The article is really targeting popcount for large bitvectors; in this scenario, call overhead is a non-issue (constant cost of <5-10 cycles vs O(n) work to be

The article is really targeting popcount for large bitvectors; in this scenario, call overhead is a non-issue (constant cost of <5-10 cycles vs O(n) work to be performed). I$ effects will also be negligible (the AVX2 instruction sequence is only a few cachelines long, and you wouldn't inline it).