Thanks - I hadn't thought about that but it seems to be the main downside of this approach. The benefit of NaN-boxing is that it reassigns values that are otherwise unused - floating-point calculations will never generate NaNs with those bit patterns.
