In defence of object orientation

So, rather randomly, I was discussing with @PetMac about the merits of a particular engine being split up into multiple libraries. We’d suffered from the other extreme: one gigantic project that contained everything (and had several minute link times as a result). I opined that it was, by and large, a good thing, even if it was inevitable that a lot of time be spent splitting off chunks of functionality and pushing them up or down the hierarchy of libraries to avoid circular dependencies. The alternative being of course that libraries end up tightly coupled, and even though they are two separate units, they are effectively indivisible. That is, not quite spaghetti code, but certainly a tightly snarled up ball of functionality that it would take many man-hours to pull apart. And as soon as libraries start to stick together like that, the rot sets in quickly; one  reference turns to dozens, and even if it might have been possible to separate them again before, it isn’t feasible any longer. I think (and he can correct me if I’m misstating his opinion) that Pete agreed on that front.

Why is that relevant to object orientation? Well, because the means by which you most commonly ‘fix’ a circular dependency is to abstract one side so that it becomes unaware of the other. So your character system knows about your audio system because characters carry around audio references; but instead of your audio system being aware of characters (so that they can, say, position sounds), you rewrite your audio system in terms of ‘character-like things’. Or more cleanly, ‘things with position’. Because that’s all the audio system really needs to know about. In an object oriented system, you’d use an interface definition to say that characters can satisfy the definition of a ‘thing with position’; of course that’s not the only way to achieve the same goal, but it’s certainly a nice easy way to do it. What’s important is that the library has a nice clean interface, that expresses exactly what it needs, in a succinct way. Ideally, it is also written without explicit knowledge of any other libraries. Having a clean and clear interface is what helps you keep that lovely de-coupled code, and lets you re-use it elsewhere.

Personally, I’ve never had a problem with using interfaces or other object-oriented mechanisms. But recently Pete has been trying to persuade me that object orientation is the dark side, and that our code would be much better if we only thought about things in terms of data transforms. There’s been a lot of eminently sensible stuff written on it, including stuff by @noel_llopis over on his blog, and by @TonyAlbrecht in a talk for GCAP. I’ve read their pieces, and don’t really disagree with most of it. If I have an issue at all, it is that their concerns about OO (and C++ specifically) primarily relate to performance, and when I’m coding, performance is only one factor; an equally pressing factor is how easy the code is to write and maintain.

Here’s the thing though; object-orientation can be really bad for performance, sure. And used badly, it can be really bad for design as well. C++ has a whole lot of cruft that means that expressing the code design you want, without locking yourself into bad assumptions, is hard. Not impossible, just hard. But there are a whole lot of code design needs I have which are very hard to satisfy without the basic features of C++. Interfaces and polymorphism straight off, and probably more. Really though, my problem lies with anyone that tells me that we should all go back to using C instead of C++, because it will avoid all of that bad stuff. Well, sure. I could go back to writing in assembly and never worry about variable aliasing as well, but I’m not going to. I’ll use C-style interfaces when they help, and C++ when they help, thank you very much. Whatever gets me the simplest, cleanest, most maintainable interface, that still lets me do the work.

I have no doubt that using C-style library interfaces would avoid a lot of unnecessary object-orientation. @PetMac is trying to persuade me though that a C-style interface is just plain better, and not only that, but that the inputs and outputs should only be structures defined in the library interface. So an audio transform would be ProcessAudioEmitters, and if you want to process a bunch of positional audio emitters, one for each character, you have to marshal an array of audio emitter structures, and copy the position from your character into its audio emitter. Which doesn’t sound so terrible, if it leads to a cleaner interface. I’d probably be fine with that. At a simple level, for core systems like audio or rendering, where the inputs and outputs are clear and rarely change, I think that would probably work well. Best of all it makes the audio library completely independent – it knows nothing of the things that it’s working with, except the data the other systems choose to feed it.

My problem comes when I consider how I would make that approach scale to all the other systems I need to build. The example I posed to Pete was one of an AI system. To use Pete’s preferred paradigm, and think of data transforms, the AI system would be a DecideWhatToDo transform. Great. What are the inputs and outputs? Well, that depends on the AI. One type of AI might want just a list of character types and positions. Another might want to know about the environment as well. Smarter AI might want character positions, damage levels, movement histories, factional allegiances; as well as the ability to co-operate with other characters. The outputs of the AI are just as bad – they can affect everything from the desired target position, to queuing animations, in fact pretty much anything a character can do might be an output of the AI.

I would describe Pete’s system as a ‘push’ system. Everything the system needs has to be fed to it explicitly, in terms it can understand. The problem with push systems though is that when the number of inputs goes up, the amount of code you have to maintain just for the marshalling of the push grows with it. You find yourself implementing the same code several times: you add the notion of damage to the character, then you have to add the ability to marshal the damage information into a structure the AI system would understand, then you have to add the notion of damage to every single AI interface that wants to know about damage. And in a system with dozens of different sorts of AI, that might be a lot of interfaces.

To me that smells wrong. It means that you’re baking implementation details (like AI ‘X’ cares about damage) into the interface. Conversely, the ‘pull’ system stays relatively clean. You simply pass the list of characters to the AI, or the environment, and allow the AI system to ask the character interface for only the data it needs. Characters might provide a vast array of query-able state, and the AI can pick and choose what it asks for. Of course this comes with a down side. The AI system now has to have knowledge of the character system (or at least, provide abstractions which the character system can fulfil). It’s no longer truly independent. The performance impact of lugging over the entire character object, when perhaps you only want to access a few small parts of it, is very real. But in terms of the ability to write clean code, without a massive amount of interface book-keeping, it’s a big win. That said, I’m open to persuasion. If someone can describe to me how they would write a succinct AI library interface in a C-style, for a few dozen varied and sophisticated character AI, without giving the AI library knowledge of the character interfaces, I’d be happy to change my point of view.

There will be those who say that if your structures are that complex, you’ve already done something wrong. That’s very idealistic thinking. The simple fact is that we are often writing fantastically complex simulations. Sometimes the ‘pure’ systems that you’d need to build to support the level of complexity the design calls for are just far more effort than the benefits they would give. When it comes down to it, we need to write code effectively more than anything else. We need to be able to code quickly, cleanly, and flexibly; especially when the game design is changing quickly as well. It’s of no benefit at all to spend months building a fantastically clean engine to support one game design, only to find that in the time it took you to build it, design changes have rendered it obsolete.

To sum up, because I’ve gone on for a long time: the one thing I like less than being accused of ‘drinking the OO kool-aid’, is the notion that there’s only one right way to do things. As a coder, you should be constantly and critically evaluating all your systems and interfaces. Sometimes a data oriented approach is better: consider the purity of the interface and the vastly improved ability to parallelise and minimise your memory accesses. Other times the structures and inter-dependencies are simply too complex, and object orientation is the most effective tool at keeping your code clean and versatile. I won’t claim to always get it right (as Pete and Tim have both at various times pointed out, I tend to over-structure my code), but I’d hope I always aim for clean code as best I can.

One Response to “In defence of object orientation”

  1. MrCranky Says:

    In response to some further discussion on Twitter, to give some examples: I’m talking about the difference between:

    vector<Command> DecideWhatToDo(const vector<CharacterInstance>& characters, const Environment& environment, const Factions& factions);


    struct AICharacterInput {
    Vector3 m_position;
    float m_damage;
    float m_damageReceivedInLast30Seconds;
    vector m_movementHistory;
    //information about any other detail the AI might want to make its decisions.

    struct AIEnvironmentKnowledge {
    //Everything that you'd need to describe the environment

    struct AIFactions {
    //Everything that you'd need to describe the factions

    struct AICommand {
    //Something to describe the output of the AI processing

    vector<AICommand> DecideWhatToDo(const vector<AICharacterInput>& characters, const AIEnvironmentKnowledge& environment, const AIFactions& factions);

    You see what I mean? The former doesn’t change when the details of the AI implementation requires another part of the character information: a character is still a character, the environment is still the environment. Everything that a character can be is encapsulated in the character interface.

    The latter essentially ends up duplicating everything in the other libraries. Add something to the Environment that you want to query in the AI, you have to add it to the AI interface as well, and you have to add code in the caller to copy information from the character instance into the AICharacterInput. Yes, you gain performance in the actual processing of the AI, but you also lose performance when you have to convert the information from its original form (e.g. the version contained in the character structure) into the form the AI wants to use.

    Worse, if different AI processing needs different sets of input, then you’re faced with the choice of either a) duplicating the interface structures and having different input layouts for each AI, or b) copying information into an AI structure that the AI won’t actually need.

Black Company Studios Limited, 14 Belford Road, Edinburgh, EH4 3BL
Registered in Scotland (SC283017) VAT Reg. No.: 886 4592 64
Last modified: June 17 2014.