[flang] bug passing non-contiguous array to MPI procedure #138471
Description
I've run into errors with flang 20.1.0 when passing a non-contiguous dummy array to MPI procedures (MPICH 4.3.0), where it needs to be doing copy-in/copy-out. It appears that the address of the initial array element is being passed instead of the address of a contiguous copy of the array. I'm not sure if this is a bug with flang, or a problem with MPICH where it doesn't know how to provide a correct mpi module for flang.
I've pared it down to a small (serial) reproducer that doesn't use MPI at all, but mimics the mpi module from MPICH 4.3.0. There are two files mpi_bcast.c:
// Dummy version of MPI_Bcast to dump the received buffer
#include <stdio.h>
void mpi_bcast_(void *buffer, int *count) {
float *vector = (float*) buffer;
printf(" in mpi_bcast: buffer=");
for (int i = 0; i < *count; ++i) {
printf(" %f", vector[i]);
}
printf("\n");
}and the main program flang-20240504.F90:
module mpi_dummy
!! Actual form of the interface from the MPICH 4.3.0 mpi module
interface MPI_Bcast
subroutine MPI_Bcast(buffer, count)!, datatype, root, comm, ierror)
implicit none
#ifdef __flang__
!DIR$ IGNORE_TKR buffer
#endif
#ifdef __INTEL_COMPILER
!DEC$ ATTRIBUTES NO_ARG_CHECK :: buffer
#endif
#ifdef __GFORTRAN__
!GCC$ ATTRIBUTES NO_ARG_CHECK :: buffer
#endif
real :: buffer
integer :: count
!integer :: datatype
!integer :: root
!integer :: comm
!integer :: ierror
end subroutine
end interface
end module
program main
use mpi_dummy
real :: x(5)
x = [1,2,3,4,5]
print *, 'x=', x
call bcast(x(1::2))
contains
subroutine bcast(buffer)
real, intent(inout) :: buffer(:)
print *, ' in bcast: buffer=', buffer
! A CONTIGUOUS COPY OF BUFFER MUST BE PASSED HERE
call MPI_Bcast(buffer, size(buffer))
end subroutine
end programTo compile:
clang -c mpi_bcast.c
flang flang-20240504.F90 mpi_bcast.oThe expected output from running should be
$ ./a.out
x= 1. 2. 3. 4. 5.
in bcast: buffer= 1. 3. 5.
in mpi_bcast: buffer= 1.000000 3.000000 5.000000
But with flang I'm getting
$ ./a.out
x= 1. 2. 3. 4. 5.
in bcast: buffer= 1. 3. 5.
in mpi_bcast: buffer= 1.000000 2.000000 3.000000
Both Intel ifx and gfortran produce the expected results.