Re: [PATCH] forwprop: Further fixes for simplify_rotate [PR108440]

[Aldy Hernandez <aldyh@redhat.com>]

On 1/19/23 09:41, Jakub Jelinek wrote:

> +	  range_query *q = get_range_query (cfun);
> +	  if (q == get_global_range_query ())
> +	    q = enable_ranger (cfun);

Oh, neat.  Clever.  I hadn't thought about that.

> +	  if (!q->range_of_expr (r, rotcnt, check_range_stmt))
> +	    {
> +	      if (check_range > 0)
> +		return false;
> +	      r.set_varying (TREE_TYPE (rotcnt));
> +	    }
>   	  int prec = TYPE_PRECISION (TREE_TYPE (rotcnt));
>   	  signop sign = TYPE_SIGN (TREE_TYPE (rotcnt));
>   	  wide_int min = wide_int::from (TYPE_PRECISION (rtype), prec, sign);
>   	  wide_int max = wide_int::from (wider_prec - 1, prec, sign);
> -	  int_range<2> r2 (TREE_TYPE (rotcnt), min, max);
> +	  if (check_range < 0)
> +	    max = min;
> +	  int_range<1> r2 (TREE_TYPE (rotcnt), min, max);
 >   	  r.intersect (r2);

Currently int_range<1> is a legacy range (anti ranges and such 
internally).  It's better to use <2> as the use of r2 will have to be 
converted to a multi-range before intersecting.  FYI, <2> is the 
smallest multi-range.

This is really an implementation detail, so don't bother changing it, 
even though it's slightly slower.  In the next release we'll nuke 
legacy, and <1> will mean what you think it means...the smallest range 
with one sub-range (and none of that anti range business internally).

Thanks.
Aldy