Generating instructions for floating point arithmetic