MOVD.W	-8(R2), R6
	LSR	R4, R6, R5	// return value
	LSL	R3, R6, R8	// x[i] << s
	SUB	$1, R1
one:	TBZ	$0, R1, two
	MOVD.W	-8(R2), R6
	LSR	R4, R6, R7
	ORR	R8, R7
	LSL	R3, R6, R8
	SUB	$1, R1
	MOVD.W	R7, -8(R0)
two:
	TBZ	$1, R1, loop
	LDP.W	-16(R2), (R6, R7)
	LSR	R4, R7, R10
	ORR	R8, R10
	LSL	R3, R7
	LSR	R4, R6, R9
	ORR	R7, R9
	LSL	R3, R6, R8
	SUB	$2, R1

	MUL	R1, R0

	// Code stolen from compiler output for:
	//
	//	var x uint64
	//	func f() (sec uint64, nsec uint32) { return x / 1000000000, uint32(x % 1000000000) }
	//
	LSR	$1, R0, R1
	MOVD	$-8543223759426509416, R2
	UMULH	R1, R2, R1
	LSR	$28, R1, R1
	MOVD	R1, sec+0(FP)
	MOVD	$1000000000, R2
	MSUB	R1, R0, R2, R0
	MOVW	R0, nsec+8(FP)

// license that can be found in the LICENSE file.

//go:build !purego

#include "textflag.h"

// carryPropagate works exactly like carryPropagateGeneric and uses the
// same AND, ADD, and LSR+MADD instructions emitted by the compiler, but
// avoids loading R0-R4 twice and uses LDP and STP.
//
// See https://golang.org/issues/43145 for the main compiler issue.
//
// func carryPropagate(v *Element)

func mask32(x uint32) uint32 {
	return (x << 29) >> 29 // arm64:"AND\t[$]7, R[0-9]+",-"LSR",-"LSL"
}

// check 16-bit shift masking
func mask16(x uint16) uint16 {
	return (x << 14) >> 14 // arm64:"AND\t[$]3, R[0-9]+",-"LSR",-"LSL"
}

// check 8-bit shift masking
func mask8(x uint8) uint8 {
	return (x << 7) >> 7 // arm64:"AND\t[$]1, R[0-9]+",-"LSR",-"LSL"
}

func maskshift(x uint64) uint64 {

	if shift >= 0 && shift < 64 {
		// arm64:"LSR",-"CSEL"
		r1 = val64 >> shift
	}
	if shift >= 0 && shift < 32 {
		// arm64:"LSR",-"CSEL"
		r2 = val32 >> shift
	}
	if shift >= 0 && shift < 16 {
		// arm64:"LSR",-"CSEL"
		r3 = val16 >> shift
	}
	if shift >= 0 && shift < 8 {
		// arm64:"LSR",-"CSEL"
		r4 = val8 >> shift
	}
	return r1, r2, r3, r4
}

	// Only reduce, no multiplications are needed
	// First reduction step
	ADDS	acc0<<32, acc1, acc1
	LSR	$32, acc0, t0
	MUL	acc0, const1, t1
	UMULH	acc0, const1, acc0
	ADCS	t0, acc2
	ADCS	t1, acc3
	ADC	$0, acc0
	// Second reduction step
	ADDS	acc1<<32, acc2, acc2
	LSR	$32, acc1, t0
	MUL	acc1, const1, t1
	UMULH	acc1, const1, acc1
	ADCS	t0, acc3
	ADCS	t1, acc0
	ADC	$0, acc1

	VAND	V5.B16, V3.B16, V3.B16
	VAND	V5.B16, V4.B16, V4.B16
	VADDP	V4.B16, V3.B16, V6.B16 // 256->128
	VADDP	V6.B16, V6.B16, V6.B16 // 128->64
	VMOV	V6.D[0], R6
	// Clear the irrelevant lower bits
	LSL	$1, R9, R4
	LSR	R4, R6, R6
	LSL	R4, R6, R6
	// The first block can also be the last
	BLS	masklast
	// Have we found something already?
	CBNZ	R6, tail

loop:
	VLD1.P	(R3), [V1.B16, V2.B16]
	SUBS	$0x20, R2, R2

	FSUBD	F15, F1, F16	// F16 = y = 1-((lo-(r*c)/(2-c))-hi)
	// inline Ldexp(y, k), benefit:
	// 1, no parameter pass overhead.
	// 2, skip unnecessary checks for Inf/NaN/Zero
	FMOVD	F16, R0
	AND	$FracMask, R0, R2	// fraction
	LSR	$52, R0, R5	// exponent
	ADD	R1, R5		// R1 = int(k)
	CMP	$1, R5
	BGE	normal
	ADD	$52, R5		// denormal
	MOVD	$C1, R8
	FMOVD	R8, F1		// m = 2**-52
normal:
	ORR	R5<<52, R2, R0
	FMOVD	R0, F0

	"LDORLB",
	"LDORLD",
	"LDORLH",
	"LDORLW",
	"LDORW",
	"LDP",
	"LDPSW",
	"LDPW",
	"LDXP",
	"LDXPW",
	"LDXR",
	"LDXRB",
	"LDXRH",
	"LDXRW",
	"LSL",
	"LSLW",
	"LSR",
	"LSRW",
	"MADD",
	"MADDW",
	"MNEG",
	"MNEGW",
	"MOVB",
	"MOVBU",
	"MOVD",
	"MOVH",
	"MOVHU",
	"MOVK",
	"MOVKW",
	"MOVN",
	"MOVNW",
	"MOVP",
	"MOVPD",
	"MOVPQ",

func LenDiv1(a []int) int {
	// 386:"SHRL\t[$]10"
	// amd64:"SHRQ\t[$]10"
	// arm64:"LSR\t[$]10",-"SDIV"
	// arm:"SRL\t[$]10",-".*udiv"
	// ppc64x:"SRD"\t[$]10"
	return len(a) / 1024
}

func LenDiv2(s string) int {
	// 386:"SHRL\t[$]11"
	// amd64:"SHRQ\t[$]11"
	// arm64:"LSR\t[$]11",-"SDIV"
	// arm:"SRL\t[$]11",-".*udiv"
	// ppc64x:"SRD\t[$]11"
	return len(s) / (4097 >> 1)
}