I was optimizing the Lua transpiler for my modding language grug and ran into a really weird LuaJIT performance bug. The same benchmark could randomly run 20× slower, and it turned out a LuaJIT NYI (Not Yet Implemented) could silently blacklist an unrelated hot loop.

I wrote up the investigation here.

It goes from the benchmark mystery through LuaJIT's trace recorder internals, and ends with a PR to get unpack off LuaJIT's NYI list. Feedback is very welcome! :)