Hi Immanuel,
Fix follows below (MIT license). As I'm trying to understand the design descisions, what speaks against the use of size_t in that chain of functions, when the only variable (except of the const 1) used as a parameter is of size_t? Or the other way around: why is the count in nn_dist of size_t when it is used as a uint32_t? Sorry, I'm just being curious :)
There are two different values involved here.The first one is the reference count of the message. It's encoded as uint32_t as it is altered using CPU atomic operations and fixed-size type makes that simpler.
Then there's number of connections connected to the socket (as it appers in nn_dist class). Given that it's a number of in-memory objects, size_t seems appropriate.
Finally, the two values meet when a message is sent to *all* peers using a PUB or SURVEYOR socket. The algorithm takes the number of connections (size_t) and, instead of copying the message for every connection, it just increases the reference count of the message (uint32_t) by an appropriate amount.
HTH, Martin
