Re: [GHC] #5113: Huge performance regression of 7.0.2, 7.0.3 and HEAD over 7.0.1 and 6.12 (MonoLocalBinds)

11 Feb 2013


      #5113: Huge performance regression of 7.0.2, 7.0.3 and HEAD over 7.0.1 and 6.12
(MonoLocalBinds)
--------------------------------------+-------------------------------------
  Reporter:  daniel.is.fischer        |          Owner:                             
      Type:  bug                      |         Status:  closed                     
  Priority:  normal                   |      Milestone:                             
 Component:  Compiler                 |        Version:  7.0.3                      
Resolution:  fixed                    |       Keywords:  performance, MonoLocalBinds
        Os:  Linux                    |   Architecture:  x86                        
   Failure:  Runtime performance bug  |     Difficulty:  Unknown                    
  Testcase:  perf/should_run/T5113    |      Blockedby:                             
  Blocking:                           |        Related:                             
--------------------------------------+-------------------------------------
Changes (by simonpj):

  * difficulty:  => Unknown


Comment:

 This patch is the crucial one that fixes the original problem
 {{{
 commit b5c18c91da911a7729563207c7b95f7e452cca7e
 Author: Simon Peyton Jones 
 Date:   Fri Feb 8 17:29:40 2013 +0000

     Fix an old and egregious specialisation bug (Trac #5113)

     The specialiser needs to know if a dictionay has some structure,
     so that it can decide whether to specialise a function. Eg
      (A)    let d = $dfblah d1
             in ....(f d)....

      (B)    \d. ....(f d)....

     In (A) it's probably worth specialising f; in (B) it isn't.
     Previously we were relying on d's unfolding, but the specialiser
     does cloning as it goes, which discards the unfolding. So we
     were simply discarding all specialisations for functions with
     local dictionary bindings!  This bug seems to have been there
     for a long time.

     This is what originally caused Trac #5113.  Then we went through a
 phase
     where local bindings were not generalised, and that meant there was
     no locally overloaded f to specialise; so the performance problem
 appeared
     to be fixed.  But now we are generalising local bindings again, so it
     re-appeared.

     This patch fixes the original problem.

  compiler/specialise/Specialise.lhs |  390
 ++++++++++++++++++++----------------
  1 files changed, 214 insertions(+), 176 deletions(-)
 }}}

-- 
Ticket URL: http://hackage.haskell.org/trac/ghc/ticket/5113#comment:6
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler