VMRGEW V4, V5, V29
	VMRGEW V6, V7, V30

	XXPERMDI VS32, VS34, $0, VS33
	XXPERMDI VS32, VS34, $3, VS35
	XXPERMDI VS59, VS60, $0, VS32
	XXPERMDI VS59, VS60, $3, VS34

	VMRGOW V4, V5, V4
	VMRGOW V6, V7, V6

	VMRGEW V8, V9, V27
	VMRGEW V10, V11, V28

	XXPERMDI VS36, VS38, $0, VS37
	XXPERMDI VS36, VS38, $3, VS39
	XXPERMDI VS61, VS62, $0, VS36
	XXPERMDI VS61, VS62, $3, VS38

	VMRGOW V8, V9, V8

	LXVD2X (R16)(x_ptr), X1

	XXPERMDI X0, X0, $2, X0
	XXPERMDI X1, X1, $2, X1

	LXVD2X (R0)(y_ptr), Y0
	LXVD2X (R16)(y_ptr), Y1

	XXPERMDI Y0, Y0, $2, Y0
	XXPERMDI Y1, Y1, $2, Y1

	LXVD2X (R16)(CPOOL), P1
	LXVD2X (R0)(CPOOL), P0

	CALL p256MulInternal<>(SB)

	MOVD $p256mul<>+0x00(SB), CPOOL

	XXPERMDI T0, T0, $2, T0
	XXPERMDI T1, T1, $2, T1
	STXVD2X T0, (R0)(res_ptr)

	"XXLANDC",
	"XXLEQV",
	"XXLNAND",
	"XXLOR",
	"XXLORC",
	"XXLNOR",
	"XXLORQ",
	"XXLXOR",
	"XXSEL",
	"XXMRGHW",
	"XXMRGLW",
	"XXSPLTW",
	"XXSPLTIB",
	"XXPERM",
	"XXPERMDI",
	"XXSLDWI",
	"XXBRQ",
	"XXBRD",
	"XXBRW",
	"XXBRH",
	"XSCVDPSP",
	"XSCVSPDP",
	"XSCVDPSPN",
	"XSCVSPDPN",
	"XVCVDPSP",
	"XVCVSPDP",
	"XSCVDPSXDS",
	"XSCVDPSXWS",

#ifndef GOPPC64_power9
	VBPERMQ	V6,V0,V6
	MFVRD	V6,R4
	CNTLZW	R4,R4
#else
#ifdef GOARCH_ppc64le
	// Put the value back into LE ordering by swapping doublewords.
	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]

	XXSLDWI V1, V2, $1, V3          // f0611117
	XXSLDWI V1, $1, V2, V3          // f0611117
	XXSLDWI VS33, VS34, $1, VS35    // f0611117
	XXSLDWI VS33, $1, VS34, VS35    // f0611117
	XXPERMDI VS33, VS34, $1, VS35   // f0611157
	XXPERMDI VS33, $1, VS34, VS35   // f0611157
	XSCVDPSP VS1, VS2               // f0400c24
	XVCVDPSP VS1, VS2               // f0400e24
	XSCVSXDDP VS1, VS2              // f0400de0

	XXLAND:         "xxland",
	XXLANDC:        "xxlandc",
	XXLNOR:         "xxlnor",
	XXLOR:          "xxlor",
	XXLXOR:         "xxlxor",
	XXMRGHW:        "xxmrghw",
	XXMRGLW:        "xxmrglw",
	XXPERMDI:       "xxpermdi",
	XXSEL:          "xxsel",
	XXSLDWI:        "xxsldwi",
	XXSPLTW:        "xxspltw",
	CMPB:           "cmpb",
	DADD:           "dadd",
	DADDCC:         "dadd.",
	DADDQ:          "daddq",

			opset(AXXMRGLW, r0)

		case AXXSPLTW: /* xxspltw */
			opset(AXXSPLTW, r0)

		case AXXSPLTIB: /* xxspltib */
			opset(AXXSPLTIB, r0)

		case AXXPERM: /* xxpermdi */
			opset(AXXPERM, r0)

		case AXXSLDWI: /* xxsldwi */
			opset(AXXPERMDI, r0)
			opset(AXXSLDWI, r0)

		case AXXBRQ: /* xxbrq, xxbrd, xxbrw, xxbrh */
			opset(AXXBRD, r0)