Re: [GHC] #15318: Core Lint error involving newtype family instances with wrappers

28 Jun 2018

      #15318: Core Lint error involving newtype family instances with wrappers
-------------------------------------+-------------------------------------
        Reporter:  RyanGlScott       |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  8.6.1
       Component:  Compiler (Type    |              Version:  8.4.3
  checker)                           |
      Resolution:                    |             Keywords:  TypeFamilies
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Compile-time      |  Unknown/Multiple
  crash or panic                     |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):  Phab:D4902
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by simonpj):

 Yes, Phab:D4902 will fix the bug, but it perpetuates a mistake!

 Consider
 {{{
   data family S a
   data family T a

   data    instance S [a] = MkS a
   newtype instance T [a] = MkT a
 }}}
 Currently, we get this:
 {{{
   data SList a = MkS a
   axiom coS a :: SList a ~ S [a]

   -- Wrapper for MkS
   $WMkS :: a -> S [a]
   $WMkS x = MkS x |> coS a

   -- newtype TList a = MkT a
   axiom coTList a :: a ~ TList a
   axiom coT a :: TList a ~ T [a]

   -- Worker for MkT
   MkT :: a -> T [a]
   MkT x = x |> coTList a |> coT a
 }}}
 Notice the inconsistency: the cast that takes us from the
 representation type to the final user type is done in
 the ''wrapper'' for data types, but in the ''worker''
 for a newtype.  (Reminder: for data types the worker isn't
 an executable function, it's the Core data constructor; but
 for a newtype the "worker" is an executable function that
 expands to a cast.)

 This inconsistency shows up in the `MkId` functions `wrapNewTypeBody` and
 `unwrapNewTypeBody`.
 The former wraps two casts (as above) while the latter unwraps only one!

 I think we can readily remove the inconsistency:

 * Don't export `wrapNewTypeBody` from `MkId`; it is not used outside.

 * Remove the extra cast from `wrapNewTypeBody`

 * Do not make the change to `mkDataConRep` proposed in Phab:D4902.

 * Make a wrapper for family-instance newtypes.  Do this by making
   `wwrapper_reqd` return `True` for ''all'' types for which
   `isFamInstTyCon` holds. (Currently that predicate is not tested for
   newtypes.)

 I think that might be it.  It solves the problem at source,
 produces less code and less to explain (because it's consistent).

-- 
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/15318#comment:2
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler