VCLRRB V1, R2, V3                       // 106111cd
	VCLZDM V1, V2, V3                       // 10611784
	VCMPEQUQ V1, V2, V3                     // 106111c7
	VCMPEQUQCC V1, V2, V3                   // 106115c7
	VCMPGTSQ V1, V2, V3                     // 10611387
	VCMPGTSQCC V1, V2, V3                   // 10611787
	VCMPGTUQ V1, V2, V3                     // 10611287
	VCMPGTUQCC V1, V2, V3                   // 10611687

	VCFUX:          "vcfux",
	VCMPBFP:        "vcmpbfp",
	VCMPBFPCC:      "vcmpbfp.",
	VCMPEQFP:       "vcmpeqfp",
	VCMPEQFPCC:     "vcmpeqfp.",
	VCMPEQUB:       "vcmpequb",
	VCMPEQUBCC:     "vcmpequb.",
	VCMPEQUH:       "vcmpequh",
	VCMPEQUHCC:     "vcmpequh.",
	VCMPEQUW:       "vcmpequw",
	VCMPEQUWCC:     "vcmpequw.",
	VCMPGEFP:       "vcmpgefp",
	VCMPGEFPCC:     "vcmpgefp.",
	VCMPGTFP:       "vcmpgtfp",

	"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",
	"VCMPGTUW",
	"VCMPGTUWCC",
	"VCMPGTUD",

	"VDIVEUQ",
	"VDIVEUD",
	"VDIVESW",
	"VDIVESQ",
	"VDIVESD",
	"VCTZDM",
	"VCNTMBW",
	"VCNTMBH",
	"VCNTMBD",
	"VCNTMBB",
	"VCMPUQ",
	"VCMPSQ",
	"VCMPGTUQCC",
	"VCMPGTUQ",
	"VCMPGTSQCC",
	"VCMPGTSQ",
	"VCMPEQUQCC",
	"VCMPEQUQ",
	"VCLZDM",
	"VCLRRB",
	"VCLRLB",
	"VCFUGED",
	"STXVRWX",
	"STXVRHX",
	"STXVRDX",
	"STXVRBX",
	"STXVPX",
	"STXVP",

	VSLDOI   $2, V2, V3, V9     // V9=(V2:V3)<<2
	VSLDOI   $3, V2, V3, V10    // V10=(V2:v3)<<3
	VCMPEQUW V1, V2, V5         // compare index 0, 4, ... with sep
	VCMPEQUW V1, V4, V6         // compare index 1, 5, ... with sep
	VCMPEQUW V1, V9, V11        // compare index 2, 6, ... with sep
	VCMPEQUW V1, V10, V12       // compare index 3, 7, ... with sep
	VSEL     V6, V5, V29, V13   // merge index 0, 1, 4, 5, using mask

	AVSOI
	AVSLDOI
	AVCLZ
	AVCLZB
	AVCLZH
	AVCLZW
	AVCLZD
	AVPOPCNT
	AVPOPCNTB
	AVPOPCNTH
	AVPOPCNTW
	AVPOPCNTD
	AVCMPEQ
	AVCMPEQUB
	AVCMPEQUBCC
	AVCMPEQUH
	AVCMPEQUHCC
	AVCMPEQUW
	AVCMPEQUWCC
	AVCMPEQUD
	AVCMPEQUDCC
	AVCMPGT
	AVCMPGTUB
	AVCMPGTUBCC
	AVCMPGTUH
	AVCMPGTUHCC
	AVCMPGTUW
	AVCMPGTUWCC
	AVCMPGTUD
	AVCMPGTUDCC
	AVCMPGTSB

	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.

	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:
		return OPVC(4, 134, 0, 0) /* vcmpequw - v2.03 */
	case AVCMPEQUWCC:

	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
	VCMPEQUWCC V1, V2, V3           // 10611486
	VCMPEQUD V1, V2, V3             // 106110c7
	VCMPEQUDCC V1, V2, V3           // 106114c7

	BR	notfound


cmp8:	// Length 8 - 15
#ifdef GOPPC64_power10
	// Load all the bytes into a single VSR in BE order.
	SLD	$56,R4,R5
	LXVLL	R3,R5,V2
	// Compare and count the number which don't match.
	VCMPEQUB	V2,V1,V6
	VCLZLSBB	V6,R3
	// If count is the number of bytes, or more. No matches are found.
	CMPU	R3,R4
	MOVD	$-1,R5
	// Otherwise, the count is the index of the first match.
	ISEL	CR0LT,R3,R5,R3
	RET