Re: [Firebird-devel] What is expected Vulcan (or SQL) behavior with NaNs?

[Fred P. <fp...@us...>]

Tom Cole wrote:
> A little update, while some of you are scratching your heads and wondering "what IS he babbling about?  NaNs are just ignored or propagated, aren't they?"  
> 
> It turns out this is likely just my problem, since my callers (of Vulcan in embedded mode) have changed the process state to make NaN operations throw exceptions more aggressively on Windows.  Cans open, worms everywhere.
> 
> I'd still be delighted for pointers to standards that might guide me in ensuring that I don't break any existing assumptions about NaNs, though, while I muddle through this.
> 
> Thanks.
> ...[snip]

Hi Tom,

If this is still an issue, you might have a look at the <math.h>
comparison macros, specified by C99/POSIX/etc, to see if they'll suit
your needs. On some platforms, these macros may map to a few inline
CPU instructions. Here is a excerpt of the "isgreater" MAN page from
an openSUSE 10.2 system:
	...
        The normal relation operations (like less) will
        fail if one of the operands is NaN. This will
        cause an exception. To avoid this, C99 defines
        these macros. The macros are guaranteed to
        evaluate their operands only once.  The operand
        can be of any real floating-point type.

        isgreater()
               determines (x) > (y) without an exception if x or y is NaN.
        isgreaterequal()
               determines (x) >= (y) without an exception if x or y is NaN.
        ...
        isunordered() is true if x or y is NaN and false
               otherwise.

        NOTE Not all hardware supports these functions,
	      and where it doesn't, they will be emulated by
               macros. This will give you a performance
	      penalty.  Don't use these functions if NaN
	      is of no concern for you

GLIBC, at least, redefines these macros as optimized assembler instructions
for many of the architectures it supports, but not all. Other architectures
fall back to using the GCC provided builtin functions for their
implementation. I think if you look in <bits/mathinline.h> in your GLIBC
installation, you should be able to see if there are optimized assembler
versions of these macros for your target platform.

These macros are also supported by SunOS 5.10. I don't know if MS/Windows
compilers support these "standard" macros, but since the word "standard" is
involved, I doubt it. ;-) ;-) ;-)

Regards,
---Fred P.