[Cfp-interest] exceptions and flags

[David Hough CFP]

> What user problems would arise if some implementations detected underflow exceptions as tiny and inexact?


A more suggestive way to pose the question is


 What user problems would arise if some implementations missed exact underflow
 exceptions?


In order for subsequent scaling to work as expected, an algorithm might
depend on the normal case producing only normalized data, relying on catching
overflow and underflow exceptions to go into more complicated processing
in the exceptional case.    That's why 754-1985 defines underflow to trap
whether or not exact.

I think I objected both during 1985 and 2008 development, to no avail.
It's the way it is on purpose, not by accident, whether or not we agree
that the purpose is convincing.