[PATCH] Fix set of even probabilities (PR middle-end/89737).

[PATCH] Fix set of even probabilities (PR middle-end/89737).

[Martin Liška]

[-- Attachment #1: Type: text/plain, Size: 839 bytes --]

Hi.

When calling set_even_probabilities, the function assumes that an edge
can't live in both sets ({un,}likely_edges). When such situation happens,
clear just the sets.

Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

Ready to be installed?
Thanks,
Martin

gcc/ChangeLog:

2019-03-18  Martin Liska  <mliska@suse.cz>

	PR middle-end/89737
	* predict.c (combine_predictions_for_bb): Empty likely_edges and
	unlikely_edges if there's an edge that belongs to both these sets.

gcc/testsuite/ChangeLog:

2019-03-18  Martin Liska  <mliska@suse.cz>

	PR middle-end/89737
	* gcc.dg/pr89737.c: New test.
---
 gcc/predict.c                  | 17 ++++++++++++++---
 gcc/testsuite/gcc.dg/pr89737.c | 17 +++++++++++++++++
 2 files changed, 31 insertions(+), 3 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/pr89737.c



[-- Attachment #2: 0001-Fix-set-of-even-probabilities-PR-middle-end-89737.patch --]
[-- Type: text/x-patch, Size: 1734 bytes --]

diff --git a/gcc/predict.c b/gcc/predict.c
index 43ee91a5b13..60a19d7edd1 100644
--- a/gcc/predict.c
+++ b/gcc/predict.c
@@ -1229,12 +1229,23 @@ combine_predictions_for_bb (basic_block bb, bool dry_run)
 	    if (pred->ep_probability <= PROB_VERY_UNLIKELY
 		|| pred->ep_predictor == PRED_COLD_LABEL)
 	      unlikely_edges.add (pred->ep_edge);
-	    if (pred->ep_probability >= PROB_VERY_LIKELY
-		|| pred->ep_predictor == PRED_BUILTIN_EXPECT
-		|| pred->ep_predictor == PRED_HOT_LABEL)
+	    else if (pred->ep_probability >= PROB_VERY_LIKELY
+		     || pred->ep_predictor == PRED_BUILTIN_EXPECT
+		     || pred->ep_predictor == PRED_HOT_LABEL)
 	      likely_edges.add (pred);
 	  }
 
+      /* It can happen that an edge is both in likely_edges and unlikely_edges.
+	 Clear both sets in that situation.  */
+      for (hash_set<edge_prediction *>::iterator it = likely_edges.begin ();
+	   it != likely_edges.end (); ++it)
+	if (unlikely_edges.contains ((*it)->ep_edge))
+	  {
+	    likely_edges.empty ();
+	    unlikely_edges.empty ();
+	    break;
+	  }
+
       if (!dry_run)
 	set_even_probabilities (bb, &unlikely_edges, &likely_edges);
       clear_bb_predictions (bb);
diff --git a/gcc/testsuite/gcc.dg/pr89737.c b/gcc/testsuite/gcc.dg/pr89737.c
new file mode 100644
index 00000000000..cd3dc81769e
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr89737.c
@@ -0,0 +1,17 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-profile_estimate" } */
+
+int a, b;
+
+void c() {
+  &&d;
+  void *e = &&f, *g = &&h;
+f:
+  __attribute__((hot)) h : __attribute__((cold)) for (; a;) goto *g;
+d:
+  for (; b;)
+    goto *e;
+}
+
+/* { dg-final { scan-tree-dump-times "predicted to even probabilities" 4 "profile_estimate"} } */
+

Re: [PATCH] Fix set of even probabilities (PR middle-end/89737).

[Jeff Law]

On 3/19/19 2:02 AM, Martin Liška wrote:
> Hi.
> 
> When calling set_even_probabilities, the function assumes that an edge
> can't live in both sets ({un,}likely_edges). When such situation happens,
> clear just the sets.
> 
> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
> 
> Ready to be installed?
> Thanks,
> Martin
> 
> gcc/ChangeLog:
> 
> 2019-03-18  Martin Liska  <mliska@suse.cz>
> 
> 	PR middle-end/89737
> 	* predict.c (combine_predictions_for_bb): Empty likely_edges and
> 	unlikely_edges if there's an edge that belongs to both these sets.
> 
> gcc/testsuite/ChangeLog:
> 
> 2019-03-18  Martin Liska  <mliska@suse.cz>
> 
> 	PR middle-end/89737
> 	* gcc.dg/pr89737.c: New test.
But is having the edge in both sets a valid state?  You know this code
better than I, so I'll go with your recommendation here.

jeff

Re: [PATCH] Fix set of even probabilities (PR middle-end/89737).

[Jan Hubicka]

> On 3/19/19 2:02 AM, Martin Liška wrote:
> > Hi.
> > 
> > When calling set_even_probabilities, the function assumes that an edge
> > can't live in both sets ({un,}likely_edges). When such situation happens,
> > clear just the sets.
> > 
> > Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
> > 
> > Ready to be installed?
> > Thanks,
> > Martin
> > 
> > gcc/ChangeLog:
> > 
> > 2019-03-18  Martin Liska  <mliska@suse.cz>
> > 
> > 	PR middle-end/89737
> > 	* predict.c (combine_predictions_for_bb): Empty likely_edges and
> > 	unlikely_edges if there's an edge that belongs to both these sets.
> > 
> > gcc/testsuite/ChangeLog:
> > 
> > 2019-03-18  Martin Liska  <mliska@suse.cz>
> > 
> > 	PR middle-end/89737
> > 	* gcc.dg/pr89737.c: New test.
> But is having the edge in both sets a valid state?  You know this code
> better than I, so I'll go with your recommendation here.

It is the situation you add both hot and cold attribute which I guess
can happen.
I think it is OK to ignore both hint then, so patch is OK.

Honza
> 
> jeff

Re: [PATCH] Fix set of even probabilities (PR middle-end/89737).

[Jakub Jelinek]

On Tue, Mar 19, 2019 at 09:02:27AM +0100, Martin Liška wrote:
> 2019-03-18  Martin Liska  <mliska@suse.cz>
> 
> 	PR middle-end/89737
> 	* predict.c (combine_predictions_for_bb): Empty likely_edges and
> 	unlikely_edges if there's an edge that belongs to both these sets.

I admit I don't know this code at all, just wonder if if you find
an intersection between likely_edges and unlikely_edges, couldn't you just
remove the intersection from both, rather than all other edges too?

	Jakub

Re: [PATCH] Fix set of even probabilities (PR middle-end/89737).

[Martin Liška]

On 3/19/19 3:52 PM, Jakub Jelinek wrote:
> On Tue, Mar 19, 2019 at 09:02:27AM +0100, Martin Liška wrote:
>> 2019-03-18  Martin Liska  <mliska@suse.cz>
>>
>> 	PR middle-end/89737
>> 	* predict.c (combine_predictions_for_bb): Empty likely_edges and
>> 	unlikely_edges if there's an edge that belongs to both these sets.
> 
> I admit I don't know this code at all, just wonder if if you find
> an intersection between likely_edges and unlikely_edges, couldn't you just
> remove the intersection from both, rather than all other edges too?

Having a contradict edge makes any assumption about the likeliness of the
others misleading. Plus it's very rare case this happens.

I'm going to install the patch.
Thanks,
Martin

> 
> 	Jakub
>