OTOH it might be a good argument in favor: custom, completely
optimizable-away allocators might be the only way to go beyond GMP
performance.
There are some interface changes underway in GMP that eliminate some of
the more egregious allocation performance
