Properly support LUB for ReErased

[compiler-errors]

This is needed to properly check obligations on opaque types, since instead of returning ReEmpty, we now return ReErased for unconstrained regions in MIR borrowck.

This shouldn't be unsafe for now, but it does kinda risk future bugs if we accidentally allow ReEmpty to leak into other inference contexts, since LUB would cause an ICE that is very obvious to see otherwise. I'm not really certain how to fix this other than reintroducing a ReRootEmpty region for opaque type inference that properly represents an unconstrained opaque lifetime..

cc @oli-obk and @jackh726
not sure which of y'all to request review from, so I flipped a coin and it landed on heads. therefore r? @oli-obk, feel free to reassign though.

Fixes #102649
Fixes #102510

Originally caused by #98559

[jackh726]

Yeah this is probably the easiest way out of this.

I don't think it's a "proper" fix though, but I'm not sure what that would look like. As you said, this almost merits a ReRootEmpty on its own. I'm guessing returning a ReVar here doesn't work, probably because we complain about unconstrained variables?

[compiler-errors]

I'm guessing returning a ReVar here doesn't work

Yeah, we can't do that, since this needs to be a lifetime that we can return in the type_of query, so it can't be associated to a specific inference context, for example.

Other choices like using ReStatic for these unconstrained regions causes outlives errors, and nothing else seems to be obviously correct. It would take a lot more modifications, but we could return a binder or something, but that would be a much larger scale change.

[jackh726]

r=me with a FIXME comment added referencing the issue. I think its worth merging this now, but it's not the right long-term fix. I can also see a case to actually escalate a bit more and open an issue to actually track a "proper" fix, but I'm not sure.

[aliemjay]

I'm not really comfortable with this. I believe a more acceptable short-term fix is to instantiate ReErased with fresh region variables. This way the hack would be more localized.

For the long-term fix I imagine wrapping the hidden type with an existential binder and adding sufficient assertions that the existential variables only appear in closure/generator parent substs.

[oli-obk]

Yea, I'd also prefer to limit this to just opaque types, so instantiating the ReErased to a new inference var after getting the hidden type in

rust/compiler/rustc_hir_analysis/src/check/check.rs

Line 727 in 3e01d07

let hidden_type = tcx.bound_type_of(def_id.to_def_id()).subst(tcx, substs);