https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94421
--- Comment #3 from otcmaf <xuemaosheng at huawei dot com> --- (In reply to Andrew Pinski from comment #1) > change_spec_dep_to_hard is called from two places: > * in sched_analyze_insn, it is conditionalized on (current_sched_info->flags > & DO_SPECULATION) already. > * in update_dep, it is conditionalized on if was_spec was true and if no > longer dep_spec_p (dep) is true. > > dep_spec_p is true when current_sched_info->flags & DO_SPECULATION is true > (or when a few other cases. > > But DEP_STATUS (dep) can only be SPECULATIVE when current_sched_info->flags > & DO_SPECULATION is true. > > So If anything you could in theory put an assert there for > current_sched_info->flags & DO_SPECULATION and get the same effect if I read > the code correctly. In our case, the compile bug is called from function update_dep. Variable was_spec is true, dep_spec_p (dep) is false, current_sched_info->flags & DO_SPECULATION is zero, DEP_STATUS (dep) & SPECULATIVE is zero. So in this case, we should avoid to free the memory of spec_dependency_cache. 1191 if (was_spec && !dep_spec_p (dep)) 1192 /* The old dep was speculative, but now it isn't. */ 1193 change_spec_dep_to_hard (sd_it); 1194 (gdb) p was_spec $11 = true (gdb) p dep_spec_p (dep) $12 = false (gdb) p current_sched_info->flags & DO_SPECULATION $13 = 0 (gdb) p DEP_STATUS (dep) & SPECULATIVE $14 = 0 (gdb) p !dep_spec_p (dep) $15 = true
