diff options
author | 2017-04-06 10:22:26 -0400 | |
---|---|---|
committer | 2017-04-06 16:14:09 +0000 | |
commit | 14987ebb97f491f1b5bc776252b5ddbf65b8fca0 (patch) | |
tree | 404a2dc7c6e57766313bcb44161e820fb5526a9e /src/jumper/SkJumper_generated_win.S | |
parent | 8a823faeba2da8a77740b0cd9eaf1acf473067b1 (diff) |
jumper, add load_f32()
Change-Id: I71d85ffe29bc11678ff1e696fa4a2c93d0b4fcbe
Reviewed-on: https://skia-review.googlesource.com/11446
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/jumper/SkJumper_generated_win.S')
-rw-r--r-- | src/jumper/SkJumper_generated_win.S | 183 |
1 files changed, 166 insertions, 17 deletions
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S index 4ec22db65b..2fa69f5c8b 100644 --- a/src/jumper/SkJumper_generated_win.S +++ b/src/jumper/SkJumper_generated_win.S @@ -1607,7 +1607,7 @@ _sk_load_4444_hsw LABEL PROC DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) - DB 233,255,255,255,225 ; jmpq ffffffffe20016ec <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff563> + DB 233,255,255,255,225 ; jmpq ffffffffe20016ec <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff4a4> DB 255 ; (bad) DB 255 ; (bad) DB 255 ; (bad) @@ -2022,6 +2022,55 @@ _sk_store_u16_be_hsw LABEL PROC DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) DB 235,174 ; jmp 1cf3 <_sk_store_u16_be_hsw+0xef> +PUBLIC _sk_load_f32_hsw +_sk_load_f32_hsw LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,131,249,7 ; cmp $0x7,%rcx + DB 119,110 ; ja 1dbb <_sk_load_f32_hsw+0x76> + DB 76,139,0 ; mov (%rax),%r8 + DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 + DB 76,141,21,133,0,0,0 ; lea 0x85(%rip),%r10 # 1de4 <_sk_load_f32_hsw+0x9f> + DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,3,125,24,68,136,112,1 ; vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8 + DB 196,131,125,24,92,136,96,1 ; vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3 + DB 196,131,125,24,76,136,80,1 ; vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1 + DB 196,131,125,24,84,136,64,1 ; vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2 + DB 196,129,121,16,68,136,48 ; vmovupd 0x30(%r8,%r9,4),%xmm0 + DB 196,195,125,13,192,12 ; vblendpd $0xc,%ymm8,%ymm0,%ymm0 + DB 196,1,121,16,68,136,32 ; vmovupd 0x20(%r8,%r9,4),%xmm8 + DB 196,99,61,13,203,12 ; vblendpd $0xc,%ymm3,%ymm8,%ymm9 + DB 196,129,121,16,92,136,16 ; vmovupd 0x10(%r8,%r9,4),%xmm3 + DB 196,99,101,13,209,12 ; vblendpd $0xc,%ymm1,%ymm3,%ymm10 + DB 196,129,121,16,12,136 ; vmovupd (%r8,%r9,4),%xmm1 + DB 196,227,117,13,202,12 ; vblendpd $0xc,%ymm2,%ymm1,%ymm1 + DB 196,193,116,20,210 ; vunpcklps %ymm10,%ymm1,%ymm2 + DB 196,193,116,21,218 ; vunpckhps %ymm10,%ymm1,%ymm3 + DB 197,180,20,200 ; vunpcklps %ymm0,%ymm9,%ymm1 + DB 197,52,21,192 ; vunpckhps %ymm0,%ymm9,%ymm8 + DB 197,237,20,193 ; vunpcklpd %ymm1,%ymm2,%ymm0 + DB 197,237,21,201 ; vunpckhpd %ymm1,%ymm2,%ymm1 + DB 196,193,101,20,208 ; vunpcklpd %ymm8,%ymm3,%ymm2 + DB 196,193,101,21,216 ; vunpckhpd %ymm8,%ymm3,%ymm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 144 ; nop + DB 132,255 ; test %bh,%bh + DB 255 ; (bad) + DB 255,203 ; dec %ebx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 190,255,255,255,177 ; mov $0xb1ffffff,%esi + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,164,255,255,255,156,255 ; jmpq *-0x630001(%rdi,%rdi,8) + DB 255 ; (bad) + DB 255,148,255,255,255,140,255 ; callq *-0x730001(%rdi,%rdi,8) + DB 255 ; (bad) + DB 255 ; .byte 0xff + PUBLIC _sk_store_f32_hsw _sk_store_f32_hsw LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -2036,7 +2085,7 @@ _sk_store_f32_hsw LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 1db2 <_sk_store_f32_hsw+0x6d> + DB 117,55 ; jne 1e71 <_sk_store_f32_hsw+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -2049,22 +2098,22 @@ _sk_store_f32_hsw LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 1dae <_sk_store_f32_hsw+0x69> + DB 116,240 ; je 1e6d <_sk_store_f32_hsw+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 1dae <_sk_store_f32_hsw+0x69> + DB 114,227 ; jb 1e6d <_sk_store_f32_hsw+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 1dae <_sk_store_f32_hsw+0x69> + DB 116,218 ; je 1e6d <_sk_store_f32_hsw+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 1dae <_sk_store_f32_hsw+0x69> + DB 114,205 ; jb 1e6d <_sk_store_f32_hsw+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 1dae <_sk_store_f32_hsw+0x69> + DB 116,195 ; je 1e6d <_sk_store_f32_hsw+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 1dae <_sk_store_f32_hsw+0x69> + DB 114,181 ; jb 1e6d <_sk_store_f32_hsw+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 1dae <_sk_store_f32_hsw+0x69> + DB 235,171 ; jmp 1e6d <_sk_store_f32_hsw+0x69> PUBLIC _sk_clamp_x_hsw _sk_clamp_x_hsw LABEL PROC @@ -4804,6 +4853,56 @@ _sk_store_u16_be_avx LABEL PROC DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8) DB 235,174 ; jmp 25f6 <_sk_store_u16_be_avx+0xf6> +PUBLIC _sk_load_f32_avx +_sk_load_f32_avx LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,131,249,7 ; cmp $0x7,%rcx + DB 119,110 ; ja 26be <_sk_load_f32_avx+0x76> + DB 76,139,0 ; mov (%rax),%r8 + DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9 + DB 76,141,21,134,0,0,0 ; lea 0x86(%rip),%r10 # 26e8 <_sk_load_f32_avx+0xa0> + DB 73,99,4,138 ; movslq (%r10,%rcx,4),%rax + DB 76,1,208 ; add %r10,%rax + DB 255,224 ; jmpq *%rax + DB 196,3,125,24,68,136,112,1 ; vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8 + DB 196,131,125,24,92,136,96,1 ; vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3 + DB 196,131,125,24,76,136,80,1 ; vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1 + DB 196,131,125,24,84,136,64,1 ; vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2 + DB 196,129,121,16,68,136,48 ; vmovupd 0x30(%r8,%r9,4),%xmm0 + DB 196,195,125,13,192,12 ; vblendpd $0xc,%ymm8,%ymm0,%ymm0 + DB 196,1,121,16,68,136,32 ; vmovupd 0x20(%r8,%r9,4),%xmm8 + DB 196,99,61,13,203,12 ; vblendpd $0xc,%ymm3,%ymm8,%ymm9 + DB 196,129,121,16,92,136,16 ; vmovupd 0x10(%r8,%r9,4),%xmm3 + DB 196,99,101,13,209,12 ; vblendpd $0xc,%ymm1,%ymm3,%ymm10 + DB 196,129,121,16,12,136 ; vmovupd (%r8,%r9,4),%xmm1 + DB 196,227,117,13,202,12 ; vblendpd $0xc,%ymm2,%ymm1,%ymm1 + DB 196,193,116,20,210 ; vunpcklps %ymm10,%ymm1,%ymm2 + DB 196,193,116,21,218 ; vunpckhps %ymm10,%ymm1,%ymm3 + DB 197,180,20,200 ; vunpcklps %ymm0,%ymm9,%ymm1 + DB 197,52,21,192 ; vunpckhps %ymm0,%ymm9,%ymm8 + DB 197,237,20,193 ; vunpcklpd %ymm1,%ymm2,%ymm0 + DB 197,237,21,201 ; vunpckhpd %ymm1,%ymm2,%ymm1 + DB 196,193,101,20,208 ; vunpcklpd %ymm8,%ymm3,%ymm2 + DB 196,193,101,21,216 ; vunpckhpd %ymm8,%ymm3,%ymm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + DB 102,144 ; xchg %ax,%ax + DB 131,255,255 ; cmp $0xffffffff,%edi + DB 255,202 ; dec %edx + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; (bad) + DB 189,255,255,255,176 ; mov $0xb0ffffff,%ebp + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,163,255,255,255,155 ; jmpq *-0x64000001(%rbx) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255,147,255,255,255,139 ; callq *-0x74000001(%rbx) + DB 255 ; (bad) + DB 255 ; (bad) + DB 255 ; .byte 0xff + PUBLIC _sk_store_f32_avx _sk_store_f32_avx LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -4818,7 +4917,7 @@ _sk_store_f32_avx LABEL PROC DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8 DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11 DB 72,133,201 ; test %rcx,%rcx - DB 117,55 ; jne 26b5 <_sk_store_f32_avx+0x6d> + DB 117,55 ; jne 2775 <_sk_store_f32_avx+0x6d> DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 @@ -4831,22 +4930,22 @@ _sk_store_f32_avx LABEL PROC DB 255,224 ; jmpq *%rax DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4) DB 72,131,249,1 ; cmp $0x1,%rcx - DB 116,240 ; je 26b1 <_sk_store_f32_avx+0x69> + DB 116,240 ; je 2771 <_sk_store_f32_avx+0x69> DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4) DB 72,131,249,3 ; cmp $0x3,%rcx - DB 114,227 ; jb 26b1 <_sk_store_f32_avx+0x69> + DB 114,227 ; jb 2771 <_sk_store_f32_avx+0x69> DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4) - DB 116,218 ; je 26b1 <_sk_store_f32_avx+0x69> + DB 116,218 ; je 2771 <_sk_store_f32_avx+0x69> DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4) DB 72,131,249,5 ; cmp $0x5,%rcx - DB 114,205 ; jb 26b1 <_sk_store_f32_avx+0x69> + DB 114,205 ; jb 2771 <_sk_store_f32_avx+0x69> DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) - DB 116,195 ; je 26b1 <_sk_store_f32_avx+0x69> + DB 116,195 ; je 2771 <_sk_store_f32_avx+0x69> DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) DB 72,131,249,7 ; cmp $0x7,%rcx - DB 114,181 ; jb 26b1 <_sk_store_f32_avx+0x69> + DB 114,181 ; jb 2771 <_sk_store_f32_avx+0x69> DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) - DB 235,171 ; jmp 26b1 <_sk_store_f32_avx+0x69> + DB 235,171 ; jmp 2771 <_sk_store_f32_avx+0x69> PUBLIC _sk_clamp_x_avx _sk_clamp_x_avx LABEL PROC @@ -7088,6 +7187,31 @@ _sk_store_u16_be_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_load_f32_sse41 +_sk_load_f32_sse41 LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,139,0 ; mov (%rax),%rax + DB 72,137,249 ; mov %rdi,%rcx + DB 72,193,225,4 ; shl $0x4,%rcx + DB 68,15,16,4,8 ; movups (%rax,%rcx,1),%xmm8 + DB 15,16,68,8,16 ; movups 0x10(%rax,%rcx,1),%xmm0 + DB 15,16,92,8,32 ; movups 0x20(%rax,%rcx,1),%xmm3 + DB 68,15,16,76,8,48 ; movups 0x30(%rax,%rcx,1),%xmm9 + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 65,15,20,201 ; unpcklps %xmm9,%xmm1 + DB 68,15,21,192 ; unpckhps %xmm0,%xmm8 + DB 65,15,21,217 ; unpckhps %xmm9,%xmm3 + DB 15,40,194 ; movaps %xmm2,%xmm0 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 15,18,202 ; movhlps %xmm2,%xmm1 + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 65,15,18,216 ; movhlps %xmm8,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_store_f32_sse41 _sk_store_f32_sse41 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax @@ -9477,6 +9601,31 @@ _sk_store_u16_be_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax DB 255,224 ; jmpq *%rax +PUBLIC _sk_load_f32_sse2 +_sk_load_f32_sse2 LABEL PROC + DB 72,173 ; lods %ds:(%rsi),%rax + DB 72,139,0 ; mov (%rax),%rax + DB 72,137,249 ; mov %rdi,%rcx + DB 72,193,225,4 ; shl $0x4,%rcx + DB 68,15,16,4,8 ; movups (%rax,%rcx,1),%xmm8 + DB 15,16,68,8,16 ; movups 0x10(%rax,%rcx,1),%xmm0 + DB 15,16,92,8,32 ; movups 0x20(%rax,%rcx,1),%xmm3 + DB 68,15,16,76,8,48 ; movups 0x30(%rax,%rcx,1),%xmm9 + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 15,20,208 ; unpcklps %xmm0,%xmm2 + DB 15,40,203 ; movaps %xmm3,%xmm1 + DB 65,15,20,201 ; unpcklps %xmm9,%xmm1 + DB 68,15,21,192 ; unpckhps %xmm0,%xmm8 + DB 65,15,21,217 ; unpckhps %xmm9,%xmm3 + DB 15,40,194 ; movaps %xmm2,%xmm0 + DB 102,15,20,193 ; unpcklpd %xmm1,%xmm0 + DB 15,18,202 ; movhlps %xmm2,%xmm1 + DB 65,15,40,208 ; movaps %xmm8,%xmm2 + DB 102,15,20,211 ; unpcklpd %xmm3,%xmm2 + DB 65,15,18,216 ; movhlps %xmm8,%xmm3 + DB 72,173 ; lods %ds:(%rsi),%rax + DB 255,224 ; jmpq *%rax + PUBLIC _sk_store_f32_sse2 _sk_store_f32_sse2 LABEL PROC DB 72,173 ; lods %ds:(%rsi),%rax |