[SYCL] Fix abi issues caused by reduction implementation

The fix also removes the field handler::MReductionsStorage and re-uses
the existing MSharedPtrStorage to keep reductions buffers alive
until the execution on device/host code using those buffers finishes.

Signed-off-by: Vyacheslav N Klochkov <[email protected]>

Author: v-klochkov

sycl/include/CL/sycl/detail/cg.hpp

@@ -456,7 +456,6 @@ class CGExecKernel : public CG {
   string_class MKernelName;
   detail::OSModuleHandle MOSModuleHandle;
   vector_class<shared_ptr_class<detail::stream_impl>> MStreams;
-  vector_class<shared_ptr_class<void>> MReductions;
 
   CGExecKernel(NDRDescT NDRDesc, unique_ptr_class<HostKernelBase> HKernel,
                shared_ptr_class<detail::kernel_impl> SyclKernel,
@@ -468,15 +467,14 @@ class CGExecKernel : public CG {
                vector_class<ArgDesc> Args, string_class KernelName,
                detail::OSModuleHandle OSModuleHandle,
                vector_class<shared_ptr_class<detail::stream_impl>> Streams,
-               vector_class<shared_ptr_class<void>> Reductions, CGTYPE Type,
-               detail::code_location loc = {})
+               CGTYPE Type, detail::code_location loc = {})
       : CG(Type, std::move(ArgsStorage), std::move(AccStorage),
            std::move(SharedPtrStorage), std::move(Requirements),
            std::move(Events), std::move(loc)),
         MNDRDesc(std::move(NDRDesc)), MHostKernel(std::move(HKernel)),
         MSyclKernel(std::move(SyclKernel)), MArgs(std::move(Args)),
         MKernelName(std::move(KernelName)), MOSModuleHandle(OSModuleHandle),
-        MStreams(std::move(Streams)), MReductions(std::move(Reductions)) {
+        MStreams(std::move(Streams)) {
     assert((getType() == RUN_ON_HOST_INTEL || getType() == KERNEL) &&
            "Wrong type of exec kernel CG.");
   }

sycl/include/CL/sycl/handler.hpp

@@ -255,14 +255,14 @@ class __SYCL_EXPORT handler {
     MStreamStorage.push_back(std::move(Stream));
   }
 
-  /// Saves buffers and scalars associated with reduction to handler.
-  /// They are then forwarded to command group later and destroyed
-  /// only after the command group finishes the work on device/host.
+  /// Saves buffers created by handling reduction feature in handler.
+  /// They are then forwarded to command group and destroyed only after
+  /// the command group finishes the work on device/host.
+  /// The 'MSharedPtrStorage' suits that need.
   ///
-  /// @param ReduObj is a pointer to object that must be preserved
-  /// for reduction until the .
-  void addReduction(shared_ptr_class<void> ReduObj) {
-    MReductionStorage.push_back(std::move(ReduObj));
+  /// @param ReduObj is a pointer to object that must be stored.
+  void addReduction(shared_ptr_class<const void> ReduObj) {
+    MSharedPtrStorage.push_back(std::move(ReduObj));
   }
 
   ~handler() = default;
@@ -1714,7 +1714,6 @@ class __SYCL_EXPORT handler {
   vector_class<detail::AccessorImplPtr> MAccStorage;
   vector_class<detail::LocalAccessorImplPtr> MLocalAccStorage;
   vector_class<shared_ptr_class<detail::stream_impl>> MStreamStorage;
-  vector_class<shared_ptr_class<void>> MReductionStorage;
   vector_class<shared_ptr_class<const void>> MSharedPtrStorage;
   /// The list of arguments for the kernel.
   vector_class<detail::ArgDesc> MArgs;

sycl/source/handler.cpp

@@ -40,8 +40,8 @@ event handler::finalize() {
         std::move(MArgsStorage), std::move(MAccStorage),
         std::move(MSharedPtrStorage), std::move(MRequirements),
         std::move(MEvents), std::move(MArgs), std::move(MKernelName),
-        std::move(MOSModuleHandle), std::move(MStreamStorage),
-        std::move(MReductionStorage), MCGType, MCodeLoc));
+        std::move(MOSModuleHandle), std::move(MStreamStorage), MCGType,
+        MCodeLoc));
     break;
   }
   case detail::CG::INTEROP_TASK_CODEPLAY:

sycl/test/abi/sycl_symbols_linux.dump

@@ -3232,7 +3232,8 @@ _ZN2cl4sycl7handler10processArgEPvRKNS0_6detail19kernel_param_kind_tEimRmb
 _ZN2cl4sycl7handler13getKernelNameB5cxx11Ev
 _ZN2cl4sycl7handler18extractArgsAndReqsEv
 _ZN2cl4sycl7handler28extractArgsAndReqsFromLambdaEPcmPKNS0_6detail19kernel_param_desc_tE
-_ZN2cl4sycl7handler8finalizeERKNS0_6detail13code_locationE
+_ZN2cl4sycl7handler15addEventToQueueESt10shared_ptrINS0_6detail10queue_implEENS0_5eventE
+_ZN2cl4sycl7handler8finalizeEv
 _ZN2cl4sycl7program17build_with_sourceENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEES7_
 _ZN2cl4sycl7program19compile_with_sourceENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEES7_
 _ZN2cl4sycl7program22build_with_kernel_nameENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEES7_l

sycl/test/abi/symbol_size.cpp

@@ -43,7 +43,7 @@ int main() {
   check_size<device_selector, 8>();
   check_size<event, 16>();
   check_size<gpu_selector, 8>();
-  check_size<handler, 472>();
+  check_size<handler, 528>();
   check_size<image<1>, 16>();
   check_size<kernel, 16>();
   check_size<platform, 16>();