Re: Optimizer pattern language

[Tom Lord]

       > I think I may have caught the gist of Tom's suggestion.


I think so.  It isn't a very _specific_ suggestion -- it's not like I
have a clear picture of what the result will be.

It's just that:

        1) you can capture the (un)memoization and optimization rules
           at a very high level

        2) you can probably process that very high level notation
           quite effectively

It's something to look into before leaping into, for example, trying
to make up some abstractions for walking memoized trees.

In my experience, this kind of approach is really hard to get started,
but then once you do, it's far more tractable than other approaches.

(One fun, vaguely related example was the original guile pattern
matching macro:  it was written in a subset of its own language, with
a bootstrapping implementation, then recompiled itself.   It's
drawback was that I didn't ever buckle down to make the bootstrapping
happen off-line, so it was very slow on start-up.)

Hmm...maybe I should eventually get around to putting together a
bibliography of relevent materials.

-t