Subj : Re: Memory visibility and MS Interlocked instructions
To   : comp.programming.threads
From : Peter Dimov
Date : Thu Sep 01 2005 11:47 am

Sean Kelly wrote:
> Alexander Terekhov wrote:
> > Sean Kelly wrote:
> > [...]
> > >                                                            msync.acq
> > > and msync.rel operations (to use atomic<> semantics) would both need
> > > the LOCK prefix.  Is this correct?
> >
> > No.
> >
> > >                                    I'll admit that until this I was
> > > beginning to think that only msync.acq stores needed a LOCK.
> >
> > That's not true either.
>
> Help me out here :)  It's been established that loads can be migrated
> above preceding stores if no membars are present, so LOCK must be used
> in some cases to prevent this, correct?

The unusual load with msync_rel would need a LOCK or an MFENCE, true,
as would a store with msync_acq.

But the common load.acq and store.rel operations map to MOV, no LOCK or
MFENCE. A load.acq is allowed to migrate above a store.rel.

.