Re: negative argument in ALOG bug in FEFF825 on CRAY PVP machines

[Matt Newville <newville@cars.uchicago.edu>]

Dear Feffers,

There WAS an occasional sign problem in padlib.f that only
revealed itself on the cray.  The problem boiled down to
determining the sign of x (double precision) as an integer:
       i = abs(x)/x

On all systems I've tried, this reliably gives -1 or 1, but
on the cray it occasionally gave 0.  This is probably an issue
related to precision. I would definitely characterize it as a
compiler problem, though there's probably a cray engineer who
would not agree with that.

Anyway, padlib.f now does (essentially)
       i = 1
       if (x.le.0) i = -1
and later assigns any number with (abs(x).le.1.d-39) to 0.d0.
This appears to satisfy the cray.

--Matt Newville