nix: no more yolo on ghc 9.4; test suites generally work!

[lf-]

I tried using HLS again in one of my projects and found that the nix packaging
isn't possible to use with other projects due to overriding mkDerivation
and causing world-rebuilds. (work vendored most of the HLS flake
because of this but I want to actually fix it upstream)

It turns out that the GHC 9.4 set is stable enough to remove the yolo anyway.

I've also added an overlay attribute that uses the nixpkgs set without
the other overrides to mkDerivation, for use as an overlay in other flakes.

[lf-]

hls-cabal-plugin> src/Ide/Plugin/Cabal/Parse.hs:21:1: error:
hls-cabal-plugin>     Ambiguous module name ‘Distribution.Types.Version’:
hls-cabal-plugin>       it was found in multiple packages:
hls-cabal-plugin>       Cabal-3.6.3.0 Cabal-syntax-3.8.1.0

how does that even happen?! There are no mentions of Cabal-syntax anywhere I can find and my checking finds it is version 3.6, with version 3.8.1.0 as a separate version there.

[lf-]

Ohhhh. Gross. It's nixpkgs.

  # 2022-10-27: implicit-hie 0.1.3.0 needs a newer version of Cabal-syntax.
  implicit-hie = super.implicit-hie.override {
    Cabal-syntax = self.Cabal-syntax_3_8_1_0;
  };

[lf-]

Yeah the build was jacked before I touched it, that explains a lot 😇

Explanation of the eldritch nix horrors that have been invoked here: what happens when you override boot packages is "you can't". So you can't override cabal to be newer without forcing the package set to be inconsistent and have multiple versions of the same package wind up in a package database and cause build failures. That is just what happens if you override a boot package. Nix/cabal goes all zalgo-text on you :(

[fendor]

Changes to the cabal plugin look good to me!

[michaelpj]

Hmm, the Nix CI jobs are still failing, would you expect them to succeed?

[guibou]

Great.

Note for the future: it is possible to change the YOLO thing so the jailbreak and doCheck are only added if the package is broken. This improves a lot the caching.

[lf-]

Hmm, the Nix CI jobs are still failing, would you expect them to succeed?

nope, they were broken before i changed anything

[lf-]

Great.

Note for the future: it is possible to change the YOLO thing so the jailbreak and doCheck are only added if the package is broken. This improves a lot the caching.

this is actually done already in flake.nix :) so the yolo is unnecessary

[lf-]

I believe that this doesn't make the CI state any worse, as far as I can tell. So it should be mergeable, with further CI fixes as future work.

[lf-]

Aah, no, the benchmark build gets silly bounds problems since cabal-syntax and cabal have to be the same version, and cabal syntax 3.8 allows building on old ghc even though this causes this problem.

[lf-]

I am not sure how to inject an extra constraint into the Cabal 3.6 cabal file. So I'm not sure how to fix this benchmark. Does anyone have any ideas?

[cydparser]

@lf- Try removing Cabal from build-depends.

[lf-]

yep you're right @cydparser. Going to see if this fixes the older CI as well.

[michaelpj]

Let's see if we can get #3383 in, then?

[cydparser]

@lf-, Cabal-syntax was removed as a dependency in the mainline branch. Try removing commit d5afada and rebasing.

[fendor]

Is this still valuable to have?
EDIT:
What I mean is, should we just rebase this PR and then it is ready to go?

[lf-]

I think yes because it would utilize nixpkgs binary caches better to not disable all test suites. I'm not sure if I'll get around to rebase super soon.

[cydparser]

@fendor, these changes should already be in master: https://github.com/haskell/haskell-language-server/pull/3503/commits

[michaelpj]

I'm assuming this is stale, please reopen if not.