Adaptive Instruction Scheduling for the Itanium

Many problems that compilers want to solve are NP-Hard to solve optimally. Therefore, compilers authors typically use fast ad-hoc methods that strike a trade-off between good code quality and fast compile time. We investigated taking the decision on the trade-off out of the compiler author’s hands: based on profile information or static estimates, the compiler will tilt the balance in favour of code quality on hot segments of code, compensating by going for fast compile time on cold segments. The particular problem we decided to attack was instruction scheduling on the Itanium. We find inconclusive results, but we have an excuse: our machine model was inaccurate.