Re: inlining + optimization = nuisance bugs

Zalman Stern wrote:


> You can say things like "Floating-point is imprecise anyway" and
> "Well if you'd just write robust code, this wouldn't happen" but it's
> just dancing around the fact that the tools make it (very) difficult
> to express a precise algorithm.


Kahan's term for people promoting this point of view ("if you'd just
write robust code") was "the guilt squad". Luddy's counter- example
in favor of inconsistent use of high precision arithmetic is not a
good one, since every numerical analysis course I ever taken explained
the concept of "machine epsilon" and what would happen if you added
two numbers differing by a more extreme ratio. Machine epsilon
figures into proofs of convergence for many numerical algorithms. (I
took these courses as an undergradual and gradual student, and worked
through some of the proofs.)


In contrast, inconsistent use of higher precision (what you can get
with fused-multiply-add) does not. That it is faster, does not make
it better.


David Chase
NaturalBridge LLC
--

Return to the comp.compilers page.
Search the comp.compilers archives again.