Hi!
As mentioned in the PR, range_of_expr returns false if the type
of the expression isn't suitable for corresponding range type,
but doesn't if the range is undefined for other reasons. Still,
lower/upper_bound is defined only for ranges which actually have
at least one pair of subranges, VR_UNDEFINED range doesn't have it.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux,
ok for trunk?
2023-03-14 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/109115
* tree-vect-patterns.cc (vect_recog_divmod_pattern): Don't use
r.upper_bound () on r.undefined_p () range.
* gcc.dg/pr109115.c: New test.
--- gcc/tree-vect-patterns.cc.jj 2023-03-12 22:36:06.388177607 +0100
+++ gcc/tree-vect-patterns.cc 2023-03-13 22:49:18.278476093 +0100
@@ -3973,7 +3973,7 @@ vect_recog_divmod_pattern (vec_info *vin
/* Check that no overflow will occur. If we don't have range
information we can't perform the optimization. */
- if (ranger.range_of_expr (r, oprnd0, stmt))
+ if (ranger.range_of_expr (r, oprnd0, stmt) && !r.undefined_p ())
{
wide_int max = r.upper_bound ();
wide_int one = wi::shwi (1, prec);
--- gcc/testsuite/gcc.dg/pr109115.c.jj 2023-03-13 22:56:27.269428198 +0100
+++ gcc/testsuite/gcc.dg/pr109115.c 2023-03-13 22:56:04.174753778 +0100
@@ -0,0 +1,20 @@
+/* PR tree-optimization/109115 */
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+int a, b;
+
+int
+main ()
+{
+ unsigned short c = a, e = -1;
+ if (b)
+ {
+ unsigned d = (a ^ 1U) / a & c;
+ int f = (~d >> ~a) / e;
+ if (a)
+ f = a;
+ a = f;
+ }
+ return 0;
+}
Jakub
