Lambda the Ultimate

Folks,

Those who will find this totally OT will be kind enough to pardon me, but very sincerely, I would love this to be true, as I could then elaborate on a great PLT application we could make out of this, I'm really serious. But I don't want to bug people about the latter yet, without knowing if I can assume that sort of foundational formula I'd like to base the other idea upon.

So, first things first: can someone be kind to point me out where is the big mistake I made at the link above? It must be so big I still can't see it..

Thank you bunch in advance for your understanding and/or your help.
CJ

[Edit] Flaw found... Hmph. well, unsurprisingly.. Just made a quick update, though; it was a just first shot, anyway. I'll try have a couple others in the same strategy before I definitely surrender.

[Update @ 2010/12/03, 3pm PST] Second (or rather, third) attempt; I believe it's better... unless I missed something else again. Same link as above.

Dear all, this is an old subject which I touched upon trying to find a new compilation method for Hi code.

Common folklore has it that tracing collection is preferable over reference counting for functional language implementations due to the high allocation rate of nodes. So, my language has a Cheney-style garbage collector.

To get rid of some performance problems, and some technical concerns, I am considering switching to reference counting, but I am a bit wary of the performance impact.

Are there any solid studies out there which compare different schemes for garbage collection in an abstract manner?

(More concretely, it seems to me that there should be a tipping point where 'doing a lot of local work' must become cheaper than 'doing a lot of global work', but I don't have a good feeling where that tipping point would be. I.e, at 1MB, 1GB, 1TB of heap data?)