https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55409
--- Comment #14 from Jonathan Wakely <redi at gcc dot gnu.org> --- (In reply to David Krauss from comment #13) > 1. Line 304: rebind isn’t an official member of allocator_traits. Prefer > rebind_alloc. That isn't allocator_traits, it's __gnu_cxx::__alloc_traits, which is a compatibility layer that works the same in C++03 and C++11. It has rebind, not rebind_alloc, see <ext/alloc_traits.h> > 2. _M_put_node is assuming no fancy pointers: its parameter is a > _Node_alloc_traits::pointer but its argument is a _Node*. I think the intent > with fancy pointers is that they get stored in nodes, I agree it's desirable, but it's not really clear that it's required, see http://cplusplus.github.io/LWG/lwg-active.html#2261 > but doing so would > change the ABI for fancy pointers that implicitly convert to raw. Perhaps > that’s acceptable, with these sweeping changes? These changes shouldn't affect the ABI. I have patches coming some time for fancy pointers in all containers, for PR57272, which also don't affect the ABI. My plan is to use the existing node types for allocators without fancy pointers, but use a new family of types (_List_ptr_node, _List_ptr_iterator etc.) that use fancy pointers. That will preserve ABI compatibility for non-fancy pointers, but also fully support fancy pointers. But it's a lot of work, due to the backward-compatibility requirements. > 3. Line 554: should call _Node_alloc_traits::construct. (This breaks > allocators which need a rebind to do construction, but so does the C++11 > standard. To be both backward- and forward-compatible would require > rebinding, converting, constructing via traits, converting back, and > assigning. According to my reading, anyway.) For what it’s worth, the > revision already uses _Node_alloc_traits::destroy, although that’s > customized less in practice. Line 504 is only compiled in C++98 mode, where allocators are required to provide a construct member.
