Forbid object types which are not object safe

[nrc]

We currently only check for object safety when coercing a concrete object to a trait object. We should prevent the programmer writing object types which are not object safe.

[nrc]

backwards incompatible

[nikomatsakis]

True but I think it can only break dead code -- you can't produce instances of those types. Seems like this doesn't have to be fixed for 1.0.

[brson]

P-low. Seems like it won't break any useful code. Justabug?

[nikomatsakis]

(That is, it's probably the kind of bug we can fix at some point. Not that I don't want to fix it.)

[ftxqxd]

I think this would also break public functions that take non-object-safe trait objects as parameters (e.g., pub fn foo(x: Box<ObjectUnsafe>)), so long as they aren’t called within the crate.

[steveklabnik]

Traige: we did this, didn't we?

[jonas-schievink]

Looks like #21953 is a duplicate which was closed

[bluss]

Is FnOnce object safe? Its methods can't be called through an object. Add is the same way (currently object safe but not callable).

[steveklabnik]

@jonas-schievink yup!

[nikomatsakis]

@bluss FnOnce is considered object safe. Basically, the restriction against calling call_once(self) is enforced on the caller side by requiring that callee be Sized (and hopefully we will lift this restruction eventually). Put another way, if I write fn foo<T:FnOnce()>(t: T), I can't have T=FnOnce() anyway because T: Sized would not hold. If I write fn foo<T:FnOnce()+?Sized>(t: T), then T=FnOnce() is possible, but t() will fail to compile because T: Sized does not hold. So it works out either way.