I haven't read about the rest of your problem much, but I usually think in terms of _cache pressure_. A lookup table can be very fast, but adds to the cache pre

I haven't read about the rest of your problem much, but I usually think in terms of _cache pressure_. A lookup table can be very fast, but adds to the cache pressure. A large implementation with many special smart cases can also be very fast, but also adds to the cache pressure (instructions also need to be loaded from memory). Another thing I don't believe you mention is that not every n-instruction solution is equal (even if we only use single cycle instructions as possibilities). This is because of instruction-level parallelism. (a _ b) _ (c _ d) only takes two cycles to evaluate on a modern CPU. a _ (b _ (c _ d)) takes three.