for bitwise ops: look at 5.2, it has it now. 32-bit operations.
floating point arithmetic: in practice, I haven't seen it as a problem as soon as I stay within the 32-bit range - then the numbers can be represented exactly, so the abstraction works well.
OOP: can't say for it, I did not do it much.
For the "fast" - check luajit (http://luajit.org/) - it is near-native speed while remaining reasonably slow.
And the FFI in luajit is simply fantastic. Take a look at https://github.com/justincormack/ljsyscall which binds the syscalls.
The annoyance that I'd thought of is that the array indexing starts with one, not zero - this brings some headaches during the index calculations. But this also forces to rethink the logic to use more of iterators, and abstract away from the indices. So, after all, it is not that big of a deal.