%1 = "tf.ReadVariableOp"(%0) {device = "/CPU:0"} : (tensor<!tf_type.resource<tensor<1x1xf32>>>) -> tensor<1x1xf32>
  %2:2 = "tf.Split"(%cst, %arg0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x4xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)
  %3:2 = "tf.Split"(%cst, %2#0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x2xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>)

    return %0#0, %0#1, %0#2, %0#3, %0#4 : tensor<?xi32>, tensor<?x!tf_type.string>, tensor<?x2xi32>, tensor<3x!tf_type.string>, tensor<?x3xi32>
  }

  // CHECK-LABEL: func @deplicated_return_from_host_and_tpu

For example, if we have the code

```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

  %5 = "quantfork.stats"(%4) {layerStats = dense<[-56.2916565, 122.922478]> : tensor<2xf32>} : (tensor<1x4xf32>) -> tensor<1x4xf32>
  %6 = "tfl.svdf"(%0, %1, %2, %3, %5) {fused_activation_function = "RELU", rank = 1 : i32} : (tensor<1x3xf32>, tensor<2x3xf32>, tensor<2x1xf32>, tensor<2xf32>, tensor<1x4xf32>) -> tensor<1x2xf32>

  func.return %2 : tensor<1x128x32xf32>

// CHECK-DAG:  [[VAL_1:%.*]] = arith.constant dense<0> : tensor<3xi32>
// CHECK-DAG:  [[VAL_2:%.*]] = arith.constant dense<[1, 128, 32]> : tensor<3xi32>
// CHECK:  [[VAL_3:%.*]] = "tfl.slice"(%arg0, [[VAL_1]], [[VAL_2]]) : (tensor<4x128x32xf32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x128x32xf32>
}

// -----

// CHECK-LABEL: @WhileCanonicalizeBug

}

func.func @snapshot(%arg0: tensor<3xi32>) -> tensor<3xi32> {
  %0 = "tf.Snapshot"(%arg0) : (tensor<3xi32>) -> tensor<3xi32>
  func.return %0 : tensor<3xi32>
  // Should be converted to Identity and then from Identity to value
  // CHECK-LABEL: snapshot
  // CHECK:  return %arg0 : tensor<3xi32>
}

func.func @stop_gradient(%arg0: tensor<3xi32>) -> tensor<3xi32> {

// CHECK-DAG: %[[SCALE:.*]] = "tf.Const"() <{value = dense<0.0787401571> : tensor<f32>
}

// -----

module {
  func.func @quantize_xlagather(%arg0: tensor<10x2xi32>) -> tensor<1x300x10xf32> {
    %cst_0 = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<200x100x300xf32>} : () -> tensor<200x100x300xf32>

    %cst = "tf.Const"() {value = dense<0> : tensor<1x2xi32>} : () -> tensor<1x2xi32>
    %cst_0 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
    %cst_1 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
    %cst_2 = "tf.Const"() {value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>
    %cst_3 = "tf.Const"() {value = dense<4> : tensor<1xi32>} : () -> tensor<1xi32>

    // CHECK:      %[[ONE:.*]] = "tf.Const"() <{value = dense<1> : tensor<i32>}>
    // CHECK:      %[[RES_READ_VAL:[0-9]*]] = "tf.ReadVariableOp"
    // CHECK-SAME: (tensor<*x!tf_type.resource<tensor<2x8xi32>>>) -> tensor<2x8xi32>
    // CHECK:      "tf.AddV2"(%[[RES_READ_VAL]], %[[ONE]])
    // CHECK-SAME: (tensor<2x8xi32>, tensor<i32>) -> tensor<2x8xi32>
    // CHECK:      "tf.AssignVariableOp"

  %0 = "tf.Const"() {value = dense<20> : tensor<1xi32>} : () -> tensor<1xi32>
  %1 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
  %2 = "tf.Shape"(%arg0) : (tensor<?x1x2x3xf32>) -> tensor<4xi32>
  %3 = "tf.StridedSlice"(%2, %1, %0, %1) {begin_mask = 0 : i64, ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 0 : i64} : (tensor<4xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<3xi32>