Jason,
I got my firefox tree building again and noticed that my patch to enable
ODR merging on non-class types caused false positives:
/aux/hubicka/trunk-install/include/c++/6.0.0/ext/new_allocator.h:66:26:
warning: type ‘(anonymous namespace)::ObservationWithStack const&’ violates one
definition rule [-Wodr]
typedef const _Tp& const_reference;
^
/aux/hubicka/trunk-install/include/c++/6.0.0/ext/alloc_traits.h:110:53: note:
it is defined as a pointer to different type in another translation unit
typedef const value_type& const_reference;
^
/aux/hubicka/firefox8/xpcom/build/MainThreadIOLogger.cpp:28:8: note: type
‘const struct ObservationWithStack’ should match type ‘const struct value_type’
struct ObservationWithStack
^
/aux/hubicka/trunk-install/include/c++/6.0.0/ext/alloc_traits.h:103:53: note:
the incompatible type is defined here
typedef typename _Base_type::value_type value_type;
^
Here the obvious problem is that we try to merge type in anonymous namespace.
This is because at LTO time odr_type_p returns true and
type_in_anonymous_namespace returns false. odr_type_p basically checks that we
computed mangled type name for it which is done in free_lang_data as follows:
{
/* We use DECL_ASSEMBLER_NAME to hold mangled type names for One Definition
Rule
merging. */
if (flag_lto_odr_type_mering
&& TREE_CODE (decl) == TYPE_DECL
&& DECL_NAME (decl)
&& decl == TYPE_NAME (TREE_TYPE (decl))
&& !is_lang_specific (TREE_TYPE (decl))
/* Save some work. Names of builtin types are always derived from
properties of its main variant. A special case are integer types
where mangling do make differences between char/signed char/unsigned
char etc. Storing name for these makes e.g.
-fno-signed-char/-fsigned-char mismatches to be handled well.
See cp/mangle.c:write_builtin_type for details. */
&& (TREE_CODE (TREE_TYPE (decl)) != VOID_TYPE
&& TREE_CODE (TREE_TYPE (decl)) != BOOLEAN_TYPE
&& TREE_CODE (TREE_TYPE (decl)) != REAL_TYPE
&& TREE_CODE (TREE_TYPE (decl)) != FIXED_POINT_TYPE)
&& !TYPE_ARTIFICIAL (TREE_TYPE (decl))
&& !variably_modified_type_p (TREE_TYPE (decl), NULL_TREE)
&& !type_in_anonymous_namespace_p (TREE_TYPE (decl)))
return !DECL_ASSEMBLER_NAME_SET_P (decl);
and obviously we are not intended to mangle tyhis type and it should be
caught by type_in_anonymous_namespace_p check:
bool
type_in_anonymous_namespace_p (const_tree t)
{
/* TREE_PUBLIC of TYPE_STUB_DECL may not be properly set for
bulitin types; those have CONTEXT NULL. */
if (!TYPE_CONTEXT (t))
return false;
return (TYPE_STUB_DECL (t) && !TREE_PUBLIC (TYPE_STUB_DECL (t)));
}
We already discussed earlier that type_in_anonymous_namespace_p is not working
on compund types, because these do not have TYPE_STUB_DECL. I tought those are
!TYPE_NAME types. What is reason for !TYPE_NAME type with no TYPE_STUB_DECL?
Is it always a compound type with typedef name?
One possible solution is to avoid producing mangled names so these won't
be odr_type_p at LTO time:
Index: tree.c
===================================================================
--- tree.c (revision 222991)
+++ tree.c (working copy)
@@ -5173,6 +5173,7 @@ need_assembler_name_p (tree decl)
&& TREE_CODE (TREE_TYPE (decl)) != REAL_TYPE
&& TREE_CODE (TREE_TYPE (decl)) != FIXED_POINT_TYPE)
&& !TYPE_ARTIFICIAL (TREE_TYPE (decl))
+ && TYPE_STUB_DECL (TREE_TYPE (decl))
&& !variably_modified_type_p (TREE_TYPE (decl), NULL_TREE)
&& !type_in_anonymous_namespace_p (TREE_TYPE (decl)))
return !DECL_ASSEMBLER_NAME_SET_P (decl);
which silence the warnings, but I think it also may cause false negatives.
Ohter option is to make type_in_anonymous_namespace_p work. The types in
question are anonymous because they are defined within anonymous template,
would the following work in type_in_anonymous_namespace_p?
+ /* Types (such as pointer_type) defined within class types may not
+ have their own TYPE_STUB_DECL. Look for the outer type. */
+ while (!TYPE_STUB_DECL (t) && TYPE_NAME (t)
+ && TREE_CODE (TYPE_NAME (t)) == TYPE_DECL
+ && DECL_CONTEXT (TYPE_NAME (t))
+ && TYPE_P (DECL_CONTEXT (TYPE_NAME (t))))
+ t = DECL_CONTEXT (TYPE_NAME (t));
Last option would be make anonymous_namespace_p to walk to compound type bases.
I made variant of this in
https://gcc.gnu.org/ml/gcc-patches/2015-05/msg00880.html
(odr_or_derived_type_p, modifying type_in_anonymous_namespace_p to work same
way is definitly easy)
What would be a preffered fix for this?
Honza
