https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107767
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #10)
> Looks like a general CFG optimization if we have N forwarders to a PHI with
> the same value then we can merge them to one (it's enough to have a single
> forwarder which we can use as the remaining one even). Note that's kind-of
> a reverse
> mergephi for same values.
>
> PHI <a, b, c, c>
>
> ->
>
> forwarder
> |
> PHI <a, b, c>
>
> CFG cleanup will then eventually elide forwarders into the added forwarder.
Are those actually forwarder blocks though?
Doesn't the GIMPLE_PREDICT statement at the start of each one of them prevent
tree_forwarder_block_p from returning true?
As a hack I've removed them manually:
FOR_EACH_BB_FN (bb, cfun)
{
gimple_stmt_iterator gsi = gsi_after_labels (bb);
if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) == GIMPLE_PREDICT)
gsi_remove (&gsi, true);
}
in pass_if_to_switch::execute before return TODO_cleanup_cfg;, but that didn't
help.
