%cst_1 = "tf.Const"() {value = dense<[2, 0, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
  %cst_2 = "tf.Const"() {value = dense<[512, 64]> : tensor<2xi32>} : () -> tensor<2xi32>
  %0 = "tf.ExpandDims"(%arg0, %cst) {device = ""} : (tensor<8x64xf32>, tensor<i32>) -> tensor<8x64x1xf32>
  %1 = "tf.AddV2"(%0, %arg1) {device = ""} : (tensor<8x64x1xf32>, tensor<8x64x64xf32>) -> tensor<8x64x64xf32>

  tfg.func @test_func_name0(%test_func_name0: tensor<*x!tf_type.resource> {tf._output_shapes = #tf_type.shape<*>, tf._resource_arg_unique_id = 0 : i64, tfg.name = "test_func_name0"},...

  func.return %0 : tensor<!tf_type.variant<tensor<2xi32>>>

module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
  func.func @main() -> () {
     %s0 = "tf.Const"() {value = dense<[4, 1]> : tensor<2xi32>} : () -> tensor<2xi32>
    func.return
  }

func.func @side_effect_free_user() -> tensor<2xi32> {
  // CHECK: [[handle:%.*]] = "tf.VarHandleOp"
  // CHECK-SAME: shared_name = "x"
  // CHECK: [[value0:%.*]] = "tf.ReadVariableOp"([[handle]])
  // CHECK: [[value1:%.*]] = "tf.ReadVariableOp"([[handle]])
  // CHECK: "tf.ConcatV2"([[value1]], [[value0]]
  // CHECK-SAME: (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<2xi32>

  // CHECK-DAG: %[[IND_SLICE0_START:.*]] = "tf.Const"() <{value = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK-DAG: %[[IND_SLICE0_SIZE:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK: %[[IND_SLICE0:.*]] = "tf.Slice"(%[[INDS]], %[[IND_SLICE0_START]], %[[IND_SLICE0_SIZE]]) : (tensor<2xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>

// CHECK: %[[reshape_i8:.*]] = "tf.Reshape"(%[[maxpool_i8]]
// CHECK-SAME: (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>

// -----

// CHECK-LABEL: standalone_avgpool_test
func.func @standalone_avgpool_test(%arg0: tensor<*xf32>) -> tensor<*xf32> {
  %cst = arith.constant dense<[-1, 144]> : tensor<2xi32>
  %0 = "quantfork.qcast"(%arg0) : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>

module {
func.func @main(%arg0: tensor<21x32x32x128xf32>, %arg1: tensor<2xi32>) -> tensor<1x64x64x128xf32> {
  %0 = "tf.ResizeBilinear"(%arg0, %arg1) {align_corners = false, device = "", half_pixel_centers = true} : (tensor<21x32x32x128xf32>, tensor<2xi32>) -> tensor<1x64x64x128xf32>
  func.return %0 : tensor<1x64x64x128xf32>
}

      %outputs_9, %control_10 = tf_executor.island(%control_4) wraps "tf.Identity"(%outputs_7) {device = ""} : (tensor<2x3xi32>) -> tensor<2x3xi32>
      tf_executor.fetch %outputs_9 : tensor<2x3xi32>
    }
    return %0 : tensor<2x3xi32>
  }
}

// CHECK: module attributes
// CHECK-SAME:  tfl.metadata = {{{.*}}keep_stablehlo_constant = "true"{{.*}}}

  // Reduce over spatial dimensions (NCHW)
  %1 = "tf.Const"() {value = dense<[2, 3]> : tensor<2xi32>} : () -> tensor<2xi32>

  // Transpose input: NHWC -> NCHW
  %2 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %3 = "tf.Transpose"(%arg0, %2) : (tensor<?x224x224x3xf32>, tensor<4xi32>) -> tensor<?x3x224x224xf32>

  // Pad spatial dimensions.