[SYCL] Do not lock unconditionally while access queue_iml::MMissedCleanupRequests

sycl/source/detail/queue_impl.cpp

@@ -618,12 +618,12 @@ void queue_impl::wait(const detail::code_location &CodeLoc) {
     WeakEvents.swap(MEventsWeak);
     SharedEvents.swap(MEventsShared);
 
-    {
-      std::lock_guard<std::mutex> RequestLock(MMissedCleanupRequestsMtx);
-      for (auto &UpdatedGraph : MMissedCleanupRequests)
-        doUnenqueuedCommandCleanup(UpdatedGraph);
-      MMissedCleanupRequests.clear();
-    }
+    MMissedCleanupRequests.unset(
+        [&](MissedCleanupRequestsType &MissedCleanupRequests) {
+          for (auto &UpdatedGraph : MissedCleanupRequests)
+            doUnenqueuedCommandCleanup(UpdatedGraph);
+          MissedCleanupRequests.clear();
+        });
   }
   // If the queue is either a host one or does not support OOO (and we use
   // multiple in-order queues as a result of that), wait for each event
@@ -800,8 +800,10 @@ void queue_impl::revisitUnenqueuedCommandsState(
   if (Lock.owns_lock())
     doUnenqueuedCommandCleanup(CompletedHostTask->getCommandGraph());
   else {
-    std::lock_guard<std::mutex> RequestLock(MMissedCleanupRequestsMtx);
-    MMissedCleanupRequests.push_back(CompletedHostTask->getCommandGraph());
+    MMissedCleanupRequests.set(
+        [&](MissedCleanupRequestsType &MissedCleanupRequests) {
+          MissedCleanupRequests.push_back(CompletedHostTask->getCommandGraph());
+        });
   }
 }
 

sycl/source/detail/queue_impl.hpp

@@ -838,12 +838,12 @@ class queue_impl {
     // dependency so in the case where some commands were not enqueued
     // (blocked), we track them to prevent barrier from being enqueued
     // earlier.
-    {
-      std::lock_guard<std::mutex> RequestLock(MMissedCleanupRequestsMtx);
-      for (auto &UpdatedGraph : MMissedCleanupRequests)
-        doUnenqueuedCommandCleanup(UpdatedGraph);
-      MMissedCleanupRequests.clear();
-    }
+    MMissedCleanupRequests.unset(
+        [&](MissedCleanupRequestsType &MissedCleanupRequests) {
+          for (auto &UpdatedGraph : MissedCleanupRequests)
+            doUnenqueuedCommandCleanup(UpdatedGraph);
+          MissedCleanupRequests.clear();
+        });
     auto &Deps = MGraph.expired() ? MDefaultGraphDeps : MExtGraphDeps;
     if (Type == CGType::Barrier && !Deps.UnenqueuedCmdEvents.empty()) {
       Handler.depends_on(Deps.UnenqueuedCmdEvents);
@@ -1025,6 +1025,36 @@ class queue_impl {
     }
   } MDefaultGraphDeps, MExtGraphDeps;
 
+  // Implement check-lock-check pattern to not lock empty MData as the locks
+  // come with runtime overhead.
+  template <typename DataType> class CheckLockCheck {
+    DataType MData;
+    std::atomic_bool MIsSet = false;
+    mutable std::mutex MDataMtx;
+
+  public:
+    template <typename F> void set(F &&func) {
+      std::lock_guard<std::mutex> Lock(MDataMtx);
+      MIsSet.store(true, std::memory_order_release);
+      std::forward<F>(func)(MData);
+    }
+    template <typename F> void unset(F &&func) {
+      if (MIsSet.load(std::memory_order_acquire)) {
+        std::lock_guard<std::mutex> Lock(MDataMtx);
+        if (MIsSet.load(std::memory_order_acquire)) {
+          std::forward<F>(func)(MData);
+          MIsSet.store(false, std::memory_order_release);
+        }
+      }
+    }
+    DataType read() {
+      if (!MIsSet.load(std::memory_order_acquire))
+        return DataType{};
+      std::lock_guard<std::mutex> Lock(MDataMtx);
+      return MData;
+    }
+  };
+
   const bool MIsInorder;
 
   std::vector<EventImplPtr> MStreamsServiceEvents;
@@ -1071,9 +1101,9 @@ class queue_impl {
   unsigned long long MQueueID;
   static std::atomic<unsigned long long> MNextAvailableQueueID;
 
-  std::deque<std::shared_ptr<ext::oneapi::experimental::detail::graph_impl>>
-      MMissedCleanupRequests;
-  std::mutex MMissedCleanupRequestsMtx;
+  using MissedCleanupRequestsType = std::deque<
+      std::shared_ptr<ext::oneapi::experimental::detail::graph_impl>>;
+  CheckLockCheck<MissedCleanupRequestsType> MMissedCleanupRequests;
 
   friend class sycl::ext::oneapi::experimental::detail::node_impl;