https://issues.apache.org/jira/browse/ARROW-3790
On Wed, Nov 14, 2018 at 9:03 AM Wes McKinney <wesmck...@gmail.com> wrote:
>
> There's a logical error here
>
> https://github.com/apache/arrow/blob/master/cpp/src/arrow/compute/kernels/cast.cc#L239
>
> The cast
>
> constexpr in_type kMax =
> static_cast<in_type>(std::numeric_limits<out_type>::max());
>
> is overflowing int32_t
>
> This code path was only written for the unsigned-to-signed case. The
> functors need to be reworked to accommodate the signed-to-unsigned
> cast
> On Wed, Nov 14, 2018 at 8:36 AM Francois Saint-Jacques
> <fsaintjacq...@networkdump.com> wrote:
> >
> > Seems like the type combination you're using (int32 -> uint32) and (int32
> > -> uint64) don't match the following pattern-matching
> >
> > https://github.com/apache/arrow/blob/master/cpp/src/arrow/compute/kernels/cast.cc#L191-L192
> >
> > which avoid using "safe" cast and revert to the following cast
> > implementation (with no explicit check):
> >
> > https://github.com/apache/arrow/blob/master/cpp/src/arrow/compute/kernels/cast.cc#L316-L318
> >
> > My take on this is that downcasting concept should only apply to same sign
> > integers. I'd propose the following refactor:
> >
> > 1. Rename `allow_int_overflow` to `allow_int_truncate`. Since truncation is
> > still applicable even if there's no downcasting.
> > 2. Refactor code such that the pattern matching is done on possible
> > truncation and not downcasting. The fast casting is only applied to
> > same-sign and size(O) >= size(I). The (possible according to
> > allow_int_truncate) slow and bound checking is done when signed-ness is not
> > matching or there's downcasting.
> >
> > Can you open a JIRA ticket?
> >
> > François
> >
> > On Wed, Nov 14, 2018 at 6:10 AM Romain Francois <rom...@purrple.cat> wrote:
> >
> > > I'm implementing Array$cast() in the R package, i.e.
> > > https://issues.apache.org/jira/browse/ARROW-3741?filter=12344983
> > >
> > > I'm seeing some weird results when casting from int32 to uint32 and uint64
> > > (I'm expecting errors here as it's supposed to be safe
> > >
> > > https://github.com/apache/arrow/pull/2959#issuecomment-438606974
> > >
> > > a <- array(-(1:3))
> > > a$cast(uint16())
> > > #> Error in Array__cast(self, target_type, options): Invalid: Integer
> > > value out of bounds
> > > a$cast(uint16())
> > > #> Error in Array__cast(self, target_type, options): Invalid: Integer
> > > value out of bounds
> > > a$cast(uint32())
> > > #> arrow::Array
> > > #> [
> > > #> 4294967295,
> > > #> 4294967294,
> > > #> 4294967293
> > > #> ]
> > > a$cast(uint32())$type()
> > > #> arrow::UInt32
> > > #> uint32
> > >
> > > a$cast(uint64())
> > > #> arrow::Array
> > > #> [
> > > #> -1,
> > > #> -2,
> > > #> -3
> > > #> ]
> > > a$cast(uint64())$type()
> > > #> arrow::UInt64
> > > #> uint64
> > >
> > > I don't think this is an R issue.
> > >
> > > Can someone using other front ends to the C++ library check if they get
> > > the same problem ?
> > >
> > >
> > >
> > >
> >
> > --
> > Sent from my jetpack.
