Little fix for cygwin&msys2 environment

[proukornew]

No description provided.

[vondele]

I'll merge this one shortly

[vondele]

seems like this again broke something now compiling under linux within fishtest. Might need another revert official-stockfish/fishtest#952

[gvreuls]

It's unfortunate this silently breaks fishtest because the solution was tested after the revert of the original patch and seemed to work perfectly.

I think this isn't our bug per se but something going wrong with what the ffishtest worker assumes is its current working directory, either because of a bug in games.py (where some directory juggling is going on) or perhaps even in the python interpreter itself (Python 3.9.5 on my workers).

[proukornew]

subprocess.check_call(
            MAKE_CMD + ARCH + " -j {}".format(concurrency) + " profile-build",
            shell=True,
            env=dict(os.environ, CXXFLAGS="-DNNUE_EMBEDDING_OFF"),
        )
        try:  # try/pass needed for backwards compatibility with older stockfish, where 'make strip' fails under mingw.
            subprocess.check_call(
                MAKE_CMD + ARCH + " -j {}".format(concurrency) + " strip", shell=True
            )
        except:
            pass
        shutil.move("stockfish" + EXE_SUFFIX, destination)

Look consistent and no directory juggling

[gvreuls]

   tmp_dir = tempfile.mkdtemp(dir=worker_dir)

    try:
        os.chdir(tmp_dir)
        with open("sf.gz", "wb+") as f:
            f.write(
                requests.get(
                    github_api(repo_url) + "/zipball/" + sha, timeout=HTTP_TIMEOUT
                ).content
            )
        zip_file = ZipFile("sf.gz")
        zip_file.extractall()
        zip_file.close()

        os.chdir(glob.glob(os.path.join(tmp_dir, "*/src/"))[0])

Looks like directory juggling to me, and I don't know if os.chdir() changes the python working directory too, I'm not a python specialist.

[gvreuls]

The error with the fishtest worker only occurs if the ccache compiler cache is installed and disabling the cache for the worker fixes the problem, so this is due to ccache incorrectly assuming that the instrumented object files created during the make profile-build process are cacheable and reusable in later builds.

The patch is fine but exposes a ccache bug that's easily mitigable by disabling the cache for the worker (by setting CCACHE_DISABLE="1" in the worker's environment).