libm: Improved integer utilities, implement shifts and bug fixes for i256 and u256

[quaternic]

i256 and u256

• operators now use the same overflow convention as primitives
• implement << and - (previously just >> and +)
• implement Ord correctly (the previous PartialOrd was broken)
• corrected i256::SIGNED to true

The above mentioned operators are now included in the MinInt-trait.
(Never mind, that would need additional changes in compiler-builtins, so I removed it for now.)

The Int-trait is extended with trailing_zeros, carrying_add, and borrowing_sub.

[tgross35]

The implementations all look pretty great to me. Would it be possible to extend the fuzz tests in libm-test/tests/u256.rs to cover these?

There are also some benchmarks, it would be good to basically copy the add and shr benchmarks for sub and shl.

compiler-builtins/libm-test/benches/icount.rs

Lines 57 to 127 in 0bbec72

	fn setup_u128_mul() -> Vec<(u128, u128)> {
	let step = u128::MAX / 300;
	let mut x = 0u128;
	let mut y = 0u128;
	let mut v = Vec::new();
	loop {
	'inner: loop {
	match y.checked_add(step) {
	Some(new) => y = new,
	None => break 'inner,
	}
	v.push((x, y))
	}
	match x.checked_add(step) {
	Some(new) => x = new,
	None => break,
	}
	}
	v
	}
	fn setup_u256_add() -> Vec<(u256, u256)> {
	let mut v = Vec::new();
	for (x, y) in setup_u128_mul() {
	// square the u128 inputs to cover most of the u256 range
	v.push((x.widen_mul(x), y.widen_mul(y)));
	}
	// Doesn't get covered by `u128:MAX^2`
	v.push((u256::MAX, u256::MAX));
	v
	}
	fn setup_u256_shift() -> Vec<(u256, u32)> {
	let mut v = Vec::new();
	for (x, _) in setup_u128_mul() {
	let x2 = x.widen_mul(x);
	for y in 0u32..256 {
	v.push((x2, y));
	}
	}
	v
	}
	#[library_benchmark]
	#[bench::linspace(setup_u128_mul())]
	fn icount_bench_u128_widen_mul(cases: Vec<(u128, u128)>) {
	for (x, y) in cases.iter().copied() {
	black_box(black_box(x).zero_widen_mul(black_box(y)));
	}
	}
	#[library_benchmark]
	#[bench::linspace(setup_u256_add())]
	fn icount_bench_u256_add(cases: Vec<(u256, u256)>) {
	for (x, y) in cases.iter().copied() {
	black_box(black_box(x) + black_box(y));
	}
	}
	#[library_benchmark]
	#[bench::linspace(setup_u256_shift())]
	fn icount_bench_u256_shr(cases: Vec<(u256, u32)>) {
	for (x, y) in cases.iter().copied() {
	black_box(black_box(x) >> black_box(y));
	}

[quaternic]

I've now added both of those.

[tgross35]

Awesome, thanks for the improvements!