Changing the gears to Generic Programming

Practical usage of generic programming can be narrowed down to two forms:

  1. Writing top level generic functions
  2. Writing generic custom types

Generic Functions

Writing top level generic functions is easier. Its more of like you write a normal function and later on you realize that this function is not restricted to any type, rather it is just an abstract algorithm that can be used by many types, you upgrade it to generic type. For example,

bool isEqual(int x, int y) {
    return x == y;
}

can be easily abstracted out to a generic form such as

template <typename T>
bool isEqual(T x, T y) {
    return x == y;
}

This generic form would work as long as T supports a == operation. Which brings us to the term you will often find with generic programming, Type Constraints.

So, the above example would work fine with all types that support == operation. Such as:

std::cout << (isEqual(10, 10) ? "Y" : "N") << std::endl;
std::cout << (isEqual(10.5, 10.5) ? "Y" : "N") << std::endl;
std::cout << (isEqual("hello", "hello") ? "Y" : "N") << std::endl;

But would easily break down when used like:

struct Vector2 {
    float x, y;
    Vector2(float xx, float yy) : x(xx), y(yy) {}
};

std::cout << (isEqual(Vector2(10, 10), Vector2(10, 10)) ? "Y" : "N") << std::endl; // output compilation failure

And the reason being, that the Type Constraint has been violated. The type Vector2 does not provides an == operation. The fix is simple, satisfy the Type Constraint

bool operator == (const Vector2 &lhs, const Vector2 &rhs) {
    return (lhs.x == rhs.x) && (lhs.y == rhs.y);
}

Another type of error that could arise with generic functions is, where the constraints are not violated syntactically, by are logically broken. For example:

const char w1[] = "world";
const char w2[] = "world";
std::cout << (isEqual(w1, w2) ? "Y" : "N") << std::endl; // output N

Here, our isEqual() sees that the types are const char * and the compiler knows how to compare two pointers. But the result might or might not be what we were expecting. If you were expecting a string comparison, you’ve to tell the compiler how to do that.

For our specific case, it can be achieved by providing an overloaded implementation that just takes const char *

bool isEqual(const char *x, const char *y) {
    return strcmp(x, y) == 0;
}

Generic Types

The main power of generic programming lies when working with generic types. Just as generic functions abstract the algorithm to a type-free level, generic types abstract our design to type free level.

Think of generic types as if us providing the type information for the compiler, and passing the responsibility of creating the actual types to the compiler.

To get an quick overview, the Vector2 type that we wrote above for float types can easily be rewritten as a generic type

template <typename T>
struct Vector2 {
    T x, y;
    Vector2(T xx, T yy) : x(xx), y(yy) {}
};

And the == implementation should just fit with the change:

template <typename T>
bool operator == (const Vector2<T> &lhs, const Vector2<T> &rhs) {
    return (lhs.x == rhs.x) && (lhs.y == rhs.y);
}

Now, where should we actually use the generic types in real world? To address this question, you’ve to switch you mindset from run-time polymorphism to compile-time polymorphism. Let’s consider an example where you have designed your types in the classical object-oriented fashion, that is using inheritence

class Shape {
public:
    virtual ~Shape() {}
    void PrintArea() {
        std::cout << "area = " << Area() << std::endl;
    }

private:
    virtual float Area() const = 0;
};

class Rectangle: public Shape {
public:
    Rectangle(float w, float h) : width(w), height(h) {}

private:
    float Area() const {
        return width * height;
    }

    float width, height;
};

class Circle: public Shape {
public:
    Circle(float r) : radius(r) {}

private:
    float Area() const {
        return 3.14 * (radius * radius);
    }

    float radius;
};

int main() {
    Rectangle *r = new Rectangle(10, 20);
    Circle *c = new Circle(5);

    std::vector<Shape *> shapes;

    shapes.push_back(r);
    shapes.push_back(c);

    std::for_each(shapes.begin(), shapes.end(), [](Shape *s) {
        s->PrintArea();
    });

    delete r;
    delete c;

    return 0;
}

Here, we have a base Shape class that does provides the public interface and leaves the actual implementation details to the subclasses. In this case the actual calculation of the area. Internally, we’re calling the Area() function on Shape and let the run-time do the actual look-up for the implementation by looking at the type information.

In another words, what we are actually doing is that delegating some part of the implementation of our Shape class to subclasses. Another way of achieving the same results is by not depending on subclasses for overrides, but delegating the implementation to some external classes.

struct ShapeImpl {
    virtual float Area() const = 0;
};

class RectangleImpl: public ShapeImpl {
public:
    RectangleImpl(float w, float h) : width(w), height(h) {}

private:
    float Area() const {
        return width * height;
    }

    float width, height;
};

class CircleImpl: public ShapeImpl {
public:
    CircleImpl(float r) : radius(r) {}

private:
    float Area() const {
        return 3.14 * (radius * radius);
    }

    float radius;
};

class Shape {
public:
    Shape(const ShapeImpl *impl) : implementation(impl) {}

    void PrintArea() const {
        std::cout << "area = " << implementation->Area() << std::endl;
    }

private:
    const ShapeImpl *implementation;
};

int main() {
    RectangleImpl *r = new RectangleImpl(10, 20);
    CircleImpl *c = new CircleImpl(5);

    std::vector<Shape> shapes;

    shapes.push_back(Shape(r));
    shapes.push_back(Shape(c));

    std::for_each(shapes.begin(), shapes.end(), [](const Shape &s) {
        s.PrintArea();
    });

    delete r;
    delete c;

    return 0;
}

I don’t know if we’ve achieved any actual improvement over the earlier implementation with this or not. But, this gives you an idea that the implementation can be dragged out of the core class while keeping the visible interface of the Shape intact.

If we can achieve this, we can even go a step further and actually provide the implementation at compile time. In other words, we can simply make the Shape class generic, and let the compiler plug-in the implementation details at compile-time.

template <typename Impl>
class Shape {
public:
    Shape(const Impl impl) : implementation(impl) {}

    void PrintArea() const {
        std::cout << "area = " << implementation.Area() << std::endl;
    }

private:
    const Impl implementation;
};

Now, we can create and use our generic shape instances as:

    Shape<RectangleImpl> rect(RectangleImpl(10, 20));
    Shape<CircleImpl> circle(CircleImpl(5));

    rect.PrintArea();
    circle.PrintArea();

The only type constraints on the RectangleImpl and CircleImpl is that they need to have a Area() function.

class RectangleImpl {
public:
    RectangleImpl(float w, float h) : width(w), height(h) {}

    float Area() const {
        return width * height;
    }

private:
    float width, height;
};

class CircleImpl {
public:
    CircleImpl(float r) : radius(r) {}

    float Area() const {
        return 3.14 * (radius * radius);
    }

private:
    float radius;
};

The C++ compiler is smart enough to see the constraints are not violated by any new delegate class. For example if we introduce a TriangleImpl as:

class TriangleImpl {
public:
    TriangleImpl(float b, float h) : base(b), height(h) {}

private:
    float base, height;
};

Shape<TriangleImpl> triangle(TriangleImpl(10, 5));
triangle.PrintArea();

The compiler will immediately throw error messages, unless you fix your implementation by providing the Area() function.

class TriangleImpl {

public:
    TriangleImpl(float b, float h) : base(b), height(h) {}

    float Area() const {
        return 0.5 * base * height;
    }

private:
    float base, height;
};

Now, if you notice that we are not using an std::vector anymore, that is because we don’t have a Shape type anymore. And if you find yourself thinking in this direction, you need to get out of the run-time polymorphic mindset. When using compile-time polymorphism you’ve to keep in mind the fact that now you’re not in control of creating the custom types, rather you’ve passed on that authority to the compiler.

So, Shape<TriangleImpl> and Shape<CircleImpl> are entirely different types with no common ancestor or anything.

When designing with compile time polymorphism you need to keep in mind that your generic type should be a leaf node or a final class in the entire type system. You should not have to subclass your generic type, rather plug-in a delegate class that provides the implementation that you would rather provide by sub-classing.

To summarize, here’s a quick list of good and bad of generic type system:

Pro

  1. Robust: No run time exceptions.
  2. Efficient: No run time lookups

Cons

  1. Not a good candidate for base class
  2. Hard to read and debug.

Now, moving on to using generics with Swift. Swift provides the same good old C++ way of writing generic code with some extra type constraint system. Where in C++ the compiler would do the type constraint checking when you actually compile the code, Swift actually makes the constraint system more explicit, such that you have to provide the constraint information with a protocol.

protocol ShapeImplementable {
    var area: Double { get }
}

struct Shape<Impl: ShapeImplementable> {
    private let implementation: ShapeImplementable

    init(impl: ShapeImplementable) {
        implementation = impl
    }

    func printArea() {
        println("area = \(implementation.area)")
    }
}

And then you can implement your delegate class as

struct RectangleImpl: ShapeImplementable {
    var area: Double {
        return width * height
    }

    init(w: Double, h: Double) {
        width = w
        height = h
    }

    private let width: Double
    private let height: Double
}

And finally use them the same way you would with C++.

let rect = Shape<RectangleImpl>(impl: RectangleImpl(w: 10, h: 20))

The entire code for this article is also available at github.com/chunkyguy/GenericsDemo

Goodbye and have a nice day!