Random optimization because I noticed I had a copy of that in C lying around in FBInk ;p.

(We can't use the unsigned wraparound trickery from the first ref. for the surrogates test because Lua :/).

For info: https://luapower.com/luajit-notes

• multiplications and additions are cheaper than memory access, so storing the results of these operations in temporary variables might actually harm performance (more register spills).
• divisions are 4x slower than multiplications on x86, so when dividing by a constant, it helps turning x / c into x * (1 / c) since the constant expression is folded -- LuaJIT does this already for power-of-2 constants where the semantics are equivalent.
• the % operator is slow (it's implemented in terms of math.floor() and division) and really kills hot loops; math.fmod() is even slower; I don't have a solution for this except for x % powers-of-two which can be computed with bit ops.