suboptimal llvm-ir due to missing noalias annotations after mir inlining

suboptimal llvm-ir due to missing noalias annotations after mir inlining

Code

I tried this code:

#[inline(never)]
#[no_mangle]
pub unsafe fn noalias_ptr(x: *mut i32, y: *mut i32) -> i32 {
    #[inline]
    fn noalias_ref(x: &mut i32, y: &mut i32) -> i32 {
        *x = 16;
        *y = 12;
        *x
    }

    let x = &mut *x;
    let y = &mut *y;
    noalias_ref(x, y)
}

I expected to see this happen: generated asm on x86-64

noalias_ptr:
  mov dword ptr [rdi], 16
  mov dword ptr [rsi], 12
  mov eax, 16
  ret

Instead, this happened: generated asm on x86-64

noalias_ptr:
  mov dword ptr [rdi], 16
  mov dword ptr [rsi], 12
  mov eax, dword ptr [rdi]
  ret

Version it worked on

It most recently worked on: rust 1.64

Version with regression

rustc --version --verbose:

rustc 1.80.0 (051478957 2024-07-21)
binary: rustc
commit-hash: 051478957371ee0084a7c0913941d2a8c4757bb9
commit-date: 2024-07-21
host: x86_64-unknown-linux-gnu
release: 1.80.0
LLVM version: 18.1.7
Compiler returned: 0

Fun fact

changing noalias_ref(x, y) to (noalias_ref as fn(_, _) -> _)(x, y) "fixes" the issue in this case, since it seems to block the mir inliner

@rustbot modify labels: +regression-from-stable-to-stable-regression-untriaged

The report names no source files or tests. Start by reproducing the Rust 1.80 behavior with the provided code and comparing MIR, LLVM IR, and x86-64 output with Rust 1.64; trace how MIR inlining affects noalias information. Done means the optimized output avoids the unnecessary reload and matches the expected result.