Re: [GHC] #10148: Optimization causes repeated computation

16 Mar 2015

      #10148: Optimization causes repeated computation
-------------------------------------+-------------------------------------
        Reporter:  akio              |                   Owner:
            Type:  bug               |                  Status:  new
        Priority:  normal            |               Milestone:
       Component:  Compiler          |                 Version:  7.10.1-rc1
      Resolution:                    |                Keywords:
Operating System:  Unknown/Multiple  |            Architecture:
 Type of failure:  Runtime           |  Unknown/Multiple
  performance bug                    |               Test Case:
      Blocked By:                    |                Blocking:
 Related Tickets:                    |  Differential Revisions:
-------------------------------------+-------------------------------------

Comment (by akio):

 It looks like the simplifier does not preserve correctness of demand
 information.

 After the worker-wrapper pass, I have this fragment of code

 {{{
             of m'_axp [Dmd=<L,U(U(U))>] { Array ipv_s1i9 [Dmd=<L,A>] ->
             Main.Machine (gstep m'_axp) t_axr
             }
 }}}

 Here, the demand info on `ipv_s1i9` is correct, it is indeed unused.

 However, after one iteration of simplifier pass, it becomes:

 {{{
     of m'_axp [Dmd=<L,U(U(U))>] { Array ipv_s1i9 [Dmd=<L,A>] ->
     (# \ (w_s3Ix :: Int) ->
          case w_s3Ix of _ [Occ=Dead] { GHC.Types.I# ww_X3IZ ->
          case $wgstep_s3II ipv_s1i9 ww_X3IZ
          of _ [Occ=Dead] { (# ww_X3Ja, ww_X3Jc #) ->
          Main.Machine ww_X3Ja ww_X3Jc
          }
          },
        ww_s3IA #)
 }}}

 Now `ipv_s1i9` is referenced and will be used, but its usage information
 has not been updated.

 I've found the note `[Case alternative occ info]` in Simplify.hs, which
 explains why the occurrence info on case-bound variables should be zapped.
 I tried zapping the usage info as well as the occurrence info here (patch
 attached). It seems to compile `repeated2.hs` fine, but I'm not sure about
 a few points:

 * Is the approach right? i.e. is it a good idea to manually modify the
 usage info in the simplifier?
 * Probably the patch is overly conservative, in that it zaps the
 strictness info as well as the usage info. However, how hard should I try
 to preserve information? For example, should the code have a special case
 for when the case binder and the scrutinee are both marked unused?

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10148#comment:8>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler