1 //! Simple variable-time big integer implementation
4 use core::marker::PhantomData;
6 const WORD_COUNT_4096: usize = 4096 / 64;
7 const WORD_COUNT_256: usize = 256 / 64;
8 const WORD_COUNT_384: usize = 384 / 64;
10 // RFC 5702 indicates RSA keys can be up to 4096 bits
11 #[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
12 pub(super) struct U4096([u64; WORD_COUNT_4096]);
14 #[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
15 pub(super) struct U256([u64; WORD_COUNT_256]);
17 #[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
18 pub(super) struct U384([u64; WORD_COUNT_384]);
20 pub(super) trait Int: Clone + Ord + Sized {
23 fn from_be_bytes(b: &[u8]) -> Result<Self, ()>;
24 fn limbs(&self) -> &[u64];
27 const ZERO: U256 = U256([0; 4]);
28 const BYTES: usize = 32;
29 fn from_be_bytes(b: &[u8]) -> Result<Self, ()> { Self::from_be_bytes(b) }
30 fn limbs(&self) -> &[u64] { &self.0 }
33 const ZERO: U384 = U384([0; 6]);
34 const BYTES: usize = 48;
35 fn from_be_bytes(b: &[u8]) -> Result<Self, ()> { Self::from_be_bytes(b) }
36 fn limbs(&self) -> &[u64] { &self.0 }
39 /// Defines a *PRIME* Modulus
40 pub(super) trait PrimeModulus<I: Int> {
42 const R_SQUARED_MOD_PRIME: I;
43 const NEGATIVE_PRIME_INV_MOD_R: I;
46 #[derive(Clone, Debug, PartialEq, Eq)] // Ord doesn't make sense cause we have an R factor
47 pub(super) struct U256Mod<M: PrimeModulus<U256>>(U256, PhantomData<M>);
49 #[derive(Clone, Debug, PartialEq, Eq)] // Ord doesn't make sense cause we have an R factor
50 pub(super) struct U384Mod<M: PrimeModulus<U384>>(U384, PhantomData<M>);
52 macro_rules! debug_unwrap { ($v: expr) => { {
54 debug_assert!(v.is_ok());
57 Err(e) => return Err(e),
61 // Various const versions of existing slice utilities
62 /// Const version of `&a[start..end]`
63 const fn const_subslice<'a, T>(a: &'a [T], start: usize, end: usize) -> &'a [T] {
64 assert!(start <= a.len());
65 assert!(end <= a.len());
66 assert!(end >= start);
67 let mut startptr = a.as_ptr();
68 startptr = unsafe { startptr.add(start) };
69 let len = end - start;
70 // The docs for from_raw_parts do not mention any requirements that the pointer be valid if the
71 // length is zero, aside from requiring proper alignment (which is met here). Thus,
72 // one-past-the-end should be an acceptable pointer for a 0-length slice.
73 unsafe { alloc::slice::from_raw_parts(startptr, len) }
76 /// Const version of `dest[dest_start..dest_end].copy_from_slice(source)`
78 /// Once `const_mut_refs` is stable we can convert this to a function
79 macro_rules! copy_from_slice {
80 ($dest: ident, $dest_start: expr, $dest_end: expr, $source: ident) => { {
81 let dest_start = $dest_start;
82 let dest_end = $dest_end;
83 assert!(dest_start <= $dest.len());
84 assert!(dest_end <= $dest.len());
85 assert!(dest_end >= dest_start);
86 assert!(dest_end - dest_start == $source.len());
88 while i < $source.len() {
89 $dest[i + dest_start] = $source[i];
95 /// Const version of a > b
96 const fn slice_greater_than(a: &[u64], b: &[u64]) -> bool {
97 debug_assert!(a.len() == b.len());
98 let len = if a.len() <= b.len() { a.len() } else { b.len() };
101 if a[i] > b[i] { return true; }
102 else if a[i] < b[i] { return false; }
108 /// Const version of a == b
109 const fn slice_equal(a: &[u64], b: &[u64]) -> bool {
110 debug_assert!(a.len() == b.len());
111 let len = if a.len() <= b.len() { a.len() } else { b.len() };
114 if a[i] != b[i] { return false; }
120 /// Adds a single u64 valuein-place, returning an overflow flag, in which case one out-of-bounds
121 /// high bit is implicitly included in the result.
123 /// Once `const_mut_refs` is stable we can convert this to a function
124 macro_rules! add_u64 { ($a: ident, $b: expr) => { {
129 let (v, carry) = $a[i].overflowing_add(add);
132 if add == 0 { break; }
140 /// Negates the given u64 slice.
142 /// Once `const_mut_refs` is stable we can convert this to a function
143 macro_rules! negate { ($v: ident) => { {
146 $v[i] ^= 0xffff_ffff_ffff_ffff;
149 let overflow = add_u64!($v, 1);
150 debug_assert!(!overflow);
153 /// Doubles in-place, returning an overflow flag, in which case one out-of-bounds high bit is
154 /// implicitly included in the result.
156 /// Once `const_mut_refs` is stable we can convert this to a function
157 macro_rules! double { ($a: ident) => { {
158 { let _: &[u64] = &$a; } // Force type resolution
160 let mut carry = false;
163 let mut next_carry = ($a[len - 1 - i] & (1 << 63)) != 0;
164 let (v, next_carry_2) = ($a[len - 1 - i] << 1).overflowing_add(carry as u64);
166 debug_assert!(!next_carry || !next_carry_2);
167 next_carry |= next_carry_2;
174 macro_rules! define_add { ($name: ident, $len: expr) => {
175 /// Adds two $len-64-bit integers together, returning a new $len-64-bit integer and an overflow
176 /// bit, with the same semantics as the std [`u64::overflowing_add`] method.
177 const fn $name(a: &[u64], b: &[u64]) -> ([u64; $len], bool) {
178 debug_assert!(a.len() == $len);
179 debug_assert!(b.len() == $len);
180 let mut r = [0; $len];
181 let mut carry = false;
184 let pos = $len - 1 - i;
185 let (v, mut new_carry) = a[pos].overflowing_add(b[pos]);
186 let (v2, new_new_carry) = v.overflowing_add(carry as u64);
187 new_carry |= new_new_carry;
196 define_add!(add_2, 2);
197 define_add!(add_3, 3);
198 define_add!(add_4, 4);
199 define_add!(add_6, 6);
200 define_add!(add_8, 8);
201 define_add!(add_12, 12);
202 define_add!(add_16, 16);
203 define_add!(add_32, 32);
204 define_add!(add_64, 64);
205 define_add!(add_128, 128);
207 macro_rules! define_sub { ($name: ident, $name_abs: ident, $len: expr) => {
208 /// Subtracts the `b` $len-64-bit integer from the `a` $len-64-bit integer, returning a new
209 /// $len-64-bit integer and an overflow bit, with the same semantics as the std
210 /// [`u64::overflowing_sub`] method.
211 const fn $name(a: &[u64], b: &[u64]) -> ([u64; $len], bool) {
212 debug_assert!(a.len() == $len);
213 debug_assert!(b.len() == $len);
214 let mut r = [0; $len];
215 let mut carry = false;
218 let pos = $len - 1 - i;
219 let (v, mut new_carry) = a[pos].overflowing_sub(b[pos]);
220 let (v2, new_new_carry) = v.overflowing_sub(carry as u64);
221 new_carry |= new_new_carry;
229 /// Subtracts the `b` $len-64-bit integer from the `a` $len-64-bit integer, returning a new
230 /// $len-64-bit integer representing the absolute value of the result, as well as a sign bit.
232 const fn $name_abs(a: &[u64], b: &[u64]) -> ([u64; $len], bool) {
233 let (mut res, neg) = $name(a, b);
241 define_sub!(sub_2, sub_abs_2, 2);
242 define_sub!(sub_3, sub_abs_3, 3);
243 define_sub!(sub_4, sub_abs_4, 4);
244 define_sub!(sub_6, sub_abs_6, 6);
245 define_sub!(sub_8, sub_abs_8, 8);
246 define_sub!(sub_12, sub_abs_12, 12);
247 define_sub!(sub_16, sub_abs_16, 16);
248 define_sub!(sub_32, sub_abs_32, 32);
249 define_sub!(sub_64, sub_abs_64, 64);
250 define_sub!(sub_128, sub_abs_128, 128);
252 /// Multiplies two 128-bit integers together, returning a new 256-bit integer.
254 /// This is the base case for our multiplication, taking advantage of Rust's native 128-bit int
255 /// types to do multiplication (potentially) natively.
256 const fn mul_2(a: &[u64], b: &[u64]) -> [u64; 4] {
257 debug_assert!(a.len() == 2);
258 debug_assert!(b.len() == 2);
260 // Gradeschool multiplication is way faster here.
261 let (a0, a1) = (a[0] as u128, a[1] as u128);
262 let (b0, b1) = (b[0] as u128, b[1] as u128);
266 let (z1, i_carry) = z1i.overflowing_add(z1j);
269 let z2a = ((z2 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
270 let z1a = ((z1 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
271 let z0a = ((z0 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
272 let z2b = (z2 & 0xffff_ffff_ffff_ffff) as u64;
273 let z1b = (z1 & 0xffff_ffff_ffff_ffff) as u64;
274 let z0b = (z0 & 0xffff_ffff_ffff_ffff) as u64;
277 let (k, j_carry) = z0a.overflowing_add(z1b);
278 let (mut j, mut second_i_carry) = z1a.overflowing_add(z2b);
281 (j, new_i_carry) = j.overflowing_add(j_carry as u64);
282 debug_assert!(!second_i_carry || !new_i_carry);
283 second_i_carry |= new_i_carry;
286 let mut spurious_overflow;
287 (i, spurious_overflow) = i.overflowing_add(i_carry as u64);
288 debug_assert!(!spurious_overflow);
289 (i, spurious_overflow) = i.overflowing_add(second_i_carry as u64);
290 debug_assert!(!spurious_overflow);
295 const fn mul_3(a: &[u64], b: &[u64]) -> [u64; 6] {
296 debug_assert!(a.len() == 3);
297 debug_assert!(b.len() == 3);
299 let (a0, a1, a2) = (a[0] as u128, a[1] as u128, a[2] as u128);
300 let (b0, b1, b2) = (b[0] as u128, b[1] as u128, b[2] as u128);
312 let r5 = ((m4 >> 0) & 0xffff_ffff_ffff_ffff) as u64;
314 let r4a = ((m4 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
315 let r4b = ((m3a >> 0) & 0xffff_ffff_ffff_ffff) as u64;
316 let r4c = ((m3b >> 0) & 0xffff_ffff_ffff_ffff) as u64;
318 let r3a = ((m3a >> 64) & 0xffff_ffff_ffff_ffff) as u64;
319 let r3b = ((m3b >> 64) & 0xffff_ffff_ffff_ffff) as u64;
320 let r3c = ((m2a >> 0 ) & 0xffff_ffff_ffff_ffff) as u64;
321 let r3d = ((m2b >> 0 ) & 0xffff_ffff_ffff_ffff) as u64;
322 let r3e = ((m2c >> 0 ) & 0xffff_ffff_ffff_ffff) as u64;
324 let r2a = ((m2a >> 64) & 0xffff_ffff_ffff_ffff) as u64;
325 let r2b = ((m2b >> 64) & 0xffff_ffff_ffff_ffff) as u64;
326 let r2c = ((m2c >> 64) & 0xffff_ffff_ffff_ffff) as u64;
327 let r2d = ((m1a >> 0 ) & 0xffff_ffff_ffff_ffff) as u64;
328 let r2e = ((m1b >> 0 ) & 0xffff_ffff_ffff_ffff) as u64;
330 let r1a = ((m1a >> 64) & 0xffff_ffff_ffff_ffff) as u64;
331 let r1b = ((m1b >> 64) & 0xffff_ffff_ffff_ffff) as u64;
332 let r1c = ((m0 >> 0 ) & 0xffff_ffff_ffff_ffff) as u64;
334 let r0a = ((m0 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
336 let (r4, r3_ca) = r4a.overflowing_add(r4b);
337 let (r4, r3_cb) = r4.overflowing_add(r4c);
338 let r3_c = r3_ca as u64 + r3_cb as u64;
340 let (r3, r2_ca) = r3a.overflowing_add(r3b);
341 let (r3, r2_cb) = r3.overflowing_add(r3c);
342 let (r3, r2_cc) = r3.overflowing_add(r3d);
343 let (r3, r2_cd) = r3.overflowing_add(r3e);
344 let (r3, r2_ce) = r3.overflowing_add(r3_c);
345 let r2_c = r2_ca as u64 + r2_cb as u64 + r2_cc as u64 + r2_cd as u64 + r2_ce as u64;
347 let (r2, r1_ca) = r2a.overflowing_add(r2b);
348 let (r2, r1_cb) = r2.overflowing_add(r2c);
349 let (r2, r1_cc) = r2.overflowing_add(r2d);
350 let (r2, r1_cd) = r2.overflowing_add(r2e);
351 let (r2, r1_ce) = r2.overflowing_add(r2_c);
352 let r1_c = r1_ca as u64 + r1_cb as u64 + r1_cc as u64 + r1_cd as u64 + r1_ce as u64;
354 let (r1, r0_ca) = r1a.overflowing_add(r1b);
355 let (r1, r0_cb) = r1.overflowing_add(r1c);
356 let (r1, r0_cc) = r1.overflowing_add(r1_c);
357 let r0_c = r0_ca as u64 + r0_cb as u64 + r0_cc as u64;
359 let (r0, must_not_overflow) = r0a.overflowing_add(r0_c);
360 debug_assert!(!must_not_overflow);
362 [r0, r1, r2, r3, r4, r5]
365 macro_rules! define_mul { ($name: ident, $len: expr, $submul: ident, $add: ident, $subadd: ident, $sub: ident, $subsub: ident) => {
366 /// Multiplies two $len-64-bit integers together, returning a new $len*2-64-bit integer.
367 const fn $name(a: &[u64], b: &[u64]) -> [u64; $len * 2] {
368 // We could probably get a bit faster doing gradeschool multiplication for some smaller
369 // sizes, but its easier to just have one variable-length multiplication, so we do
370 // Karatsuba always here.
371 debug_assert!(a.len() == $len);
372 debug_assert!(b.len() == $len);
374 let a0 = const_subslice(a, 0, $len / 2);
375 let a1 = const_subslice(a, $len / 2, $len);
376 let b0 = const_subslice(b, 0, $len / 2);
377 let b1 = const_subslice(b, $len / 2, $len);
379 let z2 = $submul(a0, b0);
380 let z0 = $submul(a1, b1);
382 let (z1a_max, z1a_min, z1a_sign) =
383 if slice_greater_than(&a1, &a0) { (a1, a0, true) } else { (a0, a1, false) };
384 let (z1b_max, z1b_min, z1b_sign) =
385 if slice_greater_than(&b1, &b0) { (b1, b0, true) } else { (b0, b1, false) };
387 let z1a = $subsub(z1a_max, z1a_min);
388 debug_assert!(!z1a.1);
389 let z1b = $subsub(z1b_max, z1b_min);
390 debug_assert!(!z1b.1);
391 let z1m_sign = z1a_sign == z1b_sign;
393 let z1m = $submul(&z1a.0, &z1b.0);
394 let z1n = $add(&z0, &z2);
395 let mut z1_carry = z1n.1;
396 let z1 = if z1m_sign {
397 let r = $sub(&z1n.0, &z1m);
398 if r.1 { z1_carry ^= true; }
401 let r = $add(&z1n.0, &z1m);
402 if r.1 { z1_carry = true; }
406 let l = const_subslice(&z0, $len / 2, $len);
407 let (k, j_carry) = $subadd(const_subslice(&z0, 0, $len / 2), const_subslice(&z1, $len / 2, $len));
408 let (mut j, mut i_carry) = $subadd(const_subslice(&z1, 0, $len / 2), const_subslice(&z2, $len / 2, $len));
410 let new_i_carry = add_u64!(j, 1);
411 debug_assert!(!i_carry || !new_i_carry);
412 i_carry |= new_i_carry;
414 let mut i = [0; $len / 2];
415 let i_source = const_subslice(&z2, 0, $len / 2);
416 copy_from_slice!(i, 0, $len / 2, i_source);
418 let spurious_carry = add_u64!(i, 1);
419 debug_assert!(!spurious_carry);
422 let spurious_carry = add_u64!(i, 1);
423 debug_assert!(!spurious_carry);
426 let mut res = [0; $len * 2];
427 copy_from_slice!(res, $len * 2 * 0 / 4, $len * 2 * 1 / 4, i);
428 copy_from_slice!(res, $len * 2 * 1 / 4, $len * 2 * 2 / 4, j);
429 copy_from_slice!(res, $len * 2 * 2 / 4, $len * 2 * 3 / 4, k);
430 copy_from_slice!(res, $len * 2 * 3 / 4, $len * 2 * 4 / 4, l);
435 define_mul!(mul_4, 4, mul_2, add_4, add_2, sub_4, sub_2);
436 define_mul!(mul_6, 6, mul_3, add_6, add_3, sub_6, sub_3);
437 define_mul!(mul_8, 8, mul_4, add_8, add_4, sub_8, sub_4);
438 define_mul!(mul_16, 16, mul_8, add_16, add_8, sub_16, sub_8);
439 define_mul!(mul_32, 32, mul_16, add_32, add_16, sub_32, sub_16);
440 define_mul!(mul_64, 64, mul_32, add_64, add_32, sub_64, sub_32);
443 /// Squares a 128-bit integer, returning a new 256-bit integer.
445 /// This is the base case for our squaring, taking advantage of Rust's native 128-bit int
446 /// types to do multiplication (potentially) natively.
447 const fn sqr_2(a: &[u64]) -> [u64; 4] {
448 debug_assert!(a.len() == 2);
450 let (a0, a1) = (a[0] as u128, a[1] as u128);
452 let mut z1 = a0 * a1;
453 let i_carry = z1 & (1u128 << 127) != 0;
457 let z2a = ((z2 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
458 let z1a = ((z1 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
459 let z0a = ((z0 >> 64) & 0xffff_ffff_ffff_ffff) as u64;
460 let z2b = (z2 & 0xffff_ffff_ffff_ffff) as u64;
461 let z1b = (z1 & 0xffff_ffff_ffff_ffff) as u64;
462 let z0b = (z0 & 0xffff_ffff_ffff_ffff) as u64;
465 let (k, j_carry) = z0a.overflowing_add(z1b);
466 let (mut j, mut second_i_carry) = z1a.overflowing_add(z2b);
469 (j, new_i_carry) = j.overflowing_add(j_carry as u64);
470 debug_assert!(!second_i_carry || !new_i_carry);
471 second_i_carry |= new_i_carry;
474 let mut spurious_overflow;
475 (i, spurious_overflow) = i.overflowing_add(i_carry as u64);
476 debug_assert!(!spurious_overflow);
477 (i, spurious_overflow) = i.overflowing_add(second_i_carry as u64);
478 debug_assert!(!spurious_overflow);
483 macro_rules! define_sqr { ($name: ident, $len: expr, $submul: ident, $subsqr: ident, $subadd: ident) => {
484 /// Squares a $len-64-bit integers, returning a new $len*2-64-bit integer.
485 const fn $name(a: &[u64]) -> [u64; $len * 2] {
486 debug_assert!(a.len() == $len);
488 let hi = const_subslice(a, 0, $len / 2);
489 let lo = const_subslice(a, $len / 2, $len);
491 let v0 = $subsqr(lo);
492 let mut v1 = $submul(hi, lo);
493 let i_carry = double!(v1);
494 let v2 = $subsqr(hi);
496 let l = const_subslice(&v0, $len / 2, $len);
497 let (k, j_carry) = $subadd(const_subslice(&v0, 0, $len / 2), const_subslice(&v1, $len / 2, $len));
498 let (mut j, mut i_carry_2) = $subadd(const_subslice(&v1, 0, $len / 2), const_subslice(&v2, $len / 2, $len));
500 let mut i = [0; $len / 2];
501 let i_source = const_subslice(&v2, 0, $len / 2);
502 copy_from_slice!(i, 0, $len / 2, i_source);
505 let new_i_carry = add_u64!(j, 1);
506 debug_assert!(!i_carry_2 || !new_i_carry);
507 i_carry_2 |= new_i_carry;
510 let spurious_carry = add_u64!(i, 1);
511 debug_assert!(!spurious_carry);
514 let spurious_carry = add_u64!(i, 1);
515 debug_assert!(!spurious_carry);
518 let mut res = [0; $len * 2];
519 copy_from_slice!(res, $len * 2 * 0 / 4, $len * 2 * 1 / 4, i);
520 copy_from_slice!(res, $len * 2 * 1 / 4, $len * 2 * 2 / 4, j);
521 copy_from_slice!(res, $len * 2 * 2 / 4, $len * 2 * 3 / 4, k);
522 copy_from_slice!(res, $len * 2 * 3 / 4, $len * 2 * 4 / 4, l);
527 // TODO: Write an optimized sqr_3 (though secp384r1 is barely used)
528 const fn sqr_3(a: &[u64]) -> [u64; 6] { mul_3(a, a) }
530 define_sqr!(sqr_4, 4, mul_2, sqr_2, add_2);
531 define_sqr!(sqr_6, 6, mul_3, sqr_3, add_3);
532 define_sqr!(sqr_8, 8, mul_4, sqr_4, add_4);
533 define_sqr!(sqr_16, 16, mul_8, sqr_8, add_8);
534 define_sqr!(sqr_32, 32, mul_16, sqr_16, add_16);
535 define_sqr!(sqr_64, 64, mul_32, sqr_32, add_32);
537 macro_rules! dummy_pre_push { ($name: ident, $len: expr) => {} }
538 macro_rules! vec_pre_push { ($name: ident, $len: expr) => { $name.push([0; $len]); } }
540 macro_rules! define_div_rem { ($name: ident, $len: expr, $sub: ident, $heap_init: expr, $pre_push: ident $(, $const_opt: tt)?) => {
541 /// Divides two $len-64-bit integers, `a` by `b`, returning the quotient and remainder
543 /// Fails iff `b` is zero.
544 $($const_opt)? fn $name(a: &[u64; $len], b: &[u64; $len]) -> Result<([u64; $len], [u64; $len]), ()> {
545 if slice_equal(b, &[0; $len]) { return Err(()); }
548 let mut pow2s = $heap_init;
549 let mut pow2s_count = 0;
550 while slice_greater_than(a, &b_pow) {
551 $pre_push!(pow2s, $len);
552 pow2s[pow2s_count] = b_pow;
554 let double_overflow = double!(b_pow);
555 if double_overflow { break; }
557 let mut quot = [0; $len];
559 let mut pow2 = pow2s_count as isize - 1;
561 let b_pow = pow2s[pow2 as usize];
562 let overflow = double!(quot);
563 debug_assert!(!overflow);
564 if slice_greater_than(&rem, &b_pow) {
565 let (r, carry) = $sub(&rem, &b_pow);
566 debug_assert!(!carry);
572 if slice_equal(&rem, b) {
573 let overflow = add_u64!(quot, 1);
574 debug_assert!(!overflow);
575 Ok((quot, [0; $len]))
583 define_div_rem!(div_rem_2, 2, sub_2, [[0; 2]; 2 * 64], dummy_pre_push, const);
584 define_div_rem!(div_rem_4, 4, sub_4, [[0; 4]; 4 * 64], dummy_pre_push, const); // Uses 8 KiB of stack
585 define_div_rem!(div_rem_6, 6, sub_6, [[0; 6]; 6 * 64], dummy_pre_push, const); // Uses 18 KiB of stack!
586 #[cfg(debug_assertions)]
587 define_div_rem!(div_rem_8, 8, sub_8, [[0; 8]; 8 * 64], dummy_pre_push, const); // Uses 32 KiB of stack!
588 #[cfg(debug_assertions)]
589 define_div_rem!(div_rem_12, 12, sub_12, [[0; 12]; 12 * 64], dummy_pre_push, const); // Uses 72 KiB of stack!
590 define_div_rem!(div_rem_64, 64, sub_64, Vec::new(), vec_pre_push); // Uses up to 2 MiB of heap
591 #[cfg(debug_assertions)]
592 define_div_rem!(div_rem_128, 128, sub_128, Vec::new(), vec_pre_push); // Uses up to 8 MiB of heap
594 macro_rules! define_mod_inv { ($name: ident, $len: expr, $div: ident, $add: ident, $sub_abs: ident, $mul: ident) => {
595 /// Calculates the modular inverse of a $len-64-bit number with respect to the given modulus,
597 const fn $name(a: &[u64; $len], m: &[u64; $len]) -> Result<[u64; $len], ()> {
598 if slice_equal(a, &[0; $len]) || slice_equal(m, &[0; $len]) { return Err(()); }
600 let (mut s, mut old_s) = ([0; $len], [0; $len]);
605 let (mut old_s_neg, mut s_neg) = (false, false);
607 while !slice_equal(&r, &[0; $len]) {
608 let (quot, new_r) = debug_unwrap!($div(&old_r, &r));
610 let new_sa = $mul(", &s);
611 debug_assert!(slice_equal(const_subslice(&new_sa, 0, $len), &[0; $len]), "S overflowed");
612 let (new_s, new_s_neg) = match (old_s_neg, s_neg) {
614 let (new_s, overflow) = $add(&old_s, const_subslice(&new_sa, $len, new_sa.len()));
615 debug_assert!(!overflow);
619 let (new_s, overflow) = $add(&old_s, const_subslice(&new_sa, $len, new_sa.len()));
620 debug_assert!(!overflow);
624 let (new_s, overflow) = $add(&old_s, const_subslice(&new_sa, $len, new_sa.len()));
625 debug_assert!(!overflow);
628 (false, false) => $sub_abs(&old_s, const_subslice(&new_sa, $len, new_sa.len())),
640 // At this point old_r contains our GCD and old_s our first Bézout's identity coefficient.
641 if !slice_equal(const_subslice(&old_r, 0, $len - 1), &[0; $len - 1]) || old_r[$len - 1] != 1 {
644 debug_assert!(slice_greater_than(m, &old_s));
646 let (modinv, underflow) = $sub_abs(m, &old_s);
647 debug_assert!(!underflow);
648 debug_assert!(slice_greater_than(m, &modinv));
657 define_mod_inv!(mod_inv_2, 2, div_rem_2, add_2, sub_abs_2, mul_2);
658 define_mod_inv!(mod_inv_4, 4, div_rem_4, add_4, sub_abs_4, mul_4);
659 define_mod_inv!(mod_inv_6, 6, div_rem_6, add_6, sub_abs_6, mul_6);
661 define_mod_inv!(mod_inv_8, 8, div_rem_8, add_8, sub_abs_8, mul_8);
664 /// Constructs a new [`U4096`] from a variable number of big-endian bytes.
665 pub(super) fn from_be_bytes(bytes: &[u8]) -> Result<U4096, ()> {
666 if bytes.len() > 4096/8 { return Err(()); }
667 let u64s = (bytes.len() + 7) / 8;
668 let mut res = [0; WORD_COUNT_4096];
671 let pos = (u64s - i) * 8;
672 let start = bytes.len().saturating_sub(pos);
673 let end = bytes.len() + 8 - pos;
674 b[8 + start - end..].copy_from_slice(&bytes[start..end]);
675 res[i + WORD_COUNT_4096 - u64s] = u64::from_be_bytes(b);
680 /// Naively multiplies `self` * `b` mod `m`, returning a new [`U4096`].
682 /// Fails iff m is 0 or self or b are greater than m.
683 #[cfg(debug_assertions)]
684 fn mulmod_naive(&self, b: &U4096, m: &U4096) -> Result<U4096, ()> {
685 if m.0 == [0; WORD_COUNT_4096] { return Err(()); }
686 if self > m || b > m { return Err(()); }
688 let mul = mul_64(&self.0, &b.0);
690 let mut m_zeros = [0; 128];
691 m_zeros[WORD_COUNT_4096..].copy_from_slice(&m.0);
692 let (_, rem) = div_rem_128(&mul, &m_zeros)?;
693 let mut res = [0; WORD_COUNT_4096];
694 debug_assert_eq!(&rem[..WORD_COUNT_4096], &[0; WORD_COUNT_4096]);
695 res.copy_from_slice(&rem[WORD_COUNT_4096..]);
699 /// Calculates `self` ^ `exp` mod `m`, returning a new [`U4096`].
701 /// Fails iff m is 0, even, or self or b are greater than m.
702 pub(super) fn expmod_odd_mod(&self, mut exp: u32, m: &U4096) -> Result<U4096, ()> {
703 #![allow(non_camel_case_types)]
705 if m.0 == [0; WORD_COUNT_4096] { return Err(()); }
706 if m.0[WORD_COUNT_4096 - 1] & 1 == 0 { return Err(()); }
707 if self > m { return Err(()); }
709 let mut t = [0; WORD_COUNT_4096];
710 if &m.0[..WORD_COUNT_4096 - 1] == &[0; WORD_COUNT_4096 - 1] && m.0[WORD_COUNT_4096 - 1] == 1 {
713 t[WORD_COUNT_4096 - 1] = 1;
714 if exp == 0 { return Ok(U4096(t)); }
716 // Because m is not even, using 2^4096 as the Montgomery R value is always safe - it is
717 // guaranteed to be co-prime with any non-even integer.
719 type mul_ty = fn(&[u64], &[u64]) -> [u64; WORD_COUNT_4096 * 2];
720 type sqr_ty = fn(&[u64]) -> [u64; WORD_COUNT_4096 * 2];
721 type add_double_ty = fn(&[u64], &[u64]) -> ([u64; WORD_COUNT_4096 * 2], bool);
722 type sub_ty = fn(&[u64], &[u64]) -> ([u64; WORD_COUNT_4096], bool);
723 let (word_count, log_bits, mul, sqr, add_double, sub) =
724 if m.0[..WORD_COUNT_4096 / 2] == [0; WORD_COUNT_4096 / 2] {
725 if m.0[..WORD_COUNT_4096 * 3 / 4] == [0; WORD_COUNT_4096 * 3 / 4] {
726 fn mul_16_subarr(a: &[u64], b: &[u64]) -> [u64; WORD_COUNT_4096 * 2] {
727 debug_assert_eq!(a.len(), WORD_COUNT_4096);
728 debug_assert_eq!(b.len(), WORD_COUNT_4096);
729 debug_assert_eq!(&a[..WORD_COUNT_4096 * 3 / 4], &[0; WORD_COUNT_4096 * 3 / 4]);
730 debug_assert_eq!(&b[..WORD_COUNT_4096 * 3 / 4], &[0; WORD_COUNT_4096 * 3 / 4]);
731 let mut res = [0; WORD_COUNT_4096 * 2];
732 res[WORD_COUNT_4096 + WORD_COUNT_4096 / 2..].copy_from_slice(
733 &mul_16(&a[WORD_COUNT_4096 * 3 / 4..], &b[WORD_COUNT_4096 * 3 / 4..]));
736 fn sqr_16_subarr(a: &[u64]) -> [u64; WORD_COUNT_4096 * 2] {
737 debug_assert_eq!(a.len(), WORD_COUNT_4096);
738 debug_assert_eq!(&a[..WORD_COUNT_4096 * 3 / 4], &[0; WORD_COUNT_4096 * 3 / 4]);
739 let mut res = [0; WORD_COUNT_4096 * 2];
740 res[WORD_COUNT_4096 + WORD_COUNT_4096 / 2..].copy_from_slice(
741 &sqr_16(&a[WORD_COUNT_4096 * 3 / 4..]));
744 fn add_32_subarr(a: &[u64], b: &[u64]) -> ([u64; WORD_COUNT_4096 * 2], bool) {
745 debug_assert_eq!(a.len(), WORD_COUNT_4096 * 2);
746 debug_assert_eq!(b.len(), WORD_COUNT_4096 * 2);
747 debug_assert_eq!(&a[..WORD_COUNT_4096 * 3 / 2], &[0; WORD_COUNT_4096 * 3 / 2]);
748 debug_assert_eq!(&b[..WORD_COUNT_4096 * 3 / 2], &[0; WORD_COUNT_4096 * 3 / 2]);
749 let (add, overflow) = add_32(&a[WORD_COUNT_4096 * 3 / 2..], &b[WORD_COUNT_4096 * 3 / 2..]);
750 let mut res = [0; WORD_COUNT_4096 * 2];
751 res[WORD_COUNT_4096 * 3 / 2..].copy_from_slice(&add);
754 fn sub_16_subarr(a: &[u64], b: &[u64]) -> ([u64; WORD_COUNT_4096], bool) {
755 debug_assert_eq!(a.len(), WORD_COUNT_4096);
756 debug_assert_eq!(b.len(), WORD_COUNT_4096);
757 debug_assert_eq!(&a[..WORD_COUNT_4096 * 3 / 4], &[0; WORD_COUNT_4096 * 3 / 4]);
758 debug_assert_eq!(&b[..WORD_COUNT_4096 * 3 / 4], &[0; WORD_COUNT_4096 * 3 / 4]);
759 let (sub, underflow) = sub_16(&a[WORD_COUNT_4096 * 3 / 4..], &b[WORD_COUNT_4096 * 3 / 4..]);
760 let mut res = [0; WORD_COUNT_4096];
761 res[WORD_COUNT_4096 * 3 / 4..].copy_from_slice(&sub);
764 (16, 10, mul_16_subarr as mul_ty, sqr_16_subarr as sqr_ty, add_32_subarr as add_double_ty, sub_16_subarr as sub_ty)
766 fn mul_32_subarr(a: &[u64], b: &[u64]) -> [u64; WORD_COUNT_4096 * 2] {
767 debug_assert_eq!(a.len(), WORD_COUNT_4096);
768 debug_assert_eq!(b.len(), WORD_COUNT_4096);
769 debug_assert_eq!(&a[..WORD_COUNT_4096 / 2], &[0; WORD_COUNT_4096 / 2]);
770 debug_assert_eq!(&b[..WORD_COUNT_4096 / 2], &[0; WORD_COUNT_4096 / 2]);
771 let mut res = [0; WORD_COUNT_4096 * 2];
772 res[WORD_COUNT_4096..].copy_from_slice(
773 &mul_32(&a[WORD_COUNT_4096 / 2..], &b[WORD_COUNT_4096 / 2..]));
776 fn sqr_32_subarr(a: &[u64]) -> [u64; WORD_COUNT_4096 * 2] {
777 debug_assert_eq!(a.len(), WORD_COUNT_4096);
778 debug_assert_eq!(&a[..WORD_COUNT_4096 / 2], &[0; WORD_COUNT_4096 / 2]);
779 let mut res = [0; WORD_COUNT_4096 * 2];
780 res[WORD_COUNT_4096..].copy_from_slice(
781 &sqr_32(&a[WORD_COUNT_4096 / 2..]));
784 fn add_64_subarr(a: &[u64], b: &[u64]) -> ([u64; WORD_COUNT_4096 * 2], bool) {
785 debug_assert_eq!(a.len(), WORD_COUNT_4096 * 2);
786 debug_assert_eq!(b.len(), WORD_COUNT_4096 * 2);
787 debug_assert_eq!(&a[..WORD_COUNT_4096], &[0; WORD_COUNT_4096]);
788 debug_assert_eq!(&b[..WORD_COUNT_4096], &[0; WORD_COUNT_4096]);
789 let (add, overflow) = add_64(&a[WORD_COUNT_4096..], &b[WORD_COUNT_4096..]);
790 let mut res = [0; WORD_COUNT_4096 * 2];
791 res[WORD_COUNT_4096..].copy_from_slice(&add);
794 fn sub_32_subarr(a: &[u64], b: &[u64]) -> ([u64; WORD_COUNT_4096], bool) {
795 debug_assert_eq!(a.len(), WORD_COUNT_4096);
796 debug_assert_eq!(b.len(), WORD_COUNT_4096);
797 debug_assert_eq!(&a[..WORD_COUNT_4096 / 2], &[0; WORD_COUNT_4096 / 2]);
798 debug_assert_eq!(&b[..WORD_COUNT_4096 / 2], &[0; WORD_COUNT_4096 / 2]);
799 let (sub, underflow) = sub_32(&a[WORD_COUNT_4096 / 2..], &b[WORD_COUNT_4096 / 2..]);
800 let mut res = [0; WORD_COUNT_4096];
801 res[WORD_COUNT_4096 / 2..].copy_from_slice(&sub);
804 (32, 11, mul_32_subarr as mul_ty, sqr_32_subarr as sqr_ty, add_64_subarr as add_double_ty, sub_32_subarr as sub_ty)
807 (64, 12, mul_64 as mul_ty, sqr_64 as sqr_ty, add_128 as add_double_ty, sub_64 as sub_ty)
810 let mut r = [0; WORD_COUNT_4096 * 2];
811 r[WORD_COUNT_4096 * 2 - word_count - 1] = 1;
813 let mut m_inv_pos = [0; WORD_COUNT_4096];
814 m_inv_pos[WORD_COUNT_4096 - 1] = 1;
815 let mut two = [0; WORD_COUNT_4096];
816 two[WORD_COUNT_4096 - 1] = 2;
817 for _ in 0..log_bits {
818 let mut m_m_inv = mul(&m_inv_pos, &m.0);
819 m_m_inv[..WORD_COUNT_4096 * 2 - word_count].fill(0);
820 let m_inv = mul(&sub(&two, &m_m_inv[WORD_COUNT_4096..]).0, &m_inv_pos);
821 m_inv_pos[WORD_COUNT_4096 - word_count..].copy_from_slice(&m_inv[WORD_COUNT_4096 * 2 - word_count..]);
823 m_inv_pos[..WORD_COUNT_4096 - word_count].fill(0);
825 // We want the negative modular inverse of m mod R, so subtract m_inv from R.
826 let mut m_inv = m_inv_pos;
828 m_inv[..WORD_COUNT_4096 - word_count].fill(0);
829 debug_assert_eq!(&mul(&m_inv, &m.0)[WORD_COUNT_4096 * 2 - word_count..],
831 &[0xffff_ffff_ffff_ffff; WORD_COUNT_4096][WORD_COUNT_4096 - word_count..]);
833 debug_assert_eq!(&m_inv[..WORD_COUNT_4096 - word_count], &[0; WORD_COUNT_4096][..WORD_COUNT_4096 - word_count]);
835 let mont_reduction = |mu: [u64; WORD_COUNT_4096 * 2]| -> [u64; WORD_COUNT_4096] {
836 debug_assert_eq!(&mu[..WORD_COUNT_4096 * 2 - word_count * 2],
837 &[0; WORD_COUNT_4096 * 2][..WORD_COUNT_4096 * 2 - word_count * 2]);
838 let mut mu_mod_r = [0; WORD_COUNT_4096];
839 mu_mod_r[WORD_COUNT_4096 - word_count..].copy_from_slice(&mu[WORD_COUNT_4096 * 2 - word_count..]);
840 let mut v = mul(&mu_mod_r, &m_inv);
841 v[..WORD_COUNT_4096 * 2 - word_count].fill(0); // mod R
842 let t0 = mul(&v[WORD_COUNT_4096..], &m.0);
843 let (t1, t1_extra_bit) = add_double(&t0, &mu);
844 let mut t1_on_r = [0; WORD_COUNT_4096];
845 debug_assert_eq!(&t1[WORD_COUNT_4096 * 2 - word_count..], &[0; WORD_COUNT_4096][WORD_COUNT_4096 - word_count..],
846 "t1 should be divisible by r");
847 t1_on_r[WORD_COUNT_4096 - word_count..].copy_from_slice(&t1[WORD_COUNT_4096 * 2 - word_count * 2..WORD_COUNT_4096 * 2 - word_count]);
848 if t1_extra_bit || t1_on_r >= m.0 {
850 (t1_on_r, underflow) = sub(&t1_on_r, &m.0);
851 debug_assert_eq!(t1_extra_bit, underflow);
856 // Calculate R^2 mod m as ((2^DOUBLES * R) mod m)^(log_bits - LOG2_DOUBLES) mod R
857 let mut r_minus_one = [0xffff_ffff_ffff_ffffu64; WORD_COUNT_4096];
858 r_minus_one[..WORD_COUNT_4096 - word_count].fill(0);
859 // While we do a full div here, in general R should be less than 2x m (assuming the RSA
860 // modulus used its full bit range and is 1024, 2048, or 4096 bits), so it should be cheap.
861 // In cases with a nonstandard RSA modulus we may end up being pretty slow here, but we'll
863 // If we ever find a problem with this we should reduce R to be tigher on m, as we're
864 // wasting extra bits of calculation if R is too far from m.
865 let (_, mut r_mod_m) = debug_unwrap!(div_rem_64(&r_minus_one, &m.0));
866 let r_mod_m_overflow = add_u64!(r_mod_m, 1);
867 if r_mod_m_overflow || r_mod_m >= m.0 {
868 (r_mod_m, _) = sub_64(&r_mod_m, &m.0);
871 let mut r2_mod_m: [u64; 64] = r_mod_m;
872 const DOUBLES: usize = 32;
873 const LOG2_DOUBLES: usize = 5;
875 for _ in 0..DOUBLES {
876 let overflow = double!(r2_mod_m);
877 if overflow || r2_mod_m > m.0 {
878 (r2_mod_m, _) = sub_64(&r2_mod_m, &m.0);
881 for _ in 0..log_bits - LOG2_DOUBLES {
882 r2_mod_m = mont_reduction(sqr(&r2_mod_m));
884 // Clear excess high bits
885 for (m_limb, r2_limb) in m.0.iter().zip(r2_mod_m.iter_mut()) {
886 let clear_bits = m_limb.leading_zeros();
887 if clear_bits == 0 { break; }
888 *r2_limb &= !(0xffff_ffff_ffff_ffffu64 << (64 - clear_bits));
889 if *m_limb != 0 { break; }
891 debug_assert!(r2_mod_m < m.0);
893 // Calculate t * R and a * R as mont multiplications by R^2 mod m
894 let mut tr = mont_reduction(mul(&r2_mod_m, &t));
895 let mut ar = mont_reduction(mul(&r2_mod_m, &self.0));
897 #[cfg(debug_assertions)] {
898 debug_assert_eq!(r2_mod_m, U4096(r_mod_m).mulmod_naive(&U4096(r_mod_m), &m).unwrap().0);
899 debug_assert_eq!(&tr, &U4096(t).mulmod_naive(&U4096(r_mod_m), &m).unwrap().0);
900 debug_assert_eq!(&ar, &self.mulmod_naive(&U4096(r_mod_m), &m).unwrap().0);
905 tr = mont_reduction(mul(&tr, &ar));
908 ar = mont_reduction(sqr(&ar));
911 ar = mont_reduction(mul(&ar, &tr));
912 let mut resr = [0; WORD_COUNT_4096 * 2];
913 resr[WORD_COUNT_4096..].copy_from_slice(&ar);
914 Ok(U4096(mont_reduction(resr)))
918 const fn u64_from_bytes_a_panicking(b: &[u8]) -> u64 {
920 [a, b, c, d, e, f, g, h, ..] => {
921 ((*a as u64) << 8*7) |
922 ((*b as u64) << 8*6) |
923 ((*c as u64) << 8*5) |
924 ((*d as u64) << 8*4) |
925 ((*e as u64) << 8*3) |
926 ((*f as u64) << 8*2) |
927 ((*g as u64) << 8*1) |
934 const fn u64_from_bytes_b_panicking(b: &[u8]) -> u64 {
936 [_, _, _, _, _, _, _, _,
937 a, b, c, d, e, f, g, h, ..] => {
938 ((*a as u64) << 8*7) |
939 ((*b as u64) << 8*6) |
940 ((*c as u64) << 8*5) |
941 ((*d as u64) << 8*4) |
942 ((*e as u64) << 8*3) |
943 ((*f as u64) << 8*2) |
944 ((*g as u64) << 8*1) |
951 const fn u64_from_bytes_c_panicking(b: &[u8]) -> u64 {
953 [_, _, _, _, _, _, _, _,
954 _, _, _, _, _, _, _, _,
955 a, b, c, d, e, f, g, h, ..] => {
956 ((*a as u64) << 8*7) |
957 ((*b as u64) << 8*6) |
958 ((*c as u64) << 8*5) |
959 ((*d as u64) << 8*4) |
960 ((*e as u64) << 8*3) |
961 ((*f as u64) << 8*2) |
962 ((*g as u64) << 8*1) |
969 const fn u64_from_bytes_d_panicking(b: &[u8]) -> u64 {
971 [_, _, _, _, _, _, _, _,
972 _, _, _, _, _, _, _, _,
973 _, _, _, _, _, _, _, _,
974 a, b, c, d, e, f, g, h, ..] => {
975 ((*a as u64) << 8*7) |
976 ((*b as u64) << 8*6) |
977 ((*c as u64) << 8*5) |
978 ((*d as u64) << 8*4) |
979 ((*e as u64) << 8*3) |
980 ((*f as u64) << 8*2) |
981 ((*g as u64) << 8*1) |
988 const fn u64_from_bytes_e_panicking(b: &[u8]) -> u64 {
990 [_, _, _, _, _, _, _, _,
991 _, _, _, _, _, _, _, _,
992 _, _, _, _, _, _, _, _,
993 _, _, _, _, _, _, _, _,
994 a, b, c, d, e, f, g, h, ..] => {
995 ((*a as u64) << 8*7) |
996 ((*b as u64) << 8*6) |
997 ((*c as u64) << 8*5) |
998 ((*d as u64) << 8*4) |
999 ((*e as u64) << 8*3) |
1000 ((*f as u64) << 8*2) |
1001 ((*g as u64) << 8*1) |
1002 ((*h as u64) << 8*0)
1008 const fn u64_from_bytes_f_panicking(b: &[u8]) -> u64 {
1010 [_, _, _, _, _, _, _, _,
1011 _, _, _, _, _, _, _, _,
1012 _, _, _, _, _, _, _, _,
1013 _, _, _, _, _, _, _, _,
1014 _, _, _, _, _, _, _, _,
1015 a, b, c, d, e, f, g, h, ..] => {
1016 ((*a as u64) << 8*7) |
1017 ((*b as u64) << 8*6) |
1018 ((*c as u64) << 8*5) |
1019 ((*d as u64) << 8*4) |
1020 ((*e as u64) << 8*3) |
1021 ((*f as u64) << 8*2) |
1022 ((*g as u64) << 8*1) |
1023 ((*h as u64) << 8*0)
1030 /// Constructs a new [`U256`] from a variable number of big-endian bytes.
1031 pub(super) fn from_be_bytes(bytes: &[u8]) -> Result<U256, ()> {
1032 if bytes.len() > 256/8 { return Err(()); }
1033 let u64s = (bytes.len() + 7) / 8;
1034 let mut res = [0; WORD_COUNT_256];
1037 let pos = (u64s - i) * 8;
1038 let start = bytes.len().saturating_sub(pos);
1039 let end = bytes.len() + 8 - pos;
1040 b[8 + start - end..].copy_from_slice(&bytes[start..end]);
1041 res[i + WORD_COUNT_256 - u64s] = u64::from_be_bytes(b);
1046 /// Constructs a new [`U256`] from a fixed number of big-endian bytes.
1047 pub(super) const fn from_32_be_bytes_panicking(bytes: &[u8; 32]) -> U256 {
1049 u64_from_bytes_a_panicking(bytes),
1050 u64_from_bytes_b_panicking(bytes),
1051 u64_from_bytes_c_panicking(bytes),
1052 u64_from_bytes_d_panicking(bytes),
1057 pub(super) const fn zero() -> U256 { U256([0, 0, 0, 0]) }
1058 pub(super) const fn one() -> U256 { U256([0, 0, 0, 1]) }
1059 pub(super) const fn three() -> U256 { U256([0, 0, 0, 3]) }
1062 impl<M: PrimeModulus<U256>> U256Mod<M> {
1063 const fn mont_reduction(mu: [u64; 8]) -> Self {
1064 #[cfg(debug_assertions)] {
1065 // Check NEGATIVE_PRIME_INV_MOD_R is correct. Since this is all const, the compiler
1066 // should be able to do it at compile time alone.
1067 let minus_one_mod_r = mul_4(&M::PRIME.0, &M::NEGATIVE_PRIME_INV_MOD_R.0);
1068 assert!(slice_equal(const_subslice(&minus_one_mod_r, 4, 8), &[0xffff_ffff_ffff_ffff; 4]));
1071 #[cfg(debug_assertions)] {
1072 // Check R_SQUARED_MOD_PRIME is correct. Since this is all const, the compiler
1073 // should be able to do it at compile time alone.
1074 let r_minus_one = [0xffff_ffff_ffff_ffff; 4];
1075 let (mut r_mod_prime, _) = sub_4(&r_minus_one, &M::PRIME.0);
1076 add_u64!(r_mod_prime, 1);
1077 let r_squared = sqr_4(&r_mod_prime);
1078 let mut prime_extended = [0; 8];
1079 let prime = M::PRIME.0;
1080 copy_from_slice!(prime_extended, 4, 8, prime);
1081 let (_, r_squared_mod_prime) = if let Ok(v) = div_rem_8(&r_squared, &prime_extended) { v } else { panic!() };
1082 assert!(slice_greater_than(&prime_extended, &r_squared_mod_prime));
1083 assert!(slice_equal(const_subslice(&r_squared_mod_prime, 4, 8), &M::R_SQUARED_MOD_PRIME.0));
1086 let mu_mod_r = const_subslice(&mu, 4, 8);
1087 let mut v = mul_4(&mu_mod_r, &M::NEGATIVE_PRIME_INV_MOD_R.0);
1088 const ZEROS: &[u64; 4] = &[0; 4];
1089 copy_from_slice!(v, 0, 4, ZEROS); // mod R
1090 let t0 = mul_4(const_subslice(&v, 4, 8), &M::PRIME.0);
1091 let (t1, t1_extra_bit) = add_8(&t0, &mu);
1092 let t1_on_r = const_subslice(&t1, 0, 4);
1093 let mut res = [0; 4];
1094 if t1_extra_bit || slice_greater_than(&t1_on_r, &M::PRIME.0) {
1096 (res, underflow) = sub_4(&t1_on_r, &M::PRIME.0);
1097 debug_assert!(t1_extra_bit == underflow);
1099 copy_from_slice!(res, 0, 4, t1_on_r);
1101 Self(U256(res), PhantomData)
1104 pub(super) const fn from_u256_panicking(v: U256) -> Self {
1105 assert!(v.0[0] <= M::PRIME.0[0]);
1106 if v.0[0] == M::PRIME.0[0] {
1107 assert!(v.0[1] <= M::PRIME.0[1]);
1108 if v.0[1] == M::PRIME.0[1] {
1109 assert!(v.0[2] <= M::PRIME.0[2]);
1110 if v.0[2] == M::PRIME.0[2] {
1111 assert!(v.0[3] < M::PRIME.0[3]);
1115 assert!(M::PRIME.0[0] != 0 || M::PRIME.0[1] != 0 || M::PRIME.0[2] != 0 || M::PRIME.0[3] != 0);
1116 Self::mont_reduction(mul_4(&M::R_SQUARED_MOD_PRIME.0, &v.0))
1119 pub(super) fn from_u256(mut v: U256) -> Self {
1120 debug_assert!(M::PRIME.0 != [0; 4]);
1121 debug_assert!(M::PRIME.0[0] > (1 << 63), "PRIME should have the top bit set");
1122 while v >= M::PRIME {
1123 let (new_v, spurious_underflow) = sub_4(&v.0, &M::PRIME.0);
1124 debug_assert!(!spurious_underflow);
1127 Self::mont_reduction(mul_4(&M::R_SQUARED_MOD_PRIME.0, &v.0))
1130 pub(super) fn from_modinv_of(v: U256) -> Result<Self, ()> {
1131 Ok(Self::from_u256(U256(mod_inv_4(&v.0, &M::PRIME.0)?)))
1134 /// Multiplies `self` * `b` mod `m`.
1136 /// Panics if `self`'s modulus is not equal to `b`'s
1137 pub(super) fn mul(&self, b: &Self) -> Self {
1138 Self::mont_reduction(mul_4(&self.0.0, &b.0.0))
1141 /// Doubles `self` mod `m`.
1142 pub(super) fn double(&self) -> Self {
1143 let mut res = self.0.0;
1144 let overflow = double!(res);
1145 if overflow || !slice_greater_than(&M::PRIME.0, &res) {
1147 (res, underflow) = sub_4(&res, &M::PRIME.0);
1148 debug_assert_eq!(overflow, underflow);
1150 Self(U256(res), PhantomData)
1153 /// Multiplies `self` by 3 mod `m`.
1154 pub(super) fn times_three(&self) -> Self {
1155 // TODO: Optimize this a lot
1156 self.mul(&U256Mod::from_u256(U256::three()))
1159 /// Multiplies `self` by 4 mod `m`.
1160 pub(super) fn times_four(&self) -> Self {
1161 // TODO: Optimize this somewhat?
1162 self.double().double()
1165 /// Multiplies `self` by 8 mod `m`.
1166 pub(super) fn times_eight(&self) -> Self {
1167 // TODO: Optimize this somewhat?
1168 self.double().double().double()
1171 /// Multiplies `self` by 8 mod `m`.
1172 pub(super) fn square(&self) -> Self {
1173 Self::mont_reduction(sqr_4(&self.0.0))
1176 /// Subtracts `b` from `self` % `m`.
1177 pub(super) fn sub(&self, b: &Self) -> Self {
1178 let (mut val, underflow) = sub_4(&self.0.0, &b.0.0);
1181 (val, overflow) = add_4(&val, &M::PRIME.0);
1182 debug_assert_eq!(overflow, underflow);
1184 Self(U256(val), PhantomData)
1187 /// Adds `b` to `self` % `m`.
1188 pub(super) fn add(&self, b: &Self) -> Self {
1189 let (mut val, overflow) = add_4(&self.0.0, &b.0.0);
1190 if overflow || !slice_greater_than(&M::PRIME.0, &val) {
1192 (val, underflow) = sub_4(&val, &M::PRIME.0);
1193 debug_assert_eq!(overflow, underflow);
1195 Self(U256(val), PhantomData)
1198 /// Returns the underlying [`U256`].
1199 pub(super) fn into_u256(self) -> U256 {
1200 let mut expanded_self = [0; 8];
1201 expanded_self[4..].copy_from_slice(&self.0.0);
1202 Self::mont_reduction(expanded_self).0
1207 /// Constructs a new [`U384`] from a variable number of big-endian bytes.
1208 pub(super) fn from_be_bytes(bytes: &[u8]) -> Result<U384, ()> {
1209 if bytes.len() > 384/8 { return Err(()); }
1210 let u64s = (bytes.len() + 7) / 8;
1211 let mut res = [0; WORD_COUNT_384];
1214 let pos = (u64s - i) * 8;
1215 let start = bytes.len().saturating_sub(pos);
1216 let end = bytes.len() + 8 - pos;
1217 b[8 + start - end..].copy_from_slice(&bytes[start..end]);
1218 res[i + WORD_COUNT_384 - u64s] = u64::from_be_bytes(b);
1223 /// Constructs a new [`U384`] from a fixed number of big-endian bytes.
1224 pub(super) const fn from_48_be_bytes_panicking(bytes: &[u8; 48]) -> U384 {
1226 u64_from_bytes_a_panicking(bytes),
1227 u64_from_bytes_b_panicking(bytes),
1228 u64_from_bytes_c_panicking(bytes),
1229 u64_from_bytes_d_panicking(bytes),
1230 u64_from_bytes_e_panicking(bytes),
1231 u64_from_bytes_f_panicking(bytes),
1236 pub(super) const fn zero() -> U384 { U384([0, 0, 0, 0, 0, 0]) }
1237 pub(super) const fn one() -> U384 { U384([0, 0, 0, 0, 0, 1]) }
1238 pub(super) const fn three() -> U384 { U384([0, 0, 0, 0, 0, 3]) }
1241 impl<M: PrimeModulus<U384>> U384Mod<M> {
1242 const fn mont_reduction(mu: [u64; 12]) -> Self {
1243 #[cfg(debug_assertions)] {
1244 // Check NEGATIVE_PRIME_INV_MOD_R is correct. Since this is all const, the compiler
1245 // should be able to do it at compile time alone.
1246 let minus_one_mod_r = mul_6(&M::PRIME.0, &M::NEGATIVE_PRIME_INV_MOD_R.0);
1247 assert!(slice_equal(const_subslice(&minus_one_mod_r, 6, 12), &[0xffff_ffff_ffff_ffff; 6]));
1250 #[cfg(debug_assertions)] {
1251 // Check R_SQUARED_MOD_PRIME is correct. Since this is all const, the compiler
1252 // should be able to do it at compile time alone.
1253 let r_minus_one = [0xffff_ffff_ffff_ffff; 6];
1254 let (mut r_mod_prime, _) = sub_6(&r_minus_one, &M::PRIME.0);
1255 add_u64!(r_mod_prime, 1);
1256 let r_squared = sqr_6(&r_mod_prime);
1257 let mut prime_extended = [0; 12];
1258 let prime = M::PRIME.0;
1259 copy_from_slice!(prime_extended, 6, 12, prime);
1260 let (_, r_squared_mod_prime) = if let Ok(v) = div_rem_12(&r_squared, &prime_extended) { v } else { panic!() };
1261 assert!(slice_greater_than(&prime_extended, &r_squared_mod_prime));
1262 assert!(slice_equal(const_subslice(&r_squared_mod_prime, 6, 12), &M::R_SQUARED_MOD_PRIME.0));
1265 let mu_mod_r = const_subslice(&mu, 6, 12);
1266 let mut v = mul_6(&mu_mod_r, &M::NEGATIVE_PRIME_INV_MOD_R.0);
1267 const ZEROS: &[u64; 6] = &[0; 6];
1268 copy_from_slice!(v, 0, 6, ZEROS); // mod R
1269 let t0 = mul_6(const_subslice(&v, 6, 12), &M::PRIME.0);
1270 let (t1, t1_extra_bit) = add_12(&t0, &mu);
1271 let t1_on_r = const_subslice(&t1, 0, 6);
1272 let mut res = [0; 6];
1273 if t1_extra_bit || slice_greater_than(&t1_on_r, &M::PRIME.0) {
1275 (res, underflow) = sub_6(&t1_on_r, &M::PRIME.0);
1276 debug_assert!(t1_extra_bit == underflow);
1278 copy_from_slice!(res, 0, 6, t1_on_r);
1280 Self(U384(res), PhantomData)
1283 pub(super) const fn from_u384_panicking(v: U384) -> Self {
1284 assert!(v.0[0] <= M::PRIME.0[0]);
1285 if v.0[0] == M::PRIME.0[0] {
1286 assert!(v.0[1] <= M::PRIME.0[1]);
1287 if v.0[1] == M::PRIME.0[1] {
1288 assert!(v.0[2] <= M::PRIME.0[2]);
1289 if v.0[2] == M::PRIME.0[2] {
1290 assert!(v.0[3] <= M::PRIME.0[3]);
1291 if v.0[3] == M::PRIME.0[3] {
1292 assert!(v.0[4] <= M::PRIME.0[4]);
1293 if v.0[4] == M::PRIME.0[4] {
1294 assert!(v.0[5] < M::PRIME.0[5]);
1300 assert!(M::PRIME.0[0] != 0 || M::PRIME.0[1] != 0 || M::PRIME.0[2] != 0
1301 || M::PRIME.0[3] != 0|| M::PRIME.0[4] != 0|| M::PRIME.0[5] != 0);
1302 Self::mont_reduction(mul_6(&M::R_SQUARED_MOD_PRIME.0, &v.0))
1305 pub(super) fn from_u384(mut v: U384) -> Self {
1306 debug_assert!(M::PRIME.0 != [0; 6]);
1307 debug_assert!(M::PRIME.0[0] > (1 << 63), "PRIME should have the top bit set");
1308 while v >= M::PRIME {
1309 let (new_v, spurious_underflow) = sub_6(&v.0, &M::PRIME.0);
1310 debug_assert!(!spurious_underflow);
1313 Self::mont_reduction(mul_6(&M::R_SQUARED_MOD_PRIME.0, &v.0))
1316 pub(super) fn from_modinv_of(v: U384) -> Result<Self, ()> {
1317 Ok(Self::from_u384(U384(mod_inv_6(&v.0, &M::PRIME.0)?)))
1320 /// Multiplies `self` * `b` mod `m`.
1322 /// Panics if `self`'s modulus is not equal to `b`'s
1323 pub(super) fn mul(&self, b: &Self) -> Self {
1324 Self::mont_reduction(mul_6(&self.0.0, &b.0.0))
1327 /// Doubles `self` mod `m`.
1328 pub(super) fn double(&self) -> Self {
1329 let mut res = self.0.0;
1330 let overflow = double!(res);
1331 if overflow || !slice_greater_than(&M::PRIME.0, &res) {
1333 (res, underflow) = sub_6(&res, &M::PRIME.0);
1334 debug_assert_eq!(overflow, underflow);
1336 Self(U384(res), PhantomData)
1339 /// Multiplies `self` by 3 mod `m`.
1340 pub(super) fn times_three(&self) -> Self {
1341 // TODO: Optimize this a lot
1342 self.mul(&U384Mod::from_u384(U384::three()))
1345 /// Multiplies `self` by 4 mod `m`.
1346 pub(super) fn times_four(&self) -> Self {
1347 // TODO: Optimize this somewhat?
1348 self.double().double()
1351 /// Multiplies `self` by 8 mod `m`.
1352 pub(super) fn times_eight(&self) -> Self {
1353 // TODO: Optimize this somewhat?
1354 self.double().double().double()
1357 /// Multiplies `self` by 8 mod `m`.
1358 pub(super) fn square(&self) -> Self {
1359 Self::mont_reduction(sqr_6(&self.0.0))
1362 /// Subtracts `b` from `self` % `m`.
1363 pub(super) fn sub(&self, b: &Self) -> Self {
1364 let (mut val, underflow) = sub_6(&self.0.0, &b.0.0);
1367 (val, overflow) = add_6(&val, &M::PRIME.0);
1368 debug_assert_eq!(overflow, underflow);
1370 Self(U384(val), PhantomData)
1373 /// Adds `b` to `self` % `m`.
1374 pub(super) fn add(&self, b: &Self) -> Self {
1375 let (mut val, overflow) = add_6(&self.0.0, &b.0.0);
1376 if overflow || !slice_greater_than(&M::PRIME.0, &val) {
1378 (val, underflow) = sub_6(&val, &M::PRIME.0);
1379 debug_assert_eq!(overflow, underflow);
1381 Self(U384(val), PhantomData)
1384 /// Returns the underlying [`U384`].
1385 pub(super) fn into_u384(self) -> U384 {
1386 let mut expanded_self = [0; 12];
1387 expanded_self[6..].copy_from_slice(&self.0.0);
1388 Self::mont_reduction(expanded_self).0
1397 impl PrimeModulus<U256> for P256 {
1398 const PRIME: U256 = U256::from_32_be_bytes_panicking(&hex_lit::hex!(
1399 "ffffffff00000001000000000000000000000000ffffffffffffffffffffffff"));
1400 const R_SQUARED_MOD_PRIME: U256 = U256::from_32_be_bytes_panicking(&hex_lit::hex!(
1401 "00000004fffffffdfffffffffffffffefffffffbffffffff0000000000000003"));
1402 const NEGATIVE_PRIME_INV_MOD_R: U256 = U256::from_32_be_bytes_panicking(&hex_lit::hex!(
1403 "ffffffff00000002000000000000000000000001000000000000000000000001"));
1407 impl PrimeModulus<U384> for P384 {
1408 const PRIME: U384 = U384::from_48_be_bytes_panicking(&hex_lit::hex!(
1409 "fffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff"));
1410 const R_SQUARED_MOD_PRIME: U384 = U384::from_48_be_bytes_panicking(&hex_lit::hex!(
1411 "000000000000000000000000000000010000000200000000fffffffe000000000000000200000000fffffffe00000001"));
1412 const NEGATIVE_PRIME_INV_MOD_R: U384 = U384::from_48_be_bytes_panicking(&hex_lit::hex!(
1413 "00000014000000140000000c00000002fffffffcfffffffafffffffbfffffffe00000000000000010000000100000001"));
1420 /// Read some bytes and use them to test bigint math by comparing results against the `ibig` crate.
1421 pub fn fuzz_math(input: &[u8]) {
1422 if input.len() < 32 || input.len() % 16 != 0 { return; }
1423 let split = core::cmp::min(input.len() / 2, 512);
1424 let (a, b) = input.split_at(core::cmp::min(input.len() / 2, 512));
1425 let b = &b[..split];
1427 let ai = ibig::UBig::from_be_bytes(&a);
1428 let bi = ibig::UBig::from_be_bytes(&b);
1430 let mut a_u64s = Vec::with_capacity(split / 8);
1431 for chunk in a.chunks(8) {
1432 a_u64s.push(u64::from_be_bytes(chunk.try_into().unwrap()));
1434 let mut b_u64s = Vec::with_capacity(split / 8);
1435 for chunk in b.chunks(8) {
1436 b_u64s.push(u64::from_be_bytes(chunk.try_into().unwrap()));
1439 macro_rules! test { ($mul: ident, $sqr: ident, $add: ident, $sub: ident, $div_rem: ident, $mod_inv: ident) => {
1440 let res = $mul(&a_u64s, &b_u64s);
1441 let mut res_bytes = Vec::with_capacity(input.len() / 2);
1443 res_bytes.extend_from_slice(&i.to_be_bytes());
1445 assert_eq!(ibig::UBig::from_be_bytes(&res_bytes), ai.clone() * bi.clone());
1447 debug_assert_eq!($mul(&a_u64s, &a_u64s), $sqr(&a_u64s));
1448 debug_assert_eq!($mul(&b_u64s, &b_u64s), $sqr(&b_u64s));
1450 let (res, carry) = $add(&a_u64s, &b_u64s);
1451 let mut res_bytes = Vec::with_capacity(input.len() / 2 + 1);
1452 if carry { res_bytes.push(1); } else { res_bytes.push(0); }
1454 res_bytes.extend_from_slice(&i.to_be_bytes());
1456 assert_eq!(ibig::UBig::from_be_bytes(&res_bytes), ai.clone() + bi.clone());
1458 let mut add_u64s = a_u64s.clone();
1459 let carry = add_u64!(add_u64s, 1);
1460 let mut res_bytes = Vec::with_capacity(input.len() / 2 + 1);
1461 if carry { res_bytes.push(1); } else { res_bytes.push(0); }
1462 for i in &add_u64s {
1463 res_bytes.extend_from_slice(&i.to_be_bytes());
1465 assert_eq!(ibig::UBig::from_be_bytes(&res_bytes), ai.clone() + 1);
1467 let mut double_u64s = b_u64s.clone();
1468 let carry = double!(double_u64s);
1469 let mut res_bytes = Vec::with_capacity(input.len() / 2 + 1);
1470 if carry { res_bytes.push(1); } else { res_bytes.push(0); }
1471 for i in &double_u64s {
1472 res_bytes.extend_from_slice(&i.to_be_bytes());
1474 assert_eq!(ibig::UBig::from_be_bytes(&res_bytes), bi.clone() * 2);
1476 let (quot, rem) = if let Ok(res) =
1477 $div_rem(&a_u64s[..].try_into().unwrap(), &b_u64s[..].try_into().unwrap()) {
1480 let mut quot_bytes = Vec::with_capacity(input.len() / 2);
1482 quot_bytes.extend_from_slice(&i.to_be_bytes());
1484 let mut rem_bytes = Vec::with_capacity(input.len() / 2);
1486 rem_bytes.extend_from_slice(&i.to_be_bytes());
1488 let (quoti, remi) = ibig::ops::DivRem::div_rem(ai.clone(), &bi);
1489 assert_eq!(ibig::UBig::from_be_bytes("_bytes), quoti);
1490 assert_eq!(ibig::UBig::from_be_bytes(&rem_bytes), remi);
1492 if ai != ibig::UBig::from(0u32) { // ibig provides a spurious modular inverse for 0
1493 let ring = ibig::modular::ModuloRing::new(&bi);
1494 let ar = ring.from(ai.clone());
1495 let invi = ar.inverse().map(|i| i.residue());
1497 if let Ok(modinv) = $mod_inv(&a_u64s[..].try_into().unwrap(), &b_u64s[..].try_into().unwrap()) {
1498 let mut modinv_bytes = Vec::with_capacity(input.len() / 2);
1500 modinv_bytes.extend_from_slice(&i.to_be_bytes());
1502 assert_eq!(invi.unwrap(), ibig::UBig::from_be_bytes(&modinv_bytes));
1504 assert!(invi.is_none());
1509 macro_rules! test_mod { ($amodp: expr, $bmodp: expr, $PRIME: expr, $len: expr, $into: ident, $div_rem_double: ident, $div_rem: ident, $mul: ident, $add: ident, $sub: ident) => {
1510 // Test the U256/U384Mod wrapper, which operates in Montgomery representation
1511 let mut p_extended = [0; $len * 2];
1512 p_extended[$len..].copy_from_slice(&$PRIME);
1514 let amodp_squared = $div_rem_double(&$mul(&a_u64s, &a_u64s), &p_extended).unwrap().1;
1515 assert_eq!(&amodp_squared[..$len], &[0; $len]);
1516 assert_eq!(&$amodp.square().$into().0, &amodp_squared[$len..]);
1518 let abmodp = $div_rem_double(&$mul(&a_u64s, &b_u64s), &p_extended).unwrap().1;
1519 assert_eq!(&abmodp[..$len], &[0; $len]);
1520 assert_eq!(&$amodp.mul(&$bmodp).$into().0, &abmodp[$len..]);
1522 let (aplusb, aplusb_overflow) = $add(&a_u64s, &b_u64s);
1523 let mut aplusb_extended = [0; $len * 2];
1524 aplusb_extended[$len..].copy_from_slice(&aplusb);
1525 if aplusb_overflow { aplusb_extended[$len - 1] = 1; }
1526 let aplusbmodp = $div_rem_double(&aplusb_extended, &p_extended).unwrap().1;
1527 assert_eq!(&aplusbmodp[..$len], &[0; $len]);
1528 assert_eq!(&$amodp.add(&$bmodp).$into().0, &aplusbmodp[$len..]);
1530 let (mut aminusb, aminusb_underflow) = $sub(&a_u64s, &b_u64s);
1531 if aminusb_underflow {
1533 (aminusb, overflow) = $add(&aminusb, &$PRIME);
1535 (aminusb, overflow) = $add(&aminusb, &$PRIME);
1539 let aminusbmodp = $div_rem(&aminusb, &$PRIME).unwrap().1;
1540 assert_eq!(&$amodp.sub(&$bmodp).$into().0, &aminusbmodp);
1543 if a_u64s.len() == 2 {
1544 test!(mul_2, sqr_2, add_2, sub_2, div_rem_2, mod_inv_2);
1545 } else if a_u64s.len() == 4 {
1546 test!(mul_4, sqr_4, add_4, sub_4, div_rem_4, mod_inv_4);
1547 let amodp = U256Mod::<fuzz_moduli::P256>::from_u256(U256(a_u64s[..].try_into().unwrap()));
1548 let bmodp = U256Mod::<fuzz_moduli::P256>::from_u256(U256(b_u64s[..].try_into().unwrap()));
1549 test_mod!(amodp, bmodp, fuzz_moduli::P256::PRIME.0, 4, into_u256, div_rem_8, div_rem_4, mul_4, add_4, sub_4);
1550 } else if a_u64s.len() == 6 {
1551 test!(mul_6, sqr_6, add_6, sub_6, div_rem_6, mod_inv_6);
1552 let amodp = U384Mod::<fuzz_moduli::P384>::from_u384(U384(a_u64s[..].try_into().unwrap()));
1553 let bmodp = U384Mod::<fuzz_moduli::P384>::from_u384(U384(b_u64s[..].try_into().unwrap()));
1554 test_mod!(amodp, bmodp, fuzz_moduli::P384::PRIME.0, 6, into_u384, div_rem_12, div_rem_6, mul_6, add_6, sub_6);
1555 } else if a_u64s.len() == 8 {
1556 test!(mul_8, sqr_8, add_8, sub_8, div_rem_8, mod_inv_8);
1557 } else if input.len() == 512*2 + 4 {
1558 let mut e_bytes = [0; 4];
1559 e_bytes.copy_from_slice(&input[512 * 2..512 * 2 + 4]);
1560 let e = u32::from_le_bytes(e_bytes);
1561 let a = U4096::from_be_bytes(&a).unwrap();
1562 let b = U4096::from_be_bytes(&b).unwrap();
1564 let res = if let Ok(r) = a.expmod_odd_mod(e, &b) { r } else { return };
1565 let mut res_bytes = Vec::with_capacity(512);
1567 res_bytes.extend_from_slice(&i.to_be_bytes());
1570 let ring = ibig::modular::ModuloRing::new(&bi);
1571 let ar = ring.from(ai.clone());
1572 assert_eq!(ar.pow(&e.into()).residue(), ibig::UBig::from_be_bytes(&res_bytes));
1581 fn mul_min_simple_tests() {
1584 let res = mul_2(&a, &b);
1585 assert_eq!(res, [0, 3, 10, 8]);
1587 let a = [0x1bad_cafe_dead_beef, 2424];
1588 let b = [0x2bad_beef_dead_cafe, 4242];
1589 let res = mul_2(&a, &b);
1590 assert_eq!(res, [340296855556511776, 15015369169016130186, 4248480538569992542, 10282608]);
1592 let a = [0xf6d9_f8eb_8b60_7a6d, 0x4b93_833e_2194_fc2e];
1593 let b = [0xfdab_0000_6952_8ab4, 0xd302_0000_8282_0000];
1594 let res = mul_2(&a, &b);
1595 assert_eq!(res, [17625486516939878681, 18390748118453258282, 2695286104209847530, 1510594524414214144]);
1597 let a = [0x8b8b_8b8b_8b8b_8b8b, 0x8b8b_8b8b_8b8b_8b8b];
1598 let b = [0x8b8b_8b8b_8b8b_8b8b, 0x8b8b_8b8b_8b8b_8b8b];
1599 let res = mul_2(&a, &b);
1600 assert_eq!(res, [5481115605507762349, 8230042173354675923, 16737530186064798, 15714555036048702841]);
1602 let a = [0x0000_0000_0000_0020, 0x002d_362c_005b_7753];
1603 let b = [0x0900_0000_0030_0003, 0xb708_00fe_0000_00cd];
1604 let res = mul_2(&a, &b);
1605 assert_eq!(res, [1, 2306290405521702946, 17647397529888728169, 10271802099389861239]);
1607 let a = [0x0000_0000_7fff_ffff, 0xffff_ffff_0000_0000];
1608 let b = [0x0000_0800_0000_0000, 0x0000_1000_0000_00e1];
1609 let res = mul_2(&a, &b);
1610 assert_eq!(res, [1024, 0, 483183816703, 18446743107341910016]);
1612 let a = [0xf6d9_f8eb_ebeb_eb6d, 0x4b93_83a0_bb35_0680];
1613 let b = [0xfd02_b9b9_b9b9_b9b9, 0xb9b9_b9b9_b9b9_b9b9];
1614 let res = mul_2(&a, &b);
1615 assert_eq!(res, [17579814114991930107, 15033987447865175985, 488855932380801351, 5453318140933190272]);
1617 let a = [u64::MAX; 2];
1618 let b = [u64::MAX; 2];
1619 let res = mul_2(&a, &b);
1620 assert_eq!(res, [18446744073709551615, 18446744073709551614, 0, 1]);
1624 fn add_simple_tests() {
1625 let a = [u64::MAX; 2];
1626 let b = [u64::MAX; 2];
1627 assert_eq!(add_2(&a, &b), ([18446744073709551615, 18446744073709551614], true));
1629 let a = [0x1bad_cafe_dead_beef, 2424];
1630 let b = [0x2bad_beef_dead_cafe, 4242];
1631 assert_eq!(add_2(&a, &b), ([5141855058045667821, 6666], false));
1635 fn mul_4_simple_tests() {
1638 assert_eq!(mul_4(&a, &b),
1639 [0, 2, 4, 6, 8, 6, 4, 2]);
1641 let a = [0x1bad_cafe_dead_beef, 2424, 0x1bad_cafe_dead_beef, 2424];
1642 let b = [0x2bad_beef_dead_cafe, 4242, 0x2bad_beef_dead_cafe, 4242];
1643 assert_eq!(mul_4(&a, &b),
1644 [340296855556511776, 15015369169016130186, 4929074249683016095, 11583994264332991364,
1645 8837257932696496860, 15015369169036695402, 4248480538569992542, 10282608]);
1647 let a = [u64::MAX; 4];
1648 let b = [u64::MAX; 4];
1649 assert_eq!(mul_4(&a, &b),
1650 [18446744073709551615, 18446744073709551615, 18446744073709551615,
1651 18446744073709551614, 0, 0, 0, 1]);
1655 fn double_simple_tests() {
1656 let mut a = [0xfff5_b32d_01ff_0000, 0x00e7_e7e7_e7e7_e7e7];
1657 assert!(double!(a));
1658 assert_eq!(a, [18440945635998695424, 130551405668716494]);
1660 let mut a = [u64::MAX, u64::MAX];
1661 assert!(double!(a));
1662 assert_eq!(a, [18446744073709551615, 18446744073709551614]);