variables;
  for (auto var_handle_op : var_handle_ops)
    variables.insert(GetResourceKey(var_handle_op));

  llvm::SmallVector<Operation*, 4> work_list;
  for (auto var_handle_op : session_init_func.getOps<TF::VarHandleOp>()) {
    if (variables.count(GetResourceKey(var_handle_op)))
      work_list.push_back(var_handle_op);
  }

    // Creates a VarHandleOp -> ReadVariableOp pair for each ConstOp.
    const auto resource_type = RankedTensorType::get(
        /*shape=*/{}, /*elementType=*/TF::ResourceType::get(
            /*subtypes=*/llvm::ArrayRef<TensorType>{const_op.getType()},
            builder.getContext()));
    auto var_handle_op =
        builder.create<TF::VarHandleOp>(const_op.getLoc(),

      auto resource_type = var_handle_op.getResource().getType();
      func_arg_types.push_back(resource_type);
      var_handle_op.getResource().replaceAllUsesWith(
          block.addArgument(resource_type, var_handle_op.getLoc()));
    } else {
      var_handle_op.getResource().replaceAllUsesWith(
          block.getArgument(it.first->getSecond()));
    }
    var_handle_op.erase();
  }

}

using LocalVarOp = std::variant<TF::VarHandleOp, TF::MlirLocalVarOp>;

Value LocalVarOp_resource(LocalVarOp &op) {
  if (auto var_handle_op = std::get_if<TF::VarHandleOp>(&op)) {
    return var_handle_op->getResource();
  } else {
    return std::get<TF::MlirLocalVarOp>(op).getResource();
  }
}

void LocalVarOp_erase(LocalVarOp &op) {
  if (auto var_handle_op = std::get_if<TF::VarHandleOp>(&op)) {
    var_handle_op->erase();

// Test that the locs are properly set to their shared_names.
// CHECK-LOC: "tf.VarHandleOp"() <{{{.*shared_name = "var_0".*}}}>
// CHECK-LOC-SAME: loc("var_0")
// CHECK-LOC: "tf.VarHandleOp"() <{{{.*shared_name = "var_1".*}}}>
// CHECK-LOC-SAME: loc("var_1")
}

// -----

// RestoreV2 op not created for `AssignVariableOp(VarHandleOp, Const)` patterns
// in the initializer function of "init_op" type.

// a single string tensor as argument, which specifies the path to the
// checkpoint to which the variable's tensor values are saved. It finds
// `tf.AssignVariableOp(tf.VarHandleOp, tf.Const)` pattern in the initializer
// function of type "restore_op" to identify the VarHandleOps that should be
// saved using the SaveV2 op.
std::unique_ptr<OperationPass<ModuleOp>> CreateInsertSaveOpPass();

  attributes {tf_saved_model.exported_names = ["main"]} {
  // CHECK: tf.VarHandleOp
  %0 = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "variable1"} : () -> tensor<!tf_type.resource<tensor<1x3xf32>>>
  %1 = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "variable2"} : () -> tensor<!tf_type.resource<tensor<1x3xf32>>>

    // CHECK: "tf.VarHandleOp"
    func.return
  }
}

// -----

// Test mutation detection propagates across function calls.

module {
  func.func @f() -> tensor<0xf32> {
    // CHECK: "tf.VarHandleOp"

  attributes {tf_saved_model.exported_names = ["handles_ambiguous_var_handles"]} {
    // CHECK: VarHandleOp
    %0 = "tf.VarHandleOp"() {container = "", shared_name = "foo"} : () -> tensor<!tf_type.resource<tensor<10xf32>>>
    func.call @assign_resource(%0) : (tensor<!tf_type.resource<tensor<10xf32>>>) -> ()
    %1 = "tf.VarHandleOp"() {container = "", shared_name = "foo"} : () -> tensor<!tf_type.resource<tensor<10xf32>>>

  // CHECK-NOT: "tf.VarHandleOp"
  // CHECK-NOT: "tf.ReadVariableOp"
  // CHECK: %[[CONST:.*]] = "tf.Const"()
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg1, %[[CONST]])
  // CHECK: %[[PACK:[0-9]*]] = "tf.Pack"(%[[CONST]], %[[ADD]])
  // CHECK: return %[[PACK]]
  %0 = "tf.Const"() {value = dense<4.200000e+01> : tensor<f32>} : () -> tensor<f32>