"VNAND",
	"VOR",
	"VORC",
	"VNOR",
	"VXOR",
	"VEQV",
	"VADDUM",
	"VADDUBM",
	"VADDUHM",
	"VADDUWM",
	"VADDUDM",
	"VADDUQM",
	"VADDCU",
	"VADDCUQ",
	"VADDCUW",
	"VADDUS",
	"VADDUBS",
	"VADDUHS",
	"VADDUWS",
	"VADDSS",
	"VADDSBS",
	"VADDSHS",
	"VADDSWS",
	"VADDE",
	"VADDEUQM",
	"VADDECUQ",
	"VSUBUM",
	"VSUBUBM",
	"VSUBUHM",
	"VSUBUWM",

	VSPLTISW $8, V29
	VSPLTISW $7, V30
	PCALIGN $16
loop_vsx:
	VADDUWM V0, V4, V0
	VADDUWM V1, V5, V1
	VADDUWM V2, V6, V2
	VADDUWM V3, V7, V3

	VPERMXOR V12, V0, V21, V12
	VPERMXOR V13, V1, V21, V13
	VPERMXOR V14, V2, V21, V14
	VPERMXOR V15, V3, V21, V15

	VADDUWM V8, V12, V8
	VADDUWM V9, V13, V9
	VADDUWM V10, V14, V10
	VADDUWM V11, V15, V11

	VXOR V4, V8, V4
	VXOR V5, V9, V5

	VCMPEQUB V0, V1, V0
	VPOPCNTD V2, V2		// A match is 0xFF or 0. Count the bits into doubleword buckets.
	VPOPCNTD V0, V0
	VADDUDM	V0, V4, V4	// Accumulate the popcounts. They are 8x the count.
	VADDUDM	V2, V5, V5	// The count will be fixed up afterwards.
	ADD	$32, R3
	BDNZ	cmploop

	VADDUDM	V4, V5, V5
	MFVSRD	V5, R18
	VSLDOI	$8, V5, V5, V5
	MFVSRD	V5, R21
	ADD	R21, R18, R18
	ANDCC	$31, R4, R4