Eliminate std::shared_ptr<detail::queue_impl> from sycl::handler.

Author: Alexandr-Konovalov

sycl/include/sycl/handler.hpp

@@ -430,7 +430,7 @@ class __SYCL_EXPORT handler {
   /// \param Queue is a SYCL queue.
   /// \param CallerNeedsEvent indicates if the event resulting from this handler
   ///        is needed by the caller.
-  handler(std::shared_ptr<detail::queue_impl> Queue, bool CallerNeedsEvent);
+  handler(std::shared_ptr<detail::queue_impl> &Queue, bool CallerNeedsEvent);
 
   /// Constructs SYCL handler from the associated queue and the submission's
   /// primary and secondary queue.
@@ -443,7 +443,7 @@ class __SYCL_EXPORT handler {
   /// \param CallerNeedsEvent indicates if the event resulting from this handler
   ///        is needed by the caller.
   handler(detail::handler_impl *HandlerImpl,
-          std::shared_ptr<detail::queue_impl> Queue);
+          std::shared_ptr<detail::queue_impl> &Queue);
 
   /// Constructs SYCL handler from Graph.
   ///
@@ -3425,7 +3425,7 @@ class __SYCL_EXPORT handler {
 private:
   std::unique_ptr<detail::handler_impl> MImplOwner;
   detail::handler_impl *impl;
-  std::shared_ptr<detail::queue_impl> MQueue;
+  std::shared_ptr<detail::queue_impl> &MQueue;
   std::vector<detail::LocalAccessorImplPtr> MLocalAccStorage;
   std::vector<std::shared_ptr<detail::stream_impl>> MStreamStorage;
   detail::string MKernelName;

sycl/source/detail/handler_impl.hpp

@@ -31,11 +31,11 @@ enum class HandlerSubmissionState : std::uint8_t {
 
 class handler_impl {
 public:
-  handler_impl(std::shared_ptr<queue_impl> SubmissionPrimaryQueue,
-               std::shared_ptr<queue_impl> SubmissionSecondaryQueue,
+  handler_impl(queue_impl *SubmissionPrimaryQueue,
+               queue_impl *SubmissionSecondaryQueue,
                bool EventNeeded)
-      : MSubmissionPrimaryQueue(std::move(SubmissionPrimaryQueue)),
-        MSubmissionSecondaryQueue(std::move(SubmissionSecondaryQueue)),
+      : MSubmissionPrimaryQueue(SubmissionPrimaryQueue),
+        MSubmissionSecondaryQueue(SubmissionSecondaryQueue),
         MEventNeeded(EventNeeded) {};
 
   handler_impl(
@@ -73,13 +73,13 @@ class handler_impl {
   /// Shared pointer to the primary queue implementation. This is different from
   /// the queue associated with the handler if the corresponding submission is
   /// a fallback from a previous submission.
-  std::shared_ptr<queue_impl> MSubmissionPrimaryQueue;
+  queue_impl *MSubmissionPrimaryQueue;
 
   /// Shared pointer to the secondary queue implementation. Nullptr if no
   /// secondary queue fallback was given in the associated submission. This is
   /// equal to the queue associated with the handler if the corresponding
   /// submission is a fallback from a previous submission.
-  std::shared_ptr<queue_impl> MSubmissionSecondaryQueue;
+  queue_impl *MSubmissionSecondaryQueue;
 
   /// Bool stores information about whether the event resulting from the
   /// corresponding work is required.

sycl/source/detail/queue_impl.cpp

@@ -364,8 +364,8 @@ event queue_impl::submit_impl(const detail::type_erased_cgfo_ty &CGF,
                               bool IsTopCodeLoc,
                               const SubmissionInfo &SubmitInfo) {
   detail::handler_impl HandlerImpl(
-      std::move(PrimaryQueue), std::move(SecondaryQueue), CallerNeedsEvent);
-  handler Handler(&HandlerImpl, Self);
+      PrimaryQueue.get(), SecondaryQueue.get(), CallerNeedsEvent);
+  handler Handler(&HandlerImpl, const_cast<std::shared_ptr<queue_impl> &>(Self));
   Handler.saveCodeLoc(Loc, IsTopCodeLoc);
 
   {
@@ -387,7 +387,7 @@ event queue_impl::submit_impl(const detail::type_erased_cgfo_ty &CGF,
 
   addEvent(Event);
 
-  auto EventImpl = detail::getSyclObjImpl(Event);
+  auto &EventImpl = detail::getSyclObjImpl(Event);
   for (auto &Stream : Streams) {
     // We don't want stream flushing to be blocking operation that is why submit
     // a host task to print stream buffer. It will fire up as soon as the kernel

sycl/source/handler.cpp

@@ -302,17 +302,19 @@ fill_copy_args(detail::handler_impl *impl,
 
 } // namespace detail
 
-handler::handler(std::shared_ptr<detail::queue_impl> Queue,
+handler::handler(std::shared_ptr<detail::queue_impl> &Queue,
                  bool CallerNeedsEvent)
-    : MImplOwner(std::make_unique<detail::handler_impl>(Queue, nullptr,
+    : MImplOwner(std::make_unique<detail::handler_impl>(Queue.get(), nullptr,
                                                         CallerNeedsEvent)),
-      impl(MImplOwner.get()), MQueue(std::move(Queue)) {}
+      impl(MImplOwner.get()), MQueue(Queue) {}
 
 handler::handler(detail::handler_impl *HandlerImpl,
-                 std::shared_ptr<detail::queue_impl> Queue)
-    : impl(HandlerImpl), MQueue(std::move(Queue)) {}
+                 std::shared_ptr<detail::queue_impl> &Queue)
+    : impl(HandlerImpl), MQueue(Queue) {}
 
-handler::handler(detail::handler_impl *HandlerImpl) : impl(HandlerImpl) {}
+static std::shared_ptr<detail::queue_impl> DummyQueue;
+
+handler::handler(detail::handler_impl *HandlerImpl) : impl(HandlerImpl), MQueue(DummyQueue) {}
 
 // Sets the submission state to indicate that an explicit kernel bundle has been
 // set. Throws a sycl::exception with errc::invalid if the current state
@@ -1660,21 +1662,17 @@ void handler::ext_oneapi_signal_external_semaphore(
 
 void handler::use_kernel_bundle(
     const kernel_bundle<bundle_state::executable> &ExecBundle) {
-  std::shared_ptr<detail::queue_impl> PrimaryQueue =
-      impl->MSubmissionPrimaryQueue;
   if ((!impl->MGraph &&
-       (PrimaryQueue->get_context() != ExecBundle.get_context())) ||
+       (impl->MSubmissionPrimaryQueue->get_context() != ExecBundle.get_context())) ||
       (impl->MGraph &&
        (impl->MGraph->getContext() != ExecBundle.get_context())))
     throw sycl::exception(
         make_error_code(errc::invalid),
         "Context associated with the primary queue is different from the "
         "context associated with the kernel bundle");
 
-  std::shared_ptr<detail::queue_impl> SecondaryQueue =
-      impl->MSubmissionSecondaryQueue;
-  if (SecondaryQueue &&
-      SecondaryQueue->get_context() != ExecBundle.get_context())
+  if (impl->MSubmissionSecondaryQueue &&
+      impl->MSubmissionSecondaryQueue->get_context() != ExecBundle.get_context())
     throw sycl::exception(
         make_error_code(errc::invalid),
         "Context associated with the secondary queue is different from the "
@@ -1816,7 +1814,7 @@ void handler::verifyDeviceHasProgressGuarantee(
 }
 
 bool handler::supportsUSMMemcpy2D() {
-  for (const std::shared_ptr<detail::queue_impl> &QueueImpl :
+  for (detail::queue_impl *QueueImpl :
        {impl->MSubmissionPrimaryQueue, impl->MSubmissionSecondaryQueue}) {
     if (QueueImpl &&
         !checkContextSupports(QueueImpl->getContextImplPtr(),
@@ -1827,7 +1825,7 @@ bool handler::supportsUSMMemcpy2D() {
 }
 
 bool handler::supportsUSMFill2D() {
-  for (const std::shared_ptr<detail::queue_impl> &QueueImpl :
+  for (detail::queue_impl *QueueImpl :
        {impl->MSubmissionPrimaryQueue, impl->MSubmissionSecondaryQueue}) {
     if (QueueImpl && !checkContextSupports(QueueImpl->getContextImplPtr(),
                                            UR_CONTEXT_INFO_USM_FILL2D_SUPPORT))
@@ -1837,7 +1835,7 @@ bool handler::supportsUSMFill2D() {
 }
 
 bool handler::supportsUSMMemset2D() {
-  for (const std::shared_ptr<detail::queue_impl> &QueueImpl :
+  for (detail::queue_impl *QueueImpl :
        {impl->MSubmissionPrimaryQueue, impl->MSubmissionSecondaryQueue}) {
     if (QueueImpl && !checkContextSupports(QueueImpl->getContextImplPtr(),
                                            UR_CONTEXT_INFO_USM_FILL2D_SUPPORT))