Subject: Re: Vortex download note (and some comments)
From: Vassily Litvinov (vass@cs.washington.edu)
Date: Tue Feb 22 2000 - 12:20:34 PST
Andy,
Thank you very much for trying out our system and taking time to learn
Cecil and give us back such insightful comments. I won't be able to
address all of them, just a couple.
> One of the questions I wondered about was the price of Cecil's expressiveness.
> My first test was of course "Hello world!". Well, a 2MB executable for that
> (with noevalstdlib, no interrupt_checking, no debug_support and shared libs)
> is something one has to get used to ...
Yes... Vortex is not tuned for small programs, as we tend to study larger
benchmarks/programs. Good news: compiling hello.world doesn't take a
whole night any more...
> Then I wrote the following micro-benchmark to check the level of optimization
> Vortex can achieve:
This is an excellent micro-benchmark for studying performance problems.
Indeed, so far we have focused more on compiler-style programs where these
kinds of problems are much more difficult to handle (and are responsible
for much smaller slow-down). But Craig Chambers has been working on
representation optimization and can probably give you a better answer when
he gets back from vacation.
The issue of object identity is really one of the big problems when
optimizing object representation. Another issue is the complexity of
analysing the program to understand what optimizations can be legal *and*
profitable. For example, in a language with features like dynamic
dispatch and first-class function objects it is very hard to understand
which objects can appear at which callsites. As you point out, if such
features are not used, it should be possible to optimize better, but
Vortex does not handle this case quite well.
Thank you for the idea of distinguishing between objects supporting and
not supporting object identity; we will think about it.
> p.s.: I tried to subscribe to the Cecil mailing list but I got the
> following error message:
I'll look into it.
Regards,
Vassily
This archive was generated by hypermail 2b25 : Tue Oct 03 2000 - 15:21:22 PDT