Re: gcj jit, timing comparison

[Thomas Johnsson <thomas at gatespacetelematics dot com>]

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