[GHC] #15009: Float equalities past local equalities

[GHC]

#15009: Float equalities past local equalities
-------------------------------------+-------------------------------------
        Reporter:  goldfire          |                Owner:  (none)
            Type:  feature request   |               Status:  closed
        Priority:  normal            |            Milestone:  8.4.3
       Component:  Compiler          |              Version:  8.2.2
      Resolution:  fixed             |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:  gadt/T15009
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by simonpj):

 Hmm.  I think you may be suggesting this transformation on constraints:
 {{{
     forall tvs1. blah1 => forall tvs2. blah2 => unsolved-stuff

 -->

     forall (tvs1, tvs2). (blah1, blah2) => unsolved-stuff
 }}}
 The hope would be that an equality in the inner `blah2` might look like
 `(a~ty)`, where
 `a` is bound by the outer `tvs1`.  If so, then in the transformed
 implication,
 the equality would fall under `Note [Let-bound skolems]`, and we could
 float
 things from `unsolved-stuff` out.

 I think that is plausible, and it's an interesting idea that I had not
 onsidered..  The side condition is that there are no "wanteds"
 as a peer to the inner implication, like this
 {{{
     forall tvs1. blah1 => ( forall tvs2. blah2 => unsolved-stuff
                           , alpha ~ Int )
 }}}
 If there was, the argument of comment:11 would apply.

 But what if there were ''two'' such implications?  Floating constraints
 out of
 either would disable floating constraints out of the other!

 This looks complicated and unprincipled to me.  I suggest just using a
 type signature.
 Extremely complicated type inference is not necessarily a boon to the
 programmer!

 Perhaps there is a simple, principled algorithm hiding in there; but I
 don't yet see it.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/15009#comment:13>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler