https://gcc.gnu.org/bugzilla/show_bug.cgi?id=23471
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Marc Glisse from comment #3) > The case a*a has been handled for a while, both in fold-const.c > (tree_binary_nonnegative_warnv_p) and in VRP. However, the case a*a*a*a*a*a > is not handled. In the .optimized dump at -O3, we still have: > > b_3 = a_2(D) * a_2(D); > b_4 = a_2(D) * b_3; > b_5 = a_2(D) * b_4; > b_6 = a_2(D) * b_5; > b_7 = a_2(D) * b_6; > > If reassoc or some similar pass had turned it into the following, the > optimization would probably work: > > b_3 = a_2(D) * a_2(D); > b_4 = a_2(D) * b_3; > b_7 = b_4 * b_4; > > Not sure how often this would be useful though. Hum, reassoc handles this for floats only it seems (__builtin_powi detection and expansion). Ah, and of course reassoc does nothing to TYPE_OVERFLOW_UNDEFINED ops. It seems to be reassoc could handle this case anyway as code-gen optimiztion.
