Suboptimal code generation for array Element by Element replace/overwrite (Uneeded stores)

[Keith-Cancel]

I noticed that the code generation on something I was working on was less than ideal. I was able to distill it down to some very simple examples here on compiler explorer: Compiler Explorer

I would expect the output to be similar to either test4() or test5() for the functions test1(), test2(), and test3().

For example here is test3():

fn test3(mut a: [u32; 3], b: &[u32; 3]) -> [u32;3] {
    *a.get_mut(0).unwrap() = b[0];
    *a.get_mut(1).unwrap() = b[1];
    *a.get_mut(2).unwrap() = b[2];
    return a;
}

The generated assembly is:

test3:
    mov rax, rdi
    mov rcx, qword ptr [rdx]
    mov qword ptr [rsi], rcx
    mov edx, dword ptr [rdx + 8]
    mov dword ptr [rsi + 8], edx
    mov dword ptr [rdi + 8], edx
    mov qword ptr [rdi], rcx
    ret

Where as if we look at test4() we can see that we can get rid of the spurious stores, by adding let mut a = a; What is interesting is this does nothing when using the index operator [], if you look at what I have on compiler explorer test2().

fn test4(mut a: [u32; 3], b: &[u32; 3]) -> [u32;3] {
    let mut a = a;
    *a.get_mut(0).unwrap() = b[0];
    *a.get_mut(1).unwrap() = b[1];
    *a.get_mut(2).unwrap() = b[2];
    return a;
}

The generated assembly for test4()

test4:
    mov rax, rdi
    mov ecx, dword ptr [rdx + 8]
    mov rdx, qword ptr [rdx]
    mov qword ptr [rdi], rdx
    mov dword ptr [rdi + 8], ecx
    ret

@rustbot label A-LLVM
@rustbot label C-optimization
@rustbot label A-codegen