*/
  void cleanUp() {
    if (threadStarted) {
      return;
    }

    Reference<?> reference;
    while ((reference = queue.poll()) != null) {
      /*
       * This is for the benefit of phantom references. Weak and soft references will have already
       * been cleared by this point.
       */
      reference.clear();
      try {
        ((FinalizableReference) reference).finalizeReferent();

   */
  void cleanUp() {
    if (threadStarted) {
      return;
    }

    Reference<?> reference;
    while ((reference = queue.poll()) != null) {
      /*
       * This is for the benefit of phantom references. Weak and soft references will have already
       * been cleared by this point.
       */
      reference.clear();
      try {
        ((FinalizableReference) reference).finalizeReferent();

   * world. Figure out what sort of space-time tradeoff we're actually going to get here with the
   * *Map variants. This class is particularly hard to benchmark, because the benefit is not only in
   * less allocation, but also having the GC do less work to scan the heap because of fewer
   * references, which is particularly hard to quantify.
   */

 * one advantage is that we can introduce two separate {@code Condition} objects, which allows us to
 * use {@code signal()} instead of {@code signalAll()}, which may be a performance benefit.
 *
 * <pre>{@code
 * public class SafeBox<V> {
 *   private V value;
 *   private final ReentrantLock lock = new ReentrantLock();
 *   private final Condition valuePresent = lock.newCondition();

   * we've chosen to instead use a signature that benefits existing users -- and is still safe.
   *
   * For convertAll, I haven't looked as closely at *how* much existing users benefit, so we should
   * keep an eye out for problems that new users encounter. Note also that convertAll could support
   * both use cases by using @PolyNull. (By contrast, we can't use @PolyNull for our superinterface

    // the monitor with the guard satisfied. Therefore as soon as tearDownLatch is triggered, we
    // should be able to enter the monitor, and then we set the guard to satisfied for the benefit
    // of any remaining waiting threads.

    tearDownLatch.countDown();
    assertTrue(
        "Monitor still occupied in tearDown()",
        monitor.enter(UNEXPECTED_HANG_DELAY_MILLIS, TimeUnit.MILLISECONDS));

   * not required to be a {@link BufferedInputStream} in order to allow implementations to simply
   * delegate to {@link #openStream()} when the stream returned by that method does not benefit from
   * additional buffering (for example, a {@code ByteArrayInputStream}). This method returns a new,
   * independent stream each time it is called.
   *

    // the monitor with the guard satisfied. Therefore as soon as tearDownLatch is triggered, we
    // should be able to enter the monitor, and then we set the guard to satisfied for the benefit
    // of any remaining waiting threads.

    tearDownLatch.countDown();
    assertTrue(
        "Monitor still occupied in tearDown()",
        monitor.enter(UNEXPECTED_HANG_DELAY_MILLIS, TimeUnit.MILLISECONDS));

        // run() on the same thread that called execute() - such as in a thread pool, and think
        // that it was happening inline. As a side benefit, avoids holding on to the Thread object
        // longer than necessary.
        submitting = null;
      }
    }

    @SuppressWarnings("ShortCircuitBoolean")
    @Override
    public void run() {

    /*
     * Padding from 40 into 64 bytes, same size as cache line. Might be beneficial to add a fourth
     * long here, to minimize chance of interference between consecutive locks, but I couldn't
     * observe any benefit from that.
     */
    long unused1;
    long unused2;
    long unused3;

    PaddedLock() {
      super(false);
    }
  }

  private static class PaddedSemaphore extends Semaphore {