I just came across this.  It was originally posted to comp.compilers.

-BobC


From: "Jason Riedy" <ejr@cs.berkeley.edu>
Subject: 754 committee looking for uses of signaling NaNs
Date: 19 Sep 2002 01:16:50 -0400

As background, IEEE754-1985 and IEEE854-1987 are being revised. The working group's page is at http://grouper.ieee.org/groups/754/

The 754 committee is contemplating making signaling NaNs optional.  In real life, that means no one would implement them.  Thus, we're searching for actual uses of signaling NaNs to see what damage would occur.

The real choice ahead is whether the committee should
a) make signaling NaNs optional, or
b) completely standardize signaling NaN behavior.

So, does anyone know of codes that actually use signaling NaNs?

We know of using it for finding uninitialized floating-point data, but general-purpose NaNs handle that task well, too.  One of signaling NaN's original purposes was to support "alternative arithmetics."  No one seems to interested, probably because signaling NaNs currently are too platform-specific and too slow.

There are some known issues with fully standardizing signaling NaNs, like different hardware uses of the significand, different behavior on copying, compiler optimization gotchas, and system trapping support. Making signaling NaNs useful may cause more pain than making them optional.

We don't know, so we'd appreciate some feedback.  Feel free to pass this request on to other interested parties.  I've set follow-ups to myseld, and I'll summarize the responses in a week or so.  The responses will also be archived somewhere off the 754 committee page. If you feel the need to post a response, please trim the Newsgroups line.  Thanks!

Jason
--