Fix data race with synthetic delay initialization

The synthetic delay macro initializes a local atomic variable to point
to a delay implementation class. Because initializing the class involves
writing data to memory which is subsequently deferences through the
atomic pointer, we need to use memory barries when updating the pointer
to make sure the data behind it is committed.

Thanks to Alexander Potapenko for spotting this.

Review URL: https://codereview.chromium.org/392933002

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@284203 0039d316-1c4b-4281-b951-d872f2087c98

base/debug/trace_event_synthetic_delay.cc

@@ -220,11 +220,11 @@ base::debug::TraceEventSyntheticDelay* GetOrCreateDelay(
     base::subtle::AtomicWord* impl_ptr) {
   base::debug::TraceEventSyntheticDelay* delay_impl =
       reinterpret_cast<base::debug::TraceEventSyntheticDelay*>(
-          base::subtle::NoBarrier_Load(impl_ptr));
+          base::subtle::Acquire_Load(impl_ptr));
   if (!delay_impl) {
     delay_impl = base::debug::TraceEventSyntheticDelayRegistry::GetInstance()
                      ->GetOrCreateDelay(name);
-    base::subtle::NoBarrier_Store(
+    base::subtle::Release_Store(
         impl_ptr, reinterpret_cast<base::subtle::AtomicWord>(delay_impl));
   }
   return delay_impl;