[haberman]

It's possible that both were "correct" but that -O2 and -O3 gave different results due to x86 extended precision. For some interesting reading, see:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=323 (a centi-bug of back and forth between "it's broken!" and "no it's not, and we're not going to change the compiler.")

http://gcc.gnu.org/ml/gcc/2003-08/msg01183.html (a really informative thread about the issue)