[SYCL] Always store last event (for ioq) if scheduler was not bypassed

sycl/source/detail/queue_impl.cpp

@@ -349,7 +349,13 @@ queue_impl::submit_impl(const detail::type_erased_cgfo_ty &CGF,
                          !requiresPostProcess;
 
   if (noLastEventPath) {
-    return finalizeHandlerInOrderNoEventsUnlocked(Handler);
+    std::unique_lock<std::mutex> Lock(MMutex);
+
+    // Check if we are still in no last event mode. There could
+    // have been a concurrent submit.
+    if (MNoLastEventMode.load(std::memory_order_relaxed)) {
+      return finalizeHandlerInOrderNoEventsUnlocked(Handler);
+    }
   }
 
   detail::EventImplPtr EventImpl;

sycl/source/detail/queue_impl.hpp

@@ -750,7 +750,14 @@ class queue_impl : public std::enable_shared_from_this<queue_impl> {
 
     synchronizeWithExternalEvent(Handler);
 
-    return parseEvent(Handler.finalize());
+    auto Event = parseEvent(Handler.finalize());
+
+    if (Event && !Scheduler::CheckEventReadiness(*MContext, Event)) {
+      MDefaultGraphDeps.LastEventPtr = Event;
+      MNoLastEventMode.store(false, std::memory_order_relaxed);
+    }
+
+    return Event;
   }
 
   template <typename HandlerType = handler>

sycl/test-e2e/InorderQueue/in_order_multi_queue_host_task.cpp

@@ -0,0 +1,66 @@
+// RUN: %{build} -o %t.out
+// RUN: %{run} %t.out
+//==-------- in_order_multi_queue_host_task.cpp ---------------------------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+#include <condition_variable>
+#include <iostream>
+
+#include <sycl/detail/core.hpp>
+#include <sycl/properties/all_properties.hpp>
+#include <sycl/usm.hpp>
+
+using namespace sycl;
+
+const int dataSize = 1024;
+
+int main() {
+  queue Queue1{property::queue::in_order()};
+  queue Queue2{property::queue::in_order()};
+
+  int *dataA = malloc_host<int>(dataSize, Queue1);
+  int *dataB = malloc_host<int>(dataSize, Queue1);
+  int *dataC = malloc_host<int>(dataSize, Queue1);
+
+  bool ready = false;
+  std::mutex host_task_mtx;
+  std::condition_variable cv;
+
+  auto Event1 = Queue1.submit([&](handler &cgh) {
+    cgh.host_task([&] {
+      std::unique_lock<std::mutex> lk(host_task_mtx);
+      cv.wait(lk, [&] { return ready; });
+      for (size_t i = 0; i < dataSize; ++i) {
+        dataA[i] = i;
+      }
+    });
+  });
+
+  Queue2.submit([&](handler &cgh) {
+    cgh.depends_on(Event1);
+    cgh.parallel_for(range<1>(dataSize),
+                     [=](id<1> idx) { dataB[idx[0]] = dataA[idx[0]]; });
+  });
+
+  {
+    std::unique_lock<std::mutex> lk(host_task_mtx);
+    ready = true;
+  }
+  cv.notify_one();
+
+  Queue2.wait();
+
+  for (size_t i = 0; i != dataSize; ++i) {
+    if (dataB[i] != i) {
+      std::cout << "Result mismatches " << dataB[i] << " vs expected " << i
+                << " for index " << i << std::endl;
+      return 1;
+    }
+  }
+
+  return 0;
+}