[Haskell-cafe] Re: Experiments with defunctionalization, church-encoding and CPS

[Heinrich Apfelmus]

David Menendez wrote:
> On Sun, Nov 1, 2009 at 7:12 AM, Heinrich Apfelmus
> <apfelmus at quantentunnel.de> wrote:
>> Even then, the results are mixed. The Church-encoding shines in GHCi as
>> it should, but loses its advantage when the code is being compiled. I
>> guess we have to look at the core if we want to know what exactly is
>> going on.
> 
> What optimization level did you compile with?

No optimization. From the bottom of the paste

  http://hpaste.org/fastcgi/hpaste.fcgi/view?id=10686#a11420


    Results:

    In ghci, the church encodings clearly win.

        *Main> force testTree
            ()
            (1.63 secs, 13948748 bytes)
        *Main> test findMaybe 0
            100000
            (1.70 secs, 15026356 bytes)
        *Main> test findChurch 0
            100000
            (0.72 secs, 15553668 bytes)
        *Main> test findChurch' 0
            100000
            (0.71 secs, 13456600 bytes)


    But when compiling, the algebraic data types wins.

    ghc --make Test.hs

        data Maybe
            mean: 90.08407 ms, lb 86.36974 ms, ub 94.13646 ms
        Church encoding
            mean: 152.0198 ms, lb 144.0916 ms, ub 161.0063 ms
        Church encoding optimised
            mean: 114.0715 ms, lb 107.3498 ms, ub 122.6881 ms


I am a bit surprised. Then again, I probably shouldn't be surprised that
the cost model is not like I imagine it to be.


Regards,
apfelmus

--
http://apfelmus.nfshost.com