hint: Splay is a benchmark that penalizes generational GCs. If you have a fast mark-sweep GC with an aggressive growth policy you are going to totally rock it.
I may slightly be showing how long it's been since I looked at Splay. :)
My memory was that originally V8 was out-performing everyone else primarily because the initial allocation (into the nursery) was way quicker than anyone else did. Or maybe it was freeing the nodes that die young?
Certainly I remember both SpiderMonkey and Carakan spending huge amounts of time on GC on Splay, and Carakan's massive speed-up came when the object representation was changed (GC actually asymptotically regressed as a result, but in almost every practical case cache-locality (and reduction of memory accesses) outweighed it).