Rollup merge of rust-lang#123302 - compiler-errors:sized-bound-first, r=estebank

Make sure to insert `Sized` bound first into clauses list

rust-lang#120323 made it so that we don't insert an implicit `Sized` bound whenever we see an *explicit* `Sized` bound. However, since the code that inserts implicit sized bounds puts the bound as the *first* in the list, that means that it had the **side-effect** of possibly meaning we check `Sized` *after* checking other trait bounds.

If those trait bounds result in ambiguity or overflow or something, it may change how we winnow candidates. (**edit: SEE** rust-lang#123303) This is likely the cause for the regression in rust-lang#123279 (comment), since the impl...

```rust
impl<T: Job + Sized> AsJob for T { // <----- changing this to `Sized + Job` or just `Job` (which turns into `Sized + Job`) will FIX the issue.
}
```

...looks incredibly suspicious.

Fixes [after beta-backport] rust-lang#123279.

Alternative is to revert rust-lang#120323. I don't have a strong opinion about this, but think it may be nice to keep the diagnostic changes around.

Author: matthiaskrgr

compiler/rustc_hir_analysis/src/bounds.rs

@@ -54,14 +54,20 @@ impl<'tcx> Bounds<'tcx> {
         span: Span,
         polarity: ty::PredicatePolarity,
     ) {
-        self.clauses.push((
+        let clause = (
             trait_ref
                 .map_bound(|trait_ref| {
                     ty::ClauseKind::Trait(ty::TraitPredicate { trait_ref, polarity })
                 })
                 .to_predicate(tcx),
             span,
-        ));
+        );
+        // FIXME(-Znext-solver): We can likely remove this hack once the new trait solver lands.
+        if tcx.lang_items().sized_trait() == Some(trait_ref.def_id()) {
+            self.clauses.insert(0, clause);
+        } else {
+            self.clauses.push(clause);
+        }
     }
 
     pub fn push_projection_bound(

tests/incremental/hashes/function_interfaces.rs

@@ -217,7 +217,7 @@ pub fn second_trait_bound<T: Eq + Clone>() {}
 pub fn second_builtin_bound<T: Send        >() {}
 
 #[cfg(not(any(cfail1,cfail4)))]
-#[rustc_clean(cfg = "cfail2", except = "opt_hir_owner_nodes, predicates_of")]
+#[rustc_clean(cfg = "cfail2", except = "opt_hir_owner_nodes")]
 #[rustc_clean(cfg = "cfail3")]
 #[rustc_clean(cfg = "cfail5", except = "opt_hir_owner_nodes, predicates_of")]
 #[rustc_clean(cfg = "cfail6")]

tests/ui/methods/probe-overflow-due-to-sized-predicate-ordering.rs

@@ -0,0 +1,30 @@
+//@ check-pass
+// Regression test due to #123279
+
+pub trait Job: AsJob {
+    fn run_once(&self);
+}
+
+impl<F: Fn()> Job for F {
+    fn run_once(&self) {
+        todo!()
+    }
+}
+
+pub trait AsJob {}
+
+// Ensure that `T: Sized + Job` by reordering the explicit `Sized` to where
+// the implicit sized pred would go.
+impl<T: Job + Sized> AsJob for T {}
+
+pub struct LoopingJobService {
+    job: Box<dyn Job>,
+}
+
+impl Job for LoopingJobService {
+    fn run_once(&self) {
+        self.job.run_once()
+    }
+}
+
+fn main() {}