Use thread-safe cache in thread_local_caching_allocator

libs/core/allocator_support/include/hpx/allocator_support/thread_local_caching_allocator.hpp

@@ -1,4 +1,4 @@
-//  Copyright (c) 2023 Hartmut Kaiser
+//  Copyright (c) 2023-2024 Hartmut Kaiser
 //
 //  SPDX-License-Identifier: BSL-1.0
 //  Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -12,7 +12,6 @@
 #include <cstddef>
 #include <memory>
 #include <new>
-#include <stack>
 #include <type_traits>
 #include <utility>
 
@@ -21,8 +20,10 @@ namespace hpx::util {
 #if defined(HPX_ALLOCATOR_SUPPORT_HAVE_CACHING) &&                             \
     !((defined(HPX_HAVE_CUDA) && defined(__CUDACC__)) ||                       \
         defined(HPX_HAVE_HIP))
+
     ///////////////////////////////////////////////////////////////////////////
-    template <typename T = char, typename Allocator = std::allocator<T>>
+    template <template <typename, typename> class Stack, typename T = char,
+        typename Allocator = std::allocator<T>>
     struct thread_local_caching_allocator
     {
         HPX_NO_UNIQUE_ADDRESS Allocator alloc;
@@ -38,7 +39,7 @@ namespace hpx::util {
         template <typename U>
         struct rebind
         {
-            using other = thread_local_caching_allocator<U,
+            using other = thread_local_caching_allocator<Stack, U,
                 typename traits::template rebind_alloc<U>>;
         };
 
@@ -56,6 +57,7 @@ namespace hpx::util {
             explicit allocated_cache(Allocator const& a) noexcept(
                 noexcept(std::is_nothrow_copy_constructible_v<Allocator>))
               : alloc(a)
+              , data(0)
             {
             }
 
@@ -72,19 +74,19 @@ namespace hpx::util {
             pointer allocate(size_type n)
             {
                 pointer p;
-                if (data.empty())
+                std::pair<T*, size_type> pair;
+                if (data.pop(pair))
+                {
+                    p = pair.first;
+                }
+                else
                 {
                     p = traits::allocate(alloc, n);
                     if (p == nullptr)
                     {
                         throw std::bad_alloc();
                     }
                 }
-                else
-                {
-                    p = data.top().first;
-                    data.pop();
-                }
 
                 ++allocated;
                 return p;
@@ -104,16 +106,15 @@ namespace hpx::util {
         private:
             void clear_cache() noexcept
             {
-                while (!data.empty())
+                std::pair<T*, size_type> p;
+                while (data.pop(p))
                 {
-                    traits::deallocate(
-                        alloc, data.top().first, data.top().second);
-                    data.pop();
+                    traits::deallocate(alloc, p.first, p.second);
                 }
             }
 
             HPX_NO_UNIQUE_ADDRESS Allocator alloc;
-            std::stack<std::pair<T*, size_type>> data;
+            Stack<std::pair<T*, size_type>, Allocator> data;
             std::size_t allocated = 0;
             std::size_t deallocated = 0;
         };
@@ -134,7 +135,7 @@ namespace hpx::util {
 
         template <typename U, typename Alloc>
         explicit thread_local_caching_allocator(
-            thread_local_caching_allocator<U, Alloc> const&
+            thread_local_caching_allocator<Stack, U, Alloc> const&
                 rhs) noexcept(noexcept(std::
                 is_nothrow_copy_constructible_v<Alloc>))
           : alloc(rhs.alloc)
@@ -198,7 +199,8 @@ namespace hpx::util {
         }
     };
 #else
-    template <typename T = char, typename Allocator = std::allocator<T>>
+    template <template <typename, typename> class Stack, typename T = char,
+        typename Allocator = std::allocator<T>>
     using thread_local_caching_allocator = Allocator;
 #endif
 }    // namespace hpx::util

libs/core/async_base/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2020-2022 The STE||AR-Group
+# Copyright (c) 2020-2024 The STE||AR-Group
 #
 # SPDX-License-Identifier: BSL-1.0
 # Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -34,6 +34,6 @@ add_hpx_module(
   COMPAT_HEADERS ${async_base_compat_headers}
   SOURCES ${async_base_sources}
   MODULE_DEPENDENCIES hpx_allocator_support hpx_concepts hpx_config
-                      hpx_coroutines hpx_tag_invoke
+                      hpx_concurrency hpx_coroutines hpx_tag_invoke
   CMAKE_SUBDIRS examples tests
 )

libs/core/async_base/include/hpx/async_base/dataflow.hpp

@@ -1,4 +1,4 @@
-//  Copyright (c) 2007-2022 Hartmut Kaiser
+//  Copyright (c) 2007-2024 Hartmut Kaiser
 //
 //  SPDX-License-Identifier: BSL-1.0
 //  Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -29,6 +29,7 @@ namespace hpx {
 #else
 
 #include <hpx/config.hpp>
+#include <hpx/concurrency/stack.hpp>
 #include <hpx/modules/allocator_support.hpp>
 #include <hpx/modules/concepts.hpp>
 #include <hpx/modules/tag_invoke.hpp>
@@ -60,12 +61,14 @@ namespace hpx {
             // clang-format on
             friend constexpr HPX_FORCEINLINE auto tag_fallback_invoke(
                 dataflow_t tag, F&& f, Ts&&... ts)
-                -> decltype(tag(hpx::util::thread_local_caching_allocator<char,
+                -> decltype(tag(hpx::util::thread_local_caching_allocator<
+                                    hpx::lockfree::variable_size_stack, char,
                                     hpx::util::internal_allocator<>>{},
                     HPX_FORWARD(F, f), HPX_FORWARD(Ts, ts)...))
             {
                 using allocator_type =
-                    hpx::util::thread_local_caching_allocator<char,
+                    hpx::util::thread_local_caching_allocator<
+                        hpx::lockfree::variable_size_stack, char,
                         hpx::util::internal_allocator<>>;
                 return hpx::functional::tag_invoke(tag, allocator_type{},
                     HPX_FORWARD(F, f), HPX_FORWARD(Ts, ts)...);

libs/core/async_combinators/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2023 The STE||AR-Group
+# Copyright (c) 2019-2024 The STE||AR-Group
 #
 # SPDX-License-Identifier: BSL-1.0
 # Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -43,6 +43,7 @@ add_hpx_module(
   EXCLUDE_FROM_GLOBAL_HEADER "hpx/async_combinators/future_wait.hpp"
   MODULE_DEPENDENCIES
     hpx_async_base
+    hpx_concurrency
     hpx_config
     hpx_errors
     hpx_futures

libs/core/async_combinators/include/hpx/async_combinators/when_all.hpp

@@ -1,4 +1,4 @@
-//  Copyright (c) 2007-2022 Hartmut Kaiser
+//  Copyright (c) 2007-2024 Hartmut Kaiser
 //  Copyright (c) 2013 Agustin Berge
 //  Copyright (c) 2017 Denis Blank
 //
@@ -131,6 +131,7 @@ namespace hpx {
 #include <hpx/config.hpp>
 #include <hpx/allocator_support/internal_allocator.hpp>
 #include <hpx/allocator_support/thread_local_caching_allocator.hpp>
+#include <hpx/concurrency/stack.hpp>
 #include <hpx/datastructures/tuple.hpp>
 #include <hpx/functional/tag_invoke.hpp>
 #include <hpx/futures/detail/future_data.hpp>
@@ -201,14 +202,16 @@ namespace hpx::lcos::detail {
             return async_visit_future(HPX_FORWARD(T, current));
         }
 
+        // clang-format off
         template <typename T, typename N>
         auto operator()(hpx::util::async_traverse_detach_tag, T&& current,
             N&& next) -> decltype(async_detach_future(HPX_FORWARD(T, current),
-            HPX_FORWARD(N, next)))
+                          HPX_FORWARD(N, next)))
         {
             return async_detach_future(
                 HPX_FORWARD(T, current), HPX_FORWARD(N, next));
         }
+        // clang-format on
 
         template <typename T>
         void operator()(hpx::util::async_traverse_complete_tag, T&& pack)
@@ -226,7 +229,8 @@ namespace hpx::lcos::detail {
         using frame_type = async_when_all_frame<result_type>;
         using no_addref = typename frame_type::base_type::init_no_addref;
 
-        using allocator_type = hpx::util::thread_local_caching_allocator<char,
+        using allocator_type = hpx::util::thread_local_caching_allocator<
+            hpx::lockfree::variable_size_stack, char,
             hpx::util::internal_allocator<>>;
         auto frame = hpx::util::traverse_pack_async_allocator(allocator_type{},
             hpx::util::async_traverse_in_place_tag<frame_type>{}, no_addref{},

libs/core/concurrency/include/hpx/concurrency/stack.hpp

@@ -1,5 +1,5 @@
 //  Copyright (C) 2008-2013 Tim Blechmann
-//  Copyright (c) 2022-2023 Hartmut Kaiser
+//  Copyright (c) 2022-2024 Hartmut Kaiser
 //
 //  SPDX-License-Identifier: BSL-1.0
 //  Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -770,4 +770,10 @@ namespace hpx::lockfree {
 
         pool_t pool;
     };
+
+    template <typename T, typename Allocator = std::allocator<T>>
+    class variable_size_stack : public stack<T, Allocator>
+    {
+        using stack<T, Allocator>::stack;
+    };
 }    // namespace hpx::lockfree

libs/core/concurrency/tests/unit/stack.cpp

@@ -73,7 +73,7 @@ void ranged_unsynchronized_push_test()
     HPX_TEST(!stk.unsynchronized_pop(out));
 }
 
-void fixed_size_stack_test()
+void variable_size_stack_test()
 {
     hpx::lockfree::stack<long, std::allocator<long>, 128> stk;
 
@@ -89,7 +89,7 @@ void fixed_size_stack_test()
     HPX_TEST(stk.empty());
 }
 
-void fixed_size_stack_test_exhausted()
+void variable_size_stack_test_exhausted()
 {
     hpx::lockfree::stack<long, std::allocator<long>, 2> stk;
 
@@ -206,8 +206,8 @@ int main()
     unsafe_stack_test();
     ranged_push_test();
     ranged_unsynchronized_push_test();
-    fixed_size_stack_test();
-    fixed_size_stack_test_exhausted();
+    variable_size_stack_test();
+    variable_size_stack_test_exhausted();
     bounded_stack_test_exhausted();
     stack_consume_one_test();
     stack_consume_all_test();

libs/core/execution/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2022 The STE||AR-Group
+# Copyright (c) 2019-2024 The STE||AR-Group
 #
 # SPDX-License-Identifier: BSL-1.0
 # Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -160,6 +160,7 @@ add_hpx_module(
   MODULE_DEPENDENCIES
     hpx_async_base
     hpx_async_combinators
+    hpx_concurrency
     hpx_config
     hpx_threading
     hpx_pack_traversal

libs/core/execution/include/hpx/execution/detail/future_exec.hpp

@@ -1,4 +1,4 @@
-//  Copyright (c) 2007-2022 Hartmut Kaiser
+//  Copyright (c) 2007-2024 Hartmut Kaiser
 //  Copyright (c) 2013 Agustin Berge
 //
 //  SPDX-License-Identifier: BSL-1.0
@@ -13,6 +13,7 @@
 #include <hpx/assert.hpp>
 #include <hpx/async_base/launch_policy.hpp>
 #include <hpx/async_base/traits/is_launch_policy.hpp>
+#include <hpx/concurrency/stack.hpp>
 #include <hpx/coroutines/thread_enums.hpp>
 #include <hpx/execution/traits/executor_traits.hpp>
 #include <hpx/execution/traits/future_then_result_exec.hpp>
@@ -64,9 +65,9 @@ namespace hpx::lcos::detail {
             using continuation_result_type =
                 hpx::util::invoke_result_t<F, Future>;
 
-            using allocator_type =
-                hpx::util::thread_local_caching_allocator<char,
-                    hpx::util::internal_allocator<>>;
+            using allocator_type = hpx::util::thread_local_caching_allocator<
+                hpx::lockfree::variable_size_stack, char,
+                hpx::util::internal_allocator<>>;
 
             hpx::traits::detail::shared_state_ptr_t<result_type> p =
                 detail::make_continuation_alloc<continuation_result_type>(

libs/core/executors/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2020-2022 The STE||AR-Group
+# Copyright (c) 2020-2024 The STE||AR-Group
 #
 # SPDX-License-Identifier: BSL-1.0
 # Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -97,6 +97,7 @@ add_hpx_module(
     hpx_allocator_support
     hpx_async_base
     hpx_concepts
+    hpx_concurrency
     hpx_config
     hpx_errors
     hpx_execution

libs/core/executors/include/hpx/executors/parallel_executor.hpp

@@ -1,5 +1,5 @@
 //  Copyright (c) 2019-2020 ETH Zurich
-//  Copyright (c) 2007-2023 Hartmut Kaiser
+//  Copyright (c) 2007-2024 Hartmut Kaiser
 //  Copyright (c) 2019 Agustin Berge
 //
 //  SPDX-License-Identifier: BSL-1.0
@@ -12,6 +12,7 @@
 #include <hpx/allocator_support/internal_allocator.hpp>
 #include <hpx/allocator_support/thread_local_caching_allocator.hpp>
 #include <hpx/async_base/launch_policy.hpp>
+#include <hpx/concurrency/stack.hpp>
 #include <hpx/execution/detail/async_launch_policy_dispatch.hpp>
 #include <hpx/execution/detail/future_exec.hpp>
 #include <hpx/execution/detail/post_policy_dispatch.hpp>
@@ -389,9 +390,9 @@ namespace hpx::execution {
                 hpx::bind_back(HPX_FORWARD(F, f), HPX_FORWARD(Ts, ts)...));
 #endif
 
-            using allocator_type =
-                hpx::util::thread_local_caching_allocator<char,
-                    hpx::util::internal_allocator<>>;
+            using allocator_type = hpx::util::thread_local_caching_allocator<
+                hpx::lockfree::variable_size_stack, char,
+                hpx::util::internal_allocator<>>;
             hpx::traits::detail::shared_state_ptr_t<result_type> p =
                 lcos::detail::make_continuation_alloc_nounwrap<result_type>(
                     allocator_type{}, HPX_FORWARD(Future, predecessor),

libs/core/futures/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2022 The STE||AR-Group
+# Copyright (c) 2019-2024 The STE||AR-Group
 #
 # SPDX-License-Identifier: BSL-1.0
 # Distributed under the Boost Software License, Version 1.0. (See accompanying
@@ -68,6 +68,7 @@ add_hpx_module(
     hpx_assertion
     hpx_async_base
     hpx_concepts
+    hpx_concurrency
     hpx_config
     hpx_errors
     hpx_functional