ivlo+32(FP), SI MOVQ ivhi+40(FP), DI MOVOU bswapMask<>+0(SB), X0 MOVQ SI, X1 PINSRQ $0x01, DI, X1 MOVAPS X1, X8 PSHUFB X0, X1 MOVQ SI, R8 ADDQ $0x07, R8 JC ctr8_slow XORQ R8, R8 INCQ R8 PXOR X9, X9 PINSRQ $0x00, R8, X9 PADDQ X9, X8 MOVAPS X8, X2 PADDQ X9, X8 MOVAPS X8, X3 PADDQ X9, X8 MOVAPS X8, X4 PADDQ X9, X8 MOVAPS X8, X5 PADDQ X9, X8 MOVAPS X8, X6 PADDQ X9, X8 MOVAPS X8, X7 PADDQ X9, X8 MOVAPS X8, X8 JMP ctr8_done ctr8_slow: ADDQ $0x01, SI ADCQ $0x00, DI MOVQ SI, X2 PINSRQ $0x01, DI, X2 ADDQ $0x01, SI...