We are ICEing in set_range_info_raw because value_range_kind cannot be
VR_VARYING, since SSA_NAME_RANGE_TYPE can only hold VR_RANGE /
VR_ANTI_RANGE. Most of the time setting a VR_VARYING as a global
range makes no sense. However, we can have a range spanning the
entire domain (VR_RANGE of [MIN,MAX] which is essentially a
VR_VARYING), if the nonzero bits are set.
This was working before because set_range_info_raw allows setting
VR_RANGE of [MIN, MAX]. However, when going through an irange, we
normalize this to a VR_VARYING, thus causing the ICE. It's
interesting that other calls to set_range_info with an irange haven't
triggered this.
One solution would be to just ignore VR_VARYING and bail, since
set_range_info* is really an update of the current range semantic
wise. After all, we keep the nonzero bits which provide additional
info. But this would be a change in behavior, so not suitable until
after GCC 12 is released. So in order to keep with current behavior
we can just denormalize the varying to VR_RANGE.
Tested on x86-64 Linux.
PR tree-optimization/105432
gcc/ChangeLog:
* tree-ssanames.cc (set_range_info): Denormalize VR_VARYING to
VR_RANGE before passing a piecewise range to set_range_info_raw.
{
gcc_assert (!POINTER_TYPE_P (TREE_TYPE (name)));
- /* A range of the entire domain is really no range at all. */
tree type = TREE_TYPE (name);
+ if (range_type == VR_VARYING)
+ {
+ /* SSA_NAME_RANGE_TYPE can only hold a VR_RANGE or
+ VR_ANTI_RANGE. Denormalize VR_VARYING to VR_RANGE. */
+ range_type = VR_RANGE;
+ gcc_checking_assert (min == wi::min_value (type));
+ gcc_checking_assert (max == wi::max_value (type));
+ }
+
+ /* A range of the entire domain is really no range at all. */
if (min == wi::min_value (TYPE_PRECISION (type), TYPE_SIGN (type))
&& max == wi::max_value (TYPE_PRECISION (type), TYPE_SIGN (type)))
{