[Rd] [External] Possible ALTREP bug

[Gabriel Becker]

Hi Jim et al,

Just to hopefully add a bit to what Luke already answered, from what I am
recalling looking back at that bioconductor thread Elt methods are used in
places where there are hard implicit assumptions that no garbage collection
will occur (ie they are called on things that aren't PROTECTed), and beyond
that, in places where there are hard assumptions that no error (longjmp)
will occur. I could be wrong, but I don't know that suspending garbage
collection would protect from the second one. Ie it is possible that an
error *ever* being raised from R code that implements an elt method could
cause all hell to break loose.

Luke or Tomas Kalibera would know more.

I was disappointed that implementing ALTREPs in R code was not in the cards
(it was in my original proposal back in 2016 to the DSC) but I trust Luke
that there are important reasons we can't safely allow that.

Best,
~G

On Fri, May 28, 2021 at 8:31 AM Jim Hester <james.f.hester using gmail.com> wrote:

> From reading the discussion on the Bioconductor issue tracker it seems like
> the reason the GC is not suspended for the non-string ALTREP Elt methods is
> primarily due to performance concerns.
>
> If this is the case perhaps an additional flag could be added to the
> `R_set_altrep_*()` functions so ALTREP authors could indicate if GC should
> be halted when that particular method is called for that particular ALTREP
> class.
>
> This would avoid the performance hit (other than a boolean check) for the
> standard case when no allocations are expected, but allow authors to
> indicate that R should pause GC if needed for methods in their class.
>
> On Fri, May 28, 2021 at 9:42 AM <luke-tierney using uiowa.edu> wrote:
>
> > integer and real Elt methods are not expected to allocate. You would
> > have to suspend GC to be able to do that. This currently can't be done
> > from package code.
> >
> > Best,
> >
> > luke
> >
> > On Fri, 28 May 2021, Gábor Csárdi wrote:
> >
> > > I have found some weird SEXP corruption behavior with ALTREP, which
> > > could be a bug. (Or I could be doing something wrong.)
> > >
> > > I have an integer ALTREP vector that calls back to R from the Elt
> > > method. When this vector is indexed in a lapply(), its first element
> > > gets corrupted. Sometimes it's just a type change to logical, but
> > > sometimes the corruption causes a crash.
> > >
> > > I saw this on macOS from R 3.5.3 to 4.2.0. I created a small package
> > > that demonstrates this: https://github.com/gaborcsardi/redfish
> > >
> > > The R callback in this package calls `loadNamespace("Matrix")`, but
> > > the same crash happens for other packages as well, and sometimes it
> > > also happens if I don't load any packages at all. (But that example
> > > was much more complicated, so I went with the package loading.)
> > >
> > > It is somewhat random, and sometimes turning off the JIT avoids the
> > > crash, but not always.
> > >
> > > Hopefully I am just doing something wrong in the ALTREP code (see
> > > https://github.com/gaborcsardi/redfish/blob/main/src/test.c), and it
> > > is not actually a bug.
> > >
> > > Thanks,
> > > Gabor
> > >
> > > ______________________________________________
> > > R-devel using r-project.org mailing list
> > > https://stat.ethz.ch/mailman/listinfo/r-devel
> > >
> >
> > --
> > Luke Tierney
> > Ralph E. Wareham Professor of Mathematical Sciences
> > University of Iowa                  Phone:             319-335-3386
> > Department of Statistics and        Fax:               319-335-3017
> >     Actuarial Science
> > 241 Schaeffer Hall                  email:   luke-tierney using uiowa.edu
> > Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu
> > ______________________________________________
> > R-devel using r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-devel
> >
>
>         [[alternative HTML version deleted]]
>
> ______________________________________________
> R-devel using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>

	[[alternative HTML version deleted]]