Add GitHub Actions CI for testing

[lukel97]

Following the discussion in #6 , it looks like we should probably move away from CircleCI at some point. This adds GitHub Actions CI workflows, reusing the infrastructure we already have for the static binaries. It should run the tests on linux, macOS and windows now.

[jneira]

We will need haskell-language-server executable (without make the exe name shorter)

[jneira]

It seems we need stack on path to make wrapper tests pass?

EDIT
In fact it is the exe wrapper itself, we would need to make the env var workaround for it too but why we are keeping the executables with shortened names around? We could use the original ones, no?

EDIT2
Ahh, it will fail in windows cause build-tool-depends uses the executable from dist-newstyle 😟

[lukel97]

@jneira I've added another environment var HLS_WRAPPER_TEST_EXE, seems to work ok for now. Hopefully we can get rid of this whenever ghc supports the windows long paths extension!

[lukel97]

Looks like those windows failures are genuine. Opened up a quick PR to hopefully fix them here: haskell/hie-bios#256

[lukel97]

Not sure about these ones though:

expected: "module T17 where\n\n-- >>> :type  +no 42\n-- parse error on input \8216+\8217\n"
but got: "module T17 where\n\n-- >>> :type  +no 42\n-- parse error on input `+'\n"

Is this ghc outputting a different encoding or can the tests set anything to encode it as the one below? @jneira

[konn]

IIRC, GHC changes quotation marks according to the locale settings. I encountered the same behaviour when I contributed to https://github.com/nomeata/inspection-testing

[jneira]

Yeah, hie had similar failing tests too.
Afair not all ghc versions has the same behaviour so ignore quotation marks usually is the best bet.

[jneira]

I think delimiters used by ghc in windows is due to lack of support for utf-8 in windows shell by default. So to avoid output weird replacements they decided to use ascii delimiters for windows.
In ghcide recently i remove the delimiters from the expected string and in hie i duplicated expected strings, trying both variants always. It is ugly but somewhat worked.

[lukel97]

@jneira thats the approach I've ended up using, I went down a rabbit hole today of trying to understand what decides the default localeEncoding but didn't get anywhere. If this gets merged we should probably squash all my trial and error commits!

[jneira]

did you add -j1 due a concrete issue?

[lukel97]

Out of precaution, the tests previously failed on circleci because without -j1 tasty attempts to run all the functional tests in parallel, which means it spins up 20 or so hls instances simultaneously. I'm not sure if the GitHub runners can handle it or not but the circleCI runners definitely couldn't!

[jneira]

Maybe it worths a comment?

[jneira]

Amazing work, as usual. Many thanks, specially for take care of minority os's.