https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108519
--- Comment #1 from Alexander Monakov <amonakov at gcc dot gnu.org> --- We diverge in sched1 due to extra calls to advance_one_cycle when scheduling a BB that is empty apart from one debug insn. The following patch adds a hexdump of automaton state to make the problem evident: diff --git a/gcc/sched-rgn.cc b/gcc/sched-rgn.cc index 420c45dff..c09398897 100644 --- a/gcc/sched-rgn.cc +++ b/gcc/sched-rgn.cc @@ -3098,8 +3098,14 @@ save_state_for_fallthru_edge (basic_block last_bb, state_t state) memcpy (bb_state[f->dest->index], state, dfa_state_size); if (sched_verbose >= 5) - fprintf (sched_dump, "saving state for edge %d->%d\n", - f->src->index, f->dest->index); + { + fprintf (sched_dump, "saving state for edge %d->%d\n", + f->src->index, f->dest->index); + for (size_t i = 0; i < dfa_state_size; i++) + fprintf (sched_dump, "%02x%c", i[(unsigned char *)state], + (i+1) % 16 ? ' ' : '\n'); + fprintf(sched_dump, "\n---\n"); + } } } With the above patch it's obvious we advance the automaton state a few extra times when scheduling BB 3, and then inherit the modified state to BB 4. I think we don't need to schedule blocks that only contain debug insns. IBM folks, care to test the following? diff --git a/gcc/haifa-sched.cc b/gcc/haifa-sched.cc index 4efaa9445..f00a92e26 100644 --- a/gcc/haifa-sched.cc +++ b/gcc/haifa-sched.cc @@ -5040,7 +5040,7 @@ no_real_insns_p (const rtx_insn *head, const rtx_insn *tail) { while (head != NEXT_INSN (tail)) { - if (!NOTE_P (head) && !LABEL_P (head)) + if (!NOTE_P (head) && !LABEL_P (head) && !DEBUG_INSN_P (head)) return 0; head = NEXT_INSN (head); } diff --git a/gcc/sched-rgn.cc b/gcc/sched-rgn.cc index 420c45dff..c09398897 100644 --- a/gcc/sched-rgn.cc +++ b/gcc/sched-rgn.cc @@ -2753,7 +2753,7 @@ free_block_dependencies (int bb) get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); - if (no_real_insns_p (head, tail)) + if (0 && no_real_insns_p (head, tail)) return; sched_free_deps (head, tail, true);