pub fn copy_u32(input: &[u8; 4], output: &mut [u8; 4])  {
    let a = &input[..4];
    let b = &mut output[..4];
    for b in b.chunks_exact_mut(4) {
        b.copy_from_slice(a);
    }
}

Complied with -C opt-level=s -C inline-threshold=300 produces (same as just -C opt-level=s)

example::copy_u32:
        mov     rax, rsi
        mov     rdx, rdi
        mov     esi, 4
        mov     ecx, 4
        mov     rdi, rax
        jmp     qword ptr [rip + core::slice::<impl [T]>::copy_from_slice@GOTPCREL]

Complied with -C opt-level=s -C llvm-args=-inline-threshold=300 produces

example::copy_u32:
        mov     eax, dword ptr [rdi]
        mov     dword ptr [rsi], eax
        ret

(Godbolt link)

The cause appears to be that LLVM keeps using a separate inlining threshold from the "primary" one given by rustc on functions with optsize attribute, but only when it didn't receive -inline-threshold itself.

Fixing this seems to either require converting the rustc-received argument to LLVM commandline one, or constructing a custom llvm::InlineParams structure instead of using one of LLVM-provided helper functions?