[gclist] Age based promotion

[Nick Barnes]

At 1998-12-17 16:10:43 UT, Simon Peyton-Jones writes:
> David
> 
> > Supposing you keep two flavors of remembered set, one for those
> > old objects modified to point to young (and hence likely to be
> > modified again to point to different young) and the other for
> > those objects promoted which happen to contain pointers to young?
> 
> Thanks for the algorithm you suggest.  By suggesting it, you imply that this
> issue is not well covered in the literature, which confirms what
> I thought.
> 
> > I don't think that transitive promotion is a good idea, since I can
> > imagine (and I think I've seen) certain programming styles for which
> > this would be a relatively disastrous policy.
> 
> Dave Ungar said this too.  Can either of you give some idea of what
> sort of programs would behave badly?  The kind where old things are
> modified often?

Compiler.  Immortal symbol table.  Transient information
(e.g. optimizer objects) which hangs off symbol table entries.

Nick B