// nodes, keep_ops, init_ops. Fills in `values` and `lengths`, each of which
// must point to an array of length at least `num_values`.
//
// The elements of values will point to addresses in `storage` which must be at
// least `storage_size` bytes in length.  `num_values` and `storage` can be
// obtained from TF_GetNodesToPreserveSize
//

// in the memory of the device name 'device_name'.
// If source and destination are the same device, then this creates a new handle
// that shares the underlying buffer. Otherwise, it currently requires at least
// one of the source or destination devices to be CPU (i.e., for the source or
// destination tensor to be placed in host memory).
// If async execution is enabled, the copy may be enqueued and the call will

    to the core.
*   As every PR requires several CPU/GPU hours of CI testing, we discourage
    submitting PRs to fix one typo, one warning,etc. We recommend fixing the
    same issue at the file level at least (e.g.: fix all typos in a file, fix
    all compiler warnings in a file, etc.)
*   Tests should follow the
    [testing best practices](https://www.tensorflow.org/community/contribute/tests)
    guide.

#### License

  ///    the file contents do not extend past `offset` and thus nothing was
  ///    placed in `out`.
  /// 4) `TF_OK` otherwise (i.e. the read succeeded, and at least one byte was
  /// placed
  ///    in `buffer`).
  ///
  /// Caller is responsible for allocating memory for `buffer`.
  /// `buffer` will be left unchanged in case of errors.

          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,

  // watched inputs.
  void Watch(int64_t tensor_id);

  // Records an operation with inputs `input_tensor_id` and outputs
  // `output_tensors` on the tape and marks all its outputs as watched if at
  // least one input of the op is watched and has trainable dtype.
  //
  // op_type is used to decide which of the incoming gradients can be left as
  // nullptr instead of building zeros when build_default_zeros_grads == true.

-		  : memset (s, c, n)))
-#  endif
-# endif
-
-/* GCC < 3.0 optimizes memset(s, 0, n) but not bzero(s, n).
-   The optimization is broken before EGCS 1.1.
-   GCC 3.0+ has __builtin_bzero as well, but at least till GCC 3.4
-   if it decides to call the library function, it calls memset
-   and not bzero.  */
-# if __GNUC_PREREQ (2, 91)
-#  define __bzero(s, n) __builtin_memset (s, '\0', n)
-# endif
-

    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {
    // Special-cased operation for packing per-device tensors into one parallel
    // tensor.

  tensorflow::condition_variable finished_join_;

  // Temporary state between `StartExecute` and `Join`.
  //
  //   Inputs; pointers are to objects not owned by the DeviceThread, but which
  //   are expected to live at least until `Join` finishes:
  TFE_Context* context_ TF_GUARDED_BY(execution_mutex_);
  const char* operation_name_ TF_GUARDED_BY(execution_mutex_);
  absl::optional<int64_t> step_id_ TF_GUARDED_BY(execution_mutex_) =