aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/jumper/SkJumper_generated_win.S
diff options
context:
space:
mode:
authorGravatar Mike Klein <mtklein@chromium.org>2017-04-27 13:36:57 -0400
committerGravatar Skia Commit-Bot <skia-commit-bot@chromium.org>2017-04-27 18:23:58 +0000
commit0aa742f15a400433bafe6350c20523b4dd062f64 (patch)
treed3650d705e32138a60c453e42c54f959a57423de /src/jumper/SkJumper_generated_win.S
parentfdd77daedbba3b7c53be74a82fb9fae891b51696 (diff)
jumper, replace _i with normal constants
So far I only seem to be encountering constant pools with float constants, so integer constants should be easy to make normal. This just removes _i. There might be a couple integer constants generated with C() too... they'll be the next CL. CQ_INCLUDE_TRYBOTS=skia.primary:Test-Win7-MSVC-Golo-CPU-AVX-x86_64-Release,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Release,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Release-SK_CPU_LIMIT_SSE41,Test-Ubuntu-Clang-GCE-CPU-AVX2-x86_64-Release-SK_CPU_LIMIT_SSE2 Change-Id: Icc82cbc660d1e33bcdb5282072fb86cb5190d901 Reviewed-on: https://skia-review.googlesource.com/14527 Reviewed-by: Herb Derby <herb@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.S2888
1 files changed, 1656 insertions, 1232 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index 2258440ca5..358346bd1f 100644
--- a/src/jumper/SkJumper_generated_win.S
+++ b/src/jumper/SkJumper_generated_win.S
@@ -106,14 +106,14 @@ _sk_seed_shader_hsw LABEL PROC
DB 197,249,110,199 ; vmovd %edi,%xmm0
DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,74,64,0,0 ; vbroadcastss 0x404a(%rip),%ymm1 # 41a4 <_sk_callback_hsw+0x11a>
+ DB 196,226,125,24,13,10,64,0,0 ; vbroadcastss 0x400a(%rip),%ymm1 # 4164 <_sk_callback_hsw+0x11a>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,88,2 ; vaddps (%rdx),%ymm0,%ymm0
DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,21,46,64,0,0 ; vbroadcastss 0x402e(%rip),%ymm2 # 41a8 <_sk_callback_hsw+0x11e>
+ DB 196,226,125,24,21,238,63,0,0 ; vbroadcastss 0x3fee(%rip),%ymm2 # 4168 <_sk_callback_hsw+0x11e>
DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3
DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4
DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5
@@ -1299,24 +1299,19 @@ _sk_load_tables_hsw LABEL PROC
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 76,3,8 ; add (%rax),%r9
DB 77,133,192 ; test %r8,%r8
- DB 117,121 ; jne 13b6 <_sk_load_tables_hsw+0x8e>
+ DB 117,109 ; jne 13aa <_sk_load_tables_hsw+0x82>
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 197,249,110,193 ; vmovd %ecx,%xmm0
- DB 196,226,125,88,208 ; vpbroadcastd %xmm0,%ymm2
- DB 197,237,219,203 ; vpand %ymm3,%ymm2,%ymm1
+ DB 197,229,219,13,246,46,0,0 ; vpand 0x2ef6(%rip),%ymm3,%ymm1 # 4240 <_sk_callback_hsw+0x1f6>
DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8
DB 72,139,72,8 ; mov 0x8(%rax),%rcx
DB 76,139,72,16 ; mov 0x10(%rax),%r9
+ DB 197,237,118,210 ; vpcmpeqd %ymm2,%ymm2,%ymm2
+ DB 196,226,109,146,4,137 ; vgatherdps %ymm2,(%rcx,%ymm1,4),%ymm0
+ DB 196,226,101,0,21,246,46,0,0 ; vpshufb 0x2ef6(%rip),%ymm3,%ymm2 # 4260 <_sk_callback_hsw+0x216>
DB 196,65,53,118,201 ; vpcmpeqd %ymm9,%ymm9,%ymm9
- DB 196,226,53,146,4,137 ; vgatherdps %ymm9,(%rcx,%ymm1,4),%ymm0
- DB 197,245,114,211,8 ; vpsrld $0x8,%ymm3,%ymm1
- DB 197,109,219,201 ; vpand %ymm1,%ymm2,%ymm9
- DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10
- DB 196,130,45,146,12,137 ; vgatherdps %ymm10,(%r9,%ymm9,4),%ymm1
+ DB 196,194,53,146,12,145 ; vgatherdps %ymm9,(%r9,%ymm2,4),%ymm1
DB 72,139,64,24 ; mov 0x18(%rax),%rax
- DB 197,181,114,211,16 ; vpsrld $0x10,%ymm3,%ymm9
- DB 196,65,109,219,201 ; vpand %ymm9,%ymm2,%ymm9
+ DB 196,98,101,0,13,254,46,0,0 ; vpshufb 0x2efe(%rip),%ymm3,%ymm9 # 4280 <_sk_callback_hsw+0x236>
DB 196,162,61,146,20,136 ; vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2
DB 197,229,114,211,24 ; vpsrld $0x18,%ymm3,%ymm3
DB 197,124,91,195 ; vcvtdq2ps %ymm3,%ymm8
@@ -1335,7 +1330,7 @@ _sk_load_tables_hsw LABEL PROC
DB 196,193,249,110,194 ; vmovq %r10,%xmm0
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
- DB 233,99,255,255,255 ; jmpq 1342 <_sk_load_tables_hsw+0x1a>
+ DB 233,111,255,255,255 ; jmpq 1342 <_sk_load_tables_hsw+0x1a>
PUBLIC _sk_load_tables_u16_be_hsw
_sk_load_tables_u16_be_hsw LABEL PROC
@@ -1343,7 +1338,7 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,216,0,0,0 ; jne 14cd <_sk_load_tables_u16_be_hsw+0xee>
+ DB 15,133,208,0,0,0 ; jne 14b9 <_sk_load_tables_u16_be_hsw+0xe6>
DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8
DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2
DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3
@@ -1354,28 +1349,26 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 196,193,97,105,217 ; vpunpckhwd %xmm9,%xmm3,%xmm3
DB 197,121,97,194 ; vpunpcklwd %xmm2,%xmm0,%xmm8
DB 197,121,105,202 ; vpunpckhwd %xmm2,%xmm0,%xmm9
- DB 197,241,97,211 ; vpunpcklwd %xmm3,%xmm1,%xmm2
+ DB 197,241,97,195 ; vpunpcklwd %xmm3,%xmm1,%xmm0
DB 197,113,105,235 ; vpunpckhwd %xmm3,%xmm1,%xmm13
- DB 197,185,108,194 ; vpunpcklqdq %xmm2,%xmm8,%xmm0
- DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0
- DB 65,184,255,0,0,0 ; mov $0xff,%r8d
- DB 196,193,121,110,200 ; vmovd %r8d,%xmm1
- DB 196,98,125,88,209 ; vpbroadcastd %xmm1,%ymm10
- DB 196,193,125,219,202 ; vpand %ymm10,%ymm0,%ymm1
+ DB 197,185,108,200 ; vpunpcklqdq %xmm0,%xmm8,%xmm1
+ DB 197,185,109,208 ; vpunpckhqdq %xmm0,%xmm8,%xmm2
+ DB 196,65,49,108,197 ; vpunpcklqdq %xmm13,%xmm9,%xmm8
+ DB 197,121,111,21,133,47,0,0 ; vmovdqa 0x2f85(%rip),%xmm10 # 43c0 <_sk_callback_hsw+0x376>
+ DB 196,193,113,219,194 ; vpand %xmm10,%xmm1,%xmm0
+ DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1
DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11
DB 76,139,64,8 ; mov 0x8(%rax),%r8
DB 76,139,72,16 ; mov 0x10(%rax),%r9
DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12
DB 196,194,29,146,4,136 ; vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0
- DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1
- DB 196,226,125,51,201 ; vpmovzxwd %xmm1,%ymm1
- DB 196,193,117,219,210 ; vpand %ymm10,%ymm1,%ymm2
- DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8
- DB 196,194,61,146,12,145 ; vgatherdps %ymm8,(%r9,%ymm2,4),%ymm1
+ DB 196,193,105,219,202 ; vpand %xmm10,%xmm2,%xmm1
+ DB 196,226,125,51,209 ; vpmovzxwd %xmm1,%ymm2
+ DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12
+ DB 196,194,29,146,12,145 ; vgatherdps %ymm12,(%r9,%ymm2,4),%ymm1
DB 72,139,64,24 ; mov 0x18(%rax),%rax
- DB 196,193,49,108,213 ; vpunpcklqdq %xmm13,%xmm9,%xmm2
- DB 196,226,125,51,210 ; vpmovzxwd %xmm2,%ymm2
- DB 196,65,109,219,194 ; vpand %ymm10,%ymm2,%ymm8
+ DB 196,193,57,219,210 ; vpand %xmm10,%xmm8,%xmm2
+ DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8
DB 196,162,37,146,20,128 ; vgatherdps %ymm11,(%rax,%ymm8,4),%ymm2
DB 184,128,0,128,55 ; mov $0x37800080,%eax
DB 197,249,110,216 ; vmovd %eax,%xmm3
@@ -1392,29 +1385,29 @@ _sk_load_tables_u16_be_hsw LABEL PROC
DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 1533 <_sk_load_tables_u16_be_hsw+0x154>
+ DB 116,85 ; je 151f <_sk_load_tables_u16_be_hsw+0x14c>
DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 1533 <_sk_load_tables_u16_be_hsw+0x154>
+ DB 114,72 ; jb 151f <_sk_load_tables_u16_be_hsw+0x14c>
DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 1540 <_sk_load_tables_u16_be_hsw+0x161>
+ DB 116,72 ; je 152c <_sk_load_tables_u16_be_hsw+0x159>
DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 1540 <_sk_load_tables_u16_be_hsw+0x161>
+ DB 114,59 ; jb 152c <_sk_load_tables_u16_be_hsw+0x159>
DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,250,254,255,255 ; je 1410 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,132,2,255,255,255 ; je 1404 <_sk_load_tables_u16_be_hsw+0x31>
DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,233,254,255,255 ; jb 1410 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 15,130,241,254,255,255 ; jb 1404 <_sk_load_tables_u16_be_hsw+0x31>
DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9
- DB 233,221,254,255,255 ; jmpq 1410 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,229,254,255,255 ; jmpq 1404 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,208,254,255,255 ; jmpq 1410 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,216,254,255,255 ; jmpq 1404 <_sk_load_tables_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,199,254,255,255 ; jmpq 1410 <_sk_load_tables_u16_be_hsw+0x31>
+ DB 233,207,254,255,255 ; jmpq 1404 <_sk_load_tables_u16_be_hsw+0x31>
PUBLIC _sk_load_tables_rgb_u16_be_hsw
_sk_load_tables_rgb_u16_be_hsw LABEL PROC
@@ -1422,7 +1415,7 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,207,0,0,0 ; jne 162a <_sk_load_tables_rgb_u16_be_hsw+0xe1>
+ DB 15,133,198,0,0,0 ; jne 160d <_sk_load_tables_rgb_u16_be_hsw+0xd8>
DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0
DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2
DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1
@@ -1437,30 +1430,28 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 197,241,97,203 ; vpunpcklwd %xmm3,%xmm1,%xmm1
DB 196,193,33,97,218 ; vpunpcklwd %xmm10,%xmm11,%xmm3
DB 197,121,97,194 ; vpunpcklwd %xmm2,%xmm0,%xmm8
- DB 197,121,105,202 ; vpunpckhwd %xmm2,%xmm0,%xmm9
+ DB 197,249,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm0
DB 197,241,97,211 ; vpunpcklwd %xmm3,%xmm1,%xmm2
- DB 197,241,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm3
- DB 197,185,108,194 ; vpunpcklqdq %xmm2,%xmm8,%xmm0
- DB 196,226,125,51,192 ; vpmovzxwd %xmm0,%ymm0
- DB 65,184,255,0,0,0 ; mov $0xff,%r8d
- DB 196,193,121,110,200 ; vmovd %r8d,%xmm1
- DB 196,98,125,88,209 ; vpbroadcastd %xmm1,%ymm10
- DB 196,193,125,219,202 ; vpand %ymm10,%ymm0,%ymm1
- DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11
+ DB 197,241,105,203 ; vpunpckhwd %xmm3,%xmm1,%xmm1
+ DB 197,185,108,218 ; vpunpcklqdq %xmm2,%xmm8,%xmm3
+ DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2
+ DB 197,121,108,193 ; vpunpcklqdq %xmm1,%xmm0,%xmm8
+ DB 197,121,111,13,31,46,0,0 ; vmovdqa 0x2e1f(%rip),%xmm9 # 43d0 <_sk_callback_hsw+0x386>
+ DB 196,193,97,219,193 ; vpand %xmm9,%xmm3,%xmm0
+ DB 196,226,125,51,200 ; vpmovzxwd %xmm0,%ymm1
+ DB 197,229,118,219 ; vpcmpeqd %ymm3,%ymm3,%ymm3
DB 76,139,64,8 ; mov 0x8(%rax),%r8
DB 76,139,72,16 ; mov 0x10(%rax),%r9
- DB 196,65,29,118,228 ; vpcmpeqd %ymm12,%ymm12,%ymm12
- DB 196,194,29,146,4,136 ; vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0
- DB 197,185,109,202 ; vpunpckhqdq %xmm2,%xmm8,%xmm1
- DB 196,226,125,51,201 ; vpmovzxwd %xmm1,%ymm1
- DB 196,193,117,219,210 ; vpand %ymm10,%ymm1,%ymm2
- DB 196,65,61,118,192 ; vpcmpeqd %ymm8,%ymm8,%ymm8
- DB 196,194,61,146,12,145 ; vgatherdps %ymm8,(%r9,%ymm2,4),%ymm1
+ DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10
+ DB 196,194,45,146,4,136 ; vgatherdps %ymm10,(%r8,%ymm1,4),%ymm0
+ DB 196,193,105,219,201 ; vpand %xmm9,%xmm2,%xmm1
+ DB 196,226,125,51,209 ; vpmovzxwd %xmm1,%ymm2
+ DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10
+ DB 196,194,45,146,12,145 ; vgatherdps %ymm10,(%r9,%ymm2,4),%ymm1
DB 72,139,64,24 ; mov 0x18(%rax),%rax
- DB 197,177,108,211 ; vpunpcklqdq %xmm3,%xmm9,%xmm2
- DB 196,226,125,51,210 ; vpmovzxwd %xmm2,%ymm2
- DB 196,193,109,219,218 ; vpand %ymm10,%ymm2,%ymm3
- DB 196,226,37,146,20,152 ; vgatherdps %ymm11,(%rax,%ymm3,4),%ymm2
+ DB 196,193,57,219,209 ; vpand %xmm9,%xmm8,%xmm2
+ DB 196,98,125,51,194 ; vpmovzxwd %xmm2,%ymm8
+ DB 196,162,101,146,20,128 ; vgatherdps %ymm3,(%rax,%ymm8,4),%ymm2
DB 184,0,0,128,63 ; mov $0x3f800000,%eax
DB 197,249,110,216 ; vmovd %eax,%xmm3
DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3
@@ -1469,36 +1460,36 @@ _sk_load_tables_rgb_u16_be_hsw LABEL PROC
DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0
DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 1643 <_sk_load_tables_rgb_u16_be_hsw+0xfa>
- DB 233,76,255,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 1626 <_sk_load_tables_rgb_u16_be_hsw+0xf1>
+ DB 233,85,255,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1
DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 1672 <_sk_load_tables_rgb_u16_be_hsw+0x129>
+ DB 114,26 ; jb 1655 <_sk_load_tables_rgb_u16_be_hsw+0x120>
DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1
DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 1677 <_sk_load_tables_rgb_u16_be_hsw+0x12e>
- DB 233,29,255,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,24,255,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 165a <_sk_load_tables_rgb_u16_be_hsw+0x125>
+ DB 233,38,255,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,33,255,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1
DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 16a6 <_sk_load_tables_rgb_u16_be_hsw+0x15d>
+ DB 114,26 ; jb 1689 <_sk_load_tables_rgb_u16_be_hsw+0x154>
DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1
DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 16ab <_sk_load_tables_rgb_u16_be_hsw+0x162>
- DB 233,233,254,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,228,254,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 168e <_sk_load_tables_rgb_u16_be_hsw+0x159>
+ DB 233,242,254,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,237,254,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3
DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 16d4 <_sk_load_tables_rgb_u16_be_hsw+0x18b>
+ DB 114,20 ; jb 16b7 <_sk_load_tables_rgb_u16_be_hsw+0x182>
DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3
DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
- DB 233,187,254,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
- DB 233,182,254,255,255 ; jmpq 158f <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,196,254,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
+ DB 233,191,254,255,255 ; jmpq 157b <_sk_load_tables_rgb_u16_be_hsw+0x46>
PUBLIC _sk_byte_tables_hsw
_sk_byte_tables_hsw LABEL PROC
@@ -1864,33 +1855,33 @@ _sk_parametric_r_hsw LABEL PROC
DB 196,66,125,168,211 ; vfmadd213ps %ymm11,%ymm0,%ymm10
DB 196,226,125,24,0 ; vbroadcastss (%rax),%ymm0
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,231,36,0,0 ; vbroadcastss 0x24e7(%rip),%ymm12 # 41ac <_sk_callback_hsw+0x122>
- DB 196,98,125,24,45,226,36,0,0 ; vbroadcastss 0x24e2(%rip),%ymm13 # 41b0 <_sk_callback_hsw+0x126>
+ DB 196,98,125,24,37,196,36,0,0 ; vbroadcastss 0x24c4(%rip),%ymm12 # 416c <_sk_callback_hsw+0x122>
+ DB 196,98,125,24,45,191,36,0,0 ; vbroadcastss 0x24bf(%rip),%ymm13 # 4170 <_sk_callback_hsw+0x126>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,216,36,0,0 ; vbroadcastss 0x24d8(%rip),%ymm13 # 41b4 <_sk_callback_hsw+0x12a>
+ DB 196,98,125,24,45,181,36,0,0 ; vbroadcastss 0x24b5(%rip),%ymm13 # 4174 <_sk_callback_hsw+0x12a>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,206,36,0,0 ; vbroadcastss 0x24ce(%rip),%ymm13 # 41b8 <_sk_callback_hsw+0x12e>
+ DB 196,98,125,24,45,171,36,0,0 ; vbroadcastss 0x24ab(%rip),%ymm13 # 4178 <_sk_callback_hsw+0x12e>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,196,36,0,0 ; vbroadcastss 0x24c4(%rip),%ymm11 # 41bc <_sk_callback_hsw+0x132>
+ DB 196,98,125,24,29,161,36,0,0 ; vbroadcastss 0x24a1(%rip),%ymm11 # 417c <_sk_callback_hsw+0x132>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,186,36,0,0 ; vbroadcastss 0x24ba(%rip),%ymm12 # 41c0 <_sk_callback_hsw+0x136>
+ DB 196,98,125,24,37,151,36,0,0 ; vbroadcastss 0x2497(%rip),%ymm12 # 4180 <_sk_callback_hsw+0x136>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,176,36,0,0 ; vbroadcastss 0x24b0(%rip),%ymm12 # 41c4 <_sk_callback_hsw+0x13a>
+ DB 196,98,125,24,37,141,36,0,0 ; vbroadcastss 0x248d(%rip),%ymm12 # 4184 <_sk_callback_hsw+0x13a>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0
DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10
DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10
- DB 196,98,125,24,29,145,36,0,0 ; vbroadcastss 0x2491(%rip),%ymm11 # 41c8 <_sk_callback_hsw+0x13e>
+ DB 196,98,125,24,29,110,36,0,0 ; vbroadcastss 0x246e(%rip),%ymm11 # 4188 <_sk_callback_hsw+0x13e>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,135,36,0,0 ; vbroadcastss 0x2487(%rip),%ymm11 # 41cc <_sk_callback_hsw+0x142>
+ DB 196,98,125,24,29,100,36,0,0 ; vbroadcastss 0x2464(%rip),%ymm11 # 418c <_sk_callback_hsw+0x142>
DB 196,98,45,172,216 ; vfnmadd213ps %ymm0,%ymm10,%ymm11
- DB 196,226,125,24,5,125,36,0,0 ; vbroadcastss 0x247d(%rip),%ymm0 # 41d0 <_sk_callback_hsw+0x146>
+ DB 196,226,125,24,5,90,36,0,0 ; vbroadcastss 0x245a(%rip),%ymm0 # 4190 <_sk_callback_hsw+0x146>
DB 196,193,124,92,194 ; vsubps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,115,36,0,0 ; vbroadcastss 0x2473(%rip),%ymm10 # 41d4 <_sk_callback_hsw+0x14a>
+ DB 196,98,125,24,21,80,36,0,0 ; vbroadcastss 0x2450(%rip),%ymm10 # 4194 <_sk_callback_hsw+0x14a>
DB 197,172,94,192 ; vdivps %ymm0,%ymm10,%ymm0
DB 197,164,88,192 ; vaddps %ymm0,%ymm11,%ymm0
- DB 196,98,125,24,21,102,36,0,0 ; vbroadcastss 0x2466(%rip),%ymm10 # 41d8 <_sk_callback_hsw+0x14e>
+ DB 196,98,125,24,21,67,36,0,0 ; vbroadcastss 0x2443(%rip),%ymm10 # 4198 <_sk_callback_hsw+0x14e>
DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0
DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -1918,33 +1909,33 @@ _sk_parametric_g_hsw LABEL PROC
DB 196,66,117,168,211 ; vfmadd213ps %ymm11,%ymm1,%ymm10
DB 196,226,125,24,8 ; vbroadcastss (%rax),%ymm1
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,237,35,0,0 ; vbroadcastss 0x23ed(%rip),%ymm12 # 41dc <_sk_callback_hsw+0x152>
- DB 196,98,125,24,45,232,35,0,0 ; vbroadcastss 0x23e8(%rip),%ymm13 # 41e0 <_sk_callback_hsw+0x156>
+ DB 196,98,125,24,37,202,35,0,0 ; vbroadcastss 0x23ca(%rip),%ymm12 # 419c <_sk_callback_hsw+0x152>
+ DB 196,98,125,24,45,197,35,0,0 ; vbroadcastss 0x23c5(%rip),%ymm13 # 41a0 <_sk_callback_hsw+0x156>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,222,35,0,0 ; vbroadcastss 0x23de(%rip),%ymm13 # 41e4 <_sk_callback_hsw+0x15a>
+ DB 196,98,125,24,45,187,35,0,0 ; vbroadcastss 0x23bb(%rip),%ymm13 # 41a4 <_sk_callback_hsw+0x15a>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,212,35,0,0 ; vbroadcastss 0x23d4(%rip),%ymm13 # 41e8 <_sk_callback_hsw+0x15e>
+ DB 196,98,125,24,45,177,35,0,0 ; vbroadcastss 0x23b1(%rip),%ymm13 # 41a8 <_sk_callback_hsw+0x15e>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,202,35,0,0 ; vbroadcastss 0x23ca(%rip),%ymm11 # 41ec <_sk_callback_hsw+0x162>
+ DB 196,98,125,24,29,167,35,0,0 ; vbroadcastss 0x23a7(%rip),%ymm11 # 41ac <_sk_callback_hsw+0x162>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,192,35,0,0 ; vbroadcastss 0x23c0(%rip),%ymm12 # 41f0 <_sk_callback_hsw+0x166>
+ DB 196,98,125,24,37,157,35,0,0 ; vbroadcastss 0x239d(%rip),%ymm12 # 41b0 <_sk_callback_hsw+0x166>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,182,35,0,0 ; vbroadcastss 0x23b6(%rip),%ymm12 # 41f4 <_sk_callback_hsw+0x16a>
+ DB 196,98,125,24,37,147,35,0,0 ; vbroadcastss 0x2393(%rip),%ymm12 # 41b4 <_sk_callback_hsw+0x16a>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1
DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10
DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
- DB 196,98,125,24,29,151,35,0,0 ; vbroadcastss 0x2397(%rip),%ymm11 # 41f8 <_sk_callback_hsw+0x16e>
+ DB 196,98,125,24,29,116,35,0,0 ; vbroadcastss 0x2374(%rip),%ymm11 # 41b8 <_sk_callback_hsw+0x16e>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,141,35,0,0 ; vbroadcastss 0x238d(%rip),%ymm11 # 41fc <_sk_callback_hsw+0x172>
+ DB 196,98,125,24,29,106,35,0,0 ; vbroadcastss 0x236a(%rip),%ymm11 # 41bc <_sk_callback_hsw+0x172>
DB 196,98,45,172,217 ; vfnmadd213ps %ymm1,%ymm10,%ymm11
- DB 196,226,125,24,13,131,35,0,0 ; vbroadcastss 0x2383(%rip),%ymm1 # 4200 <_sk_callback_hsw+0x176>
+ DB 196,226,125,24,13,96,35,0,0 ; vbroadcastss 0x2360(%rip),%ymm1 # 41c0 <_sk_callback_hsw+0x176>
DB 196,193,116,92,202 ; vsubps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,121,35,0,0 ; vbroadcastss 0x2379(%rip),%ymm10 # 4204 <_sk_callback_hsw+0x17a>
+ DB 196,98,125,24,21,86,35,0,0 ; vbroadcastss 0x2356(%rip),%ymm10 # 41c4 <_sk_callback_hsw+0x17a>
DB 197,172,94,201 ; vdivps %ymm1,%ymm10,%ymm1
DB 197,164,88,201 ; vaddps %ymm1,%ymm11,%ymm1
- DB 196,98,125,24,21,108,35,0,0 ; vbroadcastss 0x236c(%rip),%ymm10 # 4208 <_sk_callback_hsw+0x17e>
+ DB 196,98,125,24,21,73,35,0,0 ; vbroadcastss 0x2349(%rip),%ymm10 # 41c8 <_sk_callback_hsw+0x17e>
DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1
DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -1972,33 +1963,33 @@ _sk_parametric_b_hsw LABEL PROC
DB 196,66,109,168,211 ; vfmadd213ps %ymm11,%ymm2,%ymm10
DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,243,34,0,0 ; vbroadcastss 0x22f3(%rip),%ymm12 # 420c <_sk_callback_hsw+0x182>
- DB 196,98,125,24,45,238,34,0,0 ; vbroadcastss 0x22ee(%rip),%ymm13 # 4210 <_sk_callback_hsw+0x186>
+ DB 196,98,125,24,37,208,34,0,0 ; vbroadcastss 0x22d0(%rip),%ymm12 # 41cc <_sk_callback_hsw+0x182>
+ DB 196,98,125,24,45,203,34,0,0 ; vbroadcastss 0x22cb(%rip),%ymm13 # 41d0 <_sk_callback_hsw+0x186>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,228,34,0,0 ; vbroadcastss 0x22e4(%rip),%ymm13 # 4214 <_sk_callback_hsw+0x18a>
+ DB 196,98,125,24,45,193,34,0,0 ; vbroadcastss 0x22c1(%rip),%ymm13 # 41d4 <_sk_callback_hsw+0x18a>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,218,34,0,0 ; vbroadcastss 0x22da(%rip),%ymm13 # 4218 <_sk_callback_hsw+0x18e>
+ DB 196,98,125,24,45,183,34,0,0 ; vbroadcastss 0x22b7(%rip),%ymm13 # 41d8 <_sk_callback_hsw+0x18e>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,208,34,0,0 ; vbroadcastss 0x22d0(%rip),%ymm11 # 421c <_sk_callback_hsw+0x192>
+ DB 196,98,125,24,29,173,34,0,0 ; vbroadcastss 0x22ad(%rip),%ymm11 # 41dc <_sk_callback_hsw+0x192>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,198,34,0,0 ; vbroadcastss 0x22c6(%rip),%ymm12 # 4220 <_sk_callback_hsw+0x196>
+ DB 196,98,125,24,37,163,34,0,0 ; vbroadcastss 0x22a3(%rip),%ymm12 # 41e0 <_sk_callback_hsw+0x196>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,188,34,0,0 ; vbroadcastss 0x22bc(%rip),%ymm12 # 4224 <_sk_callback_hsw+0x19a>
+ DB 196,98,125,24,37,153,34,0,0 ; vbroadcastss 0x2299(%rip),%ymm12 # 41e4 <_sk_callback_hsw+0x19a>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2
DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10
DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10
- DB 196,98,125,24,29,157,34,0,0 ; vbroadcastss 0x229d(%rip),%ymm11 # 4228 <_sk_callback_hsw+0x19e>
+ DB 196,98,125,24,29,122,34,0,0 ; vbroadcastss 0x227a(%rip),%ymm11 # 41e8 <_sk_callback_hsw+0x19e>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,147,34,0,0 ; vbroadcastss 0x2293(%rip),%ymm11 # 422c <_sk_callback_hsw+0x1a2>
+ DB 196,98,125,24,29,112,34,0,0 ; vbroadcastss 0x2270(%rip),%ymm11 # 41ec <_sk_callback_hsw+0x1a2>
DB 196,98,45,172,218 ; vfnmadd213ps %ymm2,%ymm10,%ymm11
- DB 196,226,125,24,21,137,34,0,0 ; vbroadcastss 0x2289(%rip),%ymm2 # 4230 <_sk_callback_hsw+0x1a6>
+ DB 196,226,125,24,21,102,34,0,0 ; vbroadcastss 0x2266(%rip),%ymm2 # 41f0 <_sk_callback_hsw+0x1a6>
DB 196,193,108,92,210 ; vsubps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,127,34,0,0 ; vbroadcastss 0x227f(%rip),%ymm10 # 4234 <_sk_callback_hsw+0x1aa>
+ DB 196,98,125,24,21,92,34,0,0 ; vbroadcastss 0x225c(%rip),%ymm10 # 41f4 <_sk_callback_hsw+0x1aa>
DB 197,172,94,210 ; vdivps %ymm2,%ymm10,%ymm2
DB 197,164,88,210 ; vaddps %ymm2,%ymm11,%ymm2
- DB 196,98,125,24,21,114,34,0,0 ; vbroadcastss 0x2272(%rip),%ymm10 # 4238 <_sk_callback_hsw+0x1ae>
+ DB 196,98,125,24,21,79,34,0,0 ; vbroadcastss 0x224f(%rip),%ymm10 # 41f8 <_sk_callback_hsw+0x1ae>
DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2
DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -2026,33 +2017,33 @@ _sk_parametric_a_hsw LABEL PROC
DB 196,66,101,168,211 ; vfmadd213ps %ymm11,%ymm3,%ymm10
DB 196,226,125,24,24 ; vbroadcastss (%rax),%ymm3
DB 196,65,124,91,218 ; vcvtdq2ps %ymm10,%ymm11
- DB 196,98,125,24,37,249,33,0,0 ; vbroadcastss 0x21f9(%rip),%ymm12 # 423c <_sk_callback_hsw+0x1b2>
- DB 196,98,125,24,45,244,33,0,0 ; vbroadcastss 0x21f4(%rip),%ymm13 # 4240 <_sk_callback_hsw+0x1b6>
+ DB 196,98,125,24,37,214,33,0,0 ; vbroadcastss 0x21d6(%rip),%ymm12 # 41fc <_sk_callback_hsw+0x1b2>
+ DB 196,98,125,24,45,209,33,0,0 ; vbroadcastss 0x21d1(%rip),%ymm13 # 4200 <_sk_callback_hsw+0x1b6>
DB 196,65,44,84,213 ; vandps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,234,33,0,0 ; vbroadcastss 0x21ea(%rip),%ymm13 # 4244 <_sk_callback_hsw+0x1ba>
+ DB 196,98,125,24,45,199,33,0,0 ; vbroadcastss 0x21c7(%rip),%ymm13 # 4204 <_sk_callback_hsw+0x1ba>
DB 196,65,44,86,213 ; vorps %ymm13,%ymm10,%ymm10
- DB 196,98,125,24,45,224,33,0,0 ; vbroadcastss 0x21e0(%rip),%ymm13 # 4248 <_sk_callback_hsw+0x1be>
+ DB 196,98,125,24,45,189,33,0,0 ; vbroadcastss 0x21bd(%rip),%ymm13 # 4208 <_sk_callback_hsw+0x1be>
DB 196,66,37,184,236 ; vfmadd231ps %ymm12,%ymm11,%ymm13
- DB 196,98,125,24,29,214,33,0,0 ; vbroadcastss 0x21d6(%rip),%ymm11 # 424c <_sk_callback_hsw+0x1c2>
+ DB 196,98,125,24,29,179,33,0,0 ; vbroadcastss 0x21b3(%rip),%ymm11 # 420c <_sk_callback_hsw+0x1c2>
DB 196,66,45,172,221 ; vfnmadd213ps %ymm13,%ymm10,%ymm11
- DB 196,98,125,24,37,204,33,0,0 ; vbroadcastss 0x21cc(%rip),%ymm12 # 4250 <_sk_callback_hsw+0x1c6>
+ DB 196,98,125,24,37,169,33,0,0 ; vbroadcastss 0x21a9(%rip),%ymm12 # 4210 <_sk_callback_hsw+0x1c6>
DB 196,65,44,88,212 ; vaddps %ymm12,%ymm10,%ymm10
- DB 196,98,125,24,37,194,33,0,0 ; vbroadcastss 0x21c2(%rip),%ymm12 # 4254 <_sk_callback_hsw+0x1ca>
+ DB 196,98,125,24,37,159,33,0,0 ; vbroadcastss 0x219f(%rip),%ymm12 # 4214 <_sk_callback_hsw+0x1ca>
DB 196,65,28,94,210 ; vdivps %ymm10,%ymm12,%ymm10
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10
DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10
- DB 196,98,125,24,29,163,33,0,0 ; vbroadcastss 0x21a3(%rip),%ymm11 # 4258 <_sk_callback_hsw+0x1ce>
+ DB 196,98,125,24,29,128,33,0,0 ; vbroadcastss 0x2180(%rip),%ymm11 # 4218 <_sk_callback_hsw+0x1ce>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,153,33,0,0 ; vbroadcastss 0x2199(%rip),%ymm11 # 425c <_sk_callback_hsw+0x1d2>
+ DB 196,98,125,24,29,118,33,0,0 ; vbroadcastss 0x2176(%rip),%ymm11 # 421c <_sk_callback_hsw+0x1d2>
DB 196,98,45,172,219 ; vfnmadd213ps %ymm3,%ymm10,%ymm11
- DB 196,226,125,24,29,143,33,0,0 ; vbroadcastss 0x218f(%rip),%ymm3 # 4260 <_sk_callback_hsw+0x1d6>
+ DB 196,226,125,24,29,108,33,0,0 ; vbroadcastss 0x216c(%rip),%ymm3 # 4220 <_sk_callback_hsw+0x1d6>
DB 196,193,100,92,218 ; vsubps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,133,33,0,0 ; vbroadcastss 0x2185(%rip),%ymm10 # 4264 <_sk_callback_hsw+0x1da>
+ DB 196,98,125,24,21,98,33,0,0 ; vbroadcastss 0x2162(%rip),%ymm10 # 4224 <_sk_callback_hsw+0x1da>
DB 197,172,94,219 ; vdivps %ymm3,%ymm10,%ymm3
DB 197,164,88,219 ; vaddps %ymm3,%ymm11,%ymm3
- DB 196,98,125,24,21,120,33,0,0 ; vbroadcastss 0x2178(%rip),%ymm10 # 4268 <_sk_callback_hsw+0x1de>
+ DB 196,98,125,24,21,85,33,0,0 ; vbroadcastss 0x2155(%rip),%ymm10 # 4228 <_sk_callback_hsw+0x1de>
DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -2141,7 +2132,7 @@ _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 2299 <_sk_load_a8_hsw+0x42>
+ DB 117,50 ; jne 227c <_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
@@ -2164,9 +2155,9 @@ _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 22a1 <_sk_load_a8_hsw+0x4a>
+ DB 117,234 ; jne 2284 <_sk_load_a8_hsw+0x4a>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,173 ; jmp 226b <_sk_load_a8_hsw+0x14>
+ DB 235,173 ; jmp 224e <_sk_load_a8_hsw+0x14>
PUBLIC _sk_gather_a8_hsw
_sk_gather_a8_hsw LABEL PROC
@@ -2237,7 +2228,7 @@ _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 23d6 <_sk_store_a8_hsw+0x3b>
+ DB 117,10 ; jne 23b9 <_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
@@ -2245,10 +2236,10 @@ _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 23d2 <_sk_store_a8_hsw+0x37>
+ DB 119,236 ; ja 23b5 <_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 # 2438 <_sk_store_a8_hsw+0x9d>
+ DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 241c <_sk_store_a8_hsw+0x9e>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -2259,26 +2250,27 @@ _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 23d2 <_sk_store_a8_hsw+0x37>
- DB 247,255 ; idiv %edi
+ DB 235,154 ; jmp 23b5 <_sk_store_a8_hsw+0x37>
+ DB 144 ; nop
+ DB 246,255 ; idiv %bh
DB 255 ; (bad)
DB 255 ; (bad)
- DB 239 ; out %eax,(%dx)
+ DB 238 ; out %al,(%dx)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,231 ; jmpq *%rdi
+ DB 255,230 ; jmpq *%rsi
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 223,255 ; (bad)
+ DB 222,255 ; fdivrp %st,%st(7)
DB 255 ; (bad)
- DB 255,215 ; callq *%rdi
+ DB 255,214 ; callq *%rsi
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,207 ; dec %edi
+ DB 255,206 ; dec %esi
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,199 ; inc %edi
+ DB 255,198 ; inc %esi
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -2290,7 +2282,7 @@ _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 24a0 <_sk_load_g8_hsw+0x4c>
+ DB 117,60 ; jne 2484 <_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
@@ -2315,9 +2307,9 @@ _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 24a8 <_sk_load_g8_hsw+0x54>
+ DB 117,234 ; jne 248c <_sk_load_g8_hsw+0x54>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,163 ; jmp 2468 <_sk_load_g8_hsw+0x14>
+ DB 235,163 ; jmp 244c <_sk_load_g8_hsw+0x14>
PUBLIC _sk_gather_g8_hsw
_sk_gather_g8_hsw LABEL PROC
@@ -2382,9 +2374,9 @@ _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 25bb <_sk_gather_i8_hsw+0xf>
+ DB 116,5 ; je 259f <_sk_gather_i8_hsw+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 25bd <_sk_gather_i8_hsw+0x11>
+ DB 235,2 ; jmp 25a1 <_sk_gather_i8_hsw+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -2422,21 +2414,16 @@ _sk_gather_i8_hsw LABEL PROC
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 197,229,219,5,81,28,0,0 ; vpand 0x1c51(%rip),%ymm3,%ymm0 # 42a0 <_sk_callback_hsw+0x256>
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 196,226,101,0,13,81,28,0,0 ; vpshufb 0x1c51(%rip),%ymm3,%ymm1 # 42c0 <_sk_callback_hsw+0x276>
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 196,226,101,0,21,95,28,0,0 ; vpshufb 0x1c5f(%rip),%ymm3,%ymm2 # 42e0 <_sk_callback_hsw+0x296>
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
@@ -2455,7 +2442,7 @@ _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 276f <_sk_load_565_hsw+0xa3>
+ DB 15,133,149,0,0,0 ; jne 2748 <_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
@@ -2495,9 +2482,9 @@ _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 26e0 <_sk_load_565_hsw+0x14>
+ DB 15,135,89,255,255,255 ; ja 26b9 <_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 # 27dc <_sk_load_565_hsw+0x110>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 27b4 <_sk_load_565_hsw+0x10f>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -2509,27 +2496,27 @@ _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 26e0 <_sk_load_565_hsw+0x14>
- DB 144 ; nop
- DB 243,255 ; repz (bad)
+ DB 233,5,255,255,255 ; jmpq 26b9 <_sk_load_565_hsw+0x14>
+ DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 27e1 <_sk_load_565_hsw+0x115>
DB 255 ; (bad)
- DB 255,227 ; jmpq *%rbx
+ DB 236 ; in (%dx),%al
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
DB 255 ; (bad)
- DB 219,255 ; (bad)
DB 255 ; (bad)
- DB 255,211 ; callq *%rbx
DB 255 ; (bad)
+ DB 220,255 ; fdivr %st,%st(7)
+ DB 255 ; (bad)
+ DB 255,212 ; callq *%rsp
DB 255 ; (bad)
- DB 255,203 ; dec %ebx
DB 255 ; (bad)
+ DB 255,204 ; dec %esp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 191 ; .byte 0xbf
+ DB 255,192 ; inc %eax
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -2637,7 +2624,7 @@ _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 29a7 <_sk_store_565_hsw+0x6c>
+ DB 117,10 ; jne 297f <_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
@@ -2645,9 +2632,9 @@ _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 29a3 <_sk_store_565_hsw+0x68>
+ DB 119,236 ; ja 297b <_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 # 2a04 <_sk_store_565_hsw+0xc9>
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 29dc <_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
@@ -2658,7 +2645,7 @@ _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 29a3 <_sk_store_565_hsw+0x68>
+ DB 235,159 ; jmp 297b <_sk_store_565_hsw+0x68>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2687,7 +2674,7 @@ _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 2ae1 <_sk_load_4444_hsw+0xc1>
+ DB 15,133,179,0,0,0 ; jne 2ab9 <_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
@@ -2733,9 +2720,9 @@ _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 2a34 <_sk_load_4444_hsw+0x14>
+ DB 15,135,59,255,255,255 ; ja 2a0c <_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 # 2b50 <_sk_load_4444_hsw+0x130>
+ DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 2b28 <_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
@@ -2747,13 +2734,13 @@ _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 2a34 <_sk_load_4444_hsw+0x14>
+ DB 233,231,254,255,255 ; jmpq 2a0c <_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 ffffffffe2002b58 <_sk_callback_hsw+0xffffffffe1ffeace>
+ DB 233,255,255,255,225 ; jmpq ffffffffe2002b30 <_sk_callback_hsw+0xffffffffe1ffeae6>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2881,7 +2868,7 @@ _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 2d3f <_sk_store_4444_hsw+0x72>
+ DB 117,10 ; jne 2d17 <_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
@@ -2889,9 +2876,9 @@ _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 2d3b <_sk_store_4444_hsw+0x6e>
+ DB 119,236 ; ja 2d13 <_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 # 2d9c <_sk_store_4444_hsw+0xcf>
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 2d74 <_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
@@ -2902,7 +2889,7 @@ _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 2d3b <_sk_store_4444_hsw+0x6e>
+ DB 235,159 ; jmp 2d13 <_sk_store_4444_hsw+0x6e>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -2933,23 +2920,18 @@ _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 2e35 <_sk_load_8888_hsw+0x7d>
+ DB 117,93 ; jne 2e02 <_sk_load_8888_hsw+0x72>
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
- DB 196,226,125,88,208 ; vpbroadcastd %xmm0,%ymm2
- DB 197,237,219,195 ; vpand %ymm3,%ymm2,%ymm0
+ DB 197,229,219,5,78,21,0,0 ; vpand 0x154e(%rip),%ymm3,%ymm0 # 4300 <_sk_callback_hsw+0x2b6>
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 196,226,101,0,13,78,21,0,0 ; vpshufb 0x154e(%rip),%ymm3,%ymm1 # 4320 <_sk_callback_hsw+0x2d6>
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 196,226,101,0,21,92,21,0,0 ; vpshufb 0x155c(%rip),%ymm3,%ymm2 # 4340 <_sk_callback_hsw+0x2f6>
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
@@ -2966,7 +2948,7 @@ _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 2dd2 <_sk_load_8888_hsw+0x1a>
+ DB 235,130 ; jmp 2daa <_sk_load_8888_hsw+0x1a>
PUBLIC _sk_gather_8888_hsw
_sk_gather_8888_hsw LABEL PROC
@@ -2979,21 +2961,16 @@ _sk_gather_8888_hsw LABEL PROC
DB 197,245,254,192 ; vpaddd %ymm0,%ymm1,%ymm0
DB 197,245,118,201 ; vpcmpeqd %ymm1,%ymm1,%ymm1
DB 196,194,117,144,28,128 ; vpgatherdd %ymm1,(%r8,%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 197,229,219,5,10,21,0,0 ; vpand 0x150a(%rip),%ymm3,%ymm0 # 4360 <_sk_callback_hsw+0x316>
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 196,226,101,0,13,10,21,0,0 ; vpshufb 0x150a(%rip),%ymm3,%ymm1 # 4380 <_sk_callback_hsw+0x336>
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 196,226,101,0,21,24,21,0,0 ; vpshufb 0x1518(%rip),%ymm3,%ymm2 # 43a0 <_sk_callback_hsw+0x356>
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
@@ -3026,7 +3003,7 @@ _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 2f58 <_sk_store_8888_hsw+0x74>
+ DB 117,12 ; jne 2f17 <_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
@@ -3039,14 +3016,14 @@ _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 2f51 <_sk_store_8888_hsw+0x6d>
+ DB 235,211 ; jmp 2f10 <_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 2fe9 <_sk_load_f16_hsw+0x6b>
+ DB 117,97 ; jne 2fa8 <_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
@@ -3072,29 +3049,29 @@ _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 3048 <_sk_load_f16_hsw+0xca>
+ DB 116,79 ; je 3007 <_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 3048 <_sk_load_f16_hsw+0xca>
+ DB 114,67 ; jb 3007 <_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 3055 <_sk_load_f16_hsw+0xd7>
+ DB 116,68 ; je 3014 <_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 3055 <_sk_load_f16_hsw+0xd7>
+ DB 114,56 ; jb 3014 <_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 2f9f <_sk_load_f16_hsw+0x21>
+ DB 15,132,114,255,255,255 ; je 2f5e <_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 2f9f <_sk_load_f16_hsw+0x21>
+ DB 15,130,98,255,255,255 ; jb 2f5e <_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 2f9f <_sk_load_f16_hsw+0x21>
+ DB 233,87,255,255,255 ; jmpq 2f5e <_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 2f9f <_sk_load_f16_hsw+0x21>
+ DB 233,74,255,255,255 ; jmpq 2f5e <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,65,255,255,255 ; jmpq 2f9f <_sk_load_f16_hsw+0x21>
+ DB 233,65,255,255,255 ; jmpq 2f5e <_sk_load_f16_hsw+0x21>
PUBLIC _sk_gather_f16_hsw
_sk_gather_f16_hsw LABEL PROC
@@ -3148,7 +3125,7 @@ _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 314d <_sk_store_f16_hsw+0x65>
+ DB 117,27 ; jne 310c <_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)
@@ -3157,22 +3134,22 @@ _sk_store_f16_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,241 ; je 3149 <_sk_store_f16_hsw+0x61>
+ DB 116,241 ; je 3108 <_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 3149 <_sk_store_f16_hsw+0x61>
+ DB 114,229 ; jb 3108 <_sk_store_f16_hsw+0x61>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,221 ; je 3149 <_sk_store_f16_hsw+0x61>
+ DB 116,221 ; je 3108 <_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 3149 <_sk_store_f16_hsw+0x61>
+ DB 114,209 ; jb 3108 <_sk_store_f16_hsw+0x61>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,201 ; je 3149 <_sk_store_f16_hsw+0x61>
+ DB 116,201 ; je 3108 <_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 3149 <_sk_store_f16_hsw+0x61>
+ DB 114,189 ; jb 3108 <_sk_store_f16_hsw+0x61>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,181 ; jmp 3149 <_sk_store_f16_hsw+0x61>
+ DB 235,181 ; jmp 3108 <_sk_store_f16_hsw+0x61>
PUBLIC _sk_load_u16_be_hsw
_sk_load_u16_be_hsw LABEL PROC
@@ -3180,7 +3157,7 @@ _sk_load_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,205,0,0,0 ; jne 3277 <_sk_load_u16_be_hsw+0xe3>
+ DB 15,133,205,0,0,0 ; jne 3236 <_sk_load_u16_be_hsw+0xe3>
DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8
DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2
DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3
@@ -3229,29 +3206,29 @@ _sk_load_u16_be_hsw LABEL PROC
DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 32dd <_sk_load_u16_be_hsw+0x149>
+ DB 116,85 ; je 329c <_sk_load_u16_be_hsw+0x149>
DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 32dd <_sk_load_u16_be_hsw+0x149>
+ DB 114,72 ; jb 329c <_sk_load_u16_be_hsw+0x149>
DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 32ea <_sk_load_u16_be_hsw+0x156>
+ DB 116,72 ; je 32a9 <_sk_load_u16_be_hsw+0x156>
DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 32ea <_sk_load_u16_be_hsw+0x156>
+ DB 114,59 ; jb 32a9 <_sk_load_u16_be_hsw+0x156>
DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,5,255,255,255 ; je 31c5 <_sk_load_u16_be_hsw+0x31>
+ DB 15,132,5,255,255,255 ; je 3184 <_sk_load_u16_be_hsw+0x31>
DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,244,254,255,255 ; jb 31c5 <_sk_load_u16_be_hsw+0x31>
+ DB 15,130,244,254,255,255 ; jb 3184 <_sk_load_u16_be_hsw+0x31>
DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9
- DB 233,232,254,255,255 ; jmpq 31c5 <_sk_load_u16_be_hsw+0x31>
+ DB 233,232,254,255,255 ; jmpq 3184 <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,219,254,255,255 ; jmpq 31c5 <_sk_load_u16_be_hsw+0x31>
+ DB 233,219,254,255,255 ; jmpq 3184 <_sk_load_u16_be_hsw+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,210,254,255,255 ; jmpq 31c5 <_sk_load_u16_be_hsw+0x31>
+ DB 233,210,254,255,255 ; jmpq 3184 <_sk_load_u16_be_hsw+0x31>
PUBLIC _sk_load_rgb_u16_be_hsw
_sk_load_rgb_u16_be_hsw LABEL PROC
@@ -3259,7 +3236,7 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,211,0,0,0 ; jne 33d8 <_sk_load_rgb_u16_be_hsw+0xe5>
+ DB 15,133,211,0,0,0 ; jne 3397 <_sk_load_rgb_u16_be_hsw+0xe5>
DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0
DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2
DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1
@@ -3309,36 +3286,36 @@ _sk_load_rgb_u16_be_hsw LABEL PROC
DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0
DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 33f1 <_sk_load_rgb_u16_be_hsw+0xfe>
- DB 233,72,255,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,5 ; jne 33b0 <_sk_load_rgb_u16_be_hsw+0xfe>
+ DB 233,72,255,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1
DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 3420 <_sk_load_rgb_u16_be_hsw+0x12d>
+ DB 114,26 ; jb 33df <_sk_load_rgb_u16_be_hsw+0x12d>
DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1
DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 3425 <_sk_load_rgb_u16_be_hsw+0x132>
- DB 233,25,255,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,20,255,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 33e4 <_sk_load_rgb_u16_be_hsw+0x132>
+ DB 233,25,255,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,20,255,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1
DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 3454 <_sk_load_rgb_u16_be_hsw+0x161>
+ DB 114,26 ; jb 3413 <_sk_load_rgb_u16_be_hsw+0x161>
DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1
DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 3459 <_sk_load_rgb_u16_be_hsw+0x166>
- DB 233,229,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,224,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 117,10 ; jne 3418 <_sk_load_rgb_u16_be_hsw+0x166>
+ DB 233,229,254,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,224,254,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3
DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 3482 <_sk_load_rgb_u16_be_hsw+0x18f>
+ DB 114,20 ; jb 3441 <_sk_load_rgb_u16_be_hsw+0x18f>
DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3
DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
- DB 233,183,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
- DB 233,178,254,255,255 ; jmpq 3339 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,183,254,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
+ DB 233,178,254,255,255 ; jmpq 32f8 <_sk_load_rgb_u16_be_hsw+0x46>
PUBLIC _sk_store_u16_be_hsw
_sk_store_u16_be_hsw LABEL PROC
@@ -3385,7 +3362,7 @@ _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 3582 <_sk_store_u16_be_hsw+0xfb>
+ DB 117,31 ; jne 3541 <_sk_store_u16_be_hsw+0xfb>
DB 196,1,120,17,28,72 ; vmovups %xmm11,(%r8,%r9,2)
DB 196,1,120,17,84,72,16 ; vmovups %xmm10,0x10(%r8,%r9,2)
DB 196,1,120,17,76,72,32 ; vmovups %xmm9,0x20(%r8,%r9,2)
@@ -3394,31 +3371,31 @@ _sk_store_u16_be_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,1,121,214,28,72 ; vmovq %xmm11,(%r8,%r9,2)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 116,240 ; je 353d <_sk_store_u16_be_hsw+0xf7>
DB 196,1,121,23,92,72,8 ; vmovhpd %xmm11,0x8(%r8,%r9,2)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 114,227 ; jb 353d <_sk_store_u16_be_hsw+0xf7>
DB 196,1,121,214,84,72,16 ; vmovq %xmm10,0x10(%r8,%r9,2)
- DB 116,218 ; je 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 116,218 ; je 353d <_sk_store_u16_be_hsw+0xf7>
DB 196,1,121,23,84,72,24 ; vmovhpd %xmm10,0x18(%r8,%r9,2)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 114,205 ; jb 353d <_sk_store_u16_be_hsw+0xf7>
DB 196,1,121,214,76,72,32 ; vmovq %xmm9,0x20(%r8,%r9,2)
- DB 116,196 ; je 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 116,196 ; je 353d <_sk_store_u16_be_hsw+0xf7>
DB 196,1,121,23,76,72,40 ; vmovhpd %xmm9,0x28(%r8,%r9,2)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,183 ; jb 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 114,183 ; jb 353d <_sk_store_u16_be_hsw+0xf7>
DB 196,1,121,214,68,72,48 ; vmovq %xmm8,0x30(%r8,%r9,2)
- DB 235,174 ; jmp 357e <_sk_store_u16_be_hsw+0xf7>
+ DB 235,174 ; jmp 353d <_sk_store_u16_be_hsw+0xf7>
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 3646 <_sk_load_f32_hsw+0x76>
+ DB 119,110 ; ja 3605 <_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 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 3670 <_sk_load_f32_hsw+0xa0>
+ DB 76,141,21,135,0,0,0 ; lea 0x87(%rip),%r10 # 3630 <_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
@@ -3444,19 +3421,21 @@ _sk_load_f32_hsw LABEL PROC
DB 196,193,101,21,216 ; vunpckhpd %ymm8,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
- DB 102,144 ; xchg %ax,%ax
- DB 131,255,255 ; cmp $0xffffffff,%edi
- DB 255,202 ; dec %edx
+ DB 15,31,0 ; nopl (%rax)
+ DB 130 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,201 ; dec %ecx
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 189,255,255,255,176 ; mov $0xb0ffffff,%ebp
+ DB 188,255,255,255,175 ; mov $0xafffffff,%esp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,163,255,255,255,155 ; jmpq *-0x64000001(%rbx)
+ DB 255,162,255,255,255,154 ; jmpq *-0x65000001(%rdx)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,147,255,255,255,139 ; callq *-0x74000001(%rbx)
+ DB 255,146,255,255,255,138 ; callq *-0x75000001(%rdx)
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -3475,7 +3454,7 @@ _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 36fd <_sk_store_f32_hsw+0x6d>
+ DB 117,55 ; jne 36bd <_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
@@ -3488,22 +3467,22 @@ _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 36f9 <_sk_store_f32_hsw+0x69>
+ DB 116,240 ; je 36b9 <_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 36f9 <_sk_store_f32_hsw+0x69>
+ DB 114,227 ; jb 36b9 <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 36f9 <_sk_store_f32_hsw+0x69>
+ DB 116,218 ; je 36b9 <_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 36f9 <_sk_store_f32_hsw+0x69>
+ DB 114,205 ; jb 36b9 <_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 36f9 <_sk_store_f32_hsw+0x69>
+ DB 116,195 ; je 36b9 <_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 36f9 <_sk_store_f32_hsw+0x69>
+ DB 114,181 ; jb 36b9 <_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 36f9 <_sk_store_f32_hsw+0x69>
+ DB 235,171 ; jmp 36b9 <_sk_store_f32_hsw+0x69>
PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
@@ -3744,7 +3723,7 @@ _sk_linear_gradient_hsw LABEL PROC
DB 196,98,125,24,72,28 ; vbroadcastss 0x1c(%rax),%ymm9
DB 76,139,0 ; mov (%rax),%r8
DB 77,133,192 ; test %r8,%r8
- DB 15,132,143,0,0,0 ; je 3b89 <_sk_linear_gradient_hsw+0xb5>
+ DB 15,132,143,0,0,0 ; je 3b49 <_sk_linear_gradient_hsw+0xb5>
DB 72,139,64,8 ; mov 0x8(%rax),%rax
DB 72,131,192,32 ; add $0x20,%rax
DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12
@@ -3771,8 +3750,8 @@ _sk_linear_gradient_hsw LABEL PROC
DB 196,67,13,74,201,208 ; vblendvps %ymm13,%ymm9,%ymm14,%ymm9
DB 72,131,192,36 ; add $0x24,%rax
DB 73,255,200 ; dec %r8
- DB 117,140 ; jne 3b13 <_sk_linear_gradient_hsw+0x3f>
- DB 235,17 ; jmp 3b9a <_sk_linear_gradient_hsw+0xc6>
+ DB 117,140 ; jne 3ad3 <_sk_linear_gradient_hsw+0x3f>
+ DB 235,17 ; jmp 3b5a <_sk_linear_gradient_hsw+0xc6>
DB 197,244,87,201 ; vxorps %ymm1,%ymm1,%ymm1
DB 197,236,87,210 ; vxorps %ymm2,%ymm2,%ymm2
DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3
@@ -4162,10 +4141,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 41b4 <.literal4+0x10>
+ DB 127,0 ; jg 4174 <.literal4+0x10>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 422d <.literal4+0x89>
+ DB 119,115 ; ja 41ed <.literal4+0x89>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4179,10 +4158,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 41e4 <.literal4+0x40>
+ DB 127,0 ; jg 41a4 <.literal4+0x40>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 425d <.literal4+0xb9>
+ DB 119,115 ; ja 421d <.literal4+0xb9>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4196,10 +4175,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4214 <.literal4+0x70>
+ DB 127,0 ; jg 41d4 <.literal4+0x70>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 428d <_sk_callback_hsw+0x203>
+ DB 119,115 ; ja 424d <_sk_callback_hsw+0x203>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4213,10 +4192,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4244 <.literal4+0xa0>
+ DB 127,0 ; jg 4204 <.literal4+0xa0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 42bd <_sk_callback_hsw+0x233>
+ DB 119,115 ; ja 427d <_sk_callback_hsw+0x233>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -4228,6 +4207,234 @@ ALIGN 4
DB 65,0,0 ; add %al,(%r8)
DB 0 ; .byte 0x0
DB 75 ; rex.WXB
+
+ALIGN 32
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004268 <_sk_callback_hsw+0xa00021e>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004270 <_sk_callback_hsw+0x12000226>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004278 <_sk_callback_hsw+0x1a00022e>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004280 <_sk_callback_hsw+0x3000236>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,18 ; callq *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,22 ; callq *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,26 ; lcall *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,30 ; lcall *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0042c8 <_sk_callback_hsw+0xa00027e>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 120042d0 <_sk_callback_hsw+0x12000286>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a0042d8 <_sk_callback_hsw+0x1a00028e>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30042e0 <_sk_callback_hsw+0x3000296>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,18 ; callq *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,22 ; callq *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,26 ; lcall *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,30 ; lcall *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004328 <_sk_callback_hsw+0xa0002de>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004330 <_sk_callback_hsw+0x120002e6>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004338 <_sk_callback_hsw+0x1a0002ee>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 3004340 <_sk_callback_hsw+0x30002f6>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,18 ; callq *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,22 ; callq *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,26 ; lcall *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,30 ; lcall *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004388 <_sk_callback_hsw+0xa00033e>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,17 ; decl 0x11ffffff(%rip) # 12004390 <_sk_callback_hsw+0x12000346>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,21,255,255,255,25 ; callq *0x19ffffff(%rip) # 1a004398 <_sk_callback_hsw+0x1a00034e>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,29,255,255,255,2 ; lcall *0x2ffffff(%rip) # 30043a0 <_sk_callback_hsw+0x3000356>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,18 ; callq *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,22 ; callq *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,26 ; lcall *(%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,30 ; lcall *(%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
+
+ALIGN 16
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
ALIGN 32
PUBLIC _sk_start_pipeline_avx
@@ -4327,14 +4534,14 @@ _sk_seed_shader_avx LABEL PROC
DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0
DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
- DB 196,226,125,24,13,164,91,0,0 ; vbroadcastss 0x5ba4(%rip),%ymm1 # 5d04 <_sk_callback_avx+0x11a>
+ DB 196,226,125,24,13,40,91,0,0 ; vbroadcastss 0x5b28(%rip),%ymm1 # 5c88 <_sk_callback_avx+0x11a>
DB 197,252,88,193 ; vaddps %ymm1,%ymm0,%ymm0
DB 197,252,88,2 ; vaddps (%rdx),%ymm0,%ymm0
DB 196,226,125,24,16 ; vbroadcastss (%rax),%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 197,236,88,201 ; vaddps %ymm1,%ymm2,%ymm1
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 196,226,125,24,21,136,91,0,0 ; vbroadcastss 0x5b88(%rip),%ymm2 # 5d08 <_sk_callback_avx+0x11e>
+ DB 196,226,125,24,21,12,91,0,0 ; vbroadcastss 0x5b0c(%rip),%ymm2 # 5c8c <_sk_callback_avx+0x11e>
DB 197,228,87,219 ; vxorps %ymm3,%ymm3,%ymm3
DB 197,220,87,228 ; vxorps %ymm4,%ymm4,%ymm4
DB 197,212,87,237 ; vxorps %ymm5,%ymm5,%ymm5
@@ -5668,22 +5875,19 @@ _sk_lerp_565_avx LABEL PROC
PUBLIC _sk_load_tables_avx
_sk_load_tables_avx LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,139,0 ; mov (%rax),%r8
+ DB 72,133,201 ; test %rcx,%rcx
+ DB 15,133,37,2,0,0 ; jne 1863 <_sk_load_tables_avx+0x233>
+ DB 196,65,124,16,4,184 ; vmovups (%r8,%rdi,4),%ymm8
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 76,139,0 ; mov (%rax),%r8
- DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,56,2,0,0 ; jne 1880 <_sk_load_tables_avx+0x250>
- DB 196,65,124,16,4,184 ; vmovups (%r8,%rdi,4),%ymm8
- DB 187,255,0,0,0 ; mov $0xff,%ebx
- DB 197,249,110,195 ; vmovd %ebx,%xmm0
- DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0
- DB 196,99,125,24,200,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm9
- DB 196,193,52,84,192 ; vandps %ymm8,%ymm9,%ymm0
+ DB 197,124,40,13,42,71,0,0 ; vmovaps 0x472a(%rip),%ymm9 # 5d80 <_sk_callback_avx+0x212>
+ DB 196,193,60,84,193 ; vandps %ymm9,%ymm8,%ymm0
DB 196,193,249,126,193 ; vmovq %xmm0,%r9
DB 69,137,203 ; mov %r9d,%r11d
DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10
@@ -5701,22 +5905,18 @@ _sk_load_tables_avx LABEL PROC
DB 76,139,64,16 ; mov 0x10(%rax),%r8
DB 196,161,122,16,68,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm0
DB 196,163,121,33,68,165,0,16 ; vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
- DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1
- DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0
- DB 197,250,16,76,157,0 ; vmovss 0x0(%rbp,%rbx,4),%xmm1
- DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0
+ DB 196,163,121,33,68,173,0,32 ; vinsertps $0x20,0x0(%rbp,%r13,4),%xmm0,%xmm0
+ DB 196,227,121,33,68,157,0,48 ; vinsertps $0x30,0x0(%rbp,%rbx,4),%xmm0,%xmm0
DB 196,161,122,16,76,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm1
DB 196,163,113,33,76,141,0,16 ; vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
- DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3
- DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1
- DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3
- DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1
+ DB 196,163,113,33,76,181,0,32 ; vinsertps $0x20,0x0(%rbp,%r14,4),%xmm1,%xmm1
+ DB 196,163,113,33,76,149,0,48 ; vinsertps $0x30,0x0(%rbp,%r10,4),%xmm1,%xmm1
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
DB 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1
DB 196,67,125,25,194,1 ; vextractf128 $0x1,%ymm8,%xmm10
DB 196,193,105,114,210,8 ; vpsrld $0x8,%xmm10,%xmm2
DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
- DB 197,180,84,201 ; vandps %ymm1,%ymm9,%ymm1
+ DB 196,193,116,84,201 ; vandps %ymm9,%ymm1,%ymm1
DB 196,193,249,126,201 ; vmovq %xmm1,%r9
DB 69,137,203 ; mov %r9d,%r11d
DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10
@@ -5747,7 +5947,7 @@ _sk_load_tables_avx LABEL PROC
DB 196,193,105,114,208,16 ; vpsrld $0x10,%xmm8,%xmm2
DB 196,193,97,114,210,16 ; vpsrld $0x10,%xmm10,%xmm3
DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
- DB 197,180,84,210 ; vandps %ymm2,%ymm9,%ymm2
+ DB 196,193,108,84,209 ; vandps %ymm9,%ymm2,%ymm2
DB 196,193,249,126,208 ; vmovq %xmm2,%r8
DB 69,137,194 ; mov %r8d,%r10d
DB 196,195,249,22,209,1 ; vpextrq $0x1,%xmm2,%r9
@@ -5791,17 +5991,17 @@ _sk_load_tables_avx LABEL PROC
DB 65,95 ; pop %r15
DB 93 ; pop %rbp
DB 255,224 ; jmpq *%rax
- DB 137,203 ; mov %ecx,%ebx
- DB 128,227,7 ; and $0x7,%bl
+ DB 65,137,201 ; mov %ecx,%r9d
+ DB 65,128,225,7 ; and $0x7,%r9b
DB 196,65,60,87,192 ; vxorps %ymm8,%ymm8,%ymm8
- DB 254,203 ; dec %bl
- DB 128,251,6 ; cmp $0x6,%bl
- DB 15,135,185,253,255,255 ; ja 164e <_sk_load_tables_avx+0x1e>
- DB 15,182,219 ; movzbl %bl,%ebx
- DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 1928 <_sk_load_tables_avx+0x2f8>
- DB 73,99,28,153 ; movslq (%r9,%rbx,4),%rbx
- DB 76,1,203 ; add %r9,%rbx
- DB 255,227 ; jmpq *%rbx
+ DB 65,254,201 ; dec %r9b
+ DB 65,128,249,6 ; cmp $0x6,%r9b
+ DB 15,135,200,253,255,255 ; ja 1644 <_sk_load_tables_avx+0x14>
+ DB 69,15,182,201 ; movzbl %r9b,%r9d
+ DB 76,141,21,141,0,0,0 ; lea 0x8d(%rip),%r10 # 1914 <_sk_load_tables_avx+0x2e4>
+ DB 79,99,12,138 ; movslq (%r10,%r9,4),%r9
+ DB 77,1,209 ; add %r10,%r9
+ DB 65,255,225 ; jmpq *%r9
DB 196,193,121,110,68,184,24 ; vmovd 0x18(%r8,%rdi,4),%xmm0
DB 197,249,112,192,68 ; vpshufd $0x44,%xmm0,%xmm0
DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
@@ -5821,24 +6021,25 @@ _sk_load_tables_avx LABEL PROC
DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8
DB 196,195,57,34,4,184,0 ; vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0
DB 196,99,61,12,192,15 ; vblendps $0xf,%ymm0,%ymm8,%ymm8
- DB 233,38,253,255,255 ; jmpq 164e <_sk_load_tables_avx+0x1e>
- DB 238 ; out %al,(%dx)
+ DB 233,51,253,255,255 ; jmpq 1644 <_sk_load_tables_avx+0x14>
+ DB 15,31,0 ; nopl (%rax)
+ DB 235,255 ; jmp 1915 <_sk_load_tables_avx+0x2e5>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,224 ; jmpq *%rax
+ DB 221,255 ; (bad)
DB 255 ; (bad)
+ DB 255,207 ; dec %edi
DB 255 ; (bad)
- DB 255,210 ; callq *%rdx
DB 255 ; (bad)
+ DB 255,193 ; inc %ecx
DB 255 ; (bad)
- DB 255,196 ; inc %esp
DB 255 ; (bad)
+ DB 255,173,255,255,255,153 ; ljmp *-0x66000001(%rbp)
DB 255 ; (bad)
- DB 255,176,255,255,255,156 ; pushq -0x63000001(%rax)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255 ; .byte 0xff
- DB 128,255,255 ; cmp $0xff,%bh
+ DB 125,255 ; jge 192d <_sk_load_tables_avx+0x2fd>
+ DB 255 ; (bad)
DB 255 ; .byte 0xff
PUBLIC _sk_load_tables_u16_be_avx
@@ -5847,7 +6048,7 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,165,2,0,0 ; jne 1bff <_sk_load_tables_u16_be_avx+0x2bb>
+ DB 15,133,125,2,0,0 ; jne 1bc3 <_sk_load_tables_u16_be_avx+0x293>
DB 196,1,121,16,4,72 ; vmovupd (%r8,%r9,2),%xmm8
DB 196,129,121,16,84,72,16 ; vmovupd 0x10(%r8,%r9,2),%xmm2
DB 196,129,121,16,92,72,32 ; vmovupd 0x20(%r8,%r9,2),%xmm3
@@ -5862,124 +6063,117 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 197,185,105,210 ; vpunpckhwd %xmm2,%xmm8,%xmm2
DB 196,193,97,97,201 ; vpunpcklwd %xmm9,%xmm3,%xmm1
DB 196,193,97,105,217 ; vpunpckhwd %xmm9,%xmm3,%xmm3
- DB 197,121,97,218 ; vpunpcklwd %xmm2,%xmm0,%xmm11
- DB 197,121,105,202 ; vpunpckhwd %xmm2,%xmm0,%xmm9
- DB 197,241,97,211 ; vpunpcklwd %xmm3,%xmm1,%xmm2
- DB 197,113,105,211 ; vpunpckhwd %xmm3,%xmm1,%xmm10
- DB 197,161,108,194 ; vpunpcklqdq %xmm2,%xmm11,%xmm0
- DB 196,65,57,239,192 ; vpxor %xmm8,%xmm8,%xmm8
- DB 196,193,121,105,200 ; vpunpckhwd %xmm8,%xmm0,%xmm1
- DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
- DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- DB 187,255,0,0,0 ; mov $0xff,%ebx
- DB 197,249,110,203 ; vmovd %ebx,%xmm1
- DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1
- DB 196,99,117,24,225,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm12
- DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0
- DB 196,193,249,126,193 ; vmovq %xmm0,%r9
- DB 69,137,203 ; mov %r9d,%r11d
- DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10
- DB 69,137,214 ; mov %r10d,%r14d
+ DB 197,121,97,202 ; vpunpcklwd %xmm2,%xmm0,%xmm9
+ DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8
+ DB 197,241,97,195 ; vpunpcklwd %xmm3,%xmm1,%xmm0
+ DB 197,113,105,227 ; vpunpckhwd %xmm3,%xmm1,%xmm12
+ DB 197,177,108,208 ; vpunpcklqdq %xmm0,%xmm9,%xmm2
+ DB 197,177,109,200 ; vpunpckhqdq %xmm0,%xmm9,%xmm1
+ DB 196,65,57,108,212 ; vpunpcklqdq %xmm12,%xmm8,%xmm10
+ DB 197,121,111,29,94,68,0,0 ; vmovdqa 0x445e(%rip),%xmm11 # 5e00 <_sk_callback_avx+0x292>
+ DB 196,193,105,219,195 ; vpand %xmm11,%xmm2,%xmm0
+ DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
+ DB 196,193,121,105,209 ; vpunpckhwd %xmm9,%xmm0,%xmm2
+ DB 196,195,249,22,208,1 ; vpextrq $0x1,%xmm2,%r8
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 77,137,194 ; mov %r8,%r10
DB 73,193,234,32 ; shr $0x20,%r10
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0
- DB 196,193,249,126,196 ; vmovq %xmm0,%r12
- DB 69,137,231 ; mov %r12d,%r15d
- DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx
- DB 65,137,221 ; mov %ebx,%r13d
- DB 72,193,235,32 ; shr $0x20,%rbx
- DB 73,193,236,32 ; shr $0x20,%r12
+ DB 196,193,249,126,208 ; vmovq %xmm2,%r8
+ DB 69,137,195 ; mov %r8d,%r11d
+ DB 77,137,198 ; mov %r8,%r14
+ DB 73,193,238,32 ; shr $0x20,%r14
+ DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
+ DB 196,225,249,126,195 ; vmovq %xmm0,%rbx
+ DB 65,137,223 ; mov %ebx,%r15d
+ DB 72,193,235,30 ; shr $0x1e,%rbx
+ DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12
+ DB 69,137,229 ; mov %r12d,%r13d
+ DB 73,193,236,30 ; shr $0x1e,%r12
DB 72,139,104,8 ; mov 0x8(%rax),%rbp
DB 76,139,64,16 ; mov 0x10(%rax),%r8
- DB 196,161,122,16,68,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm0
- DB 196,163,121,33,68,165,0,16 ; vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
- DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1
- DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0
- DB 197,250,16,76,157,0 ; vmovss 0x0(%rbp,%rbx,4),%xmm1
- DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0
- DB 196,161,122,16,76,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm1
- DB 196,163,113,33,76,141,0,16 ; vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
- DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3
- DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1
- DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3
- DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1
- DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
- DB 197,161,109,202 ; vpunpckhqdq %xmm2,%xmm11,%xmm1
- DB 196,193,113,105,208 ; vpunpckhwd %xmm8,%xmm1,%xmm2
- DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1
- DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
- DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1
- DB 196,193,249,126,201 ; vmovq %xmm1,%r9
- DB 69,137,203 ; mov %r9d,%r11d
- DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10
- DB 69,137,214 ; mov %r10d,%r14d
- DB 73,193,234,32 ; shr $0x20,%r10
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1
- DB 196,225,249,126,205 ; vmovq %xmm1,%rbp
- DB 65,137,239 ; mov %ebp,%r15d
- DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx
- DB 65,137,220 ; mov %ebx,%r12d
- DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,161,122,16,68,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm0
+ DB 196,163,121,33,68,181,0,16 ; vinsertps $0x10,0x0(%rbp,%r14,4),%xmm0,%xmm0
+ DB 196,161,122,16,84,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm2
+ DB 196,227,121,33,194,32 ; vinsertps $0x20,%xmm2,%xmm0,%xmm0
+ DB 196,161,122,16,84,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm2
+ DB 196,227,121,33,194,48 ; vinsertps $0x30,%xmm2,%xmm0,%xmm0
+ DB 196,161,122,16,84,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm2
+ DB 196,227,105,33,84,29,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,1),%xmm2,%xmm2
+ DB 196,161,122,16,92,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm3
+ DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2
+ DB 196,161,122,16,92,37,0 ; vmovss 0x0(%rbp,%r12,1),%xmm3
+ DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2
+ DB 196,227,109,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm2,%ymm0
+ DB 196,193,113,219,203 ; vpand %xmm11,%xmm1,%xmm1
+ DB 196,193,113,105,209 ; vpunpckhwd %xmm9,%xmm1,%xmm2
+ 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 196,129,122,16,12,184 ; vmovss (%r8,%r15,4),%xmm1
- DB 196,195,113,33,12,168,16 ; vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1
- DB 196,129,122,16,20,160 ; vmovss (%r8,%r12,4),%xmm2
+ DB 196,225,249,126,211 ; vmovq %xmm2,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1
+ DB 196,193,249,126,203 ; vmovq %xmm1,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,30 ; shr $0x1e,%r11
+ DB 196,195,249,22,207,1 ; vpextrq $0x1,%xmm1,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,30 ; shr $0x1e,%r15
+ DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1
+ DB 196,195,113,33,12,152,16 ; vinsertps $0x10,(%r8,%rbx,4),%xmm1,%xmm1
+ DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2
DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1
- DB 196,193,122,16,20,152 ; vmovss (%r8,%rbx,4),%xmm2
+ DB 196,193,122,16,20,168 ; vmovss (%r8,%rbp,4),%xmm2
DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1
- DB 196,129,122,16,20,152 ; vmovss (%r8,%r11,4),%xmm2
- DB 196,131,105,33,20,136,16 ; vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2
- DB 196,129,122,16,28,176 ; vmovss (%r8,%r14,4),%xmm3
+ DB 196,129,122,16,20,176 ; vmovss (%r8,%r14,4),%xmm2
+ DB 196,131,105,33,20,24,16 ; vinsertps $0x10,(%r8,%r11,1),%xmm2,%xmm2
+ DB 196,129,122,16,28,160 ; vmovss (%r8,%r12,4),%xmm3
DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2
- DB 196,129,122,16,28,144 ; vmovss (%r8,%r10,4),%xmm3
+ DB 196,129,122,16,28,56 ; vmovss (%r8,%r15,1),%xmm3
DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2
DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1
- DB 72,139,64,24 ; mov 0x18(%rax),%rax
- DB 196,193,49,108,210 ; vpunpcklqdq %xmm10,%xmm9,%xmm2
- DB 196,193,105,105,216 ; vpunpckhwd %xmm8,%xmm2,%xmm3
- DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2
- DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
- DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2
- DB 196,193,249,126,208 ; vmovq %xmm2,%r8
- DB 69,137,194 ; mov %r8d,%r10d
- DB 196,195,249,22,209,1 ; vpextrq $0x1,%xmm2,%r9
- DB 69,137,203 ; mov %r9d,%r11d
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 73,193,232,32 ; shr $0x20,%r8
- DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2
- DB 196,225,249,126,213 ; vmovq %xmm2,%rbp
- DB 65,137,238 ; mov %ebp,%r14d
- DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx
- DB 65,137,223 ; mov %ebx,%r15d
- DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 76,139,80,24 ; mov 0x18(%rax),%r10
+ DB 196,193,41,219,211 ; vpand %xmm11,%xmm10,%xmm2
+ DB 196,193,105,105,217 ; vpunpckhwd %xmm9,%xmm2,%xmm3
+ 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 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2
- DB 196,227,105,33,20,168,16 ; vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2
- DB 196,161,122,16,28,184 ; vmovss (%rax,%r15,4),%xmm3
+ 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 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2
+ DB 196,225,249,126,208 ; vmovq %xmm2,%rax
+ DB 65,137,195 ; mov %eax,%r11d
+ DB 72,193,232,30 ; shr $0x1e,%rax
+ DB 196,195,249,22,214,1 ; vpextrq $0x1,%xmm2,%r14
+ DB 69,137,247 ; mov %r14d,%r15d
+ DB 73,193,238,30 ; shr $0x1e,%r14
+ DB 196,129,122,16,20,138 ; vmovss (%r10,%r9,4),%xmm2
+ DB 196,195,105,33,20,154,16 ; vinsertps $0x10,(%r10,%rbx,4),%xmm2,%xmm2
+ DB 196,129,122,16,28,130 ; vmovss (%r10,%r8,4),%xmm3
DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2
- DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3
- DB 196,99,105,33,219,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm11
- DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3
- DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3
- DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2
+ DB 196,193,122,16,28,170 ; vmovss (%r10,%rbp,4),%xmm3
+ DB 196,99,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm10
+ DB 196,129,122,16,28,154 ; vmovss (%r10,%r11,4),%xmm3
+ DB 196,195,97,33,28,2,16 ; vinsertps $0x10,(%r10,%rax,1),%xmm3,%xmm3
+ DB 196,129,122,16,20,186 ; vmovss (%r10,%r15,4),%xmm2
DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2
- DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3
+ DB 196,129,122,16,28,50 ; vmovss (%r10,%r14,1),%xmm3
DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2
- DB 196,195,109,24,211,1 ; vinsertf128 $0x1,%xmm11,%ymm2,%ymm2
+ DB 196,195,109,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm2,%ymm2
DB 184,128,0,128,55 ; mov $0x37800080,%eax
DB 197,249,110,216 ; vmovd %eax,%xmm3
DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3
- DB 196,99,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm11
- DB 196,193,49,109,218 ; vpunpckhqdq %xmm10,%xmm9,%xmm3
- DB 197,177,113,243,8 ; vpsllw $0x8,%xmm3,%xmm9
+ DB 196,99,101,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm10
+ DB 196,193,57,109,220 ; vpunpckhqdq %xmm12,%xmm8,%xmm3
+ DB 197,185,113,243,8 ; vpsllw $0x8,%xmm3,%xmm8
DB 197,225,113,211,8 ; vpsrlw $0x8,%xmm3,%xmm3
- DB 197,177,235,219 ; vpor %xmm3,%xmm9,%xmm3
- DB 196,65,97,105,192 ; vpunpckhwd %xmm8,%xmm3,%xmm8
+ DB 197,185,235,219 ; vpor %xmm3,%xmm8,%xmm3
+ DB 196,65,97,105,193 ; vpunpckhwd %xmm9,%xmm3,%xmm8
DB 196,226,121,51,219 ; vpmovzxwd %xmm3,%xmm3
DB 196,195,101,24,216,1 ; vinsertf128 $0x1,%xmm8,%ymm3,%ymm3
DB 197,252,91,219 ; vcvtdq2ps %ymm3,%ymm3
- DB 196,193,100,89,219 ; vmulps %ymm11,%ymm3,%ymm3
+ DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 91 ; pop %rbx
DB 65,92 ; pop %r12
@@ -5991,29 +6185,29 @@ _sk_load_tables_u16_be_avx LABEL PROC
DB 196,1,123,16,4,72 ; vmovsd (%r8,%r9,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 1c65 <_sk_load_tables_u16_be_avx+0x321>
+ DB 116,85 ; je 1c29 <_sk_load_tables_u16_be_avx+0x2f9>
DB 196,1,57,22,68,72,8 ; vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 1c65 <_sk_load_tables_u16_be_avx+0x321>
+ DB 114,72 ; jb 1c29 <_sk_load_tables_u16_be_avx+0x2f9>
DB 196,129,123,16,84,72,16 ; vmovsd 0x10(%r8,%r9,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 1c72 <_sk_load_tables_u16_be_avx+0x32e>
+ DB 116,72 ; je 1c36 <_sk_load_tables_u16_be_avx+0x306>
DB 196,129,105,22,84,72,24 ; vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 1c72 <_sk_load_tables_u16_be_avx+0x32e>
+ DB 114,59 ; jb 1c36 <_sk_load_tables_u16_be_avx+0x306>
DB 196,129,123,16,92,72,32 ; vmovsd 0x20(%r8,%r9,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,45,253,255,255 ; je 1975 <_sk_load_tables_u16_be_avx+0x31>
+ DB 15,132,85,253,255,255 ; je 1961 <_sk_load_tables_u16_be_avx+0x31>
DB 196,129,97,22,92,72,40 ; vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,28,253,255,255 ; jb 1975 <_sk_load_tables_u16_be_avx+0x31>
+ DB 15,130,68,253,255,255 ; jb 1961 <_sk_load_tables_u16_be_avx+0x31>
DB 196,1,122,126,76,72,48 ; vmovq 0x30(%r8,%r9,2),%xmm9
- DB 233,16,253,255,255 ; jmpq 1975 <_sk_load_tables_u16_be_avx+0x31>
+ DB 233,56,253,255,255 ; jmpq 1961 <_sk_load_tables_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,3,253,255,255 ; jmpq 1975 <_sk_load_tables_u16_be_avx+0x31>
+ DB 233,43,253,255,255 ; jmpq 1961 <_sk_load_tables_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,250,252,255,255 ; jmpq 1975 <_sk_load_tables_u16_be_avx+0x31>
+ DB 233,34,253,255,255 ; jmpq 1961 <_sk_load_tables_u16_be_avx+0x31>
PUBLIC _sk_load_tables_rgb_u16_be_avx
_sk_load_tables_rgb_u16_be_avx LABEL PROC
@@ -6021,7 +6215,7 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 76,141,12,127 ; lea (%rdi,%rdi,2),%r9
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,146,2,0,0 ; jne 1f1f <_sk_load_tables_rgb_u16_be_avx+0x2a4>
+ DB 15,133,105,2,0,0 ; jne 1eba <_sk_load_tables_rgb_u16_be_avx+0x27b>
DB 196,129,122,111,4,72 ; vmovdqu (%r8,%r9,2),%xmm0
DB 196,129,122,111,84,72,12 ; vmovdqu 0xc(%r8,%r9,2),%xmm2
DB 196,129,122,111,76,72,24 ; vmovdqu 0x18(%r8,%r9,2),%xmm1
@@ -6041,109 +6235,102 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 196,193,57,97,209 ; vpunpcklwd %xmm9,%xmm8,%xmm2
DB 197,241,97,203 ; vpunpcklwd %xmm3,%xmm1,%xmm1
DB 196,193,33,97,218 ; vpunpcklwd %xmm10,%xmm11,%xmm3
- DB 197,121,97,218 ; vpunpcklwd %xmm2,%xmm0,%xmm11
- DB 197,121,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm8
+ DB 197,121,97,194 ; vpunpcklwd %xmm2,%xmm0,%xmm8
+ DB 197,249,105,194 ; vpunpckhwd %xmm2,%xmm0,%xmm0
DB 197,241,97,211 ; vpunpcklwd %xmm3,%xmm1,%xmm2
- DB 197,113,105,211 ; vpunpckhwd %xmm3,%xmm1,%xmm10
- DB 197,161,108,194 ; vpunpcklqdq %xmm2,%xmm11,%xmm0
- DB 196,65,25,239,228 ; vpxor %xmm12,%xmm12,%xmm12
- DB 196,193,121,105,204 ; vpunpckhwd %xmm12,%xmm0,%xmm1
- DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
- DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- DB 187,255,0,0,0 ; mov $0xff,%ebx
- DB 197,249,110,203 ; vmovd %ebx,%xmm1
- DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1
- DB 196,99,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm9
- DB 196,193,124,84,193 ; vandps %ymm9,%ymm0,%ymm0
- DB 196,193,249,126,193 ; vmovq %xmm0,%r9
- DB 69,137,203 ; mov %r9d,%r11d
- DB 196,195,249,22,194,1 ; vpextrq $0x1,%xmm0,%r10
- DB 69,137,214 ; mov %r10d,%r14d
+ DB 197,241,105,219 ; vpunpckhwd %xmm3,%xmm1,%xmm3
+ DB 197,185,108,202 ; vpunpcklqdq %xmm2,%xmm8,%xmm1
+ DB 197,185,109,210 ; vpunpckhqdq %xmm2,%xmm8,%xmm2
+ DB 197,121,108,195 ; vpunpcklqdq %xmm3,%xmm0,%xmm8
+ DB 197,121,111,13,75,65,0,0 ; vmovdqa 0x414b(%rip),%xmm9 # 5e10 <_sk_callback_avx+0x2a2>
+ DB 196,193,113,219,193 ; vpand %xmm9,%xmm1,%xmm0
+ DB 196,65,41,239,210 ; vpxor %xmm10,%xmm10,%xmm10
+ DB 196,193,121,105,202 ; vpunpckhwd %xmm10,%xmm0,%xmm1
+ DB 196,195,249,22,200,1 ; vpextrq $0x1,%xmm1,%r8
+ DB 69,137,193 ; mov %r8d,%r9d
+ DB 77,137,194 ; mov %r8,%r10
DB 73,193,234,32 ; shr $0x20,%r10
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 196,227,125,25,192,1 ; vextractf128 $0x1,%ymm0,%xmm0
- DB 196,193,249,126,196 ; vmovq %xmm0,%r12
- DB 69,137,231 ; mov %r12d,%r15d
- DB 196,227,249,22,195,1 ; vpextrq $0x1,%xmm0,%rbx
- DB 65,137,221 ; mov %ebx,%r13d
- DB 72,193,235,32 ; shr $0x20,%rbx
- DB 73,193,236,32 ; shr $0x20,%r12
+ DB 196,193,249,126,200 ; vmovq %xmm1,%r8
+ DB 69,137,195 ; mov %r8d,%r11d
+ DB 77,137,198 ; mov %r8,%r14
+ DB 73,193,238,32 ; shr $0x20,%r14
+ DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
+ DB 196,225,249,126,195 ; vmovq %xmm0,%rbx
+ DB 65,137,223 ; mov %ebx,%r15d
+ DB 72,193,235,30 ; shr $0x1e,%rbx
+ DB 196,195,249,22,196,1 ; vpextrq $0x1,%xmm0,%r12
+ DB 69,137,229 ; mov %r12d,%r13d
+ DB 73,193,236,30 ; shr $0x1e,%r12
DB 72,139,104,8 ; mov 0x8(%rax),%rbp
DB 76,139,64,16 ; mov 0x10(%rax),%r8
- DB 196,161,122,16,68,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm0
- DB 196,163,121,33,68,165,0,16 ; vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
- DB 196,161,122,16,76,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm1
+ DB 196,161,122,16,68,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm0
+ DB 196,163,121,33,68,181,0,16 ; vinsertps $0x10,0x0(%rbp,%r14,4),%xmm0,%xmm0
+ DB 196,161,122,16,76,141,0 ; vmovss 0x0(%rbp,%r9,4),%xmm1
DB 196,227,121,33,193,32 ; vinsertps $0x20,%xmm1,%xmm0,%xmm0
- DB 197,250,16,76,157,0 ; vmovss 0x0(%rbp,%rbx,4),%xmm1
+ DB 196,161,122,16,76,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm1
DB 196,227,121,33,193,48 ; vinsertps $0x30,%xmm1,%xmm0,%xmm0
- DB 196,161,122,16,76,157,0 ; vmovss 0x0(%rbp,%r11,4),%xmm1
- DB 196,163,113,33,76,141,0,16 ; vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
- DB 196,161,122,16,92,181,0 ; vmovss 0x0(%rbp,%r14,4),%xmm3
+ DB 196,161,122,16,76,189,0 ; vmovss 0x0(%rbp,%r15,4),%xmm1
+ DB 196,227,113,33,76,29,0,16 ; vinsertps $0x10,0x0(%rbp,%rbx,1),%xmm1,%xmm1
+ DB 196,161,122,16,92,173,0 ; vmovss 0x0(%rbp,%r13,4),%xmm3
DB 196,227,113,33,203,32 ; vinsertps $0x20,%xmm3,%xmm1,%xmm1
- DB 196,161,122,16,92,149,0 ; vmovss 0x0(%rbp,%r10,4),%xmm3
+ DB 196,161,122,16,92,37,0 ; vmovss 0x0(%rbp,%r12,1),%xmm3
DB 196,227,113,33,203,48 ; vinsertps $0x30,%xmm3,%xmm1,%xmm1
DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
- DB 197,161,109,202 ; vpunpckhqdq %xmm2,%xmm11,%xmm1
- DB 196,193,113,105,212 ; vpunpckhwd %xmm12,%xmm1,%xmm2
- DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1
- DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
- DB 196,193,116,84,201 ; vandps %ymm9,%ymm1,%ymm1
- DB 196,193,249,126,201 ; vmovq %xmm1,%r9
- DB 69,137,203 ; mov %r9d,%r11d
- DB 196,195,249,22,202,1 ; vpextrq $0x1,%xmm1,%r10
- DB 69,137,214 ; mov %r10d,%r14d
- DB 73,193,234,32 ; shr $0x20,%r10
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 196,227,125,25,201,1 ; vextractf128 $0x1,%ymm1,%xmm1
- DB 196,225,249,126,205 ; vmovq %xmm1,%rbp
- DB 65,137,239 ; mov %ebp,%r15d
- DB 196,227,249,22,203,1 ; vpextrq $0x1,%xmm1,%rbx
- DB 65,137,220 ; mov %ebx,%r12d
- DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,193,105,219,201 ; vpand %xmm9,%xmm2,%xmm1
+ DB 196,193,113,105,210 ; vpunpckhwd %xmm10,%xmm1,%xmm2
+ 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 196,129,122,16,12,184 ; vmovss (%r8,%r15,4),%xmm1
- DB 196,195,113,33,12,168,16 ; vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1
- DB 196,129,122,16,20,160 ; vmovss (%r8,%r12,4),%xmm2
+ DB 196,225,249,126,211 ; vmovq %xmm2,%rbx
+ DB 65,137,218 ; mov %ebx,%r10d
+ DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 196,226,121,51,201 ; vpmovzxwd %xmm1,%xmm1
+ DB 196,193,249,126,203 ; vmovq %xmm1,%r11
+ DB 69,137,222 ; mov %r11d,%r14d
+ DB 73,193,235,30 ; shr $0x1e,%r11
+ DB 196,195,249,22,207,1 ; vpextrq $0x1,%xmm1,%r15
+ DB 69,137,252 ; mov %r15d,%r12d
+ DB 73,193,239,30 ; shr $0x1e,%r15
+ DB 196,129,122,16,12,144 ; vmovss (%r8,%r10,4),%xmm1
+ DB 196,195,113,33,12,152,16 ; vinsertps $0x10,(%r8,%rbx,4),%xmm1,%xmm1
+ DB 196,129,122,16,20,136 ; vmovss (%r8,%r9,4),%xmm2
DB 196,227,113,33,202,32 ; vinsertps $0x20,%xmm2,%xmm1,%xmm1
- DB 196,193,122,16,20,152 ; vmovss (%r8,%rbx,4),%xmm2
+ DB 196,193,122,16,20,168 ; vmovss (%r8,%rbp,4),%xmm2
DB 196,227,113,33,202,48 ; vinsertps $0x30,%xmm2,%xmm1,%xmm1
- DB 196,129,122,16,20,152 ; vmovss (%r8,%r11,4),%xmm2
- DB 196,131,105,33,20,136,16 ; vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2
- DB 196,129,122,16,28,176 ; vmovss (%r8,%r14,4),%xmm3
+ DB 196,129,122,16,20,176 ; vmovss (%r8,%r14,4),%xmm2
+ DB 196,131,105,33,20,24,16 ; vinsertps $0x10,(%r8,%r11,1),%xmm2,%xmm2
+ DB 196,129,122,16,28,160 ; vmovss (%r8,%r12,4),%xmm3
DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2
- DB 196,129,122,16,28,144 ; vmovss (%r8,%r10,4),%xmm3
+ DB 196,129,122,16,28,56 ; vmovss (%r8,%r15,1),%xmm3
DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2
DB 196,227,109,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm2,%ymm1
- DB 72,139,64,24 ; mov 0x18(%rax),%rax
- DB 196,193,57,108,210 ; vpunpcklqdq %xmm10,%xmm8,%xmm2
- DB 196,193,105,105,220 ; vpunpckhwd %xmm12,%xmm2,%xmm3
- DB 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2
- DB 196,227,109,24,211,1 ; vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
- DB 196,193,108,84,209 ; vandps %ymm9,%ymm2,%ymm2
- DB 196,193,249,126,208 ; vmovq %xmm2,%r8
- DB 69,137,194 ; mov %r8d,%r10d
- DB 196,195,249,22,209,1 ; vpextrq $0x1,%xmm2,%r9
- DB 69,137,203 ; mov %r9d,%r11d
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 73,193,232,32 ; shr $0x20,%r8
- DB 196,227,125,25,210,1 ; vextractf128 $0x1,%ymm2,%xmm2
- DB 196,225,249,126,213 ; vmovq %xmm2,%rbp
- DB 65,137,238 ; mov %ebp,%r14d
- DB 196,227,249,22,211,1 ; vpextrq $0x1,%xmm2,%rbx
- DB 65,137,223 ; mov %ebx,%r15d
- DB 72,193,235,32 ; shr $0x20,%rbx
+ DB 76,139,80,24 ; mov 0x18(%rax),%r10
+ DB 196,193,57,219,209 ; vpand %xmm9,%xmm8,%xmm2
+ DB 196,193,105,105,218 ; vpunpckhwd %xmm10,%xmm2,%xmm3
+ 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 196,161,122,16,20,176 ; vmovss (%rax,%r14,4),%xmm2
- DB 196,227,105,33,20,168,16 ; vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2
- DB 196,161,122,16,28,184 ; vmovss (%rax,%r15,4),%xmm3
+ 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 196,226,121,51,210 ; vpmovzxwd %xmm2,%xmm2
+ DB 196,225,249,126,208 ; vmovq %xmm2,%rax
+ DB 65,137,195 ; mov %eax,%r11d
+ DB 72,193,232,30 ; shr $0x1e,%rax
+ DB 196,195,249,22,214,1 ; vpextrq $0x1,%xmm2,%r14
+ DB 69,137,247 ; mov %r14d,%r15d
+ DB 73,193,238,30 ; shr $0x1e,%r14
+ DB 196,129,122,16,20,138 ; vmovss (%r10,%r9,4),%xmm2
+ DB 196,195,105,33,20,154,16 ; vinsertps $0x10,(%r10,%rbx,4),%xmm2,%xmm2
+ DB 196,129,122,16,28,130 ; vmovss (%r10,%r8,4),%xmm3
DB 196,227,105,33,211,32 ; vinsertps $0x20,%xmm3,%xmm2,%xmm2
- DB 197,250,16,28,152 ; vmovss (%rax,%rbx,4),%xmm3
+ DB 196,193,122,16,28,170 ; vmovss (%r10,%rbp,4),%xmm3
DB 196,99,105,33,195,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm8
- DB 196,161,122,16,28,144 ; vmovss (%rax,%r10,4),%xmm3
- DB 196,163,97,33,28,128,16 ; vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3
- DB 196,161,122,16,20,152 ; vmovss (%rax,%r11,4),%xmm2
+ DB 196,129,122,16,28,154 ; vmovss (%r10,%r11,4),%xmm3
+ DB 196,195,97,33,28,2,16 ; vinsertps $0x10,(%r10,%rax,1),%xmm3,%xmm3
+ DB 196,129,122,16,20,186 ; vmovss (%r10,%r15,4),%xmm2
DB 196,227,97,33,210,32 ; vinsertps $0x20,%xmm2,%xmm3,%xmm2
- DB 196,161,122,16,28,136 ; vmovss (%rax,%r9,4),%xmm3
+ DB 196,129,122,16,28,50 ; vmovss (%r10,%r14,1),%xmm3
DB 196,227,105,33,211,48 ; vinsertps $0x30,%xmm3,%xmm2,%xmm2
DB 196,195,109,24,208,1 ; vinsertf128 $0x1,%xmm8,%ymm2,%ymm2
DB 184,0,0,128,63 ; mov $0x3f800000,%eax
@@ -6161,36 +6348,36 @@ _sk_load_tables_rgb_u16_be_avx LABEL PROC
DB 196,129,121,110,4,72 ; vmovd (%r8,%r9,2),%xmm0
DB 196,129,121,196,68,72,4,2 ; vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 1f38 <_sk_load_tables_rgb_u16_be_avx+0x2bd>
- DB 233,137,253,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 117,5 ; jne 1ed3 <_sk_load_tables_rgb_u16_be_avx+0x294>
+ DB 233,178,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
DB 196,129,121,110,76,72,6 ; vmovd 0x6(%r8,%r9,2),%xmm1
DB 196,1,113,196,68,72,10,2 ; vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 1f67 <_sk_load_tables_rgb_u16_be_avx+0x2ec>
+ DB 114,26 ; jb 1f02 <_sk_load_tables_rgb_u16_be_avx+0x2c3>
DB 196,129,121,110,76,72,12 ; vmovd 0xc(%r8,%r9,2),%xmm1
DB 196,129,113,196,84,72,16,2 ; vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 1f6c <_sk_load_tables_rgb_u16_be_avx+0x2f1>
- DB 233,90,253,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
- DB 233,85,253,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 1f07 <_sk_load_tables_rgb_u16_be_avx+0x2c8>
+ DB 233,131,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,126,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
DB 196,129,121,110,76,72,18 ; vmovd 0x12(%r8,%r9,2),%xmm1
DB 196,1,113,196,76,72,22,2 ; vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 1f9b <_sk_load_tables_rgb_u16_be_avx+0x320>
+ DB 114,26 ; jb 1f36 <_sk_load_tables_rgb_u16_be_avx+0x2f7>
DB 196,129,121,110,76,72,24 ; vmovd 0x18(%r8,%r9,2),%xmm1
DB 196,129,113,196,76,72,28,2 ; vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 1fa0 <_sk_load_tables_rgb_u16_be_avx+0x325>
- DB 233,38,253,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
- DB 233,33,253,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 1f3b <_sk_load_tables_rgb_u16_be_avx+0x2fc>
+ DB 233,79,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,74,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
DB 196,129,121,110,92,72,30 ; vmovd 0x1e(%r8,%r9,2),%xmm3
DB 196,1,97,196,92,72,34,2 ; vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 1fc9 <_sk_load_tables_rgb_u16_be_avx+0x34e>
+ DB 114,20 ; jb 1f64 <_sk_load_tables_rgb_u16_be_avx+0x325>
DB 196,129,121,110,92,72,36 ; vmovd 0x24(%r8,%r9,2),%xmm3
DB 196,129,97,196,92,72,40,2 ; vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
- DB 233,248,252,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
- DB 233,243,252,255,255 ; jmpq 1cc1 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,33,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
+ DB 233,28,253,255,255 ; jmpq 1c85 <_sk_load_tables_rgb_u16_be_avx+0x46>
PUBLIC _sk_byte_tables_avx
_sk_byte_tables_avx LABEL PROC
@@ -6694,36 +6881,36 @@ _sk_parametric_r_avx LABEL PROC
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,216 ; vcvtdq2ps %ymm0,%ymm11
- DB 196,98,125,24,37,210,52,0,0 ; vbroadcastss 0x34d2(%rip),%ymm12 # 5d0c <_sk_callback_avx+0x122>
+ DB 196,98,125,24,37,187,52,0,0 ; vbroadcastss 0x34bb(%rip),%ymm12 # 5c90 <_sk_callback_avx+0x122>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,200,52,0,0 ; vbroadcastss 0x34c8(%rip),%ymm12 # 5d10 <_sk_callback_avx+0x126>
+ DB 196,98,125,24,37,177,52,0,0 ; vbroadcastss 0x34b1(%rip),%ymm12 # 5c94 <_sk_callback_avx+0x126>
DB 196,193,124,84,196 ; vandps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,190,52,0,0 ; vbroadcastss 0x34be(%rip),%ymm12 # 5d14 <_sk_callback_avx+0x12a>
+ DB 196,98,125,24,37,167,52,0,0 ; vbroadcastss 0x34a7(%rip),%ymm12 # 5c98 <_sk_callback_avx+0x12a>
DB 196,193,124,86,196 ; vorps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,180,52,0,0 ; vbroadcastss 0x34b4(%rip),%ymm12 # 5d18 <_sk_callback_avx+0x12e>
+ DB 196,98,125,24,37,157,52,0,0 ; vbroadcastss 0x349d(%rip),%ymm12 # 5c9c <_sk_callback_avx+0x12e>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,170,52,0,0 ; vbroadcastss 0x34aa(%rip),%ymm12 # 5d1c <_sk_callback_avx+0x132>
+ DB 196,98,125,24,37,147,52,0,0 ; vbroadcastss 0x3493(%rip),%ymm12 # 5ca0 <_sk_callback_avx+0x132>
DB 196,65,124,89,228 ; vmulps %ymm12,%ymm0,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,155,52,0,0 ; vbroadcastss 0x349b(%rip),%ymm12 # 5d20 <_sk_callback_avx+0x136>
+ DB 196,98,125,24,37,132,52,0,0 ; vbroadcastss 0x3484(%rip),%ymm12 # 5ca4 <_sk_callback_avx+0x136>
DB 196,193,124,88,196 ; vaddps %ymm12,%ymm0,%ymm0
- DB 196,98,125,24,37,145,52,0,0 ; vbroadcastss 0x3491(%rip),%ymm12 # 5d24 <_sk_callback_avx+0x13a>
+ DB 196,98,125,24,37,122,52,0,0 ; vbroadcastss 0x347a(%rip),%ymm12 # 5ca8 <_sk_callback_avx+0x13a>
DB 197,156,94,192 ; vdivps %ymm0,%ymm12,%ymm0
DB 197,164,92,192 ; vsubps %ymm0,%ymm11,%ymm0
DB 197,172,89,192 ; vmulps %ymm0,%ymm10,%ymm0
DB 196,99,125,8,208,1 ; vroundps $0x1,%ymm0,%ymm10
DB 196,65,124,92,210 ; vsubps %ymm10,%ymm0,%ymm10
- DB 196,98,125,24,29,117,52,0,0 ; vbroadcastss 0x3475(%rip),%ymm11 # 5d28 <_sk_callback_avx+0x13e>
+ DB 196,98,125,24,29,94,52,0,0 ; vbroadcastss 0x345e(%rip),%ymm11 # 5cac <_sk_callback_avx+0x13e>
DB 196,193,124,88,195 ; vaddps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,107,52,0,0 ; vbroadcastss 0x346b(%rip),%ymm11 # 5d2c <_sk_callback_avx+0x142>
+ DB 196,98,125,24,29,84,52,0,0 ; vbroadcastss 0x3454(%rip),%ymm11 # 5cb0 <_sk_callback_avx+0x142>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,124,92,195 ; vsubps %ymm11,%ymm0,%ymm0
- DB 196,98,125,24,29,92,52,0,0 ; vbroadcastss 0x345c(%rip),%ymm11 # 5d30 <_sk_callback_avx+0x146>
+ DB 196,98,125,24,29,69,52,0,0 ; vbroadcastss 0x3445(%rip),%ymm11 # 5cb4 <_sk_callback_avx+0x146>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,82,52,0,0 ; vbroadcastss 0x3452(%rip),%ymm11 # 5d34 <_sk_callback_avx+0x14a>
+ DB 196,98,125,24,29,59,52,0,0 ; vbroadcastss 0x343b(%rip),%ymm11 # 5cb8 <_sk_callback_avx+0x14a>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,124,88,194 ; vaddps %ymm10,%ymm0,%ymm0
- DB 196,98,125,24,21,67,52,0,0 ; vbroadcastss 0x3443(%rip),%ymm10 # 5d38 <_sk_callback_avx+0x14e>
+ DB 196,98,125,24,21,44,52,0,0 ; vbroadcastss 0x342c(%rip),%ymm10 # 5cbc <_sk_callback_avx+0x14e>
DB 196,193,124,89,194 ; vmulps %ymm10,%ymm0,%ymm0
DB 197,253,91,192 ; vcvtps2dq %ymm0,%ymm0
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -6754,36 +6941,36 @@ _sk_parametric_g_avx LABEL PROC
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,217 ; vcvtdq2ps %ymm1,%ymm11
- DB 196,98,125,24,37,188,51,0,0 ; vbroadcastss 0x33bc(%rip),%ymm12 # 5d3c <_sk_callback_avx+0x152>
+ DB 196,98,125,24,37,165,51,0,0 ; vbroadcastss 0x33a5(%rip),%ymm12 # 5cc0 <_sk_callback_avx+0x152>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,178,51,0,0 ; vbroadcastss 0x33b2(%rip),%ymm12 # 5d40 <_sk_callback_avx+0x156>
+ DB 196,98,125,24,37,155,51,0,0 ; vbroadcastss 0x339b(%rip),%ymm12 # 5cc4 <_sk_callback_avx+0x156>
DB 196,193,116,84,204 ; vandps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,168,51,0,0 ; vbroadcastss 0x33a8(%rip),%ymm12 # 5d44 <_sk_callback_avx+0x15a>
+ DB 196,98,125,24,37,145,51,0,0 ; vbroadcastss 0x3391(%rip),%ymm12 # 5cc8 <_sk_callback_avx+0x15a>
DB 196,193,116,86,204 ; vorps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,158,51,0,0 ; vbroadcastss 0x339e(%rip),%ymm12 # 5d48 <_sk_callback_avx+0x15e>
+ DB 196,98,125,24,37,135,51,0,0 ; vbroadcastss 0x3387(%rip),%ymm12 # 5ccc <_sk_callback_avx+0x15e>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,148,51,0,0 ; vbroadcastss 0x3394(%rip),%ymm12 # 5d4c <_sk_callback_avx+0x162>
+ DB 196,98,125,24,37,125,51,0,0 ; vbroadcastss 0x337d(%rip),%ymm12 # 5cd0 <_sk_callback_avx+0x162>
DB 196,65,116,89,228 ; vmulps %ymm12,%ymm1,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,133,51,0,0 ; vbroadcastss 0x3385(%rip),%ymm12 # 5d50 <_sk_callback_avx+0x166>
+ DB 196,98,125,24,37,110,51,0,0 ; vbroadcastss 0x336e(%rip),%ymm12 # 5cd4 <_sk_callback_avx+0x166>
DB 196,193,116,88,204 ; vaddps %ymm12,%ymm1,%ymm1
- DB 196,98,125,24,37,123,51,0,0 ; vbroadcastss 0x337b(%rip),%ymm12 # 5d54 <_sk_callback_avx+0x16a>
+ DB 196,98,125,24,37,100,51,0,0 ; vbroadcastss 0x3364(%rip),%ymm12 # 5cd8 <_sk_callback_avx+0x16a>
DB 197,156,94,201 ; vdivps %ymm1,%ymm12,%ymm1
DB 197,164,92,201 ; vsubps %ymm1,%ymm11,%ymm1
DB 197,172,89,201 ; vmulps %ymm1,%ymm10,%ymm1
DB 196,99,125,8,209,1 ; vroundps $0x1,%ymm1,%ymm10
DB 196,65,116,92,210 ; vsubps %ymm10,%ymm1,%ymm10
- DB 196,98,125,24,29,95,51,0,0 ; vbroadcastss 0x335f(%rip),%ymm11 # 5d58 <_sk_callback_avx+0x16e>
+ DB 196,98,125,24,29,72,51,0,0 ; vbroadcastss 0x3348(%rip),%ymm11 # 5cdc <_sk_callback_avx+0x16e>
DB 196,193,116,88,203 ; vaddps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,85,51,0,0 ; vbroadcastss 0x3355(%rip),%ymm11 # 5d5c <_sk_callback_avx+0x172>
+ DB 196,98,125,24,29,62,51,0,0 ; vbroadcastss 0x333e(%rip),%ymm11 # 5ce0 <_sk_callback_avx+0x172>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,116,92,203 ; vsubps %ymm11,%ymm1,%ymm1
- DB 196,98,125,24,29,70,51,0,0 ; vbroadcastss 0x3346(%rip),%ymm11 # 5d60 <_sk_callback_avx+0x176>
+ DB 196,98,125,24,29,47,51,0,0 ; vbroadcastss 0x332f(%rip),%ymm11 # 5ce4 <_sk_callback_avx+0x176>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,60,51,0,0 ; vbroadcastss 0x333c(%rip),%ymm11 # 5d64 <_sk_callback_avx+0x17a>
+ DB 196,98,125,24,29,37,51,0,0 ; vbroadcastss 0x3325(%rip),%ymm11 # 5ce8 <_sk_callback_avx+0x17a>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,116,88,202 ; vaddps %ymm10,%ymm1,%ymm1
- DB 196,98,125,24,21,45,51,0,0 ; vbroadcastss 0x332d(%rip),%ymm10 # 5d68 <_sk_callback_avx+0x17e>
+ DB 196,98,125,24,21,22,51,0,0 ; vbroadcastss 0x3316(%rip),%ymm10 # 5cec <_sk_callback_avx+0x17e>
DB 196,193,116,89,202 ; vmulps %ymm10,%ymm1,%ymm1
DB 197,253,91,201 ; vcvtps2dq %ymm1,%ymm1
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -6814,36 +7001,36 @@ _sk_parametric_b_avx LABEL PROC
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,218 ; vcvtdq2ps %ymm2,%ymm11
- DB 196,98,125,24,37,166,50,0,0 ; vbroadcastss 0x32a6(%rip),%ymm12 # 5d6c <_sk_callback_avx+0x182>
+ DB 196,98,125,24,37,143,50,0,0 ; vbroadcastss 0x328f(%rip),%ymm12 # 5cf0 <_sk_callback_avx+0x182>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,156,50,0,0 ; vbroadcastss 0x329c(%rip),%ymm12 # 5d70 <_sk_callback_avx+0x186>
+ DB 196,98,125,24,37,133,50,0,0 ; vbroadcastss 0x3285(%rip),%ymm12 # 5cf4 <_sk_callback_avx+0x186>
DB 196,193,108,84,212 ; vandps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,146,50,0,0 ; vbroadcastss 0x3292(%rip),%ymm12 # 5d74 <_sk_callback_avx+0x18a>
+ DB 196,98,125,24,37,123,50,0,0 ; vbroadcastss 0x327b(%rip),%ymm12 # 5cf8 <_sk_callback_avx+0x18a>
DB 196,193,108,86,212 ; vorps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,136,50,0,0 ; vbroadcastss 0x3288(%rip),%ymm12 # 5d78 <_sk_callback_avx+0x18e>
+ DB 196,98,125,24,37,113,50,0,0 ; vbroadcastss 0x3271(%rip),%ymm12 # 5cfc <_sk_callback_avx+0x18e>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,126,50,0,0 ; vbroadcastss 0x327e(%rip),%ymm12 # 5d7c <_sk_callback_avx+0x192>
+ DB 196,98,125,24,37,103,50,0,0 ; vbroadcastss 0x3267(%rip),%ymm12 # 5d00 <_sk_callback_avx+0x192>
DB 196,65,108,89,228 ; vmulps %ymm12,%ymm2,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,111,50,0,0 ; vbroadcastss 0x326f(%rip),%ymm12 # 5d80 <_sk_callback_avx+0x196>
+ DB 196,98,125,24,37,88,50,0,0 ; vbroadcastss 0x3258(%rip),%ymm12 # 5d04 <_sk_callback_avx+0x196>
DB 196,193,108,88,212 ; vaddps %ymm12,%ymm2,%ymm2
- DB 196,98,125,24,37,101,50,0,0 ; vbroadcastss 0x3265(%rip),%ymm12 # 5d84 <_sk_callback_avx+0x19a>
+ DB 196,98,125,24,37,78,50,0,0 ; vbroadcastss 0x324e(%rip),%ymm12 # 5d08 <_sk_callback_avx+0x19a>
DB 197,156,94,210 ; vdivps %ymm2,%ymm12,%ymm2
DB 197,164,92,210 ; vsubps %ymm2,%ymm11,%ymm2
DB 197,172,89,210 ; vmulps %ymm2,%ymm10,%ymm2
DB 196,99,125,8,210,1 ; vroundps $0x1,%ymm2,%ymm10
DB 196,65,108,92,210 ; vsubps %ymm10,%ymm2,%ymm10
- DB 196,98,125,24,29,73,50,0,0 ; vbroadcastss 0x3249(%rip),%ymm11 # 5d88 <_sk_callback_avx+0x19e>
+ DB 196,98,125,24,29,50,50,0,0 ; vbroadcastss 0x3232(%rip),%ymm11 # 5d0c <_sk_callback_avx+0x19e>
DB 196,193,108,88,211 ; vaddps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,63,50,0,0 ; vbroadcastss 0x323f(%rip),%ymm11 # 5d8c <_sk_callback_avx+0x1a2>
+ DB 196,98,125,24,29,40,50,0,0 ; vbroadcastss 0x3228(%rip),%ymm11 # 5d10 <_sk_callback_avx+0x1a2>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,108,92,211 ; vsubps %ymm11,%ymm2,%ymm2
- DB 196,98,125,24,29,48,50,0,0 ; vbroadcastss 0x3230(%rip),%ymm11 # 5d90 <_sk_callback_avx+0x1a6>
+ DB 196,98,125,24,29,25,50,0,0 ; vbroadcastss 0x3219(%rip),%ymm11 # 5d14 <_sk_callback_avx+0x1a6>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,38,50,0,0 ; vbroadcastss 0x3226(%rip),%ymm11 # 5d94 <_sk_callback_avx+0x1aa>
+ DB 196,98,125,24,29,15,50,0,0 ; vbroadcastss 0x320f(%rip),%ymm11 # 5d18 <_sk_callback_avx+0x1aa>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,108,88,210 ; vaddps %ymm10,%ymm2,%ymm2
- DB 196,98,125,24,21,23,50,0,0 ; vbroadcastss 0x3217(%rip),%ymm10 # 5d98 <_sk_callback_avx+0x1ae>
+ DB 196,98,125,24,21,0,50,0,0 ; vbroadcastss 0x3200(%rip),%ymm10 # 5d1c <_sk_callback_avx+0x1ae>
DB 196,193,108,89,210 ; vmulps %ymm10,%ymm2,%ymm2
DB 197,253,91,210 ; vcvtps2dq %ymm2,%ymm2
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -6874,36 +7061,36 @@ _sk_parametric_a_avx LABEL PROC
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
DB 196,98,125,24,16 ; vbroadcastss (%rax),%ymm10
DB 197,124,91,219 ; vcvtdq2ps %ymm3,%ymm11
- DB 196,98,125,24,37,144,49,0,0 ; vbroadcastss 0x3190(%rip),%ymm12 # 5d9c <_sk_callback_avx+0x1b2>
+ DB 196,98,125,24,37,121,49,0,0 ; vbroadcastss 0x3179(%rip),%ymm12 # 5d20 <_sk_callback_avx+0x1b2>
DB 196,65,36,89,220 ; vmulps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,134,49,0,0 ; vbroadcastss 0x3186(%rip),%ymm12 # 5da0 <_sk_callback_avx+0x1b6>
+ DB 196,98,125,24,37,111,49,0,0 ; vbroadcastss 0x316f(%rip),%ymm12 # 5d24 <_sk_callback_avx+0x1b6>
DB 196,193,100,84,220 ; vandps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,124,49,0,0 ; vbroadcastss 0x317c(%rip),%ymm12 # 5da4 <_sk_callback_avx+0x1ba>
+ DB 196,98,125,24,37,101,49,0,0 ; vbroadcastss 0x3165(%rip),%ymm12 # 5d28 <_sk_callback_avx+0x1ba>
DB 196,193,100,86,220 ; vorps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,114,49,0,0 ; vbroadcastss 0x3172(%rip),%ymm12 # 5da8 <_sk_callback_avx+0x1be>
+ DB 196,98,125,24,37,91,49,0,0 ; vbroadcastss 0x315b(%rip),%ymm12 # 5d2c <_sk_callback_avx+0x1be>
DB 196,65,36,88,220 ; vaddps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,104,49,0,0 ; vbroadcastss 0x3168(%rip),%ymm12 # 5dac <_sk_callback_avx+0x1c2>
+ DB 196,98,125,24,37,81,49,0,0 ; vbroadcastss 0x3151(%rip),%ymm12 # 5d30 <_sk_callback_avx+0x1c2>
DB 196,65,100,89,228 ; vmulps %ymm12,%ymm3,%ymm12
DB 196,65,36,92,220 ; vsubps %ymm12,%ymm11,%ymm11
- DB 196,98,125,24,37,89,49,0,0 ; vbroadcastss 0x3159(%rip),%ymm12 # 5db0 <_sk_callback_avx+0x1c6>
+ DB 196,98,125,24,37,66,49,0,0 ; vbroadcastss 0x3142(%rip),%ymm12 # 5d34 <_sk_callback_avx+0x1c6>
DB 196,193,100,88,220 ; vaddps %ymm12,%ymm3,%ymm3
- DB 196,98,125,24,37,79,49,0,0 ; vbroadcastss 0x314f(%rip),%ymm12 # 5db4 <_sk_callback_avx+0x1ca>
+ DB 196,98,125,24,37,56,49,0,0 ; vbroadcastss 0x3138(%rip),%ymm12 # 5d38 <_sk_callback_avx+0x1ca>
DB 197,156,94,219 ; vdivps %ymm3,%ymm12,%ymm3
DB 197,164,92,219 ; vsubps %ymm3,%ymm11,%ymm3
DB 197,172,89,219 ; vmulps %ymm3,%ymm10,%ymm3
DB 196,99,125,8,211,1 ; vroundps $0x1,%ymm3,%ymm10
DB 196,65,100,92,210 ; vsubps %ymm10,%ymm3,%ymm10
- DB 196,98,125,24,29,51,49,0,0 ; vbroadcastss 0x3133(%rip),%ymm11 # 5db8 <_sk_callback_avx+0x1ce>
+ DB 196,98,125,24,29,28,49,0,0 ; vbroadcastss 0x311c(%rip),%ymm11 # 5d3c <_sk_callback_avx+0x1ce>
DB 196,193,100,88,219 ; vaddps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,41,49,0,0 ; vbroadcastss 0x3129(%rip),%ymm11 # 5dbc <_sk_callback_avx+0x1d2>
+ DB 196,98,125,24,29,18,49,0,0 ; vbroadcastss 0x3112(%rip),%ymm11 # 5d40 <_sk_callback_avx+0x1d2>
DB 196,65,44,89,219 ; vmulps %ymm11,%ymm10,%ymm11
DB 196,193,100,92,219 ; vsubps %ymm11,%ymm3,%ymm3
- DB 196,98,125,24,29,26,49,0,0 ; vbroadcastss 0x311a(%rip),%ymm11 # 5dc0 <_sk_callback_avx+0x1d6>
+ DB 196,98,125,24,29,3,49,0,0 ; vbroadcastss 0x3103(%rip),%ymm11 # 5d44 <_sk_callback_avx+0x1d6>
DB 196,65,36,92,210 ; vsubps %ymm10,%ymm11,%ymm10
- DB 196,98,125,24,29,16,49,0,0 ; vbroadcastss 0x3110(%rip),%ymm11 # 5dc4 <_sk_callback_avx+0x1da>
+ DB 196,98,125,24,29,249,48,0,0 ; vbroadcastss 0x30f9(%rip),%ymm11 # 5d48 <_sk_callback_avx+0x1da>
DB 196,65,36,94,210 ; vdivps %ymm10,%ymm11,%ymm10
DB 196,193,100,88,218 ; vaddps %ymm10,%ymm3,%ymm3
- DB 196,98,125,24,21,1,49,0,0 ; vbroadcastss 0x3101(%rip),%ymm10 # 5dc8 <_sk_callback_avx+0x1de>
+ DB 196,98,125,24,21,234,48,0,0 ; vbroadcastss 0x30ea(%rip),%ymm10 # 5d4c <_sk_callback_avx+0x1de>
DB 196,193,100,89,218 ; vmulps %ymm10,%ymm3,%ymm3
DB 197,253,91,219 ; vcvtps2dq %ymm3,%ymm3
DB 196,98,125,24,80,20 ; vbroadcastss 0x14(%rax),%ymm10
@@ -7010,7 +7197,7 @@ _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 2eec <_sk_load_a8_avx+0x5a>
+ DB 117,74 ; jne 2e87 <_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
@@ -7037,9 +7224,9 @@ _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 2ef4 <_sk_load_a8_avx+0x62>
+ DB 117,234 ; jne 2e8f <_sk_load_a8_avx+0x62>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,149 ; jmp 2ea6 <_sk_load_a8_avx+0x14>
+ DB 235,149 ; jmp 2e41 <_sk_load_a8_avx+0x14>
PUBLIC _sk_gather_a8_avx
_sk_gather_a8_avx LABEL PROC
@@ -7116,7 +7303,7 @@ _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 304d <_sk_store_a8_avx+0x42>
+ DB 117,10 ; jne 2fe8 <_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
@@ -7124,10 +7311,10 @@ _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 3049 <_sk_store_a8_avx+0x3e>
+ DB 119,236 ; ja 2fe4 <_sk_store_a8_avx+0x3e>
DB 196,66,121,48,192 ; vpmovzxbw %xmm8,%xmm8
DB 65,15,182,192 ; movzbl %r8b,%eax
- DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 30b0 <_sk_store_a8_avx+0xa5>
+ DB 76,141,5,68,0,0,0 ; lea 0x44(%rip),%r8 # 304c <_sk_store_a8_avx+0xa6>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -7138,27 +7325,28 @@ _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 3049 <_sk_store_a8_avx+0x3e>
- DB 144 ; nop
- DB 246,255 ; idiv %bh
+ DB 235,154 ; jmp 2fe4 <_sk_store_a8_avx+0x3e>
+ DB 102,144 ; xchg %ax,%ax
+ DB 245 ; cmc
DB 255 ; (bad)
DB 255 ; (bad)
- DB 238 ; out %al,(%dx)
DB 255 ; (bad)
+ DB 237 ; in (%dx),%eax
DB 255 ; (bad)
- DB 255,230 ; jmpq *%rsi
DB 255 ; (bad)
+ DB 255,229 ; jmpq *%rbp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 222,255 ; fdivrp %st,%st(7)
DB 255 ; (bad)
- DB 255,214 ; callq *%rsi
+ DB 221,255 ; (bad)
DB 255 ; (bad)
+ DB 255,213 ; callq *%rbp
DB 255 ; (bad)
- DB 255,206 ; dec %esi
DB 255 ; (bad)
+ DB 255,205 ; dec %ebp
DB 255 ; (bad)
- DB 255,198 ; inc %esi
+ DB 255 ; (bad)
+ DB 255,197 ; inc %ebp
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -7170,7 +7358,7 @@ _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 3137 <_sk_load_g8_avx+0x6b>
+ DB 117,91 ; jne 30d3 <_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
@@ -7200,9 +7388,9 @@ _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 313f <_sk_load_g8_avx+0x73>
+ DB 117,234 ; jne 30db <_sk_load_g8_avx+0x73>
DB 196,193,249,110,193 ; vmovq %r9,%xmm0
- DB 235,132 ; jmp 30e0 <_sk_load_g8_avx+0x14>
+ DB 235,132 ; jmp 307c <_sk_load_g8_avx+0x14>
PUBLIC _sk_gather_g8_avx
_sk_gather_g8_avx LABEL PROC
@@ -7273,9 +7461,9 @@ _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 3276 <_sk_gather_i8_avx+0xf>
+ DB 116,5 ; je 3212 <_sk_gather_i8_avx+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 3278 <_sk_gather_i8_avx+0x11>
+ DB 235,2 ; jmp 3214 <_sk_gather_i8_avx+0x11>
DB 72,173 ; lods %ds:(%rsi),%rax
DB 65,87 ; push %r15
DB 65,86 ; push %r14
@@ -7337,27 +7525,24 @@ _sk_gather_i8_avx LABEL PROC
DB 196,163,121,34,4,163,2 ; vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0
DB 196,163,121,34,28,19,3 ; vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3
DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
- DB 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 197,124,40,21,98,42,0,0 ; vmovaps 0x2a62(%rip),%ymm10 # 5da0 <_sk_callback_avx+0x232>
+ DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%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 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1
+ DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2
+ DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
+ DB 196,193,116,84,202 ; vandps %ymm10,%ymm1,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1
- DB 196,193,41,114,208,16 ; vpsrld $0x10,%xmm8,%xmm10
+ DB 196,193,33,114,208,16 ; vpsrld $0x10,%xmm8,%xmm11
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 196,227,37,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm11,%ymm2
+ DB 196,193,108,84,210 ; vandps %ymm10,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2
DB 196,193,57,114,208,24 ; vpsrld $0x18,%xmm8,%xmm8
@@ -7378,7 +7563,7 @@ _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 3512 <_sk_load_565_avx+0xdf>
+ DB 15,133,209,0,0,0 ; jne 34a5 <_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
@@ -7428,9 +7613,9 @@ _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 3447 <_sk_load_565_avx+0x14>
+ DB 15,135,29,255,255,255 ; ja 33da <_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 # 3580 <_sk_load_565_avx+0x14d>
+ DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 3514 <_sk_load_565_avx+0x14e>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7442,28 +7627,26 @@ _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 3447 <_sk_load_565_avx+0x14>
- DB 102,144 ; xchg %ax,%ax
- DB 242,255 ; repnz (bad)
- DB 255 ; (bad)
+ DB 233,201,254,255,255 ; jmpq 33da <_sk_load_565_avx+0x14>
+ DB 15,31,0 ; nopl (%rax)
+ DB 241 ; icebp
DB 255 ; (bad)
- DB 234 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,226 ; jmpq *%rdx
+ DB 233,255,255,255,225 ; jmpq ffffffffe200351c <_sk_callback_avx+0xffffffffe1ffd9ae>
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 218,255 ; (bad)
+ DB 217,255 ; fcos
DB 255 ; (bad)
- DB 255,210 ; callq *%rdx
+ DB 255,209 ; callq *%rcx
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,202 ; dec %edx
+ DB 255,201 ; dec %ecx
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 190 ; .byte 0xbe
+ DB 189 ; .byte 0xbd
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -7595,7 +7778,7 @@ _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 37cb <_sk_store_565_avx+0x9e>
+ DB 117,10 ; jne 375f <_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
@@ -7603,9 +7786,9 @@ _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 37c7 <_sk_store_565_avx+0x9a>
+ DB 119,236 ; ja 375b <_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 # 3828 <_sk_store_565_avx+0xfb>
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 37bc <_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
@@ -7616,7 +7799,7 @@ _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 37c7 <_sk_store_565_avx+0x9a>
+ DB 235,159 ; jmp 375b <_sk_store_565_avx+0x9a>
DB 247,255 ; idiv %edi
DB 255 ; (bad)
DB 255 ; (bad)
@@ -7645,7 +7828,7 @@ _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 3947 <_sk_load_4444_avx+0x103>
+ DB 15,133,245,0,0,0 ; jne 38db <_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
@@ -7702,9 +7885,9 @@ _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 3858 <_sk_load_4444_avx+0x14>
+ DB 15,135,249,254,255,255 ; ja 37ec <_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 # 39b4 <_sk_load_4444_avx+0x170>
+ DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 3948 <_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
@@ -7716,12 +7899,12 @@ _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 3858 <_sk_load_4444_avx+0x14>
+ DB 233,165,254,255,255 ; jmpq 37ec <_sk_load_4444_avx+0x14>
DB 144 ; nop
DB 243,255 ; repz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 39b9 <_sk_load_4444_avx+0x175>
+ DB 235,255 ; jmp 394d <_sk_load_4444_avx+0x175>
DB 255 ; (bad)
DB 255,227 ; jmpq *%rbx
DB 255 ; (bad)
@@ -7878,7 +8061,7 @@ _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 3c34 <_sk_store_4444_avx+0xaf>
+ DB 117,10 ; jne 3bc8 <_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
@@ -7886,9 +8069,9 @@ _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 3c30 <_sk_store_4444_avx+0xab>
+ DB 119,236 ; ja 3bc4 <_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 # 3c94 <_sk_store_4444_avx+0x10f>
+ DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 3c28 <_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
@@ -7899,7 +8082,7 @@ _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 3c30 <_sk_store_4444_avx+0xab>
+ DB 235,159 ; jmp 3bc4 <_sk_store_4444_avx+0xab>
DB 15,31,0 ; nopl (%rax)
DB 244 ; hlt
DB 255 ; (bad)
@@ -7930,30 +8113,27 @@ _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 3d5b <_sk_load_8888_avx+0xab>
+ DB 15,133,147,0,0,0 ; jne 3ce5 <_sk_load_8888_avx+0xa1>
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
- DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0
- DB 196,99,125,24,216,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm11
- DB 196,193,36,84,193 ; vandps %ymm9,%ymm11,%ymm0
+ DB 197,124,40,21,96,33,0,0 ; vmovaps 0x2160(%rip),%ymm10 # 5dc0 <_sk_callback_avx+0x252>
+ DB 196,193,52,84,194 ; vandps %ymm10,%ymm9,%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 196,99,117,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm8
DB 196,193,124,89,192 ; vmulps %ymm8,%ymm0,%ymm0
- DB 196,193,41,114,209,8 ; vpsrld $0x8,%xmm9,%xmm10
+ DB 196,193,113,114,209,8 ; vpsrld $0x8,%xmm9,%xmm1
DB 196,99,125,25,203,1 ; vextractf128 $0x1,%ymm9,%xmm3
- 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 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2
+ DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
+ DB 196,193,116,84,202 ; vandps %ymm10,%ymm1,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,200 ; vmulps %ymm8,%ymm1,%ymm1
- DB 196,193,41,114,209,16 ; vpsrld $0x10,%xmm9,%xmm10
+ DB 196,193,33,114,209,16 ; vpsrld $0x10,%xmm9,%xmm11
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 196,227,37,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm11,%ymm2
+ DB 196,193,108,84,210 ; vandps %ymm10,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,208 ; vmulps %ymm8,%ymm2,%ymm2
DB 196,193,49,114,209,24 ; vpsrld $0x18,%xmm9,%xmm9
@@ -7968,9 +8148,9 @@ _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 3cc4 <_sk_load_8888_avx+0x14>
+ DB 15,135,90,255,255,255 ; ja 3c58 <_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 # 3e08 <_sk_load_8888_avx+0x158>
+ DB 76,141,13,139,0,0,0 ; lea 0x8b(%rip),%r9 # 3d94 <_sk_load_8888_avx+0x150>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -7993,24 +8173,26 @@ _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 3cc4 <_sk_load_8888_avx+0x14>
- DB 238 ; out %al,(%dx)
+ DB 233,198,254,255,255 ; jmpq 3c58 <_sk_load_8888_avx+0x14>
+ DB 102,144 ; xchg %ax,%ax
+ DB 236 ; in (%dx),%al
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,224 ; jmpq *%rax
DB 255 ; (bad)
+ DB 222,255 ; fdivrp %st,%st(7)
DB 255 ; (bad)
- DB 255,210 ; callq *%rdx
+ DB 255,208 ; callq *%rax
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,196 ; inc %esp
+ DB 255,194 ; inc %edx
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255,176,255,255,255,156 ; pushq -0x63000001(%rax)
+ DB 255,174,255,255,255,154 ; ljmp *-0x65000001(%rsi)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 255 ; .byte 0xff
- DB 128,255,255 ; cmp $0xff,%bh
+ DB 255 ; (bad)
+ DB 126,255 ; jle 3dad <_sk_load_8888_avx+0x169>
+ DB 255 ; (bad)
DB 255 ; .byte 0xff
PUBLIC _sk_gather_8888_avx
@@ -8052,27 +8234,24 @@ _sk_gather_8888_avx LABEL PROC
DB 196,131,121,34,4,152,2 ; vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0
DB 196,131,121,34,28,144,3 ; vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3
DB 196,227,61,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
- DB 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 197,124,40,21,126,31,0,0 ; vmovaps 0x1f7e(%rip),%ymm10 # 5de0 <_sk_callback_avx+0x272>
+ DB 196,193,124,84,194 ; vandps %ymm10,%ymm0,%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 196,193,113,114,208,8 ; vpsrld $0x8,%xmm8,%xmm1
+ DB 197,233,114,211,8 ; vpsrld $0x8,%xmm3,%xmm2
+ DB 196,227,117,24,202,1 ; vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
+ DB 196,193,116,84,202 ; vandps %ymm10,%ymm1,%ymm1
DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
DB 196,193,116,89,201 ; vmulps %ymm9,%ymm1,%ymm1
- DB 196,193,41,114,208,16 ; vpsrld $0x10,%xmm8,%xmm10
+ DB 196,193,33,114,208,16 ; vpsrld $0x10,%xmm8,%xmm11
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 196,227,37,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm11,%ymm2
+ DB 196,193,108,84,210 ; vandps %ymm10,%ymm2,%ymm2
DB 197,252,91,210 ; vcvtdq2ps %ymm2,%ymm2
DB 196,193,108,89,209 ; vmulps %ymm9,%ymm2,%ymm2
DB 196,193,57,114,208,24 ; vpsrld $0x18,%xmm8,%xmm8
@@ -8119,7 +8298,7 @@ _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 4009 <_sk_store_8888_avx+0xa4>
+ DB 117,10 ; jne 3f8c <_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
@@ -8127,9 +8306,9 @@ _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 4005 <_sk_store_8888_avx+0xa0>
+ DB 119,236 ; ja 3f88 <_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 # 4078 <_sk_store_8888_avx+0x113>
+ DB 76,141,5,85,0,0,0 ; lea 0x55(%rip),%r8 # 3ffc <_sk_store_8888_avx+0x114>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -8143,28 +8322,29 @@ _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 4005 <_sk_store_8888_avx+0xa0>
- DB 102,144 ; xchg %ax,%ax
- DB 246,255 ; idiv %bh
+ DB 235,143 ; jmp 3f88 <_sk_store_8888_avx+0xa0>
+ DB 15,31,0 ; nopl (%rax)
+ DB 245 ; cmc
DB 255 ; (bad)
DB 255 ; (bad)
- DB 238 ; out %al,(%dx)
DB 255 ; (bad)
+ DB 237 ; in (%dx),%eax
DB 255 ; (bad)
- DB 255,230 ; jmpq *%rsi
DB 255 ; (bad)
+ DB 255,229 ; jmpq *%rbp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 222,255 ; fdivrp %st,%st(7)
DB 255 ; (bad)
- DB 255,209 ; callq *%rcx
+ DB 221,255 ; (bad)
DB 255 ; (bad)
+ DB 255,208 ; callq *%rax
DB 255 ; (bad)
- DB 255,195 ; inc %ebx
+ DB 255 ; (bad)
+ DB 255,194 ; inc %edx
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
- DB 181,255 ; mov $0xff,%ch
+ DB 180,255 ; mov $0xff,%ah
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -8178,7 +8358,7 @@ _sk_load_f16_avx LABEL PROC
DB 197,252,17,116,36,64 ; vmovups %ymm6,0x40(%rsp)
DB 197,252,17,108,36,32 ; vmovups %ymm5,0x20(%rsp)
DB 197,254,127,36,36 ; vmovdqu %ymm4,(%rsp)
- DB 15,133,143,2,0,0 ; jne 434f <_sk_load_f16_avx+0x2bb>
+ DB 15,133,143,2,0,0 ; jne 42d3 <_sk_load_f16_avx+0x2bb>
DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8
DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2
DB 197,249,16,76,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm1
@@ -8196,13 +8376,13 @@ _sk_load_f16_avx LABEL PROC
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
- DB 196,98,125,24,37,179,28,0,0 ; vbroadcastss 0x1cb3(%rip),%ymm12 # 5dcc <_sk_callback_avx+0x1e2>
+ DB 196,98,125,24,37,179,28,0,0 ; vbroadcastss 0x1cb3(%rip),%ymm12 # 5d50 <_sk_callback_avx+0x1e2>
DB 196,193,124,84,204 ; vandps %ymm12,%ymm0,%ymm1
DB 197,252,87,193 ; vxorps %ymm1,%ymm0,%ymm0
DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14
- DB 196,98,121,24,29,159,28,0,0 ; vbroadcastss 0x1c9f(%rip),%xmm11 # 5dd0 <_sk_callback_avx+0x1e6>
+ DB 196,98,121,24,29,159,28,0,0 ; vbroadcastss 0x1c9f(%rip),%xmm11 # 5d54 <_sk_callback_avx+0x1e6>
DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3
- DB 196,98,121,24,45,149,28,0,0 ; vbroadcastss 0x1c95(%rip),%xmm13 # 5dd4 <_sk_callback_avx+0x1ea>
+ DB 196,98,121,24,45,149,28,0,0 ; vbroadcastss 0x1c95(%rip),%xmm13 # 5d58 <_sk_callback_avx+0x1ea>
DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3
DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10
DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10
@@ -8216,7 +8396,7 @@ _sk_load_f16_avx LABEL PROC
DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
DB 197,252,86,193 ; vorps %ymm1,%ymm0,%ymm0
DB 196,227,125,25,193,1 ; vextractf128 $0x1,%ymm0,%xmm1
- DB 196,226,121,24,29,75,28,0,0 ; vbroadcastss 0x1c4b(%rip),%xmm3 # 5dd8 <_sk_callback_avx+0x1ee>
+ DB 196,226,121,24,29,75,28,0,0 ; vbroadcastss 0x1c4b(%rip),%xmm3 # 5d5c <_sk_callback_avx+0x1ee>
DB 197,241,254,203 ; vpaddd %xmm3,%xmm1,%xmm1
DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0
DB 196,227,125,24,193,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
@@ -8309,29 +8489,29 @@ _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 43ae <_sk_load_f16_avx+0x31a>
+ DB 116,79 ; je 4332 <_sk_load_f16_avx+0x31a>
DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,67 ; jb 43ae <_sk_load_f16_avx+0x31a>
+ DB 114,67 ; jb 4332 <_sk_load_f16_avx+0x31a>
DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,68 ; je 43bb <_sk_load_f16_avx+0x327>
+ DB 116,68 ; je 433f <_sk_load_f16_avx+0x327>
DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,56 ; jb 43bb <_sk_load_f16_avx+0x327>
+ DB 114,56 ; jb 433f <_sk_load_f16_avx+0x327>
DB 197,251,16,76,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,68,253,255,255 ; je 40d7 <_sk_load_f16_avx+0x43>
+ DB 15,132,68,253,255,255 ; je 405b <_sk_load_f16_avx+0x43>
DB 197,241,22,76,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,52,253,255,255 ; jb 40d7 <_sk_load_f16_avx+0x43>
+ DB 15,130,52,253,255,255 ; jb 405b <_sk_load_f16_avx+0x43>
DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9
- DB 233,41,253,255,255 ; jmpq 40d7 <_sk_load_f16_avx+0x43>
+ DB 233,41,253,255,255 ; jmpq 405b <_sk_load_f16_avx+0x43>
DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,28,253,255,255 ; jmpq 40d7 <_sk_load_f16_avx+0x43>
+ DB 233,28,253,255,255 ; jmpq 405b <_sk_load_f16_avx+0x43>
DB 197,241,87,201 ; vxorpd %xmm1,%xmm1,%xmm1
- DB 233,19,253,255,255 ; jmpq 40d7 <_sk_load_f16_avx+0x43>
+ DB 233,19,253,255,255 ; jmpq 405b <_sk_load_f16_avx+0x43>
PUBLIC _sk_gather_f16_avx
_sk_gather_f16_avx LABEL PROC
@@ -8393,13 +8573,13 @@ _sk_gather_f16_avx LABEL PROC
DB 197,249,105,210 ; vpunpckhwd %xmm2,%xmm0,%xmm2
DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0
- DB 196,98,125,24,37,11,25,0,0 ; vbroadcastss 0x190b(%rip),%ymm12 # 5ddc <_sk_callback_avx+0x1f2>
+ DB 196,98,125,24,37,11,25,0,0 ; vbroadcastss 0x190b(%rip),%ymm12 # 5d60 <_sk_callback_avx+0x1f2>
DB 196,193,124,84,212 ; vandps %ymm12,%ymm0,%ymm2
DB 197,252,87,194 ; vxorps %ymm2,%ymm0,%ymm0
DB 196,195,125,25,198,1 ; vextractf128 $0x1,%ymm0,%xmm14
- DB 196,98,121,24,29,247,24,0,0 ; vbroadcastss 0x18f7(%rip),%xmm11 # 5de0 <_sk_callback_avx+0x1f6>
+ DB 196,98,121,24,29,247,24,0,0 ; vbroadcastss 0x18f7(%rip),%xmm11 # 5d64 <_sk_callback_avx+0x1f6>
DB 196,193,8,87,219 ; vxorps %xmm11,%xmm14,%xmm3
- DB 196,98,121,24,45,237,24,0,0 ; vbroadcastss 0x18ed(%rip),%xmm13 # 5de4 <_sk_callback_avx+0x1fa>
+ DB 196,98,121,24,45,237,24,0,0 ; vbroadcastss 0x18ed(%rip),%xmm13 # 5d68 <_sk_callback_avx+0x1fa>
DB 197,145,102,219 ; vpcmpgtd %xmm3,%xmm13,%xmm3
DB 196,65,120,87,211 ; vxorps %xmm11,%xmm0,%xmm10
DB 196,65,17,102,210 ; vpcmpgtd %xmm10,%xmm13,%xmm10
@@ -8413,7 +8593,7 @@ _sk_gather_f16_avx LABEL PROC
DB 196,227,125,24,195,1 ; vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
DB 197,252,86,194 ; vorps %ymm2,%ymm0,%ymm0
DB 196,227,125,25,194,1 ; vextractf128 $0x1,%ymm0,%xmm2
- DB 196,226,121,24,29,163,24,0,0 ; vbroadcastss 0x18a3(%rip),%xmm3 # 5de8 <_sk_callback_avx+0x1fe>
+ DB 196,226,121,24,29,163,24,0,0 ; vbroadcastss 0x18a3(%rip),%xmm3 # 5d6c <_sk_callback_avx+0x1fe>
DB 197,233,254,211 ; vpaddd %xmm3,%xmm2,%xmm2
DB 197,249,254,195 ; vpaddd %xmm3,%xmm0,%xmm0
DB 196,227,125,24,194,1 ; vinsertf128 $0x1,%xmm2,%ymm0,%ymm0
@@ -8515,12 +8695,12 @@ _sk_store_f16_avx LABEL PROC
DB 197,252,17,180,36,128,0,0,0 ; vmovups %ymm6,0x80(%rsp)
DB 197,252,17,108,36,96 ; vmovups %ymm5,0x60(%rsp)
DB 197,252,17,100,36,64 ; vmovups %ymm4,0x40(%rsp)
- DB 196,98,125,24,13,176,22,0,0 ; vbroadcastss 0x16b0(%rip),%ymm9 # 5dec <_sk_callback_avx+0x202>
+ DB 196,98,125,24,13,176,22,0,0 ; vbroadcastss 0x16b0(%rip),%ymm9 # 5d70 <_sk_callback_avx+0x202>
DB 196,65,124,84,209 ; vandps %ymm9,%ymm0,%ymm10
DB 197,252,17,4,36 ; vmovups %ymm0,(%rsp)
DB 196,65,124,87,218 ; vxorps %ymm10,%ymm0,%ymm11
DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12
- DB 196,98,121,24,5,150,22,0,0 ; vbroadcastss 0x1696(%rip),%xmm8 # 5df0 <_sk_callback_avx+0x206>
+ DB 196,98,121,24,5,150,22,0,0 ; vbroadcastss 0x1696(%rip),%xmm8 # 5d74 <_sk_callback_avx+0x206>
DB 196,65,57,102,236 ; vpcmpgtd %xmm12,%xmm8,%xmm13
DB 196,65,57,102,243 ; vpcmpgtd %xmm11,%xmm8,%xmm14
DB 196,67,13,24,237,1 ; vinsertf128 $0x1,%xmm13,%ymm14,%ymm13
@@ -8530,7 +8710,7 @@ _sk_store_f16_avx LABEL PROC
DB 196,67,13,24,242,1 ; vinsertf128 $0x1,%xmm10,%ymm14,%ymm14
DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11
DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12
- DB 196,98,125,24,21,93,22,0,0 ; vbroadcastss 0x165d(%rip),%ymm10 # 5df4 <_sk_callback_avx+0x20a>
+ DB 196,98,125,24,21,93,22,0,0 ; vbroadcastss 0x165d(%rip),%ymm10 # 5d78 <_sk_callback_avx+0x20a>
DB 196,65,12,86,242 ; vorps %ymm10,%ymm14,%ymm14
DB 196,67,125,25,247,1 ; vextractf128 $0x1,%ymm14,%xmm15
DB 196,65,1,254,228 ; vpaddd %xmm12,%xmm15,%xmm12
@@ -8612,7 +8792,7 @@ _sk_store_f16_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 117,75 ; jne 497e <_sk_store_f16_avx+0x270>
+ DB 117,75 ; jne 4902 <_sk_store_f16_avx+0x270>
DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8)
DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8)
DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -8628,22 +8808,22 @@ _sk_store_f16_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,193 ; je 494a <_sk_store_f16_avx+0x23c>
+ DB 116,193 ; je 48ce <_sk_store_f16_avx+0x23c>
DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,181 ; jb 494a <_sk_store_f16_avx+0x23c>
+ DB 114,181 ; jb 48ce <_sk_store_f16_avx+0x23c>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,173 ; je 494a <_sk_store_f16_avx+0x23c>
+ DB 116,173 ; je 48ce <_sk_store_f16_avx+0x23c>
DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,161 ; jb 494a <_sk_store_f16_avx+0x23c>
+ DB 114,161 ; jb 48ce <_sk_store_f16_avx+0x23c>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,153 ; je 494a <_sk_store_f16_avx+0x23c>
+ DB 116,153 ; je 48ce <_sk_store_f16_avx+0x23c>
DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,141 ; jb 494a <_sk_store_f16_avx+0x23c>
+ DB 114,141 ; jb 48ce <_sk_store_f16_avx+0x23c>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,133 ; jmp 494a <_sk_store_f16_avx+0x23c>
+ DB 235,133 ; jmp 48ce <_sk_store_f16_avx+0x23c>
PUBLIC _sk_load_u16_be_avx
_sk_load_u16_be_avx LABEL PROC
@@ -8651,7 +8831,7 @@ _sk_load_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,189,0,0,0,0 ; lea 0x0(,%rdi,4),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,5,1,0,0 ; jne 4ae0 <_sk_load_u16_be_avx+0x11b>
+ DB 15,133,5,1,0,0 ; jne 4a64 <_sk_load_u16_be_avx+0x11b>
DB 196,65,121,16,4,64 ; vmovupd (%r8,%rax,2),%xmm8
DB 196,193,121,16,84,64,16 ; vmovupd 0x10(%r8,%rax,2),%xmm2
DB 196,193,121,16,92,64,32 ; vmovupd 0x20(%r8,%rax,2),%xmm3
@@ -8710,29 +8890,29 @@ _sk_load_u16_be_avx LABEL PROC
DB 196,65,123,16,4,64 ; vmovsd (%r8,%rax,2),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,85 ; je 4b46 <_sk_load_u16_be_avx+0x181>
+ DB 116,85 ; je 4aca <_sk_load_u16_be_avx+0x181>
DB 196,65,57,22,68,64,8 ; vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,72 ; jb 4b46 <_sk_load_u16_be_avx+0x181>
+ DB 114,72 ; jb 4aca <_sk_load_u16_be_avx+0x181>
DB 196,193,123,16,84,64,16 ; vmovsd 0x10(%r8,%rax,2),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,72 ; je 4b53 <_sk_load_u16_be_avx+0x18e>
+ DB 116,72 ; je 4ad7 <_sk_load_u16_be_avx+0x18e>
DB 196,193,105,22,84,64,24 ; vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,59 ; jb 4b53 <_sk_load_u16_be_avx+0x18e>
+ DB 114,59 ; jb 4ad7 <_sk_load_u16_be_avx+0x18e>
DB 196,193,123,16,92,64,32 ; vmovsd 0x20(%r8,%rax,2),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,205,254,255,255 ; je 49f6 <_sk_load_u16_be_avx+0x31>
+ DB 15,132,205,254,255,255 ; je 497a <_sk_load_u16_be_avx+0x31>
DB 196,193,97,22,92,64,40 ; vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,188,254,255,255 ; jb 49f6 <_sk_load_u16_be_avx+0x31>
+ DB 15,130,188,254,255,255 ; jb 497a <_sk_load_u16_be_avx+0x31>
DB 196,65,122,126,76,64,48 ; vmovq 0x30(%r8,%rax,2),%xmm9
- DB 233,176,254,255,255 ; jmpq 49f6 <_sk_load_u16_be_avx+0x31>
+ DB 233,176,254,255,255 ; jmpq 497a <_sk_load_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,163,254,255,255 ; jmpq 49f6 <_sk_load_u16_be_avx+0x31>
+ DB 233,163,254,255,255 ; jmpq 497a <_sk_load_u16_be_avx+0x31>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,154,254,255,255 ; jmpq 49f6 <_sk_load_u16_be_avx+0x31>
+ DB 233,154,254,255,255 ; jmpq 497a <_sk_load_u16_be_avx+0x31>
PUBLIC _sk_load_rgb_u16_be_avx
_sk_load_rgb_u16_be_avx LABEL PROC
@@ -8740,7 +8920,7 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 76,139,0 ; mov (%rax),%r8
DB 72,141,4,127 ; lea (%rdi,%rdi,2),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,8,1,0,0 ; jne 4c76 <_sk_load_rgb_u16_be_avx+0x11a>
+ DB 15,133,8,1,0,0 ; jne 4bfa <_sk_load_rgb_u16_be_avx+0x11a>
DB 196,193,122,111,4,64 ; vmovdqu (%r8,%rax,2),%xmm0
DB 196,193,122,111,84,64,12 ; vmovdqu 0xc(%r8,%rax,2),%xmm2
DB 196,193,122,111,76,64,24 ; vmovdqu 0x18(%r8,%rax,2),%xmm1
@@ -8799,36 +8979,36 @@ _sk_load_rgb_u16_be_avx LABEL PROC
DB 196,193,121,110,4,64 ; vmovd (%r8,%rax,2),%xmm0
DB 196,193,121,196,68,64,4,2 ; vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 117,5 ; jne 4c8f <_sk_load_rgb_u16_be_avx+0x133>
- DB 233,19,255,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,5 ; jne 4c13 <_sk_load_rgb_u16_be_avx+0x133>
+ DB 233,19,255,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,76,64,6 ; vmovd 0x6(%r8,%rax,2),%xmm1
DB 196,65,113,196,68,64,10,2 ; vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,26 ; jb 4cbe <_sk_load_rgb_u16_be_avx+0x162>
+ DB 114,26 ; jb 4c42 <_sk_load_rgb_u16_be_avx+0x162>
DB 196,193,121,110,76,64,12 ; vmovd 0xc(%r8,%rax,2),%xmm1
DB 196,193,113,196,84,64,16,2 ; vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 117,10 ; jne 4cc3 <_sk_load_rgb_u16_be_avx+0x167>
- DB 233,228,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,223,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 4c47 <_sk_load_rgb_u16_be_avx+0x167>
+ DB 233,228,254,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,223,254,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,76,64,18 ; vmovd 0x12(%r8,%rax,2),%xmm1
DB 196,65,113,196,76,64,22,2 ; vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,26 ; jb 4cf2 <_sk_load_rgb_u16_be_avx+0x196>
+ DB 114,26 ; jb 4c76 <_sk_load_rgb_u16_be_avx+0x196>
DB 196,193,121,110,76,64,24 ; vmovd 0x18(%r8,%rax,2),%xmm1
DB 196,193,113,196,76,64,28,2 ; vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 117,10 ; jne 4cf7 <_sk_load_rgb_u16_be_avx+0x19b>
- DB 233,176,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,171,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 117,10 ; jne 4c7b <_sk_load_rgb_u16_be_avx+0x19b>
+ DB 233,176,254,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,171,254,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
DB 196,193,121,110,92,64,30 ; vmovd 0x1e(%r8,%rax,2),%xmm3
DB 196,65,97,196,92,64,34,2 ; vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,20 ; jb 4d20 <_sk_load_rgb_u16_be_avx+0x1c4>
+ DB 114,20 ; jb 4ca4 <_sk_load_rgb_u16_be_avx+0x1c4>
DB 196,193,121,110,92,64,36 ; vmovd 0x24(%r8,%rax,2),%xmm3
DB 196,193,97,196,92,64,40,2 ; vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
- DB 233,130,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
- DB 233,125,254,255,255 ; jmpq 4ba2 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,130,254,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
+ DB 233,125,254,255,255 ; jmpq 4b26 <_sk_load_rgb_u16_be_avx+0x46>
PUBLIC _sk_store_u16_be_avx
_sk_store_u16_be_avx LABEL PROC
@@ -8876,7 +9056,7 @@ _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 4e27 <_sk_store_u16_be_avx+0x102>
+ DB 117,31 ; jne 4dab <_sk_store_u16_be_avx+0x102>
DB 196,1,120,17,28,72 ; vmovups %xmm11,(%r8,%r9,2)
DB 196,1,120,17,84,72,16 ; vmovups %xmm10,0x10(%r8,%r9,2)
DB 196,1,120,17,76,72,32 ; vmovups %xmm9,0x20(%r8,%r9,2)
@@ -8885,31 +9065,31 @@ _sk_store_u16_be_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,1,121,214,28,72 ; vmovq %xmm11,(%r8,%r9,2)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 116,240 ; je 4da7 <_sk_store_u16_be_avx+0xfe>
DB 196,1,121,23,92,72,8 ; vmovhpd %xmm11,0x8(%r8,%r9,2)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 114,227 ; jb 4da7 <_sk_store_u16_be_avx+0xfe>
DB 196,1,121,214,84,72,16 ; vmovq %xmm10,0x10(%r8,%r9,2)
- DB 116,218 ; je 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 116,218 ; je 4da7 <_sk_store_u16_be_avx+0xfe>
DB 196,1,121,23,84,72,24 ; vmovhpd %xmm10,0x18(%r8,%r9,2)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 114,205 ; jb 4da7 <_sk_store_u16_be_avx+0xfe>
DB 196,1,121,214,76,72,32 ; vmovq %xmm9,0x20(%r8,%r9,2)
- DB 116,196 ; je 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 116,196 ; je 4da7 <_sk_store_u16_be_avx+0xfe>
DB 196,1,121,23,76,72,40 ; vmovhpd %xmm9,0x28(%r8,%r9,2)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,183 ; jb 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 114,183 ; jb 4da7 <_sk_store_u16_be_avx+0xfe>
DB 196,1,121,214,68,72,48 ; vmovq %xmm8,0x30(%r8,%r9,2)
- DB 235,174 ; jmp 4e23 <_sk_store_u16_be_avx+0xfe>
+ DB 235,174 ; jmp 4da7 <_sk_store_u16_be_avx+0xfe>
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 4eeb <_sk_load_f32_avx+0x76>
+ DB 119,110 ; ja 4e6f <_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 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 4f14 <_sk_load_f32_avx+0x9f>
+ DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 4e98 <_sk_load_f32_avx+0x9f>
DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax
DB 76,1,208 ; add %r10,%rax
DB 255,224 ; jmpq *%rax
@@ -8965,7 +9145,7 @@ _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 4fa1 <_sk_store_f32_avx+0x6d>
+ DB 117,55 ; jne 4f25 <_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
@@ -8978,22 +9158,22 @@ _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 4f9d <_sk_store_f32_avx+0x69>
+ DB 116,240 ; je 4f21 <_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 4f9d <_sk_store_f32_avx+0x69>
+ DB 114,227 ; jb 4f21 <_sk_store_f32_avx+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 4f9d <_sk_store_f32_avx+0x69>
+ DB 116,218 ; je 4f21 <_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 4f9d <_sk_store_f32_avx+0x69>
+ DB 114,205 ; jb 4f21 <_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 4f9d <_sk_store_f32_avx+0x69>
+ DB 116,195 ; je 4f21 <_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 4f9d <_sk_store_f32_avx+0x69>
+ DB 114,181 ; jb 4f21 <_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 4f9d <_sk_store_f32_avx+0x69>
+ DB 235,171 ; jmp 4f21 <_sk_store_f32_avx+0x69>
PUBLIC _sk_clamp_x_avx
_sk_clamp_x_avx LABEL PROC
@@ -9297,7 +9477,7 @@ _sk_linear_gradient_avx LABEL PROC
DB 196,226,125,24,88,28 ; vbroadcastss 0x1c(%rax),%ymm3
DB 76,139,0 ; mov (%rax),%r8
DB 77,133,192 ; test %r8,%r8
- DB 15,132,146,0,0,0 ; je 5555 <_sk_linear_gradient_avx+0xb8>
+ DB 15,132,146,0,0,0 ; je 54d9 <_sk_linear_gradient_avx+0xb8>
DB 72,139,64,8 ; mov 0x8(%rax),%rax
DB 72,131,192,32 ; add $0x20,%rax
DB 196,65,28,87,228 ; vxorps %ymm12,%ymm12,%ymm12
@@ -9324,8 +9504,8 @@ _sk_linear_gradient_avx LABEL PROC
DB 196,227,13,74,219,208 ; vblendvps %ymm13,%ymm3,%ymm14,%ymm3
DB 72,131,192,36 ; add $0x24,%rax
DB 73,255,200 ; dec %r8
- DB 117,140 ; jne 54df <_sk_linear_gradient_avx+0x42>
- DB 235,20 ; jmp 5569 <_sk_linear_gradient_avx+0xcc>
+ DB 117,140 ; jne 5463 <_sk_linear_gradient_avx+0x42>
+ DB 235,20 ; jmp 54ed <_sk_linear_gradient_avx+0xcc>
DB 196,65,36,87,219 ; vxorps %ymm11,%ymm11,%ymm11
DB 196,65,44,87,210 ; vxorps %ymm10,%ymm10,%ymm10
DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9
@@ -9783,10 +9963,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 5d14 <.literal4+0x10>
+ DB 127,0 ; jg 5c98 <.literal4+0x10>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 5d8d <.literal4+0x89>
+ DB 119,115 ; ja 5d11 <.literal4+0x89>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9800,10 +9980,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 5d44 <.literal4+0x40>
+ DB 127,0 ; jg 5cc8 <.literal4+0x40>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 5dbd <.literal4+0xb9>
+ DB 119,115 ; ja 5d41 <.literal4+0xb9>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9817,10 +9997,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 5d74 <.literal4+0x70>
+ DB 127,0 ; jg 5cf8 <.literal4+0x70>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 5ded <.literal4+0xe9>
+ DB 119,115 ; ja 5d71 <.literal4+0xe9>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9834,10 +10014,10 @@ ALIGN 4
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 5da4 <.literal4+0xa0>
+ DB 127,0 ; jg 5d28 <.literal4+0xa0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 5e1d <_sk_callback_avx+0x233>
+ DB 119,115 ; ja 5da1 <_sk_callback_avx+0x233>
DB 248 ; clc
DB 194,117,191 ; retq $0xbf75
DB 191,63,249,68,180 ; mov $0xb444f93f,%edi
@@ -9861,6 +10041,90 @@ ALIGN 4
DB 0,128,0,0,128,56 ; add %al,0x38800000(%rax)
DB 0,64,254 ; add %al,-0x2(%rax)
DB 255 ; .byte 0xff
+
+ALIGN 32
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+
+ALIGN 16
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
ALIGN 32
PUBLIC _sk_start_pipeline_sse41
@@ -9942,7 +10206,7 @@ _sk_seed_shader_sse41 LABEL PROC
DB 102,15,110,199 ; movd %edi,%xmm0
DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
- DB 15,40,21,193,62,0,0 ; movaps 0x3ec1(%rip),%xmm2 # 3fd0 <_sk_callback_sse41+0xb4>
+ DB 15,40,21,209,62,0,0 ; movaps 0x3ed1(%rip),%xmm2 # 3fe0 <_sk_callback_sse41+0xb8>
DB 15,88,202 ; addps %xmm2,%xmm1
DB 15,16,2 ; movups (%rdx),%xmm0
DB 15,88,193 ; addps %xmm1,%xmm0
@@ -9951,7 +10215,7 @@ _sk_seed_shader_sse41 LABEL PROC
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 15,88,202 ; addps %xmm2,%xmm1
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,21,176,62,0,0 ; movaps 0x3eb0(%rip),%xmm2 # 3fe0 <_sk_callback_sse41+0xc4>
+ DB 15,40,21,192,62,0,0 ; movaps 0x3ec0(%rip),%xmm2 # 3ff0 <_sk_callback_sse41+0xc8>
DB 15,87,219 ; xorps %xmm3,%xmm3
DB 15,87,228 ; xorps %xmm4,%xmm4
DB 15,87,237 ; xorps %xmm5,%xmm5
@@ -11463,56 +11727,51 @@ _sk_lerp_565_sse41 LABEL PROC
PUBLIC _sk_load_tables_sse41
_sk_load_tables_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,8 ; mov (%rax),%rcx
- DB 76,139,64,8 ; mov 0x8(%rax),%r8
- DB 243,68,15,111,4,185 ; movdqu (%rcx,%rdi,4),%xmm8
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 102,15,110,193 ; movd %ecx,%xmm0
- DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
- DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1
- DB 102,15,114,209,8 ; psrld $0x8,%xmm1
- DB 102,15,219,200 ; pand %xmm0,%xmm1
- DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2
- DB 102,15,114,210,16 ; psrld $0x10,%xmm2
- DB 102,15,219,208 ; pand %xmm0,%xmm2
+ DB 76,139,0 ; mov (%rax),%r8
+ DB 76,139,72,8 ; mov 0x8(%rax),%r9
+ DB 243,69,15,111,4,184 ; movdqu (%r8,%rdi,4),%xmm8
+ DB 102,15,111,5,232,41,0,0 ; movdqa 0x29e8(%rip),%xmm0 # 4000 <_sk_callback_sse41+0xd8>
DB 102,65,15,219,192 ; pand %xmm8,%xmm0
- DB 102,72,15,58,22,193,1 ; pextrq $0x1,%xmm0,%rcx
- DB 65,137,201 ; mov %ecx,%r9d
- DB 72,193,233,32 ; shr $0x20,%rcx
- DB 102,73,15,126,194 ; movq %xmm0,%r10
- DB 69,137,211 ; mov %r10d,%r11d
- DB 73,193,234,32 ; shr $0x20,%r10
- DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0
- DB 102,67,15,58,33,4,144,16 ; insertps $0x10,(%r8,%r10,4),%xmm0
- DB 102,67,15,58,33,4,136,32 ; insertps $0x20,(%r8,%r9,4),%xmm0
- DB 102,65,15,58,33,4,136,48 ; insertps $0x30,(%r8,%rcx,4),%xmm0
+ DB 102,73,15,58,22,192,1 ; pextrq $0x1,%xmm0,%r8
+ DB 102,72,15,126,193 ; movq %xmm0,%rcx
+ DB 68,15,182,209 ; movzbl %cl,%r10d
+ DB 72,193,233,30 ; shr $0x1e,%rcx
+ DB 69,15,182,216 ; movzbl %r8b,%r11d
+ DB 73,193,232,30 ; shr $0x1e,%r8
+ DB 243,67,15,16,4,145 ; movss (%r9,%r10,4),%xmm0
+ DB 102,65,15,58,33,4,9,16 ; insertps $0x10,(%r9,%rcx,1),%xmm0
+ DB 102,67,15,58,33,4,153,32 ; insertps $0x20,(%r9,%r11,4),%xmm0
+ DB 243,67,15,16,12,1 ; movss (%r9,%r8,1),%xmm1
+ DB 102,15,58,33,193,48 ; insertps $0x30,%xmm1,%xmm0
DB 76,139,64,16 ; mov 0x10(%rax),%r8
- DB 102,73,15,58,22,202,1 ; pextrq $0x1,%xmm1,%r10
- DB 77,137,209 ; mov %r10,%r9
- DB 73,193,233,32 ; shr $0x20,%r9
+ DB 102,65,15,111,200 ; movdqa %xmm8,%xmm1
+ DB 102,15,56,0,13,163,41,0,0 ; pshufb 0x29a3(%rip),%xmm1 # 4010 <_sk_callback_sse41+0xe8>
+ DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9
DB 102,72,15,126,201 ; movq %xmm1,%rcx
- DB 65,137,203 ; mov %ecx,%r11d
- DB 65,129,227,255,255,255,0 ; and $0xffffff,%r11d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 65,129,226,255,255,255,0 ; and $0xffffff,%r10d
- DB 243,67,15,16,12,152 ; movss (%r8,%r11,4),%xmm1
+ DB 69,15,182,217 ; movzbl %r9b,%r11d
+ DB 73,193,233,30 ; shr $0x1e,%r9
+ DB 243,67,15,16,12,144 ; movss (%r8,%r10,4),%xmm1
DB 102,65,15,58,33,12,8,16 ; insertps $0x10,(%r8,%rcx,1),%xmm1
- DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3
- DB 102,15,58,33,203,32 ; insertps $0x20,%xmm3,%xmm1
- DB 243,67,15,16,28,136 ; movss (%r8,%r9,4),%xmm3
- DB 102,15,58,33,203,48 ; insertps $0x30,%xmm3,%xmm1
- DB 76,139,72,24 ; mov 0x18(%rax),%r9
+ DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2
+ DB 102,15,58,33,202,32 ; insertps $0x20,%xmm2,%xmm1
+ DB 243,67,15,16,20,8 ; movss (%r8,%r9,1),%xmm2
+ DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1
+ DB 76,139,64,24 ; mov 0x18(%rax),%r8
+ DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2
+ DB 102,15,56,0,21,95,41,0,0 ; pshufb 0x295f(%rip),%xmm2 # 4020 <_sk_callback_sse41+0xf8>
DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx
- DB 68,15,183,193 ; movzwl %cx,%r8d
- DB 72,193,233,32 ; shr $0x20,%rcx
DB 102,72,15,126,208 ; movq %xmm2,%rax
- DB 68,15,183,208 ; movzwl %ax,%r10d
+ DB 68,15,182,200 ; movzbl %al,%r9d
DB 72,193,232,30 ; shr $0x1e,%rax
- DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2
- DB 102,65,15,58,33,20,1,16 ; insertps $0x10,(%r9,%rax,1),%xmm2
- DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3
+ DB 68,15,182,209 ; movzbl %cl,%r10d
+ DB 72,193,233,30 ; shr $0x1e,%rcx
+ DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2
+ DB 102,65,15,58,33,20,0,16 ; insertps $0x10,(%r8,%rax,1),%xmm2
+ DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3
DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2
- DB 243,65,15,16,28,137 ; movss (%r9,%rcx,4),%xmm3
+ DB 243,65,15,16,28,8 ; movss (%r8,%rcx,1),%xmm3
DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2
DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8
DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
@@ -11530,36 +11789,36 @@ _sk_load_tables_u16_be_sse41 LABEL PROC
DB 76,139,64,8 ; mov 0x8(%rax),%r8
DB 243,15,111,4,249 ; movdqu (%rcx,%rdi,8),%xmm0
DB 243,15,111,76,249,16 ; movdqu 0x10(%rcx,%rdi,8),%xmm1
- DB 102,15,111,216 ; movdqa %xmm0,%xmm3
- DB 102,15,97,217 ; punpcklwd %xmm1,%xmm3
+ DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9
+ DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9
DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0
- DB 102,15,111,203 ; movdqa %xmm3,%xmm1
+ DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1
DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1
- DB 102,15,105,216 ; punpckhwd %xmm0,%xmm3
- DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 102,15,110,209 ; movd %ecx,%xmm2
- DB 102,68,15,112,194,0 ; pshufd $0x0,%xmm2,%xmm8
+ DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9
+ DB 102,68,15,111,5,213,40,0,0 ; movdqa 0x28d5(%rip),%xmm8 # 4030 <_sk_callback_sse41+0x108>
+ DB 102,15,111,193 ; movdqa %xmm1,%xmm0
DB 102,65,15,219,192 ; pand %xmm8,%xmm0
+ DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0
DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9
DB 102,72,15,126,193 ; movq %xmm0,%rcx
- DB 65,137,202 ; mov %ecx,%r10d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 69,137,203 ; mov %r9d,%r11d
+ DB 69,15,182,217 ; movzbl %r9b,%r11d
DB 73,193,233,30 ; shr $0x1e,%r9
DB 243,67,15,16,4,144 ; movss (%r8,%r10,4),%xmm0
DB 102,65,15,58,33,4,8,16 ; insertps $0x10,(%r8,%rcx,1),%xmm0
- DB 102,67,15,58,33,4,152,32 ; insertps $0x20,(%r8,%r11,4),%xmm0
- DB 102,67,15,58,33,4,8,48 ; insertps $0x30,(%r8,%r9,1),%xmm0
+ DB 243,67,15,16,20,152 ; movss (%r8,%r11,4),%xmm2
+ DB 102,15,58,33,194,32 ; insertps $0x20,%xmm2,%xmm0
+ DB 243,67,15,16,20,8 ; movss (%r8,%r9,1),%xmm2
+ DB 102,15,58,33,194,48 ; insertps $0x30,%xmm2,%xmm0
DB 76,139,64,16 ; mov 0x10(%rax),%r8
- DB 102,15,115,217,8 ; psrldq $0x8,%xmm1
+ DB 102,15,56,0,13,136,40,0,0 ; pshufb 0x2888(%rip),%xmm1 # 4040 <_sk_callback_sse41+0x118>
DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1
- DB 102,65,15,219,200 ; pand %xmm8,%xmm1
DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9
DB 102,72,15,126,201 ; movq %xmm1,%rcx
- DB 65,137,202 ; mov %ecx,%r10d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 69,137,203 ; mov %r9d,%r11d
+ DB 69,15,182,217 ; movzbl %r9b,%r11d
DB 73,193,233,30 ; shr $0x1e,%r9
DB 243,67,15,16,12,144 ; movss (%r8,%r10,4),%xmm1
DB 102,65,15,58,33,12,8,16 ; insertps $0x10,(%r8,%rcx,1),%xmm1
@@ -11568,22 +11827,24 @@ _sk_load_tables_u16_be_sse41 LABEL PROC
DB 243,67,15,16,20,8 ; movss (%r8,%r9,1),%xmm2
DB 102,15,58,33,202,48 ; insertps $0x30,%xmm2,%xmm1
DB 76,139,64,24 ; mov 0x18(%rax),%r8
- DB 102,15,56,51,211 ; pmovzxwd %xmm3,%xmm2
- DB 102,65,15,219,208 ; pand %xmm8,%xmm2
+ DB 102,69,15,219,193 ; pand %xmm9,%xmm8
+ DB 102,65,15,56,51,208 ; pmovzxwd %xmm8,%xmm2
DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx
DB 102,72,15,126,208 ; movq %xmm2,%rax
- DB 65,137,193 ; mov %eax,%r9d
+ DB 68,15,182,200 ; movzbl %al,%r9d
DB 72,193,232,30 ; shr $0x1e,%rax
- DB 65,137,202 ; mov %ecx,%r10d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2
DB 102,65,15,58,33,20,0,16 ; insertps $0x10,(%r8,%rax,1),%xmm2
- DB 102,67,15,58,33,20,144,32 ; insertps $0x20,(%r8,%r10,4),%xmm2
- DB 102,65,15,58,33,20,8,48 ; insertps $0x30,(%r8,%rcx,1),%xmm2
+ DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3
+ DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2
+ DB 243,65,15,16,28,8 ; movss (%r8,%rcx,1),%xmm3
+ DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2
DB 184,128,0,128,55 ; mov $0x37800080,%eax
DB 102,68,15,110,192 ; movd %eax,%xmm8
DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8
- DB 102,15,112,219,78 ; pshufd $0x4e,%xmm3,%xmm3
+ DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3
DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9
DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9
DB 102,15,113,211,8 ; psrlw $0x8,%xmm3
@@ -11603,38 +11864,37 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC
DB 243,66,15,111,20,73 ; movdqu (%rcx,%r9,2),%xmm2
DB 243,66,15,111,68,73,8 ; movdqu 0x8(%rcx,%r9,2),%xmm0
DB 102,15,115,216,4 ; psrldq $0x4,%xmm0
- DB 102,15,111,202 ; movdqa %xmm2,%xmm1
- DB 102,15,115,217,6 ; psrldq $0x6,%xmm1
+ DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9
+ DB 102,65,15,115,217,6 ; psrldq $0x6,%xmm9
DB 102,15,97,208 ; punpcklwd %xmm0,%xmm2
DB 102,15,115,216,6 ; psrldq $0x6,%xmm0
- DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1
- DB 102,15,111,194 ; movdqa %xmm2,%xmm0
- DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0
- DB 102,68,15,112,192,78 ; pshufd $0x4e,%xmm0,%xmm8
- DB 102,15,105,209 ; punpckhwd %xmm1,%xmm2
+ DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9
+ DB 102,15,111,202 ; movdqa %xmm2,%xmm1
+ DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1
+ DB 102,68,15,111,5,124,39,0,0 ; movdqa 0x277c(%rip),%xmm8 # 4050 <_sk_callback_sse41+0x128>
+ DB 102,15,111,193 ; movdqa %xmm1,%xmm0
+ DB 102,65,15,219,192 ; pand %xmm8,%xmm0
DB 102,15,56,51,192 ; pmovzxwd %xmm0,%xmm0
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 102,15,110,201 ; movd %ecx,%xmm1
- DB 102,68,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm9
- DB 102,65,15,219,193 ; pand %xmm9,%xmm0
DB 102,73,15,58,22,193,1 ; pextrq $0x1,%xmm0,%r9
DB 102,72,15,126,193 ; movq %xmm0,%rcx
- DB 65,137,202 ; mov %ecx,%r10d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 69,137,203 ; mov %r9d,%r11d
+ DB 69,15,182,217 ; movzbl %r9b,%r11d
DB 73,193,233,30 ; shr $0x1e,%r9
DB 243,67,15,16,4,144 ; movss (%r8,%r10,4),%xmm0
DB 102,65,15,58,33,4,8,16 ; insertps $0x10,(%r8,%rcx,1),%xmm0
- DB 102,67,15,58,33,4,152,32 ; insertps $0x20,(%r8,%r11,4),%xmm0
- DB 102,67,15,58,33,4,8,48 ; insertps $0x30,(%r8,%r9,1),%xmm0
+ DB 243,67,15,16,28,152 ; movss (%r8,%r11,4),%xmm3
+ DB 102,15,58,33,195,32 ; insertps $0x20,%xmm3,%xmm0
+ DB 243,67,15,16,28,8 ; movss (%r8,%r9,1),%xmm3
+ DB 102,15,58,33,195,48 ; insertps $0x30,%xmm3,%xmm0
DB 76,139,64,16 ; mov 0x10(%rax),%r8
- DB 102,65,15,56,51,200 ; pmovzxwd %xmm8,%xmm1
- DB 102,65,15,219,201 ; pand %xmm9,%xmm1
+ DB 102,15,56,0,13,47,39,0,0 ; pshufb 0x272f(%rip),%xmm1 # 4060 <_sk_callback_sse41+0x138>
+ DB 102,15,56,51,201 ; pmovzxwd %xmm1,%xmm1
DB 102,73,15,58,22,201,1 ; pextrq $0x1,%xmm1,%r9
DB 102,72,15,126,201 ; movq %xmm1,%rcx
- DB 65,137,202 ; mov %ecx,%r10d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 69,137,203 ; mov %r9d,%r11d
+ DB 69,15,182,217 ; movzbl %r9b,%r11d
DB 73,193,233,30 ; shr $0x1e,%r9
DB 243,67,15,16,12,144 ; movss (%r8,%r10,4),%xmm1
DB 102,65,15,58,33,12,8,16 ; insertps $0x10,(%r8,%rcx,1),%xmm1
@@ -11643,18 +11903,21 @@ _sk_load_tables_rgb_u16_be_sse41 LABEL PROC
DB 243,67,15,16,28,8 ; movss (%r8,%r9,1),%xmm3
DB 102,15,58,33,203,48 ; insertps $0x30,%xmm3,%xmm1
DB 76,139,64,24 ; mov 0x18(%rax),%r8
+ DB 102,65,15,105,209 ; punpckhwd %xmm9,%xmm2
+ DB 102,65,15,219,208 ; pand %xmm8,%xmm2
DB 102,15,56,51,210 ; pmovzxwd %xmm2,%xmm2
- DB 102,65,15,219,209 ; pand %xmm9,%xmm2
DB 102,72,15,58,22,209,1 ; pextrq $0x1,%xmm2,%rcx
DB 102,72,15,126,208 ; movq %xmm2,%rax
- DB 65,137,193 ; mov %eax,%r9d
+ DB 68,15,182,200 ; movzbl %al,%r9d
DB 72,193,232,30 ; shr $0x1e,%rax
- DB 65,137,202 ; mov %ecx,%r10d
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2
DB 102,65,15,58,33,20,0,16 ; insertps $0x10,(%r8,%rax,1),%xmm2
- DB 102,67,15,58,33,20,144,32 ; insertps $0x20,(%r8,%r10,4),%xmm2
- DB 102,65,15,58,33,20,8,48 ; insertps $0x30,(%r8,%rcx,1),%xmm2
+ DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3
+ DB 102,15,58,33,211,32 ; insertps $0x20,%xmm3,%xmm2
+ DB 243,65,15,16,28,8 ; movss (%r8,%rcx,1),%xmm3
+ DB 102,15,58,33,211,48 ; insertps $0x30,%xmm3,%xmm2
DB 184,0,0,128,63 ; mov $0x3f800000,%eax
DB 102,15,110,216 ; movd %eax,%xmm3
DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
@@ -11948,31 +12211,31 @@ _sk_parametric_r_sse41 LABEL PROC
DB 69,15,88,208 ; addps %xmm8,%xmm10
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
DB 69,15,91,194 ; cvtdq2ps %xmm10,%xmm8
- DB 68,15,89,5,111,33,0,0 ; mulps 0x216f(%rip),%xmm8 # 3ff0 <_sk_callback_sse41+0xd4>
- DB 68,15,84,21,119,33,0,0 ; andps 0x2177(%rip),%xmm10 # 4000 <_sk_callback_sse41+0xe4>
- DB 68,15,86,21,127,33,0,0 ; orps 0x217f(%rip),%xmm10 # 4010 <_sk_callback_sse41+0xf4>
- DB 68,15,88,5,135,33,0,0 ; addps 0x2187(%rip),%xmm8 # 4020 <_sk_callback_sse41+0x104>
- DB 68,15,40,37,143,33,0,0 ; movaps 0x218f(%rip),%xmm12 # 4030 <_sk_callback_sse41+0x114>
+ DB 68,15,89,5,204,33,0,0 ; mulps 0x21cc(%rip),%xmm8 # 4070 <_sk_callback_sse41+0x148>
+ DB 68,15,84,21,212,33,0,0 ; andps 0x21d4(%rip),%xmm10 # 4080 <_sk_callback_sse41+0x158>
+ DB 68,15,86,21,220,33,0,0 ; orps 0x21dc(%rip),%xmm10 # 4090 <_sk_callback_sse41+0x168>
+ DB 68,15,88,5,228,33,0,0 ; addps 0x21e4(%rip),%xmm8 # 40a0 <_sk_callback_sse41+0x178>
+ DB 68,15,40,37,236,33,0,0 ; movaps 0x21ec(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x188>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,196 ; subps %xmm12,%xmm8
- DB 68,15,88,21,143,33,0,0 ; addps 0x218f(%rip),%xmm10 # 4040 <_sk_callback_sse41+0x124>
- DB 68,15,40,37,151,33,0,0 ; movaps 0x2197(%rip),%xmm12 # 4050 <_sk_callback_sse41+0x134>
+ DB 68,15,88,21,236,33,0,0 ; addps 0x21ec(%rip),%xmm10 # 40c0 <_sk_callback_sse41+0x198>
+ DB 68,15,40,37,244,33,0,0 ; movaps 0x21f4(%rip),%xmm12 # 40d0 <_sk_callback_sse41+0x1a8>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,196 ; subps %xmm12,%xmm8
DB 69,15,89,195 ; mulps %xmm11,%xmm8
DB 102,69,15,58,8,208,1 ; roundps $0x1,%xmm8,%xmm10
DB 69,15,40,216 ; movaps %xmm8,%xmm11
DB 69,15,92,218 ; subps %xmm10,%xmm11
- DB 68,15,88,5,132,33,0,0 ; addps 0x2184(%rip),%xmm8 # 4060 <_sk_callback_sse41+0x144>
- DB 68,15,40,21,140,33,0,0 ; movaps 0x218c(%rip),%xmm10 # 4070 <_sk_callback_sse41+0x154>
+ DB 68,15,88,5,225,33,0,0 ; addps 0x21e1(%rip),%xmm8 # 40e0 <_sk_callback_sse41+0x1b8>
+ DB 68,15,40,21,233,33,0,0 ; movaps 0x21e9(%rip),%xmm10 # 40f0 <_sk_callback_sse41+0x1c8>
DB 69,15,89,211 ; mulps %xmm11,%xmm10
DB 69,15,92,194 ; subps %xmm10,%xmm8
- DB 68,15,40,21,140,33,0,0 ; movaps 0x218c(%rip),%xmm10 # 4080 <_sk_callback_sse41+0x164>
+ DB 68,15,40,21,233,33,0,0 ; movaps 0x21e9(%rip),%xmm10 # 4100 <_sk_callback_sse41+0x1d8>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 68,15,40,29,144,33,0,0 ; movaps 0x2190(%rip),%xmm11 # 4090 <_sk_callback_sse41+0x174>
+ DB 68,15,40,29,237,33,0,0 ; movaps 0x21ed(%rip),%xmm11 # 4110 <_sk_callback_sse41+0x1e8>
DB 69,15,94,218 ; divps %xmm10,%xmm11
DB 69,15,88,216 ; addps %xmm8,%xmm11
- DB 68,15,89,29,144,33,0,0 ; mulps 0x2190(%rip),%xmm11 # 40a0 <_sk_callback_sse41+0x184>
+ DB 68,15,89,29,237,33,0,0 ; mulps 0x21ed(%rip),%xmm11 # 4120 <_sk_callback_sse41+0x1f8>
DB 102,69,15,91,211 ; cvtps2dq %xmm11,%xmm10
DB 243,68,15,16,64,20 ; movss 0x14(%rax),%xmm8
DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8
@@ -12011,31 +12274,31 @@ _sk_parametric_g_sse41 LABEL PROC
DB 68,15,88,217 ; addps %xmm1,%xmm11
DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12
- DB 68,15,89,37,255,32,0,0 ; mulps 0x20ff(%rip),%xmm12 # 40b0 <_sk_callback_sse41+0x194>
- DB 68,15,84,29,7,33,0,0 ; andps 0x2107(%rip),%xmm11 # 40c0 <_sk_callback_sse41+0x1a4>
- DB 68,15,86,29,15,33,0,0 ; orps 0x210f(%rip),%xmm11 # 40d0 <_sk_callback_sse41+0x1b4>
- DB 68,15,88,37,23,33,0,0 ; addps 0x2117(%rip),%xmm12 # 40e0 <_sk_callback_sse41+0x1c4>
- DB 15,40,13,32,33,0,0 ; movaps 0x2120(%rip),%xmm1 # 40f0 <_sk_callback_sse41+0x1d4>
+ DB 68,15,89,37,92,33,0,0 ; mulps 0x215c(%rip),%xmm12 # 4130 <_sk_callback_sse41+0x208>
+ DB 68,15,84,29,100,33,0,0 ; andps 0x2164(%rip),%xmm11 # 4140 <_sk_callback_sse41+0x218>
+ DB 68,15,86,29,108,33,0,0 ; orps 0x216c(%rip),%xmm11 # 4150 <_sk_callback_sse41+0x228>
+ DB 68,15,88,37,116,33,0,0 ; addps 0x2174(%rip),%xmm12 # 4160 <_sk_callback_sse41+0x238>
+ DB 15,40,13,125,33,0,0 ; movaps 0x217d(%rip),%xmm1 # 4170 <_sk_callback_sse41+0x248>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
- DB 68,15,88,29,32,33,0,0 ; addps 0x2120(%rip),%xmm11 # 4100 <_sk_callback_sse41+0x1e4>
- DB 15,40,13,41,33,0,0 ; movaps 0x2129(%rip),%xmm1 # 4110 <_sk_callback_sse41+0x1f4>
+ DB 68,15,88,29,125,33,0,0 ; addps 0x217d(%rip),%xmm11 # 4180 <_sk_callback_sse41+0x258>
+ DB 15,40,13,134,33,0,0 ; movaps 0x2186(%rip),%xmm1 # 4190 <_sk_callback_sse41+0x268>
DB 65,15,94,203 ; divps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10
DB 69,15,40,220 ; movaps %xmm12,%xmm11
DB 69,15,92,218 ; subps %xmm10,%xmm11
- DB 68,15,88,37,22,33,0,0 ; addps 0x2116(%rip),%xmm12 # 4120 <_sk_callback_sse41+0x204>
- DB 15,40,13,31,33,0,0 ; movaps 0x211f(%rip),%xmm1 # 4130 <_sk_callback_sse41+0x214>
+ DB 68,15,88,37,115,33,0,0 ; addps 0x2173(%rip),%xmm12 # 41a0 <_sk_callback_sse41+0x278>
+ DB 15,40,13,124,33,0,0 ; movaps 0x217c(%rip),%xmm1 # 41b0 <_sk_callback_sse41+0x288>
DB 65,15,89,203 ; mulps %xmm11,%xmm1
DB 68,15,92,225 ; subps %xmm1,%xmm12
- DB 68,15,40,21,31,33,0,0 ; movaps 0x211f(%rip),%xmm10 # 4140 <_sk_callback_sse41+0x224>
+ DB 68,15,40,21,124,33,0,0 ; movaps 0x217c(%rip),%xmm10 # 41c0 <_sk_callback_sse41+0x298>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,13,36,33,0,0 ; movaps 0x2124(%rip),%xmm1 # 4150 <_sk_callback_sse41+0x234>
+ DB 15,40,13,129,33,0,0 ; movaps 0x2181(%rip),%xmm1 # 41d0 <_sk_callback_sse41+0x2a8>
DB 65,15,94,202 ; divps %xmm10,%xmm1
DB 65,15,88,204 ; addps %xmm12,%xmm1
- DB 15,89,13,37,33,0,0 ; mulps 0x2125(%rip),%xmm1 # 4160 <_sk_callback_sse41+0x244>
+ DB 15,89,13,130,33,0,0 ; mulps 0x2182(%rip),%xmm1 # 41e0 <_sk_callback_sse41+0x2b8>
DB 102,68,15,91,209 ; cvtps2dq %xmm1,%xmm10
DB 243,15,16,72,20 ; movss 0x14(%rax),%xmm1
DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1
@@ -12074,31 +12337,31 @@ _sk_parametric_b_sse41 LABEL PROC
DB 68,15,88,218 ; addps %xmm2,%xmm11
DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12
- DB 68,15,89,37,152,32,0,0 ; mulps 0x2098(%rip),%xmm12 # 4170 <_sk_callback_sse41+0x254>
- DB 68,15,84,29,160,32,0,0 ; andps 0x20a0(%rip),%xmm11 # 4180 <_sk_callback_sse41+0x264>
- DB 68,15,86,29,168,32,0,0 ; orps 0x20a8(%rip),%xmm11 # 4190 <_sk_callback_sse41+0x274>
- DB 68,15,88,37,176,32,0,0 ; addps 0x20b0(%rip),%xmm12 # 41a0 <_sk_callback_sse41+0x284>
- DB 15,40,21,185,32,0,0 ; movaps 0x20b9(%rip),%xmm2 # 41b0 <_sk_callback_sse41+0x294>
+ DB 68,15,89,37,245,32,0,0 ; mulps 0x20f5(%rip),%xmm12 # 41f0 <_sk_callback_sse41+0x2c8>
+ DB 68,15,84,29,253,32,0,0 ; andps 0x20fd(%rip),%xmm11 # 4200 <_sk_callback_sse41+0x2d8>
+ DB 68,15,86,29,5,33,0,0 ; orps 0x2105(%rip),%xmm11 # 4210 <_sk_callback_sse41+0x2e8>
+ DB 68,15,88,37,13,33,0,0 ; addps 0x210d(%rip),%xmm12 # 4220 <_sk_callback_sse41+0x2f8>
+ DB 15,40,21,22,33,0,0 ; movaps 0x2116(%rip),%xmm2 # 4230 <_sk_callback_sse41+0x308>
DB 65,15,89,211 ; mulps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
- DB 68,15,88,29,185,32,0,0 ; addps 0x20b9(%rip),%xmm11 # 41c0 <_sk_callback_sse41+0x2a4>
- DB 15,40,21,194,32,0,0 ; movaps 0x20c2(%rip),%xmm2 # 41d0 <_sk_callback_sse41+0x2b4>
+ DB 68,15,88,29,22,33,0,0 ; addps 0x2116(%rip),%xmm11 # 4240 <_sk_callback_sse41+0x318>
+ DB 15,40,21,31,33,0,0 ; movaps 0x211f(%rip),%xmm2 # 4250 <_sk_callback_sse41+0x328>
DB 65,15,94,211 ; divps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10
DB 69,15,40,220 ; movaps %xmm12,%xmm11
DB 69,15,92,218 ; subps %xmm10,%xmm11
- DB 68,15,88,37,175,32,0,0 ; addps 0x20af(%rip),%xmm12 # 41e0 <_sk_callback_sse41+0x2c4>
- DB 15,40,21,184,32,0,0 ; movaps 0x20b8(%rip),%xmm2 # 41f0 <_sk_callback_sse41+0x2d4>
+ DB 68,15,88,37,12,33,0,0 ; addps 0x210c(%rip),%xmm12 # 4260 <_sk_callback_sse41+0x338>
+ DB 15,40,21,21,33,0,0 ; movaps 0x2115(%rip),%xmm2 # 4270 <_sk_callback_sse41+0x348>
DB 65,15,89,211 ; mulps %xmm11,%xmm2
DB 68,15,92,226 ; subps %xmm2,%xmm12
- DB 68,15,40,21,184,32,0,0 ; movaps 0x20b8(%rip),%xmm10 # 4200 <_sk_callback_sse41+0x2e4>
+ DB 68,15,40,21,21,33,0,0 ; movaps 0x2115(%rip),%xmm10 # 4280 <_sk_callback_sse41+0x358>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,21,189,32,0,0 ; movaps 0x20bd(%rip),%xmm2 # 4210 <_sk_callback_sse41+0x2f4>
+ DB 15,40,21,26,33,0,0 ; movaps 0x211a(%rip),%xmm2 # 4290 <_sk_callback_sse41+0x368>
DB 65,15,94,210 ; divps %xmm10,%xmm2
DB 65,15,88,212 ; addps %xmm12,%xmm2
- DB 15,89,21,190,32,0,0 ; mulps 0x20be(%rip),%xmm2 # 4220 <_sk_callback_sse41+0x304>
+ DB 15,89,21,27,33,0,0 ; mulps 0x211b(%rip),%xmm2 # 42a0 <_sk_callback_sse41+0x378>
DB 102,68,15,91,210 ; cvtps2dq %xmm2,%xmm10
DB 243,15,16,80,20 ; movss 0x14(%rax),%xmm2
DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2
@@ -12137,31 +12400,31 @@ _sk_parametric_a_sse41 LABEL PROC
DB 68,15,88,219 ; addps %xmm3,%xmm11
DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
DB 69,15,91,227 ; cvtdq2ps %xmm11,%xmm12
- DB 68,15,89,37,49,32,0,0 ; mulps 0x2031(%rip),%xmm12 # 4230 <_sk_callback_sse41+0x314>
- DB 68,15,84,29,57,32,0,0 ; andps 0x2039(%rip),%xmm11 # 4240 <_sk_callback_sse41+0x324>
- DB 68,15,86,29,65,32,0,0 ; orps 0x2041(%rip),%xmm11 # 4250 <_sk_callback_sse41+0x334>
- DB 68,15,88,37,73,32,0,0 ; addps 0x2049(%rip),%xmm12 # 4260 <_sk_callback_sse41+0x344>
- DB 15,40,29,82,32,0,0 ; movaps 0x2052(%rip),%xmm3 # 4270 <_sk_callback_sse41+0x354>
+ DB 68,15,89,37,142,32,0,0 ; mulps 0x208e(%rip),%xmm12 # 42b0 <_sk_callback_sse41+0x388>
+ DB 68,15,84,29,150,32,0,0 ; andps 0x2096(%rip),%xmm11 # 42c0 <_sk_callback_sse41+0x398>
+ DB 68,15,86,29,158,32,0,0 ; orps 0x209e(%rip),%xmm11 # 42d0 <_sk_callback_sse41+0x3a8>
+ DB 68,15,88,37,166,32,0,0 ; addps 0x20a6(%rip),%xmm12 # 42e0 <_sk_callback_sse41+0x3b8>
+ DB 15,40,29,175,32,0,0 ; movaps 0x20af(%rip),%xmm3 # 42f0 <_sk_callback_sse41+0x3c8>
DB 65,15,89,219 ; mulps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
- DB 68,15,88,29,82,32,0,0 ; addps 0x2052(%rip),%xmm11 # 4280 <_sk_callback_sse41+0x364>
- DB 15,40,29,91,32,0,0 ; movaps 0x205b(%rip),%xmm3 # 4290 <_sk_callback_sse41+0x374>
+ DB 68,15,88,29,175,32,0,0 ; addps 0x20af(%rip),%xmm11 # 4300 <_sk_callback_sse41+0x3d8>
+ DB 15,40,29,184,32,0,0 ; movaps 0x20b8(%rip),%xmm3 # 4310 <_sk_callback_sse41+0x3e8>
DB 65,15,94,219 ; divps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 102,69,15,58,8,212,1 ; roundps $0x1,%xmm12,%xmm10
DB 69,15,40,220 ; movaps %xmm12,%xmm11
DB 69,15,92,218 ; subps %xmm10,%xmm11
- DB 68,15,88,37,72,32,0,0 ; addps 0x2048(%rip),%xmm12 # 42a0 <_sk_callback_sse41+0x384>
- DB 15,40,29,81,32,0,0 ; movaps 0x2051(%rip),%xmm3 # 42b0 <_sk_callback_sse41+0x394>
+ DB 68,15,88,37,165,32,0,0 ; addps 0x20a5(%rip),%xmm12 # 4320 <_sk_callback_sse41+0x3f8>
+ DB 15,40,29,174,32,0,0 ; movaps 0x20ae(%rip),%xmm3 # 4330 <_sk_callback_sse41+0x408>
DB 65,15,89,219 ; mulps %xmm11,%xmm3
DB 68,15,92,227 ; subps %xmm3,%xmm12
- DB 68,15,40,21,81,32,0,0 ; movaps 0x2051(%rip),%xmm10 # 42c0 <_sk_callback_sse41+0x3a4>
+ DB 68,15,40,21,174,32,0,0 ; movaps 0x20ae(%rip),%xmm10 # 4340 <_sk_callback_sse41+0x418>
DB 69,15,92,211 ; subps %xmm11,%xmm10
- DB 15,40,29,86,32,0,0 ; movaps 0x2056(%rip),%xmm3 # 42d0 <_sk_callback_sse41+0x3b4>
+ DB 15,40,29,179,32,0,0 ; movaps 0x20b3(%rip),%xmm3 # 4350 <_sk_callback_sse41+0x428>
DB 65,15,94,218 ; divps %xmm10,%xmm3
DB 65,15,88,220 ; addps %xmm12,%xmm3
- DB 15,89,29,87,32,0,0 ; mulps 0x2057(%rip),%xmm3 # 42e0 <_sk_callback_sse41+0x3c4>
+ DB 15,89,29,180,32,0,0 ; mulps 0x20b4(%rip),%xmm3 # 4360 <_sk_callback_sse41+0x438>
DB 102,68,15,91,211 ; cvtps2dq %xmm3,%xmm10
DB 243,15,16,88,20 ; movss 0x14(%rax),%xmm3
DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
@@ -12381,9 +12644,9 @@ _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 25c0 <_sk_gather_i8_sse41+0xf>
+ DB 116,5 ; je 25e3 <_sk_gather_i8_sse41+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 25c2 <_sk_gather_i8_sse41+0x11>
+ DB 235,2 ; jmp 25e5 <_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
@@ -12414,23 +12677,19 @@ _sk_gather_i8_sse41 LABEL PROC
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 102,15,111,5,235,28,0,0 ; movdqa 0x1ceb(%rip),%xmm0 # 4370 <_sk_callback_sse41+0x448>
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 102,15,111,203 ; movdqa %xmm3,%xmm1
+ DB 102,15,56,0,13,212,28,0,0 ; pshufb 0x1cd4(%rip),%xmm1 # 4380 <_sk_callback_sse41+0x458>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 102,15,111,211 ; movdqa %xmm3,%xmm2
+ DB 102,15,56,0,21,208,28,0,0 ; pshufb 0x1cd0(%rip),%xmm2 # 4390 <_sk_callback_sse41+0x468>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12698,24 +12957,20 @@ PUBLIC _sk_load_8888_sse41
_sk_load_8888_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
- DB 243,15,111,28,184 ; movdqu (%rax,%rdi,4),%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 15,16,28,184 ; movups (%rax,%rdi,4),%xmm3
+ DB 15,40,5,194,24,0,0 ; movaps 0x18c2(%rip),%xmm0 # 43a0 <_sk_callback_sse41+0x478>
+ DB 15,84,195 ; andps %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 15,40,203 ; movaps %xmm3,%xmm1
+ DB 102,15,56,0,13,173,24,0,0 ; pshufb 0x18ad(%rip),%xmm1 # 43b0 <_sk_callback_sse41+0x488>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 15,40,211 ; movaps %xmm3,%xmm2
+ DB 102,15,56,0,21,170,24,0,0 ; pshufb 0x18aa(%rip),%xmm2 # 43c0 <_sk_callback_sse41+0x498>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12744,23 +12999,19 @@ _sk_gather_8888_sse41 LABEL PROC
DB 102,65,15,58,34,28,129,1 ; pinsrd $0x1,(%r9,%rax,4),%xmm3
DB 102,67,15,58,34,28,145,2 ; pinsrd $0x2,(%r9,%r10,4),%xmm3
DB 102,65,15,58,34,28,137,3 ; pinsrd $0x3,(%r9,%rcx,4),%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 102,15,111,5,67,24,0,0 ; movdqa 0x1843(%rip),%xmm0 # 43d0 <_sk_callback_sse41+0x4a8>
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 102,15,111,203 ; movdqa %xmm3,%xmm1
+ DB 102,15,56,0,13,44,24,0,0 ; pshufb 0x182c(%rip),%xmm1 # 43e0 <_sk_callback_sse41+0x4b8>
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
+ DB 102,15,111,211 ; movdqa %xmm3,%xmm2
+ DB 102,15,56,0,21,40,24,0,0 ; pshufb 0x1828(%rip),%xmm2 # 43f0 <_sk_callback_sse41+0x4c8>
DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
DB 65,15,89,208 ; mulps %xmm8,%xmm2
DB 102,15,114,211,24 ; psrld $0x18,%xmm3
@@ -12810,18 +13061,18 @@ _sk_load_f16_sse41 LABEL PROC
DB 102,68,15,97,216 ; punpcklwd %xmm0,%xmm11
DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9
DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1
- DB 102,68,15,111,5,114,22,0,0 ; movdqa 0x1672(%rip),%xmm8 # 42f0 <_sk_callback_sse41+0x3d4>
+ DB 102,68,15,111,5,118,23,0,0 ; movdqa 0x1776(%rip),%xmm8 # 4400 <_sk_callback_sse41+0x4d8>
DB 102,15,111,209 ; movdqa %xmm1,%xmm2
DB 102,65,15,219,208 ; pand %xmm8,%xmm2
DB 102,15,239,202 ; pxor %xmm2,%xmm1
- DB 102,15,111,29,109,22,0,0 ; movdqa 0x166d(%rip),%xmm3 # 4300 <_sk_callback_sse41+0x3e4>
+ DB 102,15,111,29,113,23,0,0 ; movdqa 0x1771(%rip),%xmm3 # 4410 <_sk_callback_sse41+0x4e8>
DB 102,15,114,242,16 ; pslld $0x10,%xmm2
DB 102,15,111,193 ; movdqa %xmm1,%xmm0
DB 102,15,56,63,195 ; pmaxud %xmm3,%xmm0
DB 102,15,118,193 ; pcmpeqd %xmm1,%xmm0
DB 102,15,114,241,13 ; pslld $0xd,%xmm1
DB 102,15,235,202 ; por %xmm2,%xmm1
- DB 102,68,15,111,21,89,22,0,0 ; movdqa 0x1659(%rip),%xmm10 # 4310 <_sk_callback_sse41+0x3f4>
+ DB 102,68,15,111,21,93,23,0,0 ; movdqa 0x175d(%rip),%xmm10 # 4420 <_sk_callback_sse41+0x4f8>
DB 102,65,15,254,202 ; paddd %xmm10,%xmm1
DB 102,15,219,193 ; pand %xmm1,%xmm0
DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11
@@ -12892,18 +13143,18 @@ _sk_gather_f16_sse41 LABEL PROC
DB 102,68,15,97,218 ; punpcklwd %xmm2,%xmm11
DB 102,68,15,105,202 ; punpckhwd %xmm2,%xmm9
DB 102,65,15,56,51,203 ; pmovzxwd %xmm11,%xmm1
- DB 102,68,15,111,5,23,21,0,0 ; movdqa 0x1517(%rip),%xmm8 # 4320 <_sk_callback_sse41+0x404>
+ DB 102,68,15,111,5,27,22,0,0 ; movdqa 0x161b(%rip),%xmm8 # 4430 <_sk_callback_sse41+0x508>
DB 102,15,111,209 ; movdqa %xmm1,%xmm2
DB 102,65,15,219,208 ; pand %xmm8,%xmm2
DB 102,15,239,202 ; pxor %xmm2,%xmm1
- DB 102,15,111,29,18,21,0,0 ; movdqa 0x1512(%rip),%xmm3 # 4330 <_sk_callback_sse41+0x414>
+ DB 102,15,111,29,22,22,0,0 ; movdqa 0x1616(%rip),%xmm3 # 4440 <_sk_callback_sse41+0x518>
DB 102,15,114,242,16 ; pslld $0x10,%xmm2
DB 102,15,111,193 ; movdqa %xmm1,%xmm0
DB 102,15,56,63,195 ; pmaxud %xmm3,%xmm0
DB 102,15,118,193 ; pcmpeqd %xmm1,%xmm0
DB 102,15,114,241,13 ; pslld $0xd,%xmm1
DB 102,15,235,202 ; por %xmm2,%xmm1
- DB 102,68,15,111,21,254,20,0,0 ; movdqa 0x14fe(%rip),%xmm10 # 4340 <_sk_callback_sse41+0x424>
+ DB 102,68,15,111,21,2,22,0,0 ; movdqa 0x1602(%rip),%xmm10 # 4450 <_sk_callback_sse41+0x528>
DB 102,65,15,254,202 ; paddd %xmm10,%xmm1
DB 102,15,219,193 ; pand %xmm1,%xmm0
DB 102,65,15,115,219,8 ; psrldq $0x8,%xmm11
@@ -12949,17 +13200,17 @@ PUBLIC _sk_store_f16_sse41
_sk_store_f16_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
- DB 102,68,15,111,21,52,20,0,0 ; movdqa 0x1434(%rip),%xmm10 # 4350 <_sk_callback_sse41+0x434>
+ DB 102,68,15,111,21,56,21,0,0 ; movdqa 0x1538(%rip),%xmm10 # 4460 <_sk_callback_sse41+0x538>
DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12
DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13
DB 102,69,15,219,234 ; pand %xmm10,%xmm13
DB 102,69,15,239,229 ; pxor %xmm13,%xmm12
- DB 102,68,15,111,13,39,20,0,0 ; movdqa 0x1427(%rip),%xmm9 # 4360 <_sk_callback_sse41+0x444>
+ DB 102,68,15,111,13,43,21,0,0 ; movdqa 0x152b(%rip),%xmm9 # 4470 <_sk_callback_sse41+0x548>
DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13
DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8
DB 102,69,15,102,196 ; pcmpgtd %xmm12,%xmm8
DB 102,65,15,114,212,13 ; psrld $0xd,%xmm12
- DB 102,68,15,111,29,24,20,0,0 ; movdqa 0x1418(%rip),%xmm11 # 4370 <_sk_callback_sse41+0x454>
+ DB 102,68,15,111,29,28,21,0,0 ; movdqa 0x151c(%rip),%xmm11 # 4480 <_sk_callback_sse41+0x558>
DB 102,69,15,235,235 ; por %xmm11,%xmm13
DB 102,69,15,254,236 ; paddd %xmm12,%xmm13
DB 102,69,15,223,197 ; pandn %xmm13,%xmm8
@@ -13548,7 +13799,7 @@ _sk_linear_gradient_sse41 LABEL PROC
DB 69,15,198,237,0 ; shufps $0x0,%xmm13,%xmm13
DB 72,139,8 ; mov (%rax),%rcx
DB 72,133,201 ; test %rcx,%rcx
- DB 15,132,4,1,0,0 ; je 399e <_sk_linear_gradient_sse41+0x13e>
+ DB 15,132,4,1,0,0 ; je 39aa <_sk_linear_gradient_sse41+0x13e>
DB 72,131,236,88 ; sub $0x58,%rsp
DB 15,41,36,36 ; movaps %xmm4,(%rsp)
DB 15,41,108,36,16 ; movaps %xmm5,0x10(%rsp)
@@ -13599,13 +13850,13 @@ _sk_linear_gradient_sse41 LABEL PROC
DB 15,40,196 ; movaps %xmm4,%xmm0
DB 72,131,192,36 ; add $0x24,%rax
DB 72,255,201 ; dec %rcx
- DB 15,133,65,255,255,255 ; jne 38c6 <_sk_linear_gradient_sse41+0x66>
+ DB 15,133,65,255,255,255 ; jne 38d2 <_sk_linear_gradient_sse41+0x66>
DB 15,40,124,36,48 ; movaps 0x30(%rsp),%xmm7
DB 15,40,116,36,32 ; movaps 0x20(%rsp),%xmm6
DB 15,40,108,36,16 ; movaps 0x10(%rsp),%xmm5
DB 15,40,36,36 ; movaps (%rsp),%xmm4
DB 72,131,196,88 ; add $0x58,%rsp
- DB 235,13 ; jmp 39ab <_sk_linear_gradient_sse41+0x14b>
+ DB 235,13 ; jmp 39b7 <_sk_linear_gradient_sse41+0x14b>
DB 15,87,201 ; xorps %xmm1,%xmm1
DB 15,87,210 ; xorps %xmm2,%xmm2
DB 15,87,219 ; xorps %xmm3,%xmm3
@@ -14048,9 +14299,67 @@ ALIGN 16
DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax)
DB 63 ; (bad)
DB 0,0 ; add %al,(%rax)
- DB 128,63,0 ; cmpb $0x0,(%rdi)
+ DB 128,63,255 ; cmpb $0xff,(%rdi)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,1 ; add %al,(%rcx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004018 <_sk_callback_sse41+0xa0000f0>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004020 <_sk_callback_sse41+0x30000f8>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 8,255 ; or %bh,%bh
+ DB 10,255 ; or %bh,%bh
+ DB 12,255 ; or $0xff,%al
+ DB 14 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
DB 0,0 ; add %al,(%rax)
- DB 52,0 ; xor $0x0,%al
+ DB 0,0 ; add %al,(%rax)
+ DB 0,8 ; add %cl,(%rax)
+ DB 128,10,128 ; orb $0x80,(%rdx)
+ DB 12,128 ; or $0x80,%al
+ DB 14 ; (bad)
+ DB 128,0,0 ; addb $0x0,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,52,0 ; add %dh,(%rax,%rax,1)
DB 0,0 ; add %al,(%rax)
DB 52,0 ; xor $0x0,%al
DB 0,0 ; add %al,(%rax)
@@ -14058,16 +14367,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4004 <.literal16+0x34>
+ DB 127,0 ; jg 4084 <.literal16+0xa4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4008 <.literal16+0x38>
+ DB 127,0 ; jg 4088 <.literal16+0xa8>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 400c <.literal16+0x3c>
+ DB 127,0 ; jg 408c <.literal16+0xac>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4010 <.literal16+0x40>
+ DB 127,0 ; jg 4090 <.literal16+0xb0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14076,7 +14385,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4095 <.literal16+0xc5>
+ DB 119,115 ; ja 4115 <.literal16+0x135>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14087,7 +14396,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 3ff9 <.literal16+0x29>
+ DB 117,191 ; jne 4079 <.literal16+0x99>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14099,7 +14408,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a3803a <_sk_callback_sse41+0xffffffffe9a3411e>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a380ba <_sk_callback_sse41+0xffffffffe9a34192>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14148,16 +14457,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 40c4 <.literal16+0xf4>
+ DB 127,0 ; jg 4144 <.literal16+0x164>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 40c8 <.literal16+0xf8>
+ DB 127,0 ; jg 4148 <.literal16+0x168>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 40cc <.literal16+0xfc>
+ DB 127,0 ; jg 414c <.literal16+0x16c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 40d0 <.literal16+0x100>
+ DB 127,0 ; jg 4150 <.literal16+0x170>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14166,7 +14475,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4155 <.literal16+0x185>
+ DB 119,115 ; ja 41d5 <.literal16+0x1f5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14177,7 +14486,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 40b9 <.literal16+0xe9>
+ DB 117,191 ; jne 4139 <.literal16+0x159>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14189,7 +14498,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a380fa <_sk_callback_sse41+0xffffffffe9a341de>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a3817a <_sk_callback_sse41+0xffffffffe9a34252>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14238,16 +14547,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4184 <.literal16+0x1b4>
+ DB 127,0 ; jg 4204 <.literal16+0x224>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4188 <.literal16+0x1b8>
+ DB 127,0 ; jg 4208 <.literal16+0x228>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 418c <.literal16+0x1bc>
+ DB 127,0 ; jg 420c <.literal16+0x22c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4190 <.literal16+0x1c0>
+ DB 127,0 ; jg 4210 <.literal16+0x230>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14256,7 +14565,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4215 <.literal16+0x245>
+ DB 119,115 ; ja 4295 <.literal16+0x2b5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14267,7 +14576,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4179 <.literal16+0x1a9>
+ DB 117,191 ; jne 41f9 <.literal16+0x219>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14279,7 +14588,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a381ba <_sk_callback_sse41+0xffffffffe9a3429e>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a3823a <_sk_callback_sse41+0xffffffffe9a34312>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14328,16 +14637,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4244 <.literal16+0x274>
+ DB 127,0 ; jg 42c4 <.literal16+0x2e4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4248 <.literal16+0x278>
+ DB 127,0 ; jg 42c8 <.literal16+0x2e8>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 424c <.literal16+0x27c>
+ DB 127,0 ; jg 42cc <.literal16+0x2ec>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4250 <.literal16+0x280>
+ DB 127,0 ; jg 42d0 <.literal16+0x2f0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -14346,7 +14655,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 42d5 <.literal16+0x305>
+ DB 119,115 ; ja 4355 <.literal16+0x375>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -14357,7 +14666,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4239 <.literal16+0x269>
+ DB 117,191 ; jne 42b9 <.literal16+0x2d9>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -14369,7 +14678,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a3827a <_sk_callback_sse41+0xffffffffe9a3435e>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a382fa <_sk_callback_sse41+0xffffffffe9a343d2>
DB 220,63 ; fdivrl (%rdi)
DB 81 ; push %rcx
DB 140,242 ; mov %?,%edx
@@ -14409,13 +14718,90 @@ ALIGN 16
DB 75,0,0 ; rex.WXB add %al,(%r8)
DB 0,75,0 ; add %cl,0x0(%rbx)
DB 0,0 ; add %al,(%rax)
- DB 75,0,128,0,0,0,128 ; rex.WXB add %al,-0x80000000(%r8)
+ DB 75,255,0 ; rex.WXB incq (%r8)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a004388 <_sk_callback_sse41+0xa000460>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 3004390 <_sk_callback_sse41+0x3000468>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0043b8 <_sk_callback_sse41+0xa000490>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30043c0 <_sk_callback_sse41+0x3000498>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 1,255 ; add %edi,%edi
+ DB 255 ; (bad)
+ DB 255,5,255,255,255,9 ; incl 0x9ffffff(%rip) # a0043e8 <_sk_callback_sse41+0xa0004c0>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,13,255,255,255,2 ; decl 0x2ffffff(%rip) # 30043f0 <_sk_callback_sse41+0x30004c8>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,6 ; incl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,10 ; decl (%rdx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,14 ; decl (%rsi)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,0 ; incl (%rax)
+ DB 128,0,0 ; addb $0x0,(%rax)
DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax)
DB 0,0 ; add %al,(%rax)
- DB 0,4,0 ; add %al,(%rax,%rax,1)
+ DB 0,128,0,0,0,4 ; add %al,0x4000000(%rax)
DB 0,0 ; add %al,(%rax)
- DB 4,0 ; add $0x0,%al
+ DB 0,4,0 ; add %al,(%rax,%rax,1)
DB 0,0 ; add %al,(%rax)
DB 4,0 ; add $0x0,%al
DB 0,0 ; add %al,(%rax)
@@ -14546,7 +14932,7 @@ _sk_seed_shader_sse2 LABEL PROC
DB 102,15,110,199 ; movd %edi,%xmm0
DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
- DB 15,40,21,177,66,0,0 ; movaps 0x42b1(%rip),%xmm2 # 43c0 <_sk_callback_sse2+0xab>
+ DB 15,40,21,161,66,0,0 ; movaps 0x42a1(%rip),%xmm2 # 43b0 <_sk_callback_sse2+0xad>
DB 15,88,202 ; addps %xmm2,%xmm1
DB 15,16,2 ; movups (%rdx),%xmm0
DB 15,88,193 ; addps %xmm1,%xmm0
@@ -14555,7 +14941,7 @@ _sk_seed_shader_sse2 LABEL PROC
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 15,88,202 ; addps %xmm2,%xmm1
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 15,40,21,160,66,0,0 ; movaps 0x42a0(%rip),%xmm2 # 43d0 <_sk_callback_sse2+0xbb>
+ DB 15,40,21,144,66,0,0 ; movaps 0x4290(%rip),%xmm2 # 43c0 <_sk_callback_sse2+0xbd>
DB 15,87,219 ; xorps %xmm3,%xmm3
DB 15,87,228 ; xorps %xmm4,%xmm4
DB 15,87,237 ; xorps %xmm5,%xmm5
@@ -16107,67 +16493,64 @@ _sk_lerp_565_sse2 LABEL PROC
PUBLIC _sk_load_tables_sse2
_sk_load_tables_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
- DB 72,139,8 ; mov (%rax),%rcx
- DB 76,139,64,8 ; mov 0x8(%rax),%r8
- DB 243,68,15,111,4,185 ; movdqu (%rcx,%rdi,4),%xmm8
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 102,15,110,193 ; movd %ecx,%xmm0
- DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
- DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9
- DB 102,65,15,114,209,8 ; psrld $0x8,%xmm9
- DB 102,68,15,219,200 ; pand %xmm0,%xmm9
- DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10
- DB 102,65,15,114,210,16 ; psrld $0x10,%xmm10
- DB 102,68,15,219,208 ; pand %xmm0,%xmm10
+ DB 76,139,0 ; mov (%rax),%r8
+ DB 76,139,72,8 ; mov 0x8(%rax),%r9
+ DB 243,69,15,111,12,184 ; movdqu (%r8,%rdi,4),%xmm9
+ DB 102,68,15,111,5,76,45,0,0 ; movdqa 0x2d4c(%rip),%xmm8 # 43d0 <_sk_callback_sse2+0xcd>
+ DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0
DB 102,65,15,219,192 ; pand %xmm8,%xmm0
- DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3
- DB 102,72,15,126,217 ; movq %xmm3,%rcx
- DB 65,137,201 ; mov %ecx,%r9d
- DB 72,193,233,32 ; shr $0x20,%rcx
- DB 102,73,15,126,194 ; movq %xmm0,%r10
- DB 69,137,211 ; mov %r10d,%r11d
- DB 73,193,234,32 ; shr $0x20,%r10
- DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3
- DB 243,65,15,16,4,136 ; movss (%r8,%rcx,4),%xmm0
- DB 15,20,216 ; unpcklps %xmm0,%xmm3
- DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0
- DB 243,67,15,16,12,136 ; movss (%r8,%r9,4),%xmm1
+ DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1
+ DB 102,73,15,126,200 ; movq %xmm1,%r8
+ DB 102,72,15,126,193 ; movq %xmm0,%rcx
+ DB 68,15,182,209 ; movzbl %cl,%r10d
+ DB 72,193,233,30 ; shr $0x1e,%rcx
+ DB 69,15,182,216 ; movzbl %r8b,%r11d
+ DB 73,193,232,30 ; shr $0x1e,%r8
+ DB 243,65,15,16,12,9 ; movss (%r9,%rcx,1),%xmm1
+ DB 243,67,15,16,4,1 ; movss (%r9,%r8,1),%xmm0
+ DB 15,20,200 ; unpcklps %xmm0,%xmm1
+ DB 243,67,15,16,4,145 ; movss (%r9,%r10,4),%xmm0
+ DB 243,67,15,16,20,153 ; movss (%r9,%r11,4),%xmm2
+ DB 15,20,194 ; unpcklps %xmm2,%xmm0
DB 15,20,193 ; unpcklps %xmm1,%xmm0
- DB 15,20,195 ; unpcklps %xmm3,%xmm0
DB 76,139,64,16 ; mov 0x10(%rax),%r8
- DB 102,65,15,112,201,78 ; pshufd $0x4e,%xmm9,%xmm1
- DB 102,73,15,126,202 ; movq %xmm1,%r10
- DB 77,137,209 ; mov %r10,%r9
- DB 73,193,233,32 ; shr $0x20,%r9
- DB 102,76,15,126,201 ; movq %xmm9,%rcx
- DB 65,137,203 ; mov %ecx,%r11d
- DB 65,129,227,255,255,255,0 ; and $0xffffff,%r11d
+ DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1
+ DB 102,15,114,209,8 ; psrld $0x8,%xmm1
+ DB 102,65,15,219,200 ; pand %xmm8,%xmm1
+ DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2
+ DB 102,73,15,126,209 ; movq %xmm2,%r9
+ DB 102,72,15,126,201 ; movq %xmm1,%rcx
+ DB 68,15,182,209 ; movzbl %cl,%r10d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 65,129,226,255,255,255,0 ; and $0xffffff,%r10d
- DB 243,65,15,16,28,8 ; movss (%r8,%rcx,1),%xmm3
- DB 243,67,15,16,12,136 ; movss (%r8,%r9,4),%xmm1
- DB 15,20,217 ; unpcklps %xmm1,%xmm3
- DB 243,67,15,16,12,152 ; movss (%r8,%r11,4),%xmm1
- DB 243,67,15,16,20,144 ; movss (%r8,%r10,4),%xmm2
- DB 15,20,202 ; unpcklps %xmm2,%xmm1
+ DB 69,15,182,217 ; movzbl %r9b,%r11d
+ DB 73,193,233,30 ; shr $0x1e,%r9
+ DB 243,65,15,16,20,8 ; movss (%r8,%rcx,1),%xmm2
+ DB 243,67,15,16,12,8 ; movss (%r8,%r9,1),%xmm1
+ DB 15,20,209 ; unpcklps %xmm1,%xmm2
+ DB 243,67,15,16,12,144 ; movss (%r8,%r10,4),%xmm1
+ DB 243,67,15,16,28,152 ; movss (%r8,%r11,4),%xmm3
DB 15,20,203 ; unpcklps %xmm3,%xmm1
- DB 76,139,72,24 ; mov 0x18(%rax),%r9
- DB 102,65,15,112,210,78 ; pshufd $0x4e,%xmm10,%xmm2
- DB 102,72,15,126,209 ; movq %xmm2,%rcx
- DB 68,15,183,193 ; movzwl %cx,%r8d
- DB 72,193,233,32 ; shr $0x20,%rcx
- DB 102,76,15,126,208 ; movq %xmm10,%rax
- DB 68,15,183,208 ; movzwl %ax,%r10d
+ DB 15,20,202 ; unpcklps %xmm2,%xmm1
+ DB 76,139,64,24 ; mov 0x18(%rax),%r8
+ DB 102,65,15,111,209 ; movdqa %xmm9,%xmm2
+ DB 102,15,114,210,16 ; psrld $0x10,%xmm2
+ DB 102,65,15,219,208 ; pand %xmm8,%xmm2
+ DB 102,15,112,218,78 ; pshufd $0x4e,%xmm2,%xmm3
+ DB 102,72,15,126,217 ; movq %xmm3,%rcx
+ DB 102,72,15,126,208 ; movq %xmm2,%rax
+ DB 68,15,182,200 ; movzbl %al,%r9d
DB 72,193,232,30 ; shr $0x1e,%rax
- DB 243,69,15,16,12,1 ; movss (%r9,%rax,1),%xmm9
- DB 243,65,15,16,20,137 ; movss (%r9,%rcx,4),%xmm2
- DB 68,15,20,202 ; unpcklps %xmm2,%xmm9
- DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2
- DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3
+ DB 68,15,182,209 ; movzbl %cl,%r10d
+ DB 72,193,233,30 ; shr $0x1e,%rcx
+ DB 243,69,15,16,4,0 ; movss (%r8,%rax,1),%xmm8
+ DB 243,65,15,16,20,8 ; movss (%r8,%rcx,1),%xmm2
+ DB 68,15,20,194 ; unpcklps %xmm2,%xmm8
+ DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2
+ DB 243,67,15,16,28,144 ; movss (%r8,%r10,4),%xmm3
DB 15,20,211 ; unpcklps %xmm3,%xmm2
- DB 65,15,20,209 ; unpcklps %xmm9,%xmm2
- DB 102,65,15,114,208,24 ; psrld $0x18,%xmm8
- DB 69,15,91,192 ; cvtdq2ps %xmm8,%xmm8
+ DB 65,15,20,208 ; unpcklps %xmm8,%xmm2
+ DB 102,65,15,114,209,24 ; psrld $0x18,%xmm9
+ DB 69,15,91,193 ; cvtdq2ps %xmm9,%xmm8
DB 184,129,128,128,59 ; mov $0x3b808081,%eax
DB 102,15,110,216 ; movd %eax,%xmm3
DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
@@ -16182,25 +16565,23 @@ _sk_load_tables_u16_be_sse2 LABEL PROC
DB 76,139,64,8 ; mov 0x8(%rax),%r8
DB 243,15,111,4,249 ; movdqu (%rcx,%rdi,8),%xmm0
DB 243,15,111,76,249,16 ; movdqu 0x10(%rcx,%rdi,8),%xmm1
- DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10
- DB 102,68,15,97,209 ; punpcklwd %xmm1,%xmm10
+ DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9
+ DB 102,68,15,97,201 ; punpcklwd %xmm1,%xmm9
DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0
- DB 102,65,15,111,202 ; movdqa %xmm10,%xmm1
+ DB 102,65,15,111,201 ; movdqa %xmm9,%xmm1
DB 102,15,97,200 ; punpcklwd %xmm0,%xmm1
- DB 102,68,15,105,208 ; punpckhwd %xmm0,%xmm10
- DB 102,69,15,239,192 ; pxor %xmm8,%xmm8
+ DB 102,68,15,105,200 ; punpckhwd %xmm0,%xmm9
+ DB 102,68,15,111,21,18,44,0,0 ; movdqa 0x2c12(%rip),%xmm10 # 43e0 <_sk_callback_sse2+0xdd>
DB 102,15,111,193 ; movdqa %xmm1,%xmm0
+ DB 102,65,15,219,194 ; pand %xmm10,%xmm0
+ DB 102,69,15,239,192 ; pxor %xmm8,%xmm8
DB 102,65,15,97,192 ; punpcklwd %xmm8,%xmm0
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 102,15,110,217 ; movd %ecx,%xmm3
- DB 102,68,15,112,203,0 ; pshufd $0x0,%xmm3,%xmm9
- DB 102,65,15,219,193 ; pand %xmm9,%xmm0
DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3
DB 102,72,15,126,217 ; movq %xmm3,%rcx
- DB 68,15,183,201 ; movzwl %cx,%r9d
+ DB 68,15,182,201 ; movzbl %cl,%r9d
DB 72,193,233,32 ; shr $0x20,%rcx
DB 102,73,15,126,194 ; movq %xmm0,%r10
- DB 69,15,183,218 ; movzwl %r10w,%r11d
+ DB 69,15,182,218 ; movzbl %r10b,%r11d
DB 73,193,234,30 ; shr $0x1e,%r10
DB 243,67,15,16,28,16 ; movss (%r8,%r10,1),%xmm3
DB 243,65,15,16,4,136 ; movss (%r8,%rcx,4),%xmm0
@@ -16211,14 +16592,14 @@ _sk_load_tables_u16_be_sse2 LABEL PROC
DB 15,20,195 ; unpcklps %xmm3,%xmm0
DB 76,139,80,16 ; mov 0x10(%rax),%r10
DB 102,15,115,217,8 ; psrldq $0x8,%xmm1
+ DB 102,65,15,219,202 ; pand %xmm10,%xmm1
DB 102,65,15,97,200 ; punpcklwd %xmm8,%xmm1
- DB 102,65,15,219,201 ; pand %xmm9,%xmm1
DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2
DB 102,73,15,126,208 ; movq %xmm2,%r8
- DB 69,15,183,200 ; movzwl %r8w,%r9d
+ DB 69,15,182,200 ; movzbl %r8b,%r9d
DB 73,193,232,32 ; shr $0x20,%r8
DB 102,72,15,126,201 ; movq %xmm1,%rcx
- DB 68,15,183,217 ; movzwl %cx,%r11d
+ DB 68,15,182,217 ; movzbl %cl,%r11d
DB 72,193,233,30 ; shr $0x1e,%rcx
DB 243,65,15,16,20,10 ; movss (%r10,%rcx,1),%xmm2
DB 243,67,15,16,12,130 ; movss (%r10,%r8,4),%xmm1
@@ -16228,33 +16609,33 @@ _sk_load_tables_u16_be_sse2 LABEL PROC
DB 15,20,203 ; unpcklps %xmm3,%xmm1
DB 15,20,202 ; unpcklps %xmm2,%xmm1
DB 76,139,72,24 ; mov 0x18(%rax),%r9
- DB 102,69,15,112,218,78 ; pshufd $0x4e,%xmm10,%xmm11
- DB 102,69,15,97,208 ; punpcklwd %xmm8,%xmm10
DB 102,69,15,219,209 ; pand %xmm9,%xmm10
+ DB 102,69,15,97,208 ; punpcklwd %xmm8,%xmm10
DB 102,65,15,112,210,78 ; pshufd $0x4e,%xmm10,%xmm2
DB 102,72,15,126,209 ; movq %xmm2,%rcx
- DB 68,15,183,193 ; movzwl %cx,%r8d
+ DB 68,15,182,193 ; movzbl %cl,%r8d
DB 72,193,233,32 ; shr $0x20,%rcx
DB 102,76,15,126,208 ; movq %xmm10,%rax
- DB 68,15,183,208 ; movzwl %ax,%r10d
+ DB 68,15,182,208 ; movzbl %al,%r10d
DB 72,193,232,30 ; shr $0x1e,%rax
- DB 243,69,15,16,12,1 ; movss (%r9,%rax,1),%xmm9
+ DB 243,69,15,16,20,1 ; movss (%r9,%rax,1),%xmm10
DB 243,65,15,16,20,137 ; movss (%r9,%rcx,4),%xmm2
- DB 68,15,20,202 ; unpcklps %xmm2,%xmm9
+ DB 68,15,20,210 ; unpcklps %xmm2,%xmm10
DB 243,67,15,16,20,145 ; movss (%r9,%r10,4),%xmm2
DB 243,67,15,16,28,129 ; movss (%r9,%r8,4),%xmm3
DB 15,20,211 ; unpcklps %xmm3,%xmm2
- DB 65,15,20,209 ; unpcklps %xmm9,%xmm2
+ DB 65,15,20,210 ; unpcklps %xmm10,%xmm2
DB 184,128,0,128,55 ; mov $0x37800080,%eax
- DB 102,68,15,110,200 ; movd %eax,%xmm9
- DB 69,15,198,201,0 ; shufps $0x0,%xmm9,%xmm9
- DB 102,65,15,111,219 ; movdqa %xmm11,%xmm3
- DB 102,15,113,243,8 ; psllw $0x8,%xmm3
- DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11
- DB 102,68,15,235,219 ; por %xmm3,%xmm11
- DB 102,69,15,97,216 ; punpcklwd %xmm8,%xmm11
- DB 65,15,91,219 ; cvtdq2ps %xmm11,%xmm3
- DB 65,15,89,217 ; mulps %xmm9,%xmm3
+ DB 102,68,15,110,208 ; movd %eax,%xmm10
+ DB 69,15,198,210,0 ; shufps $0x0,%xmm10,%xmm10
+ DB 102,65,15,112,217,78 ; pshufd $0x4e,%xmm9,%xmm3
+ DB 102,68,15,111,203 ; movdqa %xmm3,%xmm9
+ DB 102,65,15,113,241,8 ; psllw $0x8,%xmm9
+ DB 102,15,113,211,8 ; psrlw $0x8,%xmm3
+ DB 102,65,15,235,217 ; por %xmm9,%xmm3
+ DB 102,65,15,97,216 ; punpcklwd %xmm8,%xmm3
+ DB 15,91,219 ; cvtdq2ps %xmm3,%xmm3
+ DB 65,15,89,218 ; mulps %xmm10,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -16267,61 +16648,59 @@ _sk_load_tables_rgb_u16_be_sse2 LABEL PROC
DB 243,70,15,111,28,73 ; movdqu (%rcx,%r9,2),%xmm11
DB 243,66,15,111,68,73,8 ; movdqu 0x8(%rcx,%r9,2),%xmm0
DB 102,15,115,216,4 ; psrldq $0x4,%xmm0
- DB 102,65,15,111,219 ; movdqa %xmm11,%xmm3
- DB 102,15,115,219,6 ; psrldq $0x6,%xmm3
+ DB 102,69,15,111,211 ; movdqa %xmm11,%xmm10
+ DB 102,65,15,115,218,6 ; psrldq $0x6,%xmm10
DB 102,68,15,97,216 ; punpcklwd %xmm0,%xmm11
DB 102,15,115,216,6 ; psrldq $0x6,%xmm0
- DB 102,15,97,216 ; punpcklwd %xmm0,%xmm3
+ DB 102,68,15,97,208 ; punpcklwd %xmm0,%xmm10
DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0
- DB 102,15,97,195 ; punpcklwd %xmm3,%xmm0
- DB 102,68,15,112,208,78 ; pshufd $0x4e,%xmm0,%xmm10
- DB 102,68,15,105,219 ; punpckhwd %xmm3,%xmm11
- DB 102,69,15,239,192 ; pxor %xmm8,%xmm8
- DB 102,65,15,97,192 ; punpcklwd %xmm8,%xmm0
- DB 185,255,0,0,0 ; mov $0xff,%ecx
- DB 102,15,110,217 ; movd %ecx,%xmm3
- DB 102,68,15,112,203,0 ; pshufd $0x0,%xmm3,%xmm9
- DB 102,65,15,219,193 ; pand %xmm9,%xmm0
+ DB 102,65,15,97,194 ; punpcklwd %xmm10,%xmm0
+ DB 102,68,15,111,5,165,42,0,0 ; movdqa 0x2aa5(%rip),%xmm8 # 43f0 <_sk_callback_sse2+0xed>
+ DB 102,15,112,200,78 ; pshufd $0x4e,%xmm0,%xmm1
+ DB 102,65,15,219,192 ; pand %xmm8,%xmm0
+ DB 102,69,15,239,201 ; pxor %xmm9,%xmm9
+ DB 102,65,15,97,193 ; punpcklwd %xmm9,%xmm0
DB 102,15,112,216,78 ; pshufd $0x4e,%xmm0,%xmm3
DB 102,72,15,126,217 ; movq %xmm3,%rcx
- DB 68,15,183,201 ; movzwl %cx,%r9d
+ DB 68,15,182,201 ; movzbl %cl,%r9d
DB 72,193,233,32 ; shr $0x20,%rcx
DB 102,73,15,126,194 ; movq %xmm0,%r10
- DB 69,15,183,218 ; movzwl %r10w,%r11d
+ DB 69,15,182,218 ; movzbl %r10b,%r11d
DB 73,193,234,30 ; shr $0x1e,%r10
DB 243,67,15,16,28,16 ; movss (%r8,%r10,1),%xmm3
DB 243,65,15,16,4,136 ; movss (%r8,%rcx,4),%xmm0
DB 15,20,216 ; unpcklps %xmm0,%xmm3
DB 243,67,15,16,4,152 ; movss (%r8,%r11,4),%xmm0
- DB 243,67,15,16,12,136 ; movss (%r8,%r9,4),%xmm1
- DB 15,20,193 ; unpcklps %xmm1,%xmm0
+ DB 243,67,15,16,20,136 ; movss (%r8,%r9,4),%xmm2
+ DB 15,20,194 ; unpcklps %xmm2,%xmm0
DB 15,20,195 ; unpcklps %xmm3,%xmm0
DB 76,139,80,16 ; mov 0x10(%rax),%r10
- DB 102,69,15,97,208 ; punpcklwd %xmm8,%xmm10
- DB 102,69,15,219,209 ; pand %xmm9,%xmm10
- DB 102,65,15,112,202,78 ; pshufd $0x4e,%xmm10,%xmm1
- DB 102,73,15,126,200 ; movq %xmm1,%r8
- DB 69,15,183,200 ; movzwl %r8w,%r9d
+ DB 102,65,15,219,200 ; pand %xmm8,%xmm1
+ DB 102,65,15,97,201 ; punpcklwd %xmm9,%xmm1
+ DB 102,15,112,209,78 ; pshufd $0x4e,%xmm1,%xmm2
+ DB 102,73,15,126,208 ; movq %xmm2,%r8
+ DB 69,15,182,200 ; movzbl %r8b,%r9d
DB 73,193,232,32 ; shr $0x20,%r8
- DB 102,76,15,126,209 ; movq %xmm10,%rcx
- DB 68,15,183,217 ; movzwl %cx,%r11d
+ DB 102,72,15,126,201 ; movq %xmm1,%rcx
+ DB 68,15,182,217 ; movzbl %cl,%r11d
DB 72,193,233,30 ; shr $0x1e,%rcx
- DB 243,65,15,16,28,10 ; movss (%r10,%rcx,1),%xmm3
+ DB 243,65,15,16,20,10 ; movss (%r10,%rcx,1),%xmm2
DB 243,67,15,16,12,130 ; movss (%r10,%r8,4),%xmm1
- DB 15,20,217 ; unpcklps %xmm1,%xmm3
+ DB 15,20,209 ; unpcklps %xmm1,%xmm2
DB 243,67,15,16,12,154 ; movss (%r10,%r11,4),%xmm1
- DB 243,67,15,16,20,138 ; movss (%r10,%r9,4),%xmm2
- DB 15,20,202 ; unpcklps %xmm2,%xmm1
+ DB 243,67,15,16,28,138 ; movss (%r10,%r9,4),%xmm3
DB 15,20,203 ; unpcklps %xmm3,%xmm1
+ DB 15,20,202 ; unpcklps %xmm2,%xmm1
DB 76,139,72,24 ; mov 0x18(%rax),%r9
- DB 102,69,15,97,216 ; punpcklwd %xmm8,%xmm11
- DB 102,69,15,219,217 ; pand %xmm9,%xmm11
+ DB 102,69,15,105,218 ; punpckhwd %xmm10,%xmm11
+ DB 102,69,15,219,216 ; pand %xmm8,%xmm11
+ DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11
DB 102,65,15,112,211,78 ; pshufd $0x4e,%xmm11,%xmm2
DB 102,72,15,126,209 ; movq %xmm2,%rcx
- DB 68,15,183,193 ; movzwl %cx,%r8d
+ DB 68,15,182,193 ; movzbl %cl,%r8d
DB 72,193,233,32 ; shr $0x20,%rcx
DB 102,76,15,126,216 ; movq %xmm11,%rax
- DB 68,15,183,208 ; movzwl %ax,%r10d
+ DB 68,15,182,208 ; movzbl %al,%r10d
DB 72,193,232,30 ; shr $0x1e,%rax
DB 243,69,15,16,4,1 ; movss (%r9,%rax,1),%xmm8
DB 243,65,15,16,20,137 ; movss (%r9,%rcx,4),%xmm2
@@ -16675,15 +17054,15 @@ _sk_parametric_r_sse2 LABEL PROC
DB 69,15,88,209 ; addps %xmm9,%xmm10
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9
- DB 68,15,89,13,78,36,0,0 ; mulps 0x244e(%rip),%xmm9 # 43e0 <_sk_callback_sse2+0xcb>
- DB 68,15,84,21,86,36,0,0 ; andps 0x2456(%rip),%xmm10 # 43f0 <_sk_callback_sse2+0xdb>
- DB 68,15,86,21,94,36,0,0 ; orps 0x245e(%rip),%xmm10 # 4400 <_sk_callback_sse2+0xeb>
- DB 68,15,88,13,102,36,0,0 ; addps 0x2466(%rip),%xmm9 # 4410 <_sk_callback_sse2+0xfb>
- DB 68,15,40,37,110,36,0,0 ; movaps 0x246e(%rip),%xmm12 # 4420 <_sk_callback_sse2+0x10b>
+ DB 68,15,89,13,136,36,0,0 ; mulps 0x2488(%rip),%xmm9 # 4400 <_sk_callback_sse2+0xfd>
+ DB 68,15,84,21,144,36,0,0 ; andps 0x2490(%rip),%xmm10 # 4410 <_sk_callback_sse2+0x10d>
+ DB 68,15,86,21,152,36,0,0 ; orps 0x2498(%rip),%xmm10 # 4420 <_sk_callback_sse2+0x11d>
+ DB 68,15,88,13,160,36,0,0 ; addps 0x24a0(%rip),%xmm9 # 4430 <_sk_callback_sse2+0x12d>
+ DB 68,15,40,37,168,36,0,0 ; movaps 0x24a8(%rip),%xmm12 # 4440 <_sk_callback_sse2+0x13d>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,110,36,0,0 ; addps 0x246e(%rip),%xmm10 # 4430 <_sk_callback_sse2+0x11b>
- DB 68,15,40,37,118,36,0,0 ; movaps 0x2476(%rip),%xmm12 # 4440 <_sk_callback_sse2+0x12b>
+ DB 68,15,88,21,168,36,0,0 ; addps 0x24a8(%rip),%xmm10 # 4450 <_sk_callback_sse2+0x14d>
+ DB 68,15,40,37,176,36,0,0 ; movaps 0x24b0(%rip),%xmm12 # 4460 <_sk_callback_sse2+0x15d>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -16691,21 +17070,21 @@ _sk_parametric_r_sse2 LABEL PROC
DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12
- DB 68,15,84,37,96,36,0,0 ; andps 0x2460(%rip),%xmm12 # 4450 <_sk_callback_sse2+0x13b>
+ DB 68,15,84,37,154,36,0,0 ; andps 0x249a(%rip),%xmm12 # 4470 <_sk_callback_sse2+0x16d>
DB 69,15,87,210 ; xorps %xmm10,%xmm10
DB 69,15,92,220 ; subps %xmm12,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,92,227 ; subps %xmm11,%xmm12
- DB 68,15,88,13,88,36,0,0 ; addps 0x2458(%rip),%xmm9 # 4460 <_sk_callback_sse2+0x14b>
- DB 68,15,40,29,96,36,0,0 ; movaps 0x2460(%rip),%xmm11 # 4470 <_sk_callback_sse2+0x15b>
+ DB 68,15,88,13,146,36,0,0 ; addps 0x2492(%rip),%xmm9 # 4480 <_sk_callback_sse2+0x17d>
+ DB 68,15,40,29,154,36,0,0 ; movaps 0x249a(%rip),%xmm11 # 4490 <_sk_callback_sse2+0x18d>
DB 69,15,89,220 ; mulps %xmm12,%xmm11
DB 69,15,92,203 ; subps %xmm11,%xmm9
- DB 68,15,40,29,96,36,0,0 ; movaps 0x2460(%rip),%xmm11 # 4480 <_sk_callback_sse2+0x16b>
+ DB 68,15,40,29,154,36,0,0 ; movaps 0x249a(%rip),%xmm11 # 44a0 <_sk_callback_sse2+0x19d>
DB 69,15,92,220 ; subps %xmm12,%xmm11
- DB 68,15,40,37,100,36,0,0 ; movaps 0x2464(%rip),%xmm12 # 4490 <_sk_callback_sse2+0x17b>
+ DB 68,15,40,37,158,36,0,0 ; movaps 0x249e(%rip),%xmm12 # 44b0 <_sk_callback_sse2+0x1ad>
DB 69,15,94,227 ; divps %xmm11,%xmm12
DB 69,15,88,225 ; addps %xmm9,%xmm12
- DB 68,15,89,37,100,36,0,0 ; mulps 0x2464(%rip),%xmm12 # 44a0 <_sk_callback_sse2+0x18b>
+ DB 68,15,89,37,158,36,0,0 ; mulps 0x249e(%rip),%xmm12 # 44c0 <_sk_callback_sse2+0x1bd>
DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9
DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
@@ -16742,15 +17121,15 @@ _sk_parametric_g_sse2 LABEL PROC
DB 69,15,88,209 ; addps %xmm9,%xmm10
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9
- DB 68,15,89,13,213,35,0,0 ; mulps 0x23d5(%rip),%xmm9 # 44b0 <_sk_callback_sse2+0x19b>
- DB 68,15,84,21,221,35,0,0 ; andps 0x23dd(%rip),%xmm10 # 44c0 <_sk_callback_sse2+0x1ab>
- DB 68,15,86,21,229,35,0,0 ; orps 0x23e5(%rip),%xmm10 # 44d0 <_sk_callback_sse2+0x1bb>
- DB 68,15,88,13,237,35,0,0 ; addps 0x23ed(%rip),%xmm9 # 44e0 <_sk_callback_sse2+0x1cb>
- DB 68,15,40,37,245,35,0,0 ; movaps 0x23f5(%rip),%xmm12 # 44f0 <_sk_callback_sse2+0x1db>
+ DB 68,15,89,13,15,36,0,0 ; mulps 0x240f(%rip),%xmm9 # 44d0 <_sk_callback_sse2+0x1cd>
+ DB 68,15,84,21,23,36,0,0 ; andps 0x2417(%rip),%xmm10 # 44e0 <_sk_callback_sse2+0x1dd>
+ DB 68,15,86,21,31,36,0,0 ; orps 0x241f(%rip),%xmm10 # 44f0 <_sk_callback_sse2+0x1ed>
+ DB 68,15,88,13,39,36,0,0 ; addps 0x2427(%rip),%xmm9 # 4500 <_sk_callback_sse2+0x1fd>
+ DB 68,15,40,37,47,36,0,0 ; movaps 0x242f(%rip),%xmm12 # 4510 <_sk_callback_sse2+0x20d>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,245,35,0,0 ; addps 0x23f5(%rip),%xmm10 # 4500 <_sk_callback_sse2+0x1eb>
- DB 68,15,40,37,253,35,0,0 ; movaps 0x23fd(%rip),%xmm12 # 4510 <_sk_callback_sse2+0x1fb>
+ DB 68,15,88,21,47,36,0,0 ; addps 0x242f(%rip),%xmm10 # 4520 <_sk_callback_sse2+0x21d>
+ DB 68,15,40,37,55,36,0,0 ; movaps 0x2437(%rip),%xmm12 # 4530 <_sk_callback_sse2+0x22d>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -16758,21 +17137,21 @@ _sk_parametric_g_sse2 LABEL PROC
DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12
- DB 68,15,84,37,231,35,0,0 ; andps 0x23e7(%rip),%xmm12 # 4520 <_sk_callback_sse2+0x20b>
+ DB 68,15,84,37,33,36,0,0 ; andps 0x2421(%rip),%xmm12 # 4540 <_sk_callback_sse2+0x23d>
DB 69,15,87,210 ; xorps %xmm10,%xmm10
DB 69,15,92,220 ; subps %xmm12,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,92,227 ; subps %xmm11,%xmm12
- DB 68,15,88,13,223,35,0,0 ; addps 0x23df(%rip),%xmm9 # 4530 <_sk_callback_sse2+0x21b>
- DB 68,15,40,29,231,35,0,0 ; movaps 0x23e7(%rip),%xmm11 # 4540 <_sk_callback_sse2+0x22b>
+ DB 68,15,88,13,25,36,0,0 ; addps 0x2419(%rip),%xmm9 # 4550 <_sk_callback_sse2+0x24d>
+ DB 68,15,40,29,33,36,0,0 ; movaps 0x2421(%rip),%xmm11 # 4560 <_sk_callback_sse2+0x25d>
DB 69,15,89,220 ; mulps %xmm12,%xmm11
DB 69,15,92,203 ; subps %xmm11,%xmm9
- DB 68,15,40,29,231,35,0,0 ; movaps 0x23e7(%rip),%xmm11 # 4550 <_sk_callback_sse2+0x23b>
+ DB 68,15,40,29,33,36,0,0 ; movaps 0x2421(%rip),%xmm11 # 4570 <_sk_callback_sse2+0x26d>
DB 69,15,92,220 ; subps %xmm12,%xmm11
- DB 68,15,40,37,235,35,0,0 ; movaps 0x23eb(%rip),%xmm12 # 4560 <_sk_callback_sse2+0x24b>
+ DB 68,15,40,37,37,36,0,0 ; movaps 0x2425(%rip),%xmm12 # 4580 <_sk_callback_sse2+0x27d>
DB 69,15,94,227 ; divps %xmm11,%xmm12
DB 69,15,88,225 ; addps %xmm9,%xmm12
- DB 68,15,89,37,235,35,0,0 ; mulps 0x23eb(%rip),%xmm12 # 4570 <_sk_callback_sse2+0x25b>
+ DB 68,15,89,37,37,36,0,0 ; mulps 0x2425(%rip),%xmm12 # 4590 <_sk_callback_sse2+0x28d>
DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9
DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
@@ -16809,15 +17188,15 @@ _sk_parametric_b_sse2 LABEL PROC
DB 69,15,88,209 ; addps %xmm9,%xmm10
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9
- DB 68,15,89,13,92,35,0,0 ; mulps 0x235c(%rip),%xmm9 # 4580 <_sk_callback_sse2+0x26b>
- DB 68,15,84,21,100,35,0,0 ; andps 0x2364(%rip),%xmm10 # 4590 <_sk_callback_sse2+0x27b>
- DB 68,15,86,21,108,35,0,0 ; orps 0x236c(%rip),%xmm10 # 45a0 <_sk_callback_sse2+0x28b>
- DB 68,15,88,13,116,35,0,0 ; addps 0x2374(%rip),%xmm9 # 45b0 <_sk_callback_sse2+0x29b>
- DB 68,15,40,37,124,35,0,0 ; movaps 0x237c(%rip),%xmm12 # 45c0 <_sk_callback_sse2+0x2ab>
+ DB 68,15,89,13,150,35,0,0 ; mulps 0x2396(%rip),%xmm9 # 45a0 <_sk_callback_sse2+0x29d>
+ DB 68,15,84,21,158,35,0,0 ; andps 0x239e(%rip),%xmm10 # 45b0 <_sk_callback_sse2+0x2ad>
+ DB 68,15,86,21,166,35,0,0 ; orps 0x23a6(%rip),%xmm10 # 45c0 <_sk_callback_sse2+0x2bd>
+ DB 68,15,88,13,174,35,0,0 ; addps 0x23ae(%rip),%xmm9 # 45d0 <_sk_callback_sse2+0x2cd>
+ DB 68,15,40,37,182,35,0,0 ; movaps 0x23b6(%rip),%xmm12 # 45e0 <_sk_callback_sse2+0x2dd>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,124,35,0,0 ; addps 0x237c(%rip),%xmm10 # 45d0 <_sk_callback_sse2+0x2bb>
- DB 68,15,40,37,132,35,0,0 ; movaps 0x2384(%rip),%xmm12 # 45e0 <_sk_callback_sse2+0x2cb>
+ DB 68,15,88,21,182,35,0,0 ; addps 0x23b6(%rip),%xmm10 # 45f0 <_sk_callback_sse2+0x2ed>
+ DB 68,15,40,37,190,35,0,0 ; movaps 0x23be(%rip),%xmm12 # 4600 <_sk_callback_sse2+0x2fd>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -16825,21 +17204,21 @@ _sk_parametric_b_sse2 LABEL PROC
DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12
- DB 68,15,84,37,110,35,0,0 ; andps 0x236e(%rip),%xmm12 # 45f0 <_sk_callback_sse2+0x2db>
+ DB 68,15,84,37,168,35,0,0 ; andps 0x23a8(%rip),%xmm12 # 4610 <_sk_callback_sse2+0x30d>
DB 69,15,87,210 ; xorps %xmm10,%xmm10
DB 69,15,92,220 ; subps %xmm12,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,92,227 ; subps %xmm11,%xmm12
- DB 68,15,88,13,102,35,0,0 ; addps 0x2366(%rip),%xmm9 # 4600 <_sk_callback_sse2+0x2eb>
- DB 68,15,40,29,110,35,0,0 ; movaps 0x236e(%rip),%xmm11 # 4610 <_sk_callback_sse2+0x2fb>
+ DB 68,15,88,13,160,35,0,0 ; addps 0x23a0(%rip),%xmm9 # 4620 <_sk_callback_sse2+0x31d>
+ DB 68,15,40,29,168,35,0,0 ; movaps 0x23a8(%rip),%xmm11 # 4630 <_sk_callback_sse2+0x32d>
DB 69,15,89,220 ; mulps %xmm12,%xmm11
DB 69,15,92,203 ; subps %xmm11,%xmm9
- DB 68,15,40,29,110,35,0,0 ; movaps 0x236e(%rip),%xmm11 # 4620 <_sk_callback_sse2+0x30b>
+ DB 68,15,40,29,168,35,0,0 ; movaps 0x23a8(%rip),%xmm11 # 4640 <_sk_callback_sse2+0x33d>
DB 69,15,92,220 ; subps %xmm12,%xmm11
- DB 68,15,40,37,114,35,0,0 ; movaps 0x2372(%rip),%xmm12 # 4630 <_sk_callback_sse2+0x31b>
+ DB 68,15,40,37,172,35,0,0 ; movaps 0x23ac(%rip),%xmm12 # 4650 <_sk_callback_sse2+0x34d>
DB 69,15,94,227 ; divps %xmm11,%xmm12
DB 69,15,88,225 ; addps %xmm9,%xmm12
- DB 68,15,89,37,114,35,0,0 ; mulps 0x2372(%rip),%xmm12 # 4640 <_sk_callback_sse2+0x32b>
+ DB 68,15,89,37,172,35,0,0 ; mulps 0x23ac(%rip),%xmm12 # 4660 <_sk_callback_sse2+0x35d>
DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9
DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
@@ -16876,15 +17255,15 @@ _sk_parametric_a_sse2 LABEL PROC
DB 69,15,88,209 ; addps %xmm9,%xmm10
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
DB 69,15,91,202 ; cvtdq2ps %xmm10,%xmm9
- DB 68,15,89,13,227,34,0,0 ; mulps 0x22e3(%rip),%xmm9 # 4650 <_sk_callback_sse2+0x33b>
- DB 68,15,84,21,235,34,0,0 ; andps 0x22eb(%rip),%xmm10 # 4660 <_sk_callback_sse2+0x34b>
- DB 68,15,86,21,243,34,0,0 ; orps 0x22f3(%rip),%xmm10 # 4670 <_sk_callback_sse2+0x35b>
- DB 68,15,88,13,251,34,0,0 ; addps 0x22fb(%rip),%xmm9 # 4680 <_sk_callback_sse2+0x36b>
- DB 68,15,40,37,3,35,0,0 ; movaps 0x2303(%rip),%xmm12 # 4690 <_sk_callback_sse2+0x37b>
+ DB 68,15,89,13,29,35,0,0 ; mulps 0x231d(%rip),%xmm9 # 4670 <_sk_callback_sse2+0x36d>
+ DB 68,15,84,21,37,35,0,0 ; andps 0x2325(%rip),%xmm10 # 4680 <_sk_callback_sse2+0x37d>
+ DB 68,15,86,21,45,35,0,0 ; orps 0x232d(%rip),%xmm10 # 4690 <_sk_callback_sse2+0x38d>
+ DB 68,15,88,13,53,35,0,0 ; addps 0x2335(%rip),%xmm9 # 46a0 <_sk_callback_sse2+0x39d>
+ DB 68,15,40,37,61,35,0,0 ; movaps 0x233d(%rip),%xmm12 # 46b0 <_sk_callback_sse2+0x3ad>
DB 69,15,89,226 ; mulps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
- DB 68,15,88,21,3,35,0,0 ; addps 0x2303(%rip),%xmm10 # 46a0 <_sk_callback_sse2+0x38b>
- DB 68,15,40,37,11,35,0,0 ; movaps 0x230b(%rip),%xmm12 # 46b0 <_sk_callback_sse2+0x39b>
+ DB 68,15,88,21,61,35,0,0 ; addps 0x233d(%rip),%xmm10 # 46c0 <_sk_callback_sse2+0x3bd>
+ DB 68,15,40,37,69,35,0,0 ; movaps 0x2345(%rip),%xmm12 # 46d0 <_sk_callback_sse2+0x3cd>
DB 69,15,94,226 ; divps %xmm10,%xmm12
DB 69,15,92,204 ; subps %xmm12,%xmm9
DB 69,15,89,203 ; mulps %xmm11,%xmm9
@@ -16892,21 +17271,21 @@ _sk_parametric_a_sse2 LABEL PROC
DB 69,15,91,218 ; cvtdq2ps %xmm10,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,194,227,1 ; cmpltps %xmm11,%xmm12
- DB 68,15,84,37,245,34,0,0 ; andps 0x22f5(%rip),%xmm12 # 46c0 <_sk_callback_sse2+0x3ab>
+ DB 68,15,84,37,47,35,0,0 ; andps 0x232f(%rip),%xmm12 # 46e0 <_sk_callback_sse2+0x3dd>
DB 69,15,87,210 ; xorps %xmm10,%xmm10
DB 69,15,92,220 ; subps %xmm12,%xmm11
DB 69,15,40,225 ; movaps %xmm9,%xmm12
DB 69,15,92,227 ; subps %xmm11,%xmm12
- DB 68,15,88,13,237,34,0,0 ; addps 0x22ed(%rip),%xmm9 # 46d0 <_sk_callback_sse2+0x3bb>
- DB 68,15,40,29,245,34,0,0 ; movaps 0x22f5(%rip),%xmm11 # 46e0 <_sk_callback_sse2+0x3cb>
+ DB 68,15,88,13,39,35,0,0 ; addps 0x2327(%rip),%xmm9 # 46f0 <_sk_callback_sse2+0x3ed>
+ DB 68,15,40,29,47,35,0,0 ; movaps 0x232f(%rip),%xmm11 # 4700 <_sk_callback_sse2+0x3fd>
DB 69,15,89,220 ; mulps %xmm12,%xmm11
DB 69,15,92,203 ; subps %xmm11,%xmm9
- DB 68,15,40,29,245,34,0,0 ; movaps 0x22f5(%rip),%xmm11 # 46f0 <_sk_callback_sse2+0x3db>
+ DB 68,15,40,29,47,35,0,0 ; movaps 0x232f(%rip),%xmm11 # 4710 <_sk_callback_sse2+0x40d>
DB 69,15,92,220 ; subps %xmm12,%xmm11
- DB 68,15,40,37,249,34,0,0 ; movaps 0x22f9(%rip),%xmm12 # 4700 <_sk_callback_sse2+0x3eb>
+ DB 68,15,40,37,51,35,0,0 ; movaps 0x2333(%rip),%xmm12 # 4720 <_sk_callback_sse2+0x41d>
DB 69,15,94,227 ; divps %xmm11,%xmm12
DB 69,15,88,225 ; addps %xmm9,%xmm12
- DB 68,15,89,37,249,34,0,0 ; mulps 0x22f9(%rip),%xmm12 # 4710 <_sk_callback_sse2+0x3fb>
+ DB 68,15,89,37,51,35,0,0 ; mulps 0x2333(%rip),%xmm12 # 4730 <_sk_callback_sse2+0x42d>
DB 102,69,15,91,204 ; cvtps2dq %xmm12,%xmm9
DB 243,68,15,16,88,20 ; movss 0x14(%rax),%xmm11
DB 69,15,198,219,0 ; shufps $0x0,%xmm11,%xmm11
@@ -17162,9 +17541,9 @@ _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 27bd <_sk_gather_i8_sse2+0xf>
+ DB 116,5 ; je 27a3 <_sk_gather_i8_sse2+0xf>
DB 76,137,192 ; mov %r8,%rax
- DB 235,2 ; jmp 27bf <_sk_gather_i8_sse2+0x11>
+ DB 235,2 ; jmp 27a5 <_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
@@ -17209,31 +17588,30 @@ _sk_gather_i8_sse2 LABEL PROC
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 102,71,15,110,12,144 ; movd (%r8,%r10,4),%xmm9
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 102,68,15,98,201 ; punpckldq %xmm1,%xmm9
+ DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9
+ DB 102,15,111,21,195,30,0,0 ; movdqa 0x1ec3(%rip),%xmm2 # 4740 <_sk_callback_sse2+0x43d>
+ DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0
+ DB 102,15,219,194 ; pand %xmm2,%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 102,65,15,111,201 ; movdqa %xmm9,%xmm1
+ DB 102,15,114,209,8 ; psrld $0x8,%xmm1
+ DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
- DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3
+ DB 102,15,114,211,16 ; psrld $0x10,%xmm3
+ DB 102,15,219,218 ; pand %xmm2,%xmm3
+ DB 15,91,211 ; cvtdq2ps %xmm3,%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 102,65,15,114,209,24 ; psrld $0x18,%xmm9
+ DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3
DB 65,15,89,216 ; mulps %xmm8,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -17519,28 +17897,27 @@ PUBLIC _sk_load_8888_sse2
_sk_load_8888_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
- DB 243,15,111,28,184 ; movdqu (%rax,%rdi,4),%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 68,15,16,12,184 ; movups (%rax,%rdi,4),%xmm9
+ DB 15,40,21,17,26,0,0 ; movaps 0x1a11(%rip),%xmm2 # 4750 <_sk_callback_sse2+0x44d>
+ DB 65,15,40,193 ; movaps %xmm9,%xmm0
+ DB 15,84,194 ; andps %xmm2,%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 65,15,40,201 ; movaps %xmm9,%xmm1
+ DB 102,15,114,209,8 ; psrld $0x8,%xmm1
+ DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
- DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 65,15,40,217 ; movaps %xmm9,%xmm3
+ DB 102,15,114,211,16 ; psrld $0x10,%xmm3
+ DB 102,15,219,218 ; pand %xmm2,%xmm3
+ DB 15,91,211 ; cvtdq2ps %xmm3,%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 102,65,15,114,209,24 ; psrld $0x18,%xmm9
+ DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3
DB 65,15,89,216 ; mulps %xmm8,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -17570,31 +17947,30 @@ _sk_gather_8888_sse2 LABEL PROC
DB 102,65,15,110,4,137 ; movd (%r9,%rcx,4),%xmm0
DB 102,65,15,110,12,129 ; movd (%r9,%rax,4),%xmm1
DB 102,15,98,193 ; punpckldq %xmm1,%xmm0
- DB 102,67,15,110,28,145 ; movd (%r9,%r10,4),%xmm3
+ DB 102,71,15,110,12,145 ; movd (%r9,%r10,4),%xmm9
DB 102,67,15,110,12,129 ; movd (%r9,%r8,4),%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 102,68,15,98,201 ; punpckldq %xmm1,%xmm9
+ DB 102,68,15,98,200 ; punpckldq %xmm0,%xmm9
+ DB 102,15,111,21,73,25,0,0 ; movdqa 0x1949(%rip),%xmm2 # 4760 <_sk_callback_sse2+0x45d>
+ DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0
+ DB 102,15,219,194 ; pand %xmm2,%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 102,65,15,111,201 ; movdqa %xmm9,%xmm1
+ DB 102,15,114,209,8 ; psrld $0x8,%xmm1
+ DB 102,15,219,202 ; pand %xmm2,%xmm1
DB 15,91,201 ; cvtdq2ps %xmm1,%xmm1
DB 65,15,89,200 ; mulps %xmm8,%xmm1
- DB 15,91,210 ; cvtdq2ps %xmm2,%xmm2
+ DB 102,65,15,111,217 ; movdqa %xmm9,%xmm3
+ DB 102,15,114,211,16 ; psrld $0x10,%xmm3
+ DB 102,15,219,218 ; pand %xmm2,%xmm3
+ DB 15,91,211 ; cvtdq2ps %xmm3,%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 102,65,15,114,209,24 ; psrld $0x18,%xmm9
+ DB 65,15,91,217 ; cvtdq2ps %xmm9,%xmm3
DB 65,15,89,216 ; mulps %xmm8,%xmm3
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -17642,7 +18018,7 @@ _sk_load_f16_sse2 LABEL PROC
DB 102,69,15,239,210 ; pxor %xmm10,%xmm10
DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1
DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1
- DB 102,68,15,111,13,232,23,0,0 ; movdqa 0x17e8(%rip),%xmm9 # 4720 <_sk_callback_sse2+0x40b>
+ DB 102,68,15,111,13,74,24,0,0 ; movdqa 0x184a(%rip),%xmm9 # 4770 <_sk_callback_sse2+0x46d>
DB 102,15,111,193 ; movdqa %xmm1,%xmm0
DB 102,65,15,219,193 ; pand %xmm9,%xmm0
DB 102,15,239,200 ; pxor %xmm0,%xmm1
@@ -17650,11 +18026,11 @@ _sk_load_f16_sse2 LABEL PROC
DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13
DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13
DB 102,68,15,235,232 ; por %xmm0,%xmm13
- DB 102,68,15,111,29,205,23,0,0 ; movdqa 0x17cd(%rip),%xmm11 # 4730 <_sk_callback_sse2+0x41b>
+ DB 102,68,15,111,29,47,24,0,0 ; movdqa 0x182f(%rip),%xmm11 # 4780 <_sk_callback_sse2+0x47d>
DB 102,69,15,254,235 ; paddd %xmm11,%xmm13
- DB 102,68,15,111,37,207,23,0,0 ; movdqa 0x17cf(%rip),%xmm12 # 4740 <_sk_callback_sse2+0x42b>
+ DB 102,68,15,111,37,49,24,0,0 ; movdqa 0x1831(%rip),%xmm12 # 4790 <_sk_callback_sse2+0x48d>
DB 102,65,15,239,204 ; pxor %xmm12,%xmm1
- DB 102,15,111,29,210,23,0,0 ; movdqa 0x17d2(%rip),%xmm3 # 4750 <_sk_callback_sse2+0x43b>
+ DB 102,15,111,29,52,24,0,0 ; movdqa 0x1834(%rip),%xmm3 # 47a0 <_sk_callback_sse2+0x49d>
DB 102,15,111,195 ; movdqa %xmm3,%xmm0
DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0
DB 102,65,15,223,197 ; pandn %xmm13,%xmm0
@@ -17738,7 +18114,7 @@ _sk_gather_f16_sse2 LABEL PROC
DB 102,69,15,239,210 ; pxor %xmm10,%xmm10
DB 102,65,15,111,206 ; movdqa %xmm14,%xmm1
DB 102,65,15,97,202 ; punpcklwd %xmm10,%xmm1
- DB 102,68,15,111,13,96,22,0,0 ; movdqa 0x1660(%rip),%xmm9 # 4760 <_sk_callback_sse2+0x44b>
+ DB 102,68,15,111,13,194,22,0,0 ; movdqa 0x16c2(%rip),%xmm9 # 47b0 <_sk_callback_sse2+0x4ad>
DB 102,15,111,193 ; movdqa %xmm1,%xmm0
DB 102,65,15,219,193 ; pand %xmm9,%xmm0
DB 102,15,239,200 ; pxor %xmm0,%xmm1
@@ -17746,11 +18122,11 @@ _sk_gather_f16_sse2 LABEL PROC
DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13
DB 102,65,15,114,245,13 ; pslld $0xd,%xmm13
DB 102,68,15,235,232 ; por %xmm0,%xmm13
- DB 102,68,15,111,29,69,22,0,0 ; movdqa 0x1645(%rip),%xmm11 # 4770 <_sk_callback_sse2+0x45b>
+ DB 102,68,15,111,29,167,22,0,0 ; movdqa 0x16a7(%rip),%xmm11 # 47c0 <_sk_callback_sse2+0x4bd>
DB 102,69,15,254,235 ; paddd %xmm11,%xmm13
- DB 102,68,15,111,37,71,22,0,0 ; movdqa 0x1647(%rip),%xmm12 # 4780 <_sk_callback_sse2+0x46b>
+ DB 102,68,15,111,37,169,22,0,0 ; movdqa 0x16a9(%rip),%xmm12 # 47d0 <_sk_callback_sse2+0x4cd>
DB 102,65,15,239,204 ; pxor %xmm12,%xmm1
- DB 102,15,111,29,74,22,0,0 ; movdqa 0x164a(%rip),%xmm3 # 4790 <_sk_callback_sse2+0x47b>
+ DB 102,15,111,29,172,22,0,0 ; movdqa 0x16ac(%rip),%xmm3 # 47e0 <_sk_callback_sse2+0x4dd>
DB 102,15,111,195 ; movdqa %xmm3,%xmm0
DB 102,15,102,193 ; pcmpgtd %xmm1,%xmm0
DB 102,65,15,223,197 ; pandn %xmm13,%xmm0
@@ -17801,17 +18177,17 @@ PUBLIC _sk_store_f16_sse2
_sk_store_f16_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
- DB 102,68,15,111,21,114,21,0,0 ; movdqa 0x1572(%rip),%xmm10 # 47a0 <_sk_callback_sse2+0x48b>
+ DB 102,68,15,111,21,212,21,0,0 ; movdqa 0x15d4(%rip),%xmm10 # 47f0 <_sk_callback_sse2+0x4ed>
DB 102,68,15,111,224 ; movdqa %xmm0,%xmm12
DB 102,68,15,111,232 ; movdqa %xmm0,%xmm13
DB 102,69,15,219,234 ; pand %xmm10,%xmm13
DB 102,69,15,239,229 ; pxor %xmm13,%xmm12
- DB 102,68,15,111,13,101,21,0,0 ; movdqa 0x1565(%rip),%xmm9 # 47b0 <_sk_callback_sse2+0x49b>
+ DB 102,68,15,111,13,199,21,0,0 ; movdqa 0x15c7(%rip),%xmm9 # 4800 <_sk_callback_sse2+0x4fd>
DB 102,65,15,114,213,16 ; psrld $0x10,%xmm13
DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8
DB 102,69,15,102,196 ; pcmpgtd %xmm12,%xmm8
DB 102,65,15,114,212,13 ; psrld $0xd,%xmm12
- DB 102,68,15,111,29,86,21,0,0 ; movdqa 0x1556(%rip),%xmm11 # 47c0 <_sk_callback_sse2+0x4ab>
+ DB 102,68,15,111,29,184,21,0,0 ; movdqa 0x15b8(%rip),%xmm11 # 4810 <_sk_callback_sse2+0x50d>
DB 102,69,15,235,235 ; por %xmm11,%xmm13
DB 102,69,15,254,236 ; paddd %xmm12,%xmm13
DB 102,65,15,114,245,16 ; pslld $0x10,%xmm13
@@ -18111,7 +18487,7 @@ _sk_repeat_x_sse2 LABEL PROC
DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10
DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10
DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9
- DB 68,15,84,13,135,16,0,0 ; andps 0x1087(%rip),%xmm9 # 47d0 <_sk_callback_sse2+0x4bb>
+ DB 68,15,84,13,233,16,0,0 ; andps 0x10e9(%rip),%xmm9 # 4820 <_sk_callback_sse2+0x51d>
DB 69,15,92,209 ; subps %xmm9,%xmm10
DB 69,15,89,208 ; mulps %xmm8,%xmm10
DB 65,15,92,194 ; subps %xmm10,%xmm0
@@ -18131,7 +18507,7 @@ _sk_repeat_y_sse2 LABEL PROC
DB 243,69,15,91,209 ; cvttps2dq %xmm9,%xmm10
DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10
DB 69,15,194,202,1 ; cmpltps %xmm10,%xmm9
- DB 68,15,84,13,79,16,0,0 ; andps 0x104f(%rip),%xmm9 # 47e0 <_sk_callback_sse2+0x4cb>
+ DB 68,15,84,13,177,16,0,0 ; andps 0x10b1(%rip),%xmm9 # 4830 <_sk_callback_sse2+0x52d>
DB 69,15,92,209 ; subps %xmm9,%xmm10
DB 69,15,89,208 ; mulps %xmm8,%xmm10
DB 65,15,92,202 ; subps %xmm10,%xmm1
@@ -18155,7 +18531,7 @@ _sk_mirror_x_sse2 LABEL PROC
DB 243,69,15,91,218 ; cvttps2dq %xmm10,%xmm11
DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11
DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10
- DB 68,15,84,21,5,16,0,0 ; andps 0x1005(%rip),%xmm10 # 47f0 <_sk_callback_sse2+0x4db>
+ DB 68,15,84,21,103,16,0,0 ; andps 0x1067(%rip),%xmm10 # 4840 <_sk_callback_sse2+0x53d>
DB 69,15,87,228 ; xorps %xmm12,%xmm12
DB 69,15,92,218 ; subps %xmm10,%xmm11
DB 69,15,89,216 ; mulps %xmm8,%xmm11
@@ -18183,7 +18559,7 @@ _sk_mirror_y_sse2 LABEL PROC
DB 243,69,15,91,218 ; cvttps2dq %xmm10,%xmm11
DB 69,15,91,219 ; cvtdq2ps %xmm11,%xmm11
DB 69,15,194,211,1 ; cmpltps %xmm11,%xmm10
- DB 68,15,84,21,171,15,0,0 ; andps 0xfab(%rip),%xmm10 # 4800 <_sk_callback_sse2+0x4eb>
+ DB 68,15,84,21,13,16,0,0 ; andps 0x100d(%rip),%xmm10 # 4850 <_sk_callback_sse2+0x54d>
DB 69,15,87,228 ; xorps %xmm12,%xmm12
DB 69,15,92,218 ; subps %xmm10,%xmm11
DB 69,15,89,216 ; mulps %xmm8,%xmm11
@@ -18434,7 +18810,7 @@ _sk_linear_gradient_sse2 LABEL PROC
DB 69,15,198,228,0 ; shufps $0x0,%xmm12,%xmm12
DB 72,139,8 ; mov (%rax),%rcx
DB 72,133,201 ; test %rcx,%rcx
- DB 15,132,15,1,0,0 ; je 3d69 <_sk_linear_gradient_sse2+0x149>
+ DB 15,132,15,1,0,0 ; je 3d57 <_sk_linear_gradient_sse2+0x149>
DB 72,139,64,8 ; mov 0x8(%rax),%rax
DB 72,131,192,32 ; add $0x20,%rax
DB 69,15,87,192 ; xorps %xmm8,%xmm8
@@ -18495,8 +18871,8 @@ _sk_linear_gradient_sse2 LABEL PROC
DB 69,15,86,231 ; orps %xmm15,%xmm12
DB 72,131,192,36 ; add $0x24,%rax
DB 72,255,201 ; dec %rcx
- DB 15,133,8,255,255,255 ; jne 3c6f <_sk_linear_gradient_sse2+0x4f>
- DB 235,13 ; jmp 3d76 <_sk_linear_gradient_sse2+0x156>
+ DB 15,133,8,255,255,255 ; jne 3c5d <_sk_linear_gradient_sse2+0x4f>
+ DB 235,13 ; jmp 3d64 <_sk_linear_gradient_sse2+0x156>
DB 15,87,201 ; xorps %xmm1,%xmm1
DB 15,87,210 ; xorps %xmm2,%xmm2
DB 15,87,219 ; xorps %xmm3,%xmm3
@@ -18556,7 +18932,7 @@ _sk_save_xy_sse2 LABEL PROC
DB 69,15,91,210 ; cvtdq2ps %xmm10,%xmm10
DB 69,15,40,217 ; movaps %xmm9,%xmm11
DB 69,15,194,218,1 ; cmpltps %xmm10,%xmm11
- DB 68,15,40,37,210,9,0,0 ; movaps 0x9d2(%rip),%xmm12 # 4810 <_sk_callback_sse2+0x4fb>
+ DB 68,15,40,37,52,10,0,0 ; movaps 0xa34(%rip),%xmm12 # 4860 <_sk_callback_sse2+0x55d>
DB 69,15,84,220 ; andps %xmm12,%xmm11
DB 69,15,92,211 ; subps %xmm11,%xmm10
DB 69,15,92,202 ; subps %xmm10,%xmm9
@@ -18950,7 +19326,31 @@ ALIGN 16
DB 0,128,63,0,0,128 ; add %al,-0x7fffffc1(%rax)
DB 63 ; (bad)
DB 0,0 ; add %al,(%rax)
- DB 128,63,0 ; cmpb $0x0,(%rdi)
+ DB 128,63,255 ; cmpb $0xff,(%rdi)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,255 ; add %bh,%bh
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,0 ; add %al,(%rax)
DB 0,0 ; add %al,(%rax)
DB 52,0 ; xor $0x0,%al
DB 0,0 ; add %al,(%rax)
@@ -18960,16 +19360,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 43f4 <.literal16+0x34>
+ DB 127,0 ; jg 4414 <.literal16+0x64>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 43f8 <.literal16+0x38>
+ DB 127,0 ; jg 4418 <.literal16+0x68>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 43fc <.literal16+0x3c>
+ DB 127,0 ; jg 441c <.literal16+0x6c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4400 <.literal16+0x40>
+ DB 127,0 ; jg 4420 <.literal16+0x70>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -18978,7 +19378,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4485 <.literal16+0xc5>
+ DB 119,115 ; ja 44a5 <.literal16+0xf5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -18989,7 +19389,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 43e9 <.literal16+0x29>
+ DB 117,191 ; jne 4409 <.literal16+0x59>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -19001,7 +19401,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a3842a <_sk_callback_sse2+0xffffffffe9a34115>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a3844a <_sk_callback_sse2+0xffffffffe9a34147>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -19055,16 +19455,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 44c4 <.literal16+0x104>
+ DB 127,0 ; jg 44e4 <.literal16+0x134>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 44c8 <.literal16+0x108>
+ DB 127,0 ; jg 44e8 <.literal16+0x138>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 44cc <.literal16+0x10c>
+ DB 127,0 ; jg 44ec <.literal16+0x13c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 44d0 <.literal16+0x110>
+ DB 127,0 ; jg 44f0 <.literal16+0x140>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -19073,7 +19473,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4555 <.literal16+0x195>
+ DB 119,115 ; ja 4575 <.literal16+0x1c5>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -19084,7 +19484,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 44b9 <.literal16+0xf9>
+ DB 117,191 ; jne 44d9 <.literal16+0x129>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -19096,7 +19496,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a384fa <_sk_callback_sse2+0xffffffffe9a341e5>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a3851a <_sk_callback_sse2+0xffffffffe9a34217>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -19150,16 +19550,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4594 <.literal16+0x1d4>
+ DB 127,0 ; jg 45b4 <.literal16+0x204>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4598 <.literal16+0x1d8>
+ DB 127,0 ; jg 45b8 <.literal16+0x208>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 459c <.literal16+0x1dc>
+ DB 127,0 ; jg 45bc <.literal16+0x20c>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 45a0 <.literal16+0x1e0>
+ DB 127,0 ; jg 45c0 <.literal16+0x210>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -19168,7 +19568,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 4625 <.literal16+0x265>
+ DB 119,115 ; ja 4645 <.literal16+0x295>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -19179,7 +19579,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4589 <.literal16+0x1c9>
+ DB 117,191 ; jne 45a9 <.literal16+0x1f9>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -19191,7 +19591,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a385ca <_sk_callback_sse2+0xffffffffe9a342b5>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a385ea <_sk_callback_sse2+0xffffffffe9a342e7>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -19245,16 +19645,16 @@ ALIGN 16
DB 0,0 ; add %al,(%rax)
DB 52,255 ; xor $0xff,%al
DB 255 ; (bad)
- DB 127,0 ; jg 4664 <.literal16+0x2a4>
+ DB 127,0 ; jg 4684 <.literal16+0x2d4>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4668 <.literal16+0x2a8>
+ DB 127,0 ; jg 4688 <.literal16+0x2d8>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 466c <.literal16+0x2ac>
+ DB 127,0 ; jg 468c <.literal16+0x2dc>
DB 255 ; (bad)
DB 255 ; (bad)
- DB 127,0 ; jg 4670 <.literal16+0x2b0>
+ DB 127,0 ; jg 4690 <.literal16+0x2e0>
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
@@ -19263,7 +19663,7 @@ ALIGN 16
DB 0,63 ; add %bh,(%rdi)
DB 0,0 ; add %al,(%rax)
DB 0,63 ; add %bh,(%rdi)
- DB 119,115 ; ja 46f5 <.literal16+0x335>
+ DB 119,115 ; ja 4715 <.literal16+0x365>
DB 248 ; clc
DB 194,119,115 ; retq $0x7377
DB 248 ; clc
@@ -19274,7 +19674,7 @@ ALIGN 16
DB 194,117,191 ; retq $0xbf75
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
- DB 117,191 ; jne 4659 <.literal16+0x299>
+ DB 117,191 ; jne 4679 <.literal16+0x2c9>
DB 191,63,117,191,191 ; mov $0xbfbf753f,%edi
DB 63 ; (bad)
DB 249 ; stc
@@ -19286,7 +19686,7 @@ ALIGN 16
DB 249 ; stc
DB 68,180,62 ; rex.R mov $0x3e,%spl
DB 163,233,220,63,163,233,220,63,163 ; movabs %eax,0xa33fdce9a33fdce9
- DB 233,220,63,163,233 ; jmpq ffffffffe9a3869a <_sk_callback_sse2+0xffffffffe9a34385>
+ DB 233,220,63,163,233 ; jmpq ffffffffe9a386ba <_sk_callback_sse2+0xffffffffe9a343b7>
DB 220,63 ; fdivrl (%rdi)
DB 0,0 ; add %al,(%rax)
DB 128,63,0 ; cmpb $0x0,(%rdi)
@@ -19331,7 +19731,31 @@ ALIGN 16
DB 75,0,0 ; rex.WXB add %al,(%r8)
DB 0,75,0 ; add %cl,0x0(%rbx)
DB 0,0 ; add %al,(%rax)
- DB 75,0,128,0,0,0,128 ; rex.WXB add %al,-0x80000000(%r8)
+ DB 75,255,0 ; rex.WXB incq (%r8)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 255,0 ; incl (%rax)
+ DB 0,0 ; add %al,(%rax)
+ DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax)
DB 0,0 ; add %al,(%rax)
DB 0,128,0,0,0,128 ; add %al,-0x80000000(%rax)
DB 0,0 ; add %al,(%rax)