Re: [PATCH] Improve cost computations in IVOPTS

[Zdenek Dvorak <rakdver at kam dot mff dot cuni dot cz>]

Hi,

> 2009-05-25  Eric Botcazou  <ebotcazou@adacore.com>
> 
> 	* tree-ssa-loop-ivopts.c (strip_offset_1) <MULT_EXPR>: New case.
> 	(force_expr_to_var_cost) <NEGATE_EXPR>: Likewise.
> 	(ptr_difference_cost): Use affine combinations to compute it.
> 	(difference_cost): Likewise.
> 	(get_computation_cost_at): Compute more accurate cost for addresses
> 	if the ratio is a multiplier allowed in addresses.
> 	For non-addresses, consider that an additional offset or symbol is
> 	added only once.

OK (with the changes proposed by Richard), and

> @@ -3415,8 +3447,8 @@ ptr_difference_cost (struct ivopts_data 
>  		     unsigned HOST_WIDE_INT *offset, bitmap *depends_on)
>  {
>    HOST_WIDE_INT diff = 0;
> -  comp_cost cost;
> -  bool speed = optimize_loop_for_speed_p (data->current_loop);
> +  aff_tree aff_e1, aff_e2;
> +  tree type;
>  
>    gcc_assert (TREE_CODE (e1) == ADDR_EXPR);
>  
> @@ -3434,12 +3466,14 @@ ptr_difference_cost (struct ivopts_data 
>  
>    *symbol_present = false;
>    *var_present = true;
> -  
> -  cost = force_var_cost (data, e1, depends_on);
> -  cost = add_costs (cost, force_var_cost (data, e2, depends_on));
> -  cost.cost += add_cost (Pmode, speed);
>  
> -  return cost;
> +  type = signed_type_for (TREE_TYPE (e1));
> +  tree_to_aff_combination (e1, type, &aff_e1);
> +  tree_to_aff_combination (e2, type, &aff_e2);
> +  aff_combination_scale (&aff_e2, double_int_minus_one);
> +  aff_combination_add (&aff_e1, &aff_e2);
> +
> +  return force_var_cost (data, aff_combination_to_tree (&aff_e1), depends_on);
>  }

assuming that compile time and memory consumption are not significantly
affected by this (and other similar) changes -- ivopts tries rather hard
to avoid folding things and creating new expressions in the phase of the
computation of the costs, as I used to run into problems with memory requirements
otherwise.

Zdenek