从 C++ 单例 pthread_once 到 double-check、volatile、memory barrier

pthread_once 确保一个被注册的子函数至多被执行一次：

#include <pthread.h>
#include <cassert>

using namespace std;

class Singleton
{
public:
  static
    Singleton *instance();
private:
  static
    Singleton *instance_;
  static
    pthread_once_t once;
  static
    void _instance();
  Singleton();
};

Singleton::Singleton()
{
};

Singleton *Singleton::instance_ = NULL;

void Singleton::_instance()
{
  assert(NULL == instance_);
  instance_ = new Singleton();
}

pthread_once_t Singleton::once = PTHREAD_ONCE_INIT;

Singleton *Singleton::instance()
{
  pthread_once(&once, Singleton::_instance);
  return instance_;
}

下面这段代码，看起来非常和谐的 double-check，what problem?

pthread_mutex_t singleton_mutex = PTHREAD_MUTEX_INITIALIZER;

Singleton *Singleton::instance2()
{
  if (NULL == instance_) {
    pthread_mutex