Functional dependencies regression

[rl]

Testcase:

{-# LANGUAGE FlexibleInstances, UndecidableInstances, MultiParamTypeClasses, FunctionalDependencies #-}
module Foo where

class Mul x y z | x y -> z
class IsType a
class IsType a => IsSized a s | a -> s

data Array n a = Array
instance IsSized a s => IsType (Array n a)
instance (IsSized a s, Mul n s ns) => IsSized (Array n a) ns

ghc-7.0.0.20101014 rejects this with:

    Couldn't match type `s' with `s1'
      because this skolem type variable would escape: `s1'
    This skolem is bound by the instance declaration
    In the instance declaration for `IsSized (Array n a) ns'

ghc-7.0.0.20101005 and all previous versions accept it. This is from the llvm package, so is fairly critical.

[igloo]

Fixed in HEAD by:

Fri Oct 15 14:08:18 BST 2010  simonpj@microsoft.com
  * Fix Trac #4401: meta-tyvars allocated by the constraint solver are always touchable

but this causes break001(ghci) and break006(ghci) to fail, so I haven't merged it yet.

[simonpj]

I'm working on break001/6. Simon

[simonpj]

This is fixed now. Test is in place.

[rl]

This hasn't been merged yet. Should the status be changed to merge? We might forget otherwise.

[igloo]

The patch above has been merged, as has the test. Please reopen if there's another patch that needs to be merged for this ticket.

[rl]

That's weird. LLVM builds for me with the head snapshot from Oct 19 but not with the branch snapshot from the same day. Or perhaps I messed up. I'll investigate.

[igloo]

It was merged on the 19th, after the nightly builds started:  http://www.haskell.org/pipermail/cvs-ghc/2010-October/056892.html

[rl]

That explains it then. Sorry about the noise. Thanks!