ensure task completion happens-before await resumption

It should be the case that non-atomic writes to memory
performed by task A are observable by any task B who
awaits A's completion.

Author: kavon

stdlib/public/Concurrency/Task.cpp

@@ -438,6 +438,11 @@ static void completeTaskImpl(AsyncTask *task,
                              SwiftError *error) {
   assert(task && "completing task, but there is no active task registered");
 
+  // This release-fence synchronizes with the acquire-fences performed
+  // by any tasks that resume after awaiting a task's completion. It ensures
+  // non-atomic writes done by this task are observable by those tasks.
+  std::atomic_thread_fence(std::memory_order_release);
+
   // Store the error result.
   auto asyncContextPrefix = reinterpret_cast<AsyncContextPrefix *>(
       reinterpret_cast<char *>(context) - sizeof(AsyncContextPrefix));
@@ -540,6 +545,8 @@ static void future_adapter(SWIFT_ASYNC_CONTEXT AsyncContext *_context) {
 
 SWIFT_CC(swiftasync)
 static void task_wait_throwing_resume_adapter(SWIFT_ASYNC_CONTEXT AsyncContext *_context) {
+  // This acquire synchronizes with the release in `completeTaskImpl`.
+  std::atomic_thread_fence(std::memory_order_acquire);
 
   auto context = static_cast<TaskFutureWaitAsyncContext *>(_context);
   auto resumeWithError =
@@ -550,6 +557,9 @@ static void task_wait_throwing_resume_adapter(SWIFT_ASYNC_CONTEXT AsyncContext *
 SWIFT_CC(swiftasync)
 static void
 task_future_wait_resume_adapter(SWIFT_ASYNC_CONTEXT AsyncContext *_context) {
+  // This acquire synchronizes with the release in `completeTaskImpl`.
+  std::atomic_thread_fence(std::memory_order_acquire);
+
   return _context->ResumeParent(_context->Parent);
 }