Unexpected alignment of `repr(packed)` struct containing a DST

[nicholasbishop]

Example code (playground link)

#![feature(layout_for_ptr, ptr_metadata)]
use std::{mem, ptr};

#[repr(C, packed)]
struct A {
    f: u32,
}

#[repr(C, packed)]
struct B {
    f: [u32],
}

fn main() {
    let a = A { f: 0 };
    dbg!(mem::align_of_val(&a));

    let storage = [0u8; 4];
    let b: *const B = ptr::from_raw_parts(storage.as_ptr().cast(), 1);

    dbg!(unsafe { mem::align_of_val_raw(b) });
}

Running this outside of Miri gives the output I would expect; values of A and B both have an alignment of 1:

[src/main.rs:16] mem::align_of_val(&a) = 1
[src/main.rs:21] unsafe { mem::align_of_val_raw(b) } = 1

But when running under Miri, I get this instead:

[src/main.rs:16] mem::align_of_val(&a) = 1
[src/main.rs:21] unsafe { mem::align_of_val_raw(b) } = 4

I would have expected the alignment of B to be 1 since it is repr(packed). https://doc.rust-lang.org/nomicon/other-reprs.html#reprpacked says "repr(packed) forces Rust to strip any padding, and only align the type to a byte", and doesn't mention an exception for DSTs.

[RalfJung]

Good catch! This indeed seems wrong.

Miri's logic for this is here. And it seems like that subtly differs from the codegen logic. The cranelift one is here and it does contain a special case for packed types. Cc @bjorn3

Looks like we really would want some way to share this code between all the backends and Miri...