https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93582
--- Comment #35 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Unfortunately, it breaks miserably, e.g. miscompiles libcpp/macro.c.
Reduced testcase from that:
--- gcc/testsuite/gcc.c-torture/execute/pr93582.c.jj 2020-02-28
12:27:51.280925113 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr93582.c 2020-02-28
12:26:17.272332573 +0100
@@ -0,0 +1,22 @@
+/* PR tree-optimization/93582 */
+
+short a;
+int b, c;
+
+__attribute__((noipa)) void
+foo (void)
+{
+ b = c;
+ a &= 7;
+}
+
+int
+main ()
+{
+ c = 27;
+ a = 14;
+ foo ();
+ if (b != 27 || a != 6)
+ __builtin_abort ();
+ return 0;
+}
I believe the thing is that for the case with mask we need to completely avoid
the vn_reference_lookup_or_insert_for_pieces calls that finish method does, and
vn_reference_insert
call visit_reference_op_load does.
Probably instead of that we should call vn_nary_op_insert_stmt.
Now, I wonder if e.g. finish couldn't for the masked case just return the tree
itself rather than anything else, and let visit_reference_op_load not call
visit_reference_op_load if mask is non-NULL.
