What's faster: down-cast from virtual base or cross-cast?

Question

This is somewhat hypothetical as I'm not too worried about performance - just wondering which option is actually the fastest/most efficient in general, or if there is no difference whatsoever.

Suppose I have the following code for a visitor template that supports overloading:

#define IMPLEMENT_VISITOR_WITH_SUPERCLASS(superclass)  \
    typedef superclass visitor_super_t;     \
    virtual void visit(Visitor& v) { v.visit(*this); }
//-----------------------------------------------------------------------------
// Implementation detail:
// Selective dispatcher for the visitor - required to handle overloading.
//
template <typename T>
struct VisitorDispatch {
    static void dispatch(Visitor* v, T* t) { v->visit(*t); }
};
// Specalization for cases where dispatch is not defined
template <> struct VisitorDispatch<void> {
    static void dispatch(Visitor* v, void* t) { throw std::bad_cast(""); }
};

//-----------------------------------------------------------------------------
// Derive visitors from this and 'Visitor'.
template <typename T>
class VTarget
{
public:
    // Don't really need a virtual dtor.
    virtual void dispatch(T& t) = 0;
};

//-----------------------------------------------------------------------------
class Visitor
{
public:
    virtual ~Visitor() = 0;

    template <typename T>
    void visit(T& t) {
        typedef VTarget<T> target_t;
        target_t* tgt = dynamic_cast<target_t*>(this);
        if (tgt) {
            tgt->dispatch(t);
        }
        else {
            // Navigate up inhertiance hierarchy.
            // requires 'super' to be defined in all classes in hierarchy
            // applicable to this visitor.
            typedef typename T::visitor_super_t super;
            super* s = static_cast<super*>(&t);
            VisitorDispatch<super>::dispatch(this, s);
        }
    }
};

//-----------------------------------------------------------------------------
inline Visitor::~Visitor() {}

This is then used to create generic visitors:

class CommonBase { 
    IMPLEMENT_VISITOR_WITH_SUPERCLASS(void)
    virtual ~CommonBase() = 0;
};
class A : public CommonBase {
    IMPLEMENT_VISITOR_WITH_SUPERCLASS(CommonBase)
};
class B : public CommonBase {
    IMPLEMENT_VISITOR_WITH_SUPERCLASS(CommonBase)
};

class MyVisitor
    : public Visitor
    , public VTarget<CommonBase>
    , public VTarget<A>
    , public VTarget<B>
{
public:
    virtual void dispatch(CommonBase& obj);
    virtual void dispatch(A& obj);
    virtual void dispatch(B& obj);
};

Using the visitor ultimately results in dynamic_cast<>'s from Visitor to VTarget<T>, which is a cross-cast.

The other way this could be implemented is to make Visitor a virtual base of VTarget<T> - MyVisitor would then not need to inherit directly from Visitor anymore. The dynamic_cast<> in the Visitor::visit code would then result in a down-cast from the virtual base, Visitor.

Is one method faster than the other when performing the casts? Or do you only get a size penalty for having the virtual base?

Answer 1

Well, it looks like the cross-cast method is faster than the virtual base method.

With a visitation that requires 1 fallback to a superclass, over 100000000 iterations, the cross-cast method took 30.2747 seconds, and the virtual base method took 41.3999 - about 37% slower.

With no fallback to an overload for a superclass, the cross-cast was 10.733 seconds and the virtual base 19.9982 (86% slower).

I was more interested to know how the dynamic_cast would operate in either mode, really.