testsuite/tests/typecheck/should_fail/tcfail138.hs

   1 {-# LANGUAGE MultiParamTypeClasses, FunctionalDependencies,
   2              FlexibleInstances, UndecidableInstances #-}
   3 -- UndecidableInstances because (L a b) is no smaller than (C a b)
   4
   5 -- This one shows up another rather subtle functional-dependecy
   6 -- case.  The error is:
   7 --
   8 --    Could not deduce (C a b') from the context (C a b)
   9 --      arising from the superclasses of an instance declaration at Foo.hs:8:0
  10 --    Probable fix: add (C a b') to the instance declaration superclass context
  11 --    In the instance declaration for `C (Maybe a) a'
  12 --
  13 -- Since L is a superclass of the (sought) constraint (C a b'), you might
  14 -- think that we'd generate the superclasses (L a b') and (L a b), and now
  15 -- the fundep will force b=b'.  But GHC is very cautious about generating
  16 -- superclasses when doing context reduction for instance declarations,
  17 -- becasue of the danger of superclass loops.
  18 --
  19 -- So, today, this program fails.  It's trivial to fix by adding a fundep for C
  20 --      class (G a, L a b) => C a b | a -> b
  21
  22 -- Note: Sept 08: when fixing Trac #1470, tc138 started working!
  23 -- This test is a very strange one (fundeps, undecidable instances),
  24 -- so I'm just marking it as "should-succeed".  It's not very clear to
  25 -- me what the "right" answer should be; when we have the type equality
  26 -- story more worked out we might want to think about that.
  27
  28 module ShouldFail where
  29
  30 class G a
  31 class L a b | a -> b
  32 class (G a, L a b) => C a b
  33
  34 instance C a b' => G (Maybe a)
  35 instance C a b  => C (Maybe a) a
  36 instance L (Maybe a) a