Gabriele Svelto (gabriele.svelto.delete@this.gmail.com) on August 21, 2018 3:54 pm wrote:
>
> One big downside they had was that they could not be executed speculatively

The thing is, properly designed remote atomics for statistics have to be of the type where they are

(a) not ordered wrt other memory accesses (including locking!)

(b) do not return a value to the processor

and then the way you implement them is that they are "scheduled" (to whatever remote atomic engine) when the instruction retires. No serialization by the core what-so-ever, outside of just the queue size for the remote atomics or something.

Then you'd have a "wait for remote atomics queue to flush" instruction for any readers - assuming the reader even cares (which might not be the case). If you really care, you do cross-CPU calls and locking and everybody is forced to do the sync, but for statistics, that kind of caring is so rare that it might not even exist.

Of course, for people who want remote atomics for other things, the rules can be very different. If you want them for refcounting in a GC environment you might have very different synchronization needs indeed.

Linus