Jiri B wrote:
> any news?
>
> I sent requested file
>
> http://lists.gnu.org/archive/html/bug-gnulib/2012-01/msg00258.html
Thanks for the reminder. The allocator.i file reveals that the fd_set
type is being used in /usr/include/unistd.h:255 but it is only defined
later, in /usr/include/sys/select.h.
In the meantime, I got access to an OpenBSD 5.0 machine. I could reproduce
the problem with a gnulib testdir for the modules
unistd stdlib pthread signal sys_select allocator mkstemp pthread_sigmask
Analyzing the chain of includes
In file included from ./unistd.h:44,
from ./stdlib.h:90,
from ./pthread.h:38,
from ./signal.h:65,
from ./sys/select.h:58,
from /usr/include/sys/types.h:224,
from ./sys/types.h:28,
from /usr/include/stdlib.h:41,
from ./stdlib.h:35,
from allocator.c:4:
/usr/include/unistd.h:255: error: expected declaration specifiers or '...'
before 'fd_set'
The include from stdlib.h to unistd.h can hardly be avoided:
#if (@GNULIB_MKSTEMP@ || @GNULIB_MKSTEMPS@ || @GNULIB_GETSUBOPT@ || defined
GNULIB_POSIXCHECK) && ! defined __GLIBC__ && !((defined _WIN32 || defined
__WIN32__) && ! defined __CYGWIN__)
/* On MacOS X 10.3, only <unistd.h> declares mkstemp. */
/* On MacOS X 10.5, only <unistd.h> declares mkstemps. */
/* On Cygwin 1.7.1, only <unistd.h> declares getsubopt. */
/* But avoid namespace pollution on glibc systems and native Windows. */
# include <unistd.h>
#endif
Also the include from pthread.h to stdlib.h can hardly be avoided
(or at least, I'm not included to touch pthread.h right now).
Also the include from signal.h to pthread.h can hardly be avoided:
/* MacOS X 10.3, FreeBSD 6.4, OpenBSD 3.8, OSF/1 4.0, Solaris 2.6 declare
pthread_sigmask in <pthread.h>, not in <signal.h>.
But avoid namespace pollution on glibc systems.*/
#if (@GNULIB_PTHREAD_SIGMASK@ || defined GNULIB_POSIXCHECK) \
&& ((defined __APPLE__ && defined __MACH__) || defined __FreeBSD__ ||
defined __OpenBSD__ || defined __osf__ || defined __sun) \
&& ! defined __GLIBC__
# include <pthread.h>
#endif
But the include from sys/select.h to signal.h seems like overkill,
especially if invoked from sys/types.h. I'm applying this fix.
2012-02-04 Bruno Haible <br...@clisp.org>
sys_select: Avoid syntax error on OpenBSD 5.0.
* lib/sys_select.in.h: Include <signal.h> only after the include_next
<sys/select.h>, not before.
Reported by Jiri B <ji...@devio.us>.
--- lib/sys_select.in.h.orig Sat Feb 4 14:38:47 2012
+++ lib/sys_select.in.h Sat Feb 4 14:38:38 2012
@@ -51,12 +51,6 @@
in <signal.h> where it belongs. */
#include <sys/types.h>
-/* Get definition of 'sigset_t'.
- But avoid namespace pollution on glibc systems. */
-#if !(defined __GLIBC__ && !defined __UCLIBC__)
-# include <signal.h>
-#endif
-
#if @HAVE_SYS_SELECT_H@
/* On OSF/1 4.0, <sys/select.h> provides only a forward declaration
@@ -81,6 +75,14 @@
#endif
+/* Get definition of 'sigset_t'.
+ But avoid namespace pollution on glibc systems.
+ Do this after the include_next (for the sake of OpenBSD 5.0) but before
+ the split double-inclusion guard (for the sake of Solaris). */
+#if !(defined __GLIBC__ && !defined __UCLIBC__)
+# include <signal.h>
+#endif
+
#ifndef _@GUARD_PREFIX@_SYS_SELECT_H
#define _@GUARD_PREFIX@_SYS_SELECT_H
