More results and comments about gradual underflow

[Keith Bierman ADT/QED]

>By all means, let us have SEPARATE benchmarks, like the one Vaughn

I strongly disagree. Chip designers are only going to be looking at
the set of "SPEC (all variations), TPC (all variations), etc. They are
certainly not going to make implementation decisions based on a test
of what they will perceive as "corner cases".

That isn't to say that I think benchmark designers should go out of
their way to put extra "exceptional" cases into their codes. However,
I think that working to eliminate them from mainstream benchmarks
merely serves to bias implementations to do even less in hardware.

But clearly opinions vary! ;>