Another "ifort -i8" problem

[ompiteam]

From http://www.open-mpi.org/community/lists/users/2012/07/19732.php:

There is another problem using -fdefault-integer-8. I am using 1.6..

For the i8:

configure:44650: checking for the value of MPI_STATUS_SIZE
configure:44674: result: 3 Fortran INTEGERs
configure:44866: checking if Fortran compiler works
configure:44895: /usr/bin/gfortran44 -o conftest -m64  -g -fPIC -fdefault-integer-8  -fexceptions  conftest.f  >&5
configure:44895: $? = 0

MPI_STATUS_SIZE 3

For the i4:

configure:44650: checking for the value of MPI_STATUS_SIZE
configure:44674: result: 6 Fortran INTEGERs
configure:44866: checking if Fortran compiler works
configure:44895: gfortran -o conftest -m64 -g -fPIC  -fexceptions  conftest.f  >&5

MPI_STATUS_SIZE 6

For i8, I found out later in ompi/mpi/c/status_c2f.c:

#!c
for( i = 0; i < (int)(sizeof(MPI_Status) / sizeof(int)); i++ )
    &n! bsp;   f_status[i] = OMPI_INT_2_FINT(c_ints[i]);

This needs to be investigated.

[ompiteam]

Imported from trac issue 3163. Created by jsquyres on 2012-07-05T13:23:55, last modified: 2013-04-24T22:41:01

[ompiteam]

Trac comment by jsquyres on 2013-04-24 22:41:01:

This issue came up again on the mailing list in Apr 2013: http://www.open-mpi.org/community/lists/users/2013/04/21730.php.

I confirm the reports: when I compile OMPI with ifort -i8, MPI_STATUS_SIZE is 3. If I don't compile with -i8, MPI_STATUS_SIZE is 6.

This happens in both the v1.6 branch, the trunk, and the v1.7 branch.

It is this way because that's 3*8 bytes, which is 24 bytes, which is the same as the size of the C status (4 int's and a size_t).

But that obviously creates problems when you try to access individual status members from Fortran; e.g., accessing an -i8-compiled status(MPI_TAG) will access the wrong data (i.e., the data at byte offset MPI_TAG=2 -- 7 through 16), rather than where the C-based data is located (byte offset 4-7).

As such, when compiled with -i8, I guess MPI_STATUS_SIZE should be 3 + ceil((sizeof(int) + sizeof(size_t) / 8), which is likely 2, so MPI_STATUS_SIZE should be 5.

But as noted in this bug, I think there's also shared culpability with MPI_Status_f2c and MPI_Status_c2f, because they don't copy to/from C-sized integers to Fortran-sized integers.

(note, too, that there's another -i8 problem in https://svn.open-mpi.org/trac/ompi/ticket/3157)