What is virtual inheritance in C++ and when should you use it?

When we start coding in an object-oriented programming language we often think that it's about building nice inheritance hierarchies. So we do. A bit later we learn that we should rather use composition over inheritance. So we do. But we still need inheritance, and from time to time we run into problems where it seems to be the only way. At those times, we might learn about some more specialized forms of inheritance. In C++, this might mean private inheritance or virtual inheritance. Today we speak about the latter.

What is virtual inheritance?

The diamond problem

Virtual inheritance is a C++ technique that ensures that only one copy of a base class's member variables are inherited by second-level derivatives (a.k.a. grandchild derived classes). Without virtual inheritance, if two classes B and C inherit from class A, and class D inherits from both B and C, then D will contain two copies of A's member variables: one via B, and one via C. These will be accessible independently, using scope resolution.

Instead, if classes B and C inherit virtually from class A, then objects of class D will contain only one set of the member variables from class A.

As you probably guessed, this technique is useful when you have to deal with multiple inheritance and it's a way to solve the infamous diamond inheritance.

Multiple base class instances

In practice, virtual base classes are most suitable when the classes that derive from the virtual base, and especially the virtual base itself, are pure abstract classes. This means the classes above the "join class" (the one in the bottom) have very little if any data.

Consider the following class hierarchy to represent the diamond problem, though not with pure abstracts.

struct Person {
    virtual ~Person() = default;
    virtual void speak() {}
};

struct Student: Person {
    virtual void learn() {}
};

struct Worker: Person {
    virtual void work() {}
};

// A teaching assistant is both a worker and a student
struct TeachingAssistant: Student, Worker {};

TeachingAssistant ta;

As we said above, a call to aTeachingAssistant.speak() is ambiguous because there are two Person (indirect) base classes in TeachingAssistant, so any TeachingAssistant object has two different Person base class subobjects.

An attempt to directly bind a reference to the Person subobject of a TeachingAssistant object would fail, since the binding is inherently ambiguous:

TeachingAssistant ta;
Person& a = ta;  // error: which Person subobject should a TeachingAssistant cast into, 
                // a Student::Person or a Worker::Person?

To disambiguate, we would need to explicitly convert ta to any of the two base class subobjects:

TeachingAssistant ta;
Person& student = static_cast<Student&>(ta); 
Person& worker = static_cast<Worker&>(ta);

In order to call speak(), the same disambiguation, or explicit qualification is needed: static_cast<Student&>(ta).speak() or static_cast<Worker&>(ta).speak() or alternatively ta.Student::speak() and ta.Worker::speak(). Explicit qualification not only uses an easier, uniform syntax for both pointers and objects but also allows for static dispatch, so it would arguably be the preferable way to do it.

In this case, the double inheritance of Person is probably unwanted, as we want to model that the relation between TeachingAssistant and a Person exists only once. The fact that a TeachingAssistant is a Student and is a Worker at the same time does not imply that a TeachingAssistant is a Person twice (unless the TA suffers from schizophrenia): a Person base class corresponds to a contract that TeachingAssistant implements (the "is a" relationship above really means "implements the requirements of"), and a TeachingAssistant only implements the Person contract once.

There should be only one behaviour

The real-world meaning of "exists only once" is that a TeachingAssistant should have only one way of implementing speak, not two different ways.

In our degenerate case, Person::speak() is not overridden in either Student or Worker, but that could be different and then we would TeachingAssistant would have multiple implementations of the speak() method.

If we introduce virtual to our inheritance in the following way, our problems disappear:

struct Person {
    virtual ~Person() = default;
    virtual void speak() {}
};

// Two classes virtually inheriting Person:
struct Student: virtual Person {
    virtual void learn() {}
};

struct Worker: virtual Person {
    virtual void work() {}
};

// A teaching assistant is still a student and the worker
struct TeachingAssistant: Student, Worker {};

Now we can easily call speak().

The Person portion of TeachingAssistant::Worker is now the same Person instance as the one used by TeachingAssistant::Student, which is to say that a TeachingAssistant has only one - shared - Person instance in its representation and so a call to TeachingAssistant::speak is unambiguous. Additionally, a direct cast from TeachingAssistant to Person is also unambiguous, now that there exists only one Person instance which TeachingAssistant could be converted to.

This can be done through vtable pointers. Without going into details, the object size increases by two pointers, but there is only one Person object behind and no ambiguity.

You must use the virtual keyword in the middle level of the diamond. Using it in the bottom doesn't help.

You can find more detail at the Core Guidelines and here.

Should we always use virtual inheritance? If yes, why? If not, why not?

The answer is definitely no. The base of an idiomatic answer can be the most fundamental idea of C++: you only pay for what you use. And if you don't need virtual inheritance, you should rather not pay for it.

Virtual inheritance is almost never needed. It addresses the diamond inheritance problem that we saw at the beginning of the article. It can only happen if you have multiple inheritance, otherwise, you cannot have this issue.

At the same time, it has some drawbacks.

More complex dependencies

Virtual inheritance causes troubles with object initialization and copying. Since it is the "most derived" class that is responsible for these operations, it has to be familiar with all the intimate details of the structure of base classes.

Due to this, a more complex dependency appears between the classes, which complicates the project structure and forces you to make some additional revisions in all those classes during refactoring. All this leads to new bugs and makes the code less readable and thus less maintainable.

Expensive type conversions

ISO C++ guidelines also suggests that C-style downcasts cannot be used to cast a base class pointer to a derived one.

The problems can be solved by dynamic_cast, but it has its performance implications. Using too much dynamic_cast in your code can make a big hit, and it also means that your project's architecture is probably very poor.

You can always implement what you need without multiple inheritance. There is no surprise in that. After all, the feature of virtual inheritance is not present in many other major languages, yet they are used for large and complex projects.

Conclusion

Today, we discussed the diamond inheritance problem. We understood that when there are multiple paths between a base and a derived class, there are multiple base objects instantiated which is almost never desirable. C++ proposes virtual inheritance to solve this problem and letting such structures to live with only one instance of a base class.

Yet, as you should only pay for what you use, virtual inheritance should not be your default choice. Most projects can be implemented without such a language feature and if you can design your software without multiple inheritance, you don't need to deal with its downsides.

Have you ever used multiple inheritance in your production code? If yes, what was the use-case?

Connect deeper

If you found interesting this article, please subscribe to my personal blog and let's connect on Twitter!