Maybe this is why Intel give us F2XM1 (2^x - 1) as an opcode instead of, what would seem to be more useful, 2^x? Still - it requires a lot of thought from the programmer to realise he might gain some precision by removing pows and exps and splitting them up to avoid precision loss - and even more to make the jump to assembler in order to do it.
Another reason is that the operation may already be needed as a component of a trigonometric function, and it doesn't cost much to promote it to a separate opcode.
