On Tue, Nov 5, 2019 at 3:27 PM Richard Sandiford
<richard.sandif...@arm.com> wrote:
>
> vectorizable_assignment handles true SSA-to-SSA copies (which hopefully
> we don't see in practice) and no-op conversions that are required
> to maintain correct gimple, such as changes between signed and
> unsigned types. These cases shouldn't generate any code and so
> shouldn't count against either the scalar or vector costs.
>
> Later patches test this, but it seemed worth splitting out.
Hmm, but you have to adjust vect_compute_single_scalar_iteration_cost and
possibly the SLP cost walk as well, otherwise we're artificially making
those copies cheaper when vectorized.
>
> 2019-11-04 Richard Sandiford <richard.sandif...@arm.com>
>
> gcc/
> * tree-vect-stmts.c (vectorizable_assignment): Don't add a cost.
>
> Index: gcc/tree-vect-stmts.c
> ===================================================================
> --- gcc/tree-vect-stmts.c 2019-11-05 14:17:43.330141911 +0000
> +++ gcc/tree-vect-stmts.c 2019-11-05 14:18:39.169752725 +0000
> @@ -5305,7 +5305,7 @@ vectorizable_conversion (stmt_vec_info s
> static bool
> vectorizable_assignment (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
> stmt_vec_info *vec_stmt, slp_tree slp_node,
> - stmt_vector_for_cost *cost_vec)
> + stmt_vector_for_cost *)
> {
> tree vec_dest;
> tree scalar_dest;
> @@ -5313,7 +5313,6 @@ vectorizable_assignment (stmt_vec_info s
> loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
> tree new_temp;
> enum vect_def_type dt[1] = {vect_unknown_def_type};
> - int ndts = 1;
> int ncopies;
> int i, j;
> vec<tree> vec_oprnds = vNULL;
> @@ -5409,7 +5408,8 @@ vectorizable_assignment (stmt_vec_info s
> {
> STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
> DUMP_VECT_SCOPE ("vectorizable_assignment");
> - vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node,
> cost_vec);
> + /* Don't add a cost here. SSA copies and no-op conversions
> + shouldn't generate any code in either scalar or vector form. */
> return true;
> }
>
