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

	"MOVLPD",
	"MOVLPS",
	"MOVLQSX",
	"MOVLQZX",
	"MOVMSKPD",
	"MOVMSKPS",
	"MOVNTDQA",
	"MOVNTIL",
	"MOVNTIQ",
	"MOVNTO",
	"MOVNTPD",
	"MOVNTPS",
	"MOVNTQ",
	"MOVO",
	"MOVOU",
	"MOVQ",
	"MOVQL",
	"MOVQOZX",
	"MOVSB",
	"MOVSD",
	"MOVSHDUP",
	"MOVSL",
	"MOVSLDUP",
	"MOVSQ",
	"MOVSS",
	"MOVSW",
	"MOVSWW",
	"MOVUPD",
	"MOVUPS",
	"MOVW",

				// 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':