// ALWAYS-DECOMPOSE-NOT:  "tf.AssignAddVariableOp"
  "tf.AssignAddVariableOp"(%0, %1) {dtype = "tfdtype$DT_INT32"} : (tensor<*x!tf_type.resource<tensor<2x8xi32>>>, tensor<i32>) -> ()
  // CHECK: tf_device.cluster
  "tf_device.cluster"() ({
    // CHECK-NOT: "tf.AssignAddVariableOp"
    "tf.AssignAddVariableOp"(%0, %1) {dtype = "tfdtype$DT_INT32"} : (tensor<*x!tf_type.resource<tensor<2x8xi32>>>, tensor<i32>) -> ()
    tf_device.return
  }) : () -> ()

  %0 = "tf.Shape"(%arg0) {T = "tfdtype$DT_FLOAT", output = "tfdtype$DT_INT32"} : (tensor<f32>) -> tensor<0xi32>

  // Result shape need not be static. Folding harness uses TensorFlow constant
  // in that case.
  // CHECK-DAG: "tf.Const"() <{value = dense<[1, 32, 32, 16]> : tensor<4xi32>}> : () -> tensor<?xi32>
  %1 = "tf.Shape"(%arg1) {T = "tfdtype$DT_FLOAT", output = "tfdtype$DT_INT32"} : (tensor<1x32x32x16xf32>) -> tensor<?xi32>

}

func.func @const() -> tensor<2xi32> {

// -----

// CHECK-LABEL: @const
func.func @const() -> tensor<2xi32> {
  // CHECK: mhlo.constant dense<0> : tensor<2xi32>
  %0 = "tf.Const"() {device = "", name = "", dtype = "tfdtype$DT_INT32", value = dense<0> : tensor<2xi32>} : () -> (tensor<2xi32>)
  func.return %0: tensor<2xi32>
}

// -----

// CHECK-LABEL: @const_dynamic_output
func.func @const_dynamic_output() -> tensor<*xi32> {

    // CHECK: tf.BroadcastGradientArgs
    // CHECK-SAME: (tensor<*xi32>, tensor<0xi32>) -> (tensor<?xi32>, tensor<?xi32>)
    %2:2 = "tf.BroadcastGradientArgs"(%arg0, %arg1) {T = "tfdtype$DT_INT32", name = "BroadcastGradientArgs"} : (tensor<*xi32>, tensor<0xi32>) -> (tensor<?xi32>, tensor<?xi32>)
    func.return %2#0 : tensor<?xi32>
  }

  // CHECK-LABEL: func @shape_from_const_input