On Wed, Feb 13, 2019 at 10:51:07AM -0800, Steve Kargl wrote:
> To get Richard's emin/emax patch into the tree so adequate
> testing can be done prior to the 9.1 release, I'll suggest
> that we XFAIL norm2_3.f90 while we work out the details for
> norm2.
>
> Yesteraday, I looked at Jakub's patch, and thought the approach
> was fine. I haven't had time to look at the library side.
> Perhaps, we can adopt Jakub's approach for the library as well.
Yeah. I guess fastest would be to precompute as constants the
two boundaries when we might want to scale something (of course different
for each kind), then just compare the operand and/or result against
those and only if they are above that do frexp* to determine exponent,
(re)compute the scale variable (scalbn* ?) and then just multiply or divide
by that (if division is much slower, we could have two scale variables, one
for scaling down and one for scaling up).
Jakub
