http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58028
--- Comment #10 from rguenther at suse dot de <rguenther at suse dot de> --- On Wed, 29 Jan 2014, grosser at gcc dot gnu.org wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58028 > > --- Comment #9 from Tobias Grosser <grosser at gcc dot gnu.org> --- > (In reply to rguent...@suse.de from comment #8) > > On Wed, 29 Jan 2014, dominiq at lps dot ens.fr wrote: > > > > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58028 > > > > > > --- Comment #7 from Dominique d'Humieres <dominiq at lps dot ens.fr> --- > > > > The testsuite failures are. We have to do sth about them. > > > > > > Revert r200946? > > > > Probably - the > > > > + * graphite-dependences.c (carries_deps): Do not assume the > > schedule is > > + in 2D + 1 form. > > > > part looks wrong. > > > > - idx = 2 * depth + 1; > > - for (i = 0; i < idx; i++) > > + for (i = 0; i < depth - 1; i++) > > > > we now iterate over less dimensions than before. > > Only in cases where the schedule is not in 2D + 1 form. In case the schedule > is in 2D + 1 form > > int scheduling_dim = isl_set_n_dim (domain); > > will be equal to 2 * depth + 1. > > > I'd say we should simply check whether the loop _is_ in 2D + 1 form > > at > > > > + isl_set *domain = isl_set_from_cloog_domain (stmt->domain); > > + int scheduling_dim = isl_set_n_dim (domain); > > + > > if (flag_loop_parallelize_all > > - && loop_is_parallel_p (loop, bb_pbb_mapping, level)) > > + && loop_is_parallel_p (loop, bb_pbb_mapping, scheduling_dim)) > > loop->can_be_parallel = true; > > > > thus > > > > if (flag_loop_parallelize_all > > && scheduling_dim == 2 * level + 1 > > && loop_is_parallel_p (loop, bb_pbb_mapping, level) > > loop->can_be_parallel = true; > > The change you propose seems conservatively correct, as in that loops that > are not in 2D + 1 form are not detected to be parallel. This change may hide > the bug, but I don't see any bug it solves. It may solve the fallout from your change - I don't have enough graphite knowledge to fix a real bug there ;) > > no time to check whether reverting the other hunk plus this will > > resolve the bug the revision fixed and restores the testcases. > > I assume it will fix the crash, but it will not detect parallellism in none > 2D+1 loops, something the isl scheduler happily creates. Sure. We'd be much happier if the real bug was fixed, maybe you can give it a try? The FAIL: libgomp.graphite/force-parallel-4.c execution test FAIL: libgomp.graphite/force-parallel-4.c scan-tree-dump-times graphite "2 loops carried no dependency" 1 FAIL suggests that we miss some dependency now.
