[quote name='Caburé' timestamp='1352475204' post='4999314']
That was a rethorical question, I already knew how to create a static class in Java.

He's not talking about a static class, he's talking about a normal class. Static data comes with many of its own pitfalls, even when that static data isn't truly global.
[/quote]

(I mean Java)

if I put a

public static final int x = 1;

in any normal class, that class is both global and static: is immutable because must hold the final value in at least one of its fields.. even without having static fields, in Java.

Every class in Java is static. Not every instance.

[quote name='Caburé' timestamp='1352475204' post='4999314']
That was a rethorical question, I already knew how to create a static class in Java.

who said anything about static classes?
My entire point is to avoid static stuff actually . What I meant was, what do you use to store your precomputed values? A standard class, plain and simple.
[/quote]

Yes, you are right.

In Game I might have a World class which also have Update() and Draw() methods. And in World there is a Player, but the Player class should be able to call a function in Graphics when a key is pressed. I can't do that in either Update() or Draw() since they only are passed one of Input and Graphics. In this case I find it really useful for one of them to global since I don't have to bother about this issue. I can see two options to solve this without the use of globals.

1. Change Game::Update() and World::Update() to take a Graphics* parameter.
   
2. Add a Graphics* mGraphics member variable to each class that needs it, and set it with SetGraphics(Graphics* pGraphics) on initialization (what if I forget to set it?)
   

Maybe there are more options, what do you recommend doing?

Ah, but like I noted before, I would argue here that the problem is with the design whereby "the Player class should be able to call a function in Graphics when a key is pressed." That's giving the player object responsibility over two very distinct systems (graphics and input), especially when the player appearance to be nothing more in this design than an entity within the logical world. You are now conflating, and thus coupling, your logical objects with your input and your rendering. With a bunch of singletons/globals this is less obvious, but when you remove them as you have suggested here, you discover those dependencies.

Ideally all three of those (logic, rendering, input processing) should be independent of each other. Communication between them should occur via the higher level container for each subsystem (in this case, the game itself, so perhaps in your "game" class).

The input system collects a bunch of keyboard events, or whatnot. Your game translates those (maybe via some kind of action mapper). Then it asks the world for the currently-controlled entity and uses the public interface of that entity to apply the actions. Once all that's done and the rest of the world goes through it's simulation step, the game collects all the visible/renderable entities and constructs (or updates) rendering descriptions for them, and feeds those rendering descriptions to the renderer object itself.

In this system:

• The "Player" class makes the most sense as the action mapper; all it has are a pointer to the currently-controlled world entity, and a bunch of methods to translate unmapped key codes into actions based on the entity state (for example, making the "V" key either draw or stow the entity's weapons, depending on the entity's current "are weapons drawn" state). The responsibility of the player is to change entity state.
  
• World logic, input logic and rendering are all independant, tied together at the topmost level by the game itself, which uses a Player object to translate between input and world systems.
  
• Similarly, the topmost game object also feeds input to the graphics by iterating all the active and renderable game entities and creating render descriptions (sprites, for example) using that data. The game determines whether to use animation sequence 0 (weapons stowed) or 1 (weapons drawn) for an entity based on that entity's state, but the graphics system only knows about and only has to see a bunch of sprites.
  

You arrive at the quandary you posed about the player object by incorrectly taking the OO mentality too far -- OO design is about modelling objects in real-world terms, but only to extent. Your initial description of a Player does model a human player very literally -- it is the human player who, however indirectly, causes graphics to change when they press keys. But the in-code version of that object shouldn't be developed that literally, because it leads to exactly the kind of spaghetti of dependencies you encountered.

I don't see that singletons or globals are particularly bad. I've worked on commercial games where singletons and globals have been used. One used them wisely and was the easiest codebase that i've ever worked on and one used it badly and it was extremely complicated and hard to ramp up on. I have also worked on commerical projects where "correct" OOP has been used and overused and this was also one of the hardest codebases i've worked on.

You can write bad code with "correct" practices. You can write great code with "bad" practices.

Writing tight OOP code in studios where iteration time is important and year on year you're rewriting large portions of code is common because of new platforms and new titles is a complete waste of time and it is important that you make simple architectural decisions that:

a) Allow new programmers to get on with things easily.
b) Make it easy to add features to.

It is argued that you can screw things up if you start touching global variables that you shouldn't. If you hire programmers with an ounce of common sense they generally won't do that and no matter what "barriers" you put in place to protect programmers from themselves, if a shortcut is really desired then it will be found. The friend keyword comes out in no time.

Stop worrying about writing perfect OOP and just get on with writing cool stuff for your game/engine.

Writing tight OOP code in studios where iteration time is important and year on year you're rewriting large portions of code is common because of new platforms and new titles is a complete waste of time and it is important that you make simple architectural decisions that:

Except every time I hear the 'new code because of new platforms' arguement is I can't help but think it is garbage.

Consoles have been stable for many years now and it's only recently that a nod towards the new ones has happened. We've had ~7 years of stability in that domain.
Windows is an even bigger pile of fun as DX9 was stable for the better part of a decade and again it is only in the last few years that DX11 has become 'a thing'.
Other platforms come along at a much lower frequency and, with correct abstraction, are not a case of 'rewrite large portions of code' but rework layers which are needed to get them up and running.

I would argue that by NOT taking time to design things properly up front people are wasting more time with rewrites and rebuilds of systems which don't need it. The 'get it done' mindset has its uses but hacking on things is not, in itself, a good way to go unless you want to waste time over the course of years rebuilding the problem.

a) Allow new programmers to get on with things easily.
b) Make it easy to add features to.
[/quote]

Except none of this is an arguement against designing things properly nor, more importantly when it comes to this thread, is it an arguement in favour of singletons which, in my experiance, have only served to cause more harm than good.

A correctly designed system can allow for BOTH of those things IFF it is correctly designed.
And by that I don't mean the classic 'just out of university' OOP method of 'zomg! classes for everyone!' but a sensible division of code and responcibility. People get too hung up on 'OOP = class' and lose track of the real meaning of the design methodology. More to the point they focus purely on 'put everything in a class and give it methods' than correctly splitting things up into classes, structures and free functions as required and, more importantly, ignore data flows and how information needs to move around the system.

In short; most people in the world write utterly shit OOP but that doesn't mean it isn't useful to design larger subsystems for it.

It is argued that you can screw things up if you start touching global variables that you shouldn't. If you hire programmers with an ounce of common sense they generally won't do that and no matter what "barriers" you put in place to protect programmers from themselves, if a shortcut is really desired then it will be found. The friend keyword comes out in no time.
[/quote]

The problem is that most people AREN'T good enough to not make mistakes and do it wrong.
I have seen some wonderful things where people fresh out of university have commited crimes against code so bad it makes me wonder why they are allowed to even write code in the first place. (My favorite being the guy who decided to, in his branch without telling anyone, change the interface to our lua subsystem from function(luaObject, ....) syntax to luaObject->function(....) syntax while at the same time weaving his work around the changes. It took me 4 HOURS to get his code back into mainline for a merge and I no longer trust fresh faced programmers not to fuck up in some massive manner)

So, yes, people make mistakes and yes there are ways to hack around the 'protection' BUT with a properly designed system the immediate solution is often cleaner, easier to maintain and, more importantly, does NOT lead them down the path of thinking 'hey, others have done it so it's probably ok for me to do so...' because they learn by example.

Truth of the matter is that most people in the industry can't design for crap - unfortunately if we tried to staff an industry with people who were good at it we'd soon run out of people so in order to save time later its best to discourage bad code to start with.

And to that end... a story!

In our engine we were lacking a sane method of the game team to transfer values from the logic thread to the renderer thread (and thus the GPU).
We took a day to hash out WHAT we wanted the system to do (I'll spare you the details, because frankly they are a tad confusing at times), roughly how we wanted it to interact and then I got cracking on the design and implementation details.
The two senior members of the team feedback into the process every few days as we discussed things and by the end of a two month period we had a working solution (with some of the most complicated multi-pass binding/patch table setup code I've written in my life BUT...) which was created when the renderer started up and cleanly passed, by reference, to the subsystems which required it.

However we weren't finished there because the aim for this system was to replace ALL the variables which could be passed to shaders; this included built in tokens for things like shadow map textures as well as the post-processing system's frame buffer textures and variables.

The 'global texture cache' (as I believe it was called) had been written partly by a senior and finished off by a junior. Frankly the code was overly complex for what it did BUT more importantly it had been implemented as a global system and because of this it had inflitrated multiple levels of code with no clarity as to the flow. It took LONGER for me to remove the old code and clean up the mess it had made than it took me to write the addition to the Parameter System which had been written; Clean OO intergration 1 v Global Code 0.
(The new system also expanded the abilities of the GTC functionality bring it in line with the general parameters affording the game team more control).

Post Processing was the same story; the functionality added to the Parameter System took maybe 2 days to add with no major interface changes aside from some special casing due to the stupid way the Post Processing was implemented.

Finally a few months back we needed a feature adding to the Parameter System at the request of the game team. The conversation as to if I should even add the feature took LONGER than the time it took me to write the feature, integrate it, test it and push it into the publish queue, the system was that extendable.

The system itself was built on both OOD and DOD principles to use classes when it made sense and pay attention to the data when it didn't.
The system was written in Jan/Feb of this year with a couple of weeks in March to integrate the extra functionality and remove the old.

This system is a lynch pin in the game F1 All Stars which is coming out soon; if it breaks then we lose logic->renderer data flow so it's one of the most critical systems in the whole engine right now.

The system has had ZERO time spent on bug fixing; not a single bug has been filed against it and when suspicion has been cast it has been easy to prove (despite the mad binding code which while it looks complex does work flawlessly) it wasn't a problem.

Now, we could have hacked together a system in a month, thrown it in as a global and watched as everyone abused it and dealt with bugs left, right and centre BUT instead we decided to do it right and because of that the integration is clean, the code is robust and its freed me up to work on other bits of genius instead.

[quote name='larspensjo' timestamp='1352464987' post='4999261']
[quote name='Khatharr' timestamp='1352453741' post='4999196']
I'm probably gonna get creamed for saying this, but I usually use namespaces for systems that I only want one of.

I hope you won't get, or people will eventually get afraid asking things.

I am not sure I understand how the use of namespaces limits the number of instances to a maximum of 1.[/quote]

Because you can only have one namespace with the same name per program.
[/quote]

Which does nothing to enforce that a type defined within that namespace cannot be instantiated multiple times:


namespace N {
struct C {
int someMemberVariable;

void someMemberFunction ();
};
}

int main () {
C c1;
C c2;
}

// ...


The only way you could really guarantee uniqueness out of this is to do something like:


namespace N {
namespace C {
extern int someQuoteMemberQuoteVariable;

void someQuoteMemberUnquoteFunction ();
}
}


Is that what you mean?

That transform C from an aggregate type into a namespace containing free functions (and freestanding global "members" which presumably in a real implementation you'd not actually expose in such a fashion).

That means you have only a single thing that "C" represents, but it's no different from classical procedural programming there, and subject to the same pros and cons. Much of the C runtime library functions that maintain state -- like rand() -- are essentially an implementation of this technique sans the namespace (which just provides a name scope and some organizational stability), and have many of the same problems a singleton can.