This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

[PATCH] Fix aliasing miscompilation


Hi!

The following testcase is miscompiled with -O2 on i386. The reason is that
when the __restrict keyword is applied to the pointer to incomplete struct a,
get_alias_set sets up alias for struct a as whole but
record_component_aliases is never called on that type once it is layout out
fully (the testcase does not fail even without this patch if baz prototype
is moved after struct a full declaration or if the __restrict keyword is
removed from baz prototype).
The patch below seems to fix it. I'm not testing flag_strict_aliasing
because then the alias set is either 0 or -1.
Ok to commit?

2000-10-24  Jakub Jelinek  <jakub@redhat.com>

	* stor-layout.c (layout_type): Call record_component_aliases if
	TYPE_ALIAS_SET was already set on the incomplete type.

	* gcc.c-torture/execute/20001024-1.c: New test.

--- gcc/testsuite/gcc.c-torture/execute/20001024-1.c.jj	Tue Oct 24 13:32:08 2000
+++ gcc/testsuite/gcc.c-torture/execute/20001024-1.c	Tue Oct 24 10:16:57 2000
@@ -0,0 +1,34 @@
+struct a;
+
+extern int baz (struct a *__restrict x);
+
+struct a {
+  long v;
+  long w;
+};
+
+struct b {
+  struct a c;
+  struct a d;
+};
+
+int bar (int x, const struct b *__restrict y, struct b *__restrict z)
+{
+  if (y->c.v || y->c.w != 250000 || y->d.v || y->d.w != 250000)
+    abort();
+}
+
+void foo(void)
+{
+  struct b x;
+  x.c.v = 0;
+  x.c.w = 250000;
+  x.d = x.c;
+  bar(0, &x, ((void *)0));
+}
+
+int main()
+{
+  foo();
+  exit(0);
+}
--- gcc/stor-layout.c.jj	Mon Oct 23 15:24:47 2000
+++ gcc/stor-layout.c	Tue Oct 24 13:29:49 2000
@@ -1507,6 +1507,13 @@ layout_type (type)
      record it so set_sizetype can fix it up.  */
   if (! sizetype_set)
     early_type_list = tree_cons (NULL_TREE, type, early_type_list);
+
+  /* If an alias set has been set for this aggregate when it was incomplete,
+     we need to record alias sets for the fields now.  */
+  if (AGGREGATE_TYPE_P (type)
+      && TYPE_ALIAS_SET_KNOWN_P (type)
+      && TYPE_ALIAS_SET (type))
+    record_component_aliases (type);
 }
 
 /* Create and return a type for signed integers of PRECISION bits.  */

	Jakub

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]