No subject

[uunet!research.att.com!dmg]

to- (numeric-interestavalidgh.com)
from- David M. Gay
re- expression evaluation in IEEE arithmetic

Unexpected extra precision is a cause of headaches.
On x86 systems, I routinely use "round to double" mode
to avoid surprises.  And on RS6000 systems, I sometimes
disable fused multiply-add instructions (in a context --
simplifying linear programming problems -- where they
sometimes lead to trouble).

In short, I support Hough's recommendation of
the "traditional Fortran rules" -- use just the precision
implied by the operands.  (Extended precision is a boon
when used just where requested.)

-- dmg