// Was the dependence to be deleted the first/last (or both) use of this
  // tensor?
  const bool was_first_use =
      (!tensor.operations.empty() && ioperation == tensor.first_use());
  const bool was_last_use =
      (!tensor.operations.empty() && ioperation == tensor.last_use());
  Erase(ioperation, tensor.operations);
  Erase(itensor, operation.tensors);
  if (was_first_use) {
    operation.alloc -= size;
    if (!was_last_use) {

   * instance, so it can be invoked again to create multiple independent maps.
   *
   * <p>The bulk operations {@code putAll}, {@code equals}, and {@code clear} are not guaranteed to
   * be performed atomically on the returned map. Additionally, {@code size} and {@code
   * containsValue} are implemented as bulk read operations, and thus may fail to observe concurrent
   * writes.
   *

 * Table} interface.
 *
 * <p>The views returned by {@link #column}, {@link #columnKeySet()}, and {@link #columnMap()} have
 * iterators that don't support {@code remove()}. Otherwise, all optional operations are supported.
 * Null row keys, columns keys, and values are not supported.
 *
 * <p>Lookups by row key are often faster than lookups by column key, because the data is stored in

 * amortized since some operations can trigger a hash table resize.
 *
 * <p>Unlike {@code java.util.HashSet}, iteration is only proportional to the actual {@code size()},
 * which is optimal, and <i>not</i> the size of the internal hashtable, which could be much larger
 * than {@code size()}. Furthermore, this structure only depends on a fixed number of arrays; {@code

   * cartesian product is constructed, the input lists are merely copied. Only as the resulting list
   * is iterated are the individual lists created, and these are not retained after iteration.
   *
   * @param lists the lists to choose elements from, in the order that the elements chosen from
   *     those lists should appear in the resulting lists

   * iterator that traverses the elements of each iterable in {@code inputs}. The input iterators
   * are not polled until necessary.
   *
   * <p>The returned iterable's iterator supports {@code remove()} when the corresponding input
   * iterator supports it. The methods of the returned iterable may throw {@code
   * NullPointerException} if any of the input iterators is {@code null}.
   *

   * iterator that traverses the elements of each iterable in {@code inputs}. The input iterators
   * are not polled until necessary.
   *
   * <p>The returned iterable's iterator supports {@code remove()} when the corresponding input
   * iterator supports it. The methods of the returned iterable may throw {@code
   * NullPointerException} if any of the input iterators is {@code null}.
   *

 * need for a distinct {@code ImmutableBiMultimap} type.
 *
 * <p><a id="iteration"></a>
 *
 * <p><b>Key-grouped iteration.</b> All view collections follow the same iteration order. In all
 * current implementations, the iteration order always keeps multiple entries with the same key
 * together. Any creation method that would customarily respect insertion order (such as {@link

 * amortized since some operations can trigger a hash table resize.
 *
 * <p>Unlike {@code java.util.HashSet}, iteration is only proportional to the actual {@code size()},
 * which is optimal, and <i>not</i> the size of the internal hashtable, which could be much larger
 * than {@code size()}. Furthermore, this structure only depends on a fixed number of arrays; {@code

 * need for a distinct {@code ImmutableBiMultimap} type.
 *
 * <p><a id="iteration"></a>
 *
 * <p><b>Key-grouped iteration.</b> All view collections follow the same iteration order. In all
 * current implementations, the iteration order always keeps multiple entries with the same key
 * together. Any creation method that would customarily respect insertion order (such as {@link