Say we have a polymorphic class Dinosaur:

struct Dinosaur {
    virtual ~Dinosaur() = default;

And some concrete classes:

struct Diplodocus: Dinosaur {};

struct Stegosaurus : Dinosaur {};

struct Tyrannosaurus : Dinosaur {};

and a factory function that creates an instance based on some runtime value, say some string:

// If `name` equals "Diplodocus", return a Diplodocus;
// Else if `name` equals "Stegosaurus", return a Stegosaurus;
// Else if `name` equals "Tyrannosaurus", return a Tyrannosaurus;
// Else, return nullptr.
std::unique_ptr<Dinosaur> make_dinosaur(std::string_view name);

This is a common pattern. Let’s explore some ways to implement make_dinosaur.

The I-Need-to-Go-Home-in-10-Minute Solution

The most straightfoward way is chaining ifs:

std::unique_ptr<Dinosaur> make_dinosaur(std::string_view name) {
    if (name == "Diplodocus") {
        return std::make_unique<Diplodocus>();
    } else if (name == "Stegosaurus") {
        return std::make_unique<Stegosaurus>();
    } else if (name == "Tyrannosaurus") {
        return std::make_unique<Tyrannosaurus>();
    } else {
        return nullptr;

It’s all good, and I doubt it’ll take 5 minutes.

But, suddenly it starts to rain heavily and the elevator breaks. You have to unfortunately stay in the office for a bit longer. You need to kill some time anyway, so you stare at the code, and start thinking:

What can be done to improve the code? What can you do now to make life easier for the future you?

Analysis of the Solution

As we can see, there is high proportion of boilerplate. The only new information of each code block is the dinosaur name. Everything else just repeats themselves:

  • String comparison ==
  • std::make_unique

Any of the above are subject to change in business requirements:

  • Case-insensitive string comparison
  • Returning shared_ptr instead
  • Additional arguments to constructor

When any of the above happens, we would need to revise every block carefully.

Besides, the name of the type and type itself are not related, so there’s room for copy-pasta error when new dinosaurs are added.

Macros to the rescue … ?

We may use a helper macro to reduce the boilerplate:

#define MAKE_DINO_IMPL(T) if (name == #T) \
        return std::make_unique<T>()


std::unique_ptr<Dinosaur> make_dinosaur(std::string_view name) {
    return nullptr;

Problems of macros aside, this does improve the solution. It factors out all the repeating parts and leaves only the core information (the types) in the function body.

Can we do better yet? Let’s try templates! Whenever we want to replace macros, templates are our friends.


If we try to convert the macro to a function template, we might have:

template <class T>
std::string get_typename();

template <class T>
bool make_dino_impl(std::unique_ptr<Dinosaur>& out, std::string_view name) {
    if (name == get_typename<T>()) {
        out = std::make_unique<T>();
        return true;
    return false;

get_typename can be implemented by, for example, leveraging __PRETTY_FUNCTION__. Alternatively, we may add a static function type_name() to each concrete class:

struct Diplodocus: Dinosaur {
    static std::string type_name() { return "Dinosaur"; }

Bottom line is, get_typename<T> is a (mostly) solved problem.

A notion of list of types

Now that we have our templated make_dino_impl, we want to apply it to all the dinosaurs. It’s a for_each, but on types instead of on values!

And finally, I’m ready to introduce the star of the day: type_sequence.

template <class... T>
struct type_sequence{};

I first bumped into this construct in Andrei Alexandrescu’s CppCon talk, where he “urged” the committee to add the one-liner to the standard. This construct is handy, even outside heavy template meta-programming, as it turns parameter packs into types and values which we can pass around much more easily.

With it, we can define our list of types as follows:

using dinasour_types = type_sequence<

And apply the “for_each” equivalent as:

template <class... Ts>
std::unique_ptr<Dinosaur> make_dinosaur_from(type_sequence<Ts...>, std::string_view name) {
    std::unique_ptr<Dinosaur> p;
    (make_dino_impl<Ts>(p, name) || ...); // for each type in Ts ...
    return p;

std::unique_ptr<Dinosaur> make_dinosaur(std::string_view name) {
    return make_dinosaur_from(dinasour_types{}, name);

Are we making things better though?

You may wonder. Is it really worth the trouble? All the template code seems unnecessarily complex!

The answer is: it scales better.

Adding a new type to the factory function is as simple as adding it to the dinasour_types:

using dinasour_types = type_sequence<
    Parasaurolophus // I'm new!

That’s it, a type (and a comma). This is even simpler than the macro solution.

Better yet, make_dinosaur_from and make_dino_impl can be generalized into make_object:

/// Find a type `T` in `Ts` matching `name`, and return std::make_unique<T>().
template <class P, class... Ts>
P make_object(type_sequence<Ts...>, std::string_view name);

Using this generalized make_object, we can create not only Dinosaur, but also Widget, Channel, and so on.

Bonus: Using a map

So far, we’ve been searching the types one by one. Alternatively, we can use a map whose keys are the type names and values are the function pointers:

template <class T>
std::unique_ptr<Dinosaur> make_unique_dino() {
    return std::make_unique<T>();

template <class... Ts>
std::unique_ptr<Dinosaur> make_dinosaur_from(type_sequence<Ts...>, std::string_view name) {
    static std::unordered_map dinosaur_map {
        std::pair {get_typename<Ts>(), &make_unique_dino<Ts>} ...
    if (auto it = dinosaur_map.find(std::string(name)); it != dinosaur_map.end()) {
        auto fn = it->second;
        return fn();
    return nullptr;

std::unique_ptr<Dinosaur> make_dinosaur(std::string_view name) {
    return make_dinosaur_from(dinasour_types{}, name);


We have essentially splitted a function into several smaller ones, each doing a specific job:

  • make_unique_dino: actually making the object, knowing the type
  • make_dinosaur_from: search for the concrete type

It definitely takes more effort to set things up initially, and the rain has stopped long ago!