[GHC] #15151: Better Interaction Between Specialization and GND

[GHC]

#15151: Better Interaction Between Specialization and GND
-------------------------------------+-------------------------------------
        Reporter:  andrewthad        |                Owner:  (none)
            Type:  feature request   |               Status:  new
        Priority:  normal            |            Milestone:  Research
                                     |  needed
       Component:  Compiler          |              Version:  8.4.2
      Resolution:                    |             Keywords:  deriving
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------
Changes (by RyanGlScott):

 * status:  infoneeded => new
 * milestone:  8.6.1 => Research needed


Comment:

 OK. I think I'm inclined to agree with Simon here that isn't at all simple
 to achieve. GHC doesn't track whether an instance was derived or not (let
 alone whether it was GND'd or not), and moreover, it's questionable that
 we'd even want to do this. After all, why should a GND'd instance receive
 special privileges? If I defined this instance manually:

 {{{#!hs
 instance Eq PersonId where
   (==) = coerce @Int @PersonId (==)
 }}}

 Then why shouldn't GHC be able to recognize that that instance has the
 same representation the `Eq Int` instance?

 A similar problem arises when you try to `coerce` from a `Map Int ()` to a
 `Map PersonId ()`. The nominal role on `Map`'s first type argument
 prevents this from typechecking, which is rather unsatisfying, since the
 `Ord PersonId` instance has the same representation as the `Ord Int`
 instance. Again, folks have asked if there is a way we could carve out an
 exception here to allow this to typecheck if `Ord PersonId` was GND'd, but
 I have similar reservations about that idea as I do the one proposed in
 this ticket.

 In short, it feels like there ought to be a more general guiding principle
 here to determine whether two instance dictionaries are representationally
 equivalent. I'm not sure what that would be, though.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/15151#comment:6>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler