[GHC] #11525: Using a dummy typechecker plugin causes an ambiguity check error

[GHC]

#11525: Using a dummy typechecker plugin causes an ambiguity check error
-------------------------------------+-------------------------------------
        Reporter:  jme               |                Owner:
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  8.2.1
       Component:  Compiler (Type    |              Version:  8.0.1-rc1
  checker)                           |             Keywords:
      Resolution:                    |  UndecidableSuperClasses, plugin
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  GHC rejects       |  Unknown/Multiple
  valid program                      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by adamgundry):

 Okay, let me try to summarise:

 1. When a type-checker plugin is in use, we zonk the constraints before
 calling the plugin. Thus plugins don't need to worry about doing their own
 zonking.

 2. Zonking a `CDictCan` constraint (or indeed anything other than a
 `CHoleCan`) turns it into a `CNonCanonical`. Presumably this is because
 zonking may not preserve the invariants of all the canonical constraints,
 so they might need to be re-canonicalized if they are looked at again
 later.

 3. The `expandSuperClasses` step, which runs after the simple wanted and
 plugin loop, looks for `CDictCan`s. If a plugin is in use, 1 and 2 mean
 that it doesn't find any, and hence fails to expand superclasses.  This
 leads to the solving failure reported in this ticket.

 If we can make `zonkCt` preserve `CDictCan` rather than turning it into
 `CNonCanonical`, then `expandSuperClasses` should just work even in the
 presence of plugins. However, it isn't completely obvious to me whether
 the invariants of `CDictCan` (e.g. the fact that `cc_tyargs` is "function-
 free") will still hold after zonking.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11525#comment:6>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler