Missed optimization when using `split_first`

[xTachyon]

Impl 1:

pub fn f(input: &mut &[u64]) -> Option<u64> {
    match input {
        [] => None,
        [first, rest @ ..] => {
            *input = rest;
            Some(*first)
        }
    }
}

Impl 2:

pub fn f(input: &mut &[u64]) -> Option<u64> {
    let (first, rest) = input.split_first()?;
    *input = rest;
    Some(*first)
}

Godbolt: https://godbolt.org/z/aasq15z7a

These functions should be equivalent, but the second function generates one more if than the first one. This also happens when using .get(0) and then doing the slicing manually. I expected the impls to generate the same code.

[JakobDegen]

Quoting myself on Zulip:

I've seen this before too. The extra comparison is a null pointer check and is going to be the result of something creating a Option<&T>

[nikic]

This is a metadata preservation failure during SimplifyCFG speculation: https://llvm.godbolt.org/z/frs8o853f

Preserving !nonnull here used to be illegal, but thanks to https://reviews.llvm.org/D141386 we are allowed to preserve it now. Only need to drop !noundef.

[nikic]

Upstream patch: https://reviews.llvm.org/D146629

[nikic]

Fixed by #114048, needs codegen test.