[GHC] #9476: Implement late lambda-lifting

[GHC]

#9476: Implement late lambda-lifting
-------------------------------------+-------------------------------------
        Reporter:  simonpj           |                Owner:  sgraf
            Type:  feature request   |               Status:  closed
        Priority:  normal            |            Milestone:  8.8.1
       Component:  Compiler          |              Version:  7.8.2
      Resolution:  fixed             |             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):

 Replying to [comment:67 simonpj]:
 > I think there is an option to force GC to happen more frequently, but I
 can't see it in the manual.  Simon Marlow might know.

 There's `-i`, but that only seems to work in conjunction with `-h`:

 {{{
    -h       Heap residency profile (output file <program>.hp)
    -i<sec>  Time between heap profile samples (seconds, default: 0.1)
 }}}

 Also, even with `-i0.0001s -h` the sampled maximum residency doesn't come
 close to the 192/196MB above.

 > In any case, getting an accurate residency profile, by using one
 generation and frequent gcs, is a Good Thing.  when you do that, to the
 residency profiles of the two versions differ?  I didn't understand the
 figures you show above.

 Sorry for causing you trouble to understand what I wrote, I really have to
 figure out better ways to present my results.

 So, yeah, using one generation and 'just the right GC frequency' is
 basically what I did in comment:65. I defined 'just the right frequency'
 by choosing the `-A$iM` setting (which determines the points at which GC
 happens) for which maximum residency is at its maximum for each of the two
 candidates.

 For the `default`, which looks at closure growth, this setting was `-A56M`
 with a maximum residency of 192MB. For the version where we also lift said
 `go` function (which we normally would not), maximum residency was at
 196MB for `-A76M`. That's worse than the baseline, as I initially
 expected. The fuzz above was all for nothing, because I once again merely
 measured the RTS parameterisation.

 In comment:66 I remembered that despite better runtime measurements with
 the baseline, counted instructions regressed when 'turning off' GC with
 `-A128M`. That's strange, but I'll probably refer to runtime measurements
 in the paper now and ignore counted instructions completely.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9476#comment:68>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler