///    block at a higher offset (indicating that the partial block should have
  ///    been a full block).
  /// 3) `TF_OUT_OF_RANGE` if the read from the remote filesystem succeeded, but
  ///    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`).
  ///

#ifndef TENSORFLOW_C_EAGER_TAPE_H_
#define TENSORFLOW_C_EAGER_TAPE_H_

// Language-agnostic gradient tape. Does not perform backpropagation, just
// maintains the data structures required to do so.

#include <stack>
#include <unordered_map>
#include <unordered_set>
#include <vector>

#include "tensorflow/core/config/flag_defs.h"
#include "tensorflow/core/config/flags.h"

//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.
// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

-  char c;
   char *s = s1;
 
   /* Find the end of S1.  */
-  do
-    c = *s1++;
-  while (c != '\0');
-
-  /* Make S1 point before next character, so we can increment
-     it while memory is read (wins on pipelined cpus).  */
-  s1 -= 2;
+  s1 += strlen (s1);
 
-  if (n >= 4)
-    {
-      size_t n4 = n >> 2;
-      do
-	{
-	  c = *s2++;
-	  *++s1 = c;
-	  if (c == '\0')

// ----------------------------------------------------------------------------
namespace tf_gcs_filesystem {
// TODO(vnvo2409): Use partial reponse for better performance.
// TODO(vnvo2409): We could do some cleanups like `return TF_SetStatus`.
// TODO(vnvo2409): Refactor the filesystem implementation when
// https://github.com/googleapis/google-cloud-cpp/issues/4482 is done.

    TF_SetStatus(status, TF_OK, "");
}

static int64_t Tell(const TF_WritableFile* file, TF_Status* status) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);

  // POSIX's `ftell` returns `long`, do a manual cast.
  int64_t position = int64_t{ftell(posix_file->handle)};
  if (position < 0)
    TF_SetStatusFromIOError(status, errno, posix_file->filename);
  else
    TF_SetStatus(status, TF_OK, "");

    return nullptr;
  }
  std::vector<int64_t> dimvec(num_dims);
  for (int i = 0; i < num_dims; ++i) {
    dimvec[i] = static_cast<int64_t>(dims[i]);
  }

  // TODO(apassos) do we need to wrap the deallocator here to make sure to sync
  // the device?
  TF_ManagedBuffer* buf =
      new TF_ManagedBuffer(data, len, deallocator, deallocator_arg,
                           /*owns_memory=*/false);

  EXPECT_EQ(TF_GetCode(status.get()), TF_INVALID_ARGUMENT);
  EXPECT_THAT(TF_Message(status.get()), HasSubstr("assertion failed"));

  // Note that future collectives with the same context do not work at the
  // moment; once canceled, the collective executor requires the program to be
  // restarted / context to be reset.
}

TEST(PARALLEL_DEVICE_LIB, TestDifferentShapes) {

// Returns the function content in a human-readable format, with length set in
// `len`. The format is subject to change in the future.
// The returned string is heap-allocated, and caller should call free() on it.
//
// Do not return const char*, because some foreign language binding
// (e.g. swift) cannot then call free() on the returned pointer.
TF_CAPI_EXPORT extern char* TF_FunctionDebugString(TF_Function* func,

  std::vector<const Node*> control_output_nodes;
  control_output_nodes.reserve(ncontrol_outputs);
  for (int i = 0; i < ncontrol_outputs; ++i) {
    control_output_nodes.push_back(&control_outputs[i]->node);
  }

  // Do the actual function creation.
  DCHECK(append_hash_to_fn_name <= 1);
  tensorflow::FunctionDef fdef;
  status->status = tensorflow::GraphToFunctionDef(
      fn_body->graph, fn_name, append_hash_to_fn_name != 0,