package sike

// Fp implementation

// Compute z = x + y (mod 2*p).
func fpAddRdc(z, x, y *Fp) {
	var carry uint64

	// z=x+y % p503
	for i := 0; i < FP_WORDS; i++ {
		z[i], carry = addc64(carry, x[i], y[i])
	}

	// z = z - p503x2
	carry = 0
	for i := 0; i < FP_WORDS; i++ {
		z[i], carry = subc64(carry, z[i], p503x2[i])
	}

	// if z<0 add p503x2 back
	mask := uint64(0 - carry)
	carry = 0
	for i := 0; i < FP_WORDS; i++ {
		z[i], carry = addc64(carry, z[i], p503x2[i]&mask)
	}
}

// Compute z = x - y (mod 2*p).
func fpSubRdc(z, x, y *Fp) {
	var borrow uint64

	// z = z - p503x2
	for i := 0; i < FP_WORDS; i++ {
		z[i], borrow = subc64(borrow, x[i], y[i])
	}

	// if z<0 add p503x2 back
	mask := uint64(0 - borrow)
	borrow = 0
	for i := 0; i < FP_WORDS; i++ {
		z[i], borrow = addc64(borrow, z[i], p503x2[i]&mask)
	}
}

// Reduce a field element in [0, 2*p) to one in [0,p).
func fpRdcP(x *Fp) {
	var borrow, mask uint64
	for i := 0; i < FP_WORDS; i++ {
		x[i], borrow = subc64(borrow, x[i], p503[i])
	}

	// Sets all bits if borrow = 1
	mask = 0 - borrow
	borrow = 0
	for i := 0; i < FP_WORDS; i++ {
		x[i], borrow = addc64(borrow, x[i], p503[i]&mask)
	}
}

// Implementation doesn't actually depend on a prime field.
func fpSwapCond(x, y *Fp, mask uint8) {
	if mask != 0 {
		var tmp Fp
		copy(tmp[:], y[:])
		copy(y[:], x[:])
		copy(x[:], tmp[:])
	}
}

// Compute z = x * y.
func fpMul(z *FpX2, x, y *Fp) {
	var u, v, t uint64
	var carry uint64
	var uv uint128

	for i := uint64(0); i < FP_WORDS; i++ {
		for j := uint64(0); j <= i; j++ {
			uv = mul64(x[j], y[i-j])
			v, carry = addc64(0, uv.L, v)
			u, carry = addc64(carry, uv.H, u)
			t += carry
		}
		z[i] = v
		v = u
		u = t
		t = 0
	}

	for i := FP_WORDS; i < (2*FP_WORDS)-1; i++ {
		for j := i - FP_WORDS + 1; j < FP_WORDS; j++ {
			uv = mul64(x[j], y[i-j])
			v, carry = addc64(0, uv.L, v)
			u, carry = addc64(carry, uv.H, u)
			t += carry
		}
		z[i] = v
		v = u
		u = t
		t = 0
	}
	z[2*FP_WORDS-1] = v
}

// Perform Montgomery reduction: set z = x R^{-1} (mod 2*p)
// with R=2^512. Destroys the input value.
func fpMontRdc(z *Fp, x *FpX2) {
	var carry, t, u, v uint64
	var uv uint128
	var count int

	count = 3 // number of 0 digits in the least significat part of p503 + 1

	for i := 0; i < FP_WORDS; i++ {
		for j := 0; j < i; j++ {
			if j < (i - count + 1) {
				uv = mul64(z[j], p503p1[i-j])
				v, carry = addc64(0, uv.L, v)
				u, carry = addc64(carry, uv.H, u)
				t += carry
			}
		}
		v, carry = addc64(0, v, x[i])
		u, carry = addc64(carry, u, 0)
		t += carry

		z[i] = v
		v = u
		u = t
		t = 0
	}

	for i := FP_WORDS; i < 2*FP_WORDS-1; i++ {
		if count > 0 {
			count--
		}
		for j := i - FP_WORDS + 1; j < FP_WORDS; j++ {
			if j < (FP_WORDS - count) {
				uv = mul64(z[j], p503p1[i-j])
				v, carry = addc64(0, uv.L, v)
				u, carry = addc64(carry, uv.H, u)
				t += carry
			}
		}
		v, carry = addc64(0, v, x[i])
		u, carry = addc64(carry, u, 0)

		t += carry
		z[i-FP_WORDS] = v
		v = u
		u = t
		t = 0
	}
	v, carry = addc64(0, v, x[2*FP_WORDS-1])
	z[FP_WORDS-1] = v
}

// Compute z = x + y, without reducing mod p.
func fp2Add(z, x, y *FpX2) {
	var carry uint64
	for i := 0; i < 2*FP_WORDS; i++ {
		z[i], carry = addc64(carry, x[i], y[i])
	}
}

// Compute z = x - y, without reducing mod p.
func fp2Sub(z, x, y *FpX2) {
	var borrow, mask uint64
	for i := 0; i < 2*FP_WORDS; i++ {
		z[i], borrow = subc64(borrow, x[i], y[i])
	}

	// Sets all bits if borrow = 1
	mask = 0 - borrow
	borrow = 0
	for i := FP_WORDS; i < 2*FP_WORDS; i++ {
		z[i], borrow = addc64(borrow, z[i], p503[i-FP_WORDS]&mask)
	}
}

// Montgomery multiplication. Input values must be already
// in Montgomery domain.
func fpMulRdc(dest, lhs, rhs *Fp) {
	a := lhs // = a*R
	b := rhs // = b*R

	var ab FpX2
	fpMul(&ab, a, b)     // = a*b*R*R
	fpMontRdc(dest, &ab) // = a*b*R mod p
}