Did you by any chance benchmark this? It seems to have quite the impact (at least 15% if not more) and to be enabled by default. I think we should recommend users of vroom to compile with -DNDEBUG

No, I didn't. I suspect that this is related to using an unordered_map to store the counts based on operator name:

https://github.com/VROOM-Project/vroom/blob/master/src/algorithms/local_search/local_search.cpp#L548

While this is constant-time-ish, there are a lot of occurrences of those updates. Setting aside the fact that users are already advised to build against tagged versions (so with -DNDEBUG enabled), I agree we can do better. Options include:

• store counts in plain std::arrays by assigning a rank to all operators (instead of a name), this would be a bit less straightforward to follow and maintain, especially for the count of applied moves which is the reason I went for name indexing in the first place;
• since this is only really required for dev purposes, put this behind another flag that one has to add explicitly.

We should probably do both.

Setting aside the fact that users are already advised to build against tagged versions (so with -DNDEBUG enabled),

Ah I wasn't aware of that. Then it is not as much of an issue as I thought.

I suspect that this is related to using an unordered_map to store the counts based on operator name

String hashing appears to be the bottleneck according to my profiling. I think a quick fix could be to replace the string names with an enumeration

enum OperatorName {
    UnassignedExchange,
    CrossExchange,
    ...
}

and an additional array to convert the enum-values to strings for printing.

@krypt-n I've sketched some changes in #667 based on your suggestion. Feel free to review it if you can!