Suboptimal Assembly Output from a Modulo on a Power of Two

[Altanis]

I tried this code:

#[no_mangle]
pub fn index(key: usize, buckets: usize) -> usize {
    assert!(buckets.is_power_of_two());
    key % buckets
}

Since assert!(buckets.is_power_of_two()) is run (and .is_power_of_two() is an inbuilt function), the latter modulus operation should reduce to key & (buckets - 1).

Instead, the outputted assembly here shows that this reduction does not take place.

Meta

The compiler in the GodBolt link is version 1.80.0.

[scottmcm]

This is llvm/llvm-project#58996, and needs to be fixed in LLVM. Rust itself won't do this optimization.

[dtcxzyw]

@scottmcm @dianqk @nikic This issue has been fixed by llvm/llvm-project#107745. Please add llvm-fixed-upstream.

[dianqk]

You can also add it. :)
@rustbot label +llvm-fixed-upstream

[nikic]

@dtcxzyw Doesn't this issue need support for dominating condition rather than assumption?

[dtcxzyw]

@dtcxzyw Doesn't this issue need support for dominating condition rather than assumption?

I am preparing a follow-up patch for this :)

[dianqk]

@rustbot label -llvm-fixed-upstream

[veera-sivarajan]

@rustbot label +llvm-fixed-upstream

[nikic]

Fixed by LLVM 20 upgrade.

[scottmcm]

With the test added in #138472 I think we're good here 🎉