September 90 - Meta-spaghetti
Donald E. Carlile
(BILL: I FORGOT TO GET A COMPANY NAME FROM DON. WANT TO GIVE IT A SHOP?)
I'm a fairly conscientious programmer-I like to think I write good clean code; I take the trouble to use modular techniques; I even stood on my head to learn object-oriented techniques. But in spite of all this, I found out I was writing spaghetti.
Spaghetti code! Someone still uses GOTO!?
Well, possibly I exagerrate: I don't mean spaghetti code in the old BASIC GOTO sense-```I mean spaghetti code that is only possible when you are writing in an OOP language. Maybe we'd better call it meta-spaghetti.
I found I was writing meta-spaghetti when I was asked by someone if I would let him use some of my objects. I had known I had only one unit, but I hadn't realized how tangled it was. It was then that I developed the term meta-spaghetti to define the condition of my code.
What I mean by meta-spaghetti is objects which refer to each other recursively. That is, objects which in their definitions include fields of each other's type, as in
TMyDocument = OBJECT(TDocument)
fStar : TStar;
TStar = OBJECT(TObject)
fMyDocument : TMyDocument;
This leads to strange loops and to units which cannot easily be decomposed into simpler entities.
As anyone who has read even part of Hofstader's Gödel, Escher, Bach knows, it is impossible to entirely banish strange loops: you made that reference in the object definition for some reason, and it cannot be easily abandoned without some thought and work. That being said, you might still want to introduce some measure of isolation between your objects.
I can tell you from experience that it is difficult to undo the damage once it's done. As I was unraveling my code, however, I developed some guidelines for avoiding and dealing with meta-spaghetti. Perhaps they will prove helpful to you as well, as you design your code from the beginning to avoid this condition.
1. References should only to be to objects down the ownership chain.
As I was defining my set of rules, this is the first one that became apparent: when defining an object, references can be made to the internals of the objects it owns. This is what I mean by references down the ownership chain. (For a more complete discussion of the ownership chain, see the author's Chains Required, Whips Optional, an unpublished, and indeed unwritten, article.)
As an example, in my Trek-style game I had a method to add text to a textedit window. This method belonged to my document object. In order for my ship objects to add text about their status to the window, they had to have a reference to my document object. This caused a problem when I wanted my ship objects in a separate unit so that others could use them.
I solved this problem by changing the way I added text to the window. Instead of calling the document method, I added a text handle field to the ship object. I then changed my document DoIdle to check the texthandle of each ship object. In this way I followed the rule of having only references down the ownership chain.
If this rule is followed, the separation of objects into units becomes very clear and simple. There is no confusing interdependence of objects, and it is clear what the unit dependencies should be.
There is only one problem with this rule: it's not always possible or desirable to follow it. Some objects need to be interrelated; when this is the case, they probably should not be separated. This brings us to the next rule:
2. When it is impossible to remove references between objects, include them in the same unit.
The universe of my game is defined by two object types, TGalaxy and TQuadrant. These two object types refer to each other quite a lot, and it makes a great deal of sense for them to both be in the same unit. It is also unlikely that they would need to be separated to be reused. Therefore, I have put them into their own unit.
3. When it is impossible to untangle references and they must be in separate units, define a new object to make the references and push it down the ownership chain.
Don't be afraid to revise your structure. Although it will save you work to think out your structure ahead of time, sometimes this proves impossible. When you find you have painted yourself into a corner, cut a new doorway.
In my original design, my TMyDocument object owned all the lists of quadrants and other data containers. Thus, all my data objects referred back to it in order to refer to each other. When I wanted to separate out my ships, stars and quadrants into other units to make them usable elsewhere and to decrease compile time, I found the task impossible.
I then defined a new object, a TGalaxy, to contain all the objectionable fields and methods (no pun intended) and made a new field in the revised TMyDocument object to refer to an instantiation of TGalaxy. This made it possible to cleanly make the separation.
Make vanilla references when Rule 1 is impossible.
Another strategy for avoiding meta-spaghetti is to make references to the parent type rather than the type that is giving you trouble. This is not always possible, but sometimes the method or field you need to reference is contained in the object's parent or ancestor.
In my game, I call the draw method of my TQuadView object type from one of the methods of TQuadrant. I do this to avoid the flicker introduced when I use the invalid rectangle methods of MacApp. Originally TQuadrant included a field fMyDocument of type TMyDocument. TMyDocument, of course, had a reference to an fQuadView, of type TQuadView. TQuadView was a specialized descendant of TView. When I needed to do a drawing, I called fMyDocument.fQuadView.Draw. In this way, TQuadrant absolutely needed to be in the same unit as TQuadView. It was also specifically geared to my application only.
To untangle this particular mess, I added a field fQuadView, of type TView, to my TGalaxy object. Since variables of an ancestor type can contain any descendent, this field can contain an object of the specialized type TQuadView. The only limitation is that specialized fields and methods may not be called using this reference.
At the time an object of type TQuadView is made, it is passed to fGalaxy of the TMyDocument. Then I can call Draw, since Draw is one of the methods of the ancestor TView class.
5. Define proto-objects for objects up the chain when Rule 1 is impossible.
Sometimes all the above rules fail: you still want to break up objects into different units but they can't come completely apart. When this happens, it is important to remember that it is permissible to define objects or abstract classes which have no instantiation. You can define these "proto-objects" to contain those fields and methods which are necessary for the objects lower down the ownership chain and can include these classes in the same unit with those objects. Then USE that unit in the unit which defines the real objects and make them descendants.
As I mentioned earlier, TGalaxy and TQuadrant define the universe of my game. I found it impossible to completely divorce these objects from the objects I refer to, i.e., my ships, stars, bases,and torpedoes. I reasoned further that any structure which made use of my entities would also need some kind of quadrant and galaxy. I then defined a TProtoGalaxy and TProtoQuadrant, which had references for the fields and methods necessary for my entity unit, while not burdening the unit with the heavy details and the bulk of the fields and methods. I then defined TGalaxy as a descendant of TProtoGalaxy and TQuadrant as a descendant of TProtoQuadrant.
This one might be just a bit tricky to follow. Let me try and restate it thusly: the Galaxy and Quadrant objects have several fields and methods which ships and things need to reference. Two examples of such fields are the galaxy array and the list of other entities in the quad. An example of a referenced method is the one which detects when collisions occur. I therefore defined objects, which I call proto-objects, which contain references to these things. I don't define the methods-e.g., the collision procedure in the protoQuad consists of a Begin and an End, as do the rest of the methods in my proto-objects. In a different unit I then create descendants of the proto-objects with all the fields and fleshed-out methods fully defined.
I don't claim that these rules are the simple once-and-for-all solution to the problem of meta-spaghetti, but they are a start. You can probably come up with a few rules and methods of your own. Good luck!
Rules for avoiding meta-spaghetti:
- References should only to be to objects down the ownership chain.
- When it is impossible to remove references between objects, include them in the same unit.
- When it is impossible to untangle references, and they must be in separate units, feel free to define a new object.
- Make vanilla references when Rule 1 is impossible.
- Define proto-objects for objects up the chain when Rule 1 is impossible.
If you have any questions or comments for Don, he can be reached on AppleLink at N0231.