Rollup merge of rust-lang#139828 - compiler-errors:rigid-trait, r=lcnr

Don't require rigid alias's trait to hold

See test for write-up. TL;DR is that we don't need the trait bound to hold, since we enforce it during WF.

I think this is preferable to introducing (if we even could do so) a more specific hack around coroutine interiors, higher ranked types, etc, since this is just a manifestation of more pervasive issues w/ lifetime erasure in coroutines. This just doesn't manifest in the old solver b/c it doesn't try to prove `T: Trait` holds when rigidly projecting `<T as Trait>::Assoc`.

It's pretty clear that this affects quite a few traits (rust-lang#139763), so I think this needs fixing.

r? lcnr

Fixes rust-lang/trait-system-refactor-initiative#177

Author: matthiaskrgr

compiler/rustc_next_trait_solver/src/solve/normalizes_to/mod.rs

@@ -45,7 +45,6 @@ where
                             goal,
                             goal.predicate.alias,
                         );
-                        this.add_goal(GoalSource::AliasWellFormed, goal.with(cx, trait_ref));
                         this.evaluate_added_goals_and_make_canonical_response(Certainty::Yes)
                     })
                 })

tests/ui/coroutine/higher-ranked-rigid.rs

@@ -0,0 +1,41 @@
+//@ edition: 2024
+//@ revisions: current next
+//@ ignore-compare-mode-next-solver (explicit revisions)
+//@[next] compile-flags: -Znext-solver
+//@ check-pass
+
+// Regression test for <https://github.com/rust-lang/trait-system-refactor-initiative/issues/177>.
+// Coroutines erase all free lifetimes from their interior types, replacing them with higher-
+// ranked regions which act as universals, to properly represent the fact that we don't know what
+// the value of the region is within the coroutine.
+//
+// In the future in `from_request`, that means that the `'r` lifetime is being replaced in
+// `<T as FromRequest<'r>>::Assoc`, which is in present in the existential bounds of the
+// `dyn Future` that it's awaiting. Normalizing this associated type, with its free lifetimes
+// replaced, means proving `T: FromRequest<'!0>`, which doesn't hold without constraining the
+// `'!0` lifetime, which we don't do today.
+
+// Proving `T: Trait` holds when `<T as Trait>::Assoc` is rigid is not necessary for soundness,
+// at least not *yet*, and it's not even necessary for diagnostics since we have other special
+// casing for, e.g., AliasRelate goals failing in the BestObligation folder.
+
+// The old solver unintentioanlly avoids this by never checking that `T: Trait` holds when
+// `<T as Trait>::Assoc` is rigid. Introducing this additional requirement when projecting rigidly
+// in the old solver causes this (and tons of production crates) to fail. See the fallout from the
+// crater run at <https://github.com/rust-lang/rust/pull/139763>.
+
+use std::future::Future;
+use std::pin::Pin;
+
+pub trait FromRequest<'r> {
+    type Assoc;
+    fn from_request() -> Pin<Box<dyn Future<Output = Self::Assoc> + Send>>;
+}
+
+fn test<'r, T: FromRequest<'r>>() -> Pin<Box<dyn Future<Output = ()> + Send>> {
+    Box::pin(async move {
+        T::from_request().await;
+    })
+}
+
+fn main() {}