MULLD  h2, r1, t3;  \
	ADDZE  t4, h0;      \
	MULHDU r1, h1, t5;  \
	MULLD  r1, h1, t4;  \
	ADDC   t4, t2, t2;  \
	ADDE   t5, t3, t3;  \
	ADDC   h0, t2, t2;  \
	MOVD   $-4, t4;     \
	ADDZE  t3;          \
	RLDICL $0, t2, $62, h2; \
	AND    t2, t4, h0;  \
	ADDC   t0, h0, h0;  \
	ADDE   t3, t1, h1;  \
	SLD    $62, t3, t4; \
	SRD    $2, t2;      \
	ADDZE  h2;          \
	OR     t4, t2, t2;  \
	SRD    $2, t3;      \

	BLT	cmp16		// Jump to the small string case if it's <32 bytes.

	CMP	R4,$64,CR1
	MOVD	$16,R11
	MOVD	R3,R8
	BLT	CR1,cmp32	// Special case for length 32 - 63
	MOVD	$32,R12
	MOVD	$48,R6

	RLDICR  $0,R4,$63-6,R9	// R9 = len &^ 63
	ADD	R3,R9,R9	// R9 = &s[len &^ 63]
	ANDCC	$63,R4		// (len &= 63) cmp 0.

	PCALIGN	$16
loop64:
	LXVD2X	(R0)(R8),V2	// Scan 64 bytes at a time, starting at &s[0]
	VCMPEQUBCC	V2,V1,V6