48(BP), R14 MOVQ R14, 48(BP) ADDQ 56(BP), R15 MOVQ R15, 56(BP) ADDQ $0x80, SI CMPQ SI, 640(SP) JB loop end: RET DATA PSHUFFLE_BYTE_FLIP_MASK<>+0(SB)/8, $0x0001020304050607 DATA PSHUFFLE_BYTE_FLIP_MASK<>+8(SB)/8, $0x08090a0b0c0d0e0f DATA PSHUFFLE_BYTE_FLIP_MASK<>+16(SB)/8, $0x1011121314151617 DATA PSHUFFLE_BYTE_FLIP_MASK<>+24(SB)/8, $0x18191a1b1c1d1e1f GLOBL PSHUFFLE_BYTE_FLIP_MASK<>(SB), RODATA|NOPTR, $32 // func blockAVX2(dig *Digest, p []byte) // Requires: AVX, AVX2, BMI2 TEXT ·blockAVX2(SB), NOSPLIT,...