Better to do this in assembler - then you know for sure what opcodes you're getting and, therefore, how long it'll take.
No need for tuning, and no nasty surprises when the compiler suddenly decides to generate different code.

Search the Keil forum - there's been a couple of threads on this recently.

That's 1ms - lowercase 's' for seconds (uppercase 'S' is siemens - the SI unit of conductance).