Do we have a policy on how to prevent this from happening in the future? i.e. what's the right way to implement -initWithCoder and -init? I think we should have a standard pattern we use everywhere.
- We don't want them to share large chunks of copy-and-pasted code since this is error prone.
- Calling -init (or the designated initializer) directly from -initWithCoder could cause problems since it could conflict with the required call to [super initWithCoder: ];
- The only idea I can think of right now is factoring out the common code in to a static function like this:
- (id) init
self = [super init];
- (id)initWithCoder: (NSCoder*)coder
self = [super initWithCoder: coder];
if ([coder containsValueForKey: @"foo"]) [self setFoo: [coder decodeObjectForKey: @"foo"]];
if ([coder containsValueForKey: @"bar"]) [self setBar: [coder decodeObjectForKey: @"bar"]];
static void privateInit(MyClass *self)
self->someCArray = malloc(100);
self->someValue = 123;
Note that the privateInit function only sets up ivars belonging to the current class, so each superclass would have its own privateInit function to set up the ivars belonging to that superclass.
How does that pattern look?
Not bad at all ... I like the use of a static function for common initialisation code (better than a private method) ... I think it's used in a few places in gnustep but it's not, in my opinion, a general solution. I think the situations are just too variable to support a general/standard pattern.
For the vast majority of classes there is little/no special initialisation required ... the classes contain simple instance variables whose values should either default to zero/nil or be initialized by the decoding methods (or by the copying methods) directly, and we should avoid doing extra work.
In the *really* complex cases where you have complicated data structures to construct, my preference is to use a private class rather than a lot of instance variables in the public class ... in this case the code for -init, -initWithCoder:, -copyWithZone: etc will all just call something like 'ivar = [PrivateClass new];'
A good design principle here is lazy initialisation ... complex/expensive data structures are set up only at the point there they are actually used ... often we don't create them on initialisation of the object.
Between these two extremes we have a other solutions:
Where there is no superclass method which needs calling (very often the case), then all the other methods should just call the designated initializer (usually -init) ... I think this is the most 'standard' pattern as it is the way initialisation is normally done, and it's more intuitive for these odd initialisation cases to follow that standard principle.
Where there is superclass initialisation involved, as you point out, we can't just call the -init method. In this case I think your solution of using a static function or macro to initialise things can be valuable.
Finally, it's sometimes worth using a private method (eg -_init) for instance in a class cluster where we actually want subclasses to be able to modify/replace parts of the superclass initialisation ... but I'd avoid this as a rule.
In summary ... there's no standard pattern to use everywhere, but there are a few principles:
1. design classes where ivars work when they are initialised to zero by default, so no extra support is needed.
2. don't do complex data structure setup until it's actually needed, and then try to use a static function or use a private class to encapsulate the complexity.
3. keep initialisation in the designated initialiser if possible.