Kahan's Floating Point Test

[dewar@gnat.com]

<<I checked this on a linux box.  The result is due to the pow function
in the system math library.  The function is sensitive to the setting
of the fpu rounding precision.  It is not a compiler bug.
>>

Not surprising x**y is the classical example where extended precision
makes this a trivial calculation (you can compute it using logs and 
exps ONLY if you have extended precision).