%2 = "tf.Add"(%1, %1) {
    _xla_compile_device_type = "TPU", _replication_info = "cluster",
    device = "/task:0/device:TPU:0", dtype = f32
  } : (tensor<1x80xf32>, tensor<1x80xf32>) -> tensor<1x80xf32>
  %3 = "tf.ResourceGatherNd"(%arg0, %0) {
    Tindices = i32
  } : (tensor<*x!tf_type.resource<tensor<80xf32>>>, tensor<i32>) -> tensor<1x80xf32>

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir

  %0 = "mhlo.reshape"(%arg0) : (tensor<1x1x512xf32>) -> tensor<1x512xf32>
  %1 = "mhlo.dot"(%0, %arg1) : (tensor<1x512xf32>, tensor<512x13x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x13xf32>
  %2 = "mhlo.reshape"(%1) : (tensor<1x13xf32>) -> tensor<1x1x13xf32>
  func.return %2 : tensor<1x1x13xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<

  %0 = "tfl.fully_connected"(%arg0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<2x2x10xf32>, tensor<40x40xf32>, none) -> tensor<1x40xf32>
  func.return %0 : tensor<1x40xf32>
}

// -----

func.func @testFullyConnectedWith3DFilter(%arg0: tensor<1x37xf32>, %arg1: tensor<40x2x37xf32>, %arg2: tensor<40xf32>) -> tensor<1x40xf32> {

  %2 = "tf.Transpose"(%1, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %2 : tensor<2x1xf32>

// CHECK: %cst = arith.constant
// CHECK: %[[trans:.*]] = "tf.Transpose"
// CHECK-SAME: -> tensor<2x1xf32>
// CHECK: %[[q:.*]] = "tfl.quantize"(%[[trans]]) <{qtype = tensor<2x1x!quant.uniform<u8:f32, 1.000000e+00>>}>

func.func @torch_index_select(%arg0: tensor<2x1xf32>, %arg1: tensor<2xi32>) -> tensor<2x1xf32> {
  %0 = "mhlo.torch_index_select"(%arg0, %arg1) {
    batch_dims = 0 : i64, dim = 0 : i64
  } : (tensor<2x1xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %0 : tensor<2x1xf32>
}

// CHECK-LABEL:   func @lowered_cumsum(
// CHECK-SAME:      %[[VAL_0:.*]]: tensor<4x12xf32>) -> tensor<4x12xf32> {

  // CHECK: return %0
}

// CHECK-LABEL: testAddOfNegRight
func.func @testAddOfNegRight(%arg0: tensor<8x16xf32>, %arg1: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.Neg"(%arg1) : (tensor<8x16xf32>) -> tensor<8x16xf32>
  %1 = "tf.Add"(%arg0, %0) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<8x16xf32>, tensor<8x16xf32>) -> tensor<8x16xf32>

//                   (tensor<4x1xf32>, tensor<4x2xf32>, tensor<4x3xf32>)
//
// We will generate slices following slices:
// %0 = "mhlo.slice"(%input) {
//        limit_indices = dense<[4, 1]> : tensor<2xi64>,
//        start_indices = dense<0> : tensor<2xi64>,
//        strides = dense<1> : tensor<2xi64>} :
//        (tensor<4x6xf32>) -> tensor<4x1xf32>
// %1 = "mhlo.slice"(%input) {

mlir_module = '''python
func @main(%arg0 : tensor<10xf32>, %arg1 : tensor<10xf32>) -> tensor<10x10xf32> {
   %add = "magic.op"(%arg0, %arg1) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10x10xf32>
   return %ret : tensor<10x10xf32>
}
'''

@tf.function
def foo(x, y):
  return mlir_passthrough_op([x, y], mlir_module, Toutputs=[tf.float32])