Migrate to ghc-lib-parser

[neongreen]

Fixes #422, fixes #344.

Note to the reviewer

This can be reviewed commit-by-commit.

TODO

• Fix the way I work with XPat.

• Run nixpkgs tests.

• Check that the executable is buildable.

• Fix warnings.

• Format with Ormolu.

• Add a test for:

  type family G a b where
    forall x y. G [x] (Proxy y) = Double
    forall z.   G z   z         = Bool
  

• Implement support for the point above.

• Investigate:

  {-# LANGUAGE TypeFamilies, TypeOperators, NoStarIsType #-}
  
  instance PNum Nat where
    type a * b = ()
  

• Go through singletons and look for more regressions.

• Add a test and implementation for:

  {-# RULES "example" forall a. forall (x :: a). id x = x #-}
  

• Add tests for:

  type family a ! b
  type family a . b
  

• Add paren tests:

  class a + b
  data D1 = forall a b. (a + b) => D1 a b
  data D2 = forall a b.  a + b  => D2 a b -- now allowed
  
  data T = MkT1 { a :: {-# UNPACK #-} (Maybe Int && Bool) }
         | MkT2 { a :: {-# UNPACK #-}  Maybe Int && Bool  } -- now allowed
  
  data G where
    MkG1 :: {-# UNPACK #-} (Maybe Int && Bool) -> G
    MkG2 :: {-# UNPACK #-}  Maybe Int && Bool  -> G  -- now allowed
  
  Proxy '(a :: A, b :: B)
  

• Try out a rule with two levels of multiline foralls.

• Fix regression: the forall should not be broken across several lines:

  sEither ::
    forall a b c (l :: a ~> c) (r :: b ~> c) (e :: Either a b).
    (forall n. Proxy l -> Sing n -> Sing (Apply l n)) ->
    (forall n. Proxy r -> Sing n -> Sing (Apply r n)) ->
    Sing e ->
    Sing (Either_ l r e)
  sEither l _ (SLeft x) = l Proxy x
  sEither _ r (SRight x) = r Proxy x

• Add a test for an (accidental?) improvement: the :<>: declaration should remain on the same line:

  data ErrorMessage' s
    = -- | Show the text as is.
      Text s
    | -- | Pretty print the type.
      -- @ShowType :: k -> ErrorMessage@
      forall t. ShowType t
    | -- | Put two pieces of error message next
      -- to each other.
      ErrorMessage' s :<>: ErrorMessage' s
    | -- | Stack two pieces of error message on top
      -- of each other.
      ErrorMessage' s :$$: ErrorMessage' s

[neongreen]

I'm rather disappointed with GHC here.

Of course, an alternative option is swallowing the bug and doing p -> ($ p) if we get something other than an XPat. I'm not sure what is the best course of action here.

[neongreen]

This is questionable. I don't have enough experience with GHC internals to say whether it's the right thing to do.

[neongreen]

{-
Note [HsArgPar]
A HsArgPar indicates that everything to the left of this in the argument list is
enclosed in parentheses together with the function itself. It is necessary so
that we can recreate the parenthesis structure in the original source after
typechecking the arguments.

The SrcSpan is the span of the original HsPar

((f arg1) arg2 arg3) results in an input argument list of
[HsValArg arg1, HsArgPar span1, HsValArg arg2, HsValArg arg3, HsArgPar span2]

-}

This is weird. And we probably don't need to handle it because currying?

[mrkkrp]

I do not know. I'm OK with this as long as everything works in practice.

[mrkkrp]

What is the status of this one?

[neongreen]

All tests pass and it's ready to be reviewed.

[neongreen]

I don't know why CI fails, might be a timeout.

[mrkkrp]

Good job at reducing the duplication!

[mrkkrp]

This is good to go but is blocked on the CI issues.

[mrkkrp]

Nice!