MPI_Status_f082f not part of the mpi_f08 interface

[aradi]

MPI_Status_f082f not part of the mpi_f08 interface

[hppritcha]

i cannot reproduce this. I'm using gfortran 8.5.0. Which fortran compiler are you using?

[aradi]

mpifort --version
GNU Fortran (GCC) 13.2.0
Copyright (C) 2023 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Oh, that is interesting. Do you really mean, gfortran 8.5.0? Just, because I am somewhat unsure, whether gfortran 8.5 implements the Fortran 2008 constructs type(*) and dimension(..), which are in theory necessary for the mpi_f08 interface. (But I could be wrong about this.)

[bosilca]

I can't replicate either

mpifort --version
GNU Fortran (GCC) 11.4.1 20230605 (Red Hat 11.4.1-2)

Can you find your OMPI libraries and send back the output of nm libmpi_usempif08.so | grep -i MPI_Status_

[hppritcha]

unfortunately Open MPI currently does not know about assumed rank fortran option. we're still at assumed shape. hopefully this will be fixed soon.

[aradi]

@bosilca Sure, I obtain following output

nm libmpi_usempif08.so | grep -i MPI_Status_
00000000000233db T mpi_status_f082f_f08_
0000000000023419 T mpi_status_f2f08_f08_
0000000000023457 T mpi_status_set_cancelled_f08_
0000000000023495 T mpi_status_set_elements_f08_
00000000000234d8 T mpi_status_set_elements_x_f08_
                 U ompi_status_f082f_f
                 U ompi_status_f2f08_f
                 U ompi_status_set_elements_f
                 U ompi_status_set_elements_x_f
0000000000030533 T pmpi_status_f082f_f08_
0000000000030571 T pmpi_status_f2f08_f08_
                 U pmpi_status_set_cancelled_
00000000000305af T pmpi_status_set_cancelled_f08_
00000000000305ed T pmpi_status_set_elements_f08_
0000000000030630 T pmpi_status_set_elements_x_f08_

But, please note, that the problem is not, that the routine is not part of the library; the problem is, that the routine is not part of the mpi_f08 Fortran module. I get compilation error, not a linking error. If I replace

use mpi_f08, only : ..., MPI_Status_f082f, ...

with

use mpi_f08

the code compiles (but for the wrong reason):

• In the former case, the compiler explicitly checks the mpi_f08.mod module file (generated during the build of OpenMPI) whether it contains information about the signature of the MPI_Status_f082f() routine. And it fails, because it is not found.
• In the latter case, the module file is also checked and the signature is not found either, but the compiler is allowed then to assume, that MPI_Status_f082f() is an arbitrary external routine with unknown signature. The compiler builds the program then, but no checking of argument number or types is done, you just hope for the best. You can test it yourself: use use mpi_f08 without the only clause and remove one of the arguments in the call to MPI_Status_f08. The code will still happily compile (and then very likely crash on execution).

[hppritcha]

another thing to check is the mpi_f08.mod. Its actually just gzipped text in the case of the gnu compiler.
When I gunzip the module file generated by gfortran 8.5.0 i see an entry of this form

4952 'mpi_status_f082f' 'mpi_f08_interfaces' '' 1 ((PROCEDURE
UNKNOWN-INTENT UNKNOWN-PROC UNKNOWN UNKNOWN 0 0 SUBROUTINE GENERIC
ARRAY_OUTER_DEPENDENCY) () (UNKNOWN 0 0 0 0 UNKNOWN ()) 0 0 () () 0 () ()
() 0 0)

[aradi]

@hppritcha That's a good point. I've checked it and

zcat mpi_f08.mod | grep -i mpi_status_f082f

is empty.
Running

for m in *.mod; do echo "*** $m  ***"; zcat $m | grep -i mpi_status_f082f; done

returns

*** mpi.mod  ***
4135 'mpi_status_f082f' 'mpi' '' 1 ((PROCEDURE UNKNOWN-INTENT
8139 'pmpi_status_f082f' 'mpi' '' 1 ((PROCEDURE UNKNOWN-INTENT
'mpi_status' 0 4134 'mpi_status_f082f' 0 4135 'mpi_status_f2f08' 0 4140
0 8120 'pmpi_start' 0 8130 'pmpi_startall' 0 8134 'pmpi_status_f082f' 0
*** mpi_ext.mod  ***
4153 'mpi_status_f082f' 'mpi' '' 1 ((PROCEDURE UNKNOWN-INTENT
8201 'pmpi_status_f082f' 'mpi' '' 1 ((PROCEDURE UNKNOWN-INTENT
'mpi_status' 0 4152 'mpi_status_f082f' 0 4153 'mpi_status_f2f08' 0 4158
0 8182 'pmpi_start' 0 8192 'pmpi_startall' 0 8196 'pmpi_status_f082f' 0
*** mpi_f08.mod  ***
*** mpi_f08_callbacks.mod  ***
*** mpi_f08_ext.mod  ***
*** mpi_f08_interfaces.mod  ***
*** mpi_f08_interfaces_callbacks.mod  ***
*** mpi_f08_types.mod  ***
*** mpi_types.mod  ***
*** pmpi_f08_interfaces.mod  ***

so it is apparently there in mpi.mod and mpi_ext.mod.

[aradi]

Just one more info, I am using the --enable-mpi1-compatibility configure option. Could that make a difference?

[bosilca]

MPI1 compatibility should not remove this API, it only adds back some deprecated API. However, the inclusion of the status_f08 accessors is driven by OMPI_FORTRAN_HAVE_TYPE_MPI_STATUS, which is detected during configure by the macro OMPI_FORTRAN_CHECK_BIND_C_TYPE_NAME. This macro tries to detect if the compiler supports TYPE(...), BIND(C, NAME=...), and apparently your compiler fails this test. Can you please check in your config.log for 'if Fortran compiler supports` to see why the test failed.

[hppritcha]

uh oh. i forgot i was testing on our main branch. there seems to be a problem with 5.0.3. checking if its fixed in the 5.0.x branch. i can reproduce this issue using the gcc 13.2.0 and ompi 5.0.3

[hppritcha]

this seems to not be fixed on 5.0.x branch

[hppritcha]

looks like 83eb116 was not cherry picked back to v5.0.x