aboutsummaryrefslogtreecommitdiffhomepage
path: root/src
diff options
context:
space:
mode:
authorGravatar Mike Klein <mtklein@chromium.org>2017-03-31 13:52:45 -0400
committerGravatar Skia Commit-Bot <skia-commit-bot@chromium.org>2017-03-31 18:47:41 +0000
commitf809fef8280e3d6ad9d95697cd234560b49962ab (patch)
tree201f2b574edbd42f4a55c18ba22441e0afcdc97d /src
parent15bf678e08d583ebd92839c0792a5b8d89557a8c (diff)
jumper, a couple simple loads and stores
Change-Id: I217d7b562f5fa443978044e17469ba757c061209 Reviewed-on: https://skia-review.googlesource.com/10971 Reviewed-by: Mike Klein <mtklein@chromium.org> Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src')
-rw-r--r--src/jumper/SkJumper.cpp3
-rw-r--r--src/jumper/SkJumper_generated.S991
-rw-r--r--src/jumper/SkJumper_generated_win.S783
-rw-r--r--src/jumper/SkJumper_stages.cpp28
4 files changed, 1624 insertions, 181 deletions
diff --git a/src/jumper/SkJumper.cpp b/src/jumper/SkJumper.cpp
index ebb014c5b4..0ed8bd8e6c 100644
--- a/src/jumper/SkJumper.cpp
+++ b/src/jumper/SkJumper.cpp
@@ -86,8 +86,11 @@ static K kConstants = {
M(load_tables) \
M(load_a8) \
M(store_a8) \
+ M(load_g8) \
M(load_565) \
M(store_565) \
+ M(load_4444) \
+ M(store_4444) \
M(load_8888) \
M(store_8888) \
M(load_f16) \
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S
index 3f1f86b34e..b0e8f6eb0e 100644
--- a/src/jumper/SkJumper_generated.S
+++ b/src/jumper/SkJumper_generated.S
@@ -1169,6 +1169,31 @@ _sk_store_a8_aarch64:
.long 0x91004021 // add x1, x1, #0x10
.long 0xd61f0060 // br x3
+HIDDEN _sk_load_g8_aarch64
+.globl _sk_load_g8_aarch64
+_sk_load_g8_aarch64:
+ .long 0xa8c10c28 // ldp x8, x3, [x1], #16
+ .long 0x52a77009 // mov w9, #0x3b800000
+ .long 0x72901029 // movk w9, #0x8081
+ .long 0x4e040d20 // dup v0.4s, w9
+ .long 0xf9400108 // ldr x8, [x8]
+ .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
+ .long 0x8b000108 // add x8, x8, x0
+ .long 0x3940010a // ldrb w10, [x8]
+ .long 0x39400509 // ldrb w9, [x8, #1]
+ .long 0x3940090b // ldrb w11, [x8, #2]
+ .long 0x39400d08 // ldrb w8, [x8, #3]
+ .long 0x4e021d41 // mov v1.h[0], w10
+ .long 0x4e061d21 // mov v1.h[1], w9
+ .long 0x4e0a1d61 // mov v1.h[2], w11
+ .long 0x4e0e1d01 // mov v1.h[3], w8
+ .long 0x2f10a421 // uxtl v1.4s, v1.4h
+ .long 0x6e21d821 // ucvtf v1.4s, v1.4s
+ .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
+ .long 0x4ea01c01 // mov v1.16b, v0.16b
+ .long 0x4ea01c02 // mov v2.16b, v0.16b
+ .long 0xd61f0060 // br x3
+
HIDDEN _sk_load_565_aarch64
.globl _sk_load_565_aarch64
_sk_load_565_aarch64:
@@ -1227,6 +1252,71 @@ _sk_store_565_aarch64:
.long 0x91004021 // add x1, x1, #0x10
.long 0xd61f0060 // br x3
+HIDDEN _sk_load_4444_aarch64
+.globl _sk_load_4444_aarch64
+_sk_load_4444_aarch64:
+ .long 0xa8c10c28 // ldp x8, x3, [x1], #16
+ .long 0xd37ff809 // lsl x9, x0, #1
+ .long 0x4f072601 // movi v1.4s, #0xf0, lsl #8
+ .long 0x4f0025e2 // movi v2.4s, #0xf, lsl #8
+ .long 0xf9400108 // ldr x8, [x8]
+ .long 0x4f070603 // movi v3.4s, #0xf0
+ .long 0x4f0005f0 // movi v16.4s, #0xf
+ .long 0xfc696900 // ldr d0, [x8, x9]
+ .long 0x52a6f108 // mov w8, #0x37880000
+ .long 0x72911128 // movk w8, #0x8889
+ .long 0x2f10a400 // uxtl v0.4s, v0.4h
+ .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b
+ .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b
+ .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b
+ .long 0x4e301c10 // and v16.16b, v0.16b, v16.16b
+ .long 0x4e040d00 // dup v0.4s, w8
+ .long 0x52a73108 // mov w8, #0x39880000
+ .long 0x72911128 // movk w8, #0x8889
+ .long 0x4e21d821 // scvtf v1.4s, v1.4s
+ .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
+ .long 0x4e040d01 // dup v1.4s, w8
+ .long 0x52a77108 // mov w8, #0x3b880000
+ .long 0x72911128 // movk w8, #0x8889
+ .long 0x4e21d842 // scvtf v2.4s, v2.4s
+ .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s
+ .long 0x4e040d02 // dup v2.4s, w8
+ .long 0x52a7b108 // mov w8, #0x3d880000
+ .long 0x72911128 // movk w8, #0x8889
+ .long 0x4e21d863 // scvtf v3.4s, v3.4s
+ .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s
+ .long 0x4e040d03 // dup v3.4s, w8
+ .long 0x4e21da10 // scvtf v16.4s, v16.4s
+ .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s
+ .long 0xd61f0060 // br x3
+
+HIDDEN _sk_store_4444_aarch64
+.globl _sk_store_4444_aarch64
+_sk_store_4444_aarch64:
+ .long 0x4f01f5d0 // fmov v16.4s, #1.500000000000000000e+01
+ .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s
+ .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s
+ .long 0xf9400028 // ldr x8, [x1]
+ .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
+ .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
+ .long 0x4f2c5631 // shl v17.4s, v17.4s, #12
+ .long 0x4f285652 // shl v18.4s, v18.4s, #8
+ .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b
+ .long 0x6e30dc52 // fmul v18.4s, v2.4s, v16.4s
+ .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
+ .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s
+ .long 0x4f245652 // shl v18.4s, v18.4s, #4
+ .long 0xf9400108 // ldr x8, [x8]
+ .long 0x4eb21e31 // orr v17.16b, v17.16b, v18.16b
+ .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s
+ .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b
+ .long 0xd37ff809 // lsl x9, x0, #1
+ .long 0x0e612a10 // xtn v16.4h, v16.4s
+ .long 0xfc296910 // str d16, [x8, x9]
+ .long 0xf9400423 // ldr x3, [x1, #8]
+ .long 0x91004021 // add x1, x1, #0x10
+ .long 0xd61f0060 // br x3
+
HIDDEN _sk_load_8888_aarch64
.globl _sk_load_8888_aarch64
_sk_load_8888_aarch64:
@@ -2797,6 +2887,32 @@ _sk_store_a8_vfp4:
.long 0x437f0000 // .word 0x437f0000
.long 0x437f0000 // .word 0x437f0000
+HIDDEN _sk_load_g8_vfp4
+.globl _sk_load_g8_vfp4
+_sk_load_g8_vfp4:
+ .long 0xe24dd004 // sub sp, sp, #4
+ .long 0xe8911008 // ldm r1, {r3, ip}
+ .long 0xe2811008 // add r1, r1, #8
+ .long 0xf2873f10 // vmov.f32 d3, #1
+ .long 0xe5933000 // ldr r3, [r3]
+ .long 0xe0833000 // add r3, r3, r0
+ .long 0xe1d330b0 // ldrh r3, [r3]
+ .long 0xe1cd30b0 // strh r3, [sp]
+ .long 0xe1a0300d // mov r3, sp
+ .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16]
+ .long 0xf3c80a30 // vmovl.u8 q8, d16
+ .long 0xf3d00a30 // vmovl.u16 q8, d16
+ .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
+ .long 0xeddf1b05 // vldr d17, [pc, #20]
+ .long 0xf3000db1 // vmul.f32 d0, d16, d17
+ .long 0xf2201110 // vorr d1, d0, d0
+ .long 0xf2202110 // vorr d2, d0, d0
+ .long 0xe28dd004 // add sp, sp, #4
+ .long 0xe12fff1c // bx ip
+ .long 0xe320f000 // nop {0}
+ .long 0x3b808081 // .word 0x3b808081
+ .long 0x3b808081 // .word 0x3b808081
+
HIDDEN _sk_load_565_vfp4
.globl _sk_load_565_vfp4
_sk_load_565_vfp4:
@@ -2865,6 +2981,82 @@ _sk_store_565_vfp4:
.long 0x427c0000 // .word 0x427c0000
.long 0x427c0000 // .word 0x427c0000
+HIDDEN _sk_load_4444_vfp4
+.globl _sk_load_4444_vfp4
+_sk_load_4444_vfp4:
+ .long 0xe24dd004 // sub sp, sp, #4
+ .long 0xe8911008 // ldm r1, {r3, ip}
+ .long 0xf3c71210 // vmov.i32 d17, #61440
+ .long 0xf3c74010 // vmov.i32 d20, #240
+ .long 0xf2c0501f // vmov.i32 d21, #15
+ .long 0xeddf6b1d // vldr d22, [pc, #116]
+ .long 0xe5933000 // ldr r3, [r3]
+ .long 0xe2811008 // add r1, r1, #8
+ .long 0xe7933080 // ldr r3, [r3, r0, lsl #1]
+ .long 0xe58d3000 // str r3, [sp]
+ .long 0xe1a0300d // mov r3, sp
+ .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32]
+ .long 0xf3d02a30 // vmovl.u16 q9, d16
+ .long 0xf2c0021f // vmov.i32 d16, #3840
+ .long 0xf24211b1 // vand d17, d18, d17
+ .long 0xf24201b0 // vand d16, d18, d16
+ .long 0xf24241b4 // vand d20, d18, d20
+ .long 0xf24221b5 // vand d18, d18, d21
+ .long 0xeddf3b0c // vldr d19, [pc, #48]
+ .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
+ .long 0xeddf5b0c // vldr d21, [pc, #48]
+ .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
+ .long 0xf3fb4624 // vcvt.f32.s32 d20, d20
+ .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
+ .long 0xf3010db3 // vmul.f32 d0, d17, d19
+ .long 0xeddf1b0b // vldr d17, [pc, #44]
+ .long 0xf3001db5 // vmul.f32 d1, d16, d21
+ .long 0xf3042db6 // vmul.f32 d2, d20, d22
+ .long 0xf3023db1 // vmul.f32 d3, d18, d17
+ .long 0xe28dd004 // add sp, sp, #4
+ .long 0xe12fff1c // bx ip
+ .long 0xe320f000 // nop {0}
+ .long 0x37888889 // .word 0x37888889
+ .long 0x37888889 // .word 0x37888889
+ .long 0x39888889 // .word 0x39888889
+ .long 0x39888889 // .word 0x39888889
+ .long 0x3b888889 // .word 0x3b888889
+ .long 0x3b888889 // .word 0x3b888889
+ .long 0x3d888889 // .word 0x3d888889
+ .long 0x3d888889 // .word 0x3d888889
+
+HIDDEN _sk_store_4444_vfp4
+.globl _sk_store_4444_vfp4
+_sk_store_4444_vfp4:
+ .long 0xf2c20f1e // vmov.f32 d16, #15
+ .long 0xe5913000 // ldr r3, [r1]
+ .long 0xf2c3261f // vmov.i32 d18, #1056964608
+ .long 0xf2c3361f // vmov.i32 d19, #1056964608
+ .long 0xf2402c30 // vfma.f32 d18, d0, d16
+ .long 0xe5933000 // ldr r3, [r3]
+ .long 0xf2413c30 // vfma.f32 d19, d1, d16
+ .long 0xf2c3461f // vmov.i32 d20, #1056964608
+ .long 0xe0833080 // add r3, r3, r0, lsl #1
+ .long 0xf2424c30 // vfma.f32 d20, d2, d16
+ .long 0xf2c3161f // vmov.i32 d17, #1056964608
+ .long 0xf2431c30 // vfma.f32 d17, d3, d16
+ .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18
+ .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19
+ .long 0xf3fb37a4 // vcvt.u32.f32 d19, d20
+ .long 0xf2ec0530 // vshl.s32 d16, d16, #12
+ .long 0xf2e82532 // vshl.s32 d18, d18, #8
+ .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17
+ .long 0xf2e43533 // vshl.s32 d19, d19, #4
+ .long 0xf26201b0 // vorr d16, d18, d16
+ .long 0xf26001b3 // vorr d16, d16, d19
+ .long 0xf26001b1 // vorr d16, d16, d17
+ .long 0xf3f60121 // vuzp.16 d16, d17
+ .long 0xf4c3080f // vst1.32 {d16[0]}, [r3]
+ .long 0xe2813008 // add r3, r1, #8
+ .long 0xe591c004 // ldr ip, [r1, #4]
+ .long 0xe1a01003 // mov r1, r3
+ .long 0xe12fff1c // bx ip
+
HIDDEN _sk_load_8888_vfp4
.globl _sk_load_8888_vfp4
_sk_load_8888_vfp4:
@@ -4457,7 +4649,7 @@ _sk_lerp_565_hsw:
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 233,255,255,255,225 // jmpq ffffffffe2000f54 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff4f0>
+ .byte 233,255,255,255,225 // jmpq ffffffffe2000f54 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff248>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
@@ -4617,13 +4809,50 @@ _sk_store_a8_hsw:
.byte 255 // (bad)
.byte 255 // .byte 0xff
+HIDDEN _sk_load_g8_hsw
+.globl _sk_load_g8_hsw
+_sk_load_g8_hsw:
+ .byte 73,137,200 // mov %rcx,%r8
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 72,1,248 // add %rdi,%rax
+ .byte 77,133,192 // test %r8,%r8
+ .byte 117,60 // jne 118c <_sk_load_g8_hsw+0x4c>
+ .byte 197,250,126,0 // vmovq (%rax),%xmm0
+ .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
+ .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
+ .byte 184,129,128,128,59 // mov $0x3b808081,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
+ .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
+ .byte 184,0,0,128,63 // mov $0x3f800000,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,226,125,88,217 // vpbroadcastd %xmm1,%ymm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 76,137,193 // mov %r8,%rcx
+ .byte 197,252,40,200 // vmovaps %ymm0,%ymm1
+ .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
+ .byte 255,224 // jmpq *%rax
+ .byte 49,201 // xor %ecx,%ecx
+ .byte 77,137,194 // mov %r8,%r10
+ .byte 69,49,201 // xor %r9d,%r9d
+ .byte 68,15,182,24 // movzbl (%rax),%r11d
+ .byte 72,255,192 // inc %rax
+ .byte 73,211,227 // shl %cl,%r11
+ .byte 77,9,217 // or %r11,%r9
+ .byte 72,131,193,8 // add $0x8,%rcx
+ .byte 73,255,202 // dec %r10
+ .byte 117,234 // jne 1194 <_sk_load_g8_hsw+0x54>
+ .byte 196,193,249,110,193 // vmovq %r9,%xmm0
+ .byte 235,163 // jmp 1154 <_sk_load_g8_hsw+0x14>
+
HIDDEN _sk_load_565_hsw
.globl _sk_load_565_hsw
_sk_load_565_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,149,0,0,0 // jne 11e3 <_sk_load_565_hsw+0xa3>
+ .byte 15,133,149,0,0,0 // jne 1254 <_sk_load_565_hsw+0xa3>
.byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
.byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2
.byte 184,0,248,0,0 // mov $0xf800,%eax
@@ -4663,9 +4892,9 @@ _sk_load_565_hsw:
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,89,255,255,255 // ja 1154 <_sk_load_565_hsw+0x14>
+ .byte 15,135,89,255,255,255 // ja 11c5 <_sk_load_565_hsw+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 1250 <_sk_load_565_hsw+0x110>
+ .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 12c0 <_sk_load_565_hsw+0x10f>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -4677,27 +4906,27 @@ _sk_load_565_hsw:
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- .byte 233,5,255,255,255 // jmpq 1154 <_sk_load_565_hsw+0x14>
- .byte 144 // nop
- .byte 243,255 // repz (bad)
+ .byte 233,5,255,255,255 // jmpq 11c5 <_sk_load_565_hsw+0x14>
+ .byte 244 // hlt
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 235,255 // jmp 1255 <_sk_load_565_hsw+0x115>
.byte 255 // (bad)
- .byte 255,227 // jmpq *%rbx
+ .byte 236 // in (%dx),%al
.byte 255 // (bad)
.byte 255 // (bad)
+ .byte 255,228 // jmpq *%rsp
.byte 255 // (bad)
- .byte 219,255 // (bad)
.byte 255 // (bad)
- .byte 255,211 // callq *%rbx
.byte 255 // (bad)
+ .byte 220,255 // fdivr %st,%st(7)
.byte 255 // (bad)
- .byte 255,203 // dec %ebx
+ .byte 255,212 // callq *%rsp
.byte 255 // (bad)
.byte 255 // (bad)
+ .byte 255,204 // dec %esp
.byte 255 // (bad)
- .byte 191 // .byte 0xbf
+ .byte 255 // (bad)
+ .byte 255,192 // inc %eax
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // .byte 0xff
@@ -4726,7 +4955,7 @@ _sk_store_565_hsw:
.byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 12d8 <_sk_store_565_hsw+0x6c>
+ .byte 117,10 // jne 1348 <_sk_store_565_hsw+0x6c>
.byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -4734,9 +4963,9 @@ _sk_store_565_hsw:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 12d4 <_sk_store_565_hsw+0x68>
+ .byte 119,236 // ja 1344 <_sk_store_565_hsw+0x68>
.byte 65,15,182,192 // movzbl %r8b,%eax
- .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 1338 <_sk_store_565_hsw+0xcc>
+ .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 13a8 <_sk_store_565_hsw+0xcc>
.byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
.byte 76,1,192 // add %r8,%rax
.byte 255,224 // jmpq *%rax
@@ -4747,7 +4976,7 @@ _sk_store_565_hsw:
.byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
.byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
.byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
- .byte 235,159 // jmp 12d4 <_sk_store_565_hsw+0x68>
+ .byte 235,159 // jmp 1344 <_sk_store_565_hsw+0x68>
.byte 15,31,0 // nopl (%rax)
.byte 244 // hlt
.byte 255 // (bad)
@@ -4773,6 +5002,167 @@ _sk_store_565_hsw:
.byte 255 // (bad)
.byte 255 // .byte 0xff
+HIDDEN _sk_load_4444_hsw
+.globl _sk_load_4444_hsw
+_sk_load_4444_hsw:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 76,139,16 // mov (%rax),%r10
+ .byte 72,133,201 // test %rcx,%rcx
+ .byte 15,133,179,0,0,0 // jne 1485 <_sk_load_4444_hsw+0xc1>
+ .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
+ .byte 196,98,125,51,200 // vpmovzxwd %xmm0,%ymm9
+ .byte 184,0,240,0,0 // mov $0xf000,%eax
+ .byte 197,249,110,192 // vmovd %eax,%xmm0
+ .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
+ .byte 196,193,125,219,193 // vpand %ymm9,%ymm0,%ymm0
+ .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
+ .byte 184,137,136,136,55 // mov $0x37888889,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
+ .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
+ .byte 184,0,15,0,0 // mov $0xf00,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
+ .byte 196,193,117,219,201 // vpand %ymm9,%ymm1,%ymm1
+ .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
+ .byte 184,137,136,136,57 // mov $0x39888889,%eax
+ .byte 197,249,110,208 // vmovd %eax,%xmm2
+ .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
+ .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
+ .byte 184,240,0,0,0 // mov $0xf0,%eax
+ .byte 197,249,110,208 // vmovd %eax,%xmm2
+ .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
+ .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2
+ .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8
+ .byte 184,137,136,136,59 // mov $0x3b888889,%eax
+ .byte 197,249,110,208 // vmovd %eax,%xmm2
+ .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
+ .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
+ .byte 184,15,0,0,0 // mov $0xf,%eax
+ .byte 197,249,110,216 // vmovd %eax,%xmm3
+ .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
+ .byte 196,193,101,219,217 // vpand %ymm9,%ymm3,%ymm3
+ .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
+ .byte 184,137,136,136,61 // mov $0x3d888889,%eax
+ .byte 197,249,110,216 // vmovd %eax,%xmm3
+ .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
+ .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,137,200 // mov %ecx,%r8d
+ .byte 65,128,224,7 // and $0x7,%r8b
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 65,254,200 // dec %r8b
+ .byte 65,128,248,6 // cmp $0x6,%r8b
+ .byte 15,135,59,255,255,255 // ja 13d8 <_sk_load_4444_hsw+0x14>
+ .byte 69,15,182,192 // movzbl %r8b,%r8d
+ .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 14f4 <_sk_load_4444_hsw+0x130>
+ .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
+ .byte 76,1,200 // add %r9,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 233,231,254,255,255 // jmpq 13d8 <_sk_load_4444_hsw+0x14>
+ .byte 15,31,0 // nopl (%rax)
+ .byte 241 // icebp
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 233,255,255,255,225 // jmpq ffffffffe20014fc <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff7f0>
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 217,255 // fcos
+ .byte 255 // (bad)
+ .byte 255,209 // callq *%rcx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,201 // dec %ecx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 189 // .byte 0xbd
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
+
+HIDDEN _sk_store_4444_hsw
+.globl _sk_store_4444_hsw
+_sk_store_4444_hsw:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 76,139,8 // mov (%rax),%r9
+ .byte 184,0,0,112,65 // mov $0x41700000,%eax
+ .byte 197,121,110,192 // vmovd %eax,%xmm8
+ .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
+ .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
+ .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
+ .byte 196,193,53,114,241,12 // vpslld $0xc,%ymm9,%ymm9
+ .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
+ .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
+ .byte 196,193,45,114,242,8 // vpslld $0x8,%ymm10,%ymm10
+ .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9
+ .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10
+ .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
+ .byte 196,193,45,114,242,4 // vpslld $0x4,%ymm10,%ymm10
+ .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
+ .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
+ .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8
+ .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
+ .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
+ .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
+ .byte 72,133,201 // test %rcx,%rcx
+ .byte 117,10 // jne 1582 <_sk_store_4444_hsw+0x72>
+ .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,137,200 // mov %ecx,%r8d
+ .byte 65,128,224,7 // and $0x7,%r8b
+ .byte 65,254,200 // dec %r8b
+ .byte 65,128,248,6 // cmp $0x6,%r8b
+ .byte 119,236 // ja 157e <_sk_store_4444_hsw+0x6e>
+ .byte 65,15,182,192 // movzbl %r8b,%eax
+ .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 15e0 <_sk_store_4444_hsw+0xd0>
+ .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
+ .byte 76,1,192 // add %r8,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
+ .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
+ .byte 235,159 // jmp 157e <_sk_store_4444_hsw+0x6e>
+ .byte 144 // nop
+ .byte 246,255 // idiv %bh
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 238 // out %al,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,230 // jmpq *%rsi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 222,255 // fdivrp %st,%st(7)
+ .byte 255 // (bad)
+ .byte 255,214 // callq *%rsi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,206 // dec %esi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,198 // inc %esi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
+
HIDDEN _sk_load_8888_hsw
.globl _sk_load_8888_hsw
_sk_load_8888_hsw:
@@ -4781,7 +5171,7 @@ _sk_load_8888_hsw:
.byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
.byte 76,3,8 // add (%rax),%r9
.byte 77,133,192 // test %r8,%r8
- .byte 117,104 // jne 13d1 <_sk_load_8888_hsw+0x7d>
+ .byte 117,104 // jne 1679 <_sk_load_8888_hsw+0x7d>
.byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
.byte 184,255,0,0,0 // mov $0xff,%eax
.byte 197,249,110,192 // vmovd %eax,%xmm0
@@ -4814,7 +5204,7 @@ _sk_load_8888_hsw:
.byte 196,225,249,110,192 // vmovq %rax,%xmm0
.byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
.byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
- .byte 233,116,255,255,255 // jmpq 136e <_sk_load_8888_hsw+0x1a>
+ .byte 233,116,255,255,255 // jmpq 1616 <_sk_load_8888_hsw+0x1a>
HIDDEN _sk_store_8888_hsw
.globl _sk_store_8888_hsw
@@ -4841,7 +5231,7 @@ _sk_store_8888_hsw:
.byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8
.byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
.byte 77,133,192 // test %r8,%r8
- .byte 117,12 // jne 146e <_sk_store_8888_hsw+0x74>
+ .byte 117,12 // jne 1716 <_sk_store_8888_hsw+0x74>
.byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,137,193 // mov %r8,%rcx
@@ -4854,7 +5244,7 @@ _sk_store_8888_hsw:
.byte 196,97,249,110,200 // vmovq %rax,%xmm9
.byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9
.byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9)
- .byte 235,211 // jmp 1467 <_sk_store_8888_hsw+0x6d>
+ .byte 235,211 // jmp 170f <_sk_store_8888_hsw+0x6d>
HIDDEN _sk_load_f16_hsw
.globl _sk_load_f16_hsw
@@ -4862,7 +5252,7 @@ _sk_load_f16_hsw:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,97 // jne 14ff <_sk_load_f16_hsw+0x6b>
+ .byte 117,97 // jne 17a7 <_sk_load_f16_hsw+0x6b>
.byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8
.byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
.byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3
@@ -4888,29 +5278,29 @@ _sk_load_f16_hsw:
.byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,79 // je 155e <_sk_load_f16_hsw+0xca>
+ .byte 116,79 // je 1806 <_sk_load_f16_hsw+0xca>
.byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,67 // jb 155e <_sk_load_f16_hsw+0xca>
+ .byte 114,67 // jb 1806 <_sk_load_f16_hsw+0xca>
.byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,68 // je 156b <_sk_load_f16_hsw+0xd7>
+ .byte 116,68 // je 1813 <_sk_load_f16_hsw+0xd7>
.byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,56 // jb 156b <_sk_load_f16_hsw+0xd7>
+ .byte 114,56 // jb 1813 <_sk_load_f16_hsw+0xd7>
.byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,114,255,255,255 // je 14b5 <_sk_load_f16_hsw+0x21>
+ .byte 15,132,114,255,255,255 // je 175d <_sk_load_f16_hsw+0x21>
.byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,98,255,255,255 // jb 14b5 <_sk_load_f16_hsw+0x21>
+ .byte 15,130,98,255,255,255 // jb 175d <_sk_load_f16_hsw+0x21>
.byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9
- .byte 233,87,255,255,255 // jmpq 14b5 <_sk_load_f16_hsw+0x21>
+ .byte 233,87,255,255,255 // jmpq 175d <_sk_load_f16_hsw+0x21>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,74,255,255,255 // jmpq 14b5 <_sk_load_f16_hsw+0x21>
+ .byte 233,74,255,255,255 // jmpq 175d <_sk_load_f16_hsw+0x21>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,65,255,255,255 // jmpq 14b5 <_sk_load_f16_hsw+0x21>
+ .byte 233,65,255,255,255 // jmpq 175d <_sk_load_f16_hsw+0x21>
HIDDEN _sk_store_f16_hsw
.globl _sk_store_f16_hsw
@@ -4930,7 +5320,7 @@ _sk_store_f16_hsw:
.byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9
.byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,27 // jne 15d9 <_sk_store_f16_hsw+0x65>
+ .byte 117,27 // jne 1881 <_sk_store_f16_hsw+0x65>
.byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8)
.byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8)
.byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -4939,22 +5329,22 @@ _sk_store_f16_hsw:
.byte 255,224 // jmpq *%rax
.byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,241 // je 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 116,241 // je 187d <_sk_store_f16_hsw+0x61>
.byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,229 // jb 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 114,229 // jb 187d <_sk_store_f16_hsw+0x61>
.byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
- .byte 116,221 // je 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 116,221 // je 187d <_sk_store_f16_hsw+0x61>
.byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,209 // jb 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 114,209 // jb 187d <_sk_store_f16_hsw+0x61>
.byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
- .byte 116,201 // je 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 116,201 // je 187d <_sk_store_f16_hsw+0x61>
.byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,189 // jb 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 114,189 // jb 187d <_sk_store_f16_hsw+0x61>
.byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
- .byte 235,181 // jmp 15d5 <_sk_store_f16_hsw+0x61>
+ .byte 235,181 // jmp 187d <_sk_store_f16_hsw+0x61>
HIDDEN _sk_store_f32_hsw
.globl _sk_store_f32_hsw
@@ -4971,7 +5361,7 @@ _sk_store_f32_hsw:
.byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
.byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,55 // jne 168d <_sk_store_f32_hsw+0x6d>
+ .byte 117,55 // jne 1935 <_sk_store_f32_hsw+0x6d>
.byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
.byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
.byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
@@ -4984,22 +5374,22 @@ _sk_store_f32_hsw:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 1689 <_sk_store_f32_hsw+0x69>
+ .byte 116,240 // je 1931 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 1689 <_sk_store_f32_hsw+0x69>
+ .byte 114,227 // jb 1931 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
- .byte 116,218 // je 1689 <_sk_store_f32_hsw+0x69>
+ .byte 116,218 // je 1931 <_sk_store_f32_hsw+0x69>
.byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 1689 <_sk_store_f32_hsw+0x69>
+ .byte 114,205 // jb 1931 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
- .byte 116,195 // je 1689 <_sk_store_f32_hsw+0x69>
+ .byte 116,195 // je 1931 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,181 // jb 1689 <_sk_store_f32_hsw+0x69>
+ .byte 114,181 // jb 1931 <_sk_store_f32_hsw+0x69>
.byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
- .byte 235,171 // jmp 1689 <_sk_store_f32_hsw+0x69>
+ .byte 235,171 // jmp 1931 <_sk_store_f32_hsw+0x69>
HIDDEN _sk_clamp_x_hsw
.globl _sk_clamp_x_hsw
@@ -6517,7 +6907,7 @@ _sk_lerp_565_avx:
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 233,255,255,255,225 // jmpq ffffffffe2001208 <_sk_linear_gradient_2stops_avx+0xffffffffe1ffee62>
+ .byte 233,255,255,255,225 // jmpq ffffffffe2001208 <_sk_linear_gradient_2stops_avx+0xffffffffe1ffeb1e>
.byte 255 // (bad)
.byte 255 // (bad)
.byte 255 // (bad)
@@ -6812,13 +7202,55 @@ _sk_store_a8_avx:
.byte 255 // (bad)
.byte 255 // .byte 0xff
+HIDDEN _sk_load_g8_avx
+.globl _sk_load_g8_avx
+_sk_load_g8_avx:
+ .byte 73,137,200 // mov %rcx,%r8
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 72,1,248 // add %rdi,%rax
+ .byte 77,133,192 // test %r8,%r8
+ .byte 117,91 // jne 16db <_sk_load_g8_avx+0x6b>
+ .byte 197,250,126,0 // vmovq (%rax),%xmm0
+ .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
+ .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
+ .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
+ .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
+ .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
+ .byte 184,129,128,128,59 // mov $0x3b808081,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
+ .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
+ .byte 184,0,0,128,63 // mov $0x3f800000,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
+ .byte 196,227,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 76,137,193 // mov %r8,%rcx
+ .byte 197,252,40,200 // vmovaps %ymm0,%ymm1
+ .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
+ .byte 255,224 // jmpq *%rax
+ .byte 49,201 // xor %ecx,%ecx
+ .byte 77,137,194 // mov %r8,%r10
+ .byte 69,49,201 // xor %r9d,%r9d
+ .byte 68,15,182,24 // movzbl (%rax),%r11d
+ .byte 72,255,192 // inc %rax
+ .byte 73,211,227 // shl %cl,%r11
+ .byte 77,9,217 // or %r11,%r9
+ .byte 72,131,193,8 // add $0x8,%rcx
+ .byte 73,255,202 // dec %r10
+ .byte 117,234 // jne 16e3 <_sk_load_g8_avx+0x73>
+ .byte 196,193,249,110,193 // vmovq %r9,%xmm0
+ .byte 235,132 // jmp 1684 <_sk_load_g8_avx+0x14>
+
HIDDEN _sk_load_565_avx
.globl _sk_load_565_avx
_sk_load_565_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,209,0,0,0 // jne 174f <_sk_load_565_avx+0xdf>
+ .byte 15,133,209,0,0,0 // jne 17df <_sk_load_565_avx+0xdf>
.byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
.byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
.byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
@@ -6868,9 +7300,9 @@ _sk_load_565_avx:
.byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,29,255,255,255 // ja 1684 <_sk_load_565_avx+0x14>
+ .byte 15,135,29,255,255,255 // ja 1714 <_sk_load_565_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 17bc <_sk_load_565_avx+0x14c>
+ .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 184c <_sk_load_565_avx+0x14c>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -6882,12 +7314,12 @@ _sk_load_565_avx:
.byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
.byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- .byte 233,201,254,255,255 // jmpq 1684 <_sk_load_565_avx+0x14>
+ .byte 233,201,254,255,255 // jmpq 1714 <_sk_load_565_avx+0x14>
.byte 144 // nop
.byte 243,255 // repz (bad)
.byte 255 // (bad)
.byte 255 // (bad)
- .byte 235,255 // jmp 17c1 <_sk_load_565_avx+0x151>
+ .byte 235,255 // jmp 1851 <_sk_load_565_avx+0x151>
.byte 255 // (bad)
.byte 255,227 // jmpq *%rbx
.byte 255 // (bad)
@@ -6939,7 +7371,7 @@ _sk_store_565_avx:
.byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
.byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 1876 <_sk_store_565_avx+0x9e>
+ .byte 117,10 // jne 1906 <_sk_store_565_avx+0x9e>
.byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -6947,9 +7379,9 @@ _sk_store_565_avx:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 1872 <_sk_store_565_avx+0x9a>
+ .byte 119,236 // ja 1902 <_sk_store_565_avx+0x9a>
.byte 65,15,182,192 // movzbl %r8b,%eax
- .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 18d4 <_sk_store_565_avx+0xfc>
+ .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 1964 <_sk_store_565_avx+0xfc>
.byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
.byte 76,1,192 // add %r8,%rax
.byte 255,224 // jmpq *%rax
@@ -6960,7 +7392,7 @@ _sk_store_565_avx:
.byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
.byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
.byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
- .byte 235,159 // jmp 1872 <_sk_store_565_avx+0x9a>
+ .byte 235,159 // jmp 1902 <_sk_store_565_avx+0x9a>
.byte 144 // nop
.byte 246,255 // idiv %bh
.byte 255 // (bad)
@@ -6985,13 +7417,195 @@ _sk_store_565_avx:
.byte 255 // (bad)
.byte 255 // .byte 0xff
+HIDDEN _sk_load_4444_avx
+.globl _sk_load_4444_avx
+_sk_load_4444_avx:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 76,139,16 // mov (%rax),%r10
+ .byte 72,133,201 // test %rcx,%rcx
+ .byte 15,133,245,0,0,0 // jne 1a83 <_sk_load_4444_avx+0x103>
+ .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
+ .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
+ .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
+ .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
+ .byte 196,99,125,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm9
+ .byte 184,0,240,0,0 // mov $0xf000,%eax
+ .byte 197,249,110,192 // vmovd %eax,%xmm0
+ .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
+ .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
+ .byte 196,193,124,84,193 // vandps %ymm9,%ymm0,%ymm0
+ .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
+ .byte 184,137,136,136,55 // mov $0x37888889,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
+ .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
+ .byte 184,0,15,0,0 // mov $0xf00,%eax
+ .byte 197,249,110,200 // vmovd %eax,%xmm1
+ .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
+ .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1
+ .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
+ .byte 184,137,136,136,57 // mov $0x39888889,%eax
+ .byte 197,249,110,208 // vmovd %eax,%xmm2
+ .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
+ .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
+ .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
+ .byte 184,240,0,0,0 // mov $0xf0,%eax
+ .byte 197,249,110,208 // vmovd %eax,%xmm2
+ .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2
+ .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
+ .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2
+ .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8
+ .byte 184,137,136,136,59 // mov $0x3b888889,%eax
+ .byte 197,249,110,208 // vmovd %eax,%xmm2
+ .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
+ .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
+ .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
+ .byte 184,15,0,0,0 // mov $0xf,%eax
+ .byte 197,249,110,216 // vmovd %eax,%xmm3
+ .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
+ .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
+ .byte 196,193,100,84,217 // vandps %ymm9,%ymm3,%ymm3
+ .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
+ .byte 184,137,136,136,61 // mov $0x3d888889,%eax
+ .byte 197,249,110,216 // vmovd %eax,%xmm3
+ .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
+ .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
+ .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,137,200 // mov %ecx,%r8d
+ .byte 65,128,224,7 // and $0x7,%r8b
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 65,254,200 // dec %r8b
+ .byte 65,128,248,6 // cmp $0x6,%r8b
+ .byte 15,135,249,254,255,255 // ja 1994 <_sk_load_4444_avx+0x14>
+ .byte 69,15,182,192 // movzbl %r8b,%r8d
+ .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 1af0 <_sk_load_4444_avx+0x170>
+ .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
+ .byte 76,1,200 // add %r9,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
+ .byte 233,165,254,255,255 // jmpq 1994 <_sk_load_4444_avx+0x14>
+ .byte 144 // nop
+ .byte 243,255 // repz (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 235,255 // jmp 1af5 <_sk_load_4444_avx+0x175>
+ .byte 255 // (bad)
+ .byte 255,227 // jmpq *%rbx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 219,255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,211 // callq *%rbx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,203 // dec %ebx
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 191 // .byte 0xbf
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
+
+HIDDEN _sk_store_4444_avx
+.globl _sk_store_4444_avx
+_sk_store_4444_avx:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 76,139,8 // mov (%rax),%r9
+ .byte 184,0,0,112,65 // mov $0x41700000,%eax
+ .byte 197,121,110,192 // vmovd %eax,%xmm8
+ .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
+ .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
+ .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
+ .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
+ .byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10
+ .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9
+ .byte 196,193,49,114,241,12 // vpslld $0xc,%xmm9,%xmm9
+ .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
+ .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
+ .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
+ .byte 196,193,33,114,242,8 // vpslld $0x8,%xmm10,%xmm11
+ .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
+ .byte 196,193,41,114,242,8 // vpslld $0x8,%xmm10,%xmm10
+ .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
+ .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9
+ .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10
+ .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
+ .byte 196,193,33,114,242,4 // vpslld $0x4,%xmm10,%xmm11
+ .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
+ .byte 196,193,41,114,242,4 // vpslld $0x4,%xmm10,%xmm10
+ .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
+ .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
+ .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
+ .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8
+ .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8
+ .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
+ .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
+ .byte 72,133,201 // test %rcx,%rcx
+ .byte 117,10 // jne 1bbb <_sk_store_4444_avx+0xaf>
+ .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 65,137,200 // mov %ecx,%r8d
+ .byte 65,128,224,7 // and $0x7,%r8b
+ .byte 65,254,200 // dec %r8b
+ .byte 65,128,248,6 // cmp $0x6,%r8b
+ .byte 119,236 // ja 1bb7 <_sk_store_4444_avx+0xab>
+ .byte 65,15,182,192 // movzbl %r8b,%eax
+ .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 1c18 <_sk_store_4444_avx+0x10c>
+ .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
+ .byte 76,1,192 // add %r8,%rax
+ .byte 255,224 // jmpq *%rax
+ .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
+ .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
+ .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
+ .byte 235,159 // jmp 1bb7 <_sk_store_4444_avx+0xab>
+ .byte 247,255 // idiv %edi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 239 // out %eax,(%dx)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,231 // jmpq *%rdi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 223,255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,215 // callq *%rdi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,207 // dec %edi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255,199 // inc %edi
+ .byte 255 // (bad)
+ .byte 255 // (bad)
+ .byte 255 // .byte 0xff
+
HIDDEN _sk_load_8888_avx
.globl _sk_load_8888_avx
_sk_load_8888_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 76,139,16 // mov (%rax),%r10
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,157,0,0,0 // jne 199b <_sk_load_8888_avx+0xab>
+ .byte 15,133,157,0,0,0 // jne 1cdf <_sk_load_8888_avx+0xab>
.byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9
.byte 184,255,0,0,0 // mov $0xff,%eax
.byte 197,249,110,192 // vmovd %eax,%xmm0
@@ -7029,9 +7643,9 @@ _sk_load_8888_avx:
.byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 15,135,80,255,255,255 // ja 1904 <_sk_load_8888_avx+0x14>
+ .byte 15,135,80,255,255,255 // ja 1c48 <_sk_load_8888_avx+0x14>
.byte 69,15,182,192 // movzbl %r8b,%r8d
- .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 1a48 <_sk_load_8888_avx+0x158>
+ .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 1d8c <_sk_load_8888_avx+0x158>
.byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
.byte 76,1,200 // add %r9,%rax
.byte 255,224 // jmpq *%rax
@@ -7054,7 +7668,7 @@ _sk_load_8888_avx:
.byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
.byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0
.byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
- .byte 233,188,254,255,255 // jmpq 1904 <_sk_load_8888_avx+0x14>
+ .byte 233,188,254,255,255 // jmpq 1c48 <_sk_load_8888_avx+0x14>
.byte 238 // out %al,(%dx)
.byte 255 // (bad)
.byte 255 // (bad)
@@ -7107,7 +7721,7 @@ _sk_store_8888_avx:
.byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8
.byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,10 // jne 1b08 <_sk_store_8888_avx+0xa4>
+ .byte 117,10 // jne 1e4c <_sk_store_8888_avx+0xa4>
.byte 196,65,124,17,4,185 // vmovups %ymm8,(%r9,%rdi,4)
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
@@ -7115,9 +7729,9 @@ _sk_store_8888_avx:
.byte 65,128,224,7 // and $0x7,%r8b
.byte 65,254,200 // dec %r8b
.byte 65,128,248,6 // cmp $0x6,%r8b
- .byte 119,236 // ja 1b04 <_sk_store_8888_avx+0xa0>
+ .byte 119,236 // ja 1e48 <_sk_store_8888_avx+0xa0>
.byte 65,15,182,192 // movzbl %r8b,%eax
- .byte 76,141,5,85,0,0,0 // lea 0x55(%rip),%r8 # 1b78 <_sk_store_8888_avx+0x114>
+ .byte 76,141,5,85,0,0,0 // lea 0x55(%rip),%r8 # 1ebc <_sk_store_8888_avx+0x114>
.byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
.byte 76,1,192 // add %r8,%rax
.byte 255,224 // jmpq *%rax
@@ -7131,7 +7745,7 @@ _sk_store_8888_avx:
.byte 196,67,121,22,68,185,8,2 // vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4)
.byte 196,67,121,22,68,185,4,1 // vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4)
.byte 196,65,121,126,4,185 // vmovd %xmm8,(%r9,%rdi,4)
- .byte 235,143 // jmp 1b04 <_sk_store_8888_avx+0xa0>
+ .byte 235,143 // jmp 1e48 <_sk_store_8888_avx+0xa0>
.byte 15,31,0 // nopl (%rax)
.byte 245 // cmc
.byte 255 // (bad)
@@ -7163,7 +7777,7 @@ _sk_load_f16_avx:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 72,139,0 // mov (%rax),%rax
.byte 72,133,201 // test %rcx,%rcx
- .byte 15,133,2,1,0,0 // jne 1ca4 <_sk_load_f16_avx+0x110>
+ .byte 15,133,2,1,0,0 // jne 1fe8 <_sk_load_f16_avx+0x110>
.byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8
.byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
.byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3
@@ -7221,29 +7835,29 @@ _sk_load_f16_avx:
.byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8
.byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,79 // je 1d03 <_sk_load_f16_avx+0x16f>
+ .byte 116,79 // je 2047 <_sk_load_f16_avx+0x16f>
.byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,67 // jb 1d03 <_sk_load_f16_avx+0x16f>
+ .byte 114,67 // jb 2047 <_sk_load_f16_avx+0x16f>
.byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 116,68 // je 1d10 <_sk_load_f16_avx+0x17c>
+ .byte 116,68 // je 2054 <_sk_load_f16_avx+0x17c>
.byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,56 // jb 1d10 <_sk_load_f16_avx+0x17c>
+ .byte 114,56 // jb 2054 <_sk_load_f16_avx+0x17c>
.byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 15,132,209,254,255,255 // je 1bb9 <_sk_load_f16_avx+0x25>
+ .byte 15,132,209,254,255,255 // je 1efd <_sk_load_f16_avx+0x25>
.byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 15,130,193,254,255,255 // jb 1bb9 <_sk_load_f16_avx+0x25>
+ .byte 15,130,193,254,255,255 // jb 1efd <_sk_load_f16_avx+0x25>
.byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9
- .byte 233,182,254,255,255 // jmpq 1bb9 <_sk_load_f16_avx+0x25>
+ .byte 233,182,254,255,255 // jmpq 1efd <_sk_load_f16_avx+0x25>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
.byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
- .byte 233,169,254,255,255 // jmpq 1bb9 <_sk_load_f16_avx+0x25>
+ .byte 233,169,254,255,255 // jmpq 1efd <_sk_load_f16_avx+0x25>
.byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
- .byte 233,160,254,255,255 // jmpq 1bb9 <_sk_load_f16_avx+0x25>
+ .byte 233,160,254,255,255 // jmpq 1efd <_sk_load_f16_avx+0x25>
HIDDEN _sk_store_f16_avx
.globl _sk_store_f16_avx
@@ -7283,7 +7897,7 @@ _sk_store_f16_avx:
.byte 196,65,25,98,205 // vpunpckldq %xmm13,%xmm12,%xmm9
.byte 196,65,25,106,197 // vpunpckhdq %xmm13,%xmm12,%xmm8
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,31 // jne 1def <_sk_store_f16_avx+0xd6>
+ .byte 117,31 // jne 2133 <_sk_store_f16_avx+0xd6>
.byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8)
.byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8)
.byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8)
@@ -7292,22 +7906,22 @@ _sk_store_f16_avx:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 1deb <_sk_store_f16_avx+0xd2>
+ .byte 116,240 // je 212f <_sk_store_f16_avx+0xd2>
.byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 1deb <_sk_store_f16_avx+0xd2>
+ .byte 114,227 // jb 212f <_sk_store_f16_avx+0xd2>
.byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8)
- .byte 116,218 // je 1deb <_sk_store_f16_avx+0xd2>
+ .byte 116,218 // je 212f <_sk_store_f16_avx+0xd2>
.byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 1deb <_sk_store_f16_avx+0xd2>
+ .byte 114,205 // jb 212f <_sk_store_f16_avx+0xd2>
.byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8)
- .byte 116,196 // je 1deb <_sk_store_f16_avx+0xd2>
+ .byte 116,196 // je 212f <_sk_store_f16_avx+0xd2>
.byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,183 // jb 1deb <_sk_store_f16_avx+0xd2>
+ .byte 114,183 // jb 212f <_sk_store_f16_avx+0xd2>
.byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8)
- .byte 235,174 // jmp 1deb <_sk_store_f16_avx+0xd2>
+ .byte 235,174 // jmp 212f <_sk_store_f16_avx+0xd2>
HIDDEN _sk_store_f32_avx
.globl _sk_store_f32_avx
@@ -7324,7 +7938,7 @@ _sk_store_f32_avx:
.byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
.byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
.byte 72,133,201 // test %rcx,%rcx
- .byte 117,55 // jne 1eaa <_sk_store_f32_avx+0x6d>
+ .byte 117,55 // jne 21ee <_sk_store_f32_avx+0x6d>
.byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
.byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
.byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
@@ -7337,22 +7951,22 @@ _sk_store_f32_avx:
.byte 255,224 // jmpq *%rax
.byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
.byte 72,131,249,1 // cmp $0x1,%rcx
- .byte 116,240 // je 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 116,240 // je 21ea <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
.byte 72,131,249,3 // cmp $0x3,%rcx
- .byte 114,227 // jb 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 114,227 // jb 21ea <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
- .byte 116,218 // je 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 116,218 // je 21ea <_sk_store_f32_avx+0x69>
.byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
.byte 72,131,249,5 // cmp $0x5,%rcx
- .byte 114,205 // jb 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 114,205 // jb 21ea <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
- .byte 116,195 // je 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 116,195 // je 21ea <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
.byte 72,131,249,7 // cmp $0x7,%rcx
- .byte 114,181 // jb 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 114,181 // jb 21ea <_sk_store_f32_avx+0x69>
.byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
- .byte 235,171 // jmp 1ea6 <_sk_store_f32_avx+0x69>
+ .byte 235,171 // jmp 21ea <_sk_store_f32_avx+0x69>
HIDDEN _sk_clamp_x_avx
.globl _sk_clamp_x_avx
@@ -9146,6 +9760,25 @@ _sk_store_a8_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_load_g8_sse41
+.globl _sk_load_g8_sse41
+_sk_load_g8_sse41:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0
+ .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
+ .byte 184,129,128,128,59 // mov $0x3b808081,%eax
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
+ .byte 15,89,193 // mulps %xmm1,%xmm0
+ .byte 184,0,0,128,63 // mov $0x3f800000,%eax
+ .byte 102,15,110,216 // movd %eax,%xmm3
+ .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,200 // movaps %xmm0,%xmm1
+ .byte 15,40,208 // movaps %xmm0,%xmm2
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_load_565_sse41
.globl _sk_load_565_sse41
_sk_load_565_sse41:
@@ -9212,6 +9845,81 @@ _sk_store_565_sse41:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_load_4444_sse41
+.globl _sk_load_4444_sse41
+_sk_load_4444_sse41:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 102,68,15,56,51,12,120 // pmovzxwd (%rax,%rdi,2),%xmm9
+ .byte 184,0,240,0,0 // mov $0xf000,%eax
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
+ .byte 102,65,15,219,193 // pand %xmm9,%xmm0
+ .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
+ .byte 184,137,136,136,55 // mov $0x37888889,%eax
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
+ .byte 15,89,193 // mulps %xmm1,%xmm0
+ .byte 184,0,15,0,0 // mov $0xf00,%eax
+ .byte 102,15,110,200 // movd %eax,%xmm1
+ .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
+ .byte 102,65,15,219,201 // pand %xmm9,%xmm1
+ .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2
+ .byte 184,137,136,136,57 // mov $0x39888889,%eax
+ .byte 102,15,110,200 // movd %eax,%xmm1
+ .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
+ .byte 15,89,202 // mulps %xmm2,%xmm1
+ .byte 184,240,0,0,0 // mov $0xf0,%eax
+ .byte 102,15,110,208 // movd %eax,%xmm2
+ .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
+ .byte 102,65,15,219,209 // pand %xmm9,%xmm2
+ .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8
+ .byte 184,137,136,136,59 // mov $0x3b888889,%eax
+ .byte 102,15,110,208 // movd %eax,%xmm2
+ .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 184,15,0,0,0 // mov $0xf,%eax
+ .byte 102,15,110,216 // movd %eax,%xmm3
+ .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
+ .byte 102,65,15,219,217 // pand %xmm9,%xmm3
+ .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
+ .byte 184,137,136,136,61 // mov $0x3d888889,%eax
+ .byte 102,15,110,216 // movd %eax,%xmm3
+ .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
+ .byte 65,15,89,216 // mulps %xmm8,%xmm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_store_4444_sse41
+.globl _sk_store_4444_sse41
+_sk_store_4444_sse41:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 185,0,0,112,65 // mov $0x41700000,%ecx
+ .byte 102,68,15,110,193 // movd %ecx,%xmm8
+ .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 68,15,89,200 // mulps %xmm0,%xmm9
+ .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
+ .byte 102,65,15,114,241,12 // pslld $0xc,%xmm9
+ .byte 69,15,40,208 // movaps %xmm8,%xmm10
+ .byte 68,15,89,209 // mulps %xmm1,%xmm10
+ .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
+ .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10
+ .byte 102,69,15,235,209 // por %xmm9,%xmm10
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 68,15,89,202 // mulps %xmm2,%xmm9
+ .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
+ .byte 102,65,15,114,241,4 // pslld $0x4,%xmm9
+ .byte 68,15,89,195 // mulps %xmm3,%xmm8
+ .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
+ .byte 102,69,15,86,193 // orpd %xmm9,%xmm8
+ .byte 102,69,15,86,194 // orpd %xmm10,%xmm8
+ .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
+ .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2)
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_load_8888_sse41
.globl _sk_load_8888_sse41
_sk_load_8888_sse41:
@@ -11258,6 +11966,28 @@ _sk_store_a8_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_load_g8_sse2
+.globl _sk_load_g8_sse2
+_sk_load_g8_sse2:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 102,15,110,4,56 // movd (%rax,%rdi,1),%xmm0
+ .byte 102,15,239,201 // pxor %xmm1,%xmm1
+ .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0
+ .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
+ .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
+ .byte 184,129,128,128,59 // mov $0x3b808081,%eax
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
+ .byte 15,89,193 // mulps %xmm1,%xmm0
+ .byte 184,0,0,128,63 // mov $0x3f800000,%eax
+ .byte 102,15,110,216 // movd %eax,%xmm3
+ .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 15,40,200 // movaps %xmm0,%xmm1
+ .byte 15,40,208 // movaps %xmm0,%xmm2
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_load_565_sse2
.globl _sk_load_565_sse2
_sk_load_565_sse2:
@@ -11328,6 +12058,85 @@ _sk_store_565_sse2:
.byte 72,173 // lods %ds:(%rsi),%rax
.byte 255,224 // jmpq *%rax
+HIDDEN _sk_load_4444_sse2
+.globl _sk_load_4444_sse2
+_sk_load_4444_sse2:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 243,68,15,126,12,120 // movq (%rax,%rdi,2),%xmm9
+ .byte 102,15,239,192 // pxor %xmm0,%xmm0
+ .byte 102,68,15,97,200 // punpcklwd %xmm0,%xmm9
+ .byte 184,0,240,0,0 // mov $0xf000,%eax
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
+ .byte 102,65,15,219,193 // pand %xmm9,%xmm0
+ .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
+ .byte 184,137,136,136,55 // mov $0x37888889,%eax
+ .byte 102,15,110,192 // movd %eax,%xmm0
+ .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
+ .byte 15,89,193 // mulps %xmm1,%xmm0
+ .byte 184,0,15,0,0 // mov $0xf00,%eax
+ .byte 102,15,110,200 // movd %eax,%xmm1
+ .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
+ .byte 102,65,15,219,201 // pand %xmm9,%xmm1
+ .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2
+ .byte 184,137,136,136,57 // mov $0x39888889,%eax
+ .byte 102,15,110,200 // movd %eax,%xmm1
+ .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
+ .byte 15,89,202 // mulps %xmm2,%xmm1
+ .byte 184,240,0,0,0 // mov $0xf0,%eax
+ .byte 102,15,110,208 // movd %eax,%xmm2
+ .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
+ .byte 102,65,15,219,209 // pand %xmm9,%xmm2
+ .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8
+ .byte 184,137,136,136,59 // mov $0x3b888889,%eax
+ .byte 102,15,110,208 // movd %eax,%xmm2
+ .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
+ .byte 65,15,89,208 // mulps %xmm8,%xmm2
+ .byte 184,15,0,0,0 // mov $0xf,%eax
+ .byte 102,15,110,216 // movd %eax,%xmm3
+ .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
+ .byte 102,65,15,219,217 // pand %xmm9,%xmm3
+ .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
+ .byte 184,137,136,136,61 // mov $0x3d888889,%eax
+ .byte 102,15,110,216 // movd %eax,%xmm3
+ .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
+ .byte 65,15,89,216 // mulps %xmm8,%xmm3
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
+HIDDEN _sk_store_4444_sse2
+.globl _sk_store_4444_sse2
+_sk_store_4444_sse2:
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 72,139,0 // mov (%rax),%rax
+ .byte 185,0,0,112,65 // mov $0x41700000,%ecx
+ .byte 102,68,15,110,193 // movd %ecx,%xmm8
+ .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 68,15,89,200 // mulps %xmm0,%xmm9
+ .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
+ .byte 102,65,15,114,241,12 // pslld $0xc,%xmm9
+ .byte 69,15,40,208 // movaps %xmm8,%xmm10
+ .byte 68,15,89,209 // mulps %xmm1,%xmm10
+ .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
+ .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10
+ .byte 102,69,15,235,209 // por %xmm9,%xmm10
+ .byte 69,15,40,200 // movaps %xmm8,%xmm9
+ .byte 68,15,89,202 // mulps %xmm2,%xmm9
+ .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
+ .byte 102,65,15,114,241,4 // pslld $0x4,%xmm9
+ .byte 68,15,89,195 // mulps %xmm3,%xmm8
+ .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
+ .byte 102,69,15,86,193 // orpd %xmm9,%xmm8
+ .byte 102,69,15,86,194 // orpd %xmm10,%xmm8
+ .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
+ .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8
+ .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
+ .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2)
+ .byte 72,173 // lods %ds:(%rsi),%rax
+ .byte 255,224 // jmpq *%rax
+
HIDDEN _sk_load_8888_sse2
.globl _sk_load_8888_sse2
_sk_load_8888_sse2:
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index 6e0eb84fb9..2fc3c4c8a8 100644
--- a/src/jumper/SkJumper_generated_win.S
+++ b/src/jumper/SkJumper_generated_win.S
@@ -1266,12 +1266,48 @@ _sk_store_a8_hsw LABEL PROC
DB 255 ; (bad)
DB 255 ; .byte 0xff
+PUBLIC _sk_load_g8_hsw
+_sk_load_g8_hsw LABEL PROC
+ DB 73,137,200 ; mov %rcx,%r8
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 72,1,248 ; add %rdi,%rax
+ DB 77,133,192 ; test %r8,%r8
+ DB 117,60 ; jne 1228 <_sk_load_g8_hsw+0x4c>
+ DB 197,250,126,0 ; vmovq (%rax),%xmm0
+ DB 196,226,125,49,192 ; vpmovzxbd %xmm0,%ymm0
+ DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
+ DB 184,129,128,128,59 ; mov $0x3b808081,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1
+ DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
+ DB 184,0,0,128,63 ; mov $0x3f800000,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,226,125,88,217 ; vpbroadcastd %xmm1,%ymm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,137,193 ; mov %r8,%rcx
+ DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
+ DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
+ DB 255,224 ; jmpq *%rax
+ DB 49,201 ; xor %ecx,%ecx
+ DB 77,137,194 ; mov %r8,%r10
+ DB 69,49,201 ; xor %r9d,%r9d
+ DB 68,15,182,24 ; movzbl (%rax),%r11d
+ DB 72,255,192 ; inc %rax
+ DB 73,211,227 ; shl %cl,%r11
+ DB 77,9,217 ; or %r11,%r9
+ DB 72,131,193,8 ; add $0x8,%rcx
+ DB 73,255,202 ; dec %r10
+ DB 117,234 ; jne 1230 <_sk_load_g8_hsw+0x54>
+ DB 196,193,249,110,193 ; vmovq %r9,%xmm0
+ DB 235,163 ; jmp 11f0 <_sk_load_g8_hsw+0x14>
+
PUBLIC _sk_load_565_hsw
_sk_load_565_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,149,0,0,0 ; jne 127f <_sk_load_565_hsw+0xa3>
+ DB 15,133,149,0,0,0 ; jne 12f0 <_sk_load_565_hsw+0xa3>
DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
DB 196,226,125,51,208 ; vpmovzxwd %xmm0,%ymm2
DB 184,0,248,0,0 ; mov $0xf800,%eax
@@ -1311,9 +1347,9 @@ _sk_load_565_hsw LABEL PROC
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,89,255,255,255 ; ja 11f0 <_sk_load_565_hsw+0x14>
+ DB 15,135,89,255,255,255 ; ja 1261 <_sk_load_565_hsw+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 12ec <_sk_load_565_hsw+0x110>
+ DB 76,141,13,73,0,0,0 ; lea 0x49(%rip),%r9 # 135c <_sk_load_565_hsw+0x10f>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -1325,27 +1361,27 @@ _sk_load_565_hsw LABEL PROC
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- DB 233,5,255,255,255 ; jmpq 11f0 <_sk_load_565_hsw+0x14>
- DB 144 ; nop
- DB 243,255 ; repz (bad)
+ DB 233,5,255,255,255 ; jmpq 1261 <_sk_load_565_hsw+0x14>
+ DB 244 ; hlt
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 12f1 <_sk_load_565_hsw+0x115>
DB 255 ; (bad)
- DB 255,227 ; jmpq *%rbx
+ DB 236 ; in (%dx),%al
DB 255 ; (bad)
DB 255 ; (bad)
+ DB 255,228 ; jmpq *%rsp
DB 255 ; (bad)
- DB 219,255 ; (bad)
DB 255 ; (bad)
- DB 255,211 ; callq *%rbx
DB 255 ; (bad)
+ DB 220,255 ; fdivr %st,%st(7)
+ DB 255 ; (bad)
+ DB 255,212 ; callq *%rsp
DB 255 ; (bad)
- DB 255,203 ; dec %ebx
DB 255 ; (bad)
+ DB 255,204 ; dec %esp
DB 255 ; (bad)
DB 255 ; (bad)
- DB 191 ; .byte 0xbf
+ DB 255,192 ; inc %eax
DB 255 ; (bad)
DB 255 ; (bad)
DB 255 ; .byte 0xff
@@ -1373,7 +1409,7 @@ _sk_store_565_hsw LABEL PROC
DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 1374 <_sk_store_565_hsw+0x6c>
+ DB 117,10 ; jne 13e4 <_sk_store_565_hsw+0x6c>
DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -1381,9 +1417,9 @@ _sk_store_565_hsw LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 1370 <_sk_store_565_hsw+0x68>
+ DB 119,236 ; ja 13e0 <_sk_store_565_hsw+0x68>
DB 65,15,182,192 ; movzbl %r8b,%eax
- DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 13d4 <_sk_store_565_hsw+0xcc>
+ DB 76,141,5,69,0,0,0 ; lea 0x45(%rip),%r8 # 1444 <_sk_store_565_hsw+0xcc>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -1394,7 +1430,7 @@ _sk_store_565_hsw LABEL PROC
DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2)
- DB 235,159 ; jmp 1370 <_sk_store_565_hsw+0x68>
+ DB 235,159 ; jmp 13e0 <_sk_store_565_hsw+0x68>
DB 15,31,0 ; nopl (%rax)
DB 244 ; hlt
DB 255 ; (bad)
@@ -1420,6 +1456,165 @@ _sk_store_565_hsw LABEL PROC
DB 255 ; (bad)
DB 255 ; .byte 0xff
+PUBLIC _sk_load_4444_hsw
+_sk_load_4444_hsw LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,139,16 ; mov (%rax),%r10
+ DB 72,133,201 ; test %rcx,%rcx
+ DB 15,133,179,0,0,0 ; jne 1521 <_sk_load_4444_hsw+0xc1>
+ DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
+ DB 196,98,125,51,200 ; vpmovzxwd %xmm0,%ymm9
+ DB 184,0,240,0,0 ; mov $0xf000,%eax
+ DB 197,249,110,192 ; vmovd %eax,%xmm0
+ DB 196,226,125,88,192 ; vpbroadcastd %xmm0,%ymm0
+ DB 196,193,125,219,193 ; vpand %ymm9,%ymm0,%ymm0
+ DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
+ DB 184,137,136,136,55 ; mov $0x37888889,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1
+ DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
+ DB 184,0,15,0,0 ; mov $0xf00,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,226,125,88,201 ; vpbroadcastd %xmm1,%ymm1
+ DB 196,193,117,219,201 ; vpand %ymm9,%ymm1,%ymm1
+ DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
+ DB 184,137,136,136,57 ; mov $0x39888889,%eax
+ DB 197,249,110,208 ; vmovd %eax,%xmm2
+ DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2
+ DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
+ DB 184,240,0,0,0 ; mov $0xf0,%eax
+ DB 197,249,110,208 ; vmovd %eax,%xmm2
+ DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2
+ DB 196,193,109,219,209 ; vpand %ymm9,%ymm2,%ymm2
+ DB 197,124,91,194 ; vcvtdq2ps %ymm2,%ymm8
+ DB 184,137,136,136,59 ; mov $0x3b888889,%eax
+ DB 197,249,110,208 ; vmovd %eax,%xmm2
+ DB 196,226,125,88,210 ; vpbroadcastd %xmm2,%ymm2
+ DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
+ DB 184,15,0,0,0 ; mov $0xf,%eax
+ DB 197,249,110,216 ; vmovd %eax,%xmm3
+ DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3
+ DB 196,193,101,219,217 ; vpand %ymm9,%ymm3,%ymm3
+ DB 197,124,91,195 ; vcvtdq2ps %ymm3,%ymm8
+ DB 184,137,136,136,61 ; mov $0x3d888889,%eax
+ DB 197,249,110,216 ; vmovd %eax,%xmm3
+ DB 196,226,125,88,219 ; vpbroadcastd %xmm3,%ymm3
+ DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 65,128,224,7 ; and $0x7,%r8b
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 65,254,200 ; dec %r8b
+ DB 65,128,248,6 ; cmp $0x6,%r8b
+ DB 15,135,59,255,255,255 ; ja 1474 <_sk_load_4444_hsw+0x14>
+ DB 69,15,182,192 ; movzbl %r8b,%r8d
+ DB 76,141,13,76,0,0,0 ; lea 0x4c(%rip),%r9 # 1590 <_sk_load_4444_hsw+0x130>
+ DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
+ DB 76,1,200 ; add %r9,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 196,193,121,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
+ DB 233,231,254,255,255 ; jmpq 1474 <_sk_load_4444_hsw+0x14>
+ DB 15,31,0 ; nopl (%rax)
+ DB 241 ; icebp
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 233,255,255,255,225 ; jmpq ffffffffe2001598 <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff7f0>
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 217,255 ; fcos
+ DB 255 ; (bad)
+ DB 255,209 ; callq *%rcx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,201 ; dec %ecx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 189 ; .byte 0xbd
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
+
+PUBLIC _sk_store_4444_hsw
+_sk_store_4444_hsw LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,139,8 ; mov (%rax),%r9
+ DB 184,0,0,112,65 ; mov $0x41700000,%eax
+ DB 197,121,110,192 ; vmovd %eax,%xmm8
+ DB 196,66,125,88,192 ; vpbroadcastd %xmm8,%ymm8
+ DB 197,60,89,200 ; vmulps %ymm0,%ymm8,%ymm9
+ DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
+ DB 196,193,53,114,241,12 ; vpslld $0xc,%ymm9,%ymm9
+ DB 197,60,89,209 ; vmulps %ymm1,%ymm8,%ymm10
+ DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10
+ DB 196,193,45,114,242,8 ; vpslld $0x8,%ymm10,%ymm10
+ DB 196,65,45,235,201 ; vpor %ymm9,%ymm10,%ymm9
+ DB 197,60,89,210 ; vmulps %ymm2,%ymm8,%ymm10
+ DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10
+ DB 196,193,45,114,242,4 ; vpslld $0x4,%ymm10,%ymm10
+ DB 197,60,89,195 ; vmulps %ymm3,%ymm8,%ymm8
+ DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8
+ DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8
+ DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8
+ DB 196,67,125,57,193,1 ; vextracti128 $0x1,%ymm8,%xmm9
+ DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
+ DB 72,133,201 ; test %rcx,%rcx
+ DB 117,10 ; jne 161e <_sk_store_4444_hsw+0x72>
+ DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2)
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 65,128,224,7 ; and $0x7,%r8b
+ DB 65,254,200 ; dec %r8b
+ DB 65,128,248,6 ; cmp $0x6,%r8b
+ DB 119,236 ; ja 161a <_sk_store_4444_hsw+0x6e>
+ DB 65,15,182,192 ; movzbl %r8b,%eax
+ DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 167c <_sk_store_4444_hsw+0xd0>
+ DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
+ DB 76,1,192 ; add %r8,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 196,67,121,21,68,121,12,6 ; vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,10,5 ; vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,8,4 ; vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,6,3 ; vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
+ DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2)
+ DB 235,159 ; jmp 161a <_sk_store_4444_hsw+0x6e>
+ DB 144 ; nop
+ DB 246,255 ; idiv %bh
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 238 ; out %al,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,230 ; jmpq *%rsi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 222,255 ; fdivrp %st,%st(7)
+ DB 255 ; (bad)
+ DB 255,214 ; callq *%rsi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,206 ; dec %esi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,198 ; inc %esi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
+
PUBLIC _sk_load_8888_hsw
_sk_load_8888_hsw LABEL PROC
DB 73,137,200 ; mov %rcx,%r8
@@ -1427,7 +1622,7 @@ _sk_load_8888_hsw LABEL PROC
DB 76,141,12,189,0,0,0,0 ; lea 0x0(,%rdi,4),%r9
DB 76,3,8 ; add (%rax),%r9
DB 77,133,192 ; test %r8,%r8
- DB 117,104 ; jne 146d <_sk_load_8888_hsw+0x7d>
+ DB 117,104 ; jne 1715 <_sk_load_8888_hsw+0x7d>
DB 196,193,126,111,25 ; vmovdqu (%r9),%ymm3
DB 184,255,0,0,0 ; mov $0xff,%eax
DB 197,249,110,192 ; vmovd %eax,%xmm0
@@ -1460,7 +1655,7 @@ _sk_load_8888_hsw LABEL PROC
DB 196,225,249,110,192 ; vmovq %rax,%xmm0
DB 196,226,125,33,192 ; vpmovsxbd %xmm0,%ymm0
DB 196,194,125,140,25 ; vpmaskmovd (%r9),%ymm0,%ymm3
- DB 233,116,255,255,255 ; jmpq 140a <_sk_load_8888_hsw+0x1a>
+ DB 233,116,255,255,255 ; jmpq 16b2 <_sk_load_8888_hsw+0x1a>
PUBLIC _sk_store_8888_hsw
_sk_store_8888_hsw LABEL PROC
@@ -1486,7 +1681,7 @@ _sk_store_8888_hsw LABEL PROC
DB 196,65,45,235,192 ; vpor %ymm8,%ymm10,%ymm8
DB 196,65,53,235,192 ; vpor %ymm8,%ymm9,%ymm8
DB 77,133,192 ; test %r8,%r8
- DB 117,12 ; jne 150a <_sk_store_8888_hsw+0x74>
+ DB 117,12 ; jne 17b2 <_sk_store_8888_hsw+0x74>
DB 196,65,126,127,1 ; vmovdqu %ymm8,(%r9)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,137,193 ; mov %r8,%rcx
@@ -1499,14 +1694,14 @@ _sk_store_8888_hsw LABEL PROC
DB 196,97,249,110,200 ; vmovq %rax,%xmm9
DB 196,66,125,33,201 ; vpmovsxbd %xmm9,%ymm9
DB 196,66,53,142,1 ; vpmaskmovd %ymm8,%ymm9,(%r9)
- DB 235,211 ; jmp 1503 <_sk_store_8888_hsw+0x6d>
+ DB 235,211 ; jmp 17ab <_sk_store_8888_hsw+0x6d>
PUBLIC _sk_load_f16_hsw
_sk_load_f16_hsw LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 117,97 ; jne 159b <_sk_load_f16_hsw+0x6b>
+ DB 117,97 ; jne 1843 <_sk_load_f16_hsw+0x6b>
DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8
DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2
DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3
@@ -1532,29 +1727,29 @@ _sk_load_f16_hsw LABEL PROC
DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,79 ; je 15fa <_sk_load_f16_hsw+0xca>
+ DB 116,79 ; je 18a2 <_sk_load_f16_hsw+0xca>
DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,67 ; jb 15fa <_sk_load_f16_hsw+0xca>
+ DB 114,67 ; jb 18a2 <_sk_load_f16_hsw+0xca>
DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,68 ; je 1607 <_sk_load_f16_hsw+0xd7>
+ DB 116,68 ; je 18af <_sk_load_f16_hsw+0xd7>
DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,56 ; jb 1607 <_sk_load_f16_hsw+0xd7>
+ DB 114,56 ; jb 18af <_sk_load_f16_hsw+0xd7>
DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,114,255,255,255 ; je 1551 <_sk_load_f16_hsw+0x21>
+ DB 15,132,114,255,255,255 ; je 17f9 <_sk_load_f16_hsw+0x21>
DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,98,255,255,255 ; jb 1551 <_sk_load_f16_hsw+0x21>
+ DB 15,130,98,255,255,255 ; jb 17f9 <_sk_load_f16_hsw+0x21>
DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9
- DB 233,87,255,255,255 ; jmpq 1551 <_sk_load_f16_hsw+0x21>
+ DB 233,87,255,255,255 ; jmpq 17f9 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,74,255,255,255 ; jmpq 1551 <_sk_load_f16_hsw+0x21>
+ DB 233,74,255,255,255 ; jmpq 17f9 <_sk_load_f16_hsw+0x21>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,65,255,255,255 ; jmpq 1551 <_sk_load_f16_hsw+0x21>
+ DB 233,65,255,255,255 ; jmpq 17f9 <_sk_load_f16_hsw+0x21>
PUBLIC _sk_store_f16_hsw
_sk_store_f16_hsw LABEL PROC
@@ -1573,7 +1768,7 @@ _sk_store_f16_hsw LABEL PROC
DB 196,65,57,98,205 ; vpunpckldq %xmm13,%xmm8,%xmm9
DB 196,65,57,106,197 ; vpunpckhdq %xmm13,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,27 ; jne 1675 <_sk_store_f16_hsw+0x65>
+ DB 117,27 ; jne 191d <_sk_store_f16_hsw+0x65>
DB 197,120,17,28,248 ; vmovups %xmm11,(%rax,%rdi,8)
DB 197,120,17,84,248,16 ; vmovups %xmm10,0x10(%rax,%rdi,8)
DB 197,120,17,76,248,32 ; vmovups %xmm9,0x20(%rax,%rdi,8)
@@ -1582,22 +1777,22 @@ _sk_store_f16_hsw LABEL PROC
DB 255,224 ; jmpq *%rax
DB 197,121,214,28,248 ; vmovq %xmm11,(%rax,%rdi,8)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,241 ; je 1671 <_sk_store_f16_hsw+0x61>
+ DB 116,241 ; je 1919 <_sk_store_f16_hsw+0x61>
DB 197,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%rax,%rdi,8)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,229 ; jb 1671 <_sk_store_f16_hsw+0x61>
+ DB 114,229 ; jb 1919 <_sk_store_f16_hsw+0x61>
DB 197,121,214,84,248,16 ; vmovq %xmm10,0x10(%rax,%rdi,8)
- DB 116,221 ; je 1671 <_sk_store_f16_hsw+0x61>
+ DB 116,221 ; je 1919 <_sk_store_f16_hsw+0x61>
DB 197,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%rax,%rdi,8)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,209 ; jb 1671 <_sk_store_f16_hsw+0x61>
+ DB 114,209 ; jb 1919 <_sk_store_f16_hsw+0x61>
DB 197,121,214,76,248,32 ; vmovq %xmm9,0x20(%rax,%rdi,8)
- DB 116,201 ; je 1671 <_sk_store_f16_hsw+0x61>
+ DB 116,201 ; je 1919 <_sk_store_f16_hsw+0x61>
DB 197,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%rax,%rdi,8)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,189 ; jb 1671 <_sk_store_f16_hsw+0x61>
+ DB 114,189 ; jb 1919 <_sk_store_f16_hsw+0x61>
DB 197,121,214,68,248,48 ; vmovq %xmm8,0x30(%rax,%rdi,8)
- DB 235,181 ; jmp 1671 <_sk_store_f16_hsw+0x61>
+ DB 235,181 ; jmp 1919 <_sk_store_f16_hsw+0x61>
PUBLIC _sk_store_f32_hsw
_sk_store_f32_hsw LABEL PROC
@@ -1613,7 +1808,7 @@ _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 1729 <_sk_store_f32_hsw+0x6d>
+ DB 117,55 ; jne 19d1 <_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
@@ -1626,22 +1821,22 @@ _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 1725 <_sk_store_f32_hsw+0x69>
+ DB 116,240 ; je 19cd <_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 1725 <_sk_store_f32_hsw+0x69>
+ DB 114,227 ; jb 19cd <_sk_store_f32_hsw+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 1725 <_sk_store_f32_hsw+0x69>
+ DB 116,218 ; je 19cd <_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 1725 <_sk_store_f32_hsw+0x69>
+ DB 114,205 ; jb 19cd <_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 1725 <_sk_store_f32_hsw+0x69>
+ DB 116,195 ; je 19cd <_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 1725 <_sk_store_f32_hsw+0x69>
+ DB 114,181 ; jb 19cd <_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 1725 <_sk_store_f32_hsw+0x69>
+ DB 235,171 ; jmp 19cd <_sk_store_f32_hsw+0x69>
PUBLIC _sk_clamp_x_hsw
_sk_clamp_x_hsw LABEL PROC
@@ -3426,12 +3621,53 @@ _sk_store_a8_avx LABEL PROC
DB 255 ; (bad)
DB 255 ; .byte 0xff
+PUBLIC _sk_load_g8_avx
+_sk_load_g8_avx LABEL PROC
+ DB 73,137,200 ; mov %rcx,%r8
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 72,1,248 ; add %rdi,%rax
+ DB 77,133,192 ; test %r8,%r8
+ DB 117,91 ; jne 1777 <_sk_load_g8_avx+0x6b>
+ DB 197,250,126,0 ; vmovq (%rax),%xmm0
+ DB 196,226,121,49,200 ; vpmovzxbd %xmm0,%xmm1
+ DB 196,227,121,4,192,229 ; vpermilps $0xe5,%xmm0,%xmm0
+ DB 196,226,121,49,192 ; vpmovzxbd %xmm0,%xmm0
+ DB 196,227,117,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
+ DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
+ DB 184,129,128,128,59 ; mov $0x3b808081,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1
+ DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
+ DB 184,0,0,128,63 ; mov $0x3f800000,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1
+ DB 196,227,117,24,217,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,137,193 ; mov %r8,%rcx
+ DB 197,252,40,200 ; vmovaps %ymm0,%ymm1
+ DB 197,252,40,208 ; vmovaps %ymm0,%ymm2
+ DB 255,224 ; jmpq *%rax
+ DB 49,201 ; xor %ecx,%ecx
+ DB 77,137,194 ; mov %r8,%r10
+ DB 69,49,201 ; xor %r9d,%r9d
+ DB 68,15,182,24 ; movzbl (%rax),%r11d
+ DB 72,255,192 ; inc %rax
+ DB 73,211,227 ; shl %cl,%r11
+ DB 77,9,217 ; or %r11,%r9
+ DB 72,131,193,8 ; add $0x8,%rcx
+ DB 73,255,202 ; dec %r10
+ DB 117,234 ; jne 177f <_sk_load_g8_avx+0x73>
+ DB 196,193,249,110,193 ; vmovq %r9,%xmm0
+ DB 235,132 ; jmp 1720 <_sk_load_g8_avx+0x14>
+
PUBLIC _sk_load_565_avx
_sk_load_565_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,209,0,0,0 ; jne 17eb <_sk_load_565_avx+0xdf>
+ DB 15,133,209,0,0,0 ; jne 187b <_sk_load_565_avx+0xdf>
DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1
DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
@@ -3481,9 +3717,9 @@ _sk_load_565_avx LABEL PROC
DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,29,255,255,255 ; ja 1720 <_sk_load_565_avx+0x14>
+ DB 15,135,29,255,255,255 ; ja 17b0 <_sk_load_565_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1858 <_sk_load_565_avx+0x14c>
+ DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 18e8 <_sk_load_565_avx+0x14c>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -3495,12 +3731,12 @@ _sk_load_565_avx LABEL PROC
DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
- DB 233,201,254,255,255 ; jmpq 1720 <_sk_load_565_avx+0x14>
+ DB 233,201,254,255,255 ; jmpq 17b0 <_sk_load_565_avx+0x14>
DB 144 ; nop
DB 243,255 ; repz (bad)
DB 255 ; (bad)
DB 255 ; (bad)
- DB 235,255 ; jmp 185d <_sk_load_565_avx+0x151>
+ DB 235,255 ; jmp 18ed <_sk_load_565_avx+0x151>
DB 255 ; (bad)
DB 255,227 ; jmpq *%rbx
DB 255 ; (bad)
@@ -3551,7 +3787,7 @@ _sk_store_565_avx LABEL PROC
DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 1912 <_sk_store_565_avx+0x9e>
+ DB 117,10 ; jne 19a2 <_sk_store_565_avx+0x9e>
DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -3559,9 +3795,9 @@ _sk_store_565_avx LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 190e <_sk_store_565_avx+0x9a>
+ DB 119,236 ; ja 199e <_sk_store_565_avx+0x9a>
DB 65,15,182,192 ; movzbl %r8b,%eax
- DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1970 <_sk_store_565_avx+0xfc>
+ DB 76,141,5,67,0,0,0 ; lea 0x43(%rip),%r8 # 1a00 <_sk_store_565_avx+0xfc>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -3572,7 +3808,7 @@ _sk_store_565_avx LABEL PROC
DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2)
- DB 235,159 ; jmp 190e <_sk_store_565_avx+0x9a>
+ DB 235,159 ; jmp 199e <_sk_store_565_avx+0x9a>
DB 144 ; nop
DB 246,255 ; idiv %bh
DB 255 ; (bad)
@@ -3597,12 +3833,192 @@ _sk_store_565_avx LABEL PROC
DB 255 ; (bad)
DB 255 ; .byte 0xff
+PUBLIC _sk_load_4444_avx
+_sk_load_4444_avx LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,139,16 ; mov (%rax),%r10
+ DB 72,133,201 ; test %rcx,%rcx
+ DB 15,133,245,0,0,0 ; jne 1b1f <_sk_load_4444_avx+0x103>
+ DB 196,193,122,111,4,122 ; vmovdqu (%r10,%rdi,2),%xmm0
+ DB 197,241,239,201 ; vpxor %xmm1,%xmm1,%xmm1
+ DB 197,249,105,201 ; vpunpckhwd %xmm1,%xmm0,%xmm1
+ DB 196,226,121,51,192 ; vpmovzxwd %xmm0,%xmm0
+ DB 196,99,125,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm0,%ymm9
+ DB 184,0,240,0,0 ; mov $0xf000,%eax
+ DB 197,249,110,192 ; vmovd %eax,%xmm0
+ DB 197,249,112,192,0 ; vpshufd $0x0,%xmm0,%xmm0
+ DB 196,227,125,24,192,1 ; vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
+ DB 196,193,124,84,193 ; vandps %ymm9,%ymm0,%ymm0
+ DB 197,252,91,192 ; vcvtdq2ps %ymm0,%ymm0
+ DB 184,137,136,136,55 ; mov $0x37888889,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 196,227,121,4,201,0 ; vpermilps $0x0,%xmm1,%xmm1
+ DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ DB 197,252,89,193 ; vmulps %ymm1,%ymm0,%ymm0
+ DB 184,0,15,0,0 ; mov $0xf00,%eax
+ DB 197,249,110,200 ; vmovd %eax,%xmm1
+ DB 197,249,112,201,0 ; vpshufd $0x0,%xmm1,%xmm1
+ DB 196,227,117,24,201,1 ; vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
+ DB 196,193,116,84,201 ; vandps %ymm9,%ymm1,%ymm1
+ DB 197,252,91,201 ; vcvtdq2ps %ymm1,%ymm1
+ DB 184,137,136,136,57 ; mov $0x39888889,%eax
+ DB 197,249,110,208 ; vmovd %eax,%xmm2
+ DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2
+ DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
+ DB 197,244,89,202 ; vmulps %ymm2,%ymm1,%ymm1
+ DB 184,240,0,0,0 ; mov $0xf0,%eax
+ DB 197,249,110,208 ; vmovd %eax,%xmm2
+ DB 197,249,112,210,0 ; vpshufd $0x0,%xmm2,%xmm2
+ DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
+ DB 196,193,108,84,209 ; vandps %ymm9,%ymm2,%ymm2
+ DB 197,124,91,194 ; vcvtdq2ps %ymm2,%ymm8
+ DB 184,137,136,136,59 ; mov $0x3b888889,%eax
+ DB 197,249,110,208 ; vmovd %eax,%xmm2
+ DB 196,227,121,4,210,0 ; vpermilps $0x0,%xmm2,%xmm2
+ DB 196,227,109,24,210,1 ; vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
+ DB 197,188,89,210 ; vmulps %ymm2,%ymm8,%ymm2
+ DB 184,15,0,0,0 ; mov $0xf,%eax
+ DB 197,249,110,216 ; vmovd %eax,%xmm3
+ DB 197,249,112,219,0 ; vpshufd $0x0,%xmm3,%xmm3
+ DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
+ DB 196,193,100,84,217 ; vandps %ymm9,%ymm3,%ymm3
+ DB 197,124,91,195 ; vcvtdq2ps %ymm3,%ymm8
+ DB 184,137,136,136,61 ; mov $0x3d888889,%eax
+ DB 197,249,110,216 ; vmovd %eax,%xmm3
+ DB 196,227,121,4,219,0 ; vpermilps $0x0,%xmm3,%xmm3
+ DB 196,227,101,24,219,1 ; vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
+ DB 197,188,89,219 ; vmulps %ymm3,%ymm8,%ymm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 65,128,224,7 ; and $0x7,%r8b
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 65,254,200 ; dec %r8b
+ DB 65,128,248,6 ; cmp $0x6,%r8b
+ DB 15,135,249,254,255,255 ; ja 1a30 <_sk_load_4444_avx+0x14>
+ DB 69,15,182,192 ; movzbl %r8b,%r8d
+ DB 76,141,13,74,0,0,0 ; lea 0x4a(%rip),%r9 # 1b8c <_sk_load_4444_avx+0x170>
+ DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
+ DB 76,1,200 ; add %r9,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 197,249,239,192 ; vpxor %xmm0,%xmm0,%xmm0
+ DB 196,193,121,196,68,122,12,6 ; vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,10,5 ; vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,8,4 ; vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,6,3 ; vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,4,2 ; vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,68,122,2,1 ; vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
+ DB 196,193,121,196,4,122,0 ; vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
+ DB 233,165,254,255,255 ; jmpq 1a30 <_sk_load_4444_avx+0x14>
+ DB 144 ; nop
+ DB 243,255 ; repz (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 235,255 ; jmp 1b91 <_sk_load_4444_avx+0x175>
+ DB 255 ; (bad)
+ DB 255,227 ; jmpq *%rbx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 219,255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,211 ; callq *%rbx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,203 ; dec %ebx
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 191 ; .byte 0xbf
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
+
+PUBLIC _sk_store_4444_avx
+_sk_store_4444_avx LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 76,139,8 ; mov (%rax),%r9
+ DB 184,0,0,112,65 ; mov $0x41700000,%eax
+ DB 197,121,110,192 ; vmovd %eax,%xmm8
+ DB 196,67,121,4,192,0 ; vpermilps $0x0,%xmm8,%xmm8
+ DB 196,67,61,24,192,1 ; vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
+ DB 197,60,89,200 ; vmulps %ymm0,%ymm8,%ymm9
+ DB 196,65,125,91,201 ; vcvtps2dq %ymm9,%ymm9
+ DB 196,193,41,114,241,12 ; vpslld $0xc,%xmm9,%xmm10
+ DB 196,67,125,25,201,1 ; vextractf128 $0x1,%ymm9,%xmm9
+ DB 196,193,49,114,241,12 ; vpslld $0xc,%xmm9,%xmm9
+ DB 196,67,45,24,201,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
+ DB 197,60,89,209 ; vmulps %ymm1,%ymm8,%ymm10
+ DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10
+ DB 196,193,33,114,242,8 ; vpslld $0x8,%xmm10,%xmm11
+ DB 196,67,125,25,210,1 ; vextractf128 $0x1,%ymm10,%xmm10
+ DB 196,193,41,114,242,8 ; vpslld $0x8,%xmm10,%xmm10
+ DB 196,67,37,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
+ DB 196,65,45,86,201 ; vorpd %ymm9,%ymm10,%ymm9
+ DB 197,60,89,210 ; vmulps %ymm2,%ymm8,%ymm10
+ DB 196,65,125,91,210 ; vcvtps2dq %ymm10,%ymm10
+ DB 196,193,33,114,242,4 ; vpslld $0x4,%xmm10,%xmm11
+ DB 196,67,125,25,210,1 ; vextractf128 $0x1,%ymm10,%xmm10
+ DB 196,193,41,114,242,4 ; vpslld $0x4,%xmm10,%xmm10
+ DB 196,67,37,24,210,1 ; vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
+ DB 197,60,89,195 ; vmulps %ymm3,%ymm8,%ymm8
+ DB 196,65,125,91,192 ; vcvtps2dq %ymm8,%ymm8
+ DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8
+ DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8
+ DB 196,67,125,25,193,1 ; vextractf128 $0x1,%ymm8,%xmm9
+ DB 196,66,57,43,193 ; vpackusdw %xmm9,%xmm8,%xmm8
+ DB 72,133,201 ; test %rcx,%rcx
+ DB 117,10 ; jne 1c57 <_sk_store_4444_avx+0xaf>
+ DB 196,65,122,127,4,121 ; vmovdqu %xmm8,(%r9,%rdi,2)
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+ DB 65,137,200 ; mov %ecx,%r8d
+ DB 65,128,224,7 ; and $0x7,%r8b
+ DB 65,254,200 ; dec %r8b
+ DB 65,128,248,6 ; cmp $0x6,%r8b
+ DB 119,236 ; ja 1c53 <_sk_store_4444_avx+0xab>
+ DB 65,15,182,192 ; movzbl %r8b,%eax
+ DB 76,141,5,66,0,0,0 ; lea 0x42(%rip),%r8 # 1cb4 <_sk_store_4444_avx+0x10c>
+ DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
+ DB 76,1,192 ; add %r8,%rax
+ DB 255,224 ; jmpq *%rax
+ DB 196,67,121,21,68,121,12,6 ; vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,10,5 ; vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,8,4 ; vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,6,3 ; vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,4,2 ; vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
+ DB 196,67,121,21,68,121,2,1 ; vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
+ DB 196,67,121,21,4,121,0 ; vpextrw $0x0,%xmm8,(%r9,%rdi,2)
+ DB 235,159 ; jmp 1c53 <_sk_store_4444_avx+0xab>
+ DB 247,255 ; idiv %edi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 239 ; out %eax,(%dx)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,231 ; jmpq *%rdi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 223,255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,215 ; callq *%rdi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,207 ; dec %edi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255,199 ; inc %edi
+ DB 255 ; (bad)
+ DB 255 ; (bad)
+ DB 255 ; .byte 0xff
+
PUBLIC _sk_load_8888_avx
_sk_load_8888_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 76,139,16 ; mov (%rax),%r10
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,157,0,0,0 ; jne 1a37 <_sk_load_8888_avx+0xab>
+ DB 15,133,157,0,0,0 ; jne 1d7b <_sk_load_8888_avx+0xab>
DB 196,65,124,16,12,186 ; vmovups (%r10,%rdi,4),%ymm9
DB 184,255,0,0,0 ; mov $0xff,%eax
DB 197,249,110,192 ; vmovd %eax,%xmm0
@@ -3640,9 +4056,9 @@ _sk_load_8888_avx LABEL PROC
DB 196,65,52,87,201 ; vxorps %ymm9,%ymm9,%ymm9
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 15,135,80,255,255,255 ; ja 19a0 <_sk_load_8888_avx+0x14>
+ DB 15,135,80,255,255,255 ; ja 1ce4 <_sk_load_8888_avx+0x14>
DB 69,15,182,192 ; movzbl %r8b,%r8d
- DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 1ae4 <_sk_load_8888_avx+0x158>
+ DB 76,141,13,137,0,0,0 ; lea 0x89(%rip),%r9 # 1e28 <_sk_load_8888_avx+0x158>
DB 75,99,4,129 ; movslq (%r9,%r8,4),%rax
DB 76,1,200 ; add %r9,%rax
DB 255,224 ; jmpq *%rax
@@ -3665,7 +4081,7 @@ _sk_load_8888_avx LABEL PROC
DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9
DB 196,195,49,34,4,186,0 ; vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0
DB 196,99,53,12,200,15 ; vblendps $0xf,%ymm0,%ymm9,%ymm9
- DB 233,188,254,255,255 ; jmpq 19a0 <_sk_load_8888_avx+0x14>
+ DB 233,188,254,255,255 ; jmpq 1ce4 <_sk_load_8888_avx+0x14>
DB 238 ; out %al,(%dx)
DB 255 ; (bad)
DB 255 ; (bad)
@@ -3717,7 +4133,7 @@ _sk_store_8888_avx LABEL PROC
DB 196,65,45,86,192 ; vorpd %ymm8,%ymm10,%ymm8
DB 196,65,53,86,192 ; vorpd %ymm8,%ymm9,%ymm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,10 ; jne 1ba4 <_sk_store_8888_avx+0xa4>
+ DB 117,10 ; jne 1ee8 <_sk_store_8888_avx+0xa4>
DB 196,65,124,17,4,185 ; vmovups %ymm8,(%r9,%rdi,4)
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
@@ -3725,9 +4141,9 @@ _sk_store_8888_avx LABEL PROC
DB 65,128,224,7 ; and $0x7,%r8b
DB 65,254,200 ; dec %r8b
DB 65,128,248,6 ; cmp $0x6,%r8b
- DB 119,236 ; ja 1ba0 <_sk_store_8888_avx+0xa0>
+ DB 119,236 ; ja 1ee4 <_sk_store_8888_avx+0xa0>
DB 65,15,182,192 ; movzbl %r8b,%eax
- DB 76,141,5,85,0,0,0 ; lea 0x55(%rip),%r8 # 1c14 <_sk_store_8888_avx+0x114>
+ DB 76,141,5,85,0,0,0 ; lea 0x55(%rip),%r8 # 1f58 <_sk_store_8888_avx+0x114>
DB 73,99,4,128 ; movslq (%r8,%rax,4),%rax
DB 76,1,192 ; add %r8,%rax
DB 255,224 ; jmpq *%rax
@@ -3741,7 +4157,7 @@ _sk_store_8888_avx LABEL PROC
DB 196,67,121,22,68,185,8,2 ; vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4)
DB 196,67,121,22,68,185,4,1 ; vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4)
DB 196,65,121,126,4,185 ; vmovd %xmm8,(%r9,%rdi,4)
- DB 235,143 ; jmp 1ba0 <_sk_store_8888_avx+0xa0>
+ DB 235,143 ; jmp 1ee4 <_sk_store_8888_avx+0xa0>
DB 15,31,0 ; nopl (%rax)
DB 245 ; cmc
DB 255 ; (bad)
@@ -3772,7 +4188,7 @@ _sk_load_f16_avx LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 72,139,0 ; mov (%rax),%rax
DB 72,133,201 ; test %rcx,%rcx
- DB 15,133,2,1,0,0 ; jne 1d40 <_sk_load_f16_avx+0x110>
+ DB 15,133,2,1,0,0 ; jne 2084 <_sk_load_f16_avx+0x110>
DB 197,121,16,4,248 ; vmovupd (%rax,%rdi,8),%xmm8
DB 197,249,16,84,248,16 ; vmovupd 0x10(%rax,%rdi,8),%xmm2
DB 197,249,16,92,248,32 ; vmovupd 0x20(%rax,%rdi,8),%xmm3
@@ -3830,29 +4246,29 @@ _sk_load_f16_avx LABEL PROC
DB 197,123,16,4,248 ; vmovsd (%rax,%rdi,8),%xmm8
DB 196,65,49,239,201 ; vpxor %xmm9,%xmm9,%xmm9
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,79 ; je 1d9f <_sk_load_f16_avx+0x16f>
+ DB 116,79 ; je 20e3 <_sk_load_f16_avx+0x16f>
DB 197,57,22,68,248,8 ; vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,67 ; jb 1d9f <_sk_load_f16_avx+0x16f>
+ DB 114,67 ; jb 20e3 <_sk_load_f16_avx+0x16f>
DB 197,251,16,84,248,16 ; vmovsd 0x10(%rax,%rdi,8),%xmm2
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 116,68 ; je 1dac <_sk_load_f16_avx+0x17c>
+ DB 116,68 ; je 20f0 <_sk_load_f16_avx+0x17c>
DB 197,233,22,84,248,24 ; vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,56 ; jb 1dac <_sk_load_f16_avx+0x17c>
+ DB 114,56 ; jb 20f0 <_sk_load_f16_avx+0x17c>
DB 197,251,16,92,248,32 ; vmovsd 0x20(%rax,%rdi,8),%xmm3
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 15,132,209,254,255,255 ; je 1c55 <_sk_load_f16_avx+0x25>
+ DB 15,132,209,254,255,255 ; je 1f99 <_sk_load_f16_avx+0x25>
DB 197,225,22,92,248,40 ; vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 15,130,193,254,255,255 ; jb 1c55 <_sk_load_f16_avx+0x25>
+ DB 15,130,193,254,255,255 ; jb 1f99 <_sk_load_f16_avx+0x25>
DB 197,122,126,76,248,48 ; vmovq 0x30(%rax,%rdi,8),%xmm9
- DB 233,182,254,255,255 ; jmpq 1c55 <_sk_load_f16_avx+0x25>
+ DB 233,182,254,255,255 ; jmpq 1f99 <_sk_load_f16_avx+0x25>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
DB 197,233,87,210 ; vxorpd %xmm2,%xmm2,%xmm2
- DB 233,169,254,255,255 ; jmpq 1c55 <_sk_load_f16_avx+0x25>
+ DB 233,169,254,255,255 ; jmpq 1f99 <_sk_load_f16_avx+0x25>
DB 197,225,87,219 ; vxorpd %xmm3,%xmm3,%xmm3
- DB 233,160,254,255,255 ; jmpq 1c55 <_sk_load_f16_avx+0x25>
+ DB 233,160,254,255,255 ; jmpq 1f99 <_sk_load_f16_avx+0x25>
PUBLIC _sk_store_f16_avx
_sk_store_f16_avx LABEL PROC
@@ -3891,7 +4307,7 @@ _sk_store_f16_avx LABEL PROC
DB 196,65,25,98,205 ; vpunpckldq %xmm13,%xmm12,%xmm9
DB 196,65,25,106,197 ; vpunpckhdq %xmm13,%xmm12,%xmm8
DB 72,133,201 ; test %rcx,%rcx
- DB 117,31 ; jne 1e8b <_sk_store_f16_avx+0xd6>
+ DB 117,31 ; jne 21cf <_sk_store_f16_avx+0xd6>
DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8)
DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8)
DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8)
@@ -3900,22 +4316,22 @@ _sk_store_f16_avx LABEL PROC
DB 255,224 ; jmpq *%rax
DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8)
DB 72,131,249,1 ; cmp $0x1,%rcx
- DB 116,240 ; je 1e87 <_sk_store_f16_avx+0xd2>
+ DB 116,240 ; je 21cb <_sk_store_f16_avx+0xd2>
DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8)
DB 72,131,249,3 ; cmp $0x3,%rcx
- DB 114,227 ; jb 1e87 <_sk_store_f16_avx+0xd2>
+ DB 114,227 ; jb 21cb <_sk_store_f16_avx+0xd2>
DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8)
- DB 116,218 ; je 1e87 <_sk_store_f16_avx+0xd2>
+ DB 116,218 ; je 21cb <_sk_store_f16_avx+0xd2>
DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8)
DB 72,131,249,5 ; cmp $0x5,%rcx
- DB 114,205 ; jb 1e87 <_sk_store_f16_avx+0xd2>
+ DB 114,205 ; jb 21cb <_sk_store_f16_avx+0xd2>
DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8)
- DB 116,196 ; je 1e87 <_sk_store_f16_avx+0xd2>
+ DB 116,196 ; je 21cb <_sk_store_f16_avx+0xd2>
DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8)
DB 72,131,249,7 ; cmp $0x7,%rcx
- DB 114,183 ; jb 1e87 <_sk_store_f16_avx+0xd2>
+ DB 114,183 ; jb 21cb <_sk_store_f16_avx+0xd2>
DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8)
- DB 235,174 ; jmp 1e87 <_sk_store_f16_avx+0xd2>
+ DB 235,174 ; jmp 21cb <_sk_store_f16_avx+0xd2>
PUBLIC _sk_store_f32_avx
_sk_store_f32_avx LABEL PROC
@@ -3931,7 +4347,7 @@ _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 1f46 <_sk_store_f32_avx+0x6d>
+ DB 117,55 ; jne 228a <_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
@@ -3944,22 +4360,22 @@ _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 1f42 <_sk_store_f32_avx+0x69>
+ DB 116,240 ; je 2286 <_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 1f42 <_sk_store_f32_avx+0x69>
+ DB 114,227 ; jb 2286 <_sk_store_f32_avx+0x69>
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
- DB 116,218 ; je 1f42 <_sk_store_f32_avx+0x69>
+ DB 116,218 ; je 2286 <_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 1f42 <_sk_store_f32_avx+0x69>
+ DB 114,205 ; jb 2286 <_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 1f42 <_sk_store_f32_avx+0x69>
+ DB 116,195 ; je 2286 <_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 1f42 <_sk_store_f32_avx+0x69>
+ DB 114,181 ; jb 2286 <_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 1f42 <_sk_store_f32_avx+0x69>
+ DB 235,171 ; jmp 2286 <_sk_store_f32_avx+0x69>
PUBLIC _sk_clamp_x_avx
_sk_clamp_x_avx LABEL PROC
@@ -5727,6 +6143,24 @@ _sk_store_a8_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_load_g8_sse41
+_sk_load_g8_sse41 LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 102,15,56,49,4,56 ; pmovzxbd (%rax,%rdi,1),%xmm0
+ DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
+ DB 184,129,128,128,59 ; mov $0x3b808081,%eax
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
+ DB 15,89,193 ; mulps %xmm1,%xmm0
+ DB 184,0,0,128,63 ; mov $0x3f800000,%eax
+ DB 102,15,110,216 ; movd %eax,%xmm3
+ DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,200 ; movaps %xmm0,%xmm1
+ DB 15,40,208 ; movaps %xmm0,%xmm2
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_565_sse41
_sk_load_565_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -5791,6 +6225,79 @@ _sk_store_565_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_load_4444_sse41
+_sk_load_4444_sse41 LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 102,68,15,56,51,12,120 ; pmovzxwd (%rax,%rdi,2),%xmm9
+ DB 184,0,240,0,0 ; mov $0xf000,%eax
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
+ DB 102,65,15,219,193 ; pand %xmm9,%xmm0
+ DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
+ DB 184,137,136,136,55 ; mov $0x37888889,%eax
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
+ DB 15,89,193 ; mulps %xmm1,%xmm0
+ DB 184,0,15,0,0 ; mov $0xf00,%eax
+ DB 102,15,110,200 ; movd %eax,%xmm1
+ DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1
+ DB 102,65,15,219,201 ; pand %xmm9,%xmm1
+ DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2
+ DB 184,137,136,136,57 ; mov $0x39888889,%eax
+ DB 102,15,110,200 ; movd %eax,%xmm1
+ DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1
+ DB 15,89,202 ; mulps %xmm2,%xmm1
+ DB 184,240,0,0,0 ; mov $0xf0,%eax
+ DB 102,15,110,208 ; movd %eax,%xmm2
+ DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2
+ DB 102,65,15,219,209 ; pand %xmm9,%xmm2
+ DB 68,15,91,194 ; cvtdq2ps %xmm2,%xmm8
+ DB 184,137,136,136,59 ; mov $0x3b888889,%eax
+ DB 102,15,110,208 ; movd %eax,%xmm2
+ DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 184,15,0,0,0 ; mov $0xf,%eax
+ DB 102,15,110,216 ; movd %eax,%xmm3
+ DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3
+ DB 102,65,15,219,217 ; pand %xmm9,%xmm3
+ DB 68,15,91,195 ; cvtdq2ps %xmm3,%xmm8
+ DB 184,137,136,136,61 ; mov $0x3d888889,%eax
+ DB 102,15,110,216 ; movd %eax,%xmm3
+ DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
+ DB 65,15,89,216 ; mulps %xmm8,%xmm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_store_4444_sse41
+_sk_store_4444_sse41 LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 185,0,0,112,65 ; mov $0x41700000,%ecx
+ DB 102,68,15,110,193 ; movd %ecx,%xmm8
+ DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 68,15,89,200 ; mulps %xmm0,%xmm9
+ DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9
+ DB 102,65,15,114,241,12 ; pslld $0xc,%xmm9
+ DB 69,15,40,208 ; movaps %xmm8,%xmm10
+ DB 68,15,89,209 ; mulps %xmm1,%xmm10
+ DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10
+ DB 102,65,15,114,242,8 ; pslld $0x8,%xmm10
+ DB 102,69,15,235,209 ; por %xmm9,%xmm10
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 68,15,89,202 ; mulps %xmm2,%xmm9
+ DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9
+ DB 102,65,15,114,241,4 ; pslld $0x4,%xmm9
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8
+ DB 102,69,15,86,193 ; orpd %xmm9,%xmm8
+ DB 102,69,15,86,194 ; orpd %xmm10,%xmm8
+ DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8
+ DB 102,68,15,214,4,120 ; movq %xmm8,(%rax,%rdi,2)
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_8888_sse41
_sk_load_8888_sse41 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -7804,6 +8311,27 @@ _sk_store_a8_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_load_g8_sse2
+_sk_load_g8_sse2 LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 102,15,110,4,56 ; movd (%rax,%rdi,1),%xmm0
+ DB 102,15,239,201 ; pxor %xmm1,%xmm1
+ DB 102,15,96,193 ; punpcklbw %xmm1,%xmm0
+ DB 102,15,97,193 ; punpcklwd %xmm1,%xmm0
+ DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
+ DB 184,129,128,128,59 ; mov $0x3b808081,%eax
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
+ DB 15,89,193 ; mulps %xmm1,%xmm0
+ DB 184,0,0,128,63 ; mov $0x3f800000,%eax
+ DB 102,15,110,216 ; movd %eax,%xmm3
+ DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 15,40,200 ; movaps %xmm0,%xmm1
+ DB 15,40,208 ; movaps %xmm0,%xmm2
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_565_sse2
_sk_load_565_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
@@ -7872,6 +8400,83 @@ _sk_store_565_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
DB 255,224 ; jmpq *%rax
+PUBLIC _sk_load_4444_sse2
+_sk_load_4444_sse2 LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 243,68,15,126,12,120 ; movq (%rax,%rdi,2),%xmm9
+ DB 102,15,239,192 ; pxor %xmm0,%xmm0
+ DB 102,68,15,97,200 ; punpcklwd %xmm0,%xmm9
+ DB 184,0,240,0,0 ; mov $0xf000,%eax
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 102,15,112,192,0 ; pshufd $0x0,%xmm0,%xmm0
+ DB 102,65,15,219,193 ; pand %xmm9,%xmm0
+ DB 15,91,200 ; cvtdq2ps %xmm0,%xmm1
+ DB 184,137,136,136,55 ; mov $0x37888889,%eax
+ DB 102,15,110,192 ; movd %eax,%xmm0
+ DB 15,198,192,0 ; shufps $0x0,%xmm0,%xmm0
+ DB 15,89,193 ; mulps %xmm1,%xmm0
+ DB 184,0,15,0,0 ; mov $0xf00,%eax
+ DB 102,15,110,200 ; movd %eax,%xmm1
+ DB 102,15,112,201,0 ; pshufd $0x0,%xmm1,%xmm1
+ DB 102,65,15,219,201 ; pand %xmm9,%xmm1
+ DB 15,91,209 ; cvtdq2ps %xmm1,%xmm2
+ DB 184,137,136,136,57 ; mov $0x39888889,%eax
+ DB 102,15,110,200 ; movd %eax,%xmm1
+ DB 15,198,201,0 ; shufps $0x0,%xmm1,%xmm1
+ DB 15,89,202 ; mulps %xmm2,%xmm1
+ DB 184,240,0,0,0 ; mov $0xf0,%eax
+ DB 102,15,110,208 ; movd %eax,%xmm2
+ DB 102,15,112,210,0 ; pshufd $0x0,%xmm2,%xmm2
+ DB 102,65,15,219,209 ; pand %xmm9,%xmm2
+ DB 68,15,91,194 ; cvtdq2ps %xmm2,%xmm8
+ DB 184,137,136,136,59 ; mov $0x3b888889,%eax
+ DB 102,15,110,208 ; movd %eax,%xmm2
+ DB 15,198,210,0 ; shufps $0x0,%xmm2,%xmm2
+ DB 65,15,89,208 ; mulps %xmm8,%xmm2
+ DB 184,15,0,0,0 ; mov $0xf,%eax
+ DB 102,15,110,216 ; movd %eax,%xmm3
+ DB 102,15,112,219,0 ; pshufd $0x0,%xmm3,%xmm3
+ DB 102,65,15,219,217 ; pand %xmm9,%xmm3
+ DB 68,15,91,195 ; cvtdq2ps %xmm3,%xmm8
+ DB 184,137,136,136,61 ; mov $0x3d888889,%eax
+ DB 102,15,110,216 ; movd %eax,%xmm3
+ DB 15,198,219,0 ; shufps $0x0,%xmm3,%xmm3
+ DB 65,15,89,216 ; mulps %xmm8,%xmm3
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
+PUBLIC _sk_store_4444_sse2
+_sk_store_4444_sse2 LABEL PROC
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 72,139,0 ; mov (%rax),%rax
+ DB 185,0,0,112,65 ; mov $0x41700000,%ecx
+ DB 102,68,15,110,193 ; movd %ecx,%xmm8
+ DB 69,15,198,192,0 ; shufps $0x0,%xmm8,%xmm8
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 68,15,89,200 ; mulps %xmm0,%xmm9
+ DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9
+ DB 102,65,15,114,241,12 ; pslld $0xc,%xmm9
+ DB 69,15,40,208 ; movaps %xmm8,%xmm10
+ DB 68,15,89,209 ; mulps %xmm1,%xmm10
+ DB 102,69,15,91,210 ; cvtps2dq %xmm10,%xmm10
+ DB 102,65,15,114,242,8 ; pslld $0x8,%xmm10
+ DB 102,69,15,235,209 ; por %xmm9,%xmm10
+ DB 69,15,40,200 ; movaps %xmm8,%xmm9
+ DB 68,15,89,202 ; mulps %xmm2,%xmm9
+ DB 102,69,15,91,201 ; cvtps2dq %xmm9,%xmm9
+ DB 102,65,15,114,241,4 ; pslld $0x4,%xmm9
+ DB 68,15,89,195 ; mulps %xmm3,%xmm8
+ DB 102,69,15,91,192 ; cvtps2dq %xmm8,%xmm8
+ DB 102,69,15,86,193 ; orpd %xmm9,%xmm8
+ DB 102,69,15,86,194 ; orpd %xmm10,%xmm8
+ DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8
+ DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8
+ DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8
+ DB 102,68,15,214,4,120 ; movq %xmm8,(%rax,%rdi,2)
+ DB 72,173 ; lods %ds:(%rsi),%rax
+ DB 255,224 ; jmpq *%rax
+
PUBLIC _sk_load_8888_sse2
_sk_load_8888_sse2 LABEL PROC
DB 72,173 ; lods %ds:(%rsi),%rax
diff --git a/src/jumper/SkJumper_stages.cpp b/src/jumper/SkJumper_stages.cpp
index e39a496a05..6dddf9c744 100644
--- a/src/jumper/SkJumper_stages.cpp
+++ b/src/jumper/SkJumper_stages.cpp
@@ -353,6 +353,13 @@ SI void from_565(U16 _565, F* r, F* g, F* b) {
*g = cast(wide & C(63<< 5)) * C(1.0f / (63<< 5));
*b = cast(wide & C(31<< 0)) * C(1.0f / (31<< 0));
}
+SI void from_4444(U16 _4444, F* r, F* g, F* b, F* a) {
+ U32 wide = expand(_4444);
+ *r = cast(wide & C(15<<12)) * C(1.0f / (15<<12));
+ *g = cast(wide & C(15<< 8)) * C(1.0f / (15<< 8));
+ *b = cast(wide & C(15<< 4)) * C(1.0f / (15<< 4));
+ *a = cast(wide & C(15<< 0)) * C(1.0f / (15<< 0));
+}
// Sometimes we want to work with 4 floats directly, regardless of the depth of the F vector.
#if defined(JUMPER)
@@ -771,6 +778,13 @@ STAGE(store_a8) {
store(ptr, packed, tail);
}
+STAGE(load_g8) {
+ auto ptr = *(const uint8_t**)ctx + x;
+
+ r = g = b = cast(expand(load<U8>(ptr, tail))) * C(1/255.0f);
+ a = 1.0_f;
+}
+
STAGE(load_565) {
auto ptr = *(const uint16_t**)ctx + x;
@@ -786,6 +800,19 @@ STAGE(store_565) {
store(ptr, px, tail);
}
+STAGE(load_4444) {
+ auto ptr = *(const uint16_t**)ctx + x;
+ from_4444(load<U16>(ptr, tail), &r,&g,&b,&a);
+}
+STAGE(store_4444) {
+ auto ptr = *(uint16_t**)ctx + x;
+ U16 px = pack( round(r, 15.0_f) << 12
+ | round(g, 15.0_f) << 8
+ | round(b, 15.0_f) << 4
+ | round(a, 15.0_f) );
+ store(ptr, px, tail);
+}
+
STAGE(load_8888) {
auto ptr = *(const uint32_t**)ctx + x;
@@ -795,7 +822,6 @@ STAGE(load_8888) {
b = cast((px >> 16) & 0xff_i) * C(1/255.0f);
a = cast((px >> 24) ) * C(1/255.0f);
}
-
STAGE(store_8888) {
auto ptr = *(uint32_t**)ctx + x;
generated by cgit v1.2.3 (git 2.39.1) at 2024-06-22 20:50:26 +0000