comments

Author: huixie90

libcxx/docs/ReleaseNotes/22.rst

@@ -82,6 +82,8 @@ Improvements and New Features
 - The ``std::{generate, generate_n}`` and ``std::ranges::generate_n`` algorithms have been optimized for segmented
   iterators, resulting in a performance improvement for ``std::deque<short>`` and
   ``std::join_view<vector<vector<short>>>`` iterators.
+- ``std::atomic::wait`` has been refactored to accept more types to use platform native wait functions directly.
+  This is guarded behind the ABI Macro ``_LIBCPP_ABI_ATOMIC_WAIT_NATIVE_BY_SIZE``.
 
 Deprecations and Removals
 -------------------------

libcxx/include/__atomic/atomic_sync.h

@@ -21,6 +21,7 @@
 #include <__type_traits/has_unique_object_representation.h>
 #include <__type_traits/invoke.h>
 #include <__type_traits/is_same.h>
+#include <__type_traits/is_trivially_copyable.h>
 #include <__type_traits/void_t.h>
 #include <__utility/declval.h>
 #include <cstring>
@@ -139,7 +140,8 @@ _LIBCPP_HIDE_FROM_ABI constexpr bool __has_native_atomic_wait_impl(size_t __size
 
 template <class _Tp>
 concept __has_native_atomic_wait =
-    has_unique_object_representations_v<_Tp> && __has_native_atomic_wait_impl(sizeof(_Tp));
+    has_unique_object_representations_v<_Tp> && is_trivially_copyable_v<_Tp> &&
+    __has_native_atomic_wait_impl(sizeof(_Tp));
 
 #    else // _LIBCPP_ABI_ATOMIC_WAIT_NATIVE_BY_SIZE
 

libcxx/lib/abi/CHANGELOG.TXT

@@ -16,17 +16,6 @@ New entries should be added directly below the "Version" header.
 Version 22.0
 ------------
 
-* [libc++] Remove __time_get_storage::{__analyze,init} from the ABI
-
-  These functions have never been used outside the dylib, so there is no point in exporting them.
-
-  All platforms
-  -------------
-  Symbol removed: _ZNSt3__118__time_get_storageIcE4initERKNS_5ctypeIcEE
-  Symbol removed: _ZNSt3__118__time_get_storageIcE9__analyzeEcRKNS_5ctypeIcEE
-  Symbol removed: _ZNSt3__118__time_get_storageIwE4initERKNS_5ctypeIwEE
-  Symbol removed: _ZNSt3__118__time_get_storageIwE9__analyzeEcRKNS_5ctypeIwEE
-
 * [libc++] Allows any types of size 4 and 8 to use native platform ulock_wait
 
   This patch added symbols for platform wait functions with the size of the type 
@@ -41,6 +30,17 @@ Version 22.0
   Symbol added: __ZNSt3__126__atomic_notify_all_nativeILm8EEEvPKv
   Symbol added: __ZNSt3__132__atomic_notify_one_global_tableEPKv
 
+* [libc++] Remove __time_get_storage::{__analyze,init} from the ABI
+
+  These functions have never been used outside the dylib, so there is no point in exporting them.
+
+  All platforms
+  -------------
+  Symbol removed: _ZNSt3__118__time_get_storageIcE4initERKNS_5ctypeIcEE
+  Symbol removed: _ZNSt3__118__time_get_storageIcE9__analyzeEcRKNS_5ctypeIcEE
+  Symbol removed: _ZNSt3__118__time_get_storageIwE4initERKNS_5ctypeIwEE
+  Symbol removed: _ZNSt3__118__time_get_storageIwE9__analyzeEcRKNS_5ctypeIwEE
+
 ------------
 Version 21.0
 ------------

libcxx/src/atomic.cpp

@@ -240,10 +240,10 @@ __contention_wait(__cxx_atomic_contention_t* __waiter_count, void const* __addre
   __cxx_atomic_fetch_sub(__waiter_count, __cxx_contention_t(1), memory_order_release);
 }
 
-#if defined(__GCC_DESTRUCTIVE_SIZE) && defined(__GCC_CONSTRUCTIVE_SIZE)
-static constexpr size_t __cache_line_size = std::hardware_constructive_interference_size;
-#elif defined(__APPLE__) && defined(__aarch64__)
+#if defined(__APPLE__) && defined(__aarch64__)
 static constexpr size_t __cache_line_size = 128;
+#elif defined(__cpp_lib_hardware_interference_size)
+static constexpr size_t __cache_line_size = std::hardware_constructive_interference_size;
 #else
 static constexpr size_t __cache_line_size = 64;
 #endif

libcxx/test/std/atomics/atomics.types.operations/atomics.types.operations.wait/lost_wakeup.pass.cpp

@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+// UNSUPPORTED: no-threads
+// UNSUPPORTED: c++03, c++11, c++14, c++17
+
+// This is a stress test for std::atomic::wait for lost wake ups.
+
+// <atomic>
+
+#include <atomic>
+#include <functional>
+#include <thread>
+#include <vector>
+
+#include "make_test_thread.h"
+
+constexpr int num_waiters    = 8;
+constexpr int num_iterations = 10'000;
+
+void wait(std::atomic<int>& waiter_ready, const std::atomic<int>& state) {
+  for (int i = 0; i < num_iterations; ++i) {
+    auto old_state = state.load(std::memory_order_acquire);
+    waiter_ready.fetch_add(1, std::memory_order_acq_rel);
+    state.wait(old_state, std::memory_order_acquire);
+  }
+}
+
+void notify(std::atomic<int>& waiter_ready, std::atomic<int>& state) {
+  for (int i = 0; i < num_iterations; ++i) {
+    while (waiter_ready.load(std::memory_order_acquire) < num_waiters) {
+      std::this_thread::yield();
+    }
+    waiter_ready.store(0, std::memory_order_release);
+    state.fetch_add(1, std::memory_order_acq_rel);
+    state.notify_all();
+  }
+}
+
+int main(int, char**) {
+  for (int run = 0; run < 20; ++run) {
+    std::atomic<int> waiter_ready(0);
+    std::atomic<int> state(0);
+    std::vector<std::jthread> threads;
+    for (int i = 0; i < 8; ++i)
+      threads.push_back(support::make_test_jthread(wait, std::ref(waiter_ready), std::cref(state)));
+
+    threads.push_back(support::make_test_jthread(notify, std::ref(waiter_ready), std::ref(state)));
+  }
+
+  return 0;
+}