Hi!
Similar problem to the recently fixed UBSAN_VPTR lowering,
ubsan_expand_objsize_ifn sets *gsi to the first stmt in a new bb
after splitting block after UBSAN_OBJECT_SIZE, which is the next stmt
that should be processed, so we should always return no_next = true
to avoid gsi_next on it before it will be processed.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2015-02-12 Jakub Jelinek <ja...@redhat.com>
PR sanitizer/65019
* ubsan.c (ubsan_expand_objsize_ifn): Always return true.
* g++.dg/ubsan/pr65019.C: New test.
--- gcc/ubsan.c.jj 2015-02-10 22:58:55.000000000 +0100
+++ gcc/ubsan.c 2015-02-11 17:20:36.313063536 +0100
@@ -987,7 +987,7 @@ ubsan_expand_objsize_ifn (gimple_stmt_it
/* Get rid of the UBSAN_OBJECT_SIZE call from the IR. */
unlink_stmt_vdef (stmt);
gsi_remove (&gsi_orig, true);
- return gsi_end_p (*gsi);
+ return true;
}
/* Cached __ubsan_vptr_type_cache decl. */
--- gcc/testsuite/g++.dg/ubsan/pr65019.C.jj 2015-02-11 17:26:44.832959016
+0100
+++ gcc/testsuite/g++.dg/ubsan/pr65019.C 2015-02-11 17:26:23.000000000
+0100
@@ -0,0 +1,24 @@
+// PR sanitizer/65019
+// { dg-do compile }
+// { dg-options "-fsanitize=alignment,object-size,vptr -std=c++11 -O2
-fcompare-debug" }
+
+struct A { };
+struct B { };
+struct C final {
+ C (const A &, int);
+ static B *foo (const A &, int = 1);
+ virtual ~C ();
+ void *c;
+};
+
+B *
+C::foo (const A &x, int y)
+{
+ C *d = new C (x, y);
+ if (d->c == nullptr)
+ delete d;
+}
+
+C::~C ()
+{
+}
Jakub
