Avoid unnecessary peeling for gaps with LD3

vectorizable_load forces peeling for gaps if the vectorisation factor
is not a multiple of the group size, since in that case we'd normally load
beyond the original scalar accesses but drop the excess elements as part
of a following permute:

          if (loop_vinfo
              && ! STMT_VINFO_STRIDED_P (stmt_info)
              && (GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0
                  || (!slp && vf % GROUP_SIZE (vinfo_for_stmt (first_stmt)) != 0)))

This isn't necessary for LOAD_LANES though, since it loads only the
data needed and does the permute itself.

Tested on aarch64-linux-gnu and x86_64-linux-gnu.

gcc/
	* tree-vect-stmts.c (vectorizable_load): Reorder checks so that
	load_lanes/grouped_load classification comes first.  Don't check
	whether the vectorization factor is a multiple of the group size
	for load_lanes.

gcc/testsuite/
	* gcc.dg/vect/vect-load-lanes-peeling-1.c: New test.

From-SVN: r236632

1 files changed, 12 insertions, 13 deletions

diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index f66e180..1252d33 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -6303,6 +6303,17 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt,
       gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info));
 
       first_stmt = GROUP_FIRST_ELEMENT (stmt_info);
+      group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
+
+      if (!slp
+	  && !PURE_SLP_STMT (stmt_info)
+	  && !STMT_VINFO_STRIDED_P (stmt_info))
+	{
+	  if (vect_load_lanes_supported (vectype, group_size))
+	    load_lanes_p = true;
+	  else if (!vect_grouped_load_supported (vectype, group_size))
+	    return false;
+	}
 
       /* If this is single-element interleaving with an element distance
          that leaves unused vector loads around punt - we at least create
@@ -6330,7 +6341,7 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt,
       if (loop_vinfo
 	  && ! STMT_VINFO_STRIDED_P (stmt_info)
 	  && (GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0
-	      || (!slp && vf % GROUP_SIZE (vinfo_for_stmt (first_stmt)) != 0)))
+	      || (!slp && !load_lanes_p && vf % group_size != 0)))
 	{
 	  if (dump_enabled_p ())
 	    dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
@@ -6350,8 +6361,6 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt,
       if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())
 	slp_perm = true;
 
-      group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
-
       /* ???  The following is overly pessimistic (as well as the loop
          case above) in the case we can statically determine the excess
 	 elements loaded are within the bounds of a decl that is accessed.
@@ -6364,16 +6373,6 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt,
 	  return false;
 	}
 
-      if (!slp
-	  && !PURE_SLP_STMT (stmt_info)
-	  && !STMT_VINFO_STRIDED_P (stmt_info))
-	{
-	  if (vect_load_lanes_supported (vectype, group_size))
-	    load_lanes_p = true;
-	  else if (!vect_grouped_load_supported (vectype, group_size))
-	    return false;
-	}
-
       /* Invalidate assumptions made by dependence analysis when vectorization
 	 on the unrolled body effectively re-orders stmts.  */
       if (!PURE_SLP_STMT (stmt_info)