// CHECK-NEXT: %[[none_min:.*]] = "tfr.constant_tensor"(%[[N_128]])
// CHECK-NEXT: %[[none_max:.*]] = "tfr.constant_tensor"(%[[N127]])
// CHECK-NEXT: %[[relu_min:.*]] = "tfr.constant_tensor"(%[[N42]])
// CHECK-NEXT: %[[relu_max:.*]] = "tfr.constant_tensor"(%[[N127]])
// CHECK-NEXT: %[[relu6_min:.*]] = "tfr.constant_tensor"(%[[N42]])
// CHECK-NEXT: %[[relu6_max:.*]] = "tfr.constant_tensor"(%[[N102]])

  %batch = "tfr.constant_tensor"(%batch_size) : (i64) -> tensor<i64>
  %batch1 = "tfr.cast"(%batch) : (tensor<i64>) -> !tfr.tensor
  %calls = "tfr.constant_tensor"(%num_parallel_calls) : (i64) -> tensor<i64>
  %calls1 = "tfr.cast"(%calls) : (tensor<i64>) -> !tfr.tensor
  %drop = "tfr.constant_tensor"(%drop_remainder) : (i1) -> tensor<i1>

  }];
}

def TFR_ConstantTensorOp : TFR_Op<"constant_tensor", [Pure]> {
  let description = [{
    The `constant_tensor` operation converts the operand with non-built-in
    tensor type to built-in tensor type or tfr.tensor type. If it is built-in
    tensor type, the shape shouldn't be changed during the conversion.

    Example:

    ```mlir
    %1 = tfr.constant_tensor(%0) : f32 -> tensor<f32>

}

// -----

// CHECK-LABEL: constant_tensor_array
func.func @constant_tensor_array() -> !tfr.tensor {
  %0 = tfr.constant [1, -1, 3] -> !tfr.attr
  %1 = "tfr.constant_tensor"(%0) : (!tfr.attr) -> !tfr.tensor
  func.return %1 : !tfr.tensor

// CHECK-NEXT: %[[RES:.*]] = "tf.Const"() <{value = dense<[1, -1, 3]> : tensor<3xi64>}> : () -> tensor<3xi64>