VXOR     LEMASK, T0, LEMASK
	VPERM    XL, XL, LEMASK, XL
#endif
	VXOR     ZERO, ZERO, ZERO

	CMPU LEN, $64
	BGE  gcm_ghash_p8_4x

	LXVD2X (INP)(R0), VIN
	ADD    $16, INP, INP
	SUBCCC $16, LEN, LEN
#ifdef GOARCH_ppc64le
	VPERM  IN, IN, LEMASK, IN
#endif
	VXOR   IN, XL, IN
	BEQ    short

	LXVD2X (HTBL)(R8), VH2L  // load H^2
	MOVD   $16, R8
	LXVD2X (HTBL)(R9), VH2