executors: fix parallel_executor caching and fork_join stacksize

- Remove mutable mask_ cache from parallel_policy_executor_base.
  Storing a mutable member in a constexpr execution_policy causes
  UB and compilation failures (specifically, reading the mutable
  member during constant destruction in C++20). The caching has
  been entirely removed to restore correctness.
- Remove stale 'Must not be nostack' doc comments from
  fork_join_executor constructors, as stackless threads are now
  supported (yielding is simply skipped via allow_yielding).

Author: arpittkhandelwal

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

@@ -1250,24 +1250,6 @@ namespace hpx::execution::experimental {
                 HPX_FORWARD(F, f), HPX_FORWARD(Fs, fs)...);
         }
 
-        template <typename Parameters>
-            requires(hpx::traits::is_executor_parameters_v<Parameters>)
-        friend std::size_t tag_invoke(
-            hpx::execution::experimental::processing_units_count_t,
-            Parameters&&, fork_join_executor const& exec,
-            hpx::chrono::steady_duration const& = hpx::chrono::null_duration,
-            std::size_t = 0)
-        {
-            return exec.shared_data_->num_threads_;
-        }
-
-        friend std::size_t tag_invoke(
-            hpx::execution::experimental::get_first_core_t,
-            fork_join_executor const& exec) noexcept
-        {
-            return hpx::threads::find_first(exec.shared_data_->pu_mask_);
-        }
-
     public:
         bool operator==(fork_join_executor const& rhs) const noexcept
         {
@@ -1288,8 +1270,7 @@ namespace hpx::execution::experimental {
         /// \brief Construct a fork_join_executor.
         ///
         /// \param priority  The priority of the worker threads.
-        /// \param stacksize The stacksize of the worker threads. Must not be
-        ///                  nostack.
+        /// \param stacksize The stacksize of the worker threads.
         /// \param sched     The loop schedule of the parallel regions.
         /// \param yield_delay The time after which the executor yields to other
         ///        work if it has not received any new work for execution.
@@ -1301,15 +1282,6 @@ namespace hpx::execution::experimental {
             std::chrono::nanoseconds yield_delay = std::chrono::microseconds(
                 300))
         {
-            if (stacksize == threads::thread_stacksize::nostack)
-            {
-                HPX_THROW_EXCEPTION(hpx::error::bad_parameter,
-                    "fork_join_executor::fork_join_executor",
-                    "The fork_join_executor does not support using "
-                    "thread_stacksize::nostack as the stacksize (stackful "
-                    "threads are required to yield correctly when idle)");
-            }
-
             shared_data_ = std::make_shared<shared_data>(
                 priority, stacksize, sched, yield_delay);
         }
@@ -1318,8 +1290,7 @@ namespace hpx::execution::experimental {
         ///
         /// \param pu_mask   The PU-mask to use for placing the created threads
         /// \param priority  The priority of the worker threads.
-        /// \param stacksize The stacksize of the worker threads. Must not be
-        ///                  nostack.
+        /// \param stacksize The stacksize of the worker threads.
         /// \param sched     The loop schedule of the parallel regions.
         /// \param yield_delay The time after which the executor yields to other
         ///        work if it has not received any new work for execution.
@@ -1331,15 +1302,6 @@ namespace hpx::execution::experimental {
             std::chrono::nanoseconds yield_delay = std::chrono::microseconds(
                 300))
         {
-            if (stacksize == threads::thread_stacksize::nostack)
-            {
-                HPX_THROW_EXCEPTION(hpx::error::bad_parameter,
-                    "fork_join_executor::fork_join_executor",
-                    "The fork_join_executor does not support using "
-                    "thread_stacksize::nostack as the stacksize (stackful "
-                    "threads are required to yield correctly when idle)");
-            }
-
             shared_data_ = std::make_shared<shared_data>(
                 priority, stacksize, sched, yield_delay, pu_mask);
         }

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

@@ -27,7 +27,6 @@
 #include <hpx/modules/topology.hpp>
 
 #include <algorithm>
-#include <atomic>
 #include <cstddef>
 #include <cstdint>
 #include <string>
@@ -162,7 +161,6 @@ namespace hpx::execution {
           , policy_(rhs.policy_)
           , first_core_(rhs.first_core_)
           , num_cores_(rhs.num_cores_)
-          , mask_(nullptr)    // force recomputing cached pu mask
 #if defined(HPX_HAVE_THREAD_DESCRIPTION)
           , annotation_(rhs.annotation_)
 #endif
@@ -179,9 +177,6 @@ namespace hpx::execution {
                 policy_ = rhs.policy_;
                 first_core_ = rhs.first_core_;
                 num_cores_ = rhs.num_cores_;
-                auto* old_mask =
-                    mask_.exchange(nullptr, std::memory_order_relaxed);
-                delete old_mask;    // force recomputing cached pu mask
 
 #if defined(HPX_HAVE_THREAD_DESCRIPTION)
                 annotation_ = rhs.annotation_;
@@ -190,10 +185,7 @@ namespace hpx::execution {
             return *this;
         }
 
-        ~parallel_policy_executor_base()
-        {
-            delete mask_.load(std::memory_order_relaxed);
-        }
+        constexpr ~parallel_policy_executor_base() = default;
 
         // backwards compatibility support, will be removed in the future
         template <typename Parameters>
@@ -348,12 +340,6 @@ namespace hpx::execution {
 
         hpx::threads::mask_type pu_mask() const
         {
-            auto* current_mask = mask_.load(std::memory_order_acquire);
-            if (current_mask != nullptr && hpx::threads::any(*current_mask))
-            {
-                return *current_mask;
-            }
-
             auto const num_threads = get_num_cores();
             auto const available_threads = static_cast<std::uint32_t>(
                 pool()->get_active_os_thread_count());
@@ -380,16 +366,7 @@ namespace hpx::execution {
                 }
             }
 
-            auto* new_mask = new hpx::threads::mask_type(mask);
-            hpx::threads::mask_type* expected = nullptr;
-            if (mask_.compare_exchange_strong(expected, new_mask,
-                    std::memory_order_release, std::memory_order_acquire))
-            {
-                return *new_mask;
-            }
-
-            delete new_mask;
-            return *expected;
+            return mask;
         }
 
         threads::thread_pool_base* pool() const
@@ -403,7 +380,6 @@ namespace hpx::execution {
         Policy policy_;
         std::size_t first_core_ = 0;
         std::size_t num_cores_ = 0;
-        mutable std::atomic<hpx::threads::mask_type*> mask_{nullptr};
 #if defined(HPX_HAVE_THREAD_DESCRIPTION)
         char const* annotation_ = nullptr;
 #endif