• mpich FTBFS on sh4: MPIUI_Thread TLS definition mismatches tbss

    From John Paul Adrian Glaubitz@21:1/5 to All on Tue Jun 6 20:30:02 2017
    On a side note: Please avoid posting to debian-ports@ as this will cross-post to the mailing lists of *all* ports architectures. Rather use debian-ports-devel@
    if necessary.

    Thanks,
    Adrian

    --
    .''`. John Paul Adrian Glaubitz
    : :' : Debian Developer - glaubitz@debian.org
    `. `' Freie Universitaet Berlin - glaubitz@physik.fu-berlin.de
    `- GPG: 62FF 8A75 84E0 2956 9546 0006 7426 3B37 F5B5 F913

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Drew Parsons@21:1/5 to John Paul Adrian Glaubitz on Wed Jun 7 07:40:01 2017
    Sorry Adrian, I thought debian-ports was a single mailing list!

    Drew


    On Tue, 2017-06-06 at 19:25 +0200, John Paul Adrian Glaubitz wrote:
    On a side note: Please avoid posting to debian-ports@ as this will
    cross-post
    to the mailing lists of *all* ports architectures. Rather use debian- ports-devel@
    if necessary.

    Thanks,
    Adrian


    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Drew Parsons@21:1/5 to Aaron M. Ucko on Wed Jun 7 08:00:01 2017
    On Tue, 2017-06-06 at 11:58 -0400, Aaron M. Ucko wrote:
    Drew Parsons <dparsons@debian.org> writes:

    The link error suggests 2 different versions of MPIUI_Thread are
    used. 
    But I can only see a definition in src/util/thread/mpiu_thread.c. src/mpi/comm does not refer to it, and comm_rank.c does not use any
    MPIU object. So the error message doesn't make sense to me.

    comm_rank's reference to MPIUI_Thread is also present on at least
    amd64,
    and presumably due to some macro that (directly or indirectly) calls MPIU_THREADPRIV_FIELD.  However, it's not clear why multithreaded
    build
    settings would be in effect for mpiu_thread but not also
    comm_rank.  (It
    looks like MPIUI_Thread's thread-locality is conditional on MPICH_IS_THREADED, which mpichconf.h defines centrally.)

    I don't have time to dig deeper, but hope that brief analysis helps.


    That sounds about right, thanks Aaron.

    Is the problem likely to be in the toolchain or in the mpich code?  

    Given that no other arch has the problem, the toolchain seems likely.
    Unless sh4 does something special in regards to thread support that the
    code needs to account for.

    Drew

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)