XXPERMDI	V6,V6,$2,V6
#endif
	_VCZBEBB	V6,R4
#endif
	ADD	R3,R4,R3
	RET

cmp16:	// Length 16 - 31
	CMPU	R4,$16
	ADD	R4,R3,R9
	BLT	cmp8

	ADD	$-16,R9,R9		// &s[len(s)-16]

	// Bytes 0 - 15
	LXVD2X	(R0)(R3),V2
	VCMPEQUBCC	V2,V1,V6
	MOVD	R3,R8
	BNE	CR6,foundat0		// Match found at R8+32 bytes, jump out

	BEQ	notfound

GLOBL byteswap<>+0(SB), RODATA, $16
#define SWAP V21
#endif

TEXT cmpbody<>(SB),NOSPLIT|NOFRAME,$0-0
start:
	CMP	R9,$16,CR0
	CMP	R9,$32,CR1
	CMP	R9,$64,CR2
	MOVD	$16,R10
	BLT	cmp8
	BLT	CR1,cmp16
	BLT	CR2,cmp32

cmp64:	// >= 64B
	DCBT	(R5)		// optimize for size>=64
	DCBT	(R6)		// cache hint

	SRD	$6,R9,R14	// There is at least one iteration.
	MOVD	R14,CTR
	ANDCC   $63,R9,R9