Miri: avoid comparing layouts

[RalfJung]

This is a follow-up to #70532.

r? @eddyb so I tried to find a case where the "subtyping on assignment" happens with a larger type, not just immediately with a function pointer. So far, I failed.

Here's my last attempt:

use std::cell::RefCell;

type Log<'a> = &'a RefCell<Vec<u32>>;

struct Wrap<T>(T, u32, u32);

struct GaspB<'a>(Wrap<for <'b> fn(u32, &'b str, Log<'a>)>);

enum E<'a> {
    B(GaspB<'a>, bool),
}

fn g_b(_y: u32, _ctxt: &str, _log: Log) {}

pub fn main() {
    // This has a type-changing assignment.
    E::B(GaspB(Wrap(g_b, 0, 1)), true);
    // This does not.
    let w: Wrap<for <'b, 'a> fn(u32, &'b str, Log<'a>)> = Wrap(g_b, 0, 1);
    E::B(GaspB(w), true);
}

If such a case exists, the assertion here is still wrong (as the larger type might or might not be a Scalar/ScalarPair). If such a case does not exist, the fully type-based check should actually work. Thus, this PR cannot actually make things any less correct. Am I missing something?

[RalfJung]

I am slightly paranoid here about comparing the wrong thing (in an earlier version I accidentally compared mutabilities here instead of types, by using the wrong pattern). Is there any good way to assert that these things are types?

[eddyb]

You can probably use Ty::eq to call the implementation of ==, but idk if it looks pretty.
The other option is I guess ascription, i.e. (src_pointeee: Ty) == (dest_pointee: Ty).

[eddyb]

That example is extremely complex, I can't follow it. It should be trivial to use subtyping:

fn simple(x: for<'a> fn(&'a ())) -> fn(&'static ()) {
    x
}

fn nested(x: (for<'a> fn(&'a ()), String)) -> (fn(&'static ()), String) {
    x
}

[eddyb]

There's a trap you may have fallen into, looking again at your example:
You're constructing structs by passing g_b to them, without enforcing the type of the struct.
That means g_b will get an expected type passed down and subtype early, instead of the whole struct that contains g_b subtyping later. I get around that by not constructing anything myself.

[RalfJung]

I tried to get around exactly that by adding the intermediate w variable.

[RalfJung]

@eddyb ah, your nested indeed ICEs on Miri.

[RalfJung]

This PR is definitely wrong then, closing. See #70804 for the issue.