`repr(C)` enums with a variant that overflows the native enum type should warn

[ChrisDenton]

Consider this:

#[repr(C)]
enum Bar {
    A = 0x1_FFFF_FFFF,
}

Assume that the targets native enum size is 4 bytes. Obviously 0x1_FFFF_FFFF will overflow. This is accepted by the compiler, without warning. Instead it will silently use a larger underlying type.

I think this should be a future incompatibility warning with an eye to making it a hard error in a future edition. Being explicit with, say, repr(u64) is better than doing so silently (especially with it being incompatible with MSVC). However, any warning would be good if others disagree.

[workingjubilee]

I believe the result of this declaration in C would be, as they say, "implementation-defined", so yes we should warn in this case and possibly error in the future. However we must be careful because we need to actually only get shouty if this is bigger than the C int, not the actual default #[repr(C)] enum size, since some platforms define their underlying types for enums as scaling down to 8 bits by default and thus using the minimal size, a la Rust.

[ChrisDenton]

That sounds fair. It's times like these that I think the duel purpose of repr(C) as both "whatever C does on this platform" and "guaranteed layout for unsafe shenanigans" was a mistake.

Btw, I'll remove the O-windows tag because this is not really specific to any platform.

[LunarLambda]

In C, prior to C23, all enumerators must fit in int, so repr(C) and overflow should warn/error (Godbolt)

In C++, the underlying type is again implementation-defined, but must not be larger than int/unsigned int unless an enumerator demands it (Godbolt). This matches what Rust seems to do currently.

C23 adapts the C++ rules. The language used there is a lot more complicated, seeming to imply that individual enumerators may have different types since they become globally scoped "integer constants" (cppreference), and rather than having an underlying type that all enumerators share, the enum as a whole has an implementation-defined "compatible" type that is capable of storing all enumerators. (unsure how to experimentally verify this) (or this could just be an incredibly roundabout way of saying it's exactly like the C++ rules)