mirror of https://github.com/goccy/go-json.git
444 lines
21 KiB
ArmAsm
444 lines
21 KiB
ArmAsm
//+build !noasm !appengine
|
|
// AUTO-GENERATED BY C2GOASM -- DO NOT EDIT
|
|
|
|
TEXT ·_findEscapeIndex64(SB), $0-24
|
|
|
|
MOVQ buf+0(FP), DI
|
|
MOVQ len+8(FP), SI
|
|
|
|
WORD $0xf089 // mov eax, esi
|
|
WORD $0xf8c1; BYTE $0x1f // sar eax, 31
|
|
WORD $0xe8c1; BYTE $0x1d // shr eax, 29
|
|
WORD $0xf001 // add eax, esi
|
|
WORD $0xf8c1; BYTE $0x03 // sar eax, 3
|
|
WORD $0x9848 // cdqe
|
|
WORD $0xc683; BYTE $0x07 // add esi, 7
|
|
WORD $0xfe83; BYTE $0x0f // cmp esi, 15
|
|
JB LBB0_5
|
|
QUAD $0xfefefefefeffbd49; WORD $0xfefe // mov r13, -72340172838076673
|
|
WORD $0xd231 // xor edx, edx
|
|
QUAD $0x222222222222b949; WORD $0x2222 // mov r9, 2459565876494606882
|
|
QUAD $0x5c5c5c5c5c5cba49; WORD $0x5c5c // mov r10, 6655295901103053916
|
|
QUAD $0x3c3c3c3c3c3cbb49; WORD $0x3c3c // mov r11, 4340410370284600380
|
|
QUAD $0x3e3e3e3e3e3ebe49; WORD $0x3e3e // mov r14, 4485090715960753726
|
|
QUAD $0x262626262626bf49; WORD $0x2626 // mov r15, 2748926567846913574
|
|
QUAD $0x808080808080bc49; WORD $0x8080 // mov r12, -9187201950435737472
|
|
|
|
LBB0_2:
|
|
LONG $0xd71c8b48 // mov rbx, qword [rdi + 8*rdx]
|
|
QUAD $0xdfdfdfdfdfe0b948; WORD $0xdfdf // mov rcx, -2314885530818453536
|
|
LONG $0x0b348d48 // lea rsi, [rbx + rcx]
|
|
WORD $0x0948; BYTE $0xde // or rsi, rbx
|
|
WORD $0x8948; BYTE $0xd9 // mov rcx, rbx
|
|
WORD $0x314c; BYTE $0xc9 // xor rcx, r9
|
|
WORD $0x014c; BYTE $0xe9 // add rcx, r13
|
|
WORD $0x0948; BYTE $0xf1 // or rcx, rsi
|
|
WORD $0x8948; BYTE $0xde // mov rsi, rbx
|
|
WORD $0x314c; BYTE $0xd6 // xor rsi, r10
|
|
WORD $0x014c; BYTE $0xee // add rsi, r13
|
|
WORD $0x8949; BYTE $0xd8 // mov r8, rbx
|
|
WORD $0x314d; BYTE $0xd8 // xor r8, r11
|
|
WORD $0x014d; BYTE $0xe8 // add r8, r13
|
|
WORD $0x0949; BYTE $0xf0 // or r8, rsi
|
|
WORD $0x0949; BYTE $0xc8 // or r8, rcx
|
|
WORD $0x8948; BYTE $0xd9 // mov rcx, rbx
|
|
WORD $0x314c; BYTE $0xf1 // xor rcx, r14
|
|
WORD $0x014c; BYTE $0xe9 // add rcx, r13
|
|
WORD $0x314c; BYTE $0xfb // xor rbx, r15
|
|
WORD $0x014c; BYTE $0xeb // add rbx, r13
|
|
WORD $0x0948; BYTE $0xcb // or rbx, rcx
|
|
WORD $0x094c; BYTE $0xc3 // or rbx, r8
|
|
WORD $0x214c; BYTE $0xe3 // and rbx, r12
|
|
JNE LBB0_3
|
|
LONG $0x01c28348 // add rdx, 1
|
|
WORD $0x3948; BYTE $0xc2 // cmp rdx, rax
|
|
JB LBB0_2
|
|
|
|
LBB0_5:
|
|
LONG $0x03e0c148 // shl rax, 3
|
|
JMP LBB0_6
|
|
|
|
LBB0_3:
|
|
WORD $0xbc0f; BYTE $0xc3 // bsf eax, ebx
|
|
|
|
LBB0_6:
|
|
MOVQ AX, ret+16(FP)
|
|
RET
|
|
|
|
DATA LCDATA1<>+0x000(SB)/8, $0xdfdfdfdfdfdfdfe0
|
|
DATA LCDATA1<>+0x008(SB)/8, $0xdfdfdfdfdfdfdfe0
|
|
DATA LCDATA1<>+0x010(SB)/8, $0x2222222222222222
|
|
DATA LCDATA1<>+0x018(SB)/8, $0x2222222222222222
|
|
DATA LCDATA1<>+0x020(SB)/8, $0xfefefefefefefeff
|
|
DATA LCDATA1<>+0x028(SB)/8, $0xfefefefefefefeff
|
|
DATA LCDATA1<>+0x030(SB)/8, $0x5c5c5c5c5c5c5c5c
|
|
DATA LCDATA1<>+0x038(SB)/8, $0x5c5c5c5c5c5c5c5c
|
|
DATA LCDATA1<>+0x040(SB)/8, $0x3c3c3c3c3c3c3c3c
|
|
DATA LCDATA1<>+0x048(SB)/8, $0x3c3c3c3c3c3c3c3c
|
|
DATA LCDATA1<>+0x050(SB)/8, $0x3e3e3e3e3e3e3e3e
|
|
DATA LCDATA1<>+0x058(SB)/8, $0x3e3e3e3e3e3e3e3e
|
|
DATA LCDATA1<>+0x060(SB)/8, $0x2626262626262626
|
|
DATA LCDATA1<>+0x068(SB)/8, $0x2626262626262626
|
|
DATA LCDATA1<>+0x070(SB)/8, $0x8080808080808080
|
|
DATA LCDATA1<>+0x078(SB)/8, $0x8080808080808080
|
|
GLOBL LCDATA1<>(SB), 8, $128
|
|
|
|
TEXT ·_findEscapeIndex128(SB), $0-24
|
|
|
|
MOVQ buf+0(FP), DI
|
|
MOVQ len+8(FP), SI
|
|
LEAQ LCDATA1<>(SB), BP
|
|
|
|
WORD $0xf089 // mov eax, esi
|
|
WORD $0xf8c1; BYTE $0x1f // sar eax, 31
|
|
WORD $0xe8c1; BYTE $0x1c // shr eax, 28
|
|
WORD $0xf001 // add eax, esi
|
|
WORD $0xf8c1; BYTE $0x04 // sar eax, 4
|
|
WORD $0x4e8d; BYTE $0x0f // lea ecx, [rsi + 15]
|
|
WORD $0xf983; BYTE $0x1f // cmp ecx, 31
|
|
JB LBB1_5
|
|
WORD $0x6348; BYTE $0xc8 // movsxd rcx, eax
|
|
WORD $0xd231 // xor edx, edx
|
|
LONG $0x456f79c5; BYTE $0x00 // vmovdqa xmm8, oword 0[rbp] /* [rip + .LCPI1_0] */
|
|
LONG $0x4d6f79c5; BYTE $0x10 // vmovdqa xmm9, oword 16[rbp] /* [rip + .LCPI1_1] */
|
|
LONG $0x556ff9c5; BYTE $0x20 // vmovdqa xmm2, oword 32[rbp] /* [rip + .LCPI1_2] */
|
|
LONG $0x556f79c5; BYTE $0x30 // vmovdqa xmm10, oword 48[rbp] /* [rip + .LCPI1_3] */
|
|
LONG $0x5d6f79c5; BYTE $0x40 // vmovdqa xmm11, oword 64[rbp] /* [rip + .LCPI1_4] */
|
|
LONG $0x656f79c5; BYTE $0x50 // vmovdqa xmm12, oword 80[rbp] /* [rip + .LCPI1_5] */
|
|
LONG $0x6d6f79c5; BYTE $0x60 // vmovdqa xmm13, oword 96[rbp] /* [rip + .LCPI1_6] */
|
|
LONG $0x7d6ff9c5; BYTE $0x70 // vmovdqa xmm7, oword 112[rbp] /* [rip + .LCPI1_7] */
|
|
|
|
LBB1_2:
|
|
LONG $0x076ffac5 // vmovdqu xmm0, oword [rdi]
|
|
LONG $0xd479c1c4; BYTE $0xc8 // vpaddq xmm1, xmm0, xmm8
|
|
LONG $0xef79c1c4; BYTE $0xd9 // vpxor xmm3, xmm0, xmm9
|
|
LONG $0xdad4e1c5 // vpaddq xmm3, xmm3, xmm2
|
|
LONG $0xef79c1c4; BYTE $0xe2 // vpxor xmm4, xmm0, xmm10
|
|
LONG $0xe2d4d9c5 // vpaddq xmm4, xmm4, xmm2
|
|
LONG $0xef79c1c4; BYTE $0xeb // vpxor xmm5, xmm0, xmm11
|
|
LONG $0xead4d1c5 // vpaddq xmm5, xmm5, xmm2
|
|
LONG $0xe5ebd9c5 // vpor xmm4, xmm4, xmm5
|
|
LONG $0xef79c1c4; BYTE $0xec // vpxor xmm5, xmm0, xmm12
|
|
LONG $0xead4d1c5 // vpaddq xmm5, xmm5, xmm2
|
|
LONG $0xef79c1c4; BYTE $0xf5 // vpxor xmm6, xmm0, xmm13
|
|
LONG $0xf2d4c9c5 // vpaddq xmm6, xmm6, xmm2
|
|
LONG $0xeeebd1c5 // vpor xmm5, xmm5, xmm6
|
|
LONG $0xc0ebf1c5 // vpor xmm0, xmm1, xmm0
|
|
LONG $0xc3ebf9c5 // vpor xmm0, xmm0, xmm3
|
|
LONG $0xc4ebf9c5 // vpor xmm0, xmm0, xmm4
|
|
LONG $0xc5ebf9c5 // vpor xmm0, xmm0, xmm5
|
|
LONG $0xc7dbf9c5 // vpand xmm0, xmm0, xmm7
|
|
LONG $0xd8d7f9c5 // vpmovmskb ebx, xmm0
|
|
WORD $0xdb85 // test ebx, ebx
|
|
JNE LBB1_3
|
|
LONG $0x10c78348 // add rdi, 16
|
|
LONG $0x01c28348 // add rdx, 1
|
|
WORD $0x3948; BYTE $0xca // cmp rdx, rcx
|
|
JB LBB1_2
|
|
|
|
LBB1_5:
|
|
WORD $0xe0c1; BYTE $0x04 // shl eax, 4
|
|
WORD $0xc629 // sub esi, eax
|
|
WORD $0xfe83; BYTE $0x08 // cmp esi, 8
|
|
JL LBB1_11
|
|
QUAD $0xfefefefefeffba49; WORD $0xfefe // mov r10, -72340172838076673
|
|
WORD $0x8941; BYTE $0xf5 // mov r13d, esi
|
|
LONG $0x03edc141 // shr r13d, 3
|
|
WORD $0xd231 // xor edx, edx
|
|
QUAD $0x3c3c3c3c3c3cbb49; WORD $0x3c3c // mov r11, 4340410370284600380
|
|
QUAD $0x3e3e3e3e3e3ebe49; WORD $0x3e3e // mov r14, 4485090715960753726
|
|
QUAD $0x262626262626bf49; WORD $0x2626 // mov r15, 2748926567846913574
|
|
QUAD $0x808080808080bc49; WORD $0x8080 // mov r12, -9187201950435737472
|
|
|
|
LBB1_7:
|
|
LONG $0xd71c8b48 // mov rbx, qword [rdi + 8*rdx]
|
|
QUAD $0xdfdfdfdfdfe0b948; WORD $0xdfdf // mov rcx, -2314885530818453536
|
|
LONG $0x0b0c8d48 // lea rcx, [rbx + rcx]
|
|
WORD $0x0948; BYTE $0xd9 // or rcx, rbx
|
|
WORD $0x8949; BYTE $0xd8 // mov r8, rbx
|
|
QUAD $0x222222222222b848; WORD $0x2222 // mov rax, 2459565876494606882
|
|
WORD $0x3149; BYTE $0xc0 // xor r8, rax
|
|
WORD $0x014d; BYTE $0xd0 // add r8, r10
|
|
WORD $0x0949; BYTE $0xc8 // or r8, rcx
|
|
WORD $0x8948; BYTE $0xd9 // mov rcx, rbx
|
|
QUAD $0x5c5c5c5c5c5cb848; WORD $0x5c5c // mov rax, 6655295901103053916
|
|
WORD $0x3148; BYTE $0xc1 // xor rcx, rax
|
|
WORD $0x014c; BYTE $0xd1 // add rcx, r10
|
|
WORD $0x8949; BYTE $0xd9 // mov r9, rbx
|
|
WORD $0x314d; BYTE $0xd9 // xor r9, r11
|
|
WORD $0x014d; BYTE $0xd1 // add r9, r10
|
|
WORD $0x0949; BYTE $0xc9 // or r9, rcx
|
|
WORD $0x094d; BYTE $0xc1 // or r9, r8
|
|
WORD $0x8948; BYTE $0xd9 // mov rcx, rbx
|
|
WORD $0x314c; BYTE $0xf1 // xor rcx, r14
|
|
WORD $0x014c; BYTE $0xd1 // add rcx, r10
|
|
WORD $0x314c; BYTE $0xfb // xor rbx, r15
|
|
WORD $0x014c; BYTE $0xd3 // add rbx, r10
|
|
WORD $0x0948; BYTE $0xcb // or rbx, rcx
|
|
WORD $0x094c; BYTE $0xcb // or rbx, r9
|
|
WORD $0x214c; BYTE $0xe3 // and rbx, r12
|
|
JNE LBB1_8
|
|
LONG $0x01c28348 // add rdx, 1
|
|
WORD $0x394c; BYTE $0xea // cmp rdx, r13
|
|
JB LBB1_7
|
|
WORD $0xe683; BYTE $0xf8 // and esi, -8
|
|
JMP LBB1_12
|
|
|
|
LBB1_11:
|
|
WORD $0x6348; BYTE $0xf0 // movsxd rsi, eax
|
|
JMP LBB1_12
|
|
|
|
LBB1_3:
|
|
WORD $0xbc0f; BYTE $0xf3 // bsf esi, ebx
|
|
JMP LBB1_12
|
|
|
|
LBB1_8:
|
|
WORD $0xbc0f; BYTE $0xf3 // bsf esi, ebx
|
|
|
|
LBB1_12:
|
|
WORD $0x8948; BYTE $0xf0 // mov rax, rsi
|
|
MOVQ AX, ret+16(FP)
|
|
RET
|
|
|
|
DATA LCDATA2<>+0x000(SB)/8, $0xdfdfdfdfdfdfdfe0
|
|
DATA LCDATA2<>+0x008(SB)/8, $0x2222222222222222
|
|
DATA LCDATA2<>+0x010(SB)/8, $0xfefefefefefefeff
|
|
DATA LCDATA2<>+0x018(SB)/8, $0x5c5c5c5c5c5c5c5c
|
|
DATA LCDATA2<>+0x020(SB)/8, $0x3c3c3c3c3c3c3c3c
|
|
DATA LCDATA2<>+0x028(SB)/8, $0x3e3e3e3e3e3e3e3e
|
|
DATA LCDATA2<>+0x030(SB)/8, $0x2626262626262626
|
|
DATA LCDATA2<>+0x038(SB)/8, $0x0000000000000000
|
|
DATA LCDATA2<>+0x040(SB)/8, $0x8080808080808080
|
|
DATA LCDATA2<>+0x048(SB)/8, $0x8080808080808080
|
|
DATA LCDATA2<>+0x050(SB)/8, $0x8080808080808080
|
|
DATA LCDATA2<>+0x058(SB)/8, $0x8080808080808080
|
|
DATA LCDATA2<>+0x060(SB)/8, $0xdfdfdfdfdfdfdfe0
|
|
DATA LCDATA2<>+0x068(SB)/8, $0xdfdfdfdfdfdfdfe0
|
|
DATA LCDATA2<>+0x070(SB)/8, $0x2222222222222222
|
|
DATA LCDATA2<>+0x078(SB)/8, $0x2222222222222222
|
|
DATA LCDATA2<>+0x080(SB)/8, $0xfefefefefefefeff
|
|
DATA LCDATA2<>+0x088(SB)/8, $0xfefefefefefefeff
|
|
DATA LCDATA2<>+0x090(SB)/8, $0x5c5c5c5c5c5c5c5c
|
|
DATA LCDATA2<>+0x098(SB)/8, $0x5c5c5c5c5c5c5c5c
|
|
DATA LCDATA2<>+0x0a0(SB)/8, $0x3c3c3c3c3c3c3c3c
|
|
DATA LCDATA2<>+0x0a8(SB)/8, $0x3c3c3c3c3c3c3c3c
|
|
DATA LCDATA2<>+0x0b0(SB)/8, $0x3e3e3e3e3e3e3e3e
|
|
DATA LCDATA2<>+0x0b8(SB)/8, $0x3e3e3e3e3e3e3e3e
|
|
DATA LCDATA2<>+0x0c0(SB)/8, $0x2626262626262626
|
|
DATA LCDATA2<>+0x0c8(SB)/8, $0x2626262626262626
|
|
DATA LCDATA2<>+0x0d0(SB)/8, $0x8080808080808080
|
|
DATA LCDATA2<>+0x0d8(SB)/8, $0x8080808080808080
|
|
GLOBL LCDATA2<>(SB), 8, $224
|
|
|
|
TEXT ·_findEscapeIndex256(SB), $0-24
|
|
|
|
MOVQ buf+0(FP), DI
|
|
MOVQ len+8(FP), SI
|
|
LEAQ LCDATA2<>(SB), BP
|
|
|
|
WORD $0xf089 // mov eax, esi
|
|
WORD $0xf8c1; BYTE $0x1f // sar eax, 31
|
|
WORD $0xe8c1; BYTE $0x1b // shr eax, 27
|
|
WORD $0xf001 // add eax, esi
|
|
WORD $0xf8c1; BYTE $0x05 // sar eax, 5
|
|
WORD $0x4e8d; BYTE $0x1f // lea ecx, [rsi + 31]
|
|
WORD $0xf983; BYTE $0x3f // cmp ecx, 63
|
|
JB LBB2_4
|
|
WORD $0x6348; BYTE $0xc8 // movsxd rcx, eax
|
|
LONG $0x597de2c4; WORD $0x0045 // vpbroadcastq ymm0, qword 0[rbp] /* [rip + .LCPI2_0] */
|
|
LONG $0x597de2c4; WORD $0x084d // vpbroadcastq ymm1, qword 8[rbp] /* [rip + .LCPI2_1] */
|
|
LONG $0x597de2c4; WORD $0x1055 // vpbroadcastq ymm2, qword 16[rbp] /* [rip + .LCPI2_2] */
|
|
LONG $0x597de2c4; WORD $0x185d // vpbroadcastq ymm3, qword 24[rbp] /* [rip + .LCPI2_3] */
|
|
LONG $0x597de2c4; WORD $0x2065 // vpbroadcastq ymm4, qword 32[rbp] /* [rip + .LCPI2_4] */
|
|
LONG $0x597de2c4; WORD $0x286d // vpbroadcastq ymm5, qword 40[rbp] /* [rip + .LCPI2_5] */
|
|
WORD $0xd231 // xor edx, edx
|
|
LONG $0x597de2c4; WORD $0x3075 // vpbroadcastq ymm6, qword 48[rbp] /* [rip + .LCPI2_6] */
|
|
LONG $0x7d6ffdc5; BYTE $0x40 // vmovdqa ymm7, yword 64[rbp] /* [rip + .LCPI2_7] */
|
|
|
|
LBB2_2:
|
|
LONG $0x076f7ec5 // vmovdqu ymm8, yword [rdi]
|
|
LONG $0xc8d43dc5 // vpaddq ymm9, ymm8, ymm0
|
|
LONG $0xd1ef3dc5 // vpxor ymm10, ymm8, ymm1
|
|
LONG $0xd2d42dc5 // vpaddq ymm10, ymm10, ymm2
|
|
LONG $0xdbef3dc5 // vpxor ymm11, ymm8, ymm3
|
|
LONG $0xdad425c5 // vpaddq ymm11, ymm11, ymm2
|
|
LONG $0xe4ef3dc5 // vpxor ymm12, ymm8, ymm4
|
|
LONG $0xe2d41dc5 // vpaddq ymm12, ymm12, ymm2
|
|
LONG $0xeb2541c4; BYTE $0xdc // vpor ymm11, ymm11, ymm12
|
|
LONG $0xe5ef3dc5 // vpxor ymm12, ymm8, ymm5
|
|
LONG $0xe2d41dc5 // vpaddq ymm12, ymm12, ymm2
|
|
LONG $0xeeef3dc5 // vpxor ymm13, ymm8, ymm6
|
|
LONG $0xead415c5 // vpaddq ymm13, ymm13, ymm2
|
|
LONG $0xeb1d41c4; BYTE $0xe5 // vpor ymm12, ymm12, ymm13
|
|
LONG $0xeb3541c4; BYTE $0xc0 // vpor ymm8, ymm9, ymm8
|
|
LONG $0xeb3d41c4; BYTE $0xc2 // vpor ymm8, ymm8, ymm10
|
|
LONG $0xeb3d41c4; BYTE $0xc3 // vpor ymm8, ymm8, ymm11
|
|
LONG $0xeb3d41c4; BYTE $0xc4 // vpor ymm8, ymm8, ymm12
|
|
LONG $0xc7db3dc5 // vpand ymm8, ymm8, ymm7
|
|
LONG $0xd77dc1c4; BYTE $0xd8 // vpmovmskb ebx, ymm8
|
|
WORD $0xdb85 // test ebx, ebx
|
|
JNE LBB2_19
|
|
LONG $0x20c78348 // add rdi, 32
|
|
LONG $0x01c28348 // add rdx, 1
|
|
WORD $0x3948; BYTE $0xca // cmp rdx, rcx
|
|
JB LBB2_2
|
|
|
|
LBB2_4:
|
|
WORD $0xe0c1; BYTE $0x05 // shl eax, 5
|
|
WORD $0xc629 // sub esi, eax
|
|
WORD $0xfe83; BYTE $0x10 // cmp esi, 16
|
|
JL LBB2_13
|
|
WORD $0xf089 // mov eax, esi
|
|
WORD $0xe8c1; BYTE $0x04 // shr eax, 4
|
|
WORD $0xc931 // xor ecx, ecx
|
|
LONG $0x456f79c5; BYTE $0x60 // vmovdqa xmm8, oword 96[rbp] /* [rip + .LCPI2_8] */
|
|
LONG $0x4d6f79c5; BYTE $0x70 // vmovdqa xmm9, oword 112[rbp] /* [rip + .LCPI2_9] */
|
|
QUAD $0x00000080956ff9c5 // vmovdqa xmm2, oword 128[rbp] /* [rip + .LCPI2_10] */
|
|
QUAD $0x00000090956f79c5 // vmovdqa xmm10, oword 144[rbp] /* [rip + .LCPI2_11] */
|
|
QUAD $0x000000a09d6f79c5 // vmovdqa xmm11, oword 160[rbp] /* [rip + .LCPI2_12] */
|
|
QUAD $0x000000b0a56f79c5 // vmovdqa xmm12, oword 176[rbp] /* [rip + .LCPI2_13] */
|
|
QUAD $0x000000c0ad6f79c5 // vmovdqa xmm13, oword 192[rbp] /* [rip + .LCPI2_14] */
|
|
QUAD $0x000000d0bd6ff9c5 // vmovdqa xmm7, oword 208[rbp] /* [rip + .LCPI2_15] */
|
|
|
|
LBB2_6:
|
|
LONG $0x076ffac5 // vmovdqu xmm0, oword [rdi]
|
|
LONG $0xd479c1c4; BYTE $0xc8 // vpaddq xmm1, xmm0, xmm8
|
|
LONG $0xef79c1c4; BYTE $0xd9 // vpxor xmm3, xmm0, xmm9
|
|
LONG $0xdad4e1c5 // vpaddq xmm3, xmm3, xmm2
|
|
LONG $0xef79c1c4; BYTE $0xe2 // vpxor xmm4, xmm0, xmm10
|
|
LONG $0xe2d4d9c5 // vpaddq xmm4, xmm4, xmm2
|
|
LONG $0xef79c1c4; BYTE $0xeb // vpxor xmm5, xmm0, xmm11
|
|
LONG $0xead4d1c5 // vpaddq xmm5, xmm5, xmm2
|
|
LONG $0xe5ebd9c5 // vpor xmm4, xmm4, xmm5
|
|
LONG $0xef79c1c4; BYTE $0xec // vpxor xmm5, xmm0, xmm12
|
|
LONG $0xead4d1c5 // vpaddq xmm5, xmm5, xmm2
|
|
LONG $0xef79c1c4; BYTE $0xf5 // vpxor xmm6, xmm0, xmm13
|
|
LONG $0xf2d4c9c5 // vpaddq xmm6, xmm6, xmm2
|
|
LONG $0xeeebd1c5 // vpor xmm5, xmm5, xmm6
|
|
LONG $0xc0ebf1c5 // vpor xmm0, xmm1, xmm0
|
|
LONG $0xc3ebf9c5 // vpor xmm0, xmm0, xmm3
|
|
LONG $0xc4ebf9c5 // vpor xmm0, xmm0, xmm4
|
|
LONG $0xc5ebf9c5 // vpor xmm0, xmm0, xmm5
|
|
LONG $0xc7dbf9c5 // vpand xmm0, xmm0, xmm7
|
|
LONG $0xd0d7f9c5 // vpmovmskb edx, xmm0
|
|
WORD $0xd285 // test edx, edx
|
|
JNE LBB2_20
|
|
LONG $0x10c78348 // add rdi, 16
|
|
LONG $0x01c18348 // add rcx, 1
|
|
WORD $0x3948; BYTE $0xc1 // cmp rcx, rax
|
|
JB LBB2_6
|
|
WORD $0xf089 // mov eax, esi
|
|
WORD $0xe083; BYTE $0xf0 // and eax, -16
|
|
WORD $0xc629 // sub esi, eax
|
|
WORD $0xfe83; BYTE $0x08 // cmp esi, 8
|
|
JL LBB2_18
|
|
QUAD $0xfefefefefeffb848; WORD $0xfefe // mov rax, -72340172838076673
|
|
QUAD $0x5c5c5c5c5c5cba49; WORD $0x5c5c // mov r10, 6655295901103053916
|
|
QUAD $0x3e3e3e3e3e3ebb49; WORD $0x3e3e // mov r11, 4485090715960753726
|
|
QUAD $0x3c3c3c3c3c3cbe49; WORD $0x3c3c // mov r14, 4340410370284600380
|
|
QUAD $0x262626262626bf49; WORD $0x2626 // mov r15, 2748926567846913574
|
|
QUAD $0x222222222222bc49; WORD $0x2222 // mov r12, 2459565876494606882
|
|
WORD $0x8941; BYTE $0xf5 // mov r13d, esi
|
|
LONG $0x03edc141 // shr r13d, 3
|
|
WORD $0xdb31 // xor ebx, ebx
|
|
|
|
LBB2_10:
|
|
LONG $0xdf148b48 // mov rdx, qword [rdi + 8*rbx]
|
|
QUAD $0xdfdfdfdfdfe0b948; WORD $0xdfdf // mov rcx, -2314885530818453536
|
|
LONG $0x0a0c8d48 // lea rcx, [rdx + rcx]
|
|
WORD $0x0948; BYTE $0xd1 // or rcx, rdx
|
|
WORD $0x8949; BYTE $0xd0 // mov r8, rdx
|
|
WORD $0x314d; BYTE $0xe0 // xor r8, r12
|
|
WORD $0x0149; BYTE $0xc0 // add r8, rax
|
|
WORD $0x0949; BYTE $0xc8 // or r8, rcx
|
|
WORD $0x8948; BYTE $0xd1 // mov rcx, rdx
|
|
WORD $0x314c; BYTE $0xd1 // xor rcx, r10
|
|
WORD $0x0148; BYTE $0xc1 // add rcx, rax
|
|
WORD $0x8949; BYTE $0xd1 // mov r9, rdx
|
|
WORD $0x314d; BYTE $0xf1 // xor r9, r14
|
|
WORD $0x0149; BYTE $0xc1 // add r9, rax
|
|
WORD $0x0949; BYTE $0xc9 // or r9, rcx
|
|
WORD $0x094d; BYTE $0xc1 // or r9, r8
|
|
WORD $0x8948; BYTE $0xd1 // mov rcx, rdx
|
|
WORD $0x314c; BYTE $0xd9 // xor rcx, r11
|
|
WORD $0x0148; BYTE $0xc1 // add rcx, rax
|
|
WORD $0x314c; BYTE $0xfa // xor rdx, r15
|
|
WORD $0x0148; BYTE $0xc2 // add rdx, rax
|
|
WORD $0x0948; BYTE $0xca // or rdx, rcx
|
|
WORD $0x094c; BYTE $0xca // or rdx, r9
|
|
QUAD $0x808080808080b948; WORD $0x8080 // mov rcx, -9187201950435737472
|
|
WORD $0x2148; BYTE $0xca // and rdx, rcx
|
|
JNE LBB2_22
|
|
LONG $0x01c38348 // add rbx, 1
|
|
WORD $0x394c; BYTE $0xeb // cmp rbx, r13
|
|
JB LBB2_10
|
|
WORD $0xe683; BYTE $0xf8 // and esi, -8
|
|
JMP LBB2_23
|
|
|
|
LBB2_13:
|
|
WORD $0xfe83; BYTE $0x08 // cmp esi, 8
|
|
JL LBB2_18
|
|
QUAD $0xfefefefefeffbd49; WORD $0xfefe // mov r13, -72340172838076673
|
|
QUAD $0x808080808080b949; WORD $0x8080 // mov r9, -9187201950435737472
|
|
QUAD $0x5c5c5c5c5c5cba49; WORD $0x5c5c // mov r10, 6655295901103053916
|
|
QUAD $0x3e3e3e3e3e3ebb49; WORD $0x3e3e // mov r11, 4485090715960753726
|
|
QUAD $0x3c3c3c3c3c3cbe49; WORD $0x3c3c // mov r14, 4340410370284600380
|
|
QUAD $0x262626262626bf49; WORD $0x2626 // mov r15, 2748926567846913574
|
|
QUAD $0x222222222222bc49; WORD $0x2222 // mov r12, 2459565876494606882
|
|
WORD $0xeec1; BYTE $0x03 // shr esi, 3
|
|
WORD $0xe683; BYTE $0x1f // and esi, 31
|
|
WORD $0xc931 // xor ecx, ecx
|
|
|
|
LBB2_15:
|
|
LONG $0xcf148b48 // mov rdx, qword [rdi + 8*rcx]
|
|
QUAD $0xdfdfdfdfdfe0b848; WORD $0xdfdf // mov rax, -2314885530818453536
|
|
LONG $0x021c8d48 // lea rbx, [rdx + rax]
|
|
WORD $0x0948; BYTE $0xd3 // or rbx, rdx
|
|
WORD $0x8948; BYTE $0xd0 // mov rax, rdx
|
|
WORD $0x314c; BYTE $0xe0 // xor rax, r12
|
|
WORD $0x014c; BYTE $0xe8 // add rax, r13
|
|
WORD $0x0948; BYTE $0xd8 // or rax, rbx
|
|
WORD $0x8948; BYTE $0xd3 // mov rbx, rdx
|
|
WORD $0x314c; BYTE $0xd3 // xor rbx, r10
|
|
WORD $0x014c; BYTE $0xeb // add rbx, r13
|
|
WORD $0x8949; BYTE $0xd0 // mov r8, rdx
|
|
WORD $0x314d; BYTE $0xf0 // xor r8, r14
|
|
WORD $0x014d; BYTE $0xe8 // add r8, r13
|
|
WORD $0x0949; BYTE $0xd8 // or r8, rbx
|
|
WORD $0x0949; BYTE $0xc0 // or r8, rax
|
|
WORD $0x8948; BYTE $0xd0 // mov rax, rdx
|
|
WORD $0x314c; BYTE $0xd8 // xor rax, r11
|
|
WORD $0x014c; BYTE $0xe8 // add rax, r13
|
|
WORD $0x314c; BYTE $0xfa // xor rdx, r15
|
|
WORD $0x014c; BYTE $0xea // add rdx, r13
|
|
WORD $0x0948; BYTE $0xc2 // or rdx, rax
|
|
WORD $0x094c; BYTE $0xc2 // or rdx, r8
|
|
WORD $0x214c; BYTE $0xca // and rdx, r9
|
|
JNE LBB2_22
|
|
LONG $0x01c18348 // add rcx, 1
|
|
WORD $0x3948; BYTE $0xf1 // cmp rcx, rsi
|
|
JB LBB2_15
|
|
WORD $0xe6c1; BYTE $0x03 // shl esi, 3
|
|
JMP LBB2_23
|
|
|
|
LBB2_18:
|
|
WORD $0x6348; BYTE $0xf0 // movsxd rsi, eax
|
|
JMP LBB2_23
|
|
|
|
LBB2_19:
|
|
WORD $0xbc0f; BYTE $0xf3 // bsf esi, ebx
|
|
JMP LBB2_23
|
|
|
|
LBB2_20:
|
|
WORD $0xbc0f; BYTE $0xf2 // bsf esi, edx
|
|
JMP LBB2_23
|
|
|
|
LBB2_22:
|
|
WORD $0xbc0f; BYTE $0xf2 // bsf esi, edx
|
|
|
|
LBB2_23:
|
|
WORD $0x8948; BYTE $0xf0 // mov rax, rsi
|
|
VZEROUPPER
|
|
MOVQ AX, ret+16(FP)
|
|
RET
|