I worked on a game port (PS1->PC) and all (game object) strings were hashed using 16-bit hash and code directly checked their 16-bit values against them in the code, but unlike "gperf" there was no perect hash, just some standard routine (forgot what it was). So it made me wonder (initially) when some objects were named differently - like instead of chara, it was chara2 or chara5 (no 1, 3 or 4) - e.g. find a name where when hashed won't result in collision... then again it was super fast and just using 16-bits so win overall... Later in another studio we had some DB with collisions (32-bits) such that you don't introduce some string that would cause collision with existing.

While gperf (unlike, say, CMPH[1]) is also intended for small sets of short strings, I expect the difference here is in the definition of “small”: in the article, it’s six, whereas in gperf’s intended application, it’s more like six dozen (and CMPH’s “large” is thousands to millions).

[1] https://cmph.sourceforge.net/gperf.html

But how many ns per query?