E0277 for unions only mentions structs and enum variants

[apasel422]

The note issued for the following is "only the last field of a struct or enum variant may have a dynamically sized type," which is confusing, since no struct or enum variant is used:

#![feature(untagged_unions)]

union Foo<T: ?Sized> {
    value: T,
}

[durka]

Similar to #36252.

[nagisa]

Is there any reason to forbid any (and arbitrary number of) union variant from being dynamically sized? As far as layout is concerned every single field/variant of union is the last.

[apasel422]

@nagisa I'd like to see dynamically sized union fields supported as well.

[petrochenkov]

@apasel422

I'd like to see dynamically sized union fields supported as well.

What is the motivation? (Besides "Why not?")
I didn't implement them because enums don't support DST's either, and the union RFC doesn't talk about DSTs and design work (read, one more RFC) is required.

[apasel422]

@petrochenkov My original use case was something like:

union Inner<T: ?Sized> {
    value: T,
}

pub struct NoDrop<T: ?Sized>(Inner<T>);

where you want something like Box<NoDrop<Trait>>, which will let the Box's allocation be freed but prevent Trait from being dropped. But now I'm struggling to come up with a reason to do that.

[joshtriplett]

@petrochenkov Is there any non-trivial design question that arises from putting DSTs in a union, or would they just work in the obvious way? Not arguing against an additional RFC here, just asking if I've missed some unexpected complexity arising from that combination.

[petrochenkov]

@joshtriplett
I don't remember exactly, something about unsizing coercions and how to construct unsized unions (with examples). This coercion business also represented the largest part of implementation work.
+ Something about last/non-last/>1 unsized fields in unions, they are different from structs in this regard.
+ The motivational part, of course.

[Mark-Simulacrum]

Aside from the overall question of DSTs in unions, for the time being we should fix the error message here in some way, but I don't know if we have access to whether it's a union or struct at that point -- "last field" for a union doesn't really make sense.