[kaffe] JIT3 question.

[Dalibor Topic]

Hi Kiyo,

--- Kiyo Inaba <inaba at src.ricoh.co.jp> wrote:
> As someone may know, I am now trying to make
> super-h's jit3 work.
> While doing this, I go back to check how it was
> organized in m68k.
> Of course, the jit3 port of m68k does not work for
> kjc (this is the
> reason why m68k port still uses 'jit' rather than
> 'jit3'), but it
> can execute so many regression tests.

could you take a look at this patch by Richard
Zidlicky at
http://www.kaffe.org/pipermail/kaffe/2001-December/007541.html
. I don't think it has been merged in. does it improve
things on m68k ?

> After gathering several info from regression test
> results for both jit
> and jit3, I was surprised that 'jit' is faster than
> 'jit3' in m68k!

that's weird. I thought jit3 was the "next generation"
jit, so it was supposed to be better. At least that
should be the case on i386. ;) 

On the other hand, I've recently changed
System.arraycopy to pure java, so that might expose
deficiencies in a jit implementation that weren't so
visible before. You could profile a benchmark run, and
see if your code spends a lot of time in arraycopy,
and check the generated assembly in that case.

> Are there anyone who can remind me the reason why
> JIT3 is introduced?

I've searched the mailing list archives but couldn't
find anything about design differences between the
two, except for the same question being asked by
someone else ;)

cheers,

dalibor topic

__________________________________________________
Do you Yahoo!?
Yahoo! Web Hosting - Let the expert host your site
http://webhosting.yahoo.com