diff options
| author |  Florin Malita <fmalita@chromium.org> | 2017-08-13 19:38:14 -0400 |
|---|---|---|
| committer |  Skia Commit-Bot <skia-commit-bot@chromium.org> | 2017-08-14 00:00:58 +0000 |
| commit | 8ae12751f062783443b5bbe70e5645fe78922d5c (patch) | |
| tree | 79d85255b675786b548babbda06254e24637efc8 /src/jumper | |
| parent | 215922a1f10de855e0717b03321b2c125e055cd0 (diff) | |
lowp: lighten, difference, exclusion
Change-Id: I5773cf831c7e41a932bee1f2c6830085fb7db025
Reviewed-on: https://skia-review.googlesource.com/33764
Commit-Queue: Florin Malita <fmalita@chromium.org>
Reviewed-by: Mike Klein <mtklein@google.com>
Diffstat (limited to 'src/jumper')
| -rw-r--r-- | src/jumper/SkJumper.cpp | 5 | ||||
| -rw-r--r-- | src/jumper/SkJumper_generated.S | 1195 | ||||
| -rw-r--r-- | src/jumper/SkJumper_generated_win.S | 1175 | ||||
| -rw-r--r-- | src/jumper/SkJumper_stages_8bit.cpp | 28 |
4 files changed, 2167 insertions, 236 deletions
diff --git a/src/jumper/SkJumper.cpp b/src/jumper/SkJumper.cpp index 2e4dbe1215..8baa1890b6 100644 --- a/src/jumper/SkJumper.cpp +++ b/src/jumper/SkJumper.cpp @@ -131,7 +131,10 @@ using StartPipelineFn = void(size_t,size_t,size_t,size_t, void**,K*); M(multiply) \ M(screen) \ M(xor_) \ - M(darken) + M(darken) \ + M(lighten) \ + M(difference) \ + M(exclusion) #endif extern "C" { diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index 7fd41f9280..2f0417a5c2 100644 --- a/src/jumper/SkJumper_generated.S +++ b/src/jumper/SkJumper_generated.S @@ -56656,7 +56656,7 @@ HIDDEN _sk_set_rgb_hsw_8bit FUNCTION(_sk_set_rgb_hsw_8bit) _sk_set_rgb_hsw_8bit: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 197,250,16,37,194,40,0,0 // vmovss 0x28c2(%rip),%xmm4 # 2984 <_sk_darken_hsw_8bit+0x187> + .byte 197,250,16,37,66,45,0,0 // vmovss 0x2d42(%rip),%xmm4 # 2e04 <_sk_exclusion_hsw_8bit+0x16a> .byte 197,218,89,40 // vmulss (%rax),%xmm4,%xmm5 .byte 196,225,250,44,205 // vcvttss2si %xmm5,%rcx .byte 197,218,89,104,4 // vmulss 0x4(%rax),%xmm4,%xmm5 @@ -56669,7 +56669,7 @@ _sk_set_rgb_hsw_8bit: .byte 9,208 // or %edx,%eax .byte 197,249,110,224 // vmovd %eax,%xmm4 .byte 196,226,125,88,228 // vpbroadcastd %xmm4,%ymm4 - .byte 197,253,111,45,198,40,0,0 // vmovdqa 0x28c6(%rip),%ymm5 # 29c0 <_sk_darken_hsw_8bit+0x1c3> + .byte 197,253,111,45,70,45,0,0 // vmovdqa 0x2d46(%rip),%ymm5 # 2e40 <_sk_exclusion_hsw_8bit+0x1a6> .byte 197,245,219,205 // vpand %ymm5,%ymm1,%ymm1 .byte 197,253,219,197 // vpand %ymm5,%ymm0,%ymm0 .byte 197,221,235,192 // vpor %ymm0,%ymm4,%ymm0 @@ -56681,10 +56681,10 @@ HIDDEN _sk_premul_hsw_8bit .globl _sk_premul_hsw_8bit FUNCTION(_sk_premul_hsw_8bit) _sk_premul_hsw_8bit: - .byte 197,253,111,37,202,40,0,0 // vmovdqa 0x28ca(%rip),%ymm4 # 29e0 <_sk_darken_hsw_8bit+0x1e3> + .byte 197,253,111,37,74,45,0,0 // vmovdqa 0x2d4a(%rip),%ymm4 # 2e60 <_sk_exclusion_hsw_8bit+0x1c6> .byte 196,226,125,0,236 // vpshufb %ymm4,%ymm0,%ymm5 .byte 196,226,117,0,228 // vpshufb %ymm4,%ymm1,%ymm4 - .byte 197,253,111,53,216,40,0,0 // vmovdqa 0x28d8(%rip),%ymm6 # 2a00 <_sk_darken_hsw_8bit+0x203> + .byte 197,253,111,53,88,45,0,0 // vmovdqa 0x2d58(%rip),%ymm6 # 2e80 <_sk_exclusion_hsw_8bit+0x1e6> .byte 197,221,235,230 // vpor %ymm6,%ymm4,%ymm4 .byte 197,213,235,238 // vpor %ymm6,%ymm5,%ymm5 .byte 196,226,125,48,240 // vpmovzxbw %xmm0,%ymm6 @@ -56724,7 +56724,7 @@ HIDDEN _sk_swap_rb_hsw_8bit .globl _sk_swap_rb_hsw_8bit FUNCTION(_sk_swap_rb_hsw_8bit) _sk_swap_rb_hsw_8bit: - .byte 197,253,111,37,80,40,0,0 // vmovdqa 0x2850(%rip),%ymm4 # 2a20 <_sk_darken_hsw_8bit+0x223> + .byte 197,253,111,37,208,44,0,0 // vmovdqa 0x2cd0(%rip),%ymm4 # 2ea0 <_sk_exclusion_hsw_8bit+0x206> .byte 196,226,125,0,196 // vpshufb %ymm4,%ymm0,%ymm0 .byte 196,226,117,0,204 // vpshufb %ymm4,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -57074,7 +57074,7 @@ _sk_load_bgra_hsw_8bit: .byte 117,35 // jne 6b4 <_sk_load_bgra_hsw_8bit+0x44> .byte 196,161,126,111,76,130,32 // vmovdqu 0x20(%rdx,%r8,4),%ymm1 .byte 196,161,126,111,4,130 // vmovdqu (%rdx,%r8,4),%ymm0 - .byte 197,253,111,37,154,35,0,0 // vmovdqa 0x239a(%rip),%ymm4 # 2a40 <_sk_darken_hsw_8bit+0x243> + .byte 197,253,111,37,26,40,0,0 // vmovdqa 0x281a(%rip),%ymm4 # 2ec0 <_sk_exclusion_hsw_8bit+0x226> .byte 196,226,125,0,196 // vpshufb %ymm4,%ymm0,%ymm0 .byte 196,226,117,0,204 // vpshufb %ymm4,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -57189,7 +57189,7 @@ _sk_load_bgra_dst_hsw_8bit: .byte 117,35 // jne 86c <_sk_load_bgra_dst_hsw_8bit+0x44> .byte 196,161,126,111,92,130,32 // vmovdqu 0x20(%rdx,%r8,4),%ymm3 .byte 196,161,126,111,20,130 // vmovdqu (%rdx,%r8,4),%ymm2 - .byte 197,253,111,37,2,34,0,0 // vmovdqa 0x2202(%rip),%ymm4 # 2a60 <_sk_darken_hsw_8bit+0x263> + .byte 197,253,111,37,130,38,0,0 // vmovdqa 0x2682(%rip),%ymm4 # 2ee0 <_sk_exclusion_hsw_8bit+0x246> .byte 196,226,109,0,212 // vpshufb %ymm4,%ymm2,%ymm2 .byte 196,226,101,0,220 // vpshufb %ymm4,%ymm3,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -57300,7 +57300,7 @@ _sk_store_bgra_hsw_8bit: .byte 72,15,175,209 // imul %rcx,%rdx .byte 72,193,226,2 // shl $0x2,%rdx .byte 72,3,16 // add (%rax),%rdx - .byte 197,253,111,37,124,32,0,0 // vmovdqa 0x207c(%rip),%ymm4 # 2a80 <_sk_darken_hsw_8bit+0x283> + .byte 197,253,111,37,252,36,0,0 // vmovdqa 0x24fc(%rip),%ymm4 # 2f00 <_sk_exclusion_hsw_8bit+0x266> .byte 196,226,117,0,236 // vpshufb %ymm4,%ymm1,%ymm5 .byte 196,226,125,0,228 // vpshufb %ymm4,%ymm0,%ymm4 .byte 77,133,201 // test %r9,%r9 @@ -57586,10 +57586,10 @@ _sk_store_a8_hsw_8bit: .byte 72,99,87,8 // movslq 0x8(%rdi),%rdx .byte 72,15,175,209 // imul %rcx,%rdx .byte 72,3,16 // add (%rax),%rdx - .byte 197,253,111,37,168,28,0,0 // vmovdqa 0x1ca8(%rip),%ymm4 # 2aa0 <_sk_darken_hsw_8bit+0x2a3> + .byte 197,253,111,37,40,33,0,0 // vmovdqa 0x2128(%rip),%ymm4 # 2f20 <_sk_exclusion_hsw_8bit+0x286> .byte 196,226,117,0,236 // vpshufb %ymm4,%ymm1,%ymm5 .byte 196,227,253,0,237,232 // vpermq $0xe8,%ymm5,%ymm5 - .byte 197,249,111,53,181,30,0,0 // vmovdqa 0x1eb5(%rip),%xmm6 # 2cc0 <_sk_darken_hsw_8bit+0x4c3> + .byte 197,249,111,53,149,35,0,0 // vmovdqa 0x2395(%rip),%xmm6 # 31a0 <_sk_exclusion_hsw_8bit+0x506> .byte 196,226,81,0,238 // vpshufb %xmm6,%xmm5,%xmm5 .byte 196,226,125,0,228 // vpshufb %ymm4,%ymm0,%ymm4 .byte 196,227,253,0,228,232 // vpermq $0xe8,%ymm4,%ymm4 @@ -57681,10 +57681,10 @@ _sk_load_g8_hsw_8bit: .byte 196,226,125,49,200 // vpmovzxbd %xmm0,%ymm1 .byte 197,249,112,192,78 // vpshufd $0x4e,%xmm0,%xmm0 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 - .byte 196,226,125,88,37,61,26,0,0 // vpbroadcastd 0x1a3d(%rip),%ymm4 # 2988 <_sk_darken_hsw_8bit+0x18b> + .byte 196,226,125,88,37,189,30,0,0 // vpbroadcastd 0x1ebd(%rip),%ymm4 # 2e08 <_sk_exclusion_hsw_8bit+0x16e> .byte 196,226,125,64,236 // vpmulld %ymm4,%ymm0,%ymm5 .byte 196,226,117,64,196 // vpmulld %ymm4,%ymm1,%ymm0 - .byte 196,226,125,88,13,46,26,0,0 // vpbroadcastd 0x1a2e(%rip),%ymm1 # 298c <_sk_darken_hsw_8bit+0x18f> + .byte 196,226,125,88,13,174,30,0,0 // vpbroadcastd 0x1eae(%rip),%ymm1 # 2e0c <_sk_exclusion_hsw_8bit+0x172> .byte 197,253,235,193 // vpor %ymm1,%ymm0,%ymm0 .byte 197,213,235,201 // vpor %ymm1,%ymm5,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -57782,10 +57782,10 @@ _sk_load_g8_dst_hsw_8bit: .byte 196,226,125,49,218 // vpmovzxbd %xmm2,%ymm3 .byte 197,249,112,210,78 // vpshufd $0x4e,%xmm2,%xmm2 .byte 196,226,125,49,210 // vpmovzxbd %xmm2,%ymm2 - .byte 196,226,125,88,37,205,24,0,0 // vpbroadcastd 0x18cd(%rip),%ymm4 # 2990 <_sk_darken_hsw_8bit+0x193> + .byte 196,226,125,88,37,77,29,0,0 // vpbroadcastd 0x1d4d(%rip),%ymm4 # 2e10 <_sk_exclusion_hsw_8bit+0x176> .byte 196,226,109,64,236 // vpmulld %ymm4,%ymm2,%ymm5 .byte 196,226,101,64,212 // vpmulld %ymm4,%ymm3,%ymm2 - .byte 196,226,125,88,29,190,24,0,0 // vpbroadcastd 0x18be(%rip),%ymm3 # 2994 <_sk_darken_hsw_8bit+0x197> + .byte 196,226,125,88,29,62,29,0,0 // vpbroadcastd 0x1d3e(%rip),%ymm3 # 2e14 <_sk_exclusion_hsw_8bit+0x17a> .byte 197,237,235,211 // vpor %ymm3,%ymm2,%ymm2 .byte 197,213,235,219 // vpor %ymm3,%ymm5,%ymm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -57882,7 +57882,7 @@ _sk_srcover_rgba_8888_hsw_8bit: .byte 15,133,222,0,0,0 // jne 1303 <_sk_srcover_rgba_8888_hsw_8bit+0x103> .byte 196,33,126,111,76,138,32 // vmovdqu 0x20(%rdx,%r9,4),%ymm9 .byte 196,33,126,111,28,138 // vmovdqu (%rdx,%r9,4),%ymm11 - .byte 197,253,111,53,134,24,0,0 // vmovdqa 0x1886(%rip),%ymm6 # 2ac0 <_sk_darken_hsw_8bit+0x2c3> + .byte 197,253,111,53,6,29,0,0 // vmovdqa 0x1d06(%rip),%ymm6 # 2f40 <_sk_exclusion_hsw_8bit+0x2a6> .byte 196,226,117,0,254 // vpshufb %ymm6,%ymm1,%ymm7 .byte 196,226,125,0,246 // vpshufb %ymm6,%ymm0,%ymm6 .byte 196,66,125,48,195 // vpmovzxbw %xmm11,%ymm8 @@ -58090,7 +58090,7 @@ FUNCTION(_sk_scale_1_float_hsw_8bit) _sk_scale_1_float_hsw_8bit: .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,250,16,32 // vmovss (%rax),%xmm4 - .byte 197,218,89,37,2,20,0,0 // vmulss 0x1402(%rip),%xmm4,%xmm4 # 2998 <_sk_darken_hsw_8bit+0x19b> + .byte 197,218,89,37,130,24,0,0 // vmulss 0x1882(%rip),%xmm4,%xmm4 # 2e18 <_sk_exclusion_hsw_8bit+0x17e> .byte 197,250,44,196 // vcvttss2si %xmm4,%eax .byte 197,249,110,224 // vmovd %eax,%xmm4 .byte 196,226,125,120,228 // vpbroadcastb %xmm4,%ymm4 @@ -58100,7 +58100,7 @@ _sk_scale_1_float_hsw_8bit: .byte 196,226,125,48,241 // vpmovzxbw %xmm1,%ymm6 .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1 .byte 196,226,125,48,201 // vpmovzxbw %xmm1,%ymm1 - .byte 197,221,219,37,21,21,0,0 // vpand 0x1515(%rip),%ymm4,%ymm4 # 2ae0 <_sk_darken_hsw_8bit+0x2e3> + .byte 197,221,219,37,149,25,0,0 // vpand 0x1995(%rip),%ymm4,%ymm4 # 2f60 <_sk_exclusion_hsw_8bit+0x2c6> .byte 197,221,213,249 // vpmullw %ymm1,%ymm4,%ymm7 .byte 197,93,213,198 // vpmullw %ymm6,%ymm4,%ymm8 .byte 197,93,213,200 // vpmullw %ymm0,%ymm4,%ymm9 @@ -58139,7 +58139,7 @@ _sk_scale_u8_hsw_8bit: .byte 196,226,125,49,236 // vpmovzxbd %xmm4,%ymm5 .byte 197,249,112,228,78 // vpshufd $0x4e,%xmm4,%xmm4 .byte 196,226,125,49,228 // vpmovzxbd %xmm4,%ymm4 - .byte 197,253,111,53,159,20,0,0 // vmovdqa 0x149f(%rip),%ymm6 # 2b00 <_sk_darken_hsw_8bit+0x303> + .byte 197,253,111,53,31,25,0,0 // vmovdqa 0x191f(%rip),%ymm6 # 2f80 <_sk_exclusion_hsw_8bit+0x2e6> .byte 196,226,93,0,230 // vpshufb %ymm6,%ymm4,%ymm4 .byte 196,226,85,0,238 // vpshufb %ymm6,%ymm5,%ymm5 .byte 196,226,125,48,240 // vpmovzxbw %xmm0,%ymm6 @@ -58256,7 +58256,7 @@ FUNCTION(_sk_lerp_1_float_hsw_8bit) _sk_lerp_1_float_hsw_8bit: .byte 72,173 // lods %ds:(%rsi),%rax .byte 197,250,16,32 // vmovss (%rax),%xmm4 - .byte 197,218,89,37,102,17,0,0 // vmulss 0x1166(%rip),%xmm4,%xmm4 # 299c <_sk_darken_hsw_8bit+0x19f> + .byte 197,218,89,37,230,21,0,0 // vmulss 0x15e6(%rip),%xmm4,%xmm4 # 2e1c <_sk_exclusion_hsw_8bit+0x182> .byte 197,250,44,196 // vcvttss2si %xmm4,%eax .byte 197,249,110,224 // vmovd %eax,%xmm4 .byte 196,226,125,120,228 // vpbroadcastb %xmm4,%ymm4 @@ -58266,7 +58266,7 @@ _sk_lerp_1_float_hsw_8bit: .byte 196,226,125,48,241 // vpmovzxbw %xmm1,%ymm6 .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1 .byte 196,226,125,48,201 // vpmovzxbw %xmm1,%ymm1 - .byte 197,221,219,61,181,18,0,0 // vpand 0x12b5(%rip),%ymm4,%ymm7 # 2b20 <_sk_darken_hsw_8bit+0x323> + .byte 197,221,219,61,53,23,0,0 // vpand 0x1735(%rip),%ymm4,%ymm7 # 2fa0 <_sk_exclusion_hsw_8bit+0x306> .byte 197,69,213,193 // vpmullw %ymm1,%ymm7,%ymm8 .byte 197,69,213,206 // vpmullw %ymm6,%ymm7,%ymm9 .byte 197,69,213,208 // vpmullw %ymm0,%ymm7,%ymm10 @@ -58336,7 +58336,7 @@ _sk_lerp_u8_hsw_8bit: .byte 196,226,125,49,236 // vpmovzxbd %xmm4,%ymm5 .byte 197,249,112,228,78 // vpshufd $0x4e,%xmm4,%xmm4 .byte 196,226,125,49,228 // vpmovzxbd %xmm4,%ymm4 - .byte 197,253,111,53,170,17,0,0 // vmovdqa 0x11aa(%rip),%ymm6 # 2b40 <_sk_darken_hsw_8bit+0x343> + .byte 197,253,111,53,42,22,0,0 // vmovdqa 0x162a(%rip),%ymm6 # 2fc0 <_sk_exclusion_hsw_8bit+0x326> .byte 196,98,93,0,206 // vpshufb %ymm6,%ymm4,%ymm9 .byte 196,98,85,0,222 // vpshufb %ymm6,%ymm5,%ymm11 .byte 196,226,125,48,240 // vpmovzxbw %xmm0,%ymm6 @@ -58505,7 +58505,7 @@ HIDDEN _sk_black_color_hsw_8bit FUNCTION(_sk_black_color_hsw_8bit) _sk_black_color_hsw_8bit: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 196,226,125,24,5,117,13,0,0 // vbroadcastss 0xd75(%rip),%ymm0 # 29a0 <_sk_darken_hsw_8bit+0x1a3> + .byte 196,226,125,24,5,245,17,0,0 // vbroadcastss 0x11f5(%rip),%ymm0 # 2e20 <_sk_exclusion_hsw_8bit+0x186> .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 .byte 255,224 // jmpq *%rax @@ -58531,7 +58531,7 @@ HIDDEN _sk_srcatop_hsw_8bit .globl _sk_srcatop_hsw_8bit FUNCTION(_sk_srcatop_hsw_8bit) _sk_srcatop_hsw_8bit: - .byte 197,125,111,5,15,15,0,0 // vmovdqa 0xf0f(%rip),%ymm8 # 2b60 <_sk_darken_hsw_8bit+0x363> + .byte 197,125,111,5,143,19,0,0 // vmovdqa 0x138f(%rip),%ymm8 # 2fe0 <_sk_exclusion_hsw_8bit+0x346> .byte 196,194,101,0,224 // vpshufb %ymm8,%ymm3,%ymm4 .byte 196,194,109,0,232 // vpshufb %ymm8,%ymm2,%ymm5 .byte 196,98,125,48,208 // vpmovzxbw %xmm0,%ymm10 @@ -58608,7 +58608,7 @@ HIDDEN _sk_dstatop_hsw_8bit .globl _sk_dstatop_hsw_8bit FUNCTION(_sk_dstatop_hsw_8bit) _sk_dstatop_hsw_8bit: - .byte 197,125,111,5,202,13,0,0 // vmovdqa 0xdca(%rip),%ymm8 # 2b80 <_sk_darken_hsw_8bit+0x383> + .byte 197,125,111,5,74,18,0,0 // vmovdqa 0x124a(%rip),%ymm8 # 3000 <_sk_exclusion_hsw_8bit+0x366> .byte 196,194,117,0,224 // vpshufb %ymm8,%ymm1,%ymm4 .byte 196,194,125,0,232 // vpshufb %ymm8,%ymm0,%ymm5 .byte 196,98,125,48,210 // vpmovzxbw %xmm2,%ymm10 @@ -58685,7 +58685,7 @@ HIDDEN _sk_srcin_hsw_8bit .globl _sk_srcin_hsw_8bit FUNCTION(_sk_srcin_hsw_8bit) _sk_srcin_hsw_8bit: - .byte 197,253,111,37,129,12,0,0 // vmovdqa 0xc81(%rip),%ymm4 # 2ba0 <_sk_darken_hsw_8bit+0x3a3> + .byte 197,253,111,37,1,17,0,0 // vmovdqa 0x1101(%rip),%ymm4 # 3020 <_sk_exclusion_hsw_8bit+0x386> .byte 196,226,101,0,236 // vpshufb %ymm4,%ymm3,%ymm5 .byte 196,226,109,0,228 // vpshufb %ymm4,%ymm2,%ymm4 .byte 196,226,125,48,240 // vpmovzxbw %xmm0,%ymm6 @@ -58725,7 +58725,7 @@ HIDDEN _sk_dstin_hsw_8bit .globl _sk_dstin_hsw_8bit FUNCTION(_sk_dstin_hsw_8bit) _sk_dstin_hsw_8bit: - .byte 197,253,111,37,247,11,0,0 // vmovdqa 0xbf7(%rip),%ymm4 # 2bc0 <_sk_darken_hsw_8bit+0x3c3> + .byte 197,253,111,37,119,16,0,0 // vmovdqa 0x1077(%rip),%ymm4 # 3040 <_sk_exclusion_hsw_8bit+0x3a6> .byte 196,226,117,0,204 // vpshufb %ymm4,%ymm1,%ymm1 .byte 196,226,125,0,196 // vpshufb %ymm4,%ymm0,%ymm0 .byte 196,226,125,48,226 // vpmovzxbw %xmm2,%ymm4 @@ -58765,7 +58765,7 @@ HIDDEN _sk_srcout_hsw_8bit .globl _sk_srcout_hsw_8bit FUNCTION(_sk_srcout_hsw_8bit) _sk_srcout_hsw_8bit: - .byte 197,253,111,37,107,11,0,0 // vmovdqa 0xb6b(%rip),%ymm4 # 2be0 <_sk_darken_hsw_8bit+0x3e3> + .byte 197,253,111,37,235,15,0,0 // vmovdqa 0xfeb(%rip),%ymm4 # 3060 <_sk_exclusion_hsw_8bit+0x3c6> .byte 196,226,109,0,236 // vpshufb %ymm4,%ymm2,%ymm5 .byte 196,226,101,0,228 // vpshufb %ymm4,%ymm3,%ymm4 .byte 197,205,118,246 // vpcmpeqd %ymm6,%ymm6,%ymm6 @@ -58808,7 +58808,7 @@ HIDDEN _sk_dstout_hsw_8bit .globl _sk_dstout_hsw_8bit FUNCTION(_sk_dstout_hsw_8bit) _sk_dstout_hsw_8bit: - .byte 197,253,111,37,213,10,0,0 // vmovdqa 0xad5(%rip),%ymm4 # 2c00 <_sk_darken_hsw_8bit+0x403> + .byte 197,253,111,37,85,15,0,0 // vmovdqa 0xf55(%rip),%ymm4 # 3080 <_sk_exclusion_hsw_8bit+0x3e6> .byte 196,226,125,0,196 // vpshufb %ymm4,%ymm0,%ymm0 .byte 196,226,117,0,204 // vpshufb %ymm4,%ymm1,%ymm1 .byte 197,221,118,228 // vpcmpeqd %ymm4,%ymm4,%ymm4 @@ -58851,7 +58851,7 @@ HIDDEN _sk_srcover_hsw_8bit .globl _sk_srcover_hsw_8bit FUNCTION(_sk_srcover_hsw_8bit) _sk_srcover_hsw_8bit: - .byte 197,253,111,37,61,10,0,0 // vmovdqa 0xa3d(%rip),%ymm4 # 2c20 <_sk_darken_hsw_8bit+0x423> + .byte 197,253,111,37,189,14,0,0 // vmovdqa 0xebd(%rip),%ymm4 # 30a0 <_sk_exclusion_hsw_8bit+0x406> .byte 196,226,117,0,236 // vpshufb %ymm4,%ymm1,%ymm5 .byte 196,226,125,0,228 // vpshufb %ymm4,%ymm0,%ymm4 .byte 196,98,125,48,202 // vpmovzxbw %xmm2,%ymm9 @@ -58895,7 +58895,7 @@ HIDDEN _sk_dstover_hsw_8bit .globl _sk_dstover_hsw_8bit FUNCTION(_sk_dstover_hsw_8bit) _sk_dstover_hsw_8bit: - .byte 197,253,111,37,158,9,0,0 // vmovdqa 0x99e(%rip),%ymm4 # 2c40 <_sk_darken_hsw_8bit+0x443> + .byte 197,253,111,37,30,14,0,0 // vmovdqa 0xe1e(%rip),%ymm4 # 30c0 <_sk_exclusion_hsw_8bit+0x426> .byte 196,226,101,0,236 // vpshufb %ymm4,%ymm3,%ymm5 .byte 196,226,109,0,228 // vpshufb %ymm4,%ymm2,%ymm4 .byte 196,98,125,48,200 // vpmovzxbw %xmm0,%ymm9 @@ -58978,7 +58978,7 @@ FUNCTION(_sk_multiply_hsw_8bit) _sk_multiply_hsw_8bit: .byte 197,253,111,243 // vmovdqa %ymm3,%ymm6 .byte 197,253,111,218 // vmovdqa %ymm2,%ymm3 - .byte 197,125,111,13,95,8,0,0 // vmovdqa 0x85f(%rip),%ymm9 # 2c60 <_sk_darken_hsw_8bit+0x463> + .byte 197,125,111,13,223,12,0,0 // vmovdqa 0xcdf(%rip),%ymm9 # 30e0 <_sk_exclusion_hsw_8bit+0x446> .byte 196,194,101,0,225 // vpshufb %ymm9,%ymm3,%ymm4 .byte 196,194,77,0,233 // vpshufb %ymm9,%ymm6,%ymm5 .byte 196,65,45,118,210 // vpcmpeqd %ymm10,%ymm10,%ymm10 @@ -59122,7 +59122,7 @@ HIDDEN _sk_xor__hsw_8bit .globl _sk_xor__hsw_8bit FUNCTION(_sk_xor__hsw_8bit) _sk_xor__hsw_8bit: - .byte 197,125,111,13,237,5,0,0 // vmovdqa 0x5ed(%rip),%ymm9 # 2c80 <_sk_darken_hsw_8bit+0x483> + .byte 197,125,111,13,109,10,0,0 // vmovdqa 0xa6d(%rip),%ymm9 # 3100 <_sk_exclusion_hsw_8bit+0x466> .byte 196,194,109,0,225 // vpshufb %ymm9,%ymm2,%ymm4 .byte 196,194,101,0,249 // vpshufb %ymm9,%ymm3,%ymm7 .byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11 @@ -59201,7 +59201,7 @@ HIDDEN _sk_darken_hsw_8bit .globl _sk_darken_hsw_8bit FUNCTION(_sk_darken_hsw_8bit) _sk_darken_hsw_8bit: - .byte 197,125,111,5,155,4,0,0 // vmovdqa 0x49b(%rip),%ymm8 # 2ca0 <_sk_darken_hsw_8bit+0x4a3> + .byte 197,125,111,5,27,9,0,0 // vmovdqa 0x91b(%rip),%ymm8 # 3120 <_sk_exclusion_hsw_8bit+0x486> .byte 196,194,101,0,224 // vpshufb %ymm8,%ymm3,%ymm4 .byte 196,194,109,0,240 // vpshufb %ymm8,%ymm2,%ymm6 .byte 196,98,125,48,208 // vpmovzxbw %xmm0,%ymm10 @@ -59274,15 +59274,258 @@ _sk_darken_hsw_8bit: .byte 197,253,248,246 // vpsubb %ymm6,%ymm0,%ymm6 .byte 197,245,248,205 // vpsubb %ymm5,%ymm1,%ymm1 .byte 197,253,248,196 // vpsubb %ymm4,%ymm0,%ymm0 - .byte 196,226,125,88,37,48,0,0,0 // vpbroadcastd 0x30(%rip),%ymm4 # 29a4 <_sk_darken_hsw_8bit+0x1a7> + .byte 196,226,125,88,37,176,4,0,0 // vpbroadcastd 0x4b0(%rip),%ymm4 # 2e24 <_sk_exclusion_hsw_8bit+0x18a> .byte 196,227,125,76,198,64 // vpblendvb %ymm4,%ymm6,%ymm0,%ymm0 .byte 196,227,117,76,207,64 // vpblendvb %ymm4,%ymm7,%ymm1,%ymm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax +HIDDEN _sk_lighten_hsw_8bit +.globl _sk_lighten_hsw_8bit +FUNCTION(_sk_lighten_hsw_8bit) +_sk_lighten_hsw_8bit: + .byte 197,125,111,5,180,7,0,0 // vmovdqa 0x7b4(%rip),%ymm8 # 3140 <_sk_exclusion_hsw_8bit+0x4a6> + .byte 196,194,101,0,224 // vpshufb %ymm8,%ymm3,%ymm4 + .byte 196,194,109,0,240 // vpshufb %ymm8,%ymm2,%ymm6 + .byte 196,98,125,48,208 // vpmovzxbw %xmm0,%ymm10 + .byte 196,227,125,57,197,1 // vextracti128 $0x1,%ymm0,%xmm5 + .byte 196,226,125,48,237 // vpmovzxbw %xmm5,%ymm5 + .byte 196,98,125,48,201 // vpmovzxbw %xmm1,%ymm9 + .byte 196,227,125,57,207,1 // vextracti128 $0x1,%ymm1,%xmm7 + .byte 196,226,125,48,255 // vpmovzxbw %xmm7,%ymm7 + .byte 196,98,125,48,222 // vpmovzxbw %xmm6,%ymm11 + .byte 196,227,125,57,246,1 // vextracti128 $0x1,%ymm6,%xmm6 + .byte 196,226,125,48,246 // vpmovzxbw %xmm6,%ymm6 + .byte 196,98,125,48,228 // vpmovzxbw %xmm4,%ymm12 + .byte 196,227,125,57,228,1 // vextracti128 $0x1,%ymm4,%xmm4 + .byte 196,226,125,48,228 // vpmovzxbw %xmm4,%ymm4 + .byte 197,221,213,231 // vpmullw %ymm7,%ymm4,%ymm4 + .byte 196,65,29,213,225 // vpmullw %ymm9,%ymm12,%ymm12 + .byte 197,205,213,245 // vpmullw %ymm5,%ymm6,%ymm6 + .byte 196,65,37,213,218 // vpmullw %ymm10,%ymm11,%ymm11 + .byte 196,65,37,253,210 // vpaddw %ymm10,%ymm11,%ymm10 + .byte 197,205,253,237 // vpaddw %ymm5,%ymm6,%ymm5 + .byte 196,193,29,253,241 // vpaddw %ymm9,%ymm12,%ymm6 + .byte 197,221,253,231 // vpaddw %ymm7,%ymm4,%ymm4 + .byte 197,197,113,212,8 // vpsrlw $0x8,%ymm4,%ymm7 + .byte 197,205,113,214,8 // vpsrlw $0x8,%ymm6,%ymm6 + .byte 197,221,113,213,8 // vpsrlw $0x8,%ymm5,%ymm4 + .byte 196,193,85,113,210,8 // vpsrlw $0x8,%ymm10,%ymm5 + .byte 196,99,85,56,204,1 // vinserti128 $0x1,%xmm4,%ymm5,%ymm9 + .byte 196,227,85,70,228,49 // vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + .byte 197,53,103,204 // vpackuswb %ymm4,%ymm9,%ymm9 + .byte 196,227,77,56,239,1 // vinserti128 $0x1,%xmm7,%ymm6,%ymm5 + .byte 196,227,77,70,247,49 // vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + .byte 197,85,103,222 // vpackuswb %ymm6,%ymm5,%ymm11 + .byte 196,194,117,0,232 // vpshufb %ymm8,%ymm1,%ymm5 + .byte 196,194,125,0,248 // vpshufb %ymm8,%ymm0,%ymm7 + .byte 196,98,125,48,194 // vpmovzxbw %xmm2,%ymm8 + .byte 196,227,125,57,212,1 // vextracti128 $0x1,%ymm2,%xmm4 + .byte 196,226,125,48,228 // vpmovzxbw %xmm4,%ymm4 + .byte 196,98,125,48,211 // vpmovzxbw %xmm3,%ymm10 + .byte 196,227,125,57,222,1 // vextracti128 $0x1,%ymm3,%xmm6 + .byte 196,226,125,48,246 // vpmovzxbw %xmm6,%ymm6 + .byte 196,98,125,48,231 // vpmovzxbw %xmm7,%ymm12 + .byte 196,227,125,57,255,1 // vextracti128 $0x1,%ymm7,%xmm7 + .byte 196,226,125,48,255 // vpmovzxbw %xmm7,%ymm7 + .byte 196,98,125,48,237 // vpmovzxbw %xmm5,%ymm13 + .byte 196,227,125,57,237,1 // vextracti128 $0x1,%ymm5,%xmm5 + .byte 196,226,125,48,237 // vpmovzxbw %xmm5,%ymm5 + .byte 197,205,213,237 // vpmullw %ymm5,%ymm6,%ymm5 + .byte 196,65,45,213,237 // vpmullw %ymm13,%ymm10,%ymm13 + .byte 197,221,213,255 // vpmullw %ymm7,%ymm4,%ymm7 + .byte 196,65,61,213,228 // vpmullw %ymm12,%ymm8,%ymm12 + .byte 196,65,29,253,192 // vpaddw %ymm8,%ymm12,%ymm8 + .byte 197,197,253,228 // vpaddw %ymm4,%ymm7,%ymm4 + .byte 196,193,21,253,250 // vpaddw %ymm10,%ymm13,%ymm7 + .byte 197,213,253,238 // vpaddw %ymm6,%ymm5,%ymm5 + .byte 197,213,113,213,8 // vpsrlw $0x8,%ymm5,%ymm5 + .byte 197,205,113,215,8 // vpsrlw $0x8,%ymm7,%ymm6 + .byte 197,221,113,212,8 // vpsrlw $0x8,%ymm4,%ymm4 + .byte 196,193,69,113,208,8 // vpsrlw $0x8,%ymm8,%ymm7 + .byte 196,99,69,56,196,1 // vinserti128 $0x1,%xmm4,%ymm7,%ymm8 + .byte 196,227,69,70,228,49 // vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + .byte 197,189,103,228 // vpackuswb %ymm4,%ymm8,%ymm4 + .byte 196,227,77,56,253,1 // vinserti128 $0x1,%xmm5,%ymm6,%ymm7 + .byte 196,227,77,70,237,49 // vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + .byte 197,197,103,237 // vpackuswb %ymm5,%ymm7,%ymm5 + .byte 197,181,218,244 // vpminub %ymm4,%ymm9,%ymm6 + .byte 197,165,218,253 // vpminub %ymm5,%ymm11,%ymm7 + .byte 197,237,252,192 // vpaddb %ymm0,%ymm2,%ymm0 + .byte 197,229,252,201 // vpaddb %ymm1,%ymm3,%ymm1 + .byte 197,245,248,255 // vpsubb %ymm7,%ymm1,%ymm7 + .byte 197,253,248,246 // vpsubb %ymm6,%ymm0,%ymm6 + .byte 197,245,248,205 // vpsubb %ymm5,%ymm1,%ymm1 + .byte 197,253,248,196 // vpsubb %ymm4,%ymm0,%ymm0 + .byte 196,226,125,88,37,45,3,0,0 // vpbroadcastd 0x32d(%rip),%ymm4 # 2e28 <_sk_exclusion_hsw_8bit+0x18e> + .byte 196,227,125,76,198,64 // vpblendvb %ymm4,%ymm6,%ymm0,%ymm0 + .byte 196,227,117,76,207,64 // vpblendvb %ymm4,%ymm7,%ymm1,%ymm1 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_difference_hsw_8bit +.globl _sk_difference_hsw_8bit +FUNCTION(_sk_difference_hsw_8bit) +_sk_difference_hsw_8bit: + .byte 197,125,111,5,77,6,0,0 // vmovdqa 0x64d(%rip),%ymm8 # 3160 <_sk_exclusion_hsw_8bit+0x4c6> + .byte 196,194,101,0,224 // vpshufb %ymm8,%ymm3,%ymm4 + .byte 196,194,109,0,240 // vpshufb %ymm8,%ymm2,%ymm6 + .byte 196,98,125,48,208 // vpmovzxbw %xmm0,%ymm10 + .byte 196,227,125,57,197,1 // vextracti128 $0x1,%ymm0,%xmm5 + .byte 196,226,125,48,237 // vpmovzxbw %xmm5,%ymm5 + .byte 196,98,125,48,201 // vpmovzxbw %xmm1,%ymm9 + .byte 196,227,125,57,207,1 // vextracti128 $0x1,%ymm1,%xmm7 + .byte 196,226,125,48,255 // vpmovzxbw %xmm7,%ymm7 + .byte 196,98,125,48,222 // vpmovzxbw %xmm6,%ymm11 + .byte 196,227,125,57,246,1 // vextracti128 $0x1,%ymm6,%xmm6 + .byte 196,226,125,48,246 // vpmovzxbw %xmm6,%ymm6 + .byte 196,98,125,48,228 // vpmovzxbw %xmm4,%ymm12 + .byte 196,227,125,57,228,1 // vextracti128 $0x1,%ymm4,%xmm4 + .byte 196,226,125,48,228 // vpmovzxbw %xmm4,%ymm4 + .byte 197,221,213,231 // vpmullw %ymm7,%ymm4,%ymm4 + .byte 196,65,29,213,225 // vpmullw %ymm9,%ymm12,%ymm12 + .byte 197,205,213,245 // vpmullw %ymm5,%ymm6,%ymm6 + .byte 196,65,37,213,218 // vpmullw %ymm10,%ymm11,%ymm11 + .byte 196,65,37,253,210 // vpaddw %ymm10,%ymm11,%ymm10 + .byte 197,205,253,237 // vpaddw %ymm5,%ymm6,%ymm5 + .byte 196,193,29,253,241 // vpaddw %ymm9,%ymm12,%ymm6 + .byte 197,221,253,231 // vpaddw %ymm7,%ymm4,%ymm4 + .byte 197,197,113,212,8 // vpsrlw $0x8,%ymm4,%ymm7 + .byte 197,205,113,214,8 // vpsrlw $0x8,%ymm6,%ymm6 + .byte 197,221,113,213,8 // vpsrlw $0x8,%ymm5,%ymm4 + .byte 196,193,85,113,210,8 // vpsrlw $0x8,%ymm10,%ymm5 + .byte 196,99,85,56,204,1 // vinserti128 $0x1,%xmm4,%ymm5,%ymm9 + .byte 196,227,85,70,228,49 // vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + .byte 197,53,103,204 // vpackuswb %ymm4,%ymm9,%ymm9 + .byte 196,227,77,56,239,1 // vinserti128 $0x1,%xmm7,%ymm6,%ymm5 + .byte 196,227,77,70,247,49 // vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + .byte 197,85,103,222 // vpackuswb %ymm6,%ymm5,%ymm11 + .byte 196,194,117,0,232 // vpshufb %ymm8,%ymm1,%ymm5 + .byte 196,194,125,0,248 // vpshufb %ymm8,%ymm0,%ymm7 + .byte 196,98,125,48,194 // vpmovzxbw %xmm2,%ymm8 + .byte 196,227,125,57,212,1 // vextracti128 $0x1,%ymm2,%xmm4 + .byte 196,226,125,48,228 // vpmovzxbw %xmm4,%ymm4 + .byte 196,98,125,48,211 // vpmovzxbw %xmm3,%ymm10 + .byte 196,227,125,57,222,1 // vextracti128 $0x1,%ymm3,%xmm6 + .byte 196,226,125,48,246 // vpmovzxbw %xmm6,%ymm6 + .byte 196,98,125,48,231 // vpmovzxbw %xmm7,%ymm12 + .byte 196,227,125,57,255,1 // vextracti128 $0x1,%ymm7,%xmm7 + .byte 196,226,125,48,255 // vpmovzxbw %xmm7,%ymm7 + .byte 196,98,125,48,237 // vpmovzxbw %xmm5,%ymm13 + .byte 196,227,125,57,237,1 // vextracti128 $0x1,%ymm5,%xmm5 + .byte 196,226,125,48,237 // vpmovzxbw %xmm5,%ymm5 + .byte 197,205,213,237 // vpmullw %ymm5,%ymm6,%ymm5 + .byte 196,65,45,213,237 // vpmullw %ymm13,%ymm10,%ymm13 + .byte 197,221,213,255 // vpmullw %ymm7,%ymm4,%ymm7 + .byte 196,65,61,213,228 // vpmullw %ymm12,%ymm8,%ymm12 + .byte 196,65,29,253,192 // vpaddw %ymm8,%ymm12,%ymm8 + .byte 197,197,253,228 // vpaddw %ymm4,%ymm7,%ymm4 + .byte 196,193,21,253,250 // vpaddw %ymm10,%ymm13,%ymm7 + .byte 197,213,253,238 // vpaddw %ymm6,%ymm5,%ymm5 + .byte 197,213,113,213,8 // vpsrlw $0x8,%ymm5,%ymm5 + .byte 197,205,113,215,8 // vpsrlw $0x8,%ymm7,%ymm6 + .byte 197,221,113,212,8 // vpsrlw $0x8,%ymm4,%ymm4 + .byte 196,193,69,113,208,8 // vpsrlw $0x8,%ymm8,%ymm7 + .byte 196,99,69,56,196,1 // vinserti128 $0x1,%xmm4,%ymm7,%ymm8 + .byte 196,227,69,70,228,49 // vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + .byte 197,189,103,228 // vpackuswb %ymm4,%ymm8,%ymm4 + .byte 196,227,77,56,253,1 // vinserti128 $0x1,%xmm5,%ymm6,%ymm7 + .byte 196,227,77,70,237,49 // vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + .byte 197,197,103,237 // vpackuswb %ymm5,%ymm7,%ymm5 + .byte 197,165,218,245 // vpminub %ymm5,%ymm11,%ymm6 + .byte 197,181,218,252 // vpminub %ymm4,%ymm9,%ymm7 + .byte 197,197,252,255 // vpaddb %ymm7,%ymm7,%ymm7 + .byte 197,205,252,246 // vpaddb %ymm6,%ymm6,%ymm6 + .byte 197,237,252,192 // vpaddb %ymm0,%ymm2,%ymm0 + .byte 197,229,252,201 // vpaddb %ymm1,%ymm3,%ymm1 + .byte 197,245,248,246 // vpsubb %ymm6,%ymm1,%ymm6 + .byte 197,253,248,255 // vpsubb %ymm7,%ymm0,%ymm7 + .byte 197,245,248,205 // vpsubb %ymm5,%ymm1,%ymm1 + .byte 197,253,248,196 // vpsubb %ymm4,%ymm0,%ymm0 + .byte 196,226,125,88,37,162,1,0,0 // vpbroadcastd 0x1a2(%rip),%ymm4 # 2e2c <_sk_exclusion_hsw_8bit+0x192> + .byte 196,227,125,76,199,64 // vpblendvb %ymm4,%ymm7,%ymm0,%ymm0 + .byte 196,227,117,76,206,64 // vpblendvb %ymm4,%ymm6,%ymm1,%ymm1 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_exclusion_hsw_8bit +.globl _sk_exclusion_hsw_8bit +FUNCTION(_sk_exclusion_hsw_8bit) +_sk_exclusion_hsw_8bit: + .byte 196,98,125,48,192 // vpmovzxbw %xmm0,%ymm8 + .byte 196,227,125,57,196,1 // vextracti128 $0x1,%ymm0,%xmm4 + .byte 196,98,125,48,204 // vpmovzxbw %xmm4,%ymm9 + .byte 196,98,125,48,209 // vpmovzxbw %xmm1,%ymm10 + .byte 196,227,125,57,204,1 // vextracti128 $0x1,%ymm1,%xmm4 + .byte 196,98,125,48,220 // vpmovzxbw %xmm4,%ymm11 + .byte 196,226,125,48,226 // vpmovzxbw %xmm2,%ymm4 + .byte 196,227,125,57,213,1 // vextracti128 $0x1,%ymm2,%xmm5 + .byte 196,226,125,48,237 // vpmovzxbw %xmm5,%ymm5 + .byte 196,226,125,48,243 // vpmovzxbw %xmm3,%ymm6 + .byte 196,227,125,57,223,1 // vextracti128 $0x1,%ymm3,%xmm7 + .byte 196,226,125,48,255 // vpmovzxbw %xmm7,%ymm7 + .byte 196,65,69,213,227 // vpmullw %ymm11,%ymm7,%ymm12 + .byte 196,65,77,213,234 // vpmullw %ymm10,%ymm6,%ymm13 + .byte 196,65,85,213,241 // vpmullw %ymm9,%ymm5,%ymm14 + .byte 196,65,93,213,248 // vpmullw %ymm8,%ymm4,%ymm15 + .byte 196,65,5,253,192 // vpaddw %ymm8,%ymm15,%ymm8 + .byte 196,65,13,253,201 // vpaddw %ymm9,%ymm14,%ymm9 + .byte 196,65,21,253,210 // vpaddw %ymm10,%ymm13,%ymm10 + .byte 196,65,29,253,219 // vpaddw %ymm11,%ymm12,%ymm11 + .byte 196,193,37,113,211,8 // vpsrlw $0x8,%ymm11,%ymm11 + .byte 196,193,45,113,210,8 // vpsrlw $0x8,%ymm10,%ymm10 + .byte 196,193,53,113,209,8 // vpsrlw $0x8,%ymm9,%ymm9 + .byte 196,193,61,113,208,8 // vpsrlw $0x8,%ymm8,%ymm8 + .byte 196,67,61,56,225,1 // vinserti128 $0x1,%xmm9,%ymm8,%ymm12 + .byte 196,67,61,70,193,49 // vperm2i128 $0x31,%ymm9,%ymm8,%ymm8 + .byte 196,65,29,103,192 // vpackuswb %ymm8,%ymm12,%ymm8 + .byte 196,67,45,56,203,1 // vinserti128 $0x1,%xmm11,%ymm10,%ymm9 + .byte 196,67,45,70,211,49 // vperm2i128 $0x31,%ymm11,%ymm10,%ymm10 + .byte 196,65,53,103,202 // vpackuswb %ymm10,%ymm9,%ymm9 + .byte 196,65,61,252,208 // vpaddb %ymm8,%ymm8,%ymm10 + .byte 196,65,53,252,193 // vpaddb %ymm9,%ymm9,%ymm8 + .byte 197,109,252,200 // vpaddb %ymm0,%ymm2,%ymm9 + .byte 197,101,252,217 // vpaddb %ymm1,%ymm3,%ymm11 + .byte 196,65,37,248,192 // vpsubb %ymm8,%ymm11,%ymm8 + .byte 196,65,53,248,210 // vpsubb %ymm10,%ymm9,%ymm10 + .byte 197,125,111,37,32,4,0,0 // vmovdqa 0x420(%rip),%ymm12 # 3180 <_sk_exclusion_hsw_8bit+0x4e6> + .byte 196,194,117,0,204 // vpshufb %ymm12,%ymm1,%ymm1 + .byte 196,194,125,0,196 // vpshufb %ymm12,%ymm0,%ymm0 + .byte 196,98,125,48,224 // vpmovzxbw %xmm0,%ymm12 + .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 + .byte 196,226,125,48,192 // vpmovzxbw %xmm0,%ymm0 + .byte 196,98,125,48,233 // vpmovzxbw %xmm1,%ymm13 + .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1 + .byte 196,226,125,48,201 // vpmovzxbw %xmm1,%ymm1 + .byte 197,197,213,201 // vpmullw %ymm1,%ymm7,%ymm1 + .byte 196,65,77,213,237 // vpmullw %ymm13,%ymm6,%ymm13 + .byte 197,213,213,192 // vpmullw %ymm0,%ymm5,%ymm0 + .byte 196,65,93,213,228 // vpmullw %ymm12,%ymm4,%ymm12 + .byte 197,157,253,228 // vpaddw %ymm4,%ymm12,%ymm4 + .byte 197,253,253,197 // vpaddw %ymm5,%ymm0,%ymm0 + .byte 197,149,253,238 // vpaddw %ymm6,%ymm13,%ymm5 + .byte 197,245,253,207 // vpaddw %ymm7,%ymm1,%ymm1 + .byte 197,245,113,209,8 // vpsrlw $0x8,%ymm1,%ymm1 + .byte 197,213,113,213,8 // vpsrlw $0x8,%ymm5,%ymm5 + .byte 197,253,113,208,8 // vpsrlw $0x8,%ymm0,%ymm0 + .byte 197,221,113,212,8 // vpsrlw $0x8,%ymm4,%ymm4 + .byte 196,227,93,56,240,1 // vinserti128 $0x1,%xmm0,%ymm4,%ymm6 + .byte 196,227,93,70,192,49 // vperm2i128 $0x31,%ymm0,%ymm4,%ymm0 + .byte 197,205,103,192 // vpackuswb %ymm0,%ymm6,%ymm0 + .byte 196,227,85,56,225,1 // vinserti128 $0x1,%xmm1,%ymm5,%ymm4 + .byte 196,227,85,70,201,49 // vperm2i128 $0x31,%ymm1,%ymm5,%ymm1 + .byte 197,221,103,201 // vpackuswb %ymm1,%ymm4,%ymm1 + .byte 197,165,248,201 // vpsubb %ymm1,%ymm11,%ymm1 + .byte 197,181,248,192 // vpsubb %ymm0,%ymm9,%ymm0 + .byte 196,226,125,88,37,63,0,0,0 // vpbroadcastd 0x3f(%rip),%ymm4 # 2e30 <_sk_exclusion_hsw_8bit+0x196> + .byte 196,195,125,76,194,64 // vpblendvb %ymm4,%ymm10,%ymm0,%ymm0 + .byte 196,195,117,76,200,64 // vpblendvb %ymm4,%ymm8,%ymm1,%ymm1 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + BALIGN4 .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 29cb <_sk_darken_hsw_8bit+0x1ce> + .byte 127,67 // jg 2e4b <_sk_exclusion_hsw_8bit+0x1b1> .byte 1,1 // add %eax,(%rcx) .byte 1,0 // add %eax,(%rax) .byte 0,0 // add %al,(%rax) @@ -59292,14 +59535,23 @@ BALIGN4 .byte 0,0 // add %al,(%rax) .byte 0,255 // add %bh,%bh .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 29df <_sk_darken_hsw_8bit+0x1e2> + .byte 127,67 // jg 2e5f <_sk_exclusion_hsw_8bit+0x1c5> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 29e3 <_sk_darken_hsw_8bit+0x1e6> + .byte 127,67 // jg 2e63 <_sk_exclusion_hsw_8bit+0x1c9> .byte 0,0 // add %al,(%rax) .byte 0,255 // add %bh,%bh .byte 255 // (bad) .byte 255 // (bad) .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) BALIGN32 .byte 0,0 // add %al,(%rax) @@ -59754,6 +60006,72 @@ BALIGN32 .byte 11,11 // or (%rbx),%ecx .byte 11,11 // or (%rbx),%ecx .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,3,3 // lsl (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,3,3 // lsl (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,3,3 // lsl (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,3,3 // lsl (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,3,3 // lsl (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,3,3 // lsl (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) .byte 15 // .byte 0xf .byte 15 // .byte 0xf .byte 15 // .byte 0xf @@ -59851,7 +60169,7 @@ HIDDEN _sk_set_rgb_sse41_8bit FUNCTION(_sk_set_rgb_sse41_8bit) _sk_set_rgb_sse41_8bit: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,15,16,37,209,29,0,0 // movss 0x1dd1(%rip),%xmm4 # 1e94 <_sk_darken_sse41_8bit+0x170> + .byte 243,15,16,37,237,33,0,0 // movss 0x21ed(%rip),%xmm4 # 22b0 <_sk_exclusion_sse41_8bit+0x137> .byte 243,15,16,40 // movss (%rax),%xmm5 .byte 243,15,89,236 // mulss %xmm4,%xmm5 .byte 243,72,15,44,205 // cvttss2si %xmm5,%rcx @@ -59866,7 +60184,7 @@ _sk_set_rgb_sse41_8bit: .byte 9,208 // or %edx,%eax .byte 102,15,110,224 // movd %eax,%xmm4 .byte 102,15,112,228,0 // pshufd $0x0,%xmm4,%xmm4 - .byte 102,15,111,45,157,29,0,0 // movdqa 0x1d9d(%rip),%xmm5 # 1ea0 <_sk_darken_sse41_8bit+0x17c> + .byte 102,15,111,45,189,33,0,0 // movdqa 0x21bd(%rip),%xmm5 # 22c0 <_sk_exclusion_sse41_8bit+0x147> .byte 102,15,219,205 // pand %xmm5,%xmm1 .byte 102,15,219,197 // pand %xmm5,%xmm0 .byte 102,15,235,196 // por %xmm4,%xmm0 @@ -59880,12 +60198,12 @@ FUNCTION(_sk_premul_sse41_8bit) _sk_premul_sse41_8bit: .byte 102,15,111,225 // movdqa %xmm1,%xmm4 .byte 102,15,111,232 // movdqa %xmm0,%xmm5 - .byte 102,15,111,5,137,29,0,0 // movdqa 0x1d89(%rip),%xmm0 # 1eb0 <_sk_darken_sse41_8bit+0x18c> + .byte 102,15,111,5,169,33,0,0 // movdqa 0x21a9(%rip),%xmm0 # 22d0 <_sk_exclusion_sse41_8bit+0x157> .byte 102,15,111,253 // movdqa %xmm5,%xmm7 .byte 102,15,56,0,248 // pshufb %xmm0,%xmm7 .byte 102,15,111,244 // movdqa %xmm4,%xmm6 .byte 102,15,56,0,240 // pshufb %xmm0,%xmm6 - .byte 102,15,111,5,127,29,0,0 // movdqa 0x1d7f(%rip),%xmm0 # 1ec0 <_sk_darken_sse41_8bit+0x19c> + .byte 102,15,111,5,159,33,0,0 // movdqa 0x219f(%rip),%xmm0 # 22e0 <_sk_exclusion_sse41_8bit+0x167> .byte 102,15,235,240 // por %xmm0,%xmm6 .byte 102,15,235,248 // por %xmm0,%xmm7 .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 @@ -59918,7 +60236,7 @@ HIDDEN _sk_swap_rb_sse41_8bit .globl _sk_swap_rb_sse41_8bit FUNCTION(_sk_swap_rb_sse41_8bit) _sk_swap_rb_sse41_8bit: - .byte 102,15,111,37,12,29,0,0 // movdqa 0x1d0c(%rip),%xmm4 # 1ed0 <_sk_darken_sse41_8bit+0x1ac> + .byte 102,15,111,37,44,33,0,0 // movdqa 0x212c(%rip),%xmm4 # 22f0 <_sk_exclusion_sse41_8bit+0x177> .byte 102,15,56,0,196 // pshufb %xmm4,%xmm0 .byte 102,15,56,0,204 // pshufb %xmm4,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -60045,7 +60363,7 @@ _sk_load_8888_dst_sse41_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,221 // jmpq ffffffffde000354 <_sk_darken_sse41_8bit+0xffffffffddffe630> + .byte 233,255,255,255,221 // jmpq ffffffffde000354 <_sk_exclusion_sse41_8bit+0xffffffffddffe1db> .byte 255 // (bad) .byte 255 // (bad) .byte 255 // .byte 0xff @@ -60128,7 +60446,7 @@ _sk_load_bgra_sse41_8bit: .byte 117,35 // jne 448 <_sk_load_bgra_sse41_8bit+0x44> .byte 243,66,15,111,76,130,16 // movdqu 0x10(%rdx,%r8,4),%xmm1 .byte 243,66,15,111,4,130 // movdqu (%rdx,%r8,4),%xmm0 - .byte 102,15,111,37,166,26,0,0 // movdqa 0x1aa6(%rip),%xmm4 # 1ee0 <_sk_darken_sse41_8bit+0x1bc> + .byte 102,15,111,37,198,30,0,0 // movdqa 0x1ec6(%rip),%xmm4 # 2300 <_sk_exclusion_sse41_8bit+0x187> .byte 102,15,56,0,196 // pshufb %xmm4,%xmm0 .byte 102,15,56,0,204 // pshufb %xmm4,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -60191,7 +60509,7 @@ _sk_load_bgra_dst_sse41_8bit: .byte 117,35 // jne 518 <_sk_load_bgra_dst_sse41_8bit+0x44> .byte 243,66,15,111,92,130,16 // movdqu 0x10(%rdx,%r8,4),%xmm3 .byte 243,66,15,111,20,130 // movdqu (%rdx,%r8,4),%xmm2 - .byte 102,15,111,37,230,25,0,0 // movdqa 0x19e6(%rip),%xmm4 # 1ef0 <_sk_darken_sse41_8bit+0x1cc> + .byte 102,15,111,37,6,30,0,0 // movdqa 0x1e06(%rip),%xmm4 # 2310 <_sk_exclusion_sse41_8bit+0x197> .byte 102,15,56,0,212 // pshufb %xmm4,%xmm2 .byte 102,15,56,0,220 // pshufb %xmm4,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -60250,7 +60568,7 @@ _sk_store_bgra_sse41_8bit: .byte 72,15,175,209 // imul %rcx,%rdx .byte 72,193,226,2 // shl $0x2,%rdx .byte 72,3,16 // add (%rax),%rdx - .byte 102,15,111,53,56,25,0,0 // movdqa 0x1938(%rip),%xmm6 # 1f00 <_sk_darken_sse41_8bit+0x1dc> + .byte 102,15,111,53,88,29,0,0 // movdqa 0x1d58(%rip),%xmm6 # 2320 <_sk_exclusion_sse41_8bit+0x1a7> .byte 102,15,111,233 // movdqa %xmm1,%xmm5 .byte 102,15,56,0,238 // pshufb %xmm6,%xmm5 .byte 102,15,111,224 // movdqa %xmm0,%xmm4 @@ -60316,7 +60634,7 @@ _sk_load_a8_sse41_8bit: .byte 77,133,201 // test %r9,%r9 .byte 117,42 // jne 6af <_sk_load_a8_sse41_8bit+0x47> .byte 102,66,15,56,48,12,2 // pmovzxbw (%rdx,%r8,1),%xmm1 - .byte 102,15,219,13,124,24,0,0 // pand 0x187c(%rip),%xmm1 # 1f10 <_sk_darken_sse41_8bit+0x1ec> + .byte 102,15,219,13,156,28,0,0 // pand 0x1c9c(%rip),%xmm1 # 2330 <_sk_exclusion_sse41_8bit+0x1b7> .byte 102,15,239,228 // pxor %xmm4,%xmm4 .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0 .byte 102,15,105,204 // punpckhwd %xmm4,%xmm1 @@ -60364,7 +60682,7 @@ _sk_load_a8_sse41_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,222 // jmpq ffffffffdf000750 <_sk_darken_sse41_8bit+0xffffffffdeffea2c> + .byte 233,255,255,255,222 // jmpq ffffffffdf000750 <_sk_exclusion_sse41_8bit+0xffffffffdeffe5d7> .byte 255 // (bad) .byte 255 // (bad) .byte 255,211 // callq *%rbx @@ -60389,7 +60707,7 @@ _sk_load_a8_dst_sse41_8bit: .byte 77,133,201 // test %r9,%r9 .byte 117,42 // jne 7a3 <_sk_load_a8_dst_sse41_8bit+0x47> .byte 102,66,15,56,48,28,2 // pmovzxbw (%rdx,%r8,1),%xmm3 - .byte 102,15,219,29,152,23,0,0 // pand 0x1798(%rip),%xmm3 # 1f20 <_sk_darken_sse41_8bit+0x1fc> + .byte 102,15,219,29,184,27,0,0 // pand 0x1bb8(%rip),%xmm3 # 2340 <_sk_exclusion_sse41_8bit+0x1c7> .byte 102,15,239,228 // pxor %xmm4,%xmm4 .byte 102,15,56,51,211 // pmovzxwd %xmm3,%xmm2 .byte 102,15,105,220 // punpckhwd %xmm4,%xmm3 @@ -60437,7 +60755,7 @@ _sk_load_a8_dst_sse41_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,222 // jmpq ffffffffdf000844 <_sk_darken_sse41_8bit+0xffffffffdeffeb20> + .byte 233,255,255,255,222 // jmpq ffffffffdf000844 <_sk_exclusion_sse41_8bit+0xffffffffdeffe6cb> .byte 255 // (bad) .byte 255 // (bad) .byte 255,211 // callq *%rbx @@ -60459,7 +60777,7 @@ _sk_store_a8_sse41_8bit: .byte 72,99,87,8 // movslq 0x8(%rdi),%rdx .byte 72,15,175,209 // imul %rcx,%rdx .byte 72,3,16 // add (%rax),%rdx - .byte 102,15,111,45,192,22,0,0 // movdqa 0x16c0(%rip),%xmm5 # 1f30 <_sk_darken_sse41_8bit+0x20c> + .byte 102,15,111,45,224,26,0,0 // movdqa 0x1ae0(%rip),%xmm5 # 2350 <_sk_exclusion_sse41_8bit+0x1d7> .byte 102,15,111,241 // movdqa %xmm1,%xmm6 .byte 102,15,56,0,245 // pshufb %xmm5,%xmm6 .byte 102,15,111,224 // movdqa %xmm0,%xmm4 @@ -60467,7 +60785,7 @@ _sk_store_a8_sse41_8bit: .byte 102,15,108,230 // punpcklqdq %xmm6,%xmm4 .byte 77,133,201 // test %r9,%r9 .byte 117,19 // jne 89e <_sk_store_a8_sse41_8bit+0x4e> - .byte 102,15,56,0,37,204,22,0,0 // pshufb 0x16cc(%rip),%xmm4 # 1f60 <_sk_darken_sse41_8bit+0x23c> + .byte 102,15,56,0,37,236,26,0,0 // pshufb 0x1aec(%rip),%xmm4 # 2380 <_sk_exclusion_sse41_8bit+0x207> .byte 102,66,15,214,36,2 // movq %xmm4,(%rdx,%r8,1) .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax @@ -60483,13 +60801,13 @@ _sk_store_a8_sse41_8bit: .byte 102,66,15,58,20,36,2,0 // pextrb $0x0,%xmm4,(%rdx,%r8,1) .byte 235,209 // jmp 89a <_sk_store_a8_sse41_8bit+0x4a> .byte 102,66,15,58,20,100,2,2,4 // pextrb $0x4,%xmm4,0x2(%rdx,%r8,1) - .byte 102,15,56,0,37,117,22,0,0 // pshufb 0x1675(%rip),%xmm4 # 1f50 <_sk_darken_sse41_8bit+0x22c> + .byte 102,15,56,0,37,149,26,0,0 // pshufb 0x1a95(%rip),%xmm4 # 2370 <_sk_exclusion_sse41_8bit+0x1f7> .byte 102,66,15,58,21,36,2,0 // pextrw $0x0,%xmm4,(%rdx,%r8,1) .byte 235,181 // jmp 89a <_sk_store_a8_sse41_8bit+0x4a> .byte 102,66,15,58,20,100,2,6,12 // pextrb $0xc,%xmm4,0x6(%rdx,%r8,1) .byte 102,66,15,58,20,100,2,5,10 // pextrb $0xa,%xmm4,0x5(%rdx,%r8,1) .byte 102,66,15,58,20,100,2,4,8 // pextrb $0x8,%xmm4,0x4(%rdx,%r8,1) - .byte 102,15,56,0,37,55,22,0,0 // pshufb 0x1637(%rip),%xmm4 # 1f40 <_sk_darken_sse41_8bit+0x21c> + .byte 102,15,56,0,37,87,26,0,0 // pshufb 0x1a57(%rip),%xmm4 # 2360 <_sk_exclusion_sse41_8bit+0x1e7> .byte 102,66,15,126,36,2 // movd %xmm4,(%rdx,%r8,1) .byte 235,137 // jmp 89a <_sk_store_a8_sse41_8bit+0x4a> .byte 15,31,0 // nopl (%rax) @@ -60529,14 +60847,14 @@ _sk_load_g8_sse41_8bit: .byte 77,133,201 // test %r9,%r9 .byte 117,66 // jne 98f <_sk_load_g8_sse41_8bit+0x5f> .byte 102,66,15,56,48,12,2 // pmovzxbw (%rdx,%r8,1),%xmm1 - .byte 102,15,219,13,20,22,0,0 // pand 0x1614(%rip),%xmm1 # 1f70 <_sk_darken_sse41_8bit+0x24c> + .byte 102,15,219,13,52,26,0,0 // pand 0x1a34(%rip),%xmm1 # 2390 <_sk_exclusion_sse41_8bit+0x217> .byte 102,15,239,228 // pxor %xmm4,%xmm4 .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0 .byte 102,15,105,204 // punpckhwd %xmm4,%xmm1 - .byte 102,15,111,37,15,22,0,0 // movdqa 0x160f(%rip),%xmm4 # 1f80 <_sk_darken_sse41_8bit+0x25c> + .byte 102,15,111,37,47,26,0,0 // movdqa 0x1a2f(%rip),%xmm4 # 23a0 <_sk_exclusion_sse41_8bit+0x227> .byte 102,15,56,64,204 // pmulld %xmm4,%xmm1 .byte 102,15,56,64,196 // pmulld %xmm4,%xmm0 - .byte 102,15,111,37,13,22,0,0 // movdqa 0x160d(%rip),%xmm4 # 1f90 <_sk_darken_sse41_8bit+0x26c> + .byte 102,15,111,37,45,26,0,0 // movdqa 0x1a2d(%rip),%xmm4 # 23b0 <_sk_exclusion_sse41_8bit+0x237> .byte 102,15,235,196 // por %xmm4,%xmm0 .byte 102,15,235,204 // por %xmm4,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -60581,7 +60899,7 @@ _sk_load_g8_sse41_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 232,255,255,255,221 // callq ffffffffde000a34 <_sk_darken_sse41_8bit+0xffffffffddffed10> + .byte 232,255,255,255,221 // callq ffffffffde000a34 <_sk_exclusion_sse41_8bit+0xffffffffddffe8bb> .byte 255 // (bad) .byte 255 // (bad) .byte 255,210 // callq *%rdx @@ -60606,14 +60924,14 @@ _sk_load_g8_dst_sse41_8bit: .byte 77,133,201 // test %r9,%r9 .byte 117,66 // jne a9f <_sk_load_g8_dst_sse41_8bit+0x5f> .byte 102,66,15,56,48,28,2 // pmovzxbw (%rdx,%r8,1),%xmm3 - .byte 102,15,219,29,52,21,0,0 // pand 0x1534(%rip),%xmm3 # 1fa0 <_sk_darken_sse41_8bit+0x27c> + .byte 102,15,219,29,84,25,0,0 // pand 0x1954(%rip),%xmm3 # 23c0 <_sk_exclusion_sse41_8bit+0x247> .byte 102,15,239,228 // pxor %xmm4,%xmm4 .byte 102,15,56,51,211 // pmovzxwd %xmm3,%xmm2 .byte 102,15,105,220 // punpckhwd %xmm4,%xmm3 - .byte 102,15,111,37,47,21,0,0 // movdqa 0x152f(%rip),%xmm4 # 1fb0 <_sk_darken_sse41_8bit+0x28c> + .byte 102,15,111,37,79,25,0,0 // movdqa 0x194f(%rip),%xmm4 # 23d0 <_sk_exclusion_sse41_8bit+0x257> .byte 102,15,56,64,220 // pmulld %xmm4,%xmm3 .byte 102,15,56,64,212 // pmulld %xmm4,%xmm2 - .byte 102,15,111,37,45,21,0,0 // movdqa 0x152d(%rip),%xmm4 # 1fc0 <_sk_darken_sse41_8bit+0x29c> + .byte 102,15,111,37,77,25,0,0 // movdqa 0x194d(%rip),%xmm4 # 23e0 <_sk_exclusion_sse41_8bit+0x267> .byte 102,15,235,212 // por %xmm4,%xmm2 .byte 102,15,235,220 // por %xmm4,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -60658,7 +60976,7 @@ _sk_load_g8_dst_sse41_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 232,255,255,255,221 // callq ffffffffde000b44 <_sk_darken_sse41_8bit+0xffffffffddffee20> + .byte 232,255,255,255,221 // callq ffffffffde000b44 <_sk_exclusion_sse41_8bit+0xffffffffddffe9cb> .byte 255 // (bad) .byte 255 // (bad) .byte 255,210 // callq *%rdx @@ -60686,7 +61004,7 @@ _sk_srcover_rgba_8888_sse41_8bit: .byte 243,70,15,111,68,138,16 // movdqu 0x10(%rdx,%r9,4),%xmm8 .byte 243,70,15,111,12,138 // movdqu (%rdx,%r9,4),%xmm9 .byte 77,133,192 // test %r8,%r8 - .byte 102,15,111,37,67,20,0,0 // movdqa 0x1443(%rip),%xmm4 # 1fd0 <_sk_darken_sse41_8bit+0x2ac> + .byte 102,15,111,37,99,24,0,0 // movdqa 0x1863(%rip),%xmm4 # 23f0 <_sk_exclusion_sse41_8bit+0x277> .byte 102,15,111,241 // movdqa %xmm1,%xmm6 .byte 102,15,56,0,244 // pshufb %xmm4,%xmm6 .byte 102,15,111,248 // movdqa %xmm0,%xmm7 @@ -60808,7 +61126,7 @@ _sk_scale_1_float_sse41_8bit: .byte 102,15,111,232 // movdqa %xmm0,%xmm5 .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,15,16,0 // movss (%rax),%xmm0 - .byte 243,15,89,5,42,17,0,0 // mulss 0x112a(%rip),%xmm0 # 1e98 <_sk_darken_sse41_8bit+0x174> + .byte 243,15,89,5,70,21,0,0 // mulss 0x1546(%rip),%xmm0 # 22b4 <_sk_exclusion_sse41_8bit+0x13b> .byte 243,15,44,192 // cvttss2si %xmm0,%eax .byte 15,87,192 // xorps %xmm0,%xmm0 .byte 102,68,15,56,48,197 // pmovzxbw %xmm5,%xmm8 @@ -60816,7 +61134,7 @@ _sk_scale_1_float_sse41_8bit: .byte 102,68,15,56,48,204 // pmovzxbw %xmm4,%xmm9 .byte 102,15,104,224 // punpckhbw %xmm0,%xmm4 .byte 102,15,110,240 // movd %eax,%xmm6 - .byte 102,15,56,0,53,74,18,0,0 // pshufb 0x124a(%rip),%xmm6 # 1fe0 <_sk_darken_sse41_8bit+0x2bc> + .byte 102,15,56,0,53,106,22,0,0 // pshufb 0x166a(%rip),%xmm6 # 2400 <_sk_exclusion_sse41_8bit+0x287> .byte 102,15,111,206 // movdqa %xmm6,%xmm1 .byte 102,65,15,213,201 // pmullw %xmm9,%xmm1 .byte 102,15,111,198 // movdqa %xmm6,%xmm0 @@ -60851,11 +61169,11 @@ _sk_scale_u8_sse41_8bit: .byte 77,133,201 // test %r9,%r9 .byte 15,133,160,0,0,0 // jne ea7 <_sk_scale_u8_sse41_8bit+0xc1> .byte 102,66,15,56,48,52,2 // pmovzxbw (%rdx,%r8,1),%xmm6 - .byte 102,15,219,53,218,17,0,0 // pand 0x11da(%rip),%xmm6 # 1ff0 <_sk_darken_sse41_8bit+0x2cc> + .byte 102,15,219,53,250,21,0,0 // pand 0x15fa(%rip),%xmm6 # 2410 <_sk_exclusion_sse41_8bit+0x297> .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 102,15,111,254 // movdqa %xmm6,%xmm7 - .byte 102,15,56,0,61,216,17,0,0 // pshufb 0x11d8(%rip),%xmm7 # 2000 <_sk_darken_sse41_8bit+0x2dc> - .byte 102,15,56,0,53,223,17,0,0 // pshufb 0x11df(%rip),%xmm6 # 2010 <_sk_darken_sse41_8bit+0x2ec> + .byte 102,15,56,0,61,248,21,0,0 // pshufb 0x15f8(%rip),%xmm7 # 2420 <_sk_exclusion_sse41_8bit+0x2a7> + .byte 102,15,56,0,53,255,21,0,0 // pshufb 0x15ff(%rip),%xmm6 # 2430 <_sk_exclusion_sse41_8bit+0x2b7> .byte 102,68,15,56,48,200 // pmovzxbw %xmm0,%xmm9 .byte 102,65,15,104,192 // punpckhbw %xmm8,%xmm0 .byte 102,68,15,56,48,209 // pmovzxbw %xmm1,%xmm10 @@ -60940,7 +61258,7 @@ FUNCTION(_sk_lerp_1_float_sse41_8bit) _sk_lerp_1_float_sse41_8bit: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,15,16,32 // movss (%rax),%xmm4 - .byte 243,15,89,37,46,15,0,0 // mulss 0xf2e(%rip),%xmm4 # 1e9c <_sk_darken_sse41_8bit+0x178> + .byte 243,15,89,37,74,19,0,0 // mulss 0x134a(%rip),%xmm4 # 22b8 <_sk_exclusion_sse41_8bit+0x13f> .byte 243,15,44,196 // cvttss2si %xmm4,%eax .byte 102,15,110,224 // movd %eax,%xmm4 .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 @@ -60951,7 +61269,7 @@ _sk_lerp_1_float_sse41_8bit: .byte 102,65,15,104,193 // punpckhbw %xmm9,%xmm0 .byte 102,68,15,56,48,217 // pmovzxbw %xmm1,%xmm11 .byte 102,65,15,104,201 // punpckhbw %xmm9,%xmm1 - .byte 102,15,56,0,61,119,16,0,0 // pshufb 0x1077(%rip),%xmm7 # 2020 <_sk_darken_sse41_8bit+0x2fc> + .byte 102,15,56,0,61,151,20,0,0 // pshufb 0x1497(%rip),%xmm7 # 2440 <_sk_exclusion_sse41_8bit+0x2c7> .byte 102,68,15,111,231 // movdqa %xmm7,%xmm12 .byte 102,69,15,213,227 // pmullw %xmm11,%xmm12 .byte 102,68,15,111,239 // movdqa %xmm7,%xmm13 @@ -61014,11 +61332,11 @@ _sk_lerp_u8_sse41_8bit: .byte 77,133,201 // test %r9,%r9 .byte 15,133,46,1,0,0 // jne 11d1 <_sk_lerp_u8_sse41_8bit+0x14f> .byte 102,66,15,56,48,60,2 // pmovzxbw (%rdx,%r8,1),%xmm7 - .byte 102,15,219,61,126,15,0,0 // pand 0xf7e(%rip),%xmm7 # 2030 <_sk_darken_sse41_8bit+0x30c> + .byte 102,15,219,61,158,19,0,0 // pand 0x139e(%rip),%xmm7 # 2450 <_sk_exclusion_sse41_8bit+0x2d7> .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 102,15,111,247 // movdqa %xmm7,%xmm6 - .byte 102,15,56,0,53,124,15,0,0 // pshufb 0xf7c(%rip),%xmm6 # 2040 <_sk_darken_sse41_8bit+0x31c> - .byte 102,15,56,0,61,131,15,0,0 // pshufb 0xf83(%rip),%xmm7 # 2050 <_sk_darken_sse41_8bit+0x32c> + .byte 102,15,56,0,53,156,19,0,0 // pshufb 0x139c(%rip),%xmm6 # 2460 <_sk_exclusion_sse41_8bit+0x2e7> + .byte 102,15,56,0,61,163,19,0,0 // pshufb 0x13a3(%rip),%xmm7 # 2470 <_sk_exclusion_sse41_8bit+0x2f7> .byte 102,68,15,56,48,200 // pmovzxbw %xmm0,%xmm9 .byte 102,65,15,104,192 // punpckhbw %xmm8,%xmm0 .byte 102,68,15,56,48,209 // pmovzxbw %xmm1,%xmm10 @@ -61113,7 +61431,7 @@ _sk_lerp_u8_sse41_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,222 // jmpq ffffffffdf00127c <_sk_darken_sse41_8bit+0xffffffffdefff558> + .byte 233,255,255,255,222 // jmpq ffffffffdf00127c <_sk_exclusion_sse41_8bit+0xffffffffdefff103> .byte 255 // (bad) .byte 255 // (bad) .byte 255,211 // callq *%rbx @@ -61147,7 +61465,7 @@ HIDDEN _sk_black_color_sse41_8bit FUNCTION(_sk_black_color_sse41_8bit) _sk_black_color_sse41_8bit: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,5,187,13,0,0 // movaps 0xdbb(%rip),%xmm0 # 2060 <_sk_darken_sse41_8bit+0x33c> + .byte 15,40,5,219,17,0,0 // movaps 0x11db(%rip),%xmm0 # 2480 <_sk_exclusion_sse41_8bit+0x307> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 255,224 // jmpq *%rax @@ -61173,7 +61491,7 @@ HIDDEN _sk_srcatop_sse41_8bit .globl _sk_srcatop_sse41_8bit FUNCTION(_sk_srcatop_sse41_8bit) _sk_srcatop_sse41_8bit: - .byte 102,68,15,111,21,167,13,0,0 // movdqa 0xda7(%rip),%xmm10 # 2070 <_sk_darken_sse41_8bit+0x34c> + .byte 102,68,15,111,21,199,17,0,0 // movdqa 0x11c7(%rip),%xmm10 # 2490 <_sk_exclusion_sse41_8bit+0x317> .byte 102,68,15,111,219 // movdqa %xmm3,%xmm11 .byte 102,68,15,56,48,195 // pmovzxbw %xmm3,%xmm8 .byte 102,15,111,235 // movdqa %xmm3,%xmm5 @@ -61243,7 +61561,7 @@ HIDDEN _sk_dstatop_sse41_8bit .globl _sk_dstatop_sse41_8bit FUNCTION(_sk_dstatop_sse41_8bit) _sk_dstatop_sse41_8bit: - .byte 102,68,15,111,29,124,12,0,0 // movdqa 0xc7c(%rip),%xmm11 # 2080 <_sk_darken_sse41_8bit+0x35c> + .byte 102,68,15,111,29,156,16,0,0 // movdqa 0x109c(%rip),%xmm11 # 24a0 <_sk_exclusion_sse41_8bit+0x327> .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 .byte 102,69,15,56,0,235 // pshufb %xmm11,%xmm13 .byte 102,68,15,111,248 // movdqa %xmm0,%xmm15 @@ -61315,7 +61633,7 @@ FUNCTION(_sk_srcin_sse41_8bit) _sk_srcin_sse41_8bit: .byte 102,15,111,225 // movdqa %xmm1,%xmm4 .byte 102,15,111,232 // movdqa %xmm0,%xmm5 - .byte 102,15,111,5,67,11,0,0 // movdqa 0xb43(%rip),%xmm0 # 2090 <_sk_darken_sse41_8bit+0x36c> + .byte 102,15,111,5,99,15,0,0 // movdqa 0xf63(%rip),%xmm0 # 24b0 <_sk_exclusion_sse41_8bit+0x337> .byte 102,15,111,243 // movdqa %xmm3,%xmm6 .byte 102,15,56,0,240 // pshufb %xmm0,%xmm6 .byte 102,15,111,250 // movdqa %xmm2,%xmm7 @@ -61350,7 +61668,7 @@ HIDDEN _sk_dstin_sse41_8bit .globl _sk_dstin_sse41_8bit FUNCTION(_sk_dstin_sse41_8bit) _sk_dstin_sse41_8bit: - .byte 102,15,111,37,198,10,0,0 // movdqa 0xac6(%rip),%xmm4 # 20a0 <_sk_darken_sse41_8bit+0x37c> + .byte 102,15,111,37,230,14,0,0 // movdqa 0xee6(%rip),%xmm4 # 24c0 <_sk_exclusion_sse41_8bit+0x347> .byte 102,15,56,0,204 // pshufb %xmm4,%xmm1 .byte 102,15,56,0,196 // pshufb %xmm4,%xmm0 .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 @@ -61389,7 +61707,7 @@ FUNCTION(_sk_srcout_sse41_8bit) _sk_srcout_sse41_8bit: .byte 102,15,111,225 // movdqa %xmm1,%xmm4 .byte 102,15,111,232 // movdqa %xmm0,%xmm5 - .byte 102,15,111,5,57,10,0,0 // movdqa 0xa39(%rip),%xmm0 # 20b0 <_sk_darken_sse41_8bit+0x38c> + .byte 102,15,111,5,89,14,0,0 // movdqa 0xe59(%rip),%xmm0 # 24d0 <_sk_exclusion_sse41_8bit+0x357> .byte 102,15,111,250 // movdqa %xmm2,%xmm7 .byte 102,15,56,0,248 // pshufb %xmm0,%xmm7 .byte 102,15,111,243 // movdqa %xmm3,%xmm6 @@ -61427,7 +61745,7 @@ HIDDEN _sk_dstout_sse41_8bit .globl _sk_dstout_sse41_8bit FUNCTION(_sk_dstout_sse41_8bit) _sk_dstout_sse41_8bit: - .byte 102,15,111,37,176,9,0,0 // movdqa 0x9b0(%rip),%xmm4 # 20c0 <_sk_darken_sse41_8bit+0x39c> + .byte 102,15,111,37,208,13,0,0 // movdqa 0xdd0(%rip),%xmm4 # 24e0 <_sk_exclusion_sse41_8bit+0x367> .byte 102,15,56,0,196 // pshufb %xmm4,%xmm0 .byte 102,15,56,0,204 // pshufb %xmm4,%xmm1 .byte 102,15,118,228 // pcmpeqd %xmm4,%xmm4 @@ -61467,7 +61785,7 @@ HIDDEN _sk_srcover_sse41_8bit .globl _sk_srcover_sse41_8bit FUNCTION(_sk_srcover_sse41_8bit) _sk_srcover_sse41_8bit: - .byte 102,15,111,53,31,9,0,0 // movdqa 0x91f(%rip),%xmm6 # 20d0 <_sk_darken_sse41_8bit+0x3ac> + .byte 102,15,111,53,63,13,0,0 // movdqa 0xd3f(%rip),%xmm6 # 24f0 <_sk_exclusion_sse41_8bit+0x377> .byte 102,68,15,111,217 // movdqa %xmm1,%xmm11 .byte 102,68,15,56,0,222 // pshufb %xmm6,%xmm11 .byte 102,15,111,232 // movdqa %xmm0,%xmm5 @@ -61508,7 +61826,7 @@ HIDDEN _sk_dstover_sse41_8bit .globl _sk_dstover_sse41_8bit FUNCTION(_sk_dstover_sse41_8bit) _sk_dstover_sse41_8bit: - .byte 102,68,15,111,5,127,8,0,0 // movdqa 0x87f(%rip),%xmm8 # 20e0 <_sk_darken_sse41_8bit+0x3bc> + .byte 102,68,15,111,5,159,12,0,0 // movdqa 0xc9f(%rip),%xmm8 # 2500 <_sk_exclusion_sse41_8bit+0x387> .byte 102,68,15,111,209 // movdqa %xmm1,%xmm10 .byte 102,68,15,56,48,201 // pmovzxbw %xmm1,%xmm9 .byte 102,15,252,203 // paddb %xmm3,%xmm1 @@ -61587,7 +61905,7 @@ _sk_multiply_sse41_8bit: .byte 102,15,111,218 // movdqa %xmm2,%xmm3 .byte 102,15,111,209 // movdqa %xmm1,%xmm2 .byte 102,15,111,200 // movdqa %xmm0,%xmm1 - .byte 102,68,15,111,53,77,7,0,0 // movdqa 0x74d(%rip),%xmm14 # 20f0 <_sk_darken_sse41_8bit+0x3cc> + .byte 102,68,15,111,53,109,11,0,0 // movdqa 0xb6d(%rip),%xmm14 # 2510 <_sk_exclusion_sse41_8bit+0x397> .byte 102,68,15,111,195 // movdqa %xmm3,%xmm8 .byte 102,15,111,235 // movdqa %xmm3,%xmm5 .byte 102,65,15,56,0,238 // pshufb %xmm14,%xmm5 @@ -61716,7 +62034,7 @@ HIDDEN _sk_xor__sse41_8bit .globl _sk_xor__sse41_8bit FUNCTION(_sk_xor__sse41_8bit) _sk_xor__sse41_8bit: - .byte 102,68,15,111,21,27,5,0,0 // movdqa 0x51b(%rip),%xmm10 # 2100 <_sk_darken_sse41_8bit+0x3dc> + .byte 102,68,15,111,21,59,9,0,0 // movdqa 0x93b(%rip),%xmm10 # 2520 <_sk_exclusion_sse41_8bit+0x3a7> .byte 102,68,15,111,226 // movdqa %xmm2,%xmm12 .byte 102,68,15,56,48,194 // pmovzxbw %xmm2,%xmm8 .byte 102,15,111,234 // movdqa %xmm2,%xmm5 @@ -61789,7 +62107,7 @@ HIDDEN _sk_darken_sse41_8bit FUNCTION(_sk_darken_sse41_8bit) _sk_darken_sse41_8bit: .byte 102,68,15,111,240 // movdqa %xmm0,%xmm14 - .byte 102,68,15,111,5,222,3,0,0 // movdqa 0x3de(%rip),%xmm8 # 2110 <_sk_darken_sse41_8bit+0x3ec> + .byte 102,68,15,111,5,254,7,0,0 // movdqa 0x7fe(%rip),%xmm8 # 2530 <_sk_exclusion_sse41_8bit+0x3b7> .byte 102,68,15,111,219 // movdqa %xmm3,%xmm11 .byte 102,68,15,56,48,203 // pmovzxbw %xmm3,%xmm9 .byte 102,15,111,243 // movdqa %xmm3,%xmm6 @@ -61856,20 +62174,252 @@ _sk_darken_sse41_8bit: .byte 102,15,248,231 // psubb %xmm7,%xmm4 .byte 102,15,248,206 // psubb %xmm6,%xmm1 .byte 102,68,15,248,245 // psubb %xmm5,%xmm14 - .byte 15,40,5,162,2,0,0 // movaps 0x2a2(%rip),%xmm0 # 2120 <_sk_darken_sse41_8bit+0x3fc> + .byte 15,40,5,194,6,0,0 // movaps 0x6c2(%rip),%xmm0 # 2540 <_sk_exclusion_sse41_8bit+0x3c7> .byte 102,68,15,56,16,244 // pblendvb %xmm0,%xmm4,%xmm14 .byte 102,65,15,56,16,200 // pblendvb %xmm0,%xmm8,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax .byte 102,65,15,111,198 // movdqa %xmm14,%xmm0 .byte 255,224 // jmpq *%rax +HIDDEN _sk_lighten_sse41_8bit +.globl _sk_lighten_sse41_8bit +FUNCTION(_sk_lighten_sse41_8bit) +_sk_lighten_sse41_8bit: + .byte 102,68,15,111,240 // movdqa %xmm0,%xmm14 + .byte 102,68,15,111,5,175,6,0,0 // movdqa 0x6af(%rip),%xmm8 # 2550 <_sk_exclusion_sse41_8bit+0x3d7> + .byte 102,68,15,111,219 // movdqa %xmm3,%xmm11 + .byte 102,68,15,56,48,203 // pmovzxbw %xmm3,%xmm9 + .byte 102,15,111,243 // movdqa %xmm3,%xmm6 + .byte 102,65,15,56,0,240 // pshufb %xmm8,%xmm6 + .byte 102,68,15,111,234 // movdqa %xmm2,%xmm13 + .byte 102,68,15,56,48,210 // pmovzxbw %xmm2,%xmm10 + .byte 102,15,111,234 // movdqa %xmm2,%xmm5 + .byte 102,65,15,56,0,232 // pshufb %xmm8,%xmm5 + .byte 102,69,15,239,228 // pxor %xmm12,%xmm12 + .byte 102,65,15,104,196 // punpckhbw %xmm12,%xmm0 + .byte 102,15,111,225 // movdqa %xmm1,%xmm4 + .byte 102,65,15,104,228 // punpckhbw %xmm12,%xmm4 + .byte 102,69,15,56,48,254 // pmovzxbw %xmm14,%xmm15 + .byte 102,15,56,48,253 // pmovzxbw %xmm5,%xmm7 + .byte 102,65,15,104,236 // punpckhbw %xmm12,%xmm5 + .byte 102,15,213,232 // pmullw %xmm0,%xmm5 + .byte 102,15,253,232 // paddw %xmm0,%xmm5 + .byte 102,15,56,48,198 // pmovzxbw %xmm6,%xmm0 + .byte 102,65,15,104,244 // punpckhbw %xmm12,%xmm6 + .byte 102,15,213,244 // pmullw %xmm4,%xmm6 + .byte 102,15,253,244 // paddw %xmm4,%xmm6 + .byte 102,65,15,111,230 // movdqa %xmm14,%xmm4 + .byte 102,68,15,252,242 // paddb %xmm2,%xmm14 + .byte 102,65,15,213,255 // pmullw %xmm15,%xmm7 + .byte 102,65,15,253,255 // paddw %xmm15,%xmm7 + .byte 102,68,15,56,48,249 // pmovzxbw %xmm1,%xmm15 + .byte 102,65,15,213,199 // pmullw %xmm15,%xmm0 + .byte 102,65,15,253,199 // paddw %xmm15,%xmm0 + .byte 102,68,15,111,249 // movdqa %xmm1,%xmm15 + .byte 102,15,252,203 // paddb %xmm3,%xmm1 + .byte 102,15,113,213,8 // psrlw $0x8,%xmm5 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,15,103,253 // packuswb %xmm5,%xmm7 + .byte 102,15,113,214,8 // psrlw $0x8,%xmm6 + .byte 102,15,113,208,8 // psrlw $0x8,%xmm0 + .byte 102,15,103,198 // packuswb %xmm6,%xmm0 + .byte 102,69,15,56,0,248 // pshufb %xmm8,%xmm15 + .byte 102,65,15,56,0,224 // pshufb %xmm8,%xmm4 + .byte 102,69,15,104,236 // punpckhbw %xmm12,%xmm13 + .byte 102,69,15,104,220 // punpckhbw %xmm12,%xmm11 + .byte 102,15,56,48,236 // pmovzxbw %xmm4,%xmm5 + .byte 102,65,15,56,48,247 // pmovzxbw %xmm15,%xmm6 + .byte 102,65,15,104,228 // punpckhbw %xmm12,%xmm4 + .byte 102,69,15,104,252 // punpckhbw %xmm12,%xmm15 + .byte 102,69,15,213,251 // pmullw %xmm11,%xmm15 + .byte 102,65,15,213,229 // pmullw %xmm13,%xmm4 + .byte 102,65,15,213,241 // pmullw %xmm9,%xmm6 + .byte 102,65,15,213,234 // pmullw %xmm10,%xmm5 + .byte 102,65,15,253,229 // paddw %xmm13,%xmm4 + .byte 102,69,15,253,251 // paddw %xmm11,%xmm15 + .byte 102,65,15,253,234 // paddw %xmm10,%xmm5 + .byte 102,65,15,253,241 // paddw %xmm9,%xmm6 + .byte 102,65,15,113,215,8 // psrlw $0x8,%xmm15 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,15,113,214,8 // psrlw $0x8,%xmm6 + .byte 102,15,113,213,8 // psrlw $0x8,%xmm5 + .byte 102,15,103,236 // packuswb %xmm4,%xmm5 + .byte 102,65,15,103,247 // packuswb %xmm15,%xmm6 + .byte 102,15,218,253 // pminub %xmm5,%xmm7 + .byte 102,15,218,198 // pminub %xmm6,%xmm0 + .byte 102,68,15,111,193 // movdqa %xmm1,%xmm8 + .byte 102,68,15,248,192 // psubb %xmm0,%xmm8 + .byte 102,65,15,111,230 // movdqa %xmm14,%xmm4 + .byte 102,15,248,231 // psubb %xmm7,%xmm4 + .byte 102,15,248,206 // psubb %xmm6,%xmm1 + .byte 102,68,15,248,245 // psubb %xmm5,%xmm14 + .byte 15,40,5,115,5,0,0 // movaps 0x573(%rip),%xmm0 # 2560 <_sk_exclusion_sse41_8bit+0x3e7> + .byte 102,68,15,56,16,244 // pblendvb %xmm0,%xmm4,%xmm14 + .byte 102,65,15,56,16,200 // pblendvb %xmm0,%xmm8,%xmm1 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 102,65,15,111,198 // movdqa %xmm14,%xmm0 + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_difference_sse41_8bit +.globl _sk_difference_sse41_8bit +FUNCTION(_sk_difference_sse41_8bit) +_sk_difference_sse41_8bit: + .byte 102,68,15,111,240 // movdqa %xmm0,%xmm14 + .byte 102,68,15,111,5,96,5,0,0 // movdqa 0x560(%rip),%xmm8 # 2570 <_sk_exclusion_sse41_8bit+0x3f7> + .byte 102,68,15,111,219 // movdqa %xmm3,%xmm11 + .byte 102,68,15,56,48,203 // pmovzxbw %xmm3,%xmm9 + .byte 102,15,111,251 // movdqa %xmm3,%xmm7 + .byte 102,65,15,56,0,248 // pshufb %xmm8,%xmm7 + .byte 102,68,15,111,234 // movdqa %xmm2,%xmm13 + .byte 102,68,15,56,48,210 // pmovzxbw %xmm2,%xmm10 + .byte 102,15,111,234 // movdqa %xmm2,%xmm5 + .byte 102,65,15,56,0,232 // pshufb %xmm8,%xmm5 + .byte 102,69,15,239,228 // pxor %xmm12,%xmm12 + .byte 102,65,15,104,196 // punpckhbw %xmm12,%xmm0 + .byte 102,15,111,225 // movdqa %xmm1,%xmm4 + .byte 102,65,15,104,228 // punpckhbw %xmm12,%xmm4 + .byte 102,69,15,56,48,254 // pmovzxbw %xmm14,%xmm15 + .byte 102,15,56,48,245 // pmovzxbw %xmm5,%xmm6 + .byte 102,65,15,104,236 // punpckhbw %xmm12,%xmm5 + .byte 102,15,213,232 // pmullw %xmm0,%xmm5 + .byte 102,15,253,232 // paddw %xmm0,%xmm5 + .byte 102,15,56,48,199 // pmovzxbw %xmm7,%xmm0 + .byte 102,65,15,104,252 // punpckhbw %xmm12,%xmm7 + .byte 102,15,213,252 // pmullw %xmm4,%xmm7 + .byte 102,15,253,252 // paddw %xmm4,%xmm7 + .byte 102,65,15,111,230 // movdqa %xmm14,%xmm4 + .byte 102,68,15,252,242 // paddb %xmm2,%xmm14 + .byte 102,65,15,213,247 // pmullw %xmm15,%xmm6 + .byte 102,65,15,253,247 // paddw %xmm15,%xmm6 + .byte 102,68,15,56,48,249 // pmovzxbw %xmm1,%xmm15 + .byte 102,65,15,213,199 // pmullw %xmm15,%xmm0 + .byte 102,65,15,253,199 // paddw %xmm15,%xmm0 + .byte 102,68,15,111,249 // movdqa %xmm1,%xmm15 + .byte 102,15,252,203 // paddb %xmm3,%xmm1 + .byte 102,15,113,213,8 // psrlw $0x8,%xmm5 + .byte 102,15,113,214,8 // psrlw $0x8,%xmm6 + .byte 102,15,103,245 // packuswb %xmm5,%xmm6 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,15,113,208,8 // psrlw $0x8,%xmm0 + .byte 102,15,103,199 // packuswb %xmm7,%xmm0 + .byte 102,69,15,56,0,248 // pshufb %xmm8,%xmm15 + .byte 102,65,15,56,0,224 // pshufb %xmm8,%xmm4 + .byte 102,69,15,104,236 // punpckhbw %xmm12,%xmm13 + .byte 102,69,15,104,220 // punpckhbw %xmm12,%xmm11 + .byte 102,15,56,48,236 // pmovzxbw %xmm4,%xmm5 + .byte 102,65,15,56,48,255 // pmovzxbw %xmm15,%xmm7 + .byte 102,65,15,104,228 // punpckhbw %xmm12,%xmm4 + .byte 102,69,15,104,252 // punpckhbw %xmm12,%xmm15 + .byte 102,69,15,213,251 // pmullw %xmm11,%xmm15 + .byte 102,65,15,213,229 // pmullw %xmm13,%xmm4 + .byte 102,65,15,213,249 // pmullw %xmm9,%xmm7 + .byte 102,65,15,213,234 // pmullw %xmm10,%xmm5 + .byte 102,65,15,253,229 // paddw %xmm13,%xmm4 + .byte 102,69,15,253,251 // paddw %xmm11,%xmm15 + .byte 102,65,15,253,234 // paddw %xmm10,%xmm5 + .byte 102,65,15,253,249 // paddw %xmm9,%xmm7 + .byte 102,65,15,113,215,8 // psrlw $0x8,%xmm15 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,15,113,213,8 // psrlw $0x8,%xmm5 + .byte 102,15,103,236 // packuswb %xmm4,%xmm5 + .byte 102,65,15,103,255 // packuswb %xmm15,%xmm7 + .byte 102,15,218,199 // pminub %xmm7,%xmm0 + .byte 102,15,218,245 // pminub %xmm5,%xmm6 + .byte 102,15,252,246 // paddb %xmm6,%xmm6 + .byte 102,15,252,192 // paddb %xmm0,%xmm0 + .byte 102,68,15,111,193 // movdqa %xmm1,%xmm8 + .byte 102,68,15,248,192 // psubb %xmm0,%xmm8 + .byte 102,65,15,111,230 // movdqa %xmm14,%xmm4 + .byte 102,15,248,230 // psubb %xmm6,%xmm4 + .byte 102,15,248,207 // psubb %xmm7,%xmm1 + .byte 102,68,15,248,245 // psubb %xmm5,%xmm14 + .byte 15,40,5,28,4,0,0 // movaps 0x41c(%rip),%xmm0 # 2580 <_sk_exclusion_sse41_8bit+0x407> + .byte 102,68,15,56,16,244 // pblendvb %xmm0,%xmm4,%xmm14 + .byte 102,65,15,56,16,200 // pblendvb %xmm0,%xmm8,%xmm1 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 102,65,15,111,198 // movdqa %xmm14,%xmm0 + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_exclusion_sse41_8bit +.globl _sk_exclusion_sse41_8bit +FUNCTION(_sk_exclusion_sse41_8bit) +_sk_exclusion_sse41_8bit: + .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 + .byte 102,15,111,248 // movdqa %xmm0,%xmm7 + .byte 102,65,15,104,249 // punpckhbw %xmm9,%xmm7 + .byte 102,68,15,56,48,208 // pmovzxbw %xmm0,%xmm10 + .byte 102,68,15,111,218 // movdqa %xmm2,%xmm11 + .byte 102,15,111,234 // movdqa %xmm2,%xmm5 + .byte 102,65,15,104,233 // punpckhbw %xmm9,%xmm5 + .byte 102,68,15,252,216 // paddb %xmm0,%xmm11 + .byte 102,68,15,111,5,231,3,0,0 // movdqa 0x3e7(%rip),%xmm8 # 2590 <_sk_exclusion_sse41_8bit+0x417> + .byte 102,65,15,56,0,192 // pshufb %xmm8,%xmm0 + .byte 102,15,56,48,240 // pmovzxbw %xmm0,%xmm6 + .byte 102,65,15,104,193 // punpckhbw %xmm9,%xmm0 + .byte 102,15,213,197 // pmullw %xmm5,%xmm0 + .byte 102,15,253,197 // paddw %xmm5,%xmm0 + .byte 102,15,213,239 // pmullw %xmm7,%xmm5 + .byte 102,15,253,239 // paddw %xmm7,%xmm5 + .byte 102,15,56,48,250 // pmovzxbw %xmm2,%xmm7 + .byte 102,15,213,247 // pmullw %xmm7,%xmm6 + .byte 102,15,253,247 // paddw %xmm7,%xmm6 + .byte 102,65,15,213,250 // pmullw %xmm10,%xmm7 + .byte 102,65,15,253,250 // paddw %xmm10,%xmm7 + .byte 102,15,111,225 // movdqa %xmm1,%xmm4 + .byte 102,65,15,104,225 // punpckhbw %xmm9,%xmm4 + .byte 102,15,113,213,8 // psrlw $0x8,%xmm5 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,15,103,253 // packuswb %xmm5,%xmm7 + .byte 102,15,252,255 // paddb %xmm7,%xmm7 + .byte 102,65,15,111,235 // movdqa %xmm11,%xmm5 + .byte 102,15,248,239 // psubb %xmm7,%xmm5 + .byte 102,68,15,56,48,209 // pmovzxbw %xmm1,%xmm10 + .byte 102,15,113,208,8 // psrlw $0x8,%xmm0 + .byte 102,15,113,214,8 // psrlw $0x8,%xmm6 + .byte 102,15,103,240 // packuswb %xmm0,%xmm6 + .byte 102,68,15,248,222 // psubb %xmm6,%xmm11 + .byte 15,40,5,124,3,0,0 // movaps 0x37c(%rip),%xmm0 # 25a0 <_sk_exclusion_sse41_8bit+0x427> + .byte 102,68,15,56,16,221 // pblendvb %xmm0,%xmm5,%xmm11 + .byte 102,15,111,235 // movdqa %xmm3,%xmm5 + .byte 102,15,111,243 // movdqa %xmm3,%xmm6 + .byte 102,65,15,104,241 // punpckhbw %xmm9,%xmm6 + .byte 102,15,252,233 // paddb %xmm1,%xmm5 + .byte 102,65,15,56,0,200 // pshufb %xmm8,%xmm1 + .byte 102,15,56,48,249 // pmovzxbw %xmm1,%xmm7 + .byte 102,65,15,104,201 // punpckhbw %xmm9,%xmm1 + .byte 102,15,213,206 // pmullw %xmm6,%xmm1 + .byte 102,15,253,206 // paddw %xmm6,%xmm1 + .byte 102,15,213,244 // pmullw %xmm4,%xmm6 + .byte 102,15,253,244 // paddw %xmm4,%xmm6 + .byte 102,15,56,48,227 // pmovzxbw %xmm3,%xmm4 + .byte 102,15,213,252 // pmullw %xmm4,%xmm7 + .byte 102,15,253,252 // paddw %xmm4,%xmm7 + .byte 102,65,15,213,226 // pmullw %xmm10,%xmm4 + .byte 102,65,15,253,226 // paddw %xmm10,%xmm4 + .byte 102,15,113,214,8 // psrlw $0x8,%xmm6 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,15,103,230 // packuswb %xmm6,%xmm4 + .byte 102,15,252,228 // paddb %xmm4,%xmm4 + .byte 102,15,111,245 // movdqa %xmm5,%xmm6 + .byte 102,15,248,244 // psubb %xmm4,%xmm6 + .byte 102,15,113,209,8 // psrlw $0x8,%xmm1 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,15,103,249 // packuswb %xmm1,%xmm7 + .byte 102,15,248,239 // psubb %xmm7,%xmm5 + .byte 102,15,56,16,238 // pblendvb %xmm0,%xmm6,%xmm5 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 102,65,15,111,195 // movdqa %xmm11,%xmm0 + .byte 102,15,111,205 // movdqa %xmm5,%xmm1 + .byte 255,224 // jmpq *%rax + BALIGN4 .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 1edb <_sk_darken_sse41_8bit+0x1b7> + .byte 127,67 // jg 22f7 <_sk_exclusion_sse41_8bit+0x17e> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 1edf <_sk_darken_sse41_8bit+0x1bb> + .byte 127,67 // jg 22fb <_sk_exclusion_sse41_8bit+0x182> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 1ee3 <_sk_darken_sse41_8bit+0x1bf> + .byte 127,67 // jg 22ff <_sk_exclusion_sse41_8bit+0x186> BALIGN16 .byte 0,0 // add %al,(%rax) @@ -62242,6 +62792,75 @@ BALIGN16 .byte 255 // (bad) .byte 255 // (bad) .byte 255,0 // incl (%rax) + .byte 3,3 // add (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 3,3 // add (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 3,3 // add (%rbx),%eax + .byte 3,3 // add (%rbx),%eax + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 7 // (bad) + .byte 11,11 // or (%rbx),%ecx + .byte 11,11 // or (%rbx),%ecx + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15 // (bad) + .byte 15,255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) BALIGN32 HIDDEN _sk_start_pipeline_sse2_8bit @@ -62326,7 +62945,7 @@ HIDDEN _sk_set_rgb_sse2_8bit FUNCTION(_sk_set_rgb_sse2_8bit) _sk_set_rgb_sse2_8bit: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,15,16,37,233,35,0,0 // movss 0x23e9(%rip),%xmm4 # 24ac <_sk_darken_sse2_8bit+0x1ee> + .byte 243,15,16,37,97,41,0,0 // movss 0x2961(%rip),%xmm4 # 2a24 <_sk_exclusion_sse2_8bit+0x197> .byte 243,15,16,40 // movss (%rax),%xmm5 .byte 243,15,89,236 // mulss %xmm4,%xmm5 .byte 243,72,15,44,205 // cvttss2si %xmm5,%rcx @@ -62341,7 +62960,7 @@ _sk_set_rgb_sse2_8bit: .byte 9,208 // or %edx,%eax .byte 102,15,110,224 // movd %eax,%xmm4 .byte 102,15,112,228,0 // pshufd $0x0,%xmm4,%xmm4 - .byte 102,15,111,45,189,35,0,0 // movdqa 0x23bd(%rip),%xmm5 # 24c0 <_sk_darken_sse2_8bit+0x202> + .byte 102,15,111,45,45,41,0,0 // movdqa 0x292d(%rip),%xmm5 # 2a30 <_sk_exclusion_sse2_8bit+0x1a3> .byte 102,15,219,205 // pand %xmm5,%xmm1 .byte 102,15,219,197 // pand %xmm5,%xmm0 .byte 102,15,235,196 // por %xmm4,%xmm0 @@ -62366,7 +62985,7 @@ _sk_premul_sse2_8bit: .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 .byte 242,15,112,192,95 // pshuflw $0x5f,%xmm0,%xmm0 .byte 243,15,112,248,95 // pshufhw $0x5f,%xmm0,%xmm7 - .byte 102,15,111,5,113,35,0,0 // movdqa 0x2371(%rip),%xmm0 # 24d0 <_sk_darken_sse2_8bit+0x212> + .byte 102,15,111,5,225,40,0,0 // movdqa 0x28e1(%rip),%xmm0 # 2a40 <_sk_exclusion_sse2_8bit+0x1b3> .byte 102,15,235,248 // por %xmm0,%xmm7 .byte 102,15,235,240 // por %xmm0,%xmm6 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 @@ -62887,7 +63506,7 @@ _sk_load_a8_sse2_8bit: .byte 117,48 // jne 7fd <_sk_load_a8_sse2_8bit+0x4d> .byte 243,66,15,126,4,2 // movq (%rdx,%r8,1),%xmm0 .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 - .byte 102,15,84,5,1,29,0,0 // andpd 0x1d01(%rip),%xmm0 # 24e0 <_sk_darken_sse2_8bit+0x222> + .byte 102,15,84,5,113,34,0,0 // andpd 0x2271(%rip),%xmm0 # 2a50 <_sk_exclusion_sse2_8bit+0x1c3> .byte 102,15,239,228 // pxor %xmm4,%xmm4 .byte 102,15,40,200 // movapd %xmm0,%xmm1 .byte 102,15,105,204 // punpckhwd %xmm4,%xmm1 @@ -62964,7 +63583,7 @@ _sk_load_a8_dst_sse2_8bit: .byte 117,48 // jne 8f1 <_sk_load_a8_dst_sse2_8bit+0x4d> .byte 243,66,15,126,20,2 // movq (%rdx,%r8,1),%xmm2 .byte 102,15,96,208 // punpcklbw %xmm0,%xmm2 - .byte 102,15,84,21,29,28,0,0 // andpd 0x1c1d(%rip),%xmm2 # 24f0 <_sk_darken_sse2_8bit+0x232> + .byte 102,15,84,21,141,33,0,0 // andpd 0x218d(%rip),%xmm2 # 2a60 <_sk_exclusion_sse2_8bit+0x1d3> .byte 102,15,239,228 // pxor %xmm4,%xmm4 .byte 102,15,40,218 // movapd %xmm2,%xmm3 .byte 102,15,105,220 // punpckhwd %xmm4,%xmm3 @@ -63048,7 +63667,7 @@ _sk_store_a8_sse2_8bit: .byte 102,15,107,229 // packssdw %xmm5,%xmm4 .byte 77,133,201 // test %r9,%r9 .byte 117,22 // jne 9f5 <_sk_store_a8_sse2_8bit+0x5d> - .byte 102,15,219,37,25,27,0,0 // pand 0x1b19(%rip),%xmm4 # 2500 <_sk_darken_sse2_8bit+0x242> + .byte 102,15,219,37,137,32,0,0 // pand 0x2089(%rip),%xmm4 # 2a70 <_sk_exclusion_sse2_8bit+0x1e3> .byte 102,15,103,228 // packuswb %xmm4,%xmm4 .byte 102,66,15,214,36,2 // movq %xmm4,(%rdx,%r8,1) .byte 72,173 // lods %ds:(%rsi),%rax @@ -63069,7 +63688,7 @@ _sk_store_a8_sse2_8bit: .byte 102,15,127,100,36,184 // movdqa %xmm4,-0x48(%rsp) .byte 138,68,36,188 // mov -0x44(%rsp),%al .byte 66,136,68,2,2 // mov %al,0x2(%rdx,%r8,1) - .byte 102,15,219,37,195,26,0,0 // pand 0x1ac3(%rip),%xmm4 # 2500 <_sk_darken_sse2_8bit+0x242> + .byte 102,15,219,37,51,32,0,0 // pand 0x2033(%rip),%xmm4 # 2a70 <_sk_exclusion_sse2_8bit+0x1e3> .byte 102,15,103,228 // packuswb %xmm4,%xmm4 .byte 102,15,126,224 // movd %xmm4,%eax .byte 102,66,137,4,2 // mov %ax,(%rdx,%r8,1) @@ -63083,7 +63702,7 @@ _sk_store_a8_sse2_8bit: .byte 102,15,127,100,36,200 // movdqa %xmm4,-0x38(%rsp) .byte 138,68,36,208 // mov -0x30(%rsp),%al .byte 66,136,68,2,4 // mov %al,0x4(%rdx,%r8,1) - .byte 102,15,219,37,127,26,0,0 // pand 0x1a7f(%rip),%xmm4 # 2500 <_sk_darken_sse2_8bit+0x242> + .byte 102,15,219,37,239,31,0,0 // pand 0x1fef(%rip),%xmm4 # 2a70 <_sk_exclusion_sse2_8bit+0x1e3> .byte 102,15,103,228 // packuswb %xmm4,%xmm4 .byte 102,66,15,126,36,2 // movd %xmm4,(%rdx,%r8,1) .byte 233,97,255,255,255 // jmpq 9f1 <_sk_store_a8_sse2_8bit+0x59> @@ -63093,7 +63712,7 @@ _sk_store_a8_sse2_8bit: .byte 255 // (bad) .byte 255 // (bad) .byte 255 // (bad) - .byte 233,255,255,255,218 // jmpq ffffffffdb000aa0 <_sk_darken_sse2_8bit+0xffffffffdaffe7e2> + .byte 233,255,255,255,218 // jmpq ffffffffdb000aa0 <_sk_exclusion_sse2_8bit+0xffffffffdaffe213> .byte 255 // (bad) .byte 255 // (bad) .byte 255,203 // dec %ebx @@ -63120,12 +63739,12 @@ _sk_load_g8_sse2_8bit: .byte 117,116 // jne b3d <_sk_load_g8_sse2_8bit+0x91> .byte 243,66,15,126,4,2 // movq (%rdx,%r8,1),%xmm0 .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 - .byte 102,15,84,5,53,26,0,0 // andpd 0x1a35(%rip),%xmm0 # 2510 <_sk_darken_sse2_8bit+0x252> + .byte 102,15,84,5,165,31,0,0 // andpd 0x1fa5(%rip),%xmm0 # 2a80 <_sk_exclusion_sse2_8bit+0x1f3> .byte 102,15,239,201 // pxor %xmm1,%xmm1 .byte 102,15,40,224 // movapd %xmm0,%xmm4 .byte 102,15,97,225 // punpcklwd %xmm1,%xmm4 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 - .byte 102,15,111,45,45,26,0,0 // movdqa 0x1a2d(%rip),%xmm5 # 2520 <_sk_darken_sse2_8bit+0x262> + .byte 102,15,111,45,157,31,0,0 // movdqa 0x1f9d(%rip),%xmm5 # 2a90 <_sk_exclusion_sse2_8bit+0x203> .byte 102,15,112,240,245 // pshufd $0xf5,%xmm0,%xmm6 .byte 102,15,244,197 // pmuludq %xmm5,%xmm0 .byte 102,15,112,200,232 // pshufd $0xe8,%xmm0,%xmm1 @@ -63138,7 +63757,7 @@ _sk_load_g8_sse2_8bit: .byte 102,15,244,245 // pmuludq %xmm5,%xmm6 .byte 102,15,112,230,232 // pshufd $0xe8,%xmm6,%xmm4 .byte 102,15,98,196 // punpckldq %xmm4,%xmm0 - .byte 102,15,111,37,255,25,0,0 // movdqa 0x19ff(%rip),%xmm4 # 2530 <_sk_darken_sse2_8bit+0x272> + .byte 102,15,111,37,111,31,0,0 // movdqa 0x1f6f(%rip),%xmm4 # 2aa0 <_sk_exclusion_sse2_8bit+0x213> .byte 102,15,235,196 // por %xmm4,%xmm0 .byte 102,15,235,204 // por %xmm4,%xmm1 .byte 72,173 // lods %ds:(%rsi),%rax @@ -63212,12 +63831,12 @@ _sk_load_g8_dst_sse2_8bit: .byte 117,116 // jne c7d <_sk_load_g8_dst_sse2_8bit+0x91> .byte 243,66,15,126,20,2 // movq (%rdx,%r8,1),%xmm2 .byte 102,15,96,208 // punpcklbw %xmm0,%xmm2 - .byte 102,15,84,21,37,25,0,0 // andpd 0x1925(%rip),%xmm2 # 2540 <_sk_darken_sse2_8bit+0x282> + .byte 102,15,84,21,149,30,0,0 // andpd 0x1e95(%rip),%xmm2 # 2ab0 <_sk_exclusion_sse2_8bit+0x223> .byte 102,15,239,219 // pxor %xmm3,%xmm3 .byte 102,15,40,226 // movapd %xmm2,%xmm4 .byte 102,15,97,227 // punpcklwd %xmm3,%xmm4 .byte 102,15,105,211 // punpckhwd %xmm3,%xmm2 - .byte 102,15,111,45,29,25,0,0 // movdqa 0x191d(%rip),%xmm5 # 2550 <_sk_darken_sse2_8bit+0x292> + .byte 102,15,111,45,141,30,0,0 // movdqa 0x1e8d(%rip),%xmm5 # 2ac0 <_sk_exclusion_sse2_8bit+0x233> .byte 102,15,112,242,245 // pshufd $0xf5,%xmm2,%xmm6 .byte 102,15,244,213 // pmuludq %xmm5,%xmm2 .byte 102,15,112,218,232 // pshufd $0xe8,%xmm2,%xmm3 @@ -63230,7 +63849,7 @@ _sk_load_g8_dst_sse2_8bit: .byte 102,15,244,245 // pmuludq %xmm5,%xmm6 .byte 102,15,112,230,232 // pshufd $0xe8,%xmm6,%xmm4 .byte 102,15,98,212 // punpckldq %xmm4,%xmm2 - .byte 102,15,111,37,239,24,0,0 // movdqa 0x18ef(%rip),%xmm4 # 2560 <_sk_darken_sse2_8bit+0x2a2> + .byte 102,15,111,37,95,30,0,0 // movdqa 0x1e5f(%rip),%xmm4 # 2ad0 <_sk_exclusion_sse2_8bit+0x243> .byte 102,15,235,212 // por %xmm4,%xmm2 .byte 102,15,235,220 // por %xmm4,%xmm3 .byte 72,173 // lods %ds:(%rsi),%rax @@ -63447,7 +64066,7 @@ _sk_scale_1_float_sse2_8bit: .byte 102,68,15,111,200 // movdqa %xmm0,%xmm9 .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,15,16,0 // movss (%rax),%xmm0 - .byte 243,15,89,5,32,21,0,0 // mulss 0x1520(%rip),%xmm0 # 24b0 <_sk_darken_sse2_8bit+0x1f2> + .byte 243,15,89,5,152,26,0,0 // mulss 0x1a98(%rip),%xmm0 # 2a28 <_sk_exclusion_sse2_8bit+0x19b> .byte 243,15,44,192 // cvttss2si %xmm0,%eax .byte 102,15,239,246 // pxor %xmm6,%xmm6 .byte 102,65,15,111,193 // movdqa %xmm9,%xmm0 @@ -63459,7 +64078,7 @@ _sk_scale_1_float_sse2_8bit: .byte 102,15,96,246 // punpcklbw %xmm6,%xmm6 .byte 242,15,112,246,0 // pshuflw $0x0,%xmm6,%xmm6 .byte 102,15,112,246,80 // pshufd $0x50,%xmm6,%xmm6 - .byte 102,15,219,53,167,21,0,0 // pand 0x15a7(%rip),%xmm6 # 2570 <_sk_darken_sse2_8bit+0x2b2> + .byte 102,15,219,53,23,27,0,0 // pand 0x1b17(%rip),%xmm6 # 2ae0 <_sk_exclusion_sse2_8bit+0x253> .byte 102,15,111,254 // movdqa %xmm6,%xmm7 .byte 102,65,15,213,248 // pmullw %xmm8,%xmm7 .byte 102,15,111,230 // movdqa %xmm6,%xmm4 @@ -63495,7 +64114,7 @@ _sk_scale_u8_sse2_8bit: .byte 15,133,239,0,0,0 // jne 1129 <_sk_scale_u8_sse2_8bit+0x110> .byte 243,66,15,126,36,2 // movq (%rdx,%r8,1),%xmm4 .byte 102,15,96,224 // punpcklbw %xmm0,%xmm4 - .byte 102,15,84,37,52,21,0,0 // andpd 0x1534(%rip),%xmm4 # 2580 <_sk_darken_sse2_8bit+0x2c2> + .byte 102,15,84,37,164,26,0,0 // andpd 0x1aa4(%rip),%xmm4 # 2af0 <_sk_exclusion_sse2_8bit+0x263> .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 102,15,40,236 // movapd %xmm4,%xmm5 .byte 102,65,15,105,232 // punpckhwd %xmm8,%xmm5 @@ -63604,7 +64223,7 @@ FUNCTION(_sk_lerp_1_float_sse2_8bit) _sk_lerp_1_float_sse2_8bit: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,15,16,32 // movss (%rax),%xmm4 - .byte 243,15,89,37,202,18,0,0 // mulss 0x12ca(%rip),%xmm4 # 24b4 <_sk_darken_sse2_8bit+0x1f6> + .byte 243,15,89,37,66,24,0,0 // mulss 0x1842(%rip),%xmm4 # 2a2c <_sk_exclusion_sse2_8bit+0x19f> .byte 243,15,44,196 // cvttss2si %xmm4,%eax .byte 102,15,110,224 // movd %eax,%xmm4 .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 @@ -63617,7 +64236,7 @@ _sk_lerp_1_float_sse2_8bit: .byte 102,68,15,111,217 // movdqa %xmm1,%xmm11 .byte 102,69,15,96,217 // punpcklbw %xmm9,%xmm11 .byte 102,65,15,104,201 // punpckhbw %xmm9,%xmm1 - .byte 102,15,111,53,101,19,0,0 // movdqa 0x1365(%rip),%xmm6 # 2590 <_sk_darken_sse2_8bit+0x2d2> + .byte 102,15,111,53,213,24,0,0 // movdqa 0x18d5(%rip),%xmm6 # 2b00 <_sk_exclusion_sse2_8bit+0x273> .byte 102,65,15,219,240 // pand %xmm8,%xmm6 .byte 102,15,111,230 // movdqa %xmm6,%xmm4 .byte 102,15,213,225 // pmullw %xmm1,%xmm4 @@ -63685,7 +64304,7 @@ _sk_lerp_u8_sse2_8bit: .byte 15,133,141,1,0,0 // jne 14c0 <_sk_lerp_u8_sse2_8bit+0x1ae> .byte 243,66,15,126,44,2 // movq (%rdx,%r8,1),%xmm5 .byte 102,15,96,232 // punpcklbw %xmm0,%xmm5 - .byte 102,15,84,45,91,18,0,0 // andpd 0x125b(%rip),%xmm5 # 25a0 <_sk_darken_sse2_8bit+0x2e2> + .byte 102,15,84,45,203,23,0,0 // andpd 0x17cb(%rip),%xmm5 # 2b10 <_sk_exclusion_sse2_8bit+0x283> .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 102,15,40,229 // movapd %xmm5,%xmm4 .byte 102,65,15,105,224 // punpckhwd %xmm8,%xmm4 @@ -63843,7 +64462,7 @@ HIDDEN _sk_black_color_sse2_8bit FUNCTION(_sk_black_color_sse2_8bit) _sk_black_color_sse2_8bit: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 15,40,5,35,16,0,0 // movaps 0x1023(%rip),%xmm0 # 25b0 <_sk_darken_sse2_8bit+0x2f2> + .byte 15,40,5,147,21,0,0 // movaps 0x1593(%rip),%xmm0 # 2b20 <_sk_exclusion_sse2_8bit+0x293> .byte 15,40,200 // movaps %xmm0,%xmm1 .byte 255,224 // jmpq *%rax @@ -64735,7 +65354,115 @@ _sk_darken_sse2_8bit: .byte 102,65,15,248,234 // psubb %xmm10,%xmm5 .byte 102,15,248,207 // psubb %xmm7,%xmm1 .byte 102,15,248,196 // psubb %xmm4,%xmm0 - .byte 102,15,111,37,62,1,0,0 // movdqa 0x13e(%rip),%xmm4 # 25c0 <_sk_darken_sse2_8bit+0x302> + .byte 102,15,111,37,174,6,0,0 // movdqa 0x6ae(%rip),%xmm4 # 2b30 <_sk_exclusion_sse2_8bit+0x2a3> + .byte 102,15,219,236 // pand %xmm4,%xmm5 + .byte 102,15,111,252 // movdqa %xmm4,%xmm7 + .byte 102,15,223,248 // pandn %xmm0,%xmm7 + .byte 102,15,235,239 // por %xmm7,%xmm5 + .byte 102,15,219,244 // pand %xmm4,%xmm6 + .byte 102,15,223,225 // pandn %xmm1,%xmm4 + .byte 102,15,235,244 // por %xmm4,%xmm6 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 102,15,111,197 // movdqa %xmm5,%xmm0 + .byte 102,15,111,206 // movdqa %xmm6,%xmm1 + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_lighten_sse2_8bit +.globl _sk_lighten_sse2_8bit +FUNCTION(_sk_lighten_sse2_8bit) +_sk_lighten_sse2_8bit: + .byte 242,15,112,227,231 // pshuflw $0xe7,%xmm3,%xmm4 + .byte 243,15,112,228,231 // pshufhw $0xe7,%xmm4,%xmm4 + .byte 102,15,112,228,232 // pshufd $0xe8,%xmm4,%xmm4 + .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 + .byte 242,15,112,228,95 // pshuflw $0x5f,%xmm4,%xmm4 + .byte 243,15,112,228,95 // pshufhw $0x5f,%xmm4,%xmm4 + .byte 242,15,112,234,231 // pshuflw $0xe7,%xmm2,%xmm5 + .byte 243,15,112,237,231 // pshufhw $0xe7,%xmm5,%xmm5 + .byte 102,15,112,237,232 // pshufd $0xe8,%xmm5,%xmm5 + .byte 102,15,96,237 // punpcklbw %xmm5,%xmm5 + .byte 242,15,112,237,95 // pshuflw $0x5f,%xmm5,%xmm5 + .byte 243,15,112,253,95 // pshufhw $0x5f,%xmm5,%xmm7 + .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 + .byte 102,15,111,240 // movdqa %xmm0,%xmm6 + .byte 102,65,15,96,240 // punpcklbw %xmm8,%xmm6 + .byte 102,68,15,111,200 // movdqa %xmm0,%xmm9 + .byte 102,69,15,104,200 // punpckhbw %xmm8,%xmm9 + .byte 102,68,15,111,217 // movdqa %xmm1,%xmm11 + .byte 102,69,15,96,216 // punpcklbw %xmm8,%xmm11 + .byte 102,68,15,111,225 // movdqa %xmm1,%xmm12 + .byte 102,69,15,104,224 // punpckhbw %xmm8,%xmm12 + .byte 102,68,15,111,215 // movdqa %xmm7,%xmm10 + .byte 102,69,15,96,208 // punpcklbw %xmm8,%xmm10 + .byte 102,65,15,104,248 // punpckhbw %xmm8,%xmm7 + .byte 102,68,15,111,236 // movdqa %xmm4,%xmm13 + .byte 102,69,15,96,232 // punpcklbw %xmm8,%xmm13 + .byte 102,65,15,104,224 // punpckhbw %xmm8,%xmm4 + .byte 102,65,15,213,228 // pmullw %xmm12,%xmm4 + .byte 102,69,15,213,235 // pmullw %xmm11,%xmm13 + .byte 102,65,15,213,249 // pmullw %xmm9,%xmm7 + .byte 102,68,15,213,214 // pmullw %xmm6,%xmm10 + .byte 102,68,15,253,214 // paddw %xmm6,%xmm10 + .byte 102,65,15,253,249 // paddw %xmm9,%xmm7 + .byte 102,69,15,253,235 // paddw %xmm11,%xmm13 + .byte 102,65,15,253,228 // paddw %xmm12,%xmm4 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,65,15,113,213,8 // psrlw $0x8,%xmm13 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 + .byte 102,68,15,103,215 // packuswb %xmm7,%xmm10 + .byte 102,68,15,103,236 // packuswb %xmm4,%xmm13 + .byte 242,15,112,225,231 // pshuflw $0xe7,%xmm1,%xmm4 + .byte 243,15,112,228,231 // pshufhw $0xe7,%xmm4,%xmm4 + .byte 102,15,112,228,232 // pshufd $0xe8,%xmm4,%xmm4 + .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 + .byte 242,15,112,228,95 // pshuflw $0x5f,%xmm4,%xmm4 + .byte 243,68,15,112,220,95 // pshufhw $0x5f,%xmm4,%xmm11 + .byte 242,15,112,224,231 // pshuflw $0xe7,%xmm0,%xmm4 + .byte 243,15,112,228,231 // pshufhw $0xe7,%xmm4,%xmm4 + .byte 102,15,112,228,232 // pshufd $0xe8,%xmm4,%xmm4 + .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 + .byte 242,15,112,228,95 // pshuflw $0x5f,%xmm4,%xmm4 + .byte 243,68,15,112,228,95 // pshufhw $0x5f,%xmm4,%xmm12 + .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9 + .byte 102,15,252,194 // paddb %xmm2,%xmm0 + .byte 102,15,111,242 // movdqa %xmm2,%xmm6 + .byte 102,65,15,96,240 // punpcklbw %xmm8,%xmm6 + .byte 102,69,15,104,200 // punpckhbw %xmm8,%xmm9 + .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14 + .byte 102,15,252,203 // paddb %xmm3,%xmm1 + .byte 102,15,111,235 // movdqa %xmm3,%xmm5 + .byte 102,65,15,96,232 // punpcklbw %xmm8,%xmm5 + .byte 102,69,15,104,240 // punpckhbw %xmm8,%xmm14 + .byte 102,65,15,111,228 // movdqa %xmm12,%xmm4 + .byte 102,65,15,96,224 // punpcklbw %xmm8,%xmm4 + .byte 102,69,15,104,224 // punpckhbw %xmm8,%xmm12 + .byte 102,65,15,111,251 // movdqa %xmm11,%xmm7 + .byte 102,65,15,96,248 // punpcklbw %xmm8,%xmm7 + .byte 102,69,15,104,216 // punpckhbw %xmm8,%xmm11 + .byte 102,69,15,213,222 // pmullw %xmm14,%xmm11 + .byte 102,15,213,253 // pmullw %xmm5,%xmm7 + .byte 102,69,15,213,225 // pmullw %xmm9,%xmm12 + .byte 102,15,213,230 // pmullw %xmm6,%xmm4 + .byte 102,15,253,230 // paddw %xmm6,%xmm4 + .byte 102,69,15,253,225 // paddw %xmm9,%xmm12 + .byte 102,15,253,253 // paddw %xmm5,%xmm7 + .byte 102,69,15,253,222 // paddw %xmm14,%xmm11 + .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,65,15,113,212,8 // psrlw $0x8,%xmm12 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,65,15,103,228 // packuswb %xmm12,%xmm4 + .byte 102,65,15,103,251 // packuswb %xmm11,%xmm7 + .byte 102,68,15,218,212 // pminub %xmm4,%xmm10 + .byte 102,68,15,218,239 // pminub %xmm7,%xmm13 + .byte 102,15,111,241 // movdqa %xmm1,%xmm6 + .byte 102,65,15,248,245 // psubb %xmm13,%xmm6 + .byte 102,15,111,232 // movdqa %xmm0,%xmm5 + .byte 102,65,15,248,234 // psubb %xmm10,%xmm5 + .byte 102,15,248,207 // psubb %xmm7,%xmm1 + .byte 102,15,248,196 // psubb %xmm4,%xmm0 + .byte 102,15,111,37,210,4,0,0 // movdqa 0x4d2(%rip),%xmm4 # 2b40 <_sk_exclusion_sse2_8bit+0x2b3> .byte 102,15,219,236 // pand %xmm4,%xmm5 .byte 102,15,111,252 // movdqa %xmm4,%xmm7 .byte 102,15,223,248 // pandn %xmm0,%xmm7 @@ -64748,13 +65475,213 @@ _sk_darken_sse2_8bit: .byte 102,15,111,206 // movdqa %xmm6,%xmm1 .byte 255,224 // jmpq *%rax +HIDDEN _sk_difference_sse2_8bit +.globl _sk_difference_sse2_8bit +FUNCTION(_sk_difference_sse2_8bit) +_sk_difference_sse2_8bit: + .byte 242,15,112,227,231 // pshuflw $0xe7,%xmm3,%xmm4 + .byte 243,15,112,228,231 // pshufhw $0xe7,%xmm4,%xmm4 + .byte 102,15,112,228,232 // pshufd $0xe8,%xmm4,%xmm4 + .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 + .byte 242,15,112,228,95 // pshuflw $0x5f,%xmm4,%xmm4 + .byte 243,15,112,228,95 // pshufhw $0x5f,%xmm4,%xmm4 + .byte 242,15,112,234,231 // pshuflw $0xe7,%xmm2,%xmm5 + .byte 243,15,112,237,231 // pshufhw $0xe7,%xmm5,%xmm5 + .byte 102,15,112,237,232 // pshufd $0xe8,%xmm5,%xmm5 + .byte 102,15,96,237 // punpcklbw %xmm5,%xmm5 + .byte 242,15,112,237,95 // pshuflw $0x5f,%xmm5,%xmm5 + .byte 243,15,112,253,95 // pshufhw $0x5f,%xmm5,%xmm7 + .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 + .byte 102,68,15,111,200 // movdqa %xmm0,%xmm9 + .byte 102,69,15,96,200 // punpcklbw %xmm8,%xmm9 + .byte 102,68,15,111,208 // movdqa %xmm0,%xmm10 + .byte 102,69,15,104,208 // punpckhbw %xmm8,%xmm10 + .byte 102,68,15,111,217 // movdqa %xmm1,%xmm11 + .byte 102,69,15,96,216 // punpcklbw %xmm8,%xmm11 + .byte 102,68,15,111,225 // movdqa %xmm1,%xmm12 + .byte 102,69,15,104,224 // punpckhbw %xmm8,%xmm12 + .byte 102,68,15,111,239 // movdqa %xmm7,%xmm13 + .byte 102,69,15,96,232 // punpcklbw %xmm8,%xmm13 + .byte 102,65,15,104,248 // punpckhbw %xmm8,%xmm7 + .byte 102,68,15,111,244 // movdqa %xmm4,%xmm14 + .byte 102,69,15,96,240 // punpcklbw %xmm8,%xmm14 + .byte 102,65,15,104,224 // punpckhbw %xmm8,%xmm4 + .byte 102,65,15,213,228 // pmullw %xmm12,%xmm4 + .byte 102,69,15,213,243 // pmullw %xmm11,%xmm14 + .byte 102,65,15,213,250 // pmullw %xmm10,%xmm7 + .byte 102,69,15,213,233 // pmullw %xmm9,%xmm13 + .byte 102,69,15,253,233 // paddw %xmm9,%xmm13 + .byte 102,65,15,253,250 // paddw %xmm10,%xmm7 + .byte 102,69,15,253,243 // paddw %xmm11,%xmm14 + .byte 102,65,15,253,228 // paddw %xmm12,%xmm4 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,65,15,113,214,8 // psrlw $0x8,%xmm14 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,65,15,113,213,8 // psrlw $0x8,%xmm13 + .byte 102,68,15,103,239 // packuswb %xmm7,%xmm13 + .byte 102,68,15,103,244 // packuswb %xmm4,%xmm14 + .byte 242,15,112,225,231 // pshuflw $0xe7,%xmm1,%xmm4 + .byte 243,15,112,228,231 // pshufhw $0xe7,%xmm4,%xmm4 + .byte 102,15,112,228,232 // pshufd $0xe8,%xmm4,%xmm4 + .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 + .byte 242,15,112,228,95 // pshuflw $0x5f,%xmm4,%xmm4 + .byte 243,68,15,112,212,95 // pshufhw $0x5f,%xmm4,%xmm10 + .byte 242,15,112,224,231 // pshuflw $0xe7,%xmm0,%xmm4 + .byte 243,15,112,228,231 // pshufhw $0xe7,%xmm4,%xmm4 + .byte 102,15,112,228,232 // pshufd $0xe8,%xmm4,%xmm4 + .byte 102,15,96,228 // punpcklbw %xmm4,%xmm4 + .byte 242,15,112,228,95 // pshuflw $0x5f,%xmm4,%xmm4 + .byte 243,68,15,112,220,95 // pshufhw $0x5f,%xmm4,%xmm11 + .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9 + .byte 102,15,252,194 // paddb %xmm2,%xmm0 + .byte 102,15,111,242 // movdqa %xmm2,%xmm6 + .byte 102,65,15,96,240 // punpcklbw %xmm8,%xmm6 + .byte 102,69,15,104,200 // punpckhbw %xmm8,%xmm9 + .byte 102,68,15,111,227 // movdqa %xmm3,%xmm12 + .byte 102,15,252,203 // paddb %xmm3,%xmm1 + .byte 102,15,111,235 // movdqa %xmm3,%xmm5 + .byte 102,65,15,96,232 // punpcklbw %xmm8,%xmm5 + .byte 102,69,15,104,224 // punpckhbw %xmm8,%xmm12 + .byte 102,65,15,111,227 // movdqa %xmm11,%xmm4 + .byte 102,65,15,96,224 // punpcklbw %xmm8,%xmm4 + .byte 102,69,15,104,216 // punpckhbw %xmm8,%xmm11 + .byte 102,65,15,111,250 // movdqa %xmm10,%xmm7 + .byte 102,65,15,96,248 // punpcklbw %xmm8,%xmm7 + .byte 102,69,15,104,208 // punpckhbw %xmm8,%xmm10 + .byte 102,69,15,213,212 // pmullw %xmm12,%xmm10 + .byte 102,15,213,253 // pmullw %xmm5,%xmm7 + .byte 102,69,15,213,217 // pmullw %xmm9,%xmm11 + .byte 102,15,213,230 // pmullw %xmm6,%xmm4 + .byte 102,15,253,230 // paddw %xmm6,%xmm4 + .byte 102,69,15,253,217 // paddw %xmm9,%xmm11 + .byte 102,15,253,253 // paddw %xmm5,%xmm7 + .byte 102,69,15,253,212 // paddw %xmm12,%xmm10 + .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,65,15,103,227 // packuswb %xmm11,%xmm4 + .byte 102,65,15,103,250 // packuswb %xmm10,%xmm7 + .byte 102,68,15,218,247 // pminub %xmm7,%xmm14 + .byte 102,68,15,218,236 // pminub %xmm4,%xmm13 + .byte 102,69,15,252,237 // paddb %xmm13,%xmm13 + .byte 102,69,15,252,246 // paddb %xmm14,%xmm14 + .byte 102,15,111,241 // movdqa %xmm1,%xmm6 + .byte 102,65,15,248,246 // psubb %xmm14,%xmm6 + .byte 102,15,111,232 // movdqa %xmm0,%xmm5 + .byte 102,65,15,248,237 // psubb %xmm13,%xmm5 + .byte 102,15,248,207 // psubb %xmm7,%xmm1 + .byte 102,15,248,196 // psubb %xmm4,%xmm0 + .byte 102,15,111,37,235,2,0,0 // movdqa 0x2eb(%rip),%xmm4 # 2b50 <_sk_exclusion_sse2_8bit+0x2c3> + .byte 102,15,219,236 // pand %xmm4,%xmm5 + .byte 102,15,111,252 // movdqa %xmm4,%xmm7 + .byte 102,15,223,248 // pandn %xmm0,%xmm7 + .byte 102,15,235,239 // por %xmm7,%xmm5 + .byte 102,15,219,244 // pand %xmm4,%xmm6 + .byte 102,15,223,225 // pandn %xmm1,%xmm4 + .byte 102,15,235,244 // por %xmm4,%xmm6 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 102,15,111,197 // movdqa %xmm5,%xmm0 + .byte 102,15,111,206 // movdqa %xmm6,%xmm1 + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_exclusion_sse2_8bit +.globl _sk_exclusion_sse2_8bit +FUNCTION(_sk_exclusion_sse2_8bit) +_sk_exclusion_sse2_8bit: + .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 + .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 + .byte 102,68,15,111,226 // movdqa %xmm2,%xmm12 + .byte 102,68,15,252,224 // paddb %xmm0,%xmm12 + .byte 242,68,15,112,208,231 // pshuflw $0xe7,%xmm0,%xmm10 + .byte 102,15,111,224 // movdqa %xmm0,%xmm4 + .byte 102,65,15,96,225 // punpcklbw %xmm9,%xmm4 + .byte 102,69,15,104,193 // punpckhbw %xmm9,%xmm8 + .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13 + .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14 + .byte 102,68,15,252,241 // paddb %xmm1,%xmm14 + .byte 242,68,15,112,217,231 // pshuflw $0xe7,%xmm1,%xmm11 + .byte 102,68,15,111,249 // movdqa %xmm1,%xmm15 + .byte 102,69,15,96,249 // punpcklbw %xmm9,%xmm15 + .byte 102,15,111,234 // movdqa %xmm2,%xmm5 + .byte 102,15,111,250 // movdqa %xmm2,%xmm7 + .byte 102,65,15,96,249 // punpcklbw %xmm9,%xmm7 + .byte 102,65,15,104,233 // punpckhbw %xmm9,%xmm5 + .byte 243,65,15,112,194,231 // pshufhw $0xe7,%xmm10,%xmm0 + .byte 102,15,112,192,232 // pshufd $0xe8,%xmm0,%xmm0 + .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 + .byte 242,15,112,192,95 // pshuflw $0x5f,%xmm0,%xmm0 + .byte 243,68,15,112,208,95 // pshufhw $0x5f,%xmm0,%xmm10 + .byte 102,65,15,111,194 // movdqa %xmm10,%xmm0 + .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 + .byte 102,15,213,199 // pmullw %xmm7,%xmm0 + .byte 102,15,253,199 // paddw %xmm7,%xmm0 + .byte 102,15,213,252 // pmullw %xmm4,%xmm7 + .byte 102,15,253,252 // paddw %xmm4,%xmm7 + .byte 102,15,111,227 // movdqa %xmm3,%xmm4 + .byte 102,65,15,96,225 // punpcklbw %xmm9,%xmm4 + .byte 102,69,15,104,209 // punpckhbw %xmm9,%xmm10 + .byte 102,68,15,213,213 // pmullw %xmm5,%xmm10 + .byte 102,68,15,253,213 // paddw %xmm5,%xmm10 + .byte 102,65,15,213,232 // pmullw %xmm8,%xmm5 + .byte 102,65,15,253,232 // paddw %xmm8,%xmm5 + .byte 243,65,15,112,243,231 // pshufhw $0xe7,%xmm11,%xmm6 + .byte 102,15,112,246,232 // pshufd $0xe8,%xmm6,%xmm6 + .byte 102,15,96,246 // punpcklbw %xmm6,%xmm6 + .byte 242,15,112,246,95 // pshuflw $0x5f,%xmm6,%xmm6 + .byte 243,15,112,206,95 // pshufhw $0x5f,%xmm6,%xmm1 + .byte 102,15,111,241 // movdqa %xmm1,%xmm6 + .byte 102,65,15,96,241 // punpcklbw %xmm9,%xmm6 + .byte 102,15,213,244 // pmullw %xmm4,%xmm6 + .byte 102,15,253,244 // paddw %xmm4,%xmm6 + .byte 102,65,15,213,231 // pmullw %xmm15,%xmm4 + .byte 102,65,15,253,231 // paddw %xmm15,%xmm4 + .byte 102,68,15,111,195 // movdqa %xmm3,%xmm8 + .byte 102,69,15,104,233 // punpckhbw %xmm9,%xmm13 + .byte 102,69,15,104,193 // punpckhbw %xmm9,%xmm8 + .byte 102,65,15,104,201 // punpckhbw %xmm9,%xmm1 + .byte 102,65,15,213,200 // pmullw %xmm8,%xmm1 + .byte 102,65,15,253,200 // paddw %xmm8,%xmm1 + .byte 102,69,15,213,197 // pmullw %xmm13,%xmm8 + .byte 102,69,15,253,197 // paddw %xmm13,%xmm8 + .byte 102,15,113,213,8 // psrlw $0x8,%xmm5 + .byte 102,15,113,215,8 // psrlw $0x8,%xmm7 + .byte 102,15,103,253 // packuswb %xmm5,%xmm7 + .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8 + .byte 102,15,113,212,8 // psrlw $0x8,%xmm4 + .byte 102,65,15,103,224 // packuswb %xmm8,%xmm4 + .byte 102,15,252,255 // paddb %xmm7,%xmm7 + .byte 102,15,252,228 // paddb %xmm4,%xmm4 + .byte 102,65,15,111,238 // movdqa %xmm14,%xmm5 + .byte 102,15,248,236 // psubb %xmm4,%xmm5 + .byte 102,65,15,111,228 // movdqa %xmm12,%xmm4 + .byte 102,15,248,231 // psubb %xmm7,%xmm4 + .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 + .byte 102,15,113,208,8 // psrlw $0x8,%xmm0 + .byte 102,65,15,103,194 // packuswb %xmm10,%xmm0 + .byte 102,15,113,209,8 // psrlw $0x8,%xmm1 + .byte 102,15,113,214,8 // psrlw $0x8,%xmm6 + .byte 102,15,103,241 // packuswb %xmm1,%xmm6 + .byte 102,68,15,248,246 // psubb %xmm6,%xmm14 + .byte 102,68,15,248,224 // psubb %xmm0,%xmm12 + .byte 102,15,111,13,96,1,0,0 // movdqa 0x160(%rip),%xmm1 # 2b60 <_sk_exclusion_sse2_8bit+0x2d3> + .byte 102,15,111,193 // movdqa %xmm1,%xmm0 + .byte 102,65,15,223,196 // pandn %xmm12,%xmm0 + .byte 102,15,219,225 // pand %xmm1,%xmm4 + .byte 102,15,235,196 // por %xmm4,%xmm0 + .byte 102,15,219,233 // pand %xmm1,%xmm5 + .byte 102,65,15,223,206 // pandn %xmm14,%xmm1 + .byte 102,15,235,205 // por %xmm5,%xmm1 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + BALIGN4 .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 24f3 <_sk_darken_sse2_8bit+0x235> + .byte 127,67 // jg 2a6b <_sk_exclusion_sse2_8bit+0x1de> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 24f7 <_sk_darken_sse2_8bit+0x239> + .byte 127,67 // jg 2a6f <_sk_exclusion_sse2_8bit+0x1e2> .byte 0,0 // add %al,(%rax) - .byte 127,67 // jg 24fb <_sk_darken_sse2_8bit+0x23d> + .byte 127,67 // jg 2a73 <_sk_exclusion_sse2_8bit+0x1e6> BALIGN16 .byte 0,0 // add %al,(%rax) @@ -64897,6 +65824,42 @@ BALIGN16 .byte 255 // (bad) .byte 255 // (bad) .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,0 // incl (%rax) #elif defined(__i386__) BALIGN32 diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 0faff48dcb..b65679526e 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -39059,7 +39059,7 @@ _sk_uniform_color_hsw_8bit LABEL PROC PUBLIC _sk_set_rgb_hsw_8bit _sk_set_rgb_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 197,250,16,37,201,40,0,0 ; vmovss 0x28c9(%rip),%xmm4 # 2a4c <_sk_darken_hsw_8bit+0x189> + DB 197,250,16,37,69,45,0,0 ; vmovss 0x2d45(%rip),%xmm4 # 2ec8 <_sk_exclusion_hsw_8bit+0x168> DB 197,218,89,40 ; vmulss (%rax),%xmm4,%xmm5 DB 196,225,250,44,205 ; vcvttss2si %xmm5,%rcx DB 197,218,89,104,4 ; vmulss 0x4(%rax),%xmm4,%xmm5 @@ -39072,7 +39072,7 @@ _sk_set_rgb_hsw_8bit LABEL PROC DB 9,208 ; or %edx,%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 DB 196,226,125,88,228 ; vpbroadcastd %xmm4,%ymm4 - DB 197,253,111,45,197,40,0,0 ; vmovdqa 0x28c5(%rip),%ymm5 # 2a80 <_sk_darken_hsw_8bit+0x1bd> + DB 197,253,111,45,69,45,0,0 ; vmovdqa 0x2d45(%rip),%ymm5 # 2f00 <_sk_exclusion_hsw_8bit+0x1a0> DB 197,245,219,205 ; vpand %ymm5,%ymm1,%ymm1 DB 197,253,219,197 ; vpand %ymm5,%ymm0,%ymm0 DB 197,221,235,192 ; vpor %ymm0,%ymm4,%ymm0 @@ -39082,10 +39082,10 @@ _sk_set_rgb_hsw_8bit LABEL PROC PUBLIC _sk_premul_hsw_8bit _sk_premul_hsw_8bit LABEL PROC - DB 197,253,111,37,201,40,0,0 ; vmovdqa 0x28c9(%rip),%ymm4 # 2aa0 <_sk_darken_hsw_8bit+0x1dd> + DB 197,253,111,37,73,45,0,0 ; vmovdqa 0x2d49(%rip),%ymm4 # 2f20 <_sk_exclusion_hsw_8bit+0x1c0> DB 196,226,125,0,236 ; vpshufb %ymm4,%ymm0,%ymm5 DB 196,226,117,0,228 ; vpshufb %ymm4,%ymm1,%ymm4 - DB 197,253,111,53,215,40,0,0 ; vmovdqa 0x28d7(%rip),%ymm6 # 2ac0 <_sk_darken_hsw_8bit+0x1fd> + DB 197,253,111,53,87,45,0,0 ; vmovdqa 0x2d57(%rip),%ymm6 # 2f40 <_sk_exclusion_hsw_8bit+0x1e0> DB 197,221,235,230 ; vpor %ymm6,%ymm4,%ymm4 DB 197,213,235,238 ; vpor %ymm6,%ymm5,%ymm5 DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 @@ -39123,7 +39123,7 @@ _sk_premul_hsw_8bit LABEL PROC PUBLIC _sk_swap_rb_hsw_8bit _sk_swap_rb_hsw_8bit LABEL PROC - DB 197,253,111,37,79,40,0,0 ; vmovdqa 0x284f(%rip),%ymm4 # 2ae0 <_sk_darken_hsw_8bit+0x21d> + DB 197,253,111,37,207,44,0,0 ; vmovdqa 0x2ccf(%rip),%ymm4 # 2f60 <_sk_exclusion_hsw_8bit+0x200> DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -39463,7 +39463,7 @@ _sk_load_bgra_hsw_8bit LABEL PROC DB 117,35 ; jne 774 <_sk_load_bgra_hsw_8bit+0x44> DB 196,161,126,111,76,130,32 ; vmovdqu 0x20(%rdx,%r8,4),%ymm1 DB 196,161,126,111,4,130 ; vmovdqu (%rdx,%r8,4),%ymm0 - DB 197,253,111,37,154,35,0,0 ; vmovdqa 0x239a(%rip),%ymm4 # 2b00 <_sk_darken_hsw_8bit+0x23d> + DB 197,253,111,37,26,40,0,0 ; vmovdqa 0x281a(%rip),%ymm4 # 2f80 <_sk_exclusion_hsw_8bit+0x220> DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -39576,7 +39576,7 @@ _sk_load_bgra_dst_hsw_8bit LABEL PROC DB 117,35 ; jne 92c <_sk_load_bgra_dst_hsw_8bit+0x44> DB 196,161,126,111,92,130,32 ; vmovdqu 0x20(%rdx,%r8,4),%ymm3 DB 196,161,126,111,20,130 ; vmovdqu (%rdx,%r8,4),%ymm2 - DB 197,253,111,37,2,34,0,0 ; vmovdqa 0x2202(%rip),%ymm4 # 2b20 <_sk_darken_hsw_8bit+0x25d> + DB 197,253,111,37,130,38,0,0 ; vmovdqa 0x2682(%rip),%ymm4 # 2fa0 <_sk_exclusion_hsw_8bit+0x240> DB 196,226,109,0,212 ; vpshufb %ymm4,%ymm2,%ymm2 DB 196,226,101,0,220 ; vpshufb %ymm4,%ymm3,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -39685,7 +39685,7 @@ _sk_store_bgra_hsw_8bit LABEL PROC DB 72,15,175,209 ; imul %rcx,%rdx DB 72,193,226,2 ; shl $0x2,%rdx DB 72,3,16 ; add (%rax),%rdx - DB 197,253,111,37,124,32,0,0 ; vmovdqa 0x207c(%rip),%ymm4 # 2b40 <_sk_darken_hsw_8bit+0x27d> + DB 197,253,111,37,252,36,0,0 ; vmovdqa 0x24fc(%rip),%ymm4 # 2fc0 <_sk_exclusion_hsw_8bit+0x260> DB 196,226,117,0,236 ; vpshufb %ymm4,%ymm1,%ymm5 DB 196,226,125,0,228 ; vpshufb %ymm4,%ymm0,%ymm4 DB 77,133,201 ; test %r9,%r9 @@ -39965,10 +39965,10 @@ _sk_store_a8_hsw_8bit LABEL PROC DB 72,99,87,8 ; movslq 0x8(%rdi),%rdx DB 72,15,175,209 ; imul %rcx,%rdx DB 72,3,16 ; add (%rax),%rdx - DB 197,253,111,37,168,28,0,0 ; vmovdqa 0x1ca8(%rip),%ymm4 # 2b60 <_sk_darken_hsw_8bit+0x29d> + DB 197,253,111,37,40,33,0,0 ; vmovdqa 0x2128(%rip),%ymm4 # 2fe0 <_sk_exclusion_hsw_8bit+0x280> DB 196,226,117,0,236 ; vpshufb %ymm4,%ymm1,%ymm5 DB 196,227,253,0,237,232 ; vpermq $0xe8,%ymm5,%ymm5 - DB 197,249,111,53,181,30,0,0 ; vmovdqa 0x1eb5(%rip),%xmm6 # 2d80 <_sk_darken_hsw_8bit+0x4bd> + DB 197,249,111,53,149,35,0,0 ; vmovdqa 0x2395(%rip),%xmm6 # 3260 <_sk_exclusion_hsw_8bit+0x500> DB 196,226,81,0,238 ; vpshufb %xmm6,%xmm5,%xmm5 DB 196,226,125,0,228 ; vpshufb %ymm4,%ymm0,%ymm4 DB 196,227,253,0,228,232 ; vpermq $0xe8,%ymm4,%ymm4 @@ -40058,10 +40058,10 @@ _sk_load_g8_hsw_8bit LABEL PROC DB 196,226,125,49,200 ; vpmovzxbd %xmm0,%ymm1 DB 197,249,112,192,78 ; vpshufd $0x4e,%xmm0,%xmm0 DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0 - DB 196,226,125,88,37,69,26,0,0 ; vpbroadcastd 0x1a45(%rip),%ymm4 # 2a50 <_sk_darken_hsw_8bit+0x18d> + DB 196,226,125,88,37,193,30,0,0 ; vpbroadcastd 0x1ec1(%rip),%ymm4 # 2ecc <_sk_exclusion_hsw_8bit+0x16c> DB 196,226,125,64,236 ; vpmulld %ymm4,%ymm0,%ymm5 DB 196,226,117,64,196 ; vpmulld %ymm4,%ymm1,%ymm0 - DB 196,226,125,88,13,54,26,0,0 ; vpbroadcastd 0x1a36(%rip),%ymm1 # 2a54 <_sk_darken_hsw_8bit+0x191> + DB 196,226,125,88,13,178,30,0,0 ; vpbroadcastd 0x1eb2(%rip),%ymm1 # 2ed0 <_sk_exclusion_hsw_8bit+0x170> DB 197,253,235,193 ; vpor %ymm1,%ymm0,%ymm0 DB 197,213,235,201 ; vpor %ymm1,%ymm5,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -40157,10 +40157,10 @@ _sk_load_g8_dst_hsw_8bit LABEL PROC DB 196,226,125,49,218 ; vpmovzxbd %xmm2,%ymm3 DB 197,249,112,210,78 ; vpshufd $0x4e,%xmm2,%xmm2 DB 196,226,125,49,210 ; vpmovzxbd %xmm2,%ymm2 - DB 196,226,125,88,37,213,24,0,0 ; vpbroadcastd 0x18d5(%rip),%ymm4 # 2a58 <_sk_darken_hsw_8bit+0x195> + DB 196,226,125,88,37,81,29,0,0 ; vpbroadcastd 0x1d51(%rip),%ymm4 # 2ed4 <_sk_exclusion_hsw_8bit+0x174> DB 196,226,109,64,236 ; vpmulld %ymm4,%ymm2,%ymm5 DB 196,226,101,64,212 ; vpmulld %ymm4,%ymm3,%ymm2 - DB 196,226,125,88,29,198,24,0,0 ; vpbroadcastd 0x18c6(%rip),%ymm3 # 2a5c <_sk_darken_hsw_8bit+0x199> + DB 196,226,125,88,29,66,29,0,0 ; vpbroadcastd 0x1d42(%rip),%ymm3 # 2ed8 <_sk_exclusion_hsw_8bit+0x178> DB 197,237,235,211 ; vpor %ymm3,%ymm2,%ymm2 DB 197,213,235,219 ; vpor %ymm3,%ymm5,%ymm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -40255,7 +40255,7 @@ _sk_srcover_rgba_8888_hsw_8bit LABEL PROC DB 15,133,222,0,0,0 ; jne 13c3 <_sk_srcover_rgba_8888_hsw_8bit+0x103> DB 196,33,126,111,76,138,32 ; vmovdqu 0x20(%rdx,%r9,4),%ymm9 DB 196,33,126,111,28,138 ; vmovdqu (%rdx,%r9,4),%ymm11 - DB 197,253,111,53,134,24,0,0 ; vmovdqa 0x1886(%rip),%ymm6 # 2b80 <_sk_darken_hsw_8bit+0x2bd> + DB 197,253,111,53,6,29,0,0 ; vmovdqa 0x1d06(%rip),%ymm6 # 3000 <_sk_exclusion_hsw_8bit+0x2a0> DB 196,226,117,0,254 ; vpshufb %ymm6,%ymm1,%ymm7 DB 196,226,125,0,246 ; vpshufb %ymm6,%ymm0,%ymm6 DB 196,66,125,48,195 ; vpmovzxbw %xmm11,%ymm8 @@ -40461,7 +40461,7 @@ PUBLIC _sk_scale_1_float_hsw_8bit _sk_scale_1_float_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,250,16,32 ; vmovss (%rax),%xmm4 - DB 197,218,89,37,10,20,0,0 ; vmulss 0x140a(%rip),%xmm4,%xmm4 # 2a60 <_sk_darken_hsw_8bit+0x19d> + DB 197,218,89,37,134,24,0,0 ; vmulss 0x1886(%rip),%xmm4,%xmm4 # 2edc <_sk_exclusion_hsw_8bit+0x17c> DB 197,250,44,196 ; vcvttss2si %xmm4,%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 DB 196,226,125,120,228 ; vpbroadcastb %xmm4,%ymm4 @@ -40471,7 +40471,7 @@ _sk_scale_1_float_hsw_8bit LABEL PROC DB 196,226,125,48,241 ; vpmovzxbw %xmm1,%ymm6 DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,221,219,37,21,21,0,0 ; vpand 0x1515(%rip),%ymm4,%ymm4 # 2ba0 <_sk_darken_hsw_8bit+0x2dd> + DB 197,221,219,37,149,25,0,0 ; vpand 0x1995(%rip),%ymm4,%ymm4 # 3020 <_sk_exclusion_hsw_8bit+0x2c0> DB 197,221,213,249 ; vpmullw %ymm1,%ymm4,%ymm7 DB 197,93,213,198 ; vpmullw %ymm6,%ymm4,%ymm8 DB 197,93,213,200 ; vpmullw %ymm0,%ymm4,%ymm9 @@ -40508,7 +40508,7 @@ _sk_scale_u8_hsw_8bit LABEL PROC DB 196,226,125,49,236 ; vpmovzxbd %xmm4,%ymm5 DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,125,49,228 ; vpmovzxbd %xmm4,%ymm4 - DB 197,253,111,53,159,20,0,0 ; vmovdqa 0x149f(%rip),%ymm6 # 2bc0 <_sk_darken_hsw_8bit+0x2fd> + DB 197,253,111,53,31,25,0,0 ; vmovdqa 0x191f(%rip),%ymm6 # 3040 <_sk_exclusion_hsw_8bit+0x2e0> DB 196,226,93,0,230 ; vpshufb %ymm6,%ymm4,%ymm4 DB 196,226,85,0,238 ; vpshufb %ymm6,%ymm5,%ymm5 DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 @@ -40623,7 +40623,7 @@ PUBLIC _sk_lerp_1_float_hsw_8bit _sk_lerp_1_float_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 197,250,16,32 ; vmovss (%rax),%xmm4 - DB 197,218,89,37,110,17,0,0 ; vmulss 0x116e(%rip),%xmm4,%xmm4 # 2a64 <_sk_darken_hsw_8bit+0x1a1> + DB 197,218,89,37,234,21,0,0 ; vmulss 0x15ea(%rip),%xmm4,%xmm4 # 2ee0 <_sk_exclusion_hsw_8bit+0x180> DB 197,250,44,196 ; vcvttss2si %xmm4,%eax DB 197,249,110,224 ; vmovd %eax,%xmm4 DB 196,226,125,120,228 ; vpbroadcastb %xmm4,%ymm4 @@ -40633,7 +40633,7 @@ _sk_lerp_1_float_hsw_8bit LABEL PROC DB 196,226,125,48,241 ; vpmovzxbw %xmm1,%ymm6 DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 - DB 197,221,219,61,181,18,0,0 ; vpand 0x12b5(%rip),%ymm4,%ymm7 # 2be0 <_sk_darken_hsw_8bit+0x31d> + DB 197,221,219,61,53,23,0,0 ; vpand 0x1735(%rip),%ymm4,%ymm7 # 3060 <_sk_exclusion_hsw_8bit+0x300> DB 197,69,213,193 ; vpmullw %ymm1,%ymm7,%ymm8 DB 197,69,213,206 ; vpmullw %ymm6,%ymm7,%ymm9 DB 197,69,213,208 ; vpmullw %ymm0,%ymm7,%ymm10 @@ -40701,7 +40701,7 @@ _sk_lerp_u8_hsw_8bit LABEL PROC DB 196,226,125,49,236 ; vpmovzxbd %xmm4,%ymm5 DB 197,249,112,228,78 ; vpshufd $0x4e,%xmm4,%xmm4 DB 196,226,125,49,228 ; vpmovzxbd %xmm4,%ymm4 - DB 197,253,111,53,170,17,0,0 ; vmovdqa 0x11aa(%rip),%ymm6 # 2c00 <_sk_darken_hsw_8bit+0x33d> + DB 197,253,111,53,42,22,0,0 ; vmovdqa 0x162a(%rip),%ymm6 # 3080 <_sk_exclusion_hsw_8bit+0x320> DB 196,98,93,0,206 ; vpshufb %ymm6,%ymm4,%ymm9 DB 196,98,85,0,222 ; vpshufb %ymm6,%ymm5,%ymm11 DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 @@ -40864,7 +40864,7 @@ _sk_move_dst_src_hsw_8bit LABEL PROC PUBLIC _sk_black_color_hsw_8bit _sk_black_color_hsw_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 196,226,125,24,5,125,13,0,0 ; vbroadcastss 0xd7d(%rip),%ymm0 # 2a68 <_sk_darken_hsw_8bit+0x1a5> + DB 196,226,125,24,5,249,17,0,0 ; vbroadcastss 0x11f9(%rip),%ymm0 # 2ee4 <_sk_exclusion_hsw_8bit+0x184> DB 197,252,40,200 ; vmovaps %ymm0,%ymm1 DB 255,224 ; jmpq *%rax @@ -40884,7 +40884,7 @@ _sk_clear_hsw_8bit LABEL PROC PUBLIC _sk_srcatop_hsw_8bit _sk_srcatop_hsw_8bit LABEL PROC - DB 197,125,111,5,15,15,0,0 ; vmovdqa 0xf0f(%rip),%ymm8 # 2c20 <_sk_darken_hsw_8bit+0x35d> + DB 197,125,111,5,143,19,0,0 ; vmovdqa 0x138f(%rip),%ymm8 # 30a0 <_sk_exclusion_hsw_8bit+0x340> DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 DB 196,194,109,0,232 ; vpshufb %ymm8,%ymm2,%ymm5 DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 @@ -40959,7 +40959,7 @@ _sk_srcatop_hsw_8bit LABEL PROC PUBLIC _sk_dstatop_hsw_8bit _sk_dstatop_hsw_8bit LABEL PROC - DB 197,125,111,5,202,13,0,0 ; vmovdqa 0xdca(%rip),%ymm8 # 2c40 <_sk_darken_hsw_8bit+0x37d> + DB 197,125,111,5,74,18,0,0 ; vmovdqa 0x124a(%rip),%ymm8 # 30c0 <_sk_exclusion_hsw_8bit+0x360> DB 196,194,117,0,224 ; vpshufb %ymm8,%ymm1,%ymm4 DB 196,194,125,0,232 ; vpshufb %ymm8,%ymm0,%ymm5 DB 196,98,125,48,210 ; vpmovzxbw %xmm2,%ymm10 @@ -41034,7 +41034,7 @@ _sk_dstatop_hsw_8bit LABEL PROC PUBLIC _sk_srcin_hsw_8bit _sk_srcin_hsw_8bit LABEL PROC - DB 197,253,111,37,129,12,0,0 ; vmovdqa 0xc81(%rip),%ymm4 # 2c60 <_sk_darken_hsw_8bit+0x39d> + DB 197,253,111,37,1,17,0,0 ; vmovdqa 0x1101(%rip),%ymm4 # 30e0 <_sk_exclusion_hsw_8bit+0x380> DB 196,226,101,0,236 ; vpshufb %ymm4,%ymm3,%ymm5 DB 196,226,109,0,228 ; vpshufb %ymm4,%ymm2,%ymm4 DB 196,226,125,48,240 ; vpmovzxbw %xmm0,%ymm6 @@ -41072,7 +41072,7 @@ _sk_srcin_hsw_8bit LABEL PROC PUBLIC _sk_dstin_hsw_8bit _sk_dstin_hsw_8bit LABEL PROC - DB 197,253,111,37,247,11,0,0 ; vmovdqa 0xbf7(%rip),%ymm4 # 2c80 <_sk_darken_hsw_8bit+0x3bd> + DB 197,253,111,37,119,16,0,0 ; vmovdqa 0x1077(%rip),%ymm4 # 3100 <_sk_exclusion_hsw_8bit+0x3a0> DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,125,48,226 ; vpmovzxbw %xmm2,%ymm4 @@ -41110,7 +41110,7 @@ _sk_dstin_hsw_8bit LABEL PROC PUBLIC _sk_srcout_hsw_8bit _sk_srcout_hsw_8bit LABEL PROC - DB 197,253,111,37,107,11,0,0 ; vmovdqa 0xb6b(%rip),%ymm4 # 2ca0 <_sk_darken_hsw_8bit+0x3dd> + DB 197,253,111,37,235,15,0,0 ; vmovdqa 0xfeb(%rip),%ymm4 # 3120 <_sk_exclusion_hsw_8bit+0x3c0> DB 196,226,109,0,236 ; vpshufb %ymm4,%ymm2,%ymm5 DB 196,226,101,0,228 ; vpshufb %ymm4,%ymm3,%ymm4 DB 197,205,118,246 ; vpcmpeqd %ymm6,%ymm6,%ymm6 @@ -41151,7 +41151,7 @@ _sk_srcout_hsw_8bit LABEL PROC PUBLIC _sk_dstout_hsw_8bit _sk_dstout_hsw_8bit LABEL PROC - DB 197,253,111,37,213,10,0,0 ; vmovdqa 0xad5(%rip),%ymm4 # 2cc0 <_sk_darken_hsw_8bit+0x3fd> + DB 197,253,111,37,85,15,0,0 ; vmovdqa 0xf55(%rip),%ymm4 # 3140 <_sk_exclusion_hsw_8bit+0x3e0> DB 196,226,125,0,196 ; vpshufb %ymm4,%ymm0,%ymm0 DB 196,226,117,0,204 ; vpshufb %ymm4,%ymm1,%ymm1 DB 197,221,118,228 ; vpcmpeqd %ymm4,%ymm4,%ymm4 @@ -41192,7 +41192,7 @@ _sk_dstout_hsw_8bit LABEL PROC PUBLIC _sk_srcover_hsw_8bit _sk_srcover_hsw_8bit LABEL PROC - DB 197,253,111,37,61,10,0,0 ; vmovdqa 0xa3d(%rip),%ymm4 # 2ce0 <_sk_darken_hsw_8bit+0x41d> + DB 197,253,111,37,189,14,0,0 ; vmovdqa 0xebd(%rip),%ymm4 # 3160 <_sk_exclusion_hsw_8bit+0x400> DB 196,226,117,0,236 ; vpshufb %ymm4,%ymm1,%ymm5 DB 196,226,125,0,228 ; vpshufb %ymm4,%ymm0,%ymm4 DB 196,98,125,48,202 ; vpmovzxbw %xmm2,%ymm9 @@ -41234,7 +41234,7 @@ _sk_srcover_hsw_8bit LABEL PROC PUBLIC _sk_dstover_hsw_8bit _sk_dstover_hsw_8bit LABEL PROC - DB 197,253,111,37,158,9,0,0 ; vmovdqa 0x99e(%rip),%ymm4 # 2d00 <_sk_darken_hsw_8bit+0x43d> + DB 197,253,111,37,30,14,0,0 ; vmovdqa 0xe1e(%rip),%ymm4 # 3180 <_sk_exclusion_hsw_8bit+0x420> DB 196,226,101,0,236 ; vpshufb %ymm4,%ymm3,%ymm5 DB 196,226,109,0,228 ; vpshufb %ymm4,%ymm2,%ymm4 DB 196,98,125,48,200 ; vpmovzxbw %xmm0,%ymm9 @@ -41314,7 +41314,7 @@ _sk_multiply_hsw_8bit LABEL PROC DB 72,131,236,56 ; sub $0x38,%rsp DB 197,253,111,243 ; vmovdqa %ymm3,%ymm6 DB 197,253,111,218 ; vmovdqa %ymm2,%ymm3 - DB 197,125,111,13,91,8,0,0 ; vmovdqa 0x85b(%rip),%ymm9 # 2d20 <_sk_darken_hsw_8bit+0x45d> + DB 197,125,111,13,219,12,0,0 ; vmovdqa 0xcdb(%rip),%ymm9 # 31a0 <_sk_exclusion_hsw_8bit+0x440> DB 196,194,101,0,225 ; vpshufb %ymm9,%ymm3,%ymm4 DB 196,194,77,0,233 ; vpshufb %ymm9,%ymm6,%ymm5 DB 196,65,45,118,210 ; vpcmpeqd %ymm10,%ymm10,%ymm10 @@ -41455,7 +41455,7 @@ _sk_screen_hsw_8bit LABEL PROC PUBLIC _sk_xor__hsw_8bit _sk_xor__hsw_8bit LABEL PROC - DB 197,125,111,13,231,5,0,0 ; vmovdqa 0x5e7(%rip),%ymm9 # 2d40 <_sk_darken_hsw_8bit+0x47d> + DB 197,125,111,13,103,10,0,0 ; vmovdqa 0xa67(%rip),%ymm9 # 31c0 <_sk_exclusion_hsw_8bit+0x460> DB 196,194,109,0,225 ; vpshufb %ymm9,%ymm2,%ymm4 DB 196,194,101,0,249 ; vpshufb %ymm9,%ymm3,%ymm7 DB 196,65,37,118,219 ; vpcmpeqd %ymm11,%ymm11,%ymm11 @@ -41532,7 +41532,7 @@ _sk_xor__hsw_8bit LABEL PROC PUBLIC _sk_darken_hsw_8bit _sk_darken_hsw_8bit LABEL PROC - DB 197,125,111,5,149,4,0,0 ; vmovdqa 0x495(%rip),%ymm8 # 2d60 <_sk_darken_hsw_8bit+0x49d> + DB 197,125,111,5,21,9,0,0 ; vmovdqa 0x915(%rip),%ymm8 # 31e0 <_sk_exclusion_hsw_8bit+0x480> DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 @@ -41605,15 +41605,252 @@ _sk_darken_hsw_8bit LABEL PROC DB 197,253,248,246 ; vpsubb %ymm6,%ymm0,%ymm6 DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 - DB 196,226,125,88,37,50,0,0,0 ; vpbroadcastd 0x32(%rip),%ymm4 # 2a6c <_sk_darken_hsw_8bit+0x1a9> + DB 196,226,125,88,37,174,4,0,0 ; vpbroadcastd 0x4ae(%rip),%ymm4 # 2ee8 <_sk_exclusion_hsw_8bit+0x188> DB 196,227,125,76,198,64 ; vpblendvb %ymm4,%ymm6,%ymm0,%ymm0 DB 196,227,117,76,207,64 ; vpblendvb %ymm4,%ymm7,%ymm1,%ymm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_lighten_hsw_8bit +_sk_lighten_hsw_8bit LABEL PROC + DB 197,125,111,5,174,7,0,0 ; vmovdqa 0x7ae(%rip),%ymm8 # 3200 <_sk_exclusion_hsw_8bit+0x4a0> + DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 + DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 + DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 + DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,98,125,48,222 ; vpmovzxbw %xmm6,%ymm11 + DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 + DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 + DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 + DB 196,65,29,213,225 ; vpmullw %ymm9,%ymm12,%ymm12 + DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 + DB 196,65,37,213,218 ; vpmullw %ymm10,%ymm11,%ymm11 + DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 + DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 + DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 + DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 + DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 + DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 + DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,53,103,204 ; vpackuswb %ymm4,%ymm9,%ymm9 + DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,85,103,222 ; vpackuswb %ymm6,%ymm5,%ymm11 + DB 196,194,117,0,232 ; vpshufb %ymm8,%ymm1,%ymm5 + DB 196,194,125,0,248 ; vpshufb %ymm8,%ymm0,%ymm7 + DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 + DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 + DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,98,125,48,237 ; vpmovzxbw %xmm5,%ymm13 + DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 + DB 196,65,45,213,237 ; vpmullw %ymm13,%ymm10,%ymm13 + DB 197,221,213,255 ; vpmullw %ymm7,%ymm4,%ymm7 + DB 196,65,61,213,228 ; vpmullw %ymm12,%ymm8,%ymm12 + DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 + DB 197,197,253,228 ; vpaddw %ymm4,%ymm7,%ymm4 + DB 196,193,21,253,250 ; vpaddw %ymm10,%ymm13,%ymm7 + DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 + DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 + DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 + DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 + DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 + DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 + DB 197,181,218,244 ; vpminub %ymm4,%ymm9,%ymm6 + DB 197,165,218,253 ; vpminub %ymm5,%ymm11,%ymm7 + DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 + DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 + DB 197,245,248,255 ; vpsubb %ymm7,%ymm1,%ymm7 + DB 197,253,248,246 ; vpsubb %ymm6,%ymm0,%ymm6 + DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 + DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 + DB 196,226,125,88,37,43,3,0,0 ; vpbroadcastd 0x32b(%rip),%ymm4 # 2eec <_sk_exclusion_hsw_8bit+0x18c> + DB 196,227,125,76,198,64 ; vpblendvb %ymm4,%ymm6,%ymm0,%ymm0 + DB 196,227,117,76,207,64 ; vpblendvb %ymm4,%ymm7,%ymm1,%ymm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_difference_hsw_8bit +_sk_difference_hsw_8bit LABEL PROC + DB 197,125,111,5,71,6,0,0 ; vmovdqa 0x647(%rip),%ymm8 # 3220 <_sk_exclusion_hsw_8bit+0x4c0> + DB 196,194,101,0,224 ; vpshufb %ymm8,%ymm3,%ymm4 + DB 196,194,109,0,240 ; vpshufb %ymm8,%ymm2,%ymm6 + DB 196,98,125,48,208 ; vpmovzxbw %xmm0,%ymm10 + DB 196,227,125,57,197,1 ; vextracti128 $0x1,%ymm0,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,98,125,48,201 ; vpmovzxbw %xmm1,%ymm9 + DB 196,227,125,57,207,1 ; vextracti128 $0x1,%ymm1,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,98,125,48,222 ; vpmovzxbw %xmm6,%ymm11 + DB 196,227,125,57,246,1 ; vextracti128 $0x1,%ymm6,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,98,125,48,228 ; vpmovzxbw %xmm4,%ymm12 + DB 196,227,125,57,228,1 ; vextracti128 $0x1,%ymm4,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 + DB 197,221,213,231 ; vpmullw %ymm7,%ymm4,%ymm4 + DB 196,65,29,213,225 ; vpmullw %ymm9,%ymm12,%ymm12 + DB 197,205,213,245 ; vpmullw %ymm5,%ymm6,%ymm6 + DB 196,65,37,213,218 ; vpmullw %ymm10,%ymm11,%ymm11 + DB 196,65,37,253,210 ; vpaddw %ymm10,%ymm11,%ymm10 + DB 197,205,253,237 ; vpaddw %ymm5,%ymm6,%ymm5 + DB 196,193,29,253,241 ; vpaddw %ymm9,%ymm12,%ymm6 + DB 197,221,253,231 ; vpaddw %ymm7,%ymm4,%ymm4 + DB 197,197,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm7 + DB 197,205,113,214,8 ; vpsrlw $0x8,%ymm6,%ymm6 + DB 197,221,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm4 + DB 196,193,85,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm5 + DB 196,99,85,56,204,1 ; vinserti128 $0x1,%xmm4,%ymm5,%ymm9 + DB 196,227,85,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm5,%ymm4 + DB 197,53,103,204 ; vpackuswb %ymm4,%ymm9,%ymm9 + DB 196,227,77,56,239,1 ; vinserti128 $0x1,%xmm7,%ymm6,%ymm5 + DB 196,227,77,70,247,49 ; vperm2i128 $0x31,%ymm7,%ymm6,%ymm6 + DB 197,85,103,222 ; vpackuswb %ymm6,%ymm5,%ymm11 + DB 196,194,117,0,232 ; vpshufb %ymm8,%ymm1,%ymm5 + DB 196,194,125,0,248 ; vpshufb %ymm8,%ymm0,%ymm7 + DB 196,98,125,48,194 ; vpmovzxbw %xmm2,%ymm8 + DB 196,227,125,57,212,1 ; vextracti128 $0x1,%ymm2,%xmm4 + DB 196,226,125,48,228 ; vpmovzxbw %xmm4,%ymm4 + DB 196,98,125,48,211 ; vpmovzxbw %xmm3,%ymm10 + DB 196,227,125,57,222,1 ; vextracti128 $0x1,%ymm3,%xmm6 + DB 196,226,125,48,246 ; vpmovzxbw %xmm6,%ymm6 + DB 196,98,125,48,231 ; vpmovzxbw %xmm7,%ymm12 + DB 196,227,125,57,255,1 ; vextracti128 $0x1,%ymm7,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,98,125,48,237 ; vpmovzxbw %xmm5,%ymm13 + DB 196,227,125,57,237,1 ; vextracti128 $0x1,%ymm5,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 197,205,213,237 ; vpmullw %ymm5,%ymm6,%ymm5 + DB 196,65,45,213,237 ; vpmullw %ymm13,%ymm10,%ymm13 + DB 197,221,213,255 ; vpmullw %ymm7,%ymm4,%ymm7 + DB 196,65,61,213,228 ; vpmullw %ymm12,%ymm8,%ymm12 + DB 196,65,29,253,192 ; vpaddw %ymm8,%ymm12,%ymm8 + DB 197,197,253,228 ; vpaddw %ymm4,%ymm7,%ymm4 + DB 196,193,21,253,250 ; vpaddw %ymm10,%ymm13,%ymm7 + DB 197,213,253,238 ; vpaddw %ymm6,%ymm5,%ymm5 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 197,205,113,215,8 ; vpsrlw $0x8,%ymm7,%ymm6 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 + DB 196,193,69,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm7 + DB 196,99,69,56,196,1 ; vinserti128 $0x1,%xmm4,%ymm7,%ymm8 + DB 196,227,69,70,228,49 ; vperm2i128 $0x31,%ymm4,%ymm7,%ymm4 + DB 197,189,103,228 ; vpackuswb %ymm4,%ymm8,%ymm4 + DB 196,227,77,56,253,1 ; vinserti128 $0x1,%xmm5,%ymm6,%ymm7 + DB 196,227,77,70,237,49 ; vperm2i128 $0x31,%ymm5,%ymm6,%ymm5 + DB 197,197,103,237 ; vpackuswb %ymm5,%ymm7,%ymm5 + DB 197,165,218,245 ; vpminub %ymm5,%ymm11,%ymm6 + DB 197,181,218,252 ; vpminub %ymm4,%ymm9,%ymm7 + DB 197,197,252,255 ; vpaddb %ymm7,%ymm7,%ymm7 + DB 197,205,252,246 ; vpaddb %ymm6,%ymm6,%ymm6 + DB 197,237,252,192 ; vpaddb %ymm0,%ymm2,%ymm0 + DB 197,229,252,201 ; vpaddb %ymm1,%ymm3,%ymm1 + DB 197,245,248,246 ; vpsubb %ymm6,%ymm1,%ymm6 + DB 197,253,248,255 ; vpsubb %ymm7,%ymm0,%ymm7 + DB 197,245,248,205 ; vpsubb %ymm5,%ymm1,%ymm1 + DB 197,253,248,196 ; vpsubb %ymm4,%ymm0,%ymm0 + DB 196,226,125,88,37,160,1,0,0 ; vpbroadcastd 0x1a0(%rip),%ymm4 # 2ef0 <_sk_exclusion_hsw_8bit+0x190> + DB 196,227,125,76,199,64 ; vpblendvb %ymm4,%ymm7,%ymm0,%ymm0 + DB 196,227,117,76,206,64 ; vpblendvb %ymm4,%ymm6,%ymm1,%ymm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_hsw_8bit +_sk_exclusion_hsw_8bit LABEL PROC + DB 196,98,125,48,192 ; vpmovzxbw %xmm0,%ymm8 + DB 196,227,125,57,196,1 ; vextracti128 $0x1,%ymm0,%xmm4 + DB 196,98,125,48,204 ; vpmovzxbw %xmm4,%ymm9 + DB 196,98,125,48,209 ; vpmovzxbw %xmm1,%ymm10 + DB 196,227,125,57,204,1 ; vextracti128 $0x1,%ymm1,%xmm4 + DB 196,98,125,48,220 ; vpmovzxbw %xmm4,%ymm11 + DB 196,226,125,48,226 ; vpmovzxbw %xmm2,%ymm4 + DB 196,227,125,57,213,1 ; vextracti128 $0x1,%ymm2,%xmm5 + DB 196,226,125,48,237 ; vpmovzxbw %xmm5,%ymm5 + DB 196,226,125,48,243 ; vpmovzxbw %xmm3,%ymm6 + DB 196,227,125,57,223,1 ; vextracti128 $0x1,%ymm3,%xmm7 + DB 196,226,125,48,255 ; vpmovzxbw %xmm7,%ymm7 + DB 196,65,69,213,227 ; vpmullw %ymm11,%ymm7,%ymm12 + DB 196,65,77,213,234 ; vpmullw %ymm10,%ymm6,%ymm13 + DB 196,65,85,213,241 ; vpmullw %ymm9,%ymm5,%ymm14 + DB 196,65,93,213,248 ; vpmullw %ymm8,%ymm4,%ymm15 + DB 196,65,5,253,192 ; vpaddw %ymm8,%ymm15,%ymm8 + DB 196,65,13,253,201 ; vpaddw %ymm9,%ymm14,%ymm9 + DB 196,65,21,253,210 ; vpaddw %ymm10,%ymm13,%ymm10 + DB 196,65,29,253,219 ; vpaddw %ymm11,%ymm12,%ymm11 + DB 196,193,37,113,211,8 ; vpsrlw $0x8,%ymm11,%ymm11 + DB 196,193,45,113,210,8 ; vpsrlw $0x8,%ymm10,%ymm10 + DB 196,193,53,113,209,8 ; vpsrlw $0x8,%ymm9,%ymm9 + DB 196,193,61,113,208,8 ; vpsrlw $0x8,%ymm8,%ymm8 + DB 196,67,61,56,225,1 ; vinserti128 $0x1,%xmm9,%ymm8,%ymm12 + DB 196,67,61,70,193,49 ; vperm2i128 $0x31,%ymm9,%ymm8,%ymm8 + DB 196,65,29,103,192 ; vpackuswb %ymm8,%ymm12,%ymm8 + DB 196,67,45,56,203,1 ; vinserti128 $0x1,%xmm11,%ymm10,%ymm9 + DB 196,67,45,70,211,49 ; vperm2i128 $0x31,%ymm11,%ymm10,%ymm10 + DB 196,65,53,103,202 ; vpackuswb %ymm10,%ymm9,%ymm9 + DB 196,65,61,252,208 ; vpaddb %ymm8,%ymm8,%ymm10 + DB 196,65,53,252,193 ; vpaddb %ymm9,%ymm9,%ymm8 + DB 197,109,252,200 ; vpaddb %ymm0,%ymm2,%ymm9 + DB 197,101,252,217 ; vpaddb %ymm1,%ymm3,%ymm11 + DB 196,65,37,248,192 ; vpsubb %ymm8,%ymm11,%ymm8 + DB 196,65,53,248,210 ; vpsubb %ymm10,%ymm9,%ymm10 + DB 197,125,111,37,26,4,0,0 ; vmovdqa 0x41a(%rip),%ymm12 # 3240 <_sk_exclusion_hsw_8bit+0x4e0> + DB 196,194,117,0,204 ; vpshufb %ymm12,%ymm1,%ymm1 + DB 196,194,125,0,196 ; vpshufb %ymm12,%ymm0,%ymm0 + DB 196,98,125,48,224 ; vpmovzxbw %xmm0,%ymm12 + DB 196,227,125,57,192,1 ; vextracti128 $0x1,%ymm0,%xmm0 + DB 196,226,125,48,192 ; vpmovzxbw %xmm0,%ymm0 + DB 196,98,125,48,233 ; vpmovzxbw %xmm1,%ymm13 + DB 196,227,125,57,201,1 ; vextracti128 $0x1,%ymm1,%xmm1 + DB 196,226,125,48,201 ; vpmovzxbw %xmm1,%ymm1 + DB 197,197,213,201 ; vpmullw %ymm1,%ymm7,%ymm1 + DB 196,65,77,213,237 ; vpmullw %ymm13,%ymm6,%ymm13 + DB 197,213,213,192 ; vpmullw %ymm0,%ymm5,%ymm0 + DB 196,65,93,213,228 ; vpmullw %ymm12,%ymm4,%ymm12 + DB 197,157,253,228 ; vpaddw %ymm4,%ymm12,%ymm4 + DB 197,253,253,197 ; vpaddw %ymm5,%ymm0,%ymm0 + DB 197,149,253,238 ; vpaddw %ymm6,%ymm13,%ymm5 + DB 197,245,253,207 ; vpaddw %ymm7,%ymm1,%ymm1 + DB 197,245,113,209,8 ; vpsrlw $0x8,%ymm1,%ymm1 + DB 197,213,113,213,8 ; vpsrlw $0x8,%ymm5,%ymm5 + DB 197,253,113,208,8 ; vpsrlw $0x8,%ymm0,%ymm0 + DB 197,221,113,212,8 ; vpsrlw $0x8,%ymm4,%ymm4 + DB 196,227,93,56,240,1 ; vinserti128 $0x1,%xmm0,%ymm4,%ymm6 + DB 196,227,93,70,192,49 ; vperm2i128 $0x31,%ymm0,%ymm4,%ymm0 + DB 197,205,103,192 ; vpackuswb %ymm0,%ymm6,%ymm0 + DB 196,227,85,56,225,1 ; vinserti128 $0x1,%xmm1,%ymm5,%ymm4 + DB 196,227,85,70,201,49 ; vperm2i128 $0x31,%ymm1,%ymm5,%ymm1 + DB 197,221,103,201 ; vpackuswb %ymm1,%ymm4,%ymm1 + DB 197,165,248,201 ; vpsubb %ymm1,%ymm11,%ymm1 + DB 197,181,248,192 ; vpsubb %ymm0,%ymm9,%ymm0 + DB 196,226,125,88,37,61,0,0,0 ; vpbroadcastd 0x3d(%rip),%ymm4 # 2ef4 <_sk_exclusion_hsw_8bit+0x194> + DB 196,195,125,76,194,64 ; vpblendvb %ymm4,%ymm10,%ymm0,%ymm0 + DB 196,195,117,76,200,64 ; vpblendvb %ymm4,%ymm8,%ymm1,%ymm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + ALIGN 4 DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2a93 <_sk_darken_hsw_8bit+0x1d0> + DB 127,67 ; jg 2f0f <_sk_exclusion_hsw_8bit+0x1af> DB 1,1 ; add %eax,(%rcx) DB 1,0 ; add %eax,(%rax) DB 0,0 ; add %al,(%rax) @@ -41623,14 +41860,23 @@ ALIGN 4 DB 0,0 ; add %al,(%rax) DB 0,255 ; add %bh,%bh DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2aa7 <_sk_darken_hsw_8bit+0x1e4> + DB 127,67 ; jg 2f23 <_sk_exclusion_hsw_8bit+0x1c3> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 2aab <_sk_darken_hsw_8bit+0x1e8> + DB 127,67 ; jg 2f27 <_sk_exclusion_hsw_8bit+0x1c7> DB 0,0 ; add %al,(%rax) DB 0,255 ; add %bh,%bh DB 255 ; (bad) DB 255 ; (bad) DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) ALIGN 32 DB 0,0 ; add %al,(%rax) @@ -42085,6 +42331,72 @@ ALIGN 32 DB 11,11 ; or (%rbx),%ecx DB 11,11 ; or (%rbx),%ecx DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,3,3 ; lsl (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,3,3 ; lsl (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,3,3 ; lsl (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,3,3 ; lsl (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,3,3 ; lsl (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,3,3 ; lsl (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) DB 15 ; .byte 0xf DB 15 ; .byte 0xf DB 15 ; .byte 0xf @@ -42198,7 +42510,7 @@ _sk_uniform_color_sse41_8bit LABEL PROC PUBLIC _sk_set_rgb_sse41_8bit _sk_set_rgb_sse41_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,15,16,37,208,29,0,0 ; movss 0x1dd0(%rip),%xmm4 # 1f50 <_sk_darken_sse41_8bit+0x170> + DB 243,15,16,37,236,33,0,0 ; movss 0x21ec(%rip),%xmm4 # 236c <_sk_exclusion_sse41_8bit+0x137> DB 243,15,16,40 ; movss (%rax),%xmm5 DB 243,15,89,236 ; mulss %xmm4,%xmm5 DB 243,72,15,44,205 ; cvttss2si %xmm5,%rcx @@ -42213,7 +42525,7 @@ _sk_set_rgb_sse41_8bit LABEL PROC DB 9,208 ; or %edx,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 102,15,111,45,160,29,0,0 ; movdqa 0x1da0(%rip),%xmm5 # 1f60 <_sk_darken_sse41_8bit+0x180> + DB 102,15,111,45,192,33,0,0 ; movdqa 0x21c0(%rip),%xmm5 # 2380 <_sk_exclusion_sse41_8bit+0x14b> DB 102,15,219,205 ; pand %xmm5,%xmm1 DB 102,15,219,197 ; pand %xmm5,%xmm0 DB 102,15,235,196 ; por %xmm4,%xmm0 @@ -42225,12 +42537,12 @@ PUBLIC _sk_premul_sse41_8bit _sk_premul_sse41_8bit LABEL PROC DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,140,29,0,0 ; movdqa 0x1d8c(%rip),%xmm0 # 1f70 <_sk_darken_sse41_8bit+0x190> + DB 102,15,111,5,172,33,0,0 ; movdqa 0x21ac(%rip),%xmm0 # 2390 <_sk_exclusion_sse41_8bit+0x15b> DB 102,15,111,253 ; movdqa %xmm5,%xmm7 DB 102,15,56,0,248 ; pshufb %xmm0,%xmm7 DB 102,15,111,244 ; movdqa %xmm4,%xmm6 DB 102,15,56,0,240 ; pshufb %xmm0,%xmm6 - DB 102,15,111,5,130,29,0,0 ; movdqa 0x1d82(%rip),%xmm0 # 1f80 <_sk_darken_sse41_8bit+0x1a0> + DB 102,15,111,5,162,33,0,0 ; movdqa 0x21a2(%rip),%xmm0 # 23a0 <_sk_exclusion_sse41_8bit+0x16b> DB 102,15,235,240 ; por %xmm0,%xmm6 DB 102,15,235,248 ; por %xmm0,%xmm7 DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 @@ -42261,7 +42573,7 @@ _sk_premul_sse41_8bit LABEL PROC PUBLIC _sk_swap_rb_sse41_8bit _sk_swap_rb_sse41_8bit LABEL PROC - DB 102,15,111,37,15,29,0,0 ; movdqa 0x1d0f(%rip),%xmm4 # 1f90 <_sk_darken_sse41_8bit+0x1b0> + DB 102,15,111,37,47,33,0,0 ; movdqa 0x212f(%rip),%xmm4 # 23b0 <_sk_exclusion_sse41_8bit+0x17b> DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -42382,7 +42694,7 @@ _sk_load_8888_dst_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,221 ; jmpq ffffffffde000410 <_sk_darken_sse41_8bit+0xffffffffddffe630> + DB 233,255,255,255,221 ; jmpq ffffffffde000410 <_sk_exclusion_sse41_8bit+0xffffffffddffe1db> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; .byte 0xff @@ -42461,7 +42773,7 @@ _sk_load_bgra_sse41_8bit LABEL PROC DB 117,35 ; jne 504 <_sk_load_bgra_sse41_8bit+0x44> DB 243,66,15,111,76,130,16 ; movdqu 0x10(%rdx,%r8,4),%xmm1 DB 243,66,15,111,4,130 ; movdqu (%rdx,%r8,4),%xmm0 - DB 102,15,111,37,170,26,0,0 ; movdqa 0x1aaa(%rip),%xmm4 # 1fa0 <_sk_darken_sse41_8bit+0x1c0> + DB 102,15,111,37,202,30,0,0 ; movdqa 0x1eca(%rip),%xmm4 # 23c0 <_sk_exclusion_sse41_8bit+0x18b> DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -42522,7 +42834,7 @@ _sk_load_bgra_dst_sse41_8bit LABEL PROC DB 117,35 ; jne 5d4 <_sk_load_bgra_dst_sse41_8bit+0x44> DB 243,66,15,111,92,130,16 ; movdqu 0x10(%rdx,%r8,4),%xmm3 DB 243,66,15,111,20,130 ; movdqu (%rdx,%r8,4),%xmm2 - DB 102,15,111,37,234,25,0,0 ; movdqa 0x19ea(%rip),%xmm4 # 1fb0 <_sk_darken_sse41_8bit+0x1d0> + DB 102,15,111,37,10,30,0,0 ; movdqa 0x1e0a(%rip),%xmm4 # 23d0 <_sk_exclusion_sse41_8bit+0x19b> DB 102,15,56,0,212 ; pshufb %xmm4,%xmm2 DB 102,15,56,0,220 ; pshufb %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -42579,7 +42891,7 @@ _sk_store_bgra_sse41_8bit LABEL PROC DB 72,15,175,209 ; imul %rcx,%rdx DB 72,193,226,2 ; shl $0x2,%rdx DB 72,3,16 ; add (%rax),%rdx - DB 102,15,111,53,60,25,0,0 ; movdqa 0x193c(%rip),%xmm6 # 1fc0 <_sk_darken_sse41_8bit+0x1e0> + DB 102,15,111,53,92,29,0,0 ; movdqa 0x1d5c(%rip),%xmm6 # 23e0 <_sk_exclusion_sse41_8bit+0x1ab> DB 102,15,111,233 ; movdqa %xmm1,%xmm5 DB 102,15,56,0,238 ; pshufb %xmm6,%xmm5 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 @@ -42643,7 +42955,7 @@ _sk_load_a8_sse41_8bit LABEL PROC DB 77,133,201 ; test %r9,%r9 DB 117,42 ; jne 76b <_sk_load_a8_sse41_8bit+0x47> DB 102,66,15,56,48,12,2 ; pmovzxbw (%rdx,%r8,1),%xmm1 - DB 102,15,219,13,128,24,0,0 ; pand 0x1880(%rip),%xmm1 # 1fd0 <_sk_darken_sse41_8bit+0x1f0> + DB 102,15,219,13,160,28,0,0 ; pand 0x1ca0(%rip),%xmm1 # 23f0 <_sk_exclusion_sse41_8bit+0x1bb> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 @@ -42691,7 +43003,7 @@ _sk_load_a8_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf00080c <_sk_darken_sse41_8bit+0xffffffffdeffea2c> + DB 233,255,255,255,222 ; jmpq ffffffffdf00080c <_sk_exclusion_sse41_8bit+0xffffffffdeffe5d7> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -42714,7 +43026,7 @@ _sk_load_a8_dst_sse41_8bit LABEL PROC DB 77,133,201 ; test %r9,%r9 DB 117,42 ; jne 85f <_sk_load_a8_dst_sse41_8bit+0x47> DB 102,66,15,56,48,28,2 ; pmovzxbw (%rdx,%r8,1),%xmm3 - DB 102,15,219,29,156,23,0,0 ; pand 0x179c(%rip),%xmm3 # 1fe0 <_sk_darken_sse41_8bit+0x200> + DB 102,15,219,29,188,27,0,0 ; pand 0x1bbc(%rip),%xmm3 # 2400 <_sk_exclusion_sse41_8bit+0x1cb> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,56,51,211 ; pmovzxwd %xmm3,%xmm2 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 @@ -42762,7 +43074,7 @@ _sk_load_a8_dst_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf000900 <_sk_darken_sse41_8bit+0xffffffffdeffeb20> + DB 233,255,255,255,222 ; jmpq ffffffffdf000900 <_sk_exclusion_sse41_8bit+0xffffffffdeffe6cb> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -42782,7 +43094,7 @@ _sk_store_a8_sse41_8bit LABEL PROC DB 72,99,87,8 ; movslq 0x8(%rdi),%rdx DB 72,15,175,209 ; imul %rcx,%rdx DB 72,3,16 ; add (%rax),%rdx - DB 102,15,111,45,196,22,0,0 ; movdqa 0x16c4(%rip),%xmm5 # 1ff0 <_sk_darken_sse41_8bit+0x210> + DB 102,15,111,45,228,26,0,0 ; movdqa 0x1ae4(%rip),%xmm5 # 2410 <_sk_exclusion_sse41_8bit+0x1db> DB 102,15,111,241 ; movdqa %xmm1,%xmm6 DB 102,15,56,0,245 ; pshufb %xmm5,%xmm6 DB 102,15,111,224 ; movdqa %xmm0,%xmm4 @@ -42790,7 +43102,7 @@ _sk_store_a8_sse41_8bit LABEL PROC DB 102,15,108,230 ; punpcklqdq %xmm6,%xmm4 DB 77,133,201 ; test %r9,%r9 DB 117,19 ; jne 95a <_sk_store_a8_sse41_8bit+0x4e> - DB 102,15,56,0,37,208,22,0,0 ; pshufb 0x16d0(%rip),%xmm4 # 2020 <_sk_darken_sse41_8bit+0x240> + DB 102,15,56,0,37,240,26,0,0 ; pshufb 0x1af0(%rip),%xmm4 # 2440 <_sk_exclusion_sse41_8bit+0x20b> DB 102,66,15,214,36,2 ; movq %xmm4,(%rdx,%r8,1) DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax @@ -42806,13 +43118,13 @@ _sk_store_a8_sse41_8bit LABEL PROC DB 102,66,15,58,20,36,2,0 ; pextrb $0x0,%xmm4,(%rdx,%r8,1) DB 235,209 ; jmp 956 <_sk_store_a8_sse41_8bit+0x4a> DB 102,66,15,58,20,100,2,2,4 ; pextrb $0x4,%xmm4,0x2(%rdx,%r8,1) - DB 102,15,56,0,37,121,22,0,0 ; pshufb 0x1679(%rip),%xmm4 # 2010 <_sk_darken_sse41_8bit+0x230> + DB 102,15,56,0,37,153,26,0,0 ; pshufb 0x1a99(%rip),%xmm4 # 2430 <_sk_exclusion_sse41_8bit+0x1fb> DB 102,66,15,58,21,36,2,0 ; pextrw $0x0,%xmm4,(%rdx,%r8,1) DB 235,181 ; jmp 956 <_sk_store_a8_sse41_8bit+0x4a> DB 102,66,15,58,20,100,2,6,12 ; pextrb $0xc,%xmm4,0x6(%rdx,%r8,1) DB 102,66,15,58,20,100,2,5,10 ; pextrb $0xa,%xmm4,0x5(%rdx,%r8,1) DB 102,66,15,58,20,100,2,4,8 ; pextrb $0x8,%xmm4,0x4(%rdx,%r8,1) - DB 102,15,56,0,37,59,22,0,0 ; pshufb 0x163b(%rip),%xmm4 # 2000 <_sk_darken_sse41_8bit+0x220> + DB 102,15,56,0,37,91,26,0,0 ; pshufb 0x1a5b(%rip),%xmm4 # 2420 <_sk_exclusion_sse41_8bit+0x1eb> DB 102,66,15,126,36,2 ; movd %xmm4,(%rdx,%r8,1) DB 235,137 ; jmp 956 <_sk_store_a8_sse41_8bit+0x4a> DB 15,31,0 ; nopl (%rax) @@ -42850,14 +43162,14 @@ _sk_load_g8_sse41_8bit LABEL PROC DB 77,133,201 ; test %r9,%r9 DB 117,66 ; jne a4b <_sk_load_g8_sse41_8bit+0x5f> DB 102,66,15,56,48,12,2 ; pmovzxbw (%rdx,%r8,1),%xmm1 - DB 102,15,219,13,24,22,0,0 ; pand 0x1618(%rip),%xmm1 # 2030 <_sk_darken_sse41_8bit+0x250> + DB 102,15,219,13,56,26,0,0 ; pand 0x1a38(%rip),%xmm1 # 2450 <_sk_exclusion_sse41_8bit+0x21b> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,56,51,193 ; pmovzxwd %xmm1,%xmm0 DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 - DB 102,15,111,37,19,22,0,0 ; movdqa 0x1613(%rip),%xmm4 # 2040 <_sk_darken_sse41_8bit+0x260> + DB 102,15,111,37,51,26,0,0 ; movdqa 0x1a33(%rip),%xmm4 # 2460 <_sk_exclusion_sse41_8bit+0x22b> DB 102,15,56,64,204 ; pmulld %xmm4,%xmm1 DB 102,15,56,64,196 ; pmulld %xmm4,%xmm0 - DB 102,15,111,37,17,22,0,0 ; movdqa 0x1611(%rip),%xmm4 # 2050 <_sk_darken_sse41_8bit+0x270> + DB 102,15,111,37,49,26,0,0 ; movdqa 0x1a31(%rip),%xmm4 # 2470 <_sk_exclusion_sse41_8bit+0x23b> DB 102,15,235,196 ; por %xmm4,%xmm0 DB 102,15,235,204 ; por %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -42902,7 +43214,7 @@ _sk_load_g8_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 232,255,255,255,221 ; callq ffffffffde000af0 <_sk_darken_sse41_8bit+0xffffffffddffed10> + DB 232,255,255,255,221 ; callq ffffffffde000af0 <_sk_exclusion_sse41_8bit+0xffffffffddffe8bb> DB 255 ; (bad) DB 255 ; (bad) DB 255,210 ; callq *%rdx @@ -42925,14 +43237,14 @@ _sk_load_g8_dst_sse41_8bit LABEL PROC DB 77,133,201 ; test %r9,%r9 DB 117,66 ; jne b5b <_sk_load_g8_dst_sse41_8bit+0x5f> DB 102,66,15,56,48,28,2 ; pmovzxbw (%rdx,%r8,1),%xmm3 - DB 102,15,219,29,56,21,0,0 ; pand 0x1538(%rip),%xmm3 # 2060 <_sk_darken_sse41_8bit+0x280> + DB 102,15,219,29,88,25,0,0 ; pand 0x1958(%rip),%xmm3 # 2480 <_sk_exclusion_sse41_8bit+0x24b> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,56,51,211 ; pmovzxwd %xmm3,%xmm2 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 - DB 102,15,111,37,51,21,0,0 ; movdqa 0x1533(%rip),%xmm4 # 2070 <_sk_darken_sse41_8bit+0x290> + DB 102,15,111,37,83,25,0,0 ; movdqa 0x1953(%rip),%xmm4 # 2490 <_sk_exclusion_sse41_8bit+0x25b> DB 102,15,56,64,220 ; pmulld %xmm4,%xmm3 DB 102,15,56,64,212 ; pmulld %xmm4,%xmm2 - DB 102,15,111,37,49,21,0,0 ; movdqa 0x1531(%rip),%xmm4 # 2080 <_sk_darken_sse41_8bit+0x2a0> + DB 102,15,111,37,81,25,0,0 ; movdqa 0x1951(%rip),%xmm4 # 24a0 <_sk_exclusion_sse41_8bit+0x26b> DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,220 ; por %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -42977,7 +43289,7 @@ _sk_load_g8_dst_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 232,255,255,255,221 ; callq ffffffffde000c00 <_sk_darken_sse41_8bit+0xffffffffddffee20> + DB 232,255,255,255,221 ; callq ffffffffde000c00 <_sk_exclusion_sse41_8bit+0xffffffffddffe9cb> DB 255 ; (bad) DB 255 ; (bad) DB 255,210 ; callq *%rdx @@ -43003,7 +43315,7 @@ _sk_srcover_rgba_8888_sse41_8bit LABEL PROC DB 243,70,15,111,68,138,16 ; movdqu 0x10(%rdx,%r9,4),%xmm8 DB 243,70,15,111,12,138 ; movdqu (%rdx,%r9,4),%xmm9 DB 77,133,192 ; test %r8,%r8 - DB 102,15,111,37,71,20,0,0 ; movdqa 0x1447(%rip),%xmm4 # 2090 <_sk_darken_sse41_8bit+0x2b0> + DB 102,15,111,37,103,24,0,0 ; movdqa 0x1867(%rip),%xmm4 # 24b0 <_sk_exclusion_sse41_8bit+0x27b> DB 102,15,111,241 ; movdqa %xmm1,%xmm6 DB 102,15,56,0,244 ; pshufb %xmm4,%xmm6 DB 102,15,111,248 ; movdqa %xmm0,%xmm7 @@ -43123,7 +43435,7 @@ _sk_scale_1_float_sse41_8bit LABEL PROC DB 102,15,111,232 ; movdqa %xmm0,%xmm5 DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,0 ; movss (%rax),%xmm0 - DB 243,15,89,5,42,17,0,0 ; mulss 0x112a(%rip),%xmm0 # 1f54 <_sk_darken_sse41_8bit+0x174> + DB 243,15,89,5,70,21,0,0 ; mulss 0x1546(%rip),%xmm0 # 2370 <_sk_exclusion_sse41_8bit+0x13b> DB 243,15,44,192 ; cvttss2si %xmm0,%eax DB 15,87,192 ; xorps %xmm0,%xmm0 DB 102,68,15,56,48,197 ; pmovzxbw %xmm5,%xmm8 @@ -43131,7 +43443,7 @@ _sk_scale_1_float_sse41_8bit LABEL PROC DB 102,68,15,56,48,204 ; pmovzxbw %xmm4,%xmm9 DB 102,15,104,224 ; punpckhbw %xmm0,%xmm4 DB 102,15,110,240 ; movd %eax,%xmm6 - DB 102,15,56,0,53,78,18,0,0 ; pshufb 0x124e(%rip),%xmm6 # 20a0 <_sk_darken_sse41_8bit+0x2c0> + DB 102,15,56,0,53,110,22,0,0 ; pshufb 0x166e(%rip),%xmm6 # 24c0 <_sk_exclusion_sse41_8bit+0x28b> DB 102,15,111,206 ; movdqa %xmm6,%xmm1 DB 102,65,15,213,201 ; pmullw %xmm9,%xmm1 DB 102,15,111,198 ; movdqa %xmm6,%xmm0 @@ -43164,11 +43476,11 @@ _sk_scale_u8_sse41_8bit LABEL PROC DB 77,133,201 ; test %r9,%r9 DB 15,133,160,0,0,0 ; jne f63 <_sk_scale_u8_sse41_8bit+0xc1> DB 102,66,15,56,48,52,2 ; pmovzxbw (%rdx,%r8,1),%xmm6 - DB 102,15,219,53,222,17,0,0 ; pand 0x11de(%rip),%xmm6 # 20b0 <_sk_darken_sse41_8bit+0x2d0> + DB 102,15,219,53,254,21,0,0 ; pand 0x15fe(%rip),%xmm6 # 24d0 <_sk_exclusion_sse41_8bit+0x29b> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,15,111,254 ; movdqa %xmm6,%xmm7 - DB 102,15,56,0,61,220,17,0,0 ; pshufb 0x11dc(%rip),%xmm7 # 20c0 <_sk_darken_sse41_8bit+0x2e0> - DB 102,15,56,0,53,227,17,0,0 ; pshufb 0x11e3(%rip),%xmm6 # 20d0 <_sk_darken_sse41_8bit+0x2f0> + DB 102,15,56,0,61,252,21,0,0 ; pshufb 0x15fc(%rip),%xmm7 # 24e0 <_sk_exclusion_sse41_8bit+0x2ab> + DB 102,15,56,0,53,3,22,0,0 ; pshufb 0x1603(%rip),%xmm6 # 24f0 <_sk_exclusion_sse41_8bit+0x2bb> DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 @@ -43251,7 +43563,7 @@ PUBLIC _sk_lerp_1_float_sse41_8bit _sk_lerp_1_float_sse41_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,32 ; movss (%rax),%xmm4 - DB 243,15,89,37,46,15,0,0 ; mulss 0xf2e(%rip),%xmm4 # 1f58 <_sk_darken_sse41_8bit+0x178> + DB 243,15,89,37,74,19,0,0 ; mulss 0x134a(%rip),%xmm4 # 2374 <_sk_exclusion_sse41_8bit+0x13f> DB 243,15,44,196 ; cvttss2si %xmm4,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 @@ -43262,7 +43574,7 @@ _sk_lerp_1_float_sse41_8bit LABEL PROC DB 102,65,15,104,193 ; punpckhbw %xmm9,%xmm0 DB 102,68,15,56,48,217 ; pmovzxbw %xmm1,%xmm11 DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 - DB 102,15,56,0,61,123,16,0,0 ; pshufb 0x107b(%rip),%xmm7 # 20e0 <_sk_darken_sse41_8bit+0x300> + DB 102,15,56,0,61,155,20,0,0 ; pshufb 0x149b(%rip),%xmm7 # 2500 <_sk_exclusion_sse41_8bit+0x2cb> DB 102,68,15,111,231 ; movdqa %xmm7,%xmm12 DB 102,69,15,213,227 ; pmullw %xmm11,%xmm12 DB 102,68,15,111,239 ; movdqa %xmm7,%xmm13 @@ -43323,11 +43635,11 @@ _sk_lerp_u8_sse41_8bit LABEL PROC DB 77,133,201 ; test %r9,%r9 DB 15,133,46,1,0,0 ; jne 128d <_sk_lerp_u8_sse41_8bit+0x14f> DB 102,66,15,56,48,60,2 ; pmovzxbw (%rdx,%r8,1),%xmm7 - DB 102,15,219,61,130,15,0,0 ; pand 0xf82(%rip),%xmm7 # 20f0 <_sk_darken_sse41_8bit+0x310> + DB 102,15,219,61,162,19,0,0 ; pand 0x13a2(%rip),%xmm7 # 2510 <_sk_exclusion_sse41_8bit+0x2db> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,15,111,247 ; movdqa %xmm7,%xmm6 - DB 102,15,56,0,53,128,15,0,0 ; pshufb 0xf80(%rip),%xmm6 # 2100 <_sk_darken_sse41_8bit+0x320> - DB 102,15,56,0,61,135,15,0,0 ; pshufb 0xf87(%rip),%xmm7 # 2110 <_sk_darken_sse41_8bit+0x330> + DB 102,15,56,0,53,160,19,0,0 ; pshufb 0x13a0(%rip),%xmm6 # 2520 <_sk_exclusion_sse41_8bit+0x2eb> + DB 102,15,56,0,61,167,19,0,0 ; pshufb 0x13a7(%rip),%xmm7 # 2530 <_sk_exclusion_sse41_8bit+0x2fb> DB 102,68,15,56,48,200 ; pmovzxbw %xmm0,%xmm9 DB 102,65,15,104,192 ; punpckhbw %xmm8,%xmm0 DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 @@ -43422,7 +43734,7 @@ _sk_lerp_u8_sse41_8bit LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,222 ; jmpq ffffffffdf001338 <_sk_darken_sse41_8bit+0xffffffffdefff558> + DB 233,255,255,255,222 ; jmpq ffffffffdf001338 <_sk_exclusion_sse41_8bit+0xffffffffdefff103> DB 255 ; (bad) DB 255 ; (bad) DB 255,211 ; callq *%rbx @@ -43450,7 +43762,7 @@ _sk_move_dst_src_sse41_8bit LABEL PROC PUBLIC _sk_black_color_sse41_8bit _sk_black_color_sse41_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,5,191,13,0,0 ; movaps 0xdbf(%rip),%xmm0 # 2120 <_sk_darken_sse41_8bit+0x340> + DB 15,40,5,223,17,0,0 ; movaps 0x11df(%rip),%xmm0 # 2540 <_sk_exclusion_sse41_8bit+0x30b> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 255,224 ; jmpq *%rax @@ -43470,7 +43782,7 @@ _sk_clear_sse41_8bit LABEL PROC PUBLIC _sk_srcatop_sse41_8bit _sk_srcatop_sse41_8bit LABEL PROC - DB 102,68,15,111,21,171,13,0,0 ; movdqa 0xdab(%rip),%xmm10 # 2130 <_sk_darken_sse41_8bit+0x350> + DB 102,68,15,111,21,203,17,0,0 ; movdqa 0x11cb(%rip),%xmm10 # 2550 <_sk_exclusion_sse41_8bit+0x31b> DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 DB 102,68,15,56,48,195 ; pmovzxbw %xmm3,%xmm8 DB 102,15,111,235 ; movdqa %xmm3,%xmm5 @@ -43538,7 +43850,7 @@ _sk_srcatop_sse41_8bit LABEL PROC PUBLIC _sk_dstatop_sse41_8bit _sk_dstatop_sse41_8bit LABEL PROC - DB 102,68,15,111,29,128,12,0,0 ; movdqa 0xc80(%rip),%xmm11 # 2140 <_sk_darken_sse41_8bit+0x360> + DB 102,68,15,111,29,160,16,0,0 ; movdqa 0x10a0(%rip),%xmm11 # 2560 <_sk_exclusion_sse41_8bit+0x32b> DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 DB 102,69,15,56,0,235 ; pshufb %xmm11,%xmm13 DB 102,68,15,111,248 ; movdqa %xmm0,%xmm15 @@ -43608,7 +43920,7 @@ PUBLIC _sk_srcin_sse41_8bit _sk_srcin_sse41_8bit LABEL PROC DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,71,11,0,0 ; movdqa 0xb47(%rip),%xmm0 # 2150 <_sk_darken_sse41_8bit+0x370> + DB 102,15,111,5,103,15,0,0 ; movdqa 0xf67(%rip),%xmm0 # 2570 <_sk_exclusion_sse41_8bit+0x33b> DB 102,15,111,243 ; movdqa %xmm3,%xmm6 DB 102,15,56,0,240 ; pshufb %xmm0,%xmm6 DB 102,15,111,250 ; movdqa %xmm2,%xmm7 @@ -43641,7 +43953,7 @@ _sk_srcin_sse41_8bit LABEL PROC PUBLIC _sk_dstin_sse41_8bit _sk_dstin_sse41_8bit LABEL PROC - DB 102,15,111,37,202,10,0,0 ; movdqa 0xaca(%rip),%xmm4 # 2160 <_sk_darken_sse41_8bit+0x380> + DB 102,15,111,37,234,14,0,0 ; movdqa 0xeea(%rip),%xmm4 # 2580 <_sk_exclusion_sse41_8bit+0x34b> DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,69,15,239,210 ; pxor %xmm10,%xmm10 @@ -43678,7 +43990,7 @@ PUBLIC _sk_srcout_sse41_8bit _sk_srcout_sse41_8bit LABEL PROC DB 102,15,111,225 ; movdqa %xmm1,%xmm4 DB 102,15,111,232 ; movdqa %xmm0,%xmm5 - DB 102,15,111,5,61,10,0,0 ; movdqa 0xa3d(%rip),%xmm0 # 2170 <_sk_darken_sse41_8bit+0x390> + DB 102,15,111,5,93,14,0,0 ; movdqa 0xe5d(%rip),%xmm0 # 2590 <_sk_exclusion_sse41_8bit+0x35b> DB 102,15,111,250 ; movdqa %xmm2,%xmm7 DB 102,15,56,0,248 ; pshufb %xmm0,%xmm7 DB 102,15,111,243 ; movdqa %xmm3,%xmm6 @@ -43714,7 +44026,7 @@ _sk_srcout_sse41_8bit LABEL PROC PUBLIC _sk_dstout_sse41_8bit _sk_dstout_sse41_8bit LABEL PROC - DB 102,15,111,37,180,9,0,0 ; movdqa 0x9b4(%rip),%xmm4 # 2180 <_sk_darken_sse41_8bit+0x3a0> + DB 102,15,111,37,212,13,0,0 ; movdqa 0xdd4(%rip),%xmm4 # 25a0 <_sk_exclusion_sse41_8bit+0x36b> DB 102,15,56,0,196 ; pshufb %xmm4,%xmm0 DB 102,15,56,0,204 ; pshufb %xmm4,%xmm1 DB 102,15,118,228 ; pcmpeqd %xmm4,%xmm4 @@ -43752,7 +44064,7 @@ _sk_dstout_sse41_8bit LABEL PROC PUBLIC _sk_srcover_sse41_8bit _sk_srcover_sse41_8bit LABEL PROC - DB 102,15,111,53,35,9,0,0 ; movdqa 0x923(%rip),%xmm6 # 2190 <_sk_darken_sse41_8bit+0x3b0> + DB 102,15,111,53,67,13,0,0 ; movdqa 0xd43(%rip),%xmm6 # 25b0 <_sk_exclusion_sse41_8bit+0x37b> DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 DB 102,68,15,56,0,222 ; pshufb %xmm6,%xmm11 DB 102,15,111,232 ; movdqa %xmm0,%xmm5 @@ -43791,7 +44103,7 @@ _sk_srcover_sse41_8bit LABEL PROC PUBLIC _sk_dstover_sse41_8bit _sk_dstover_sse41_8bit LABEL PROC - DB 102,68,15,111,5,131,8,0,0 ; movdqa 0x883(%rip),%xmm8 # 21a0 <_sk_darken_sse41_8bit+0x3c0> + DB 102,68,15,111,5,163,12,0,0 ; movdqa 0xca3(%rip),%xmm8 # 25c0 <_sk_exclusion_sse41_8bit+0x38b> DB 102,68,15,111,209 ; movdqa %xmm1,%xmm10 DB 102,68,15,56,48,201 ; pmovzxbw %xmm1,%xmm9 DB 102,15,252,203 ; paddb %xmm3,%xmm1 @@ -43866,7 +44178,7 @@ _sk_multiply_sse41_8bit LABEL PROC DB 102,15,111,218 ; movdqa %xmm2,%xmm3 DB 102,15,111,209 ; movdqa %xmm1,%xmm2 DB 102,15,111,200 ; movdqa %xmm0,%xmm1 - DB 102,68,15,111,53,81,7,0,0 ; movdqa 0x751(%rip),%xmm14 # 21b0 <_sk_darken_sse41_8bit+0x3d0> + DB 102,68,15,111,53,113,11,0,0 ; movdqa 0xb71(%rip),%xmm14 # 25d0 <_sk_exclusion_sse41_8bit+0x39b> DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 DB 102,15,111,235 ; movdqa %xmm3,%xmm5 DB 102,65,15,56,0,238 ; pshufb %xmm14,%xmm5 @@ -43991,7 +44303,7 @@ _sk_screen_sse41_8bit LABEL PROC PUBLIC _sk_xor__sse41_8bit _sk_xor__sse41_8bit LABEL PROC - DB 102,68,15,111,21,31,5,0,0 ; movdqa 0x51f(%rip),%xmm10 # 21c0 <_sk_darken_sse41_8bit+0x3e0> + DB 102,68,15,111,21,63,9,0,0 ; movdqa 0x93f(%rip),%xmm10 # 25e0 <_sk_exclusion_sse41_8bit+0x3ab> DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 DB 102,68,15,56,48,194 ; pmovzxbw %xmm2,%xmm8 DB 102,15,111,234 ; movdqa %xmm2,%xmm5 @@ -44062,7 +44374,7 @@ _sk_xor__sse41_8bit LABEL PROC PUBLIC _sk_darken_sse41_8bit _sk_darken_sse41_8bit LABEL PROC DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 - DB 102,68,15,111,5,226,3,0,0 ; movdqa 0x3e2(%rip),%xmm8 # 21d0 <_sk_darken_sse41_8bit+0x3f0> + DB 102,68,15,111,5,2,8,0,0 ; movdqa 0x802(%rip),%xmm8 # 25f0 <_sk_exclusion_sse41_8bit+0x3bb> DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 DB 102,15,111,243 ; movdqa %xmm3,%xmm6 @@ -44129,20 +44441,246 @@ _sk_darken_sse41_8bit LABEL PROC DB 102,15,248,231 ; psubb %xmm7,%xmm4 DB 102,15,248,206 ; psubb %xmm6,%xmm1 DB 102,68,15,248,245 ; psubb %xmm5,%xmm14 - DB 15,40,5,166,2,0,0 ; movaps 0x2a6(%rip),%xmm0 # 21e0 <_sk_darken_sse41_8bit+0x400> + DB 15,40,5,198,6,0,0 ; movaps 0x6c6(%rip),%xmm0 # 2600 <_sk_exclusion_sse41_8bit+0x3cb> DB 102,68,15,56,16,244 ; pblendvb %xmm0,%xmm4,%xmm14 DB 102,65,15,56,16,200 ; pblendvb %xmm0,%xmm8,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 DB 255,224 ; jmpq *%rax +PUBLIC _sk_lighten_sse41_8bit +_sk_lighten_sse41_8bit LABEL PROC + DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 + DB 102,68,15,111,5,179,6,0,0 ; movdqa 0x6b3(%rip),%xmm8 # 2610 <_sk_exclusion_sse41_8bit+0x3db> + DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 + DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,65,15,56,0,240 ; pshufb %xmm8,%xmm6 + DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 + DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,65,15,56,0,232 ; pshufb %xmm8,%xmm5 + DB 102,69,15,239,228 ; pxor %xmm12,%xmm12 + DB 102,65,15,104,196 ; punpckhbw %xmm12,%xmm0 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 + DB 102,69,15,56,48,254 ; pmovzxbw %xmm14,%xmm15 + DB 102,15,56,48,253 ; pmovzxbw %xmm5,%xmm7 + DB 102,65,15,104,236 ; punpckhbw %xmm12,%xmm5 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,253,232 ; paddw %xmm0,%xmm5 + DB 102,15,56,48,198 ; pmovzxbw %xmm6,%xmm0 + DB 102,65,15,104,244 ; punpckhbw %xmm12,%xmm6 + DB 102,15,213,244 ; pmullw %xmm4,%xmm6 + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 + DB 102,68,15,252,242 ; paddb %xmm2,%xmm14 + DB 102,65,15,213,255 ; pmullw %xmm15,%xmm7 + DB 102,65,15,253,255 ; paddw %xmm15,%xmm7 + DB 102,68,15,56,48,249 ; pmovzxbw %xmm1,%xmm15 + DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 + DB 102,65,15,253,199 ; paddw %xmm15,%xmm0 + DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,253 ; packuswb %xmm5,%xmm7 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,198 ; packuswb %xmm6,%xmm0 + DB 102,69,15,56,0,248 ; pshufb %xmm8,%xmm15 + DB 102,65,15,56,0,224 ; pshufb %xmm8,%xmm4 + DB 102,69,15,104,236 ; punpckhbw %xmm12,%xmm13 + DB 102,69,15,104,220 ; punpckhbw %xmm12,%xmm11 + DB 102,15,56,48,236 ; pmovzxbw %xmm4,%xmm5 + DB 102,65,15,56,48,247 ; pmovzxbw %xmm15,%xmm6 + DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 + DB 102,69,15,104,252 ; punpckhbw %xmm12,%xmm15 + DB 102,69,15,213,251 ; pmullw %xmm11,%xmm15 + DB 102,65,15,213,229 ; pmullw %xmm13,%xmm4 + DB 102,65,15,213,241 ; pmullw %xmm9,%xmm6 + DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 + DB 102,65,15,253,229 ; paddw %xmm13,%xmm4 + DB 102,69,15,253,251 ; paddw %xmm11,%xmm15 + DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 + DB 102,65,15,253,241 ; paddw %xmm9,%xmm6 + DB 102,65,15,113,215,8 ; psrlw $0x8,%xmm15 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,236 ; packuswb %xmm4,%xmm5 + DB 102,65,15,103,247 ; packuswb %xmm15,%xmm6 + DB 102,15,218,253 ; pminub %xmm5,%xmm7 + DB 102,15,218,198 ; pminub %xmm6,%xmm0 + DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 + DB 102,68,15,248,192 ; psubb %xmm0,%xmm8 + DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 + DB 102,15,248,231 ; psubb %xmm7,%xmm4 + DB 102,15,248,206 ; psubb %xmm6,%xmm1 + DB 102,68,15,248,245 ; psubb %xmm5,%xmm14 + DB 15,40,5,119,5,0,0 ; movaps 0x577(%rip),%xmm0 # 2620 <_sk_exclusion_sse41_8bit+0x3eb> + DB 102,68,15,56,16,244 ; pblendvb %xmm0,%xmm4,%xmm14 + DB 102,65,15,56,16,200 ; pblendvb %xmm0,%xmm8,%xmm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_difference_sse41_8bit +_sk_difference_sse41_8bit LABEL PROC + DB 102,68,15,111,240 ; movdqa %xmm0,%xmm14 + DB 102,68,15,111,5,100,5,0,0 ; movdqa 0x564(%rip),%xmm8 # 2630 <_sk_exclusion_sse41_8bit+0x3fb> + DB 102,68,15,111,219 ; movdqa %xmm3,%xmm11 + DB 102,68,15,56,48,203 ; pmovzxbw %xmm3,%xmm9 + DB 102,15,111,251 ; movdqa %xmm3,%xmm7 + DB 102,65,15,56,0,248 ; pshufb %xmm8,%xmm7 + DB 102,68,15,111,234 ; movdqa %xmm2,%xmm13 + DB 102,68,15,56,48,210 ; pmovzxbw %xmm2,%xmm10 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,65,15,56,0,232 ; pshufb %xmm8,%xmm5 + DB 102,69,15,239,228 ; pxor %xmm12,%xmm12 + DB 102,65,15,104,196 ; punpckhbw %xmm12,%xmm0 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 + DB 102,69,15,56,48,254 ; pmovzxbw %xmm14,%xmm15 + DB 102,15,56,48,245 ; pmovzxbw %xmm5,%xmm6 + DB 102,65,15,104,236 ; punpckhbw %xmm12,%xmm5 + DB 102,15,213,232 ; pmullw %xmm0,%xmm5 + DB 102,15,253,232 ; paddw %xmm0,%xmm5 + DB 102,15,56,48,199 ; pmovzxbw %xmm7,%xmm0 + DB 102,65,15,104,252 ; punpckhbw %xmm12,%xmm7 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 + DB 102,68,15,252,242 ; paddb %xmm2,%xmm14 + DB 102,65,15,213,247 ; pmullw %xmm15,%xmm6 + DB 102,65,15,253,247 ; paddw %xmm15,%xmm6 + DB 102,68,15,56,48,249 ; pmovzxbw %xmm1,%xmm15 + DB 102,65,15,213,199 ; pmullw %xmm15,%xmm0 + DB 102,65,15,253,199 ; paddw %xmm15,%xmm0 + DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,245 ; packuswb %xmm5,%xmm6 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,103,199 ; packuswb %xmm7,%xmm0 + DB 102,69,15,56,0,248 ; pshufb %xmm8,%xmm15 + DB 102,65,15,56,0,224 ; pshufb %xmm8,%xmm4 + DB 102,69,15,104,236 ; punpckhbw %xmm12,%xmm13 + DB 102,69,15,104,220 ; punpckhbw %xmm12,%xmm11 + DB 102,15,56,48,236 ; pmovzxbw %xmm4,%xmm5 + DB 102,65,15,56,48,255 ; pmovzxbw %xmm15,%xmm7 + DB 102,65,15,104,228 ; punpckhbw %xmm12,%xmm4 + DB 102,69,15,104,252 ; punpckhbw %xmm12,%xmm15 + DB 102,69,15,213,251 ; pmullw %xmm11,%xmm15 + DB 102,65,15,213,229 ; pmullw %xmm13,%xmm4 + DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 + DB 102,65,15,213,234 ; pmullw %xmm10,%xmm5 + DB 102,65,15,253,229 ; paddw %xmm13,%xmm4 + DB 102,69,15,253,251 ; paddw %xmm11,%xmm15 + DB 102,65,15,253,234 ; paddw %xmm10,%xmm5 + DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 + DB 102,65,15,113,215,8 ; psrlw $0x8,%xmm15 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,103,236 ; packuswb %xmm4,%xmm5 + DB 102,65,15,103,255 ; packuswb %xmm15,%xmm7 + DB 102,15,218,199 ; pminub %xmm7,%xmm0 + DB 102,15,218,245 ; pminub %xmm5,%xmm6 + DB 102,15,252,246 ; paddb %xmm6,%xmm6 + DB 102,15,252,192 ; paddb %xmm0,%xmm0 + DB 102,68,15,111,193 ; movdqa %xmm1,%xmm8 + DB 102,68,15,248,192 ; psubb %xmm0,%xmm8 + DB 102,65,15,111,230 ; movdqa %xmm14,%xmm4 + DB 102,15,248,230 ; psubb %xmm6,%xmm4 + DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,68,15,248,245 ; psubb %xmm5,%xmm14 + DB 15,40,5,32,4,0,0 ; movaps 0x420(%rip),%xmm0 # 2640 <_sk_exclusion_sse41_8bit+0x40b> + DB 102,68,15,56,16,244 ; pblendvb %xmm0,%xmm4,%xmm14 + DB 102,65,15,56,16,200 ; pblendvb %xmm0,%xmm8,%xmm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 102,65,15,111,198 ; movdqa %xmm14,%xmm0 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_sse41_8bit +_sk_exclusion_sse41_8bit LABEL PROC + DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 + DB 102,15,111,248 ; movdqa %xmm0,%xmm7 + DB 102,65,15,104,249 ; punpckhbw %xmm9,%xmm7 + DB 102,68,15,56,48,208 ; pmovzxbw %xmm0,%xmm10 + DB 102,68,15,111,218 ; movdqa %xmm2,%xmm11 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,65,15,104,233 ; punpckhbw %xmm9,%xmm5 + DB 102,68,15,252,216 ; paddb %xmm0,%xmm11 + DB 102,68,15,111,5,235,3,0,0 ; movdqa 0x3eb(%rip),%xmm8 # 2650 <_sk_exclusion_sse41_8bit+0x41b> + DB 102,65,15,56,0,192 ; pshufb %xmm8,%xmm0 + DB 102,15,56,48,240 ; pmovzxbw %xmm0,%xmm6 + DB 102,65,15,104,193 ; punpckhbw %xmm9,%xmm0 + DB 102,15,213,197 ; pmullw %xmm5,%xmm0 + DB 102,15,253,197 ; paddw %xmm5,%xmm0 + DB 102,15,213,239 ; pmullw %xmm7,%xmm5 + DB 102,15,253,239 ; paddw %xmm7,%xmm5 + DB 102,15,56,48,250 ; pmovzxbw %xmm2,%xmm7 + DB 102,15,213,247 ; pmullw %xmm7,%xmm6 + DB 102,15,253,247 ; paddw %xmm7,%xmm6 + DB 102,65,15,213,250 ; pmullw %xmm10,%xmm7 + DB 102,65,15,253,250 ; paddw %xmm10,%xmm7 + DB 102,15,111,225 ; movdqa %xmm1,%xmm4 + DB 102,65,15,104,225 ; punpckhbw %xmm9,%xmm4 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,253 ; packuswb %xmm5,%xmm7 + DB 102,15,252,255 ; paddb %xmm7,%xmm7 + DB 102,65,15,111,235 ; movdqa %xmm11,%xmm5 + DB 102,15,248,239 ; psubb %xmm7,%xmm5 + DB 102,68,15,56,48,209 ; pmovzxbw %xmm1,%xmm10 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,240 ; packuswb %xmm0,%xmm6 + DB 102,68,15,248,222 ; psubb %xmm6,%xmm11 + DB 15,40,5,128,3,0,0 ; movaps 0x380(%rip),%xmm0 # 2660 <_sk_exclusion_sse41_8bit+0x42b> + DB 102,68,15,56,16,221 ; pblendvb %xmm0,%xmm5,%xmm11 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,15,111,243 ; movdqa %xmm3,%xmm6 + DB 102,65,15,104,241 ; punpckhbw %xmm9,%xmm6 + DB 102,15,252,233 ; paddb %xmm1,%xmm5 + DB 102,65,15,56,0,200 ; pshufb %xmm8,%xmm1 + DB 102,15,56,48,249 ; pmovzxbw %xmm1,%xmm7 + DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 + DB 102,15,213,206 ; pmullw %xmm6,%xmm1 + DB 102,15,253,206 ; paddw %xmm6,%xmm1 + DB 102,15,213,244 ; pmullw %xmm4,%xmm6 + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,15,56,48,227 ; pmovzxbw %xmm3,%xmm4 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,65,15,213,226 ; pmullw %xmm10,%xmm4 + DB 102,65,15,253,226 ; paddw %xmm10,%xmm4 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,15,103,230 ; packuswb %xmm6,%xmm4 + DB 102,15,252,228 ; paddb %xmm4,%xmm4 + DB 102,15,111,245 ; movdqa %xmm5,%xmm6 + DB 102,15,248,244 ; psubb %xmm4,%xmm6 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,249 ; packuswb %xmm1,%xmm7 + DB 102,15,248,239 ; psubb %xmm7,%xmm5 + DB 102,15,56,16,238 ; pblendvb %xmm0,%xmm6,%xmm5 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 102,65,15,111,195 ; movdqa %xmm11,%xmm0 + DB 102,15,111,205 ; movdqa %xmm5,%xmm1 + DB 255,224 ; jmpq *%rax + ALIGN 4 DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 1f97 <_sk_darken_sse41_8bit+0x1b7> + DB 127,67 ; jg 23b3 <_sk_exclusion_sse41_8bit+0x17e> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 1f9b <_sk_darken_sse41_8bit+0x1bb> + DB 127,67 ; jg 23b7 <_sk_exclusion_sse41_8bit+0x182> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 1f9f <_sk_darken_sse41_8bit+0x1bf> + DB 127,67 ; jg 23bb <_sk_exclusion_sse41_8bit+0x186> ALIGN 16 DB 0,0 ; add %al,(%rax) @@ -44515,6 +45053,75 @@ ALIGN 16 DB 255 ; (bad) DB 255 ; (bad) DB 255,0 ; incl (%rax) + DB 3,3 ; add (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 3,3 ; add (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 3,3 ; add (%rbx),%eax + DB 3,3 ; add (%rbx),%eax + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 7 ; (bad) + DB 11,11 ; or (%rbx),%ecx + DB 11,11 ; or (%rbx),%ecx + DB 15 ; (bad) + DB 15 ; (bad) + DB 15 ; (bad) + DB 15,255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) ALIGN 32 PUBLIC _sk_start_pipeline_sse2_8bit @@ -44615,7 +45222,7 @@ _sk_uniform_color_sse2_8bit LABEL PROC PUBLIC _sk_set_rgb_sse2_8bit _sk_set_rgb_sse2_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,15,16,37,244,35,0,0 ; movss 0x23f4(%rip),%xmm4 # 2574 <_sk_darken_sse2_8bit+0x1ec> + DB 243,15,16,37,108,41,0,0 ; movss 0x296c(%rip),%xmm4 # 2aec <_sk_exclusion_sse2_8bit+0x195> DB 243,15,16,40 ; movss (%rax),%xmm5 DB 243,15,89,236 ; mulss %xmm4,%xmm5 DB 243,72,15,44,205 ; cvttss2si %xmm5,%rcx @@ -44630,7 +45237,7 @@ _sk_set_rgb_sse2_8bit LABEL PROC DB 9,208 ; or %edx,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,112,228,0 ; pshufd $0x0,%xmm4,%xmm4 - DB 102,15,111,45,192,35,0,0 ; movdqa 0x23c0(%rip),%xmm5 # 2580 <_sk_darken_sse2_8bit+0x1f8> + DB 102,15,111,45,64,41,0,0 ; movdqa 0x2940(%rip),%xmm5 # 2b00 <_sk_exclusion_sse2_8bit+0x1a9> DB 102,15,219,205 ; pand %xmm5,%xmm1 DB 102,15,219,197 ; pand %xmm5,%xmm0 DB 102,15,235,196 ; por %xmm4,%xmm0 @@ -44653,7 +45260,7 @@ _sk_premul_sse2_8bit LABEL PROC DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 DB 243,15,112,248,95 ; pshufhw $0x5f,%xmm0,%xmm7 - DB 102,15,111,5,116,35,0,0 ; movdqa 0x2374(%rip),%xmm0 # 2590 <_sk_darken_sse2_8bit+0x208> + DB 102,15,111,5,244,40,0,0 ; movdqa 0x28f4(%rip),%xmm0 # 2b10 <_sk_exclusion_sse2_8bit+0x1b9> DB 102,15,235,248 ; por %xmm0,%xmm7 DB 102,15,235,240 ; por %xmm0,%xmm6 DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 @@ -45155,7 +45762,7 @@ _sk_load_a8_sse2_8bit LABEL PROC DB 117,48 ; jne 8b9 <_sk_load_a8_sse2_8bit+0x4d> DB 243,66,15,126,4,2 ; movq (%rdx,%r8,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 102,15,84,5,5,29,0,0 ; andpd 0x1d05(%rip),%xmm0 # 25a0 <_sk_darken_sse2_8bit+0x218> + DB 102,15,84,5,133,34,0,0 ; andpd 0x2285(%rip),%xmm0 # 2b20 <_sk_exclusion_sse2_8bit+0x1c9> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,40,200 ; movapd %xmm0,%xmm1 DB 102,15,105,204 ; punpckhwd %xmm4,%xmm1 @@ -45230,7 +45837,7 @@ _sk_load_a8_dst_sse2_8bit LABEL PROC DB 117,48 ; jne 9ad <_sk_load_a8_dst_sse2_8bit+0x4d> DB 243,66,15,126,20,2 ; movq (%rdx,%r8,1),%xmm2 DB 102,15,96,208 ; punpcklbw %xmm0,%xmm2 - DB 102,15,84,21,33,28,0,0 ; andpd 0x1c21(%rip),%xmm2 # 25b0 <_sk_darken_sse2_8bit+0x228> + DB 102,15,84,21,161,33,0,0 ; andpd 0x21a1(%rip),%xmm2 # 2b30 <_sk_exclusion_sse2_8bit+0x1d9> DB 102,15,239,228 ; pxor %xmm4,%xmm4 DB 102,15,40,218 ; movapd %xmm2,%xmm3 DB 102,15,105,220 ; punpckhwd %xmm4,%xmm3 @@ -45313,7 +45920,7 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 102,15,107,229 ; packssdw %xmm5,%xmm4 DB 77,133,201 ; test %r9,%r9 DB 117,26 ; jne ab9 <_sk_store_a8_sse2_8bit+0x65> - DB 102,15,219,37,25,27,0,0 ; pand 0x1b19(%rip),%xmm4 # 25c0 <_sk_darken_sse2_8bit+0x238> + DB 102,15,219,37,153,32,0,0 ; pand 0x2099(%rip),%xmm4 # 2b40 <_sk_exclusion_sse2_8bit+0x1e9> DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,66,15,214,36,2 ; movq %xmm4,(%rdx,%r8,1) DB 72,173 ; lods %ds:(%rsi),%rax @@ -45335,7 +45942,7 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 102,15,127,100,36,16 ; movdqa %xmm4,0x10(%rsp) DB 138,68,36,20 ; mov 0x14(%rsp),%al DB 66,136,68,2,2 ; mov %al,0x2(%rdx,%r8,1) - DB 102,15,219,37,193,26,0,0 ; pand 0x1ac1(%rip),%xmm4 # 25c0 <_sk_darken_sse2_8bit+0x238> + DB 102,15,219,37,65,32,0,0 ; pand 0x2041(%rip),%xmm4 # 2b40 <_sk_exclusion_sse2_8bit+0x1e9> DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,15,126,224 ; movd %xmm4,%eax DB 102,66,137,4,2 ; mov %ax,(%rdx,%r8,1) @@ -45349,7 +45956,7 @@ _sk_store_a8_sse2_8bit LABEL PROC DB 102,15,127,100,36,32 ; movdqa %xmm4,0x20(%rsp) DB 138,68,36,40 ; mov 0x28(%rsp),%al DB 66,136,68,2,4 ; mov %al,0x4(%rdx,%r8,1) - DB 102,15,219,37,125,26,0,0 ; pand 0x1a7d(%rip),%xmm4 # 25c0 <_sk_darken_sse2_8bit+0x238> + DB 102,15,219,37,253,31,0,0 ; pand 0x1ffd(%rip),%xmm4 # 2b40 <_sk_exclusion_sse2_8bit+0x1e9> DB 102,15,103,228 ; packuswb %xmm4,%xmm4 DB 102,66,15,126,36,2 ; movd %xmm4,(%rdx,%r8,1) DB 233,95,255,255,255 ; jmpq ab1 <_sk_store_a8_sse2_8bit+0x5d> @@ -45387,12 +45994,12 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 117,116 ; jne c01 <_sk_load_g8_sse2_8bit+0x91> DB 243,66,15,126,4,2 ; movq (%rdx,%r8,1),%xmm0 DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 - DB 102,15,84,5,49,26,0,0 ; andpd 0x1a31(%rip),%xmm0 # 25d0 <_sk_darken_sse2_8bit+0x248> + DB 102,15,84,5,177,31,0,0 ; andpd 0x1fb1(%rip),%xmm0 # 2b50 <_sk_exclusion_sse2_8bit+0x1f9> DB 102,15,239,201 ; pxor %xmm1,%xmm1 DB 102,15,40,224 ; movapd %xmm0,%xmm4 DB 102,15,97,225 ; punpcklwd %xmm1,%xmm4 DB 102,15,105,193 ; punpckhwd %xmm1,%xmm0 - DB 102,15,111,45,41,26,0,0 ; movdqa 0x1a29(%rip),%xmm5 # 25e0 <_sk_darken_sse2_8bit+0x258> + DB 102,15,111,45,169,31,0,0 ; movdqa 0x1fa9(%rip),%xmm5 # 2b60 <_sk_exclusion_sse2_8bit+0x209> DB 102,15,112,240,245 ; pshufd $0xf5,%xmm0,%xmm6 DB 102,15,244,197 ; pmuludq %xmm5,%xmm0 DB 102,15,112,200,232 ; pshufd $0xe8,%xmm0,%xmm1 @@ -45405,7 +46012,7 @@ _sk_load_g8_sse2_8bit LABEL PROC DB 102,15,244,245 ; pmuludq %xmm5,%xmm6 DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 DB 102,15,98,196 ; punpckldq %xmm4,%xmm0 - DB 102,15,111,37,251,25,0,0 ; movdqa 0x19fb(%rip),%xmm4 # 25f0 <_sk_darken_sse2_8bit+0x268> + DB 102,15,111,37,123,31,0,0 ; movdqa 0x1f7b(%rip),%xmm4 # 2b70 <_sk_exclusion_sse2_8bit+0x219> DB 102,15,235,196 ; por %xmm4,%xmm0 DB 102,15,235,204 ; por %xmm4,%xmm1 DB 72,173 ; lods %ds:(%rsi),%rax @@ -45477,12 +46084,12 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 117,116 ; jne d41 <_sk_load_g8_dst_sse2_8bit+0x91> DB 243,66,15,126,20,2 ; movq (%rdx,%r8,1),%xmm2 DB 102,15,96,208 ; punpcklbw %xmm0,%xmm2 - DB 102,15,84,21,33,25,0,0 ; andpd 0x1921(%rip),%xmm2 # 2600 <_sk_darken_sse2_8bit+0x278> + DB 102,15,84,21,161,30,0,0 ; andpd 0x1ea1(%rip),%xmm2 # 2b80 <_sk_exclusion_sse2_8bit+0x229> DB 102,15,239,219 ; pxor %xmm3,%xmm3 DB 102,15,40,226 ; movapd %xmm2,%xmm4 DB 102,15,97,227 ; punpcklwd %xmm3,%xmm4 DB 102,15,105,211 ; punpckhwd %xmm3,%xmm2 - DB 102,15,111,45,25,25,0,0 ; movdqa 0x1919(%rip),%xmm5 # 2610 <_sk_darken_sse2_8bit+0x288> + DB 102,15,111,45,153,30,0,0 ; movdqa 0x1e99(%rip),%xmm5 # 2b90 <_sk_exclusion_sse2_8bit+0x239> DB 102,15,112,242,245 ; pshufd $0xf5,%xmm2,%xmm6 DB 102,15,244,213 ; pmuludq %xmm5,%xmm2 DB 102,15,112,218,232 ; pshufd $0xe8,%xmm2,%xmm3 @@ -45495,7 +46102,7 @@ _sk_load_g8_dst_sse2_8bit LABEL PROC DB 102,15,244,245 ; pmuludq %xmm5,%xmm6 DB 102,15,112,230,232 ; pshufd $0xe8,%xmm6,%xmm4 DB 102,15,98,212 ; punpckldq %xmm4,%xmm2 - DB 102,15,111,37,235,24,0,0 ; movdqa 0x18eb(%rip),%xmm4 # 2620 <_sk_darken_sse2_8bit+0x298> + DB 102,15,111,37,107,30,0,0 ; movdqa 0x1e6b(%rip),%xmm4 # 2ba0 <_sk_exclusion_sse2_8bit+0x249> DB 102,15,235,212 ; por %xmm4,%xmm2 DB 102,15,235,220 ; por %xmm4,%xmm3 DB 72,173 ; lods %ds:(%rsi),%rax @@ -45708,7 +46315,7 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,0 ; movss (%rax),%xmm0 - DB 243,15,89,5,36,21,0,0 ; mulss 0x1524(%rip),%xmm0 # 2578 <_sk_darken_sse2_8bit+0x1f0> + DB 243,15,89,5,156,26,0,0 ; mulss 0x1a9c(%rip),%xmm0 # 2af0 <_sk_exclusion_sse2_8bit+0x199> DB 243,15,44,192 ; cvttss2si %xmm0,%eax DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 102,65,15,111,193 ; movdqa %xmm9,%xmm0 @@ -45720,7 +46327,7 @@ _sk_scale_1_float_sse2_8bit LABEL PROC DB 102,15,96,246 ; punpcklbw %xmm6,%xmm6 DB 242,15,112,246,0 ; pshuflw $0x0,%xmm6,%xmm6 DB 102,15,112,246,80 ; pshufd $0x50,%xmm6,%xmm6 - DB 102,15,219,53,163,21,0,0 ; pand 0x15a3(%rip),%xmm6 # 2630 <_sk_darken_sse2_8bit+0x2a8> + DB 102,15,219,53,35,27,0,0 ; pand 0x1b23(%rip),%xmm6 # 2bb0 <_sk_exclusion_sse2_8bit+0x259> DB 102,15,111,254 ; movdqa %xmm6,%xmm7 DB 102,65,15,213,248 ; pmullw %xmm8,%xmm7 DB 102,15,111,230 ; movdqa %xmm6,%xmm4 @@ -45754,7 +46361,7 @@ _sk_scale_u8_sse2_8bit LABEL PROC DB 15,133,239,0,0,0 ; jne 11ed <_sk_scale_u8_sse2_8bit+0x110> DB 243,66,15,126,36,2 ; movq (%rdx,%r8,1),%xmm4 DB 102,15,96,224 ; punpcklbw %xmm0,%xmm4 - DB 102,15,84,37,48,21,0,0 ; andpd 0x1530(%rip),%xmm4 # 2640 <_sk_darken_sse2_8bit+0x2b8> + DB 102,15,84,37,176,26,0,0 ; andpd 0x1ab0(%rip),%xmm4 # 2bc0 <_sk_exclusion_sse2_8bit+0x269> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,15,40,236 ; movapd %xmm4,%xmm5 DB 102,65,15,105,232 ; punpckhwd %xmm8,%xmm5 @@ -45861,7 +46468,7 @@ PUBLIC _sk_lerp_1_float_sse2_8bit _sk_lerp_1_float_sse2_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,15,16,32 ; movss (%rax),%xmm4 - DB 243,15,89,37,206,18,0,0 ; mulss 0x12ce(%rip),%xmm4 # 257c <_sk_darken_sse2_8bit+0x1f4> + DB 243,15,89,37,70,24,0,0 ; mulss 0x1846(%rip),%xmm4 # 2af4 <_sk_exclusion_sse2_8bit+0x19d> DB 243,15,44,196 ; cvttss2si %xmm4,%eax DB 102,15,110,224 ; movd %eax,%xmm4 DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 @@ -45874,7 +46481,7 @@ _sk_lerp_1_float_sse2_8bit LABEL PROC DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 DB 102,69,15,96,217 ; punpcklbw %xmm9,%xmm11 DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 - DB 102,15,111,53,97,19,0,0 ; movdqa 0x1361(%rip),%xmm6 # 2650 <_sk_darken_sse2_8bit+0x2c8> + DB 102,15,111,53,225,24,0,0 ; movdqa 0x18e1(%rip),%xmm6 # 2bd0 <_sk_exclusion_sse2_8bit+0x279> DB 102,65,15,219,240 ; pand %xmm8,%xmm6 DB 102,15,111,230 ; movdqa %xmm6,%xmm4 DB 102,15,213,225 ; pmullw %xmm1,%xmm4 @@ -45940,7 +46547,7 @@ _sk_lerp_u8_sse2_8bit LABEL PROC DB 15,133,141,1,0,0 ; jne 1584 <_sk_lerp_u8_sse2_8bit+0x1ae> DB 243,66,15,126,44,2 ; movq (%rdx,%r8,1),%xmm5 DB 102,15,96,232 ; punpcklbw %xmm0,%xmm5 - DB 102,15,84,45,87,18,0,0 ; andpd 0x1257(%rip),%xmm5 # 2660 <_sk_darken_sse2_8bit+0x2d8> + DB 102,15,84,45,215,23,0,0 ; andpd 0x17d7(%rip),%xmm5 # 2be0 <_sk_exclusion_sse2_8bit+0x289> DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 102,15,40,229 ; movapd %xmm5,%xmm4 DB 102,65,15,105,224 ; punpckhwd %xmm8,%xmm4 @@ -46092,7 +46699,7 @@ _sk_move_dst_src_sse2_8bit LABEL PROC PUBLIC _sk_black_color_sse2_8bit _sk_black_color_sse2_8bit LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 15,40,5,31,16,0,0 ; movaps 0x101f(%rip),%xmm0 # 2670 <_sk_darken_sse2_8bit+0x2e8> + DB 15,40,5,159,21,0,0 ; movaps 0x159f(%rip),%xmm0 # 2bf0 <_sk_exclusion_sse2_8bit+0x299> DB 15,40,200 ; movaps %xmm0,%xmm1 DB 255,224 ; jmpq *%rax @@ -46956,7 +47563,113 @@ _sk_darken_sse2_8bit LABEL PROC DB 102,65,15,248,234 ; psubb %xmm10,%xmm5 DB 102,15,248,207 ; psubb %xmm7,%xmm1 DB 102,15,248,196 ; psubb %xmm4,%xmm0 - DB 102,15,111,37,52,1,0,0 ; movdqa 0x134(%rip),%xmm4 # 2680 <_sk_darken_sse2_8bit+0x2f8> + DB 102,15,111,37,180,6,0,0 ; movdqa 0x6b4(%rip),%xmm4 # 2c00 <_sk_exclusion_sse2_8bit+0x2a9> + DB 102,15,219,236 ; pand %xmm4,%xmm5 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 + DB 102,15,223,248 ; pandn %xmm0,%xmm7 + DB 102,15,235,239 ; por %xmm7,%xmm5 + DB 102,15,219,244 ; pand %xmm4,%xmm6 + DB 102,15,223,225 ; pandn %xmm1,%xmm4 + DB 102,15,235,244 ; por %xmm4,%xmm6 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_lighten_sse2_8bit +_sk_lighten_sse2_8bit LABEL PROC + DB 242,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 + DB 242,15,112,234,231 ; pshuflw $0xe7,%xmm2,%xmm5 + DB 243,15,112,237,231 ; pshufhw $0xe7,%xmm5,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 + DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 + DB 243,15,112,253,95 ; pshufhw $0x5f,%xmm5,%xmm7 + DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 + DB 102,15,111,240 ; movdqa %xmm0,%xmm6 + DB 102,65,15,96,240 ; punpcklbw %xmm8,%xmm6 + DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 + DB 102,69,15,104,200 ; punpckhbw %xmm8,%xmm9 + DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 + DB 102,69,15,96,216 ; punpcklbw %xmm8,%xmm11 + DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 + DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 + DB 102,68,15,111,215 ; movdqa %xmm7,%xmm10 + DB 102,69,15,96,208 ; punpcklbw %xmm8,%xmm10 + DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 + DB 102,68,15,111,236 ; movdqa %xmm4,%xmm13 + DB 102,69,15,96,232 ; punpcklbw %xmm8,%xmm13 + DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 + DB 102,65,15,213,228 ; pmullw %xmm12,%xmm4 + DB 102,69,15,213,235 ; pmullw %xmm11,%xmm13 + DB 102,65,15,213,249 ; pmullw %xmm9,%xmm7 + DB 102,68,15,213,214 ; pmullw %xmm6,%xmm10 + DB 102,68,15,253,214 ; paddw %xmm6,%xmm10 + DB 102,65,15,253,249 ; paddw %xmm9,%xmm7 + DB 102,69,15,253,235 ; paddw %xmm11,%xmm13 + DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 + DB 102,68,15,103,215 ; packuswb %xmm7,%xmm10 + DB 102,68,15,103,236 ; packuswb %xmm4,%xmm13 + DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,68,15,112,220,95 ; pshufhw $0x5f,%xmm4,%xmm11 + DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,68,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm12 + DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 + DB 102,15,252,194 ; paddb %xmm2,%xmm0 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,65,15,96,240 ; punpcklbw %xmm8,%xmm6 + DB 102,69,15,104,200 ; punpckhbw %xmm8,%xmm9 + DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,65,15,96,232 ; punpcklbw %xmm8,%xmm5 + DB 102,69,15,104,240 ; punpckhbw %xmm8,%xmm14 + DB 102,65,15,111,228 ; movdqa %xmm12,%xmm4 + DB 102,65,15,96,224 ; punpcklbw %xmm8,%xmm4 + DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 + DB 102,65,15,111,251 ; movdqa %xmm11,%xmm7 + DB 102,65,15,96,248 ; punpcklbw %xmm8,%xmm7 + DB 102,69,15,104,216 ; punpckhbw %xmm8,%xmm11 + DB 102,69,15,213,222 ; pmullw %xmm14,%xmm11 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,69,15,213,225 ; pmullw %xmm9,%xmm12 + DB 102,15,213,230 ; pmullw %xmm6,%xmm4 + DB 102,15,253,230 ; paddw %xmm6,%xmm4 + DB 102,69,15,253,225 ; paddw %xmm9,%xmm12 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,69,15,253,222 ; paddw %xmm14,%xmm11 + DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,113,212,8 ; psrlw $0x8,%xmm12 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,65,15,103,228 ; packuswb %xmm12,%xmm4 + DB 102,65,15,103,251 ; packuswb %xmm11,%xmm7 + DB 102,68,15,218,212 ; pminub %xmm4,%xmm10 + DB 102,68,15,218,239 ; pminub %xmm7,%xmm13 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,65,15,248,245 ; psubb %xmm13,%xmm6 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,65,15,248,234 ; psubb %xmm10,%xmm5 + DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,248,196 ; psubb %xmm4,%xmm0 + DB 102,15,111,37,216,4,0,0 ; movdqa 0x4d8(%rip),%xmm4 # 2c10 <_sk_exclusion_sse2_8bit+0x2b9> DB 102,15,219,236 ; pand %xmm4,%xmm5 DB 102,15,111,252 ; movdqa %xmm4,%xmm7 DB 102,15,223,248 ; pandn %xmm0,%xmm7 @@ -46969,13 +47682,209 @@ _sk_darken_sse2_8bit LABEL PROC DB 102,15,111,206 ; movdqa %xmm6,%xmm1 DB 255,224 ; jmpq *%rax +PUBLIC _sk_difference_sse2_8bit +_sk_difference_sse2_8bit LABEL PROC + DB 242,15,112,227,231 ; pshuflw $0xe7,%xmm3,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,15,112,228,95 ; pshufhw $0x5f,%xmm4,%xmm4 + DB 242,15,112,234,231 ; pshuflw $0xe7,%xmm2,%xmm5 + DB 243,15,112,237,231 ; pshufhw $0xe7,%xmm5,%xmm5 + DB 102,15,112,237,232 ; pshufd $0xe8,%xmm5,%xmm5 + DB 102,15,96,237 ; punpcklbw %xmm5,%xmm5 + DB 242,15,112,237,95 ; pshuflw $0x5f,%xmm5,%xmm5 + DB 243,15,112,253,95 ; pshufhw $0x5f,%xmm5,%xmm7 + DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 + DB 102,68,15,111,200 ; movdqa %xmm0,%xmm9 + DB 102,69,15,96,200 ; punpcklbw %xmm8,%xmm9 + DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10 + DB 102,69,15,104,208 ; punpckhbw %xmm8,%xmm10 + DB 102,68,15,111,217 ; movdqa %xmm1,%xmm11 + DB 102,69,15,96,216 ; punpcklbw %xmm8,%xmm11 + DB 102,68,15,111,225 ; movdqa %xmm1,%xmm12 + DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 + DB 102,68,15,111,239 ; movdqa %xmm7,%xmm13 + DB 102,69,15,96,232 ; punpcklbw %xmm8,%xmm13 + DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 + DB 102,68,15,111,244 ; movdqa %xmm4,%xmm14 + DB 102,69,15,96,240 ; punpcklbw %xmm8,%xmm14 + DB 102,65,15,104,224 ; punpckhbw %xmm8,%xmm4 + DB 102,65,15,213,228 ; pmullw %xmm12,%xmm4 + DB 102,69,15,213,243 ; pmullw %xmm11,%xmm14 + DB 102,65,15,213,250 ; pmullw %xmm10,%xmm7 + DB 102,69,15,213,233 ; pmullw %xmm9,%xmm13 + DB 102,69,15,253,233 ; paddw %xmm9,%xmm13 + DB 102,65,15,253,250 ; paddw %xmm10,%xmm7 + DB 102,69,15,253,243 ; paddw %xmm11,%xmm14 + DB 102,65,15,253,228 ; paddw %xmm12,%xmm4 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,65,15,113,214,8 ; psrlw $0x8,%xmm14 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,113,213,8 ; psrlw $0x8,%xmm13 + DB 102,68,15,103,239 ; packuswb %xmm7,%xmm13 + DB 102,68,15,103,244 ; packuswb %xmm4,%xmm14 + DB 242,15,112,225,231 ; pshuflw $0xe7,%xmm1,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,68,15,112,212,95 ; pshufhw $0x5f,%xmm4,%xmm10 + DB 242,15,112,224,231 ; pshuflw $0xe7,%xmm0,%xmm4 + DB 243,15,112,228,231 ; pshufhw $0xe7,%xmm4,%xmm4 + DB 102,15,112,228,232 ; pshufd $0xe8,%xmm4,%xmm4 + DB 102,15,96,228 ; punpcklbw %xmm4,%xmm4 + DB 242,15,112,228,95 ; pshuflw $0x5f,%xmm4,%xmm4 + DB 243,68,15,112,220,95 ; pshufhw $0x5f,%xmm4,%xmm11 + DB 102,68,15,111,202 ; movdqa %xmm2,%xmm9 + DB 102,15,252,194 ; paddb %xmm2,%xmm0 + DB 102,15,111,242 ; movdqa %xmm2,%xmm6 + DB 102,65,15,96,240 ; punpcklbw %xmm8,%xmm6 + DB 102,69,15,104,200 ; punpckhbw %xmm8,%xmm9 + DB 102,68,15,111,227 ; movdqa %xmm3,%xmm12 + DB 102,15,252,203 ; paddb %xmm3,%xmm1 + DB 102,15,111,235 ; movdqa %xmm3,%xmm5 + DB 102,65,15,96,232 ; punpcklbw %xmm8,%xmm5 + DB 102,69,15,104,224 ; punpckhbw %xmm8,%xmm12 + DB 102,65,15,111,227 ; movdqa %xmm11,%xmm4 + DB 102,65,15,96,224 ; punpcklbw %xmm8,%xmm4 + DB 102,69,15,104,216 ; punpckhbw %xmm8,%xmm11 + DB 102,65,15,111,250 ; movdqa %xmm10,%xmm7 + DB 102,65,15,96,248 ; punpcklbw %xmm8,%xmm7 + DB 102,69,15,104,208 ; punpckhbw %xmm8,%xmm10 + DB 102,69,15,213,212 ; pmullw %xmm12,%xmm10 + DB 102,15,213,253 ; pmullw %xmm5,%xmm7 + DB 102,69,15,213,217 ; pmullw %xmm9,%xmm11 + DB 102,15,213,230 ; pmullw %xmm6,%xmm4 + DB 102,15,253,230 ; paddw %xmm6,%xmm4 + DB 102,69,15,253,217 ; paddw %xmm9,%xmm11 + DB 102,15,253,253 ; paddw %xmm5,%xmm7 + DB 102,69,15,253,212 ; paddw %xmm12,%xmm10 + DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,65,15,113,211,8 ; psrlw $0x8,%xmm11 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,65,15,103,227 ; packuswb %xmm11,%xmm4 + DB 102,65,15,103,250 ; packuswb %xmm10,%xmm7 + DB 102,68,15,218,247 ; pminub %xmm7,%xmm14 + DB 102,68,15,218,236 ; pminub %xmm4,%xmm13 + DB 102,69,15,252,237 ; paddb %xmm13,%xmm13 + DB 102,69,15,252,246 ; paddb %xmm14,%xmm14 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,65,15,248,246 ; psubb %xmm14,%xmm6 + DB 102,15,111,232 ; movdqa %xmm0,%xmm5 + DB 102,65,15,248,237 ; psubb %xmm13,%xmm5 + DB 102,15,248,207 ; psubb %xmm7,%xmm1 + DB 102,15,248,196 ; psubb %xmm4,%xmm0 + DB 102,15,111,37,241,2,0,0 ; movdqa 0x2f1(%rip),%xmm4 # 2c20 <_sk_exclusion_sse2_8bit+0x2c9> + DB 102,15,219,236 ; pand %xmm4,%xmm5 + DB 102,15,111,252 ; movdqa %xmm4,%xmm7 + DB 102,15,223,248 ; pandn %xmm0,%xmm7 + DB 102,15,235,239 ; por %xmm7,%xmm5 + DB 102,15,219,244 ; pand %xmm4,%xmm6 + DB 102,15,223,225 ; pandn %xmm1,%xmm4 + DB 102,15,235,244 ; por %xmm4,%xmm6 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 102,15,111,197 ; movdqa %xmm5,%xmm0 + DB 102,15,111,206 ; movdqa %xmm6,%xmm1 + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_exclusion_sse2_8bit +_sk_exclusion_sse2_8bit LABEL PROC + DB 102,69,15,239,201 ; pxor %xmm9,%xmm9 + DB 102,68,15,111,192 ; movdqa %xmm0,%xmm8 + DB 102,68,15,111,226 ; movdqa %xmm2,%xmm12 + DB 102,68,15,252,224 ; paddb %xmm0,%xmm12 + DB 242,68,15,112,208,231 ; pshuflw $0xe7,%xmm0,%xmm10 + DB 102,15,111,224 ; movdqa %xmm0,%xmm4 + DB 102,65,15,96,225 ; punpcklbw %xmm9,%xmm4 + DB 102,69,15,104,193 ; punpckhbw %xmm9,%xmm8 + DB 102,68,15,111,233 ; movdqa %xmm1,%xmm13 + DB 102,68,15,111,243 ; movdqa %xmm3,%xmm14 + DB 102,68,15,252,241 ; paddb %xmm1,%xmm14 + DB 242,68,15,112,217,231 ; pshuflw $0xe7,%xmm1,%xmm11 + DB 102,68,15,111,249 ; movdqa %xmm1,%xmm15 + DB 102,69,15,96,249 ; punpcklbw %xmm9,%xmm15 + DB 102,15,111,234 ; movdqa %xmm2,%xmm5 + DB 102,15,111,250 ; movdqa %xmm2,%xmm7 + DB 102,65,15,96,249 ; punpcklbw %xmm9,%xmm7 + DB 102,65,15,104,233 ; punpckhbw %xmm9,%xmm5 + DB 243,65,15,112,194,231 ; pshufhw $0xe7,%xmm10,%xmm0 + DB 102,15,112,192,232 ; pshufd $0xe8,%xmm0,%xmm0 + DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 + DB 242,15,112,192,95 ; pshuflw $0x5f,%xmm0,%xmm0 + DB 243,68,15,112,208,95 ; pshufhw $0x5f,%xmm0,%xmm10 + DB 102,65,15,111,194 ; movdqa %xmm10,%xmm0 + DB 102,65,15,96,193 ; punpcklbw %xmm9,%xmm0 + DB 102,15,213,199 ; pmullw %xmm7,%xmm0 + DB 102,15,253,199 ; paddw %xmm7,%xmm0 + DB 102,15,213,252 ; pmullw %xmm4,%xmm7 + DB 102,15,253,252 ; paddw %xmm4,%xmm7 + DB 102,15,111,227 ; movdqa %xmm3,%xmm4 + DB 102,65,15,96,225 ; punpcklbw %xmm9,%xmm4 + DB 102,69,15,104,209 ; punpckhbw %xmm9,%xmm10 + DB 102,68,15,213,213 ; pmullw %xmm5,%xmm10 + DB 102,68,15,253,213 ; paddw %xmm5,%xmm10 + DB 102,65,15,213,232 ; pmullw %xmm8,%xmm5 + DB 102,65,15,253,232 ; paddw %xmm8,%xmm5 + DB 243,65,15,112,243,231 ; pshufhw $0xe7,%xmm11,%xmm6 + DB 102,15,112,246,232 ; pshufd $0xe8,%xmm6,%xmm6 + DB 102,15,96,246 ; punpcklbw %xmm6,%xmm6 + DB 242,15,112,246,95 ; pshuflw $0x5f,%xmm6,%xmm6 + DB 243,15,112,206,95 ; pshufhw $0x5f,%xmm6,%xmm1 + DB 102,15,111,241 ; movdqa %xmm1,%xmm6 + DB 102,65,15,96,241 ; punpcklbw %xmm9,%xmm6 + DB 102,15,213,244 ; pmullw %xmm4,%xmm6 + DB 102,15,253,244 ; paddw %xmm4,%xmm6 + DB 102,65,15,213,231 ; pmullw %xmm15,%xmm4 + DB 102,65,15,253,231 ; paddw %xmm15,%xmm4 + DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 + DB 102,69,15,104,233 ; punpckhbw %xmm9,%xmm13 + DB 102,69,15,104,193 ; punpckhbw %xmm9,%xmm8 + DB 102,65,15,104,201 ; punpckhbw %xmm9,%xmm1 + DB 102,65,15,213,200 ; pmullw %xmm8,%xmm1 + DB 102,65,15,253,200 ; paddw %xmm8,%xmm1 + DB 102,69,15,213,197 ; pmullw %xmm13,%xmm8 + DB 102,69,15,253,197 ; paddw %xmm13,%xmm8 + DB 102,15,113,213,8 ; psrlw $0x8,%xmm5 + DB 102,15,113,215,8 ; psrlw $0x8,%xmm7 + DB 102,15,103,253 ; packuswb %xmm5,%xmm7 + DB 102,65,15,113,208,8 ; psrlw $0x8,%xmm8 + DB 102,15,113,212,8 ; psrlw $0x8,%xmm4 + DB 102,65,15,103,224 ; packuswb %xmm8,%xmm4 + DB 102,15,252,255 ; paddb %xmm7,%xmm7 + DB 102,15,252,228 ; paddb %xmm4,%xmm4 + DB 102,65,15,111,238 ; movdqa %xmm14,%xmm5 + DB 102,15,248,236 ; psubb %xmm4,%xmm5 + DB 102,65,15,111,228 ; movdqa %xmm12,%xmm4 + DB 102,15,248,231 ; psubb %xmm7,%xmm4 + DB 102,65,15,113,210,8 ; psrlw $0x8,%xmm10 + DB 102,15,113,208,8 ; psrlw $0x8,%xmm0 + DB 102,65,15,103,194 ; packuswb %xmm10,%xmm0 + DB 102,15,113,209,8 ; psrlw $0x8,%xmm1 + DB 102,15,113,214,8 ; psrlw $0x8,%xmm6 + DB 102,15,103,241 ; packuswb %xmm1,%xmm6 + DB 102,68,15,248,246 ; psubb %xmm6,%xmm14 + DB 102,68,15,248,224 ; psubb %xmm0,%xmm12 + DB 102,15,111,13,102,1,0,0 ; movdqa 0x166(%rip),%xmm1 # 2c30 <_sk_exclusion_sse2_8bit+0x2d9> + DB 102,15,111,193 ; movdqa %xmm1,%xmm0 + DB 102,65,15,223,196 ; pandn %xmm12,%xmm0 + DB 102,15,219,225 ; pand %xmm1,%xmm4 + DB 102,15,235,196 ; por %xmm4,%xmm0 + DB 102,15,219,233 ; pand %xmm1,%xmm5 + DB 102,65,15,223,206 ; pandn %xmm14,%xmm1 + DB 102,15,235,205 ; por %xmm5,%xmm1 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + ALIGN 4 DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 25bb <_sk_darken_sse2_8bit+0x233> + DB 127,67 ; jg 2b33 <_sk_exclusion_sse2_8bit+0x1dc> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 25bf <_sk_darken_sse2_8bit+0x237> + DB 127,67 ; jg 2b37 <_sk_exclusion_sse2_8bit+0x1e0> DB 0,0 ; add %al,(%rax) - DB 127,67 ; jg 25c3 <_sk_darken_sse2_8bit+0x23b> + DB 127,67 ; jg 2b3b <_sk_exclusion_sse2_8bit+0x1e4> ALIGN 16 DB 0,0 ; add %al,(%rax) @@ -47118,6 +48027,42 @@ ALIGN 16 DB 255 ; (bad) DB 255 ; (bad) DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,0 ; incl (%rax) ELSE .MODEL FLAT,C _text32 SEGMENT ALIGN(32) 'CODE' diff --git a/src/jumper/SkJumper_stages_8bit.cpp b/src/jumper/SkJumper_stages_8bit.cpp index bceb872747..cf037b0209 100644 --- a/src/jumper/SkJumper_stages_8bit.cpp +++ b/src/jumper/SkJumper_stages_8bit.cpp @@ -110,6 +110,7 @@ SI V operator*(V x, V y) { } SI V inv(V v) { return 0xff - v; } +SI V two(V v) { return v + v; } SI V lerp(V from, V to, V t) { return to*t + from*inv(t); } SI V alpha(V v) { @@ -145,6 +146,11 @@ SI V max(V a, V b) { return (a.u8x4 & gt) | (b.u8x4 &~gt); } +SI V min(V a, V b) { + auto gt = a.u8x4 > b.u8x4; + return (a.u8x4 & ~gt) | (b.u8x4 >); +} + struct Params { size_t x,y,tail; }; @@ -407,8 +413,22 @@ STAGE(multiply) { src = src*inv(alpha(dst)) + dst*inv(alpha(src)) + src*dst; } STAGE(screen) { src = src + inv(src)*dst; } STAGE(xor_) { src = src*inv(alpha(dst)) + dst*inv(alpha(src)); } -STAGE(darken) { - V rgb = src + (dst - max(src*alpha(dst), dst*alpha(src))); - V a = src + (dst - dst*alpha(src)); - src = (rgb.u32 & 0x00ffffff) | (a.u32 & 0xff000000); +SI V srcover_alpha(V src, V dst, V rgb) { + V a = src + (dst - dst*alpha(src)); + return (rgb.u32 & 0x00ffffff) | (a.u32 & 0xff000000); } + +STAGE(darken) { src = srcover_alpha(src, dst, src + (dst - max(src*alpha(dst), dst*alpha(src)))); } +STAGE(lighten) { src = srcover_alpha(src, dst, src + (dst - min(src*alpha(dst), dst*alpha(src)))); } + +STAGE(difference) { + V min_ = min(src*alpha(dst), dst*alpha(src)); + src = srcover_alpha(src, dst, (src - min_) + (dst - min_)); +} + +STAGE(exclusion) { + V sd = src*dst; + src = srcover_alpha(src, dst, (src - sd) + (dst - sd)); +} + +#undef BLEND_MODE |