build on 9.2+

[pillowtrucker]

Hello again! I've been trying to build haskus-system and its dependencies on ghc 9.8.1 and I believe I've already managed to patch most of the ghc API incompatibilities in the haskus-binary package here: https://github.com/pillowtrucker/haskus-packages/commit/176709d46ca6956ed4753e7a325018a61e39ce70 but there's still one error thrown by cabal that I just don't understand:

src/lib/Haskus/Memory/Buffer.hs:167:18: error: [GHC-18872]
    • Couldn't match a lifted type with an unlifted type
      When matching types
        a0 :: *
        (# (# #) | MutVar# RealWorld [IO ()] #) :: TYPE
                                                     (SumRep [ZeroBitRep, UnliftedRep])
      Expected: a0
        Actual: Finalizers# RealWorld
    • In the first argument of ‘mkWeak#’, namely ‘fin’
      In the expression: mkWeak# fin b (unIO $ runFinalizers fin) s
      In the expression:
        case mkWeak# fin b (unIO $ runFinalizers fin) s of
          (# s1, _wk #) -> (# s1, () #)
    |
167 |     case mkWeak# fin b (unIO $ runFinalizers fin) s of

I've tried to work around it by using compatMkWeak# from the basement package, but that seems to be an even earlier implementation of mkWeak# and it doesn't help. I'm afraid I don't really understand what has changed there any how/why

[pillowtrucker]

Would something like this do what was originally intended ?

addFinalizer :: Buffer -> IO () -> IO ()
addFinalizer b f = do
   let !fin = getFinalizers b
   case fin of
     Finalizers rfs -> do
       wasEmpty <- insertFinalizer fin f
     -- add the weak reference to the finalizer IORef (not to Addr#/byteArray#/...)
       when wasEmpty $ IO \s ->
         case mkWeak# (IORef (STRef rfs)) b (unIO $ runFinalizers fin) s of
           (# s1, _wk #) -> (# s1, () #)
     NoFinalizers -> error "nonsense"

The main difference seems to be that mkWeak# does not take unlifted types since 9.x, but I don't know what I'm doing here really. I've also considered changing fin into \_ -> fin, but that seems more obviously wrong than the above

[pillowtrucker]

As mentioned in https://github.com/haskus/haskus-system/issues/3 — I think I have at least basic functionality working under ghc 9.8.1 now. Would you like me to try to refactor the changes to this repository and to haskus-system as a real PR ? https://github.com/haskus/packages/compare/master...pillowtrucker:haskus-packages:mine

[pillowtrucker]

haskus-ui, haskus-maths and haskus-calculus seem to have problems with https://gitlab.haskell.org/ghc/ghc/-/issues/18806, so I'm just replacing the macro with the expansion as per https://github.com/haskus/haskus-manual/blob/master/source/eadt/basics.rst, but letting the compiler infer what the signature actually is there

[hsyl20]

Hello! Thanks for the nudge :D

I've pushed some fixes: I can now build locally with 8.10, 9.0, 9.2, 9.4, 9.6, 9.8

• Finalizers: the code was wrong I think. I've made the weak pointer on a cons cell of the finalizers list instead.
• eadtPattern was generating linear pattern synonyms which don't work (cf the issue you linked). But there was no reason to generate linear pattern synonyms in the first place, so i've fixed that
• change to numeric types: I've fixed those similarly to what you did but using CPP to keep backward compatibility

[pillowtrucker]

that's great :D thank you. I will test again with the new version

[pillowtrucker]

works great, cheers!