SmallVector<int64_t> output_shape(input_shape.getShape().begin(),
                                    input_shape.getShape().end());
  for (int i : spatial_dims) {
    output_shape[i] += padding_values[2 * i] + padding_values[2 * i + 1];
  }

  return builder.create<TF::PadV2Op>(
      loc, RankedTensorType::get(output_shape, builder.getI8Type()), input,
      temp_padding,

    const mlir::TensorType cluster_func_output_type,
    const xla::OpSharding& output_sharding,
    mlir::Type* tiled_logical_computation_type) {
  const auto output_shape = cluster_func_output_type.getShape();
  auto new_output_shape = llvm::to_vector<4>(output_shape);
  auto dimension_to_splits_map =
      GetDimensionIndicesAndNumSplitsFromSharding(output_sharding);
  if (!dimension_to_splits_map.ok()) {

  %17 = "tf.If"(%2, %2, %13, %13) {_lower_using_switch_merge = true, _read_only_resource_inputs = [], device = "", else_branch = @WhitespaceTokenize_RaggedConcat_assert_equal_1_Assert_AssertGuard_false_3210, is_stateless = false, output_shapes = [#tf_type.shape<>], then_branch = @WhitespaceTokenize_RaggedConcat_assert_equal_1_Assert_AssertGuard_true_3200} : (tensor<i1>, tensor<i1>, tensor<i64>, tensor<i64>) -> tensor<i1>

  int64_t out_row_dim = output_shape[output_shape.size() - 2];
  int64_t out_col_dim = output_shape[output_shape.size() - 1];

  int64_t expected_out_row_dim = op.getAdjX() ? x_col_dim : x_row_dim;
  int64_t expected_out_col_dim = op.getAdjY() ? y_row_dim : y_col_dim;

  if (expected_out_row_dim != ShapedType::kDynamic &&
      out_row_dim != ShapedType::kDynamic &&
      out_row_dim != expected_out_row_dim)

// Converts a shape from MLIR to a TensorFlow tensor shape proto.
void ConvertToTensorShapeProto(llvm::ArrayRef<int64_t> shape,
                               TensorShapeProto* output_shape);

// Converts an MLIR type to a TensorFlow tensor shape.
PartialTensorShape ConvertTypeToTensorShape(const mlir::Type& type);

// Converts an MLIR shaped type to a TensorFlow shape attribute.

    value {
      func {
        name: "body"
      }
    }
  }
  attr {
    key: "cond"
    value {
      func {
        name: "cond"
      }
    }
  }
  attr {
    key: "output_shapes"
    value {
      list {
        shape {
        }
        shape {
          unknown_rank: true
        }
      }
    }
  }
  experimental_debug_info {
  }
}
node {

  }

  // Gather shapes for output.
  for (auto v : ddn.lhs_batch_dimensions()) {
    output_shape.push_back(lhs_shape[v]);
  }

  // Batch dimension is gathered from the right side.
  if (output_shape.empty()) {
    for (auto v : ddn.rhs_batch_dimensions()) {
      output_shape.push_back(rhs_shape[v]);
    }
  }

  // Gather remaining dimensions.

            saved_concrete_function.output_signature().tensor_spec_value();
        PartialTensorShape output_shape = output_tensor_spec_proto.shape();
        DataType output_dtype = output_tensor_spec_proto.dtype();
        TaggedValue output_tensor_spec(output_shape, output_dtype);

        // Register the function trace.
        //

  %cst = "tf.Const"() <{value = dense<1> : tensor<i32>}> {device = ""} : () -> tensor<i32>
  %0 = "tf.MultiDeviceIteratorFromStringHandle"(%arg0) <{output_shapes = [#tf_type.shape<>], output_types = [!tf_type.string]}> {device = ""} : (tensor<!tf_type.string>) -> tensor<!tf_type.resource>

      device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0",
      finalize_func = @finalize,
      init_func = @init,
      next_func = @next,
      operandSegmentSizes = array<i32: 1, 2, 1>,
      output_shapes = [#tf_type.shape<>],
      output_types = [!tf_type.string],
      metadata = ""} : (
              tensor<4xf32>,
              tensor<3xf32>,
              tensor<!tf_type.resource>,