factories.add(factory);
      }
    }
    if (!Modifier.isAbstract(type.getRawType().getModifiers())) {
      for (Constructor<?> constructor : type.getRawType().getDeclaredConstructors()) {
        Invokable<T, T> invokable = type.constructor(constructor);
        if (!invokable.isPrivate() && !invokable.isSynthetic()) {
          factories.add(invokable);
        }
      }
    }

    private int distinctElements;
    private long totalCount;
    private int height;
    private @Nullable AvlNode<E> left;
    private @Nullable AvlNode<E> right;
    /*
     * pred and succ are nullable after construction, but we always call successor() to initialize
     * them immediately thereafter.
     *
     * They may be subsequently nulled out by TreeMultiset.clear(). I think that the only place that

 * instance should be created every time the method is called.
 *
 * <p>For example, the subclass could pass a {@link java.util.TreeMap} during construction, and
 * {@link #createCollection()} could return a {@link java.util.TreeSet}, in which case the
 * multimap's iterators would propagate through the keys and values in sorted order.
 *

            "issue at https://github.com/tensorflow/tensorflow/issues.");
        return;
      }
      // We have a convoluted scheme here: Using the C++ graph construction API
      // to add potentially many nodes to the graph without running the checks
      // (such as uniqueness of the names of nodes) we run with other functions

   * is {@code 60}, and the concurrency level is {@code 8}, then eight segments are created, each
   * having a hash table of size eight. Providing a large enough estimate at construction time
   * avoids the need for expensive resizing operations later, but setting this value unnecessarily
   * high wastes memory.
   *
   * @return this {@code CacheBuilder} instance (for chaining)

	// limit (8KB), we stop growing the source string once the limit
	// is reached and keep reusing the same source string - that
	// should therefore be always resident in the L1 cache - until we
	// have completed the construction of the result.
	// This yields significant speedups (up to +100%) in cases where
	// the result length is large (roughly, over L2 cache size).
	const chunkLimit = 8 * 1024
	chunkMax := n
	if chunkMax > chunkLimit {

   * appears that this pattern is not directly applicable. But we can use the ridiculous, devious
   * trick of throwing an exception in the middle of the construction of another exception. Hotspot
   * is fine with that.
   */

  /**
   * Ensures that {@code index} specifies a valid <i>element</i> in an array, list or string of size

### `tf.keras`:

*   Improvements from the Functional API refactoring:
    *   Functional model construction does not need to maintain a global
        workspace graph, removing memory leaks especially when building many
        models or very large models.
    *   Functional model construction should be ~8-10% faster on average.
    *   Functional models can now contain non-symbolic values in their call

// TODO(jeff,sanjay):
// - export functions to set Config fields

// --------------------------------------------------------------------------
// The new graph construction API, still under development.

// Represents a computation graph.  Graphs may be shared between sessions.
// Graphs are thread-safe when used as directed below.
typedef struct TF_Graph TF_Graph;

  }

  // Inner Classes

  /**
   * Segments are specialized versions of hash tables. This subclass inherits from ReentrantLock
   * opportunistically, just to simplify some locking and avoid separate construction.
   */
  @SuppressWarnings("serial") // This class is never serialized.
  abstract static class Segment<
          K, V, E extends InternalEntry<K, V, E>, S extends Segment<K, V, E, S>>