2017-10-06 Richard Biener <rguenther@suse.de>

	PR tree-optimization/82436
	* tree-vect-slp.c (vect_supported_load_permutation_p): More
	conservatively choose the vectorization factor when checking
	whether we can perform the required load permutation.
	(vect_transform_slp_perm_load): Assert when we may not fail.

	* gcc.dg/vect/pr82436.c: New testcase.


git-svn-id: https://gcc.gnu.org/svn/gcc/trunk@253473 138bc75d-0d04-0410-961f-82ee72b054a4

4 files changed, 61 insertions, 5 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 38a86abaa39..06aaa06f75d 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,11 @@
+2017-10-06  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/82436
+	* tree-vect-slp.c (vect_supported_load_permutation_p): More
+	conservatively choose the vectorization factor when checking
+	whether we can perform the required load permutation.
+	(vect_transform_slp_perm_load): Assert when we may not fail.
+
 2017-10-05  Segher Boessenkool  <segher@kernel.crashing.org>
 
 	* config/rs6000/sysv4.h (SUBTARGET_OVERRIDE_OPTIONS): Correct error
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 9a3a5c8920a..dd39b78e7ab 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2017-10-06  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/82436
+	* gcc.dg/vect/pr82436.c: New testcase.
+
 2017-10-05  Bernd Edlinger  <bernd.edlinger@hotmail.de>
 
 	* gcc.dg/Wincompatible-pointer-types-1.c: New test.
diff --git a/gcc/testsuite/gcc.dg/vect/pr82436.c b/gcc/testsuite/gcc.dg/vect/pr82436.c
new file mode 100644
index 00000000000..4129e15447f
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr82436.c
@@ -0,0 +1,35 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-Ofast -fno-tree-scev-cprop" } */
+/* { dg-additional-options "-mavx2" { target { x86_64-*-* i?86-*-* } } } */
+
+struct reflection_type
+{
+  int h;
+  int k;
+  int l;
+  double f_exp;
+  double f_sigma;
+  _Complex double f_calc;
+  double f_pred;
+  double i_exp;
+  double i_sigma;
+  double i_pred;
+};
+
+double y, w;
+int foo (struct reflection_type *r, int n, unsigned s)
+{
+  int i;
+  y = 0;
+  w = 0;
+  for (i = 1; i < n; ++i)
+    {
+      struct reflection_type *x = &r[i*s];
+      double fpred = x->f_pred;
+      double fexp = x->f_exp;
+      double tem = (fpred - fexp);
+      y += __builtin_fabs (tem / x->f_sigma);
+      w += __builtin_fabs (tem / fexp);
+    }
+  return i;
+}
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index 5e40a374180..bc81b3d4b04 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -1629,14 +1629,20 @@ vect_supported_load_permutation_p (slp_instance slp_instn)
       return true;
     }
 
-  /* For loop vectorization verify we can generate the permutation.  */
+  /* For loop vectorization verify we can generate the permutation.  Be
+     conservative about the vectorization factor, there are permutations
+     that will use three vector inputs only starting from a specific factor
+     and the vectorization factor is not yet final.
+     ???  The SLP instance unrolling factor might not be the maximum one.  */
   unsigned n_perms;
+  unsigned test_vf
+    = least_common_multiple (SLP_INSTANCE_UNROLLING_FACTOR (slp_instn),
+			     LOOP_VINFO_VECT_FACTOR
+			       (STMT_VINFO_LOOP_VINFO (vinfo_for_stmt (stmt))));
   FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
     if (node->load_permutation.exists ()
-	&& !vect_transform_slp_perm_load
-	      (node, vNULL, NULL,
-	       SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true,
-	       &n_perms))
+	&& !vect_transform_slp_perm_load (node, vNULL, NULL, test_vf,
+					  slp_instn, true, &n_perms))
       return false;
 
   return true;
@@ -3613,6 +3619,7 @@ vect_transform_slp_perm_load (slp_tree node, vec<tree> dr_chain,
 		  dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM,
 				    stmt, 0);
 		}
+	      gcc_assert (analyze_only);
 	      return false;
 	    }
 
@@ -3636,6 +3643,7 @@ vect_transform_slp_perm_load (slp_tree node, vec<tree> dr_chain,
 			dump_printf (MSG_MISSED_OPTIMIZATION, "%d ", mask[i]);
 		      dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
 		    }
+		  gcc_assert (analyze_only);
 		  return false;
 		}