"tf_saved_model.global_tensor"() {sym_name = "v2", type = tensor<f32>, value = dense<1.0> : tensor<f32> } : () -> ()

  func.func @f(%arg0: tensor<!tf_type.resource<tensor<f32>>> {tf_saved_model.bound_input = @v})
  attributes {tf_saved_model.exported_names = ["f"]} {
    %val = "tf.ReadVariableOp"(%arg0) : (tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32>
    func.return

    %3 = "tf.NotEqual"(%1, %2) : (tensor<f32>, tensor<f32>) -> tensor<i1>
    "tf.Yield"(%3) : (tensor<i1>) -> ()
  },  {
  ^bb0(%arg3: tensor<f32>, %arg4: tensor<*xf32>):
    %cst = "tf.Const"() {value = dense<1.0> : tensor<f32>} : () -> tensor<f32>
    %1 = "tf.Sub"(%arg3, %cst) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    "tf.Yield"(%1, %arg4) : (tensor<f32>, tensor<*xf32>) -> ()

  %2 = "tf.ReadVariableOp"(%arg2) : (!tf_res_f32) -> tensor<f32>
  %execute_output:3 = "tf_device.parallel_execute"() ({
    %3:3 = "tf_device.cluster_func"(%0, %1, %2) {_xla_compile_device_type = "TPU", _replication_info = "tpu", func = @device_func_5} : (tensor<f32>, tensor<f32>, tensor<f32>) -> (tensor<f32>, tensor<f32>, tensor<f32>)
    tf_device.return %3#0, %3#1, %3#2 : tensor<f32>, tensor<f32>, tensor<f32>

// CHECK:         %[[ARG_TUPLE]] = (f32[128,10]{1,0}, f32[10,1024]{1,0}, f32[128,1024]{1,0}) parameter(0)
// CHECK-SAME:    sharding={
// CHECK-SAME:    {devices=[1,2]0,1}
// CHECK-SAME:    {maximal device=0}
// CHECK-SAME:    {replicated}
// CHECK-SAME:    }

// CHECK-LABEL: dequantize
func.func @dequantize(%arg0: tensor<2x3x!tf_type.qint8>, %min_range: tensor<f32>, %max_range: tensor<f32>) -> tensor<2x3xf32> {
  // CHECK-DAG: %[[HALF_RANGE:.*]] = "tf.Const"() <{value = dense<1.280000e+02> : tensor<f32>}>
  // CHECK-DAG: %[[C255:.*]] = "tf.Const"() <{value = dense<2.550000e+02> : tensor<f32>}>
  // CHECK-DAG: %[[CAST:.*]] = "tf.Cast"(%arg0) <{Truncate = false}>

    %0 = "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<f32>
    %1 = "tf.ReadVariableOp"(%arg1) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<f32>
    %2 = "tf.ReadVariableOp"(%arg2) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<f32>
    "tf.OpA"(%0, %1, %2) : (tensor<f32>, tensor<f32>, tensor<f32>) -> ()
  }
  func.return

  %q2 = "quantfork.qcast"(%q1) {volatile} : (tensor<3x2x!quant.uniform<i8:f32, 6.000000e-03:-128>>) -> tensor<3x2x!quant.uniform<i8:f32, 0.013075299590241675:-64>>
  %dq1 = "quantfork.dcast"(%q2) : (tensor<3x2x!quant.uniform<i8:f32, 0.013075299590241675:-64>>) -> tensor<3x2xf32>

  %1 = "tfl.quantize"(%arg1) {qtype = tensor<2x1x!quant.uniform<u8:f32, 1.0:128>>}: (tensor<2x1xf32>) -> tensor<2x1x!quant.uniform<u8:f32, 1.0:128>>
  %2 = "tfl.dequantize"(%0) : (tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>) -> tensor<2x2xf32>
  %3 = "tfl.dequantize"(%1) : (tensor<2x1x!quant.uniform<u8:f32, 1.0:128>>) -> tensor<2x1xf32>

func.func @removeDeadReadVariableOp(%arg0: tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32> {
  %0 = "tf.ReadVariableOp"(%arg0) : (tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32>
  %1 = "tf.ReadVariableOp"(%arg0) : (tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32>
  func.return %0: tensor<f32>

 // CHECK: %0 = "tf.ReadVariableOp"(%arg0) : (tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32>
 // CHECK: return %0

      "tfl.yield"(%1) : (tensor<i1>) -> ()
    },
    // body
    {
    ^bb0(%arg0: tensor<f32>, %arg1: tensor<f32>, %arg2: tensor<i32>):
      %2 = "tfl.add"(%arg0, %arg0) {fused_activation_function = "NONE"} : (tensor<f32>, tensor<f32>) -> tensor<f32>
      %3 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "NONE"} : (tensor<f32>, tensor<f32>) -> tensor<f32>
      %cst_4 = "arith.constant" () {value = dense<1> : tensor<i32>} : () -> tensor<i32>