Benchmarks that test only speed on a fixed set of problems encourages what amounts to trying to psych-out the test, cram for the exam, etc.
Tune the compiler, control the test set, etc. (Kahan's ``Baleful'' paper)..
where to put this section in??
