> The additional refcount takes space that could be used for the payload that now either needs to spill into the next cacheline or needs to be compressed further.

It could, but since the cachelines are adjacents the prefetcher will pull subsequent cache lines anyways, and since the objects typically used in doubly-linked lists are very large (much larger than 64-bit), then this doesn't matter on any practical application that I am aware of. Our results of 4% performance increase is for the worst-case that we found. Usually is 1% or less.

If you have a real-world benchmark that shows otherwise, can you please share it?

AFAIK there does not exist any hardware for which the performance model for:

- non-intrusive doubly-linked lists,

- typical object sizes used on these lists (>> 64-bit)

- typical operations done on these lists (splice, etc.)

would predict a performance difference, and in our case, having replaced unsafe lists with safe list in many large applications, we never were able to measure a difference in application performance, only in the synthetic worst-case micro-benchmarks.

So I am truly interested to learn about your use case.