MOVD	R20, CTR
	MOVD	$16, R21
	XXLXOR	V4, V4, V4
	XXLXOR	V5, V5, V5

	PCALIGN	$16
cmploop:
	LXVD2X	(R0)(R3), V0	// Count 32B per loop with two vector accumulators.
	LXVD2X	(R21)(R3), V2
	VCMPEQUB V2, V1, V2
	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.

	"VSRAW",
	"VSRAD",
	"VSOI",
	"VSLDOI",
	"VCLZ",
	"VCLZB",
	"VCLZH",
	"VCLZW",
	"VCLZD",
	"VPOPCNT",
	"VPOPCNTB",
	"VPOPCNTH",
	"VPOPCNTW",
	"VPOPCNTD",
	"VCMPEQ",
	"VCMPEQUB",
	"VCMPEQUBCC",
	"VCMPEQUH",
	"VCMPEQUHCC",
	"VCMPEQUW",
	"VCMPEQUWCC",
	"VCMPEQUD",
	"VCMPEQUDCC",
	"VCMPGT",
	"VCMPGTUB",
	"VCMPGTUBCC",
	"VCMPGTUH",
	"VCMPGTUHCC",

	VPOPCNTB V1, V2                 // 10400f03
	VPOPCNTH V1, V2                 // 10400f43
	VPOPCNTW V1, V2                 // 10400f83
	VPOPCNTD V1, V2                 // 10400fc3
	VCMPEQUB V1, V2, V3             // 10611006
	VCMPEQUBCC V1, V2, V3           // 10611406
	VCMPEQUH V1, V2, V3             // 10611046
	VCMPEQUHCC V1, V2, V3           // 10611446
	VCMPEQUW V1, V2, V3             // 10611086

		return OPVX(4, 1923, 0, 0) /* vpopcntw - v2.07 */
	case AVPOPCNTD:
		return OPVX(4, 1987, 0, 0) /* vpopcntd - v2.07 */

	case AVCMPEQUB:
		return OPVC(4, 6, 0, 0) /* vcmpequb - v2.03 */
	case AVCMPEQUBCC:
		return OPVC(4, 6, 0, 1) /* vcmpequb. - v2.03 */
	case AVCMPEQUH:
		return OPVC(4, 70, 0, 0) /* vcmpequh - v2.03 */
	case AVCMPEQUHCC:
		return OPVC(4, 70, 0, 1) /* vcmpequh. - v2.03 */
	case AVCMPEQUW: