Re: libatomic-ops-7.4.0 not compatible with ppc64le architecture

On Fri, Dec 20, 2013 at 10:48:14AM +0530, Deepak C Shetty wrote:
> On 12/19/2013 03:27 AM, Lennart Sorensen wrote:
> >On Wed, Dec 18, 2013 at 04:30:54PM +0530, Madhu Pavan wrote:
> >>Hi Ivan Maidanski,
> >>
> >>We are in the process of adding support for new arch (ppc64le) and
> >>as part of that we wanted to send patches for config.guess and
> >>configure as appropriate.
> >>
> >>libatomic-ops-7.4.0 could not be built on ppc64le because it had an
> >>outdated configure/config.guess/libtool.m4. Since your pkg uses
> >>libtool and the latest tarball of your pkg didn't build on ppc64le
> >>due to outdated libtool present on your system. We wanted to request
> >>you to upgrade your libtool to the alpha release of libtool which
> >>has all the bits and pieces needed for ppc64le support.
> >>
> >>In that context, I wanted to confirm the below
> >>
> >>1) Should we send patches to 'gc-V9/[email protected]' mailing list or
> >>there is some other mailing list ?
> >>
> >>2) New libtool alpha release which supports ppc64le arch is
> >>available, you may  pick up the alpha release from either of the
> >>below :
> >>
> >>ftp://alpha.gnu.org/gnu/libtool/libtool-2.4.2.418.tar.gz (1.6MB)
> >>
> >>ftp://alpha.gnu.org/gnu/libtool/libtool-2.4.2.418.tar.xz (920KB)
> >>
> >>It would be great if you can install this new libtool which supports
> >>ppc64le, and spin the pkg and let us know the new tar availability.
> >>We can test it on ppc64le architecture for you. Please let me know
> >>if you are planning to release or spin newer version.
> >>
> >>Currently I used
> >>
> >>http://www.hpl.hp.com/research/linux/atomic_ops/download.php4
> >>
> >>Which does not have ppc64le support.
> >>
> >>Your response is highly appreciated.
> >Well in my experience libatomic-ops doesn't work at all on powerpc,
> >at least if run on a power7.  The testsuite fails.  It seems to somehow
> >get by on power6 and older.  Many things have had bugs exposed by power7
> >since it seems the power7 really relies on the memory model rules of
> >the power architecture in ways previous designs did not.  Use of memory
> >barriers have to be done correctly to work on the power7, otherwise
> >different CPUs will often see different values in memory due to caching.
> 
> Hi Lennart,
>     Thanks for letting us know. We were not aware of it.
> 
> Would you know if someone from community/IBM is looking/working on
> this and/or there is any effort to get this working on Power7
> systems ?

I am not aware of anyone working on it.  I looked at it, but it is beyond
my skill set at this time.  I think given the complexity of doing atomic
ops safely on a multi cpu system, it might need the help of someone at
IBM that really knows the architecture.

I just know about it because we have a power6 and a power7 IBM system
we run Debian on as a build system, and pixman failed the testsuite on
the power7, but passed on the power6.

I have been tempted to "fix" the problem by just putting locks in place
and not caring that it completely defeats the purpose of the library.
At least it would work.

-- 
Len Sorensen