L2:
	MOVW.P	4(R2), R5
	MOVW.P	4(R3), R6
	SBC.S	R6, R5
	MOVW.P	R5, 4(R1)
E2:
	TEQ	R1, R4
	BNE L2

	MOVW	$0, R0
	MOVW.CC	$1, R0
	MOVW	R0, c+36(FP)
	RET


// func addVW(z, x []Word, y Word) (c Word)
TEXT ·addVW(SB),NOSPLIT,$0
	MOVW	z+0(FP), R1
	MOVW	z_len+4(FP), R4
	MOVW	x+12(FP), R2
	MOVW	y+24(FP), R3
	ADD	R4<<2, R1, R4
	TEQ	R1, R4
	BNE L3a
	MOVW	R3, c+28(FP)
	RET
L3a:
	MOVW.P	4(R2), R5

	SBBQ CX, CX		// save CF

	ADDQ $1, SI		// i++
	SUBQ $1, DI		// n--
	JG L2			// if n > 0 goto L2

E2:	NEGQ CX
	MOVQ CX, c+72(FP)	// return c
	RET


// func addVW(z, x []Word, y Word) (c Word)
TEXT ·addVW(SB),NOSPLIT,$0
	MOVQ z_len+8(FP), DI
	CMPQ DI, $32
	JG large
	MOVQ x+24(FP), R8
	MOVQ y+48(FP), CX	// c = y
	MOVQ z+0(FP), R10

	MOVQ $0, SI		// i = 0

	for i := 0; i < len(z) && i < len(x); i++ {
		zi, cc := bits.Add(uint(x[i]), uint(c), 0)
		z[i] = Word(zi)
		c = Word(cc)
	}
	return
}

// addVWlarge is addVW, but intended for large z.
// The only difference is that we check on every iteration
// whether we are done with carries,
// and if so, switch to a much faster copy instead.
// This is only a good idea for large z,

	MOVD  8(R8), R11
	MOVD  8(R9), R16
	SUBE  R16, R11, R20
	MOVD  R20, 8(R10)

final:
	ADDZE R4
	XOR   $1, R4

done:
	MOVD  R4, c+72(FP)
	RET

// func addVW(z, x []Word, y Word) (c Word)
TEXT ·addVW(SB), NOSPLIT, $0
	MOVD z+0(FP), R10	// R10 = z[]
	MOVD x+24(FP), R8	// R8 = x[]
	MOVD y+48(FP), R4	// R4 = y = c
	MOVD z_len+8(FP), R11	// R11 = z_len

			"flag.kind",
			"flag.ro",
		},
		"regexp": {
			"(*bitState).push",
		},
		"math/big": {
			"bigEndianWord",
			// The following functions require the math_big_pure_go build tag.
			"addVW",
			"subVW",
		},
		"math/rand": {
			"(*rngSource).Int63",
			"(*rngSource).Uint64",
		},
		"net": {
			"(*UDPConn).ReadFromUDP",
		},
		"sync": {

		// n == 0 because m >= n; result is 0
		return z[:0]
	case n == 0:
		// result is x
		return z.set(x)
	}
	// m > 0

	z = z.make(m + 1)
	c := addVV(z[0:n], x, y)
	if m > n {
		c = addVW(z[n:m], x[n:], c)
	}
	z[m] = c

	return z.norm()
}

func (z nat) sub(x, y nat) nat {
	m := len(x)
	n := len(y)

	switch {
	case m < n:
		panic("underflow")
	case m == 0:

	MOVD R0, R4
	SUBE R4, R4            // save CF

	ADD $8, R10 // i++
	SUB $1, R3  // n--
	BGT L1      // if n > 0 goto L1

E1:
	NEG  R4, R4
	MOVD R4, c+72(FP) // return c
	RET

TEXT ·addVW(SB), NOSPLIT, $0
	MOVD z_len+8(FP), R5 // length of z
	MOVD x+24(FP), R6
	MOVD y+48(FP), R7    // c = y
	MOVD z+0(FP), R8

	CMPBEQ R5, $0, returnC // if len(z) == 0, we can have an early return

		// and it's Below if we truncate (Exact results require no rounding).
		// For a negative result (z.neg) it is exactly the opposite.
		z.acc = makeAcc(inc != z.neg)

		if inc {
			// add 1 to mantissa
			if addVW(z.mant, z.mant, lsb) != 0 {
				// mantissa overflow => adjust exponent
				if z.exp >= MaxExp {
					// exponent overflow
					z.form = inf
					return
				}
				z.exp++