Move implementation details away from handler to handler_impl

Author: lbushi25

sycl/include/sycl/handler.hpp

@@ -1871,16 +1871,7 @@ class __SYCL_EXPORT handler {
     setArgHelper(ArgIndex, std::move(Arg));
   }
 
-  template <typename DataT, typename PropertyListT =
-                                ext::oneapi::experimental::empty_properties_t>
-  void set_arg(
-      int ArgIndex,
-      detail::work_group_memory_impl &Arg) {
-    MWorkGroupMemoryObjects.push_back(
-        std::make_shared<detail::work_group_memory_impl>(Arg));
-    addArg(detail::kernel_param_kind_t::kind_work_group_memory,
-           MWorkGroupMemoryObjects.back().get(), 0, ArgIndex);
-  }
+  void set_arg(int ArgIndex, detail::work_group_memory_impl &Arg);
 
   // set_arg for graph dynamic_parameters
   template <typename T>
@@ -3246,7 +3237,6 @@ class __SYCL_EXPORT handler {
 
   std::vector<detail::LocalAccessorImplPtr> MLocalAccStorage;
   std::vector<std::shared_ptr<detail::stream_impl>> MStreamStorage;
-  std::vector<std::shared_ptr<detail::work_group_memory_impl>> MWorkGroupMemoryObjects;
   detail::string MKernelName;
   /// Storage for a sycl::kernel object.
   std::shared_ptr<detail::kernel_impl> MKernel;

sycl/source/detail/handler_impl.hpp

@@ -13,6 +13,7 @@
 #include <detail/kernel_bundle_impl.hpp>
 #include <memory>
 #include <sycl/ext/oneapi/experimental/graph.hpp>
+#include <sycl/ext/oneapi/experimental/work_group_memory.hpp>
 
 namespace sycl {
 inline namespace _V1 {
@@ -197,6 +198,9 @@ class handler_impl {
 
   /// True if MCodeLoc is sycl entry point code location
   bool MIsTopCodeLoc = true;
+
+  /// List of work group memory objects associated with this handler
+  std::vector<std::shared_ptr<detail::work_group_memory_impl>> MWorkGroupMemoryObjects;
 };
 
 } // namespace detail

sycl/source/handler.cpp

@@ -34,7 +34,6 @@
 #include <sycl/stream.hpp>
 
 #include <sycl/ext/oneapi/bindless_images_memory.hpp>
-#include <sycl/ext/oneapi/experimental/work_group_memory.hpp>
 #include <sycl/ext/oneapi/memcpy2d.hpp>
 
 namespace sycl {
@@ -819,6 +818,13 @@ void handler::processArg(void *Ptr, const detail::kernel_param_kind_t &Kind,
   }
 }
 
+void handler::set_arg(int ArgIndex, detail::work_group_memory_impl &Arg) {
+  impl->MWorkGroupMemoryObjects.push_back(
+      std::make_shared<detail::work_group_memory_impl>(Arg));
+  addArg(detail::kernel_param_kind_t::kind_work_group_memory,
+         impl->MWorkGroupMemoryObjects.back().get(), 0, ArgIndex);
+}
+
 // The argument can take up more space to store additional information about
 // MAccessRange, MMemoryRange, and MOffset added with addArgsForGlobalAccessor.
 // We use the worst-case estimate because the lifetime of the vector is short.