The RBox SharedStableAbi implementation should have IsNonZeroType == True

[afranchuk]

Unless I'm missing something, isn't it the case that RBox should have IsNonZeroType as True rather than False?

[rodrimati1992]

The reason I haven't done it yet is because I haven't seen documentation guaranteeing that it has a stable representation across compiler versions.

#[repr(C)]
pub struct RBox<T>{
    ptr: std::ptr::NonNull<T>,
    vtable: *const [T; 0],
}

extern {
    pub fn foo(_: RBox<usize>);
    pub fn bar(_: Option<RBox<usize>>);
}

And the compiler warns with this message:

warning: `extern` block uses type `std::option::Option<RBox<usize>>`, which is not FFI-safe
 --> src/lib.rs:9:19
  |
9 |     pub fn bar(_: Option<RBox<usize>>);
  |                   ^^^^^^^^^^^^^^^^^^^ not FFI-safe
  |
  = note: `#[warn(improper_ctypes)]` on by default
  = help: consider adding a `#[repr(C)]`, `#[repr(transparent)]`, or integer `#[repr(...)]` attribute to this enum
  = note: enum has no representation hint

warning: 1 warning emitted

But this lint is known to have some false positives, so it might actually be ffi-safe.

[afranchuk]

Ah, understood. Yeah I think that they are being overly cautious there. After reading the documentation of IsNonZeroType, I just expected RBox to be able to satisfy that (it has a unique representation for zero). Otherwise the documentation for NonNull states that Option<NonNull<T>> is the same as *mut T, so that seems like it declares a stable representation across compiler versions (though it'd be best to verify against the standard library specification).

[rodrimati1992]

I decided to ask the question in the zulip unsafe-code-guidelines chat ,it looks like it's not guaranteed.

[afranchuk]

Ah, thanks for looking further into it, that's a revealing discussion.