fdlibm which uses macros for manipulating the high and low words of a
double, like the rest of fdlibm functions in clib2. The previous version
would give bogus results when compiled with -O3 in clib2 which lead to
"interesting" results (and lots of fun while searching for the problem)
in Ghostscript. Wish I managed to track this down yesterday for 1.196
release... <aantonijevic>
