> On 05/21/2015 04:32 PM, Jan Hubicka wrote: > >I think at LTO time it is useful to have two things > > - be able to say what type comply C++ ODR rule, because we special case > > these > > for ODR warnings > > - be able to say at LTO time what types are anonymous, that is they are > > not > > compatible with any type from other unit. > > ...where the second group is a subset of the first; in languages > that use structural compatibility, anonymous types can be > compatible. > > >Either a special flag TYPE_ODR_P set by C++ FE on all types that comply ODR > >rule > >or making TYPE_PUBLIC (TYPE_STUB_DECL (t)) to reliably identify that type is > >anonymous (or an equivalent) would fix the two issues. > > It seems that we've been establishing which types have linkage by > setting their DECL_ASSEMBLER_NAME. It seems that the problem is > coming from leaving it unset for types with internal (anonymous) > linkage, so you're trying to check TREE_PUBLIC instead, which breaks > with other front ends. Perhaps we should set DECL_ASSEMBLER_NAME for > internal C++ types to some magic value?
This would also work, yes. We can set it into something like "<anonymous>". One problem would be that type_with_linkage_p/type_in_anonymous_namespace_p would not work on non-C++ types without LTO (because then we do not produce the type manglings). I suppose it is not really a problem because only place I use them is the devirtualization machinery and that won't get any polymorphic types for other languages. I suppose I can drop type_in_anonymous_namespcae_p in tree.c and make mangle_decl to set "anonymous" for those? Honza
