slice::get_mut() followed by slice::copy_from_slice generates unreachable panic branch

[bugadani]

In code like the following, the compiler misses a possible optimization: in f1, the length of dst is equal to the length of bytes, yet the compiler generates a call to len_mismatch_fail.

pub fn f1(a: &mut [u8], offset: usize, bytes: &[u8]) {
    if let Some(dst) = a.get_mut(offset..offset + bytes.len()) {
        dst.copy_from_slice(bytes);
    }
}

The compiler knows the lengths are equal (and thus, the panic is unreachable), because the following snippet optimizes the panic away:

pub fn f2(a: &mut [u8], offset: usize, bytes: &[u8]) {
    if let Some(dst) = a.get_mut(offset..offset + bytes.len()) {
        assert!(dst.len() == bytes.len());
        dst.copy_from_slice(bytes);
    }
}

This is a regression between rustc versions 1.51 and 1.52.

https://rust.godbolt.org/z/YhEY78E9P

[steffahn]

@rustbot label I-slow, E-needs-bisection

[steffahn]

This seems to be fixed on nightly. @rustbot label E-needs-test

Bisection would still be interesting IMO, both for the regression and the fix.

[steffahn]

Alright, that’s pretty boring in terms of rustc-bisection: it was fixed by #99464 (Update to LLVM 15), and it regressed in nightly-2021-03-05 which includes #81451 (Upgrade¹ to LLVM 12).

@rustbot label -E-needs-bisection, +A-llvm.

Footnotes

1. what’s up with the inconsistent naming, upgrade vs. update? ↩