[gclist] zillion objects and gc

[Charles Fiterman]

>That only helps for known leaf objects.  When you drop the last
>reference to an object, the refcounter must scan it to find all the
>pointers, decrement all the associated reference counts, and recurse
>if necessary.  If this object was the root of a large structure,
>that's a lot of memory which a tracing collector would never have
>touched.  Even if you have code to avoid traversing leaf objects, this
>is still a big lose for reference counting.

In a strongly typed language you could isolate objects which could never be
part of a reference cycle. This works even better if the language has
special knowledge of some data structures.

What interests me even more is having data structures such as hash tables
etc. that scale well to multiprocessor use. That is if a lot of processors
are using the structure they aren't forced to go single file.