Re: sending notifications in -dealloc

[Richard Frith-Macdonald]

On Friday, May 31, 2002, at 03:16 PM, Helge Hess wrote:

> Helge Hess wrote:
> > This doesn't work on gstep-base, because the NSNotification created 
> > retains and afterwards releases the "object:self". This somehow 
> > results in a duplicate call to -dealloc (because the RC apparently 
> > reaches "0" twice).
> > Is anyone aware of differences with regards to RC in -dealloc or has 
> > any other ideas ?
>
> some more info, NSObject -release looks like this in libFoundation:
> ---snip---
>     if (NSExtraRefCount(self) == 1)
>         [self dealloc];
>     else
>         NSDecrementExtraRefCountWasZero(self);
> ---snap---
>
> and like that in gstep-base:
> ---snip---
>   if (NSDecrementExtraRefCountWasZero(self))
>     {
>           [self dealloc];
>     }
> ---snap---
>
> (some code removed).
>
> Actually I think that libFoundation is correct (and apparently it's 
> implemented in the same way in MacOSX Foundation, I'll check that in 
> the evening at home).
>
> Gstep-base is implemented like it's written down in the documentation 
> "If anObject didn't have an extra reference count, it returns YES, 
> indicating that the object should be deallocated (with dealloc)". But 
> this breaks if the count isn't checked (resulting unnecessarily in 
> multiple -deallocs).
>
> Opinions ?

While GNUstep is of course 'right' :-) (ie as specified/documented), but 
we want compatibility.

I think that the fact that NSDecrementExtraRefCountWasZero() returns a 
value implies it is
intended to be used - which suggests that the libFoundation behavior is 
not quite in line
with MacOS-X either.

Perhaps NSDecrementExtraRefCountWasZero() and NSIncrementExtraRefCount() 
should simply not
operate once the extra ref count goes to (unsigned)-1  (with the former 
returning NO) ?