I am trying to convert a double to a string in a native NT application, i.e. an application that only depends on ntdll.dll. Unfortunately, ntdll’s version of vsnprintf does not support %f et al., forcing me to implement the conversion on my own.
The aforementioned ntdll.dll exports only a few of the math.h functions (floor, ceil, log, pow, …). However, I am reasonably sure that I can implement any of the unavailable math.h functions if necessary.
There is an implementation of floating point conversion in GNU’s libc, but the code is extremely dense and difficult to comprehent (the GNU indentation style does not help here).
I’ve already implemented the conversion by normalizing the number (i.e. multiplying/dividing the number by 10 until it’s in the interval [1, 10)) and then generating each digit by cutting the integral part off with modf and multiplying the fractional part by 10. This works, but there is a loss of precision (only the first 15 digits are correct). The loss of precision is, of course, inherent to the algorithm.
I’d settle with 17 digits, but an algorithm that would be able to generate an arbitrary number of digits correctly would be preferred.
Could you please suggest an algorithm or point me to a good resource?
After a lot of research, I found a paper titled Printing Floating-Point Numbers Quickly and Accurately. It uses exact rational arithmetic to avoid precision loss. It cites a little older paper: How to Print Floating-Point Numbers Accurately, which however seems to require ACM subscription to access.
Since the former paper was reprinted in 2006, I am inclined to believe that it is still current. The exact rational arithmetic (which requires dynamic allocation) seems to be a necessary evil.