[mpich-discuss] MPICH 3.1.4 segfaults Clang 3.6.1 and 3.7

Torbjörn Klatt opensource at torbjoern-klatt.de
Thu Sep 3 23:42:32 CDT 2015


Hello,

thanks Rob and Pavan for the reply.

Yes, it's on Linux (openSUSE 13.2 64bit to be precise, Clang compiled with 
distro's default GCC 4.8.3).

After testing Clang 3.4.2 and seeing it to fail with the same segfault, I 
assume users how want to use MPICH with Clang are stuck with Clang 3.3. You 
may think about writing that down somewhere for future users.

For now I'll keep using Clang 3.3 and GCC 5.x for my MPI development but keep 
on testing future Clang and MPICH versions.

Torbjörn


On Thursday 03 September 2015 01:11:54 Balaji, Pavan wrote:
> This is only an issue with clang on Linux, right?  (I couldn't quite tell
> from the email chain).  On Mac OS/X, we already disable weak symbols, so
> this shouldn't occur.
 
> Can someone remind me why we didn't simply add a check in our configure to
> say if compiler == clang, then disable weak symbols?
 
>   -- Pavan
> 
> 
> 
> 
> 
> On 9/2/15, 4:57 PM, "Rob Latham" <robl at mcs.anl.gov> wrote:
> 
> 
> >
> >
> >On 09/02/2015 02:29 AM, Torbjörn Klatt wrote:
> >
> >> Hello,
> >>
> >>
> >>
> >> I'm trying to use MPICH 3.1.4 with Clang and tried Clang 3.6.1 and 3.7.
> >> With
 both versions Clang segfaults at the same point:
> >>
> >>
> >>
> >> pt2pt/greq_start.c:239:62: error: alias must point to a defined variable
> >> or
 function
> >> 
> >>          MPIX_Grequest_class *greq_class)
> >> 
> >> __attribute__((weak,alias("MPIX_Grequest_class_create")));
> >>
> >>
> >>
> >> I've filed a LLVM bug [1] with more details, however without response
> >> yet.
> >>
> >>
> >>
> >>
> >> MPICH was configured with:
> >>
> >>
> >>
> >> CC=clang CXX=clang++ /opt/toolchains/src/mpi/mpich/3.1.4/configure --
> >> srcdir=/opt/toolchains/src/mpi/mpich/3.1.4 --
> >> prefix=/opt/toolchains/clang/3.7/libraries/mpi/mpich/3.1.4
> >> --enable-g=most --
 enable-fast=all --enable-fortran=all --enable-cxx
> >>
> >>
> >>
> >>
> >> Is this a known MPICH problem or am I doing something wrong?
> >
> >
> >It's been a long-standing annoyance on our end as well:
> >
> >We sort of fixed this two+ years ago:
> >http://trac.mpich.org/projects/mpich/ticket/1815
> >
> >then something happened about 6-ish months ago:
> >http://trac.mpich.org/projects/mpich/ticket/2236
> >
> >My ubuntu bug report appears to have garnered exactly zero attention:
> >https://bugs.launchpad.net/ubuntu/+source/llvm-toolchain-snapshot/+bug/1439
> >865
 
> >As it's a clang compiler error, it's hard to know what the workaround 
> >would be.  Thanks for trying 3.6.1 and 3.7.    At least we can rule out 
> >"fixed in newer versions"
> >
> >You will find this same segfault in mpich-master: we haven't done 
> >anything on our end to work around it, unfortunately.
> >
> >==rob
> >
-- 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20150904/e787d73f/attachment.sig>
-------------- next part --------------
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss


More information about the discuss mailing list