* [PATCH] Fix LIM PHI movement cost
@ 2016-05-18 10:45 Richard Biener
0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2016-05-18 10:45 UTC (permalink / raw)
To: gcc-patches
The PHI movement penalty is bogously biased by adding the cost of moving
the condition - but the condition is computed unconditionally and thus
its cost should be added to the PHI cost.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
This patch is required to avoid regressing with a followup that will
exchange LIM and PRE (PRE can cleanup after LIM which avoids one
copyprop pass and it can perform CSE which increases vectorization
opportunities - failed to find the PR right now).
Richard.
2016-05-18 Richard Biener <rguenther@suse.de>
* tree-ssa-loop-im.c (determine_max_movement): Properly add
condition cost to PHI cost instead of total_cost.
Index: gcc/tree-ssa-loop-im.c
===================================================================
--- gcc/tree-ssa-loop-im.c (revision 236361)
+++ gcc/tree-ssa-loop-im.c (working copy)
@@ -717,7 +717,7 @@ determine_max_movement (gimple *stmt, bo
return false;
def_data = get_lim_data (SSA_NAME_DEF_STMT (val));
if (def_data)
- total_cost += def_data->cost;
+ lim_data->cost += def_data->cost;
}
/* We want to avoid unconditionally executing very expensive
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2016-05-18 10:45 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-05-18 10:45 [PATCH] Fix LIM PHI movement cost Richard Biener
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).