On Tue, Mar 08, 2016 at 07:11:45PM +0100, Richard Biener wrote:
> I believe the safest fix is to re-instantiate the compatibility check by
> refactoring operand_equal_p to perform it on the full ref (but not recursions
> where it would be redundant and maybe too conservative).
> I've noticed this as well when doing the last operand_equal_p surgery, esp.
> The incomplete and bogus half-way type checking done at its top.
I've tried to add types_compatible_p check to operand_equal_p for all
toplevel expressions, but that affected 25x more operand_equal_p calls
during x86_64 and i686-linux bootstrap/regtest than just doing it
for *MEM_REF only - details in the PR; after discussions on IRC with Richard
and Honza I've committed reversion of the October change, and we'll need to
start with operand_equal_p changes early during stage1 next time, rather
than at the end of stage1.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.
2016-03-09 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/70127
* fold-const.c (operand_equal_p): Revert the 2015-10-28 change.
* gcc.c-torture/execute/pr70127.c: New test.
--- gcc/fold-const.c.jj 2016-03-09 15:06:21.000000000 +0100
+++ gcc/fold-const.c 2016-03-09 18:25:07.429926750 +0100
@@ -3032,6 +3032,9 @@ operand_equal_p (const_tree arg0, const_
TYPE_SIZE (TREE_TYPE (arg1)),
flags)))
return 0;
+ /* Verify that access happens in similar types. */
+ if (!types_compatible_p (TREE_TYPE (arg0), TREE_TYPE (arg1)))
+ return 0;
/* Verify that accesses are TBAA compatible. */
if (!alias_ptr_types_compatible_p
(TREE_TYPE (TREE_OPERAND (arg0, 1)),
--- gcc/testsuite/gcc.c-torture/execute/pr70127.c.jj 2016-03-08
12:11:11.890835632 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr70127.c 2016-03-08
12:10:58.000000000 +0100
@@ -0,0 +1,23 @@
+/* PR tree-optimization/70127 */
+
+struct S { int f; signed int g : 2; } a[1], c = {5, 1}, d;
+short b;
+
+__attribute__((noinline, noclone)) void
+foo (int x)
+{
+ if (x != 1)
+ __builtin_abort ();
+}
+
+int
+main ()
+{
+ while (b++ <= 0)
+ {
+ struct S e = {1, 1};
+ d = e = a[0] = c;
+ }
+ foo (a[0].g);
+ return 0;
+}
Jakub
