Path: utzoo!attcan!uunet!lll-winken!lll-tis!helios.ee.lbl.gov!pasteur!ucbvax!BRL.ARPA!glennrp From: glennrp@BRL.ARPA (Glenn Randers-Pehrson, WMB) Newsgroups: comp.sys.sgi Subject: exp(-100.)=-.43e34 on IRIS Message-ID: <8807070852.aa22044@TBD.BRL.ARPA> Date: 7 Jul 88 12:52:49 GMT Sender: daemon@ucbvax.BERKELEY.EDU Organization: The Internet Lines: 36 The behaviour of the EXP() function for large negative numbers on the IRIS is shameful. I've tried the following program on a number of machines (VAX f77, Gould f77, Alliant FORTRAN, Cray-2 cft77) and all return 0.0 for out-of-range negative numbers. The IRIS returns not-a-number if the floating point board isn't used. Worse, if the floating point board is used it returns wrong numbers. This on a 3130 running 3.5r2 and a 2500T running 3.6. program bugexp dimension trial(10) data trial/-1., -20., -50., -80., -85., -88.02968, -88.02970, & -100., -1000., -10000/ do 10 i=1,10 x=trial(i) xx=exp(x) write(*,'('' exp('',f13.6,'')='',g15.8)')x,xx 10 continue stop end Results with f77 -Zf: exp( -1.000000)= .36787950 exp( -20.000000)= .20611540E-08 exp( -50.000000)= .19287500E-21 exp( -80.000000)= .18048510E-34 exp( -85.000000)= .12160990E-36 exp( -88.029680)= .15974800E-42 exp( -88.029700)=??????????????? exp( -100.000000)= -.43075530E+34 exp( -1000.000000)= .12234650E+29 exp(-10000.000000)= .41732680E-27 Programmed STOP