[jamesu]

Having implemented both a stack and a register-based VM I'd agree with the sentiment in the title. The register-based implementation (based partly off LUA 5 bytecode) was notably faster which I attributed to improved use of CPU cache combined with the instruction doing more work in per memory-fetch.

However I'd also point out the register-based VMs was far more difficult to debug, plus the compiler was much more long-winded to account for register allocation. Plus by the time you add on function dispatch overhead and real-world usage patterns, the gap where the register machine is technically better (i.e. besides just doing fibonacci) narrows somewhat.