I really like this idea — I use it in a couple of you languages I maintain for work — but have never seen a good write up for the idea. Do you have any pointers to papers?
Basic block arguments are a good thing. Cranelift wrote some stuff on this and mlton is well documented. Huge semantic simplification in my opinion - the simultaneous evaluation hassle of phi nodes disappears, leaving behind needing to deal with calling conventions across jumps. Which are a hassle, but it's the same thing as a function call, and you've probably got those anyway.
