Hi,
we upgraded from mpich 3.1.2 to 3.2 and our nightly builds under valgrind show a new valgrind warning, which I suppose originates from the mpi library, rather than our code (given the stack trace below and the fact that they are absent with 3.1):
There are two slightly distinct traces (alltoall and alltoallv):
==30165== Conditional jump or move depends on uninitialised value(s)
==30165== at 0x6E00056: MPIR_Process_status (mpiimpl.h:4394)
==30165== by 0x6E00056: MPIC_Waitall (helper_fns.c:774)
==30165== by 0x6D5A853: MPIR_Alltoall_intra (alltoall.c:369)
==30165== by 0x6D5B63C: MPIR_Alltoall (alltoall.c:564)
==30165== by 0x6D5B6EF: MPIR_Alltoall_impl (alltoall.c:599)
==30165== by 0x6D5BF07: PMPI_Alltoall (alltoall.c:722)
==30165== by 0x6AFF0DC: PMPI_ALLTOALL (alltoallf.c:276)
==30165== by 0x1D04699: __message_passing_MOD_mp_alltoall_i22
==30201== Conditional jump or move depends on uninitialised value(s)
==30201== at 0x6E00056: MPIR_Process_status (mpiimpl.h:4394)
==30201== by 0x6E00056: MPIC_Waitall (helper_fns.c:774)
==30201== by 0x6D5C4EA: MPIR_Alltoallv_intra (alltoallv.c:194)
==30201== by 0x6D5CEC9: MPIR_Alltoallv (alltoallv.c:339)
==30201== by 0x6D5CFAA: MPIR_Alltoallv_impl (alltoallv.c:376)
==30201== by 0x6D5D8CA: PMPI_Alltoallv (alltoallv.c:527)
==30201== by 0x6AFF13E: PMPI_ALLTOALLV (alltoallvf.c:276)
==30201== by 0x1CCE180: __message_passing_MOD_mp_alltoall_z22v (message_passing_z.f90:223)
Is this issue already known/fixed ?
I saw there was a somewhat related change (which is part of 3.2):
https://trac.mpich.org/projects/mpich/changeset/be7c17b3177ccb0ede49761b44a438952360b061
but maybe that change was not fully complete ?
Best regards,
Joost VandeVondele
Prof. Dr. Joost VandeVondele
Nanoscale Simulations
ETH Zurich
HIT G 41.4
Wolfgang-Pauli-Strasse 27
8093 Zurich, Switzerland
Joost.VandeVondele@mat.ethz.ch
http://www.nanosim.mat.ethz.ch/
Tel.: +41 (0)44 6338451