[PATCH] tree-optimization/106860 - fix profile scaling in split_loop

[PATCH] tree-optimization/106860 - fix profile scaling in split_loop

[Richard Biener]

The following fixes a mistake in loop splitting which assumes loop
latches have a single predecessor and that edge is from the exit
test.  Instead work from the single exit edge we have to find the
edge towards the latch.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	PR tree-optimization/106860
	* tree-ssa-loop-split.cc (split_loop): Find the exit to
	latch edge from the loop exit edge instead of from the
	latch.  Verify we're going to find it.

	* g++.dg/opt/pr106860.C: New testcase.
---
 gcc/testsuite/g++.dg/opt/pr106860.C | 23 +++++++++++++++++++++++
 gcc/tree-ssa-loop-split.cc          | 16 ++++++++++------
 2 files changed, 33 insertions(+), 6 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/opt/pr106860.C

diff --git a/gcc/testsuite/g++.dg/opt/pr106860.C b/gcc/testsuite/g++.dg/opt/pr106860.C
new file mode 100644
index 00000000000..a0209dcf9da
--- /dev/null
+++ b/gcc/testsuite/g++.dg/opt/pr106860.C
@@ -0,0 +1,23 @@
+// { dg-do compile }
+// { dg-options "-Ofast -ftrapv -fnon-call-exceptions -fno-tree-fre" }
+
+static const int N = 12;
+int nSlip;
+
+int main ()
+{
+  int i, j, k, fdot = 0;
+  int a[N][N];
+
+  for ( i = 1; i < nSlip; i++)
+    {
+      for ( j = i+1; j < nSlip; j++)
+        {
+          for ( k = 0; k < i; k++)
+            fdot += a[i][k] * a[k][j];
+          a[i][j] = a[i][j] - fdot;
+        }
+   }
+
+  return 0;
+}
diff --git a/gcc/tree-ssa-loop-split.cc b/gcc/tree-ssa-loop-split.cc
index bccf621493b..fad4e8361b0 100644
--- a/gcc/tree-ssa-loop-split.cc
+++ b/gcc/tree-ssa-loop-split.cc
@@ -531,16 +531,17 @@ split_loop (class loop *loop1)
   tree guard_iv;
   tree border = NULL_TREE;
   affine_iv iv;
+  edge exit1;
 
-  if (!single_exit (loop1)
+  if (!(exit1 = single_exit (loop1))
+      || EDGE_COUNT (exit1->src->succs) != 2
       /* ??? We could handle non-empty latches when we split the latch edge
 	 (not the exit edge), and put the new exit condition in the new block.
 	 OTOH this executes some code unconditionally that might have been
 	 skipped by the original exit before.  */
       || !empty_block_p (loop1->latch)
       || !easy_exit_values (loop1)
-      || !number_of_iterations_exit (loop1, single_exit (loop1), &niter,
-				     false, true)
+      || !number_of_iterations_exit (loop1, exit1, &niter, false, true)
       || niter.cmp == ERROR_MARK
       /* We can't yet handle loops controlled by a != predicate.  */
       || niter.cmp == NE_EXPR)
@@ -644,10 +645,13 @@ split_loop (class loop *loop1)
 	fix_loop_bb_probability (loop1, loop2, true_edge, false_edge);
 
 	/* Fix first loop's exit probability after scaling.  */
-	edge exit_to_latch1 = single_pred_edge (loop1->latch);
+	edge exit_to_latch1;
+	if (EDGE_SUCC (exit1->src, 0) == exit1)
+	  exit_to_latch1 = EDGE_SUCC (exit1->src, 1);
+	else
+	  exit_to_latch1 = EDGE_SUCC (exit1->src, 0);
 	exit_to_latch1->probability *= true_edge->probability;
-	single_exit (loop1)->probability
-	  = exit_to_latch1->probability.invert ();
+	exit1->probability = exit_to_latch1->probability.invert ();
 
 	/* Finally patch out the two copies of the condition to be always
 	   true/false (or opposite).  */
-- 
2.37.3