Rollup merge of #130617 - lcnr:nalgebra-hang-3, r=compiler-errors

bail if there are too many non-region infer vars in the query response

A minimal fix for the hang in nalgebra. If the query response would result in too many distinct non-region inference variables, simply overwrite the result with overflow. This should either happen if the result already has too many distinct type inference variables, or if evaluating the query encountered a lot of ambiguous associated types. In both cases it's straightforward to wait until the aliases are no longer ambiguous and then try again.

r? `@compiler-errors`

1 files changed, 11 insertions, 0 deletions

diff --git a/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs b/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs
index 5acfec3dee3..4d743c05190 100644
--- a/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs
+++ b/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs
@@ -157,6 +157,17 @@ where
             },
         );
 
+        // HACK: We bail with overflow if the response would have too many non-region
+        // inference variables. This tends to only happen if we encounter a lot of
+        // ambiguous alias types which get replaced with fresh inference variables
+        // during generalization. This prevents a hang in nalgebra.
+        let num_non_region_vars = canonical.variables.iter().filter(|c| !c.is_region()).count();
+        if num_non_region_vars > self.cx().recursion_limit() {
+            return Ok(self.make_ambiguous_response_no_constraints(MaybeCause::Overflow {
+                suggest_increasing_limit: true,
+            }));
+        }
+
         Ok(canonical)
     }