[RFC] Reduce memory with Key and Keys (and improve performance)

[Stephen J. Turnbull]

John Arbash Meinel writes:
 > My guess is that by not having these 900k object participating in the
 > garbage collector, we don't:
 >   1) Have as many garbage checks running. (IIRC gc only runs after so
 >      many gc objects have been allocated without being deallocated. But
 >      our objects aren't allocated as part of the gc set.)

It might be worth looking at reducing the frequency of GC and see how
much PeakWorking increases.

 >   2) A given cycle check is now cheaper, because it doesn't have to walk
 >      the 900k objects that won't be a part of a cycle anyway.

Obviously this is very important, so I would describe this as an
opportunity to grab even more speed at the expense of some of your
size gains, not as an alternative to further work on data structures
that allow more efficient GC.