我应该在何时使用 dynamic_cast 进行向下转换？

Question

I noticed that compiler will optimize out some dynamic_cast when the following operation is non-polymorphic, for example the following code:

#include <iostream>

using namespace std;

struct A
{
    virtual ~A() = default;
    virtual int f()
    {
        return 1;
    };
    virtual int g() = 0;
};

struct B : A
{
    int g() const
    {
        return 2;
    }
    int g() override
    {
        return 3;
    }
    int h()
    {
        return 4;
    }
};

int main()
{
    A*   p  = new B();
    auto u  = p->f();
    auto v1 = static_cast<B*>(p)->f();
    auto v2 = dynamic_cast<B*>(p)->f();
    auto w  = p->g();
    auto x1 = static_cast<const B*>(p)->g();
    auto x2 = dynamic_cast<B*>(p)->g();
    auto x3 = dynamic_cast<const B*>(p)->g();
    auto y  = dynamic_cast<B*>(p)->h();
    cout << u << v1 << v2 << w << x1 << x2 << x3 << y << endl;
    delete p;
    return 0;
}

There are only two dynamic_cast calls compiled with g++ -O2, that means that equals to static_cast, so should I always use dynamic_cast for downcast for no extra overhead to be considered?

Answer 1

The main issue with dynamic_cast is that they are very slow and complicated. Compilers can indeed optimize it when they are aware of the actual type in compile time but not always. Technically, your code should have 0 dynamic casts if compilers knew how to do it properly. So don't trust too much on the exact mechanisms of optimization that compilers use.

Certain parts of your code could've been optimized via abuse of undefined behavior. For example:

dynamic_cast<B*>(p)->f(); 
// this is optimized instantly to p->f(); 
// if dynamic_cast returns nullptr it would be undefined behavior IIRC, 
// so optimizer can assume that the cast is successful and it becomes equivalent to
// static_cast<B*>(p)->f() after optimization,
// regardless of whether p is actually of type B or not

In general, it is difficult to provide concrete methodology on dynamic casts, if performance allows it, always dynamic cast on down casting. Not doing so is possible undefined behavior and security leak.

If you have no guarantee that the derived class is of this specific type - you have no choice but to use dynamic cast.

If you have a guarantee but there might be bugs in the code, consider making a static assert with dynamic cast inside, and using static cast in optimized versions.

Answer 2

In fact, I don't see any effective difference in using static_cast and dynamic_cast here. First, if you call a virtual function via a casted pointer, all the following calls will have the very same effect — they will trigger dynamic dispatching :

auto v0 = p->f();
auto v1 = static_cast<B*>(p)->f();
auto v2 = dynamic_cast<B*>(p)->f();

dynamic_cast might add some overhead, but the observable effect will be the same. Namely. B::f will be called (if it's overridden).

As for non-virtual functions from a derived class, you need to cast:

auto x1 = static_cast<const B*>(p)->g();
auto x3 = dynamic_cast<const B*>(p)->g();

If those casts are invalid, you will get undefined behavior in both cases. If they are valid, there will be effectively the same. Just again, with some additional overhead of dynamic_cast .

Note that, IMO, there is no dynamic_cast optimized away in your program. You can observe two dynamic_cast calls in your assembly, and you have two different forms of dynamic_cast in your code. I assume a compiler is doing dynamic_cast<B*> only once and uses the result 3 times.

---

Note that you can bypass the dynamic dispatch by manually selecting which f you want to call by the following syntax:

auto y1 = static_cast<B*>(p)->A::f();
auto y2 = static_cast<B*>(p)->B::f();

Or, with dynamic_cast the very same way.

Live demo: https://wandbox.org/permlink/CZRLPWxHjSMk8dFK .