instantiate response: no unnecessary new universe

compiler/rustc_trait_selection/src/solve/eval_ctxt/canonical.rs

@@ -215,7 +215,7 @@ impl<'tcx> EvalCtxt<'_, 'tcx> {
         // created inside of the query directly instead of returning them to the
         // caller.
         let prev_universe = self.infcx.universe();
-        let universes_created_in_query = response.max_universe.index() + 1;
+        let universes_created_in_query = response.max_universe.index();
         for _ in 0..universes_created_in_query {
             self.infcx.create_next_universe();
         }

tests/ui/traits/new-solver/generalize/generalize-proj-new-universe-index-1.rs

@@ -0,0 +1,73 @@
+// compile-flags: -Ztrait-solver=next
+// check-pass
+
+// A minimization of an ambiguity when using typenum. See
+// https://github.com/rust-lang/trait-system-refactor-initiative/issues/55
+// for more details.
+trait Id {
+    type Assoc: ?Sized;
+}
+impl<T: ?Sized> Id for T {
+    type Assoc = T;
+}
+
+trait WithAssoc<T: ?Sized> {
+    type Assoc: ?Sized;
+}
+
+
+struct Leaf;
+struct Wrapper<U: ?Sized>(U);
+
+impl<U: ?Sized> WithAssoc<U> for Leaf {
+    type Assoc = U;
+}
+
+impl<Ul: ?Sized, Ur: ?Sized> WithAssoc<Wrapper<Ur>> for Wrapper<Ul>
+where
+    Ul: WithAssoc<Ur>,
+{
+    type Assoc = <<Ul as WithAssoc<Ur>>::Assoc as Id>::Assoc;
+}
+
+fn bound<T: ?Sized, U: ?Sized, V: ?Sized>()
+where
+    T: WithAssoc<U, Assoc = V>,
+{
+}
+
+// normalize self type to `Wrapper<Leaf>`
+//   This succeeds, HOWEVER, instantiating the query response previously
+//   incremented the universe index counter.
+// equate impl headers:
+//      <Wrapper<Leaf> as WithAssoc<<Wrapper<Leaf> as Id>::Assoc>>
+//      <Wrapper<?2t> as WithAssoc<Wrapper<?3t>>>
+// ~> AliasRelate(<Wrapper<Leaf> as Id>::Assoc, Equate, Wrapper<?3t>)
+// add where bounds:
+// ~> Leaf: WithAssoc<?3t>
+// equate with assoc type:
+//      ?0t
+//      <Leaf as WithAssoc<?3t>>::Assoc as Id>::Assoc
+// ~> AliasRelate(
+//      <<Leaf as WithAssoc<?3t>>::Assoc as Id>::Assoc,
+//      Equate,
+//      <<Leaf as WithAssoc<?4t>>::Assoc as Id>::Assoc,
+//    )
+//
+// We do not reuse `?3t` during generalization because `?0t` cannot name `?4t` as we created
+// it after incrementing the universe index while normalizing the self type.
+//
+// evaluate_added_goals_and_make_query_response:
+//    AliasRelate(<Wrapper<Leaf> as Id>::Assoc, Equate, Wrapper<?3t>)
+//      YES, constrains ?3t to Leaf
+//    AliasRelate(
+//      <<Leaf as WithAssoc<Leaf>>::Assoc as Id>::Assoc,
+//      Equate,
+//      <<Leaf as WithAssoc<?4t>>::Assoc as Id>::Assoc,
+//    )
+//
+// Normalizing <<Leaf as WithAssoc<?4t>>::Assoc as Id>::Assoc then *correctly*
+// results in ambiguity.
+fn main() {
+    bound::<<Wrapper<Leaf> as Id>::Assoc, <Wrapper<Leaf> as Id>::Assoc, _>()
+}

tests/ui/traits/new-solver/generalize/generalize-proj-new-universe-index-2.rs

@@ -0,0 +1,75 @@
+// compile-flags: -Ztrait-solver=next
+// known-bug: trait-system-refactor-initiative#60
+
+// Generalizing a projection containing an inference variable
+// which cannot be named by the `root_vid` can result in ambiguity.
+//
+// Because we do not decrement the universe index when exiting a forall,
+// this can cause unexpected failures.
+//
+// See generalize-proj-new-universe-index-1.rs for more details.
+
+// For this reproduction we need:
+// - an inference variable with a lower universe
+// - enter a binder to increment the current universe
+// - create a new inference variable which is constrained by proving a goal
+// - equate a projection containing the new variable with the first variable
+// - generalization creates yet another inference variable which is then
+//   part of an alias-relate, resulting this to fail with ambiguity.
+//
+// Because we need to enter the binder in-between the creation of the first
+// and second inference variable, this is easiest via
+// `assemble_candidates_after_normalizing_self_ty` because eagerly call
+// `try_evaluate_added_goals` there before creating the inference variables
+// for the impl parameters.
+trait Id {
+    type Assoc: ?Sized;
+}
+impl<T: ?Sized> Id for T {
+    type Assoc = T;
+}
+
+// By adding an higher ranked bound to the impl we currently
+// propagate this bound to the caller, forcing us to create a new
+// universe.
+trait IdHigherRankedBound {
+    type Assoc: ?Sized;
+}
+
+impl<T: ?Sized> IdHigherRankedBound for T
+where
+    for<'a> T: 'a,
+{
+    type Assoc = T;
+}
+
+trait WithAssoc<T: ?Sized> {
+    type Assoc: ?Sized;
+}
+
+
+struct Leaf;
+struct Wrapper<U: ?Sized>(U);
+struct Rigid;
+
+impl<U: ?Sized> WithAssoc<U> for Leaf {
+    type Assoc = U;
+}
+
+
+impl<Ur: ?Sized> WithAssoc<Wrapper<Ur>> for Rigid
+where
+    Leaf: WithAssoc<Ur>,
+{
+    type Assoc = <<Leaf as WithAssoc<Ur>>::Assoc as Id>::Assoc;
+}
+
+fn bound<T: ?Sized, U: ?Sized, V: ?Sized>()
+where
+    T: WithAssoc<U, Assoc = V>,
+{
+}
+
+fn main() {
+    bound::<<Rigid as IdHigherRankedBound>::Assoc, <Wrapper<Leaf> as Id>::Assoc, _>()
+}

tests/ui/traits/new-solver/generalize/generalize-proj-new-universe-index-2.stderr

@@ -0,0 +1,9 @@
+error[E0282]: type annotations needed
+  --> $DIR/generalize-proj-new-universe-index-2.rs:74:5
+   |
+LL |     bound::<<Rigid as IdHigherRankedBound>::Assoc, <Wrapper<Leaf> as Id>::Assoc, _>()
+   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ cannot infer type of the type parameter `V` declared on the function `bound`
+
+error: aborting due to previous error
+
+For more information about this error, try `rustc --explain E0282`.