[gclist] ref-counting performance cost

[Simon Peyton-Jones]

| But this allocation cost would also be paid in a non-compacting 
| or conservative mark-and-sweep, not just reference counting
| (i.e., it is not a unique problem to RC unlike the problems
| of space, time and cycles).

Indeed, that is true.   But it's a problem that cannot be fixed for RC,
whereas it is relatively easy to adapt the mark/sweep story to do
compaction.   (Mind you, compaction involves changing pointers, which
is hard in a concurrent setting.)

| To quote some advantages of RC from the GC Book of Jones & Lins:

Don't get me wrong!  I'm *not* knocking RC!  It's an elegant technique with
many advantages, as you enumerate. I was saying "people often
seem to ignore the allocation cost of non-compacting techniques", not
"the allocation costs of non-compacting techniques make them useless".

Simon