[GHC] #4965: 60% performance regression in continuation-heavy code between 6.12 and 7

[GHC]

#4965: 60% performance regression in continuation-heavy code between 6.12 and 7
---------------------------------+------------------------------------------
    Reporter:  bos               |       Owner:                         
        Type:  bug               |      Status:  new                    
    Priority:  normal            |   Component:  Compiler               
     Version:  7.0.1             |    Keywords:                         
    Testcase:                    |   Blockedby:                         
          Os:  Unknown/Multiple  |    Blocking:                         
Architecture:  Unknown/Multiple  |     Failure:  Runtime performance bug
---------------------------------+------------------------------------------
 I have a JSON parsing library whose performance has regressed by a little
 over 60% between 6.12.3 and the current head of the 7.0 branch.

 The library uses my attoparsec parsing library, which is written in a
 heavily continuation-based style.

 == Building the code ==

 Reproducing the problem is pretty easy.

 You need two git repos:

 {{{
 git clone git://github.com/bos/attoparsec
 git clone git://github.com/mailrank/aeson
 }}}

 They're easily built:

 {{{
 (cd attoparsec && cabal install)
 (cd aeson && cabal install)
 }}}

 == Measuring performance ==

 Inside the aeson repo are some benchmarks, of which {{{AesonParse}}} is
 the only interesting one for our purposes.

 {{{
 cd aeson/benchmarks && make AesonParse
 }}}

 There's also prepackaged test data, so you can run a benchmark on the
 spot.

 {{{
 ./AesonParse 1000 json-data/twitter100.json
 }}}

 This program accepts a number of loop iterations as its first argument,
 and a data file as its second. It prints its execution time, so a smaller
 number indicates better performance.

 == Performance with 6.12.3 ==

 Here's what I see with 6.12.3, on my MacBook Pro running OS X 10.6.5.

 {{{
 $ ./AesonParse 1000 json-data/twitter100.json
 json-data/twitter100.json:
   1000 good, 1.847631s
 }}}

 This number is good - it's about as fast as Python's C-based json parser.

 == Performance with GHC 7 ==

 You can pass a {{{ghc}}} argument to {{{make}}} to build the benchmark
 with a different compiler.

 {{{
 make clean
 make AesonParse ghc=/usr/local/ghc/7/bin/ghc
 }}}

 And with the benchmark rebuilt, here's the performance:

 {{{
 $ ./AesonParse7 1000 json-data/twitter100.json
 json-data/twitter100.json:
   1000 good, 2.939818s
 }}}

 This is one of the very few instances I've seen of a performance
 regression with GHC 7, and unfortunately it's very large.

 Please let me know what other information I can provide.

-- 
Ticket URL: <http://hackage.haskell.org/trac/ghc/ticket/4965>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler