%outputs_22 =  "tf.Const"() {device = "", value = dense<16> : tensor<i32>} : () -> tensor<i32>
  %outputs_24 =  "tf.Const"() {device = "", value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %outputs_26 =  "tf.Range"(%outputs_24, %outputs_22, %outputs_20) {device = ""} : (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<*xi32>
  // CHECK: [[read_only_tensor_list:%.*]] = "tf.TensorListFromTensor"

  %0 = "tfl.while"(%arg0) ({
  ^bb0(%block: tensor<i32>):
    %1 = "tfl.add"(%arg0, %arg0) { fused_activation_function = "NONE", tac.device = "CPU", tac.inference_type = "FLOAT"} : (tensor<i32>, tensor<i32>) -> (tensor<i32>)
    %2 = "tfl.add"(%1, %1) { fused_activation_function = "NONE", tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<i32>, tensor<i32>) -> (tensor<i32>)
    "tfl.yield"(%2) : (tensor<i32>) -> ()
  },{

      name = "avgpool", filter_height = 3 : i32, filter_width = 6 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 3 : i32, stride_w = 1 : i32
    } : (tensor<1x224x224x3xf32>) -> tensor<1x224x224x3xf32>
  %6 = "tfl.conv_2d"(%5, %4, %cst) {
      dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32

// CHECK: %[[ADD_0:.+]] = stablehlo.add %[[DOT_GENERAL_0]], %[[DYNAMIC_BROADCAST_IN_DIM_0]] : tensor<?x3x!quant.uniform<i32:f32:1, {{.*}}>>
// CHECK: %[[UNIFORM_QUANTIZE_1:.+]] = stablehlo.uniform_quantize %[[ADD_0]] : (tensor<?x3x!quant.uniform<i32:f32:1, {{.*}}>>)

  %2 = "tf.opC"() {is_stateless = true} : () -> tensor<f32>
  %3 = "tf.opD"(%ri_0, %ri_1, %arg2, %2) {_xla_compile_device_type = "TPU", _replication_info = "replicate", is_stateless = true} : (tensor<i1>, tensor<i32>, tensor<f32>, tensor<f32>) -> tensor<i32>
  %ro_0:2 = "tf.TPUReplicatedOutput"(%3) : (tensor<i32>) -> (tensor<i32>, tensor<i32>)

    // CHECK: %[[RES_READ_VAL:[0-9]*]] = "tf.ReadVariableOp"
    // CHECK: "tf.Sub"(%[[RES_READ_VAL]], %[[ONE]])
    // CHECK: "tf.AssignVariableOp"

    %1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
    "tf.AssignSubVariableOp"(%0, %1) {dtype = "tfdtype$DT_INT32"} : (tensor<!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()

    tf_device.return

  // CHECK:  %[[CAST:.*]] = "tf.Cast"(%arg0) <{Truncate = false}> : (tensor<ui32>) -> tensor<i32>
  // CHECK:  %[[CAST1:.*]] = "tf.Cast"(%arg1) <{Truncate = false}> : (tensor<ui32>) -> tensor<i32>
  // CHECK:  %[[ADD:.*]] = "tf.AddV2"(%[[CAST]], %[[CAST1]]) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK:  %[[CAST2:.*]] = "tf.Cast"(%[[ADD]]) <{Truncate = false}> : (tensor<i32>) -> tensor<ui32>
  // CHECK:  return %[[CAST2]] : tensor<ui32>
}

// CHECK-DAG: %[[input_12:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x!quant.uniform<i32:f32, 1.1776238018224695E-4>>) -> tensor<2xf32>
// CHECK-DAG: %[[input_13:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x!quant.uniform<i32:f32, 1.9420648637759368E-4>>) -> tensor<2xf32>
// CHECK-DAG: %[[input_14:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x!quant.uniform<i32:f32, 1.1827185827747504E-4>>) -> tensor<2xf32>

// %13 = stablehlo.convert %11  // i8->i32 cast for q1.
// %14 = stablehlo.convert %3  // [Optional] i8->i32 cast for z1.
// %15 = stablehlo.broadcast_in_dim %14  // Operand = %4 if no `convert` above.
// %16 = stablehlo.subtract %13, %15  // q1 - z1
// %17 = stablehlo.convert %12  // i8->i32 cast for q2.
// %18 = stablehlo.convert %6  // [Optional] i8->i32 cast for z2.

//   %start = "tf.Const"() {value = dense<0> : tensor<i32>}
//   %limit = "tf.Const"() {value = dense<5> : tensor<i32>}
//   %delta = "tf.Const"() {value = dense<1> : tensor<i32>}
//   %updates = "tf.Range"(%start, %limit, %delta) :
//     (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<5xi32>
//   %shape = "tf.Const"() {value = dense<[5, 1]> : tensor<2xi32>}