[GHC] #9117: Coercible constraint solver misses one

[GHC]

#9117: Coercible constraint solver misses one
-------------------------------------+------------------------------------
        Reporter:  goldfire          |            Owner:  nomeata
            Type:  bug               |           Status:  new
        Priority:  normal            |        Milestone:
       Component:  Compiler          |          Version:  7.8.2
      Resolution:                    |         Keywords:
Operating System:  Unknown/Multiple  |     Architecture:  Unknown/Multiple
 Type of failure:  None/Unknown      |       Difficulty:  Unknown
       Test Case:                    |       Blocked By:
        Blocking:                    |  Related Tickets:
-------------------------------------+------------------------------------

Comment (by nomeata):

 In the paper comments you write

 > Suppose module `A` uses `coerce` somewhere and contains a newtype whose
 constructor is not imported. We now modify `A` to import the constructor.
 Is it guaranteed that the use(s) of `coerce` will still work? We would
 want to offer such a guarantee I think.

 We are unable to give this guarantee with either of the two solving
 strategies. Consider:

 {{{#!haskell
 module Foo where
 newtype Rec a = Rec a
 newtype Hide a = Hide (Rec a)

 module Bar where
 import Foo (Hide(Hide), Rec)

 ex :: Rec Age -> Hide Int
 ex = coerce
 }}}

 This works (`Hide` is unrolled, then we can lift through `Rec` and are
 done). If we also import `Rec`, then `Coercible (Rec Age) (Hide Int)`
 (where no lifting is possible yet, so the order does not matter) will
 cause it to loop.

 I couldn’t construct it with a less pathological newtype yet, though, as
 these will cause a proper data tycon to appear, then the newtypes on the
 RHS get unwrapped and eventually the terms on both sides match.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9117#comment:24>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler