MOVOU y1in(16*0), X11
	MOVOU y1in(16*1), X12
	MOVOU z1in(16*0), X13
	MOVOU z1in(16*1), X14

	PAND X15, X0
	PAND X15, X1
	PAND X15, X2
	PAND X15, X3
	PAND X15, X4
	PAND X15, X5

	PAND X6, X9
	PAND X6, X10
	PAND X6, X11
	PAND X6, X12
	PAND X6, X13
	PAND X6, X14

	PXOR X9, X0
	PXOR X10, X1
	PXOR X11, X2
	PXOR X12, X3
	PXOR X13, X4
	PXOR X14, X5

	MOVOU (16*13)(KS), T0
	AESENC T0, B0
	MOVOU (16*14)(KS), T0
initEncLast:
	AESENCLAST T0, B0

	PSHUFB BSWAP, B0
	// H * 2
	PSHUFD $0xff, B0, T0
	MOVOU B0, T1
	PSRAL $31, T0
	PAND POLY, T0
	PSRLL $31, T1
	PSLLDQ $4, T1
	PSLLL $1, B0
	PXOR T0, B0
	PXOR T1, B0
	// Karatsuba pre-computations
	MOVOU B0, (16*14)(dst)
	PSHUFD $78, B0, B1
	PXOR B0, B1
	MOVOU B1, (16*15)(dst)

	"PABSD",
	"PABSW",
	"PACKSSLW",
	"PACKSSWB",
	"PACKUSDW",
	"PACKUSWB",
	"PADDB",
	"PADDL",
	"PADDQ",
	"PADDSB",
	"PADDSW",
	"PADDUSB",
	"PADDUSW",
	"PADDW",
	"PALIGNR",
	"PAND",
	"PANDN",
	"PAUSE",
	"PAVGB",
	"PAVGW",
	"PBLENDVB",
	"PBLENDW",
	"PCLMULQDQ",
	"PCMPEQB",
	"PCMPEQL",
	"PCMPEQQ",
	"PCMPEQW",
	"PCMPESTRI",
	"PCMPESTRM",
	"PCMPGTB",

	JE	aes0

	ADDL	$16, AX
	TESTW	$0xff0, AX
	JE	endofpage

	// 16 bytes loaded at this address won't cross
	// a page boundary, so we can load it directly.
	MOVOU	-16(AX), X1
	ADDL	BX, BX
	PAND	masks<>(SB)(BX*8), X1

final1:
	PXOR	X0, X1	// xor data with seed
	AESENC	X1, X1  // scramble combo 3 times
	AESENC	X1, X1
	AESENC	X1, X1
	MOVL	X1, (DX)
	RET

endofpage: