extended precision hazards

[Jerome T. Coonen]

Samuel A. Figueroa wrote (regarding performance of numerical Java code
on machines other than orthogonal float/double architectures):
> 
> The 1.5X-10X range (I think it's closer to 10X myself, at least for double
> precision) takes into account the fact that on an x86...

I included the 1.5X in my recent posting because that's what some
practitioners have reported as the degradation suffered on PowerPC
in certain loops when use of fused mul-add is disallowed.

-Jerome Coonen