On Thu, Aug 11, 2016 at 04:06:10PM +0530, Prathamesh Kulkarni wrote:
> +static void
> +replace_call_by_cmp (gimple_stmt_iterator *gsi, location_t loc,
> + tree type, tree arg1, tree arg2,
> + tree res_type, enum tree_code cmp_code)
> +{
> + tree ptrtype = build_pointer_type_for_mode (char_type_node, ptr_mode,
> true);
> + tree off = build_int_cst (ptrtype, 0);
> + arg1 = build2_loc (loc, MEM_REF, type, arg1, off);
> + tree tem1 = fold_const_aggregate_ref (arg1);
> + if (tem1)
> + arg1 = tem1;
> +
> + if (POINTER_TYPE_P (TREE_TYPE (arg2)))
> + {
> + arg2 = build2_loc (loc, MEM_REF, type, arg2, off);
> + tree tem2 = fold_const_aggregate_ref (arg2);
> + if (tem2)
> + arg2 = tem2;
> + }
> + else
> + gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (arg2)));
> +
> + tree res = fold_convert_loc (loc, res_type,
> + fold_build2_loc (loc, cmp_code,
> + boolean_type_node,
> + arg1, arg2));
> +
> + gimplify_and_update_call_from_tree (gsi, res);
I know it is pre-existing, but do you really need to create everything as
trees first and then gimplify? Can't you emit GIMPLE right away?
> @@ -2505,7 +2558,7 @@ const pass_data pass_data_strlen =
> 0, /* properties_provided */
> 0, /* properties_destroyed */
> 0, /* todo_flags_start */
> - 0, /* todo_flags_finish */
> + TODO_update_ssa_only_virtuals, /* todo_flags_finish */
> };
>
> class pass_strlen : public gimple_opt_pass
And if you do, I'd hope you can avoid this. The strlen call has
a vuse, just use the same virtual operand in vuse on the mem read.
Jakub
