Unknown types implement all traits

[lnicola]

[flodiebold]

iterate_method_candidates_with_autoref should probably just return an empty result when called with a free variable as receiver.

[iDawer]

I guess this happened after #10373 ? 😄
I bet that here
https://github.com/rust-analyzer/rust-analyzer/blob/cf1ea9d0b9241265369b3ba869795ac9bda9f146/crates/hir/src/lib.rs#L2592-L2602
the {error} type gets replaced with bound variable and that may implement all(?) traits in scope. May be shouldn't call replace_errors_with_variables on the {error} type itself?

[flodiebold]

I guess this happened after #10373 ?

Yes.

May be shouldn't call replace_errors_with_variables on the {error} type itself?

No, that's fine. Note that the type of x in the example is a type variable originally. The problem here is just trying to do method resolution on a method receiver that's a variable, which rustc will just refuse to do.

[lnicola]

Something like:

diff --git i/crates/hir_ty/src/method_resolution.rs w/crates/hir_ty/src/method_resolution.rs
index 8e6ab8af0..5ba637ef4 100644
--- i/crates/hir_ty/src/method_resolution.rs
+++ w/crates/hir_ty/src/method_resolution.rs
@@ -581,6 +581,10 @@ fn iterate_method_candidates_by_receiver(
     name: Option<&Name>,
     mut callback: &mut dyn FnMut(&Canonical<Ty>, AssocItemId) -> ControlFlow<()>,
 ) -> ControlFlow<()> {
+    match receiver_ty.value.kind(&Interner) {
+        chalk_ir::TyKind::BoundVar(_) => return ControlFlow::Break(()),
+        _ => {}
+    }
     // We're looking for methods with *receiver* type receiver_ty. These could
     // be found in any of the derefs of receiver_ty, so we have to go through
     // that.

Helps in this case, but seems to break for integral types.

(Not sure if that's supposed to be a BoundVar).

[flodiebold]

Yeah, it's supposed to be a bound var. We can check on the binders of the canonical whether it's a general or an int/float variable.

[iDawer]

From my limited understanding Chalk treats bound vars as existentially quantified and placeholders as universally quantified, right? Since hir_ty::replace_errors_with_variables uses bound vars it getting ambiguous solution for queries like exists<T> { T: AnyTrait }.

Cant remember but I saw somewhere talking about special handling of sized bound on type vars (or errors?) on Chalk side. So could we hack it on RA side with FromEnv({error}: Sized)? I applied this in hir::Type::iterate_method_candidates_dyn and it seems solved both this and #10297 issues.

Thinking on how the hack could break, the {error} type will satisfy : Sized and : ?Sized bounds. Luckily, : !Sized bounds are impossible 😄

[flodiebold]

Bound vars can be existentially or universally quantified, it depends on what they refer to. And no, we shouldn't add hacks for the error type; replacing the errors back with variables, which is what they were during type inference, is the correct thing to do.

[kornelski]

Please consider special-casing this for #8468 by disabling auto-imports. Auto-complete suggestions of potentially irrelevant methods isn't a big deal. However, auto-import of definitely incorrect traits that break correctly-picked inherent methods is a real problem.