Arrows.hs Measurements

Arrows.hs

{-# LANGUAGE Arrows #-}

import Control.Arrow (returnA)

f :: () -> ()
f =
  proc () ->
    let _ = () in
    let _ = () in
    let _ = () in
    let _ = () in
    -- and then on line 284168
    returnA -< ()

Parsing time

Command:

  $ inplace/bin/ghc-stage2 -ddump-timings ~/Arrows.hs
  
Measurements:

  master
    Parser [Main]: alloc=9560778976 time=7956.168
    Parser [Main]: alloc=9560778976 time=7911.255
    Parser [Main]: alloc=9560778976 time=7978.155
    Parser [Main]: alloc=9560778976 time=8072.056
    Parser [Main]: alloc=9560778976 time=7867.061

  wip/exp-pat-frame (deep embedding)
    Parser [Main]: alloc=9574418648 time=8050.242
    Parser [Main]: alloc=9574418648 time=7928.873
    Parser [Main]: alloc=9574418648 time=7896.294
    Parser [Main]: alloc=9574418648 time=7917.393
    Parser [Main]: alloc=9574418648 time=7840.901

  wip/exp-cmd-frame (shallow embedding, without --coerce)
    Parser [Main]: alloc=9899586040 time=8769.317
    Parser [Main]: alloc=9899586040 time=8843.270
    Parser [Main]: alloc=9899586040 time=8698.141
    Parser [Main]: alloc=9899586040 time=8757.309
    Parser [Main]: alloc=9899586040 time=8771.368
    
  wip/exp-cmd-frame (shallow embedding, with --coerce)
    Parser [Main]: alloc=9706356832 time=7963.211
    Parser [Main]: alloc=9706356832 time=8002.884
    Parser [Main]: alloc=9706356832 time=8085.704
    Parser [Main]: alloc=9706356832 time=8042.423
    Parser [Main]: alloc=9706356832 time=8007.699

NB. Thermal throttling of my CPU seemed to have an effect on the execution time, changing the results by up to 1500 units. The performance numbers were much worse right after a full rebuild (in the ballpark of 9500).

When interleaving tests for master, wip/exp-pat-frame (deep embedding), and wip/exp-cmd-frame (shallow embedding (with --coerce), I observe very similar numbers:

1. master: Parser [Main]: alloc=9560778976 time=7973.300
2. wip/exp-pat-frame: Parser [Main]: alloc=9574418648 time=7878.008
3. wip/exp-cmd-frame: Parser [Main]: alloc=9706356832 time=7957.351

It is therefore safe to conclude that neither of the approaches has any significant impact on parsing performance. Omitting --coerce has some negative performance impact, about 10%.

Using a GADT instead of a CPS encoding in the shallow embedding had no effect: time spent is still approximately equal to that in master:

1. master: Parser [Main]: alloc=9560778976 time=8090.421
2. wip/exp-pat-frame: Parser [Main]: alloc=9574418648 time=7921.118
3. wip/exp-cmd-frame: Parser [Main]: alloc=9679077216 time=8006.478

Allocations are apparently slightly reduced.