Subj : Re: std::msync
To   : comp.programming.threads
From : Alexander Terekhov
Date : Sat Apr 02 2005 05:07 pm

Peter Dimov wrote:
> 
> Alexander Terekhov wrote:
> > bool lock() throw() {
> >   return use_count_.increment_if_not_min(); // naked
> > }
> 
> msync::naked_competing, IIUC? 

Yep.

>                               Is cmpxchg on x86 a proper implementation of
> naked_competing, or does it need a lock prefix? 

Without a lock it is naked_MP_noncompeting so to speak. Without a 
lock it isn't atomic with respect to other processors (and probably 
HT threads, cores, etc.) at all AFAIK.


>                                                 And if lockless cmpxchg is
> fine,

It might work on a uniprocessor or then all potentially contending 
thread are bound to the same processor.

> 
> > void add_ref() throw() {
> >   use_count_.increment(); // naked
> > }
> 
> should this use cmpxchg too instead of lock inc/lock xadd?

Given that cmpxchg without lock is MP unsafe,  lock inc/lock xadd 
is probably beter.

regards,
alexander.

.