C++ Does casting create a new object?

Question

As indicated in the title above, my question is simply whether or not a C++ cast does create a new object of the target class. Of course, I have used Google, MSDN, IBM and stackoverflow's search tool before asking this but I can't find an appropriate answer to my question.

Lets consider the following implementation of the diamond problem solved by using virtual inheritance:

#include <iostream>
#include <cstdlib>

struct A
{
  int a;

  A(): a(2) {  }
};

struct B: virtual public A
{
  int b;

  B(): b(7) {  }
};

struct C: virtual public A
{
  int c;

  C(): c(1) {  }
};

struct END: virtual public B, virtual public C
{
  int end;

  END(): end(8) {  }
};

int main()
{
  END *end = new END();
  A *a = dynamic_cast<A*>(end);
  B *b = dynamic_cast<B*>(end);
  C *c = dynamic_cast<C*>(end);

  std::cout << "Values of a:\na->a: " << a->a << "\n\n";
  std::cout << "Values of b:\nb->a: " << b->a << "\nb->b: " << b->b << "\n\n";
  std::cout << "Values of c:\nc->a: " << c->a << "\nc->c: " << c->c << "\n\n";

  std::cout << "Handle of end: " << end << "\n";
  std::cout << "Handle of a: " << a << "\n";
  std::cout << "Handle of b: " << b << "\n";
  std::cout << "Handle of c: " << c << "\n\n";
  system("PAUSE");
  return 0;
}

As I understood, the actual structure of B and C, which normally consists of both an embedded instance of A and variables of B resp. C, is destroyed since the virtual A of B and C is merged to one embedded object in END to avoid ambiguities. Since (as I always thought) dynamic_cast usually only increases the address stored by a pointer by the offset of the embedded (cast's) target class there will be a problem due to the fact that the target (B or C) class is divided into several parts.

But if I run the example with MSVC++ 2011 Express everything will happen as expected (i.e. it will run, all *.a output 2), the pointers only slightly differ. Therefor, I suspect that the casts nevertheless only move the addresses of the source pointers by the internal offset of B's / C's instance.

But how? How does the resulting instance of B / C know the position of the shared A object. Since there is only one A object inside the END object but normally an A object in B and C, either B or C must not have an instance of A, but, indeed, both seem to have an instance of it.

Or does virtual only delegate calls to A's members to a central A object without deleting the respective A objects of each base class which inherits virtual from A (i.e. does virtual actually not destroy the internal structure of inherited and therefor embedded objects but only not using their virtualized (= shared) members)?

Or does virtual create a new "offset map" (i.e. the map which tells the address offsets of all members relative to the pointer to a class instance, I dunno the actual term) for such casted objects to handle their "distributedness"?

I hope I have clarified everything, many thanks in advance
BlueBlobb

PS:
I'm sorry if there are some grammar mistakes, I'm only a beer loving Bavarian, not a native speaker :P

Edit:
If have added these lines to output the addresses of all int a's:

  std::cout << "Handle of end.a: " << &end->a << "\n";
  std::cout << "Handle of a.a: " << &a->a << "\n";
  std::cout << "Handle of a.b: " << &b->a << "\n";
  std::cout << "Handle of a.c: " << &c->a << "\n\n";

They are the same implying that there is indeed only one A object.

Answer 1

my question is simply whether or not a C++ cast does create a new object of the target class.

Yes, a cast to a class type would create new temporary object of that type.

Note that your example doesn't cast to a class anywhere: the only casts it performs are to pointer types. Those casts do create new instances of pointers - but not of the objects pointed to. I'm not sure what your example was supposed to demonstrate, nor how it is related to your stated question.

Also, dynamic_cast is unnecessary where you use it; an implicit conversion would work just as well.

Since (as I always thought) dynamic_cast usually only increases the address stored by a pointer by the offset of the embedded (cast's) target class

You must be thinking of static_cast or something. dynamic_cast is much more powerful. For example, it can cast from B* to C*, even though they are unrelated at compile time, by going down to END* and then back up the other branch. dynamic_cast utilizes run-time type information.

How does the resulting instance of B / C know the position of the shared A object.

This is implementation-dependent. A typical implementation would reserve space within the derived class instance to store an offset to its virtual base class instance. The constructor of the most-derived class initializes all those offsets.

Answer 2

No, you're just seeing the effects of multiple inheritance. In order for a pointer to be cast to a different base type, it has to be adjusted to the part of the object that represents that exact type. The compiler knows the original type of the pointer and the result type, so it can apply the necessary offsets. In order for the derived type to satisfy the "is-a" requirement it must have the necessary structure built in to emulate all of the base types.

There's one case where a cast can create a new object, and that's when you're casting to a type other than a pointer or reference type. Often that won't be possible unless you've defined a cast operator for that type.

Answer 3

The example you gave uses pointers.

A* a = dynamic_cast<A*>(end);

So the only "new" thing created here is another pointer, which will point to the "A" vtable of the object to which "end" points. It does not actually construct a new object of the class/struct types you are using.

Contrast with

A a;
B b(a);

Here a new object is created. But otherwise, casting does not create a new object of the destination cast type.

The reason the pointers differ is because they are pointing to the different vtables that preceed the data section of the underlying object.

Example:

#include <iostream>

using namespace std;

struct A {
    int a[64];
    A() { cout << "A()" << endl; }
    A(const A&) { cout << "A(A&)" << endl; }
    A& operator = (const A&) { cout << "A=A" << endl; return *this; }
};

struct B : virtual public A {
    int b[64];
    B() { cout << "B()" << endl; }
    B(const B&) { cout << "B(B&)" << endl; }
    B(const A&) { cout << "B(A&)" << endl; }
    B& operator = (const B&) { cout << "B=B" << endl; return *this; }
    B& operator = (const A&) { cout << "B=A" << endl; return *this; }
};

struct C : virtual public A {
    int c[64];
    C() { cout << "C()" << endl; }
    C(const C&) { cout << "C(C&)" << endl; }
    C(const B&) { cout << "C(B&)" << endl; }
    C(const A&) { cout << "C(A&)" << endl; }
    C& operator = (const C&) { cout << "C=C" << endl; return *this; }
    C& operator = (const B&) { cout << "C=B" << endl; return *this; }
    C& operator = (const A&) { cout << "C=A" << endl; return *this; }
};

struct END : virtual public B, C {
    int end[64];
    END() { cout << "END()" << endl; }
    END(const END&) { cout << "END(END&)" << endl; }
    END(const C&) { cout << "END(C&)" << endl; }
    END(const B&) { cout << "END(B&)" << endl; }
    END(const A&) { cout << "END(A&)" << endl; }
    END& operator = (const END&) { cout << "END=END" << endl; return *this; }
    END& operator = (const C&) { cout << "END=C" << endl; return *this; }
    END& operator = (const B&) { cout << "END=B" << endl; return *this; }
    END& operator = (const A&) { cout << "END=A" << endl; return *this; }
};

int main() {
    END* end = new END();

    A *a = dynamic_cast<A*>(end);
    B *b = dynamic_cast<B*>(end);
    C *c = dynamic_cast<C*>(end);

    std::cout << "end = " << (void*)end << std::endl;
    std::cout << "a = " << (void*)a << std::endl;
    std::cout << "b = " << (void*)b << std::endl;
    std::cout << "c = " << (void*)c << std::endl;

    // the direct pointers are going to have to differ
    // to point to the correct vtable. what about 'a' in all cases?
    std::cout << "end->a = " << (void*)&(end->a) << std::endl;
    std::cout << "a->a = " << (void*)&(a->a) << std::endl;
    std::cout << "b->a = " << (void*)&(b->a) << std::endl;
    std::cout << "c->a = " << (void*)&(c->a) << std::endl;


}

Which you can see running here: http://ideone.com/0QAoWE

Answer 4

At least with MSVC in VS 2017, the answer is a definite maybe.

// Value is a struct that contains a member: std::string _string;
// _value is a std::variant<> containing a Value as one member

template <> std::string const &Get<std::string>() const
{
    // Required pre-condition: _value.index() == TYPE_VALUE
    Value const &value = std::get<TYPE_VALUE>(_value);
    return static_cast<std::string>(value._string);
}

std::string const &test()
{
    static std::string x = "hello world";
    return static_cast<std::string>(x);
}

Get() is a very small snippet from a much larger project, and won't operate without the support of several hundred other lines of code. test() is something I quickly threw together to investigate.

As written, Get()generates the following warning:

warning C4172: returning address of local variable or temporary

while test() compiles clean. If I remove the static_cast<> from Get(), it also compiles cleanly.

P.S. in hindsight, I ought to rename _value to something like _payload, since it can contain a lot more than a Value.