Looks like clang is unhappy due to https://stackoverflow.com/a/46115028.

Note on benchmarks

TL;DR superseding the usual Exchange operator with the new SWAP*-adjusted implementation does improve results in most cases. This means generally reducing average/median/worst gap to optimal while stretching computing time by around 5 to 10% usually, which is definitely acceptable for the sake of quality improvement.

Results vary across exploration levels and benchmarks classes so I'll just report general trends.

CVRP

• More instances solved to full number of jobs
• Steady reduction of average optimal gap around -0.5 / -0.6 point

This may not sound like a big deal but it is significant since gaps are already quite low. For example with -x 5, average gap drops from 1.79% to 1.32% while median gap drops from 1.05% down to 0.68%.

MDVRP and HVRP

Not that many instances so probably not very significant, but same trend as above with more volatility in average gap reduction (from ~-0.1 to -0.8 point).

VRPTW

Steady reduction in costs for Homberger instances while average gap actually slightly increases for Solomon instances. The latter is a bit frustrating and a closer inspection shows that most Solomon instances are actually slightly improved but a few of them (in the R and RC classes) generate quite poorer solutions. I did not come up with a full explanation for this, other than the fact that the "random" instances are tougher to solve (especially with tight constraints) so any local search change is more prone to lead to getting trapped in a worse local minima.