From: Richard Biener Date: Tue, 2 Nov 2021 17:47:14 +0000 (+0100) Subject: tree-optimization/103029 - ensure vect loop versioning constraint on PHIs X-Git-Tag: basepoints/gcc-13~3463 X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=164bbf701ff10ff44e272525e8f462ed3ff1cf43;p=thirdparty%2Fgcc.git tree-optimization/103029 - ensure vect loop versioning constraint on PHIs PHI nodes in vectorizer loop versioning need to maintain the same order of PHI arguments to not disturb SLP discovery. The following adds an assertion and mitigation in case loop versioning breaks this which happens more often after the recent reorg. 2021-11-02 Richard Biener PR tree-optimization/103029 * tree-vect-loop-manip.c (vect_loop_versioning): Ensure the PHI nodes in the loop maintain their original operand order. --- diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c index 72d583189c47..f788deb3d72c 100644 --- a/gcc/tree-vect-loop-manip.c +++ b/gcc/tree-vect-loop-manip.c @@ -3562,12 +3562,29 @@ vect_loop_versioning (loop_vec_info loop_vinfo, "applying loop versioning to outer loop %d\n", loop_to_version->num); + unsigned orig_pe_idx = loop_preheader_edge (loop)->dest_idx; + initialize_original_copy_tables (); nloop = loop_version (loop_to_version, cond_expr, &condition_bb, prob, prob.invert (), prob, prob.invert (), true); gcc_assert (nloop); nloop = get_loop_copy (loop); + /* For cycle vectorization with SLP we rely on the PHI arguments + appearing in the same order as the SLP node operands which for the + loop PHI nodes means the preheader edge dest index needs to remain + the same for the analyzed loop which also becomes the vectorized one. + Make it so in case the state after versioning differs by redirecting + the first edge into the header to the same destination which moves + it last. */ + if (loop_preheader_edge (loop)->dest_idx != orig_pe_idx) + { + edge e = EDGE_PRED (loop->header, 0); + ssa_redirect_edge (e, e->dest); + flush_pending_stmts (e); + } + gcc_assert (loop_preheader_edge (loop)->dest_idx == orig_pe_idx); + /* Kill off IFN_LOOP_VECTORIZED_CALL in the copy, nobody will reap those otherwise; they also refer to the original loops. */