| ((source[offset + 1] & 0xFF) << 8)
        | ((source[offset + 2] & 0xFF) << 16)
        | ((source[offset + 3] & 0xFF) << 24);
  }

  /**
   * Indicates that the load and store operations will be very efficient because of use of VarHandle
   * or Unsafe. May be useful for calling code to fall back on an alternative implementation that is
   * slower than those implementations but faster than the pure-Java mask-and-shift.
   */

  /** The concurrency level. */
  final int concurrencyLevel;

  /** Strategy for comparing keys. */
  final Equivalence<Object> keyEquivalence;

  /** Strategy for handling entries and segments in a type-safe and efficient manner. */
  final transient InternalEntryHelper<K, V, E, S> entryHelper;

  /**
   * Creates a new, empty map with the specified strategy, initial capacity and concurrency level.
   */

  /** The concurrency level. */
  final int concurrencyLevel;

  /** Strategy for comparing keys. */
  final Equivalence<Object> keyEquivalence;

  /** Strategy for handling entries and segments in a type-safe and efficient manner. */
  final transient InternalEntryHelper<K, V, E, S> entryHelper;

  /**
   * Creates a new, empty map with the specified strategy, initial capacity and concurrency level.
   */

    `tf.contrib.data.SqlDataset`.
*   Distributed Mutex / CriticalSection added to
    `tf.contrib.framework.CriticalSection`.
*   Better text processing with `tf.regex_replace`.
*   Easy, efficient sequence input with
    `tf.contrib.data.bucket_by_sequence_length`
*   Initial support for `tf.contrib.tensorrt` that enables native TensorRT in
    TensorFlow.

## Bug Fixes and Other Changes

in mind. It is strongly typed and garbage-collected and has explicit
support for concurrent programming.  Programs are constructed from
<i>packages</i>, whose properties allow efficient management of
dependencies.
</p>

<p>
The syntax is compact and simple to parse, allowing for easy analysis
by automatic tools such as integrated development environments.
</p>

   * en]}.
   *
   * <p><b>{@code Stream} equivalent:</b> {@code stream.collect(Collectors.joining(", ", "[", "]"))}
   * or (less efficiently) {@code stream.collect(Collectors.toList()).toString()}.
   */
  @Override
  public String toString() {
    return Iterables.toString(getDelegate());
  }

  /**

   * en]}.
   *
   * <p><b>{@code Stream} equivalent:</b> {@code stream.collect(Collectors.joining(", ", "[", "]"))}
   * or (less efficiently) {@code stream.collect(Collectors.toList()).toString()}.
   */
  @Override
  public String toString() {
    return Iterables.toString(getDelegate());
  }

  /**

        // a cyclic edge: it's possible for two threads to simultaneous find
        // "safe" edges which together form a cycle. Preventing this race
        // condition efficiently without _introducing_ deadlock is probably
        // tricky. For now, just accept the race condition---missing a warning
        // now and then is still better than having no deadlock detection.

      HashMap<K, V> newHashMapWithExpectedSize(int expectedSize) {
    return new HashMap<>(capacity(expectedSize));
  }

  /**
   * Returns a capacity that is sufficient to keep the map from being resized as long as it grows no
   * larger than expectedSize and the load factor is ≥ its default (0.75).
   */
  static int capacity(int expectedSize) {
    if (expectedSize < 3) {

      HashMap<K, V> newHashMapWithExpectedSize(int expectedSize) {
    return new HashMap<>(capacity(expectedSize));
  }

  /**
   * Returns a capacity that is sufficient to keep the map from being resized as long as it grows no
   * larger than expectedSize and the load factor is ≥ its default (0.75).
   */
  static int capacity(int expectedSize) {
    if (expectedSize < 3) {