[Haskell-cafe] -O2 compile option can give speed increase over -O. Fasta shootout program test runs.

[Derek Elkins]

On Tue, 2007-07-17 at 11:24 +1000, Donald Bruce Stewart wrote:
> r.kelsall:
> > I have been playing with the Fasta program in the shootout to see if
> > I can make it umm faster. Starting from dons program on this page and
> > adding some timing calculations as suggested on this wiki page
> > 
> > http://shootout.alioth.debian.org/gp4/benchmark.php?test=fasta〈=ghc&id=2
> > http://www.haskell.org/haskellwiki/Timing_computations
> > 
> > I added different OPTIONS into the top line of the program did a
> > ghc --make fasta.hs   and ran it each time with  fasta 2500000
> > (This is one tenth of the shootout figure.) These runs all keep the
> > existing OPTIONS of  -fbang-patterns -fexcess-precision
> > 
> >   Seconds   OPTIONS Added
> >   -------   -------------
> >    40.5
> >    40.5        -funbox-strict-fields
> >    40.4                      {-# INLINE rand #-}
> >    17.2        -O
> >    17.0        -O  -fvia-C
> >    14.4        -O  -optc-march=pentium4
> >    11.5        -O2
> >    11.2        -O3
> >    11.5        -O3           {-# INLINE rand #-}
> >    11.3        -O2 -optc-march=pentium4
> > 
> > There was a bit of variation, I've averaged over two runs. This is on
> > an Intel Pentium D 2.66GHz running W2K and GHC 6.6.1.
> > 
> > It seems the -O2 option can give a significant speed increase relative
> > to just the -O option. This is contrary to the documentation which says
> > 
> > http://www.haskell.org/ghc/docs/latest/html/users_guide/options-optimise.html
> > http://www.haskell.org/ghc/docs/latest/html/users_guide/faster.html
> > 
> > it won't make any difference. I guess it's program, architecture and
> > operating system specific, but according to these figures the -O2 option
> > seems well worth a try for programs that need speed. It may be that
> > we sacrifice bounds checking or something important with -O2, I don't
> > know.
> 
> Yes, -O2 is getting better, as new optimisations like SpecConstr are
> enabled by it. For shootout problems, I'd selectively test with -O2, and
> if it is better, use that.
> 
> Good work! And yes, I see that it is currently compiled with:
> 
>     -O fbang-patterns -fexcess-precision  -fglasgow-exts  -optc-march=pentium4
> 
> if -O2 is consistently better here, then we could happily switch.

Just to add as this was not addressed. -O2 -does not- turn off bounds
checking or any other obvious safety mechanism.