Remove optimizations that the compiler will likely do
This commit is contained in:
Eirik Ogilvie-Wigley
parent
687d5f18c6
commit
e4b3a57e16
123
src/fq.rs
123
src/fq.rs
|
|
@ -64,13 +64,6 @@ fn adc(a: u64, b: u64, carry: u64) -> (u64, u64) {
|
|||
(adc as u64, (adc >> 64) as u64)
|
||||
}
|
||||
|
||||
/// Compute a + b, returning the result and the carry over.
|
||||
#[inline(always)]
|
||||
fn overflowing_add(a: u64, b: u64) -> (u64, u64) {
|
||||
let sum = u128::from(a) + u128::from(b);
|
||||
(sum as u64, (sum >> 64) as u64)
|
||||
}
|
||||
|
||||
/// Compute a - (b + borrow), returning the result and setting borrow to
|
||||
/// the borrow value.
|
||||
#[inline(always)]
|
||||
|
|
@ -79,23 +72,9 @@ fn sbb(a: u64, b: u64, borrow: &mut u128) -> u64 {
|
|||
*borrow as u64
|
||||
}
|
||||
|
||||
/// Compute (b * c), returning the result and the carry over.
|
||||
#[inline(always)]
|
||||
fn overflowing_mul(b: u64, c: u64) -> (u64, u64) {
|
||||
let mac = u128::from(b) * u128::from(c);
|
||||
(mac as u64, (mac >> 64) as u64)
|
||||
}
|
||||
|
||||
/// Compute a + (b * c), returning the result and the carry over.
|
||||
#[inline(always)]
|
||||
fn mac(a: u64, b: u64, c: u64) -> (u64, u64) {
|
||||
let mac = u128::from(a) + (u128::from(b) * u128::from(c));
|
||||
(mac as u64, (mac >> 64) as u64)
|
||||
}
|
||||
|
||||
/// Compute a + (b * c) + carry, returning the result and the new carry over.
|
||||
#[inline(always)]
|
||||
fn mac_with_carry(a: u64, b: u64, c: u64, carry: u64) -> (u64, u64) {
|
||||
fn mac(a: u64, b: u64, c: u64, carry: u64) -> (u64, u64) {
|
||||
let mac = u128::from(a) + (u128::from(b) * u128::from(c)) + u128::from(carry);
|
||||
(mac as u64, (mac >> 64) as u64)
|
||||
}
|
||||
|
|
@ -132,18 +111,18 @@ impl<'b> SubAssign<&'b Fq> for Fq {
|
|||
// borrow = 0x000...000. Thus, we use it as a mask to conditionally add the modulus.
|
||||
let borrow_mask = (borrow >> 64) as u64;
|
||||
|
||||
let (d0, carry) = overflowing_add(self.0[0], MODULUS.0[0] & borrow_mask);
|
||||
let (d0, carry) = adc(self.0[0], MODULUS.0[0] & borrow_mask, 0);
|
||||
let (d1, carry) = adc(self.0[1], MODULUS.0[1] & borrow_mask, carry);
|
||||
let (d2, carry) = adc(self.0[2], MODULUS.0[2] & borrow_mask, carry);
|
||||
let (d3, _) = adc(self.0[3], MODULUS.0[3] & borrow_mask, carry);
|
||||
|
||||
|
||||
self.0 = [d0, d1, d2, d3];
|
||||
}
|
||||
}
|
||||
|
||||
impl<'b> AddAssign<&'b Fq> for Fq {
|
||||
fn add_assign(&mut self, rhs: &'b Fq) {
|
||||
let (d0, carry) = overflowing_add(self.0[0], rhs.0[0]);
|
||||
let (d0, carry) = adc(self.0[0], rhs.0[0], 0);
|
||||
let (d1, carry) = adc(self.0[1], rhs.0[1], carry);
|
||||
let (d2, carry) = adc(self.0[2], rhs.0[2], carry);
|
||||
let (d3, _) = adc(self.0[3], rhs.0[3], carry);
|
||||
|
|
@ -160,25 +139,25 @@ impl<'b> MulAssign<&'b Fq> for Fq {
|
|||
fn mul_assign(&mut self, rhs: &'b Fq) {
|
||||
// Schoolbook multiplication
|
||||
|
||||
let (r0, carry) = overflowing_mul(self.0[0], rhs.0[0]);
|
||||
let (r1, carry) = mac(carry, self.0[0], rhs.0[1]);
|
||||
let (r2, carry) = mac(carry, self.0[0], rhs.0[2]);
|
||||
let (r3, r4) = mac(carry, self.0[0], rhs.0[3]);
|
||||
let (r0, carry) = mac(0, self.0[0], rhs.0[0], 0);
|
||||
let (r1, carry) = mac(0, self.0[0], rhs.0[1], carry);
|
||||
let (r2, carry) = mac(0, self.0[0], rhs.0[2], carry);
|
||||
let (r3, r4) = mac(0, self.0[0], rhs.0[3], carry);
|
||||
|
||||
let (r1, carry) = mac(r1, self.0[1], rhs.0[0]);
|
||||
let (r2, carry) = mac_with_carry(r2, self.0[1], rhs.0[1], carry);
|
||||
let (r3, carry) = mac_with_carry(r3, self.0[1], rhs.0[2], carry);
|
||||
let (r4, r5) = mac_with_carry(r4, self.0[1], rhs.0[3], carry);
|
||||
let (r1, carry) = mac(r1, self.0[1], rhs.0[0], 0);
|
||||
let (r2, carry) = mac(r2, self.0[1], rhs.0[1], carry);
|
||||
let (r3, carry) = mac(r3, self.0[1], rhs.0[2], carry);
|
||||
let (r4, r5) = mac(r4, self.0[1], rhs.0[3], carry);
|
||||
|
||||
let (r2, carry) = mac(r2, self.0[2], rhs.0[0]);
|
||||
let (r3, carry) = mac_with_carry(r3, self.0[2], rhs.0[1], carry);
|
||||
let (r4, carry) = mac_with_carry(r4, self.0[2], rhs.0[2], carry);
|
||||
let (r5, r6) = mac_with_carry(r5, self.0[2], rhs.0[3], carry);
|
||||
let (r2, carry) = mac(r2, self.0[2], rhs.0[0], 0);
|
||||
let (r3, carry) = mac(r3, self.0[2], rhs.0[1], carry);
|
||||
let (r4, carry) = mac(r4, self.0[2], rhs.0[2], carry);
|
||||
let (r5, r6) = mac(r5, self.0[2], rhs.0[3], carry);
|
||||
|
||||
let (r3, carry) = mac(r3, self.0[3], rhs.0[0]);
|
||||
let (r4, carry) = mac_with_carry(r4, self.0[3], rhs.0[1], carry);
|
||||
let (r5, carry) = mac_with_carry(r5, self.0[3], rhs.0[2], carry);
|
||||
let (r6, r7) = mac_with_carry(r6, self.0[3], rhs.0[3], carry);
|
||||
let (r3, carry) = mac(r3, self.0[3], rhs.0[0], 0);
|
||||
let (r4, carry) = mac(r4, self.0[3], rhs.0[1], carry);
|
||||
let (r5, carry) = mac(r5, self.0[3], rhs.0[2], carry);
|
||||
let (r6, r7) = mac(r6, self.0[3], rhs.0[3], carry);
|
||||
|
||||
self.montgomery_reduce(r0, r1, r2, r3, r4, r5, r6, r7);
|
||||
}
|
||||
|
|
@ -263,14 +242,14 @@ impl Fq {
|
|||
|
||||
/// Squares this element.
|
||||
pub fn square_assign(&mut self) {
|
||||
let (r1, carry) = overflowing_mul(self.0[0], self.0[1]);
|
||||
let (r2, carry) = mac(carry, self.0[0], self.0[2]);
|
||||
let (r3, r4) = mac(carry, self.0[0], self.0[3]);
|
||||
let (r1, carry) = mac(0, self.0[0], self.0[1], 0);
|
||||
let (r2, carry) = mac(0, self.0[0], self.0[2], carry);
|
||||
let (r3, r4) = mac(0, self.0[0], self.0[3], carry);
|
||||
|
||||
let (r3, carry) = mac(r3, self.0[1], self.0[2]);
|
||||
let (r4, r5) = mac_with_carry(r4, self.0[1], self.0[3], carry);
|
||||
let (r3, carry) = mac(r3, self.0[1], self.0[2], 0);
|
||||
let (r4, r5) = mac(r4, self.0[1], self.0[3], carry);
|
||||
|
||||
let (r5, r6) = mac(r5, self.0[2], self.0[3]);
|
||||
let (r5, r6) = mac(r5, self.0[2], self.0[3], 0);
|
||||
|
||||
let r7 = r6 >> 63;
|
||||
let r6 = (r6 << 1) | (r5 >> 63);
|
||||
|
|
@ -280,14 +259,14 @@ impl Fq {
|
|||
let r2 = (r2 << 1) | (r1 >> 63);
|
||||
let r1 = r1 << 1;
|
||||
|
||||
let (r0, carry) = overflowing_mul(self.0[0], self.0[0]);
|
||||
let (r1, carry) = overflowing_add(r1, carry);
|
||||
let (r2, carry) = mac_with_carry(r2, self.0[1], self.0[1], carry);
|
||||
let (r3, carry) = overflowing_add(r3, carry);
|
||||
let (r4, carry) = mac_with_carry(r4, self.0[2], self.0[2], carry);
|
||||
let (r5, carry) = overflowing_add(r5, carry);
|
||||
let (r6, carry) = mac_with_carry(r6, self.0[3], self.0[3], carry);
|
||||
let r7 = r7 + carry;
|
||||
let (r0, carry) = mac(0, self.0[0], self.0[0], 0);
|
||||
let (r1, carry) = adc(0, r1, carry);
|
||||
let (r2, carry) = mac(r2, self.0[1], self.0[1], carry);
|
||||
let (r3, carry) = adc(0, r3, carry);
|
||||
let (r4, carry) = mac(r4, self.0[2], self.0[2], carry);
|
||||
let (r5, carry) = adc(0, r5, carry);
|
||||
let (r6, carry) = mac(r6, self.0[3], self.0[3], carry);
|
||||
let (r7, _) = adc(0, r7, carry);
|
||||
|
||||
self.montgomery_reduce(r0, r1, r2, r3, r4, r5, r6, r7);
|
||||
}
|
||||
|
|
@ -337,32 +316,32 @@ impl Fq {
|
|||
// <http://cacr.uwaterloo.ca/hac/about/chap14.pdf>.
|
||||
|
||||
let k = r0.wrapping_mul(INV);
|
||||
let (_, carry) = mac(r0, k, MODULUS.0[0]);
|
||||
let (r1, carry) = mac_with_carry(r1, k, MODULUS.0[1], carry);
|
||||
let (r2, carry) = mac_with_carry(r2, k, MODULUS.0[2], carry);
|
||||
let (r3, carry) = mac_with_carry(r3, k, MODULUS.0[3], carry);
|
||||
let (_, carry) = mac(r0, k, MODULUS.0[0], 0);
|
||||
let (r1, carry) = mac(r1, k, MODULUS.0[1], carry);
|
||||
let (r2, carry) = mac(r2, k, MODULUS.0[2], carry);
|
||||
let (r3, carry) = mac(r3, k, MODULUS.0[3], carry);
|
||||
let (r4, carry2) = adc(r4, 0, carry);
|
||||
|
||||
let k = r1.wrapping_mul(INV);
|
||||
let (_, carry) = mac(r1, k, MODULUS.0[0]);
|
||||
let (r2, carry) = mac_with_carry(r2, k, MODULUS.0[1], carry);
|
||||
let (r3, carry) = mac_with_carry(r3, k, MODULUS.0[2], carry);
|
||||
let (r4, carry) = mac_with_carry(r4, k, MODULUS.0[3], carry);
|
||||
let (_, carry) = mac(r1, k, MODULUS.0[0], 0);
|
||||
let (r2, carry) = mac(r2, k, MODULUS.0[1], carry);
|
||||
let (r3, carry) = mac(r3, k, MODULUS.0[2], carry);
|
||||
let (r4, carry) = mac(r4, k, MODULUS.0[3], carry);
|
||||
let (r5, carry2) = adc(r5, carry2, carry);
|
||||
|
||||
let k = r2.wrapping_mul(INV);
|
||||
let (_, carry) = mac(r2, k, MODULUS.0[0]);
|
||||
let (r3, carry) = mac_with_carry(r3, k, MODULUS.0[1], carry);
|
||||
let (r4, carry) = mac_with_carry(r4, k, MODULUS.0[2], carry);
|
||||
let (r5, carry) = mac_with_carry(r5, k, MODULUS.0[3], carry);
|
||||
let (_, carry) = mac(r2, k, MODULUS.0[0], 0);
|
||||
let (r3, carry) = mac(r3, k, MODULUS.0[1], carry);
|
||||
let (r4, carry) = mac(r4, k, MODULUS.0[2], carry);
|
||||
let (r5, carry) = mac(r5, k, MODULUS.0[3], carry);
|
||||
let (r6, carry2) = adc(r6, carry2, carry);
|
||||
|
||||
let k = r3.wrapping_mul(INV);
|
||||
let (_, carry) = mac(r3, k, MODULUS.0[0]);
|
||||
let (r4, carry) = mac_with_carry(r4, k, MODULUS.0[1], carry);
|
||||
let (r5, carry) = mac_with_carry(r5, k, MODULUS.0[2], carry);
|
||||
let (r6, carry) = mac_with_carry(r6, k, MODULUS.0[3], carry);
|
||||
let r7 = r7 + carry2 + carry;
|
||||
let (_, carry) = mac(r3, k, MODULUS.0[0], 0);
|
||||
let (r4, carry) = mac(r4, k, MODULUS.0[1], carry);
|
||||
let (r5, carry) = mac(r5, k, MODULUS.0[2], carry);
|
||||
let (r6, carry) = mac(r6, k, MODULUS.0[3], carry);
|
||||
let (r7, _) = adc(r7, carry2, carry);
|
||||
|
||||
self.0[0] = r4;
|
||||
self.0[1] = r5;
|
||||
|
|
|
|||
Loading…
Reference in New Issue