Hi!
This transformation turns it into a false or true if TREE_OVERFLOW
and TYPE_OVERFLOW_UNDEFINED, but if TREE_OVERFLOW is set, but not
TYPE_OVERFLOW_UNDEFINED, we leak the overflow bit into the IL, which then
confuses the VRP pass.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
2017-11-27 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/80788
* match.pd (X +- C1 CMP C2 -> X CMP C2 -+ C1): If res
has TREE_OVERFLOW set, call drop_tree_overflow.
* gcc.dg/pr80788.c: New test.
--- gcc/match.pd.jj 2017-11-24 22:38:53.000000000 +0100
+++ gcc/match.pd 2017-11-27 13:02:37.385280024 +0100
@@ -4393,7 +4393,8 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
&& TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
{ constant_boolean_node (cmp == NE_EXPR, type); }
(if (single_use (@3))
- (cmp @0 { res; }))))))))
+ (cmp @0 { TREE_OVERFLOW (res)
+ ? drop_tree_overflow (res) : res; }))))))))
(for cmp (lt le gt ge)
(for op (plus minus)
rop (minus plus)
--- gcc/testsuite/gcc.dg/pr80788.c.jj 2017-11-27 13:05:08.554378216 +0100
+++ gcc/testsuite/gcc.dg/pr80788.c 2017-11-27 13:04:42.000000000 +0100
@@ -0,0 +1,13 @@
+/* PR tree-optimization/80788 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fwrapv" } */
+
+void
+foo (signed char x)
+{
+ signed char a = (x + 1) ^ 128;
+ x &= !!a;
+ if (x != 0)
+ for (;;)
+ ;
+}
Jakub
