[cpp-threads] Failed compare-and-swap

[Paul E. McKenney]

On Fri, Aug 03, 2007 at 12:07:08AM -0000, Boehm, Hans wrote:
> I'm confused.  I thought we had success and failure orders, not read and
> write orders.  If you go with read/write, you can't drop the acquire
> constraint on failure, which is probably the most (only?) interesting
> case in practice, right?

Color me confused.  :-/  I let the argument names of "r" and "w" distract
me from the true semantics of the arguments.

> If we have success and failure orders, it seems more natural to me to
> list the success case first, which also seems optimal in terms of
> overloads.

I agree.  The success-order argument should precede the failure-order
argument.

						Thanx, Paul

> Hans
> 
> > -----Original Message-----
> > From: cpp-threads-bounces at decadentplace.org.uk 
> > [mailto:cpp-threads-bounces at decadentplace.org.uk] On Behalf 
> > Of Paul E. McKenney
> > Sent: Thursday, August 02, 2007 4:55 PM
> > To: C++ threads standardisation
> > Subject: Re: [cpp-threads] Failed compare-and-swap
> > 
> > On Thu, Aug 02, 2007 at 03:36:35PM -0700, Lawrence Crowl wrote:
> > > On 8/2/07, Peter Dimov <pdimov at mmltd.net> wrote:
> > > > Lawrence Crowl wrote:
> > > > > On 8/2/07, Boehm, Hans <hans.boehm at hp.com> wrote:
> > > >
> > > > >> Does it make sense to use an additional overload instead of a 
> > > > >> default value, so that failure_order can effectively 
> > default to 
> > > > >> success_order? I think that's clearly the correct default for
> > > > >>
> > > > >> seq_cst, relaxed, and release
> > > > >>
> > > > >> since release becomes vacuous in the failure case, 
> > since there is 
> > > > >> no store.
> > > > >
> > > > > That could work, but we'd need three overloads.
> > > >
> > > > Two are enough for that default.
> > > >
> > > > compare_swap( int& e, int d, memory_order w, memory_order r );
> > > >
> > > > compare_swap( int& e, int d, memory_order w = 
> > memory_order_seq_cst ) 
> > > > {
> > > >     compare_swap( e, d, w, w );
> > > > }
> > > 
> > > Ah, yes, if you switch the order of parameters.  Hans has 
> > already said 
> > > he likes the switched order.  Any other opinions?
> > 
> > Given that the read happens before the write (in principle, 
> > anyway), I believe that we would come to deeply regret 
> > putting the write-side memory-order argument before the 
> > read-side memory-order argument.
> > 
> > An extra overload seems to me to be a small price to pay in this case.
> > 
> > > > The first overload can be made to strip release from r.
> > > 
> > > I think that if the parameter is explicit, it should be 
> > exactly right.
> > > We currently require programmers to use the right form, and 
> > I'd like 
> > > to leave it that way.
> > 
> > Since many programs are read much more frequently than they 
> > are written, I agree with requiring the correct form.
> > 
> > 							Thanx, Paul
> > 
> > --
> > cpp-threads mailing list
> > cpp-threads at decadentplace.org.uk
> > http://www.decadentplace.org.uk/cgi-bin/mailman/listinfo/cpp-threads
> >