ObjC++: fix testcase failures on Darwin

[Nicola Pero]

>  imagine if C where defined so that int i; in the presence of setjmp/ 
> longjmp were required to work.

I was thinking about the same thing :-)

Having a flag for the C compiler that enables that behaviour would be  
great.  There's plenty of C users
who use setjmp/longjmp and would like a flag that makes them safe (at  
the expense of some speed obviously).

By the way, as far as I can see, older GCCs (up to 2004) had  
setjmp_protect() and setjmp_protect_args()
that you could use to protect a function against setjmp clobbering.   
In fact, until 2002, if -traditional
was passed, GCC would automatically call setjmp_protect() on all  
functions that used
setjmp(), which is exactly what the flag above would need to do. ;-)

Maybe we could reimplement setjmp_protect() ... a vanilla  
implementation that just marks all local variables
and arguments in the function as volatile ?  Then we can call it in  
finish_function() if a -fsafe-setjmp flag is
passed and the function contains a call to sejmp ... assuming it works  
to mark the variables as volatile after
the fact, you'd get no spurious warnings since the volatile warnings  
are produced earlier.

Anyway, just brainstorming.  Maybe someone who knows more about the  
internals can chip in and suggest
if there is a good way of doing this. ;-)

Thanks