orphan check incorrectly handles projections

[lcnr]

We consider projections to be foreign types during orphan check

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

Lines 780 to 782 in 7ed044c

	| ty::Alias(ty::Projection | ty::Inherent | ty::Weak, ..) => {
	self.found_non_local_ty(ty)
	}

but do not consider them to be parameters when checking for uncovered types

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

Line 621 in ceeb5ad

} else if let ty::Param(_) = input_ty.kind() {

This is more obvious after #99552

this means that the following impl passes the orphan check even though it shouldn't

// crate a
pub trait Foreign<T, U> {
    type Assoc;
}

// crate b
use a::Foreign;

trait Id {
    type Assoc;
}

impl<T> Id for T {
    type Assoc = T;
}

pub struct B;
impl<T> Foreign<B, T> for <T as Id>::Assoc {
    type Assoc = usize;
}

The impl in b overlaps with an impl impl<U> Foreign<T, LocalTy> for LocalTy in another crate c which passes the orphan check.

While I wasn't able to cause runtime UB with this, I was able to get an ICE during codegen_fulfill_obligation: https://github.com/lcnr/orphan-check-ub

cc @rust-lang/types

[lcnr]

after #99552 has landed, we should be able to fix this by:

• rename found_param_ty to found_uncovered_ty
• also use that function for ty::Projection
• add tests + update the error message to also make sense for projections
• run crater

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-high

[atsuzaki]

@rustbot claim

[nikomatsakis]

Discussed in the types team on-site meetup thing:

Concluded that we should explore having the orphan check normalize first but having the "is knowable" check remain syntactic (not normalize).

[steffahn]

Adding some labels from #107377… @rustbot label A-associated-items, A-coherence

[fmease]

Reopening because the soundness issue still exists. However, since #117164 we now emit a future-incompat dont-report-in-deps warn-by-default lint for it. Tracked separately in #124559.