}

  void Run(const std::vector<TF_Output>& run_outputs,
           std::initializer_list<int> input_values) {
    DCHECK_EQ(inputs_.size(), input_values.size());
    std::vector<std::pair<TF_Operation*, TF_Tensor*>> inputs(inputs_.size());
    int i = 0;
    for (int v : input_values) {
      inputs[i] = {inputs_[i].oper, Int32Tensor(v)};
      ++i;
    }

      return mlir::TypeAttr::get(type);
    }
    case AttrValue::kShape:
      return ConvertTensorShapeProto(value.shape(), builder->getContext());
    case AttrValue::kTensor:
      return ConvertTensorProto(value.tensor(), builder);
    case AttrValue::kList: {
      absl::InlinedVector<mlir::Attribute, 8> attrs;
      for (const auto& item : value.list().i())
        attrs.push_back(builder->getI64IntegerAttr(item));

  // Inference with non-broadcastable shape

  );

  TF_DerivedOperandTypeListAttr Tin = TF_DerivedOperandTypeListAttr<0>;
}

def TF_DumpTensorOp : TF_Op<"DumpTensor", []> {
  let summary = "Dump tensor proto.";

  let arguments = (ins
    Arg<TF_Tensor>:$input,

    StrAttr:$log_dir_path,
    StrAttr:$file_name,
    BoolAttr:$enabled,
    StrAttr:$func_name,
    StrAttr:$node_name
  );

  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpAddInput(op, var_handle, status);

  // Convert 'value' to a TF_Tensor then a TFE_TensorHandle.
  std::unique_ptr<TF_Tensor, decltype(&TF_DeleteTensor)> t(
      TF_AllocateTensor(TF_FLOAT, nullptr, 0, sizeof(value)), TF_DeleteTensor);
  memcpy(TF_TensorData(t.get()), &value, TF_TensorByteSize(t.get()));

}

/// Parse a type registered to this dialect.
Type FallbackDialect::parseType(DialectAsmParser &parser) const {
  StringRef keyword;
  if (parser.parseKeyword(&keyword)) return Type();

  if (keyword == "tf_tensor") return TFTensorType::get(getContext());
  if (keyword == "tf_allocator") return TFAllocatorType::get(getContext());

  parser.emitError(parser.getNameLoc(), "unknown type: ") << keyword;
  return Type();
}

    case AttrValue::kType:
      return DataType_Name(attr_value.type());
    case AttrValue::kShape:
      return PrintTensorShape(attr_value.shape());
    case AttrValue::kTensor:
      return PrintTensorProto(attr_value.tensor());
    case AttrValue::kList: {
      string ret = "{";
      if (attr_value.list().s_size() > 0) {
        for (int i = 0; i < attr_value.list().s_size(); ++i) {

          tf_device.return %20#0, %20#1, %20#2, %20#3, %20#4, %20#5, %20#6, %20#7, %20#8, %20#9 : tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>, tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>...

    // CHECK: %[[ADD:.*]] = "tf.Add"(%[[MUL]], %[[MUL]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    // CHECK: return %[[ADD]] : tensor<1xf32>
    %0 = "tf.Mul"(%arg0, %arg0) : (tensor<1xf32>, tensor<1xf32>) -> tensor<*xf32>
    %1 = "tf.Add"(%0, %0) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %1 : tensor<*xf32>
  }

  // CHECK-LABEL: func @simple_chain_with_broadcast

  func.return %0 : tensor<?...