Profiling optimised Haskell - causal analysis and implementation

At the present time, performance optimisation of real-life Haskell programs is a bit of a “black art”. Programmers that can do so reliably are highly esteemed, doubly so if they manage to do it without sacrificing the character of the language by falling back to an “imperative style”. The reason is that while programming at a high-level does not need to result in slow performance, it must rely on a delicate mix of optimisations and transformations to work out just right. Predicting how all these cogs will turn is hard enough – but where something goes wrong, the various transformations will have mangled the program to the point where even finding the crucial locations in the code can become a game of cat-and-mouse. In this work we will lift the veil on the performance of heavily transformed Haskell programs: Using a formal causality analysis we will track source code links from square one, and maintain the connection all the way to the final costs generated by the program. This will allow us to implement a profiling solution that can measure performance at high accuracy while explaining in detail how we got to the point in question. Furthermore, we will directly support the performance analysis process by developing an interactive profiling user interface that allows rapid theory forming and evaluation, as well as deep analysis where required.