Question

I’m learning about loop unrolling to avoid stalls caused by dependencies. I found many examples on internet and in literature, but I found no explanation on how the algorithm used to obtain the optimized code works (in case there is one such algorithm, of course). In particular, I don’t know how to determinate how many times should the loop be unrolled. Can it be calculated beforehand?

Answer 1

The rule of thumb is that you unwind so that:

• operations are done on “natural” boundries 4,8,16,32.. bytes
• you don’t introduce excessive register pressure (i.e. you don’t start juggling registers to memory)
• you don’t start repeating the exact same instruction sequence over and over

Basically you unroll as long as you can put more resources to work and you stop when you no longer can measure any performance gains.