Auto merge of #146329 - lcnr:opaque-type-infer-alias-candidates, r=BoxyUwU

consider item bounds for non-yet-defined opaque types

Based on rust-lang/rust#140405.

fixes https://github.com/rust-lang/trait-system-refactor-initiative/issues/182
fixes https://github.com/rust-lang/trait-system-refactor-initiative/issues/196
fixes https://github.com/rust-lang/trait-system-refactor-initiative/issues/205

there's some jank here, see https://github.com/rust-lang/trait-system-refactor-initiative/issues/229

## Design

If the self type is an inference variable which has been sub-unified with am opaque type, we need to incompletely guide inference to avoid breakage.

In this case, we
- look at the item bounds of all sub-unified opaque types, and
- blanket impls which do not constrain the self type

Even if there are applicable candidates, we always force their certainty to be `Maybe`, so they will always have to be reproven once we've constrained the inference variable.

This is a bit iffy, see the added tests.

r? `@BoxyUwU`

1 files changed, 2 insertions, 6 deletions

diff --git a/compiler/rustc_next_trait_solver/src/solve/mod.rs b/compiler/rustc_next_trait_solver/src/solve/mod.rs
index db3460c4ff6..cd27c9c26c1 100644
--- a/compiler/rustc_next_trait_solver/src/solve/mod.rs
+++ b/compiler/rustc_next_trait_solver/src/solve/mod.rs
@@ -24,7 +24,7 @@ mod trait_goals;
 use derive_where::derive_where;
 use rustc_type_ir::inherent::*;
 pub use rustc_type_ir::solve::*;
-use rustc_type_ir::{self as ty, Interner, TypingMode};
+use rustc_type_ir::{self as ty, Interner, TyVid, TypingMode};
 use tracing::instrument;
 
 pub use self::eval_ctxt::{
@@ -418,11 +418,6 @@ pub struct GoalEvaluation<I: Interner> {
     pub has_changed: HasChanged,
     /// If the [`Certainty`] was `Maybe`, then keep track of whether the goal has changed
     /// before rerunning it.
-    ///
-    /// We knowingly ignore the `sub_root` of our inference variables here. This means we
-    /// may not reevaluate a goal even though a change to the `sub_root` could cause a goal
-    /// to make progress. Tracking them adds additional complexity for an incredibly minor
-    /// type inference improvement. We could look into properly handling this in the future.
     pub stalled_on: Option<GoalStalledOn<I>>,
 }
 
@@ -431,6 +426,7 @@ pub struct GoalEvaluation<I: Interner> {
 pub struct GoalStalledOn<I: Interner> {
     pub num_opaques: usize,
     pub stalled_vars: Vec<I::GenericArg>,
+    pub sub_roots: Vec<TyVid>,
     /// The cause that will be returned on subsequent evaluations if this goal remains stalled.
     pub stalled_cause: MaybeCause,
 }