Re: [PATCH] Fix IV-opts so it no longer produces MEM[index:]

[Zdenek Dvorak <rakdver@kam.mff.cuni.cz>]

Hello,

> Index: tree-ssa-loop-manip.c
> ===================================================================
> --- tree-ssa-loop-manip.c	(revision 127066)
> +++ tree-ssa-loop-manip.c	(working copy)
> @@ -56,12 +56,15 @@ create_iv (tree base, tree step, tree va
>    tree vb, va;
>    enum tree_code incr_op = PLUS_EXPR;
>    edge pe = loop_preheader_edge (loop);
> +  tree var1;
>  
>    if (!var)
>      {
>        var = create_tmp_var (TREE_TYPE (base), "ivtmp");
>        add_referenced_var (var);
>      }
> +  var1 = create_tmp_var (TREE_TYPE (step), "ivtmp");
> +  add_referenced_var (var1);
>  
>    vb = make_ssa_name (var, NULL_TREE);
>    if (var_before)
> @@ -104,7 +107,7 @@ create_iv (tree base, tree step, tree va
>      }
>    /* Gimplify the step if necessary.  We put the computations in front of the
>       loop (i.e. the step should be loop invariant).  */
> -  step = force_gimple_operand (step, &stmts, true, var);
> +  step = force_gimple_operand (step, &stmts, true, var1);

just change this to
step = force_gimple_operand (step, &stmts, true, NULL_TREE);

> @@ -1934,7 +1939,7 @@ add_candidate_1 (struct ivopts_data *dat
>      {
>        orig_type = TREE_TYPE (base);
>        type = generic_type_for (orig_type);
> -      if (type != orig_type)
> +      if (type != orig_type && !POINTER_TYPE_P (orig_type))
>  	{
>  	  base = fold_convert (type, base);
>  	  step = fold_convert (type, step);

You still need to convert base to type.

> Index: tree-affine.c
> ===================================================================
> --- tree-affine.c	(revision 127066)
> +++ tree-affine.c	(working copy)
> @@ -350,29 +349,40 @@ add_elt_to_tree (tree expr, tree type, t
>  		 aff_tree *comb)
>  {
>    enum tree_code code;
> +  tree type1 = type;
> +  if (POINTER_TYPE_P (type))
> +    type1 = sizetype;
>  
>    scale = double_int_ext_for_comb (scale, comb);
> -  elt = fold_convert (type, elt);
> +  elt = fold_convert (type1, elt);
>  
>    if (double_int_one_p (scale))
>      {
>        if (!expr)
> -	return elt;
> +	return fold_convert (type, elt);

These changes look wrong; for example, if aff_combination_to_tree is
called with combination with elements i and p (where type of p = int *
and type of i is sizetype), this would end up producing
(int *) i PPLUS (sizetype) p.  I think aff_combination_to_tree needs
to be rewritten to deal with the pointer type elements of the combination
separately.

Zdenek