Re: bug in NSPortCoder

[Richard Frith-Macdonald]

On Friday, April 11, 2003, at 03:50  pm, Enrico Sersale wrote:

> With the last cvs DO doesn't work anymore.
> I'm not able to find exactly where the bug is but it must be in
> NSPortCoder.m because changing only this file with a copy taken from 
> the
> yesterday cvs all is working again.
> I don't know if this can help, but, in the ChangeLog there is:
>
> * Headers/gnustep/base/NSPortCoder.h: Add array for pointers
>
> but NSPortCoder.h is identical to the old one.

I've fixed that ... I inadvertently committed incorrect versions of 
NSPortCoder.[hm],
noticed and reverted them, but forgot to fix the ChangeLog.  I've 
corrected the
ChangeLog entry now.

> This is the output of nsconnection_client in Testing:

<snip>

> ot (Ny String 3)
> Small Struct:
>   sending ptr to c got 3b
> Struct:
> Segmentation fault (core dumped)

Well ... I've been doing quite a bit on this, to do with field 
alignment in structures.
In gcc-3 __alignof__ has been changed so that it no longer returns the 
alignment
used to build structures.  In particular, on intel it now returns 8 
rather than 4 for
the alignment of a double (but still uses 4 to lay out its structures).

Someone added some functions to libobjc (quite a long time ago) to work 
out
alignments, and I've modified the base library archiving code to use 
this stuff.

I modified the nsconnection tests to pass a structure which would 
demonstrate
this alignment problem.

While I was looking at this I found a bug in handling pointers to 
structures, which
could cause memory to be scribbled on, and fixing that exposed another 
obscure
memory management problem when returning a c-string as both an argument 
and
the return value of the same DO method.

All that should be pretty much ok.