MOVOU (8*16 + 1*16)(SP), B1
		MOVOU (8*16 + 2*16)(SP), B2
		MOVOU (8*16 + 3*16)(SP), B3
		MOVOU (8*16 + 4*16)(SP), B4
		MOVOU (8*16 + 5*16)(SP), B5
		MOVOU (8*16 + 6*16)(SP), B6
		MOVOU (8*16 + 7*16)(SP), B7

		MOVOU (16*0)(SP), T0
		PSHUFD $78, T0, T1
		PXOR T0, T1

		MOVOU (16*0)(pTbl), ACC0
		MOVOU (16*1)(pTbl), ACCM
		MOVOU ACC0, ACC1

		PCLMULQDQ $0x00, T1, ACCM
		PCLMULQDQ $0x00, T0, ACC0

	MOVQ zero+40(FP), t2

	MOVOU (16*0)(BX), X0
	MOVOU (16*1)(BX), X1
	MOVOU (16*2)(BX), X2
	MOVOU (16*3)(BX), X3
	MOVOU (16*4)(BX), X4
	MOVOU (16*5)(BX), X5

	MOVOU X0, x1in(16*0)
	MOVOU X1, x1in(16*1)
	MOVOU X2, y1in(16*0)
	MOVOU X3, y1in(16*1)
	MOVOU X4, z1in(16*0)
	MOVOU X5, z1in(16*1)

	MOVOU (16*0)(CX), X0
	MOVOU (16*1)(CX), X1

	MOVOU X0, x2in(16*0)

	AESENC	X3, X3
	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

	// load data
	MOVOU	(AX), X8
	MOVOU	16(AX), X9
	MOVOU	32(AX), X10
	MOVOU	48(AX), X11
	MOVOU	-64(AX)(CX*1), X12
	MOVOU	-48(AX)(CX*1), X13
	MOVOU	-32(AX)(CX*1), X14
	MOVOU	-16(AX)(CX*1), X15

	// xor with seed
	PXOR	X0, X8
	PXOR	X1, X9
	PXOR	X2, X10
	PXOR	X3, X11
	PXOR	X4, X12

	PXOR	runtime·aeskeysched+16(SB), X1
	PXOR	runtime·aeskeysched+32(SB), X2
	PXOR	runtime·aeskeysched+48(SB), X3
	AESENC	X1, X1
	AESENC	X2, X2
	AESENC	X3, X3

	MOVOU	(AX), X4
	MOVOU	16(AX), X5
	MOVOU	-32(AX)(BX*1), X6
	MOVOU	-16(AX)(BX*1), X7

	PXOR	X0, X4
	PXOR	X1, X5
	PXOR	X2, X6
	PXOR	X3, X7

	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

				// MOVSD, SQRTSD, etc
				src = 8
				break
			}
			if strings.HasSuffix(op, "SS") {
				// MOVSS, SQRTSS, etc
				src = 4
				break
			}
			if op == "MOVO" || op == "MOVOU" {
				src = 16
				break
			}
			if strings.HasPrefix(op, "SET") {
				// SETEQ, etc
				src = 1
				break
			}
			switch op[len(op)-1] {
			case 'B':
				src = 1
			case 'W':