https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83158
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Indeed it's related. We now have range-info for
x_16 = x_15(D) + 2;
and thus:
Visiting conditional with predicate: if (x_16 == 0)
With known ranges
x_16: [-2147483646, +INF]
...
Intersecting
~[0, 0] EQUIVALENCES: { x_16 } (1 elements)
and
[-2147483646, +INF]
to
[-2147483646, +INF] EQUIVALENCES: { x_16 } (1 elements)
arguably the newly chosen range is smaller ...
we don't have the ability (limitation of the equivalence representation)
to keep ~[0, 0] as equivalence and it wouldn't help in this case given
the info is later supposed to be taken from SSA_NAME_RANGE_INFO.
We do special-case ~[0, 0] in the range intersection code but only for
pointer-type sized ranges...
/* Choose the anti-range if it is ~[0,0], that range is special
enough to special case when vr1's range is relatively wide. */
else if (*vr0min == *vr0max
&& integer_zerop (*vr0min)
&& (TYPE_PRECISION (TREE_TYPE (*vr0min))
== TYPE_PRECISION (ptr_type_node))
&& TREE_CODE (vr1max) == INTEGER_CST
&& TREE_CODE (vr1min) == INTEGER_CST
&& (wi::clz (wi::to_wide (vr1max) - wi::to_wide (vr1min))
< TYPE_PRECISION (TREE_TYPE (*vr0min)) / 2))
;
thus a "fix" is (I guess I'm ok with that if it doesn't regress any testcase
we have):
Index: gcc/tree-vrp.c
===================================================================
--- gcc/tree-vrp.c (revision 255167)
+++ gcc/tree-vrp.c (working copy)
@@ -6024,8 +6024,6 @@ intersect_ranges (enum value_range_type
enough to special case when vr1's range is relatively wide. */
else if (*vr0min == *vr0max
&& integer_zerop (*vr0min)
- && (TYPE_PRECISION (TREE_TYPE (*vr0min))
- == TYPE_PRECISION (ptr_type_node))
&& TREE_CODE (vr1max) == INTEGER_CST
&& TREE_CODE (vr1min) == INTEGER_CST
&& (wi::clz (wi::to_wide (vr1max) - wi::to_wide (vr1min))
