@mjk@hachyderm.io For me the canonical "efficient stack machine" is the J1 Forth CPU: https://excamera.com/files/j1.pdf J1 has a dedicated hardware register for top-of-stack to avoid having to pump multiple arguments from the stack memory. On a modern OoO it's also worth thinking about the critical dependency paths: ALU -> store -> load -> ALU (with store->load forwarding) vs ALU -> zero-extend (ALU) -> ALU. The latter has lower latency.