GC, the simple solution

[Frank Benoit]

> * The cost of removing the last pointer to an object is unbounded
delete can be unbound? I didn't know that.

> * Even if the general latency is good, the total overhead of adjusting
> references is significantly greater than that of a tracing GC


> * It has a substantial space overhead, making it less useful for smaller
> heaps
In codesize yes, but in heap size?
I though the tracing GC does need much more heap because it should be
called very seldom.

> * Cyclic references (already mentioned)
> 
> I think these makes RC less suited for the general case, whereas it might
> fit very well in a tightly controlled system where low latency is required
> (which I'm well aware that you need ;).
> 
> IMO, the best general GC (the default) would be one which can handle as many
> cases as possible, and which is manipulable in those (preferably known)
> cases which it don't handle out of the box. I think that one need to
> combine features from several GC techniques (possibly also RC) to achieve
> this. 
> 

agreed.