tree-optimization/102155 - fix LIM fill_always_executed_in CFG walk
authorRichard Biener <rguenther@suse.de>
Wed, 1 Sep 2021 07:51:45 +0000 (09:51 +0200)
committerRichard Biener <rguenther@suse.de>
Thu, 2 Sep 2021 05:55:29 +0000 (07:55 +0200)
commitf482bf2af86990329b4df660f8c1eb9e094de9f9
treef0b1271a7b6cf34f7e2c5cc3b393103ee4bdf4e4
parent6e16b2123dd1cf5bab23cda0ce65223e5d55eded
tree-optimization/102155 - fix LIM fill_always_executed_in CFG walk

This fixes the CFG walk order of fill_always_executed_in to use
RPO oder rather than the dominator based order computed by
get_loop_body_in_dom_order.  That fixes correctness issues with
unordered dominator children.

The RPO order computed by rev_post_order_and_mark_dfs_back_seme in
its for-iteration mode is a good match for the algorithm.

2021-09-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102155
* tree-ssa-loop-im.c (fill_always_executed_in_1): Iterate
over a part of the RPO array and do not recurse here.
Dump blocks marked as always executed.
(fill_always_executed_in): Walk over the RPO array and
process loops whose header we run into.
(loop_invariant_motion_in_fun): Compute the first RPO
using rev_post_order_and_mark_dfs_back_seme in iteration
order and pass that to fill_always_executed_in.
gcc/tree-ssa-loop-im.c