[1/2] Improve compute_control_dep_chain documentation

Message ID 20220826123409.524BF13A7E@imap2.suse-dmz.suse.de
State New, archived
Headers
Series [1/2] Improve compute_control_dep_chain documentation |

Commit Message

Richard Biener Aug. 26, 2022, 12:34 p.m. UTC
  In the quest to understand how compute_control_dep_chain works I've
produced the following two changes, documenting PR106754 on the
way.

Bootstrap and regtest is running on x86_64-unknown-linux-gnu,
the changes should be no-ops but hopefully improve understanding
of the code.

--


The following refactors compute_control_dep_chain slightly by
inlining is_loop_exit and factoring the check on the loop
invariant condition.  It also adds a comment as of how I
understand the code and it's current problem.

	* gimple-predicate-analysis.cc (compute_control_dep_chain):
	Inline is_loop_exit and refactor, add comment about
	loop exits.
---
 gcc/gimple-predicate-analysis.cc | 27 +++++++++++++++++++++++----
 1 file changed, 23 insertions(+), 4 deletions(-)
  

Comments

Jeff Law Aug. 26, 2022, 4:19 p.m. UTC | #1
On 8/26/2022 6:34 AM, Richard Biener via Gcc-patches wrote:
> In the quest to understand how compute_control_dep_chain works I've
> produced the following two changes, documenting PR106754 on the
> way.
>
> Bootstrap and regtest is running on x86_64-unknown-linux-gnu,
> the changes should be no-ops but hopefully improve understanding
> of the code.
Thanks so much for this.  Every time I've had to get into that code it 
makes my head hurt -- anything that clarifies is an improvement.

Jeff
  

Patch

diff --git a/gcc/gimple-predicate-analysis.cc b/gcc/gimple-predicate-analysis.cc
index 32542f93057..934e9516e7b 100644
--- a/gcc/gimple-predicate-analysis.cc
+++ b/gcc/gimple-predicate-analysis.cc
@@ -1110,6 +1110,10 @@  compute_control_dep_chain (basic_block dom_bb, const_basic_block dep_bb,
 			   vec<edge> &cur_cd_chain, unsigned *num_calls,
 			   unsigned in_region = 0, unsigned depth = 0)
 {
+  /* In our recursive calls this doesn't happen.  */
+  if (single_succ_p (dom_bb))
+    return false;
+
   if (*num_calls > (unsigned)param_uninit_control_dep_attempts)
     {
       if (dump_file)
@@ -1167,7 +1171,21 @@  compute_control_dep_chain (basic_block dom_bb, const_basic_block dep_bb,
       basic_block cd_bb = e->dest;
       cur_cd_chain.safe_push (e);
       while (!dominated_by_p (CDI_POST_DOMINATORS, dom_bb, cd_bb)
-	     || is_loop_exit (dom_bb, cd_bb))
+	     /* We want to stop when the CFG merges back from the
+		branch in dom_bb.  The post-dominance check alone
+		falls foul of the case of a loop exit test branch
+		where the path on the loop exit post-dominates
+		the branch block.
+		The following catches this but will not allow
+		exploring the post-dom path further.  For the
+		outermost recursion this means we will fail to
+		reach dep_bb while for others it means at least
+		dropping the loop exit predicate from the path
+		which is problematic as it increases the domain
+		spanned by the resulting predicate.
+		See gcc.dg/uninit-pred-11.c for the first case
+		and PR106754 for the second.  */
+	     || single_pred_p (cd_bb))
 	{
 	  if (cd_bb == dep_bb)
 	    {
@@ -1187,9 +1205,10 @@  compute_control_dep_chain (basic_block dom_bb, const_basic_block dep_bb,
 	    break;
 
 	  /* Check if DEP_BB is indirectly control-dependent on DOM_BB.  */
-	  if (compute_control_dep_chain (cd_bb, dep_bb, cd_chains,
-					 num_chains, cur_cd_chain,
-					 num_calls, in_region, depth + 1))
+	  if (!single_succ_p (cd_bb)
+	      && compute_control_dep_chain (cd_bb, dep_bb, cd_chains,
+					    num_chains, cur_cd_chain,
+					    num_calls, in_region, depth + 1))
 	    {
 	      found_cd_chain = true;
 	      break;