costs of avoiding extended-precision register allocation problems

[David Hough]

Turns out my memory of this stuff after five years was somewhat defective.
 From the SunOS 3.2 Floating-Point Programmer's Guide,
for a Sun-3/160, compiling -O -f68881, double precision
linpack ran

100 KFLOPS		-O -f68881
 90 KFLOPS		-O -f68881 -fstore
 82 KFLOPS		-O -f68881          setting double-precision rounding
					    in fpsr

So the penalties weren't too high for that system.  They might be more
noticeable on a 3/470 with 33 MHz 68030/68882, but I'll leave that
investigation to somebody with a pressing interest in it.