gcj jit, timing comparison

Mon Jun 5 12:18:00 GMT 2006

Andrew Haley wrote:
> Thomas Johnsson writes:
> > 
> > I'm toying with gcj jit, ie running gcj as a jit:er for gij -- great stuff!
> > For embedded caffeinemark it is called like this:
> > 
> > $GIJ -Dgnu.gcj.jit.compiler=$GCJ -Dgnu.gcj.jit.cachedir=${CACHEDIR} 
> > -Dgnu.gcj.jit.options=$OPT CaffeineMarkEmbeddedApp
> > (OPT=-O2)
> > 
> > I've run this both for gcj/gij 4.0.2 and 4.1.0, and statically compiled 
> > gcj, with the following numbers:
> > 
> > gij/gcj 'jit' 4.0.2: 39000-39300
> > gij/gcj 'jit' 4.1.0: 44700-44900 
> > gcj -O2 static 4.1.0: 56700-57200
> > 
> > Ie, there's an improvement by about 14% 4.0.2->4.1.0, presumably due to 
> > better code from gcc/gcj. (?)
> > 
> > And now to my question: where does the difference 'jit' vs static come from?
>> Indirect dispatch, mostly.
>> > It is interesting to look at the cache after the run: The .so files are 
> > about 10 times bigger than the .class files.
>> Try using `strip' on those files. They'll still be bigger than the
> .class files, though: .class is a very efficient encoding.
>
Seems I was overly pessimistic, a more typical factor for larger 
programs (based on du -ck ), seems to be 5-6, or 3-4 stripped (depending 
on the sizes of classes people are in th habit of writing)
-- Thomas