%cst_2 = "tf.Const"() {value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>
  %cst_3 = "tf.Const"() {value = dense<4> : tensor<1xi32>} : () -> tensor<1xi32>
  %0 = "tf.XlaReduceWindow"(%arg0, %arg1, %cst_0, %cst_1, %cst_2, %cst_3, %cst) {computation = @sum_reducer} : (tensor<7xf32>, tensor<f32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<10xf32>

  func.return %1 : tensor<1xf32>
}

func.func @cond_true(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
  %0 = tfl.add %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

func.func @cond_false(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {

    %outputs_6:2, %control_7 = tf_executor.island wraps "tf.While"(%arg0, %arg1) {body = @body, cond = @cond, is_stateless = false, shape_invariant} : (tensor<i32>, tensor<5xf32>) -> (tensor<i32>, tensor<5xf32>) loc("WhileWithOutputShapes")
    tf_executor.fetch %outputs_2#1, %outputs_4#1, %outputs_6#1 : tensor<5xf32>, tensor<5xf32>, tensor<5xf32>
  }

// CHECK-NOT: tf.BatchFunction
// CHECK: %[[ADD0:.*]] = "tf.AddV2"(%[[ARG0]], %[[ARG1]])
// CHECK: return %[[ADD0]] : tensor<1xf32>

func.func private @batched_func(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Identity"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
  %1 = "tf.Identity"(%arg1) : (tensor<1xf32>) -> tensor<1xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

    //body
    ^bb0(%arg_body: tensor<1xf32>):
    %result_body = "tfl.add"(%arg_body, %arg_body) { fused_activation_function = "NONE" } : (tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>)
    "tfl.yield"(%result_body) : (tensor<1xf32>) -> ()
  }) : (tensor<1xf32>) -> (tensor<1xf32>)
  %tmp5 = "tfl.add"(%tmp4, %tmp2) { fused_activation_function = "NONE" } : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

func.func @multiply(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "vhlo.multiply_v1"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0 : tensor<1xf32>
}

// CHECK:func.func private @multiply(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
// CHECK: %0 = "vhlo.multiply_v1"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
// CHECK: return %0 : tensor<1xf32>
// CHECK:}

    return
  }
  // CHECK-LABEL: func private @foo.impl
  func.func private @foo.impl(%arg0: tensor<1xf32>, %arg1: tensor<2xf32>) -> (tensor<2xf32>, tensor<1xf32>) {
    return %arg1, %arg0 : tensor<2xf32>, tensor<1xf32>
  }

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

  %fill = "tf.Fill"(%cst_0, %cst) : (tensor<1xi64>, tensor<f32>) -> tensor<1xf32>
  tf_device.replicate([%0, %fill] as %arg_r0: tensor<1xf32>) {n = 2 : i32} {
    %1 = "tf_device.launch"() <{device = "TPU_REPLICATED_HOST_0"}> ({
      %2 = "tf.Identity"(%arg_r0) : (tensor<1xf32>) -> tensor<1xf32>
      tf_device.return %2 : tensor<1xf32>
    }) : () -> tensor<1xf32>