exp(large numbers) on IRIS

[John D. McCalpin]

Silicon Graphics has known about the pathetic behavior of the exp
function for a long time, but they do not seem at all interested in
doing anything about it.... As I understand it, they bought the
compiler from Silicon Valley Software, and did an in-house rewrite of
the intrinsic transcendental functions.  

The behaviour can be modified (though not corrected) by initializing
the floating-point exception handler. You will need to add the
following lines to the example code:

#include <fortfpe.h>				! somewhere in declarations

	call setfpe (NOCOREDUMP, ZEROVALS)	! first executable statement

This makes MOST underflows go gently to zero, but it does not 
fix the exp() problem. I don't know why the floating-point exception
handler is not active by default....

I have to uses a fix to the exp() problem, I use
	x = min(x,88.0)
	x = max(-88.0,x)
	result=exp(x)
or, in one line:
	result = exp(max(-88.0,min(x,88.0)))
The constant 88.0 is the largest value which returns a correct result
with 32-bit precision. I can't recall the number for 64-bit precision
--- I think it is about 300.

Now that the 4D's are the SGI's main interest, I would not count on
getting any improvements to the SVS Fortran on the IRIS 3000's....
especially since they did not respond when the IRIS 3000 was their 
only entry.  I suggest that you try a SUN machine.  They have troubles, 
but at least the Fortran numerics are quite robust.

John D. McCalpin
mccalpin at nu.cs.fsu.edu