aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/jumper/SkJumper_generated_win.S
diff options
context:
space:
mode:
authorGravatar Mike Klein <mtklein@chromium.org>2017-04-07 13:09:29 -0400
committerGravatar Skia Commit-Bot <skia-commit-bot@chromium.org>2017-04-07 17:54:11 +0000
commit40de6dad46874d03cc33a883797f20c665d7aa39 (patch)
treeab4a89b3b8d6aca42fadda2164c2448b9148e7c4 /src/jumper/SkJumper_generated_win.S
parent4e3abc1ad5f078ed55cbc0c0ef0e14062a39bd13 (diff)
jumper, byte_tables + byte_tables_rgb
Factors out a function F from_byte(U8) too. Change-Id: Ib739ccbd509ddf25d2bfb7751ba6eaf51b16c12f Reviewed-on: https://skia-review.googlesource.com/11791 Reviewed-by: Matt Sarett <msarett@google.com> Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/jumper/SkJumper_generated_win.S')
-rw-r--r--src/jumper/SkJumper_generated_win.S1313
1 files changed, 1130 insertions, 183 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index d1ef1a42af..c158c806d0 100644
--- a/src/jumper/SkJumper_generated_win.S
+++ b/src/jumper/SkJumper_generated_win.S
@@ -1252,6 +1252,293 @@ _sk_load_tables_hsw LABEL PROC
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
DB 233,99,255,255,255 ; jmpq 1172 <_sk_load_tables_hsw+0x1a>
+PUBLIC _sk_byte_tables_hsw
+_sk_byte_tables_hsw LABEL PROC
+ 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 72,173 ; lods %ds:(%rsi),%rax
+ DB 65,184,0,0,127,67 ; mov $0x437f0000,%r8d
+ DB 196,65,121,110,192 ; vmovd %r8d,%xmm8
+ DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8
+ DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
+ DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 68,137,197 ; mov %r8d,%ebp
+ DB 77,137,194 ; mov %r8,%r10
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,192 ; vmovq %xmm0,%r8
+ DB 69,137,195 ; mov %r8d,%r11d
+ DB 77,137,199 ; mov %r8,%r15
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 69,137,198 ; mov %r8d,%r14d
+ DB 77,137,196 ; mov %r8,%r12
+ DB 73,193,236,32 ; shr $0x20,%r12
+ DB 196,225,249,126,195 ; vmovq %xmm0,%rbx
+ DB 65,137,221 ; mov %ebx,%r13d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 76,139,8 ; mov (%rax),%r9
+ DB 76,139,64,8 ; mov 0x8(%rax),%r8
+ DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
+ DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
+ DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp
+ DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp
+ DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,41 ; movzbl (%r9,%r13,1),%ebp
+ DB 196,227,121,32,197,4 ; vpinsrb $0x4,%ebp,%xmm0,%xmm0
+ DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp
+ DB 196,227,121,32,197,5 ; vpinsrb $0x5,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp
+ DB 196,227,121,32,197,6 ; vpinsrb $0x6,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp
+ DB 196,227,121,32,197,7 ; vpinsrb $0x7,%ebp,%xmm0,%xmm0
+ DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
+ DB 197,124,91,208 ; vcvtdq2ps %ymm0,%ymm10
+ DB 189,129,128,128,59 ; mov $0x3b808081,%ebp
+ DB 197,249,110,197 ; vmovd %ebp,%xmm0
+ DB 196,98,125,88,200 ; vpbroadcastd %xmm0,%ymm9
+ DB 196,193,44,89,193 ; vmulps %ymm9,%ymm10,%ymm0
+ DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
+ DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1
+ DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp
+ DB 65,137,233 ; mov %ebp,%r9d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,203 ; vmovq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1
+ DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 196,193,249,126,207 ; vmovq %xmm1,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
+ DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
+ DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx
+ DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1
+ DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp
+ DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp
+ DB 196,227,113,32,205,4 ; vpinsrb $0x4,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,56 ; movzbl (%r8,%r15,1),%ebp
+ DB 196,227,113,32,205,5 ; vpinsrb $0x5,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp
+ DB 196,227,113,32,205,6 ; vpinsrb $0x6,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp
+ DB 196,227,113,32,205,7 ; vpinsrb $0x7,%ebp,%xmm1,%xmm1
+ DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1
+ DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
+ DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1
+ DB 76,139,64,16 ; mov 0x10(%rax),%r8
+ DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
+ DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2
+ DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp
+ DB 65,137,233 ; mov %ebp,%r9d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,211 ; vmovq %xmm2,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,57,210,1 ; vextracti128 $0x1,%ymm2,%xmm2
+ DB 196,195,249,22,211,1 ; vpextrq $0x1,%xmm2,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 196,193,249,126,215 ; vmovq %xmm2,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2
+ DB 196,195,105,32,20,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2
+ DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx
+ DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2
+ DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp
+ DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2
+ DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp
+ DB 196,227,105,32,213,4 ; vpinsrb $0x4,%ebp,%xmm2,%xmm2
+ DB 67,15,182,44,56 ; movzbl (%r8,%r15,1),%ebp
+ DB 196,227,105,32,213,5 ; vpinsrb $0x5,%ebp,%xmm2,%xmm2
+ DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp
+ DB 196,227,105,32,213,6 ; vpinsrb $0x6,%ebp,%xmm2,%xmm2
+ DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp
+ DB 196,227,105,32,213,7 ; vpinsrb $0x7,%ebp,%xmm2,%xmm2
+ DB 196,226,125,49,210 ; vpmovzxbd %xmm2,%ymm2
+ DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 72,139,64,24 ; mov 0x18(%rax),%rax
+ DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3
+ DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3
+ DB 196,227,249,22,221,1 ; vpextrq $0x1,%xmm3,%rbp
+ DB 65,137,232 ; mov %ebp,%r8d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,219 ; vmovq %xmm3,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,57,219,1 ; vextracti128 $0x1,%ymm3,%xmm3
+ DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10
+ DB 69,137,211 ; mov %r10d,%r11d
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,222 ; vmovq %xmm3,%r14
+ DB 69,137,247 ; mov %r14d,%r15d
+ DB 73,193,238,32 ; shr $0x20,%r14
+ DB 196,163,121,32,28,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3
+ DB 196,227,97,32,28,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3
+ DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx
+ DB 196,227,97,32,219,2 ; vpinsrb $0x2,%ebx,%xmm3,%xmm3
+ DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp
+ DB 196,227,97,32,221,3 ; vpinsrb $0x3,%ebp,%xmm3,%xmm3
+ DB 66,15,182,44,56 ; movzbl (%rax,%r15,1),%ebp
+ DB 196,227,97,32,221,4 ; vpinsrb $0x4,%ebp,%xmm3,%xmm3
+ DB 66,15,182,44,48 ; movzbl (%rax,%r14,1),%ebp
+ DB 196,227,97,32,221,5 ; vpinsrb $0x5,%ebp,%xmm3,%xmm3
+ DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp
+ DB 196,227,97,32,221,6 ; vpinsrb $0x6,%ebp,%xmm3,%xmm3
+ DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax
+ DB 196,227,97,32,216,7 ; vpinsrb $0x7,%eax,%xmm3,%xmm3
+ DB 196,226,125,49,219 ; vpmovzxbd %xmm3,%ymm3
+ DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
+ DB 197,180,89,219 ; vmulps %ymm3,%ymm9,%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 93 ; pop %rbp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_byte_tables_rgb_hsw
+_sk_byte_tables_rgb_hsw LABEL PROC
+ 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 72,173 ; lods %ds:(%rsi),%rax
+ DB 68,139,64,24 ; mov 0x18(%rax),%r8d
+ DB 65,255,200 ; dec %r8d
+ DB 196,65,121,110,192 ; vmovd %r8d,%xmm8
+ DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8
+ DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
+ DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
+ DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 68,137,197 ; mov %r8d,%ebp
+ DB 77,137,194 ; mov %r8,%r10
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,192 ; vmovq %xmm0,%r8
+ DB 69,137,195 ; mov %r8d,%r11d
+ DB 77,137,199 ; mov %r8,%r15
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 69,137,198 ; mov %r8d,%r14d
+ DB 77,137,196 ; mov %r8,%r12
+ DB 73,193,236,32 ; shr $0x20,%r12
+ DB 196,225,249,126,195 ; vmovq %xmm0,%rbx
+ DB 65,137,221 ; mov %ebx,%r13d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 76,139,8 ; mov (%rax),%r9
+ DB 76,139,64,8 ; mov 0x8(%rax),%r8
+ DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
+ DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
+ DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp
+ DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp
+ DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,41 ; movzbl (%r9,%r13,1),%ebp
+ DB 196,227,121,32,197,4 ; vpinsrb $0x4,%ebp,%xmm0,%xmm0
+ DB 65,15,182,44,25 ; movzbl (%r9,%rbx,1),%ebp
+ DB 196,227,121,32,197,5 ; vpinsrb $0x5,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp
+ DB 196,227,121,32,197,6 ; vpinsrb $0x6,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp
+ DB 196,227,121,32,197,7 ; vpinsrb $0x7,%ebp,%xmm0,%xmm0
+ DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
+ DB 197,124,91,208 ; vcvtdq2ps %ymm0,%ymm10
+ DB 189,129,128,128,59 ; mov $0x3b808081,%ebp
+ DB 197,249,110,197 ; vmovd %ebp,%xmm0
+ DB 196,98,125,88,200 ; vpbroadcastd %xmm0,%ymm9
+ DB 196,193,44,89,193 ; vmulps %ymm9,%ymm10,%ymm0
+ DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
+ DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1
+ DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp
+ DB 65,137,233 ; mov %ebp,%r9d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,203 ; vmovq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1
+ DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 196,193,249,126,207 ; vmovq %xmm1,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
+ DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
+ DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx
+ DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1
+ DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp
+ DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,32 ; movzbl (%r8,%r12,1),%ebp
+ DB 196,227,113,32,205,4 ; vpinsrb $0x4,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,56 ; movzbl (%r8,%r15,1),%ebp
+ DB 196,227,113,32,205,5 ; vpinsrb $0x5,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp
+ DB 196,227,113,32,205,6 ; vpinsrb $0x6,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp
+ DB 196,227,113,32,205,7 ; vpinsrb $0x7,%ebp,%xmm1,%xmm1
+ DB 196,226,125,49,201 ; vpmovzxbd %xmm1,%ymm1
+ DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
+ DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1
+ DB 72,139,64,16 ; mov 0x10(%rax),%rax
+ DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
+ DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2
+ DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp
+ DB 65,137,232 ; mov %ebp,%r8d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,211 ; vmovq %xmm2,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,57,210,1 ; vextracti128 $0x1,%ymm2,%xmm2
+ DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10
+ DB 69,137,211 ; mov %r10d,%r11d
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,214 ; vmovq %xmm2,%r14
+ DB 69,137,247 ; mov %r14d,%r15d
+ DB 73,193,238,32 ; shr $0x20,%r14
+ DB 196,163,121,32,20,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2
+ DB 196,227,105,32,20,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2
+ DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx
+ DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2
+ DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp
+ DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2
+ DB 66,15,182,44,56 ; movzbl (%rax,%r15,1),%ebp
+ DB 196,227,105,32,213,4 ; vpinsrb $0x4,%ebp,%xmm2,%xmm2
+ DB 66,15,182,44,48 ; movzbl (%rax,%r14,1),%ebp
+ DB 196,227,105,32,213,5 ; vpinsrb $0x5,%ebp,%xmm2,%xmm2
+ DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp
+ DB 196,227,105,32,213,6 ; vpinsrb $0x6,%ebp,%xmm2,%xmm2
+ DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax
+ DB 196,227,105,32,208,7 ; vpinsrb $0x7,%eax,%xmm2,%xmm2
+ DB 196,226,125,49,210 ; vpmovzxbd %xmm2,%ymm2
+ DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ 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 93 ; pop %rbp
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_a8_hsw
_sk_load_a8_hsw LABEL PROC
DB 73,137,200 ; mov %rcx,%r8
@@ -1259,7 +1546,7 @@ _sk_load_a8_hsw LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,50 ; jne 1251 <_sk_load_a8_hsw+0x42>
+ DB 117,50 ; jne 173e <_sk_load_a8_hsw+0x42>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
@@ -1282,9 +1569,9 @@ _sk_load_a8_hsw LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 1259 <_sk_load_a8_hsw+0x4a>
+ DB 117,234 ; jne 1746 <_sk_load_a8_hsw+0x4a>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,173 ; jmp 1223 <_sk_load_a8_hsw+0x14>
+ DB 235,173 ; jmp 1710 <_sk_load_a8_hsw+0x14>
PUBLIC _sk_gather_a8_hsw
_sk_gather_a8_hsw LABEL PROC
@@ -1355,7 +1642,7 @@ _sk_store_a8_hsw LABEL PROC
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 138e <_sk_store_a8_hsw+0x3b>
+ DB 117,10 ; jne 187b <_sk_store_a8_hsw+0x3b>
DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1363,10 +1650,10 @@ _sk_store_a8_hsw LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 138a <_sk_store_a8_hsw+0x37>
+ DB 119,236 ; ja 1877 <_sk_store_a8_hsw+0x37>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 65,15,182,192 ; movzbl %r8b,%eax
- DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 13f0 <_sk_store_a8_hsw+0x9d>
+ DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 18e0 <_sk_store_a8_hsw+0xa0>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -1377,26 +1664,28 @@ _sk_store_a8_hsw LABEL PROC
DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1)
DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1)
DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1)
- DB 235,154 ; jmp 138a <_sk_store_a8_hsw+0x37>
- DB 247,255 ; idiv %edi
+ DB 235,154 ; jmp 1877 <_sk_store_a8_hsw+0x37>
+ DB 15,31,0 ; nopl (%rax)
+ DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
- DB 239 ; out %eax,(%dx)
DB 255 ; (bad)
+ DB 236 ; in (%dx),%al
DB 255 ; (bad)
- DB 255,231 ; jmpq *%rdi
DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 223,255 ; (bad)
DB 255 ; (bad)
- DB 255,215 ; callq *%rdi
+ DB 220,255 ; fdivr %st,%st(7)
DB 255 ; (bad)
+ DB 255,212 ; callq *%rsp
DB 255 ; (bad)
- DB 255,207 ; dec %edi
DB 255 ; (bad)
+ DB 255,204 ; dec %esp
DB 255 ; (bad)
- DB 255,199 ; inc %edi
+ DB 255 ; (bad)
+ DB 255,196 ; inc %esp
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -1408,7 +1697,7 @@ _sk_load_g8_hsw LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,60 ; jne 1458 <_sk_load_g8_hsw+0x4c>
+ DB 117,60 ; jne 1948 <_sk_load_g8_hsw+0x4c>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
@@ -1433,9 +1722,9 @@ _sk_load_g8_hsw LABEL PROC
DB 77,9,217 ; or %r11,%r9
DB 72,131,193,8 ; add $0x8,%rcx
DB 73,255,202 ; dec %r10
- DB 117,234 ; jne 1460 <_sk_load_g8_hsw+0x54>
+ DB 117,234 ; jne 1950 <_sk_load_g8_hsw+0x54>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,163 ; jmp 1420 <_sk_load_g8_hsw+0x14>
+ DB 235,163 ; jmp 1910 <_sk_load_g8_hsw+0x14>
PUBLIC _sk_gather_g8_hsw
_sk_gather_g8_hsw LABEL PROC
@@ -1500,9 +1789,9 @@ _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 116,5 ; je 1a63 <_sk_gather_i8_hsw+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 1575 <_sk_gather_i8_hsw+0x11>
+ DB 235,2 ; jmp 1a65 <_sk_gather_i8_hsw+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -1573,7 +1862,7 @@ _sk_load_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,149,0,0,0 ; jne 1727 <_sk_load_565_hsw+0xa3>
+ DB 15,133,149,0,0,0 ; jne 1c17 <_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
@@ -1613,9 +1902,9 @@ _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 1698 <_sk_load_565_hsw+0x14>
+ DB 15,135,89,255,255,255 ; ja 1b88 <_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 # 1794 <_sk_load_565_hsw+0x110>
+ DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1c84 <_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
@@ -1627,12 +1916,12 @@ _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 1698 <_sk_load_565_hsw+0x14>
+ DB 233,5,255,255,255 ; jmpq 1b88 <_sk_load_565_hsw+0x14>
DB 144 ; nop
DB 243,255 ; repz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 1799 <_sk_load_565_hsw+0x115>
+ DB 235,255 ; jmp 1c89 <_sk_load_565_hsw+0x115>
DB 255 ; (bad)
DB 255,227 ; jmpq *%rbx
DB 255 ; (bad)
@@ -1755,7 +2044,7 @@ _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 195f <_sk_store_565_hsw+0x6c>
+ DB 117,10 ; jne 1e4f <_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
@@ -1763,9 +2052,9 @@ _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 195b <_sk_store_565_hsw+0x68>
+ DB 119,236 ; ja 1e4b <_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 # 19bc <_sk_store_565_hsw+0xc9>
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1eac <_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
@@ -1776,7 +2065,7 @@ _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 195b <_sk_store_565_hsw+0x68>
+ DB 235,159 ; jmp 1e4b <_sk_store_565_hsw+0x68>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -1805,7 +2094,7 @@ _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 1a99 <_sk_load_4444_hsw+0xc1>
+ DB 15,133,179,0,0,0 ; jne 1f89 <_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
@@ -1851,9 +2140,9 @@ _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 19ec <_sk_load_4444_hsw+0x14>
+ DB 15,135,59,255,255,255 ; ja 1edc <_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 # 1b08 <_sk_load_4444_hsw+0x130>
+ DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1ff8 <_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
@@ -1865,13 +2154,13 @@ _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 19ec <_sk_load_4444_hsw+0x14>
+ DB 233,231,254,255,255 ; jmpq 1edc <_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 ffffffffe2001b10 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff234>
+ DB 233,255,255,255,225 ; jmpq ffffffffe2002000 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff234>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
@@ -1999,7 +2288,7 @@ _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 1cf7 <_sk_store_4444_hsw+0x72>
+ DB 117,10 ; jne 21e7 <_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
@@ -2007,9 +2296,9 @@ _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 1cf3 <_sk_store_4444_hsw+0x6e>
+ DB 119,236 ; ja 21e3 <_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 # 1d54 <_sk_store_4444_hsw+0xcf>
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2244 <_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
@@ -2020,7 +2309,7 @@ _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 1cf3 <_sk_store_4444_hsw+0x6e>
+ DB 235,159 ; jmp 21e3 <_sk_store_4444_hsw+0x6e>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2051,7 +2340,7 @@ _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 1ded <_sk_load_8888_hsw+0x7d>
+ DB 117,104 ; jne 22dd <_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
@@ -2084,7 +2373,7 @@ _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 1d8a <_sk_load_8888_hsw+0x1a>
+ DB 233,116,255,255,255 ; jmpq 227a <_sk_load_8888_hsw+0x1a>
PUBLIC _sk_gather_8888_hsw
_sk_gather_8888_hsw LABEL PROC
@@ -2144,7 +2433,7 @@ _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 1f10 <_sk_store_8888_hsw+0x74>
+ DB 117,12 ; jne 2400 <_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
@@ -2157,14 +2446,14 @@ _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 1f09 <_sk_store_8888_hsw+0x6d>
+ DB 235,211 ; jmp 23f9 <_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 1fa1 <_sk_load_f16_hsw+0x6b>
+ DB 117,97 ; jne 2491 <_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
@@ -2190,29 +2479,29 @@ _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 2000 <_sk_load_f16_hsw+0xca>
+ DB 116,79 ; je 24f0 <_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 2000 <_sk_load_f16_hsw+0xca>
+ DB 114,67 ; jb 24f0 <_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 200d <_sk_load_f16_hsw+0xd7>
+ DB 116,68 ; je 24fd <_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 200d <_sk_load_f16_hsw+0xd7>
+ DB 114,56 ; jb 24fd <_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 1f57 <_sk_load_f16_hsw+0x21>
+ DB 15,132,114,255,255,255 ; je 2447 <_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 1f57 <_sk_load_f16_hsw+0x21>
+ DB 15,130,98,255,255,255 ; jb 2447 <_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 1f57 <_sk_load_f16_hsw+0x21>
+ DB 233,87,255,255,255 ; jmpq 2447 <_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 1f57 <_sk_load_f16_hsw+0x21>
+ DB 233,74,255,255,255 ; jmpq 2447 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,65,255,255,255 ; jmpq 1f57 <_sk_load_f16_hsw+0x21>
+ DB 233,65,255,255,255 ; jmpq 2447 <_sk_load_f16_hsw+0x21>
PUBLIC _sk_gather_f16_hsw
_sk_gather_f16_hsw LABEL PROC
@@ -2266,7 +2555,7 @@ _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 2105 <_sk_store_f16_hsw+0x65>
+ DB 117,27 ; jne 25f5 <_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)
@@ -2275,29 +2564,29 @@ _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 2101 <_sk_store_f16_hsw+0x61>
+ DB 116,241 ; je 25f1 <_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 2101 <_sk_store_f16_hsw+0x61>
+ DB 114,229 ; jb 25f1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,221 ; je 2101 <_sk_store_f16_hsw+0x61>
+ DB 116,221 ; je 25f1 <_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 2101 <_sk_store_f16_hsw+0x61>
+ DB 114,209 ; jb 25f1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,201 ; je 2101 <_sk_store_f16_hsw+0x61>
+ DB 116,201 ; je 25f1 <_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 2101 <_sk_store_f16_hsw+0x61>
+ DB 114,189 ; jb 25f1 <_sk_store_f16_hsw+0x61>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,181 ; jmp 2101 <_sk_store_f16_hsw+0x61>
+ DB 235,181 ; jmp 25f1 <_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 2223 <_sk_load_u16_be_hsw+0xd7>
+ DB 15,133,201,0,0,0 ; jne 2713 <_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
@@ -2346,29 +2635,29 @@ _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 2282 <_sk_load_u16_be_hsw+0x136>
+ DB 116,79 ; je 2772 <_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 2282 <_sk_load_u16_be_hsw+0x136>
+ DB 114,67 ; jb 2772 <_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 228f <_sk_load_u16_be_hsw+0x143>
+ DB 116,68 ; je 277f <_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 228f <_sk_load_u16_be_hsw+0x143>
+ DB 114,56 ; jb 277f <_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 2171 <_sk_load_u16_be_hsw+0x25>
+ DB 15,132,10,255,255,255 ; je 2661 <_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 2171 <_sk_load_u16_be_hsw+0x25>
+ DB 15,130,250,254,255,255 ; jb 2661 <_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 2171 <_sk_load_u16_be_hsw+0x25>
+ DB 233,239,254,255,255 ; jmpq 2661 <_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 2171 <_sk_load_u16_be_hsw+0x25>
+ DB 233,226,254,255,255 ; jmpq 2661 <_sk_load_u16_be_hsw+0x25>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,217,254,255,255 ; jmpq 2171 <_sk_load_u16_be_hsw+0x25>
+ DB 233,217,254,255,255 ; jmpq 2661 <_sk_load_u16_be_hsw+0x25>
PUBLIC _sk_store_u16_be_hsw
_sk_store_u16_be_hsw LABEL PROC
@@ -2414,7 +2703,7 @@ _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 238b <_sk_store_u16_be_hsw+0xf3>
+ DB 117,31 ; jne 287b <_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)
@@ -2423,31 +2712,31 @@ _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 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 116,240 ; je 2877 <_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 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 114,227 ; jb 2877 <_sk_store_u16_be_hsw+0xef>
DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8)
- DB 116,218 ; je 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 116,218 ; je 2877 <_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 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 114,205 ; jb 2877 <_sk_store_u16_be_hsw+0xef>
DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8)
- DB 116,196 ; je 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 116,196 ; je 2877 <_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 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 114,183 ; jb 2877 <_sk_store_u16_be_hsw+0xef>
DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8)
- DB 235,174 ; jmp 2387 <_sk_store_u16_be_hsw+0xef>
+ DB 235,174 ; jmp 2877 <_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 244f <_sk_load_f32_hsw+0x76>
+ DB 119,110 ; ja 293f <_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 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 2478 <_sk_load_f32_hsw+0x9f>
+ DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 2968 <_sk_load_f32_hsw+0x9f>
DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -2503,7 +2792,7 @@ _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 2505 <_sk_store_f32_hsw+0x6d>
+ DB 117,55 ; jne 29f5 <_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
@@ -2516,22 +2805,22 @@ _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 2501 <_sk_store_f32_hsw+0x69>
+ DB 116,240 ; je 29f1 <_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 2501 <_sk_store_f32_hsw+0x69>
+ DB 114,227 ; jb 29f1 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 2501 <_sk_store_f32_hsw+0x69>
+ DB 116,218 ; je 29f1 <_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 2501 <_sk_store_f32_hsw+0x69>
+ DB 114,205 ; jb 29f1 <_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 2501 <_sk_store_f32_hsw+0x69>
+ DB 116,195 ; je 29f1 <_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 2501 <_sk_store_f32_hsw+0x69>
+ DB 114,181 ; jb 29f1 <_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 2501 <_sk_store_f32_hsw+0x69>
+ DB 235,171 ; jmp 29f1 <_sk_store_f32_hsw+0x69>
PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
@@ -4297,6 +4586,297 @@ _sk_load_tables_avx LABEL PROC
DB 128,255,255 ; cmp $0xff,%bh
DB 255 ; .byte 0xff
+PUBLIC _sk_byte_tables_avx
+_sk_byte_tables_avx LABEL PROC
+ 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 72,173 ; lods %ds:(%rsi),%rax
+ DB 65,184,0,0,127,67 ; mov $0x437f0000,%r8d
+ DB 196,65,121,110,192 ; vmovd %r8d,%xmm8
+ DB 196,67,121,4,192,0 ; vpermilps $0x0,%xmm8,%xmm8
+ DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
+ DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
+ DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 68,137,197 ; mov %r8d,%ebp
+ DB 77,137,194 ; mov %r8,%r10
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,192 ; vmovq %xmm0,%r8
+ DB 69,137,195 ; mov %r8d,%r11d
+ DB 77,137,199 ; mov %r8,%r15
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 69,137,198 ; mov %r8d,%r14d
+ DB 77,137,196 ; mov %r8,%r12
+ DB 73,193,236,32 ; shr $0x20,%r12
+ DB 196,225,249,126,195 ; vmovq %xmm0,%rbx
+ DB 65,137,221 ; mov %ebx,%r13d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 76,139,8 ; mov (%rax),%r9
+ DB 76,139,64,8 ; mov 0x8(%rax),%r8
+ DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
+ DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
+ DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp
+ DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp
+ DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0
+ DB 196,98,121,49,200 ; vpmovzxbd %xmm0,%xmm9
+ DB 196,131,121,32,4,41,0 ; vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0
+ DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0
+ DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp
+ DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp
+ DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0
+ DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
+ DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0
+ DB 197,124,91,208 ; vcvtdq2ps %ymm0,%ymm10
+ DB 189,129,128,128,59 ; mov $0x3b808081,%ebp
+ DB 197,249,110,197 ; vmovd %ebp,%xmm0
+ DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0
+ DB 196,99,125,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm9
+ DB 196,193,44,89,193 ; vmulps %ymm9,%ymm10,%ymm0
+ DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
+ DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1
+ DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp
+ DB 65,137,233 ; mov %ebp,%r9d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,203 ; vmovq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1
+ DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 196,193,249,126,207 ; vmovq %xmm1,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
+ DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
+ DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx
+ DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1
+ DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp
+ DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1
+ DB 196,98,121,49,209 ; vpmovzxbd %xmm1,%xmm10
+ DB 196,131,121,32,12,32,0 ; vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1
+ DB 196,131,113,32,12,56,1 ; vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1
+ DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp
+ DB 196,227,113,32,205,2 ; vpinsrb $0x2,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp
+ DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1
+ DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1
+ DB 196,227,45,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
+ DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
+ DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1
+ DB 76,139,64,16 ; mov 0x10(%rax),%r8
+ DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
+ DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2
+ DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp
+ DB 65,137,233 ; mov %ebp,%r9d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,211 ; vmovq %xmm2,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2
+ DB 196,195,249,22,211,1 ; vpextrq $0x1,%xmm2,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 196,193,249,126,215 ; vmovq %xmm2,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,131,121,32,20,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2
+ DB 196,195,105,32,20,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2
+ DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx
+ DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2
+ DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp
+ DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2
+ DB 196,98,121,49,210 ; vpmovzxbd %xmm2,%xmm10
+ DB 196,131,121,32,20,32,0 ; vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm2
+ DB 196,131,105,32,20,56,1 ; vpinsrb $0x1,(%r8,%r15,1),%xmm2,%xmm2
+ DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp
+ DB 196,227,105,32,213,2 ; vpinsrb $0x2,%ebp,%xmm2,%xmm2
+ DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp
+ DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2
+ DB 196,226,121,49,210 ; vpmovzxbd %xmm2,%xmm2
+ DB 196,227,45,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm10,%ymm2
+ DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ DB 72,139,64,24 ; mov 0x18(%rax),%rax
+ DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3
+ DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3
+ DB 196,227,249,22,221,1 ; vpextrq $0x1,%xmm3,%rbp
+ DB 65,137,232 ; mov %ebp,%r8d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,219 ; vmovq %xmm3,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,25,219,1 ; vextractf128 $0x1,%ymm3,%xmm3
+ DB 196,195,249,22,218,1 ; vpextrq $0x1,%xmm3,%r10
+ DB 69,137,211 ; mov %r10d,%r11d
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,222 ; vmovq %xmm3,%r14
+ DB 69,137,247 ; mov %r14d,%r15d
+ DB 73,193,238,32 ; shr $0x20,%r14
+ DB 196,163,121,32,28,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3
+ DB 196,227,97,32,28,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3
+ DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx
+ DB 196,227,97,32,219,2 ; vpinsrb $0x2,%ebx,%xmm3,%xmm3
+ DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp
+ DB 196,227,97,32,221,3 ; vpinsrb $0x3,%ebp,%xmm3,%xmm3
+ DB 196,98,121,49,195 ; vpmovzxbd %xmm3,%xmm8
+ DB 196,163,121,32,28,56,0 ; vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm3
+ DB 196,163,97,32,28,48,1 ; vpinsrb $0x1,(%rax,%r14,1),%xmm3,%xmm3
+ DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp
+ DB 196,227,97,32,221,2 ; vpinsrb $0x2,%ebp,%xmm3,%xmm3
+ DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax
+ DB 196,227,97,32,216,3 ; vpinsrb $0x3,%eax,%xmm3,%xmm3
+ DB 196,226,121,49,219 ; vpmovzxbd %xmm3,%xmm3
+ DB 196,227,61,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
+ DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
+ DB 197,180,89,219 ; vmulps %ymm3,%ymm9,%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 93 ; pop %rbp
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_byte_tables_rgb_avx
+_sk_byte_tables_rgb_avx LABEL PROC
+ 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 72,173 ; lods %ds:(%rsi),%rax
+ DB 68,139,64,24 ; mov 0x18(%rax),%r8d
+ DB 65,255,200 ; dec %r8d
+ DB 196,65,121,110,192 ; vmovd %r8d,%xmm8
+ DB 196,65,121,112,192,0 ; vpshufd $0x0,%xmm8,%xmm8
+ DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
+ DB 196,65,124,91,192 ; vcvtdq2ps %ymm8,%ymm8
+ DB 197,188,89,192 ; vmulps %ymm0,%ymm8,%ymm0
+ DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 68,137,197 ; mov %r8d,%ebp
+ DB 77,137,194 ; mov %r8,%r10
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,192 ; vmovq %xmm0,%r8
+ DB 69,137,195 ; mov %r8d,%r11d
+ DB 77,137,199 ; mov %r8,%r15
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0
+ DB 196,195,249,22,192,1 ; vpextrq $0x1,%xmm0,%r8
+ DB 69,137,198 ; mov %r8d,%r14d
+ DB 77,137,196 ; mov %r8,%r12
+ DB 73,193,236,32 ; shr $0x20,%r12
+ DB 196,225,249,126,195 ; vmovq %xmm0,%rbx
+ DB 65,137,221 ; mov %ebx,%r13d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 76,139,8 ; mov (%rax),%r9
+ DB 76,139,64,8 ; mov 0x8(%rax),%r8
+ DB 196,131,121,32,4,25,0 ; vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
+ DB 196,131,121,32,4,57,1 ; vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
+ DB 65,15,182,44,41 ; movzbl (%r9,%rbp,1),%ebp
+ DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,17 ; movzbl (%r9,%r10,1),%ebp
+ DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0
+ DB 196,98,121,49,200 ; vpmovzxbd %xmm0,%xmm9
+ DB 196,131,121,32,4,41,0 ; vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0
+ DB 196,195,121,32,4,25,1 ; vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0
+ DB 67,15,182,44,49 ; movzbl (%r9,%r14,1),%ebp
+ DB 196,227,121,32,197,2 ; vpinsrb $0x2,%ebp,%xmm0,%xmm0
+ DB 67,15,182,44,33 ; movzbl (%r9,%r12,1),%ebp
+ DB 196,227,121,32,197,3 ; vpinsrb $0x3,%ebp,%xmm0,%xmm0
+ DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
+ DB 196,227,53,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm9,%ymm0
+ DB 197,124,91,208 ; vcvtdq2ps %ymm0,%ymm10
+ DB 189,129,128,128,59 ; mov $0x3b808081,%ebp
+ DB 197,249,110,197 ; vmovd %ebp,%xmm0
+ DB 196,227,121,4,192,0 ; vpermilps $0x0,%xmm0,%xmm0
+ DB 196,99,125,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm9
+ DB 196,193,44,89,193 ; vmulps %ymm9,%ymm10,%ymm0
+ DB 197,188,89,201 ; vmulps %ymm1,%ymm8,%ymm1
+ DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1
+ DB 196,227,249,22,205,1 ; vpextrq $0x1,%xmm1,%rbp
+ DB 65,137,233 ; mov %ebp,%r9d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,203 ; vmovq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1
+ DB 196,195,249,22,203,1 ; vpextrq $0x1,%xmm1,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 196,193,249,126,207 ; vmovq %xmm1,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,32 ; shr $0x20,%r15
+ DB 196,131,121,32,12,16,0 ; vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
+ DB 196,195,113,32,12,24,1 ; vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
+ DB 67,15,182,28,8 ; movzbl (%r8,%r9,1),%ebx
+ DB 196,227,113,32,203,2 ; vpinsrb $0x2,%ebx,%xmm1,%xmm1
+ DB 65,15,182,44,40 ; movzbl (%r8,%rbp,1),%ebp
+ DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1
+ DB 196,98,121,49,209 ; vpmovzxbd %xmm1,%xmm10
+ DB 196,131,121,32,12,32,0 ; vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1
+ DB 196,131,113,32,12,56,1 ; vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1
+ DB 67,15,182,44,48 ; movzbl (%r8,%r14,1),%ebp
+ DB 196,227,113,32,205,2 ; vpinsrb $0x2,%ebp,%xmm1,%xmm1
+ DB 67,15,182,44,24 ; movzbl (%r8,%r11,1),%ebp
+ DB 196,227,113,32,205,3 ; vpinsrb $0x3,%ebp,%xmm1,%xmm1
+ DB 196,226,121,49,201 ; vpmovzxbd %xmm1,%xmm1
+ DB 196,227,45,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
+ DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
+ DB 197,180,89,201 ; vmulps %ymm1,%ymm9,%ymm1
+ DB 72,139,64,16 ; mov 0x10(%rax),%rax
+ DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
+ DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2
+ DB 196,227,249,22,213,1 ; vpextrq $0x1,%xmm2,%rbp
+ DB 65,137,232 ; mov %ebp,%r8d
+ DB 72,193,237,32 ; shr $0x20,%rbp
+ DB 196,225,249,126,211 ; vmovq %xmm2,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2
+ DB 196,195,249,22,210,1 ; vpextrq $0x1,%xmm2,%r10
+ DB 69,137,211 ; mov %r10d,%r11d
+ DB 73,193,234,32 ; shr $0x20,%r10
+ DB 196,193,249,126,214 ; vmovq %xmm2,%r14
+ DB 69,137,247 ; mov %r14d,%r15d
+ DB 73,193,238,32 ; shr $0x20,%r14
+ DB 196,163,121,32,20,8,0 ; vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2
+ DB 196,227,105,32,20,24,1 ; vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2
+ DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx
+ DB 196,227,105,32,211,2 ; vpinsrb $0x2,%ebx,%xmm2,%xmm2
+ DB 15,182,44,40 ; movzbl (%rax,%rbp,1),%ebp
+ DB 196,227,105,32,213,3 ; vpinsrb $0x3,%ebp,%xmm2,%xmm2
+ DB 196,98,121,49,194 ; vpmovzxbd %xmm2,%xmm8
+ DB 196,163,121,32,20,56,0 ; vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm2
+ DB 196,163,105,32,20,48,1 ; vpinsrb $0x1,(%rax,%r14,1),%xmm2,%xmm2
+ DB 66,15,182,44,24 ; movzbl (%rax,%r11,1),%ebp
+ DB 196,227,105,32,213,2 ; vpinsrb $0x2,%ebp,%xmm2,%xmm2
+ DB 66,15,182,4,16 ; movzbl (%rax,%r10,1),%eax
+ DB 196,227,105,32,208,3 ; vpinsrb $0x3,%eax,%xmm2,%xmm2
+ DB 196,226,121,49,210 ; vpmovzxbd %xmm2,%xmm2
+ DB 196,227,61,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm8,%ymm2
+ DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
+ DB 197,180,89,210 ; vmulps %ymm2,%ymm9,%ymm2
+ 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 93 ; pop %rbp
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_a8_avx
_sk_load_a8_avx LABEL PROC
DB 73,137,200 ; mov %rcx,%r8
@@ -4304,7 +4884,7 @@ _sk_load_a8_avx LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,74 ; jne 177a <_sk_load_a8_avx+0x5a>
+ DB 117,74 ; jne 1c98 <_sk_load_a8_avx+0x5a>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
@@ -4331,9 +4911,9 @@ _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 1782 <_sk_load_a8_avx+0x62>
+ DB 117,234 ; jne 1ca0 <_sk_load_a8_avx+0x62>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,149 ; jmp 1734 <_sk_load_a8_avx+0x14>
+ DB 235,149 ; jmp 1c52 <_sk_load_a8_avx+0x14>
PUBLIC _sk_gather_a8_avx
_sk_gather_a8_avx LABEL PROC
@@ -4410,7 +4990,7 @@ _sk_store_a8_avx LABEL PROC
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 196,65,57,103,192 ; vpackuswb %xmm8,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 18db <_sk_store_a8_avx+0x42>
+ DB 117,10 ; jne 1df9 <_sk_store_a8_avx+0x42>
DB 196,65,123,17,4,57 ; vmovsd %xmm8,(%r9,%rdi,1)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -4418,10 +4998,10 @@ _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 18d7 <_sk_store_a8_avx+0x3e>
+ DB 119,236 ; ja 1df5 <_sk_store_a8_avx+0x3e>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 65,15,182,192 ; movzbl %r8b,%eax
- DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 1940 <_sk_store_a8_avx+0xa7>
+ DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1e5c <_sk_store_a8_avx+0xa5>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -4432,28 +5012,27 @@ _sk_store_a8_avx LABEL PROC
DB 196,67,121,20,68,57,2,4 ; vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1)
DB 196,67,121,20,68,57,1,2 ; vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1)
DB 196,67,121,20,4,57,0 ; vpextrb $0x0,%xmm8,(%r9,%rdi,1)
- DB 235,154 ; jmp 18d7 <_sk_store_a8_avx+0x3e>
- DB 15,31,0 ; nopl (%rax)
- DB 244 ; hlt
- DB 255 ; (bad)
+ DB 235,154 ; jmp 1df5 <_sk_store_a8_avx+0x3e>
+ DB 144 ; nop
+ DB 246,255 ; idiv %bh
DB 255 ; (bad)
DB 255 ; (bad)
- DB 236 ; in (%dx),%al
+ DB 238 ; out %al,(%dx)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,228 ; jmpq *%rsp
+ DB 255,230 ; jmpq *%rsi
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 220,255 ; fdivr %st,%st(7)
+ DB 222,255 ; fdivrp %st,%st(7)
DB 255 ; (bad)
- DB 255,212 ; callq *%rsp
+ DB 255,214 ; callq *%rsi
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,204 ; dec %esp
+ DB 255,206 ; dec %esi
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,196 ; inc %esp
+ DB 255,198 ; inc %esi
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -4465,7 +5044,7 @@ _sk_load_g8_avx LABEL PROC
DB 72,139,0 ; mov (%rax),%rax
DB 72,1,248 ; add %rdi,%rax
DB 77,133,192 ; test %r8,%r8
- DB 117,91 ; jne 19c7 <_sk_load_g8_avx+0x6b>
+ DB 117,91 ; jne 1ee3 <_sk_load_g8_avx+0x6b>
DB 197,250,126,0 ; vmovq (%rax),%xmm0
DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
@@ -4495,9 +5074,9 @@ _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 19cf <_sk_load_g8_avx+0x73>
+ DB 117,234 ; jne 1eeb <_sk_load_g8_avx+0x73>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,132 ; jmp 1970 <_sk_load_g8_avx+0x14>
+ DB 235,132 ; jmp 1e8c <_sk_load_g8_avx+0x14>
PUBLIC _sk_gather_g8_avx
_sk_gather_g8_avx LABEL PROC
@@ -4568,9 +5147,9 @@ _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 1b06 <_sk_gather_i8_avx+0xf>
+ DB 116,5 ; je 2022 <_sk_gather_i8_avx+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 1b08 <_sk_gather_i8_avx+0x11>
+ DB 235,2 ; jmp 2024 <_sk_gather_i8_avx+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -4673,7 +5252,7 @@ _sk_load_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,209,0,0,0 ; jne 1da2 <_sk_load_565_avx+0xdf>
+ DB 15,133,209,0,0,0 ; jne 22be <_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
@@ -4723,9 +5302,9 @@ _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 1cd7 <_sk_load_565_avx+0x14>
+ DB 15,135,29,255,255,255 ; ja 21f3 <_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 # 1e10 <_sk_load_565_avx+0x14d>
+ DB 76,141,13,75,0,0,0 ; lea 0x4b(%rip),%r9 # 232c <_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
@@ -4737,7 +5316,7 @@ _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 1cd7 <_sk_load_565_avx+0x14>
+ DB 233,201,254,255,255 ; jmpq 21f3 <_sk_load_565_avx+0x14>
DB 102,144 ; xchg %ax,%ax
DB 242,255 ; repnz (bad)
DB 255 ; (bad)
@@ -4890,7 +5469,7 @@ _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 205b <_sk_store_565_avx+0x9e>
+ DB 117,10 ; jne 2577 <_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
@@ -4898,9 +5477,9 @@ _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 2057 <_sk_store_565_avx+0x9a>
+ DB 119,236 ; ja 2573 <_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 # 20b8 <_sk_store_565_avx+0xfb>
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 25d4 <_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
@@ -4911,7 +5490,7 @@ _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 2057 <_sk_store_565_avx+0x9a>
+ DB 235,159 ; jmp 2573 <_sk_store_565_avx+0x9a>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -4940,7 +5519,7 @@ _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 21d7 <_sk_load_4444_avx+0x103>
+ DB 15,133,245,0,0,0 ; jne 26f3 <_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
@@ -4997,9 +5576,9 @@ _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 20e8 <_sk_load_4444_avx+0x14>
+ DB 15,135,249,254,255,255 ; ja 2604 <_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 # 2244 <_sk_load_4444_avx+0x170>
+ DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 2760 <_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
@@ -5011,12 +5590,12 @@ _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 20e8 <_sk_load_4444_avx+0x14>
+ DB 233,165,254,255,255 ; jmpq 2604 <_sk_load_4444_avx+0x14>
DB 144 ; nop
DB 243,255 ; repz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 2249 <_sk_load_4444_avx+0x175>
+ DB 235,255 ; jmp 2765 <_sk_load_4444_avx+0x175>
DB 255 ; (bad)
DB 255,227 ; jmpq *%rbx
DB 255 ; (bad)
@@ -5173,7 +5752,7 @@ _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 24c4 <_sk_store_4444_avx+0xaf>
+ DB 117,10 ; jne 29e0 <_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
@@ -5181,9 +5760,9 @@ _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 24c0 <_sk_store_4444_avx+0xab>
+ DB 119,236 ; ja 29dc <_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 # 2524 <_sk_store_4444_avx+0x10f>
+ DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 2a40 <_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
@@ -5194,7 +5773,7 @@ _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 24c0 <_sk_store_4444_avx+0xab>
+ DB 235,159 ; jmp 29dc <_sk_store_4444_avx+0xab>
DB 15,31,0 ; nopl (%rax)
DB 244 ; hlt
DB 255 ; (bad)
@@ -5225,7 +5804,7 @@ _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 25eb <_sk_load_8888_avx+0xab>
+ DB 15,133,157,0,0,0 ; jne 2b07 <_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
@@ -5263,9 +5842,9 @@ _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 2554 <_sk_load_8888_avx+0x14>
+ DB 15,135,80,255,255,255 ; ja 2a70 <_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 # 2698 <_sk_load_8888_avx+0x158>
+ DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 2bb4 <_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
@@ -5288,7 +5867,7 @@ _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 2554 <_sk_load_8888_avx+0x14>
+ DB 233,188,254,255,255 ; jmpq 2a70 <_sk_load_8888_avx+0x14>
DB 238 ; out %al,(%dx)
DB 255 ; (bad)
DB 255 ; (bad)
@@ -5414,7 +5993,7 @@ _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 2899 <_sk_store_8888_avx+0xa4>
+ DB 117,10 ; jne 2db5 <_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
@@ -5422,9 +6001,9 @@ _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 2895 <_sk_store_8888_avx+0xa0>
+ DB 119,236 ; ja 2db1 <_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 # 2908 <_sk_store_8888_avx+0x113>
+ DB 76,141,5,84,0,0,0 ; lea 0x54(%rip),%r8 # 2e24 <_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
@@ -5438,7 +6017,7 @@ _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 2895 <_sk_store_8888_avx+0xa0>
+ DB 235,143 ; jmp 2db1 <_sk_store_8888_avx+0xa0>
DB 102,144 ; xchg %ax,%ax
DB 246,255 ; idiv %bh
DB 255 ; (bad)
@@ -5468,7 +6047,7 @@ _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 2a43 <_sk_load_f16_avx+0x11f>
+ DB 15,133,17,1,0,0 ; jne 2f5f <_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
@@ -5530,29 +6109,29 @@ _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 2aa2 <_sk_load_f16_avx+0x17e>
+ DB 116,79 ; je 2fbe <_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 2aa2 <_sk_load_f16_avx+0x17e>
+ DB 114,67 ; jb 2fbe <_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 2aaf <_sk_load_f16_avx+0x18b>
+ DB 116,68 ; je 2fcb <_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 2aaf <_sk_load_f16_avx+0x18b>
+ DB 114,56 ; jb 2fcb <_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 2949 <_sk_load_f16_avx+0x25>
+ DB 15,132,194,254,255,255 ; je 2e65 <_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 2949 <_sk_load_f16_avx+0x25>
+ DB 15,130,178,254,255,255 ; jb 2e65 <_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 2949 <_sk_load_f16_avx+0x25>
+ DB 233,167,254,255,255 ; jmpq 2e65 <_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 2949 <_sk_load_f16_avx+0x25>
+ DB 233,154,254,255,255 ; jmpq 2e65 <_sk_load_f16_avx+0x25>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,145,254,255,255 ; jmpq 2949 <_sk_load_f16_avx+0x25>
+ DB 233,145,254,255,255 ; jmpq 2e65 <_sk_load_f16_avx+0x25>
PUBLIC _sk_gather_f16_avx
_sk_gather_f16_avx LABEL PROC
@@ -5692,7 +6271,7 @@ _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 2d38 <_sk_store_f16_avx+0xd2>
+ DB 117,31 ; jne 3254 <_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)
@@ -5701,29 +6280,29 @@ _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 2d34 <_sk_store_f16_avx+0xce>
+ DB 116,240 ; je 3250 <_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 2d34 <_sk_store_f16_avx+0xce>
+ DB 114,227 ; jb 3250 <_sk_store_f16_avx+0xce>
DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8)
- DB 116,218 ; je 2d34 <_sk_store_f16_avx+0xce>
+ DB 116,218 ; je 3250 <_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 2d34 <_sk_store_f16_avx+0xce>
+ DB 114,205 ; jb 3250 <_sk_store_f16_avx+0xce>
DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8)
- DB 116,196 ; je 2d34 <_sk_store_f16_avx+0xce>
+ DB 116,196 ; je 3250 <_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 2d34 <_sk_store_f16_avx+0xce>
+ DB 114,183 ; jb 3250 <_sk_store_f16_avx+0xce>
DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8)
- DB 235,174 ; jmp 2d34 <_sk_store_f16_avx+0xce>
+ DB 235,174 ; jmp 3250 <_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 2e95 <_sk_load_u16_be_avx+0x10f>
+ DB 15,133,1,1,0,0 ; jne 33b1 <_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
@@ -5782,29 +6361,29 @@ _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 2ef4 <_sk_load_u16_be_avx+0x16e>
+ DB 116,79 ; je 3410 <_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 2ef4 <_sk_load_u16_be_avx+0x16e>
+ DB 114,67 ; jb 3410 <_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 2f01 <_sk_load_u16_be_avx+0x17b>
+ DB 116,68 ; je 341d <_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 2f01 <_sk_load_u16_be_avx+0x17b>
+ DB 114,56 ; jb 341d <_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 2dab <_sk_load_u16_be_avx+0x25>
+ DB 15,132,210,254,255,255 ; je 32c7 <_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 2dab <_sk_load_u16_be_avx+0x25>
+ DB 15,130,194,254,255,255 ; jb 32c7 <_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 2dab <_sk_load_u16_be_avx+0x25>
+ DB 233,183,254,255,255 ; jmpq 32c7 <_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 2dab <_sk_load_u16_be_avx+0x25>
+ DB 233,170,254,255,255 ; jmpq 32c7 <_sk_load_u16_be_avx+0x25>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,161,254,255,255 ; jmpq 2dab <_sk_load_u16_be_avx+0x25>
+ DB 233,161,254,255,255 ; jmpq 32c7 <_sk_load_u16_be_avx+0x25>
PUBLIC _sk_store_u16_be_avx
_sk_store_u16_be_avx LABEL PROC
@@ -5851,7 +6430,7 @@ _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 3004 <_sk_store_u16_be_avx+0xfa>
+ DB 117,31 ; jne 3520 <_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)
@@ -5860,31 +6439,31 @@ _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 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 116,240 ; je 351c <_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 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 114,227 ; jb 351c <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8)
- DB 116,218 ; je 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 116,218 ; je 351c <_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 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 114,205 ; jb 351c <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8)
- DB 116,196 ; je 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 116,196 ; je 351c <_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 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 114,183 ; jb 351c <_sk_store_u16_be_avx+0xf6>
DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8)
- DB 235,174 ; jmp 3000 <_sk_store_u16_be_avx+0xf6>
+ DB 235,174 ; jmp 351c <_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 30c8 <_sk_load_f32_avx+0x76>
+ DB 119,110 ; ja 35e4 <_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 76,141,21,132,0,0,0 ; lea 0x84(%rip),%r10 # 30f0 <_sk_load_f32_avx+0x9e>
+ DB 76,141,21,132,0,0,0 ; lea 0x84(%rip),%r10 # 360c <_sk_load_f32_avx+0x9e>
DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -5941,7 +6520,7 @@ _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 317d <_sk_store_f32_avx+0x6d>
+ DB 117,55 ; jne 3699 <_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
@@ -5954,22 +6533,22 @@ _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 3179 <_sk_store_f32_avx+0x69>
+ DB 116,240 ; je 3695 <_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 3179 <_sk_store_f32_avx+0x69>
+ DB 114,227 ; jb 3695 <_sk_store_f32_avx+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 3179 <_sk_store_f32_avx+0x69>
+ DB 116,218 ; je 3695 <_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 3179 <_sk_store_f32_avx+0x69>
+ DB 114,205 ; jb 3695 <_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 3179 <_sk_store_f32_avx+0x69>
+ DB 116,195 ; je 3695 <_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 3179 <_sk_store_f32_avx+0x69>
+ DB 114,181 ; jb 3695 <_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 3179 <_sk_store_f32_avx+0x69>
+ DB 235,171 ; jmp 3695 <_sk_store_f32_avx+0x69>
PUBLIC _sk_clamp_x_avx
_sk_clamp_x_avx LABEL PROC
@@ -7793,6 +8372,168 @@ _sk_load_tables_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_byte_tables_sse41
+_sk_byte_tables_sse41 LABEL PROC
+ DB 65,86 ; push %r14
+ DB 83 ; push %rbx
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 185,0,0,127,67 ; mov $0x437f0000,%ecx
+ DB 102,68,15,110,193 ; movd %ecx,%xmm8
+ DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0
+ DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,73,15,126,193 ; movq %xmm0,%r9
+ DB 69,137,202 ; mov %r9d,%r10d
+ DB 77,137,203 ; mov %r9,%r11
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 76,139,48 ; mov (%rax),%r14
+ DB 76,139,72,8 ; mov 0x8(%rax),%r9
+ DB 102,67,15,58,32,4,22,0 ; pinsrb $0x0,(%r14,%r10,1),%xmm0
+ DB 102,67,15,58,32,4,30,1 ; pinsrb $0x1,(%r14,%r11,1),%xmm0
+ DB 67,15,182,28,6 ; movzbl (%r14,%r8,1),%ebx
+ DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0
+ DB 65,15,182,12,14 ; movzbl (%r14,%rcx,1),%ecx
+ DB 102,15,58,32,193,3 ; pinsrb $0x3,%ecx,%xmm0
+ DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0
+ DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
+ DB 185,129,128,128,59 ; mov $0x3b808081,%ecx
+ DB 102,68,15,110,201 ; movd %ecx,%xmm9
+ DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9
+ DB 65,15,89,193 ; mulps %xmm9,%xmm0
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
+ DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,72,15,126,203 ; movq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 102,67,15,58,32,12,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm1
+ DB 102,65,15,58,32,12,25,1 ; pinsrb $0x1,(%r9,%rbx,1),%xmm1
+ DB 67,15,182,28,1 ; movzbl (%r9,%r8,1),%ebx
+ DB 102,15,58,32,203,2 ; pinsrb $0x2,%ebx,%xmm1
+ DB 65,15,182,12,9 ; movzbl (%r9,%rcx,1),%ecx
+ DB 102,15,58,32,201,3 ; pinsrb $0x3,%ecx,%xmm1
+ DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1
+ DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 76,139,72,16 ; mov 0x10(%rax),%r9
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2
+ DB 102,72,15,58,22,211,1 ; pextrq $0x1,%xmm2,%rbx
+ DB 65,137,216 ; mov %ebx,%r8d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 102,72,15,126,209 ; movq %xmm2,%rcx
+ DB 65,137,202 ; mov %ecx,%r10d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,67,15,58,32,20,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm2
+ DB 102,65,15,58,32,20,9,1 ; pinsrb $0x1,(%r9,%rcx,1),%xmm2
+ DB 67,15,182,12,1 ; movzbl (%r9,%r8,1),%ecx
+ DB 102,15,58,32,209,2 ; pinsrb $0x2,%ecx,%xmm2
+ DB 65,15,182,12,25 ; movzbl (%r9,%rbx,1),%ecx
+ DB 102,15,58,32,209,3 ; pinsrb $0x3,%ecx,%xmm2
+ DB 102,15,56,49,210 ; pmovzxbd %xmm2,%xmm2
+ DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 65,15,89,209 ; mulps %xmm9,%xmm2
+ DB 72,139,64,24 ; mov 0x18(%rax),%rax
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 102,65,15,91,216 ; cvtps2dq %xmm8,%xmm3
+ DB 102,72,15,58,22,217,1 ; pextrq $0x1,%xmm3,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,72,15,126,219 ; movq %xmm3,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 102,66,15,58,32,28,8,0 ; pinsrb $0x0,(%rax,%r9,1),%xmm3
+ DB 102,15,58,32,28,24,1 ; pinsrb $0x1,(%rax,%rbx,1),%xmm3
+ DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx
+ DB 102,15,58,32,219,2 ; pinsrb $0x2,%ebx,%xmm3
+ DB 15,182,4,8 ; movzbl (%rax,%rcx,1),%eax
+ DB 102,15,58,32,216,3 ; pinsrb $0x3,%eax,%xmm3
+ DB 102,15,56,49,219 ; pmovzxbd %xmm3,%xmm3
+ DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
+ DB 65,15,89,217 ; mulps %xmm9,%xmm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 91 ; pop %rbx
+ DB 65,94 ; pop %r14
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_byte_tables_rgb_sse41
+_sk_byte_tables_rgb_sse41 LABEL PROC
+ DB 65,86 ; push %r14
+ DB 83 ; push %rbx
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 139,72,24 ; mov 0x18(%rax),%ecx
+ DB 255,201 ; dec %ecx
+ DB 102,68,15,110,193 ; movd %ecx,%xmm8
+ DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8
+ DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0
+ DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,73,15,126,193 ; movq %xmm0,%r9
+ DB 69,137,202 ; mov %r9d,%r10d
+ DB 77,137,203 ; mov %r9,%r11
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 76,139,48 ; mov (%rax),%r14
+ DB 76,139,72,8 ; mov 0x8(%rax),%r9
+ DB 102,67,15,58,32,4,22,0 ; pinsrb $0x0,(%r14,%r10,1),%xmm0
+ DB 102,67,15,58,32,4,30,1 ; pinsrb $0x1,(%r14,%r11,1),%xmm0
+ DB 67,15,182,28,6 ; movzbl (%r14,%r8,1),%ebx
+ DB 102,15,58,32,195,2 ; pinsrb $0x2,%ebx,%xmm0
+ DB 65,15,182,12,14 ; movzbl (%r14,%rcx,1),%ecx
+ DB 102,15,58,32,193,3 ; pinsrb $0x3,%ecx,%xmm0
+ DB 102,15,56,49,192 ; pmovzxbd %xmm0,%xmm0
+ DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
+ DB 185,129,128,128,59 ; mov $0x3b808081,%ecx
+ DB 102,68,15,110,201 ; movd %ecx,%xmm9
+ DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9
+ DB 65,15,89,193 ; mulps %xmm9,%xmm0
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
+ DB 102,72,15,58,22,201,1 ; pextrq $0x1,%xmm1,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,72,15,126,203 ; movq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 102,67,15,58,32,12,17,0 ; pinsrb $0x0,(%r9,%r10,1),%xmm1
+ DB 102,65,15,58,32,12,25,1 ; pinsrb $0x1,(%r9,%rbx,1),%xmm1
+ DB 67,15,182,28,1 ; movzbl (%r9,%r8,1),%ebx
+ DB 102,15,58,32,203,2 ; pinsrb $0x2,%ebx,%xmm1
+ DB 65,15,182,12,9 ; movzbl (%r9,%rcx,1),%ecx
+ DB 102,15,58,32,201,3 ; pinsrb $0x3,%ecx,%xmm1
+ DB 102,15,56,49,201 ; pmovzxbd %xmm1,%xmm1
+ DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
+ DB 65,15,89,201 ; mulps %xmm9,%xmm1
+ DB 72,139,64,16 ; mov 0x10(%rax),%rax
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2
+ DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,72,15,126,211 ; movq %xmm2,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 102,66,15,58,32,20,8,0 ; pinsrb $0x0,(%rax,%r9,1),%xmm2
+ DB 102,15,58,32,20,24,1 ; pinsrb $0x1,(%rax,%rbx,1),%xmm2
+ DB 66,15,182,28,0 ; movzbl (%rax,%r8,1),%ebx
+ DB 102,15,58,32,211,2 ; pinsrb $0x2,%ebx,%xmm2
+ DB 15,182,4,8 ; movzbl (%rax,%rcx,1),%eax
+ DB 102,15,58,32,208,3 ; pinsrb $0x3,%eax,%xmm2
+ DB 102,15,56,49,210 ; pmovzxbd %xmm2,%xmm2
+ DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 65,15,89,209 ; mulps %xmm9,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 91 ; pop %rbx
+ DB 65,94 ; pop %r14
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_a8_sse41
_sk_load_a8_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -7917,9 +8658,9 @@ _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 116,5 ; je 18d8 <_sk_gather_i8_sse41+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 1618 <_sk_gather_i8_sse41+0x11>
+ DB 235,2 ; jmp 18da <_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
@@ -10531,6 +11272,212 @@ _sk_load_tables_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_byte_tables_sse2
+_sk_byte_tables_sse2 LABEL PROC
+ DB 65,86 ; push %r14
+ DB 83 ; push %rbx
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 185,0,0,127,67 ; mov $0x437f0000,%ecx
+ DB 102,68,15,110,193 ; movd %ecx,%xmm8
+ DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0
+ DB 102,72,15,126,193 ; movq %xmm0,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0
+ DB 102,73,15,126,193 ; movq %xmm0,%r9
+ DB 69,137,202 ; mov %r9d,%r10d
+ DB 77,137,203 ; mov %r9,%r11
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 76,139,48 ; mov (%rax),%r14
+ DB 76,139,72,8 ; mov 0x8(%rax),%r9
+ DB 71,15,182,20,22 ; movzbl (%r14,%r10,1),%r10d
+ DB 67,15,182,28,30 ; movzbl (%r14,%r11,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,211 ; or %r10d,%ebx
+ DB 71,15,182,4,6 ; movzbl (%r14,%r8,1),%r8d
+ DB 65,15,182,12,14 ; movzbl (%r14,%rcx,1),%ecx
+ DB 193,225,8 ; shl $0x8,%ecx
+ DB 68,9,193 ; or %r8d,%ecx
+ DB 102,15,196,193,0 ; pinsrw $0x0,%ecx,%xmm0
+ DB 102,15,196,195,1 ; pinsrw $0x1,%ebx,%xmm0
+ DB 102,69,15,239,201 ; pxor %xmm9,%xmm9
+ DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0
+ DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0
+ DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
+ DB 185,129,128,128,59 ; mov $0x3b808081,%ecx
+ DB 102,68,15,110,209 ; movd %ecx,%xmm10
+ DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
+ DB 65,15,89,194 ; mulps %xmm10,%xmm0
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
+ DB 102,72,15,126,201 ; movq %xmm1,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1
+ DB 102,72,15,126,203 ; movq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d
+ DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,211 ; or %r10d,%ebx
+ DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d
+ DB 65,15,182,12,9 ; movzbl (%r9,%rcx,1),%ecx
+ DB 193,225,8 ; shl $0x8,%ecx
+ DB 68,9,193 ; or %r8d,%ecx
+ DB 102,15,196,201,0 ; pinsrw $0x0,%ecx,%xmm1
+ DB 102,15,196,203,1 ; pinsrw $0x1,%ebx,%xmm1
+ DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1
+ DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1
+ DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
+ DB 65,15,89,202 ; mulps %xmm10,%xmm1
+ DB 76,139,72,16 ; mov 0x10(%rax),%r9
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2
+ DB 102,72,15,126,211 ; movq %xmm2,%rbx
+ DB 65,137,216 ; mov %ebx,%r8d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 102,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm2
+ DB 102,72,15,126,209 ; movq %xmm2,%rcx
+ DB 65,137,202 ; mov %ecx,%r10d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d
+ DB 65,15,182,12,9 ; movzbl (%r9,%rcx,1),%ecx
+ DB 193,225,8 ; shl $0x8,%ecx
+ DB 68,9,209 ; or %r10d,%ecx
+ DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d
+ DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,195 ; or %r8d,%ebx
+ DB 102,15,196,211,0 ; pinsrw $0x0,%ebx,%xmm2
+ DB 102,15,196,209,1 ; pinsrw $0x1,%ecx,%xmm2
+ DB 102,65,15,96,209 ; punpcklbw %xmm9,%xmm2
+ DB 102,65,15,97,209 ; punpcklwd %xmm9,%xmm2
+ DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 65,15,89,210 ; mulps %xmm10,%xmm2
+ DB 72,139,64,24 ; mov 0x18(%rax),%rax
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 102,65,15,91,216 ; cvtps2dq %xmm8,%xmm3
+ DB 102,72,15,126,217 ; movq %xmm3,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,15,112,219,78 ; pshufd $0x4e,%xmm3,%xmm3
+ DB 102,72,15,126,219 ; movq %xmm3,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 70,15,182,12,8 ; movzbl (%rax,%r9,1),%r9d
+ DB 15,182,28,24 ; movzbl (%rax,%rbx,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,203 ; or %r9d,%ebx
+ DB 70,15,182,4,0 ; movzbl (%rax,%r8,1),%r8d
+ DB 15,182,4,8 ; movzbl (%rax,%rcx,1),%eax
+ DB 193,224,8 ; shl $0x8,%eax
+ DB 68,9,192 ; or %r8d,%eax
+ DB 102,15,196,216,0 ; pinsrw $0x0,%eax,%xmm3
+ DB 102,15,196,219,1 ; pinsrw $0x1,%ebx,%xmm3
+ DB 102,65,15,96,217 ; punpcklbw %xmm9,%xmm3
+ DB 102,65,15,97,217 ; punpcklwd %xmm9,%xmm3
+ DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
+ DB 65,15,89,218 ; mulps %xmm10,%xmm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 91 ; pop %rbx
+ DB 65,94 ; pop %r14
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_byte_tables_rgb_sse2
+_sk_byte_tables_rgb_sse2 LABEL PROC
+ DB 65,86 ; push %r14
+ DB 83 ; push %rbx
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 139,72,24 ; mov 0x18(%rax),%ecx
+ DB 255,201 ; dec %ecx
+ DB 102,68,15,110,193 ; movd %ecx,%xmm8
+ DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8
+ DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
+ DB 65,15,89,192 ; mulps %xmm8,%xmm0
+ DB 102,15,91,192 ; cvtps2dq %xmm0,%xmm0
+ DB 102,72,15,126,193 ; movq %xmm0,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm0
+ DB 102,73,15,126,193 ; movq %xmm0,%r9
+ DB 69,137,202 ; mov %r9d,%r10d
+ DB 77,137,203 ; mov %r9,%r11
+ DB 73,193,235,32 ; shr $0x20,%r11
+ DB 76,139,48 ; mov (%rax),%r14
+ DB 76,139,72,8 ; mov 0x8(%rax),%r9
+ DB 71,15,182,20,22 ; movzbl (%r14,%r10,1),%r10d
+ DB 67,15,182,28,30 ; movzbl (%r14,%r11,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,211 ; or %r10d,%ebx
+ DB 71,15,182,4,6 ; movzbl (%r14,%r8,1),%r8d
+ DB 65,15,182,12,14 ; movzbl (%r14,%rcx,1),%ecx
+ DB 193,225,8 ; shl $0x8,%ecx
+ DB 68,9,193 ; or %r8d,%ecx
+ DB 102,15,196,193,0 ; pinsrw $0x0,%ecx,%xmm0
+ DB 102,15,196,195,1 ; pinsrw $0x1,%ebx,%xmm0
+ DB 102,69,15,239,201 ; pxor %xmm9,%xmm9
+ DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0
+ DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0
+ DB 15,91,192 ; cvtdq2ps %xmm0,%xmm0
+ DB 185,129,128,128,59 ; mov $0x3b808081,%ecx
+ DB 102,68,15,110,209 ; movd %ecx,%xmm10
+ DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
+ DB 65,15,89,194 ; mulps %xmm10,%xmm0
+ DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 102,15,91,201 ; cvtps2dq %xmm1,%xmm1
+ DB 102,72,15,126,201 ; movq %xmm1,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,15,112,201,78 ; pshufd $0x4e,%xmm1,%xmm1
+ DB 102,72,15,126,203 ; movq %xmm1,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 71,15,182,20,17 ; movzbl (%r9,%r10,1),%r10d
+ DB 65,15,182,28,25 ; movzbl (%r9,%rbx,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,211 ; or %r10d,%ebx
+ DB 71,15,182,4,1 ; movzbl (%r9,%r8,1),%r8d
+ DB 65,15,182,12,9 ; movzbl (%r9,%rcx,1),%ecx
+ DB 193,225,8 ; shl $0x8,%ecx
+ DB 68,9,193 ; or %r8d,%ecx
+ DB 102,15,196,201,0 ; pinsrw $0x0,%ecx,%xmm1
+ DB 102,15,196,203,1 ; pinsrw $0x1,%ebx,%xmm1
+ DB 102,65,15,96,201 ; punpcklbw %xmm9,%xmm1
+ DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1
+ DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
+ DB 65,15,89,202 ; mulps %xmm10,%xmm1
+ DB 72,139,64,16 ; mov 0x10(%rax),%rax
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 102,15,91,210 ; cvtps2dq %xmm2,%xmm2
+ DB 102,72,15,126,209 ; movq %xmm2,%rcx
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 72,193,233,32 ; shr $0x20,%rcx
+ DB 102,15,112,210,78 ; pshufd $0x4e,%xmm2,%xmm2
+ DB 102,72,15,126,211 ; movq %xmm2,%rbx
+ DB 65,137,217 ; mov %ebx,%r9d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 70,15,182,12,8 ; movzbl (%rax,%r9,1),%r9d
+ DB 15,182,28,24 ; movzbl (%rax,%rbx,1),%ebx
+ DB 193,227,8 ; shl $0x8,%ebx
+ DB 68,9,203 ; or %r9d,%ebx
+ DB 70,15,182,4,0 ; movzbl (%rax,%r8,1),%r8d
+ DB 15,182,4,8 ; movzbl (%rax,%rcx,1),%eax
+ DB 193,224,8 ; shl $0x8,%eax
+ DB 68,9,192 ; or %r8d,%eax
+ DB 102,15,196,208,0 ; pinsrw $0x0,%eax,%xmm2
+ DB 102,15,196,211,1 ; pinsrw $0x1,%ebx,%xmm2
+ DB 102,65,15,96,209 ; punpcklbw %xmm9,%xmm2
+ DB 102,65,15,97,209 ; punpcklwd %xmm9,%xmm2
+ DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 65,15,89,210 ; mulps %xmm10,%xmm2
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 91 ; pop %rbx
+ DB 65,94 ; pop %r14
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_a8_sse2
_sk_load_a8_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -10687,9 +11634,9 @@ _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 116,5 ; je 1a49 <_sk_gather_i8_sse2+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 172b <_sk_gather_i8_sse2+0x11>
+ DB 235,2 ; jmp 1a4b <_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