Compile issue with GCC 10.1.0

[ndkeen]

Wanted to try newer GNU compiler, but hit an issue here:

cime/src/share/timing/perf_utils.F90

  282 |    call MPI_BCAST(vec,lsize,MPI_INTEGER,0,comm,ierr)
      |                  1
......
  314 |    call MPI_BCAST(vec,lsize,MPI_LOGICAL,0,comm,ierr)
      |                  2
Error: Type mismatch between actual argument at (1) and actual argument at (2) (INTEGER(4)/LOGICAL(4)).

This is on cori. I created simple tester:

module boilerplate
  implicit none
  private
#include <mpif.h>
  save
  
  interface shr_mpi_bcast ; module procedure &
       shr_mpi_bcastl0, &
       shr_mpi_bcasti0
  end interface
  
contains
  
  SUBROUTINE shr_mpi_bcasti0(vec,comm)
    IMPLICIT none
    integer, intent(inout):: vec
    integer, intent(in)   :: comm
    integer               :: ierr,lsize
    lsize = 1
    call MPI_BCAST(vec,lsize,MPI_INTEGER,0,comm,ierr)
  END SUBROUTINE shr_mpi_bcasti0
  
  SUBROUTINE shr_mpi_bcastl0(vec,comm)
    IMPLICIT none
    logical, intent(inout):: vec
    integer, intent(in)   :: comm
    integer               :: ierr, lsize
    lsize = 1
    call MPI_BCAST(vec,lsize,MPI_LOGICAL,0,comm,ierr)
  END SUBROUTINE shr_mpi_bcastl0
end module boilerplate


program mpitest

#include <mpif.h>
  integer::rank,nprocs,ierr

  call MPI_INIT(ierr)
  call MPI_COMM_RANK(MPI_COMM_WORLD, rank, ierr)
  call MPI_COMM_SIZE(MPI_COMM_WORLD, nprocs, ierr)
  
  print*, "rank ", rank, " out of ", nprocs
  
  call MPI_FINALIZE(ierr)
end program mpitest

module swap PrgEnv-intel PrgEnv-gnu
module swap gcc gcc/10.1.0
And then build with ftn.

cori10% ftn module_interface.F90
module_interface.F90:7:6:

    7 |   save
      |      1
Warning: Legacy Extension: Blanket SAVE statement at (1) follows previous SAVE statement
module_interface.F90:22:19:

   22 |     call MPI_BCAST(vec,lsize,MPI_INTEGER,0,comm,ierr)
      |                   1
......
   31 |     call MPI_BCAST(vec,lsize,MPI_LOGICAL,0,comm,ierr)
      |                   2
Error: Type mismatch between actual argument at (1) and actual argument at (2) (INTEGER(4)/LOGICAL(4)).

[ambrad]

I googled a bit, and it seems adding the Fortran flag -fallow-argument-mismatch might fix this.

[whannah1]

@ndkeen I've used the argument suggested by @ambrad with good results. I've got E3SM building and running locally on my mac with GNU, which also required using the -fallow-invalid-boz flag to fix a similar issue.

[ndkeen]

Thanks Andrew! That does de-escalate that sort of error to a warning and let the build continue. (Though there are quite a few warnings that could be cleaned up).

I then ran into another issue with
integer, parameter :: gen_hash_key_offset = z'000053db'

Might need -fallow-invalid-boz to work-around this oddity. Just now reading Walter's comment...

[ndkeen]

I still get an error at link time when using gcc 10.1 on Cori. Even if I build with intel, but use the gcc/10.1.0 module, the build will not complete.

[ndkeen]

Noting the link error here, which isn't very helpful. But as I noted above, I still get this error even if I use Intel compiler, but still have gcc/10.1.0 module loaded.

[ 76%] Linking Fortran static library libatm.a

...

/usr/bin/ranlib libatm.a
make[2]: Leaving directory '/global/cscratch1/sd/ndk/e3sm_scratch/cori-haswell/m49-nov19/f-ne4.m49-nov19.gnu.mpt.n001p032t32x1.2d.gnu10.DEBUG/bld/cmake-bld'
[ 76%] Built target atm
make[1]: Leaving directory '/global/cscratch1/sd/ndk/e3sm_scratch/cori-haswell/m49-nov19/f-ne4.m49-nov19.gnu.mpt.n001p032t32x1.2d.gnu10.DEBUG/bld/cmake-bld'
gmake: *** [Makefile:87: all] Error 2

/global/cscratch1/sd/ndk/e3sm_scratch/cori-haswell/m49-nov19/f-ne4.m49-nov19.gnu.mpt.n001p032t32x1.2d.gnu10.DEBUG/bld/e3sm.bldlog.201210-132354

[rljacob]

This is a problem with GNU 10, right? It should allow both instances of MPI_bcast by default without any special flags.

[rljacob]

This discussion suggests its an MPI problem: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91731

[ndkeen]

I created a branch ndk/machinefiles/cori-test-gnu9 that allows for building with gnu9 or gnu10.
Using --compiler=gnu works as now in master (ie 8.3). But if you try --compiler=gnu9 or --compiler=gnu10, will build with those.

To complete the build with gnu10, will need a change to components/elm/src/external_models/sbetr/src/betr/betr_math/ODEMod.F90
that is described here:
https://www.gitmemory.com/issue/E3SM-Project/E3SM/4151/794906780

I'm not sure if there is a PR for that change yet. Yes there is: #4364

I was hoping we could use this branch to test and fix any issues with gnu9/gnu10 before then deciding how to proceed.
Note that the new machine Perlumutter only has gnu version 9 or 10.