Created attachment 26886 [details]
gzip compressed preprocessed source

Strange, with the 4.7-RH cc1plus I can reproduce the segfault even on
inline void *operator new (__SIZE_TYPE__, void *p) throw ()
{
  return p;
}
struct B
{
  virtual ~B ();
  B ();
};
struct A : B
{
  A () : B () {}
  virtual void bar ();
};
void
foo ()
{
  char a[64];
  B *b = new (&a) A ();
  b->~B ();
}

(at -O2, both -m32 and -m64), but I can't reproduce it with a cross from x86_64-linux to powerpc64-linux.

Ah, #define HAVE_AS_GNU_ATTRIBUTE 1 in auto-host.h is essential for this.
This is devirt in action, for some reason the devirtualized A::~A () doesn't have DECL_EXTERNAL set (not sure what should set it, the FE, the devirtualizer?)
and likely the rs6000 backend just should be more forgiving and if it can't find a cgraph node, IMHO call_ABI_of_interest should just return true as a conservative answer.
So, for 4.7.0 IMHO best would be to do:
--- gcc/config/rs6000/rs6000.c 2012-03-13 19:58:59.342625117 +0100
+++ gcc/config/rs6000/rs6000.c 2012-03-13 23:14:15.693828006 +0100
@@ -7452,6 +7452,9 @@ call_ABI_of_interest (tree fndecl)
       /* Interesting functions that we are emitting in this object file.  */
       c_node = cgraph_get_node (fndecl);
       c_node = cgraph_function_or_thunk_node (c_node, NULL);
+      if (c_node == NULL)
+        return true;
+
       return !cgraph_only_called_directly_p (c_node);
     }
   return false;
and for 4.7.1/4.8 (while this can perhaps stay) we should find out why DECL_EXTERNAL isn't set.

Does

Index: gcc/gimple-fold.c
===================================================================
--- gcc/gimple-fold.c   (revision 185376)
+++ gcc/gimple-fold.c   (working copy)
@@ -143,6 +143,8 @@ canonicalize_constructor_val (tree cval)
          if (cfun && gimple_referenced_vars (cfun))
            add_referenced_var (base);
        }
+      else if (base && TREE_CODE (base) == FUNCTION_DECL)
+       cgraph_get_create_node (base);
       /* Fixup types in global initializers.  */
       if (TREE_TYPE (TREE_TYPE (cval)) != TREE_TYPE (TREE_OPERAND (cval, 0)))
        cval = build_fold_addr_expr (TREE_OPERAND (cval, 0));
@@ -3115,6 +3117,11 @@ gimple_get_virt_method_for_binfo (HOST_W
   if (!can_refer_decl_in_current_unit_p (fn))
     return NULL_TREE;
 
+  /* Make sure we create a cgraph node for functions we'll reference.
+     They can be non-existent if the reference comes from an entry
+     of an external vtable for example.  */
+  cgraph_get_create_node (fn);
+
   return fn;
 }
 
help?

Author: rguenth
Date: Wed Mar 14 13:01:46 2012
New Revision: 185381

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185381
Log:
2012-03-14  Richard Guenther  <rguenther@suse.de>

	PR middle-end/52582
	* gimple-fold.c (canonicalize_constructor_val): Make sure
	we have a cgraph node for a FUNCTION_DECL that comes from
	a constructor.
	(gimple_get_virt_method_for_binfo): Likewise.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/gimple-fold.c

Author: jakub
Date: Wed Mar 14 14:21:12 2012
New Revision: 185383

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185383
Log:
	PR c++/52582
	* config/rs6000/rs6000.c (call_ABI_of_interest): Return true
	if c_node is NULL.

	* g++.dg/opt/pr52582.C: New test.

Added:
    branches/gcc-4_7-branch/gcc/testsuite/g++.dg/opt/pr52582.C
Modified:
    branches/gcc-4_7-branch/gcc/ChangeLog
    branches/gcc-4_7-branch/gcc/config/rs6000/rs6000.c
    branches/gcc-4_7-branch/gcc/testsuite/ChangeLog

Author: jakub
Date: Wed Mar 14 14:24:32 2012
New Revision: 185384

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185384
Log:
	PR c++/52582
	* config/rs6000/rs6000.c (call_ABI_of_interest): Return true
	if c_node is NULL.

	* g++.dg/opt/pr52582.C: New test.

Added:
    trunk/gcc/testsuite/g++.dg/opt/pr52582.C
Modified:
    trunk/gcc/testsuite/ChangeLog

Temporarily worked around for 4.7.0.  We really should find out why we aren't setting DECL_EXTERNAL on this.

Created attachment 26893 [details]
patch

Does this fix it?

On this testcase?  Yes, it does.  Haven't done bootstrap/regtest with it though.

Bootstrapped/regtested now on x86_64-linux and i686-linux (trunk) and {x86_64,i686,powerpc,powerpc64}-linux (4.7 branch, with the rs6000.c change backed out).  No regressions.  So this looks ok for 4.8/4.7.1 to me, for 4.7.0 I think we should just keep the rs6000 hack.  Thanks.

Author: jason
Date: Thu Mar 15 21:22:38 2012
New Revision: 185443

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185443
Log:
	PR c++/52582
	* method.c (implicitly_declare_fn): Set DECL_EXTERNAL.

Added:
    trunk/gcc/testsuite/g++.dg/torture/pr52582.C
Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/method.c
    trunk/gcc/testsuite/ChangeLog

GCC 4.7.0 is being released, adjusting target milestone.

Author: jason
Date: Thu Mar 22 18:34:24 2012
New Revision: 185705

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185705
Log:
	PR c++/52582
	* method.c (implicitly_declare_fn): Set DECL_EXTERNAL.

Added:
    branches/gcc-4_7-branch/gcc/testsuite/g++.dg/torture/pr52582.C
Modified:
    branches/gcc-4_7-branch/gcc/cp/ChangeLog
    branches/gcc-4_7-branch/gcc/cp/method.c
    branches/gcc-4_7-branch/gcc/testsuite/ChangeLog

Fixed.