From 8e200787b769ba9ee0bd4bd37d27ded1fafc06e6 Mon Sep 17 00:00:00 2001 From: Mike Klein Date: Mon, 5 Jun 2017 19:45:15 -0400 Subject: [PATCH] more easy lowp stages Change-Id: I8a292bc98135b41ceedb4242451436c3657616fc Reviewed-on: https://skia-review.googlesource.com/18722 Reviewed-by: Mike Reed Commit-Queue: Mike Klein --- src/jumper/SkJumper.cpp | 5 + src/jumper/SkJumper_generated.S | 337 ++++++++++++++++++++++++---- src/jumper/SkJumper_generated_win.S | 333 +++++++++++++++++++++++---- src/jumper/SkJumper_stages_lowp.cpp | 34 +++ 4 files changed, 611 insertions(+), 98 deletions(-) diff --git a/src/jumper/SkJumper.cpp b/src/jumper/SkJumper.cpp index f536f4c741..82248001f8 100644 --- a/src/jumper/SkJumper.cpp +++ b/src/jumper/SkJumper.cpp @@ -72,8 +72,13 @@ using StartPipelineFn = void(size_t,size_t,size_t,void**,K*); // Some stages have low-precision (~15 bit) versions from SkJumper_stages_lowp.cpp. #define LOWP_STAGES(M) \ M(constant_color) \ + M(set_rgb) \ + M(premul) \ M(load_8888) \ M(store_8888) \ + M(load_a8) \ + M(store_a8) \ + M(load_g8) \ M(srcover_rgba_8888) \ M(lerp_1_float) \ M(lerp_u8) \ diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S index 286d348e61..7328536fa0 100644 --- a/src/jumper/SkJumper_generated.S +++ b/src/jumper/SkJumper_generated.S @@ -36805,7 +36805,7 @@ HIDDEN _sk_constant_color_ssse3_lowp FUNCTION(_sk_constant_color_ssse3_lowp) _sk_constant_color_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax - .byte 243,15,16,29,34,14,0,0 // movss 0xe22(%rip),%xmm3 # ee0 <_sk_xor__ssse3_lowp+0x9b> + .byte 243,15,16,29,26,17,0,0 // movss 0x111a(%rip),%xmm3 # 11d8 <_sk_xor__ssse3_lowp+0x9b> .byte 243,15,16,0 // movss (%rax),%xmm0 .byte 243,15,89,195 // mulss %xmm3,%xmm0 .byte 243,68,15,44,200 // cvttss2si %xmm0,%r9d @@ -36832,6 +36832,45 @@ _sk_constant_color_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 255,224 // jmpq *%rax +HIDDEN _sk_set_rgb_ssse3_lowp +.globl _sk_set_rgb_ssse3_lowp +FUNCTION(_sk_set_rgb_ssse3_lowp) +_sk_set_rgb_ssse3_lowp: + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 243,15,16,21,163,16,0,0 // movss 0x10a3(%rip),%xmm2 # 11dc <_sk_xor__ssse3_lowp+0x9f> + .byte 243,15,16,0 // movss (%rax),%xmm0 + .byte 243,15,89,194 // mulss %xmm2,%xmm0 + .byte 243,68,15,44,200 // cvttss2si %xmm0,%r9d + .byte 102,65,15,110,193 // movd %r9d,%xmm0 + .byte 242,15,112,192,0 // pshuflw $0x0,%xmm0,%xmm0 + .byte 102,15,112,192,80 // pshufd $0x50,%xmm0,%xmm0 + .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 + .byte 243,15,89,202 // mulss %xmm2,%xmm1 + .byte 243,68,15,44,201 // cvttss2si %xmm1,%r9d + .byte 102,65,15,110,201 // movd %r9d,%xmm1 + .byte 242,15,112,201,0 // pshuflw $0x0,%xmm1,%xmm1 + .byte 102,15,112,201,80 // pshufd $0x50,%xmm1,%xmm1 + .byte 243,15,89,80,8 // mulss 0x8(%rax),%xmm2 + .byte 243,15,44,194 // cvttss2si %xmm2,%eax + .byte 102,15,110,208 // movd %eax,%xmm2 + .byte 242,15,112,210,0 // pshuflw $0x0,%xmm2,%xmm2 + .byte 102,15,112,210,80 // pshufd $0x50,%xmm2,%xmm2 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + +HIDDEN _sk_premul_ssse3_lowp +.globl _sk_premul_ssse3_lowp +FUNCTION(_sk_premul_ssse3_lowp) +_sk_premul_ssse3_lowp: + .byte 102,15,56,11,195 // pmulhrsw %xmm3,%xmm0 + .byte 102,15,56,29,192 // pabsw %xmm0,%xmm0 + .byte 102,15,56,11,203 // pmulhrsw %xmm3,%xmm1 + .byte 102,15,56,29,201 // pabsw %xmm1,%xmm1 + .byte 102,15,56,11,211 // pmulhrsw %xmm3,%xmm2 + .byte 102,15,56,29,210 // pabsw %xmm2,%xmm2 + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + HIDDEN _sk_load_8888_ssse3_lowp .globl _sk_load_8888_ssse3_lowp FUNCTION(_sk_load_8888_ssse3_lowp) @@ -36839,10 +36878,10 @@ _sk_load_8888_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 .byte 77,133,192 // test %r8,%r8 - .byte 117,113 // jne 1aa <_sk_load_8888_ssse3_lowp+0x7b> + .byte 117,113 // jne 22a <_sk_load_8888_ssse3_lowp+0x7b> .byte 69,15,16,76,147,16 // movups 0x10(%r11,%rdx,4),%xmm9 .byte 69,15,16,4,147 // movups (%r11,%rdx,4),%xmm8 - .byte 102,15,111,5,164,13,0,0 // movdqa 0xda4(%rip),%xmm0 # ef0 <_sk_xor__ssse3_lowp+0xab> + .byte 102,15,111,5,36,16,0,0 // movdqa 0x1024(%rip),%xmm0 # 11f0 <_sk_xor__ssse3_lowp+0xb3> .byte 102,68,15,56,0,192 // pshufb %xmm0,%xmm8 .byte 102,68,15,56,0,200 // pshufb %xmm0,%xmm9 .byte 102,65,15,111,208 // movdqa %xmm8,%xmm2 @@ -36856,7 +36895,7 @@ _sk_load_8888_ssse3_lowp: .byte 102,15,239,210 // pxor %xmm2,%xmm2 .byte 102,65,15,96,208 // punpcklbw %xmm8,%xmm2 .byte 102,65,15,104,216 // punpckhbw %xmm8,%xmm3 - .byte 102,68,15,111,5,110,13,0,0 // movdqa 0xd6e(%rip),%xmm8 # f00 <_sk_xor__ssse3_lowp+0xbb> + .byte 102,68,15,111,5,238,15,0,0 // movdqa 0xfee(%rip),%xmm8 # 1200 <_sk_xor__ssse3_lowp+0xc3> .byte 102,65,15,228,192 // pmulhuw %xmm8,%xmm0 .byte 102,65,15,228,200 // pmulhuw %xmm8,%xmm1 .byte 102,65,15,228,208 // pmulhuw %xmm8,%xmm2 @@ -36869,9 +36908,9 @@ _sk_load_8888_ssse3_lowp: .byte 69,15,87,192 // xorps %xmm8,%xmm8 .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 119,129 // ja 144 <_sk_load_8888_ssse3_lowp+0x15> + .byte 119,129 // ja 1c4 <_sk_load_8888_ssse3_lowp+0x15> .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,130,0,0,0 // lea 0x82(%rip),%r10 # 250 <_sk_load_8888_ssse3_lowp+0x121> + .byte 76,141,21,130,0,0,0 // lea 0x82(%rip),%r10 # 2d0 <_sk_load_8888_ssse3_lowp+0x121> .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -36896,7 +36935,7 @@ _sk_load_8888_ssse3_lowp: .byte 68,15,40,192 // movaps %xmm0,%xmm8 .byte 243,65,15,16,4,147 // movss (%r11,%rdx,4),%xmm0 .byte 243,68,15,16,192 // movss %xmm0,%xmm8 - .byte 233,244,254,255,255 // jmpq 144 <_sk_load_8888_ssse3_lowp+0x15> + .byte 233,244,254,255,255 // jmpq 1c4 <_sk_load_8888_ssse3_lowp+0x15> .byte 240,255 // lock (bad) .byte 255 // (bad) .byte 255 // (bad) @@ -36937,7 +36976,7 @@ _sk_store_8888_ssse3_lowp: .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 .byte 102,69,15,105,202 // punpckhwd %xmm10,%xmm9 .byte 77,133,192 // test %r8,%r8 - .byte 117,17 // jne 2db <_sk_store_8888_ssse3_lowp+0x6f> + .byte 117,17 // jne 35b <_sk_store_8888_ssse3_lowp+0x6f> .byte 243,69,15,127,76,147,16 // movdqu %xmm9,0x10(%r11,%rdx,4) .byte 243,69,15,127,4,147 // movdqu %xmm8,(%r11,%rdx,4) .byte 72,173 // lods %ds:(%rsi),%rax @@ -36946,9 +36985,9 @@ _sk_store_8888_ssse3_lowp: .byte 65,128,225,7 // and $0x7,%r9b .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 119,236 // ja 2d7 <_sk_store_8888_ssse3_lowp+0x6b> + .byte 119,236 // ja 357 <_sk_store_8888_ssse3_lowp+0x6b> .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,90,0,0,0 // lea 0x5a(%rip),%r10 # 350 <_sk_store_8888_ssse3_lowp+0xe4> + .byte 76,141,21,90,0,0,0 // lea 0x5a(%rip),%r10 # 3d0 <_sk_store_8888_ssse3_lowp+0xe4> .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -36964,7 +37003,7 @@ _sk_store_8888_ssse3_lowp: .byte 102,69,15,112,200,229 // pshufd $0xe5,%xmm8,%xmm9 .byte 102,69,15,126,76,147,4 // movd %xmm9,0x4(%r11,%rdx,4) .byte 102,69,15,126,4,147 // movd %xmm8,(%r11,%rdx,4) - .byte 235,136 // jmp 2d7 <_sk_store_8888_ssse3_lowp+0x6b> + .byte 235,136 // jmp 357 <_sk_store_8888_ssse3_lowp+0x6b> .byte 144 // nop .byte 247,255 // idiv %edi .byte 255 // (bad) @@ -36987,6 +37026,199 @@ _sk_store_8888_ssse3_lowp: .byte 255 // (bad) .byte 255 // .byte 0xff +HIDDEN _sk_load_a8_ssse3_lowp +.globl _sk_load_a8_ssse3_lowp +FUNCTION(_sk_load_a8_ssse3_lowp) +_sk_load_a8_ssse3_lowp: + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,139,24 // mov (%rax),%r11 + .byte 77,133,192 // test %r8,%r8 + .byte 117,37 // jne 41b <_sk_load_a8_ssse3_lowp+0x2f> + .byte 243,65,15,126,28,19 // movq (%r11,%rdx,1),%xmm3 + .byte 102,15,96,216 // punpcklbw %xmm0,%xmm3 + .byte 102,15,113,243,8 // psllw $0x8,%xmm3 + .byte 102,15,228,29,3,14,0,0 // pmulhuw 0xe03(%rip),%xmm3 # 1210 <_sk_xor__ssse3_lowp+0xd3> + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 102,15,239,192 // pxor %xmm0,%xmm0 + .byte 15,87,201 // xorps %xmm1,%xmm1 + .byte 15,87,210 // xorps %xmm2,%xmm2 + .byte 255,224 // jmpq *%rax + .byte 69,137,193 // mov %r8d,%r9d + .byte 65,128,225,7 // and $0x7,%r9b + .byte 102,15,239,219 // pxor %xmm3,%xmm3 + .byte 65,254,201 // dec %r9b + .byte 65,128,249,6 // cmp $0x6,%r9b + .byte 119,209 // ja 400 <_sk_load_a8_ssse3_lowp+0x14> + .byte 69,15,182,201 // movzbl %r9b,%r9d + .byte 76,141,21,94,0,0,0 // lea 0x5e(%rip),%r10 # 498 <_sk_load_a8_ssse3_lowp+0xac> + .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax + .byte 76,1,208 // add %r10,%rax + .byte 255,224 // jmpq *%rax + .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax + .byte 102,15,239,219 // pxor %xmm3,%xmm3 + .byte 102,15,196,216,6 // pinsrw $0x6,%eax,%xmm3 + .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax + .byte 102,15,196,216,5 // pinsrw $0x5,%eax,%xmm3 + .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax + .byte 102,15,196,216,4 // pinsrw $0x4,%eax,%xmm3 + .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax + .byte 102,15,196,216,3 // pinsrw $0x3,%eax,%xmm3 + .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax + .byte 102,15,196,216,2 // pinsrw $0x2,%eax,%xmm3 + .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax + .byte 102,15,196,216,1 // pinsrw $0x1,%eax,%xmm3 + .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax + .byte 102,15,196,216,0 // pinsrw $0x0,%eax,%xmm3 + .byte 233,104,255,255,255 // jmpq 400 <_sk_load_a8_ssse3_lowp+0x14> + .byte 241 // icebp + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,230 // jmpq *%rsi + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 219,255 // (bad) + .byte 255 // (bad) + .byte 255,208 // callq *%rax + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,197 // inc %ebp + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 186,255,255,255,171 // mov $0xabffffff,%edx + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // .byte 0xff + +HIDDEN _sk_store_a8_ssse3_lowp +.globl _sk_store_a8_ssse3_lowp +FUNCTION(_sk_store_a8_ssse3_lowp) +_sk_store_a8_ssse3_lowp: + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,139,24 // mov (%rax),%r11 + .byte 102,68,15,111,195 // movdqa %xmm3,%xmm8 + .byte 102,65,15,113,208,7 // psrlw $0x7,%xmm8 + .byte 102,69,15,103,192 // packuswb %xmm8,%xmm8 + .byte 77,133,192 // test %r8,%r8 + .byte 117,10 // jne 4d8 <_sk_store_a8_ssse3_lowp+0x24> + .byte 242,69,15,17,4,19 // movsd %xmm8,(%r11,%rdx,1) + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 255,224 // jmpq *%rax + .byte 69,137,193 // mov %r8d,%r9d + .byte 65,128,225,7 // and $0x7,%r9b + .byte 65,254,201 // dec %r9b + .byte 65,128,249,6 // cmp $0x6,%r9b + .byte 119,236 // ja 4d4 <_sk_store_a8_ssse3_lowp+0x20> + .byte 102,68,15,96,192 // punpcklbw %xmm0,%xmm8 + .byte 69,15,182,201 // movzbl %r9b,%r9d + .byte 76,141,21,128,0,0,0 // lea 0x80(%rip),%r10 # 578 <_sk_store_a8_ssse3_lowp+0xc4> + .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax + .byte 76,1,208 // add %r10,%rax + .byte 255,224 // jmpq *%rax + .byte 102,68,15,127,68,36,232 // movdqa %xmm8,-0x18(%rsp) + .byte 138,68,36,244 // mov -0xc(%rsp),%al + .byte 65,136,68,19,6 // mov %al,0x6(%r11,%rdx,1) + .byte 102,68,15,127,68,36,216 // movdqa %xmm8,-0x28(%rsp) + .byte 138,68,36,226 // mov -0x1e(%rsp),%al + .byte 65,136,68,19,5 // mov %al,0x5(%r11,%rdx,1) + .byte 102,68,15,127,68,36,200 // movdqa %xmm8,-0x38(%rsp) + .byte 138,68,36,208 // mov -0x30(%rsp),%al + .byte 65,136,68,19,4 // mov %al,0x4(%r11,%rdx,1) + .byte 102,68,15,127,68,36,184 // movdqa %xmm8,-0x48(%rsp) + .byte 138,68,36,190 // mov -0x42(%rsp),%al + .byte 65,136,68,19,3 // mov %al,0x3(%r11,%rdx,1) + .byte 102,68,15,127,68,36,168 // movdqa %xmm8,-0x58(%rsp) + .byte 138,68,36,172 // mov -0x54(%rsp),%al + .byte 65,136,68,19,2 // mov %al,0x2(%r11,%rdx,1) + .byte 102,68,15,127,68,36,152 // movdqa %xmm8,-0x68(%rsp) + .byte 138,68,36,154 // mov -0x66(%rsp),%al + .byte 65,136,68,19,1 // mov %al,0x1(%r11,%rdx,1) + .byte 102,68,15,127,68,36,136 // movdqa %xmm8,-0x78(%rsp) + .byte 138,68,36,136 // mov -0x78(%rsp),%al + .byte 65,136,4,19 // mov %al,(%r11,%rdx,1) + .byte 233,95,255,255,255 // jmpq 4d4 <_sk_store_a8_ssse3_lowp+0x20> + .byte 15,31,0 // nopl (%rax) + .byte 233,255,255,255,217 // jmpq ffffffffda00057c <_sk_xor__ssse3_lowp+0xffffffffd9fff43f> + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,201 // dec %ecx + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 185,255,255,255,169 // mov $0xa9ffffff,%ecx + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,153,255,255,255,137 // lcall *-0x76000001(%rcx) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // .byte 0xff + +HIDDEN _sk_load_g8_ssse3_lowp +.globl _sk_load_g8_ssse3_lowp +FUNCTION(_sk_load_g8_ssse3_lowp) +_sk_load_g8_ssse3_lowp: + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 76,139,24 // mov (%rax),%r11 + .byte 77,133,192 // test %r8,%r8 + .byte 117,42 // jne 5c8 <_sk_load_g8_ssse3_lowp+0x34> + .byte 243,65,15,126,4,19 // movq (%r11,%rdx,1),%xmm0 + .byte 102,15,96,192 // punpcklbw %xmm0,%xmm0 + .byte 102,15,113,240,8 // psllw $0x8,%xmm0 + .byte 102,15,228,5,107,12,0,0 // pmulhuw 0xc6b(%rip),%xmm0 # 1220 <_sk_xor__ssse3_lowp+0xe3> + .byte 72,173 // lods %ds:(%rsi),%rax + .byte 15,40,29,114,12,0,0 // movaps 0xc72(%rip),%xmm3 # 1230 <_sk_xor__ssse3_lowp+0xf3> + .byte 102,15,111,200 // movdqa %xmm0,%xmm1 + .byte 102,15,111,208 // movdqa %xmm0,%xmm2 + .byte 255,224 // jmpq *%rax + .byte 69,137,193 // mov %r8d,%r9d + .byte 65,128,225,7 // and $0x7,%r9b + .byte 102,15,239,192 // pxor %xmm0,%xmm0 + .byte 65,254,201 // dec %r9b + .byte 65,128,249,6 // cmp $0x6,%r9b + .byte 119,204 // ja 5a8 <_sk_load_g8_ssse3_lowp+0x14> + .byte 69,15,182,201 // movzbl %r9b,%r9d + .byte 76,141,21,97,0,0,0 // lea 0x61(%rip),%r10 # 648 <_sk_load_g8_ssse3_lowp+0xb4> + .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax + .byte 76,1,208 // add %r10,%rax + .byte 255,224 // jmpq *%rax + .byte 65,15,182,68,19,6 // movzbl 0x6(%r11,%rdx,1),%eax + .byte 102,15,239,192 // pxor %xmm0,%xmm0 + .byte 102,15,196,192,6 // pinsrw $0x6,%eax,%xmm0 + .byte 65,15,182,68,19,5 // movzbl 0x5(%r11,%rdx,1),%eax + .byte 102,15,196,192,5 // pinsrw $0x5,%eax,%xmm0 + .byte 65,15,182,68,19,4 // movzbl 0x4(%r11,%rdx,1),%eax + .byte 102,15,196,192,4 // pinsrw $0x4,%eax,%xmm0 + .byte 65,15,182,68,19,3 // movzbl 0x3(%r11,%rdx,1),%eax + .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0 + .byte 65,15,182,68,19,2 // movzbl 0x2(%r11,%rdx,1),%eax + .byte 102,15,196,192,2 // pinsrw $0x2,%eax,%xmm0 + .byte 65,15,182,68,19,1 // movzbl 0x1(%r11,%rdx,1),%eax + .byte 102,15,196,192,1 // pinsrw $0x1,%eax,%xmm0 + .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax + .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0 + .byte 233,99,255,255,255 // jmpq 5a8 <_sk_load_g8_ssse3_lowp+0x14> + .byte 15,31,0 // nopl (%rax) + .byte 238 // out %al,(%dx) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,227 // jmpq *%rbx + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 216,255 // fdivr %st(7),%st + .byte 255 // (bad) + .byte 255,205 // dec %ebp + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,194 // inc %edx + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255,183,255,255,255,168 // pushq -0x57000001(%rdi) + .byte 255 // (bad) + .byte 255 // (bad) + .byte 255 // .byte 0xff + HIDDEN _sk_srcover_rgba_8888_ssse3_lowp .globl _sk_srcover_rgba_8888_ssse3_lowp FUNCTION(_sk_srcover_rgba_8888_ssse3_lowp) @@ -36994,11 +37226,11 @@ _sk_srcover_rgba_8888_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 .byte 77,133,192 // test %r8,%r8 - .byte 15,133,66,1,0,0 // jne 4bc <_sk_srcover_rgba_8888_ssse3_lowp+0x150> + .byte 15,133,66,1,0,0 // jne 7b4 <_sk_srcover_rgba_8888_ssse3_lowp+0x150> .byte 69,15,16,76,147,16 // movups 0x10(%r11,%rdx,4),%xmm9 .byte 69,15,16,4,147 // movups (%r11,%rdx,4),%xmm8 .byte 77,133,192 // test %r8,%r8 - .byte 102,15,111,37,128,11,0,0 // movdqa 0xb80(%rip),%xmm4 # f10 <_sk_xor__ssse3_lowp+0xcb> + .byte 102,15,111,37,184,11,0,0 // movdqa 0xbb8(%rip),%xmm4 # 1240 <_sk_xor__ssse3_lowp+0x103> .byte 102,68,15,56,0,196 // pshufb %xmm4,%xmm8 .byte 102,68,15,56,0,204 // pshufb %xmm4,%xmm9 .byte 102,65,15,111,240 // movdqa %xmm8,%xmm6 @@ -37012,12 +37244,12 @@ _sk_srcover_rgba_8888_ssse3_lowp: .byte 102,15,239,246 // pxor %xmm6,%xmm6 .byte 102,65,15,96,240 // punpcklbw %xmm8,%xmm6 .byte 102,65,15,104,248 // punpckhbw %xmm8,%xmm7 - .byte 102,68,15,111,5,74,11,0,0 // movdqa 0xb4a(%rip),%xmm8 # f20 <_sk_xor__ssse3_lowp+0xdb> + .byte 102,68,15,111,5,130,11,0,0 // movdqa 0xb82(%rip),%xmm8 # 1250 <_sk_xor__ssse3_lowp+0x113> .byte 102,65,15,228,224 // pmulhuw %xmm8,%xmm4 .byte 102,65,15,228,232 // pmulhuw %xmm8,%xmm5 .byte 102,65,15,228,240 // pmulhuw %xmm8,%xmm6 .byte 102,65,15,228,248 // pmulhuw %xmm8,%xmm7 - .byte 102,68,15,111,29,61,11,0,0 // movdqa 0xb3d(%rip),%xmm11 # f30 <_sk_xor__ssse3_lowp+0xeb> + .byte 102,68,15,111,29,117,11,0,0 // movdqa 0xb75(%rip),%xmm11 # 1260 <_sk_xor__ssse3_lowp+0x123> .byte 102,68,15,249,219 // psubw %xmm3,%xmm11 .byte 102,68,15,111,196 // movdqa %xmm4,%xmm8 .byte 102,69,15,56,11,195 // pmulhrsw %xmm11,%xmm8 @@ -37050,7 +37282,7 @@ _sk_srcover_rgba_8888_ssse3_lowp: .byte 102,15,111,193 // movdqa %xmm1,%xmm0 .byte 102,15,97,194 // punpcklwd %xmm2,%xmm0 .byte 102,15,105,202 // punpckhwd %xmm2,%xmm1 - .byte 15,133,207,0,0,0 // jne 566 <_sk_srcover_rgba_8888_ssse3_lowp+0x1fa> + .byte 15,133,207,0,0,0 // jne 85e <_sk_srcover_rgba_8888_ssse3_lowp+0x1fa> .byte 243,65,15,127,76,147,16 // movdqu %xmm1,0x10(%r11,%rdx,4) .byte 243,65,15,127,4,147 // movdqu %xmm0,(%r11,%rdx,4) .byte 72,173 // lods %ds:(%rsi),%rax @@ -37065,9 +37297,9 @@ _sk_srcover_rgba_8888_ssse3_lowp: .byte 69,15,87,192 // xorps %xmm8,%xmm8 .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 15,135,172,254,255,255 // ja 385 <_sk_srcover_rgba_8888_ssse3_lowp+0x19> + .byte 15,135,172,254,255,255 // ja 67d <_sk_srcover_rgba_8888_ssse3_lowp+0x19> .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,248,0,0,0 // lea 0xf8(%rip),%r10 # 5dc <_sk_srcover_rgba_8888_ssse3_lowp+0x270> + .byte 76,141,21,248,0,0,0 // lea 0xf8(%rip),%r10 # 8d4 <_sk_srcover_rgba_8888_ssse3_lowp+0x270> .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -37092,14 +37324,14 @@ _sk_srcover_rgba_8888_ssse3_lowp: .byte 68,15,40,196 // movaps %xmm4,%xmm8 .byte 243,65,15,16,36,147 // movss (%r11,%rdx,4),%xmm4 .byte 243,68,15,16,196 // movss %xmm4,%xmm8 - .byte 233,31,254,255,255 // jmpq 385 <_sk_srcover_rgba_8888_ssse3_lowp+0x19> + .byte 233,31,254,255,255 // jmpq 67d <_sk_srcover_rgba_8888_ssse3_lowp+0x19> .byte 69,137,193 // mov %r8d,%r9d .byte 65,128,225,7 // and $0x7,%r9b .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 15,135,42,255,255,255 // ja 4a4 <_sk_srcover_rgba_8888_ssse3_lowp+0x138> + .byte 15,135,42,255,255,255 // ja 79c <_sk_srcover_rgba_8888_ssse3_lowp+0x138> .byte 65,15,182,193 // movzbl %r9b,%eax - .byte 76,141,13,115,0,0,0 // lea 0x73(%rip),%r9 # 5f8 <_sk_srcover_rgba_8888_ssse3_lowp+0x28c> + .byte 76,141,13,115,0,0,0 // lea 0x73(%rip),%r9 # 8f0 <_sk_srcover_rgba_8888_ssse3_lowp+0x28c> .byte 73,99,4,129 // movslq (%r9,%rax,4),%rax .byte 76,1,200 // add %r9,%rax .byte 255,224 // jmpq *%rax @@ -37115,8 +37347,8 @@ _sk_srcover_rgba_8888_ssse3_lowp: .byte 102,15,112,200,229 // pshufd $0xe5,%xmm0,%xmm1 .byte 102,65,15,126,76,147,4 // movd %xmm1,0x4(%r11,%rdx,4) .byte 102,65,15,126,4,147 // movd %xmm0,(%r11,%rdx,4) - .byte 233,200,254,255,255 // jmpq 4a4 <_sk_srcover_rgba_8888_ssse3_lowp+0x138> - .byte 122,255 // jp 5dd <_sk_srcover_rgba_8888_ssse3_lowp+0x271> + .byte 233,200,254,255,255 // jmpq 79c <_sk_srcover_rgba_8888_ssse3_lowp+0x138> + .byte 122,255 // jp 8d5 <_sk_srcover_rgba_8888_ssse3_lowp+0x271> .byte 255 // (bad) .byte 255,101,255 // jmpq *-0x1(%rbp) .byte 255 // (bad) @@ -37155,7 +37387,7 @@ FUNCTION(_sk_scale_1_float_ssse3_lowp) _sk_scale_1_float_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 243,68,15,89,5,192,8,0,0 // mulss 0x8c0(%rip),%xmm8 # ee4 <_sk_xor__ssse3_lowp+0x9f> + .byte 243,68,15,89,5,196,8,0,0 // mulss 0x8c4(%rip),%xmm8 # 11e0 <_sk_xor__ssse3_lowp+0xa3> .byte 243,65,15,44,192 // cvttss2si %xmm8,%eax .byte 102,68,15,110,192 // movd %eax,%xmm8 .byte 242,69,15,112,192,0 // pshuflw $0x0,%xmm8,%xmm8 @@ -37178,11 +37410,11 @@ _sk_scale_u8_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 .byte 77,133,192 // test %r8,%r8 - .byte 117,74 // jne 6be <_sk_scale_u8_ssse3_lowp+0x54> + .byte 117,74 // jne 9b6 <_sk_scale_u8_ssse3_lowp+0x54> .byte 243,69,15,126,4,19 // movq (%r11,%rdx,1),%xmm8 .byte 102,68,15,96,192 // punpcklbw %xmm0,%xmm8 .byte 102,65,15,113,240,8 // psllw $0x8,%xmm8 - .byte 102,68,15,228,5,178,8,0,0 // pmulhuw 0x8b2(%rip),%xmm8 # f40 <_sk_xor__ssse3_lowp+0xfb> + .byte 102,68,15,228,5,234,8,0,0 // pmulhuw 0x8ea(%rip),%xmm8 # 1270 <_sk_xor__ssse3_lowp+0x133> .byte 102,65,15,56,11,192 // pmulhrsw %xmm8,%xmm0 .byte 102,15,56,29,192 // pabsw %xmm0,%xmm0 .byte 102,65,15,56,11,200 // pmulhrsw %xmm8,%xmm1 @@ -37198,9 +37430,9 @@ _sk_scale_u8_ssse3_lowp: .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 119,172 // ja 67f <_sk_scale_u8_ssse3_lowp+0x15> + .byte 119,172 // ja 977 <_sk_scale_u8_ssse3_lowp+0x15> .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,102,0,0,0 // lea 0x66(%rip),%r10 # 744 <_sk_scale_u8_ssse3_lowp+0xda> + .byte 76,141,21,102,0,0,0 // lea 0x66(%rip),%r10 # a3c <_sk_scale_u8_ssse3_lowp+0xda> .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -37219,7 +37451,7 @@ _sk_scale_u8_ssse3_lowp: .byte 102,68,15,196,192,1 // pinsrw $0x1,%eax,%xmm8 .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax .byte 102,68,15,196,192,0 // pinsrw $0x0,%eax,%xmm8 - .byte 233,59,255,255,255 // jmpq 67f <_sk_scale_u8_ssse3_lowp+0x15> + .byte 233,59,255,255,255 // jmpq 977 <_sk_scale_u8_ssse3_lowp+0x15> .byte 240,255 // lock (bad) .byte 255 // (bad) .byte 255,228 // jmpq *%rsp @@ -37244,14 +37476,14 @@ FUNCTION(_sk_lerp_1_float_ssse3_lowp) _sk_lerp_1_float_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 243,68,15,16,0 // movss (%rax),%xmm8 - .byte 243,68,15,89,5,120,7,0,0 // mulss 0x778(%rip),%xmm8 # ee8 <_sk_xor__ssse3_lowp+0xa3> + .byte 243,68,15,89,5,124,7,0,0 // mulss 0x77c(%rip),%xmm8 # 11e4 <_sk_xor__ssse3_lowp+0xa7> .byte 243,65,15,44,192 // cvttss2si %xmm8,%eax .byte 102,68,15,110,192 // movd %eax,%xmm8 .byte 242,69,15,112,192,0 // pshuflw $0x0,%xmm8,%xmm8 .byte 102,69,15,112,192,80 // pshufd $0x50,%xmm8,%xmm8 .byte 102,65,15,56,11,192 // pmulhrsw %xmm8,%xmm0 .byte 102,68,15,56,29,200 // pabsw %xmm0,%xmm9 - .byte 102,68,15,111,21,181,7,0,0 // movdqa 0x7b5(%rip),%xmm10 # f50 <_sk_xor__ssse3_lowp+0x10b> + .byte 102,68,15,111,21,237,7,0,0 // movdqa 0x7ed(%rip),%xmm10 # 1280 <_sk_xor__ssse3_lowp+0x143> .byte 102,69,15,249,208 // psubw %xmm8,%xmm10 .byte 102,15,111,196 // movdqa %xmm4,%xmm0 .byte 102,65,15,56,11,194 // pmulhrsw %xmm10,%xmm0 @@ -37284,14 +37516,14 @@ _sk_lerp_u8_ssse3_lowp: .byte 72,173 // lods %ds:(%rsi),%rax .byte 76,139,24 // mov (%rax),%r11 .byte 77,133,192 // test %r8,%r8 - .byte 15,133,169,0,0,0 // jne 8cc <_sk_lerp_u8_ssse3_lowp+0xb7> + .byte 15,133,169,0,0,0 // jne bc4 <_sk_lerp_u8_ssse3_lowp+0xb7> .byte 243,69,15,126,4,19 // movq (%r11,%rdx,1),%xmm8 .byte 102,68,15,96,192 // punpcklbw %xmm0,%xmm8 .byte 102,65,15,113,240,8 // psllw $0x8,%xmm8 - .byte 102,68,15,228,5,35,7,0,0 // pmulhuw 0x723(%rip),%xmm8 # f60 <_sk_xor__ssse3_lowp+0x11b> + .byte 102,68,15,228,5,91,7,0,0 // pmulhuw 0x75b(%rip),%xmm8 # 1290 <_sk_xor__ssse3_lowp+0x153> .byte 102,65,15,56,11,192 // pmulhrsw %xmm8,%xmm0 .byte 102,68,15,56,29,200 // pabsw %xmm0,%xmm9 - .byte 102,68,15,111,21,30,7,0,0 // movdqa 0x71e(%rip),%xmm10 # f70 <_sk_xor__ssse3_lowp+0x12b> + .byte 102,68,15,111,21,86,7,0,0 // movdqa 0x756(%rip),%xmm10 # 12a0 <_sk_xor__ssse3_lowp+0x163> .byte 102,69,15,249,208 // psubw %xmm8,%xmm10 .byte 102,15,111,196 // movdqa %xmm4,%xmm0 .byte 102,65,15,56,11,194 // pmulhrsw %xmm10,%xmm0 @@ -37321,9 +37553,9 @@ _sk_lerp_u8_ssse3_lowp: .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 .byte 65,254,201 // dec %r9b .byte 65,128,249,6 // cmp $0x6,%r9b - .byte 15,135,73,255,255,255 // ja 82e <_sk_lerp_u8_ssse3_lowp+0x19> + .byte 15,135,73,255,255,255 // ja b26 <_sk_lerp_u8_ssse3_lowp+0x19> .byte 69,15,182,201 // movzbl %r9b,%r9d - .byte 76,141,21,104,0,0,0 // lea 0x68(%rip),%r10 # 958 <_sk_lerp_u8_ssse3_lowp+0x143> + .byte 76,141,21,104,0,0,0 // lea 0x68(%rip),%r10 # c50 <_sk_lerp_u8_ssse3_lowp+0x143> .byte 75,99,4,138 // movslq (%r10,%r9,4),%rax .byte 76,1,208 // add %r10,%rax .byte 255,224 // jmpq *%rax @@ -37342,7 +37574,7 @@ _sk_lerp_u8_ssse3_lowp: .byte 102,68,15,196,192,1 // pinsrw $0x1,%eax,%xmm8 .byte 65,15,182,4,19 // movzbl (%r11,%rdx,1),%eax .byte 102,68,15,196,192,0 // pinsrw $0x0,%eax,%xmm8 - .byte 233,216,254,255,255 // jmpq 82e <_sk_lerp_u8_ssse3_lowp+0x19> + .byte 233,216,254,255,255 // jmpq b26 <_sk_lerp_u8_ssse3_lowp+0x19> .byte 102,144 // xchg %ax,%ax .byte 238 // out %al,(%dx) .byte 255 // (bad) @@ -37434,7 +37666,7 @@ FUNCTION(_sk_srcatop_ssse3_lowp) _sk_srcatop_ssse3_lowp: .byte 102,15,56,11,199 // pmulhrsw %xmm7,%xmm0 .byte 102,68,15,56,29,192 // pabsw %xmm0,%xmm8 - .byte 102,68,15,111,13,137,5,0,0 // movdqa 0x589(%rip),%xmm9 # f80 <_sk_xor__ssse3_lowp+0x13b> + .byte 102,68,15,111,13,193,5,0,0 // movdqa 0x5c1(%rip),%xmm9 # 12b0 <_sk_xor__ssse3_lowp+0x173> .byte 102,68,15,249,203 // psubw %xmm3,%xmm9 .byte 102,15,111,196 // movdqa %xmm4,%xmm0 .byte 102,65,15,56,11,193 // pmulhrsw %xmm9,%xmm0 @@ -37467,7 +37699,7 @@ _sk_dstatop_ssse3_lowp: .byte 102,68,15,111,196 // movdqa %xmm4,%xmm8 .byte 102,68,15,56,11,195 // pmulhrsw %xmm3,%xmm8 .byte 102,69,15,56,29,192 // pabsw %xmm8,%xmm8 - .byte 102,68,15,111,13,8,5,0,0 // movdqa 0x508(%rip),%xmm9 # f90 <_sk_xor__ssse3_lowp+0x14b> + .byte 102,68,15,111,13,64,5,0,0 // movdqa 0x540(%rip),%xmm9 # 12c0 <_sk_xor__ssse3_lowp+0x183> .byte 102,68,15,249,207 // psubw %xmm7,%xmm9 .byte 102,65,15,56,11,193 // pmulhrsw %xmm9,%xmm0 .byte 102,15,56,29,192 // pabsw %xmm0,%xmm0 @@ -37530,7 +37762,7 @@ HIDDEN _sk_srcout_ssse3_lowp .globl _sk_srcout_ssse3_lowp FUNCTION(_sk_srcout_ssse3_lowp) _sk_srcout_ssse3_lowp: - .byte 102,68,15,111,5,46,4,0,0 // movdqa 0x42e(%rip),%xmm8 # fa0 <_sk_xor__ssse3_lowp+0x15b> + .byte 102,68,15,111,5,102,4,0,0 // movdqa 0x466(%rip),%xmm8 # 12d0 <_sk_xor__ssse3_lowp+0x193> .byte 102,68,15,249,199 // psubw %xmm7,%xmm8 .byte 102,65,15,56,11,192 // pmulhrsw %xmm8,%xmm0 .byte 102,15,56,29,192 // pabsw %xmm0,%xmm0 @@ -37547,7 +37779,7 @@ HIDDEN _sk_dstout_ssse3_lowp .globl _sk_dstout_ssse3_lowp FUNCTION(_sk_dstout_ssse3_lowp) _sk_dstout_ssse3_lowp: - .byte 102,68,15,111,5,255,3,0,0 // movdqa 0x3ff(%rip),%xmm8 # fb0 <_sk_xor__ssse3_lowp+0x16b> + .byte 102,68,15,111,5,55,4,0,0 // movdqa 0x437(%rip),%xmm8 # 12e0 <_sk_xor__ssse3_lowp+0x1a3> .byte 102,68,15,249,195 // psubw %xmm3,%xmm8 .byte 102,15,111,196 // movdqa %xmm4,%xmm0 .byte 102,65,15,56,11,192 // pmulhrsw %xmm8,%xmm0 @@ -37567,7 +37799,7 @@ HIDDEN _sk_srcover_ssse3_lowp .globl _sk_srcover_ssse3_lowp FUNCTION(_sk_srcover_ssse3_lowp) _sk_srcover_ssse3_lowp: - .byte 102,68,15,111,5,196,3,0,0 // movdqa 0x3c4(%rip),%xmm8 # fc0 <_sk_xor__ssse3_lowp+0x17b> + .byte 102,68,15,111,5,252,3,0,0 // movdqa 0x3fc(%rip),%xmm8 # 12f0 <_sk_xor__ssse3_lowp+0x1b3> .byte 102,68,15,249,195 // psubw %xmm3,%xmm8 .byte 102,68,15,111,204 // movdqa %xmm4,%xmm9 .byte 102,69,15,56,11,200 // pmulhrsw %xmm8,%xmm9 @@ -37591,7 +37823,7 @@ HIDDEN _sk_dstover_ssse3_lowp .globl _sk_dstover_ssse3_lowp FUNCTION(_sk_dstover_ssse3_lowp) _sk_dstover_ssse3_lowp: - .byte 102,68,15,111,5,111,3,0,0 // movdqa 0x36f(%rip),%xmm8 # fd0 <_sk_xor__ssse3_lowp+0x18b> + .byte 102,68,15,111,5,167,3,0,0 // movdqa 0x3a7(%rip),%xmm8 # 1300 <_sk_xor__ssse3_lowp+0x1c3> .byte 102,68,15,249,199 // psubw %xmm7,%xmm8 .byte 102,65,15,56,11,192 // pmulhrsw %xmm8,%xmm0 .byte 102,15,56,29,192 // pabsw %xmm0,%xmm0 @@ -37627,7 +37859,7 @@ HIDDEN _sk_multiply_ssse3_lowp .globl _sk_multiply_ssse3_lowp FUNCTION(_sk_multiply_ssse3_lowp) _sk_multiply_ssse3_lowp: - .byte 102,68,15,111,5,4,3,0,0 // movdqa 0x304(%rip),%xmm8 # fe0 <_sk_xor__ssse3_lowp+0x19b> + .byte 102,68,15,111,5,60,3,0,0 // movdqa 0x33c(%rip),%xmm8 # 1310 <_sk_xor__ssse3_lowp+0x1d3> .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 .byte 102,68,15,249,207 // psubw %xmm7,%xmm9 .byte 102,68,15,111,208 // movdqa %xmm0,%xmm10 @@ -37687,7 +37919,7 @@ HIDDEN _sk_screen_ssse3_lowp .globl _sk_screen_ssse3_lowp FUNCTION(_sk_screen_ssse3_lowp) _sk_screen_ssse3_lowp: - .byte 102,68,15,111,5,22,2,0,0 // movdqa 0x216(%rip),%xmm8 # ff0 <_sk_xor__ssse3_lowp+0x1ab> + .byte 102,68,15,111,5,78,2,0,0 // movdqa 0x24e(%rip),%xmm8 # 1320 <_sk_xor__ssse3_lowp+0x1e3> .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 .byte 102,68,15,249,200 // psubw %xmm0,%xmm9 .byte 102,68,15,56,11,204 // pmulhrsw %xmm4,%xmm9 @@ -37714,7 +37946,7 @@ HIDDEN _sk_xor__ssse3_lowp .globl _sk_xor__ssse3_lowp FUNCTION(_sk_xor__ssse3_lowp) _sk_xor__ssse3_lowp: - .byte 102,68,15,111,5,178,1,0,0 // movdqa 0x1b2(%rip),%xmm8 # 1000 <_sk_xor__ssse3_lowp+0x1bb> + .byte 102,68,15,111,5,234,1,0,0 // movdqa 0x1ea(%rip),%xmm8 # 1330 <_sk_xor__ssse3_lowp+0x1f3> .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 .byte 102,68,15,249,207 // psubw %xmm7,%xmm9 .byte 102,65,15,56,11,193 // pmulhrsw %xmm9,%xmm0 @@ -37749,7 +37981,8 @@ BALIGN4 .byte 0,71,0 // add %al,0x0(%rdi) .byte 0,0 // add %al,(%rax) .byte 71,0,0 // rex.RXB add %r8b,(%r8) - .byte 0 // .byte 0x0 + .byte 0,71,0 // add %al,0x0(%rdi) + .byte 0,0 // add %al,(%rax) .byte 71 // rex.RXB BALIGN16 @@ -37760,8 +37993,14 @@ BALIGN16 .byte 3,7 // add (%rdi),%eax .byte 11,15 // or (%rdi),%ecx .byte 129,128,129,128,129,128,129,128,129,128// addl $0x80818081,-0x7f7e7f7f(%rax) - .byte 129,128,129,128,129,128,0,4,8,12 // addl $0xc080400,-0x7f7e7f7f(%rax) - .byte 1,5,9,13,2,6 // add %eax,0x6020d09(%rip) # 6021c23 <_sk_xor__ssse3_lowp+0x6020dde> + .byte 129,128,129,128,129,128,129,128,129,128// addl $0x80818081,-0x7f7e7f7f(%rax) + .byte 129,128,129,128,129,128,129,128,129,128// addl $0x80818081,-0x7f7e7f7f(%rax) + .byte 129,128,129,128,129,128,129,128,129,128// addl $0x80818081,-0x7f7e7f7f(%rax) + .byte 129,128,129,128,129,128,129,128,0,128// addl $0x80008081,-0x7f7e7f7f(%rax) + .byte 0,128,0,128,0,128 // add %al,-0x7fff8000(%rax) + .byte 0,128,0,128,0,128 // add %al,-0x7fff8000(%rax) + .byte 0,128,0,4,8,12 // add %al,0xc080400(%rax) + .byte 1,5,9,13,2,6 // add %eax,0x6020d09(%rip) # 6021f53 <_sk_xor__ssse3_lowp+0x6020e16> .byte 10,14 // or (%rsi),%cl .byte 3,7 // add (%rdi),%eax .byte 11,15 // or (%rdi),%ecx diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 4c82299b8f..5c0624f477 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -26226,7 +26226,7 @@ _sk_just_return_ssse3_lowp LABEL PROC PUBLIC _sk_constant_color_ssse3_lowp _sk_constant_color_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax - DB 243,15,16,29,45,14,0,0 ; movss 0xe2d(%rip),%xmm3 # f8c <_sk_xor__ssse3_lowp+0x9b> + DB 243,15,16,29,45,17,0,0 ; movss 0x112d(%rip),%xmm3 # 128c <_sk_xor__ssse3_lowp+0x9b> DB 243,15,16,0 ; movss (%rax),%xmm0 DB 243,15,89,195 ; mulss %xmm3,%xmm0 DB 243,68,15,44,200 ; cvttss2si %xmm0,%r9d @@ -26253,15 +26253,50 @@ _sk_constant_color_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_set_rgb_ssse3_lowp +_sk_set_rgb_ssse3_lowp LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 243,15,16,21,182,16,0,0 ; movss 0x10b6(%rip),%xmm2 # 1290 <_sk_xor__ssse3_lowp+0x9f> + DB 243,15,16,0 ; movss (%rax),%xmm0 + DB 243,15,89,194 ; mulss %xmm2,%xmm0 + DB 243,68,15,44,200 ; cvttss2si %xmm0,%r9d + DB 102,65,15,110,193 ; movd %r9d,%xmm0 + DB 242,15,112,192,0 ; pshuflw $0x0,%xmm0,%xmm0 + DB 102,15,112,192,80 ; pshufd $0x50,%xmm0,%xmm0 + DB 243,15,16,72,4 ; movss 0x4(%rax),%xmm1 + DB 243,15,89,202 ; mulss %xmm2,%xmm1 + DB 243,68,15,44,201 ; cvttss2si %xmm1,%r9d + DB 102,65,15,110,201 ; movd %r9d,%xmm1 + DB 242,15,112,201,0 ; pshuflw $0x0,%xmm1,%xmm1 + DB 102,15,112,201,80 ; pshufd $0x50,%xmm1,%xmm1 + DB 243,15,89,80,8 ; mulss 0x8(%rax),%xmm2 + DB 243,15,44,194 ; cvttss2si %xmm2,%eax + DB 102,15,110,208 ; movd %eax,%xmm2 + DB 242,15,112,210,0 ; pshuflw $0x0,%xmm2,%xmm2 + DB 102,15,112,210,80 ; pshufd $0x50,%xmm2,%xmm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + +PUBLIC _sk_premul_ssse3_lowp +_sk_premul_ssse3_lowp LABEL PROC + DB 102,15,56,11,195 ; pmulhrsw %xmm3,%xmm0 + DB 102,15,56,29,192 ; pabsw %xmm0,%xmm0 + DB 102,15,56,11,203 ; pmulhrsw %xmm3,%xmm1 + DB 102,15,56,29,201 ; pabsw %xmm1,%xmm1 + DB 102,15,56,11,211 ; pmulhrsw %xmm3,%xmm2 + DB 102,15,56,29,210 ; pabsw %xmm2,%xmm2 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_load_8888_ssse3_lowp _sk_load_8888_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 DB 77,133,192 ; test %r8,%r8 - DB 117,113 ; jne 24b <_sk_load_8888_ssse3_lowp+0x7b> + DB 117,113 ; jne 2cb <_sk_load_8888_ssse3_lowp+0x7b> DB 69,15,16,76,147,16 ; movups 0x10(%r11,%rdx,4),%xmm9 DB 69,15,16,4,147 ; movups (%r11,%rdx,4),%xmm8 - DB 102,15,111,5,179,13,0,0 ; movdqa 0xdb3(%rip),%xmm0 # fa0 <_sk_xor__ssse3_lowp+0xaf> + DB 102,15,111,5,51,16,0,0 ; movdqa 0x1033(%rip),%xmm0 # 12a0 <_sk_xor__ssse3_lowp+0xaf> DB 102,68,15,56,0,192 ; pshufb %xmm0,%xmm8 DB 102,68,15,56,0,200 ; pshufb %xmm0,%xmm9 DB 102,65,15,111,208 ; movdqa %xmm8,%xmm2 @@ -26275,7 +26310,7 @@ _sk_load_8888_ssse3_lowp LABEL PROC DB 102,15,239,210 ; pxor %xmm2,%xmm2 DB 102,65,15,96,208 ; punpcklbw %xmm8,%xmm2 DB 102,65,15,104,216 ; punpckhbw %xmm8,%xmm3 - DB 102,68,15,111,5,125,13,0,0 ; movdqa 0xd7d(%rip),%xmm8 # fb0 <_sk_xor__ssse3_lowp+0xbf> + DB 102,68,15,111,5,253,15,0,0 ; movdqa 0xffd(%rip),%xmm8 # 12b0 <_sk_xor__ssse3_lowp+0xbf> DB 102,65,15,228,192 ; pmulhuw %xmm8,%xmm0 DB 102,65,15,228,200 ; pmulhuw %xmm8,%xmm1 DB 102,65,15,228,208 ; pmulhuw %xmm8,%xmm2 @@ -26288,9 +26323,9 @@ _sk_load_8888_ssse3_lowp LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,129 ; ja 1e5 <_sk_load_8888_ssse3_lowp+0x15> + DB 119,129 ; ja 265 <_sk_load_8888_ssse3_lowp+0x15> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 2f4 <_sk_load_8888_ssse3_lowp+0x124> + DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 374 <_sk_load_8888_ssse3_lowp+0x124> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -26315,7 +26350,7 @@ _sk_load_8888_ssse3_lowp LABEL PROC DB 68,15,40,192 ; movaps %xmm0,%xmm8 DB 243,65,15,16,4,147 ; movss (%r11,%rdx,4),%xmm0 DB 243,68,15,16,192 ; movss %xmm0,%xmm8 - DB 233,244,254,255,255 ; jmpq 1e5 <_sk_load_8888_ssse3_lowp+0x15> + DB 233,244,254,255,255 ; jmpq 265 <_sk_load_8888_ssse3_lowp+0x15> DB 15,31,0 ; nopl (%rax) DB 237 ; in (%dx),%eax DB 255 ; (bad) @@ -26355,7 +26390,7 @@ _sk_store_8888_ssse3_lowp LABEL PROC DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8 DB 102,69,15,105,202 ; punpckhwd %xmm10,%xmm9 DB 77,133,192 ; test %r8,%r8 - DB 117,17 ; jne 37f <_sk_store_8888_ssse3_lowp+0x6f> + DB 117,17 ; jne 3ff <_sk_store_8888_ssse3_lowp+0x6f> DB 243,69,15,127,76,147,16 ; movdqu %xmm9,0x10(%r11,%rdx,4) DB 243,69,15,127,4,147 ; movdqu %xmm8,(%r11,%rdx,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -26364,9 +26399,9 @@ _sk_store_8888_ssse3_lowp LABEL PROC DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,236 ; ja 37b <_sk_store_8888_ssse3_lowp+0x6b> + DB 119,236 ; ja 3fb <_sk_store_8888_ssse3_lowp+0x6b> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,90,0,0,0 ; lea 0x5a(%rip),%r10 # 3f4 <_sk_store_8888_ssse3_lowp+0xe4> + DB 76,141,21,90,0,0,0 ; lea 0x5a(%rip),%r10 # 474 <_sk_store_8888_ssse3_lowp+0xe4> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -26382,7 +26417,7 @@ _sk_store_8888_ssse3_lowp LABEL PROC DB 102,69,15,112,200,229 ; pshufd $0xe5,%xmm8,%xmm9 DB 102,69,15,126,76,147,4 ; movd %xmm9,0x4(%r11,%rdx,4) DB 102,69,15,126,4,147 ; movd %xmm8,(%r11,%rdx,4) - DB 235,136 ; jmp 37b <_sk_store_8888_ssse3_lowp+0x6b> + DB 235,136 ; jmp 3fb <_sk_store_8888_ssse3_lowp+0x6b> DB 144 ; nop DB 247,255 ; idiv %edi DB 255 ; (bad) @@ -26405,16 +26440,209 @@ _sk_store_8888_ssse3_lowp LABEL PROC DB 255 ; (bad) DB 255 ; .byte 0xff +PUBLIC _sk_load_a8_ssse3_lowp +_sk_load_a8_ssse3_lowp LABEL PROC + DB 80 ; push %rax + DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 117,39 ; jne 4c2 <_sk_load_a8_ssse3_lowp+0x32> + DB 243,65,15,126,28,19 ; movq (%r11,%rdx,1),%xmm3 + DB 102,15,96,216 ; punpcklbw %xmm0,%xmm3 + DB 102,15,113,243,8 ; psllw $0x8,%xmm3 + DB 102,15,228,29,14,14,0,0 ; pmulhuw 0xe0e(%rip),%xmm3 # 12c0 <_sk_xor__ssse3_lowp+0xcf> + DB 72,173 ; lods %ds:(%rsi),%rax + DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 15,87,201 ; xorps %xmm1,%xmm1 + DB 15,87,210 ; xorps %xmm2,%xmm2 + DB 65,89 ; pop %r9 + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 102,15,239,219 ; pxor %xmm3,%xmm3 + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,207 ; ja 4a5 <_sk_load_a8_ssse3_lowp+0x15> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,95,0,0,0 ; lea 0x5f(%rip),%r10 # 540 <_sk_load_a8_ssse3_lowp+0xb0> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax + DB 102,15,239,219 ; pxor %xmm3,%xmm3 + DB 102,15,196,216,6 ; pinsrw $0x6,%eax,%xmm3 + DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax + DB 102,15,196,216,5 ; pinsrw $0x5,%eax,%xmm3 + DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax + DB 102,15,196,216,4 ; pinsrw $0x4,%eax,%xmm3 + DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax + DB 102,15,196,216,3 ; pinsrw $0x3,%eax,%xmm3 + DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax + DB 102,15,196,216,2 ; pinsrw $0x2,%eax,%xmm3 + DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax + DB 102,15,196,216,1 ; pinsrw $0x1,%eax,%xmm3 + DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax + DB 102,15,196,216,0 ; pinsrw $0x0,%eax,%xmm3 + DB 233,102,255,255,255 ; jmpq 4a5 <_sk_load_a8_ssse3_lowp+0x15> + DB 144 ; nop + DB 240,255 ; lock (bad) + DB 255 ; (bad) + DB 255,229 ; jmpq *%rbp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 218,255 ; (bad) + DB 255 ; (bad) + DB 255,207 ; dec %edi + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,196 ; inc %esp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 185,255,255,255,170 ; mov $0xaaffffff,%ecx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; .byte 0xff + +PUBLIC _sk_store_a8_ssse3_lowp +_sk_store_a8_ssse3_lowp LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,139,24 ; mov (%rax),%r11 + DB 102,68,15,111,195 ; movdqa %xmm3,%xmm8 + DB 102,65,15,113,208,7 ; psrlw $0x7,%xmm8 + DB 102,69,15,103,192 ; packuswb %xmm8,%xmm8 + DB 77,133,192 ; test %r8,%r8 + DB 117,10 ; jne 580 <_sk_store_a8_ssse3_lowp+0x24> + DB 242,69,15,17,4,19 ; movsd %xmm8,(%r11,%rdx,1) + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,236 ; ja 57c <_sk_store_a8_ssse3_lowp+0x20> + DB 72,131,236,120 ; sub $0x78,%rsp + DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,128,0,0,0 ; lea 0x80(%rip),%r10 # 624 <_sk_store_a8_ssse3_lowp+0xc8> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 102,68,15,127,68,36,96 ; movdqa %xmm8,0x60(%rsp) + DB 138,68,36,108 ; mov 0x6c(%rsp),%al + DB 65,136,68,19,6 ; mov %al,0x6(%r11,%rdx,1) + DB 102,68,15,127,68,36,80 ; movdqa %xmm8,0x50(%rsp) + DB 138,68,36,90 ; mov 0x5a(%rsp),%al + DB 65,136,68,19,5 ; mov %al,0x5(%r11,%rdx,1) + DB 102,68,15,127,68,36,64 ; movdqa %xmm8,0x40(%rsp) + DB 138,68,36,72 ; mov 0x48(%rsp),%al + DB 65,136,68,19,4 ; mov %al,0x4(%r11,%rdx,1) + DB 102,68,15,127,68,36,48 ; movdqa %xmm8,0x30(%rsp) + DB 138,68,36,54 ; mov 0x36(%rsp),%al + DB 65,136,68,19,3 ; mov %al,0x3(%r11,%rdx,1) + DB 102,68,15,127,68,36,32 ; movdqa %xmm8,0x20(%rsp) + DB 138,68,36,36 ; mov 0x24(%rsp),%al + DB 65,136,68,19,2 ; mov %al,0x2(%r11,%rdx,1) + DB 102,68,15,127,68,36,16 ; movdqa %xmm8,0x10(%rsp) + DB 138,68,36,18 ; mov 0x12(%rsp),%al + DB 65,136,68,19,1 ; mov %al,0x1(%r11,%rdx,1) + DB 102,68,15,127,4,36 ; movdqa %xmm8,(%rsp) + DB 138,4,36 ; mov (%rsp),%al + DB 65,136,4,19 ; mov %al,(%r11,%rdx,1) + DB 72,131,196,120 ; add $0x78,%rsp + DB 233,89,255,255,255 ; jmpq 57c <_sk_store_a8_ssse3_lowp+0x20> + DB 144 ; nop + DB 233,255,255,255,217 ; jmpq ffffffffda000628 <_sk_xor__ssse3_lowp+0xffffffffd9fff437> + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,201 ; dec %ecx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 185,255,255,255,169 ; mov $0xa9ffffff,%ecx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,153,255,255,255,137 ; lcall *-0x76000001(%rcx) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; .byte 0xff + +PUBLIC _sk_load_g8_ssse3_lowp +_sk_load_g8_ssse3_lowp LABEL PROC + DB 80 ; push %rax + DB 72,173 ; lods %ds:(%rsi),%rax + DB 76,139,24 ; mov (%rax),%r11 + DB 77,133,192 ; test %r8,%r8 + DB 117,44 ; jne 677 <_sk_load_g8_ssse3_lowp+0x37> + DB 243,65,15,126,4,19 ; movq (%r11,%rdx,1),%xmm0 + DB 102,15,96,192 ; punpcklbw %xmm0,%xmm0 + DB 102,15,113,240,8 ; psllw $0x8,%xmm0 + DB 102,15,228,5,110,12,0,0 ; pmulhuw 0xc6e(%rip),%xmm0 # 12d0 <_sk_xor__ssse3_lowp+0xdf> + DB 72,173 ; lods %ds:(%rsi),%rax + DB 15,40,29,117,12,0,0 ; movaps 0xc75(%rip),%xmm3 # 12e0 <_sk_xor__ssse3_lowp+0xef> + DB 102,15,111,200 ; movdqa %xmm0,%xmm1 + DB 102,15,111,208 ; movdqa %xmm0,%xmm2 + DB 65,89 ; pop %r9 + DB 255,224 ; jmpq *%rax + DB 69,137,193 ; mov %r8d,%r9d + DB 65,128,225,7 ; and $0x7,%r9b + DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 65,254,201 ; dec %r9b + DB 65,128,249,6 ; cmp $0x6,%r9b + DB 119,202 ; ja 655 <_sk_load_g8_ssse3_lowp+0x15> + DB 69,15,182,201 ; movzbl %r9b,%r9d + DB 76,141,21,94,0,0,0 ; lea 0x5e(%rip),%r10 # 6f4 <_sk_load_g8_ssse3_lowp+0xb4> + DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 65,15,182,68,19,6 ; movzbl 0x6(%r11,%rdx,1),%eax + DB 102,15,239,192 ; pxor %xmm0,%xmm0 + DB 102,15,196,192,6 ; pinsrw $0x6,%eax,%xmm0 + DB 65,15,182,68,19,5 ; movzbl 0x5(%r11,%rdx,1),%eax + DB 102,15,196,192,5 ; pinsrw $0x5,%eax,%xmm0 + DB 65,15,182,68,19,4 ; movzbl 0x4(%r11,%rdx,1),%eax + DB 102,15,196,192,4 ; pinsrw $0x4,%eax,%xmm0 + DB 65,15,182,68,19,3 ; movzbl 0x3(%r11,%rdx,1),%eax + DB 102,15,196,192,3 ; pinsrw $0x3,%eax,%xmm0 + DB 65,15,182,68,19,2 ; movzbl 0x2(%r11,%rdx,1),%eax + DB 102,15,196,192,2 ; pinsrw $0x2,%eax,%xmm0 + DB 65,15,182,68,19,1 ; movzbl 0x1(%r11,%rdx,1),%eax + DB 102,15,196,192,1 ; pinsrw $0x1,%eax,%xmm0 + DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax + DB 102,15,196,192,0 ; pinsrw $0x0,%eax,%xmm0 + DB 233,97,255,255,255 ; jmpq 655 <_sk_load_g8_ssse3_lowp+0x15> + DB 241 ; icebp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,230 ; jmpq *%rsi + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 219,255 ; (bad) + DB 255 ; (bad) + DB 255,208 ; callq *%rax + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,197 ; inc %ebp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 186,255,255,255,171 ; mov $0xabffffff,%edx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; .byte 0xff + PUBLIC _sk_srcover_rgba_8888_ssse3_lowp _sk_srcover_rgba_8888_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 DB 77,133,192 ; test %r8,%r8 - DB 15,133,66,1,0,0 ; jne 560 <_sk_srcover_rgba_8888_ssse3_lowp+0x150> + DB 15,133,66,1,0,0 ; jne 860 <_sk_srcover_rgba_8888_ssse3_lowp+0x150> DB 69,15,16,76,147,16 ; movups 0x10(%r11,%rdx,4),%xmm9 DB 69,15,16,4,147 ; movups (%r11,%rdx,4),%xmm8 DB 77,133,192 ; test %r8,%r8 - DB 102,15,111,37,140,11,0,0 ; movdqa 0xb8c(%rip),%xmm4 # fc0 <_sk_xor__ssse3_lowp+0xcf> + DB 102,15,111,37,188,11,0,0 ; movdqa 0xbbc(%rip),%xmm4 # 12f0 <_sk_xor__ssse3_lowp+0xff> DB 102,68,15,56,0,196 ; pshufb %xmm4,%xmm8 DB 102,68,15,56,0,204 ; pshufb %xmm4,%xmm9 DB 102,65,15,111,240 ; movdqa %xmm8,%xmm6 @@ -26428,12 +26656,12 @@ _sk_srcover_rgba_8888_ssse3_lowp LABEL PROC DB 102,15,239,246 ; pxor %xmm6,%xmm6 DB 102,65,15,96,240 ; punpcklbw %xmm8,%xmm6 DB 102,65,15,104,248 ; punpckhbw %xmm8,%xmm7 - DB 102,68,15,111,5,86,11,0,0 ; movdqa 0xb56(%rip),%xmm8 # fd0 <_sk_xor__ssse3_lowp+0xdf> + DB 102,68,15,111,5,134,11,0,0 ; movdqa 0xb86(%rip),%xmm8 # 1300 <_sk_xor__ssse3_lowp+0x10f> DB 102,65,15,228,224 ; pmulhuw %xmm8,%xmm4 DB 102,65,15,228,232 ; pmulhuw %xmm8,%xmm5 DB 102,65,15,228,240 ; pmulhuw %xmm8,%xmm6 DB 102,65,15,228,248 ; pmulhuw %xmm8,%xmm7 - DB 102,68,15,111,29,73,11,0,0 ; movdqa 0xb49(%rip),%xmm11 # fe0 <_sk_xor__ssse3_lowp+0xef> + DB 102,68,15,111,29,121,11,0,0 ; movdqa 0xb79(%rip),%xmm11 # 1310 <_sk_xor__ssse3_lowp+0x11f> DB 102,68,15,249,219 ; psubw %xmm3,%xmm11 DB 102,68,15,111,196 ; movdqa %xmm4,%xmm8 DB 102,69,15,56,11,195 ; pmulhrsw %xmm11,%xmm8 @@ -26466,7 +26694,7 @@ _sk_srcover_rgba_8888_ssse3_lowp LABEL PROC DB 102,15,111,193 ; movdqa %xmm1,%xmm0 DB 102,15,97,194 ; punpcklwd %xmm2,%xmm0 DB 102,15,105,202 ; punpckhwd %xmm2,%xmm1 - DB 15,133,207,0,0,0 ; jne 60a <_sk_srcover_rgba_8888_ssse3_lowp+0x1fa> + DB 15,133,207,0,0,0 ; jne 90a <_sk_srcover_rgba_8888_ssse3_lowp+0x1fa> DB 243,65,15,127,76,147,16 ; movdqu %xmm1,0x10(%r11,%rdx,4) DB 243,65,15,127,4,147 ; movdqu %xmm0,(%r11,%rdx,4) DB 72,173 ; lods %ds:(%rsi),%rax @@ -26481,9 +26709,9 @@ _sk_srcover_rgba_8888_ssse3_lowp LABEL PROC DB 69,15,87,192 ; xorps %xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,172,254,255,255 ; ja 429 <_sk_srcover_rgba_8888_ssse3_lowp+0x19> + DB 15,135,172,254,255,255 ; ja 729 <_sk_srcover_rgba_8888_ssse3_lowp+0x19> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,248,0,0,0 ; lea 0xf8(%rip),%r10 # 680 <_sk_srcover_rgba_8888_ssse3_lowp+0x270> + DB 76,141,21,248,0,0,0 ; lea 0xf8(%rip),%r10 # 980 <_sk_srcover_rgba_8888_ssse3_lowp+0x270> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -26508,14 +26736,14 @@ _sk_srcover_rgba_8888_ssse3_lowp LABEL PROC DB 68,15,40,196 ; movaps %xmm4,%xmm8 DB 243,65,15,16,36,147 ; movss (%r11,%rdx,4),%xmm4 DB 243,68,15,16,196 ; movss %xmm4,%xmm8 - DB 233,31,254,255,255 ; jmpq 429 <_sk_srcover_rgba_8888_ssse3_lowp+0x19> + DB 233,31,254,255,255 ; jmpq 729 <_sk_srcover_rgba_8888_ssse3_lowp+0x19> DB 69,137,193 ; mov %r8d,%r9d DB 65,128,225,7 ; and $0x7,%r9b DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,42,255,255,255 ; ja 548 <_sk_srcover_rgba_8888_ssse3_lowp+0x138> + DB 15,135,42,255,255,255 ; ja 848 <_sk_srcover_rgba_8888_ssse3_lowp+0x138> DB 65,15,182,193 ; movzbl %r9b,%eax - DB 76,141,13,115,0,0,0 ; lea 0x73(%rip),%r9 # 69c <_sk_srcover_rgba_8888_ssse3_lowp+0x28c> + DB 76,141,13,115,0,0,0 ; lea 0x73(%rip),%r9 # 99c <_sk_srcover_rgba_8888_ssse3_lowp+0x28c> DB 73,99,4,129 ; movslq (%r9,%rax,4),%rax DB 76,1,200 ; add %r9,%rax DB 255,224 ; jmpq *%rax @@ -26531,8 +26759,8 @@ _sk_srcover_rgba_8888_ssse3_lowp LABEL PROC DB 102,15,112,200,229 ; pshufd $0xe5,%xmm0,%xmm1 DB 102,65,15,126,76,147,4 ; movd %xmm1,0x4(%r11,%rdx,4) DB 102,65,15,126,4,147 ; movd %xmm0,(%r11,%rdx,4) - DB 233,200,254,255,255 ; jmpq 548 <_sk_srcover_rgba_8888_ssse3_lowp+0x138> - DB 122,255 ; jp 681 <_sk_srcover_rgba_8888_ssse3_lowp+0x271> + DB 233,200,254,255,255 ; jmpq 848 <_sk_srcover_rgba_8888_ssse3_lowp+0x138> + DB 122,255 ; jp 981 <_sk_srcover_rgba_8888_ssse3_lowp+0x271> DB 255 ; (bad) DB 255,101,255 ; jmpq *-0x1(%rbp) DB 255 ; (bad) @@ -26569,7 +26797,7 @@ PUBLIC _sk_scale_1_float_ssse3_lowp _sk_scale_1_float_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 243,68,15,89,5,200,8,0,0 ; mulss 0x8c8(%rip),%xmm8 # f90 <_sk_xor__ssse3_lowp+0x9f> + DB 243,68,15,89,5,204,8,0,0 ; mulss 0x8cc(%rip),%xmm8 # 1294 <_sk_xor__ssse3_lowp+0xa3> DB 243,65,15,44,192 ; cvttss2si %xmm8,%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 242,69,15,112,192,0 ; pshuflw $0x0,%xmm8,%xmm8 @@ -26591,11 +26819,11 @@ _sk_scale_u8_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 DB 77,133,192 ; test %r8,%r8 - DB 117,76 ; jne 765 <_sk_scale_u8_ssse3_lowp+0x57> + DB 117,76 ; jne a65 <_sk_scale_u8_ssse3_lowp+0x57> DB 243,69,15,126,4,19 ; movq (%r11,%rdx,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,65,15,113,240,8 ; psllw $0x8,%xmm8 - DB 102,68,15,228,5,189,8,0,0 ; pmulhuw 0x8bd(%rip),%xmm8 # ff0 <_sk_xor__ssse3_lowp+0xff> + DB 102,68,15,228,5,237,8,0,0 ; pmulhuw 0x8ed(%rip),%xmm8 # 1320 <_sk_xor__ssse3_lowp+0x12f> DB 102,65,15,56,11,192 ; pmulhrsw %xmm8,%xmm0 DB 102,15,56,29,192 ; pabsw %xmm0,%xmm0 DB 102,65,15,56,11,200 ; pmulhrsw %xmm8,%xmm1 @@ -26612,9 +26840,9 @@ _sk_scale_u8_ssse3_lowp LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 119,170 ; ja 724 <_sk_scale_u8_ssse3_lowp+0x16> + DB 119,170 ; ja a24 <_sk_scale_u8_ssse3_lowp+0x16> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,103,0,0,0 ; lea 0x67(%rip),%r10 # 7ec <_sk_scale_u8_ssse3_lowp+0xde> + DB 76,141,21,103,0,0,0 ; lea 0x67(%rip),%r10 # aec <_sk_scale_u8_ssse3_lowp+0xde> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -26633,7 +26861,7 @@ _sk_scale_u8_ssse3_lowp LABEL PROC DB 102,68,15,196,192,1 ; pinsrw $0x1,%eax,%xmm8 DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax DB 102,68,15,196,192,0 ; pinsrw $0x0,%eax,%xmm8 - DB 233,57,255,255,255 ; jmpq 724 <_sk_scale_u8_ssse3_lowp+0x16> + DB 233,57,255,255,255 ; jmpq a24 <_sk_scale_u8_ssse3_lowp+0x16> DB 144 ; nop DB 239 ; out %eax,(%dx) DB 255 ; (bad) @@ -26661,14 +26889,14 @@ PUBLIC _sk_lerp_1_float_ssse3_lowp _sk_lerp_1_float_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 243,68,15,16,0 ; movss (%rax),%xmm8 - DB 243,68,15,89,5,124,7,0,0 ; mulss 0x77c(%rip),%xmm8 # f94 <_sk_xor__ssse3_lowp+0xa3> + DB 243,68,15,89,5,128,7,0,0 ; mulss 0x780(%rip),%xmm8 # 1298 <_sk_xor__ssse3_lowp+0xa7> DB 243,65,15,44,192 ; cvttss2si %xmm8,%eax DB 102,68,15,110,192 ; movd %eax,%xmm8 DB 242,69,15,112,192,0 ; pshuflw $0x0,%xmm8,%xmm8 DB 102,69,15,112,192,80 ; pshufd $0x50,%xmm8,%xmm8 DB 102,65,15,56,11,192 ; pmulhrsw %xmm8,%xmm0 DB 102,68,15,56,29,200 ; pabsw %xmm0,%xmm9 - DB 102,68,15,111,21,189,7,0,0 ; movdqa 0x7bd(%rip),%xmm10 # 1000 <_sk_xor__ssse3_lowp+0x10f> + DB 102,68,15,111,21,237,7,0,0 ; movdqa 0x7ed(%rip),%xmm10 # 1330 <_sk_xor__ssse3_lowp+0x13f> DB 102,69,15,249,208 ; psubw %xmm8,%xmm10 DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,65,15,56,11,194 ; pmulhrsw %xmm10,%xmm0 @@ -26700,14 +26928,14 @@ _sk_lerp_u8_ssse3_lowp LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 76,139,24 ; mov (%rax),%r11 DB 77,133,192 ; test %r8,%r8 - DB 15,133,171,0,0,0 ; jne 977 <_sk_lerp_u8_ssse3_lowp+0xba> + DB 15,133,171,0,0,0 ; jne c77 <_sk_lerp_u8_ssse3_lowp+0xba> DB 243,69,15,126,4,19 ; movq (%r11,%rdx,1),%xmm8 DB 102,68,15,96,192 ; punpcklbw %xmm0,%xmm8 DB 102,65,15,113,240,8 ; psllw $0x8,%xmm8 - DB 102,68,15,228,5,42,7,0,0 ; pmulhuw 0x72a(%rip),%xmm8 # 1010 <_sk_xor__ssse3_lowp+0x11f> + DB 102,68,15,228,5,90,7,0,0 ; pmulhuw 0x75a(%rip),%xmm8 # 1340 <_sk_xor__ssse3_lowp+0x14f> DB 102,65,15,56,11,192 ; pmulhrsw %xmm8,%xmm0 DB 102,68,15,56,29,200 ; pabsw %xmm0,%xmm9 - DB 102,68,15,111,21,37,7,0,0 ; movdqa 0x725(%rip),%xmm10 # 1020 <_sk_xor__ssse3_lowp+0x12f> + DB 102,68,15,111,21,85,7,0,0 ; movdqa 0x755(%rip),%xmm10 # 1350 <_sk_xor__ssse3_lowp+0x15f> DB 102,69,15,249,208 ; psubw %xmm8,%xmm10 DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,65,15,56,11,194 ; pmulhrsw %xmm10,%xmm0 @@ -26738,9 +26966,9 @@ _sk_lerp_u8_ssse3_lowp LABEL PROC DB 102,69,15,239,192 ; pxor %xmm8,%xmm8 DB 65,254,201 ; dec %r9b DB 65,128,249,6 ; cmp $0x6,%r9b - DB 15,135,71,255,255,255 ; ja 8d7 <_sk_lerp_u8_ssse3_lowp+0x1a> + DB 15,135,71,255,255,255 ; ja bd7 <_sk_lerp_u8_ssse3_lowp+0x1a> DB 69,15,182,201 ; movzbl %r9b,%r9d - DB 76,141,21,105,0,0,0 ; lea 0x69(%rip),%r10 # a04 <_sk_lerp_u8_ssse3_lowp+0x147> + DB 76,141,21,105,0,0,0 ; lea 0x69(%rip),%r10 # d04 <_sk_lerp_u8_ssse3_lowp+0x147> DB 75,99,4,138 ; movslq (%r10,%r9,4),%rax DB 76,1,208 ; add %r10,%rax DB 255,224 ; jmpq *%rax @@ -26759,7 +26987,7 @@ _sk_lerp_u8_ssse3_lowp LABEL PROC DB 102,68,15,196,192,1 ; pinsrw $0x1,%eax,%xmm8 DB 65,15,182,4,19 ; movzbl (%r11,%rdx,1),%eax DB 102,68,15,196,192,0 ; pinsrw $0x0,%eax,%xmm8 - DB 233,214,254,255,255 ; jmpq 8d7 <_sk_lerp_u8_ssse3_lowp+0x1a> + DB 233,214,254,255,255 ; jmpq bd7 <_sk_lerp_u8_ssse3_lowp+0x1a> DB 15,31,0 ; nopl (%rax) DB 237 ; in (%dx),%eax DB 255 ; (bad) @@ -26839,7 +27067,7 @@ PUBLIC _sk_srcatop_ssse3_lowp _sk_srcatop_ssse3_lowp LABEL PROC DB 102,15,56,11,199 ; pmulhrsw %xmm7,%xmm0 DB 102,68,15,56,29,192 ; pabsw %xmm0,%xmm8 - DB 102,68,15,111,13,141,5,0,0 ; movdqa 0x58d(%rip),%xmm9 # 1030 <_sk_xor__ssse3_lowp+0x13f> + DB 102,68,15,111,13,189,5,0,0 ; movdqa 0x5bd(%rip),%xmm9 # 1360 <_sk_xor__ssse3_lowp+0x16f> DB 102,68,15,249,203 ; psubw %xmm3,%xmm9 DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,65,15,56,11,193 ; pmulhrsw %xmm9,%xmm0 @@ -26870,7 +27098,7 @@ _sk_dstatop_ssse3_lowp LABEL PROC DB 102,68,15,111,196 ; movdqa %xmm4,%xmm8 DB 102,68,15,56,11,195 ; pmulhrsw %xmm3,%xmm8 DB 102,69,15,56,29,192 ; pabsw %xmm8,%xmm8 - DB 102,68,15,111,13,12,5,0,0 ; movdqa 0x50c(%rip),%xmm9 # 1040 <_sk_xor__ssse3_lowp+0x14f> + DB 102,68,15,111,13,60,5,0,0 ; movdqa 0x53c(%rip),%xmm9 # 1370 <_sk_xor__ssse3_lowp+0x17f> DB 102,68,15,249,207 ; psubw %xmm7,%xmm9 DB 102,65,15,56,11,193 ; pmulhrsw %xmm9,%xmm0 DB 102,15,56,29,192 ; pabsw %xmm0,%xmm0 @@ -26927,7 +27155,7 @@ _sk_dstin_ssse3_lowp LABEL PROC PUBLIC _sk_srcout_ssse3_lowp _sk_srcout_ssse3_lowp LABEL PROC - DB 102,68,15,111,5,50,4,0,0 ; movdqa 0x432(%rip),%xmm8 # 1050 <_sk_xor__ssse3_lowp+0x15f> + DB 102,68,15,111,5,98,4,0,0 ; movdqa 0x462(%rip),%xmm8 # 1380 <_sk_xor__ssse3_lowp+0x18f> DB 102,68,15,249,199 ; psubw %xmm7,%xmm8 DB 102,65,15,56,11,192 ; pmulhrsw %xmm8,%xmm0 DB 102,15,56,29,192 ; pabsw %xmm0,%xmm0 @@ -26942,7 +27170,7 @@ _sk_srcout_ssse3_lowp LABEL PROC PUBLIC _sk_dstout_ssse3_lowp _sk_dstout_ssse3_lowp LABEL PROC - DB 102,68,15,111,5,3,4,0,0 ; movdqa 0x403(%rip),%xmm8 # 1060 <_sk_xor__ssse3_lowp+0x16f> + DB 102,68,15,111,5,51,4,0,0 ; movdqa 0x433(%rip),%xmm8 # 1390 <_sk_xor__ssse3_lowp+0x19f> DB 102,68,15,249,195 ; psubw %xmm3,%xmm8 DB 102,15,111,196 ; movdqa %xmm4,%xmm0 DB 102,65,15,56,11,192 ; pmulhrsw %xmm8,%xmm0 @@ -26960,7 +27188,7 @@ _sk_dstout_ssse3_lowp LABEL PROC PUBLIC _sk_srcover_ssse3_lowp _sk_srcover_ssse3_lowp LABEL PROC - DB 102,68,15,111,5,200,3,0,0 ; movdqa 0x3c8(%rip),%xmm8 # 1070 <_sk_xor__ssse3_lowp+0x17f> + DB 102,68,15,111,5,248,3,0,0 ; movdqa 0x3f8(%rip),%xmm8 # 13a0 <_sk_xor__ssse3_lowp+0x1af> DB 102,68,15,249,195 ; psubw %xmm3,%xmm8 DB 102,68,15,111,204 ; movdqa %xmm4,%xmm9 DB 102,69,15,56,11,200 ; pmulhrsw %xmm8,%xmm9 @@ -26982,7 +27210,7 @@ _sk_srcover_ssse3_lowp LABEL PROC PUBLIC _sk_dstover_ssse3_lowp _sk_dstover_ssse3_lowp LABEL PROC - DB 102,68,15,111,5,115,3,0,0 ; movdqa 0x373(%rip),%xmm8 # 1080 <_sk_xor__ssse3_lowp+0x18f> + DB 102,68,15,111,5,163,3,0,0 ; movdqa 0x3a3(%rip),%xmm8 # 13b0 <_sk_xor__ssse3_lowp+0x1bf> DB 102,68,15,249,199 ; psubw %xmm7,%xmm8 DB 102,65,15,56,11,192 ; pmulhrsw %xmm8,%xmm0 DB 102,15,56,29,192 ; pabsw %xmm0,%xmm0 @@ -27014,7 +27242,7 @@ _sk_modulate_ssse3_lowp LABEL PROC PUBLIC _sk_multiply_ssse3_lowp _sk_multiply_ssse3_lowp LABEL PROC - DB 102,68,15,111,5,8,3,0,0 ; movdqa 0x308(%rip),%xmm8 # 1090 <_sk_xor__ssse3_lowp+0x19f> + DB 102,68,15,111,5,56,3,0,0 ; movdqa 0x338(%rip),%xmm8 # 13c0 <_sk_xor__ssse3_lowp+0x1cf> DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,68,15,249,207 ; psubw %xmm7,%xmm9 DB 102,68,15,111,208 ; movdqa %xmm0,%xmm10 @@ -27070,7 +27298,7 @@ _sk_plus__ssse3_lowp LABEL PROC PUBLIC _sk_screen_ssse3_lowp _sk_screen_ssse3_lowp LABEL PROC - DB 102,68,15,111,5,26,2,0,0 ; movdqa 0x21a(%rip),%xmm8 # 10a0 <_sk_xor__ssse3_lowp+0x1af> + DB 102,68,15,111,5,74,2,0,0 ; movdqa 0x24a(%rip),%xmm8 # 13d0 <_sk_xor__ssse3_lowp+0x1df> DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,68,15,249,200 ; psubw %xmm0,%xmm9 DB 102,68,15,56,11,204 ; pmulhrsw %xmm4,%xmm9 @@ -27095,7 +27323,7 @@ _sk_screen_ssse3_lowp LABEL PROC PUBLIC _sk_xor__ssse3_lowp _sk_xor__ssse3_lowp LABEL PROC - DB 102,68,15,111,5,182,1,0,0 ; movdqa 0x1b6(%rip),%xmm8 # 10b0 <_sk_xor__ssse3_lowp+0x1bf> + DB 102,68,15,111,5,230,1,0,0 ; movdqa 0x1e6(%rip),%xmm8 # 13e0 <_sk_xor__ssse3_lowp+0x1ef> DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9 DB 102,68,15,249,207 ; psubw %xmm7,%xmm9 DB 102,65,15,56,11,193 ; pmulhrsw %xmm9,%xmm0 @@ -27130,7 +27358,8 @@ ALIGN 4 DB 0,71,0 ; add %al,0x0(%rdi) DB 0,0 ; add %al,(%rax) DB 71,0,0 ; rex.RXB add %r8b,(%r8) - DB 0 ; .byte 0x0 + DB 0,71,0 ; add %al,0x0(%rdi) + DB 0,0 ; add %al,(%rax) DB 71 ; rex.RXB ALIGN 16 @@ -27141,8 +27370,14 @@ ALIGN 16 DB 3,7 ; add (%rdi),%eax DB 11,15 ; or (%rdi),%ecx DB 129,128,129,128,129,128,129,128,129,128; addl $0x80818081,-0x7f7e7f7f(%rax) - DB 129,128,129,128,129,128,0,4,8,12 ; addl $0xc080400,-0x7f7e7f7f(%rax) - DB 1,5,9,13,2,6 ; add %eax,0x6020d09(%rip) # 6021cd3 <_sk_xor__ssse3_lowp+0x6020de2> + DB 129,128,129,128,129,128,129,128,129,128; addl $0x80818081,-0x7f7e7f7f(%rax) + DB 129,128,129,128,129,128,129,128,129,128; addl $0x80818081,-0x7f7e7f7f(%rax) + DB 129,128,129,128,129,128,129,128,129,128; addl $0x80818081,-0x7f7e7f7f(%rax) + DB 129,128,129,128,129,128,129,128,0,128; addl $0x80008081,-0x7f7e7f7f(%rax) + DB 0,128,0,128,0,128 ; add %al,-0x7fff8000(%rax) + DB 0,128,0,128,0,128 ; add %al,-0x7fff8000(%rax) + DB 0,128,0,4,8,12 ; add %al,0xc080400(%rax) + DB 1,5,9,13,2,6 ; add %eax,0x6020d09(%rip) # 6022003 <_sk_xor__ssse3_lowp+0x6020e12> DB 10,14 ; or (%rsi),%cl DB 3,7 ; add (%rdi),%eax DB 11,15 ; or (%rdi),%ecx diff --git a/src/jumper/SkJumper_stages_lowp.cpp b/src/jumper/SkJumper_stages_lowp.cpp index 38273e1dea..b8d3cdae47 100644 --- a/src/jumper/SkJumper_stages_lowp.cpp +++ b/src/jumper/SkJumper_stages_lowp.cpp @@ -170,6 +170,11 @@ SI U32 to_8888(F r, F g, F b, F a) { memcpy((uint32_t*)&px + 4, &hi, sizeof(hi)); return px; } +SI U8 to_byte(F v) { + // See to_8888() just above. + U16 packed = _mm_packus_epi16(v>>7, v>>7); // Doesn't really matter what we pack on top. + return unaligned_load(&packed); +} // Stages! @@ -181,6 +186,19 @@ STAGE(constant_color) { a = rgba[3]; } +STAGE(set_rgb) { + auto rgb = (const float*)ctx; + r = rgb[0]; + g = rgb[1]; + b = rgb[2]; +} + +STAGE(premul) { + r = r * a; + g = g * a; + b = b * a; +} + STAGE(load_8888) { auto ptr = *(const uint32_t**)ctx + x; from_8888(load(ptr, tail), &r,&g,&b,&a); @@ -190,6 +208,22 @@ STAGE(store_8888) { store(ptr, to_8888(r,g,b,a), tail); } +STAGE(load_a8) { + auto ptr = *(const uint8_t**)ctx + x; + r = g = b = 0.0f; + a = from_byte(load(ptr, tail)); +} +STAGE(store_a8) { + auto ptr = *(uint8_t**)ctx + x; + store(ptr, to_byte(a), tail); +} + +STAGE(load_g8) { + auto ptr = *(const uint8_t**)ctx + x; + r = g = b = from_byte(load(ptr, tail)); + a = 1.0f; +} + STAGE(srcover_rgba_8888) { auto ptr = *(uint32_t**)ctx + x;