Skip to content

[Rcpp-devel] Rcpp-FAQ vignette and OS X entries: Small update

4 messages · Dirk Eddelbuettel, Kevin Ushey

#
I noticed a post cross-referenced on R Bloggers about Rcpp(Armadillo), OS X
and the need for an additional Fortran packages from Simon's site -- so I
took this as opportunity to update the OS X answers in the Rcpp FAQ vignette.

As I don't use OS X, I'd appreciate a second set of eyes by anyone with a
moment to spare and some mad OS X "skillz".

Current Rnw file
  https://github.com/RcppCore/Rcpp/blob/master/vignettes/Rcpp-FAQ.Rnw

Link to the committed changes
  https://github.com/RcppCore/Rcpp/commit/e892622ce1c741f47b2e43b1d44ece32d44ee239

Corrections and improvements to other parts are of course welcome as well :)

Dirk
#
I think this bit:

The \texttt{clang} and \texttt{clang++} compilers from the LLVM project can
also be used as they are inter-operable with \texttt{gcc} et al.  The
\texttt{clang++} compiler is interesting as it emits much more
comprehensible error messages than \texttt{g++} (though \texttt{g++} 4.8
and 4.9
have caught up).


needs to be amended. The standard library implementation used with gcc
(libstdc++) is not ABI compatible with the one used with clang (libc++).
This has been a common pain point for users running Snow Leopard who
updated their operating system to Mavericks, and also installed Mavericks
R, without re-installing any previously installed packages -- having a
clang-compiled package with C++ code try to link to an older gcc-compiled
package with C++ code caused many headaches for users, either with weird
segfaults or errors on loading packages or (if lucky) errors at link time.

In addition:

OS X used to be a little more conservative with compiler versions as Apple
stuck with gcc-4.2. Following the 'Mavericks' release, it is the opposite as
only llvm is provide in Xcode---while the R build provided by CRAN still has
hardwired settings for the gcc/g++ combination.

Until that is resolved, users either need to create softlinks (say, in,
\code{/usr/local/bin}) or override the \code{CC} and \code{CXX} variables
via a
file \code{~/.R/Makevars} or its system-equivalent in R's \code{etc/}
directory. Also see \faq{q:OSXArma} below. In general, consult the
\code{r-sig-mac} mailing list for further details.


There are now two separate CRAN binaries, one compiled for Snow Leopard and
above (using gcc-4.2.1), and one compiled for Mavericks (using the latest
clang release distributed with Mavericks). Fortunately, since R doesn't
include any C++ code, one can use Snow Leopard R with clang-compiled
packages containing C++ code with no problems. However, packages containing
C++ code should be compiled with the same compiler suite. This is a bit of
a pain point for users of Snow Leopard R who might be using clang -- the
binaries distributed by CRAN are compiled with gcc-4.2 and libstdc++ (by
default) and those won't be compatible with packages compiled from source
with clang++ and libc++.

Professor Ripley compiled a set of recommendations in a post on R-SIG-Mac
that we could link to:
https://stat.ethz.ch/pipermail/r-sig-mac/2014-April/010835.html

We might also link to the BioConductor guidelines for Mavericks + C++11
development: http://www.bioconductor.org/developers/how-to/mavericks-howto/;
it's a nice overview of the state of the world there.

The general prescription, I think:

1. If you want to mix and match CRAN binaries with packages installed from
source, make sure your compiler suite matches that of the CRAN R you've
installed. Otherwise, the safer route is to compile all packages from
source.

2. If you're running OS X Mavericks (or greater), and want to compile
packages from source (ostensibly using clang), use the Mavericks CRAN R
binary, otherwise you will run into frustrating ABI compatibility issues
(as Snow Leopard R will grab CRAN package binaries compiled with gcc 4.2;
again incompatible with those compiled with clang)*.

I can try to amend the vignette to this effect later and submit a PR. If
anyone else on the list thinks I missed something here, let me know too.

Cheers,
Kevin

*I suppose someone running Mavericks with a Snow Leopard R could set the
'pkgType' option (see ?install.packages) to grab Mavericks binaries of
packages, but I doubt this is recommended.

PS: A useful utility one could write would check what C++ standard library
implementation each installed package links to. Maybe I'll write that...
On Sun, Aug 3, 2014 at 9:17 AM, Dirk Eddelbuettel <edd at debian.org> wrote:

            
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rcpp-devel/attachments/20140803/8ed1c06a/attachment.html>
#
As promised:

otool <- Sys.which("otool")
if (otool == "") {
  stop("This utility requires 'otool' to run")
}

allPackages <- list.files(.libPaths(), full.names = TRUE)
stdLibsUsed <- lapply(allPackages, function(path) {
  pkgName <- basename(path)
  libPath <- file.path(path, "libs", paste0(pkgName, ".so"))
  if (!file.exists(libPath)) {
    return(character())
  }
  otoolOutput <- system2("otool", args = c("-L", libPath), stdout = TRUE)
  c(
    "libc++"[any(grepl("libc++", otoolOutput, fixed = TRUE))],
    "libstdc++"[any(grepl("libstdc++", otoolOutput, fixed = TRUE))]
  )
})

pkgsWithLibStdCpp <- allPackages[sapply(stdLibsUsed, function(x) {
  "libstdc++" %in% x
})]

pkgsWithLibCpp <- allPackages[sapply(stdLibsUsed, function(x) {
  "libc++" %in% x
})]

list(
  "libstdc++" = pkgsWithLibStdCpp,
  "libc++" = pkgsWithLibCpp
)


I'm not sure where this could live, but basically, Mavericks users who see
any packages linking to `libstdc++` would probably want to reinstall those
packages. There may be some packages that need to explicitly link to
libstdc++ even on a Mavericks system for other reasons, though...

That said, most of the initial migration headaches seem to be over so this
probably would have been more useful half a year ago ;)

Cheers,
Kevin
On Sun, Aug 3, 2014 at 1:10 PM, Kevin Ushey <kevinushey at gmail.com> wrote:

            
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rcpp-devel/attachments/20140803/3989ab4e/attachment-0001.html>
#
On 3 August 2014 at 13:10, Kevin Ushey wrote:
| I think this bit:
| 
|     The \texttt{clang} and \texttt{clang++} compilers from the LLVM project can
|     also be used as they are inter-operable with \texttt{gcc} et al. ?The
|     \texttt{clang++} compiler is interesting as it emits much more
|     comprehensible error messages than \texttt{g++} (though \texttt{g++} 4.8
|     and 4.9
|     have caught up).
| 
| needs to be amended. The standard library implementation used with gcc
| (libstdc++) is not ABI compatible with the one used with clang (libc++). This

My bad. On Linux it is. Will qualify. 
 
| Professor Ripley compiled a set of recommendations in a post on R-SIG-Mac that
| we could link to:?https://stat.ethz.ch/pipermail/r-sig-mac/2014-April/
| 010835.html

Yes, thanks for the rmeinder. I had meant to link to this post as well. Will do.
 
| We might also link to the BioConductor guidelines for Mavericks + C++11
| development:?http://www.bioconductor.org/developers/how-to/mavericks-howto/;
| it's a nice overview of the state of the world there.

Good one. Had seen that as well as promptly forgotten. 

| I can try to amend the vignette to this effect later and submit a PR. If anyone
| else on the list thinks I missed something here, let me know too.

I'd welcome that. I'll make a quick set of changes now, but there will always
be room for improvement.

Thanks!

Dirk