Rollup merge of #124936 - lcnr:cool-beans, r=compiler-errors

analyse visitor: build proof tree in probe

see inline comments

fixes #124791
fixes #124702

r? ```@compiler-errors```

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

@@ -409,6 +409,7 @@ pub(in crate::solve) fn make_canonical_state<'tcx, T: TypeFoldable<TyCtxt<'tcx>>
 /// This currently assumes that unifying the var values trivially succeeds.
 /// Adding any inference constraints which weren't present when originally
 /// computing the canonical query can result in bugs.
+#[instrument(level = "debug", skip(infcx, span, param_env))]
 pub(in crate::solve) fn instantiate_canonical_state<'tcx, T: TypeFoldable<TyCtxt<'tcx>>>(
     infcx: &InferCtxt<'tcx>,
     span: Span,

compiler/rustc_trait_selection/src/solve/inspect/analyse.rs

@@ -146,6 +146,11 @@ impl<'a, 'tcx> InspectCandidate<'a, 'tcx> {
     /// inference constraints, and optionally the args of an impl if this candidate
     /// came from a `CandidateSource::Impl`. This function modifies the state of the
     /// `infcx`.
+    #[instrument(
+        level = "debug",
+        skip_all,
+        fields(goal = ?self.goal.goal, nested_goals = ?self.nested_goals)
+    )]
     pub fn instantiate_nested_goals_and_opt_impl_args(
         &self,
         span: Span,
@@ -213,10 +218,23 @@ impl<'a, 'tcx> InspectCandidate<'a, 'tcx> {
                     };
                     let goal =
                         goal.with(infcx.tcx, ty::NormalizesTo { alias, term: unconstrained_term });
-                    let proof_tree = EvalCtxt::enter_root(infcx, GenerateProofTree::Yes, |ecx| {
-                        ecx.evaluate_goal_raw(GoalEvaluationKind::Root, GoalSource::Misc, goal)
-                    })
-                    .1;
+                    // We have to use a `probe` here as evaluating a `NormalizesTo` can constrain the
+                    // expected term. This means that candidates which only fail due to nested goals
+                    // and which normalize to a different term then the final result could ICE: when
+                    // building their proof tree, the expected term was unconstrained, but when
+                    // instantiating the candidate it is already constrained to the result of another
+                    // candidate.
+                    let proof_tree = infcx
+                        .probe(|_| {
+                            EvalCtxt::enter_root(infcx, GenerateProofTree::Yes, |ecx| {
+                                ecx.evaluate_goal_raw(
+                                    GoalEvaluationKind::Root,
+                                    GoalSource::Misc,
+                                    goal,
+                                )
+                            })
+                        })
+                        .1;
                     InspectGoal::new(
                         infcx,
                         self.goal.depth + 1,
@@ -225,13 +243,17 @@ impl<'a, 'tcx> InspectCandidate<'a, 'tcx> {
                         source,
                     )
                 }
-                _ => InspectGoal::new(
-                    infcx,
-                    self.goal.depth + 1,
-                    infcx.evaluate_root_goal(goal, GenerateProofTree::Yes).1.unwrap(),
-                    None,
-                    source,
-                ),
+                _ => {
+                    // We're using a probe here as evaluating a goal could constrain
+                    // inference variables by choosing one candidate. If we then recurse
+                    // into another candidate who ends up with different inference
+                    // constraints, we get an ICE if we already applied the constraints
+                    // from the chosen candidate.
+                    let proof_tree = infcx
+                        .probe(|_| infcx.evaluate_root_goal(goal, GenerateProofTree::Yes).1)
+                        .unwrap();
+                    InspectGoal::new(infcx, self.goal.depth + 1, proof_tree, None, source)
+                }
             })
             .collect();
 

tests/ui/traits/next-solver/coherence/ambiguity-causes-canonical-state-ice-1.rs

@@ -0,0 +1,43 @@
+//@ compile-flags: -Znext-solver=coherence
+//@ check-pass
+
+// A regression test for #124791. Computing ambiguity causes
+// for the overlap of the `ToString` impls caused an ICE.
+#![crate_type = "lib"]
+#![feature(min_specialization)]
+trait Display {}
+
+trait ToOwned {
+    type Owned;
+}
+
+impl<T> ToOwned for T {
+    type Owned = T;
+}
+
+struct Cow<B: ?Sized>(B);
+
+impl<B: ?Sized> Display for Cow<B>
+where
+    B: ToOwned,
+    B::Owned: Display,
+{
+}
+
+impl Display for () {}
+
+trait ToString {
+    fn to_string();
+}
+
+impl<T: Display + ?Sized> ToString for T {
+    default fn to_string() {}
+}
+
+impl ToString for Cow<str> {
+    fn to_string() {}
+}
+
+impl ToOwned for str {
+    type Owned = ();
+}

tests/ui/traits/next-solver/coherence/ambiguity-causes-canonical-state-ice-2.rs

@@ -0,0 +1,19 @@
+//@ compile-flags: -Znext-solver=coherence
+
+// A regression test for #124791. Computing ambiguity causes
+// for the overlap of the `ToString` impls caused an ICE.
+#![crate_type = "lib"]
+trait ToOwned {
+    type Owned;
+}
+impl<T> ToOwned for T {
+    type Owned = u8;
+}
+impl ToOwned for str {
+    type Owned = i8;
+}
+
+trait Overlap {}
+impl<T: ToOwned<Owned = i8> + ?Sized> Overlap for T {}
+impl Overlap for str {}
+//~^ ERROR conflicting implementations of trait `Overlap`

tests/ui/traits/next-solver/coherence/ambiguity-causes-canonical-state-ice-2.stderr

@@ -0,0 +1,11 @@
+error[E0119]: conflicting implementations of trait `Overlap` for type `str`
+  --> $DIR/ambiguity-causes-canonical-state-ice-2.rs:18:1
+   |
+LL | impl<T: ToOwned<Owned = i8> + ?Sized> Overlap for T {}
+   | --------------------------------------------------- first implementation here
+LL | impl Overlap for str {}
+   | ^^^^^^^^^^^^^^^^^^^^ conflicting implementation for `str`
+
+error: aborting due to 1 previous error
+
+For more information about this error, try `rustc --explain E0119`.

tests/ui/traits/next-solver/cycles/coinduction/incompleteness-unstable-result.rs

@@ -61,7 +61,7 @@ where
     // entering the cycle from `A` fails, but would work if we were to use the cache
     // result of `B<X>`.
     impls_trait::<A<X>, _, _, _>();
-    //~^ ERROR the trait bound `A<X>: Trait<i32, u8, u8>` is not satisfied
+    //~^ ERROR the trait bound `A<X>: Trait<_, _, _>` is not satisfied
 }
 
 fn main() {

tests/ui/traits/next-solver/cycles/coinduction/incompleteness-unstable-result.stderr

@@ -1,10 +1,11 @@
-error[E0277]: the trait bound `A<X>: Trait<i32, u8, u8>` is not satisfied
+error[E0277]: the trait bound `A<X>: Trait<_, _, _>` is not satisfied
   --> $DIR/incompleteness-unstable-result.rs:63:19
    |
 LL |     impls_trait::<A<X>, _, _, _>();
-   |                   ^^^^ the trait `Trait<i32, u8, u8>` is not implemented for `A<X>`, which is required by `A<X>: Trait<_, _, _>`
+   |                   ^^^^ the trait `Trait<_, _, _>` is not implemented for `A<X>`, which is required by `A<X>: Trait<_, _, _>`
    |
-note: required for `A<X>` to implement `Trait<i32, u8, u8>`
+   = help: the trait `Trait<U, V, D>` is implemented for `A<T>`
+note: required for `A<X>` to implement `Trait<_, _, _>`
   --> $DIR/incompleteness-unstable-result.rs:32:50
    |
 LL | impl<T: ?Sized, U: ?Sized, V: ?Sized, D: ?Sized> Trait<U, V, D> for A<T>
@@ -13,16 +14,12 @@ LL | impl<T: ?Sized, U: ?Sized, V: ?Sized, D: ?Sized> Trait<U, V, D> for A<T>
 LL |     A<T>: Trait<U, D, V>,
    |           -------------- unsatisfied trait bound introduced here
    = note: 8 redundant requirements hidden
-   = note: required for `A<X>` to implement `Trait<i32, u8, u8>`
+   = note: required for `A<X>` to implement `Trait<_, _, _>`
 note: required by a bound in `impls_trait`
   --> $DIR/incompleteness-unstable-result.rs:51:28
    |
 LL | fn impls_trait<T: ?Sized + Trait<U, V, D>, U: ?Sized, V: ?Sized, D: ?Sized>() {}
    |                            ^^^^^^^^^^^^^^ required by this bound in `impls_trait`
-help: consider extending the `where` clause, but there might be an alternative better way to express this requirement
-   |
-LL |     X: IncompleteGuidance<u32, i16>, A<X>: Trait<i32, u8, u8>
-   |                                    ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 error: aborting due to 1 previous error