That'd just leave you with wasted space rather than making more space available for storing things inline. For the real-world tests I would not be at all surprised if the reduced number of heap allocations was a bigger source of speedups than the more cache-friendly reads.