re PR tree-optimization/69882 (Excessive reduction statements generated by SLP)

2016-02-22  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/69882
	* tree-vect-slp.c (vect_attempt_slp_rearrange_stmts): Properly
	preserve permutations present because of gaps.
	(vect_supported_load_permutation_p): Always continue checking
	permutations after vect_attempt_slp_rearrange_stmts.

	* gfortran.dg/vect/pr69882.f90: New testcase.

From-SVN: r233605

gcc/ChangeLog

+2016-02-22  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/69882
+	* tree-vect-slp.c (vect_attempt_slp_rearrange_stmts): Properly
+	preserve permutations present because of gaps.
+	(vect_supported_load_permutation_p): Always continue checking
+	permutations after vect_attempt_slp_rearrange_stmts.
+
 2016-02-22  Bin Cheng  <bin.cheng@arm.com>
 
 	* tree-vect-loop.c (vect_estimate_min_profitable_iters): Dump

gcc/testsuite/ChangeLog

+2016-02-22  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/69882
+	* gfortran.dg/vect/pr69882.f90: New testcase.
+
 2016-02-22  Jakub Jelinek  <jakub@redhat.com>
 
 	PR target/69885

gcc/testsuite/gfortran.dg/vect/pr69882.f90

+! { dg-additional-options "-Ofast" }
+! { dg-additional-options "-mavx" { target avx_runtime } }
+
+subroutine foo(a, x)
+  implicit none
+
+  integer, parameter :: XX=4, YY=26
+  integer, intent(in) :: x
+  real *8, intent(in) :: a(XX,YY)
+  real *8 :: c(XX)
+
+  integer i, k
+
+  c = 0
+
+  do k=x,YY
+     do i=1,2
+        c(i) = max(c(i), a(i,k))
+     end do
+  end do
+
+  PRINT *, "c=", c
+
+  IF (c(1) .gt. 0.0) THEN
+     CALL ABORT
+  END IF
+
+  IF (c(2) .gt. 0.0) THEN
+     CALL ABORT
+  END IF
+end subroutine foo
+
+PROGRAM MAIN
+  real *8 a(4, 26)
+
+  a = 0
+  a(3,1) = 100.0
+  a(4,1) = 100.0
+  
+  CALL FOO(a, 1)
+END PROGRAM

gcc/tree-vect-slp.c

@@ -1332,8 +1332,19 @@ vect_attempt_slp_rearrange_stmts (slp_instance slp_instn)
 			    node->load_permutation);
 
   /* We are done, no actual permutations need to be generated.  */
+  unsigned int unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_instn);
   FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
-    SLP_TREE_LOAD_PERMUTATION (node).release ();
+    {
+      gimple *first_stmt = SLP_TREE_SCALAR_STMTS (node)[0];
+      first_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt));
+      /* But we have to keep those permutations that are required because
+         of handling of gaps.  */
+      if (unrolling_factor == 1
+	  || (group_size == GROUP_SIZE (vinfo_for_stmt (first_stmt))
+	      && GROUP_GAP (vinfo_for_stmt (first_stmt)) == 0))
+	SLP_TREE_LOAD_PERMUTATION (node).release ();
+    }
+
   return true;
 }
 
@@ -1381,12 +1392,7 @@ vect_supported_load_permutation_p (slp_instance slp_instn)
      In reduction chain the order of the loads is not important.  */
   if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))
       && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
-    {
-      if (vect_attempt_slp_rearrange_stmts (slp_instn))
-	return true;
-
-      /* Fallthru to general load permutation handling.  */
-    }
+    vect_attempt_slp_rearrange_stmts (slp_instn);
 
   /* In basic block vectorization we allow any subchain of an interleaving
      chain.