On 8/29/19 4:45 AM, m.djorov@... [firebird-support] wrote:

>  
> In binary representation of the values maybe 4.72 is the closest, but
> we are talking about mathematic.
> This is from the description of the ROUND function in the documentation:
> Rounds a number to the nearest integer. If the fractional part is
> exactly 0.5, rounding is upward for positive numbers and downward for
> negative numbers.
> So .725 is all cases is rounded up to .73 as it should be, but in one
> case it is not.
>

The issue is that when you write: cast(2.725 as double precision) you
aren't dealing with 'mathematics' anymore, or even the value 2.725

The value that you get at that point will be the nearest value to 2.725
that is expressible as a double precision floating point number which
will be slightly different since 2.725 is NOT exactly representable as
double precision floating point number. The number you get is allowed to
be either the representable value just below or just above, the value,
though the preference is the closer one. If the number you get is
something like 2.72499...xx then round needs to round down, if you get
something like 2.72500..xx then round needs to round up.

If the numbers you had WERE exactly representable, like 2.625 (21/8)
then the rules on how to round would matter, but since the number you
have, after being made representable, isn't exactly 0.5 in the
fractional part, that clause doesn't apply.


--
Richard Damon