// CHECK: "tf.TPUExecuteAndUpdateVariables"(%[[ID]], %[[R1]], %[[COMPILE]]#1)
            "tf_device.launch"() ({
              "tf.TPUExecuteAndUpdateVariables"(%id, %arg31, %compile#1)
                    {device_var_reads_indices = [0, 1], device_var_updates_indices = [0, 1]}
                      : (tensor<*x!tf_type.resource<tensor<f32>>>, tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>>, tensor<2x!tf_type.string>) -> ()
              tf_device.return

  llvm::SmallVector<int64_t, 8> device_var_reads_indices;
  llvm::SmallVector<int64_t, 8> device_var_updates_indices;
  for (auto resource : var_access_info.resources_read) {
    auto info_it = var_access_info.per_resource_info.find(resource);
    if (info_it == var_access_info.per_resource_info.end()) continue;
    device_var_reads_indices.push_back(info_it->second.execute_input_index);

  // CHECK-SAME: <{device = "/job:localhost/replica:0/task:0/device:TPU:0"}>
  // CHECK-NEXT: "tf.TPUExecuteAndUpdateVariables"(%[[ID_0]], %[[ARG_1]], %[[READ_2]], %[[COMPILE]]#1)
  // CHECK-SAME: device_var_reads_indices = [0, 1],
  // CHECK-SAME: device_var_updates_indices = [0, -1]
  %execute:2 = "tf_device.launch"() ({
    %0:2 = "tf.TPUExecute"(%read0, %read1, %read2, %compile#1) {

      %3 = "tf.AssignVariableOp"(%arg0, %2)
    ```

    will be transformed into

    ```mlir
      %2 = "tf.TPUExecuteAndUpdateVariables"(%arg0, %arg1, %compile)
        { device_var_reads_indices = [0, 1],
          device_var_updates_indices = [0, -1] }
    ````

    The transformation happens only for on-device variables. The above