Fix performance regressions due to lazy initialization.

src/platform.h

Index: src/platform.h
diff --git a/src/platform.h b/src/platform.h
index 4ec6057c5df01d563e39d3b8a07457d5bc8e8d96..3b2aa3c8bafdf7c2a11ce1190769ef9eeac168b6 100644
--- a/src/platform.h
+++ b/src/platform.h
@@ -119,6 +119,10 @@ class OS {
   // Initializes the platform OS support. Called once at VM startup.
   static void SetUp();
 
+  // Initializes the platform OS support that depend on CPU features. This is
+  // called after CPU initialization.
+  static void PostSetUp();
+
   // Returns the accumulated user time for thread. This routine
   // can be used for profiling. The implementation should
   // strive for high-precision timer resolution, preferable
@@ -545,7 +549,8 @@ struct CreateMutexTrait {
 //     // Do something.
 //   }
 //
-typedef LazyDynamicInstance<Mutex, CreateMutexTrait>::type LazyMutex;
+typedef LazyDynamicInstance<
+    Mutex, CreateMutexTrait, ThreadSafeInitOnceTrait>::type LazyMutex;
 
 #define LAZY_MUTEX_INITIALIZER LAZY_DYNAMIC_INSTANCE_INITIALIZER
 
@@ -616,7 +621,8 @@ struct CreateSemaphoreTrait {
 template <int InitialValue>
 struct LazySemaphore {
   typedef typename LazyDynamicInstance<
-      Semaphore, CreateSemaphoreTrait<InitialValue> >::type type;
+      Semaphore, CreateSemaphoreTrait<InitialValue>,
+      ThreadSafeInitOnceTrait>::type type;
 };
 
 #define LAZY_SEMAPHORE_INITIALIZER LAZY_DYNAMIC_INSTANCE_INITIALIZER