[Bug libobjc/38307] Calling of the +initialize method is not completely thread-safe

["rfm at gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38307

--- Comment #4 from rfm at gnu dot org 2011-03-16 12:43:39 UTC ---
I just searched this old bug report up, and found that I'd missed seeing (or
forgotten about) the latest response :-(

I've been running with my patch for a couple of years without problems, but
when I tried David's patch out, practically all my gnustep code simply started
crashing on startup ... so there's something wrong there but I haven't had time
to see what it might be.

However, to address points in the last comment ...

> I still have a hard time groking what was intended with the receiver.  It all
> seems very intertwined and I think there is a more straight forward way to
> implement this.  

I can't argue with that ... it *is* complex :-(

> Also, with this patch get_imp fails on class methods.  (get_imp also has the
> nasty effect of installing the dispatch table without calling +initialize and
> the same goes for __objc_responds_to).

Very good point ... it hasn't been an issue for me in practice, but that
clearly is something which needs to be addressed.  However, I'm not sure that's
a new bug ... haven't those functions always bypassed +initialize?  I'd have to
check, but I agree that they really should call +initialize

> I'm not to fond of introducing InitializingList as special type. I think should
> be fine with using the existing hash map tables for this. I don't think we
> really need to introduce a new type.  Do you really think that method dispatch
> for partially installed dispatch tables is performance critical?

Certainly there's no performance issue with method dispatch in initialize ...
it should be rare enough that we can use a really slow implementation. 
However, we have to build the dispatch table for the class, so it seemed
reasonable to use it within the thread calling +initialize even though we don't
want to install it for other threads until +initialize has completed.  Perhaps
we can store the uninstalled dispatch tables in a hash map more simply than
using a new type ... that sounds reasonable.