class FreeQueue
{
  const static int BUFFERS = 2;
  const static int SIZE = 128;
  FixedVector<void*,SIZE> ptrs[BUFFERS];
  int currentQueue;
  Atomic ownedByJob[BUFFERS];
public:
  FreeQueue() : currentQueue() {}

  void Push(void* ptr)
  {
    while(1)
    {
      if( currentQueue==-1 ) // worst case, all buffers are full and jobs haven't completed yet
      {
        ThreadPool::WaitUntil( IsFalse(ownedByJob, BUFFERS) ); // go idle and/or run available jobs
        for( int i=0; i!=BUFFERS && currentQueue==-1; ++i )
          if( ownedByJob[i] == 0 )
            currentQueue = i;
      }
      assert( currentQueue!=-1 );
      if( ptrs[currentQueue]->Push(ptr) )
        return;//actualy pushed the ptr into a buffer
      else//failed to push, this buffer is full
      {
        //mark this buffer as being owned by a job, kick of the async call to actually free the pointers in that buffer
        ownedByJob[currentQueue] = 1;
        ThreadPool::PushJob( ActuallyCallFree(ptrs[currentQueue], ownedByJob[currentQueue]) );
        // ActuallyCallFree does: foreach ptr in arg0: _real_free_(ptr); arg0.clear(); arg1 = 0;

        currentQueue = (currentQueue+1) % BUFFERS;//try to move to the next buffer
        if( ownedByJob[currentQueue] )// next buffer is still waiting for it's job to complete, try find another one
        {
          currentQueue = -1;
          for( int i=0; i!=BUFFERS && currentQueue==-1; ++i )
            if( ownedByJob[i] == 0 )
              currentQueue = i;
        }
      }
    }
  }
};

static ThreadLocal<FreeQueue> g_toFree;

void free(void* p)
{
  g_toFree->push(p);
}