Restore Opaque behavior to coherence check

[compiler-errors]

Fixes #99663.

This broke in 84c3fcd. I'm not exactly certain that adding this behavior back is necessarily correct, but at least the UI test I provided may stimulate some thoughts.

I think delaying a bug here is certainly not correct in the case of opaques -- if we want to change coherence behavior for opaques, then we should at least be emitting a new error.

r? @lcnr

[compiler-errors]

#99552 (comment)

The algorithm replicates the existing behaviour, so...

Hm, guess not?

[lcnr]

Hm, guess not?

@compiler-errors is the Opaque difference the only reason why you wrote this?

[compiler-errors]

@lcnr that's the only difference I could find, but I didn't look further than the opaque change that I touched in this PR.

Just pointed out that comment because it seems #99552 did change behavior, not just replicated the old behavior like the comment suggested, unless I'm misunderstanding the point of that comment.

[lcnr]

please add the following as an additional test:

#![feature(type_alias_impl_trait)]

struct Outer<T: ?Sized> {
    i: InnerSend<T>,
}

type InnerSend<T: ?Sized> = impl Send;

fn constrain<T: ?Sized>() -> InnerSend<T> {
    ()
}

trait SendMustNotImplDrop {}
impl<T: ?Sized + Send + Drop> SendMustNotImplDrop for T {}
impl<T: ?Sized> SendMustNotImplDrop for Outer<T> {}

fn main() {}

this triggers the ice using auto traits.

It was surprising to me that we never look into the opaque types during coherence, causing an ICE for closures etc as well, (by changing constrain to || ()). The reason for that is that is_knowable fails for auto trait/sized impls on Opaque types as they always result in Conflict::Upstream.

r=me after adding this test and fixing my nit

[lcnr]

Just pointed out that comment because it seems #99552 did change behavior, not just replicated the old behavior like the comment suggested, unless I'm misunderstanding the point of that comment.

It did change the behavior for Opaque types because we thought that case would be unreachable.

I interpreted the comment by oli to be about #99552 my new FIXME. While orphan_check_trait_ref is now "visibly wrong" (or at least in need of some further thought), it does keep the previous behavior, so merging it as is was fine.

rust/compiler/rustc_trait_selection/src/traits/coherence.rs

Lines 720 to 723 in 7f93d4a

	// FIXME: Constants should participate in orphan checking.
	fn visit_const(&mut self, _c: ty::Const<'tcx>) -> ControlFlow<Self::BreakTy> {
	ControlFlow::CONTINUE
	}

[compiler-errors]

@bors r=lcnr

[bors]

📌 Commit b7cf9f7 has been approved by lcnr

It is now in the queue for this repository.