Missed optimization: re-checking enum variants after `std::mem::replace` is not eliminated (somtimes)

[oxalica]

Minified example: (godbolt)

use std::task::{Poll, Waker};

pub enum State<T> {
    Inactive,
    Active(Waker),
    Signalled(T),
}

#[unsafe(no_mangle)]
pub fn poll_state(st: &mut State<String>, w: &Waker) -> Poll<String> { // a
    match st {
        State::Signalled(_) => {
            // Just checked the variant, take the value out.
            let State::Signalled(v) = std::mem::replace(st, State::Inactive) else {
                unreachable!() // This panic should be eliminated.
            };
            Poll::Ready(v)
        }
        _ => {
            *st = State::Active(w.clone()); // b
            Poll::Pending
        }
    }
}

The optimization is fragile. If we remove the state assignment on the second branch (b), or change function signature to operate on State<u8>, then the unreachable panic in the first branch will be correctly eliminated.

[Berrysoft]

Seems that the panic won't be eliminated if assignment (b) is removed.

[ashivaram23]

I've come across something similar. I think MIR optimizations don't propagate known enum discriminants across blocks so it's left to LLVM, which might be unable to do more with it. In the case I've seen, LLVM's inability to optimize further could be traced down to how it handles the enum's niche optimization, but it could be something totally different here.

A MIR pass that propagates enum discriminants (through dataflow analysis or just by checking the direct predecessors of each block) would probably fix all these cases.

[oxalica]

known enum discriminants

So is it related to that known value ranges are not propagated through some boundaries (functions, for example)? I also encountered an issue about the discriminant of Option<NonZero<usize>> in #49572 (comment)? Not sure if they are related.

[ashivaram23]

I'm mainly thinking of propagating across blocks within a single function. Some kind of logic that would allow the first match arm block to consider the panic block unreachable since it can only be reached if (a copy of) st is not Signalled, which can't possibly be true at that point.

The example in the comment you linked should also be fixed by that logic since you set to Some and unwrap within a single block. If the function pointer call had range metadata saying it's not zero, or if it was a separate function with a return value range attribute, then LLVM would probably eliminate the unwrap failure call on its own. But I think the benefit of doing this optimization in MIR is that it wouldn't depend on the frontend giving LLVM the right metadata for all these cases and LLVM managing to understand all the different kinds of niches.

[nikic]

This fails to optimize because LLVM currently fails to track loads through memcpys.