CI failures with GHC 8.8

[ndmitchell]

E.g. https://github.com/ndmitchell/hlint/runs/3707891836?check_suite_focus=true says:

TEST FAILURE (BAD REFACTORING)
SRC: /home/runner/.cabal/store/ghc-8.8.4/hlint-3.3.4-ff2d1f010e79360de5348a757d7f6c211a5c35a3d44832c024922102288aa22b/share/hlint.yaml:1301:0
INPUT: test1196 = map ((,) (+ 1)) "123"
Refactoring failed: exit code 1

/opt/ghc/8.8.4/lib/ghc-8.8.4/settings: openFile: does not exist (No such file or directory)

@zliu41 - any changes that would have caused that?

[zliu41]

/opt/ghc/8.8.4/lib/ghc-8.8.4/settings: openFile: does not exist (No such file or directory)

This indicates refactor is probably built on a different machine than the one running it (haskell/haskell-language-server#591).

I'm not sure where the refactor binary used in CI comes from; it is not mentioned in ci.yml. My guess is if we cabal install it from ci.yml, the problem would go away.

[zliu41]

#1300 is the fix.

[ndmitchell]

We are now cabal installing it and it still doesn't work - any ideas? The first fix didn't work as it corrupted the cache on subsequent builds, so I moved the install line, and now 8.8 fails once more

[zliu41]

I'm not very familiar with actions/cache, but looks like if you put the install after the cache then it was still using the cached, outdated refactor binary. Perhaps we can add the output of ghc --print-libdir to the cache key, so that whenever libdir changes (which would cause the "openFile: does not exist" error) then the cache is invalidated.

[zliu41]

Hmm I see that in the subsequent builds, only 8.8 succeeded and all others failed. I don't understand how that could happen, since the cache key includes matrix.ghc and so caches for different GHC versions should be independent.

[zliu41]

#1301 seems to work; I tried the build multiple times. Although I still don't know what caused the cache corruption.

[ndmitchell]

Thanks for all your work on this @zliu41! Much appreciated :)