Re: [PATCH] Use CAS instead of atomic swap to implement spinlock

[Rich Felker <dalias@libc.org>]

On Sun, Apr 19, 2015 at 08:50:18AM +0300, Alexander Monakov wrote:
> On Sat, 18 Apr 2015, Rich Felker wrote:
> 
> > On Wed, Apr 15, 2015 at 01:44:53AM +0300, Alexander Monakov wrote:
> > > This should allow spinning without constantly dirtying cache lines holding the
> > > spinlock value.  On architectures without native atomic swap, musl implement
> > > a_swap by looping around a_cas.
> > > ---
> > > If I'm not mistaken this was also suggested by nsz on IRC.
> > > 
> > >  src/thread/pthread_spin_lock.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/src/thread/pthread_spin_lock.c b/src/thread/pthread_spin_lock.c
> > > index df575f0..dabcb31 100644
> > > --- a/src/thread/pthread_spin_lock.c
> > > +++ b/src/thread/pthread_spin_lock.c
> > > @@ -2,6 +2,6 @@
> > >  
> > >  int pthread_spin_lock(pthread_spinlock_t *s)
> > >  {
> > > -	while (a_swap(s, 1)) a_spin();
> > > +	while (a_cas(s, 0, 1)) a_spin();
> > >  	return 0;
> > >  }
> > 
> > Would it perhaps be better to do something like this?
> > 
> > while (*(volatile int *)s || a_cas(s, 0, 1)) a_spin();
> 
> I think so, Yes.  Is the cast required, or is it possible to change the
> pthread_spinlock_t typedef to 'volatile int'?

For C++ ABI purposes, I think switching to volatile int would be a
different type. :(

I wouldn't really be opposed to changing it for C and just having the
ABI-compat type used when __cplusplus is defined. We already do that
for pthread_t.

Rich