Subj : Re: sync vs. eieio on powerpc
To   : comp.programming.threads
From : Alexander Terekhov
Date : Tue Mar 01 2005 03:15 pm


Joe Seigh wrote:
> 
> On Tue, 01 Mar 2005 14:37:09 +0100, Alexander Terekhov <terekhov@web.de> wrote:
> 
> >
> > Joe Seigh wrote:
> > [...]
> >> > Nope. You can have ordinary loads in between LR and SC+isync. They
> >> > are unordered (data and control dependencies aside for a moment). And
> >> > think of moving isync above SC (which is a store, not a load).
> >> >
> >>
> >> That's what I said.
> >
> > No. SC+isync != isync+SC, so to speak.
> >
> 
> isync is not a store/store when used with store conditional, e.g.
> if one thread does
> 
>     stwcx
>     isync
>     stw

In order to have msync effect, isync needs preceding conditional 
branch. In the case of lock acquisition, that conditional branch is 
related to store, not load. So it IS store/* thing.

[...]
> It is a really bad idea to define isync as an "acquire" barrier
> even though it may appear to work as one in certain situations.

I agree. They should have provided straightforward SC with a whole 
bunch of associated combinations (if any) of unidirectional hoist 
and sink constraints affecting (only in the case of successful SC 
of course) subsequent and/or preceding (in the program order) 
loads and/or stores respectively.

regards,
alexander.

.