`sentinel' gcc-3.x/OpenBSD compat. [Re: [patch 1/3] Make obconcat use stdarg]

[Jan Kratochvil <jan.kratochvil@redhat.com>]

On Fri, 30 Apr 2010 21:08:48 +0200, Mark Kettenis wrote:
> > From: Jan Kratochvil <jan.kratochvil@redhat.com>
> > 
> > __attribute__ ((sentinel)) availability for gcc >= 4.0 I have copied from
> > <glib-2.0/glib/gmacros.h>.  It roughly matches the GCC ChangeLog dates.
> 
> The OpenBSD system compile, which is based on GCC 3.3.5, already has
> the sentinel attribute.

FSF GCC 3.3.5 does not support this attribute:
	void f (int i, ...) __attribute__ ((__sentinel__));
	sentinel.c:1: warning: `__sentinel__' attribute directive ignored
	(+ there are no traces of any `sentinel' in the gcc-3.3.5 sources)

OpenBSD-4.6-amd64 from md5sum
	28edaa65b16bd5381ebf645d27039a94  openbsd-install46.iso
does not boot for me on Fedora 12 x86_64 KVM using
	kernel-2.6.32.11-99.fc12.x86_64
	qemu-system-x86-0.11.0-13.fc12.x86_64
as it hangs on "setting tty flags"; I have not found any OpenBSD Live image.

Moreover I do not consider relevant to support this attribute on so obsolete
compiler.  Not using the attribute has no effect on users building GDB.
This attribute is useful only for the GDB developers.


> I'm afraid you'll need an explicit cast for the NULLs used as the
> sentinel value, otherwise platforms that
> 
>   #define NULL 0L
> 
> will generate warnings like:
> 
>   sentinel.c: In function 'foo':
>   sentinel.c:8: warning: missing sentinel in function call
> 
> and with -Werror, that's not a good :(.

Which platforms?  Could you provide a login there?  As NULL is commonly
defined using (such as in <dbus-1.0/dbus/dbus-macros.h>):

#ifndef NULL
#  ifdef __cplusplus
#    define NULL        (0L)
#  else /* !__cplusplus */
#    define NULL        ((void*) 0)
#  endif /* !__cplusplus */
#endif

forcing NULL to be ((void *) 0) may make GDB C++ compatibility problematic
(I do not understand reasons for the conditional above, though).

OTOH a requirement of some special sentinel expression instead of simple `NULL'
for obconcat would nullify its stdarg convenience.

Therefore something could be placed to configure.ac for NULL redefinition
depending on the test on the specific build system.


Thanks,
Jan