Comment by stefanha

Thanks for the reply. The FAQ doesn't really answer my specific questions, but I understand that HVM is still young and there aren't answers to everything yet.

The FAQ mentions that every reduction rule can be compiled to machine code in theory. That's the core of what I'm asking - can reduction be implemented efficiently? I suspect it's not enough to have fast reduction rules, because that still requires a runtime that spends CPU cycles and memory accesses on reduction housekeeping. To be fast you need to minimize explicit reduction at runtime because that's not what CPUs are good at.

I guess an aggressive HVM compiler would find sub-programs that can be treated as leaf nodes and replaced entirely with compiled machine code. The machine code doesn't use interaction nets but it computes an equivalent result to the interaction net. This doesn't mean that reduction and interaction nets aren't used, they still serve a purpose for lazy evaluation and parallelism at higher levels of the program. The compiler would have to figure out where to draw the line between the high-level reduction runtime and the leaf nodes.

Anyway, this is just what comes to mind. I don't really know how HVM works, but thanks for sharing it.