Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!munnari!comp.vuw.ac.nz!dsiramd!ray
From: ray@dsiramd.nz (Ray Brownrigg)
Newsgroups: comp.bugs.sys5,comp.sys.att
Subject: 3B2 floating point libraries
Message-ID: <178@dsiramd.nz>
Date: Wed, 14-Oct-87 19:03:53 EDT
Article-I.D.: dsiramd.178
Posted: Wed Oct 14 19:03:53 1987
Date-Received: Sat, 17-Oct-87 02:09:54 EDT
Reply-To: ray@dsiramd.UUCP (Ray Brownrigg)
Organization: DSIR Applied Mathematics Division, Wellington, NZ
Lines: 19
Xref: mnetor comp.bugs.sys5:260 comp.sys.att:1497

I have uncovered a strange bug in the floating point routines for the
Fortran XLA+ and C-FP+ compilers for an AT&T 3B2 (with MAU), relating
to non-integral powers of 0.5 (and perhaps other numbers as well). For
example the expression 0.5**Y returns the value 0.5**(Y-1) when Y is
2.5, 4.5, 6.5, 8.5, ... Further, using C-FP+, the statement
	printf("%10.8f\n", pow(0.5, 2.5))
results in a zero being output.

I have reported the bugs (to Olivetti), but would be interested if anybody
else has seen the same problem.

Otherwise, beware. One workaround it to use exp(0.5*log(y)), but this is
not easy in a package as large as the S statistical package.

-- 
Ray Brownrigg		UUCP: {utai!calgary,uunet}!vuwcomp!dsiramd!ray
Applied Maths Div, DSIR		ACSnet:	ray@dsiramd.nz[@munnari]
PO Box 1335			System:	OLIVETTI/AT&T 3B2/400B+, System V R3.0
Wellington, New Zealand			"UNX -rules -OK"