Discard round-to-even code?

[ledvinap]

Round-to-even in %f is not working correctly, maybe it would be best to simply discard it completely? Opinions?

  unsigned long whole = (unsigned long)value;
  double tmp = (value - whole) * pow10[prec];
  unsigned long frac = (unsigned long)tmp;
  diff = tmp - frac;

The problem is that tmp value is multiplied by pow10.

With %.1f, 0.95:
0.95 is represented as 0.94999999999999996 in binary. This should be rounded down.
But (0.95 - 0) * 10 is 9.5, exactly representable in binary and tie-decision code takes place, rounding to even and outputting "1.0".

It may be possible to fix this double-rounding (at least in some cases), but IMO it's not worth the complexity. It may be simpler to just remove the tie-handling code and live with it

[ledvinap]

One edge case is %.0f - current rounding computation is correct and .5 is exactly representable in binary, so rounding should apply.