Hacker News Reader

Comment by rwmj | Hacker News Reader

I had a fun floating-point related bug the other day: I wanted to sort a list of doubles in C, and my comparison function was in essence:

    static int compare (void *dpa, void *dpb)
    {
      return *(double*)dpa - *(double*)dpb;
    }

The doubles represented times in nanoseconds, and the program failed to sort properly when the total runtime was larger than (IIRC) 2 seconds. I wonder if you can work out why.

(Spoiler alert) the fix was: https://github.com/libguestfs/libguestfs/commit/1f4a0bd90df3...

View on HN

Your fix requires a lot of mental CPU cycles on the part of whoever reads the code. Are you sure it wouldn't have been better to do it the simple, obvious way?

   double a = *(double *) dpa;   
   double b = *(double *) dpb;

   if (a > b) return 1;
   if (b < a) return -1;
   return 0;

Oh my gosh ... the fortran computed goto is back ... See http://www.lahey.com/docs/lfprohelp/F95ARComputed_GOTOStmt.h...

Yup I think you are right -- the (fixed) code is very obscure now.

I think it's actually quite elegant. The issue is that it always does two comparisons, even when one would suffice.

Depending on how the compiler computes a>b, it may be branchless, though, and I expect that to be more important on modern CPUs.

That code returns 0 for all NaN inputs.

True enough, but if you're passing NaNs to qsort(), something has already gone wrong in your professional life.

How come the argument types are void* instead of double*?

man qsort

http://linux.die.net/man/3/qsort

That's C's idea of generics.

Reply on news.ycombinator.com