Boxy-type ASSERT failure in 6.10: panic in xmonad-contrib

[dons]

Testing xmonad-contrib-0.8 with ghc 6.10 RC. Probably a release-blocking bug.

[ 35 of 138] Compiling XMonad.Layout.MultiToggle ( XMonad/Layout/MultiToggle.hs, dist/build/XMonad/Layout/MultiToggle.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 6.10.0.20081007 for x86_64-unknown-linux):
	ASSERT failed! file typecheck/TcUnify.lhs line 1000
a{tv aKF8} [box]

Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug

To reproduced, grab cabal-install 0.6 and:

cabal install xmonad-contrib-0.8

[igloo]

You need a DEBUG compiler to reproduce this (so it may not be a regression, but we should take a look anyway):

{-# LANGUAGE ExistentialQuantification, MultiParamTypeClasses,
             FunctionalDependencies, Rank2Types #-}

module Q where

class Transformer t a | t -> a where
    transform :: t -> l a -> (forall l'. l' a -> b) -> b

data EL a = forall l. EL (l a)

unEL :: EL a -> (forall l. l a -> b) -> b
unEL = undefined

transform' :: (Transformer t a) => t -> EL a -> EL a
transform' = undefined

data MultiToggleS ts a = MultiToggleS ts

data MultiToggle = MultiToggle

expand :: HList ts a => MultiToggleS ts a -> MultiToggle
expand (MultiToggleS ts) =
    resolve ts
        (\x mt ->
            let g = transform' x in
            mt
        )
        MultiToggle

class HList c a where
    resolve :: c -> (forall t. (Transformer t a) => t -> b) -> b

$ ghc -fforce-recomp -c q.hs
ghc: panic! (the 'impossible' happened)
  (GHC version 6.11.20081010 for x86_64-unknown-linux):
        ASSERT failed! file typecheck/TcUnify.lhs line 1000 a{tv ahb} [box]

Line 1000 is the second ASSERT here:

unifyType :: TcTauType -> TcTauType -> TcM CoercionI
-- No boxes expected inside these types
-- Acutal and expected types
unifyType ty1 ty2       -- ty1 expected, ty2 inferred
  = ASSERT2( not (isBoxyTy ty1), ppr ty1 )
    ASSERT2( not (isBoxyTy ty2), ppr ty2 )
    addErrCtxtM (unifyCtxt ty1 ty2) $
    uTysOuter True ty1 True ty2

[simonpj]

The assertion is to do with "boxy types" which Dimitrios and I are about (well, over the next few months) to overhaul completely. Can you see if matters are ok without DEBUG on? I bet they will be, and the ASSERT is over-cautious. In which case, we'll let it ride for now.

Simon

[dons]

@simon

xmonad-contrib builds ok without the DEBUG compiler, so might be ok to let it ride.

[simonmar]

almost certainly not arch/os specific

[simonpj]

OK changing to 6.12; I think 6.10 is fine.

Simon

[igloo]

Let's check this is fixed after the typechecker branch is merged.

[simonpj]

I believe this works with the new type checker, at lesat with Ian's cut-down case (which I have added as a regression test).

DonS: can you try with xmonad?

I'll move to 'infoneeded'.

Simon

[simonpj]

No response from Don, so closing. The code that has the ASSERT is gone altogether.