Re: NSConnection bugs

[Philippe Roussel]

Le mardi 29 mars 2011 à 09:25 +0100, Richard Frith-Macdonald a écrit :
> On 29 Mar 2011, at 01:01, David Chisnall wrote:
> 
> > Hello the list,
> > 
> > I spent some time with valgrind and found out why NSConnection is so flakey 
> > on 64-bit platforms.  It seems that we use the same incorrect pattern in a 
> > lot of places throughout the class.  We probably use it in other places, so 
> > it would be worth doing a brief audit for this bug before the next release. 
> >  I had a quick look, and I think that NSConnection is the only place where 
> > we use GSIMap with something other than pointers.
> > 
> > The problem is in statements like this:
> > 
> >      node = GSIMapNodeForKey(IreplyMap, (GSIMapKey)seq);
> > 
> > Looks fine?  The problem is that (GSIMapKey) is a union (something 
> > concealed by the fact that it is a macro - unions are difficult to use 
> > safely, and anything that hides the fact that you are using a union is 
> > generally a bad thing).  Specifically, it's a problem that GSIMapKey is a 
> > union of void*, id, unsigned, and int.  On most current 64-bit platforms, 
> > unsigned and int are 32-bit quantities, while void* and id are 64-bit 
> > quantities. 
> > 
> > The problem, therefore, is that the we are defining 4 bytes of an 8-byte 
> > quantity and then (in the callee) accessing all 8 bytes.  This results in 
> > some random behaviour.  The connection.m test was failing because sometimes 
> > the uninitialised bytes were 0 (so the map was finding the value it was 
> > looking for) and sometimes they were some other value from the stack.
> 
> Well spotted.  I've implemented a fix for that throughout base.

Seems to break on x86-64 gcc 4.4.3 :

GSAttributedString.m:147: error: cast to union type from type not present in 
union

Philippe