Better Optimiziers?

[grunwald at uiucdcsb.UUCP]

As I hack away at a modified version of the Berkley Pascal compiler and look at
the kind of code that is emitted by /lib/f1, I keep asking myself "Why didn't
anyone build a better optimizer for the PPCIR?"

   Lets admit it, /lib/c2 gives a marginal improvement for programs generated
by PC and F77. If there was an optimizer pass which did dataflow anaylsis
shared code detecttion, and register assignment on the PCCIR forest of trees,
it would be fairly machine independent, and I think that it would be fairly
efficient too.
   Tweaking every compiler to include an optimizing phase, as was apparently
done with F77, is too expensive. A filter before /lib/f1 is all that's
really needed.

   Typically, C programs are not optimized on anything but the assembler level,
but I think that an optimizer which did the optimizations mentioned would even
be able to improve many a C program as well as PC and F77 programs.

   I'm certain that someone out that must have done this at some time in the
past, particularly since optimizing heavily used utilities (CSH, troff, cc)
might provide a substantial savings of CPU time. Tweaking a utility like
this by hand is usually not worth the effort, but an optimizer should be able
to get at least 5 -> 10% with the minimal effort of recompiling the utilties.

   I heard a that someone at Zilog had worked on something like this, and that
someone at MIT might have something along the lines of what I'm looking for.
Does anyone know if this is true and if any of this is in the public domain?

   What about the Graham-Glanville style code generators written at UCBerkley?
Are they distributed and do they provide any significantly increased
performance? I've only seen the articles in SIGPLAN Trans. on Comp. Const.,
an the figures made it appear that they give higher code-density.

Dirk Grunwald
University of Illinois,
{pur-ee,ihnp4}!uiucdcs!grunwald		UUCP domain
grunwald.uiuc at Csnet-relay.CSNET		Internet domain