"STXVW4X",
	"STXVH8X",
	"STXVB16X",
	"STXVX",
	"LXSDX",
	"STXSDX",
	"LXSIWAX",
	"LXSIWZX",
	"STXSIWX",
	"MFVSRD",
	"MFFPRD",
	"MFVRD",
	"MFVSRWZ",
	"MFVSRLD",
	"MTVSRD",
	"MTFPRD",
	"MTVRD",
	"MTVSRWA",
	"MTVSRWZ",
	"MTVSRDD",
	"MTVSRWS",
	"XXLAND",
	"XXLANDC",
	"XXLEQV",
	"XXLNAND",
	"XXLOR",
	"XXLORC",
	"XXLNOR",
	"XXLORQ",
	"XXLXOR",

	VADDUDM	V0, V4, V4	// Accumulate the popcounts. They are 8x the count.
	VADDUDM	V2, V5, V5	// The count will be fixed up afterwards.
	ADD	$32, R3
	BDNZ	cmploop

	VADDUDM	V4, V5, V5
	MFVSRD	V5, R18
	VSLDOI	$8, V5, V5, V5
	MFVSRD	V5, R21
	ADD	R21, R18, R18
	ANDCC	$31, R4, R4
	// Skip the tail processing if no bytes remaining.
	BEQ	tail_0

#ifdef GOPPC64_power10

	VPERM	V3,V3,SWAP,V3
	VPERM	V4,V4,SWAP,V4
#endif

	MFVSRD	VS35,R16	// move upper doublewords of A and B into GPR for comparison
	MFVSRD	VS36,R10

	CMPU	R16,R10
	BEQ	lower
	SETB_CR0_NE(R3)
	RET

	PCALIGN $16
lower:
	VSLDOI	$8,V3,V3,V3	// move lower doublewords of A and B into GPR for comparison
	MFVSRD	VS35,R16
	VSLDOI	$8,V4,V4,V4
	MFVSRD	VS36,R10

	CMPU	R16,R10
	SETB_CR0_NE(R3)
	RET

func fnma(x, y, z float64) float64 {
	// riscv64:"FNMADDD",-"FNMSUBD"
	return math.FMA(x, -y, -z)
}

func fromFloat64(f64 float64) uint64 {
	// amd64:"MOVQ\tX.*, [^X].*"
	// arm64:"FMOVD\tF.*, R.*"
	// ppc64x:"MFVSRD"
	// mips64/hardfloat:"MOVV\tF.*, R.*"
	return math.Float64bits(f64+1) + 1
}

func fromFloat32(f32 float32) uint32 {
	// amd64:"MOVL\tX.*, [^X].*"
	// arm64:"FMOVS\tF.*, R.*"

	BR	vfound
foundat0:
	SUB	R3,R8,R3
	ADD	$0+ADJUST_FOR_CNTLZW,R3
vfound:
	// Map equal values into a 16 bit value with earlier matches setting higher bits.
#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