[Patch] mt_allocator: spare mem & fix alignment problems

[B. Kosnik]

>the below does two, in principle unrelated, things:
>
>  1- Changes struct block_record to a union: in fact next and
>     thread_id are never used at the same time. The former only
>     when the block is unused and linked to the others of the
>     same thread, the latter only when used. This allows to spare
>     a good amount of memory (see attachments)
>
>  2- Adds a configurable _M_min_bin, which sets the smallest bin
>     size. The current 1 byte, frankly, doesn't make much sense
>     and, anyway, raising it to 8 bytes allows to fix the alignment
>     problems we were experiencing on ia64 (now the alignment is
>     the same used by default by pool_allocator).

Thanks. All this looks great!

>Overall, the performance seem slightly improved (see attachments), but
>perhaps the heuristics in deallocate should be tweaked again.

After you check this in I'll take it for a spin as well.

>Tested x86/x86_64/ia64 linux: now same results on all three, with the
>only unsolved problem (wrt new/pool-allocator) of pthread7-rope.

This is really great news!

I'm not quite sure what is up with pthread7-rope. Do you know if this
fails if std::string is used instead of rope? I'm just curious.

-benjamin