Would it be possible to re-derive some of the LRS info by analyzing the final code? The less information you need to carry through the optimizations, the better, it seems to me.