[GHC] #9476: Implement late lambda-lifting

[GHC]

#9476: Implement late lambda-lifting
-------------------------------------+-------------------------------------
        Reporter:  simonpj           |                Owner:  sgraf
            Type:  feature request   |               Status:  patch
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  7.8.2
      Resolution:                    |             Keywords:  LateLamLift
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Runtime           |  Unknown/Multiple
  performance bug                    |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:  #8763 #13286      |  Differential Rev(s):  Phab:D5224
       Wiki Page:  LateLamLift       |
-------------------------------------+-------------------------------------

Comment (by sgraf):

 Nevermind, the regression was due to lifting binders occuring in a nullary
 applications. Example from `CS`:

 {{{
 main_$s$wgo =
     \r [sc sc1]
         case sc1 of wild {
           __DEFAULT ->
               let {
                 lvl =
                     \u []
                         case -# [wild 1#] of sat { __DEFAULT ->
 main_$s$wgo sc sat; }; } in
               let {
                 sat =
                     \r [s1] case plusInteger s1 ds of s' { __DEFAULT ->
 lvl s'; };
               } in  sat;
           0# -> sc ();
         };
 ==>
 main_$s$wgo =
     \r [sc sc1]
         case sc1 of wild {
           __DEFAULT ->
               let {
                 lvl =
                     \u []
                         case -# [wild 1#] of sat { __DEFAULT ->
 main_$s$wgo sc sat; };
               } in  $lsat lvl;
           0# -> sc ();
         };
 $lsat =
     \r [lvl s1] case plusInteger s1 ds of s' { __DEFAULT -> lvl s'; };
 }}}

 (`main_$s$wgo` is a ternary function)

 This is a beneficial lift from the perspective of closure growth, but
 actually allocates more because the nullary application `sat` turns into a
 partial application `$lsat lvl` which allocates.

 Fixing this by regarding nullary applications as argument occurrences of
 the binder (e.g. a no-go) has the same effect to nofib benchmarks as just
 disallowing ''any'' undersaturated calls to the binder to lift. The latter
 is much simpler to check for (just look at the `StgBinderInfo` of the RHS)
 and is the status quo in Phab:D5224. So, no news, basically.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9476#comment:47>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler