Rollup merge of rust-lang#64420 - nnethercote:inline-mark_neighbours_…

…as_waiting_from, r=Mark-Simulacrum

Inline `mark_neighbours_as_waiting_from`.

This function is very hot, doesn't get inlined because it's recursive,
and the function calls are significant.

This commit splits it into inlined and uninlined variants, and uses the
inlined variant for the hot call site. This wins several percent on a
few benchmarks.

r? @nikomatsakis

src/librustc_data_structures/obligation_forest/mod.rs

@@ -559,13 +559,20 @@ impl<O: ForestObligation> ObligationForest<O> {
         trace
     }
 
-    #[inline]
-    fn mark_neighbors_as_waiting_from(&self, node: &Node<O>) {
+    // This always-inlined function is for the hot call site.
+    #[inline(always)]
+    fn inlined_mark_neighbors_as_waiting_from(&self, node: &Node<O>) {
         for dependent in node.parent.iter().chain(node.dependents.iter()) {
             self.mark_as_waiting_from(&self.nodes[dependent.get()]);
         }
     }
 
+    // This never-inlined function is for the cold call site.
+    #[inline(never)]
+    fn uninlined_mark_neighbors_as_waiting_from(&self, node: &Node<O>) {
+        self.inlined_mark_neighbors_as_waiting_from(node)
+    }
+
     /// Marks all nodes that depend on a pending node as `NodeState::Waiting`.
     fn mark_as_waiting(&self) {
         for node in &self.nodes {
@@ -576,7 +583,8 @@ impl<O: ForestObligation> ObligationForest<O> {
 
         for node in &self.nodes {
             if node.state.get() == NodeState::Pending {
-                self.mark_neighbors_as_waiting_from(node);
+                // This call site is hot.
+                self.inlined_mark_neighbors_as_waiting_from(node);
             }
         }
     }
@@ -588,7 +596,8 @@ impl<O: ForestObligation> ObligationForest<O> {
             NodeState::Pending | NodeState::Done => {},
         }
 
-        self.mark_neighbors_as_waiting_from(node);
+        // This call site is cold.
+        self.uninlined_mark_neighbors_as_waiting_from(node);
     }
 
     /// Compresses the vector, removing all popped nodes. This adjusts