Re: [Sbcl-commits] master: Correctly round floats converted from ratios

On Wed, Jan 28, 2026 at 12:10 AM Christophe Rhodes via Sbcl-devel
<[email protected]> wrote:
>
> stassats via Sbcl-commits <[email protected]> writes:
>
> > +(defun check-ratio-to-float (ratio type)
> > +  (declare (ratio ratio))
> > +  (let* ((result (float ratio type))
> > +         (new-ratio (rational result)))
> > +    (multiple-value-bind (sig exp sign) (integer-decode-float result)
> > +      (let* ((prev-float (scale-float (float (* sign (1- sig)) type) exp))
>
> I think this line is off-by-one if the float in question is a
> non-denormal and exactly at a power-of-two boundary, where the
> resolution is twice as fine below the float as above the float.
Corrected.

> Also, how should conversion to float behave above
> most-positive-fooble-float?  Is there an implicit interval above the
> maximum where we convert down and don't error, or does anything above
> the rational value of the most-positive float signal overflow?

The float-ratio functions call scale-float, which then decides whether
to overflow or not.


_______________________________________________
Sbcl-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sbcl-devel