[FFmpeg-devel] [PATCH 2/3] avutil/opt: Add AV_OPT_TYPE_UINT64

[Ronald S. Bultje]

Hi,

On Sun, Nov 20, 2016 at 9:16 AM, Michael Niedermayer <michael at niedermayer.cc
> wrote:

> On Sun, Nov 20, 2016 at 08:56:15AM -0500, Ronald S. Bultje wrote:
> > Hi,
> >
> > On Sun, Nov 20, 2016 at 6:57 AM, Michael Niedermayer
> <michael at niedermayer.cc
> > > wrote:
> >
> > > @@ -131,6 +132,20 @@ static int write_number(void *obj, const AVOption
> *o,
> > > void *dst, double num, int
> > >          if (intnum == 1 && d == (double)INT64_MAX) *(int64_t *)dst =
> > > INT64_MAX;
> > >          else                                       *(int64_t *)dst =
> > > llrint(d) * intnum;
> > >          break;}
> > > +    case AV_OPT_TYPE_UINT64:{
> > > +        double d = num / den;
> > > +        // We must special case uint64_t here as llrint() does not
> > > support values
> > > +        // outside the int64_t range and there is no portable function
> > > which does
> > > +        // "INT64_MAX + 1ULL" is used as it is representable exactly
> as
> > > IEEE double
> > > +        // while INT64_MAX is not
> > > +        if (intnum == 1 && d == (double)UINT64_MAX) {
> > > +            *(int64_t *)dst = UINT64_MAX;
> > > +        } else if (o->max > INT64_MAX + 1ULL && d > INT64_MAX + 1ULL)
> {
> > > +            *(uint64_t *)dst = (llrint(d - (INT64_MAX + 1ULL)) +
> > > (INT64_MAX + 1ULL))*intnum;
> > > +        } else {
> > > +            *(int64_t *)dst = llrint(d) * intnum;
> > > +        }
> > > +        break;}
> > >      case AV_OPT_TYPE_FLOAT:
> > >          *(float *)dst = num * intnum / den;
> > >          break;
> >
> >
> > For the stupid, like me: what does this do? More specifically, this seems
> > an integer codepath, but there is a double in there. Why?
>
> write_number() has a num/den double argument. If this is used to
> set a uint64_t to UINT64_MAX things fail as double cannot exactly
> represent UINT64_MAX.  (the closest it can represent is "UINT64_MAX + 1"
> So it needs to be handled as a special case. Otherwise it turns into 0


This sounds incredibly shaky. Is write_number() so fringe that we don't
need an integer codepath?

Ronald