Useless copies of alignment padding

[glandium]

See the following extreme test-case:

#[repr(align(1024))]
pub struct Foo(usize);

pub fn foo() -> Box<Foo> {
    Box::new(Foo(42))
}

This is what godbolt says rust stable does with it with -O:

        push    rbx
        sub     rsp, 1024
        mov     edi, 1024
        mov     esi, 1024
        call    __rust_alloc@PLT
        test    rax, rax
        je      .LBB0_1
        mov     rbx, rax
        mov     qword ptr [rax], 42
        mov     rdi, rax
        add     rdi, 8
        lea     rsi, [rsp + 8]
        mov     edx, 1016
        call    memcpy@PLT
        mov     rax, rbx
        add     rsp, 1024
        pop     rbx
        ret
.LBB0_1:
        mov     edi, 1024
        mov     esi, 1024
        call    alloc::alloc::handle_alloc_error@PLT

That is, it allocates 1024 bytes (expected), writes 42 directly at the allocated location, and then proceeds to copy 1016 (uninitialized) bytes from the stack.

[nikic]

Similar to #56191, but not quite the same issue.

[klensy]

Looks resolved around 1.52 https://rust.godbolt.org/z/n14acj8sG

[kadiwa4]

@rustbot label: +E-needs-test -A-LLVM
(--emit=llvm-ir already has the optimization and as I understand that hasn't been touched by LLVM's optimizer yet. So it isn't LLVM-specific, right?)