Re: sync operations: where's the barrier?

[Geoffrey Keating <geoffk at apple dot com>]

On 18/04/2005, at 6:13 PM, David Edelsohn wrote:

> > > > > > Geoffrey Keating writes:
>
> Geoff> The documentation for the atomic operation patterns says things 
> like:
>
> > > This pattern must issue any memory barrier instructions such that the
> > > pattern as a whole acts as a full barrier.
>
> Geoff> Should the barrier happen before the operation, after the 
> operation,
> Geoff> are there two barriers, or is it undefined?
>
> 	On PowerPC, this has a lot to do with the cooperation of the
> various functions referencing the memory atomically.  I am most 
> familiar
> with emitting sync (or lwsync) before the atomic operation.

That's what I'd planned to do, emit sync before the operation; but 
obviously it does matter, one way lets you use the intrinsic to acquire 
an object and the other way lets you use the intrinsic to release an 
object.

(Ideally it would be lwsync; actually, ideally there would be no sync 
at all, and users would be required to output one if they wanted one; 
but at present the routines are documented to have a 'sync'.  The 
question is where...)

Attachment: smime.p7s
Description: S/MIME cryptographic signature