Support template-haskell 2.20.0 and unix-compat 0.7

[locallycompact]

Caught by https://gitlab.horizon-haskell.net/package-sets/horizon-core/-/jobs/197815

[Gabriella439]

Can you explain the unix-related change a bit more? The reason I ask is because I thought the unix package was unavailable on Windows (thus the dependence on unix-compat)

[locallycompact]

Oh, right. In that case this fix is probably wrong. I'm not sure why PosixCompat.User was removed in 0.7, then. The Changelog doesn't say.

[locallycompact]

Regarding haskell-pkg-janitors/unix-compat#2 (comment) . What is expected of these codepaths on Windows?

[Gabriella439]

I'm going to check if the corresponding dhall subcommand is broken on windows; if so in the worst case scenario we can just remove support for that subcommand on Windows

[Gabriella439]

Yeah, so it looks like some of the functionality that dhall to-directory-tree depended on from System.Posix.Compat was broken on Windows. In the common case users don't hit this code path (it's a rarely-used feature of this subcommand) which is probably why nobody noticed this problem just yet.

I think the best approach here is to disable that code path on Windows and I can put up the change for that.

[Gabriella439]

Actually, it seems like there are some tests that depended on Unix permissions that were running successfully on Windows. That seems surprising because as far as I can tell those tests would have transitively run System.PosixCompat.User.get{User,Group}EntryFromName and failed:

dhall-haskell/dhall/tests/Dhall/Test/DirectoryTree.hs

Line 32 in 7b40b1c

, fixpointedUserGroup

cc: @mmhat: I wanted to know if you had any insight here.

[mmhat]

@Gabriella439 I'll have a look.

[mmhat]

Ok, the test didn't fail because we never tried to actually set the user or group in fixpointedUserGroup: This test just checks that we can parse the encoded directory tree and doesn't interpret it. That is due to the restrictions on the various platforms: On Windows, the implementation of unix-compat simply doesn't work while on Linux the restriction of chown(2) makes it difficult to test. From the man page:

The owner of a file may change the group of the file to any group of which that owner is a member.

That means, writing a proper test for that would mean that we make assumptions about the groups the user running the test is a member of... Unless I am missing something here.

Since metadata is broken on Windows I think the best is to disable it there and throw an error if the user attempts to set it explicitly: #2507

[Gabriella439]

@locallycompact: If you merge locallycompact#1 into your branch then I think it will fix the build failure on Windows

[locallycompact]

Awesome thanks.

Incidentally though this th fix is wrong though because it's 2.11.0 that should be the future-proof bound. However, PvP isn't respected for unreleased commits on ghc master, 2.11.0 doesn't exist yet, so I'm rethinking how to approach future proofing.

I'll resubmit what can be merged.

[ysangkok]

@locallycompact Are you still working on this? Dhall was remove from Stackage because of this issue.

[locallycompact]

I will tidy this up, I'm a little indisposed atm but by all means feel free to do it before me.

[locallycompact]

Closed in favor of #2541