Re: [RFC] Bare bones of virtual call tracking

[Jan Hubicka <hubicka at ucw dot cz>]

> On 08/14/2013 02:14 AM, Jan Hubicka wrote:
> >As a temporary hack I suppose I can rely on assembler name of virtual table
> >to be unique for each templated class?
> 
> Actually, that seems like a fine solution for devirtualization; just
> compare the mangled name of the vtable to establish type identity.
OK,
does this seem to make sense?  I checked that it gets rid of my
false ODR violation warnings (I disabled the code for types not
having virtual tables).

So if it looks OK, I would like to go ahead with this patch.
(I am testing it now in isolation at x86_64-linux).

Honza

Index: tree.c
===================================================================
--- tree.c	(revision 201764)
+++ tree.c	(working copy)
@@ -11833,15 +11833,15 @@ types_same_for_odr (tree type1, tree typ
   if (type1 == type2)
     return true;
 
-  /* If types are not structuraly same, do not bother to contnue.
-     Match in the remainder of code would mean ODR violation.  */
-  if (!types_compatible_p (type1, type2))
-    return false;
-
+  /* Without LTO main variants of types are unique.  */
 #ifndef ENABLE_CHECKING
   if (!in_lto_p)
     return false;
 #endif
+   /* If types are not structuraly same, do not bother to contnue.
+      Match in the remainder of code would mean ODR violation.  */
+   if (!types_compatible_p (type1, type2))
+     return false;
 
   /* Check for anonymous namespaces. Those have !TREE_PUBLIC
      on the corresponding TYPE_STUB_DECL.  */
@@ -11852,6 +11852,34 @@ types_same_for_odr (tree type1, tree typ
 	  || !TREE_PUBLIC (TYPE_STUB_DECL (type2))))
     return false;
 
+  /* When assembler name of virtual table is available, it is
+     easy to compare types for equivalence.
+     FIXME: the code bellow consider all instantiations of the same
+     template to have same name.  This is because we have no access
+     to template parameters.
+     To avoid false positives that may lead to wrong devirtualizations,
+     compoare also representative virtual tables.
+     We can still return false positive positives for types without
+     virtual tables, but for the moment we do not care.  */
+  if (TYPE_BINFO (type1) && TYPE_BINFO (type2)
+      && BINFO_VTABLE (TYPE_BINFO (type1))
+      && BINFO_VTABLE (TYPE_BINFO (type2)))
+    {
+      tree v1 = BINFO_VTABLE (TYPE_BINFO (type1));
+      tree v2 = BINFO_VTABLE (TYPE_BINFO (type2));
+
+      if (TREE_CODE (v1) == POINTER_PLUS_EXPR)
+	{
+	  if (TREE_CODE (v2) != POINTER_PLUS_EXPR
+	      || !operand_equal_p (TREE_OPERAND (v1, 1),
+				   TREE_OPERAND (v2, 1), 0))
+	    return false;
+	  v1 = TREE_OPERAND (TREE_OPERAND (v1, 0), 0);
+	  v2 = TREE_OPERAND (TREE_OPERAND (v2, 0), 0);
+	}
+      return DECL_ASSEMBLER_NAME (v1) == DECL_ASSEMBLER_NAME (v2);
+    }
+
   if (!TYPE_NAME (type1))
     return false;
   if (!decls_same_for_odr (TYPE_NAME (type1), TYPE_NAME (type2)))