suboptimal codegen when initializing an array of padded structs

[iximeow]

in the following code, reset_implicit_padding generates somewhat worse assembly under -C opt-level=3 than reset_explicit_padding. with explicit padding generates to a memset - i'd hoped both would memset:

#[repr(Rust, align(2))]
#[derive(Default, Copy, Clone)]
pub struct ImplicitPadding {
    v: u8,
}

#[repr(Rust, align(2))]
#[derive(Default, Copy, Clone)]
pub struct ExplicitPadding {
    v: u8,
    _pad: [u8; 1],
}

#[no_mangle]
pub fn reset_implicit_padding(array: &mut [ImplicitPadding; 160]) {
    *array = [ImplicitPadding::default(); 160];
}

#[no_mangle]
pub fn reset_explicit_padding(array: &mut [ExplicitPadding; 160]) {
    *array = [ExplicitPadding::default(); 160];
}

which compiles to these two implementations:

reset_implicit_padding:
        mov     eax, 7
.LBB0_1:
        mov     byte ptr [rdi + 2*rax - 14], 0
        mov     byte ptr [rdi + 2*rax - 12], 0
        mov     byte ptr [rdi + 2*rax - 10], 0
        mov     byte ptr [rdi + 2*rax - 8], 0
        mov     byte ptr [rdi + 2*rax - 6], 0
        mov     byte ptr [rdi + 2*rax - 4], 0
        mov     byte ptr [rdi + 2*rax - 2], 0
        mov     byte ptr [rdi + 2*rax], 0
        add     rax, 8
        cmp     rax, 167
        jne     .LBB0_1
        ret

reset_explicit_padding:
        mov     edx, 320
        xor     esi, esi
        jmp     qword ptr [rip + memset@GOTPCREL]

godbolt for reference.

it turns out that when an array like this does become a memset it's seemingly often because llvm realized in loop-idiom that the stores can be turned into a memset, rather than ever being emitted as a memset from rustc?

i don't know enough to know if this would need rustc to initialize padding bytes as undef (?), or if they're already undef and there's some other reason llvm doesn't figure out this could be a memset. i also can't tell if this would be in conflict with this issue about padding copies or the regression test for it. there are certainly cases where copying large amounts of padding would be counterproductive, though if writing a padding byte could turn 4+2+1 bytes of stores into one 8-byte store, that's probably always a net win.

(the original code i've minimized had 6-byte or 14-byte structs aligned to 8 and 16 respectively, and for entirely indecipherable reasons also got the assignment fully unrolled, yielding ~600 mov rather than a memset. i would much rather that be a call to memset, and have added a _pad field for the time being.)

[clubby789]

Looks somewhat similar to https://github.com/rust-lang/rust/issues/101685. iirc, LLVM eliminates store undef (which is what ends up being emitted by this) before running the opt that transforms them to memset/cpy.

store i8 val, ptr
store i8 val, ptr + 2
store i8 val, ptr + 4

etc. can't be turned into a memset at that point as LLVM only sees the holey stores