[GHC] #14111: strange error when using data families with levity polymorphism and unboxed sums and data families

[GHC]

#14111: strange error when using data families with levity polymorphism and unboxed
sums and data families
-------------------------------------+-------------------------------------
        Reporter:  carter            |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  8.2.1
      Resolution:                    |             Keywords:  TypeFamilies
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:  13737             |             Blocking:
 Related Tickets:  #14457            |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by goldfire):

 I think we have a way forward here, actually.

 First off, the problem isn't the main type-checker knot. The problem is
 the littler knot in `tcDataDefn` which retrieves the `tycon` from the
 future in order to, well, build the tycon. This tycon is used in three
 places:

 1. In the `res_ty` passed to `tcConDecls`. This `res_ty` is, in turn, used
 in two places:
  a. In the call to `buildDataCon`, where it is used to set up the
 `dcOrigResTy` field of a `DataCon`.
  b. To do the result-type matching in `rejigConRes`.
 2. As the `TyCon` passed to `tcConDecls`. That is used only when building
 the `DataCon`, so that a `DataCon` knows what `TyCon` it has come from.
 3. In `mkNewTyConRhs`, so that a newtype `CoAxiom` knows what `TyCon` it
 has come from.

 Usages 1a, 2, and 3 absolutely need the final, correct, fully zonked
 `TyCon`. However, usage 1b does ''not''. Conveniently, usage 1b is the one
 causing us trouble.

 The solution is to use `kcLookupTcTyCon` to get the `TcTyCon` of interest
 and build the `res_ty` with un-knot-tied types. Pass this (as a new
 parameter) to `tcConDecls`. And away you go!

 Remaining to do:
  1. Send all the stuff in `TyCon`s through Core Lint. I'm pretty sure we
 have some bits and pieces that are not fully zonked sneaking into
 `TyCon`s. For example, the `final_bndrs` in `tcDataDefn` look not to have
 ever been zonked (by `zonkTcTypeToType` or a friend).
  2. Fix the errors that arise in (1).

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14111#comment:20>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler