#370 Problems with IEEE-float NaN and INF

[Stavros Macrakis]

flinf: 2.0^1024$

flinf => Error: Can't print a non-number (lisp break)
Should print something special, e.g. INFe0

flnan: flinf-flinf$

flnan => Bind stack overflow (lisp break)
Should print something special, e.g. NaNe0

is(flnan=flnan) => TRUE NO! NaN is not equal to itself;
cf ?=(flnan,flnan).

is(flnan>0) => TRUE NO!
is(flnan<0) => TRUE NO!

flnan*0 => 0 NO!
flinf*0 => 0 NO!

GCL also can't print flnan/flinf.

Maxima 5.9.0 gcl 2.5.0 mingw Windows 2000 Athlon

[Robert Dodier]

Logged In: YES
user_id=501686

In reference to the examples given in the original write-up,
it is worth pointing out that GCL does indeed compute inf
and nan; it just can't display them.

Clisp refuses to compute the inf or nan, and there doesn't
appear to be a way to compel it to do so.

CMUCL is happy with inf and nan after :lisp
(extensions::set-floating-point-modes :traps nil) is executed.

[Robert Dodier]

Logged In: YES
user_id=501686

Clisp does not recognize floating point INF and NAN; those 2
values cannot be produced or recognized by Clisp's floating
point implementation (which is all software for greater
portability). So I can't see any way, short of modifying
Clisp, for Maxima to accomodate INF and NAN in a way that is
workable across Lisp implementations.

We could probably convince the GCL team to modify GCL to
handle INF and NAN; that seems a lesser challenge than
getting Clisp modified.

[Robert Dodier]

Just tried it with GCL 2.6.12 (near 2.6.13) + post-5.46, and I get 2.0^1024 resulting in false. Not sure what to think about that.

[Camm Maguire]

Sorry for the terse reply!

2.0^1024.1; => #<inf>, with which we can test the original submitters
results (to whom I refer as "he").</inf>

2.0^1024; => false, which I have traced to a failure somewhere in
maxima's #'simpexpt (not simp-expt, sorry for the typo), and appears
independent of running the code compiled or not.

I seem to have one small remaining issue in expt, but others seem to
remain in maxima:

=============================================================================
(%i16) flinf:2.0^1024;
(%o16) false
(%i17) flinf:2.0^1024.1;
(%o17) #<inf>
(%i18) flnan:flinf-flinf;
(%o18) #<-nan>
(%i19) flnan;
(%o19) #<-nan>
(%i20) is(flnan=flnan);
(%o20) true
(%i21) :lisp (= $flnan $flnan)</inf>

NIL
(%i21) cf?=(flnan,flnan);

incorrect syntax: = is not an infix operator
cf?=(
^
(%i21) is(flnan>0);
(%o21) true
(%i22) is(flnan<0);
(%o22) true
(%i23) :lisp (< $flnan 0)

1> (< #<-nan> 0)
<1 (< NIL)
NIL
(%i23) :lisp (> $flnan 0)

NIL
(%i23) flnan^0;
(%o23) 1.0
(%i24) flinf^0;
(%o24) 1.0
(%i25) :lisp (expt $flnan 0)

1.0
(%i25) :lisp (expt $flinf 0)

1.0
(%i25) flinf;
(%o25) #<inf>
(%i26)
(%i29) :lisp (trace simpexpt)</inf>

(SIMPEXPT)
(%i29) flinf:2.0^1024;

1> (SIMPEXPT ((MEXPT) 2.0 1024) 1 NIL)
<1 (SIMPEXPT NIL)
(%o29) false
(%i30) flinf:2.0^1024.1;

1> (SIMPEXPT ((MEXPT) 2.0 1024.1) 1 NIL)
<1 (SIMPEXPT #<inf>)
(%o30) #<inf>
(%i31)
=============================================================================</inf></inf>

[Camm Maguire]

This clears the 'false' result noticed recently:

modified src/simp.lisp
@@ -1003,9 +1003,9 @@
(t
(setq b (expt a (- b)))
(*red 1 b)))))

• (if (float-inf-p result) ;; needed for gcl - no trap of overflow
• (signal 'floating-point-overflow)
• result)))
• (when (float-inf-p result) ;; needed for gcl - no trap of overflow
• (signal 'floating-point-overflow))
• result))