is virtual inheritance from pure abstract classes (interfaces) necessary

Question

Why is it that in the code below the compiler complains that PureAbstractBase is an ambiguous base class of MultiplyInheritedClass? I realize I have two copies of the PureAbstractBase in MultiplyInheritedClass and that FirstConreteClass and SecondConreteClass should be derived virtually because they're the middle row of the diamond (and that does indeed fix the problem with the code below). But even though I have two copies of the interface why is it that the code in MultiplyInheritedClass does not just override both and unambiguously pick the interface class defined in MultiplyInheritedClass?

#include <iostream>
using namespace std;

class PureAbstractBase {
  public:
    virtual void interface() = 0;
};

// I know that changing the following line to:
// class FirstConcreteClass : public virtual PureAbstractBase {
// fixes the problem with this hierarchy
class FirstConcreteClass : public PureAbstractBase {
  public:
    virtual void interface() { implementation(); }
  private:
    void implementation() { cout << "This is object FirstConcreteClass\n"; }
};

// I know that changing the following line to:
// class SecondConcreteClass : public virtual PureAbstractBase {
// fixes the problem with this hierarchy
class SecondConcreteClass : public PureAbstractBase {
  public:
    virtual void interface() { implementation(); }
  private:
    void implementation() { cout << "This is object SecondConcreteClass\n"; }
};

class MultiplyInheritedClass : public FirstConcreteClass,
                               public SecondConcreteClass {
  public:
    virtual void interface() { implementation(); }
  private:
    void implementation() { cout << "This is object MultiplyInheritedClass\n"; }
};

Further, why do I not have issues with the following hierarchy? Doesn't the ConcreteHandler class have three copies of the AbstractTaggingInterface in this case? So why doesn't it have the same issue as the example above?

#include <iostream>
using namespace std;

class AbstractTaggingInterface {
  public:
    virtual void taggingInterface() = 0;
};

class FirstAbstractHandler : public AbstractTaggingInterface {
  public:
    virtual void taggingInterface() { cout << "FirstAbstractHandler\n"; }
    virtual void handleFirst() = 0;
};

class SecondAbstractHandler : public AbstractTaggingInterface {
  public:
    virtual void taggingInterface() { cout << "SecondAbstractHandler\n"; }
    virtual void handleSecond() = 0;
};

class ThirdAbstractHandler : public AbstractTaggingInterface {
  public:
    virtual void taggingInterface() { cout << "ThridAbstractHandler\n"; }
    virtual void handleThird() = 0;
};

class ConcreteHandler : public FirstAbstractHandler,
                        public SecondAbstractHandler,
                        public ThirdAbstractHandler {
  public:
    virtual void taggingInterface() = { cout << "ConcreteHandler\n"; }
    virtual void handleFirst() {}
    virtual void handleSecond() {}
    virtual void handleThird() {}
};

I am trying to wrap my head around all of this because I had a conversation with a colleague recently where he claimed that if you were inheriting from pure virtual classes (interfaces) without any data members then virtual inheritance was not necessary. I think understanding why the former code example does not work and the latter does would go a long way to getting this straight in my head (and clear up what exactly he meant by his comment). Thanks in advance.

Answer 1

You need virtual inheritance to overcome the diamond-ambiguity:

class FirstConcreteClass  : public virtual PureAbstractBase { ... };
class SecondConcreteClass : public virtual PureAbstractBase { ... };

---

Long-winded explanation: Suppose you have this:

// *** Example with errrors! *** //
struct A { virtual int foo(); };
struct B1 : public A { virtual int foo(); };
struct B2 : public A { virtual int foo(); };
struct C: public B1, public B2 { /* ... */ };  // ambiguous base class A!

int main() {
  A * px = new C;                              // error, ambiguous base!
  px->foo();                                   // error, ambiguous override!
}

The inheritance of the virtual function foo is ambiguous because it comes in three ways: from B1, from B2 and from A. The inheritance diagram forms a "diamond":

   /-> B1 >-\
A->          ->C
   \-> B2 >-/

By making the inheritance virtual, struct B1 : public virtual A; etc., you allow any baseclass of C* to call the correct member:

struct A { virtual int foo(); };
struct B1 : public virtual A { virtual int foo(); };
struct B2 : public virtual A { virtual int foo(); };
struct C: public B1, public B2 { virtual int foo(); };

We must also define C::foo() for this to make sense, as otherwise C would not have a well-defined member foo.

Some more details: Suppose we now have a properly virtually-inheriting class C as above. We can access all the various virtual members as desired:

int main() {
  A * pa = new C;
  pa->foo();      // the most derived one
  pa->A::foo();   // the original A's foo

  B1 * pb1 = new C;
  pb1->foo();     // the most derived one
  pb1->A::foo();  // A's foo
  pb1->B1::foo(); // B1's foo

  C * pc = new C;
  pc->foo();      // the most derived one
  pc->A::foo();   // A's foo
  pc->B1::foo();  // B1's foo
  pc->B2::foo();  // B2's foo
  pc->C::foo();   // C's foo, same as "pc->foo()"
}

 

Update: As David says in the comment, the important point here is that the intermediate classes B1 and B2 inherit virtually so that further classes (in this case C) can inherit from them while simultaneously keeping the inheritance from A unambiguous. Sorry for the initial mistake and thanks for the correction!

Answer 2

Your first example fails because the compiler cannot disambiguate between the three implementations of implementation(). You are overriding that method in MultiplyInheritedClass, which actually overrides both FirstConcreteClass::implementation and SecondConcreteClass::implementation (once virtual, always virtual). However, both virtual calls still exist in the interface of MultiplyInheritedClass, which makes the call ambiguous at the call site.

The reason that your example works without virtual inheritance is that there is no conflicting implementation of the common base class. Put another way:

class Base
{
public:
    void DoSomething() {
    std::cout << "TADA!";
    }
}

class One : public Base
{
    //...
}

class Two : public Base
{
    //...
}

class Mixed : public One, public Two
{
    //...
}

int main()
{
    Mixed abc;
    abc.DoSomething(); //Fails because the compiler doesn't know whether to call
                       // One::DoSomething or Two::DoSomething, because they both
                       // have implementations.

    //In response to comment:
    abc.One::DoSomething(); //Succeeds! You removed the ambiguity.
}

Because your example has all pure virtual functions, there's no multiple implementations which the compiler needs to disambiguate. Therefore, only one implementation exists, and the call is unambiguous.

Answer 3

I tried both of the question codes and they worked fine when instantiating an object of the multi-inherited class. It didn't work only with polymorphism, like this for example:

PureAbstractBase* F;
F = new MultiplyInheritedClass();

And the reason is clear: it doesn't know to which copy of the Abstract base class it should be linked (sorry for bad expressions, I understand the idea but can't express it). And since inherting virtaully makes only one copy exist in the derived class, then it's fine.

Also the code of Billy ONeal is not clear at all, what should we place instead of the comments?

If we place:

public:    
void DoSomething() 
{    std::cout << "TADA!";    }

it works fine, because of no virtuality.

I work on Visual Studio 2008.

Answer 4

Why not do it like this (suggested in Benjamin Supnik's blog entry):

#include <iostream>

class PureAbstractBase {
public:
    virtual void interface() = 0;
};

class FirstConcreteClass : public PureAbstractBase {
public:
    virtual void interface() { implementation(); }
private:
    void implementation() { std::cout << "Fisrt" << std::endl; }
};

class SecondConcreteClass : public PureAbstractBase {
public:
    virtual void interface() { implementation(); }
private:
    void implementation() { std::cout << "Second" << std::endl; }
};

class MultiplyInheritedClass : public FirstConcreteClass,
                               public SecondConcreteClass 
{
public:
    virtual void interface() { implementation(); }
private:
    void implementation() { std::cout << "Multiple" << std::endl; }
};

int main() {
MultiplyInheritedClass mic;
mic.interface();

FirstConcreteClass *fc = &mic; //disambiguate to FirstConcreteClass 
PureAbstractBase *pab1 = fc;
pab1->interface();

SecondConcreteClass *sc = &mic; //disambiguate to SecondConcreteClass 
PureAbstractBase *pab2 = sc;
pab2->interface();
}

which gives:

Multiple
Multiple
Multiple    

This way:

• no virtual bases are involved (do you really need them?)
• you can call the overriden function via a an instance of the MultiplyInheritedClass
• ambiguity is removed by a two-stage conversion