https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96685
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Ok, so for GENERIC it seems to be the associate: in fold_binary_loc that converts ~x + y created by this patch into (y - x) + 1, and we don't have an equivalent for that in GIMPLE. So, shall I restrict this match.pd optimization to #if GIMPLE only, or shall I canonicalize not to ~x + y but to (y - x) + 1, or should we implement the associate: optimization on GIMPLE? I guess the last one might be too much in stage3. I guess the middle option wouldn't help for the testcase in the patch, because we wouldn't canonicalize ~x + y to (y - x) + 1 and the first option would not consider ~x + y or ~(x - y) equivalent to user-written (y - x) + 1.
