[SYCL] Fix command cleanup invoked from multiple threads

sycl/source/detail/event_impl.cpp

@@ -95,16 +95,14 @@ event_impl::event_impl(QueueImplPtr Queue) : MQueue(Queue) {
 
 void event_impl::wait(
     std::shared_ptr<cl::sycl::detail::event_impl> Self) const {
-
   if (MEvent)
     // presence of MEvent means the command has been enqueued, so no need to
     // go via the slow path event waiting in the scheduler
     waitInternal();
   else if (MCommand)
-    detail::Scheduler::getInstance().waitForEvent(std::move(Self));
+    detail::Scheduler::getInstance().waitForEvent(Self);
   if (MCommand && !SYCLConfig<SYCL_DISABLE_EXECUTION_GRAPH_CLEANUP>::get())
-    detail::Scheduler::getInstance().cleanupFinishedCommands(
-        static_cast<Command *>(MCommand));
+    detail::Scheduler::getInstance().cleanupFinishedCommands(std::move(Self));
 }
 
 void event_impl::wait_and_throw(

sycl/source/detail/scheduler/scheduler.cpp

@@ -123,9 +123,11 @@ void Scheduler::waitForEvent(EventImplPtr Event) {
   GraphProcessor::waitForEvent(std::move(Event));
 }
 
-void Scheduler::cleanupFinishedCommands(Command *FinishedCmd) {
+void Scheduler::cleanupFinishedCommands(EventImplPtr FinishedEvent) {
   std::lock_guard<std::mutex> lock(MGraphLock);
-  MGraphBuilder.cleanupFinishedCommands(FinishedCmd);
+  Command *FinishedCmd = static_cast<Command *>(FinishedEvent->getCommand());
+  if (FinishedCmd)
+    MGraphBuilder.cleanupFinishedCommands(FinishedCmd);
 }
 
 void Scheduler::removeMemoryObject(detail::SYCLMemObjI *MemObj) {

sycl/source/detail/scheduler/scheduler.hpp

@@ -79,7 +79,7 @@ class Scheduler {
 
   // Removes finished non-leaf non-alloca commands from the subgraph (assuming
   // that all its commands have been waited for).
-  void cleanupFinishedCommands(Command *FinishedCmd);
+  void cleanupFinishedCommands(EventImplPtr FinishedEvent);
 
   // Creates nodes in the graph, that update Req with the pointer to the host
   // memory which contains the latest data of the memory object. New

sycl/test/scheduler/CommandCleanupThreadSafety.cpp

@@ -0,0 +1,38 @@
+// RUN: %clangxx -fsycl %s -o %t.out -lpthread
+// RUN: %CPU_RUN_PLACEHOLDER %t.out
+#include <CL/sycl.hpp>
+
+#include <cassert>
+#include <cstddef>
+#include <thread>
+#include <vector>
+
+// This test checks that the command graph cleanup works properly when
+// invoked from multiple threads.
+using namespace cl::sycl;
+
+class Foo;
+
+event submitTask(queue &Q, buffer<int, 1> &Buf) {
+  return Q.submit([&](handler &Cgh) {
+    auto Acc = Buf.get_access<access::mode::read_write>(Cgh);
+    Cgh.single_task<Foo>([=]() { Acc[0] = 42; });
+  });
+}
+
+int main() {
+  queue Q;
+  buffer<int, 1> Buf(range<1>(1));
+
+  // Create multiple commands, each one dependent on the previous
+  std::vector<event> Events;
+  for (std::size_t I = 0; I < 16; ++I)
+    Events.push_back(submitTask(Q, Buf));
+
+  // Initiate cleanup from multiple threads
+  std::vector<std::thread> Threads;
+  for (event &E : Events)
+    Threads.emplace_back([&]() { E.wait(); });
+  for (std::thread &T : Threads)
+    T.join();
+}

sycl/test/scheduler/FinishedCmdCleanup.cpp

@@ -1,6 +1,7 @@
 // RUN: %clangxx -fsycl -I %sycl_source_dir %s -o %t.out
 // RUN: %t.out
 #include <CL/sycl.hpp>
+#include <detail/event_impl.hpp>
 #include <detail/scheduler/scheduler.hpp>
 
 #include <algorithm>
@@ -76,7 +77,9 @@ int main() {
   addEdge(InnerA, &LeafA, &AllocaA);
   addEdge(InnerA, InnerB, &AllocaB);
 
-  TS.cleanupFinishedCommands(InnerA);
+  std::shared_ptr<detail::event_impl> Event{new detail::event_impl{}};
+  Event->setCommand(InnerA);
+  TS.cleanupFinishedCommands(Event);
   TS.removeRecordForMemObj(detail::getSyclObjImpl(BufC).get());
 
   assert(NInnerCommandsAlive == 0);