Don’t inherit from standard types

Puzzle of the day: Here we have two different class types, both of which conceptually wrap up a sequence of integers into something that can be “decremented” (popping the last element) and tested for emptiness via operator bool. The only difference between them is that one is expressed via composition and the other via inheritance.

struct Composition {
    std::vector<int> m_;
public:
    Composition(std::initializer_list<int> il) : m_(il) {}
    operator bool() const { return !m_.empty(); }
    void operator--() { m_.pop_back(); }
};

struct Inheritance : std::vector<int> {
public:
    Inheritance(std::initializer_list<int> il) : vector(il) {}
    operator bool() const { return !this->empty(); }
    void operator--() { this->pop_back(); }
};

Now we feed them to this function:

template<class T>
int countdown_to_zero(T t)
{
    std::vector results = { t };
    for ( ; t; --t)
        results.push_back(t);
    return results.size();
}

int main() {
    auto x = Composition{ 1, 2, 3 };
    auto y = Inheritance{ 1, 2, 3 };
    std::cout << countdown_to_zero(x) << "\n";
    std::cout << countdown_to_zero(y) << "\n";
}

And we feed this to the compiler, and we run the resulting program, and we get:

4
6

The puzzle is: Why the difference?


Suppose I told you that the keyword struct was relevant…

(Recall that members and inheritance relationships are private-by-default for classes declared with the class keyword, but public-by-default for classes declared with the struct keyword.)

Change the definitions to

class Composition {

class Inheritance : std::vector<int> {

and you’ll see that you now get a compiler error.

prog.cc:23:29: error: cannot cast 'const Inheritance' to its
private base class 'const std::__1::vector<int, std::__1::allocator<int> >'
    std::vector results = { t };
                            ^

That’s right — I snuck in some CTAD when you weren’t looking! (Or, if you were watching for uses of CTAD, this probably wasn’t much of a puzzle, eh?)

In the Composition case, what we have here is a vector<Composition>. When we push_back onto it, we’re pushing back Composition objects.

In the Inheritance case, what we have here is a vector<int>. When we push_back onto it, we’re pushing back int objects, which we get by calling Inheritance’s implicit operator bool. And when we try to hide our inheritance relationship by making it private, we end up hiding the relationship from countdown_to_zero — but we can’t hide it from CTAD!


Any time we see a C++ “puzzle” — after we finish enjoying its puzzling aspect — we should pause and think: Real production codebases should never be puzzling. What guidelines can we follow in our daily work, to make sure that we don’t accidentally leave a puzzle for our coworkers? What simple guidelines, if followed punctiliously in this case, would have led to no puzzle at all?

This puzzle becomes puzzling by breaking two of my simple guidelines.

  1. Never, ever use CTAD. Not even by accident! (I hope to see -Wctad in Clang soon.)

  2. Never, ever inherit from any std:: type. Not even privately!

(Except of course for the standard types you’re supposed to inherit from. std::iterator has rightly been deprecated, but std::enable_shared_from_this<T> is certainly an exception to this general rule. Another exception to the rule — fittingly! — is std::exception, which should be at the root of all your exception hierarchies.)

I recently said in a mailing-list discussion, and I stick by it:

Standard types are like boxes of chocolates: You never know what you’re going to get.

If you think you do know “what you get” when your type T inherits from std::vector: quick, does your type T have a member function ==, and what does it do? Does it have emplace_back, and if so, what is emplace_back’s return type? Does it have CTAD deduction guides? And so on. It’s not that these questions don’t have answers; it’s that you don’t know the answers (and neither do your coworkers). And even if you know everything you get with std::vector this year — which means you had no trouble with the puzzler, right? — well, in C++2a everything’s going to change again, and you’ll have to go re-audit your code to make sure it still does the right thing.

Rather than audit your code every time the definition of std::vector changes, I confidently assert that it’s vastly more reliable to simply not inherit from std::vector in the first place. And this goes for all the types you didn’t write yourself: not just vector, not just “standard containers,” not even “standard types.” If you (or your project, or your company) didn’t write class Foo, then class Foo should not be granted control over the API of your own class. And that’s what you’re doing when you inherit from a base class: you’re granting that class control over your API.

I’m sure we’ll revisit this topic again.

Posted 2018-12-11