libcxx/src/atomic.cpp

//===----------------------------------------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//

#include <__thread/timed_backoff_policy.h>
#include <atomic>
#include <climits>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <functional>
#include <new>
#include <thread>
#include <type_traits>

// Determine whether to use the public os_sync API on Apple platforms based on the deployment target.
// Otherwise we fall back to the private __ulock API.
#if defined(__APPLE__)
#  if (defined(__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) &&                                                       \
       __ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__ >= 140400) ||                                                     \
      (defined(__ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__) &&                                                      \
       __ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__ >= 170400) ||                                                    \
      (defined(__ENVIRONMENT_TV_OS_VERSION_MIN_REQUIRED__) && __ENVIRONMENT_TV_OS_VERSION_MIN_REQUIRED__ >= 170400) || \
      (defined(__ENVIRONMENT_WATCH_OS_VERSION_MIN_REQUIRED__) &&                                                       \
       __ENVIRONMENT_WATCH_OS_VERSION_MIN_REQUIRED__ >= 100400)
#    define _LIBCPP_USE_OS_SYNC
#  endif
#endif

#ifdef __linux__

#  include <linux/futex.h>
#  include <sys/syscall.h>
#  include <unistd.h>

// libc++ uses SYS_futex as a universal syscall name. However, on 32 bit architectures
// with a 64 bit time_t, we need to specify SYS_futex_time64.
#  if !defined(SYS_futex) && defined(SYS_futex_time64)
#    define SYS_futex SYS_futex_time64
#  endif
#  define _LIBCPP_FUTEX(...) syscall(SYS_futex, __VA_ARGS__)

#elif defined(__FreeBSD__)

#  include <sys/types.h>
#  include <sys/umtx.h>

#  define _LIBCPP_FUTEX(...) syscall(SYS_futex, __VA_ARGS__)

#elif defined(__OpenBSD__)

#  include <sys/futex.h>

// OpenBSD has no indirect syscalls
#  define _LIBCPP_FUTEX(...) futex(__VA_ARGS__)

#elif defined(_WIN32)

#  include <memory>
#  include <windows.h>

#elif defined(__APPLE__)

#  if defined(_LIBCPP_USE_OS_SYNC)
#    include <os/os_sync_wait_on_address.h>
#  endif

#elif defined(__Fuchsia__)

#  include <zircon/syscalls.h>

#else // <- Add other operating systems here

// Baseline needs no new headers

#  define _LIBCPP_FUTEX(...) syscall(SYS_futex, __VA_ARGS__)

#endif

_LIBCPP_PUSH_MACROS
#include <__undef_macros>

_LIBCPP_BEGIN_NAMESPACE_STD
_LIBCPP_BEGIN_EXPLICIT_ABI_ANNOTATIONS

struct NoTimeout {};

#ifdef __linux__

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  static_assert(_Size == 4, "Can only wait on 4 bytes value");
  alignas(__cxx_contention_t) char buffer[_Size];
  std::memcpy(&buffer, const_cast<const void*>(__val), _Size);
  static constexpr timespec __default_timeout = {2, 0};
  timespec __timeout;
  if constexpr (is_same_v<MaybeTimeout, NoTimeout>) {
    __timeout = __default_timeout;
  } else {
    __timeout.tv_sec  = maybe_timeout_ns / 1'000'000'000;
    __timeout.tv_nsec = maybe_timeout_ns % 1'000'000'000;
  }
  _LIBCPP_FUTEX(__ptr, FUTEX_WAIT_PRIVATE, *reinterpret_cast<__cxx_contention_t const*>(&buffer), &__timeout, 0, 0);
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const* __ptr, bool __notify_one) {
  static_assert(_Size == 4, "Can only wake up on 4 bytes value");
  _LIBCPP_FUTEX(__ptr, FUTEX_WAKE_PRIVATE, __notify_one ? 1 : INT_MAX, 0, 0, 0);
}

#elif defined(__APPLE__)

#  if defined(_LIBCPP_USE_OS_SYNC)

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  std::uint64_t __value = [__val]() -> std::uint64_t {
    if constexpr (_Size == 4) {
      std::uint32_t __result;
      std::memcpy(&__result, __val, _Size);
      return __result;
    } else if constexpr (_Size == 8) {
      std::uint64_t __result;
      std::memcpy(&__result, __val, _Size);
      return __result;
    } else {
      static_assert(false, "Can only wait on 8 bytes or 4 bytes value");
    }
  }();

  if constexpr (is_same_v<MaybeTimeout, NoTimeout>) {
    os_sync_wait_on_address(const_cast<void*>(__ptr), __value, _Size, OS_SYNC_WAIT_ON_ADDRESS_NONE);
  } else {
    // os_sync_wait_on_address_with_timeout returns EINVAL on a 0-ns timeout, so clamp to 1ns
    // when an already-expired deadline produced a zero-duration timeout.
    uint64_t __timeout_ns = std::max<uint64_t>(maybe_timeout_ns, 1);
    os_sync_wait_on_address_with_timeout(
        const_cast<void*>(__ptr),
        __value,
        _Size,
        OS_SYNC_WAIT_ON_ADDRESS_NONE,
        OS_CLOCK_MACH_ABSOLUTE_TIME,
        __timeout_ns);
  }
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const* __ptr, bool __notify_one) {
  static_assert(_Size == 8 || _Size == 4, "Can only wake up on 8 bytes or 4 bytes value");
  if (__notify_one)
    os_sync_wake_by_address_any(const_cast<void*>(__ptr), _Size, OS_SYNC_WAKE_BY_ADDRESS_NONE);
  else
    os_sync_wake_by_address_all(const_cast<void*>(__ptr), _Size, OS_SYNC_WAKE_BY_ADDRESS_NONE);
}

#  else // !_LIBCPP_USE_OS_SYNC -- fall back to the private __ulock API

extern "C" int __ulock_wait(
    uint32_t operation, void* addr, uint64_t value, uint32_t timeout); /* timeout is specified in microseconds */
extern "C" int __ulock_wake(uint32_t operation, void* addr, uint64_t wake_value);

// https://github.com/apple/darwin-xnu/blob/2ff845c2e033bd0ff64b5b6aa6063a1f8f65aa32/bsd/sys/ulock.h#L82
#    define UL_COMPARE_AND_WAIT 1
#    define UL_COMPARE_AND_WAIT64 5
#    define ULF_WAKE_ALL 0x00000100

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  auto __timeout_us = [&] {
    if constexpr (is_same_v<MaybeTimeout, NoTimeout>) {
      return uint32_t(0);
    } else {
      return std::max(static_cast<uint32_t>(maybe_timeout_ns / 1000), uint32_t(1));
    }
  }();

  std::uint64_t __value = [__val]() -> std::uint64_t {
    if constexpr (_Size == 4) {
      std::uint32_t __result;
      std::memcpy(&__result, __val, _Size);
      return __result;
    } else if constexpr (_Size == 8) {
      std::uint64_t __result;
      std::memcpy(&__result, __val, _Size);
      return __result;
    } else {
      static_assert(false, "Can only wait on 8 bytes or 4 bytes value");
    }
  }();

  __ulock_wait(
      _Size == 4 ? UL_COMPARE_AND_WAIT : UL_COMPARE_AND_WAIT64, const_cast<void*>(__ptr), __value, __timeout_us);
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const* __ptr, bool __notify_one) {
  static_assert(_Size == 8 || _Size == 4, "Can only wake up on 8 bytes or 4 bytes value");

  if constexpr (_Size == 4)
    __ulock_wake(UL_COMPARE_AND_WAIT | (__notify_one ? 0 : ULF_WAKE_ALL), const_cast<void*>(__ptr), 0);
  else
    __ulock_wake(UL_COMPARE_AND_WAIT64 | (__notify_one ? 0 : ULF_WAKE_ALL), const_cast<void*>(__ptr), 0);
}

#  endif // _LIBCPP_USE_OS_SYNC

#elif defined(__FreeBSD__) && __SIZEOF_LONG__ == 8
/*
 * Since __cxx_contention_t is int64_t even on 32bit FreeBSD
 * platforms, we have to use umtx ops that work on the long type, and
 * limit its use to architectures where long and int64_t are synonyms.
 */

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  static_assert(_Size == 8, "Can only wait on 8 bytes value");
  // At the moment, FreeBSD is stuck on stable ABI, which only supports platform wait with __cxx_contention_t
  // It is safe to reinterpret_cast the val as it is ever going to be passed a __cxx_contention_t under this ABI
  // If in the future FreeBSD decides to experiment unstable ABI to support more types, this cast will no longer be
  // safe.
  __cxx_contention_t value = *reinterpret_cast<const __cxx_contention_t*>(__val);
  if constexpr (is_same_v<MaybeTimeout, NoTimeout>) {
    _umtx_op(const_cast<void*>(__ptr), UMTX_OP_WAIT, value, nullptr, nullptr);
  } else {
    timespec timeout{};
    timeout.tv_sec  = maybe_timeout_ns / 1'000'000'000;
    timeout.tv_nsec = maybe_timeout_ns % 1'000'000'000;

    _umtx_op(const_cast<void*>(__ptr),
             UMTX_OP_WAIT,
             value,
             reinterpret_cast<void*>(static_cast<uintptr_t>(sizeof(timeout))),
             &timeout);
  }
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const* __ptr, bool __notify_one) {
  static_assert(_Size == 8, "Can only wake up on 8 bytes value");
  _umtx_op(const_cast<void*>(__ptr), UMTX_OP_WAKE, __notify_one ? 1 : INT_MAX, nullptr, nullptr);
}

#elif defined(_WIN32)

static void* win32_get_synch_api_function(const char* function_name) {
  // Attempt to load the API set. Note that as per the Microsoft STL implementation, we assume this API is already
  // loaded and accessible. While this isn't explicitly guaranteed by publicly available Win32 API documentation, it is
  // true in practice, and may be guaranteed by internal documentation not released publicly. In any case the fact that
  // the Microsoft STL made this assumption is reasonable basis to say that we can too. The alternative to this would be
  // to use LoadLibrary, but then leak the module handle. We can't call FreeLibrary, as this would have to be triggered
  // by a global static destructor, which would hang off DllMain, and calling FreeLibrary from DllMain is explicitly
  // mentioned as not being allowed:
  // https://learn.microsoft.com/en-us/windows/win32/dlls/dllmain
  // Given the range of bad options here, we have chosen to mirror what Microsoft did, as it seems fair to assume that
  // Microsoft will guarantee compatibility for us, as we are exposed to the same conditions as all existing Windows
  // apps using the Microsoft STL VS2015/2017/2019/2022 runtimes, where Windows 7 support has not been excluded at
  // compile time.
  static auto module_handle = GetModuleHandleW(L"api-ms-win-core-synch-l1-2-0.dll");
  if (module_handle == nullptr) {
    return nullptr;
  }

  // Attempt to locate the function in the API and return the result to the caller. Note that the NULL return from this
  // method is documented as being interchangeable with nullptr.
  // https://devblogs.microsoft.com/oldnewthing/20180307-00/?p=98175
  return reinterpret_cast<void*>(GetProcAddress(module_handle, function_name));
}

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  static_assert(_Size == 8, "Can only wait on 8 bytes value");
  // WaitOnAddress was added in Windows 8 (build 9200)
  static auto wait_on_address =
      reinterpret_cast<BOOL(WINAPI*)(void*, PVOID, SIZE_T, DWORD)>(win32_get_synch_api_function("WaitOnAddress"));
  if (wait_on_address != nullptr) {
    auto timeout_ms = [&]() -> DWORD {
      if constexpr (is_same_v<MaybeTimeout, NoTimeout>) {
        return INFINITE;
      } else {
        uint64_t ms = maybe_timeout_ns / 1'000'000;
        if (ms == 0 && maybe_timeout_ns > 100'000)
          // Round up to 1ms if requested between 100us - 1ms
          return 1;

        return static_cast<DWORD>(std::min(static_cast<uint64_t>(INFINITE), ms));
      }
    }();
    wait_on_address(const_cast<void*>(__ptr), const_cast<void*>(__val), _Size, timeout_ms);
  } else {
    std::chrono::nanoseconds timeout = std::chrono::nanoseconds(0);
    if constexpr (!is_same_v<MaybeTimeout, NoTimeout>) {
      timeout = std::chrono::nanoseconds(maybe_timeout_ns);
    }
    __libcpp_thread_poll_with_backoff(
        [=]() -> bool { return std::memcmp(const_cast<const void*>(__ptr), __val, _Size) != 0; },
        __libcpp_timed_backoff_policy(),
        timeout);
  }
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const* __ptr, bool __notify_one) {
  static_assert(_Size == 8, "Can only wake up on 8 bytes value");
  if (__notify_one) {
    // WakeByAddressSingle was added in Windows 8 (build 9200)
    static auto wake_by_address_single =
        reinterpret_cast<void(WINAPI*)(PVOID)>(win32_get_synch_api_function("WakeByAddressSingle"));
    if (wake_by_address_single != nullptr) {
      wake_by_address_single(const_cast<void*>(__ptr));
    } else {
      // The fallback implementation of waking does nothing, as the fallback wait implementation just does polling, so
      // there's nothing to do here.
    }
  } else {
    // WakeByAddressAll was added in Windows 8 (build 9200)
    static auto wake_by_address_all =
        reinterpret_cast<void(WINAPI*)(PVOID)>(win32_get_synch_api_function("WakeByAddressAll"));
    if (wake_by_address_all != nullptr) {
      wake_by_address_all(const_cast<void*>(__ptr));
    } else {
      // The fallback implementation of waking does nothing, as the fallback wait implementation just does polling, so
      // there's nothing to do here.
    }
  }
}

#elif defined(__Fuchsia__)

template <std::size_t _Size>
static inline zx_futex_t const* __get_zx_futex(void const* __ptr) {
  static_assert(_Size == sizeof(zx_futex_t), "Can only wait/wake on zx_futex_t-compatible value");

  // Implicitly link against the vDSO system call ABI without requiring the
  // final link to specify -lzircon explicitly when statically linking libc++.
#  pragma comment(lib, "zircon")

  return reinterpret_cast<zx_futex_t const*>(__ptr);
}

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  zx_futex_t val;
  std::memcpy(&val, __val, _Size);
  zx_instant_mono_t deadline;
  if constexpr (is_same_v<MaybeTimeout, NoTimeout>) {
    deadline = ZX_TIME_INFINITE;
  } else {
    deadline = _zx_deadline_after(maybe_timeout_ns);
  }
  _zx_futex_wait(__get_zx_futex<_Size>(__ptr), val, ZX_HANDLE_INVALID, deadline);
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const* __ptr, bool __notify_one) {
  _zx_futex_wake(__get_zx_futex<_Size>(__ptr), __notify_one ? 1 : UINT32_MAX);
}

#else // <- Add other operating systems here

// Baseline is just a timed backoff

template <std::size_t _Size, class MaybeTimeout>
static void __platform_wait_on_address(void const* __ptr, void const* __val, MaybeTimeout maybe_timeout_ns) {
  std::chrono::nanoseconds timeout = std::chrono::nanoseconds(0);
  if constexpr (!is_same_v<MaybeTimeout, NoTimeout>) {
    timeout = std::chrono::nanoseconds(maybe_timeout_ns);
  }
  __libcpp_thread_poll_with_backoff(
      [=]() -> bool { return std::memcmp(const_cast<const void*>(__ptr), __val, _Size) != 0; },
      __libcpp_timed_backoff_policy(),
      timeout);
}

template <std::size_t _Size>
static void __platform_wake_by_address(void const*, bool) {}

#endif // __linux__

// =============================
// Local hidden helper functions
// =============================

/* Given an atomic to track contention and an atomic to actually wait on, which may be
   the same atomic, we try to detect contention to avoid spuriously calling the platform. */

template <std::size_t _Size>
static void
__contention_notify(__cxx_atomic_contention_t* __waiter_count, void const* __address_to_notify, bool __notify_one) {
  if (0 != __cxx_atomic_load(__waiter_count, memory_order_seq_cst))
    // We only call 'wake' if we consumed a contention bit here.
    __platform_wake_by_address<_Size>(__address_to_notify, __notify_one);
}

template <std::size_t _Size, class MaybeTimeout>
static void __contention_wait(__cxx_atomic_contention_t* __waiter_count,
                              void const* __address_to_wait,
                              void const* __old_value,
                              MaybeTimeout maybe_timeout_ns) {
  __cxx_atomic_fetch_add(__waiter_count, __cxx_contention_t(1), memory_order_relaxed);
  // https://llvm.org/PR109290
  // There are no platform guarantees of a memory barrier in the platform wait implementation
  __cxx_atomic_thread_fence(memory_order_seq_cst);
  // We sleep as long as the monitored value hasn't changed.
  __platform_wait_on_address<_Size>(__address_to_wait, __old_value, maybe_timeout_ns);
  __cxx_atomic_fetch_sub(__waiter_count, __cxx_contention_t(1), memory_order_release);
}

static constexpr size_t __contention_table_size = (1 << 8); /* < there's no magic in this number */

static constexpr hash<void const*> __contention_hasher;

// Unified contention table shared by both the native path (atomics whose size matches the platform
// wait size and are waited on directly) and the global path (atomics with the wrong size, which use
// the global table's atomic as the wait/notify address). The native path only reads/writes the native
// waiter count, and the global path uses the global waiter count and the platform state.
//
// We use a single table for both the native and non-native states to minimize the size impact, since
// we would have plenty of unused padding otherwise. This technically means that if both a native and
// a non-native atomic fall into the same bucket and are reading/writing waiter counts on the same
// cache line, we would end up in a "false-sharing" situation. However, in practice, it would require
// both native and non-native atomics to be used concurrently, and to fall in the same bucket, which
// is expected to be unlikely.
struct alignas(std::hardware_destructive_interference_size) /* aim to avoid false sharing */ __contention_state {
  __cxx_atomic_contention_t __waiter_count_native;
  __cxx_atomic_contention_t __waiter_count_global;
  __cxx_atomic_contention_t __platform_state;
  constexpr __contention_state() : __waiter_count_native(0), __waiter_count_global(0), __platform_state(0) {}
};

static constinit __contention_state __contention_table[__contention_table_size];

static __contention_state* __get_contention_state(void const* p) {
  return &__contention_table[__contention_hasher(p) & (__contention_table_size - 1)];
}

/* When the incoming atomic is the wrong size for the platform wait size, need to
   launder the value sequence through an atomic from our table. */

static void __atomic_notify_global_table(void const* __location) {
  auto const __entry = __get_contention_state(__location);
  // The value sequence laundering happens on the next line below.
  __cxx_atomic_fetch_add(&__entry->__platform_state, __cxx_contention_t(1), memory_order_seq_cst);
  __contention_notify<sizeof(__cxx_atomic_contention_t)>(
      &__entry->__waiter_count_global,
      &__entry->__platform_state,
      false /* when laundering, we can't handle notify_one */);
}

// =============================
// New dylib exported symbols
// =============================

// global
_LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t __atomic_monitor_global(void const* __location) noexcept {
  auto const __entry = __get_contention_state(__location);
  return __cxx_atomic_load(&__entry->__platform_state, memory_order_acquire);
}

_LIBCPP_EXPORTED_FROM_ABI void
__atomic_wait_global_table(void const* __location, __cxx_contention_t __old_value) noexcept {
  auto const __entry = __get_contention_state(__location);
  __contention_wait<sizeof(__cxx_atomic_contention_t)>(
      &__entry->__waiter_count_global, &__entry->__platform_state, &__old_value, NoTimeout{});
}

_LIBCPP_EXPORTED_FROM_ABI void __atomic_wait_global_table_with_timeout(
    void const* __location, __cxx_contention_t __old_value, uint64_t __timeout_ns) noexcept {
  auto const __entry = __get_contention_state(__location);
  __contention_wait<sizeof(__cxx_atomic_contention_t)>(
      &__entry->__waiter_count_global, &__entry->__platform_state, &__old_value, __timeout_ns);
}

_LIBCPP_EXPORTED_FROM_ABI void __atomic_notify_one_global_table(void const* __location) noexcept {
  __atomic_notify_global_table(__location);
}

_LIBCPP_EXPORTED_FROM_ABI void __atomic_notify_all_global_table(void const* __location) noexcept {
  __atomic_notify_global_table(__location);
}

// native

template <std::size_t _Size>
_LIBCPP_EXPORTED_FROM_ABI void __atomic_wait_native(void const* __address, void const* __old_value) noexcept {
  __contention_wait<_Size>(
      &__get_contention_state(__address)->__waiter_count_native, __address, __old_value, NoTimeout{});
}

template <std::size_t _Size>
_LIBCPP_EXPORTED_FROM_ABI void
__atomic_wait_native_with_timeout(void const* __address, void const* __old_value, uint64_t __timeout_ns) noexcept {
  __contention_wait<_Size>(
      &__get_contention_state(__address)->__waiter_count_native, __address, __old_value, __timeout_ns);
}

template <std::size_t _Size>
_LIBCPP_EXPORTED_FROM_ABI void __atomic_notify_one_native(void const* __location) noexcept {
  __contention_notify<_Size>(&__get_contention_state(__location)->__waiter_count_native, __location, true);
}

template <std::size_t _Size>
_LIBCPP_EXPORTED_FROM_ABI void __atomic_notify_all_native(void const* __location) noexcept {
  __contention_notify<_Size>(&__get_contention_state(__location)->__waiter_count_native, __location, false);
}

// ==================================================
// Instantiation of the templates with supported size
// ==================================================

#if defined(_LIBCPP_ABI_ATOMIC_WAIT_NATIVE_BY_SIZE)

#  define _INSTANTIATE(_SIZE)                                                                                          \
    template _LIBCPP_EXPORTED_FROM_ABI void __atomic_wait_native<_SIZE>(void const*, void const*) noexcept;            \
    template _LIBCPP_EXPORTED_FROM_ABI void __atomic_wait_native_with_timeout<_SIZE>(                                  \
        void const*, void const*, uint64_t) noexcept;                                                                  \
    template _LIBCPP_EXPORTED_FROM_ABI void __atomic_notify_one_native<_SIZE>(void const*) noexcept;                   \
    template _LIBCPP_EXPORTED_FROM_ABI void __atomic_notify_all_native<_SIZE>(void const*) noexcept;

_LIBCPP_NATIVE_PLATFORM_WAIT_SIZES(_INSTANTIATE)

#  undef _INSTANTIATE

#else // _LIBCPP_ABI_ATOMIC_WAIT_NATIVE_BY_SIZE

template _LIBCPP_EXPORTED_FROM_ABI void
__atomic_wait_native<sizeof(__cxx_contention_t)>(void const* __address, void const* __old_value) noexcept;

template _LIBCPP_EXPORTED_FROM_ABI void __atomic_wait_native_with_timeout<sizeof(__cxx_contention_t)>(
    void const* __address, void const* __old_value, uint64_t) noexcept;

template _LIBCPP_EXPORTED_FROM_ABI void
__atomic_notify_one_native<sizeof(__cxx_contention_t)>(void const* __location) noexcept;

template _LIBCPP_EXPORTED_FROM_ABI void
__atomic_notify_all_native<sizeof(__cxx_contention_t)>(void const* __location) noexcept;

#endif // _LIBCPP_ABI_ATOMIC_WAIT_NATIVE_BY_SIZE

// =============================================================
// Old dylib exported symbols, for backwards compatibility
// =============================================================
_LIBCPP_DIAGNOSTIC_PUSH
_LIBCPP_CLANG_DIAGNOSTIC_IGNORED("-Wmissing-prototypes")

_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(void const volatile* __location) noexcept {
  __atomic_notify_global_table(const_cast<void const*>(__location));
}

_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_all(void const volatile* __location) noexcept {
  __atomic_notify_global_table(const_cast<void const*>(__location));
}

_LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t __libcpp_atomic_monitor(void const volatile* __location) noexcept {
  auto const __entry = __get_contention_state(const_cast<void const*>(__location));
  return __cxx_atomic_load(&__entry->__platform_state, memory_order_acquire);
}

_LIBCPP_EXPORTED_FROM_ABI void
__libcpp_atomic_wait(void const volatile* __location, __cxx_contention_t __old_value) noexcept {
  auto const __entry = __get_contention_state(const_cast<void const*>(__location));
  __contention_wait<sizeof(__cxx_atomic_contention_t)>(
      &__entry->__waiter_count_global, &__entry->__platform_state, &__old_value, NoTimeout{});
}

_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(__cxx_atomic_contention_t const volatile* __location) noexcept {
  auto __location_cast = const_cast<const void*>(static_cast<const volatile void*>(__location));
  __contention_notify<sizeof(__cxx_atomic_contention_t)>(
      &__get_contention_state(__location_cast)->__waiter_count_native, __location_cast, true);
}

_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_all(__cxx_atomic_contention_t const volatile* __location) noexcept {
  auto __location_cast = const_cast<const void*>(static_cast<const volatile void*>(__location));
  __contention_notify<sizeof(__cxx_atomic_contention_t)>(
      &__get_contention_state(__location_cast)->__waiter_count_native, __location_cast, false);
}

_LIBCPP_EXPORTED_FROM_ABI void
__libcpp_atomic_wait(__cxx_atomic_contention_t const volatile* __location, __cxx_contention_t __old_value) noexcept {
  auto __location_cast = const_cast<const void*>(static_cast<const volatile void*>(__location));
  __contention_wait<sizeof(__cxx_atomic_contention_t)>(
      &__get_contention_state(__location_cast)->__waiter_count_native, __location_cast, &__old_value, NoTimeout{});
}

// this function is even unused in the old ABI
_LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t
__libcpp_atomic_monitor(__cxx_atomic_contention_t const volatile* __location) noexcept {
  return __cxx_atomic_load(__location, memory_order_acquire);
}

_LIBCPP_DIAGNOSTIC_POP

_LIBCPP_END_EXPLICIT_ABI_ANNOTATIONS
_LIBCPP_END_NAMESPACE_STD

_LIBCPP_POP_MACROS