[GHC] #13221: OccurAnal fails to rediscover join points

[GHC]

#13221: OccurAnal fails to rediscover join points
-------------------------------------+-------------------------------------
        Reporter:  lukemaurer        |                Owner:  lukemaurer
            Type:  bug               |               Status:  patch
        Priority:  high              |            Milestone:  8.2.1
       Component:  Compiler          |              Version:  8.1
      Resolution:                    |             Keywords:  JoinPoints
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):  D3080
       Wiki Page:                    |
-------------------------------------+-------------------------------------
Changes (by lukemaurer):

 * status:  new => patch
 * differential:   => D3080


Comment:

 Found two causes:

 * In several files, the problem was a mistake with occurrence-analyzing
 rules:

 {{{
 let $sj = \y ys -> ...
     {-# RULES "SC:j" forall y ys. j (y:ys) = $sj y ys #-}
     j = \xs -> ...
     in ...
 }}}

   Here `j` can be a join point of join arity 1. Since its rule matches
 exactly one argument (same as the join arity), the body of its RHS can
 contain tail calls (see Note [Rules and join points] in OccurAnal). Thus
 `$sj` can also be a join point, since its only occurrence is a tail call
 with two arguments.

   But by mistake, OccurAnal is counting the binders of the rule, not the
 arguments on its RHS; since there are two binders and `j`'s join arity is
 1, `$sj` gets rejected as a potential join point.

   The warning appears because in the case that `j` is a join point to
 begin with and SpecConstr specializes it, `$sj` will be created as a join
 point, so if OccurAnal doesn't think it can be a join point it issues the
 warning.

 * In T7796, the problem is that we're zapping too much when looking at
 stable unfoldings; we should be allowing unfoldings for join points to
 have tail calls.

   {{{
   let j1 = \x -> ...
       j2 [Unf=\y -> j1 y]
       j2 = \y -> j1 y
   in ... j2 a {- tail call -} ...
   }}}

   Here both `j1` and `j2` should be made join points, but to know that we
 have to see that the call to `j1` from the unfolding of `j2` is a tail
 call. Again, the warning arises from a join point getting specialized; if
 `j1` is a specialization of `j2`, failing to discover `j1` as a join point
 produces the warning.

 Submitted a patch to Phab. This doesn't fix //all// the instances of the
 warning, but it fixes the noisy ones that show up in test cases. (The
 remaining source of noise is beta redexes arising from the simple
 optimizer; that's an ongoing issue. Fortunately, when that causes a
 warning, it'll only show up once before the simplifier gets rid of the
 beta redex.)

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13221#comment:3>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler