On Thu, May 03, 2018 at 11:44:09AM +0200, Richard Biener wrote:
> > + *RESULT = nextafter (*ARG0, *ARG1)
> > +
> > + or
> > +
> > + *RESULT = nexttoward (*ARG0, *ARG1)
> > +
> > + in format FORMAT. Return true on success. */
> > +
> > +static bool
> > +fold_const_nextafter (real_value *result, const real_value *arg0,
> > + const real_value *arg1, const real_format *format)
> > +{
> > + if (flag_signaling_nans
>
> HONOR_SNANS ()?
That requires a machine_mode or tree, but I don't have either of those,
nor the caller (fold_const_call_sss) has those.
I could change it to:
if (flag_signalling_nans
&& !flag_finite_math_only
&& format->has_nans
so that it would better duplicate what HONOR_SNANS actually tests.
Though, I think it is ok to punt if one of the operands is a signalling nan
even if flag_signalling_nans.
> > + && (REAL_VALUE_ISSIGNALING_NAN (*arg0)
> > + || REAL_VALUE_ISSIGNALING_NAN (*arg1)))
> > + return false;
> > +
> > + /* Don't handle composite modes, nor decimal, nor modes without
> > + inf or denorm at least for now. */
> > + if (format->pnan < format->p
> > + || format->b == 10
> > + || !format->has_inf
> > + || !format->has_denorm)
> > + return false;
>
> I wonder if we should assert in real_nextafter that we can actually
> handle the input? It likely returns garbage if fed with decimal float
> stuff?
Yes, but nobody would call it even if we removed the || format->b == 10
condition, the only builtins we fold this for have non-decimal
float/double/long double arguments.
I can surely add an assert with the same set of checks as the above ones.
Jakub
