On Tue, Mar 25, 2025 at 08:47:49AM +0100, Richard Biener wrote:
> On Mon, 24 Mar 2025, Jakub Jelinek wrote:
>
> > On Mon, Mar 24, 2025 at 10:55:44PM +0100, Jakub Jelinek wrote:
> > > If it was HOST_WIDE_INT_MAX + (size_t) 1 to ~(size_t) 0, previously it
> > > would
> > > be false and now is false.
> >
> > Sorry, this case used to be false and now is true.
>
> Just to add, when writing this I wondered whether a
>
> bool real_is_integer (const REAL_VALUE_TYPE *, wide_int *, int);
>
> would be useful, or adding an optional bool *exact arg to the
> existing real_to_integer.
Maybe the latter would be better.
> Semantics of -0.0 vs. 0.0 and INF/NAN (which previously, when
> casting to (size_t) was UB, likewise for out-of-bound values?)
> is of course details that need to be documented.
Yes, it was UB at compile time.
Note, my fmodf128 replacements with mpfr most likely don't handle
infinities and NaNs right either, but I think the code was again
UB at those points because it used casts to integers from _Float128
after that.
> I think the main issue with my transform is that it lost the
> non-negative check.
And also the cases of unsigned which fits into host size_t but
doesn't fit into HOST_WIDE_INT. But what exactly the code wants to
do is unclear.
> The problem with the orignal code is that
> it lacks documentation on the intent of the check.
Yes.
Jakub
