If maximizing streaming behavior, minimizing indirection, and minimizing passes over data is what you're looking for, you could do worse than to read papers from the 50's and 60's. Expensive memory meant tiny memories, which pretty much forced streaming, and high latencies pretty much forced minimizing passes. (I've read about a two-pass PDP assembler which required physically feeding the program tape through the reader twice! Tape for data also meant that indirection was minimized because everyone took great pains to bring relevant data together, which come to think of it was probably already habitual from the card era.)