do it for new solver too

compiler/rustc_trait_selection/src/solve/mod.rs

@@ -197,8 +197,27 @@ impl<'a, 'tcx> EvalCtxt<'a, InferCtxt<'tcx>> {
         goal: Goal<'tcx, (ty::Const<'tcx>, Ty<'tcx>)>,
     ) -> QueryResult<'tcx> {
         let (ct, ty) = goal.predicate;
-        self.eq(goal.param_env, ct.ty(), ty)?;
-        self.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
+
+        // FIXME(BoxyUwU): Really we should not be calling `ct.ty()` for any variant
+        // other than `ConstKind::Value`. Unfortunately this would require looking in the
+        // env for any `ConstArgHasType` assumptions for parameters and placeholders. I
+        // have not yet gotten around to implementing this though.
+        //
+        // We do still stall on infer vars though as otherwise a goal like:
+        // `ConstArgHasType(?x: usize, usize)` can succeed even though it might later
+        // get unified with some const that is not of type `usize`.
+        match ct.kind() {
+            ty::ConstKind::Infer(_) => {
+                self.evaluate_added_goals_and_make_canonical_response(Certainty::AMBIGUOUS)
+            }
+            ty::ConstKind::Error(_) => {
+                self.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
+            }
+            _ => {
+                self.eq(goal.param_env, ct.ty(), ty)?;
+                self.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
+            }
+        }
     }
 }