%10 = "tfl.pseudo_const"() {value = dense<0.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %11 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %recurrent_input = "tfl.pseudo_const"() {value = dense<0.000000e+00> : tensor<1x3xf32>} : () -> tensor<1x3xf32>
  %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>

// CHECK:           return %[[VAL_2]], %[[VAL_3]] : tensor<4xf32>, tensor<4xf32>
// CHECK:         }
func.func @broadcast_atan2(%arg0: tensor<1xf32>, %arg1: tensor<4xf32>) -> (tensor<4xf32>, tensor<4xf32>) {
  %0 = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<[0]> : tensor<1xi64>}> : (tensor<1xf32>) -> tensor<4xf32>
  %1 = mhlo.atan2 %0, %arg1 : tensor<4xf32>
  %2 = mhlo.atan2 %arg1, %0 : tensor<4xf32>

  func.return %0 : tensor<*xf32>
}

// 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>
}

  %7 = "tfl.add"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<  f32>) -> tensor<4xf32>
  %8 = "tfl.add"(%2, %3) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  %9 = "tfl.add"(%2, %3) {fused_activation_function = "SIGN_BIT"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

  %1:2 = "tfl.split"(%cst, %0) {num_splits = 2 : i32} : (tensor<i32>, tensor<4xf32>) -> (tensor<2xf32>, tensor<2xf32>)
  %2 = "tfl.quantize"(%1#0) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>
  %3 = "tfl.quantize"(%1#1) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>

// CHECK-LABEL: floor_mod
func.func @floor_mod(%arg0: tensor<5xf32>, %arg1: tensor<5xf32>) -> tensor<5xf32> {
  %0 = "tf.FloorMod"(%arg0, %arg1) : (tensor<5xf32>, tensor<5xf32>) -> tensor<5xf32>
  func.return %0 : tensor<5xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0, %arg1) <{custom_code = "FlexFloorMod", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<5xf32>, tensor<5xf32>) -> tensor<5xf32>
// CHECK: return %[[CUSTOM_0]] : tensor<5xf32>
}

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>

  %1:4 = "tfl.split"(%cst, %0) {num_splits = 4 : i32} : (tensor<i32>, tensor<4xf32>) -> (tensor<2xf32>, tensor<2xf32>,tensor<2xf32>, tensor<2xf32>)
  %2 = "tfl.quantize"(%1#0) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>
  %3 = "tfl.quantize"(%1#1) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<3xf32>) -> tensor<*xf32>
    %2 = "tf.PartitionedCall"(%arg0, %cst_2) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn_1} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>
    func.return %1, %2: tensor<*xf32>, tensor<*xf32>

  %2 = "tf.Relu6"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
  func.return %2 : tensor<*xf32>
}

// CHECK-LABEL: func @conv_with_non_constant_filter
// CHECK-DAG: %[[CONST_0:.*]] = "tf.Const"() <{value = dense<0.000000e+00> : tensor<2xf32>}> : () -> tensor<2xf32>
// CHECK: %[[PARTITIONEDCALL_0:.*]] = "tf.PartitionedCall"(%arg0, %arg1, %[[CONST_0]])
// CHECK-SAME: f = @composite_conv2d_with_bias_and_relu6_fn_1