Speeding up GC

[Joe Buck]

> > This is kind of the "next level" of what I was proposing.  Once you
> > have this "enter short-term mode" thing, you can have the "exit
> > short-term" call just reset the state of all the temporary GC pools as
> > if GC has been run on them and no references had been found.
> 
> Just doesn't this bring all the bugs back that got fixed by the GC introduction?

You make the short-term area just a pool of the GC, and you have two
modes, settable by a flag.  In the production mode, when "exit short-term"
is called you just nuke the short-term pool.  In the GC-debug mode, when
exit short-term is called you do a GC and abort if there's anything live
in the short-term pool.  You then run the full regression suite in the
GC-debug mode.  GC-debug is shipped, so that if any user gets a mysterious
ICE you can see if this is the cause.