Is it possible to make install.packages compile source code on Unix but use shipped binary on Windows?

Hi,

is it possible to make install.packages install a source package by
compiling it on Unix, but skip compilation and use a pre-built windows
binary (shipped in exec/) if installing on Windows?

The package should be installable on Windows machines that do not have
Rtools installed. This is to make the update process easier for users,
because the C++ code does not change often, meaning that I can just
send source packages built on Linux, with a binary executable built
once in a while on win-builder.

Note that this is for a package that is not meant for CRAN, so no
policy issue here.

Thank you.

Bests,
Renaud

A 'pre-built windows binary' of what?

You can easily ship a DLL or .exe for use on Windows: just make use of
configure.win and/or Makefile.win.  If you have a Makefile.win it will
override the normal procedures for directory src.

OTOH, shipping something in exec will be installed everwhere.

Watch out for sub-architectures: 'Windows' is in fact two platforms.  We
have over the years seen a lot of problem with people shipping 64-bit
Windows binaries: 32-bit Windows does not recognize those.

Thank you for the replies and hints.

A 'pre-built windows binary' of what?

Sorry this was not clear, src/ contains source code for a binary
executable, which produces no .lib or .dll.
It is all fine in a R development environment:
   * On Linux, the executable is compiled and copied to
R_PACKAGE_DIR/binR_ARCH using a custom install.libs.R script.
   * Same thing on Windows, if Rtools is installed, install.packages
compiles and installs both 32 and 64 bits .exe.

You can easily ship a DLL or .exe for use on Windows: just make use of
configure.win and/or Makefile.win.  If you have a Makefile.win it will
override the normal procedures for directory src.

I have tried using Makevars.win, but this requires make to be
installed (as by Rtools). So I guess using Makefile.win would also not
work, on a classic non-development geared Windows host.
Would using configure.win work without Rtools?

OTOH, shipping something in exec will be installed everywhere.

True, although it seems that sub-directories of exec/ are actually not
installed though. Is this intentional?

Watch out for sub-architectures: 'Windows' is in fact two platforms.  We
have over the years seen a lot of problem with people shipping 64-bit
Windows binaries: 32-bit Windows does not recognize those.

Indeed, and I think I properly take care of this in install.libs.R.

@Grabriel: the package is provided as a source package. This one is
installable locally from file, although in may case users actually get
it through a personal CRAN-like repository, using argument
type='both'.

Thank you.

Bests,
Renaud

Maybe.  Read the documentation and sources for yourself (see below).

Yes, *and documented*

(including that it should not be used for Windows executables).

Maybe.  Read the documentation and sources for yourself (see below).

Not working, at least in my hands, as it requires `sh`.

Yes, *and documented*

True. I overlooked the beginning of the NB point.

(including that it should not be used for Windows executables).

Yes, that's why I use the suggested procedure that uses
src/install.libs.R to copy compiled .exe files into bin<R_ARCH>/.

So, eventually, I guess the answer to the original question is: no,
one cannot make install.packages skip compilation of a source package,
only if on Windows, without having Rtools installed -- and in PATH.

Renaud

It isn't hard to set up Rtools on Windows and you only need to do it
once. Then you build a binary package on your development system to
distribute to your users.

Without even considering any technical details there is a purely
strategic issue here. If a system has been set up that is robust and
widely tested, like the R packaging system, you are much better off
working with it than trying to subvert it.