Re: [gpsd-dev] gpsd fails to build on mac 10.7

[Eric S. Raymond]

Greg Troxel <address@hidden>:
> So it seems the immediate fix is that barrier() should just be defined
> as a no-op if there isn't compiler-specific fix.

Current code could certainly stand to be cleaned up a little:

#ifndef S_SPLINT_S
#if defined(NTPSHM_ENABLE) || defined(SHM_EXPORT_ENABLE)
/* memory barriers */
static /address@hidden@*/ inline void barrier(void) {
#if (__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 7)
       __atomic_thread_fence(__ATOMIC_SEQ_CST);
#elif (__GNUC__ == 4 && __GNUC_MINOR__ >= 4)
       __sync_synchronize();
#else
# error Memory barriers are not supported. Disable NTPSHM and SHM export.
#endif /* defined(NTPSHM_ENABLE) || defined(SHM_EXPORT_ENABLE) */
#endif /* defined(NTPSHM_ENABLE) || defined(SHM_EXPORT_ENABLE) */
#else
#define barrier()
#endif /* S_SPLINT_S */

Maybe this would be better?

/* memory barriers */
static /address@hidden@*/ inline void barrier(void)
{
#ifndef S_SPLINT_S
#if (__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 7)
       __atomic_thread_fence(__ATOMIC_SEQ_CST);
#elif (__GNUC__ == 4 && __GNUC_MINOR__ >= 4)
       __sync_synchronize();
#endif
#endif /* S_SPLINT_S */
}

> Has anyone actually seen a bug due to missing barriers?  (I get the
> theory.)

I haven't.  But I don't view that as determinative; they could show up
as data corruption subtle enough to be hard to detect.

> What is one supposed to do while complying with C99 and POSIX?  Relying
> on gcc extensions is very 00s.

I know, but I think it's better to quietly use these if present than
to ignore them.
-- 
                <a href="http://www.catb.org/~esr/";>Eric S. Raymond</a>

signature.asc
Description: Digital signature