[GHC] #9476: Implement late lambda-lifting

[GHC]

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

Comment (by sgraf):

 I reimplemented this as an STG pass and just rebased it onto a recent 8.6
 alpha commit. You can track the progress on my
 [https://github.com/sgraf812/ghc/tree/22ce14be00f73b63c24e584dda1ad5e1252e57bd
 `lift-lambda` branch, of which this is the current HEAD]. It passes
 `./validate`, but due to nofib's `lambda` benchmark not made fit for the
 next step of the MonadFail proposal, we can't run benchmarks.

 I ran some benchmarks before the rebase, which was relative to
 [https://github.com/ghc/ghc/tree/bb539cfe335e this commit]. Here are the
 numbers, eliding some uninteresting ones:

 {{{
 --------------------------------------------------------------------------------
         Program         Allocs    Instrs
 --------------------------------------------------------------------------------
             VSD          -0.3%     +0.0%
             VSM           0.0%     -0.1%
            atom          -0.8%     -1.2%
    binary-trees          -0.0%     -1.5%
         circsim          -1.0%     -0.6%
        clausify          -1.9%     -0.2%
     constraints          -0.9%     -0.7%
    cryptarithm1          -2.8%     -8.0%
    cryptarithm2          -4.0%     -2.9%
     exact-reals          -2.1%     -0.1%
            fft2          -1.0%     -0.4%
           fluid          -0.9%     -0.6%
          hidden          -1.0%     -0.6%
    k-nucleotide          -0.0%     +2.4%
            lcss          -0.1%     -1.3%
            mate          -3.2%     -0.6%
         mkhprog          -1.3%     +1.0%
           power          -0.7%     -1.0%
          puzzle          -0.0%     +0.0%
          queens         -17.7%     -0.9%
         reptile          -0.3%    -17.1%
         rewrite          -1.7%     -0.1%
       typecheck          -2.7%     -1.9%
 --------------------------------------------------------------------------------
             Min         -17.7%    -17.1%
             Max           0.0%     +4.2%
  Geometric Mean          -0.5%     -0.3%
 }}}

 So, no regressions in allocation (which I'd consider a bug) and some small
 gains in general. Note that because of ticket:15333#comment:2 I'm not
 really trusting many of the regressions in counted instructions. I'd test
 with the proposed `-fllvm -optlo -Os` configuration, but that will have to
 wait until nofib is fixed.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9476#comment:19>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler