On Fri, Mar 19, 2021 at 10:01:50AM +0100, Aldy Hernandez via Gcc-patches wrote:
> --- a/gcc/value-range.cc
> +++ b/gcc/value-range.cc
> @@ -184,12 +184,46 @@ irange::irange_set (tree min, tree max)
> verify_range ();
> }
>
> +void
> +irange::irange_set_1bit_anti_range (tree min, tree max)
> +{
> + tree type = TREE_TYPE (min);
> + gcc_checking_assert (TYPE_PRECISION (type) == 1);
> +
> + if (operand_equal_p (min, max))
> + {
> + // Since these are 1-bit quantities, they can only be [MIN,MIN]
> + // or [MAX,MAX].
> + if (vrp_val_is_min (min))
> + min = max = vrp_val_max (type);
> + else
> + min = max = vrp_val_min (type);
> + set (min, max);
> + }
> + else
> + {
> + // The only alternative is [MIN,MAX], which is the empty range,
> + // but legacy code treats this as varying.
> + gcc_checking_assert (vrp_val_is_min (min));
> + gcc_checking_assert (vrp_val_is_max (max));
> + set_varying (type);
Doesn't this mean we handle the precision 1 ranges inconsistently with
the precision 2+ ranges? Because irange_set_anti_range for min equal
to minimum and max equal to maximum (i.e. VARYING) means
m_num_ranges = 0; (empty range).
> @@ -291,16 +325,10 @@ irange::set (tree min, tree max, value_range_kind kind)
> set_varying (type);
> return;
> }
> - else if (TYPE_PRECISION (TREE_TYPE (min)) == 1
> - && (is_min || is_max))
> + else if (TYPE_PRECISION (TREE_TYPE (min)) == 1)
> {
> - /* Non-empty boolean ranges can always be represented
> - as a singleton range. */
> - if (is_min)
> - min = max = vrp_val_max (TREE_TYPE (min));
> - else
> - min = max = vrp_val_min (TREE_TYPE (min));
> - kind = VR_RANGE;
> + irange_set_1bit_anti_range (min, max);
> + return;
This is in else branch of is_min && is_max which results in
varying, so irange_set_1bit_anti_range would not need to deal with that case
for this caller.
Jakub
