[Boost-testing] [math] floating point classification - testing help wanted

From: John Maddock (john_at_[hidden])
Date: 2005-12-01 08:29:19

For a while now I've been meaning to clean up my floating-point 
classification code to provide a more reliable way of testing for Nan's and 
infinities.

The new TR1 complex number code could really use this to improved it's 
reliability.  I believe Robert Ramey was looking for this functionality for 
serialisation, and frankly it's hard to write any serious floating point 
code without it.

So...  attached is a first version of an fpclassify implementation for Boost 
(if there's interest).  As long as this one works (not a given 
unfortunately) then it's obviously trivial to sit isnan, isinfinity etc on 
top of it.

So.... I'm looking for testers to run the attached program on as many 
different platforms/compilers as possible, basically I want to see if 
there's any mileage in this or if I'm on to a looser with this code!

Please build the test program with optimisations turned on: many compilers 
make non-IEEE arithmetic assumptions once you tell them to optimise the 
code.

Tested so far:

Win32:
VC6: Can't detect signaling NaN's (_isnan is broken).
VC7, 7.1, 8: OK
Intel 7, 8, 9: OK provided you're not using the VC6 runtime.
GCC cygwin: OK.
Borland 5.6.4: So broken I don't know where to begin :-(

HP-UX:
gcc-3.4.2: OK.
aCC: OK (native fpclassify broken for long doubles, but portable version 
works).

Tru64:
cxx : OK (but can't test infinities/NaN's as they're not supported?

Suse Linux:
gcc-3.3.3 : OK.

Many thanks, John.