Enforce predicates which link bivariant opaque lifetimes

[compiler-errors]

During predicates_of, we install outlives obligations which link the bivariant lifetimes of an opaque to their invariant (captured) copies. During NLL, we don't enforce well-formedness of an opaque that we are trying to define, and therefore we never actually enforce that these lifetimes are equal. This means that we have strange errors (#122307), or ICEs (#121512) from unconstrained lifetimes that should be constrained.

Fix this by requiring the opaque in add_item_bounds_for_hidden_type to be well-formed, which will enforce its own predicates.

r? @aliemjay cc @lcnr @oli-obk

You have any thoughts about this approach? I guess f72f5d8 is unnecessary, because it's still possible to ICE with RPITs and uncaptured lifetimes, e.g.

fn test<'a: 'a>() -> impl Sized {
    let _: () = test::<'a>();
}

---

side-note: I'm feeling a déjà-vu writing this PR, like this has been attempted before... if it has, could you remind me where this was, and why this approach ultimately wasn't taken?

[aliemjay]

If the parent lifetimes are relevant in any way then why they are bivariant in the first place? I do not expect the WF of an opaque type to constrain its bivariant lifetimes. A bit off-topic but the current state does not look right to me.

[aliemjay]

I guess f72f5d8 is unnecessary, because it's still possible to ICE with RPITs and uncaptured lifetimes, e.g.

This ICE should be fixed by #116891.

[lcnr]

why is this necessary? I'd expect that we require the opaque to be well-formed in the location which introduces it. We have to check the hidden type, as the structural recursion in wf.rs does not check it, so there are types in "relevant positions" which are not wf.

[lcnr]

This is confusing for me 🤔 I think I now get what's going on:

• calling test::<'a>() returns Opaque<'a, 'a>, we then assign this to a local of type () which uses subtyping.
• generalizing Opaque<'a, 'a> results in Opaque<'whatever, 'a> because the first lifetime param is bivariant (would be the same if it were co/contravariant)
• we then never check that Opaque<'whatever, 'a> is wf, leaving 'whatever unconstrained

That's also not quite right, as even changing Generalizer::regions to always set self.has_unconstrained_ty_var when creating a new region var does not fix this 🤔

[lcnr]

very much in favor of this 😁 wanted to actually open a PR for that myself soon.

Can you merge the behavior of all ty::Alias here?

[compiler-errors]

This PR is unnecessary since (afaict) this bug is fixed by #116891, or at least it was fixed when I rebased and tested that one. Another fix (one with less fallout) would also be to just mark all TAIT args as invariant, since they're always captured anyways.