In D60835#1395293, @dteske wrote:

You see, R supports more than just source tarballs. It also supports binary tarballs and wheel-like tarballs (those are binary tarballs with the ELF executables patched with patchelf to reference libs that have been rolled into the tarball alongside those executables -- for library packages that do not depend on system libraries).

In D60835#1395251, @dteske wrote:

In D60835#1395063, @uzsolt wrote:

Only an idea: can we separate the build and install stage in "R CMD INSTALL"? I'd prefer it. Today I don't have time to test but maybe the "R CMD INSTALL --build" (with fine tuning) works (I did only one small test).
If exist a solution can introduce into cran.mk.
I don't know but maybe there are cran ports that have real build-time dependencies (required only building e.g. gmake in other regular ports).

What do you think?

(IMHO the cran.mk it isn't perfect e.g. doesn't respect TEST_ENV variable)

I have many years experience here and can offer a way to separate the build and install stage. Some notes:

@uzsolt --build is not a "build-only" flag. It means "build binaries of the installed package(s)"; the install still runs. R CMD INSTALL is a thin wrapper around tools:::.install_packages(). That function has no compile-only entry. It does however fork on the Built: field in DESCRIPTION. A source tree has no Built: field and so the first call to .install_packages() is a full install and writes Built:. An installed directory has Built: set, and a second call to .install_packages() on that directory takes the binary path, which is a copy (what would be our install stage).

The split is two calls. The first is the build, into a temporary library, with --no-test-load and the --install-tests that cran.mk always passes (--no-docs and --no-html too, when the DOCS option is off). It needs the toolchain and the packages a source install checks: Depends, Imports, and LinkingTo. The second call points R CMD INSTALL at that temporary installed directory and a different -l, the stage library (${STAGEDIR}${PREFIX}/lib/R/library). That one does not compile.

The operation is to install once into a library, then install that installed directory into another one. It would be a change to cran.mk, separate from the BUILD_DEPENDS cleanup.

Only an idea: can we separate the build and install stage in "R CMD INSTALL"? I'd prefer it. Today I don't have time to test but maybe the "R CMD INSTALL --build" (with fine tuning) works (I did only one small test).
If exist a solution can introduce into cran.mk.
I don't know but maybe there are cran ports that have real build-time dependencies (required only building e.g. gmake in other regular ports).

Are the changes of devel/biome part of commit?