heuristic optimizations [was "Death by error checks."]

hbaker@netcom.com (Henry Baker) writes:
|This is why a lot of the 'heuristic' optimizations found in compilers
|work mainly for the standard benchmarks and some of the compiler
|vendor's internal code, (or for some of the compiler vendor's most
|favored customers!), but are pretty much a waste of effort for nearly
|every one else.


I agree that there are many optimizations based on heuristics, and
that every heuristic will probably break down in certain situations.
I think it would probably be a mistake to do away with all
heuristic-based optimizations, however. A good global register
allocator would be one example of a optimization that I'd rather not
do without. Sure, it's based on heuristics, but the benefits greatly
outweigh the drawbacks (at least for most of the code I write).


I think that the reason why some optimizations fail so spectacularly
is that it's very difficult to develop "good" heuristics for them.
Loop unrolling is one example that springs to mind. A good heuristic
to determine how much and when to unroll is incredibly difficult to
develop. This doesn't seem to stop compiler writers, however, since as
you point out, even a sloppy implemenation will probably improve
performance for the SPEC benchmarks.


Nat
--
Nathaniel McIntosh | Department of Computer Science
mcintosh@cs.rice.edu | Rice University, Mail Stop 41
n.mcintosh@ieee.org | 6100 South Main, Houston, TX 77005 USA


--

Return to the comp.compilers page.
Search the comp.compilers archives again.