A long time ago, in a post far far away… we seen how constexpr can be used to retrieve and store type names at compile time, so C++ types can be tagged and indexed without relying on RTTI.

The goal of this series is to build a simple reflection engine. Since the engine is dynamic, we need a way to handle objects of heterogenous types at runtime. A kind of Boost.Any, but focused on our future reflection needs.

Manu unchained

There are situations where I would like to live in that wonderful world of JavaScript, where everybody can become everything they want: A string, an integer, an elephant… Sounds great, but sadly I’m living as a slave of the C++ type system.

Thankfully C programmers were getting rid of the C type system since the very birth of the language, doing safe generic programming in terms of void*. For us, void* is a simple way to instance and reference any kind of C++ object in an homogeneous manner, regardless of its type:

template<typename T>
void* create()
{
    return new T();
}

template<typename T>
void destroy(void* object)
{
    delete reinterpret_cast<T*>(object);
}

int main()
{
    void* intObject = create<int>();
    void* boolObject = create<bool>();

    // Welcome to Java
    std::vector<void*> arrayList{intObject, boolObject};

    destroy<int>(intObject);
    destroy<bool>(boolObject);
}

Simple, like most C-like APIs look like. It has some caveats though:

Anyway, the thing still holds. Let’s see how we can improve this simple idea with lightweight C++ abstractions.

Metatype

The first issue with the simple API above is that by referencing everything through a void* we have lost that precious information that a type system gives to us. Well, not exactly. Maybe the compiler had lost that information, but we know exactly what type was used and could access that info at runtime.

The first abstraction on the object API is a class in charge of manipulating objects of an specific type at runtime. That is, a class that represents a C++ type at runtime. I call this class MetaType:

class MetaType
{
public:
    void* create();
    void destroy(void* object);
    const TypeInfo& typeInfo() const;

private:
    TypeInfo _typeInfo;
};

bool operator==(const MetaType& lhs, const MetaType& rhs);
bool operator!=(const MetaType& lhs, const MetaType& rhs);

A MetaType variable represents a particular instance of the API above, keeping track of the type being used with that particular API:

MetaType intType{int}, boolType{bool};
assert(intType != boolType);

void* intObject = intType.create();
void* boolObject = boolType.create();

...

intType.destroy(intObject);
boolType.destroy(boolObject);

But Manu, that doesn’t event compile. Ok, ok; I was cheating you. Sadly C++ is not (yet) that cool, so we have to do some extra work. We have to find a way to store all type behaviors we need (By behaviors I mean how we manipulate instances of a type), but all that behaviors should be kept in variables of the same type. void* is a way of type punning as we’ve seen, but C++ has another cool one: Inheritance-based polymorphism:

class MetaType
{
    ...

private:
    class TypeBehavior
    {
    public:
        TypeBehavior(const TypeInfo& typeInfo);
        virtual ~TypeBehavior() = default;

        const TypeInfo& typeInfo() const;

        virtual void* create() = 0;
        virtual void destroy(void* object) = 0;

    private:
        TypeInfo _typeInfo;
    };

    std::shared_ptr<TypeBehavior> _behavior;

    template<typename T>
    class TypeBehaviorFor : public TypeBehavior
    {
    public:
        TypeBehaviorFor() : TypeBehavior{TypeInfo::get<T>()}
        {}

        void* create() override
        {
            return new T();
        }

        void destroy(void* object) override
        {
            delete reinterpret_cast<T*>(object);
        }
    };
};

My beloved templates to the rescue. What we have done here is to define a common interface for type behaviors, then use a template to define the behavior of each requested (instanced) type.

TypeBehavior class above is rather minimalistic, you may want to instance objects given constructor parameters, assign objects, copy them, etc. The post only shows the most simple interface, but the real world code both the reader and I would write includes more overloads of create() method and others such as copy(), assign(), etc.

Also note I’m calling raw new/delete. Of course this is not optimal but simple for the post too. I suggest you to use an object pool instead. With a pool this works great since most allocations are just recycling old dead objects.

Now any MetaType object just asks its internal behavior about what to do:

void* MetaType::create()
{
    return _behavior->create();
}

void MetaType::destroy(void* object)
{
    return _behavor->destroy(object);
}

const TypeInfo& MetaType::typeInfo() const
{
    return _behavior->typeInfo();
}

bool operator==(const MetaType& lhs, const MetaType& rhs)
{
    return lhs.typeInfo() == rhs.typeInfo();
}

bool operator!=(const MetaType& lhs, const MetaType& rhs)
{
    return !(lhs == rhs);
}

Finally write the usual factory template and we are done:

class MetaType
{
public:
    template<typename T>
    MetaType get()
    {
        return {std::make_shared<TypeBehaviorFor<T>>()};
    }

    ...

private:
    MetaType(const std::shared_ptr<TypeBehavior>& behavior) :
        _behavior{behavior}
    {}

    ...
};

The user code looks more or less like this:

auto intType = MetaType::get<int>();

void* integer = intType.create();
...
intType.destroy(integer);

Metatype for reflection

MetaType class shown above is the minimal needed to write something in the lines of Boost.Any. But for dynamic reflection one may need more features, specifically, asking for C++ types given their name:

void* createObject(const std::string& typeName)
{
    auto type = MetaType::get(typeName);
    return type.create();
}

To do that, MetaType should be able to register known types. The simplest way by holding a typeName -> behavior table as part of MetaType class:

class MetaType
{
public:
    template<typename T>
    MetaType get()
    {
        static std::shared_ptr<TypeBehavior> behavior = []
        {
            auto behavior = std::make_shared<TypeBehaviorFor<T>>();
            _registry[ctti::unnamed_type_id<T>()] = behavior;
            return behavior;
        };

        return {behavior};
    }

    MetaType get(const std::string& typeName)
    {
        return {_registry[ctti::id_from_name(typeName)]};
    }

    ...

private:
    std::shared_ptr<TypeBehavior> _behavior;
    ...
    static std::unordered_map<
        ctti::unnamed_type_id_t,
        std::shared_ptr<TypeBehavior>
    > _registry;
}

Now as long as you’ve statically used a type once, you would be able to ask for it by name:

MetaType::get<int>(); // "register" the type

auto intType = MetaType::get("int");
void* integer = intType.create();
intType.destroy(integer);

If you ever worked with C++ libraries interacting with dynamic type systems (Boost.Python and luabind come to my mind), you might have noticed those always involve some kind of “registration” step. That registration is a bootstrapping process where the library sets up all the information that it needs to map, among other things, the C++ static type system to its dynamic counterparts. Is in this boring registration code were an external tool may help a lot. But that’s a topic for following posts on libclang API…

MetaObject

So far we’ve solved one of our two problems with the simple C-like object API. But we still have to create and destroy objects manually (And don’t forget to use the correct MetaType! Was the first problem solved at all?!!).

This is solved by reintroducing objects into the RAII principle, wrapping them with a handle class in charge of managing their lifetime by means of constructors and destructors:

class MetaObject // Aka "Any"
{
public:
    template<typename T>
    MetaObject(T&& value);

    template<typename T>
    const T& get() const
    template<typename T>
    T& get();

    template<typename T>
    operator const T&() const;
    template<typename T>
    operator T&();

    const MetaType& type() const;

private:
    MetaType _type;
    std::shared_ptr<void> _object;
};

MetaObject class keeps together both the object and its type, so objects are always manipulated using the right type. A couple of get() methods give access to the underlying object:

template<typename T>
const T& MetaObject::get() const
{
    assert(MetaType::get<T>() == _type);
    return *reinterpret_cast<const T*>(_object.get());
}

template<typename T>
T& MetaObject::get()
{
    assert(MetaType::get<T>() == _type);
    return *reinterpret_cast<T*>(_object.get());
}

Usage of metaobjects is simplified by automatic conversion operators:

template<typename T>
MetaObject::operator const T&() const
{
    return get<T>();
}

template<typename T>
MetaObject::operator T&()
{
    return get<T>();
}
int add(int a, int b)
{
    return a + b;
}

int main()
{
    MetaObject a{1}, b{2};

    f(a, b);
}

“Simplified”. Many people don’t like automatic conversions, but I find them useful in situations like this to make user code more pleasant to write and read. I always try to be pragmatic instead of dogmatic and pedant, but I’m sure I don’t always act this way. In fact many of my best friends are sick of me, in their words, “Oh shit let’s run away, he’s starting a pedant C++ discourse again!” while having beers after the work/university… I’m sorry guys :)

A std::shared_ptr is used to manage object lifetime. We could have written a custom destructor, constructors, and assignment operators; fut I feel simpler to use a standar library smart pointer with a custom deleter:

class Deleter
{
public:
    Deleter(const MetaType& type) :
        _type{type}
    {}

    void operator()(void* object)
    {
        _type.destroy(object);
    }

private:
    MetaType _type;
};

template<typename T>
MetaObject::MetaObject(T&& value) :
    _type{MetaType::get<std::decay_t<T>>()},
    _object{
        _type.create(std::forward<T>(value)),
        Deleter{_type}
    }
{}

AFAIK std::shared_ptr doesn’t give public access to its deleter, so I had to store type twice (One for the MetaObject checks and other for the deleter).

Conclusion

Having a generic type gives us the chance to manipulate any C++ object in an homogeneous way, which will be very useful when accessing and manipulating elements of classes with dynamic reflection features.

In next posts we will introduce the reflection runtime model, with the classes needed to access class member functions, class fields, and finally, a registry with all class information.

Appendix: Playing with metaobjects

The conversion operators defined as part of MetaObject makes simple to pass values back and forth this type, like we’ve seen in the example:

MetaObject integer{1};
int i = integer;

On top of these we can write some interesting tools:

Heterogeneous vector container

Thanks to generic types std::vector could work as a container of heterogeneous values. Like a type erased std::tuple:

template<typename... Ts>
std::vector<MetaObject> pack_to_vector(Ts&&... values)
{
    return { MetaObject{std::forward<Ts>(values)}... };
}

Using the indices trick a tuple alternative looks like this:

template<typename... Ts, std::size_t... Is>
std::vector<MetaObject> tuple_to_vector(const std::tuple<Ts...>& tuple,
                                        std::index_sequence<Is...>)
{
    return pack_to_vector(std::get<Is>(tuple)...);
}

template<typenbame... Ts>
std::vector<MetaObject> tuple_to_vector(const std::tuple<Ts...>& tuple)
{
    return tuple_to_vector(tuple, std::make_index_sequence_for<Ts...>{});
}

You might want to add two overloads (tuple lvalue and tuple rvalue) to reduce copies.

A vector_to_tuple() function is left as an exercise for the reader.

Invoking functions

Following with the heterogeneous std::vector container, we could write an utility to invoke functions using a vector of arguments, in the line of tuple_call():

template<typename F, std::size_t... Is>
auto vector_call(const std::vector<MetaObject>& args, F function,
                 std::index_sequence<Is...>)
{
    return function(args[Is].get<function_argument_t<Is, F>>()...);
}

template<typename F>
auto vector_call(const std::vector<MetaObject>& args, F function)
{
    return vector_call(
        args,
        function,
        std::make_index_sequence<function_arguments<F>::size>{}
    );
}

Here we query the function signature expand vector elements to the corresponding function argument types.

Getting the signature if a function in a generic way is not a simple task, C++ supports many flavors of function-like entities. This interesting topic deserves an entire post (Or a talk for a future Madrid C++ meetup as one our members suggested!). Just for completeness I will leave here a link to the utility header being used in my reflection engine: #include <siplasplas/utility/[function_traits.hpp] (https://github.com/GueimUCM/siplasplas/blob/master/include/siplasplas/utility/function_traits.hpp)>

Given the function signature (A type list with all the function arguments types), to invoke the function with the vector we “unpack” it by getting the value of each MetaObject using the corresponding type.

Having fun

All the tools above can be combined to do some interesing things. Here’s a little spoiler/example: As part of the reflection engine I wrote an attribute API for functions inspired in C# attributes syntax.

That API is based in an Attribute interface as follows:

class Attribute
{
public:
    virtual std::vector<MetaObject> processArguments(const std::vector<MetaObject>& args) = 0;
    virtual MetaObject processReturnValue(const MetaObject& value) = 0;
};

Attribute instances are injected before and after function invocation:

ReturnType invoke(const std::vector<MetaObject>& args, Attribute& attr, F f)
{
    auto processedArgs = attr.processArguments(args);
    auto returnValue = vector_call(processedArgs, f);
    return attr.processReturnValue(returnValue);
}

This works great to add decorators to member functions of a class, such as precondition checking or logging function calls. You can also troll your users by changing the sign of integer arguments, clearing strings, etc. You’re free.

That’s enough. More on attributes in following posts!