Interesting. Another thing I noticed is that code runs faster if floating point and integer arithmetic instructions are interleaved rather than "blocked" together.