Weird fundep behavior (with -fallow-undecidable-instances)

I encounter a strange behavior with functional dependencies. Consider this program

class Foo x y | x -> y where
 foo :: x -> y

class Bar x y where
 bar :: x -> y -> Int

instance (Foo x y, Bar y z) => Bar x z where
 bar x z = bar (foo x) z

Compiling (with 6.4.2, -fallow-undecidable-instances and -fglasgow-exts) I get the following error message:

Foo.hs:12:0:
    Context reduction stack overflow; size = 21
    Use -fcontext-stack20 to increase stack size to (e.g.) 20
        `$dBar :: Bar y z' arising from use of `bar' at Foo.hs:13:11-13
        [... same thing 20 times ...]
        `$dBar :: Bar y z' arising from use of `bar' at Foo.hs:13:11-13
        `bar :: {bar at [y z]}' arising from use of `bar' at Foo.hs:13:11-13
    When trying to generalise the type inferred for `bar'
      Signature type:     forall x y z. (Foo x y, Bar y z) => x -> z -> Int
      Type to generalise: x -> z -> Int
    In the instance declaration for `Bar x z'

The declaration requires undecidable instances, but I doubt that the problem comes from that. What makes it even more weird is that I can get this to compile, and behave as expected, if I do one of a) declare an instance of Bar for any type, or b) add an explicit type signature (foo x :: y) in the definition of Bar. The first seems weird since how could a different, unrelated instance affect the typeability of the second instance? The second, b), is weird since by the FD x -> y we should already know that foo x :: y.

Same behavior in GHC 6.4.1. Hugs (with -98 +O) accepts the code.