perf(semaphore): avoid the queue lock for positive-balance releases

CHANGELOG.md

@@ -32,6 +32,10 @@ All notable changes to this project will be documented in this file.
 * Prevent deadlocks in `singleflight::Group::work` and `try_work` when a duplicate key's destructor calls back into the same group.
 * Fix undefined behavior when sending or dropping a oneshot sender races with receiver completion or cancellation.
 
+### Improvements
+
+* Add a lock-free fast path for releasing `Semaphore` permits while permits remain available, which speeds up dropping `RwLock` read guards, releasing multi-permit `Semaphore` permits, and returning objects to a bounded pool that is not at capacity. `Mutex` releases are unchanged.
+
 ### Notable changes
 
 * Executor waker operations, including cloning, waking, and dropping, are expected not to panic; recovery from panicking waker callbacks is no longer supported.

asyncband/src/broadcast/mpmc/bounded/mod.rs

@@ -195,7 +195,7 @@ impl<T> Shared<T> {
         // subscription.
         let waiting = self.waiting_senders();
         if freed > 0 && waiting > 0 {
-            self.tx_permits.release_if_nonempty(freed.min(waiting));
+            self.tx_permits.release_if_waiters(freed.min(waiting));
         }
     }
 
@@ -323,7 +323,7 @@ impl<T> BoundedSender<T> {
                 };
 
                 loop {
-                    // Enqueue before rechecking. `release_if_nonempty` adds nothing when no
+                    // Enqueue before rechecking. `release_if_waiters` adds nothing when no
                     // producer is queued, so a reclaim landing between the recheck and the
                     // registration would otherwise drop its wake-up and park this producer for
                     // good. Registering first orders this producer's semaphore acquisition ahead