Subj : Re: Adding thread support for C...
To   : comp.programming.threads
From : Chris Friesen
Date : Tue Aug 02 2005 11:29 pm

David Schwartz wrote:
> "Chris Friesen" <cfriesen@nortel.com> wrote

>>Really?  How would you make the lock/modify/unlock cycle faster than a 
>>single atomic increment?

>     Who says the CPU *has* an atomic increment? It may only have an atomic 
> compare-and-swap.

Okay, point.

>>In the optimal uncontended case each of the lock/unlock will be an atomic 
>>operation, so you end up with one extra atomic as well as whatever 
>>modification that you were trying to do.

>     Perhaps the CPU has heavily optimized atomic 'lock' and 'unlock' 
> operations and its atomic increment is implemented as some kind of 
> read/increment/compare-exchange/loop-on-fail thing.

Again, I admit it could be done this way.  Out of curiosity, do you know 
of any such architectures?

Chris

.