------- Comment #3 from dominiq at lps dot ens dot fr 2007-05-31 21:31 ------- > This behavior with -ffast-math is expected and is hoped for.
Yes indeed: I said I would like to see it at some level of -On. > That is PR 323. If I understood correctly the PR, it deals with side effects to do the extra precision in x87. I have done the tests on a PPC (sorry for not mentionning it). In the later case it is due to the fact that a/b may have a different round-off error as a*(1.0/b). > *** This bug has been marked as a duplicate of 28684 *** Related, but a different issue. My understanding of the Fortran standard (7.1.8.3) is that reordering is allowed: 'a+b+c' can be computed as '(a+c)+b' (note 7.18), even if it is usually computed as '(a+b)+c', but "7.1.8.2 Integrity of parentheses" forbids to compute '(a+f)-f' as 'a'. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32172
