https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93820
--- Comment #7 from rguenther at suse dot de <rguenther at suse dot de> --- On Tue, 25 Feb 2020, jakub at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93820 > > --- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> --- > On the original testcase, the reason why dse3 doesn't help is we have: > a[4] = 1; > _4 = *c.0_1; > a[5] = _4; > a[3] = 1; > e = 0; > a[8] = 0; > MEM <vector(4) int> [(int *)&a + 16B] = { 0, 0, 0, 0 }; > and c.0_1 might alias the a array, so dse3 removes a[5] store as redundant, > but > only later cddce removes the _4 = *c.0_1 load and only then the a[4] = 1; > store > could be also removed (i.e. it would require iteration, or performing some > kind > of (cheap) dce in the dse pass after removing each store). A simple immediate-use / worklist based DCE would be possible after each DSE but then this always felt like too ad-hoc to me ... It would also have hidden this bug in store-merging (which is also why I prefer the GIMPLE testcase at least for the affected function since there's 100s of passes before store-merging)
