The “OO” Antipattern

For those who only read above the fold: I don’t say that all object-orientation is bad! OOP, especially classical polymorphic OOP, has a well-deserved place in real code. I’m going to talk about a very specific antipattern that I see with some frequency: the use of class for things that should be simple free functions.

I originally wrote about this specific piece of code on Code Review StackExchange (February 2018). I gave an unrecorded lightning talk based on the same material in April 2018.

If this kind of thing appeals to you, I recommend with the highest possible recommendation that you get yourself a copy of Kernighan and Plauger’s The Elements of Programming Style (1974, reissued 1978).

The original code

Students who’ve learned C++ in a particular kind of academic setting often approach it with the “everything is an object” mindset. Suppose they’re given a task to compute some value; then their first step is to create a ValueComputer object with a member function vc.computeResult().

For example: A student is asked to use dynamic programming to count the number of distinct domino tilings of a rectangle. The student writes:

int main()
{
    DominoTilingCounter tc(4, 7);  // in a 4x7 rectangle
    std::cout << tc.count() << '\n';
}

Having framed the problem, they go on to implement the DominoTilingCounter class. The clever student even adds memoization, so that the count() member function won’t be so slow the second time it’s called:

class DominoTilingCounter {
    int height, width;
    bool done = false;
    int tilingCount;
    int computeCount(int h, int w, std::string_view prevRow, int rowIdx) {
        [...recursive solution omitted...]
    }
public:
    explicit DominoTilingCounter(int h, int w) : height(h), width(w) {
        if (h == 0 || w == 0 || (w*h) % 2 != 0) {
            tilingCount = 0;
            done = true;
        }
    }
    int count() {
        if (!done) {
            tilingCount = computeCount(height, width, "", 0);
            done = true;
        }
        return tilingCount;
    }
};

(Experienced programmers may be cringing, here. That’s the point.)

Unfortunately, this code fails the const-correctness test: the count() member function sounds like it should be non-modifying, but in fact it needs to update member data and thus cannot be const.

std::cout << DominoTilingCounter(4, 7).count() << '\n';
    // Fails to compile!

(Hacker News points out that this does compile; class prvalues are not const the way scalar prvalues are. Oops.) Anyway, we can fix this issue incidentally, just by applying a little more logic.

The logical leap

Look: when you construct a DominoTilingCounter object tc, it is specifically for the purpose of computing tc.count(), right? There’s no other purpose that tc could serve?

Again: When you construct a DominoTilingCounter object, it is specifically for the purpose of computing tc.count().

So put the computation in the constructor!

class DominoTilingCounter {
    int height, width;
    int tilingCount;
    int computeCount(int h, int w, std::string_view prevRow, int rowIdx) {
        [...recursive solution omitted...]
    }
public:
    explicit DominoTilingCounter(int h, int w) : height(h), width(w) {
        if (h == 0 || w == 0 || (w*h) % 2 != 0) {
            tilingCount = 0;
        } else {
            tilingCount = computeCount(height, width, "", 0);
        }
    }
    int count() const {
        return tilingCount;
    }
};

This refactoring does several things:

  • It permits count() to be a const member function.

  • It eliminates the “empty, uncomputed” state from your program.

  • It eliminates the data member done, whose whole purpose was to track that empty state. (In C++17 we might use a std::optional for that purpose; but look, now we don’t have to!)

In fact, the private data members height and width are now unused as well. It turns out that we were using them only to pass data from the constructor to the computation in the count() method; and now that the computation is taking place in the constructor, we don’t need those data members anymore. Our code shrinks drastically.

The final leap

In the original code, the student’s computeCount member function happened to take w and h as function parameters, rather than reading them from the height and width data members. That was a fortunate accident: computeCount doesn’t use any of the data members of the DominoTilingCounter object, and so we can mark it static. Our code is now something like this:

class DominoTilingCounter {
    int tilingCount;
    static int computeCount(int h, int w, std::string_view prevRow, int rowIdx) {
        if (h == 0 || w == 0 || (w*h) % 2 != 0) {
            return 0;
        }
        [...recursive solution omitted...]
    }
public:
    explicit DominoTilingCounter(int h, int w) {
        tilingCount = computeCount(h, w, "", 0);
    }
    int count() const {
        return tilingCount;
    }
};

The final leap is to observe that this entire class does nothing but wrap an assignment to an int!

Avoid classes that are tantamount to int.

What we should write is a simple free function countDominoTilings(w, h):

int countDominoTilingsImpl(int h, int w, std::string_view prevRow, int rowIdx) {
    if (h == 0 || w == 0 || (w*h) % 2 != 0) {
        return 0;
    }
    [...recursive solution omitted...]
}

int countDominoTilings(int h, int w) {
    return countDominoTilingsImpl(h, w, "", 0);
}

int main() {
    int tc = countDominoTilings(4, 7);  // in a 4x7 rectangle
    std::cout << tc << '\n';
}

No more class, no more worrying about const, no more worrying about memoization (it becomes the caller’s problem, for better or worse). Our original DominoTilingCounter object wasn’t thread-safe, but now we don’t have to worry about that, either. And our code is about a dozen lines shorter.


Again, this is not to say that all classes are bad! In fact, the antipattern discussed here is very close to the Builder pattern, and there’s nothing wrong with the Builder pattern — when it’s needed, that is. All I’m saying is:

When you have to compute a value, don’t write a ValueComputer class.
Write a compute_value function instead.


For other takes on this topic, see:

Posted 2020-05-28