Sounds very interesting, I've occasionally wondered what it would be like to have a profiler "in the loop" which could direct improvements. Not everything can be known at compile-time, and even one run can differ greatly from another (ETL vs normal interactive use, or day-to-day load).