use ghcup on windows (clone of #192)

[andreasabel]

Rebased clone of #192 to run latest CI.

• use ghcup from PATH on windows #192

Please also consider the discussion at #192.

TODO (correctness):

• Fix unintended breakage of apt install method.
• Write cabal config so that symlinking is off on Windows by default.
  • Windows: cabal-3.0/2/?: symlinking feature not available ghcup-hs#808
• Exhaustively test different versions of cabal/ghc in CI for launch of the ghcup/Windows install method.
• Blocker: stack-building Agda with cpphs on Windows fails with ghcup-installed ghcs 9.2 and 9.4 ghcup-hs#823

TODO (features):

• Bring back choco install method (by option and/or as fallback). E.g. we could have an input setup-method which is a list of install methods that are tried in order (if available for that platform). The default could be for now setup-method: [ghcup, choco, apt], but one could opt out of fallback by e.g. just setup-method: [ghcup].
  (Bikeshedding: method, install-method.)

Closes #70. Closes #97. Closes #112. Closes #191.

• Consider ghcup for windows #70
• Make ghc installation consistent across all os's (use ghcup as fallback for windows) #97
• Option to use ghcup on Windows #112
• ghcup not found on windows-2022 version 20230214.1 #191

[andreasabel]

Atm, this PR also seems to destroy the apt installation method (fallback on linux): https://github.com/haskell/actions/actions/runs/4410084290/jobs/7727167607#step:3:146
We might give up apt soon when ubuntu-18.04 is removed at the end of the month:

• The Ubuntu 18.04 Actions runner image will begin deprecation on 2022/08/08 and will be fully unsupported by 2023/04/03 actions/runner-images#6002

Ubuntu-20.04 will only support a few legacy GHCs: relevant are 7.4.2 7.2.2 7.0.4 (and maybe 7.10.3). https://launchpad.net/~hvr/+archive/ubuntu/ghc?field.series_filter=focal

[andreasabel]

Windows CI with cabal 3.0 and 3.2 fails with error cabal: symlinking feature not available on Windows. Not sure if this should be fixed in ghcup or whether the setup action needs to work around this problem.

• Windows: cabal-3.0/2/?: symlinking feature not available ghcup-hs#808

[ulysses4ever]

The symlinking issue has a long history on cabal bugtracker (e.g. haskell/cabal#5748). Chocolatey fixes it here: https://github.com/Mistuke/CabalChoco/blob/656b693745acd0f91d2ef06b3ecfb2ab9051d65d/3.2.0.0/cabal/tools/chocolateyInstall.ps1#L286 If ghcup doesn't include the same fix, then either the action has to include it (it's simply a matter of cabal user-config update -a "install-method: copy") or just stop trying to fix this issue. I'm not sure how useful it is to support/patch earlier cabal-installs, so the second option doesn't sound too bad to me.

[andreasabel]

haskell/cabal#5870

@ulysses4ever wrote

(it's simply a matter of cabal user-config update -a "install-method: copy")

Thanks for the pointers. It seems that these config file settings are available since cabal 3.2.0.0 only:

• bindir-method/install-method, installdir, new-install flags in SavedConfig -- 3.0 version cabal#5870

So it is good to test relevant cabal versions on Windows (3.0, 3.2). Added to CI.

[ulysses4ever]

Good to see the progress here, thank you.

More generally, I wanted to share a doubt I was contemplating over the weekend after I learned a bit about the CabalChoco/GhcChoco repos. It looks like supporting Windows required a lot of downstream patching. The fact that CI caught the symlinking issue may be a sheer luck, and more issues may go unnoticed until the switch from Chocolatey to ghcup goes public. So, maybe the change in its current form is a bit radical. Maybe something more cautious like switching to ghcup only when newer cabals are requested (e.g. starting from 3.10 or something). It may also be easier to convince the ghcup maintainer to patch new releases when/if such a need pops up than it is for old releases.

This is just some rough thought, and it's up to the maintainers of respective facilities to decide, of course.

[andreasabel]

I brought back Choco as default, for the sake of backwards compatibility.
CI passes, but I have some regression in the wild that could be caused by switching to ghcup:

• Agda build failure with cpphs and GHC 9.4.5 on Windows commercialhaskell/stack#6111