OC 1.9.0: allocatable component in derived type coarray can't be allocated

[MichaelSiehl]

Defect/Bug Report: OpenCoarrays 1.9.0: allocatable component in derived type coarray can't be allocated#

Are allocatable components in derived type coarrays are really supported with the current OpenCoarrays 1.9.0 master download? I tried it out but the allocate statement itself does (seem to) crash the application. I have a small example program here: https://github.com/MichaelSiehl/Coarray-with-Allocatable-Component-Example
The following ALLOCATE statements did crash my example program:
ALLOCATE (Coarray_Object % reaDataArray (LBOUND (reaDataArray, DIM=1) : UBOUND (reaDataArray, DIM=1)))
and also the simpler form
ALLOCATE (Coarray_Object % reaDataArray (1 : 2))
did crash my example program. Therefore, it must be the ALLOCATE statement itself that makes the program hang.

• OpenCoarrays Version: 1.9.0, the current master
• Fortran Compiler: gfortran 8.0.0
• C compiler used for building lib: gcc 8.0.0
• Installation method: make
• MPI library being used: MPICH 3.2
• Machine architecture and number of physical cores: Linux Ubuntu 14.04 on an intel icore, 4 cores

[zbeekman]

Hi Michael,

OpenCoarrays Version: 1.9.0, the current master
Sorry... which is it master or 1.9.0?

• Fortran Compiler: gfortran 8.0.0
• C compiler used for building lib: gcc 8.0.0

Does this mean you're building GCC and GFortran from source using the current GCC trunk? I ask, because the latest stable release is 7.1.0 not 8.0.0. Using OpenCoarrays with GCC trunk is not officially supported.

• Installation method: make

So you're not using CMake, you're just typing make? Can you try using either the "Advanced CMake based installation" or the installer script?

I'm tempted to completely remove the Makefiles from the src tree, since I think they are used almost exclusively by Alessandro (@afanfa) and are not really intended for end users.

[zbeekman]

Hi @MichaelSiehl,

When I try to run your code locally using MPICH and OpenCoarrays (master) with GFortran/GCC 7.1 on MacOS, the program hangs. I'm still digging around a bit, but this does look like it's either a bug, or due to some yet to be implemented feature. @vehre any chance you could take a look at Michael's code? (https://github.com/MichaelSiehl/Coarray-with-Allocatable-Component-Example)

[zbeekman]

@vehre Here is a minimalistic reproducer:

program main

  implicit none
  type :: nonsymmetric
     real, dimension(:), allocatable :: arr
  end type

  type(nonsymmetric), codimension[*] :: parent_obj

  sync all

  if (this_image() == 2) then
     allocate(parent_obj%arr(3))
  end if

  sync all

  if (this_image() == 1) then
     print*, 'Image 2 has size ', size(parent_obj[2]%arr), ' asymmetric allocation'
  end if
end program

Any ideas why this hangs? Tested with master @ 232d234

[afanfa]

@zbeekman please do not remove the Makefile. I know several users that rely on that and they would not be happy to deal with CMake.

[zbeekman]

I've confirmed this bug happens using CMake too, so it's not the Makefile's fault 😄

[zbeekman]

@afanfa would it be acceptable to keep the Makefiles in the git repo, but exclude them from release archives? That way most users who get official releases won't be tempted into running make?

[MichaelSiehl]

Hi Zaak,
thanks for investigating on this. Also with your short example, it seems to be the ALLOCATE statement that causes the failure.
I did download the master directly and did assume it was 1.9.0. Also, I was using the GGC 8 trunk (which is, apart from this, well working with OpenCoarrays). Until now, I did just use make to compile OpenCoarrays (because it worked well and was simple to use -one can easily compile and use different versions of the OpenCoarrays library on the same system, even with the same terminal window-). Nevertheless, it's no problem to use cmake instead. So, for me it would be just fine if the make file would be in the git repo only and not in the release.
cheers

[vehre]

There is a branch now issue-399 which addresses the issue for gcc 7+ unfortunately does is break some things on gcc 6.x. Still investigating.

[rouson]

I have found that there are almost as many preferred installation methods as there are users. While we can't provide every installation method to every user, keeping the Makefile in the repository but out of releases seems like a good compromise. Thanks for the suggestion, @afanfa. We might need to hang a sign on the Makefile that says, "Abandon all hope, ye who enter here." ;)