merge llvm libcxxabi 16.0.6
authorrobert <robert@openbsd.org>
Fri, 26 Jan 2024 11:45:17 +0000 (11:45 +0000)
committerrobert <robert@openbsd.org>
Fri, 26 Jan 2024 11:45:17 +0000 (11:45 +0000)
gnu/llvm/libcxxabi/cmake/Modules/HandleCompilerRT.cmake [deleted file]
gnu/llvm/libcxxabi/src/cxa_guard_impl.h
gnu/llvm/libcxxabi/src/include/atomic_support.h [deleted file]

diff --git a/gnu/llvm/libcxxabi/cmake/Modules/HandleCompilerRT.cmake b/gnu/llvm/libcxxabi/cmake/Modules/HandleCompilerRT.cmake
deleted file mode 100644 (file)
index 1f95327..0000000
+++ /dev/null
@@ -1,64 +0,0 @@
-function(find_compiler_rt_library name dest)
-  if (NOT DEFINED LIBCXXABI_COMPILE_FLAGS)
-    message(FATAL_ERROR "LIBCXXABI_COMPILE_FLAGS must be defined when using this function")
-  endif()
-  set(dest "" PARENT_SCOPE)
-  set(CLANG_COMMAND ${CMAKE_CXX_COMPILER} ${LIBCXXABI_COMPILE_FLAGS}
-      "--rtlib=compiler-rt" "--print-libgcc-file-name")
-  if (CMAKE_CXX_COMPILER_ID MATCHES Clang AND CMAKE_CXX_COMPILER_TARGET)
-    list(APPEND CLANG_COMMAND "--target=${CMAKE_CXX_COMPILER_TARGET}")
-  endif()
-  get_property(LIBCXXABI_CXX_FLAGS CACHE CMAKE_CXX_FLAGS PROPERTY VALUE)
-  string(REPLACE " " ";" LIBCXXABI_CXX_FLAGS "${LIBCXXABI_CXX_FLAGS}")
-  list(APPEND CLANG_COMMAND ${LIBCXXABI_CXX_FLAGS})
-  execute_process(
-      COMMAND ${CLANG_COMMAND}
-      RESULT_VARIABLE HAD_ERROR
-      OUTPUT_VARIABLE LIBRARY_FILE
-  )
-  string(STRIP "${LIBRARY_FILE}" LIBRARY_FILE)
-  file(TO_CMAKE_PATH "${LIBRARY_FILE}" LIBRARY_FILE)
-  string(REPLACE "builtins" "${name}" LIBRARY_FILE "${LIBRARY_FILE}")
-  if (NOT HAD_ERROR AND EXISTS "${LIBRARY_FILE}")
-    message(STATUS "Found compiler-rt library: ${LIBRARY_FILE}")
-    set(${dest} "${LIBRARY_FILE}" PARENT_SCOPE)
-  else()
-    message(STATUS "Failed to find compiler-rt library")
-  endif()
-endfunction()
-
-function(find_compiler_rt_dir dest)
-  if (NOT DEFINED LIBCXXABI_COMPILE_FLAGS)
-    message(FATAL_ERROR "LIBCXXABI_COMPILE_FLAGS must be defined when using this function")
-  endif()
-  set(dest "" PARENT_SCOPE)
-  if (APPLE)
-    set(CLANG_COMMAND ${CMAKE_CXX_COMPILER} ${LIBCXXABI_COMPILE_FLAGS}
-        "-print-file-name=lib")
-    execute_process(
-        COMMAND ${CLANG_COMMAND}
-        RESULT_VARIABLE HAD_ERROR
-        OUTPUT_VARIABLE LIBRARY_DIR
-    )
-    string(STRIP "${LIBRARY_DIR}" LIBRARY_DIR)
-    file(TO_CMAKE_PATH "${LIBRARY_DIR}" LIBRARY_DIR)
-    set(LIBRARY_DIR "${LIBRARY_DIR}/darwin")
-  else()
-    set(CLANG_COMMAND ${CMAKE_CXX_COMPILER} ${LIBCXXABI_COMPILE_FLAGS}
-        "--rtlib=compiler-rt" "--print-libgcc-file-name")
-    execute_process(
-        COMMAND ${CLANG_COMMAND}
-        RESULT_VARIABLE HAD_ERROR
-        OUTPUT_VARIABLE LIBRARY_FILE
-    )
-    string(STRIP "${LIBRARY_FILE}" LIBRARY_FILE)
-    file(TO_CMAKE_PATH "${LIBRARY_FILE}" LIBRARY_FILE)
-    get_filename_component(LIBRARY_DIR "${LIBRARY_FILE}" DIRECTORY)
-  endif()
-  if (NOT HAD_ERROR AND EXISTS "${LIBRARY_DIR}")
-    message(STATUS "Found compiler-rt directory: ${LIBRARY_DIR}")
-    set(${dest} "${LIBRARY_DIR}" PARENT_SCOPE)
-  else()
-    message(STATUS "Failed to find compiler-rt directory")
-  endif()
-endfunction()
index 5ca4c57..7c17467 100644 (file)
@@ -5,6 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+
 #ifndef LIBCXXABI_SRC_INCLUDE_CXA_GUARD_IMPL_H
 #define LIBCXXABI_SRC_INCLUDE_CXA_GUARD_IMPL_H
 
  * the thread currently performing initialization is stored in the second word.
  *
  *  Guard Object Layout:
- * -------------------------------------------------------------------------
- * |a: guard byte | a+1: init byte | a+2 : unused ... | a+4: thread-id ... |
- * ------------------------------------------------------------------------
+ * ---------------------------------------------------------------------------
+ * | a+0: guard byte | a+1: init byte | a+2: unused ... | a+4: thread-id ... |
+ * ---------------------------------------------------------------------------
+ *
+ * Note that we don't do what the ABI docs suggest (put a mutex in the guard
+ * object which we acquire in cxa_guard_acquire and release in
+ * cxa_guard_release). Instead we use the init byte to imitate that behaviour,
+ * but without actually holding anything mutex related between aquire and
+ * release/abort.
  *
  *  Access Protocol:
  *    For each implementation the guard byte is checked and set before accessing
  */
 
 #include "__cxxabi_config.h"
-#include "include/atomic_support.h"
-#include <unistd.h>
+#include "include/atomic_support.h" // from libc++
 #if defined(__has_include)
-# if __has_include(<sys/syscall.h>)
-#   include <sys/syscall.h>
-# endif
+#  if __has_include(<sys/syscall.h>)
+#    include <sys/syscall.h>
+#  endif
+#  if __has_include(<unistd.h>)
+#    include <unistd.h>
+#  endif
 #endif
 
-#include <stdlib.h>
 #include <__threading_support>
+#include <cstdint>
+#include <cstring>
+#include <limits.h>
+#include <stdlib.h>
+
 #ifndef _LIBCXXABI_HAS_NO_THREADS
-#if defined(__ELF__) && defined(_LIBCXXABI_LINK_PTHREAD_LIB)
-#pragma comment(lib, "pthread")
-#endif
+#  if defined(__ELF__) && defined(_LIBCXXABI_LINK_PTHREAD_LIB)
+#    pragma comment(lib, "pthread")
+#  endif
 #endif
 
 #if defined(__clang__)
-# pragma clang diagnostic push
-# pragma clang diagnostic ignored "-Wtautological-pointer-compare"
+#  pragma clang diagnostic push
+#  pragma clang diagnostic ignored "-Wtautological-pointer-compare"
 #elif defined(__GNUC__)
-# pragma GCC diagnostic push
-# pragma GCC diagnostic ignored "-Waddress"
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Waddress"
 #endif
 
 // To make testing possible, this header is included from both cxa_guard.cpp
 // defined when including this file. Only `src/cxa_guard.cpp` should define
 // the former.
 #ifdef BUILDING_CXA_GUARD
-# include "abort_message.h"
-# define ABORT_WITH_MESSAGE(...) ::abort_message(__VA_ARGS__)
+#  include "abort_message.h"
+#  define ABORT_WITH_MESSAGE(...) ::abort_message(__VA_ARGS__)
 #elif defined(TESTING_CXA_GUARD)
-# define ABORT_WITH_MESSAGE(...) ::abort()
+#  define ABORT_WITH_MESSAGE(...) ::abort()
 #else
-# error "Either BUILDING_CXA_GUARD or TESTING_CXA_GUARD must be defined"
+#  error "Either BUILDING_CXA_GUARD or TESTING_CXA_GUARD must be defined"
 #endif
 
 #if __has_feature(thread_sanitizer)
 extern "C" void __tsan_acquire(void*);
 extern "C" void __tsan_release(void*);
 #else
-#define __tsan_acquire(addr) ((void)0)
-#define __tsan_release(addr) ((void)0)
+#  define __tsan_acquire(addr) ((void)0)
+#  define __tsan_release(addr) ((void)0)
 #endif
 
 namespace __cxxabiv1 {
@@ -99,7 +112,7 @@ namespace {
 //                          Misc Utilities
 //===----------------------------------------------------------------------===//
 
-template <class T, T(*Init)()>
+template <class T, T (*Init)()>
 struct LazyValue {
   LazyValue() : is_init(false) {}
 
@@ -110,7 +123,8 @@ struct LazyValue {
     }
     return value;
   }
- private:
+
+private:
   T value;
   bool is_init = false;
 };
@@ -120,25 +134,19 @@ class AtomicInt {
 public:
   using MemoryOrder = std::__libcpp_atomic_order;
 
-  explicit AtomicInt(IntType *b) : b_(b) {}
+  explicit AtomicInt(IntTypeb) : b_(b) {}
   AtomicInt(AtomicInt const&) = delete;
   AtomicInt& operator=(AtomicInt const&) = delete;
 
-  IntType load(MemoryOrder ord) {
-    return std::__libcpp_atomic_load(b_, ord);
-  }
-  void store(IntType val, MemoryOrder ord) {
-    std::__libcpp_atomic_store(b_, val, ord);
-  }
-  IntType exchange(IntType new_val, MemoryOrder ord) {
-    return std::__libcpp_atomic_exchange(b_, new_val, ord);
-  }
-  bool compare_exchange(IntType *expected, IntType desired, MemoryOrder ord_success, MemoryOrder ord_failure) {
+  IntType load(MemoryOrder ord) { return std::__libcpp_atomic_load(b_, ord); }
+  void store(IntType val, MemoryOrder ord) { std::__libcpp_atomic_store(b_, val, ord); }
+  IntType exchange(IntType new_val, MemoryOrder ord) { return std::__libcpp_atomic_exchange(b_, new_val, ord); }
+  bool compare_exchange(IntType* expected, IntType desired, MemoryOrder ord_success, MemoryOrder ord_failure) {
     return std::__libcpp_atomic_compare_exchange(b_, expected, desired, ord_success, ord_failure);
   }
 
 private:
-  IntType *b_;
+  IntTypeb_;
 };
 
 //===----------------------------------------------------------------------===//
@@ -148,8 +156,7 @@ private:
 #if defined(__APPLE__) && defined(_LIBCPP_HAS_THREAD_API_PTHREAD)
 uint32_t PlatformThreadID() {
   static_assert(sizeof(mach_port_t) == sizeof(uint32_t), "");
-  return static_cast<uint32_t>(
-      pthread_mach_thread_np(std::__libcpp_thread_get_current_id()));
+  return static_cast<uint32_t>(pthread_mach_thread_np(std::__libcpp_thread_get_current_id()));
 }
 #elif defined(SYS_gettid) && defined(_LIBCPP_HAS_THREAD_API_PTHREAD)
 uint32_t PlatformThreadID() {
@@ -160,99 +167,108 @@ uint32_t PlatformThreadID() {
 constexpr uint32_t (*PlatformThreadID)() = nullptr;
 #endif
 
-
-constexpr bool PlatformSupportsThreadID() {
-  return +PlatformThreadID != nullptr;
-}
-
 //===----------------------------------------------------------------------===//
-//                          GuardBase
+//                          GuardByte
 //===----------------------------------------------------------------------===//
 
-enum class AcquireResult {
-  INIT_IS_DONE,
-  INIT_IS_PENDING,
-};
-constexpr AcquireResult INIT_IS_DONE = AcquireResult::INIT_IS_DONE;
-constexpr AcquireResult INIT_IS_PENDING = AcquireResult::INIT_IS_PENDING;
-
 static constexpr uint8_t UNSET = 0;
 static constexpr uint8_t COMPLETE_BIT = (1 << 0);
 static constexpr uint8_t PENDING_BIT = (1 << 1);
 static constexpr uint8_t WAITING_BIT = (1 << 2);
 
-template <class Derived>
-struct GuardObject {
-  GuardObject() = delete;
-  GuardObject(GuardObject const&) = delete;
-  GuardObject& operator=(GuardObject const&) = delete;
+/// Manages reads and writes to the guard byte.
+struct GuardByte {
+  GuardByte() = delete;
+  GuardByte(GuardByte const&) = delete;
+  GuardByte& operator=(GuardByte const&) = delete;
 
-  explicit GuardObject(uint32_t* g)
-      : base_address(g), guard_byte_address(reinterpret_cast<uint8_t*>(g)),
-        init_byte_address(reinterpret_cast<uint8_t*>(g) + 1),
-        thread_id_address(nullptr) {}
-
-  explicit GuardObject(uint64_t* g)
-      : base_address(g), guard_byte_address(reinterpret_cast<uint8_t*>(g)),
-        init_byte_address(reinterpret_cast<uint8_t*>(g) + 1),
-        thread_id_address(reinterpret_cast<uint32_t*>(g) + 1) {}
+  explicit GuardByte(uint8_t* const guard_byte_address) : guard_byte(guard_byte_address) {}
 
 public:
-  /// Implements __cxa_guard_acquire
-  AcquireResult cxa_guard_acquire() {
-    AtomicInt<uint8_t> guard_byte(guard_byte_address);
-    if (guard_byte.load(std::_AO_Acquire) != UNSET)
-      return INIT_IS_DONE;
-    return derived()->acquire_init_byte();
+  /// The guard byte portion of cxa_guard_acquire. Returns true if
+  /// initialization has already been completed.
+  bool acquire() {
+    // if guard_byte is non-zero, we have already completed initialization
+    // (i.e. release has been called)
+    return guard_byte.load(std::_AO_Acquire) != UNSET;
   }
 
-  /// Implements __cxa_guard_release
-  void cxa_guard_release() {
-    AtomicInt<uint8_t> guard_byte(guard_byte_address);
-    // Store complete first, so that when release wakes other folks, they see
-    // it as having been completed.
-    guard_byte.store(COMPLETE_BIT, std::_AO_Release);
-    derived()->release_init_byte();
-  }
+  /// The guard byte portion of cxa_guard_release.
+  void release() { guard_byte.store(COMPLETE_BIT, std::_AO_Release); }
 
-  /// Implements __cxa_guard_abort
-  void cxa_guard_abort() { derived()->abort_init_byte(); }
-
-public:
-  /// base_address - the address of the original guard object.
-  void* const base_address;
-  /// The address of the guard byte at offset 0.
-  uint8_t* const guard_byte_address;
-  /// The address of the byte used by the implementation during initialization.
-  uint8_t* const init_byte_address;
-  /// An optional address storing an identifier for the thread performing initialization.
-  /// It's used to detect recursive initialization.
-  uint32_t* const thread_id_address;
+  /// The guard byte portion of cxa_guard_abort.
+  void abort() {} // Nothing to do
 
 private:
-  Derived* derived() { return static_cast<Derived*>(this); }
+  AtomicInt<uint8_t> guard_byte;
 };
 
+//===----------------------------------------------------------------------===//
+//                       InitByte Implementations
+//===----------------------------------------------------------------------===//
+//
+// Each initialization byte implementation supports the following methods:
+//
+//  InitByte(uint8_t* _init_byte_address, uint32_t* _thread_id_address)
+//    Construct the InitByte object, initializing our member variables
+//
+//  bool acquire()
+//    Called before we start the initialization. Check if someone else has already started, and if
+//    not to signal our intent to start it ourselves. We determine the current status from the init
+//    byte, which is one of 4 possible values:
+//      COMPLETE:           Initialization was finished by somebody else. Return true.
+//      PENDING:            Somebody has started the initialization already, set the WAITING bit,
+//                          then wait for the init byte to get updated with a new value.
+//      (PENDING|WAITING):  Somebody has started the initialization already, and we're not the
+//                          first one waiting. Wait for the init byte to get updated.
+//      UNSET:              Initialization hasn't successfully completed, and nobody is currently
+//                          performing the initialization. Set the PENDING bit to indicate our
+//                          intention to start the initialization, and return false.
+//    The return value indicates whether initialization has already been completed.
+//
+//  void release()
+//    Called after successfully completing the initialization. Update the init byte to reflect
+//    that, then if anybody else is waiting, wake them up.
+//
+//  void abort()
+//    Called after an error is thrown during the initialization. Reset the init byte to UNSET to
+//    indicate that we're no longer performing the initialization, then if anybody is waiting, wake
+//    them up so they can try performing the initialization.
+//
+
 //===----------------------------------------------------------------------===//
 //                    Single Threaded Implementation
 //===----------------------------------------------------------------------===//
 
-struct InitByteNoThreads : GuardObject<InitByteNoThreads> {
-  using GuardObject::GuardObject;
+/// InitByteNoThreads - Doesn't use any inter-thread synchronization when
+/// managing reads and writes to the init byte.
+struct InitByteNoThreads {
+  InitByteNoThreads() = delete;
+  InitByteNoThreads(InitByteNoThreads const&) = delete;
+  InitByteNoThreads& operator=(InitByteNoThreads const&) = delete;
+
+  explicit InitByteNoThreads(uint8_t* _init_byte_address, uint32_t*) : init_byte_address(_init_byte_address) {}
 
-  AcquireResult acquire_init_byte() {
+  /// The init byte portion of cxa_guard_acquire. Returns true if
+  /// initialization has already been completed.
+  bool acquire() {
     if (*init_byte_address == COMPLETE_BIT)
-      return INIT_IS_DONE;
+      return true;
     if (*init_byte_address & PENDING_BIT)
       ABORT_WITH_MESSAGE("__cxa_guard_acquire detected recursive initialization");
     *init_byte_address = PENDING_BIT;
-    return INIT_IS_PENDING;
+    return false;
   }
 
-  void release_init_byte() { *init_byte_address = COMPLETE_BIT; }
-  void abort_init_byte() { *init_byte_address = UNSET; }
-};
+  /// The init byte portion of cxa_guard_release.
+  void release() { *init_byte_address = COMPLETE_BIT; }
+  /// The init byte portion of cxa_guard_abort.
+  void abort() { *init_byte_address = UNSET; }
 
+private:
+  /// The address of the byte used during initialization.
+  uint8_t* const init_byte_address;
+};
 
 //===----------------------------------------------------------------------===//
 //                     Global Mutex Implementation
@@ -280,9 +296,7 @@ struct LibcppCondVar {
   LibcppCondVar(LibcppCondVar const&) = delete;
   LibcppCondVar& operator=(LibcppCondVar const&) = delete;
 
-  bool wait(LibcppMutex& mut) {
-    return std::__libcpp_condvar_wait(&cond, &mut.mutex);
-  }
+  bool wait(LibcppMutex& mut) { return std::__libcpp_condvar_wait(&cond, &mut.mutex); }
   bool broadcast() { return std::__libcpp_condvar_broadcast(&cond); }
 
 private:
@@ -293,28 +307,25 @@ struct LibcppMutex {};
 struct LibcppCondVar {};
 #endif // !defined(_LIBCXXABI_HAS_NO_THREADS)
 
-
+/// InitByteGlobalMutex - Uses a global mutex and condition variable (common to
+/// all static local variables) to manage reads and writes to the init byte.
 template <class Mutex, class CondVar, Mutex& global_mutex, CondVar& global_cond,
           uint32_t (*GetThreadID)() = PlatformThreadID>
-struct InitByteGlobalMutex
-    : GuardObject<InitByteGlobalMutex<Mutex, CondVar, global_mutex, global_cond,
-                                    GetThreadID>> {
-
-  using BaseT = typename InitByteGlobalMutex::GuardObject;
-  using BaseT::BaseT;
+struct InitByteGlobalMutex {
 
-  explicit InitByteGlobalMutex(uint32_t *g)
-    : BaseT(g), has_thread_id_support(false) {}
-  explicit InitByteGlobalMutex(uint64_t *g)
-    : BaseT(g), has_thread_id_support(PlatformSupportsThreadID()) {}
+  explicit InitByteGlobalMutex(uint8_t* _init_byte_address, uint32_t* _thread_id_address)
+      : init_byte_address(_init_byte_address), thread_id_address(_thread_id_address),
+        has_thread_id_support(_thread_id_address != nullptr && GetThreadID != nullptr) {}
 
 public:
-  AcquireResult acquire_init_byte() {
+  /// The init byte portion of cxa_guard_acquire. Returns true if
+  /// initialization has already been completed.
+  bool acquire() {
     LockGuard g("__cxa_guard_acquire");
     // Check for possible recursive initialization.
     if (has_thread_id_support && (*init_byte_address & PENDING_BIT)) {
       if (*thread_id_address == current_thread_id.get())
-       ABORT_WITH_MESSAGE("__cxa_guard_acquire detected recursive initialization");
+        ABORT_WITH_MESSAGE("__cxa_guard_acquire detected recursive initialization");
     }
 
     // Wait until the pending bit is not set.
@@ -324,16 +335,17 @@ public:
     }
 
     if (*init_byte_address == COMPLETE_BIT)
-      return INIT_IS_DONE;
+      return true;
 
     if (has_thread_id_support)
       *thread_id_address = current_thread_id.get();
 
     *init_byte_address = PENDING_BIT;
-    return INIT_IS_PENDING;
+    return false;
   }
 
-  void release_init_byte() {
+  /// The init byte portion of cxa_guard_release.
+  void release() {
     bool has_waiting;
     {
       LockGuard g("__cxa_guard_release");
@@ -347,7 +359,8 @@ public:
     }
   }
 
-  void abort_init_byte() {
+  /// The init byte portion of cxa_guard_abort.
+  void abort() {
     bool has_waiting;
     {
       LockGuard g("__cxa_guard_abort");
@@ -364,8 +377,12 @@ public:
   }
 
 private:
-  using BaseT::init_byte_address;
-  using BaseT::thread_id_address;
+  /// The address of the byte used during initialization.
+  uint8_t* const init_byte_address;
+  /// An optional address storing an identifier for the thread performing initialization.
+  /// It's used to detect recursive initialization.
+  uint32_t* const thread_id_address;
+
   const bool has_thread_id_support;
   LazyValue<uint32_t, GetThreadID> current_thread_id;
 
@@ -375,8 +392,7 @@ private:
     LockGuard(LockGuard const&) = delete;
     LockGuard& operator=(LockGuard const&) = delete;
 
-    explicit LockGuard(const char* calling_func)
-        : calling_func_(calling_func)  {
+    explicit LockGuard(const char* calling_func) : calling_func_(calling_func) {
       if (global_mutex.lock())
         ABORT_WITH_MESSAGE("%s failed to acquire mutex", calling_func_);
     }
@@ -424,50 +440,40 @@ constexpr void (*PlatformFutexWait)(int*, int) = nullptr;
 constexpr void (*PlatformFutexWake)(int*) = nullptr;
 #endif
 
-constexpr bool PlatformSupportsFutex() {
-  return +PlatformFutexWait != nullptr;
-}
+constexpr bool PlatformSupportsFutex() { return +PlatformFutexWait != nullptr; }
 
-/// InitByteFutex - Manages initialization using atomics and the futex syscall
-/// for waiting and waking.
-template <void (*Wait)(int*, int) = PlatformFutexWait,
-          void (*Wake)(int*) = PlatformFutexWake,
+/// InitByteFutex - Uses a futex to manage reads and writes to the init byte.
+template <void (*Wait)(int*, int) = PlatformFutexWait, void (*Wake)(int*) = PlatformFutexWake,
           uint32_t (*GetThreadIDArg)() = PlatformThreadID>
-struct InitByteFutex : GuardObject<InitByteFutex<Wait, Wake, GetThreadIDArg>> {
-  using BaseT = typename InitByteFutex::GuardObject;
+struct InitByteFutex {
 
-  /// ARM Constructor
-  explicit InitByteFutex(uint32_t *g) : BaseT(g),
-    init_byte(this->init_byte_address),
-    has_thread_id_support(this->thread_id_address && GetThreadIDArg),
-    thread_id(this->thread_id_address) {}
-
-  /// Itanium Constructor
-  explicit InitByteFutex(uint64_t *g) : BaseT(g),
-    init_byte(this->init_byte_address),
-    has_thread_id_support(this->thread_id_address && GetThreadIDArg),
-    thread_id(this->thread_id_address) {}
+  explicit InitByteFutex(uint8_t* _init_byte_address, uint32_t* _thread_id_address)
+      : init_byte(_init_byte_address),
+        has_thread_id_support(_thread_id_address != nullptr && GetThreadIDArg != nullptr),
+        thread_id(_thread_id_address),
+        base_address(reinterpret_cast<int*>(/*_init_byte_address & ~0x3*/ _init_byte_address - 1)) {}
 
 public:
-  AcquireResult acquire_init_byte() {
+  /// The init byte portion of cxa_guard_acquire. Returns true if
+  /// initialization has already been completed.
+  bool acquire() {
     while (true) {
       uint8_t last_val = UNSET;
-      if (init_byte.compare_exchange(&last_val, PENDING_BIT, std::_AO_Acq_Rel,
-                                     std::_AO_Acquire)) {
+      if (init_byte.compare_exchange(&last_val, PENDING_BIT, std::_AO_Acq_Rel, std::_AO_Acquire)) {
         if (has_thread_id_support) {
           thread_id.store(current_thread_id.get(), std::_AO_Relaxed);
         }
-        return INIT_IS_PENDING;
+        return false;
       }
 
       if (last_val == COMPLETE_BIT)
-        return INIT_IS_DONE;
+        return true;
 
       if (last_val & PENDING_BIT) {
 
         // Check for recursive initialization
         if (has_thread_id_support && thread_id.load(std::_AO_Relaxed) == current_thread_id.get()) {
-            ABORT_WITH_MESSAGE("__cxa_guard_acquire detected recursive initialization");
+          ABORT_WITH_MESSAGE("__cxa_guard_acquire detected recursive initialization");
         }
 
         if ((last_val & WAITING_BIT) == 0) {
@@ -475,11 +481,10 @@ public:
           // (1) another thread finished the whole thing before we got here
           // (2) another thread set the waiting bit we were trying to thread
           // (3) another thread had an exception and failed to finish
-          if (!init_byte.compare_exchange(&last_val, PENDING_BIT | WAITING_BIT,
-                                          std::_AO_Acq_Rel, std::_AO_Release)) {
+          if (!init_byte.compare_exchange(&last_val, PENDING_BIT | WAITING_BIT, std::_AO_Acq_Rel, std::_AO_Release)) {
             // (1) success, via someone else's work!
             if (last_val == COMPLETE_BIT)
-              return INIT_IS_DONE;
+              return true;
 
             // (3) someone else, bailed on doing the work, retry from the start!
             if (last_val == UNSET)
@@ -493,30 +498,30 @@ public:
     }
   }
 
-  void release_init_byte() {
+  /// The init byte portion of cxa_guard_release.
+  void release() {
     uint8_t old = init_byte.exchange(COMPLETE_BIT, std::_AO_Acq_Rel);
     if (old & WAITING_BIT)
       wake_all();
   }
 
-  void abort_init_byte() {
+  /// The init byte portion of cxa_guard_abort.
+  void abort() {
     if (has_thread_id_support)
       thread_id.store(0, std::_AO_Relaxed);
 
-    uint8_t old = init_byte.exchange(0, std::_AO_Acq_Rel);
+    uint8_t old = init_byte.exchange(UNSET, std::_AO_Acq_Rel);
     if (old & WAITING_BIT)
       wake_all();
   }
 
 private:
   /// Use the futex to wait on the current guard variable. Futex expects a
-  /// 32-bit 4-byte aligned address as the first argument, so we have to use use
-  /// the base address of the guard variable (not the init byte).
-  void wait_on_initialization() {
-    Wait(static_cast<int*>(this->base_address),
-         expected_value_for_futex(PENDING_BIT | WAITING_BIT));
-  }
-  void wake_all() { Wake(static_cast<int*>(this->base_address)); }
+  /// 32-bit 4-byte aligned address as the first argument, so we use the 4-byte
+  /// aligned address that encompasses the init byte (i.e. the address of the
+  /// raw guard object that was passed to __cxa_guard_acquire/release/abort).
+  void wait_on_initialization() { Wait(base_address, expected_value_for_futex(PENDING_BIT | WAITING_BIT)); }
+  void wake_all() { Wake(base_address); }
 
 private:
   AtomicInt<uint8_t> init_byte;
@@ -526,6 +531,10 @@ private:
   AtomicInt<uint32_t> thread_id;
   LazyValue<uint32_t, GetThreadIDArg> current_thread_id;
 
+  /// the 4-byte-aligned address that encompasses the init byte (i.e. the
+  /// address of the raw guard object).
+  int* const base_address;
+
   /// Create the expected integer value for futex `wait(int* addr, int expected)`.
   /// We pass the base address as the first argument, So this function creates
   /// an zero-initialized integer  with `b` copied at the correct offset.
@@ -538,6 +547,86 @@ private:
   static_assert(Wait != nullptr && Wake != nullptr, "");
 };
 
+//===----------------------------------------------------------------------===//
+//                          GuardObject
+//===----------------------------------------------------------------------===//
+
+enum class AcquireResult {
+  INIT_IS_DONE,
+  INIT_IS_PENDING,
+};
+constexpr AcquireResult INIT_IS_DONE = AcquireResult::INIT_IS_DONE;
+constexpr AcquireResult INIT_IS_PENDING = AcquireResult::INIT_IS_PENDING;
+
+/// Co-ordinates between GuardByte and InitByte.
+template <class InitByteT>
+struct GuardObject {
+  GuardObject() = delete;
+  GuardObject(GuardObject const&) = delete;
+  GuardObject& operator=(GuardObject const&) = delete;
+
+private:
+  GuardByte guard_byte;
+  InitByteT init_byte;
+
+public:
+  /// ARM Constructor
+  explicit GuardObject(uint32_t* raw_guard_object)
+      : guard_byte(reinterpret_cast<uint8_t*>(raw_guard_object)),
+        init_byte(reinterpret_cast<uint8_t*>(raw_guard_object) + 1, nullptr) {}
+
+  /// Itanium Constructor
+  explicit GuardObject(uint64_t* raw_guard_object)
+      : guard_byte(reinterpret_cast<uint8_t*>(raw_guard_object)),
+        init_byte(reinterpret_cast<uint8_t*>(raw_guard_object) + 1, reinterpret_cast<uint32_t*>(raw_guard_object) + 1) {
+  }
+
+  /// Implements __cxa_guard_acquire.
+  AcquireResult cxa_guard_acquire() {
+    // Use short-circuit evaluation to avoid calling init_byte.acquire when
+    // guard_byte.acquire returns true. (i.e. don't call it when we know from
+    // the guard byte that initialization has already been completed)
+    if (guard_byte.acquire() || init_byte.acquire())
+      return INIT_IS_DONE;
+    return INIT_IS_PENDING;
+  }
+
+  /// Implements __cxa_guard_release.
+  void cxa_guard_release() {
+    // Update guard byte first, so if somebody is woken up by init_byte.release
+    // and comes all the way back around to __cxa_guard_acquire again, they see
+    // it as having completed initialization.
+    guard_byte.release();
+    init_byte.release();
+  }
+
+  /// Implements __cxa_guard_abort.
+  void cxa_guard_abort() {
+    guard_byte.abort();
+    init_byte.abort();
+  }
+};
+
+//===----------------------------------------------------------------------===//
+//                          Convenience Classes
+//===----------------------------------------------------------------------===//
+
+/// NoThreadsGuard - Manages initialization without performing any inter-thread
+/// synchronization.
+using NoThreadsGuard = GuardObject<InitByteNoThreads>;
+
+/// GlobalMutexGuard - Manages initialization using a global mutex and
+/// condition variable.
+template <class Mutex, class CondVar, Mutex& global_mutex, CondVar& global_cond,
+          uint32_t (*GetThreadID)() = PlatformThreadID>
+using GlobalMutexGuard = GuardObject<InitByteGlobalMutex<Mutex, CondVar, global_mutex, global_cond, GetThreadID>>;
+
+/// FutexGuard - Manages initialization using atomics and the futex syscall for
+/// waiting and waking.
+template <void (*Wait)(int*, int) = PlatformFutexWait, void (*Wake)(int*) = PlatformFutexWake,
+          uint32_t (*GetThreadIDArg)() = PlatformThreadID>
+using FutexGuard = GuardObject<InitByteFutex<Wait, Wake, GetThreadIDArg>>;
+
 //===----------------------------------------------------------------------===//
 //
 //===----------------------------------------------------------------------===//
@@ -547,33 +636,27 @@ struct GlobalStatic {
   static T instance;
 };
 template <class T>
-_LIBCPP_SAFE_STATIC T GlobalStatic<T>::instance = {};
+_LIBCPP_CONSTINIT T GlobalStatic<T>::instance = {};
 
-enum class Implementation {
-  NoThreads,
-  GlobalLock,
-  Futex
-};
+enum class Implementation { NoThreads, GlobalMutex, Futex };
 
 template <Implementation Impl>
 struct SelectImplementation;
 
 template <>
 struct SelectImplementation<Implementation::NoThreads> {
-  using type = InitByteNoThreads;
+  using type = NoThreadsGuard;
 };
 
 template <>
-struct SelectImplementation<Implementation::GlobalLock> {
-  using type = InitByteGlobalMutex<
-      LibcppMutex, LibcppCondVar, GlobalStatic<LibcppMutex>::instance,
-      GlobalStatic<LibcppCondVar>::instance, PlatformThreadID>;
+struct SelectImplementation<Implementation::GlobalMutex> {
+  using type = GlobalMutexGuard<LibcppMutex, LibcppCondVar, GlobalStatic<LibcppMutex>::instance,
+                                GlobalStatic<LibcppCondVar>::instance, PlatformThreadID>;
 };
 
 template <>
 struct SelectImplementation<Implementation::Futex> {
-  using type =
-      InitByteFutex<PlatformFutexWait, PlatformFutexWake, PlatformThreadID>;
+  using type = FutexGuard<PlatformFutexWait, PlatformFutexWake, PlatformThreadID>;
 };
 
 // TODO(EricWF): We should prefer the futex implementation when available. But
@@ -584,22 +667,21 @@ constexpr Implementation CurrentImplementation =
 #elif defined(_LIBCXXABI_USE_FUTEX)
     Implementation::Futex;
 #else
-   Implementation::GlobalLock;
+    Implementation::GlobalMutex;
 #endif
 
-static_assert(CurrentImplementation != Implementation::Futex
-           || PlatformSupportsFutex(), "Futex selected but not supported");
+static_assert(CurrentImplementation != Implementation::Futex || PlatformSupportsFutex(),
+              "Futex selected but not supported");
 
-using SelectedImplementation =
-    SelectImplementation<CurrentImplementation>::type;
+using SelectedImplementation = SelectImplementation<CurrentImplementation>::type;
 
 } // end namespace
 } // end namespace __cxxabiv1
 
 #if defined(__clang__)
-# pragma clang diagnostic pop
+#  pragma clang diagnostic pop
 #elif defined(__GNUC__)
-# pragma GCC diagnostic pop
+#  pragma GCC diagnostic pop
 #endif
 
 #endif // LIBCXXABI_SRC_INCLUDE_CXA_GUARD_IMPL_H
diff --git a/gnu/llvm/libcxxabi/src/include/atomic_support.h b/gnu/llvm/libcxxabi/src/include/atomic_support.h
deleted file mode 100644 (file)
index 6f9dbf4..0000000
+++ /dev/null
@@ -1,180 +0,0 @@
-//===----------------------------------------------------------------------===////
-//
-// 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
-//
-//===----------------------------------------------------------------------===////
-
-// FIXME: This file is copied from libcxx/src/include/atomic_support.h. Instead
-// of duplicating the file in libc++abi we should require that the libc++
-// sources are available when building libc++abi.
-
-#ifndef ATOMIC_SUPPORT_H
-#define ATOMIC_SUPPORT_H
-
-#include "__config"
-#include "memory" // for __libcpp_relaxed_load
-
-#if defined(__clang__) && __has_builtin(__atomic_load_n)             \
-                       && __has_builtin(__atomic_store_n)            \
-                       && __has_builtin(__atomic_add_fetch)          \
-                       && __has_builtin(__atomic_exchange_n)         \
-                       && __has_builtin(__atomic_compare_exchange_n) \
-                       && defined(__ATOMIC_RELAXED)                  \
-                       && defined(__ATOMIC_CONSUME)                  \
-                       && defined(__ATOMIC_ACQUIRE)                  \
-                       && defined(__ATOMIC_RELEASE)                  \
-                       && defined(__ATOMIC_ACQ_REL)                  \
-                       && defined(__ATOMIC_SEQ_CST)
-#   define _LIBCXXABI_HAS_ATOMIC_BUILTINS
-#elif !defined(__clang__) && defined(_GNUC_VER) && _GNUC_VER >= 407
-#   define _LIBCXXABI_HAS_ATOMIC_BUILTINS
-#endif
-
-#if !defined(_LIBCXXABI_HAS_ATOMIC_BUILTINS) && !defined(_LIBCXXABI_HAS_NO_THREADS)
-# if defined(_LIBCPP_WARNING)
-    _LIBCPP_WARNING("Building libc++ without __atomic builtins is unsupported")
-# else
-#   warning Building libc++ without __atomic builtins is unsupported
-# endif
-#endif
-
-_LIBCPP_BEGIN_NAMESPACE_STD
-
-namespace {
-
-#if defined(_LIBCXXABI_HAS_ATOMIC_BUILTINS) && !defined(_LIBCXXABI_HAS_NO_THREADS)
-
-enum __libcpp_atomic_order {
-    _AO_Relaxed = __ATOMIC_RELAXED,
-    _AO_Consume = __ATOMIC_CONSUME,
-    _AO_Acquire = __ATOMIC_ACQUIRE,
-    _AO_Release = __ATOMIC_RELEASE,
-    _AO_Acq_Rel = __ATOMIC_ACQ_REL,
-    _AO_Seq     = __ATOMIC_SEQ_CST
-};
-
-template <class _ValueType, class _FromType>
-inline _LIBCPP_INLINE_VISIBILITY
-void __libcpp_atomic_store(_ValueType* __dest, _FromType __val,
-                           int __order = _AO_Seq)
-{
-    __atomic_store_n(__dest, __val, __order);
-}
-
-template <class _ValueType, class _FromType>
-inline _LIBCPP_INLINE_VISIBILITY
-void __libcpp_relaxed_store(_ValueType* __dest, _FromType __val)
-{
-    __atomic_store_n(__dest, __val, _AO_Relaxed);
-}
-
-template <class _ValueType>
-inline _LIBCPP_INLINE_VISIBILITY
-_ValueType __libcpp_atomic_load(_ValueType const* __val,
-                                int __order = _AO_Seq)
-{
-    return __atomic_load_n(__val, __order);
-}
-
-template <class _ValueType, class _AddType>
-inline _LIBCPP_INLINE_VISIBILITY
-_ValueType __libcpp_atomic_add(_ValueType* __val, _AddType __a,
-                               int __order = _AO_Seq)
-{
-    return __atomic_add_fetch(__val, __a, __order);
-}
-
-template <class _ValueType>
-inline _LIBCPP_INLINE_VISIBILITY
-_ValueType __libcpp_atomic_exchange(_ValueType* __target,
-                                    _ValueType __value, int __order = _AO_Seq)
-{
-    return __atomic_exchange_n(__target, __value, __order);
-}
-
-template <class _ValueType>
-inline _LIBCPP_INLINE_VISIBILITY
-bool __libcpp_atomic_compare_exchange(_ValueType* __val,
-    _ValueType* __expected, _ValueType __after,
-    int __success_order = _AO_Seq,
-    int __fail_order = _AO_Seq)
-{
-    return __atomic_compare_exchange_n(__val, __expected, __after, true,
-                                       __success_order, __fail_order);
-}
-
-#else // _LIBCPP_HAS_NO_THREADS
-
-enum __libcpp_atomic_order {
-    _AO_Relaxed,
-    _AO_Consume,
-    _AO_Acquire,
-    _AO_Release,
-    _AO_Acq_Rel,
-    _AO_Seq
-};
-
-template <class _ValueType, class _FromType>
-inline _LIBCPP_INLINE_VISIBILITY
-void __libcpp_atomic_store(_ValueType* __dest, _FromType __val,
-                           int = 0)
-{
-    *__dest = __val;
-}
-
-template <class _ValueType, class _FromType>
-inline _LIBCPP_INLINE_VISIBILITY
-void __libcpp_relaxed_store(_ValueType* __dest, _FromType __val)
-{
-    *__dest = __val;
-}
-
-template <class _ValueType>
-inline _LIBCPP_INLINE_VISIBILITY
-_ValueType __libcpp_atomic_load(_ValueType const* __val,
-                                int = 0)
-{
-    return *__val;
-}
-
-template <class _ValueType, class _AddType>
-inline _LIBCPP_INLINE_VISIBILITY
-_ValueType __libcpp_atomic_add(_ValueType* __val, _AddType __a,
-                               int = 0)
-{
-    return *__val += __a;
-}
-
-template <class _ValueType>
-inline _LIBCPP_INLINE_VISIBILITY
-_ValueType __libcpp_atomic_exchange(_ValueType* __target,
-                                    _ValueType __value, int  = _AO_Seq)
-{
-    _ValueType old = *__target;
-    *__target = __value;
-    return old;
-}
-
-template <class _ValueType>
-inline _LIBCPP_INLINE_VISIBILITY
-bool __libcpp_atomic_compare_exchange(_ValueType* __val,
-    _ValueType* __expected, _ValueType __after,
-    int = 0, int = 0)
-{
-    if (*__val == *__expected) {
-        *__val = __after;
-        return true;
-    }
-    *__expected = *__val;
-    return false;
-}
-
-#endif // _LIBCPP_HAS_NO_THREADS
-
-} // end namespace
-
-_LIBCPP_END_NAMESPACE_STD
-
-#endif // ATOMIC_SUPPORT_H