[CI] Run on Windows

[giordano]

Test in view of #43, I don't know if this will work, let's see

[giordano]

Some progress: compilation is successful on Windows, but tests are failing because tdms.exe can't be found.

[giordano]

I'm completely stuck at this point and I have no clue of what to do. There are two orthogonal issues:

1. tdms.exe is built successfully but it isn't installed. This could be solved by running cmake --install . --config ${{ matrix.build_type }}, but this would install the program in a directory (I think it was something like C:\Program Files (x86)\tdms\bin) which isn't in the PATH and so the executable can't be found anyway, which isn't much progress. I see two options:
   1. Search the executable also in a Release/ subdirectory of the current folder. I don't like this option much, but might be OK-ish (but still very meh) for the purpose of running the tests in CI
   2. Add the installation directory to the PATH. I'd say this is the "proper" solution in the Unix world, but I'm not sure people on Windows are accustomed to fiddling with these settings
2. For the life of me I can't make the dynamic loader on Windows find the needed dynamic libraries, like fftw3.dll or libmat.dll/libmex.dll. I tried to add the relevant directories to PATH (the Windows equivalent of LD_LIBRARY_PATH), to no avail. The only working solution is to...cd to directory where fftw3.dll is, but we definitely can't expect users to move around (and I also have opinions, not positive ones, about this behaviour of the loader). Also, we can't expect fftw and matlab libraries to be in the same directory

[t-young31]

I've also just spent an hour or two on this and I am also stuck.

1. I think moving the executable from build/Release to build/ would be a reasonable solution to combat the hilariously inconsistent behaviour
2. Is more troubling.. I also found I needed to copy the .dlls into the same directory 😕 , which is not very user friendly. There also doesn't seem to be a way of using any kind of RPATH so I'm also unsure of the solution.

I suppose a static link to fftw and no matlab dependency would solve this, depending on our priorities we may want to leave this for a bit

[prmunro]

I'm not sure whether my review is required here or not. Are you asking whether to persist with this? I don't think I can offer a solution to this. I only ever compiled it under windows using Visual Studio. I understand why you're using cmake though.

[giordano]

Note that we aren't having difficulties with compiling the program, but with running it: it can't find fftw and matlab shared libraries

[prmunro]

I'm happy to approve this noting that at present the code won't run under windows. Can we add this as something to come back to?

[t-young31]

Can we add this as something to come back to?

I'd be happy to do that

---

Comment for (potentially) the future: Alessandro mentioned on Friday it's apparently common to copy all the dependencies into the target directory. I also found this while googling dynamically linking on Windows. There is a CMake command in v3.21 to do just that but it requires SHARED libraries, and doesn't seem to work with the current CMake setup. With ${FFTW_ROOT} and ${Matlab_ROOT_DIR} it is possible to find the required dlls in a semi portable way, and copy them to the target directory. Ugly but maybe required – open to other better options!

[t-young31]

TDMS is now successfully building on GitHub actions under Windows! Turns out setting $PATH can work to find runtime .dlls (and is required because the MATLAB shared libraries have dependencies beyond just libmat.dll libmex.dll and libmx.dll). Due to some oddness with the zip file handling not all the system tests are run, but have instead been checked using an interactive session (following the CI build config/compile). Given this now closes #43 I'd be happy to merge this in and add an issue for running a useful system test on Windows CI, if required.