Notes on Expression Evaluation and Exceptions in SPARC

[Larry Yang]

>3) If the loop
>	s = sum.i x(i)*y(i)
>can be accumulated in higher precision, even 1.5X slower than in working
>precision, it (and similar loops) should be compiled that way by default.

I don't remember the issues exactly, but I recall learning in the famous
Floating-Point Indoctrination Lectures that one of the problems with
the x87 and the 68881/2 floating-point processors was that computations
were performed in 80-bit extended internally, then rounded when the final
result was to be delivered to a software-visible location. Doesn't the
recommendation to accumulate in higher precision suffer from the same
problems?

Again, this issue seems to fall in the domain of "more accurate vs. 
more consistent."

-- Larry