Re: libstdc++ PR 57272 Fancy pointer support in Hashtable

[Jonathan Wakely <jwakely@redhat.com>]

On 28/09/20 22:37 +0200, François Dumont via Libstdc++ wrote:
>Following recent changes on _Hashtable I rebase the patch and 
>completely review it.
>
>I managed to integrate the allocator custom pointer type without 
>touching to _Hashtable base types like _Hash_code_base or 
>_Hashtable_base. However I cannot see how to use the custom pointer 
>type without impacting the node types like _Hash_node_base which now 
>takes a template parameter, the custom pointer type.
>
>On an abi point of view node types are different however the data 
>structure is the same. The only difference is that the _Hash_node_base 
>_M_nxt is now a _Hash_node<> custom pointer rather than a simple 
>_Hash_node_base*.
>
>Even if this patch can't go in because of the abi breaking change I am 
>going to adapt some of the code simplifications for master. Especially 
>the _Hash_code_base and _Local_iterator_base simplifications.
>
>Let me know if you can think of a way to integrate the custom pointer 
>without impacting abi. Unless impacting node types and associated 
>iterator types is fine even if I already noticed that pretty printer 
>tests are broken with those changes.

The approach I used for the other containers (which was never
completed and committed) is something like:

struct _Node_base
{
   _Node_base* _M_next;
};

template<typename _Ptr>
struct _Fancy_node_base
{
   _Ptr _M_next;
};

template<typename _Ptr>
   using node_base = conditional_t<is_pointer<_Ptr>::value,
                                   _Node_base,
                                   _Fancy_node_base<_Ptr>>;

This way all existing code that has allocators with non-fancy pointers
continues to use the same type. Code using fancy pointers (which
doesn't currently work properly anyway) changes to use the new types
that depend on the pointer type.