ian.piumarta at inria.fr
Tue Feb 18 01:08:57 UTC 2003
On Mon, 17 Feb 2003, Eric Merritt wrote:
> Hello All,
> Just out of curiosity does anyone know where development of the Jitter
> is at?
It's stalled. I wanted make a version that worked on Mac for Dan to play
with and the easiest route appeared to be to make it work on OS X which he
was planning to install. Making j5 work on OS X meant making the Unix VM
work on OS X, but that particular "subproject" got slightly out of hand
and finally turned into a significant part of the hacking I've done during
the last 5 or 6 months.
The current non-optimising compiler generates native code with all the
hooks, profiling, recompilation triggers and support in place to do
adaptive compilation based on runtime type feedback. (I was about to
embark on the first round of adaptive compilation when I became distracted
by OS X.) The emphasis is on reducing compilation time to an absolute
minimum: what's important is generating code for lots of methods quickly
(compilation shouldn't introduce any detectable pauses in execution) but
which includes profiling (to record type, sender and frequency information
to drive an optimising compiler). The real benefits come from
recompilation using this optimising compiler, once runtime type behaviour
is known, whose emphasis is on producing the fastest possible code at the
expense of compilation time. Since this compiler is invoked far less
frequently than the non-optimising one, its longer compilation times
shouldn't introduce detectable pauses either.
The non-optimising compiler already produces stable code (where "stable"
means I don't know how to crash the code by writing bizarre methods,
abusing contexts or exceptions, etc., within the image) for PPC and 386.
The first part of the optimising compiler was the next thing on the to-do
list when OS X came along and preempted it.
> I heard that Ian was working on J5 and I did a quick search of the
> archive and couldn't find anything interesting.
I don't think there's been any discussion of it here.
> I am really just curious about it.
To satisfy your curiosity I checked it out for you here:
Although I don't suggest you try this at home, the above could be dropped
into a 3.2 tree (next to the "platforms" and "src" dirs) and will build a
working j5 (provided the regular VM has already been built in a dir called
"bld" at the same level). It generates code for PPC on Unix and 386 on
Unix or Windows. (It probably works with the 3.5 tree too, modulo the
Quartz code not bothering to look for "j_interpret" before launching the
interpreter, so my objective of making it work on Mac OS might already
have been met. Some trivial additions are required to the win32 tree to
make it work, but they're outside the tree that I checked out. If you're
curiosity is instatiable then I could check out my modified win32 tree
The above directory contains the answers to any and all questions you
might have regarding j5, albeit in an _extremely_ undigested form. ;)
I'm hearing encouraging things about Anthony's latest run at his closure
work. When I get round to picking up j5 from where I left off I'll
probably start by looking at what he's been up to and adapt j5 to work
with it instead of with an unmodified "traditional" image.
More information about the Squeak-dev