[parisc-linux] missing barrier in _raw_spin_lock?

[John David Anglin]

> > #define __ldcw(a) ({ \
> >   unsigned int __ret;                                                   \
> >   __asm__ __volatile__("ldcw 0(%2),%0"                                  \
> > 		      : "=r" (__ret), "=m" (*(a)) : "r" (a));           \
> >   __ret;                                                                \
> > })

> I suppose the memory operand specification is required here. Newer 
> compilers (especially gcc-3.4) can optimize away local variables if you 
> only access the address but not the contents. I think even your 
> pthreads version is not really correct, because it specifies (*(a)) as
> output only instead of inout.

I agree.

> No, putting the barrier into __ldcw is wrong because it would impact all
> other uses of __ldcw that don't need the barrier. AFAICS, the 

The only other uses for __ldcw are the SPIN_LOCK macro in atomic.h
and the _raw_spin_trylock in spinlock.h (i.e., it is only used
acquire locks).  If all these need barriers, then it might as well
be in __ldcw.

Dave
-- 
J. David Anglin                                  dave.anglin at nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)