tree-vect-stmts.c (get_group_load_store_type): Avoid peeling for gaps by loading…

tree-vect-stmts.c (get_group_load_store_type): Avoid peeling for gaps by loading only lower halves of vectors if possible.

2019-05-03  Richard Biener  <rguenther@suse.de>

	* tree-vect-stmts.c (get_group_load_store_type): Avoid
	peeling for gaps by loading only lower halves of vectors
	if possible.
	(vectorizable_load): Likewise.

	* gcc.dg/vect/slp-reduc-sad-2.c: New testcase.

From-SVN: r270847

gcc/ChangeLog

 2019-05-03  Richard Biener  <rguenther@suse.de>
 
+	* tree-vect-stmts.c (get_group_load_store_type): Avoid
+	peeling for gaps by loading only lower halves of vectors
+	if possible.
+	(vectorizable_load): Likewise.
+
+2019-05-03  Richard Biener  <rguenther@suse.de>
+
 	PR middle-end/89518
 	* match.pd: Add pattern to optimize (A / B) * B + (A % B) to A.
 

gcc/testsuite/ChangeLog

 2019-05-03  Richard Biener  <rguenther@suse.de>
 
+	* gcc.dg/vect/slp-reduc-sad-2.c: New testcase.
+
+2019-05-03  Richard Biener  <rguenther@suse.de>
+
 	PR middle-end/89518
 	* gcc.dg/pr89518.c: New testcase.
 

gcc/testsuite/gcc.dg/vect/slp-reduc-sad-2.c

+/* { dg-do compile } */
+/* { dg-require-effective-target vect_usad_char } */
+/* With AVX256 or more we do not pull off the trick eliding the epilogue.  */
+/* { dg-additional-options "-mprefer-avx128" { target { x86_64-*-* i?86-*-* } } } */
+
+typedef unsigned char uint8_t;
+int x264_pixel_sad_8x8( uint8_t *pix1, uint8_t *pix2, int i_stride_pix2 )
+{
+  int i_sum = 0;
+  for( int y = 0; y < 8; y++ )
+    {
+      i_sum += __builtin_abs( pix1[0] - pix2[0] );
+      i_sum += __builtin_abs( pix1[1] - pix2[1] );
+      i_sum += __builtin_abs( pix1[2] - pix2[2] );
+      i_sum += __builtin_abs( pix1[3] - pix2[3] );
+      i_sum += __builtin_abs( pix1[4] - pix2[4] );
+      i_sum += __builtin_abs( pix1[5] - pix2[5] );
+      i_sum += __builtin_abs( pix1[6] - pix2[6] );
+      i_sum += __builtin_abs( pix1[7] - pix2[7] );
+      pix1 += 16;
+      pix2 += i_stride_pix2;
+    }
+  return i_sum;
+}
+
+/* { dg-final { scan-tree-dump "vect_recog_sad_pattern: detected" "vect" } } */
+/* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" } } */
+/* { dg-final { scan-tree-dump-not "access with gaps requires scalar epilogue loop" "vect" } } */
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */

gcc/tree-vect-stmts.c

@@ -2258,6 +2258,29 @@ get_group_load_store_type (stmt_vec_info stmt_info, tree vectype, bool slp,
 	      && gap < (vect_known_alignment_in_bytes (first_dr_info)
 			/ vect_get_scalar_dr_size (first_dr_info)))
 	    overrun_p = false;
+
+	  /* If the gap splits the vector in half and the target
+	     can do half-vector operations avoid the epilogue peeling
+	     by simply loading half of the vector only.  Usually
+	     the construction with an upper zero half will be elided.  */
+	  dr_alignment_support alignment_support_scheme;
+	  scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype));
+	  machine_mode vmode;
+	  if (overrun_p
+	      && !masked_p
+	      && (((alignment_support_scheme
+		      = vect_supportable_dr_alignment (first_dr_info, false)))
+		   == dr_aligned
+		  || alignment_support_scheme == dr_unaligned_supported)
+	      && known_eq (nunits, (group_size - gap) * 2)
+	      && mode_for_vector (elmode, (group_size - gap)).exists (&vmode)
+	      && VECTOR_MODE_P (vmode)
+	      && targetm.vector_mode_supported_p (vmode)
+	      && (convert_optab_handler (vec_init_optab,
+					 TYPE_MODE (vectype), vmode)
+		  != CODE_FOR_nothing))
+	    overrun_p = false;
+
 	  if (overrun_p && !can_overrun_p)
 	    {
 	      if (dump_enabled_p ())
@@ -8516,8 +8539,24 @@ vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 		      }
 		    else
 		      {
+			tree ltype = vectype;
+			/* If there's no peeling for gaps but we have a gap
+			   with slp loads then load the lower half of the
+			   vector only.  See get_group_load_store_type for
+			   when we apply this optimization.  */
+			if (slp
+			    && loop_vinfo
+			    && !LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)
+			    && DR_GROUP_GAP (first_stmt_info) != 0
+			    && known_eq (nunits,
+					 (group_size
+					  - DR_GROUP_GAP (first_stmt_info)) * 2))
+			  ltype = build_vector_type (TREE_TYPE (vectype),
+						     (group_size
+						      - DR_GROUP_GAP
+						          (first_stmt_info)));
 			data_ref
-			  = fold_build2 (MEM_REF, vectype, dataref_ptr,
+			  = fold_build2 (MEM_REF, ltype, dataref_ptr,
 					 dataref_offset
 					 ? dataref_offset
 					 : build_int_cst (ref_type, 0));
@@ -8531,6 +8570,23 @@ vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 			  TREE_TYPE (data_ref)
 			    = build_aligned_type (TREE_TYPE (data_ref),
 						  TYPE_ALIGN (elem_type));
+			if (ltype != vectype)
+			  {
+			    vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr));
+			    tree tem = make_ssa_name (ltype);
+			    new_stmt = gimple_build_assign (tem, data_ref);
+			    vect_finish_stmt_generation (stmt_info, new_stmt, gsi);
+			    data_ref = NULL;
+			    vec<constructor_elt, va_gc> *v;
+			    vec_alloc (v, 2);
+			    CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem);
+			    CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
+						    build_zero_cst (ltype));
+			    new_stmt
+			      = gimple_build_assign (vec_dest,
+						     build_constructor
+						       (vectype, v));
+			  }
 		      }
 		    break;
 		  }