https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127085
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that you made a former irreducible region reducible, so a new loop
appears. I wonder why you ever want to duplicate the cond over a backedge?
One way out would be to loops_state_set (LOOPS_NEED_FIXUP) which allows
new loop discovery.
In preserves_loop_structure_p you say "requires EDGE_DFS_BACK to be current",
but the pass doesn't mark backedges. Instead of checking EDGE_DFS_BACK
simply never duplicate loop headers. There's also duplicating blocks
with exits which requires re-scanning loop exits (but phiopt doesn't have
recorded exits).
So I'd change preserves_loop_structure_p to just reject
/* Never duplicate loop headers. */
if (e->dest->loop_father->header == e->dest)
return false;
and do loops_state_set (LOOPS_NEED_FIXUP). If you want to avoid altering
irreducible regions you'd have to avoid duplicating all blocks with
incoming backedges, thus mark_dfs_back_edges ().
Usually making irreducible regions reducible is good. You can of course
end up making it even more irreducible.
