diff options
Diffstat (limited to 'vendor/github.com/minio/md5-simd/md5block_amd64.s')
| -rw-r--r-- | vendor/github.com/minio/md5-simd/md5block_amd64.s | 714 |
1 files changed, 714 insertions, 0 deletions
diff --git a/vendor/github.com/minio/md5-simd/md5block_amd64.s b/vendor/github.com/minio/md5-simd/md5block_amd64.s new file mode 100644 index 00000000..fbc4a21f --- /dev/null +++ b/vendor/github.com/minio/md5-simd/md5block_amd64.s @@ -0,0 +1,714 @@ +// Code generated by command: go run gen.go -out ../md5block_amd64.s -stubs ../md5block_amd64.go -pkg=md5simd. DO NOT EDIT. + +// +build !appengine +// +build !noasm +// +build gc + +// func blockScalar(dig *[4]uint32, p []byte) +TEXT ·blockScalar(SB), $0-32 + MOVQ p_len+16(FP), AX + MOVQ dig+0(FP), CX + MOVQ p_base+8(FP), DX + SHRQ $0x06, AX + SHLQ $0x06, AX + LEAQ (DX)(AX*1), AX + CMPQ DX, AX + JEQ end + MOVL (CX), BX + MOVL 4(CX), BP + MOVL 8(CX), SI + MOVL 12(CX), CX + MOVL $0xffffffff, DI + +loop: + MOVL (DX), R8 + MOVL CX, R9 + MOVL BX, R10 + MOVL BP, R11 + MOVL SI, R12 + MOVL CX, R13 + + // ROUND1 + XORL SI, R9 + ADDL $0xd76aa478, BX + ADDL R8, BX + ANDL BP, R9 + XORL CX, R9 + MOVL 4(DX), R8 + ADDL R9, BX + ROLL $0x07, BX + MOVL SI, R9 + ADDL BP, BX + XORL BP, R9 + ADDL $0xe8c7b756, CX + ADDL R8, CX + ANDL BX, R9 + XORL SI, R9 + MOVL 8(DX), R8 + ADDL R9, CX + ROLL $0x0c, CX + MOVL BP, R9 + ADDL BX, CX + XORL BX, R9 + ADDL $0x242070db, SI + ADDL R8, SI + ANDL CX, R9 + XORL BP, R9 + MOVL 12(DX), R8 + ADDL R9, SI + ROLL $0x11, SI + MOVL BX, R9 + ADDL CX, SI + XORL CX, R9 + ADDL $0xc1bdceee, BP + ADDL R8, BP + ANDL SI, R9 + XORL BX, R9 + MOVL 16(DX), R8 + ADDL R9, BP + ROLL $0x16, BP + MOVL CX, R9 + ADDL SI, BP + XORL SI, R9 + ADDL $0xf57c0faf, BX + ADDL R8, BX + ANDL BP, R9 + XORL CX, R9 + MOVL 20(DX), R8 + ADDL R9, BX + ROLL $0x07, BX + MOVL SI, R9 + ADDL BP, BX + XORL BP, R9 + ADDL $0x4787c62a, CX + ADDL R8, CX + ANDL BX, R9 + XORL SI, R9 + MOVL 24(DX), R8 + ADDL R9, CX + ROLL $0x0c, CX + MOVL BP, R9 + ADDL BX, CX + XORL BX, R9 + ADDL $0xa8304613, SI + ADDL R8, SI + ANDL CX, R9 + XORL BP, R9 + MOVL 28(DX), R8 + ADDL R9, SI + ROLL $0x11, SI + MOVL BX, R9 + ADDL CX, SI + XORL CX, R9 + ADDL $0xfd469501, BP + ADDL R8, BP + ANDL SI, R9 + XORL BX, R9 + MOVL 32(DX), R8 + ADDL R9, BP + ROLL $0x16, BP + MOVL CX, R9 + ADDL SI, BP + XORL SI, R9 + ADDL $0x698098d8, BX + ADDL R8, BX + ANDL BP, R9 + XORL CX, R9 + MOVL 36(DX), R8 + ADDL R9, BX + ROLL $0x07, BX + MOVL SI, R9 + ADDL BP, BX + XORL BP, R9 + ADDL $0x8b44f7af, CX + ADDL R8, CX + ANDL BX, R9 + XORL SI, R9 + MOVL 40(DX), R8 + ADDL R9, CX + ROLL $0x0c, CX + MOVL BP, R9 + ADDL BX, CX + XORL BX, R9 + ADDL $0xffff5bb1, SI + ADDL R8, SI + ANDL CX, R9 + XORL BP, R9 + MOVL 44(DX), R8 + ADDL R9, SI + ROLL $0x11, SI + MOVL BX, R9 + ADDL CX, SI + XORL CX, R9 + ADDL $0x895cd7be, BP + ADDL R8, BP + ANDL SI, R9 + XORL BX, R9 + MOVL 48(DX), R8 + ADDL R9, BP + ROLL $0x16, BP + MOVL CX, R9 + ADDL SI, BP + XORL SI, R9 + ADDL $0x6b901122, BX + ADDL R8, BX + ANDL BP, R9 + XORL CX, R9 + MOVL 52(DX), R8 + ADDL R9, BX + ROLL $0x07, BX + MOVL SI, R9 + ADDL BP, BX + XORL BP, R9 + ADDL $0xfd987193, CX + ADDL R8, CX + ANDL BX, R9 + XORL SI, R9 + MOVL 56(DX), R8 + ADDL R9, CX + ROLL $0x0c, CX + MOVL BP, R9 + ADDL BX, CX + XORL BX, R9 + ADDL $0xa679438e, SI + ADDL R8, SI + ANDL CX, R9 + XORL BP, R9 + MOVL 60(DX), R8 + ADDL R9, SI + ROLL $0x11, SI + MOVL BX, R9 + ADDL CX, SI + XORL CX, R9 + ADDL $0x49b40821, BP + ADDL R8, BP + ANDL SI, R9 + XORL BX, R9 + MOVL 4(DX), R8 + ADDL R9, BP + ROLL $0x16, BP + MOVL CX, R9 + ADDL SI, BP + + // ROUND2 + MOVL CX, R9 + MOVL CX, R14 + XORL DI, R9 + ADDL $0xf61e2562, BX + ADDL R8, BX + ANDL BP, R14 + ANDL SI, R9 + MOVL 24(DX), R8 + ORL R9, R14 + MOVL SI, R9 + ADDL R14, BX + MOVL SI, R14 + ROLL $0x05, BX + ADDL BP, BX + XORL DI, R9 + ADDL $0xc040b340, CX + ADDL R8, CX + ANDL BX, R14 + ANDL BP, R9 + MOVL 44(DX), R8 + ORL R9, R14 + MOVL BP, R9 + ADDL R14, CX + MOVL BP, R14 + ROLL $0x09, CX + ADDL BX, CX + XORL DI, R9 + ADDL $0x265e5a51, SI + ADDL R8, SI + ANDL CX, R14 + ANDL BX, R9 + MOVL (DX), R8 + ORL R9, R14 + MOVL BX, R9 + ADDL R14, SI + MOVL BX, R14 + ROLL $0x0e, SI + ADDL CX, SI + XORL DI, R9 + ADDL $0xe9b6c7aa, BP + ADDL R8, BP + ANDL SI, R14 + ANDL CX, R9 + MOVL 20(DX), R8 + ORL R9, R14 + MOVL CX, R9 + ADDL R14, BP + MOVL CX, R14 + ROLL $0x14, BP + ADDL SI, BP + XORL DI, R9 + ADDL $0xd62f105d, BX + ADDL R8, BX + ANDL BP, R14 + ANDL SI, R9 + MOVL 40(DX), R8 + ORL R9, R14 + MOVL SI, R9 + ADDL R14, BX + MOVL SI, R14 + ROLL $0x05, BX + ADDL BP, BX + XORL DI, R9 + ADDL $0x02441453, CX + ADDL R8, CX + ANDL BX, R14 + ANDL BP, R9 + MOVL 60(DX), R8 + ORL R9, R14 + MOVL BP, R9 + ADDL R14, CX + MOVL BP, R14 + ROLL $0x09, CX + ADDL BX, CX + XORL DI, R9 + ADDL $0xd8a1e681, SI + ADDL R8, SI + ANDL CX, R14 + ANDL BX, R9 + MOVL 16(DX), R8 + ORL R9, R14 + MOVL BX, R9 + ADDL R14, SI + MOVL BX, R14 + ROLL $0x0e, SI + ADDL CX, SI + XORL DI, R9 + ADDL $0xe7d3fbc8, BP + ADDL R8, BP + ANDL SI, R14 + ANDL CX, R9 + MOVL 36(DX), R8 + ORL R9, R14 + MOVL CX, R9 + ADDL R14, BP + MOVL CX, R14 + ROLL $0x14, BP + ADDL SI, BP + XORL DI, R9 + ADDL $0x21e1cde6, BX + ADDL R8, BX + ANDL BP, R14 + ANDL SI, R9 + MOVL 56(DX), R8 + ORL R9, R14 + MOVL SI, R9 + ADDL R14, BX + MOVL SI, R14 + ROLL $0x05, BX + ADDL BP, BX + XORL DI, R9 + ADDL $0xc33707d6, CX + ADDL R8, CX + ANDL BX, R14 + ANDL BP, R9 + MOVL 12(DX), R8 + ORL R9, R14 + MOVL BP, R9 + ADDL R14, CX + MOVL BP, R14 + ROLL $0x09, CX + ADDL BX, CX + XORL DI, R9 + ADDL $0xf4d50d87, SI + ADDL R8, SI + ANDL CX, R14 + ANDL BX, R9 + MOVL 32(DX), R8 + ORL R9, R14 + MOVL BX, R9 + ADDL R14, SI + MOVL BX, R14 + ROLL $0x0e, SI + ADDL CX, SI + XORL DI, R9 + ADDL $0x455a14ed, BP + ADDL R8, BP + ANDL SI, R14 + ANDL CX, R9 + MOVL 52(DX), R8 + ORL R9, R14 + MOVL CX, R9 + ADDL R14, BP + MOVL CX, R14 + ROLL $0x14, BP + ADDL SI, BP + XORL DI, R9 + ADDL $0xa9e3e905, BX + ADDL R8, BX + ANDL BP, R14 + ANDL SI, R9 + MOVL 8(DX), R8 + ORL R9, R14 + MOVL SI, R9 + ADDL R14, BX + MOVL SI, R14 + ROLL $0x05, BX + ADDL BP, BX + XORL DI, R9 + ADDL $0xfcefa3f8, CX + ADDL R8, CX + ANDL BX, R14 + ANDL BP, R9 + MOVL 28(DX), R8 + ORL R9, R14 + MOVL BP, R9 + ADDL R14, CX + MOVL BP, R14 + ROLL $0x09, CX + ADDL BX, CX + XORL DI, R9 + ADDL $0x676f02d9, SI + ADDL R8, SI + ANDL CX, R14 + ANDL BX, R9 + MOVL 48(DX), R8 + ORL R9, R14 + MOVL BX, R9 + ADDL R14, SI + MOVL BX, R14 + ROLL $0x0e, SI + ADDL CX, SI + XORL DI, R9 + ADDL $0x8d2a4c8a, BP + ADDL R8, BP + ANDL SI, R14 + ANDL CX, R9 + MOVL 20(DX), R8 + ORL R9, R14 + MOVL CX, R9 + ADDL R14, BP + MOVL CX, R14 + ROLL $0x14, BP + ADDL SI, BP + + // ROUND3 + MOVL SI, R9 + ADDL $0xfffa3942, BX + ADDL R8, BX + MOVL 32(DX), R8 + XORL CX, R9 + XORL BP, R9 + ADDL R9, BX + ROLL $0x04, BX + MOVL BP, R9 + ADDL BP, BX + ADDL $0x8771f681, CX + ADDL R8, CX + MOVL 44(DX), R8 + XORL SI, R9 + XORL BX, R9 + ADDL R9, CX + ROLL $0x0b, CX + MOVL BX, R9 + ADDL BX, CX + ADDL $0x6d9d6122, SI + ADDL R8, SI + MOVL 56(DX), R8 + XORL BP, R9 + XORL CX, R9 + ADDL R9, SI + ROLL $0x10, SI + MOVL CX, R9 + ADDL CX, SI + ADDL $0xfde5380c, BP + ADDL R8, BP + MOVL 4(DX), R8 + XORL BX, R9 + XORL SI, R9 + ADDL R9, BP + ROLL $0x17, BP + MOVL SI, R9 + ADDL SI, BP + ADDL $0xa4beea44, BX + ADDL R8, BX + MOVL 16(DX), R8 + XORL CX, R9 + XORL BP, R9 + ADDL R9, BX + ROLL $0x04, BX + MOVL BP, R9 + ADDL BP, BX + ADDL $0x4bdecfa9, CX + ADDL R8, CX + MOVL 28(DX), R8 + XORL SI, R9 + XORL BX, R9 + ADDL R9, CX + ROLL $0x0b, CX + MOVL BX, R9 + ADDL BX, CX + ADDL $0xf6bb4b60, SI + ADDL R8, SI + MOVL 40(DX), R8 + XORL BP, R9 + XORL CX, R9 + ADDL R9, SI + ROLL $0x10, SI + MOVL CX, R9 + ADDL CX, SI + ADDL $0xbebfbc70, BP + ADDL R8, BP + MOVL 52(DX), R8 + XORL BX, R9 + XORL SI, R9 + ADDL R9, BP + ROLL $0x17, BP + MOVL SI, R9 + ADDL SI, BP + ADDL $0x289b7ec6, BX + ADDL R8, BX + MOVL (DX), R8 + XORL CX, R9 + XORL BP, R9 + ADDL R9, BX + ROLL $0x04, BX + MOVL BP, R9 + ADDL BP, BX + ADDL $0xeaa127fa, CX + ADDL R8, CX + MOVL 12(DX), R8 + XORL SI, R9 + XORL BX, R9 + ADDL R9, CX + ROLL $0x0b, CX + MOVL BX, R9 + ADDL BX, CX + ADDL $0xd4ef3085, SI + ADDL R8, SI + MOVL 24(DX), R8 + XORL BP, R9 + XORL CX, R9 + ADDL R9, SI + ROLL $0x10, SI + MOVL CX, R9 + ADDL CX, SI + ADDL $0x04881d05, BP + ADDL R8, BP + MOVL 36(DX), R8 + XORL BX, R9 + XORL SI, R9 + ADDL R9, BP + ROLL $0x17, BP + MOVL SI, R9 + ADDL SI, BP + ADDL $0xd9d4d039, BX + ADDL R8, BX + MOVL 48(DX), R8 + XORL CX, R9 + XORL BP, R9 + ADDL R9, BX + ROLL $0x04, BX + MOVL BP, R9 + ADDL BP, BX + ADDL $0xe6db99e5, CX + ADDL R8, CX + MOVL 60(DX), R8 + XORL SI, R9 + XORL BX, R9 + ADDL R9, CX + ROLL $0x0b, CX + MOVL BX, R9 + ADDL BX, CX + ADDL $0x1fa27cf8, SI + ADDL R8, SI + MOVL 8(DX), R8 + XORL BP, R9 + XORL CX, R9 + ADDL R9, SI + ROLL $0x10, SI + MOVL CX, R9 + ADDL CX, SI + ADDL $0xc4ac5665, BP + ADDL R8, BP + MOVL (DX), R8 + XORL BX, R9 + XORL SI, R9 + ADDL R9, BP + ROLL $0x17, BP + MOVL SI, R9 + ADDL SI, BP + + // ROUND4 + MOVL DI, R9 + XORL CX, R9 + ADDL $0xf4292244, BX + ADDL R8, BX + ORL BP, R9 + XORL SI, R9 + ADDL R9, BX + MOVL 28(DX), R8 + MOVL DI, R9 + ROLL $0x06, BX + XORL SI, R9 + ADDL BP, BX + ADDL $0x432aff97, CX + ADDL R8, CX + ORL BX, R9 + XORL BP, R9 + ADDL R9, CX + MOVL 56(DX), R8 + MOVL DI, R9 + ROLL $0x0a, CX + XORL BP, R9 + ADDL BX, CX + ADDL $0xab9423a7, SI + ADDL R8, SI + ORL CX, R9 + XORL BX, R9 + ADDL R9, SI + MOVL 20(DX), R8 + MOVL DI, R9 + ROLL $0x0f, SI + XORL BX, R9 + ADDL CX, SI + ADDL $0xfc93a039, BP + ADDL R8, BP + ORL SI, R9 + XORL CX, R9 + ADDL R9, BP + MOVL 48(DX), R8 + MOVL DI, R9 + ROLL $0x15, BP + XORL CX, R9 + ADDL SI, BP + ADDL $0x655b59c3, BX + ADDL R8, BX + ORL BP, R9 + XORL SI, R9 + ADDL R9, BX + MOVL 12(DX), R8 + MOVL DI, R9 + ROLL $0x06, BX + XORL SI, R9 + ADDL BP, BX + ADDL $0x8f0ccc92, CX + ADDL R8, CX + ORL BX, R9 + XORL BP, R9 + ADDL R9, CX + MOVL 40(DX), R8 + MOVL DI, R9 + ROLL $0x0a, CX + XORL BP, R9 + ADDL BX, CX + ADDL $0xffeff47d, SI + ADDL R8, SI + ORL CX, R9 + XORL BX, R9 + ADDL R9, SI + MOVL 4(DX), R8 + MOVL DI, R9 + ROLL $0x0f, SI + XORL BX, R9 + ADDL CX, SI + ADDL $0x85845dd1, BP + ADDL R8, BP + ORL SI, R9 + XORL CX, R9 + ADDL R9, BP + MOVL 32(DX), R8 + MOVL DI, R9 + ROLL $0x15, BP + XORL CX, R9 + ADDL SI, BP + ADDL $0x6fa87e4f, BX + ADDL R8, BX + ORL BP, R9 + XORL SI, R9 + ADDL R9, BX + MOVL 60(DX), R8 + MOVL DI, R9 + ROLL $0x06, BX + XORL SI, R9 + ADDL BP, BX + ADDL $0xfe2ce6e0, CX + ADDL R8, CX + ORL BX, R9 + XORL BP, R9 + ADDL R9, CX + MOVL 24(DX), R8 + MOVL DI, R9 + ROLL $0x0a, CX + XORL BP, R9 + ADDL BX, CX + ADDL $0xa3014314, SI + ADDL R8, SI + ORL CX, R9 + XORL BX, R9 + ADDL R9, SI + MOVL 52(DX), R8 + MOVL DI, R9 + ROLL $0x0f, SI + XORL BX, R9 + ADDL CX, SI + ADDL $0x4e0811a1, BP + ADDL R8, BP + ORL SI, R9 + XORL CX, R9 + ADDL R9, BP + MOVL 16(DX), R8 + MOVL DI, R9 + ROLL $0x15, BP + XORL CX, R9 + ADDL SI, BP + ADDL $0xf7537e82, BX + ADDL R8, BX + ORL BP, R9 + XORL SI, R9 + ADDL R9, BX + MOVL 44(DX), R8 + MOVL DI, R9 + ROLL $0x06, BX + XORL SI, R9 + ADDL BP, BX + ADDL $0xbd3af235, CX + ADDL R8, CX + ORL BX, R9 + XORL BP, R9 + ADDL R9, CX + MOVL 8(DX), R8 + MOVL DI, R9 + ROLL $0x0a, CX + XORL BP, R9 + ADDL BX, CX + ADDL $0x2ad7d2bb, SI + ADDL R8, SI + ORL CX, R9 + XORL BX, R9 + ADDL R9, SI + MOVL 36(DX), R8 + MOVL DI, R9 + ROLL $0x0f, SI + XORL BX, R9 + ADDL CX, SI + ADDL $0xeb86d391, BP + ADDL R8, BP + ORL SI, R9 + XORL CX, R9 + ADDL R9, BP + ROLL $0x15, BP + ADDL SI, BP + ADDL R10, BX + ADDL R11, BP + ADDL R12, SI + ADDL R13, CX + + // Prepare next loop + ADDQ $0x40, DX + CMPQ DX, AX + JB loop + + // Write output + MOVQ dig+0(FP), AX + MOVL BX, (AX) + MOVL BP, 4(AX) + MOVL SI, 8(AX) + MOVL CX, 12(AX) + +end: + RET |