use-mpi-f08: fix prototypes and more

[hppritcha]

In the course of checking out a user reported problem:

https://www.mail-archive.com/[email protected]//msg35382.html

it was discovered that many of the f08 entry points were improperly treating 'c' int args to the internal ompi*_f functions as fortran integers.

This commit fixes that and also addressed the problem reported by
the user. It was a special case where they were building the fortran
bindings with non-default integer size which happened to no longer match
the 'c' int.

The fact that this problem wasn't seen before just shows that probably 99% of the time the default Fortran integer kind is compatible with 'c' int.

[hppritcha]

For the record, the compiler flags with gnu to use to reproduce the user problem are

./configure  FFLAGS="-m64 -fdefault-integer-8" FCFLAGS="-m64 -fdefault-integer-8" CFLAGS=-m64 CXXFLAGS=-m64

[hppritcha]

i should point out that the fortran code generated by the python binding code in PR #12226 does this in a much more elegant manner but only for a few of the files touched in this PR that are impacted by big count. If time allows or I find an energetic intern, we could convert all of the remaining *.F90 files - even those not impacted by big count - to this approach and dispense with the coupling to the 'c' code in mpif-h.

Here's for instance how the code for MPI_Pack_external_size_f08 generated by the python code looks in that PR. Note no need to pass funny end of list of args 'c' ints:

subroutine MPI_Pack_external_size_f08(datarep,  &
                                      incount,  &
                                      datatype, &
                                      size,     &
                                      ierror)
    use :: iso_c_binding, only: c_char
    use :: mpi_f08_types, only: MPI_COUNT_KIND, MPI_Datatype, MPI_ADDRESS_KIND
    implicit none
    CHARACTER(LEN=*), INTENT(IN) :: datarep
    INTEGER, INTENT(IN) :: incount
    TYPE(MPI_Datatype), INTENT(IN) :: datatype
    INTEGER(KIND=MPI_ADDRESS_KIND), INTENT(OUT) :: size
    INTEGER, OPTIONAL, INTENT(OUT) :: ierror
    INTEGER :: c_ierr

    interface
        subroutine ompi_pack_external_size_wrapper_f08(datarep,  &
                                                       incount,  &
                                                       datatype, &
                                                       size,     &
                                                       ierror) &
            BIND(C, name="ompi_pack_external_size_wrapper_f08")
            use :: iso_c_binding, only: c_char
            use :: mpi_f08_types, only: MPI_COUNT_KIND, MPI_Datatype, MPI_ADDRESS_KIND
            implicit none
            CHARACTER(KIND=C_CHAR), INTENT(IN) :: datarep(*)
            INTEGER, INTENT(IN) :: incount
            INTEGER, INTENT(IN) :: datatype
            INTEGER(KIND=MPI_ADDRESS_KIND), INTENT(OUT) :: size
            INTEGER, INTENT(OUT) :: ierror
        end subroutine ompi_pack_external_size_wrapper_f08
    end interface

    call ompi_pack_external_size_wrapper_f08(datarep,          &
                                             incount,          &
                                             datatype%MPI_VAL, &
                                             size,             &
                                             c_ierr)
    if (present(ierror)) ierror = c_ierr
end subroutine MPI_Pack_external_size_f08

subroutine MPI_Pack_external_size_f08_c(datarep,  &
                                        incount,  &
                                        datatype, &
                                        size,     &
                                        ierror)
    use :: iso_c_binding, only: c_char
    use :: mpi_f08_types, only: MPI_COUNT_KIND, MPI_Datatype
    implicit none
    CHARACTER(LEN=*), INTENT(IN) :: datarep
    INTEGER(KIND=MPI_COUNT_KIND), INTENT(IN) :: incount
    TYPE(MPI_Datatype), INTENT(IN) :: datatype
    INTEGER(KIND=MPI_COUNT_KIND), INTENT(OUT) :: size
    INTEGER, OPTIONAL, INTENT(OUT) :: ierror
    INTEGER :: c_ierr

    interface
        subroutine ompi_pack_external_size_wrapper_f08_c(datarep,  &
                                                         incount,  &
                                                         datatype, &
                                                         size,     &
                                                         ierror) &
            BIND(C, name="ompi_pack_external_size_wrapper_f08_c")
            use :: iso_c_binding, only: c_char
            use :: mpi_f08_types, only: MPI_COUNT_KIND, MPI_Datatype
            implicit none
            CHARACTER(KIND=C_CHAR), INTENT(IN) :: datarep(*)
            INTEGER(KIND=MPI_COUNT_KIND), INTENT(IN) :: incount
            INTEGER, INTENT(IN) :: datatype
            INTEGER(KIND=MPI_COUNT_KIND), INTENT(OUT) :: size
            INTEGER, INTENT(OUT) :: ierror
        end subroutine ompi_pack_external_size_wrapper_f08_c
    end interface

    call ompi_pack_external_size_wrapper_f08_c(datarep,          &
                                               incount,          &
                                               datatype%MPI_VAL, &
                                               size,             &
                                               c_ierr)
    if (present(ierror)) ierror = c_ierr
end subroutine MPI_Pack_external_size_f08_c

[ggouaillardet]

Looks good to me, kuddos for the clever use of LEN(...) to skip declaring a local variable!

[jsquyres]

@hppritcha In the call to ompi_pack_external_size_wrapper_f08, how does the C routine receive the string length argument -- does it have an additional int argument at the end of the list?

Also, if the fortran code is compiled with -fdefault-integer-8, is the implicit string length argument 4 bytes or 8?

[jsquyres]

clever use of LEN(...)

Agreed -- that's cool!

I notice that 4 files were deleted in this PR. Was that intentional?

[hppritcha]

yes the deletion was intentional. they should not have been there. i vaguely recall these were also found and removed as part of #12226

[hppritcha]

@hppritcha In the call to ompi_pack_external_size_wrapper_f08, how does the C routine receive the string length argument -- does it have an additional int argument at the end of the list?

Also, if the fortran code is compiled with -fdefault-integer-8, is the implicit string length argument 4 bytes or 8?

the 'c' routine doesn't need this pasted on string length arguments (generated by the compiler when it doesn't know language is of the invoked routine) because of the interface definition

           CHARACTER(KIND=C_CHAR), INTENT(IN) :: datarep(*)

the fortran compiler generates a 'c' null terminated string under the covers which gets passed to ompi_pack_external_size_wrapper_f08.

[hppritcha]

add on to the above comment. Since the compiler (fortran one) generates a properly formatted 'c' string then there is no 4/8 byte length argument problem.

After we get this in, and i have some time, i'm going to check out replacing the old crufty 'c' code in mpif-h with something like this but adjusted for the mpif.h/use mpi integer args for handles and lack of optional args. obviously no big count for those interfaces.

[jsquyres]

TIL:

the fortran compiler generates a 'c' null terminated string under the covers

Cool!

[jsquyres]

After we get this in, and i have some time, i'm going to check out replacing the old crufty 'c' code in mpif-h with something like this but adjusted for the mpif.h/use mpi integer args for handles and lack of optional args. obviously no big count for those interfaces.

Should this be part of the ABI effort this summer, and/or a reason to examine slurping in the MPI Forum JSON and/or Python companion library?

[hppritcha]

we'll see. i don't want to overwhelm interns with too many projects.

[hppritcha]

@jsquyres odd a rebase to fix conflicts necessitates a re-review I guess.

[hppritcha]

@jsquyres when you have time could you do quick re-review? no changes since you last approved just a rebase on top of main.