Re: support for bitwise comparison of floats

[Paul Eggert]

The "right" way to test for -0 versus +0 is to use copysign.
For example, instead of this:

        if (arg < 0.0L)
          {
            sign = -1;
            arg = -arg;
          }
        else if (arg == 0.0L)
          {
            /* Distinguish 0.0L and -0.0L.  */
            static long double plus_zero = 0.0L;
            long double arg_mem = arg;
            if (memcmp (&plus_zero, &arg_mem, SIZEOF_LDBL) != 0)
              {
                sign = -1;
                arg = -arg;
              }
          }


vasnprintf.c should do this:

        if (copysignl (1, arg) < 0)
          {
            sign = -1;
            arg = -arg;
          }

This is a valid transformation since we know at this point that ARG is
not a NaN.

I think it's much cleaner.  And the code is smaller on my platform,
anyway: no function call is generated for copysignl.  It should be
fast enough for vasnprintf.

This needs copysignl for older platforms, but another gnulib module
should suffice for that....