netcdf-fortran does not honour variant specifications of compiler, but requires g95 instead

[platipodium (Carsten Lemmen)]

When trying to install netcdf-fortran (with openmpi) I get g95 as a requirement and *not* the selected gcc10 or gccdevel on MacOS aarch64

> port install netcdf-fortran +gccdevel+openmpi

--->  Computing dependencies for netcdf-fortran
The following dependencies will be installed:  g95

> port install netcdf-fortran +gcc10+openmpi

--->  Computing dependencies for netcdf-fortran
The following dependencies will be installed:  g95

[jmroot (Joshua Root)]

For some reason the gcc* variants aren't offered when using +openmpi:

% port variants netcdf-fortran +openmpi
netcdf-fortran has the variants:
[+]g95: Build using the g95 Fortran compiler
   mpich: Build using the MPICH compiler
     * conflicts with openmpi
  +openmpi: Build using the OpenMPI compiler
     * conflicts with mpich

[jmroot (Joshua Root)]

I don't see where the problem is on initial inspection, but I guess it must be in the mpi or compilers portgroups. Adding some people who might know more.

[mascguy (Christopher Nielsen)]

It looks like the following commit is disabling most of the compiler support.

​mpi PG: Skip missing compilers on arm

The problem is that for non-ARM builds, it disables almost everything for openmpi:

       # Disable compilers not support on arm
        if {${os.arch} eq "arm" || !$is_mpich} {
            lappend disabled -gcc5 -gcc6 -gcc7 -gcc8 -gcc9 -gcc10
            lappend disabled -clang60 -clang70 -clang80 -clang90 -clang10
        }

With the latest mpi portgroup, all compiler variants are missing when selecting openmpi:

$ port info --variants netcdf-fortran +openmpi
variants: g95, mpich, openmpi, universal

Reverting to a version of the portgroup prior to that commit:

$ port info --variants netcdf-fortran +openmpi
variants: g95, gcc10, gcc5, gcc6, gcc7, gcc8, gcc9, mpich, openmpi, universal

@CJones, do you want to go ahead and fix?

[mascguy (Christopher Nielsen)]

Hmmmm, well, I believe clang90 and clang10 are working on ARM, along with gcc10. So the logic needs to be revisited.

However, it also disables everything for openmpi, when running on x86. And that's definitely a bug.

[mascguy (Christopher Nielsen)]

The logic in question should be the following, I believe:

        # Disable compilers not support on arm
        if {${os.arch} eq "arm" && !$is_mpich} {
            lappend disabled -gcc5 -gcc6 -gcc7 -gcc8 -gcc9 
            lappend disabled -clang60 -clang70 -clang80 
        }

[platipodium (Carsten Lemmen)]

The above solution by @mascguy is confirmed working for me, but still does not solve +gccdevel +openmpi (which may not be supported at this time?)

[Christopher Nielsen <mascguy@…>]

In b57ec390c8239d4ce47353f5adb1b43f60da4431/macports-ports (master):

mpi pg: fix ARM support, along with x86

Fixes: #62835

[mascguy (Christopher Nielsen)]

Replying to platipodium:

The above solution by @mascguy is confirmed working for me, but still does not solve +gccdevel +openmpi (which may not be supported at this time?)

Carsten, is gccdevel necessary for your work, or are you able to use gcc10?

[platipodium (Carsten Lemmen)]

Carsten, is gccdevel necessary for your work, or are you able to use gcc10?

As I am developing on an M1 machine, I'd prefer (native) gcc11 (gccdevel).... but mpich works so I don't depend on openmpi variant

[cjones051073 (Chris Jones)]

Replying to mascguy:

Hmmmm, well, I believe clang90 and clang10 are working on ARM, along with gcc10. So the logic needs to be revisited.

However, it also disables everything for openmpi, when running on x86. And that's definitely a bug.

Clang 9 and 10 build on arm but they are not reliable enough to use really, I have seen them ICE etc. Clang 11 is really the best one to use on arm

[cjones051073 (Chris Jones)]

Replying to Christopher Nielsen <mascguy@…>:

In b57ec390c8239d4ce47353f5adb1b43f60da4431/macports-ports (master):

mpi pg: fix ARM support, along with x86

Fixes: #62835

Allowing gcc10 is fine as that has been fixed on arm since that list was first made, but it is a mistake to allow clang9.0 and 10 on arm as these are not reliable on this platform. Why would you use them when clang 11 is available and has good arm support ?

[mascguy (Christopher Nielsen)]

If the consensus is that Clang 9 and 10 are unreliable on ARM, I'll update the portgroup accordingly.

[mascguy (Christopher Nielsen)]

Never mind, looks like that was already done. :-)

[jmroot (Joshua Root)]

Replying to Christopher Nielsen <mascguy@…>:

In b57ec390c8239d4ce47353f5adb1b43f60da4431/macports-ports (master):

mpi pg: fix ARM support, along with x86

Fixes: #62835

It looks like this change just shifts the problem to mpich instead of openmpi? I don't see why $is_mpich should be relevant to this decision at all.

[mascguy (Christopher Nielsen)]

Replying to jmroot:

It looks like this change just shifts the problem to mpich instead of openmpi? I don't see why $is_mpich should be relevant to this decision at all.

Good catch. @CJones, do you want to fix, or shall I?

[mascguy (Christopher Nielsen)]

Fixed in commit:

​https://github.com/macports/macports-ports/commit/620c2cb33a325c7166caaa40a88b157f323c53b1

[platipodium (Carsten Lemmen)]

And there is more to it:

sh-3.2# port install netcdf-fortran +gccdevel+openmpi
--->  Computing dependencies for netcdf-fortran
Error: Can't install libgcc because conflicting ports are active: libgcc-devel

Is this possibly fixed by your above commits? Otherwise maybe make this a new ticket?

[kencu (Ken)]

Just try doing what it says; deactivate libgcc-devel and install libgcc.

[platipodium (Carsten Lemmen)]

Replying to kencu:

Just try doing what it says; deactivate libgcc-devel and install libgcc.

Cannot work as libgcc-devel is a dependency of gccdevel. So when I request +gccdevel, it should correctly depend on libgcc-devel, *not* libgcc

[kencu (Ken)]

so just disable gcc-devel :>

Edit: oh, I see what you mean now. Requesting netcdf-fortran +gccdevel gives you this libgcc issue. Right.