Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!sun-barr!newstop!exodus!exodus-bb!khb
From: khb@chiba.Eng.Sun.COM (Keith Bierman fpgroup)
Newsgroups: comp.lang.fortran
Subject: Re: detecting Nan's
Message-ID: <KHB.91Jan10195225@chiba.Eng.Sun.COM>
Date: 11 Jan 91 03:52:25 GMT
References: <9101110259.AA28451@ucbvax.Berkeley.EDU>
Sender: news@exodus.Eng.Sun.COM
Organization: Sun MegaSystems
Lines: 74
In-reply-to: JBS@IBM.COM's message of 11 Jan 91 02:56:52 GMT


In article <9101110259.AA28451@ucbvax.Berkeley.EDU> JBS@IBM.COM writes:

   Newsgroups: comp.lang.fortran
   Date: 11 Jan 91 02:56:52 GMT
   Sender: daemon@ucbvax.BERKELEY.EDU
   Lines: 5

   >      if(dnum*0d0 .ne. 0d0) write(*,1)
   >1     format(' The variable dnum is equal to NAN!')

	  Is this correct?  Suppose dnum is Inf.

No, it isn't. Aside from reading the standard, let's try code!

      real dnum/1.0/,x/0.0/
      dnum=dnum/x
      print*,dnum
      if(dnum*0d0 .ne. 0d0) print*," is nan?",dnum
      end
      
f77 bork.f && a.out
bork.f:
 MAIN:
 Inf           
 is nan? Inf           

On some systems, vendors will have been kind and provided library
functions. For example, on a Sun one might code something like

      real dnum/1.0/,x/0.0/
      integer ir_isnan,ir_isinf  ! libm function
      dnum=dnum/x
      print*,dnum
      
      if (ir_isnan(dnum)) print*,"dnum IS NaN"
      if (ir_isinf(dnum)) print*,"dnum IS Inf"

      end
      
f77 bork.f && a.out
bork.f:
 MAIN:
 Inf           
dnum IS Inf

It would be more in keeping with the notion of standard compliance, to
not treat ir_* as both a logical and an integer, so one might quite
reasonably prefer to either declare ir* as logicals (despite the libm
definition) or to test the value

	if (ir() .eq. 1)

On systems which are ieee hw compliant, but have not provided the sw
hooks, one can easily code these functions in whatever language seems
most appropriate. To test for NaN it is probably best to use the
result of the appropriate mask ANDed with the value under test.

note that on a Sun, unless user overridden there will also be a
message from ieee_retrospective, ala

 Warning: the following IEEE floating-point arithmetic exceptions 
 occurred in this program and were never cleared: 
 Division by Zero; 

Final note, if memory serves, the right mask is 0x7ff00000

cheers

--
----------------------------------------------------------------
Keith H. Bierman    kbierman@Eng.Sun.COM | khb@chiba.Eng.Sun.COM
SMI 2550 Garcia 12-33			 | (415 336 2648)   
    Mountain View, CA 94043