```mlir
      %2 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
      %3 = "tf.B"(%2) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
      %4 = "tf.C"(%2, %3) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
      %5 = "tf.D"(%4) : (tensor<?xi32>) -> tensor<?xi32>
    ```

    After the pass, we will have:

    ```mlir
      %0 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>

// CHECK-LABEL: floor_mod
func.func @floor_mod(%arg0: tensor<5xf32>, %arg1: tensor<5xf32>) -> tensor<5xf32> {
  %0 = "tf.FloorMod"(%arg0, %arg1) : (tensor<5xf32>, tensor<5xf32>) -> tensor<5xf32>
  func.return %0 : tensor<5xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0, %arg1) <{custom_code = "FlexFloorMod", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<5xf32>, tensor<5xf32>) -> tensor<5xf32>
// CHECK: return %[[CUSTOM_0]] : tensor<5xf32>
}

#loc3])\0Amodule {\0A  func.func @host_func(%arg0: tensor<3360x?xi32> loc(fused[#loc1, #loc2, #loc3]), %arg1: tensor<3xi32> loc(fused[#loc1, #loc2, #loc3]), %arg2: tensor<i32> loc(fused[#loc1, #loc2, #loc3]), %arg3: tensor<2xi32> loc(fused[#loc1, #loc2, #loc3])) -> (tensor<1x1120x?xi32>, tensor<1x1120x?xi32>, tensor<1120x?xi32>, tensor<2xi32>) {\0A    %0 = \22tf.Reshape\22(%arg0, %arg1) {_xla_outside_compilation = \220\22} : (tensor<3360x?xi32>, tensor<3xi32>) -> tensor<3x1120x?xi32> loc(#loc9)\0A    %1:3 = ...

# CHECK-NEXT:   "tfl.yield"(%[[OUTPUTS]]) : (tensor<*xi32>) -> ()
# CHECK-NEXT: }) {T = i32, device = ""} : (tensor<4xi32>, tensor<4xi32>) -> tensor<*xi32>
# CHECK-NEXT: return %[[CUSTOM]] : tensor<*xi32>

    %3 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %4 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %5 = "tf.Const"() {value = dense<0> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %6 = "tf.Const"() {value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>

// CHECK-DAG:   [[ZERO_1:%cst.*]] = arith.constant dense<0> : tensor<i32>
// CHECK-DAG:   [[PARTIAL_INDEX:%.*]] = "tf.Fill"([[VECTOR_RANK]], [[ZERO_1]]) : (tensor<1xi32>, tensor<i32>) -> tensor<?xi32>
// CHECK-DAG:   [[EXPANDED_INDIX:%.*]] = "tf.ExpandDims"([[INDEX]], [[ZERO]]) : (tensor<i32>, tensor<i32>) -> tensor<1xi32>

        %5:2 = "tf.C"(%4) : (tensor<2xi32>) -> (tensor<2xi32>, tensor<2xi32>)
        tf_device.return %5#0, %5#1 : tensor<2xi32>, tensor<2xi32>
      }) {num_cores_per_replica = 1, topology =  "", device_assignment =  []} : () -> (tensor<2xi32>, tensor<2xi32>)
      tf_device.return %2#1 : tensor<2xi32>
    }
    func.return %1 : tensor<2xi32>
  }

    func.func @tpu_partitioned_call_func(%arg0: tensor<?xi32>) -> (tensor<?xi32>) {
      func.return %arg0 : tensor<?xi32>
    }

    func.func @main(%arg0: tensor<20xi32>, %arg1: tensor<?xi32>) -> tensor<*xi32> {
      %2 = "tf.TPUPartitionedCall"(%arg0, %arg1) {f = @tpu_partitioned_call_func} : (tensor<20xi32>, tensor<?xi32>) -> tensor<*xi32>
      func.return %2 : tensor<*xi32>
    }
  })";

  %1 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi32>} : () -> tensor<3xi32>
  %2 = "tfl.pseudo_const"() {value = dense<[0, 2, 0]> : tensor<3xi32>} : () -> tensor<3xi32>
  %3 = "tfl.pseudo_const"() {value = dense<1> : tensor<3xi32>} : () -> tensor<3xi32>

  %4 = "tfl.var_handle"() {container = "", shared_name = "read_assign/states0"} : () -> tensor<!tf_type.resource>

  %cst_0 = "tf.Const"() {value = dense<-3> : tensor<i32>} : () -> tensor<i32>
  %cst_1 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
  %cst_2 = "tf.Const"() {value = dense<4> : tensor<1x2xi32>} : () -> tensor<1x2xi32>
  %0 = "tf.SpaceToBatchND"(%arg0, %cst_1, %cst_2) : (tensor<1x128x3xf32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<2x68x3xf32>
  %1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<2x68x3xf32>, tensor<i32>) -> tensor<2x1x68x3xf32>