if (atomic.compareAndSet(oldValue, newValue)) {
          return oldValue;
        }
        // value changed
      }
    }
  }

  /**
   * Copies all of the mappings from the specified map to this map. The effect of this call is
   * equivalent to that of calling {@code put(k, v)} on this map once for each mapping from key
   * {@code k} to value {@code v} in the specified map. The behavior of this operation is undefined

      for (PerListenerQueue<L> queue : listeners) {
        queue.add(event, label);
      }
    }
  }

  /**
   * Dispatches all events enqueued prior to this call, serially and in order, for every listener.
   *
   * <p>Note: this method is idempotent and safe to call from any thread
   */
  public void dispatch() {
    // iterate by index to avoid concurrent modification exceptions

   *       </ul>
   *   <li>another java.util.Map delegate implementation. In most modern JDKs, normal java.util hash
   *       collections intelligently fall back to a binary search tree if hash table collisions are
   *       detected. Rather than going to all the trouble of reimplementing this ourselves, we
   *       simply switch over to use the JDK implementation wholesale if probable hash flooding is

   * skipped element by immediately calling {@code remove()} on the returned fluent iterable's
   * iterator, as the {@code Iterator} contract states that a call to {@code * remove()} before a
   * call to {@code next()} will throw an {@link IllegalStateException}.
   *
   * <p><b>{@code Stream} equivalent:</b> {@link Stream#skip} (same).
   */
  public final FluentIterable<E> skip(int numberToSkip) {

 * value and the old value.
 *
 * <p>Keys and values may be null. All optional multimap methods are supported, and all returned
 * views are modifiable.
 *
 * <p>The lists returned by {@link #get}, {@link #removeAll}, and {@link #replaceValues} all
 * implement {@link java.util.RandomAccess}.
 *

   *       </ul>
   *   <li>another java.util.Set delegate implementation. In most modern JDKs, normal java.util hash
   *       collections intelligently fall back to a binary search tree if hash table collisions are
   *       detected. Rather than going to all the trouble of reimplementing this ourselves, we
   *       simply switch over to use the JDK implementation wholesale if probable hash flooding is

   *
   * <p>A subcollection, such as a sublist, contains some of the values for a given key. Its
   * ancestor field points to the full wrapped collection with all values for the key. The
   * subcollection {@code refreshIfEmpty}, {@code removeIfEmpty}, and {@code addToMap} methods call
   * the corresponding methods of the full wrapped collection.
   */
  @WeakOuter
  class WrappedCollection extends AbstractCollection<V> {

   * specified element (or equivalently, if this queue changed as a result of the call).
   *
   * @param o element to be removed from this queue, if present
   * @return {@code true} if this queue changed as a result of the call
   */
  @CanIgnoreReturnValue // pushed down from class to method
  @Override
  public boolean remove(@Nullable Object o) {

                 * the map to change between that call and this one. But if that happens, the
                 * behavior is undefined because of the concurrent mutation.)
                 *
                 * (Our prototype checker happens to be "smart" enough to understand this for the
                 * *get* call in getValue but not for the *put* call here.)
                 *

  // via delegate()); it seems inappropriate to throw ISE on this method.
  @Override
  public int hashCode() {
    return delegate.hashCode();
  }

  private void validate() {
    // Don't use checkState(), because we don't want the overhead of generating the error message
    // unless it's actually going to be used; validate() is called for all set method calls, so it
    // needs to be fast.