c++ reinterpret_cast, virtual, and templates ok?

Question

In C++, assume following class hierarchy:

class BaseClass { };
class ChildClass : public BaseClass { };

Further assume factory classes for these two classes with a common, templated base class:

template<typename T>
class Factory {
public:
  virtual T* create() = 0;
};

class BaseClassFactory : public Factory<BaseClass> {
public:
  virtual BaseClass* create() {
    return new BaseClass(&m_field);
  }
private:
  SomeClass m_field;
};

class ChildClassFactory : public Factory<ChildClass> {
public:
  virtual ChildClass* create() {
    return new ChildClass(&m_field);
  }
private:
  SomeOtherClass m_field; // Different class than SomeClass
};

Note that the size/internal structure of ChildClassFactory and BaseClassFactory is different due to their different fields.

Now, if a have an instance of ChildClassFactory (or Factory<ChildClass>), can I safely cast it to Factory<BaseClass> (via reinterpret_cast)?

Factory<ChildClass>* childFactory = new ChildClassFactory();

// static_cast doesn't work - need to use reinterpret_cast
Factory<BaseClass>* baseFactory = reinterpret_cast<Factory<BaseClass>*>(childFactory);

// Does this work correctly? (i.e. is "cls" of type "ChildClass"?)
BaseClass* cls = baseFactory->create();

I know that you can't always cast templated classes this way, but in this special case a cast should be safe, shouldn't it?

I've tested it with Visual C++ 2010 and it does work. My question now is whether this is portable to other compilers?

Update: Since there has been some confusion let me clarify some more what's (supposed to be) important in my example:

• ChildClass is a child class of BaseClass
• A user of Factory<BaseClass> doesn't know what child class of BaseClass will be created. All he knows is that BaseClass is created.
• Factory<T> has no fields of its own (other than the vtable).
• Factory::create() is virtual

Answer 1

No, it is not. You may not use the result of a reinterpret_cast other than to cast stuff back, except for a few special cases:

ISO14882:2011(e) 5.2.10-7:

An object pointer can be explicitly converted to an object pointer of a different type.70 When a prvalue v of type “pointer to T1” is converted to the type “pointer to cv T2”, the result is static_cast(static_cast(v)) if both T1 and T2 are standard-layout types (3.9) and the alignment requirements of T2 are no stricter than those of T1, or if either type is void. Converting a prvalue of type “pointer to T1” to the type “pointer to T2” (where T1 and T2 are object types and where the alignment requirements of T2 are no stricter than those of T1) and back to its original type yields the original pointer value. The result of any other such pointer conversion is unspecified.

To make a possible failure scenario more clear, consider multiple inheritance, where using a static_cast or dynamic_cast would sometimes adjust the pointer value, but reinterpret_cast will not. Consider casting in this example from A* to B*:

struct A { int x; };
struct B { int y; };
struct C : A, B { };

To understand how your code fails in a different way too, consider how most compilers implement the virtual function call mechanism: With virtual pointers. Your instance of ChildClassFactory will have a virtual pointer, pointing to the virtual table of ChildClassFactory. Now when you reinterpret_cast this beast, it just happens to "work" incidentally, because the compiler expects some virtual pointer, pointing to a virtual table that will have the same/similar layout. But it will still contain the values pointing to the ChildCLassFactory virtual functions, thus these functions would be called. All of this is long after invoking undefined behaviour. It is as if you are jumping with a car into a large canyon and thinking "hey, everything is driving fine" just because you have not hit the ground yet.

Answer 2

No, reinterpret_cast is only to be used for lowlevel code since it will not perfrom the correct address manipulation. Use static_cast or dynamic_cast instead,

Why do you want two factories this does not fit in the GoF factory pattern.

reinterpret_cast is not the way to do it since it is slow (runtime checks) and is not a nice OO design (you want to use the polymophisme build into the language).

Instead make constructors in the factory class that produces the types you are after and then have these call the constructor of the individual types.

The factory pattern allows you to be by ignorant of changes in your implementaion, which is a good thing since you minimize your dependencies, and allows for easier maintainance in the future of the code.

Answer 3

I've ticked the original answer above (to give him the credit), but I thought I'd sum up what I've learned here.

So, the basic problem is that it's not defined how dispatching virtual calls must be implemented.

This means that the data structure(s) that are internally used for virtual call dispatching (e.g. vtables) may or may not be bit compatible among template instantiations created from the same template.