| # Copyright 2017 Google Inc. |
| # |
| # Use of this source code is governed by a BSD-style license that can be |
| # found in the LICENSE file. |
| |
| # This file is generated semi-automatically with this command: |
| # $ src/jumper/build_stages.py |
| |
| #if defined(__MACH__) |
| #define HIDDEN .private_extern |
| #define FUNCTION(name) |
| #else |
| #define HIDDEN .hidden |
| #define FUNCTION(name) .type name,%function |
| .section .note.GNU-stack,"",%progbits |
| #endif |
| .text |
| #if defined(__aarch64__) |
| .balign 4 |
| |
| HIDDEN _sk_start_pipeline_aarch64 |
| .globl _sk_start_pipeline_aarch64 |
| FUNCTION(_sk_start_pipeline_aarch64) |
| _sk_start_pipeline_aarch64: |
| .long 0xa9bd5bf7 // stp x23, x22, [sp, #-48]! |
| .long 0xa90153f5 // stp x21, x20, [sp, #16] |
| .long 0xa9027bf3 // stp x19, x30, [sp, #32] |
| .long 0xaa0103f4 // mov x20, x1 |
| .long 0xf8408697 // ldr x23, [x20], #8 |
| .long 0xaa0003f5 // mov x21, x0 |
| .long 0xaa0303f3 // mov x19, x3 |
| .long 0x910012a8 // add x8, x21, #0x4 |
| .long 0xeb13011f // cmp x8, x19 |
| .long 0xaa0203f6 // mov x22, x2 |
| .long 0x54000069 // b.ls 34 <sk_start_pipeline_aarch64+0x34> // b.plast |
| .long 0xaa1503e0 // mov x0, x21 |
| .long 0x14000012 // b 78 <sk_start_pipeline_aarch64+0x78> |
| .long 0x6f00e400 // movi v0.2d, #0x0 |
| .long 0x6f00e401 // movi v1.2d, #0x0 |
| .long 0x6f00e402 // movi v2.2d, #0x0 |
| .long 0x6f00e403 // movi v3.2d, #0x0 |
| .long 0x6f00e404 // movi v4.2d, #0x0 |
| .long 0x6f00e405 // movi v5.2d, #0x0 |
| .long 0x6f00e406 // movi v6.2d, #0x0 |
| .long 0x6f00e407 // movi v7.2d, #0x0 |
| .long 0xaa1503e0 // mov x0, x21 |
| .long 0xaa1403e1 // mov x1, x20 |
| .long 0xaa1603e2 // mov x2, x22 |
| .long 0xd63f02e0 // blr x23 |
| .long 0x910012a0 // add x0, x21, #0x4 |
| .long 0x910022a8 // add x8, x21, #0x8 |
| .long 0xeb13011f // cmp x8, x19 |
| .long 0xaa0003f5 // mov x21, x0 |
| .long 0x54fffe09 // b.ls 34 <sk_start_pipeline_aarch64+0x34> // b.plast |
| .long 0xa9427bf3 // ldp x19, x30, [sp, #32] |
| .long 0xa94153f5 // ldp x21, x20, [sp, #16] |
| .long 0xa8c35bf7 // ldp x23, x22, [sp], #48 |
| .long 0xd65f03c0 // ret |
| |
| HIDDEN _sk_just_return_aarch64 |
| .globl _sk_just_return_aarch64 |
| FUNCTION(_sk_just_return_aarch64) |
| _sk_just_return_aarch64: |
| .long 0xd65f03c0 // ret |
| |
| HIDDEN _sk_seed_shader_aarch64 |
| .globl _sk_seed_shader_aarch64 |
| FUNCTION(_sk_seed_shader_aarch64) |
| _sk_seed_shader_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x3dc00046 // ldr q6, [x2] |
| .long 0x4e040c00 // dup v0.4s, w0 |
| .long 0x4f0167e7 // movi v7.4s, #0x3f, lsl #24 |
| .long 0x4d40c901 // ld1r {v1.4s}, [x8] |
| .long 0x4e21d800 // scvtf v0.4s, v0.4s |
| .long 0x4e27d400 // fadd v0.4s, v0.4s, v7.4s |
| .long 0x4f03f602 // fmov v2.4s, #1.000000000000000000e+00 |
| .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| .long 0x6f00e403 // movi v3.2d, #0x0 |
| .long 0x6f00e404 // movi v4.2d, #0x0 |
| .long 0x6f00e405 // movi v5.2d, #0x0 |
| .long 0x4e26d400 // fadd v0.4s, v0.4s, v6.4s |
| .long 0x6f00e406 // movi v6.2d, #0x0 |
| .long 0x4e27d421 // fadd v1.4s, v1.4s, v7.4s |
| .long 0x6f00e407 // movi v7.2d, #0x0 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_constant_color_aarch64 |
| .globl _sk_constant_color_aarch64 |
| FUNCTION(_sk_constant_color_aarch64) |
| _sk_constant_color_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xaa0803ea // mov x10, x8 |
| .long 0x4ddfc940 // ld1r {v0.4s}, [x10], #4 |
| .long 0x91002109 // add x9, x8, #0x8 |
| .long 0x91003108 // add x8, x8, #0xc |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0x4d40c903 // ld1r {v3.4s}, [x8] |
| .long 0x4d40c941 // ld1r {v1.4s}, [x10] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_clear_aarch64 |
| .globl _sk_clear_aarch64 |
| FUNCTION(_sk_clear_aarch64) |
| _sk_clear_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6f00e400 // movi v0.2d, #0x0 |
| .long 0x6f00e401 // movi v1.2d, #0x0 |
| .long 0x6f00e402 // movi v2.2d, #0x0 |
| .long 0x6f00e403 // movi v3.2d, #0x0 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_srcatop_aarch64 |
| .globl _sk_srcatop_aarch64 |
| FUNCTION(_sk_srcatop_aarch64) |
| _sk_srcatop_aarch64: |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x6e27dc00 // fmul v0.4s, v0.4s, v7.4s |
| .long 0x6e27dc21 // fmul v1.4s, v1.4s, v7.4s |
| .long 0x6e27dc42 // fmul v2.4s, v2.4s, v7.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| .long 0x4e30ccc2 // fmla v2.4s, v6.4s, v16.4s |
| .long 0x6e27de10 // fmul v16.4s, v16.4s, v7.4s |
| .long 0x4e23ccf0 // fmla v16.4s, v7.4s, v3.4s |
| .long 0x4eb01e03 // mov v3.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_dstatop_aarch64 |
| .globl _sk_dstatop_aarch64 |
| FUNCTION(_sk_dstatop_aarch64) |
| _sk_dstatop_aarch64: |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4ea7d610 // fsub v16.4s, v16.4s, v7.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| .long 0x6e23de10 // fmul v16.4s, v16.4s, v3.4s |
| .long 0x4e23ccf0 // fmla v16.4s, v7.4s, v3.4s |
| .long 0x4e23cc80 // fmla v0.4s, v4.4s, v3.4s |
| .long 0x4e23cca1 // fmla v1.4s, v5.4s, v3.4s |
| .long 0x4e23ccc2 // fmla v2.4s, v6.4s, v3.4s |
| .long 0x4eb01e03 // mov v3.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_srcin_aarch64 |
| .globl _sk_srcin_aarch64 |
| FUNCTION(_sk_srcin_aarch64) |
| _sk_srcin_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e27dc00 // fmul v0.4s, v0.4s, v7.4s |
| .long 0x6e27dc21 // fmul v1.4s, v1.4s, v7.4s |
| .long 0x6e27dc42 // fmul v2.4s, v2.4s, v7.4s |
| .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_dstin_aarch64 |
| .globl _sk_dstin_aarch64 |
| FUNCTION(_sk_dstin_aarch64) |
| _sk_dstin_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e24dc60 // fmul v0.4s, v3.4s, v4.4s |
| .long 0x6e25dc61 // fmul v1.4s, v3.4s, v5.4s |
| .long 0x6e26dc62 // fmul v2.4s, v3.4s, v6.4s |
| .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_srcout_aarch64 |
| .globl _sk_srcout_aarch64 |
| FUNCTION(_sk_srcout_aarch64) |
| _sk_srcout_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4ea7d610 // fsub v16.4s, v16.4s, v7.4s |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_dstout_aarch64 |
| .globl _sk_dstout_aarch64 |
| FUNCTION(_sk_dstout_aarch64) |
| _sk_dstout_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00 |
| .long 0x4ea3d403 // fsub v3.4s, v0.4s, v3.4s |
| .long 0x6e24dc60 // fmul v0.4s, v3.4s, v4.4s |
| .long 0x6e25dc61 // fmul v1.4s, v3.4s, v5.4s |
| .long 0x6e26dc62 // fmul v2.4s, v3.4s, v6.4s |
| .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_srcover_aarch64 |
| .globl _sk_srcover_aarch64 |
| FUNCTION(_sk_srcover_aarch64) |
| _sk_srcover_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x4e24ce00 // fmla v0.4s, v16.4s, v4.4s |
| .long 0x4e25ce01 // fmla v1.4s, v16.4s, v5.4s |
| .long 0x4e26ce02 // fmla v2.4s, v16.4s, v6.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_dstover_aarch64 |
| .globl _sk_dstover_aarch64 |
| FUNCTION(_sk_dstover_aarch64) |
| _sk_dstover_aarch64: |
| .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea41c90 // mov v16.16b, v4.16b |
| .long 0x4ea7d634 // fsub v20.4s, v17.4s, v7.4s |
| .long 0x4ea51cb1 // mov v17.16b, v5.16b |
| .long 0x4ea61cd2 // mov v18.16b, v6.16b |
| .long 0x4ea71cf3 // mov v19.16b, v7.16b |
| .long 0x4e20ce90 // fmla v16.4s, v20.4s, v0.4s |
| .long 0x4e21ce91 // fmla v17.4s, v20.4s, v1.4s |
| .long 0x4e22ce92 // fmla v18.4s, v20.4s, v2.4s |
| .long 0x4e23ce93 // fmla v19.4s, v20.4s, v3.4s |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0x4eb21e42 // mov v2.16b, v18.16b |
| .long 0x4eb31e63 // mov v3.16b, v19.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_modulate_aarch64 |
| .globl _sk_modulate_aarch64 |
| FUNCTION(_sk_modulate_aarch64) |
| _sk_modulate_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e24dc00 // fmul v0.4s, v0.4s, v4.4s |
| .long 0x6e25dc21 // fmul v1.4s, v1.4s, v5.4s |
| .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_multiply_aarch64 |
| .globl _sk_multiply_aarch64 |
| FUNCTION(_sk_multiply_aarch64) |
| _sk_multiply_aarch64: |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4ea7d613 // fsub v19.4s, v16.4s, v7.4s |
| .long 0x4ea3d614 // fsub v20.4s, v16.4s, v3.4s |
| .long 0x6e20de70 // fmul v16.4s, v19.4s, v0.4s |
| .long 0x6e21de71 // fmul v17.4s, v19.4s, v1.4s |
| .long 0x6e22de72 // fmul v18.4s, v19.4s, v2.4s |
| .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e34cc90 // fmla v16.4s, v4.4s, v20.4s |
| .long 0x4e34ccb1 // fmla v17.4s, v5.4s, v20.4s |
| .long 0x4e34ccd2 // fmla v18.4s, v6.4s, v20.4s |
| .long 0x4e34ccf3 // fmla v19.4s, v7.4s, v20.4s |
| .long 0x4e20cc90 // fmla v16.4s, v4.4s, v0.4s |
| .long 0x4e21ccb1 // fmla v17.4s, v5.4s, v1.4s |
| .long 0x4e22ccd2 // fmla v18.4s, v6.4s, v2.4s |
| .long 0x4e23ccf3 // fmla v19.4s, v7.4s, v3.4s |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0x4eb21e42 // mov v2.16b, v18.16b |
| .long 0x4eb31e63 // mov v3.16b, v19.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_plus__aarch64 |
| .globl _sk_plus__aarch64 |
| FUNCTION(_sk_plus__aarch64) |
| _sk_plus__aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| .long 0x4e27d463 // fadd v3.4s, v3.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_screen_aarch64 |
| .globl _sk_screen_aarch64 |
| FUNCTION(_sk_screen_aarch64) |
| _sk_screen_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e24d410 // fadd v16.4s, v0.4s, v4.4s |
| .long 0x4e25d431 // fadd v17.4s, v1.4s, v5.4s |
| .long 0x4e26d452 // fadd v18.4s, v2.4s, v6.4s |
| .long 0x4e27d473 // fadd v19.4s, v3.4s, v7.4s |
| .long 0x4ea4cc10 // fmls v16.4s, v0.4s, v4.4s |
| .long 0x4ea5cc31 // fmls v17.4s, v1.4s, v5.4s |
| .long 0x4ea6cc52 // fmls v18.4s, v2.4s, v6.4s |
| .long 0x4ea7cc73 // fmls v19.4s, v3.4s, v7.4s |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0x4eb21e42 // mov v2.16b, v18.16b |
| .long 0x4eb31e63 // mov v3.16b, v19.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_xor__aarch64 |
| .globl _sk_xor__aarch64 |
| FUNCTION(_sk_xor__aarch64) |
| _sk_xor__aarch64: |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea7d611 // fsub v17.4s, v16.4s, v7.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s |
| .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s |
| .long 0x6e22de22 // fmul v2.4s, v17.4s, v2.4s |
| .long 0x6e23de23 // fmul v3.4s, v17.4s, v3.4s |
| .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| .long 0x4e30ccc2 // fmla v2.4s, v6.4s, v16.4s |
| .long 0x4e30cce3 // fmla v3.4s, v7.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_darken_aarch64 |
| .globl _sk_darken_aarch64 |
| FUNCTION(_sk_darken_aarch64) |
| _sk_darken_aarch64: |
| .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s |
| .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s |
| .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s |
| .long 0x4e31f610 // fmax v16.4s, v16.4s, v17.4s |
| .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s |
| .long 0x4e33f652 // fmax v18.4s, v18.4s, v19.4s |
| .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s |
| .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4e33f631 // fmax v17.4s, v17.4s, v19.4s |
| .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s |
| .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_lighten_aarch64 |
| .globl _sk_lighten_aarch64 |
| FUNCTION(_sk_lighten_aarch64) |
| _sk_lighten_aarch64: |
| .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s |
| .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s |
| .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s |
| .long 0x4eb1f610 // fmin v16.4s, v16.4s, v17.4s |
| .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s |
| .long 0x4eb3f652 // fmin v18.4s, v18.4s, v19.4s |
| .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s |
| .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4eb3f631 // fmin v17.4s, v17.4s, v19.4s |
| .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s |
| .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_difference_aarch64 |
| .globl _sk_difference_aarch64 |
| FUNCTION(_sk_difference_aarch64) |
| _sk_difference_aarch64: |
| .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s |
| .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s |
| .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s |
| .long 0x4eb1f610 // fmin v16.4s, v16.4s, v17.4s |
| .long 0x4eb3f652 // fmin v18.4s, v18.4s, v19.4s |
| .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| .long 0x4e30d610 // fadd v16.4s, v16.4s, v16.4s |
| .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s |
| .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s |
| .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| .long 0x4e32d650 // fadd v16.4s, v18.4s, v18.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4eb3f631 // fmin v17.4s, v17.4s, v19.4s |
| .long 0x4eb0d421 // fsub v1.4s, v1.4s, v16.4s |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| .long 0x4e31d631 // fadd v17.4s, v17.4s, v17.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_exclusion_aarch64 |
| .globl _sk_exclusion_aarch64 |
| FUNCTION(_sk_exclusion_aarch64) |
| _sk_exclusion_aarch64: |
| .long 0x4e24d410 // fadd v16.4s, v0.4s, v4.4s |
| .long 0x6e24dc00 // fmul v0.4s, v0.4s, v4.4s |
| .long 0x4e20d400 // fadd v0.4s, v0.4s, v0.4s |
| .long 0x4ea0d600 // fsub v0.4s, v16.4s, v0.4s |
| .long 0x4e25d430 // fadd v16.4s, v1.4s, v5.4s |
| .long 0x6e25dc21 // fmul v1.4s, v1.4s, v5.4s |
| .long 0x4e21d421 // fadd v1.4s, v1.4s, v1.4s |
| .long 0x4ea1d601 // fsub v1.4s, v16.4s, v1.4s |
| .long 0x4e26d450 // fadd v16.4s, v2.4s, v6.4s |
| .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea2d602 // fsub v2.4s, v16.4s, v2.4s |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_colorburn_aarch64 |
| .globl _sk_colorburn_aarch64 |
| FUNCTION(_sk_colorburn_aarch64) |
| _sk_colorburn_aarch64: |
| .long 0x4ea4d4f3 // fsub v19.4s, v7.4s, v4.4s |
| .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s |
| .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| .long 0x6e20fe73 // fdiv v19.4s, v19.4s, v0.4s |
| .long 0x4ea7d634 // fsub v20.4s, v17.4s, v7.4s |
| .long 0x4eb3f4f3 // fmin v19.4s, v7.4s, v19.4s |
| .long 0x6e20de95 // fmul v21.4s, v20.4s, v0.4s |
| .long 0x4eb3d4f3 // fsub v19.4s, v7.4s, v19.4s |
| .long 0x4e24d6b6 // fadd v22.4s, v21.4s, v4.4s |
| .long 0x4e33cc75 // fmla v21.4s, v3.4s, v19.4s |
| .long 0x4ea5d4f3 // fsub v19.4s, v7.4s, v5.4s |
| .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s |
| .long 0x6e21fe73 // fdiv v19.4s, v19.4s, v1.4s |
| .long 0x4ea0d812 // fcmeq v18.4s, v0.4s, #0.0 |
| .long 0x4eb3f4f3 // fmin v19.4s, v7.4s, v19.4s |
| .long 0x6e751c12 // bsl v18.16b, v0.16b, v21.16b |
| .long 0x6e21de80 // fmul v0.4s, v20.4s, v1.4s |
| .long 0x4eb3d4f3 // fsub v19.4s, v7.4s, v19.4s |
| .long 0x4e25d415 // fadd v21.4s, v0.4s, v5.4s |
| .long 0x4e33cc60 // fmla v0.4s, v3.4s, v19.4s |
| .long 0x4ea0d833 // fcmeq v19.4s, v1.4s, #0.0 |
| .long 0x6e601c33 // bsl v19.16b, v1.16b, v0.16b |
| .long 0x4ea6d4e0 // fsub v0.4s, v7.4s, v6.4s |
| .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s |
| .long 0x6e22fc00 // fdiv v0.4s, v0.4s, v2.4s |
| .long 0x4ea0f4e0 // fmin v0.4s, v7.4s, v0.4s |
| .long 0x6e22de81 // fmul v1.4s, v20.4s, v2.4s |
| .long 0x4ea0d4e0 // fsub v0.4s, v7.4s, v0.4s |
| .long 0x4e26d434 // fadd v20.4s, v1.4s, v6.4s |
| .long 0x4e20cc61 // fmla v1.4s, v3.4s, v0.4s |
| .long 0x4ea0d840 // fcmeq v0.4s, v2.4s, #0.0 |
| .long 0x4ea3d631 // fsub v17.4s, v17.4s, v3.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e27e490 // fcmeq v16.4s, v4.4s, v7.4s |
| .long 0x6e611c40 // bsl v0.16b, v2.16b, v1.16b |
| .long 0x4e31cc92 // fmla v18.4s, v4.4s, v17.4s |
| .long 0x4e27e4a1 // fcmeq v1.4s, v5.4s, v7.4s |
| .long 0x4e27e4c2 // fcmeq v2.4s, v6.4s, v7.4s |
| .long 0x4e31ccb3 // fmla v19.4s, v5.4s, v17.4s |
| .long 0x4e31ccc0 // fmla v0.4s, v6.4s, v17.4s |
| .long 0x6e721ed0 // bsl v16.16b, v22.16b, v18.16b |
| .long 0x6e731ea1 // bsl v1.16b, v21.16b, v19.16b |
| .long 0x6e601e82 // bsl v2.16b, v20.16b, v0.16b |
| .long 0x4e27ce23 // fmla v3.4s, v17.4s, v7.4s |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_colordodge_aarch64 |
| .globl _sk_colordodge_aarch64 |
| FUNCTION(_sk_colordodge_aarch64) |
| _sk_colordodge_aarch64: |
| .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00 |
| .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| .long 0x4ea0d474 // fsub v20.4s, v3.4s, v0.4s |
| .long 0x6e25dc75 // fmul v21.4s, v3.4s, v5.4s |
| .long 0x4ea1d476 // fsub v22.4s, v3.4s, v1.4s |
| .long 0x4ea7d657 // fsub v23.4s, v18.4s, v7.4s |
| .long 0x6e34fe31 // fdiv v17.4s, v17.4s, v20.4s |
| .long 0x6e36feb4 // fdiv v20.4s, v21.4s, v22.4s |
| .long 0x6e20def5 // fmul v21.4s, v23.4s, v0.4s |
| .long 0x4eb1f4f1 // fmin v17.4s, v7.4s, v17.4s |
| .long 0x4e23e413 // fcmeq v19.4s, v0.4s, v3.4s |
| .long 0x4e24d6b6 // fadd v22.4s, v21.4s, v4.4s |
| .long 0x4e31cc75 // fmla v21.4s, v3.4s, v17.4s |
| .long 0x6e751c13 // bsl v19.16b, v0.16b, v21.16b |
| .long 0x6e21dee0 // fmul v0.4s, v23.4s, v1.4s |
| .long 0x4eb4f4f4 // fmin v20.4s, v7.4s, v20.4s |
| .long 0x4e25d415 // fadd v21.4s, v0.4s, v5.4s |
| .long 0x4e34cc60 // fmla v0.4s, v3.4s, v20.4s |
| .long 0x4e23e434 // fcmeq v20.4s, v1.4s, v3.4s |
| .long 0x6e601c34 // bsl v20.16b, v1.16b, v0.16b |
| .long 0x6e26dc60 // fmul v0.4s, v3.4s, v6.4s |
| .long 0x4ea2d461 // fsub v1.4s, v3.4s, v2.4s |
| .long 0x6e21fc00 // fdiv v0.4s, v0.4s, v1.4s |
| .long 0x6e22dee1 // fmul v1.4s, v23.4s, v2.4s |
| .long 0x4ea0f4e0 // fmin v0.4s, v7.4s, v0.4s |
| .long 0x4e26d437 // fadd v23.4s, v1.4s, v6.4s |
| .long 0x4e20cc61 // fmla v1.4s, v3.4s, v0.4s |
| .long 0x4e23e440 // fcmeq v0.4s, v2.4s, v3.4s |
| .long 0x6e611c40 // bsl v0.16b, v2.16b, v1.16b |
| .long 0x4ea3d641 // fsub v1.4s, v18.4s, v3.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea0d890 // fcmeq v16.4s, v4.4s, #0.0 |
| .long 0x4ea0d8b1 // fcmeq v17.4s, v5.4s, #0.0 |
| .long 0x4e21cc93 // fmla v19.4s, v4.4s, v1.4s |
| .long 0x4e21ccb4 // fmla v20.4s, v5.4s, v1.4s |
| .long 0x4ea0d8c2 // fcmeq v2.4s, v6.4s, #0.0 |
| .long 0x4e21ccc0 // fmla v0.4s, v6.4s, v1.4s |
| .long 0x6e731ed0 // bsl v16.16b, v22.16b, v19.16b |
| .long 0x6e741eb1 // bsl v17.16b, v21.16b, v20.16b |
| .long 0x6e601ee2 // bsl v2.16b, v23.16b, v0.16b |
| .long 0x4e27cc23 // fmla v3.4s, v1.4s, v7.4s |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_hardlight_aarch64 |
| .globl _sk_hardlight_aarch64 |
| FUNCTION(_sk_hardlight_aarch64) |
| _sk_hardlight_aarch64: |
| .long 0x4ea4d4f4 // fsub v20.4s, v7.4s, v4.4s |
| .long 0x4ea0d475 // fsub v21.4s, v3.4s, v0.4s |
| .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s |
| .long 0x4e20d411 // fadd v17.4s, v0.4s, v0.4s |
| .long 0x6e24dc12 // fmul v18.4s, v0.4s, v4.4s |
| .long 0x6e27dc73 // fmul v19.4s, v3.4s, v7.4s |
| .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s |
| .long 0x6e31e471 // fcmge v17.4s, v3.4s, v17.4s |
| .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| .long 0x4eb4d674 // fsub v20.4s, v19.4s, v20.4s |
| .long 0x6e741e51 // bsl v17.16b, v18.16b, v20.16b |
| .long 0x4ea5d4f2 // fsub v18.4s, v7.4s, v5.4s |
| .long 0x4ea1d474 // fsub v20.4s, v3.4s, v1.4s |
| .long 0x6e32de92 // fmul v18.4s, v20.4s, v18.4s |
| .long 0x4e21d436 // fadd v22.4s, v1.4s, v1.4s |
| .long 0x6e25dc35 // fmul v21.4s, v1.4s, v5.4s |
| .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| .long 0x6e36e476 // fcmge v22.4s, v3.4s, v22.4s |
| .long 0x4e35d6b5 // fadd v21.4s, v21.4s, v21.4s |
| .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x6e721eb6 // bsl v22.16b, v21.16b, v18.16b |
| .long 0x4ea6d4f2 // fsub v18.4s, v7.4s, v6.4s |
| .long 0x4ea2d475 // fsub v21.4s, v3.4s, v2.4s |
| .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| .long 0x4ea7d615 // fsub v21.4s, v16.4s, v7.4s |
| .long 0x4e22d454 // fadd v20.4s, v2.4s, v2.4s |
| .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s |
| .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s |
| .long 0x6e22deb5 // fmul v21.4s, v21.4s, v2.4s |
| .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x6e34e474 // fcmge v20.4s, v3.4s, v20.4s |
| .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s |
| .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| .long 0x4e30ccd5 // fmla v21.4s, v6.4s, v16.4s |
| .long 0x6e721c54 // bsl v20.16b, v2.16b, v18.16b |
| .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s |
| .long 0x4e36d421 // fadd v1.4s, v1.4s, v22.4s |
| .long 0x4e34d6a2 // fadd v2.4s, v21.4s, v20.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_overlay_aarch64 |
| .globl _sk_overlay_aarch64 |
| FUNCTION(_sk_overlay_aarch64) |
| _sk_overlay_aarch64: |
| .long 0x4ea4d4f4 // fsub v20.4s, v7.4s, v4.4s |
| .long 0x4ea0d475 // fsub v21.4s, v3.4s, v0.4s |
| .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s |
| .long 0x4e24d491 // fadd v17.4s, v4.4s, v4.4s |
| .long 0x6e24dc12 // fmul v18.4s, v0.4s, v4.4s |
| .long 0x6e27dc73 // fmul v19.4s, v3.4s, v7.4s |
| .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s |
| .long 0x6e31e4f1 // fcmge v17.4s, v7.4s, v17.4s |
| .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| .long 0x4eb4d674 // fsub v20.4s, v19.4s, v20.4s |
| .long 0x6e741e51 // bsl v17.16b, v18.16b, v20.16b |
| .long 0x4ea5d4f2 // fsub v18.4s, v7.4s, v5.4s |
| .long 0x4ea1d474 // fsub v20.4s, v3.4s, v1.4s |
| .long 0x6e32de92 // fmul v18.4s, v20.4s, v18.4s |
| .long 0x4e25d4b6 // fadd v22.4s, v5.4s, v5.4s |
| .long 0x6e25dc35 // fmul v21.4s, v1.4s, v5.4s |
| .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| .long 0x6e36e4f6 // fcmge v22.4s, v7.4s, v22.4s |
| .long 0x4e35d6b5 // fadd v21.4s, v21.4s, v21.4s |
| .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x6e721eb6 // bsl v22.16b, v21.16b, v18.16b |
| .long 0x4ea6d4f2 // fsub v18.4s, v7.4s, v6.4s |
| .long 0x4ea2d475 // fsub v21.4s, v3.4s, v2.4s |
| .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| .long 0x4ea7d615 // fsub v21.4s, v16.4s, v7.4s |
| .long 0x4e26d4d4 // fadd v20.4s, v6.4s, v6.4s |
| .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s |
| .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s |
| .long 0x6e22deb5 // fmul v21.4s, v21.4s, v2.4s |
| .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| .long 0x6e34e4f4 // fcmge v20.4s, v7.4s, v20.4s |
| .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s |
| .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| .long 0x4e30ccd5 // fmla v21.4s, v6.4s, v16.4s |
| .long 0x6e721c54 // bsl v20.16b, v2.16b, v18.16b |
| .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s |
| .long 0x4e36d421 // fadd v1.4s, v1.4s, v22.4s |
| .long 0x4e34d6a2 // fadd v2.4s, v21.4s, v20.4s |
| .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_softlight_aarch64 |
| .globl _sk_softlight_aarch64 |
| FUNCTION(_sk_softlight_aarch64) |
| _sk_softlight_aarch64: |
| .long 0x4ea0c8f5 // fcmgt v21.4s, v7.4s, #0.0 |
| .long 0x6e27fc96 // fdiv v22.4s, v4.4s, v7.4s |
| .long 0x6e27fcb8 // fdiv v24.4s, v5.4s, v7.4s |
| .long 0x6e27fcd9 // fdiv v25.4s, v6.4s, v7.4s |
| .long 0x4e351ed6 // and v22.16b, v22.16b, v21.16b |
| .long 0x4e351f18 // and v24.16b, v24.16b, v21.16b |
| .long 0x4e351f35 // and v21.16b, v25.16b, v21.16b |
| .long 0x6ea1dad9 // frsqrte v25.4s, v22.4s |
| .long 0x6e39df3d // fmul v29.4s, v25.4s, v25.4s |
| .long 0x4ebdfedd // frsqrts v29.4s, v22.4s, v29.4s |
| .long 0x6e3ddf39 // fmul v25.4s, v25.4s, v29.4s |
| .long 0x4ea1db3d // frecpe v29.4s, v25.4s |
| .long 0x6ea0fada // fneg v26.4s, v22.4s |
| .long 0x6ea1db1b // frsqrte v27.4s, v24.4s |
| .long 0x4e3dff39 // frecps v25.4s, v25.4s, v29.4s |
| .long 0x4e3dcf3a // fmla v26.4s, v25.4s, v29.4s |
| .long 0x6e3bdf7d // fmul v29.4s, v27.4s, v27.4s |
| .long 0x4ebdff1d // frsqrts v29.4s, v24.4s, v29.4s |
| .long 0x6e3ddf7b // fmul v27.4s, v27.4s, v29.4s |
| .long 0x4ea1db7d // frecpe v29.4s, v27.4s |
| .long 0x6ea0fb1c // fneg v28.4s, v24.4s |
| .long 0x6ea1dab9 // frsqrte v25.4s, v21.4s |
| .long 0x4e3dff7b // frecps v27.4s, v27.4s, v29.4s |
| .long 0x4e3dcf7c // fmla v28.4s, v27.4s, v29.4s |
| .long 0x6e39df3d // fmul v29.4s, v25.4s, v25.4s |
| .long 0x4ebdfebd // frsqrts v29.4s, v21.4s, v29.4s |
| .long 0x6e3ddf39 // fmul v25.4s, v25.4s, v29.4s |
| .long 0x4ea1db3d // frecpe v29.4s, v25.4s |
| .long 0x6ea0fabb // fneg v27.4s, v21.4s |
| .long 0x4e3dff39 // frecps v25.4s, v25.4s, v29.4s |
| .long 0x4e3dcf3b // fmla v27.4s, v25.4s, v29.4s |
| .long 0x4e36d6d9 // fadd v25.4s, v22.4s, v22.4s |
| .long 0x4f07f613 // fmov v19.4s, #-1.000000000000000000e+00 |
| .long 0x4e39d739 // fadd v25.4s, v25.4s, v25.4s |
| .long 0x4e24d497 // fadd v23.4s, v4.4s, v4.4s |
| .long 0x4e33d6dd // fadd v29.4s, v22.4s, v19.4s |
| .long 0x4e39cf39 // fmla v25.4s, v25.4s, v25.4s |
| .long 0x4f00f794 // fmov v20.4s, #7.000000000000000000e+00 |
| .long 0x6e39dfb9 // fmul v25.4s, v29.4s, v25.4s |
| .long 0x4e37d6f7 // fadd v23.4s, v23.4s, v23.4s |
| .long 0x6e37e4f7 // fcmge v23.4s, v7.4s, v23.4s |
| .long 0x4e36ce99 // fmla v25.4s, v20.4s, v22.4s |
| .long 0x6e7a1f37 // bsl v23.16b, v25.16b, v26.16b |
| .long 0x4e38d719 // fadd v25.4s, v24.4s, v24.4s |
| .long 0x4e39d739 // fadd v25.4s, v25.4s, v25.4s |
| .long 0x4e33d71a // fadd v26.4s, v24.4s, v19.4s |
| .long 0x4e39cf39 // fmla v25.4s, v25.4s, v25.4s |
| .long 0x6e39df59 // fmul v25.4s, v26.4s, v25.4s |
| .long 0x4e25d4ba // fadd v26.4s, v5.4s, v5.4s |
| .long 0x4e3ad75a // fadd v26.4s, v26.4s, v26.4s |
| .long 0x6e3ae4fa // fcmge v26.4s, v7.4s, v26.4s |
| .long 0x4e38ce99 // fmla v25.4s, v20.4s, v24.4s |
| .long 0x6e7c1f3a // bsl v26.16b, v25.16b, v28.16b |
| .long 0x4e35d6bc // fadd v28.4s, v21.4s, v21.4s |
| .long 0x4e3cd79c // fadd v28.4s, v28.4s, v28.4s |
| .long 0x4e33d6b3 // fadd v19.4s, v21.4s, v19.4s |
| .long 0x4e3ccf9c // fmla v28.4s, v28.4s, v28.4s |
| .long 0x6e3cde73 // fmul v19.4s, v19.4s, v28.4s |
| .long 0x4e35ce93 // fmla v19.4s, v20.4s, v21.4s |
| .long 0x4e26d4d4 // fadd v20.4s, v6.4s, v6.4s |
| .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s |
| .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00 |
| .long 0x6e34e4f4 // fcmge v20.4s, v7.4s, v20.4s |
| .long 0x4e20d411 // fadd v17.4s, v0.4s, v0.4s |
| .long 0x6e7b1e74 // bsl v20.16b, v19.16b, v27.16b |
| .long 0x4ea7d65b // fsub v27.4s, v18.4s, v7.4s |
| .long 0x4ea31c70 // mov v16.16b, v3.16b |
| .long 0x4e21d43d // fadd v29.4s, v1.4s, v1.4s |
| .long 0x4e22d45c // fadd v28.4s, v2.4s, v2.4s |
| .long 0x6e20df60 // fmul v0.4s, v27.4s, v0.4s |
| .long 0x6e21df61 // fmul v1.4s, v27.4s, v1.4s |
| .long 0x6e22df62 // fmul v2.4s, v27.4s, v2.4s |
| .long 0x4ea3d63b // fsub v27.4s, v17.4s, v3.4s |
| .long 0x4eb6d656 // fsub v22.4s, v18.4s, v22.4s |
| .long 0x4ea31c79 // mov v25.16b, v3.16b |
| .long 0x4e3bced0 // fmla v16.4s, v22.4s, v27.4s |
| .long 0x4ea3d7b6 // fsub v22.4s, v29.4s, v3.4s |
| .long 0x4eb8d658 // fsub v24.4s, v18.4s, v24.4s |
| .long 0x4ea31c73 // mov v19.16b, v3.16b |
| .long 0x4e36cf19 // fmla v25.4s, v24.4s, v22.4s |
| .long 0x4ea3d798 // fsub v24.4s, v28.4s, v3.4s |
| .long 0x4eb5d655 // fsub v21.4s, v18.4s, v21.4s |
| .long 0x4e38ceb3 // fmla v19.4s, v21.4s, v24.4s |
| .long 0x6e27df7b // fmul v27.4s, v27.4s, v7.4s |
| .long 0x6e27ded6 // fmul v22.4s, v22.4s, v7.4s |
| .long 0x6e27df18 // fmul v24.4s, v24.4s, v7.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e37df77 // fmul v23.4s, v27.4s, v23.4s |
| .long 0x6e3aded6 // fmul v22.4s, v22.4s, v26.4s |
| .long 0x6e34df14 // fmul v20.4s, v24.4s, v20.4s |
| .long 0x4ea3d652 // fsub v18.4s, v18.4s, v3.4s |
| .long 0x6e31e471 // fcmge v17.4s, v3.4s, v17.4s |
| .long 0x6e3de475 // fcmge v21.4s, v3.4s, v29.4s |
| .long 0x6e3ce47c // fcmge v28.4s, v3.4s, v28.4s |
| .long 0x6e24de10 // fmul v16.4s, v16.4s, v4.4s |
| .long 0x6e25df39 // fmul v25.4s, v25.4s, v5.4s |
| .long 0x6e26de73 // fmul v19.4s, v19.4s, v6.4s |
| .long 0x4e23cc97 // fmla v23.4s, v4.4s, v3.4s |
| .long 0x4e23ccb6 // fmla v22.4s, v5.4s, v3.4s |
| .long 0x4e23ccd4 // fmla v20.4s, v6.4s, v3.4s |
| .long 0x4e32cc80 // fmla v0.4s, v4.4s, v18.4s |
| .long 0x4e32cca1 // fmla v1.4s, v5.4s, v18.4s |
| .long 0x4e32ccc2 // fmla v2.4s, v6.4s, v18.4s |
| .long 0x6e771e11 // bsl v17.16b, v16.16b, v23.16b |
| .long 0x6e761f35 // bsl v21.16b, v25.16b, v22.16b |
| .long 0x6e741e7c // bsl v28.16b, v19.16b, v20.16b |
| .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s |
| .long 0x4e35d421 // fadd v1.4s, v1.4s, v21.4s |
| .long 0x4e3cd442 // fadd v2.4s, v2.4s, v28.4s |
| .long 0x4e27ce43 // fmla v3.4s, v18.4s, v7.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_clamp_0_aarch64 |
| .globl _sk_clamp_0_aarch64 |
| FUNCTION(_sk_clamp_0_aarch64) |
| _sk_clamp_0_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6f00e410 // movi v16.2d, #0x0 |
| .long 0x4e30f400 // fmax v0.4s, v0.4s, v16.4s |
| .long 0x4e30f421 // fmax v1.4s, v1.4s, v16.4s |
| .long 0x4e30f442 // fmax v2.4s, v2.4s, v16.4s |
| .long 0x4e30f463 // fmax v3.4s, v3.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_clamp_1_aarch64 |
| .globl _sk_clamp_1_aarch64 |
| FUNCTION(_sk_clamp_1_aarch64) |
| _sk_clamp_1_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s |
| .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s |
| .long 0x4eb0f442 // fmin v2.4s, v2.4s, v16.4s |
| .long 0x4eb0f463 // fmin v3.4s, v3.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_clamp_a_aarch64 |
| .globl _sk_clamp_a_aarch64 |
| FUNCTION(_sk_clamp_a_aarch64) |
| _sk_clamp_a_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| .long 0x4eb0f463 // fmin v3.4s, v3.4s, v16.4s |
| .long 0x4ea3f400 // fmin v0.4s, v0.4s, v3.4s |
| .long 0x4ea3f421 // fmin v1.4s, v1.4s, v3.4s |
| .long 0x4ea3f442 // fmin v2.4s, v2.4s, v3.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_set_rgb_aarch64 |
| .globl _sk_set_rgb_aarch64 |
| FUNCTION(_sk_set_rgb_aarch64) |
| _sk_set_rgb_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0x4ddfc920 // ld1r {v0.4s}, [x9], #4 |
| .long 0x91002108 // add x8, x8, #0x8 |
| .long 0x4d40c902 // ld1r {v2.4s}, [x8] |
| .long 0x4d40c921 // ld1r {v1.4s}, [x9] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_swap_rb_aarch64 |
| .globl _sk_swap_rb_aarch64 |
| FUNCTION(_sk_swap_rb_aarch64) |
| _sk_swap_rb_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea01c10 // mov v16.16b, v0.16b |
| .long 0x4ea21c40 // mov v0.16b, v2.16b |
| .long 0x4eb01e02 // mov v2.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_swap_aarch64 |
| .globl _sk_swap_aarch64 |
| FUNCTION(_sk_swap_aarch64) |
| _sk_swap_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea31c70 // mov v16.16b, v3.16b |
| .long 0x4ea21c51 // mov v17.16b, v2.16b |
| .long 0x4ea11c32 // mov v18.16b, v1.16b |
| .long 0x4ea01c13 // mov v19.16b, v0.16b |
| .long 0x4ea41c80 // mov v0.16b, v4.16b |
| .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| .long 0x4eb31e64 // mov v4.16b, v19.16b |
| .long 0x4eb21e45 // mov v5.16b, v18.16b |
| .long 0x4eb11e26 // mov v6.16b, v17.16b |
| .long 0x4eb01e07 // mov v7.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_move_src_dst_aarch64 |
| .globl _sk_move_src_dst_aarch64 |
| FUNCTION(_sk_move_src_dst_aarch64) |
| _sk_move_src_dst_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea01c04 // mov v4.16b, v0.16b |
| .long 0x4ea11c25 // mov v5.16b, v1.16b |
| .long 0x4ea21c46 // mov v6.16b, v2.16b |
| .long 0x4ea31c67 // mov v7.16b, v3.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_move_dst_src_aarch64 |
| .globl _sk_move_dst_src_aarch64 |
| FUNCTION(_sk_move_dst_src_aarch64) |
| _sk_move_dst_src_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea41c80 // mov v0.16b, v4.16b |
| .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_premul_aarch64 |
| .globl _sk_premul_aarch64 |
| FUNCTION(_sk_premul_aarch64) |
| _sk_premul_aarch64: |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s |
| .long 0x6e23dc21 // fmul v1.4s, v1.4s, v3.4s |
| .long 0x6e23dc42 // fmul v2.4s, v2.4s, v3.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_unpremul_aarch64 |
| .globl _sk_unpremul_aarch64 |
| FUNCTION(_sk_unpremul_aarch64) |
| _sk_unpremul_aarch64: |
| .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4ea0d870 // fcmeq v16.4s, v3.4s, #0.0 |
| .long 0x6e23fe31 // fdiv v17.4s, v17.4s, v3.4s |
| .long 0x4e701e30 // bic v16.16b, v17.16b, v16.16b |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_from_srgb_aarch64 |
| .globl _sk_from_srgb_aarch64 |
| FUNCTION(_sk_from_srgb_aarch64) |
| _sk_from_srgb_aarch64: |
| .long 0x52a7d328 // mov w8, #0x3e990000 |
| .long 0x72933348 // movk w8, #0x999a |
| .long 0x4e040d10 // dup v16.4s, w8 |
| .long 0x52a7e648 // mov w8, #0x3f320000 |
| .long 0x7291eb88 // movk w8, #0x8f5c |
| .long 0x4e040d11 // dup v17.4s, w8 |
| .long 0x52a76468 // mov w8, #0x3b230000 |
| .long 0x729ae148 // movk w8, #0xd70a |
| .long 0x4e040d12 // dup v18.4s, w8 |
| .long 0x52a7b3c8 // mov w8, #0x3d9e0000 |
| .long 0x72907228 // movk w8, #0x8391 |
| .long 0x6e22dc54 // fmul v20.4s, v2.4s, v2.4s |
| .long 0x4eb11e35 // mov v21.16b, v17.16b |
| .long 0x4eb11e37 // mov v23.16b, v17.16b |
| .long 0x4e22ce11 // fmla v17.4s, v16.4s, v2.4s |
| .long 0x4eb21e56 // mov v22.16b, v18.16b |
| .long 0x4eb21e58 // mov v24.16b, v18.16b |
| .long 0x4e34ce32 // fmla v18.4s, v17.4s, v20.4s |
| .long 0x4e040d11 // dup v17.4s, w8 |
| .long 0x52a7ac28 // mov w8, #0x3d610000 |
| .long 0x6e20dc13 // fmul v19.4s, v0.4s, v0.4s |
| .long 0x7288f5c8 // movk w8, #0x47ae |
| .long 0x4e20ce15 // fmla v21.4s, v16.4s, v0.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e21dc34 // fmul v20.4s, v1.4s, v1.4s |
| .long 0x4e33ceb6 // fmla v22.4s, v21.4s, v19.4s |
| .long 0x4e040d13 // dup v19.4s, w8 |
| .long 0x4e21ce17 // fmla v23.4s, v16.4s, v1.4s |
| .long 0x6e31dc15 // fmul v21.4s, v0.4s, v17.4s |
| .long 0x6ea0e660 // fcmgt v0.4s, v19.4s, v0.4s |
| .long 0x6e31dc30 // fmul v16.4s, v1.4s, v17.4s |
| .long 0x6ea1e661 // fcmgt v1.4s, v19.4s, v1.4s |
| .long 0x6e31dc51 // fmul v17.4s, v2.4s, v17.4s |
| .long 0x6ea2e662 // fcmgt v2.4s, v19.4s, v2.4s |
| .long 0x4e34cef8 // fmla v24.4s, v23.4s, v20.4s |
| .long 0x6e761ea0 // bsl v0.16b, v21.16b, v22.16b |
| .long 0x6e781e01 // bsl v1.16b, v16.16b, v24.16b |
| .long 0x6e721e22 // bsl v2.16b, v17.16b, v18.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_to_srgb_aarch64 |
| .globl _sk_to_srgb_aarch64 |
| FUNCTION(_sk_to_srgb_aarch64) |
| _sk_to_srgb_aarch64: |
| .long 0x52a828e8 // mov w8, #0x41470000 |
| .long 0x728b8528 // movk w8, #0x5c29 |
| .long 0x4e040d12 // dup v18.4s, w8 |
| .long 0x52a7e608 // mov w8, #0x3f300000 |
| .long 0x728df9c8 // movk w8, #0x6fce |
| .long 0x6ea1d811 // frsqrte v17.4s, v0.4s |
| .long 0x4e040d13 // dup v19.4s, w8 |
| .long 0x52b7b948 // mov w8, #0xbdca0000 |
| .long 0x728af508 // movk w8, #0x57a8 |
| .long 0x6ea1d834 // frsqrte v20.4s, v1.4s |
| .long 0x6e31de36 // fmul v22.4s, v17.4s, v17.4s |
| .long 0x4e040d10 // dup v16.4s, w8 |
| .long 0x52a77188 // mov w8, #0x3b8c0000 |
| .long 0x6ea1d855 // frsqrte v21.4s, v2.4s |
| .long 0x6e34de98 // fmul v24.4s, v20.4s, v20.4s |
| .long 0x4eb6fc16 // frsqrts v22.4s, v0.4s, v22.4s |
| .long 0x729ce088 // movk w8, #0xe704 |
| .long 0x6e35deb9 // fmul v25.4s, v21.4s, v21.4s |
| .long 0x4eb8fc38 // frsqrts v24.4s, v1.4s, v24.4s |
| .long 0x6e36de31 // fmul v17.4s, v17.4s, v22.4s |
| .long 0x4e040d17 // dup v23.4s, w8 |
| .long 0x4eb9fc59 // frsqrts v25.4s, v2.4s, v25.4s |
| .long 0x6e38de94 // fmul v20.4s, v20.4s, v24.4s |
| .long 0x4ea1da36 // frecpe v22.4s, v17.4s |
| .long 0x6e32dc1a // fmul v26.4s, v0.4s, v18.4s |
| .long 0x6ea0e6e0 // fcmgt v0.4s, v23.4s, v0.4s |
| .long 0x6e32dc3c // fmul v28.4s, v1.4s, v18.4s |
| .long 0x6ea1e6e1 // fcmgt v1.4s, v23.4s, v1.4s |
| .long 0x6e32dc52 // fmul v18.4s, v2.4s, v18.4s |
| .long 0x6ea2e6e2 // fcmgt v2.4s, v23.4s, v2.4s |
| .long 0x6e39deb5 // fmul v21.4s, v21.4s, v25.4s |
| .long 0x4ea1da97 // frecpe v23.4s, v20.4s |
| .long 0x4e36fe39 // frecps v25.4s, v17.4s, v22.4s |
| .long 0x4ea1dab8 // frecpe v24.4s, v21.4s |
| .long 0x6e39ded6 // fmul v22.4s, v22.4s, v25.4s |
| .long 0x4e37fe99 // frecps v25.4s, v20.4s, v23.4s |
| .long 0x4eb01e1b // mov v27.16b, v16.16b |
| .long 0x6e39def7 // fmul v23.4s, v23.4s, v25.4s |
| .long 0x4e38feb9 // frecps v25.4s, v21.4s, v24.4s |
| .long 0x6e39df18 // fmul v24.4s, v24.4s, v25.4s |
| .long 0x4eb01e19 // mov v25.16b, v16.16b |
| .long 0x4e36ce7b // fmla v27.4s, v19.4s, v22.4s |
| .long 0x6ea1da36 // frsqrte v22.4s, v17.4s |
| .long 0x4e37ce79 // fmla v25.4s, v19.4s, v23.4s |
| .long 0x6ea1da97 // frsqrte v23.4s, v20.4s |
| .long 0x4e38ce70 // fmla v16.4s, v19.4s, v24.4s |
| .long 0x6e36ded8 // fmul v24.4s, v22.4s, v22.4s |
| .long 0x6ea1dab3 // frsqrte v19.4s, v21.4s |
| .long 0x4eb8fe31 // frsqrts v17.4s, v17.4s, v24.4s |
| .long 0x6e37def8 // fmul v24.4s, v23.4s, v23.4s |
| .long 0x4eb8fe94 // frsqrts v20.4s, v20.4s, v24.4s |
| .long 0x6e33de78 // fmul v24.4s, v19.4s, v19.4s |
| .long 0x52a7da48 // mov w8, #0x3ed20000 |
| .long 0x4eb8feb5 // frsqrts v21.4s, v21.4s, v24.4s |
| .long 0x7290f848 // movk w8, #0x87c2 |
| .long 0x6e31ded1 // fmul v17.4s, v22.4s, v17.4s |
| .long 0x6e34def4 // fmul v20.4s, v23.4s, v20.4s |
| .long 0x6e35de73 // fmul v19.4s, v19.4s, v21.4s |
| .long 0x4e040d15 // dup v21.4s, w8 |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e31cebb // fmla v27.4s, v21.4s, v17.4s |
| .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| .long 0x4e34ceb9 // fmla v25.4s, v21.4s, v20.4s |
| .long 0x4e33ceb0 // fmla v16.4s, v21.4s, v19.4s |
| .long 0x4ebbf633 // fmin v19.4s, v17.4s, v27.4s |
| .long 0x4eb9f634 // fmin v20.4s, v17.4s, v25.4s |
| .long 0x4eb0f630 // fmin v16.4s, v17.4s, v16.4s |
| .long 0x6e731f40 // bsl v0.16b, v26.16b, v19.16b |
| .long 0x6e741f81 // bsl v1.16b, v28.16b, v20.16b |
| .long 0x6e701e42 // bsl v2.16b, v18.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_from_2dot2_aarch64 |
| .globl _sk_from_2dot2_aarch64 |
| FUNCTION(_sk_from_2dot2_aarch64) |
| _sk_from_2dot2_aarch64: |
| .long 0x6ea1d810 // frsqrte v16.4s, v0.4s |
| .long 0x6ea1d832 // frsqrte v18.4s, v1.4s |
| .long 0x6e30de15 // fmul v21.4s, v16.4s, v16.4s |
| .long 0x6e20dc11 // fmul v17.4s, v0.4s, v0.4s |
| .long 0x6ea1d854 // frsqrte v20.4s, v2.4s |
| .long 0x6e32de56 // fmul v22.4s, v18.4s, v18.4s |
| .long 0x4eb5fc00 // frsqrts v0.4s, v0.4s, v21.4s |
| .long 0x6e21dc33 // fmul v19.4s, v1.4s, v1.4s |
| .long 0x6e34de97 // fmul v23.4s, v20.4s, v20.4s |
| .long 0x4eb6fc21 // frsqrts v1.4s, v1.4s, v22.4s |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x4eb7fc55 // frsqrts v21.4s, v2.4s, v23.4s |
| .long 0x6e21de41 // fmul v1.4s, v18.4s, v1.4s |
| .long 0x6ea1d812 // frsqrte v18.4s, v0.4s |
| .long 0x6e35de90 // fmul v16.4s, v20.4s, v21.4s |
| .long 0x6ea1d834 // frsqrte v20.4s, v1.4s |
| .long 0x6e32de56 // fmul v22.4s, v18.4s, v18.4s |
| .long 0x6ea1da15 // frsqrte v21.4s, v16.4s |
| .long 0x6e34de97 // fmul v23.4s, v20.4s, v20.4s |
| .long 0x4eb6fc00 // frsqrts v0.4s, v0.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x4eb7fc21 // frsqrts v1.4s, v1.4s, v23.4s |
| .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6e21de81 // fmul v1.4s, v20.4s, v1.4s |
| .long 0x6ea1d812 // frsqrte v18.4s, v0.4s |
| .long 0x6e30deb0 // fmul v16.4s, v21.4s, v16.4s |
| .long 0x6ea1d834 // frsqrte v20.4s, v1.4s |
| .long 0x6e32de56 // fmul v22.4s, v18.4s, v18.4s |
| .long 0x6ea1da15 // frsqrte v21.4s, v16.4s |
| .long 0x6e34de97 // fmul v23.4s, v20.4s, v20.4s |
| .long 0x4eb6fc00 // frsqrts v0.4s, v0.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x4eb7fc21 // frsqrts v1.4s, v1.4s, v23.4s |
| .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6e21de81 // fmul v1.4s, v20.4s, v1.4s |
| .long 0x6ea1d812 // frsqrte v18.4s, v0.4s |
| .long 0x6e30deb0 // fmul v16.4s, v21.4s, v16.4s |
| .long 0x6ea1d834 // frsqrte v20.4s, v1.4s |
| .long 0x6e32de56 // fmul v22.4s, v18.4s, v18.4s |
| .long 0x6ea1da15 // frsqrte v21.4s, v16.4s |
| .long 0x6e34de97 // fmul v23.4s, v20.4s, v20.4s |
| .long 0x4eb6fc00 // frsqrts v0.4s, v0.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x4eb7fc21 // frsqrts v1.4s, v1.4s, v23.4s |
| .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6e21de81 // fmul v1.4s, v20.4s, v1.4s |
| .long 0x6ea1d812 // frsqrte v18.4s, v0.4s |
| .long 0x6e20dc14 // fmul v20.4s, v0.4s, v0.4s |
| .long 0x6e30deb0 // fmul v16.4s, v21.4s, v16.4s |
| .long 0x6ea1d835 // frsqrte v21.4s, v1.4s |
| .long 0x6e21dc36 // fmul v22.4s, v1.4s, v1.4s |
| .long 0x6e32de57 // fmul v23.4s, v18.4s, v18.4s |
| .long 0x6e34dc14 // fmul v20.4s, v0.4s, v20.4s |
| .long 0x4eb7fc00 // frsqrts v0.4s, v0.4s, v23.4s |
| .long 0x6ea1da17 // frsqrte v23.4s, v16.4s |
| .long 0x6e34de31 // fmul v17.4s, v17.4s, v20.4s |
| .long 0x6e35deb4 // fmul v20.4s, v21.4s, v21.4s |
| .long 0x6e36dc36 // fmul v22.4s, v1.4s, v22.4s |
| .long 0x4eb4fc21 // frsqrts v1.4s, v1.4s, v20.4s |
| .long 0x6e30de14 // fmul v20.4s, v16.4s, v16.4s |
| .long 0x6e36de73 // fmul v19.4s, v19.4s, v22.4s |
| .long 0x6e37def6 // fmul v22.4s, v23.4s, v23.4s |
| .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s |
| .long 0x6e34de14 // fmul v20.4s, v16.4s, v20.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6e22dc42 // fmul v2.4s, v2.4s, v2.4s |
| .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s |
| .long 0x6ea1d812 // frsqrte v18.4s, v0.4s |
| .long 0x6e34dc42 // fmul v2.4s, v2.4s, v20.4s |
| .long 0x6e30def0 // fmul v16.4s, v23.4s, v16.4s |
| .long 0x6ea1d834 // frsqrte v20.4s, v1.4s |
| .long 0x6e32de56 // fmul v22.4s, v18.4s, v18.4s |
| .long 0x6ea1da15 // frsqrte v21.4s, v16.4s |
| .long 0x4eb6fc00 // frsqrts v0.4s, v0.4s, v22.4s |
| .long 0x6e34de96 // fmul v22.4s, v20.4s, v20.4s |
| .long 0x4eb6fc21 // frsqrts v1.4s, v1.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e20de40 // fmul v0.4s, v18.4s, v0.4s |
| .long 0x6e21de81 // fmul v1.4s, v20.4s, v1.4s |
| .long 0x6e30deb0 // fmul v16.4s, v21.4s, v16.4s |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s |
| .long 0x6e21de61 // fmul v1.4s, v19.4s, v1.4s |
| .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s |
| .long 0x4e32f400 // fmax v0.4s, v0.4s, v18.4s |
| .long 0x4e32f421 // fmax v1.4s, v1.4s, v18.4s |
| .long 0x4e32f442 // fmax v2.4s, v2.4s, v18.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_to_2dot2_aarch64 |
| .globl _sk_to_2dot2_aarch64 |
| FUNCTION(_sk_to_2dot2_aarch64) |
| _sk_to_2dot2_aarch64: |
| .long 0x6ea1d810 // frsqrte v16.4s, v0.4s |
| .long 0x6e30de13 // fmul v19.4s, v16.4s, v16.4s |
| .long 0x6ea1d831 // frsqrte v17.4s, v1.4s |
| .long 0x4eb3fc00 // frsqrts v0.4s, v0.4s, v19.4s |
| .long 0x6ea1d852 // frsqrte v18.4s, v2.4s |
| .long 0x6e31de34 // fmul v20.4s, v17.4s, v17.4s |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6e32de55 // fmul v21.4s, v18.4s, v18.4s |
| .long 0x4eb4fc21 // frsqrts v1.4s, v1.4s, v20.4s |
| .long 0x6ea1d810 // frsqrte v16.4s, v0.4s |
| .long 0x4eb5fc42 // frsqrts v2.4s, v2.4s, v21.4s |
| .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s |
| .long 0x4ea1d811 // frecpe v17.4s, v0.4s |
| .long 0x6e30de16 // fmul v22.4s, v16.4s, v16.4s |
| .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s |
| .long 0x6ea1d832 // frsqrte v18.4s, v1.4s |
| .long 0x4eb6fc16 // frsqrts v22.4s, v0.4s, v22.4s |
| .long 0x4e31fc00 // frecps v0.4s, v0.4s, v17.4s |
| .long 0x4ea1d833 // frecpe v19.4s, v1.4s |
| .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s |
| .long 0x6e32de51 // fmul v17.4s, v18.4s, v18.4s |
| .long 0x6ea1d854 // frsqrte v20.4s, v2.4s |
| .long 0x4eb1fc31 // frsqrts v17.4s, v1.4s, v17.4s |
| .long 0x4e33fc21 // frecps v1.4s, v1.4s, v19.4s |
| .long 0x6e21de61 // fmul v1.4s, v19.4s, v1.4s |
| .long 0x6e34de93 // fmul v19.4s, v20.4s, v20.4s |
| .long 0x4eb3fc53 // frsqrts v19.4s, v2.4s, v19.4s |
| .long 0x6e36de10 // fmul v16.4s, v16.4s, v22.4s |
| .long 0x6e31de51 // fmul v17.4s, v18.4s, v17.4s |
| .long 0x6e33de92 // fmul v18.4s, v20.4s, v19.4s |
| .long 0x6ea1da13 // frsqrte v19.4s, v16.4s |
| .long 0x4ea1d855 // frecpe v21.4s, v2.4s |
| .long 0x6e33de76 // fmul v22.4s, v19.4s, v19.4s |
| .long 0x4e35fc42 // frecps v2.4s, v2.4s, v21.4s |
| .long 0x6ea1da34 // frsqrte v20.4s, v17.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6e22dea2 // fmul v2.4s, v21.4s, v2.4s |
| .long 0x6ea1da55 // frsqrte v21.4s, v18.4s |
| .long 0x6e34de96 // fmul v22.4s, v20.4s, v20.4s |
| .long 0x6e30de70 // fmul v16.4s, v19.4s, v16.4s |
| .long 0x4eb6fe31 // frsqrts v17.4s, v17.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x6ea1da13 // frsqrte v19.4s, v16.4s |
| .long 0x4eb6fe52 // frsqrts v18.4s, v18.4s, v22.4s |
| .long 0x6e31de91 // fmul v17.4s, v20.4s, v17.4s |
| .long 0x6e33de76 // fmul v22.4s, v19.4s, v19.4s |
| .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| .long 0x6ea1da34 // frsqrte v20.4s, v17.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6ea1da55 // frsqrte v21.4s, v18.4s |
| .long 0x6e34de96 // fmul v22.4s, v20.4s, v20.4s |
| .long 0x6e30de70 // fmul v16.4s, v19.4s, v16.4s |
| .long 0x4eb6fe31 // frsqrts v17.4s, v17.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x6ea1da13 // frsqrte v19.4s, v16.4s |
| .long 0x4eb6fe52 // frsqrts v18.4s, v18.4s, v22.4s |
| .long 0x6e31de91 // fmul v17.4s, v20.4s, v17.4s |
| .long 0x6e33de76 // fmul v22.4s, v19.4s, v19.4s |
| .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| .long 0x6ea1da34 // frsqrte v20.4s, v17.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6ea1da55 // frsqrte v21.4s, v18.4s |
| .long 0x6e34de96 // fmul v22.4s, v20.4s, v20.4s |
| .long 0x6e30de70 // fmul v16.4s, v19.4s, v16.4s |
| .long 0x4eb6fe31 // frsqrts v17.4s, v17.4s, v22.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x6ea1da13 // frsqrte v19.4s, v16.4s |
| .long 0x4eb6fe52 // frsqrts v18.4s, v18.4s, v22.4s |
| .long 0x6e31de91 // fmul v17.4s, v20.4s, v17.4s |
| .long 0x6e33de76 // fmul v22.4s, v19.4s, v19.4s |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6ea1da34 // frsqrte v20.4s, v17.4s |
| .long 0x4eb6fe10 // frsqrts v16.4s, v16.4s, v22.4s |
| .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| .long 0x6e34de96 // fmul v22.4s, v20.4s, v20.4s |
| .long 0x6e30de70 // fmul v16.4s, v19.4s, v16.4s |
| .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s |
| .long 0x6ea1da55 // frsqrte v21.4s, v18.4s |
| .long 0x4eb6fe31 // frsqrts v17.4s, v17.4s, v22.4s |
| .long 0x4ea1da13 // frecpe v19.4s, v16.4s |
| .long 0x6e35deb6 // fmul v22.4s, v21.4s, v21.4s |
| .long 0x6e31de91 // fmul v17.4s, v20.4s, v17.4s |
| .long 0x4e33fe10 // frecps v16.4s, v16.4s, v19.4s |
| .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s |
| .long 0x4eb6fe52 // frsqrts v18.4s, v18.4s, v22.4s |
| .long 0x6e30de70 // fmul v16.4s, v19.4s, v16.4s |
| .long 0x4ea1da33 // frecpe v19.4s, v17.4s |
| .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| .long 0x4e33fe31 // frecps v17.4s, v17.4s, v19.4s |
| .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s |
| .long 0x4ea1da53 // frecpe v19.4s, v18.4s |
| .long 0x4e33fe52 // frecps v18.4s, v18.4s, v19.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6e32de72 // fmul v18.4s, v19.4s, v18.4s |
| .long 0x6f00e413 // movi v19.2d, #0x0 |
| .long 0x6e30dc00 // fmul v0.4s, v0.4s, v16.4s |
| .long 0x6e31dc21 // fmul v1.4s, v1.4s, v17.4s |
| .long 0x6e32dc42 // fmul v2.4s, v2.4s, v18.4s |
| .long 0x4e33f400 // fmax v0.4s, v0.4s, v19.4s |
| .long 0x4e33f421 // fmax v1.4s, v1.4s, v19.4s |
| .long 0x4e33f442 // fmax v2.4s, v2.4s, v19.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_rgb_to_hsl_aarch64 |
| .globl _sk_rgb_to_hsl_aarch64 |
| FUNCTION(_sk_rgb_to_hsl_aarch64) |
| _sk_rgb_to_hsl_aarch64: |
| .long 0x4e21f410 // fmax v16.4s, v0.4s, v1.4s |
| .long 0x4ea1f411 // fmin v17.4s, v0.4s, v1.4s |
| .long 0x6ea1e454 // fcmgt v20.4s, v2.4s, v1.4s |
| .long 0x4f00f715 // fmov v21.4s, #6.000000000000000000e+00 |
| .long 0x4e22f610 // fmax v16.4s, v16.4s, v2.4s |
| .long 0x4ea2f631 // fmin v17.4s, v17.4s, v2.4s |
| .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00 |
| .long 0x4e341eb4 // and v20.16b, v21.16b, v20.16b |
| .long 0x4eb1d615 // fsub v21.4s, v16.4s, v17.4s |
| .long 0x4ea2d433 // fsub v19.4s, v1.4s, v2.4s |
| .long 0x4ea0d456 // fsub v22.4s, v2.4s, v0.4s |
| .long 0x4f026417 // movi v23.4s, #0x40, lsl #24 |
| .long 0x6e35fe42 // fdiv v2.4s, v18.4s, v21.4s |
| .long 0x4ea1d418 // fsub v24.4s, v0.4s, v1.4s |
| .long 0x4f00f619 // fmov v25.4s, #4.000000000000000000e+00 |
| .long 0x4f0167fa // movi v26.4s, #0x3f, lsl #24 |
| .long 0x4eb0d6f2 // fsub v18.4s, v23.4s, v16.4s |
| .long 0x4e36cc57 // fmla v23.4s, v2.4s, v22.4s |
| .long 0x4e31e616 // fcmeq v22.4s, v16.4s, v17.4s |
| .long 0x4e20e600 // fcmeq v0.4s, v16.4s, v0.4s |
| .long 0x4e21e601 // fcmeq v1.4s, v16.4s, v1.4s |
| .long 0x4e31d610 // fadd v16.4s, v16.4s, v17.4s |
| .long 0x52a7c548 // mov w8, #0x3e2a0000 |
| .long 0x4e33cc54 // fmla v20.4s, v2.4s, v19.4s |
| .long 0x4e38cc59 // fmla v25.4s, v2.4s, v24.4s |
| .long 0x6e3ade02 // fmul v2.4s, v16.4s, v26.4s |
| .long 0x72955568 // movk w8, #0xaaab |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4eb1d651 // fsub v17.4s, v18.4s, v17.4s |
| .long 0x6ebae452 // fcmgt v18.4s, v2.4s, v26.4s |
| .long 0x6e791ee1 // bsl v1.16b, v23.16b, v25.16b |
| .long 0x4e040d13 // dup v19.4s, w8 |
| .long 0x6e701e32 // bsl v18.16b, v17.16b, v16.16b |
| .long 0x6e611e80 // bsl v0.16b, v20.16b, v1.16b |
| .long 0x6e32fea1 // fdiv v1.4s, v21.4s, v18.4s |
| .long 0x6e33dc00 // fmul v0.4s, v0.4s, v19.4s |
| .long 0x4e761c00 // bic v0.16b, v0.16b, v22.16b |
| .long 0x4e761c21 // bic v1.16b, v1.16b, v22.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_hsl_to_rgb_aarch64 |
| .globl _sk_hsl_to_rgb_aarch64 |
| FUNCTION(_sk_hsl_to_rgb_aarch64) |
| _sk_hsl_to_rgb_aarch64: |
| .long 0x52a7d548 // mov w8, #0x3eaa0000 |
| .long 0x72955568 // movk w8, #0xaaab |
| .long 0x4e040d17 // dup v23.4s, w8 |
| .long 0x52a7c548 // mov w8, #0x3e2a0000 |
| .long 0x72955568 // movk w8, #0xaaab |
| .long 0x4e040d13 // dup v19.4s, w8 |
| .long 0x52a7e548 // mov w8, #0x3f2a0000 |
| .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00 |
| .long 0x4f07f616 // fmov v22.4s, #-1.000000000000000000e+00 |
| .long 0x72955568 // movk w8, #0xaaab |
| .long 0x4e22d435 // fadd v21.4s, v1.4s, v2.4s |
| .long 0x4e040d1a // dup v26.4s, w8 |
| .long 0x52b7d548 // mov w8, #0xbeaa0000 |
| .long 0x6eb2e41d // fcmgt v29.4s, v0.4s, v18.4s |
| .long 0x4e36d41e // fadd v30.4s, v0.4s, v22.4s |
| .long 0x4f0167f1 // movi v17.4s, #0x3f, lsl #24 |
| .long 0x4ea0d830 // fcmeq v16.4s, v1.4s, #0.0 |
| .long 0x4ea0e819 // fcmlt v25.4s, v0.4s, #0.0 |
| .long 0x72955568 // movk w8, #0xaaab |
| .long 0x4e32d43c // fadd v28.4s, v1.4s, v18.4s |
| .long 0x4ea2cc35 // fmls v21.4s, v1.4s, v2.4s |
| .long 0x4e32d401 // fadd v1.4s, v0.4s, v18.4s |
| .long 0x6e601fdd // bsl v29.16b, v30.16b, v0.16b |
| .long 0x4e37d417 // fadd v23.4s, v0.4s, v23.4s |
| .long 0x6ea2e63b // fcmgt v27.4s, v17.4s, v2.4s |
| .long 0x4e040d1e // dup v30.4s, w8 |
| .long 0x6e22df9c // fmul v28.4s, v28.4s, v2.4s |
| .long 0x6e7d1c39 // bsl v25.16b, v1.16b, v29.16b |
| .long 0x6eb2e6e1 // fcmgt v1.4s, v23.4s, v18.4s |
| .long 0x4e36d6fd // fadd v29.4s, v23.4s, v22.4s |
| .long 0x4e3ed41e // fadd v30.4s, v0.4s, v30.4s |
| .long 0x6e751f9b // bsl v27.16b, v28.16b, v21.16b |
| .long 0x4ea0eaf5 // fcmlt v21.4s, v23.4s, #0.0 |
| .long 0x4e32d6fc // fadd v28.4s, v23.4s, v18.4s |
| .long 0x6e771fa1 // bsl v1.16b, v29.16b, v23.16b |
| .long 0x4f026414 // movi v20.4s, #0x40, lsl #24 |
| .long 0x6e611f95 // bsl v21.16b, v28.16b, v1.16b |
| .long 0x4e32d7c1 // fadd v1.4s, v30.4s, v18.4s |
| .long 0x6eb2e7d2 // fcmgt v18.4s, v30.4s, v18.4s |
| .long 0x4e36d7d6 // fadd v22.4s, v30.4s, v22.4s |
| .long 0x6ea0fb7c // fneg v28.4s, v27.4s |
| .long 0x4ea0ebdd // fcmlt v29.4s, v30.4s, #0.0 |
| .long 0x6e7e1ed2 // bsl v18.16b, v22.16b, v30.16b |
| .long 0x4e22ce9c // fmla v28.4s, v20.4s, v2.4s |
| .long 0x4f00f718 // fmov v24.4s, #6.000000000000000000e+00 |
| .long 0x6e721c3d // bsl v29.16b, v1.16b, v18.16b |
| .long 0x4ebcd761 // fsub v1.4s, v27.4s, v28.4s |
| .long 0x4eb5d752 // fsub v18.4s, v26.4s, v21.4s |
| .long 0x4ebc1f94 // mov v20.16b, v28.16b |
| .long 0x6e38dc38 // fmul v24.4s, v1.4s, v24.4s |
| .long 0x4eb9d756 // fsub v22.4s, v26.4s, v25.4s |
| .long 0x4ebc1f9f // mov v31.16b, v28.16b |
| .long 0x4e32cf14 // fmla v20.4s, v24.4s, v18.4s |
| .long 0x4ebc1f81 // mov v1.16b, v28.16b |
| .long 0x4ebc1f92 // mov v18.16b, v28.16b |
| .long 0x4e38cc1f // fmla v31.4s, v0.4s, v24.4s |
| .long 0x4e36cf01 // fmla v1.4s, v24.4s, v22.4s |
| .long 0x4ebdd740 // fsub v0.4s, v26.4s, v29.4s |
| .long 0x4e3ecf12 // fmla v18.4s, v24.4s, v30.4s |
| .long 0x4ebc1f96 // mov v22.16b, v28.16b |
| .long 0x6eb5e75e // fcmgt v30.4s, v26.4s, v21.4s |
| .long 0x4e20cf16 // fmla v22.4s, v24.4s, v0.4s |
| .long 0x6e7c1e9e // bsl v30.16b, v20.16b, v28.16b |
| .long 0x6eb9e754 // fcmgt v20.4s, v26.4s, v25.4s |
| .long 0x6ebde75a // fcmgt v26.4s, v26.4s, v29.4s |
| .long 0x6e7c1c34 // bsl v20.16b, v1.16b, v28.16b |
| .long 0x6e7c1eda // bsl v26.16b, v22.16b, v28.16b |
| .long 0x4e37cf1c // fmla v28.4s, v24.4s, v23.4s |
| .long 0x6eb9e637 // fcmgt v23.4s, v17.4s, v25.4s |
| .long 0x6eb5e678 // fcmgt v24.4s, v19.4s, v21.4s |
| .long 0x6eb5e635 // fcmgt v21.4s, v17.4s, v21.4s |
| .long 0x6ebde631 // fcmgt v17.4s, v17.4s, v29.4s |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x6eb9e676 // fcmgt v22.4s, v19.4s, v25.4s |
| .long 0x6ebde673 // fcmgt v19.4s, v19.4s, v29.4s |
| .long 0x6e7a1f71 // bsl v17.16b, v27.16b, v26.16b |
| .long 0x6e7e1f75 // bsl v21.16b, v27.16b, v30.16b |
| .long 0x6e741f77 // bsl v23.16b, v27.16b, v20.16b |
| .long 0x6e711e53 // bsl v19.16b, v18.16b, v17.16b |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb01e01 // mov v1.16b, v16.16b |
| .long 0x6e751f98 // bsl v24.16b, v28.16b, v21.16b |
| .long 0x6e771ff6 // bsl v22.16b, v31.16b, v23.16b |
| .long 0x6e731c50 // bsl v16.16b, v2.16b, v19.16b |
| .long 0x6e781c40 // bsl v0.16b, v2.16b, v24.16b |
| .long 0x6e761c41 // bsl v1.16b, v2.16b, v22.16b |
| .long 0x4eb01e02 // mov v2.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_scale_1_float_aarch64 |
| .globl _sk_scale_1_float_aarch64 |
| FUNCTION(_sk_scale_1_float_aarch64) |
| _sk_scale_1_float_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xbd400110 // ldr s16, [x8] |
| .long 0x4f909000 // fmul v0.4s, v0.4s, v16.s[0] |
| .long 0x4f909021 // fmul v1.4s, v1.4s, v16.s[0] |
| .long 0x4f909042 // fmul v2.4s, v2.4s, v16.s[0] |
| .long 0x4f909063 // fmul v3.4s, v3.4s, v16.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_scale_u8_aarch64 |
| .globl _sk_scale_u8_aarch64 |
| FUNCTION(_sk_scale_u8_aarch64) |
| _sk_scale_u8_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x8b000108 // add x8, x8, x0 |
| .long 0x39400109 // ldrb w9, [x8] |
| .long 0x3940050a // ldrb w10, [x8, #1] |
| .long 0x3940090b // ldrb w11, [x8, #2] |
| .long 0x39400d08 // ldrb w8, [x8, #3] |
| .long 0x4e021d31 // mov v17.h[0], w9 |
| .long 0x4e061d51 // mov v17.h[1], w10 |
| .long 0x4e0a1d71 // mov v17.h[2], w11 |
| .long 0x4e0e1d11 // mov v17.h[3], w8 |
| .long 0x2f10a631 // uxtl v17.4s, v17.4h |
| .long 0x6e21da31 // ucvtf v17.4s, v17.4s |
| .long 0x6e30de30 // fmul v16.4s, v17.4s, v16.4s |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_lerp_1_float_aarch64 |
| .globl _sk_lerp_1_float_aarch64 |
| FUNCTION(_sk_lerp_1_float_aarch64) |
| _sk_lerp_1_float_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x4ea4d411 // fsub v17.4s, v0.4s, v4.4s |
| .long 0x4ea41c80 // mov v0.16b, v4.16b |
| .long 0x4ea5d432 // fsub v18.4s, v1.4s, v5.4s |
| .long 0xbd400110 // ldr s16, [x8] |
| .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| .long 0x4f901220 // fmla v0.4s, v17.4s, v16.s[0] |
| .long 0x4ea6d451 // fsub v17.4s, v2.4s, v6.4s |
| .long 0x4f901241 // fmla v1.4s, v18.4s, v16.s[0] |
| .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| .long 0x4ea7d472 // fsub v18.4s, v3.4s, v7.4s |
| .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| .long 0x4f901222 // fmla v2.4s, v17.4s, v16.s[0] |
| .long 0x4f901243 // fmla v3.4s, v18.4s, v16.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_lerp_u8_aarch64 |
| .globl _sk_lerp_u8_aarch64 |
| FUNCTION(_sk_lerp_u8_aarch64) |
| _sk_lerp_u8_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x4ea4d412 // fsub v18.4s, v0.4s, v4.4s |
| .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 0x4e021d51 // mov v17.h[0], w10 |
| .long 0x4e061d31 // mov v17.h[1], w9 |
| .long 0x4e0a1d71 // mov v17.h[2], w11 |
| .long 0x4e0e1d11 // mov v17.h[3], w8 |
| .long 0x2f10a620 // uxtl v0.4s, v17.4h |
| .long 0x6e21d800 // ucvtf v0.4s, v0.4s |
| .long 0x6e30dc10 // fmul v16.4s, v0.4s, v16.4s |
| .long 0x4ea41c80 // mov v0.16b, v4.16b |
| .long 0x4ea5d431 // fsub v17.4s, v1.4s, v5.4s |
| .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| .long 0x4e32ce00 // fmla v0.4s, v16.4s, v18.4s |
| .long 0x4ea6d452 // fsub v18.4s, v2.4s, v6.4s |
| .long 0x4e31ce01 // fmla v1.4s, v16.4s, v17.4s |
| .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| .long 0x4ea7d471 // fsub v17.4s, v3.4s, v7.4s |
| .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| .long 0x4e32ce02 // fmla v2.4s, v16.4s, v18.4s |
| .long 0x4e31ce03 // fmla v3.4s, v16.4s, v17.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_lerp_565_aarch64 |
| .globl _sk_lerp_565_aarch64 |
| FUNCTION(_sk_lerp_565_aarch64) |
| _sk_lerp_565_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xd37ff809 // lsl x9, x0, #1 |
| .long 0x4f072710 // movi v16.4s, #0xf8, lsl #8 |
| .long 0x4ea4d413 // fsub v19.4s, v0.4s, v4.4s |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0xfc696903 // ldr d3, [x8, x9] |
| .long 0x52a6f088 // mov w8, #0x37840000 |
| .long 0x72842108 // movk w8, #0x2108 |
| .long 0x4e040d11 // dup v17.4s, w8 |
| .long 0x2f10a463 // uxtl v3.4s, v3.4h |
| .long 0x321b17e8 // orr w8, wzr, #0x7e0 |
| .long 0x4e301c60 // and v0.16b, v3.16b, v16.16b |
| .long 0x4e040d12 // dup v18.4s, w8 |
| .long 0x52a74048 // mov w8, #0x3a020000 |
| .long 0x4e21d800 // scvtf v0.4s, v0.4s |
| .long 0x72810428 // movk w8, #0x821 |
| .long 0x6e31dc10 // fmul v16.4s, v0.4s, v17.4s |
| .long 0x4ea41c80 // mov v0.16b, v4.16b |
| .long 0x4e33ce00 // fmla v0.4s, v16.4s, v19.4s |
| .long 0x4f0007f0 // movi v16.4s, #0x1f |
| .long 0x4e040d11 // dup v17.4s, w8 |
| .long 0x52a7a088 // mov w8, #0x3d040000 |
| .long 0x4e321c72 // and v18.16b, v3.16b, v18.16b |
| .long 0x72842108 // movk w8, #0x2108 |
| .long 0x4e301c63 // and v3.16b, v3.16b, v16.16b |
| .long 0x4ea6d450 // fsub v16.4s, v2.4s, v6.4s |
| .long 0x4e21da42 // scvtf v2.4s, v18.4s |
| .long 0x6e31dc51 // fmul v17.4s, v2.4s, v17.4s |
| .long 0x4e040d02 // dup v2.4s, w8 |
| .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| .long 0x4ea5d433 // fsub v19.4s, v1.4s, v5.4s |
| .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| .long 0x6e22dc63 // fmul v3.4s, v3.4s, v2.4s |
| .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| .long 0x4e33ce21 // fmla v1.4s, v17.4s, v19.4s |
| .long 0x4e30cc62 // fmla v2.4s, v3.4s, v16.4s |
| .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_tables_aarch64 |
| .globl _sk_load_tables_aarch64 |
| FUNCTION(_sk_load_tables_aarch64) |
| _sk_load_tables_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xd37ef409 // lsl x9, x0, #2 |
| .long 0x6f00e620 // movi v0.2d, #0xff000000ff |
| .long 0x52a7700b // mov w11, #0x3b800000 |
| .long 0xa940310a // ldp x10, x12, [x8] |
| .long 0x7290102b // movk w11, #0x8081 |
| .long 0x4e040d63 // dup v3.4s, w11 |
| .long 0x3ce96942 // ldr q2, [x10, x9] |
| .long 0xa9412109 // ldp x9, x8, [x8, #16] |
| .long 0x4e201c41 // and v1.16b, v2.16b, v0.16b |
| .long 0x1e26002e // fmov w14, s1 |
| .long 0x6f380450 // ushr v16.4s, v2.4s, #8 |
| .long 0x6f300451 // ushr v17.4s, v2.4s, #16 |
| .long 0x8b2e498e // add x14, x12, w14, uxtw #2 |
| .long 0x0e0c3c2a // mov w10, v1.s[1] |
| .long 0x0e143c2b // mov w11, v1.s[2] |
| .long 0x0e1c3c2d // mov w13, v1.s[3] |
| .long 0x4e201e01 // and v1.16b, v16.16b, v0.16b |
| .long 0x4e201e30 // and v16.16b, v17.16b, v0.16b |
| .long 0x0d4081c0 // ld1 {v0.s}[0], [x14] |
| .long 0x8b2a498a // add x10, x12, w10, uxtw #2 |
| .long 0xbc6b5991 // ldr s17, [x12, w11, uxtw #2] |
| .long 0xbc6d5992 // ldr s18, [x12, w13, uxtw #2] |
| .long 0x0e0c3c2b // mov w11, v1.s[1] |
| .long 0x0e143c2c // mov w12, v1.s[2] |
| .long 0x0e1c3c2d // mov w13, v1.s[3] |
| .long 0x1e26002e // fmov w14, s1 |
| .long 0x8b2e492e // add x14, x9, w14, uxtw #2 |
| .long 0xbc6c5933 // ldr s19, [x9, w12, uxtw #2] |
| .long 0xbc6d5934 // ldr s20, [x9, w13, uxtw #2] |
| .long 0x8b2b4929 // add x9, x9, w11, uxtw #2 |
| .long 0x1e26020b // fmov w11, s16 |
| .long 0x6f280442 // ushr v2.4s, v2.4s, #24 |
| .long 0x0d409140 // ld1 {v0.s}[1], [x10] |
| .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| .long 0x8b2b490a // add x10, x8, w11, uxtw #2 |
| .long 0x0d4081c1 // ld1 {v1.s}[0], [x14] |
| .long 0x6e23dc43 // fmul v3.4s, v2.4s, v3.4s |
| .long 0x0d408142 // ld1 {v2.s}[0], [x10] |
| .long 0x0e0c3e0f // mov w15, v16.s[1] |
| .long 0x0e143e0c // mov w12, v16.s[2] |
| .long 0x8b2f490a // add x10, x8, w15, uxtw #2 |
| .long 0x0e1c3e0d // mov w13, v16.s[3] |
| .long 0xbc6c5910 // ldr s16, [x8, w12, uxtw #2] |
| .long 0x0d409121 // ld1 {v1.s}[1], [x9] |
| .long 0x0d409142 // ld1 {v2.s}[1], [x10] |
| .long 0x6e140620 // mov v0.s[2], v17.s[0] |
| .long 0xbc6d5911 // ldr s17, [x8, w13, uxtw #2] |
| .long 0x6e140661 // mov v1.s[2], v19.s[0] |
| .long 0x6e140602 // mov v2.s[2], v16.s[0] |
| .long 0x6e1c0640 // mov v0.s[3], v18.s[0] |
| .long 0x6e1c0681 // mov v1.s[3], v20.s[0] |
| .long 0x6e1c0622 // mov v2.s[3], v17.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_tables_u16_be_aarch64 |
| .globl _sk_load_tables_u16_be_aarch64 |
| FUNCTION(_sk_load_tables_u16_be_aarch64) |
| _sk_load_tables_u16_be_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x52a6f00a // mov w10, #0x37800000 |
| .long 0x7280100a // movk w10, #0x80 |
| .long 0x4e040d50 // dup v16.4s, w10 |
| .long 0xa9402d09 // ldp x9, x11, [x8] |
| .long 0x8b000d29 // add x9, x9, x0, lsl #3 |
| .long 0x0c400520 // ld4 {v0.4h-v3.4h}, [x9] |
| .long 0xa9412109 // ldp x9, x8, [x8, #16] |
| .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8 |
| .long 0x2f10a411 // uxtl v17.4s, v0.4h |
| .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8 |
| .long 0x1e26022e // fmov w14, s17 |
| .long 0x0e0c3e2a // mov w10, v17.s[1] |
| .long 0x0e143e2c // mov w12, v17.s[2] |
| .long 0x0e1c3e2d // mov w13, v17.s[3] |
| .long 0x8b2e496e // add x14, x11, w14, uxtw #2 |
| .long 0x2f10a431 // uxtl v17.4s, v1.4h |
| .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8 |
| .long 0x2f10a453 // uxtl v19.4s, v2.4h |
| .long 0x2f180462 // ushr v2.4h, v3.4h, #8 |
| .long 0x0d4081c0 // ld1 {v0.s}[0], [x14] |
| .long 0x0f185474 // shl v20.4h, v3.4h, #8 |
| .long 0x0ea21e82 // orr v2.8b, v20.8b, v2.8b |
| .long 0x8b2a496a // add x10, x11, w10, uxtw #2 |
| .long 0x1e26022e // fmov w14, s17 |
| .long 0x1e26026f // fmov w15, s19 |
| .long 0x2f10a442 // uxtl v2.4s, v2.4h |
| .long 0x8b2e492e // add x14, x9, w14, uxtw #2 |
| .long 0x0d409140 // ld1 {v0.s}[1], [x10] |
| .long 0x8b2f490a // add x10, x8, w15, uxtw #2 |
| .long 0x6e21d842 // ucvtf v2.4s, v2.4s |
| .long 0x0d4081c1 // ld1 {v1.s}[0], [x14] |
| .long 0x6e30dc43 // fmul v3.4s, v2.4s, v16.4s |
| .long 0x0d408142 // ld1 {v2.s}[0], [x10] |
| .long 0xbc6c5972 // ldr s18, [x11, w12, uxtw #2] |
| .long 0xbc6d5975 // ldr s21, [x11, w13, uxtw #2] |
| .long 0x0e0c3e2b // mov w11, v17.s[1] |
| .long 0x0e143e2c // mov w12, v17.s[2] |
| .long 0x0e1c3e2d // mov w13, v17.s[3] |
| .long 0x8b2b492b // add x11, x9, w11, uxtw #2 |
| .long 0xbc6c5931 // ldr s17, [x9, w12, uxtw #2] |
| .long 0xbc6d5936 // ldr s22, [x9, w13, uxtw #2] |
| .long 0x0e0c3e69 // mov w9, v19.s[1] |
| .long 0x0e143e6c // mov w12, v19.s[2] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0x0e1c3e6d // mov w13, v19.s[3] |
| .long 0xbc6c5913 // ldr s19, [x8, w12, uxtw #2] |
| .long 0x0d409161 // ld1 {v1.s}[1], [x11] |
| .long 0x0d409122 // ld1 {v2.s}[1], [x9] |
| .long 0xbc6d5910 // ldr s16, [x8, w13, uxtw #2] |
| .long 0x6e140640 // mov v0.s[2], v18.s[0] |
| .long 0x6e140621 // mov v1.s[2], v17.s[0] |
| .long 0x6e140662 // mov v2.s[2], v19.s[0] |
| .long 0x6e1c06a0 // mov v0.s[3], v21.s[0] |
| .long 0x6e1c06c1 // mov v1.s[3], v22.s[0] |
| .long 0x6e1c0602 // mov v2.s[3], v16.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_tables_rgb_u16_be_aarch64 |
| .globl _sk_load_tables_rgb_u16_be_aarch64 |
| FUNCTION(_sk_load_tables_rgb_u16_be_aarch64) |
| _sk_load_tables_rgb_u16_be_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x321f07ea // orr w10, wzr, #0x6 |
| .long 0xa9402d09 // ldp x9, x11, [x8] |
| .long 0x9b0a2409 // madd x9, x0, x10, x9 |
| .long 0x0c404521 // ld3 {v1.4h-v3.4h}, [x9] |
| .long 0xa9412109 // ldp x9, x8, [x8, #16] |
| .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8 |
| .long 0x2f10a420 // uxtl v0.4s, v1.4h |
| .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8 |
| .long 0x1e26000e // fmov w14, s0 |
| .long 0x0e143c0c // mov w12, v0.s[2] |
| .long 0x8b2e496e // add x14, x11, w14, uxtw #2 |
| .long 0x2f10a450 // uxtl v16.4s, v2.4h |
| .long 0x0e0c3c0a // mov w10, v0.s[1] |
| .long 0x0e1c3c0d // mov w13, v0.s[3] |
| .long 0xbc6c5971 // ldr s17, [x11, w12, uxtw #2] |
| .long 0x0d4081c0 // ld1 {v0.s}[0], [x14] |
| .long 0x0e143e0c // mov w12, v16.s[2] |
| .long 0x1e26020e // fmov w14, s16 |
| .long 0x2f07b7e3 // bic v3.4h, #0xff, lsl #8 |
| .long 0x8b2a496a // add x10, x11, w10, uxtw #2 |
| .long 0xbc6d5972 // ldr s18, [x11, w13, uxtw #2] |
| .long 0x0e0c3e0b // mov w11, v16.s[1] |
| .long 0x0e1c3e0d // mov w13, v16.s[3] |
| .long 0xbc6c5930 // ldr s16, [x9, w12, uxtw #2] |
| .long 0x8b2e492c // add x12, x9, w14, uxtw #2 |
| .long 0x2f10a462 // uxtl v2.4s, v3.4h |
| .long 0xbc6d5923 // ldr s3, [x9, w13, uxtw #2] |
| .long 0x0d408181 // ld1 {v1.s}[0], [x12] |
| .long 0x0e143c4c // mov w12, v2.s[2] |
| .long 0x1e26004d // fmov w13, s2 |
| .long 0xbc6c5913 // ldr s19, [x8, w12, uxtw #2] |
| .long 0x8b2d490c // add x12, x8, w13, uxtw #2 |
| .long 0x8b2b492b // add x11, x9, w11, uxtw #2 |
| .long 0x0e0c3c49 // mov w9, v2.s[1] |
| .long 0x0d409140 // ld1 {v0.s}[1], [x10] |
| .long 0x0e1c3c4a // mov w10, v2.s[3] |
| .long 0x0d408182 // ld1 {v2.s}[0], [x12] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0x0d409161 // ld1 {v1.s}[1], [x11] |
| .long 0x6e140620 // mov v0.s[2], v17.s[0] |
| .long 0x0d409122 // ld1 {v2.s}[1], [x9] |
| .long 0xbc6a5911 // ldr s17, [x8, w10, uxtw #2] |
| .long 0x6e140601 // mov v1.s[2], v16.s[0] |
| .long 0x6e1c0640 // mov v0.s[3], v18.s[0] |
| .long 0x6e140662 // mov v2.s[2], v19.s[0] |
| .long 0x6e1c0461 // mov v1.s[3], v3.s[0] |
| .long 0x6e1c0622 // mov v2.s[3], v17.s[0] |
| .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_byte_tables_aarch64 |
| .globl _sk_byte_tables_aarch64 |
| FUNCTION(_sk_byte_tables_aarch64) |
| _sk_byte_tables_aarch64: |
| .long 0xd10083ff // sub sp, sp, #0x20 |
| .long 0xaa0103e8 // mov x8, x1 |
| .long 0x91002109 // add x9, x8, #0x8 |
| .long 0xa9014ff4 // stp x20, x19, [sp, #16] |
| .long 0xf90007e9 // str x9, [sp, #8] |
| .long 0xf8410429 // ldr x9, [x1], #16 |
| .long 0x52a86fea // mov w10, #0x437f0000 |
| .long 0x4e040d51 // dup v17.4s, w10 |
| .long 0x52a7700b // mov w11, #0x3b800000 |
| .long 0xa9405126 // ldp x6, x20, [x9] |
| .long 0x6e31dc00 // fmul v0.4s, v0.4s, v17.4s |
| .long 0x7290102b // movk w11, #0x8081 |
| .long 0x6e21a800 // fcvtnu v0.4s, v0.4s |
| .long 0x4e040d70 // dup v16.4s, w11 |
| .long 0x0e0c3c0a // mov w10, v0.s[1] |
| .long 0x0e143c0b // mov w11, v0.s[2] |
| .long 0x0e1c3c0c // mov w12, v0.s[3] |
| .long 0x1e26000d // fmov w13, s0 |
| .long 0x386d48cd // ldrb w13, [x6, w13, uxtw] |
| .long 0x386a48ca // ldrb w10, [x6, w10, uxtw] |
| .long 0x386b48cb // ldrb w11, [x6, w11, uxtw] |
| .long 0x386c48cc // ldrb w12, [x6, w12, uxtw] |
| .long 0xa9412526 // ldp x6, x9, [x9, #16] |
| .long 0x6e31dc42 // fmul v2.4s, v2.4s, v17.4s |
| .long 0x6e31dc21 // fmul v1.4s, v1.4s, v17.4s |
| .long 0x6e31dc63 // fmul v3.4s, v3.4s, v17.4s |
| .long 0x6e21a842 // fcvtnu v2.4s, v2.4s |
| .long 0x6e21a821 // fcvtnu v1.4s, v1.4s |
| .long 0x6e21a863 // fcvtnu v3.4s, v3.4s |
| .long 0x0e0c3c52 // mov w18, v2.s[1] |
| .long 0x0e143c43 // mov w3, v2.s[2] |
| .long 0x0e1c3c44 // mov w4, v2.s[3] |
| .long 0x1e260045 // fmov w5, s2 |
| .long 0x1e260031 // fmov w17, s1 |
| .long 0x386548c5 // ldrb w5, [x6, w5, uxtw] |
| .long 0x387248d2 // ldrb w18, [x6, w18, uxtw] |
| .long 0x386348c3 // ldrb w3, [x6, w3, uxtw] |
| .long 0x386448c4 // ldrb w4, [x6, w4, uxtw] |
| .long 0x1e260066 // fmov w6, s3 |
| .long 0x0e0c3c2e // mov w14, v1.s[1] |
| .long 0x0e0c3c67 // mov w7, v3.s[1] |
| .long 0x38714a91 // ldrb w17, [x20, w17, uxtw] |
| .long 0x38664926 // ldrb w6, [x9, w6, uxtw] |
| .long 0x0e143c2f // mov w15, v1.s[2] |
| .long 0x0e1c3c30 // mov w16, v1.s[3] |
| .long 0x0e143c73 // mov w19, v3.s[2] |
| .long 0x386e4a8e // ldrb w14, [x20, w14, uxtw] |
| .long 0x38674927 // ldrb w7, [x9, w7, uxtw] |
| .long 0x386f4a8f // ldrb w15, [x20, w15, uxtw] |
| .long 0x38704a90 // ldrb w16, [x20, w16, uxtw] |
| .long 0x0e1c3c74 // mov w20, v3.s[3] |
| .long 0x38734933 // ldrb w19, [x9, w19, uxtw] |
| .long 0x38744929 // ldrb w9, [x9, w20, uxtw] |
| .long 0x4e021da0 // mov v0.h[0], w13 |
| .long 0x4e021e21 // mov v1.h[0], w17 |
| .long 0x4e021ca2 // mov v2.h[0], w5 |
| .long 0x4e021cc3 // mov v3.h[0], w6 |
| .long 0x4e061d40 // mov v0.h[1], w10 |
| .long 0x4e061dc1 // mov v1.h[1], w14 |
| .long 0x4e061e42 // mov v2.h[1], w18 |
| .long 0x4e061ce3 // mov v3.h[1], w7 |
| .long 0x4e0a1d60 // mov v0.h[2], w11 |
| .long 0x4e0a1de1 // mov v1.h[2], w15 |
| .long 0x4e0a1c62 // mov v2.h[2], w3 |
| .long 0x4e0a1e63 // mov v3.h[2], w19 |
| .long 0x4e0e1d80 // mov v0.h[3], w12 |
| .long 0x4e0e1e01 // mov v1.h[3], w16 |
| .long 0x4e0e1c82 // mov v2.h[3], w4 |
| .long 0x4e0e1d23 // mov v3.h[3], w9 |
| .long 0xf9400505 // ldr x5, [x8, #8] |
| .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8 |
| .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8 |
| .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8 |
| .long 0x2f07b7e3 // bic v3.4h, #0xff, lsl #8 |
| .long 0xa9414ff4 // ldp x20, x19, [sp, #16] |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x2f10a421 // uxtl v1.4s, v1.4h |
| .long 0x2f10a442 // uxtl v2.4s, v2.4h |
| .long 0x2f10a463 // uxtl v3.4s, v3.4h |
| .long 0x6e21d800 // ucvtf v0.4s, v0.4s |
| .long 0x6e21d821 // ucvtf v1.4s, v1.4s |
| .long 0x6e21d842 // ucvtf v2.4s, v2.4s |
| .long 0x6e21d863 // ucvtf v3.4s, v3.4s |
| .long 0x6e30dc00 // fmul v0.4s, v0.4s, v16.4s |
| .long 0x6e30dc21 // fmul v1.4s, v1.4s, v16.4s |
| .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s |
| .long 0x6e30dc63 // fmul v3.4s, v3.4s, v16.4s |
| .long 0x910083ff // add sp, sp, #0x20 |
| .long 0xd61f00a0 // br x5 |
| |
| HIDDEN _sk_byte_tables_rgb_aarch64 |
| .globl _sk_byte_tables_rgb_aarch64 |
| FUNCTION(_sk_byte_tables_rgb_aarch64) |
| _sk_byte_tables_rgb_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0xb9401909 // ldr w9, [x8, #24] |
| .long 0xa9402d0a // ldp x10, x11, [x8] |
| .long 0xf9400908 // ldr x8, [x8, #16] |
| .long 0x51000529 // sub w9, w9, #0x1 |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x4e21da31 // scvtf v17.4s, v17.4s |
| .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s |
| .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s |
| .long 0x6e22de22 // fmul v2.4s, v17.4s, v2.4s |
| .long 0x6e21a821 // fcvtnu v1.4s, v1.4s |
| .long 0x6e21a800 // fcvtnu v0.4s, v0.4s |
| .long 0x6e21a842 // fcvtnu v2.4s, v2.4s |
| .long 0x0e0c3c2f // mov w15, v1.s[1] |
| .long 0x0e143c30 // mov w16, v1.s[2] |
| .long 0x0e1c3c31 // mov w17, v1.s[3] |
| .long 0x1e260032 // fmov w18, s1 |
| .long 0x1e26000e // fmov w14, s0 |
| .long 0x38724972 // ldrb w18, [x11, w18, uxtw] |
| .long 0x386f496f // ldrb w15, [x11, w15, uxtw] |
| .long 0x38704970 // ldrb w16, [x11, w16, uxtw] |
| .long 0x3871496b // ldrb w11, [x11, w17, uxtw] |
| .long 0x1e260051 // fmov w17, s2 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x386e494e // ldrb w14, [x10, w14, uxtw] |
| .long 0x0e0c3c44 // mov w4, v2.s[1] |
| .long 0x38714911 // ldrb w17, [x8, w17, uxtw] |
| .long 0x0e143c0c // mov w12, v0.s[2] |
| .long 0x0e1c3c0d // mov w13, v0.s[3] |
| .long 0x0e143c45 // mov w5, v2.s[2] |
| .long 0x38694949 // ldrb w9, [x10, w9, uxtw] |
| .long 0x38644904 // ldrb w4, [x8, w4, uxtw] |
| .long 0x386c494c // ldrb w12, [x10, w12, uxtw] |
| .long 0x386d494a // ldrb w10, [x10, w13, uxtw] |
| .long 0x0e1c3c4d // mov w13, v2.s[3] |
| .long 0x38654905 // ldrb w5, [x8, w5, uxtw] |
| .long 0x386d4908 // ldrb w8, [x8, w13, uxtw] |
| .long 0x4e021dc0 // mov v0.h[0], w14 |
| .long 0x4e021e41 // mov v1.h[0], w18 |
| .long 0x4e021e22 // mov v2.h[0], w17 |
| .long 0x4e061d20 // mov v0.h[1], w9 |
| .long 0x4e061de1 // mov v1.h[1], w15 |
| .long 0x4e061c82 // mov v2.h[1], w4 |
| .long 0x4e0a1d80 // mov v0.h[2], w12 |
| .long 0x4e0a1e01 // mov v1.h[2], w16 |
| .long 0x4e0a1ca2 // mov v2.h[2], w5 |
| .long 0x4e0e1d40 // mov v0.h[3], w10 |
| .long 0x4e0e1d61 // mov v1.h[3], w11 |
| .long 0x4e0e1d02 // mov v2.h[3], w8 |
| .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8 |
| .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8 |
| .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8 |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x2f10a421 // uxtl v1.4s, v1.4h |
| .long 0x2f10a442 // uxtl v2.4s, v2.4h |
| .long 0x6e21d800 // ucvtf v0.4s, v0.4s |
| .long 0x6e21d821 // ucvtf v1.4s, v1.4s |
| .long 0x6e21d842 // ucvtf v2.4s, v2.4s |
| .long 0x6e30dc00 // fmul v0.4s, v0.4s, v16.4s |
| .long 0x6e30dc21 // fmul v1.4s, v1.4s, v16.4s |
| .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_table_r_aarch64 |
| .globl _sk_table_r_aarch64 |
| FUNCTION(_sk_table_r_aarch64) |
| _sk_table_r_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xb9400909 // ldr w9, [x8, #8] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x51000529 // sub w9, w9, #0x1 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| .long 0x6e21a810 // fcvtnu v16.4s, v0.4s |
| .long 0x1e26020b // fmov w11, s16 |
| .long 0x8b2b490b // add x11, x8, w11, uxtw #2 |
| .long 0x0d408160 // ld1 {v0.s}[0], [x11] |
| .long 0x0e0c3e09 // mov w9, v16.s[1] |
| .long 0x0e143e0a // mov w10, v16.s[2] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0x0e1c3e0b // mov w11, v16.s[3] |
| .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2] |
| .long 0x0d409120 // ld1 {v0.s}[1], [x9] |
| .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2] |
| .long 0x6e140600 // mov v0.s[2], v16.s[0] |
| .long 0x6e1c0620 // mov v0.s[3], v17.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_table_g_aarch64 |
| .globl _sk_table_g_aarch64 |
| FUNCTION(_sk_table_g_aarch64) |
| _sk_table_g_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xb9400909 // ldr w9, [x8, #8] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x51000529 // sub w9, w9, #0x1 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| .long 0x6e21a830 // fcvtnu v16.4s, v1.4s |
| .long 0x1e26020b // fmov w11, s16 |
| .long 0x8b2b490b // add x11, x8, w11, uxtw #2 |
| .long 0x0d408161 // ld1 {v1.s}[0], [x11] |
| .long 0x0e0c3e09 // mov w9, v16.s[1] |
| .long 0x0e143e0a // mov w10, v16.s[2] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0x0e1c3e0b // mov w11, v16.s[3] |
| .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2] |
| .long 0x0d409121 // ld1 {v1.s}[1], [x9] |
| .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2] |
| .long 0x6e140601 // mov v1.s[2], v16.s[0] |
| .long 0x6e1c0621 // mov v1.s[3], v17.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_table_b_aarch64 |
| .globl _sk_table_b_aarch64 |
| FUNCTION(_sk_table_b_aarch64) |
| _sk_table_b_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xb9400909 // ldr w9, [x8, #8] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x51000529 // sub w9, w9, #0x1 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| .long 0x6e21a850 // fcvtnu v16.4s, v2.4s |
| .long 0x1e26020b // fmov w11, s16 |
| .long 0x8b2b490b // add x11, x8, w11, uxtw #2 |
| .long 0x0d408162 // ld1 {v2.s}[0], [x11] |
| .long 0x0e0c3e09 // mov w9, v16.s[1] |
| .long 0x0e143e0a // mov w10, v16.s[2] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0x0e1c3e0b // mov w11, v16.s[3] |
| .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2] |
| .long 0x0d409122 // ld1 {v2.s}[1], [x9] |
| .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2] |
| .long 0x6e140602 // mov v2.s[2], v16.s[0] |
| .long 0x6e1c0622 // mov v2.s[3], v17.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_table_a_aarch64 |
| .globl _sk_table_a_aarch64 |
| FUNCTION(_sk_table_a_aarch64) |
| _sk_table_a_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xb9400909 // ldr w9, [x8, #8] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x51000529 // sub w9, w9, #0x1 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s |
| .long 0x6e21a870 // fcvtnu v16.4s, v3.4s |
| .long 0x1e26020b // fmov w11, s16 |
| .long 0x8b2b490b // add x11, x8, w11, uxtw #2 |
| .long 0x0d408163 // ld1 {v3.s}[0], [x11] |
| .long 0x0e0c3e09 // mov w9, v16.s[1] |
| .long 0x0e143e0a // mov w10, v16.s[2] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0x0e1c3e0b // mov w11, v16.s[3] |
| .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2] |
| .long 0x0d409123 // ld1 {v3.s}[1], [x9] |
| .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2] |
| .long 0x6e140603 // mov v3.s[2], v16.s[0] |
| .long 0x6e1c0623 // mov v3.s[3], v17.s[0] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_parametric_r_aarch64 |
| .globl _sk_parametric_r_aarch64 |
| FUNCTION(_sk_parametric_r_aarch64) |
| _sk_parametric_r_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x9100610a // add x10, x8, #0x18 |
| .long 0x4d40c932 // ld1r {v18.4s}, [x9] |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0xbd400d11 // ldr s17, [x8, #12] |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 |
| .long 0x9100210a // add x10, x8, #0x8 |
| .long 0x4d40c954 // ld1r {v20.4s}, [x10] |
| .long 0x4f911010 // fmla v16.4s, v0.4s, v17.s[0] |
| .long 0xbd400135 // ldr s21, [x9] |
| .long 0x52b85f09 // mov w9, #0xc2f80000 |
| .long 0x728e6ee9 // movk w9, #0x7377 |
| .long 0x4e040d37 // dup v23.4s, w9 |
| .long 0x52a7f7e9 // mov w9, #0x3fbf0000 |
| .long 0x7297eea9 // movk w9, #0xbf75 |
| .long 0x4f951014 // fmla v20.4s, v0.4s, v21.s[0] |
| .long 0x6e20e640 // fcmge v0.4s, v18.4s, v0.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a7d689 // mov w9, #0x3eb40000 |
| .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 |
| .long 0x72889f29 // movk w9, #0x44f9 |
| .long 0x4e21da95 // scvtf v21.4s, v20.4s |
| .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7fb89 // mov w9, #0x3fdc0000 |
| .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s |
| .long 0x729d3469 // movk w9, #0xe9a3 |
| .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 |
| .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a85e49 // mov w9, #0x42f20000 |
| .long 0x72918a29 // movk w9, #0x8c51 |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7f7c9 // mov w9, #0x3fbe0000 |
| .long 0x729791a9 // movk w9, #0xbc8d |
| .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a81349 // mov w9, #0x409a0000 |
| .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s |
| .long 0x729ebf09 // movk w9, #0xf5f8 |
| .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x52a83ba9 // mov w9, #0x41dd0000 |
| .long 0x4e219a33 // frintm v19.4s, v17.4s |
| .long 0x729a5fc9 // movk w9, #0xd2fe |
| .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s |
| .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s |
| .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s |
| .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x91005108 // add x8, x8, #0x14 |
| .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s |
| .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s |
| .long 0x4d40c914 // ld1r {v20.4s}, [x8] |
| .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 |
| .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x6e711e00 // bsl v0.16b, v16.16b, v17.16b |
| .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 |
| .long 0x4e32f400 // fmax v0.4s, v0.4s, v18.4s |
| .long 0x4eb5f400 // fmin v0.4s, v0.4s, v21.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_parametric_g_aarch64 |
| .globl _sk_parametric_g_aarch64 |
| FUNCTION(_sk_parametric_g_aarch64) |
| _sk_parametric_g_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x9100610a // add x10, x8, #0x18 |
| .long 0x4d40c932 // ld1r {v18.4s}, [x9] |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0xbd400d11 // ldr s17, [x8, #12] |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 |
| .long 0x9100210a // add x10, x8, #0x8 |
| .long 0x4d40c954 // ld1r {v20.4s}, [x10] |
| .long 0x4f911030 // fmla v16.4s, v1.4s, v17.s[0] |
| .long 0xbd400135 // ldr s21, [x9] |
| .long 0x52b85f09 // mov w9, #0xc2f80000 |
| .long 0x728e6ee9 // movk w9, #0x7377 |
| .long 0x4e040d37 // dup v23.4s, w9 |
| .long 0x52a7f7e9 // mov w9, #0x3fbf0000 |
| .long 0x7297eea9 // movk w9, #0xbf75 |
| .long 0x4f951034 // fmla v20.4s, v1.4s, v21.s[0] |
| .long 0x6e21e641 // fcmge v1.4s, v18.4s, v1.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a7d689 // mov w9, #0x3eb40000 |
| .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 |
| .long 0x72889f29 // movk w9, #0x44f9 |
| .long 0x4e21da95 // scvtf v21.4s, v20.4s |
| .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7fb89 // mov w9, #0x3fdc0000 |
| .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s |
| .long 0x729d3469 // movk w9, #0xe9a3 |
| .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 |
| .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a85e49 // mov w9, #0x42f20000 |
| .long 0x72918a29 // movk w9, #0x8c51 |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7f7c9 // mov w9, #0x3fbe0000 |
| .long 0x729791a9 // movk w9, #0xbc8d |
| .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a81349 // mov w9, #0x409a0000 |
| .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s |
| .long 0x729ebf09 // movk w9, #0xf5f8 |
| .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x52a83ba9 // mov w9, #0x41dd0000 |
| .long 0x4e219a33 // frintm v19.4s, v17.4s |
| .long 0x729a5fc9 // movk w9, #0xd2fe |
| .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s |
| .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s |
| .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s |
| .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x91005108 // add x8, x8, #0x14 |
| .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s |
| .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s |
| .long 0x4d40c914 // ld1r {v20.4s}, [x8] |
| .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 |
| .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x6e711e01 // bsl v1.16b, v16.16b, v17.16b |
| .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 |
| .long 0x4e32f421 // fmax v1.4s, v1.4s, v18.4s |
| .long 0x4eb5f421 // fmin v1.4s, v1.4s, v21.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_parametric_b_aarch64 |
| .globl _sk_parametric_b_aarch64 |
| FUNCTION(_sk_parametric_b_aarch64) |
| _sk_parametric_b_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x9100610a // add x10, x8, #0x18 |
| .long 0x4d40c932 // ld1r {v18.4s}, [x9] |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0xbd400d11 // ldr s17, [x8, #12] |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 |
| .long 0x9100210a // add x10, x8, #0x8 |
| .long 0x4d40c954 // ld1r {v20.4s}, [x10] |
| .long 0x4f911050 // fmla v16.4s, v2.4s, v17.s[0] |
| .long 0xbd400135 // ldr s21, [x9] |
| .long 0x52b85f09 // mov w9, #0xc2f80000 |
| .long 0x728e6ee9 // movk w9, #0x7377 |
| .long 0x4e040d37 // dup v23.4s, w9 |
| .long 0x52a7f7e9 // mov w9, #0x3fbf0000 |
| .long 0x7297eea9 // movk w9, #0xbf75 |
| .long 0x4f951054 // fmla v20.4s, v2.4s, v21.s[0] |
| .long 0x6e22e642 // fcmge v2.4s, v18.4s, v2.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a7d689 // mov w9, #0x3eb40000 |
| .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 |
| .long 0x72889f29 // movk w9, #0x44f9 |
| .long 0x4e21da95 // scvtf v21.4s, v20.4s |
| .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7fb89 // mov w9, #0x3fdc0000 |
| .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s |
| .long 0x729d3469 // movk w9, #0xe9a3 |
| .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 |
| .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a85e49 // mov w9, #0x42f20000 |
| .long 0x72918a29 // movk w9, #0x8c51 |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7f7c9 // mov w9, #0x3fbe0000 |
| .long 0x729791a9 // movk w9, #0xbc8d |
| .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a81349 // mov w9, #0x409a0000 |
| .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s |
| .long 0x729ebf09 // movk w9, #0xf5f8 |
| .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x52a83ba9 // mov w9, #0x41dd0000 |
| .long 0x4e219a33 // frintm v19.4s, v17.4s |
| .long 0x729a5fc9 // movk w9, #0xd2fe |
| .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s |
| .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s |
| .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s |
| .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x91005108 // add x8, x8, #0x14 |
| .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s |
| .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s |
| .long 0x4d40c914 // ld1r {v20.4s}, [x8] |
| .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 |
| .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x6e711e02 // bsl v2.16b, v16.16b, v17.16b |
| .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 |
| .long 0x4e32f442 // fmax v2.4s, v2.4s, v18.4s |
| .long 0x4eb5f442 // fmin v2.4s, v2.4s, v21.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_parametric_a_aarch64 |
| .globl _sk_parametric_a_aarch64 |
| FUNCTION(_sk_parametric_a_aarch64) |
| _sk_parametric_a_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4f016696 // movi v22.4s, #0x34, lsl #24 |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x9100610a // add x10, x8, #0x18 |
| .long 0x4d40c932 // ld1r {v18.4s}, [x9] |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0xbd400d11 // ldr s17, [x8, #12] |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 |
| .long 0x9100210a // add x10, x8, #0x8 |
| .long 0x4d40c954 // ld1r {v20.4s}, [x10] |
| .long 0x4f911070 // fmla v16.4s, v3.4s, v17.s[0] |
| .long 0xbd400135 // ldr s21, [x9] |
| .long 0x52b85f09 // mov w9, #0xc2f80000 |
| .long 0x728e6ee9 // movk w9, #0x7377 |
| .long 0x4e040d37 // dup v23.4s, w9 |
| .long 0x52a7f7e9 // mov w9, #0x3fbf0000 |
| .long 0x7297eea9 // movk w9, #0xbf75 |
| .long 0x4f951074 // fmla v20.4s, v3.4s, v21.s[0] |
| .long 0x6e23e643 // fcmge v3.4s, v18.4s, v3.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a7d689 // mov w9, #0x3eb40000 |
| .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16 |
| .long 0x72889f29 // movk w9, #0x44f9 |
| .long 0x4e21da95 // scvtf v21.4s, v20.4s |
| .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7fb89 // mov w9, #0x3fdc0000 |
| .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s |
| .long 0x729d3469 // movk w9, #0xe9a3 |
| .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24 |
| .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a85e49 // mov w9, #0x42f20000 |
| .long 0x72918a29 // movk w9, #0x8c51 |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x4e040d34 // dup v20.4s, w9 |
| .long 0x52a7f7c9 // mov w9, #0x3fbe0000 |
| .long 0x729791a9 // movk w9, #0xbc8d |
| .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x52a81349 // mov w9, #0x409a0000 |
| .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s |
| .long 0x729ebf09 // movk w9, #0xf5f8 |
| .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x52a83ba9 // mov w9, #0x41dd0000 |
| .long 0x4e219a33 // frintm v19.4s, v17.4s |
| .long 0x729a5fc9 // movk w9, #0xd2fe |
| .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s |
| .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s |
| .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s |
| .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s |
| .long 0x4e040d35 // dup v21.4s, w9 |
| .long 0x91005108 // add x8, x8, #0x14 |
| .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s |
| .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s |
| .long 0x4d40c914 // ld1r {v20.4s}, [x8] |
| .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24 |
| .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x6e711e03 // bsl v3.16b, v16.16b, v17.16b |
| .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00 |
| .long 0x4e32f463 // fmax v3.4s, v3.4s, v18.4s |
| .long 0x4eb5f463 // fmin v3.4s, v3.4s, v21.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_a8_aarch64 |
| .globl _sk_load_a8_aarch64 |
| FUNCTION(_sk_load_a8_aarch64) |
| _sk_load_a8_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4e040d22 // dup v2.4s, w9 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x6f00e400 // movi v0.2d, #0x0 |
| .long 0x6f00e401 // movi v1.2d, #0x0 |
| .long 0x8b000108 // add x8, x8, x0 |
| .long 0x3940010a // ldrb w10, [x8] |
| .long 0x3940050b // ldrb w11, [x8, #1] |
| .long 0x3940090c // ldrb w12, [x8, #2] |
| .long 0x39400d08 // ldrb w8, [x8, #3] |
| .long 0x4e021d43 // mov v3.h[0], w10 |
| .long 0x4e061d63 // mov v3.h[1], w11 |
| .long 0x4e0a1d83 // mov v3.h[2], w12 |
| .long 0x4e0e1d03 // mov v3.h[3], w8 |
| .long 0x2f10a463 // uxtl v3.4s, v3.4h |
| .long 0x6e21d863 // ucvtf v3.4s, v3.4s |
| .long 0x6e22dc63 // fmul v3.4s, v3.4s, v2.4s |
| .long 0x6f00e402 // movi v2.2d, #0x0 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_gather_a8_aarch64 |
| .globl _sk_gather_a8_aarch64 |
| FUNCTION(_sk_gather_a8_aarch64) |
| _sk_gather_a8_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x4e040d23 // dup v3.4s, w9 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x386c490c // ldrb w12, [x8, w12, uxtw] |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x38694909 // ldrb w9, [x8, w9, uxtw] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x386a490a // ldrb w10, [x8, w10, uxtw] |
| .long 0x386b4908 // ldrb w8, [x8, w11, uxtw] |
| .long 0x4e021d82 // mov v2.h[0], w12 |
| .long 0x4e061d22 // mov v2.h[1], w9 |
| .long 0x4e0a1d42 // mov v2.h[2], w10 |
| .long 0x4e0e1d02 // mov v2.h[3], w8 |
| .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8 |
| .long 0x2f10a442 // uxtl v2.4s, v2.4h |
| .long 0x6e21d842 // ucvtf v2.4s, v2.4s |
| .long 0x6f00e400 // movi v0.2d, #0x0 |
| .long 0x6f00e401 // movi v1.2d, #0x0 |
| .long 0x6e23dc43 // fmul v3.4s, v2.4s, v3.4s |
| .long 0x6f00e402 // movi v2.2d, #0x0 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_store_a8_aarch64 |
| .globl _sk_store_a8_aarch64 |
| FUNCTION(_sk_store_a8_aarch64) |
| _sk_store_a8_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52a86fe9 // mov w9, #0x437f0000 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| .long 0x0e612a10 // xtn v16.4h, v16.4s |
| .long 0x0e0e3e09 // umov w9, v16.h[3] |
| .long 0x8b000108 // add x8, x8, x0 |
| .long 0x39000d09 // strb w9, [x8, #3] |
| .long 0x0e0a3e09 // umov w9, v16.h[2] |
| .long 0x39000909 // strb w9, [x8, #2] |
| .long 0x0e063e09 // umov w9, v16.h[1] |
| .long 0x39000509 // strb w9, [x8, #1] |
| .long 0x0e023e09 // umov w9, v16.h[0] |
| .long 0x39000109 // strb w9, [x8] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_g8_aarch64 |
| .globl _sk_load_g8_aarch64 |
| FUNCTION(_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_gather_g8_aarch64 |
| .globl _sk_gather_g8_aarch64 |
| FUNCTION(_sk_gather_g8_aarch64) |
| _sk_gather_g8_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x4e040d23 // dup v3.4s, w9 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x386c490c // ldrb w12, [x8, w12, uxtw] |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x38694909 // ldrb w9, [x8, w9, uxtw] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x386a490a // ldrb w10, [x8, w10, uxtw] |
| .long 0x386b4908 // ldrb w8, [x8, w11, uxtw] |
| .long 0x4e021d80 // mov v0.h[0], w12 |
| .long 0x4e061d20 // mov v0.h[1], w9 |
| .long 0x4e0a1d40 // mov v0.h[2], w10 |
| .long 0x4e0e1d00 // mov v0.h[3], w8 |
| .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8 |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x6e21d800 // ucvtf v0.4s, v0.4s |
| .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s |
| .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| .long 0x4ea01c01 // mov v1.16b, v0.16b |
| .long 0x4ea01c02 // mov v2.16b, v0.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_gather_i8_aarch64 |
| .globl _sk_gather_i8_aarch64 |
| FUNCTION(_sk_gather_i8_aarch64) |
| _sk_gather_i8_aarch64: |
| .long 0xaa0103e8 // mov x8, x1 |
| .long 0xf8408429 // ldr x9, [x1], #8 |
| .long 0xb4000069 // cbz x9, 1f40 <sk_gather_i8_aarch64+0x14> |
| .long 0xaa0903ea // mov x10, x9 |
| .long 0x14000003 // b 1f48 <sk_gather_i8_aarch64+0x1c> |
| .long 0xf940050a // ldr x10, [x8, #8] |
| .long 0x91004101 // add x1, x8, #0x10 |
| .long 0xf8410548 // ldr x8, [x10], #16 |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0xf9400529 // ldr x9, [x9, #8] |
| .long 0x4d40c942 // ld1r {v2.4s}, [x10] |
| .long 0x6f00e623 // movi v3.2d, #0xff000000ff |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x1e26000d // fmov w13, s0 |
| .long 0x0e0c3c0a // mov w10, v0.s[1] |
| .long 0x386d490d // ldrb w13, [x8, w13, uxtw] |
| .long 0x0e143c0b // mov w11, v0.s[2] |
| .long 0x386a490a // ldrb w10, [x8, w10, uxtw] |
| .long 0x0e1c3c0c // mov w12, v0.s[3] |
| .long 0x386b490b // ldrb w11, [x8, w11, uxtw] |
| .long 0x386c4908 // ldrb w8, [x8, w12, uxtw] |
| .long 0x4e021da0 // mov v0.h[0], w13 |
| .long 0x4e061d40 // mov v0.h[1], w10 |
| .long 0x4e0a1d60 // mov v0.h[2], w11 |
| .long 0x4e0e1d00 // mov v0.h[3], w8 |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x4e231c00 // and v0.16b, v0.16b, v3.16b |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x8b2c492c // add x12, x9, w12, uxtw #2 |
| .long 0x0e0c3c08 // mov w8, v0.s[1] |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x0d408180 // ld1 {v0.s}[0], [x12] |
| .long 0x8b284928 // add x8, x9, w8, uxtw #2 |
| .long 0xb86a592a // ldr w10, [x9, w10, uxtw #2] |
| .long 0x52a7700c // mov w12, #0x3b800000 |
| .long 0x0d409100 // ld1 {v0.s}[1], [x8] |
| .long 0xb86b5928 // ldr w8, [x9, w11, uxtw #2] |
| .long 0x7290102c // movk w12, #0x8081 |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x4e141d40 // mov v0.s[2], w10 |
| .long 0x4e1c1d00 // mov v0.s[3], w8 |
| .long 0x4e231c01 // and v1.16b, v0.16b, v3.16b |
| .long 0x6f380402 // ushr v2.4s, v0.4s, #8 |
| .long 0x6f300411 // ushr v17.4s, v0.4s, #16 |
| .long 0x4e040d90 // dup v16.4s, w12 |
| .long 0x6f280400 // ushr v0.4s, v0.4s, #24 |
| .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| .long 0x4e231c42 // and v2.16b, v2.16b, v3.16b |
| .long 0x4e231e23 // and v3.16b, v17.16b, v3.16b |
| .long 0x4e21d811 // scvtf v17.4s, v0.4s |
| .long 0x6e30dc20 // fmul v0.4s, v1.4s, v16.4s |
| .long 0x4e21d841 // scvtf v1.4s, v2.4s |
| .long 0x4e21d862 // scvtf v2.4s, v3.4s |
| .long 0x6e30dc21 // fmul v1.4s, v1.4s, v16.4s |
| .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s |
| .long 0x6e30de23 // fmul v3.4s, v17.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_565_aarch64 |
| .globl _sk_load_565_aarch64 |
| FUNCTION(_sk_load_565_aarch64) |
| _sk_load_565_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xd37ff809 // lsl x9, x0, #1 |
| .long 0x4f072701 // movi v1.4s, #0xf8, lsl #8 |
| .long 0x4f0007e3 // movi v3.4s, #0x1f |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0xfc696900 // ldr d0, [x8, x9] |
| .long 0x321b17e8 // orr w8, wzr, #0x7e0 |
| .long 0x4e040d02 // dup v2.4s, w8 |
| .long 0x52a6f088 // mov w8, #0x37840000 |
| .long 0x72842108 // movk w8, #0x2108 |
| .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 0x4e040d00 // dup v0.4s, w8 |
| .long 0x52a74048 // mov w8, #0x3a020000 |
| .long 0x72810428 // movk w8, #0x821 |
| .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s |
| .long 0x4e040d01 // dup v1.4s, w8 |
| .long 0x52a7a088 // mov w8, #0x3d040000 |
| .long 0x72842108 // movk w8, #0x2108 |
| .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s |
| .long 0x4e040d02 // dup v2.4s, w8 |
| .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s |
| .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_gather_565_aarch64 |
| .globl _sk_gather_565_aarch64 |
| FUNCTION(_sk_gather_565_aarch64) |
| _sk_gather_565_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x321b17e9 // orr w9, wzr, #0x7e0 |
| .long 0x4e040d23 // dup v3.4s, w9 |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x8b2c450c // add x12, x8, w12, uxtw #1 |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x0d404180 // ld1 {v0.h}[0], [x12] |
| .long 0x78695909 // ldrh w9, [x8, w9, uxtw #1] |
| .long 0x786a590a // ldrh w10, [x8, w10, uxtw #1] |
| .long 0x786b5908 // ldrh w8, [x8, w11, uxtw #1] |
| .long 0x4f072701 // movi v1.4s, #0xf8, lsl #8 |
| .long 0x4e061d20 // mov v0.h[1], w9 |
| .long 0x4e0a1d40 // mov v0.h[2], w10 |
| .long 0x4e0e1d00 // mov v0.h[3], w8 |
| .long 0x52a6f08b // mov w11, #0x37840000 |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x7284210b // movk w11, #0x2108 |
| .long 0x52a74049 // mov w9, #0x3a020000 |
| .long 0x4f0007e2 // movi v2.4s, #0x1f |
| .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b |
| .long 0x72810429 // movk w9, #0x821 |
| .long 0x52a7a08a // mov w10, #0x3d040000 |
| .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b |
| .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b |
| .long 0x4e040d60 // dup v0.4s, w11 |
| .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| .long 0x7284210a // movk w10, #0x2108 |
| .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s |
| .long 0x4e040d21 // dup v1.4s, w9 |
| .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| .long 0x6e21dc61 // fmul v1.4s, v3.4s, v1.4s |
| .long 0x4e040d43 // dup v3.4s, w10 |
| .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| .long 0x6e23dc42 // fmul v2.4s, v2.4s, v3.4s |
| .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_store_565_aarch64 |
| .globl _sk_store_565_aarch64 |
| FUNCTION(_sk_store_565_aarch64) |
| _sk_store_565_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52a84f8a // mov w10, #0x427c0000 |
| .long 0x4f01f7f0 // fmov v16.4s, #3.100000000000000000e+01 |
| .long 0x4e040d52 // dup v18.4s, w10 |
| .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| .long 0x6e32dc32 // fmul v18.4s, v1.4s, v18.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| .long 0x6e30dc50 // fmul v16.4s, v2.4s, v16.4s |
| .long 0x4f2b5631 // shl v17.4s, v17.4s, #11 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x4f255652 // shl v18.4s, v18.4s, #5 |
| .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b |
| .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_4444_aarch64 |
| .globl _sk_load_4444_aarch64 |
| FUNCTION(_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_gather_4444_aarch64 |
| .globl _sk_gather_4444_aarch64 |
| FUNCTION(_sk_gather_4444_aarch64) |
| _sk_gather_4444_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0x4f070603 // movi v3.4s, #0xf0 |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x4f0005f0 // movi v16.4s, #0xf |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x8b2c450c // add x12, x8, w12, uxtw #1 |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x0d404180 // ld1 {v0.h}[0], [x12] |
| .long 0x78695909 // ldrh w9, [x8, w9, uxtw #1] |
| .long 0x786a590a // ldrh w10, [x8, w10, uxtw #1] |
| .long 0x786b5908 // ldrh w8, [x8, w11, uxtw #1] |
| .long 0x4f072601 // movi v1.4s, #0xf0, lsl #8 |
| .long 0x4e061d20 // mov v0.h[1], w9 |
| .long 0x4e0a1d40 // mov v0.h[2], w10 |
| .long 0x4e0e1d00 // mov v0.h[3], w8 |
| .long 0x52a6f10b // mov w11, #0x37880000 |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x7291112b // movk w11, #0x8889 |
| .long 0x4f0025e2 // movi v2.4s, #0xf, lsl #8 |
| .long 0x52a73109 // mov w9, #0x39880000 |
| .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b |
| .long 0x72911129 // movk w9, #0x8889 |
| .long 0x52a7710a // mov w10, #0x3b880000 |
| .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 0x4e040d60 // dup v0.4s, w11 |
| .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| .long 0x7291112a // movk w10, #0x8889 |
| .long 0x52a7b108 // mov w8, #0x3d880000 |
| .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s |
| .long 0x4e040d21 // dup v1.4s, w9 |
| .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| .long 0x72911128 // movk w8, #0x8889 |
| .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s |
| .long 0x4e040d42 // dup v2.4s, w10 |
| .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 |
| FUNCTION(_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 |
| FUNCTION(_sk_load_8888_aarch64) |
| _sk_load_8888_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xd37ef409 // lsl x9, x0, #2 |
| .long 0x6f00e621 // movi v1.2d, #0xff000000ff |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x3ce96900 // ldr q0, [x8, x9] |
| .long 0x52a77008 // mov w8, #0x3b800000 |
| .long 0x72901028 // movk w8, #0x8081 |
| .long 0x4e040d02 // dup v2.4s, w8 |
| .long 0x6f380410 // ushr v16.4s, v0.4s, #8 |
| .long 0x6f300411 // ushr v17.4s, v0.4s, #16 |
| .long 0x4e211c03 // and v3.16b, v0.16b, v1.16b |
| .long 0x6f280400 // ushr v0.4s, v0.4s, #24 |
| .long 0x4e211e10 // and v16.16b, v16.16b, v1.16b |
| .long 0x4e211e21 // and v1.16b, v17.16b, v1.16b |
| .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| .long 0x4e21d811 // scvtf v17.4s, v0.4s |
| .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| .long 0x4e21d832 // scvtf v18.4s, v1.4s |
| .long 0x6e22dc60 // fmul v0.4s, v3.4s, v2.4s |
| .long 0x6e22de23 // fmul v3.4s, v17.4s, v2.4s |
| .long 0x6e22de01 // fmul v1.4s, v16.4s, v2.4s |
| .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_gather_8888_aarch64 |
| .globl _sk_gather_8888_aarch64 |
| FUNCTION(_sk_gather_8888_aarch64) |
| _sk_gather_8888_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x8b2c490c // add x12, x8, w12, uxtw #2 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x0d408180 // ld1 {v0.s}[0], [x12] |
| .long 0x8b294909 // add x9, x8, w9, uxtw #2 |
| .long 0xb86a590a // ldr w10, [x8, w10, uxtw #2] |
| .long 0xb86b5908 // ldr w8, [x8, w11, uxtw #2] |
| .long 0x0d409120 // ld1 {v0.s}[1], [x9] |
| .long 0x6f00e621 // movi v1.2d, #0xff000000ff |
| .long 0x52a77009 // mov w9, #0x3b800000 |
| .long 0x72901029 // movk w9, #0x8081 |
| .long 0x4e141d40 // mov v0.s[2], w10 |
| .long 0x4e1c1d00 // mov v0.s[3], w8 |
| .long 0x6f380410 // ushr v16.4s, v0.4s, #8 |
| .long 0x6f300411 // ushr v17.4s, v0.4s, #16 |
| .long 0x4e211c03 // and v3.16b, v0.16b, v1.16b |
| .long 0x6f280400 // ushr v0.4s, v0.4s, #24 |
| .long 0x4e211e10 // and v16.16b, v16.16b, v1.16b |
| .long 0x4e211e21 // and v1.16b, v17.16b, v1.16b |
| .long 0x4e040d22 // dup v2.4s, w9 |
| .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| .long 0x4e21d811 // scvtf v17.4s, v0.4s |
| .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| .long 0x4e21d832 // scvtf v18.4s, v1.4s |
| .long 0x6e22dc60 // fmul v0.4s, v3.4s, v2.4s |
| .long 0x6e22de23 // fmul v3.4s, v17.4s, v2.4s |
| .long 0x6e22de01 // fmul v1.4s, v16.4s, v2.4s |
| .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_store_8888_aarch64 |
| .globl _sk_store_8888_aarch64 |
| FUNCTION(_sk_store_8888_aarch64) |
| _sk_store_8888_aarch64: |
| .long 0x52a86fea // mov w10, #0x437f0000 |
| .long 0x4e040d50 // dup v16.4s, w10 |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s |
| .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .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 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| .long 0x4f305652 // shl v18.4s, v18.4s, #16 |
| .long 0x4eb21e31 // orr v17.16b, v17.16b, v18.16b |
| .long 0x4f385610 // shl v16.4s, v16.4s, #24 |
| .long 0xd37ef409 // lsl x9, x0, #2 |
| .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b |
| .long 0x3ca96910 // str q16, [x8, x9] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_f16_aarch64 |
| .globl _sk_load_f16_aarch64 |
| FUNCTION(_sk_load_f16_aarch64) |
| _sk_load_f16_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| .long 0x0c400510 // ld4 {v16.4h-v19.4h}, [x8] |
| .long 0x0e217a00 // fcvtl v0.4s, v16.4h |
| .long 0x0e217a21 // fcvtl v1.4s, v17.4h |
| .long 0x0e217a42 // fcvtl v2.4s, v18.4h |
| .long 0x0e217a63 // fcvtl v3.4s, v19.4h |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_gather_f16_aarch64 |
| .globl _sk_gather_f16_aarch64 |
| FUNCTION(_sk_gather_f16_aarch64) |
| _sk_gather_f16_aarch64: |
| .long 0xa9bf7bfd // stp x29, x30, [sp, #-16]! |
| .long 0xd100c3e9 // sub x9, sp, #0x30 |
| .long 0x910003fd // mov x29, sp |
| .long 0x927be93f // and sp, x9, #0xffffffffffffffe0 |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s |
| .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s |
| .long 0x91004109 // add x9, x8, #0x10 |
| .long 0x4d40c922 // ld1r {v2.4s}, [x9] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s |
| .long 0x0e143c0a // mov w10, v0.s[2] |
| .long 0x1e26000c // fmov w12, s0 |
| .long 0x8b2c4d0c // add x12, x8, w12, uxtw #3 |
| .long 0x8b2a4d0a // add x10, x8, w10, uxtw #3 |
| .long 0x0e0c3c09 // mov w9, v0.s[1] |
| .long 0x0e1c3c0b // mov w11, v0.s[3] |
| .long 0x0d408540 // ld1 {v0.d}[0], [x10] |
| .long 0x0d408581 // ld1 {v1.d}[0], [x12] |
| .long 0x8b294d09 // add x9, x8, w9, uxtw #3 |
| .long 0x8b2b4d08 // add x8, x8, w11, uxtw #3 |
| .long 0x4d408500 // ld1 {v0.d}[1], [x8] |
| .long 0x4d408521 // ld1 {v1.d}[1], [x9] |
| .long 0x910003e8 // mov x8, sp |
| .long 0xad0003e1 // stp q1, q0, [sp] |
| .long 0x0c400510 // ld4 {v16.4h-v19.4h}, [x8] |
| .long 0xf9400428 // ldr x8, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0x0e217a00 // fcvtl v0.4s, v16.4h |
| .long 0x0e217a21 // fcvtl v1.4s, v17.4h |
| .long 0x0e217a42 // fcvtl v2.4s, v18.4h |
| .long 0x0e217a63 // fcvtl v3.4s, v19.4h |
| .long 0xd63f0100 // blr x8 |
| .long 0x910003bf // mov sp, x29 |
| .long 0xa8c17bfd // ldp x29, x30, [sp], #16 |
| .long 0xd65f03c0 // ret |
| |
| HIDDEN _sk_store_f16_aarch64 |
| .globl _sk_store_f16_aarch64 |
| FUNCTION(_sk_store_f16_aarch64) |
| _sk_store_f16_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x0e216810 // fcvtn v16.4h, v0.4s |
| .long 0x0e216831 // fcvtn v17.4h, v1.4s |
| .long 0x0e216852 // fcvtn v18.4h, v2.4s |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x0e216873 // fcvtn v19.4h, v3.4s |
| .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| .long 0x0c000510 // st4 {v16.4h-v19.4h}, [x8] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_u16_be_aarch64 |
| .globl _sk_load_u16_be_aarch64 |
| FUNCTION(_sk_load_u16_be_aarch64) |
| _sk_load_u16_be_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| .long 0x0c400500 // ld4 {v0.4h-v3.4h}, [x8] |
| .long 0x52a6f008 // mov w8, #0x37800000 |
| .long 0x72801008 // movk w8, #0x80 |
| .long 0x0f185410 // shl v16.4h, v0.4h, #8 |
| .long 0x2f180411 // ushr v17.4h, v0.4h, #8 |
| .long 0x0f185432 // shl v18.4h, v1.4h, #8 |
| .long 0x2f180433 // ushr v19.4h, v1.4h, #8 |
| .long 0x0f185454 // shl v20.4h, v2.4h, #8 |
| .long 0x2f180455 // ushr v21.4h, v2.4h, #8 |
| .long 0x0f185476 // shl v22.4h, v3.4h, #8 |
| .long 0x2f180460 // ushr v0.4h, v3.4h, #8 |
| .long 0x0eb11e01 // orr v1.8b, v16.8b, v17.8b |
| .long 0x0eb31e42 // orr v2.8b, v18.8b, v19.8b |
| .long 0x0eb51e90 // orr v16.8b, v20.8b, v21.8b |
| .long 0x0ea01ec0 // orr v0.8b, v22.8b, v0.8b |
| .long 0x2f10a421 // uxtl v1.4s, v1.4h |
| .long 0x2f10a442 // uxtl v2.4s, v2.4h |
| .long 0x2f10a610 // uxtl v16.4s, v16.4h |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x4e040d03 // dup v3.4s, w8 |
| .long 0x6e21d821 // ucvtf v1.4s, v1.4s |
| .long 0x6e21d842 // ucvtf v2.4s, v2.4s |
| .long 0x6e21da10 // ucvtf v16.4s, v16.4s |
| .long 0x6e21d811 // ucvtf v17.4s, v0.4s |
| .long 0x6e23dc20 // fmul v0.4s, v1.4s, v3.4s |
| .long 0x6e23dc41 // fmul v1.4s, v2.4s, v3.4s |
| .long 0x6e23de02 // fmul v2.4s, v16.4s, v3.4s |
| .long 0x6e23de23 // fmul v3.4s, v17.4s, v3.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_rgb_u16_be_aarch64 |
| .globl _sk_load_rgb_u16_be_aarch64 |
| FUNCTION(_sk_load_rgb_u16_be_aarch64) |
| _sk_load_rgb_u16_be_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x321f07e9 // orr w9, wzr, #0x6 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x9b092008 // madd x8, x0, x9, x8 |
| .long 0x0c404500 // ld3 {v0.4h-v2.4h}, [x8] |
| .long 0x52a6f008 // mov w8, #0x37800000 |
| .long 0x72801008 // movk w8, #0x80 |
| .long 0x0f185403 // shl v3.4h, v0.4h, #8 |
| .long 0x2f180410 // ushr v16.4h, v0.4h, #8 |
| .long 0x0f185431 // shl v17.4h, v1.4h, #8 |
| .long 0x2f180432 // ushr v18.4h, v1.4h, #8 |
| .long 0x0f185453 // shl v19.4h, v2.4h, #8 |
| .long 0x2f180440 // ushr v0.4h, v2.4h, #8 |
| .long 0x0eb01c61 // orr v1.8b, v3.8b, v16.8b |
| .long 0x0eb21e23 // orr v3.8b, v17.8b, v18.8b |
| .long 0x0ea01e60 // orr v0.8b, v19.8b, v0.8b |
| .long 0x2f10a421 // uxtl v1.4s, v1.4h |
| .long 0x2f10a463 // uxtl v3.4s, v3.4h |
| .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| .long 0x4e040d02 // dup v2.4s, w8 |
| .long 0x6e21d821 // ucvtf v1.4s, v1.4s |
| .long 0x6e21d863 // ucvtf v3.4s, v3.4s |
| .long 0x6e21d810 // ucvtf v16.4s, v0.4s |
| .long 0x6e22dc20 // fmul v0.4s, v1.4s, v2.4s |
| .long 0x6e22dc61 // fmul v1.4s, v3.4s, v2.4s |
| .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_store_u16_be_aarch64 |
| .globl _sk_store_u16_be_aarch64 |
| FUNCTION(_sk_store_u16_be_aarch64) |
| _sk_store_u16_be_aarch64: |
| .long 0x52a8efe9 // mov w9, #0x477f0000 |
| .long 0x729fe009 // movk w9, #0xff00 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0x0e612a31 // xtn v17.4h, v17.4s |
| .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s |
| .long 0x0f185633 // shl v19.4h, v17.4h, #8 |
| .long 0x2f180631 // ushr v17.4h, v17.4h, #8 |
| .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| .long 0x0eb11e75 // orr v21.8b, v19.8b, v17.8b |
| .long 0x6e30dc51 // fmul v17.4s, v2.4s, v16.4s |
| .long 0x0e612a52 // xtn v18.4h, v18.4s |
| .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x0f185654 // shl v20.4h, v18.4h, #8 |
| .long 0x2f180652 // ushr v18.4h, v18.4h, #8 |
| .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| .long 0x0e612a31 // xtn v17.4h, v17.4s |
| .long 0x0eb21e96 // orr v22.8b, v20.8b, v18.8b |
| .long 0x0e612a10 // xtn v16.4h, v16.4s |
| .long 0x0f185632 // shl v18.4h, v17.4h, #8 |
| .long 0x2f180631 // ushr v17.4h, v17.4h, #8 |
| .long 0x0eb11e57 // orr v23.8b, v18.8b, v17.8b |
| .long 0x0f185611 // shl v17.4h, v16.4h, #8 |
| .long 0x2f180610 // ushr v16.4h, v16.4h, #8 |
| .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| .long 0x0eb01e38 // orr v24.8b, v17.8b, v16.8b |
| .long 0x0c000515 // st4 {v21.4h-v24.4h}, [x8] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_load_f32_aarch64 |
| .globl _sk_load_f32_aarch64 |
| FUNCTION(_sk_load_f32_aarch64) |
| _sk_load_f32_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x8b001108 // add x8, x8, x0, lsl #4 |
| .long 0x4c400900 // ld4 {v0.4s-v3.4s}, [x8] |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_store_f32_aarch64 |
| .globl _sk_store_f32_aarch64 |
| FUNCTION(_sk_store_f32_aarch64) |
| _sk_store_f32_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0xf9400108 // ldr x8, [x8] |
| .long 0x8b001108 // add x8, x8, x0, lsl #4 |
| .long 0x4c000900 // st4 {v0.4s-v3.4s}, [x8] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_clamp_x_aarch64 |
| .globl _sk_clamp_x_aarch64 |
| FUNCTION(_sk_clamp_x_aarch64) |
| _sk_clamp_x_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x6f00e411 // movi v17.2d, #0x0 |
| .long 0x4e20f620 // fmax v0.4s, v17.4s, v0.4s |
| .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| .long 0x4d40c910 // ld1r {v16.4s}, [x8] |
| .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s |
| .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_clamp_y_aarch64 |
| .globl _sk_clamp_y_aarch64 |
| FUNCTION(_sk_clamp_y_aarch64) |
| _sk_clamp_y_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x6f00e411 // movi v17.2d, #0x0 |
| .long 0x4e21f621 // fmax v1.4s, v17.4s, v1.4s |
| .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| .long 0x4d40c910 // ld1r {v16.4s}, [x8] |
| .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s |
| .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_repeat_x_aarch64 |
| .globl _sk_repeat_x_aarch64 |
| FUNCTION(_sk_repeat_x_aarch64) |
| _sk_repeat_x_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| .long 0xbd400110 // ldr s16, [x8] |
| .long 0x4e040612 // dup v18.4s, v16.s[0] |
| .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s |
| .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s |
| .long 0x4e219a52 // frintm v18.4s, v18.4s |
| .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0] |
| .long 0x4eb1f400 // fmin v0.4s, v0.4s, v17.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_repeat_y_aarch64 |
| .globl _sk_repeat_y_aarch64 |
| FUNCTION(_sk_repeat_y_aarch64) |
| _sk_repeat_y_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| .long 0xbd400110 // ldr s16, [x8] |
| .long 0x4e040612 // dup v18.4s, v16.s[0] |
| .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s |
| .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s |
| .long 0x4e219a52 // frintm v18.4s, v18.4s |
| .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0] |
| .long 0x4eb1f421 // fmin v1.4s, v1.4s, v17.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_mirror_x_aarch64 |
| .globl _sk_mirror_x_aarch64 |
| FUNCTION(_sk_mirror_x_aarch64) |
| _sk_mirror_x_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xbd400110 // ldr s16, [x8] |
| .long 0x4e040611 // dup v17.4s, v16.s[0] |
| .long 0x1e302a10 // fadd s16, s16, s16 |
| .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s |
| .long 0x4e040612 // dup v18.4s, v16.s[0] |
| .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s |
| .long 0x4e219a52 // frintm v18.4s, v18.4s |
| .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0] |
| .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff |
| .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s |
| .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s |
| .long 0x4ea0f800 // fabs v0.4s, v0.4s |
| .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_mirror_y_aarch64 |
| .globl _sk_mirror_y_aarch64 |
| FUNCTION(_sk_mirror_y_aarch64) |
| _sk_mirror_y_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xbd400110 // ldr s16, [x8] |
| .long 0x4e040611 // dup v17.4s, v16.s[0] |
| .long 0x1e302a10 // fadd s16, s16, s16 |
| .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s |
| .long 0x4e040612 // dup v18.4s, v16.s[0] |
| .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s |
| .long 0x4e219a52 // frintm v18.4s, v18.4s |
| .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0] |
| .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff |
| .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s |
| .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s |
| .long 0x4ea0f821 // fabs v1.4s, v1.4s |
| .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_luminance_to_alpha_aarch64 |
| .globl _sk_luminance_to_alpha_aarch64 |
| FUNCTION(_sk_luminance_to_alpha_aarch64) |
| _sk_luminance_to_alpha_aarch64: |
| .long 0x52a7cb28 // mov w8, #0x3e590000 |
| .long 0x72967a08 // movk w8, #0xb3d0 |
| .long 0x4e040d11 // dup v17.4s, w8 |
| .long 0x52a7e6e8 // mov w8, #0x3f370000 |
| .long 0x7282eb28 // movk w8, #0x1759 |
| .long 0x4ea01c10 // mov v16.16b, v0.16b |
| .long 0x4e040d00 // dup v0.4s, w8 |
| .long 0x52a7b268 // mov w8, #0x3d930000 |
| .long 0xf8408423 // ldr x3, [x1], #8 |
| .long 0x729bb308 // movk w8, #0xdd98 |
| .long 0x6e20dc23 // fmul v3.4s, v1.4s, v0.4s |
| .long 0x4e30ce23 // fmla v3.4s, v17.4s, v16.4s |
| .long 0x4e040d10 // dup v16.4s, w8 |
| .long 0x6f00e400 // movi v0.2d, #0x0 |
| .long 0x6f00e401 // movi v1.2d, #0x0 |
| .long 0x4e22ce03 // fmla v3.4s, v16.4s, v2.4s |
| .long 0x6f00e402 // movi v2.2d, #0x0 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_matrix_2x3_aarch64 |
| .globl _sk_matrix_2x3_aarch64 |
| FUNCTION(_sk_matrix_2x3_aarch64) |
| _sk_matrix_2x3_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0x9100410a // add x10, x8, #0x10 |
| .long 0x4ddfc932 // ld1r {v18.4s}, [x9], #4 |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x2d415113 // ldp s19, s20, [x8, #8] |
| .long 0x9100510a // add x10, x8, #0x14 |
| .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| .long 0x4f931030 // fmla v16.4s, v1.4s, v19.s[0] |
| .long 0xbd400133 // ldr s19, [x9] |
| .long 0x4f941031 // fmla v17.4s, v1.4s, v20.s[0] |
| .long 0x4e20ce50 // fmla v16.4s, v18.4s, v0.4s |
| .long 0x4f931011 // fmla v17.4s, v0.4s, v19.s[0] |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_matrix_3x4_aarch64 |
| .globl _sk_matrix_3x4_aarch64 |
| FUNCTION(_sk_matrix_3x4_aarch64) |
| _sk_matrix_3x4_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0x9100910a // add x10, x8, #0x24 |
| .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x9100a10a // add x10, x8, #0x28 |
| .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| .long 0x9100b10a // add x10, x8, #0x2c |
| .long 0x2d435514 // ldp s20, s21, [x8, #24] |
| .long 0xbd402116 // ldr s22, [x8, #32] |
| .long 0x4d40c952 // ld1r {v18.4s}, [x10] |
| .long 0x4f941050 // fmla v16.4s, v2.4s, v20.s[0] |
| .long 0x4f951051 // fmla v17.4s, v2.4s, v21.s[0] |
| .long 0x4f961052 // fmla v18.4s, v2.4s, v22.s[0] |
| .long 0x2d425502 // ldp s2, s21, [x8, #16] |
| .long 0x2d415d14 // ldp s20, s23, [x8, #8] |
| .long 0x4f821031 // fmla v17.4s, v1.4s, v2.s[0] |
| .long 0xbd400122 // ldr s2, [x9] |
| .long 0x4f971030 // fmla v16.4s, v1.4s, v23.s[0] |
| .long 0x4f951032 // fmla v18.4s, v1.4s, v21.s[0] |
| .long 0x4e20ce70 // fmla v16.4s, v19.4s, v0.4s |
| .long 0x4f941012 // fmla v18.4s, v0.4s, v20.s[0] |
| .long 0x4f821011 // fmla v17.4s, v0.4s, v2.s[0] |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0x4eb21e42 // mov v2.16b, v18.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_matrix_4x5_aarch64 |
| .globl _sk_matrix_4x5_aarch64 |
| FUNCTION(_sk_matrix_4x5_aarch64) |
| _sk_matrix_4x5_aarch64: |
| .long 0xf9400029 // ldr x9, [x1] |
| .long 0xaa0903e8 // mov x8, x9 |
| .long 0x9101012a // add x10, x9, #0x40 |
| .long 0x4ddfc914 // ld1r {v20.4s}, [x8], #4 |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x9101112a // add x10, x9, #0x44 |
| .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| .long 0x9101212a // add x10, x9, #0x48 |
| .long 0x4d40c952 // ld1r {v18.4s}, [x10] |
| .long 0x2d465533 // ldp s19, s21, [x9, #48] |
| .long 0x2d475d36 // ldp s22, s23, [x9, #56] |
| .long 0x9101312a // add x10, x9, #0x4c |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f931070 // fmla v16.4s, v3.4s, v19.s[0] |
| .long 0x4d40c953 // ld1r {v19.4s}, [x10] |
| .long 0x4f951071 // fmla v17.4s, v3.4s, v21.s[0] |
| .long 0x4f961072 // fmla v18.4s, v3.4s, v22.s[0] |
| .long 0x2d445935 // ldp s21, s22, [x9, #32] |
| .long 0x4f971073 // fmla v19.4s, v3.4s, v23.s[0] |
| .long 0x2d455d23 // ldp s3, s23, [x9, #40] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0x4f951050 // fmla v16.4s, v2.4s, v21.s[0] |
| .long 0x4f961051 // fmla v17.4s, v2.4s, v22.s[0] |
| .long 0x2d425935 // ldp s21, s22, [x9, #16] |
| .long 0x4f971053 // fmla v19.4s, v2.4s, v23.s[0] |
| .long 0x4f831052 // fmla v18.4s, v2.4s, v3.s[0] |
| .long 0x2d410d22 // ldp s2, s3, [x9, #8] |
| .long 0x4f951030 // fmla v16.4s, v1.4s, v21.s[0] |
| .long 0x2d435d35 // ldp s21, s23, [x9, #24] |
| .long 0x4f961031 // fmla v17.4s, v1.4s, v22.s[0] |
| .long 0xbd400116 // ldr s22, [x8] |
| .long 0x4e20ce90 // fmla v16.4s, v20.4s, v0.4s |
| .long 0x4f951032 // fmla v18.4s, v1.4s, v21.s[0] |
| .long 0x4f971033 // fmla v19.4s, v1.4s, v23.s[0] |
| .long 0x4f821012 // fmla v18.4s, v0.4s, v2.s[0] |
| .long 0x4f831013 // fmla v19.4s, v0.4s, v3.s[0] |
| .long 0x4f961011 // fmla v17.4s, v0.4s, v22.s[0] |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0x4eb11e21 // mov v1.16b, v17.16b |
| .long 0x4eb21e42 // mov v2.16b, v18.16b |
| .long 0x4eb31e63 // mov v3.16b, v19.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_matrix_perspective_aarch64 |
| .globl _sk_matrix_perspective_aarch64 |
| FUNCTION(_sk_matrix_perspective_aarch64) |
| _sk_matrix_perspective_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0x9100510a // add x10, x8, #0x14 |
| .long 0x4ddfc930 // ld1r {v16.4s}, [x9], #4 |
| .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| .long 0x9100810a // add x10, x8, #0x20 |
| .long 0x4d40c952 // ld1r {v18.4s}, [x10] |
| .long 0x2d41d113 // ldp s19, s20, [x8, #12] |
| .long 0x2d435915 // ldp s21, s22, [x8, #24] |
| .long 0x91002108 // add x8, x8, #0x8 |
| .long 0x4f941031 // fmla v17.4s, v1.4s, v20.s[0] |
| .long 0x4d40c914 // ld1r {v20.4s}, [x8] |
| .long 0x4f961032 // fmla v18.4s, v1.4s, v22.s[0] |
| .long 0xbd400136 // ldr s22, [x9] |
| .long 0x4f951012 // fmla v18.4s, v0.4s, v21.s[0] |
| .long 0x4f931011 // fmla v17.4s, v0.4s, v19.s[0] |
| .long 0x4f961034 // fmla v20.4s, v1.4s, v22.s[0] |
| .long 0x4ea1da41 // frecpe v1.4s, v18.4s |
| .long 0x4e21fe52 // frecps v18.4s, v18.4s, v1.4s |
| .long 0x6e32dc32 // fmul v18.4s, v1.4s, v18.4s |
| .long 0x4e20ce14 // fmla v20.4s, v16.4s, v0.4s |
| .long 0x6e32de21 // fmul v1.4s, v17.4s, v18.4s |
| .long 0x6e32de80 // fmul v0.4s, v20.4s, v18.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_linear_gradient_aarch64 |
| .globl _sk_linear_gradient_aarch64 |
| FUNCTION(_sk_linear_gradient_aarch64) |
| _sk_linear_gradient_aarch64: |
| .long 0xf9400029 // ldr x9, [x1] |
| .long 0x91004128 // add x8, x9, #0x10 |
| .long 0x9100512a // add x10, x9, #0x14 |
| .long 0x4d40c910 // ld1r {v16.4s}, [x8] |
| .long 0x91006128 // add x8, x9, #0x18 |
| .long 0x4d40c941 // ld1r {v1.4s}, [x10] |
| .long 0x9100712a // add x10, x9, #0x1c |
| .long 0x4d40c902 // ld1r {v2.4s}, [x8] |
| .long 0xf9400128 // ldr x8, [x9] |
| .long 0x4d40c943 // ld1r {v3.4s}, [x10] |
| .long 0xb40006c8 // cbz x8, 2b14 <sk_linear_gradient_aarch64+0x100> |
| .long 0x6dbf23e9 // stp d9, d8, [sp, #-16]! |
| .long 0xf9400529 // ldr x9, [x9, #8] |
| .long 0x6f00e413 // movi v19.2d, #0x0 |
| .long 0x6f00e411 // movi v17.2d, #0x0 |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x91004129 // add x9, x9, #0x10 |
| .long 0x6f00e414 // movi v20.2d, #0x0 |
| .long 0xd100412a // sub x10, x9, #0x10 |
| .long 0x4d40c955 // ld1r {v21.4s}, [x10] |
| .long 0xd100312b // sub x11, x9, #0xc |
| .long 0xd100212a // sub x10, x9, #0x8 |
| .long 0x4d40c976 // ld1r {v22.4s}, [x11] |
| .long 0xd100112b // sub x11, x9, #0x4 |
| .long 0x4d40c957 // ld1r {v23.4s}, [x10] |
| .long 0xaa0903ea // mov x10, x9 |
| .long 0x4d40c978 // ld1r {v24.4s}, [x11] |
| .long 0x4ddfc959 // ld1r {v25.4s}, [x10], #4 |
| .long 0x9100412b // add x11, x9, #0x10 |
| .long 0x4ea31c7b // mov v27.16b, v3.16b |
| .long 0x6ea0e6a3 // fcmgt v3.4s, v21.4s, v0.4s |
| .long 0x4d40c97a // ld1r {v26.4s}, [x11] |
| .long 0x4eb41e95 // mov v21.16b, v20.16b |
| .long 0x4ea31c74 // mov v20.16b, v3.16b |
| .long 0x9100212b // add x11, x9, #0x8 |
| .long 0x4eb31e69 // mov v9.16b, v19.16b |
| .long 0x4ea31c73 // mov v19.16b, v3.16b |
| .long 0x6e771eb4 // bsl v20.16b, v21.16b, v23.16b |
| .long 0x4d40c975 // ld1r {v21.4s}, [x11] |
| .long 0x9100312b // add x11, x9, #0xc |
| .long 0x6e761d33 // bsl v19.16b, v9.16b, v22.16b |
| .long 0x4d40c976 // ld1r {v22.4s}, [x11] |
| .long 0x4d40c957 // ld1r {v23.4s}, [x10] |
| .long 0x4eb21e5c // mov v28.16b, v18.16b |
| .long 0x4eb11e3d // mov v29.16b, v17.16b |
| .long 0x4eb01e1e // mov v30.16b, v16.16b |
| .long 0x4ea11c3f // mov v31.16b, v1.16b |
| .long 0x4ea21c48 // mov v8.16b, v2.16b |
| .long 0x4ea31c72 // mov v18.16b, v3.16b |
| .long 0x4ea31c71 // mov v17.16b, v3.16b |
| .long 0x4ea31c70 // mov v16.16b, v3.16b |
| .long 0x4ea31c61 // mov v1.16b, v3.16b |
| .long 0x4ea31c62 // mov v2.16b, v3.16b |
| .long 0x6e7a1f63 // bsl v3.16b, v27.16b, v26.16b |
| .long 0x6e781f92 // bsl v18.16b, v28.16b, v24.16b |
| .long 0x6e791fb1 // bsl v17.16b, v29.16b, v25.16b |
| .long 0x6e751fe1 // bsl v1.16b, v31.16b, v21.16b |
| .long 0x6e761d02 // bsl v2.16b, v8.16b, v22.16b |
| .long 0xd1000508 // sub x8, x8, #0x1 |
| .long 0x6e771fd0 // bsl v16.16b, v30.16b, v23.16b |
| .long 0x91009129 // add x9, x9, #0x24 |
| .long 0xb5fffaa8 // cbnz x8, 2a5c <sk_linear_gradient_aarch64+0x48> |
| .long 0x6cc123e9 // ldp d9, d8, [sp], #16 |
| .long 0x14000005 // b 2b24 <sk_linear_gradient_aarch64+0x110> |
| .long 0x6f00e414 // movi v20.2d, #0x0 |
| .long 0x6f00e412 // movi v18.2d, #0x0 |
| .long 0x6f00e411 // movi v17.2d, #0x0 |
| .long 0x6f00e413 // movi v19.2d, #0x0 |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4e20ce70 // fmla v16.4s, v19.4s, v0.4s |
| .long 0x4e20ce81 // fmla v1.4s, v20.4s, v0.4s |
| .long 0x4e20ce42 // fmla v2.4s, v18.4s, v0.4s |
| .long 0x4e20ce23 // fmla v3.4s, v17.4s, v0.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_linear_gradient_2stops_aarch64 |
| .globl _sk_linear_gradient_2stops_aarch64 |
| FUNCTION(_sk_linear_gradient_2stops_aarch64) |
| _sk_linear_gradient_2stops_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0xaa0803e9 // mov x9, x8 |
| .long 0x9100410a // add x10, x8, #0x10 |
| .long 0x4ddfc931 // ld1r {v17.4s}, [x9], #4 |
| .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| .long 0x9100510a // add x10, x8, #0x14 |
| .long 0x4d40c941 // ld1r {v1.4s}, [x10] |
| .long 0x9100610a // add x10, x8, #0x18 |
| .long 0x4d40c942 // ld1r {v2.4s}, [x10] |
| .long 0x9100710a // add x10, x8, #0x1c |
| .long 0x2d414d12 // ldp s18, s19, [x8, #8] |
| .long 0x4d40c943 // ld1r {v3.4s}, [x10] |
| .long 0x4e20ce30 // fmla v16.4s, v17.4s, v0.4s |
| .long 0xbd400131 // ldr s17, [x9] |
| .long 0x4f921002 // fmla v2.4s, v0.4s, v18.s[0] |
| .long 0x4f931003 // fmla v3.4s, v0.4s, v19.s[0] |
| .long 0x4f911001 // fmla v1.4s, v0.4s, v17.s[0] |
| .long 0x4eb01e00 // mov v0.16b, v16.16b |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_save_xy_aarch64 |
| .globl _sk_save_xy_aarch64 |
| FUNCTION(_sk_save_xy_aarch64) |
| _sk_save_xy_aarch64: |
| .long 0x4f0167f0 // movi v16.4s, #0x3f, lsl #24 |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4e30d411 // fadd v17.4s, v0.4s, v16.4s |
| .long 0x4e30d430 // fadd v16.4s, v1.4s, v16.4s |
| .long 0x4e219a32 // frintm v18.4s, v17.4s |
| .long 0x4eb2d631 // fsub v17.4s, v17.4s, v18.4s |
| .long 0x4e219a12 // frintm v18.4s, v16.4s |
| .long 0x4eb2d610 // fsub v16.4s, v16.4s, v18.4s |
| .long 0x3d800100 // str q0, [x8] |
| .long 0x3d800901 // str q1, [x8, #32] |
| .long 0x3d801111 // str q17, [x8, #64] |
| .long 0x3d801910 // str q16, [x8, #96] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_accumulate_aarch64 |
| .globl _sk_accumulate_aarch64 |
| FUNCTION(_sk_accumulate_aarch64) |
| _sk_accumulate_aarch64: |
| .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| .long 0x3dc02110 // ldr q16, [x8, #128] |
| .long 0x3dc02911 // ldr q17, [x8, #160] |
| .long 0x6e31de10 // fmul v16.4s, v16.4s, v17.4s |
| .long 0x4e30cc04 // fmla v4.4s, v0.4s, v16.4s |
| .long 0x4e30cc25 // fmla v5.4s, v1.4s, v16.4s |
| .long 0x4e30cc46 // fmla v6.4s, v2.4s, v16.4s |
| .long 0x4e30cc67 // fmla v7.4s, v3.4s, v16.4s |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bilinear_nx_aarch64 |
| .globl _sk_bilinear_nx_aarch64 |
| FUNCTION(_sk_bilinear_nx_aarch64) |
| _sk_bilinear_nx_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| .long 0x3dc01100 // ldr q0, [x8, #64] |
| .long 0x3dc00110 // ldr q16, [x8] |
| .long 0x4ea0d620 // fsub v0.4s, v17.4s, v0.4s |
| .long 0x3d802100 // str q0, [x8, #128] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f0567e0 // movi v0.4s, #0xbf, lsl #24 |
| .long 0x4e20d600 // fadd v0.4s, v16.4s, v0.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bilinear_px_aarch64 |
| .globl _sk_bilinear_px_aarch64 |
| FUNCTION(_sk_bilinear_px_aarch64) |
| _sk_bilinear_px_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x3dc01100 // ldr q0, [x8, #64] |
| .long 0x3dc00110 // ldr q16, [x8] |
| .long 0x3d802100 // str q0, [x8, #128] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f0167e0 // movi v0.4s, #0x3f, lsl #24 |
| .long 0x4e20d600 // fadd v0.4s, v16.4s, v0.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bilinear_ny_aarch64 |
| .globl _sk_bilinear_ny_aarch64 |
| FUNCTION(_sk_bilinear_ny_aarch64) |
| _sk_bilinear_ny_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| .long 0x3dc01901 // ldr q1, [x8, #96] |
| .long 0x3dc00910 // ldr q16, [x8, #32] |
| .long 0x4ea1d621 // fsub v1.4s, v17.4s, v1.4s |
| .long 0x3d802901 // str q1, [x8, #160] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f0567e1 // movi v1.4s, #0xbf, lsl #24 |
| .long 0x4e21d601 // fadd v1.4s, v16.4s, v1.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bilinear_py_aarch64 |
| .globl _sk_bilinear_py_aarch64 |
| FUNCTION(_sk_bilinear_py_aarch64) |
| _sk_bilinear_py_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x3dc01901 // ldr q1, [x8, #96] |
| .long 0x3dc00910 // ldr q16, [x8, #32] |
| .long 0x3d802901 // str q1, [x8, #160] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f0167e1 // movi v1.4s, #0x3f, lsl #24 |
| .long 0x4e21d601 // fadd v1.4s, v16.4s, v1.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_n3x_aarch64 |
| .globl _sk_bicubic_n3x_aarch64 |
| FUNCTION(_sk_bicubic_n3x_aarch64) |
| _sk_bicubic_n3x_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52a7d8e9 // mov w9, #0x3ec70000 |
| .long 0x72838e49 // movk w9, #0x1c72 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x3dc01111 // ldr q17, [x8, #64] |
| .long 0x52b7d549 // mov w9, #0xbeaa0000 |
| .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00 |
| .long 0x72955569 // movk w9, #0xaaab |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s |
| .long 0x6e20dc11 // fmul v17.4s, v0.4s, v0.4s |
| .long 0x4e20ce12 // fmla v18.4s, v16.4s, v0.4s |
| .long 0x6e32de20 // fmul v0.4s, v17.4s, v18.4s |
| .long 0x3dc00113 // ldr q19, [x8] |
| .long 0x3d802100 // str q0, [x8, #128] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f07f700 // fmov v0.4s, #-1.500000000000000000e+00 |
| .long 0x4e20d660 // fadd v0.4s, v19.4s, v0.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_n1x_aarch64 |
| .globl _sk_bicubic_n1x_aarch64 |
| FUNCTION(_sk_bicubic_n1x_aarch64) |
| _sk_bicubic_n1x_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52b7f2a9 // mov w9, #0xbf950000 |
| .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00 |
| .long 0x728aaaa9 // movk w9, #0x5555 |
| .long 0x3dc01110 // ldr q16, [x8, #64] |
| .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00 |
| .long 0x4f0167f2 // movi v18.4s, #0x3f, lsl #24 |
| .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x52a7ac69 // mov w9, #0x3d630000 |
| .long 0x7291c729 // movk w9, #0x8e39 |
| .long 0x4e20ce11 // fmla v17.4s, v16.4s, v0.4s |
| .long 0x4e20ce32 // fmla v18.4s, v17.4s, v0.4s |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x4e20ce51 // fmla v17.4s, v18.4s, v0.4s |
| .long 0x3dc00110 // ldr q16, [x8] |
| .long 0x3d802111 // str q17, [x8, #128] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f0567e0 // movi v0.4s, #0xbf, lsl #24 |
| .long 0x4e20d600 // fadd v0.4s, v16.4s, v0.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_p1x_aarch64 |
| .globl _sk_bicubic_p1x_aarch64 |
| FUNCTION(_sk_bicubic_p1x_aarch64) |
| _sk_bicubic_p1x_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52b7f2a9 // mov w9, #0xbf950000 |
| .long 0x728aaaa9 // movk w9, #0x5555 |
| .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00 |
| .long 0x3dc01112 // ldr q18, [x8, #64] |
| .long 0x3dc00100 // ldr q0, [x8] |
| .long 0x4e040d33 // dup v19.4s, w9 |
| .long 0x52a7ac69 // mov w9, #0x3d630000 |
| .long 0x4f0167f0 // movi v16.4s, #0x3f, lsl #24 |
| .long 0x7291c729 // movk w9, #0x8e39 |
| .long 0x4e32ce71 // fmla v17.4s, v19.4s, v18.4s |
| .long 0x4e30d400 // fadd v0.4s, v0.4s, v16.4s |
| .long 0x4e32ce30 // fmla v16.4s, v17.4s, v18.4s |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x4e32ce11 // fmla v17.4s, v16.4s, v18.4s |
| .long 0x3d802111 // str q17, [x8, #128] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_p3x_aarch64 |
| .globl _sk_bicubic_p3x_aarch64 |
| FUNCTION(_sk_bicubic_p3x_aarch64) |
| _sk_bicubic_p3x_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52a7d8e9 // mov w9, #0x3ec70000 |
| .long 0x72838e49 // movk w9, #0x1c72 |
| .long 0x4e040d20 // dup v0.4s, w9 |
| .long 0x3dc01110 // ldr q16, [x8, #64] |
| .long 0x52b7d549 // mov w9, #0xbeaa0000 |
| .long 0x72955569 // movk w9, #0xaaab |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x6e30de13 // fmul v19.4s, v16.4s, v16.4s |
| .long 0x4e30cc11 // fmla v17.4s, v0.4s, v16.4s |
| .long 0x6e31de60 // fmul v0.4s, v19.4s, v17.4s |
| .long 0x3dc00112 // ldr q18, [x8] |
| .long 0x3d802100 // str q0, [x8, #128] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f03f700 // fmov v0.4s, #1.500000000000000000e+00 |
| .long 0x4e20d640 // fadd v0.4s, v18.4s, v0.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_n3y_aarch64 |
| .globl _sk_bicubic_n3y_aarch64 |
| FUNCTION(_sk_bicubic_n3y_aarch64) |
| _sk_bicubic_n3y_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52a7d8e9 // mov w9, #0x3ec70000 |
| .long 0x72838e49 // movk w9, #0x1c72 |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x3dc01911 // ldr q17, [x8, #96] |
| .long 0x52b7d549 // mov w9, #0xbeaa0000 |
| .long 0x4f03f601 // fmov v1.4s, #1.000000000000000000e+00 |
| .long 0x72955569 // movk w9, #0xaaab |
| .long 0x4e040d32 // dup v18.4s, w9 |
| .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s |
| .long 0x6e21dc31 // fmul v17.4s, v1.4s, v1.4s |
| .long 0x4e21ce12 // fmla v18.4s, v16.4s, v1.4s |
| .long 0x6e32de21 // fmul v1.4s, v17.4s, v18.4s |
| .long 0x3dc00913 // ldr q19, [x8, #32] |
| .long 0x3d802901 // str q1, [x8, #160] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f07f701 // fmov v1.4s, #-1.500000000000000000e+00 |
| .long 0x4e21d661 // fadd v1.4s, v19.4s, v1.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_n1y_aarch64 |
| .globl _sk_bicubic_n1y_aarch64 |
| FUNCTION(_sk_bicubic_n1y_aarch64) |
| _sk_bicubic_n1y_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52b7f2a9 // mov w9, #0xbf950000 |
| .long 0x4f03f601 // fmov v1.4s, #1.000000000000000000e+00 |
| .long 0x728aaaa9 // movk w9, #0x5555 |
| .long 0x3dc01910 // ldr q16, [x8, #96] |
| .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00 |
| .long 0x4f0167f2 // movi v18.4s, #0x3f, lsl #24 |
| .long 0x4eb0d421 // fsub v1.4s, v1.4s, v16.4s |
| .long 0x4e040d30 // dup v16.4s, w9 |
| .long 0x52a7ac69 // mov w9, #0x3d630000 |
| .long 0x7291c729 // movk w9, #0x8e39 |
| .long 0x4e21ce11 // fmla v17.4s, v16.4s, v1.4s |
| .long 0x4e21ce32 // fmla v18.4s, v17.4s, v1.4s |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x4e21ce51 // fmla v17.4s, v18.4s, v1.4s |
| .long 0x3dc00910 // ldr q16, [x8, #32] |
| .long 0x3d802911 // str q17, [x8, #160] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f0567e1 // movi v1.4s, #0xbf, lsl #24 |
| .long 0x4e21d601 // fadd v1.4s, v16.4s, v1.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_p1y_aarch64 |
| .globl _sk_bicubic_p1y_aarch64 |
| FUNCTION(_sk_bicubic_p1y_aarch64) |
| _sk_bicubic_p1y_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52b7f2a9 // mov w9, #0xbf950000 |
| .long 0x728aaaa9 // movk w9, #0x5555 |
| .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00 |
| .long 0x3dc01912 // ldr q18, [x8, #96] |
| .long 0x3dc00901 // ldr q1, [x8, #32] |
| .long 0x4e040d33 // dup v19.4s, w9 |
| .long 0x52a7ac69 // mov w9, #0x3d630000 |
| .long 0x4f0167f0 // movi v16.4s, #0x3f, lsl #24 |
| .long 0x7291c729 // movk w9, #0x8e39 |
| .long 0x4e32ce71 // fmla v17.4s, v19.4s, v18.4s |
| .long 0x4e30d421 // fadd v1.4s, v1.4s, v16.4s |
| .long 0x4e32ce30 // fmla v16.4s, v17.4s, v18.4s |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x4e32ce11 // fmla v17.4s, v16.4s, v18.4s |
| .long 0x3d802911 // str q17, [x8, #160] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_bicubic_p3y_aarch64 |
| .globl _sk_bicubic_p3y_aarch64 |
| FUNCTION(_sk_bicubic_p3y_aarch64) |
| _sk_bicubic_p3y_aarch64: |
| .long 0xf9400028 // ldr x8, [x1] |
| .long 0x52a7d8e9 // mov w9, #0x3ec70000 |
| .long 0x72838e49 // movk w9, #0x1c72 |
| .long 0x4e040d21 // dup v1.4s, w9 |
| .long 0x3dc01910 // ldr q16, [x8, #96] |
| .long 0x52b7d549 // mov w9, #0xbeaa0000 |
| .long 0x72955569 // movk w9, #0xaaab |
| .long 0x4e040d31 // dup v17.4s, w9 |
| .long 0x6e30de13 // fmul v19.4s, v16.4s, v16.4s |
| .long 0x4e30cc31 // fmla v17.4s, v1.4s, v16.4s |
| .long 0x6e31de61 // fmul v1.4s, v19.4s, v17.4s |
| .long 0x3dc00912 // ldr q18, [x8, #32] |
| .long 0x3d802901 // str q1, [x8, #160] |
| .long 0xf9400423 // ldr x3, [x1, #8] |
| .long 0x4f03f701 // fmov v1.4s, #1.500000000000000000e+00 |
| .long 0x4e21d641 // fadd v1.4s, v18.4s, v1.4s |
| .long 0x91004021 // add x1, x1, #0x10 |
| .long 0xd61f0060 // br x3 |
| |
| HIDDEN _sk_callback_aarch64 |
| .globl _sk_callback_aarch64 |
| FUNCTION(_sk_callback_aarch64) |
| _sk_callback_aarch64: |
| .long 0xd10283ff // sub sp, sp, #0xa0 |
| .long 0xa90853f5 // stp x21, x20, [sp, #128] |
| .long 0xa9097bf3 // stp x19, x30, [sp, #144] |
| .long 0xad031fe6 // stp q6, q7, [sp, #96] |
| .long 0xad0217e4 // stp q4, q5, [sp, #64] |
| .long 0xad010fe2 // stp q2, q3, [sp, #32] |
| .long 0xad0007e0 // stp q0, q1, [sp] |
| .long 0xaa0103f4 // mov x20, x1 |
| .long 0xf9400288 // ldr x8, [x20] |
| .long 0xaa0003f5 // mov x21, x0 |
| .long 0x321e03e1 // orr w1, wzr, #0x4 |
| .long 0xaa0203f3 // mov x19, x2 |
| .long 0xa9402109 // ldp x9, x8, [x8] |
| .long 0xaa0803e0 // mov x0, x8 |
| .long 0xd63f0120 // blr x9 |
| .long 0xf9400683 // ldr x3, [x20, #8] |
| .long 0x91004281 // add x1, x20, #0x10 |
| .long 0xaa1503e0 // mov x0, x21 |
| .long 0xaa1303e2 // mov x2, x19 |
| .long 0xad4007e0 // ldp q0, q1, [sp] |
| .long 0xad410fe2 // ldp q2, q3, [sp, #32] |
| .long 0xad4217e4 // ldp q4, q5, [sp, #64] |
| .long 0xad431fe6 // ldp q6, q7, [sp, #96] |
| .long 0xa9497bf3 // ldp x19, x30, [sp, #144] |
| .long 0xa94853f5 // ldp x21, x20, [sp, #128] |
| .long 0x910283ff // add sp, sp, #0xa0 |
| .long 0xd61f0060 // br x3 |
| #elif defined(__arm__) |
| .balign 4 |
| |
| HIDDEN _sk_start_pipeline_vfp4 |
| .globl _sk_start_pipeline_vfp4 |
| FUNCTION(_sk_start_pipeline_vfp4) |
| _sk_start_pipeline_vfp4: |
| .long 0xe92d41f0 // push {r4, r5, r6, r7, r8, lr} |
| .long 0xe1a04000 // mov r4, r0 |
| .long 0xe2840002 // add r0, r4, #2 |
| .long 0xe1a05003 // mov r5, r3 |
| .long 0xe1a08002 // mov r8, r2 |
| .long 0xe1a07001 // mov r7, r1 |
| .long 0xe1500005 // cmp r0, r5 |
| .long 0x8a000010 // bhi 64 <sk_start_pipeline_vfp4+0x64> |
| .long 0xe4976004 // ldr r6, [r7], #4 |
| .long 0xf2800010 // vmov.i32 d0, #0 |
| .long 0xe1a00004 // mov r0, r4 |
| .long 0xf2801010 // vmov.i32 d1, #0 |
| .long 0xe1a01007 // mov r1, r7 |
| .long 0xf2802010 // vmov.i32 d2, #0 |
| .long 0xe1a02008 // mov r2, r8 |
| .long 0xf2803010 // vmov.i32 d3, #0 |
| .long 0xf2804010 // vmov.i32 d4, #0 |
| .long 0xf2805010 // vmov.i32 d5, #0 |
| .long 0xf2806010 // vmov.i32 d6, #0 |
| .long 0xf2807010 // vmov.i32 d7, #0 |
| .long 0xe12fff36 // blx r6 |
| .long 0xe2840004 // add r0, r4, #4 |
| .long 0xe2844002 // add r4, r4, #2 |
| .long 0xe1500005 // cmp r0, r5 |
| .long 0x9affffef // bls 24 <sk_start_pipeline_vfp4+0x24> |
| .long 0xe1a00004 // mov r0, r4 |
| .long 0xe8bd81f0 // pop {r4, r5, r6, r7, r8, pc} |
| |
| HIDDEN _sk_just_return_vfp4 |
| .globl _sk_just_return_vfp4 |
| FUNCTION(_sk_just_return_vfp4) |
| _sk_just_return_vfp4: |
| .long 0xe12fff1e // bx lr |
| |
| HIDDEN _sk_seed_shader_vfp4 |
| .globl _sk_seed_shader_vfp4 |
| FUNCTION(_sk_seed_shader_vfp4) |
| _sk_seed_shader_vfp4: |
| .long 0xee800b90 // vdup.32 d16, r0 |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xedd23b00 // vldr d19, [r2] |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf2872f10 // vmov.f32 d2, #1 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2400da1 // vadd.f32 d16, d16, d17 |
| .long 0xf2803010 // vmov.i32 d3, #0 |
| .long 0xf2804010 // vmov.i32 d4, #0 |
| .long 0xf2021da1 // vadd.f32 d1, d18, d17 |
| .long 0xf2000da3 // vadd.f32 d0, d16, d19 |
| .long 0xf2805010 // vmov.i32 d5, #0 |
| .long 0xf2806010 // vmov.i32 d6, #0 |
| .long 0xf2807010 // vmov.i32 d7, #0 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_constant_color_vfp4 |
| .globl _sk_constant_color_vfp4 |
| FUNCTION(_sk_constant_color_vfp4) |
| _sk_constant_color_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe283400c // add r4, r3, #12 |
| .long 0xe1a0e003 // mov lr, r3 |
| .long 0xe2833008 // add r3, r3, #8 |
| .long 0xf4ae0c9d // vld1.32 {d0[]}, [lr :32]! |
| .long 0xf4a43c9f // vld1.32 {d3[]}, [r4 :32] |
| .long 0xf4a32c9f // vld1.32 {d2[]}, [r3 :32] |
| .long 0xf4ae1c9f // vld1.32 {d1[]}, [lr :32] |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_clear_vfp4 |
| .globl _sk_clear_vfp4 |
| FUNCTION(_sk_clear_vfp4) |
| _sk_clear_vfp4: |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2800010 // vmov.i32 d0, #0 |
| .long 0xf2801010 // vmov.i32 d1, #0 |
| .long 0xf2802010 // vmov.i32 d2, #0 |
| .long 0xf2803010 // vmov.i32 d3, #0 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_srcatop_vfp4 |
| .globl _sk_srcatop_vfp4 |
| FUNCTION(_sk_srcatop_vfp4) |
| _sk_srcatop_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2603d83 // vsub.f32 d19, d16, d3 |
| .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| .long 0xf3430d94 // vmul.f32 d16, d19, d4 |
| .long 0xf3431d95 // vmul.f32 d17, d19, d5 |
| .long 0xf3432d96 // vmul.f32 d18, d19, d6 |
| .long 0xf2400c17 // vfma.f32 d16, d0, d7 |
| .long 0xf2411c17 // vfma.f32 d17, d1, d7 |
| .long 0xf2422c17 // vfma.f32 d18, d2, d7 |
| .long 0xf2033c97 // vfma.f32 d3, d19, d7 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_dstatop_vfp4 |
| .globl _sk_dstatop_vfp4 |
| FUNCTION(_sk_dstatop_vfp4) |
| _sk_dstatop_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3431d15 // vmul.f32 d17, d3, d5 |
| .long 0xf2604d87 // vsub.f32 d20, d16, d7 |
| .long 0xf3430d14 // vmul.f32 d16, d3, d4 |
| .long 0xf3432d16 // vmul.f32 d18, d3, d6 |
| .long 0xf3433d17 // vmul.f32 d19, d3, d7 |
| .long 0xf2440c90 // vfma.f32 d16, d20, d0 |
| .long 0xf2441c91 // vfma.f32 d17, d20, d1 |
| .long 0xf2442c92 // vfma.f32 d18, d20, d2 |
| .long 0xf2443c93 // vfma.f32 d19, d20, d3 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xf22331b3 // vorr d3, d19, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_srcin_vfp4 |
| .globl _sk_srcin_vfp4 |
| FUNCTION(_sk_srcin_vfp4) |
| _sk_srcin_vfp4: |
| .long 0xf3000d17 // vmul.f32 d0, d0, d7 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3011d17 // vmul.f32 d1, d1, d7 |
| .long 0xf3022d17 // vmul.f32 d2, d2, d7 |
| .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_dstin_vfp4 |
| .globl _sk_dstin_vfp4 |
| FUNCTION(_sk_dstin_vfp4) |
| _sk_dstin_vfp4: |
| .long 0xf3030d14 // vmul.f32 d0, d3, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3031d15 // vmul.f32 d1, d3, d5 |
| .long 0xf3032d16 // vmul.f32 d2, d3, d6 |
| .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_srcout_vfp4 |
| .globl _sk_srcout_vfp4 |
| FUNCTION(_sk_srcout_vfp4) |
| _sk_srcout_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2600d87 // vsub.f32 d16, d16, d7 |
| .long 0xf3000d90 // vmul.f32 d0, d16, d0 |
| .long 0xf3001d91 // vmul.f32 d1, d16, d1 |
| .long 0xf3002d92 // vmul.f32 d2, d16, d2 |
| .long 0xf3003d93 // vmul.f32 d3, d16, d3 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_dstout_vfp4 |
| .globl _sk_dstout_vfp4 |
| FUNCTION(_sk_dstout_vfp4) |
| _sk_dstout_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf3000d94 // vmul.f32 d0, d16, d4 |
| .long 0xf3001d95 // vmul.f32 d1, d16, d5 |
| .long 0xf3002d96 // vmul.f32 d2, d16, d6 |
| .long 0xf3003d97 // vmul.f32 d3, d16, d7 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_srcover_vfp4 |
| .globl _sk_srcover_vfp4 |
| FUNCTION(_sk_srcover_vfp4) |
| _sk_srcover_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf2040c30 // vfma.f32 d0, d4, d16 |
| .long 0xf2051c30 // vfma.f32 d1, d5, d16 |
| .long 0xf2062c30 // vfma.f32 d2, d6, d16 |
| .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_dstover_vfp4 |
| .globl _sk_dstover_vfp4 |
| FUNCTION(_sk_dstover_vfp4) |
| _sk_dstover_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2651115 // vorr d17, d5, d5 |
| .long 0xf2604d87 // vsub.f32 d20, d16, d7 |
| .long 0xf2640114 // vorr d16, d4, d4 |
| .long 0xf2662116 // vorr d18, d6, d6 |
| .long 0xf2673117 // vorr d19, d7, d7 |
| .long 0xf2400c34 // vfma.f32 d16, d0, d20 |
| .long 0xf2411c34 // vfma.f32 d17, d1, d20 |
| .long 0xf2422c34 // vfma.f32 d18, d2, d20 |
| .long 0xf2433c34 // vfma.f32 d19, d3, d20 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xf22331b3 // vorr d3, d19, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_modulate_vfp4 |
| .globl _sk_modulate_vfp4 |
| FUNCTION(_sk_modulate_vfp4) |
| _sk_modulate_vfp4: |
| .long 0xf3000d14 // vmul.f32 d0, d0, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3011d15 // vmul.f32 d1, d1, d5 |
| .long 0xf3022d16 // vmul.f32 d2, d2, d6 |
| .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_multiply_vfp4 |
| .globl _sk_multiply_vfp4 |
| FUNCTION(_sk_multiply_vfp4) |
| _sk_multiply_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2601d87 // vsub.f32 d17, d16, d7 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf3412d90 // vmul.f32 d18, d17, d0 |
| .long 0xf3403d94 // vmul.f32 d19, d16, d4 |
| .long 0xf3414d91 // vmul.f32 d20, d17, d1 |
| .long 0xf3405d95 // vmul.f32 d21, d16, d5 |
| .long 0xf3416d92 // vmul.f32 d22, d17, d2 |
| .long 0xf3418d93 // vmul.f32 d24, d17, d3 |
| .long 0xf3407d96 // vmul.f32 d23, d16, d6 |
| .long 0xf3409d97 // vmul.f32 d25, d16, d7 |
| .long 0xf2430da2 // vadd.f32 d16, d19, d18 |
| .long 0xf2451da4 // vadd.f32 d17, d21, d20 |
| .long 0xf2472da6 // vadd.f32 d18, d23, d22 |
| .long 0xf2493da8 // vadd.f32 d19, d25, d24 |
| .long 0xf2400c14 // vfma.f32 d16, d0, d4 |
| .long 0xf2411c15 // vfma.f32 d17, d1, d5 |
| .long 0xf2422c16 // vfma.f32 d18, d2, d6 |
| .long 0xf2433c17 // vfma.f32 d19, d3, d7 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xf22331b3 // vorr d3, d19, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_plus__vfp4 |
| .globl _sk_plus__vfp4 |
| FUNCTION(_sk_plus__vfp4) |
| _sk_plus__vfp4: |
| .long 0xf2000d04 // vadd.f32 d0, d0, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2011d05 // vadd.f32 d1, d1, d5 |
| .long 0xf2022d06 // vadd.f32 d2, d2, d6 |
| .long 0xf2033d07 // vadd.f32 d3, d3, d7 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_screen_vfp4 |
| .globl _sk_screen_vfp4 |
| FUNCTION(_sk_screen_vfp4) |
| _sk_screen_vfp4: |
| .long 0xf2400d04 // vadd.f32 d16, d0, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2411d05 // vadd.f32 d17, d1, d5 |
| .long 0xf2422d06 // vadd.f32 d18, d2, d6 |
| .long 0xf2433d07 // vadd.f32 d19, d3, d7 |
| .long 0xf2600c14 // vfms.f32 d16, d0, d4 |
| .long 0xf2611c15 // vfms.f32 d17, d1, d5 |
| .long 0xf2622c16 // vfms.f32 d18, d2, d6 |
| .long 0xf2633c17 // vfms.f32 d19, d3, d7 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xf22331b3 // vorr d3, d19, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_xor__vfp4 |
| .globl _sk_xor__vfp4 |
| FUNCTION(_sk_xor__vfp4) |
| _sk_xor__vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2603d83 // vsub.f32 d19, d16, d3 |
| .long 0xf2604d87 // vsub.f32 d20, d16, d7 |
| .long 0xf3430d94 // vmul.f32 d16, d19, d4 |
| .long 0xf3431d95 // vmul.f32 d17, d19, d5 |
| .long 0xf3432d96 // vmul.f32 d18, d19, d6 |
| .long 0xf3433d97 // vmul.f32 d19, d19, d7 |
| .long 0xf2440c90 // vfma.f32 d16, d20, d0 |
| .long 0xf2441c91 // vfma.f32 d17, d20, d1 |
| .long 0xf2442c92 // vfma.f32 d18, d20, d2 |
| .long 0xf2443c93 // vfma.f32 d19, d20, d3 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xf22331b3 // vorr d3, d19, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_darken_vfp4 |
| .globl _sk_darken_vfp4 |
| FUNCTION(_sk_darken_vfp4) |
| _sk_darken_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3431d14 // vmul.f32 d17, d3, d4 |
| .long 0xf3402d17 // vmul.f32 d18, d0, d7 |
| .long 0xf3433d15 // vmul.f32 d19, d3, d5 |
| .long 0xf3414d17 // vmul.f32 d20, d1, d7 |
| .long 0xf3435d16 // vmul.f32 d21, d3, d6 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf3426d17 // vmul.f32 d22, d2, d7 |
| .long 0xf2421fa1 // vmax.f32 d17, d18, d17 |
| .long 0xf2407d04 // vadd.f32 d23, d0, d4 |
| .long 0xf2443fa3 // vmax.f32 d19, d20, d19 |
| .long 0xf2412d05 // vadd.f32 d18, d1, d5 |
| .long 0xf2424d06 // vadd.f32 d20, d2, d6 |
| .long 0xf2465fa5 // vmax.f32 d21, d22, d21 |
| .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| .long 0xf2270da1 // vsub.f32 d0, d23, d17 |
| .long 0xf2221da3 // vsub.f32 d1, d18, d19 |
| .long 0xf2242da5 // vsub.f32 d2, d20, d21 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_lighten_vfp4 |
| .globl _sk_lighten_vfp4 |
| FUNCTION(_sk_lighten_vfp4) |
| _sk_lighten_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3431d14 // vmul.f32 d17, d3, d4 |
| .long 0xf3402d17 // vmul.f32 d18, d0, d7 |
| .long 0xf3433d15 // vmul.f32 d19, d3, d5 |
| .long 0xf3414d17 // vmul.f32 d20, d1, d7 |
| .long 0xf3435d16 // vmul.f32 d21, d3, d6 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf3426d17 // vmul.f32 d22, d2, d7 |
| .long 0xf2621fa1 // vmin.f32 d17, d18, d17 |
| .long 0xf2407d04 // vadd.f32 d23, d0, d4 |
| .long 0xf2643fa3 // vmin.f32 d19, d20, d19 |
| .long 0xf2412d05 // vadd.f32 d18, d1, d5 |
| .long 0xf2424d06 // vadd.f32 d20, d2, d6 |
| .long 0xf2665fa5 // vmin.f32 d21, d22, d21 |
| .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| .long 0xf2270da1 // vsub.f32 d0, d23, d17 |
| .long 0xf2221da3 // vsub.f32 d1, d18, d19 |
| .long 0xf2242da5 // vsub.f32 d2, d20, d21 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_difference_vfp4 |
| .globl _sk_difference_vfp4 |
| FUNCTION(_sk_difference_vfp4) |
| _sk_difference_vfp4: |
| .long 0xf3430d14 // vmul.f32 d16, d3, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3401d17 // vmul.f32 d17, d0, d7 |
| .long 0xf3432d15 // vmul.f32 d18, d3, d5 |
| .long 0xf3413d17 // vmul.f32 d19, d1, d7 |
| .long 0xf3434d16 // vmul.f32 d20, d3, d6 |
| .long 0xf3425d17 // vmul.f32 d21, d2, d7 |
| .long 0xf2c76f10 // vmov.f32 d22, #1 |
| .long 0xf2610fa0 // vmin.f32 d16, d17, d16 |
| .long 0xf2631fa2 // vmin.f32 d17, d19, d18 |
| .long 0xf2662d83 // vsub.f32 d18, d22, d3 |
| .long 0xf2653fa4 // vmin.f32 d19, d21, d20 |
| .long 0xf2404d04 // vadd.f32 d20, d0, d4 |
| .long 0xf2400da0 // vadd.f32 d16, d16, d16 |
| .long 0xf2073c32 // vfma.f32 d3, d7, d18 |
| .long 0xf2415d05 // vadd.f32 d21, d1, d5 |
| .long 0xf2411da1 // vadd.f32 d17, d17, d17 |
| .long 0xf2426d06 // vadd.f32 d22, d2, d6 |
| .long 0xf2432da3 // vadd.f32 d18, d19, d19 |
| .long 0xf2240da0 // vsub.f32 d0, d20, d16 |
| .long 0xf2251da1 // vsub.f32 d1, d21, d17 |
| .long 0xf2262da2 // vsub.f32 d2, d22, d18 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_exclusion_vfp4 |
| .globl _sk_exclusion_vfp4 |
| FUNCTION(_sk_exclusion_vfp4) |
| _sk_exclusion_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3401d14 // vmul.f32 d17, d0, d4 |
| .long 0xf3412d15 // vmul.f32 d18, d1, d5 |
| .long 0xf3423d16 // vmul.f32 d19, d2, d6 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf2404d04 // vadd.f32 d20, d0, d4 |
| .long 0xf2411da1 // vadd.f32 d17, d17, d17 |
| .long 0xf2415d05 // vadd.f32 d21, d1, d5 |
| .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| .long 0xf2426d06 // vadd.f32 d22, d2, d6 |
| .long 0xf2433da3 // vadd.f32 d19, d19, d19 |
| .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| .long 0xf2240da1 // vsub.f32 d0, d20, d17 |
| .long 0xf2251da2 // vsub.f32 d1, d21, d18 |
| .long 0xf2262da3 // vsub.f32 d2, d22, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_colorburn_vfp4 |
| .globl _sk_colorburn_vfp4 |
| FUNCTION(_sk_colorburn_vfp4) |
| _sk_colorburn_vfp4: |
| .long 0xed2d8b08 // vpush {d8-d11} |
| .long 0xf2670d04 // vsub.f32 d16, d7, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2671d06 // vsub.f32 d17, d7, d6 |
| .long 0xf2672d05 // vsub.f32 d18, d7, d5 |
| .long 0xf3008d93 // vmul.f32 d8, d16, d3 |
| .long 0xf3019d93 // vmul.f32 d9, d17, d3 |
| .long 0xf302ad93 // vmul.f32 d10, d18, d3 |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xeec8baa0 // vdiv.f32 s23, s17, s1 |
| .long 0xee88ba00 // vdiv.f32 s22, s16, s0 |
| .long 0xeec98aa2 // vdiv.f32 s17, s19, s5 |
| .long 0xee898a02 // vdiv.f32 s16, s18, s4 |
| .long 0xeeca9aa1 // vdiv.f32 s19, s21, s3 |
| .long 0xee8a9a01 // vdiv.f32 s18, s20, s2 |
| .long 0xf2672f08 // vmin.f32 d18, d7, d8 |
| .long 0xf2673f09 // vmin.f32 d19, d7, d9 |
| .long 0xf2670f0b // vmin.f32 d16, d7, d11 |
| .long 0xf2614d87 // vsub.f32 d20, d17, d7 |
| .long 0xf2672d22 // vsub.f32 d18, d7, d18 |
| .long 0xf2673d23 // vsub.f32 d19, d7, d19 |
| .long 0xf2611d83 // vsub.f32 d17, d17, d3 |
| .long 0xf2670d20 // vsub.f32 d16, d7, d16 |
| .long 0xf3445d90 // vmul.f32 d21, d20, d0 |
| .long 0xf3446d92 // vmul.f32 d22, d20, d2 |
| .long 0xf3422d93 // vmul.f32 d18, d18, d3 |
| .long 0xf3444d91 // vmul.f32 d20, d20, d1 |
| .long 0xf3433d93 // vmul.f32 d19, d19, d3 |
| .long 0xf3400d93 // vmul.f32 d16, d16, d3 |
| .long 0xf3417d95 // vmul.f32 d23, d17, d5 |
| .long 0xf3418d94 // vmul.f32 d24, d17, d4 |
| .long 0xf3419d96 // vmul.f32 d25, d17, d6 |
| .long 0xf2443da3 // vadd.f32 d19, d20, d19 |
| .long 0xf2462da2 // vadd.f32 d18, d22, d18 |
| .long 0xf245ada0 // vadd.f32 d26, d21, d16 |
| .long 0xf247bd81 // vadd.f32 d27, d23, d1 |
| .long 0xf248cd80 // vadd.f32 d28, d24, d0 |
| .long 0xf249dd82 // vadd.f32 d29, d25, d2 |
| .long 0xf2073c31 // vfma.f32 d3, d7, d17 |
| .long 0xf2499da2 // vadd.f32 d25, d25, d18 |
| .long 0xf2473da3 // vadd.f32 d19, d23, d19 |
| .long 0xf3f97501 // vceq.f32 d23, d1, #0 |
| .long 0xf2455d84 // vadd.f32 d21, d21, d4 |
| .long 0xf2444d85 // vadd.f32 d20, d20, d5 |
| .long 0xf2440e07 // vceq.f32 d16, d4, d7 |
| .long 0xf2466d86 // vadd.f32 d22, d22, d6 |
| .long 0xf2451e07 // vceq.f32 d17, d5, d7 |
| .long 0xf2462e07 // vceq.f32 d18, d6, d7 |
| .long 0xf35b71b3 // vbsl d23, d27, d19 |
| .long 0xf3f93500 // vceq.f32 d19, d0, #0 |
| .long 0xf2488daa // vadd.f32 d24, d24, d26 |
| .long 0xf35c31b8 // vbsl d19, d28, d24 |
| .long 0xf3f98502 // vceq.f32 d24, d2, #0 |
| .long 0xf35d81b9 // vbsl d24, d29, d25 |
| .long 0xf35501b3 // vbsl d16, d21, d19 |
| .long 0xf35411b7 // vbsl d17, d20, d23 |
| .long 0xf35621b8 // vbsl d18, d22, d24 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xecbd8b08 // vpop {d8-d11} |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_colordodge_vfp4 |
| .globl _sk_colordodge_vfp4 |
| FUNCTION(_sk_colordodge_vfp4) |
| _sk_colordodge_vfp4: |
| .long 0xed2d8b0e // vpush {d8-d14} |
| .long 0xf2238d02 // vsub.f32 d8, d3, d2 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3039d16 // vmul.f32 d9, d3, d6 |
| .long 0xf223ad01 // vsub.f32 d10, d3, d1 |
| .long 0xf303bd15 // vmul.f32 d11, d3, d5 |
| .long 0xf223cd00 // vsub.f32 d12, d3, d0 |
| .long 0xf303dd14 // vmul.f32 d13, d3, d4 |
| .long 0xeec9eaa8 // vdiv.f32 s29, s19, s17 |
| .long 0xee89ea08 // vdiv.f32 s28, s18, s16 |
| .long 0xeecb8aaa // vdiv.f32 s17, s23, s21 |
| .long 0xeecd9aac // vdiv.f32 s19, s27, s25 |
| .long 0xee8b8a0a // vdiv.f32 s16, s22, s20 |
| .long 0xee8d9a0c // vdiv.f32 s18, s26, s24 |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xf2672f0e // vmin.f32 d18, d7, d14 |
| .long 0xf2601d87 // vsub.f32 d17, d16, d7 |
| .long 0xf2673f08 // vmin.f32 d19, d7, d8 |
| .long 0xf2674f09 // vmin.f32 d20, d7, d9 |
| .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| .long 0xf3415d92 // vmul.f32 d21, d17, d2 |
| .long 0xf3422d93 // vmul.f32 d18, d18, d3 |
| .long 0xf3416d91 // vmul.f32 d22, d17, d1 |
| .long 0xf3433d93 // vmul.f32 d19, d19, d3 |
| .long 0xf3411d90 // vmul.f32 d17, d17, d0 |
| .long 0xf3444d93 // vmul.f32 d20, d20, d3 |
| .long 0xf3407d95 // vmul.f32 d23, d16, d5 |
| .long 0xf3408d94 // vmul.f32 d24, d16, d4 |
| .long 0xf3409d96 // vmul.f32 d25, d16, d6 |
| .long 0xf2452da2 // vadd.f32 d18, d21, d18 |
| .long 0xf2463da3 // vadd.f32 d19, d22, d19 |
| .long 0xf2414da4 // vadd.f32 d20, d17, d20 |
| .long 0xf241ae03 // vceq.f32 d26, d1, d3 |
| .long 0xf247bd81 // vadd.f32 d27, d23, d1 |
| .long 0xf3b91505 // vceq.f32 d1, d5, #0 |
| .long 0xf240ce03 // vceq.f32 d28, d0, d3 |
| .long 0xf248dd80 // vadd.f32 d29, d24, d0 |
| .long 0xf3b90504 // vceq.f32 d0, d4, #0 |
| .long 0xf242ee03 // vceq.f32 d30, d2, d3 |
| .long 0xf249fd82 // vadd.f32 d31, d25, d2 |
| .long 0xf3b92506 // vceq.f32 d2, d6, #0 |
| .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| .long 0xf2410d84 // vadd.f32 d16, d17, d4 |
| .long 0xf2491da2 // vadd.f32 d17, d25, d18 |
| .long 0xf2462d85 // vadd.f32 d18, d22, d5 |
| .long 0xf2455d86 // vadd.f32 d21, d21, d6 |
| .long 0xf2473da3 // vadd.f32 d19, d23, d19 |
| .long 0xf2484da4 // vadd.f32 d20, d24, d20 |
| .long 0xf35fe1b1 // vbsl d30, d31, d17 |
| .long 0xf35ba1b3 // vbsl d26, d27, d19 |
| .long 0xf35dc1b4 // vbsl d28, d29, d20 |
| .long 0xf31001bc // vbsl d0, d16, d28 |
| .long 0xf31211ba // vbsl d1, d18, d26 |
| .long 0xf31521be // vbsl d2, d21, d30 |
| .long 0xecbd8b0e // vpop {d8-d14} |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_hardlight_vfp4 |
| .globl _sk_hardlight_vfp4 |
| FUNCTION(_sk_hardlight_vfp4) |
| _sk_hardlight_vfp4: |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2670d04 // vsub.f32 d16, d7, d4 |
| .long 0xf2617d87 // vsub.f32 d23, d17, d7 |
| .long 0xf2611d83 // vsub.f32 d17, d17, d3 |
| .long 0xf2672d05 // vsub.f32 d18, d7, d5 |
| .long 0xf2674d06 // vsub.f32 d20, d7, d6 |
| .long 0xf2633d00 // vsub.f32 d19, d3, d0 |
| .long 0xf2635d01 // vsub.f32 d21, d3, d1 |
| .long 0xf2636d02 // vsub.f32 d22, d3, d2 |
| .long 0xf347bd90 // vmul.f32 d27, d23, d0 |
| .long 0xf341cd94 // vmul.f32 d28, d17, d4 |
| .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| .long 0xf3463db4 // vmul.f32 d19, d22, d20 |
| .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| .long 0xf2404d00 // vadd.f32 d20, d0, d0 |
| .long 0xf3405d14 // vmul.f32 d21, d0, d4 |
| .long 0xf2416d01 // vadd.f32 d22, d1, d1 |
| .long 0xf3418d15 // vmul.f32 d24, d1, d5 |
| .long 0xf2429d02 // vadd.f32 d25, d2, d2 |
| .long 0xf342ad16 // vmul.f32 d26, d2, d6 |
| .long 0xf347dd91 // vmul.f32 d29, d23, d1 |
| .long 0xf341fd95 // vmul.f32 d31, d17, d5 |
| .long 0xf24cbdab // vadd.f32 d27, d28, d27 |
| .long 0xf3477d92 // vmul.f32 d23, d23, d2 |
| .long 0xf341cd96 // vmul.f32 d28, d17, d6 |
| .long 0xf2400da0 // vadd.f32 d16, d16, d16 |
| .long 0xf343ed17 // vmul.f32 d30, d3, d7 |
| .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| .long 0xf2433da3 // vadd.f32 d19, d19, d19 |
| .long 0xf3434e24 // vcge.f32 d20, d3, d20 |
| .long 0xf2455da5 // vadd.f32 d21, d21, d21 |
| .long 0xf3436e26 // vcge.f32 d22, d3, d22 |
| .long 0xf3439e29 // vcge.f32 d25, d3, d25 |
| .long 0xf2488da8 // vadd.f32 d24, d24, d24 |
| .long 0xf24aadaa // vadd.f32 d26, d26, d26 |
| .long 0xf2073c31 // vfma.f32 d3, d7, d17 |
| .long 0xf24fddad // vadd.f32 d29, d31, d29 |
| .long 0xf24c1da7 // vadd.f32 d17, d28, d23 |
| .long 0xf26e0da0 // vsub.f32 d16, d30, d16 |
| .long 0xf26e2da2 // vsub.f32 d18, d30, d18 |
| .long 0xf26e3da3 // vsub.f32 d19, d30, d19 |
| .long 0xf35541b0 // vbsl d20, d21, d16 |
| .long 0xf35861b2 // vbsl d22, d24, d18 |
| .long 0xf35a91b3 // vbsl d25, d26, d19 |
| .long 0xf20b0da4 // vadd.f32 d0, d27, d20 |
| .long 0xf20d1da6 // vadd.f32 d1, d29, d22 |
| .long 0xf2012da9 // vadd.f32 d2, d17, d25 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_overlay_vfp4 |
| .globl _sk_overlay_vfp4 |
| FUNCTION(_sk_overlay_vfp4) |
| _sk_overlay_vfp4: |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2670d04 // vsub.f32 d16, d7, d4 |
| .long 0xf2617d87 // vsub.f32 d23, d17, d7 |
| .long 0xf2611d83 // vsub.f32 d17, d17, d3 |
| .long 0xf2672d05 // vsub.f32 d18, d7, d5 |
| .long 0xf2674d06 // vsub.f32 d20, d7, d6 |
| .long 0xf2633d00 // vsub.f32 d19, d3, d0 |
| .long 0xf2635d01 // vsub.f32 d21, d3, d1 |
| .long 0xf2636d02 // vsub.f32 d22, d3, d2 |
| .long 0xf347bd90 // vmul.f32 d27, d23, d0 |
| .long 0xf341cd94 // vmul.f32 d28, d17, d4 |
| .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| .long 0xf3463db4 // vmul.f32 d19, d22, d20 |
| .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| .long 0xf2444d04 // vadd.f32 d20, d4, d4 |
| .long 0xf3405d14 // vmul.f32 d21, d0, d4 |
| .long 0xf2456d05 // vadd.f32 d22, d5, d5 |
| .long 0xf3418d15 // vmul.f32 d24, d1, d5 |
| .long 0xf2469d06 // vadd.f32 d25, d6, d6 |
| .long 0xf342ad16 // vmul.f32 d26, d2, d6 |
| .long 0xf347dd91 // vmul.f32 d29, d23, d1 |
| .long 0xf341fd95 // vmul.f32 d31, d17, d5 |
| .long 0xf24cbdab // vadd.f32 d27, d28, d27 |
| .long 0xf3477d92 // vmul.f32 d23, d23, d2 |
| .long 0xf341cd96 // vmul.f32 d28, d17, d6 |
| .long 0xf343ed17 // vmul.f32 d30, d3, d7 |
| .long 0xf2400da0 // vadd.f32 d16, d16, d16 |
| .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| .long 0xf2433da3 // vadd.f32 d19, d19, d19 |
| .long 0xf3474e24 // vcge.f32 d20, d7, d20 |
| .long 0xf2455da5 // vadd.f32 d21, d21, d21 |
| .long 0xf3476e26 // vcge.f32 d22, d7, d22 |
| .long 0xf2488da8 // vadd.f32 d24, d24, d24 |
| .long 0xf3479e29 // vcge.f32 d25, d7, d25 |
| .long 0xf24aadaa // vadd.f32 d26, d26, d26 |
| .long 0xf2073c31 // vfma.f32 d3, d7, d17 |
| .long 0xf24fddad // vadd.f32 d29, d31, d29 |
| .long 0xf24c1da7 // vadd.f32 d17, d28, d23 |
| .long 0xf26e0da0 // vsub.f32 d16, d30, d16 |
| .long 0xf26e2da2 // vsub.f32 d18, d30, d18 |
| .long 0xf26e3da3 // vsub.f32 d19, d30, d19 |
| .long 0xf35541b0 // vbsl d20, d21, d16 |
| .long 0xf35861b2 // vbsl d22, d24, d18 |
| .long 0xf35a91b3 // vbsl d25, d26, d19 |
| .long 0xf20b0da4 // vadd.f32 d0, d27, d20 |
| .long 0xf20d1da6 // vadd.f32 d1, d29, d22 |
| .long 0xf2012da9 // vadd.f32 d2, d17, d25 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_softlight_vfp4 |
| .globl _sk_softlight_vfp4 |
| FUNCTION(_sk_softlight_vfp4) |
| _sk_softlight_vfp4: |
| .long 0xed2d8b06 // vpush {d8-d10} |
| .long 0xeec58aa7 // vdiv.f32 s17, s11, s15 |
| .long 0xf3f90407 // vcgt.f32 d16, d7, #0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeec49aa7 // vdiv.f32 s19, s9, s15 |
| .long 0xeec6aaa7 // vdiv.f32 s21, s13, s15 |
| .long 0xee858a07 // vdiv.f32 s16, s10, s14 |
| .long 0xee849a07 // vdiv.f32 s18, s8, s14 |
| .long 0xee86aa07 // vdiv.f32 s20, s12, s14 |
| .long 0xf26021b0 // vorr d18, d16, d16 |
| .long 0xf2c01010 // vmov.i32 d17, #0 |
| .long 0xf3582131 // vbsl d18, d8, d17 |
| .long 0xf26031b0 // vorr d19, d16, d16 |
| .long 0xf3fb45a2 // vrsqrte.f32 d20, d18 |
| .long 0xf3593131 // vbsl d19, d9, d17 |
| .long 0xf35a0131 // vbsl d16, d10, d17 |
| .long 0xf3fb15a3 // vrsqrte.f32 d17, d19 |
| .long 0xf3fb55a0 // vrsqrte.f32 d21, d16 |
| .long 0xf3446db4 // vmul.f32 d22, d20, d20 |
| .long 0xf243ada3 // vadd.f32 d26, d19, d19 |
| .long 0xf240bda0 // vadd.f32 d27, d16, d16 |
| .long 0xf3417db1 // vmul.f32 d23, d17, d17 |
| .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| .long 0xf2626fb6 // vrsqrts.f32 d22, d18, d22 |
| .long 0xf2429da2 // vadd.f32 d25, d18, d18 |
| .long 0xf2637fb7 // vrsqrts.f32 d23, d19, d23 |
| .long 0xf2608fb8 // vrsqrts.f32 d24, d16, d24 |
| .long 0xf2818f1c // vmov.f32 d8, #7 |
| .long 0xf2499da9 // vadd.f32 d25, d25, d25 |
| .long 0xf3444db6 // vmul.f32 d20, d20, d22 |
| .long 0xf24a6daa // vadd.f32 d22, d26, d26 |
| .long 0xf24badab // vadd.f32 d26, d27, d27 |
| .long 0xf3411db7 // vmul.f32 d17, d17, d23 |
| .long 0xf3455db8 // vmul.f32 d21, d21, d24 |
| .long 0xf3fb7524 // vrecpe.f32 d23, d20 |
| .long 0xf3498db9 // vmul.f32 d24, d25, d25 |
| .long 0xf3fbd521 // vrecpe.f32 d29, d17 |
| .long 0xf34aedba // vmul.f32 d30, d26, d26 |
| .long 0xf3fbf525 // vrecpe.f32 d31, d21 |
| .long 0xf2444fb7 // vrecps.f32 d20, d20, d23 |
| .long 0xf346cdb6 // vmul.f32 d28, d22, d22 |
| .long 0xf2411fbd // vrecps.f32 d17, d17, d29 |
| .long 0xf3c7bf10 // vmov.f32 d27, #-1 |
| .long 0xf2455fbf // vrecps.f32 d21, d21, d31 |
| .long 0xf24aadae // vadd.f32 d26, d26, d30 |
| .long 0xf2498da8 // vadd.f32 d24, d25, d24 |
| .long 0xf2429dab // vadd.f32 d25, d18, d27 |
| .long 0xf2466dac // vadd.f32 d22, d22, d28 |
| .long 0xf243cdab // vadd.f32 d28, d19, d27 |
| .long 0xf240bdab // vadd.f32 d27, d16, d27 |
| .long 0xf3474db4 // vmul.f32 d20, d23, d20 |
| .long 0xf2c7ef10 // vmov.f32 d30, #1 |
| .long 0xf34d1db1 // vmul.f32 d17, d29, d17 |
| .long 0xf34badba // vmul.f32 d26, d27, d26 |
| .long 0xf242bd02 // vadd.f32 d27, d2, d2 |
| .long 0xf26edda0 // vsub.f32 d29, d30, d16 |
| .long 0xf3498db8 // vmul.f32 d24, d25, d24 |
| .long 0xf3429d98 // vmul.f32 d25, d18, d8 |
| .long 0xf34f5db5 // vmul.f32 d21, d31, d21 |
| .long 0xf26efda2 // vsub.f32 d31, d30, d18 |
| .long 0xf2642da2 // vsub.f32 d18, d20, d18 |
| .long 0xf26b4d83 // vsub.f32 d20, d27, d3 |
| .long 0xf2498da8 // vadd.f32 d24, d25, d24 |
| .long 0xf34c6db6 // vmul.f32 d22, d28, d22 |
| .long 0xf3437d98 // vmul.f32 d23, d19, d8 |
| .long 0xf3449dbd // vmul.f32 d25, d20, d29 |
| .long 0xf245dd05 // vadd.f32 d29, d5, d5 |
| .long 0xf340cd98 // vmul.f32 d28, d16, d8 |
| .long 0xf2476da6 // vadd.f32 d22, d23, d22 |
| .long 0xf2611da3 // vsub.f32 d17, d17, d19 |
| .long 0xf24dddad // vadd.f32 d29, d29, d29 |
| .long 0xf24c7daa // vadd.f32 d23, d28, d26 |
| .long 0xf2650da0 // vsub.f32 d16, d21, d16 |
| .long 0xf26e3da3 // vsub.f32 d19, d30, d19 |
| .long 0xf347de2d // vcge.f32 d29, d7, d29 |
| .long 0xf241ad01 // vadd.f32 d26, d1, d1 |
| .long 0xf3444d97 // vmul.f32 d20, d20, d7 |
| .long 0xf358d1b2 // vbsl d29, d24, d18 |
| .long 0xf2448d04 // vadd.f32 d24, d4, d4 |
| .long 0xf2462d06 // vadd.f32 d18, d6, d6 |
| .long 0xf26a5d83 // vsub.f32 d21, d26, d3 |
| .long 0xf2488da8 // vadd.f32 d24, d24, d24 |
| .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| .long 0xf345cdbf // vmul.f32 d28, d21, d31 |
| .long 0xf3455d97 // vmul.f32 d21, d21, d7 |
| .long 0xf3478e28 // vcge.f32 d24, d7, d24 |
| .long 0xf3472e22 // vcge.f32 d18, d7, d18 |
| .long 0xf343fd14 // vmul.f32 d31, d3, d4 |
| .long 0xf3455dbd // vmul.f32 d21, d21, d29 |
| .long 0xf35681b1 // vbsl d24, d22, d17 |
| .long 0xf2401d00 // vadd.f32 d17, d0, d0 |
| .long 0xf35721b0 // vbsl d18, d23, d16 |
| .long 0xf24c0d83 // vadd.f32 d16, d28, d3 |
| .long 0xf2496d83 // vadd.f32 d22, d25, d3 |
| .long 0xf2617d83 // vsub.f32 d23, d17, d3 |
| .long 0xf3442db2 // vmul.f32 d18, d20, d18 |
| .long 0xf3434e2a // vcge.f32 d20, d3, d26 |
| .long 0xf343ae2b // vcge.f32 d26, d3, d27 |
| .long 0xf3473db3 // vmul.f32 d19, d23, d19 |
| .long 0xf3477d97 // vmul.f32 d23, d23, d7 |
| .long 0xf3431e21 // vcge.f32 d17, d3, d17 |
| .long 0xf3400d95 // vmul.f32 d16, d16, d5 |
| .long 0xf2433d83 // vadd.f32 d19, d19, d3 |
| .long 0xf3477db8 // vmul.f32 d23, d23, d24 |
| .long 0xf26e8d87 // vsub.f32 d24, d30, d7 |
| .long 0xf26eed83 // vsub.f32 d30, d30, d3 |
| .long 0xf3433d94 // vmul.f32 d19, d19, d4 |
| .long 0xf24f7da7 // vadd.f32 d23, d31, d23 |
| .long 0xf3489d91 // vmul.f32 d25, d24, d1 |
| .long 0xf348cd90 // vmul.f32 d28, d24, d0 |
| .long 0xf34edd94 // vmul.f32 d29, d30, d4 |
| .long 0xf34ebd95 // vmul.f32 d27, d30, d5 |
| .long 0xf3488d92 // vmul.f32 d24, d24, d2 |
| .long 0xf34efd96 // vmul.f32 d31, d30, d6 |
| .long 0xf24dcdac // vadd.f32 d28, d29, d28 |
| .long 0xf343dd15 // vmul.f32 d29, d3, d5 |
| .long 0xf24b9da9 // vadd.f32 d25, d27, d25 |
| .long 0xf343bd16 // vmul.f32 d27, d3, d6 |
| .long 0xf3466d96 // vmul.f32 d22, d22, d6 |
| .long 0xf24f8da8 // vadd.f32 d24, d31, d24 |
| .long 0xf24d5da5 // vadd.f32 d21, d29, d21 |
| .long 0xf24b2da2 // vadd.f32 d18, d27, d18 |
| .long 0xf35311b7 // vbsl d17, d19, d23 |
| .long 0xf35041b5 // vbsl d20, d16, d21 |
| .long 0xf356a1b2 // vbsl d26, d22, d18 |
| .long 0xf2073c3e // vfma.f32 d3, d7, d30 |
| .long 0xf20c0da1 // vadd.f32 d0, d28, d17 |
| .long 0xf2091da4 // vadd.f32 d1, d25, d20 |
| .long 0xf2082daa // vadd.f32 d2, d24, d26 |
| .long 0xecbd8b06 // vpop {d8-d10} |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_clamp_0_vfp4 |
| .globl _sk_clamp_0_vfp4 |
| FUNCTION(_sk_clamp_0_vfp4) |
| _sk_clamp_0_vfp4: |
| .long 0xf2c00010 // vmov.i32 d16, #0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2000f20 // vmax.f32 d0, d0, d16 |
| .long 0xf2011f20 // vmax.f32 d1, d1, d16 |
| .long 0xf2022f20 // vmax.f32 d2, d2, d16 |
| .long 0xf2033f20 // vmax.f32 d3, d3, d16 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_clamp_1_vfp4 |
| .globl _sk_clamp_1_vfp4 |
| FUNCTION(_sk_clamp_1_vfp4) |
| _sk_clamp_1_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2200f20 // vmin.f32 d0, d0, d16 |
| .long 0xf2211f20 // vmin.f32 d1, d1, d16 |
| .long 0xf2222f20 // vmin.f32 d2, d2, d16 |
| .long 0xf2233f20 // vmin.f32 d3, d3, d16 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_clamp_a_vfp4 |
| .globl _sk_clamp_a_vfp4 |
| FUNCTION(_sk_clamp_a_vfp4) |
| _sk_clamp_a_vfp4: |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf2233f20 // vmin.f32 d3, d3, d16 |
| .long 0xf2200f03 // vmin.f32 d0, d0, d3 |
| .long 0xf2211f03 // vmin.f32 d1, d1, d3 |
| .long 0xf2222f03 // vmin.f32 d2, d2, d3 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_set_rgb_vfp4 |
| .globl _sk_set_rgb_vfp4 |
| FUNCTION(_sk_set_rgb_vfp4) |
| _sk_set_rgb_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe283e008 // add lr, r3, #8 |
| .long 0xf4a30c9d // vld1.32 {d0[]}, [r3 :32]! |
| .long 0xf4ae2c9f // vld1.32 {d2[]}, [lr :32] |
| .long 0xf4a31c9f // vld1.32 {d1[]}, [r3 :32] |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_swap_rb_vfp4 |
| .globl _sk_swap_rb_vfp4 |
| FUNCTION(_sk_swap_rb_vfp4) |
| _sk_swap_rb_vfp4: |
| .long 0xeef00b40 // vmov.f64 d16, d0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeeb00b42 // vmov.f64 d0, d2 |
| .long 0xeeb02b60 // vmov.f64 d2, d16 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_swap_vfp4 |
| .globl _sk_swap_vfp4 |
| FUNCTION(_sk_swap_vfp4) |
| _sk_swap_vfp4: |
| .long 0xeef00b43 // vmov.f64 d16, d3 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeef01b42 // vmov.f64 d17, d2 |
| .long 0xeef02b41 // vmov.f64 d18, d1 |
| .long 0xeef03b40 // vmov.f64 d19, d0 |
| .long 0xeeb00b44 // vmov.f64 d0, d4 |
| .long 0xeeb01b45 // vmov.f64 d1, d5 |
| .long 0xeeb02b46 // vmov.f64 d2, d6 |
| .long 0xeeb03b47 // vmov.f64 d3, d7 |
| .long 0xeeb04b63 // vmov.f64 d4, d19 |
| .long 0xeeb05b62 // vmov.f64 d5, d18 |
| .long 0xeeb06b61 // vmov.f64 d6, d17 |
| .long 0xeeb07b60 // vmov.f64 d7, d16 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_move_src_dst_vfp4 |
| .globl _sk_move_src_dst_vfp4 |
| FUNCTION(_sk_move_src_dst_vfp4) |
| _sk_move_src_dst_vfp4: |
| .long 0xeeb04b40 // vmov.f64 d4, d0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeeb05b41 // vmov.f64 d5, d1 |
| .long 0xeeb06b42 // vmov.f64 d6, d2 |
| .long 0xeeb07b43 // vmov.f64 d7, d3 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_move_dst_src_vfp4 |
| .globl _sk_move_dst_src_vfp4 |
| FUNCTION(_sk_move_dst_src_vfp4) |
| _sk_move_dst_src_vfp4: |
| .long 0xeeb00b44 // vmov.f64 d0, d4 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeeb01b45 // vmov.f64 d1, d5 |
| .long 0xeeb02b46 // vmov.f64 d2, d6 |
| .long 0xeeb03b47 // vmov.f64 d3, d7 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_premul_vfp4 |
| .globl _sk_premul_vfp4 |
| FUNCTION(_sk_premul_vfp4) |
| _sk_premul_vfp4: |
| .long 0xf3000d13 // vmul.f32 d0, d0, d3 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3011d13 // vmul.f32 d1, d1, d3 |
| .long 0xf3022d13 // vmul.f32 d2, d2, d3 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_unpremul_vfp4 |
| .globl _sk_unpremul_vfp4 |
| FUNCTION(_sk_unpremul_vfp4) |
| _sk_unpremul_vfp4: |
| .long 0xed2d8b04 // vpush {d8-d9} |
| .long 0xeeb78a00 // vmov.f32 s16, #112 |
| .long 0xf3f91503 // vceq.f32 d17, d3, #0 |
| .long 0xf2c00010 // vmov.i32 d16, #0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeec89a23 // vdiv.f32 s19, s16, s7 |
| .long 0xee889a03 // vdiv.f32 s18, s16, s6 |
| .long 0xf3501199 // vbsl d17, d16, d9 |
| .long 0xf3010d90 // vmul.f32 d0, d17, d0 |
| .long 0xf3011d91 // vmul.f32 d1, d17, d1 |
| .long 0xf3012d92 // vmul.f32 d2, d17, d2 |
| .long 0xecbd8b04 // vpop {d8-d9} |
| .long 0xe12fff13 // bx r3 |
| .long 0xe320f000 // nop {0} |
| |
| HIDDEN _sk_from_srgb_vfp4 |
| .globl _sk_from_srgb_vfp4 |
| FUNCTION(_sk_from_srgb_vfp4) |
| _sk_from_srgb_vfp4: |
| .long 0xeddf3b20 // vldr d19, [pc, #128] |
| .long 0xf3408d10 // vmul.f32 d24, d0, d0 |
| .long 0xeddf0b1c // vldr d16, [pc, #112] |
| .long 0xf26341b3 // vorr d20, d19, d19 |
| .long 0xf26351b3 // vorr d21, d19, d19 |
| .long 0xeddf9b1f // vldr d25, [pc, #124] |
| .long 0xf2404c30 // vfma.f32 d20, d0, d16 |
| .long 0xeddf2b1b // vldr d18, [pc, #108] |
| .long 0xf2415c30 // vfma.f32 d21, d1, d16 |
| .long 0xeddfcb1d // vldr d28, [pc, #116] |
| .long 0xf2423c30 // vfma.f32 d19, d2, d16 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3426d12 // vmul.f32 d22, d2, d2 |
| .long 0xf3417d11 // vmul.f32 d23, d1, d1 |
| .long 0xf3620e80 // vcgt.f32 d16, d18, d0 |
| .long 0xf3621e81 // vcgt.f32 d17, d18, d1 |
| .long 0xf341ad39 // vmul.f32 d26, d1, d25 |
| .long 0xf342bd39 // vmul.f32 d27, d2, d25 |
| .long 0xf3622e82 // vcgt.f32 d18, d18, d2 |
| .long 0xf3409d39 // vmul.f32 d25, d0, d25 |
| .long 0xf26cd1bc // vorr d29, d28, d28 |
| .long 0xf248dcb4 // vfma.f32 d29, d24, d20 |
| .long 0xf26c41bc // vorr d20, d28, d28 |
| .long 0xf2474cb5 // vfma.f32 d20, d23, d21 |
| .long 0xf246ccb3 // vfma.f32 d28, d22, d19 |
| .long 0xf35901bd // vbsl d16, d25, d29 |
| .long 0xf35a11b4 // vbsl d17, d26, d20 |
| .long 0xf35b21bc // vbsl d18, d27, d28 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xe12fff13 // bx r3 |
| .long 0x3e99999a // .word 0x3e99999a |
| .long 0x3e99999a // .word 0x3e99999a |
| .long 0x3f328f5c // .word 0x3f328f5c |
| .long 0x3f328f5c // .word 0x3f328f5c |
| .long 0x3d6147ae // .word 0x3d6147ae |
| .long 0x3d6147ae // .word 0x3d6147ae |
| .long 0x3d9e8391 // .word 0x3d9e8391 |
| .long 0x3d9e8391 // .word 0x3d9e8391 |
| .long 0x3b23d70a // .word 0x3b23d70a |
| .long 0x3b23d70a // .word 0x3b23d70a |
| |
| HIDDEN _sk_to_srgb_vfp4 |
| .globl _sk_to_srgb_vfp4 |
| FUNCTION(_sk_to_srgb_vfp4) |
| _sk_to_srgb_vfp4: |
| .long 0xf3fb0582 // vrsqrte.f32 d16, d2 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3fb1581 // vrsqrte.f32 d17, d1 |
| .long 0xf3fb2580 // vrsqrte.f32 d18, d0 |
| .long 0xf3403db0 // vmul.f32 d19, d16, d16 |
| .long 0xf3414db1 // vmul.f32 d20, d17, d17 |
| .long 0xf3425db2 // vmul.f32 d21, d18, d18 |
| .long 0xf2623f33 // vrsqrts.f32 d19, d2, d19 |
| .long 0xf2614f34 // vrsqrts.f32 d20, d1, d20 |
| .long 0xf2605f35 // vrsqrts.f32 d21, d0, d21 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| .long 0xf3422db5 // vmul.f32 d18, d18, d21 |
| .long 0xf3fb3520 // vrecpe.f32 d19, d16 |
| .long 0xf3fb4521 // vrecpe.f32 d20, d17 |
| .long 0xf3fb6522 // vrecpe.f32 d22, d18 |
| .long 0xf3fb55a0 // vrsqrte.f32 d21, d16 |
| .long 0xf3fb75a1 // vrsqrte.f32 d23, d17 |
| .long 0xf3fb85a2 // vrsqrte.f32 d24, d18 |
| .long 0xf2409fb3 // vrecps.f32 d25, d16, d19 |
| .long 0xf241afb4 // vrecps.f32 d26, d17, d20 |
| .long 0xf242bfb6 // vrecps.f32 d27, d18, d22 |
| .long 0xf345cdb5 // vmul.f32 d28, d21, d21 |
| .long 0xf347ddb7 // vmul.f32 d29, d23, d23 |
| .long 0xf348edb8 // vmul.f32 d30, d24, d24 |
| .long 0xf2600fbc // vrsqrts.f32 d16, d16, d28 |
| .long 0xf2611fbd // vrsqrts.f32 d17, d17, d29 |
| .long 0xf2622fbe // vrsqrts.f32 d18, d18, d30 |
| .long 0xf3433db9 // vmul.f32 d19, d19, d25 |
| .long 0xeddf9b21 // vldr d25, [pc, #132] |
| .long 0xf3444dba // vmul.f32 d20, d20, d26 |
| .long 0xeddfab21 // vldr d26, [pc, #132] |
| .long 0xf3466dbb // vmul.f32 d22, d22, d27 |
| .long 0xf26ab1ba // vorr d27, d26, d26 |
| .long 0xf243bcb9 // vfma.f32 d27, d19, d25 |
| .long 0xf26a31ba // vorr d19, d26, d26 |
| .long 0xf2443cb9 // vfma.f32 d19, d20, d25 |
| .long 0xeddf4b1d // vldr d20, [pc, #116] |
| .long 0xf246acb9 // vfma.f32 d26, d22, d25 |
| .long 0xf3450db0 // vmul.f32 d16, d21, d16 |
| .long 0xeddf5b1c // vldr d21, [pc, #112] |
| .long 0xf3471db1 // vmul.f32 d17, d23, d17 |
| .long 0xf3482db2 // vmul.f32 d18, d24, d18 |
| .long 0xf3406d35 // vmul.f32 d22, d0, d21 |
| .long 0xf240bcb4 // vfma.f32 d27, d16, d20 |
| .long 0xf2413cb4 // vfma.f32 d19, d17, d20 |
| .long 0xf242acb4 // vfma.f32 d26, d18, d20 |
| .long 0xeddf2b17 // vldr d18, [pc, #92] |
| .long 0xf3417d35 // vmul.f32 d23, d1, d21 |
| .long 0xf3620e80 // vcgt.f32 d16, d18, d0 |
| .long 0xf3621e81 // vcgt.f32 d17, d18, d1 |
| .long 0xf3622e82 // vcgt.f32 d18, d18, d2 |
| .long 0xf3425d35 // vmul.f32 d21, d2, d21 |
| .long 0xf2c74f10 // vmov.f32 d20, #1 |
| .long 0xf2648faa // vmin.f32 d24, d20, d26 |
| .long 0xf2643fa3 // vmin.f32 d19, d20, d19 |
| .long 0xf2644fab // vmin.f32 d20, d20, d27 |
| .long 0xf35601b8 // vbsl d16, d22, d24 |
| .long 0xf35711b3 // vbsl d17, d23, d19 |
| .long 0xf35521b4 // vbsl d18, d21, d20 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22221b2 // vorr d2, d18, d18 |
| .long 0xe12fff13 // bx r3 |
| .long 0x3f306fce // .word 0x3f306fce |
| .long 0x3f306fce // .word 0x3f306fce |
| .long 0xbdca57a8 // .word 0xbdca57a8 |
| .long 0xbdca57a8 // .word 0xbdca57a8 |
| .long 0x3ed287c2 // .word 0x3ed287c2 |
| .long 0x3ed287c2 // .word 0x3ed287c2 |
| .long 0x41475c29 // .word 0x41475c29 |
| .long 0x41475c29 // .word 0x41475c29 |
| .long 0x3b8ce704 // .word 0x3b8ce704 |
| .long 0x3b8ce704 // .word 0x3b8ce704 |
| |
| HIDDEN _sk_from_2dot2_vfp4 |
| .globl _sk_from_2dot2_vfp4 |
| FUNCTION(_sk_from_2dot2_vfp4) |
| _sk_from_2dot2_vfp4: |
| .long 0xf3fb0580 // vrsqrte.f32 d16, d0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3fb1581 // vrsqrte.f32 d17, d1 |
| .long 0xf3fb2582 // vrsqrte.f32 d18, d2 |
| .long 0xf3403db0 // vmul.f32 d19, d16, d16 |
| .long 0xf3414db1 // vmul.f32 d20, d17, d17 |
| .long 0xf3425db2 // vmul.f32 d21, d18, d18 |
| .long 0xf2603f33 // vrsqrts.f32 d19, d0, d19 |
| .long 0xf2614f34 // vrsqrts.f32 d20, d1, d20 |
| .long 0xf2625f35 // vrsqrts.f32 d21, d2, d21 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| .long 0xf3422db5 // vmul.f32 d18, d18, d21 |
| .long 0xf3fb35a0 // vrsqrte.f32 d19, d16 |
| .long 0xf3fb45a1 // vrsqrte.f32 d20, d17 |
| .long 0xf3fb55a2 // vrsqrte.f32 d21, d18 |
| .long 0xf3436db3 // vmul.f32 d22, d19, d19 |
| .long 0xf3447db4 // vmul.f32 d23, d20, d20 |
| .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| .long 0xf2600fb6 // vrsqrts.f32 d16, d16, d22 |
| .long 0xf2611fb7 // vrsqrts.f32 d17, d17, d23 |
| .long 0xf2622fb8 // vrsqrts.f32 d18, d18, d24 |
| .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| .long 0xf3441db1 // vmul.f32 d17, d20, d17 |
| .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| .long 0xf3fb35a0 // vrsqrte.f32 d19, d16 |
| .long 0xf3fb45a1 // vrsqrte.f32 d20, d17 |
| .long 0xf3fb55a2 // vrsqrte.f32 d21, d18 |
| .long 0xf3436db3 // vmul.f32 d22, d19, d19 |
| .long 0xf3447db4 // vmul.f32 d23, d20, d20 |
| .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| .long 0xf2600fb6 // vrsqrts.f32 d16, d16, d22 |
| .long 0xf2611fb7 // vrsqrts.f32 d17, d17, d23 |
| .long 0xf2622fb8 // vrsqrts.f32 d18, d18, d24 |
| .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| .long 0xf3441db1 // vmul.f32 d17, d20, d17 |
| .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| .long 0xf3fb35a0 // vrsqrte.f32 d19, d16 |
| .long 0xf3fb45a1 // vrsqrte.f32 d20, d17 |
| .long 0xf3fb55a2 // vrsqrte.f32 d21, d18 |
| .long 0xf3436db3 // vmul.f32 d22, d19, d19 |
| .long 0xf3447db4 // vmul.f32 d23, d20, d20 |
| .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| .long 0xf2600fb6 // vrsqrts.f32 d16, d16, d22 |
| .long 0xf2611fb7 // vrsqrts.f32 d17, d17, d23 |
| .long 0xf2622fb8 // vrsqrts.f32 d18, d18, d24 |
| .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| .long 0xf3441db1 // vmul.f32 d17, d20, d17 |
| .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| .long 0xf3fb35a0 // vrsqrte.f32 d19, d16 |
| .long 0xf3fb45a1 // vrsqrte.f32 d20, d17 |
| .long 0xf3fb55a2 // vrsqrte.f32 d21, d18 |
| .long 0xf340bdb0 // vmul.f32 d27, d16, d16 |
| .long 0xf341ddb1 // vmul.f32 d29, d17, d17 |
| .long 0xf3436db3 // vmul.f32 d22, d19, d19 |
| .long 0xf3447db4 // vmul.f32 d23, d20, d20 |
| .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| .long 0xf2606fb6 // vrsqrts.f32 d22, d16, d22 |
| .long 0xf2617fb7 // vrsqrts.f32 d23, d17, d23 |
| .long 0xf2628fb8 // vrsqrts.f32 d24, d18, d24 |
| .long 0xf3400dbb // vmul.f32 d16, d16, d27 |
| .long 0xf3411dbd // vmul.f32 d17, d17, d29 |
| .long 0xf341bd11 // vmul.f32 d27, d1, d1 |
| .long 0xf3433db6 // vmul.f32 d19, d19, d22 |
| .long 0xf3444db7 // vmul.f32 d20, d20, d23 |
| .long 0xf3455db8 // vmul.f32 d21, d21, d24 |
| .long 0xf34b1db1 // vmul.f32 d17, d27, d17 |
| .long 0xf3fb65a3 // vrsqrte.f32 d22, d19 |
| .long 0xf3fb75a4 // vrsqrte.f32 d23, d20 |
| .long 0xf3fb85a5 // vrsqrte.f32 d24, d21 |
| .long 0xf3469db6 // vmul.f32 d25, d22, d22 |
| .long 0xf347adb7 // vmul.f32 d26, d23, d23 |
| .long 0xf348cdb8 // vmul.f32 d28, d24, d24 |
| .long 0xf2633fb9 // vrsqrts.f32 d19, d19, d25 |
| .long 0xf2644fba // vrsqrts.f32 d20, d20, d26 |
| .long 0xf3429db2 // vmul.f32 d25, d18, d18 |
| .long 0xf2655fbc // vrsqrts.f32 d21, d21, d28 |
| .long 0xf340ad10 // vmul.f32 d26, d0, d0 |
| .long 0xf3422db9 // vmul.f32 d18, d18, d25 |
| .long 0xf3429d12 // vmul.f32 d25, d2, d2 |
| .long 0xf3463db3 // vmul.f32 d19, d22, d19 |
| .long 0xf3474db4 // vmul.f32 d20, d23, d20 |
| .long 0xf34a0db0 // vmul.f32 d16, d26, d16 |
| .long 0xf3485db5 // vmul.f32 d21, d24, d21 |
| .long 0xf3492db2 // vmul.f32 d18, d25, d18 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| .long 0xf3422db5 // vmul.f32 d18, d18, d21 |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2000fa3 // vmax.f32 d0, d16, d19 |
| .long 0xf2011fa3 // vmax.f32 d1, d17, d19 |
| .long 0xf2022fa3 // vmax.f32 d2, d18, d19 |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_to_2dot2_vfp4 |
| .globl _sk_to_2dot2_vfp4 |
| FUNCTION(_sk_to_2dot2_vfp4) |
| _sk_to_2dot2_vfp4: |
| .long 0xf3fb0580 // vrsqrte.f32 d16, d0 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3fb1581 // vrsqrte.f32 d17, d1 |
| .long 0xf3fb3582 // vrsqrte.f32 d19, d2 |
| .long 0xf3402db0 // vmul.f32 d18, d16, d16 |
| .long 0xf3414db1 // vmul.f32 d20, d17, d17 |
| .long 0xf3435db3 // vmul.f32 d21, d19, d19 |
| .long 0xf2602f32 // vrsqrts.f32 d18, d0, d18 |
| .long 0xf2614f34 // vrsqrts.f32 d20, d1, d20 |
| .long 0xf2625f35 // vrsqrts.f32 d21, d2, d21 |
| .long 0xf3402db2 // vmul.f32 d18, d16, d18 |
| .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| .long 0xf3430db5 // vmul.f32 d16, d19, d21 |
| .long 0xf3fb35a2 // vrsqrte.f32 d19, d18 |
| .long 0xf3fb45a1 // vrsqrte.f32 d20, d17 |
| .long 0xf3fb55a0 // vrsqrte.f32 d21, d16 |
| .long 0xf3fbc522 // vrecpe.f32 d28, d18 |
| .long 0xf3436db3 // vmul.f32 d22, d19, d19 |
| .long 0xf3447db4 // vmul.f32 d23, d20, d20 |
| .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| .long 0xf2626fb6 // vrsqrts.f32 d22, d18, d22 |
| .long 0xf2617fb7 // vrsqrts.f32 d23, d17, d23 |
| .long 0xf2608fb8 // vrsqrts.f32 d24, d16, d24 |
| .long 0xf2422fbc // vrecps.f32 d18, d18, d28 |
| .long 0xf3433db6 // vmul.f32 d19, d19, d22 |
| .long 0xf3444db7 // vmul.f32 d20, d20, d23 |
| .long 0xf3455db8 // vmul.f32 d21, d21, d24 |
| .long 0xf34c2db2 // vmul.f32 d18, d28, d18 |
| .long 0xf3fb65a3 // vrsqrte.f32 d22, d19 |
| .long 0xf3fb75a4 // vrsqrte.f32 d23, d20 |
| .long 0xf3fb85a5 // vrsqrte.f32 d24, d21 |
| .long 0xf3469db6 // vmul.f32 d25, d22, d22 |
| .long 0xf347adb7 // vmul.f32 d26, d23, d23 |
| .long 0xf348bdb8 // vmul.f32 d27, d24, d24 |
| .long 0xf2633fb9 // vrsqrts.f32 d19, d19, d25 |
| .long 0xf2644fba // vrsqrts.f32 d20, d20, d26 |
| .long 0xf2655fbb // vrsqrts.f32 d21, d21, d27 |
| .long 0xf3463db3 // vmul.f32 d19, d22, d19 |
| .long 0xf3474db4 // vmul.f32 d20, d23, d20 |
| .long 0xf3485db5 // vmul.f32 d21, d24, d21 |
| .long 0xf3fb65a3 // vrsqrte.f32 d22, d19 |
| .long 0xf3fb75a4 // vrsqrte.f32 d23, d20 |
| .long 0xf3fb85a5 // vrsqrte.f32 d24, d21 |
| .long 0xf3469db6 // vmul.f32 d25, d22, d22 |
| .long 0xf347adb7 // vmul.f32 d26, d23, d23 |
| .long 0xf348bdb8 // vmul.f32 d27, d24, d24 |
| .long 0xf2633fb9 // vrsqrts.f32 d19, d19, d25 |
| .long 0xf2644fba // vrsqrts.f32 d20, d20, d26 |
| .long 0xf2655fbb // vrsqrts.f32 d21, d21, d27 |
| .long 0xf3463db3 // vmul.f32 d19, d22, d19 |
| .long 0xf3474db4 // vmul.f32 d20, d23, d20 |
| .long 0xf3485db5 // vmul.f32 d21, d24, d21 |
| .long 0xf3fb65a3 // vrsqrte.f32 d22, d19 |
| .long 0xf3fb75a4 // vrsqrte.f32 d23, d20 |
| .long 0xf3fb85a5 // vrsqrte.f32 d24, d21 |
| .long 0xf3469db6 // vmul.f32 d25, d22, d22 |
| .long 0xf347adb7 // vmul.f32 d26, d23, d23 |
| .long 0xf348bdb8 // vmul.f32 d27, d24, d24 |
| .long 0xf2633fb9 // vrsqrts.f32 d19, d19, d25 |
| .long 0xf2644fba // vrsqrts.f32 d20, d20, d26 |
| .long 0xf2655fbb // vrsqrts.f32 d21, d21, d27 |
| .long 0xf3463db3 // vmul.f32 d19, d22, d19 |
| .long 0xf3474db4 // vmul.f32 d20, d23, d20 |
| .long 0xf3485db5 // vmul.f32 d21, d24, d21 |
| .long 0xf3fb65a3 // vrsqrte.f32 d22, d19 |
| .long 0xf3fb75a4 // vrsqrte.f32 d23, d20 |
| .long 0xf3fb85a5 // vrsqrte.f32 d24, d21 |
| .long 0xf3432db2 // vmul.f32 d18, d19, d18 |
| .long 0xf3469db6 // vmul.f32 d25, d22, d22 |
| .long 0xf347adb7 // vmul.f32 d26, d23, d23 |
| .long 0xf348bdb8 // vmul.f32 d27, d24, d24 |
| .long 0xf2639fb9 // vrsqrts.f32 d25, d19, d25 |
| .long 0xf264afba // vrsqrts.f32 d26, d20, d26 |
| .long 0xf265bfbb // vrsqrts.f32 d27, d21, d27 |
| .long 0xf3466db9 // vmul.f32 d22, d22, d25 |
| .long 0xf3fb9521 // vrecpe.f32 d25, d17 |
| .long 0xf3477dba // vmul.f32 d23, d23, d26 |
| .long 0xf3fba520 // vrecpe.f32 d26, d16 |
| .long 0xf3488dbb // vmul.f32 d24, d24, d27 |
| .long 0xf2411fb9 // vrecps.f32 d17, d17, d25 |
| .long 0xf3fbb526 // vrecpe.f32 d27, d22 |
| .long 0xf3fbd527 // vrecpe.f32 d29, d23 |
| .long 0xf2400fba // vrecps.f32 d16, d16, d26 |
| .long 0xf3fbe528 // vrecpe.f32 d30, d24 |
| .long 0xf2466fbb // vrecps.f32 d22, d22, d27 |
| .long 0xf2477fbd // vrecps.f32 d23, d23, d29 |
| .long 0xf2488fbe // vrecps.f32 d24, d24, d30 |
| .long 0xf3491db1 // vmul.f32 d17, d25, d17 |
| .long 0xf34a0db0 // vmul.f32 d16, d26, d16 |
| .long 0xf34b6db6 // vmul.f32 d22, d27, d22 |
| .long 0xf3441db1 // vmul.f32 d17, d20, d17 |
| .long 0xf34d3db7 // vmul.f32 d19, d29, d23 |
| .long 0xf34e4db8 // vmul.f32 d20, d30, d24 |
| .long 0xf3450db0 // vmul.f32 d16, d21, d16 |
| .long 0xf3422db6 // vmul.f32 d18, d18, d22 |
| .long 0xf3411db3 // vmul.f32 d17, d17, d19 |
| .long 0xf3400db4 // vmul.f32 d16, d16, d20 |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2020fa3 // vmax.f32 d0, d18, d19 |
| .long 0xf2011fa3 // vmax.f32 d1, d17, d19 |
| .long 0xf2002fa3 // vmax.f32 d2, d16, d19 |
| .long 0xe12fff13 // bx r3 |
| .long 0xe320f000 // nop {0} |
| |
| HIDDEN _sk_rgb_to_hsl_vfp4 |
| .globl _sk_rgb_to_hsl_vfp4 |
| FUNCTION(_sk_rgb_to_hsl_vfp4) |
| _sk_rgb_to_hsl_vfp4: |
| .long 0xed2d8b08 // vpush {d8-d11} |
| .long 0xf2401f01 // vmax.f32 d17, d0, d1 |
| .long 0xeddf9b2c // vldr d25, [pc, #176] |
| .long 0xf2600f01 // vmin.f32 d16, d0, d1 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xeeb78a00 // vmov.f32 s16, #112 |
| .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| .long 0xf2411f82 // vmax.f32 d17, d17, d2 |
| .long 0xf2602f82 // vmin.f32 d18, d16, d2 |
| .long 0xf2c45610 // vmov.i32 d21, #1073741824 |
| .long 0xf2607d01 // vsub.f32 d23, d0, d1 |
| .long 0xf2656da1 // vsub.f32 d22, d21, d17 |
| .long 0xf221ada2 // vsub.f32 d10, d17, d18 |
| .long 0xf2413da2 // vadd.f32 d19, d17, d18 |
| .long 0xf2c08010 // vmov.i32 d24, #0 |
| .long 0xf2666da2 // vsub.f32 d22, d22, d18 |
| .long 0xf241ae80 // vceq.f32 d26, d17, d0 |
| .long 0xeec8ba2a // vdiv.f32 s23, s16, s21 |
| .long 0xf3430db4 // vmul.f32 d16, d19, d20 |
| .long 0xee88ba0a // vdiv.f32 s22, s16, s20 |
| .long 0xf3209ea4 // vcgt.f32 d9, d16, d20 |
| .long 0xf2614d02 // vsub.f32 d20, d1, d2 |
| .long 0xf3477d9b // vmul.f32 d23, d23, d11 |
| .long 0xf31691b3 // vbsl d9, d22, d19 |
| .long 0xf2623d00 // vsub.f32 d19, d2, d0 |
| .long 0xf3626e01 // vcgt.f32 d22, d2, d1 |
| .long 0xeeca8aa9 // vdiv.f32 s17, s21, s19 |
| .long 0xee8a8a09 // vdiv.f32 s16, s20, s18 |
| .long 0xf3433d9b // vmul.f32 d19, d19, d11 |
| .long 0xf3444d9b // vmul.f32 d20, d20, d11 |
| .long 0xf35961b8 // vbsl d22, d25, d24 |
| .long 0xf2419e81 // vceq.f32 d25, d17, d1 |
| .long 0xf2011ea2 // vceq.f32 d1, d17, d18 |
| .long 0xf2433da5 // vadd.f32 d19, d19, d21 |
| .long 0xf2c15f10 // vmov.f32 d21, #4 |
| .long 0xf2464da4 // vadd.f32 d20, d22, d20 |
| .long 0xf2471da5 // vadd.f32 d17, d23, d21 |
| .long 0xf35391b1 // vbsl d25, d19, d17 |
| .long 0xeddf1b0a // vldr d17, [pc, #40] |
| .long 0xf2612111 // vorr d18, d1, d1 |
| .long 0xf354a1b9 // vbsl d26, d20, d25 |
| .long 0xf35821ba // vbsl d18, d24, d26 |
| .long 0xf3181198 // vbsl d1, d24, d8 |
| .long 0xf3020db1 // vmul.f32 d0, d18, d17 |
| .long 0xf22021b0 // vorr d2, d16, d16 |
| .long 0xecbd8b08 // vpop {d8-d11} |
| .long 0xe12fff13 // bx r3 |
| .long 0xe320f000 // nop {0} |
| .long 0x40c00000 // .word 0x40c00000 |
| .long 0x40c00000 // .word 0x40c00000 |
| .long 0x3e2aaaab // .word 0x3e2aaaab |
| .long 0x3e2aaaab // .word 0x3e2aaaab |
| |
| HIDDEN _sk_hsl_to_rgb_vfp4 |
| .globl _sk_hsl_to_rgb_vfp4 |
| FUNCTION(_sk_hsl_to_rgb_vfp4) |
| _sk_hsl_to_rgb_vfp4: |
| .long 0xf2c72f10 // vmov.f32 d18, #1 |
| .long 0xeddf0b4f // vldr d16, [pc, #316] |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xeddf9b4f // vldr d25, [pc, #316] |
| .long 0xf2415d22 // vadd.f32 d21, d1, d18 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3414d12 // vmul.f32 d20, d1, d2 |
| .long 0xf2416d02 // vadd.f32 d22, d1, d2 |
| .long 0xf2407d20 // vadd.f32 d23, d0, d16 |
| .long 0xf3610e82 // vcgt.f32 d16, d17, d2 |
| .long 0xf3455d92 // vmul.f32 d21, d21, d2 |
| .long 0xf2664da4 // vsub.f32 d20, d22, d20 |
| .long 0xf2426d02 // vadd.f32 d22, d2, d2 |
| .long 0xf3c73f10 // vmov.f32 d19, #-1 |
| .long 0xf35501b4 // vbsl d16, d21, d20 |
| .long 0xf2409d29 // vadd.f32 d25, d0, d25 |
| .long 0xf2408d23 // vadd.f32 d24, d0, d19 |
| .long 0xf3f9e629 // vclt.f32 d30, d25, #0 |
| .long 0xf360ae22 // vcgt.f32 d26, d0, d18 |
| .long 0xf247cda3 // vadd.f32 d28, d23, d19 |
| .long 0xf367dea2 // vcgt.f32 d29, d23, d18 |
| .long 0xf240bd22 // vadd.f32 d27, d0, d18 |
| .long 0xf2666da0 // vsub.f32 d22, d22, d16 |
| .long 0xf2474da2 // vadd.f32 d20, d23, d18 |
| .long 0xf358a190 // vbsl d26, d24, d0 |
| .long 0xf3f98600 // vclt.f32 d24, d0, #0 |
| .long 0xf3695ea2 // vcgt.f32 d21, d25, d18 |
| .long 0xf2493da3 // vadd.f32 d19, d25, d19 |
| .long 0xf35b81ba // vbsl d24, d27, d26 |
| .long 0xf3f9a627 // vclt.f32 d26, d23, #0 |
| .long 0xf35cd1b7 // vbsl d29, d28, d23 |
| .long 0xeddfcb35 // vldr d28, [pc, #212] |
| .long 0xf2492da2 // vadd.f32 d18, d25, d18 |
| .long 0xf260bda6 // vsub.f32 d27, d16, d22 |
| .long 0xf354a1bd // vbsl d26, d20, d29 |
| .long 0xf2c14f18 // vmov.f32 d20, #6 |
| .long 0xf35351b9 // vbsl d21, d19, d25 |
| .long 0xf26cddaa // vsub.f32 d29, d28, d26 |
| .long 0xf352e1b5 // vbsl d30, d18, d21 |
| .long 0xf34b2db4 // vmul.f32 d18, d27, d20 |
| .long 0xf26c3da8 // vsub.f32 d19, d28, d24 |
| .long 0xf26c4dae // vsub.f32 d20, d28, d30 |
| .long 0xf36cbeaa // vcgt.f32 d27, d28, d26 |
| .long 0xf3425dbd // vmul.f32 d21, d18, d29 |
| .long 0xf3477db2 // vmul.f32 d23, d23, d18 |
| .long 0xf3423db3 // vmul.f32 d19, d18, d19 |
| .long 0xf3444db2 // vmul.f32 d20, d20, d18 |
| .long 0xf2465da5 // vadd.f32 d21, d22, d21 |
| .long 0xf342dd90 // vmul.f32 d29, d18, d0 |
| .long 0xf3210eaa // vcgt.f32 d0, d17, d26 |
| .long 0xf3492db2 // vmul.f32 d18, d25, d18 |
| .long 0xf355b1b6 // vbsl d27, d21, d22 |
| .long 0xeddf5b22 // vldr d21, [pc, #136] |
| .long 0xf36cfea8 // vcgt.f32 d31, d28, d24 |
| .long 0xf2463da3 // vadd.f32 d19, d22, d19 |
| .long 0xf36cceae // vcgt.f32 d28, d28, d30 |
| .long 0xf2464da4 // vadd.f32 d20, d22, d20 |
| .long 0xf365aeaa // vcgt.f32 d26, d21, d26 |
| .long 0xf2467da7 // vadd.f32 d23, d22, d23 |
| .long 0xf3619ea8 // vcgt.f32 d25, d17, d24 |
| .long 0xf3611eae // vcgt.f32 d17, d17, d30 |
| .long 0xf31001bb // vbsl d0, d16, d27 |
| .long 0xf353f1b6 // vbsl d31, d19, d22 |
| .long 0xf354c1b6 // vbsl d28, d20, d22 |
| .long 0xf357a190 // vbsl d26, d23, d0 |
| .long 0xf3b90501 // vceq.f32 d0, d1, #0 |
| .long 0xf3658ea8 // vcgt.f32 d24, d21, d24 |
| .long 0xf246ddad // vadd.f32 d29, d22, d29 |
| .long 0xf3653eae // vcgt.f32 d19, d21, d30 |
| .long 0xf2462da2 // vadd.f32 d18, d22, d18 |
| .long 0xf35091bf // vbsl d25, d16, d31 |
| .long 0xf35011bc // vbsl d17, d16, d28 |
| .long 0xf2600110 // vorr d16, d0, d0 |
| .long 0xf2201110 // vorr d1, d0, d0 |
| .long 0xf352013a // vbsl d16, d2, d26 |
| .long 0xf35d81b9 // vbsl d24, d29, d25 |
| .long 0xf35231b1 // vbsl d19, d18, d17 |
| .long 0xf3121138 // vbsl d1, d2, d24 |
| .long 0xf3120133 // vbsl d0, d2, d19 |
| .long 0xf22021b0 // vorr d2, d16, d16 |
| .long 0xe12fff13 // bx r3 |
| .long 0xe320f000 // nop {0} |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| .long 0x3eaaaaab // .word 0x3eaaaaab |
| .long 0x3eaaaaab // .word 0x3eaaaaab |
| .long 0x3f2aaaab // .word 0x3f2aaaab |
| .long 0x3f2aaaab // .word 0x3f2aaaab |
| .long 0x3e2aaaab // .word 0x3e2aaaab |
| .long 0x3e2aaaab // .word 0x3e2aaaab |
| |
| HIDDEN _sk_scale_1_float_vfp4 |
| .globl _sk_scale_1_float_vfp4 |
| FUNCTION(_sk_scale_1_float_vfp4) |
| _sk_scale_1_float_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xf3000d90 // vmul.f32 d0, d16, d0 |
| .long 0xf3001d91 // vmul.f32 d1, d16, d1 |
| .long 0xf3002d92 // vmul.f32 d2, d16, d2 |
| .long 0xf3003d93 // vmul.f32 d3, d16, d3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_scale_u8_vfp4 |
| .globl _sk_scale_u8_vfp4 |
| FUNCTION(_sk_scale_u8_vfp4) |
| _sk_scale_u8_vfp4: |
| .long 0xe24dd004 // sub sp, sp, #4 |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .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 0xeddf1b06 // vldr d17, [pc, #24] |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf3000d90 // vmul.f32 d0, d16, d0 |
| .long 0xf3001d91 // vmul.f32 d1, d16, d1 |
| .long 0xf3002d92 // vmul.f32 d2, d16, d2 |
| .long 0xf3003d93 // vmul.f32 d3, d16, d3 |
| .long 0xe28dd004 // add sp, sp, #4 |
| .long 0xe12fff1c // bx ip |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_lerp_1_float_vfp4 |
| .globl _sk_lerp_1_float_vfp4 |
| FUNCTION(_sk_lerp_1_float_vfp4) |
| _sk_lerp_1_float_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2600d04 // vsub.f32 d16, d0, d4 |
| .long 0xf2611d05 // vsub.f32 d17, d1, d5 |
| .long 0xf2622d06 // vsub.f32 d18, d2, d6 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2633d07 // vsub.f32 d19, d3, d7 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf2240114 // vorr d0, d4, d4 |
| .long 0xf2251115 // vorr d1, d5, d5 |
| .long 0xf2262116 // vorr d2, d6, d6 |
| .long 0xf2273117 // vorr d3, d7, d7 |
| .long 0xf2000cb4 // vfma.f32 d0, d16, d20 |
| .long 0xf2011cb4 // vfma.f32 d1, d17, d20 |
| .long 0xf2022cb4 // vfma.f32 d2, d18, d20 |
| .long 0xf2033cb4 // vfma.f32 d3, d19, d20 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_lerp_u8_vfp4 |
| .globl _sk_lerp_u8_vfp4 |
| FUNCTION(_sk_lerp_u8_vfp4) |
| _sk_lerp_u8_vfp4: |
| .long 0xe24dd004 // sub sp, sp, #4 |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2602d04 // vsub.f32 d18, d0, d4 |
| .long 0xf2623d06 // vsub.f32 d19, d2, d6 |
| .long 0xf2634d07 // vsub.f32 d20, d3, d7 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf2240114 // vorr d0, d4, d4 |
| .long 0xf2262116 // vorr d2, d6, d6 |
| .long 0xe0833000 // add r3, r3, r0 |
| .long 0xf2273117 // vorr d3, d7, d7 |
| .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 0xeddf1b08 // vldr d17, [pc, #32] |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf2611d05 // vsub.f32 d17, d1, d5 |
| .long 0xf2251115 // vorr d1, d5, d5 |
| .long 0xf2020cb0 // vfma.f32 d0, d18, d16 |
| .long 0xf2011cb0 // vfma.f32 d1, d17, d16 |
| .long 0xf2032cb0 // vfma.f32 d2, d19, d16 |
| .long 0xf2043cb0 // vfma.f32 d3, d20, d16 |
| .long 0xe28dd004 // add sp, sp, #4 |
| .long 0xe12fff1c // bx ip |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_lerp_565_vfp4 |
| .globl _sk_lerp_565_vfp4 |
| FUNCTION(_sk_lerp_565_vfp4) |
| _sk_lerp_565_vfp4: |
| .long 0xe24dd004 // sub sp, sp, #4 |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3c72218 // vmov.i32 d18, #63488 |
| .long 0xf2c1101f // vmov.i32 d17, #31 |
| .long 0xf2603d04 // vsub.f32 d19, d0, d4 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf2616d05 // vsub.f32 d22, d1, d5 |
| .long 0xf2240114 // vorr d0, d4, d4 |
| .long 0xf2251115 // vorr d1, d5, d5 |
| .long 0xe7933080 // ldr r3, [r3, r0, lsl #1] |
| .long 0xf2873f10 // vmov.f32 d3, #1 |
| .long 0xe58d3000 // str r3, [sp] |
| .long 0xe1a0300d // mov r3, sp |
| .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32] |
| .long 0xe3a03e7e // mov r3, #2016 |
| .long 0xf3d04a30 // vmovl.u16 q10, d16 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf24421b2 // vand d18, d20, d18 |
| .long 0xf24411b1 // vand d17, d20, d17 |
| .long 0xeddf5b12 // vldr d21, [pc, #72] |
| .long 0xf24401b0 // vand d16, d20, d16 |
| .long 0xeddf4b0e // vldr d20, [pc, #56] |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3422db4 // vmul.f32 d18, d18, d20 |
| .long 0xeddf4b0d // vldr d20, [pc, #52] |
| .long 0xf3400db5 // vmul.f32 d16, d16, d21 |
| .long 0xf2625d06 // vsub.f32 d21, d2, d6 |
| .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| .long 0xf2262116 // vorr d2, d6, d6 |
| .long 0xf2030cb2 // vfma.f32 d0, d19, d18 |
| .long 0xf2061cb0 // vfma.f32 d1, d22, d16 |
| .long 0xf2052cb1 // vfma.f32 d2, d21, d17 |
| .long 0xe28dd004 // add sp, sp, #4 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x37842108 // .word 0x37842108 |
| .long 0x37842108 // .word 0x37842108 |
| .long 0x3a020821 // .word 0x3a020821 |
| .long 0x3a020821 // .word 0x3a020821 |
| .long 0x3d042108 // .word 0x3d042108 |
| .long 0x3d042108 // .word 0x3d042108 |
| |
| HIDDEN _sk_load_tables_vfp4 |
| .globl _sk_load_tables_vfp4 |
| FUNCTION(_sk_load_tables_vfp4) |
| _sk_load_tables_vfp4: |
| .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3c7001f // vmov.i32 d16, #255 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe99300b0 // ldmib r3, {r4, r5, r7} |
| .long 0xe08e3100 // add r3, lr, r0, lsl #2 |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xf24121b0 // vand d18, d17, d16 |
| .long 0xf3f83031 // vshr.u32 d19, d17, #8 |
| .long 0xee323b90 // vmov.32 r3, d18[1] |
| .long 0xee126b90 // vmov.32 r6, d18[0] |
| .long 0xf3f02031 // vshr.u32 d18, d17, #16 |
| .long 0xf24221b0 // vand d18, d18, d16 |
| .long 0xf24301b0 // vand d16, d19, d16 |
| .long 0xe0843103 // add r3, r4, r3, lsl #2 |
| .long 0xedd30a00 // vldr s1, [r3] |
| .long 0xe0843106 // add r3, r4, r6, lsl #2 |
| .long 0xee326b90 // vmov.32 r6, d18[1] |
| .long 0xed930a00 // vldr s0, [r3] |
| .long 0xee303b90 // vmov.32 r3, d16[1] |
| .long 0xee104b90 // vmov.32 r4, d16[0] |
| .long 0xf3e80031 // vshr.u32 d16, d17, #24 |
| .long 0xeddf1b0d // vldr d17, [pc, #52] |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| .long 0xe087e106 // add lr, r7, r6, lsl #2 |
| .long 0xee126b90 // vmov.32 r6, d18[0] |
| .long 0xe0853103 // add r3, r5, r3, lsl #2 |
| .long 0xedde2a00 // vldr s5, [lr] |
| .long 0xedd31a00 // vldr s3, [r3] |
| .long 0xe0853104 // add r3, r5, r4, lsl #2 |
| .long 0xed931a00 // vldr s2, [r3] |
| .long 0xe0873106 // add r3, r7, r6, lsl #2 |
| .long 0xed932a00 // vldr s4, [r3] |
| .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_load_tables_u16_be_vfp4 |
| .globl _sk_load_tables_u16_be_vfp4 |
| FUNCTION(_sk_load_tables_u16_be_vfp4) |
| _sk_load_tables_u16_be_vfp4: |
| .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3c7801f // vmov.i32 d24, #255 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5936000 // ldr r6, [r3] |
| .long 0xe9934030 // ldmib r3, {r4, r5, lr} |
| .long 0xe0863180 // add r3, r6, r0, lsl #3 |
| .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]! |
| .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3] |
| .long 0xee903bb0 // vmov.u16 r3, d16[0] |
| .long 0xee926bb0 // vmov.u16 r6, d18[0] |
| .long 0xee043b90 // vmov.32 d20[0], r3 |
| .long 0xee913bb0 // vmov.u16 r3, d17[0] |
| .long 0xee056b90 // vmov.32 d21[0], r6 |
| .long 0xee906bf0 // vmov.u16 r6, d16[1] |
| .long 0xee063b90 // vmov.32 d22[0], r3 |
| .long 0xee933bb0 // vmov.u16 r3, d19[0] |
| .long 0xee246b90 // vmov.32 d20[1], r6 |
| .long 0xee926bf0 // vmov.u16 r6, d18[1] |
| .long 0xf24441b8 // vand d20, d20, d24 |
| .long 0xee073b90 // vmov.32 d23[0], r3 |
| .long 0xee913bf0 // vmov.u16 r3, d17[1] |
| .long 0xee256b90 // vmov.32 d21[1], r6 |
| .long 0xee936bf0 // vmov.u16 r6, d19[1] |
| .long 0xf24501b8 // vand d16, d21, d24 |
| .long 0xee107b90 // vmov.32 r7, d16[0] |
| .long 0xee263b90 // vmov.32 d22[1], r3 |
| .long 0xee343b90 // vmov.32 r3, d20[1] |
| .long 0xf24611b8 // vand d17, d22, d24 |
| .long 0xee276b90 // vmov.32 d23[1], r6 |
| .long 0xee146b90 // vmov.32 r6, d20[0] |
| .long 0xf2e82537 // vshl.s32 d18, d23, #8 |
| .long 0xe0843103 // add r3, r4, r3, lsl #2 |
| .long 0xedd30a00 // vldr s1, [r3] |
| .long 0xe0843106 // add r3, r4, r6, lsl #2 |
| .long 0xee304b90 // vmov.32 r4, d16[1] |
| .long 0xee116b90 // vmov.32 r6, d17[0] |
| .long 0xf3c70c1f // vmov.i32 d16, #65535 |
| .long 0xed930a00 // vldr s0, [r3] |
| .long 0xee313b90 // vmov.32 r3, d17[1] |
| .long 0xf24711b0 // vand d17, d23, d16 |
| .long 0xf3f81031 // vshr.u32 d17, d17, #8 |
| .long 0xf26211b1 // vorr d17, d18, d17 |
| .long 0xf24101b0 // vand d16, d17, d16 |
| .long 0xeddf1b0c // vldr d17, [pc, #48] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| .long 0xe08e4104 // add r4, lr, r4, lsl #2 |
| .long 0xedd42a00 // vldr s5, [r4] |
| .long 0xe0853103 // add r3, r5, r3, lsl #2 |
| .long 0xedd31a00 // vldr s3, [r3] |
| .long 0xe0853106 // add r3, r5, r6, lsl #2 |
| .long 0xed931a00 // vldr s2, [r3] |
| .long 0xe08e3107 // add r3, lr, r7, lsl #2 |
| .long 0xed932a00 // vldr s4, [r3] |
| .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x37800080 // .word 0x37800080 |
| .long 0x37800080 // .word 0x37800080 |
| |
| HIDDEN _sk_load_tables_rgb_u16_be_vfp4 |
| .globl _sk_load_tables_rgb_u16_be_vfp4 |
| FUNCTION(_sk_load_tables_rgb_u16_be_vfp4) |
| _sk_load_tables_rgb_u16_be_vfp4: |
| .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3c7601f // vmov.i32 d22, #255 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2873f10 // vmov.f32 d3, #1 |
| .long 0xe5936000 // ldr r6, [r3] |
| .long 0xe9934030 // ldmib r3, {r4, r5, lr} |
| .long 0xe0803080 // add r3, r0, r0, lsl #1 |
| .long 0xe0863083 // add r3, r6, r3, lsl #1 |
| .long 0xf4e3060d // vld3.16 {d16[0],d17[0],d18[0]}, [r3]! |
| .long 0xf4e3064f // vld3.16 {d16[1],d17[1],d18[1]}, [r3] |
| .long 0xee903bb0 // vmov.u16 r3, d16[0] |
| .long 0xee926bb0 // vmov.u16 r6, d18[0] |
| .long 0xee043b90 // vmov.32 d20[0], r3 |
| .long 0xee903bf0 // vmov.u16 r3, d16[1] |
| .long 0xee056b90 // vmov.32 d21[0], r6 |
| .long 0xee916bb0 // vmov.u16 r6, d17[0] |
| .long 0xee243b90 // vmov.32 d20[1], r3 |
| .long 0xee923bf0 // vmov.u16 r3, d18[1] |
| .long 0xf24441b6 // vand d20, d20, d22 |
| .long 0xee076b90 // vmov.32 d23[0], r6 |
| .long 0xee916bf0 // vmov.u16 r6, d17[1] |
| .long 0xee253b90 // vmov.32 d21[1], r3 |
| .long 0xee343b90 // vmov.32 r3, d20[1] |
| .long 0xf24501b6 // vand d16, d21, d22 |
| .long 0xee276b90 // vmov.32 d23[1], r6 |
| .long 0xee146b90 // vmov.32 r6, d20[0] |
| .long 0xf24711b6 // vand d17, d23, d22 |
| .long 0xee107b90 // vmov.32 r7, d16[0] |
| .long 0xe0843103 // add r3, r4, r3, lsl #2 |
| .long 0xedd30a00 // vldr s1, [r3] |
| .long 0xe0843106 // add r3, r4, r6, lsl #2 |
| .long 0xee304b90 // vmov.32 r4, d16[1] |
| .long 0xee116b90 // vmov.32 r6, d17[0] |
| .long 0xed930a00 // vldr s0, [r3] |
| .long 0xee313b90 // vmov.32 r3, d17[1] |
| .long 0xe08e4104 // add r4, lr, r4, lsl #2 |
| .long 0xedd42a00 // vldr s5, [r4] |
| .long 0xe0853103 // add r3, r5, r3, lsl #2 |
| .long 0xedd31a00 // vldr s3, [r3] |
| .long 0xe0853106 // add r3, r5, r6, lsl #2 |
| .long 0xed931a00 // vldr s2, [r3] |
| .long 0xe08e3107 // add r3, lr, r7, lsl #2 |
| .long 0xed932a00 // vldr s4, [r3] |
| .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_byte_tables_vfp4 |
| .globl _sk_byte_tables_vfp4 |
| FUNCTION(_sk_byte_tables_vfp4) |
| _sk_byte_tables_vfp4: |
| .long 0xe92d4bf0 // push {r4, r5, r6, r7, r8, r9, fp, lr} |
| .long 0xeddf0b37 // vldr d16, [pc, #220] |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| .long 0xe8911010 // ldm r1, {r4, ip} |
| .long 0xf2422c30 // vfma.f32 d18, d2, d16 |
| .long 0xf2413c30 // vfma.f32 d19, d1, d16 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| .long 0xe89402e0 // ldm r4, {r5, r6, r7, r9} |
| .long 0xf2404c30 // vfma.f32 d20, d0, d16 |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xf2431c30 // vfma.f32 d17, d3, d16 |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf3fb37a3 // vcvt.u32.f32 d19, d19 |
| .long 0xf3fb47a4 // vcvt.u32.f32 d20, d20 |
| .long 0xee123b90 // vmov.32 r3, d18[0] |
| .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| .long 0xee13eb90 // vmov.32 lr, d19[0] |
| .long 0xee144b90 // vmov.32 r4, d20[0] |
| .long 0xe7d78003 // ldrb r8, [r7, r3] |
| .long 0xe7d6300e // ldrb r3, [r6, lr] |
| .long 0xee053b90 // vmov.32 d21[0], r3 |
| .long 0xe7d53004 // ldrb r3, [r5, r4] |
| .long 0xee344b90 // vmov.32 r4, d20[1] |
| .long 0xee013b90 // vmov.32 d17[0], r3 |
| .long 0xee103b90 // vmov.32 r3, d16[0] |
| .long 0xee048b90 // vmov.32 d20[0], r8 |
| .long 0xe7d5e004 // ldrb lr, [r5, r4] |
| .long 0xee334b90 // vmov.32 r4, d19[1] |
| .long 0xee325b90 // vmov.32 r5, d18[1] |
| .long 0xf3c7201f // vmov.i32 d18, #255 |
| .long 0xe7d93003 // ldrb r3, [r9, r3] |
| .long 0xee21eb90 // vmov.32 d17[1], lr |
| .long 0xf24111b2 // vand d17, d17, d18 |
| .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17 |
| .long 0xe7d64004 // ldrb r4, [r6, r4] |
| .long 0xee306b90 // vmov.32 r6, d16[1] |
| .long 0xee003b90 // vmov.32 d16[0], r3 |
| .long 0xee254b90 // vmov.32 d21[1], r4 |
| .long 0xe7d74005 // ldrb r4, [r7, r5] |
| .long 0xf24531b2 // vand d19, d21, d18 |
| .long 0xee244b90 // vmov.32 d20[1], r4 |
| .long 0xf24441b2 // vand d20, d20, d18 |
| .long 0xf3fb46a4 // vcvt.f32.u32 d20, d20 |
| .long 0xe7d93006 // ldrb r3, [r9, r6] |
| .long 0xee203b90 // vmov.32 d16[1], r3 |
| .long 0xf24001b2 // vand d16, d16, d18 |
| .long 0xf3fb26a3 // vcvt.f32.u32 d18, d19 |
| .long 0xeddf3b09 // vldr d19, [pc, #36] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3010db3 // vmul.f32 d0, d17, d19 |
| .long 0xf3042db3 // vmul.f32 d2, d20, d19 |
| .long 0xf3021db3 // vmul.f32 d1, d18, d19 |
| .long 0xf3003db3 // vmul.f32 d3, d16, d19 |
| .long 0xe8bd4bf0 // pop {r4, r5, r6, r7, r8, r9, fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x437f0000 // .word 0x437f0000 |
| .long 0x437f0000 // .word 0x437f0000 |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_byte_tables_rgb_vfp4 |
| .globl _sk_byte_tables_rgb_vfp4 |
| FUNCTION(_sk_byte_tables_rgb_vfp4) |
| _sk_byte_tables_rgb_vfp4: |
| .long 0xe92d41f0 // push {r4, r5, r6, r7, r8, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xf3c7301f // vmov.i32 d19, #255 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe9930110 // ldmib r3, {r4, r8} |
| .long 0xe593300c // ldr r3, [r3, #12] |
| .long 0xe2433001 // sub r3, r3, #1 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf2402c30 // vfma.f32 d18, d0, d16 |
| .long 0xf2411c30 // vfma.f32 d17, d1, d16 |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| .long 0xee123b90 // vmov.32 r3, d18[0] |
| .long 0xee326b90 // vmov.32 r6, d18[1] |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2422c30 // vfma.f32 d18, d2, d16 |
| .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18 |
| .long 0xee107b90 // vmov.32 r7, d16[0] |
| .long 0xee305b90 // vmov.32 r5, d16[1] |
| .long 0xe7de3003 // ldrb r3, [lr, r3] |
| .long 0xe7dee006 // ldrb lr, [lr, r6] |
| .long 0xee116b90 // vmov.32 r6, d17[0] |
| .long 0xee023b90 // vmov.32 d18[0], r3 |
| .long 0xee313b90 // vmov.32 r3, d17[1] |
| .long 0xee22eb90 // vmov.32 d18[1], lr |
| .long 0xf24221b3 // vand d18, d18, d19 |
| .long 0xf3fb26a2 // vcvt.f32.u32 d18, d18 |
| .long 0xe7d87007 // ldrb r7, [r8, r7] |
| .long 0xee017b90 // vmov.32 d17[0], r7 |
| .long 0xe7d46006 // ldrb r6, [r4, r6] |
| .long 0xe7d43003 // ldrb r3, [r4, r3] |
| .long 0xee006b90 // vmov.32 d16[0], r6 |
| .long 0xe7d84005 // ldrb r4, [r8, r5] |
| .long 0xee203b90 // vmov.32 d16[1], r3 |
| .long 0xee214b90 // vmov.32 d17[1], r4 |
| .long 0xf24001b3 // vand d16, d16, d19 |
| .long 0xf24111b3 // vand d17, d17, d19 |
| .long 0xeddf3b07 // vldr d19, [pc, #28] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17 |
| .long 0xf3020db3 // vmul.f32 d0, d18, d19 |
| .long 0xf3001db3 // vmul.f32 d1, d16, d19 |
| .long 0xf3012db3 // vmul.f32 d2, d17, d19 |
| .long 0xe8bd41f0 // pop {r4, r5, r6, r7, r8, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_table_r_vfp4 |
| .globl _sk_table_r_vfp4 |
| FUNCTION(_sk_table_r_vfp4) |
| _sk_table_r_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe5933004 // ldr r3, [r3, #4] |
| .long 0xe2433001 // sub r3, r3, #1 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| .long 0xee303b90 // vmov.32 r3, d16[1] |
| .long 0xee104b90 // vmov.32 r4, d16[0] |
| .long 0xe08e3103 // add r3, lr, r3, lsl #2 |
| .long 0xedd30a00 // vldr s1, [r3] |
| .long 0xe08e3104 // add r3, lr, r4, lsl #2 |
| .long 0xed930a00 // vldr s0, [r3] |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_table_g_vfp4 |
| .globl _sk_table_g_vfp4 |
| FUNCTION(_sk_table_g_vfp4) |
| _sk_table_g_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe5933004 // ldr r3, [r3, #4] |
| .long 0xe2433001 // sub r3, r3, #1 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf2411c30 // vfma.f32 d17, d1, d16 |
| .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| .long 0xee303b90 // vmov.32 r3, d16[1] |
| .long 0xee104b90 // vmov.32 r4, d16[0] |
| .long 0xe08e3103 // add r3, lr, r3, lsl #2 |
| .long 0xedd31a00 // vldr s3, [r3] |
| .long 0xe08e3104 // add r3, lr, r4, lsl #2 |
| .long 0xed931a00 // vldr s2, [r3] |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_table_b_vfp4 |
| .globl _sk_table_b_vfp4 |
| FUNCTION(_sk_table_b_vfp4) |
| _sk_table_b_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe5933004 // ldr r3, [r3, #4] |
| .long 0xe2433001 // sub r3, r3, #1 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf2421c30 // vfma.f32 d17, d2, d16 |
| .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| .long 0xee303b90 // vmov.32 r3, d16[1] |
| .long 0xee104b90 // vmov.32 r4, d16[0] |
| .long 0xe08e3103 // add r3, lr, r3, lsl #2 |
| .long 0xedd32a00 // vldr s5, [r3] |
| .long 0xe08e3104 // add r3, lr, r4, lsl #2 |
| .long 0xed932a00 // vldr s4, [r3] |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_table_a_vfp4 |
| .globl _sk_table_a_vfp4 |
| FUNCTION(_sk_table_a_vfp4) |
| _sk_table_a_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe5933004 // ldr r3, [r3, #4] |
| .long 0xe2433001 // sub r3, r3, #1 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf2431c30 // vfma.f32 d17, d3, d16 |
| .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| .long 0xee303b90 // vmov.32 r3, d16[1] |
| .long 0xee104b90 // vmov.32 r4, d16[0] |
| .long 0xe08e3103 // add r3, lr, r3, lsl #2 |
| .long 0xedd33a00 // vldr s7, [r3] |
| .long 0xe08e3104 // add r3, lr, r4, lsl #2 |
| .long 0xed933a00 // vldr s6, [r3] |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_parametric_r_vfp4 |
| .globl _sk_parametric_r_vfp4 |
| FUNCTION(_sk_parametric_r_vfp4) |
| _sk_parametric_r_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xed2d8b06 // vpush {d8-d10} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xeddf3b41 // vldr d19, [pc, #260] |
| .long 0xed9f8a4e // vldr s16, [pc, #312] |
| .long 0xe1a0300e // mov r3, lr |
| .long 0xeddf4b42 // vldr d20, [pc, #264] |
| .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf2412c90 // vfma.f32 d18, d17, d0 |
| .long 0xf2c71d1f // vmov.i32 d17, #8388607 |
| .long 0xf24211b1 // vand d17, d18, d17 |
| .long 0xf2c3171f // vorr.i32 d17, #1056964608 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf2019da3 // vadd.f32 d9, d17, d19 |
| .long 0xf2c33614 // vmov.i32 d19, #872415232 |
| .long 0xf3422db3 // vmul.f32 d18, d18, d19 |
| .long 0xeddf3b30 // vldr d19, [pc, #192] |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf3411db3 // vmul.f32 d17, d17, d19 |
| .long 0xed9f8a39 // vldr s16, [pc, #228] |
| .long 0xf2422da4 // vadd.f32 d18, d18, d20 |
| .long 0xeddf4b2e // vldr d20, [pc, #184] |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2621da1 // vsub.f32 d17, d18, d17 |
| .long 0xf2611d8a // vsub.f32 d17, d17, d10 |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf3fb1720 // vcvt.s32.f32 d17, d16 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 |
| .long 0xf35421b3 // vbsl d18, d20, d19 |
| .long 0xeddf4b2b // vldr d20, [pc, #172] |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xeddf2b25 // vldr d18, [pc, #148] |
| .long 0xf2601da1 // vsub.f32 d17, d16, d17 |
| .long 0xf2400da4 // vadd.f32 d16, d16, d20 |
| .long 0xf2229da1 // vsub.f32 d9, d18, d17 |
| .long 0xeddf2b23 // vldr d18, [pc, #140] |
| .long 0xf3411db2 // vmul.f32 d17, d17, d18 |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| .long 0xf2c4161b // vmov.i32 d17, #1258291200 |
| .long 0xf2400d8a // vadd.f32 d16, d16, d10 |
| .long 0xf2402cb1 // vfma.f32 d18, d16, d17 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3018 // add r3, lr, #24 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf2401c90 // vfma.f32 d17, d16, d0 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf3400e80 // vcge.f32 d16, d16, d0 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf2442da2 // vadd.f32 d18, d20, d18 |
| .long 0xf35101b2 // vbsl d16, d17, d18 |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xf2400fa3 // vmax.f32 d16, d16, d19 |
| .long 0xf2200fa1 // vmin.f32 d0, d16, d17 |
| .long 0xecbd8b06 // vpop {d8-d10} |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x3fdce9a3 // .word 0x3fdce9a3 |
| .long 0x41ddd2fe // .word 0x41ddd2fe |
| |
| HIDDEN _sk_parametric_g_vfp4 |
| .globl _sk_parametric_g_vfp4 |
| FUNCTION(_sk_parametric_g_vfp4) |
| _sk_parametric_g_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xed2d8b06 // vpush {d8-d10} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xeddf3b41 // vldr d19, [pc, #260] |
| .long 0xed9f8a4e // vldr s16, [pc, #312] |
| .long 0xe1a0300e // mov r3, lr |
| .long 0xeddf4b42 // vldr d20, [pc, #264] |
| .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf2412c91 // vfma.f32 d18, d17, d1 |
| .long 0xf2c71d1f // vmov.i32 d17, #8388607 |
| .long 0xf24211b1 // vand d17, d18, d17 |
| .long 0xf2c3171f // vorr.i32 d17, #1056964608 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf2019da3 // vadd.f32 d9, d17, d19 |
| .long 0xf2c33614 // vmov.i32 d19, #872415232 |
| .long 0xf3422db3 // vmul.f32 d18, d18, d19 |
| .long 0xeddf3b30 // vldr d19, [pc, #192] |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf3411db3 // vmul.f32 d17, d17, d19 |
| .long 0xed9f8a39 // vldr s16, [pc, #228] |
| .long 0xf2422da4 // vadd.f32 d18, d18, d20 |
| .long 0xeddf4b2e // vldr d20, [pc, #184] |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2621da1 // vsub.f32 d17, d18, d17 |
| .long 0xf2611d8a // vsub.f32 d17, d17, d10 |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf3fb1720 // vcvt.s32.f32 d17, d16 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 |
| .long 0xf35421b3 // vbsl d18, d20, d19 |
| .long 0xeddf4b2b // vldr d20, [pc, #172] |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xeddf2b25 // vldr d18, [pc, #148] |
| .long 0xf2601da1 // vsub.f32 d17, d16, d17 |
| .long 0xf2400da4 // vadd.f32 d16, d16, d20 |
| .long 0xf2229da1 // vsub.f32 d9, d18, d17 |
| .long 0xeddf2b23 // vldr d18, [pc, #140] |
| .long 0xf3411db2 // vmul.f32 d17, d17, d18 |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| .long 0xf2c4161b // vmov.i32 d17, #1258291200 |
| .long 0xf2400d8a // vadd.f32 d16, d16, d10 |
| .long 0xf2402cb1 // vfma.f32 d18, d16, d17 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3018 // add r3, lr, #24 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf2401c91 // vfma.f32 d17, d16, d1 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf3400e81 // vcge.f32 d16, d16, d1 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf2442da2 // vadd.f32 d18, d20, d18 |
| .long 0xf35101b2 // vbsl d16, d17, d18 |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xf2400fa3 // vmax.f32 d16, d16, d19 |
| .long 0xf2201fa1 // vmin.f32 d1, d16, d17 |
| .long 0xecbd8b06 // vpop {d8-d10} |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x3fdce9a3 // .word 0x3fdce9a3 |
| .long 0x41ddd2fe // .word 0x41ddd2fe |
| |
| HIDDEN _sk_parametric_b_vfp4 |
| .globl _sk_parametric_b_vfp4 |
| FUNCTION(_sk_parametric_b_vfp4) |
| _sk_parametric_b_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xed2d8b06 // vpush {d8-d10} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xeddf3b41 // vldr d19, [pc, #260] |
| .long 0xed9f8a4e // vldr s16, [pc, #312] |
| .long 0xe1a0300e // mov r3, lr |
| .long 0xeddf4b42 // vldr d20, [pc, #264] |
| .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf2412c92 // vfma.f32 d18, d17, d2 |
| .long 0xf2c71d1f // vmov.i32 d17, #8388607 |
| .long 0xf24211b1 // vand d17, d18, d17 |
| .long 0xf2c3171f // vorr.i32 d17, #1056964608 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf2019da3 // vadd.f32 d9, d17, d19 |
| .long 0xf2c33614 // vmov.i32 d19, #872415232 |
| .long 0xf3422db3 // vmul.f32 d18, d18, d19 |
| .long 0xeddf3b30 // vldr d19, [pc, #192] |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf3411db3 // vmul.f32 d17, d17, d19 |
| .long 0xed9f8a39 // vldr s16, [pc, #228] |
| .long 0xf2422da4 // vadd.f32 d18, d18, d20 |
| .long 0xeddf4b2e // vldr d20, [pc, #184] |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2621da1 // vsub.f32 d17, d18, d17 |
| .long 0xf2611d8a // vsub.f32 d17, d17, d10 |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf3fb1720 // vcvt.s32.f32 d17, d16 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 |
| .long 0xf35421b3 // vbsl d18, d20, d19 |
| .long 0xeddf4b2b // vldr d20, [pc, #172] |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xeddf2b25 // vldr d18, [pc, #148] |
| .long 0xf2601da1 // vsub.f32 d17, d16, d17 |
| .long 0xf2400da4 // vadd.f32 d16, d16, d20 |
| .long 0xf2229da1 // vsub.f32 d9, d18, d17 |
| .long 0xeddf2b23 // vldr d18, [pc, #140] |
| .long 0xf3411db2 // vmul.f32 d17, d17, d18 |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| .long 0xf2c4161b // vmov.i32 d17, #1258291200 |
| .long 0xf2400d8a // vadd.f32 d16, d16, d10 |
| .long 0xf2402cb1 // vfma.f32 d18, d16, d17 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3018 // add r3, lr, #24 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf2401c92 // vfma.f32 d17, d16, d2 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf3400e82 // vcge.f32 d16, d16, d2 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf2442da2 // vadd.f32 d18, d20, d18 |
| .long 0xf35101b2 // vbsl d16, d17, d18 |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xf2400fa3 // vmax.f32 d16, d16, d19 |
| .long 0xf2202fa1 // vmin.f32 d2, d16, d17 |
| .long 0xecbd8b06 // vpop {d8-d10} |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x3fdce9a3 // .word 0x3fdce9a3 |
| .long 0x41ddd2fe // .word 0x41ddd2fe |
| |
| HIDDEN _sk_parametric_a_vfp4 |
| .globl _sk_parametric_a_vfp4 |
| FUNCTION(_sk_parametric_a_vfp4) |
| _sk_parametric_a_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xed2d8b06 // vpush {d8-d10} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xeddf3b41 // vldr d19, [pc, #260] |
| .long 0xed9f8a4e // vldr s16, [pc, #312] |
| .long 0xe1a0300e // mov r3, lr |
| .long 0xeddf4b42 // vldr d20, [pc, #264] |
| .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf2412c93 // vfma.f32 d18, d17, d3 |
| .long 0xf2c71d1f // vmov.i32 d17, #8388607 |
| .long 0xf24211b1 // vand d17, d18, d17 |
| .long 0xf2c3171f // vorr.i32 d17, #1056964608 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf2019da3 // vadd.f32 d9, d17, d19 |
| .long 0xf2c33614 // vmov.i32 d19, #872415232 |
| .long 0xf3422db3 // vmul.f32 d18, d18, d19 |
| .long 0xeddf3b30 // vldr d19, [pc, #192] |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf3411db3 // vmul.f32 d17, d17, d19 |
| .long 0xed9f8a39 // vldr s16, [pc, #228] |
| .long 0xf2422da4 // vadd.f32 d18, d18, d20 |
| .long 0xeddf4b2e // vldr d20, [pc, #184] |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2621da1 // vsub.f32 d17, d18, d17 |
| .long 0xf2611d8a // vsub.f32 d17, d17, d10 |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf3fb1720 // vcvt.s32.f32 d17, d16 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3612ea0 // vcgt.f32 d18, d17, d16 |
| .long 0xf35421b3 // vbsl d18, d20, d19 |
| .long 0xeddf4b2b // vldr d20, [pc, #172] |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xeddf2b25 // vldr d18, [pc, #148] |
| .long 0xf2601da1 // vsub.f32 d17, d16, d17 |
| .long 0xf2400da4 // vadd.f32 d16, d16, d20 |
| .long 0xf2229da1 // vsub.f32 d9, d18, d17 |
| .long 0xeddf2b23 // vldr d18, [pc, #140] |
| .long 0xf3411db2 // vmul.f32 d17, d17, d18 |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xeec8aa29 // vdiv.f32 s21, s16, s19 |
| .long 0xee88aa09 // vdiv.f32 s20, s16, s18 |
| .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| .long 0xf2c4161b // vmov.i32 d17, #1258291200 |
| .long 0xf2400d8a // vadd.f32 d16, d16, d10 |
| .long 0xf2402cb1 // vfma.f32 d18, d16, d17 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3018 // add r3, lr, #24 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf2401c93 // vfma.f32 d17, d16, d3 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf3400e83 // vcge.f32 d16, d16, d3 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf2442da2 // vadd.f32 d18, d20, d18 |
| .long 0xf35101b2 // vbsl d16, d17, d18 |
| .long 0xf2c71f10 // vmov.f32 d17, #1 |
| .long 0xf2400fa3 // vmax.f32 d16, d16, d19 |
| .long 0xf2203fa1 // vmin.f32 d3, d16, d17 |
| .long 0xecbd8b06 // vpop {d8-d10} |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3eb444f9 // .word 0x3eb444f9 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0x3fbfbf75 // .word 0x3fbfbf75 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0xc2f87377 // .word 0xc2f87377 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x409af5f8 // .word 0x409af5f8 |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x3fbebc8d // .word 0x3fbebc8d |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x42f28c51 // .word 0x42f28c51 |
| .long 0x3fdce9a3 // .word 0x3fdce9a3 |
| .long 0x41ddd2fe // .word 0x41ddd2fe |
| |
| HIDDEN _sk_load_a8_vfp4 |
| .globl _sk_load_a8_vfp4 |
| FUNCTION(_sk_load_a8_vfp4) |
| _sk_load_a8_vfp4: |
| .long 0xe24dd004 // sub sp, sp, #4 |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2800010 // vmov.i32 d0, #0 |
| .long 0xf2801010 // vmov.i32 d1, #0 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf2802010 // vmov.i32 d2, #0 |
| .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 0xeddf1b03 // vldr d17, [pc, #12] |
| .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| .long 0xe28dd004 // add sp, sp, #4 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_gather_a8_vfp4 |
| .globl _sk_gather_a8_vfp4 |
| FUNCTION(_sk_gather_a8_vfp4) |
| _sk_gather_a8_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2800010 // vmov.i32 d0, #0 |
| .long 0xe493e008 // ldr lr, [r3], #8 |
| .long 0xf2801010 // vmov.i32 d1, #0 |
| .long 0xf2802010 // vmov.i32 d2, #0 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee314b90 // vmov.32 r4, d17[1] |
| .long 0xf3c7101f // vmov.i32 d17, #255 |
| .long 0xe7de3003 // ldrb r3, [lr, r3] |
| .long 0xe7de4004 // ldrb r4, [lr, r4] |
| .long 0xee003b90 // vmov.32 d16[0], r3 |
| .long 0xee204b90 // vmov.32 d16[1], r4 |
| .long 0xf24001b1 // vand d16, d16, d17 |
| .long 0xeddf1b03 // vldr d17, [pc, #12] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_store_a8_vfp4 |
| .globl _sk_store_a8_vfp4 |
| FUNCTION(_sk_store_a8_vfp4) |
| _sk_store_a8_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xeddf0b0d // vldr d16, [pc, #52] |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xf2431c30 // vfma.f32 d17, d3, d16 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| .long 0xee10eb90 // vmov.32 lr, d16[0] |
| .long 0xee30cb90 // vmov.32 ip, d16[1] |
| .long 0xe7e3e000 // strb lr, [r3, r0]! |
| .long 0xe5c3c001 // strb ip, [r3, #1] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x437f0000 // .word 0x437f0000 |
| .long 0x437f0000 // .word 0x437f0000 |
| |
| HIDDEN _sk_load_g8_vfp4 |
| .globl _sk_load_g8_vfp4 |
| FUNCTION(_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_gather_g8_vfp4 |
| .globl _sk_gather_g8_vfp4 |
| FUNCTION(_sk_gather_g8_vfp4) |
| _sk_gather_g8_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2873f10 // vmov.f32 d3, #1 |
| .long 0xe493e008 // ldr lr, [r3], #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee314b90 // vmov.32 r4, d17[1] |
| .long 0xf3c7101f // vmov.i32 d17, #255 |
| .long 0xe7de3003 // ldrb r3, [lr, r3] |
| .long 0xe7de4004 // ldrb r4, [lr, r4] |
| .long 0xee003b90 // vmov.32 d16[0], r3 |
| .long 0xee204b90 // vmov.32 d16[1], r4 |
| .long 0xf24001b1 // vand d16, d16, d17 |
| .long 0xeddf1b05 // vldr d17, [pc, #20] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3000db1 // vmul.f32 d0, d16, d17 |
| .long 0xf2201110 // vorr d1, d0, d0 |
| .long 0xf2202110 // vorr d2, d0, d0 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_gather_i8_vfp4 |
| .globl _sk_gather_i8_vfp4 |
| FUNCTION(_sk_gather_i8_vfp4) |
| _sk_gather_i8_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe1a0e001 // mov lr, r1 |
| .long 0xe491c004 // ldr ip, [r1], #4 |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xe35c0000 // cmp ip, #0 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xe1a0300c // mov r3, ip |
| .long 0x028e1008 // addeq r1, lr, #8 |
| .long 0x059e3004 // ldreq r3, [lr, #4] |
| .long 0xe493e008 // ldr lr, [r3], #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee314b90 // vmov.32 r4, d17[1] |
| .long 0xf3c7101f // vmov.i32 d17, #255 |
| .long 0xe7de3003 // ldrb r3, [lr, r3] |
| .long 0xe7de4004 // ldrb r4, [lr, r4] |
| .long 0xee003b90 // vmov.32 d16[0], r3 |
| .long 0xee204b90 // vmov.32 d16[1], r4 |
| .long 0xe59c4004 // ldr r4, [ip, #4] |
| .long 0xf24001b1 // vand d16, d16, d17 |
| .long 0xee103b90 // vmov.32 r3, d16[0] |
| .long 0xee30eb90 // vmov.32 lr, d16[1] |
| .long 0xe0843103 // add r3, r4, r3, lsl #2 |
| .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32] |
| .long 0xe084310e // add r3, r4, lr, lsl #2 |
| .long 0xf4e308bf // vld1.32 {d16[1]}, [r3 :32] |
| .long 0xf24021b1 // vand d18, d16, d17 |
| .long 0xf3f83030 // vshr.u32 d19, d16, #8 |
| .long 0xf3e84030 // vshr.u32 d20, d16, #24 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3f00030 // vshr.u32 d16, d16, #16 |
| .long 0xf24331b1 // vand d19, d19, d17 |
| .long 0xf24001b1 // vand d16, d16, d17 |
| .long 0xeddf1b0a // vldr d17, [pc, #40] |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb4624 // vcvt.f32.s32 d20, d20 |
| .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3020db1 // vmul.f32 d0, d18, d17 |
| .long 0xf3043db1 // vmul.f32 d3, d20, d17 |
| .long 0xf3031db1 // vmul.f32 d1, d19, d17 |
| .long 0xf3002db1 // vmul.f32 d2, d16, d17 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff13 // bx r3 |
| .long 0xe320f000 // nop {0} |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_load_565_vfp4 |
| .globl _sk_load_565_vfp4 |
| FUNCTION(_sk_load_565_vfp4) |
| _sk_load_565_vfp4: |
| .long 0xe24dd004 // sub sp, sp, #4 |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c1101f // vmov.i32 d17, #31 |
| .long 0xf3c72218 // vmov.i32 d18, #63488 |
| .long 0xeddf3b16 // vldr d19, [pc, #88] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf2873f10 // vmov.f32 d3, #1 |
| .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 0xe3a03e7e // mov r3, #2016 |
| .long 0xf3d04a30 // vmovl.u16 q10, d16 |
| .long 0xee803b90 // vdup.32 d16, r3 |
| .long 0xf24411b1 // vand d17, d20, d17 |
| .long 0xeddf5b0e // vldr d21, [pc, #56] |
| .long 0xf24421b2 // vand d18, d20, d18 |
| .long 0xf24401b0 // vand d16, d20, d16 |
| .long 0xeddf4b09 // vldr d20, [pc, #36] |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3020db3 // vmul.f32 d0, d18, d19 |
| .long 0xf3001db4 // vmul.f32 d1, d16, d20 |
| .long 0xf3012db5 // vmul.f32 d2, d17, d21 |
| .long 0xe28dd004 // add sp, sp, #4 |
| .long 0xe12fff1c // bx ip |
| .long 0x37842108 // .word 0x37842108 |
| .long 0x37842108 // .word 0x37842108 |
| .long 0x3a020821 // .word 0x3a020821 |
| .long 0x3a020821 // .word 0x3a020821 |
| .long 0x3d042108 // .word 0x3d042108 |
| .long 0x3d042108 // .word 0x3d042108 |
| |
| HIDDEN _sk_gather_565_vfp4 |
| .globl _sk_gather_565_vfp4 |
| FUNCTION(_sk_gather_565_vfp4) |
| _sk_gather_565_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xeddf4b20 // vldr d20, [pc, #128] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe493e008 // ldr lr, [r3], #8 |
| .long 0xf2873f10 // vmov.f32 d3, #1 |
| .long 0xeddf5b1e // vldr d21, [pc, #120] |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xf2c1201f // vmov.i32 d18, #31 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee314b90 // vmov.32 r4, d17[1] |
| .long 0xf3c71218 // vmov.i32 d17, #63488 |
| .long 0xe08e3083 // add r3, lr, r3, lsl #1 |
| .long 0xe08e4084 // add r4, lr, r4, lsl #1 |
| .long 0xe1d330b0 // ldrh r3, [r3] |
| .long 0xe1d440b0 // ldrh r4, [r4] |
| .long 0xee003b90 // vmov.32 d16[0], r3 |
| .long 0xe3a03e7e // mov r3, #2016 |
| .long 0xee833b90 // vdup.32 d19, r3 |
| .long 0xee204b90 // vmov.32 d16[1], r4 |
| .long 0xf24011b1 // vand d17, d16, d17 |
| .long 0xf24031b3 // vand d19, d16, d19 |
| .long 0xf24001b2 // vand d16, d16, d18 |
| .long 0xf3fb2623 // vcvt.f32.s32 d18, d19 |
| .long 0xeddf3b07 // vldr d19, [pc, #28] |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3021db4 // vmul.f32 d1, d18, d20 |
| .long 0xf3010db3 // vmul.f32 d0, d17, d19 |
| .long 0xf3002db5 // vmul.f32 d2, d16, d21 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x37842108 // .word 0x37842108 |
| .long 0x37842108 // .word 0x37842108 |
| .long 0x3a020821 // .word 0x3a020821 |
| .long 0x3a020821 // .word 0x3a020821 |
| .long 0x3d042108 // .word 0x3d042108 |
| .long 0x3d042108 // .word 0x3d042108 |
| |
| HIDDEN _sk_store_565_vfp4 |
| .globl _sk_store_565_vfp4 |
| FUNCTION(_sk_store_565_vfp4) |
| _sk_store_565_vfp4: |
| .long 0xf2c30f1f // vmov.f32 d16, #31 |
| .long 0xeddf1b15 // vldr d17, [pc, #84] |
| .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2413c31 // vfma.f32 d19, d1, d17 |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2422c30 // vfma.f32 d18, d2, d16 |
| .long 0xe0833080 // add r3, r3, r0, lsl #1 |
| .long 0xf3fb07a3 // vcvt.u32.f32 d16, d19 |
| .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| .long 0xf2e50530 // vshl.s32 d16, d16, #5 |
| .long 0xf2eb1531 // vshl.s32 d17, d17, #11 |
| .long 0xf26001b1 // vorr d16, d16, d17 |
| .long 0xf26001b2 // vorr d16, d16, d18 |
| .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 |
| .long 0x427c0000 // .word 0x427c0000 |
| .long 0x427c0000 // .word 0x427c0000 |
| |
| HIDDEN _sk_load_4444_vfp4 |
| .globl _sk_load_4444_vfp4 |
| FUNCTION(_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_gather_4444_vfp4 |
| .globl _sk_gather_4444_vfp4 |
| FUNCTION(_sk_gather_4444_vfp4) |
| _sk_gather_4444_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xf3c73010 // vmov.i32 d19, #240 |
| .long 0xeddf5b21 // vldr d21, [pc, #132] |
| .long 0xe493e008 // ldr lr, [r3], #8 |
| .long 0xf2c0401f // vmov.i32 d20, #15 |
| .long 0xeddf6b20 // vldr d22, [pc, #128] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xf2c0221f // vmov.i32 d18, #3840 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee314b90 // vmov.32 r4, d17[1] |
| .long 0xf3c71210 // vmov.i32 d17, #61440 |
| .long 0xe08e3083 // add r3, lr, r3, lsl #1 |
| .long 0xe08e4084 // add r4, lr, r4, lsl #1 |
| .long 0xe1d330b0 // ldrh r3, [r3] |
| .long 0xe1d440b0 // ldrh r4, [r4] |
| .long 0xee003b90 // vmov.32 d16[0], r3 |
| .long 0xee204b90 // vmov.32 d16[1], r4 |
| .long 0xf24011b1 // vand d17, d16, d17 |
| .long 0xf24021b2 // vand d18, d16, d18 |
| .long 0xf24031b3 // vand d19, d16, d19 |
| .long 0xf24001b4 // vand d16, d16, d20 |
| .long 0xeddf4b0a // vldr d20, [pc, #40] |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3010db4 // vmul.f32 d0, d17, d20 |
| .long 0xeddf1b0a // vldr d17, [pc, #40] |
| .long 0xf3021db5 // vmul.f32 d1, d18, d21 |
| .long 0xf3032db6 // vmul.f32 d2, d19, d22 |
| .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| .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 |
| FUNCTION(_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 |
| FUNCTION(_sk_load_8888_vfp4) |
| _sk_load_8888_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3c7001f // vmov.i32 d16, #255 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xe0833100 // add r3, r3, r0, lsl #2 |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xf24121b0 // vand d18, d17, d16 |
| .long 0xf3f83031 // vshr.u32 d19, d17, #8 |
| .long 0xf3e84031 // vshr.u32 d20, d17, #24 |
| .long 0xf3f01031 // vshr.u32 d17, d17, #16 |
| .long 0xf24331b0 // vand d19, d19, d16 |
| .long 0xf24101b0 // vand d16, d17, d16 |
| .long 0xeddf1b08 // vldr d17, [pc, #32] |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb4624 // vcvt.f32.s32 d20, d20 |
| .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3020db1 // vmul.f32 d0, d18, d17 |
| .long 0xf3043db1 // vmul.f32 d3, d20, d17 |
| .long 0xf3031db1 // vmul.f32 d1, d19, d17 |
| .long 0xf3002db1 // vmul.f32 d2, d16, d17 |
| .long 0xe12fff1c // bx ip |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_gather_8888_vfp4 |
| .globl _sk_gather_8888_vfp4 |
| FUNCTION(_sk_gather_8888_vfp4) |
| _sk_gather_8888_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe493e008 // ldr lr, [r3], #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee314b90 // vmov.32 r4, d17[1] |
| .long 0xf3c7101f // vmov.i32 d17, #255 |
| .long 0xe08e3103 // add r3, lr, r3, lsl #2 |
| .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32] |
| .long 0xe08e3104 // add r3, lr, r4, lsl #2 |
| .long 0xf4e308bf // vld1.32 {d16[1]}, [r3 :32] |
| .long 0xf24021b1 // vand d18, d16, d17 |
| .long 0xf3f83030 // vshr.u32 d19, d16, #8 |
| .long 0xf3e84030 // vshr.u32 d20, d16, #24 |
| .long 0xf3f00030 // vshr.u32 d16, d16, #16 |
| .long 0xf24331b1 // vand d19, d19, d17 |
| .long 0xf24001b1 // vand d16, d16, d17 |
| .long 0xeddf1b09 // vldr d17, [pc, #36] |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb4624 // vcvt.f32.s32 d20, d20 |
| .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3020db1 // vmul.f32 d0, d18, d17 |
| .long 0xf3043db1 // vmul.f32 d3, d20, d17 |
| .long 0xf3031db1 // vmul.f32 d1, d19, d17 |
| .long 0xf3002db1 // vmul.f32 d2, d16, d17 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x3b808081 // .word 0x3b808081 |
| .long 0x3b808081 // .word 0x3b808081 |
| |
| HIDDEN _sk_store_8888_vfp4 |
| .globl _sk_store_8888_vfp4 |
| FUNCTION(_sk_store_8888_vfp4) |
| _sk_store_8888_vfp4: |
| .long 0xeddf0b1a // vldr d16, [pc, #104] |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2412c30 // vfma.f32 d18, d1, d16 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xf2423c30 // vfma.f32 d19, d2, d16 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| .long 0xe0833100 // add r3, r3, r0, lsl #2 |
| .long 0xf2434c30 // vfma.f32 d20, d3, d16 |
| .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18 |
| .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19 |
| .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| .long 0xf3fb37a4 // vcvt.u32.f32 d19, d20 |
| .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| .long 0xf2f02532 // vshl.s32 d18, d18, #16 |
| .long 0xf26001b1 // vorr d16, d16, d17 |
| .long 0xf2f81533 // vshl.s32 d17, d19, #24 |
| .long 0xf26001b2 // vorr d16, d16, d18 |
| .long 0xf26001b1 // vorr d16, d16, d17 |
| .long 0xedc30b00 // vstr d16, [r3] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x437f0000 // .word 0x437f0000 |
| .long 0x437f0000 // .word 0x437f0000 |
| |
| HIDDEN _sk_load_f16_vfp4 |
| .globl _sk_load_f16_vfp4 |
| FUNCTION(_sk_load_f16_vfp4) |
| _sk_load_f16_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]! |
| .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3] |
| .long 0xf3b60720 // vcvt.f32.f16 q0, d16 |
| .long 0xf3b62722 // vcvt.f32.f16 q1, d18 |
| .long 0xf3f64721 // vcvt.f32.f16 q10, d17 |
| .long 0xf3f60723 // vcvt.f32.f16 q8, d19 |
| .long 0xf22411b4 // vorr d1, d20, d20 |
| .long 0xf22031b0 // vorr d3, d16, d16 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_gather_f16_vfp4 |
| .globl _sk_gather_f16_vfp4 |
| FUNCTION(_sk_gather_f16_vfp4) |
| _sk_gather_f16_vfp4: |
| .long 0xe92d4c10 // push {r4, sl, fp, lr} |
| .long 0xe28db008 // add fp, sp, #8 |
| .long 0xe24dd010 // sub sp, sp, #16 |
| .long 0xe7c3d01f // bfc sp, #0, #4 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf3fb0701 // vcvt.s32.f32 d16, d1 |
| .long 0xf3fb1700 // vcvt.s32.f32 d17, d0 |
| .long 0xe493c008 // ldr ip, [r3], #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26219a0 // vmla.i32 d17, d18, d16 |
| .long 0xee113b90 // vmov.32 r3, d17[0] |
| .long 0xee31eb90 // vmov.32 lr, d17[1] |
| .long 0xe08c4183 // add r4, ip, r3, lsl #3 |
| .long 0xe08c318e // add r3, ip, lr, lsl #3 |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xe1a0300d // mov r3, sp |
| .long 0xedd40b00 // vldr d16, [r4] |
| .long 0xf4430aef // vst1.64 {d16-d17}, [r3 :128] |
| .long 0xf4e3071f // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3 :64] |
| .long 0xe3833008 // orr r3, r3, #8 |
| .long 0xf4e3075f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3 :64] |
| .long 0xf3b60720 // vcvt.f32.f16 q0, d16 |
| .long 0xf3b62722 // vcvt.f32.f16 q1, d18 |
| .long 0xe5913004 // ldr r3, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf3f64721 // vcvt.f32.f16 q10, d17 |
| .long 0xf3f60723 // vcvt.f32.f16 q8, d19 |
| .long 0xf22411b4 // vorr d1, d20, d20 |
| .long 0xf22031b0 // vorr d3, d16, d16 |
| .long 0xe12fff33 // blx r3 |
| .long 0xe24bd008 // sub sp, fp, #8 |
| .long 0xe8bd8c10 // pop {r4, sl, fp, pc} |
| |
| HIDDEN _sk_store_f16_vfp4 |
| .globl _sk_store_f16_vfp4 |
| FUNCTION(_sk_store_f16_vfp4) |
| _sk_store_f16_vfp4: |
| .long 0xf2630113 // vorr d16, d3, d3 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2612111 // vorr d18, d1, d1 |
| .long 0xf3f67620 // vcvt.f16.f32 d23, q8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf3f66602 // vcvt.f16.f32 d22, q1 |
| .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| .long 0xf3f65622 // vcvt.f16.f32 d21, q9 |
| .long 0xf3f64600 // vcvt.f16.f32 d20, q0 |
| .long 0xf22211b2 // vorr d1, d18, d18 |
| .long 0xf22031b0 // vorr d3, d16, d16 |
| .long 0xf4c3470d // vst4.16 {d20[0],d21[0],d22[0],d23[0]}, [r3]! |
| .long 0xf4c3474f // vst4.16 {d20[1],d21[1],d22[1],d23[1]}, [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_u16_be_vfp4 |
| .globl _sk_load_u16_be_vfp4 |
| FUNCTION(_sk_load_u16_be_vfp4) |
| _sk_load_u16_be_vfp4: |
| .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]! |
| .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3] |
| .long 0xee903bb0 // vmov.u16 r3, d16[0] |
| .long 0xee92ebb0 // vmov.u16 lr, d18[0] |
| .long 0xee914bb0 // vmov.u16 r4, d17[0] |
| .long 0xee937bb0 // vmov.u16 r7, d19[0] |
| .long 0xee905bf0 // vmov.u16 r5, d16[1] |
| .long 0xee926bf0 // vmov.u16 r6, d18[1] |
| .long 0xee043b90 // vmov.32 d20[0], r3 |
| .long 0xee05eb90 // vmov.32 d21[0], lr |
| .long 0xee93ebf0 // vmov.u16 lr, d19[1] |
| .long 0xee913bf0 // vmov.u16 r3, d17[1] |
| .long 0xf3c71c1f // vmov.i32 d17, #65535 |
| .long 0xee004b90 // vmov.32 d16[0], r4 |
| .long 0xee027b90 // vmov.32 d18[0], r7 |
| .long 0xee245b90 // vmov.32 d20[1], r5 |
| .long 0xf24431b1 // vand d19, d20, d17 |
| .long 0xee256b90 // vmov.32 d21[1], r6 |
| .long 0xf2e84534 // vshl.s32 d20, d20, #8 |
| .long 0xf24561b1 // vand d22, d21, d17 |
| .long 0xf3f83033 // vshr.u32 d19, d19, #8 |
| .long 0xf2e85535 // vshl.s32 d21, d21, #8 |
| .long 0xf26431b3 // vorr d19, d20, d19 |
| .long 0xf3f86036 // vshr.u32 d22, d22, #8 |
| .long 0xf24331b1 // vand d19, d19, d17 |
| .long 0xf26551b6 // vorr d21, d21, d22 |
| .long 0xf3fb36a3 // vcvt.f32.u32 d19, d19 |
| .long 0xee22eb90 // vmov.32 d18[1], lr |
| .long 0xee203b90 // vmov.32 d16[1], r3 |
| .long 0xf24281b1 // vand d24, d18, d17 |
| .long 0xf2e82532 // vshl.s32 d18, d18, #8 |
| .long 0xf24071b1 // vand d23, d16, d17 |
| .long 0xf3f84038 // vshr.u32 d20, d24, #8 |
| .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| .long 0xf3f87037 // vshr.u32 d23, d23, #8 |
| .long 0xf26221b4 // vorr d18, d18, d20 |
| .long 0xf26001b7 // vorr d16, d16, d23 |
| .long 0xf24541b1 // vand d20, d21, d17 |
| .long 0xf24001b1 // vand d16, d16, d17 |
| .long 0xf24211b1 // vand d17, d18, d17 |
| .long 0xeddf2b09 // vldr d18, [pc, #36] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3fb46a4 // vcvt.f32.u32 d20, d20 |
| .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17 |
| .long 0xf3030db2 // vmul.f32 d0, d19, d18 |
| .long 0xf3001db2 // vmul.f32 d1, d16, d18 |
| .long 0xf3042db2 // vmul.f32 d2, d20, d18 |
| .long 0xf3013db2 // vmul.f32 d3, d17, d18 |
| .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x37800080 // .word 0x37800080 |
| .long 0x37800080 // .word 0x37800080 |
| |
| HIDDEN _sk_load_rgb_u16_be_vfp4 |
| .globl _sk_load_rgb_u16_be_vfp4 |
| FUNCTION(_sk_load_rgb_u16_be_vfp4) |
| _sk_load_rgb_u16_be_vfp4: |
| .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2873f10 // vmov.f32 d3, #1 |
| .long 0xe593e000 // ldr lr, [r3] |
| .long 0xe0803080 // add r3, r0, r0, lsl #1 |
| .long 0xe08e3083 // add r3, lr, r3, lsl #1 |
| .long 0xf4e3060d // vld3.16 {d16[0],d17[0],d18[0]}, [r3]! |
| .long 0xf4e3064f // vld3.16 {d16[1],d17[1],d18[1]}, [r3] |
| .long 0xee90ebb0 // vmov.u16 lr, d16[0] |
| .long 0xee913bb0 // vmov.u16 r3, d17[0] |
| .long 0xee924bb0 // vmov.u16 r4, d18[0] |
| .long 0xee927bf0 // vmov.u16 r7, d18[1] |
| .long 0xf3c73c1f // vmov.i32 d19, #65535 |
| .long 0xee905bf0 // vmov.u16 r5, d16[1] |
| .long 0xee916bf0 // vmov.u16 r6, d17[1] |
| .long 0xee00eb90 // vmov.32 d16[0], lr |
| .long 0xee013b90 // vmov.32 d17[0], r3 |
| .long 0xee024b90 // vmov.32 d18[0], r4 |
| .long 0xee205b90 // vmov.32 d16[1], r5 |
| .long 0xee216b90 // vmov.32 d17[1], r6 |
| .long 0xf24041b3 // vand d20, d16, d19 |
| .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| .long 0xee227b90 // vmov.32 d18[1], r7 |
| .long 0xf24151b3 // vand d21, d17, d19 |
| .long 0xf3f84034 // vshr.u32 d20, d20, #8 |
| .long 0xf24261b3 // vand d22, d18, d19 |
| .long 0xf2e81531 // vshl.s32 d17, d17, #8 |
| .long 0xf3f85035 // vshr.u32 d21, d21, #8 |
| .long 0xf2e82532 // vshl.s32 d18, d18, #8 |
| .long 0xf3f86036 // vshr.u32 d22, d22, #8 |
| .long 0xf26001b4 // vorr d16, d16, d20 |
| .long 0xf26111b5 // vorr d17, d17, d21 |
| .long 0xf26221b6 // vorr d18, d18, d22 |
| .long 0xf24001b3 // vand d16, d16, d19 |
| .long 0xf24111b3 // vand d17, d17, d19 |
| .long 0xf24221b3 // vand d18, d18, d19 |
| .long 0xeddf3b07 // vldr d19, [pc, #28] |
| .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17 |
| .long 0xf3fb26a2 // vcvt.f32.u32 d18, d18 |
| .long 0xf3000db3 // vmul.f32 d0, d16, d19 |
| .long 0xf3011db3 // vmul.f32 d1, d17, d19 |
| .long 0xf3022db3 // vmul.f32 d2, d18, d19 |
| .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| .long 0xe12fff1c // bx ip |
| .long 0x37800080 // .word 0x37800080 |
| .long 0x37800080 // .word 0x37800080 |
| |
| HIDDEN _sk_store_u16_be_vfp4 |
| .globl _sk_store_u16_be_vfp4 |
| FUNCTION(_sk_store_u16_be_vfp4) |
| _sk_store_u16_be_vfp4: |
| .long 0xeddf0b2a // vldr d16, [pc, #168] |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2432c30 // vfma.f32 d18, d3, d16 |
| .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| .long 0xf2423c30 // vfma.f32 d19, d2, d16 |
| .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| .long 0xf2414c30 // vfma.f32 d20, d1, d16 |
| .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18 |
| .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19 |
| .long 0xf3c73c1f // vmov.i32 d19, #65535 |
| .long 0xf3fb47a4 // vcvt.u32.f32 d20, d20 |
| .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| .long 0xf24051b3 // vand d21, d16, d19 |
| .long 0xf24261b3 // vand d22, d18, d19 |
| .long 0xf24471b3 // vand d23, d20, d19 |
| .long 0xf24131b3 // vand d19, d17, d19 |
| .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| .long 0xf3f85035 // vshr.u32 d21, d21, #8 |
| .long 0xf2e82532 // vshl.s32 d18, d18, #8 |
| .long 0xf3f86036 // vshr.u32 d22, d22, #8 |
| .long 0xf260b1b5 // vorr d27, d16, d21 |
| .long 0xf2e84534 // vshl.s32 d20, d20, #8 |
| .long 0xf3f87037 // vshr.u32 d23, d23, #8 |
| .long 0xf262a1b6 // vorr d26, d18, d22 |
| .long 0xf2e81531 // vshl.s32 d17, d17, #8 |
| .long 0xf3f83033 // vshr.u32 d19, d19, #8 |
| .long 0xf26491b7 // vorr d25, d20, d23 |
| .long 0xf26181b3 // vorr d24, d17, d19 |
| .long 0xf3f6b120 // vuzp.16 d27, d16 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xf3f6a120 // vuzp.16 d26, d16 |
| .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| .long 0xf3f69120 // vuzp.16 d25, d16 |
| .long 0xf3f68120 // vuzp.16 d24, d16 |
| .long 0xf4c3870d // vst4.16 {d24[0],d25[0],d26[0],d27[0]}, [r3]! |
| .long 0xf4c3874f // vst4.16 {d24[1],d25[1],d26[1],d27[1]}, [r3] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x477fff00 // .word 0x477fff00 |
| .long 0x477fff00 // .word 0x477fff00 |
| |
| HIDDEN _sk_load_f32_vfp4 |
| .globl _sk_load_f32_vfp4 |
| FUNCTION(_sk_load_f32_vfp4) |
| _sk_load_f32_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xe0833200 // add r3, r3, r0, lsl #4 |
| .long 0xf423008f // vld4.32 {d0-d3}, [r3] |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_store_f32_vfp4 |
| .globl _sk_store_f32_vfp4 |
| FUNCTION(_sk_store_f32_vfp4) |
| _sk_store_f32_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xe5933000 // ldr r3, [r3] |
| .long 0xe0833200 // add r3, r3, r0, lsl #4 |
| .long 0xf403008f // vst4.32 {d0-d3}, [r3] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_clamp_x_vfp4 |
| .globl _sk_clamp_x_vfp4 |
| FUNCTION(_sk_clamp_x_vfp4) |
| _sk_clamp_x_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c00010 // vmov.i32 d16, #0 |
| .long 0xf3c71e1f // vmov.i8 d17, #255 |
| .long 0xf2400f80 // vmax.f32 d16, d16, d0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| .long 0xf2200fa1 // vmin.f32 d0, d16, d17 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_clamp_y_vfp4 |
| .globl _sk_clamp_y_vfp4 |
| FUNCTION(_sk_clamp_y_vfp4) |
| _sk_clamp_y_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c00010 // vmov.i32 d16, #0 |
| .long 0xf3c71e1f // vmov.i8 d17, #255 |
| .long 0xf2400f81 // vmax.f32 d16, d16, d1 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| .long 0xf2201fa1 // vmin.f32 d1, d16, d17 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_repeat_x_vfp4 |
| .globl _sk_repeat_x_vfp4 |
| FUNCTION(_sk_repeat_x_vfp4) |
| _sk_repeat_x_vfp4: |
| .long 0xed2d8b04 // vpush {d8-d9} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c02010 // vmov.i32 d18, #0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xeddf3b10 // vldr d19, [pc, #64] |
| .long 0xed938a00 // vldr s16, [r3] |
| .long 0xeec09a88 // vdiv.f32 s19, s1, s16 |
| .long 0xee809a08 // vdiv.f32 s18, s0, s16 |
| .long 0xf3fb0709 // vcvt.s32.f32 d16, d9 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3601e89 // vcgt.f32 d17, d16, d9 |
| .long 0xf35311b2 // vbsl d17, d19, d18 |
| .long 0xf3f42c08 // vdup.32 d18, d8[0] |
| .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| .long 0xf3c71e1f // vmov.i8 d17, #255 |
| .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0] |
| .long 0xf2600d20 // vsub.f32 d16, d0, d16 |
| .long 0xf2200fa1 // vmin.f32 d0, d16, d17 |
| .long 0xecbd8b04 // vpop {d8-d9} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| |
| HIDDEN _sk_repeat_y_vfp4 |
| .globl _sk_repeat_y_vfp4 |
| FUNCTION(_sk_repeat_y_vfp4) |
| _sk_repeat_y_vfp4: |
| .long 0xed2d8b04 // vpush {d8-d9} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c02010 // vmov.i32 d18, #0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xeddf3b10 // vldr d19, [pc, #64] |
| .long 0xed938a00 // vldr s16, [r3] |
| .long 0xeec19a88 // vdiv.f32 s19, s3, s16 |
| .long 0xee819a08 // vdiv.f32 s18, s2, s16 |
| .long 0xf3fb0709 // vcvt.s32.f32 d16, d9 |
| .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| .long 0xf3601e89 // vcgt.f32 d17, d16, d9 |
| .long 0xf35311b2 // vbsl d17, d19, d18 |
| .long 0xf3f42c08 // vdup.32 d18, d8[0] |
| .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| .long 0xf3c71e1f // vmov.i8 d17, #255 |
| .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0] |
| .long 0xf2610d20 // vsub.f32 d16, d1, d16 |
| .long 0xf2201fa1 // vmin.f32 d1, d16, d17 |
| .long 0xecbd8b04 // vpop {d8-d9} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| |
| HIDDEN _sk_mirror_x_vfp4 |
| .globl _sk_mirror_x_vfp4 |
| FUNCTION(_sk_mirror_x_vfp4) |
| _sk_mirror_x_vfp4: |
| .long 0xed2d8b04 // vpush {d8-d9} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xeddf4b14 // vldr d20, [pc, #80] |
| .long 0xed938a00 // vldr s16, [r3] |
| .long 0xee389a08 // vadd.f32 s18, s16, s16 |
| .long 0xf3f40c08 // vdup.32 d16, d8[0] |
| .long 0xf2200d20 // vsub.f32 d0, d0, d16 |
| .long 0xeec08a89 // vdiv.f32 s17, s1, s18 |
| .long 0xee808a09 // vdiv.f32 s16, s0, s18 |
| .long 0xf3fb1708 // vcvt.s32.f32 d17, d8 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3612e88 // vcgt.f32 d18, d17, d8 |
| .long 0xf35421b3 // vbsl d18, d20, d19 |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xf3c72e1f // vmov.i8 d18, #255 |
| .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0] |
| .long 0xf2601d21 // vsub.f32 d17, d0, d17 |
| .long 0xf2611da0 // vsub.f32 d17, d17, d16 |
| .long 0xf26008a2 // vadd.i32 d16, d16, d18 |
| .long 0xf3f91721 // vabs.f32 d17, d17 |
| .long 0xf2210fa0 // vmin.f32 d0, d17, d16 |
| .long 0xecbd8b04 // vpop {d8-d9} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| |
| HIDDEN _sk_mirror_y_vfp4 |
| .globl _sk_mirror_y_vfp4 |
| FUNCTION(_sk_mirror_y_vfp4) |
| _sk_mirror_y_vfp4: |
| .long 0xed2d8b04 // vpush {d8-d9} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xeddf4b14 // vldr d20, [pc, #80] |
| .long 0xed938a00 // vldr s16, [r3] |
| .long 0xee389a08 // vadd.f32 s18, s16, s16 |
| .long 0xf3f40c08 // vdup.32 d16, d8[0] |
| .long 0xf2211d20 // vsub.f32 d1, d1, d16 |
| .long 0xeec18a89 // vdiv.f32 s17, s3, s18 |
| .long 0xee818a09 // vdiv.f32 s16, s2, s18 |
| .long 0xf3fb1708 // vcvt.s32.f32 d17, d8 |
| .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| .long 0xf3612e88 // vcgt.f32 d18, d17, d8 |
| .long 0xf35421b3 // vbsl d18, d20, d19 |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xf3c72e1f // vmov.i8 d18, #255 |
| .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0] |
| .long 0xf2611d21 // vsub.f32 d17, d1, d17 |
| .long 0xf2611da0 // vsub.f32 d17, d17, d16 |
| .long 0xf26008a2 // vadd.i32 d16, d16, d18 |
| .long 0xf3f91721 // vabs.f32 d17, d17 |
| .long 0xf2211fa0 // vmin.f32 d1, d17, d16 |
| .long 0xecbd8b04 // vpop {d8-d9} |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| |
| HIDDEN _sk_luminance_to_alpha_vfp4 |
| .globl _sk_luminance_to_alpha_vfp4 |
| FUNCTION(_sk_luminance_to_alpha_vfp4) |
| _sk_luminance_to_alpha_vfp4: |
| .long 0xeddf0b0a // vldr d16, [pc, #40] |
| .long 0xeddf1b0b // vldr d17, [pc, #44] |
| .long 0xf3410d30 // vmul.f32 d16, d1, d16 |
| .long 0xe4913004 // ldr r3, [r1], #4 |
| .long 0xf3401d31 // vmul.f32 d17, d0, d17 |
| .long 0xf2800010 // vmov.i32 d0, #0 |
| .long 0xf2801010 // vmov.i32 d1, #0 |
| .long 0xf2013da0 // vadd.f32 d3, d17, d16 |
| .long 0xeddf0b06 // vldr d16, [pc, #24] |
| .long 0xf2023c30 // vfma.f32 d3, d2, d16 |
| .long 0xf2802010 // vmov.i32 d2, #0 |
| .long 0xe12fff13 // bx r3 |
| .long 0x3f371759 // .word 0x3f371759 |
| .long 0x3f371759 // .word 0x3f371759 |
| .long 0x3e59b3d0 // .word 0x3e59b3d0 |
| .long 0x3e59b3d0 // .word 0x3e59b3d0 |
| .long 0x3d93dd98 // .word 0x3d93dd98 |
| .long 0x3d93dd98 // .word 0x3d93dd98 |
| |
| HIDDEN _sk_matrix_2x3_vfp4 |
| .globl _sk_matrix_2x3_vfp4 |
| FUNCTION(_sk_matrix_2x3_vfp4) |
| _sk_matrix_2x3_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf2410c31 // vfma.f32 d16, d1, d17 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xf2411c32 // vfma.f32 d17, d1, d18 |
| .long 0xf4ee2c9d // vld1.32 {d18[]}, [lr :32]! |
| .long 0xf4ee3c9f // vld1.32 {d19[]}, [lr :32] |
| .long 0xf2400c32 // vfma.f32 d16, d0, d18 |
| .long 0xf2401c33 // vfma.f32 d17, d0, d19 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_matrix_3x4_vfp4 |
| .globl _sk_matrix_3x4_vfp4 |
| FUNCTION(_sk_matrix_3x4_vfp4) |
| _sk_matrix_3x4_vfp4: |
| .long 0xe92d4800 // push {fp, lr} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe28e3020 // add r3, lr, #32 |
| .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| .long 0xe28e302c // add r3, lr, #44 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e301c // add r3, lr, #28 |
| .long 0xf2420c33 // vfma.f32 d16, d2, d19 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xe28e3018 // add r3, lr, #24 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e3024 // add r3, lr, #36 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3028 // add r3, lr, #40 |
| .long 0xf2421c32 // vfma.f32 d17, d2, d18 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf2422c34 // vfma.f32 d18, d2, d20 |
| .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf2411c34 // vfma.f32 d17, d1, d20 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf2410c34 // vfma.f32 d16, d1, d20 |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf2412c33 // vfma.f32 d18, d1, d19 |
| .long 0xf4ee3c9d // vld1.32 {d19[]}, [lr :32]! |
| .long 0xf4ee4c9f // vld1.32 {d20[]}, [lr :32] |
| .long 0xf2401c33 // vfma.f32 d17, d0, d19 |
| .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| .long 0xf2400c33 // vfma.f32 d16, d0, d19 |
| .long 0xf2402c34 // vfma.f32 d18, d0, d20 |
| .long 0xf22101b1 // vorr d0, d17, d17 |
| .long 0xf22021b0 // vorr d2, d16, d16 |
| .long 0xf22211b2 // vorr d1, d18, d18 |
| .long 0xe8bd4800 // pop {fp, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_matrix_4x5_vfp4 |
| .globl _sk_matrix_4x5_vfp4 |
| FUNCTION(_sk_matrix_4x5_vfp4) |
| _sk_matrix_4x5_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xf2620112 // vorr d16, d2, d2 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe2834014 // add r4, r3, #20 |
| .long 0xe1a0e003 // mov lr, r3 |
| .long 0xf4e45c9f // vld1.32 {d21[]}, [r4 :32] |
| .long 0xe2834028 // add r4, r3, #40 |
| .long 0xf4e46c9f // vld1.32 {d22[]}, [r4 :32] |
| .long 0xe2834038 // add r4, r3, #56 |
| .long 0xf4e47c9f // vld1.32 {d23[]}, [r4 :32] |
| .long 0xe2834048 // add r4, r3, #72 |
| .long 0xf4a42c9f // vld1.32 {d2[]}, [r4 :32] |
| .long 0xe2834034 // add r4, r3, #52 |
| .long 0xf2032c37 // vfma.f32 d2, d3, d23 |
| .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32] |
| .long 0xe2834044 // add r4, r3, #68 |
| .long 0xf4e41c9f // vld1.32 {d17[]}, [r4 :32] |
| .long 0xe2834030 // add r4, r3, #48 |
| .long 0xf2431c38 // vfma.f32 d17, d3, d24 |
| .long 0xf4e49c9f // vld1.32 {d25[]}, [r4 :32] |
| .long 0xe283403c // add r4, r3, #60 |
| .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| .long 0xe283404c // add r4, r3, #76 |
| .long 0xf2002cb6 // vfma.f32 d2, d16, d22 |
| .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32] |
| .long 0xe2834040 // add r4, r3, #64 |
| .long 0xf2432c33 // vfma.f32 d18, d3, d19 |
| .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| .long 0xe2834020 // add r4, r3, #32 |
| .long 0xf2433c39 // vfma.f32 d19, d3, d25 |
| .long 0xf4e47c9f // vld1.32 {d23[]}, [r4 :32] |
| .long 0xe283402c // add r4, r3, #44 |
| .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32] |
| .long 0xe2834024 // add r4, r3, #36 |
| .long 0xf2402cb8 // vfma.f32 d18, d16, d24 |
| .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32] |
| .long 0xf2401cb8 // vfma.f32 d17, d16, d24 |
| .long 0xe2834010 // add r4, r3, #16 |
| .long 0xf2403cb7 // vfma.f32 d19, d16, d23 |
| .long 0xf4ee4c9d // vld1.32 {d20[]}, [lr :32]! |
| .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| .long 0xe283401c // add r4, r3, #28 |
| .long 0xf4e46c9f // vld1.32 {d22[]}, [r4 :32] |
| .long 0xe2834018 // add r4, r3, #24 |
| .long 0xf2412c36 // vfma.f32 d18, d1, d22 |
| .long 0xf2411c35 // vfma.f32 d17, d1, d21 |
| .long 0xf4ee5c9f // vld1.32 {d21[]}, [lr :32] |
| .long 0xf2413c30 // vfma.f32 d19, d1, d16 |
| .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| .long 0xe2834008 // add r4, r3, #8 |
| .long 0xe283300c // add r3, r3, #12 |
| .long 0xf2012c30 // vfma.f32 d2, d1, d16 |
| .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| .long 0xf2401c35 // vfma.f32 d17, d0, d21 |
| .long 0xf2403c34 // vfma.f32 d19, d0, d20 |
| .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| .long 0xf2402c34 // vfma.f32 d18, d0, d20 |
| .long 0xf2002c30 // vfma.f32 d2, d0, d16 |
| .long 0xf22111b1 // vorr d1, d17, d17 |
| .long 0xf22301b3 // vorr d0, d19, d19 |
| .long 0xf22231b2 // vorr d3, d18, d18 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_matrix_perspective_vfp4 |
| .globl _sk_matrix_perspective_vfp4 |
| FUNCTION(_sk_matrix_perspective_vfp4) |
| _sk_matrix_perspective_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe28e301c // add r3, lr, #28 |
| .long 0xe28e4010 // add r4, lr, #16 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe28e3020 // add r3, lr, #32 |
| .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| .long 0xe28e3018 // add r3, lr, #24 |
| .long 0xf2411c30 // vfma.f32 d17, d1, d16 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe1a0300e // mov r3, lr |
| .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32] |
| .long 0xe28e4008 // add r4, lr, #8 |
| .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! |
| .long 0xf4e35c9f // vld1.32 {d21[]}, [r3 :32] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xf2413c35 // vfma.f32 d19, d1, d21 |
| .long 0xf4e35c9f // vld1.32 {d21[]}, [r3 :32] |
| .long 0xe28e300c // add r3, lr, #12 |
| .long 0xf2415c32 // vfma.f32 d21, d1, d18 |
| .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| .long 0xf3fb4521 // vrecpe.f32 d20, d17 |
| .long 0xf2403c30 // vfma.f32 d19, d0, d16 |
| .long 0xf2411fb4 // vrecps.f32 d17, d17, d20 |
| .long 0xf2405c32 // vfma.f32 d21, d0, d18 |
| .long 0xf3440db1 // vmul.f32 d16, d20, d17 |
| .long 0xf3030db0 // vmul.f32 d0, d19, d16 |
| .long 0xf3051db0 // vmul.f32 d1, d21, d16 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_linear_gradient_vfp4 |
| .globl _sk_linear_gradient_vfp4 |
| FUNCTION(_sk_linear_gradient_vfp4) |
| _sk_linear_gradient_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe591e000 // ldr lr, [r1] |
| .long 0xe28e3014 // add r3, lr, #20 |
| .long 0xe1a0400e // mov r4, lr |
| .long 0xf4a33c9f // vld1.32 {d3[]}, [r3 :32] |
| .long 0xe28e3010 // add r3, lr, #16 |
| .long 0xf4a32c9f // vld1.32 {d2[]}, [r3 :32] |
| .long 0xe28e3008 // add r3, lr, #8 |
| .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| .long 0xe494c00c // ldr ip, [r4], #12 |
| .long 0xf4a41c9f // vld1.32 {d1[]}, [r4 :32] |
| .long 0xe35c0000 // cmp ip, #0 |
| .long 0x0a000036 // beq 2f78 <sk_linear_gradient_vfp4+0x110> |
| .long 0xe59e3004 // ldr r3, [lr, #4] |
| .long 0xf2c01010 // vmov.i32 d17, #0 |
| .long 0xf2c07010 // vmov.i32 d23, #0 |
| .long 0xf2c08010 // vmov.i32 d24, #0 |
| .long 0xe2833020 // add r3, r3, #32 |
| .long 0xf2c06010 // vmov.i32 d22, #0 |
| .long 0xe2434018 // sub r4, r3, #24 |
| .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| .long 0xe25cc001 // subs ip, ip, #1 |
| .long 0xf4e4dc9f // vld1.32 {d29[]}, [r4 :32] |
| .long 0xe2434014 // sub r4, r3, #20 |
| .long 0xf4e45c9f // vld1.32 {d21[]}, [r4 :32] |
| .long 0xe243400c // sub r4, r3, #12 |
| .long 0xf4e44c9f // vld1.32 {d20[]}, [r4 :32] |
| .long 0xe2434020 // sub r4, r3, #32 |
| .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32] |
| .long 0xe2434004 // sub r4, r3, #4 |
| .long 0xf3622e80 // vcgt.f32 d18, d18, d0 |
| .long 0xf4e4bc9f // vld1.32 {d27[]}, [r4 :32] |
| .long 0xe2434008 // sub r4, r3, #8 |
| .long 0xf4e4cc9f // vld1.32 {d28[]}, [r4 :32] |
| .long 0xe2434010 // sub r4, r3, #16 |
| .long 0xf262a1b2 // vorr d26, d18, d18 |
| .long 0xf4e4ec9f // vld1.32 {d30[]}, [r4 :32] |
| .long 0xe243401c // sub r4, r3, #28 |
| .long 0xf352a13b // vbsl d26, d2, d27 |
| .long 0xe2833024 // add r3, r3, #36 |
| .long 0xf262b1b2 // vorr d27, d18, d18 |
| .long 0xf26291b2 // vorr d25, d18, d18 |
| .long 0xf351b13c // vbsl d27, d1, d28 |
| .long 0xf262c1b2 // vorr d28, d18, d18 |
| .long 0xf3539133 // vbsl d25, d3, d19 |
| .long 0xf350c1b4 // vbsl d28, d16, d20 |
| .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| .long 0xf26241b2 // vorr d20, d18, d18 |
| .long 0xf26231b2 // vorr d19, d18, d18 |
| .long 0xf35841b5 // vbsl d20, d24, d21 |
| .long 0xf26251b2 // vorr d21, d18, d18 |
| .long 0xf35121b0 // vbsl d18, d17, d16 |
| .long 0xf35731be // vbsl d19, d23, d30 |
| .long 0xf35651bd // vbsl d21, d22, d29 |
| .long 0xf26211b2 // vorr d17, d18, d18 |
| .long 0xf22931b9 // vorr d3, d25, d25 |
| .long 0xf22a21ba // vorr d2, d26, d26 |
| .long 0xf22b11bb // vorr d1, d27, d27 |
| .long 0xf26c01bc // vorr d16, d28, d28 |
| .long 0xf26371b3 // vorr d23, d19, d19 |
| .long 0xf26481b4 // vorr d24, d20, d20 |
| .long 0xf26561b5 // vorr d22, d21, d21 |
| .long 0x1affffd3 // bne 2eb4 <sk_linear_gradient_vfp4+0x4c> |
| .long 0xf26c01bc // vorr d16, d28, d28 |
| .long 0xf22b11bb // vorr d1, d27, d27 |
| .long 0xf22a21ba // vorr d2, d26, d26 |
| .long 0xf22931b9 // vorr d3, d25, d25 |
| .long 0xea000003 // b 2f88 <sk_linear_gradient_vfp4+0x120> |
| .long 0xf2c05010 // vmov.i32 d21, #0 |
| .long 0xf2c04010 // vmov.i32 d20, #0 |
| .long 0xf2c03010 // vmov.i32 d19, #0 |
| .long 0xf2c02010 // vmov.i32 d18, #0 |
| .long 0xf2400c32 // vfma.f32 d16, d0, d18 |
| .long 0xe5913004 // ldr r3, [r1, #4] |
| .long 0xf2001c35 // vfma.f32 d1, d0, d21 |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xf2002c34 // vfma.f32 d2, d0, d20 |
| .long 0xf2003c33 // vfma.f32 d3, d0, d19 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff13 // bx r3 |
| |
| HIDDEN _sk_linear_gradient_2stops_vfp4 |
| .globl _sk_linear_gradient_2stops_vfp4 |
| FUNCTION(_sk_linear_gradient_2stops_vfp4) |
| _sk_linear_gradient_2stops_vfp4: |
| .long 0xe92d4010 // push {r4, lr} |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xe283400c // add r4, r3, #12 |
| .long 0xe1a0e003 // mov lr, r3 |
| .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32] |
| .long 0xe2834008 // add r4, r3, #8 |
| .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| .long 0xe2834018 // add r4, r3, #24 |
| .long 0xf4a42c9f // vld1.32 {d2[]}, [r4 :32] |
| .long 0xe2834010 // add r4, r3, #16 |
| .long 0xf2002c33 // vfma.f32 d2, d0, d19 |
| .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| .long 0xe283401c // add r4, r3, #28 |
| .long 0xe2833014 // add r3, r3, #20 |
| .long 0xf4ee1c9d // vld1.32 {d17[]}, [lr :32]! |
| .long 0xf2400c31 // vfma.f32 d16, d0, d17 |
| .long 0xf4a43c9f // vld1.32 {d3[]}, [r4 :32] |
| .long 0xf2003c32 // vfma.f32 d3, d0, d18 |
| .long 0xf4ee1c9f // vld1.32 {d17[]}, [lr :32] |
| .long 0xf4a31c9f // vld1.32 {d1[]}, [r3 :32] |
| .long 0xf2001c31 // vfma.f32 d1, d0, d17 |
| .long 0xf22001b0 // vorr d0, d16, d16 |
| .long 0xe8bd4010 // pop {r4, lr} |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_save_xy_vfp4 |
| .globl _sk_save_xy_vfp4 |
| FUNCTION(_sk_save_xy_vfp4) |
| _sk_save_xy_vfp4: |
| .long 0xf2c3061f // vmov.i32 d16, #1056964608 |
| .long 0xeddf7b17 // vldr d23, [pc, #92] |
| .long 0xf2c06010 // vmov.i32 d22, #0 |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2401d20 // vadd.f32 d17, d0, d16 |
| .long 0xf2410d20 // vadd.f32 d16, d1, d16 |
| .long 0xed830b00 // vstr d0, [r3] |
| .long 0xed831b08 // vstr d1, [r3, #32] |
| .long 0xf3fb2721 // vcvt.s32.f32 d18, d17 |
| .long 0xf3fb3720 // vcvt.s32.f32 d19, d16 |
| .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 |
| .long 0xf3624ea1 // vcgt.f32 d20, d18, d17 |
| .long 0xf3635ea0 // vcgt.f32 d21, d19, d16 |
| .long 0xf35741b6 // vbsl d20, d23, d22 |
| .long 0xf35751b6 // vbsl d21, d23, d22 |
| .long 0xf2622da4 // vsub.f32 d18, d18, d20 |
| .long 0xf2633da5 // vsub.f32 d19, d19, d21 |
| .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| .long 0xf2600da3 // vsub.f32 d16, d16, d19 |
| .long 0xedc31b10 // vstr d17, [r3, #64] |
| .long 0xedc30b18 // vstr d16, [r3, #96] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0x3f800000 // .word 0x3f800000 |
| .long 0x3f800000 // .word 0x3f800000 |
| |
| HIDDEN _sk_accumulate_vfp4 |
| .globl _sk_accumulate_vfp4 |
| FUNCTION(_sk_accumulate_vfp4) |
| _sk_accumulate_vfp4: |
| .long 0xe8911008 // ldm r1, {r3, ip} |
| .long 0xe2811008 // add r1, r1, #8 |
| .long 0xedd31b28 // vldr d17, [r3, #160] |
| .long 0xedd30b20 // vldr d16, [r3, #128] |
| .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| .long 0xf2004c90 // vfma.f32 d4, d16, d0 |
| .long 0xf2005c91 // vfma.f32 d5, d16, d1 |
| .long 0xf2006c92 // vfma.f32 d6, d16, d2 |
| .long 0xf2007c93 // vfma.f32 d7, d16, d3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_bilinear_nx_vfp4 |
| .globl _sk_bilinear_nx_vfp4 |
| FUNCTION(_sk_bilinear_nx_vfp4) |
| _sk_bilinear_nx_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xedd32b10 // vldr d18, [r3, #64] |
| .long 0xf2600da2 // vsub.f32 d16, d16, d18 |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xf3c3261f // vmov.i32 d18, #-1090519040 |
| .long 0xf2010da2 // vadd.f32 d0, d17, d18 |
| .long 0xedc30b20 // vstr d16, [r3, #128] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_bilinear_px_vfp4 |
| .globl _sk_bilinear_px_vfp4 |
| FUNCTION(_sk_bilinear_px_vfp4) |
| _sk_bilinear_px_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c3061f // vmov.i32 d16, #1056964608 |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xedd32b10 // vldr d18, [r3, #64] |
| .long 0xf2010da0 // vadd.f32 d0, d17, d16 |
| .long 0xedc32b20 // vstr d18, [r3, #128] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_bilinear_ny_vfp4 |
| .globl _sk_bilinear_ny_vfp4 |
| FUNCTION(_sk_bilinear_ny_vfp4) |
| _sk_bilinear_ny_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xedd32b18 // vldr d18, [r3, #96] |
| .long 0xf2600da2 // vsub.f32 d16, d16, d18 |
| .long 0xedd31b08 // vldr d17, [r3, #32] |
| .long 0xf3c3261f // vmov.i32 d18, #-1090519040 |
| .long 0xf2011da2 // vadd.f32 d1, d17, d18 |
| .long 0xedc30b28 // vstr d16, [r3, #160] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_bilinear_py_vfp4 |
| .globl _sk_bilinear_py_vfp4 |
| FUNCTION(_sk_bilinear_py_vfp4) |
| _sk_bilinear_py_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c3061f // vmov.i32 d16, #1056964608 |
| .long 0xedd31b08 // vldr d17, [r3, #32] |
| .long 0xedd32b18 // vldr d18, [r3, #96] |
| .long 0xf2011da0 // vadd.f32 d1, d17, d16 |
| .long 0xedc32b28 // vstr d18, [r3, #160] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| |
| HIDDEN _sk_bicubic_n3x_vfp4 |
| .globl _sk_bicubic_n3x_vfp4 |
| FUNCTION(_sk_bicubic_n3x_vfp4) |
| _sk_bicubic_n3x_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xeddf3b10 // vldr d19, [pc, #64] |
| .long 0xedd32b10 // vldr d18, [r3, #64] |
| .long 0xf2600da2 // vsub.f32 d16, d16, d18 |
| .long 0xeddf2b0b // vldr d18, [pc, #44] |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xf2403cb2 // vfma.f32 d19, d16, d18 |
| .long 0xf3400db0 // vmul.f32 d16, d16, d16 |
| .long 0xf3c72f18 // vmov.f32 d18, #-1.5 |
| .long 0xf2010da2 // vadd.f32 d0, d17, d18 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xedc30b20 // vstr d16, [r3, #128] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| |
| HIDDEN _sk_bicubic_n1x_vfp4 |
| .globl _sk_bicubic_n1x_vfp4 |
| FUNCTION(_sk_bicubic_n1x_vfp4) |
| _sk_bicubic_n1x_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xf2c73f18 // vmov.f32 d19, #1.5 |
| .long 0xedd32b10 // vldr d18, [r3, #64] |
| .long 0xf2600da2 // vsub.f32 d16, d16, d18 |
| .long 0xeddf2b0d // vldr d18, [pc, #52] |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xf2403cb2 // vfma.f32 d19, d16, d18 |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2402cb3 // vfma.f32 d18, d16, d19 |
| .long 0xeddf3b0a // vldr d19, [pc, #40] |
| .long 0xf2403cb2 // vfma.f32 d19, d16, d18 |
| .long 0xf3c3061f // vmov.i32 d16, #-1090519040 |
| .long 0xf2010da0 // vadd.f32 d0, d17, d16 |
| .long 0xedc33b20 // vstr d19, [r3, #128] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| |
| HIDDEN _sk_bicubic_p1x_vfp4 |
| .globl _sk_bicubic_p1x_vfp4 |
| FUNCTION(_sk_bicubic_p1x_vfp4) |
| _sk_bicubic_p1x_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c71f18 // vmov.f32 d17, #1.5 |
| .long 0xeddf0b0c // vldr d16, [pc, #48] |
| .long 0xedd33b10 // vldr d19, [r3, #64] |
| .long 0xf2431cb0 // vfma.f32 d17, d19, d16 |
| .long 0xedd32b00 // vldr d18, [r3] |
| .long 0xf2c3061f // vmov.i32 d16, #1056964608 |
| .long 0xf2020da0 // vadd.f32 d0, d18, d16 |
| .long 0xf2430cb1 // vfma.f32 d16, d19, d17 |
| .long 0xeddf1b07 // vldr d17, [pc, #28] |
| .long 0xf2431cb0 // vfma.f32 d17, d19, d16 |
| .long 0xedc31b20 // vstr d17, [r3, #128] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| |
| HIDDEN _sk_bicubic_p3x_vfp4 |
| .globl _sk_bicubic_p3x_vfp4 |
| FUNCTION(_sk_bicubic_p3x_vfp4) |
| _sk_bicubic_p3x_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xeddf0b0d // vldr d16, [pc, #52] |
| .long 0xeddf3b0e // vldr d19, [pc, #56] |
| .long 0xedd32b10 // vldr d18, [r3, #64] |
| .long 0xf2423cb0 // vfma.f32 d19, d18, d16 |
| .long 0xedd31b00 // vldr d17, [r3] |
| .long 0xf3420db2 // vmul.f32 d16, d18, d18 |
| .long 0xf2c72f18 // vmov.f32 d18, #1.5 |
| .long 0xf2010da2 // vadd.f32 d0, d17, d18 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xedc30b20 // vstr d16, [r3, #128] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| |
| HIDDEN _sk_bicubic_n3y_vfp4 |
| .globl _sk_bicubic_n3y_vfp4 |
| FUNCTION(_sk_bicubic_n3y_vfp4) |
| _sk_bicubic_n3y_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xeddf3b10 // vldr d19, [pc, #64] |
| .long 0xedd32b18 // vldr d18, [r3, #96] |
| .long 0xf2600da2 // vsub.f32 d16, d16, d18 |
| .long 0xeddf2b0b // vldr d18, [pc, #44] |
| .long 0xedd31b08 // vldr d17, [r3, #32] |
| .long 0xf2403cb2 // vfma.f32 d19, d16, d18 |
| .long 0xf3400db0 // vmul.f32 d16, d16, d16 |
| .long 0xf3c72f18 // vmov.f32 d18, #-1.5 |
| .long 0xf2011da2 // vadd.f32 d1, d17, d18 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xedc30b28 // vstr d16, [r3, #160] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| |
| HIDDEN _sk_bicubic_n1y_vfp4 |
| .globl _sk_bicubic_n1y_vfp4 |
| FUNCTION(_sk_bicubic_n1y_vfp4) |
| _sk_bicubic_n1y_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c70f10 // vmov.f32 d16, #1 |
| .long 0xf2c73f18 // vmov.f32 d19, #1.5 |
| .long 0xedd32b18 // vldr d18, [r3, #96] |
| .long 0xf2600da2 // vsub.f32 d16, d16, d18 |
| .long 0xeddf2b0d // vldr d18, [pc, #52] |
| .long 0xedd31b08 // vldr d17, [r3, #32] |
| .long 0xf2403cb2 // vfma.f32 d19, d16, d18 |
| .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| .long 0xf2402cb3 // vfma.f32 d18, d16, d19 |
| .long 0xeddf3b0a // vldr d19, [pc, #40] |
| .long 0xf2403cb2 // vfma.f32 d19, d16, d18 |
| .long 0xf3c3061f // vmov.i32 d16, #-1090519040 |
| .long 0xf2011da0 // vadd.f32 d1, d17, d16 |
| .long 0xedc33b28 // vstr d19, [r3, #160] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| |
| HIDDEN _sk_bicubic_p1y_vfp4 |
| .globl _sk_bicubic_p1y_vfp4 |
| FUNCTION(_sk_bicubic_p1y_vfp4) |
| _sk_bicubic_p1y_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xf2c71f18 // vmov.f32 d17, #1.5 |
| .long 0xeddf0b0c // vldr d16, [pc, #48] |
| .long 0xedd33b18 // vldr d19, [r3, #96] |
| .long 0xf2431cb0 // vfma.f32 d17, d19, d16 |
| .long 0xedd32b08 // vldr d18, [r3, #32] |
| .long 0xf2c3061f // vmov.i32 d16, #1056964608 |
| .long 0xf2021da0 // vadd.f32 d1, d18, d16 |
| .long 0xf2430cb1 // vfma.f32 d16, d19, d17 |
| .long 0xeddf1b07 // vldr d17, [pc, #28] |
| .long 0xf2431cb0 // vfma.f32 d17, d19, d16 |
| .long 0xedc31b28 // vstr d17, [r3, #160] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0xbf955555 // .word 0xbf955555 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| .long 0x3d638e39 // .word 0x3d638e39 |
| |
| HIDDEN _sk_bicubic_p3y_vfp4 |
| .globl _sk_bicubic_p3y_vfp4 |
| FUNCTION(_sk_bicubic_p3y_vfp4) |
| _sk_bicubic_p3y_vfp4: |
| .long 0xe5913000 // ldr r3, [r1] |
| .long 0xeddf0b0d // vldr d16, [pc, #52] |
| .long 0xeddf3b0e // vldr d19, [pc, #56] |
| .long 0xedd32b18 // vldr d18, [r3, #96] |
| .long 0xf2423cb0 // vfma.f32 d19, d18, d16 |
| .long 0xedd31b08 // vldr d17, [r3, #32] |
| .long 0xf3420db2 // vmul.f32 d16, d18, d18 |
| .long 0xf2c72f18 // vmov.f32 d18, #1.5 |
| .long 0xf2011da2 // vadd.f32 d1, d17, d18 |
| .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| .long 0xedc30b28 // vstr d16, [r3, #160] |
| .long 0xe2813008 // add r3, r1, #8 |
| .long 0xe591c004 // ldr ip, [r1, #4] |
| .long 0xe1a01003 // mov r1, r3 |
| .long 0xe12fff1c // bx ip |
| .long 0xe320f000 // nop {0} |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0x3ec71c72 // .word 0x3ec71c72 |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| .long 0xbeaaaaab // .word 0xbeaaaaab |
| |
| HIDDEN _sk_callback_vfp4 |
| .globl _sk_callback_vfp4 |
| FUNCTION(_sk_callback_vfp4) |
| _sk_callback_vfp4: |
| .long 0xe92d4070 // push {r4, r5, r6, lr} |
| .long 0xed2d8b10 // vpush {d8-d15} |
| .long 0xe1a05001 // mov r5, r1 |
| .long 0xe1a06000 // mov r6, r0 |
| .long 0xe5950000 // ldr r0, [r5] |
| .long 0xe1a04002 // mov r4, r2 |
| .long 0xe3a01002 // mov r1, #2 |
| .long 0xeeb08b47 // vmov.f64 d8, d7 |
| .long 0xe5902000 // ldr r2, [r0] |
| .long 0xe5900004 // ldr r0, [r0, #4] |
| .long 0xeeb09b46 // vmov.f64 d9, d6 |
| .long 0xeeb0ab45 // vmov.f64 d10, d5 |
| .long 0xeeb0bb44 // vmov.f64 d11, d4 |
| .long 0xeeb0cb43 // vmov.f64 d12, d3 |
| .long 0xeeb0db42 // vmov.f64 d13, d2 |
| .long 0xeeb0eb41 // vmov.f64 d14, d1 |
| .long 0xeeb0fb40 // vmov.f64 d15, d0 |
| .long 0xe12fff32 // blx r2 |
| .long 0xe2851008 // add r1, r5, #8 |
| .long 0xe5953004 // ldr r3, [r5, #4] |
| .long 0xe1a00006 // mov r0, r6 |
| .long 0xe1a02004 // mov r2, r4 |
| .long 0xeeb00b4f // vmov.f64 d0, d15 |
| .long 0xeeb01b4e // vmov.f64 d1, d14 |
| .long 0xeeb02b4d // vmov.f64 d2, d13 |
| .long 0xeeb03b4c // vmov.f64 d3, d12 |
| .long 0xeeb04b4b // vmov.f64 d4, d11 |
| .long 0xeeb05b4a // vmov.f64 d5, d10 |
| .long 0xeeb06b49 // vmov.f64 d6, d9 |
| .long 0xeeb07b48 // vmov.f64 d7, d8 |
| .long 0xecbd8b10 // vpop {d8-d15} |
| .long 0xe8bd4070 // pop {r4, r5, r6, lr} |
| .long 0xe12fff13 // bx r3 |
| #elif defined(__x86_64__) |
| |
| HIDDEN _sk_start_pipeline_hsw |
| .globl _sk_start_pipeline_hsw |
| FUNCTION(_sk_start_pipeline_hsw) |
| _sk_start_pipeline_hsw: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 73,137,205 // mov %rcx,%r13 |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 72,137,251 // mov %rdi,%rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,199 // mov %rax,%r15 |
| .byte 73,137,244 // mov %rsi,%r12 |
| .byte 72,141,67,8 // lea 0x8(%rbx),%rax |
| .byte 76,57,232 // cmp %r13,%rax |
| .byte 118,5 // jbe 28 <_sk_start_pipeline_hsw+0x28> |
| .byte 72,137,223 // mov %rbx,%rdi |
| .byte 235,65 // jmp 69 <_sk_start_pipeline_hsw+0x69> |
| .byte 185,0,0,0,0 // mov $0x0,%ecx |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| .byte 72,137,223 // mov %rbx,%rdi |
| .byte 76,137,230 // mov %r12,%rsi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 65,255,215 // callq *%r15 |
| .byte 72,141,123,8 // lea 0x8(%rbx),%rdi |
| .byte 72,131,195,16 // add $0x10,%rbx |
| .byte 76,57,235 // cmp %r13,%rbx |
| .byte 72,137,251 // mov %rdi,%rbx |
| .byte 118,191 // jbe 28 <_sk_start_pipeline_hsw+0x28> |
| .byte 76,137,233 // mov %r13,%rcx |
| .byte 72,41,249 // sub %rdi,%rcx |
| .byte 116,41 // je 9a <_sk_start_pipeline_hsw+0x9a> |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| .byte 76,137,230 // mov %r12,%rsi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 65,255,215 // callq *%r15 |
| .byte 76,137,232 // mov %r13,%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 197,248,119 // vzeroupper |
| .byte 195 // retq |
| |
| HIDDEN _sk_just_return_hsw |
| .globl _sk_just_return_hsw |
| FUNCTION(_sk_just_return_hsw) |
| _sk_just_return_hsw: |
| .byte 195 // retq |
| |
| HIDDEN _sk_seed_shader_hsw |
| .globl _sk_seed_shader_hsw |
| FUNCTION(_sk_seed_shader_hsw) |
| _sk_seed_shader_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,249,110,199 // vmovd %edi,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 |
| .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_constant_color_hsw |
| .globl _sk_constant_color_hsw |
| FUNCTION(_sk_constant_color_hsw) |
| _sk_constant_color_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clear_hsw |
| .globl _sk_clear_hsw |
| FUNCTION(_sk_clear_hsw) |
| _sk_clear_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcatop_hsw |
| .globl _sk_srcatop_hsw |
| FUNCTION(_sk_srcatop_hsw) |
| _sk_srcatop_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 |
| .byte 196,194,69,168,193 // vfmadd213ps %ymm9,%ymm7,%ymm0 |
| .byte 197,60,89,205 // vmulps %ymm5,%ymm8,%ymm9 |
| .byte 196,194,69,168,201 // vfmadd213ps %ymm9,%ymm7,%ymm1 |
| .byte 197,60,89,206 // vmulps %ymm6,%ymm8,%ymm9 |
| .byte 196,194,69,168,209 // vfmadd213ps %ymm9,%ymm7,%ymm2 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstatop_hsw |
| .globl _sk_dstatop_hsw |
| FUNCTION(_sk_dstatop_hsw) |
| _sk_dstatop_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 196,226,101,184,196 // vfmadd231ps %ymm4,%ymm3,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 196,226,101,184,205 // vfmadd231ps %ymm5,%ymm3,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 196,226,101,184,214 // vfmadd231ps %ymm6,%ymm3,%ymm2 |
| .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcin_hsw |
| .globl _sk_srcin_hsw |
| FUNCTION(_sk_srcin_hsw) |
| _sk_srcin_hsw: |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstin_hsw |
| .globl _sk_dstin_hsw |
| FUNCTION(_sk_dstin_hsw) |
| _sk_dstin_hsw: |
| .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcout_hsw |
| .globl _sk_srcout_hsw |
| FUNCTION(_sk_srcout_hsw) |
| _sk_srcout_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstout_hsw |
| .globl _sk_dstout_hsw |
| FUNCTION(_sk_dstout_hsw) |
| _sk_dstout_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 |
| .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcover_hsw |
| .globl _sk_srcover_hsw |
| FUNCTION(_sk_srcover_hsw) |
| _sk_srcover_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0 |
| .byte 196,194,85,184,200 // vfmadd231ps %ymm8,%ymm5,%ymm1 |
| .byte 196,194,77,184,208 // vfmadd231ps %ymm8,%ymm6,%ymm2 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstover_hsw |
| .globl _sk_dstover_hsw |
| FUNCTION(_sk_dstover_hsw) |
| _sk_dstover_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 |
| .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 |
| .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2 |
| .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_modulate_hsw |
| .globl _sk_modulate_hsw |
| FUNCTION(_sk_modulate_hsw) |
| _sk_modulate_hsw: |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_multiply_hsw |
| .globl _sk_multiply_hsw |
| FUNCTION(_sk_multiply_hsw) |
| _sk_multiply_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 |
| .byte 196,98,53,184,208 // vfmadd231ps %ymm0,%ymm9,%ymm10 |
| .byte 196,194,93,168,194 // vfmadd213ps %ymm10,%ymm4,%ymm0 |
| .byte 197,52,89,209 // vmulps %ymm1,%ymm9,%ymm10 |
| .byte 196,98,61,184,213 // vfmadd231ps %ymm5,%ymm8,%ymm10 |
| .byte 196,194,85,168,202 // vfmadd213ps %ymm10,%ymm5,%ymm1 |
| .byte 197,52,89,210 // vmulps %ymm2,%ymm9,%ymm10 |
| .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10 |
| .byte 196,194,77,168,210 // vfmadd213ps %ymm10,%ymm6,%ymm2 |
| .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| .byte 196,66,69,168,193 // vfmadd213ps %ymm9,%ymm7,%ymm8 |
| .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_plus__hsw |
| .globl _sk_plus__hsw |
| FUNCTION(_sk_plus__hsw) |
| _sk_plus__hsw: |
| .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_screen_hsw |
| .globl _sk_screen_hsw |
| FUNCTION(_sk_screen_hsw) |
| _sk_screen_hsw: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 196,194,93,172,192 // vfnmadd213ps %ymm8,%ymm4,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 196,194,85,172,200 // vfnmadd213ps %ymm8,%ymm5,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 196,194,77,172,208 // vfnmadd213ps %ymm8,%ymm6,%ymm2 |
| .byte 197,100,88,199 // vaddps %ymm7,%ymm3,%ymm8 |
| .byte 196,194,69,172,216 // vfnmadd213ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_xor__hsw |
| .globl _sk_xor__hsw |
| FUNCTION(_sk_xor__hsw) |
| _sk_xor__hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 |
| .byte 196,194,53,168,194 // vfmadd213ps %ymm10,%ymm9,%ymm0 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 196,226,61,184,205 // vfmadd231ps %ymm5,%ymm8,%ymm1 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 196,226,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm2 |
| .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| .byte 196,98,69,168,195 // vfmadd213ps %ymm3,%ymm7,%ymm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,195 // vmovaps %ymm8,%ymm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_darken_hsw |
| .globl _sk_darken_hsw |
| FUNCTION(_sk_darken_hsw) |
| _sk_darken_hsw: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| .byte 196,193,124,95,193 // vmaxps %ymm9,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| .byte 196,193,116,95,201 // vmaxps %ymm9,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lighten_hsw |
| .globl _sk_lighten_hsw |
| FUNCTION(_sk_lighten_hsw) |
| _sk_lighten_hsw: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_difference_hsw |
| .globl _sk_difference_hsw |
| FUNCTION(_sk_difference_hsw) |
| _sk_difference_hsw: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_exclusion_hsw |
| .globl _sk_exclusion_hsw |
| FUNCTION(_sk_exclusion_hsw) |
| _sk_exclusion_hsw: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colorburn_hsw |
| .globl _sk_colorburn_hsw |
| FUNCTION(_sk_colorburn_hsw) |
| _sk_colorburn_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 |
| .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,228 // vmulps %ymm4,%ymm8,%ymm12 |
| .byte 197,68,92,236 // vsubps %ymm4,%ymm7,%ymm13 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 197,20,94,232 // vdivps %ymm0,%ymm13,%ymm13 |
| .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| .byte 196,66,101,168,235 // vfmadd213ps %ymm11,%ymm3,%ymm13 |
| .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| .byte 197,28,88,224 // vaddps %ymm0,%ymm12,%ymm12 |
| .byte 196,193,124,194,194,0 // vcmpeqps %ymm10,%ymm0,%ymm0 |
| .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0 |
| .byte 197,92,194,231,0 // vcmpeqps %ymm7,%ymm4,%ymm12 |
| .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| .byte 197,52,89,217 // vmulps %ymm1,%ymm9,%ymm11 |
| .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| .byte 197,68,92,237 // vsubps %ymm5,%ymm7,%ymm13 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 197,20,94,233 // vdivps %ymm1,%ymm13,%ymm13 |
| .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| .byte 196,66,101,168,235 // vfmadd213ps %ymm11,%ymm3,%ymm13 |
| .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| .byte 197,28,88,225 // vaddps %ymm1,%ymm12,%ymm12 |
| .byte 196,193,116,194,202,0 // vcmpeqps %ymm10,%ymm1,%ymm1 |
| .byte 196,195,21,74,204,16 // vblendvps %ymm1,%ymm12,%ymm13,%ymm1 |
| .byte 197,84,194,231,0 // vcmpeqps %ymm7,%ymm5,%ymm12 |
| .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9 |
| .byte 196,65,108,194,210,0 // vcmpeqps %ymm10,%ymm2,%ymm10 |
| .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| .byte 197,68,92,230 // vsubps %ymm6,%ymm7,%ymm12 |
| .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12 |
| .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 |
| .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| .byte 196,65,68,92,228 // vsubps %ymm12,%ymm7,%ymm12 |
| .byte 196,66,101,168,225 // vfmadd213ps %ymm9,%ymm3,%ymm12 |
| .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| .byte 196,227,37,74,210,160 // vblendvps %ymm10,%ymm2,%ymm11,%ymm2 |
| .byte 197,76,194,215,0 // vcmpeqps %ymm7,%ymm6,%ymm10 |
| .byte 197,52,88,206 // vaddps %ymm6,%ymm9,%ymm9 |
| .byte 196,195,109,74,209,160 // vblendvps %ymm10,%ymm9,%ymm2,%ymm2 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colordodge_hsw |
| .globl _sk_colordodge_hsw |
| FUNCTION(_sk_colordodge_hsw) |
| _sk_colordodge_hsw: |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 |
| .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 |
| .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 |
| .byte 197,100,89,228 // vmulps %ymm4,%ymm3,%ymm12 |
| .byte 197,100,92,232 // vsubps %ymm0,%ymm3,%ymm13 |
| .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| .byte 197,52,89,236 // vmulps %ymm4,%ymm9,%ymm13 |
| .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| .byte 196,66,101,168,227 // vfmadd213ps %ymm11,%ymm3,%ymm12 |
| .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| .byte 197,20,88,232 // vaddps %ymm0,%ymm13,%ymm13 |
| .byte 197,252,194,195,0 // vcmpeqps %ymm3,%ymm0,%ymm0 |
| .byte 196,195,29,74,197,0 // vblendvps %ymm0,%ymm13,%ymm12,%ymm0 |
| .byte 196,65,92,194,224,0 // vcmpeqps %ymm8,%ymm4,%ymm12 |
| .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| .byte 197,100,89,229 // vmulps %ymm5,%ymm3,%ymm12 |
| .byte 197,100,92,233 // vsubps %ymm1,%ymm3,%ymm13 |
| .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| .byte 197,52,89,237 // vmulps %ymm5,%ymm9,%ymm13 |
| .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| .byte 196,66,101,168,227 // vfmadd213ps %ymm11,%ymm3,%ymm12 |
| .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| .byte 197,20,88,233 // vaddps %ymm1,%ymm13,%ymm13 |
| .byte 197,244,194,203,0 // vcmpeqps %ymm3,%ymm1,%ymm1 |
| .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1 |
| .byte 196,65,84,194,224,0 // vcmpeqps %ymm8,%ymm5,%ymm12 |
| .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11 |
| .byte 197,100,92,226 // vsubps %ymm2,%ymm3,%ymm12 |
| .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 |
| .byte 197,52,89,230 // vmulps %ymm6,%ymm9,%ymm12 |
| .byte 196,65,68,93,219 // vminps %ymm11,%ymm7,%ymm11 |
| .byte 196,66,101,168,218 // vfmadd213ps %ymm10,%ymm3,%ymm11 |
| .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12 |
| .byte 197,236,194,211,0 // vcmpeqps %ymm3,%ymm2,%ymm2 |
| .byte 196,195,37,74,212,32 // vblendvps %ymm2,%ymm12,%ymm11,%ymm2 |
| .byte 196,65,76,194,192,0 // vcmpeqps %ymm8,%ymm6,%ymm8 |
| .byte 197,44,88,214 // vaddps %ymm6,%ymm10,%ymm10 |
| .byte 196,195,109,74,210,128 // vblendvps %ymm8,%ymm10,%ymm2,%ymm2 |
| .byte 196,194,69,184,217 // vfmadd231ps %ymm9,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hardlight_hsw |
| .globl _sk_hardlight_hsw |
| FUNCTION(_sk_hardlight_hsw) |
| _sk_hardlight_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11 |
| .byte 197,124,88,200 // vaddps %ymm0,%ymm0,%ymm9 |
| .byte 197,52,194,227,2 // vcmpleps %ymm3,%ymm9,%ymm12 |
| .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| .byte 196,98,61,184,221 // vfmadd231ps %ymm5,%ymm8,%ymm11 |
| .byte 197,116,88,225 // vaddps %ymm1,%ymm1,%ymm12 |
| .byte 197,28,194,227,2 // vcmpleps %ymm3,%ymm12,%ymm12 |
| .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10 |
| .byte 197,108,88,218 // vaddps %ymm2,%ymm2,%ymm11 |
| .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_overlay_hsw |
| .globl _sk_overlay_hsw |
| FUNCTION(_sk_overlay_hsw) |
| _sk_overlay_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11 |
| .byte 197,92,88,204 // vaddps %ymm4,%ymm4,%ymm9 |
| .byte 197,52,194,231,2 // vcmpleps %ymm7,%ymm9,%ymm12 |
| .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| .byte 196,98,61,184,221 // vfmadd231ps %ymm5,%ymm8,%ymm11 |
| .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12 |
| .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10 |
| .byte 197,76,88,222 // vaddps %ymm6,%ymm6,%ymm11 |
| .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11 |
| .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_softlight_hsw |
| .globl _sk_softlight_hsw |
| FUNCTION(_sk_softlight_hsw) |
| _sk_softlight_hsw: |
| .byte 197,252,17,84,36,200 // vmovups %ymm2,-0x38(%rsp) |
| .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| .byte 197,44,194,223,1 // vcmpltps %ymm7,%ymm10,%ymm11 |
| .byte 197,92,94,199 // vdivps %ymm7,%ymm4,%ymm8 |
| .byte 196,67,45,74,224,176 // vblendvps %ymm11,%ymm8,%ymm10,%ymm12 |
| .byte 196,65,28,88,196 // vaddps %ymm12,%ymm12,%ymm8 |
| .byte 196,65,60,88,232 // vaddps %ymm8,%ymm8,%ymm13 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,66,21,168,237 // vfmadd213ps %ymm13,%ymm13,%ymm13 |
| .byte 196,65,28,92,240 // vsubps %ymm8,%ymm12,%ymm14 |
| .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 196,65,28,89,249 // vmulps %ymm9,%ymm12,%ymm15 |
| .byte 196,66,21,184,254 // vfmadd231ps %ymm14,%ymm13,%ymm15 |
| .byte 196,65,124,82,236 // vrsqrtps %ymm12,%ymm13 |
| .byte 196,65,124,83,237 // vrcpps %ymm13,%ymm13 |
| .byte 196,65,20,92,236 // vsubps %ymm12,%ymm13,%ymm13 |
| .byte 197,92,88,244 // vaddps %ymm4,%ymm4,%ymm14 |
| .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14 |
| .byte 197,12,194,247,2 // vcmpleps %ymm7,%ymm14,%ymm14 |
| .byte 196,67,21,74,239,224 // vblendvps %ymm14,%ymm15,%ymm13,%ymm13 |
| .byte 197,124,88,240 // vaddps %ymm0,%ymm0,%ymm14 |
| .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| .byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12 |
| .byte 196,98,5,168,227 // vfmadd213ps %ymm3,%ymm15,%ymm12 |
| .byte 197,28,89,228 // vmulps %ymm4,%ymm12,%ymm12 |
| .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13 |
| .byte 196,98,101,184,236 // vfmadd231ps %ymm4,%ymm3,%ymm13 |
| .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| .byte 196,195,21,74,212,224 // vblendvps %ymm14,%ymm12,%ymm13,%ymm2 |
| .byte 197,84,94,239 // vdivps %ymm7,%ymm5,%ymm13 |
| .byte 196,67,45,74,237,176 // vblendvps %ymm11,%ymm13,%ymm10,%ymm13 |
| .byte 196,65,20,88,245 // vaddps %ymm13,%ymm13,%ymm14 |
| .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14 |
| .byte 196,66,13,168,246 // vfmadd213ps %ymm14,%ymm14,%ymm14 |
| .byte 196,65,20,92,248 // vsubps %ymm8,%ymm13,%ymm15 |
| .byte 196,65,4,89,246 // vmulps %ymm14,%ymm15,%ymm14 |
| .byte 196,66,53,184,245 // vfmadd231ps %ymm13,%ymm9,%ymm14 |
| .byte 196,65,124,82,253 // vrsqrtps %ymm13,%ymm15 |
| .byte 196,65,124,83,255 // vrcpps %ymm15,%ymm15 |
| .byte 196,65,4,92,253 // vsubps %ymm13,%ymm15,%ymm15 |
| .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| .byte 196,67,5,74,230,192 // vblendvps %ymm12,%ymm14,%ymm15,%ymm12 |
| .byte 197,116,88,241 // vaddps %ymm1,%ymm1,%ymm14 |
| .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13 |
| .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| .byte 196,98,5,168,235 // vfmadd213ps %ymm3,%ymm15,%ymm13 |
| .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| .byte 196,65,4,89,228 // vmulps %ymm12,%ymm15,%ymm12 |
| .byte 197,20,89,237 // vmulps %ymm5,%ymm13,%ymm13 |
| .byte 196,98,101,184,229 // vfmadd231ps %ymm5,%ymm3,%ymm12 |
| .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| .byte 196,67,29,74,237,224 // vblendvps %ymm14,%ymm13,%ymm12,%ymm13 |
| .byte 197,76,94,231 // vdivps %ymm7,%ymm6,%ymm12 |
| .byte 196,67,45,74,212,176 // vblendvps %ymm11,%ymm12,%ymm10,%ymm10 |
| .byte 196,65,44,88,218 // vaddps %ymm10,%ymm10,%ymm11 |
| .byte 196,65,36,88,219 // vaddps %ymm11,%ymm11,%ymm11 |
| .byte 196,66,37,168,219 // vfmadd213ps %ymm11,%ymm11,%ymm11 |
| .byte 196,65,44,92,224 // vsubps %ymm8,%ymm10,%ymm12 |
| .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11 |
| .byte 196,66,45,168,203 // vfmadd213ps %ymm11,%ymm10,%ymm9 |
| .byte 196,65,124,82,218 // vrsqrtps %ymm10,%ymm11 |
| .byte 196,65,124,83,219 // vrcpps %ymm11,%ymm11 |
| .byte 196,65,36,92,218 // vsubps %ymm10,%ymm11,%ymm11 |
| .byte 197,76,88,230 // vaddps %ymm6,%ymm6,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| .byte 196,67,37,74,201,192 // vblendvps %ymm12,%ymm9,%ymm11,%ymm9 |
| .byte 197,124,16,116,36,200 // vmovups -0x38(%rsp),%ymm14 |
| .byte 196,65,12,88,222 // vaddps %ymm14,%ymm14,%ymm11 |
| .byte 197,36,92,227 // vsubps %ymm3,%ymm11,%ymm12 |
| .byte 196,65,60,92,210 // vsubps %ymm10,%ymm8,%ymm10 |
| .byte 196,98,29,168,211 // vfmadd213ps %ymm3,%ymm12,%ymm10 |
| .byte 197,28,89,231 // vmulps %ymm7,%ymm12,%ymm12 |
| .byte 196,65,28,89,201 // vmulps %ymm9,%ymm12,%ymm9 |
| .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10 |
| .byte 196,98,101,184,206 // vfmadd231ps %ymm6,%ymm3,%ymm9 |
| .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| .byte 196,67,53,74,202,176 // vblendvps %ymm11,%ymm10,%ymm9,%ymm9 |
| .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11 |
| .byte 196,193,108,88,195 // vaddps %ymm11,%ymm2,%ymm0 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,226,61,184,205 // vfmadd231ps %ymm5,%ymm8,%ymm1 |
| .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1 |
| .byte 196,193,44,89,214 // vmulps %ymm14,%ymm10,%ymm2 |
| .byte 196,226,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm2 |
| .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 |
| .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_0_hsw |
| .globl _sk_clamp_0_hsw |
| FUNCTION(_sk_clamp_0_hsw) |
| _sk_clamp_0_hsw: |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_1_hsw |
| .globl _sk_clamp_1_hsw |
| FUNCTION(_sk_clamp_1_hsw) |
| _sk_clamp_1_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 |
| .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_a_hsw |
| .globl _sk_clamp_a_hsw |
| FUNCTION(_sk_clamp_a_hsw) |
| _sk_clamp_a_hsw: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 |
| .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 |
| .byte 197,236,93,211 // vminps %ymm3,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_set_rgb_hsw |
| .globl _sk_set_rgb_hsw |
| FUNCTION(_sk_set_rgb_hsw) |
| _sk_set_rgb_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_rb_hsw |
| .globl _sk_swap_rb_hsw |
| FUNCTION(_sk_swap_rb_hsw) |
| _sk_swap_rb_hsw: |
| .byte 197,124,40,192 // vmovaps %ymm0,%ymm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,194 // vmovaps %ymm2,%ymm0 |
| .byte 197,124,41,194 // vmovaps %ymm8,%ymm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_hsw |
| .globl _sk_swap_hsw |
| FUNCTION(_sk_swap_hsw) |
| _sk_swap_hsw: |
| .byte 197,124,40,195 // vmovaps %ymm3,%ymm8 |
| .byte 197,124,40,202 // vmovaps %ymm2,%ymm9 |
| .byte 197,124,40,209 // vmovaps %ymm1,%ymm10 |
| .byte 197,124,40,216 // vmovaps %ymm0,%ymm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| .byte 197,124,41,220 // vmovaps %ymm11,%ymm4 |
| .byte 197,124,41,213 // vmovaps %ymm10,%ymm5 |
| .byte 197,124,41,206 // vmovaps %ymm9,%ymm6 |
| .byte 197,124,41,199 // vmovaps %ymm8,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_src_dst_hsw |
| .globl _sk_move_src_dst_hsw |
| FUNCTION(_sk_move_src_dst_hsw) |
| _sk_move_src_dst_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,224 // vmovaps %ymm0,%ymm4 |
| .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 |
| .byte 197,252,40,242 // vmovaps %ymm2,%ymm6 |
| .byte 197,252,40,251 // vmovaps %ymm3,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_dst_src_hsw |
| .globl _sk_move_dst_src_hsw |
| FUNCTION(_sk_move_dst_src_hsw) |
| _sk_move_dst_src_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_premul_hsw |
| .globl _sk_premul_hsw |
| FUNCTION(_sk_premul_hsw) |
| _sk_premul_hsw: |
| .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 |
| .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_unpremul_hsw |
| .globl _sk_unpremul_hsw |
| FUNCTION(_sk_unpremul_hsw) |
| _sk_unpremul_hsw: |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 |
| .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_srgb_hsw |
| .globl _sk_from_srgb_hsw |
| FUNCTION(_sk_from_srgb_hsw) |
| _sk_from_srgb_hsw: |
| .byte 184,145,131,158,61 // mov $0x3d9e8391,%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 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 |
| .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| .byte 196,66,125,88,228 // vpbroadcastd %xmm12,%ymm12 |
| .byte 196,65,125,111,235 // vmovdqa %ymm11,%ymm13 |
| .byte 196,66,125,168,236 // vfmadd213ps %ymm12,%ymm0,%ymm13 |
| .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| .byte 196,66,125,88,246 // vpbroadcastd %xmm14,%ymm14 |
| .byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13 |
| .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 196,193,124,194,194,1 // vcmpltps %ymm10,%ymm0,%ymm0 |
| .byte 196,195,21,74,193,0 // vblendvps %ymm0,%ymm9,%ymm13,%ymm0 |
| .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9 |
| .byte 197,116,89,233 // vmulps %ymm1,%ymm1,%ymm13 |
| .byte 196,65,125,111,251 // vmovdqa %ymm11,%ymm15 |
| .byte 196,66,117,168,252 // vfmadd213ps %ymm12,%ymm1,%ymm15 |
| .byte 196,66,21,168,254 // vfmadd213ps %ymm14,%ymm13,%ymm15 |
| .byte 196,193,116,194,202,1 // vcmpltps %ymm10,%ymm1,%ymm1 |
| .byte 196,195,5,74,201,16 // vblendvps %ymm1,%ymm9,%ymm15,%ymm1 |
| .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| .byte 197,108,89,202 // vmulps %ymm2,%ymm2,%ymm9 |
| .byte 196,66,109,168,220 // vfmadd213ps %ymm12,%ymm2,%ymm11 |
| .byte 196,66,53,168,222 // vfmadd213ps %ymm14,%ymm9,%ymm11 |
| .byte 196,193,108,194,210,1 // vcmpltps %ymm10,%ymm2,%ymm2 |
| .byte 196,195,37,74,208,32 // vblendvps %ymm2,%ymm8,%ymm11,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_srgb_hsw |
| .globl _sk_to_srgb_hsw |
| FUNCTION(_sk_to_srgb_hsw) |
| _sk_to_srgb_hsw: |
| .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| .byte 196,65,124,83,216 // vrcpps %ymm8,%ymm11 |
| .byte 196,65,124,82,224 // vrsqrtps %ymm8,%ymm12 |
| .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,89,232 // vmulps %ymm0,%ymm8,%ymm13 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| .byte 196,66,125,88,246 // vpbroadcastd %xmm14,%ymm14 |
| .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| .byte 197,121,110,248 // vmovd %eax,%xmm15 |
| .byte 196,66,125,88,255 // vpbroadcastd %xmm15,%ymm15 |
| .byte 196,66,13,168,223 // vfmadd213ps %ymm15,%ymm14,%ymm11 |
| .byte 196,66,45,184,220 // vfmadd231ps %ymm12,%ymm10,%ymm11 |
| .byte 196,65,52,93,219 // vminps %ymm11,%ymm9,%ymm11 |
| .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| .byte 196,66,125,88,228 // vpbroadcastd %xmm12,%ymm12 |
| .byte 196,193,124,194,196,1 // vcmpltps %ymm12,%ymm0,%ymm0 |
| .byte 196,195,37,74,197,0 // vblendvps %ymm0,%ymm13,%ymm11,%ymm0 |
| .byte 197,124,82,217 // vrsqrtps %ymm1,%ymm11 |
| .byte 196,65,124,83,235 // vrcpps %ymm11,%ymm13 |
| .byte 196,65,124,82,219 // vrsqrtps %ymm11,%ymm11 |
| .byte 196,66,13,168,239 // vfmadd213ps %ymm15,%ymm14,%ymm13 |
| .byte 196,66,45,184,235 // vfmadd231ps %ymm11,%ymm10,%ymm13 |
| .byte 197,60,89,217 // vmulps %ymm1,%ymm8,%ymm11 |
| .byte 196,65,52,93,237 // vminps %ymm13,%ymm9,%ymm13 |
| .byte 196,193,116,194,204,1 // vcmpltps %ymm12,%ymm1,%ymm1 |
| .byte 196,195,21,74,203,16 // vblendvps %ymm1,%ymm11,%ymm13,%ymm1 |
| .byte 197,124,82,218 // vrsqrtps %ymm2,%ymm11 |
| .byte 196,65,124,83,235 // vrcpps %ymm11,%ymm13 |
| .byte 196,66,13,168,239 // vfmadd213ps %ymm15,%ymm14,%ymm13 |
| .byte 196,65,124,82,219 // vrsqrtps %ymm11,%ymm11 |
| .byte 196,66,45,184,235 // vfmadd231ps %ymm11,%ymm10,%ymm13 |
| .byte 196,65,52,93,205 // vminps %ymm13,%ymm9,%ymm9 |
| .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| .byte 196,193,108,194,212,1 // vcmpltps %ymm12,%ymm2,%ymm2 |
| .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_2dot2_hsw |
| .globl _sk_from_2dot2_hsw |
| FUNCTION(_sk_from_2dot2_hsw) |
| _sk_from_2dot2_hsw: |
| .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,200 // vrsqrtps %ymm8,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 197,252,89,192 // vmulps %ymm0,%ymm0,%ymm0 |
| .byte 196,65,60,89,208 // vmulps %ymm8,%ymm8,%ymm10 |
| .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 196,65,124,82,210 // vrsqrtps %ymm10,%ymm10 |
| .byte 197,244,89,201 // vmulps %ymm1,%ymm1,%ymm1 |
| .byte 196,65,52,89,217 // vmulps %ymm9,%ymm9,%ymm11 |
| .byte 196,65,52,89,203 // vmulps %ymm11,%ymm9,%ymm9 |
| .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| .byte 197,124,82,202 // vrsqrtps %ymm2,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 196,65,124,82,210 // vrsqrtps %ymm10,%ymm10 |
| .byte 197,236,89,210 // vmulps %ymm2,%ymm2,%ymm2 |
| .byte 196,65,52,89,217 // vmulps %ymm9,%ymm9,%ymm11 |
| .byte 196,65,52,89,203 // vmulps %ymm11,%ymm9,%ymm9 |
| .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2 |
| .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_2dot2_hsw |
| .globl _sk_to_2dot2_hsw |
| FUNCTION(_sk_to_2dot2_hsw) |
| _sk_to_2dot2_hsw: |
| .byte 197,252,82,192 // vrsqrtps %ymm0,%ymm0 |
| .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,200 // vrsqrtps %ymm8,%ymm9 |
| .byte 197,252,83,192 // vrcpps %ymm0,%ymm0 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 196,65,124,83,193 // vrcpps %ymm9,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| .byte 197,252,82,201 // vrsqrtps %ymm1,%ymm1 |
| .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 197,252,83,201 // vrcpps %ymm1,%ymm1 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 196,65,124,83,202 // vrcpps %ymm10,%ymm9 |
| .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 |
| .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| .byte 197,252,82,210 // vrsqrtps %ymm2,%ymm2 |
| .byte 197,124,82,202 // vrsqrtps %ymm2,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 197,252,83,210 // vrcpps %ymm2,%ymm2 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 196,65,124,83,202 // vrcpps %ymm10,%ymm9 |
| .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2 |
| .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_rgb_to_hsl_hsw |
| .globl _sk_rgb_to_hsl_hsw |
| FUNCTION(_sk_rgb_to_hsl_hsw) |
| _sk_rgb_to_hsl_hsw: |
| .byte 197,252,17,124,36,200 // vmovups %ymm7,-0x38(%rsp) |
| .byte 197,252,40,254 // vmovaps %ymm6,%ymm7 |
| .byte 197,252,40,245 // vmovaps %ymm5,%ymm6 |
| .byte 197,252,40,236 // vmovaps %ymm4,%ymm5 |
| .byte 197,252,40,227 // vmovaps %ymm3,%ymm4 |
| .byte 197,252,40,216 // vmovaps %ymm0,%ymm3 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d |
| .byte 184,0,0,192,64 // mov $0x40c00000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 65,185,0,0,0,64 // mov $0x40000000,%r9d |
| .byte 184,0,0,128,64 // mov $0x40800000,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 197,100,95,217 // vmaxps %ymm1,%ymm3,%ymm11 |
| .byte 197,36,95,218 // vmaxps %ymm2,%ymm11,%ymm11 |
| .byte 197,100,93,225 // vminps %ymm1,%ymm3,%ymm12 |
| .byte 197,28,93,226 // vminps %ymm2,%ymm12,%ymm12 |
| .byte 196,65,36,92,236 // vsubps %ymm12,%ymm11,%ymm13 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,65,60,94,197 // vdivps %ymm13,%ymm8,%ymm8 |
| .byte 197,116,194,242,1 // vcmpltps %ymm2,%ymm1,%ymm14 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 196,65,4,87,255 // vxorps %ymm15,%ymm15,%ymm15 |
| .byte 196,67,5,74,201,224 // vblendvps %ymm14,%ymm9,%ymm15,%ymm9 |
| .byte 197,116,92,242 // vsubps %ymm2,%ymm1,%ymm14 |
| .byte 196,66,61,168,241 // vfmadd213ps %ymm9,%ymm8,%ymm14 |
| .byte 197,236,92,195 // vsubps %ymm3,%ymm2,%ymm0 |
| .byte 197,100,92,201 // vsubps %ymm1,%ymm3,%ymm9 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 196,66,61,168,202 // vfmadd213ps %ymm10,%ymm8,%ymm9 |
| .byte 196,193,121,110,209 // vmovd %r9d,%xmm2 |
| .byte 196,98,125,88,210 // vpbroadcastd %xmm2,%ymm10 |
| .byte 196,194,61,168,194 // vfmadd213ps %ymm10,%ymm8,%ymm0 |
| .byte 197,164,194,201,0 // vcmpeqps %ymm1,%ymm11,%ymm1 |
| .byte 196,227,53,74,192,16 // vblendvps %ymm1,%ymm0,%ymm9,%ymm0 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 197,164,194,211,0 // vcmpeqps %ymm3,%ymm11,%ymm2 |
| .byte 196,195,125,74,198,32 // vblendvps %ymm2,%ymm14,%ymm0,%ymm0 |
| .byte 196,193,36,88,220 // vaddps %ymm12,%ymm11,%ymm3 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,228,89,209 // vmulps %ymm1,%ymm3,%ymm2 |
| .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1 |
| .byte 196,65,44,92,195 // vsubps %ymm11,%ymm10,%ymm8 |
| .byte 196,65,60,92,196 // vsubps %ymm12,%ymm8,%ymm8 |
| .byte 196,195,101,74,200,16 // vblendvps %ymm1,%ymm8,%ymm3,%ymm1 |
| .byte 196,193,36,194,220,0 // vcmpeqps %ymm12,%ymm11,%ymm3 |
| .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 |
| .byte 196,195,125,74,199,48 // vblendvps %ymm3,%ymm15,%ymm0,%ymm0 |
| .byte 196,195,117,74,207,48 // vblendvps %ymm3,%ymm15,%ymm1,%ymm1 |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,220 // vmovaps %ymm4,%ymm3 |
| .byte 197,252,40,229 // vmovaps %ymm5,%ymm4 |
| .byte 197,252,40,238 // vmovaps %ymm6,%ymm5 |
| .byte 197,252,40,247 // vmovaps %ymm7,%ymm6 |
| .byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hsl_to_rgb_hsw |
| .globl _sk_hsl_to_rgb_hsw |
| FUNCTION(_sk_hsl_to_rgb_hsw) |
| _sk_hsl_to_rgb_hsw: |
| .byte 72,131,236,56 // sub $0x38,%rsp |
| .byte 197,252,17,60,36 // vmovups %ymm7,(%rsp) |
| .byte 197,252,17,116,36,224 // vmovups %ymm6,-0x20(%rsp) |
| .byte 197,252,17,108,36,192 // vmovups %ymm5,-0x40(%rsp) |
| .byte 197,252,17,100,36,160 // vmovups %ymm4,-0x60(%rsp) |
| .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) |
| .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 |
| .byte 196,193,108,194,200,1 // vcmpltps %ymm8,%ymm2,%ymm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,224 // vmovd %eax,%xmm4 |
| .byte 196,98,125,88,212 // vpbroadcastd %xmm4,%ymm10 |
| .byte 197,172,88,229 // vaddps %ymm5,%ymm10,%ymm4 |
| .byte 197,220,89,226 // vmulps %ymm2,%ymm4,%ymm4 |
| .byte 197,84,88,202 // vaddps %ymm2,%ymm5,%ymm9 |
| .byte 196,98,85,188,202 // vfnmadd231ps %ymm2,%ymm5,%ymm9 |
| .byte 196,99,53,74,204,16 // vblendvps %ymm1,%ymm4,%ymm9,%ymm9 |
| .byte 65,184,0,0,0,64 // mov $0x40000000,%r8d |
| .byte 184,171,170,170,62 // mov $0x3eaaaaab,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 |
| .byte 197,148,88,224 // vaddps %ymm0,%ymm13,%ymm4 |
| .byte 184,0,0,0,0 // mov $0x0,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12 |
| .byte 197,172,194,204,1 // vcmpltps %ymm4,%ymm10,%ymm1 |
| .byte 196,65,92,92,218 // vsubps %ymm10,%ymm4,%ymm11 |
| .byte 196,195,93,74,203,16 // vblendvps %ymm1,%ymm11,%ymm4,%ymm1 |
| .byte 196,65,92,194,220,1 // vcmpltps %ymm12,%ymm4,%ymm11 |
| .byte 197,44,88,244 // vaddps %ymm4,%ymm10,%ymm14 |
| .byte 196,195,117,74,206,176 // vblendvps %ymm11,%ymm14,%ymm1,%ymm1 |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 |
| .byte 196,66,109,170,217 // vfmsub213ps %ymm9,%ymm2,%ymm11 |
| .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d |
| .byte 184,0,0,192,64 // mov $0x40c00000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,65,52,92,243 // vsubps %ymm11,%ymm9,%ymm14 |
| .byte 197,12,89,243 // vmulps %ymm3,%ymm14,%ymm14 |
| .byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,98,125,88,251 // vpbroadcastd %xmm3,%ymm15 |
| .byte 197,132,92,217 // vsubps %ymm1,%ymm15,%ymm3 |
| .byte 196,194,13,168,219 // vfmadd213ps %ymm11,%ymm14,%ymm3 |
| .byte 196,193,116,194,255,1 // vcmpltps %ymm15,%ymm1,%ymm7 |
| .byte 196,227,37,74,219,112 // vblendvps %ymm7,%ymm3,%ymm11,%ymm3 |
| .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 |
| .byte 196,195,101,74,249,112 // vblendvps %ymm7,%ymm9,%ymm3,%ymm7 |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,244,194,203,1 // vcmpltps %ymm3,%ymm1,%ymm1 |
| .byte 196,194,13,168,227 // vfmadd213ps %ymm11,%ymm14,%ymm4 |
| .byte 196,227,69,74,228,16 // vblendvps %ymm1,%ymm4,%ymm7,%ymm4 |
| .byte 197,172,194,200,1 // vcmpltps %ymm0,%ymm10,%ymm1 |
| .byte 196,193,124,92,250 // vsubps %ymm10,%ymm0,%ymm7 |
| .byte 196,227,125,74,207,16 // vblendvps %ymm1,%ymm7,%ymm0,%ymm1 |
| .byte 196,193,124,194,252,1 // vcmpltps %ymm12,%ymm0,%ymm7 |
| .byte 197,172,88,240 // vaddps %ymm0,%ymm10,%ymm6 |
| .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1 |
| .byte 197,132,92,241 // vsubps %ymm1,%ymm15,%ymm6 |
| .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6 |
| .byte 196,193,116,194,255,1 // vcmpltps %ymm15,%ymm1,%ymm7 |
| .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6 |
| .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 |
| .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6 |
| .byte 197,244,194,203,1 // vcmpltps %ymm3,%ymm1,%ymm1 |
| .byte 196,193,124,92,253 // vsubps %ymm13,%ymm0,%ymm7 |
| .byte 196,194,13,168,195 // vfmadd213ps %ymm11,%ymm14,%ymm0 |
| .byte 196,227,77,74,200,16 // vblendvps %ymm1,%ymm0,%ymm6,%ymm1 |
| .byte 197,172,194,199,1 // vcmpltps %ymm7,%ymm10,%ymm0 |
| .byte 196,193,68,92,242 // vsubps %ymm10,%ymm7,%ymm6 |
| .byte 196,227,69,74,198,0 // vblendvps %ymm0,%ymm6,%ymm7,%ymm0 |
| .byte 196,193,68,194,244,1 // vcmpltps %ymm12,%ymm7,%ymm6 |
| .byte 197,44,88,215 // vaddps %ymm7,%ymm10,%ymm10 |
| .byte 196,195,125,74,194,96 // vblendvps %ymm6,%ymm10,%ymm0,%ymm0 |
| .byte 196,194,13,168,251 // vfmadd213ps %ymm11,%ymm14,%ymm7 |
| .byte 197,132,92,240 // vsubps %ymm0,%ymm15,%ymm6 |
| .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6 |
| .byte 196,65,124,194,215,1 // vcmpltps %ymm15,%ymm0,%ymm10 |
| .byte 196,227,37,74,246,160 // vblendvps %ymm10,%ymm6,%ymm11,%ymm6 |
| .byte 196,65,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm8 |
| .byte 196,195,77,74,241,128 // vblendvps %ymm8,%ymm9,%ymm6,%ymm6 |
| .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0 |
| .byte 196,227,77,74,223,0 // vblendvps %ymm0,%ymm7,%ymm6,%ymm3 |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,212,194,232,0 // vcmpeqps %ymm0,%ymm5,%ymm5 |
| .byte 196,227,93,74,194,80 // vblendvps %ymm5,%ymm2,%ymm4,%ymm0 |
| .byte 196,227,117,74,202,80 // vblendvps %ymm5,%ymm2,%ymm1,%ymm1 |
| .byte 196,227,101,74,210,80 // vblendvps %ymm5,%ymm2,%ymm3,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,16,92,36,128 // vmovups -0x80(%rsp),%ymm3 |
| .byte 197,252,16,100,36,160 // vmovups -0x60(%rsp),%ymm4 |
| .byte 197,252,16,108,36,192 // vmovups -0x40(%rsp),%ymm5 |
| .byte 197,252,16,116,36,224 // vmovups -0x20(%rsp),%ymm6 |
| .byte 197,252,16,60,36 // vmovups (%rsp),%ymm7 |
| .byte 72,131,196,56 // add $0x38,%rsp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_1_float_hsw |
| .globl _sk_scale_1_float_hsw |
| FUNCTION(_sk_scale_1_float_hsw) |
| _sk_scale_1_float_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_u8_hsw |
| .globl _sk_scale_u8_hsw |
| FUNCTION(_sk_scale_u8_hsw) |
| _sk_scale_u8_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 117,56 // jne 11bc <_sk_scale_u8_hsw+0x48> |
| .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,193 // mov %r8,%rcx |
| .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 11c4 <_sk_scale_u8_hsw+0x50> |
| .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| .byte 235,167 // jmp 1188 <_sk_scale_u8_hsw+0x14> |
| |
| HIDDEN _sk_lerp_1_float_hsw |
| .globl _sk_lerp_1_float_hsw |
| FUNCTION(_sk_lerp_1_float_hsw) |
| _sk_lerp_1_float_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 |
| .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 |
| .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2 |
| .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_u8_hsw |
| .globl _sk_lerp_u8_hsw |
| FUNCTION(_sk_lerp_u8_hsw) |
| _sk_lerp_u8_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 117,76 // jne 126c <_sk_lerp_u8_hsw+0x5c> |
| .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 |
| .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 |
| .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2 |
| .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,193 // mov %r8,%rcx |
| .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 1274 <_sk_lerp_u8_hsw+0x64> |
| .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| .byte 235,147 // jmp 1224 <_sk_lerp_u8_hsw+0x14> |
| |
| HIDDEN _sk_lerp_565_hsw |
| .globl _sk_lerp_565_hsw |
| FUNCTION(_sk_lerp_565_hsw) |
| _sk_lerp_565_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,16 // mov (%rax),%r10 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,179,0,0,0 // jne 1352 <_sk_lerp_565_hsw+0xc1> |
| .byte 196,193,122,111,28,122 // vmovdqu (%r10,%rdi,2),%xmm3 |
| .byte 196,98,125,51,195 // vpmovzxwd %xmm3,%ymm8 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 |
| .byte 197,124,91,203 // vcvtdq2ps %ymm3,%ymm9 |
| .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 |
| .byte 197,124,91,211 // vcvtdq2ps %ymm3,%ymm10 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 |
| .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%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 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| .byte 196,226,53,168,196 // vfmadd213ps %ymm4,%ymm9,%ymm0 |
| .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| .byte 196,226,45,168,205 // vfmadd213ps %ymm5,%ymm10,%ymm1 |
| .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| .byte 196,226,101,168,214 // vfmadd213ps %ymm6,%ymm3,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%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 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| .byte 65,254,200 // dec %r8b |
| .byte 65,128,248,6 // cmp $0x6,%r8b |
| .byte 15,135,59,255,255,255 // ja 12a5 <_sk_lerp_565_hsw+0x14> |
| .byte 69,15,182,192 // movzbl %r8b,%r8d |
| .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 13c0 <_sk_lerp_565_hsw+0x12f> |
| .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| .byte 76,1,200 // add %r9,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| .byte 196,193,97,196,92,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 196,193,97,196,92,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 196,193,97,196,92,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 196,193,97,196,92,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 196,193,97,196,92,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 196,193,97,196,92,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 196,193,97,196,28,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3 |
| .byte 233,231,254,255,255 // jmpq 12a5 <_sk_lerp_565_hsw+0x14> |
| .byte 102,144 // xchg %ax,%ax |
| .byte 242,255 // repnz (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 234 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,226 // jmpq *%rdx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 218,255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,210 // callq *%rdx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,202 // dec %edx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 190 // .byte 0xbe |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_load_tables_hsw |
| .globl _sk_load_tables_hsw |
| FUNCTION(_sk_load_tables_hsw) |
| _sk_load_tables_hsw: |
| .byte 73,137,200 // mov %rcx,%r8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .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 117,121 // jne 146a <_sk_load_tables_hsw+0x8e> |
| .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 197,249,110,193 // vmovd %ecx,%xmm0 |
| .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2 |
| .byte 197,237,219,203 // vpand %ymm3,%ymm2,%ymm1 |
| .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| .byte 72,139,72,8 // mov 0x8(%rax),%rcx |
| .byte 76,139,72,16 // mov 0x10(%rax),%r9 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,226,53,146,4,137 // vgatherdps %ymm9,(%rcx,%ymm1,4),%ymm0 |
| .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1 |
| .byte 197,109,219,201 // vpand %ymm1,%ymm2,%ymm9 |
| .byte 196,65,45,118,210 // vpcmpeqd %ymm10,%ymm10,%ymm10 |
| .byte 196,130,45,146,12,137 // vgatherdps %ymm10,(%r9,%ymm9,4),%ymm1 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9 |
| .byte 196,65,109,219,201 // vpand %ymm9,%ymm2,%ymm9 |
| .byte 196,162,61,146,20,136 // vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2 |
| .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 |
| .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%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 76,137,193 // mov %r8,%rcx |
| .byte 255,224 // jmpq *%rax |
| .byte 185,8,0,0,0 // mov $0x8,%ecx |
| .byte 68,41,193 // sub %r8d,%ecx |
| .byte 192,225,3 // shl $0x3,%cl |
| .byte 73,199,194,255,255,255,255 // mov $0xffffffffffffffff,%r10 |
| .byte 73,211,234 // shr %cl,%r10 |
| .byte 196,193,249,110,194 // vmovq %r10,%xmm0 |
| .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 |
| .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 |
| .byte 233,99,255,255,255 // jmpq 13f6 <_sk_load_tables_hsw+0x1a> |
| |
| HIDDEN _sk_load_tables_u16_be_hsw |
| .globl _sk_load_tables_u16_be_hsw |
| FUNCTION(_sk_load_tables_u16_be_hsw) |
| _sk_load_tables_u16_be_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,216,0,0,0 // jne 1581 <_sk_load_tables_u16_be_hsw+0xee> |
| .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8 |
| .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2 |
| .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3 |
| .byte 196,1,122,111,76,72,48 // vmovdqu 0x30(%r8,%r9,2),%xmm9 |
| .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,113,105,235 // vpunpckhwd %xmm3,%xmm1,%xmm13 |
| .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 |
| .byte 65,184,255,0,0,0 // mov $0xff,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,209 // vpbroadcastd %xmm1,%ymm10 |
| .byte 196,193,125,219,202 // vpand %ymm10,%ymm0,%ymm1 |
| .byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11 |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 76,139,72,16 // mov 0x10(%rax),%r9 |
| .byte 196,65,29,118,228 // vpcmpeqd %ymm12,%ymm12,%ymm12 |
| .byte 196,194,29,146,4,136 // vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0 |
| .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1 |
| .byte 196,193,117,219,210 // vpand %ymm10,%ymm1,%ymm2 |
| .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| .byte 196,194,61,146,12,145 // vgatherdps %ymm8,(%r9,%ymm2,4),%ymm1 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 196,193,49,108,213 // vpunpcklqdq %xmm13,%xmm9,%xmm2 |
| .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2 |
| .byte 196,65,109,219,194 // vpand %ymm10,%ymm2,%ymm8 |
| .byte 196,162,37,146,20,128 // vgatherdps %ymm11,(%rax,%ymm8,4),%ymm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,98,125,88,195 // vpbroadcastd %xmm3,%ymm8 |
| .byte 196,193,49,109,221 // vpunpckhqdq %xmm13,%xmm9,%xmm3 |
| .byte 197,177,113,243,8 // vpsllw $0x8,%xmm3,%xmm9 |
| .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3 |
| .byte 197,177,235,219 // vpor %xmm3,%xmm9,%xmm3 |
| .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8 |
| .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 116,85 // je 15e7 <_sk_load_tables_u16_be_hsw+0x154> |
| .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,72 // jb 15e7 <_sk_load_tables_u16_be_hsw+0x154> |
| .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 116,72 // je 15f4 <_sk_load_tables_u16_be_hsw+0x161> |
| .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,59 // jb 15f4 <_sk_load_tables_u16_be_hsw+0x161> |
| .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 15,132,250,254,255,255 // je 14c4 <_sk_load_tables_u16_be_hsw+0x31> |
| .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 15,130,233,254,255,255 // jb 14c4 <_sk_load_tables_u16_be_hsw+0x31> |
| .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9 |
| .byte 233,221,254,255,255 // jmpq 14c4 <_sk_load_tables_u16_be_hsw+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| .byte 233,208,254,255,255 // jmpq 14c4 <_sk_load_tables_u16_be_hsw+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 233,199,254,255,255 // jmpq 14c4 <_sk_load_tables_u16_be_hsw+0x31> |
| |
| HIDDEN _sk_load_tables_rgb_u16_be_hsw |
| .globl _sk_load_tables_rgb_u16_be_hsw |
| FUNCTION(_sk_load_tables_rgb_u16_be_hsw) |
| _sk_load_tables_rgb_u16_be_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,207,0,0,0 // jne 16de <_sk_load_tables_rgb_u16_be_hsw+0xe1> |
| .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0 |
| .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2 |
| .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1 |
| .byte 196,129,122,111,92,72,32 // vmovdqu 0x20(%r8,%r9,2),%xmm3 |
| .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3 |
| .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8 |
| .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9 |
| .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11 |
| .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10 |
| .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0 |
| .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2 |
| .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1 |
| .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3 |
| .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3 |
| .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 |
| .byte 65,184,255,0,0,0 // mov $0xff,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,209 // vpbroadcastd %xmm1,%ymm10 |
| .byte 196,193,125,219,202 // vpand %ymm10,%ymm0,%ymm1 |
| .byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11 |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 76,139,72,16 // mov 0x10(%rax),%r9 |
| .byte 196,65,29,118,228 // vpcmpeqd %ymm12,%ymm12,%ymm12 |
| .byte 196,194,29,146,4,136 // vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0 |
| .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1 |
| .byte 196,193,117,219,210 // vpand %ymm10,%ymm1,%ymm2 |
| .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| .byte 196,194,61,146,12,145 // vgatherdps %ymm8,(%r9,%ymm2,4),%ymm1 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2 |
| .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2 |
| .byte 196,193,109,219,218 // vpand %ymm10,%ymm2,%ymm3 |
| .byte 196,226,37,146,20,152 // vgatherdps %ymm11,(%rax,%ymm3,4),%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0 |
| .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 117,5 // jne 16f7 <_sk_load_tables_rgb_u16_be_hsw+0xfa> |
| .byte 233,76,255,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1 |
| .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,26 // jb 1726 <_sk_load_tables_rgb_u16_be_hsw+0x129> |
| .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1 |
| .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 117,10 // jne 172b <_sk_load_tables_rgb_u16_be_hsw+0x12e> |
| .byte 233,29,255,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| .byte 233,24,255,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1 |
| .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,26 // jb 175a <_sk_load_tables_rgb_u16_be_hsw+0x15d> |
| .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1 |
| .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 117,10 // jne 175f <_sk_load_tables_rgb_u16_be_hsw+0x162> |
| .byte 233,233,254,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| .byte 233,228,254,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3 |
| .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,20 // jb 1788 <_sk_load_tables_rgb_u16_be_hsw+0x18b> |
| .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3 |
| .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 |
| .byte 233,187,254,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| .byte 233,182,254,255,255 // jmpq 1643 <_sk_load_tables_rgb_u16_be_hsw+0x46> |
| |
| HIDDEN _sk_byte_tables_hsw |
| .globl _sk_byte_tables_hsw |
| FUNCTION(_sk_byte_tables_hsw) |
| _sk_byte_tables_hsw: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,127,67 // mov $0x437f0000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 68,137,197 // mov %r8d,%ebp |
| .byte 77,137,194 // mov %r8,%r10 |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,192 // vmovq %xmm0,%r8 |
| .byte 69,137,195 // mov %r8d,%r11d |
| .byte 77,137,199 // mov %r8,%r15 |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 69,137,198 // mov %r8d,%r14d |
| .byte 77,137,196 // mov %r8,%r12 |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 |
| .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp |
| .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp |
| .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,41 // movzbl (%r9,%r13,1),%ebp |
| .byte 196,227,121,32,197,4 // vpinsrb $0x4,%ebp,%xmm0,%xmm0 |
| .byte 65,15,182,44,25 // movzbl (%r9,%rbx,1),%ebp |
| .byte 196,227,121,32,197,5 // vpinsrb $0x5,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp |
| .byte 196,227,121,32,197,6 // vpinsrb $0x6,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp |
| .byte 196,227,121,32,197,7 // vpinsrb $0x7,%ebp,%xmm0,%xmm0 |
| .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 |
| .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10 |
| .byte 189,129,128,128,59 // mov $0x3b808081,%ebp |
| .byte 197,249,110,197 // vmovd %ebp,%xmm0 |
| .byte 196,98,125,88,200 // vpbroadcastd %xmm0,%ymm9 |
| .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp |
| .byte 65,137,233 // mov %ebp,%r9d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,203 // vmovq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1 |
| .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,193,249,126,207 // vmovq %xmm1,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 |
| .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1 |
| .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp |
| .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,32 // movzbl (%r8,%r12,1),%ebp |
| .byte 196,227,113,32,205,4 // vpinsrb $0x4,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,56 // movzbl (%r8,%r15,1),%ebp |
| .byte 196,227,113,32,205,5 // vpinsrb $0x5,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp |
| .byte 196,227,113,32,205,6 // vpinsrb $0x6,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp |
| .byte 196,227,113,32,205,7 // vpinsrb $0x7,%ebp,%xmm1,%xmm1 |
| .byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp |
| .byte 65,137,233 // mov %ebp,%r9d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,211 // vmovq %xmm2,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,57,210,1 // vextracti128 $0x1,%ymm2,%xmm2 |
| .byte 196,195,249,22,211,1 // vpextrq $0x1,%xmm2,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,193,249,126,215 // vmovq %xmm2,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,20,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 |
| .byte 196,195,105,32,20,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2 |
| .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp |
| .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2 |
| .byte 67,15,182,44,32 // movzbl (%r8,%r12,1),%ebp |
| .byte 196,227,105,32,213,4 // vpinsrb $0x4,%ebp,%xmm2,%xmm2 |
| .byte 67,15,182,44,56 // movzbl (%r8,%r15,1),%ebp |
| .byte 196,227,105,32,213,5 // vpinsrb $0x5,%ebp,%xmm2,%xmm2 |
| .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp |
| .byte 196,227,105,32,213,6 // vpinsrb $0x6,%ebp,%xmm2,%xmm2 |
| .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp |
| .byte 196,227,105,32,213,7 // vpinsrb $0x7,%ebp,%xmm2,%xmm2 |
| .byte 196,226,125,49,210 // vpmovzxbd %xmm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 |
| .byte 196,227,249,22,221,1 // vpextrq $0x1,%xmm3,%rbp |
| .byte 65,137,232 // mov %ebp,%r8d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,219 // vmovq %xmm3,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,57,219,1 // vextracti128 $0x1,%ymm3,%xmm3 |
| .byte 196,195,249,22,218,1 // vpextrq $0x1,%xmm3,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,222 // vmovq %xmm3,%r14 |
| .byte 69,137,247 // mov %r14d,%r15d |
| .byte 73,193,238,32 // shr $0x20,%r14 |
| .byte 196,163,121,32,28,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3 |
| .byte 196,227,97,32,28,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3 |
| .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx |
| .byte 196,227,97,32,219,2 // vpinsrb $0x2,%ebx,%xmm3,%xmm3 |
| .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp |
| .byte 196,227,97,32,221,3 // vpinsrb $0x3,%ebp,%xmm3,%xmm3 |
| .byte 66,15,182,44,56 // movzbl (%rax,%r15,1),%ebp |
| .byte 196,227,97,32,221,4 // vpinsrb $0x4,%ebp,%xmm3,%xmm3 |
| .byte 66,15,182,44,48 // movzbl (%rax,%r14,1),%ebp |
| .byte 196,227,97,32,221,5 // vpinsrb $0x5,%ebp,%xmm3,%xmm3 |
| .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp |
| .byte 196,227,97,32,221,6 // vpinsrb $0x6,%ebp,%xmm3,%xmm3 |
| .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax |
| .byte 196,227,97,32,216,7 // vpinsrb $0x7,%eax,%xmm3,%xmm3 |
| .byte 196,226,125,49,219 // vpmovzxbd %xmm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_byte_tables_rgb_hsw |
| .globl _sk_byte_tables_rgb_hsw |
| FUNCTION(_sk_byte_tables_rgb_hsw) |
| _sk_byte_tables_rgb_hsw: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 68,139,64,24 // mov 0x18(%rax),%r8d |
| .byte 65,255,200 // dec %r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 68,137,197 // mov %r8d,%ebp |
| .byte 77,137,194 // mov %r8,%r10 |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,192 // vmovq %xmm0,%r8 |
| .byte 69,137,195 // mov %r8d,%r11d |
| .byte 77,137,199 // mov %r8,%r15 |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 69,137,198 // mov %r8d,%r14d |
| .byte 77,137,196 // mov %r8,%r12 |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 |
| .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp |
| .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp |
| .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,41 // movzbl (%r9,%r13,1),%ebp |
| .byte 196,227,121,32,197,4 // vpinsrb $0x4,%ebp,%xmm0,%xmm0 |
| .byte 65,15,182,44,25 // movzbl (%r9,%rbx,1),%ebp |
| .byte 196,227,121,32,197,5 // vpinsrb $0x5,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp |
| .byte 196,227,121,32,197,6 // vpinsrb $0x6,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp |
| .byte 196,227,121,32,197,7 // vpinsrb $0x7,%ebp,%xmm0,%xmm0 |
| .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 |
| .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10 |
| .byte 189,129,128,128,59 // mov $0x3b808081,%ebp |
| .byte 197,249,110,197 // vmovd %ebp,%xmm0 |
| .byte 196,98,125,88,200 // vpbroadcastd %xmm0,%ymm9 |
| .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp |
| .byte 65,137,233 // mov %ebp,%r9d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,203 // vmovq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1 |
| .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,193,249,126,207 // vmovq %xmm1,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 |
| .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1 |
| .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp |
| .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,32 // movzbl (%r8,%r12,1),%ebp |
| .byte 196,227,113,32,205,4 // vpinsrb $0x4,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,56 // movzbl (%r8,%r15,1),%ebp |
| .byte 196,227,113,32,205,5 // vpinsrb $0x5,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp |
| .byte 196,227,113,32,205,6 // vpinsrb $0x6,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp |
| .byte 196,227,113,32,205,7 // vpinsrb $0x7,%ebp,%xmm1,%xmm1 |
| .byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 72,139,64,16 // mov 0x10(%rax),%rax |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp |
| .byte 65,137,232 // mov %ebp,%r8d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,211 // vmovq %xmm2,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,57,210,1 // vextracti128 $0x1,%ymm2,%xmm2 |
| .byte 196,195,249,22,210,1 // vpextrq $0x1,%xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,214 // vmovq %xmm2,%r14 |
| .byte 69,137,247 // mov %r14d,%r15d |
| .byte 73,193,238,32 // shr $0x20,%r14 |
| .byte 196,163,121,32,20,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2 |
| .byte 196,227,105,32,20,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2 |
| .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx |
| .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2 |
| .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp |
| .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2 |
| .byte 66,15,182,44,56 // movzbl (%rax,%r15,1),%ebp |
| .byte 196,227,105,32,213,4 // vpinsrb $0x4,%ebp,%xmm2,%xmm2 |
| .byte 66,15,182,44,48 // movzbl (%rax,%r14,1),%ebp |
| .byte 196,227,105,32,213,5 // vpinsrb $0x5,%ebp,%xmm2,%xmm2 |
| .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp |
| .byte 196,227,105,32,213,6 // vpinsrb $0x6,%ebp,%xmm2,%xmm2 |
| .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax |
| .byte 196,227,105,32,208,7 // vpinsrb $0x7,%eax,%xmm2,%xmm2 |
| .byte 196,226,125,49,210 // vpmovzxbd %xmm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_r_hsw |
| .globl _sk_table_r_hsw |
| FUNCTION(_sk_table_r_hsw) |
| _sk_table_r_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,125,91,192 // vcvtps2dq %ymm0,%ymm8 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,130,53,146,4,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_g_hsw |
| .globl _sk_table_g_hsw |
| FUNCTION(_sk_table_g_hsw) |
| _sk_table_g_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,125,91,193 // vcvtps2dq %ymm1,%ymm8 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,130,53,146,12,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_b_hsw |
| .globl _sk_table_b_hsw |
| FUNCTION(_sk_table_b_hsw) |
| _sk_table_b_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,125,91,194 // vcvtps2dq %ymm2,%ymm8 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,130,53,146,20,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_a_hsw |
| .globl _sk_table_a_hsw |
| FUNCTION(_sk_table_a_hsw) |
| _sk_table_a_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 197,125,91,195 // vcvtps2dq %ymm3,%ymm8 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,130,53,146,28,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_r_hsw |
| .globl _sk_parametric_r_hsw |
| FUNCTION(_sk_parametric_r_hsw) |
| _sk_parametric_r_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,124,194,192,2 // vcmpleps %ymm8,%ymm0,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,66,125,168,202 // vfmadd213ps %ymm10,%ymm0,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 196,66,125,168,211 // vfmadd213ps %ymm11,%ymm0,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 196,65,125,219,210 // vpand %ymm10,%ymm0,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,45,235,208 // vpor %ymm0,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,240 // vpbroadcastd %xmm0,%ymm14 |
| .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,216 // vpbroadcastd %xmm0,%ymm11 |
| .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0 |
| .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 |
| .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 |
| .byte 197,28,89,216 // vmulps %ymm0,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,224 // vpbroadcastd %xmm0,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 196,65,124,88,219 // vaddps %ymm11,%ymm0,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,232 // vpbroadcastd %xmm0,%ymm13 |
| .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,216 // vpbroadcastd %xmm0,%ymm11 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 |
| .byte 197,164,94,192 // vdivps %ymm0,%ymm11,%ymm0 |
| .byte 197,148,88,192 // vaddps %ymm0,%ymm13,%ymm0 |
| .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 |
| .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,124,95,192 // vmaxps %ymm8,%ymm0,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_g_hsw |
| .globl _sk_parametric_g_hsw |
| FUNCTION(_sk_parametric_g_hsw) |
| _sk_parametric_g_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,116,194,192,2 // vcmpleps %ymm8,%ymm1,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,66,117,168,202 // vfmadd213ps %ymm10,%ymm1,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 196,66,117,168,211 // vfmadd213ps %ymm11,%ymm1,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 196,65,117,219,210 // vpand %ymm10,%ymm1,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,45,235,209 // vpor %ymm1,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,241 // vpbroadcastd %xmm1,%ymm14 |
| .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,217 // vpbroadcastd %xmm1,%ymm11 |
| .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 |
| .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 |
| .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 |
| .byte 197,28,89,217 // vmulps %ymm1,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13 |
| .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,217 // vpbroadcastd %xmm1,%ymm11 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 |
| .byte 197,164,94,201 // vdivps %ymm1,%ymm11,%ymm1 |
| .byte 197,148,88,201 // vaddps %ymm1,%ymm13,%ymm1 |
| .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 |
| .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_b_hsw |
| .globl _sk_parametric_b_hsw |
| FUNCTION(_sk_parametric_b_hsw) |
| _sk_parametric_b_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,108,194,192,2 // vcmpleps %ymm8,%ymm2,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,66,109,168,202 // vfmadd213ps %ymm10,%ymm2,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 196,66,109,168,211 // vfmadd213ps %ymm11,%ymm2,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 196,65,109,219,210 // vpand %ymm10,%ymm2,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 197,45,235,210 // vpor %ymm2,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,242 // vpbroadcastd %xmm2,%ymm14 |
| .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,218 // vpbroadcastd %xmm2,%ymm11 |
| .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 |
| .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 |
| .byte 197,28,89,218 // vmulps %ymm2,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,226 // vpbroadcastd %xmm2,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 196,65,108,88,219 // vaddps %ymm11,%ymm2,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,234 // vpbroadcastd %xmm2,%ymm13 |
| .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,98,125,88,218 // vpbroadcastd %xmm2,%ymm11 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2 |
| .byte 197,164,94,210 // vdivps %ymm2,%ymm11,%ymm2 |
| .byte 197,148,88,210 // vaddps %ymm2,%ymm13,%ymm2 |
| .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,108,95,192 // vmaxps %ymm8,%ymm2,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| .byte 197,188,93,210 // vminps %ymm2,%ymm8,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_a_hsw |
| .globl _sk_parametric_a_hsw |
| FUNCTION(_sk_parametric_a_hsw) |
| _sk_parametric_a_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,100,194,192,2 // vcmpleps %ymm8,%ymm3,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,66,101,168,202 // vfmadd213ps %ymm10,%ymm3,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 196,66,101,168,211 // vfmadd213ps %ymm11,%ymm3,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,65,101,219,210 // vpand %ymm10,%ymm3,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,45,235,211 // vpor %ymm3,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,243 // vpbroadcastd %xmm3,%ymm14 |
| .byte 196,66,37,186,245 // vfmsub231ps %ymm13,%ymm11,%ymm14 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 |
| .byte 196,66,45,172,222 // vfnmadd213ps %ymm14,%ymm10,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,172,88,219 // vaddps %ymm3,%ymm10,%ymm3 |
| .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 |
| .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 |
| .byte 197,28,89,219 // vmulps %ymm3,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,227 // vpbroadcastd %xmm3,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,235 // vpbroadcastd %xmm3,%ymm13 |
| .byte 196,66,45,172,235 // vfnmadd213ps %ymm11,%ymm10,%ymm13 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3 |
| .byte 197,164,94,219 // vdivps %ymm3,%ymm11,%ymm3 |
| .byte 197,148,88,219 // vaddps %ymm3,%ymm13,%ymm3 |
| .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 |
| .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 |
| .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,100,95,192 // vmaxps %ymm8,%ymm3,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,188,93,219 // vminps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_a8_hsw |
| .globl _sk_load_a8_hsw |
| FUNCTION(_sk_load_a8_hsw) |
| _sk_load_a8_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 117,50 // jne 2358 <_sk_load_a8_hsw+0x42> |
| .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 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 76,137,193 // mov %r8,%rcx |
| .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 2360 <_sk_load_a8_hsw+0x4a> |
| .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| .byte 235,173 // jmp 232a <_sk_load_a8_hsw+0x14> |
| |
| HIDDEN _sk_gather_a8_hsw |
| .globl _sk_gather_a8_hsw |
| FUNCTION(_sk_gather_a8_hsw) |
| _sk_gather_a8_hsw: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,194 // vmovq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,199 // vmovq %xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,4,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0 |
| .byte 71,15,182,12,8 // movzbl (%r8,%r9,1),%r9d |
| .byte 196,195,121,32,193,2 // vpinsrb $0x2,%r9d,%xmm0,%xmm0 |
| .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax |
| .byte 196,227,121,32,192,3 // vpinsrb $0x3,%eax,%xmm0,%xmm0 |
| .byte 67,15,182,4,32 // movzbl (%r8,%r12,1),%eax |
| .byte 196,227,121,32,192,4 // vpinsrb $0x4,%eax,%xmm0,%xmm0 |
| .byte 67,15,182,4,56 // movzbl (%r8,%r15,1),%eax |
| .byte 196,227,121,32,192,5 // vpinsrb $0x5,%eax,%xmm0,%xmm0 |
| .byte 67,15,182,4,48 // movzbl (%r8,%r14,1),%eax |
| .byte 196,227,121,32,192,6 // vpinsrb $0x6,%eax,%xmm0,%xmm0 |
| .byte 65,15,182,4,24 // movzbl (%r8,%rbx,1),%eax |
| .byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%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 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,237,239,210 // vpxor %ymm2,%ymm2,%ymm2 |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_a8_hsw |
| .globl _sk_store_a8_hsw |
| FUNCTION(_sk_store_a8_hsw) |
| _sk_store_a8_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 184,0,0,127,67 // mov $0x437f0000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 117,10 // jne 2495 <_sk_store_a8_hsw+0x3b> |
| .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) |
| .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 2491 <_sk_store_a8_hsw+0x37> |
| .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 24f8 <_sk_store_a8_hsw+0x9e> |
| .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| .byte 76,1,192 // add %r8,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,67,121,20,68,57,6,12 // vpextrb $0xc,%xmm8,0x6(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,5,10 // vpextrb $0xa,%xmm8,0x5(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,4,8 // vpextrb $0x8,%xmm8,0x4(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,3,6 // vpextrb $0x6,%xmm8,0x3(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) |
| .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) |
| .byte 235,154 // jmp 2491 <_sk_store_a8_hsw+0x37> |
| .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_g8_hsw |
| .globl _sk_load_g8_hsw |
| FUNCTION(_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 2560 <_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 2568 <_sk_load_g8_hsw+0x54> |
| .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| .byte 235,163 // jmp 2528 <_sk_load_g8_hsw+0x14> |
| |
| HIDDEN _sk_gather_g8_hsw |
| .globl _sk_gather_g8_hsw |
| FUNCTION(_sk_gather_g8_hsw) |
| _sk_gather_g8_hsw: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,194 // vmovq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,199 // vmovq %xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,4,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0 |
| .byte 71,15,182,12,8 // movzbl (%r8,%r9,1),%r9d |
| .byte 196,195,121,32,193,2 // vpinsrb $0x2,%r9d,%xmm0,%xmm0 |
| .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax |
| .byte 196,227,121,32,192,3 // vpinsrb $0x3,%eax,%xmm0,%xmm0 |
| .byte 67,15,182,4,32 // movzbl (%r8,%r12,1),%eax |
| .byte 196,227,121,32,192,4 // vpinsrb $0x4,%eax,%xmm0,%xmm0 |
| .byte 67,15,182,4,56 // movzbl (%r8,%r15,1),%eax |
| .byte 196,227,121,32,192,5 // vpinsrb $0x5,%eax,%xmm0,%xmm0 |
| .byte 67,15,182,4,48 // movzbl (%r8,%r14,1),%eax |
| .byte 196,227,121,32,192,6 // vpinsrb $0x6,%eax,%xmm0,%xmm0 |
| .byte 65,15,182,4,24 // movzbl (%r8,%rbx,1),%eax |
| .byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%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 197,252,40,200 // vmovaps %ymm0,%ymm1 |
| .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_i8_hsw |
| .globl _sk_gather_i8_hsw |
| FUNCTION(_sk_gather_i8_hsw) |
| _sk_gather_i8_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,192 // mov %rax,%r8 |
| .byte 77,133,192 // test %r8,%r8 |
| .byte 116,5 // je 267b <_sk_gather_i8_hsw+0xf> |
| .byte 76,137,192 // mov %r8,%rax |
| .byte 235,2 // jmp 267d <_sk_gather_i8_hsw+0x11> |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,194 // mov %eax,%r10d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,195 // vmovq %xmm0,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,223 // mov %ebx,%r15d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,196 // vmovq %xmm0,%r12 |
| .byte 69,137,229 // mov %r12d,%r13d |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 196,131,121,32,4,49,0 // vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,25,1 // vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,17,2 // vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,1,3 // vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,41,4 // vpinsrb $0x4,(%r9,%r13,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,33,5 // vpinsrb $0x5,(%r9,%r12,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,57,6 // vpinsrb $0x6,(%r9,%r15,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,25,7 // vpinsrb $0x7,(%r9,%rbx,1),%xmm0,%xmm0 |
| .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 |
| .byte 73,139,64,8 // mov 0x8(%r8),%rax |
| .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 |
| .byte 196,226,117,144,28,128 // vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2 |
| .byte 197,237,219,195 // vpand %ymm3,%ymm2,%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 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1 |
| .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9 |
| .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_565_hsw |
| .globl _sk_load_565_hsw |
| FUNCTION(_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 282f <_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 |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,253,219,194 // vpand %ymm2,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,245,219,202 // vpand %ymm2,%ymm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%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 15,135,89,255,255,255 // ja 27a0 <_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 # 289c <_sk_load_565_hsw+0x110> |
| .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 233,5,255,255,255 // jmpq 27a0 <_sk_load_565_hsw+0x14> |
| .byte 144 // nop |
| .byte 243,255 // repz (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 235,255 // jmp 28a1 <_sk_load_565_hsw+0x115> |
| .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_gather_565_hsw |
| .globl _sk_gather_565_hsw |
| FUNCTION(_sk_gather_565_hsw) |
| _sk_gather_565_hsw: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,194 // vmovq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,199 // vmovq %xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 71,15,183,20,80 // movzwl (%r8,%r10,2),%r10d |
| .byte 71,15,183,28,88 // movzwl (%r8,%r11,2),%r11d |
| .byte 196,193,121,110,195 // vmovd %r11d,%xmm0 |
| .byte 196,193,121,196,194,1 // vpinsrw $0x1,%r10d,%xmm0,%xmm0 |
| .byte 71,15,183,12,72 // movzwl (%r8,%r9,2),%r9d |
| .byte 196,193,121,196,193,2 // vpinsrw $0x2,%r9d,%xmm0,%xmm0 |
| .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax |
| .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0 |
| .byte 67,15,183,4,96 // movzwl (%r8,%r12,2),%eax |
| .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0 |
| .byte 67,15,183,4,120 // movzwl (%r8,%r15,2),%eax |
| .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0 |
| .byte 67,15,183,4,112 // movzwl (%r8,%r14,2),%eax |
| .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0 |
| .byte 65,15,183,4,88 // movzwl (%r8,%rbx,2),%eax |
| .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0 |
| .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,253,219,194 // vpand %ymm2,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,245,219,202 // vpand %ymm2,%ymm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_565_hsw |
| .globl _sk_store_565_hsw |
| FUNCTION(_sk_store_565_hsw) |
| _sk_store_565_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 184,0,0,248,65 // mov $0x41f80000,%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 196,193,53,114,241,11 // vpslld $0xb,%ymm9,%ymm9 |
| .byte 184,0,0,124,66 // mov $0x427c0000,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| .byte 196,193,45,114,242,5 // vpslld $0x5,%ymm10,%ymm10 |
| .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9 |
| .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%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 2a67 <_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 |
| .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 2a63 <_sk_store_565_hsw+0x68> |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2ac4 <_sk_store_565_hsw+0xc9> |
| .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 2a63 <_sk_store_565_hsw+0x68> |
| .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_4444_hsw |
| .globl _sk_load_4444_hsw |
| FUNCTION(_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 2ba1 <_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 2af4 <_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 # 2c10 <_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 2af4 <_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 ffffffffe2002c18 <_sk_callback_hsw+0xffffffffe1ffeace> |
| .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_gather_4444_hsw |
| .globl _sk_gather_4444_hsw |
| FUNCTION(_sk_gather_4444_hsw) |
| _sk_gather_4444_hsw: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,194 // vmovq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,199 // vmovq %xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 71,15,183,20,80 // movzwl (%r8,%r10,2),%r10d |
| .byte 71,15,183,28,88 // movzwl (%r8,%r11,2),%r11d |
| .byte 196,193,121,110,195 // vmovd %r11d,%xmm0 |
| .byte 196,193,121,196,194,1 // vpinsrw $0x1,%r10d,%xmm0,%xmm0 |
| .byte 71,15,183,12,72 // movzwl (%r8,%r9,2),%r9d |
| .byte 196,193,121,196,193,2 // vpinsrw $0x2,%r9d,%xmm0,%xmm0 |
| .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax |
| .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0 |
| .byte 67,15,183,4,96 // movzwl (%r8,%r12,2),%eax |
| .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0 |
| .byte 67,15,183,4,120 // movzwl (%r8,%r15,2),%eax |
| .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0 |
| .byte 67,15,183,4,112 // movzwl (%r8,%r14,2),%eax |
| .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0 |
| .byte 65,15,183,4,88 // movzwl (%r8,%rbx,2),%eax |
| .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%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 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_4444_hsw |
| .globl _sk_store_4444_hsw |
| FUNCTION(_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 2dff <_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 2dfb <_sk_store_4444_hsw+0x6e> |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2e5c <_sk_store_4444_hsw+0xcf> |
| .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 2dfb <_sk_store_4444_hsw+0x6e> |
| .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_hsw |
| .globl _sk_load_8888_hsw |
| FUNCTION(_sk_load_8888_hsw) |
| _sk_load_8888_hsw: |
| .byte 73,137,200 // mov %rcx,%r8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .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 2ef5 <_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 |
| .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2 |
| .byte 197,237,219,195 // vpand %ymm3,%ymm2,%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 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1 |
| .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9 |
| .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,193 // mov %r8,%rcx |
| .byte 255,224 // jmpq *%rax |
| .byte 185,8,0,0,0 // mov $0x8,%ecx |
| .byte 68,41,193 // sub %r8d,%ecx |
| .byte 192,225,3 // shl $0x3,%cl |
| .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax |
| .byte 72,211,232 // shr %cl,%rax |
| .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 2e92 <_sk_load_8888_hsw+0x1a> |
| |
| HIDDEN _sk_gather_8888_hsw |
| .globl _sk_gather_8888_hsw |
| FUNCTION(_sk_gather_8888_hsw) |
| _sk_gather_8888_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 |
| .byte 196,194,117,144,28,128 // vpgatherdd %ymm1,(%r8,%ymm0,4),%ymm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2 |
| .byte 197,237,219,195 // vpand %ymm3,%ymm2,%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 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1 |
| .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9 |
| .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_8888_hsw |
| .globl _sk_store_8888_hsw |
| FUNCTION(_sk_store_8888_hsw) |
| _sk_store_8888_hsw: |
| .byte 73,137,200 // mov %rcx,%r8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 76,3,8 // add (%rax),%r9 |
| .byte 184,0,0,127,67 // mov $0x437f0000,%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 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 196,193,45,114,242,16 // vpslld $0x10,%ymm10,%ymm10 |
| .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| .byte 196,193,61,114,240,24 // vpslld $0x18,%ymm8,%ymm8 |
| .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 3018 <_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 |
| .byte 255,224 // jmpq *%rax |
| .byte 185,8,0,0,0 // mov $0x8,%ecx |
| .byte 68,41,193 // sub %r8d,%ecx |
| .byte 192,225,3 // shl $0x3,%cl |
| .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax |
| .byte 72,211,232 // shr %cl,%rax |
| .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 3011 <_sk_store_8888_hsw+0x6d> |
| |
| HIDDEN _sk_load_f16_hsw |
| .globl _sk_load_f16_hsw |
| FUNCTION(_sk_load_f16_hsw) |
| _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 30a9 <_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 |
| .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9 |
| .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3 |
| .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| .byte 196,226,125,19,192 // vcvtph2ps %xmm0,%ymm0 |
| .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| .byte 196,226,125,19,201 // vcvtph2ps %xmm1,%ymm1 |
| .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2 |
| .byte 196,226,125,19,210 // vcvtph2ps %xmm2,%ymm2 |
| .byte 197,177,109,219 // vpunpckhqdq %xmm3,%xmm9,%xmm3 |
| .byte 196,226,125,19,219 // vcvtph2ps %xmm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .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 3108 <_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 3108 <_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 3115 <_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 3115 <_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 305f <_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 305f <_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 305f <_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 305f <_sk_load_f16_hsw+0x21> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 233,65,255,255,255 // jmpq 305f <_sk_load_f16_hsw+0x21> |
| |
| HIDDEN _sk_gather_f16_hsw |
| .globl _sk_gather_f16_hsw |
| FUNCTION(_sk_gather_f16_hsw) |
| _sk_gather_f16_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2 |
| .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0 |
| .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 |
| .byte 197,237,118,210 // vpcmpeqd %ymm2,%ymm2,%ymm2 |
| .byte 196,194,237,144,28,192 // vpgatherdq %ymm2,(%r8,%xmm0,8),%ymm3 |
| .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0 |
| .byte 196,194,245,144,20,192 // vpgatherdq %ymm1,(%r8,%xmm0,8),%ymm2 |
| .byte 196,227,125,57,216,1 // vextracti128 $0x1,%ymm3,%xmm0 |
| .byte 196,227,125,57,209,1 // vextracti128 $0x1,%ymm2,%xmm1 |
| .byte 197,97,97,192 // vpunpcklwd %xmm0,%xmm3,%xmm8 |
| .byte 197,225,105,192 // vpunpckhwd %xmm0,%xmm3,%xmm0 |
| .byte 197,233,97,217 // vpunpcklwd %xmm1,%xmm2,%xmm3 |
| .byte 197,233,105,201 // vpunpckhwd %xmm1,%xmm2,%xmm1 |
| .byte 197,57,97,200 // vpunpcklwd %xmm0,%xmm8,%xmm9 |
| .byte 197,57,105,192 // vpunpckhwd %xmm0,%xmm8,%xmm8 |
| .byte 197,225,97,209 // vpunpcklwd %xmm1,%xmm3,%xmm2 |
| .byte 197,225,105,217 // vpunpckhwd %xmm1,%xmm3,%xmm3 |
| .byte 197,177,108,194 // vpunpcklqdq %xmm2,%xmm9,%xmm0 |
| .byte 196,226,125,19,192 // vcvtph2ps %xmm0,%ymm0 |
| .byte 197,177,109,202 // vpunpckhqdq %xmm2,%xmm9,%xmm1 |
| .byte 196,226,125,19,201 // vcvtph2ps %xmm1,%ymm1 |
| .byte 197,185,108,211 // vpunpcklqdq %xmm3,%xmm8,%xmm2 |
| .byte 196,226,125,19,210 // vcvtph2ps %xmm2,%ymm2 |
| .byte 197,185,109,219 // vpunpckhqdq %xmm3,%xmm8,%xmm3 |
| .byte 196,226,125,19,219 // vcvtph2ps %xmm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_f16_hsw |
| .globl _sk_store_f16_hsw |
| FUNCTION(_sk_store_f16_hsw) |
| _sk_store_f16_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 196,195,125,29,192,4 // vcvtps2ph $0x4,%ymm0,%xmm8 |
| .byte 196,195,125,29,201,4 // vcvtps2ph $0x4,%ymm1,%xmm9 |
| .byte 196,195,125,29,210,4 // vcvtps2ph $0x4,%ymm2,%xmm10 |
| .byte 196,195,125,29,219,4 // vcvtps2ph $0x4,%ymm3,%xmm11 |
| .byte 196,65,57,97,225 // vpunpcklwd %xmm9,%xmm8,%xmm12 |
| .byte 196,65,57,105,193 // vpunpckhwd %xmm9,%xmm8,%xmm8 |
| .byte 196,65,41,97,203 // vpunpcklwd %xmm11,%xmm10,%xmm9 |
| .byte 196,65,41,105,235 // vpunpckhwd %xmm11,%xmm10,%xmm13 |
| .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| .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 320d <_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) |
| .byte 197,122,127,68,248,48 // vmovdqu %xmm8,0x30(%rax,%rdi,8) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .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 3209 <_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 3209 <_sk_store_f16_hsw+0x61> |
| .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) |
| .byte 116,221 // je 3209 <_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 3209 <_sk_store_f16_hsw+0x61> |
| .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) |
| .byte 116,201 // je 3209 <_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 3209 <_sk_store_f16_hsw+0x61> |
| .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) |
| .byte 235,181 // jmp 3209 <_sk_store_f16_hsw+0x61> |
| |
| HIDDEN _sk_load_u16_be_hsw |
| .globl _sk_load_u16_be_hsw |
| FUNCTION(_sk_load_u16_be_hsw) |
| _sk_load_u16_be_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,205,0,0,0 // jne 3337 <_sk_load_u16_be_hsw+0xe3> |
| .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 |
| .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 |
| .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 |
| .byte 196,65,122,111,76,64,48 // vmovdqu 0x30(%r8,%rax,2),%xmm9 |
| .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,98,125,88,208 // vpbroadcastd %xmm0,%ymm10 |
| .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1 |
| .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0 |
| .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 |
| .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 |
| .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1 |
| .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1 |
| .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,193,49,108,211 // vpunpcklqdq %xmm11,%xmm9,%xmm2 |
| .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3 |
| .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2 |
| .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2 |
| .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 196,193,49,109,219 // vpunpckhqdq %xmm11,%xmm9,%xmm3 |
| .byte 197,185,113,243,8 // vpsllw $0x8,%xmm3,%xmm8 |
| .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3 |
| .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3 |
| .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 |
| .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 116,85 // je 339d <_sk_load_u16_be_hsw+0x149> |
| .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,72 // jb 339d <_sk_load_u16_be_hsw+0x149> |
| .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 116,72 // je 33aa <_sk_load_u16_be_hsw+0x156> |
| .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,59 // jb 33aa <_sk_load_u16_be_hsw+0x156> |
| .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 15,132,5,255,255,255 // je 3285 <_sk_load_u16_be_hsw+0x31> |
| .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 15,130,244,254,255,255 // jb 3285 <_sk_load_u16_be_hsw+0x31> |
| .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 |
| .byte 233,232,254,255,255 // jmpq 3285 <_sk_load_u16_be_hsw+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| .byte 233,219,254,255,255 // jmpq 3285 <_sk_load_u16_be_hsw+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 233,210,254,255,255 // jmpq 3285 <_sk_load_u16_be_hsw+0x31> |
| |
| HIDDEN _sk_load_rgb_u16_be_hsw |
| .globl _sk_load_rgb_u16_be_hsw |
| FUNCTION(_sk_load_rgb_u16_be_hsw) |
| _sk_load_rgb_u16_be_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,211,0,0,0 // jne 3498 <_sk_load_rgb_u16_be_hsw+0xe5> |
| .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 |
| .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 |
| .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 |
| .byte 196,193,122,111,92,64,32 // vmovdqu 0x20(%r8,%rax,2),%xmm3 |
| .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3 |
| .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8 |
| .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9 |
| .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11 |
| .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10 |
| .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0 |
| .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2 |
| .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1 |
| .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3 |
| .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,98,125,88,208 // vpbroadcastd %xmm0,%ymm10 |
| .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1 |
| .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0 |
| .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 |
| .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 |
| .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1 |
| .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1 |
| .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2 |
| .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3 |
| .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2 |
| .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2 |
| .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 |
| .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 117,5 // jne 34b1 <_sk_load_rgb_u16_be_hsw+0xfe> |
| .byte 233,72,255,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 |
| .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,26 // jb 34e0 <_sk_load_rgb_u16_be_hsw+0x12d> |
| .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 |
| .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 117,10 // jne 34e5 <_sk_load_rgb_u16_be_hsw+0x132> |
| .byte 233,25,255,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| .byte 233,20,255,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 |
| .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,26 // jb 3514 <_sk_load_rgb_u16_be_hsw+0x161> |
| .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 |
| .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 117,10 // jne 3519 <_sk_load_rgb_u16_be_hsw+0x166> |
| .byte 233,229,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| .byte 233,224,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 |
| .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,20 // jb 3542 <_sk_load_rgb_u16_be_hsw+0x18f> |
| .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 |
| .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 |
| .byte 233,183,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| .byte 233,178,254,255,255 // jmpq 33f9 <_sk_load_rgb_u16_be_hsw+0x46> |
| |
| HIDDEN _sk_store_u16_be_hsw |
| .globl _sk_store_u16_be_hsw |
| FUNCTION(_sk_store_u16_be_hsw) |
| _sk_store_u16_be_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 184,0,255,127,71 // mov $0x477fff00,%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 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 |
| .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9 |
| .byte 196,193,41,113,241,8 // vpsllw $0x8,%xmm9,%xmm10 |
| .byte 196,193,49,113,209,8 // vpsrlw $0x8,%xmm9,%xmm9 |
| .byte 196,65,41,235,201 // vpor %xmm9,%xmm10,%xmm9 |
| .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11 |
| .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10 |
| .byte 196,193,33,113,242,8 // vpsllw $0x8,%xmm10,%xmm11 |
| .byte 196,193,41,113,210,8 // vpsrlw $0x8,%xmm10,%xmm10 |
| .byte 196,65,33,235,210 // vpor %xmm10,%xmm11,%xmm10 |
| .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11 |
| .byte 196,65,125,91,219 // vcvtps2dq %ymm11,%ymm11 |
| .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 |
| .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11 |
| .byte 196,193,25,113,243,8 // vpsllw $0x8,%xmm11,%xmm12 |
| .byte 196,193,33,113,211,8 // vpsrlw $0x8,%xmm11,%xmm11 |
| .byte 196,65,25,235,219 // vpor %xmm11,%xmm12,%xmm11 |
| .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12 |
| .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8 |
| .byte 196,193,25,113,240,8 // vpsllw $0x8,%xmm8,%xmm12 |
| .byte 196,193,57,113,208,8 // vpsrlw $0x8,%xmm8,%xmm8 |
| .byte 196,65,25,235,192 // vpor %xmm8,%xmm12,%xmm8 |
| .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12 |
| .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13 |
| .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9 |
| .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8 |
| .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 |
| .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 117,31 // jne 3642 <_sk_store_u16_be_hsw+0xfb> |
| .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2) |
| .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2) |
| .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2) |
| .byte 196,1,122,127,68,72,48 // vmovdqu %xmm8,0x30(%r8,%r9,2) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2) |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 116,240 // je 363e <_sk_store_u16_be_hsw+0xf7> |
| .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2) |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,227 // jb 363e <_sk_store_u16_be_hsw+0xf7> |
| .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2) |
| .byte 116,218 // je 363e <_sk_store_u16_be_hsw+0xf7> |
| .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2) |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,205 // jb 363e <_sk_store_u16_be_hsw+0xf7> |
| .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2) |
| .byte 116,196 // je 363e <_sk_store_u16_be_hsw+0xf7> |
| .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2) |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,183 // jb 363e <_sk_store_u16_be_hsw+0xf7> |
| .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2) |
| .byte 235,174 // jmp 363e <_sk_store_u16_be_hsw+0xf7> |
| |
| HIDDEN _sk_load_f32_hsw |
| .globl _sk_load_f32_hsw |
| FUNCTION(_sk_load_f32_hsw) |
| _sk_load_f32_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 119,110 // ja 3706 <_sk_load_f32_hsw+0x76> |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 3730 <_sk_load_f32_hsw+0xa0> |
| .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax |
| .byte 76,1,208 // add %r10,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,3,125,24,68,136,112,1 // vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8 |
| .byte 196,131,125,24,92,136,96,1 // vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3 |
| .byte 196,131,125,24,76,136,80,1 // vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1 |
| .byte 196,131,125,24,84,136,64,1 // vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2 |
| .byte 196,129,121,16,68,136,48 // vmovupd 0x30(%r8,%r9,4),%xmm0 |
| .byte 196,195,125,13,192,12 // vblendpd $0xc,%ymm8,%ymm0,%ymm0 |
| .byte 196,1,121,16,68,136,32 // vmovupd 0x20(%r8,%r9,4),%xmm8 |
| .byte 196,99,61,13,203,12 // vblendpd $0xc,%ymm3,%ymm8,%ymm9 |
| .byte 196,129,121,16,92,136,16 // vmovupd 0x10(%r8,%r9,4),%xmm3 |
| .byte 196,99,101,13,209,12 // vblendpd $0xc,%ymm1,%ymm3,%ymm10 |
| .byte 196,129,121,16,12,136 // vmovupd (%r8,%r9,4),%xmm1 |
| .byte 196,227,117,13,202,12 // vblendpd $0xc,%ymm2,%ymm1,%ymm1 |
| .byte 196,193,116,20,210 // vunpcklps %ymm10,%ymm1,%ymm2 |
| .byte 196,193,116,21,218 // vunpckhps %ymm10,%ymm1,%ymm3 |
| .byte 197,180,20,200 // vunpcklps %ymm0,%ymm9,%ymm1 |
| .byte 197,52,21,192 // vunpckhps %ymm0,%ymm9,%ymm8 |
| .byte 197,237,20,193 // vunpcklpd %ymm1,%ymm2,%ymm0 |
| .byte 197,237,21,201 // vunpckhpd %ymm1,%ymm2,%ymm1 |
| .byte 196,193,101,20,208 // vunpcklpd %ymm8,%ymm3,%ymm2 |
| .byte 196,193,101,21,216 // vunpckhpd %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 102,144 // xchg %ax,%ax |
| .byte 131,255,255 // cmp $0xffffffff,%edi |
| .byte 255,202 // dec %edx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 189,255,255,255,176 // mov $0xb0ffffff,%ebp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,163,255,255,255,155 // jmpq *-0x64000001(%rbx) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,147,255,255,255,139 // callq *-0x74000001(%rbx) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_store_f32_hsw |
| .globl _sk_store_f32_hsw |
| FUNCTION(_sk_store_f32_hsw) |
| _sk_store_f32_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax |
| .byte 197,124,20,193 // vunpcklps %ymm1,%ymm0,%ymm8 |
| .byte 197,124,21,217 // vunpckhps %ymm1,%ymm0,%ymm11 |
| .byte 197,108,20,203 // vunpcklps %ymm3,%ymm2,%ymm9 |
| .byte 197,108,21,227 // vunpckhps %ymm3,%ymm2,%ymm12 |
| .byte 196,65,61,20,209 // vunpcklpd %ymm9,%ymm8,%ymm10 |
| .byte 196,65,61,21,201 // vunpckhpd %ymm9,%ymm8,%ymm9 |
| .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 37bd <_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 |
| .byte 196,67,61,6,195,49 // vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 |
| .byte 196,65,125,17,36,128 // vmovupd %ymm12,(%r8,%rax,4) |
| .byte 196,65,125,17,108,128,32 // vmovupd %ymm13,0x20(%r8,%rax,4) |
| .byte 196,65,125,17,76,128,64 // vmovupd %ymm9,0x40(%r8,%rax,4) |
| .byte 196,65,125,17,68,128,96 // vmovupd %ymm8,0x60(%r8,%rax,4) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .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 37b9 <_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 37b9 <_sk_store_f32_hsw+0x69> |
| .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) |
| .byte 116,218 // je 37b9 <_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 37b9 <_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 37b9 <_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 37b9 <_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 37b9 <_sk_store_f32_hsw+0x69> |
| |
| HIDDEN _sk_clamp_x_hsw |
| .globl _sk_clamp_x_hsw |
| FUNCTION(_sk_clamp_x_hsw) |
| _sk_clamp_x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0 |
| .byte 196,98,125,88,0 // vpbroadcastd (%rax),%ymm8 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,65,61,254,193 // vpaddd %ymm9,%ymm8,%ymm8 |
| .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_y_hsw |
| .globl _sk_clamp_y_hsw |
| FUNCTION(_sk_clamp_y_hsw) |
| _sk_clamp_y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,188,95,201 // vmaxps %ymm1,%ymm8,%ymm1 |
| .byte 196,98,125,88,0 // vpbroadcastd (%rax),%ymm8 |
| .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| .byte 196,65,61,254,193 // vpaddd %ymm9,%ymm8,%ymm8 |
| .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_x_hsw |
| .globl _sk_repeat_x_hsw |
| FUNCTION(_sk_repeat_x_hsw) |
| _sk_repeat_x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9 |
| .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| .byte 196,98,61,172,200 // vfnmadd213ps %ymm0,%ymm8,%ymm9 |
| .byte 197,253,118,192 // vpcmpeqd %ymm0,%ymm0,%ymm0 |
| .byte 197,189,254,192 // vpaddd %ymm0,%ymm8,%ymm0 |
| .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_y_hsw |
| .globl _sk_repeat_y_hsw |
| FUNCTION(_sk_repeat_y_hsw) |
| _sk_repeat_y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9 |
| .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| .byte 196,98,61,172,201 // vfnmadd213ps %ymm1,%ymm8,%ymm9 |
| .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 |
| .byte 197,189,254,201 // vpaddd %ymm1,%ymm8,%ymm1 |
| .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_x_hsw |
| .globl _sk_mirror_x_hsw |
| FUNCTION(_sk_mirror_x_hsw) |
| _sk_mirror_x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,122,16,0 // vmovss (%rax),%xmm8 |
| .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9 |
| .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10 |
| .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0 |
| .byte 196,226,125,24,192 // vbroadcastss %xmm0,%ymm0 |
| .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8 |
| .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| .byte 196,66,125,172,194 // vfnmadd213ps %ymm10,%ymm0,%ymm8 |
| .byte 196,193,60,92,193 // vsubps %ymm9,%ymm8,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8 |
| .byte 197,188,84,192 // vandps %ymm0,%ymm8,%ymm0 |
| .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8 |
| .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_y_hsw |
| .globl _sk_mirror_y_hsw |
| FUNCTION(_sk_mirror_y_hsw) |
| _sk_mirror_y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,122,16,0 // vmovss (%rax),%xmm8 |
| .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9 |
| .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10 |
| .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1 |
| .byte 196,226,125,24,201 // vbroadcastss %xmm1,%ymm1 |
| .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8 |
| .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| .byte 196,66,117,172,194 // vfnmadd213ps %ymm10,%ymm1,%ymm8 |
| .byte 196,193,60,92,201 // vsubps %ymm9,%ymm8,%ymm1 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8 |
| .byte 197,188,84,201 // vandps %ymm1,%ymm8,%ymm1 |
| .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8 |
| .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_luminance_to_alpha_hsw |
| .globl _sk_luminance_to_alpha_hsw |
| FUNCTION(_sk_luminance_to_alpha_hsw) |
| _sk_luminance_to_alpha_hsw: |
| .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,98,125,88,195 // vpbroadcastd %xmm3,%ymm8 |
| .byte 184,89,23,55,63 // mov $0x3f371759,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| .byte 197,228,89,201 // vmulps %ymm1,%ymm3,%ymm1 |
| .byte 196,98,125,168,193 // vfmadd213ps %ymm1,%ymm0,%ymm8 |
| .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,226,125,88,216 // vpbroadcastd %xmm0,%ymm3 |
| .byte 196,194,109,168,216 // vfmadd213ps %ymm8,%ymm2,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,253,239,192 // vpxor %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_2x3_hsw |
| .globl _sk_matrix_2x3_hsw |
| FUNCTION(_sk_matrix_2x3_hsw) |
| _sk_matrix_2x3_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 |
| .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8 |
| .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm11 |
| .byte 196,98,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm9 |
| .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9 |
| .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_3x4_hsw |
| .globl _sk_matrix_3x4_hsw |
| FUNCTION(_sk_matrix_3x4_hsw) |
| _sk_matrix_3x4_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 |
| .byte 196,98,125,24,80,12 // vbroadcastss 0xc(%rax),%ymm10 |
| .byte 196,98,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm11 |
| .byte 196,98,125,24,64,36 // vbroadcastss 0x24(%rax),%ymm8 |
| .byte 196,66,109,184,195 // vfmadd231ps %ymm11,%ymm2,%ymm8 |
| .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8 |
| .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,16 // vbroadcastss 0x10(%rax),%ymm11 |
| .byte 196,98,125,24,96,28 // vbroadcastss 0x1c(%rax),%ymm12 |
| .byte 196,98,125,24,72,40 // vbroadcastss 0x28(%rax),%ymm9 |
| .byte 196,66,109,184,204 // vfmadd231ps %ymm12,%ymm2,%ymm9 |
| .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9 |
| .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 196,98,125,24,96,20 // vbroadcastss 0x14(%rax),%ymm12 |
| .byte 196,98,125,24,104,32 // vbroadcastss 0x20(%rax),%ymm13 |
| .byte 196,98,125,24,80,44 // vbroadcastss 0x2c(%rax),%ymm10 |
| .byte 196,66,109,184,213 // vfmadd231ps %ymm13,%ymm2,%ymm10 |
| .byte 196,66,117,184,212 // vfmadd231ps %ymm12,%ymm1,%ymm10 |
| .byte 196,66,125,184,211 // vfmadd231ps %ymm11,%ymm0,%ymm10 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| .byte 197,124,41,210 // vmovaps %ymm10,%ymm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_4x5_hsw |
| .globl _sk_matrix_4x5_hsw |
| FUNCTION(_sk_matrix_4x5_hsw) |
| _sk_matrix_4x5_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 |
| .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| .byte 196,98,125,24,88,32 // vbroadcastss 0x20(%rax),%ymm11 |
| .byte 196,98,125,24,96,48 // vbroadcastss 0x30(%rax),%ymm12 |
| .byte 196,98,125,24,64,64 // vbroadcastss 0x40(%rax),%ymm8 |
| .byte 196,66,101,184,196 // vfmadd231ps %ymm12,%ymm3,%ymm8 |
| .byte 196,66,109,184,195 // vfmadd231ps %ymm11,%ymm2,%ymm8 |
| .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8 |
| .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| .byte 196,98,125,24,96,36 // vbroadcastss 0x24(%rax),%ymm12 |
| .byte 196,98,125,24,104,52 // vbroadcastss 0x34(%rax),%ymm13 |
| .byte 196,98,125,24,72,68 // vbroadcastss 0x44(%rax),%ymm9 |
| .byte 196,66,101,184,205 // vfmadd231ps %ymm13,%ymm3,%ymm9 |
| .byte 196,66,109,184,204 // vfmadd231ps %ymm12,%ymm2,%ymm9 |
| .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9 |
| .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 196,98,125,24,96,24 // vbroadcastss 0x18(%rax),%ymm12 |
| .byte 196,98,125,24,104,40 // vbroadcastss 0x28(%rax),%ymm13 |
| .byte 196,98,125,24,112,56 // vbroadcastss 0x38(%rax),%ymm14 |
| .byte 196,98,125,24,80,72 // vbroadcastss 0x48(%rax),%ymm10 |
| .byte 196,66,101,184,214 // vfmadd231ps %ymm14,%ymm3,%ymm10 |
| .byte 196,66,109,184,213 // vfmadd231ps %ymm13,%ymm2,%ymm10 |
| .byte 196,66,117,184,212 // vfmadd231ps %ymm12,%ymm1,%ymm10 |
| .byte 196,66,125,184,211 // vfmadd231ps %ymm11,%ymm0,%ymm10 |
| .byte 196,98,125,24,96,12 // vbroadcastss 0xc(%rax),%ymm12 |
| .byte 196,98,125,24,104,28 // vbroadcastss 0x1c(%rax),%ymm13 |
| .byte 196,98,125,24,112,44 // vbroadcastss 0x2c(%rax),%ymm14 |
| .byte 196,98,125,24,120,60 // vbroadcastss 0x3c(%rax),%ymm15 |
| .byte 196,98,125,24,88,76 // vbroadcastss 0x4c(%rax),%ymm11 |
| .byte 196,66,101,184,223 // vfmadd231ps %ymm15,%ymm3,%ymm11 |
| .byte 196,66,109,184,222 // vfmadd231ps %ymm14,%ymm2,%ymm11 |
| .byte 196,66,117,184,221 // vfmadd231ps %ymm13,%ymm1,%ymm11 |
| .byte 196,66,125,184,220 // vfmadd231ps %ymm12,%ymm0,%ymm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| .byte 197,124,41,210 // vmovaps %ymm10,%ymm2 |
| .byte 197,124,41,219 // vmovaps %ymm11,%ymm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_perspective_hsw |
| .globl _sk_matrix_perspective_hsw |
| FUNCTION(_sk_matrix_perspective_hsw) |
| _sk_matrix_perspective_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| .byte 196,66,117,184,209 // vfmadd231ps %ymm9,%ymm1,%ymm10 |
| .byte 196,66,125,184,208 // vfmadd231ps %ymm8,%ymm0,%ymm10 |
| .byte 196,98,125,24,64,12 // vbroadcastss 0xc(%rax),%ymm8 |
| .byte 196,98,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm9 |
| .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| .byte 196,66,117,184,217 // vfmadd231ps %ymm9,%ymm1,%ymm11 |
| .byte 196,66,125,184,216 // vfmadd231ps %ymm8,%ymm0,%ymm11 |
| .byte 196,98,125,24,64,24 // vbroadcastss 0x18(%rax),%ymm8 |
| .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 |
| .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12 |
| .byte 196,66,117,184,225 // vfmadd231ps %ymm9,%ymm1,%ymm12 |
| .byte 196,66,125,184,224 // vfmadd231ps %ymm8,%ymm0,%ymm12 |
| .byte 196,193,124,83,204 // vrcpps %ymm12,%ymm1 |
| .byte 197,172,89,193 // vmulps %ymm1,%ymm10,%ymm0 |
| .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_hsw |
| .globl _sk_linear_gradient_hsw |
| FUNCTION(_sk_linear_gradient_hsw) |
| _sk_linear_gradient_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 77,133,192 // test %r8,%r8 |
| .byte 15,132,143,0,0,0 // je 3c49 <_sk_linear_gradient_hsw+0xb5> |
| .byte 72,139,64,8 // mov 0x8(%rax),%rax |
| .byte 72,131,192,32 // add $0x20,%rax |
| .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 196,98,125,24,104,224 // vbroadcastss -0x20(%rax),%ymm13 |
| .byte 196,65,124,194,237,1 // vcmpltps %ymm13,%ymm0,%ymm13 |
| .byte 196,98,125,24,112,228 // vbroadcastss -0x1c(%rax),%ymm14 |
| .byte 196,67,13,74,228,208 // vblendvps %ymm13,%ymm12,%ymm14,%ymm12 |
| .byte 196,98,125,24,112,232 // vbroadcastss -0x18(%rax),%ymm14 |
| .byte 196,227,13,74,201,208 // vblendvps %ymm13,%ymm1,%ymm14,%ymm1 |
| .byte 196,98,125,24,112,236 // vbroadcastss -0x14(%rax),%ymm14 |
| .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2 |
| .byte 196,98,125,24,112,240 // vbroadcastss -0x10(%rax),%ymm14 |
| .byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3 |
| .byte 196,98,125,24,112,244 // vbroadcastss -0xc(%rax),%ymm14 |
| .byte 196,67,13,74,192,208 // vblendvps %ymm13,%ymm8,%ymm14,%ymm8 |
| .byte 196,98,125,24,112,248 // vbroadcastss -0x8(%rax),%ymm14 |
| .byte 196,67,13,74,219,208 // vblendvps %ymm13,%ymm11,%ymm14,%ymm11 |
| .byte 196,98,125,24,112,252 // vbroadcastss -0x4(%rax),%ymm14 |
| .byte 196,67,13,74,210,208 // vblendvps %ymm13,%ymm10,%ymm14,%ymm10 |
| .byte 196,98,125,24,48 // vbroadcastss (%rax),%ymm14 |
| .byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9 |
| .byte 72,131,192,36 // add $0x24,%rax |
| .byte 73,255,200 // dec %r8 |
| .byte 117,140 // jne 3bd3 <_sk_linear_gradient_hsw+0x3f> |
| .byte 235,17 // jmp 3c5a <_sk_linear_gradient_hsw+0xc6> |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 |
| .byte 196,66,125,184,196 // vfmadd231ps %ymm12,%ymm0,%ymm8 |
| .byte 196,194,125,168,203 // vfmadd213ps %ymm11,%ymm0,%ymm1 |
| .byte 196,194,125,168,210 // vfmadd213ps %ymm10,%ymm0,%ymm2 |
| .byte 196,194,125,168,217 // vfmadd213ps %ymm9,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_2stops_hsw |
| .globl _sk_linear_gradient_2stops_hsw |
| FUNCTION(_sk_linear_gradient_2stops_hsw) |
| _sk_linear_gradient_2stops_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1 |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,98,125,184,193 // vfmadd231ps %ymm1,%ymm0,%ymm8 |
| .byte 196,226,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm2 |
| .byte 196,226,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm1 |
| .byte 196,226,125,184,202 // vfmadd231ps %ymm2,%ymm0,%ymm1 |
| .byte 196,226,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm3 |
| .byte 196,226,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm2 |
| .byte 196,226,125,184,211 // vfmadd231ps %ymm3,%ymm0,%ymm2 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3 |
| .byte 196,194,125,184,217 // vfmadd231ps %ymm9,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_save_xy_hsw |
| .globl _sk_save_xy_hsw |
| FUNCTION(_sk_save_xy_hsw) |
| _sk_save_xy_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,88,200 // vaddps %ymm0,%ymm8,%ymm9 |
| .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10 |
| .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,88,193 // vaddps %ymm1,%ymm8,%ymm8 |
| .byte 196,67,125,8,208,1 // vroundps $0x1,%ymm8,%ymm10 |
| .byte 196,65,60,92,194 // vsubps %ymm10,%ymm8,%ymm8 |
| .byte 197,252,17,0 // vmovups %ymm0,(%rax) |
| .byte 197,252,17,72,32 // vmovups %ymm1,0x20(%rax) |
| .byte 197,124,17,72,64 // vmovups %ymm9,0x40(%rax) |
| .byte 197,124,17,64,96 // vmovups %ymm8,0x60(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_accumulate_hsw |
| .globl _sk_accumulate_hsw |
| FUNCTION(_sk_accumulate_hsw) |
| _sk_accumulate_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,16,128,128,0,0,0 // vmovups 0x80(%rax),%ymm8 |
| .byte 197,60,89,128,160,0,0,0 // vmulps 0xa0(%rax),%ymm8,%ymm8 |
| .byte 196,226,61,184,224 // vfmadd231ps %ymm0,%ymm8,%ymm4 |
| .byte 196,226,61,184,233 // vfmadd231ps %ymm1,%ymm8,%ymm5 |
| .byte 196,226,61,184,242 // vfmadd231ps %ymm2,%ymm8,%ymm6 |
| .byte 196,98,101,168,199 // vfmadd213ps %ymm7,%ymm3,%ymm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,199 // vmovaps %ymm8,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_nx_hsw |
| .globl _sk_bilinear_nx_hsw |
| FUNCTION(_sk_bilinear_nx_hsw) |
| _sk_bilinear_nx_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_px_hsw |
| .globl _sk_bilinear_px_hsw |
| FUNCTION(_sk_bilinear_px_hsw) |
| _sk_bilinear_px_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_ny_hsw |
| .globl _sk_bilinear_ny_hsw |
| FUNCTION(_sk_bilinear_ny_hsw) |
| _sk_bilinear_ny_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_py_hsw |
| .globl _sk_bilinear_py_hsw |
| FUNCTION(_sk_bilinear_py_hsw) |
| _sk_bilinear_py_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3x_hsw |
| .globl _sk_bicubic_n3x_hsw |
| FUNCTION(_sk_bicubic_n3x_hsw) |
| _sk_bicubic_n3x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10 |
| .byte 196,65,44,89,193 // vmulps %ymm9,%ymm10,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1x_hsw |
| .globl _sk_bicubic_n1x_hsw |
| FUNCTION(_sk_bicubic_n1x_hsw) |
| _sk_bicubic_n1x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,200 // vmovd %r8d,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 196,66,61,168,202 // vfmadd213ps %ymm10,%ymm8,%ymm9 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,200 // vmovd %r8d,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9 |
| .byte 197,124,17,136,128,0,0,0 // vmovups %ymm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1x_hsw |
| .globl _sk_bicubic_p1x_hsw |
| FUNCTION(_sk_bicubic_p1x_hsw) |
| _sk_bicubic_p1x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,98,125,88,192 // vpbroadcastd %xmm0,%ymm8 |
| .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 |
| .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 196,66,53,168,211 // vfmadd213ps %ymm11,%ymm9,%ymm10 |
| .byte 196,66,53,168,208 // vfmadd213ps %ymm8,%ymm9,%ymm10 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,66,53,184,194 // vfmadd231ps %ymm10,%ymm9,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3x_hsw |
| .globl _sk_bicubic_p3x_hsw |
| FUNCTION(_sk_bicubic_p3x_hsw) |
| _sk_bicubic_p3x_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10 |
| .byte 196,65,52,89,194 // vmulps %ymm10,%ymm9,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3y_hsw |
| .globl _sk_bicubic_n3y_hsw |
| FUNCTION(_sk_bicubic_n3y_hsw) |
| _sk_bicubic_n3y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10 |
| .byte 196,65,44,89,193 // vmulps %ymm9,%ymm10,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1y_hsw |
| .globl _sk_bicubic_n1y_hsw |
| FUNCTION(_sk_bicubic_n1y_hsw) |
| _sk_bicubic_n1y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,200 // vmovd %r8d,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 196,66,61,168,202 // vfmadd213ps %ymm10,%ymm8,%ymm9 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,200 // vmovd %r8d,%xmm9 |
| .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9 |
| .byte 197,124,17,136,160,0,0,0 // vmovups %ymm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1y_hsw |
| .globl _sk_bicubic_p1y_hsw |
| FUNCTION(_sk_bicubic_p1y_hsw) |
| _sk_bicubic_p1y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 |
| .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 |
| .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 196,66,53,168,211 // vfmadd213ps %ymm11,%ymm9,%ymm10 |
| .byte 196,66,53,168,208 // vfmadd213ps %ymm8,%ymm9,%ymm10 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| .byte 196,66,53,184,194 // vfmadd231ps %ymm10,%ymm9,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3y_hsw |
| .globl _sk_bicubic_p3y_hsw |
| FUNCTION(_sk_bicubic_p3y_hsw) |
| _sk_bicubic_p3y_hsw: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10 |
| .byte 196,65,52,89,194 // vmulps %ymm10,%ymm9,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_callback_hsw |
| .globl _sk_callback_hsw |
| FUNCTION(_sk_callback_hsw) |
| _sk_callback_hsw: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,129,236,24,1,0,0 // sub $0x118,%rsp |
| .byte 197,252,17,188,36,224,0,0,0 // vmovups %ymm7,0xe0(%rsp) |
| .byte 197,252,17,180,36,192,0,0,0 // vmovups %ymm6,0xc0(%rsp) |
| .byte 197,252,17,172,36,160,0,0,0 // vmovups %ymm5,0xa0(%rsp) |
| .byte 197,252,17,164,36,128,0,0,0 // vmovups %ymm4,0x80(%rsp) |
| .byte 197,252,17,92,36,96 // vmovups %ymm3,0x60(%rsp) |
| .byte 197,252,17,84,36,64 // vmovups %ymm2,0x40(%rsp) |
| .byte 197,252,17,76,36,32 // vmovups %ymm1,0x20(%rsp) |
| .byte 197,252,17,4,36 // vmovups %ymm0,(%rsp) |
| .byte 72,137,203 // mov %rcx,%rbx |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 73,137,255 // mov %rdi,%r15 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,244 // mov %rsi,%r12 |
| .byte 72,139,120,8 // mov 0x8(%rax),%rdi |
| .byte 72,133,219 // test %rbx,%rbx |
| .byte 190,8,0,0,0 // mov $0x8,%esi |
| .byte 15,69,243 // cmovne %ebx,%esi |
| .byte 197,248,119 // vzeroupper |
| .byte 255,16 // callq *(%rax) |
| .byte 76,137,230 // mov %r12,%rsi |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,255 // mov %r15,%rdi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 72,137,217 // mov %rbx,%rcx |
| .byte 197,252,16,4,36 // vmovups (%rsp),%ymm0 |
| .byte 197,252,16,76,36,32 // vmovups 0x20(%rsp),%ymm1 |
| .byte 197,252,16,84,36,64 // vmovups 0x40(%rsp),%ymm2 |
| .byte 197,252,16,92,36,96 // vmovups 0x60(%rsp),%ymm3 |
| .byte 197,252,16,164,36,128,0,0,0 // vmovups 0x80(%rsp),%ymm4 |
| .byte 197,252,16,172,36,160,0,0,0 // vmovups 0xa0(%rsp),%ymm5 |
| .byte 197,252,16,180,36,192,0,0,0 // vmovups 0xc0(%rsp),%ymm6 |
| .byte 197,252,16,188,36,224,0,0,0 // vmovups 0xe0(%rsp),%ymm7 |
| .byte 72,129,196,24,1,0,0 // add $0x118,%rsp |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_start_pipeline_avx |
| .globl _sk_start_pipeline_avx |
| FUNCTION(_sk_start_pipeline_avx) |
| _sk_start_pipeline_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 73,137,205 // mov %rcx,%r13 |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 72,137,251 // mov %rdi,%rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,199 // mov %rax,%r15 |
| .byte 73,137,244 // mov %rsi,%r12 |
| .byte 72,141,67,8 // lea 0x8(%rbx),%rax |
| .byte 76,57,232 // cmp %r13,%rax |
| .byte 118,5 // jbe 28 <_sk_start_pipeline_avx+0x28> |
| .byte 72,137,223 // mov %rbx,%rdi |
| .byte 235,65 // jmp 69 <_sk_start_pipeline_avx+0x69> |
| .byte 185,0,0,0,0 // mov $0x0,%ecx |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| .byte 72,137,223 // mov %rbx,%rdi |
| .byte 76,137,230 // mov %r12,%rsi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 65,255,215 // callq *%r15 |
| .byte 72,141,123,8 // lea 0x8(%rbx),%rdi |
| .byte 72,131,195,16 // add $0x10,%rbx |
| .byte 76,57,235 // cmp %r13,%rbx |
| .byte 72,137,251 // mov %rdi,%rbx |
| .byte 118,191 // jbe 28 <_sk_start_pipeline_avx+0x28> |
| .byte 76,137,233 // mov %r13,%rcx |
| .byte 72,41,249 // sub %rdi,%rcx |
| .byte 116,41 // je 9a <_sk_start_pipeline_avx+0x9a> |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| .byte 76,137,230 // mov %r12,%rsi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 65,255,215 // callq *%r15 |
| .byte 76,137,232 // mov %r13,%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 197,248,119 // vzeroupper |
| .byte 195 // retq |
| |
| HIDDEN _sk_just_return_avx |
| .globl _sk_just_return_avx |
| FUNCTION(_sk_just_return_avx) |
| _sk_just_return_avx: |
| .byte 195 // retq |
| |
| HIDDEN _sk_seed_shader_avx |
| .globl _sk_seed_shader_avx |
| FUNCTION(_sk_seed_shader_avx) |
| _sk_seed_shader_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,249,110,199 // vmovd %edi,%xmm0 |
| .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 |
| .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_constant_color_avx |
| .globl _sk_constant_color_avx |
| FUNCTION(_sk_constant_color_avx) |
| _sk_constant_color_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clear_avx |
| .globl _sk_clear_avx |
| FUNCTION(_sk_clear_avx) |
| _sk_clear_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcatop_avx |
| .globl _sk_srcatop_avx |
| FUNCTION(_sk_srcatop_avx) |
| _sk_srcatop_avx: |
| .byte 197,124,89,199 // vmulps %ymm7,%ymm0,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,124,92,203 // vsubps %ymm3,%ymm0,%ymm9 |
| .byte 197,180,89,196 // vmulps %ymm4,%ymm9,%ymm0 |
| .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,52,89,197 // vmulps %ymm5,%ymm9,%ymm8 |
| .byte 196,193,116,88,200 // vaddps %ymm8,%ymm1,%ymm1 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,52,89,198 // vmulps %ymm6,%ymm9,%ymm8 |
| .byte 196,193,108,88,208 // vaddps %ymm8,%ymm2,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 197,52,89,199 // vmulps %ymm7,%ymm9,%ymm8 |
| .byte 196,193,100,88,216 // vaddps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstatop_avx |
| .globl _sk_dstatop_avx |
| FUNCTION(_sk_dstatop_avx) |
| _sk_dstatop_avx: |
| .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9 |
| .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 |
| .byte 197,100,89,197 // vmulps %ymm5,%ymm3,%ymm8 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 197,188,88,201 // vaddps %ymm1,%ymm8,%ymm1 |
| .byte 197,100,89,198 // vmulps %ymm6,%ymm3,%ymm8 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 197,188,88,210 // vaddps %ymm2,%ymm8,%ymm2 |
| .byte 197,100,89,199 // vmulps %ymm7,%ymm3,%ymm8 |
| .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcin_avx |
| .globl _sk_srcin_avx |
| FUNCTION(_sk_srcin_avx) |
| _sk_srcin_avx: |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstin_avx |
| .globl _sk_dstin_avx |
| FUNCTION(_sk_dstin_avx) |
| _sk_dstin_avx: |
| .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcout_avx |
| .globl _sk_srcout_avx |
| FUNCTION(_sk_srcout_avx) |
| _sk_srcout_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstout_avx |
| .globl _sk_dstout_avx |
| FUNCTION(_sk_dstout_avx) |
| _sk_dstout_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 |
| .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcover_avx |
| .globl _sk_srcover_avx |
| FUNCTION(_sk_srcover_avx) |
| _sk_srcover_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 |
| .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 |
| .byte 197,60,89,205 // vmulps %ymm5,%ymm8,%ymm9 |
| .byte 197,180,88,201 // vaddps %ymm1,%ymm9,%ymm1 |
| .byte 197,60,89,206 // vmulps %ymm6,%ymm8,%ymm9 |
| .byte 197,180,88,210 // vaddps %ymm2,%ymm9,%ymm2 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstover_avx |
| .globl _sk_dstover_avx |
| FUNCTION(_sk_dstover_avx) |
| _sk_dstover_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_modulate_avx |
| .globl _sk_modulate_avx |
| FUNCTION(_sk_modulate_avx) |
| _sk_modulate_avx: |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_multiply_avx |
| .globl _sk_multiply_avx |
| FUNCTION(_sk_multiply_avx) |
| _sk_multiply_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 |
| .byte 197,52,89,209 // vmulps %ymm1,%ymm9,%ymm10 |
| .byte 197,60,89,221 // vmulps %ymm5,%ymm8,%ymm11 |
| .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 |
| .byte 197,52,89,210 // vmulps %ymm2,%ymm9,%ymm10 |
| .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 196,193,100,88,216 // vaddps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_plus__avx |
| .globl _sk_plus__avx |
| FUNCTION(_sk_plus__avx) |
| _sk_plus__avx: |
| .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_screen_avx |
| .globl _sk_screen_avx |
| FUNCTION(_sk_screen_avx) |
| _sk_screen_avx: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 197,100,88,199 // vaddps %ymm7,%ymm3,%ymm8 |
| .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| .byte 197,188,92,219 // vsubps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_xor__avx |
| .globl _sk_xor__avx |
| FUNCTION(_sk_xor__avx) |
| _sk_xor__avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 |
| .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 197,60,89,213 // vmulps %ymm5,%ymm8,%ymm10 |
| .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 197,60,89,214 // vmulps %ymm6,%ymm8,%ymm10 |
| .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_darken_avx |
| .globl _sk_darken_avx |
| FUNCTION(_sk_darken_avx) |
| _sk_darken_avx: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| .byte 196,193,124,95,193 // vmaxps %ymm9,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| .byte 196,193,116,95,201 // vmaxps %ymm9,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lighten_avx |
| .globl _sk_lighten_avx |
| FUNCTION(_sk_lighten_avx) |
| _sk_lighten_avx: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_difference_avx |
| .globl _sk_difference_avx |
| FUNCTION(_sk_difference_avx) |
| _sk_difference_avx: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_exclusion_avx |
| .globl _sk_exclusion_avx |
| FUNCTION(_sk_exclusion_avx) |
| _sk_exclusion_avx: |
| .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colorburn_avx |
| .globl _sk_colorburn_avx |
| FUNCTION(_sk_colorburn_avx) |
| _sk_colorburn_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 |
| .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,228 // vmulps %ymm4,%ymm8,%ymm12 |
| .byte 197,68,92,236 // vsubps %ymm4,%ymm7,%ymm13 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 197,20,94,232 // vdivps %ymm0,%ymm13,%ymm13 |
| .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 196,65,20,88,235 // vaddps %ymm11,%ymm13,%ymm13 |
| .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| .byte 197,28,88,224 // vaddps %ymm0,%ymm12,%ymm12 |
| .byte 196,193,124,194,194,0 // vcmpeqps %ymm10,%ymm0,%ymm0 |
| .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0 |
| .byte 197,92,194,231,0 // vcmpeqps %ymm7,%ymm4,%ymm12 |
| .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| .byte 197,52,89,217 // vmulps %ymm1,%ymm9,%ymm11 |
| .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| .byte 197,68,92,237 // vsubps %ymm5,%ymm7,%ymm13 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 197,20,94,233 // vdivps %ymm1,%ymm13,%ymm13 |
| .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 196,65,36,88,237 // vaddps %ymm13,%ymm11,%ymm13 |
| .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| .byte 197,28,88,225 // vaddps %ymm1,%ymm12,%ymm12 |
| .byte 196,193,116,194,202,0 // vcmpeqps %ymm10,%ymm1,%ymm1 |
| .byte 196,195,21,74,204,16 // vblendvps %ymm1,%ymm12,%ymm13,%ymm1 |
| .byte 197,84,194,231,0 // vcmpeqps %ymm7,%ymm5,%ymm12 |
| .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9 |
| .byte 196,65,108,194,210,0 // vcmpeqps %ymm10,%ymm2,%ymm10 |
| .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| .byte 197,68,92,230 // vsubps %ymm6,%ymm7,%ymm12 |
| .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12 |
| .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 |
| .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| .byte 196,65,68,92,228 // vsubps %ymm12,%ymm7,%ymm12 |
| .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| .byte 196,65,52,88,228 // vaddps %ymm12,%ymm9,%ymm12 |
| .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| .byte 196,227,37,74,210,160 // vblendvps %ymm10,%ymm2,%ymm11,%ymm2 |
| .byte 197,76,194,215,0 // vcmpeqps %ymm7,%ymm6,%ymm10 |
| .byte 197,52,88,206 // vaddps %ymm6,%ymm9,%ymm9 |
| .byte 196,195,109,74,209,160 // vblendvps %ymm10,%ymm9,%ymm2,%ymm2 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colordodge_avx |
| .globl _sk_colordodge_avx |
| FUNCTION(_sk_colordodge_avx) |
| _sk_colordodge_avx: |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 |
| .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 |
| .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 |
| .byte 197,100,89,228 // vmulps %ymm4,%ymm3,%ymm12 |
| .byte 197,100,92,232 // vsubps %ymm0,%ymm3,%ymm13 |
| .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| .byte 197,52,89,236 // vmulps %ymm4,%ymm9,%ymm13 |
| .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| .byte 196,65,28,88,227 // vaddps %ymm11,%ymm12,%ymm12 |
| .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| .byte 197,20,88,232 // vaddps %ymm0,%ymm13,%ymm13 |
| .byte 197,252,194,195,0 // vcmpeqps %ymm3,%ymm0,%ymm0 |
| .byte 196,195,29,74,197,0 // vblendvps %ymm0,%ymm13,%ymm12,%ymm0 |
| .byte 196,65,92,194,224,0 // vcmpeqps %ymm8,%ymm4,%ymm12 |
| .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| .byte 197,100,89,229 // vmulps %ymm5,%ymm3,%ymm12 |
| .byte 197,100,92,233 // vsubps %ymm1,%ymm3,%ymm13 |
| .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| .byte 197,52,89,237 // vmulps %ymm5,%ymm9,%ymm13 |
| .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| .byte 196,65,28,88,227 // vaddps %ymm11,%ymm12,%ymm12 |
| .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| .byte 197,20,88,233 // vaddps %ymm1,%ymm13,%ymm13 |
| .byte 197,244,194,203,0 // vcmpeqps %ymm3,%ymm1,%ymm1 |
| .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1 |
| .byte 196,65,84,194,224,0 // vcmpeqps %ymm8,%ymm5,%ymm12 |
| .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11 |
| .byte 197,100,92,226 // vsubps %ymm2,%ymm3,%ymm12 |
| .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 |
| .byte 197,52,89,230 // vmulps %ymm6,%ymm9,%ymm12 |
| .byte 196,65,68,93,219 // vminps %ymm11,%ymm7,%ymm11 |
| .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 |
| .byte 196,65,44,88,219 // vaddps %ymm11,%ymm10,%ymm11 |
| .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12 |
| .byte 197,236,194,211,0 // vcmpeqps %ymm3,%ymm2,%ymm2 |
| .byte 196,195,37,74,212,32 // vblendvps %ymm2,%ymm12,%ymm11,%ymm2 |
| .byte 196,65,76,194,192,0 // vcmpeqps %ymm8,%ymm6,%ymm8 |
| .byte 197,44,88,214 // vaddps %ymm6,%ymm10,%ymm10 |
| .byte 196,195,109,74,210,128 // vblendvps %ymm8,%ymm10,%ymm2,%ymm2 |
| .byte 197,52,89,199 // vmulps %ymm7,%ymm9,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hardlight_avx |
| .globl _sk_hardlight_avx |
| FUNCTION(_sk_hardlight_avx) |
| _sk_hardlight_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,65,52,88,219 // vaddps %ymm11,%ymm9,%ymm11 |
| .byte 197,124,88,200 // vaddps %ymm0,%ymm0,%ymm9 |
| .byte 197,52,194,227,2 // vcmpleps %ymm3,%ymm9,%ymm12 |
| .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| .byte 197,116,88,225 // vaddps %ymm1,%ymm1,%ymm12 |
| .byte 197,28,194,227,2 // vcmpleps %ymm3,%ymm12,%ymm12 |
| .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| .byte 197,108,88,218 // vaddps %ymm2,%ymm2,%ymm11 |
| .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_overlay_avx |
| .globl _sk_overlay_avx |
| FUNCTION(_sk_overlay_avx) |
| _sk_overlay_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,65,52,88,219 // vaddps %ymm11,%ymm9,%ymm11 |
| .byte 197,92,88,204 // vaddps %ymm4,%ymm4,%ymm9 |
| .byte 197,52,194,231,2 // vcmpleps %ymm7,%ymm9,%ymm12 |
| .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12 |
| .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| .byte 197,76,88,222 // vaddps %ymm6,%ymm6,%ymm11 |
| .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11 |
| .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_softlight_avx |
| .globl _sk_softlight_avx |
| FUNCTION(_sk_softlight_avx) |
| _sk_softlight_avx: |
| .byte 197,252,17,84,36,200 // vmovups %ymm2,-0x38(%rsp) |
| .byte 197,252,40,209 // vmovaps %ymm1,%ymm2 |
| .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 |
| .byte 197,52,194,215,1 // vcmpltps %ymm7,%ymm9,%ymm10 |
| .byte 197,92,94,199 // vdivps %ymm7,%ymm4,%ymm8 |
| .byte 196,67,53,74,216,160 // vblendvps %ymm10,%ymm8,%ymm9,%ymm11 |
| .byte 196,65,36,88,195 // vaddps %ymm11,%ymm11,%ymm8 |
| .byte 196,65,60,88,224 // vaddps %ymm8,%ymm8,%ymm12 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 196,65,28,89,236 // vmulps %ymm12,%ymm12,%ymm13 |
| .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| .byte 196,65,36,92,232 // vsubps %ymm8,%ymm11,%ymm13 |
| .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13 |
| .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| .byte 196,67,121,4,228,0 // vpermilps $0x0,%xmm12,%xmm12 |
| .byte 196,67,29,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm12,%ymm12 |
| .byte 196,65,36,89,244 // vmulps %ymm12,%ymm11,%ymm14 |
| .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 |
| .byte 196,65,124,82,243 // vrsqrtps %ymm11,%ymm14 |
| .byte 196,65,124,83,246 // vrcpps %ymm14,%ymm14 |
| .byte 196,65,12,92,243 // vsubps %ymm11,%ymm14,%ymm14 |
| .byte 197,92,88,252 // vaddps %ymm4,%ymm4,%ymm15 |
| .byte 196,65,4,88,255 // vaddps %ymm15,%ymm15,%ymm15 |
| .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15 |
| .byte 196,67,13,74,237,240 // vblendvps %ymm15,%ymm13,%ymm14,%ymm13 |
| .byte 197,124,88,240 // vaddps %ymm0,%ymm0,%ymm14 |
| .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| .byte 196,65,60,92,219 // vsubps %ymm11,%ymm8,%ymm11 |
| .byte 196,65,4,89,219 // vmulps %ymm11,%ymm15,%ymm11 |
| .byte 197,36,88,219 // vaddps %ymm3,%ymm11,%ymm11 |
| .byte 197,36,89,220 // vmulps %ymm4,%ymm11,%ymm11 |
| .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13 |
| .byte 197,100,89,252 // vmulps %ymm4,%ymm3,%ymm15 |
| .byte 196,65,4,88,237 // vaddps %ymm13,%ymm15,%ymm13 |
| .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| .byte 196,195,21,74,203,224 // vblendvps %ymm14,%ymm11,%ymm13,%ymm1 |
| .byte 197,84,94,239 // vdivps %ymm7,%ymm5,%ymm13 |
| .byte 196,67,53,74,237,160 // vblendvps %ymm10,%ymm13,%ymm9,%ymm13 |
| .byte 196,65,20,88,245 // vaddps %ymm13,%ymm13,%ymm14 |
| .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14 |
| .byte 196,65,12,89,254 // vmulps %ymm14,%ymm14,%ymm15 |
| .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14 |
| .byte 196,65,20,92,248 // vsubps %ymm8,%ymm13,%ymm15 |
| .byte 196,65,4,89,246 // vmulps %ymm14,%ymm15,%ymm14 |
| .byte 196,65,28,89,253 // vmulps %ymm13,%ymm12,%ymm15 |
| .byte 196,65,4,88,246 // vaddps %ymm14,%ymm15,%ymm14 |
| .byte 196,65,124,82,253 // vrsqrtps %ymm13,%ymm15 |
| .byte 196,65,124,83,255 // vrcpps %ymm15,%ymm15 |
| .byte 196,65,4,92,253 // vsubps %ymm13,%ymm15,%ymm15 |
| .byte 197,84,88,221 // vaddps %ymm5,%ymm5,%ymm11 |
| .byte 196,65,36,88,219 // vaddps %ymm11,%ymm11,%ymm11 |
| .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11 |
| .byte 196,67,5,74,222,176 // vblendvps %ymm11,%ymm14,%ymm15,%ymm11 |
| .byte 197,108,88,242 // vaddps %ymm2,%ymm2,%ymm14 |
| .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13 |
| .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13 |
| .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| .byte 196,65,4,89,219 // vmulps %ymm11,%ymm15,%ymm11 |
| .byte 197,100,89,253 // vmulps %ymm5,%ymm3,%ymm15 |
| .byte 196,65,4,88,219 // vaddps %ymm11,%ymm15,%ymm11 |
| .byte 197,20,88,235 // vaddps %ymm3,%ymm13,%ymm13 |
| .byte 197,20,89,237 // vmulps %ymm5,%ymm13,%ymm13 |
| .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| .byte 196,67,37,74,237,224 // vblendvps %ymm14,%ymm13,%ymm11,%ymm13 |
| .byte 197,76,94,223 // vdivps %ymm7,%ymm6,%ymm11 |
| .byte 196,67,53,74,203,160 // vblendvps %ymm10,%ymm11,%ymm9,%ymm9 |
| .byte 196,65,52,88,209 // vaddps %ymm9,%ymm9,%ymm10 |
| .byte 196,65,44,88,210 // vaddps %ymm10,%ymm10,%ymm10 |
| .byte 196,65,44,89,218 // vmulps %ymm10,%ymm10,%ymm11 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 196,65,52,92,216 // vsubps %ymm8,%ymm9,%ymm11 |
| .byte 196,65,36,89,210 // vmulps %ymm10,%ymm11,%ymm10 |
| .byte 196,65,28,89,217 // vmulps %ymm9,%ymm12,%ymm11 |
| .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| .byte 196,65,124,82,217 // vrsqrtps %ymm9,%ymm11 |
| .byte 196,65,124,83,219 // vrcpps %ymm11,%ymm11 |
| .byte 196,65,36,92,217 // vsubps %ymm9,%ymm11,%ymm11 |
| .byte 197,76,88,230 // vaddps %ymm6,%ymm6,%ymm12 |
| .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| .byte 196,67,37,74,210,192 // vblendvps %ymm12,%ymm10,%ymm11,%ymm10 |
| .byte 197,124,16,116,36,200 // vmovups -0x38(%rsp),%ymm14 |
| .byte 196,65,12,88,222 // vaddps %ymm14,%ymm14,%ymm11 |
| .byte 197,36,92,227 // vsubps %ymm3,%ymm11,%ymm12 |
| .byte 196,65,60,92,201 // vsubps %ymm9,%ymm8,%ymm9 |
| .byte 196,65,28,89,201 // vmulps %ymm9,%ymm12,%ymm9 |
| .byte 197,28,89,231 // vmulps %ymm7,%ymm12,%ymm12 |
| .byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10 |
| .byte 197,100,89,230 // vmulps %ymm6,%ymm3,%ymm12 |
| .byte 196,65,28,88,210 // vaddps %ymm10,%ymm12,%ymm10 |
| .byte 197,52,88,203 // vaddps %ymm3,%ymm9,%ymm9 |
| .byte 197,52,89,206 // vmulps %ymm6,%ymm9,%ymm9 |
| .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| .byte 196,67,45,74,201,176 // vblendvps %ymm11,%ymm9,%ymm10,%ymm9 |
| .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| .byte 197,244,88,192 // vaddps %ymm0,%ymm1,%ymm0 |
| .byte 197,172,89,202 // vmulps %ymm2,%ymm10,%ymm1 |
| .byte 197,188,89,213 // vmulps %ymm5,%ymm8,%ymm2 |
| .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1 |
| .byte 196,193,44,89,214 // vmulps %ymm14,%ymm10,%ymm2 |
| .byte 197,60,89,214 // vmulps %ymm6,%ymm8,%ymm10 |
| .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 |
| .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_0_avx |
| .globl _sk_clamp_0_avx |
| FUNCTION(_sk_clamp_0_avx) |
| _sk_clamp_0_avx: |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_1_avx |
| .globl _sk_clamp_1_avx |
| FUNCTION(_sk_clamp_1_avx) |
| _sk_clamp_1_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 |
| .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_a_avx |
| .globl _sk_clamp_a_avx |
| FUNCTION(_sk_clamp_a_avx) |
| _sk_clamp_a_avx: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 |
| .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 |
| .byte 197,236,93,211 // vminps %ymm3,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_set_rgb_avx |
| .globl _sk_set_rgb_avx |
| FUNCTION(_sk_set_rgb_avx) |
| _sk_set_rgb_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_rb_avx |
| .globl _sk_swap_rb_avx |
| FUNCTION(_sk_swap_rb_avx) |
| _sk_swap_rb_avx: |
| .byte 197,124,40,192 // vmovaps %ymm0,%ymm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,194 // vmovaps %ymm2,%ymm0 |
| .byte 197,124,41,194 // vmovaps %ymm8,%ymm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_avx |
| .globl _sk_swap_avx |
| FUNCTION(_sk_swap_avx) |
| _sk_swap_avx: |
| .byte 197,124,40,195 // vmovaps %ymm3,%ymm8 |
| .byte 197,124,40,202 // vmovaps %ymm2,%ymm9 |
| .byte 197,124,40,209 // vmovaps %ymm1,%ymm10 |
| .byte 197,124,40,216 // vmovaps %ymm0,%ymm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| .byte 197,124,41,220 // vmovaps %ymm11,%ymm4 |
| .byte 197,124,41,213 // vmovaps %ymm10,%ymm5 |
| .byte 197,124,41,206 // vmovaps %ymm9,%ymm6 |
| .byte 197,124,41,199 // vmovaps %ymm8,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_src_dst_avx |
| .globl _sk_move_src_dst_avx |
| FUNCTION(_sk_move_src_dst_avx) |
| _sk_move_src_dst_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,224 // vmovaps %ymm0,%ymm4 |
| .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 |
| .byte 197,252,40,242 // vmovaps %ymm2,%ymm6 |
| .byte 197,252,40,251 // vmovaps %ymm3,%ymm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_dst_src_avx |
| .globl _sk_move_dst_src_avx |
| FUNCTION(_sk_move_dst_src_avx) |
| _sk_move_dst_src_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_premul_avx |
| .globl _sk_premul_avx |
| FUNCTION(_sk_premul_avx) |
| _sk_premul_avx: |
| .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 |
| .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_unpremul_avx |
| .globl _sk_unpremul_avx |
| FUNCTION(_sk_unpremul_avx) |
| _sk_unpremul_avx: |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 |
| .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_srgb_avx |
| .globl _sk_from_srgb_avx |
| FUNCTION(_sk_from_srgb_avx) |
| _sk_from_srgb_avx: |
| .byte 184,145,131,158,61 // mov $0x3d9e8391,%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 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 |
| .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| .byte 196,67,121,4,228,0 // vpermilps $0x0,%xmm12,%xmm12 |
| .byte 196,67,29,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm12,%ymm12 |
| .byte 197,36,89,232 // vmulps %ymm0,%ymm11,%ymm13 |
| .byte 196,65,20,88,236 // vaddps %ymm12,%ymm13,%ymm13 |
| .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14 |
| .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14 |
| .byte 196,65,44,89,213 // vmulps %ymm13,%ymm10,%ymm10 |
| .byte 196,65,12,88,210 // vaddps %ymm10,%ymm14,%ymm10 |
| .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| .byte 197,121,110,232 // vmovd %eax,%xmm13 |
| .byte 196,67,121,4,237,0 // vpermilps $0x0,%xmm13,%xmm13 |
| .byte 196,67,21,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm13,%ymm13 |
| .byte 196,193,124,194,197,1 // vcmpltps %ymm13,%ymm0,%ymm0 |
| .byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0 |
| .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9 |
| .byte 197,116,89,209 // vmulps %ymm1,%ymm1,%ymm10 |
| .byte 197,36,89,249 // vmulps %ymm1,%ymm11,%ymm15 |
| .byte 196,65,28,88,255 // vaddps %ymm15,%ymm12,%ymm15 |
| .byte 196,65,44,89,215 // vmulps %ymm15,%ymm10,%ymm10 |
| .byte 196,65,12,88,210 // vaddps %ymm10,%ymm14,%ymm10 |
| .byte 196,193,116,194,205,1 // vcmpltps %ymm13,%ymm1,%ymm1 |
| .byte 196,195,45,74,201,16 // vblendvps %ymm1,%ymm9,%ymm10,%ymm1 |
| .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| .byte 197,108,89,202 // vmulps %ymm2,%ymm2,%ymm9 |
| .byte 197,36,89,210 // vmulps %ymm2,%ymm11,%ymm10 |
| .byte 196,65,28,88,210 // vaddps %ymm10,%ymm12,%ymm10 |
| .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 |
| .byte 196,65,12,88,201 // vaddps %ymm9,%ymm14,%ymm9 |
| .byte 196,193,108,194,213,1 // vcmpltps %ymm13,%ymm2,%ymm2 |
| .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_srgb_avx |
| .globl _sk_to_srgb_avx |
| FUNCTION(_sk_to_srgb_avx) |
| _sk_to_srgb_avx: |
| .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| .byte 196,65,124,83,232 // vrcpps %ymm8,%ymm13 |
| .byte 196,65,124,82,240 // vrsqrtps %ymm8,%ymm14 |
| .byte 184,41,92,71,65 // mov $0x41475c29,%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 197,60,89,224 // vmulps %ymm0,%ymm8,%ymm12 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| .byte 197,121,110,248 // vmovd %eax,%xmm15 |
| .byte 196,67,121,4,255,0 // vpermilps $0x0,%xmm15,%xmm15 |
| .byte 196,67,5,24,255,1 // vinsertf128 $0x1,%xmm15,%ymm15,%ymm15 |
| .byte 196,65,20,89,235 // vmulps %ymm11,%ymm13,%ymm13 |
| .byte 196,65,20,88,239 // vaddps %ymm15,%ymm13,%ymm13 |
| .byte 196,65,12,89,242 // vmulps %ymm10,%ymm14,%ymm14 |
| .byte 196,65,12,88,237 // vaddps %ymm13,%ymm14,%ymm13 |
| .byte 196,65,52,93,237 // vminps %ymm13,%ymm9,%ymm13 |
| .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14 |
| .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14 |
| .byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0 |
| .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0 |
| .byte 197,124,82,225 // vrsqrtps %ymm1,%ymm12 |
| .byte 196,65,124,83,236 // vrcpps %ymm12,%ymm13 |
| .byte 196,65,124,82,228 // vrsqrtps %ymm12,%ymm12 |
| .byte 196,65,36,89,237 // vmulps %ymm13,%ymm11,%ymm13 |
| .byte 196,65,4,88,237 // vaddps %ymm13,%ymm15,%ymm13 |
| .byte 196,65,44,89,228 // vmulps %ymm12,%ymm10,%ymm12 |
| .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| .byte 197,60,89,233 // vmulps %ymm1,%ymm8,%ymm13 |
| .byte 196,65,52,93,228 // vminps %ymm12,%ymm9,%ymm12 |
| .byte 196,193,116,194,206,1 // vcmpltps %ymm14,%ymm1,%ymm1 |
| .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1 |
| .byte 197,124,82,226 // vrsqrtps %ymm2,%ymm12 |
| .byte 196,65,124,83,236 // vrcpps %ymm12,%ymm13 |
| .byte 196,65,36,89,221 // vmulps %ymm13,%ymm11,%ymm11 |
| .byte 196,65,4,88,219 // vaddps %ymm11,%ymm15,%ymm11 |
| .byte 196,65,124,82,228 // vrsqrtps %ymm12,%ymm12 |
| .byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 196,65,52,93,202 // vminps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| .byte 196,193,108,194,214,1 // vcmpltps %ymm14,%ymm2,%ymm2 |
| .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_2dot2_avx |
| .globl _sk_from_2dot2_avx |
| FUNCTION(_sk_from_2dot2_avx) |
| _sk_from_2dot2_avx: |
| .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,200 // vrsqrtps %ymm8,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 197,252,89,192 // vmulps %ymm0,%ymm0,%ymm0 |
| .byte 196,65,60,89,208 // vmulps %ymm8,%ymm8,%ymm10 |
| .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 196,65,124,82,210 // vrsqrtps %ymm10,%ymm10 |
| .byte 197,244,89,201 // vmulps %ymm1,%ymm1,%ymm1 |
| .byte 196,65,52,89,217 // vmulps %ymm9,%ymm9,%ymm11 |
| .byte 196,65,52,89,203 // vmulps %ymm11,%ymm9,%ymm9 |
| .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| .byte 197,124,82,202 // vrsqrtps %ymm2,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 196,65,124,82,210 // vrsqrtps %ymm10,%ymm10 |
| .byte 197,236,89,210 // vmulps %ymm2,%ymm2,%ymm2 |
| .byte 196,65,52,89,217 // vmulps %ymm9,%ymm9,%ymm11 |
| .byte 196,65,52,89,203 // vmulps %ymm11,%ymm9,%ymm9 |
| .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2 |
| .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_2dot2_avx |
| .globl _sk_to_2dot2_avx |
| FUNCTION(_sk_to_2dot2_avx) |
| _sk_to_2dot2_avx: |
| .byte 197,252,82,192 // vrsqrtps %ymm0,%ymm0 |
| .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,192 // vrsqrtps %ymm8,%ymm8 |
| .byte 196,65,124,82,200 // vrsqrtps %ymm8,%ymm9 |
| .byte 197,252,83,192 // vrcpps %ymm0,%ymm0 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 196,65,124,83,193 // vrcpps %ymm9,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| .byte 197,252,82,201 // vrsqrtps %ymm1,%ymm1 |
| .byte 197,124,82,201 // vrsqrtps %ymm1,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 197,252,83,201 // vrcpps %ymm1,%ymm1 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 196,65,124,83,202 // vrcpps %ymm10,%ymm9 |
| .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 |
| .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| .byte 197,252,82,210 // vrsqrtps %ymm2,%ymm2 |
| .byte 197,124,82,202 // vrsqrtps %ymm2,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,201 // vrsqrtps %ymm9,%ymm9 |
| .byte 196,65,124,82,209 // vrsqrtps %ymm9,%ymm10 |
| .byte 197,252,83,210 // vrcpps %ymm2,%ymm2 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 196,65,124,83,202 // vrcpps %ymm10,%ymm9 |
| .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2 |
| .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_rgb_to_hsl_avx |
| .globl _sk_rgb_to_hsl_avx |
| FUNCTION(_sk_rgb_to_hsl_avx) |
| _sk_rgb_to_hsl_avx: |
| .byte 197,124,95,193 // vmaxps %ymm1,%ymm0,%ymm8 |
| .byte 197,60,95,194 // vmaxps %ymm2,%ymm8,%ymm8 |
| .byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9 |
| .byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9 |
| .byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11 |
| .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d |
| .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12 |
| .byte 196,65,28,89,227 // vmulps %ymm11,%ymm12,%ymm12 |
| .byte 65,185,0,0,192,64 // mov $0x40c00000,%r9d |
| .byte 197,108,92,232 // vsubps %ymm0,%ymm2,%ymm13 |
| .byte 196,65,20,89,235 // vmulps %ymm11,%ymm13,%ymm13 |
| .byte 65,186,0,0,0,64 // mov $0x40000000,%r10d |
| .byte 197,124,92,241 // vsubps %ymm1,%ymm0,%ymm14 |
| .byte 196,65,12,89,219 // vmulps %ymm11,%ymm14,%ymm11 |
| .byte 184,0,0,128,64 // mov $0x40800000,%eax |
| .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14 |
| .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14 |
| .byte 196,65,36,88,222 // vaddps %ymm14,%ymm11,%ymm11 |
| .byte 196,65,121,110,242 // vmovd %r10d,%xmm14 |
| .byte 197,244,194,210,1 // vcmpltps %ymm2,%ymm1,%ymm2 |
| .byte 197,188,194,201,0 // vcmpeqps %ymm1,%ymm8,%ymm1 |
| .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14 |
| .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14 |
| .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 |
| .byte 196,67,37,74,221,16 // vblendvps %ymm1,%ymm13,%ymm11,%ymm11 |
| .byte 196,193,121,110,201 // vmovd %r9d,%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 196,65,20,87,237 // vxorps %ymm13,%ymm13,%ymm13 |
| .byte 196,227,21,74,201,32 // vblendvps %ymm2,%ymm1,%ymm13,%ymm1 |
| .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0 |
| .byte 196,227,37,74,193,0 // vblendvps %ymm0,%ymm1,%ymm11,%ymm0 |
| .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1 |
| .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| .byte 196,99,109,24,218,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm11 |
| .byte 196,193,116,89,211 // vmulps %ymm11,%ymm1,%ymm2 |
| .byte 197,36,194,218,1 // vcmpltps %ymm2,%ymm11,%ymm11 |
| .byte 196,65,12,92,224 // vsubps %ymm8,%ymm14,%ymm12 |
| .byte 196,65,28,92,225 // vsubps %ymm9,%ymm12,%ymm12 |
| .byte 196,195,117,74,204,176 // vblendvps %ymm11,%ymm12,%ymm1,%ymm1 |
| .byte 196,65,60,194,193,0 // vcmpeqps %ymm9,%ymm8,%ymm8 |
| .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1 |
| .byte 196,67,125,74,205,128 // vblendvps %ymm8,%ymm13,%ymm0,%ymm9 |
| .byte 196,195,117,74,205,128 // vblendvps %ymm8,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hsl_to_rgb_avx |
| .globl _sk_hsl_to_rgb_avx |
| FUNCTION(_sk_hsl_to_rgb_avx) |
| _sk_hsl_to_rgb_avx: |
| .byte 72,131,236,56 // sub $0x38,%rsp |
| .byte 197,252,17,60,36 // vmovups %ymm7,(%rsp) |
| .byte 197,252,17,116,36,224 // vmovups %ymm6,-0x20(%rsp) |
| .byte 197,252,17,108,36,192 // vmovups %ymm5,-0x40(%rsp) |
| .byte 197,252,17,100,36,160 // vmovups %ymm4,-0x60(%rsp) |
| .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp) |
| .byte 197,252,40,226 // vmovaps %ymm2,%ymm4 |
| .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 |
| .byte 197,252,40,216 // vmovaps %ymm0,%ymm3 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm8 |
| .byte 196,193,92,194,192,1 // vcmpltps %ymm8,%ymm4,%ymm0 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| .byte 196,99,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm10 |
| .byte 197,172,88,213 // vaddps %ymm5,%ymm10,%ymm2 |
| .byte 197,236,89,212 // vmulps %ymm4,%ymm2,%ymm2 |
| .byte 197,84,88,204 // vaddps %ymm4,%ymm5,%ymm9 |
| .byte 197,84,89,220 // vmulps %ymm4,%ymm5,%ymm11 |
| .byte 196,65,52,92,203 // vsubps %ymm11,%ymm9,%ymm9 |
| .byte 196,99,53,74,202,0 // vblendvps %ymm0,%ymm2,%ymm9,%ymm9 |
| .byte 65,184,0,0,0,64 // mov $0x40000000,%r8d |
| .byte 184,171,170,170,62 // mov $0x3eaaaaab,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 |
| .byte 197,28,88,251 // vaddps %ymm3,%ymm12,%ymm15 |
| .byte 184,0,0,0,0 // mov $0x0,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 |
| .byte 196,193,44,194,199,1 // vcmpltps %ymm15,%ymm10,%ymm0 |
| .byte 196,193,4,92,210 // vsubps %ymm10,%ymm15,%ymm2 |
| .byte 196,227,5,74,194,0 // vblendvps %ymm0,%ymm2,%ymm15,%ymm0 |
| .byte 196,193,4,194,213,1 // vcmpltps %ymm13,%ymm15,%ymm2 |
| .byte 196,65,44,88,223 // vaddps %ymm15,%ymm10,%ymm11 |
| .byte 196,195,125,74,203,32 // vblendvps %ymm2,%ymm11,%ymm0,%ymm1 |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| .byte 196,65,124,92,217 // vsubps %ymm9,%ymm0,%ymm11 |
| .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d |
| .byte 184,0,0,192,64 // mov $0x40c00000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 196,193,52,92,211 // vsubps %ymm11,%ymm9,%ymm2 |
| .byte 197,108,89,240 // vmulps %ymm0,%ymm2,%ymm14 |
| .byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2 |
| .byte 197,140,89,210 // vmulps %ymm2,%ymm14,%ymm2 |
| .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 |
| .byte 197,244,194,248,1 // vcmpltps %ymm0,%ymm1,%ymm7 |
| .byte 196,227,37,74,210,112 // vblendvps %ymm7,%ymm2,%ymm11,%ymm2 |
| .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 |
| .byte 196,195,109,74,249,112 // vblendvps %ymm7,%ymm9,%ymm2,%ymm7 |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1 |
| .byte 196,65,4,89,254 // vmulps %ymm14,%ymm15,%ymm15 |
| .byte 196,65,36,88,255 // vaddps %ymm15,%ymm11,%ymm15 |
| .byte 196,67,69,74,255,16 // vblendvps %ymm1,%ymm15,%ymm7,%ymm15 |
| .byte 197,172,194,203,1 // vcmpltps %ymm3,%ymm10,%ymm1 |
| .byte 196,193,100,92,250 // vsubps %ymm10,%ymm3,%ymm7 |
| .byte 196,227,101,74,207,16 // vblendvps %ymm1,%ymm7,%ymm3,%ymm1 |
| .byte 196,193,100,194,253,1 // vcmpltps %ymm13,%ymm3,%ymm7 |
| .byte 197,172,88,243 // vaddps %ymm3,%ymm10,%ymm6 |
| .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1 |
| .byte 197,252,92,241 // vsubps %ymm1,%ymm0,%ymm6 |
| .byte 197,140,89,246 // vmulps %ymm6,%ymm14,%ymm6 |
| .byte 197,164,88,246 // vaddps %ymm6,%ymm11,%ymm6 |
| .byte 197,244,194,248,1 // vcmpltps %ymm0,%ymm1,%ymm7 |
| .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6 |
| .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7 |
| .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6 |
| .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1 |
| .byte 197,140,89,251 // vmulps %ymm3,%ymm14,%ymm7 |
| .byte 197,164,88,255 // vaddps %ymm7,%ymm11,%ymm7 |
| .byte 196,227,77,74,207,16 // vblendvps %ymm1,%ymm7,%ymm6,%ymm1 |
| .byte 196,193,100,92,220 // vsubps %ymm12,%ymm3,%ymm3 |
| .byte 197,172,194,243,1 // vcmpltps %ymm3,%ymm10,%ymm6 |
| .byte 196,193,100,92,250 // vsubps %ymm10,%ymm3,%ymm7 |
| .byte 196,227,101,74,247,96 // vblendvps %ymm6,%ymm7,%ymm3,%ymm6 |
| .byte 196,193,100,194,253,1 // vcmpltps %ymm13,%ymm3,%ymm7 |
| .byte 197,44,88,211 // vaddps %ymm3,%ymm10,%ymm10 |
| .byte 196,195,77,74,242,112 // vblendvps %ymm7,%ymm10,%ymm6,%ymm6 |
| .byte 197,204,194,248,1 // vcmpltps %ymm0,%ymm6,%ymm7 |
| .byte 197,252,92,198 // vsubps %ymm6,%ymm0,%ymm0 |
| .byte 197,140,89,192 // vmulps %ymm0,%ymm14,%ymm0 |
| .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 |
| .byte 196,227,37,74,192,112 // vblendvps %ymm7,%ymm0,%ymm11,%ymm0 |
| .byte 196,193,76,194,248,1 // vcmpltps %ymm8,%ymm6,%ymm7 |
| .byte 196,195,125,74,193,112 // vblendvps %ymm7,%ymm9,%ymm0,%ymm0 |
| .byte 197,204,194,210,1 // vcmpltps %ymm2,%ymm6,%ymm2 |
| .byte 196,193,100,89,222 // vmulps %ymm14,%ymm3,%ymm3 |
| .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 |
| .byte 196,227,125,74,211,32 // vblendvps %ymm2,%ymm3,%ymm0,%ymm2 |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,212,194,216,0 // vcmpeqps %ymm0,%ymm5,%ymm3 |
| .byte 196,227,5,74,196,48 // vblendvps %ymm3,%ymm4,%ymm15,%ymm0 |
| .byte 196,227,117,74,204,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm1 |
| .byte 196,227,109,74,212,48 // vblendvps %ymm3,%ymm4,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,16,92,36,128 // vmovups -0x80(%rsp),%ymm3 |
| .byte 197,252,16,100,36,160 // vmovups -0x60(%rsp),%ymm4 |
| .byte 197,252,16,108,36,192 // vmovups -0x40(%rsp),%ymm5 |
| .byte 197,252,16,116,36,224 // vmovups -0x20(%rsp),%ymm6 |
| .byte 197,252,16,60,36 // vmovups (%rsp),%ymm7 |
| .byte 72,131,196,56 // add $0x38,%rsp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_1_float_avx |
| .globl _sk_scale_1_float_avx |
| FUNCTION(_sk_scale_1_float_avx) |
| _sk_scale_1_float_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_u8_avx |
| .globl _sk_scale_u8_avx |
| FUNCTION(_sk_scale_u8_avx) |
| _sk_scale_u8_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 117,80 // jne 1456 <_sk_scale_u8_avx+0x60> |
| .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 |
| .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 |
| .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 |
| .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,193 // mov %r8,%rcx |
| .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 145e <_sk_scale_u8_avx+0x68> |
| .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| .byte 235,143 // jmp 140a <_sk_scale_u8_avx+0x14> |
| |
| HIDDEN _sk_lerp_1_float_avx |
| .globl _sk_lerp_1_float_avx |
| FUNCTION(_sk_lerp_1_float_avx) |
| _sk_lerp_1_float_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_u8_avx |
| .globl _sk_lerp_u8_avx |
| FUNCTION(_sk_lerp_u8_avx) |
| _sk_lerp_u8_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 117,116 // jne 153e <_sk_lerp_u8_avx+0x84> |
| .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 |
| .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 |
| .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 |
| .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,193 // mov %r8,%rcx |
| .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 1546 <_sk_lerp_u8_avx+0x8c> |
| .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| .byte 233,104,255,255,255 // jmpq 14ce <_sk_lerp_u8_avx+0x14> |
| |
| HIDDEN _sk_lerp_565_avx |
| .globl _sk_lerp_565_avx |
| FUNCTION(_sk_lerp_565_avx) |
| _sk_lerp_565_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,16 // mov (%rax),%r10 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,250,0,0,0 // jne 166e <_sk_lerp_565_avx+0x108> |
| .byte 196,65,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm8 |
| .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| .byte 197,185,105,219 // vpunpckhwd %xmm3,%xmm8,%xmm3 |
| .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 |
| .byte 196,99,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm8 |
| .byte 184,0,248,0,0 // mov $0xf800,%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 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 |
| .byte 197,124,91,203 // vcvtdq2ps %ymm3,%ymm9 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| .byte 184,224,7,0,0 // mov $0x7e0,%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 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 |
| .byte 197,124,91,211 // vcvtdq2ps %ymm3,%ymm10 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%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 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10 |
| .byte 184,31,0,0,0 // mov $0x1f,%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 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 |
| .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%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 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 |
| .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 |
| .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 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 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8 |
| .byte 65,254,200 // dec %r8b |
| .byte 65,128,248,6 // cmp $0x6,%r8b |
| .byte 15,135,243,254,255,255 // ja 157a <_sk_lerp_565_avx+0x14> |
| .byte 69,15,182,192 // movzbl %r8b,%r8d |
| .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 16dc <_sk_lerp_565_avx+0x176> |
| .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| .byte 76,1,200 // add %r9,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| .byte 196,65,97,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm8 |
| .byte 196,65,57,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm8,%xmm8 |
| .byte 196,65,57,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm8,%xmm8 |
| .byte 196,65,57,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm8,%xmm8 |
| .byte 196,65,57,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 |
| .byte 196,65,57,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 |
| .byte 196,65,57,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 |
| .byte 233,159,254,255,255 // jmpq 157a <_sk_lerp_565_avx+0x14> |
| .byte 144 // nop |
| .byte 243,255 // repz (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 235,255 // jmp 16e1 <_sk_lerp_565_avx+0x17b> |
| .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_load_tables_avx |
| .globl _sk_load_tables_avx |
| FUNCTION(_sk_load_tables_avx) |
| _sk_load_tables_avx: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,56,2,0,0 // jne 1948 <_sk_load_tables_avx+0x250> |
| .byte 196,65,124,16,4,184 // vmovups (%r8,%rdi,4),%ymm8 |
| .byte 187,255,0,0,0 // mov $0xff,%ebx |
| .byte 197,249,110,195 // vmovd %ebx,%xmm0 |
| .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9 |
| .byte 196,193,52,84,192 // vandps %ymm8,%ymm9,%ymm0 |
| .byte 196,193,249,126,193 // vmovq %xmm0,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10 |
| .byte 69,137,214 // mov %r10d,%r14d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| .byte 196,193,249,126,196 // vmovq %xmm0,%r12 |
| .byte 69,137,231 // mov %r12d,%r15d |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 72,139,104,8 // mov 0x8(%rax),%rbp |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0 |
| .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0 |
| .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1 |
| .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0 |
| .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1 |
| .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0 |
| .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1 |
| .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 |
| .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3 |
| .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1 |
| .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3 |
| .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1 |
| .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 |
| .byte 196,67,125,25,194,1 // vextractf128 $0x1,%ymm8,%xmm10 |
| .byte 196,193,105,114,210,8 // vpsrld $0x8,%xmm10,%xmm2 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 197,180,84,201 // vandps %ymm1,%ymm9,%ymm1 |
| .byte 196,193,249,126,201 // vmovq %xmm1,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10 |
| .byte 69,137,214 // mov %r10d,%r14d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,225,249,126,205 // vmovq %xmm1,%rbp |
| .byte 65,137,239 // mov %ebp,%r15d |
| .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx |
| .byte 65,137,220 // mov %ebx,%r12d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1 |
| .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1 |
| .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2 |
| .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1 |
| .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2 |
| .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1 |
| .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2 |
| .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2 |
| .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3 |
| .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3 |
| .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 196,193,105,114,208,16 // vpsrld $0x10,%xmm8,%xmm2 |
| .byte 196,193,97,114,210,16 // vpsrld $0x10,%xmm10,%xmm3 |
| .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 |
| .byte 197,180,84,210 // vandps %ymm2,%ymm9,%ymm2 |
| .byte 196,193,249,126,208 // vmovq %xmm2,%r8 |
| .byte 69,137,194 // mov %r8d,%r10d |
| .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 73,193,232,32 // shr $0x20,%r8 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,225,249,126,213 // vmovq %xmm2,%rbp |
| .byte 65,137,238 // mov %ebp,%r14d |
| .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx |
| .byte 65,137,223 // mov %ebx,%r15d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2 |
| .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2 |
| .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3 |
| .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3 |
| .byte 196,99,105,33,203,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm9 |
| .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3 |
| .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 |
| .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2 |
| .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2 |
| .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3 |
| .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| .byte 196,195,109,24,209,1 // vinsertf128 $0x1,%xmm9,%ymm2,%ymm2 |
| .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8 |
| .byte 196,193,97,114,210,24 // vpsrld $0x18,%xmm10,%xmm3 |
| .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 |
| .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%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 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| .byte 137,203 // mov %ecx,%ebx |
| .byte 128,227,7 // and $0x7,%bl |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 254,203 // dec %bl |
| .byte 128,251,6 // cmp $0x6,%bl |
| .byte 15,135,185,253,255,255 // ja 1716 <_sk_load_tables_avx+0x1e> |
| .byte 15,182,219 // movzbl %bl,%ebx |
| .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 19f0 <_sk_load_tables_avx+0x2f8> |
| .byte 73,99,28,153 // movslq (%r9,%rbx,4),%rbx |
| .byte 76,1,203 // add %r9,%rbx |
| .byte 255,227 // jmpq *%rbx |
| .byte 196,193,121,110,68,184,24 // vmovd 0x18(%r8,%rdi,4),%xmm0 |
| .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 196,99,117,12,192,64 // vblendps $0x40,%ymm0,%ymm1,%ymm8 |
| .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| .byte 196,195,121,34,68,184,20,1 // vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0 |
| .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 |
| .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| .byte 196,195,121,34,68,184,16,0 // vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0 |
| .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 |
| .byte 196,195,57,34,68,184,12,3 // vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm8,%xmm0 |
| .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| .byte 196,195,57,34,68,184,8,2 // vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm8,%xmm0 |
| .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| .byte 196,195,57,34,68,184,4,1 // vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm8,%xmm0 |
| .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| .byte 196,195,57,34,4,184,0 // vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0 |
| .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| .byte 233,38,253,255,255 // jmpq 1716 <_sk_load_tables_avx+0x1e> |
| .byte 238 // out %al,(%dx) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,224 // jmpq *%rax |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,210 // callq *%rdx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,196 // inc %esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,176,255,255,255,156 // pushq -0x63000001(%rax) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| .byte 128,255,255 // cmp $0xff,%bh |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_load_tables_u16_be_avx |
| .globl _sk_load_tables_u16_be_avx |
| FUNCTION(_sk_load_tables_u16_be_avx) |
| _sk_load_tables_u16_be_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,165,2,0,0 // jne 1cc7 <_sk_load_tables_u16_be_avx+0x2bb> |
| .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8 |
| .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2 |
| .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3 |
| .byte 196,1,122,111,76,72,48 // vmovdqu 0x30(%r8,%r9,2),%xmm9 |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| .byte 197,121,97,218 // vpunpcklwd %xmm2,%xmm0,%xmm11 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,113,105,211 // vpunpckhwd %xmm3,%xmm1,%xmm10 |
| .byte 197,161,108,194 // vpunpcklqdq %xmm2,%xmm11,%xmm0 |
| .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8 |
| .byte 196,193,121,105,200 // vpunpckhwd %xmm8,%xmm0,%xmm1 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| .byte 187,255,0,0,0 // mov $0xff,%ebx |
| .byte 197,249,110,203 // vmovd %ebx,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 |
| .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0 |
| .byte 196,193,249,126,193 // vmovq %xmm0,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10 |
| .byte 69,137,214 // mov %r10d,%r14d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| .byte 196,193,249,126,196 // vmovq %xmm0,%r12 |
| .byte 69,137,231 // mov %r12d,%r15d |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 72,139,104,8 // mov 0x8(%rax),%rbp |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0 |
| .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0 |
| .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1 |
| .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0 |
| .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1 |
| .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0 |
| .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1 |
| .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 |
| .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3 |
| .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1 |
| .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3 |
| .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1 |
| .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| .byte 197,161,109,202 // vpunpckhqdq %xmm2,%xmm11,%xmm1 |
| .byte 196,193,113,105,208 // vpunpckhwd %xmm8,%xmm1,%xmm2 |
| .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1 |
| .byte 196,193,249,126,201 // vmovq %xmm1,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10 |
| .byte 69,137,214 // mov %r10d,%r14d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,225,249,126,205 // vmovq %xmm1,%rbp |
| .byte 65,137,239 // mov %ebp,%r15d |
| .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx |
| .byte 65,137,220 // mov %ebx,%r12d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1 |
| .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1 |
| .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2 |
| .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1 |
| .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2 |
| .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1 |
| .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2 |
| .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2 |
| .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3 |
| .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3 |
| .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 196,193,49,108,210 // vpunpcklqdq %xmm10,%xmm9,%xmm2 |
| .byte 196,193,105,105,216 // vpunpckhwd %xmm8,%xmm2,%xmm3 |
| .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 |
| .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2 |
| .byte 196,193,249,126,208 // vmovq %xmm2,%r8 |
| .byte 69,137,194 // mov %r8d,%r10d |
| .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 73,193,232,32 // shr $0x20,%r8 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,225,249,126,213 // vmovq %xmm2,%rbp |
| .byte 65,137,238 // mov %ebp,%r14d |
| .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx |
| .byte 65,137,223 // mov %ebx,%r15d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2 |
| .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2 |
| .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3 |
| .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3 |
| .byte 196,99,105,33,219,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm11 |
| .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3 |
| .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 |
| .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2 |
| .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2 |
| .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3 |
| .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| .byte 196,195,109,24,211,1 // vinsertf128 $0x1,%xmm11,%ymm2,%ymm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| .byte 196,99,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm11 |
| .byte 196,193,49,109,218 // vpunpckhqdq %xmm10,%xmm9,%xmm3 |
| .byte 197,177,113,243,8 // vpsllw $0x8,%xmm3,%xmm9 |
| .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3 |
| .byte 197,177,235,219 // vpor %xmm3,%xmm9,%xmm3 |
| .byte 196,65,97,105,192 // vpunpckhwd %xmm8,%xmm3,%xmm8 |
| .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3 |
| .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,219 // vmulps %ymm11,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8 |
| .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 116,85 // je 1d2d <_sk_load_tables_u16_be_avx+0x321> |
| .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,72 // jb 1d2d <_sk_load_tables_u16_be_avx+0x321> |
| .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 116,72 // je 1d3a <_sk_load_tables_u16_be_avx+0x32e> |
| .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,59 // jb 1d3a <_sk_load_tables_u16_be_avx+0x32e> |
| .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 15,132,45,253,255,255 // je 1a3d <_sk_load_tables_u16_be_avx+0x31> |
| .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 15,130,28,253,255,255 // jb 1a3d <_sk_load_tables_u16_be_avx+0x31> |
| .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9 |
| .byte 233,16,253,255,255 // jmpq 1a3d <_sk_load_tables_u16_be_avx+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| .byte 233,3,253,255,255 // jmpq 1a3d <_sk_load_tables_u16_be_avx+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 233,250,252,255,255 // jmpq 1a3d <_sk_load_tables_u16_be_avx+0x31> |
| |
| HIDDEN _sk_load_tables_rgb_u16_be_avx |
| .globl _sk_load_tables_rgb_u16_be_avx |
| FUNCTION(_sk_load_tables_rgb_u16_be_avx) |
| _sk_load_tables_rgb_u16_be_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,146,2,0,0 // jne 1fe7 <_sk_load_tables_rgb_u16_be_avx+0x2a4> |
| .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0 |
| .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2 |
| .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1 |
| .byte 196,129,122,111,92,72,32 // vmovdqu 0x20(%r8,%r9,2),%xmm3 |
| .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3 |
| .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8 |
| .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9 |
| .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11 |
| .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10 |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0 |
| .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2 |
| .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1 |
| .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3 |
| .byte 197,121,97,218 // vpunpcklwd %xmm2,%xmm0,%xmm11 |
| .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,113,105,211 // vpunpckhwd %xmm3,%xmm1,%xmm10 |
| .byte 197,161,108,194 // vpunpcklqdq %xmm2,%xmm11,%xmm0 |
| .byte 196,65,25,239,228 // vpxor %xmm12,%xmm12,%xmm12 |
| .byte 196,193,121,105,204 // vpunpckhwd %xmm12,%xmm0,%xmm1 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| .byte 187,255,0,0,0 // mov $0xff,%ebx |
| .byte 197,249,110,203 // vmovd %ebx,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9 |
| .byte 196,193,124,84,193 // vandps %ymm9,%ymm0,%ymm0 |
| .byte 196,193,249,126,193 // vmovq %xmm0,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10 |
| .byte 69,137,214 // mov %r10d,%r14d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| .byte 196,193,249,126,196 // vmovq %xmm0,%r12 |
| .byte 69,137,231 // mov %r12d,%r15d |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 72,139,104,8 // mov 0x8(%rax),%rbp |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0 |
| .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0 |
| .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1 |
| .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0 |
| .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1 |
| .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0 |
| .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1 |
| .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 |
| .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3 |
| .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1 |
| .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3 |
| .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1 |
| .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| .byte 197,161,109,202 // vpunpckhqdq %xmm2,%xmm11,%xmm1 |
| .byte 196,193,113,105,212 // vpunpckhwd %xmm12,%xmm1,%xmm2 |
| .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1 |
| .byte 196,193,249,126,201 // vmovq %xmm1,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10 |
| .byte 69,137,214 // mov %r10d,%r14d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,225,249,126,205 // vmovq %xmm1,%rbp |
| .byte 65,137,239 // mov %ebp,%r15d |
| .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx |
| .byte 65,137,220 // mov %ebx,%r12d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1 |
| .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1 |
| .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2 |
| .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1 |
| .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2 |
| .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1 |
| .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2 |
| .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2 |
| .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3 |
| .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3 |
| .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 196,193,57,108,210 // vpunpcklqdq %xmm10,%xmm8,%xmm2 |
| .byte 196,193,105,105,220 // vpunpckhwd %xmm12,%xmm2,%xmm3 |
| .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 |
| .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2 |
| .byte 196,193,249,126,208 // vmovq %xmm2,%r8 |
| .byte 69,137,194 // mov %r8d,%r10d |
| .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9 |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 73,193,232,32 // shr $0x20,%r8 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,225,249,126,213 // vmovq %xmm2,%rbp |
| .byte 65,137,238 // mov %ebp,%r14d |
| .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx |
| .byte 65,137,223 // mov %ebx,%r15d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2 |
| .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2 |
| .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3 |
| .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3 |
| .byte 196,99,105,33,195,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm8 |
| .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3 |
| .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 |
| .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2 |
| .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2 |
| .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3 |
| .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| .byte 196,195,109,24,208,1 // vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0 |
| .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 117,5 // jne 2000 <_sk_load_tables_rgb_u16_be_avx+0x2bd> |
| .byte 233,137,253,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1 |
| .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,26 // jb 202f <_sk_load_tables_rgb_u16_be_avx+0x2ec> |
| .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1 |
| .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 117,10 // jne 2034 <_sk_load_tables_rgb_u16_be_avx+0x2f1> |
| .byte 233,90,253,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| .byte 233,85,253,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1 |
| .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,26 // jb 2063 <_sk_load_tables_rgb_u16_be_avx+0x320> |
| .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1 |
| .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 117,10 // jne 2068 <_sk_load_tables_rgb_u16_be_avx+0x325> |
| .byte 233,38,253,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| .byte 233,33,253,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3 |
| .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,20 // jb 2091 <_sk_load_tables_rgb_u16_be_avx+0x34e> |
| .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3 |
| .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3 |
| .byte 233,248,252,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| .byte 233,243,252,255,255 // jmpq 1d89 <_sk_load_tables_rgb_u16_be_avx+0x46> |
| |
| HIDDEN _sk_byte_tables_avx |
| .globl _sk_byte_tables_avx |
| FUNCTION(_sk_byte_tables_avx) |
| _sk_byte_tables_avx: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,127,67 // mov $0x437f0000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 68,137,197 // mov %r8d,%ebp |
| .byte 77,137,194 // mov %r8,%r10 |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,192 // vmovq %xmm0,%r8 |
| .byte 69,137,195 // mov %r8d,%r11d |
| .byte 77,137,199 // mov %r8,%r15 |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 69,137,198 // mov %r8d,%r14d |
| .byte 77,137,196 // mov %r8,%r12 |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 |
| .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp |
| .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp |
| .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0 |
| .byte 196,98,121,49,200 // vpmovzxbd %xmm0,%xmm9 |
| .byte 196,131,121,32,4,41,0 // vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,25,1 // vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 |
| .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp |
| .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp |
| .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0 |
| .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 |
| .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10 |
| .byte 189,129,128,128,59 // mov $0x3b808081,%ebp |
| .byte 197,249,110,197 // vmovd %ebp,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9 |
| .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp |
| .byte 65,137,233 // mov %ebp,%r9d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,203 // vmovq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,193,249,126,207 // vmovq %xmm1,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 |
| .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1 |
| .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp |
| .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1 |
| .byte 196,98,121,49,209 // vpmovzxbd %xmm1,%xmm10 |
| .byte 196,131,121,32,12,32,0 // vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1 |
| .byte 196,131,113,32,12,56,1 // vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1 |
| .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp |
| .byte 196,227,113,32,205,2 // vpinsrb $0x2,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp |
| .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1 |
| .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 |
| .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp |
| .byte 65,137,233 // mov %ebp,%r9d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,211 // vmovq %xmm2,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,195,249,22,211,1 // vpextrq $0x1,%xmm2,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,193,249,126,215 // vmovq %xmm2,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,20,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2 |
| .byte 196,195,105,32,20,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2 |
| .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp |
| .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2 |
| .byte 196,98,121,49,210 // vpmovzxbd %xmm2,%xmm10 |
| .byte 196,131,121,32,20,32,0 // vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm2 |
| .byte 196,131,105,32,20,56,1 // vpinsrb $0x1,(%r8,%r15,1),%xmm2,%xmm2 |
| .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp |
| .byte 196,227,105,32,213,2 // vpinsrb $0x2,%ebp,%xmm2,%xmm2 |
| .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp |
| .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2 |
| .byte 196,226,121,49,210 // vpmovzxbd %xmm2,%xmm2 |
| .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 |
| .byte 196,227,249,22,221,1 // vpextrq $0x1,%xmm3,%rbp |
| .byte 65,137,232 // mov %ebp,%r8d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,219 // vmovq %xmm3,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,25,219,1 // vextractf128 $0x1,%ymm3,%xmm3 |
| .byte 196,195,249,22,218,1 // vpextrq $0x1,%xmm3,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,222 // vmovq %xmm3,%r14 |
| .byte 69,137,247 // mov %r14d,%r15d |
| .byte 73,193,238,32 // shr $0x20,%r14 |
| .byte 196,163,121,32,28,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3 |
| .byte 196,227,97,32,28,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3 |
| .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx |
| .byte 196,227,97,32,219,2 // vpinsrb $0x2,%ebx,%xmm3,%xmm3 |
| .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp |
| .byte 196,227,97,32,221,3 // vpinsrb $0x3,%ebp,%xmm3,%xmm3 |
| .byte 196,98,121,49,195 // vpmovzxbd %xmm3,%xmm8 |
| .byte 196,163,121,32,28,56,0 // vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm3 |
| .byte 196,163,97,32,28,48,1 // vpinsrb $0x1,(%rax,%r14,1),%xmm3,%xmm3 |
| .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp |
| .byte 196,227,97,32,221,2 // vpinsrb $0x2,%ebp,%xmm3,%xmm3 |
| .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax |
| .byte 196,227,97,32,216,3 // vpinsrb $0x3,%eax,%xmm3,%xmm3 |
| .byte 196,226,121,49,219 // vpmovzxbd %xmm3,%xmm3 |
| .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_byte_tables_rgb_avx |
| .globl _sk_byte_tables_rgb_avx |
| FUNCTION(_sk_byte_tables_rgb_avx) |
| _sk_byte_tables_rgb_avx: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 68,139,64,24 // mov 0x18(%rax),%r8d |
| .byte 65,255,200 // dec %r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%xmm8 |
| .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 68,137,197 // mov %r8d,%ebp |
| .byte 77,137,194 // mov %r8,%r10 |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,192 // vmovq %xmm0,%r8 |
| .byte 69,137,195 // mov %r8d,%r11d |
| .byte 77,137,199 // mov %r8,%r15 |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8 |
| .byte 69,137,198 // mov %r8d,%r14d |
| .byte 77,137,196 // mov %r8,%r12 |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,221 // mov %ebx,%r13d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0 |
| .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp |
| .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp |
| .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0 |
| .byte 196,98,121,49,200 // vpmovzxbd %xmm0,%xmm9 |
| .byte 196,131,121,32,4,41,0 // vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,25,1 // vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 |
| .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp |
| .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0 |
| .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp |
| .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0 |
| .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 |
| .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10 |
| .byte 189,129,128,128,59 // mov $0x3b808081,%ebp |
| .byte 197,249,110,197 // vmovd %ebp,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9 |
| .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp |
| .byte 65,137,233 // mov %ebp,%r9d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,203 // vmovq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 196,193,249,126,207 // vmovq %xmm1,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1 |
| .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1 |
| .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp |
| .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1 |
| .byte 196,98,121,49,209 // vpmovzxbd %xmm1,%xmm10 |
| .byte 196,131,121,32,12,32,0 // vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1 |
| .byte 196,131,113,32,12,56,1 // vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1 |
| .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp |
| .byte 196,227,113,32,205,2 // vpinsrb $0x2,%ebp,%xmm1,%xmm1 |
| .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp |
| .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1 |
| .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 |
| .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 72,139,64,16 // mov 0x10(%rax),%rax |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp |
| .byte 65,137,232 // mov %ebp,%r8d |
| .byte 72,193,237,32 // shr $0x20,%rbp |
| .byte 196,225,249,126,211 // vmovq %xmm2,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,195,249,22,210,1 // vpextrq $0x1,%xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,193,249,126,214 // vmovq %xmm2,%r14 |
| .byte 69,137,247 // mov %r14d,%r15d |
| .byte 73,193,238,32 // shr $0x20,%r14 |
| .byte 196,163,121,32,20,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2 |
| .byte 196,227,105,32,20,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2 |
| .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx |
| .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2 |
| .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp |
| .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2 |
| .byte 196,98,121,49,194 // vpmovzxbd %xmm2,%xmm8 |
| .byte 196,163,121,32,20,56,0 // vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm2 |
| .byte 196,163,105,32,20,48,1 // vpinsrb $0x1,(%rax,%r14,1),%xmm2,%xmm2 |
| .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp |
| .byte 196,227,105,32,213,2 // vpinsrb $0x2,%ebp,%xmm2,%xmm2 |
| .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax |
| .byte 196,227,105,32,208,3 // vpinsrb $0x3,%eax,%xmm2,%xmm2 |
| .byte 196,226,121,49,210 // vpmovzxbd %xmm2,%xmm2 |
| .byte 196,227,61,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm8,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_r_avx |
| .globl _sk_table_r_avx |
| FUNCTION(_sk_table_r_avx) |
| _sk_table_r_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,194 // vmovq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,199 // vmovq %xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,129,122,16,4,160 // vmovss (%r8,%r12,4),%xmm0 |
| .byte 196,3,121,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm0,%xmm8 |
| .byte 196,129,122,16,4,176 // vmovss (%r8,%r14,4),%xmm0 |
| .byte 196,99,57,33,192,32 // vinsertps $0x20,%xmm0,%xmm8,%xmm8 |
| .byte 196,193,122,16,4,152 // vmovss (%r8,%rbx,4),%xmm0 |
| .byte 196,99,57,33,192,48 // vinsertps $0x30,%xmm0,%xmm8,%xmm8 |
| .byte 196,129,122,16,4,152 // vmovss (%r8,%r11,4),%xmm0 |
| .byte 196,3,121,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm0,%xmm9 |
| .byte 196,129,122,16,4,136 // vmovss (%r8,%r9,4),%xmm0 |
| .byte 196,99,49,33,200,32 // vinsertps $0x20,%xmm0,%xmm9,%xmm9 |
| .byte 196,193,122,16,4,128 // vmovss (%r8,%rax,4),%xmm0 |
| .byte 196,227,49,33,192,48 // vinsertps $0x30,%xmm0,%xmm9,%xmm0 |
| .byte 196,195,125,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm0,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_g_avx |
| .globl _sk_table_g_avx |
| FUNCTION(_sk_table_g_avx) |
| _sk_table_g_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,227,249,22,200,1 // vpextrq $0x1,%xmm1,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,202 // vmovq %xmm1,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,207 // vmovq %xmm1,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,129,122,16,12,160 // vmovss (%r8,%r12,4),%xmm1 |
| .byte 196,3,113,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm1,%xmm8 |
| .byte 196,129,122,16,12,176 // vmovss (%r8,%r14,4),%xmm1 |
| .byte 196,99,57,33,193,32 // vinsertps $0x20,%xmm1,%xmm8,%xmm8 |
| .byte 196,193,122,16,12,152 // vmovss (%r8,%rbx,4),%xmm1 |
| .byte 196,99,57,33,193,48 // vinsertps $0x30,%xmm1,%xmm8,%xmm8 |
| .byte 196,129,122,16,12,152 // vmovss (%r8,%r11,4),%xmm1 |
| .byte 196,3,113,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm1,%xmm9 |
| .byte 196,129,122,16,12,136 // vmovss (%r8,%r9,4),%xmm1 |
| .byte 196,99,49,33,201,32 // vinsertps $0x20,%xmm1,%xmm9,%xmm9 |
| .byte 196,193,122,16,12,128 // vmovss (%r8,%rax,4),%xmm1 |
| .byte 196,227,49,33,201,48 // vinsertps $0x30,%xmm1,%xmm9,%xmm1 |
| .byte 196,195,117,24,200,1 // vinsertf128 $0x1,%xmm8,%ymm1,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_b_avx |
| .globl _sk_table_b_avx |
| FUNCTION(_sk_table_b_avx) |
| _sk_table_b_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,210 // vmovq %xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,215 // vmovq %xmm2,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2 |
| .byte 196,3,105,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm2,%xmm8 |
| .byte 196,129,122,16,20,176 // vmovss (%r8,%r14,4),%xmm2 |
| .byte 196,99,57,33,194,32 // vinsertps $0x20,%xmm2,%xmm8,%xmm8 |
| .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2 |
| .byte 196,99,57,33,194,48 // vinsertps $0x30,%xmm2,%xmm8,%xmm8 |
| .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2 |
| .byte 196,3,105,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm2,%xmm9 |
| .byte 196,129,122,16,20,136 // vmovss (%r8,%r9,4),%xmm2 |
| .byte 196,99,49,33,202,32 // vinsertps $0x20,%xmm2,%xmm9,%xmm9 |
| .byte 196,193,122,16,20,128 // vmovss (%r8,%rax,4),%xmm2 |
| .byte 196,227,49,33,210,48 // vinsertps $0x30,%xmm2,%xmm9,%xmm2 |
| .byte 196,195,109,24,208,1 // vinsertf128 $0x1,%xmm8,%ymm2,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_a_avx |
| .globl _sk_table_a_avx |
| FUNCTION(_sk_table_a_avx) |
| _sk_table_a_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 |
| .byte 196,227,249,22,216,1 // vpextrq $0x1,%xmm3,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,218 // vmovq %xmm3,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 196,227,125,25,219,1 // vextractf128 $0x1,%ymm3,%xmm3 |
| .byte 196,227,249,22,219,1 // vpextrq $0x1,%xmm3,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,249,126,223 // vmovq %xmm3,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,129,122,16,28,160 // vmovss (%r8,%r12,4),%xmm3 |
| .byte 196,3,97,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm3,%xmm8 |
| .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3 |
| .byte 196,99,57,33,195,32 // vinsertps $0x20,%xmm3,%xmm8,%xmm8 |
| .byte 196,193,122,16,28,152 // vmovss (%r8,%rbx,4),%xmm3 |
| .byte 196,99,57,33,195,48 // vinsertps $0x30,%xmm3,%xmm8,%xmm8 |
| .byte 196,129,122,16,28,152 // vmovss (%r8,%r11,4),%xmm3 |
| .byte 196,3,97,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm3,%xmm9 |
| .byte 196,129,122,16,28,136 // vmovss (%r8,%r9,4),%xmm3 |
| .byte 196,99,49,33,203,32 // vinsertps $0x20,%xmm3,%xmm9,%xmm9 |
| .byte 196,193,122,16,28,128 // vmovss (%r8,%rax,4),%xmm3 |
| .byte 196,227,49,33,219,48 // vinsertps $0x30,%xmm3,%xmm9,%xmm3 |
| .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_r_avx |
| .globl _sk_parametric_r_avx |
| FUNCTION(_sk_parametric_r_avx) |
| _sk_parametric_r_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,124,194,192,2 // vcmpleps %ymm8,%ymm0,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 196,65,124,88,211 // vaddps %ymm11,%ymm0,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,36,89,216 // vmulps %ymm0,%ymm11,%ymm11 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 196,65,124,84,210 // vandps %ymm10,%ymm0,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,44,86,208 // vorps %ymm0,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,172,88,192 // vaddps %ymm0,%ymm10,%ymm0 |
| .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 |
| .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0 |
| .byte 197,28,89,216 // vmulps %ymm0,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 196,65,124,88,219 // vaddps %ymm11,%ymm0,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0 |
| .byte 197,36,92,216 // vsubps %ymm0,%ymm11,%ymm11 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm13 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0 |
| .byte 197,148,94,192 // vdivps %ymm0,%ymm13,%ymm0 |
| .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0 |
| .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 |
| .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 |
| .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,124,95,192 // vmaxps %ymm8,%ymm0,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_g_avx |
| .globl _sk_parametric_g_avx |
| FUNCTION(_sk_parametric_g_avx) |
| _sk_parametric_g_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,116,194,192,2 // vcmpleps %ymm8,%ymm1,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,65,116,88,211 // vaddps %ymm11,%ymm1,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,36,89,217 // vmulps %ymm1,%ymm11,%ymm11 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| .byte 196,65,116,84,210 // vandps %ymm10,%ymm1,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| .byte 197,44,86,209 // vorps %ymm1,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,233,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 |
| .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 |
| .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1 |
| .byte 197,28,89,217 // vmulps %ymm1,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 196,65,116,88,219 // vaddps %ymm11,%ymm1,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 |
| .byte 197,36,92,217 // vsubps %ymm1,%ymm11,%ymm11 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,233,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm13 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1 |
| .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1 |
| .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 |
| .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 |
| .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8 |
| .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 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_b_avx |
| .globl _sk_parametric_b_avx |
| FUNCTION(_sk_parametric_b_avx) |
| _sk_parametric_b_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,108,194,192,2 // vcmpleps %ymm8,%ymm2,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 196,65,108,88,211 // vaddps %ymm11,%ymm2,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 |
| .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 |
| .byte 196,65,108,84,210 // vandps %ymm10,%ymm2,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 |
| .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 |
| .byte 197,44,86,210 // vorps %ymm2,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| .byte 196,99,109,24,234,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 |
| .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2 |
| .byte 197,28,89,218 // vmulps %ymm2,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| .byte 196,99,109,24,226,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 196,65,108,88,219 // vaddps %ymm11,%ymm2,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2 |
| .byte 197,36,92,218 // vsubps %ymm2,%ymm11,%ymm11 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%xmm2 |
| .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| .byte 196,99,109,24,234,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm13 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,208 // vmovd %r8d,%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 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2 |
| .byte 197,148,94,210 // vdivps %ymm2,%ymm13,%ymm2 |
| .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 |
| .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,108,95,192 // vmaxps %ymm8,%ymm2,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,188,93,210 // vminps %ymm2,%ymm8,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_a_avx |
| .globl _sk_parametric_a_avx |
| FUNCTION(_sk_parametric_a_avx) |
| _sk_parametric_a_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,65,100,194,192,2 // vcmpleps %ymm8,%ymm3,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 |
| .byte 196,65,100,88,211 // vaddps %ymm11,%ymm3,%ymm10 |
| .byte 196,98,125,24,32 // vbroadcastss (%rax),%ymm12 |
| .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11 |
| .byte 65,184,0,0,0,52 // mov $0x34000000,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 |
| .byte 65,184,255,255,127,0 // mov $0x7fffff,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| .byte 196,65,100,84,210 // vandps %ymm10,%ymm3,%ymm10 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| .byte 197,44,86,211 // vorps %ymm3,%ymm10,%ymm10 |
| .byte 65,184,119,115,248,66 // mov $0x42f87377,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 |
| .byte 65,184,117,191,191,63 // mov $0x3fbfbf75,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 |
| .byte 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 |
| .byte 65,184,163,233,220,63 // mov $0x3fdce9a3,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| .byte 196,99,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 |
| .byte 65,184,249,68,180,62 // mov $0x3eb444f9,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 197,172,88,219 // vaddps %ymm3,%ymm10,%ymm3 |
| .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 |
| .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3 |
| .byte 197,28,89,219 // vmulps %ymm3,%ymm12,%ymm11 |
| .byte 196,67,125,8,211,1 // vroundps $0x1,%ymm11,%ymm10 |
| .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| .byte 196,99,101,24,227,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm12 |
| .byte 65,184,81,140,242,66 // mov $0x42f28c51,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 196,65,100,88,219 // vaddps %ymm11,%ymm3,%ymm11 |
| .byte 65,184,141,188,190,63 // mov $0x3fbebc8d,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3 |
| .byte 197,36,92,219 // vsubps %ymm3,%ymm11,%ymm11 |
| .byte 65,184,254,210,221,65 // mov $0x41ddd2fe,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%xmm3 |
| .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| .byte 196,99,101,24,235,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm13 |
| .byte 65,184,248,245,154,64 // mov $0x409af5f8,%r8d |
| .byte 196,193,121,110,216 // vmovd %r8d,%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 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3 |
| .byte 197,148,94,219 // vdivps %ymm3,%ymm13,%ymm3 |
| .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3 |
| .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 |
| .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3 |
| .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 196,65,100,95,192 // vmaxps %ymm8,%ymm3,%ymm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 197,188,93,219 // vminps %ymm3,%ymm8,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_a8_avx |
| .globl _sk_load_a8_avx |
| FUNCTION(_sk_load_a8_avx) |
| _sk_load_a8_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 117,74 // jne 309e <_sk_load_a8_avx+0x5a> |
| .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 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 76,137,193 // mov %r8,%rcx |
| .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 30a6 <_sk_load_a8_avx+0x62> |
| .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| .byte 235,149 // jmp 3058 <_sk_load_a8_avx+0x14> |
| |
| HIDDEN _sk_gather_a8_avx |
| .globl _sk_gather_a8_avx |
| FUNCTION(_sk_gather_a8_avx) |
| _sk_gather_a8_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 |
| .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1 |
| .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3 |
| .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,210 // vmovq %xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,131,121,32,4,48,0 // vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0 |
| .byte 67,15,182,28,32 // movzbl (%r8,%r12,1),%ebx |
| .byte 196,227,121,32,195,2 // vpinsrb $0x2,%ebx,%xmm0,%xmm0 |
| .byte 67,15,182,28,56 // movzbl (%r8,%r15,1),%ebx |
| .byte 196,227,121,32,195,3 // vpinsrb $0x3,%ebx,%xmm0,%xmm0 |
| .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| .byte 196,131,121,32,12,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1 |
| .byte 196,131,113,32,12,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1 |
| .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax |
| .byte 196,227,113,32,200,3 // vpinsrb $0x3,%eax,%xmm1,%xmm1 |
| .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%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 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_a8_avx |
| .globl _sk_store_a8_avx |
| FUNCTION(_sk_store_a8_avx) |
| _sk_store_a8_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 184,0,0,127,67 // mov $0x437f0000,%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 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 117,10 // jne 31ff <_sk_store_a8_avx+0x42> |
| .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) |
| .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 31fb <_sk_store_a8_avx+0x3e> |
| .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3264 <_sk_store_a8_avx+0xa7> |
| .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| .byte 76,1,192 // add %r8,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,67,121,20,68,57,6,12 // vpextrb $0xc,%xmm8,0x6(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,5,10 // vpextrb $0xa,%xmm8,0x5(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,4,8 // vpextrb $0x8,%xmm8,0x4(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,3,6 // vpextrb $0x6,%xmm8,0x3(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) |
| .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) |
| .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) |
| .byte 235,154 // jmp 31fb <_sk_store_a8_avx+0x3e> |
| .byte 15,31,0 // nopl (%rax) |
| .byte 244 // hlt |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 236 // in (%dx),%al |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,228 // jmpq *%rsp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 220,255 // fdivr %st,%st(7) |
| .byte 255 // (bad) |
| .byte 255,212 // callq *%rsp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,204 // dec %esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,196 // inc %esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_load_g8_avx |
| .globl _sk_load_g8_avx |
| FUNCTION(_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 32eb <_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 32f3 <_sk_load_g8_avx+0x73> |
| .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| .byte 235,132 // jmp 3294 <_sk_load_g8_avx+0x14> |
| |
| HIDDEN _sk_gather_g8_avx |
| .globl _sk_gather_g8_avx |
| FUNCTION(_sk_gather_g8_avx) |
| _sk_gather_g8_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 |
| .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1 |
| .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3 |
| .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,210 // vmovq %xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,131,121,32,4,48,0 // vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0 |
| .byte 67,15,182,28,32 // movzbl (%r8,%r12,1),%ebx |
| .byte 196,227,121,32,195,2 // vpinsrb $0x2,%ebx,%xmm0,%xmm0 |
| .byte 67,15,182,28,56 // movzbl (%r8,%r15,1),%ebx |
| .byte 196,227,121,32,195,3 // vpinsrb $0x3,%ebx,%xmm0,%xmm0 |
| .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| .byte 196,131,121,32,12,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1 |
| .byte 196,131,113,32,12,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1 |
| .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx |
| .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1 |
| .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax |
| .byte 196,227,113,32,200,3 // vpinsrb $0x3,%eax,%xmm1,%xmm1 |
| .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%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 197,252,40,200 // vmovaps %ymm0,%ymm1 |
| .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_i8_avx |
| .globl _sk_gather_i8_avx |
| FUNCTION(_sk_gather_i8_avx) |
| _sk_gather_i8_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,192 // mov %rax,%r8 |
| .byte 77,133,192 // test %r8,%r8 |
| .byte 116,5 // je 342a <_sk_gather_i8_avx+0xf> |
| .byte 76,137,192 // mov %r8,%rax |
| .byte 235,2 // jmp 342c <_sk_gather_i8_avx+0x11> |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 |
| .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1 |
| .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3 |
| .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,194 // mov %eax,%r10d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,211 // vmovq %xmm2,%r11 |
| .byte 69,137,222 // mov %r11d,%r14d |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,223 // mov %ebx,%r15d |
| .byte 196,195,249,22,196,1 // vpextrq $0x1,%xmm0,%r12 |
| .byte 69,137,229 // mov %r12d,%r13d |
| .byte 73,193,236,32 // shr $0x20,%r12 |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,131,121,32,4,49,0 // vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,25,1 // vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,17,2 // vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,1,3 // vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0 |
| .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10 |
| .byte 196,193,249,126,195 // vmovq %xmm0,%r11 |
| .byte 196,131,121,32,4,57,0 // vpinsrb $0x0,(%r9,%r15,1),%xmm0,%xmm0 |
| .byte 196,195,121,32,4,25,1 // vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,41,2 // vpinsrb $0x2,(%r9,%r13,1),%xmm0,%xmm0 |
| .byte 196,131,121,32,4,33,3 // vpinsrb $0x3,(%r9,%r12,1),%xmm0,%xmm0 |
| .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| .byte 73,139,88,8 // mov 0x8(%r8),%rbx |
| .byte 196,193,249,126,193 // vmovq %xmm0,%r9 |
| .byte 69,137,200 // mov %r9d,%r8d |
| .byte 73,193,233,30 // shr $0x1e,%r9 |
| .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax |
| .byte 65,137,198 // mov %eax,%r14d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 69,137,223 // mov %r11d,%r15d |
| .byte 73,193,235,30 // shr $0x1e,%r11 |
| .byte 69,137,212 // mov %r10d,%r12d |
| .byte 73,193,234,30 // shr $0x1e,%r10 |
| .byte 196,161,121,110,4,131 // vmovd (%rbx,%r8,4),%xmm0 |
| .byte 196,163,121,34,4,11,1 // vpinsrd $0x1,(%rbx,%r9,1),%xmm0,%xmm0 |
| .byte 196,163,121,34,4,179,2 // vpinsrd $0x2,(%rbx,%r14,4),%xmm0,%xmm0 |
| .byte 196,99,121,34,4,3,3 // vpinsrd $0x3,(%rbx,%rax,1),%xmm0,%xmm8 |
| .byte 196,161,121,110,4,187 // vmovd (%rbx,%r15,4),%xmm0 |
| .byte 196,163,121,34,4,27,1 // vpinsrd $0x1,(%rbx,%r11,1),%xmm0,%xmm0 |
| .byte 196,163,121,34,4,163,2 // vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0 |
| .byte 196,163,121,34,28,19,3 // vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3 |
| .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm11 |
| .byte 197,164,84,192 // vandps %ymm0,%ymm11,%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 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9 |
| .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 |
| .byte 196,193,41,114,208,8 // vpsrld $0x8,%xmm8,%xmm10 |
| .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1 |
| .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 |
| .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 |
| .byte 196,193,41,114,208,16 // vpsrld $0x10,%xmm8,%xmm10 |
| .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2 |
| .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 |
| .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2 |
| .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8 |
| .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3 |
| .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,217 // vmulps %ymm9,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_565_avx |
| .globl _sk_load_565_avx |
| FUNCTION(_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 36c6 <_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 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%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 197,252,84,194 // vandps %ymm2,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%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 197,244,84,202 // vandps %ymm2,%ymm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%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 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%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 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%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 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 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 15,135,29,255,255,255 // ja 35fb <_sk_load_565_avx+0x14> |
| .byte 69,15,182,192 // movzbl %r8b,%r8d |
| .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 3734 <_sk_load_565_avx+0x14d> |
| .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 233,201,254,255,255 // jmpq 35fb <_sk_load_565_avx+0x14> |
| .byte 102,144 // xchg %ax,%ax |
| .byte 242,255 // repnz (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 234 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,226 // jmpq *%rdx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 218,255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,210 // callq *%rdx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,202 // dec %edx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 190 // .byte 0xbe |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_gather_565_avx |
| .globl _sk_gather_565_avx |
| FUNCTION(_sk_gather_565_avx) |
| _sk_gather_565_avx: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 |
| .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1 |
| .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3 |
| .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,210 // vmovq %xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 65,15,183,28,88 // movzwl (%r8,%rbx,2),%ebx |
| .byte 67,15,183,44,112 // movzwl (%r8,%r14,2),%ebp |
| .byte 197,249,110,197 // vmovd %ebp,%xmm0 |
| .byte 197,249,196,195,1 // vpinsrw $0x1,%ebx,%xmm0,%xmm0 |
| .byte 67,15,183,28,96 // movzwl (%r8,%r12,2),%ebx |
| .byte 197,249,196,195,2 // vpinsrw $0x2,%ebx,%xmm0,%xmm0 |
| .byte 67,15,183,28,120 // movzwl (%r8,%r15,2),%ebx |
| .byte 197,249,196,195,3 // vpinsrw $0x3,%ebx,%xmm0,%xmm0 |
| .byte 67,15,183,44,88 // movzwl (%r8,%r11,2),%ebp |
| .byte 197,249,196,197,4 // vpinsrw $0x4,%ebp,%xmm0,%xmm0 |
| .byte 67,15,183,44,80 // movzwl (%r8,%r10,2),%ebp |
| .byte 197,249,196,197,5 // vpinsrw $0x5,%ebp,%xmm0,%xmm0 |
| .byte 67,15,183,44,72 // movzwl (%r8,%r9,2),%ebp |
| .byte 197,249,196,197,6 // vpinsrw $0x6,%ebp,%xmm0,%xmm0 |
| .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax |
| .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%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 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%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 197,252,84,194 // vandps %ymm2,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%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 197,244,84,202 // vandps %ymm2,%ymm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%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 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%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 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%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 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_565_avx |
| .globl _sk_store_565_avx |
| FUNCTION(_sk_store_565_avx) |
| _sk_store_565_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 184,0,0,248,65 // mov $0x41f80000,%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 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10 |
| .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 |
| .byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9 |
| .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 |
| .byte 184,0,0,124,66 // mov $0x427c0000,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| .byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11 |
| .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| .byte 196,193,41,114,242,5 // vpslld $0x5,%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 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%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 397f <_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 |
| .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 397b <_sk_store_565_avx+0x9a> |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 39dc <_sk_store_565_avx+0xfb> |
| .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 397b <_sk_store_565_avx+0x9a> |
| .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_4444_avx |
| .globl _sk_load_4444_avx |
| FUNCTION(_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 3afb <_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 3a0c <_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 # 3b68 <_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 3a0c <_sk_load_4444_avx+0x14> |
| .byte 144 // nop |
| .byte 243,255 // repz (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 235,255 // jmp 3b6d <_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_gather_4444_avx |
| .globl _sk_gather_4444_avx |
| FUNCTION(_sk_gather_4444_avx) |
| _sk_gather_4444_avx: |
| .byte 85 // push %rbp |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 |
| .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1 |
| .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3 |
| .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,210 // vmovq %xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 65,15,183,28,88 // movzwl (%r8,%rbx,2),%ebx |
| .byte 67,15,183,44,112 // movzwl (%r8,%r14,2),%ebp |
| .byte 197,249,110,197 // vmovd %ebp,%xmm0 |
| .byte 197,249,196,195,1 // vpinsrw $0x1,%ebx,%xmm0,%xmm0 |
| .byte 67,15,183,28,96 // movzwl (%r8,%r12,2),%ebx |
| .byte 197,249,196,195,2 // vpinsrw $0x2,%ebx,%xmm0,%xmm0 |
| .byte 67,15,183,28,120 // movzwl (%r8,%r15,2),%ebx |
| .byte 197,249,196,195,3 // vpinsrw $0x3,%ebx,%xmm0,%xmm0 |
| .byte 67,15,183,44,88 // movzwl (%r8,%r11,2),%ebp |
| .byte 197,249,196,197,4 // vpinsrw $0x4,%ebp,%xmm0,%xmm0 |
| .byte 67,15,183,44,80 // movzwl (%r8,%r10,2),%ebp |
| .byte 197,249,196,197,5 // vpinsrw $0x5,%ebp,%xmm0,%xmm0 |
| .byte 67,15,183,44,72 // movzwl (%r8,%r9,2),%ebp |
| .byte 197,249,196,197,6 // vpinsrw $0x6,%ebp,%xmm0,%xmm0 |
| .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax |
| .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%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 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 93 // pop %rbp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_4444_avx |
| .globl _sk_store_4444_avx |
| FUNCTION(_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 3de8 <_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 3de4 <_sk_store_4444_avx+0xab> |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3e48 <_sk_store_4444_avx+0x10f> |
| .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 3de4 <_sk_store_4444_avx+0xab> |
| .byte 15,31,0 // nopl (%rax) |
| .byte 244 // hlt |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 236 // in (%dx),%al |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,228 // jmpq *%rsp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 220,255 // fdivr %st,%st(7) |
| .byte 255 // (bad) |
| .byte 255,212 // callq *%rsp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,204 // dec %esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,196 // inc %esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_load_8888_avx |
| .globl _sk_load_8888_avx |
| FUNCTION(_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 3f0f <_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 |
| .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,216,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm11 |
| .byte 196,193,36,84,193 // vandps %ymm9,%ymm11,%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 196,99,117,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm8 |
| .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| .byte 196,193,41,114,209,8 // vpsrld $0x8,%xmm9,%xmm10 |
| .byte 196,99,125,25,203,1 // vextractf128 $0x1,%ymm9,%xmm3 |
| .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1 |
| .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 |
| .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| .byte 196,193,41,114,209,16 // vpsrld $0x10,%xmm9,%xmm10 |
| .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2 |
| .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 |
| .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| .byte 196,193,49,114,209,24 // vpsrld $0x18,%xmm9,%xmm9 |
| .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3 |
| .byte 196,227,53,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%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 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 3e78 <_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 # 3fbc <_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 |
| .byte 196,193,121,110,68,186,24 // vmovd 0x18(%r10,%rdi,4),%xmm0 |
| .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 196,99,117,12,200,64 // vblendps $0x40,%ymm0,%ymm1,%ymm9 |
| .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 |
| .byte 196,195,121,34,68,186,20,1 // vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0 |
| .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 |
| .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 |
| .byte 196,195,121,34,68,186,16,0 // vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0 |
| .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 |
| .byte 196,195,49,34,68,186,12,3 // vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm9,%xmm0 |
| .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 |
| .byte 196,195,49,34,68,186,8,2 // vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm9,%xmm0 |
| .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 |
| .byte 196,195,49,34,68,186,4,1 // vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm9,%xmm0 |
| .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 3e78 <_sk_load_8888_avx+0x14> |
| .byte 238 // out %al,(%dx) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,224 // jmpq *%rax |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,210 // callq *%rdx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,196 // inc %esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,176,255,255,255,156 // pushq -0x63000001(%rax) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| .byte 128,255,255 // cmp $0xff,%bh |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_gather_8888_avx |
| .globl _sk_gather_8888_avx |
| FUNCTION(_sk_gather_8888_avx) |
| _sk_gather_8888_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1 |
| .byte 197,249,110,80,16 // vmovd 0x10(%rax),%xmm2 |
| .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 |
| .byte 196,226,105,64,217 // vpmulld %xmm1,%xmm2,%xmm3 |
| .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| .byte 196,226,105,64,201 // vpmulld %xmm1,%xmm2,%xmm1 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,194,1 // vextractf128 $0x1,%ymm0,%xmm2 |
| .byte 197,241,254,202 // vpaddd %xmm2,%xmm1,%xmm1 |
| .byte 196,225,249,126,200 // vmovq %xmm1,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 197,225,254,192 // vpaddd %xmm0,%xmm3,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 196,129,121,110,4,176 // vmovd (%r8,%r14,4),%xmm0 |
| .byte 196,195,121,34,4,152,1 // vpinsrd $0x1,(%r8,%rbx,4),%xmm0,%xmm0 |
| .byte 196,131,121,34,4,160,2 // vpinsrd $0x2,(%r8,%r12,4),%xmm0,%xmm0 |
| .byte 196,3,121,34,4,184,3 // vpinsrd $0x3,(%r8,%r15,4),%xmm0,%xmm8 |
| .byte 196,129,121,110,4,136 // vmovd (%r8,%r9,4),%xmm0 |
| .byte 196,195,121,34,4,128,1 // vpinsrd $0x1,(%r8,%rax,4),%xmm0,%xmm0 |
| .byte 196,131,121,34,4,152,2 // vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0 |
| .byte 196,131,121,34,28,144,3 // vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3 |
| .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm11 |
| .byte 197,164,84,192 // vandps %ymm0,%ymm11,%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 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9 |
| .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 |
| .byte 196,193,41,114,208,8 // vpsrld $0x8,%xmm8,%xmm10 |
| .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1 |
| .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 |
| .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1 |
| .byte 196,193,41,114,208,16 // vpsrld $0x10,%xmm8,%xmm10 |
| .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2 |
| .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 |
| .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2 |
| .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8 |
| .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3 |
| .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 196,193,100,89,217 // vmulps %ymm9,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_8888_avx |
| .globl _sk_store_8888_avx |
| FUNCTION(_sk_store_8888_avx) |
| _sk_store_8888_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 184,0,0,127,67 // mov $0x437f0000,%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 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 196,193,33,114,242,16 // vpslld $0x10,%xmm10,%xmm11 |
| .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| .byte 196,193,41,114,242,16 // vpslld $0x10,%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 196,193,33,114,240,24 // vpslld $0x18,%xmm8,%xmm11 |
| .byte 196,67,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm8 |
| .byte 196,193,57,114,240,24 // vpslld $0x18,%xmm8,%xmm8 |
| .byte 196,67,37,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm11,%ymm8 |
| .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 41bd <_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 |
| .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 41b9 <_sk_store_8888_avx+0xa0> |
| .byte 65,15,182,192 // movzbl %r8b,%eax |
| .byte 76,141,5,84,0,0,0 // lea 0x54(%rip),%r8 # 422c <_sk_store_8888_avx+0x113> |
| .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| .byte 76,1,192 // add %r8,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| .byte 196,67,121,22,76,185,24,2 // vpextrd $0x2,%xmm9,0x18(%r9,%rdi,4) |
| .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| .byte 196,67,121,22,76,185,20,1 // vpextrd $0x1,%xmm9,0x14(%r9,%rdi,4) |
| .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| .byte 196,65,122,17,76,185,16 // vmovss %xmm9,0x10(%r9,%rdi,4) |
| .byte 196,67,121,22,68,185,12,3 // vpextrd $0x3,%xmm8,0xc(%r9,%rdi,4) |
| .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 41b9 <_sk_store_8888_avx+0xa0> |
| .byte 102,144 // xchg %ax,%ax |
| .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 255,209 // callq *%rcx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,195 // inc %ebx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| .byte 181,255 // mov $0xff,%ch |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_load_f16_avx |
| .globl _sk_load_f16_avx |
| FUNCTION(_sk_load_f16_avx) |
| _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 197,252,17,124,36,200 // vmovups %ymm7,-0x38(%rsp) |
| .byte 197,252,17,116,36,168 // vmovups %ymm6,-0x58(%rsp) |
| .byte 197,252,17,108,36,136 // vmovups %ymm5,-0x78(%rsp) |
| .byte 15,133,101,2,0,0 // jne 44cd <_sk_load_f16_avx+0x285> |
| .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 197,249,16,76,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm1 |
| .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9 |
| .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| .byte 196,193,113,97,217 // vpunpcklwd %xmm9,%xmm1,%xmm3 |
| .byte 196,193,113,105,201 // vpunpckhwd %xmm9,%xmm1,%xmm1 |
| .byte 197,121,97,242 // vpunpcklwd %xmm2,%xmm0,%xmm14 |
| .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8 |
| .byte 197,97,97,249 // vpunpcklwd %xmm1,%xmm3,%xmm15 |
| .byte 197,97,105,209 // vpunpckhwd %xmm1,%xmm3,%xmm10 |
| .byte 196,193,9,108,199 // vpunpcklqdq %xmm15,%xmm14,%xmm0 |
| .byte 196,65,25,239,228 // vpxor %xmm12,%xmm12,%xmm12 |
| .byte 196,193,121,105,204 // vpunpckhwd %xmm12,%xmm0,%xmm1 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| .byte 184,0,128,0,0 // mov $0x8000,%eax |
| .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9 |
| .byte 196,193,124,84,201 // vandps %ymm9,%ymm0,%ymm1 |
| .byte 184,0,124,0,0 // mov $0x7c00,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| .byte 196,99,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm11 |
| .byte 196,193,124,84,219 // vandps %ymm11,%ymm0,%ymm3 |
| .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0 |
| .byte 196,227,125,25,218,1 // vextractf128 $0x1,%ymm3,%xmm2 |
| .byte 196,193,105,118,212 // vpcmpeqd %xmm12,%xmm2,%xmm2 |
| .byte 196,193,97,118,220 // vpcmpeqd %xmm12,%xmm3,%xmm3 |
| .byte 196,227,101,24,242,1 // vinsertf128 $0x1,%xmm2,%ymm3,%ymm6 |
| .byte 196,227,125,25,203,1 // vextractf128 $0x1,%ymm1,%xmm3 |
| .byte 197,145,114,243,16 // vpslld $0x10,%xmm3,%xmm13 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,114,243,13 // vpslld $0xd,%xmm3,%xmm2 |
| .byte 184,0,0,0,56 // mov $0x38000000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| .byte 197,145,254,251 // vpaddd %xmm3,%xmm13,%xmm7 |
| .byte 197,193,254,210 // vpaddd %xmm2,%xmm7,%xmm2 |
| .byte 197,241,114,241,16 // vpslld $0x10,%xmm1,%xmm1 |
| .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0 |
| .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 |
| .byte 196,65,20,87,237 // vxorps %ymm13,%ymm13,%ymm13 |
| .byte 196,195,125,74,197,96 // vblendvps %ymm6,%ymm13,%ymm0,%ymm0 |
| .byte 196,193,9,109,207 // vpunpckhqdq %xmm15,%xmm14,%xmm1 |
| .byte 196,193,113,105,212 // vpunpckhwd %xmm12,%xmm1,%xmm2 |
| .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 196,193,116,84,209 // vandps %ymm9,%ymm1,%ymm2 |
| .byte 196,193,116,84,243 // vandps %ymm11,%ymm1,%ymm6 |
| .byte 197,244,87,202 // vxorps %ymm2,%ymm1,%ymm1 |
| .byte 196,227,125,25,247,1 // vextractf128 $0x1,%ymm6,%xmm7 |
| .byte 196,193,65,118,252 // vpcmpeqd %xmm12,%xmm7,%xmm7 |
| .byte 196,193,73,118,244 // vpcmpeqd %xmm12,%xmm6,%xmm6 |
| .byte 196,99,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm14 |
| .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7 |
| .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 |
| .byte 196,227,125,25,206,1 // vextractf128 $0x1,%ymm1,%xmm6 |
| .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 |
| .byte 197,193,254,251 // vpaddd %xmm3,%xmm7,%xmm7 |
| .byte 197,193,254,246 // vpaddd %xmm6,%xmm7,%xmm6 |
| .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2 |
| .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 197,233,254,201 // vpaddd %xmm1,%xmm2,%xmm1 |
| .byte 196,227,117,24,206,1 // vinsertf128 $0x1,%xmm6,%ymm1,%ymm1 |
| .byte 196,195,117,74,205,224 // vblendvps %ymm14,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,57,108,210 // vpunpcklqdq %xmm10,%xmm8,%xmm2 |
| .byte 196,193,105,105,244 // vpunpckhwd %xmm12,%xmm2,%xmm6 |
| .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2 |
| .byte 196,193,108,84,243 // vandps %ymm11,%ymm2,%ymm6 |
| .byte 196,227,125,25,247,1 // vextractf128 $0x1,%ymm6,%xmm7 |
| .byte 196,193,65,118,252 // vpcmpeqd %xmm12,%xmm7,%xmm7 |
| .byte 196,193,73,118,244 // vpcmpeqd %xmm12,%xmm6,%xmm6 |
| .byte 196,99,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm14 |
| .byte 196,193,108,84,249 // vandps %ymm9,%ymm2,%ymm7 |
| .byte 197,236,87,215 // vxorps %ymm7,%ymm2,%ymm2 |
| .byte 196,227,125,25,254,1 // vextractf128 $0x1,%ymm7,%xmm6 |
| .byte 197,129,114,246,16 // vpslld $0x10,%xmm6,%xmm15 |
| .byte 196,227,125,25,214,1 // vextractf128 $0x1,%ymm2,%xmm6 |
| .byte 197,209,114,246,13 // vpslld $0xd,%xmm6,%xmm5 |
| .byte 197,129,254,243 // vpaddd %xmm3,%xmm15,%xmm6 |
| .byte 197,201,254,237 // vpaddd %xmm5,%xmm6,%xmm5 |
| .byte 197,201,114,247,16 // vpslld $0x10,%xmm7,%xmm6 |
| .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2 |
| .byte 197,201,254,243 // vpaddd %xmm3,%xmm6,%xmm6 |
| .byte 197,201,254,210 // vpaddd %xmm2,%xmm6,%xmm2 |
| .byte 196,227,109,24,213,1 // vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 |
| .byte 196,195,109,74,213,224 // vblendvps %ymm14,%ymm13,%ymm2,%ymm2 |
| .byte 196,193,57,109,234 // vpunpckhqdq %xmm10,%xmm8,%xmm5 |
| .byte 196,193,81,105,244 // vpunpckhwd %xmm12,%xmm5,%xmm6 |
| .byte 196,226,121,51,237 // vpmovzxwd %xmm5,%xmm5 |
| .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 |
| .byte 196,193,84,84,243 // vandps %ymm11,%ymm5,%ymm6 |
| .byte 196,227,125,25,247,1 // vextractf128 $0x1,%ymm6,%xmm7 |
| .byte 196,193,65,118,252 // vpcmpeqd %xmm12,%xmm7,%xmm7 |
| .byte 196,193,73,118,244 // vpcmpeqd %xmm12,%xmm6,%xmm6 |
| .byte 196,65,84,84,193 // vandps %ymm9,%ymm5,%ymm8 |
| .byte 196,193,84,87,232 // vxorps %ymm8,%ymm5,%ymm5 |
| .byte 196,99,77,24,207,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm9 |
| .byte 196,99,125,25,199,1 // vextractf128 $0x1,%ymm8,%xmm7 |
| .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 |
| .byte 196,193,73,114,240,16 // vpslld $0x10,%xmm8,%xmm6 |
| .byte 197,201,254,243 // vpaddd %xmm3,%xmm6,%xmm6 |
| .byte 197,193,254,219 // vpaddd %xmm3,%xmm7,%xmm3 |
| .byte 196,227,125,25,239,1 // vextractf128 $0x1,%ymm5,%xmm7 |
| .byte 197,193,114,247,13 // vpslld $0xd,%xmm7,%xmm7 |
| .byte 197,225,254,223 // vpaddd %xmm7,%xmm3,%xmm3 |
| .byte 197,209,114,245,13 // vpslld $0xd,%xmm5,%xmm5 |
| .byte 197,201,254,237 // vpaddd %xmm5,%xmm6,%xmm5 |
| .byte 196,227,85,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm5,%ymm3 |
| .byte 196,195,101,74,221,144 // vblendvps %ymm9,%ymm13,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,16,108,36,136 // vmovups -0x78(%rsp),%ymm5 |
| .byte 197,252,16,116,36,168 // vmovups -0x58(%rsp),%ymm6 |
| .byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7 |
| .byte 255,224 // jmpq *%rax |
| .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 452c <_sk_load_f16_avx+0x2e4> |
| .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 452c <_sk_load_f16_avx+0x2e4> |
| .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 4539 <_sk_load_f16_avx+0x2f1> |
| .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 4539 <_sk_load_f16_avx+0x2f1> |
| .byte 197,251,16,76,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm1 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 15,132,110,253,255,255 // je 427f <_sk_load_f16_avx+0x37> |
| .byte 197,241,22,76,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 15,130,94,253,255,255 // jb 427f <_sk_load_f16_avx+0x37> |
| .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 |
| .byte 233,83,253,255,255 // jmpq 427f <_sk_load_f16_avx+0x37> |
| .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 |
| .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| .byte 233,70,253,255,255 // jmpq 427f <_sk_load_f16_avx+0x37> |
| .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1 |
| .byte 233,61,253,255,255 // jmpq 427f <_sk_load_f16_avx+0x37> |
| |
| HIDDEN _sk_gather_f16_avx |
| .globl _sk_gather_f16_avx |
| FUNCTION(_sk_gather_f16_avx) |
| _sk_gather_f16_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,131,236,24 // sub $0x18,%rsp |
| .byte 197,252,17,124,36,224 // vmovups %ymm7,-0x20(%rsp) |
| .byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp) |
| .byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp) |
| .byte 197,252,17,100,36,128 // vmovups %ymm4,-0x80(%rsp) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2 |
| .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1 |
| .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3 |
| .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1 |
| .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2 |
| .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 196,193,249,126,210 // vmovq %xmm2,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0 |
| .byte 196,225,249,126,195 // vmovq %xmm0,%rbx |
| .byte 65,137,222 // mov %ebx,%r14d |
| .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15 |
| .byte 69,137,252 // mov %r15d,%r12d |
| .byte 73,193,239,32 // shr $0x20,%r15 |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 196,193,122,126,4,216 // vmovq (%r8,%rbx,8),%xmm0 |
| .byte 196,129,122,126,12,240 // vmovq (%r8,%r14,8),%xmm1 |
| .byte 197,113,108,200 // vpunpcklqdq %xmm0,%xmm1,%xmm9 |
| .byte 196,129,122,126,12,248 // vmovq (%r8,%r15,8),%xmm1 |
| .byte 196,129,122,126,20,224 // vmovq (%r8,%r12,8),%xmm2 |
| .byte 197,233,108,201 // vpunpcklqdq %xmm1,%xmm2,%xmm1 |
| .byte 196,129,122,126,20,208 // vmovq (%r8,%r10,8),%xmm2 |
| .byte 196,129,122,126,28,216 // vmovq (%r8,%r11,8),%xmm3 |
| .byte 197,97,108,210 // vpunpcklqdq %xmm2,%xmm3,%xmm10 |
| .byte 196,65,122,126,4,192 // vmovq (%r8,%rax,8),%xmm8 |
| .byte 196,129,122,126,28,200 // vmovq (%r8,%r9,8),%xmm3 |
| .byte 196,193,97,108,216 // vpunpcklqdq %xmm8,%xmm3,%xmm3 |
| .byte 197,177,97,193 // vpunpcklwd %xmm1,%xmm9,%xmm0 |
| .byte 197,177,105,201 // vpunpckhwd %xmm1,%xmm9,%xmm1 |
| .byte 197,169,97,211 // vpunpcklwd %xmm3,%xmm10,%xmm2 |
| .byte 197,169,105,219 // vpunpckhwd %xmm3,%xmm10,%xmm3 |
| .byte 197,121,97,241 // vpunpcklwd %xmm1,%xmm0,%xmm14 |
| .byte 197,121,105,193 // vpunpckhwd %xmm1,%xmm0,%xmm8 |
| .byte 197,105,97,251 // vpunpcklwd %xmm3,%xmm2,%xmm15 |
| .byte 197,105,105,211 // vpunpckhwd %xmm3,%xmm2,%xmm10 |
| .byte 196,193,9,108,199 // vpunpcklqdq %xmm15,%xmm14,%xmm0 |
| .byte 196,65,25,239,228 // vpxor %xmm12,%xmm12,%xmm12 |
| .byte 196,193,121,105,212 // vpunpckhwd %xmm12,%xmm0,%xmm2 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0 |
| .byte 184,0,128,0,0 // mov $0x8000,%eax |
| .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 |
| .byte 196,99,109,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm9 |
| .byte 196,193,124,84,209 // vandps %ymm9,%ymm0,%ymm2 |
| .byte 184,0,124,0,0 // mov $0x7c00,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| .byte 196,99,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm11 |
| .byte 196,193,124,84,219 // vandps %ymm11,%ymm0,%ymm3 |
| .byte 197,252,87,194 // vxorps %ymm2,%ymm0,%ymm0 |
| .byte 196,227,125,25,217,1 // vextractf128 $0x1,%ymm3,%xmm1 |
| .byte 196,193,113,118,204 // vpcmpeqd %xmm12,%xmm1,%xmm1 |
| .byte 196,193,97,118,220 // vpcmpeqd %xmm12,%xmm3,%xmm3 |
| .byte 196,227,101,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm3,%ymm4 |
| .byte 196,227,125,25,211,1 // vextractf128 $0x1,%ymm2,%xmm3 |
| .byte 197,145,114,243,16 // vpslld $0x10,%xmm3,%xmm13 |
| .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3 |
| .byte 197,241,114,243,13 // vpslld $0xd,%xmm3,%xmm1 |
| .byte 184,0,0,0,56 // mov $0x38000000,%eax |
| .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| .byte 197,145,254,251 // vpaddd %xmm3,%xmm13,%xmm7 |
| .byte 197,193,254,201 // vpaddd %xmm1,%xmm7,%xmm1 |
| .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2 |
| .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 197,233,254,192 // vpaddd %xmm0,%xmm2,%xmm0 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| .byte 196,65,20,87,237 // vxorps %ymm13,%ymm13,%ymm13 |
| .byte 196,195,125,74,197,64 // vblendvps %ymm4,%ymm13,%ymm0,%ymm0 |
| .byte 196,193,9,109,207 // vpunpckhqdq %xmm15,%xmm14,%xmm1 |
| .byte 196,193,113,105,212 // vpunpckhwd %xmm12,%xmm1,%xmm2 |
| .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 196,193,116,84,209 // vandps %ymm9,%ymm1,%ymm2 |
| .byte 196,193,116,84,227 // vandps %ymm11,%ymm1,%ymm4 |
| .byte 197,244,87,202 // vxorps %ymm2,%ymm1,%ymm1 |
| .byte 196,227,125,25,231,1 // vextractf128 $0x1,%ymm4,%xmm7 |
| .byte 196,193,65,118,252 // vpcmpeqd %xmm12,%xmm7,%xmm7 |
| .byte 196,193,89,118,228 // vpcmpeqd %xmm12,%xmm4,%xmm4 |
| .byte 196,227,93,24,231,1 // vinsertf128 $0x1,%xmm7,%ymm4,%ymm4 |
| .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7 |
| .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 |
| .byte 196,227,125,25,206,1 // vextractf128 $0x1,%ymm1,%xmm6 |
| .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 |
| .byte 197,193,254,251 // vpaddd %xmm3,%xmm7,%xmm7 |
| .byte 197,193,254,246 // vpaddd %xmm6,%xmm7,%xmm6 |
| .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2 |
| .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1 |
| .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2 |
| .byte 197,233,254,201 // vpaddd %xmm1,%xmm2,%xmm1 |
| .byte 196,227,117,24,206,1 // vinsertf128 $0x1,%xmm6,%ymm1,%ymm1 |
| .byte 196,195,117,74,205,64 // vblendvps %ymm4,%ymm13,%ymm1,%ymm1 |
| .byte 196,193,57,108,210 // vpunpcklqdq %xmm10,%xmm8,%xmm2 |
| .byte 196,193,105,105,228 // vpunpckhwd %xmm12,%xmm2,%xmm4 |
| .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| .byte 196,227,109,24,212,1 // vinsertf128 $0x1,%xmm4,%ymm2,%ymm2 |
| .byte 196,193,108,84,227 // vandps %ymm11,%ymm2,%ymm4 |
| .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6 |
| .byte 196,193,73,118,244 // vpcmpeqd %xmm12,%xmm6,%xmm6 |
| .byte 196,193,89,118,228 // vpcmpeqd %xmm12,%xmm4,%xmm4 |
| .byte 196,227,93,24,230,1 // vinsertf128 $0x1,%xmm6,%ymm4,%ymm4 |
| .byte 196,193,108,84,241 // vandps %ymm9,%ymm2,%ymm6 |
| .byte 197,236,87,214 // vxorps %ymm6,%ymm2,%ymm2 |
| .byte 196,227,125,25,247,1 // vextractf128 $0x1,%ymm6,%xmm7 |
| .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 |
| .byte 196,227,125,25,213,1 // vextractf128 $0x1,%ymm2,%xmm5 |
| .byte 197,209,114,245,13 // vpslld $0xd,%xmm5,%xmm5 |
| .byte 197,193,254,251 // vpaddd %xmm3,%xmm7,%xmm7 |
| .byte 197,193,254,237 // vpaddd %xmm5,%xmm7,%xmm5 |
| .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6 |
| .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2 |
| .byte 197,201,254,243 // vpaddd %xmm3,%xmm6,%xmm6 |
| .byte 197,201,254,210 // vpaddd %xmm2,%xmm6,%xmm2 |
| .byte 196,227,109,24,213,1 // vinsertf128 $0x1,%xmm5,%ymm2,%ymm2 |
| .byte 196,195,109,74,213,64 // vblendvps %ymm4,%ymm13,%ymm2,%ymm2 |
| .byte 196,193,57,109,226 // vpunpckhqdq %xmm10,%xmm8,%xmm4 |
| .byte 196,193,89,105,236 // vpunpckhwd %xmm12,%xmm4,%xmm5 |
| .byte 196,226,121,51,228 // vpmovzxwd %xmm4,%xmm4 |
| .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4 |
| .byte 196,193,92,84,235 // vandps %ymm11,%ymm4,%ymm5 |
| .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6 |
| .byte 196,193,73,118,244 // vpcmpeqd %xmm12,%xmm6,%xmm6 |
| .byte 196,193,81,118,236 // vpcmpeqd %xmm12,%xmm5,%xmm5 |
| .byte 196,193,92,84,249 // vandps %ymm9,%ymm4,%ymm7 |
| .byte 197,220,87,231 // vxorps %ymm7,%ymm4,%ymm4 |
| .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5 |
| .byte 196,227,125,25,254,1 // vextractf128 $0x1,%ymm7,%xmm6 |
| .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6 |
| .byte 197,193,114,247,16 // vpslld $0x10,%xmm7,%xmm7 |
| .byte 197,193,254,251 // vpaddd %xmm3,%xmm7,%xmm7 |
| .byte 197,201,254,219 // vpaddd %xmm3,%xmm6,%xmm3 |
| .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6 |
| .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6 |
| .byte 197,225,254,222 // vpaddd %xmm6,%xmm3,%xmm3 |
| .byte 197,217,114,244,13 // vpslld $0xd,%xmm4,%xmm4 |
| .byte 197,193,254,228 // vpaddd %xmm4,%xmm7,%xmm4 |
| .byte 196,227,93,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm4,%ymm3 |
| .byte 196,195,101,74,221,80 // vblendvps %ymm5,%ymm13,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,16,100,36,128 // vmovups -0x80(%rsp),%ymm4 |
| .byte 197,252,16,108,36,160 // vmovups -0x60(%rsp),%ymm5 |
| .byte 197,252,16,116,36,192 // vmovups -0x40(%rsp),%ymm6 |
| .byte 197,252,16,124,36,224 // vmovups -0x20(%rsp),%ymm7 |
| .byte 72,131,196,24 // add $0x18,%rsp |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_f16_avx |
| .globl _sk_store_f16_avx |
| FUNCTION(_sk_store_f16_avx) |
| _sk_store_f16_avx: |
| .byte 72,131,236,88 // sub $0x58,%rsp |
| .byte 197,252,17,124,36,32 // vmovups %ymm7,0x20(%rsp) |
| .byte 197,252,17,52,36 // vmovups %ymm6,(%rsp) |
| .byte 197,252,17,108,36,224 // vmovups %ymm5,-0x20(%rsp) |
| .byte 197,252,17,100,36,192 // vmovups %ymm4,-0x40(%rsp) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 184,0,0,0,128 // mov $0x80000000,%eax |
| .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| .byte 196,67,61,24,200,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm9 |
| .byte 197,52,84,208 // vandps %ymm0,%ymm9,%ymm10 |
| .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp) |
| .byte 196,65,124,87,218 // vxorps %ymm10,%ymm0,%ymm11 |
| .byte 184,0,0,128,56 // mov $0x38800000,%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 196,65,36,194,224,1 // vcmpltps %ymm8,%ymm11,%ymm12 |
| .byte 196,67,125,25,213,1 // vextractf128 $0x1,%ymm10,%xmm13 |
| .byte 196,193,17,114,213,16 // vpsrld $0x10,%xmm13,%xmm13 |
| .byte 196,193,9,114,210,16 // vpsrld $0x10,%xmm10,%xmm14 |
| .byte 196,193,1,114,211,13 // vpsrld $0xd,%xmm11,%xmm15 |
| .byte 196,67,125,25,218,1 // vextractf128 $0x1,%ymm11,%xmm10 |
| .byte 196,193,33,114,210,13 // vpsrld $0xd,%xmm10,%xmm11 |
| .byte 184,0,192,1,0 // mov $0x1c000,%eax |
| .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| .byte 196,65,121,112,210,0 // vpshufd $0x0,%xmm10,%xmm10 |
| .byte 196,65,9,250,242 // vpsubd %xmm10,%xmm14,%xmm14 |
| .byte 196,65,17,250,234 // vpsubd %xmm10,%xmm13,%xmm13 |
| .byte 196,65,17,254,219 // vpaddd %xmm11,%xmm13,%xmm11 |
| .byte 196,65,9,254,239 // vpaddd %xmm15,%xmm14,%xmm13 |
| .byte 196,67,21,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm13,%ymm13 |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 196,99,21,74,224,192 // vblendvps %ymm12,%ymm0,%ymm13,%ymm12 |
| .byte 197,52,84,233 // vandps %ymm1,%ymm9,%ymm13 |
| .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp) |
| .byte 196,65,116,87,245 // vxorps %ymm13,%ymm1,%ymm14 |
| .byte 196,67,125,25,239,1 // vextractf128 $0x1,%ymm13,%xmm15 |
| .byte 196,193,1,114,215,16 // vpsrld $0x10,%xmm15,%xmm15 |
| .byte 196,67,125,25,243,1 // vextractf128 $0x1,%ymm14,%xmm11 |
| .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11 |
| .byte 196,193,1,250,250 // vpsubd %xmm10,%xmm15,%xmm7 |
| .byte 196,193,65,254,251 // vpaddd %xmm11,%xmm7,%xmm7 |
| .byte 196,193,73,114,213,16 // vpsrld $0x10,%xmm13,%xmm6 |
| .byte 196,193,73,250,242 // vpsubd %xmm10,%xmm6,%xmm6 |
| .byte 196,193,81,114,214,13 // vpsrld $0xd,%xmm14,%xmm5 |
| .byte 197,201,254,237 // vpaddd %xmm5,%xmm6,%xmm5 |
| .byte 196,193,12,194,240,1 // vcmpltps %ymm8,%ymm14,%ymm6 |
| .byte 196,227,85,24,239,1 // vinsertf128 $0x1,%xmm7,%ymm5,%ymm5 |
| .byte 196,99,85,74,232,96 // vblendvps %ymm6,%ymm0,%ymm5,%ymm13 |
| .byte 197,180,84,234 // vandps %ymm2,%ymm9,%ymm5 |
| .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6 |
| .byte 197,201,114,214,16 // vpsrld $0x10,%xmm6,%xmm6 |
| .byte 197,236,87,253 // vxorps %ymm5,%ymm2,%ymm7 |
| .byte 196,227,125,25,252,1 // vextractf128 $0x1,%ymm7,%xmm4 |
| .byte 197,217,114,212,13 // vpsrld $0xd,%xmm4,%xmm4 |
| .byte 196,193,73,250,242 // vpsubd %xmm10,%xmm6,%xmm6 |
| .byte 197,201,254,228 // vpaddd %xmm4,%xmm6,%xmm4 |
| .byte 197,209,114,213,16 // vpsrld $0x10,%xmm5,%xmm5 |
| .byte 196,193,81,250,234 // vpsubd %xmm10,%xmm5,%xmm5 |
| .byte 197,201,114,215,13 // vpsrld $0xd,%xmm7,%xmm6 |
| .byte 197,209,254,238 // vpaddd %xmm6,%xmm5,%xmm5 |
| .byte 196,227,85,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm5,%ymm4 |
| .byte 196,193,68,194,232,1 // vcmpltps %ymm8,%ymm7,%ymm5 |
| .byte 196,227,93,74,224,80 // vblendvps %ymm5,%ymm0,%ymm4,%ymm4 |
| .byte 197,180,84,235 // vandps %ymm3,%ymm9,%ymm5 |
| .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6 |
| .byte 197,201,114,214,16 // vpsrld $0x10,%xmm6,%xmm6 |
| .byte 197,193,114,213,16 // vpsrld $0x10,%xmm5,%xmm7 |
| .byte 196,193,65,250,250 // vpsubd %xmm10,%xmm7,%xmm7 |
| .byte 196,193,73,250,242 // vpsubd %xmm10,%xmm6,%xmm6 |
| .byte 197,228,87,237 // vxorps %ymm5,%ymm3,%ymm5 |
| .byte 196,227,125,25,233,1 // vextractf128 $0x1,%ymm5,%xmm1 |
| .byte 197,241,114,209,13 // vpsrld $0xd,%xmm1,%xmm1 |
| .byte 197,201,254,201 // vpaddd %xmm1,%xmm6,%xmm1 |
| .byte 196,193,84,194,240,1 // vcmpltps %ymm8,%ymm5,%ymm6 |
| .byte 197,209,114,213,13 // vpsrld $0xd,%xmm5,%xmm5 |
| .byte 197,193,254,237 // vpaddd %xmm5,%xmm7,%xmm5 |
| .byte 196,227,85,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm5,%ymm1 |
| .byte 196,227,117,74,192,96 // vblendvps %ymm6,%ymm0,%ymm1,%ymm0 |
| .byte 196,99,125,25,225,1 // vextractf128 $0x1,%ymm12,%xmm1 |
| .byte 196,226,25,43,201 // vpackusdw %xmm1,%xmm12,%xmm1 |
| .byte 196,99,125,25,237,1 // vextractf128 $0x1,%ymm13,%xmm5 |
| .byte 196,226,17,43,237 // vpackusdw %xmm5,%xmm13,%xmm5 |
| .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6 |
| .byte 196,226,89,43,230 // vpackusdw %xmm6,%xmm4,%xmm4 |
| .byte 196,227,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm6 |
| .byte 196,226,121,43,198 // vpackusdw %xmm6,%xmm0,%xmm0 |
| .byte 197,241,97,245 // vpunpcklwd %xmm5,%xmm1,%xmm6 |
| .byte 197,241,105,205 // vpunpckhwd %xmm5,%xmm1,%xmm1 |
| .byte 197,217,97,232 // vpunpcklwd %xmm0,%xmm4,%xmm5 |
| .byte 197,217,105,192 // vpunpckhwd %xmm0,%xmm4,%xmm0 |
| .byte 197,73,98,221 // vpunpckldq %xmm5,%xmm6,%xmm11 |
| .byte 197,73,106,213 // vpunpckhdq %xmm5,%xmm6,%xmm10 |
| .byte 197,113,98,200 // vpunpckldq %xmm0,%xmm1,%xmm9 |
| .byte 197,113,106,192 // vpunpckhdq %xmm0,%xmm1,%xmm8 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 117,70 // jne 4aa5 <_sk_store_f16_avx+0x23d> |
| .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) |
| .byte 196,65,122,127,68,248,48 // vmovdqu %xmm8,0x30(%r8,%rdi,8) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0 |
| .byte 197,252,16,76,36,160 // vmovups -0x60(%rsp),%ymm1 |
| .byte 197,252,16,100,36,192 // vmovups -0x40(%rsp),%ymm4 |
| .byte 197,252,16,108,36,224 // vmovups -0x20(%rsp),%ymm5 |
| .byte 197,252,16,52,36 // vmovups (%rsp),%ymm6 |
| .byte 197,252,16,124,36,32 // vmovups 0x20(%rsp),%ymm7 |
| .byte 72,131,196,88 // add $0x58,%rsp |
| .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 116,201 // je 4a7a <_sk_store_f16_avx+0x212> |
| .byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8) |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,188 // jb 4a7a <_sk_store_f16_avx+0x212> |
| .byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8) |
| .byte 116,179 // je 4a7a <_sk_store_f16_avx+0x212> |
| .byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8) |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,166 // jb 4a7a <_sk_store_f16_avx+0x212> |
| .byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8) |
| .byte 116,157 // je 4a7a <_sk_store_f16_avx+0x212> |
| .byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8) |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,144 // jb 4a7a <_sk_store_f16_avx+0x212> |
| .byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8) |
| .byte 235,135 // jmp 4a7a <_sk_store_f16_avx+0x212> |
| |
| HIDDEN _sk_load_u16_be_avx |
| .globl _sk_load_u16_be_avx |
| FUNCTION(_sk_load_u16_be_avx) |
| _sk_load_u16_be_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,5,1,0,0 // jne 4c0e <_sk_load_u16_be_avx+0x11b> |
| .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8 |
| .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2 |
| .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3 |
| .byte 196,65,122,111,76,64,48 // vmovdqu 0x30(%r8,%rax,2),%xmm9 |
| .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| .byte 197,121,97,210 // vpunpcklwd %xmm2,%xmm0,%xmm10 |
| .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,113,105,203 // vpunpckhwd %xmm3,%xmm1,%xmm9 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 |
| .byte 197,169,108,194 // vpunpcklqdq %xmm2,%xmm10,%xmm0 |
| .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1 |
| .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0 |
| .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 |
| .byte 196,65,33,239,219 // vpxor %xmm11,%xmm11,%xmm11 |
| .byte 196,193,121,105,203 // vpunpckhwd %xmm11,%xmm0,%xmm1 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 |
| .byte 197,169,109,202 // vpunpckhqdq %xmm2,%xmm10,%xmm1 |
| .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 |
| .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1 |
| .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1 |
| .byte 196,193,113,105,211 // vpunpckhwd %xmm11,%xmm1,%xmm2 |
| .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2 |
| .byte 197,169,113,242,8 // vpsllw $0x8,%xmm2,%xmm10 |
| .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2 |
| .byte 197,169,235,210 // vpor %xmm2,%xmm10,%xmm2 |
| .byte 196,65,105,105,211 // vpunpckhwd %xmm11,%xmm2,%xmm10 |
| .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| .byte 196,195,109,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| .byte 196,193,57,109,217 // vpunpckhqdq %xmm9,%xmm8,%xmm3 |
| .byte 197,185,113,243,8 // vpsllw $0x8,%xmm3,%xmm8 |
| .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3 |
| .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3 |
| .byte 196,65,97,105,195 // vpunpckhwd %xmm11,%xmm3,%xmm8 |
| .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3 |
| .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 |
| .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8 |
| .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 116,85 // je 4c74 <_sk_load_u16_be_avx+0x181> |
| .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,72 // jb 4c74 <_sk_load_u16_be_avx+0x181> |
| .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 116,72 // je 4c81 <_sk_load_u16_be_avx+0x18e> |
| .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,59 // jb 4c81 <_sk_load_u16_be_avx+0x18e> |
| .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 15,132,205,254,255,255 // je 4b24 <_sk_load_u16_be_avx+0x31> |
| .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 15,130,188,254,255,255 // jb 4b24 <_sk_load_u16_be_avx+0x31> |
| .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9 |
| .byte 233,176,254,255,255 // jmpq 4b24 <_sk_load_u16_be_avx+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| .byte 233,163,254,255,255 // jmpq 4b24 <_sk_load_u16_be_avx+0x31> |
| .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| .byte 233,154,254,255,255 // jmpq 4b24 <_sk_load_u16_be_avx+0x31> |
| |
| HIDDEN _sk_load_rgb_u16_be_avx |
| .globl _sk_load_rgb_u16_be_avx |
| FUNCTION(_sk_load_rgb_u16_be_avx) |
| _sk_load_rgb_u16_be_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,133,8,1,0,0 // jne 4da4 <_sk_load_rgb_u16_be_avx+0x11a> |
| .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0 |
| .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2 |
| .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1 |
| .byte 196,193,122,111,92,64,32 // vmovdqu 0x20(%r8,%rax,2),%xmm3 |
| .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3 |
| .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8 |
| .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9 |
| .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11 |
| .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10 |
| .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0 |
| .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2 |
| .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1 |
| .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3 |
| .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| .byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,208,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm10 |
| .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1 |
| .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0 |
| .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 |
| .byte 196,65,25,239,228 // vpxor %xmm12,%xmm12,%xmm12 |
| .byte 196,193,121,105,204 // vpunpckhwd %xmm12,%xmm0,%xmm1 |
| .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 |
| .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1 |
| .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1 |
| .byte 196,193,113,105,212 // vpunpckhwd %xmm12,%xmm1,%xmm2 |
| .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,193,49,108,211 // vpunpcklqdq %xmm11,%xmm9,%xmm2 |
| .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3 |
| .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2 |
| .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2 |
| .byte 196,193,105,105,220 // vpunpckhwd %xmm12,%xmm2,%xmm3 |
| .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 |
| .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%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 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0 |
| .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0 |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 117,5 // jne 4dbd <_sk_load_rgb_u16_be_avx+0x133> |
| .byte 233,19,255,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1 |
| .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,26 // jb 4dec <_sk_load_rgb_u16_be_avx+0x162> |
| .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1 |
| .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2 |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 117,10 // jne 4df1 <_sk_load_rgb_u16_be_avx+0x167> |
| .byte 233,228,254,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| .byte 233,223,254,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1 |
| .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,26 // jb 4e20 <_sk_load_rgb_u16_be_avx+0x196> |
| .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1 |
| .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1 |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 117,10 // jne 4e25 <_sk_load_rgb_u16_be_avx+0x19b> |
| .byte 233,176,254,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| .byte 233,171,254,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3 |
| .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11 |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,20 // jb 4e4e <_sk_load_rgb_u16_be_avx+0x1c4> |
| .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3 |
| .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3 |
| .byte 233,130,254,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| .byte 233,125,254,255,255 // jmpq 4cd0 <_sk_load_rgb_u16_be_avx+0x46> |
| |
| HIDDEN _sk_store_u16_be_avx |
| .globl _sk_store_u16_be_avx |
| FUNCTION(_sk_store_u16_be_avx) |
| _sk_store_u16_be_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 184,0,255,127,71 // mov $0x477fff00,%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 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 |
| .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9 |
| .byte 196,193,41,113,241,8 // vpsllw $0x8,%xmm9,%xmm10 |
| .byte 196,193,49,113,209,8 // vpsrlw $0x8,%xmm9,%xmm9 |
| .byte 196,65,41,235,201 // vpor %xmm9,%xmm10,%xmm9 |
| .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11 |
| .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10 |
| .byte 196,193,33,113,242,8 // vpsllw $0x8,%xmm10,%xmm11 |
| .byte 196,193,41,113,210,8 // vpsrlw $0x8,%xmm10,%xmm10 |
| .byte 196,65,33,235,210 // vpor %xmm10,%xmm11,%xmm10 |
| .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11 |
| .byte 196,65,125,91,219 // vcvtps2dq %ymm11,%ymm11 |
| .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 |
| .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11 |
| .byte 196,193,25,113,243,8 // vpsllw $0x8,%xmm11,%xmm12 |
| .byte 196,193,33,113,211,8 // vpsrlw $0x8,%xmm11,%xmm11 |
| .byte 196,65,25,235,219 // vpor %xmm11,%xmm12,%xmm11 |
| .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12 |
| .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8 |
| .byte 196,193,25,113,240,8 // vpsllw $0x8,%xmm8,%xmm12 |
| .byte 196,193,57,113,208,8 // vpsrlw $0x8,%xmm8,%xmm8 |
| .byte 196,65,25,235,192 // vpor %xmm8,%xmm12,%xmm8 |
| .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12 |
| .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13 |
| .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9 |
| .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8 |
| .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 |
| .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 117,31 // jne 4f55 <_sk_store_u16_be_avx+0x102> |
| .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2) |
| .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2) |
| .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2) |
| .byte 196,1,122,127,68,72,48 // vmovdqu %xmm8,0x30(%r8,%r9,2) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2) |
| .byte 72,131,249,1 // cmp $0x1,%rcx |
| .byte 116,240 // je 4f51 <_sk_store_u16_be_avx+0xfe> |
| .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2) |
| .byte 72,131,249,3 // cmp $0x3,%rcx |
| .byte 114,227 // jb 4f51 <_sk_store_u16_be_avx+0xfe> |
| .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2) |
| .byte 116,218 // je 4f51 <_sk_store_u16_be_avx+0xfe> |
| .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2) |
| .byte 72,131,249,5 // cmp $0x5,%rcx |
| .byte 114,205 // jb 4f51 <_sk_store_u16_be_avx+0xfe> |
| .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2) |
| .byte 116,196 // je 4f51 <_sk_store_u16_be_avx+0xfe> |
| .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2) |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 114,183 // jb 4f51 <_sk_store_u16_be_avx+0xfe> |
| .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2) |
| .byte 235,174 // jmp 4f51 <_sk_store_u16_be_avx+0xfe> |
| |
| HIDDEN _sk_load_f32_avx |
| .globl _sk_load_f32_avx |
| FUNCTION(_sk_load_f32_avx) |
| _sk_load_f32_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,131,249,7 // cmp $0x7,%rcx |
| .byte 119,110 // ja 5019 <_sk_load_f32_avx+0x76> |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 5044 <_sk_load_f32_avx+0xa1> |
| .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax |
| .byte 76,1,208 // add %r10,%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 196,3,125,24,68,136,112,1 // vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8 |
| .byte 196,131,125,24,92,136,96,1 // vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3 |
| .byte 196,131,125,24,76,136,80,1 // vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1 |
| .byte 196,131,125,24,84,136,64,1 // vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2 |
| .byte 196,129,121,16,68,136,48 // vmovupd 0x30(%r8,%r9,4),%xmm0 |
| .byte 196,195,125,13,192,12 // vblendpd $0xc,%ymm8,%ymm0,%ymm0 |
| .byte 196,1,121,16,68,136,32 // vmovupd 0x20(%r8,%r9,4),%xmm8 |
| .byte 196,99,61,13,203,12 // vblendpd $0xc,%ymm3,%ymm8,%ymm9 |
| .byte 196,129,121,16,92,136,16 // vmovupd 0x10(%r8,%r9,4),%xmm3 |
| .byte 196,99,101,13,209,12 // vblendpd $0xc,%ymm1,%ymm3,%ymm10 |
| .byte 196,129,121,16,12,136 // vmovupd (%r8,%r9,4),%xmm1 |
| .byte 196,227,117,13,202,12 // vblendpd $0xc,%ymm2,%ymm1,%ymm1 |
| .byte 196,193,116,20,210 // vunpcklps %ymm10,%ymm1,%ymm2 |
| .byte 196,193,116,21,218 // vunpckhps %ymm10,%ymm1,%ymm3 |
| .byte 197,180,20,200 // vunpcklps %ymm0,%ymm9,%ymm1 |
| .byte 197,52,21,192 // vunpckhps %ymm0,%ymm9,%ymm8 |
| .byte 197,237,20,193 // vunpcklpd %ymm1,%ymm2,%ymm0 |
| .byte 197,237,21,201 // vunpckhpd %ymm1,%ymm2,%ymm1 |
| .byte 196,193,101,20,208 // vunpcklpd %ymm8,%ymm3,%ymm2 |
| .byte 196,193,101,21,216 // vunpckhpd %ymm8,%ymm3,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| .byte 15,31,0 // nopl (%rax) |
| .byte 130 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,201 // dec %ecx |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 188,255,255,255,175 // mov $0xafffffff,%esp |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,162,255,255,255,154 // jmpq *-0x65000001(%rdx) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255,146,255,255,255,138 // callq *-0x75000001(%rdx) |
| .byte 255 // (bad) |
| .byte 255 // (bad) |
| .byte 255 // .byte 0xff |
| |
| HIDDEN _sk_store_f32_avx |
| .globl _sk_store_f32_avx |
| FUNCTION(_sk_store_f32_avx) |
| _sk_store_f32_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax |
| .byte 197,124,20,193 // vunpcklps %ymm1,%ymm0,%ymm8 |
| .byte 197,124,21,217 // vunpckhps %ymm1,%ymm0,%ymm11 |
| .byte 197,108,20,203 // vunpcklps %ymm3,%ymm2,%ymm9 |
| .byte 197,108,21,227 // vunpckhps %ymm3,%ymm2,%ymm12 |
| .byte 196,65,61,20,209 // vunpcklpd %ymm9,%ymm8,%ymm10 |
| .byte 196,65,61,21,201 // vunpckhpd %ymm9,%ymm8,%ymm9 |
| .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 50d1 <_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 |
| .byte 196,67,61,6,195,49 // vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 |
| .byte 196,65,125,17,36,128 // vmovupd %ymm12,(%r8,%rax,4) |
| .byte 196,65,125,17,108,128,32 // vmovupd %ymm13,0x20(%r8,%rax,4) |
| .byte 196,65,125,17,76,128,64 // vmovupd %ymm9,0x40(%r8,%rax,4) |
| .byte 196,65,125,17,68,128,96 // vmovupd %ymm8,0x60(%r8,%rax,4) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .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 50cd <_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 50cd <_sk_store_f32_avx+0x69> |
| .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) |
| .byte 116,218 // je 50cd <_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 50cd <_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 50cd <_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 50cd <_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 50cd <_sk_store_f32_avx+0x69> |
| |
| HIDDEN _sk_clamp_x_avx |
| .globl _sk_clamp_x_avx |
| FUNCTION(_sk_clamp_x_avx) |
| _sk_clamp_x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,60,95,200 // vmaxps %ymm0,%ymm8,%ymm9 |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 |
| .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0 |
| .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_y_avx |
| .globl _sk_clamp_y_avx |
| FUNCTION(_sk_clamp_y_avx) |
| _sk_clamp_y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,60,95,201 // vmaxps %ymm1,%ymm8,%ymm9 |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1 |
| .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 |
| .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 |
| .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_x_avx |
| .globl _sk_repeat_x_avx |
| FUNCTION(_sk_repeat_x_avx) |
| _sk_repeat_x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9 |
| .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 |
| .byte 196,65,124,92,201 // vsubps %ymm9,%ymm0,%ymm9 |
| .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 |
| .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0 |
| .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_y_avx |
| .globl _sk_repeat_y_avx |
| FUNCTION(_sk_repeat_y_avx) |
| _sk_repeat_y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9 |
| .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 |
| .byte 196,65,116,92,201 // vsubps %ymm9,%ymm1,%ymm9 |
| .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1 |
| .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 |
| .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 |
| .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_x_avx |
| .globl _sk_mirror_x_avx |
| FUNCTION(_sk_mirror_x_avx) |
| _sk_mirror_x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,121,110,0 // vmovd (%rax),%xmm8 |
| .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10 |
| .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8 |
| .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| .byte 197,172,92,192 // vsubps %ymm0,%ymm10,%ymm0 |
| .byte 196,193,124,92,193 // vsubps %ymm9,%ymm0,%ymm0 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8 |
| .byte 197,60,84,192 // vandps %ymm0,%ymm8,%ymm8 |
| .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 |
| .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 |
| .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 |
| .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 |
| .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_y_avx |
| .globl _sk_mirror_y_avx |
| FUNCTION(_sk_mirror_y_avx) |
| _sk_mirror_y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,121,110,0 // vmovd (%rax),%xmm8 |
| .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10 |
| .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%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 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8 |
| .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1 |
| .byte 196,193,116,92,201 // vsubps %ymm9,%ymm1,%ymm1 |
| .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8 |
| .byte 197,60,84,193 // vandps %ymm1,%ymm8,%ymm8 |
| .byte 196,99,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm1 |
| .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 |
| .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 |
| .byte 196,227,53,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm9,%ymm1 |
| .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_luminance_to_alpha_avx |
| .globl _sk_luminance_to_alpha_avx |
| FUNCTION(_sk_luminance_to_alpha_avx) |
| _sk_luminance_to_alpha_avx: |
| .byte 184,208,179,89,62 // mov $0x3e59b3d0,%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 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 |
| .byte 184,89,23,55,63 // mov $0x3f371759,%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 197,228,89,201 // vmulps %ymm1,%ymm3,%ymm1 |
| .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| .byte 184,152,221,147,61 // mov $0x3d93dd98,%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 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 |
| .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_2x3_avx |
| .globl _sk_matrix_2x3_avx |
| FUNCTION(_sk_matrix_2x3_avx) |
| _sk_matrix_2x3_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,98,125,24,72,8 // vbroadcastss 0x8(%rax),%ymm9 |
| .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| .byte 196,98,125,24,80,12 // vbroadcastss 0xc(%rax),%ymm10 |
| .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_3x4_avx |
| .globl _sk_matrix_3x4_avx |
| FUNCTION(_sk_matrix_3x4_avx) |
| _sk_matrix_3x4_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,98,125,24,88,36 // vbroadcastss 0x24(%rax),%ymm11 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| .byte 196,98,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm11 |
| .byte 196,98,125,24,96,40 // vbroadcastss 0x28(%rax),%ymm12 |
| .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 |
| .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12 |
| .byte 196,98,125,24,104,44 // vbroadcastss 0x2c(%rax),%ymm13 |
| .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| .byte 196,193,108,88,213 // vaddps %ymm13,%ymm2,%ymm2 |
| .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 |
| .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,252,88,209 // vaddps %ymm1,%ymm0,%ymm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_4x5_avx |
| .globl _sk_matrix_4x5_avx |
| FUNCTION(_sk_matrix_4x5_avx) |
| _sk_matrix_4x5_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,98,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm9 |
| .byte 196,98,125,24,80,32 // vbroadcastss 0x20(%rax),%ymm10 |
| .byte 196,98,125,24,88,48 // vbroadcastss 0x30(%rax),%ymm11 |
| .byte 196,98,125,24,96,64 // vbroadcastss 0x40(%rax),%ymm12 |
| .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 |
| .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| .byte 196,98,125,24,88,36 // vbroadcastss 0x24(%rax),%ymm11 |
| .byte 196,98,125,24,96,52 // vbroadcastss 0x34(%rax),%ymm12 |
| .byte 196,98,125,24,104,68 // vbroadcastss 0x44(%rax),%ymm13 |
| .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 |
| .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| .byte 196,98,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm11 |
| .byte 196,98,125,24,96,40 // vbroadcastss 0x28(%rax),%ymm12 |
| .byte 196,98,125,24,104,56 // vbroadcastss 0x38(%rax),%ymm13 |
| .byte 196,98,125,24,112,72 // vbroadcastss 0x48(%rax),%ymm14 |
| .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 |
| .byte 197,28,89,226 // vmulps %ymm2,%ymm12,%ymm12 |
| .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| .byte 197,36,89,217 // vmulps %ymm1,%ymm11,%ymm11 |
| .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| .byte 197,44,89,208 // vmulps %ymm0,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 196,98,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm11 |
| .byte 196,98,125,24,96,28 // vbroadcastss 0x1c(%rax),%ymm12 |
| .byte 196,98,125,24,104,44 // vbroadcastss 0x2c(%rax),%ymm13 |
| .byte 196,98,125,24,112,60 // vbroadcastss 0x3c(%rax),%ymm14 |
| .byte 196,98,125,24,120,76 // vbroadcastss 0x4c(%rax),%ymm15 |
| .byte 197,140,89,219 // vmulps %ymm3,%ymm14,%ymm3 |
| .byte 196,193,100,88,223 // vaddps %ymm15,%ymm3,%ymm3 |
| .byte 197,148,89,210 // vmulps %ymm2,%ymm13,%ymm2 |
| .byte 197,236,88,211 // vaddps %ymm3,%ymm2,%ymm2 |
| .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1 |
| .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0 |
| .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| .byte 197,124,41,210 // vmovaps %ymm10,%ymm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_perspective_avx |
| .globl _sk_matrix_perspective_avx |
| FUNCTION(_sk_matrix_perspective_avx) |
| _sk_matrix_perspective_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| .byte 196,98,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm11 |
| .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12 |
| .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 |
| .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 |
| .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| .byte 197,252,83,200 // vrcpps %ymm0,%ymm1 |
| .byte 197,188,89,193 // vmulps %ymm1,%ymm8,%ymm0 |
| .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_avx |
| .globl _sk_linear_gradient_avx |
| FUNCTION(_sk_linear_gradient_avx) |
| _sk_linear_gradient_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| .byte 196,226,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm1 |
| .byte 196,226,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm2 |
| .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3 |
| .byte 76,139,0 // mov (%rax),%r8 |
| .byte 77,133,192 // test %r8,%r8 |
| .byte 15,132,146,0,0,0 // je 5685 <_sk_linear_gradient_avx+0xb8> |
| .byte 72,139,64,8 // mov 0x8(%rax),%rax |
| .byte 72,131,192,32 // add $0x20,%rax |
| .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 |
| .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 |
| .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| .byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11 |
| .byte 196,98,125,24,104,224 // vbroadcastss -0x20(%rax),%ymm13 |
| .byte 196,65,124,194,237,1 // vcmpltps %ymm13,%ymm0,%ymm13 |
| .byte 196,98,125,24,112,228 // vbroadcastss -0x1c(%rax),%ymm14 |
| .byte 196,67,13,74,228,208 // vblendvps %ymm13,%ymm12,%ymm14,%ymm12 |
| .byte 196,98,125,24,112,232 // vbroadcastss -0x18(%rax),%ymm14 |
| .byte 196,67,13,74,219,208 // vblendvps %ymm13,%ymm11,%ymm14,%ymm11 |
| .byte 196,98,125,24,112,236 // vbroadcastss -0x14(%rax),%ymm14 |
| .byte 196,67,13,74,210,208 // vblendvps %ymm13,%ymm10,%ymm14,%ymm10 |
| .byte 196,98,125,24,112,240 // vbroadcastss -0x10(%rax),%ymm14 |
| .byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9 |
| .byte 196,98,125,24,112,244 // vbroadcastss -0xc(%rax),%ymm14 |
| .byte 196,67,13,74,192,208 // vblendvps %ymm13,%ymm8,%ymm14,%ymm8 |
| .byte 196,98,125,24,112,248 // vbroadcastss -0x8(%rax),%ymm14 |
| .byte 196,227,13,74,201,208 // vblendvps %ymm13,%ymm1,%ymm14,%ymm1 |
| .byte 196,98,125,24,112,252 // vbroadcastss -0x4(%rax),%ymm14 |
| .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2 |
| .byte 196,98,125,24,48 // vbroadcastss (%rax),%ymm14 |
| .byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3 |
| .byte 72,131,192,36 // add $0x24,%rax |
| .byte 73,255,200 // dec %r8 |
| .byte 117,140 // jne 560f <_sk_linear_gradient_avx+0x42> |
| .byte 235,20 // jmp 5699 <_sk_linear_gradient_avx+0xcc> |
| .byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11 |
| .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 |
| .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12 |
| .byte 197,28,89,224 // vmulps %ymm0,%ymm12,%ymm12 |
| .byte 196,65,60,88,196 // vaddps %ymm12,%ymm8,%ymm8 |
| .byte 197,36,89,216 // vmulps %ymm0,%ymm11,%ymm11 |
| .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1 |
| .byte 197,44,89,208 // vmulps %ymm0,%ymm10,%ymm10 |
| .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| .byte 197,252,88,219 // vaddps %ymm3,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_2stops_avx |
| .globl _sk_linear_gradient_2stops_avx |
| FUNCTION(_sk_linear_gradient_2stops_avx) |
| _sk_linear_gradient_2stops_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1 |
| .byte 196,226,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm2 |
| .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1 |
| .byte 197,116,88,194 // vaddps %ymm2,%ymm1,%ymm8 |
| .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| .byte 196,226,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm2 |
| .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1 |
| .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1 |
| .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| .byte 196,226,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm3 |
| .byte 197,236,89,208 // vmulps %ymm0,%ymm2,%ymm2 |
| .byte 197,236,88,211 // vaddps %ymm3,%ymm2,%ymm2 |
| .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3 |
| .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 |
| .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 |
| .byte 196,193,124,88,217 // vaddps %ymm9,%ymm0,%ymm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_save_xy_avx |
| .globl _sk_save_xy_avx |
| FUNCTION(_sk_save_xy_avx) |
| _sk_save_xy_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,88,200 // vaddps %ymm0,%ymm8,%ymm9 |
| .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10 |
| .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9 |
| .byte 197,60,88,193 // vaddps %ymm1,%ymm8,%ymm8 |
| .byte 196,67,125,8,208,1 // vroundps $0x1,%ymm8,%ymm10 |
| .byte 196,65,60,92,194 // vsubps %ymm10,%ymm8,%ymm8 |
| .byte 197,252,17,0 // vmovups %ymm0,(%rax) |
| .byte 197,252,17,72,32 // vmovups %ymm1,0x20(%rax) |
| .byte 197,124,17,72,64 // vmovups %ymm9,0x40(%rax) |
| .byte 197,124,17,64,96 // vmovups %ymm8,0x60(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_accumulate_avx |
| .globl _sk_accumulate_avx |
| FUNCTION(_sk_accumulate_avx) |
| _sk_accumulate_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 197,124,16,128,128,0,0,0 // vmovups 0x80(%rax),%ymm8 |
| .byte 197,60,89,128,160,0,0,0 // vmulps 0xa0(%rax),%ymm8,%ymm8 |
| .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| .byte 197,180,88,228 // vaddps %ymm4,%ymm9,%ymm4 |
| .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9 |
| .byte 197,180,88,237 // vaddps %ymm5,%ymm9,%ymm5 |
| .byte 197,60,89,202 // vmulps %ymm2,%ymm8,%ymm9 |
| .byte 197,180,88,246 // vaddps %ymm6,%ymm9,%ymm6 |
| .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| .byte 197,188,88,255 // vaddps %ymm7,%ymm8,%ymm7 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_nx_avx |
| .globl _sk_bilinear_nx_avx |
| FUNCTION(_sk_bilinear_nx_avx) |
| _sk_bilinear_nx_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_px_avx |
| .globl _sk_bilinear_px_avx |
| FUNCTION(_sk_bilinear_px_avx) |
| _sk_bilinear_px_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_ny_avx |
| .globl _sk_bilinear_ny_avx |
| FUNCTION(_sk_bilinear_ny_avx) |
| _sk_bilinear_ny_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_py_avx |
| .globl _sk_bilinear_py_avx |
| FUNCTION(_sk_bilinear_py_avx) |
| _sk_bilinear_py_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 |
| .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3x_avx |
| .globl _sk_bicubic_n3x_avx |
| FUNCTION(_sk_bicubic_n3x_avx) |
| _sk_bicubic_n3x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,44,89,192 // vmulps %ymm8,%ymm10,%ymm8 |
| .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 |
| .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1x_avx |
| .globl _sk_bicubic_n1x_avx |
| FUNCTION(_sk_bicubic_n1x_avx) |
| _sk_bicubic_n1x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,200 // vmovd %r8d,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 |
| .byte 196,65,44,88,201 // vaddps %ymm9,%ymm10,%ymm9 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1x_avx |
| .globl _sk_bicubic_p1x_avx |
| FUNCTION(_sk_bicubic_p1x_avx) |
| _sk_bicubic_p1x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,99,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm8 |
| .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0 |
| .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 |
| .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 |
| .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3x_avx |
| .globl _sk_bicubic_p3x_avx |
| FUNCTION(_sk_bicubic_p3x_avx) |
| _sk_bicubic_p3x_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,193,121,110,192 // vmovd %r8d,%xmm0 |
| .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0 |
| .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 |
| .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 |
| .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 |
| .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3y_avx |
| .globl _sk_bicubic_n3y_avx |
| FUNCTION(_sk_bicubic_n3y_avx) |
| _sk_bicubic_n3y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,44,89,192 // vmulps %ymm8,%ymm10,%ymm8 |
| .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 |
| .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1y_avx |
| .globl _sk_bicubic_n1y_avx |
| FUNCTION(_sk_bicubic_n1y_avx) |
| _sk_bicubic_n1y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d |
| .byte 196,65,121,110,192 // vmovd %r8d,%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 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,200 // vmovd %r8d,%xmm9 |
| .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 |
| .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9 |
| .byte 196,65,44,88,201 // vaddps %ymm9,%ymm10,%ymm9 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1y_avx |
| .globl _sk_bicubic_p1y_avx |
| FUNCTION(_sk_bicubic_p1y_avx) |
| _sk_bicubic_p1y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| .byte 196,99,117,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm8 |
| .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1 |
| .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9 |
| .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 |
| .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10 |
| .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8 |
| .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 |
| .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3y_avx |
| .globl _sk_bicubic_p3y_avx |
| FUNCTION(_sk_bicubic_p3y_avx) |
| _sk_bicubic_p3y_avx: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d |
| .byte 196,193,121,110,200 // vmovd %r8d,%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 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1 |
| .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8 |
| .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9 |
| .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d |
| .byte 196,65,121,110,208 // vmovd %r8d,%xmm10 |
| .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d |
| .byte 196,65,121,110,216 // vmovd %r8d,%xmm11 |
| .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8 |
| .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8 |
| .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8 |
| .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_callback_avx |
| .globl _sk_callback_avx |
| FUNCTION(_sk_callback_avx) |
| _sk_callback_avx: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 72,129,236,24,1,0,0 // sub $0x118,%rsp |
| .byte 197,252,17,188,36,224,0,0,0 // vmovups %ymm7,0xe0(%rsp) |
| .byte 197,252,17,180,36,192,0,0,0 // vmovups %ymm6,0xc0(%rsp) |
| .byte 197,252,17,172,36,160,0,0,0 // vmovups %ymm5,0xa0(%rsp) |
| .byte 197,252,17,164,36,128,0,0,0 // vmovups %ymm4,0x80(%rsp) |
| .byte 197,252,17,92,36,96 // vmovups %ymm3,0x60(%rsp) |
| .byte 197,252,17,84,36,64 // vmovups %ymm2,0x40(%rsp) |
| .byte 197,252,17,76,36,32 // vmovups %ymm1,0x20(%rsp) |
| .byte 197,252,17,4,36 // vmovups %ymm0,(%rsp) |
| .byte 72,137,203 // mov %rcx,%rbx |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 73,137,255 // mov %rdi,%r15 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,244 // mov %rsi,%r12 |
| .byte 72,139,120,8 // mov 0x8(%rax),%rdi |
| .byte 72,133,219 // test %rbx,%rbx |
| .byte 190,8,0,0,0 // mov $0x8,%esi |
| .byte 15,69,243 // cmovne %ebx,%esi |
| .byte 197,248,119 // vzeroupper |
| .byte 255,16 // callq *(%rax) |
| .byte 76,137,230 // mov %r12,%rsi |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,255 // mov %r15,%rdi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 72,137,217 // mov %rbx,%rcx |
| .byte 197,252,16,4,36 // vmovups (%rsp),%ymm0 |
| .byte 197,252,16,76,36,32 // vmovups 0x20(%rsp),%ymm1 |
| .byte 197,252,16,84,36,64 // vmovups 0x40(%rsp),%ymm2 |
| .byte 197,252,16,92,36,96 // vmovups 0x60(%rsp),%ymm3 |
| .byte 197,252,16,164,36,128,0,0,0 // vmovups 0x80(%rsp),%ymm4 |
| .byte 197,252,16,172,36,160,0,0,0 // vmovups 0xa0(%rsp),%ymm5 |
| .byte 197,252,16,180,36,192,0,0,0 // vmovups 0xc0(%rsp),%ymm6 |
| .byte 197,252,16,188,36,224,0,0,0 // vmovups 0xe0(%rsp),%ymm7 |
| .byte 72,129,196,24,1,0,0 // add $0x118,%rsp |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_start_pipeline_sse41 |
| .globl _sk_start_pipeline_sse41 |
| FUNCTION(_sk_start_pipeline_sse41) |
| _sk_start_pipeline_sse41: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 73,137,207 // mov %rcx,%r15 |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 72,137,251 // mov %rdi,%rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,196 // mov %rax,%r12 |
| .byte 73,137,245 // mov %rsi,%r13 |
| .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| .byte 76,57,248 // cmp %r15,%rax |
| .byte 118,5 // jbe 28 <_sk_start_pipeline_sse41+0x28> |
| .byte 72,137,216 // mov %rbx,%rax |
| .byte 235,52 // jmp 5c <_sk_start_pipeline_sse41+0x5c> |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 15,87,228 // xorps %xmm4,%xmm4 |
| .byte 15,87,237 // xorps %xmm5,%xmm5 |
| .byte 15,87,246 // xorps %xmm6,%xmm6 |
| .byte 15,87,255 // xorps %xmm7,%xmm7 |
| .byte 72,137,223 // mov %rbx,%rdi |
| .byte 76,137,238 // mov %r13,%rsi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 65,255,212 // callq *%r12 |
| .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| .byte 72,131,195,8 // add $0x8,%rbx |
| .byte 76,57,251 // cmp %r15,%rbx |
| .byte 72,137,195 // mov %rax,%rbx |
| .byte 118,204 // jbe 28 <_sk_start_pipeline_sse41+0x28> |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 195 // retq |
| |
| HIDDEN _sk_just_return_sse41 |
| .globl _sk_just_return_sse41 |
| FUNCTION(_sk_just_return_sse41) |
| _sk_just_return_sse41: |
| .byte 195 // retq |
| |
| HIDDEN _sk_seed_shader_sse41 |
| .globl _sk_seed_shader_sse41 |
| FUNCTION(_sk_seed_shader_sse41) |
| _sk_seed_shader_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 102,15,110,199 // movd %edi,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 15,16,2 // movups (%rdx),%xmm0 |
| .byte 15,88,193 // addps %xmm1,%xmm0 |
| .byte 102,15,110,8 // movd (%rax),%xmm1 |
| .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 15,87,228 // xorps %xmm4,%xmm4 |
| .byte 15,87,237 // xorps %xmm5,%xmm5 |
| .byte 15,87,246 // xorps %xmm6,%xmm6 |
| .byte 15,87,255 // xorps %xmm7,%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_constant_color_sse41 |
| .globl _sk_constant_color_sse41 |
| FUNCTION(_sk_constant_color_sse41) |
| _sk_constant_color_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clear_sse41 |
| .globl _sk_clear_sse41 |
| FUNCTION(_sk_clear_sse41) |
| _sk_clear_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcatop_sse41 |
| .globl _sk_srcatop_sse41 |
| FUNCTION(_sk_srcatop_sse41) |
| _sk_srcatop_sse41: |
| .byte 15,89,199 // mulps %xmm7,%xmm0 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| .byte 15,89,207 // mulps %xmm7,%xmm1 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstatop_sse41 |
| .globl _sk_dstatop_sse41 |
| FUNCTION(_sk_dstatop_sse41) |
| _sk_dstatop_sse41: |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,198 // mulps %xmm6,%xmm8 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 65,15,88,208 // addps %xmm8,%xmm2 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcin_sse41 |
| .globl _sk_srcin_sse41 |
| FUNCTION(_sk_srcin_sse41) |
| _sk_srcin_sse41: |
| .byte 15,89,199 // mulps %xmm7,%xmm0 |
| .byte 15,89,207 // mulps %xmm7,%xmm1 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstin_sse41 |
| .globl _sk_dstin_sse41 |
| FUNCTION(_sk_dstin_sse41) |
| _sk_dstin_sse41: |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 15,40,211 // movaps %xmm3,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcout_sse41 |
| .globl _sk_srcout_sse41 |
| FUNCTION(_sk_srcout_sse41) |
| _sk_srcout_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstout_sse41 |
| .globl _sk_dstout_sse41 |
| FUNCTION(_sk_dstout_sse41) |
| _sk_dstout_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcover_sse41 |
| .globl _sk_srcover_sse41 |
| FUNCTION(_sk_srcover_sse41) |
| _sk_srcover_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstover_sse41 |
| .globl _sk_dstover_sse41 |
| FUNCTION(_sk_dstover_sse41) |
| _sk_dstover_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_modulate_sse41 |
| .globl _sk_modulate_sse41 |
| FUNCTION(_sk_modulate_sse41) |
| _sk_modulate_sse41: |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_multiply_sse41 |
| .globl _sk_multiply_sse41 |
| FUNCTION(_sk_multiply_sse41) |
| _sk_multiply_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,220 // mulps %xmm4,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 65,15,88,195 // addps %xmm11,%xmm0 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,221 // mulps %xmm5,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,222 // mulps %xmm6,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_plus__sse41 |
| .globl _sk_plus__sse41 |
| FUNCTION(_sk_plus__sse41) |
| _sk_plus__sse41: |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_screen_sse41 |
| .globl _sk_screen_sse41 |
| FUNCTION(_sk_screen_sse41) |
| _sk_screen_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 68,15,88,196 // addps %xmm4,%xmm8 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,88,205 // addps %xmm5,%xmm9 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| .byte 68,15,88,214 // addps %xmm6,%xmm10 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,92,210 // subps %xmm2,%xmm10 |
| .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| .byte 68,15,88,223 // addps %xmm7,%xmm11 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_xor__sse41 |
| .globl _sk_xor__sse41 |
| FUNCTION(_sk_xor__sse41) |
| _sk_xor__sse41: |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .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 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,92,216 // subps %xmm8,%xmm3 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,212 // mulps %xmm4,%xmm10 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,213 // mulps %xmm5,%xmm10 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_darken_sse41 |
| .globl _sk_darken_sse41 |
| FUNCTION(_sk_darken_sse41) |
| _sk_darken_sse41: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 68,15,95,201 // maxps %xmm1,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,95,193 // maxps %xmm9,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,95,209 // maxps %xmm9,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lighten_sse41 |
| .globl _sk_lighten_sse41 |
| FUNCTION(_sk_lighten_sse41) |
| _sk_lighten_sse41: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_difference_sse41 |
| .globl _sk_difference_sse41 |
| FUNCTION(_sk_difference_sse41) |
| _sk_difference_sse41: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_exclusion_sse41 |
| .globl _sk_exclusion_sse41 |
| FUNCTION(_sk_exclusion_sse41) |
| _sk_exclusion_sse41: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 15,40,200 // movaps %xmm0,%xmm1 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colorburn_sse41 |
| .globl _sk_colorburn_sse41 |
| FUNCTION(_sk_colorburn_sse41) |
| _sk_colorburn_sse41: |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 15,40,199 // movaps %xmm7,%xmm0 |
| .byte 15,92,196 // subps %xmm4,%xmm0 |
| .byte 15,89,195 // mulps %xmm3,%xmm0 |
| .byte 65,15,94,193 // divps %xmm9,%xmm0 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,93,232 // minps %xmm0,%xmm13 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 |
| .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 102,69,15,56,20,241 // blendvps %xmm0,%xmm9,%xmm14 |
| .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 68,15,88,228 // addps %xmm4,%xmm12 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0 |
| .byte 69,15,88,206 // addps %xmm14,%xmm9 |
| .byte 102,69,15,56,20,204 // blendvps %xmm0,%xmm12,%xmm9 |
| .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| .byte 68,15,89,225 // mulps %xmm1,%xmm12 |
| .byte 15,40,199 // movaps %xmm7,%xmm0 |
| .byte 15,92,197 // subps %xmm5,%xmm0 |
| .byte 15,89,195 // mulps %xmm3,%xmm0 |
| .byte 15,94,193 // divps %xmm1,%xmm0 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,93,232 // minps %xmm0,%xmm13 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 15,40,193 // movaps %xmm1,%xmm0 |
| .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 |
| .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 102,68,15,56,20,241 // blendvps %xmm0,%xmm1,%xmm14 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 68,15,88,229 // addps %xmm5,%xmm12 |
| .byte 65,15,88,206 // addps %xmm14,%xmm1 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0 |
| .byte 102,65,15,56,20,204 // blendvps %xmm0,%xmm12,%xmm1 |
| .byte 15,40,199 // movaps %xmm7,%xmm0 |
| .byte 15,92,198 // subps %xmm6,%xmm0 |
| .byte 15,89,195 // mulps %xmm3,%xmm0 |
| .byte 15,94,194 // divps %xmm2,%xmm0 |
| .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| .byte 68,15,93,224 // minps %xmm0,%xmm12 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 68,15,194,194,0 // cmpeqps %xmm2,%xmm8 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 102,68,15,56,20,234 // blendvps %xmm0,%xmm2,%xmm13 |
| .byte 68,15,88,222 // addps %xmm6,%xmm11 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 65,15,88,213 // addps %xmm13,%xmm2 |
| .byte 15,40,198 // movaps %xmm6,%xmm0 |
| .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0 |
| .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colordodge_sse41 |
| .globl _sk_colordodge_sse41 |
| FUNCTION(_sk_colordodge_sse41) |
| _sk_colordodge_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| .byte 69,15,89,224 // mulps %xmm8,%xmm12 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 65,15,92,192 // subps %xmm8,%xmm0 |
| .byte 68,15,94,200 // divps %xmm0,%xmm9 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 69,15,93,241 // minps %xmm9,%xmm14 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0 |
| .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 102,69,15,56,20,240 // blendvps %xmm0,%xmm8,%xmm14 |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| .byte 68,15,88,228 // addps %xmm4,%xmm12 |
| .byte 69,15,88,198 // addps %xmm14,%xmm8 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0 |
| .byte 102,69,15,56,20,196 // blendvps %xmm0,%xmm12,%xmm8 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 68,15,89,229 // mulps %xmm5,%xmm12 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 68,15,94,224 // divps %xmm0,%xmm12 |
| .byte 69,15,40,243 // movaps %xmm11,%xmm14 |
| .byte 68,15,89,241 // mulps %xmm1,%xmm14 |
| .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| .byte 15,40,193 // movaps %xmm1,%xmm0 |
| .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 68,15,88,245 // addps %xmm5,%xmm14 |
| .byte 65,15,88,205 // addps %xmm13,%xmm1 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0 |
| .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 68,15,89,230 // mulps %xmm6,%xmm12 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,92,194 // subps %xmm2,%xmm0 |
| .byte 68,15,94,224 // divps %xmm0,%xmm12 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| .byte 102,68,15,56,20,234 // blendvps %xmm0,%xmm2,%xmm13 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 65,15,88,213 // addps %xmm13,%xmm2 |
| .byte 68,15,194,206,0 // cmpeqps %xmm6,%xmm9 |
| .byte 68,15,88,222 // addps %xmm6,%xmm11 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hardlight_sse41 |
| .globl _sk_hardlight_sse41 |
| FUNCTION(_sk_hardlight_sse41) |
| _sk_hardlight_sse41: |
| .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp) |
| .byte 68,15,40,229 // movaps %xmm5,%xmm12 |
| .byte 15,40,244 // movaps %xmm4,%xmm6 |
| .byte 15,40,227 // movaps %xmm3,%xmm4 |
| .byte 15,40,234 // movaps %xmm2,%xmm5 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,92,215 // subps %xmm7,%xmm2 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 68,15,92,212 // subps %xmm4,%xmm10 |
| .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| .byte 68,15,89,198 // mulps %xmm6,%xmm8 |
| .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| .byte 68,15,40,252 // movaps %xmm4,%xmm15 |
| .byte 69,15,92,249 // subps %xmm9,%xmm15 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 15,40,199 // movaps %xmm7,%xmm0 |
| .byte 15,92,198 // subps %xmm6,%xmm0 |
| .byte 65,15,89,199 // mulps %xmm15,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 68,15,40,251 // movaps %xmm3,%xmm15 |
| .byte 68,15,92,248 // subps %xmm0,%xmm15 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,196,2 // cmpleps %xmm4,%xmm0 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| .byte 102,69,15,56,20,249 // blendvps %xmm0,%xmm9,%xmm15 |
| .byte 68,15,40,218 // movaps %xmm2,%xmm11 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 69,15,40,204 // movaps %xmm12,%xmm9 |
| .byte 69,15,92,233 // subps %xmm9,%xmm13 |
| .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| .byte 15,40,193 // movaps %xmm1,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,196,2 // cmpleps %xmm4,%xmm0 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 69,15,40,233 // movaps %xmm9,%xmm13 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 102,68,15,56,20,225 // blendvps %xmm0,%xmm1,%xmm12 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 69,15,88,199 // addps %xmm15,%xmm8 |
| .byte 65,15,89,205 // mulps %xmm13,%xmm1 |
| .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| .byte 65,15,88,204 // addps %xmm12,%xmm1 |
| .byte 15,89,213 // mulps %xmm5,%xmm2 |
| .byte 68,15,40,92,36,232 // movaps -0x18(%rsp),%xmm11 |
| .byte 69,15,89,203 // mulps %xmm11,%xmm9 |
| .byte 68,15,88,202 // addps %xmm2,%xmm9 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,196,2 // cmpleps %xmm4,%xmm0 |
| .byte 15,40,212 // movaps %xmm4,%xmm2 |
| .byte 15,92,213 // subps %xmm5,%xmm2 |
| .byte 65,15,89,235 // mulps %xmm11,%xmm5 |
| .byte 15,88,237 // addps %xmm5,%xmm5 |
| .byte 69,15,92,243 // subps %xmm11,%xmm14 |
| .byte 68,15,89,242 // mulps %xmm2,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 65,15,92,222 // subps %xmm14,%xmm3 |
| .byte 102,15,56,20,221 // blendvps %xmm0,%xmm5,%xmm3 |
| .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| .byte 65,15,88,226 // addps %xmm10,%xmm4 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 15,40,220 // movaps %xmm4,%xmm3 |
| .byte 15,40,230 // movaps %xmm6,%xmm4 |
| .byte 65,15,40,237 // movaps %xmm13,%xmm5 |
| .byte 65,15,40,243 // movaps %xmm11,%xmm6 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_overlay_sse41 |
| .globl _sk_overlay_sse41 |
| FUNCTION(_sk_overlay_sse41) |
| _sk_overlay_sse41: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,240 // movaps %xmm0,%xmm14 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| .byte 65,15,89,198 // mulps %xmm14,%xmm0 |
| .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| .byte 68,15,40,235 // movaps %xmm3,%xmm13 |
| .byte 69,15,92,238 // subps %xmm14,%xmm13 |
| .byte 68,15,89,244 // mulps %xmm4,%xmm14 |
| .byte 15,40,207 // movaps %xmm7,%xmm1 |
| .byte 15,92,204 // subps %xmm4,%xmm1 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 68,15,89,231 // mulps %xmm7,%xmm12 |
| .byte 65,15,89,205 // mulps %xmm13,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 69,15,40,236 // movaps %xmm12,%xmm13 |
| .byte 68,15,92,233 // subps %xmm1,%xmm13 |
| .byte 102,69,15,56,20,238 // blendvps %xmm0,%xmm14,%xmm13 |
| .byte 69,15,88,197 // addps %xmm13,%xmm8 |
| .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 15,88,200 // addps %xmm0,%xmm1 |
| .byte 68,15,40,235 // movaps %xmm3,%xmm13 |
| .byte 69,15,92,233 // subps %xmm9,%xmm13 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 68,15,92,245 // subps %xmm5,%xmm14 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| .byte 69,15,89,245 // mulps %xmm13,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 69,15,40,236 // movaps %xmm12,%xmm13 |
| .byte 69,15,92,238 // subps %xmm14,%xmm13 |
| .byte 102,69,15,56,20,233 // blendvps %xmm0,%xmm9,%xmm13 |
| .byte 65,15,88,205 // addps %xmm13,%xmm1 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 69,15,88,203 // addps %xmm11,%xmm9 |
| .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| .byte 68,15,92,218 // subps %xmm2,%xmm11 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,92,238 // subps %xmm6,%xmm13 |
| .byte 15,40,198 // movaps %xmm6,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 69,15,89,235 // mulps %xmm11,%xmm13 |
| .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| .byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12 |
| .byte 69,15,88,204 // addps %xmm12,%xmm9 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_softlight_sse41 |
| .globl _sk_softlight_sse41 |
| FUNCTION(_sk_softlight_sse41) |
| _sk_softlight_sse41: |
| .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp) |
| .byte 15,40,244 // movaps %xmm4,%xmm6 |
| .byte 15,41,84,36,232 // movaps %xmm2,-0x18(%rsp) |
| .byte 68,15,40,225 // movaps %xmm1,%xmm12 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 15,87,228 // xorps %xmm4,%xmm4 |
| .byte 15,194,231,1 // cmpltps %xmm7,%xmm4 |
| .byte 15,40,198 // movaps %xmm6,%xmm0 |
| .byte 15,94,199 // divps %xmm7,%xmm0 |
| .byte 15,84,196 // andps %xmm4,%xmm0 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| .byte 68,15,92,208 // subps %xmm0,%xmm10 |
| .byte 68,15,40,240 // movaps %xmm0,%xmm14 |
| .byte 68,15,40,248 // movaps %xmm0,%xmm15 |
| .byte 15,82,208 // rsqrtps %xmm0,%xmm2 |
| .byte 68,15,83,218 // rcpps %xmm2,%xmm11 |
| .byte 68,15,92,216 // subps %xmm0,%xmm11 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,40,208 // movaps %xmm0,%xmm2 |
| .byte 15,89,210 // mulps %xmm2,%xmm2 |
| .byte 15,88,208 // addps %xmm0,%xmm2 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| .byte 68,15,92,241 // subps %xmm1,%xmm14 |
| .byte 68,15,89,242 // mulps %xmm2,%xmm14 |
| .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,89,253 // mulps %xmm13,%xmm15 |
| .byte 69,15,88,254 // addps %xmm14,%xmm15 |
| .byte 15,40,198 // movaps %xmm6,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| .byte 102,69,15,56,20,223 // blendvps %xmm0,%xmm15,%xmm11 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 15,92,195 // subps %xmm3,%xmm0 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 68,15,88,211 // addps %xmm3,%xmm10 |
| .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| .byte 15,40,211 // movaps %xmm3,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,89,199 // mulps %xmm7,%xmm0 |
| .byte 68,15,89,216 // mulps %xmm0,%xmm11 |
| .byte 68,15,88,218 // addps %xmm2,%xmm11 |
| .byte 68,15,194,203,2 // cmpleps %xmm3,%xmm9 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 102,69,15,56,20,218 // blendvps %xmm0,%xmm10,%xmm11 |
| .byte 68,15,40,213 // movaps %xmm5,%xmm10 |
| .byte 68,15,94,215 // divps %xmm7,%xmm10 |
| .byte 68,15,84,212 // andps %xmm4,%xmm10 |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 69,15,92,202 // subps %xmm10,%xmm9 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 15,89,192 // mulps %xmm0,%xmm0 |
| .byte 15,88,194 // addps %xmm2,%xmm0 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,92,209 // subps %xmm1,%xmm2 |
| .byte 15,89,208 // mulps %xmm0,%xmm2 |
| .byte 65,15,82,194 // rsqrtps %xmm10,%xmm0 |
| .byte 68,15,83,240 // rcpps %xmm0,%xmm14 |
| .byte 69,15,92,242 // subps %xmm10,%xmm14 |
| .byte 69,15,89,213 // mulps %xmm13,%xmm10 |
| .byte 68,15,88,210 // addps %xmm2,%xmm10 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| .byte 102,69,15,56,20,242 // blendvps %xmm0,%xmm10,%xmm14 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,40,208 // movaps %xmm0,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp) |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,89,242 // mulps %xmm2,%xmm14 |
| .byte 15,40,211 // movaps %xmm3,%xmm2 |
| .byte 15,89,213 // mulps %xmm5,%xmm2 |
| .byte 68,15,88,242 // addps %xmm2,%xmm14 |
| .byte 68,15,40,249 // movaps %xmm1,%xmm15 |
| .byte 15,194,195,2 // cmpleps %xmm3,%xmm0 |
| .byte 102,69,15,56,20,241 // blendvps %xmm0,%xmm9,%xmm14 |
| .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| .byte 15,40,108,36,216 // movaps -0x28(%rsp),%xmm5 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 15,94,199 // divps %xmm7,%xmm0 |
| .byte 15,84,196 // andps %xmm4,%xmm0 |
| .byte 15,40,208 // movaps %xmm0,%xmm2 |
| .byte 15,92,209 // subps %xmm1,%xmm2 |
| .byte 15,92,200 // subps %xmm0,%xmm1 |
| .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| .byte 15,82,224 // rsqrtps %xmm0,%xmm4 |
| .byte 68,15,83,204 // rcpps %xmm4,%xmm9 |
| .byte 68,15,92,200 // subps %xmm0,%xmm9 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,40,224 // movaps %xmm0,%xmm4 |
| .byte 15,89,228 // mulps %xmm4,%xmm4 |
| .byte 15,88,224 // addps %xmm0,%xmm4 |
| .byte 15,89,226 // mulps %xmm2,%xmm4 |
| .byte 68,15,88,236 // addps %xmm4,%xmm13 |
| .byte 15,40,197 // movaps %xmm5,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| .byte 102,69,15,56,20,205 // blendvps %xmm0,%xmm13,%xmm9 |
| .byte 68,15,40,108,36,232 // movaps -0x18(%rsp),%xmm13 |
| .byte 65,15,40,197 // movaps %xmm13,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 15,40,208 // movaps %xmm0,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,202 // mulps %xmm2,%xmm1 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 15,40,211 // movaps %xmm3,%xmm2 |
| .byte 15,89,213 // mulps %xmm5,%xmm2 |
| .byte 68,15,88,202 // addps %xmm2,%xmm9 |
| .byte 15,88,203 // addps %xmm3,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 15,40,213 // movaps %xmm5,%xmm2 |
| .byte 15,194,195,2 // cmpleps %xmm3,%xmm0 |
| .byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9 |
| .byte 68,15,92,255 // subps %xmm7,%xmm15 |
| .byte 69,15,89,199 // mulps %xmm15,%xmm8 |
| .byte 69,15,89,231 // mulps %xmm15,%xmm12 |
| .byte 69,15,89,253 // mulps %xmm13,%xmm15 |
| .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 15,89,198 // mulps %xmm6,%xmm0 |
| .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| .byte 69,15,88,195 // addps %xmm11,%xmm8 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5 |
| .byte 15,89,197 // mulps %xmm5,%xmm0 |
| .byte 68,15,88,224 // addps %xmm0,%xmm12 |
| .byte 69,15,88,230 // addps %xmm14,%xmm12 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 15,89,194 // mulps %xmm2,%xmm0 |
| .byte 65,15,88,199 // addps %xmm15,%xmm0 |
| .byte 68,15,88,200 // addps %xmm0,%xmm9 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,230 // movaps %xmm6,%xmm4 |
| .byte 15,40,242 // movaps %xmm2,%xmm6 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,204 // movaps %xmm12,%xmm1 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_0_sse41 |
| .globl _sk_clamp_0_sse41 |
| FUNCTION(_sk_clamp_0_sse41) |
| _sk_clamp_0_sse41: |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 65,15,95,192 // maxps %xmm8,%xmm0 |
| .byte 65,15,95,200 // maxps %xmm8,%xmm1 |
| .byte 65,15,95,208 // maxps %xmm8,%xmm2 |
| .byte 65,15,95,216 // maxps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_1_sse41 |
| .globl _sk_clamp_1_sse41 |
| FUNCTION(_sk_clamp_1_sse41) |
| _sk_clamp_1_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,93,192 // minps %xmm8,%xmm0 |
| .byte 65,15,93,200 // minps %xmm8,%xmm1 |
| .byte 65,15,93,208 // minps %xmm8,%xmm2 |
| .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_a_sse41 |
| .globl _sk_clamp_a_sse41 |
| FUNCTION(_sk_clamp_a_sse41) |
| _sk_clamp_a_sse41: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| .byte 15,93,195 // minps %xmm3,%xmm0 |
| .byte 15,93,203 // minps %xmm3,%xmm1 |
| .byte 15,93,211 // minps %xmm3,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_set_rgb_sse41 |
| .globl _sk_set_rgb_sse41 |
| FUNCTION(_sk_set_rgb_sse41) |
| _sk_set_rgb_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_rb_sse41 |
| .globl _sk_swap_rb_sse41 |
| FUNCTION(_sk_swap_rb_sse41) |
| _sk_swap_rb_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_sse41 |
| .globl _sk_swap_sse41 |
| FUNCTION(_sk_swap_sse41) |
| _sk_swap_sse41: |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,40,202 // movaps %xmm2,%xmm9 |
| .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| .byte 68,15,40,216 // movaps %xmm0,%xmm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,40,205 // movaps %xmm5,%xmm1 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 15,40,223 // movaps %xmm7,%xmm3 |
| .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| .byte 65,15,40,234 // movaps %xmm10,%xmm5 |
| .byte 65,15,40,241 // movaps %xmm9,%xmm6 |
| .byte 65,15,40,248 // movaps %xmm8,%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_src_dst_sse41 |
| .globl _sk_move_src_dst_sse41 |
| FUNCTION(_sk_move_src_dst_sse41) |
| _sk_move_src_dst_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,224 // movaps %xmm0,%xmm4 |
| .byte 15,40,233 // movaps %xmm1,%xmm5 |
| .byte 15,40,242 // movaps %xmm2,%xmm6 |
| .byte 15,40,251 // movaps %xmm3,%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_dst_src_sse41 |
| .globl _sk_move_dst_src_sse41 |
| FUNCTION(_sk_move_dst_src_sse41) |
| _sk_move_dst_src_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,40,205 // movaps %xmm5,%xmm1 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 15,40,223 // movaps %xmm7,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_premul_sse41 |
| .globl _sk_premul_sse41 |
| FUNCTION(_sk_premul_sse41) |
| _sk_premul_sse41: |
| .byte 15,89,195 // mulps %xmm3,%xmm0 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_unpremul_sse41 |
| .globl _sk_unpremul_sse41 |
| FUNCTION(_sk_unpremul_sse41) |
| _sk_unpremul_sse41: |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,94,203 // divps %xmm3,%xmm9 |
| .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8 |
| .byte 69,15,84,193 // andps %xmm9,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_srgb_sse41 |
| .globl _sk_from_srgb_sse41 |
| FUNCTION(_sk_from_srgb_sse41) |
| _sk_from_srgb_sse41: |
| .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 68,15,40,240 // movaps %xmm0,%xmm14 |
| .byte 69,15,89,246 // mulps %xmm14,%xmm14 |
| .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| .byte 69,15,88,204 // addps %xmm12,%xmm9 |
| .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,89,206 // mulps %xmm14,%xmm9 |
| .byte 69,15,88,205 // addps %xmm13,%xmm9 |
| .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| .byte 102,68,15,110,240 // movd %eax,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9 |
| .byte 69,15,40,251 // movaps %xmm11,%xmm15 |
| .byte 68,15,89,249 // mulps %xmm1,%xmm15 |
| .byte 15,40,193 // movaps %xmm1,%xmm0 |
| .byte 15,89,192 // mulps %xmm0,%xmm0 |
| .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 69,15,88,212 // addps %xmm12,%xmm10 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 69,15,88,213 // addps %xmm13,%xmm10 |
| .byte 65,15,194,206,1 // cmpltps %xmm14,%xmm1 |
| .byte 15,40,193 // movaps %xmm1,%xmm0 |
| .byte 102,69,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm10 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 15,89,192 // mulps %xmm0,%xmm0 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 69,15,88,196 // addps %xmm12,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 69,15,88,197 // addps %xmm13,%xmm8 |
| .byte 65,15,194,214,1 // cmpltps %xmm14,%xmm2 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_srgb_sse41 |
| .globl _sk_to_srgb_sse41 |
| FUNCTION(_sk_to_srgb_sse41) |
| _sk_to_srgb_sse41: |
| .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp) |
| .byte 15,40,254 // movaps %xmm6,%xmm7 |
| .byte 15,40,245 // movaps %xmm5,%xmm6 |
| .byte 15,40,236 // movaps %xmm4,%xmm5 |
| .byte 15,40,227 // movaps %xmm3,%xmm4 |
| .byte 15,40,218 // movaps %xmm2,%xmm3 |
| .byte 15,40,209 // movaps %xmm1,%xmm2 |
| .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 |
| .byte 69,15,83,200 // rcpps %xmm8,%xmm9 |
| .byte 69,15,82,248 // rsqrtps %xmm8,%xmm15 |
| .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| .byte 102,68,15,110,240 // movd %eax,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,89,205 // mulps %xmm13,%xmm9 |
| .byte 69,15,88,206 // addps %xmm14,%xmm9 |
| .byte 69,15,89,252 // mulps %xmm12,%xmm15 |
| .byte 69,15,88,249 // addps %xmm9,%xmm15 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 69,15,93,207 // minps %xmm15,%xmm9 |
| .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| .byte 102,68,15,110,248 // movd %eax,%xmm15 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 |
| .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9 |
| .byte 68,15,82,210 // rsqrtps %xmm2,%xmm10 |
| .byte 65,15,83,194 // rcpps %xmm10,%xmm0 |
| .byte 69,15,82,210 // rsqrtps %xmm10,%xmm10 |
| .byte 65,15,89,197 // mulps %xmm13,%xmm0 |
| .byte 65,15,88,198 // addps %xmm14,%xmm0 |
| .byte 69,15,89,212 // mulps %xmm12,%xmm10 |
| .byte 68,15,88,208 // addps %xmm0,%xmm10 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 65,15,93,202 // minps %xmm10,%xmm1 |
| .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| .byte 65,15,194,215,1 // cmpltps %xmm15,%xmm2 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 102,65,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm1 |
| .byte 15,82,195 // rsqrtps %xmm3,%xmm0 |
| .byte 15,83,208 // rcpps %xmm0,%xmm2 |
| .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| .byte 65,15,88,214 // addps %xmm14,%xmm2 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 65,15,89,196 // mulps %xmm12,%xmm0 |
| .byte 15,88,194 // addps %xmm2,%xmm0 |
| .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| .byte 68,15,89,219 // mulps %xmm3,%xmm11 |
| .byte 65,15,194,223,1 // cmpltps %xmm15,%xmm3 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 15,40,220 // movaps %xmm4,%xmm3 |
| .byte 15,40,229 // movaps %xmm5,%xmm4 |
| .byte 15,40,238 // movaps %xmm6,%xmm5 |
| .byte 15,40,247 // movaps %xmm7,%xmm6 |
| .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_2dot2_sse41 |
| .globl _sk_from_2dot2_sse41 |
| FUNCTION(_sk_from_2dot2_sse41) |
| _sk_from_2dot2_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 65,15,82,192 // rsqrtps %xmm8,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 68,15,82,200 // rsqrtps %xmm0,%xmm9 |
| .byte 65,15,82,193 // rsqrtps %xmm9,%xmm0 |
| .byte 68,15,82,208 // rsqrtps %xmm0,%xmm10 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 15,89,192 // mulps %xmm0,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| .byte 65,15,95,194 // maxps %xmm10,%xmm0 |
| .byte 68,15,82,193 // rsqrtps %xmm1,%xmm8 |
| .byte 69,15,82,192 // rsqrtps %xmm8,%xmm8 |
| .byte 69,15,82,192 // rsqrtps %xmm8,%xmm8 |
| .byte 69,15,82,200 // rsqrtps %xmm8,%xmm9 |
| .byte 69,15,82,193 // rsqrtps %xmm9,%xmm8 |
| .byte 69,15,82,216 // rsqrtps %xmm8,%xmm11 |
| .byte 15,89,201 // mulps %xmm1,%xmm1 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 69,15,89,195 // mulps %xmm11,%xmm8 |
| .byte 69,15,95,194 // maxps %xmm10,%xmm8 |
| .byte 15,82,202 // rsqrtps %xmm2,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 68,15,82,217 // rsqrtps %xmm1,%xmm11 |
| .byte 65,15,82,203 // rsqrtps %xmm11,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,89,210 // mulps %xmm2,%xmm2 |
| .byte 69,15,40,203 // movaps %xmm11,%xmm9 |
| .byte 69,15,89,201 // mulps %xmm9,%xmm9 |
| .byte 69,15,89,203 // mulps %xmm11,%xmm9 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 69,15,95,202 // maxps %xmm10,%xmm9 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_2dot2_sse41 |
| .globl _sk_to_2dot2_sse41 |
| FUNCTION(_sk_to_2dot2_sse41) |
| _sk_to_2dot2_sse41: |
| .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 |
| .byte 65,15,82,192 // rsqrtps %xmm8,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 68,15,82,200 // rsqrtps %xmm0,%xmm9 |
| .byte 69,15,83,192 // rcpps %xmm8,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 65,15,83,193 // rcpps %xmm9,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 65,15,95,192 // maxps %xmm8,%xmm0 |
| .byte 68,15,82,201 // rsqrtps %xmm1,%xmm9 |
| .byte 65,15,82,201 // rsqrtps %xmm9,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 68,15,82,209 // rsqrtps %xmm1,%xmm10 |
| .byte 69,15,83,201 // rcpps %xmm9,%xmm9 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 65,15,83,202 // rcpps %xmm10,%xmm1 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 65,15,95,200 // maxps %xmm8,%xmm1 |
| .byte 68,15,82,202 // rsqrtps %xmm2,%xmm9 |
| .byte 65,15,82,209 // rsqrtps %xmm9,%xmm2 |
| .byte 15,82,210 // rsqrtps %xmm2,%xmm2 |
| .byte 15,82,210 // rsqrtps %xmm2,%xmm2 |
| .byte 15,82,210 // rsqrtps %xmm2,%xmm2 |
| .byte 68,15,82,210 // rsqrtps %xmm2,%xmm10 |
| .byte 69,15,83,201 // rcpps %xmm9,%xmm9 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 65,15,83,210 // rcpps %xmm10,%xmm2 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 65,15,95,208 // maxps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_rgb_to_hsl_sse41 |
| .globl _sk_rgb_to_hsl_sse41 |
| FUNCTION(_sk_rgb_to_hsl_sse41) |
| _sk_rgb_to_hsl_sse41: |
| .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp) |
| .byte 15,40,254 // movaps %xmm6,%xmm7 |
| .byte 15,40,245 // movaps %xmm5,%xmm6 |
| .byte 15,40,236 // movaps %xmm4,%xmm5 |
| .byte 15,40,227 // movaps %xmm3,%xmm4 |
| .byte 15,40,218 // movaps %xmm2,%xmm3 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d |
| .byte 65,185,0,0,192,64 // mov $0x40c00000,%r9d |
| .byte 184,0,0,0,64 // mov $0x40000000,%eax |
| .byte 185,0,0,128,64 // mov $0x40800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 68,15,40,224 // movaps %xmm0,%xmm12 |
| .byte 68,15,95,225 // maxps %xmm1,%xmm12 |
| .byte 68,15,95,227 // maxps %xmm3,%xmm12 |
| .byte 68,15,40,232 // movaps %xmm0,%xmm13 |
| .byte 68,15,93,233 // minps %xmm1,%xmm13 |
| .byte 68,15,93,235 // minps %xmm3,%xmm13 |
| .byte 69,15,40,204 // movaps %xmm12,%xmm9 |
| .byte 68,15,194,200,0 // cmpeqps %xmm0,%xmm9 |
| .byte 68,15,40,241 // movaps %xmm1,%xmm14 |
| .byte 68,15,92,243 // subps %xmm3,%xmm14 |
| .byte 68,15,40,249 // movaps %xmm1,%xmm15 |
| .byte 68,15,194,251,1 // cmpltps %xmm3,%xmm15 |
| .byte 69,15,40,212 // movaps %xmm12,%xmm10 |
| .byte 68,15,194,209,0 // cmpeqps %xmm1,%xmm10 |
| .byte 15,92,216 // subps %xmm0,%xmm3 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 65,15,40,212 // movaps %xmm12,%xmm2 |
| .byte 65,15,92,213 // subps %xmm13,%xmm2 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,94,218 // divps %xmm2,%xmm11 |
| .byte 65,15,89,195 // mulps %xmm11,%xmm0 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 65,15,89,219 // mulps %xmm11,%xmm3 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 15,88,217 // addps %xmm1,%xmm3 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 102,68,15,56,20,195 // blendvps %xmm0,%xmm3,%xmm8 |
| .byte 69,15,89,243 // mulps %xmm11,%xmm14 |
| .byte 102,65,15,110,217 // movd %r9d,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,84,223 // andps %xmm15,%xmm3 |
| .byte 65,15,88,222 // addps %xmm14,%xmm3 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 102,68,15,56,20,195 // blendvps %xmm0,%xmm3,%xmm8 |
| .byte 65,15,40,220 // movaps %xmm12,%xmm3 |
| .byte 65,15,92,204 // subps %xmm12,%xmm1 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,40,204 // movaps %xmm12,%xmm9 |
| .byte 69,15,89,202 // mulps %xmm10,%xmm9 |
| .byte 69,15,194,209,1 // cmpltps %xmm9,%xmm10 |
| .byte 65,15,92,205 // subps %xmm13,%xmm1 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 102,68,15,56,20,225 // blendvps %xmm0,%xmm1,%xmm12 |
| .byte 65,15,194,221,4 // cmpneqps %xmm13,%xmm3 |
| .byte 102,65,15,110,192 // movd %r8d,%xmm0 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 68,15,84,195 // andps %xmm3,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 65,15,94,212 // divps %xmm12,%xmm2 |
| .byte 15,84,211 // andps %xmm3,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 15,40,202 // movaps %xmm2,%xmm1 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 15,40,220 // movaps %xmm4,%xmm3 |
| .byte 15,40,229 // movaps %xmm5,%xmm4 |
| .byte 15,40,238 // movaps %xmm6,%xmm5 |
| .byte 15,40,247 // movaps %xmm7,%xmm6 |
| .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hsl_to_rgb_sse41 |
| .globl _sk_hsl_to_rgb_sse41 |
| FUNCTION(_sk_hsl_to_rgb_sse41) |
| _sk_hsl_to_rgb_sse41: |
| .byte 72,131,236,24 // sub $0x18,%rsp |
| .byte 15,41,60,36 // movaps %xmm7,(%rsp) |
| .byte 15,41,116,36,240 // movaps %xmm6,-0x10(%rsp) |
| .byte 15,41,108,36,224 // movaps %xmm5,-0x20(%rsp) |
| .byte 15,41,100,36,208 // movaps %xmm4,-0x30(%rsp) |
| .byte 15,41,92,36,192 // movaps %xmm3,-0x40(%rsp) |
| .byte 68,15,40,208 // movaps %xmm0,%xmm10 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,41,92,36,128 // movaps %xmm3,-0x80(%rsp) |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 15,194,195,1 // cmpltps %xmm3,%xmm0 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,40,224 // movaps %xmm8,%xmm4 |
| .byte 15,88,225 // addps %xmm1,%xmm4 |
| .byte 15,89,226 // mulps %xmm2,%xmm4 |
| .byte 15,40,217 // movaps %xmm1,%xmm3 |
| .byte 15,40,249 // movaps %xmm1,%xmm7 |
| .byte 15,88,250 // addps %xmm2,%xmm7 |
| .byte 15,89,218 // mulps %xmm2,%xmm3 |
| .byte 15,40,234 // movaps %xmm2,%xmm5 |
| .byte 15,92,251 // subps %xmm3,%xmm7 |
| .byte 102,15,56,20,252 // blendvps %xmm0,%xmm4,%xmm7 |
| .byte 184,0,0,0,64 // mov $0x40000000,%eax |
| .byte 185,171,170,170,62 // mov $0x3eaaaaab,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,41,84,36,176 // movaps %xmm2,-0x50(%rsp) |
| .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| .byte 185,0,0,0,0 // mov $0x0,%ecx |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 15,194,194,1 // cmpltps %xmm2,%xmm0 |
| .byte 15,40,218 // movaps %xmm2,%xmm3 |
| .byte 65,15,92,216 // subps %xmm8,%xmm3 |
| .byte 68,15,40,226 // movaps %xmm2,%xmm12 |
| .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12 |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| .byte 68,15,41,116,36,160 // movaps %xmm14,-0x60(%rsp) |
| .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12 |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 15,40,229 // movaps %xmm5,%xmm4 |
| .byte 15,41,100,36,144 // movaps %xmm4,-0x70(%rsp) |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax |
| .byte 15,40,199 // movaps %xmm7,%xmm0 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 185,0,0,192,64 // mov $0x40c00000,%ecx |
| .byte 102,15,110,241 // movd %ecx,%xmm6 |
| .byte 15,198,246,0 // shufps $0x0,%xmm6,%xmm6 |
| .byte 15,89,240 // mulps %xmm0,%xmm6 |
| .byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,40,235 // movaps %xmm3,%xmm5 |
| .byte 65,15,92,236 // subps %xmm12,%xmm5 |
| .byte 69,15,40,236 // movaps %xmm12,%xmm13 |
| .byte 69,15,40,252 // movaps %xmm12,%xmm15 |
| .byte 68,15,194,227,1 // cmpltps %xmm3,%xmm12 |
| .byte 15,89,238 // mulps %xmm6,%xmm5 |
| .byte 65,15,88,233 // addps %xmm9,%xmm5 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 102,68,15,56,20,221 // blendvps %xmm0,%xmm5,%xmm11 |
| .byte 68,15,194,124,36,128,1 // cmpltps -0x80(%rsp),%xmm15 |
| .byte 65,15,40,199 // movaps %xmm15,%xmm0 |
| .byte 102,68,15,56,20,223 // blendvps %xmm0,%xmm7,%xmm11 |
| .byte 102,15,110,232 // movd %eax,%xmm5 |
| .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5 |
| .byte 68,15,194,237,1 // cmpltps %xmm5,%xmm13 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| .byte 65,15,40,197 // movaps %xmm13,%xmm0 |
| .byte 102,68,15,56,20,218 // blendvps %xmm0,%xmm2,%xmm11 |
| .byte 69,15,87,228 // xorps %xmm12,%xmm12 |
| .byte 68,15,194,225,0 // cmpeqps %xmm1,%xmm12 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 102,68,15,56,20,220 // blendvps %xmm0,%xmm4,%xmm11 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,194,194,1 // cmpltps %xmm10,%xmm0 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13 |
| .byte 68,15,40,243 // movaps %xmm3,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 65,15,40,229 // movaps %xmm13,%xmm4 |
| .byte 69,15,40,253 // movaps %xmm13,%xmm15 |
| .byte 68,15,194,235,1 // cmpltps %xmm3,%xmm13 |
| .byte 68,15,89,246 // mulps %xmm6,%xmm14 |
| .byte 69,15,88,241 // addps %xmm9,%xmm14 |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 65,15,40,197 // movaps %xmm13,%xmm0 |
| .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1 |
| .byte 68,15,40,116,36,128 // movaps -0x80(%rsp),%xmm14 |
| .byte 69,15,194,254,1 // cmpltps %xmm14,%xmm15 |
| .byte 65,15,40,199 // movaps %xmm15,%xmm0 |
| .byte 102,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm1 |
| .byte 15,194,229,1 // cmpltps %xmm5,%xmm4 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 65,15,89,210 // mulps %xmm10,%xmm2 |
| .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 68,15,40,124,36,144 // movaps -0x70(%rsp),%xmm15 |
| .byte 102,65,15,56,20,207 // blendvps %xmm0,%xmm15,%xmm1 |
| .byte 68,15,92,84,36,176 // subps -0x50(%rsp),%xmm10 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 65,15,92,208 // subps %xmm8,%xmm2 |
| .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| .byte 69,15,194,194,1 // cmpltps %xmm10,%xmm8 |
| .byte 65,15,40,226 // movaps %xmm10,%xmm4 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 102,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm4 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 15,194,68,36,160,1 // cmpltps -0x60(%rsp),%xmm0 |
| .byte 69,15,88,234 // addps %xmm10,%xmm13 |
| .byte 102,65,15,56,20,229 // blendvps %xmm0,%xmm13,%xmm4 |
| .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,194,195,1 // cmpltps %xmm3,%xmm0 |
| .byte 15,92,220 // subps %xmm4,%xmm3 |
| .byte 15,89,222 // mulps %xmm6,%xmm3 |
| .byte 69,15,88,209 // addps %xmm9,%xmm10 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 102,68,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm9 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| .byte 102,68,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm9 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,194,197,1 // cmpltps %xmm5,%xmm0 |
| .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 102,69,15,56,20,207 // blendvps %xmm0,%xmm15,%xmm9 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 15,40,92,36,192 // movaps -0x40(%rsp),%xmm3 |
| .byte 15,40,100,36,208 // movaps -0x30(%rsp),%xmm4 |
| .byte 15,40,108,36,224 // movaps -0x20(%rsp),%xmm5 |
| .byte 15,40,116,36,240 // movaps -0x10(%rsp),%xmm6 |
| .byte 15,40,60,36 // movaps (%rsp),%xmm7 |
| .byte 72,131,196,24 // add $0x18,%rsp |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_1_float_sse41 |
| .globl _sk_scale_1_float_sse41 |
| FUNCTION(_sk_scale_1_float_sse41) |
| _sk_scale_1_float_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_u8_sse41 |
| .globl _sk_scale_u8_sse41 |
| FUNCTION(_sk_scale_u8_sse41) |
| _sk_scale_u8_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_1_float_sse41 |
| .globl _sk_lerp_1_float_sse41 |
| FUNCTION(_sk_lerp_1_float_sse41) |
| _sk_lerp_1_float_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,92,196 // subps %xmm4,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,92,205 // subps %xmm5,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,92,214 // subps %xmm6,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,92,223 // subps %xmm7,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_u8_sse41 |
| .globl _sk_lerp_u8_sse41 |
| FUNCTION(_sk_lerp_u8_sse41) |
| _sk_lerp_u8_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 15,92,196 // subps %xmm4,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,92,205 // subps %xmm5,%xmm1 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,92,214 // subps %xmm6,%xmm2 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,92,223 // subps %xmm7,%xmm3 |
| .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_565_sse41 |
| .globl _sk_lerp_565_sse41 |
| FUNCTION(_sk_lerp_565_sse41) |
| _sk_lerp_565_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 102,68,15,56,51,4,120 // pmovzxwd (%rax,%rdi,2),%xmm8 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%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 15,92,196 // subps %xmm4,%xmm0 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,92,205 // subps %xmm5,%xmm1 |
| .byte 65,15,89,203 // mulps %xmm11,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,92,214 // subps %xmm6,%xmm2 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_tables_sse41 |
| .globl _sk_load_tables_sse41 |
| FUNCTION(_sk_load_tables_sse41) |
| _sk_load_tables_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 243,68,15,111,4,185 // movdqu (%rcx,%rdi,4),%xmm8 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 102,15,110,193 // movd %ecx,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,65,15,111,208 // movdqa %xmm8,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,65,15,219,192 // pand %xmm8,%xmm0 |
| .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx |
| .byte 65,137,201 // mov %ecx,%r9d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,73,15,126,194 // movq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0 |
| .byte 102,67,15,58,33,4,144,16 // insertps $0x10,(%r8,%r10,4),%xmm0 |
| .byte 102,67,15,58,33,4,136,32 // insertps $0x20,(%r8,%r9,4),%xmm0 |
| .byte 102,65,15,58,33,4,136,48 // insertps $0x30,(%r8,%rcx,4),%xmm0 |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 102,73,15,58,22,202,1 // pextrq $0x1,%xmm1,%r10 |
| .byte 77,137,209 // mov %r10,%r9 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 65,137,203 // mov %ecx,%r11d |
| .byte 65,129,227,255,255,255,0 // and $0xffffff,%r11d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 65,129,226,255,255,255,0 // and $0xffffff,%r10d |
| .byte 243,67,15,16,12,152 // movss (%r8,%r11,4),%xmm1 |
| .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1 |
| .byte 243,67,15,16,28,144 // movss (%r8,%r10,4),%xmm3 |
| .byte 102,15,58,33,203,32 // insertps $0x20,%xmm3,%xmm1 |
| .byte 243,67,15,16,28,136 // movss (%r8,%r9,4),%xmm3 |
| .byte 102,15,58,33,203,48 // insertps $0x30,%xmm3,%xmm1 |
| .byte 76,139,72,24 // mov 0x18(%rax),%r9 |
| .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx |
| .byte 68,15,183,193 // movzwl %cx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,72,15,126,208 // movq %xmm2,%rax |
| .byte 68,15,183,208 // movzwl %ax,%r10d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| .byte 102,65,15,58,33,20,1,16 // insertps $0x10,(%r9,%rax,1),%xmm2 |
| .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3 |
| .byte 102,15,58,33,211,32 // insertps $0x20,%xmm3,%xmm2 |
| .byte 243,65,15,16,28,137 // movss (%r9,%rcx,4),%xmm3 |
| .byte 102,15,58,33,211,48 // insertps $0x30,%xmm3,%xmm2 |
| .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%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_load_tables_u16_be_sse41 |
| .globl _sk_load_tables_u16_be_sse41 |
| FUNCTION(_sk_load_tables_u16_be_sse41) |
| _sk_load_tables_u16_be_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 243,15,111,4,249 // movdqu (%rcx,%rdi,8),%xmm0 |
| .byte 243,15,111,76,249,16 // movdqu 0x10(%rcx,%rdi,8),%xmm1 |
| .byte 102,15,111,216 // movdqa %xmm0,%xmm3 |
| .byte 102,15,97,217 // punpcklwd %xmm1,%xmm3 |
| .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,15,105,216 // punpckhwd %xmm0,%xmm3 |
| .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 102,68,15,112,194,0 // pshufd $0x0,%xmm2,%xmm8 |
| .byte 102,65,15,219,192 // pand %xmm8,%xmm0 |
| .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,30 // shr $0x1e,%r9 |
| .byte 243,67,15,16,4,144 // movss (%r8,%r10,4),%xmm0 |
| .byte 102,65,15,58,33,4,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm0 |
| .byte 102,67,15,58,33,4,152,32 // insertps $0x20,(%r8,%r11,4),%xmm0 |
| .byte 102,67,15,58,33,4,8,48 // insertps $0x30,(%r8,%r9,1),%xmm0 |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 102,15,115,217,8 // psrldq $0x8,%xmm1 |
| .byte 102,15,56,51,201 // pmovzxwd %xmm1,%xmm1 |
| .byte 102,65,15,219,200 // pand %xmm8,%xmm1 |
| .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,30 // shr $0x1e,%r9 |
| .byte 243,67,15,16,12,144 // movss (%r8,%r10,4),%xmm1 |
| .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1 |
| .byte 243,67,15,16,20,152 // movss (%r8,%r11,4),%xmm2 |
| .byte 102,15,58,33,202,32 // insertps $0x20,%xmm2,%xmm1 |
| .byte 243,67,15,16,20,8 // movss (%r8,%r9,1),%xmm2 |
| .byte 102,15,58,33,202,48 // insertps $0x30,%xmm2,%xmm1 |
| .byte 76,139,64,24 // mov 0x18(%rax),%r8 |
| .byte 102,15,56,51,211 // pmovzxwd %xmm3,%xmm2 |
| .byte 102,65,15,219,208 // pand %xmm8,%xmm2 |
| .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx |
| .byte 102,72,15,126,208 // movq %xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 243,67,15,16,20,136 // movss (%r8,%r9,4),%xmm2 |
| .byte 102,65,15,58,33,20,0,16 // insertps $0x10,(%r8,%rax,1),%xmm2 |
| .byte 102,67,15,58,33,20,144,32 // insertps $0x20,(%r8,%r10,4),%xmm2 |
| .byte 102,65,15,58,33,20,8,48 // insertps $0x30,(%r8,%rcx,1),%xmm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 102,15,112,219,78 // pshufd $0x4e,%xmm3,%xmm3 |
| .byte 102,68,15,111,203 // movdqa %xmm3,%xmm9 |
| .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,65,15,235,217 // por %xmm9,%xmm3 |
| .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_tables_rgb_u16_be_sse41 |
| .globl _sk_load_tables_rgb_u16_be_sse41 |
| FUNCTION(_sk_load_tables_rgb_u16_be_sse41) |
| _sk_load_tables_rgb_u16_be_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 243,66,15,111,20,73 // movdqu (%rcx,%r9,2),%xmm2 |
| .byte 243,66,15,111,68,73,8 // movdqu 0x8(%rcx,%r9,2),%xmm0 |
| .byte 102,15,115,216,4 // psrldq $0x4,%xmm0 |
| .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| .byte 102,15,115,217,6 // psrldq $0x6,%xmm1 |
| .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 |
| .byte 102,15,115,216,6 // psrldq $0x6,%xmm0 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,15,111,194 // movdqa %xmm2,%xmm0 |
| .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 |
| .byte 102,68,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm8 |
| .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2 |
| .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 102,68,15,112,201,0 // pshufd $0x0,%xmm1,%xmm9 |
| .byte 102,65,15,219,193 // pand %xmm9,%xmm0 |
| .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,30 // shr $0x1e,%r9 |
| .byte 243,67,15,16,4,144 // movss (%r8,%r10,4),%xmm0 |
| .byte 102,65,15,58,33,4,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm0 |
| .byte 102,67,15,58,33,4,152,32 // insertps $0x20,(%r8,%r11,4),%xmm0 |
| .byte 102,67,15,58,33,4,8,48 // insertps $0x30,(%r8,%r9,1),%xmm0 |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 102,65,15,56,51,200 // pmovzxwd %xmm8,%xmm1 |
| .byte 102,65,15,219,201 // pand %xmm9,%xmm1 |
| .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9 |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 69,137,203 // mov %r9d,%r11d |
| .byte 73,193,233,30 // shr $0x1e,%r9 |
| .byte 243,67,15,16,12,144 // movss (%r8,%r10,4),%xmm1 |
| .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1 |
| .byte 243,67,15,16,28,152 // movss (%r8,%r11,4),%xmm3 |
| .byte 102,15,58,33,203,32 // insertps $0x20,%xmm3,%xmm1 |
| .byte 243,67,15,16,28,8 // movss (%r8,%r9,1),%xmm3 |
| .byte 102,15,58,33,203,48 // insertps $0x30,%xmm3,%xmm1 |
| .byte 76,139,64,24 // mov 0x18(%rax),%r8 |
| .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2 |
| .byte 102,65,15,219,209 // pand %xmm9,%xmm2 |
| .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx |
| .byte 102,72,15,126,208 // movq %xmm2,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 243,67,15,16,20,136 // movss (%r8,%r9,4),%xmm2 |
| .byte 102,65,15,58,33,20,0,16 // insertps $0x10,(%r8,%rax,1),%xmm2 |
| .byte 102,67,15,58,33,20,144,32 // insertps $0x20,(%r8,%r10,4),%xmm2 |
| .byte 102,65,15,58,33,20,8,48 // insertps $0x30,(%r8,%rcx,1),%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_byte_tables_sse41 |
| .globl _sk_byte_tables_sse41 |
| FUNCTION(_sk_byte_tables_sse41) |
| _sk_byte_tables_sse41: |
| .byte 65,86 // push %r14 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 |
| .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,73,15,126,193 // movq %xmm0,%r9 |
| .byte 69,137,202 // mov %r9d,%r10d |
| .byte 77,137,203 // mov %r9,%r11 |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 76,139,48 // mov (%rax),%r14 |
| .byte 76,139,72,8 // mov 0x8(%rax),%r9 |
| .byte 102,67,15,58,32,4,22,0 // pinsrb $0x0,(%r14,%r10,1),%xmm0 |
| .byte 102,67,15,58,32,4,30,1 // pinsrb $0x1,(%r14,%r11,1),%xmm0 |
| .byte 67,15,182,28,6 // movzbl (%r14,%r8,1),%ebx |
| .byte 102,15,58,32,195,2 // pinsrb $0x2,%ebx,%xmm0 |
| .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx |
| .byte 102,15,58,32,193,3 // pinsrb $0x3,%ecx,%xmm0 |
| .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 185,129,128,128,59 // mov $0x3b808081,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 |
| .byte 102,72,15,58,22,201,1 // pextrq $0x1,%xmm1,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,72,15,126,203 // movq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 102,67,15,58,32,12,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm1 |
| .byte 102,65,15,58,32,12,25,1 // pinsrb $0x1,(%r9,%rbx,1),%xmm1 |
| .byte 67,15,182,28,1 // movzbl (%r9,%r8,1),%ebx |
| .byte 102,15,58,32,203,2 // pinsrb $0x2,%ebx,%xmm1 |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 102,15,58,32,201,3 // pinsrb $0x3,%ecx,%xmm1 |
| .byte 102,15,56,49,201 // pmovzxbd %xmm1,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 76,139,72,16 // mov 0x10(%rax),%r9 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2 |
| .byte 102,72,15,58,22,211,1 // pextrq $0x1,%xmm2,%rbx |
| .byte 65,137,216 // mov %ebx,%r8d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,58,32,20,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm2 |
| .byte 102,65,15,58,32,20,9,1 // pinsrb $0x1,(%r9,%rcx,1),%xmm2 |
| .byte 67,15,182,12,1 // movzbl (%r9,%r8,1),%ecx |
| .byte 102,15,58,32,209,2 // pinsrb $0x2,%ecx,%xmm2 |
| .byte 65,15,182,12,25 // movzbl (%r9,%rbx,1),%ecx |
| .byte 102,15,58,32,209,3 // pinsrb $0x3,%ecx,%xmm2 |
| .byte 102,15,56,49,210 // pmovzxbd %xmm2,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,65,15,91,216 // cvtps2dq %xmm8,%xmm3 |
| .byte 102,72,15,58,22,217,1 // pextrq $0x1,%xmm3,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,72,15,126,219 // movq %xmm3,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 102,66,15,58,32,28,8,0 // pinsrb $0x0,(%rax,%r9,1),%xmm3 |
| .byte 102,15,58,32,28,24,1 // pinsrb $0x1,(%rax,%rbx,1),%xmm3 |
| .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx |
| .byte 102,15,58,32,219,2 // pinsrb $0x2,%ebx,%xmm3 |
| .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax |
| .byte 102,15,58,32,216,3 // pinsrb $0x3,%eax,%xmm3 |
| .byte 102,15,56,49,219 // pmovzxbd %xmm3,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,94 // pop %r14 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_byte_tables_rgb_sse41 |
| .globl _sk_byte_tables_rgb_sse41 |
| FUNCTION(_sk_byte_tables_rgb_sse41) |
| _sk_byte_tables_rgb_sse41: |
| .byte 65,86 // push %r14 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 139,72,24 // mov 0x18(%rax),%ecx |
| .byte 255,201 // dec %ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 |
| .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,73,15,126,193 // movq %xmm0,%r9 |
| .byte 69,137,202 // mov %r9d,%r10d |
| .byte 77,137,203 // mov %r9,%r11 |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 76,139,48 // mov (%rax),%r14 |
| .byte 76,139,72,8 // mov 0x8(%rax),%r9 |
| .byte 102,67,15,58,32,4,22,0 // pinsrb $0x0,(%r14,%r10,1),%xmm0 |
| .byte 102,67,15,58,32,4,30,1 // pinsrb $0x1,(%r14,%r11,1),%xmm0 |
| .byte 67,15,182,28,6 // movzbl (%r14,%r8,1),%ebx |
| .byte 102,15,58,32,195,2 // pinsrb $0x2,%ebx,%xmm0 |
| .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx |
| .byte 102,15,58,32,193,3 // pinsrb $0x3,%ecx,%xmm0 |
| .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 185,129,128,128,59 // mov $0x3b808081,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 |
| .byte 102,72,15,58,22,201,1 // pextrq $0x1,%xmm1,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,72,15,126,203 // movq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 102,67,15,58,32,12,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm1 |
| .byte 102,65,15,58,32,12,25,1 // pinsrb $0x1,(%r9,%rbx,1),%xmm1 |
| .byte 67,15,182,28,1 // movzbl (%r9,%r8,1),%ebx |
| .byte 102,15,58,32,203,2 // pinsrb $0x2,%ebx,%xmm1 |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 102,15,58,32,201,3 // pinsrb $0x3,%ecx,%xmm1 |
| .byte 102,15,56,49,201 // pmovzxbd %xmm1,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 72,139,64,16 // mov 0x10(%rax),%rax |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2 |
| .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,72,15,126,211 // movq %xmm2,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 102,66,15,58,32,20,8,0 // pinsrb $0x0,(%rax,%r9,1),%xmm2 |
| .byte 102,15,58,32,20,24,1 // pinsrb $0x1,(%rax,%rbx,1),%xmm2 |
| .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx |
| .byte 102,15,58,32,211,2 // pinsrb $0x2,%ebx,%xmm2 |
| .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax |
| .byte 102,15,58,32,208,3 // pinsrb $0x3,%eax,%xmm2 |
| .byte 102,15,56,49,210 // pmovzxbd %xmm2,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,94 // pop %r14 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_r_sse41 |
| .globl _sk_table_r_sse41 |
| FUNCTION(_sk_table_r_sse41) |
| _sk_table_r_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 102,65,15,91,192 // cvtps2dq %xmm8,%xmm0 |
| .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,67,15,16,4,145 // movss (%r9,%r10,4),%xmm0 |
| .byte 102,65,15,58,33,4,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm0 |
| .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8 |
| .byte 102,65,15,58,33,192,32 // insertps $0x20,%xmm8,%xmm0 |
| .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8 |
| .byte 102,65,15,58,33,192,48 // insertps $0x30,%xmm8,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_g_sse41 |
| .globl _sk_table_g_sse41 |
| FUNCTION(_sk_table_g_sse41) |
| _sk_table_g_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 102,65,15,91,200 // cvtps2dq %xmm8,%xmm1 |
| .byte 102,72,15,58,22,200,1 // pextrq $0x1,%xmm1,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,67,15,16,12,145 // movss (%r9,%r10,4),%xmm1 |
| .byte 102,65,15,58,33,12,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm1 |
| .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8 |
| .byte 102,65,15,58,33,200,32 // insertps $0x20,%xmm8,%xmm1 |
| .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8 |
| .byte 102,65,15,58,33,200,48 // insertps $0x30,%xmm8,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_b_sse41 |
| .globl _sk_table_b_sse41 |
| FUNCTION(_sk_table_b_sse41) |
| _sk_table_b_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 102,65,15,91,208 // cvtps2dq %xmm8,%xmm2 |
| .byte 102,72,15,58,22,208,1 // pextrq $0x1,%xmm2,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| .byte 102,65,15,58,33,20,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm2 |
| .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8 |
| .byte 102,65,15,58,33,208,32 // insertps $0x20,%xmm8,%xmm2 |
| .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8 |
| .byte 102,65,15,58,33,208,48 // insertps $0x30,%xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_a_sse41 |
| .globl _sk_table_a_sse41 |
| FUNCTION(_sk_table_a_sse41) |
| _sk_table_a_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,65,15,91,216 // cvtps2dq %xmm8,%xmm3 |
| .byte 102,72,15,58,22,216,1 // pextrq $0x1,%xmm3,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,217 // movq %xmm3,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,67,15,16,28,145 // movss (%r9,%r10,4),%xmm3 |
| .byte 102,65,15,58,33,28,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm3 |
| .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8 |
| .byte 102,65,15,58,33,216,32 // insertps $0x20,%xmm8,%xmm3 |
| .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8 |
| .byte 102,65,15,58,33,216,48 // insertps $0x30,%xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_r_sse41 |
| .globl _sk_parametric_r_sse41 |
| FUNCTION(_sk_parametric_r_sse41) |
| _sk_parametric_r_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,64,16 // movss 0x10(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,216 // mulps %xmm0,%xmm11 |
| .byte 65,15,194,192,2 // cmpleps %xmm8,%xmm0 |
| .byte 243,68,15,16,64,24 // movss 0x18(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,88,200 // addps %xmm8,%xmm9 |
| .byte 243,68,15,16,16 // movss (%rax),%xmm10 |
| .byte 243,68,15,16,64,8 // movss 0x8(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,88,216 // addps %xmm8,%xmm11 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,196 // mulps %xmm12,%xmm8 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 102,69,15,112,228,0 // pshufd $0x0,%xmm12,%xmm12 |
| .byte 102,69,15,219,227 // pand %xmm11,%xmm12 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 |
| .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,92,196 // subps %xmm12,%xmm8 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,89,227 // mulps %xmm11,%xmm12 |
| .byte 69,15,92,196 // subps %xmm12,%xmm8 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| .byte 69,15,94,229 // divps %xmm13,%xmm12 |
| .byte 69,15,92,196 // subps %xmm12,%xmm8 |
| .byte 69,15,89,194 // mulps %xmm10,%xmm8 |
| .byte 102,69,15,58,8,216,1 // roundps $0x1,%xmm8,%xmm11 |
| .byte 185,0,0,0,75 // mov $0x4b000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,224 // addps %xmm8,%xmm12 |
| .byte 69,15,92,195 // subps %xmm11,%xmm8 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,216 // mulps %xmm8,%xmm11 |
| .byte 69,15,92,227 // subps %xmm11,%xmm12 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,92,232 // subps %xmm8,%xmm13 |
| .byte 69,15,94,221 // divps %xmm13,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,218 // mulps %xmm10,%xmm11 |
| .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 |
| .byte 243,68,15,16,64,20 // movss 0x14(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,88,194 // addps %xmm10,%xmm8 |
| .byte 102,69,15,56,20,193 // blendvps %xmm0,%xmm9,%xmm8 |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 68,15,95,192 // maxps %xmm0,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_g_sse41 |
| .globl _sk_parametric_g_sse41 |
| FUNCTION(_sk_parametric_g_sse41) |
| _sk_parametric_g_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 15,40,193 // movaps %xmm1,%xmm0 |
| .byte 65,15,194,194,2 // cmpleps %xmm10,%xmm0 |
| .byte 243,15,16,72,24 // movss 0x18(%rax),%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 68,15,88,201 // addps %xmm1,%xmm9 |
| .byte 243,68,15,16,16 // movss (%rax),%xmm10 |
| .byte 243,15,16,72,8 // movss 0x8(%rax),%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 68,15,88,217 // addps %xmm1,%xmm11 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,89,236 // mulps %xmm12,%xmm13 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 102,68,15,112,225,0 // pshufd $0x0,%xmm1,%xmm12 |
| .byte 102,69,15,219,227 // pand %xmm11,%xmm12 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 102,68,15,112,217,0 // pshufd $0x0,%xmm1,%xmm11 |
| .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 68,15,92,233 // subps %xmm1,%xmm13 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 65,15,89,203 // mulps %xmm11,%xmm1 |
| .byte 68,15,92,233 // subps %xmm1,%xmm13 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| .byte 68,15,94,225 // divps %xmm1,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 |
| .byte 185,0,0,0,75 // mov $0x4b000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,92,235 // subps %xmm11,%xmm13 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 65,15,89,205 // mulps %xmm13,%xmm1 |
| .byte 68,15,92,225 // subps %xmm1,%xmm12 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,15,110,201 // movd %ecx,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 65,15,92,205 // subps %xmm13,%xmm1 |
| .byte 68,15,94,217 // divps %xmm1,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,218 // mulps %xmm10,%xmm11 |
| .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 |
| .byte 243,15,16,72,20 // movss 0x14(%rax),%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 102,65,15,56,20,201 // blendvps %xmm0,%xmm9,%xmm1 |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,95,200 // maxps %xmm0,%xmm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,93,200 // minps %xmm0,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_b_sse41 |
| .globl _sk_parametric_b_sse41 |
| FUNCTION(_sk_parametric_b_sse41) |
| _sk_parametric_b_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 65,15,194,194,2 // cmpleps %xmm10,%xmm0 |
| .byte 243,15,16,80,24 // movss 0x18(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 68,15,88,202 // addps %xmm2,%xmm9 |
| .byte 243,68,15,16,16 // movss (%rax),%xmm10 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 68,15,88,218 // addps %xmm2,%xmm11 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,89,236 // mulps %xmm12,%xmm13 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 102,68,15,112,226,0 // pshufd $0x0,%xmm2,%xmm12 |
| .byte 102,69,15,219,227 // pand %xmm11,%xmm12 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 |
| .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 68,15,92,234 // subps %xmm2,%xmm13 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,89,211 // mulps %xmm11,%xmm2 |
| .byte 68,15,92,234 // subps %xmm2,%xmm13 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| .byte 68,15,94,226 // divps %xmm2,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 |
| .byte 185,0,0,0,75 // mov $0x4b000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,92,235 // subps %xmm11,%xmm13 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| .byte 68,15,92,226 // subps %xmm2,%xmm12 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,92,213 // subps %xmm13,%xmm2 |
| .byte 68,15,94,218 // divps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,218 // mulps %xmm10,%xmm11 |
| .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 |
| .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| .byte 102,65,15,56,20,209 // blendvps %xmm0,%xmm9,%xmm2 |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,95,208 // maxps %xmm0,%xmm2 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,93,208 // minps %xmm0,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_a_sse41 |
| .globl _sk_parametric_a_sse41 |
| FUNCTION(_sk_parametric_a_sse41) |
| _sk_parametric_a_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,219 // mulps %xmm3,%xmm11 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 65,15,194,194,2 // cmpleps %xmm10,%xmm0 |
| .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| .byte 243,68,15,16,16 // movss (%rax),%xmm10 |
| .byte 243,15,16,88,8 // movss 0x8(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 68,15,88,219 // addps %xmm3,%xmm11 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,89,236 // mulps %xmm12,%xmm13 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 102,68,15,112,227,0 // pshufd $0x0,%xmm3,%xmm12 |
| .byte 102,69,15,219,227 // pand %xmm11,%xmm12 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 102,68,15,112,219,0 // pshufd $0x0,%xmm3,%xmm11 |
| .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 68,15,92,235 // subps %xmm3,%xmm13 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,89,219 // mulps %xmm11,%xmm3 |
| .byte 68,15,92,235 // subps %xmm3,%xmm13 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,88,219 // addps %xmm11,%xmm3 |
| .byte 68,15,94,227 // divps %xmm3,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| .byte 102,69,15,58,8,221,1 // roundps $0x1,%xmm13,%xmm11 |
| .byte 185,0,0,0,75 // mov $0x4b000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,92,235 // subps %xmm11,%xmm13 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,89,221 // mulps %xmm13,%xmm3 |
| .byte 68,15,92,227 // subps %xmm3,%xmm12 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,92,221 // subps %xmm13,%xmm3 |
| .byte 68,15,94,219 // divps %xmm3,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,218 // mulps %xmm10,%xmm11 |
| .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10 |
| .byte 243,15,16,88,20 // movss 0x14(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| .byte 102,65,15,56,20,217 // blendvps %xmm0,%xmm9,%xmm3 |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,95,216 // maxps %xmm0,%xmm3 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,93,216 // minps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_a8_sse41 |
| .globl _sk_load_a8_sse41 |
| FUNCTION(_sk_load_a8_sse41) |
| _sk_load_a8_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 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_a8_sse41 |
| .globl _sk_gather_a8_sse41 |
| FUNCTION(_sk_gather_a8_sse41) |
| _sk_gather_a8_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,58,32,4,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm0 |
| .byte 102,65,15,58,32,4,9,1 // pinsrb $0x1,(%r9,%rcx,1),%xmm0 |
| .byte 67,15,182,12,1 // movzbl (%r9,%r8,1),%ecx |
| .byte 102,15,58,32,193,2 // pinsrb $0x2,%ecx,%xmm0 |
| .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax |
| .byte 102,15,58,32,192,3 // pinsrb $0x3,%eax,%xmm0 |
| .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 102,15,239,201 // pxor %xmm1,%xmm1 |
| .byte 102,15,239,210 // pxor %xmm2,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_a8_sse41 |
| .globl _sk_store_a8_sse41 |
| FUNCTION(_sk_store_a8_sse41) |
| _sk_store_a8_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| .byte 102,69,15,103,192 // packuswb %xmm8,%xmm8 |
| .byte 102,68,15,126,4,56 // movd %xmm8,(%rax,%rdi,1) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_g8_sse41 |
| .globl _sk_load_g8_sse41 |
| FUNCTION(_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_gather_g8_sse41 |
| .globl _sk_gather_g8_sse41 |
| FUNCTION(_sk_gather_g8_sse41) |
| _sk_gather_g8_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,58,32,4,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm0 |
| .byte 102,65,15,58,32,4,9,1 // pinsrb $0x1,(%r9,%rcx,1),%xmm0 |
| .byte 67,15,182,12,1 // movzbl (%r9,%r8,1),%ecx |
| .byte 102,15,58,32,193,2 // pinsrb $0x2,%ecx,%xmm0 |
| .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax |
| .byte 102,15,58,32,192,3 // pinsrb $0x3,%eax,%xmm0 |
| .byte 102,15,56,49,192 // pmovzxbd %xmm0,%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_gather_i8_sse41 |
| .globl _sk_gather_i8_sse41 |
| FUNCTION(_sk_gather_i8_sse41) |
| _sk_gather_i8_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,192 // mov %rax,%r8 |
| .byte 77,133,192 // test %r8,%r8 |
| .byte 116,5 // je 26a1 <_sk_gather_i8_sse41+0xf> |
| .byte 76,137,192 // mov %r8,%rax |
| .byte 235,2 // jmp 26a3 <_sk_gather_i8_sse41+0x11> |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,16 // mov (%rax),%r10 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,203 // mov %ecx,%r11d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,58,32,4,26,0 // pinsrb $0x0,(%r10,%r11,1),%xmm0 |
| .byte 102,65,15,58,32,4,10,1 // pinsrb $0x1,(%r10,%rcx,1),%xmm0 |
| .byte 102,67,15,58,32,4,10,2 // pinsrb $0x2,(%r10,%r9,1),%xmm0 |
| .byte 102,65,15,58,32,4,2,3 // pinsrb $0x3,(%r10,%rax,1),%xmm0 |
| .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0 |
| .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 73,139,64,8 // mov 0x8(%r8),%rax |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 69,137,202 // mov %r9d,%r10d |
| .byte 73,193,233,30 // shr $0x1e,%r9 |
| .byte 102,66,15,110,28,128 // movd (%rax,%r8,4),%xmm3 |
| .byte 102,15,58,34,28,8,1 // pinsrd $0x1,(%rax,%rcx,1),%xmm3 |
| .byte 102,66,15,58,34,28,144,2 // pinsrd $0x2,(%rax,%r10,4),%xmm3 |
| .byte 102,66,15,58,34,28,8,3 // pinsrd $0x3,(%rax,%r9,1),%xmm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_565_sse41 |
| .globl _sk_load_565_sse41 |
| FUNCTION(_sk_load_565_sse41) |
| _sk_load_565_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 102,15,56,51,20,120 // pmovzxwd (%rax,%rdi,2),%xmm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,219,194 // pand %xmm2,%xmm0 |
| .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| .byte 102,15,219,202 // pand %xmm2,%xmm1 |
| .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,15,219,218 // pand %xmm2,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_565_sse41 |
| .globl _sk_gather_565_sse41 |
| FUNCTION(_sk_gather_565_sse41) |
| _sk_gather_565_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,196,4,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm0 |
| .byte 102,65,15,196,4,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm0 |
| .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx |
| .byte 102,15,196,193,2 // pinsrw $0x2,%ecx,%xmm0 |
| .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax |
| .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0 |
| .byte 102,15,56,51,208 // pmovzxwd %xmm0,%xmm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,219,194 // pand %xmm2,%xmm0 |
| .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| .byte 102,15,219,202 // pand %xmm2,%xmm1 |
| .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,15,219,218 // pand %xmm2,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_565_sse41 |
| .globl _sk_store_565_sse41 |
| FUNCTION(_sk_store_565_sse41) |
| _sk_store_565_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,248,65 // mov $0x41f80000,%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 102,65,15,114,241,11 // pslld $0xb,%xmm9 |
| .byte 185,0,0,124,66 // mov $0x427c0000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10 |
| .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%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_4444_sse41 |
| .globl _sk_load_4444_sse41 |
| FUNCTION(_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_gather_4444_sse41 |
| .globl _sk_gather_4444_sse41 |
| FUNCTION(_sk_gather_4444_sse41) |
| _sk_gather_4444_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,196,4,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm0 |
| .byte 102,65,15,196,4,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm0 |
| .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx |
| .byte 102,15,196,193,2 // pinsrw $0x2,%ecx,%xmm0 |
| .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax |
| .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0 |
| .byte 102,68,15,56,51,200 // pmovzxwd %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_sse41 |
| .globl _sk_store_4444_sse41 |
| FUNCTION(_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 |
| FUNCTION(_sk_load_8888_sse41) |
| _sk_load_8888_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,111,28,184 // movdqu (%rax,%rdi,4),%xmm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_8888_sse41 |
| .globl _sk_gather_8888_sse41 |
| FUNCTION(_sk_gather_8888_sse41) |
| _sk_gather_8888_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,126,192 // movq %xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,110,28,129 // movd (%r9,%r8,4),%xmm3 |
| .byte 102,65,15,58,34,28,129,1 // pinsrd $0x1,(%r9,%rax,4),%xmm3 |
| .byte 102,67,15,58,34,28,145,2 // pinsrd $0x2,(%r9,%r10,4),%xmm3 |
| .byte 102,65,15,58,34,28,137,3 // pinsrd $0x3,(%r9,%rcx,4),%xmm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %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_8888_sse41 |
| .globl _sk_store_8888_sse41 |
| FUNCTION(_sk_store_8888_sse41) |
| _sk_store_8888_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,127,67 // mov $0x437f0000,%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 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 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,65,15,114,240,24 // pslld $0x18,%xmm8 |
| .byte 102,69,15,235,193 // por %xmm9,%xmm8 |
| .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| .byte 243,68,15,127,4,184 // movdqu %xmm8,(%rax,%rdi,4) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_f16_sse41 |
| .globl _sk_load_f16_sse41 |
| FUNCTION(_sk_load_f16_sse41) |
| _sk_load_f16_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| .byte 102,68,15,111,200 // movdqa %xmm0,%xmm9 |
| .byte 102,68,15,97,201 // punpcklwd %xmm1,%xmm9 |
| .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| .byte 102,68,15,97,224 // punpcklwd %xmm0,%xmm12 |
| .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9 |
| .byte 102,69,15,56,51,236 // pmovzxwd %xmm12,%xmm13 |
| .byte 184,0,128,0,0 // mov $0x8000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,68,15,112,192,0 // pshufd $0x0,%xmm0,%xmm8 |
| .byte 102,65,15,111,213 // movdqa %xmm13,%xmm2 |
| .byte 102,65,15,219,208 // pand %xmm8,%xmm2 |
| .byte 184,0,124,0,0 // mov $0x7c00,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 |
| .byte 102,65,15,111,197 // movdqa %xmm13,%xmm0 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 102,68,15,239,234 // pxor %xmm2,%xmm13 |
| .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 |
| .byte 102,15,114,242,16 // pslld $0x10,%xmm2 |
| .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 |
| .byte 184,0,0,0,56 // mov $0x38000000,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 102,68,15,112,217,0 // pshufd $0x0,%xmm1,%xmm11 |
| .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 |
| .byte 102,65,15,254,213 // paddd %xmm13,%xmm2 |
| .byte 102,65,15,118,194 // pcmpeqd %xmm10,%xmm0 |
| .byte 102,15,223,194 // pandn %xmm2,%xmm0 |
| .byte 102,65,15,115,220,8 // psrldq $0x8,%xmm12 |
| .byte 102,69,15,56,51,228 // pmovzxwd %xmm12,%xmm12 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,65,15,219,208 // pand %xmm8,%xmm2 |
| .byte 102,65,15,111,204 // movdqa %xmm12,%xmm1 |
| .byte 102,15,219,203 // pand %xmm3,%xmm1 |
| .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 |
| .byte 102,15,114,242,16 // pslld $0x10,%xmm2 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 |
| .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 |
| .byte 102,65,15,118,202 // pcmpeqd %xmm10,%xmm1 |
| .byte 102,15,223,202 // pandn %xmm2,%xmm1 |
| .byte 102,69,15,56,51,225 // pmovzxwd %xmm9,%xmm12 |
| .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 |
| .byte 102,69,15,219,232 // pand %xmm8,%xmm13 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,15,219,211 // pand %xmm3,%xmm2 |
| .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 |
| .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 |
| .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 |
| .byte 102,65,15,118,210 // pcmpeqd %xmm10,%xmm2 |
| .byte 102,65,15,223,213 // pandn %xmm13,%xmm2 |
| .byte 102,65,15,115,217,8 // psrldq $0x8,%xmm9 |
| .byte 102,69,15,56,51,201 // pmovzxwd %xmm9,%xmm9 |
| .byte 102,69,15,219,193 // pand %xmm9,%xmm8 |
| .byte 102,65,15,219,217 // pand %xmm9,%xmm3 |
| .byte 102,69,15,239,200 // pxor %xmm8,%xmm9 |
| .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| .byte 102,65,15,114,241,13 // pslld $0xd,%xmm9 |
| .byte 102,69,15,254,195 // paddd %xmm11,%xmm8 |
| .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 |
| .byte 102,65,15,118,218 // pcmpeqd %xmm10,%xmm3 |
| .byte 102,65,15,223,216 // pandn %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_f16_sse41 |
| .globl _sk_gather_f16_sse41 |
| FUNCTION(_sk_gather_f16_sse41) |
| _sk_gather_f16_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,194 // paddd %xmm2,%xmm0 |
| .byte 102,72,15,126,192 // movq %xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,65,15,126,4,201 // movq (%r9,%rcx,8),%xmm0 |
| .byte 243,67,15,126,12,209 // movq (%r9,%r10,8),%xmm1 |
| .byte 102,15,108,200 // punpcklqdq %xmm0,%xmm1 |
| .byte 243,65,15,126,4,193 // movq (%r9,%rax,8),%xmm0 |
| .byte 243,67,15,126,20,193 // movq (%r9,%r8,8),%xmm2 |
| .byte 102,15,108,208 // punpcklqdq %xmm0,%xmm2 |
| .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9 |
| .byte 102,68,15,97,201 // punpcklwd %xmm1,%xmm9 |
| .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2 |
| .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| .byte 102,68,15,97,226 // punpcklwd %xmm2,%xmm12 |
| .byte 102,68,15,105,202 // punpckhwd %xmm2,%xmm9 |
| .byte 102,69,15,56,51,236 // pmovzxwd %xmm12,%xmm13 |
| .byte 184,0,128,0,0 // mov $0x8000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,68,15,112,192,0 // pshufd $0x0,%xmm0,%xmm8 |
| .byte 102,65,15,111,213 // movdqa %xmm13,%xmm2 |
| .byte 102,65,15,219,208 // pand %xmm8,%xmm2 |
| .byte 184,0,124,0,0 // mov $0x7c00,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 |
| .byte 102,65,15,111,197 // movdqa %xmm13,%xmm0 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 102,68,15,239,234 // pxor %xmm2,%xmm13 |
| .byte 102,69,15,239,210 // pxor %xmm10,%xmm10 |
| .byte 102,15,114,242,16 // pslld $0x10,%xmm2 |
| .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 |
| .byte 184,0,0,0,56 // mov $0x38000000,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 102,68,15,112,217,0 // pshufd $0x0,%xmm1,%xmm11 |
| .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 |
| .byte 102,65,15,254,213 // paddd %xmm13,%xmm2 |
| .byte 102,65,15,118,194 // pcmpeqd %xmm10,%xmm0 |
| .byte 102,15,223,194 // pandn %xmm2,%xmm0 |
| .byte 102,65,15,115,220,8 // psrldq $0x8,%xmm12 |
| .byte 102,69,15,56,51,228 // pmovzxwd %xmm12,%xmm12 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,65,15,219,208 // pand %xmm8,%xmm2 |
| .byte 102,65,15,111,204 // movdqa %xmm12,%xmm1 |
| .byte 102,15,219,203 // pand %xmm3,%xmm1 |
| .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 |
| .byte 102,15,114,242,16 // pslld $0x10,%xmm2 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 |
| .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 |
| .byte 102,65,15,118,202 // pcmpeqd %xmm10,%xmm1 |
| .byte 102,15,223,202 // pandn %xmm2,%xmm1 |
| .byte 102,69,15,56,51,225 // pmovzxwd %xmm9,%xmm12 |
| .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 |
| .byte 102,69,15,219,232 // pand %xmm8,%xmm13 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,15,219,211 // pand %xmm3,%xmm2 |
| .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 |
| .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 |
| .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 |
| .byte 102,65,15,118,210 // pcmpeqd %xmm10,%xmm2 |
| .byte 102,65,15,223,213 // pandn %xmm13,%xmm2 |
| .byte 102,65,15,115,217,8 // psrldq $0x8,%xmm9 |
| .byte 102,69,15,56,51,201 // pmovzxwd %xmm9,%xmm9 |
| .byte 102,69,15,219,193 // pand %xmm9,%xmm8 |
| .byte 102,65,15,219,217 // pand %xmm9,%xmm3 |
| .byte 102,69,15,239,200 // pxor %xmm8,%xmm9 |
| .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| .byte 102,65,15,114,241,13 // pslld $0xd,%xmm9 |
| .byte 102,69,15,254,195 // paddd %xmm11,%xmm8 |
| .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 |
| .byte 102,65,15,118,218 // pcmpeqd %xmm10,%xmm3 |
| .byte 102,65,15,223,216 // pandn %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_f16_sse41 |
| .globl _sk_store_f16_sse41 |
| FUNCTION(_sk_store_f16_sse41) |
| _sk_store_f16_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,0,128 // mov $0x80000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9 |
| .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| .byte 102,68,15,219,224 // pand %xmm0,%xmm12 |
| .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 |
| .byte 102,69,15,239,196 // pxor %xmm12,%xmm8 |
| .byte 185,0,0,128,56 // mov $0x38800000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 102,65,15,114,212,16 // psrld $0x10,%xmm12 |
| .byte 102,69,15,111,232 // movdqa %xmm8,%xmm13 |
| .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 |
| .byte 185,0,192,1,0 // mov $0x1c000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 |
| .byte 102,69,15,250,227 // psubd %xmm11,%xmm12 |
| .byte 102,69,15,254,229 // paddd %xmm13,%xmm12 |
| .byte 69,15,194,194,5 // cmpnltps %xmm10,%xmm8 |
| .byte 69,15,84,196 // andps %xmm12,%xmm8 |
| .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13 |
| .byte 102,68,15,219,233 // pand %xmm1,%xmm13 |
| .byte 102,68,15,111,225 // movdqa %xmm1,%xmm12 |
| .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 |
| .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13 |
| .byte 102,69,15,111,244 // movdqa %xmm12,%xmm14 |
| .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 |
| .byte 102,69,15,250,235 // psubd %xmm11,%xmm13 |
| .byte 102,69,15,254,238 // paddd %xmm14,%xmm13 |
| .byte 69,15,194,226,5 // cmpnltps %xmm10,%xmm12 |
| .byte 69,15,84,229 // andps %xmm13,%xmm12 |
| .byte 102,69,15,56,43,228 // packusdw %xmm12,%xmm12 |
| .byte 102,69,15,111,241 // movdqa %xmm9,%xmm14 |
| .byte 102,68,15,219,242 // pand %xmm2,%xmm14 |
| .byte 102,68,15,111,234 // movdqa %xmm2,%xmm13 |
| .byte 102,69,15,239,238 // pxor %xmm14,%xmm13 |
| .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14 |
| .byte 102,69,15,111,253 // movdqa %xmm13,%xmm15 |
| .byte 102,65,15,114,215,13 // psrld $0xd,%xmm15 |
| .byte 102,69,15,250,243 // psubd %xmm11,%xmm14 |
| .byte 102,69,15,254,247 // paddd %xmm15,%xmm14 |
| .byte 69,15,194,234,5 // cmpnltps %xmm10,%xmm13 |
| .byte 69,15,84,238 // andps %xmm14,%xmm13 |
| .byte 102,69,15,56,43,237 // packusdw %xmm13,%xmm13 |
| .byte 102,68,15,219,203 // pand %xmm3,%xmm9 |
| .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14 |
| .byte 102,69,15,239,241 // pxor %xmm9,%xmm14 |
| .byte 102,65,15,114,209,16 // psrld $0x10,%xmm9 |
| .byte 102,69,15,111,254 // movdqa %xmm14,%xmm15 |
| .byte 102,65,15,114,215,13 // psrld $0xd,%xmm15 |
| .byte 102,69,15,250,203 // psubd %xmm11,%xmm9 |
| .byte 102,69,15,254,207 // paddd %xmm15,%xmm9 |
| .byte 69,15,194,242,5 // cmpnltps %xmm10,%xmm14 |
| .byte 69,15,84,241 // andps %xmm9,%xmm14 |
| .byte 102,69,15,56,43,246 // packusdw %xmm14,%xmm14 |
| .byte 102,69,15,97,196 // punpcklwd %xmm12,%xmm8 |
| .byte 102,69,15,97,238 // punpcklwd %xmm14,%xmm13 |
| .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| .byte 102,69,15,98,205 // punpckldq %xmm13,%xmm9 |
| .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| .byte 102,69,15,106,197 // punpckhdq %xmm13,%xmm8 |
| .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_u16_be_sse41 |
| .globl _sk_load_u16_be_sse41 |
| FUNCTION(_sk_load_u16_be_sse41) |
| _sk_load_u16_be_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| .byte 102,15,111,208 // movdqa %xmm0,%xmm2 |
| .byte 102,15,97,209 // punpcklwd %xmm1,%xmm2 |
| .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,15,105,208 // punpckhwd %xmm0,%xmm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 102,15,111,193 // movdqa %xmm1,%xmm0 |
| .byte 102,15,113,240,8 // psllw $0x8,%xmm0 |
| .byte 102,15,112,217,78 // pshufd $0x4e,%xmm1,%xmm3 |
| .byte 102,15,113,209,8 // psrlw $0x8,%xmm1 |
| .byte 102,15,235,200 // por %xmm0,%xmm1 |
| .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,15,235,217 // por %xmm1,%xmm3 |
| .byte 102,15,56,51,203 // pmovzxwd %xmm3,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9 |
| .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9 |
| .byte 102,15,112,218,78 // pshufd $0x4e,%xmm2,%xmm3 |
| .byte 102,15,113,210,8 // psrlw $0x8,%xmm2 |
| .byte 102,65,15,235,209 // por %xmm9,%xmm2 |
| .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,68,15,111,203 // movdqa %xmm3,%xmm9 |
| .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,65,15,235,217 // por %xmm9,%xmm3 |
| .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_rgb_u16_be_sse41 |
| .globl _sk_load_rgb_u16_be_sse41 |
| FUNCTION(_sk_load_rgb_u16_be_sse41) |
| _sk_load_rgb_u16_be_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 72,141,12,127 // lea (%rdi,%rdi,2),%rcx |
| .byte 243,15,111,20,72 // movdqu (%rax,%rcx,2),%xmm2 |
| .byte 243,15,111,68,72,8 // movdqu 0x8(%rax,%rcx,2),%xmm0 |
| .byte 102,15,115,216,4 // psrldq $0x4,%xmm0 |
| .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| .byte 102,15,115,217,6 // psrldq $0x6,%xmm1 |
| .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 |
| .byte 102,15,115,216,6 // psrldq $0x6,%xmm0 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,15,111,194 // movdqa %xmm2,%xmm0 |
| .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 |
| .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3 |
| .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 102,15,111,200 // movdqa %xmm0,%xmm1 |
| .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| .byte 102,15,113,208,8 // psrlw $0x8,%xmm0 |
| .byte 102,15,235,193 // por %xmm1,%xmm0 |
| .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,15,235,217 // por %xmm1,%xmm3 |
| .byte 102,15,56,51,203 // pmovzxwd %xmm3,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,15,111,218 // movdqa %xmm2,%xmm3 |
| .byte 102,15,113,243,8 // psllw $0x8,%xmm3 |
| .byte 102,15,113,210,8 // psrlw $0x8,%xmm2 |
| .byte 102,15,235,211 // por %xmm3,%xmm2 |
| .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_u16_be_sse41 |
| .globl _sk_store_u16_be_sse41 |
| FUNCTION(_sk_store_u16_be_sse41) |
| _sk_store_u16_be_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,255,127,71 // mov $0x477fff00,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 |
| .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8 |
| .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| .byte 102,69,15,56,43,210 // packusdw %xmm10,%xmm10 |
| .byte 102,69,15,111,218 // movdqa %xmm10,%xmm11 |
| .byte 102,65,15,113,243,8 // psllw $0x8,%xmm11 |
| .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 |
| .byte 102,69,15,235,211 // por %xmm11,%xmm10 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 102,69,15,91,219 // cvtps2dq %xmm11,%xmm11 |
| .byte 102,69,15,56,43,219 // packusdw %xmm11,%xmm11 |
| .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12 |
| .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 |
| .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| .byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9 |
| .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| .byte 102,65,15,113,209,8 // psrlw $0x8,%xmm9 |
| .byte 102,69,15,235,204 // por %xmm12,%xmm9 |
| .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 |
| .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11 |
| .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9 |
| .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8 |
| .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_f32_sse41 |
| .globl _sk_load_f32_sse41 |
| FUNCTION(_sk_load_f32_sse41) |
| _sk_load_f32_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 72,137,249 // mov %rdi,%rcx |
| .byte 72,193,225,4 // shl $0x4,%rcx |
| .byte 68,15,16,4,8 // movups (%rax,%rcx,1),%xmm8 |
| .byte 15,16,68,8,16 // movups 0x10(%rax,%rcx,1),%xmm0 |
| .byte 15,16,92,8,32 // movups 0x20(%rax,%rcx,1),%xmm3 |
| .byte 68,15,16,76,8,48 // movups 0x30(%rax,%rcx,1),%xmm9 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 15,20,208 // unpcklps %xmm0,%xmm2 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 65,15,20,201 // unpcklps %xmm9,%xmm1 |
| .byte 68,15,21,192 // unpckhps %xmm0,%xmm8 |
| .byte 65,15,21,217 // unpckhps %xmm9,%xmm3 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 102,15,20,193 // unpcklpd %xmm1,%xmm0 |
| .byte 15,18,202 // movhlps %xmm2,%xmm1 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 102,15,20,211 // unpcklpd %xmm3,%xmm2 |
| .byte 65,15,18,216 // movhlps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_f32_sse41 |
| .globl _sk_store_f32_sse41 |
| FUNCTION(_sk_store_f32_sse41) |
| _sk_store_f32_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 72,137,249 // mov %rdi,%rcx |
| .byte 72,193,225,4 // shl $0x4,%rcx |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 68,15,20,201 // unpcklps %xmm1,%xmm9 |
| .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| .byte 68,15,40,218 // movaps %xmm2,%xmm11 |
| .byte 68,15,20,219 // unpcklps %xmm3,%xmm11 |
| .byte 68,15,21,193 // unpckhps %xmm1,%xmm8 |
| .byte 68,15,21,211 // unpckhps %xmm3,%xmm10 |
| .byte 69,15,40,225 // movaps %xmm9,%xmm12 |
| .byte 102,69,15,20,227 // unpcklpd %xmm11,%xmm12 |
| .byte 69,15,18,217 // movhlps %xmm9,%xmm11 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 102,69,15,20,202 // unpcklpd %xmm10,%xmm9 |
| .byte 69,15,18,208 // movhlps %xmm8,%xmm10 |
| .byte 102,68,15,17,36,8 // movupd %xmm12,(%rax,%rcx,1) |
| .byte 68,15,17,92,8,16 // movups %xmm11,0x10(%rax,%rcx,1) |
| .byte 102,68,15,17,76,8,32 // movupd %xmm9,0x20(%rax,%rcx,1) |
| .byte 68,15,17,84,8,48 // movups %xmm10,0x30(%rax,%rcx,1) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_x_sse41 |
| .globl _sk_clamp_x_sse41 |
| FUNCTION(_sk_clamp_x_sse41) |
| _sk_clamp_x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,95,192 // maxps %xmm0,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 102,15,118,192 // pcmpeqd %xmm0,%xmm0 |
| .byte 102,65,15,254,193 // paddd %xmm9,%xmm0 |
| .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_y_sse41 |
| .globl _sk_clamp_y_sse41 |
| FUNCTION(_sk_clamp_y_sse41) |
| _sk_clamp_y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,95,193 // maxps %xmm1,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 102,15,118,201 // pcmpeqd %xmm1,%xmm1 |
| .byte 102,65,15,254,201 // paddd %xmm9,%xmm1 |
| .byte 68,15,93,193 // minps %xmm1,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_x_sse41 |
| .globl _sk_repeat_x_sse41 |
| FUNCTION(_sk_repeat_x_sse41) |
| _sk_repeat_x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| .byte 65,15,93,193 // minps %xmm9,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_y_sse41 |
| .globl _sk_repeat_y_sse41 |
| FUNCTION(_sk_repeat_y_sse41) |
| _sk_repeat_y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 65,15,92,201 // subps %xmm9,%xmm1 |
| .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| .byte 65,15,93,201 // minps %xmm9,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_x_sse41 |
| .globl _sk_mirror_x_sse41 |
| FUNCTION(_sk_mirror_x_sse41) |
| _sk_mirror_x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 243,69,15,88,192 // addss %xmm8,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,40,208 // movaps %xmm0,%xmm10 |
| .byte 69,15,94,208 // divps %xmm8,%xmm10 |
| .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 65,15,92,194 // subps %xmm10,%xmm0 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| .byte 65,15,84,192 // andps %xmm8,%xmm0 |
| .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 |
| .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 |
| .byte 65,15,93,192 // minps %xmm8,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_y_sse41 |
| .globl _sk_mirror_y_sse41 |
| FUNCTION(_sk_mirror_y_sse41) |
| _sk_mirror_y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 65,15,92,201 // subps %xmm9,%xmm1 |
| .byte 243,69,15,88,192 // addss %xmm8,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| .byte 69,15,94,208 // divps %xmm8,%xmm10 |
| .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 65,15,92,202 // subps %xmm10,%xmm1 |
| .byte 65,15,92,201 // subps %xmm9,%xmm1 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,92,193 // subps %xmm1,%xmm8 |
| .byte 65,15,84,200 // andps %xmm8,%xmm1 |
| .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 |
| .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 |
| .byte 65,15,93,200 // minps %xmm8,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_luminance_to_alpha_sse41 |
| .globl _sk_luminance_to_alpha_sse41 |
| FUNCTION(_sk_luminance_to_alpha_sse41) |
| _sk_luminance_to_alpha_sse41: |
| .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 184,89,23,55,63 // mov $0x3f371759,%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 15,88,195 // addps %xmm3,%xmm0 |
| .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,218 // mulps %xmm2,%xmm3 |
| .byte 15,88,216 // addps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_2x3_sse41 |
| .globl _sk_matrix_2x3_sse41 |
| FUNCTION(_sk_matrix_2x3_sse41) |
| _sk_matrix_2x3_sse41: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,16 // movss 0x10(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_3x4_sse41 |
| .globl _sk_matrix_3x4_sse41 |
| FUNCTION(_sk_matrix_3x4_sse41) |
| _sk_matrix_3x4_sse41: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,36 // movss 0x24(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_4x5_sse41 |
| .globl _sk_matrix_4x5_sse41 |
| FUNCTION(_sk_matrix_4x5_sse41) |
| _sk_matrix_4x5_sse41: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,32 // movss 0x20(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,48 // movss 0x30(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,64 // movss 0x40(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,36 // movss 0x24(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,52 // movss 0x34(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,68 // movss 0x44(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,56 // movss 0x38(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 243,68,15,16,112,72 // movss 0x48(%rax),%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 243,68,15,16,88,12 // movss 0xc(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 243,68,15,16,112,60 // movss 0x3c(%rax),%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 243,68,15,16,120,76 // movss 0x4c(%rax),%xmm15 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| .byte 69,15,88,247 // addps %xmm15,%xmm14 |
| .byte 68,15,89,234 // mulps %xmm2,%xmm13 |
| .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,89,216 // mulps %xmm8,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_perspective_sse41 |
| .globl _sk_matrix_perspective_sse41 |
| FUNCTION(_sk_matrix_perspective_sse41) |
| _sk_matrix_perspective_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 243,68,15,16,80,24 // movss 0x18(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,83,202 // rcpps %xmm10,%xmm1 |
| .byte 15,89,193 // mulps %xmm1,%xmm0 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_sse41 |
| .globl _sk_linear_gradient_sse41 |
| FUNCTION(_sk_linear_gradient_sse41) |
| _sk_linear_gradient_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,24 // movss 0x18(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,28 // movss 0x1c(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,132,254,0,0,0 // je 3abe <_sk_linear_gradient_sse41+0x138> |
| .byte 15,41,100,36,168 // movaps %xmm4,-0x58(%rsp) |
| .byte 15,41,108,36,184 // movaps %xmm5,-0x48(%rsp) |
| .byte 15,41,116,36,200 // movaps %xmm6,-0x38(%rsp) |
| .byte 15,41,124,36,216 // movaps %xmm7,-0x28(%rsp) |
| .byte 72,139,64,8 // mov 0x8(%rax),%rax |
| .byte 72,131,192,32 // add $0x20,%rax |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,40,233 // movaps %xmm1,%xmm5 |
| .byte 15,40,242 // movaps %xmm2,%xmm6 |
| .byte 15,40,251 // movaps %xmm3,%xmm7 |
| .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| .byte 69,15,40,243 // movaps %xmm11,%xmm14 |
| .byte 69,15,40,252 // movaps %xmm12,%xmm15 |
| .byte 68,15,41,108,36,232 // movaps %xmm13,-0x18(%rsp) |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 243,15,16,80,224 // movss -0x20(%rax),%xmm2 |
| .byte 243,68,15,16,72,228 // movss -0x1c(%rax),%xmm9 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,40,224 // movaps %xmm0,%xmm4 |
| .byte 15,194,194,1 // cmpltps %xmm2,%xmm0 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9 |
| .byte 243,15,16,72,232 // movss -0x18(%rax),%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 102,15,56,20,205 // blendvps %xmm0,%xmm5,%xmm1 |
| .byte 243,15,16,80,236 // movss -0x14(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 102,15,56,20,214 // blendvps %xmm0,%xmm6,%xmm2 |
| .byte 243,15,16,88,240 // movss -0x10(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 102,15,56,20,223 // blendvps %xmm0,%xmm7,%xmm3 |
| .byte 243,68,15,16,80,244 // movss -0xc(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 102,69,15,56,20,208 // blendvps %xmm0,%xmm8,%xmm10 |
| .byte 243,68,15,16,88,248 // movss -0x8(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 102,69,15,56,20,222 // blendvps %xmm0,%xmm14,%xmm11 |
| .byte 243,68,15,16,96,252 // movss -0x4(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 102,69,15,56,20,231 // blendvps %xmm0,%xmm15,%xmm12 |
| .byte 243,68,15,16,40 // movss (%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 102,68,15,56,20,108,36,232 // blendvps %xmm0,-0x18(%rsp),%xmm13 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 72,131,192,36 // add $0x24,%rax |
| .byte 72,255,201 // dec %rcx |
| .byte 15,133,65,255,255,255 // jne 39e9 <_sk_linear_gradient_sse41+0x63> |
| .byte 15,40,124,36,216 // movaps -0x28(%rsp),%xmm7 |
| .byte 15,40,116,36,200 // movaps -0x38(%rsp),%xmm6 |
| .byte 15,40,108,36,184 // movaps -0x48(%rsp),%xmm5 |
| .byte 15,40,100,36,168 // movaps -0x58(%rsp),%xmm4 |
| .byte 235,13 // jmp 3acb <_sk_linear_gradient_sse41+0x145> |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 15,89,200 // mulps %xmm0,%xmm1 |
| .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| .byte 15,89,208 // mulps %xmm0,%xmm2 |
| .byte 65,15,88,212 // addps %xmm12,%xmm2 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 65,15,88,221 // addps %xmm13,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_2stops_sse41 |
| .globl _sk_linear_gradient_2stops_sse41 |
| FUNCTION(_sk_linear_gradient_2stops_sse41) |
| _sk_linear_gradient_2stops_sse41: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,15,16,80,16 // movss 0x10(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,88,194 // addps %xmm2,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 15,88,211 // addps %xmm3,%xmm2 |
| .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 243,68,15,16,72,28 // movss 0x1c(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_save_xy_sse41 |
| .globl _sk_save_xy_sse41 |
| FUNCTION(_sk_save_xy_sse41) |
| _sk_save_xy_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%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 68,15,88,200 // addps %xmm0,%xmm9 |
| .byte 102,69,15,58,8,209,1 // roundps $0x1,%xmm9,%xmm10 |
| .byte 69,15,92,202 // subps %xmm10,%xmm9 |
| .byte 68,15,88,193 // addps %xmm1,%xmm8 |
| .byte 102,69,15,58,8,208,1 // roundps $0x1,%xmm8,%xmm10 |
| .byte 69,15,92,194 // subps %xmm10,%xmm8 |
| .byte 15,17,0 // movups %xmm0,(%rax) |
| .byte 15,17,72,32 // movups %xmm1,0x20(%rax) |
| .byte 68,15,17,72,64 // movups %xmm9,0x40(%rax) |
| .byte 68,15,17,64,96 // movups %xmm8,0x60(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_accumulate_sse41 |
| .globl _sk_accumulate_sse41 |
| FUNCTION(_sk_accumulate_sse41) |
| _sk_accumulate_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 68,15,16,128,128,0,0,0 // movups 0x80(%rax),%xmm8 |
| .byte 68,15,16,136,160,0,0,0 // movups 0xa0(%rax),%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 65,15,88,224 // addps %xmm8,%xmm4 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 65,15,88,232 // addps %xmm8,%xmm5 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 65,15,88,240 // addps %xmm8,%xmm6 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 65,15,88,249 // addps %xmm9,%xmm7 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_nx_sse41 |
| .globl _sk_bilinear_nx_sse41 |
| FUNCTION(_sk_bilinear_nx_sse41) |
| _sk_bilinear_nx_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_px_sse41 |
| .globl _sk_bilinear_px_sse41 |
| FUNCTION(_sk_bilinear_px_sse41) |
| _sk_bilinear_px_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_ny_sse41 |
| .globl _sk_bilinear_ny_sse41 |
| FUNCTION(_sk_bilinear_ny_sse41) |
| _sk_bilinear_ny_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_py_sse41 |
| .globl _sk_bilinear_py_sse41 |
| FUNCTION(_sk_bilinear_py_sse41) |
| _sk_bilinear_py_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3x_sse41 |
| .globl _sk_bicubic_n3x_sse41 |
| FUNCTION(_sk_bicubic_n3x_sse41) |
| _sk_bicubic_n3x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1x_sse41 |
| .globl _sk_bicubic_n1x_sse41 |
| FUNCTION(_sk_bicubic_n1x_sse41) |
| _sk_bicubic_n1x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1x_sse41 |
| .globl _sk_bicubic_p1x_sse41 |
| FUNCTION(_sk_bicubic_p1x_sse41) |
| _sk_bicubic_p1x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 68,15,17,144,128,0,0,0 // movups %xmm10,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3x_sse41 |
| .globl _sk_bicubic_p3x_sse41 |
| FUNCTION(_sk_bicubic_p3x_sse41) |
| _sk_bicubic_p3x_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 69,15,89,201 // mulps %xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,194 // addps %xmm10,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3y_sse41 |
| .globl _sk_bicubic_n3y_sse41 |
| FUNCTION(_sk_bicubic_n3y_sse41) |
| _sk_bicubic_n3y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1y_sse41 |
| .globl _sk_bicubic_n1y_sse41 |
| FUNCTION(_sk_bicubic_n1y_sse41) |
| _sk_bicubic_n1y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1y_sse41 |
| .globl _sk_bicubic_p1y_sse41 |
| FUNCTION(_sk_bicubic_p1y_sse41) |
| _sk_bicubic_p1y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 68,15,17,144,160,0,0,0 // movups %xmm10,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3y_sse41 |
| .globl _sk_bicubic_p3y_sse41 |
| FUNCTION(_sk_bicubic_p3y_sse41) |
| _sk_bicubic_p3y_sse41: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 69,15,89,201 // mulps %xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,194 // addps %xmm10,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_callback_sse41 |
| .globl _sk_callback_sse41 |
| FUNCTION(_sk_callback_sse41) |
| _sk_callback_sse41: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 83 // push %rbx |
| .byte 72,129,236,128,0,0,0 // sub $0x80,%rsp |
| .byte 15,41,124,36,112 // movaps %xmm7,0x70(%rsp) |
| .byte 15,41,116,36,96 // movaps %xmm6,0x60(%rsp) |
| .byte 15,41,108,36,80 // movaps %xmm5,0x50(%rsp) |
| .byte 15,41,100,36,64 // movaps %xmm4,0x40(%rsp) |
| .byte 15,41,92,36,48 // movaps %xmm3,0x30(%rsp) |
| .byte 15,41,84,36,32 // movaps %xmm2,0x20(%rsp) |
| .byte 15,41,76,36,16 // movaps %xmm1,0x10(%rsp) |
| .byte 15,41,4,36 // movaps %xmm0,(%rsp) |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 73,137,255 // mov %rdi,%r15 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,137,243 // mov %rsi,%rbx |
| .byte 72,139,120,8 // mov 0x8(%rax),%rdi |
| .byte 190,4,0,0,0 // mov $0x4,%esi |
| .byte 255,16 // callq *(%rax) |
| .byte 72,137,222 // mov %rbx,%rsi |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,255 // mov %r15,%rdi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 15,40,4,36 // movaps (%rsp),%xmm0 |
| .byte 15,40,76,36,16 // movaps 0x10(%rsp),%xmm1 |
| .byte 15,40,84,36,32 // movaps 0x20(%rsp),%xmm2 |
| .byte 15,40,92,36,48 // movaps 0x30(%rsp),%xmm3 |
| .byte 15,40,100,36,64 // movaps 0x40(%rsp),%xmm4 |
| .byte 15,40,108,36,80 // movaps 0x50(%rsp),%xmm5 |
| .byte 15,40,116,36,96 // movaps 0x60(%rsp),%xmm6 |
| .byte 15,40,124,36,112 // movaps 0x70(%rsp),%xmm7 |
| .byte 72,129,196,128,0,0,0 // add $0x80,%rsp |
| .byte 91 // pop %rbx |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_start_pipeline_sse2 |
| .globl _sk_start_pipeline_sse2 |
| FUNCTION(_sk_start_pipeline_sse2) |
| _sk_start_pipeline_sse2: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 65,85 // push %r13 |
| .byte 65,84 // push %r12 |
| .byte 83 // push %rbx |
| .byte 73,137,207 // mov %rcx,%r15 |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 72,137,251 // mov %rdi,%rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,196 // mov %rax,%r12 |
| .byte 73,137,245 // mov %rsi,%r13 |
| .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| .byte 76,57,248 // cmp %r15,%rax |
| .byte 118,5 // jbe 28 <_sk_start_pipeline_sse2+0x28> |
| .byte 72,137,216 // mov %rbx,%rax |
| .byte 235,52 // jmp 5c <_sk_start_pipeline_sse2+0x5c> |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 15,87,228 // xorps %xmm4,%xmm4 |
| .byte 15,87,237 // xorps %xmm5,%xmm5 |
| .byte 15,87,246 // xorps %xmm6,%xmm6 |
| .byte 15,87,255 // xorps %xmm7,%xmm7 |
| .byte 72,137,223 // mov %rbx,%rdi |
| .byte 76,137,238 // mov %r13,%rsi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 65,255,212 // callq *%r12 |
| .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| .byte 72,131,195,8 // add $0x8,%rbx |
| .byte 76,57,251 // cmp %r15,%rbx |
| .byte 72,137,195 // mov %rax,%rbx |
| .byte 118,204 // jbe 28 <_sk_start_pipeline_sse2+0x28> |
| .byte 91 // pop %rbx |
| .byte 65,92 // pop %r12 |
| .byte 65,93 // pop %r13 |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 195 // retq |
| |
| HIDDEN _sk_just_return_sse2 |
| .globl _sk_just_return_sse2 |
| FUNCTION(_sk_just_return_sse2) |
| _sk_just_return_sse2: |
| .byte 195 // retq |
| |
| HIDDEN _sk_seed_shader_sse2 |
| .globl _sk_seed_shader_sse2 |
| FUNCTION(_sk_seed_shader_sse2) |
| _sk_seed_shader_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 102,15,110,199 // movd %edi,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,15,110,209 // movd %ecx,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 15,16,2 // movups (%rdx),%xmm0 |
| .byte 15,88,193 // addps %xmm1,%xmm0 |
| .byte 102,15,110,8 // movd (%rax),%xmm1 |
| .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 15,87,228 // xorps %xmm4,%xmm4 |
| .byte 15,87,237 // xorps %xmm5,%xmm5 |
| .byte 15,87,246 // xorps %xmm6,%xmm6 |
| .byte 15,87,255 // xorps %xmm7,%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_constant_color_sse2 |
| .globl _sk_constant_color_sse2 |
| FUNCTION(_sk_constant_color_sse2) |
| _sk_constant_color_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clear_sse2 |
| .globl _sk_clear_sse2 |
| FUNCTION(_sk_clear_sse2) |
| _sk_clear_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcatop_sse2 |
| .globl _sk_srcatop_sse2 |
| FUNCTION(_sk_srcatop_sse2) |
| _sk_srcatop_sse2: |
| .byte 15,89,199 // mulps %xmm7,%xmm0 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| .byte 15,89,207 // mulps %xmm7,%xmm1 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstatop_sse2 |
| .globl _sk_dstatop_sse2 |
| FUNCTION(_sk_dstatop_sse2) |
| _sk_dstatop_sse2: |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,198 // mulps %xmm6,%xmm8 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 65,15,88,208 // addps %xmm8,%xmm2 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcin_sse2 |
| .globl _sk_srcin_sse2 |
| FUNCTION(_sk_srcin_sse2) |
| _sk_srcin_sse2: |
| .byte 15,89,199 // mulps %xmm7,%xmm0 |
| .byte 15,89,207 // mulps %xmm7,%xmm1 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstin_sse2 |
| .globl _sk_dstin_sse2 |
| FUNCTION(_sk_dstin_sse2) |
| _sk_dstin_sse2: |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 15,40,211 // movaps %xmm3,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcout_sse2 |
| .globl _sk_srcout_sse2 |
| FUNCTION(_sk_srcout_sse2) |
| _sk_srcout_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstout_sse2 |
| .globl _sk_dstout_sse2 |
| FUNCTION(_sk_dstout_sse2) |
| _sk_dstout_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_srcover_sse2 |
| .globl _sk_srcover_sse2 |
| FUNCTION(_sk_srcover_sse2) |
| _sk_srcover_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_dstover_sse2 |
| .globl _sk_dstover_sse2 |
| FUNCTION(_sk_dstover_sse2) |
| _sk_dstover_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_modulate_sse2 |
| .globl _sk_modulate_sse2 |
| FUNCTION(_sk_modulate_sse2) |
| _sk_modulate_sse2: |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_multiply_sse2 |
| .globl _sk_multiply_sse2 |
| FUNCTION(_sk_multiply_sse2) |
| _sk_multiply_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,220 // mulps %xmm4,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 65,15,88,195 // addps %xmm11,%xmm0 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,221 // mulps %xmm5,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,222 // mulps %xmm6,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_plus__sse2 |
| .globl _sk_plus__sse2 |
| FUNCTION(_sk_plus__sse2) |
| _sk_plus__sse2: |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_screen_sse2 |
| .globl _sk_screen_sse2 |
| FUNCTION(_sk_screen_sse2) |
| _sk_screen_sse2: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 68,15,88,196 // addps %xmm4,%xmm8 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,88,205 // addps %xmm5,%xmm9 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| .byte 68,15,88,214 // addps %xmm6,%xmm10 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,92,210 // subps %xmm2,%xmm10 |
| .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| .byte 68,15,88,223 // addps %xmm7,%xmm11 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_xor__sse2 |
| .globl _sk_xor__sse2 |
| FUNCTION(_sk_xor__sse2) |
| _sk_xor__sse2: |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .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 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,92,216 // subps %xmm8,%xmm3 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,212 // mulps %xmm4,%xmm10 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,213 // mulps %xmm5,%xmm10 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 15,89,223 // mulps %xmm7,%xmm3 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_darken_sse2 |
| .globl _sk_darken_sse2 |
| FUNCTION(_sk_darken_sse2) |
| _sk_darken_sse2: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 68,15,95,201 // maxps %xmm1,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,95,193 // maxps %xmm9,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,95,209 // maxps %xmm9,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lighten_sse2 |
| .globl _sk_lighten_sse2 |
| FUNCTION(_sk_lighten_sse2) |
| _sk_lighten_sse2: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_difference_sse2 |
| .globl _sk_difference_sse2 |
| FUNCTION(_sk_difference_sse2) |
| _sk_difference_sse2: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_exclusion_sse2 |
| .globl _sk_exclusion_sse2 |
| FUNCTION(_sk_exclusion_sse2) |
| _sk_exclusion_sse2: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 15,40,200 // movaps %xmm0,%xmm1 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,89,215 // mulps %xmm7,%xmm2 |
| .byte 15,88,218 // addps %xmm2,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colorburn_sse2 |
| .globl _sk_colorburn_sse2 |
| FUNCTION(_sk_colorburn_sse2) |
| _sk_colorburn_sse2: |
| .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| .byte 68,15,40,224 // movaps %xmm0,%xmm12 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| .byte 65,15,89,196 // mulps %xmm12,%xmm0 |
| .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| .byte 15,40,207 // movaps %xmm7,%xmm1 |
| .byte 15,92,204 // subps %xmm4,%xmm1 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 65,15,94,204 // divps %xmm12,%xmm1 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,93,233 // minps %xmm1,%xmm13 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 65,15,40,204 // movaps %xmm12,%xmm1 |
| .byte 65,15,194,202,0 // cmpeqps %xmm10,%xmm1 |
| .byte 68,15,92,203 // subps %xmm3,%xmm9 |
| .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| .byte 68,15,88,240 // addps %xmm0,%xmm14 |
| .byte 68,15,84,225 // andps %xmm1,%xmm12 |
| .byte 65,15,85,206 // andnps %xmm14,%xmm1 |
| .byte 69,15,40,233 // movaps %xmm9,%xmm13 |
| .byte 68,15,89,236 // mulps %xmm4,%xmm13 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 65,15,86,204 // orps %xmm12,%xmm1 |
| .byte 68,15,40,228 // movaps %xmm4,%xmm12 |
| .byte 68,15,194,231,0 // cmpeqps %xmm7,%xmm12 |
| .byte 65,15,88,205 // addps %xmm13,%xmm1 |
| .byte 65,15,84,196 // andps %xmm12,%xmm0 |
| .byte 68,15,85,225 // andnps %xmm1,%xmm12 |
| .byte 65,15,86,196 // orps %xmm12,%xmm0 |
| .byte 65,15,40,203 // movaps %xmm11,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| .byte 68,15,92,229 // subps %xmm5,%xmm12 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 69,15,94,224 // divps %xmm8,%xmm12 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| .byte 69,15,194,234,0 // cmpeqps %xmm10,%xmm13 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 68,15,88,225 // addps %xmm1,%xmm12 |
| .byte 69,15,84,197 // andps %xmm13,%xmm8 |
| .byte 69,15,85,236 // andnps %xmm12,%xmm13 |
| .byte 69,15,86,232 // orps %xmm8,%xmm13 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 69,15,88,232 // addps %xmm8,%xmm13 |
| .byte 68,15,40,197 // movaps %xmm5,%xmm8 |
| .byte 68,15,194,199,0 // cmpeqps %xmm7,%xmm8 |
| .byte 65,15,84,200 // andps %xmm8,%xmm1 |
| .byte 69,15,85,197 // andnps %xmm13,%xmm8 |
| .byte 65,15,86,200 // orps %xmm8,%xmm1 |
| .byte 68,15,40,199 // movaps %xmm7,%xmm8 |
| .byte 68,15,92,198 // subps %xmm6,%xmm8 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 68,15,94,194 // divps %xmm2,%xmm8 |
| .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| .byte 69,15,93,224 // minps %xmm8,%xmm12 |
| .byte 68,15,40,199 // movaps %xmm7,%xmm8 |
| .byte 69,15,92,196 // subps %xmm12,%xmm8 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 68,15,194,210,0 // cmpeqps %xmm2,%xmm10 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 69,15,88,195 // addps %xmm11,%xmm8 |
| .byte 65,15,84,210 // andps %xmm10,%xmm2 |
| .byte 69,15,85,208 // andnps %xmm8,%xmm10 |
| .byte 69,15,40,195 // movaps %xmm11,%xmm8 |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 68,15,86,210 // orps %xmm2,%xmm10 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,88,210 // addps %xmm2,%xmm10 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 15,194,215,0 // cmpeqps %xmm7,%xmm2 |
| .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| .byte 65,15,85,210 // andnps %xmm10,%xmm2 |
| .byte 68,15,86,194 // orps %xmm2,%xmm8 |
| .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_colordodge_sse2 |
| .globl _sk_colordodge_sse2 |
| FUNCTION(_sk_colordodge_sse2) |
| _sk_colordodge_sse2: |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 69,15,92,225 // subps %xmm9,%xmm12 |
| .byte 69,15,94,196 // divps %xmm12,%xmm8 |
| .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 69,15,93,232 // minps %xmm8,%xmm13 |
| .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| .byte 68,15,194,243,0 // cmpeqps %xmm3,%xmm14 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 68,15,88,232 // addps %xmm0,%xmm13 |
| .byte 69,15,84,206 // andps %xmm14,%xmm9 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| .byte 69,15,86,241 // orps %xmm9,%xmm14 |
| .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 69,15,88,241 // addps %xmm9,%xmm14 |
| .byte 68,15,40,204 // movaps %xmm4,%xmm9 |
| .byte 69,15,194,200,0 // cmpeqps %xmm8,%xmm9 |
| .byte 65,15,84,193 // andps %xmm9,%xmm0 |
| .byte 69,15,85,206 // andnps %xmm14,%xmm9 |
| .byte 65,15,86,193 // orps %xmm9,%xmm0 |
| .byte 68,15,40,235 // movaps %xmm3,%xmm13 |
| .byte 68,15,89,237 // mulps %xmm5,%xmm13 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| .byte 69,15,94,233 // divps %xmm9,%xmm13 |
| .byte 69,15,40,203 // movaps %xmm11,%xmm9 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 69,15,93,229 // minps %xmm13,%xmm12 |
| .byte 68,15,40,233 // movaps %xmm1,%xmm13 |
| .byte 68,15,194,235,0 // cmpeqps %xmm3,%xmm13 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 69,15,88,225 // addps %xmm9,%xmm12 |
| .byte 65,15,84,205 // andps %xmm13,%xmm1 |
| .byte 69,15,85,236 // andnps %xmm12,%xmm13 |
| .byte 68,15,86,233 // orps %xmm1,%xmm13 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 68,15,88,205 // addps %xmm5,%xmm9 |
| .byte 68,15,88,233 // addps %xmm1,%xmm13 |
| .byte 15,40,205 // movaps %xmm5,%xmm1 |
| .byte 65,15,194,200,0 // cmpeqps %xmm8,%xmm1 |
| .byte 68,15,84,201 // andps %xmm1,%xmm9 |
| .byte 65,15,85,205 // andnps %xmm13,%xmm1 |
| .byte 68,15,86,201 // orps %xmm1,%xmm9 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 68,15,89,230 // mulps %xmm6,%xmm12 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,92,202 // subps %xmm2,%xmm1 |
| .byte 68,15,94,225 // divps %xmm1,%xmm12 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| .byte 15,40,202 // movaps %xmm2,%xmm1 |
| .byte 15,194,203,0 // cmpeqps %xmm3,%xmm1 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| .byte 15,84,209 // andps %xmm1,%xmm2 |
| .byte 65,15,85,205 // andnps %xmm13,%xmm1 |
| .byte 15,86,202 // orps %xmm2,%xmm1 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 68,15,194,198,0 // cmpeqps %xmm6,%xmm8 |
| .byte 68,15,88,222 // addps %xmm6,%xmm11 |
| .byte 69,15,84,216 // andps %xmm8,%xmm11 |
| .byte 68,15,85,193 // andnps %xmm1,%xmm8 |
| .byte 69,15,86,195 // orps %xmm11,%xmm8 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hardlight_sse2 |
| .globl _sk_hardlight_sse2 |
| FUNCTION(_sk_hardlight_sse2) |
| _sk_hardlight_sse2: |
| .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp) |
| .byte 15,40,245 // movaps %xmm5,%xmm6 |
| .byte 15,40,236 // movaps %xmm4,%xmm5 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| .byte 68,15,92,215 // subps %xmm7,%xmm10 |
| .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| .byte 69,15,40,203 // movaps %xmm11,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 69,15,88,200 // addps %xmm8,%xmm9 |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| .byte 15,40,227 // movaps %xmm3,%xmm4 |
| .byte 15,89,231 // mulps %xmm7,%xmm4 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| .byte 68,15,40,255 // movaps %xmm7,%xmm15 |
| .byte 68,15,92,253 // subps %xmm5,%xmm15 |
| .byte 69,15,89,248 // mulps %xmm8,%xmm15 |
| .byte 69,15,88,255 // addps %xmm15,%xmm15 |
| .byte 68,15,40,228 // movaps %xmm4,%xmm12 |
| .byte 69,15,92,231 // subps %xmm15,%xmm12 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| .byte 68,15,194,195,2 // cmpleps %xmm3,%xmm8 |
| .byte 15,89,197 // mulps %xmm5,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 65,15,84,192 // andps %xmm8,%xmm0 |
| .byte 69,15,85,196 // andnps %xmm12,%xmm8 |
| .byte 68,15,86,192 // orps %xmm0,%xmm8 |
| .byte 69,15,40,251 // movaps %xmm11,%xmm15 |
| .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| .byte 68,15,89,223 // mulps %xmm7,%xmm11 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 15,89,193 // mulps %xmm1,%xmm0 |
| .byte 68,15,89,254 // mulps %xmm6,%xmm15 |
| .byte 68,15,88,248 // addps %xmm0,%xmm15 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 68,15,92,238 // subps %xmm6,%xmm13 |
| .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 65,15,92,197 // subps %xmm13,%xmm0 |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| .byte 68,15,194,203,2 // cmpleps %xmm3,%xmm9 |
| .byte 15,89,206 // mulps %xmm6,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 65,15,84,201 // andps %xmm9,%xmm1 |
| .byte 68,15,85,200 // andnps %xmm0,%xmm9 |
| .byte 68,15,86,201 // orps %xmm1,%xmm9 |
| .byte 69,15,88,207 // addps %xmm15,%xmm9 |
| .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| .byte 68,15,40,108,36,232 // movaps -0x18(%rsp),%xmm13 |
| .byte 69,15,89,229 // mulps %xmm13,%xmm12 |
| .byte 69,15,88,226 // addps %xmm10,%xmm12 |
| .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| .byte 69,15,88,210 // addps %xmm10,%xmm10 |
| .byte 68,15,194,211,2 // cmpleps %xmm3,%xmm10 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,92,194 // subps %xmm2,%xmm0 |
| .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 68,15,89,240 // mulps %xmm0,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 65,15,92,230 // subps %xmm14,%xmm4 |
| .byte 65,15,84,210 // andps %xmm10,%xmm2 |
| .byte 68,15,85,212 // andnps %xmm4,%xmm10 |
| .byte 68,15,86,210 // orps %xmm2,%xmm10 |
| .byte 69,15,88,212 // addps %xmm12,%xmm10 |
| .byte 65,15,88,219 // addps %xmm11,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,40,229 // movaps %xmm5,%xmm4 |
| .byte 15,40,238 // movaps %xmm6,%xmm5 |
| .byte 65,15,40,245 // movaps %xmm13,%xmm6 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_overlay_sse2 |
| .globl _sk_overlay_sse2 |
| FUNCTION(_sk_overlay_sse2) |
| _sk_overlay_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,220 // mulps %xmm4,%xmm11 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| .byte 68,15,92,224 // subps %xmm0,%xmm12 |
| .byte 15,89,196 // mulps %xmm4,%xmm0 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,92,236 // subps %xmm4,%xmm13 |
| .byte 68,15,40,244 // movaps %xmm4,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 68,15,194,247,2 // cmpleps %xmm7,%xmm14 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| .byte 69,15,89,236 // mulps %xmm12,%xmm13 |
| .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| .byte 69,15,40,226 // movaps %xmm10,%xmm12 |
| .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| .byte 65,15,84,198 // andps %xmm14,%xmm0 |
| .byte 69,15,85,244 // andnps %xmm12,%xmm14 |
| .byte 65,15,86,198 // orps %xmm14,%xmm0 |
| .byte 65,15,88,195 // addps %xmm11,%xmm0 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 69,15,40,224 // movaps %xmm8,%xmm12 |
| .byte 68,15,89,229 // mulps %xmm5,%xmm12 |
| .byte 69,15,88,227 // addps %xmm11,%xmm12 |
| .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| .byte 68,15,92,217 // subps %xmm1,%xmm11 |
| .byte 15,89,205 // mulps %xmm5,%xmm1 |
| .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| .byte 68,15,92,237 // subps %xmm5,%xmm13 |
| .byte 68,15,40,245 // movaps %xmm5,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 68,15,194,247,2 // cmpleps %xmm7,%xmm14 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 69,15,89,235 // mulps %xmm11,%xmm13 |
| .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| .byte 69,15,92,221 // subps %xmm13,%xmm11 |
| .byte 65,15,84,206 // andps %xmm14,%xmm1 |
| .byte 69,15,85,243 // andnps %xmm11,%xmm14 |
| .byte 65,15,86,206 // orps %xmm14,%xmm1 |
| .byte 65,15,88,204 // addps %xmm12,%xmm1 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 68,15,89,222 // mulps %xmm6,%xmm11 |
| .byte 69,15,88,217 // addps %xmm9,%xmm11 |
| .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| .byte 68,15,92,202 // subps %xmm2,%xmm9 |
| .byte 15,89,214 // mulps %xmm6,%xmm2 |
| .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| .byte 68,15,92,230 // subps %xmm6,%xmm12 |
| .byte 68,15,40,238 // movaps %xmm6,%xmm13 |
| .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| .byte 68,15,194,239,2 // cmpleps %xmm7,%xmm13 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| .byte 69,15,88,228 // addps %xmm12,%xmm12 |
| .byte 69,15,92,212 // subps %xmm12,%xmm10 |
| .byte 65,15,84,213 // andps %xmm13,%xmm2 |
| .byte 69,15,85,234 // andnps %xmm10,%xmm13 |
| .byte 65,15,86,213 // orps %xmm13,%xmm2 |
| .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_softlight_sse2 |
| .globl _sk_softlight_sse2 |
| FUNCTION(_sk_softlight_sse2) |
| _sk_softlight_sse2: |
| .byte 15,41,84,36,232 // movaps %xmm2,-0x18(%rsp) |
| .byte 15,40,209 // movaps %xmm1,%xmm2 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 69,15,87,228 // xorps %xmm12,%xmm12 |
| .byte 68,15,194,231,1 // cmpltps %xmm7,%xmm12 |
| .byte 68,15,40,212 // movaps %xmm4,%xmm10 |
| .byte 68,15,94,215 // divps %xmm7,%xmm10 |
| .byte 69,15,84,212 // andps %xmm12,%xmm10 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| .byte 69,15,92,242 // subps %xmm10,%xmm14 |
| .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| .byte 65,15,82,194 // rsqrtps %xmm10,%xmm0 |
| .byte 68,15,83,248 // rcpps %xmm0,%xmm15 |
| .byte 69,15,92,250 // subps %xmm10,%xmm15 |
| .byte 69,15,88,210 // addps %xmm10,%xmm10 |
| .byte 69,15,88,210 // addps %xmm10,%xmm10 |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 15,89,192 // mulps %xmm0,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 69,15,92,217 // subps %xmm9,%xmm11 |
| .byte 68,15,89,216 // mulps %xmm0,%xmm11 |
| .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| .byte 15,40,204 // movaps %xmm4,%xmm1 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 15,88,228 // addps %xmm4,%xmm4 |
| .byte 15,88,228 // addps %xmm4,%xmm4 |
| .byte 15,194,231,2 // cmpleps %xmm7,%xmm4 |
| .byte 68,15,84,236 // andps %xmm4,%xmm13 |
| .byte 65,15,85,231 // andnps %xmm15,%xmm4 |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 15,88,192 // addps %xmm0,%xmm0 |
| .byte 65,15,86,229 // orps %xmm13,%xmm4 |
| .byte 68,15,40,232 // movaps %xmm0,%xmm13 |
| .byte 68,15,92,235 // subps %xmm3,%xmm13 |
| .byte 69,15,89,245 // mulps %xmm13,%xmm14 |
| .byte 68,15,89,239 // mulps %xmm7,%xmm13 |
| .byte 65,15,89,229 // mulps %xmm13,%xmm4 |
| .byte 65,15,88,227 // addps %xmm11,%xmm4 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| .byte 69,15,40,251 // movaps %xmm11,%xmm15 |
| .byte 15,41,76,36,216 // movaps %xmm1,-0x28(%rsp) |
| .byte 68,15,89,249 // mulps %xmm1,%xmm15 |
| .byte 69,15,40,233 // movaps %xmm9,%xmm13 |
| .byte 68,15,92,239 // subps %xmm7,%xmm13 |
| .byte 69,15,89,197 // mulps %xmm13,%xmm8 |
| .byte 69,15,88,199 // addps %xmm15,%xmm8 |
| .byte 68,15,88,243 // addps %xmm3,%xmm14 |
| .byte 68,15,89,241 // mulps %xmm1,%xmm14 |
| .byte 15,194,195,2 // cmpleps %xmm3,%xmm0 |
| .byte 68,15,84,240 // andps %xmm0,%xmm14 |
| .byte 15,85,196 // andnps %xmm4,%xmm0 |
| .byte 65,15,86,198 // orps %xmm14,%xmm0 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp) |
| .byte 68,15,40,197 // movaps %xmm5,%xmm8 |
| .byte 68,15,94,199 // divps %xmm7,%xmm8 |
| .byte 69,15,84,196 // andps %xmm12,%xmm8 |
| .byte 69,15,40,240 // movaps %xmm8,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| .byte 65,15,40,230 // movaps %xmm14,%xmm4 |
| .byte 15,89,228 // mulps %xmm4,%xmm4 |
| .byte 65,15,88,230 // addps %xmm14,%xmm4 |
| .byte 69,15,40,248 // movaps %xmm8,%xmm15 |
| .byte 69,15,92,249 // subps %xmm9,%xmm15 |
| .byte 68,15,89,252 // mulps %xmm4,%xmm15 |
| .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| .byte 69,15,92,240 // subps %xmm8,%xmm14 |
| .byte 65,15,82,224 // rsqrtps %xmm8,%xmm4 |
| .byte 15,83,228 // rcpps %xmm4,%xmm4 |
| .byte 65,15,92,224 // subps %xmm8,%xmm4 |
| .byte 69,15,89,194 // mulps %xmm10,%xmm8 |
| .byte 69,15,88,199 // addps %xmm15,%xmm8 |
| .byte 68,15,40,253 // movaps %xmm5,%xmm15 |
| .byte 69,15,88,255 // addps %xmm15,%xmm15 |
| .byte 69,15,88,255 // addps %xmm15,%xmm15 |
| .byte 68,15,194,255,2 // cmpleps %xmm7,%xmm15 |
| .byte 69,15,84,199 // andps %xmm15,%xmm8 |
| .byte 68,15,85,252 // andnps %xmm4,%xmm15 |
| .byte 69,15,86,248 // orps %xmm8,%xmm15 |
| .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| .byte 65,15,40,224 // movaps %xmm8,%xmm4 |
| .byte 15,92,227 // subps %xmm3,%xmm4 |
| .byte 68,15,89,244 // mulps %xmm4,%xmm14 |
| .byte 15,89,231 // mulps %xmm7,%xmm4 |
| .byte 68,15,89,252 // mulps %xmm4,%xmm15 |
| .byte 15,40,227 // movaps %xmm3,%xmm4 |
| .byte 15,89,229 // mulps %xmm5,%xmm4 |
| .byte 68,15,88,252 // addps %xmm4,%xmm15 |
| .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| .byte 15,89,229 // mulps %xmm5,%xmm4 |
| .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| .byte 15,88,212 // addps %xmm4,%xmm2 |
| .byte 68,15,88,243 // addps %xmm3,%xmm14 |
| .byte 68,15,89,245 // mulps %xmm5,%xmm14 |
| .byte 68,15,194,195,2 // cmpleps %xmm3,%xmm8 |
| .byte 69,15,84,240 // andps %xmm8,%xmm14 |
| .byte 69,15,85,199 // andnps %xmm15,%xmm8 |
| .byte 69,15,86,198 // orps %xmm14,%xmm8 |
| .byte 68,15,88,194 // addps %xmm2,%xmm8 |
| .byte 68,15,40,246 // movaps %xmm6,%xmm14 |
| .byte 65,15,40,206 // movaps %xmm14,%xmm1 |
| .byte 15,94,207 // divps %xmm7,%xmm1 |
| .byte 65,15,84,204 // andps %xmm12,%xmm1 |
| .byte 15,40,225 // movaps %xmm1,%xmm4 |
| .byte 65,15,92,225 // subps %xmm9,%xmm4 |
| .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 15,82,241 // rsqrtps %xmm1,%xmm6 |
| .byte 15,83,246 // rcpps %xmm6,%xmm6 |
| .byte 15,92,241 // subps %xmm1,%xmm6 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 15,40,233 // movaps %xmm1,%xmm5 |
| .byte 15,89,237 // mulps %xmm5,%xmm5 |
| .byte 15,88,233 // addps %xmm1,%xmm5 |
| .byte 15,89,236 // mulps %xmm4,%xmm5 |
| .byte 68,15,88,213 // addps %xmm5,%xmm10 |
| .byte 65,15,40,238 // movaps %xmm14,%xmm5 |
| .byte 15,40,205 // movaps %xmm5,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 15,88,201 // addps %xmm1,%xmm1 |
| .byte 15,194,207,2 // cmpleps %xmm7,%xmm1 |
| .byte 68,15,84,209 // andps %xmm1,%xmm10 |
| .byte 15,85,206 // andnps %xmm6,%xmm1 |
| .byte 15,40,84,36,232 // movaps -0x18(%rsp),%xmm2 |
| .byte 68,15,89,234 // mulps %xmm2,%xmm13 |
| .byte 15,88,210 // addps %xmm2,%xmm2 |
| .byte 65,15,86,202 // orps %xmm10,%xmm1 |
| .byte 15,40,226 // movaps %xmm2,%xmm4 |
| .byte 15,92,227 // subps %xmm3,%xmm4 |
| .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| .byte 15,89,231 // mulps %xmm7,%xmm4 |
| .byte 15,89,204 // mulps %xmm4,%xmm1 |
| .byte 15,40,227 // movaps %xmm3,%xmm4 |
| .byte 15,89,229 // mulps %xmm5,%xmm4 |
| .byte 15,88,204 // addps %xmm4,%xmm1 |
| .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| .byte 15,89,229 // mulps %xmm5,%xmm4 |
| .byte 65,15,88,229 // addps %xmm13,%xmm4 |
| .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| .byte 15,40,245 // movaps %xmm5,%xmm6 |
| .byte 15,194,211,2 // cmpleps %xmm3,%xmm2 |
| .byte 68,15,84,202 // andps %xmm2,%xmm9 |
| .byte 15,85,209 // andnps %xmm1,%xmm2 |
| .byte 65,15,86,209 // orps %xmm9,%xmm2 |
| .byte 15,88,212 // addps %xmm4,%xmm2 |
| .byte 68,15,89,223 // mulps %xmm7,%xmm11 |
| .byte 65,15,88,219 // addps %xmm11,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,100,36,216 // movaps -0x28(%rsp),%xmm4 |
| .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5 |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_0_sse2 |
| .globl _sk_clamp_0_sse2 |
| FUNCTION(_sk_clamp_0_sse2) |
| _sk_clamp_0_sse2: |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 65,15,95,192 // maxps %xmm8,%xmm0 |
| .byte 65,15,95,200 // maxps %xmm8,%xmm1 |
| .byte 65,15,95,208 // maxps %xmm8,%xmm2 |
| .byte 65,15,95,216 // maxps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_1_sse2 |
| .globl _sk_clamp_1_sse2 |
| FUNCTION(_sk_clamp_1_sse2) |
| _sk_clamp_1_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,93,192 // minps %xmm8,%xmm0 |
| .byte 65,15,93,200 // minps %xmm8,%xmm1 |
| .byte 65,15,93,208 // minps %xmm8,%xmm2 |
| .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_a_sse2 |
| .globl _sk_clamp_a_sse2 |
| FUNCTION(_sk_clamp_a_sse2) |
| _sk_clamp_a_sse2: |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| .byte 15,93,195 // minps %xmm3,%xmm0 |
| .byte 15,93,203 // minps %xmm3,%xmm1 |
| .byte 15,93,211 // minps %xmm3,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_set_rgb_sse2 |
| .globl _sk_set_rgb_sse2 |
| FUNCTION(_sk_set_rgb_sse2) |
| _sk_set_rgb_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_rb_sse2 |
| .globl _sk_swap_rb_sse2 |
| FUNCTION(_sk_swap_rb_sse2) |
| _sk_swap_rb_sse2: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_swap_sse2 |
| .globl _sk_swap_sse2 |
| FUNCTION(_sk_swap_sse2) |
| _sk_swap_sse2: |
| .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| .byte 68,15,40,202 // movaps %xmm2,%xmm9 |
| .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| .byte 68,15,40,216 // movaps %xmm0,%xmm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,40,205 // movaps %xmm5,%xmm1 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 15,40,223 // movaps %xmm7,%xmm3 |
| .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| .byte 65,15,40,234 // movaps %xmm10,%xmm5 |
| .byte 65,15,40,241 // movaps %xmm9,%xmm6 |
| .byte 65,15,40,248 // movaps %xmm8,%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_src_dst_sse2 |
| .globl _sk_move_src_dst_sse2 |
| FUNCTION(_sk_move_src_dst_sse2) |
| _sk_move_src_dst_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,224 // movaps %xmm0,%xmm4 |
| .byte 15,40,233 // movaps %xmm1,%xmm5 |
| .byte 15,40,242 // movaps %xmm2,%xmm6 |
| .byte 15,40,251 // movaps %xmm3,%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_move_dst_src_sse2 |
| .globl _sk_move_dst_src_sse2 |
| FUNCTION(_sk_move_dst_src_sse2) |
| _sk_move_dst_src_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,40,205 // movaps %xmm5,%xmm1 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 15,40,223 // movaps %xmm7,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_premul_sse2 |
| .globl _sk_premul_sse2 |
| FUNCTION(_sk_premul_sse2) |
| _sk_premul_sse2: |
| .byte 15,89,195 // mulps %xmm3,%xmm0 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_unpremul_sse2 |
| .globl _sk_unpremul_sse2 |
| FUNCTION(_sk_unpremul_sse2) |
| _sk_unpremul_sse2: |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,94,203 // divps %xmm3,%xmm9 |
| .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8 |
| .byte 69,15,84,193 // andps %xmm9,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_srgb_sse2 |
| .globl _sk_from_srgb_sse2 |
| FUNCTION(_sk_from_srgb_sse2) |
| _sk_from_srgb_sse2: |
| .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| .byte 68,15,40,224 // movaps %xmm0,%xmm12 |
| .byte 69,15,89,228 // mulps %xmm12,%xmm12 |
| .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| .byte 68,15,89,240 // mulps %xmm0,%xmm14 |
| .byte 69,15,88,242 // addps %xmm10,%xmm14 |
| .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,244 // mulps %xmm12,%xmm14 |
| .byte 69,15,88,243 // addps %xmm11,%xmm14 |
| .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 65,15,194,196,1 // cmpltps %xmm12,%xmm0 |
| .byte 68,15,84,232 // andps %xmm0,%xmm13 |
| .byte 65,15,85,198 // andnps %xmm14,%xmm0 |
| .byte 65,15,86,197 // orps %xmm13,%xmm0 |
| .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| .byte 68,15,89,233 // mulps %xmm1,%xmm13 |
| .byte 68,15,40,241 // movaps %xmm1,%xmm14 |
| .byte 69,15,89,246 // mulps %xmm14,%xmm14 |
| .byte 69,15,40,249 // movaps %xmm9,%xmm15 |
| .byte 68,15,89,249 // mulps %xmm1,%xmm15 |
| .byte 69,15,88,250 // addps %xmm10,%xmm15 |
| .byte 69,15,89,254 // mulps %xmm14,%xmm15 |
| .byte 69,15,88,251 // addps %xmm11,%xmm15 |
| .byte 65,15,194,204,1 // cmpltps %xmm12,%xmm1 |
| .byte 68,15,84,233 // andps %xmm1,%xmm13 |
| .byte 65,15,85,207 // andnps %xmm15,%xmm1 |
| .byte 65,15,86,205 // orps %xmm13,%xmm1 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 68,15,40,234 // movaps %xmm2,%xmm13 |
| .byte 69,15,89,237 // mulps %xmm13,%xmm13 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 69,15,89,205 // mulps %xmm13,%xmm9 |
| .byte 69,15,88,203 // addps %xmm11,%xmm9 |
| .byte 65,15,194,212,1 // cmpltps %xmm12,%xmm2 |
| .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| .byte 65,15,85,209 // andnps %xmm9,%xmm2 |
| .byte 65,15,86,208 // orps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_srgb_sse2 |
| .globl _sk_to_srgb_sse2 |
| FUNCTION(_sk_to_srgb_sse2) |
| _sk_to_srgb_sse2: |
| .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 |
| .byte 69,15,83,248 // rcpps %xmm8,%xmm15 |
| .byte 69,15,82,232 // rsqrtps %xmm8,%xmm13 |
| .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,40,240 // movaps %xmm8,%xmm14 |
| .byte 68,15,89,240 // mulps %xmm0,%xmm14 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,89,251 // mulps %xmm11,%xmm15 |
| .byte 69,15,88,252 // addps %xmm12,%xmm15 |
| .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| .byte 69,15,88,239 // addps %xmm15,%xmm13 |
| .byte 69,15,40,249 // movaps %xmm9,%xmm15 |
| .byte 69,15,93,253 // minps %xmm13,%xmm15 |
| .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 65,15,194,197,1 // cmpltps %xmm13,%xmm0 |
| .byte 68,15,84,240 // andps %xmm0,%xmm14 |
| .byte 65,15,85,199 // andnps %xmm15,%xmm0 |
| .byte 65,15,86,198 // orps %xmm14,%xmm0 |
| .byte 68,15,82,241 // rsqrtps %xmm1,%xmm14 |
| .byte 69,15,83,254 // rcpps %xmm14,%xmm15 |
| .byte 69,15,82,246 // rsqrtps %xmm14,%xmm14 |
| .byte 69,15,89,251 // mulps %xmm11,%xmm15 |
| .byte 69,15,88,252 // addps %xmm12,%xmm15 |
| .byte 69,15,89,242 // mulps %xmm10,%xmm14 |
| .byte 69,15,88,247 // addps %xmm15,%xmm14 |
| .byte 69,15,40,249 // movaps %xmm9,%xmm15 |
| .byte 69,15,93,254 // minps %xmm14,%xmm15 |
| .byte 69,15,40,240 // movaps %xmm8,%xmm14 |
| .byte 68,15,89,241 // mulps %xmm1,%xmm14 |
| .byte 65,15,194,205,1 // cmpltps %xmm13,%xmm1 |
| .byte 68,15,84,241 // andps %xmm1,%xmm14 |
| .byte 65,15,85,207 // andnps %xmm15,%xmm1 |
| .byte 65,15,86,206 // orps %xmm14,%xmm1 |
| .byte 68,15,82,242 // rsqrtps %xmm2,%xmm14 |
| .byte 69,15,83,254 // rcpps %xmm14,%xmm15 |
| .byte 69,15,89,251 // mulps %xmm11,%xmm15 |
| .byte 69,15,88,252 // addps %xmm12,%xmm15 |
| .byte 69,15,82,222 // rsqrtps %xmm14,%xmm11 |
| .byte 69,15,89,218 // mulps %xmm10,%xmm11 |
| .byte 69,15,88,223 // addps %xmm15,%xmm11 |
| .byte 69,15,93,203 // minps %xmm11,%xmm9 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 65,15,194,213,1 // cmpltps %xmm13,%xmm2 |
| .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| .byte 65,15,85,209 // andnps %xmm9,%xmm2 |
| .byte 65,15,86,208 // orps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_from_2dot2_sse2 |
| .globl _sk_from_2dot2_sse2 |
| FUNCTION(_sk_from_2dot2_sse2) |
| _sk_from_2dot2_sse2: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 65,15,82,192 // rsqrtps %xmm8,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 68,15,82,200 // rsqrtps %xmm0,%xmm9 |
| .byte 65,15,82,193 // rsqrtps %xmm9,%xmm0 |
| .byte 68,15,82,208 // rsqrtps %xmm0,%xmm10 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 15,89,192 // mulps %xmm0,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| .byte 65,15,95,194 // maxps %xmm10,%xmm0 |
| .byte 68,15,82,193 // rsqrtps %xmm1,%xmm8 |
| .byte 69,15,82,192 // rsqrtps %xmm8,%xmm8 |
| .byte 69,15,82,192 // rsqrtps %xmm8,%xmm8 |
| .byte 69,15,82,200 // rsqrtps %xmm8,%xmm9 |
| .byte 69,15,82,193 // rsqrtps %xmm9,%xmm8 |
| .byte 69,15,82,216 // rsqrtps %xmm8,%xmm11 |
| .byte 15,89,201 // mulps %xmm1,%xmm1 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 69,15,89,195 // mulps %xmm11,%xmm8 |
| .byte 69,15,95,194 // maxps %xmm10,%xmm8 |
| .byte 15,82,202 // rsqrtps %xmm2,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 68,15,82,217 // rsqrtps %xmm1,%xmm11 |
| .byte 65,15,82,203 // rsqrtps %xmm11,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,89,210 // mulps %xmm2,%xmm2 |
| .byte 69,15,40,203 // movaps %xmm11,%xmm9 |
| .byte 69,15,89,201 // mulps %xmm9,%xmm9 |
| .byte 69,15,89,203 // mulps %xmm11,%xmm9 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 69,15,95,202 // maxps %xmm10,%xmm9 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_to_2dot2_sse2 |
| .globl _sk_to_2dot2_sse2 |
| FUNCTION(_sk_to_2dot2_sse2) |
| _sk_to_2dot2_sse2: |
| .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 |
| .byte 65,15,82,192 // rsqrtps %xmm8,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| .byte 68,15,82,200 // rsqrtps %xmm0,%xmm9 |
| .byte 69,15,83,192 // rcpps %xmm8,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 65,15,83,193 // rcpps %xmm9,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 65,15,95,192 // maxps %xmm8,%xmm0 |
| .byte 68,15,82,201 // rsqrtps %xmm1,%xmm9 |
| .byte 65,15,82,201 // rsqrtps %xmm9,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 15,82,201 // rsqrtps %xmm1,%xmm1 |
| .byte 68,15,82,209 // rsqrtps %xmm1,%xmm10 |
| .byte 69,15,83,201 // rcpps %xmm9,%xmm9 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 65,15,83,202 // rcpps %xmm10,%xmm1 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 65,15,95,200 // maxps %xmm8,%xmm1 |
| .byte 68,15,82,202 // rsqrtps %xmm2,%xmm9 |
| .byte 65,15,82,209 // rsqrtps %xmm9,%xmm2 |
| .byte 15,82,210 // rsqrtps %xmm2,%xmm2 |
| .byte 15,82,210 // rsqrtps %xmm2,%xmm2 |
| .byte 15,82,210 // rsqrtps %xmm2,%xmm2 |
| .byte 68,15,82,210 // rsqrtps %xmm2,%xmm10 |
| .byte 69,15,83,201 // rcpps %xmm9,%xmm9 |
| .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| .byte 65,15,83,210 // rcpps %xmm10,%xmm2 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 65,15,95,208 // maxps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_rgb_to_hsl_sse2 |
| .globl _sk_rgb_to_hsl_sse2 |
| FUNCTION(_sk_rgb_to_hsl_sse2) |
| _sk_rgb_to_hsl_sse2: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| .byte 69,15,95,209 // maxps %xmm9,%xmm10 |
| .byte 68,15,95,210 // maxps %xmm2,%xmm10 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 69,15,93,217 // minps %xmm9,%xmm11 |
| .byte 68,15,93,218 // minps %xmm2,%xmm11 |
| .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| .byte 65,15,92,203 // subps %xmm11,%xmm1 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,94,225 // divps %xmm1,%xmm12 |
| .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d |
| .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 |
| .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| .byte 68,15,194,242,1 // cmpltps %xmm2,%xmm14 |
| .byte 185,0,0,192,64 // mov $0x40c00000,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,84,238 // andps %xmm14,%xmm13 |
| .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| .byte 68,15,92,242 // subps %xmm2,%xmm14 |
| .byte 69,15,89,244 // mulps %xmm12,%xmm14 |
| .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| .byte 69,15,40,242 // movaps %xmm10,%xmm14 |
| .byte 69,15,194,241,0 // cmpeqps %xmm9,%xmm14 |
| .byte 65,15,92,208 // subps %xmm8,%xmm2 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 65,15,89,212 // mulps %xmm12,%xmm2 |
| .byte 185,0,0,0,64 // mov $0x40000000,%ecx |
| .byte 69,15,89,196 // mulps %xmm12,%xmm8 |
| .byte 184,0,0,128,64 // mov $0x40800000,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,88,200 // addps %xmm8,%xmm9 |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,88,208 // addps %xmm8,%xmm2 |
| .byte 65,15,84,214 // andps %xmm14,%xmm2 |
| .byte 69,15,85,241 // andnps %xmm9,%xmm14 |
| .byte 68,15,86,242 // orps %xmm2,%xmm14 |
| .byte 68,15,84,232 // andps %xmm0,%xmm13 |
| .byte 65,15,85,198 // andnps %xmm14,%xmm0 |
| .byte 102,65,15,110,208 // movd %r8d,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,86,197 // orps %xmm13,%xmm0 |
| .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| .byte 69,15,194,203,4 // cmpneqps %xmm11,%xmm9 |
| .byte 65,15,84,193 // andps %xmm9,%xmm0 |
| .byte 15,89,194 // mulps %xmm2,%xmm0 |
| .byte 69,15,92,194 // subps %xmm10,%xmm8 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,92,195 // subps %xmm11,%xmm8 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 65,15,89,212 // mulps %xmm12,%xmm2 |
| .byte 68,15,194,226,1 // cmpltps %xmm2,%xmm12 |
| .byte 69,15,84,196 // andps %xmm12,%xmm8 |
| .byte 69,15,85,226 // andnps %xmm10,%xmm12 |
| .byte 69,15,86,224 // orps %xmm8,%xmm12 |
| .byte 65,15,94,204 // divps %xmm12,%xmm1 |
| .byte 65,15,84,201 // andps %xmm9,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_hsl_to_rgb_sse2 |
| .globl _sk_hsl_to_rgb_sse2 |
| FUNCTION(_sk_hsl_to_rgb_sse2) |
| _sk_hsl_to_rgb_sse2: |
| .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp) |
| .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp) |
| .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp) |
| .byte 15,41,100,36,184 // movaps %xmm4,-0x48(%rsp) |
| .byte 15,41,92,36,168 // movaps %xmm3,-0x58(%rsp) |
| .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| .byte 15,40,209 // movaps %xmm1,%xmm2 |
| .byte 15,40,240 // movaps %xmm0,%xmm6 |
| .byte 184,0,0,0,63 // mov $0x3f000000,%eax |
| .byte 102,68,15,110,240 // movd %eax,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| .byte 69,15,194,206,1 // cmpltps %xmm14,%xmm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 15,88,194 // addps %xmm2,%xmm0 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 15,40,218 // movaps %xmm2,%xmm3 |
| .byte 69,15,87,219 // xorps %xmm11,%xmm11 |
| .byte 68,15,194,218,0 // cmpeqps %xmm2,%xmm11 |
| .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| .byte 65,15,89,218 // mulps %xmm10,%xmm3 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 65,15,84,193 // andps %xmm9,%xmm0 |
| .byte 68,15,85,202 // andnps %xmm2,%xmm9 |
| .byte 68,15,86,200 // orps %xmm0,%xmm9 |
| .byte 184,0,0,0,64 // mov $0x40000000,%eax |
| .byte 185,171,170,170,62 // mov $0x3eaaaaab,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,41,68,36,152 // movaps %xmm8,-0x68(%rsp) |
| .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| .byte 185,0,0,0,0 // mov $0x0,%ecx |
| .byte 102,15,110,233 // movd %ecx,%xmm5 |
| .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 65,15,194,192,1 // cmpltps %xmm8,%xmm0 |
| .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| .byte 65,15,92,220 // subps %xmm12,%xmm3 |
| .byte 15,84,216 // andps %xmm0,%xmm3 |
| .byte 65,15,85,192 // andnps %xmm8,%xmm0 |
| .byte 15,86,195 // orps %xmm3,%xmm0 |
| .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| .byte 15,194,221,1 // cmpltps %xmm5,%xmm3 |
| .byte 65,15,40,212 // movaps %xmm12,%xmm2 |
| .byte 65,15,88,208 // addps %xmm8,%xmm2 |
| .byte 15,84,211 // andps %xmm3,%xmm2 |
| .byte 15,85,216 // andnps %xmm0,%xmm3 |
| .byte 15,86,218 // orps %xmm2,%xmm3 |
| .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| .byte 69,15,92,233 // subps %xmm9,%xmm13 |
| .byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 65,15,92,197 // subps %xmm13,%xmm0 |
| .byte 185,0,0,192,64 // mov $0x40c00000,%ecx |
| .byte 102,68,15,110,249 // movd %ecx,%xmm15 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 68,15,89,248 // mulps %xmm0,%xmm15 |
| .byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx |
| .byte 102,15,110,225 // movd %ecx,%xmm4 |
| .byte 15,198,228,0 // shufps $0x0,%xmm4,%xmm4 |
| .byte 15,40,212 // movaps %xmm4,%xmm2 |
| .byte 15,92,211 // subps %xmm3,%xmm2 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 15,40,195 // movaps %xmm3,%xmm0 |
| .byte 15,194,220,1 // cmpltps %xmm4,%xmm3 |
| .byte 65,15,89,215 // mulps %xmm15,%xmm2 |
| .byte 65,15,88,213 // addps %xmm13,%xmm2 |
| .byte 15,84,211 // andps %xmm3,%xmm2 |
| .byte 65,15,85,221 // andnps %xmm13,%xmm3 |
| .byte 15,86,218 // orps %xmm2,%xmm3 |
| .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| .byte 15,84,208 // andps %xmm0,%xmm2 |
| .byte 15,85,195 // andnps %xmm3,%xmm0 |
| .byte 15,86,194 // orps %xmm2,%xmm0 |
| .byte 102,15,110,248 // movd %eax,%xmm7 |
| .byte 15,198,255,0 // shufps $0x0,%xmm7,%xmm7 |
| .byte 15,194,207,1 // cmpltps %xmm7,%xmm1 |
| .byte 69,15,89,199 // mulps %xmm15,%xmm8 |
| .byte 69,15,88,197 // addps %xmm13,%xmm8 |
| .byte 68,15,84,193 // andps %xmm1,%xmm8 |
| .byte 15,85,200 // andnps %xmm0,%xmm1 |
| .byte 65,15,86,200 // orps %xmm8,%xmm1 |
| .byte 69,15,40,195 // movaps %xmm11,%xmm8 |
| .byte 68,15,85,193 // andnps %xmm1,%xmm8 |
| .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| .byte 15,194,198,1 // cmpltps %xmm6,%xmm0 |
| .byte 15,40,206 // movaps %xmm6,%xmm1 |
| .byte 65,15,92,204 // subps %xmm12,%xmm1 |
| .byte 15,84,200 // andps %xmm0,%xmm1 |
| .byte 15,85,198 // andnps %xmm6,%xmm0 |
| .byte 15,86,193 // orps %xmm1,%xmm0 |
| .byte 15,40,206 // movaps %xmm6,%xmm1 |
| .byte 15,194,205,1 // cmpltps %xmm5,%xmm1 |
| .byte 65,15,40,212 // movaps %xmm12,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,84,209 // andps %xmm1,%xmm2 |
| .byte 15,85,200 // andnps %xmm0,%xmm1 |
| .byte 15,86,202 // orps %xmm2,%xmm1 |
| .byte 15,40,196 // movaps %xmm4,%xmm0 |
| .byte 15,92,193 // subps %xmm1,%xmm0 |
| .byte 15,40,217 // movaps %xmm1,%xmm3 |
| .byte 15,40,209 // movaps %xmm1,%xmm2 |
| .byte 15,194,204,1 // cmpltps %xmm4,%xmm1 |
| .byte 65,15,89,199 // mulps %xmm15,%xmm0 |
| .byte 65,15,88,197 // addps %xmm13,%xmm0 |
| .byte 15,84,193 // andps %xmm1,%xmm0 |
| .byte 65,15,85,205 // andnps %xmm13,%xmm1 |
| .byte 15,86,200 // orps %xmm0,%xmm1 |
| .byte 65,15,194,214,1 // cmpltps %xmm14,%xmm2 |
| .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| .byte 15,84,194 // andps %xmm2,%xmm0 |
| .byte 15,85,209 // andnps %xmm1,%xmm2 |
| .byte 15,86,208 // orps %xmm0,%xmm2 |
| .byte 15,194,223,1 // cmpltps %xmm7,%xmm3 |
| .byte 65,15,40,199 // movaps %xmm15,%xmm0 |
| .byte 15,89,198 // mulps %xmm6,%xmm0 |
| .byte 65,15,88,197 // addps %xmm13,%xmm0 |
| .byte 15,84,195 // andps %xmm3,%xmm0 |
| .byte 15,85,218 // andnps %xmm2,%xmm3 |
| .byte 15,86,216 // orps %xmm0,%xmm3 |
| .byte 65,15,40,203 // movaps %xmm11,%xmm1 |
| .byte 15,85,203 // andnps %xmm3,%xmm1 |
| .byte 15,92,116,36,152 // subps -0x68(%rsp),%xmm6 |
| .byte 15,40,198 // movaps %xmm6,%xmm0 |
| .byte 15,194,197,1 // cmpltps %xmm5,%xmm0 |
| .byte 15,40,214 // movaps %xmm6,%xmm2 |
| .byte 65,15,92,212 // subps %xmm12,%xmm2 |
| .byte 65,15,40,220 // movaps %xmm12,%xmm3 |
| .byte 68,15,194,230,1 // cmpltps %xmm6,%xmm12 |
| .byte 65,15,84,212 // andps %xmm12,%xmm2 |
| .byte 68,15,85,230 // andnps %xmm6,%xmm12 |
| .byte 68,15,86,226 // orps %xmm2,%xmm12 |
| .byte 15,88,222 // addps %xmm6,%xmm3 |
| .byte 15,84,216 // andps %xmm0,%xmm3 |
| .byte 65,15,85,196 // andnps %xmm12,%xmm0 |
| .byte 15,86,195 // orps %xmm3,%xmm0 |
| .byte 15,40,232 // movaps %xmm0,%xmm5 |
| .byte 15,194,239,1 // cmpltps %xmm7,%xmm5 |
| .byte 15,40,208 // movaps %xmm0,%xmm2 |
| .byte 15,194,212,1 // cmpltps %xmm4,%xmm2 |
| .byte 15,92,224 // subps %xmm0,%xmm4 |
| .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| .byte 65,15,89,247 // mulps %xmm15,%xmm6 |
| .byte 65,15,89,231 // mulps %xmm15,%xmm4 |
| .byte 65,15,88,245 // addps %xmm13,%xmm6 |
| .byte 65,15,88,229 // addps %xmm13,%xmm4 |
| .byte 15,84,226 // andps %xmm2,%xmm4 |
| .byte 65,15,85,213 // andnps %xmm13,%xmm2 |
| .byte 15,86,212 // orps %xmm4,%xmm2 |
| .byte 68,15,84,200 // andps %xmm0,%xmm9 |
| .byte 15,85,194 // andnps %xmm2,%xmm0 |
| .byte 65,15,86,193 // orps %xmm9,%xmm0 |
| .byte 15,84,245 // andps %xmm5,%xmm6 |
| .byte 15,85,232 // andnps %xmm0,%xmm5 |
| .byte 15,86,238 // orps %xmm6,%xmm5 |
| .byte 69,15,84,211 // andps %xmm11,%xmm10 |
| .byte 68,15,85,221 // andnps %xmm5,%xmm11 |
| .byte 69,15,86,194 // orps %xmm10,%xmm8 |
| .byte 65,15,86,202 // orps %xmm10,%xmm1 |
| .byte 69,15,86,211 // orps %xmm11,%xmm10 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 15,40,92,36,168 // movaps -0x58(%rsp),%xmm3 |
| .byte 15,40,100,36,184 // movaps -0x48(%rsp),%xmm4 |
| .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5 |
| .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6 |
| .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_1_float_sse2 |
| .globl _sk_scale_1_float_sse2 |
| FUNCTION(_sk_scale_1_float_sse2) |
| _sk_scale_1_float_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_scale_u8_sse2 |
| .globl _sk_scale_u8_sse2 |
| FUNCTION(_sk_scale_u8_sse2) |
| _sk_scale_u8_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 102,68,15,110,4,56 // movd (%rax,%rdi,1),%xmm8 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8 |
| .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_1_float_sse2 |
| .globl _sk_lerp_1_float_sse2 |
| FUNCTION(_sk_lerp_1_float_sse2) |
| _sk_lerp_1_float_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,92,196 // subps %xmm4,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,92,205 // subps %xmm5,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,92,214 // subps %xmm6,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,92,223 // subps %xmm7,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_u8_sse2 |
| .globl _sk_lerp_u8_sse2 |
| FUNCTION(_sk_lerp_u8_sse2) |
| _sk_lerp_u8_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 102,68,15,110,4,56 // movd (%rax,%rdi,1),%xmm8 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8 |
| .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 15,92,196 // subps %xmm4,%xmm0 |
| .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,92,205 // subps %xmm5,%xmm1 |
| .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,92,214 // subps %xmm6,%xmm2 |
| .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .byte 15,92,223 // subps %xmm7,%xmm3 |
| .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| .byte 15,88,223 // addps %xmm7,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_lerp_565_sse2 |
| .globl _sk_lerp_565_sse2 |
| FUNCTION(_sk_lerp_565_sse2) |
| _sk_lerp_565_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,68,15,126,4,120 // movq (%rax,%rdi,2),%xmm8 |
| .byte 102,15,239,219 // pxor %xmm3,%xmm3 |
| .byte 102,68,15,97,195 // punpcklwd %xmm3,%xmm8 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%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 15,92,196 // subps %xmm4,%xmm0 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 15,88,196 // addps %xmm4,%xmm0 |
| .byte 15,92,205 // subps %xmm5,%xmm1 |
| .byte 65,15,89,203 // mulps %xmm11,%xmm1 |
| .byte 15,88,205 // addps %xmm5,%xmm1 |
| .byte 15,92,214 // subps %xmm6,%xmm2 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .byte 15,88,214 // addps %xmm6,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_tables_sse2 |
| .globl _sk_load_tables_sse2 |
| FUNCTION(_sk_load_tables_sse2) |
| _sk_load_tables_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 243,68,15,111,4,185 // movdqu (%rcx,%rdi,4),%xmm8 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 102,15,110,193 // movd %ecx,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| .byte 102,65,15,114,209,8 // psrld $0x8,%xmm9 |
| .byte 102,68,15,219,200 // pand %xmm0,%xmm9 |
| .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 |
| .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10 |
| .byte 102,68,15,219,208 // pand %xmm0,%xmm10 |
| .byte 102,65,15,219,192 // pand %xmm8,%xmm0 |
| .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3 |
| .byte 102,72,15,126,217 // movq %xmm3,%rcx |
| .byte 65,137,201 // mov %ecx,%r9d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,73,15,126,194 // movq %xmm0,%r10 |
| .byte 69,137,211 // mov %r10d,%r11d |
| .byte 73,193,234,32 // shr $0x20,%r10 |
| .byte 243,67,15,16,28,144 // movss (%r8,%r10,4),%xmm3 |
| .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0 |
| .byte 15,20,216 // unpcklps %xmm0,%xmm3 |
| .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0 |
| .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1 |
| .byte 15,20,193 // unpcklps %xmm1,%xmm0 |
| .byte 15,20,195 // unpcklps %xmm3,%xmm0 |
| .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| .byte 102,65,15,112,201,78 // pshufd $0x4e,%xmm9,%xmm1 |
| .byte 102,73,15,126,202 // movq %xmm1,%r10 |
| .byte 77,137,209 // mov %r10,%r9 |
| .byte 73,193,233,32 // shr $0x20,%r9 |
| .byte 102,76,15,126,201 // movq %xmm9,%rcx |
| .byte 65,137,203 // mov %ecx,%r11d |
| .byte 65,129,227,255,255,255,0 // and $0xffffff,%r11d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 65,129,226,255,255,255,0 // and $0xffffff,%r10d |
| .byte 243,65,15,16,28,8 // movss (%r8,%rcx,1),%xmm3 |
| .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1 |
| .byte 15,20,217 // unpcklps %xmm1,%xmm3 |
| .byte 243,67,15,16,12,152 // movss (%r8,%r11,4),%xmm1 |
| .byte 243,67,15,16,20,144 // movss (%r8,%r10,4),%xmm2 |
| .byte 15,20,202 // unpcklps %xmm2,%xmm1 |
| .byte 15,20,203 // unpcklps %xmm3,%xmm1 |
| .byte 76,139,72,24 // mov 0x18(%rax),%r9 |
| .byte 102,65,15,112,210,78 // pshufd $0x4e,%xmm10,%xmm2 |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 68,15,183,193 // movzwl %cx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,76,15,126,208 // movq %xmm10,%rax |
| .byte 68,15,183,208 // movzwl %ax,%r10d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 243,69,15,16,12,1 // movss (%r9,%rax,1),%xmm9 |
| .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2 |
| .byte 68,15,20,202 // unpcklps %xmm2,%xmm9 |
| .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3 |
| .byte 15,20,211 // unpcklps %xmm3,%xmm2 |
| .byte 65,15,20,209 // unpcklps %xmm9,%xmm2 |
| .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%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_load_tables_u16_be_sse2 |
| .globl _sk_load_tables_u16_be_sse2 |
| FUNCTION(_sk_load_tables_u16_be_sse2) |
| _sk_load_tables_u16_be_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 243,15,111,4,249 // movdqu (%rcx,%rdi,8),%xmm0 |
| .byte 243,15,111,76,249,16 // movdqu 0x10(%rcx,%rdi,8),%xmm1 |
| .byte 102,68,15,111,208 // movdqa %xmm0,%xmm10 |
| .byte 102,68,15,97,209 // punpcklwd %xmm1,%xmm10 |
| .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| .byte 102,65,15,111,202 // movdqa %xmm10,%xmm1 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,68,15,105,208 // punpckhwd %xmm0,%xmm10 |
| .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 |
| .byte 102,15,111,193 // movdqa %xmm1,%xmm0 |
| .byte 102,65,15,97,192 // punpcklwd %xmm8,%xmm0 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 102,68,15,112,203,0 // pshufd $0x0,%xmm3,%xmm9 |
| .byte 102,65,15,219,193 // pand %xmm9,%xmm0 |
| .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3 |
| .byte 102,72,15,126,217 // movq %xmm3,%rcx |
| .byte 68,15,183,201 // movzwl %cx,%r9d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,73,15,126,194 // movq %xmm0,%r10 |
| .byte 69,15,183,218 // movzwl %r10w,%r11d |
| .byte 73,193,234,30 // shr $0x1e,%r10 |
| .byte 243,67,15,16,28,16 // movss (%r8,%r10,1),%xmm3 |
| .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0 |
| .byte 15,20,216 // unpcklps %xmm0,%xmm3 |
| .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0 |
| .byte 243,67,15,16,20,136 // movss (%r8,%r9,4),%xmm2 |
| .byte 15,20,194 // unpcklps %xmm2,%xmm0 |
| .byte 15,20,195 // unpcklps %xmm3,%xmm0 |
| .byte 76,139,80,16 // mov 0x10(%rax),%r10 |
| .byte 102,15,115,217,8 // psrldq $0x8,%xmm1 |
| .byte 102,65,15,97,200 // punpcklwd %xmm8,%xmm1 |
| .byte 102,65,15,219,201 // pand %xmm9,%xmm1 |
| .byte 102,15,112,209,78 // pshufd $0x4e,%xmm1,%xmm2 |
| .byte 102,73,15,126,208 // movq %xmm2,%r8 |
| .byte 69,15,183,200 // movzwl %r8w,%r9d |
| .byte 73,193,232,32 // shr $0x20,%r8 |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 68,15,183,217 // movzwl %cx,%r11d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 243,65,15,16,20,10 // movss (%r10,%rcx,1),%xmm2 |
| .byte 243,67,15,16,12,130 // movss (%r10,%r8,4),%xmm1 |
| .byte 15,20,209 // unpcklps %xmm1,%xmm2 |
| .byte 243,67,15,16,12,154 // movss (%r10,%r11,4),%xmm1 |
| .byte 243,67,15,16,28,138 // movss (%r10,%r9,4),%xmm3 |
| .byte 15,20,203 // unpcklps %xmm3,%xmm1 |
| .byte 15,20,202 // unpcklps %xmm2,%xmm1 |
| .byte 76,139,72,24 // mov 0x18(%rax),%r9 |
| .byte 102,69,15,112,218,78 // pshufd $0x4e,%xmm10,%xmm11 |
| .byte 102,69,15,97,208 // punpcklwd %xmm8,%xmm10 |
| .byte 102,69,15,219,209 // pand %xmm9,%xmm10 |
| .byte 102,65,15,112,210,78 // pshufd $0x4e,%xmm10,%xmm2 |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 68,15,183,193 // movzwl %cx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,76,15,126,208 // movq %xmm10,%rax |
| .byte 68,15,183,208 // movzwl %ax,%r10d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 243,69,15,16,12,1 // movss (%r9,%rax,1),%xmm9 |
| .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2 |
| .byte 68,15,20,202 // unpcklps %xmm2,%xmm9 |
| .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3 |
| .byte 15,20,211 // unpcklps %xmm3,%xmm2 |
| .byte 65,15,20,209 // unpcklps %xmm9,%xmm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 102,65,15,111,219 // movdqa %xmm11,%xmm3 |
| .byte 102,15,113,243,8 // psllw $0x8,%xmm3 |
| .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 |
| .byte 102,68,15,235,219 // por %xmm3,%xmm11 |
| .byte 102,69,15,97,216 // punpcklwd %xmm8,%xmm11 |
| .byte 65,15,91,219 // cvtdq2ps %xmm11,%xmm3 |
| .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_tables_rgb_u16_be_sse2 |
| .globl _sk_load_tables_rgb_u16_be_sse2 |
| FUNCTION(_sk_load_tables_rgb_u16_be_sse2) |
| _sk_load_tables_rgb_u16_be_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9 |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| .byte 243,70,15,111,28,73 // movdqu (%rcx,%r9,2),%xmm11 |
| .byte 243,66,15,111,68,73,8 // movdqu 0x8(%rcx,%r9,2),%xmm0 |
| .byte 102,15,115,216,4 // psrldq $0x4,%xmm0 |
| .byte 102,65,15,111,219 // movdqa %xmm11,%xmm3 |
| .byte 102,15,115,219,6 // psrldq $0x6,%xmm3 |
| .byte 102,68,15,97,216 // punpcklwd %xmm0,%xmm11 |
| .byte 102,15,115,216,6 // psrldq $0x6,%xmm0 |
| .byte 102,15,97,216 // punpcklwd %xmm0,%xmm3 |
| .byte 102,65,15,111,195 // movdqa %xmm11,%xmm0 |
| .byte 102,15,97,195 // punpcklwd %xmm3,%xmm0 |
| .byte 102,68,15,112,208,78 // pshufd $0x4e,%xmm0,%xmm10 |
| .byte 102,68,15,105,219 // punpckhwd %xmm3,%xmm11 |
| .byte 102,69,15,239,192 // pxor %xmm8,%xmm8 |
| .byte 102,65,15,97,192 // punpcklwd %xmm8,%xmm0 |
| .byte 185,255,0,0,0 // mov $0xff,%ecx |
| .byte 102,15,110,217 // movd %ecx,%xmm3 |
| .byte 102,68,15,112,203,0 // pshufd $0x0,%xmm3,%xmm9 |
| .byte 102,65,15,219,193 // pand %xmm9,%xmm0 |
| .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3 |
| .byte 102,72,15,126,217 // movq %xmm3,%rcx |
| .byte 68,15,183,201 // movzwl %cx,%r9d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,73,15,126,194 // movq %xmm0,%r10 |
| .byte 69,15,183,218 // movzwl %r10w,%r11d |
| .byte 73,193,234,30 // shr $0x1e,%r10 |
| .byte 243,67,15,16,28,16 // movss (%r8,%r10,1),%xmm3 |
| .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0 |
| .byte 15,20,216 // unpcklps %xmm0,%xmm3 |
| .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0 |
| .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1 |
| .byte 15,20,193 // unpcklps %xmm1,%xmm0 |
| .byte 15,20,195 // unpcklps %xmm3,%xmm0 |
| .byte 76,139,80,16 // mov 0x10(%rax),%r10 |
| .byte 102,69,15,97,208 // punpcklwd %xmm8,%xmm10 |
| .byte 102,69,15,219,209 // pand %xmm9,%xmm10 |
| .byte 102,65,15,112,202,78 // pshufd $0x4e,%xmm10,%xmm1 |
| .byte 102,73,15,126,200 // movq %xmm1,%r8 |
| .byte 69,15,183,200 // movzwl %r8w,%r9d |
| .byte 73,193,232,32 // shr $0x20,%r8 |
| .byte 102,76,15,126,209 // movq %xmm10,%rcx |
| .byte 68,15,183,217 // movzwl %cx,%r11d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 243,65,15,16,28,10 // movss (%r10,%rcx,1),%xmm3 |
| .byte 243,67,15,16,12,130 // movss (%r10,%r8,4),%xmm1 |
| .byte 15,20,217 // unpcklps %xmm1,%xmm3 |
| .byte 243,67,15,16,12,154 // movss (%r10,%r11,4),%xmm1 |
| .byte 243,67,15,16,20,138 // movss (%r10,%r9,4),%xmm2 |
| .byte 15,20,202 // unpcklps %xmm2,%xmm1 |
| .byte 15,20,203 // unpcklps %xmm3,%xmm1 |
| .byte 76,139,72,24 // mov 0x18(%rax),%r9 |
| .byte 102,69,15,97,216 // punpcklwd %xmm8,%xmm11 |
| .byte 102,69,15,219,217 // pand %xmm9,%xmm11 |
| .byte 102,65,15,112,211,78 // pshufd $0x4e,%xmm11,%xmm2 |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 68,15,183,193 // movzwl %cx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,76,15,126,216 // movq %xmm11,%rax |
| .byte 68,15,183,208 // movzwl %ax,%r10d |
| .byte 72,193,232,30 // shr $0x1e,%rax |
| .byte 243,69,15,16,4,1 // movss (%r9,%rax,1),%xmm8 |
| .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2 |
| .byte 68,15,20,194 // unpcklps %xmm2,%xmm8 |
| .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3 |
| .byte 15,20,211 // unpcklps %xmm3,%xmm2 |
| .byte 65,15,20,208 // unpcklps %xmm8,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_byte_tables_sse2 |
| .globl _sk_byte_tables_sse2 |
| FUNCTION(_sk_byte_tables_sse2) |
| _sk_byte_tables_sse2: |
| .byte 65,86 // push %r14 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0 |
| .byte 102,73,15,126,193 // movq %xmm0,%r9 |
| .byte 69,137,202 // mov %r9d,%r10d |
| .byte 77,137,203 // mov %r9,%r11 |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 76,139,48 // mov (%rax),%r14 |
| .byte 76,139,72,8 // mov 0x8(%rax),%r9 |
| .byte 71,15,182,20,22 // movzbl (%r14,%r10,1),%r10d |
| .byte 67,15,182,28,30 // movzbl (%r14,%r11,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,211 // or %r10d,%ebx |
| .byte 71,15,182,4,6 // movzbl (%r14,%r8,1),%r8d |
| .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,193 // or %r8d,%ecx |
| .byte 102,15,196,193,0 // pinsrw $0x0,%ecx,%xmm0 |
| .byte 102,15,196,195,1 // pinsrw $0x1,%ebx,%xmm0 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 |
| .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 185,129,128,128,59 // mov $0x3b808081,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,15,112,201,78 // pshufd $0x4e,%xmm1,%xmm1 |
| .byte 102,72,15,126,203 // movq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d |
| .byte 65,15,182,28,25 // movzbl (%r9,%rbx,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,211 // or %r10d,%ebx |
| .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,193 // or %r8d,%ecx |
| .byte 102,15,196,201,0 // pinsrw $0x0,%ecx,%xmm1 |
| .byte 102,15,196,203,1 // pinsrw $0x1,%ebx,%xmm1 |
| .byte 102,65,15,96,201 // punpcklbw %xmm9,%xmm1 |
| .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,202 // mulps %xmm10,%xmm1 |
| .byte 76,139,72,16 // mov 0x10(%rax),%r9 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2 |
| .byte 102,72,15,126,211 // movq %xmm2,%rbx |
| .byte 65,137,216 // mov %ebx,%r8d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 102,15,112,210,78 // pshufd $0x4e,%xmm2,%xmm2 |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,209 // or %r10d,%ecx |
| .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d |
| .byte 65,15,182,28,25 // movzbl (%r9,%rbx,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,195 // or %r8d,%ebx |
| .byte 102,15,196,211,0 // pinsrw $0x0,%ebx,%xmm2 |
| .byte 102,15,196,209,1 // pinsrw $0x1,%ecx,%xmm2 |
| .byte 102,65,15,96,209 // punpcklbw %xmm9,%xmm2 |
| .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,210 // mulps %xmm10,%xmm2 |
| .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,65,15,91,216 // cvtps2dq %xmm8,%xmm3 |
| .byte 102,72,15,126,217 // movq %xmm3,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,15,112,219,78 // pshufd $0x4e,%xmm3,%xmm3 |
| .byte 102,72,15,126,219 // movq %xmm3,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 70,15,182,12,8 // movzbl (%rax,%r9,1),%r9d |
| .byte 15,182,28,24 // movzbl (%rax,%rbx,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,203 // or %r9d,%ebx |
| .byte 70,15,182,4,0 // movzbl (%rax,%r8,1),%r8d |
| .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax |
| .byte 193,224,8 // shl $0x8,%eax |
| .byte 68,9,192 // or %r8d,%eax |
| .byte 102,15,196,216,0 // pinsrw $0x0,%eax,%xmm3 |
| .byte 102,15,196,219,1 // pinsrw $0x1,%ebx,%xmm3 |
| .byte 102,65,15,96,217 // punpcklbw %xmm9,%xmm3 |
| .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,218 // mulps %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,94 // pop %r14 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_byte_tables_rgb_sse2 |
| .globl _sk_byte_tables_rgb_sse2 |
| FUNCTION(_sk_byte_tables_rgb_sse2) |
| _sk_byte_tables_rgb_sse2: |
| .byte 65,86 // push %r14 |
| .byte 83 // push %rbx |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 139,72,24 // mov 0x18(%rax),%ecx |
| .byte 255,201 // dec %ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0 |
| .byte 102,73,15,126,193 // movq %xmm0,%r9 |
| .byte 69,137,202 // mov %r9d,%r10d |
| .byte 77,137,203 // mov %r9,%r11 |
| .byte 73,193,235,32 // shr $0x20,%r11 |
| .byte 76,139,48 // mov (%rax),%r14 |
| .byte 76,139,72,8 // mov 0x8(%rax),%r9 |
| .byte 71,15,182,20,22 // movzbl (%r14,%r10,1),%r10d |
| .byte 67,15,182,28,30 // movzbl (%r14,%r11,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,211 // or %r10d,%ebx |
| .byte 71,15,182,4,6 // movzbl (%r14,%r8,1),%r8d |
| .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,193 // or %r8d,%ecx |
| .byte 102,15,196,193,0 // pinsrw $0x0,%ecx,%xmm0 |
| .byte 102,15,196,195,1 // pinsrw $0x1,%ebx,%xmm0 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0 |
| .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 185,129,128,128,59 // mov $0x3b808081,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1 |
| .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,15,112,201,78 // pshufd $0x4e,%xmm1,%xmm1 |
| .byte 102,72,15,126,203 // movq %xmm1,%rbx |
| .byte 65,137,218 // mov %ebx,%r10d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d |
| .byte 65,15,182,28,25 // movzbl (%r9,%rbx,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,211 // or %r10d,%ebx |
| .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,193 // or %r8d,%ecx |
| .byte 102,15,196,201,0 // pinsrw $0x0,%ecx,%xmm1 |
| .byte 102,15,196,203,1 // pinsrw $0x1,%ebx,%xmm1 |
| .byte 102,65,15,96,201 // punpcklbw %xmm9,%xmm1 |
| .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,202 // mulps %xmm10,%xmm1 |
| .byte 72,139,64,16 // mov 0x10(%rax),%rax |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2 |
| .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| .byte 65,137,200 // mov %ecx,%r8d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,15,112,210,78 // pshufd $0x4e,%xmm2,%xmm2 |
| .byte 102,72,15,126,211 // movq %xmm2,%rbx |
| .byte 65,137,217 // mov %ebx,%r9d |
| .byte 72,193,235,32 // shr $0x20,%rbx |
| .byte 70,15,182,12,8 // movzbl (%rax,%r9,1),%r9d |
| .byte 15,182,28,24 // movzbl (%rax,%rbx,1),%ebx |
| .byte 193,227,8 // shl $0x8,%ebx |
| .byte 68,9,203 // or %r9d,%ebx |
| .byte 70,15,182,4,0 // movzbl (%rax,%r8,1),%r8d |
| .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax |
| .byte 193,224,8 // shl $0x8,%eax |
| .byte 68,9,192 // or %r8d,%eax |
| .byte 102,15,196,208,0 // pinsrw $0x0,%eax,%xmm2 |
| .byte 102,15,196,211,1 // pinsrw $0x1,%ebx,%xmm2 |
| .byte 102,65,15,96,209 // punpcklbw %xmm9,%xmm2 |
| .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,210 // mulps %xmm10,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 91 // pop %rbx |
| .byte 65,94 // pop %r14 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_r_sse2 |
| .globl _sk_table_r_sse2 |
| FUNCTION(_sk_table_r_sse2) |
| _sk_table_r_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,65,15,112,192,78 // pshufd $0x4e,%xmm8,%xmm0 |
| .byte 102,72,15,126,192 // movq %xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,76,15,126,193 // movq %xmm8,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8 |
| .byte 243,65,15,16,4,129 // movss (%r9,%rax,4),%xmm0 |
| .byte 68,15,20,192 // unpcklps %xmm0,%xmm8 |
| .byte 243,67,15,16,4,145 // movss (%r9,%r10,4),%xmm0 |
| .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9 |
| .byte 65,15,20,193 // unpcklps %xmm9,%xmm0 |
| .byte 65,15,20,192 // unpcklps %xmm8,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_g_sse2 |
| .globl _sk_table_g_sse2 |
| FUNCTION(_sk_table_g_sse2) |
| _sk_table_g_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,65,15,112,200,78 // pshufd $0x4e,%xmm8,%xmm1 |
| .byte 102,72,15,126,200 // movq %xmm1,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,76,15,126,193 // movq %xmm8,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8 |
| .byte 243,65,15,16,12,129 // movss (%r9,%rax,4),%xmm1 |
| .byte 68,15,20,193 // unpcklps %xmm1,%xmm8 |
| .byte 243,67,15,16,12,145 // movss (%r9,%r10,4),%xmm1 |
| .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9 |
| .byte 65,15,20,201 // unpcklps %xmm9,%xmm1 |
| .byte 65,15,20,200 // unpcklps %xmm8,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_b_sse2 |
| .globl _sk_table_b_sse2 |
| FUNCTION(_sk_table_b_sse2) |
| _sk_table_b_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,65,15,112,208,78 // pshufd $0x4e,%xmm8,%xmm2 |
| .byte 102,72,15,126,208 // movq %xmm2,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,76,15,126,193 // movq %xmm8,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8 |
| .byte 243,65,15,16,20,129 // movss (%r9,%rax,4),%xmm2 |
| .byte 68,15,20,194 // unpcklps %xmm2,%xmm8 |
| .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9 |
| .byte 65,15,20,209 // unpcklps %xmm9,%xmm2 |
| .byte 65,15,20,208 // unpcklps %xmm8,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_table_a_sse2 |
| .globl _sk_table_a_sse2 |
| FUNCTION(_sk_table_a_sse2) |
| _sk_table_a_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 139,64,8 // mov 0x8(%rax),%eax |
| .byte 255,200 // dec %eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8 |
| .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,65,15,112,216,78 // pshufd $0x4e,%xmm8,%xmm3 |
| .byte 102,72,15,126,216 // movq %xmm3,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,76,15,126,193 // movq %xmm8,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8 |
| .byte 243,65,15,16,28,129 // movss (%r9,%rax,4),%xmm3 |
| .byte 68,15,20,195 // unpcklps %xmm3,%xmm8 |
| .byte 243,67,15,16,28,145 // movss (%r9,%r10,4),%xmm3 |
| .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9 |
| .byte 65,15,20,217 // unpcklps %xmm9,%xmm3 |
| .byte 65,15,20,216 // unpcklps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_r_sse2 |
| .globl _sk_parametric_r_sse2 |
| FUNCTION(_sk_parametric_r_sse2) |
| _sk_parametric_r_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,216 // mulps %xmm0,%xmm11 |
| .byte 65,15,194,193,2 // cmpleps %xmm9,%xmm0 |
| .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,212 // mulps %xmm12,%xmm10 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 |
| .byte 102,69,15,219,235 // pand %xmm11,%xmm13 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 |
| .byte 102,69,15,235,229 // por %xmm13,%xmm12 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,220 // mulps %xmm12,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,249 // movd %ecx,%xmm15 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 69,15,94,238 // divps %xmm14,%xmm13 |
| .byte 69,15,92,213 // subps %xmm13,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 |
| .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 |
| .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 69,15,88,250 // addps %xmm10,%xmm15 |
| .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,84,211 // andps %xmm11,%xmm10 |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 69,15,92,226 // subps %xmm10,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 102,69,15,110,208 // movd %r8d,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,89,229 // mulps %xmm13,%xmm12 |
| .byte 69,15,92,252 // subps %xmm12,%xmm15 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 69,15,94,230 // divps %xmm14,%xmm12 |
| .byte 69,15,88,231 // addps %xmm15,%xmm12 |
| .byte 69,15,89,226 // mulps %xmm10,%xmm12 |
| .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 |
| .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,226 // addps %xmm10,%xmm12 |
| .byte 68,15,84,192 // andps %xmm0,%xmm8 |
| .byte 65,15,85,196 // andnps %xmm12,%xmm0 |
| .byte 65,15,86,192 // orps %xmm8,%xmm0 |
| .byte 65,15,95,193 // maxps %xmm9,%xmm0 |
| .byte 65,15,93,195 // minps %xmm11,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_g_sse2 |
| .globl _sk_parametric_g_sse2 |
| FUNCTION(_sk_parametric_g_sse2) |
| _sk_parametric_g_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 65,15,194,201,2 // cmpleps %xmm9,%xmm1 |
| .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,212 // mulps %xmm12,%xmm10 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 |
| .byte 102,69,15,219,235 // pand %xmm11,%xmm13 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 |
| .byte 102,69,15,235,229 // por %xmm13,%xmm12 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,220 // mulps %xmm12,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,249 // movd %ecx,%xmm15 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 69,15,94,238 // divps %xmm14,%xmm13 |
| .byte 69,15,92,213 // subps %xmm13,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 |
| .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 |
| .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 69,15,88,250 // addps %xmm10,%xmm15 |
| .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,84,211 // andps %xmm11,%xmm10 |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 69,15,92,226 // subps %xmm10,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 102,69,15,110,208 // movd %r8d,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,89,229 // mulps %xmm13,%xmm12 |
| .byte 69,15,92,252 // subps %xmm12,%xmm15 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 69,15,94,230 // divps %xmm14,%xmm12 |
| .byte 69,15,88,231 // addps %xmm15,%xmm12 |
| .byte 69,15,89,226 // mulps %xmm10,%xmm12 |
| .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 |
| .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,226 // addps %xmm10,%xmm12 |
| .byte 68,15,84,193 // andps %xmm1,%xmm8 |
| .byte 65,15,85,204 // andnps %xmm12,%xmm1 |
| .byte 65,15,86,200 // orps %xmm8,%xmm1 |
| .byte 65,15,95,201 // maxps %xmm9,%xmm1 |
| .byte 65,15,93,203 // minps %xmm11,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_b_sse2 |
| .globl _sk_parametric_b_sse2 |
| FUNCTION(_sk_parametric_b_sse2) |
| _sk_parametric_b_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 65,15,194,209,2 // cmpleps %xmm9,%xmm2 |
| .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,212 // mulps %xmm12,%xmm10 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 |
| .byte 102,69,15,219,235 // pand %xmm11,%xmm13 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 |
| .byte 102,69,15,235,229 // por %xmm13,%xmm12 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,220 // mulps %xmm12,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,249 // movd %ecx,%xmm15 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 69,15,94,238 // divps %xmm14,%xmm13 |
| .byte 69,15,92,213 // subps %xmm13,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 |
| .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 |
| .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 69,15,88,250 // addps %xmm10,%xmm15 |
| .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,84,211 // andps %xmm11,%xmm10 |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 69,15,92,226 // subps %xmm10,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 102,69,15,110,208 // movd %r8d,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,89,229 // mulps %xmm13,%xmm12 |
| .byte 69,15,92,252 // subps %xmm12,%xmm15 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 69,15,94,230 // divps %xmm14,%xmm12 |
| .byte 69,15,88,231 // addps %xmm15,%xmm12 |
| .byte 69,15,89,226 // mulps %xmm10,%xmm12 |
| .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 |
| .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,226 // addps %xmm10,%xmm12 |
| .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| .byte 65,15,85,212 // andnps %xmm12,%xmm2 |
| .byte 65,15,86,208 // orps %xmm8,%xmm2 |
| .byte 65,15,95,209 // maxps %xmm9,%xmm2 |
| .byte 65,15,93,211 // minps %xmm11,%xmm2 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_parametric_a_sse2 |
| .globl _sk_parametric_a_sse2 |
| FUNCTION(_sk_parametric_a_sse2) |
| _sk_parametric_a_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,219 // mulps %xmm3,%xmm11 |
| .byte 65,15,194,217,2 // cmpleps %xmm9,%xmm3 |
| .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12 |
| .byte 185,0,0,0,52 // mov $0x34000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,212 // mulps %xmm12,%xmm10 |
| .byte 185,255,255,127,0 // mov $0x7fffff,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 102,69,15,112,236,0 // pshufd $0x0,%xmm12,%xmm13 |
| .byte 102,69,15,219,235 // pand %xmm11,%xmm13 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,227,0 // pshufd $0x0,%xmm11,%xmm12 |
| .byte 102,69,15,235,229 // por %xmm13,%xmm12 |
| .byte 185,119,115,248,66 // mov $0x42f87377,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,117,191,191,63 // mov $0x3fbfbf75,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,220 // mulps %xmm12,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 185,163,233,220,63 // mov $0x3fdce9a3,%ecx |
| .byte 102,68,15,110,233 // movd %ecx,%xmm13 |
| .byte 185,249,68,180,62 // mov $0x3eb444f9,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 65,184,0,0,0,75 // mov $0x4b000000,%r8d |
| .byte 185,81,140,242,66 // mov $0x42f28c51,%ecx |
| .byte 102,68,15,110,249 // movd %ecx,%xmm15 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| .byte 69,15,94,238 // divps %xmm14,%xmm13 |
| .byte 69,15,92,213 // subps %xmm13,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 243,69,15,91,202 // cvttps2dq %xmm10,%xmm9 |
| .byte 69,15,91,225 // cvtdq2ps %xmm9,%xmm12 |
| .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 69,15,88,250 // addps %xmm10,%xmm15 |
| .byte 69,15,194,212,1 // cmpltps %xmm12,%xmm10 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,84,211 // andps %xmm11,%xmm10 |
| .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| .byte 69,15,92,226 // subps %xmm10,%xmm12 |
| .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| .byte 102,69,15,110,208 // movd %r8d,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,141,188,190,63 // mov $0x3fbebc8d,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,89,229 // mulps %xmm13,%xmm12 |
| .byte 69,15,92,252 // subps %xmm12,%xmm15 |
| .byte 185,254,210,221,65 // mov $0x41ddd2fe,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 185,248,245,154,64 // mov $0x409af5f8,%ecx |
| .byte 102,68,15,110,241 // movd %ecx,%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| .byte 69,15,94,230 // divps %xmm14,%xmm12 |
| .byte 69,15,88,231 // addps %xmm15,%xmm12 |
| .byte 69,15,89,226 // mulps %xmm10,%xmm12 |
| .byte 102,69,15,91,212 // cvtps2dq %xmm12,%xmm10 |
| .byte 243,68,15,16,96,20 // movss 0x14(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,88,226 // addps %xmm10,%xmm12 |
| .byte 68,15,84,195 // andps %xmm3,%xmm8 |
| .byte 65,15,85,220 // andnps %xmm12,%xmm3 |
| .byte 65,15,86,216 // orps %xmm8,%xmm3 |
| .byte 65,15,95,217 // maxps %xmm9,%xmm3 |
| .byte 65,15,93,219 // minps %xmm11,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_a8_sse2 |
| .globl _sk_load_a8_sse2 |
| FUNCTION(_sk_load_a8_sse2) |
| _sk_load_a8_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 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 102,15,239,201 // pxor %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_a8_sse2 |
| .globl _sk_gather_a8_sse2 |
| FUNCTION(_sk_gather_a8_sse2) |
| _sk_gather_a8_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,72,15,126,192 // movq %xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,209 // or %r10d,%ecx |
| .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d |
| .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax |
| .byte 193,224,8 // shl $0x8,%eax |
| .byte 68,9,192 // or %r8d,%eax |
| .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0 |
| .byte 102,15,196,193,1 // pinsrw $0x1,%ecx,%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 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 102,15,239,201 // pxor %xmm1,%xmm1 |
| .byte 102,15,239,210 // pxor %xmm2,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_a8_sse2 |
| .globl _sk_store_a8_sse2 |
| FUNCTION(_sk_store_a8_sse2) |
| _sk_store_a8_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%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 102,69,15,103,192 // packuswb %xmm8,%xmm8 |
| .byte 102,68,15,126,4,56 // movd %xmm8,(%rax,%rdi,1) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_g8_sse2 |
| .globl _sk_load_g8_sse2 |
| FUNCTION(_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_gather_g8_sse2 |
| .globl _sk_gather_g8_sse2 |
| FUNCTION(_sk_gather_g8_sse2) |
| _sk_gather_g8_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,72,15,126,192 // movq %xmm0,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d |
| .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,209 // or %r10d,%ecx |
| .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d |
| .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax |
| .byte 193,224,8 // shl $0x8,%eax |
| .byte 68,9,192 // or %r8d,%eax |
| .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0 |
| .byte 102,15,196,193,1 // pinsrw $0x1,%ecx,%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_gather_i8_sse2 |
| .globl _sk_gather_i8_sse2 |
| FUNCTION(_sk_gather_i8_sse2) |
| _sk_gather_i8_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 73,137,192 // mov %rax,%r8 |
| .byte 77,133,192 // test %r8,%r8 |
| .byte 116,5 // je 28b8 <_sk_gather_i8_sse2+0xf> |
| .byte 76,137,192 // mov %r8,%rax |
| .byte 235,2 // jmp 28ba <_sk_gather_i8_sse2+0x11> |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,16 // mov (%rax),%r10 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,72,15,126,192 // movq %xmm0,%rax |
| .byte 65,137,193 // mov %eax,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0 |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,203 // mov %ecx,%r11d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 71,15,182,28,26 // movzbl (%r10,%r11,1),%r11d |
| .byte 65,15,182,12,10 // movzbl (%r10,%rcx,1),%ecx |
| .byte 193,225,8 // shl $0x8,%ecx |
| .byte 68,9,217 // or %r11d,%ecx |
| .byte 71,15,182,12,10 // movzbl (%r10,%r9,1),%r9d |
| .byte 65,15,182,4,2 // movzbl (%r10,%rax,1),%eax |
| .byte 193,224,8 // shl $0x8,%eax |
| .byte 68,9,200 // or %r9d,%eax |
| .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0 |
| .byte 102,15,196,193,1 // pinsrw $0x1,%ecx,%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 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 |
| .byte 102,72,15,126,200 // movq %xmm1,%rax |
| .byte 68,15,182,200 // movzbl %al,%r9d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 77,139,64,8 // mov 0x8(%r8),%r8 |
| .byte 68,15,182,209 // movzbl %cl,%r10d |
| .byte 72,193,233,30 // shr $0x1e,%rcx |
| .byte 102,65,15,110,4,8 // movd (%r8,%rcx,1),%xmm0 |
| .byte 102,65,15,110,12,128 // movd (%r8,%rax,4),%xmm1 |
| .byte 102,15,98,193 // punpckldq %xmm1,%xmm0 |
| .byte 102,67,15,110,28,144 // movd (%r8,%r10,4),%xmm3 |
| .byte 102,67,15,110,12,136 // movd (%r8,%r9,4),%xmm1 |
| .byte 102,15,98,217 // punpckldq %xmm1,%xmm3 |
| .byte 102,15,98,216 // punpckldq %xmm0,%xmm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_565_sse2 |
| .globl _sk_load_565_sse2 |
| FUNCTION(_sk_load_565_sse2) |
| _sk_load_565_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,126,20,120 // movq (%rax,%rdi,2),%xmm2 |
| .byte 102,15,239,192 // pxor %xmm0,%xmm0 |
| .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,219,194 // pand %xmm2,%xmm0 |
| .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| .byte 102,15,219,202 // pand %xmm2,%xmm1 |
| .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,15,219,218 // pand %xmm2,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_565_sse2 |
| .globl _sk_gather_565_sse2 |
| FUNCTION(_sk_gather_565_sse2) |
| _sk_gather_565_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 |
| .byte 102,72,15,126,200 // movq %xmm1,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,67,15,196,20,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm2 |
| .byte 102,65,15,196,20,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm2 |
| .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx |
| .byte 102,15,196,209,2 // pinsrw $0x2,%ecx,%xmm2 |
| .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax |
| .byte 102,15,196,208,3 // pinsrw $0x3,%eax,%xmm2 |
| .byte 102,15,239,192 // pxor %xmm0,%xmm0 |
| .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 |
| .byte 184,0,248,0,0 // mov $0xf800,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,219,194 // pand %xmm2,%xmm0 |
| .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| .byte 184,8,33,132,55 // mov $0x37842108,%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 184,224,7,0,0 // mov $0x7e0,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| .byte 102,15,219,202 // pand %xmm2,%xmm1 |
| .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3 |
| .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| .byte 102,15,110,200 // movd %eax,%xmm1 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 15,89,203 // mulps %xmm3,%xmm1 |
| .byte 184,31,0,0,0 // mov $0x1f,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| .byte 102,15,219,218 // pand %xmm2,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 15,89,211 // mulps %xmm3,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_565_sse2 |
| .globl _sk_store_565_sse2 |
| FUNCTION(_sk_store_565_sse2) |
| _sk_store_565_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,248,65 // mov $0x41f80000,%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 102,65,15,114,241,11 // pslld $0xb,%xmm9 |
| .byte 185,0,0,124,66 // mov $0x427c0000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10 |
| .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%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_4444_sse2 |
| .globl _sk_load_4444_sse2 |
| FUNCTION(_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_gather_4444_sse2 |
| .globl _sk_gather_4444_sse2 |
| FUNCTION(_sk_gather_4444_sse2) |
| _sk_gather_4444_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 |
| .byte 102,72,15,126,200 // movq %xmm1,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,71,15,196,12,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm9 |
| .byte 102,69,15,196,12,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm9 |
| .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx |
| .byte 102,68,15,196,201,2 // pinsrw $0x2,%ecx,%xmm9 |
| .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax |
| .byte 102,68,15,196,200,3 // pinsrw $0x3,%eax,%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 |
| FUNCTION(_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 |
| FUNCTION(_sk_load_8888_sse2) |
| _sk_load_8888_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,111,28,184 // movdqu (%rax,%rdi,4),%xmm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_8888_sse2 |
| .globl _sk_gather_8888_sse2 |
| FUNCTION(_sk_gather_8888_sse2) |
| _sk_gather_8888_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 |
| .byte 102,72,15,126,200 // movq %xmm1,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 102,65,15,110,4,137 // movd (%r9,%rcx,4),%xmm0 |
| .byte 102,65,15,110,12,129 // movd (%r9,%rax,4),%xmm1 |
| .byte 102,15,98,193 // punpckldq %xmm1,%xmm0 |
| .byte 102,67,15,110,28,145 // movd (%r9,%r10,4),%xmm3 |
| .byte 102,67,15,110,12,129 // movd (%r9,%r8,4),%xmm1 |
| .byte 102,15,98,217 // punpckldq %xmm1,%xmm3 |
| .byte 102,15,98,216 // punpckldq %xmm0,%xmm3 |
| .byte 184,255,0,0,0 // mov $0xff,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %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_8888_sse2 |
| .globl _sk_store_8888_sse2 |
| FUNCTION(_sk_store_8888_sse2) |
| _sk_store_8888_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,127,67 // mov $0x437f0000,%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 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 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| .byte 102,65,15,114,240,24 // pslld $0x18,%xmm8 |
| .byte 102,69,15,235,193 // por %xmm9,%xmm8 |
| .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| .byte 243,68,15,127,4,184 // movdqu %xmm8,(%rax,%rdi,4) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_f16_sse2 |
| .globl _sk_load_f16_sse2 |
| FUNCTION(_sk_load_f16_sse2) |
| _sk_load_f16_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 |
| .byte 102,68,15,97,193 // punpcklwd %xmm1,%xmm8 |
| .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 |
| .byte 102,68,15,97,224 // punpcklwd %xmm0,%xmm12 |
| .byte 102,68,15,105,192 // punpckhwd %xmm0,%xmm8 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 |
| .byte 102,69,15,97,233 // punpcklwd %xmm9,%xmm13 |
| .byte 184,0,128,0,0 // mov $0x8000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,68,15,112,208,0 // pshufd $0x0,%xmm0,%xmm10 |
| .byte 102,65,15,111,205 // movdqa %xmm13,%xmm1 |
| .byte 102,65,15,219,202 // pand %xmm10,%xmm1 |
| .byte 184,0,124,0,0 // mov $0x7c00,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 |
| .byte 102,65,15,111,197 // movdqa %xmm13,%xmm0 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 102,68,15,239,233 // pxor %xmm1,%xmm13 |
| .byte 102,15,114,241,16 // pslld $0x10,%xmm1 |
| .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 |
| .byte 184,0,0,0,56 // mov $0x38000000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 |
| .byte 102,65,15,254,203 // paddd %xmm11,%xmm1 |
| .byte 102,65,15,254,205 // paddd %xmm13,%xmm1 |
| .byte 102,65,15,118,193 // pcmpeqd %xmm9,%xmm0 |
| .byte 102,15,223,193 // pandn %xmm1,%xmm0 |
| .byte 102,65,15,115,220,8 // psrldq $0x8,%xmm12 |
| .byte 102,69,15,97,225 // punpcklwd %xmm9,%xmm12 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,65,15,219,210 // pand %xmm10,%xmm2 |
| .byte 102,65,15,111,204 // movdqa %xmm12,%xmm1 |
| .byte 102,15,219,203 // pand %xmm3,%xmm1 |
| .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 |
| .byte 102,15,114,242,16 // pslld $0x10,%xmm2 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 |
| .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 |
| .byte 102,65,15,118,201 // pcmpeqd %xmm9,%xmm1 |
| .byte 102,15,223,202 // pandn %xmm2,%xmm1 |
| .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 |
| .byte 102,69,15,97,225 // punpcklwd %xmm9,%xmm12 |
| .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 |
| .byte 102,69,15,219,234 // pand %xmm10,%xmm13 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,15,219,211 // pand %xmm3,%xmm2 |
| .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 |
| .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 |
| .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 |
| .byte 102,65,15,118,209 // pcmpeqd %xmm9,%xmm2 |
| .byte 102,65,15,223,213 // pandn %xmm13,%xmm2 |
| .byte 102,65,15,115,216,8 // psrldq $0x8,%xmm8 |
| .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 |
| .byte 102,69,15,219,208 // pand %xmm8,%xmm10 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 102,69,15,239,194 // pxor %xmm10,%xmm8 |
| .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10 |
| .byte 102,65,15,114,240,13 // pslld $0xd,%xmm8 |
| .byte 102,69,15,254,211 // paddd %xmm11,%xmm10 |
| .byte 102,69,15,254,208 // paddd %xmm8,%xmm10 |
| .byte 102,65,15,118,217 // pcmpeqd %xmm9,%xmm3 |
| .byte 102,65,15,223,218 // pandn %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_gather_f16_sse2 |
| .globl _sk_gather_f16_sse2 |
| FUNCTION(_sk_gather_f16_sse2) |
| _sk_gather_f16_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,139,8 // mov (%rax),%r9 |
| .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1 |
| .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2 |
| .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3 |
| .byte 102,15,244,218 // pmuludq %xmm2,%xmm3 |
| .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3 |
| .byte 102,15,244,209 // pmuludq %xmm1,%xmm2 |
| .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1 |
| .byte 102,15,98,203 // punpckldq %xmm3,%xmm1 |
| .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0 |
| .byte 102,15,254,193 // paddd %xmm1,%xmm0 |
| .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1 |
| .byte 102,72,15,126,200 // movq %xmm1,%rax |
| .byte 65,137,192 // mov %eax,%r8d |
| .byte 72,193,232,32 // shr $0x20,%rax |
| .byte 102,72,15,126,193 // movq %xmm0,%rcx |
| .byte 65,137,202 // mov %ecx,%r10d |
| .byte 72,193,233,32 // shr $0x20,%rcx |
| .byte 243,65,15,126,4,201 // movq (%r9,%rcx,8),%xmm0 |
| .byte 243,67,15,126,12,209 // movq (%r9,%r10,8),%xmm1 |
| .byte 102,15,108,200 // punpcklqdq %xmm0,%xmm1 |
| .byte 243,65,15,126,4,193 // movq (%r9,%rax,8),%xmm0 |
| .byte 243,67,15,126,20,193 // movq (%r9,%r8,8),%xmm2 |
| .byte 102,15,108,208 // punpcklqdq %xmm0,%xmm2 |
| .byte 102,68,15,111,193 // movdqa %xmm1,%xmm8 |
| .byte 102,68,15,97,194 // punpcklwd %xmm2,%xmm8 |
| .byte 102,15,105,202 // punpckhwd %xmm2,%xmm1 |
| .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 |
| .byte 102,68,15,97,225 // punpcklwd %xmm1,%xmm12 |
| .byte 102,68,15,105,193 // punpckhwd %xmm1,%xmm8 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 |
| .byte 102,69,15,97,233 // punpcklwd %xmm9,%xmm13 |
| .byte 184,0,128,0,0 // mov $0x8000,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,68,15,112,208,0 // pshufd $0x0,%xmm0,%xmm10 |
| .byte 102,65,15,111,205 // movdqa %xmm13,%xmm1 |
| .byte 102,65,15,219,202 // pand %xmm10,%xmm1 |
| .byte 184,0,124,0,0 // mov $0x7c00,%eax |
| .byte 102,15,110,192 // movd %eax,%xmm0 |
| .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 |
| .byte 102,65,15,111,197 // movdqa %xmm13,%xmm0 |
| .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| .byte 102,68,15,239,233 // pxor %xmm1,%xmm13 |
| .byte 102,15,114,241,16 // pslld $0x10,%xmm1 |
| .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13 |
| .byte 184,0,0,0,56 // mov $0x38000000,%eax |
| .byte 102,15,110,208 // movd %eax,%xmm2 |
| .byte 102,68,15,112,218,0 // pshufd $0x0,%xmm2,%xmm11 |
| .byte 102,65,15,254,203 // paddd %xmm11,%xmm1 |
| .byte 102,65,15,254,205 // paddd %xmm13,%xmm1 |
| .byte 102,65,15,118,193 // pcmpeqd %xmm9,%xmm0 |
| .byte 102,15,223,193 // pandn %xmm1,%xmm0 |
| .byte 102,65,15,115,220,8 // psrldq $0x8,%xmm12 |
| .byte 102,69,15,97,225 // punpcklwd %xmm9,%xmm12 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,65,15,219,210 // pand %xmm10,%xmm2 |
| .byte 102,65,15,111,204 // movdqa %xmm12,%xmm1 |
| .byte 102,15,219,203 // pand %xmm3,%xmm1 |
| .byte 102,68,15,239,226 // pxor %xmm2,%xmm12 |
| .byte 102,15,114,242,16 // pslld $0x10,%xmm2 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,65,15,254,211 // paddd %xmm11,%xmm2 |
| .byte 102,65,15,254,212 // paddd %xmm12,%xmm2 |
| .byte 102,65,15,118,201 // pcmpeqd %xmm9,%xmm1 |
| .byte 102,15,223,202 // pandn %xmm2,%xmm1 |
| .byte 102,69,15,111,224 // movdqa %xmm8,%xmm12 |
| .byte 102,69,15,97,225 // punpcklwd %xmm9,%xmm12 |
| .byte 102,69,15,111,236 // movdqa %xmm12,%xmm13 |
| .byte 102,69,15,219,234 // pand %xmm10,%xmm13 |
| .byte 102,65,15,111,212 // movdqa %xmm12,%xmm2 |
| .byte 102,15,219,211 // pand %xmm3,%xmm2 |
| .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 |
| .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 |
| .byte 102,65,15,114,244,13 // pslld $0xd,%xmm12 |
| .byte 102,69,15,254,235 // paddd %xmm11,%xmm13 |
| .byte 102,69,15,254,236 // paddd %xmm12,%xmm13 |
| .byte 102,65,15,118,209 // pcmpeqd %xmm9,%xmm2 |
| .byte 102,65,15,223,213 // pandn %xmm13,%xmm2 |
| .byte 102,65,15,115,216,8 // psrldq $0x8,%xmm8 |
| .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 |
| .byte 102,69,15,219,208 // pand %xmm8,%xmm10 |
| .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| .byte 102,69,15,239,194 // pxor %xmm10,%xmm8 |
| .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10 |
| .byte 102,65,15,114,240,13 // pslld $0xd,%xmm8 |
| .byte 102,69,15,254,211 // paddd %xmm11,%xmm10 |
| .byte 102,69,15,254,208 // paddd %xmm8,%xmm10 |
| .byte 102,65,15,118,217 // pcmpeqd %xmm9,%xmm3 |
| .byte 102,65,15,223,218 // pandn %xmm10,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_f16_sse2 |
| .globl _sk_store_f16_sse2 |
| FUNCTION(_sk_store_f16_sse2) |
| _sk_store_f16_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,0,0,128 // mov $0x80000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9 |
| .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| .byte 102,68,15,219,224 // pand %xmm0,%xmm12 |
| .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 |
| .byte 102,69,15,239,196 // pxor %xmm12,%xmm8 |
| .byte 185,0,0,128,56 // mov $0x38800000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 102,65,15,114,212,16 // psrld $0x10,%xmm12 |
| .byte 102,69,15,111,232 // movdqa %xmm8,%xmm13 |
| .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13 |
| .byte 185,0,192,1,0 // mov $0x1c000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 102,69,15,112,219,0 // pshufd $0x0,%xmm11,%xmm11 |
| .byte 102,69,15,250,227 // psubd %xmm11,%xmm12 |
| .byte 102,69,15,254,229 // paddd %xmm13,%xmm12 |
| .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12 |
| .byte 102,65,15,114,228,16 // psrad $0x10,%xmm12 |
| .byte 69,15,194,194,5 // cmpnltps %xmm10,%xmm8 |
| .byte 69,15,84,196 // andps %xmm12,%xmm8 |
| .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 |
| .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13 |
| .byte 102,68,15,219,233 // pand %xmm1,%xmm13 |
| .byte 102,68,15,111,225 // movdqa %xmm1,%xmm12 |
| .byte 102,69,15,239,229 // pxor %xmm13,%xmm12 |
| .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13 |
| .byte 102,69,15,111,244 // movdqa %xmm12,%xmm14 |
| .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14 |
| .byte 102,69,15,250,235 // psubd %xmm11,%xmm13 |
| .byte 102,69,15,254,238 // paddd %xmm14,%xmm13 |
| .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13 |
| .byte 102,65,15,114,229,16 // psrad $0x10,%xmm13 |
| .byte 69,15,194,226,5 // cmpnltps %xmm10,%xmm12 |
| .byte 69,15,84,229 // andps %xmm13,%xmm12 |
| .byte 102,69,15,107,228 // packssdw %xmm12,%xmm12 |
| .byte 102,69,15,111,241 // movdqa %xmm9,%xmm14 |
| .byte 102,68,15,219,242 // pand %xmm2,%xmm14 |
| .byte 102,68,15,111,234 // movdqa %xmm2,%xmm13 |
| .byte 102,69,15,239,238 // pxor %xmm14,%xmm13 |
| .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14 |
| .byte 102,69,15,111,253 // movdqa %xmm13,%xmm15 |
| .byte 102,65,15,114,215,13 // psrld $0xd,%xmm15 |
| .byte 102,69,15,250,243 // psubd %xmm11,%xmm14 |
| .byte 102,69,15,254,247 // paddd %xmm15,%xmm14 |
| .byte 102,65,15,114,246,16 // pslld $0x10,%xmm14 |
| .byte 102,65,15,114,230,16 // psrad $0x10,%xmm14 |
| .byte 69,15,194,234,5 // cmpnltps %xmm10,%xmm13 |
| .byte 69,15,84,238 // andps %xmm14,%xmm13 |
| .byte 102,69,15,107,237 // packssdw %xmm13,%xmm13 |
| .byte 102,68,15,219,203 // pand %xmm3,%xmm9 |
| .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14 |
| .byte 102,69,15,239,241 // pxor %xmm9,%xmm14 |
| .byte 102,65,15,114,209,16 // psrld $0x10,%xmm9 |
| .byte 102,69,15,111,254 // movdqa %xmm14,%xmm15 |
| .byte 102,65,15,114,215,13 // psrld $0xd,%xmm15 |
| .byte 102,69,15,250,203 // psubd %xmm11,%xmm9 |
| .byte 102,69,15,254,207 // paddd %xmm15,%xmm9 |
| .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| .byte 102,65,15,114,225,16 // psrad $0x10,%xmm9 |
| .byte 69,15,194,242,5 // cmpnltps %xmm10,%xmm14 |
| .byte 69,15,84,241 // andps %xmm9,%xmm14 |
| .byte 102,69,15,107,246 // packssdw %xmm14,%xmm14 |
| .byte 102,69,15,97,196 // punpcklwd %xmm12,%xmm8 |
| .byte 102,69,15,97,238 // punpcklwd %xmm14,%xmm13 |
| .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| .byte 102,69,15,98,205 // punpckldq %xmm13,%xmm9 |
| .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| .byte 102,69,15,106,197 // punpckhdq %xmm13,%xmm8 |
| .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_u16_be_sse2 |
| .globl _sk_load_u16_be_sse2 |
| FUNCTION(_sk_load_u16_be_sse2) |
| _sk_load_u16_be_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| .byte 102,15,111,208 // movdqa %xmm0,%xmm2 |
| .byte 102,15,97,209 // punpcklwd %xmm1,%xmm2 |
| .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,15,105,208 // punpckhwd %xmm0,%xmm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 102,15,111,193 // movdqa %xmm1,%xmm0 |
| .byte 102,15,113,240,8 // psllw $0x8,%xmm0 |
| .byte 102,15,112,217,78 // pshufd $0x4e,%xmm1,%xmm3 |
| .byte 102,15,113,209,8 // psrlw $0x8,%xmm1 |
| .byte 102,15,235,200 // por %xmm0,%xmm1 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 |
| .byte 15,91,193 // cvtdq2ps %xmm1,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,15,235,217 // por %xmm1,%xmm3 |
| .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| .byte 15,91,203 // cvtdq2ps %xmm3,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,68,15,111,210 // movdqa %xmm2,%xmm10 |
| .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| .byte 102,15,112,218,78 // pshufd $0x4e,%xmm2,%xmm3 |
| .byte 102,15,113,210,8 // psrlw $0x8,%xmm2 |
| .byte 102,65,15,235,210 // por %xmm10,%xmm2 |
| .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 102,68,15,111,211 // movdqa %xmm3,%xmm10 |
| .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,65,15,235,218 // por %xmm10,%xmm3 |
| .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_rgb_u16_be_sse2 |
| .globl _sk_load_rgb_u16_be_sse2 |
| FUNCTION(_sk_load_rgb_u16_be_sse2) |
| _sk_load_rgb_u16_be_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 72,141,12,127 // lea (%rdi,%rdi,2),%rcx |
| .byte 243,15,111,20,72 // movdqu (%rax,%rcx,2),%xmm2 |
| .byte 243,15,111,68,72,8 // movdqu 0x8(%rax,%rcx,2),%xmm0 |
| .byte 102,15,115,216,4 // psrldq $0x4,%xmm0 |
| .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| .byte 102,15,115,217,6 // psrldq $0x6,%xmm1 |
| .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 |
| .byte 102,15,115,216,6 // psrldq $0x6,%xmm0 |
| .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| .byte 102,15,111,194 // movdqa %xmm2,%xmm0 |
| .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 |
| .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3 |
| .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2 |
| .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 102,15,111,200 // movdqa %xmm0,%xmm1 |
| .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| .byte 102,15,113,208,8 // psrlw $0x8,%xmm0 |
| .byte 102,15,235,193 // por %xmm1,%xmm0 |
| .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 |
| .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| .byte 102,15,235,217 // por %xmm1,%xmm3 |
| .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| .byte 15,91,203 // cvtdq2ps %xmm3,%xmm1 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 102,15,111,218 // movdqa %xmm2,%xmm3 |
| .byte 102,15,113,243,8 // psllw $0x8,%xmm3 |
| .byte 102,15,113,210,8 // psrlw $0x8,%xmm2 |
| .byte 102,15,235,211 // por %xmm3,%xmm2 |
| .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2 |
| .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .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 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_u16_be_sse2 |
| .globl _sk_store_u16_be_sse2 |
| FUNCTION(_sk_store_u16_be_sse2) |
| _sk_store_u16_be_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 185,0,255,127,71 // mov $0x477fff00,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 102,69,15,91,192 // cvtps2dq %xmm8,%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 102,69,15,111,208 // movdqa %xmm8,%xmm10 |
| .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8 |
| .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10 |
| .byte 102,65,15,114,226,16 // psrad $0x10,%xmm10 |
| .byte 102,69,15,107,210 // packssdw %xmm10,%xmm10 |
| .byte 102,69,15,111,218 // movdqa %xmm10,%xmm11 |
| .byte 102,65,15,113,243,8 // psllw $0x8,%xmm11 |
| .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 |
| .byte 102,69,15,235,211 // por %xmm11,%xmm10 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 102,69,15,91,219 // cvtps2dq %xmm11,%xmm11 |
| .byte 102,65,15,114,243,16 // pslld $0x10,%xmm11 |
| .byte 102,65,15,114,227,16 // psrad $0x10,%xmm11 |
| .byte 102,69,15,107,219 // packssdw %xmm11,%xmm11 |
| .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12 |
| .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 |
| .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| .byte 102,65,15,114,225,16 // psrad $0x10,%xmm9 |
| .byte 102,69,15,107,201 // packssdw %xmm9,%xmm9 |
| .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| .byte 102,65,15,113,209,8 // psrlw $0x8,%xmm9 |
| .byte 102,69,15,235,204 // por %xmm12,%xmm9 |
| .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 |
| .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11 |
| .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9 |
| .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8 |
| .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_load_f32_sse2 |
| .globl _sk_load_f32_sse2 |
| FUNCTION(_sk_load_f32_sse2) |
| _sk_load_f32_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 72,137,249 // mov %rdi,%rcx |
| .byte 72,193,225,4 // shl $0x4,%rcx |
| .byte 68,15,16,4,8 // movups (%rax,%rcx,1),%xmm8 |
| .byte 15,16,68,8,16 // movups 0x10(%rax,%rcx,1),%xmm0 |
| .byte 15,16,92,8,32 // movups 0x20(%rax,%rcx,1),%xmm3 |
| .byte 68,15,16,76,8,48 // movups 0x30(%rax,%rcx,1),%xmm9 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 15,20,208 // unpcklps %xmm0,%xmm2 |
| .byte 15,40,203 // movaps %xmm3,%xmm1 |
| .byte 65,15,20,201 // unpcklps %xmm9,%xmm1 |
| .byte 68,15,21,192 // unpckhps %xmm0,%xmm8 |
| .byte 65,15,21,217 // unpckhps %xmm9,%xmm3 |
| .byte 15,40,194 // movaps %xmm2,%xmm0 |
| .byte 102,15,20,193 // unpcklpd %xmm1,%xmm0 |
| .byte 15,18,202 // movhlps %xmm2,%xmm1 |
| .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| .byte 102,15,20,211 // unpcklpd %xmm3,%xmm2 |
| .byte 65,15,18,216 // movhlps %xmm8,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_store_f32_sse2 |
| .globl _sk_store_f32_sse2 |
| FUNCTION(_sk_store_f32_sse2) |
| _sk_store_f32_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,139,0 // mov (%rax),%rax |
| .byte 72,137,249 // mov %rdi,%rcx |
| .byte 72,193,225,4 // shl $0x4,%rcx |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 68,15,20,201 // unpcklps %xmm1,%xmm9 |
| .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| .byte 68,15,40,218 // movaps %xmm2,%xmm11 |
| .byte 68,15,20,219 // unpcklps %xmm3,%xmm11 |
| .byte 68,15,21,193 // unpckhps %xmm1,%xmm8 |
| .byte 68,15,21,211 // unpckhps %xmm3,%xmm10 |
| .byte 69,15,40,225 // movaps %xmm9,%xmm12 |
| .byte 102,69,15,20,227 // unpcklpd %xmm11,%xmm12 |
| .byte 69,15,18,217 // movhlps %xmm9,%xmm11 |
| .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| .byte 102,69,15,20,202 // unpcklpd %xmm10,%xmm9 |
| .byte 69,15,18,208 // movhlps %xmm8,%xmm10 |
| .byte 102,68,15,17,36,8 // movupd %xmm12,(%rax,%rcx,1) |
| .byte 68,15,17,92,8,16 // movups %xmm11,0x10(%rax,%rcx,1) |
| .byte 102,68,15,17,76,8,32 // movupd %xmm9,0x20(%rax,%rcx,1) |
| .byte 68,15,17,84,8,48 // movups %xmm10,0x30(%rax,%rcx,1) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_x_sse2 |
| .globl _sk_clamp_x_sse2 |
| FUNCTION(_sk_clamp_x_sse2) |
| _sk_clamp_x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,95,192 // maxps %xmm0,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 102,15,118,192 // pcmpeqd %xmm0,%xmm0 |
| .byte 102,65,15,254,193 // paddd %xmm9,%xmm0 |
| .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_clamp_y_sse2 |
| .globl _sk_clamp_y_sse2 |
| FUNCTION(_sk_clamp_y_sse2) |
| _sk_clamp_y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,95,193 // maxps %xmm1,%xmm8 |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 102,15,118,201 // pcmpeqd %xmm1,%xmm1 |
| .byte 102,65,15,254,201 // paddd %xmm9,%xmm1 |
| .byte 68,15,93,193 // minps %xmm1,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_x_sse2 |
| .globl _sk_repeat_x_sse2 |
| FUNCTION(_sk_repeat_x_sse2) |
| _sk_repeat_x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 |
| .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 |
| .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,84,217 // andps %xmm9,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 65,15,92,194 // subps %xmm10,%xmm0 |
| .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| .byte 65,15,93,193 // minps %xmm9,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_repeat_y_sse2 |
| .globl _sk_repeat_y_sse2 |
| FUNCTION(_sk_repeat_y_sse2) |
| _sk_repeat_y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 |
| .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 |
| .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,84,217 // andps %xmm9,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 65,15,92,202 // subps %xmm10,%xmm1 |
| .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| .byte 65,15,93,201 // minps %xmm9,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_x_sse2 |
| .globl _sk_mirror_x_sse2 |
| FUNCTION(_sk_mirror_x_sse2) |
| _sk_mirror_x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,92,192 // subps %xmm8,%xmm0 |
| .byte 243,69,15,88,201 // addss %xmm9,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,40,208 // movaps %xmm0,%xmm10 |
| .byte 69,15,94,209 // divps %xmm9,%xmm10 |
| .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 |
| .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 |
| .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,84,226 // andps %xmm10,%xmm12 |
| .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| .byte 69,15,92,220 // subps %xmm12,%xmm11 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 65,15,92,195 // subps %xmm11,%xmm0 |
| .byte 65,15,92,192 // subps %xmm8,%xmm0 |
| .byte 68,15,92,208 // subps %xmm0,%xmm10 |
| .byte 65,15,84,194 // andps %xmm10,%xmm0 |
| .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| .byte 65,15,93,193 // minps %xmm9,%xmm0 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_mirror_y_sse2 |
| .globl _sk_mirror_y_sse2 |
| FUNCTION(_sk_mirror_y_sse2) |
| _sk_mirror_y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 243,69,15,88,201 // addss %xmm9,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| .byte 69,15,94,209 // divps %xmm9,%xmm10 |
| .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 |
| .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 |
| .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 |
| .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,84,226 // andps %xmm10,%xmm12 |
| .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| .byte 69,15,92,220 // subps %xmm12,%xmm11 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 65,15,92,203 // subps %xmm11,%xmm1 |
| .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| .byte 68,15,92,209 // subps %xmm1,%xmm10 |
| .byte 65,15,84,202 // andps %xmm10,%xmm1 |
| .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| .byte 65,15,93,201 // minps %xmm9,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_luminance_to_alpha_sse2 |
| .globl _sk_luminance_to_alpha_sse2 |
| FUNCTION(_sk_luminance_to_alpha_sse2) |
| _sk_luminance_to_alpha_sse2: |
| .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 184,89,23,55,63 // mov $0x3f371759,%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 15,88,195 // addps %xmm3,%xmm0 |
| .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| .byte 102,15,110,216 // movd %eax,%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 15,89,218 // mulps %xmm2,%xmm3 |
| .byte 15,88,216 // addps %xmm0,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 15,87,192 // xorps %xmm0,%xmm0 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_2x3_sse2 |
| .globl _sk_matrix_2x3_sse2 |
| FUNCTION(_sk_matrix_2x3_sse2) |
| _sk_matrix_2x3_sse2: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,16 // movss 0x10(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_3x4_sse2 |
| .globl _sk_matrix_3x4_sse2 |
| FUNCTION(_sk_matrix_3x4_sse2) |
| _sk_matrix_3x4_sse2: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,36 // movss 0x24(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_4x5_sse2 |
| .globl _sk_matrix_4x5_sse2 |
| FUNCTION(_sk_matrix_4x5_sse2) |
| _sk_matrix_4x5_sse2: |
| .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,32 // movss 0x20(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,48 // movss 0x30(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,64 // movss 0x40(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,36 // movss 0x24(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,52 // movss 0x34(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,68 // movss 0x44(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,56 // movss 0x38(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 243,68,15,16,112,72 // movss 0x48(%rax),%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 243,68,15,16,88,12 // movss 0xc(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 243,68,15,16,112,60 // movss 0x3c(%rax),%xmm14 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 243,68,15,16,120,76 // movss 0x4c(%rax),%xmm15 |
| .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| .byte 69,15,88,247 // addps %xmm15,%xmm14 |
| .byte 68,15,89,234 // mulps %xmm2,%xmm13 |
| .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| .byte 69,15,89,216 // mulps %xmm8,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_matrix_perspective_sse2 |
| .globl _sk_matrix_perspective_sse2 |
| FUNCTION(_sk_matrix_perspective_sse2) |
| _sk_matrix_perspective_sse2: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 243,68,15,16,80,24 // movss 0x18(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 65,15,83,202 // rcpps %xmm10,%xmm1 |
| .byte 15,89,193 // mulps %xmm1,%xmm0 |
| .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_sse2 |
| .globl _sk_linear_gradient_sse2 |
| FUNCTION(_sk_linear_gradient_sse2) |
| _sk_linear_gradient_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 72,139,8 // mov (%rax),%rcx |
| .byte 72,133,201 // test %rcx,%rcx |
| .byte 15,132,15,1,0,0 // je 3e99 <_sk_linear_gradient_sse2+0x149> |
| .byte 72,139,64,8 // mov 0x8(%rax),%rax |
| .byte 72,131,192,32 // add $0x20,%rax |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 243,68,15,16,112,224 // movss -0x20(%rax),%xmm14 |
| .byte 243,68,15,16,104,228 // movss -0x1c(%rax),%xmm13 |
| .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| .byte 69,15,40,252 // movaps %xmm12,%xmm15 |
| .byte 68,15,40,224 // movaps %xmm0,%xmm12 |
| .byte 69,15,194,230,1 // cmpltps %xmm14,%xmm12 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 69,15,84,196 // andps %xmm12,%xmm8 |
| .byte 69,15,86,198 // orps %xmm14,%xmm8 |
| .byte 243,68,15,16,104,232 // movss -0x18(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 65,15,84,204 // andps %xmm12,%xmm1 |
| .byte 65,15,86,206 // orps %xmm14,%xmm1 |
| .byte 243,68,15,16,104,236 // movss -0x14(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 65,15,84,212 // andps %xmm12,%xmm2 |
| .byte 65,15,86,214 // orps %xmm14,%xmm2 |
| .byte 243,68,15,16,104,240 // movss -0x10(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 65,15,84,220 // andps %xmm12,%xmm3 |
| .byte 65,15,86,222 // orps %xmm14,%xmm3 |
| .byte 243,68,15,16,104,244 // movss -0xc(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 69,15,84,204 // andps %xmm12,%xmm9 |
| .byte 69,15,86,206 // orps %xmm14,%xmm9 |
| .byte 243,68,15,16,104,248 // movss -0x8(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 69,15,84,212 // andps %xmm12,%xmm10 |
| .byte 69,15,86,214 // orps %xmm14,%xmm10 |
| .byte 243,68,15,16,104,252 // movss -0x4(%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,40,244 // movaps %xmm12,%xmm14 |
| .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| .byte 69,15,84,220 // andps %xmm12,%xmm11 |
| .byte 69,15,86,222 // orps %xmm14,%xmm11 |
| .byte 243,68,15,16,40 // movss (%rax),%xmm13 |
| .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| .byte 69,15,84,252 // andps %xmm12,%xmm15 |
| .byte 69,15,85,229 // andnps %xmm13,%xmm12 |
| .byte 69,15,86,231 // orps %xmm15,%xmm12 |
| .byte 72,131,192,36 // add $0x24,%rax |
| .byte 72,255,201 // dec %rcx |
| .byte 15,133,8,255,255,255 // jne 3d9f <_sk_linear_gradient_sse2+0x4f> |
| .byte 235,13 // jmp 3ea6 <_sk_linear_gradient_sse2+0x156> |
| .byte 15,87,201 // xorps %xmm1,%xmm1 |
| .byte 15,87,210 // xorps %xmm2,%xmm2 |
| .byte 15,87,219 // xorps %xmm3,%xmm3 |
| .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| .byte 15,89,200 // mulps %xmm0,%xmm1 |
| .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| .byte 15,89,208 // mulps %xmm0,%xmm2 |
| .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| .byte 15,89,216 // mulps %xmm0,%xmm3 |
| .byte 65,15,88,220 // addps %xmm12,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_linear_gradient_2stops_sse2 |
| .globl _sk_linear_gradient_2stops_sse2 |
| FUNCTION(_sk_linear_gradient_2stops_sse2) |
| _sk_linear_gradient_2stops_sse2: |
| .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 243,15,16,0 // movss (%rax),%xmm0 |
| .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| .byte 243,15,16,80,16 // movss 0x10(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| .byte 15,88,194 // addps %xmm2,%xmm0 |
| .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| .byte 15,88,202 // addps %xmm2,%xmm1 |
| .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| .byte 15,88,211 // addps %xmm3,%xmm2 |
| .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3 |
| .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| .byte 243,68,15,16,72,28 // movss 0x1c(%rax),%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_save_xy_sse2 |
| .globl _sk_save_xy_sse2 |
| FUNCTION(_sk_save_xy_sse2) |
| _sk_save_xy_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%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 68,15,88,200 // addps %xmm0,%xmm9 |
| .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 |
| .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 |
| .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| .byte 69,15,194,218,1 // cmpltps %xmm10,%xmm11 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,225 // movd %ecx,%xmm12 |
| .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| .byte 69,15,84,220 // andps %xmm12,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 69,15,92,202 // subps %xmm10,%xmm9 |
| .byte 68,15,88,193 // addps %xmm1,%xmm8 |
| .byte 243,69,15,91,208 // cvttps2dq %xmm8,%xmm10 |
| .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 |
| .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| .byte 69,15,194,218,1 // cmpltps %xmm10,%xmm11 |
| .byte 69,15,84,220 // andps %xmm12,%xmm11 |
| .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| .byte 69,15,92,194 // subps %xmm10,%xmm8 |
| .byte 15,17,0 // movups %xmm0,(%rax) |
| .byte 15,17,72,32 // movups %xmm1,0x20(%rax) |
| .byte 68,15,17,72,64 // movups %xmm9,0x40(%rax) |
| .byte 68,15,17,64,96 // movups %xmm8,0x60(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_accumulate_sse2 |
| .globl _sk_accumulate_sse2 |
| FUNCTION(_sk_accumulate_sse2) |
| _sk_accumulate_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 68,15,16,128,128,0,0,0 // movups 0x80(%rax),%xmm8 |
| .byte 68,15,16,136,160,0,0,0 // movups 0xa0(%rax),%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| .byte 65,15,88,224 // addps %xmm8,%xmm4 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,193 // mulps %xmm1,%xmm8 |
| .byte 65,15,88,232 // addps %xmm8,%xmm5 |
| .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| .byte 65,15,88,240 // addps %xmm8,%xmm6 |
| .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| .byte 65,15,88,249 // addps %xmm9,%xmm7 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_nx_sse2 |
| .globl _sk_bilinear_nx_sse2 |
| FUNCTION(_sk_bilinear_nx_sse2) |
| _sk_bilinear_nx_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_px_sse2 |
| .globl _sk_bilinear_px_sse2 |
| FUNCTION(_sk_bilinear_px_sse2) |
| _sk_bilinear_px_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_ny_sse2 |
| .globl _sk_bilinear_ny_sse2 |
| FUNCTION(_sk_bilinear_ny_sse2) |
| _sk_bilinear_ny_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bilinear_py_sse2 |
| .globl _sk_bilinear_py_sse2 |
| FUNCTION(_sk_bilinear_py_sse2) |
| _sk_bilinear_py_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3x_sse2 |
| .globl _sk_bicubic_n3x_sse2 |
| FUNCTION(_sk_bicubic_n3x_sse2) |
| _sk_bicubic_n3x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1x_sse2 |
| .globl _sk_bicubic_n1x_sse2 |
| FUNCTION(_sk_bicubic_n1x_sse2) |
| _sk_bicubic_n1x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1x_sse2 |
| .globl _sk_bicubic_p1x_sse2 |
| FUNCTION(_sk_bicubic_p1x_sse2) |
| _sk_bicubic_p1x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 68,15,17,144,128,0,0,0 // movups %xmm10,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3x_sse2 |
| .globl _sk_bicubic_p3x_sse2 |
| FUNCTION(_sk_bicubic_p3x_sse2) |
| _sk_bicubic_p3x_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,0 // movups (%rax),%xmm0 |
| .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9 |
| .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 69,15,89,201 // mulps %xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,194 // addps %xmm10,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n3y_sse2 |
| .globl _sk_bicubic_n3y_sse2 |
| FUNCTION(_sk_bicubic_n3y_sse2) |
| _sk_bicubic_n3y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,89,192 // mulps %xmm8,%xmm8 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_n1y_sse2 |
| .globl _sk_bicubic_n1y_sse2 |
| FUNCTION(_sk_bicubic_n1y_sse2) |
| _sk_bicubic_n1y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,191 // mov $0xbf000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,0,0,128,63 // mov $0x3f800000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,92,193 // subps %xmm9,%xmm8 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p1y_sse2 |
| .globl _sk_bicubic_p1y_sse2 |
| FUNCTION(_sk_bicubic_p1y_sse2) |
| _sk_bicubic_p1y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,85,85,149,191 // mov $0xbf955555,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,217 // movd %ecx,%xmm11 |
| .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| .byte 69,15,88,208 // addps %xmm8,%xmm10 |
| .byte 68,15,17,144,160,0,0,0 // movups %xmm10,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_bicubic_p3y_sse2 |
| .globl _sk_bicubic_p3y_sse2 |
| FUNCTION(_sk_bicubic_p3y_sse2) |
| _sk_bicubic_p3y_sse2: |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 15,16,72,32 // movups 0x20(%rax),%xmm1 |
| .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9 |
| .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx |
| .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 69,15,89,201 // mulps %xmm9,%xmm9 |
| .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx |
| .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| .byte 69,15,88,194 // addps %xmm10,%xmm8 |
| .byte 69,15,89,193 // mulps %xmm9,%xmm8 |
| .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax) |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 255,224 // jmpq *%rax |
| |
| HIDDEN _sk_callback_sse2 |
| .globl _sk_callback_sse2 |
| FUNCTION(_sk_callback_sse2) |
| _sk_callback_sse2: |
| .byte 65,87 // push %r15 |
| .byte 65,86 // push %r14 |
| .byte 83 // push %rbx |
| .byte 72,129,236,128,0,0,0 // sub $0x80,%rsp |
| .byte 15,41,124,36,112 // movaps %xmm7,0x70(%rsp) |
| .byte 15,41,116,36,96 // movaps %xmm6,0x60(%rsp) |
| .byte 15,41,108,36,80 // movaps %xmm5,0x50(%rsp) |
| .byte 15,41,100,36,64 // movaps %xmm4,0x40(%rsp) |
| .byte 15,41,92,36,48 // movaps %xmm3,0x30(%rsp) |
| .byte 15,41,84,36,32 // movaps %xmm2,0x20(%rsp) |
| .byte 15,41,76,36,16 // movaps %xmm1,0x10(%rsp) |
| .byte 15,41,4,36 // movaps %xmm0,(%rsp) |
| .byte 73,137,214 // mov %rdx,%r14 |
| .byte 73,137,255 // mov %rdi,%r15 |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 72,137,243 // mov %rsi,%rbx |
| .byte 72,139,120,8 // mov 0x8(%rax),%rdi |
| .byte 190,4,0,0,0 // mov $0x4,%esi |
| .byte 255,16 // callq *(%rax) |
| .byte 72,137,222 // mov %rbx,%rsi |
| .byte 72,173 // lods %ds:(%rsi),%rax |
| .byte 76,137,255 // mov %r15,%rdi |
| .byte 76,137,242 // mov %r14,%rdx |
| .byte 15,40,4,36 // movaps (%rsp),%xmm0 |
| .byte 15,40,76,36,16 // movaps 0x10(%rsp),%xmm1 |
| .byte 15,40,84,36,32 // movaps 0x20(%rsp),%xmm2 |
| .byte 15,40,92,36,48 // movaps 0x30(%rsp),%xmm3 |
| .byte 15,40,100,36,64 // movaps 0x40(%rsp),%xmm4 |
| .byte 15,40,108,36,80 // movaps 0x50(%rsp),%xmm5 |
| .byte 15,40,116,36,96 // movaps 0x60(%rsp),%xmm6 |
| .byte 15,40,124,36,112 // movaps 0x70(%rsp),%xmm7 |
| .byte 72,129,196,128,0,0,0 // add $0x80,%rsp |
| .byte 91 // pop %rbx |
| .byte 65,94 // pop %r14 |
| .byte 65,95 // pop %r15 |
| .byte 255,224 // jmpq *%rax |
| #endif |
