Poly

Crash Course: poly

Introduction

Static polymorphism is a very powerful tool in C++, albeit sometimes cumbersome to obtain.
This module aims to make it simple and easy to use.

The library allows to define concepts as interfaces to fulfill with concrete classes without having to inherit from a common base.
Among others, this is one of the advantages of static polymorphism in general and of a generic wrapper like that offered by the poly class template in particular.
The result is an object to pass around as such and not through a reference or a pointer, as it happens when it comes to working with dynamic polymorphism.

Since the poly class template makes use of entt::any internally, it also supports most of its feature. For example, the possibility to create aliases to existing and thus unmanaged objects. This allows users to exploit the static polymorphism while maintaining ownership of objects.
Likewise, the poly class template also benefits from the small buffer optimization offered by the entt::any class and therefore minimizes the number of allocations, avoiding them altogether where possible.

Other libraries

There are some very interesting libraries regarding static polymorphism.
The ones that I like more are:

dyno: runtime polymorphism done right.
Poly: a class template that makes it easy to define a type-erasing polymorphic object wrapper.

The former is admittedly an experimental library, with many interesting ideas. I've some doubts about the usefulness of some feature in real world projects, but perhaps my lack of experience comes into play here. In my opinion, its only flaw is the API which I find slightly more cumbersome than other solutions.
The latter was undoubtedly a source of inspiration for this module, although I opted for different choices in the implementation of both the final API and some feature.

Either way, the authors are gurus of the C++ community, people I only have to learn from.

Concept and implementation

The first thing to do to create a type-erasing polymorphic object wrapper (to use the terminology introduced by Eric Niebler) is to define a concept that types will have to adhere to.
For this purpose, the library offers a single class that supports both deduced and fully defined interfaces. Although having interfaces deduced automatically is convenient and allows users to write less code in most cases, it has some limitations and it's therefore useful to be able to get around the deduction by providing a custom definition for the static virtual table.

Once the interface is defined, a generic implementation is needed to fulfill the concept itself.
Also in this case, the library allows customizations based on types or families of types, so as to be able to go beyond the generic case where necessary.

Deduced interface

This is how a concept with a deduced interface is defined:

struct Drawable: entt::type_list<> {
    template<typename Base>
    struct type: Base {
        void draw() { this->template invoke<0>(*this); }
    };

    // ...
};

It's recognizable by the fact that it inherits from an empty type list.
Functions can also be const, accept any number of parameters and return a type other than void:

struct Drawable: entt::type_list<> {
    template<typename Base>
    struct type: Base {
        bool draw(int pt) const { return this->template invoke<0>(*this, pt); }
    };

    // ...
};

In this case, all parameters are passed to invoke after the reference to this and the return value is whatever the internal call returns.
As for invoke, this is a name that is injected into the concept through Base, from which one must necessarily inherit. Since it's also a dependent name, the this-> template form is unfortunately necessary due to the rules of the language. However, there also exists an alternative that goes through an external call:

struct Drawable: entt::type_list<> {
    template<typename Base>
    struct type: Base {
        void draw() const { entt::poly_call<0>(*this); }
    };

    // ...
};

Once the concept is defined, users must provide a generic implementation of it in order to tell the system how any type can satisfy its requirements. This is done via an alias template within the concept itself.
The index passed as a template parameter to either invoke or poly_call refers to how this alias is defined.

Defined interface

A fully defined concept is no different to one for which the interface is deduced, with the only difference that the list of types is not empty this time:

struct Drawable: entt::type_list<void()> {
    template<typename Base>
    struct type: Base {
        void draw() { entt::poly_call<0>(*this); }
    };

    // ...
};

Again, parameters and return values other than void are allowed. Also, the function type must be const when the method to bind to it is const:

struct Drawable: entt::type_list<bool(int) const> {
    template<typename Base>
    struct type: Base {
        bool draw(int pt) const { return entt::poly_call<0>(*this, pt); }
    };

    // ...
};

Why should a user fully define a concept if the function types are the same as the deduced ones?
In fact, this is the limitation that can be worked around by manually defining the static virtual table.

When things are deduced, there is an implicit constraint.
If the concept exposes a member function called draw with function type void(), a concept is satisfied:

Either by a class that exposes a member function with the same name and the same signature.
Or through a lambda that makes use of existing member functions from the interface itself.

In other words, it's not possible to make use of functions not belonging to the interface, even if they're part of the types that fulfill the concept.
Similarly, it's not possible to deduce a function in the static virtual table with a function type different from that of the associated member function in the interface itself.

Explicitly defining a static virtual table suppresses the deduction step and allows maximum flexibility when providing the implementation for a concept.

Fulfill a concept

The impl alias template of a concept is used to define how it's fulfilled:

struct Drawable: entt::type_list<> {
    // ...

    template<typename Type>
    using impl = entt::value_list<&Type::draw>;
};

In this case, it's stated that the draw method of a generic type is enough to satisfy the requirements of the Drawable concept.
Both member functions and free functions are supported to fulfill concepts:

template<typename Type>
void print(Type &self) { self.print(); }

struct Drawable: entt::type_list<void()> {
    // ...

    template<typename Type>
    using impl = entt::value_list<&print<Type>>;
};

Likewise, as long as the parameter types and return type support conversions to and from those of the function type referenced in the static virtual table, the actual implementation may differ in its function type since it's erased internally.
Moreover, the self parameter isn't strictly required by the system and can be left out for free functions if not required.

Refer to the inline documentation for more details.

Inheritance

Concept inheritance is straightforward due to how poly looks like in EnTT. Therefore, it's quite easy to build hierarchies of concepts if necessary.
The only constraint is that all concepts in a hierarchy must belong to the same family, that is, they must be either all deduced or all defined.

For a deduced concept, inheritance is achieved in a few steps:

struct DrawableAndErasable: entt::type_list<> {
    template<typename Base>
    struct type: typename Drawable::type<Base> {
        static constexpr auto base = Drawable::impl<Drawable::type<entt::poly_inspector>>::size;
        void erase() { entt::poly_call<base + 0>(*this); }
    };

    template<typename Type>
    using impl = entt::value_list_cat_t<
        typename Drawable::impl<Type>,
        entt::value_list<&Type::erase>
    >;
};

The static virtual table is empty and must remain so.
On the other hand, type no longer inherits from Base. Instead, it forwards its template parameter to the type exposed by the base class. Internally, the size of the static virtual table of the base class is used as an offset for the local indexes.
Finally, by means of the value_list_cat_t utility, the implementation consists in appending the new functions to the previous list.

As for a defined concept instead, the list of types is extended in a similar way to what is shown for the implementation of the above concept.
To do this, it's useful to declare a function that allows to convert a concept into its underlying type_list object:

template<typename... Type>
entt::type_list<Type...> as_type_list(const entt::type_list<Type...> &);

The definition isn't strictly required, since the function is only used through a decltype as it follows:

struct DrawableAndErasable: entt::type_list_cat_t<
    decltype(as_type_list(std::declval<Drawable>())),
    entt::type_list<void()>> {
    // ...
};

Similar to above, type_list_cat_t is used to concatenate the underlying static virtual table with the new function types.
Everything else is the same as already shown instead.

Static polymorphism in the wild

Once the concept and implementation are defined, it's possible to use the poly class template to wrap instances that meet the requirements:

using drawable = entt::poly<Drawable>;

struct circle {
    void draw() { /* ... */ }
};

struct square {
    void draw() { /* ... */ }
};

// ...

drawable instance{circle{}};
instance->draw();

instance = square{};
instance->draw();

This class offers a wide range of constructors, from the default one (which returns an uninitialized poly object) to the copy and move constructors, as well as the ability to create objects in-place.
Among others, there is also a constructor that allows users to wrap unmanaged objects in a poly instance (either const or non-const ones):

circle shape;
drawable instance{std::in_place_type<circle &>, shape};

Similarly, it's possible to create non-owning copies of poly from an existing object:

drawable other = instance.as_ref();

In both cases, although the interface of the poly object doesn't change, it doesn't construct any element or take care of destroying the referenced objects.

Note also how the underlying concept is accessed via a call to operator-> and not directly as instance.draw().
This allows users to decouple the API of the wrapper from that of the concept. Therefore, where instance.data() invokes the data member function of the poly object, instance->data() maps directly to the functionality exposed by the underlying concept.

Storage size and alignment requirement

Under the hood, the poly class template makes use of entt::any. Therefore, it can take advantage of the possibility of defining at compile-time the size of the storage suitable for the small buffer optimization as well as the alignment requirements:

entt::basic_poly<Drawable, sizeof(double[4]), alignof(double[4])>

The default size is sizeof(double[2]), which seems like a good compromise between a buffer that is too large and one unable to hold anything larger than an integer. The alignment requirement is optional and by default such that it's the most stringent (the largest) for any object whose size is at most equal to the one provided.
It's worth noting that providing a size of 0 (which is an accepted value in all respects) will force the system to dynamically allocate the contained objects in all cases.