tail:	// Count the last 0 - 31 bytes.
	CMP	R4, $16
	BLE	small_tail_p10
	LXV	0(R3), V0
	VCMPEQUB V0, V1, V0
	VCNTMBB	V0, $1, R14	// Sum the value of bit 0 of each byte of the compare into R14.
	SRD	$56, R14, R14	// The result of VCNTMBB is shifted. Unshift it.
	ADD	R14, R18, R18
	ADD	$16, R3, R3
	ANDCC	$15, R4, R4

small_tail_p10:
	SLD	$56, R4, R6
	LXVLL	R3, R6, V0
	VCMPEQUB V0, V1, V0

	VCMPGTUQ V1, V2, V3                     // 10611287
	VCMPGTUQCC V1, V2, V3                   // 10611687
	VCMPSQ V1, V2, CR2                      // 11011141
	VCMPUQ V1, V2, CR3                      // 11811101
	VCNTMBB V1, $1, R3                      // 10790e42
	VCNTMBD V1, $1, R3                      // 107f0e42
	VCNTMBH V1, $1, R3                      // 107b0e42
	VCNTMBW V1, $1, R3                      // 107d0e42

	"VDIVUD",
	"VDIVSW",
	"VDIVSQ",
	"VDIVSD",
	"VDIVEUW",
	"VDIVEUQ",
	"VDIVEUD",
	"VDIVESW",
	"VDIVESQ",
	"VDIVESD",
	"VCTZDM",
	"VCNTMBW",
	"VCNTMBH",
	"VCNTMBD",
	"VCNTMBB",
	"VCMPUQ",
	"VCMPSQ",
	"VCMPGTUQCC",
	"VCMPGTUQ",
	"VCMPGTSQCC",
	"VCMPGTSQ",
	"VCMPEQUQCC",
	"VCMPEQUQ",
	"VCLZDM",
	"VCLRRB",
	"VCLRLB",
	"VCFUGED",
	"STXVRWX",

	VCMPGTSQ:       "vcmpgtsq",
	VCMPGTSQCC:     "vcmpgtsq.",
	VCMPGTUQ:       "vcmpgtuq",
	VCMPGTUQCC:     "vcmpgtuq.",
	VCMPSQ:         "vcmpsq",
	VCMPUQ:         "vcmpuq",
	VCNTMBB:        "vcntmbb",
	VCNTMBD:        "vcntmbd",
	VCNTMBH:        "vcntmbh",
	VCNTMBW:        "vcntmbw",
	VCTZDM:         "vctzdm",
	VDIVESD:        "vdivesd",
	VDIVESQ:        "vdivesq",