Dyno: Runtime polymorphism done right

DISCLAIMER

At this point, this library is experimental and it is a pure curiosity. No stability of interface or quality of implementation is guaranteed. Use at your own risks.

Overview

Dyno solves the problem of runtime polymorphism better than vanilla C++ does. It provides a way to define interfaces that can be fulfilled non-intrusively, and it provides a fully customizable way of storing polymorphic objects and dispatching to virtual methods. It does not require inheritance, heap allocation or leaving the comfortable world of value semantics, and it can do so while outperforming vanilla C++.

Dyno is pure-library implementation of what's also known as Rust trait objects, Go interfaces, Haskell type classes, and virtual concepts. Under the hood, it uses a C++ technique known as type erasure, which is the idea behind std::any, std::function and many other useful types.

#include <dyno.hpp>
#include <iostream>
using namespace dyno::literals;

// Define the interface of something that can be drawn
struct Drawable : decltype(dyno::requires_(
  "draw"_s = dyno::method<void (std::ostream&) const>
)) { };

// Define how concrete types can fulfill that interface
template <typename T>
auto const dyno::default_concept_map<Drawable, T> = dyno::make_concept_map(
  "draw"_s = [](T const& self, std::ostream& out) { self.draw(out); }
);

// Define an object that can hold anything that can be drawn.
struct drawable {
  template <typename T>
  drawable(T x) : poly_{x} { }

  void draw(std::ostream& out) const
  { poly_.virtual_("draw"_s)(out); }

private:
  dyno::poly<Drawable> poly_;
};

struct Square {
  void draw(std::ostream& out) const { out << "Square"; }
};

struct Circle {
  void draw(std::ostream& out) const { out << "Circle"; }
};

void f(drawable const& d) {
  d.draw(std::cout);
}

int main() {
  f(Square{}); // prints Square
  f(Circle{}); // prints Circle
}

Alternatively, if you find this to be too much boilerplate and you can stand using a macro, the following is equivalent:

#include <dyno.hpp>
#include <iostream>

// Define the interface of something that can be drawn
DYNO_INTERFACE(Drawable,
  (draw, void (std::ostream&) const)
);

struct Square {
  void draw(std::ostream& out) const { out << "Square"; }
};

struct Circle {
  void draw(std::ostream& out) const { out << "Circle"; }
};

void f(Drawable const& d) {
  d.draw(std::cout);
}

int main() {
  f(Square{}); // prints Square
  f(Circle{}); // prints Circle
}

Compiler requirements

This is a C++17 library. No efforts will be made to support older compilers (sorry). The library is known to work with the following compilers:

Compiler	Version
GCC	>= 7
Clang	>= 4.0
Apple Clang	>= 9.1

Dependencies

The library depends on Boost.Hana and Boost.CallableTraits. The unit tests depend on libawful and the benchmarks depend on Google Benchmark, Boost.TypeErasure and Mpark.Variant, but you don't need them to use the library. For local development, the dependencies/install.sh script can be used to install all the dependencies automatically.

Building the library

Dyno is a header-only library, so there's nothing to build per-se. Just add the include/ directory to your compiler's header search path (and make sure the dependencies are satisfied), and you're good to go. However, there are unit tests, examples and benchmarks that can be built:

(cd dependencies && ./install.sh) # Install dependencies; will print a path to add to CMAKE_PREFIX_PATH
mkdir build
(cd build && cmake .. -DCMAKE_PREFIX_PATH="${PWD}/../dependencies/install") # Setup the build directory

cmake --build build --target examples   # Build and run the examples
cmake --build build --target tests      # Build and run the unit tests
cmake --build build --target check      # Does both examples and tests
cmake --build build --target benchmarks # Build and run the benchmarks

Introduction

In programming, the need for manipulating objects with a common interface but with a different dynamic type arises very frequently. C++ solves this with inheritance:

struct Drawable {
  virtual void draw(std::ostream& out) const = 0;
};

struct Square : Drawable {
  virtual void draw(std::ostream& out) const override final { ... }
};

struct Circle : Drawable {
  virtual void draw(std::ostream& out) const override final { ... }
};

void f(Drawable const* drawable) {
  drawable->draw(std::cout);
}

However, this approach has several drawbacks. It is

1. Intrusive
   In order for Square and Circle to fulfill the Drawable interface, they both need to inherit from the Drawable base class. This requires having the license to modify those classes, which makes inheritance very inextensible. For example, how would you make a std::vector<int> fulfill the Drawable interface? You simply can't.

2. Incompatible with value semantics
   Inheritance requires you to pass polymorphic pointers or references to objects instead of the objects themselves, which plays very badly with the rest of the language and the standard library. For example, how would you copy a vector of Drawables? You'd need to provide a virtual clone() method, but now you've just messed up your interface.

3. Tightly coupled with dynamic storage
   Because of the lack of value semantics, we usually end up allocating these polymorphic objects on the heap. This is both horribly inefficient and semantically wrong, since chances are we did not need the dynamic storage duration at all, and an object with automatic storage duration (e.g. on the stack) would have been enough.

4. Prevents inlining
   95% of the time, we end up calling a virtual method through a polymorphic pointer or reference. That requires three indirections: one for loading the pointer to the vtable inside the object, one for loading the right entry in the vtable, and one for the indirect call to the function pointer. All this jumping around makes it difficult for the compiler to make good inlining decisions. However, it turns out that all of these indirections except the indirect call can be avoided.

Unfortunately, this is the choice that C++ has made for us, and these are the rules that we are bound to when we need dynamic polymorphism. Or is it really?

So, what is this library?

Dyno solves the problem of runtime polymorphism in C++ without any of the drawbacks listed above, and many more goodies. It is:

1. Non-intrusive
   An interface can be fulfilled by a type without requiring any modification to that type. Heck, a type can even fulfill the same interface in different ways! With Dyno, you can kiss ridiculous class hierarchies goodbye.

2. 100% based on value semantics
   Polymorphic objects can be passed as-is, with their natural value semantics. You need to copy your polymorphic objects? Sure, just make sure they have a copy constructor. You want to make sure they don't get copied? Sure, mark it as deleted. With Dyno, silly clone() methods and the proliferation of pointers in APIs are things of the past.

3. Not coupled with any specific storage strategy
   The way a polymorphic object is stored is really an implementation detail, and it should not interfere with the way you use that object. Dyno gives you complete control over the way your objects are stored. You have a lot of small polymorphic objects? Sure, let's store them in a local buffer and avoid any allocation. Or maybe it makes sense for you to store things on the heap? Sure, go ahead.

4. Flexible dispatch mechanism to achieve best possible performance
   Storing a pointer to a vtable is just one of many different implementation strategies for performing dynamic dispatch. Dyno gives you complete control over how dynamic dispatch happens, and can in fact beat vtables in some cases. If you have a function that's called in a hot loop, you can for example store it directly in the object and skip the vtable indirection. You can also use application-specific knowledge the compiler could never have to optimize some dynamic calls — library-level devirtualization.

Using the library

First, you start by defining a generic interface and giving it a name. Dyno provides a simple domain specific language to do that. For example, let's define an interface Drawable that describes types that can be drawn:

#include <dyno.hpp>
using namespace dyno::literals;

struct Drawable : decltype(dyno::requires_(
  "draw"_s = dyno::method<void (std::ostream&) const>
)) { };

This defines Drawable as representing an interface for anything that has a method called draw taking a reference to a std::ostream. Dyno calls these interfaces dynamic concepts, since they describe sets of requirements to be fulfilled by a type (like C++ concepts). However, unlike C++ concepts, these dynamic concepts are used to generate runtime interfaces, hence the name dynamic. The above definition is basically equivalent to the following:

struct Drawable {
  virtual void draw(std::ostream&) const = 0;
};

Once the interface is defined, the next step is to actually create a type that satisfies this interface. With inheritance, you would write something like this:

struct Square : Drawable {
  virtual void draw(std::ostream& out) const override final {
    out << "square" << std::endl;
  }
};

With Dyno, the polymorphism is non-intrusive and it is instead provided via what is called a concept map (after C++0x Concept Maps):

struct Square { /* ... */ };

template <>
auto const dyno::concept_map<Drawable, Square> = dyno::make_concept_map(
  "draw"_s = [](Square const& square, std::ostream& out) {
    out << "square" << std::endl;
  }
);

This construct is the specialization of a C++14 variable template named concept_map defined in the dyno:: namespace. We then initialize that specialization with dyno::make_concept_map(...).

The first parameter of the lambda is the implicit *this parameter that is implied when we declared draw as a method above. It's also possible to erase non-member functions (see the relevant section).

This concept map defines how the type Square satisfies the Drawable concept. In a sense, it maps the type Square to its implementation of the concept, which motivates the appellation. When a type satisfies the requirements of a concept, we say that the type models (or is a model of) that concept. Now that Square is a model of the Drawable concept, we'd like to use a Square polymorphically as a Drawable. With traditional inheritance, we would use a pointer to a base class like this:

void f(Drawable const* d) {
  d->draw(std::cout);
}

f(new Square{});

With Dyno, polymorphism and value semantics are compatible, and the way polymorphic types are passed around can be highly customized. To do this, we'll need to define a type that can hold anything that's Drawable. It is that type, instead of a Drawable*, that we'll be passing around to and from polymorphic functions. To help define this wrapper, Dyno provides the dyno::poly container, which can hold an arbitrary object satisfying a given concept. As you will see, dyno::poly has a dual role: it stores the polymorphic object and takes care of the dynamic dispatching of methods. All you need to do is write a thin wrapper over dyno::poly to give it exactly the desired interface:

struct drawable {
  template <typename T>
  drawable(T x) : poly_{x} { }

  void draw(std::ostream& out) const
  { poly_.virtual_("draw"_s)(out); }

private:
  dyno::poly<Drawable> poly_;
};

Note: You could technically use dyno::poly directly in your interfaces. However, it is much more convenient to use a wrapper with real methods than dyno::poly, and so writing a wrapper is recommended.

Let's break this down. First, we define a member poly_ that is a polymorphic container for anything that models the Drawable concept:

dyno::poly<Drawable> poly_;

Then, we define a constructor that allows constructing this container from an arbitrary type T:

template <typename T>
drawable(T x) : poly_{x} { }

The unsaid assumption here is that T actually models the Drawable concept. Indeed, when you create a dyno::poly from an object of type T, Dyno will go and look at the concept map defined for Drawable and T, if any. If there's no such concept map, the library will report that we're trying to create a dyno::poly from a type that does not support it, and your program won't compile.

Finally, the strangest and most important part of the definition above is that of the draw method:

void draw(std::ostream& out) const
{ poly_.virtual_("draw"_s)(out); }

What happens here is that when .draw is called on our drawable object, we'll actually perform a dynamic dispatch to the implementation of the "draw" function for the object currently stored in the dyno::poly, and call that. Now, to create a function that accepts anything that's Drawable, no need to worry about pointers and ownership in your interface anymore:

void f(drawable d) {
  d.draw(std::cout);
}

f(Square{});

By the way, if you're thinking that this is all stupid and you should have been using a template, you're right. However, consider the following, where you really do need runtime polymorphism:

drawable get_drawable() {
  if (some_user_input())
    return Square{};
  else
    return Circle{};
}

f(get_drawable());

Strictly speaking, you don't need to wrap dyno::poly, but doing so puts a nice barrier between Dyno and the rest of your code, which never has to worry about how your polymorphic layer is implemented. Also, we largely ignored how dyno::poly was implemented in the above definition. However, dyno::poly is a very powerful policy-based container for polymorphic objects that can be customized to one's needs for performance. Creating a drawable wrapper makes it easy to tweak the implementation strategy used by dyno::poly for performance without impacting the rest of your code.

Customizing the polymorphic storage

The first aspect that can be customized in a dyno::poly is the way the object is stored inside the container. By default, we simply store a pointer to the actual object, like one would do with inheritance-based polymorphism. However, this is often not the most efficient implementation, and that's why dyno::poly allows customizing it. To do so, simply pass a storage policy to dyno::poly. For example, let's define our drawable wrapper so that it tries to store objects up to 16 bytes in a local buffer, but then falls back to the heap if the object is larger:

struct drawable {
  template <typename T>
  drawable(T x) : poly_{x} { }

  void draw(std::ostream& out) const
  { poly_.virtual_("draw"_s)(out); }

private:
  dyno::poly<Drawable, dyno::sbo_storage<16>> poly_;
  //                   ^^^^^^^^^^^^^^^^^^^^^ storage policy
};

Notice that nothing except the policy changed in our definition. That is one very important tenet of Dyno; these policies are implementation details, and they should not change the way you write your code. With the above definition, you can now create drawables just like you did before, and no allocation will happen when the object you're creating the drawable from fits in 16 bytes. When it does not fit, however, dyno::poly will allocate a large enough buffer on the heap.

Let's say you actually never want to do an allocation. No problem, just change the policy to dyno::local_storage<16>. If you try to construct a drawable from an object that's too large to fit in the local storage, your program won't compile. Not only are we saving an allocation, but we're also saving a pointer indirection every time we access the polymorphic object if we compare to the traditional inheritance-based approach. By tweaking these (important) implementation details for you specific use case, you can make your program much more efficient than with classic inheritance.

Other storage policies are also provided, like dyno::remote_storage and dyno::non_owning_storage. dyno::remote_storage is the default one, which always stores a pointer to a heap-allocated object. dyno::non_owning_storage stores a pointer to an object that already exists, without worrying about the lifetime of that object. It allows implementing non-owning polymorphic views over objects, which is very useful.

Custom storage policies can also be created quite easily. See <dyno/storage.hpp> for details.

Customizing the dynamic dispatch

When we introduced dyno::poly, we mentioned that it had two roles; the first is to store the polymorphic object, and the second one is to perform dynamic dispatch. Just like the storage can be customized, the way dynamic dispatching is performed can also be customized using policies. For example, let's define our drawable wrapper so that instead of storing a pointer to the vtable, it instead stores the vtable in the drawable object itself. This way, we'll avoid one indirection each time we access a virtual function:

struct drawable {
  template <typename T>
  drawable(T x) : poly_{x} { }

  void draw(std::ostream& out) const
  { poly_.virtual_("draw"_s)(out); }

private:
  using Storage = dyno::sbo_storage<16>;                      // storage policy
  using VTable = dyno::vtable<dyno::local<dyno::everything>>; // vtable policy
  dyno::poly<Drawable, Storage, VTable> poly_;
};

Notice that nothing besides the vtable policy needs to change in the definition of our drawable type. Furthermore, if we wanted, we could change the storage policy independently from the vtable policy. With the above, even though we are saving all indirections, we are paying for it by making our drawable object larger (since it needs to hold the vtable locally). This could be prohibitive if we had many functions in the vtable. Instead, it would make more sense to store most of the vtable remotely, but only inline those few functions that we call heavily. Dyno makes it very easy to do so by using Selectors, which can be used to customize what functions a policy applies to:

struct drawable {
  template <typename T>
  drawable(T x) : poly_{x} { }

  void draw(std::ostream& out) const
  { poly_.virtual_("draw"_s)(out); }

private:
  using Storage = dyno::sbo_storage<16>;
  using VTable = dyno::vtable<
    dyno::local<dyno::only<decltype("draw"_s)>>,
    dyno::remote<dyno::everything_else>
  >;
  dyno::poly<Drawable, Storage, VTable> poly_;
};

Given this definition, the vtable is actually split in two. The first part is local to the drawable object and contains only the draw method. The second part is a pointer to a vtable in static storage that holds the remaining methods (the destructor, for example).

Dyno provides two vtable policies, dyno::local<> and dyno::remote<>. Both of these policies must be customized using a Selector. The selectors supported by the library are dyno::only<functions...>, dyno::except<...>, and dyno::everything_else (which can also be spelled dyno::everything).

Defaulted concept maps

When defining a concept, it is often the case that one can provide a default definition for at least some functions associated to the concept. For example, by default, it would probably make sense to use a member function named draw (if any) to implement the abstract "draw" method of the Drawable concept. For this, one can use dyno::default_concept_map:

template <typename T>
auto const dyno::default_concept_map<Drawable, T> = dyno::make_concept_map(
  "draw"_s = [](auto const& self, std::ostream& out) { self.draw(out); }
);

Now, whenever we try to look at how some type T fulfills the Drawable concept, we'll fall back to the default concept map if no concept map was defined. For example, we can create a new type Circle:

struct Circle {
  void draw(std::ostream& out) const {
    out << "circle" << std::endl;
  }
};

f(Circle{}); // prints "circle"

Circle is automatically a model of Drawable, even though we did not explicitly define a concept map for Circle. On the other hand, if we were to define such a concept map, it would have precedence over the default one:

template <>
auto dyno::concept_map<Drawable, Circle> = dyno::make_concept_map(
  "draw"_s = [](Circle const& circle, std::ostream& out) {
    out << "triangle" << std::endl;
  }
);

f(Circle{}); // prints "triangle"

Parametric concept maps

It is sometimes useful to define a concept map for a complete family of types all at once. For example, we might want to make std::vector<T> a model of Drawable, but only when T can be printed to a stream. This is easily achieved by using this (not so) secret trick:

template <typename T>
auto const dyno::concept_map<Drawable, std::vector<T>, std::void_t<decltype(
  std::cout << std::declval<T>()
)>> = dyno::make_concept_map(
  "draw"_s = [](std::vector<T> const& v, std::ostream& out) {
    for (auto const& x : v)
      out << x << ' ';
  }
);

f(std::vector<int>{1, 2, 3}) // prints "1 2 3 "

Notice how we do not have to modify std::vector at all. How could we do this with classic polymorphism? Answer: no can do.

Erasing non-member functions

Dyno allows erasing non-member functions and functions that are dispatched on an arbitrary argument (but only one argument) too. To do this, simply define the concept using dyno::function instead of dyno::method, and use the dyno::T placeholder to denote the argument being erased:

// Define the interface of something that can be drawn
struct Drawable : decltype(dyno::requires_(
  "draw"_s = dyno::function<void (dyno::T const&, std::ostream&)>
)) { };

The dyno::T const& parameter used above represents the type of the object on which the function is being called. However, it does not have to be the first parameter:

struct Drawable : decltype(dyno::requires_(
  "draw"_s = dyno::function<void (std::ostream&, dyno::T const&)>
)) { };

The fulfillment of the concept does not change whether the concept uses a method or a function, but make sure that the parameters of your function implementation match that of the function declared in the concept:

// Define how concrete types can fulfill that interface
template <typename T>
auto const dyno::default_concept_map<Drawable, T> = dyno::make_concept_map(
  "draw"_s = [](std::ostream& out, T const& self) { self.draw(out); }
  //            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ matches the concept definition
);

Finally, when calling a function on a dyno::poly, you'll have to pass in all the parameters explicitly, since Dyno can't guess which one you want to dispatch on. The parameter that was declared with a dyno::T placeholder in the concept should be passed the dyno::poly itself:

// Define an object that can hold anything that can be drawn.
struct drawable {
  template <typename T>
  drawable(T x) : poly_{x} { }

  void draw(std::ostream& out) const
  { poly_.virtual_("draw"_s)(out, poly_); }
  //                              ^^^^^ passing the poly explicitly

private:
  dyno::poly<Drawable> poly_;
};