Fix gac to work with /bin/sh that is not bash

[fingolfin]

No description provided.

[ChrisJefferson]

need a cd .. here

[codecov]

Codecov Report

Merging #1292 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1292   +/-   ##
=======================================
  Coverage   61.36%   61.36%           
=======================================
  Files         916      916           
  Lines      277218   277218           
  Branches    15335    15335           
=======================================
  Hits       170118   170118           
  Misses     103181   103181           
  Partials     3919     3919

[fingolfin]

Actually, trying to test build Browse like that won't work, because it also needs a source code change. So I switched this to edim now.

[frankluebeck]

Without this PR it was not possible to gac-compile the kernel modules of the EDIM and Browse packages on a Debian 8 or on an Ubuntu 16.04 system (which start /bin/sh scripts in strict sh-compatibility mode). With this PR it works again, thanks.

I noticed that for example for the EDIM package after compilation (on Debian 8, x86_64) the bin
directory contains 76 kB. With GAP 4.8.6 it contains only 16 kB. It seems that in the new setup one can delete everything except the generated *.so file (26 kB, which in turn can be striped to 8 kB). Shouldn't the additional .libs directory and the *.la file be cleaned up by gac?

[fingolfin]

Yes, perhaps these should be cleared (although we might have to reconsider that when implementing make install -- but that's of course not an argument about cleaning these files for now, I just mention it for completeness).

However, this PR is strictly meant to fix the bug; I'd prefer to not also add other changes in here, but instead get it merged ASAP. So perhaps this could be filed as a feature request issue? Or feel free to submit a PR implementing it.

To this end, I just pushed a new revision, in an attempt to make the CI tests pass (the previous iteration failed because gac does not currently work on HPC-GAP, see #1295. The short term workaround is to just not test it on HPC-GAP).

[fingolfin]

The failure on Windows is another gac bug; I think I know how to fix it, will look into it when I have some time.

[fingolfin]

On Windows, we need to link the module (e.g. edim.so resp. edim.dll) against gap.dll (resp. libgap.la). The main "problem" doing that is to teach gac that. So we'll either need to let gac know when we are on Windows (by tweaking how we generate it), or else pass the required linker options directly to it from configure. Either works, it's a matter of taste in the end, I guess.

But I don't have time right now to work on it.

[fingolfin]

The new revision of this PR disables the edim test on Windows, too; and I created issue #1298 to log the problem.

[fingolfin]

From my POV this could be merged now.