>>105603497
You could try out interaction combinators, you compute it basically using a stack and it handles GC by itself, but it's entirely parallel and order agnotic so you can get some of the speedup you'd get from hardwiring instructions into the VM itself.
Also if you look up HVM you can see a version that hardwires a few more instructions like arithmatic and environments. So you can pretty easily start out completely vanilla and then arbitrarily keep adding more and more hardwired shortcuts without really changing the structure of the VM.