Missed optimization: regression 1.29 -> 1.30, makes unnecessary load from memory

[safinaskar]

Consider this code: https://godbolt.org/z/QpiBK3 .
Code:

pub fn f(a: &mut i32, b: &mut i32) -> (i32, i32, i32) {
    let c = *a;
    let d = *b;
    *a = 0;
    (c, d, *b)
}

Asm (1.30):

example::f:
        movl    (%rsi), %eax
        movl    (%rdx), %ecx
        movl    $0, (%rsi)
        movl    (%rdx), %edx
        movl    %eax, (%rdi)
        movl    %ecx, 4(%rdi)
        movl    %edx, 8(%rdi)
        movq    %rdi, %rax
        retq

We read from memory at line let d = *b (from %rdx) and then we read from memory again at line (c, d, *b) (again from %rdx).

But in Rust (unlike C and C++) we know that two mutable references cannot point to same location. So, we missed optimization. This is regression

[jonas-schievink]

This is likely due to the known issues with LLVM's noalias annotation. See #54878.

[Mark-Simulacrum]

This has since been fixed. Current nightly output doesn't load from memory twice for rdx.

f:
        mov     rax, rdi
        mov     ecx, dword ptr [rsi]
        mov     edx, dword ptr [rdx]
        mov     dword ptr [rsi], 0
        mov     dword ptr [rdi], ecx
        mov     dword ptr [rdi + 4], edx
        mov     dword ptr [rdi + 8], edx
        ret