[a1369209993]

> What limitation of doubles are you overcoming when you use the range 1-2 and then subtract one instead of just generating in the range 0-1?

The limitation is that floating point numbers have different exponent depending on where in the range [0,1) they are, while every float in [1,2) has a exponent of 3FF. Generating a number in the latter range and subtracting 1.0 lets the floating point hardware do the work of figuring out the exponent.

On futher testing, though, it looks like you can just do ( rand53() * 1p-53 ), which will be hands down better as long as your floating-point hardware has a good int64->double conversion.