diff options
author | Mike Klein <mtklein@chromium.org> | 2017-04-06 17:53:18 -0400 |
---|---|---|
committer | Skia Commit-Bot <skia-commit-bot@chromium.org> | 2017-04-06 23:06:04 +0000 |
commit | 7d3d8723319038d16456137ba932f238c1e65dbf (patch) | |
tree | b9615f69226ae8fe3b25a27293c3cb58c08f9fdd /src/jumper/SkJumper_generated_win.S | |
parent | b9a02a131eba272fbcae320c2bd6c0d668162c2e (diff) |
jumper, gather_i8
Change-Id: Iefa8044bac0555c5fff370217a6270b4f3c64300
Reviewed-on: https://skia-review.googlesource.com/11582
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/jumper/SkJumper_generated_win.S')
-rw-r--r-- | src/jumper/SkJumper_generated_win.S | 600 |
1 files changed, 461 insertions, 139 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 5a154e8295..2351b36baf 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -1495,12 +1495,85 @@ _sk_gather_g8_hsw LABEL PROC DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax +PUBLIC _sk_gather_i8_hsw +_sk_gather_i8_hsw 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 1573 <_sk_gather_i8_hsw+0xf> + DB 76,137,192 ; mov %r8,%rax + DB 235,2 ; jmp 1575 <_sk_gather_i8_hsw+0x11> + DB 72,173 ; lods %ds:(%rsi),%rax + 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 76,139,8 ; mov (%rax),%r9 + DB 197,254,91,201 ; vcvttps2dq %ymm1,%ymm1 + DB 196,226,125,88,80,16 ; vpbroadcastd 0x10(%rax),%ymm2 + DB 196,226,109,64,201 ; vpmulld %ymm1,%ymm2,%ymm1 + DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 + DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0 + DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 65,137,194 ; mov %eax,%r10d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx + DB 65,137,223 ; mov %ebx,%r15d + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,193,249,126,196 ; vmovq %xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 196,131,121,32,4,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 + DB 196,131,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 + DB 196,131,121,32,4,17,2 ; vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0 + DB 196,195,121,32,4,1,3 ; vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0 + DB 196,131,121,32,4,41,4 ; vpinsrb $0x4,(%r9,%r13,1),%xmm0,%xmm0 + DB 196,131,121,32,4,33,5 ; vpinsrb $0x5,(%r9,%r12,1),%xmm0,%xmm0 + DB 196,131,121,32,4,57,6 ; vpinsrb $0x6,(%r9,%r15,1),%xmm0,%xmm0 + DB 196,195,121,32,4,25,7 ; vpinsrb $0x7,(%r9,%rbx,1),%xmm0,%xmm0 + DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 + DB 73,139,64,8 ; mov 0x8(%r8),%rax + DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1 + DB 196,226,117,144,28,128 ; vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 + DB 184,255,0,0,0 ; mov $0xff,%eax + DB 197,249,110,192 ; vmovd %eax,%xmm0 + DB 196,226,125,88,208 ; vpbroadcastd %xmm0,%ymm2 + DB 197,237,219,195 ; vpand %ymm3,%ymm2,%ymm0 + DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 + DB 184,129,128,128,59 ; mov $0x3b808081,%eax + DB 197,249,110,200 ; vmovd %eax,%xmm1 + DB 196,98,125,88,193 ; vpbroadcastd %xmm1,%ymm8 + DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0 + DB 197,245,114,211,8 ; vpsrld $0x8,%ymm3,%ymm1 + DB 197,237,219,201 ; vpand %ymm1,%ymm2,%ymm1 + DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 + DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1 + DB 197,181,114,211,16 ; vpsrld $0x10,%ymm3,%ymm9 + DB 196,193,109,219,209 ; vpand %ymm9,%ymm2,%ymm2 + DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 + DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2 + DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3 + 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 91 ; pop %rbx + DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 + DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 + DB 255,224 ; jmpq *%rax + PUBLIC _sk_load_565_hsw _sk_load_565_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,149,0,0,0 ; jne 1607 <_sk_load_565_hsw+0xa3> + DB 15,133,149,0,0,0 ; jne 1727 <_sk_load_565_hsw+0xa3> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2 DB 184,0,248,0,0 ; mov $0xf800,%eax @@ -1540,9 +1613,9 @@ _sk_load_565_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,89,255,255,255 ; ja 1578 <_sk_load_565_hsw+0x14> + DB 15,135,89,255,255,255 ; ja 1698 <_sk_load_565_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1674 <_sk_load_565_hsw+0x110> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1794 <_sk_load_565_hsw+0x110> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -1554,12 +1627,12 @@ _sk_load_565_hsw LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,5,255,255,255 ; jmpq 1578 <_sk_load_565_hsw+0x14> + DB 233,5,255,255,255 ; jmpq 1698 <_sk_load_565_hsw+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 1679 <_sk_load_565_hsw+0x115> + DB 235,255 ; jmp 1799 <_sk_load_565_hsw+0x115> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -1682,7 +1755,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 183f <_sk_store_565_hsw+0x6c> + DB 117,10 ; jne 195f <_sk_store_565_hsw+0x6c> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1690,9 +1763,9 @@ _sk_store_565_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 183b <_sk_store_565_hsw+0x68> + DB 119,236 ; ja 195b <_sk_store_565_hsw+0x68> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 189c <_sk_store_565_hsw+0xc9> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 19bc <_sk_store_565_hsw+0xc9> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -1703,7 +1776,7 @@ _sk_store_565_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 183b <_sk_store_565_hsw+0x68> + DB 235,159 ; jmp 195b <_sk_store_565_hsw+0x68> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -1732,7 +1805,7 @@ _sk_load_4444_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,179,0,0,0 ; jne 1979 <_sk_load_4444_hsw+0xc1> + DB 15,133,179,0,0,0 ; jne 1a99 <_sk_load_4444_hsw+0xc1> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 196,98,125,51,200 ; vpmovzxwd %xmm0,%ymm9 DB 184,0,240,0,0 ; mov $0xf000,%eax @@ -1778,9 +1851,9 @@ _sk_load_4444_hsw LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,59,255,255,255 ; ja 18cc <_sk_load_4444_hsw+0x14> + DB 15,135,59,255,255,255 ; ja 19ec <_sk_load_4444_hsw+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 19e8 <_sk_load_4444_hsw+0x130> + DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1b08 <_sk_load_4444_hsw+0x130> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -1792,13 +1865,13 @@ _sk_load_4444_hsw LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,231,254,255,255 ; jmpq 18cc <_sk_load_4444_hsw+0x14> + DB 233,231,254,255,255 ; jmpq 19ec <_sk_load_4444_hsw+0x14> DB 15,31,0 ; nopl (%rax) DB 241 ; icebp DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe20019f0 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff2bc> + DB 233,255,255,255,225 ; jmpq ffffffffe2001b10 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff2bc> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -1926,7 +1999,7 @@ _sk_store_4444_hsw LABEL PROC DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9 DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 1bd7 <_sk_store_4444_hsw+0x72> + DB 117,10 ; jne 1cf7 <_sk_store_4444_hsw+0x72> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -1934,9 +2007,9 @@ _sk_store_4444_hsw LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 1bd3 <_sk_store_4444_hsw+0x6e> + DB 119,236 ; ja 1cf3 <_sk_store_4444_hsw+0x6e> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1c34 <_sk_store_4444_hsw+0xcf> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1d54 <_sk_store_4444_hsw+0xcf> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -1947,7 +2020,7 @@ _sk_store_4444_hsw LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 1bd3 <_sk_store_4444_hsw+0x6e> + DB 235,159 ; jmp 1cf3 <_sk_store_4444_hsw+0x6e> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -1978,7 +2051,7 @@ _sk_load_8888_hsw LABEL PROC DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 DB 76,3,8 ; add (%rax),%r9 DB 77,133,192 ; test %r8,%r8 - DB 117,104 ; jne 1ccd <_sk_load_8888_hsw+0x7d> + DB 117,104 ; jne 1ded <_sk_load_8888_hsw+0x7d> DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -2011,7 +2084,7 @@ _sk_load_8888_hsw LABEL PROC DB 196,225,249,110,192 ; vmovq %rax,%xmm0 DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0 DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3 - DB 233,116,255,255,255 ; jmpq 1c6a <_sk_load_8888_hsw+0x1a> + DB 233,116,255,255,255 ; jmpq 1d8a <_sk_load_8888_hsw+0x1a> PUBLIC _sk_gather_8888_hsw _sk_gather_8888_hsw LABEL PROC @@ -2071,7 +2144,7 @@ _sk_store_8888_hsw LABEL PROC DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8 DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8 DB 77,133,192 ; test %r8,%r8 - DB 117,12 ; jne 1df0 <_sk_store_8888_hsw+0x74> + DB 117,12 ; jne 1f10 <_sk_store_8888_hsw+0x74> DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9) DB 72,173 ; lods %ds:(%rsi),%rax DB 76,137,193 ; mov %r8,%rcx @@ -2084,14 +2157,14 @@ _sk_store_8888_hsw LABEL PROC DB 196,97,249,110,200 ; vmovq %rax,%xmm9 DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9 DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9) - DB 235,211 ; jmp 1de9 <_sk_store_8888_hsw+0x6d> + DB 235,211 ; jmp 1f09 <_sk_store_8888_hsw+0x6d> PUBLIC _sk_load_f16_hsw _sk_load_f16_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 117,97 ; jne 1e81 <_sk_load_f16_hsw+0x6b> + DB 117,97 ; jne 1fa1 <_sk_load_f16_hsw+0x6b> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -2117,29 +2190,29 @@ _sk_load_f16_hsw LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 1ee0 <_sk_load_f16_hsw+0xca> + DB 116,79 ; je 2000 <_sk_load_f16_hsw+0xca> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 1ee0 <_sk_load_f16_hsw+0xca> + DB 114,67 ; jb 2000 <_sk_load_f16_hsw+0xca> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 1eed <_sk_load_f16_hsw+0xd7> + DB 116,68 ; je 200d <_sk_load_f16_hsw+0xd7> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 1eed <_sk_load_f16_hsw+0xd7> + DB 114,56 ; jb 200d <_sk_load_f16_hsw+0xd7> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,114,255,255,255 ; je 1e37 <_sk_load_f16_hsw+0x21> + DB 15,132,114,255,255,255 ; je 1f57 <_sk_load_f16_hsw+0x21> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,98,255,255,255 ; jb 1e37 <_sk_load_f16_hsw+0x21> + DB 15,130,98,255,255,255 ; jb 1f57 <_sk_load_f16_hsw+0x21> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,87,255,255,255 ; jmpq 1e37 <_sk_load_f16_hsw+0x21> + DB 233,87,255,255,255 ; jmpq 1f57 <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,74,255,255,255 ; jmpq 1e37 <_sk_load_f16_hsw+0x21> + DB 233,74,255,255,255 ; jmpq 1f57 <_sk_load_f16_hsw+0x21> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,65,255,255,255 ; jmpq 1e37 <_sk_load_f16_hsw+0x21> + DB 233,65,255,255,255 ; jmpq 1f57 <_sk_load_f16_hsw+0x21> PUBLIC _sk_store_f16_hsw _sk_store_f16_hsw LABEL PROC @@ -2158,7 +2231,7 @@ _sk_store_f16_hsw LABEL PROC DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9 DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,27 ; jne 1f5b <_sk_store_f16_hsw+0x65> + DB 117,27 ; jne 207b <_sk_store_f16_hsw+0x65> DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8) DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8) DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8) @@ -2167,29 +2240,29 @@ _sk_store_f16_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,241 ; je 1f57 <_sk_store_f16_hsw+0x61> + DB 116,241 ; je 2077 <_sk_store_f16_hsw+0x61> DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,229 ; jb 1f57 <_sk_store_f16_hsw+0x61> + DB 114,229 ; jb 2077 <_sk_store_f16_hsw+0x61> DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8) - DB 116,221 ; je 1f57 <_sk_store_f16_hsw+0x61> + DB 116,221 ; je 2077 <_sk_store_f16_hsw+0x61> DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,209 ; jb 1f57 <_sk_store_f16_hsw+0x61> + DB 114,209 ; jb 2077 <_sk_store_f16_hsw+0x61> DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8) - DB 116,201 ; je 1f57 <_sk_store_f16_hsw+0x61> + DB 116,201 ; je 2077 <_sk_store_f16_hsw+0x61> DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,189 ; jb 1f57 <_sk_store_f16_hsw+0x61> + DB 114,189 ; jb 2077 <_sk_store_f16_hsw+0x61> DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8) - DB 235,181 ; jmp 1f57 <_sk_store_f16_hsw+0x61> + DB 235,181 ; jmp 2077 <_sk_store_f16_hsw+0x61> PUBLIC _sk_load_u16_be_hsw _sk_load_u16_be_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,201,0,0,0 ; jne 2079 <_sk_load_u16_be_hsw+0xd7> + DB 15,133,201,0,0,0 ; jne 2199 <_sk_load_u16_be_hsw+0xd7> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -2238,29 +2311,29 @@ _sk_load_u16_be_hsw LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 20d8 <_sk_load_u16_be_hsw+0x136> + DB 116,79 ; je 21f8 <_sk_load_u16_be_hsw+0x136> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 20d8 <_sk_load_u16_be_hsw+0x136> + DB 114,67 ; jb 21f8 <_sk_load_u16_be_hsw+0x136> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 20e5 <_sk_load_u16_be_hsw+0x143> + DB 116,68 ; je 2205 <_sk_load_u16_be_hsw+0x143> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 20e5 <_sk_load_u16_be_hsw+0x143> + DB 114,56 ; jb 2205 <_sk_load_u16_be_hsw+0x143> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,10,255,255,255 ; je 1fc7 <_sk_load_u16_be_hsw+0x25> + DB 15,132,10,255,255,255 ; je 20e7 <_sk_load_u16_be_hsw+0x25> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,250,254,255,255 ; jb 1fc7 <_sk_load_u16_be_hsw+0x25> + DB 15,130,250,254,255,255 ; jb 20e7 <_sk_load_u16_be_hsw+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,239,254,255,255 ; jmpq 1fc7 <_sk_load_u16_be_hsw+0x25> + DB 233,239,254,255,255 ; jmpq 20e7 <_sk_load_u16_be_hsw+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,226,254,255,255 ; jmpq 1fc7 <_sk_load_u16_be_hsw+0x25> + DB 233,226,254,255,255 ; jmpq 20e7 <_sk_load_u16_be_hsw+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,217,254,255,255 ; jmpq 1fc7 <_sk_load_u16_be_hsw+0x25> + DB 233,217,254,255,255 ; jmpq 20e7 <_sk_load_u16_be_hsw+0x25> PUBLIC _sk_store_u16_be_hsw _sk_store_u16_be_hsw LABEL PROC @@ -2306,7 +2379,7 @@ _sk_store_u16_be_hsw LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 21e1 <_sk_store_u16_be_hsw+0xf3> + DB 117,31 ; jne 2301 <_sk_store_u16_be_hsw+0xf3> DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8) DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8) DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8) @@ -2315,31 +2388,31 @@ _sk_store_u16_be_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 21dd <_sk_store_u16_be_hsw+0xef> + DB 116,240 ; je 22fd <_sk_store_u16_be_hsw+0xef> DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 21dd <_sk_store_u16_be_hsw+0xef> + DB 114,227 ; jb 22fd <_sk_store_u16_be_hsw+0xef> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 21dd <_sk_store_u16_be_hsw+0xef> + DB 116,218 ; je 22fd <_sk_store_u16_be_hsw+0xef> DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 21dd <_sk_store_u16_be_hsw+0xef> + DB 114,205 ; jb 22fd <_sk_store_u16_be_hsw+0xef> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 21dd <_sk_store_u16_be_hsw+0xef> + DB 116,196 ; je 22fd <_sk_store_u16_be_hsw+0xef> DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 21dd <_sk_store_u16_be_hsw+0xef> + DB 114,183 ; jb 22fd <_sk_store_u16_be_hsw+0xef> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 21dd <_sk_store_u16_be_hsw+0xef> + DB 235,174 ; jmp 22fd <_sk_store_u16_be_hsw+0xef> PUBLIC _sk_load_f32_hsw _sk_load_f32_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,131,249,7 ; cmp $0x7,%rcx - DB 119,110 ; ja 22a5 <_sk_load_f32_hsw+0x76> + DB 119,110 ; ja 23c5 <_sk_load_f32_hsw+0x76> DB 76,139,0 ; mov (%rax),%r8 DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 - DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 22d0 <_sk_load_f32_hsw+0xa1> + DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 23f0 <_sk_load_f32_hsw+0xa1> DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -2398,7 +2471,7 @@ _sk_store_f32_hsw LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 235d <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 247d <_sk_store_f32_hsw+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -2411,22 +2484,22 @@ _sk_store_f32_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 2359 <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 2479 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 2359 <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 2479 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 2359 <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 2479 <_sk_store_f32_hsw+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 2359 <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 2479 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 2359 <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 2479 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 2359 <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 2479 <_sk_store_f32_hsw+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 2359 <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 2479 <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -4458,12 +4531,117 @@ _sk_gather_g8_avx LABEL PROC DB 65,95 ; pop %r15 DB 255,224 ; jmpq *%rax +PUBLIC _sk_gather_i8_avx +_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 1b06 <_sk_gather_i8_avx+0xf> + DB 76,137,192 ; mov %r8,%rax + DB 235,2 ; jmp 1b08 <_sk_gather_i8_avx+0x11> + DB 72,173 ; lods %ds:(%rsi),%rax + 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 76,139,8 ; mov (%rax),%r9 + DB 197,254,91,209 ; vcvttps2dq %ymm1,%ymm2 + DB 197,249,110,72,16 ; vmovd 0x10(%rax),%xmm1 + DB 197,249,112,217,0 ; vpshufd $0x0,%xmm1,%xmm3 + DB 196,226,97,64,202 ; vpmulld %xmm2,%xmm3,%xmm1 + DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2 + DB 196,226,97,64,210 ; vpmulld %xmm2,%xmm3,%xmm2 + DB 197,254,91,192 ; vcvttps2dq %ymm0,%ymm0 + DB 196,227,125,25,195,1 ; vextractf128 $0x1,%ymm0,%xmm3 + DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2 + DB 196,227,249,22,208,1 ; vpextrq $0x1,%xmm2,%rax + DB 65,137,194 ; mov %eax,%r10d + DB 72,193,232,32 ; shr $0x20,%rax + DB 196,193,249,126,211 ; vmovq %xmm2,%r11 + DB 69,137,222 ; mov %r11d,%r14d + DB 73,193,235,32 ; shr $0x20,%r11 + DB 197,241,254,192 ; vpaddd %xmm0,%xmm1,%xmm0 + DB 196,225,249,126,195 ; vmovq %xmm0,%rbx + DB 65,137,223 ; mov %ebx,%r15d + DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12 + DB 69,137,229 ; mov %r12d,%r13d + DB 73,193,236,32 ; shr $0x20,%r12 + DB 72,193,235,32 ; shr $0x20,%rbx + DB 196,131,121,32,4,49,0 ; vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 + DB 196,131,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 + DB 196,131,121,32,4,17,2 ; vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0 + DB 196,195,121,32,4,1,3 ; vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0 + DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 + DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10 + DB 196,193,249,126,195 ; vmovq %xmm0,%r11 + DB 196,131,121,32,4,57,0 ; vpinsrb $0x0,(%r9,%r15,1),%xmm0,%xmm0 + DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 + DB 196,131,121,32,4,41,2 ; vpinsrb $0x2,(%r9,%r13,1),%xmm0,%xmm0 + DB 196,131,121,32,4,33,3 ; vpinsrb $0x3,(%r9,%r12,1),%xmm0,%xmm0 + DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0 + DB 73,139,88,8 ; mov 0x8(%r8),%rbx + DB 196,193,249,126,193 ; vmovq %xmm0,%r9 + DB 69,137,200 ; mov %r9d,%r8d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 196,227,249,22,192,1 ; vpextrq $0x1,%xmm0,%rax + DB 65,137,198 ; mov %eax,%r14d + DB 72,193,232,30 ; shr $0x1e,%rax + DB 69,137,223 ; mov %r11d,%r15d + DB 73,193,235,30 ; shr $0x1e,%r11 + DB 69,137,212 ; mov %r10d,%r12d + DB 73,193,234,30 ; shr $0x1e,%r10 + DB 196,161,121,110,4,131 ; vmovd (%rbx,%r8,4),%xmm0 + DB 196,163,121,34,4,11,1 ; vpinsrd $0x1,(%rbx,%r9,1),%xmm0,%xmm0 + DB 196,163,121,34,4,179,2 ; vpinsrd $0x2,(%rbx,%r14,4),%xmm0,%xmm0 + DB 196,99,121,34,4,3,3 ; vpinsrd $0x3,(%rbx,%rax,1),%xmm0,%xmm8 + DB 196,161,121,110,4,187 ; vmovd (%rbx,%r15,4),%xmm0 + DB 196,163,121,34,4,27,1 ; vpinsrd $0x1,(%rbx,%r11,1),%xmm0,%xmm0 + 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 184,255,0,0,0 ; mov $0xff,%eax + DB 197,249,110,200 ; vmovd %eax,%xmm1 + DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1 + DB 196,99,117,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm11 + DB 197,164,84,192 ; vandps %ymm0,%ymm11,%ymm0 + DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0 + DB 184,129,128,128,59 ; mov $0x3b808081,%eax + DB 197,249,110,200 ; vmovd %eax,%xmm1 + DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1 + DB 196,99,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm9 + DB 196,193,124,89,193 ; vmulps %ymm9,%ymm0,%ymm0 + DB 196,193,41,114,208,8 ; vpsrld $0x8,%xmm8,%xmm10 + DB 197,241,114,211,8 ; vpsrld $0x8,%xmm3,%xmm1 + DB 196,227,45,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 + DB 197,164,84,201 ; vandps %ymm1,%ymm11,%ymm1 + DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1 + DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1 + DB 196,193,41,114,208,16 ; vpsrld $0x10,%xmm8,%xmm10 + DB 197,233,114,211,16 ; vpsrld $0x10,%xmm3,%xmm2 + DB 196,227,45,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 + DB 197,164,84,210 ; vandps %ymm2,%ymm11,%ymm2 + DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2 + DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2 + DB 196,193,57,114,208,24 ; vpsrld $0x18,%xmm8,%xmm8 + DB 197,225,114,211,24 ; vpsrld $0x18,%xmm3,%xmm3 + DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 + DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3 + DB 196,193,100,89,217 ; vmulps %ymm9,%ymm3,%ymm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 91 ; pop %rbx + DB 65,92 ; pop %r12 + DB 65,93 ; pop %r13 + DB 65,94 ; pop %r14 + DB 65,95 ; pop %r15 + DB 255,224 ; jmpq *%rax + PUBLIC _sk_load_565_avx _sk_load_565_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,209,0,0,0 ; jne 1bd6 <_sk_load_565_avx+0xdf> + DB 15,133,209,0,0,0 ; jne 1da2 <_sk_load_565_avx+0xdf> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -4513,9 +4691,9 @@ _sk_load_565_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,29,255,255,255 ; ja 1b0b <_sk_load_565_avx+0x14> + DB 15,135,29,255,255,255 ; ja 1cd7 <_sk_load_565_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 1c44 <_sk_load_565_avx+0x14d> + DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 1e10 <_sk_load_565_avx+0x14d> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -4527,7 +4705,7 @@ _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 233,201,254,255,255 ; jmpq 1b0b <_sk_load_565_avx+0x14> + DB 233,201,254,255,255 ; jmpq 1cd7 <_sk_load_565_avx+0x14> DB 102,144 ; xchg %ax,%ax DB 242,255 ; repnz (bad) DB 255 ; (bad) @@ -4680,7 +4858,7 @@ _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 1e8f <_sk_store_565_avx+0x9e> + DB 117,10 ; jne 205b <_sk_store_565_avx+0x9e> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -4688,9 +4866,9 @@ _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 1e8b <_sk_store_565_avx+0x9a> + DB 119,236 ; ja 2057 <_sk_store_565_avx+0x9a> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1eec <_sk_store_565_avx+0xfb> + DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 20b8 <_sk_store_565_avx+0xfb> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -4701,7 +4879,7 @@ _sk_store_565_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 1e8b <_sk_store_565_avx+0x9a> + DB 235,159 ; jmp 2057 <_sk_store_565_avx+0x9a> DB 247,255 ; idiv %edi DB 255 ; (bad) DB 255 ; (bad) @@ -4730,7 +4908,7 @@ _sk_load_4444_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,245,0,0,0 ; jne 200b <_sk_load_4444_avx+0x103> + DB 15,133,245,0,0,0 ; jne 21d7 <_sk_load_4444_avx+0x103> DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0 DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1 DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1 @@ -4787,9 +4965,9 @@ _sk_load_4444_avx LABEL PROC DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,249,254,255,255 ; ja 1f1c <_sk_load_4444_avx+0x14> + DB 15,135,249,254,255,255 ; ja 20e8 <_sk_load_4444_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 2078 <_sk_load_4444_avx+0x170> + DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 2244 <_sk_load_4444_avx+0x170> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -4801,12 +4979,12 @@ _sk_load_4444_avx LABEL PROC DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 - DB 233,165,254,255,255 ; jmpq 1f1c <_sk_load_4444_avx+0x14> + DB 233,165,254,255,255 ; jmpq 20e8 <_sk_load_4444_avx+0x14> DB 144 ; nop DB 243,255 ; repz (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 235,255 ; jmp 207d <_sk_load_4444_avx+0x175> + DB 235,255 ; jmp 2249 <_sk_load_4444_avx+0x175> DB 255 ; (bad) DB 255,227 ; jmpq *%rbx DB 255 ; (bad) @@ -4963,7 +5141,7 @@ _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 22f8 <_sk_store_4444_avx+0xaf> + DB 117,10 ; jne 24c4 <_sk_store_4444_avx+0xaf> DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -4971,9 +5149,9 @@ _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 22f4 <_sk_store_4444_avx+0xab> + DB 119,236 ; ja 24c0 <_sk_store_4444_avx+0xab> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 2358 <_sk_store_4444_avx+0x10f> + DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 2524 <_sk_store_4444_avx+0x10f> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -4984,7 +5162,7 @@ _sk_store_4444_avx LABEL PROC DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2) - DB 235,159 ; jmp 22f4 <_sk_store_4444_avx+0xab> + DB 235,159 ; jmp 24c0 <_sk_store_4444_avx+0xab> DB 15,31,0 ; nopl (%rax) DB 244 ; hlt DB 255 ; (bad) @@ -5015,7 +5193,7 @@ _sk_load_8888_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,16 ; mov (%rax),%r10 DB 72,133,201 ; test %rcx,%rcx - DB 15,133,157,0,0,0 ; jne 241f <_sk_load_8888_avx+0xab> + DB 15,133,157,0,0,0 ; jne 25eb <_sk_load_8888_avx+0xab> DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9 DB 184,255,0,0,0 ; mov $0xff,%eax DB 197,249,110,192 ; vmovd %eax,%xmm0 @@ -5053,9 +5231,9 @@ _sk_load_8888_avx LABEL PROC DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9 DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 15,135,80,255,255,255 ; ja 2388 <_sk_load_8888_avx+0x14> + DB 15,135,80,255,255,255 ; ja 2554 <_sk_load_8888_avx+0x14> DB 69,15,182,192 ; movzbl %r8b,%r8d - DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 24cc <_sk_load_8888_avx+0x158> + DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 2698 <_sk_load_8888_avx+0x158> DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -5078,7 +5256,7 @@ _sk_load_8888_avx LABEL PROC DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 DB 196,195,49,34,4,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9 - DB 233,188,254,255,255 ; jmpq 2388 <_sk_load_8888_avx+0x14> + DB 233,188,254,255,255 ; jmpq 2554 <_sk_load_8888_avx+0x14> DB 238 ; out %al,(%dx) DB 255 ; (bad) DB 255 ; (bad) @@ -5204,7 +5382,7 @@ _sk_store_8888_avx LABEL PROC DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8 DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,10 ; jne 26cd <_sk_store_8888_avx+0xa4> + DB 117,10 ; jne 2899 <_sk_store_8888_avx+0xa4> DB 196,65,124,17,4,185 ; vmovups %ymm8,(%r9,%rdi,4) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -5212,9 +5390,9 @@ _sk_store_8888_avx LABEL PROC DB 65,128,224,7 ; and $0x7,%r8b DB 65,254,200 ; dec %r8b DB 65,128,248,6 ; cmp $0x6,%r8b - DB 119,236 ; ja 26c9 <_sk_store_8888_avx+0xa0> + DB 119,236 ; ja 2895 <_sk_store_8888_avx+0xa0> DB 65,15,182,192 ; movzbl %r8b,%eax - DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 273c <_sk_store_8888_avx+0x113> + DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 2908 <_sk_store_8888_avx+0x113> DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax DB 76,1,192 ; add %r8,%rax DB 255,224 ; jmpq *%rax @@ -5228,7 +5406,7 @@ _sk_store_8888_avx LABEL PROC DB 196,67,121,22,68,185,8,2 ; vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) DB 196,67,121,22,68,185,4,1 ; vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) DB 196,65,121,126,4,185 ; vmovd %xmm8,(%r9,%rdi,4) - DB 235,143 ; jmp 26c9 <_sk_store_8888_avx+0xa0> + DB 235,143 ; jmp 2895 <_sk_store_8888_avx+0xa0> DB 102,144 ; xchg %ax,%ax DB 246,255 ; idiv %bh DB 255 ; (bad) @@ -5258,7 +5436,7 @@ _sk_load_f16_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,17,1,0,0 ; jne 2877 <_sk_load_f16_avx+0x11f> + DB 15,133,17,1,0,0 ; jne 2a43 <_sk_load_f16_avx+0x11f> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -5320,29 +5498,29 @@ _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 28d6 <_sk_load_f16_avx+0x17e> + DB 116,79 ; je 2aa2 <_sk_load_f16_avx+0x17e> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 28d6 <_sk_load_f16_avx+0x17e> + DB 114,67 ; jb 2aa2 <_sk_load_f16_avx+0x17e> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 28e3 <_sk_load_f16_avx+0x18b> + DB 116,68 ; je 2aaf <_sk_load_f16_avx+0x18b> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 28e3 <_sk_load_f16_avx+0x18b> + DB 114,56 ; jb 2aaf <_sk_load_f16_avx+0x18b> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,194,254,255,255 ; je 277d <_sk_load_f16_avx+0x25> + DB 15,132,194,254,255,255 ; je 2949 <_sk_load_f16_avx+0x25> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,178,254,255,255 ; jb 277d <_sk_load_f16_avx+0x25> + DB 15,130,178,254,255,255 ; jb 2949 <_sk_load_f16_avx+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,167,254,255,255 ; jmpq 277d <_sk_load_f16_avx+0x25> + DB 233,167,254,255,255 ; jmpq 2949 <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,154,254,255,255 ; jmpq 277d <_sk_load_f16_avx+0x25> + DB 233,154,254,255,255 ; jmpq 2949 <_sk_load_f16_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,145,254,255,255 ; jmpq 277d <_sk_load_f16_avx+0x25> + DB 233,145,254,255,255 ; jmpq 2949 <_sk_load_f16_avx+0x25> PUBLIC _sk_store_f16_avx _sk_store_f16_avx LABEL PROC @@ -5381,7 +5559,7 @@ _sk_store_f16_avx LABEL PROC DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9 DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8 DB 72,133,201 ; test %rcx,%rcx - DB 117,31 ; jne 29be <_sk_store_f16_avx+0xd2> + DB 117,31 ; jne 2b8a <_sk_store_f16_avx+0xd2> DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8) DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8) DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8) @@ -5390,29 +5568,29 @@ _sk_store_f16_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 29ba <_sk_store_f16_avx+0xce> + DB 116,240 ; je 2b86 <_sk_store_f16_avx+0xce> DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 29ba <_sk_store_f16_avx+0xce> + DB 114,227 ; jb 2b86 <_sk_store_f16_avx+0xce> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 29ba <_sk_store_f16_avx+0xce> + DB 116,218 ; je 2b86 <_sk_store_f16_avx+0xce> DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 29ba <_sk_store_f16_avx+0xce> + DB 114,205 ; jb 2b86 <_sk_store_f16_avx+0xce> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 29ba <_sk_store_f16_avx+0xce> + DB 116,196 ; je 2b86 <_sk_store_f16_avx+0xce> DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 29ba <_sk_store_f16_avx+0xce> + DB 114,183 ; jb 2b86 <_sk_store_f16_avx+0xce> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 29ba <_sk_store_f16_avx+0xce> + DB 235,174 ; jmp 2b86 <_sk_store_f16_avx+0xce> PUBLIC _sk_load_u16_be_avx _sk_load_u16_be_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 72,139,0 ; mov (%rax),%rax DB 72,133,201 ; test %rcx,%rcx - DB 15,133,1,1,0,0 ; jne 2b1b <_sk_load_u16_be_avx+0x10f> + DB 15,133,1,1,0,0 ; jne 2ce7 <_sk_load_u16_be_avx+0x10f> DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8 DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2 DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3 @@ -5471,29 +5649,29 @@ _sk_load_u16_be_avx LABEL PROC DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8 DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9 DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,79 ; je 2b7a <_sk_load_u16_be_avx+0x16e> + DB 116,79 ; je 2d46 <_sk_load_u16_be_avx+0x16e> DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,67 ; jb 2b7a <_sk_load_u16_be_avx+0x16e> + DB 114,67 ; jb 2d46 <_sk_load_u16_be_avx+0x16e> DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2 DB 72,131,249,3 ; cmp $0x3,%rcx - DB 116,68 ; je 2b87 <_sk_load_u16_be_avx+0x17b> + DB 116,68 ; je 2d53 <_sk_load_u16_be_avx+0x17b> DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,56 ; jb 2b87 <_sk_load_u16_be_avx+0x17b> + DB 114,56 ; jb 2d53 <_sk_load_u16_be_avx+0x17b> DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3 DB 72,131,249,5 ; cmp $0x5,%rcx - DB 15,132,210,254,255,255 ; je 2a31 <_sk_load_u16_be_avx+0x25> + DB 15,132,210,254,255,255 ; je 2bfd <_sk_load_u16_be_avx+0x25> DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 DB 72,131,249,7 ; cmp $0x7,%rcx - DB 15,130,194,254,255,255 ; jb 2a31 <_sk_load_u16_be_avx+0x25> + DB 15,130,194,254,255,255 ; jb 2bfd <_sk_load_u16_be_avx+0x25> DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9 - DB 233,183,254,255,255 ; jmpq 2a31 <_sk_load_u16_be_avx+0x25> + DB 233,183,254,255,255 ; jmpq 2bfd <_sk_load_u16_be_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2 - DB 233,170,254,255,255 ; jmpq 2a31 <_sk_load_u16_be_avx+0x25> + DB 233,170,254,255,255 ; jmpq 2bfd <_sk_load_u16_be_avx+0x25> DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3 - DB 233,161,254,255,255 ; jmpq 2a31 <_sk_load_u16_be_avx+0x25> + DB 233,161,254,255,255 ; jmpq 2bfd <_sk_load_u16_be_avx+0x25> PUBLIC _sk_store_u16_be_avx _sk_store_u16_be_avx LABEL PROC @@ -5540,7 +5718,7 @@ _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 2c8a <_sk_store_u16_be_avx+0xfa> + DB 117,31 ; jne 2e56 <_sk_store_u16_be_avx+0xfa> DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8) DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8) DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8) @@ -5549,31 +5727,31 @@ _sk_store_u16_be_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 2c86 <_sk_store_u16_be_avx+0xf6> + DB 116,240 ; je 2e52 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 2c86 <_sk_store_u16_be_avx+0xf6> + DB 114,227 ; jb 2e52 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8) - DB 116,218 ; je 2c86 <_sk_store_u16_be_avx+0xf6> + DB 116,218 ; je 2e52 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 2c86 <_sk_store_u16_be_avx+0xf6> + DB 114,205 ; jb 2e52 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8) - DB 116,196 ; je 2c86 <_sk_store_u16_be_avx+0xf6> + DB 116,196 ; je 2e52 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,183 ; jb 2c86 <_sk_store_u16_be_avx+0xf6> + DB 114,183 ; jb 2e52 <_sk_store_u16_be_avx+0xf6> DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) - DB 235,174 ; jmp 2c86 <_sk_store_u16_be_avx+0xf6> + DB 235,174 ; jmp 2e52 <_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 2d4e <_sk_load_f32_avx+0x76> + DB 119,110 ; ja 2f1a <_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 # 2d78 <_sk_load_f32_avx+0xa0> + DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 2f44 <_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 @@ -5630,7 +5808,7 @@ _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 2e05 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 2fd1 <_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 @@ -5643,22 +5821,22 @@ _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 2e01 <_sk_store_f32_avx+0x69> + DB 116,240 ; je 2fcd <_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 2e01 <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 2fcd <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 2e01 <_sk_store_f32_avx+0x69> + DB 116,218 ; je 2fcd <_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 2e01 <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 2fcd <_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 2e01 <_sk_store_f32_avx+0x69> + DB 116,195 ; je 2fcd <_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 2e01 <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 2fcd <_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 2e01 <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 2fcd <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -7601,6 +7779,69 @@ _sk_gather_g8_sse41 LABEL PROC DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax +PUBLIC _sk_gather_i8_sse41 +_sk_gather_i8_sse41 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 1616 <_sk_gather_i8_sse41+0xf> + DB 76,137,192 ; mov %r8,%rax + DB 235,2 ; jmp 1618 <_sk_gather_i8_sse41+0x11> + DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,139,16 ; mov (%rax),%r10 + DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 + DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 + DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 + DB 102,15,56,64,209 ; pmulld %xmm1,%xmm2 + DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 + DB 102,15,254,194 ; paddd %xmm2,%xmm0 + DB 102,72,15,58,22,192,1 ; pextrq $0x1,%xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 102,72,15,126,193 ; movq %xmm0,%rcx + DB 65,137,203 ; mov %ecx,%r11d + DB 72,193,233,32 ; shr $0x20,%rcx + DB 102,67,15,58,32,4,26,0 ; pinsrb $0x0,(%r10,%r11,1),%xmm0 + DB 102,65,15,58,32,4,10,1 ; pinsrb $0x1,(%r10,%rcx,1),%xmm0 + DB 102,67,15,58,32,4,10,2 ; pinsrb $0x2,(%r10,%r9,1),%xmm0 + DB 102,65,15,58,32,4,2,3 ; pinsrb $0x3,(%r10,%rax,1),%xmm0 + DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0 + DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9 + DB 102,72,15,126,193 ; movq %xmm0,%rcx + DB 73,139,64,8 ; mov 0x8(%r8),%rax + DB 65,137,200 ; mov %ecx,%r8d + DB 72,193,233,30 ; shr $0x1e,%rcx + DB 69,137,202 ; mov %r9d,%r10d + DB 73,193,233,30 ; shr $0x1e,%r9 + DB 102,66,15,110,28,128 ; movd (%rax,%r8,4),%xmm3 + DB 102,15,58,34,28,8,1 ; pinsrd $0x1,(%rax,%rcx,1),%xmm3 + DB 102,66,15,58,34,28,144,2 ; pinsrd $0x2,(%rax,%r10,4),%xmm3 + DB 102,66,15,58,34,28,8,3 ; pinsrd $0x3,(%rax,%r9,1),%xmm3 + DB 184,255,0,0,0 ; mov $0xff,%eax + DB 102,15,110,192 ; movd %eax,%xmm0 + DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,114,209,8 ; psrld $0x8,%xmm1 + DB 102,15,219,200 ; pand %xmm0,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,114,210,16 ; psrld $0x10,%xmm2 + DB 102,15,219,208 ; pand %xmm0,%xmm2 + DB 102,15,219,195 ; pand %xmm3,%xmm0 + DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 + DB 184,129,128,128,59 ; mov $0x3b808081,%eax + DB 102,68,15,110,192 ; movd %eax,%xmm8 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 + DB 65,15,89,192 ; mulps %xmm8,%xmm0 + DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 + DB 65,15,89,200 ; mulps %xmm8,%xmm1 + DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 + DB 65,15,89,208 ; mulps %xmm8,%xmm2 + DB 102,15,114,211,24 ; psrld $0x18,%xmm3 + DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 + DB 65,15,89,216 ; mulps %xmm8,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_load_565_sse41 _sk_load_565_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -10246,6 +10487,87 @@ _sk_gather_g8_sse2 LABEL PROC DB 15,40,208 ; movaps %xmm0,%xmm2 DB 255,224 ; jmpq *%rax +PUBLIC _sk_gather_i8_sse2 +_sk_gather_i8_sse2 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 1729 <_sk_gather_i8_sse2+0xf> + DB 76,137,192 ; mov %r8,%rax + DB 235,2 ; jmp 172b <_sk_gather_i8_sse2+0x11> + DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,139,16 ; mov (%rax),%r10 + DB 243,15,91,201 ; cvttps2dq %xmm1,%xmm1 + DB 102,15,110,80,16 ; movd 0x10(%rax),%xmm2 + DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2 + DB 102,15,112,217,245 ; pshufd $0xf5,%xmm1,%xmm3 + DB 102,15,244,218 ; pmuludq %xmm2,%xmm3 + DB 102,15,112,219,232 ; pshufd $0xe8,%xmm3,%xmm3 + DB 102,15,244,209 ; pmuludq %xmm1,%xmm2 + DB 102,15,112,202,232 ; pshufd $0xe8,%xmm2,%xmm1 + DB 102,15,98,203 ; punpckldq %xmm3,%xmm1 + DB 243,15,91,192 ; cvttps2dq %xmm0,%xmm0 + DB 102,15,254,193 ; paddd %xmm1,%xmm0 + DB 102,72,15,126,192 ; movq %xmm0,%rax + DB 65,137,193 ; mov %eax,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0 + DB 102,72,15,126,193 ; movq %xmm0,%rcx + DB 65,137,203 ; mov %ecx,%r11d + DB 72,193,233,32 ; shr $0x20,%rcx + DB 71,15,182,28,26 ; movzbl (%r10,%r11,1),%r11d + DB 65,15,182,12,10 ; movzbl (%r10,%rcx,1),%ecx + DB 193,225,8 ; shl $0x8,%ecx + DB 68,9,217 ; or %r11d,%ecx + DB 71,15,182,12,10 ; movzbl (%r10,%r9,1),%r9d + DB 65,15,182,4,2 ; movzbl (%r10,%rax,1),%eax + DB 193,224,8 ; shl $0x8,%eax + DB 68,9,200 ; or %r9d,%eax + DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 + DB 102,15,196,193,1 ; pinsrw $0x1,%ecx,%xmm0 + DB 102,15,239,201 ; pxor %xmm1,%xmm1 + DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0 + DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0 + DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1 + DB 102,72,15,126,200 ; movq %xmm1,%rax + DB 68,15,182,200 ; movzbl %al,%r9d + DB 72,193,232,32 ; shr $0x20,%rax + DB 102,72,15,126,193 ; movq %xmm0,%rcx + DB 77,139,64,8 ; mov 0x8(%r8),%r8 + DB 68,15,182,209 ; movzbl %cl,%r10d + DB 72,193,233,30 ; shr $0x1e,%rcx + DB 102,65,15,110,4,8 ; movd (%r8,%rcx,1),%xmm0 + DB 102,65,15,110,12,128 ; movd (%r8,%rax,4),%xmm1 + DB 102,15,98,193 ; punpckldq %xmm1,%xmm0 + DB 102,67,15,110,28,144 ; movd (%r8,%r10,4),%xmm3 + DB 102,67,15,110,12,136 ; movd (%r8,%r9,4),%xmm1 + DB 102,15,98,217 ; punpckldq %xmm1,%xmm3 + DB 102,15,98,216 ; punpckldq %xmm0,%xmm3 + DB 184,255,0,0,0 ; mov $0xff,%eax + DB 102,15,110,192 ; movd %eax,%xmm0 + DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0 + DB 102,15,111,203 ; movdqa %xmm3,%xmm1 + DB 102,15,114,209,8 ; psrld $0x8,%xmm1 + DB 102,15,219,200 ; pand %xmm0,%xmm1 + DB 102,15,111,211 ; movdqa %xmm3,%xmm2 + DB 102,15,114,210,16 ; psrld $0x10,%xmm2 + DB 102,15,219,208 ; pand %xmm0,%xmm2 + DB 102,15,219,195 ; pand %xmm3,%xmm0 + DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0 + DB 184,129,128,128,59 ; mov $0x3b808081,%eax + DB 102,68,15,110,192 ; movd %eax,%xmm8 + DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8 + DB 65,15,89,192 ; mulps %xmm8,%xmm0 + DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1 + DB 65,15,89,200 ; mulps %xmm8,%xmm1 + DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2 + DB 65,15,89,208 ; mulps %xmm8,%xmm2 + DB 102,15,114,211,24 ; psrld $0x18,%xmm3 + DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3 + DB 65,15,89,216 ; mulps %xmm8,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_load_565_sse2 _sk_load_565_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax |