Sat Oct 31 19:15:19 EDT 2015

A linear core

The basic idea is to embed a linear core inside a nonlinear one.

- For variables "seen" by the nonlinear core, the linear RC management
  has no effect.

- For priviate variables, the RC management accomplishes early release
  and locality.

The main question is: which types are RC-managed?

RC++ on copy to graph has another problem: an object might already be
in graph memory, so it's possible to multi-increment RCs.

It seems this idea is quite flawed..

It's probably best to completely separate the memory models, and copy
data on the boundaries.

The previous implementation for RCs was correct: one wrapper per copy,
so multiple copies have multiple wrappers, and each will decrement
refcount on GC.