mpi portgroup: should not offer gcc variants that won't work

[SebKeen]

Hi,

I just tried to install a new port, and had to update mac port to version 2.5.4, and also Xcode to version 10. The port fails to update because of dependencies related to gcc49 and message that the compiler is not supported for Xcode 9.0 and higher. Thus, I tried to port install gcc9, and this again fails because of a dependency on hdf5-18 @1.8.17+cxx+gcc49+hl.

Here is the log:

bash-3.2$ sudo port install gcc9 
--->  Computing dependencies for gcc9
--->  Cleaning gcc9
--->  Scanning binaries for linking errors
--->  Found 9 broken files, matching files to ports
--->  Found 5 broken ports, determining rebuild order
You can always run 'port rev-upgrade' again to fix errors.
The following ports will be rebuilt:
 hdf5-18 @1.8.17+cxx+gcc49+hl
 cdo @1.7.2+szip
 ncview @2.1.1
 wgrib2 @2.0.5
 grads @2.0.2
Continue? [Y/n]: y
--->  hdf5-18 is replaced by hdf5
--->  Fetching distfiles for gcc49
Error: building gcc49 is not supported with Xcode 9 or greater
Error: Failed to fetch gcc49: unsupported platform
Error: See /opt/local/var/macports/logs/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_gcc49/gcc49/main.log for details.
Error: Problem while installing gcc49
Error: rev-upgrade failed: Error rebuilding hdf5-18
Error: Follow https://guide.macports.org/#project.tickets to report a bug.

I hope someone can point in the right direction.

Noel

[jmroot (Joshua Root)]

You installed gcc9 successfully; the failure was in rebuilding hdf5-18, which was found to be broken by rev-upgrade (which runs automatically after each installation and checks all installed ports). hdf5-18 has been replaced by hdf5, so hdf5 was built with the same variants that hdf5-18 was previously installed with, but the gcc49 variant no longer works with your new Xcode version.

You can fix this by simply uninstalling hdf5-18 and installing hdf5 with some newer gcc variant. However it seems like hdf5 shouldn't offer the variant in the first place when it won't work.

[ryandesign (Ryan Carsten Schmidt)]

Should it really be the responsibility of every port that offers gcc variants to know which gcc versions are supported on which systems? The gcc ports already know this information, and thus the correct information was already displayed to the user.

The hdf5 port doesn't define these variants itself; it looks like it uses the mpi portgroup to do so. The compilers portgroup also exists which does something similar. If we really want these compiler variants not to be declared on systems where they won't work, that logic should be in the portgroups, not in the individual portfiles.

[jmroot (Joshua Root)]

Additionally, the portgroup adds variants in some cases that depend on ports that simply don't exist. For example:

% port lint py37-mpi4py +gcc49
--->  Verifying Portfile for py37-mpi4py
Error: Unknown dependency: mpich-gcc49
--->  1 errors and 0 warnings found.

% port lint py37-mpi4py +clang37
--->  Verifying Portfile for py37-mpi4py
Error: Unknown dependency: mpich-clang37
--->  1 errors and 0 warnings found.

[jmroot (Joshua Root)]

In dcc2ab5d0ae8bfbd6b6a476b73386a99a34d5256/macports-ports (master):

mpi pg: only allow compiler versions that work

Fixes: #57415

[ryandesign (Ryan Carsten Schmidt)]

Replying to jmroot:

Additionally, the portgroup adds variants in some cases that depend on ports that simply don't exist.

Well that's another matter. I don't think anyone expected ports just to disappear on certain OS versions. Certainly our mirroring script doesn't expect that. See #59309.

[jmroot (Joshua Root)]

In dcc2ab5d0ae8bfbd6b6a476b73386a99a34d5256/macports-ports (dar, master, py38-reproject, revert-6945-rust-1.43.0, wireshark):

mpi pg: only allow compiler versions that work

Fixes: #57415