Incrementing &mut Option<u32> generates verbose instructions 1.94.0 -> 1.95.0

Code

[Godbolt Link] This is a follow-up to #149970 which, at the time of this writing, has not been resolved yet.

Source:

pub fn f(x: &mut Option<u32>) {
    let mut y = *x;
    y = y.map(|a| a + 1);
    *x = y;
}
// OR
pub fn f(x: &mut Option<u32>) {
    *x = x.map(|a| a + 1);
}

1.94.0's more optimally generated assembly:

f:
        inc     dword ptr [rdi + 4]
        ret

1.95.0's less optimally generated assembly:

f:
        mov     eax, dword ptr [rdi + 4]
        inc     eax
        mov     dword ptr [rdi + 4], eax
        ret

Author's Insights

Within the optimization pipeline, I noticed that the LLVM IR was the same up until the X86 DAG->DAG Instruction Selection (x86-isel) step. The Compiler Explorer link contains a pane to see this.
Although not shown in Godbolt, the regression also exists for x86_32 (i686). See it with Overrides -> Target architecture -> i686-unknown-linux-gnu.

Compilation Info

Rust Edition: 2024
Flags: -C opt-level=3

@rustbot modify labels: +regression-from-stable-to-stable +T-compiler +C-optimization +A-codegen +A-LLVM +E-needs-bisection +S-has-mcve +O-x86_64 +O-x86_32 -C-bug -regression-untriaged

Start with the provided Godbolt link and compare the LLVM IR through the x86 DAG->DAG instruction-selection step, then reproduce the regression for both x86_64 and i686 with opt-level=3. Review follow-up issue #149970 and bisect the change between Rust 1.94.0 and 1.95.0. Done means the affected cases once again produce the single-instruction increment shown for 1.94.0.