More results and comments about gradual underflow

[Keith Bierman ADT/QED]

>The moral for benchmarkers is: you need to try multiple optimization

Ah, that's a moral that is self-fulfilling. That is, if benchmarks
don't include such things, designers will make them run slower.

Benchmarks shouldn't be purely a measure of flat out speed. To use the
hack automobile analogy, that would be like measuring car performance
completely by drag-races (exceptions being like turns, they aren't
"normal" but they aren't all that unusual in real workloads either).

If the folks constructing benchmarks come up with reasonable
approximations to real workloads, then processor architects are more
likely to make good choices.