return failure();
  }

  // Fetch all VarHandleOp.
  llvm::StringSet<> variable_names;
  llvm::SmallVector<TF::VarHandleOp, 4> var_ops;
  for (auto func_op : module.getOps<func::FuncOp>()) {
    for (auto var_handle_op : func_op.getOps<TF::VarHandleOp>()) {
      auto variable_name = GetVariableName(var_handle_op);
      if (variable_names.count(variable_name)) continue;

  auto resource_operand = assign_var_op.getOperand(0);
  auto var_handle_op =
      llvm::dyn_cast<mlir::TF::VarHandleOp>(resource_operand.getDefiningOp());
  if (!var_handle_op) {
    assign_var_op->emitRemark(
        "Operand idx 0 is not a tf.VarHandleOp. The initializing tensor is not "
        "saved to checkpoint.");
    return "";
  }

// CHECK: func.func private @tf_quant__save(%[[ARG:.*]]: tensor<!tf_type.string>)
// CHECK: %[[VAR_HANDLE_0:.*]] = "tf.VarHandleOp"()
// CHECK-SAME: {{.*shared_name = "var_0".*}}
// CHECK: %[[VAR_HANDLE_1:.*]] = "tf.VarHandleOp"()
// CHECK-SAME: {{.*shared_name = "var_1".*}}

// ReadVariableOps are inserted for each VarHandleOp to read the tensor values.

    }
  }
  return nullptr;
}

Type GetGlobalType(Operation* source) {
  if (auto var_handle_op = dyn_cast<TF::VarHandleOp>(source)) {
    // Resources are represented as tensor<resource<tensor<...>>>, so
    // unwrap until we get to the inner tensor<...>.
    auto tensor =
        llvm::dyn_cast<TensorType>(var_handle_op.getResource().getType());
    if (!tensor) return nullptr;
    TF::ResourceType resource =

    StringRef variable_name =
        GetNodeNameFromClassAttrOrSharedNameAttr(variable_v2_op);
    if (variable_name.empty()) {
      return signalPassFailure();
    }
    VarHandleOp var_handle_op = builder.create<VarHandleOp>(
        variable_v2_op.getLoc(),
        ArrayRef<Type>{RankedTensorType::get(
            {}, TF::ResourceType::get(ArrayRef<TensorType>{tensor_type},

    %0 = "tf.VarHandleOp"() {_class = ["loc:@dense/kernel"], allowed_devices = [], container = "", device = "/job:worker/replica:0/task:1/device:CPU:0", shared_name = "var1"} : () -> tensor<!tf_type.resource<tensor<100x50xf32>>>
    %1 = "tf.ReadVariableOp"(%0) {device = ""} : (tensor<!tf_type.resource<tensor<100x50xf32>>>) -> tensor<100x50xf32>

// CHECK-NOT: "tf.VarHandleOp"
func.func @no_args() {
  %0 = "tf.VarHandleOp"() {container = "", shape = "tfshape$", shared_name = "x"} : () -> tensor<!tf_type.resource<tensor<f32>>>
  func.return
}

// CHECK-LABEL: func @some_args
// CHECK-SAME: (%arg0: tensor<i1>, %arg1: tensor<!tf_type.resource<tensor<f32>>> {tf.resource_name = "x"})
// CHECK-NOT: "tf.VarHandleOp"
func.func @some_args(%arg0: tensor<i1>) {

  // CHECK: func @init()
  // CHECK-NOT: tf.VarHandleOp
  // CHECK-NOT: tf.Const
  // CHECK-NOT: tf.AssignAddVariableOp
  // CHECK: return
  "tf_saved_model.session_initializer"() { initializers = [@init] } : () -> ()
  func.func @init() attributes {tf_saved_model.exported_names = ["__tf_saved_model_session_initializer"], tf_saved_model.initializer_type = "restore_op"} {

    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
    return
  }
  // Nothing has been removed.
  // CHECK: @init_func_init_op
  // CHECK-NEXT: "tf.Const"
  // CHECK-NEXT: "tf.VarHandleOp"
  // CHECK-NEXT: "tf.AssignVariableOp"

    // CHECK: %[[VAR0:.*]] = "tf.VarHandleOp"()
    %0 = "tf.VarHandleOp"() {container = "", shared_name = "var0"} : () -> tensor<!tf_type.resource<tensor<!tf_type.string>>>

    // CHECK: "tf.AssignVariableOp"(%[[VAR0]], %[[ASSET0]])
    "tf.AssignVariableOp"(%0, %arg0) : (tensor<!tf_type.resource<tensor<!tf_type.string>>>, tensor<!tf_type.string>) -> ()

    // CHECK: %[[VAR1:.*]] = "tf.VarHandleOp"()