Missed optimization with `*x += 0;`

[theemathas]

This issue was discovered by @ruriww

I tried this code:

#[unsafe(no_mangle)]
pub fn f(x: &mut u32) {
    *x = *x;
    if *x != 0 {
        *x = 0;
    }
}

#[unsafe(no_mangle)]
pub fn g(x: &mut u32) {
    *x = *x;
    if *x != 0 {
        *x = 0;
    } else {
        *x += 0;
    }
}

I expected both functions to compile to the same assembly. Note that the *x = *x; gives the compiler license to write to *x (without running into opsem questions about whether spurious writes are allowed).

Instead I got the following with -Copt-level=3 (Godbolt):

f:
        mov     dword ptr [rdi], 0
        ret

g:
        cmp     dword ptr [rdi], 0
        je      .LBB1_2
        mov     dword ptr [rdi], 0
.LBB1_2:
        ret

Meta

Godbolt compiler version:

rustc 1.94.0-nightly (fa5eda19b 2025-12-12)
binary: rustc
commit-hash: fa5eda19b95201468f5b1c5c035ec2fc06fccd66
commit-date: 2025-12-12
host: x86_64-unknown-linux-gnu
release: 1.94.0-nightly
LLVM version: 21.1.5
Internal compiler ID: nightly

[saethlin]

This is well-known as "noalias is not enough" #53105

[theemathas]

@saethlin How do you explain the fact that f gets optimized? I would have thought that *x = *x; would have been sufficient to get around that issue you linked.

[nikic]

This is a phase ordering problem. In the second example, the store that would allow the optimization is optimized out too early.

Unless you have examples of this highly specific code structure occurring in the wild, I think this is won't fix.