More comments

compiler/rustc_trait_selection/src/solve/alias_relate.rs

@@ -2,15 +2,29 @@
 //! Doing this via a separate goal is called "deferred alias relation" and part
 //! of our more general approach to "lazy normalization".
 //!
-//! This goal, e.g. `A alias-relate B`, may be satisfied by one of three branches:
+//! This is done by first normalizing both sides of the goal, ending up in
-//! * normalizes-to: If `A` is a projection, we can prove the equivalent
+//! either a concrete type, rigid projection, opaque, or an infer variable.
-//!   projection predicate with B as the right-hand side of the projection.
+//! These are related further according to the rules below:
-//!   This goal is computed in both directions, if both are aliases.
+//!
-//! * subst-relate: Equate `A` and `B` by their substs, if they're both
+//! (1.) If we end up with a rigid projection and a rigid projection, then we
-//!   aliases with the same def-id.
+//! relate those projections structurally.
-//! * bidirectional-normalizes-to: If `A` and `B` are both projections, and both
+//!
-//!   may apply, then we can compute the "intersection" of both normalizes-to by
+//! (2.) If we end up with a rigid projection and an alias, then the opaque will
-//!   performing them together. This is used specifically to resolve ambiguities.
+//! have its hidden type defined to be that rigid projection.
+//!
+//! (3.) If we end up with an opaque and an opaque, then we assemble two
+//! candidates, one defining the LHS to be the hidden type of the RHS, and vice
+//! versa.
+//!
+//! (4.) If we end up with an infer var and an opaque or rigid projection, then
+//! we assign the alias to the infer var.
+//!
+//! (5.) If we end up with an opaque and a rigid (non-projection) type, then we
+//! define the hidden type of the opaque to be the rigid type.
+//!
+//! (6.) Otherwise, if we end with two rigid (non-projection) or infer types,
+//! relate them structurally.
+
 use super::{EvalCtxt, GoalSource};
 use rustc_infer::infer::DefineOpaqueTypes;
 use rustc_infer::traits::query::NoSolution;
@@ -50,6 +64,7 @@ impl<'tcx> EvalCtxt<'_, 'tcx> {
                     self.relate(param_env, lhs, variance, rhs)?;
                     self.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                 } else if alias.is_opaque(tcx) {
+                    // FIXME: This doesn't account for variance.
                     self.define_opaque(param_env, alias, rhs)
                 } else {
                     Err(NoSolution)
@@ -60,6 +75,7 @@ impl<'tcx> EvalCtxt<'_, 'tcx> {
                     self.relate(param_env, lhs, variance, rhs)?;
                     self.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                 } else if alias.is_opaque(tcx) {
+                    // FIXME: This doesn't account for variance.
                     self.define_opaque(param_env, alias, lhs)
                 } else {
                     Err(NoSolution)
@@ -72,6 +88,7 @@ impl<'tcx> EvalCtxt<'_, 'tcx> {
         }
     }
 
+    // FIXME: This needs a name that reflects that it's okay to bottom-out with an inference var.
     /// Normalize the `term` to equate it later. This does not define opaque types.
     #[instrument(level = "debug", skip(self, param_env), ret)]
     fn try_normalize_term(

compiler/rustc_trait_selection/src/solve/mod.rs

@@ -317,7 +317,10 @@ impl<'tcx> EvalCtxt<'_, 'tcx> {
             return Some(ty);
         };
 
-        // We do no always define opaque types eagerly to allow non-defining uses in the defining scope.
+        // We do no always define opaque types eagerly to allow non-defining uses
+        // in the defining scope. However, if we can unify this opaque to an existing
+        // opaque, then we should attempt to eagerly reveal the opaque, and we fall
+        // through.
         if let DefineOpaqueTypes::No = define_opaque_types
             && let Reveal::UserFacing = param_env.reveal()
             && let ty::Opaque = kind