From patchwork Tue Sep 11 12:25:12 2012
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Subject: shrink-wrapping duplicates BBs across partitions.
Date: Tue, 11 Sep 2012 02:25:12 -0000
From: Christian Bruel <christian.bruel@st.com>
X-Patchwork-Id: 183087
Message-Id: <504F2DA8.3050208@st.com>
To: "gcc-patches@gcc.gnu.org" <gcc-patches@gcc.gnu.org>

Hello,

While testing the patch to enable shrink-wrapping on SH [PR54546], we
hit an the "error: EDGE_CROSSING missing across section boundary"

Indeed, shrink-wrap duplicates a bb with successors (containing the
return sequence) into an unlikely section. I first thought about setting
the EDGE_CROSSING on flag on those edge, but I feel that this block
duplication doesn't go in the direction of this optimization. Not
duplicating BBs between partitions solves the problem.

Does this restriction look right to you ? (regression tests are still
running on x86 and sh)

Thanks a lot for any comment.

Christian


Index: function.c
===================================================================
--- function.c	(revision 191177)
+++ function.c	(working copy)
@@ -6063,6 +6063,7 @@
 	  FOR_EACH_EDGE (e, ei, tmp_bb->preds)
 	    if (single_succ_p (e->src)
 		&& !bitmap_bit_p (&bb_on_list, e->src->index)
+		&& (BB_PARTITION (e->src) == BB_PARTITION (e->dest))
 		&& can_duplicate_block_p (e->src))
 	      {
 		edge pe;