[R-pkg-devel] R package with external C++ library

[Duncan Murdoch]

On 03/08/2016 5:36 PM, Ege Rubak wrote:
> Hi,
>
> I would like to port Google's s2-library for spherical geometry (see
> e.g. https://github.com/micolous/s2-geometry-library for a fork on
> GitHub). It is not a standard library that can easily be installed on
> various systems, so I would like to include the source code in the R
> package. The catch is that I would like to modify the source code as
> little as possible :-)
>
> I have package everything and added configure scripts and a tiny
> R-function that calls one of the C++-functions (using the antiquated .C
> interface for now -- that will of course be changed) in this repo:
> https://github.com/spatstat/s2
>
> It compiles into a working package on Ubuntu (travis-ci + my laptop),
> OSX (travis-ci), and Windows (appveyor + my surface pro), but R CMD
> check produces some warnings (and a note about the size of the shared
> object, but I assume that is less important).
>
> The main things seem to be related to (travis log is at
> https://travis-ci.org/spatstat/s2/jobs/149578339):
>
> 1. Deprecated C++ headers <ext/hash_set> and <ext/hash_map>.
>
> 2. Compiled code that calls entry points which might terminate R or
> write to stdout/stderr.
>
> Is it hopeless to get on CRAN with warnings like these?

I don't set CRAN policy, but I would say yes.  Problem 1 limits your 
package to systems using compilers that support those antiquated 
headers; R tries very hard to be portable across many systems.  Problem 
2 makes R potentially unstable.

Duncan Murdoch

> I'm not very used to writing C/C++ code, but I guess 1. can be fixed by
> a few sed commands with the replacement headers and corresponding new
> function names. Point 2. can probably also be fixed with a reasonable
> effort, but I haven't investigated yet, and I would like an opinion from
> the list before spending more time on this. In more generality the
> question could be phrased something like:
>
> "When including C++ code from an upstream library which you do not
> control should R CMD check be completely spotless or is some flexibility
> to be expected in these circumstances?"
>
> Cheers,
> Ege
>
> PS: Extra question (prehaps particularly aimed at Dirk): When I will
> actually start to use the C++ library I expect it could be beneficial to
> use Rcpp. I have seen RcppModules mentioned somewhere, and I wonder if
> such an external C++ library would make sense to interface via
> RcppModules (again aiming at changing upstream sources as little as
> possible)?
>