[cpp-threads] Failed compare-and-swap

[Peter Dimov]

Paul E. McKenney wrote:

> o release: seems like no memory fences needed in failure case, at
> least at first glance. If so, this might be what people who
> want cheap CAS failure should use.

This makes no sense to me; how can you choose to use release instead of 
acquire just because the failures are cheaper? You use acquire because you 
need acquire. If you don't care about the ordering, you can just use 
relaxed...