Discriminant update fails to optimize

[joboet]

The following code should optimize to a single store of the new discriminant, but doesn't (Godbolt).

enum State {
    A([u8; 753]),
    B([u8; 753]),
    //C,
}

unsafe fn update(s: *mut State) {
    let S::A(v) = s.read() else { unreachable_unchecked() };
    s.write(S::B(v));
}

Instead, LLVM copies the value to the stack, copies it again and then writes it back. Interestingly, LLVM optimizes this correctly if a third state without value is added.

Meta

rustc --version --verbose:

rustc 1.78.0-nightly (c67326b06 2024-03-15)
binary: rustc
commit-hash: c67326b063bd27ed04f306ba2e372cd92e0a8751
commit-date: 2024-03-15
host: x86_64-unknown-linux-gnu
release: 1.78.0-nightly
LLVM version: 18.1.0

[matthiaskrgr]

lol, looks like this optimizes with -Zmir-opt=1/0 but fails to with anything equals or above 2, I think 2 or 3 is the default for release builds?

[clubby789]

Replacing unreachable_unchecked with return also optimizes better:

init:
        test    byte ptr [rdi], 1
        jne     .LBB0_2
        mov     byte ptr [rdi], 1
.LBB0_2:
        ret

Presumably some issue with MIR inlining;

#[inline(never)]
fn unreachable2() -> ! {
    unsafe { unreachable_unchecked() }
}

#[no_mangle]
unsafe fn init(s: *mut State) {
    let State::A(v) = read(s) else { unreachable2() };
    write(s, State::B(v));
}

produces the optimal code

[saethlin]

The problem is not MIR inlining per se. I think this is the problem @dianqk was trying to fix in #122282; if you write your own unreachable_unchecked function without a runtime check, it optimizes fine: https://godbolt.org/z/6oYqrhEs9

[dianqk]

Upstream issue: llvm/llvm-project#85560

@rustbot claim
:p

[rustbot]

Error: Parsing assign command in comment failed: ...' claim' | error: expected end of command at >| ' :p'...

Please file an issue on GitHub at triagebot if there's a problem with this bot, or reach out on #t-infra on Zulip.