Sunny> Look what have I found! Consider a C#:

public class T
{
  public T free;
}

public void NewTest()
{
  T cache = new T();

  Stopwatch timer = new Stopwatch();

  timer.Reset();
  timer.Start();

  for(int i = 0; i < 10000000; ++i)
  {
    // Get from cache.
    T t;

    if (cache.free == null)
    {
      cache.free = new T();
    }

    t = cache.free;

    // Release
    cache.free = t;
    t = null;
  }

  timer.Stop();

  long cacheTicks = timer.ElapsedTicks;

  timer.Reset();
  timer.Start();

  for(int i = 0; i < 10000000; ++i)
  {
    new T();
  }

  timer.Stop();

  long newTicks = timer.ElapsedTicks;

  Console.WriteLine("cache: {0}, new: {1}", cacheTicks, newTicks);
}

Gloomy> And?

Sunny> Tests show that new T() is almost as fast as caching! GC's "new" probably has a fast route, where it shifts free memory border in an atomic way, thus allocation takes just several cycles.

Gloomy> Well, you're probably right, there is a fast route. I, however, have a different opinion. To track references, a generational garbage collector implements field assign as a call rather than a mov. This routine, except move itself, marks touched memory page in a special card table (who said GC is cheap?); thus, I think, a reference field setter is almost as slow as the "new" call.