> > Given the new code for BIT_AND_EXPR in edge_info::derive_equivalences for
> > boolean types, I think that the same special treatment must be added for
> > boolean types in the BIT_NOT_EXPR case to preserve the 0-or-1-value
> > invariant.
> >
> > Bootstrapped/regtested on x86_64-suse-linux, OK for mainline and 8 branch?
>
> OK.
Thanks. However, as reported under PR tree-opt/89536, there is an annoying
oversight in the reasoning: the predicate to be used is not integer_zerop but
whether bit #0 is 0 or 1. I have applied the attached fixlet as obviously
more correct than the current code, but Jakub has a different opinion on the
whole change so this will probably be revisited in the near future.
PR tree-optimization/89536
* tree-ssa-dom.c (edge_info::derive_equivalences) <BIT_NOT_EXPR>: Test
only whether bit #0 of the value is 0 instead of the entire value.
* gcc.c-torture/execute/20190228-1.c: New test.
--
Eric Botcazou/* PR tree-optimization/89536 */
/* Testcase by Zhendong Su <s...@cs.ucdavis.edu> */
int a = 1;
int main (void)
{
a = ~(a && 1);
if (a < -1)
a = ~a;
if (!a)
__builtin_abort ();
return 0;
}
Index: tree-ssa-dom.c
===================================================================
--- tree-ssa-dom.c (revision 269211)
+++ tree-ssa-dom.c (working copy)
@@ -348,7 +348,7 @@ edge_info::derive_equivalences (tree nam
&& TREE_CODE (rhs) == SSA_NAME
&& ssa_name_has_boolean_range (rhs))
{
- if (integer_zerop (value))
+ if ((TREE_INT_CST_LOW (value) & 1) == 0)
res = build_one_cst (TREE_TYPE (rhs));
else
res = build_zero_cst (TREE_TYPE (rhs));
