[GHC] #14808: GHC HEAD regression: GADT constructors no longer quantify tyvars in topological order

[GHC]

#14808: GHC HEAD regression: GADT constructors no longer quantify tyvars in
topological order
-------------------------------------+-------------------------------------
        Reporter:  RyanGlScott       |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  highest           |            Milestone:  8.6.1
       Component:  Compiler (Type    |              Version:  8.5
  checker)                           |
      Resolution:                    |             Keywords:  GADTs
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:  #14529, #14796    |  Differential Rev(s):  Phab:D4413
       Wiki Page:                    |
-------------------------------------+-------------------------------------
Changes (by simonpj):

 * status:  closed => new
 * resolution:  fixed =>


Comment:

 Richard, I'm disturbed by this.

 * The [http://downloads.haskell.org/~ghc/master/users-
 guide/glasgow_exts.html#visible-type-application spec for visible type
 application] says that we quantify implicit variables in left-to-right
 order; but that we perturb that order to respect dependency, speaking of a
 "table topological sort".

 * But I can't find a single comment or Note in the source code to point
 out where these things are implemented.

 * In the case in point for #14808, we have
 [https://phabricator.haskell.org/D4413 this fix] in `RnSource`.  Although
 the comment doesn't say ''why'' the order is important, I infer (from
 detective work) that:
   * `extractHsTysRdrTyVarsDups` guarantees to return variables in left-to-
 right order
   * The `hsq_implicit` field of `LHsQTyVars` is an ''ordered'' list,
 specifically in left to right order.  Can it contain duplicates?  I think
 not, but again that is not duplicated.
   * There is a `toposortTyVars` in `TcImplicitTKBndrsX` which does the
 perturbing to respect dependency.  But again there is zero documentation
 that it must be a stable topo-sort, or why.

 It should not take detective work to infer this info. If my suppositions
 are correct, could you
 * Document the key invariants on data types
 * Add a Note to explain the moving parts and
 * Refer to it from appropriate places.
 E.g. I bet the same ordering constraints apply to `hsib_ars` field of
 `HsImplicitBndrs`?

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14808#comment:4>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler