To: J3                                                     10-234
From: R. Bader, R. Rabenseifner
Subject: Reconsider CHARACTER(*) and BIND(C)
Date: 2010 October 08
References: 10-165r2, N1814, N1820


TR feature request: add support for CHARACTER(*):
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

MPI has a number of routines which take arguments which are
presently declared as CHARACTER*(*) in the interface. The
MPI-3 effort to implement the complete interface
* in an backward compatible manner
* using BIND(C) for most (if not all) calls
runs into trouble at this point. It is not possible
to convert CHARACTER(*) :: arg to CHARACTER :: arg(*) since
this would require appending C_NULL_CHAR to each actual
argument i.e., to modify existing MPI-based applications.

Therefore, it is requested that the following feature be added
to the functionality covered by the Interop TR:

A dummy argument in a BIND(C) interface may be declared
CHARACTER(kind=C_CHAR, len=*). In this case, the length information
must be passed to the C interface as an extra argument of type
size_t. For example, the Fortran interface

interface
  subroutine fpc(str, x) BIND(C)
    use, intrinsic :: iso_c_binding
    character(kind=c_char, len=*) :: str
    real :: x(*)
  end subroutine
end interface

matches the C prototype

void fpc(char *str, float *x, size_t len);

The extra arguments are passed by value, and are appended at the
end of the argument list in the same order as that of the set of
string arguments which are declared CHARACTER(*) in the Fortran
interface.


COMMENTS:

(1) Support for assumed-length CHARACTER variables was
    explicitly excluded from consideration in N1820 (item 4a, based
    on an earlier WG5 vote). However, this poses the following
    problem for the MPI-3 implementation: Assuming that passing
    an assumed-size actual argument to an assumed-rank dummy is
    supported at best with BIND(C) interfaces, it will not be
    possible to implement routines with both CHARACTER(*) arguments
    and choice buffer arguments of TYPE(*), DIMENSION(..) in a
    backward compatible manner, since passing an assumed size
    actual buffer argument will not be allowed. Hence, WG5 is
    asked to reconsider this decision.

(2) Concerning the possibility of a mismatch between default and
    C-interoperable character kinds in Fortran: For MPI libraries
    it would be more useful to define dummy arguments only as
    CHARACTER(*) and to postpone the conversion problem from
    Fortran default to C characters to the C-written routine.
    The reason for this again is that MPI-based codes would
    require an explicit conversion of character entities
    if the above-mentioned mismatch surfaces. One could, for example,
    consider providing a conversion function to the C programmer
    and making the str argument in the above C prototype a pointer
    to void.