Haha ok. Cache optimization makes it pretty complicated to reason about true costs. Anyway, it’s not obvious to me even that the run lengths would need to be descending for best performace. I’m not even starting to think about how the merging order might affect galloping boosts on realistic data ;-). I didn’t get to that power thing at this point, but it looks like it’s unrelated to this consideration, except perhaps by chance. I might have time tonight to have a look at that. Surely we don’t want an algorithm that’s optimized for what they call ”Timsort drag” sequences in the literature ;-).