http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58742
--- Comment #16 from rguenther at suse dot de <rguenther at suse dot de> --- On Tue, 21 Jan 2014, glisse at gcc dot gnu.org wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58742 > > --- Comment #15 from Marc Glisse <glisse at gcc dot gnu.org> --- > Another example: http://stackoverflow.com/q/21253690/1918193 > > where we have (-DVERSION=2): > _128 = img$_M_impl$_M_start_130 + 4000000; > pretmp_146 = (long intD.12) _128; > pretmp_145 = (long intD.12) img$_M_impl$_M_start_130; > pretmp_76 = pretmp_146 - pretmp_145; > pretmp_121 = pretmp_76 /[ex] 4; > pretmp_120 = (size_typeD.24047) pretmp_121; > > We miss that pretmp_120 is a constant, VRP thus fails to eliminate the range > checks, vectorization doesn't happen, and the code is more that 4 times slower > than it should be. > > If the desired reassoc version is hard, a simple forwprop pattern matching > would already go a long way to alleviate this issue. I agree. I have a patch.
