# C++ Template specialisation via base class

This topic is 827 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

## Recommended Posts

So as the title (hopefully somewhat) says, I'm trying to achieve a spezialisation of a template class, based on whether or not the template-parameter is derived off of another (templated) class:

// base class, specializations must match this signature
template<typename Object>
class ObjectSupplier
{
public:

static constexpr size_t objectOffset = 1;

static Object& GetClassObject(CallState& state)
{
return *state.GetAttribute<Object>(0);
}
};

// possible specialisation for all "Data"-classes,
// which are actually derived classes using CRTP
template<typename Derived>
class ObjectSupplier<ecs::Data<Derived>>
{
public:

static constexpr size_t objectOffset = 1;

static Derived& GetClassObject(CallState& state)
{
return state.GetEntityAttribute(0)->GetData<Derived>();
}
};

// ... now here's the problem:

// this would work ...
ObjectSupplier<ecs::Data<Transform>>::GetClassObject(state);
// unfornately the actual object is "Transform", which is merely derived
// of "ecs::Data<Transform>" and thus it calls the incorrect overload
ObjectSupplier<Transform>::GetClassObject(state); 

The last two lines show the actual problem. I'm using this ObjectSupplier-class as part of my script-binding system, and thus its not possible/feasable to input the required base-class manually:

template<typename Return, typename Class, typename... Args>
void registerBindFunction(ClassFunction<Return, Class, Args...> pFunction, sys::StringView strFilter, std::wstring&& strDefinition)
{
using Supplier = ObjectSupplier<Class>;
using RegistryEntry = BindFunctionRegistryEntry<ClassFunctionBindCommand<Supplier, Class, Return, Args...>>;

registerClassBindEntry<RegistryEntry, Supplier, Return, Args...>(pFunction, false, strFilter, std::move(strDefinition), DEFAULT_INPUT, DEFAULT_OUTPUT);
}

registerBindFunction(&GenericObject::Func, ...);
// needs to access the ObjectSupplier<ecs::Data<Derived>> overload
registerBindFunction(&Transform::GetAbsoluteTransform, ...);

// thats how it used to be before:

registerObjectBindFunction(&GenericObject::Func, ...);
// a few overloads that internally used a "EntityDataSupplier"-class
registerEntityDataBindFunction(&GenericObject::GetAbsoluteTransform, ...); 

(well, it were possible, but I want this binding-code to be as simple as humanly possible; which is why I'm trying to not have to manually specify anything other than "I want to bind a function here").

So, thats the gist of it. Any ideas on how to get this to work? I don't want to (again) have to create manual overloads for "registerBindFunctions", which there would have to be at least 5 (and all have a complex signature); but I'm certainly open to different approaches, if what I'm trying to achieve via template specialization isn't possible.

Thanks!

##### Share on other sites

This is as close as I was able to get:

I'm not sure if you're able to get rid Type in the base class, since you need access to that type in the specialization to determine if T is a derived class. However someone else with better template-foo can certainly prove me wrong

##### Share on other sites

Ah, seems promising! I actually tried SFINEA/enable_if, but didn't know how to put it base-struct. Putting the "Type" as part of the base class is not a problem, I'll have to try it out later, but thanks so far

##### Share on other sites

I'm not sure how many specializations you want to have for Object Supplier, but if you only want to support ecs::Data types as the specialization, and everything else to use the standard template, then you can just use std::conditional when declaring the Supplier type in registerBindFunction to select the ecs::Data type:

using Supplier = ObjectSupplier<typename std::conditional<std::is_base_of<ecs::Data<Class>, Class>::value, ecs::Data<Class>, Class>::type;

If you don't mind creating a single new overload for registerBindFunction then you can template the above to get rid of the explicit ecs::Data type if you want to support any other wrapped/CRTP base classes.

Otherwise, you can use Zipsters approach. You can also attempt to use the type directly in registerBindFunction instead of using enable_if. Not sure how much sense that makes since I'm making this up as I go, but:

using Supplier = ObjectSupplier<Class::UnderlyingType>;

Of course, now you need to make sure any class used with ObjectSupplier has that type, which I'm not particularly happy about. You can use some macro to simplify it, or you can make it more explicit by requiring a new derived class that handles it for you:

template <typename Object>
{
using UnderlyingType = Object;
};

namespace ecs
{
template <typename T>
class Data : ObjectReceiver<Data<T>>
{
};
}

Or something like that, which makes the relationship to ObjectSupplier more clear. But I'm still not too fond of it because it involves even more boilerplate and adds complexity to the usage of ObjectSupplier (now you need the Receiver class base...). However you can use ObjectReceiver directly in registerBindFunction and statically assert that the type input is derived from it, which gives you a way to guarantee its use with a clear error message.

That said, if you have some common function in your classes, then you can also completely bypass the need to provide that base type with a little bit of metaprogramming:

template<typename T, typename R>
T BaseOf(R T::*);

// baseFunc is a common function in your base class, or some such
using Supplier = ObjectSupplier<decltype(BaseOf(&Class::baseFunc))>;

If this is enough, then you won't even need to add Type to the base, and all is well (you still need some consistent function or variable to point at to deduce the type of the base class though).

There are probably more complex ways of getting the base type without requiring such a thing as "baseFunc", but I haven't looked into it much.

This is all just off the top of my head stuff. std::conditional would be the simplest but Zipsters approach is probably more scalable.

Unfortunately I don't have time to exercise "real template-foo" to get around the type declaration, but frankly it's not that important. The above is just food for thought (i.e. I wrote it up while I was eating a midnight snack, code probably doesn't even compile :P).

Edited by Styves

##### Share on other sites

What do you expect the function to look like when passing a "ObjectSupplier" class that seems to me as if you want some type of anonymous objects passing into your script, is it?

Either you create a fixed base class any of your bind function parameter has to inherit from or if that is not in your intention just try to restructure your "ObjectSupplier" class where maybe the solution would be to make some kind of anonymous type wrapping structure that also delivers the necessary informations to your binder function. Some time ago, I added a C# like object struct to my engine code (but never used it since ) to achive exactly this. Passing any type into it with an inner template abstraction while the outer struct is absolutely aware of

struct object
{
private:
struct AbstractType
{
virtual void Free(void** ptr, Allocator& allocator) = 0;

virtual void CopyByVal(void const* src, void** dest) = 0;
virtual void CopyByRef(void* const* src, void** dest) = 0;

virtual void* Value(void** ptr) = 0;

virtual const type_info& GetType() = 0;
virtual uint32 Size() = 0;
};
template<typename T> struct val_type : AbstractType
{
//override for value types here that fit into sizeof(void*)
};
template<typename T> struct ref_type : AbstractType
{
//override for ref types here that do not fit into sizeof(void*) or const char*
};

//do some pointer magic
template<typename T> static AbstractType* DiffType()
{
if(sizeof(T) <= sizeof(void*))
{
/**
the passed value matches into a pointer address memory block so use it directly
instead of creating a pointer based accessing manager. This saves a lot of memory.
*/
static val_type<T> handler;
return &handler;
}
else
{
static ref_type<T> handler;
return &handler;
}
};

AbstractType* handler; //an interface to the underlaying value
void* value; //the memory address hold value is stored

public:
inline object& Assign(object const& obj)
{
Clear();

handler = obj.handler;
if(handler) handler->CopyByRef(&obj.value, &value, GetAllocator());

return *this;
}
template<typename T> inline object& Assign(T const& ptr)
{
if(handler) handler->Free(&value, GetAllocator());

handler = DiffType<T>();
handler->CopyByVal(&ptr, &value, GetAllocator());

return *this;
}

template<typename T> inline T Cast()
{
if(!handler) return typename const_trait<T>::type();

T* r = reinterpret_cast<T*>(handler->Value(&value));
return *r;
}
template<typename T> inline bool TryCast(T& ptr)
{
if(!handler || handler->GetType() != typeof(T))
return false;

ptr = *reinterpret_cast<T*>(handler->Value(&value));
return true;
}

template<typename T> inline bool Is() { return (GetType() == typeof(T)); }
inline const std::type_info& GetType() { return ((handler) ? handler->GetType() : typeof(void)); }

inline const uint32 Size() { return ((handler) ? handler->Size() : 0); }

inline void Clear()
{
if(handler) handler->Free(&value);
handler = 0;
}
};

Left out some (I lied, a lot) of construction and operator code but reduced the basic mechanisms to the core ones to illustrate how it works. You add your any kind value into this cointainer as an anonymous plain pointer for the type erasure and add some kind of overhead information class to it additionally. Any template parameter is abstracted out by inheriting from an untemplated base class offering an interface but the deriving class will override it and so have access to the original type T again.

I added two types of those overhead classes where the value type one treats the passed pointer "as is" (usefull for saving an allocation for small types and primitives such as bool, char, int ...) while the reference type one allocates a copy of the passed memory address (so you may create an object from a type on stack without playing with invalid pointers when stack got cleared) for the cost of only two pointers

##### Share on other sites
7 hours ago, Styves said:

I'm not sure how many specializations you want to have for Object Supplier, but if you only want to support ecs::Data types as the specialization, and everything else to use the standard template, then you can just use std::conditional when declaring the Supplier type in registerBindFunction to select the ecs::Data type:

There are actually many more suppliers, which I want to be able to declare/add at different modules w/o affecting the code using it.

7 hours ago, Styves said:

Otherwise, you can use Zipsters approach. You can also attempt to use the type directly in registerBindFunction instead of using enable_if. Not sure how much sense that makes since I'm making this up as I go, but:

I don't think that would work, enable_if is used for SFINEA and probably required to only compile the specialization when the condition is actually met.

7 hours ago, Styves said:

Or something like that, which makes the relationship to ObjectSupplier more clear. But I'm still not too fond of it because it involves even more boilerplate and adds complexity to the usage of ObjectSupplier (now you need the Receiver class base...). However you can use ObjectReceiver directly in registerBindFunction and statically assert that the type input is derived from it, which gives you a way to guarantee its use with a clear error message.

Deriving the classes used for specialization is not a (good) option; first of all I try to minimze inheritance anyways and prefer external solutions via such "traits" normally; also there's a specific non-templated version for a large range of base-objects, which really should not have to be derived, if that makes any sense.

8 hours ago, Styves said:

If this is enough, then you won't even need to add Type to the base, and all is well (you still need some consistent function or variable to point at to deduce the type of the base class though).

I'm going to post my actual solution at the end of this post, where you should see that I don't even need to add "Type" to the base at all, even in Zipsters solution

2 hours ago, Shaarigan said:

What do you expect the function to look like when passing a "ObjectSupplier" class that seems to me as if you want some type of anonymous objects passing into your script, is it?

The ObjectSupplier is there to handle object-aquision for the script-calls. Long story short, I'm using a visual scripting system that is rather high level, and I want to be able to call certain functions w/o first having to aquire the actual C++-object (the type system won't even know it exists). Enter ObjectSupplier, which in case of my example will tell the script-system: "In case you bind a function of a ecs::Data<>-object, the function in script should take a Entity-reference instead and aquire the data-object off that before calling the function-pointer.
A function using tihs might look like that:

void Call(double dt)
{
CallState callState(*this);
Class& object = Supplier::GetClassObject(callState);

ClassCallHelper<false, Supplier, Class, Function, Return, Args...>::Call(object, m_pFunction, callState);
}

Now depending on the specialization of Supplier, it will eigther just get the actual object from Stack; or do something completely different as the specialization dictates. (if thats what you asked)

2 hours ago, Shaarigan said:

Either you create a fixed base class any of your bind function parameter has to inherit from or if that is not in your intention just try to restructure your "ObjectSupplier" class where maybe the solution would be to make some kind of anonymous type wrapping structure that also delivers the necessary informations to your binder function. Some time ago, I added a C# like object struct to my engine code (but never used it since ) to achive exactly this. Passing any type into it with an inner template abstraction while the outer struct is absolutely aware of

Haven't dealt with type-erasure much I think though, and I'm not sure if that would actually be superior to the current solution Zipster offered, but I'll have a proper read of it later, thanks

As for the solution I ended up using: Its pretty much as zipster wrote, just a tad simpler:

// base supplier
template<typename Object, typename Enable = void>
class ObjectSupplier
{
public:

static Object& GetClassObject(CallState& state)
{
return *state.GetAttribute<Object>(0);
}
};

template<bool Test>
using EnableIf = typename std::enable_if<Test>::type;

template<typename T, typename T2>
using CheckIsBase = EnableIf<std::is_base_of<T, T2>::value>;

template<typename Derived>
class ObjectSupplier<Derived, sys::CheckIsBase<ecs::Data<Derived>, Derived>>
{
public:

static Derived& GetClassObject(CallState& state)
{
return state.GetEntityAttribute(0)->GetData<Derived>();
}
};

sys::CheckIsBase is a typedef based on a typedef of std::enable_if using is_base_of, that I've actually been using for some time now. Also, as you can see I don't have to use Derived::Type, but can just use Derived directly. Which makes sense, if I added Type to ecs::Data<Derived>, Type would just be "Derived" in the end, which I don't have to lookup, since its already part of the template.

One problem I had is that ou can only use enable_if with no custom type supplies as second template parameter (my typedef used "int" as a default) for some reason, otherwise it will fail to correctly specialize. Other than that, this solution is pretty perfect especially since I don't have to add "Type" to the base classes; it allows me to add the ObjectSuppliers away from the code that uses it, and it also lets me add specialization based on different traits (ie. I have a "isBaseObject<>" trait for my type-system which doesn't involve inheritance).
So thanks to Zipster for providing me this neat solution, and thanks to everyone for their input

• 10
• 16
• 10
• 56