RE: [abc] oopsla outline

Thanks, Eric!

> I would still like to add some information about the spec patterns
> survey by Dwyer et al. if you don't mind. It is really very interesting,
> especially w.r.t. future work (i.e. static optimization) because it
> turns out that by optimizing only a few very common patterns, you can
> already optimize 90% of all the use cases they encountered. So this
> seems quite encouraging.

I think we might cover this in the "future work" on
static analysis, along with similar results of
Dawson Engler et al, Manuvir Das (ESP), SLAM, etc.

To avoid the criticism "this is an unfinished piece of work",
the paper should stress that we are after cheap techniques
that give reasonable performance *without* whole-program
analysis. It's dubious when programming language features
rely on complex analyses to be tractable, as that
makes their use very brittle (a small change in the
base program could suddenly make your system run 50 times
slower). So here we focus on just carefully analysing the
tracematch.

> I have now a version of Alpha which compiles and also kind of runs. I am
> just not 100% sure how to correctly run the examples they provide. I
> will try to contact Christoph Bockisch about that.

That would be superb! Do you think the safeenum example is
expressible in Alpha? [I don't even know whether it can instrument
things like jhotdraw]

> What's weka?

A base program we've used for benchmarks in our pldi paper.
I seem to remember Laurie was trying some tracematches on it,
but I don't know what the outcome was, or indeed whether
abc could process it at all. You could ask her about it?

> tracematch would (under certain circumstances). Stangely after a long
> time of debugging it turned out that the ReferenceMap implementation of
> Jakarta Commons Collections was the problem! I would never have thought
> that this class has a bug but it quite seems so: Reproducably, entries
> were deleted too early. I then replaced the map by the WeakHashMap of

Urk. The new indexing scheme uses apache commons weak identity hashmaps.
Do you have a small example demonstrating the bug?

> expression is so large. So I guess I will have to verify this though
> further runs... I am confident that we get reliable numbers within the
> next day or two. But anyway it was quite difficult to get the native AJ
> aspect right. So at least in terms of usability and maintainability, TMs
> win here quite clearly.

Hm, that would be nice!

-Oege
Received on Thu Mar 02 09:46:04 2006