Messages in this thread | | | Date | Mon, 25 Jun 2018 11:06:48 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH 2/2] tools/memory-model: Add write ordering by release-acquire and by locks |
| |
On Mon, Jun 25, 2018 at 10:29:23AM +0200, Andrea Parri wrote: > On Mon, Jun 25, 2018 at 09:32:29AM +0200, Peter Zijlstra wrote: > > > > I have yet to digest the rest of the discussion, however: > > > > On Fri, Jun 22, 2018 at 02:09:04PM -0400, Alan Stern wrote: > > > The LKMM uses the same CAT code for acquire/release and lock/unlock. > > > (In essence, it considers a lock to be an acquire and an unlock to be a > > > release; everything else follows from that.) Treating one differently > > > from the other in these tests would require some significant changes. > > > It wouldn't be easy. > > > > That is problematic, acquire+release are very much simpler operations > > than lock+unlock. > > > > At the very least, lock includes a control-dependency, where acquire > > does not. > > I don't see how this is relevant here; roughly, "if something is guaranteed > by a control-dependency, that is also guaranteed by an acquire". Right? ;)
Right, you are, clearly I needs me a wake up drink :-).. So lock does very fundamentally involve a RmW, and it has the whole wait-until loop thing in. But yes, now I'm strugging to better express how it's different from a memory ordering pov.
But still, the lock case will/must disallow the re-ordering (since we rely on it), whereas the pure acquire/release seems to be struggling.
Personally I prefer a stronger model over a weaker one (as does Linus IIRC) but clearly people have different opinions on that.
| |