module {

  func.func private @internal_dequantize_f32(
                           %input : tensor<*xf32>, %weight_scale : tensor<*xf32>) -> tensor<*xf32> {
    %mul = "tf.Mul"(%input, %weight_scale) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %mul : tensor<*xf32>
  }

  func.func private @internal_conv3d_fn(
                         %input : tensor<*xf32>, %filter : tensor<*xi8>) -> tensor<*xf32> {

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>
// CHECK:  [[MUL:%.*]] = tfl.mul(%arg0, [[CST]]) <{fused_activation_function = "NONE"}> : (tensor<3xbf16>, tensor<3x3xbf16>) -> tensor<3x3xbf16>
// CHECK:  return [[MUL]] : tensor<3x3xbf16>

// CHECK-NEXT:  signature_defs: [ ]
// CHECK-NEXT:}

  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  // tf.MyCustomOp is the result of conversion to a Custom op

  %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")
  %3 = "tfl.div"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

  %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

  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<3x!quant.uniform<i8:f32, 0.1>>, tensor<3x!quant.uniform<i8:f32, 0.1>>) -> tensor<3x!quant.uniform<i8:f32, 0.1>> loc("mul")
  func.return %1 : tensor<3x!quant.uniform<i8:f32, 0.1>>

  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<3x!quant.uniform<i8:f32, 1.0>>, tensor<3x!quant.uniform<i8:f32, 1.0>>) -> tensor<3x!quant.uniform<i8:f32, 1.0>> loc("mul")
  func.return %1 : tensor<3x!quant.uniform<i8:f32, 1.0>>

// CHECK-NEXT:  signature_defs: [ ]
// CHECK-NEXT:}

  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  // tf.div is the result of conversion to a Flex TF op
  %2 = "tf.Div"(%1, %0)  : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

// conditions are also hoisted (i.e., `tf.Add` and `tf.Mul`).

// CHECK-LABEL: readvariableop_is_hoisted_if_readonly
// CHECK:       [[CST_0:%.*]] = "tf.Const"
// CHECK:       [[VAR:%.*]] = "tf.VarHandleOp"
// CHECK:       [[CST_1:%.*]] = "tf.Const"
// CHECK:       [[VAR_VAL:%.*]] = "tf.ReadVariableOp"
// CHECK:       [[RES_1:%.*]] = "tf.Add"([[VAR_VAL]], [[CST_0]])
// CHECK:       [[RES_2:%.*]] = "tf.Mul"([[RES_1]], [[CST_1]])

  // CHECK:  %[[IS_ZERO:.*]] = "tf.Equal"(%[[Y]], %[[ZERO]]) <{incompatible_shape_error = true}> : (tensor<3xf32>, tensor<f32>) -> tensor<3xi1>
  // CHECK:  %[[MUL:.*]] = "tf.Mul"(%[[X]], %[[Y]]) : (tensor<2x3xf32>, tensor<3xf32>) -> tensor<2x3xf32>
  // CHECK:  %[[RESULT:.*]] = "tf.SelectV2"(%[[IS_ZERO]], %[[ZERO]], %[[MUL]]) : (tensor<3xi1>, tensor<f32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  %0 = "tf.MulNoNan"(%arg0, %arg1) : (tensor<2x3xf32>, tensor<3xf32>) -> tensor<2x3xf32>