Looking at the outliers in terms of L1i MPKI makes me wonder how different the results might be under peak optimization with profile guidance, LTO, and BOLT. Does the "cactus" program have a hot region of code that's too large to effectively cache, or is it just poorly laid out by the toolchain?