// 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
      // that add a node to the graph (like TF_FinishOperation).
      if (!g->name_map.insert(std::make_pair(n->name(), n)).second) {
        status->status = tensorflow::errors::Internal(

  EXPECT_EQ(value, 2);
  EXPECT_TRUE(cache2.Lookup("c", &value));
  EXPECT_EQ(value, 3);
  EXPECT_TRUE(cache2.Lookup("d", &value));
  EXPECT_EQ(value, 4);
  // Insertion of "e" causes "a" to be evicted, but the other entries are still
  // there.
  cache2.Insert("e", 5);
  EXPECT_FALSE(cache2.Lookup("a", &value));
  EXPECT_TRUE(cache2.Lookup("b", &value));
  EXPECT_EQ(value, 2);
  EXPECT_TRUE(cache2.Lookup("c", &value));

// op they are called with.
TFE_Op* CloneOp(const TFE_Op* other) {
  TF_Status* status = TF_NewStatus();
  TFE_Context* context = TFE_OpGetContext(other, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  const char* op_name = TFE_OpGetName(other, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

TEST_P(ModularFileSystemTest, TestGetMatchingPaths) {
  const std::vector<std::string> matching_filenames = {
      GetURIForPath("a_file"),
      GetURIForPath("another_file"),
  };
  const std::vector<std::string> other_filenames = {
      GetURIForPath("some_file"),
      GetURIForPath("yet_another_file"),
  };

  for (const auto& filename : matching_filenames) {
    std::unique_ptr<WritableFile> file;

                     char* buffer, TF_Status* status) -> int64_t {
    EXPECT_EQ(n, block_size);
    EXPECT_EQ(offset, 0);
    num_requests++;
    memset(buffer, 'x', n);
    notification.Notify();
    // Wait for other thread to issue read.
    Env::Default()->SleepForMicroseconds(100000);  // 0.1 secs
    TF_SetStatus(status, TF_OK, "");
    return n;
  };
  tf_gcs_filesystem::RamFileBlockCache cache(block_size, block_size, 0,

  EXPECT_EQ(6 * TF_DataTypeSize(TF_UINT64), TF_TensorByteSize(a));
  EXPECT_EQ(6 * TF_DataTypeSize(TF_UINT64), TF_TensorByteSize(b));

  // Check that a write to one tensor shows up in the other.
  *(static_cast<int64_t*>(TF_TensorData(a))) = 4;
  EXPECT_EQ(4, *(static_cast<int64_t*>(TF_TensorData(b))));
  *(static_cast<int64_t*>(TF_TensorData(b))) = 6;
  EXPECT_EQ(6, *(static_cast<int64_t*>(TF_TensorData(a))));

  return absl::OkStatus();
}

TapeTensor::TapeTensor(AbstractTensorHandle* handle) : handle_(handle) {
  handle_->Ref();
}
TapeTensor::TapeTensor(const TapeTensor& other) {
  handle_ = other.handle_;
  handle_->Ref();
}
TapeTensor::~TapeTensor() { handle_->Unref(); }

int64_t TapeTensor::GetID() const { return ToId(handle_); }

tensorflow::DataType TapeTensor::GetDType() const {

  // disable_model_pruning is turned off by default.
  if (tp_configs.disable_model_pruning == TF_TriState_On)
    configs.disable_model_pruning = true;
  else
    configs.disable_model_pruning = false;
  // The other configs are turned on by default.
  CONFIG_TOGGLE(implementation_selector);
  CONFIG_TOGGLE(function_optimization);
  CONFIG_TOGGLE(common_subgraph_elimination);
  CONFIG_TOGGLE(arithmetic_optimization);

        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,
        // but need special handling for gradients (gradient of copy-on is not
        // just copy-off but includes a sum) and consideration of performance.

  if (dev->strict_scope_placement && *requested_placement == '\0') {
    TF_SetStatus(s, TF_INTERNAL,
                 "Ops must be placed on the device explicitly, or their inputs "
                 "first copied to other devices.");
    return;
  }
  TFE_Context* context = TFE_OpGetContext(original_op, s);
  if (TF_GetCode(s) != TF_OK) return;
  const char* operation_name = TFE_OpGetName(original_op, s);