[PATCH] Improve SLP codegen, avoiding unnecessary TREE_ADDRESSABLE

[Richard Biener <rguenther@suse.de>]

The following adjusts vectorizer code generation to avoid splitting
out address increments for invariant addresses which causes objects
to get TREE_ADDRESSABLE when not necessary.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	* tree-vect-data-refs.cc (bump_vector_ptr): Return an
	invariant updated address when the input was invariant.
---
 gcc/tree-vect-data-refs.cc | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index 609cacc4971..b279a82551e 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -5195,6 +5195,14 @@ bump_vector_ptr (vec_info *vinfo,
 
   if (TREE_CODE (dataref_ptr) == SSA_NAME)
     new_dataref_ptr = copy_ssa_name (dataref_ptr);
+  else if (is_gimple_min_invariant (dataref_ptr))
+    /* When possible avoid emitting a separate increment stmt that will
+       force the addressed object addressable.  */
+    return build1 (ADDR_EXPR, TREE_TYPE (dataref_ptr),
+		   fold_build2 (MEM_REF,
+				TREE_TYPE (TREE_TYPE (dataref_ptr)),
+				dataref_ptr,
+				fold_convert (ptr_type_node, update)));
   else
     new_dataref_ptr = make_ssa_name (TREE_TYPE (dataref_ptr));
   incr_stmt = gimple_build_assign (new_dataref_ptr, POINTER_PLUS_EXPR,
-- 
2.35.3