AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

pretty up some SkSplicer code

Browse Source 
- build_stages.py uses gobjdump for both architectures, minor formatting changes
  - let byte arrays be written inline in splice()

Change-Id: I84bd47c18e5ae0b34b35f8c2f0a329fb1ea58f60
Reviewed-on: https://skia-review.googlesource.com/6833
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Herb Derby <herb@google.com>
This commit is contained in:
Mike Klein 2017-01-10 16:11:15 -05:00 committed by Skia Commit-Bot
parent 3444cadedd
commit fb00390b46
2 changed files with 474 additions and 507 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

412
src/splicer/SkSplicer_generated.h
View File

@ -285,248 +285,248 @@ static const unsigned int kSplice_store_f16[] = {
#else #else
static const unsigned char kSplice_clear[] = { static const unsigned char kSplice_clear[] = {
0xc5,0xfc,0x57,0xc0, // vxorps %ymm0, %ymm0, %ymm0 0xc5,0xfc,0x57,0xc0, // vxorps %ymm0,%ymm0,%ymm0
0xc5,0xf4,0x57,0xc9, // vxorps %ymm1, %ymm1, %ymm1 0xc5,0xf4,0x57,0xc9, // vxorps %ymm1,%ymm1,%ymm1
0xc5,0xec,0x57,0xd2, // vxorps %ymm2, %ymm2, %ymm2 0xc5,0xec,0x57,0xd2, // vxorps %ymm2,%ymm2,%ymm2
0xc5,0xe4,0x57,0xdb, // vxorps %ymm3, %ymm3, %ymm3 0xc5,0xe4,0x57,0xdb, // vxorps %ymm3,%ymm3,%ymm3
}; };
static const unsigned char kSplice_plus[] = { static const unsigned char kSplice_plus[] = {
0xc5,0xfc,0x58,0xc4, // vaddps %ymm4, %ymm0, %ymm0 0xc5,0xfc,0x58,0xc4, // vaddps %ymm4,%ymm0,%ymm0
0xc5,0xf4,0x58,0xcd, // vaddps %ymm5, %ymm1, %ymm1 0xc5,0xf4,0x58,0xcd, // vaddps %ymm5,%ymm1,%ymm1
0xc5,0xec,0x58,0xd6, // vaddps %ymm6, %ymm2, %ymm2 0xc5,0xec,0x58,0xd6, // vaddps %ymm6,%ymm2,%ymm2
0xc5,0xe4,0x58,0xdf, // vaddps %ymm7, %ymm3, %ymm3 0xc5,0xe4,0x58,0xdf, // vaddps %ymm7,%ymm3,%ymm3
}; };
static const unsigned char kSplice_srcover[] = { static const unsigned char kSplice_srcover[] = {
0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx),%ymm8
0xc5,0x3c,0x5c,0xc3, // vsubps %ymm3, %ymm8, %ymm8 0xc5,0x3c,0x5c,0xc3, // vsubps %ymm3,%ymm8,%ymm8
0xc4,0xc2,0x5d,0xb8,0xc0, // vfmadd231ps %ymm8, %ymm4, %ymm0 0xc4,0xc2,0x5d,0xb8,0xc0, // vfmadd231ps %ymm8,%ymm4,%ymm0
0xc4,0xc2,0x55,0xb8,0xc8, // vfmadd231ps %ymm8, %ymm5, %ymm1 0xc4,0xc2,0x55,0xb8,0xc8, // vfmadd231ps %ymm8,%ymm5,%ymm1
0xc4,0xc2,0x4d,0xb8,0xd0, // vfmadd231ps %ymm8, %ymm6, %ymm2 0xc4,0xc2,0x4d,0xb8,0xd0, // vfmadd231ps %ymm8,%ymm6,%ymm2
0xc4,0xc2,0x4d,0xb8,0xd8, // vfmadd231ps %ymm8, %ymm6, %ymm3 0xc4,0xc2,0x4d,0xb8,0xd8, // vfmadd231ps %ymm8,%ymm6,%ymm3
}; };
static const unsigned char kSplice_dstover[] = { static const unsigned char kSplice_dstover[] = {
0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx),%ymm8
0xc5,0x3c,0x5c,0xc7, // vsubps %ymm7, %ymm8, %ymm8 0xc5,0x3c,0x5c,0xc7, // vsubps %ymm7,%ymm8,%ymm8
0xc4,0xc2,0x7d,0xb8,0xe0, // vfmadd231ps %ymm8, %ymm0, %ymm4 0xc4,0xc2,0x7d,0xb8,0xe0, // vfmadd231ps %ymm8,%ymm0,%ymm4
0xc4,0xc2,0x75,0xb8,0xe8, // vfmadd231ps %ymm8, %ymm1, %ymm5 0xc4,0xc2,0x75,0xb8,0xe8, // vfmadd231ps %ymm8,%ymm1,%ymm5
0xc4,0xc2,0x6d,0xb8,0xf0, // vfmadd231ps %ymm8, %ymm2, %ymm6 0xc4,0xc2,0x6d,0xb8,0xf0, // vfmadd231ps %ymm8,%ymm2,%ymm6
0xc4,0xc2,0x6d,0xb8,0xf8, // vfmadd231ps %ymm8, %ymm2, %ymm7 0xc4,0xc2,0x6d,0xb8,0xf8, // vfmadd231ps %ymm8,%ymm2,%ymm7
}; };
static const unsigned char kSplice_clamp_0[] = { static const unsigned char kSplice_clamp_0[] = {
0xc4,0x41,0x3c,0x57,0xc0, // vxorps %ymm8, %ymm8, %ymm8 0xc4,0x41,0x3c,0x57,0xc0, // vxorps %ymm8,%ymm8,%ymm8
0xc4,0xc1,0x7c,0x5f,0xc0, // vmaxps %ymm8, %ymm0, %ymm0 0xc4,0xc1,0x7c,0x5f,0xc0, // vmaxps %ymm8,%ymm0,%ymm0
0xc4,0xc1,0x74,0x5f,0xc8, // vmaxps %ymm8, %ymm1, %ymm1 0xc4,0xc1,0x74,0x5f,0xc8, // vmaxps %ymm8,%ymm1,%ymm1
0xc4,0xc1,0x6c,0x5f,0xd0, // vmaxps %ymm8, %ymm2, %ymm2 0xc4,0xc1,0x6c,0x5f,0xd0, // vmaxps %ymm8,%ymm2,%ymm2
0xc4,0xc1,0x64,0x5f,0xd8, // vmaxps %ymm8, %ymm3, %ymm3 0xc4,0xc1,0x64,0x5f,0xd8, // vmaxps %ymm8,%ymm3,%ymm3
}; };
static const unsigned char kSplice_clamp_1[] = { static const unsigned char kSplice_clamp_1[] = {
0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx),%ymm8
0xc4,0xc1,0x7c,0x5d,0xc0, // vminps %ymm8, %ymm0, %ymm0 0xc4,0xc1,0x7c,0x5d,0xc0, // vminps %ymm8,%ymm0,%ymm0
0xc4,0xc1,0x74,0x5d,0xc8, // vminps %ymm8, %ymm1, %ymm1 0xc4,0xc1,0x74,0x5d,0xc8, // vminps %ymm8,%ymm1,%ymm1
0xc4,0xc1,0x6c,0x5d,0xd0, // vminps %ymm8, %ymm2, %ymm2 0xc4,0xc1,0x6c,0x5d,0xd0, // vminps %ymm8,%ymm2,%ymm2
0xc4,0xc1,0x64,0x5d,0xd8, // vminps %ymm8, %ymm3, %ymm3 0xc4,0xc1,0x64,0x5d,0xd8, // vminps %ymm8,%ymm3,%ymm3
}; };
static const unsigned char kSplice_clamp_a[] = { static const unsigned char kSplice_clamp_a[] = {
0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x04, // vbroadcastss 0x4(%rcx),%ymm8
0xc4,0xc1,0x64,0x5d,0xd8, // vminps %ymm8, %ymm3, %ymm3 0xc4,0xc1,0x64,0x5d,0xd8, // vminps %ymm8,%ymm3,%ymm3
0xc5,0xfc,0x5d,0xc3, // vminps %ymm3, %ymm0, %ymm0 0xc5,0xfc,0x5d,0xc3, // vminps %ymm3,%ymm0,%ymm0
0xc5,0xf4,0x5d,0xcb, // vminps %ymm3, %ymm1, %ymm1 0xc5,0xf4,0x5d,0xcb, // vminps %ymm3,%ymm1,%ymm1
0xc5,0xec,0x5d,0xd3, // vminps %ymm3, %ymm2, %ymm2 0xc5,0xec,0x5d,0xd3, // vminps %ymm3,%ymm2,%ymm2
}; };
static const unsigned char kSplice_swap[] = { static const unsigned char kSplice_swap[] = {
0xc5,0x7c,0x28,0xc3, // vmovaps %ymm3, %ymm8 0xc5,0x7c,0x28,0xc3, // vmovaps %ymm3,%ymm8
0xc5,0x7c,0x28,0xca, // vmovaps %ymm2, %ymm9 0xc5,0x7c,0x28,0xca, // vmovaps %ymm2,%ymm9
0xc5,0x7c,0x28,0xd1, // vmovaps %ymm1, %ymm10 0xc5,0x7c,0x28,0xd1, // vmovaps %ymm1,%ymm10
0xc5,0x7c,0x28,0xd8, // vmovaps %ymm0, %ymm11 0xc5,0x7c,0x28,0xd8, // vmovaps %ymm0,%ymm11
0xc5,0xfc,0x28,0xc4, // vmovaps %ymm4, %ymm0 0xc5,0xfc,0x28,0xc4, // vmovaps %ymm4,%ymm0
0xc5,0xfc,0x28,0xcd, // vmovaps %ymm5, %ymm1 0xc5,0xfc,0x28,0xcd, // vmovaps %ymm5,%ymm1
0xc5,0xfc,0x28,0xd6, // vmovaps %ymm6, %ymm2 0xc5,0xfc,0x28,0xd6, // vmovaps %ymm6,%ymm2
0xc5,0xfc,0x28,0xdf, // vmovaps %ymm7, %ymm3 0xc5,0xfc,0x28,0xdf, // vmovaps %ymm7,%ymm3
0xc5,0x7c,0x29,0xdc, // vmovaps %ymm11, %ymm4 0xc5,0x7c,0x29,0xdc, // vmovaps %ymm11,%ymm4
0xc5,0x7c,0x29,0xd5, // vmovaps %ymm10, %ymm5 0xc5,0x7c,0x29,0xd5, // vmovaps %ymm10,%ymm5
0xc5,0x7c,0x29,0xce, // vmovaps %ymm9, %ymm6 0xc5,0x7c,0x29,0xce, // vmovaps %ymm9,%ymm6
0xc5,0x7c,0x29,0xc7, // vmovaps %ymm8, %ymm7 0xc5,0x7c,0x29,0xc7, // vmovaps %ymm8,%ymm7
}; };
static const unsigned char kSplice_move_src_dst[] = { static const unsigned char kSplice_move_src_dst[] = {
0xc5,0xfc,0x28,0xe0, // vmovaps %ymm0, %ymm4 0xc5,0xfc,0x28,0xe0, // vmovaps %ymm0,%ymm4
0xc5,0xfc,0x28,0xe9, // vmovaps %ymm1, %ymm5 0xc5,0xfc,0x28,0xe9, // vmovaps %ymm1,%ymm5
0xc5,0xfc,0x28,0xf2, // vmovaps %ymm2, %ymm6 0xc5,0xfc,0x28,0xf2, // vmovaps %ymm2,%ymm6
0xc5,0xfc,0x28,0xfb, // vmovaps %ymm3, %ymm7 0xc5,0xfc,0x28,0xfb, // vmovaps %ymm3,%ymm7
}; };
static const unsigned char kSplice_move_dst_src[] = { static const unsigned char kSplice_move_dst_src[] = {
0xc5,0xfc,0x28,0xc4, // vmovaps %ymm4, %ymm0 0xc5,0xfc,0x28,0xc4, // vmovaps %ymm4,%ymm0
0xc5,0xfc,0x28,0xcd, // vmovaps %ymm5, %ymm1 0xc5,0xfc,0x28,0xcd, // vmovaps %ymm5,%ymm1
0xc5,0xfc,0x28,0xd6, // vmovaps %ymm6, %ymm2 0xc5,0xfc,0x28,0xd6, // vmovaps %ymm6,%ymm2
0xc5,0xfc,0x28,0xdf, // vmovaps %ymm7, %ymm3 0xc5,0xfc,0x28,0xdf, // vmovaps %ymm7,%ymm3
}; };
static const unsigned char kSplice_premul[] = { static const unsigned char kSplice_premul[] = {
0xc5,0xfc,0x59,0xc3, // vmulps %ymm3, %ymm0, %ymm0 0xc5,0xfc,0x59,0xc3, // vmulps %ymm3,%ymm0,%ymm0
0xc5,0xf4,0x59,0xcb, // vmulps %ymm3, %ymm1, %ymm1 0xc5,0xf4,0x59,0xcb, // vmulps %ymm3,%ymm1,%ymm1
0xc5,0xec,0x59,0xd3, // vmulps %ymm3, %ymm2, %ymm2 0xc5,0xec,0x59,0xd3, // vmulps %ymm3,%ymm2,%ymm2
}; };
static const unsigned char kSplice_unpremul[] = { static const unsigned char kSplice_unpremul[] = {
0xc4,0x41,0x3c,0x57,0xc0, // vxorps %ymm8, %ymm8, %ymm8 0xc4,0x41,0x3c,0x57,0xc0, // vxorps %ymm8,%ymm8,%ymm8
0xc4,0x41,0x64,0xc2,0xc8,0x00, // vcmpeqps %ymm8, %ymm3, %ymm9 0xc4,0x41,0x64,0xc2,0xc8,0x00, // vcmpeqps %ymm8,%ymm3,%ymm9
0xc4,0x62,0x7d,0x18,0x51,0x04, // vbroadcastss 0x4(%rcx), %ymm10 0xc4,0x62,0x7d,0x18,0x51,0x04, // vbroadcastss 0x4(%rcx),%ymm10
0xc5,0x2c,0x5e,0xd3, // vdivps %ymm3, %ymm10, %ymm10 0xc5,0x2c,0x5e,0xd3, // vdivps %ymm3,%ymm10,%ymm10
0xc4,0x43,0x2d,0x4a,0xc0,0x90, // vblendvps %ymm9, %ymm8, %ymm10, %ymm8 0xc4,0x43,0x2d,0x4a,0xc0,0x90, // vblendvps %ymm9,%ymm8,%ymm10,%ymm8
0xc5,0xbc,0x59,0xc0, // vmulps %ymm0, %ymm8, %ymm0 0xc5,0xbc,0x59,0xc0, // vmulps %ymm0,%ymm8,%ymm0
0xc5,0xbc,0x59,0xc9, // vmulps %ymm1, %ymm8, %ymm1 0xc5,0xbc,0x59,0xc9, // vmulps %ymm1,%ymm8,%ymm1
0xc5,0xbc,0x59,0xd2, // vmulps %ymm2, %ymm8, %ymm2 0xc5,0xbc,0x59,0xd2, // vmulps %ymm2,%ymm8,%ymm2
}; };
static const unsigned char kSplice_from_srgb[] = { static const unsigned char kSplice_from_srgb[] = {
0xc4,0x62,0x7d,0x18,0x41,0x1c, // vbroadcastss 0x1c(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x1c, // vbroadcastss 0x1c(%rcx),%ymm8
0xc5,0x3c,0x59,0xc8, // vmulps %ymm0, %ymm8, %ymm9 0xc5,0x3c,0x59,0xc8, // vmulps %ymm0,%ymm8,%ymm9
0xc5,0x7c,0x59,0xd0, // vmulps %ymm0, %ymm0, %ymm10 0xc5,0x7c,0x59,0xd0, // vmulps %ymm0,%ymm0,%ymm10
0xc4,0x62,0x7d,0x18,0x59,0x18, // vbroadcastss 0x18(%rcx), %ymm11 0xc4,0x62,0x7d,0x18,0x59,0x18, // vbroadcastss 0x18(%rcx),%ymm11
0xc4,0x62,0x7d,0x18,0x61,0x14, // vbroadcastss 0x14(%rcx), %ymm12 0xc4,0x62,0x7d,0x18,0x61,0x14, // vbroadcastss 0x14(%rcx),%ymm12
0xc4,0x41,0x7c,0x28,0xeb, // vmovaps %ymm11, %ymm13 0xc4,0x41,0x7c,0x28,0xeb, // vmovaps %ymm11,%ymm13
0xc4,0x42,0x7d,0xa8,0xec, // vfmadd213ps %ymm12, %ymm0, %ymm13 0xc4,0x42,0x7d,0xa8,0xec, // vfmadd213ps %ymm12,%ymm0,%ymm13
0xc4,0x62,0x7d,0x18,0x71,0x10, // vbroadcastss 0x10(%rcx), %ymm14 0xc4,0x62,0x7d,0x18,0x71,0x10, // vbroadcastss 0x10(%rcx),%ymm14
0xc4,0x42,0x2d,0xa8,0xee, // vfmadd213ps %ymm14, %ymm10, %ymm13 0xc4,0x42,0x2d,0xa8,0xee, // vfmadd213ps %ymm14,%ymm10,%ymm13
0xc4,0x62,0x7d,0x18,0x51,0x20, // vbroadcastss 0x20(%rcx), %ymm10 0xc4,0x62,0x7d,0x18,0x51,0x20, // vbroadcastss 0x20(%rcx),%ymm10
0xc4,0xc1,0x7c,0xc2,0xc2,0x01, // vcmpltps %ymm10, %ymm0, %ymm0 0xc4,0xc1,0x7c,0xc2,0xc2,0x01, // vcmpltps %ymm10,%ymm0,%ymm0
0xc4,0xc3,0x15,0x4a,0xc1,0x00, // vblendvps %ymm0, %ymm9, %ymm13, %ymm0 0xc4,0xc3,0x15,0x4a,0xc1,0x00, // vblendvps %ymm0,%ymm9,%ymm13,%ymm0
0xc5,0x3c,0x59,0xc9, // vmulps %ymm1, %ymm8, %ymm9 0xc5,0x3c,0x59,0xc9, // vmulps %ymm1,%ymm8,%ymm9
0xc5,0x74,0x59,0xe9, // vmulps %ymm1, %ymm1, %ymm13 0xc5,0x74,0x59,0xe9, // vmulps %ymm1,%ymm1,%ymm13
0xc4,0x41,0x7c,0x28,0xfb, // vmovaps %ymm11, %ymm15 0xc4,0x41,0x7c,0x28,0xfb, // vmovaps %ymm11,%ymm15
0xc4,0x42,0x75,0xa8,0xfc, // vfmadd213ps %ymm12, %ymm1, %ymm15 0xc4,0x42,0x75,0xa8,0xfc, // vfmadd213ps %ymm12,%ymm1,%ymm15
0xc4,0x42,0x15,0xa8,0xfe, // vfmadd213ps %ymm14, %ymm13, %ymm15 0xc4,0x42,0x15,0xa8,0xfe, // vfmadd213ps %ymm14,%ymm13,%ymm15
0xc4,0xc1,0x74,0xc2,0xca,0x01, // vcmpltps %ymm10, %ymm1, %ymm1 0xc4,0xc1,0x74,0xc2,0xca,0x01, // vcmpltps %ymm10,%ymm1,%ymm1
0xc4,0xc3,0x05,0x4a,0xc9,0x10, // vblendvps %ymm1, %ymm9, %ymm15, %ymm1 0xc4,0xc3,0x05,0x4a,0xc9,0x10, // vblendvps %ymm1,%ymm9,%ymm15,%ymm1
0xc5,0x3c,0x59,0xc2, // vmulps %ymm2, %ymm8, %ymm8 0xc5,0x3c,0x59,0xc2, // vmulps %ymm2,%ymm8,%ymm8
0xc5,0x6c,0x59,0xca, // vmulps %ymm2, %ymm2, %ymm9 0xc5,0x6c,0x59,0xca, // vmulps %ymm2,%ymm2,%ymm9
0xc4,0x42,0x6d,0xa8,0xdc, // vfmadd213ps %ymm12, %ymm2, %ymm11 0xc4,0x42,0x6d,0xa8,0xdc, // vfmadd213ps %ymm12,%ymm2,%ymm11
0xc4,0x42,0x35,0xa8,0xde, // vfmadd213ps %ymm14, %ymm9, %ymm11 0xc4,0x42,0x35,0xa8,0xde, // vfmadd213ps %ymm14,%ymm9,%ymm11
0xc4,0xc1,0x6c,0xc2,0xd2,0x01, // vcmpltps %ymm10, %ymm2, %ymm2 0xc4,0xc1,0x6c,0xc2,0xd2,0x01, // vcmpltps %ymm10,%ymm2,%ymm2
0xc4,0xc3,0x25,0x4a,0xd0,0x20, // vblendvps %ymm2, %ymm8, %ymm11, %ymm2 0xc4,0xc3,0x25,0x4a,0xd0,0x20, // vblendvps %ymm2,%ymm8,%ymm11,%ymm2
}; };
static const unsigned char kSplice_to_srgb[] = { static const unsigned char kSplice_to_srgb[] = {
0xc5,0x7c,0x52,0xc0, // vrsqrtps %ymm0, %ymm8 0xc5,0x7c,0x52,0xc0, // vrsqrtps %ymm0,%ymm8
0xc4,0x41,0x7c,0x53,0xc8, // vrcpps %ymm8, %ymm9 0xc4,0x41,0x7c,0x53,0xc8, // vrcpps %ymm8,%ymm9
0xc4,0x41,0x7c,0x52,0xd0, // vrsqrtps %ymm8, %ymm10 0xc4,0x41,0x7c,0x52,0xd0, // vrsqrtps %ymm8,%ymm10
0xc4,0x62,0x7d,0x18,0x41,0x24, // vbroadcastss 0x24(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x24, // vbroadcastss 0x24(%rcx),%ymm8
0xc5,0x3c,0x59,0xd8, // vmulps %ymm0, %ymm8, %ymm11 0xc5,0x3c,0x59,0xd8, // vmulps %ymm0,%ymm8,%ymm11
0xc4,0x62,0x7d,0x18,0x61,0x04, // vbroadcastss 0x4(%rcx), %ymm12 0xc4,0x62,0x7d,0x18,0x61,0x04, // vbroadcastss 0x4(%rcx),%ymm12
0xc4,0x62,0x7d,0x18,0x69,0x28, // vbroadcastss 0x28(%rcx), %ymm13 0xc4,0x62,0x7d,0x18,0x69,0x28, // vbroadcastss 0x28(%rcx),%ymm13
0xc4,0x62,0x7d,0x18,0x71,0x2c, // vbroadcastss 0x2c(%rcx), %ymm14 0xc4,0x62,0x7d,0x18,0x71,0x2c, // vbroadcastss 0x2c(%rcx),%ymm14
0xc4,0x62,0x7d,0x18,0x79,0x30, // vbroadcastss 0x30(%rcx), %ymm15 0xc4,0x62,0x7d,0x18,0x79,0x30, // vbroadcastss 0x30(%rcx),%ymm15
0xc4,0x42,0x0d,0xa8,0xcf, // vfmadd213ps %ymm15, %ymm14, %ymm9 0xc4,0x42,0x0d,0xa8,0xcf, // vfmadd213ps %ymm15,%ymm14,%ymm9
0xc4,0x42,0x15,0xb8,0xca, // vfmadd231ps %ymm10, %ymm13, %ymm9 0xc4,0x42,0x15,0xb8,0xca, // vfmadd231ps %ymm10,%ymm13,%ymm9
0xc4,0x41,0x1c,0x5d,0xc9, // vminps %ymm9, %ymm12, %ymm9 0xc4,0x41,0x1c,0x5d,0xc9, // vminps %ymm9,%ymm12,%ymm9
0xc4,0x62,0x7d,0x18,0x51,0x34, // vbroadcastss 0x34(%rcx), %ymm10 0xc4,0x62,0x7d,0x18,0x51,0x34, // vbroadcastss 0x34(%rcx),%ymm10
0xc4,0xc1,0x7c,0xc2,0xc2,0x01, // vcmpltps %ymm10, %ymm0, %ymm0 0xc4,0xc1,0x7c,0xc2,0xc2,0x01, // vcmpltps %ymm10,%ymm0,%ymm0
0xc4,0xc3,0x35,0x4a,0xc3,0x00, // vblendvps %ymm0, %ymm11, %ymm9, %ymm0 0xc4,0xc3,0x35,0x4a,0xc3,0x00, // vblendvps %ymm0,%ymm11,%ymm9,%ymm0
0xc5,0x7c,0x52,0xc9, // vrsqrtps %ymm1, %ymm9 0xc5,0x7c,0x52,0xc9, // vrsqrtps %ymm1,%ymm9
0xc4,0x41,0x7c,0x53,0xd9, // vrcpps %ymm9, %ymm11 0xc4,0x41,0x7c,0x53,0xd9, // vrcpps %ymm9,%ymm11
0xc4,0x41,0x7c,0x52,0xc9, // vrsqrtps %ymm9, %ymm9 0xc4,0x41,0x7c,0x52,0xc9, // vrsqrtps %ymm9,%ymm9
0xc4,0x42,0x0d,0xa8,0xdf, // vfmadd213ps %ymm15, %ymm14, %ymm11 0xc4,0x42,0x0d,0xa8,0xdf, // vfmadd213ps %ymm15,%ymm14,%ymm11
0xc4,0x42,0x15,0xb8,0xd9, // vfmadd231ps %ymm9, %ymm13, %ymm11 0xc4,0x42,0x15,0xb8,0xd9, // vfmadd231ps %ymm9,%ymm13,%ymm11
0xc5,0x3c,0x59,0xc9, // vmulps %ymm1, %ymm8, %ymm9 0xc5,0x3c,0x59,0xc9, // vmulps %ymm1,%ymm8,%ymm9
0xc4,0x41,0x1c,0x5d,0xdb, // vminps %ymm11, %ymm12, %ymm11 0xc4,0x41,0x1c,0x5d,0xdb, // vminps %ymm11,%ymm12,%ymm11
0xc4,0xc1,0x74,0xc2,0xca,0x01, // vcmpltps %ymm10, %ymm1, %ymm1 0xc4,0xc1,0x74,0xc2,0xca,0x01, // vcmpltps %ymm10,%ymm1,%ymm1
0xc4,0xc3,0x25,0x4a,0xc9,0x10, // vblendvps %ymm1, %ymm9, %ymm11, %ymm1 0xc4,0xc3,0x25,0x4a,0xc9,0x10, // vblendvps %ymm1,%ymm9,%ymm11,%ymm1
0xc5,0x7c,0x52,0xca, // vrsqrtps %ymm2, %ymm9 0xc5,0x7c,0x52,0xca, // vrsqrtps %ymm2,%ymm9
0xc4,0x41,0x7c,0x53,0xd9, // vrcpps %ymm9, %ymm11 0xc4,0x41,0x7c,0x53,0xd9, // vrcpps %ymm9,%ymm11
0xc4,0x42,0x0d,0xa8,0xdf, // vfmadd213ps %ymm15, %ymm14, %ymm11 0xc4,0x42,0x0d,0xa8,0xdf, // vfmadd213ps %ymm15,%ymm14,%ymm11
0xc4,0x41,0x7c,0x52,0xc9, // vrsqrtps %ymm9, %ymm9 0xc4,0x41,0x7c,0x52,0xc9, // vrsqrtps %ymm9,%ymm9
0xc4,0x42,0x15,0xb8,0xd9, // vfmadd231ps %ymm9, %ymm13, %ymm11 0xc4,0x42,0x15,0xb8,0xd9, // vfmadd231ps %ymm9,%ymm13,%ymm11
0xc4,0x41,0x1c,0x5d,0xcb, // vminps %ymm11, %ymm12, %ymm9 0xc4,0x41,0x1c,0x5d,0xcb, // vminps %ymm11,%ymm12,%ymm9
0xc5,0x3c,0x59,0xc2, // vmulps %ymm2, %ymm8, %ymm8 0xc5,0x3c,0x59,0xc2, // vmulps %ymm2,%ymm8,%ymm8
0xc4,0xc1,0x6c,0xc2,0xd2,0x01, // vcmpltps %ymm10, %ymm2, %ymm2 0xc4,0xc1,0x6c,0xc2,0xd2,0x01, // vcmpltps %ymm10,%ymm2,%ymm2
0xc4,0xc3,0x35,0x4a,0xd0,0x20, // vblendvps %ymm2, %ymm8, %ymm9, %ymm2 0xc4,0xc3,0x35,0x4a,0xd0,0x20, // vblendvps %ymm2,%ymm8,%ymm9,%ymm2
}; };
static const unsigned char kSplice_scale_u8[] = { static const unsigned char kSplice_scale_u8[] = {
0x48,0x8b,0x02, // movq (%rdx), %rax 0x48,0x8b,0x02, // mov (%rdx),%rax
0xc4,0x62,0x7d,0x31,0x04,0x38, // vpmovzxbd (%rax,%rdi), %ymm8 0xc4,0x62,0x7d,0x31,0x04,0x38, // vpmovzxbd (%rax,%rdi,1),%ymm8
0xc4,0x41,0x7c,0x5b,0xc0, // vcvtdq2ps %ymm8, %ymm8 0xc4,0x41,0x7c,0x5b,0xc0, // vcvtdq2ps %ymm8,%ymm8
0xc4,0x62,0x7d,0x18,0x49,0x0c, // vbroadcastss 0xc(%rcx), %ymm9 0xc4,0x62,0x7d,0x18,0x49,0x0c, // vbroadcastss 0xc(%rcx),%ymm9
0xc4,0x41,0x3c,0x59,0xc1, // vmulps %ymm9, %ymm8, %ymm8 0xc4,0x41,0x3c,0x59,0xc1, // vmulps %ymm9,%ymm8,%ymm8
0xc5,0xbc,0x59,0xc0, // vmulps %ymm0, %ymm8, %ymm0 0xc5,0xbc,0x59,0xc0, // vmulps %ymm0,%ymm8,%ymm0
0xc5,0xbc,0x59,0xc9, // vmulps %ymm1, %ymm8, %ymm1 0xc5,0xbc,0x59,0xc9, // vmulps %ymm1,%ymm8,%ymm1
0xc5,0xbc,0x59,0xd2, // vmulps %ymm2, %ymm8, %ymm2 0xc5,0xbc,0x59,0xd2, // vmulps %ymm2,%ymm8,%ymm2
0xc5,0xbc,0x59,0xdb, // vmulps %ymm3, %ymm8, %ymm3 0xc5,0xbc,0x59,0xdb, // vmulps %ymm3,%ymm8,%ymm3
}; };
static const unsigned char kSplice_load_8888[] = { static const unsigned char kSplice_load_8888[] = {
0x48,0x8b,0x02, // movq (%rdx), %rax 0x48,0x8b,0x02, // mov (%rdx),%rax
0xc5,0xfc,0x10,0x1c,0xb8, // vmovups (%rax,%rdi,4), %ymm3 0xc5,0xfc,0x10,0x1c,0xb8, // vmovups (%rax,%rdi,4),%ymm3
0xc4,0xe2,0x7d,0x18,0x11, // vbroadcastss (%rcx), %ymm2 0xc4,0xe2,0x7d,0x18,0x11, // vbroadcastss (%rcx),%ymm2
0xc5,0xec,0x54,0xc3, // vandps %ymm3, %ymm2, %ymm0 0xc5,0xec,0x54,0xc3, // vandps %ymm3,%ymm2,%ymm0
0xc5,0xfc,0x5b,0xc0, // vcvtdq2ps %ymm0, %ymm0 0xc5,0xfc,0x5b,0xc0, // vcvtdq2ps %ymm0,%ymm0
0xc4,0x62,0x7d,0x18,0x41,0x0c, // vbroadcastss 0xc(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x0c, // vbroadcastss 0xc(%rcx),%ymm8
0xc5,0xbc,0x59,0xc0, // vmulps %ymm0, %ymm8, %ymm0 0xc5,0xbc,0x59,0xc0, // vmulps %ymm0,%ymm8,%ymm0
0xc5,0xf5,0x72,0xd3,0x08, // vpsrld $0x8, %ymm3, %ymm1 0xc5,0xf5,0x72,0xd3,0x08, // vpsrld $0x8,%ymm3,%ymm1
0xc5,0xec,0x54,0xc9, // vandps %ymm1, %ymm2, %ymm1 0xc5,0xec,0x54,0xc9, // vandps %ymm1,%ymm2,%ymm1
0xc5,0xfc,0x5b,0xc9, // vcvtdq2ps %ymm1, %ymm1 0xc5,0xfc,0x5b,0xc9, // vcvtdq2ps %ymm1,%ymm1
0xc5,0xbc,0x59,0xc9, // vmulps %ymm1, %ymm8, %ymm1 0xc5,0xbc,0x59,0xc9, // vmulps %ymm1,%ymm8,%ymm1
0xc5,0xb5,0x72,0xd3,0x10, // vpsrld $0x10, %ymm3, %ymm9 0xc5,0xb5,0x72,0xd3,0x10, // vpsrld $0x10,%ymm3,%ymm9
0xc4,0xc1,0x6c,0x54,0xd1, // vandps %ymm9, %ymm2, %ymm2 0xc4,0xc1,0x6c,0x54,0xd1, // vandps %ymm9,%ymm2,%ymm2
0xc5,0xfc,0x5b,0xd2, // vcvtdq2ps %ymm2, %ymm2 0xc5,0xfc,0x5b,0xd2, // vcvtdq2ps %ymm2,%ymm2
0xc5,0xbc,0x59,0xd2, // vmulps %ymm2, %ymm8, %ymm2 0xc5,0xbc,0x59,0xd2, // vmulps %ymm2,%ymm8,%ymm2
0xc5,0xe5,0x72,0xd3,0x18, // vpsrld $0x18, %ymm3, %ymm3 0xc5,0xe5,0x72,0xd3,0x18, // vpsrld $0x18,%ymm3,%ymm3
0xc5,0xfc,0x5b,0xdb, // vcvtdq2ps %ymm3, %ymm3 0xc5,0xfc,0x5b,0xdb, // vcvtdq2ps %ymm3,%ymm3
0xc4,0xc1,0x64,0x59,0xd8, // vmulps %ymm8, %ymm3, %ymm3 0xc4,0xc1,0x64,0x59,0xd8, // vmulps %ymm8,%ymm3,%ymm3
}; };
static const unsigned char kSplice_store_8888[] = { static const unsigned char kSplice_store_8888[] = {
0x48,0x8b,0x02, // movq (%rdx), %rax 0x48,0x8b,0x02, // mov (%rdx),%rax
0xc4,0x62,0x7d,0x18,0x41,0x08, // vbroadcastss 0x8(%rcx), %ymm8 0xc4,0x62,0x7d,0x18,0x41,0x08, // vbroadcastss 0x8(%rcx),%ymm8
0xc5,0x3c,0x59,0xc8, // vmulps %ymm0, %ymm8, %ymm9 0xc5,0x3c,0x59,0xc8, // vmulps %ymm0,%ymm8,%ymm9
0xc4,0x41,0x7d,0x5b,0xc9, // vcvtps2dq %ymm9, %ymm9 0xc4,0x41,0x7d,0x5b,0xc9, // vcvtps2dq %ymm9,%ymm9
0xc5,0x3c,0x59,0xd1, // vmulps %ymm1, %ymm8, %ymm10 0xc5,0x3c,0x59,0xd1, // vmulps %ymm1,%ymm8,%ymm10
0xc4,0x41,0x7d,0x5b,0xd2, // vcvtps2dq %ymm10, %ymm10 0xc4,0x41,0x7d,0x5b,0xd2, // vcvtps2dq %ymm10,%ymm10
0xc4,0xc1,0x2d,0x72,0xf2,0x08, // vpslld $0x8, %ymm10, %ymm10 0xc4,0xc1,0x2d,0x72,0xf2,0x08, // vpslld $0x8,%ymm10,%ymm10
0xc4,0x41,0x2d,0xeb,0xc9, // vpor %ymm9, %ymm10, %ymm9 0xc4,0x41,0x2d,0xeb,0xc9, // vpor %ymm9,%ymm10,%ymm9
0xc5,0x3c,0x59,0xd2, // vmulps %ymm2, %ymm8, %ymm10 0xc5,0x3c,0x59,0xd2, // vmulps %ymm2,%ymm8,%ymm10
0xc4,0x41,0x7d,0x5b,0xd2, // vcvtps2dq %ymm10, %ymm10 0xc4,0x41,0x7d,0x5b,0xd2, // vcvtps2dq %ymm10,%ymm10
0xc4,0xc1,0x2d,0x72,0xf2,0x10, // vpslld $0x10, %ymm10, %ymm10 0xc4,0xc1,0x2d,0x72,0xf2,0x10, // vpslld $0x10,%ymm10,%ymm10
0xc5,0x3c,0x59,0xc3, // vmulps %ymm3, %ymm8, %ymm8 0xc5,0x3c,0x59,0xc3, // vmulps %ymm3,%ymm8,%ymm8
0xc4,0x41,0x7d,0x5b,0xc0, // vcvtps2dq %ymm8, %ymm8 0xc4,0x41,0x7d,0x5b,0xc0, // vcvtps2dq %ymm8,%ymm8
0xc4,0xc1,0x3d,0x72,0xf0,0x18, // vpslld $0x18, %ymm8, %ymm8 0xc4,0xc1,0x3d,0x72,0xf0,0x18, // vpslld $0x18,%ymm8,%ymm8
0xc4,0x41,0x2d,0xeb,0xc0, // vpor %ymm8, %ymm10, %ymm8 0xc4,0x41,0x2d,0xeb,0xc0, // vpor %ymm8,%ymm10,%ymm8
0xc4,0x41,0x35,0xeb,0xc0, // vpor %ymm8, %ymm9, %ymm8 0xc4,0x41,0x35,0xeb,0xc0, // vpor %ymm8,%ymm9,%ymm8
0xc5,0x7e,0x7f,0x04,0xb8, // vmovdqu %ymm8, (%rax,%rdi,4) 0xc5,0x7e,0x7f,0x04,0xb8, // vmovdqu %ymm8,(%rax,%rdi,4)
}; };
static const unsigned char kSplice_load_f16[] = { static const unsigned char kSplice_load_f16[] = {
0x48,0x8b,0x02, // movq (%rdx), %rax 0x48,0x8b,0x02, // mov (%rdx),%rax
0xc5,0xfa,0x6f,0x04,0xf8, // vmovdqu (%rax,%rdi,8), %xmm0 0xc5,0xfa,0x6f,0x04,0xf8, // vmovdqu (%rax,%rdi,8),%xmm0
0xc5,0xfa,0x6f,0x4c,0xf8,0x10, // vmovdqu 0x10(%rax,%rdi,8), %xmm1 0xc5,0xfa,0x6f,0x4c,0xf8,0x10, // vmovdqu 0x10(%rax,%rdi,8),%xmm1
0xc5,0xfa,0x6f,0x54,0xf8,0x20, // vmovdqu 0x20(%rax,%rdi,8), %xmm2 0xc5,0xfa,0x6f,0x54,0xf8,0x20, // vmovdqu 0x20(%rax,%rdi,8),%xmm2
0xc5,0xfa,0x6f,0x5c,0xf8,0x30, // vmovdqu 0x30(%rax,%rdi,8), %xmm3 0xc5,0xfa,0x6f,0x5c,0xf8,0x30, // vmovdqu 0x30(%rax,%rdi,8),%xmm3
0xc5,0x79,0x61,0xc1, // vpunpcklwd %xmm1, %xmm0, %xmm8 0xc5,0x79,0x61,0xc1, // vpunpcklwd %xmm1,%xmm0,%xmm8
0xc5,0xf9,0x69,0xc1, // vpunpckhwd %xmm1, %xmm0, %xmm0 0xc5,0xf9,0x69,0xc1, // vpunpckhwd %xmm1,%xmm0,%xmm0
0xc5,0xe9,0x61,0xcb, // vpunpcklwd %xmm3, %xmm2, %xmm1 0xc5,0xe9,0x61,0xcb, // vpunpcklwd %xmm3,%xmm2,%xmm1
0xc5,0xe9,0x69,0xd3, // vpunpckhwd %xmm3, %xmm2, %xmm2 0xc5,0xe9,0x69,0xd3, // vpunpckhwd %xmm3,%xmm2,%xmm2
0xc5,0x39,0x61,0xc8, // vpunpcklwd %xmm0, %xmm8, %xmm9 0xc5,0x39,0x61,0xc8, // vpunpcklwd %xmm0,%xmm8,%xmm9
0xc5,0x39,0x69,0xc0, // vpunpckhwd %xmm0, %xmm8, %xmm8 0xc5,0x39,0x69,0xc0, // vpunpckhwd %xmm0,%xmm8,%xmm8
0xc5,0xf1,0x61,0xda, // vpunpcklwd %xmm2, %xmm1, %xmm3 0xc5,0xf1,0x61,0xda, // vpunpcklwd %xmm2,%xmm1,%xmm3
0xc5,0x71,0x69,0xd2, // vpunpckhwd %xmm2, %xmm1, %xmm10 0xc5,0x71,0x69,0xd2, // vpunpckhwd %xmm2,%xmm1,%xmm10
0xc5,0xb1,0x6c,0xc3, // vpunpcklqdq %xmm3, %xmm9, %xmm0 0xc5,0xb1,0x6c,0xc3, // vpunpcklqdq %xmm3,%xmm9,%xmm0
0xc4,0xe2,0x7d,0x13,0xc0, // vcvtph2ps %xmm0, %ymm0 0xc4,0xe2,0x7d,0x13,0xc0, // vcvtph2ps %xmm0,%ymm0
0xc5,0xb1,0x6d,0xcb, // vpunpckhqdq %xmm3, %xmm9, %xmm1 0xc5,0xb1,0x6d,0xcb, // vpunpckhqdq %xmm3,%xmm9,%xmm1
0xc4,0xe2,0x7d,0x13,0xc9, // vcvtph2ps %xmm1, %ymm1 0xc4,0xe2,0x7d,0x13,0xc9, // vcvtph2ps %xmm1,%ymm1
0xc4,0xc1,0x39,0x6c,0xd2, // vpunpcklqdq %xmm10, %xmm8, %xmm2 0xc4,0xc1,0x39,0x6c,0xd2, // vpunpcklqdq %xmm10,%xmm8,%xmm2
0xc4,0xe2,0x7d,0x13,0xd2, // vcvtph2ps %xmm2, %ymm2 0xc4,0xe2,0x7d,0x13,0xd2, // vcvtph2ps %xmm2,%ymm2
0xc4,0xc1,0x39,0x6d,0xda, // vpunpckhqdq %xmm10, %xmm8, %xmm3 0xc4,0xc1,0x39,0x6d,0xda, // vpunpckhqdq %xmm10,%xmm8,%xmm3
0xc4,0xe2,0x7d,0x13,0xdb, // vcvtph2ps %xmm3, %ymm3 0xc4,0xe2,0x7d,0x13,0xdb, // vcvtph2ps %xmm3,%ymm3
}; };
static const unsigned char kSplice_store_f16[] = { static const unsigned char kSplice_store_f16[] = {
0x48,0x8b,0x02, // movq (%rdx), %rax 0x48,0x8b,0x02, // mov (%rdx),%rax
0xc4,0xc3,0x7d,0x1d,0xc0,0x04, // vcvtps2ph $0x4, %ymm0, %xmm8 0xc4,0xc3,0x7d,0x1d,0xc0,0x04, // vcvtps2ph $0x4,%ymm0,%xmm8
0xc4,0xc3,0x7d,0x1d,0xc9,0x04, // vcvtps2ph $0x4, %ymm1, %xmm9 0xc4,0xc3,0x7d,0x1d,0xc9,0x04, // vcvtps2ph $0x4,%ymm1,%xmm9
0xc4,0xc3,0x7d,0x1d,0xd2,0x04, // vcvtps2ph $0x4, %ymm2, %xmm10 0xc4,0xc3,0x7d,0x1d,0xd2,0x04, // vcvtps2ph $0x4,%ymm2,%xmm10
0xc4,0xc3,0x7d,0x1d,0xdb,0x04, // vcvtps2ph $0x4, %ymm3, %xmm11 0xc4,0xc3,0x7d,0x1d,0xdb,0x04, // vcvtps2ph $0x4,%ymm3,%xmm11
0xc4,0x41,0x39,0x61,0xe1, // vpunpcklwd %xmm9, %xmm8, %xmm12 0xc4,0x41,0x39,0x61,0xe1, // vpunpcklwd %xmm9,%xmm8,%xmm12
0xc4,0x41,0x39,0x69,0xc1, // vpunpckhwd %xmm9, %xmm8, %xmm8 0xc4,0x41,0x39,0x69,0xc1, // vpunpckhwd %xmm9,%xmm8,%xmm8
0xc4,0x41,0x29,0x61,0xcb, // vpunpcklwd %xmm11, %xmm10, %xmm9 0xc4,0x41,0x29,0x61,0xcb, // vpunpcklwd %xmm11,%xmm10,%xmm9
0xc4,0x41,0x29,0x69,0xd3, // vpunpckhwd %xmm11, %xmm10, %xmm10 0xc4,0x41,0x29,0x69,0xd3, // vpunpckhwd %xmm11,%xmm10,%xmm10
0xc4,0x41,0x19,0x62,0xd9, // vpunpckldq %xmm9, %xmm12, %xmm11 0xc4,0x41,0x19,0x62,0xd9, // vpunpckldq %xmm9,%xmm12,%xmm11
0xc5,0x7a,0x7f,0x1c,0xf8, // vmovdqu %xmm11, (%rax,%rdi,8) 0xc5,0x7a,0x7f,0x1c,0xf8, // vmovdqu %xmm11,(%rax,%rdi,8)
0xc4,0x41,0x19,0x6a,0xc9, // vpunpckhdq %xmm9, %xmm12, %xmm9 0xc4,0x41,0x19,0x6a,0xc9, // vpunpckhdq %xmm9,%xmm12,%xmm9
0xc5,0x7a,0x7f,0x4c,0xf8,0x10, // vmovdqu %xmm9, 0x10(%rax,%rdi,8) 0xc5,0x7a,0x7f,0x4c,0xf8,0x10, // vmovdqu %xmm9,0x10(%rax,%rdi,8)
0xc4,0x41,0x39,0x62,0xca, // vpunpckldq %xmm10, %xmm8, %xmm9 0xc4,0x41,0x39,0x62,0xca, // vpunpckldq %xmm10,%xmm8,%xmm9
0xc5,0x7a,0x7f,0x4c,0xf8,0x20, // vmovdqu %xmm9, 0x20(%rax,%rdi,8) 0xc5,0x7a,0x7f,0x4c,0xf8,0x20, // vmovdqu %xmm9,0x20(%rax,%rdi,8)
0xc4,0x41,0x39,0x6a,0xc2, // vpunpckhdq %xmm10, %xmm8, %xmm8 0xc4,0x41,0x39,0x6a,0xc2, // vpunpckhdq %xmm10,%xmm8,%xmm8
0xc5,0x7a,0x7f,0x44,0xf8,0x30, // vmovdqu %xmm8, 0x30(%rax,%rdi,8) 0xc5,0x7a,0x7f,0x44,0xf8,0x30, // vmovdqu %xmm8,0x30(%rax,%rdi,8)
}; };
#endif #endif

109
src/splicer/build_stages.py
View File

@ -25,6 +25,39 @@ subprocess.check_call(['clang++'] + cflags + aarch64 +
['-c', 'src/splicer/SkSplicer_stages.cpp'] + ['-c', 'src/splicer/SkSplicer_stages.cpp'] +
['-o', 'aarch64.o']) ['-o', 'aarch64.o'])
def parse_object_file(dot_o, array_type, done):
for line in subprocess.check_output(['gobjdump', '-d', dot_o]).split('\n'):
line = line.strip()
if not line or line.startswith(dot_o) or line.startswith('Disassembly'):
continue
# E.g. 00000000000003a4 <_load_f16>:
m = re.match('''................ <_?(.*)>:''', line)
if m:
print 'static const', array_type, 'kSplice_' + m.group(1) + '[] = {'
continue
columns = line.split('\t')
code = columns[1]
if len(columns) == 4:
inst = columns[2]
args = columns[3]
else:
inst, args = columns[2].split(' ', 1)
code, inst, args = code.strip(), inst.strip(), args.strip()
# We can't splice code that uses ip-relative addressing.
for arg in args:
assert 'rip' not in arg # TODO: detect on aarch64 too
if code == done:
print '};'
continue
hexed = ''.join('0x'+x+',' for x in code.split(' '))
print ' ' + hexed + ' '*(44-len(hexed)) + \
'// ' + inst + ' '*(14-len(inst)) + args
print '''/* print '''/*
* Copyright 2017 Google Inc. * Copyright 2017 Google Inc.
* *
@ -40,74 +73,8 @@ print '''/*
#if defined(__aarch64__) #if defined(__aarch64__)
''' '''
for line in subprocess.check_output(['gobjdump', '-d', parse_object_file('aarch64.o', 'unsigned int', '14000000')
'aarch64.o']).split('\n'): print '\n#else\n'
line = line.strip() parse_object_file('hsw.o', 'unsigned char', 'e9 00 00 00 00')
if not line or line.startswith('aarch64.o') or line.startswith('Disassembly'): print '\n#endif\n'
continue print '#endif//SkSplicer_generated_DEFINED'
m = re.match('''................ <(.*)>:''', line)
if m:
print 'static const unsigned int kSplice_' + m.group(1) + '[] = {'
continue
_, code, inst, args = line.split('\t')
code = code.strip()
# b done, where done has not yet been filled in by the linker.
if code == '14000000':
print '};'
continue
print ' ', '0x'+code+',' + ' // ' + inst + ' ' + args
print '''
#else
'''
# TODO: port this to gobjdump too
for line in subprocess.check_output(['otool', '-tvj', 'hsw.o']).split('\n'):
line = line.strip()
if line == '' or line == 'hsw.o:' or line == '(__TEXT,__text) section':
continue
m = re.match('_(.*):', line)
if m:
name = m.group(1)
print 'static const unsigned char kSplice_' + m.group(1) + '[] = {'
continue
# Skip the leading 16 byte address and a tab,
# leaving the code, instruction mnemonic, and its arguments.
line = line[17:]
columns = line.split('\t')
code = columns[0].strip()
inst = columns[1]
args = columns[2:]
# We can't splice code that uses rip relative addressing.
for arg in args:
assert 'rip' not in arg
# jmp done, the end of each stage (the address of done is not yet filled in)
if code == 'e9 00 00 00 00':
print '};'
continue
sys.stdout.write(' ')
_bytes = code.split(' ')
# This is the meat of things: copy the code to a C unsigned char array.
for byte in _bytes:
sys.stdout.write('0x' + byte + ',')
# From here on we're just making the generated file readable and pretty.
sys.stdout.write(' ' * (44 - 5*len(_bytes)))
sys.stdout.write('// ' + inst)
if args:
sys.stdout.write(' ' * (13 - len(inst)))
sys.stdout.write(' '.join(args))
sys.stdout.write('\n')
print '''
#endif
'''
print '''#endif//SkSplicer_generated_DEFINED'''