Pool allocators, the Next Generation

[Gabriel Dos Reis]

Phil Edwards <phil@jaj.com> writes:

| A brief progress report:  One experimental version is done and working.
| It maintains parity in the current allocator settings, and allows us to
| delete pthread_allocimpl.h entirely.  A new typedef, __gthread_alloc, sits
| alongside __alloc and __single_client_alloc, with all the major features of
| pthread_alloc and the added bonus that it follows the --enable-threads=xxx
| decisions.
| 
| Bunch of other cleanups, e.g., more behind-the-scenes stuff moved from
| std into __gnu_cxx.
| 
| The major downside:  There are virtual function calls in the critical path.
| I still need to do lots of timing tests before I'll even begin to suggest
| this approach, and even then it would be with a plea for assistance to
| help get rid of them.

[Thinking loudly]  If the decision of which allocator category to use
is made at configure time or compile time, then I will suggest we use
traits to get rid of the virtual calls. 

-- Gaby