Fix base::internal::Singleton to use acquire loads instead of nobarrier ones.

base/memory/singleton.cc

 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "base/memory/singleton.h"
 #include "base/threading/platform_thread.h"
 
 namespace base {
 namespace internal {
 
 subtle::AtomicWord WaitForInstance(subtle::AtomicWord* instance) {
   // Handle the race. Another thread beat us and either:
   // - Has the object in BeingCreated state
   // - Already has the object created...
   // We know value != NULL.  It could be kBeingCreatedMarker, or a valid ptr.
   // Unless your constructor can be very time consuming, it is very unlikely
   // to hit this race.  When it does, we just spin and yield the thread until
   // the object has been created.
   subtle::AtomicWord value;
   while (true) {
-    value = subtle::NoBarrier_Load(instance);
+    // The load has acquire memory ordering as the thread which reads the
+    // instance pointer must acquire visibility over the associated data.
+    // The pairing Release_Store operation is in Singleton::get().
+    value = subtle::Acquire_Load(instance);
     if (value != kBeingCreatedMarker)
       break;
     PlatformThread::YieldCurrentThread();
   }
   return value;
 }
 
 }  // namespace internal
 }  // namespace base