re PR tree-optimization/68892 (Excessive dead loads produced by BB vectorization)

2015-12-16  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/68892
	* tree-vect-slp.c (vect_analyze_slp_cost_1): Properly compute
	cost for permuted loads.

	* gcc.dg/vect/bb-slp-pr68892.c: New testcase.

From-SVN: r231674

4 files changed, 48 insertions, 27 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index c68fd13..6450d52 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2015-12-16  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/68892
+	* tree-vect-slp.c (vect_analyze_slp_cost_1): Properly compute
+	cost for permuted loads.
+
 2015-12-16  Jakub Jelinek  <jakub@redhat.com>
 
 	PR rtl-optimization/65980
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index a0b8fda..7a3c5e4 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2015-12-16  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/68892
+	* gcc.dg/vect/bb-slp-pr68892.c: New testcase.
+
 2015-12-16  Jakub Jelinek  <jakub@redhat.com>
 
 	PR rtl-optimization/65980
diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr68892.c b/gcc/testsuite/gcc.dg/vect/bb-slp-pr68892.c
new file mode 100644
index 0000000..648fe481
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr68892.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-fvect-cost-model=dynamic" } */
+
+double a[128][128];
+double b[128];
+
+void foo(void)
+{
+  b[0] = a[0][0];
+  b[1] = a[1][0];
+  b[2] = a[2][0];
+  b[3] = a[3][0];
+}
+
+/* { dg-final { scan-tree-dump "not profitable" "slp2" } } */
+/* { dg-final { scan-tree-dump-times "Basic block will be vectorized" 0 "slp2" } } */
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index 3fdc988..f57c859 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -1402,10 +1402,9 @@ vect_analyze_slp_cost_1 (slp_instance instance, slp_tree node,
 {
   unsigned i, j;
   slp_tree child;
-  gimple *stmt, *s;
+  gimple *stmt;
   stmt_vec_info stmt_info;
   tree lhs;
-  unsigned group_size = SLP_INSTANCE_GROUP_SIZE (instance);
 
   /* Recurse down the SLP tree.  */
   FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
@@ -1424,44 +1423,39 @@ vect_analyze_slp_cost_1 (slp_instance instance, slp_tree node,
 			       node, prologue_cost_vec, body_cost_vec);
       else
 	{
-	  int i;
 	  gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)));
-	  /* If the load is permuted then the alignment is determined by
-	     the first group element not by the first scalar stmt DR.  */
 	  if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
 	    {
+	      /* If the load is permuted then the alignment is determined by
+		 the first group element not by the first scalar stmt DR.  */
 	      stmt = GROUP_FIRST_ELEMENT (stmt_info);
 	      stmt_info = vinfo_for_stmt (stmt);
+	      /* Record the cost for the permutation.  */
+	      record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
+				stmt_info, 0, vect_body);
+	      /* And adjust the number of loads performed.  */
+	      unsigned nunits
+		= TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
+	      ncopies_for_cost
+	        = (GROUP_SIZE (stmt_info) - GROUP_GAP (stmt_info)
+		   + nunits - 1) / nunits;
+	      ncopies_for_cost *= SLP_INSTANCE_UNROLLING_FACTOR (instance);
 	    }
+	  /* Record the cost for the vector loads.  */
 	  vect_model_load_cost (stmt_info, ncopies_for_cost, false,
 				node, prologue_cost_vec, body_cost_vec);
-	  /* If the load is permuted record the cost for the permutation.
-	     ???  Loads from multiple chains are let through here only
-	     for a single special case involving complex numbers where
-	     in the end no permutation is necessary.  */
-	  FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, s)
-	    if ((STMT_VINFO_GROUP_FIRST_ELEMENT (vinfo_for_stmt (s))
-		 == STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info))
-		&& vect_get_place_in_interleaving_chain
-		     (s, STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info)) != i)
-	      {
-		record_stmt_cost (body_cost_vec, group_size, vec_perm,
-				  stmt_info, 0, vect_body);
-		break;
-	      }
 	}
+      return;
     }
-  else
+
+  record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
+		    stmt_info, 0, vect_body);
+  if (SLP_TREE_TWO_OPERATORS (node))
     {
       record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
 			stmt_info, 0, vect_body);
-      if (SLP_TREE_TWO_OPERATORS (node))
-	{
-	  record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
-			    stmt_info, 0, vect_body);
-	  record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
-			    stmt_info, 0, vect_body);
-	}
+      record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
+			stmt_info, 0, vect_body);
     }
 
   /* Push SLP node def-type to stmts.  */