Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!mit-eddie!snorkelwacker!think!zaphod.mps.ohio-state.edu!usc!elroy.jpl.nasa.gov!ames!pacbell!att!cbnewsh!daw
From: daw@cbnewsh.ATT.COM (David Wolverton)
Newsgroups: comp.lang.c
Subject: Re: C Community's Cavalier Attitude On Software Reliability
Summary: a note on math libs accuracy
Message-ID: <8512@cbnewsh.ATT.COM>
Date: 27 Feb 90 20:53:02 GMT
References: <8147@hubcap.clemson.edu> <2910@goanna.oz.au>
Organization: AT&T Bell Labs, Lincroft, NJ
Lines: 15

> In article <8147@hubcap.clemson.edu>, wtwolfe@hubcap.clemson.edu (Bill Wolfe) writes:
> >    SIN(3M)  The value of 'tan' for arguments greater than about 2**31
> >             is garbage.

Seeing this reminded me to mention to the net that math libraries need
not be bad.  While it is probably true that even some recent UNIXes
contain some bad math functions, they need not.  Anyone who is responsible
for a C implementation should do at least as well as the 4.3 BSD libm,
which is very careful about things like loss of accuracy.
The math library in SVR4 uses the 4.3 libm, plus some additional
accuracy improvements (see the paper in the winter Usenix proceedings)
from various places.

Dave Wolverton
daw@attunix.att.com