Improve comments in `needs_process_obligation`.

And a couple of other places.

compiler/rustc_data_structures/src/obligation_forest/mod.rs

@@ -426,6 +426,7 @@ impl<O: ForestObligation> ObligationForest<O> {
             // nodes. Therefore we use a `while` loop.
             let mut index = 0;
             while let Some(node) = self.nodes.get_mut(index) {
+                // This test is extremely hot.
                 if node.state.get() != NodeState::Pending
                     || !processor.needs_process_obligation(&node.obligation)
                 {
@@ -439,6 +440,7 @@ impl<O: ForestObligation> ObligationForest<O> {
                 // out of sync with `nodes`. It's not very common, but it does
                 // happen, and code in `compress` has to allow for it.
 
+                // This code is much less hot.
                 match processor.process_obligation(&mut node.obligation) {
                     ProcessResult::Unchanged => {
                         // No change in state.

compiler/rustc_trait_selection/src/traits/fulfill.rs

@@ -212,36 +212,44 @@ impl<'a, 'tcx> ObligationProcessor for FulfillProcessor<'a, 'tcx> {
 
     /// Identifies whether a predicate obligation needs processing.
     ///
-    /// This is always inlined, despite its size, because it has a single
+    /// This is always inlined because it has a single callsite and it is
-    /// callsite and it is called *very* frequently.
+    /// called *very* frequently. Be careful modifying this code! Several
+    /// compile-time benchmarks are very sensitive to even small changes.
     #[inline(always)]
     fn needs_process_obligation(&self, pending_obligation: &Self::Obligation) -> bool {
         // If we were stalled on some unresolved variables, first check whether
         // any of them have been resolved; if not, don't bother doing more work
         // yet.
-        match pending_obligation.stalled_on.len() {
+        let stalled_on = &pending_obligation.stalled_on;
-            // Match arms are in order of frequency, which matters because this
+        match stalled_on.len() {
-            // code is so hot. 1 and 0 dominate; 2+ is fairly rare.
+            // This case is the hottest most of the time, being hit up to 99%
-            1 => {
+            // of the time. `keccak` and `cranelift-codegen-0.82.1` are
-                let infer_var = pending_obligation.stalled_on[0];
+            // benchmarks that particularly stress this path.
-                self.selcx.infcx.ty_or_const_infer_var_changed(infer_var)
+            1 => self.selcx.infcx.ty_or_const_infer_var_changed(stalled_on[0]),
-            }
+
-            0 => {
+            // In this case we haven't changed, but wish to make a change. Note
-                // In this case we haven't changed, but wish to make a change.
+            // that this is a special case, and is not equivalent to the `_`
-                true
+            // case below, which would return `false` for an empty `stalled_on`
-            }
+            // vector.
-            _ => {
+            //
-                // This `for` loop was once a call to `all()`, but this lower-level
+            // This case is usually hit only 1% of the time or less, though it
-                // form was a perf win. See #64545 for details.
+            // reaches 20% in `wasmparser-0.101.0`.
-                (|| {
+            0 => true,
-                    for &infer_var in &pending_obligation.stalled_on {
+
-                        if self.selcx.infcx.ty_or_const_infer_var_changed(infer_var) {
+            // This case is usually hit only 1% of the time or less, though it
-                            return true;
+            // reaches 95% in `mime-0.3.16`, 64% in `wast-54.0.0`, and 12% in
-                        }
+            // `inflate-0.4.5`.
+            //
+            // The obvious way of writing this, with a call to `any()` and no
+            // closure, is currently slower than this version.
+            _ => (|| {
+                for &infer_var in stalled_on {
+                    if self.selcx.infcx.ty_or_const_infer_var_changed(infer_var) {
+                        return true;
                     }
-                    false
+                }
-                })()
+                false
-            }
+            })(),
         }
     }