4 ms·

Polymorphism in C

8 points by gwoplock 5y ago

kazinator 5y ago
ShapeInterface *CircleAsShape = &(ShapeInterface) { .Area = (double (*)(void *)) circle_Area }; Phooey! Just ShapeInterface CircleAsShape = { .Area = (double (*)(void *)) circle_Area }; then refer to &CircleAsShape. This is error-prone because of the cast. circle_Area could have the wrong parameters, or the wrong function could be used by accident and it will compile. Conformance wise, it is undefined behavior: a function that takes a Circle * is being called via a pointer to a function that takes a double *. You really want the circle_Area to have a Shape * argument. Firstly: double shape_Area(Shape *shape) { return (shape->interface->Area)(shape); // we lose shape->instance here, just shape } Then: double circle_Area(Shape *shape) { Circle *circle = (Circle *) shape->instance; ... } We now have a cast here; but that is now new. The same conversion was being done through the function pointer punning, without a cast (formally undefined in ISO C). Now we have well-defined behavior: the shape->instance pointer is really pointing to a Circle; it is valid to convert that pointer back to a Circle *. Then we have: ShapeInterface CircleAsShape = { .Area = circle_Area }; So instead of a function pointer cast with undefined behavior, we have a well-defined function pointer treatment, with a correct cast between data pointers. Another nice thing now is that type-specific implementation like circle_Area has access to the Shape pointer, and can invoke shape operations on it. double rectangle_Area(Shape *shape) { double length = shape_NominalLength(shape); double width = shape_NominalWidth(shape); return length * width; } Now you can have subtypes of rectangle which have a completely different internal organization, and don't have to reimplement Area; rectangle_Area works for them. E.g. a PNG image could be a kind of rectangle, using rectangle_Area for its Area implementation. It just provides NominalLength and NominalWidth: double pngImage_NominalLength(Shape *shape) { PNGImage *image = (PNGImage *) shape->instance; return PNGIMage_width(image); // PNGImage is an opaque type from some third party lib. } Whereas if rectangle_Area is like this: double rectangle_Area(Rectangle *rect) { return rect->length * rect->width; } that is not possible, even if you're prepared to modify rectangle_Area somehow to try to make it work. It doesn't have access to the shape container to be able to invoke the abstract shape call; there is no way it can work with PNGImage.

8 points by gwoplock 5y ago

ShapeInterface *CircleAsShape = &(ShapeInterface) { .Area = (double (*)(void *)) circle_Area }; Phooey! Just ShapeInterface CircleAsShape = { .Area = (double (*)(void *)) circle_Area }; then refer to &CircleAsShape. This is error-prone because of the cast. circle_Area could have the wrong parameters, or the wrong function could be used by accident and it will compile. Conformance wise, it is undefined behavior: a function that takes a Circle * is being called via a pointer to a function that takes a double *. You really want the circle_Area to have a Shape * argument. Firstly: double shape_Area(Shape *shape) { return (shape->interface->Area)(shape); // we lose shape->instance here, just shape } Then: double circle_Area(Shape *shape) { Circle *circle = (Circle *) shape->instance; ... } We now have a cast here; but that is now new. The same conversion was being done through the function pointer punning, without a cast (formally undefined in ISO C). Now we have well-defined behavior: the shape->instance pointer is really pointing to a Circle; it is valid to convert that pointer back to a Circle *. Then we have: ShapeInterface CircleAsShape = { .Area = circle_Area }; So instead of a function pointer cast with undefined behavior, we have a well-defined function pointer treatment, with a correct cast between data pointers. Another nice thing now is that type-specific implementation like circle_Area has access to the Shape pointer, and can invoke shape operations on it. double rectangle_Area(Shape *shape) { double length = shape_NominalLength(shape); double width = shape_NominalWidth(shape); return length * width; } Now you can have subtypes of rectangle which have a completely different internal organization, and don't have to reimplement Area; rectangle_Area works for them. E.g. a PNG image could be a kind of rectangle, using rectangle_Area for its Area implementation. It just provides NominalLength and NominalWidth: double pngImage_NominalLength(Shape *shape) { PNGImage *image = (PNGImage *) shape->instance; return PNGIMage_width(image); // PNGImage is an opaque type from some third party lib. } Whereas if rectangle_Area is like this: double rectangle_Area(Rectangle *rect) { return rect->length * rect->width; } that is not possible, even if you're prepared to modify rectangle_Area somehow to try to make it work. It doesn't have access to the shape container to be able to invoke the abstract shape call; there is no way it can work with PNGImage.