https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80213
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|REOPENED |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
We fail to handle the final reduction loop-closed PHI because the SESE region
doesn't include the BB with that PHI.
Ah, so with the --param we say the inner loop alone isn't a valid SESE
so we build a SESE for the outer loop directly while in the non-param
case we go through merge_sese of the inner loop SESE with the outer loop
one (that sounds pointless, but anyway). _This_ merge_sese adds an
extra block via
/* FIXME: We should remove this piece of code once
canonicalize_loop_closed_ssa has been removed, because that function
adds a BB with single exit. */
if (!trivially_empty_bb_p (get_exit_bb (combined)))
{
/* Find the first empty succ (with single exit) of combined.exit. */
basic_block imm_succ = combined.exit->dest;
if (single_succ_p (imm_succ)
&& single_pred_p (imm_succ)
&& trivially_empty_bb_p (imm_succ))
combined.exit = single_succ_edge (imm_succ);
else
{
DEBUG_PRINT (dp << "[scop-detection-fail] Discarding SCoP because "
<< "no single exit (empty succ) for sese exit";
print_sese (dump_file, combined));
return invalid_sese;
}
}
which happens to be the block with the loop-closed PHI nodes. Note that
canonicalize_loop_closed_ssa makes sure those are in a separate BB.
I think the proper change here is to make get_sese include the BB with
the LC PHI nodes anyhow.
Testing fix.
