llvm dev - Apr 2011 - [LLVMdev] -fplugin-arg-dragonegg-enable-gcc-optzns impact

Now that dragoneegg is robust in its default usage and the dragonegg svn
is moderately stable with -fplugin-arg-dragonegg-enable-gcc-optzns, it is
possible to gauge the impact of that feature. Comparing clang 2.9, FSF gcc
4.5.3svn,
FSF gcc 4.6.0 and dragonegg svn with FSF gcc 4.5.3svn using the himenoBMTxpa
benchmark,
the enhancement to code performance from
-fplugin-arg-dragonegg-enable-gcc-optzns is
clear on x86_64-apple-darwin10 for -fomit-frame-pointer -O3 -ffast-math
-funroll-loops...

compiler                             MFLOPS
clang 2.9                           230.529385
gcc 4.5.3                           246.064891
de-gcc 4.5.3                        204.845864
de-gcc 4.5.3 optzns                 259.672855
gcc 4.6.0                           260.344783

Pretty impressive.
          Jack

Hi Jack,
>    Now that dragoneegg is robust in its default usage and the dragonegg svn
> is moderately stable with -fplugin-arg-dragonegg-enable-gcc-optzns, it is
> possible to gauge the impact of that feature. Comparing clang 2.9, FSF gcc
4.5.3svn,
> FSF gcc 4.6.0 and dragonegg svn with FSF gcc 4.5.3svn using the
himenoBMTxpa benchmark,
> the enhancement to code performance from
-fplugin-arg-dragonegg-enable-gcc-optzns is
> clear on x86_64-apple-darwin10 for -fomit-frame-pointer -O3 -ffast-math
-funroll-loops...
>
> compiler                             MFLOPS
> clang 2.9                           230.529385
> gcc 4.5.3                           246.064891
> de-gcc 4.5.3                        204.845864
> de-gcc 4.5.3 optzns                 259.672855
> gcc 4.6.0                           260.344783
>
> Pretty impressive.
interesting results.  It needs some analysis to work out where the extra juice
is coming from though, in particular to distinguish between effects coming from
LLVM's IR level optimizations and those coming from its code generators.

Ciao, Duncan.

On Fri, Apr 15, 2011 at 08:53:19AM +0200, Duncan Sands
wrote:
> Hi Jack,
> 
> >    Now that dragoneegg is robust in its default usage and the
dragonegg svn
> > is moderately stable with -fplugin-arg-dragonegg-enable-gcc-optzns, it
is
> > possible to gauge the impact of that feature. Comparing clang 2.9, FSF
gcc 4.5.3svn,
> > FSF gcc 4.6.0 and dragonegg svn with FSF gcc 4.5.3svn using the
himenoBMTxpa benchmark,
> > the enhancement to code performance from
-fplugin-arg-dragonegg-enable-gcc-optzns is
> > clear on x86_64-apple-darwin10 for -fomit-frame-pointer -O3
-ffast-math -funroll-loops...
> >
> > compiler                             MFLOPS
> > clang 2.9                           230.529385
> > gcc 4.5.3                           246.064891
> > de-gcc 4.5.3                        204.845864
> > de-gcc 4.5.3 optzns                 259.672855
> > gcc 4.6.0                           260.344783
> >
> > Pretty impressive.
> 
> interesting results.  It needs some analysis to work out where the extra
juice
> is coming from though, in particular to distinguish between effects coming
from
> LLVM's IR level optimizations and those coming from its code
generators.
Duncan,
   This gets even more interesting.

de-gcc45 -fomit-frame-pointer -O3                                           
205.738266
de-gcc45 -fomit-frame-pointer -O3 -fplugin-arg-dragonegg-enable-gcc-optzns  
267.066124
de-gcc45 -fomit-frame-pointer -O2                                           
206.015974
de-gcc45 -fomit-frame-pointer -O2 -fplugin-arg-dragonegg-enable-gcc-optzns  
276.676232
gcc-fsf-4.5 -fomit-frame-pointer -O2                                        
239.868551
gcc-fsf-4.6 -fomit-frame-pointer -O2                                        
248.147753
llvm-clang -fomit-frame-pointer -O2                                         
226.756189

So the enhancement from -fplugin-arg-dragonegg-enable-gcc-optzns doesn't
appear due
to optimizations added between -O2 and -O3. It also is interesting that -O2
outperforms
-O3 with -fplugin-arg-dragonegg-enable-gcc-optzns. I assume that the llvm and
FSF gcc
optimizations must be at cross purposes somewhere there.
              Jack
> 
> Ciao, Duncan.
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev