[R-pkg-devel] Question about Debian installation error

[Ivan Krylov]

Hi Yu and welcome to R-package-devel! You're now part of the team as
well.

В Mon, 30 Sep 2024 21:30:44 -0400
Yu Yang <yuyang.stat using gmail.com> пишет:

>   I recently updated my package “glmtlp” to meet the “_PACKAGE”
> requirement and to deal with the “noRemap" requirement. I only
> modified the DESCRIPTION and glmtlp-package.R file and fixed a few
> typos. R CMD check showed no errors in my local machine (MacOS), but
> when I submitted to CRAN, it said installation error on Debian. 
> 
>   The log file is at:
> https://win-builder.r-project.org/incoming_pretest/glmtlp_2.0.2_20240930_002619/Debian/00install.out. 

Thank you for providing the link to the failing example. The
screenshot(?) didn't come through [1]; this mailing list only accepts
plain text messages and a restricted set of attachment types, so it's
best to compose in plain text as well.

The reason your package is failing is because your C++ code calls R API
functions by their short ("remapped") names, and this will no longer
work in the next version of R. This has been mentioned in the news for
R-devel [2,3]. The remapping is described in "Writing R Extensions",
chapter 6 [4] (and a bit more at the end, shortly before chapter 7).

In short, to reproduce the error outside of R-devel, you need to
#define R_NO_REMAP before #including R API (or provide the -DR_NO_REMAP
command-line switch to the compiler), and then fix it by finding every
remapped name in your code and replacing it by the full name (typically
by prepending Rf_ to it).

You're not the first package author to have problems with C++ code due
to R_NO_REMAP [5]. There's now an incomplete, unofficial set of
manpage-style API documentation (made possible by Luke Tierney's effort
to export a list of API functions programmatically via tools:::funAPI)
where you can find both full and remapped names for most of "ordinary"
R API and slightly less than a half of experimental R API [6]. Thanks
to Erez Shomron, there's also information about when a lot of the
functions had been introduced.

The reason we have remapping in the first place is that for historical
reasons, R API contained a lot of functions with "ordinary" names, such
as "length", which conflicted with other C functions named "length" in
the same namespace. This problem has been solved at the cost of
introducing a more manageable problem by remapping: the "real" names of
the conflicting functions now start with Rf_ (e.g. "Rf_length"), and to
avoid breaking all existing code (including R itself), R uses the C
preprocessor to replace the short names with the full names on the fly:
"#define length Rf_length" and so on.

Unfortunately, that still leads to conflicts when you need to use
something named "length" (e.g. [7]) in a source file that #includes the
R API. You type s.length(), and the preprocessor patches that to
s.Rf_length(), and everything breaks. Since C++ code is the most
affected by this, the R developers are now disabling it for all C++
code.

Hope this helps!

-- 
Best regards,
Ivan

[1]
https://stat.ethz.ch/pipermail/r-package-devel/2024q4/011121.html

[2]
https://developer.r-project.org/blosxom.cgi/R-devel/NEWS/2024/04/14#n2024-04-14

[3]
https://stat.ethz.ch/R-manual/R-devel/doc/html/NEWS.html

[4]
https://cran.r-project.org/doc/manuals/R-exts.html#The-R-API

[5]
https://stat.ethz.ch/pipermail/r-package-devel/2024q2/010913.html

[6]
https://aitap.codeberg.page/R-api/#nrows
https://aitap.codeberg.page/R-api/#XLENGTH
https://aitap.codeberg.page/R-api/#allocVector

[7]
https://en.cppreference.com/w/cpp/string/basic_string/size