymm2,%ymm2 - DB 196,226,125,24,5,189,51,0,0 ; vbroadcastss 0x33bd(%rip),%ymm0 # 6a90 <_sk_callback_avx+0x368> + DB 196,226,125,24,5,26,51,0,0 ; vbroadcastss 0x331a(%rip),%ymm0 # 68c8 <_sk_callback_avx+0x368> DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,180,51,0,0 ; vbroadcastss 0x33b4(%rip),%ymm8 # 6a94 <_sk_callback_avx+0x36c> + DB 196,98,125,24,5,17,51,0,0 ; vbroadcastss 0x3311(%rip),%ymm8 # 68cc <_sk_callback_avx+0x36c> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8356,14 +8270,14 @@ _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 117,62 ; jne 3737 <_sk_load_a8_avx+0x4e> + DB 117,62 ; jne 3612 <_sk_load_a8_avx+0x4e> 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 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,120,51,0,0 ; vbroadcastss 0x3378(%rip),%ymm1 # 6a98 <_sk_callback_avx+0x370> + DB 196,226,125,24,13,213,50,0,0 ; vbroadcastss 0x32d5(%rip),%ymm1 # 68d0 <_sk_callback_avx+0x370> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -8380,9 +8294,9 @@ _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 373f <_sk_load_a8_avx+0x56> + DB 117,234 ; jne 361a <_sk_load_a8_avx+0x56> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,161 ; jmp 36fd <_sk_load_a8_avx+0x14> + DB 235,161 ; jmp 35d8 <_sk_load_a8_avx+0x14> PUBLIC _sk_gather_a8_avx _sk_gather_a8_avx LABEL PROC @@ -8430,7 +8344,7 @@ _sk_gather_a8_avx LABEL PROC DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,109,50,0,0 ; vbroadcastss 0x326d(%rip),%ymm1 # 6a9c <_sk_callback_avx+0x374> + DB 196,226,125,24,13,202,49,0,0 ; vbroadcastss 0x31ca(%rip),%ymm1 # 68d4 <_sk_callback_avx+0x374> DB 197,252,89,217 ; vmulps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,87,192 ; vxorps %ymm0,%ymm0,%ymm0 @@ -8446,14 +8360,14 @@ PUBLIC _sk_store_a8_avx _sk_store_a8_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,72,50,0,0 ; vbroadcastss 0x3248(%rip),%ymm8 # 6aa0 <_sk_callback_avx+0x378> + DB 196,98,125,24,5,165,49,0,0 ; vbroadcastss 0x31a5(%rip),%ymm8 # 68d8 <_sk_callback_avx+0x378> DB 196,65,100,89,192 ; vmulps %ymm8,%ymm3,%ymm8 DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8 DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 3881 <_sk_store_a8_avx+0x37> + DB 117,10 ; jne 375c <_sk_store_a8_avx+0x37> DB 196,65,123,17,4,58 ; vmovsd %xmm8,(%r10,%rdi,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8461,10 +8375,10 @@ _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 387d <_sk_store_a8_avx+0x33> + DB 119,236 ; ja 3758 <_sk_store_a8_avx+0x33> DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8 DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,67,0,0,0 ; lea 0x43(%rip),%r9 # 38e4 <_sk_store_a8_avx+0x9a> + DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 37c0 <_sk_store_a8_avx+0x9b> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -8475,27 +8389,28 @@ _sk_store_a8_avx LABEL PROC DB 196,67,121,20,68,58,2,4 ; vpextrb $0x4,%xmm8,0x2(%r10,%rdi,1) DB 196,67,121,20,68,58,1,2 ; vpextrb $0x2,%xmm8,0x1(%r10,%rdi,1) DB 196,67,121,20,4,58,0 ; vpextrb $0x0,%xmm8,(%r10,%rdi,1) - DB 235,154 ; jmp 387d <_sk_store_a8_avx+0x33> - DB 144 ; nop - DB 246,255 ; idiv %bh + DB 235,154 ; jmp 3758 <_sk_store_a8_avx+0x33> + DB 102,144 ; xchg %ax,%ax + DB 245 ; cmc DB 255 ; (bad) DB 255 ; (bad) - DB 238 ; out %al,(%dx) DB 255 ; (bad) + DB 237 ; in (%dx),%eax DB 255 ; (bad) - DB 255,230 ; jmpq *%rsi DB 255 ; (bad) + DB 255,229 ; jmpq *%rbp DB 255 ; (bad) DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) DB 255 ; (bad) - DB 255,214 ; callq *%rsi + DB 221,255 ; (bad) DB 255 ; (bad) + DB 255,213 ; callq *%rbp DB 255 ; (bad) - DB 255,206 ; dec %esi DB 255 ; (bad) + DB 255,205 ; dec %ebp DB 255 ; (bad) - DB 255,198 ; inc %esi + DB 255 ; (bad) + DB 255,197 ; inc %ebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -8507,17 +8422,17 @@ _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 117,67 ; jne 3953 <_sk_load_g8_avx+0x53> + DB 117,67 ; jne 382f <_sk_load_g8_avx+0x53> 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 DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,109,49,0,0 ; vbroadcastss 0x316d(%rip),%ymm1 # 6aa4 <_sk_callback_avx+0x37c> + DB 196,226,125,24,13,201,48,0,0 ; vbroadcastss 0x30c9(%rip),%ymm1 # 68dc <_sk_callback_avx+0x37c> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,98,49,0,0 ; vbroadcastss 0x3162(%rip),%ymm3 # 6aa8 <_sk_callback_avx+0x380> + DB 196,226,125,24,29,190,48,0,0 ; vbroadcastss 0x30be(%rip),%ymm3 # 68e0 <_sk_callback_avx+0x380> DB 76,137,193 ; mov %r8,%rcx DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 @@ -8531,9 +8446,9 @@ _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 395b <_sk_load_g8_avx+0x5b> + DB 117,234 ; jne 3837 <_sk_load_g8_avx+0x5b> DB 196,193,249,110,193 ; vmovq %r9,%xmm0 - DB 235,156 ; jmp 3914 <_sk_load_g8_avx+0x14> + DB 235,156 ; jmp 37f0 <_sk_load_g8_avx+0x14> PUBLIC _sk_gather_g8_avx _sk_gather_g8_avx LABEL PROC @@ -8581,10 +8496,10 @@ _sk_gather_g8_avx LABEL PROC DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,97,48,0,0 ; vbroadcastss 0x3061(%rip),%ymm1 # 6aac <_sk_callback_avx+0x384> + DB 196,226,125,24,13,189,47,0,0 ; vbroadcastss 0x2fbd(%rip),%ymm1 # 68e4 <_sk_callback_avx+0x384> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,86,48,0,0 ; vbroadcastss 0x3056(%rip),%ymm3 # 6ab0 <_sk_callback_avx+0x388> + DB 196,226,125,24,29,178,47,0,0 ; vbroadcastss 0x2fb2(%rip),%ymm3 # 68e8 <_sk_callback_avx+0x388> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 197,252,40,208 ; vmovaps %ymm0,%ymm2 DB 91 ; pop %rbx @@ -8598,9 +8513,9 @@ _sk_gather_i8_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 73,137,192 ; mov %rax,%r8 DB 77,133,192 ; test %r8,%r8 - DB 116,5 ; je 3a7a <_sk_gather_i8_avx+0xf> + DB 116,5 ; je 3956 <_sk_gather_i8_avx+0xf> DB 76,137,192 ; mov %r8,%rax - DB 235,2 ; jmp 3a7c <_sk_gather_i8_avx+0x11> + DB 235,2 ; jmp 3958 <_sk_gather_i8_avx+0x11> DB 72,173 ; lods %ds:(%rsi),%rax DB 65,87 ; push %r15 DB 65,86 ; push %r14 @@ -8662,10 +8577,10 @@ _sk_gather_i8_avx LABEL PROC DB 196,163,121,34,4,163,2 ; vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0 DB 196,163,121,34,28,19,3 ; vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,26,49,0,0 ; vmovaps 0x311a(%rip),%ymm10 # 6cc0 <_sk_callback_avx+0x598> + DB 197,124,40,21,30,49,0,0 ; vmovaps 0x311e(%rip),%ymm10 # 6ba0 <_sk_callback_avx+0x640> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,252,46,0,0 ; vbroadcastss 0x2efc(%rip),%ymm9 # 6ab4 <_sk_callback_avx+0x38c> + DB 196,98,125,24,13,88,46,0,0 ; vbroadcastss 0x2e58(%rip),%ymm9 # 68ec <_sk_callback_avx+0x38c> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -8697,38 +8612,38 @@ _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 15,133,128,0,0,0 ; jne 3cb0 <_sk_load_565_avx+0x8e> + DB 15,133,128,0,0,0 ; jne 3b8c <_sk_load_565_avx+0x8e> 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 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,102,46,0,0 ; vbroadcastss 0x2e66(%rip),%ymm0 # 6ab8 <_sk_callback_avx+0x390> + DB 196,226,125,24,5,194,45,0,0 ; vbroadcastss 0x2dc2(%rip),%ymm0 # 68f0 <_sk_callback_avx+0x390> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,89,46,0,0 ; vbroadcastss 0x2e59(%rip),%ymm1 # 6abc <_sk_callback_avx+0x394> + DB 196,226,125,24,13,181,45,0,0 ; vbroadcastss 0x2db5(%rip),%ymm1 # 68f4 <_sk_callback_avx+0x394> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,80,46,0,0 ; vbroadcastss 0x2e50(%rip),%ymm1 # 6ac0 <_sk_callback_avx+0x398> + DB 196,226,125,24,13,172,45,0,0 ; vbroadcastss 0x2dac(%rip),%ymm1 # 68f8 <_sk_callback_avx+0x398> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,67,46,0,0 ; vbroadcastss 0x2e43(%rip),%ymm3 # 6ac4 <_sk_callback_avx+0x39c> + DB 196,226,125,24,29,159,45,0,0 ; vbroadcastss 0x2d9f(%rip),%ymm3 # 68fc <_sk_callback_avx+0x39c> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,58,46,0,0 ; vbroadcastss 0x2e3a(%rip),%ymm3 # 6ac8 <_sk_callback_avx+0x3a0> + DB 196,226,125,24,29,150,45,0,0 ; vbroadcastss 0x2d96(%rip),%ymm3 # 6900 <_sk_callback_avx+0x3a0> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,45,46,0,0 ; vbroadcastss 0x2e2d(%rip),%ymm3 # 6acc <_sk_callback_avx+0x3a4> + DB 196,226,125,24,29,137,45,0,0 ; vbroadcastss 0x2d89(%rip),%ymm3 # 6904 <_sk_callback_avx+0x3a4> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,34,46,0,0 ; vbroadcastss 0x2e22(%rip),%ymm3 # 6ad0 <_sk_callback_avx+0x3a8> + DB 196,226,125,24,29,126,45,0,0 ; vbroadcastss 0x2d7e(%rip),%ymm3 # 6908 <_sk_callback_avx+0x3a8> DB 255,224 ; jmpq *%rax DB 65,137,200 ; mov %ecx,%r8d DB 65,128,224,7 ; and $0x7,%r8b DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,110,255,255,255 ; ja 3c36 <_sk_load_565_avx+0x14> + DB 15,135,110,255,255,255 ; ja 3b12 <_sk_load_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 3d1c <_sk_load_565_avx+0xfa> + DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 3bf8 <_sk_load_565_avx+0xfa> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -8740,7 +8655,7 @@ _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 233,26,255,255,255 ; jmpq 3c36 <_sk_load_565_avx+0x14> + DB 233,26,255,255,255 ; jmpq 3b12 <_sk_load_565_avx+0x14> DB 244 ; hlt DB 255 ; (bad) DB 255 ; (bad) @@ -8816,23 +8731,23 @@ _sk_gather_565_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,209,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 - DB 196,226,125,24,5,194,44,0,0 ; vbroadcastss 0x2cc2(%rip),%ymm0 # 6ad4 <_sk_callback_avx+0x3ac> + DB 196,226,125,24,5,30,44,0,0 ; vbroadcastss 0x2c1e(%rip),%ymm0 # 690c <_sk_callback_avx+0x3ac> DB 197,236,84,192 ; vandps %ymm0,%ymm2,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,181,44,0,0 ; vbroadcastss 0x2cb5(%rip),%ymm1 # 6ad8 <_sk_callback_avx+0x3b0> + DB 196,226,125,24,13,17,44,0,0 ; vbroadcastss 0x2c11(%rip),%ymm1 # 6910 <_sk_callback_avx+0x3b0> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,172,44,0,0 ; vbroadcastss 0x2cac(%rip),%ymm1 # 6adc <_sk_callback_avx+0x3b4> + DB 196,226,125,24,13,8,44,0,0 ; vbroadcastss 0x2c08(%rip),%ymm1 # 6914 <_sk_callback_avx+0x3b4> DB 197,236,84,201 ; vandps %ymm1,%ymm2,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,29,159,44,0,0 ; vbroadcastss 0x2c9f(%rip),%ymm3 # 6ae0 <_sk_callback_avx+0x3b8> + DB 196,226,125,24,29,251,43,0,0 ; vbroadcastss 0x2bfb(%rip),%ymm3 # 6918 <_sk_callback_avx+0x3b8> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 - DB 196,226,125,24,29,150,44,0,0 ; vbroadcastss 0x2c96(%rip),%ymm3 # 6ae4 <_sk_callback_avx+0x3bc> + DB 196,226,125,24,29,242,43,0,0 ; vbroadcastss 0x2bf2(%rip),%ymm3 # 691c <_sk_callback_avx+0x3bc> DB 197,236,84,211 ; vandps %ymm3,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,226,125,24,29,137,44,0,0 ; vbroadcastss 0x2c89(%rip),%ymm3 # 6ae8 <_sk_callback_avx+0x3c0> + DB 196,226,125,24,29,229,43,0,0 ; vbroadcastss 0x2be5(%rip),%ymm3 # 6920 <_sk_callback_avx+0x3c0> DB 197,236,89,211 ; vmulps %ymm3,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,126,44,0,0 ; vbroadcastss 0x2c7e(%rip),%ymm3 # 6aec <_sk_callback_avx+0x3c4> + DB 196,226,125,24,29,218,43,0,0 ; vbroadcastss 0x2bda(%rip),%ymm3 # 6924 <_sk_callback_avx+0x3c4> DB 91 ; pop %rbx DB 65,92 ; pop %r12 DB 65,94 ; pop %r14 @@ -8844,14 +8759,14 @@ PUBLIC _sk_store_565_avx _sk_store_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,106,44,0,0 ; vbroadcastss 0x2c6a(%rip),%ymm8 # 6af0 <_sk_callback_avx+0x3c8> + DB 196,98,125,24,5,198,43,0,0 ; vbroadcastss 0x2bc6(%rip),%ymm8 # 6928 <_sk_callback_avx+0x3c8> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,11 ; vpslld $0xb,%xmm9,%xmm10 DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9 DB 196,193,49,114,241,11 ; vpslld $0xb,%xmm9,%xmm9 DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 - DB 196,98,125,24,21,67,44,0,0 ; vbroadcastss 0x2c43(%rip),%ymm10 # 6af4 <_sk_callback_avx+0x3cc> + DB 196,98,125,24,21,159,43,0,0 ; vbroadcastss 0x2b9f(%rip),%ymm10 # 692c <_sk_callback_avx+0x3cc> DB 196,65,116,89,210 ; vmulps %ymm10,%ymm1,%ymm10 DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10 DB 196,193,33,114,242,5 ; vpslld $0x5,%xmm10,%xmm11 @@ -8865,7 +8780,7 @@ _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 3f01 <_sk_store_565_avx+0x89> + DB 117,10 ; jne 3ddd <_sk_store_565_avx+0x89> DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8873,9 +8788,9 @@ _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 3efd <_sk_store_565_avx+0x85> + DB 119,236 ; ja 3dd9 <_sk_store_565_avx+0x85> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3f60 <_sk_store_565_avx+0xe8> + DB 76,141,13,68,0,0,0 ; lea 0x44(%rip),%r9 # 3e3c <_sk_store_565_avx+0xe8> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -8886,7 +8801,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2) - DB 235,159 ; jmp 3efd <_sk_store_565_avx+0x85> + DB 235,159 ; jmp 3dd9 <_sk_store_565_avx+0x85> DB 102,144 ; xchg %ax,%ax DB 245 ; cmc DB 255 ; (bad) @@ -8917,31 +8832,31 @@ _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 15,133,152,0,0,0 ; jne 4022 <_sk_load_4444_avx+0xa6> + DB 15,133,152,0,0,0 ; jne 3efe <_sk_load_4444_avx+0xa6> 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 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,76,43,0,0 ; vbroadcastss 0x2b4c(%rip),%ymm0 # 6af8 <_sk_callback_avx+0x3d0> + DB 196,226,125,24,5,168,42,0,0 ; vbroadcastss 0x2aa8(%rip),%ymm0 # 6930 <_sk_callback_avx+0x3d0> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,63,43,0,0 ; vbroadcastss 0x2b3f(%rip),%ymm1 # 6afc <_sk_callback_avx+0x3d4> + DB 196,226,125,24,13,155,42,0,0 ; vbroadcastss 0x2a9b(%rip),%ymm1 # 6934 <_sk_callback_avx+0x3d4> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,54,43,0,0 ; vbroadcastss 0x2b36(%rip),%ymm1 # 6b00 <_sk_callback_avx+0x3d8> + DB 196,226,125,24,13,146,42,0,0 ; vbroadcastss 0x2a92(%rip),%ymm1 # 6938 <_sk_callback_avx+0x3d8> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,41,43,0,0 ; vbroadcastss 0x2b29(%rip),%ymm2 # 6b04 <_sk_callback_avx+0x3dc> + DB 196,226,125,24,21,133,42,0,0 ; vbroadcastss 0x2a85(%rip),%ymm2 # 693c <_sk_callback_avx+0x3dc> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,32,43,0,0 ; vbroadcastss 0x2b20(%rip),%ymm2 # 6b08 <_sk_callback_avx+0x3e0> + DB 196,226,125,24,21,124,42,0,0 ; vbroadcastss 0x2a7c(%rip),%ymm2 # 6940 <_sk_callback_avx+0x3e0> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,19,43,0,0 ; vbroadcastss 0x2b13(%rip),%ymm8 # 6b0c <_sk_callback_avx+0x3e4> + DB 196,98,125,24,5,111,42,0,0 ; vbroadcastss 0x2a6f(%rip),%ymm8 # 6944 <_sk_callback_avx+0x3e4> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,9,43,0,0 ; vbroadcastss 0x2b09(%rip),%ymm8 # 6b10 <_sk_callback_avx+0x3e8> + DB 196,98,125,24,5,101,42,0,0 ; vbroadcastss 0x2a65(%rip),%ymm8 # 6948 <_sk_callback_avx+0x3e8> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,251,42,0,0 ; vbroadcastss 0x2afb(%rip),%ymm8 # 6b14 <_sk_callback_avx+0x3ec> + DB 196,98,125,24,5,87,42,0,0 ; vbroadcastss 0x2a57(%rip),%ymm8 # 694c <_sk_callback_avx+0x3ec> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -8950,9 +8865,9 @@ _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 15,135,86,255,255,255 ; ja 3f90 <_sk_load_4444_avx+0x14> + DB 15,135,86,255,255,255 ; ja 3e6c <_sk_load_4444_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 4090 <_sk_load_4444_avx+0x114> + DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 3f6c <_sk_load_4444_avx+0x114> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -8964,7 +8879,7 @@ _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 233,2,255,255,255 ; jmpq 3f90 <_sk_load_4444_avx+0x14> + DB 233,2,255,255,255 ; jmpq 3e6c <_sk_load_4444_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 242,255 ; repnz (bad) DB 255 ; (bad) @@ -9041,25 +8956,25 @@ _sk_gather_4444_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm3 - DB 196,226,125,24,5,146,41,0,0 ; vbroadcastss 0x2992(%rip),%ymm0 # 6b18 <_sk_callback_avx+0x3f0> + DB 196,226,125,24,5,238,40,0,0 ; vbroadcastss 0x28ee(%rip),%ymm0 # 6950 <_sk_callback_avx+0x3f0> DB 197,228,84,192 ; vandps %ymm0,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,226,125,24,13,133,41,0,0 ; vbroadcastss 0x2985(%rip),%ymm1 # 6b1c <_sk_callback_avx+0x3f4> + DB 196,226,125,24,13,225,40,0,0 ; vbroadcastss 0x28e1(%rip),%ymm1 # 6954 <_sk_callback_avx+0x3f4> DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,124,41,0,0 ; vbroadcastss 0x297c(%rip),%ymm1 # 6b20 <_sk_callback_avx+0x3f8> + DB 196,226,125,24,13,216,40,0,0 ; vbroadcastss 0x28d8(%rip),%ymm1 # 6958 <_sk_callback_avx+0x3f8> DB 197,228,84,201 ; vandps %ymm1,%ymm3,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 - DB 196,226,125,24,21,111,41,0,0 ; vbroadcastss 0x296f(%rip),%ymm2 # 6b24 <_sk_callback_avx+0x3fc> + DB 196,226,125,24,21,203,40,0,0 ; vbroadcastss 0x28cb(%rip),%ymm2 # 695c <_sk_callback_avx+0x3fc> DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1 - DB 196,226,125,24,21,102,41,0,0 ; vbroadcastss 0x2966(%rip),%ymm2 # 6b28 <_sk_callback_avx+0x400> + DB 196,226,125,24,21,194,40,0,0 ; vbroadcastss 0x28c2(%rip),%ymm2 # 6960 <_sk_callback_avx+0x400> DB 197,228,84,210 ; vandps %ymm2,%ymm3,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 - DB 196,98,125,24,5,89,41,0,0 ; vbroadcastss 0x2959(%rip),%ymm8 # 6b2c <_sk_callback_avx+0x404> + DB 196,98,125,24,5,181,40,0,0 ; vbroadcastss 0x28b5(%rip),%ymm8 # 6964 <_sk_callback_avx+0x404> DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,98,125,24,5,79,41,0,0 ; vbroadcastss 0x294f(%rip),%ymm8 # 6b30 <_sk_callback_avx+0x408> + DB 196,98,125,24,5,171,40,0,0 ; vbroadcastss 0x28ab(%rip),%ymm8 # 6968 <_sk_callback_avx+0x408> DB 196,193,100,84,216 ; vandps %ymm8,%ymm3,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 - DB 196,98,125,24,5,65,41,0,0 ; vbroadcastss 0x2941(%rip),%ymm8 # 6b34 <_sk_callback_avx+0x40c> + DB 196,98,125,24,5,157,40,0,0 ; vbroadcastss 0x289d(%rip),%ymm8 # 696c <_sk_callback_avx+0x40c> DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax DB 91 ; pop %rbx @@ -9073,7 +8988,7 @@ PUBLIC _sk_store_4444_avx _sk_store_4444_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,38,41,0,0 ; vbroadcastss 0x2926(%rip),%ymm8 # 6b38 <_sk_callback_avx+0x410> + DB 196,98,125,24,5,130,40,0,0 ; vbroadcastss 0x2882(%rip),%ymm8 # 6970 <_sk_callback_avx+0x410> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,193,41,114,241,12 ; vpslld $0xc,%xmm9,%xmm10 @@ -9100,7 +9015,7 @@ _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 42ab <_sk_store_4444_avx+0xa7> + DB 117,10 ; jne 4187 <_sk_store_4444_avx+0xa7> DB 196,65,122,127,4,122 ; vmovdqu %xmm8,(%r10,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -9108,9 +9023,9 @@ _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 42a7 <_sk_store_4444_avx+0xa3> + DB 119,236 ; ja 4183 <_sk_store_4444_avx+0xa3> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,66,0,0,0 ; lea 0x42(%rip),%r9 # 4308 <_sk_store_4444_avx+0x104> + DB 76,141,13,66,0,0,0 ; lea 0x42(%rip),%r9 # 41e4 <_sk_store_4444_avx+0x104> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -9121,7 +9036,7 @@ _sk_store_4444_avx LABEL PROC DB 196,67,121,21,68,122,4,2 ; vpextrw $0x2,%xmm8,0x4(%r10,%rdi,2) DB 196,67,121,21,68,122,2,1 ; vpextrw $0x1,%xmm8,0x2(%r10,%rdi,2) DB 196,67,121,21,4,122,0 ; vpextrw $0x0,%xmm8,(%r10,%rdi,2) - DB 235,159 ; jmp 42a7 <_sk_store_4444_avx+0xa3> + DB 235,159 ; jmp 4183 <_sk_store_4444_avx+0xa3> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -9147,87 +9062,55 @@ _sk_store_4444_avx LABEL PROC PUBLIC _sk_load_8888_avx _sk_load_8888_avx LABEL PROC + DB 80 ; push %rax + DB 73,137,200 ; mov %rcx,%r8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 72,133,201 ; test %rcx,%rcx - DB 15,133,135,0,0,0 ; jne 43b9 <_sk_load_8888_avx+0x95> - DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9 - DB 197,124,40,21,160,41,0,0 ; vmovaps 0x29a0(%rip),%ymm10 # 6ce0 <_sk_callback_avx+0x5b8> - DB 196,193,52,84,194 ; vandps %ymm10,%ymm9,%ymm0 + 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 15,133,139,0,0,0 ; jne 42a5 <_sk_load_8888_avx+0xa5> + DB 196,193,124,16,25 ; vmovups (%r9),%ymm3 + DB 197,124,40,21,153,41,0,0 ; vmovaps 0x2999(%rip),%ymm10 # 6bc0 <_sk_callback_avx+0x660> + DB 196,193,100,84,194 ; vandps %ymm10,%ymm3,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,5,234,39,0,0 ; vbroadcastss 0x27ea(%rip),%ymm8 # 6b3c <_sk_callback_avx+0x414> + DB 196,98,125,24,5,59,39,0,0 ; vbroadcastss 0x273b(%rip),%ymm8 # 6974 <_sk_callback_avx+0x414> DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 - DB 196,193,113,114,209,8 ; vpsrld $0x8,%xmm9,%xmm1 - DB 196,99,125,25,203,1 ; vextractf128 $0x1,%ymm9,%xmm3 - DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 + DB 197,241,114,211,8 ; vpsrld $0x8,%xmm3,%xmm1 + DB 196,195,125,25,217,1 ; vextractf128 $0x1,%ymm3,%xmm9 + DB 196,193,105,114,209,8 ; vpsrld $0x8,%xmm9,%xmm2 DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 196,193,116,84,202 ; vandps %ymm10,%ymm1,%ymm1 DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 - DB 196,193,33,114,209,16 ; vpsrld $0x10,%xmm9,%xmm11 - DB 197,233,114,211,16 ; vpsrld $0x10,%xmm3,%xmm2 + DB 197,161,114,211,16 ; vpsrld $0x10,%xmm3,%xmm11 + DB 196,193,105,114,209,16 ; vpsrld $0x10,%xmm9,%xmm2 DB 196,227,37,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm11,%ymm2 DB 196,193,108,84,210 ; vandps %ymm10,%ymm2,%ymm2 DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 - DB 196,193,49,114,209,24 ; vpsrld $0x18,%xmm9,%xmm9 - DB 197,225,114,211,24 ; vpsrld $0x18,%xmm3,%xmm3 - DB 196,227,53,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 + DB 197,169,114,211,24 ; vpsrld $0x18,%xmm3,%xmm10 + DB 196,193,97,114,209,24 ; vpsrld $0x18,%xmm9,%xmm3 + DB 196,227,45,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm10,%ymm3 DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 DB 196,193,100,89,216 ; vmulps %ymm8,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,137,193 ; mov %r8,%rcx + DB 65,88 ; pop %r8 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - DB 65,128,224,7 ; and $0x7,%r8b - 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 15,135,102,255,255,255 ; ja 4338 <_sk_load_8888_avx+0x14> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,139,0,0,0 ; lea 0x8b(%rip),%r9 # 4468 <_sk_load_8888_avx+0x144> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,193,121,110,68,186,24 ; vmovd 0x18(%r10,%rdi,4),%xmm0 - DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0 - DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 - DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 - DB 196,99,117,12,200,64 ; vblendps $0x40,%ymm0,%ymm1,%ymm9 - DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 - DB 196,195,121,34,68,186,20,1 ; vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0 - DB 196,99,53,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 - DB 196,99,125,25,200,1 ; vextractf128 $0x1,%ymm9,%xmm0 - DB 196,195,121,34,68,186,16,0 ; vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0 - DB 196,99,53,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 - DB 196,195,49,34,68,186,12,3 ; vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm9,%xmm0 - DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 - DB 196,195,49,34,68,186,8,2 ; vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm9,%xmm0 - DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 - DB 196,195,49,34,68,186,4,1 ; vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm9,%xmm0 - 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 233,210,254,255,255 ; jmpq 4338 <_sk_load_8888_avx+0x14> - DB 102,144 ; xchg %ax,%ax - DB 236 ; in (%dx),%al - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 222,255 ; fdivrp %st,%st(7) - DB 255 ; (bad) - DB 255,208 ; callq *%rax - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,194 ; inc %edx - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,174,255,255,255,154 ; ljmp *-0x65000001(%rsi) - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 126,255 ; jle 4481 <_sk_load_8888_avx+0x15d> - DB 255 ; (bad) - DB 255 ; .byte 0xff + DB 185,8,0,0,0 ; mov $0x8,%ecx + DB 68,41,193 ; sub %r8d,%ecx + DB 192,225,3 ; shl $0x3,%cl + DB 72,199,192,255,255,255,255 ; mov $0xffffffffffffffff,%rax + DB 72,211,232 ; shr %cl,%rax + DB 196,225,249,110,192 ; vmovq %rax,%xmm0 + DB 196,226,121,48,192 ; vpmovzxbw %xmm0,%xmm0 + DB 196,226,121,0,13,67,40,0,0 ; vpshufb 0x2843(%rip),%xmm0,%xmm1 # 6b10 <_sk_callback_avx+0x5b0> + DB 196,226,121,33,201 ; vpmovsxbd %xmm1,%xmm1 + DB 196,226,121,0,5,69,40,0,0 ; vpshufb 0x2845(%rip),%xmm0,%xmm0 # 6b20 <_sk_callback_avx+0x5c0> + DB 196,226,121,33,192 ; vpmovsxbd %xmm0,%xmm0 + DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 + DB 196,194,125,44,25 ; vmaskmovps (%r9),%ymm0,%ymm3 + DB 233,47,255,255,255 ; jmpq 421f <_sk_load_8888_avx+0x1f> PUBLIC _sk_gather_8888_avx _sk_gather_8888_avx LABEL PROC @@ -9268,10 +9151,10 @@ _sk_gather_8888_avx LABEL PROC DB 196,131,121,34,4,152,2 ; vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 DB 196,131,121,34,28,144,3 ; vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 - DB 197,124,40,21,202,39,0,0 ; vmovaps 0x27ca(%rip),%ymm10 # 6d00 <_sk_callback_avx+0x5d8> + DB 197,124,40,21,62,40,0,0 ; vmovaps 0x283e(%rip),%ymm10 # 6be0 <_sk_callback_avx+0x680> DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,13,248,37,0,0 ; vbroadcastss 0x25f8(%rip),%ymm9 # 6b40 <_sk_callback_avx+0x418> + DB 196,98,125,24,13,196,37,0,0 ; vbroadcastss 0x25c4(%rip),%ymm9 # 6978 <_sk_callback_avx+0x418> DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1 DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2 @@ -9299,9 +9182,12 @@ _sk_gather_8888_avx LABEL PROC PUBLIC _sk_store_8888_avx _sk_store_8888_avx LABEL PROC + DB 80 ; push %rax + DB 73,137,200 ; mov %rcx,%r8 DB 72,173 ; lods %ds:(%rsi),%rax - DB 76,139,16 ; mov (%rax),%r10 - DB 196,98,125,24,5,134,37,0,0 ; vbroadcastss 0x2586(%rip),%ymm8 # 6b44 <_sk_callback_avx+0x41c> + DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 + DB 76,3,8 ; add (%rax),%r9 + DB 196,98,125,24,5,70,37,0,0 ; vbroadcastss 0x2546(%rip),%ymm8 # 697c <_sk_callback_avx+0x41c> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,65,116,89,208 ; vmulps %ymm8,%ymm1,%ymm10 @@ -9325,56 +9211,27 @@ _sk_store_8888_avx LABEL PROC DB 196,67,37,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm11,%ymm8 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 464c <_sk_store_8888_avx+0x9c> - DB 196,65,124,17,4,186 ; vmovups %ymm8,(%r10,%rdi,4) + DB 77,133,192 ; test %r8,%r8 + DB 117,14 ; jne 44c8 <_sk_store_8888_avx+0xac> + DB 196,65,124,17,1 ; vmovups %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,137,193 ; mov %r8,%rcx + DB 65,88 ; pop %r8 DB 255,224 ; jmpq *%rax - DB 65,137,200 ; mov %ecx,%r8d - 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 4648 <_sk_store_8888_avx+0x98> - DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,85,0,0,0 ; lea 0x55(%rip),%r9 # 46bc <_sk_store_8888_avx+0x10c> - DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax - DB 76,1,200 ; add %r9,%rax - DB 255,224 ; jmpq *%rax - DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 196,67,121,22,76,186,24,2 ; vpextrd $0x2,%xmm9,0x18(%r10,%rdi,4) - DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 196,67,121,22,76,186,20,1 ; vpextrd $0x1,%xmm9,0x14(%r10,%rdi,4) - DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9 - DB 196,65,122,17,76,186,16 ; vmovss %xmm9,0x10(%r10,%rdi,4) - DB 196,67,121,22,68,186,12,3 ; vpextrd $0x3,%xmm8,0xc(%r10,%rdi,4) - DB 196,67,121,22,68,186,8,2 ; vpextrd $0x2,%xmm8,0x8(%r10,%rdi,4) - DB 196,67,121,22,68,186,4,1 ; vpextrd $0x1,%xmm8,0x4(%r10,%rdi,4) - DB 196,65,121,126,4,186 ; vmovd %xmm8,(%r10,%rdi,4) - DB 235,143 ; jmp 4648 <_sk_store_8888_avx+0x98> - DB 15,31,0 ; nopl (%rax) - DB 245 ; cmc - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 237 ; in (%dx),%eax - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,229 ; jmpq *%rbp - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; (bad) - DB 221,255 ; (bad) - DB 255 ; (bad) - DB 255,208 ; callq *%rax - DB 255 ; (bad) - DB 255 ; (bad) - DB 255,194 ; inc %edx - DB 255 ; (bad) - DB 255 ; (bad) - DB 255 ; .byte 0xff - DB 180,255 ; mov $0xff,%ah - DB 255 ; (bad) - DB 255 ; .byte 0xff + DB 185,8,0,0,0 ; mov $0x8,%ecx + DB 68,41,193 ; sub %r8d,%ecx + DB 192,225,3 ; shl $0x3,%cl + DB 72,199,192,255,255,255,255 ; mov $0xffffffffffffffff,%rax + DB 72,211,232 ; shr %cl,%rax + DB 196,97,249,110,200 ; vmovq %rax,%xmm9 + DB 196,66,121,48,201 ; vpmovzxbw %xmm9,%xmm9 + DB 196,98,49,0,21,64,38,0,0 ; vpshufb 0x2640(%rip),%xmm9,%xmm10 # 6b30 <_sk_callback_avx+0x5d0> + DB 196,66,121,33,210 ; vpmovsxbd %xmm10,%xmm10 + DB 196,98,49,0,13,66,38,0,0 ; vpshufb 0x2642(%rip),%xmm9,%xmm9 # 6b40 <_sk_callback_avx+0x5e0> + DB 196,66,121,33,201 ; vpmovsxbd %xmm9,%xmm9 + DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 + DB 196,66,53,46,1 ; vmaskmovps %ymm8,%ymm9,(%r9) + DB 235,175 ; jmp 44bf <_sk_store_8888_avx+0xa3> PUBLIC _sk_load_f16_avx _sk_load_f16_avx LABEL PROC @@ -9386,7 +9243,7 @@ _sk_load_f16_avx LABEL PROC DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp) DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp) DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp) - DB 15,133,143,2,0,0 ; jne 4993 <_sk_load_f16_avx+0x2bb> + DB 15,133,143,2,0,0 ; jne 47cb <_sk_load_f16_avx+0x2bb> 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 197,249,16,76,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm1 @@ -9404,13 +9261,13 @@ _sk_load_f16_avx LABEL PROC DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 - DB 196,98,125,24,37,235,35,0,0 ; vbroadcastss 0x23eb(%rip),%ymm12 # 6b48 <_sk_callback_avx+0x420> + DB 196,98,125,24,37,235,35,0,0 ; vbroadcastss 0x23eb(%rip),%ymm12 # 6980 <_sk_callback_avx+0x420> DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1 DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0 DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,215,35,0,0 ; vbroadcastss 0x23d7(%rip),%xmm11 # 6b4c <_sk_callback_avx+0x424> + DB 196,98,121,24,29,215,35,0,0 ; vbroadcastss 0x23d7(%rip),%xmm11 # 6984 <_sk_callback_avx+0x424> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,205,35,0,0 ; vbroadcastss 0x23cd(%rip),%xmm13 # 6b50 <_sk_callback_avx+0x428> + DB 196,98,121,24,45,205,35,0,0 ; vbroadcastss 0x23cd(%rip),%xmm13 # 6988 <_sk_callback_avx+0x428> DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -9424,7 +9281,7 @@ _sk_load_f16_avx LABEL PROC DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0 DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1 - DB 196,226,121,24,29,131,35,0,0 ; vbroadcastss 0x2383(%rip),%xmm3 # 6b54 <_sk_callback_avx+0x42c> + DB 196,226,121,24,29,131,35,0,0 ; vbroadcastss 0x2383(%rip),%xmm3 # 698c <_sk_callback_avx+0x42c> DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1 DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 @@ -9517,29 +9374,29 @@ _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 49f2 <_sk_load_f16_avx+0x31a> + DB 116,79 ; je 482a <_sk_load_f16_avx+0x31a> 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 49f2 <_sk_load_f16_avx+0x31a> + DB 114,67 ; jb 482a <_sk_load_f16_avx+0x31a> 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 49ff <_sk_load_f16_avx+0x327> + DB 116,68 ; je 4837 <_sk_load_f16_avx+0x327> 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 49ff <_sk_load_f16_avx+0x327> + DB 114,56 ; jb 4837 <_sk_load_f16_avx+0x327> DB 197,251,16,76,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm1 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,68,253,255,255 ; je 471b <_sk_load_f16_avx+0x43> + DB 15,132,68,253,255,255 ; je 4553 <_sk_load_f16_avx+0x43> DB 197,241,22,76,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,52,253,255,255 ; jb 471b <_sk_load_f16_avx+0x43> + DB 15,130,52,253,255,255 ; jb 4553 <_sk_load_f16_avx+0x43> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,41,253,255,255 ; jmpq 471b <_sk_load_f16_avx+0x43> + DB 233,41,253,255,255 ; jmpq 4553 <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,28,253,255,255 ; jmpq 471b <_sk_load_f16_avx+0x43> + DB 233,28,253,255,255 ; jmpq 4553 <_sk_load_f16_avx+0x43> DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1 - DB 233,19,253,255,255 ; jmpq 471b <_sk_load_f16_avx+0x43> + DB 233,19,253,255,255 ; jmpq 4553 <_sk_load_f16_avx+0x43> PUBLIC _sk_gather_f16_avx _sk_gather_f16_avx LABEL PROC @@ -9601,13 +9458,13 @@ _sk_gather_f16_avx LABEL PROC DB 197,249,105,210 ; vpunpckhwd %xmm2,%xmm0,%xmm2 DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 - DB 196,98,125,24,37,67,32,0,0 ; vbroadcastss 0x2043(%rip),%ymm12 # 6b58 <_sk_callback_avx+0x430> + DB 196,98,125,24,37,67,32,0,0 ; vbroadcastss 0x2043(%rip),%ymm12 # 6990 <_sk_callback_avx+0x430> DB 196,193,124,84,212 ; vandps %ymm12,%ymm0,%ymm2 DB 197,252,87,194 ; vxorps %ymm2,%ymm0,%ymm0 DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14 - DB 196,98,121,24,29,47,32,0,0 ; vbroadcastss 0x202f(%rip),%xmm11 # 6b5c <_sk_callback_avx+0x434> + DB 196,98,121,24,29,47,32,0,0 ; vbroadcastss 0x202f(%rip),%xmm11 # 6994 <_sk_callback_avx+0x434> DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3 - DB 196,98,121,24,45,37,32,0,0 ; vbroadcastss 0x2025(%rip),%xmm13 # 6b60 <_sk_callback_avx+0x438> + DB 196,98,121,24,45,37,32,0,0 ; vbroadcastss 0x2025(%rip),%xmm13 # 6998 <_sk_callback_avx+0x438> DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3 DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10 DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10 @@ -9621,7 +9478,7 @@ _sk_gather_f16_avx LABEL PROC DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0 DB 197,252,86,194 ; vorps %ymm2,%ymm0,%ymm0 DB 196,227,125,25,194,1 ; vextractf128 $0x1,%ymm0,%xmm2 - DB 196,226,121,24,29,219,31,0,0 ; vbroadcastss 0x1fdb(%rip),%xmm3 # 6b64 <_sk_callback_avx+0x43c> + DB 196,226,121,24,29,219,31,0,0 ; vbroadcastss 0x1fdb(%rip),%xmm3 # 699c <_sk_callback_avx+0x43c> DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0 DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 @@ -9723,12 +9580,12 @@ _sk_store_f16_avx LABEL PROC DB 197,252,17,180,36,128,0,0,0 ; vmovups %ymm6,0x80(%rsp) DB 197,252,17,108,36,96 ; vmovups %ymm5,0x60(%rsp) DB 197,252,17,100,36,64 ; vmovups %ymm4,0x40(%rsp) - DB 196,98,125,24,13,232,29,0,0 ; vbroadcastss 0x1de8(%rip),%ymm9 # 6b68 <_sk_callback_avx+0x440> + DB 196,98,125,24,13,232,29,0,0 ; vbroadcastss 0x1de8(%rip),%ymm9 # 69a0 <_sk_callback_avx+0x440> DB 196,65,124,84,209 ; vandps %ymm9,%ymm0,%ymm10 DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp) DB 196,65,124,87,218 ; vxorps %ymm10,%ymm0,%ymm11 DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12 - DB 196,98,121,24,5,206,29,0,0 ; vbroadcastss 0x1dce(%rip),%xmm8 # 6b6c <_sk_callback_avx+0x444> + DB 196,98,121,24,5,206,29,0,0 ; vbroadcastss 0x1dce(%rip),%xmm8 # 69a4 <_sk_callback_avx+0x444> DB 196,65,57,102,236 ; vpcmpgtd %xmm12,%xmm8,%xmm13 DB 196,65,57,102,243 ; vpcmpgtd %xmm11,%xmm8,%xmm14 DB 196,67,13,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm14,%ymm13 @@ -9738,7 +9595,7 @@ _sk_store_f16_avx LABEL PROC DB 196,67,13,24,242,1 ; vinsertf128 $0x1,%xmm10,%ymm14,%ymm14 DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11 DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12 - DB 196,98,125,24,21,149,29,0,0 ; vbroadcastss 0x1d95(%rip),%ymm10 # 6b70 <_sk_callback_avx+0x448> + DB 196,98,125,24,21,149,29,0,0 ; vbroadcastss 0x1d95(%rip),%ymm10 # 69a8 <_sk_callback_avx+0x448> DB 196,65,12,86,242 ; vorps %ymm10,%ymm14,%ymm14 DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15 DB 196,65,1,254,228 ; vpaddd %xmm12,%xmm15,%xmm12 @@ -9820,7 +9677,7 @@ _sk_store_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 117,75 ; jne 4fc2 <_sk_store_f16_avx+0x270> + DB 117,75 ; jne 4dfa <_sk_store_f16_avx+0x270> 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) @@ -9836,22 +9693,22 @@ _sk_store_f16_avx 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 116,193 ; je 4f8e <_sk_store_f16_avx+0x23c> + DB 116,193 ; je 4dc6 <_sk_store_f16_avx+0x23c> DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,181 ; jb 4f8e <_sk_store_f16_avx+0x23c> + DB 114,181 ; jb 4dc6 <_sk_store_f16_avx+0x23c> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,173 ; je 4f8e <_sk_store_f16_avx+0x23c> + DB 116,173 ; je 4dc6 <_sk_store_f16_avx+0x23c> DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,161 ; jb 4f8e <_sk_store_f16_avx+0x23c> + DB 114,161 ; jb 4dc6 <_sk_store_f16_avx+0x23c> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,153 ; je 4f8e <_sk_store_f16_avx+0x23c> + DB 116,153 ; je 4dc6 <_sk_store_f16_avx+0x23c> DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,141 ; jb 4f8e <_sk_store_f16_avx+0x23c> + DB 114,141 ; jb 4dc6 <_sk_store_f16_avx+0x23c> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,133 ; jmp 4f8e <_sk_store_f16_avx+0x23c> + DB 235,133 ; jmp 4dc6 <_sk_store_f16_avx+0x23c> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC @@ -9859,7 +9716,7 @@ _sk_load_u16_be_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,253,0,0,0 ; jne 511c <_sk_load_u16_be_avx+0x113> + DB 15,133,253,0,0,0 ; jne 4f54 <_sk_load_u16_be_avx+0x113> DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8 DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2 DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3 @@ -9881,7 +9738,7 @@ _sk_load_u16_be_avx LABEL PROC DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,228,26,0,0 ; vbroadcastss 0x1ae4(%rip),%ymm11 # 6b74 <_sk_callback_avx+0x44c> + DB 196,98,125,24,29,228,26,0,0 ; vbroadcastss 0x1ae4(%rip),%ymm11 # 69ac <_sk_callback_avx+0x44c> DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,177,109,202 ; vpunpckhqdq %xmm2,%xmm9,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -9915,29 +9772,29 @@ _sk_load_u16_be_avx LABEL PROC DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,85 ; je 5182 <_sk_load_u16_be_avx+0x179> + DB 116,85 ; je 4fba <_sk_load_u16_be_avx+0x179> DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,72 ; jb 5182 <_sk_load_u16_be_avx+0x179> + DB 114,72 ; jb 4fba <_sk_load_u16_be_avx+0x179> DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,72 ; je 518f <_sk_load_u16_be_avx+0x186> + DB 116,72 ; je 4fc7 <_sk_load_u16_be_avx+0x186> DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,59 ; jb 518f <_sk_load_u16_be_avx+0x186> + DB 114,59 ; jb 4fc7 <_sk_load_u16_be_avx+0x186> DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,213,254,255,255 ; je 503a <_sk_load_u16_be_avx+0x31> + DB 15,132,213,254,255,255 ; je 4e72 <_sk_load_u16_be_avx+0x31> DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,196,254,255,255 ; jb 503a <_sk_load_u16_be_avx+0x31> + DB 15,130,196,254,255,255 ; jb 4e72 <_sk_load_u16_be_avx+0x31> DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9 - DB 233,184,254,255,255 ; jmpq 503a <_sk_load_u16_be_avx+0x31> + DB 233,184,254,255,255 ; jmpq 4e72 <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,171,254,255,255 ; jmpq 503a <_sk_load_u16_be_avx+0x31> + DB 233,171,254,255,255 ; jmpq 4e72 <_sk_load_u16_be_avx+0x31> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,162,254,255,255 ; jmpq 503a <_sk_load_u16_be_avx+0x31> + DB 233,162,254,255,255 ; jmpq 4e72 <_sk_load_u16_be_avx+0x31> PUBLIC _sk_load_rgb_u16_be_avx _sk_load_rgb_u16_be_avx LABEL PROC @@ -9945,7 +9802,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,243,0,0,0 ; jne 529d <_sk_load_rgb_u16_be_avx+0x105> + DB 15,133,243,0,0,0 ; jne 50d5 <_sk_load_rgb_u16_be_avx+0x105> DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0 DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2 DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1 @@ -9972,7 +9829,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0 DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 - DB 196,98,125,24,29,68,25,0,0 ; vbroadcastss 0x1944(%rip),%ymm11 # 6b78 <_sk_callback_avx+0x450> + DB 196,98,125,24,29,68,25,0,0 ; vbroadcastss 0x1944(%rip),%ymm11 # 69b0 <_sk_callback_avx+0x450> DB 196,193,124,89,195 ; vmulps %ymm11,%ymm0,%ymm0 DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1 DB 197,233,113,241,8 ; vpsllw $0x8,%xmm1,%xmm2 @@ -9993,48 +9850,48 @@ _sk_load_rgb_u16_be_avx LABEL PROC DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 DB 196,193,108,89,211 ; vmulps %ymm11,%ymm2,%ymm2 DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,29,225,24,0,0 ; vbroadcastss 0x18e1(%rip),%ymm3 # 6b7c <_sk_callback_avx+0x454> + DB 196,226,125,24,29,225,24,0,0 ; vbroadcastss 0x18e1(%rip),%ymm3 # 69b4 <_sk_callback_avx+0x454> DB 255,224 ; jmpq *%rax DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0 DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 117,5 ; jne 52b6 <_sk_load_rgb_u16_be_avx+0x11e> - DB 233,40,255,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> + DB 117,5 ; jne 50ee <_sk_load_rgb_u16_be_avx+0x11e> + DB 233,40,255,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1 DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,26 ; jb 52e5 <_sk_load_rgb_u16_be_avx+0x14d> + DB 114,26 ; jb 511d <_sk_load_rgb_u16_be_avx+0x14d> DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1 DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 117,10 ; jne 52ea <_sk_load_rgb_u16_be_avx+0x152> - DB 233,249,254,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> - DB 233,244,254,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 5122 <_sk_load_rgb_u16_be_avx+0x152> + DB 233,249,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,244,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1 DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,26 ; jb 5319 <_sk_load_rgb_u16_be_avx+0x181> + DB 114,26 ; jb 5151 <_sk_load_rgb_u16_be_avx+0x181> DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1 DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 117,10 ; jne 531e <_sk_load_rgb_u16_be_avx+0x186> - DB 233,197,254,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> - DB 233,192,254,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> + DB 117,10 ; jne 5156 <_sk_load_rgb_u16_be_avx+0x186> + DB 233,197,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,192,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3 DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,20 ; jb 5347 <_sk_load_rgb_u16_be_avx+0x1af> + DB 114,20 ; jb 517f <_sk_load_rgb_u16_be_avx+0x1af> DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3 DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 - DB 233,151,254,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> - DB 233,146,254,255,255 ; jmpq 51de <_sk_load_rgb_u16_be_avx+0x46> + DB 233,151,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> + DB 233,146,254,255,255 ; jmpq 5016 <_sk_load_rgb_u16_be_avx+0x46> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,0 ; mov (%rax),%r8 DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax - DB 196,98,125,24,5,30,24,0,0 ; vbroadcastss 0x181e(%rip),%ymm8 # 6b80 <_sk_callback_avx+0x458> + DB 196,98,125,24,5,30,24,0,0 ; vbroadcastss 0x181e(%rip),%ymm8 # 69b8 <_sk_callback_avx+0x458> DB 196,65,124,89,200 ; vmulps %ymm8,%ymm0,%ymm9 DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9 DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10 @@ -10072,7 +9929,7 @@ _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 5446 <_sk_store_u16_be_avx+0xfa> + DB 117,31 ; jne 527e <_sk_store_u16_be_avx+0xfa> DB 196,65,120,17,28,64 ; vmovups %xmm11,(%r8,%rax,2) DB 196,65,120,17,84,64,16 ; vmovups %xmm10,0x10(%r8,%rax,2) DB 196,65,120,17,76,64,32 ; vmovups %xmm9,0x20(%r8,%rax,2) @@ -10081,31 +9938,31 @@ _sk_store_u16_be_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,64 ; vmovq %xmm11,(%r8,%rax,2) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 5442 <_sk_store_u16_be_avx+0xf6> + DB 116,240 ; je 527a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,23,92,64,8 ; vmovhpd %xmm11,0x8(%r8,%rax,2) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 5442 <_sk_store_u16_be_avx+0xf6> + DB 114,227 ; jb 527a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,84,64,16 ; vmovq %xmm10,0x10(%r8,%rax,2) - DB 116,218 ; je 5442 <_sk_store_u16_be_avx+0xf6> + DB 116,218 ; je 527a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,23,84,64,24 ; vmovhpd %xmm10,0x18(%r8,%rax,2) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 5442 <_sk_store_u16_be_avx+0xf6> + DB 114,205 ; jb 527a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,76,64,32 ; vmovq %xmm9,0x20(%r8,%rax,2) - DB 116,196 ; je 5442 <_sk_store_u16_be_avx+0xf6> + DB 116,196 ; je 527a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,23,76,64,40 ; vmovhpd %xmm9,0x28(%r8,%rax,2) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 5442 <_sk_store_u16_be_avx+0xf6> + DB 114,183 ; jb 527a <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,68,64,48 ; vmovq %xmm8,0x30(%r8,%rax,2) - DB 235,174 ; jmp 5442 <_sk_store_u16_be_avx+0xf6> + DB 235,174 ; jmp 527a <_sk_store_u16_be_avx+0xf6> PUBLIC _sk_load_f32_avx _sk_load_f32_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,131,249,7 ; cmp $0x7,%rcx - DB 119,110 ; ja 550a <_sk_load_f32_avx+0x76> + DB 119,110 ; ja 5342 <_sk_load_f32_avx+0x76> DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 5534 <_sk_load_f32_avx+0xa0> + DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 536c <_sk_load_f32_avx+0xa0> DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -10162,7 +10019,7 @@ _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 55c1 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 53f9 <_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 @@ -10175,22 +10032,22 @@ _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 55bd <_sk_store_f32_avx+0x69> + DB 116,240 ; je 53f5 <_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 55bd <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 53f5 <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 55bd <_sk_store_f32_avx+0x69> + DB 116,218 ; je 53f5 <_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 55bd <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 53f5 <_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 55bd <_sk_store_f32_avx+0x69> + DB 116,195 ; je 53f5 <_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 55bd <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 53f5 <_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 55bd <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 53f5 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -10280,7 +10137,7 @@ PUBLIC _sk_clamp_x_1_avx _sk_clamp_x_1_avx LABEL PROC DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8 DB 197,188,95,192 ; vmaxps %ymm0,%ymm8,%ymm0 - DB 196,98,125,24,5,78,20,0,0 ; vbroadcastss 0x144e(%rip),%ymm8 # 6b84 <_sk_callback_avx+0x45c> + DB 196,98,125,24,5,78,20,0,0 ; vbroadcastss 0x144e(%rip),%ymm8 # 69bc <_sk_callback_avx+0x45c> DB 196,193,124,93,192 ; vminps %ymm8,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -10294,9 +10151,9 @@ _sk_repeat_x_1_avx LABEL PROC PUBLIC _sk_mirror_x_1_avx _sk_mirror_x_1_avx LABEL PROC - DB 196,98,125,24,5,49,20,0,0 ; vbroadcastss 0x1431(%rip),%ymm8 # 6b88 <_sk_callback_avx+0x460> + DB 196,98,125,24,5,49,20,0,0 ; vbroadcastss 0x1431(%rip),%ymm8 # 69c0 <_sk_callback_avx+0x460> DB 196,193,124,88,192 ; vaddps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,13,39,20,0,0 ; vbroadcastss 0x1427(%rip),%ymm9 # 6b8c <_sk_callback_avx+0x464> + DB 196,98,125,24,13,39,20,0,0 ; vbroadcastss 0x1427(%rip),%ymm9 # 69c4 <_sk_callback_avx+0x464> DB 196,65,124,89,201 ; vmulps %ymm9,%ymm0,%ymm9 DB 196,67,125,8,201,1 ; vroundps $0x1,%ymm9,%ymm9 DB 196,65,52,88,201 ; vaddps %ymm9,%ymm9,%ymm9 @@ -10310,12 +10167,12 @@ _sk_mirror_x_1_avx LABEL PROC PUBLIC _sk_luminance_to_alpha_avx _sk_luminance_to_alpha_avx LABEL PROC - DB 196,226,125,24,29,247,19,0,0 ; vbroadcastss 0x13f7(%rip),%ymm3 # 6b90 <_sk_callback_avx+0x468> + DB 196,226,125,24,29,247,19,0,0 ; vbroadcastss 0x13f7(%rip),%ymm3 # 69c8 <_sk_callback_avx+0x468> DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,29,238,19,0,0 ; vbroadcastss 0x13ee(%rip),%ymm3 # 6b94 <_sk_callback_avx+0x46c> + DB 196,226,125,24,29,238,19,0,0 ; vbroadcastss 0x13ee(%rip),%ymm3 # 69cc <_sk_callback_avx+0x46c> DB 197,244,89,203 ; vmulps %ymm3,%ymm1,%ymm1 DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 - DB 196,226,125,24,13,225,19,0,0 ; vbroadcastss 0x13e1(%rip),%ymm1 # 6b98 <_sk_callback_avx+0x470> + DB 196,226,125,24,13,225,19,0,0 ; vbroadcastss 0x13e1(%rip),%ymm1 # 69d0 <_sk_callback_avx+0x470> DB 197,236,89,201 ; vmulps %ymm1,%ymm2,%ymm1 DB 197,252,88,217 ; vaddps %ymm1,%ymm0,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -10522,9 +10379,9 @@ _sk_evenly_spaced_gradient_avx LABEL PROC DB 72,139,24 ; mov (%rax),%rbx DB 72,139,104,8 ; mov 0x8(%rax),%rbp DB 72,255,203 ; dec %rbx - DB 120,7 ; js 5b0d <_sk_evenly_spaced_gradient_avx+0x1f> + DB 120,7 ; js 5945 <_sk_evenly_spaced_gradient_avx+0x1f> DB 196,225,242,42,203 ; vcvtsi2ss %rbx,%xmm1,%xmm1 - DB 235,21 ; jmp 5b22 <_sk_evenly_spaced_gradient_avx+0x34> + DB 235,21 ; jmp 595a <_sk_evenly_spaced_gradient_avx+0x34> DB 73,137,216 ; mov %rbx,%r8 DB 73,209,232 ; shr %r8 DB 131,227,1 ; and $0x1,%ebx @@ -10679,18 +10536,18 @@ _sk_evenly_spaced_gradient_avx LABEL PROC PUBLIC _sk_gauss_a_to_rgba_avx _sk_gauss_a_to_rgba_avx LABEL PROC - DB 196,226,125,24,5,76,13,0,0 ; vbroadcastss 0xd4c(%rip),%ymm0 # 6b9c <_sk_callback_avx+0x474> + DB 196,226,125,24,5,76,13,0,0 ; vbroadcastss 0xd4c(%rip),%ymm0 # 69d4 <_sk_callback_avx+0x474> DB 197,228,89,192 ; vmulps %ymm0,%ymm3,%ymm0 - DB 196,226,125,24,13,67,13,0,0 ; vbroadcastss 0xd43(%rip),%ymm1 # 6ba0 <_sk_callback_avx+0x478> + DB 196,226,125,24,13,67,13,0,0 ; vbroadcastss 0xd43(%rip),%ymm1 # 69d8 <_sk_callback_avx+0x478> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,54,13,0,0 ; vbroadcastss 0xd36(%rip),%ymm1 # 6ba4 <_sk_callback_avx+0x47c> + DB 196,226,125,24,13,54,13,0,0 ; vbroadcastss 0xd36(%rip),%ymm1 # 69dc <_sk_callback_avx+0x47c> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,41,13,0,0 ; vbroadcastss 0xd29(%rip),%ymm1 # 6ba8 <_sk_callback_avx+0x480> + DB 196,226,125,24,13,41,13,0,0 ; vbroadcastss 0xd29(%rip),%ymm1 # 69e0 <_sk_callback_avx+0x480> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 197,252,89,195 ; vmulps %ymm3,%ymm0,%ymm0 - DB 196,226,125,24,13,28,13,0,0 ; vbroadcastss 0xd1c(%rip),%ymm1 # 6bac <_sk_callback_avx+0x484> + DB 196,226,125,24,13,28,13,0,0 ; vbroadcastss 0xd1c(%rip),%ymm1 # 69e4 <_sk_callback_avx+0x484> DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0 DB 72,173 ; lods %ds:(%rsi),%rax DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 @@ -10710,12 +10567,12 @@ _sk_gradient_avx LABEL PROC DB 76,139,0 ; mov (%rax),%r8 DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 73,131,248,2 ; cmp $0x2,%r8 - DB 114,80 ; jb 5f0d <_sk_gradient_avx+0x69> + DB 114,80 ; jb 5d45 <_sk_gradient_avx+0x69> DB 72,139,88,72 ; mov 0x48(%rax),%rbx DB 73,255,200 ; dec %r8 DB 72,131,195,4 ; add $0x4,%rbx DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 - DB 196,98,125,24,21,218,12,0,0 ; vbroadcastss 0xcda(%rip),%ymm10 # 6bb0 <_sk_callback_avx+0x488> + DB 196,98,125,24,21,218,12,0,0 ; vbroadcastss 0xcda(%rip),%ymm10 # 69e8 <_sk_callback_avx+0x488> DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1 DB 196,98,125,24,3 ; vbroadcastss (%rbx),%ymm8 DB 197,60,194,192,2 ; vcmpleps %ymm0,%ymm8,%ymm8 @@ -10727,7 +10584,7 @@ _sk_gradient_avx LABEL PROC DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 DB 72,131,195,4 ; add $0x4,%rbx DB 73,255,200 ; dec %r8 - DB 117,205 ; jne 5eda <_sk_gradient_avx+0x36> + DB 117,205 ; jne 5d12 <_sk_gradient_avx+0x36> DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8 DB 69,137,193 ; mov %r8d,%r9d DB 73,193,232,32 ; shr $0x20,%r8 @@ -10905,27 +10762,27 @@ _sk_xy_to_unit_angle_avx LABEL PROC DB 196,65,52,95,226 ; vmaxps %ymm10,%ymm9,%ymm12 DB 196,65,36,94,220 ; vdivps %ymm12,%ymm11,%ymm11 DB 196,65,36,89,227 ; vmulps %ymm11,%ymm11,%ymm12 - DB 196,98,125,24,45,254,8,0,0 ; vbroadcastss 0x8fe(%rip),%ymm13 # 6bb4 <_sk_callback_avx+0x48c> + DB 196,98,125,24,45,254,8,0,0 ; vbroadcastss 0x8fe(%rip),%ymm13 # 69ec <_sk_callback_avx+0x48c> DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,244,8,0,0 ; vbroadcastss 0x8f4(%rip),%ymm14 # 6bb8 <_sk_callback_avx+0x490> + DB 196,98,125,24,53,244,8,0,0 ; vbroadcastss 0x8f4(%rip),%ymm14 # 69f0 <_sk_callback_avx+0x490> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,237 ; vmulps %ymm13,%ymm12,%ymm13 - DB 196,98,125,24,53,229,8,0,0 ; vbroadcastss 0x8e5(%rip),%ymm14 # 6bbc <_sk_callback_avx+0x494> + DB 196,98,125,24,53,229,8,0,0 ; vbroadcastss 0x8e5(%rip),%ymm14 # 69f4 <_sk_callback_avx+0x494> DB 196,65,20,88,238 ; vaddps %ymm14,%ymm13,%ymm13 DB 196,65,28,89,229 ; vmulps %ymm13,%ymm12,%ymm12 - DB 196,98,125,24,45,214,8,0,0 ; vbroadcastss 0x8d6(%rip),%ymm13 # 6bc0 <_sk_callback_avx+0x498> + DB 196,98,125,24,45,214,8,0,0 ; vbroadcastss 0x8d6(%rip),%ymm13 # 69f8 <_sk_callback_avx+0x498> DB 196,65,28,88,229 ; vaddps %ymm13,%ymm12,%ymm12 DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11 DB 196,65,52,194,202,1 ; vcmpltps %ymm10,%ymm9,%ymm9 - DB 196,98,125,24,21,193,8,0,0 ; vbroadcastss 0x8c1(%rip),%ymm10 # 6bc4 <_sk_callback_avx+0x49c> + DB 196,98,125,24,21,193,8,0,0 ; vbroadcastss 0x8c1(%rip),%ymm10 # 69fc <_sk_callback_avx+0x49c> DB 196,65,44,92,211 ; vsubps %ymm11,%ymm10,%ymm10 DB 196,67,37,74,202,144 ; vblendvps %ymm9,%ymm10,%ymm11,%ymm9 DB 196,193,124,194,192,1 ; vcmpltps %ymm8,%ymm0,%ymm0 - DB 196,98,125,24,21,171,8,0,0 ; vbroadcastss 0x8ab(%rip),%ymm10 # 6bc8 <_sk_callback_avx+0x4a0> + DB 196,98,125,24,21,171,8,0,0 ; vbroadcastss 0x8ab(%rip),%ymm10 # 6a00 <_sk_callback_avx+0x4a0> DB 196,65,44,92,209 ; vsubps %ymm9,%ymm10,%ymm10 DB 196,195,53,74,194,0 ; vblendvps %ymm0,%ymm10,%ymm9,%ymm0 DB 196,65,116,194,200,1 ; vcmpltps %ymm8,%ymm1,%ymm9 - DB 196,98,125,24,21,149,8,0,0 ; vbroadcastss 0x895(%rip),%ymm10 # 6bcc <_sk_callback_avx+0x4a4> + DB 196,98,125,24,21,149,8,0,0 ; vbroadcastss 0x895(%rip),%ymm10 # 6a04 <_sk_callback_avx+0x4a4> DB 197,44,92,208 ; vsubps %ymm0,%ymm10,%ymm10 DB 196,195,125,74,194,144 ; vblendvps %ymm9,%ymm10,%ymm0,%ymm0 DB 196,65,124,194,200,3 ; vcmpunordps %ymm8,%ymm0,%ymm9 @@ -10945,7 +10802,7 @@ _sk_xy_to_radius_avx LABEL PROC PUBLIC _sk_save_xy_avx _sk_save_xy_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,95,8,0,0 ; vbroadcastss 0x85f(%rip),%ymm8 # 6bd0 <_sk_callback_avx+0x4a8> + DB 196,98,125,24,5,95,8,0,0 ; vbroadcastss 0x85f(%rip),%ymm8 # 6a08 <_sk_callback_avx+0x4a8> DB 196,65,124,88,200 ; vaddps %ymm8,%ymm0,%ymm9 DB 196,67,125,8,209,1 ; vroundps $0x1,%ymm9,%ymm10 DB 196,65,52,92,202 ; vsubps %ymm10,%ymm9,%ymm9 @@ -10978,9 +10835,9 @@ _sk_accumulate_avx LABEL PROC PUBLIC _sk_bilinear_nx_avx _sk_bilinear_nx_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,235,7,0,0 ; vbroadcastss 0x7eb(%rip),%ymm0 # 6bd4 <_sk_callback_avx+0x4ac> + DB 196,226,125,24,5,235,7,0,0 ; vbroadcastss 0x7eb(%rip),%ymm0 # 6a0c <_sk_callback_avx+0x4ac> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,226,7,0,0 ; vbroadcastss 0x7e2(%rip),%ymm8 # 6bd8 <_sk_callback_avx+0x4b0> + DB 196,98,125,24,5,226,7,0,0 ; vbroadcastss 0x7e2(%rip),%ymm8 # 6a10 <_sk_callback_avx+0x4b0> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -10989,7 +10846,7 @@ _sk_bilinear_nx_avx LABEL PROC PUBLIC _sk_bilinear_px_avx _sk_bilinear_px_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,202,7,0,0 ; vbroadcastss 0x7ca(%rip),%ymm0 # 6bdc <_sk_callback_avx+0x4b4> + DB 196,226,125,24,5,202,7,0,0 ; vbroadcastss 0x7ca(%rip),%ymm0 # 6a14 <_sk_callback_avx+0x4b4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -10999,9 +10856,9 @@ _sk_bilinear_px_avx LABEL PROC PUBLIC _sk_bilinear_ny_avx _sk_bilinear_ny_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,174,7,0,0 ; vbroadcastss 0x7ae(%rip),%ymm1 # 6be0 <_sk_callback_avx+0x4b8> + DB 196,226,125,24,13,174,7,0,0 ; vbroadcastss 0x7ae(%rip),%ymm1 # 6a18 <_sk_callback_avx+0x4b8> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,164,7,0,0 ; vbroadcastss 0x7a4(%rip),%ymm8 # 6be4 <_sk_callback_avx+0x4bc> + DB 196,98,125,24,5,164,7,0,0 ; vbroadcastss 0x7a4(%rip),%ymm8 # 6a1c <_sk_callback_avx+0x4bc> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11010,7 +10867,7 @@ _sk_bilinear_ny_avx LABEL PROC PUBLIC _sk_bilinear_py_avx _sk_bilinear_py_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,140,7,0,0 ; vbroadcastss 0x78c(%rip),%ymm1 # 6be8 <_sk_callback_avx+0x4c0> + DB 196,226,125,24,13,140,7,0,0 ; vbroadcastss 0x78c(%rip),%ymm1 # 6a20 <_sk_callback_avx+0x4c0> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -11020,14 +10877,14 @@ _sk_bilinear_py_avx LABEL PROC PUBLIC _sk_bicubic_n3x_avx _sk_bicubic_n3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,111,7,0,0 ; vbroadcastss 0x76f(%rip),%ymm0 # 6bec <_sk_callback_avx+0x4c4> + DB 196,226,125,24,5,111,7,0,0 ; vbroadcastss 0x76f(%rip),%ymm0 # 6a24 <_sk_callback_avx+0x4c4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,102,7,0,0 ; vbroadcastss 0x766(%rip),%ymm8 # 6bf0 <_sk_callback_avx+0x4c8> + DB 196,98,125,24,5,102,7,0,0 ; vbroadcastss 0x766(%rip),%ymm8 # 6a28 <_sk_callback_avx+0x4c8> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,87,7,0,0 ; vbroadcastss 0x757(%rip),%ymm10 # 6bf4 <_sk_callback_avx+0x4cc> + DB 196,98,125,24,21,87,7,0,0 ; vbroadcastss 0x757(%rip),%ymm10 # 6a2c <_sk_callback_avx+0x4cc> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,77,7,0,0 ; vbroadcastss 0x74d(%rip),%ymm10 # 6bf8 <_sk_callback_avx+0x4d0> + DB 196,98,125,24,21,77,7,0,0 ; vbroadcastss 0x74d(%rip),%ymm10 # 6a30 <_sk_callback_avx+0x4d0> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -11037,19 +10894,19 @@ _sk_bicubic_n3x_avx LABEL PROC PUBLIC _sk_bicubic_n1x_avx _sk_bicubic_n1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,48,7,0,0 ; vbroadcastss 0x730(%rip),%ymm0 # 6bfc <_sk_callback_avx+0x4d4> + DB 196,226,125,24,5,48,7,0,0 ; vbroadcastss 0x730(%rip),%ymm0 # 6a34 <_sk_callback_avx+0x4d4> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 - DB 196,98,125,24,5,39,7,0,0 ; vbroadcastss 0x727(%rip),%ymm8 # 6c00 <_sk_callback_avx+0x4d8> + DB 196,98,125,24,5,39,7,0,0 ; vbroadcastss 0x727(%rip),%ymm8 # 6a38 <_sk_callback_avx+0x4d8> DB 197,60,92,64,64 ; vsubps 0x40(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,29,7,0,0 ; vbroadcastss 0x71d(%rip),%ymm9 # 6c04 <_sk_callback_avx+0x4dc> + DB 196,98,125,24,13,29,7,0,0 ; vbroadcastss 0x71d(%rip),%ymm9 # 6a3c <_sk_callback_avx+0x4dc> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,19,7,0,0 ; vbroadcastss 0x713(%rip),%ymm10 # 6c08 <_sk_callback_avx+0x4e0> + DB 196,98,125,24,21,19,7,0,0 ; vbroadcastss 0x713(%rip),%ymm10 # 6a40 <_sk_callback_avx+0x4e0> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,4,7,0,0 ; vbroadcastss 0x704(%rip),%ymm10 # 6c0c <_sk_callback_avx+0x4e4> + DB 196,98,125,24,21,4,7,0,0 ; vbroadcastss 0x704(%rip),%ymm10 # 6a44 <_sk_callback_avx+0x4e4> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,245,6,0,0 ; vbroadcastss 0x6f5(%rip),%ymm9 # 6c10 <_sk_callback_avx+0x4e8> + DB 196,98,125,24,13,245,6,0,0 ; vbroadcastss 0x6f5(%rip),%ymm9 # 6a48 <_sk_callback_avx+0x4e8> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11058,17 +10915,17 @@ _sk_bicubic_n1x_avx LABEL PROC PUBLIC _sk_bicubic_p1x_avx _sk_bicubic_p1x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,221,6,0,0 ; vbroadcastss 0x6dd(%rip),%ymm8 # 6c14 <_sk_callback_avx+0x4ec> + DB 196,98,125,24,5,221,6,0,0 ; vbroadcastss 0x6dd(%rip),%ymm8 # 6a4c <_sk_callback_avx+0x4ec> DB 197,188,88,0 ; vaddps (%rax),%ymm8,%ymm0 DB 197,124,16,72,64 ; vmovups 0x40(%rax),%ymm9 - DB 196,98,125,24,21,207,6,0,0 ; vbroadcastss 0x6cf(%rip),%ymm10 # 6c18 <_sk_callback_avx+0x4f0> + DB 196,98,125,24,21,207,6,0,0 ; vbroadcastss 0x6cf(%rip),%ymm10 # 6a50 <_sk_callback_avx+0x4f0> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,197,6,0,0 ; vbroadcastss 0x6c5(%rip),%ymm11 # 6c1c <_sk_callback_avx+0x4f4> + DB 196,98,125,24,29,197,6,0,0 ; vbroadcastss 0x6c5(%rip),%ymm11 # 6a54 <_sk_callback_avx+0x4f4> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,172,6,0,0 ; vbroadcastss 0x6ac(%rip),%ymm9 # 6c20 <_sk_callback_avx+0x4f8> + DB 196,98,125,24,13,172,6,0,0 ; vbroadcastss 0x6ac(%rip),%ymm9 # 6a58 <_sk_callback_avx+0x4f8> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11077,13 +10934,13 @@ _sk_bicubic_p1x_avx LABEL PROC PUBLIC _sk_bicubic_p3x_avx _sk_bicubic_p3x_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,148,6,0,0 ; vbroadcastss 0x694(%rip),%ymm0 # 6c24 <_sk_callback_avx+0x4fc> + DB 196,226,125,24,5,148,6,0,0 ; vbroadcastss 0x694(%rip),%ymm0 # 6a5c <_sk_callback_avx+0x4fc> DB 197,252,88,0 ; vaddps (%rax),%ymm0,%ymm0 DB 197,124,16,64,64 ; vmovups 0x40(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm10 # 6c28 <_sk_callback_avx+0x500> + DB 196,98,125,24,21,129,6,0,0 ; vbroadcastss 0x681(%rip),%ymm10 # 6a60 <_sk_callback_avx+0x500> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,119,6,0,0 ; vbroadcastss 0x677(%rip),%ymm10 # 6c2c <_sk_callback_avx+0x504> + DB 196,98,125,24,21,119,6,0,0 ; vbroadcastss 0x677(%rip),%ymm10 # 6a64 <_sk_callback_avx+0x504> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,128,0,0,0 ; vmovups %ymm8,0x80(%rax) @@ -11093,14 +10950,14 @@ _sk_bicubic_p3x_avx LABEL PROC PUBLIC _sk_bicubic_n3y_avx _sk_bicubic_n3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,90,6,0,0 ; vbroadcastss 0x65a(%rip),%ymm1 # 6c30 <_sk_callback_avx+0x508> + DB 196,226,125,24,13,90,6,0,0 ; vbroadcastss 0x65a(%rip),%ymm1 # 6a68 <_sk_callback_avx+0x508> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,80,6,0,0 ; vbroadcastss 0x650(%rip),%ymm8 # 6c34 <_sk_callback_avx+0x50c> + DB 196,98,125,24,5,80,6,0,0 ; vbroadcastss 0x650(%rip),%ymm8 # 6a6c <_sk_callback_avx+0x50c> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,65,6,0,0 ; vbroadcastss 0x641(%rip),%ymm10 # 6c38 <_sk_callback_avx+0x510> + DB 196,98,125,24,21,65,6,0,0 ; vbroadcastss 0x641(%rip),%ymm10 # 6a70 <_sk_callback_avx+0x510> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,55,6,0,0 ; vbroadcastss 0x637(%rip),%ymm10 # 6c3c <_sk_callback_avx+0x514> + DB 196,98,125,24,21,55,6,0,0 ; vbroadcastss 0x637(%rip),%ymm10 # 6a74 <_sk_callback_avx+0x514> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -11110,19 +10967,19 @@ _sk_bicubic_n3y_avx LABEL PROC PUBLIC _sk_bicubic_n1y_avx _sk_bicubic_n1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,26,6,0,0 ; vbroadcastss 0x61a(%rip),%ymm1 # 6c40 <_sk_callback_avx+0x518> + DB 196,226,125,24,13,26,6,0,0 ; vbroadcastss 0x61a(%rip),%ymm1 # 6a78 <_sk_callback_avx+0x518> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 - DB 196,98,125,24,5,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm8 # 6c44 <_sk_callback_avx+0x51c> + DB 196,98,125,24,5,16,6,0,0 ; vbroadcastss 0x610(%rip),%ymm8 # 6a7c <_sk_callback_avx+0x51c> DB 197,60,92,64,96 ; vsubps 0x60(%rax),%ymm8,%ymm8 - DB 196,98,125,24,13,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm9 # 6c48 <_sk_callback_avx+0x520> + DB 196,98,125,24,13,6,6,0,0 ; vbroadcastss 0x606(%rip),%ymm9 # 6a80 <_sk_callback_avx+0x520> DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm10 # 6c4c <_sk_callback_avx+0x524> + DB 196,98,125,24,21,252,5,0,0 ; vbroadcastss 0x5fc(%rip),%ymm10 # 6a84 <_sk_callback_avx+0x524> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,201 ; vmulps %ymm9,%ymm8,%ymm9 - DB 196,98,125,24,21,237,5,0,0 ; vbroadcastss 0x5ed(%rip),%ymm10 # 6c50 <_sk_callback_avx+0x528> + DB 196,98,125,24,21,237,5,0,0 ; vbroadcastss 0x5ed(%rip),%ymm10 # 6a88 <_sk_callback_avx+0x528> DB 196,65,52,88,202 ; vaddps %ymm10,%ymm9,%ymm9 DB 196,65,60,89,193 ; vmulps %ymm9,%ymm8,%ymm8 - DB 196,98,125,24,13,222,5,0,0 ; vbroadcastss 0x5de(%rip),%ymm9 # 6c54 <_sk_callback_avx+0x52c> + DB 196,98,125,24,13,222,5,0,0 ; vbroadcastss 0x5de(%rip),%ymm9 # 6a8c <_sk_callback_avx+0x52c> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11131,17 +10988,17 @@ _sk_bicubic_n1y_avx LABEL PROC PUBLIC _sk_bicubic_p1y_avx _sk_bicubic_p1y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,98,125,24,5,198,5,0,0 ; vbroadcastss 0x5c6(%rip),%ymm8 # 6c58 <_sk_callback_avx+0x530> + DB 196,98,125,24,5,198,5,0,0 ; vbroadcastss 0x5c6(%rip),%ymm8 # 6a90 <_sk_callback_avx+0x530> DB 197,188,88,72,32 ; vaddps 0x20(%rax),%ymm8,%ymm1 DB 197,124,16,72,96 ; vmovups 0x60(%rax),%ymm9 - DB 196,98,125,24,21,183,5,0,0 ; vbroadcastss 0x5b7(%rip),%ymm10 # 6c5c <_sk_callback_avx+0x534> + DB 196,98,125,24,21,183,5,0,0 ; vbroadcastss 0x5b7(%rip),%ymm10 # 6a94 <_sk_callback_avx+0x534> DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 - DB 196,98,125,24,29,173,5,0,0 ; vbroadcastss 0x5ad(%rip),%ymm11 # 6c60 <_sk_callback_avx+0x538> + DB 196,98,125,24,29,173,5,0,0 ; vbroadcastss 0x5ad(%rip),%ymm11 # 6a98 <_sk_callback_avx+0x538> DB 196,65,44,88,211 ; vaddps %ymm11,%ymm10,%ymm10 DB 196,65,52,89,210 ; vmulps %ymm10,%ymm9,%ymm10 DB 196,65,44,88,192 ; vaddps %ymm8,%ymm10,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 - DB 196,98,125,24,13,148,5,0,0 ; vbroadcastss 0x594(%rip),%ymm9 # 6c64 <_sk_callback_avx+0x53c> + DB 196,98,125,24,13,148,5,0,0 ; vbroadcastss 0x594(%rip),%ymm9 # 6a9c <_sk_callback_avx+0x53c> DB 196,65,60,88,193 ; vaddps %ymm9,%ymm8,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) DB 72,173 ; lods %ds:(%rsi),%rax @@ -11150,13 +11007,13 @@ _sk_bicubic_p1y_avx LABEL PROC PUBLIC _sk_bicubic_p3y_avx _sk_bicubic_p3y_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,13,124,5,0,0 ; vbroadcastss 0x57c(%rip),%ymm1 # 6c68 <_sk_callback_avx+0x540> + DB 196,226,125,24,13,124,5,0,0 ; vbroadcastss 0x57c(%rip),%ymm1 # 6aa0 <_sk_callback_avx+0x540> DB 197,244,88,72,32 ; vaddps 0x20(%rax),%ymm1,%ymm1 DB 197,124,16,64,96 ; vmovups 0x60(%rax),%ymm8 DB 196,65,60,89,200 ; vmulps %ymm8,%ymm8,%ymm9 - DB 196,98,125,24,21,104,5,0,0 ; vbroadcastss 0x568(%rip),%ymm10 # 6c6c <_sk_callback_avx+0x544> + DB 196,98,125,24,21,104,5,0,0 ; vbroadcastss 0x568(%rip),%ymm10 # 6aa4 <_sk_callback_avx+0x544> DB 196,65,60,89,194 ; vmulps %ymm10,%ymm8,%ymm8 - DB 196,98,125,24,21,94,5,0,0 ; vbroadcastss 0x55e(%rip),%ymm10 # 6c70 <_sk_callback_avx+0x548> + DB 196,98,125,24,21,94,5,0,0 ; vbroadcastss 0x55e(%rip),%ymm10 # 6aa8 <_sk_callback_avx+0x548> DB 196,65,60,88,194 ; vaddps %ymm10,%ymm8,%ymm8 DB 196,65,52,89,192 ; vmulps %ymm8,%ymm9,%ymm8 DB 197,124,17,128,160,0,0,0 ; vmovups %ymm8,0xa0(%rax) @@ -11270,25 +11127,25 @@ ALIGN 4 DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 68f5 <.literal4+0xb1> + DB 71,225,61 ; rex.RXB loope 672d <.literal4+0xb1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6905 <.literal4+0xc1> + DB 71,225,61 ; rex.RXB loope 673d <.literal4+0xc1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6915 <.literal4+0xd1> + DB 71,225,61 ; rex.RXB loope 674d <.literal4+0xd1> DB 0,0 ; add %al,(%rax) DB 128,63,154 ; cmpb $0x9a,(%rdi) DB 153 ; cltd DB 153 ; cltd DB 62,61,10,23,63,174 ; ds cmp $0xae3f170a,%eax - DB 71,225,61 ; rex.RXB loope 6925 <.literal4+0xe1> + DB 71,225,61 ; rex.RXB loope 675d <.literal4+0xe1> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,63,0,0,127 ; add %al,0x7f00003f(%rax) @@ -11340,7 +11197,7 @@ ALIGN 4 DB 190,129,128,128,59 ; mov $0x3b808081,%esi DB 129,128,128,59,0,248,0,0,8,33 ; addl $0x21080000,-0x7ffc480(%rax) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6979 <.literal4+0x135> + DB 224,7 ; loopne 67b1 <.literal4+0x135> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11356,10 +11213,10 @@ ALIGN 4 DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) DB 0,52,255 ; add %dh,(%rdi,%rdi,8) DB 255 ; (bad) - DB 127,0 ; jg 69a0 <.literal4+0x15c> + DB 127,0 ; jg 67d8 <.literal4+0x15c> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6a19 <.literal4+0x1d5> + DB 119,115 ; ja 6851 <.literal4+0x1d5> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11373,10 +11230,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 69d4 <.literal4+0x190> + DB 127,0 ; jg 680c <.literal4+0x190> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6a4d <.literal4+0x209> + DB 119,115 ; ja 6885 <.literal4+0x209> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11390,10 +11247,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 6a08 <.literal4+0x1c4> + DB 127,0 ; jg 6840 <.literal4+0x1c4> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6a81 <.literal4+0x23d> + DB 119,115 ; ja 68b9 <.literal4+0x23d> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11407,10 +11264,10 @@ ALIGN 4 DB 0,128,63,0,0,0 ; add %al,0x3f(%rax) DB 52,255 ; xor $0xff,%al DB 255 ; (bad) - DB 127,0 ; jg 6a3c <.literal4+0x1f8> + DB 127,0 ; jg 6874 <.literal4+0x1f8> DB 0,0 ; add %al,(%rax) DB 0,63 ; add %bh,(%rdi) - DB 119,115 ; ja 6ab5 <.literal4+0x271> + DB 119,115 ; ja 68ed <.literal4+0x271> DB 248 ; clc DB 194,117,191 ; retq $0xbf75 DB 191,63,249,68,180 ; mov $0xb444f93f,%edi @@ -11423,7 +11280,7 @@ ALIGN 4 DB 0,75,0 ; add %cl,0x0(%rbx) DB 0,128,63,0,0,200 ; add %al,-0x37ffffc1(%rax) DB 66,0,0 ; rex.X add %al,(%rax) - DB 127,67 ; jg 6ab3 <.literal4+0x26f> + DB 127,67 ; jg 68eb <.literal4+0x26f> DB 0,0 ; add %al,(%rax) DB 0,195 ; add %al,%bl DB 0,0 ; add %al,(%rax) @@ -11435,10 +11292,10 @@ ALIGN 4 DB 190,80,128,3,62 ; mov $0x3e038050,%esi DB 31 ; (bad) DB 215 ; xlat %ds:(%rbx) - DB 118,63 ; jbe 6ad3 <.literal4+0x28f> + DB 118,63 ; jbe 690b <.literal4+0x28f> DB 246,64,83,63 ; testb $0x3f,0x53(%rax) DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 6ae7 <.literal4+0x2a3> + DB 127,67 ; jg 691f <.literal4+0x2a3> DB 129,128,128,59,0,0,128,63,129,128 ; addl $0x80813f80,0x3b80(%rax) DB 128,59,0 ; cmpb $0x0,(%rbx) DB 0,128,63,129,128,128 ; add %al,-0x7f7f7ec1(%rax) @@ -11447,7 +11304,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6ac9 <.literal4+0x285> + DB 224,7 ; loopne 6901 <.literal4+0x285> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11459,7 +11316,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 8,33 ; or %ah,(%rcx) DB 132,55 ; test %dh,(%rdi) - DB 224,7 ; loopne 6ae5 <.literal4+0x2a1> + DB 224,7 ; loopne 691d <.literal4+0x2a1> DB 0,0 ; add %al,(%rax) DB 33,8 ; and %ecx,(%rax) DB 2,58 ; add (%rdx),%bh @@ -11470,7 +11327,7 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 248 ; clc DB 65,0,0 ; add %al,(%r8) - DB 124,66 ; jl 6b3a <.literal4+0x2f6> + DB 124,66 ; jl 6972 <.literal4+0x2f6> DB 0,240 ; add %dh,%al DB 0,0 ; add %al,(%rax) DB 137,136,136,55,0,15 ; mov %ecx,0xf003788(%rax) @@ -11488,9 +11345,9 @@ ALIGN 4 DB 137,136,136,59,15,0 ; mov %ecx,0xf3b88(%rax) DB 0,0 ; add %al,(%rax) DB 137,136,136,61,0,0 ; mov %ecx,0x3d88(%rax) - DB 112,65 ; jo 6b7d <.literal4+0x339> + DB 112,65 ; jo 69b5 <.literal4+0x339> DB 129,128,128,59,129,128,128,59,0,0 ; addl $0x3b80,-0x7f7ec480(%rax) - DB 127,67 ; jg 6b8b <.literal4+0x347> + DB 127,67 ; jg 69c3 <.literal4+0x347> DB 0,128,0,0,0,0 ; add %al,0x0(%rax) DB 0,128,0,4,0,128 ; add %al,-0x7ffffc00(%rax) DB 0,0 ; add %al,(%rax) @@ -11506,7 +11363,7 @@ ALIGN 4 DB 0,128,55,0,0,128 ; add %al,-0x7fffffc9(%rax) DB 63 ; (bad) DB 0,255 ; add %bh,%bh - DB 127,71 ; jg 6bcb <.literal4+0x387> + DB 127,71 ; jg 6a03 <.literal4+0x387> DB 0,0 ; add %al,(%rax) DB 128,63,0 ; cmpb $0x0,(%rdi) DB 0,128,191,0,0,0 ; add %al,0xbf(%rax) @@ -11602,39 +11459,89 @@ ALIGN 4 DB 170 ; stos %al,%es:(%rdi) DB 190 ; .byte 0xbe -ALIGN 32 - DB 255,0 ; incl (%rax) +ALIGN 16 + DB 0,2 ; add %al,(%rdx) + DB 4,6 ; add $0x6,%al DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) + DB 8,10 ; or %cl,(%rdx) + DB 12,14 ; or $0xe,%al DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) + DB 0,2 ; add %al,(%rdx) + DB 4,6 ; add $0x6,%al DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 8,10 ; or %cl,(%rdx) + DB 12,14 ; or $0xe,%al + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 255,0 ; incl (%rax) + DB 0,2 ; add %al,(%rdx) + DB 4,6 ; add $0x6,%al + DB 0,0 ; add %al,(%rax) DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 8,10 ; or %cl,(%rdx) + DB 12,14 ; or $0xe,%al + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,2 ; add %al,(%rdx) + DB 4,6 ; add $0x6,%al + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 8,10 ; or %cl,(%rdx) + DB 12,14 ; or $0xe,%al + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + DB 0,0 ; add %al,(%rax) + +ALIGN 32 DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) @@ -11683,24 +11590,38 @@ ALIGN 32 DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) DB 0,0 ; add %al,(%rax) - -ALIGN 16 DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) DB 255,0 ; incl (%rax) + DB 0,0 ; add %al,(%rax) ALIGN 32 PUBLIC _sk_start_pipeline_sse41 diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp index 4b506ef17b..e570af105b 100644 --- a/src/jumper/SkJumper_stages.cpp +++ b/src/jumper/SkJumper_stages.cpp @@ -176,8 +176,8 @@ SI void store(T* dst, V v, size_t tail) { } #endif -// AVX2 adds some mask loads and stores that make for shorter, faster code. -#if defined(JUMPER) && defined(__AVX2__) +// AVX adds some mask loads and stores that make for shorter, faster code. +#if defined(JUMPER) && defined(__AVX__) SI U32 mask(size_t tail) { // We go a little out of our way to avoid needing large constant values here. @@ -186,14 +186,16 @@ SI void store(T* dst, V v, size_t tail) { uint64_t mask = 0xffffffffffffffff >> 8*(kStride-tail); // Sign-extend each mask lane to its full width, 0x00000000 or 0xffffffff. - return _mm256_cvtepi8_epi32(_mm_cvtsi64_si128((int64_t)mask)); + using S8 = int8_t __attribute__((ext_vector_type(8))); + using S32 = int32_t __attribute__((ext_vector_type(8))); + return (U32)__builtin_convertvector(unaligned_load<S8>(&mask), S32); } template <> inline U32 load(const uint32_t* src, size_t tail) { __builtin_assume(tail < kStride); if (__builtin_expect(tail, 0)) { - return _mm256_maskload_epi32((const int*)src, mask(tail)); + return (U32)_mm256_maskload_ps((const float*)src, mask(tail)); } return unaligned_load<U32>(src); } @@ -202,7 +204,7 @@ SI void store(T* dst, V v, size_t tail) { inline void store(uint32_t* dst, U32 v, size_t tail) { __builtin_assume(tail < kStride); if (__builtin_expect(tail, 0)) { - return _mm256_maskstore_epi32((int*)dst, mask(tail), v); + return _mm256_maskstore_ps((float*)dst, mask(tail), (F)v); } unaligned_store(dst, v); } |