https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101419
--- Comment #16 from rguenther at suse dot de <rguenther at suse dot de> --- On Mon, 12 Jul 2021, jakub at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101419 > > --- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> --- > I agree about most of the passes you are moving, but I have an (albeit > artificial) testcase that proves that cunrolli does affect objsz: > __SIZE_TYPE__ a[10]; > > void > foo (void) > { > char *p = __builtin_malloc (64); > for (int i = 0; i < 4; i++) > { > a[i] = __builtin_object_size (p, 0); > p += 6; > } > } > > When objsz is done before cunrolli, a[0] to a[3] will all be 64, while > when it is done after cunrolli, it will be 64, 58, 52, 46. > > So, what about a mix of your and my patch, add > --- gcc/passes.def > +++ gcc/passes.def > @@ -194,14 +194,14 @@ along with GCC; see the file COPYING3. If not see > They ensure memory accesses are not indirect wherever possible. */ > NEXT_PASS (pass_strip_predict_hints, false /* early_p */); > NEXT_PASS (pass_ccp, true /* nonzero_p */); > - NEXT_PASS (pass_post_ipa_warn); > /* After CCP we rewrite no longer addressed locals into SSA > form if possible. */ > + NEXT_PASS (pass_post_ipa_warn); > NEXT_PASS (pass_complete_unrolli); > + NEXT_PASS (pass_object_sizes, false /* insert_min_max_p */); > NEXT_PASS (pass_backprop); > NEXT_PASS (pass_phiprop); > NEXT_PASS (pass_forwprop); > - NEXT_PASS (pass_object_sizes, false /* insert_min_max_p */); > /* pass_build_alias is a dummy pass that ensures that we > execute TODO_rebuild_alias at this point. */ > NEXT_PASS (pass_build_alias); > to my patch? Well, my point was to avoid pessimizing the VN done from cunrolli ;) Of course any duplication / threading can improve __bos precision, but then any transform also risks breaking it. Your example is IMHO too artificial as good argument.
