Intel position on software workaround for FDIV

[uunet!mathworks.com!moler]

Here is a copy of an Internet posting from Richard Wirt at Intel.
He asked me to pass it on to this mailing list.
  -- Cleve

------------

From: "wirt (Richard Wirt)" <wirtaintel.com>
Newsgroups: sci.math.num-analysis
Subject: Software workaround for FDIV
Date: 8 Dec 1994 02:38:51 GMT
Organization: Intel

Intel is actively working with leading members of the software 
community to define and implement a uniform, efficient software 
workaround for the floating point divide flaw on the current version
 of the Pentium (TM) processor.

To create this patch we are have formed an informal industry working 
group -- including Cleve Moler of Mathworks, Tim Coe of Vitesse 
Semiconductor, Peter Tang of Argonne National Laboratories and Terje 
Mathisen of Norsk Hydro -- to work with Intel's engineers.

The patch will include an FDIV workaround based on identifying the 
divisor and dividend for regions that can yield imprecise results. 
This approach will provide the best solution with the least 
performance impact to the end user.

Many of you have been concerned about the transcendental instructions 
that use FDIV. Sine, cosine and the exponentials have been proven not 
to be impacted by the floating point flaw. We will include a library 
for tangent, arc-tangent and the remainder, all of which use the FDIV 
operation.The proof for the log functions' not being impacted by the 
flaw is still being pursued.

Intel is now working with compiler vendors to incorporate the patch 
into their compilers and compiler libraries. After we have completed 
validation of the patch, and the major compiler vendors have completed 
validation with their test suites, we are committed to posting the 
workaround on the Intel WWW server (http://www.intel.com). We expect 
to have this posted by the end of 1994.

Dr. Richard Wirt
Intel Fellow
Director, Software Technology Laboratory