When is dynamic_cast from parent to child useful in practice? Is this always a bad practice?

Question

When is dynamic_cast from parent to child useful in practice? Is this always a bad practice?

This is NOT a duplicate of: C++ Polymorphism: from parent class to child Run-time type information in C++

First, let's start with a code example and then let's draw some conclusions.

#include<iostream> 
using namespace std; 
class B { public: virtual void fun() {cout<<"!class B!\n";} }; 
class D: public B { public: void fun() {cout<<"!class D!\n";} }; 

int main() 
{ 
/////////   1. Dynamic initialization parent from child

B *bAux = new D; 
D *d = dynamic_cast<D*>(bAux); 
cout<<"1.1. for d is "; d->fun();   //class D

B *b = dynamic_cast<B*>(d);
cout<<"1.2. for b is "; b->fun();   //class D

/////////   2. Dynamic initialization child from parent

//D *dAux4 = new B;  //invalid conversion from ‘B*’ to ‘D*’

/////////   3. Casting the parent to child

B *bAux2 = new B;
cout<<"3.1. for bAux2 is "; bAux2->fun();   //class B

D *d2 = dynamic_cast<D*>(bAux2); 
if (d2 != NULL){ cout<<"3.2. for d2 is "; d2->fun(); }  //cannot cast parent to child
else cout<<"3.2. cannot cast B* to D* \n";

/////////   4. Casting the child to parent

D *dAux3 = new D;
cout<<"4.1. for dAux3 is "; dAux3->fun();   //class D

B *b3 = dynamic_cast<B*>(dAux3); 
cout<<"4.2. for b3 is "; b3->fun();   //class D

getchar(); 
return 0; 
} 

Second, let's draw the conclusions. The output is:

1.1. for d is !class D!                                                                                                                      
1.2. for b is !class D!                                                                                                                      
3.1. for bAux2 is !class B!                                                                                                                  
3.2. cannot cast B* to D*                                                                                                                    
4.1. for dAux3 is !class D!                                                                                                                  
4.2. for b3 is !class D!

But for case 3, we actually try to dynamic cast from B* to D* (as mentioned, not possible). I wonder in which cases is this useful? You can break the cast typing (as a parallel, in the case of static type checking, there is the well-known problem: Since a Bear is an Animal, shouldn't a set of Bears be a set of Animals? No, because then you can insert a Wolf into the set and have a Wolf among the Bears).

So, the question is, in which case is the dynamic cast from parent to child useful? Is this a bad practice?

EDIT: Following the discussion, a particular case:

/////////   5. Try to cast the parent initialized from a child to another child

B *bAux = new D; 
F *f = dynamic_cast<F*>(bAux); 
if(f == NULL)
    cout<<"5.1. Cannot cast the parent initialized from a child to another child \n";// f->fun();   //segmentation fault

B *b = dynamic_cast<B*>(f);
if(b == NULL)
    cout<<"5.2. Cannot cast the NULL pointer back to parent \n"; //b->fun();   //segmentation fault

Answer 1

dynamic_cast can be useful for the case you want to get derived objects interface: if there is f() in derived and no f() function in base than you can't call f() via B*. but you can dynamic_cast to D* and call f() (if B* points to D object). (Remembering that dynamic_cast is for polymorphic types) Consider the following code:

class B {
public:
    virtual void g()
    {
        std::cout << "Base::g()\n";
    }
};
class D : public B
{
public:
    void f()
    {
        std::cout << "f()\n";
    }
    void g()
    {
        std::cout << "Derived::g()\n";
    }
};
int main()
{
    B* p = new D;
    //p->f(); // class B has no member f
    D* pd = dynamic_cast<D*>(p);
    pd->f(); // ok f() will be called

    return 0;
}

Answer 2

One example is when there are multiple base classes of the same type. Only dynamic_cast can downcast from a non-unique base class to the most derived:

struct A { virtual ~A() = 0; };
struct B1 : A {};
struct B2 : A {};
struct C : B1, B2 {};
C& f(A& a) { return dynamic_cast<C&>(a); } // Only dynamic_cast can cast A& to C&.

---

Another example is that only dynamic_cast<void*> can return a pointer to the address of the complete object, no other cast can do.