https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69989
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
So - we have an irreducible region that has a reducible part (a memset loop).
The entry into that reducible part is marked as EDGE_IRREDUCIBLE_LOOP.
Loop distribution re-directs the src of the exit of that reducible part
(also marked irreducible) but ends up removing the entry. So we trigger
the "redundant" (?) edge flag check (if the edge is part of an irreducible
region both the src and dest should be marked irreducible as well).
Now... with an embedded reducible sub-CFG that seems no longer true.
Restricting ourselves to
void
remove_edge (edge e)
{
if (current_loops != NULL)
{
rescan_loop_exit (e, false, true);
/* If we remove an edge that is part of an irreducible region
or we remove an entry into an irreducible region we may expose
new natural loops. */
if (! loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
|| e->dest->flags & BB_IRREDUCIBLE_LOOP)
loops_state_set (LOOPS_NEED_FIXUP);
}
"fixes" this particular regression. But of course as soon as the reducible
sub-CFG were removed in other ways (redirecting the entry and removing the
exit edge) we'd be back at the very same issue.
So I think we can't rely on those flags given the odd sub-CFG marking.
But in fact what we could do is instead of
static inline void
checking_verify_loop_structure (void)
{
if (flag_checking)
verify_loop_structure ();
}
do
static inline void
checking_verify_loop_structure (void)
{
loops_state_clear (LOOPS_NEED_FIXUP);
if (flag_checking)
verify_loop_structure ();
}
because that's essentially what those checking_verify_loop_structure will
assert (and it'll be useful then to avoid excessive fixups as well).
