[13/n] PR85694: Try to avoid vectorising casts of invariants

vect_recog_rotate_pattern had code to prevent operations
on invariants being vectorised unnecessarily:

  if (dt == vect_external_def
      && TREE_CODE (oprnd1) == SSA_NAME
      && is_a <loop_vec_info> (vinfo))
    {
      struct loop *loop = as_a <loop_vec_info> (vinfo)->loop;
      ext_def = loop_preheader_edge (loop);
      if (!SSA_NAME_IS_DEFAULT_DEF (oprnd1))
        {
          basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (oprnd1));
          if (bb == NULL
              || !dominated_by_p (CDI_DOMINATORS, ext_def->dest, bb))
            ext_def = NULL;
        }
    }
  [..]
      if (ext_def)
        {
          basic_block new_bb
            = gsi_insert_on_edge_immediate (ext_def, def_stmt);
          gcc_assert (!new_bb);
        }

This patch reuses the same idea for casts of invariants created
during widening optimisations.

One hitch was that vect_loop_versioning asserted that the vector loop
preheader was still empty, although the cfg transformation it's doing
should be correct either way.

2018-06-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* tree-vect-patterns.c (vect_get_external_def_edge): New function,
	split out from...
	(vect_recog_rotate_pattern): ...here.
	(vect_convert_input): Try to insert casts of invariants in the
	preheader.
	* tree-vect-loop-manip.c (vect_loop_versioning): Don't require the
	preheader to be empty.

gcc/testsuite/
	* gcc.dg/vect/vect-widen-mult-extern-1.c: New test.

From-SVN: r262277

1 files changed, 3 insertions, 2 deletions

diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c
index ea648f2..01d1850 100644
--- a/gcc/tree-vect-loop-manip.c
+++ b/gcc/tree-vect-loop-manip.c
@@ -3038,8 +3038,9 @@ vect_loop_versioning (loop_vec_info loop_vinfo,
 	 while we need to move it above LOOP's preheader.  */
       e = loop_preheader_edge (loop);
       scalar_e = loop_preheader_edge (scalar_loop);
-      gcc_assert (empty_block_p (e->src)
-		  && single_pred_p (e->src));
+      /* The vector loop preheader might not be empty, since new
+	 invariants could have been created while analyzing the loop.  */
+      gcc_assert (single_pred_p (e->src));
       gcc_assert (empty_block_p (scalar_e->src)
 		  && single_pred_p (scalar_e->src));
       gcc_assert (single_pred_p (condition_bb));