%2 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  func.return %2 : tensor<4xf32>

// CHECK-LABEL: main
// CHECK-NOT: tfl.squared_difference
// CHECK: tfl.mul %[[CONST:.*]], %arg0

^bb0(%arg0: tensor<4xf32>):
  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  // CHECK: %[[MUL:.*]] = tfl.mul
  %2 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")

  // Confirm that attributes that cannot be stored in the flatbuffer options
  // for a given operator are dropped silently.
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")

  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  // CHECK: %[[DIV:.*]] = tfl.div
  %2 = "tfl.div"(%1, %arg0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

  // CHECK-NEXT: }

  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")

^bb0(%arg0: tensor<4xf32>):
  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  // CHECK: %[[MUL:.*]] = tfl.mul
  %2 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")

// RUN: tac-translate -input-mlir -output-mlir -device-specs=NNAPI %s -o - 2>&1 | FileCheck %s

module {
  // CHECK-LABEL: main
  func.func @main(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
    %0 = "tfl.squared_difference"(%arg0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
    func.return %0 : tensor<4xf32>
    // CHECK:  [[VAL_0:%.*]] = tfl.sub %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<4xf32>

}

TargetHardwareRegistration<NNAPIHardware> nnapi_hardware(
    "Target device for NNAPI", CreateNNAPIHardware);

// Currently used for these ops:
// tfl.squared_difference
class NNAPIBasicSupportedOpNoCost : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override { return 0; }

  bool IsOpSupported(mlir::Operation* op) const override {
    return true;
  }
};

// tfl.custom / tfl.mean / tfl.minimum / tfl.pad / tfl.pow / tfl.prelu /
// tfl.relu / tfl.relu6 / tfl.rsqrt / tfl.sin / tfl.slice / tfl.softmax /
// tfl.space_to_depth / tfl.sqrt / tfl.square / tfl.squared_difference /
// tfl.strided_slice / tfl.tanh / tfl.transpose / tfl.transpose_conv
class GpuBasicSupportedOpNoCost : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override { return 0; }

}

// CHECK-NOT: "tfl.reshape"
// CHECK: "tfl.pack"

// -----

func.func @squaredDifference(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
  %0 = "tfl.squared_difference"(%arg0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  func.return %0 : tensor<4xf32>
}

// CHECK:       func @squaredDifference(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {