re PR ipa/65701 (r221530 makes 187.facerec drop with -Ofast -flto on bdver2)

2015-05-22  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/65701
	* tree-vect-data-refs.c (vect_enhance_data_refs_alignment):
	Move peeling cost models into one place.  Peel for alignment
	for single loads only if an aligned load is cheaper than
	an unaligned load.

From-SVN: r223528

gcc/ChangeLog

+2015-05-22  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/65701
+	* tree-vect-data-refs.c (vect_enhance_data_refs_alignment):
+	Move peeling cost models into one place.  Peel for alignment
+	for single loads only if an aligned load is cheaper than
+	an unaligned load.
+
 2015-05-22  Marek Polacek  <polacek@redhat.com>
 
 	PR c/47043

gcc/tree-vect-data-refs.c

@@ -1541,16 +1541,6 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
       || !slpeel_can_duplicate_loop_p (loop, single_exit (loop)))
     do_peeling = false;
 
-  /* If we don't know how many times the peeling loop will run
-     assume it will run VF-1 times and disable peeling if the remaining
-     iters are less than the vectorization factor.  */
-  if (do_peeling
-      && all_misalignments_unknown
-      && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
-      && (LOOP_VINFO_INT_NITERS (loop_vinfo)
-	  < 2 * (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 1))
-    do_peeling = false;
-
   if (do_peeling
       && all_misalignments_unknown
       && vect_supportable_dr_alignment (dr0, false))
@@ -1619,12 +1609,17 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
         }
 
       /* In case there are only loads with different unknown misalignments, use
-         peeling only if it may help to align other accesses in the loop.  */
+         peeling only if it may help to align other accesses in the loop or
+	 if it may help improving load bandwith when we'd end up using
+	 unaligned loads.  */
+      tree dr0_vt = STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr0)));
       if (!first_store
 	  && !STMT_VINFO_SAME_ALIGN_REFS (
 		  vinfo_for_stmt (DR_STMT (dr0))).length ()
-          && vect_supportable_dr_alignment (dr0, false)
-              != dr_unaligned_supported)
+	  && (vect_supportable_dr_alignment (dr0, false)
+	      != dr_unaligned_supported
+	      || (builtin_vectorization_cost (vector_load, dr0_vt, 0)
+		  == builtin_vectorization_cost (unaligned_load, dr0_vt, -1))))
         do_peeling = false;
     }
 
@@ -1641,14 +1636,6 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
 						   &body_cost_vec);
       if (!dr0 || !npeel)
         do_peeling = false;
-
-      /* If peeling by npeel will result in a remaining loop not iterating
-         enough to be vectorized then do not peel.  */
-      if (do_peeling
-	  && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
-	  && (LOOP_VINFO_INT_NITERS (loop_vinfo)
-	      < LOOP_VINFO_VECT_FACTOR (loop_vinfo) + npeel))
-	do_peeling = false;
     }
 
   if (do_peeling)
@@ -1733,6 +1720,7 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
 	    }
         }
 
+      /* Cost model #1 - honor --param vect-max-peeling-for-alignment.  */
       if (do_peeling)
         {
           unsigned max_allowed_peel
@@ -1757,6 +1745,18 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
             }
         }
 
+      /* Cost model #2 - if peeling may result in a remaining loop not
+	 iterating enough to be vectorized then do not peel.  */
+      if (do_peeling
+	  && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
+	{
+	  unsigned max_peel
+	    = npeel == 0 ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 1 : npeel;
+	  if (LOOP_VINFO_INT_NITERS (loop_vinfo)
+	      < LOOP_VINFO_VECT_FACTOR (loop_vinfo) + max_peel)
+	    do_peeling = false;
+	}
+
       if (do_peeling)
         {
           /* (1.2) Update the DR_MISALIGNMENT of each data reference DR_i.