Dead variant removal in #[repr(C)] enums

[GoldsteinE]

Consider this enum:

#[repr(C)]
enum Fun<T> {
    Fun(T),
    // maybe some other variants here
}

For unsafe code it could be useful to assume that Fun<T> is layout-compatible with Fun<MaybeUninit<T>> for any choice of T. That’s not currently the case: if T is uninhabited, Fun<T> would be zero-sized (or, more generally, the ::Fun variant will not affect enum size), but Fun<MaybeUninit<T>> wouldn’t.

Reference currently documents the above enum as being rough equivalent to something like

#[repr(C)]
enum FunDiscriminant {
    Fun,
}

#[repr(C)]
union FunData<T> {
    fun: T,
}

#[repr(C)]
struct Fun<T> {
    discriminant: FunDiscriminant,
    data: FunData<T>,
}

That doesn’t hold for uninhabited types: this “desugaring” preserves the invariant above.

I think we should do one of the following:

1. Either dead variant removal should be disabled for #[repr(C)] enums;
2. or it should be documented in the reference (I can write the patch if this option is selected).

Related: #443
Playground demonstrating the issue: https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=d2995ff99cea9c2e60d28cbf0f970176

[RalfJung]

It's interesting what happens when you remove the repr(C) from FunDiscriminant:

[src/main.rs:29:5] std::mem::size_of::<Fun<Void>>() = 0
[src/main.rs:30:5] std::mem::size_of::<Fun<std::mem::MaybeUninit<Void>>>() = 4
[src/main.rs:31:5] std::mem::size_of::<desugared::Fun<Void>>() = 0
[src/main.rs:32:5] std::mem::size_of::<desugared::Fun<std::mem::MaybeUninit<Void>>>() = 0

That's more what I expected: single-variant enums can have a zero-sized discriminant.

Also interestingly, zero-variant enums with repr(C) are rejected.

---

My first reaction is that this is a bug in rustc, we shouldn't do dead variant removal on repr(C) enums.

[chorman0773]

I agree with that. Even if the variant is uninhabited, we guarantee that there's a target-dependant discriminant field in the layout of a repr(C) enum. The same will be true of repr(Int), and repr(C,Int) (where the discriminant field is fixed to Int).

[GoldsteinE]

Assuming that’s a rustc bug, I sent a patch that fixes it.

[GoldsteinE]

The fix is stable now, so I think this could be closed.