Writing R Extensions

This is a guide to extending R, describing the process of creating R add-on packages, writing R documentation, R’s system and foreign language interfaces, and the R API.

This manual is for R, version 4.5.0 Under development (2024-12-01).

Copyright © 1999–2024 R Core Team

Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies.

Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one.

Permission is granted to copy and distribute translations of this manual into another language, under the above conditions for modified versions, except that this permission notice may be stated in a translation approved by the R Core Team.

Table of Contents


Acknowledgements

The contributions to early versions of this manual by Saikat DebRoy (who wrote the first draft of a guide to using .Call and .External) and Adrian Trapletti (who provided information on the C++ interface) are gratefully acknowledged.


1 Creating R packages

Packages provide a mechanism for loading optional code, data and documentation as needed. The R distribution itself includes about 30 packages.

In the following, we assume that you know the library() command, including its lib.loc argument, and we also assume basic knowledge of the R CMD INSTALL utility. Otherwise, please look at R’s help pages on

?library
?INSTALL

before reading on.

For packages which contain code to be compiled, a computing environment including a number of tools is assumed; the “R Installation and Administration” manual describes what is needed for each OS.

Once a source package is created, it must be installed by the command R CMD INSTALL. See Add-on packages in R Installation and Administration.

Other types of extensions are supported (but rare): See Package types.

Some notes on terminology complete this introduction. These will help with the reading of this manual, and also in describing concepts accurately when asking for help.

A package is a directory of files which extend R, a source package (the master files of a package), or a tarball containing the files of a source package, or an installed package, the result of running R CMD INSTALL on a source package. On some platforms (notably macOS and ‘x86_64’ Windows) there are also binary packages, a zip file or tarball containing the files of an installed package which can be unpacked rather than installing from sources.

A package is not1 a library. The latter is used in two senses in R documentation.

There are a number of well-defined operations on source packages.

The concept of lazy loading of code or data is mentioned at several points. This is part of the installation, always selected for R code but optional for data. When used the R objects of the package are created at installation time and stored in a database in the R directory of the installed package, being loaded into the session at first use. This makes the R session start up faster and use less (virtual) memory. (For technical details, see Lazy loading in R Internals.)

CRAN is a network of WWW sites holding the R distributions and contributed code, especially R packages. Users of R are encouraged to join in the collaborative project and to submit their own packages to CRAN: current instructions are linked from https://CRAN.R-project.org/banner.shtml#submitting.


1.1 Package structure

The sources of an R package consist of a subdirectory containing the files DESCRIPTION and NAMESPACE, and the subdirectories R, data, demo, exec, inst, man, po, src, tests, tools and vignettes (some of which can be missing, but which should not be empty). The package subdirectory may also contain files INDEX, configure, cleanup, LICENSE, LICENCE and NEWS. Other files such as INSTALL (for non-standard installation instructions), README/README.md2, or ChangeLog will be ignored by R, but may be useful to end users. The utility R CMD build may add files in a build directory (but this should not be used for other purposes).

Except where specifically mentioned,3 packages should not contain Unix-style ‘hidden’ files/directories (that is, those whose name starts with a dot).

The DESCRIPTION and INDEX files are described in the subsections below. The NAMESPACE file is described in the section on Package namespaces.

The optional files configure and cleanup are (Bourne) shell scripts which are, respectively, executed before and (if option --clean was given) after installation on Unix-alikes, see Configure and cleanup. The analogues on Windows are configure.win and cleanup.win. Since R 4.2.0 on Windows, configure.ucrt and cleanup.ucrt are supported and take precedence over configure.win and cleanup.win. They can hence be used to provide content specific to UCRT or Rtools42 and newer, if needed, but the support for .ucrt files may be removed in future when building packages from source on the older versions of R will no longer be needed, and hence the files may be renamed back to .win.

For the conventions for files NEWS and ChangeLog in the GNU project see https://www.gnu.org/prep/standards/standards.html#Documentation.

The package subdirectory should be given the same name as the package. Because some file systems (e.g., those on Windows and by default on macOS) are not case-sensitive, to maintain portability it is strongly recommended that case distinctions not be used to distinguish different packages. For example, if you have a package named foo, do not also create a package named Foo.

To ensure that file names are valid across file systems and supported operating systems, the ASCII control characters as well as the characters ‘"’, ‘*’, ‘:’, ‘/’, ‘<’, ‘>’, ‘?’, ‘\’, and ‘|’ are not allowed in file names. In addition, files with names ‘con’, ‘prn’, ‘aux’, ‘clock$’, ‘nul’, ‘com1’ to ‘com9’, and ‘lpt1’ to ‘lpt9’ after conversion to lower case and stripping possible “extensions” (e.g., ‘lpt5.foo.bar’), are disallowed. Also, file names in the same directory must not differ only by case (see the previous paragraph). In addition, the basenames of ‘.Rd’ files may be used in URLs and so must be ASCII and not contain %. For maximal portability filenames should only contain only ASCII characters not excluded already (that is A-Za-z0-9._!#$%&+,;=@^(){}'[] — we exclude space as many utilities do not accept spaces in file paths): non-English alphabetic characters cannot be guaranteed to be supported in all locales. It would be good practice to avoid the shell metacharacters (){}'[]$~: ~ is also used as part of ‘8.3’ filenames on Windows. In addition, some applications on Windows can only work with path names of certain length, following an earlier limit in the Windows operating system. Packages are normally distributed as tarballs, and these have a limit on path lengths. So, to be friendly to users who themselves may want to use a relatively long path where they extract the package, and for maximal portability, 100 bytes.

A source package if possible should not contain binary executable files: they are not portable, and a security risk if they are of the appropriate architecture. R CMD check will warn about them4 unless they are listed (one filepath per line) in a file BinaryFiles at the top level of the package. Note that CRAN will not accept submissions containing binary files even if they are listed.

The R function package.skeleton can help to create the structure for a new package: see its help page for details.


1.1.1 The DESCRIPTION file

The DESCRIPTION file contains basic information about the package in the following format:

Package: pkgname
Version: 0.5-1
Date: 2015-01-01
Title: My First Collection of Functions
Authors@R: c(person("Joe", "Developer", role = c("aut", "cre"),
                     email = "Joe.Developer@some.domain.net",
                     comment = c(ORCID = "nnnn-nnnn-nnnn-nnnn")),
              person("Pat", "Developer", role = "aut"),
              person("A.", "User", role = "ctb",
                     email = "A.User@whereever.net"))
Author: Joe Developer [aut, cre],
  Pat Developer [aut],
  A. User [ctb]
Maintainer: Joe Developer <Joe.Developer@some.domain.net>
Depends: R (>= 3.1.0), nlme
Suggests: MASS
Description: A (one paragraph) description of what
  the package does and why it may be useful.
License: GPL (>= 2)
URL: https://www.r-project.org, http://www.another.url
BugReports: https://pkgname.bugtracker.url

The format is that of a version of a ‘Debian Control File’ (see the help for ‘read.dcf’ and https://www.debian.org/doc/debian-policy/ch-controlfields.html: R does not require encoding in UTF-8 and does not support comments starting with ‘#’). Fields start with an ASCII name immediately followed by a colon: the value starts after the colon and a space. Continuation lines (for example, for descriptions longer than one line) start with a space or tab. Field names are case-sensitive: all those used by R are capitalized.

For maximal portability, the DESCRIPTION file should be written entirely in ASCII — if this is not possible it must contain an ‘Encoding’ field (see below).

Several optional fields take logical values: these can be specified as ‘yes’, ‘true’, ‘no’ or ‘false’: capitalized values are also accepted.

The ‘Package’, ‘Version’, ‘License’, ‘Description’, ‘Title’, ‘Author’, and ‘Maintainer’ fields are mandatory, all other fields are optional. Fields ‘Author’ and ‘Maintainer’ can be auto-generated from ‘Authors@R’, and may be omitted if the latter is provided: however if they are not ASCII we recommend that they are provided.

The mandatory ‘Package’ field gives the name of the package. This should contain only (ASCII) letters, numbers and dot, have at least two characters and start with a letter and not end in a dot. If it needs explaining, this should be done in the ‘Description’ field (and not the ‘Title’ field).

The mandatory ‘Version’ field gives the version of the package. This is a sequence of at least two (and usually three) non-negative integers separated by single ‘.’ or ‘-’ characters. The canonical form is as shown in the example, and a version such as ‘0.01’ or ‘0.01.0’ will be handled as if it were ‘0.1-0’. It is not a decimal number, so for example 0.9 < 0.75 since 9 < 75.

The mandatory ‘License’ field is discussed in the next subsection.

The mandatory ‘Title’ field should give a short description of the package. Some package listings may truncate the title to 65 characters. It should use title case (that is, use capitals for the principal words: tools::toTitleCase can help you with this), not use any markup, not have any continuation lines, and not end in a period (unless part of …). Do not repeat the package name: it is often used prefixed by the name. Refer to other packages and external software in single quotes, and to book titles (and similar) in double quotes.

The mandatory ‘Description’ field should give a comprehensive description of what the package does. One can use several (complete) sentences, but only one paragraph. It should be intelligible to all the intended readership (e.g. for a CRAN package to all CRAN users). It is good practice not to start with the package name, ‘This package’ or similar. As with the ‘Title’ field, double quotes should be used for quotations (including titles of books and articles), and single quotes for non-English usage, including names of other packages and external software. This field should also be used for explaining the package name if necessary. URLs should be enclosed in angle brackets, e.g. ‘<https://www.r-project.org>’: see also Specifying URLs.

The mandatory ‘Author’ field describes who wrote the package. It is a plain text field intended for human readers, but not for automatic processing (such as extracting the email addresses of all listed contributors: for that use ‘Authors@R’). Note that all significant contributors must be included: if you wrote an R wrapper for the work of others included in the src directory, you are not the sole (and maybe not even the main) author.

The mandatory ‘Maintainer’ field should give a single name followed by a valid (RFC 2822) email address in angle brackets. It should not end in a period or comma. This field is what is reported by the maintainer function and used by bug.report. For a CRAN package it should be a person, not a mailing list and not a corporate entity: do ensure that it is valid and will remain valid for the lifetime of the package.

Note that the display name (the part before the address in angle brackets) should be enclosed in double quotes if it contains non-alphanumeric characters such as comma or period. (The current standard, RFC 5322, allows periods but RFC 2822 did not.)

Both ‘Author’ and ‘Maintainer’ fields can be omitted if a suitable ‘Authors@R’ field is given. This field can be used to provide a refined and machine-readable description of the package “authors” (in particular specifying their precise roles), via suitable R code. It should create an object of class "person", by either a call to person or a series of calls (one per “author”) concatenated by c(): see the example DESCRIPTION file above. The roles can include ‘"aut"’ (author) for full authors, ‘"cre"’ (creator) for the package maintainer, and ‘"ctb"’ (contributor) for other contributors, ‘"cph"’ (copyright holder, which should be the legal name for an institution or corporate body), among others. See ?person for more information. Note that no role is assumed by default. Auto-generated package citation information takes advantage of this specification. The ‘Author’ and ‘Maintainer’ fields are auto-generated from it if needed when building5 or installing. Note that for CRAN submissions, providing ‘Authors@R’ is required, and providing ORCID identifiers (see https://orcid.org/) where possible is strongly encouraged.

An optional ‘Copyright’ field can be used where the copyright holder(s) are not the authors. If necessary, this can refer to an installed file: the convention is to use file inst/COPYRIGHTS.

The optional ‘Date’ field gives the release date of the current version of the package. It is strongly recommended6 to use the ‘yyyy-mm-dd’ format conforming to the ISO 8601 standard.

The ‘Depends’, ‘Imports’, ‘Suggests’, ‘Enhances’, ‘LinkingTo’ and ‘Additional_repositories’ fields are discussed in a later subsection.

Dependencies external to the R system should be listed in the ‘SystemRequirements’ field, possibly amplified in a separate README file. This includes specifying a non-default C++ standard and the need for GNU make.

The ‘URL’ field may give a list of URLs separated by commas or whitespace, for example the homepage of the author or a page where additional material describing the software can be found. These URLs are converted to active hyperlinks in CRAN package listings. See Specifying URLs.

The ‘BugReports’ field may contain a single URL to which bug reports about the package should be submitted. This URL will be used by bug.report instead of sending an email to the maintainer. A browser is opened for a ‘http://’ or ‘https://URL. To specify another email address for bug reports, use ‘Contact’ instead: however bug.report will try to extract an email address (preferably from a ‘mailto:’ URL or enclosed in angle brackets) from ‘BugReports’.

Base and recommended packages (i.e., packages contained in the R source distribution or available from CRAN and recommended to be included in every binary distribution of R) have a ‘Priority’ field with value ‘base’ or ‘recommended’, respectively. These priorities must not be used by other packages.

A ‘Collate’ field can be used for controlling the collation order for the R code files in a package when these are processed for package installation. The default is to collate according to the ‘C’ locale. If present, the collate specification must list all R code files in the package (taking possible OS-specific subdirectories into account, see Package subdirectories) as a whitespace separated list of file paths relative to the R subdirectory. Paths containing white space or quotes need to be quoted. An OS-specific collation field (‘Collate.unix’ or ‘Collate.windows’) will be used in preference to ‘Collate’.

The ‘LazyData’ logical field controls whether the R datasets use lazy-loading. A ‘LazyLoad’ field was used in versions prior to 2.14.0, but now is ignored.

The ‘KeepSource’ logical field controls if the package code is sourced using keep.source = TRUE or FALSE: it might be needed exceptionally for a package designed to always be used with keep.source = TRUE.

The ‘ByteCompile’ logical field controls if the package R code is to be byte-compiled on installation: the default is to byte-compile. This can be overridden by installing with flag --no-byte-compile.

The ‘UseLTO’ logical field is used to indicate if source code in the package7 is to be compiled with Link-Time Optimization (see Using Link-time Optimization) if R was installed with --enable-lto (default true) or --enable-lto=R (default false) (or on Windows8 if LTO_OPT is set in MkRules). This can be overridden by the flags --use-LTO and --no-use-LTO. LTO is said to give most size and performance improvements for large and complex (heavily templated) C++ projects.

The ‘StagedInstall’ logical field controls if package installation is ‘staged’, that is done to a temporary location and moved to the final location when successfully completed. This field was introduced in R 3.6.0 and it true by default: it is considered to be a temporary measure which may be withdrawn in future.

The ‘ZipData’ logical field has been ignored since R 2.13.0.

The ‘Biarch’ logical field is used on Windows to select the INSTALL option --force-biarch for this package. Not currently relevant.

The ‘BuildVignettes’ logical field can be set to a false value to stop R CMD build from attempting to build the vignettes, as well as preventing9 R CMD check from testing this. This should only be used exceptionally, for example if the PDFs include large figures which are not part of the package sources (and hence only in packages which do not have an Open Source license).

The ‘VignetteBuilder’ field names (in a comma-separated list) packages that provide an engine for building vignettes. These may include the current package, or ones listed in ‘Depends’, ‘Suggests’ or ‘Imports’. The utils package is always implicitly appended. See Non-Sweave vignettes for details. Note that if, for example, a vignette has engine ‘knitr::rmarkdown’, then knitr provides the engine but both knitr and rmarkdown are needed for using it, so both these packages need to be in the ‘VignetteBuilder’ field and at least suggested (as rmarkdown is only suggested by knitr, and hence not available automatically along with it). Many packages using knitr also need the package formatR which it suggests and so the user package needs to do so too and include this in ‘VignetteBuilder’.

If the DESCRIPTION file is not entirely in ASCII it should contain an ‘Encoding’ field specifying an encoding. This is used as the encoding of the DESCRIPTION file itself and of the R and NAMESPACE files, and as the default encoding of .Rd files. The examples are assumed to be in this encoding when running R CMD check, and it is used for the encoding of the CITATION file. Only encoding names latin1 and and UTF-8 are known to be portable. (Do not specify an encoding unless one is actually needed: doing so makes the package less portable. If a package has a specified encoding, you should run R CMD build etc in a locale using that encoding.)

The ‘NeedsCompilation’ field should be set to "yes" if the package contains native code which needs to be compiled, otherwise "no" (when the package could be installed from source on any platform without additional tools). This is used by install.packages(type = "both") in R >= 2.15.2 on platforms where binary packages are the norm: it is normally set by R CMD build or the repository assuming compilation is required if and only if the package has a src directory.

The ‘OS_type’ field specifies the OS(es) for which the package is intended. If present, it should be one of unix or windows, and indicates that the package can only be installed on a platform with ‘.Platform$OS.type’ having that value.

The ‘Type’ field specifies the type of the package: see Package types.

One can add subject classifications for the content of the package using the fields ‘Classification/ACM’ or ‘Classification/ACM-2012’ (using the Computing Classification System of the Association for Computing Machinery, https://www.acm.org/publications/class-2012; the former refers to the 1998 version), ‘Classification/JEL’ (the Journal of Economic Literature Classification System, https://www.aeaweb.org/econlit/jelCodes.php, or ‘Classification/MSC’ or ‘Classification/MSC-2010’ (the Mathematics Subject Classification of the American Mathematical Society, https://mathscinet.ams.org/msc/msc2010.html; the former refers to the 2000 version). The subject classifications should be comma-separated lists of the respective classification codes, e.g., ‘Classification/ACM: G.4, H.2.8, I.5.1’.

A ‘Language’ field can be used to indicate if the package documentation is not in English: this should be a comma-separated list of standard (not private use or grandfathered) IETF language tags as currently defined by RFC 5646 (https://www.rfc-editor.org/rfc/rfc5646, see also https://en.wikipedia.org/wiki/IETF_language_tag), i.e., use language subtags which in essence are 2-letter ISO 639-1 (https://en.wikipedia.org/wiki/ISO_639-1) or 3-letter ISO 639-3 (https://en.wikipedia.org/wiki/ISO_639-3) language codes.

An ‘RdMacros’ field can be used to hold a comma-separated list of packages from which the current package will import Rd macro definitions. These package should also be listed in ‘Imports’ (or ‘Depends’). The macros in these packages will be imported after the system macros, in the order listed in the ‘RdMacros’ field, before any macro definitions in the current package are loaded. Macro definitions in individual .Rd files in the man directory are loaded last, and are local to later parts of that file. In case of duplicates, the last loaded definition will be used.10 Both R CMD Rd2pdf and R CMD Rdconv have an optional flag --RdMacros=pkglist. The option is also a comma-separated list of package names, and has priority over the value given in DESCRIPTION. Packages using Rd macros should depend on R 3.2.0 or later.

Note: There should be no ‘Built’ or ‘Packaged’ fields, as these are added by the package management tools.

There is no restriction on the use of other fields not mentioned here (but using other capitalizations of these field names would cause confusion). Fields Note, Contact (for contacting the authors/developers11) and MailingList are in common use. Some repositories (including CRAN and R-forge) add their own fields.


1.1.2 Licensing

Licensing for a package which might be distributed is an important but potentially complex subject.

It is very important that you include license information! Otherwise, it may not even be legally correct for others to distribute copies of the package, let alone use it.

The package management tools use the concept of ‘free or open source software’ (FOSS, e.g., https://en.wikipedia.org/wiki/FOSS) licenses: the idea being that some users of R and its packages want to restrict themselves to such software. Others need to ensure that there are no restrictions stopping them using a package, e.g. forbidding commercial or military use. It is a central tenet of FOSS software that there are no restrictions on users nor usage.

Do not use the ‘License’ field for information on copyright holders: if needed, use a ‘Copyright’ field.

The mandatory ‘License’ field in the DESCRIPTION file should specify the license of the package in a standardized form. Alternatives are indicated via vertical bars. Individual specifications must be one of

  • One of the “standard” short specifications
    GPL-2 GPL-3 LGPL-2 LGPL-2.1 LGPL-3 AGPL-3 Artistic-2.0
    BSD_2_clause BSD_3_clause MIT
    

    as made available via https://www.R-project.org/Licenses/ and contained in subdirectory share/licenses of the R source or home directory.

  • The names or abbreviations of other licenses contained in the license data base in file share/licenses/license.db in the R source or home directory, possibly (for versioned licenses) followed by a version restriction of the form ‘(op v)’ with ‘op’ one of the comparison operators ‘<’, ‘<=’, ‘>’, ‘>=’, ‘==’, or ‘!=’ and ‘v’ a numeric version specification (strings of non-negative integers separated by ‘.’), possibly combined via,’ (see below for an example). For versioned licenses, one can also specify the name followed by the version, or combine an existing abbreviation and the version with a ‘-’.

    Abbreviations GPL and LGPL are ambiguous and usually12 taken to mean any version of the license: but it is better not to use them.

  • One of the strings ‘file LICENSE’ or ‘file LICENCE’ referring to a file named LICENSE or LICENCE in the package (source and installation) top-level directory.
  • The string ‘Unlimited’, meaning that there are no restrictions on distribution or use other than those imposed by relevant laws (including copyright laws).

Multiple licences can be specified separated by ‘|’ (surrounded by spaces) in which case the user can choose any of the alternatives.

If a package license restricts a base license (where permitted, e.g., using GPL-3 or AGPL-3 with an attribution clause), the additional terms should be placed in file LICENSE (or LICENCE), and the string ‘+ file LICENSE’ (or ‘+ file LICENCE’, respectively) should be appended to the corresponding individual license specification (preferably with the ‘+’ surrounded by spaces). Note that several commonly used licenses do not permit restrictions: this includes GPL-2 and hence any specification which includes it.

Examples of standardized specifications include

License: GPL-2
License: LGPL (>= 2.0, < 3) | Mozilla Public License
License: GPL-2 | file LICENCE
License: GPL (>= 2) | BSD_3_clause + file LICENSE
License: Artistic-2.0 | AGPL-3 + file LICENSE

Please note in particular that “Public domain” is not a valid license, since it is not recognized in some jurisdictions.

Please ensure that the license you choose also covers any dependencies (including system dependencies) of your package: it is particularly important that any restrictions on the use of such dependencies are evident to people reading your DESCRIPTION file.

Fields ‘License_is_FOSS’ and ‘License_restricts_use’ may be added by repositories where information cannot be computed from the name of the license. ‘License_is_FOSS: yes’ is used for licenses which are known to be FOSS, and ‘License_restricts_use’ can have values ‘yes’ or ‘no’ if the LICENSE file is known to restrict users or usage, or known not to. These are used by, e.g., the available.packages filters.

The optional file LICENSE/LICENCE contains a copy of the license of the package. To avoid any confusion only include such a file if it is referred to in the ‘License’ field of the DESCRIPTION file.

Whereas you should feel free to include a license file in your source distribution, please do not arrange to install yet another copy of the GNU COPYING or COPYING.LIB files but refer to the copies on https://www.R-project.org/Licenses/ and included in the R distribution (in directory share/licenses). Since files named LICENSE or LICENCE will be installed, do not use these names for standard license files. To include comments about the licensing rather than the body of a license, use a file named something like LICENSE.note.

A few “standard” licenses are rather license templates which need additional information to be completed via+ file LICENSE’ (with the ‘+’ surrounded by spaces). Where the additional information is ‘COPYRIGHT HOLDER’, this must give the actual legal entities (not something vague like ‘Name-of-package authors’): if more than one they should be listed in decreasing order of contribution.


1.1.3 Package Dependencies

The ‘Depends’ field gives a comma-separated list of package names which this package depends on. Those packages will be attached before the current package when library or require is called. Each package name may be optionally followed by a comment in parentheses specifying a version requirement. The comment should contain a comparison operator, whitespace and a valid version number, e.g. ‘MASS (>= 3.1-20)’.

The ‘Depends’ field can also specify a dependence on a certain version of R — e.g., if the package works only with R version 4.0.0 or later, include ‘R (>= 4.0)’ in the ‘Depends’ field. (As here, trailing zeroes can be dropped and it is recommended that they are.) You can also require a certain SVN revision for R-devel or R-patched, e.g. ‘R (>= 2.14.0), R (>= r56550)’ requires a version later than R-devel of late July 2011 (including released versions of 2.14.0).

It makes no sense to declare a dependence on R without a version specification, nor on the package base: this is an R package and package base is always available.

A package or ‘R’ can appear more than once in the ‘Depends’ field, for example to give upper and lower bounds on acceptable versions.

It is inadvisable to use a dependence on R with patch level (the third digit) other than zero. Doing so with packages which others depend on will cause the other packages to become unusable under earlier versions in the series, and e.g. versions 4.x.1 are widely used throughout the Northern Hemisphere academic year.

Both library and the R package checking facilities use this field: hence it is an error to use improper syntax or misuse the ‘Depends’ field for comments on other software that might be needed. The R INSTALL facilities check if the version of R used is recent enough for the package being installed, and the list of packages which is specified will be attached (after checking version requirements) before the current package.

The ‘Imports’ field lists packages whose namespaces are imported from (as specified in the NAMESPACE file) but which do not need to be attached. Namespaces accessed by the ‘::’ and ‘:::’ operators must be listed here, or in ‘Suggests’ or ‘Enhances’ (see below). Ideally this field will include all the standard packages that are used, and it is important to include S4-using packages (as their class definitions can change and the DESCRIPTION file is used to decide which packages to re-install when this happens). Packages declared in the ‘Depends’ field should not also be in the ‘Imports’ field. Version requirements can be specified and are checked when the namespace is loaded.

The ‘Suggests’ field uses the same syntax as ‘Depends’ and lists packages that are not necessarily needed. This includes packages used only in examples, tests or vignettes (see Writing package vignettes), and packages loaded in the body of functions. E.g., suppose an example13 from package foo uses a dataset from package bar. Then it is not necessary to have bar use foo unless one wants to execute all the examples/tests/vignettes: it is useful to have bar, but not necessary. Version requirements can be specified but should be checked by the code which uses the package.

Finally, the ‘Enhances’ field lists packages “enhanced” by the package at hand, e.g., by providing methods for classes from these packages, or ways to handle objects from these packages (so several packages have ‘Enhances: chron’ because they can handle datetime objects from chron even though they prefer R’s native datetime functions). Version requirements can be specified, but are currently not used. Such packages cannot be required to check the package: any tests which use them must be conditional on the presence of the package. (If your tests use e.g. a dataset from another package it should be in ‘Suggests’ and not ‘Enhances’.)

The general rules are

  • A package should be listed in only one of these fields.
  • Packages whose namespace only is needed to load the package using library(pkgname) should be listed in the ‘Imports’ field and not in the ‘Depends’ field. Packages listed in import or importFrom directives in the NAMESPACE file should almost always be in ‘Imports’ and not ‘Depends’.
  • Packages that need to be attached to successfully load the package using library(pkgname) must be listed in the ‘Depends’ field.
  • All packages that are needed14 to successfully run R CMD check on the package must be listed in one of ‘Depends’ or ‘Suggests’ or ‘Imports’. Packages used to run examples or tests conditionally (e.g. via if(require(pkgname))) should be listed in ‘Suggests’ or ‘Enhances’. (This allows checkers to ensure that all the packages needed for a complete check are installed.)
  • Packages needed to use datasets from the package should be in ‘Imports’: this includes those needed to define S4 classes used.

In particular, packages providing “only” data for examples or vignettes should be listed in ‘Suggests’ rather than ‘Depends’ in order to make lean installations possible.

Version dependencies in the ‘Depends’ and ‘Imports’ fields are used by library when it loads the package, and install.packages checks versions for the ‘Depends’, ‘Imports’ and (for dependencies = TRUE) ‘Suggests’ fields.

It is important that the information in these fields is complete and accurate: it is for example used to compute which packages depend on an updated package and which packages can safely be installed in parallel.

This scheme was developed before all packages had namespaces (R 2.14.0 in October 2011), and good practice changed once that was in place.

Field ‘Depends’ should nowadays be used rarely, only for packages which are intended to be put on the search path to make their facilities available to the end user (and not to the package itself): for example it makes sense that a user of package latticeExtra would want the functions of package lattice made available.

Almost always packages mentioned in ‘Depends’ should also be imported from in the NAMESPACE file: this ensures that any needed parts of those packages are available when some other package imports the current package.

The ‘Imports’ field should not contain packages which are not imported from (via the NAMESPACE file or :: or ::: operators), as all the packages listed in that field need to be installed for the current package to be installed. (This is checked by R CMD check.)

R code in the package should call library or require only exceptionally. Such calls are never needed for packages listed in ‘Depends’ as they will already be on the search path. It used to be common practice to use require calls for packages listed in ‘Suggests’ in functions which used their functionality, but nowadays it is better to access such functionality via :: calls.

A package that wishes to make use of header files in other packages to compile its C/C++ code needs to declare them as a comma-separated list in the field ‘LinkingTo’ in the DESCRIPTION file. For example

LinkingTo: link1, link2

The ‘LinkingTo’ field can have a version requirement which is checked at installation.

Specifying a package in ‘LinkingTo’ suffices if these are C/C++ headers containing source code or static linking is done at installation: the packages do not need to be (and usually should not be) listed in the ‘Depends’ or ‘Imports’ fields. This includes CRAN package BH and almost all users of RcppArmadillo and RcppEigen. Note that ‘LinkingTo’ applies only to installation: if a packages wishes to use headers to compile code in tests or vignettes the package providing them needs to be listed in ‘Suggests’ or perhaps ‘Depends’.

For another use of ‘LinkingTo’ see Linking to native routines in other packages.

The ‘Additional_repositories’ field is a comma-separated list of repository URLs where the packages named in the other fields may be found. It is currently used by R CMD check to check that the packages can be found, at least as source packages (which can be installed on any platform).


1.1.3.1 Suggested packages

Note that someone wanting to run the examples/tests/vignettes may not have a suggested package available (and it may not even be possible to install it for that platform). The recommendation used to be to make their use conditional via if(require("pkgname")): this is OK if that conditioning is done in examples/tests/vignettes, although using if(requireNamespace("pkgname")) is preferred, if possible.

However, using require for conditioning in package code is not good practice as it alters the search path for the rest of the session and relies on functions in that package not being masked by other require or library calls. It is better practice to use code like

   if (requireNamespace("rgl", quietly = TRUE)) {
      rgl::plot3d(...)
   } else {
      ## do something else not involving rgl.
   }

Note the use of rgl:: as that object would not necessarily be visible (and if it is, it need not be the one from that namespace: plot3d occurs in several other packages). If the intention is to give an error if the suggested package is not available, simply use e.g. rgl::plot3d.

If the conditional code produces print output, function withAutoprint can be useful.

Note that the recommendation to use suggested packages conditionally in tests does also apply to packages used to manage test suites: a notorious example was testthat which in version 1.0.0 contained illegal C++ code and hence could not be installed on standards-compliant platforms.

Some people have assumed that a ‘recommended’ package in ‘Suggests’ can safely be used unconditionally, but this is not so. (R can be installed without recommended packages, and which packages are ‘recommended’ may change.)

As noted above, packages in ‘Enhancesmust be used conditionally and hence objects within them should always be accessed via ::.

On most systems, R CMD check can be run with only those packages declared in ‘Depends’ and ‘Imports’ by setting environment variable _R_CHECK_DEPENDS_ONLY_=true, whereas setting _R_CHECK_SUGGESTS_ONLY_=true also allows suggested packages, but not those in ‘Enhances’ nor those not mentioned in the DESCRIPTION file. It is recommended that a package is checked with each of these set, as well as with neither.

WARNING: Be extremely careful if you do things which would be run at installation time depending on whether suggested packages are available or not—this includes top-level code in R code files, .onLoad functions and the definitions of S4 classes and methods. The problem is that once a namespace of a suggested package is loaded, references to it may be captured in the installed package (most commonly in S4 methods), but the suggested package may not be available when the installed package is used (which especially for binary packages might be on a different machine). Even worse, the problems might not be confined to your package, for the namespaces of your suggested packages will also be loaded whenever any package which imports yours is installed and so may be captured there.


1.1.4 The INDEX file

The optional file INDEX contains a line for each sufficiently interesting object in the package, giving its name and a description (functions such as print methods not usually called explicitly might not be included). Normally this file is missing and the corresponding information is automatically generated from the documentation sources (using tools::Rdindex()) when installing from source.

The file is part of the information given by library(help = pkgname).

Rather than editing this file, it is preferable to put customized information about the package into an overview help page (see Documenting packages) and/or a vignette (see Writing package vignettes).


1.1.5 Package subdirectories

The R subdirectory contains R code files, only. The code files to be installed must start with an ASCII (lower or upper case) letter or digit and have one of the extensions15 .R, .S, .q, .r, or .s. We recommend using .R, as this extension seems to be not used by any other software. It should be possible to read in the files using source(), so R objects must be created by assignments. Note that there need be no connection between the name of the file and the R objects created by it. Ideally, the R code files should only directly assign R objects and definitely should not call functions with side effects such as require and options. If computations are required to create objects these can use code ‘earlier’ in the package (see the ‘Collate’ field) plus functions in the ‘Depends’ packages provided that the objects created do not depend on those packages except via namespace imports.

Extreme care is needed if top-level computations are made to depend on availability or not of other packages. In particular this applies to setMethods and setClass calls. Nor should they depend on the availability of external resources such as downloads.

Two exceptions are allowed: if the R subdirectory contains a file sysdata.rda (a saved image of one or more R objects: please use suitable compression as suggested by tools::resaveRdaFiles, and see also the ‘SysDataCompressionDESCRIPTION field.) this will be lazy-loaded into the namespace environment – this is intended for system datasets that are not intended to be user-accessible via data. Also, files ending in ‘.in’ will be allowed in the R directory to allow a configure script to generate suitable files.

Only ASCII characters (and the control characters tab, form feed, LF and CR) should be used in code files. Other characters are accepted in comments16, but then the comments may not be readable in e.g. a UTF-8 locale. Non-ASCII characters in object names will normally17 fail when the package is installed. Any byte will be allowed in a quoted character string but ‘\uxxxx’ escapes should be used for non-ASCII characters. However, non-ASCII character strings may not be usable in some locales and may display incorrectly in others.

Various R functions in a package can be used to initialize and clean up. See Load hooks.

The man subdirectory should contain (only) documentation files for the objects in the package in R documentation (Rd) format. The documentation filenames must start with an ASCII (lower or upper case) letter or digit and have the extension .Rd (the default) or .rd. Further, the names must be valid in ‘file://’ URLs, which means18 they must be entirely ASCII and not contain ‘%’. See Writing R documentation files, for more information. Note that all user-level objects in a package should be documented; if a package pkg contains user-level objects which are for “internal” use only, it should provide a file pkg-internal.Rd which documents all such objects, and clearly states that these are not meant to be called by the user. See e.g. the sources for package grid in the R distribution. Note that packages which use internal objects extensively should not export those objects from their namespace, when they do not need to be documented (see Package namespaces).

Having a man directory containing no documentation files may give an installation error.

The man subdirectory may contain a subdirectory named macros; this will contain source for user-defined Rd macros. (See User-defined macros.) These use the Rd format, but may not contain anything but macro definitions, comments and whitespace.

The R and man subdirectories may contain OS-specific subdirectories named unix or windows.

The sources and headers for the compiled code are in src, plus optionally a file Makevars or Makefile (or for use on Windows, with extension .win or .ucrt). When a package is installed using R CMD INSTALL, make is used to control compilation and linking into a shared object for loading into R. There are default make variables and rules for this (determined when R is configured and recorded in R_HOME/etcR_ARCH/Makeconf), providing support for C, C++, fixed- or free-form Fortran, Objective C and Objective C++19 with associated extensions .c, .cc or .cpp, .f, .f90 or .f95,20 .m, and .mm, respectively. We recommend using .h for headers, also for C++21 or Fortran include files. (Use of extension .C for C++ is no longer supported.) Files in the src directory should not be hidden (start with a dot), and hidden files will under some versions of R be ignored.

It is not portable (and may not be possible at all) to mix all these languages in a single package. Because R itself uses it, we know that C and fixed-form Fortran can be used together, and mixing C, C++ and Fortran usually work for the platform’s native compilers.

If your code needs to depend on the platform there are certain defines which can be used in C or C++. On all Windows builds (even 64-bit ones) ‘_WIN32’ will be defined: on 64-bit Windows builds also ‘_WIN64’. For Windows on ARM, test for ‘_M_ARM64’ or both ‘_WIN32’ and ‘__aarch64__’. On macOS ‘__APPLE__’ is defined22; for an ‘Apple Silicon’ platform, test for both ‘__APPLE__’ and ‘__arm64__’.

The default rules can be tweaked by setting macros23 in a file src/Makevars (see Using Makevars). Note that this mechanism should be general enough to eliminate the need for a package-specific src/Makefile. If such a file is to be distributed, considerable care is needed to make it general enough to work on all R platforms. If it has any targets at all, it should have an appropriate first target named ‘all’ and a (possibly empty) target ‘clean’ which removes all files generated by running make (to be used by ‘R CMD INSTALL --clean’ and ‘R CMD INSTALL --preclean’). There are platform-specific file names on Windows: src/Makevars.win takes precedence over src/Makevars and src/Makefile.win must be used. Since R 4.2.0, src/Makevars.ucrt takes precedence over src/Makevars.win and src/Makefile.ucrt takes precedence over src/Makefile.win. src/Makevars.ucrt and src/Makefile.ucrt will be ignored by earlier versions of R, and hence can be used to provide content specific to UCRT or Rtools42 and newer, but the support for .ucrt files may be removed in the future when building packages from source on the older versions of R will no longer be needed, and hence the files may be renamed back to .win. Some make programs require makefiles to have a complete final line, including a newline.

A few packages use the src directory for purposes other than making a shared object (e.g. to create executables). Such packages should have files src/Makefile and src/Makefile.win or src/Makefile.ucrt (unless intended for only Unix-alikes or only Windows). Note that on Unix such makefiles are included after R_HOME/etc/R_ARCH/Makeconf so all the usual R macros and make rules are available – for example C compilation will by default use the C compiler and flags with which R was configured. This also applies on Windows as from R 4.3.0: packages intended to be used with earlier versions should include that file themselves.

The order of inclusion of makefiles for a package which does not have a src/Makefile file is

Unix-alikeWindows
src/Makevarssrc/Makevars.ucrt, src/Makevars.win
R_HOME/etc/R_ARCH/MakeconfR_HOME/etc/R_ARCH/Makeconf
R_MAKEVARS_SITE, R_HOME/etc/R_ARCH/Makevars.siteR_MAKEVARS_SITE, R_HOME/etc/R_ARCH/Makevars.site
R_HOME/share/make/shlib.mkR_HOME/share/make/winshlib.mk
R_MAKEVARS_USER, ~/.R/Makevars-platform, ~/.R/MakevarsR_MAKEVARS_USER, ~/.R/Makevars.ucrt, ~/.R/Makevars.win64, ~/.R/Makevars.win

For those which do, it is

R_HOME/etc/R_ARCH/MakeconfR_HOME/etc/R_ARCH/Makeconf
R_MAKEVARS_SITE, R_HOME/etc/R_ARCH/Makevars.siteR_MAKEVARS_SITE, R_HOME/etc/R_ARCH/Makevars.site
src/Makefilesrc/Makefile.ucrt, src/Makefile.win
R_MAKEVARS_USER, ~/.R/Makevars-platform, ~/.R/MakevarsR_MAKEVARS_USER, ~/.R/Makevars.ucrt, ~/.R/Makevars.win64, ~/.R/Makevars.win

Items in capitals are environment variables: those separated by commas are alternatives looked for in the order shown.

In very special cases packages may create binary files other than the shared objects/DLLs in the src directory. Such files will not be installed in a multi-architecture setting since R CMD INSTALL --libs-only is used to merge multiple sub-architectures and it only copies shared objects/DLLs. If a package wants to install other binaries (for example executable programs), it should provide an R script src/install.libs.R which will be run as part of the installation in the src build directory instead of copying the shared objects/DLLs. The script is run in a separate R environment containing the following variables: R_PACKAGE_NAME (the name of the package), R_PACKAGE_SOURCE (the path to the source directory of the package), R_PACKAGE_DIR (the path of the target installation directory of the package), R_ARCH (the arch-dependent part of the path, often empty), SHLIB_EXT (the extension of shared objects) and WINDOWS (TRUE on Windows, FALSE elsewhere). Something close to the default behavior could be replicated with the following src/install.libs.R file:

files <- Sys.glob(paste0("*", SHLIB_EXT))
dest <- file.path(R_PACKAGE_DIR, paste0('libs', R_ARCH))
dir.create(dest, recursive = TRUE, showWarnings = FALSE)
file.copy(files, dest, overwrite = TRUE)
if(file.exists("symbols.rds"))
    file.copy("symbols.rds", dest, overwrite = TRUE)

On the other hand, executable programs could be installed along the lines of

execs <- c("one", "two", "three")
if(WINDOWS) execs <- paste0(execs, ".exe")
if ( any(file.exists(execs)) ) {
  dest <- file.path(R_PACKAGE_DIR,  paste0('bin', R_ARCH))
  dir.create(dest, recursive = TRUE, showWarnings = FALSE)
  file.copy(execs, dest, overwrite = TRUE)
}

Note the use of architecture-specific subdirectories of bin where needed. (Executables should installed under a bin directory and not under libs. It is good practice to check that they can be executed as part of the installation script, so a broken package is not installed.)

The data subdirectory is for data files: See Data in packages.

The demo subdirectory is for R scripts (for running via demo()) that demonstrate some of the functionality of the package. Demos may be interactive and are not checked automatically, so if testing is desired use code in the tests directory to achieve this. The script files must start with a (lower or upper case) letter and have one of the extensions .R or .r. If present, the demo subdirectory should also have a 00Index file with one line for each demo, giving its name and a description separated by a tab or at least three spaces. (This index file is not generated automatically.) Note that a demo does not have a specified encoding and so should be an ASCII file (see Encoding issues). Function demo() will use the package encoding if there is one, but this is mainly useful for non-ASCII comments.

The contents of the inst subdirectory will be copied recursively to the installation directory. Subdirectories of inst should not interfere with those used by R (currently, R, data, demo, exec, libs, man, help, html and Meta, and earlier versions used latex, R-ex). The copying of the inst happens after src is built so its Makefile can create files to be installed. To exclude files from being installed, one can specify a list of exclude patterns in file .Rinstignore in the top-level source directory. These patterns should be Perl-like regular expressions (see the help for regexp in R for the precise details), one per line, to be matched case-insensitively against the file and directory paths, e.g. doc/.*[.]png$ will exclude all PNG files in inst/doc based on the extension.

Note that with the exceptions of INDEX, LICENSE/LICENCE and NEWS, information files at the top level of the package will not be installed and so not be known to users of Windows and macOS compiled packages (and not seen by those who use R CMD INSTALL or install.packages() on the tarball). So any information files you wish an end user to see should be included in inst. Note that if the named exceptions also occur in inst, the version in inst will be that seen in the installed package.

Things you might like to add to inst are a CITATION file for use by the citation function, and a NEWS.Rd file for use by the news function. See its help page for the specific format restrictions of the NEWS.Rd file.

Another file sometimes needed in inst is AUTHORS or COPYRIGHTS to specify the authors or copyright holders when this is too complex to put in the DESCRIPTION file.

Subdirectory tests is for additional package-specific test code, similar to the specific tests that come with the R distribution. Test code can either be provided directly in a .R (or .r as from R 3.4.0) file, or via a .Rin file containing code which in turn creates the corresponding .R file (e.g., by collecting all function objects in the package and then calling them with the strangest arguments). The results of running a .R file are written to a .Rout file. If there is a corresponding24 .Rout.save file, these two are compared, with differences being reported but not causing an error. The directory tests is copied to the check area, and the tests are run with the copy as the working directory and with R_LIBS set to ensure that the copy of the package installed during testing will be found by library(pkg_name). Note that the package-specific tests are run in a vanilla R session without setting the random-number seed, so tests which use random numbers will need to set the seed to obtain reproducible results (and it can be helpful to do so in all cases, to avoid occasional failures when tests are run).

If directory tests has a subdirectory Examples containing a file pkg-Ex.Rout.save, this is compared to the output file for running the examples when the latter are checked. Reference output should be produced without having the --timings option set (and note that --as-cran sets it).

If reference output is included for examples, tests or vignettes do make sure that it is fully reproducible, as it will be compared verbatim to that produced in a check run, unless the ‘IGNORE_RDIFF’ markup is used. Things which trip up maintainers include displayed version numbers from loading other packages, printing numerical results to an unreproducibly high precision and printing timings. Another trap is small values which are in fact rounding error from zero: consider using zapsmall.

Subdirectory exec could contain additional executable scripts the package needs, typically scripts for interpreters such as the shell, Perl, or Tcl. NB: only files (and not directories) under exec are installed (and those with names starting with a dot are ignored), and they are all marked as executable (mode 755, moderated by ‘umask’) on POSIX platforms. Note too that this is not suitable for executable programs since some platforms support multiple architectures using the same installed package directory.

Subdirectory po is used for files related to localization: see Internationalization.

Subdirectory tools is the preferred place for auxiliary files needed during configuration, and also for sources need to re-create scripts (e.g. M4 files for autoconf: some prefer to put those in a subdirectory m4 of tools).


1.1.6 Data in packages

The data subdirectory is for data files, either to be made available via lazy-loading or for loading using data(). (The choice is made by the ‘LazyData’ field in the DESCRIPTION file: the default is not to do so.) It should not be used for other data files needed by the package, and the convention has grown up to use directory inst/extdata for such files.

Data files can have one of three types as indicated by their extension: plain R code (.R or .r), tables (.tab, .txt, or .csv, see ?data for the file formats, and note that .csv is not the standard25 CSV format), or save() images (.RData or .rda). The files should not be hidden (have names starting with a dot). Note that R code should be if possible “self-sufficient” and not make use of extra functionality provided by the package, so that the data file can also be used without having to load the package or its namespace: it should run as silently as possible and not change the search() path by attaching packages or other environments.

Images (extensions .RData26 or .rda) can contain references to the namespaces of packages that were used to create them. Preferably there should be no such references in data files, and in any case they should only be to packages listed in the Depends and Imports fields, as otherwise it may be impossible to install the package. To check for such references, load all the images into a vanilla R session, run str() on all the datasets, and look at the output of loadedNamespaces().

Particular care is needed where a dataset or one of its components is of an S4 class, especially if the class is defined in a different package. First, the package containing the class definition has to be available to do useful things with the dataset, so that package must be listed in Imports or Depends (even if this gives a check warning about unused imports). Second, the definition of an S4 class can change, and often is unnoticed when in a package with a different author. So it may be wiser to use the .R form and use that to create the dataset object when needed (loading package namespaces but not attaching them by using requireNamespace(pkg, quietly = TRUE) and using pkg:: to refer to objects in the namespace).

If you are not using ‘LazyData’ and either your data files are large or e.g., you use data/foo.R scripts to produce your data, loading your namespace, you can speed up installation by providing a file datalist in the data subdirectory. This should have one line per topic that data() will find, in the format ‘foo’ if data(foo) provides ‘foo’, or ‘foo: bar bah’ if data(foo) provides ‘bar’ and ‘bah’. R CMD build will automatically add a datalist file to data directories of over 1Mb, using the function tools::add_datalist.

Tables (.tab, .txt, or .csv files) can be compressed by gzip, bzip2 or xz, optionally with additional extension .gz, .bz2 or .xz.

If your package is to be distributed, do consider the resource implications of large datasets for your users: they can make packages very slow to download and use up unwelcome amounts of storage space, as well as taking many seconds to load. It is normally best to distribute large datasets as .rda images prepared by save(, compress = TRUE) (the default). Using bzip2 or xz compression will usually reduce the size of both the package tarball and the installed package, in some cases by a factor of two or more.

Package tools has a couple of functions to help with data images: checkRdaFiles reports on the way the image was saved, and resaveRdaFiles will re-save with a different type of compression, including choosing the best type for that particular image.

Many packages using ‘LazyData’ will benefit from using a form of compression other than gzip in the installed lazy-loading database. This can be selected by the --data-compress option to R CMD INSTALL or by using the ‘LazyDataCompression’ field in the DESCRIPTION file. Useful values are bzip2, xz and the default, gzip: value none is also accepted. The only way to discover which is best is to try them all and look at the size of the pkgname/data/Rdata.rdb file. A function to do that (quoting sizes in KB) is

CheckLazyDataCompression <- function(pkg)
{
    pkg_name <- sub("_.*", "", pkg)
    lib <- tempfile(); dir.create(lib)
    zs <- c("gzip", "bzip2", "xz")
    res <- integer(3); names(res) <- zs
    for (z in zs) {
        opts <- c(paste0("--data-compress=", z),
                  "--no-libs", "--no-help", "--no-demo", "--no-exec", "--no-test-load")
        install.packages(pkg, lib, INSTALL_opts = opts, repos = NULL, quiet = TRUE)
        res[z] <- file.size(file.path(lib, pkg_name, "data", "Rdata.rdb"))
    }
    ceiling(res/1024)
}

(applied to a source package without any ‘LazyDataCompression’ field). R CMD check will warn if it finds a pkgname/data/Rdata.rdb file of more than 5MB without ‘LazyDataCompression’ being set. If you see that, run CheckLazyDataCompression() and set the field – to gzip in the unlikely event27 that is the best choice.

The analogue for sysdata.rda is field ‘SysDataCompression’: the default is xz for files bigger than 1MB otherwise gzip.

Lazy-loading is not supported for very large datasets (those which when serialized exceed 2GB, the limit for the format on 32-bit platforms).


1.1.7 Non-R scripts in packages

Code which needs to be compiled (C, C++, Fortran …) is included in the src subdirectory and discussed elsewhere in this document.

Subdirectory exec could be used for scripts for interpreters such as the shell, BUGS, JavaScript, Matlab, Perl, PHP (amap), Python or Tcl (Simile), or even R. However, it seems more common to use the inst directory, for example WriteXLS/inst/Perl, NMF/inst/m-files, RnavGraph/inst/tcl, RProtoBuf/inst/python and emdbook/inst/BUGS and gridSVG/inst/js.

Java code is a special case: except for very small programs, .java files should be byte-compiled (to a .class file) and distributed as part of a .jar file: the conventional location for the .jar file(s) is inst/java. It is desirable (and required under an Open Source license) to make the Java source files available: this is best done in a top-level java directory in the package—the source files should not be installed.

If your package requires one of these interpreters or an extension then this should be declared in the ‘SystemRequirements’ field of its DESCRIPTION file. (Users of Java most often do so via rJava, when depending on/importing that suffices unless there is a version requirement on Java code in the package.)

Windows and Mac users should be aware that the Tcl extensions ‘BWidget’ and ‘Tktable’ (which have sometimes been included in the Windows28 and macOS R installers) are extensions and do need to be declared (and that ‘Tktable’ is less widely available than it used to be, including not in the main repositories for major Linux distributions). ‘BWidget’ needs to be installed by the user on other OSes. This is fairly easy to do: first find the Tcl search path:

library(tcltk)
strsplit(tclvalue('auto_path'), " ")[[1]]

then download the sources from https://sourceforge.net/projects/tcllib/files/BWidget/ and in a terminal run something like

tar xf bwidget-1.9.14.tar.gz
sudo mv bwidget-1.9.14 /usr/local/lib

substituting a location on the Tcl search path for /usr/local/lib if needed. (If no location on that search path is writeable, you will need to add one each time ‘BWidget’ is to be used with tcltk::addTclPath().)

To (silently) test for the presence of ‘Tktable’ one can use

library(tcltk)
have_tktable <- !isFALSE(suppressWarnings(tclRequire('Tktable')))

Installing ‘Tktable’ needs a C compiler and the Tk headers (not necessarily installed with Tcl/Tk). At the time of writing the latest sources (from 2008) were available from https://sourceforge.net/projects/tktable/files/tktable/2.10/Tktable2.10.tar.gz/download, but needed patching for current Tk (8.6.11, but not 8.6.10) – a patch can be found at https://www.stats.ox.ac.uk/pub/bdr/Tktable/. For a system installation of Tk you may need to install ‘Tktable’ as ‘root’ as on e.g. Fedora all the locations on auto_path are owned by ‘root’.


1.1.8 Specifying URLs

URLs in many places in the package documentation will be converted to clickable hyperlinks in at least some of their renderings. So care is needed that their forms are correct and portable.

The full URL should be given, including the scheme (often ‘http://’ or ‘https://’) and a final ‘/’ for references to directories.

Spaces in URLs are not portable and how they are handled does vary by HTTP server and by client. There should be no space in the host part of an ‘http://’ URL, and spaces in the remainder should be encoded, with each space replaced by ‘%20’.

Reserved characters should be encoded unless used in their reserved sense: see the help on URLencode().

The canonical URL for a CRAN package is

https://cran.r-project.org/package=pkgname

and not a version starting ‘https://cran.r-project.org/web/packages/pkgname’.


1.2 Configure and cleanup

Note that most of this section is specific to Unix-alikes: see the comments later on about the Windows port of R.

If your package needs some system-dependent configuration before installation you can include an executable (Bourne29 shell script configure in your package which (if present) is executed by R CMD INSTALL before any other action is performed. This can be a script created by the Autoconf mechanism, but may also be a script written by yourself. Use this to detect if any nonstandard libraries are present such that corresponding code in the package can be disabled at install time rather than giving error messages when the package is compiled or used. To summarize, the full power of Autoconf is available for your extension package (including variable substitution, searching for libraries, etc.). Background and useful tips on Autoconf and related tools (including pkg-config described below) can be found at https://autotools.info/.

A configure script is run in an environment which has all the environment variables set for an R session (see R_HOME/etc/Renviron) plus R_PACKAGE_NAME (the name of the package), R_PACKAGE_DIR (the path of the target installation directory of the package, a temporary location for staged installs) and R_ARCH (the arch-dependent part of the path, often empty).

Under a Unix-alike only, an executable (Bourne shell) script cleanup is executed as the last thing by R CMD INSTALL if option --clean was given, and by R CMD build when preparing the package for building from its source.

As an example consider we want to use functionality provided by a (C or Fortran) library foo. Using Autoconf, we can create a configure script which checks for the library, sets variable HAVE_FOO to TRUE if it was found and to FALSE otherwise, and then substitutes this value into output files (by replacing instances of ‘@HAVE_FOO@’ in input files with the value of HAVE_FOO). For example, if a function named bar is to be made available by linking against library foo (i.e., using -lfoo), one could use

AC_CHECK_LIB(foo, fun, [HAVE_FOO=TRUE], [HAVE_FOO=FALSE])
AC_SUBST(HAVE_FOO)
......
AC_CONFIG_FILES([foo.R])
AC_OUTPUT

in configure.ac (assuming Autoconf 2.50 or later).

The definition of the respective R function in foo.R.in could be

foo <- function(x) {
    if(!@HAVE_FOO@)
      stop("Sorry, library 'foo' is not available")
    ...

From this file configure creates the actual R source file foo.R looking like

foo <- function(x) {
    if(!FALSE)
      stop("Sorry, library 'foo' is not available")
    ...

if library foo was not found (with the desired functionality). In this case, the above R code effectively disables the function.

One could also use different file fragments for available and missing functionality, respectively.

You will very likely need to ensure that the same C compiler and compiler flags are used in the configure tests as when compiling R or your package. Under a Unix-alike, you can achieve this by including the following fragment early in configure.ac (before calling AC_PROG_CC or anything which calls it)

: ${R_HOME=`R RHOME`}
if test -z "${R_HOME}"; then
  echo "could not determine R_HOME"
  exit 1
fi
CC=`"${R_HOME}/bin/R" CMD config CC`
CFLAGS=`"${R_HOME}/bin/R" CMD config CFLAGS`
CPPFLAGS=`"${R_HOME}/bin/R" CMD config CPPFLAGS`

(Using ‘${R_HOME}/bin/R’ rather than just ‘R’ is necessary in order to use the correct version of R when running the script as part of R CMD INSTALL, and the quotes since ‘${R_HOME}’ might contain spaces.)

If your code does load checks (for example, to check for an entry point in a library or to run code) then you will also need

LDFLAGS=`"${R_HOME}/bin/R" CMD config LDFLAGS`

Packages written with C++ need to pick up the details for the C++ compiler and switch the current language to C++ by something like

CXX=`"${R_HOME}/bin/R" CMD config CXX`
if test -z "$CXX"; then
  AC_MSG_ERROR([No C++ compiler is available])
fi
CXXFLAGS=`"${R_HOME}/bin/R" CMD config CXXFLAGS`
CPPFLAGS=`"${R_HOME}/bin/R" CMD config CPPFLAGS`
AC_LANG(C++)

The latter is important, as for example C headers may not be available to C++ programs or may not be written to avoid C++ name-mangling. Note that an R installation is not required to have a C++ compiler so ‘CXX’ may be empty. If the package specifies a non-default C++ standard, use the config variable names (such as CXX17) appropriate to the standard, but still set CXX and CXXFLAGS.

You can use R CMD config to get the value of the basic configuration variables, and also the header and library flags necessary for linking a front-end executable program against R, see R CMD config --help for details. If you do, it is essential that you use both the command and the appropriate flags, so that for example ‘CC’ must always be used with ‘CFLAGS’ and (for code to be linked into a shared library) ‘CPICFLAGS’. For Fortran, be careful to use ‘FC FFLAGS FPICFLAGS’ for fixed-form Fortran and ‘FC FCFLAGS FPICFLAGS’ for free-form Fortran.

As from R 4.3.0, variables

CC CFLAGS CXX CXXFLAGS CPPFLAGS LDFLAGS FC FCFLAGS

are set in the environment (if not already set) when configure is called from R CMD INSTALL, in case the script forgets to set them as described above. This includes making use of the selected C standard (but not the C++ standard as that is selected at a later stage by R CMD SHLIB).

To check for an external BLAS library using the AX_BLAS macro from the official Autoconf Macro Archive30, one can use

FC=`"${R_HOME}/bin/R" CMD config FC`
FCLAGS=`"${R_HOME}/bin/R" CMD config FFLAGS`
AC_PROG_FC
FLIBS=`"${R_HOME}/bin/R" CMD config FLIBS`
AX_BLAS([], AC_MSG_ERROR([could not find your BLAS library], 1))

Note that FLIBS as determined by R must be used to ensure that Fortran code works on all R platforms.

N.B.: If the configure script creates files, e.g. src/Makevars, you do need a cleanup script to remove them. Otherwise R CMD build may ship the files that are created. For example, package RODBC has

#!/bin/sh

rm -f config.* src/Makevars src/config.h

As this example shows, configure often creates working files such as config.log. If you use a hand-crafted script rather than one created by autoconf, it is highly recommended that you log its actions to file config.log.

If your configure script needs auxiliary files, it is recommended that you ship them in a tools directory (as R itself does).

You should bear in mind that the configure script will not be used on Windows systems. If your package is to be made publicly available, please give enough information for a user on a non-Unix-alike platform to configure it manually, or provide a configure.win script (or configure.ucrt) to be used on that platform. (Optionally, there can be a cleanup.win script (or cleanup.ucrt). Both should be shell scripts to be executed by ash, which is a minimal version of Bourne-style sh. As from R 4.2.0, bash is used. When configure.win (or configure.ucrt) is run the environment variables R_HOME (which uses ‘/’ as the file separator), R_ARCH and R_ARCH_BIN will be set. Use R_ARCH to decide if this is a 64-bit build for Intel (its value there is ‘/x64’) and to install DLLs to the correct place (${R_HOME}/libs${R_ARCH}). Use R_ARCH_BIN to find the correct place under the bin directory, e.g. ${R_HOME}/bin${R_ARCH_BIN}/Rscript.exe. If a configure.win script does compilation (including calling R CMD SHLIB), most of the considerations above apply.

As the scripts on Windows are executed as sh ./configure.win and similar, any ’shebang’ first line (such as #! /bin/bash) is treated as a comment.

In some rare circumstances, the configuration and cleanup scripts need to know the location into which the package is being installed. An example of this is a package that uses C code and creates two shared object/DLLs. Usually, the object that is dynamically loaded by R is linked against the second, dependent, object. On some systems, we can add the location of this dependent object to the object that is dynamically loaded by R. This means that each user does not have to set the value of the LD_LIBRARY_PATH (or equivalent) environment variable, but that the secondary object is automatically resolved. Another example is when a package installs support files that are required at run time, and their location is substituted into an R data structure at installation time. The names of the top-level library directory (i.e., specifiable via the ‘-l’ argument) and the directory of the package itself are made available to the installation scripts via the two shell/environment variables R_LIBRARY_DIR and R_PACKAGE_DIR. Additionally, the name of the package (e.g. ‘survival’ or ‘MASS’) being installed is available from the environment variable R_PACKAGE_NAME. (Currently the value of R_PACKAGE_DIR is always ${R_LIBRARY_DIR}/${R_PACKAGE_NAME}, but this used not to be the case when versioned installs were allowed. Its main use is in configure.win (or configure.ucrt) scripts for the installation path of external software’s DLLs.) Note that the value of R_PACKAGE_DIR may contain spaces and other shell-unfriendly characters, and so should be quoted in makefiles and configure scripts.

One of the more tricky tasks can be to find the headers and libraries of external software. One tool which is increasingly available on Unix-alikes (but not by default31 on macOS) to do this is pkg-config. The configure script will need to test for the presence of the command itself32 (see for example package tiff), and if present it can be asked if the software is installed, of a suitable version and for compilation/linking flags by e.g.

$ pkg-config --exists 'libtiff-4 >= 4.1.0' --print-errors # check the status
$ pkg-config --modversion libtiff-4
4.3.0
$ pkg-config --cflags libtiff-4
-I/usr/local/include
$ pkg-config --libs libtiff-4
-L/usr/local/lib -ltiff
$ pkg-config --static --libs libtiff-4
-L/usr/local/lib -ltiff -lwebp -llzma -ljpeg -lz

Note that pkg-config --libs gives the information required to link against the default version33 of that library (usually the dynamic one), and pkg-config --static --libs may be needed if the static library is to be used.

Static libraries are commonly used on macOS and Windows to facilitate bundling external software with binary distributions of packages. This means that portable (source) packages need to allow for this. It is not safe to just use pkg-config --static --libs, as that will often include further libraries that are not necessarily installed on the user’s system (or maybe only the versioned library such as libjbig.so.2.1 is installed and not libjbig.so which would be needed to use -ljbig sometimes included in pkg-config --static --libs libtiff-4).

Another issue is that pkg-config --exists may not be reliable. It checks not only that the ‘module’ is available but all of the dependencies, including those in principle needed for static linking. (XQuartz 2.8.x only distributed dynamic libraries and not some of the .pc files needed for --exists.)

Sometimes the name by which the software is known to pkg-config is not what one might expect (e.g. ‘libxml-2.0’ even for 2.9.x). To get a complete list use

pkg-config --list-all | sort

Some external software provides a -config command to do a similar job to pkg-config, including

curl-config   freetype-config  gdal-config    geos-config
gsl-config    iodbc-config     libpng-config  nc-config
pcre-config   pcre2-config     xml2-config    xslt-config

(curl-config is for libcurl not curl. nc-config is for netcdf.) Most have an option to use static libraries.

N.B. These commands indicate what header paths and libraries are needed, but they do not obviate the need to check that the recipes they give actually work. (This is especially necessary for platforms which use static linking.)

If using Autoconf it is good practice to include all the Autoconf sources in the package (and required for an Open Source package and tested by R CMD check --as-cran). This will include the file configure.ac34 in the top-level directory of the package. If extensions written in m4 are needed, these should be included under the directory tools and included from configure.ac via e.g.,

m4_include([tools/ax_pthread.m4])

Alternatively, Autoconf can be asked to search all .m4 files in a directory by including something like35

AC_CONFIG_MACRO_DIR([tools/m4])

One source of such extensions is the ‘Autoconf Archive’ (https://www.gnu.org/software/autoconf-archive/. It is not safe to assume this is installed on users’ machines, so the extension should be shipped with the package (taking care to comply with its licence).


1.2.1 Using Makevars

Sometimes writing your own configure script can be avoided by supplying a file Makevars: also one of the most common uses of a configure script is to make Makevars from Makevars.in.

A Makevars file is a makefile and is used as one of several makefiles by R CMD SHLIB (which is called by R CMD INSTALL to compile code in the src directory). It should be written if at all possible in a portable style, in particular (except for Makevars.win and Makevars.ucrt) without the use of GNU extensions.

The most common use of a Makevars file is to set additional preprocessor options (for example include paths and definitions) for C/C++ files via PKG_CPPFLAGS, and additional compiler flags by setting PKG_CFLAGS, PKG_CXXFLAGS or PKG_FFLAGS, for C, C++ or Fortran respectively (see Creating shared objects).

N.B.: Include paths are preprocessor options, not compiler options, and must be set in PKG_CPPFLAGS as otherwise platform-specific paths (e.g. ‘-I/usr/local/include’) will take precedence. PKG_CPPFLAGS should contain ‘-I’, ‘-D’, ‘-U’ and (where supported) ‘-include’ and ‘-pthread’ options: everything else should be a compiler flag. The order of flags matters, and using ‘-I’ in PKG_CFLAGS or PKG_CXXFLAGS has led to hard-to-debug platform-specific errors.

Makevars can also be used to set flags for the linker, for example ‘-L’ and ‘-l’ options, via PKG_LIBS.

When writing a Makevars file for a package you intend to distribute, take care to ensure that it is not specific to your compiler: flags such as -O2 -Wall -pedantic (and all other -W flags: for the Oracle compilers these were used to pass arguments to compiler phases) are all specific to GCC (and compilers such as clang which aim to be options-compatible with it).

Also, do not set variables such as CPPFLAGS, CFLAGS etc.: these should be settable by users (sites) through appropriate personal (site-wide) Makevars files. See Customizing package compilation in R Installation and Administration for more information.

There are some macros36 which are set whilst configuring the building of R itself and are stored in R_HOME/etcR_ARCH/Makeconf. That makefile is included as a Makefile after Makevars[.win], and the macros it defines can be used in macro assignments and make command lines in the latter. These include

FLIBS

A macro containing the set of libraries need to link Fortran code. This may need to be included in PKG_LIBS: it will normally be included automatically if the package contains Fortran source files in the src directory.

BLAS_LIBS

A macro containing the BLAS libraries used when building R. This may need to be included in PKG_LIBS. Beware that if it is empty then the R executable will contain all the double-precision and double-complex BLAS routines, but no single-precision nor complex routines. If BLAS_LIBS is included, then FLIBS also needs to be37 included following it, as most BLAS libraries are written at least partially in Fortran. However, it can be omitted if the package contains Fortran source code as that will add FLIBS to the link line.

LAPACK_LIBS

A macro containing the LAPACK libraries (and paths where appropriate) used when building R. This may need to be included in PKG_LIBS. It may point to a dynamic library libRlapack which contains the main double-precision LAPACK routines as well as those double-complex LAPACK routines needed to build R, or it may point to an external LAPACK library, or may be empty if an external BLAS library also contains LAPACK.

[libRlapack includes all the double-precision LAPACK routines which were current in 2003 and a few more recent ones: a list of which routines are included is in file src/modules/lapack/README. Note that an external LAPACK/BLAS library need not do so, as some were ‘deprecated’ (and not compiled by default) in LAPACK 3.6.0 in late 2015.]

For portability, the macros BLAS_LIBS and FLIBS should always be included after LAPACK_LIBS (and in that order).

SAFE_FFLAGS

A macro containing flags which are needed to circumvent over-optimization of FORTRAN code: it is might be ‘-g -O2 -ffloat-store’ or ‘-g -O2 -msse2 -mfpmath=sse’ on ‘ix86’ platforms using gfortran. Note that this is not an additional flag to be used as part of PKG_FFLAGS, but a replacement for FFLAGS. See the example later in this section.

Setting certain macros in Makevars will prevent R CMD SHLIB setting them: in particular if Makevars sets ‘OBJECTS’ it will not be set on the make command line. This can be useful in conjunction with implicit rules to allow other types of source code to be compiled and included in the shared object. It can also be used to control the set of files which are compiled, either by excluding some files in src or including some files in subdirectories. For example

OBJECTS = 4dfp/endianio.o 4dfp/Getifh.o R4dfp-object.o

Note that Makevars should not normally contain targets, as it is included before the default makefile and make will call the first target, intended to be all in the default makefile. If you really need to circumvent that, use a suitable (phony) target all before any actual targets in Makevars.[win]: for example package fastICA used to have

PKG_LIBS = @BLAS_LIBS@

SLAMC_FFLAGS=$(R_XTRA_FFLAGS) $(FPICFLAGS) $(SHLIB_FFLAGS) $(SAFE_FFLAGS)

all: $(SHLIB)

slamc.o: slamc.f
        $(FC) $(SLAMC_FFLAGS) -c -o slamc.o slamc.f

needed to ensure that the LAPACK routines find some constants without infinite looping. The Windows equivalent was

all: $(SHLIB)

slamc.o: slamc.f
        $(FC) $(SAFE_FFLAGS) -c -o slamc.o slamc.f

(since the other macros are all empty on that platform, and R’s internal BLAS was not used). Note that the first target in Makevars will be called, but for back-compatibility it is best named all.

If you want to create and then link to a library, say using code in a subdirectory, use something like

.PHONY: all mylibs

all: $(SHLIB)
$(SHLIB): mylibs

mylibs:
        (cd subdir; $(MAKE))

Be careful to create all the necessary dependencies, as there is no guarantee that the dependencies of all will be run in a particular order (and some of the CRAN build machines use multiple CPUs and parallel makes). In particular,

all: mylibs

does not suffice. GNU make does allow the construct

.NOTPARALLEL: all
all: mylibs $(SHLIB)

but that is not portable. dmake and pmake allow the similar .NO_PARALLEL, also not portable: some variants of pmake accept .NOTPARALLEL as an alias for .NO_PARALLEL.

Note that on Windows it is required that Makevars[.win, .ucrt] does create a DLL: this is needed as it is the only reliable way to ensure that building a DLL succeeded. If you want to use the src directory for some purpose other than building a DLL, use a Makefile.win or Makefile.ucrt file.

It is sometimes useful to have a target ‘clean’ in Makevars, Makevars.win or Makevars.ucrt: this will be used by R CMD build to clean up (a copy of) the package sources. When it is run by build it will have fewer macros set, in particular not $(SHLIB), nor $(OBJECTS) unless set in the file itself. It would also be possible to add tasks to the target ‘shlib-clean’ which is run by R CMD INSTALL and R CMD SHLIB with options --clean and --preclean.

Avoid the use of default (also known as ‘implicit’ rules) in makefiles, as these are make-specific. Even when mandated by POSIX – GNU make does not comply and this has broken package installation.

An unfortunately common error is to have

all: $(SHLIB) clean

which asks make to clean in parallel with compiling the code. Not only does this lead to hard-to-debug installation errors, it wipes out all the evidence of any error (from a parallel make or not). It is much better to leave cleaning to the end user using the facilities in the previous paragraph.

If you want to run R code in Makevars, e.g. to find configuration information, please do ensure that you use the correct copy of R or Rscript: there might not be one in the path at all, or it might be the wrong version or architecture. The correct way to do this is via

"$(R_HOME)/bin$(R_ARCH_BIN)/Rscript" filename
"$(R_HOME)/bin$(R_ARCH_BIN)/Rscript" -e 'R expression'

where $(R_ARCH_BIN) is only needed currently on Windows.

Environment or make variables can be used to select different macros for Intel 64-bit code or code for other architectures, for example (GNU make syntax, allowed on Windows)

ifeq "$(WIN)" "64"
PKG_LIBS = value for 64-bit Intel Windows
else
PKG_LIBS = value for unknown Windows architectures
endif

On Windows there is normally a choice between linking to an import library or directly to a DLL. Where possible, the latter is much more reliable: import libraries are tied to a specific toolchain, and in particular on 64-bit Windows two different conventions have been commonly used. So for example instead of

PKG_LIBS = -L$(XML_DIR)/lib -lxml2

one can use

PKG_LIBS = -L$(XML_DIR)/bin -lxml2

since on Windows -lxxx will look in turn for

libxxx.dll.a
xxx.dll.a
libxxx.a
xxx.lib
libxxx.dll
xxx.dll

where the first and second are conventionally import libraries, the third and fourth often static libraries (with .lib intended for Visual C++), but might be import libraries. See for example https://sourceware.org/binutils/docs-2.20/ld/WIN32.html#WIN32.

The fly in the ointment is that the DLL might not be named libxxx.dll, and in fact on 32-bit Windows there was a libxml2.dll whereas on one build for 64-bit Windows the DLL is called libxml2-2.dll. Using import libraries can cover over these differences but can cause equal difficulties.

If static libraries are available they can save a lot of problems with run-time finding of DLLs, especially when binary packages are to be distributed and even more when these support both architectures. Where using DLLs is unavoidable we normally arrange (via configure.win or configure.ucrt) to ship them in the same directory as the package DLL.


1.2.1.1 OpenMP support

There is some support for packages which wish to use OpenMP38. The make macros

SHLIB_OPENMP_CFLAGS
SHLIB_OPENMP_CXXFLAGS
SHLIB_OPENMP_FFLAGS

are available for use in src/Makevars, src/Makevars.win or Makevars.ucrt. Include the appropriate macro in PKG_CFLAGS, PKG_CXXFLAGS and so on, and also in PKG_LIBS (but see below for Fortran). C/C++ code that needs to be conditioned on the use of OpenMP can be used inside #ifdef _OPENMP: note that some toolchains used for R (including Apple’s for macOS39 and some others using clang40) have no OpenMP support at all, not even omp.h.

For example, a package with C code written for OpenMP should have in src/Makevars the lines

PKG_CFLAGS = $(SHLIB_OPENMP_CFLAGS)
PKG_LIBS = $(SHLIB_OPENMP_CFLAGS)

Note that the macro SHLIB_OPENMP_CXXFLAGS applies to the default C++ compiler and not necessarily to the C++17/20/23 compiler: users of the latter should do their own configure checks. If you do use your own checks, make sure that OpenMP support is complete by compiling and linking an OpenMP-using program: on some platforms the runtime library is optional and on others that library depends on other optional libraries.

Some care is needed when compilers are from different families which may use different OpenMP runtimes (e.g. clang vs GCC including gfortran, although it is often possible to use the clang runtime with GCC but not vice versa: however gfortran >= 9 may generate calls not in the clang runtime). For a package with Fortran code using OpenMP the appropriate lines are

PKG_FFLAGS = $(SHLIB_OPENMP_FFLAGS)
PKG_LIBS = $(SHLIB_OPENMP_CFLAGS)

as the C compiler will be used to link the package code. There are platforms on which this does not work for some OpenMP-using code and installation will fail. Since R >= 3.6.2 the best alternative for a package with only Fortran sources using OpenMP is to use

USE_FC_TO_LINK =
PKG_FFLAGS = $(SHLIB_OPENMP_FFLAGS)
PKG_LIBS = $(SHLIB_OPENMP_FFLAGS)

in src/Makevars, src/Makevars.win or Makevars.ucrt. Note however, that when this is used $(FLIBS) should not be included in PKG_LIBS since it is for linking Fortran-compiled code by the C compiler.

Common platforms may inline all OpenMP calls and so tolerate the omission of the OpenMP flag from PKG_LIBS, but this usually results in an installation failure with a different compiler or compilation flags. So cross-check that e.g. -fopenmp appears in the linking line in the installation logs.

It is not portable to use OpenMP with more than one of C, C++ and Fortran in a single package since it is not uncommon that the compilers are of different families.

For portability, any C/C++ code using the omp_* functions should include the omp.h header: some compilers (but not all) include it when OpenMP mode is switched on (e.g. via flag -fopenmp).

There is nothing41 to say what version of OpenMP is supported: version 4.0 (and much of 4.5 or 5.0) is supported by recent versions of the Linux and Windows platforms, but portable packages cannot assume that end users have recent versions. Apple clang on macOS has no OpenMP support. https://www.openmp.org/resources/openmp-compilers-tools/ gives some idea of what compilers support what versions. Note that support for Fortran compilers is often less up-to-date and that page suggests it is unwise to rely on a version later than 3.1. Which introduced a Fortran OpenMP module, so Fortran users of OpenMP should include

use omp_lib

Rarely, using OpenMP with clang on Linux generates calls in libatomic, resulting in loading messages like

  undefined symbol: __atomic_compare_exchange
  undefined symbol: __atomic_load

The workaround is to link with -latomic (having checked it exists).

The performance of OpenMP varies substantially between platforms. The Windows implementation has substantial overheads, so is only beneficial if quite substantial tasks are run in parallel. Also, on Windows new threads are started with the default42 FPU control word, so computations done on OpenMP threads will not make use of extended-precision arithmetic which is the default for the main process.

Do not include these macros unless your code does make use of OpenMP (possibly for C++ via included external headers): this can result in the OpenMP runtime being linked in, threads being started, ….

Calling any of the R API from threaded code is ‘for experts only’ and strongly discouraged. Many functions in the R API modify internal R data structures and might corrupt these data structures if called simultaneously from multiple threads. Most R API functions can signal errors, which must only happen on the R main thread. Also, external libraries (e.g. LAPACK) may not be thread-safe.

Packages are not standard-alone programs, and an R process could contain more than one OpenMP-enabled package as well as other components (for example, an optimized BLAS) making use of OpenMP. So careful consideration needs to be given to resource usage. OpenMP works with parallel regions, and for most implementations the default is to use as many threads as ‘CPUs’ for such regions. Parallel regions can be nested, although it is common to use only a single thread below the first level. The correctness of the detected number of ‘CPUs’ and the assumption that the R process is entitled to use them all are both dubious assumptions. One way to limit resources is to limit the overall number of threads available to OpenMP in the R process: this can be done via environment variable OMP_THREAD_LIMIT, where implemented.43 Alternatively, the number of threads per region can be limited by the environment variable OMP_NUM_THREADS or API call omp_set_num_threads, or, better, for the regions in your code as part of their specification. E.g. R uses44

#pragma omp parallel for num_threads(nthreads) ...

That way you only control your own code and not that of other OpenMP users.

Note that setting environment variables to control OpenMP is implementation-dependent and may need to be done outside the R process or before any use of OpenMP (which might be by another process or R itself). Also, implementation-specific variables such as KMP_THREAD_LIMIT might take precedence.


1.2.1.2 Using pthreads

There is no direct support for the POSIX threads (more commonly known as pthreads): by the time we considered adding it several packages were using it unconditionally so it seems that nowadays it is universally available on POSIX operating systems.

For reasonably recent versions of gcc and clang the correct specification is

PKG_CPPFLAGS = -pthread
PKG_LIBS = -pthread

(and the plural version is also accepted on some systems/versions). For other platforms the specification is

PKG_CPPFLAGS = -D_REENTRANT
PKG_LIBS = -lpthread

(and note that the library name is singular). This is what -pthread does on all known current platforms (although earlier versions of OpenBSD used a different library name).

For a tutorial see https://hpc-tutorials.llnl.gov/posix/.

POSIX threads are not normally used on Windows which has its own native concepts of threads: however, recent toolchains do provide the pthreads header and library.

The presence of a working pthreads implementation cannot be unambiguously determined without testing for yourself: however, that ‘_REENTRANT’ is defined in C/C++ code is a good indication.

Note that not all pthreads implementations are equivalent as parts are optional (see https://pubs.opengroup.org/onlinepubs/009695399/basedefs/pthread.h.html): for example, macOS lacks the ‘Barriers’ option.

See also the comments on thread-safety and performance under OpenMP: on all known R platforms OpenMP is implemented via pthreads and the known performance issues are in the latter.


1.2.1.3 Compiling in sub-directories

Package authors fairly often want to organize code in sub-directories of src, for example if they are including a separate piece of external software to which this is an R interface.

One simple way is simply to set OBJECTS to be all the objects that need to be compiled, including in sub-directories. For example, CRAN package RSiena has

SOURCES = $(wildcard data/*.cpp network/*.cpp utils/*.cpp model/*.cpp model/*/*.cpp model/*/*/*.cpp)

OBJECTS = siena07utilities.o siena07internals.o siena07setup.o siena07models.o $(SOURCES:.cpp=.o)

One problem with that approach is that unless GNU make extensions are used, the source files need to be listed and kept up-to-date. As in the following from CRAN package lossDev:

OBJECTS.samplers = samplers/ExpandableArray.o samplers/Knots.o \
  samplers/RJumpSpline.o samplers/RJumpSplineFactory.o \
  samplers/RealSlicerOV.o samplers/SliceFactoryOV.o samplers/MNorm.o
OBJECTS.distributions = distributions/DSpline.o \
  distributions/DChisqrOV.o distributions/DTOV.o \
  distributions/DNormOV.o distributions/DUnifOV.o distributions/RScalarDist.o
OBJECTS.root = RJump.o

OBJECTS = $(OBJECTS.samplers) $(OBJECTS.distributions) $(OBJECTS.root)

Where the subdirectory is self-contained code with a suitable makefile, the best approach is something like

PKG_LIBS = -LCsdp/lib -lsdp $(LAPACK_LIBS) $(BLAS_LIBS) $(FLIBS)

$(SHLIB): Csdp/lib/libsdp.a

Csdp/lib/libsdp.a:
        @(cd Csdp/lib && $(MAKE) libsdp.a \
          CC="$(CC)" CFLAGS="$(CFLAGS) $(CPICFLAGS)" AR="$(AR)" RANLIB="$(RANLIB)")

Note the quotes: the macros can contain spaces, e.g. CC = "gcc -m64 -std=gnu99". Several authors have forgotten about parallel makes: the static library in the subdirectory must be made before the shared object ($(SHLIB)) and so the latter must depend on the former. Others forget the need45 for position-independent code.

We really do not recommend using src/Makefile instead of src/Makevars, and as the example above shows, it is not necessary.


1.2.2 Configure example

It may be helpful to give an extended example of using a configure script to create a src/Makevars file: this is based on that in the RODBC package.

The configure.ac file follows: configure is created from this by running autoconf in the top-level package directory (containing configure.ac).

AC_INIT([RODBC], 1.1.8) dnl package name, version

dnl A user-specifiable option
odbc_mgr=""
AC_ARG_WITH([odbc-manager],
            AC_HELP_STRING([--with-odbc-manager=MGR],
                           [specify the ODBC manager, e.g. odbc or iodbc]),
            [odbc_mgr=$withval])

if test "$odbc_mgr" = "odbc" ; then
  AC_PATH_PROGS(ODBC_CONFIG, odbc_config)
fi

dnl Select an optional include path, from a configure option
dnl or from an environment variable.
AC_ARG_WITH([odbc-include],
            AC_HELP_STRING([--with-odbc-include=INCLUDE_PATH],
                           [the location of ODBC header files]),
            [odbc_include_path=$withval])
RODBC_CPPFLAGS="-I."
if test [ -n "$odbc_include_path" ] ; then
   RODBC_CPPFLAGS="-I. -I${odbc_include_path}"
else
  if test [ -n "${ODBC_INCLUDE}" ] ; then
     RODBC_CPPFLAGS="-I. -I${ODBC_INCLUDE}"
  fi
fi

dnl ditto for a library path
AC_ARG_WITH([odbc-lib],
            AC_HELP_STRING([--with-odbc-lib=LIB_PATH],
                           [the location of ODBC libraries]),
            [odbc_lib_path=$withval])
if test [ -n "$odbc_lib_path" ] ; then
   LIBS="-L$odbc_lib_path ${LIBS}"
else
  if test [ -n "${ODBC_LIBS}" ] ; then
     LIBS="-L${ODBC_LIBS} ${LIBS}"
  else
    if test -n "${ODBC_CONFIG}"; then
      odbc_lib_path=`odbc_config --libs | sed s/-lodbc//`
      LIBS="${odbc_lib_path} ${LIBS}"
    fi
  fi
fi

dnl Now find the compiler and compiler flags to use
: ${R_HOME=`R RHOME`}
if test -z "${R_HOME}"; then
  echo "could not determine R_HOME"
  exit 1
fi
CC=`"${R_HOME}/bin/R" CMD config CC`
CFLAGS=`"${R_HOME}/bin/R" CMD config CFLAGS`
CPPFLAGS=`"${R_HOME}/bin/R" CMD config CPPFLAGS`

if test -n "${ODBC_CONFIG}"; then
  RODBC_CPPFLAGS=`odbc_config --cflags`
fi
CPPFLAGS="${CPPFLAGS} ${RODBC_CPPFLAGS}"

dnl Check the headers can be found
AC_CHECK_HEADERS(sql.h sqlext.h)
if test "${ac_cv_header_sql_h}" = no ||
   test "${ac_cv_header_sqlext_h}" = no; then
   AC_MSG_ERROR("ODBC headers sql.h and sqlext.h not found")
fi

dnl search for a library containing an ODBC function
if test [ -n "${odbc_mgr}" ] ; then
  AC_SEARCH_LIBS(SQLTables, ${odbc_mgr}, ,
      AC_MSG_ERROR("ODBC driver manager ${odbc_mgr} not found"))
else
  AC_SEARCH_LIBS(SQLTables, odbc odbc32 iodbc, ,
      AC_MSG_ERROR("no ODBC driver manager found"))
fi

dnl for 64-bit ODBC need SQL[U]LEN, and it is unclear where they are defined.
AC_CHECK_TYPES([SQLLEN, SQLULEN], , , [# include <sql.h>])
dnl for unixODBC header
AC_CHECK_SIZEOF(long, 4)

dnl substitute RODBC_CPPFLAGS and LIBS
AC_SUBST(RODBC_CPPFLAGS)
AC_SUBST(LIBS)
AC_CONFIG_HEADERS([src/config.h])
dnl and do substitution in the src/Makevars.in and src/config.h
AC_CONFIG_FILES([src/Makevars])
AC_OUTPUT

where src/Makevars.in would be simply

PKG_CPPFLAGS = @RODBC_CPPFLAGS@
PKG_LIBS = @LIBS@

A user can then be advised to specify the location of the ODBC driver manager files by options like (lines broken for easier reading)

R CMD INSTALL \
  --configure-args='--with-odbc-include=/opt/local/include \
  --with-odbc-lib=/opt/local/lib --with-odbc-manager=iodbc' \
  RODBC

or by setting the environment variables ODBC_INCLUDE and ODBC_LIBS.


1.2.3 Using modern Fortran code

R assumes that source files with extension .f are fixed-form Fortran 90 (which includes Fortran 77), and passes them to the compiler specified by macro ‘FC’. The Fortran compiler will also accept free-form Fortran 90/95 code with extension .f90 or (most46) .f95.

The same compiler is used for both fixed-form and free-form Fortran code (with different file extensions and possibly different flags). Macro PKG_FFLAGS can be used for package-specific flags: for the un-encountered case that both are included in a single package and that different flags are needed for the two forms, macro PKG_FCFLAGS is also available for free-form Fortran.

The code used to build R allows a ‘Fortran 90’ compiler to be selected as ‘FC’, so platforms might be encountered which only support Fortran 90. However, Fortran 95 is supported on all known platforms.

Most compilers specified by ‘FC’ will accept most Fortran 2003, 2008 or 2018 code: such code should still use file extension .f90. Most current platforms use gfortran where you might need to include -std=f2003, -std=f2008 or (from version 8) -std=f2018 in PKG_FFLAGS or PKG_FCFLAGS: the default is ‘GNU Fortran’, currently Fortran 2018 (but Fortran 95 prior to gfortran 8) with non-standard extensions. The other compilers in current use (LLVM’s flang-new and Intel’s ifx) default to Fortran 201847.

It is good practice to describe a Fortran version requirement in DESCRIPTION’s ‘SystemRequirements’ field. Note that this is purely for information: the package also needs a configure script to determine the compiler and set appropriate option(s) and test that the features needed from the standard are actually supported.

The Fortran 2023 released in Nov 2023: as usual compiler vendors are introducing support incrementally. For Intel’s ifx see https://www.intel.com/content/www/us/en/developer/articles/technical/fortran-language-and-openmp-features-in-ifx.html#Fortran%20Standards. For LLVM’s flang-new see https://flang.llvm.org/docs/F202X.html. gfortran does not have complete support even for the 2008 and 2018 standards, but the option -std=f2023 is supported from version 14.1.

Modern versions of Fortran support modules, whereby compiling one source file creates a module file which is then included in others. (Module files typically have a .mod extension: they do depend on the compiler used and so should never be included in a package.) This creates a dependence which make will not know about and often causes installation with a parallel make to fail. Thus it is necessary to add explicit dependencies to src/Makevars to tell make the constraints on the order of compilation. For example, if file iface.f90 creates a module ‘iface’ used by files cmi.f90 and dmi.f90 then src/Makevars needs to contain something like

cmi.o dmi.o: iface.o

Some maintainers have found it difficult to find all the module dependencies which leads to hard-to-reproduce installation failures. There are tools available to find these, including the Intel compiler’s flag -gen-dep and makedepf90.

Note that it is not portable (although some platforms do accept it) to define a module of the same name in multiple source files.


1.2.4 Using C++ code

R can be built without a C++ compiler although one is available (but not necessarily installed) on all known R platforms. As from R 4.0.0 a C++ compiler will be selected only if it conforms to the 2011 standard (‘C++11’). A minor update48 (‘C++14’) was published in December 2014 and was used by default as from R 4.1.0 if supported. Further revisions ‘C++17’ (in December 2017) and ‘C++20’ (with many new features in December 2020) have been published since. The next revision, ‘C++23’, is expected in 2024 and several compilers already have extensive partial support for the final draft.

The default standard for compiling R packages was changed to C++17 in R 4.3.0 if supported, and from R 4.4.0 only a C++17 compiler will be selected as the default C++ compiler.

What standard a C++ compiler aims to support can be hard to determine: the value49 of __cplusplus may help but some compilers use it to denote a standard which is partially supported and some the latest standard which is (almost) fully supported. On a Unix-alike configure will try to identify a compiler and flags for each of the standards: this relies heavily on the reported values of __cplusplus.

The webpage https://en.cppreference.com/w/cpp/compiler_support gives some information on which compilers are known to support recent C++ features.

C++ standards have deprecated and later removed features. Be aware that some current compilers still accept removed features in C++17 mode, such as std::unary_function (deprecated in C++11, removed in C++17).

Different versions of R have used different default C++ standards, so for maximal portability a package should specify the standard it requires. In order to specify C++14 code in a package with a Makevars file (or Makevars.win or Makevars.ucrt on Windows) should include the line

CXX_STD = CXX14

Compilation and linking will then be done with the C++14 compiler (if any). Analogously for other standards (details below). On the other hand, specifying C++1150 when the code is valid under C++14 or C++17 reduces future portability.

Packages without a src/Makevars or src/Makefile file may specify a C++ standard for code in the src directory by including something like ‘C++14’ in the ‘SystemRequirements’ field of the DESCRIPTION file, e.g.

SystemRequirements: C++14

If a package does have a src/Makevars[.win] file then also setting the make variable ‘CXX_STD’ there is recommended, as it allows R CMD SHLIB to work correctly in the package’s src directory.

A requirement of C++17 or later should always be declared in the ‘SystemRequirements’ field (as well as in src/Makevars or src/Makefile) so this is shown on the package’s summary pages on CRAN or similar. This is also good practice for a requirement of C++14. Note that support of C++14 or C++17 is only available from R 3.4.0, so if the package has an R version requirement it needs to take that into account.

Essentially complete C++14 support is available from GCC 5, LLVM clang 3.4 and currently-used versions of Apple clang.

Code needing C++14 features can check for their presence via ‘SD-6 feature tests’51. Such a check could be

#include <memory> // header where this is defined
#if defined(__cpp_lib_make_unique) && (__cpp_lib_make_unique >= 201304)
using std::make_unique;
#else
// your emulation
#endif

C++17 (as from R 3.4.0), C++20 (as from R 4.0.0) and C++23 (as from R 4.3.0) can be specified in an analogous way (replacing 14 by 17, 20 or 23) but compiler/OS support is platform-dependent. Some C++17 and C++20 support is available with the default builds of R on macOS and Windows as from R 4.0.0. Much of g++’s support for C++17 needs version 7 or later: that is more recent than some still-current Linux distributions but often packages for later compilers are available: for RHEL/Centos 7 look for ‘devtoolset’.

Note that C++17 or later ‘support’ does not mean complete support: use feature tests as well as resources such as https://en.cppreference.com/w/cpp/compiler_support, https://gcc.gnu.org/projects/cxx-status.html and https://clang.llvm.org/cxx_status.html to see if the features you want to use are widely implemented. In particular, for C++23 R’s configure script only checks for a compiler claiming to be later than C++20.52

Attempts to specify an unknown C++ standard are silently ignored: recent versions of R throw an error for C++98 and for known standards for which no compiler+flags has been detected.

If a package using C++ has a configure script it is essential that the script selects the correct C++ compiler and standard, via something like

CXX17=`"${R_HOME}/bin/R" CMD config CXX17`
if test -z "$CXX17"; then
  AC_MSG_ERROR([No C++17 compiler is available])
fi
CXX17STD=`"${R_HOME}/bin/R" CMD config CXX17STD`
CXX="${CXX17} ${CXX17STD}"
CXXFLAGS=`"${R_HOME}/bin/R" CMD config CXX17FLAGS`
## for an configure.ac file
AC_LANG(C++)

if C++17 was specified, but using

CXX=`"${R_HOME}/bin/R" CMD config CXX`
CXXFLAGS=`"${R_HOME}/bin/R" CMD config CXXFLAGS`
## for an configure.ac file
AC_LANG(C++)

if no standard was specified.

If you want to compile C++ code in a subdirectory, make sure you pass down the macros to specify the appropriate compiler, e.g. in src/Makevars

sublibs:
         @(cd libs && $(MAKE) \
            CXX="$(CXX17) $(CXX17STD)" CXXFLAGS="$(CXX17FLAGS) $(CXX17PICFLAGS)")

The discussion above is about the standard R ways of compiling C++: it will not apply to packages using src/Makefile or building in a subdirectory that do not set the C++ standard. And compilers’ default C++ standards varies widely and gets changed frequently by vendors – for example Apple clang up to at least 16 defaults to C++98, LLVM clang 14–15 to C++14, LLVM clang 16–18 to C++17 and g++ 11–14 to C++17.

For a package with a src/Makefile (or a Windows analogue), a non-default C++ compiler can be selected by including something like

CXX14 = `"${R_HOME}/bin/R" CMD config CXX14`
CXX14STD = `"${R_HOME}/bin/R" CMD config CXX14STD`
CXX = ${CXX14} ${CXX14STD}
CXXFLAGS = `"${R_HOME}/bin/R" CMD config CXX14FLAGS`
CXXPICFLAGS = `"${R_HOME}/bin/R" CMD config CXX14PICFLAGS`
SHLIB_LD = "${R_HOME}/bin/R" CMD config SHLIB_CXX14LD`
SHLIB_LDFLAGS = "${R_HOME}/bin/R" CMD config SHLIB_CXX14LDFLAGS`

and ensuring these values are used in relevant compilations, after checking they are non-empty. A common use of src/Makefile is to compile an executable, when likely something like (for example for C++14)

CXX14 = `"${R_HOME}/bin/R" CMD config CXX14`
CXX14STD = `"${R_HOME}/bin/R" CMD config CXX14STD`
CXX = ${CXX14} ${CXX14STD}
CXXFLAGS = `"${R_HOME}/bin/R" CMD config CXX14FLAGS`

suffices. On Unix (and on Windows from R 4.3.0) this can be simplified to

CXX = ${CXX14} ${CXX14STD}
CXXFLAGS = ${CXX14FLAGS}

On a Unix-alike C++ compilation defaulted to C++11 from R 3.6.0, to C++14 from R 4.1.0 and to C++17 from R 4.3.0. However, only ‘if available’, so platforms using very old OSes might have used the previous default. Even older versions of R defaulted to the compiler’s default, almost certainly C++98 for compilers of comparable vintage.

On Windows the default was changed from C++98 to C++11 in R 3.6.2, to C++14 in R 4.2.3 and to C++17 in R 4.3.0.

The C++11 standard could be specified as from R 3.1.0 and C++14 or C++17 as from R 3.4.0, for C++20 from R 4.0.0 and for C++23 from R 4.3.0 (although they may not be supported by the compilers in use). C++11 support became mandatory in R 4.0.0 and C++17 support in R 4.4.0.

The .so/.dll in a package may need to be linked by the C++ compiler if it or any library it links to contains compiled C++ code. Dynamic linking usually brings in the C++ runtime library (commonly libstdc++ but can be, for example, libc++) but static linking (as used for external libraries on Windows and macOS) will not. R CMD INSTALL will link with the C++ compiler if there are any top-level C++ files in src, but not if these are all in subdirectories. The simplest way to force linking by the C++ compiler is to include an empty C++ file in src..


1.2.5 C standards

C has had standards C89/C90, C99, C11, C17 (also known as C18), and C23 (published in 2024). C11 was a minor change to C99 which introduced some new features and made others optional, and C17 is a ‘bug-fix’ update to C11. On the other hand, C23 makes extensive changes, including making bool, true and false reserved words, finally disallowing K&R-style function declarations and clarifying the formerly deprecated meaning of function declarations with an empty parameter list to mean zero parameters. (There are many other additions: see for example https://en.cppreference.com/w/c/23.)

The configure script in recent versions of R aims to choose a C compiler which supports C11: as the default in recent versions of gcc, LLVM clang and Apple clang is C17, that is what is likely to be chosen. On the other hand, until R 4.3.0 the makefiles for the Windows build specified C99. They now use the compiler default which for the recommended compiler is C17.

Packages may want to either avoid or embrace the changes in C23, and can do so via specifying ‘USE_Cnn’ for 17, 23, 90 or 99 in the ‘SystemRequirements’ field of their DESCRIPTION file of a package depending on ‘R (>= 4.3.0)’. Those using a configure script should set the corresponding compiler and flags, for example using

CC=`"${R_HOME}/bin/R" CMD config CC23`
CFLAGS=`"${R_HOME}/bin/R" CMD config C23FLAGS`
CPPFLAGS=`"${R_HOME}/bin/R" CMD config CPPFLAGS`
LDFLAGS=`"${R_HOME}/bin/R" CMD config LDFLAGS`

The (claimed) C standard in use can be checked by the macro __STDC_VERSION__. This is undefined in C89/C90 and should have values 199901L, 201112L and 201710L for C99, C11 and C17. The definitive value for C23 is 202311L but some compilers53 are currently using 202000L and requiring the standard to be specified as c2x. C23 has macros similar to C++ ‘feature tests’ for many of its changes, for example __STDC_VERSION_LIMITS_H__.

However, note the ‘claimed’ as no compiler had 100% conformance, and it is better to use configure to test for the feature you want to use than to condition on the value of __STDC_VERSION__. In particular, C11 alignment functionality such as _Alignas and aligned_alloc is not implemented on Windows.

End users can specify a standard by something like R CMD INSTALL --use-C17. This overrides the ‘SystemRequirements’ field, but not for any configure file.


1.2.6 Using cmake

Packages often wish to include the sources of other software and compile that for inclusion in their .so or .dll, which is normally done by including (or unpacking) the sources in a subdirectory of src, as considered above.

Further issues arise when the external software uses another build system such as cmake, principally to ensure that all the settings for compilers, include and load paths etc are made. This section has already mentioned the need to set at least some of

CC CFLAGS CXX CXXFLAGS CPPFLAGS LDFLAGS

CFLAGS and CXXFLAGS will need to include CPICFLAGS and CXXPICFLAGS respectively unless (as below) cmake is asked to generate PIC code.

Setting these (and more) as environment variables controls the behaviour of cmake (https://cmake.org/cmake/help/latest/manual/cmake-env-variables.7.html#manual:cmake-env-variables(7)), but it may be desirable to translate these into native settings such as

CMAKE_C_COMPILER
CMAKE_C_FLAGS
CMAKE_CXX_COMPILER
CMAKE_CXX_FLAGS
CMAKE_INCLUDE_PATH
CMAKE_LIBRARY_PATH
CMAKE_SHARED_LINKER_FLAGS_INIT
CMAKE_OSX_DEPLOYMENT_TARGET

and it is often necessary to ensure a static library of PIC code is built by

-DBUILD_SHARED_LIBS:bool=OFF
-DCMAKE_POSITION_INDEPENDENT_CODE:bool=ON

If R is to be detected or used, this must be the build being used for package installation – "${R_HOME}"/bin/R.

To fix ideas, consider a package with sources for a library myLib under src/libs. Two approaches have been used. It is often most convenient to build the external software in a directory other than its sources (particularly during development when the build directory can be removed between builds rather than attempting to clean the sources) – this is illustrated in the first approach.

  1. Use the package’s configure script to create a static library src/build/libmyLib.a. This can then be treated in the same way as external software, for example having in src/Makevars
    PKG_CPPFLAGS = -Ilibs/include
    PKG_LIBS = build/libmyLib.a
    

    (-Lbuild -lmyLib could also be used but this explicit specification avoids any confusion with dynamic libraries of the same name.)

    The configure script will need to contain something like (for C code)

    : ${R_HOME=`R RHOME`}
    if test -z "${R_HOME}"; then
      echo "could not determine R_HOME"
      exit 1
    fi
    CC=`"${R_HOME}/bin/R" CMD config CC`
    CFLAGS=`"${R_HOME}/bin/R" CMD config CFLAGS`
    CPPFLAGS=`"${R_HOME}/bin/R" CMD config CPPFLAGS`
    LDFLAGS=`"${R_HOME}/bin/R" CMD config LDFLAGS`
    
    cd src
    mkdir build && cd build
    cmake -S ../libs \
      -DCMAKE_BUILD_TYPE=Release \
      -DBUILD_SHARED_LIBS:bool=OFF \
      -DCMAKE_POSITION_INDEPENDENT_CODE:bool=ON
    ${MAKE}
    
  2. Use src/Makevars (or src/Makevars.win or Makevars.ucrt) to build within the subdirectory. This could be something like (for C code)
    PKG_CPPFLAGS = -Ilibs/include
    PKG_LIBS = libs/libmyLib.a
    
    $(SHLIB): mylibs
    
    mylibs:
            (cd libs; \
              CC="$(CC)" CFLAGS="$(CFLAGS)" \
              CPPFLAGS="$(CPPFLAGS)" LDFLAGS="$(LDFLAGS)" \
              cmake . \
                -DCMAKE_BUILD_TYPE=Release \
                -DBUILD_SHARED_LIBS:bool=OFF \
                -DCMAKE_POSITION_INDEPENDENT_CODE:bool=ON; \
              $(MAKE))
    
    

    the compiler and other settings having been set as Make variables by an R makefile included by INSTALL before src/Makevars.

A complication is that on macOS cmake (where installed) is commonly not on the path but at /Applications/CMake.app/Contents/bin/cmake. One way to work around this is for the package’s configure script to include

if test -z "$CMAKE"; then CMAKE="`which cmake`"; fi
if test -z "$CMAKE"; then CMAKE=/Applications/CMake.app/Contents/bin/cmake; fi
if test -f "$CMAKE"; then echo "no 'cmake' command found"; exit 1; fi

and for the second approach to substitute CMAKE into src/Makevars. This also applies to the ancillary command ctest, if used.


1.3 Checking and building packages

Before using these tools, please check that your package can be installed. R CMD check will inter alia do this, but you may get more detailed error messages doing the install directly.

If your package specifies an encoding in its DESCRIPTION file, you should run these tools in a locale which makes use of that encoding: they may not work at all or may work incorrectly in other locales (although UTF-8 locales will most likely work).

Note: R CMD check and R CMD build run R processes with --vanilla in which none of the user’s startup files are read. If you need R_LIBS set (to find packages in a non-standard library) you can set it in the environment: also you can use the check and build environment files (as specified by the environment variables R_CHECK_ENVIRON and R_BUILD_ENVIRON; if unset, files54 ~/.R/check.Renviron and ~/.R/build.Renviron are used) to set environment variables when using these utilities.

Note to Windows users: R CMD build may make use of the Windows toolset (see The Windows toolset in R Installation and Administration) if present and in your path, and it is required for packages which need it to install (including those with configure.win, cleanup.win, configure.ucrt or cleanup.ucrt scripts or a src directory) and e.g. need vignettes built.

You may need to set the environment variable TMPDIR to point to a suitable writable directory with a path not containing spaces – use forward slashes for the separators. Also, the directory needs to be on a case-honouring file system (some network-mounted file systems are not).


1.3.1 Checking packages

Using R CMD check, the R package checker, one can test whether source R packages work correctly. It can be run on one or more directories, or compressed package tar archives with extension .tar.gz, .tgz, .tar.bz2 or .tar.xz.

It is strongly recommended that the final checks are run on a tar archive prepared by R CMD build.

This runs a series of checks, including

  1. The package is installed. This will warn about missing cross-references and duplicate aliases in help files.
  2. The file names are checked to be valid across file systems and supported operating system platforms.
  3. The files and directories are checked for sufficient permissions (Unix-alikes only).
  4. The files are checked for binary executables, using a suitable version of file if available55. (There may be rare false positives.)
  5. The DESCRIPTION file is checked for completeness, and some of its entries for correctness. Unless installation tests are skipped, checking is aborted if the package dependencies cannot be resolved at run time. (You may need to set R_LIBS in the environment if dependent packages are in a separate library tree.) One check is that the package name is not that of a standard package, nor one of the defunct standard packages (‘ctest’, ‘eda’, ‘lqs’, ‘mle’, ‘modreg’, ‘mva’, ‘nls’, ‘stepfun’ and ‘ts’). Another check is that all packages mentioned in library or requires or from which the NAMESPACE file imports or are called via :: or ::: are listed (in ‘Depends’, ‘Imports’, ‘Suggests’): this is not an exhaustive check of the actual imports.
  6. Available index information (in particular, for demos and vignettes) is checked for completeness.
  7. The package subdirectories are checked for suitable file names and for not being empty. The checks on file names are controlled by the option --check-subdirs=value. This defaults to ‘default’, which runs the checks only if checking a tarball: the default can be overridden by specifying the value as ‘yes’ or ‘no’. Further, the check on the src directory is only run if the package does not contain a configure script (which corresponds to the value ‘yes-maybe’) and there is no src/Makefile or src/Makefile.in.

    To allow a configure script to generate suitable files, files ending in ‘.in’ will be allowed in the R directory.

    A warning is given for directory names that look like R package check directories – many packages have been submitted to CRAN containing these.

  8. The R files are checked for syntax errors. Bytes which are non-ASCII are reported as warnings, but these should be regarded as errors unless it is known that the package will always be used in the same locale.
  9. It is checked that the package can be loaded, first with the usual default packages and then only with package base already loaded. It is checked that the namespace can be loaded in an empty session with only the base namespace loaded. (Namespaces and packages can be loaded very early in the session, before the default packages are available, so packages should work then.)
  10. The R files are checked for correct calls to library.dynam. Package startup functions are checked for correct argument lists and (incorrect) calls to functions which modify the search path or inappropriately generate messages. The R code is checked for possible problems using codetools. In addition, it is checked whether S3 methods have all the arguments of the corresponding generic, and whether the final argument of replacement functions is called ‘value’. All foreign function calls (.C, .Fortran, .Call and .External calls) are tested to see if they have a PACKAGE argument, and if not, whether the appropriate DLL might be deduced from the namespace of the package. Any other calls are reported. (The check is generous, and users may want to supplement this by examining the output of tools::checkFF("mypkg", verbose=TRUE), especially if the intention were to always use a PACKAGE argument)
  11. The Rd files are checked for correct syntax and metadata, including the presence of the mandatory fields (\name, \alias, \title and \description). The Rd name and title are checked for being non-empty, and there is a check for missing cross-references (links).
  12. A check is made for missing documentation entries, such as undocumented user-level objects in the package.
  13. Documentation for functions, data sets, and S4 classes is checked for consistency with the corresponding code.
  14. It is checked whether all function arguments given in \usage sections of Rd files are documented in the corresponding \arguments section.
  15. The data directory is checked for non-ASCII characters and for the use of reasonable levels of compression.
  16. C, C++ and Fortran source and header files56 are tested for portable (LF-only) line endings. If there is a Makefile or Makefile.in or Makevars or Makevars.in file under the src directory, it is checked for portable line endings and the correct use of ‘$(BLAS_LIBS)’ and ‘$(LAPACK_LIBS)

    Compiled code is checked for symbols corresponding to functions which might terminate R or write to stdout/stderr instead of the console. Note that the latter might give false positives in that the symbols might be pulled in with external libraries and could never be called. Windows57 users should note that the Fortran and C++ runtime libraries are examples of such external libraries.

  17. Some checks are made of the contents of the inst/doc directory. These always include checking for files that look like leftovers, and if suitable tools (such as qpdf) are available, checking that the PDF documentation is of minimal size.
  18. The examples provided by the package’s documentation are run. (see Writing R documentation files, for information on using \examples to create executable example code.) If there is a file tests/Examples/pkg-Ex.Rout.save, the output of running the examples is compared to that file.

    Of course, released packages should be able to run at least their own examples. Each example is run in a ‘clean’ environment (so earlier examples cannot be assumed to have been run), and with the variables T and F redefined to generate an error unless they are set in the example: See Logical vectors in An Introduction to R.

  19. If the package sources contain a tests directory then the tests specified in that directory are run. (Typically they will consist of a set of .R source files and target output files .Rout.save.) Please note that the comparison will be done in the end user’s locale, so the target output files should be ASCII if at all possible. (The command line option --test-dir=foo may be used to specify tests in a non-standard location. For example, unusually slow tests could be placed in inst/slowTests and then R CMD check --test-dir=inst/slowTests would be used to run them. Other names that have been suggested are, for example, inst/testWithOracle for tests that require Oracle to be installed, inst/randomTests for tests which use random values and may occasionally fail by chance, etc.)
  20. The R code in package vignettes (see Writing package vignettes) is executed, and the vignettes re-made from their sources as a check of completeness of the sources (unless there is a ‘BuildVignettes’ field in the package’s DESCRIPTION file with a false value). If there is a target output file .Rout.save in the vignette source directory, the output from running the code in that vignette is compared with the target output file and any differences are reported (but not recorded in the log file). (If the vignette sources are in the deprecated location inst/doc, do mark such target output files to not be installed in .Rinstignore.)

    If there is an error58 in executing the R code in vignette foo.ext, a log file foo.ext.log is created in the check directory. The vignettes are re-made in a copy of the package sources in the vign_test subdirectory of the check directory, so for further information on errors look in directory pkgname/vign_test/vignettes. (It is only retained if there are errors or if environment variable _R_CHECK_CLEAN_VIGN_TEST_ is set to a false value.)

  21. The PDF version of the package’s manual is created (to check that the Rd files can be converted successfully). This needs LaTeX and suitable fonts and LaTeX packages to be installed. See Making the manuals in R Installation and Administration for further details.
  22. Optionally (including by R CMD check --as-cran) the HTML version of the manual is created and checked for compliance with the HTML5 standard. This requires a recent version59 of ‘HTML Tidy’, either on the path or at a location specified by environment variable R_TIDYCMD. Up-to-date versions can be installed from http://binaries.html-tidy.org/.

All these tests are run with collation set to the C locale, and for the examples and tests with environment variable LANGUAGE=en: this is to minimize differences between platforms.

Use R CMD check --help to obtain more information about the usage of the R package checker. A subset of the checking steps can be selected by adding command-line options. It also allows customization by setting environment variables _R_CHECK_*_ as described in Tools in R Internals: a set of these customizations similar to those used by CRAN can be selected by the option --as-cran (which works best if Internet access is available). Some Windows users may need to set environment variable R_WIN_NO_JUNCTIONS to a non-empty value. The test of cyclic declarations60in DESCRIPTION files needs repositories (including CRAN) set: do this in ~/.Rprofile, by e.g.

options(repos = c(CRAN="https://cran.r-project.org"))

One check customization which can be revealing is

_R_CHECK_CODETOOLS_PROFILE_="suppressLocalUnused=FALSE"

which reports unused local assignments. Not only does this point out computations which are unnecessary because their results are unused, it also can uncover errors. (Two such are to intend to update an object by assigning a value but mistype its name or assign in the wrong scope, for example using <- where <<- was intended.) This can give false positives, most commonly because of non-standard evaluation for formulae and because the intention is to return objects in the environment of a function for later use.

Complete checking of a package which contains a file README.md needs a reasonably current version of pandoc installed: see https://pandoc.org/installing.html.

You do need to ensure that the package is checked in a suitable locale if it contains non-ASCII characters. Such packages are likely to fail some of the checks in a C locale, and R CMD check will warn if it spots the problem. You should be able to check any package in a UTF-8 locale (if one is available). Beware that although a C locale is rarely used at a console, it may be the default if logging in remotely or for batch jobs.

Often R CMD check will need to consult a CRAN repository to check details of uninstalled packages. Normally this defaults to the CRAN main site, but a mirror can be specified by setting environment variables R_CRAN_WEB and (rarely needed) R_CRAN_SRC to the URL of a CRAN mirror.

It is possible to install a package and then check the installed package. To do so first install the package and keep a log of the installation:

R CMD INSTALL -l libdir pkg > pkg.log 2>&1

and then use

Rdev CMD check -l libdir --install=check:pkg.log pkg

(Specifying the library is required: it ensures that the just-installed package is the one checked. If you know for sure only one copy is installed you can use --install=skip: this is used for R installation’s make check-recommended.)


1.3.2 Building package tarballs

Packages may be distributed in source form as “tarballs” (.tar.gz files) or in binary form. The source form can be installed on all platforms with suitable tools and is the usual form for Unix-like systems; the binary form is platform-specific, and is the more common distribution form for the macOS and ‘x86_64’ Windows platforms.

Using R CMD build, the R package builder, one can build R package tarballs from their sources (for example, for subsequent release). It is recommended that packages are built for release by the current release version of R or ‘r-patched’, to avoid inadvertently picking up new features of a development version of R.

Prior to actually building the package in the standard gzipped tar file format, a few diagnostic checks and cleanups are performed. In particular, it is tested whether object indices exist and can be assumed to be up-to-date, and C, C++ and Fortran source files and relevant makefiles in a src directory are tested and converted to LF line-endings if necessary.

Run-time checks whether the package works correctly should be performed using R CMD check prior to invoking the final build procedure.

To exclude files from being put into the package, one can specify a list of exclude patterns in file .Rbuildignore in the top-level source directory. These patterns should be Perl-like regular expressions (see the help for regexp in R for the precise details), one per line, to be matched case-insensitively against the file and directory names relative to the top-level package source directory. In addition, directories from source control systems61 or from eclipse62, directories with names check, chm, or ending .Rcheck or Old or old and files GNUMakefile63, Read-and-delete-me or with base names starting with ‘.#’, or starting and ending with ‘#’, or ending in ‘~’, ‘.bak’ or ‘.swp’, are excluded by default64. In addition, same-package tarballs (from previous builds) and their binary forms will be excluded from the top-level directory, as well as those files in the R, demo and man directories which are flagged by R CMD check as having invalid names.

Use R CMD build --help to obtain more information about the usage of the R package builder.

Unless R CMD build is invoked with the --no-build-vignettes option (or the package’s DESCRIPTION contains ‘BuildVignettes: no’ or similar), it will attempt to (re)build the vignettes (see Writing package vignettes) in the package. To do so it installs the current package into a temporary library tree, but any dependent packages need to be installed in an available library tree (see the Note: at the top of this section).

Similarly, if the .Rd documentation files contain any \Sexpr macros (see Dynamic pages), the package will be temporarily installed to execute them. Post-execution binary copies of those pages containing build-time macros will be saved in build/partial.rdb. If there are any install-time or render-time macros, a .pdf version of the package manual will be built and installed in the build subdirectory. (This allows CRAN or other repositories to display the manual even if they are unable to install the package.) This can be suppressed by the option --no-manual or if package’s DESCRIPTION contains ‘BuildManual: no’ or similar.

One of the checks that R CMD build runs is for empty source directories. These are in most (but not all) cases unintentional, if they are intentional use the option --keep-empty-dirs (or set the environment variable _R_BUILD_KEEP_EMPTY_DIRS_ to ‘TRUE’, or have a ‘BuildKeepEmpty’ field with a true value in the DESCRIPTION file).

The --resave-data option allows saved images (.rda and .RData files) in the data directory to be optimized for size. It will also compress tabular files and convert .R files to saved images. It can take values no, gzip (the default if this option is not supplied, which can be changed by setting the environment variable _R_BUILD_RESAVE_DATA_) and best (equivalent to giving it without a value), which chooses the most effective compression. Using best adds a dependence on R (>= 2.10) to the DESCRIPTION file if bzip2 or xz compression is selected for any of the files. If this is thought undesirable, --resave-data=gzip (which is the default if that option is not supplied) will do what compression it can with gzip. A package can control how its data is resaved by supplying a ‘BuildResaveData’ field (with one of the values given earlier in this paragraph) in its DESCRIPTION file.

The --compact-vignettes option will run tools::compactPDF over the PDF files in inst/doc (and its subdirectories) to losslessly compress them. This is not enabled by default (it can be selected by environment variable _R_BUILD_COMPACT_VIGNETTES_) and needs qpdf (https://qpdf.sourceforge.io/) to be available.

It can be useful to run R CMD check --check-subdirs=yes on the built tarball as a final check on the contents.

Where a non-POSIX file system is in use which does not utilize execute permissions, some care is needed with permissions. This applies on Windows and to e.g. FAT-formatted drives and SMB-mounted file systems on other OSes. The ‘mode’ of the file recorded in the tarball will be whatever file.info() returns. On Windows this will record only directories as having execute permission and on other OSes it is likely that all files have reported ‘mode’ 0777. A particular issue is packages being built on Windows which are intended to contain executable scripts such as configure and cleanup: R CMD build ensures those two are recorded with execute permission.

Directory build of the package sources is reserved for use by R CMD build: it contains information which may not easily be created when the package is installed, including index information on the vignettes and, rarely, information on the help pages and perhaps a copy of the PDF reference manual (see above).


1.3.3 Building binary packages

Binary packages are compressed copies of installed versions of packages. They contain compiled shared libraries rather than C, C++ or Fortran source code, and the R functions are included in their installed form. The format and filename are platform-specific; for example, a binary package for Windows is usually supplied as a .zip file, and for the macOS platform the default binary package file extension is .tgz.

The recommended method of building binary packages is to use

R CMD INSTALL --build pkg

where pkg is either the name of a source tarball (in the usual .tar.gz format) or the location of the directory of the package source to be built. This operates by first installing the package and then packing the installed binaries into the appropriate binary package file for the particular platform.

By default, R CMD INSTALL --build will attempt to install the package into the default library tree for the local installation of R. This has two implications:

  • If the installation is successful, it will overwrite any existing installation of the same package.
  • The default library tree must have write permission; if not, the package will not install and the binary will not be created.

To prevent changes to the present working installation or to provide an install location with write access, create a suitably located directory with write access and use the -l option to build the package in the chosen location. The usage is then

R CMD INSTALL -l location --build pkg

where location is the chosen directory with write access. The package will be installed as a subdirectory of location, and the package binary will be created in the current directory.

Other options for R CMD INSTALL can be found using R CMD INSTALL --help, and platform-specific details for special cases are discussed in the platform-specific FAQs.

Finally, at least one web-based service is available for building binary packages from (checked) source code: WinBuilder (see https://win-builder.R-project.org/) is able to build ‘x86_64’ Windows binaries. Note that this is intended for developers on other platforms who do not have access to Windows but wish to provide binaries for the Windows platform.


1.4 Writing package vignettes

In addition to the help files in Rd format, R packages allow the inclusion of documents in arbitrary other formats. The standard location for these is subdirectory inst/doc of a source package, the contents will be copied to subdirectory doc when the package is installed. Pointers from package help indices to the installed documents are automatically created. Documents in inst/doc can be in arbitrary format, however we strongly recommend providing them in PDF format, so users on almost all platforms can easily read them. To ensure that they can be accessed from a browser (as an HTML index is provided), the file names should start with an ASCII letter and be comprised entirely of ASCII letters or digits or hyphen or underscore.

A special case is package vignettes. Vignettes are documents in PDF or HTML format obtained from plain-text literate source files from which R knows how to extract R code and create output (in PDF/HTML or intermediate LaTeX). Vignette engines do this work, using “tangle” and “weave” functions respectively. Sweave, provided by the R distribution, is the default engine. Other vignette engines besides Sweave are supported; see Non-Sweave vignettes.

Package vignettes have their sources in subdirectory vignettes of the package sources. Note that the location of the vignette sources only affects R CMD build and R CMD check: the tarball built by R CMD build includes in inst/doc the components intended to be installed.

Sweave vignette sources are normally given the file extension .Rnw or .Rtex, but for historical reasons extensions65 .Snw and .Stex are also recognized. Sweave allows the integration of LaTeX documents: see the Sweave help page in R and the Sweave vignette in package utils for details on the source document format.

Package vignettes are tested by R CMD check by executing all R code chunks they contain (except those marked for non-evaluation, e.g., with option eval=FALSE for Sweave). The R working directory for all vignette tests in R CMD check is a copy of the vignette source directory. Make sure all files needed to run the R code in the vignette (data sets, …) are accessible by either placing them in the inst/doc hierarchy of the source package or by using calls to system.file(). All other files needed to re-make the vignettes (such as LaTeX style files, BibTeX input files and files for any figures not created by running the code in the vignette) must be in the vignette source directory. R CMD check will check that vignette production has succeeded by comparing modification times of output files in inst/doc with the source in vignettes.

R CMD build will automatically66 create the (PDF or HTML versions of the) vignettes in inst/doc for distribution with the package sources. By including the vignette outputs in the package sources it is not necessary that these can be re-built at install time, i.e., the package author can use private R packages, screen snapshots and LaTeX extensions which are only available on their machine.67

By default R CMD build will run Sweave on all Sweave vignette source files in vignettes. If Makefile is found in the vignette source directory, then R CMD build will try to run make after the Sweave runs, otherwise texi2pdf is run on each .tex file produced.

The first target in the Makefile should take care of both creation of PDF/HTML files and cleaning up afterwards (including after Sweave), i.e., delete all files that shall not appear in the final package archive. Note that if the make step runs R it needs to be careful to respect the environment values of R_LIBS and R_HOME68. Finally, if there is a Makefile and it has a ‘clean:’ target, make clean is run.

All the usual caveats about including a Makefile apply. It must be portable (no GNU extensions), use LF line endings and must work correctly with a parallel make: too many authors have written things like

## BAD EXAMPLE
all: pdf clean

pdf: ABC-intro.pdf ABC-details.pdf

%.pdf:  %.tex
        texi2dvi --pdf $*

clean:
        rm *.tex ABC-details-*.pdf

which will start removing the source files whilst pdflatex is working.

Metadata lines can be placed in the source file, preferably in LaTeX comments in the preamble. One such is a \VignetteIndexEntry of the form

%\VignetteIndexEntry{Using Animal}

Others you may see are \VignettePackage (currently ignored), \VignetteDepends (a comma-separated list of package names) and \VignetteKeyword (which replaced \VignetteKeywords). These are processed at package installation time to create the saved data frame Meta/vignette.rds. The \VignetteEngine statement is described in Non-Sweave vignettes. Vignette metadata can be extracted from a source file using tools::vignetteInfo.

At install time an HTML index for all vignettes in the package is automatically created from the \VignetteIndexEntry statements unless a file index.html exists in directory inst/doc. This index is linked from the HTML help index for the package. If you do supply a inst/doc/index.html file it should contain relative links only to files under the installed doc directory, or perhaps (not really an index) to HTML help files or to the DESCRIPTION file, and be valid HTML as confirmed via the W3C Markup Validation Service or Validator.nu.

Sweave/Stangle allows the document to specify the split=TRUE option to create a single R file for each code chunk: this will not work for vignettes where it is assumed that each vignette source generates a single file with the vignette extension replaced by .R.

Do watch that PDFs are not too large – one in a CRAN package was 72MB! This is usually caused by the inclusion of overly detailed figures, which will not render well in PDF viewers. Sometimes it is much better to generate fairly high resolution bitmap (PNG, JPEG) figures and include those in the PDF document.

When R CMD build builds the vignettes, it copies these and the vignette sources from directory vignettes to inst/doc. To install any other files from the vignettes directory, include a file vignettes/.install_extras which specifies these as Perl-like regular expressions on one or more lines. (See the description of the .Rinstignore file for full details.)


1.4.1 Encodings and vignettes

Vignettes will in general include descriptive text, R input, R output and figures, LaTeX include files and bibliographic references. As any of these may contain non-ASCII characters, the handling of encodings can become very complicated.

The vignette source file should be written in ASCII or contain a declaration of the encoding (see below). This applies even to comments within the source file, since vignette engines process comments to look for options and metadata lines. When an engine’s weave and tangle functions are called on the vignette source, it will be converted to the encoding of the current R session.

Stangle() will produce an R code file in the current locale’s encoding: for a non-ASCII vignette what that is is recorded in a comment at the top of the file.

Sweave() will produce a .tex file in the current encoding, or in UTF-8 if that is declared. Non-ASCII encodings need to be declared to LaTeX via a line like

\usepackage[utf8]{inputenc}

(It is also possible to use the more recent ‘inputenx’ LaTeX package.) For files where this line is not needed (e.g. chapters included within the body of a larger document, or non-Sweave vignettes), the encoding may be declared using a comment like

%\VignetteEncoding{UTF-8}

If the encoding is UTF-8, this can also be declared using the declaration

%\SweaveUTF8

If no declaration is given in the vignette, it will be assumed to be in the encoding declared for the package. If there is no encoding declared in either place, then it is an error to use non-ASCII characters in the vignette.

In any case, be aware that LaTeX may require the ‘usepackage’ declaration.

Sweave() will also parse and evaluate the R code in each chunk. The R output will also be in the current locale (or UTF-8 if so declared), and should be covered by the ‘inputenc’ declaration. One thing people often forget is that the R output may not be ASCII even for ASCII R sources, for many possible reasons. One common one is the use of ‘fancy’ quotes: see the R help on sQuote: note carefully that it is not portable to declare UTF-8 or CP1252 to cover such quotes, as their encoding will depend on the locale used to run Sweave(): this can be circumvented by setting options(useFancyQuotes="UTF-8") in the vignette.

The final issue is the encoding of figures – this applies only to PDF figures and not PNG etc. The PDF figures will contain declarations for their encoding, but the Sweave option pdf.encoding may need to be set appropriately: see the help for the pdf() graphics device.

As a real example of the complexities, consider the fortunes package version ‘1.4-0’. That package did not have a declared encoding, and its vignette was in ASCII. However, the data it displays are read from a UTF-8 CSV file and will be assumed to be in the current encoding, so fortunes.tex will be in UTF-8 in any locale. Had read.table been told the data were UTF-8, fortunes.tex would have been in the locale’s encoding.


1.4.2 Non-Sweave vignettes

Vignettes in formats other than Sweave are supported via “vignette engines”. For example knitr version 1.1 or later can create .tex files from a variation on Sweave format, and .html files from a variation on “markdown” format. These engines replace the Sweave() function with other functions to convert vignette source files into LaTeX files for processing into .pdf, or directly into .pdf or .html files. The Stangle() function is replaced with a function that extracts the R source from a vignette.

R recognizes non-Sweave vignettes using filename extensions specified by the engine. For example, the knitr package supports the extension .Rmd (standing for “R markdown”). The user indicates the vignette engine within the vignette source using a \VignetteEngine line, for example

%\VignetteEngine{knitr::knitr}

This specifies the name of a package and an engine to use in place of Sweave in processing the vignette. As Sweave is the only engine supplied with the R distribution, the package providing any other engine must be specified in the ‘VignetteBuilder’ field of the package DESCRIPTION file, and also specified in the ‘Suggests’, ‘Imports’ or ‘Depends’ field (since its namespace must be available to build or check your package). If more than one package is specified as a builder, they will be searched in the order given there. The utils package is always implicitly appended to the list of builder packages, but may be included earlier to change the search order.

Note that a package with non-Sweave vignettes should always have a ‘VignetteBuilder’ field in the DESCRIPTION file, since this is how R CMD check recognizes that there are vignettes to be checked: packages listed there are required when the package is checked.

The vignette engine can produce .tex, .pdf, or .html files as output. If it produces .tex files, R will call texi2pdf to convert them to .pdf for display to the user (unless there is a Makefile in the vignettes directory).

Package writers who would like to supply vignette engines need to register those engines in the package .onLoad function. For example, that function could make the call

tools::vignetteEngine("knitr", weave = vweave, tangle = vtangle,
                      pattern = "[.]Rmd$", package = "knitr")

(The actual registration in knitr is more complicated, because it supports other input formats.) See the ?tools::vignetteEngine help topic for details on engine registration.


1.5 Package namespaces

R has a namespace management system for code in packages. This system allows the package writer to specify which variables in the package should be exported to make them available to package users, and which variables should be imported from other packages.

The namespace for a package is specified by the NAMESPACE file in the top level package directory. This file contains namespace directives describing the imports and exports of the namespace. Additional directives register any shared objects to be loaded and any S3-style methods that are provided. Note that although the file looks like R code (and often has R-style comments) it is not processed as R code. Only very simple conditional processing of if statements is implemented.

Packages are loaded and attached to the search path by calling library or require. Only the exported variables are placed in the attached frame. Loading a package that imports variables from other packages will cause these other packages to be loaded as well (unless they have already been loaded), but they will not be placed on the search path by these implicit loads. Thus code in the package can only depend on objects in its own namespace and its imports (including the base namespace) being visible69.

Namespaces are sealed once they are loaded. Sealing means that imports and exports cannot be changed and that internal variable bindings cannot be changed. Sealing allows a simpler implementation strategy for the namespace mechanism and allows code analysis and compilation tools to accurately identify the definition corresponding to a global variable reference in a function body.

The namespace controls the search strategy for variables used by functions in the package. If not found locally, R searches the package namespace first, then the imports, then the base namespace and then the normal search path (so the base namespace precedes the normal search rather than being at the end of it).


1.5.1 Specifying imports and exports

Exports are specified using the export directive in the NAMESPACE file. A directive of the form

export(f, g)

specifies that the variables f and g are to be exported. (Note that variable names may be quoted, and reserved words and non-standard names such as [<-.fractions must be.)

For packages with many variables to export it may be more convenient to specify the names to export with a regular expression using exportPattern. The directive

exportPattern("^[^.]")

exports all variables that do not start with a period. However, such broad patterns are not recommended for production code: it is better to list all exports or use narrowly-defined groups. (This pattern applies to S4 classes.) Beware of patterns which include names starting with a period: some of these are internal-only variables and should never be exported, e.g. ‘.__S3MethodsTable__.’ (and loading excludes known cases).

Packages implicitly import the base namespace. Variables exported from other packages with namespaces need to be imported explicitly using the directives import and importFrom. The import directive imports all exported variables from the specified package(s). Thus the directives

import(foo, bar)

specifies that all exported variables in the packages foo and bar are to be imported. If only some of the exported variables from a package are needed, then they can be imported using importFrom. The directive

importFrom(foo, f, g)

specifies that the exported variables f and g of the package foo are to be imported. Using importFrom selectively rather than import is good practice and recommended notably when importing from packages with more than a dozen exports and especially from those written by others (so what they export can change in future).

To import every symbol from a package but for a few exceptions, pass the except argument to import. The directive

import(foo, except=c(bar, baz))

imports every symbol from foo except bar and baz. The value of except should evaluate to something coercible to a character vector, after substituting each symbol for its corresponding string.

It is possible to export variables from a namespace which it has imported from other namespaces: this has to be done explicitly and not via exportPattern.

If a package only needs a few objects from another package it can use a fully qualified variable reference in the code instead of a formal import. A fully-qualified reference to the function f in package foo is of the form foo::f. This is slightly less efficient than a formal import and also loses the advantage of recording all dependencies in the NAMESPACE file (but they still need to be recorded in the DESCRIPTION file). Evaluating foo::f will cause package foo to be loaded, but not attached, if it was not loaded already—this can be an advantage in delaying the loading of a rarely used package. However, if foo is listed only in ‘Suggests’ or ‘Enhances’ this also delays the check that it is installed: it is good practice to use such imports conditionally (e.g. via requireNamespace("foo", quietly = TRUE)).

Using the foo::f form will be necessary when a package needs to use a function of the same name from more than one namespace.

Using foo:::f instead of foo::f allows access to unexported objects. This is generally not recommended, as the existence or semantics of unexported objects may be changed by the package author in routine maintenance.


1.5.2 Registering S3 methods

The standard method for S3-style UseMethod dispatching might fail to locate methods defined in a package that is imported but not attached to the search path. To ensure that these methods are available the packages defining the methods should ensure that the generics are imported and register the methods using S3method directives. If a package defines a function print.foo intended to be used as a print method for class foo, then the directive

S3method(print, foo)

ensures that the method is registered and available for UseMethod dispatch, and the function print.foo does not need to be exported. Since the generic print is defined in base it does not need to be imported explicitly.

(Note that function and class names may be quoted, and reserved words and non-standard names such as [<- and function must be.)

It is possible to specify a third argument to S3method, the function to be used as the method, for example

S3method(print, check_so_symbols, .print.via.format)

when print.check_so_symbols is not needed.

As from R 3.6.0 one can also use S3method() directives to perform delayed registration. With

if(getRversion() >= "3.6.0") {
    S3method(pkg::gen, cls)
}

function gen.cls will get registered as an S3 method for class cls and generic gen from package pkg only when the namespace of pkg is loaded. This can be employed to deal with situations where the method is not “immediately” needed, and having to pre-load the namespace of pkg (and all its strong dependencies) in order to perform immediate registration is considered too onerous.


1.5.3 Load hooks

There are a number of hooks called as packages are loaded, attached, detached, and unloaded. See help(".onLoad") for more details.

Since loading and attaching are distinct operations, separate hooks are provided for each. These hook functions are called .onLoad and .onAttach. They both take arguments70 libname and pkgname; they should be defined in the namespace but not exported.

Packages can use a .onDetach or .Last.lib function (provided the latter is exported from the namespace) when detach is called on the package. It is called with a single argument, the full path to the installed package. There is also a hook .onUnload which is called when the namespace is unloaded (via a call to unloadNamespace, perhaps called by detach(unload = TRUE)) with argument the full path to the installed package’s directory. Functions .onUnload and .onDetach should be defined in the namespace and not exported, but .Last.lib does need to be exported.

Packages are not likely to need .onAttach (except perhaps for a start-up banner); code to set options and load shared objects should be placed in a .onLoad function, or use made of the useDynLib directive described next.

User-level hooks are also available: see the help on function setHook.

These hooks are often used incorrectly. People forget to export .Last.lib. Compiled code should be loaded in .onLoad (or via a useDynLb directive: see below) and unloaded in .onUnload. Do remember that a package’s namespace can be loaded without the namespace being attached (e.g. by pkgname::fun) and that a package can be detached and re-attached whilst its namespace remains loaded.

It is good practice for these functions to be quiet. Any messages should use packageStartupMessage so users (include check scripts) can suppress them if desired.


1.5.4 useDynLib

A NAMESPACE file can contain one or more useDynLib directives which allows shared objects that need to be loaded.71 The directive

useDynLib(foo)

registers the shared object foo72 for loading with library.dynam. Loading of registered object(s) occurs after the package code has been loaded and before running the load hook function. Packages that would only need a load hook function to load a shared object can use the useDynLib directive instead.

The useDynLib directive also accepts the names of the native routines that are to be used in R via the .C, .Call, .Fortran and .External interface functions. These are given as additional arguments to the directive, for example,

useDynLib(foo, myRoutine, myOtherRoutine)

By specifying these names in the useDynLib directive, the native symbols are resolved when the package is loaded and R variables identifying these symbols are added to the package’s namespace with these names. These can be used in the .C, .Call, .Fortran and .External calls in place of the name of the routine and the PACKAGE argument. For instance, we can call the routine myRoutine from R with the code

 .Call(myRoutine, x, y)

rather than

 .Call("myRoutine", x, y, PACKAGE = "foo")

There are at least two benefits to this approach. Firstly, the symbol lookup is done just once for each symbol rather than each time the routine is invoked. Secondly, this removes any ambiguity in resolving symbols that might be present in more than one DLL. However, this approach is nowadays deprecated in favour of supplying registration information (see below).

In some circumstances, there will already be an R variable in the package with the same name as a native symbol. For example, we may have an R function in the package named myRoutine. In this case, it is necessary to map the native symbol to a different R variable name. This can be done in the useDynLib directive by using named arguments. For instance, to map the native symbol name myRoutine to the R variable myRoutine_sym, we would use

useDynLib(foo, myRoutine_sym = myRoutine, myOtherRoutine)

We could then call that routine from R using the command

 .Call(myRoutine_sym, x, y)

Symbols without explicit names are assigned to the R variable with that name.

In some cases, it may be preferable not to create R variables in the package’s namespace that identify the native routines. It may be too costly to compute these for many routines when the package is loaded if many of these routines are not likely to be used. In this case, one can still perform the symbol resolution correctly using the DLL, but do this each time the routine is called. Given a reference to the DLL as an R variable, say dll, we can call the routine myRoutine using the expression

 .Call(dll$myRoutine, x, y)

The $ operator resolves the routine with the given name in the DLL using a call to getNativeSymbol. This is the same computation as above where we resolve the symbol when the package is loaded. The only difference is that this is done each time in the case of dll$myRoutine.

In order to use this dynamic approach (e.g., dll$myRoutine), one needs the reference to the DLL as an R variable in the package. The DLL can be assigned to a variable by using the variable = dllName format used above for mapping symbols to R variables. For example, if we wanted to assign the DLL reference for the DLL foo in the example above to the variable myDLL, we would use the following directive in the NAMESPACE file:

myDLL = useDynLib(foo, myRoutine_sym = myRoutine, myOtherRoutine)

Then, the R variable myDLL is in the package’s namespace and available for calls such as myDLL$dynRoutine to access routines that are not explicitly resolved at load time.

If the package has registration information (see Registering native routines), then we can use that directly rather than specifying the list of symbols again in the useDynLib directive in the NAMESPACE file. Each routine in the registration information is specified by giving a name by which the routine is to be specified along with the address of the routine and any information about the number and type of the parameters. Using the .registration argument of useDynLib, we can instruct the namespace mechanism to create R variables for these symbols. For example, suppose we have the following registration information for a DLL named myDLL:

static R_NativePrimitiveArgType foo_t[] = {
    REALSXP, INTSXP, STRSXP, LGLSXP
};

static const R_CMethodDef cMethods[] = {
   {"foo", (DL_FUNC) &foo, 4, foo_t},
   {"bar_sym", (DL_FUNC) &bar, 0},
   {NULL, NULL, 0, NULL}
};

static const R_CallMethodDef callMethods[] = {
   {"R_call_sym", (DL_FUNC) &R_call, 4},
   {"R_version_sym", (DL_FUNC) &R_version, 0},
   {NULL, NULL, 0}
};

Then, the directive in the NAMESPACE file

useDynLib(myDLL, .registration = TRUE)

causes the DLL to be loaded and also for the R variables foo, bar_sym, R_call_sym and R_version_sym to be defined in the package’s namespace.

Note that the names for the R variables are taken from the entry in the registration information and do not need to be the same as the name of the native routine. This allows the creator of the registration information to map the native symbols to non-conflicting variable names in R, e.g. R_version to R_version_sym for use in an R function such as

R_version <- function()
{
  .Call(R_version_sym)
}

Using argument .fixes allows an automatic prefix to be added to the registered symbols, which can be useful when working with an existing package. For example, package KernSmooth has

useDynLib(KernSmooth, .registration = TRUE, .fixes = "F_")

which makes the R variables corresponding to the Fortran symbols F_bkde and so on, and so avoid clashes with R code in the namespace.

NB: Using these arguments for a package which does not register native symbols merely slows down the package loading (although many CRAN packages have done so). Once symbols are registered, check that the corresponding R variables are not accidentally exported by a pattern in the NAMESPACE file.


1.5.5 An example

As an example consider two packages named foo and bar. The R code for package foo in file foo.R is

x <- 1
f <- function(y) c(x,y)
foo <- function(x) .Call("foo", x, PACKAGE="foo")
print.foo <- function(x, ...) cat("<a foo>\n")

Some C code defines a C function compiled into DLL foo (with an appropriate extension). The NAMESPACE file for this package is

useDynLib(foo)
export(f, foo)
S3method(print, foo)

The second package bar has code file bar.R

c <- function(...) sum(...)
g <- function(y) f(c(y, 7))
h <- function(y) y+9

and NAMESPACE file

import(foo)
export(g, h)

Calling library(bar) loads bar and attaches its exports to the search path. Package foo is also loaded but not attached to the search path. A call to g produces

> g(6)
[1]  1 13

This is consistent with the definitions of c in the two settings: in bar the function c is defined to be equivalent to sum, but in foo the variable c refers to the standard function c in base.


1.5.6 Namespaces with S4 classes and methods

Some additional steps are needed for packages which make use of formal (S4-style) classes and methods (unless these are purely used internally). The package should have Depends: methods73 in its DESCRIPTION and import(methods) or importFrom(methods, ...) plus any classes and methods which are to be exported need to be declared in the NAMESPACE file. For example, the stats4 package has

export(mle) # exporting methods implicitly exports the generic
importFrom("stats", approx, optim, pchisq, predict, qchisq, qnorm, spline)
## For these, we define methods or (AIC, BIC, nobs) an implicit generic:
importFrom("stats", AIC, BIC, coef, confint, logLik, nobs, profile,
           update, vcov)
exportClasses(mle, profile.mle, summary.mle)
## All methods for imported generics:
exportMethods(coef, confint, logLik, plot, profile, summary,
              show, update, vcov)
## implicit generics which do not have any methods here
export(AIC, BIC, nobs)

All S4 classes to be used outside the package need to be listed in an exportClasses directive. Alternatively, they can be specified using exportClassPattern74 in the same style as for exportPattern. To export methods for generics from other packages an exportMethods directive can be used.

Note that exporting methods on a generic in the namespace will also export the generic, and exporting a generic in the namespace will also export its methods. If the generic function is not local to this package, either because it was imported as a generic function or because the non-generic version has been made generic solely to add S4 methods to it (as for functions such as coef in the example above), it can be declared via either or both of export or exportMethods, but the latter is clearer (and is used in the stats4 example above). In particular, for primitive functions there is no generic function, so export would export the primitive, which makes no sense. On the other hand, if the generic is local to this package, it is more natural to export the function itself using export(), and this must be done if an implicit generic is created without setting any methods for it (as is the case for AIC in stats4).

A non-local generic function is only exported to ensure that calls to the function will dispatch the methods from this package (and that is not done or required when the methods are for primitive functions). For this reason, you do not need to document such implicitly created generic functions, and undoc in package tools will not report them.

If a package uses S4 classes and methods exported from another package, but does not import the entire namespace of the other package75, it needs to import the classes and methods explicitly, with directives

importClassesFrom(package, ...)
importMethodsFrom(package, ...)

listing the classes and functions with methods respectively. Suppose we had two small packages A and B with B using A. Then they could have NAMESPACE files

export(f1, ng1)
exportMethods("[")
exportClasses(c1)

and

importFrom(A, ng1)
importClassesFrom(A, c1)
importMethodsFrom(A, f1)
export(f4, f5)
exportMethods(f6, "[")
exportClasses(c1, c2)

respectively.

Note that importMethodsFrom will also import any generics defined in the namespace on those methods.

It is important if you export S4 methods that the corresponding generics are available. You may for example need to import coef from stats to make visible a function to be converted into its implicit generic. But it is better practice to make use of the generics exported by stats4 as this enables multiple packages to unambiguously set methods on those generics.


1.6 Writing portable packages

This section contains advice on writing packages to be used on multiple platforms or for distribution (for example to be submitted to a package repository such as CRAN).

Portable packages should have simple file names: use only alphanumeric ASCII characters and period (.), and avoid those names not allowed under Windows (see Package structure).

Many of the graphics devices are platform-specific: even X11() (aka x11()) which although emulated on Windows may not be available on a Unix-alike (and is not the preferred screen device on OS X). It is rarely necessary for package code or examples to open a new device, but if essential,76 use dev.new().

Use R CMD build to make the release .tar.gz file.

R CMD check provides a basic set of checks, but often further problems emerge when people try to install and use packages submitted to CRAN – many of these involve compiled code. Here are some further checks that you can do to make your package more portable.

  • If your package has a configure script, provide a configure.win or configure.ucrt script to be used on Windows (an empty configure.win file if no actions are needed).
  • If your package has a Makevars or Makefile file, make sure that you use only portable make features. Such files should be LF-terminated77 (including the final line of the file) and not make use of GNU extensions. (The POSIX specification is available at https://pubs.opengroup.org/onlinepubs/9699919799/utilities/make.html; anything not documented there should be regarded as an extension to be avoided. Further advice can be found at https://www.gnu.org/software/autoconf/manual/autoconf.html#Portable-Make. ) Commonly misused GNU extensions are conditional inclusions (ifeq and the like), ${shell ...}, ${wildcard ...} and similar, and the use of +=78 and :=. Also, the use of $< other than in implicit rules is a GNU extension, as is the $^ macro. As is the use of .PHONY (some other makes ignore it). Unfortunately makefiles which use GNU extensions often run on other platforms but do not have the intended results.

    Note that the -C flag for make is not included in the POSIX specification and is not implemented by some of the makes which have been used with R. However, it is more commonly implemented (e.g. by FreeBSD make) than the GNU-specific --directory=.

    You should not rely on built-in/default make rules, even when specified by POSIX, as some makes do not have the POSIX ones and others have altered them.

    The use of ${shell ...} can be avoided by using backticks, e.g.

    PKG_CPPFLAGS = `gsl-config --cflags`
    

    which works in all versions of make known79 to be used with R.

    If you really must require GNU make, declare it in the DESCRIPTION file by

    SystemRequirements: GNU make
    

    and ensure that you use the value of environment variable MAKE (and not just make) in your scripts. (On some platforms GNU make is available under a name such as gmake, and there SystemRequirements is used to set MAKE.) Your configure script (or similar) does need to check that the executable pointed to by MAKE is indeed GNU make.

    If you only need GNU make for parts of the package which are rarely needed (for example to create bibliography files under vignettes), use a file called GNUmakefile rather than Makefile as GNU make (only) will use the former.

    macOS has used GNU make for many years (it previously used BSD make), but the version has been frozen at 3.81 (from 2006).

    Since the only viable make for Windows is GNU make, it is permissible to use GNU extensions in files Makevars.win, Makevars.ucrt, Makefile.win or Makefile.ucrt.

  • If you use src/Makevars to compile code in a subdirectory, ensure that you have followed all the advice above. In particular
    • Anticipate a parallel make. See Using Makevars.
    • Pass macros down to the makefile in the subdirectory, including all the needed compiler flags (including PIC and visibility flags). If they are used (even by a default rule) in the subdirectory’s Makefile, this includes macros ‘AR’ and ‘RANLIB’. See Compiling in sub-directories, which has a C example. A C++ example:
      pkg/libpkg.a:
              (cd pkg && $(MAKE) -f make_pkg libpkg.a \
               CXX="$(CXX)" CXXFLAGS="$(CXXFLAGS) $(CXXPICFLAGS) $(C_VISIBILITY)" \
               AR="$(AR)" RANLIB="$(RANLIB)")
      
    • Ensure that cleanup will be performed by R CMD build, for example in a cleanup script or a ‘clean’ target.
  • If your package uses a src/Makefile file to compile code to be linked into R, ensure that it uses exactly the same compiler and flag settings that R uses when compiling such code: people often forget ‘PIC’ flags. If R CMD config is used, this needs something like (for C++)
    RBIN = `"${R_HOME}/bin/R"`
    CXX = `"${RBIN}" CMD config CXX`
    CXXFLAGS = `"${RBIN}" CMD config CXXFLAGS` `"${RBIN}" CMD config CXXPICFLAGS`
    
  • Names of source files including = (such as src/complex_Sig=gen.c) will confuse some make programs and should be avoided.
  • Bash extensions also need to be avoided in shell scripts, including expressions in Makefiles (which are passed to the shell for processing). Some R platforms use strict80 Bourne shells: an earlier R toolset on Windows81 and some Unix-alike OSes use ash (https://en.wikipedia.org/wiki/Almquist_shell, a ‘lightweight shell with few builtins) or derivatives such as dash. Beware of assuming that all the POSIX command-line utilities are available, especially on Windows where only a subset (which has changed by version of Rtools) is provided for use with R. One particular issue is the use of echo, for which two behaviours are allowed (https://pubs.opengroup.org/onlinepubs/9699919799/utilities/echo.html) and both have occurred as defaults on R platforms: portable applications should use neither -n (as the first argument) nor escape sequences. The recommended replacement for echo -n is the command printf. Another common issue is the construction
    export FOO=value
    

    which is bash-specific (first set the variable then export it by name).

    Using test -e (or [ -e ]) in shell scripts is not fully portable82: -f is normally what is intended. Flags -a and -o are nowadays declared obsolescent by POSIX and should not be used.

    Use of ‘brace expansion’, e.g.,

    rm -f src/*.{o,so,d}
    

    is not portable.

    The -o flag for set in shell scripts is optional in POSIX and not supported on all the platforms R is used on.

    The variable ‘OSTYPE’ is shell-specific and its values are rather unpredictable and may include a version such as ‘darwin19.0’: `uname` is often what is intended (with common values ‘Darwin’, ‘Linux’ and ‘SunOS’).

    On macOS which shell /bin/sh invokes is user- and platform-dependent: it might be bash version 3.2, dash or zsh (for new accounts it is zsh, for accounts ported from Mojave or earlier it is usually bash).

  • R is not built by default as a shared library on non-Windows platforms (although it commonly is on macOS to support the GUI), so there need not be a file libR.so nor libR.dylib. Users of cmake or rust have all too frequently assumed otherwise, so do ensure your package is checked under a vanilla R build. See Configuration options in R Installation and Administration for more information.
  • Make use of the abilities of your compilers to check the standards-conformance of your code. For example, gcc, clang and gfortran83 can be used with options -Wall -pedantic to alert you to potential problems. This is particularly important for C++, where g++ -Wall -pedantic will alert you to the use of some of the GNU extensions which fail to compile on most other C++ compilers. If R was not configured accordingly, one can achieve this via personal Makevars files. See Customizing package compilation in R Installation and Administration for more information.

    Portable C++ code needs to follow all of the 2011, 2014 and 2017 standards (including not using deprecated/removed features) or to specify C+11/14/17/20/23 where available (which is not the case on all R platforms). Currently C++20 support is patchy across R platforms.

    If using Fortran with the GNU compiler, use the flags -std=f95 -Wall -pedantic which reject most GNU extensions and features from later standards. (Although R only requires Fortran 90, gfortran does not have a way to specify that standard.) Also consider -std=f2008 as some recent compilers have Fortran 2008 or even 2018 as the minimum supported standard.

    As from macOS 11 (late 2020), its C compiler sets the flag -Werror=implicit-function-declaration by default which forces stricter conformance to C99. This can be used on other platforms with gcc or clang. If your package has a (autoconf-generated) configure script, try installing it whilst using this flag, and read through the config.log file — compilation warnings and errors can lead to features which are present not being detected. (If possible do this on several platforms.)

  • R CMD check performs some checks for non-portable compiler/linker flags in src/Makevars. However, it cannot check the meaning of such flags, and some are commonly accepted but with compiler-specific meanings. There are other non-portable flags which are not checked, nor are src/Makefile files and makefiles in sub-directories. As a comment in the code says

    It is hard to think of anything apart from -I* and -D* that is safe for general use …

    although -pthread is pretty close to portable. (Option -U is portable but little use on the command line as it will only cancel built-in defines (not portable) and those defined earlier on the command line (R does not use any).)

    The GNU option -pipe used to be widely accepted by C/C++/Fortran compilers, but has been removed in flang-new 18. In any case, it should not be used in distributed code as it may lead to excessive memory use.

    People have used configure to customize src/Makevars, including for specific compilers. This is unsafe for several reasons. First, unintended compilers might meet the check—for example, several compilers other than GCC identify themselves as ‘GCC’ whilst being only partially conformant. Second, future versions of compilers may behave differently (including updates to quite old series) so for example -Werror (and specializations) can make a package non-installable under a future version. Third, using flags to suppress diagnostic messages can hide important information for debugging on a platform not tested by the package maintainer. (R CMD check can optionally report on unsafe flags which were used.)

    Avoid the use of -march and especially -march=native. This allows the compiler to generate code that will only run on a particular class of CPUs (that of the compiling machine for ‘native’). People assume this is a ‘minimum’ CPU specification, but that is not how it is documented for gcc (it is accepted by clang but apparently it is undocumented what precisely it does, and it can be accepted and may be ignored for other compilers). (For personal use -mtune is safer, but still not portable enough to be used in a public package.) Not even gcc supports ‘native’ for all CPUs, and it can do surprising things if it finds a CPU released later than its version.

  • Do be very careful with passing arguments between R, C and Fortran code. In particular, long in C will be 32-bit on some R platforms (including 64-bit Windows), but 64-bit on most modern Unix and Linux platforms. It is rather unlikely that the use of long in C code has been thought through: if you need a longer type than int you should use a configure test for a C99/C++11 type such as int_fast64_t (and failing that, long long) and typedef your own type, or use another suitable type (such as size_t, but beware that is unsigned and ssize_t is not portable).

    It is not safe to assume that long and pointer types are the same size, and they are not on 64-bit Windows. If you need to convert pointers to and from integers use the C99/C++11 integer types intptr_t and uintptr_t (in the headers <stdint.h> and <cstdint>: they are not required to be implemented by the standards but are used in C code by R itself).

    Note that integer in Fortran corresponds to int in C on all R platforms. There is no such guarantee for Fortran logical, and recent gfortran maps it to int_least32_t on most platforms.

  • Under no circumstances should your compiled code ever call abort or exit84: these terminate the user’s R process, quite possibly losing all unsaved work. One usage that could call abort is the assert macro in C or C++ functions, which should never be active in production code. The normal way to ensure that is to define the macro NDEBUG, and R CMD INSTALL does so as part of the compilation flags. Beware of including headers (including from other packages) which could undefine it, now or in future versions. If you wish to use assert during development. you can include -UNDEBUG in PKG_CPPFLAGS or #undef it in your headers or code files. Note that your own src/Makefile or makefiles in sub-directories may also need to define NDEBUG.

    This applies not only to your own code but to any external software you compile in or link to.

    Nor should Fortran code call STOP nor EXIT (a GNU extension).

  • Compiled code should not write to stdout or stderr and C++ and Fortran I/O should not be used. As with the previous item such calls may come from external software and may never be called, but package authors are often mistaken about that.
  • Compiled code should not call the system random number generators such as rand, drand48 and random85, but rather use the interfaces to R’s RNGs described in Random number generation. In particular, if more than one package initializes a system RNG (e.g. via srand), they will interfere with each other. This applies also to Fortran 90’s random_number and random_seed, and Fortran 2018’s random_init. And to GNU Fortran’s rand, irand and srand. Except for drand48, what PRNG these functions use is implementation-dependent.

    Nor should the C++11 random number library be used nor any other third-party random number generators such as those in GSL.

  • Use of sprintf and vsprintf is regarded as a potential security risk and warned about on some platforms.86 R CMD check reports if any calls are found.
  • Errors in memory allocation and reading/writing outside arrays are very common causes of crashes (e.g., segfaults) on some machines. See Checking memory access for tools which can be used to look for this.
  • Many platforms will allow unsatisfied entry points in compiled code, but will crash the application (here R) if they are ever used. Some (notably Windows) will not. Looking at the output of
    nm -pg mypkg.so
    

    and checking if any of the symbols marked U is unexpected is a good way to avoid this.

  • Linkers have a lot of freedom in how to resolve entry points in dynamically-loaded code, so the results may differ by platform. One area that has caused grief is packages including copies of standard system software such as libz (especially those already linked into R). In the case in point, entry point gzgets was sometimes resolved against the old version compiled into the package, sometimes against the copy compiled into R and sometimes against the system dynamic library. The only safe solution is to rename the entry points in the copy in the package. We have even seen problems with entry point name myprintf, which is a system entry point87 on some Linux systems.

    A related issue is the naming of libraries built as part of the package installation. macOS and Windows have case-insensitive file systems, so using

    -L. -lLZ4
    

    in PKG_LIBS will match liblz4. And -L. only appends to the list of searched locations, and liblz4 might be found in an earlier-searched location (and has been). The only safe way is to give an explicit path, for example

    ./libLZ4.a
    
  • Conflicts between symbols in DLLs are handled in very platform-specific ways. Good ways to avoid trouble are to make as many symbols as possible static (check with nm -pg), and to use names which are clearly tied to your package (which also helps users if anything does go wrong). Note that symbol names starting with R_ are regarded as part of R’s namespace and should not be used in packages.
  • It is good practice for DLLs to register their symbols (see Registering native routines), restrict visibility (see Controlling visibility) and not allow symbol search (see Registering native routines). It should be possible for a DLL to have only one visible symbol, R_init_pkgname, on suitable platforms88, which would completely avoid symbol conflicts.
  • It is not portable to call compiled code in R or other packages via .Internal, .C, .Fortran, .Call or .External, since such interfaces are subject to change without notice and will probably result in your code terminating the R process.
  • Do not use (hard or symbolic) file links in your package sources. Where possible R CMD build will replace them by copies.
  • If you do not yourself have a Windows system, consider submitting your source package to WinBuilder (https://win-builder.r-project.org/) before distribution. If you need to check on an M1 Mac, there is a check service at https://mac.r-project.org/macbuilder/submit.html.
  • It is bad practice for package code to alter the search path using library, require or attach and this often does not work as intended. For alternatives, see Suggested packages and with().
  • Examples can be run interactively via example as well as in batch mode when checking. So they should behave appropriately in both scenarios, conditioning by interactive() the parts which need an operator or observer. For instance, progress bars89 are only appropriate in interactive use, as is displaying help pages or calling View() (see below).
  • Be careful with the order of entries in macros such as PKG_LIBS. Some linkers will re-order the entries, and behaviour can differ between dynamic and static libraries. Generally -L options should precede90 the libraries (typically specified by -l options) to be found from those directories, and libraries are searched once in the order they are specified. Not all linkers allow a space after -L .
  • Care is needed with the use of LinkingTo. This puts one or more directories on the include search path ahead of system headers but (prior to R 3.4.0) after those specified in the CPPFLAGS macro of the R build (which normally includes -I/usr/local/include, but most platforms ignore that and include it with the system headers).

    Any confusion would be avoided by having LinkingTo headers in a directory named after the package. In any case, name conflicts of headers and directories under package include directories should be avoided, both between packages and between a package and system and third-party software.

  • The ar utility is often used in makefiles to make static libraries. Its modifier u is defined by POSIX but is disabled in GNU ar on some Linux distributions which use ‘deterministic mode’. The safest way to make a static library is to first remove any existing file of that name then use $(AR) -cr and then $(RANLIB) if needed (which is system-dependent: on most systems91 ar always maintains a symbol table). The POSIX standard says options should be preceded by a hyphen (as in -cr), although most OSes accept them without. Note that on some systems ar -cr must have at least one file specified.

    The s modifier (to replace a separate call to ranlib) is required by X/OPEN but not POSIX, so ar -crs is not portable.

    For portability the AR and RANLIB macros should always be used – some builds require wrappers such as gcc-ar or extra arguments to specify plugins.

  • The strip utility is platform-specific (and CRAN prohibits removing debug symbols). For example the options --strip-debug and --strip-unneeded of the GNU version are not supported on macOS: the POSIX standard for strip does not mention any options, and what calling it without options does is platform-dependent. Stripping a .so file could even prevent it being dynamically loaded into R on an untested platform.

    ld -S invokes strip --strip-debug for GNU ld (and similarly on macOS) but is not portable: in particular on Solaris it did something completely different and took an argument.

  • Some people have a need to set a locale. Locale names are not portable, and e.g. ‘fr_FR.utf8’ is commonly used on Linux but not accepted on macOS. ‘fr_FR.UTF-8’ is more portable, being accepted on recent Linux, AIX, FreeBSD, macOS and Solaris (at least). However, some Linux distributions micro-package, so locales defined by glibc (including these examples) may not be installed.
  • Avoid spaces in file names, not least as they can cause difficulties for external tools. An example was a package with a knitr vignette that used spaces in plot names: this caused some older versions of pandoc to fail with a baffling error message.

    Non-ASCII filenames can also cause problems (particularly in non-UTF-8 locales).

  • Take care in naming LaTeX macros (also known as ‘commands’) in vignette sources: if these are also defined in a future version of one of the LaTeX packages used there will be a fatal error. One instance in 2021 was package ‘hyperref’ newly defining ‘\C’, ‘\F’, ‘\G’, ‘\U’ and ‘\textapprox’. If you are confident that your definitions will be the only ones relevant you can use ‘\renewcommand’ but it is better to use names clearly associated with your package.
  • Make sure that any version requirement for Java code is both declared in the ‘SystemRequirements’ field92 and tested at runtime (not least as the Java installation when the package is installed might not be the same as when the package is run and will not be for binary packages).

    When specifying a minimum Java version please use the official version names, which are (confusingly)

    1.1 1.2 1.3 1.4 5.0 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
    

    and as from 2018 a year.month scheme such as ‘18.9’ is also in use. Fortunately only the integer values are likely to be relevant. If at all possible, use one of the LTS versions (8, 11, 17, 21 …) as the minimum version. The preferred form of version specification is

    SystemRequirements: Java (>= 11)
    

    A suitable test for Java at least version 8 for packages using rJava would be something like

    .jinit()
    jv <- .jcall("java/lang/System", "S", "getProperty", "java.runtime.version")
    if(substr(jv, 1L, 2L) == "1.") {
      jvn <- as.numeric(paste0(strsplit(jv, "[.]")[[1L]][1:2], collapse = "."))
      if(jvn < 1.8) stop("Java >= 8 is needed for this package but not available")
    }
    

    Java 9 changed the format of this string (which used to be something like ‘1.8.0_292-b10’); Java 11 gave jv as ‘11+28’ whereas Java 11.0.11 gave ‘11.0.11+9’. (https://openjdk.org:443/jeps/322 details the current scheme. Note that it is necessary to allow for pre-releases like ‘11-ea+22’.)

    Note too that the compiler used to produce a jar can impose a minimum Java version, often resulting in an arcane message like

    java.lang.UnsupportedClassVersionError: ... Unsupported major.minor version 52.0
    

    (Where https://en.wikipedia.org/wiki/Java_class_file maps class-file version numbers to Java versions.) Compile with something like javac -target 11 to ensure this is avoided. Note this also applies to packages distributing (or even downloading) compiled Java code produced by others, so their requirements need to be checked (they are often not documented accurately) and accounted for. It should be possible to check the class-file version via command-line utility javap, if necessary after extracting the .class files from a .jar archive. For example,

    jar xvf some.jar
    javap -verbose path/to/some.class | grep major
    

    Some packages have stated a requirement on a particular JDK, but a package should only be requiring a JRE unless providing its own Java interface.

    Java 8 is still in widespread use (and may remain so because of licence changes and support on older OSes: OpenJDK has security support until March 2026). On the other hand, newer platforms may only have support for recent versions of Java: for ‘arm64’ macOS the first officially supported version was 17.

  • A package with a hard-to-satisfy system requirement is by definition not portable, annoyingly so if this is not declared in the ‘SystemRequirements’ field. The most common example is the use of pandoc, which is only available for a very limited range of platforms (and has onerous requirements to install from source) and has capabilities93 that vary by build but are not documented. Several recent versions of pandoc for macOS did not work on R’s then target of High Sierra (and this too was undocumented). Another example is the Rust compilation system (cargo and rustc).

    Usage of external commands should always be conditional on a test for presence (perhaps using Sys.which), as well as declared in the ‘SystemRequirements’ field. A package should pass its checks without warnings nor errors without the external command being present.

    An external command can be a (possibly optional) requirement for an imported or suggested package but needed for examples, tests or vignettes in the package itself. Such usages should always be declared and conditional.

    Interpreters for scripting languages such as Perl, Python and Ruby need to be declared as system requirements and used conditionally: for example macOS 10.16 was announced not to have them (but released as macOS 11 with them); later it was announced that macOS 12.3 does not have Python 2 and only a minimal install of Python 3 is included. Python 2 has passed end-of-life and been removed from many major distributions. Support for Rust or Go cannot be assumed.

    Command cmake is not commonly installed, and where it is, it might not be on the path. In particular, the most common location on macOS is /Applications/CMake.app/Contents/bin/cmake and that should be looked for if cmake is not found on the path.

  • Be sure to use portable encoding names: none of utf8, mac and macroman is portable. See the help for file for more details.
  • Do not invoke R by plain R, Rscript or (on Windows) Rterm in your examples, tests, vignettes, makefiles or other scripts. As pointed out in several places earlier in this manual, use something like
    "$(R_HOME)/bin/Rscript"
    "$(R_HOME)/bin$(R_ARCH_BIN)/Rterm"
    

    with appropriate quotes (as, although not recommended, R_HOME can contain spaces).

  • Do not use R_HOME in makefiles except when passing them to the shell. Specifically, do not use R_HOME in the argument to include, as R_HOME can contain spaces. Quoting the argument to include does not help. A portable and the recommended way to avoid the problem of spaces in ${R_HOME} is using option -f of make. This is easy to do with recursive invocation of make, which is also the only usual situation when R_HOME is needed in the argument for include.
    $(MAKE) -f "${R_HOME}/etc${R_ARCH}/Makeconf" -f Makefile.inner
    
  • If distributing datasets involving date-times, consider if a time zone needs to be specified. The most portable way to distribute date-times is as objects of class "POSIXct" and as these record the time in UTC, the time represented is independent of the time zone: but how it is printed may not be. Objects of class "POSIXlt" should have a "tzone" attribute. Dates (e.g, birthdays) are conventionally considered independently of time zone.
  • If at all possible avoid any Internet access during package installation. Installation and use may well be on different machines/accounts and those allowed to install software may have no Internet access, and being self-contained helps ensure long-term reproducibility.

Do be careful in what your tests (and examples) actually test. Bad practice seen in distributed packages include:

  • It is not reasonable to test the time taken by a command: you cannot know how fast or how heavily loaded an R platform might be. At best you can test a ratio of times, and even that is fraught with difficulties and not advisable: for example, the garbage collector may trigger at unpredictable times following heuristics that may change without notice.
  • Do not test the exact format of R messages (from R itself or from other packages): They change, and they can be translated.

    Packages have even tested the exact format of system error messages, which are platform-dependent and perhaps locale-dependent. For example, in late 2021 libcurl changed its warning/error messages, including when URLs are not found.

  • Do not test for the absence of warnings (something users of testthat are fond of). Future changes in either R or packages you make use of can create new warnings, and your tests should not make these into errors. (Some deprecation notices may be intended to remain as warnings for a long time.)
  • If you use functions such as View, remember that in testing there is no one to look at the output. It is better to use something like one of
    if(interactive()) View(obj) else print(head(obj))
    if(interactive()) View(obj) else str(obj)
    
  • Be careful when comparing file paths. There can be multiple paths to a single file, and some of these can be very long character strings. If possible canonicalize paths before comparisons, but study ?normalizePath to be aware of the pitfalls.
  • Only test the accuracy of results if you have done a formal error analysis. Things such as checking that probabilities numerically sum to one are silly: numerical tests should always have a tolerance. That the tests on your platform achieve a particular tolerance says little about other platforms. R is configured by default to make use of long doubles where available, but they may not be available or be too slow for routine use. Most R platforms use ‘ix86’ or ‘x86_64’ CPUs: these may use extended precision registers on some but not all of their FPU instructions. Thus the achieved precision can depend on the compiler version and optimization flags—our experience is that 32-bit builds tend to be less precise than 64-bit ones. But not all platforms use those CPUs, and not all94 which use them configure them to allow the use of extended precision. In particular, current ARM CPUs do not have extended precision nor long doubles, and clang currently has long double the same as double on all ARM CPUs. On the other hand some CPUs have higher-precision modes which may be used for long double, notably 64-bit PowerPC and Sparc.

    If you must try to establish a tolerance empirically, configure and build R with --disable-long-double and use appropriate compiler flags (such as -ffloat-store and -fexcess-precision=standard for gcc, depending on the CPU type95) to mitigate the effects of extended-precision calculations. The platform most often seen to give different numerical results is ‘arm64’ macOS, so be sure to include that in any empirical determination.

    Tests which involve random inputs or non-deterministic algorithms should normally set a seed or be tested for many seeds.

  • Tests should use options(warn = 1) as reporting
    There were 22 warnings (use warnings() to see them)
    

    is pointless, especially for automated checking systems.

  • If your package uses dates/times, ensure that it works in all timezones, especially those near boundaries (problems have most often be seen in ‘Europe/London’ (zero offset in Winter) and ‘Pacific/Auckland’, near enough the International Date line) and with offsets not in whole hours (Adelaide, Chatham Islands, …). More extreme examples are ‘Africa/Conakry’ (permanent UTC), ‘Asia/Calcutta’ (no DST, permanent half-hour offset) and ‘Pacific/Kiritimati’(no DST, more than 12 hours ahead of UTC).

1.6.1 PDF size

There are a several tools available to reduce the size of PDF files: often the size can be reduced substantially with no or minimal loss in quality. Not only do large files take up space: they can stress the PDF viewer and take many minutes to print (if they can be printed at all).

qpdf (https://qpdf.sourceforge.io/) can compress losslessly. It is fairly readily available (e.g. it is included in rtools, has packages in Debian/Ubuntu/Fedora, and is installed as part of the CRAN macOS distribution of R). R CMD build has an option to run qpdf over PDF files under inst/doc and replace them if at least 10Kb and 10% is saved. The full path to the qpdf command can be supplied as environment variable R_QPDF (and is on the CRAN binary of R for macOS). It seems MiKTeX does not use PDF object compression and so qpdf can reduce considerably the sizes of files it outputs: MiKTeX’s defaults can be overridden by code in the preamble of an Sweave or LaTeX file — see how this is done for the R reference manual at https://svn.r-project.org/R/trunk/doc/manual/refman.top.

Other tools can reduce the size of PDFs containing bitmap images at excessively high resolution. These are often best re-generated (for example Sweave defaults to 300 ppi, and 100–150 is more appropriate for a package manual). These tools include Adobe Acrobat (not Reader), Apple’s Preview96 and Ghostscript (which converts PDF to PDF by

ps2pdf options -dAutoRotatePages=/None -dPrinted=false in.pdf out.pdf

and suitable options might be

-dPDFSETTINGS=/ebook
-dPDFSETTINGS=/screen

See https://ghostscript.readthedocs.io/en/latest/VectorDevices.html for more such and consider all the options for image downsampling). There have been examples in CRAN packages for which current versions of Ghostscript produced much bigger reductions than earlier ones (e.g. at the upgrades from 9.50 to 9.52, from 9.55 to 9.56 and then to 10.00.0).

We come across occasionally large PDF files containing excessively complicated figures using PDF vector graphics: such figures are often best redesigned or failing that, output as PNG files.

Option --compact-vignettes to R CMD build defaults to value ‘qpdf’: use ‘both’ to try harder to reduce the size, provided you have Ghostscript available (see the help for tools::compactPDF).


1.6.2 Check timing

There are several ways to find out where time is being spent in the check process. Start by setting the environment variable _R_CHECK_TIMINGS_ to ‘0’. This will report the total CPU times (not Windows) and elapsed times for installation and running examples, tests and vignettes, under each sub-architecture if appropriate. For tests and vignettes, it reports the time for each as well as the total.

Setting _R_CHECK_TIMINGS_ to a positive value sets a threshold (in seconds elapsed time) for reporting timings.

If you need to look in more detail at the timings for examples, use option --timings to R CMD check (this is set by --as-cran). This adds a summary to the check output for all the examples with CPU or elapsed time of more than 5 seconds. It produces a file mypkg.Rcheck/mypkg-Ex.timings containing timings for each help file: it is a tab-delimited file which can be read into R for further analysis.

Timings for the tests and vignette runs are given at the bottom of the corresponding log file: note that log files for successful vignette runs are only retained if environment variable _R_CHECK_ALWAYS_LOG_VIGNETTE_OUTPUT_ is set to a true value.


1.6.3 Encoding issues

The issues in this subsection have been much alleviated by the change in R 4.2.0 to running the Windows port of R in a UTF-8 locale where available. However, Windows users might be running an earlier version of R on an earlier version of Windows which does not support UTF-8 locales.

Care is needed if your package contains non-ASCII text, and in particular if it is intended to be used in more than one locale. It is possible to mark the encoding used in the DESCRIPTION file and in .Rd files, as discussed elsewhere in this manual.

First, consider carefully if you really need non-ASCII text. Some users of R will only be able to view correctly text in their native language group (e.g. Western European, Eastern European, Simplified Chinese) and ASCII.97. Other characters may not be rendered at all, rendered incorrectly, or cause your R code to give an error. For .Rd documentation, marking the encoding and including ASCII transliterations is likely to do a reasonable job. The set of characters which is commonly supported is wider than it used to be around 2000, but non-Latin alphabets (Greek, Russian, Georgian, …) are still often problematic and those with double-width characters (Chinese, Japanese, Korean, emoji) often need specialist fonts to render correctly.

Several CRAN packages have messages in their R code in French (and a few in German). A better way to tackle this is to use the internationalization facilities discussed elsewhere in this manual.

Function showNonASCIIfile in package tools can help in finding non-ASCII bytes in files.

There is a portable way to have arbitrary text in character strings (only) in your R code, which is to supply them in Unicode as ‘\uxxxx’ escapes (or, rarely needed except for emojis, ‘\Uxxxxxxxx’ escapes). If there are any characters not in the current encoding the parser will encode the character string as UTF-8 and mark it as such. This applies also to character strings in datasets: they can be prepared using ‘\uxxxx’ escapes or encoded in UTF-8 in a UTF-8 locale, or even converted to UTF-8 via iconv(). If you do this, make sure you have ‘R (>= 2.10)’ (or later) in the ‘Depends’ field of the DESCRIPTION file.

R sessions running in non-UTF-8 locales will if possible re-encode such strings for display (and this is done by RGui on older versions of Windows, for example). Suitable fonts will need to be selected or made available98 both for the console/terminal and graphics devices such as ‘X11()’ and ‘windows()’. Using ‘postscript’ or ‘pdf’ will choose a default 8-bit encoding depending on the language of the UTF-8 locale, and your users would need to be told how to select the ‘encoding’ argument.

Note that the previous two paragraphs only apply to character strings in R code. Non-ASCII characters are particularly prevalent in comments (in the R code of the package, in examples, tests, vignettes and even in the NAMESPACE file) but should be avoided there. Most commonly people use the Windows extensions to Latin-1 (often directional single and double quotes, ellipsis, bullet and en and em dashes) which are not supported in strict Latin-1 locales nor in CJK locales on Windows. A surprisingly common misuse is to use a right quote in ‘don't’ instead of the correct apostrophe.

Datasets can include marked UTF-8 or Latin-1 character strings. As R is nowadays unlikely to be run in a Latin-1 or Windows’ CP1252 locale, for performance reasons these should be converted to UTF-8.

If you want to run R CMD check on a Unix-alike over a package that sets a package encoding in its DESCRIPTION file and do not use a UTF-8 locale you may need to specify a suitable locale via environment variable R_ENCODING_LOCALES. The default is equivalent to the value

"latin1=en_US:latin2=pl_PL:UTF-8=en_US.UTF-8:latin9=fr_FR.iso885915@euro"

(which is appropriate for a system based on glibc: macOS requires latin9=fr_FR.ISO8859-15) except that if the current locale is UTF-8 then the package code is translated to UTF-8 for syntax checking, so it is strongly recommended to check in a UTF-8 locale.


1.6.4 Portable C and C++ code

Writing portable C and C++ code is mainly a matter of observing the standards (C99, C++14 or where declared C++11/17/20) and testing that extensions (such as POSIX functions) are supported. Do make maximal use of your compiler diagnostics — this typically means using flags -Wall and -pedantic for both C and C++ and additionally -Werror=implicit-function-declaration and -Wstrict-prototypes for C (on some platforms and compiler versions) these are part of -Wall or -pedantic).

C++ standards: From version 3.6.0 (3.6.2 on Windows), R defaulted to C++11 where available99; from R 4.1.0 to C++14 and from R 4.3.0 to C++17 (where available). However, in earlier versions the default standard was that of the compiler used, often C++98 or C++14, and the default is likely to change in future. For maximal portability a package should either specify a standard (see Using C++ code) or be tested under all of C++11, C++98, C++14 and C++17. (Specifying C++14 or later will limit portability.)

Note that the ‘TR1’ C++ extensions are not part of any of these standards and the <tr1/name> headers are not supplied by some of the compilers used for R, including on macOS. (Use the C++11 versions instead.)

A common error is to assume recent versions of compilers or OSes. In production environments ‘long term support’ versions of OSes may be in use for many years,100 and their compilers may not be updated during that time. For example, GCC 4.8 was still in use in 2022 and could be (in RHEL 7) until 2028: that supports neither C++14 nor C++17.

The POSIX standards only require recently-defined functions to be declared if certain macros are defined with large enough values, and on some compiler/OS combinations101 they are not declared otherwise. So you may need to include something like one of

#define _XOPEN_SOURCE 600

or

#ifdef __GLIBC__
# define _POSIX_C_SOURCE 200809L
#endif

before any headers. (strdup, strncasecmp and strnlen are such functions – there were several older platforms which did not have the POSIX 2008 function strnlen.)

However, some common errors are worth pointing out here. It can be helpful to look up functions at https://cplusplus.com/reference/ or https://en.cppreference.com/w/ and compare what is defined in the various standards.

More care is needed for functions such as mallinfo which are not specified by any of these standards—hopefully the man page on your system will tell you so. Searching online for such pages for various OSes (preferably at least Linux and macOS, and the FreeBSD manual pages at https://man.freebsd.org/cgi/man.cgi allow you to select many OSes) should reveal useful information but a configure script is likely to be needed to check availability and functionality.

Both the compiler and OS (via system header files, which may differ by architecture even for nominally the same OS) affect the compilability of C/C++ code. Compilers from the GCC, LLVM (clang and flang) Intel and Oracle Developer Studio suites have been used with R, and both LLVM clang and Oracle have more than one implementation of C++ headers and library. The range of possibilities makes comprehensive empirical checking impossible, and regrettably compilers are patchy at best on warning about non-standard code.

  • Mathematical functions such as sqrt are defined in C++11 for floating-point arguments: float, double, long double and possibly more. The standard specifies what happens with an argument of integer type but this is not always implemented, resulting in a report of ‘overloading ambiguity’: this was commonly seen on Solaris, but for pow also seen on macOS and other platforms using clang++.

    A not-uncommonly-seen problem is to mistakenly call floor(x/y) or ceil(x/y) for int arguments x and y. Since x/y does integer division, the result is of type int and ‘overloading ambiguity’ may be reported. Some people have (pointlessly) called floor and ceil on arguments of integer type, which may have an ‘overloading ambiguity’.

    A surprising common misuse is things like pow(10, -3): this should be the constant 1e-3. Note that there are constants such as M_SQRT2 defined via Rmath.h102 for sqrt(2.0), frequently mis-coded as sqrt(2).

  • Function fabs is defined only for floating-point types, except in C++11 and later which have overloads for std::fabs in <cmath> for integer types. Function abs is defined in C99’s <stdlib.h> for int and in C++’s <cstdlib> for integer types, overloaded in <cmath> for floating-point types. C++11 has additional overloads for std::abs in <cmath> for integer types. The effect of calling abs with a floating-point type is implementation-specific: it may truncate to an integer. For clarity and to avoid compiler warnings, use abs for integer types and fabs for double values, and when using C++ include <cmath> and use the std:: prefix.
  • It is an error (and make little sense, although has been seen) to call macros/functions isnan, isinf and isfinite for integer arguments: a few compilers give a compilation error. Function finite is obsolete, and some compilers will warn about its use103.
  • The GNU C/C++ compilers support a large number of non-portable extensions. For example, INFINITY (which is a float value in C99 and C++11), for which R provides the portable double value R_PosInf (and R_NegInf for -INFINITY). And NAN104 is just one NaN float value: for use with R, NA_REAL is often what is intended, but R_NaN is also available.

    Some (but not all) extensions are listed at https://gcc.gnu.org/onlinedocs/gcc/C-Extensions.html and https://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Extensions.html.

    Other GNU extensions which have bitten package writers are the use of non-portable characters such as ‘$’ in identifiers and use of C++ headers under ext.

  • Including C-style headers in C++ code is not portable. Including the legacy header105 math.h in C++ code may conflict with cmath which may be included by other headers. In C++11, functions like sqrt and isnan are defined for double arguments in math.h and for a range of types including double in cmath. Similar issues have been seen for stdlib.h and cstdlib. Including the C++ header first used to be a sufficient workaround but for some 2016 compilers only one could be included.
  • Be careful to include the headers which define the functions you use. Some compilers/OSes include other system headers in their headers which are not required by the standards, and so code may compile on such systems and not on others. (A prominent example is the C++ header <random> which is indirectly included by <algorithm> by g++. Another issue is the C header <time.h> which is included by other headers on Linux and Windows but not macOS.) g++ 11 often needs explicit inclusion of the C++ headers <limits> (for numeric_limits) or <exception> (for set_terminate and similar), whereas earlier versions included these in other headers. g++ 13 requires the explicit inclusion of <cstdint> for types such as uint32_t which was previously included implicitly. (For more such, see https://gcc.gnu.org/gcc-13/porting_to.html.)

    Note that malloc, calloc, realloc and free are defined by C99 in the header stdlib.h and (in the std:: namespace) by C++ header cstdlib. Some earlier implementations used a header malloc.h, but that is not portable and does not exist on macOS.

    This also applies to types such as ssize_t. The POSIX standards say that is declared in headers unistd.h and sys/types.h, and the latter is often included indirectly by other headers on some but not all systems.

    Similarly for constants: for example SIZE_MAX is defined in stdint.h alongside size_t.

  • Some headers are not portable: we have just mentioned malloc.h and often CRAN submissions attempt to use endian.h. The latter is a glibc extension: some OSes have machine/endian.h or sys/endian.h but some have neither.
  • Use #include "my.h" not #include <my.h> for headers in your package. The second form is intended for system headers and the search order for such headers is platform-dependent (and may not include the current directory). For extra safety, name headers in a way that cannot be confused with a system header so not, for example, types.h.
  • For C++ code, be careful to specify namespaces where needed. Many functions are defined by the standards to be in the std namespace, but g++ puts many such also in the C++ main namespace. One way to do so is to use declarations such as
    using std::floor;
    

    but it is usually preferable to use explicit namespace prefixes in the code.

    Examples seen in CRAN packages include

    abs acos atan bind calloc ceil div exp fabs floor fmod free log malloc
    memcpy memset pow printf qsort round sin sprintf sqrt strcmp strcpy
    strerror strlen strncmp strtol tan trunc
    

    This problem is less common than it used to be, but in 2019 LLVM clang did not have bind in the main namespace. Also seen has been type size_t defined only in the std namespace.

    NB: These functions are only guaranteed to be in the std namespace if the correct C++ header is included, e.g. <cmath> rather than <math.h>.

    If you define functions in C++ which are inspired by later standards, put them in a namespace and refer to them using the namespace. We have seen conflicts with std::make_unique from C++14 and std::byte, std::data, std::sample and std::size from C++17.

  • In C++ code
    using namespace std;
    

    is not good practice, and has caused platform-dependent errors if included before headers, especially system headers (which may be included by other headers). The best practice is to use explicit std:: prefixes for all functions declared by the C++ standard to be in that namespace. It is an error to use using namespace std before including any C++ headers, and some recent compilers will warn if this is done.

  • Some C++ compilers refuse to compile constructs such as
          if(ptr > 0) { ....}
    

    which compares a pointer to the integer 0. This could just use if(ptr) (pointer addresses cannot be negative) but if needed pointers can be tested against nullptr (C++11) or NULL.

  • Macros defined by the compiler/OS can cause problems. Identifiers starting with an underscore followed by an upper-case letter or another underscore are reserved for system macros and should not be used in portable code (including not as guards in C/C++ headers). Other macros, typically upper-case, may be defined by the compiler or system headers and can cause problems. Some of these can be avoided by defining _POSIX_C_SOURCE before including any system headers, but it is better to only use all-upper-case names which have a unique prefix such as the package name.
  • typedefs in OS headers can conflict with those in the package: examples have included ulong, index_t, single and thread. (Note that these may conflict with other uses as identifiers, e.g. defining a C++ function called single.) The POSIX standard reserves (in §2.2.2) all identifiers ending in _t.
  • Some compilers do not allow a space between -D and the macro to be defined. Similarly for -U.
  • If you use OpenMP, check carefully that you have followed the advice in the subsection on OpenMP support. In particular, any use of OpenMP in C/C++ code will need to use
    #ifdef _OPENMP
    # include <omp.h>
    #endif
    

    Any use of OpenMP functions, e.g. omp_set_num_threads, also needs to be conditioned. To avoid incessant warnings such as

    warning: ignoring #pragma omp parallel [-Wunknown-pragmas]
    

    uses of such pragmas should also be conditioned (or commented out if they are used in code in a package not enabling OpenMP on any platform).

    Do not hardcode -lgomp: not only is that specific to the GCC family of compilers, using the correct linker flag often sets up the run-time path to the library.

  • Package authors commonly assume things are part of C/C++ when they are not: the most common example is POSIX106 function strdup. The most common C library on Linux, glibc, will hide the declarations of such extensions unless a ‘feature-test macro’ is defined before (almost) any system header is included. So for strdup you need
    #define _POSIX_C_SOURCE 200809L
    ...
    #include <string.h>
    ...
    strdup call(s)
    

    where the appropriate value can be found by man strdup on Linux. (Use of strncasecmp is similar.)

    However, modes of gcc with ‘GNU EXTENSIONS’ (which are the default, either -std=gnu99 or -std=gnu11) declare enough macros to ensure that missing declarations are rarely seen.

    This applies also to constants such as M_PI and M_LN2, which are part of the X/Open standard: to use these define _XOPEN_SOURCE before including any headers, or include the R header Rmath.h.

  • Using alloca portably is tricky: it is neither an ISO C/C++ nor a POSIX function. An adequately portable preamble is
    #ifdef __GNUC__
    /* Includes GCC, clang and Intel compilers */
    # undef alloca
    # define alloca(x) __builtin_alloca((x))
    #elif defined(__sun) || defined(_AIX)
    /* this was necessary (and sufficient) for Solaris 10 and AIX 6: */
    # include <alloca.h>
    #endif
    
  • Compiler writers feel free to implement features from later standards than the one specified, so for example they may implement or warn on C++14/17/20 features when C++11 is specified. Portable code will not use such features – it can be hard to know what they are but the most common warnings are
    'register' storage class specifier is deprecated and incompatible with C++17
    
    ISO C++11 does not allow conversion from string literal to 'char *'
    

    (where conversion should be to const char *). Keyword register was not mentioned in C++98, deprecated in C++11 and removed in C++17.

    There are quite a lot of other C++98 features deprecated in C++11 and removed in C++17, and LLVM clang 9 and later warn about them (and as from version 16 they have been removed). Examples include bind1st/bind2nd (use std::bind or lambdas107) std::auto_ptr (replaced by std::unique_ptr), std::mem_fun_ref and std::ptr_fun.

  • Later versions of standards may add reserved words: for example bool, false and true became keywords in C23 and are no longer available as variable names. As noted above, C++17 uses byte, data, sample and size.

    So avoid common words and keywords from other programming languages.

  • Be careful about including C headers in C++ code. Issues include
    • Use of the register storage class specifier (see the previous but one item).
    • The C99 keyword restrict is not part of108 any C++ standard and is rejected by some C++ compilers.
    • Inclusion by such headers of C-style headers such as math.h (see above).

    The most portable way to interface to other software with a C API is to use C code (which can normally be mixed with C++ code in a package).

  • Include only what is essential in extern "C" {} blocks in C++ code. In particular it is not portable to include R headers in such blocks (although they are themselves C code, they may include C++ system headers and the public ones already enclose their declarations in such a block). And maintainers have include R headers from other headers included in such a block.
  • reinterpret_cast in C++ is not safe for pointers: for example the types may have different alignment requirements. Use memcpy to copy the contents to a fresh variable of the destination type.
  • Avoid platform-specific code if at all possible, but if you need to test for a platform ensure that all platforms are covered. For example, __unix__ is not defined on all Unix-alikes, in particular not on macOS. A reasonably portable way to condition code for a Unix-alike is
    #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
    #endif
    

    but

    #ifdef _WIN32
    // Windows-specific code
    # if defined(_M_ARM64) || defined(__aarch64__)
    // for ARM
    # else
    // for Intel
    # endif
    #else
    // Unix-alike code
    #endif
    

    would be better. For a Unix-alike it is much better to use configure to test for the functionality needed than make assumptions about OSes (and people all too frequently forget R is used on platforms other than Linux, Windows and macOS — and some forget macOS).

  • Headers in subdirectories are often not portable. For C++, this includes bits/, tr1/ and tr2/, none of which exist on macOS (and ext/ exists there but with different content from g++-based platforms). Header bits/stdc++.h is both not portable and not recommended for end-user code even on platforms which include it.
  • Be careful if using malloc or calloc. First, their return value must always be checked to see if the allocation succeeded – it is almost always easier to use R’s R_Calloc, which does check. Second, the first argument is of type size_t109 and some recent compilers warn about passing int (signed) arguments (which could get promoted to ridiculously large values).
  • For C code, consider using the flag -Wstrict-prototypes which is supported by gcc and LLVM and Apple clang. This has found quite a number of errors where functions have been declared without arguments and is likely to become the default in future compilers. (It already is for Apple clang and for LLVM clang in C23 mode.) Note that using f() for a function without any parameters was deprecated in C99 and C11, but it became non-deprecated in C23. However, f(void) is supported by all standards and avoids any uncertainty.

    LLVM clang has a separate warning -Wdeprecated-non-prototype which is enabled by -Wstrict-prototypes. This warns on K&R-style usage, which will not be accepted in C23.

  • Several C entry points are warned against in their man pages on most systems, often in very strong terms such as ’Do not use these functions’. macOS has started to warn110 if these are used for sprintf, vsprintf, gets, mktemp, tempmam and tmpnam. It is highly recommended that you use safer alternatives (on any platform) but the warning can be avoided by defining ‘_POSIX_C_SOURCE’ to for example ‘200809L’ before including the (C or C++) header which defines them. (However, this may hide other extensions.)
  • Compilers may interpret comments in source code, so it is necessary to remove any intended for a compiler to interpret. The main example has been comments for Visual Fortran (as the Intel Fortran compiler has been known on Windows111) like
    !DEC$ ATTRIBUTES DLLEXPORT,C,REFERENCE,ALIAS:'kdenestmlcvb' :: kdenestmlcvb
    

    which are interpreted by Intel Fortran on all platforms (and are inappropriate for use with R on Windows). gfortran has similar forms starting with !GCC$.

  • The C++ new operator takes argument std::size_t size, which is unsigned. Using a signed integer type such as int may lead to compiler warnings such as
    warning: argument 1 value '18446744073709551615' exceeds maximum object
      size 9223372036854775807  [-Walloc-size-larger-than=]
    

    (especially if LTO is used). So don’t do that!

  • Some authors feel the need to print (using Rprintf or similar) vector lengths or indices which are of type R_xlen_t. That may be a 32-bit or (most commonly) 64-bit type but which integer type it is mapped to is platform-specific. The safest way is to cast the length to double and use a double format. So one could use something like
        SEXP Robj; R_xlen_t nelem;
        Rf_error("Actual: %0.f; Expected %0.f\n", (double) XLENGTH(Robj), (double) nelem);
    

    (This could print to full precision, lengths well beyond the address space limits of current OSes, let alone practical limits.)

    If you do want to use an integer format, be aware that R_xlen_t is implemented by the int, long or long long type on current platforms and even on 64-bit ones need not be the same type as int64_t. So the values will need to be cast to the type assumed by the format (and %lld was not supported on Windows until R 4.2.0).

  • Variadic macros in C or C++ only portably allow a non-zero number of arguments, although some compilers have allowed zero, often with a warning. The latter was standardized in C++20 using the __VA_OPT__ macro. C23 also allows zero arguments in a similar way.

Some additional information for C++ is available at https://journal.r-project.org/archive/2011-2/RJournal_2011-2_Plummer.pdf by Martyn Plummer.


1.6.4.1 Common symbols

Most OSes (including all those commonly used for R) have the concept of ‘tentative definitions’ where global C variables are defined without an initializer. Traditionally the linker resolves all tentative definitions of the same variable in different object files to the same object, or to a non-tentative definition. However, gcc 10112 and LLVM clang 11113 changed their default so that tentative definitions cannot be merged and the linker will give an error if the same variable is defined in more than one object file. To avoid this, all but one of the C source files should declare the variable extern — which means that any such variables included in header files need to be declared extern. A commonly used idiom (including by R itself) is to define all global variables as extern in a header, say globals.h (and nowhere else), and then in one (and one only) source file use

#define extern
# include "globals.h"
#undef extern

A cleaner approach is not to have global variables at all, but to place in a single file common variables (declared static) followed by all the functions which make use of them: this may result in more efficient code.

The ‘modern’ behaviour can be seen114 by using compiler flag -fno-common as part of ‘CFLAGS’ in earlier versions of gcc and clang.

-fno-common is said to be particularly beneficial for ARM CPUs.

This is not pertinent to C++ which does not permit tentative definitions.


1.6.4.2 C++17 issues

R 4.3.0 and later default to C++17 when compiling C++, and that finally removed many C++98 features which were deprecated as long ago as C++11. Compiler/runtime authors have been slow to remove these, but LLVM clang with its libc++ runtime library finally started to do so in 2023 – some others warn but some do not.

The principal offender is the ‘Boost’ collection of C++ headers and libraries. There are two little-documented ways to work around aspects of its outdated code. One is to add

-D_HAS_AUTO_PTR_ETC=0

to PKG_CPPLAGS in src/Makevars, src/Makevars.win and src/Makevars.ucrt. This covers the removal of

std::auto_ptr
std::unary_function
std::binary_function
std::random_shuffle
std::binder1st
std::binder2nd

with most issues seen with code that includes boost/functional.hpp, usually indirectly.

A rarer issue is the use of illegal values for enum types, usually negative ones such as

BOOST_MPL_AUX_STATIC_CAST(AUX_WRAPPER_VALUE_TYPE, (value - 1));

in boost/mpl/aux_/integral_wrapper.hpp. Adding

 -Wno-error=enum-constexpr-conversion

to PKG_CXXFLAGS will allow this, but that flag is only accepted by recent versions of LLVM clang (and will not be in future) so needs a configure test.

Pre=built versions of current clang/libc++ are usually available from https://github.com/llvm/llvm-project/releases for a wide range of platforms (but the Windows builds there are not compatible with Rtools and the macOS ones are unsigned). To select libc++ add -stdlib=libc++ to CXX, for example by having

CXX="/path/to/clang/clang++ -std=gnu++17 -stdlib=libc++"

in ~/.R/Makevars.

Another build for Windows which may be sufficiently compatible with Rtools can be found at https://github.com/mstorsjo/llvm-mingw: this uses libc++.


1.6.5 Portable Fortran code

For many years almost all known R platforms used gfortran as their Fortran compiler, but now there are LLVM and ‘classic’ flang and the Intel compilers ifort115 and ifx are now free-of-change.

There is still a lot of Fortran code in CRAN packages which predates Fortran 77. Modern Fortran compilers are being written to target a minimum standard of Fortran 2018. and it is desirable that Fortran code in packages complies with that standard. For gfortran this can be checked by adding -std=f2018 to FFLAGS. The most commonly seen issues are

  • The use of DFLOAT, which was superseded by DBLE in Fortran 77. Also, use of DCMPLX, DCONJG, DIMAG and similar.
  • Use of what gfortran calls ‘Fortran 2018 deleted features’, although most were ‘deleted’ in earlier standards: those itemized here were deleted in Fortran 2008. (In the Fortran standards ‘deleted’ means features that compilers are not required to implement.) These include
    • Arithmetic IF statements.
    • DO loops which are not terminated with a END DO or CONTINUE statement. (Unlabelled DO loops terminated by END DO are preferred for readability.)
    • Labelled DO loops sharing a terminating CONTINUE statement.
  • The use of GNU Fortran extensions. Some are listed at https://gcc.gnu.org/onlinedocs/gfortran/Extensions-implemented-in-GNU-Fortran.html. Others which have caused problems include etime, getpid, isnan116 and sizeof.

    One that frequently catches package writers is that it allows out-of-order declarations: in standard-conformant Fortran variables must be declared (explicitly or implicitly) before use in other declarations such as dimensions.

Unfortunately this flags extensions such as DOUBLE COMPLEX and COMPLEX*16. R has tested that DOUBLE COMPLEX works and so is preferred to COMPLEX*16. (One can also use something like COMPLEX(KIND=KIND(0.0D0)).)

GNU Fortran 10 and later give a compilation error for the previously widespread practice of passing a Fortran array element where an array is expected, or a scalar instead of a length-one array. See https://gcc.gnu.org/gcc-10/porting_to.html. As do the Intel Fortran compilers, and they can be stricter.

The use of IMPLICIT NONE is highly recommended – Intel compilers with -warn will warn on variables without an explicit type.

Common non-portable constructions include

  • The use of Fortran types such as REAL(KIND=8) is very far from portable. According to the standards this merely enumerates different supported types, so DOUBLE PRECISION might be REAL(KIND=3) (and is on an actual compiler). Even if for a particular compiler the value indicates the size in bytes, which values are supported is platform-specific — for example gfortran supports values of 4 and 8 on all current platforms and 10 and 16 on a few (but not for example on all ‘arm’ CPUs).

    The same applies to INTEGER(KIND=4) and COMPLEX(KIND=16).

    Many uses of integer and real variable in Fortran code in packages will interwork with C (for example .Fortran is written in C), and R has checked that INTEGER and DOUBLE PRECISION correspond to the C types int and double. To make this explicit, from Fortran 2003 one can use the named constants c_int, c_double and c_double_complex from module iso_c_binding.

  • The Intel compilers only recognize the extensions .f (fixed-form) and .f90 (free-form) and not .f95. R CMD INSTALL works around this for packages without a src/Makefile.
  • Use of extensions .F and .F90 to indicate source code to be preprocessed: the preprocessor used is compiler-specific and may or may not be cpp. Compilers may even preprocess files with extension .f or .f90 (Intel does).
  • Fixed form Fortran (with extension .f) should only use 72 columns, and free-form at most 132 columns. This includes trailing comments. Over-long lines may be silently truncated or give a warning.
  • Tabs are not part of the Fortran character set: compilers tend to accept them but how they are interpreted is compiler-specific.
  • Fortran-66-style Hollerith constants.

As well as ‘deleted features’, Fortran standards have ‘obsolescent features’. These are similar to ‘deprecated’ in other languages, but the Fortran standards committee has said it will only move them to ‘deleted’ status when they are no longer much used. These include

  • ENTTRY statements.
  • FORALL statements.
  • Labelled DO statements.
  • COMMON and EQUIVALENCE statements, and BLOCK DATA units.
  • Computed GOTO statements, replaced by SELECT CASE.
  • Statement functions.
  • DATA statements after executable statements.
  • Specific (rather than generic) names for intrinsic functions.

gfortran with option -std=f2018 will warn about these: R will report only in the installation log.


1.6.6 Binary distribution

If you want to distribute a binary version of a package on Windows or macOS, there are further checks you need to do to check it is portable: it is all too easy to depend on external software on your own machine that other users will not have.

For Windows, check what other DLLs your package’s DLL depends on (‘imports’ from in the DLL tools’ parlance). A convenient GUI-based tool to do so is ‘Dependency Walker’ (https://www.dependencywalker.com/) for both 32-bit and 64-bit DLLs – note that this will report as missing links to R’s own DLLs such as R.dll and Rblas.dll. The command-line tool objdump in the appropriate toolchain will also reveal what DLLs are imported from. If you use a toolchain other than one provided by the R developers or use your own makefiles, watch out in particular for dependencies on the toolchain’s runtime DLLs such as libgfortran, libstdc++ and libgcc_s.

For macOS, using R CMD otool -L on the package’s shared object(s) in the libs directory will show what they depend on: watch for any dependencies in /usr/local/lib or /usr/local/gfortran/lib, notably libgfortran.?.dylib and libquadmath.0.dylib. (For ways to fix these, see Building binary packages in R Installation and Administration.)

Many people (including the CRAN package repository) will not accept source packages containing binary files as the latter are a security risk. If you want to distribute a source package which needs external software on Windows or macOS, options include

  • To arrange for installation of the package to download the additional software from a URL, as e.g. package Cairo used to.
  • To negotiate with Tomas Kalibera to include Windows software in Rtools or with Simon Urbanek to include macOS software in his ‘recipes’ system.
  • (For CRAN.) To negotiate with Uwe Ligges to host the additional components on WinBuilder, and write a configure.win file to install them.

Be aware that license requirements you may require you to supply the sources for the additional components (and will if your package has a GPL-like license).


1.7 Diagnostic messages

Diagnostic messages can be made available for translation, so it is important to write them in a consistent style. Using the tools described in the next section to extract all the messages can give a useful overview of your consistency (or lack of it). Some guidelines follow.

  • Messages are sentence fragments, and not viewed in isolation. So it is conventional not to capitalize the first word and not to end with a period (or other punctuation).
  • Try not to split up messages into small pieces. In C error messages use a single format string containing all English words in the messages.

    In R error messages do not construct a message with paste (such messages will not be translated) but via multiple arguments to stop or warning, or via gettextf.

  • Do not use colloquialisms such as “can’t” and “don’t”.
  • Conventionally single quotation marks are used for quotations such as
    'ord' must be a positive integer, at most the number of knots
    

    and double quotation marks when referring to an R character string or a class, such as

    'format' must be "normal" or "short" - using "normal"
    

    Since ASCII does not contain directional quotation marks, it is best to use ‘'’ and let the translator (including automatic translation) use directional quotations where available. The range of quotation styles is immense: unfortunately we cannot reproduce them in a portable texinfo document. But as a taster, some languages use ‘up’ and ‘down’ (comma) quotes rather than left or right quotes, and some use guillemets (and some use what Adobe calls ‘guillemotleft’ to start and others use it to end).

    In R messages it is also possible to use sQuote or dQuote as in

            stop(gettextf("object must be of class %s or %s",
                          dQuote("manova"), dQuote("maov")),
                 domain = NA)
    
  • Occasionally messages need to be singular or plural (and in other languages there may be no such concept or several plural forms – Slovenian has four). So avoid constructions such as was once used in library
    if((length(nopkgs) > 0) && !missing(lib.loc)) {
        if(length(nopkgs) > 1)
            warning("libraries ",
                    paste(sQuote(nopkgs), collapse = ", "),
                    " contain no packages")
        else
            warning("library ", paste(sQuote(nopkgs)),
                    " contains no package")
    }
    

    and was replaced by

    if((length(nopkgs) > 0) && !missing(lib.loc)) {
        pkglist <- paste(sQuote(nopkgs), collapse = ", ")
        msg <- sprintf(ngettext(length(nopkgs),
                                "library %s contains no packages",
                                "libraries %s contain no packages",
                                domain = "R-base"),
                       pkglist)
        warning(msg, domain=NA)
    }
    

    Note that it is much better to have complete clauses as here, since in another language one might need to say ‘There is no package in library %s’ or ‘There are no packages in libraries %s’.


1.8 Internationalization

There are mechanisms to translate the R- and C-level error and warning messages. There are only available if R is compiled with NLS support (which is requested by configure option --enable-nls, the default).

The procedures make use of msgfmt and xgettext which are part of GNU gettext and this will need to be installed: ‘x86_64’ Windows users can find pre-compiled binaries at https://www.stats.ox.ac.uk/pub/Rtools/goodies/gettext-tools.zip.


1.8.1 C-level messages

The process of enabling translations is

  • In a header file that will be included in all the C (or C++ or Objective C/C++) files containing messages that should be translated, declare
    #include <R.h>  /* to include Rconfig.h */
    
    #ifdef ENABLE_NLS
    #include <libintl.h>
    #define _(String) dgettext ("pkg", String)
    /* replace pkg as appropriate */
    #else
    #define _(String) (String)
    #endif
    
  • For each message that should be translated, wrap it in _(...), for example
    error(_("'ord' must be a positive integer"));
    

    If you want to use different messages for singular and plural forms, you need to add

    #ifndef ENABLE_NLS
    #define dngettext(pkg, String, StringP, N) (N == 1 ? String : StringP)
    #endif
    

    and mark strings by

    dngettext("pkg", <singular string>, <plural string>, n)
    
  • In the package’s src directory run
    xgettext --keyword=_ -o pkg.pot *.c
    

The file src/pkg.pot is the template file, and conventionally this is shipped as po/pkg.pot.


1.8.2 R messages

Mechanisms are also available to support the automatic translation of R stop, warning and message messages. They make use of message catalogs in the same way as C-level messages, but using domain R-pkg rather than pkg. Translation of character strings inside stop, warning and message calls is automatically enabled, as well as other messages enclosed in calls to gettext or gettextf. (To suppress this, use argument domain=NA.)

Tools to prepare the R-pkg.pot file are provided in package tools: xgettext2pot will prepare a file from all strings occurring inside gettext/gettextf, stop, warning and message calls. Some of these are likely to be spurious and so the file is likely to need manual editing. xgettext extracts the actual calls and so is more useful when tidying up error messages.

The R function ngettext provides an interface to the C function of the same name: see example in the previous section. It is safest to use domain="R-pkg" explicitly in calls to ngettext, and necessary for earlier versions of R unless they are calls directly from a function in the package.


1.8.3 Preparing translations

Once the template files have been created, translations can be made. Conventional translations have file extension .po and are placed in the po subdirectory of the package with a name that is either ‘ll.po’ or ‘R-ll.po’ for translations of the C and R messages respectively to language with code ‘ll’.

See Localization of messages in R Installation and Administration for details of language codes.

There is an R function, update_pkg_po in package tools, to automate much of the maintenance of message translations. See its help for what it does in detail.

If this is called on a package with no existing translations, it creates the directory pkgdir/po, creates a template file of R messages, pkgdir/po/R-pkg.pot, within it, creates the ‘en@quot’ translation and installs that. (The ‘en@quot’ pseudo-language interprets quotes in their directional forms in suitable (e.g. UTF-8) locales.)

If the package has C source files in its src directory that are marked for translation, use

touch pkgdir/po/pkg.pot

to create a dummy template file, then call update_pkg_po again (this can also be done before it is called for the first time).

When translations to new languages are added in the pkgdir/po directory, running the same command will check and then install the translations.

If the package sources are updated, the same command will update the template files, merge the changes into the translation .po files and then installed the updated translations. You will often see that merging marks translations as ‘fuzzy’ and this is reported in the coverage statistics. As fuzzy translations are not used, this is an indication that the translation files need human attention.

The merged translations are run through tools::checkPofile to check that C-style formats are used correctly: if not the mismatches are reported and the broken translations are not installed.

This function needs the GNU gettext-tools installed and on the path: see its help page.


1.9 CITATION files

An installed file named CITATION will be used by the citation() function. (It should be in the inst subdirectory of the package sources.)

The CITATION file is parsed as R code (in the package’s declared encoding, or in ASCII if none is declared). It will contain calls to function bibentry. Here is that for nlme:

## R package reference generated from DESCRIPTION metadata
citation(auto = meta)

## NLME book
bibentry(bibtype = "Book",
         title = "Mixed-Effects Models in S and S-PLUS",
         author = c(person(c("José", "C."), "Pinheiro"),
                    person(c("Douglas", "M."), "Bates")),
         year = "2000", publisher = "Springer", address = "New York",
         doi = "10.1007/b98882")

Note how the first call auto-generates citation information from object meta, a parsed version of the DESCRIPTION file – it is tempting to hardcode such information, but it normally then gets outdated. How the first entry would look like as a bibentry call can be seen from print(citation("pkgname", auto = TRUE), style = "R") for any installed package. Auto-generated information is returned by default if no CITATION file is present.

See ?bibentry for further details of the information which can be provided. In case a bibentry contains LaTeX markup (e.g., for accented characters or mathematical symbols), it may be necessary to provide a text representation to be used for printing via the textVersion argument to bibentry. E.g., earlier versions of nlme additionally used something like

         textVersion =
         paste0("Jose Pinheiro, Douglas Bates, Saikat DebRoy, ",
                "Deepayan Sarkar and the R Core Team (",
                sub("-.*", "", meta$Date),
                "). nlme: Linear and Nonlinear Mixed Effects Models. ",
                sprintf("R package version %s", meta$Version), ".")

The CITATION file should itself produce no output when source-d.

It is desirable (and essential for CRAN) that the CITATION file does not contain calls to functions such as packageDescription which assume the package is installed in a library tree on the package search path.


1.10 Package types

The DESCRIPTION file has an optional field Type which if missing is assumed to be ‘Package’, the sort of extension discussed so far in this chapter. Currently one other type is recognized; there used also to be a ‘Translation’ type.


1.10.1 Frontend

This is a rather general mechanism, designed for adding new front-ends such as the former gnomeGUI package (see the Archive area on CRAN). If a configure file is found in the top-level directory of the package it is executed, and then if a Makefile is found (often generated by configure), make is called. If R CMD INSTALL --clean is used make clean is called. No other action is taken.

R CMD build can package up this type of extension, but R CMD check will check the type and skip it.

Many packages of this type need write permission for the R installation directory.


1.11 Services

Several members of the R project have set up services to assist those writing R packages, particularly those intended for public distribution.

win-builder.r-project.org offers the automated preparation of (‘x86_64’) Windows binaries from well-tested source packages.

R-Forge (R-Forge.r-project.org) and RForge (www.rforge.net) are similar services with similar names. Both provide source-code management through SVN, daily building and checking, mailing lists and a repository that can be accessed via install.packages (they can be selected by setRepositories and the GUI menus that use it). Package developers have the opportunity to present their work on the basis of project websites or news announcements. Mailing lists, forums or wikis provide useRs with convenient instruments for discussions and for exchanging information between developers and/or interested useRs.


2 Writing R documentation files


2.1 Rd format

R objects are documented in files written in “R documentation” (Rd) format, a simple markup language much of which closely resembles (La)TeX, which can be processed into a variety of formats, including LaTeX, HTML and plain text. The translation is carried out by functions in the tools package called by the script Rdconv in R_HOME/bin and by the installation scripts for packages.

The R distribution contains more than 1400 such files which can be found in the src/library/pkg/man directories of the R source tree, where pkg stands for one of the standard packages which are included in the R distribution.

As an example, let us look at a simplified version of src/library/base/man/load.Rd which documents the R function load.

% File src/library/base/man/load.Rd
\name{load}
\alias{load}
\title{Reload Saved Datasets}
\description{
  Reload datasets written with the function \code{save}.
}
\usage{
load(file, envir = parent.frame(), verbose = FALSE)
}
\arguments{
  \item{file}{a (readable binary-mode) \link{connection}
    or a character string giving the name of the file to load
    (when \link{tilde expansion} is done).}
  \item{envir}{the environment where the data should be loaded.}
  \item{verbose}{should item names be printed during loading?}
}
\value{
  A character vector of the names of objects created, invisibly.
}
\seealso{
  \code{\link{save}}.
}
\examples{
## save all data
save(list = ls(all.names = TRUE), file = "all.RData")

## restore the saved values to the current environment
load("all.RData")
}
\keyword{file}

An Rd file consists of three parts. The header gives basic information about the name of the file, the topics documented, a title, a short textual description and R usage information for the objects documented. The body gives further information (for example, on the function’s arguments and return value, as in the above example). Finally, there is an optional footer with keyword information. The header is mandatory.

Information is given within a series of sections with standard names (and user-defined sections are also allowed). Unless otherwise specified117 these should occur only once in an Rd file (in any order), and the processing software will retain only the first occurrence of a standard section in the file, with a warning.

See “Guidelines for Rd files” for guidelines for writing documentation in Rd format which should be useful for package writers. The R generic function prompt is used to construct a bare-bones Rd file ready for manual editing. Methods are defined for documenting functions (which fill in the proper function and argument names) and data frames. There are also functions promptData, promptPackage, promptClass, and promptMethods for other types of Rd files.

The general syntax of Rd files is summarized below. For a detailed technical discussion of current Rd syntax, see “Parsing Rd files”.

Rd files consist of four types of text input. The most common is LaTeX-like, with the backslash used as a prefix on markup (e.g. \alias), and braces used to indicate arguments (e.g. {load}). The least common type of text is ‘verbatim’ text, where no markup other than the comment marker (%) is processed. There is also a rare variant of ‘verbatim’ text (used in \eqn, \deqn, \figure, and \newcommand) where comment markers need not be escaped. The final type is R-like, intended for R code, but allowing some embedded macros. Quoted strings within R-like text are handled specially: regular character escapes such as \n may be entered as-is. Only markup starting with \l (e.g. \link) or \v (e.g. \var) will be recognized within quoted strings. The rarely used vertical tab \v must be entered as \\v.

Each macro defines the input type for its argument. For example, the file initially uses LaTeX-like syntax, and this is also used in the \description section, but the \usage section uses R-like syntax, and the \alias macro uses ‘verbatim’ syntax. Comments run from a percent symbol % to the end of the line in all types of text except the rare ‘verbatim’ variant (as on the first line of the load example).

Because backslashes, braces and percent symbols have special meaning, to enter them into text sometimes requires escapes using a backslash. In general balanced braces do not need to be escaped, but percent symbols always do, except in the ‘verbatim’ variant. For the complete list of macros and rules for escapes, see “Parsing Rd files”.


2.1.1 Documenting functions

The basic markup commands used for documenting R objects (in particular, functions) are given in this subsection.

\name{name}

name typically118 is the basename of the Rd file containing the documentation. It is the “name” of the Rd object represented by the file and has to be unique in a package. To avoid problems with indexing the package manual, it may not contain ‘!’ ‘|’ nor ‘@’. (LaTeX special characters are allowed, but may not be collated correctly in the index.) There can only be one \name entry in a file, and it must not contain any markup and should only contain printable ASCII characters. Entries in the package manual will be in alphabetic119 order of the \name entries.

\alias{topic}

The \alias sections specify all “topics” the file documents. This information is collected into index data bases for lookup by the on-line (plain text and HTML) help systems. The topic can contain spaces, but (for historical reasons) leading and trailing spaces will be stripped. Percent and left brace need to be escaped by a backslash.

There may be several \alias entries. Quite often it is convenient to document several R objects in one file. For example, file Normal.Rd documents the density, distribution function, quantile function and generation of random variates for the normal distribution, and hence starts with

\name{Normal}
\alias{Normal}
\alias{dnorm}
\alias{pnorm}
\alias{qnorm}
\alias{rnorm}

Also, it is often convenient to have several different ways to refer to an R object, and an \alias does not need to be the name of an object.

Note that the \name is not necessarily a topic documented, and if so desired it needs to have an explicit \alias entry (as in this example).

\title{Title}

Title information for the Rd file. This should be capitalized and not end in a period; try to limit its length to at most 65 characters for widest compatibility.

Markup is supported in the text, but use of characters other than English text and punctuation (e.g., ‘<’) may limit portability.

There must be one (and only one) \title section in a help file.

\description{…}

A short description of what the function(s) do(es) (one paragraph, a few lines only). (If a description is too long and cannot easily be shortened, the file probably tries to document too much at once.) This is mandatory except for package-overview files.

\usage{fun(arg1, arg2, …)}

One or more lines showing the synopsis of the function(s) and variables documented in the file. These are set in typewriter font. This is an R-like command.

The usage information specified should match the function definition exactly (such that automatic checking for consistency between code and documentation is possible).

To indicate that a function can be used in several different ways, depending on the named arguments specified, use section \details. E.g., abline.Rd contains

\details{
  Typical usages are
\preformatted{abline(a, b, ...)
......
}

Use \method{generic}{class} to indicate the name of an S3 method for the generic function generic for objects inheriting from class "class". In the printed versions, this will come out as generic (reflecting the understanding that methods should not be invoked directly but via method dispatch), but codoc() and other QC tools always have access to the full name.

For example, print.ts.Rd contains

\usage{
\method{print}{ts}(x, calendar, \dots)
}

which will print as

Usage:

     ## S3 method for class 'ts':
     print(x, calendar, ...)

Usage for replacement functions should be given in the style of dim(x) <- value rather than explicitly indicating the name of the replacement function ("dim<-" in the above). Similarly, one can use \method{generic}{class}(arglist) <- value to indicate the usage of an S3 replacement method for the generic replacement function "generic<-" for objects inheriting from class "class".

Usage for S3 methods for extracting or replacing parts of an object, S3 methods for members of the Ops group, and S3 methods for user-defined (binary) infix operators (‘%xxx%’) follows the above rules, using the appropriate function names. E.g., Extract.factor.Rd contains

\usage{
\method{[}{factor}(x, \dots, drop = FALSE)
\method{[[}{factor}(x, \dots)
\method{[}{factor}(x, \dots) <- value
}

which will print as

Usage:

     ## S3 method for class 'factor':
     x[..., drop = FALSE]
     ## S3 method for class 'factor':
     x[[...]]
     ## S3 replacement method for class 'factor':
     x[...] <- value

\S3method is accepted as an alternative to \method.

\arguments{…}

Description of the function’s arguments, using an entry of the form

\item{arg_i}{Description of arg_i.}

for each element of the argument list. (Note that there is no whitespace between the three parts of the entry.) Arguments can also be described jointly by separating their names with commas (and optional whitespace) in the \item label. There may be optional text outside the \item entries, for example to give general information about groups of parameters.

\details{…}

A detailed if possible precise description of the functionality provided, extending the basic information in the \description slot.

\value{…}

Description of the function’s return value.

If a list with multiple values is returned, you can use entries of the form

\item{comp_i}{Description of comp_i.}

for each component of the list returned. There may be optional text outside the \item entries (see for example the joint help for rle and inverse.rle, or the sets of items in l10n_info). Note that \value is implicitly a \describe environment, so that environment should not be used for listing components, just individual \item{}{} entries.120

\references{…}

A section with references to the literature. Use \url{} or \href{}{} for web pointers, and \doi{} for DOIs (this needs R >= 3.3, see User-defined macros for more info).

\note{...}

Use this for a special note you want to have pointed out. Multiple \note sections are allowed, but might be confusing to the end users.

For example, pie.Rd contains

\note{
  Pie charts are a very bad way of displaying information.
  The eye is good at judging linear measures and bad at
  judging relative areas.
  ......
}
\author{…}

Information about the author(s) of the Rd file. Use \email{} without extra delimiters (such as ‘( )’ or ‘< >’) to specify email addresses, or \url{} or \href{}{} for web pointers.

\seealso{…}

Pointers to related R objects, using \code{\link{...}} to refer to them (\code is the correct markup for R object names, and \link produces hyperlinks in output formats which support this. See Marking text, and Cross-references).

\examples{…}

Examples of how to use the function. Code in this section is set in typewriter font without reformatting and is run by example() unless marked otherwise (see below).

Examples are not only useful for documentation purposes, but also provide test code used for diagnostic checking of R code. By default, text inside \examples{} will be displayed in the output of the help page and run by example() and by R CMD check. You can use \dontrun{} for text that should only be shown, but not run, and \dontshow{} for extra commands for testing that should not be shown to users, but will be run by example(). (Previously this was called \testonly, and that is still accepted.)

Text inside \dontrun{} is ‘verbatim’, but the other parts of the \examples section are R-like text.

For example,

x <- runif(10)       # Shown and run.
\dontrun{plot(x)}    # Only shown.
\dontshow{log(x)}    # Only run.

Thus, example code not included in \dontrun must be executable! In addition, it should not use any system-specific features or require special facilities (such as Internet access or write permission to specific directories). Text included in \dontrun is indicated by comments in the processed help files: it need not be valid R code but the escapes must still be used for %, \ and unpaired braces as in other ‘verbatim’ text.

Example code must be capable of being run by example, which uses source. This means that it should not access stdin, e.g. to scan() data from the example file.

Data needed for making the examples executable can be obtained by random number generation (for example, x <- rnorm(100)), or by using standard data sets listed by data() (see ?data for more info).

Finally, there is \donttest, used (at the beginning of a separate line) to mark code that should be run by example() but not by R CMD check (by default: the option --run-donttest can be used). This should be needed only occasionally but can be used for code which might fail in circumstances that are hard to test for, for example in some locales. (Use e.g. capabilities() or nzchar(Sys.which("someprogram")) to test for features needed in the examples wherever possible, and you can also use try() or tryCatch(). Use interactive() to condition examples which need someone to interact with.) Note that code included in \donttest must be correct R code, and any packages used should be declared in the DESCRIPTION file. It is good practice to include a comment in the \donttest section explaining why it is needed.

Output from code marked with \dontdiff (requires R >= 4.4.0) or between comment lines

## IGNORE_RDIFF_BEGIN
## IGNORE_RDIFF_END

is ignored when comparing check output to reference output (a pkg-Ex.Rout.save file). The comment-based markup can also be used for scripts under tests.

\keyword{key}

There can be zero or more \keyword sections per file. Each \keyword section should specify a single keyword, preferably one of the standard keywords as listed in file KEYWORDS in the R documentation directory (default R_HOME/doc). Use e.g. RShowDoc("KEYWORDS") to inspect the standard keywords from within R. There can be more than one \keyword entry if the R object being documented falls into more than one category, or none.

Do strongly consider using \concept (see Indices) instead of \keyword if you are about to use more than very few non-standard keywords.

The special keyword ‘internal’ marks a page of internal topics (typically, objects that are not part of the package’s API). If the help page for topic foo has keyword ‘internal’, then help(foo) gives this help page, but foo is excluded from several topic indices, including the alphabetical list of topics in the HTML help system.

help.search() can search by keyword, including user-defined values: however the ‘Search Engine & Keywords’ HTML page accessed via help.start() provides single-click access only to a pre-defined list of keywords.


2.1.2 Documenting data sets

The structure of Rd files which document R data sets is slightly different. Sections such as \arguments and \value are not needed but the format and source of the data should be explained.

As an example, let us look at src/library/datasets/man/rivers.Rd which documents the standard R data set rivers.

\name{rivers}
\docType{data}
\alias{rivers}
\title{Lengths of Major North American Rivers}
\description{
  This data set gives the lengths (in miles) of 141 \dQuote{major}
  rivers in North America, as compiled by the US Geological
  Survey.
}
\usage{rivers}
\format{A vector containing 141 observations.}
\source{World Almanac and Book of Facts, 1975, page 406.}
\references{
  McNeil, D. R. (1977) \emph{Interactive Data Analysis}.
  New York: Wiley.
}
\keyword{datasets}

This uses the following additional markup commands.

\docType{…}

Indicates the “type” of the documentation object. Always ‘data’ for data sets, and ‘package’ for pkg-package.Rd overview files. Documentation for S4 methods and classes uses ‘methods’ (from promptMethods()) and ‘class’ (from promptClass()).

\format{…}

A description of the format of the data set (as a vector, matrix, data frame, time series, …). For matrices and data frames this should give a description of each column, preferably as a list or table. See Lists and tables, for more information.

\source{…}

Details of the original source (a reference or URL, see Specifying URLs). In addition, section \references could give secondary sources and usages.

Note also that when documenting data set bar,

  • The \usage entry is always bar or (for packages which do not use lazy-loading of data) data(bar). (In particular, only document a single data object per Rd file.)
  • The \keyword entry should always be ‘datasets’.

If bar is a data frame, documenting it as a data set can be initiated via prompt(bar). Otherwise, the promptData function may be used.


2.1.3 Documenting S4 classes and methods

There are special ways to use the ‘?’ operator, namely ‘class?topic’ and ‘methods?topic’, to access documentation for S4 classes and methods, respectively. This mechanism depends on conventions for the topic names used in \alias entries. The topic names for S4 classes and methods respectively are of the form

class-class
generic,signature_list-method

where signature_list contains the names of the classes in the signature of the method (without quotes) separated by ‘,’ (without whitespace), with ‘ANY’ used for arguments without an explicit specification. E.g., ‘genericFunction-class’ is the topic name for documentation for the S4 class "genericFunction", and ‘coerce,ANY,NULL-method’ is the topic name for documentation for the S4 method for coerce for signature c("ANY", "NULL").

Skeletons of documentation for S4 classes and methods can be generated by using the functions promptClass() and promptMethods() from package methods. If it is necessary or desired to provide an explicit function declaration (in a \usage section) for an S4 method (e.g., if it has “surprising arguments” to be mentioned explicitly), one can use the special markup

\S4method{generic}{signature_list}(argument_list)

(e.g., ‘\S4method{coerce}{ANY,NULL}(from, to)’).

To make full use of the potential of the on-line documentation system, all user-visible S4 classes and methods in a package should at least have a suitable \alias entry in one of the package’s Rd files. If a package has methods for a function defined originally somewhere else, and does not change the underlying default method for the function, the package is responsible for documenting the methods it creates, but not for the function itself or the default method.

An S4 replacement method is documented in the same way as an S3 one: see the description of \method in Documenting functions.

See help("Documentation", package = "methods") for more information on using and creating on-line documentation for S4 classes and methods.


2.1.4 Documenting packages

Packages may have an overview help page with an \alias pkgname-package, e.g. ‘utils-package’ for the utils package, when package?pkgname will open that help page. If a topic named pkgname does not exist in another Rd file, it is helpful to use this as an additional \alias.

Skeletons of documentation for a package can be generated using the function promptPackage(). If the final = TRUE argument is used, then the Rd file will be generated in final form, containing only basic information from the DESCRIPTION file. Otherwise (the default) comments will be inserted giving suggestions for content.

Apart from the mandatory \name and \title and the pkgname-package alias, the only requirement for the package overview page is that it include a \docType{package} statement. All other content is optional. We suggest that it should be a short overview, to give a reader unfamiliar with the package enough information to get started. More extensive documentation is better placed into a package vignette (see Writing package vignettes) and referenced from this page, or into individual man pages for the functions, datasets, or classes.


2.2 Sectioning

To begin a new paragraph or leave a blank line in an example, just insert an empty line (as in (La)TeX). To break a line, use \cr.

In addition to the predefined sections (such as \description{}, \value{}, etc.), you can “define” arbitrary ones by \section{section_title}{…}. For example

\section{Warning}{
  You must not call this function unless ...
}

For consistency with the pre-assigned sections, the section name (the first argument to \section) should be capitalized (but not all upper case) and not end in a period. Whitespace between the first and second braced expressions is not allowed. Markup (e.g. \code) within the section title may cause problems with the latex conversion (depending on the version of macro packages such as ‘hyperref’) and so should be avoided.

The \subsection macro takes arguments in the same format as \section, but is used within a section, so it may be used to nest subsections within sections or other subsections. There is no predefined limit on the nesting level, but formatting is not designed for more than 3 levels (i.e. subsections within subsections within sections).

Note that additional named sections are always inserted at a fixed position in the output (before \note, \seealso and the examples), no matter where they appear in the input (but in the same order amongst themselves as in the input).


2.3 Marking text

The following logical markup commands are available for emphasizing or quoting text.

\emph{text}
\strong{text}

Emphasize text using italic and bold font if possible; \strong is regarded as stronger (more emphatic).

\bold{text}

Set text in bold font where possible.

\sQuote{text}
\dQuote{text}

Portably single or double quote text (without hard-wiring the characters used for quotation marks).

Each of the above commands takes LaTeX-like input, so other macros may be used within text.

The following logical markup commands are available for indicating specific kinds of text. Except as noted, these take ‘verbatim’ text input, and so other macros may not be used within them. Some characters will need to be escaped (see Insertions).

\code{text}

Indicate text that is a literal example of a piece of an R program, e.g., a fragment of R code or the name of an R object. Text is entered in R-like syntax, and displayed using typewriter font where possible. Macros \var and \link are interpreted within text.

\preformatted{text}

Indicate text that is a literal example of a piece of a program. Text is displayed using typewriter font where possible. Formatting, e.g. line breaks, is preserved. (Note that this includes a line break after the initial {, so typically text should start on the same line as the command.)

Due to limitations in LaTeX as of this writing, this macro may not be nested within other markup macros other than \dQuote and \sQuote, as errors or bad formatting may result.

\kbd{keyboard-characters}

Indicate keyboard input, using slanted typewriter font if possible, so users can distinguish the characters they are supposed to type from computer output. Text is entered ‘verbatim’.

\samp{text}

Indicate text that is a literal example of a sequence of characters, entered ‘verbatim’, to be included within word-wrapped text. Displayed within single quotation marks and using typewriter font where possible.

\verb{text}

Indicate text that is a literal example of a sequence of characters, entered ‘verbatim’. No wrapping or reformatting will occur. Displayed using typewriter font where possible.

\pkg{package_name}

Indicate the name of an R package. LaTeX-like.

\file{file_name}

Indicate the name of a file. Text is LaTeX-like, so backslash needs to be escaped. Displayed using a distinct font where possible.

\email{email_address}

Indicate an electronic mail address. LaTeX-like, will be rendered as a hyperlink in HTML and PDF conversion. Displayed using typewriter font where possible.

\url{uniform_resource_locator}

Indicate a uniform resource locator (URL) for the World Wide Web. The argument is handled as ‘verbatim’ text (with percent and braces escaped by backslash), and rendered as a hyperlink in HTML and PDF conversion. Line feeds are removed, and leading and trailing whitespace121 is removed. See Specifying URLs.

Displayed using typewriter font where possible.

\href{uniform_resource_locator}{text}

Indicate a hyperlink to the World Wide Web. The first argument is handled as ‘verbatim’ text (with percent and braces escaped by backslash) and is used as the URL in the hyperlink, with the second argument of LaTeX-like text displayed to the user. Line feeds are removed from the first argument, and leading and trailing whitespace is removed.

Note that RFC3986-encoded URLs (e.g. using ‘%28VS.85%29’ in place of ‘(VS.85)’) may not work correctly in versions of R before 3.1.3 and are best avoided—use URLdecode() to decode them.

\var{metasyntactic_variable}

Indicate a metasyntactic variable. In most cases this will be rendered distinctly, e.g. in italic (PDF/HTML) or wrapped in ‘<…>’ (text), but not in all122. LaTeX-like.

\env{environment_variable}

Indicate an environment variable. ‘Verbatim’. Displayed using typewriter font where possible

\option{option}

Indicate a command-line option. ‘Verbatim’. Displayed using typewriter font where possible.

\command{command_name}

Indicate the name of a command. LaTeX-like, so \var is interpreted. Displayed using typewriter font where possible.

\dfn{term}

Indicate the introductory or defining use of a term. LaTeX-like.

\cite{reference}

Indicate a reference without a direct cross-reference via \link (see Cross-references), such as the name of a book. LaTeX-like.

\acronym{acronym}

Indicate an acronym (an abbreviation written in all capital letters), such as GNU. LaTeX-like.

\abbr{abbr}

Indicates an abbreviation. LaTeX-like.


2.4 Lists and tables

The \itemize and \enumerate commands take a single argument, within which there may be one or more \item commands. The text following each \item is formatted as one or more paragraphs, suitably indented and with the first paragraph marked with a bullet point (\itemize) or a number (\enumerate).

Note that unlike argument lists, \item in these formats is followed by a space and the text (not enclosed in braces). For example

  \enumerate{
    \item A database consists of one or more records, each with one or
    more named fields.
    \item Regular lines start with a non-whitespace character.
    \item Records are separated by one or more empty lines.
  }

\itemize and \enumerate commands may be nested.

The \describe command is similar to \itemize but allows initial labels to be specified. Each \item takes two arguments, the label and the body of the item, in exactly the same way as an argument or value \item. \describe commands are mapped to <DL> lists in HTML and \description lists in LaTeX.

Using these without any \items may cause problems with some conversions and makes little sense.

The \tabular command takes two arguments. The first gives for each of the columns the required alignment (‘l’ for left-justification, ‘r’ for right-justification or ‘c’ for centring.) The second argument consists of an arbitrary number of lines separated by \cr, and with fields separated by \tab. For example:

  \tabular{rlll}{
    [,1] \tab Ozone   \tab numeric \tab Ozone (ppb)\cr
    [,2] \tab Solar.R \tab numeric \tab Solar R (lang)\cr
    [,3] \tab Wind    \tab numeric \tab Wind (mph)\cr
    [,4] \tab Temp    \tab numeric \tab Temperature (degrees F)\cr
    [,5] \tab Month   \tab numeric \tab Month (1--12)\cr
    [,6] \tab Day     \tab numeric \tab Day of month (1--31)
  }

There must be the same number of fields on each line as there are alignments in the first argument, and they must be non-empty (but can contain only spaces). (There is no whitespace between \tabular and the first argument, nor between the two arguments.)


2.5 Cross-references

The markup \link{foo} (usually in the combination \code{\link{foo}}) produces a hyperlink to the help for foo. Here foo is a topic, that is the argument of \alias markup in another Rd file (possibly in another package). Hyperlinks are supported in some of the formats to which Rd files are converted, for example HTML and PDF, but ignored in others, e.g. the text format.

One main usage of \link is in the \seealso section of the help page, see Rd format.

Note that whereas leading and trailing spaces are stripped when extracting a topic from a \alias, they are not stripped when looking up the topic of a \link.

You can specify a link to a different topic than its name by \link[=dest]{name} which links to topic dest with name name. This can be used to refer to the documentation for S3/4 classes, for example \code{"\link[=abc-class]{abc}"} would be a way to refer to the documentation of an S4 class "abc" defined in your package, and \code{"\link[=terms.object]{terms}"} to the S3 "terms" class (in package stats). To make these easy to read in the source file, \code{"\linkS4class{abc}"} expands to the form given above.

There are two other forms with an optional ‘anchor’ argument, specified as \link[pkg]{foo} and \link[pkg:bar]{foo}, to link to topics foo and bar respectively in the package pkg. They are currently only used in HTML help (and ignored for hyperlinks in LaTeX conversions of help pages). One should be careful about topics containing special characters (such as arithmetic operators) as they may result in unresolvable links, and preferably use a safer alias in the same help page.

Historically (before R version 4.1.0), links of the form \link[pkg]{foo} and \link[pkg:bar]{foo} used to be interpreted as links to files foo.html and bar.html in package pkg, respectively. For this reason, the HTML help system looks for file foo.html in package pkg if it does not find topic foo, and then searches for the topic in other installed packages. To test that links work both with both old and new systems, the pre-4.1.0 behaviour can be restored by setting the environment variable _R_HELP_LINKS_TO_TOPICS_=false.

Packages referred to by these ‘other forms’ should be declared in the DESCRIPTION file, in the ‘Depends’, ‘Imports’, ‘Suggests’ or ‘Enhances’ fields.


2.6 Mathematics

Mathematical formulae should be set beautifully for printed documentation and in KaTeX/MathJax-enhanced HTML help (as from R 4.2.0) yet we still want something useful for plain-text (and legacy HTML) help. To this end, the two commands \eqn{latex}{ascii} and \deqn{latex}{ascii} are used. Whereas \eqn is used for “inline” formulae (corresponding to TeX’s $…$), \deqn gives “displayed equations” (as in LaTeX’s displaymath environment, or TeX’s $$…$$). Both arguments are treated as ‘verbatim’ text.

Both commands can also be used as \eqn{latexascii} (only one argument) which then is used for both latex and ascii. No whitespace is allowed between command and the first argument, nor between the first and second arguments.

The following example is from Poisson.Rd:

  \deqn{p(x) = \frac{\lambda^x e^{-\lambda}}{x!}}{%
        p(x) = \lambda^x exp(-\lambda)/x!}
  for \eqn{x = 0, 1, 2, \ldots}.

In plain-text help we get

    p(x) = lambda^x exp(-lambda)/x!

for x = 0, 1, 2, ....

In legacy HTML help, Greek letters (both cases) will be rendered if preceded by a backslash, \dots and \ldots will be rendered as ellipses and \sqrt, \ge and \le as mathematical symbols.

Note that only basic LaTeX can be used, there being no provision to specify LaTeX style files, but AMS extensions are supported as from R 4.2.2.


2.7 Figures

To include figures in help pages, use the \figure markup. There are three forms.

The two commonly used simple forms are \figure{filename} and \figure{filename}{alternate text}. This will include a copy of the figure in either HTML or LaTeX output. In text output, the alternate text will be displayed instead. (When the second argument is omitted, the filename will be used.) Both the filename and the alternate text will be parsed verbatim, and should not include special characters that are significant in HTML or LaTeX.

The expert form is \figure{filename}{options: string}. (The word ‘options:’ must be typed exactly as shown and followed by at least one space.) In this form, the string is copied into the HTML img tag as attributes following the src attribute, or into the second argument of the \Figure macro in LaTeX, which by default is used as options to an \includegraphics call. As it is unlikely that any single string would suffice for both display modes, the expert form would normally be wrapped in conditionals. It is up to the author to make sure that legal HTML/LaTeX is used. For example, to include a logo in both HTML (using the simple form) and LaTeX (using the expert form), the following could be used:

\if{html}{\figure{Rlogo.svg}{options: width=100 alt="R logo"}}
\if{latex}{\figure{Rlogo.pdf}{options: width=0.5in}}

The files containing the figures should be stored in the directory man/figures. Files with extensions .jpg, .jpeg, .pdf, .png and .svg from that directory will be copied to the help/figures directory at install time. (Figures in PDF format will not display in most HTML browsers, but might be the best choice in reference manuals.) Specify the filename relative to man/figures in the \figure directive.


2.8 Insertions

Use \R for the R system itself. The \dots macro is a historical alternative to using literal ‘...’ for the dots in function argument lists; use \ldots for ellipsis dots in ordinary text.123 These macros can be followed by {}, and should be unless followed by whitespace.

After an unescaped ‘%’, you can put your own comments regarding the help text. The rest of the line (but not the newline at the end) will be completely disregarded. Therefore, you can also use it to make part of the “help” invisible.

You can produce a backslash (‘\’) by escaping it by another backslash. (Note that \cr is used for generating line breaks.)

The “comment” character ‘%’ and unpaired braces124 almost always need to be escaped by ‘\’, and ‘\\’ can be used for backslash and needs to be when there are two or more adjacent backslashes. In R-like code quoted strings are handled slightly differently; see “Parsing Rd files” for details – in particular braces should not be escaped in quoted strings.

All of ‘% { } \’ should be escaped in LaTeX-like text.

Text which might need to be represented differently in different encodings should be marked by \enc, e.g. \enc{Jöreskog}{Joreskog} (with no whitespace between the braces) where the first argument will be used where encodings are allowed and the second should be ASCII (and is used for e.g. the text conversion in locales that cannot represent the encoded form). (This is intended to be used for individual words, not whole sentences or paragraphs.)


2.9 Indices

The \alias command (see Documenting functions) is used to specify the “topics” documented, which should include all R objects in a package such as functions and variables, data sets, and S4 classes and methods (see Documenting S4 classes and methods). The on-line help system searches the index data base consisting of all alias topics.

In addition, it is possible to provide “concept index entries” using \concept, which can be used for help.search() lookups. E.g., file cor.test.Rd in the standard package stats contains

\concept{Kendall correlation coefficient}
\concept{Pearson correlation coefficient}
\concept{Spearman correlation coefficient}

so that e.g. ??Spearman will succeed in finding the help page for the test for association between paired samples using Spearman’s rho.

(Note that help.search() only uses “sections” of documentation objects with no additional markup.)

Each \concept entry should give a single index term (word or phrase), and not use any Rd markup.

If you want to cross reference such items from other help files via \link, you need to use \alias and not \concept.


2.10 Platform-specific documentation

Sometimes the documentation needs to differ by platform. Currently two OS-specific options are available, ‘unix’ and ‘windows’, and lines in the help source file can be enclosed in

#ifdef OS
   ...
#endif

or

#ifndef OS
   ...
#endif

for OS-specific inclusion or exclusion. Such blocks should not be nested, and should be entirely within a block (that, is between the opening and closing brace of a section or item), or at top-level contain one or more complete sections.

If the differences between platforms are extensive or the R objects documented are only relevant to one platform, platform-specific Rd files can be put in a unix or windows subdirectory.


2.11 Conditional text

Occasionally the best content for one output format is different from the best content for another. For this situation, the \if{format}{text} or \ifelse{format}{text}{alternate} markup is used. Here format is a comma separated list of formats in which the text should be rendered. The alternate will be rendered if the format does not match. Both text and alternate may be any sequence of text and markup.

Currently the following formats are recognized: example, html, latex and text. These select output for the corresponding targets. (Note that example refers to extracted example code rather than the displayed example in some other format.) Also accepted are TRUE (matching all formats) and FALSE (matching no formats). These could be the output of the \Sexpr macro (see Dynamic pages).

The \out{literal} macro would usually be used within the text part of \if{format}{text}. It causes the renderer to output the literal text exactly, with no attempt to escape special characters. For example, use the following to output the markup necessary to display the Greek letter in LaTeX or HTML, and the text string alpha in other formats:

\ifelse{latex}{\out{$\alpha$}}{\ifelse{html}{\out{&alpha;}}{alpha}}

2.12 Dynamic pages

Two macros supporting dynamically generated man pages are \Sexpr and \RdOpts. These are modelled after Sweave, and are intended to contain executable R expressions in the Rd file.

The main argument to \Sexpr must be valid R code that can be executed. It may also take options in square brackets before the main argument. Depending on the options, the code may be executed at package build time, package install time, or man page rendering time.

The options follow the same format as in Sweave, but different options are supported. Currently the allowed options and their defaults are:

  • eval=TRUE Whether the R code should be evaluated.
  • echo=FALSE Whether the R code should be echoed. If TRUE and results=verbatim, a display will be given in a preformatted block. For example, \Sexpr[echo=TRUE,results=verbatim]{ x <- 1 } will be displayed as
    > x <- 1
    
  • keep.source=TRUE Whether to keep the author’s formatting when displaying the code, or throw it away and use a deparsed version.
  • results=text How should the results be displayed? The possibilities are:
    • results=text Apply as.character() to the result of the code, and insert it as a text element.
    • results=verbatim Print the results of the code just as if it was executed at the console, and include the printed results verbatim. (Invisible results will not print.)
    • results=rd The result is assumed to be a character vector containing markup to be passed to parse_Rd(), with the result inserted in place. This could be used to insert computed aliases, for instance. parse_Rd() is called first with fragment = FALSE to allow a single Rd section macro to be inserted. If that fails, it is called again with fragment = TRUE, the older behavior.
    • results=hide Insert no output.
  • strip.white=true Remove leading and trailing blank lines in verbatim output if strip.white=true (or TRUE). With strip.white=all, remove all blank lines.
  • stage=install Control when this macro is run. Possible values are
    • stage=build The macro is run when building a source tarball.
    • stage=install The macro is run when installing from source.
    • stage=render The macro is run when displaying the help page.

    Conditionals such as #ifdef (see Platform-specific documentation) are applied after the build macros but before the install macros. In some situations (e.g. installing directly from a source directory without a tarball, or building a binary package) the above description is not literally accurate, but authors can rely on the sequence being build, #ifdef, install, render, with all stages executed.

    Code is only run once in each stage, so a \Sexpr[results=rd] macro can output an \Sexpr macro designed for a later stage, but not for the current one or any earlier stage.

  • width, height, fig These options are currently allowed but ignored.

The \RdOpts macro is used to set new defaults for options to apply to following uses of \Sexpr.

For more details, see the online document “Parsing Rd files”.


2.13 User-defined macros

The \newcommand and \renewcommand macros allow new macros to be defined within an Rd file. These are similar but not identical to the same-named LaTeX macros.

They each take two arguments which are parsed verbatim. The first is the name of the new macro including the initial backslash, and the second is the macro definition. As in LaTeX, \newcommand requires that the new macro not have been previously defined, whereas \renewcommand allows existing macros (including all built-in ones) to be replaced. (This test is disabled by default, but may be enabled by setting the environment variable _R_WARN_DUPLICATE_RD_MACROS_ to a true value.)

Also as in LaTeX, the new macro may be defined to take arguments, and numeric placeholders such as #1 are used in the macro definition. However, unlike LaTeX, the number of arguments is determined automatically from the highest placeholder number seen in the macro definition. For example, a macro definition containing #1 and #3 (but no other placeholders) will define a three argument macro (whose second argument will be ignored). As in LaTeX, at most 9 arguments may be defined. If the # character is followed by a non-digit it will have no special significance. All arguments to user-defined macros will be parsed as verbatim text, and simple text-substitution will be used to replace the place-holders, after which the replacement text will be parsed.

A number of macros are defined in the file share/Rd/macros/system.Rd of the R source or home directory, and these will normally be available in all .Rd files. For example, that file contains the definition

\newcommand{\PR}{\Sexpr[results=rd]{tools:::Rd_expr_PR(#1)}}

which defines \PR to be a single argument macro; then code (typically used in the NEWS.Rd file) like

\PR{1234}

will expand to

\Sexpr[results=rd]{tools:::Rd_expr_PR(1234)}

when parsed.

Some macros that might be of general use are:

\CRANpkg{pkg}

A package on CRAN

\sspace

A single space (used after a period that does not end a sentence).

\doi{numbers}

A digital object identifier (DOI).

See the system.Rd file in share/Rd/macros for more details and macro definitions, including macros \packageTitle, \packageDescription, \packageAuthor, \packageMaintainer, \packageDESCRIPTION and \packageIndices.

Packages may also define their own common macros; these would be stored in an .Rd file in man/macros in the package source and will be installed into help/macros when the package is installed. A package may also use the macros from a different package by listing the other package in the ‘RdMacros’ field in the DESCRIPTION file.


2.14 Encoding

Rd files are text files and so it is impossible to deduce the encoding they are written in unless ASCII: files with 8-bit characters could be UTF-8, Latin-1, Latin-9, KOI8-R, EUC-JP, etc. So an \encoding{} section must be used to specify the encoding if it is not ASCII. (The \encoding{} section must be on a line by itself, and in particular one containing no non-ASCII characters. The encoding declared in the DESCRIPTION file will be used if none is declared in the file.) The Rd files are converted to UTF-8 before parsing and so the preferred encoding for the files themselves is now UTF-8.

Wherever possible, avoid non-ASCII chars in Rd files, and even symbols such as ‘<’, ‘>’, ‘$’, ‘^’, ‘&’, ‘|’, ‘@’, ‘~’, and ‘*’ outside ‘verbatim’ environments (since they may disappear in fonts designed to render text). (Function showNonASCIIfile in package tools can help in finding non-ASCII bytes in the files.)

For convenience, encoding names ‘latin1’ and ‘latin2’ are always recognized: these and ‘UTF-8’ are likely to work fairly widely. However, this does not mean that all characters in UTF-8 will be recognized, and the coverage of non-Latin characters125 is fairly low. Using LaTeX inputenx (see ?Rd2pdf in R) will give greater coverage of UTF-8.

The \enc command (see Insertions) can be used to provide transliterations which will be used in conversions that do not support the declared encoding.

The LaTeX conversion converts the file to UTF-8 from the declared encoding, and includes a

\inputencoding{utf8}

command, and this needs to be matched by a suitable invocation of the \usepackage{inputenc} command. The R utility R CMD Rd2pdf looks at the converted code and includes the encodings used: it might for example use

\usepackage[utf8]{inputenc}

(Use of utf8 as an encoding requires LaTeX dated 2003/12/01 or later. Also, the use of Cyrillic characters in ‘UTF-8’ appears to also need ‘\usepackage[T2A]{fontenc}’, and R CMD Rd2pdf includes this conditionally on the file t2aenc.def being present and environment variable _R_CYRILLIC_TEX_ being set.)

Note that this mechanism works best with Latin letters: the coverage of UTF-8 in LaTeX is quite low.


2.15 Processing documentation files

There are several commands to process Rd files from the system command line.

Using R CMD Rdconv one can convert R documentation format to other formats, or extract the executable examples for run-time testing. The currently supported conversions are to plain text, HTML and LaTeX as well as extraction of the examples.

R CMD Rd2pdf generates PDF output from documentation in Rd files, which can be specified either explicitly or by the path to a directory with the sources of a package. In the latter case, a reference manual for all documented objects in the package is created, including the information in the DESCRIPTION files.

R CMD Sweave and R CMD Stangle process vignette-like documentation files (e.g. Sweave vignettes with extension ‘.Snw’ or ‘.Rnw’, or other non-Sweave vignettes). R CMD Stangle is used to extract the R code fragments.

The exact usage and a detailed list of available options for all of these commands can be obtained by running R CMD command --help, e.g., R CMD Rdconv --help. All available commands can be listed using R --help (or Rcmd --help under Windows).

All of these work under Windows. You may need to have installed the the tools to build packages from source as described in the “R Installation and Administration” manual, although typically all that is needed is a LaTeX installation.


2.16 Editing Rd files

It can be very helpful to prepare .Rd files using a editor which knows about their syntax and will highlight commands, indent to show the structure and detect mis-matched braces, and so on.

The system most commonly used for this is some version of Emacs (including XEmacs) with the ESS package (https://ESS.R-project.org/: it is often is installed with Emacs but may need to be loaded, or even installed, separately).

Another is the Eclipse IDE with the Stat-ET plugin (https://projects.eclipse.org/projects/science.statet), and (on Windows only) Tinn-R (https://sourceforge.net/projects/tinn-r/).

People have also used LaTeX mode in a editor, as .Rd files are rather similar to LaTeX files.

Some R front-ends provide editing support for .Rd files, for example RStudio (https://posit.co/).


3 Tidying and profiling R code

R code which is worth preserving in a package and perhaps making available for others to use is worth documenting, tidying up and perhaps optimizing. The last two of these activities are the subject of this chapter.


3.1 Tidying R code

R treats function code loaded from packages and code entered by users differently. By default code entered by users has the source code stored internally, and when the function is listed, the original source is reproduced. Loading code from a package (by default) discards the source code, and the function listing is re-created from the parse tree of the function.

Normally keeping the source code is a good idea, and in particular it avoids comments being removed from the source. However, we can make use of the ability to re-create a function listing from its parse tree to produce a tidy version of the function, for example with consistent indentation and spaces around operators. If the original source does not follow the standard format this tidied version can be much easier to read.

We can subvert the keeping of source in two ways.

  1. The option keep.source can be set to FALSE before the code is loaded into R.
  2. The stored source code can be removed by calling the removeSource() function, for example by
    myfun <- removeSource(myfun)
    

In each case if we then list the function we will get the standard layout.

Suppose we have a file of functions myfuns.R that we want to tidy up. Create a file tidy.R containing

source("myfuns.R", keep.source = FALSE)
dump(ls(all.names = TRUE), file = "new.myfuns.R")

and run R with this as the source file, for example by R --vanilla < tidy.R or by pasting into an R session. Then the file new.myfuns.R will contain the functions in alphabetical order in the standard layout. Warning: comments in your functions will be lost.

The standard format provides a good starting point for further tidying. Although the deparsing cannot do so, we recommend the consistent use of the preferred assignment operator ‘<-’ (rather than ‘=’) for assignment. Many package authors use a version of Emacs (on a Unix-alike or Windows) to edit R code, using the ESS[S] mode of the ESS Emacs package. See R coding standards in R Internals for style options within the ESS[S] mode recommended for the source code of R itself.


3.2 Profiling R code for speed

It is possible to profile R code on Windows and most126 Unix-alike versions of R.

The command Rprof is used to control profiling, and its help page can be consulted for full details. Profiling works by recording at fixed intervals127 (by default every 20 msecs) which line in which R function is being used, and recording the results in a file (default Rprof.out in the working directory). Then the function summaryRprof or the command-line utility R CMD Rprof Rprof.out can be used to summarize the activity.

As an example, consider the following code (from Venables & Ripley, 2002, pp. 225–6).

library(MASS); library(boot)
storm.fm <- nls(Time ~ b*Viscosity/(Wt - c), stormer,
                start = c(b=30.401, c=2.2183))
st <- cbind(stormer, fit=fitted(storm.fm))
storm.bf <- function(rs, i) {
    st$Time <-  st$fit + rs[i]
    tmp <- nls(Time ~ (b * Viscosity)/(Wt - c), st,
               start = coef(storm.fm))
    tmp$m$getAllPars()
}
rs <- scale(resid(storm.fm), scale = FALSE) # remove the mean
Rprof("boot.out")
storm.boot <- boot(rs, storm.bf, R = 4999) # slow enough to profile
Rprof(NULL)

Having run this we can summarize the results by

R CMD Rprof boot.out

Each sample represents 0.02 seconds.
Total run time: 22.52 seconds.

Total seconds: time spent in function and callees.
Self seconds: time spent in function alone.

   %       total       %        self
 total    seconds     self    seconds    name
 100.0     25.22       0.2      0.04     "boot"
  99.8     25.18       0.6      0.16     "statistic"
  96.3     24.30       4.0      1.02     "nls"
  33.9      8.56       2.2      0.56     "<Anonymous>"
  32.4      8.18       1.4      0.36     "eval"
  31.8      8.02       1.4      0.34     ".Call"
  28.6      7.22       0.0      0.00     "eval.parent"
  28.5      7.18       0.3      0.08     "model.frame"
  28.1      7.10       3.5      0.88     "model.frame.default"
  17.4      4.38       0.7      0.18     "sapply"
  15.0      3.78       3.2      0.80     "nlsModel"
  12.5      3.16       1.8      0.46     "lapply"
  12.3      3.10       2.7      0.68     "assign"
 ...

   %        self        %      total
  self    seconds     total   seconds    name
   5.7      1.44       7.5      1.88     "inherits"
   4.0      1.02      96.3     24.30     "nls"
   3.6      0.92       3.6      0.92     "$"
   3.5      0.88      28.1      7.10     "model.frame.default"
   3.2      0.80      15.0      3.78     "nlsModel"
   2.8      0.70       9.8      2.46     "qr.coef"
   2.7      0.68      12.3      3.10     "assign"
   2.5      0.64       2.5      0.64     ".Fortran"
   2.5      0.62       7.1      1.80     "qr.default"
   2.2      0.56      33.9      8.56     "<Anonymous>"
   2.1      0.54       5.9      1.48     "unlist"
   2.1      0.52       7.9      2.00     "FUN"
  ...

This often produces surprising results and can be used to identify bottlenecks or pieces of R code that could benefit from being replaced by compiled code.

Two warnings: profiling does impose a small performance penalty, and the output files can be very large if long runs are profiled at the default sampling interval.

Profiling short runs can sometimes give misleading results. R from time to time performs garbage collection to reclaim unused memory, and this takes an appreciable amount of time which profiling will charge to whichever function happens to provoke it. It may be useful to compare profiling code immediately after a call to gc() with a profiling run without a preceding call to gc.

More detailed analysis of the output can be achieved by the tools in the CRAN packages proftools and profr: in particular these allow call graphs to be studied.


3.3 Profiling R code for memory use

Measuring memory use in R code is useful either when the code takes more memory than is conveniently available or when memory allocation and copying of objects is responsible for slow code. There are three ways to profile memory use over time in R code. The second and third require R to have been compiled with --enable-memory-profiling, which is not the default, but is currently used for the macOS and Windows binary distributions. All can be misleading, for different reasons.

In understanding the memory profiles it is useful to know a little more about R’s memory allocation. Looking at the results of gc() shows a division of memory into Vcells used to store the contents of vectors and Ncells used to store everything else, including all the administrative overhead for vectors such as type and length information. In fact the vector contents are divided into two pools. Memory for small vectors (by default 128 bytes or less) is obtained in large chunks and then parcelled out by R; memory for larger vectors is obtained directly from the operating system.

Some memory allocation is obvious in interpreted code, for example,

y <- x + 1

allocates memory for a new vector y. Other memory allocation is less obvious and occurs because R is forced to make good on its promise of ‘call-by-value’ argument passing. When an argument is passed to a function it is not immediately copied. Copying occurs (if necessary) only when the argument is modified. This can lead to surprising memory use. For example, in the ‘survey’ package we have

print.svycoxph <- function (x, ...)
{
    print(x$survey.design, varnames = FALSE, design.summaries = FALSE, ...)
    x$call <- x$printcall
    NextMethod()
}

It may not be obvious that the assignment to x$call will cause the entire object x to be copied. This copying to preserve the call-by-value illusion is usually done by the internal C function duplicate.

The main reason that memory-use profiling is difficult is garbage collection. Memory is allocated at well-defined times in an R program, but is freed whenever the garbage collector happens to run.


3.3.1 Memory statistics from Rprof

The sampling profiler Rprof described in the previous section can be given the option memory.profiling=TRUE. It then writes out the total R memory allocation in small vectors, large vectors, and cons cells or nodes at each sampling interval. It also writes out the number of calls to the internal function duplicate, which is called to copy R objects. summaryRprof provides summaries of this information. The main reason that this can be misleading is that the memory use is attributed to the function running at the end of the sampling interval. A second reason is that garbage collection can make the amount of memory in use decrease, so a function appears to use little memory. Running under gctorture helps with both problems: it slows down the code to effectively increase the sampling frequency and it makes each garbage collection release a smaller amount of memory.


3.3.2 Tracking memory allocations

The second method of memory profiling uses a memory-allocation profiler, Rprofmem(), which writes out a stack trace to an output file every time a large vector is allocated (with a user-specified threshold for ‘large’) or a new page of memory is allocated for the R heap. Summary functions for this output are still being designed.

Running the example from the previous section with

> Rprofmem("boot.memprof",threshold=1000)
> storm.boot <- boot(rs, storm.bf, R = 4999)
> Rprofmem(NULL)

shows that apart from some initial and final work in boot there are no vector allocations over 1000 bytes.


3.3.3 Tracing copies of an object

The third method of memory profiling involves tracing copies made of a specific (presumably large) R object. Calling tracemem on an object marks it so that a message is printed to standard output when the object is copied via duplicate or coercion to another type, or when a new object of the same size is created in arithmetic operations. The main reason that this can be misleading is that copying of subsets or components of an object is not tracked. It may be helpful to use tracemem on these components.

In the example above we can run tracemem on the data frame st

> tracemem(st)
[1] "<0x9abd5e0>"
> storm.boot <- boot(rs, storm.bf, R = 4)
memtrace[0x9abd5e0->0x92a6d08]: statistic boot
memtrace[0x92a6d08->0x92a6d80]: $<-.data.frame $<- statistic boot
memtrace[0x92a6d80->0x92a6df8]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x9271318]: statistic boot
memtrace[0x9271318->0x9271390]: $<-.data.frame $<- statistic boot
memtrace[0x9271390->0x9271408]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x914f558]: statistic boot
memtrace[0x914f558->0x914f5f8]: $<-.data.frame $<- statistic boot
memtrace[0x914f5f8->0x914f670]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x972cbf0]: statistic boot
memtrace[0x972cbf0->0x972cc68]: $<-.data.frame $<- statistic boot
memtrace[0x972cc68->0x972cd08]: $<-.data.frame $<- statistic boot
memtrace[0x9abd5e0->0x98ead98]: statistic boot
memtrace[0x98ead98->0x98eae10]: $<-.data.frame $<- statistic boot
memtrace[0x98eae10->0x98eae88]: $<-.data.frame $<- statistic boot

The object is duplicated fifteen times, three times for each of the R+1 calls to storm.bf. This is surprising, since none of the duplications happen inside nls. Stepping through storm.bf in the debugger shows that all three happen in the line

st$Time <- st$fit + rs[i]

Data frames are slower than matrices and this is an example of why. Using tracemem(st$Viscosity) does not reveal any additional copying.


3.4 Profiling compiled code

Profiling compiled code is highly system-specific, but this section contains some hints gleaned from various R users. Some methods need to be different for a compiled executable and for dynamic/shared libraries/objects as used by R packages.

This chapter is based on reports from users and the information may not be current.


3.4.1 Profiling on Linux

Options include using sprof for a shared object, and oprofile (see https://oprofile.sourceforge.io/news/) and perf (see https://perf.wiki.kernel.org/index.php/Tutorial) for any executable or shared object. These seem less widely supplied than they used to be. There is also ‘Google Performance Tools’, also known as gperftools or google-perftools.

All of these work best when R and any packages have been built with debugging symbols.

3.4.1.1 perf

This seems the most widely distributed tool. Here is an example on x86_64 Linux using R 4.3.1 built with LTO.

At its simplest

perf record R -f tests/Examples/stats-Ex.R
perf report --sort=dso
perf report --sort=srcfile
rm perf.data*

The first report is

  75.67%  R
   9.25%  libc.so.6
   4.87%  [unknown]
   3.75%  libz.so.1.2.11
   3.37%  stats.so
   1.17%  libm.so.6
   0.63%  libtirpc.so.3.0.0
   0.41%  graphics.so
   0.30%  grDevices.so
   0.20%  libRblas.so
   0.09%  libpcre2-8.so.0.11.0
   0.07%  methods.so
   ...

which shows which shared libraries (DSOs) the time was spent in.

perf annotate can be used on an application built with GCC and -ggdb: it interleaves disassembled and source code.

3.4.1.2 oprofile and operf

The oprofile project has two modes of operation. Since version 0.9.8 (August 2012), the preferred mode is to use operf, so we discuss only that.

Let us look at the boot example from §3.2 on x86_64 Linux using R 4.3.1.

This can be run under operf and analysed by commands like

operf R -f boot.R
opreport
opreport -l /path/to/R_HOME/bin/exec/R
opreport -l /path/to/R_HOME/library/stats/src/stats.so
opannotate --source /path/to/R_HOME/bin/exec/R

The first line had to be done as root.

The first report shows in which library (etc) the time was spent:

        CPU_CLK_UNHALT...|
          samples|      %|
        ------------------
           278341 91.9947 R
            18290  6.0450 libc.so.6
             2277  0.7526 kallsyms
             1426  0.4713 stats.so
              739  0.2442 libRblas.so
              554  0.1831 libz.so.1.2.11
              373  0.1233 libm.so.6
              352  0.1163 libtirpc.so.3.0.0
              153  0.0506 ld-linux-x86-64.so.2
               12  0.0040 methods.so

(kallsyms is the kernel.)

The rest of the output is voluminous, and only extracts are shown.

Most of the time within R is spent in

samples  %        image name symbol name
52955    19.0574  R                        bcEval.lto_priv.0
16484     5.9322  R                        Rf_allocVector3
14224     5.1189  R                        Rf_findVarInFrame3
12581     4.5276  R                        CONS_NR
8289      2.9830  R                        Rf_matchArgs_NR
8034      2.8913  R                        Rf_cons
7114      2.5602  R                        R_gc_internal.lto_priv.0
6552      2.3579  R                        Rf_eval
5969      2.1481  R                        VECTOR_ELT
5684      2.0456  R                        Rf_applyClosure
5497      1.9783  R                        findVarLocInFrame.part.0.lto_priv.0
4827      1.7371  R                        Rf_mkPROMISE
4609      1.6587  R                        Rf_install
4317      1.5536  R                        Rf_findFun3
4035      1.4521  R                        getvar.lto_priv.0
3491      1.2563  R                        SETCAR
3179      1.1441  R                        Rf_defineVar
2892      1.0408  R                        duplicate1.lto_priv.0

and in stats.so

samples  %        image name               symbol name
285      24.4845  stats.so                 termsform
284      24.3986  stats.so                 numeric_deriv
213      18.2990  stats.so                 modelframe
114       9.7938  stats.so                 nls_iter
55        4.7251  stats.so                 ExtractVars
47        4.0378  stats.so                 EncodeVars
37        3.1787  stats.so                 getListElement
32        2.7491  stats.so                 TrimRepeats
25        2.1478  stats.so                 InstallVar
20        1.7182  stats.so                 MatchVar
20        1.7182  stats.so                 isZeroOne
15        1.2887  stats.so                 ConvInfoMsg.isra.0

The profiling data is by default stored in sub-directory oprofile_data of the current directory, which can be removed at the end of the session.

3.4.1.3 sprof

You can select shared objects to be profiled with sprof by setting the environment variable LD_PROFILE. For example

% setenv LD_PROFILE /path/to/R_HOME/library/stats/libs/stats.so
% R -f boot.R
% sprof /path/to/R_HOME/library/stats/libs/stats.so \
  /var/tmp/path/to/R_HOME/library/stats/libs/stats.so.profile

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total
 time   seconds   seconds    calls  us/call  us/call  name
 76.19      0.32     0.32        0     0.00           numeric_deriv
 16.67      0.39     0.07        0     0.00           nls_iter
  7.14      0.42     0.03        0     0.00           getListElement

... to clean up ...
rm /var/tmp/path/to/R_HOME/library/stats/libs/stats.so.profile

It is possible that root access will be needed to create the directories used for the profile data.


3.4.2 Profiling on macOS

Developers have recommended Instruments (part of Xcode, see https://help.apple.com/instruments/mac/current/), This had a command-line version prior to macOS 12.


3.4.3 Profiling on Windows

Very Sleepy (https://github.com/VerySleepy/verysleepy) has been used on ‘x86_64’ Windows. There were problems with accessing the debug information, but the best results which included function names were obtained by attaching the profiler to an existing Rterm process, either via GUI or using /a: (PID obtained via Sys.getpid()).


4 Debugging

This chapter covers the debugging of R extensions, starting with the ways to get useful error information and moving on to how to deal with errors that crash R.


4.1 Browsing

Most of the R-level debugging facilities are based around the built-in browser. This can be used directly by inserting a call to browser() into the code of a function (for example, using fix(my_function) ). When code execution reaches that point in the function, control returns to the R console with a special prompt. For example

> fix(summary.data.frame) ## insert browser() call after for() loop
> summary(women)
Called from: summary.data.frame(women)
Browse[1]> ls()
 [1] "digits" "i"      "lbs"    "lw"     "maxsum" "ncw"    "nm"     "nr"
 [9] "nv"     "object" "sms"    "z"
Browse[1]> maxsum
[1] 7
Browse[1]> c
     height         weight
 Min.   :58.0   Min.   :115.0
 1st Qu.:61.5   1st Qu.:124.5
 Median :65.0   Median :135.0
 Mean   :65.0   Mean   :136.7
 3rd Qu.:68.5   3rd Qu.:148.0
 Max.   :72.0   Max.   :164.0
> rm(summary.data.frame)

At the browser prompt one can enter any R expression, so for example ls() lists the objects in the current frame, and entering the name of an object will128 print it. The following commands are also accepted

  • n

    Enter ‘step-through’ mode. In this mode, hitting the return key (RET) executes the next line of code (more precisely one line and any continuation lines). Typing c will continue to the end of the current context, e.g. to the end of the current loop or function.

  • c

    In normal mode, this quits the browser and continues execution, and just return works in the same way. cont is a synonym.

  • where

    This prints the call stack. For example

    > summary(women)
    Called from: summary.data.frame(women)
    Browse[1]> where
    where 1: summary.data.frame(women)
    where 2: summary(women)
    
    Browse[1]>
    
  • Q

    Quit both the browser and the current expression, and return to the top-level prompt.

Errors in code executed at the browser prompt will normally return control to the browser prompt. Objects can be altered by assignment, and will keep their changed values when the browser is exited. If really necessary, objects can be assigned to the workspace from the browser prompt (by using <<- if the name is not already in scope).


4.2 Debugging R code

Suppose your R program gives an error message. The first thing to find out is what R was doing at the time of the error, and the most useful tool is traceback(). We suggest that this is run whenever the cause of the error is not immediately obvious. Errors are often reported to the R mailing lists as being in some package when traceback() would show that the error was being reported by some other package or base R. Here is an example from the regression suite.

> success <- c(13,12,11,14,14,11,13,11,12)
> failure <- c(0,0,0,0,0,0,0,2,2)
> resp <- cbind(success, failure)
> predictor <- c(0, 5^(0:7))
> glm(resp ~ 0+predictor, family = binomial(link="log"))
Error: no valid set of coefficients has been found: please supply starting values
> traceback()
3: stop("no valid set of coefficients has been found: please supply
         starting values", call. = FALSE)
2: glm.fit(x = X, y = Y, weights = weights, start = start, etastart = etastart,
       mustart = mustart, offset = offset, family = family, control = control,
       intercept = attr(mt, "intercept") > 0)
1: glm(resp ~ 0 + predictor, family = binomial(link ="log"))

The calls to the active frames are given in reverse order (starting with the innermost). So we see the error message comes from an explicit check in glm.fit. (traceback() shows you all the lines of the function calls, which can be limited by setting option "deparse.max.lines".)

Sometimes the traceback will indicate that the error was detected inside compiled code, for example (from ?nls)

Error in nls(y ~ a + b * x, start = list(a = 0.12345, b = 0.54321), trace = TRUE) :
        step factor 0.000488281 reduced below 'minFactor' of 0.000976563
>  traceback()
2: .Call(R_nls_iter, m, ctrl, trace)
1: nls(y ~ a + b * x, start = list(a = 0.12345, b = 0.54321), trace = TRUE)

This will be the case if the innermost call is to .C, .Fortran, .Call, .External or .Internal, but as it is also possible for such code to evaluate R expressions, this need not be the innermost call, as in

> traceback()
9: gm(a, b, x)
8: .Call(R_numeric_deriv, expr, theta, rho, dir)
7: numericDeriv(form[[3]], names(ind), env)
6: getRHS()
5: assign("rhs", getRHS(), envir = thisEnv)
4: assign("resid", .swts * (lhs - assign("rhs", getRHS(), envir = thisEnv)),
       envir = thisEnv)
3: function (newPars)
   {
       setPars(newPars)
       assign("resid", .swts * (lhs - assign("rhs", getRHS(), envir = thisEnv)),
           envir = thisEnv)
       assign("dev", sum(resid^2), envir = thisEnv)
       assign("QR", qr(.swts * attr(rhs, "gradient")), envir = thisEnv)
       return(QR$rank < min(dim(QR$qr)))
   }(c(-0.00760232418963883, 1.00119632515036))
2: .Call(R_nls_iter, m, ctrl, trace)
1: nls(yeps ~ gm(a, b, x), start = list(a = 0.12345, b = 0.54321))

Occasionally traceback() does not help, and this can be the case if S4 method dispatch is involved. Consider the following example

> xyd <- new("xyloc", x=runif(20), y=runif(20))
Error in as.environment(pkg) : no item called "package:S4nswv"
on the search list
Error in initialize(value, ...) : S language method selection got
an error when called from internal dispatch for function 'initialize'
> traceback()
2: initialize(value, ...)
1: new("xyloc", x = runif(20), y = runif(20))

which does not help much, as there is no call to as.environment in initialize (and the note “called from internal dispatch” tells us so). In this case we searched the R sources for the quoted call, which occurred in only one place, methods:::.asEnvironmentPackage. So now we knew where the error was occurring. (This was an unusually opaque example.)

The error message

evaluation nested too deeply: infinite recursion / options(expressions=)?

can be hard to handle with the default value (5000). Unless you know that there actually is deep recursion going on, it can help to set something like

options(expressions=500)

and re-run the example showing the error.

Sometimes there is warning that clearly is the precursor to some later error, but it is not obvious where it is coming from. Setting options(warn = 2) (which turns warnings into errors) can help here.

Once we have located the error, we have some choices. One way to proceed is to find out more about what was happening at the time of the crash by looking a post-mortem dump. To do so, set options(error=dump.frames) and run the code again. Then invoke debugger() and explore the dump. Continuing our example:

> options(error = dump.frames)
> glm(resp ~ 0 + predictor, family = binomial(link ="log"))
Error: no valid set of coefficients has been found: please supply starting values

which is the same as before, but an object called last.dump has appeared in the workspace. (Such objects can be large, so remove it when it is no longer needed.) We can examine this at a later time by calling the function debugger.

> debugger()
Message:  Error: no valid set of coefficients has been found: please supply starting values
Available environments had calls:
1: glm(resp ~ 0 + predictor, family = binomial(link = "log"))
2: glm.fit(x = X, y = Y, weights = weights, start = start, etastart = etastart, mus
3: stop("no valid set of coefficients has been found: please supply starting values
Enter an environment number, or 0 to exit  Selection:

which gives the same sequence of calls as traceback, but in outer-first order and with only the first line of the call, truncated to the current width. However, we can now examine in more detail what was happening at the time of the error. Selecting an environment opens the browser in that frame. So we select the function call which spawned the error message, and explore some of the variables (and execute two function calls).

Enter an environment number, or 0 to exit  Selection: 2
Browsing in the environment with call:
   glm.fit(x = X, y = Y, weights = weights, start = start, etas
Called from: debugger.look(ind)
Browse[1]> ls()
 [1] "aic"        "boundary"   "coefold"    "control"    "conv"
 [6] "dev"        "dev.resids" "devold"     "EMPTY"      "eta"
[11] "etastart"   "family"     "fit"        "good"       "intercept"
[16] "iter"       "linkinv"    "mu"         "mu.eta"     "mu.eta.val"
[21] "mustart"    "n"          "ngoodobs"   "nobs"       "nvars"
[26] "offset"     "start"      "valideta"   "validmu"    "variance"
[31] "varmu"      "w"          "weights"    "x"          "xnames"
[36] "y"          "ynames"     "z"
Browse[1]> eta
            1             2             3             4             5
 0.000000e+00 -2.235357e-06 -1.117679e-05 -5.588393e-05 -2.794197e-04
            6             7             8             9
-1.397098e-03 -6.985492e-03 -3.492746e-02 -1.746373e-01
Browse[1]> valideta(eta)
[1] TRUE
Browse[1]> mu
        1         2         3         4         5         6         7         8
1.0000000 0.9999978 0.9999888 0.9999441 0.9997206 0.9986039 0.9930389 0.9656755
        9
0.8397616
Browse[1]> validmu(mu)
[1] FALSE
Browse[1]> c
Available environments had calls:
1: glm(resp ~ 0 + predictor, family = binomial(link = "log"))
2: glm.fit(x = X, y = Y, weights = weights, start = start, etastart = etastart
3: stop("no valid set of coefficients has been found: please supply starting v

Enter an environment number, or 0 to exit  Selection: 0
> rm(last.dump)

Because last.dump can be looked at later or even in another R session, post-mortem debugging is possible even for batch usage of R. We do need to arrange for the dump to be saved: this can be done either using the command-line flag --save to save the workspace at the end of the run, or via a setting such as

> options(error = quote({dump.frames(to.file=TRUE); q()}))

See the help on dump.frames for further options and a worked example.

An alternative error action is to use the function recover():

> options(error = recover)
> glm(resp ~ 0 + predictor, family = binomial(link = "log"))
Error: no valid set of coefficients has been found: please supply starting values

Enter a frame number, or 0 to exit

1: glm(resp ~ 0 + predictor, family = binomial(link = "log"))
2: glm.fit(x = X, y = Y, weights = weights, start = start, etastart = etastart

Selection:

which is very similar to dump.frames. However, we can examine the state of the program directly, without dumping and re-loading the dump. As its help page says, recover can be routinely used as the error action in place of dump.calls and dump.frames, since it behaves like dump.frames in non-interactive use.

Post-mortem debugging is good for finding out exactly what went wrong, but not necessarily why. An alternative approach is to take a closer look at what was happening just before the error, and a good way to do that is to use debug. This inserts a call to the browser at the beginning of the function, starting in step-through mode. So in our example we could use

> debug(glm.fit)
> glm(resp ~ 0 + predictor, family = binomial(link ="log"))
debugging in: glm.fit(x = X, y = Y, weights = weights, start = start, etastart = etastart,
    mustart = mustart, offset = offset, family = family, control = control,
    intercept = attr(mt, "intercept") > 0)
debug: {
## lists the whole function
Browse[1]>
debug: x <- as.matrix(x)
...
Browse[1]> start
[1] -2.235357e-06
debug: eta <- drop(x %*% start)
Browse[1]> eta
            1             2             3             4             5
 0.000000e+00 -2.235357e-06 -1.117679e-05 -5.588393e-05 -2.794197e-04
            6             7             8             9
-1.397098e-03 -6.985492e-03 -3.492746e-02 -1.746373e-01
Browse[1]>
debug: mu <- linkinv(eta <- eta + offset)
Browse[1]> mu
        1         2         3         4         5         6         7         8
1.0000000 0.9999978 0.9999888 0.9999441 0.9997206 0.9986039 0.9930389 0.9656755
        9
0.8397616

(The prompt Browse[1]> indicates that this is the first level of browsing: it is possible to step into another function that is itself being debugged or contains a call to browser().)

debug can be used for hidden functions and S3 methods by e.g. debug(stats:::predict.Arima). (It cannot be used for S4 methods, but an alternative is given on the help page for debug.) Sometimes you want to debug a function defined inside another function, e.g. the function arimafn defined inside arima. To do so, set debug on the outer function (here arima) and step through it until the inner function has been defined. Then call debug on the inner function (and use c to get out of step-through mode in the outer function).

To remove debugging of a function, call undebug with the argument previously given to debug; debugging otherwise lasts for the rest of the R session (or until the function is edited or otherwise replaced).

trace can be used to temporarily insert debugging code into a function, for example to insert a call to browser() just before the point of the error. To return to our running example

## first get a numbered listing of the expressions of the function
> page(as.list(body(glm.fit)), method="print")
> trace(glm.fit, browser, at=22)
Tracing function "glm.fit" in package "stats"
[1] "glm.fit"
> glm(resp ~ 0 + predictor, family = binomial(link ="log"))
Tracing glm.fit(x = X, y = Y, weights = weights, start = start,
   etastart = etastart,  .... step 22
Called from: eval(expr, envir, enclos)
Browse[1]> n
## and single-step from here.
> untrace(glm.fit)

For your own functions, it may be as easy to use fix to insert temporary code, but trace can help with functions in a namespace (as can fixInNamespace). Alternatively, use trace(,edit=TRUE) to insert code visually.


4.3 Checking memory access

Errors in memory allocation and reading/writing outside arrays are very common causes of crashes (e.g., segfaults) on some machines. Often the crash appears long after the invalid memory access: in particular damage to the structures which R itself has allocated may only become apparent at the next garbage collection (or even at later garbage collections after objects have been deleted).

Note that memory access errors may be seen with LAPACK, BLAS, OpenMP and Java-using packages: some at least of these seem to be intentional, and some are related to passing characters to Fortran.

Some of these tools can detect mismatched allocation and deallocation. C++ programmers should note that memory allocated by new [] must be freed by delete [], other uses of new by delete, and memory allocated by malloc, calloc and realloc by free. Some platforms will tolerate mismatches (perhaps with memory leaks) but others will segfault.


4.3.1 Using gctorture

We can help to detect memory problems in R objects earlier by running garbage collection as often as possible. This is achieved by gctorture(TRUE), which as described on its help page

Provokes garbage collection on (nearly) every memory allocation. Intended to ferret out memory protection bugs. Also makes R run very slowly, unfortunately.

The reference to ‘memory protection’ is to missing C-level calls to PROTECT/UNPROTECT (see Handling the effects of garbage collection) which if missing allow R objects to be garbage-collected when they are still in use. But it can also help with other memory-related errors.

Normally running under gctorture(TRUE) will just produce a crash earlier in the R program, hopefully close to the actual cause. See the next section for how to decipher such crashes.

It is possible to run all the examples, tests and vignettes covered by R CMD check under gctorture(TRUE) by using the option --use-gct.

The function gctorture2 provides more refined control over the GC torture process. Its arguments step, wait and inhibit_release are documented on its help page. Environment variables can also be used at the start of the R session to turn on GC torture: R_GCTORTURE corresponds to the step argument to gctorture2, R_GCTORTURE_WAIT to wait, and R_GCTORTURE_INHIBIT_RELEASE to inhibit_release.

If R is configured with --enable-strict-barrier then a variety of tests for the integrity of the write barrier are enabled. In addition tests to help detect protect issues are enabled:

  • All GCs are full GCs.
  • New nodes in small node pages are marked as NEWSXP on creation.
  • After a GC all free nodes that are not of type NEWSXP are marked as type FREESXP and their previous type is recorded.
  • Most calls to accessor functions check their SEXP inputs and SEXP outputs and signal an error if a FREESXP is found. The address of the node and the old type are included in the error message.

R CMD check --use-gct can be set to use gctorture2(n) rather than gctorture(TRUE) by setting environment variable _R_CHECK_GCT_N_ to a positive integer value to be used as n.

Used with a debugger and with gctorture or gctorture2 this mechanism can be helpful in isolating memory protect problems.


4.3.2 Using Valgrind

If you have access to Linux on a common CPU type or supported versions of FreeBSD or Solaris129 you can use valgrind (https://valgrind.org/, pronounced to rhyme with ‘tinned’) to check for possible problems. To run some examples under valgrind use something like

R -d valgrind --vanilla < mypkg-Ex.R
R -d "valgrind --tool=memcheck --leak-check=full" --vanilla < mypkg-Ex.R

where mypkg-Ex.R is a set of examples, e.g. the file created in mypkg.Rcheck by R CMD check. Occasionally this reports memory reads of ‘uninitialised values’ that are the result of compiler optimization, so can be worth checking under an unoptimized compile: for maximal information use a build with debugging symbols. We know there will be some small memory leaks from readline and R itself — these are memory areas that are in use right up to the end of the R session. Expect this to run around 20x slower than without valgrind, and in some cases much slower than that. Several versions of valgrind were not happy with some optimized BLAS libraries that use CPU-specific instructions so you may need to build a version of R specifically to use with valgrind.

On platforms where valgrind and its headers130 are installed you can build a version of R with extra instrumentation to help valgrind detect errors in the use of memory allocated from the R heap. The configure option is --with-valgrind-instrumentation=level, where level is 0, 1 or 2. Level 0 is the default and does not add anything. Level 1 will detect some uses131 of uninitialised memory and has little impact on speed (compared to level 0). Level 2 will detect many other memory-use bugs132 but make R much slower when running under valgrind. Using this in conjunction with gctorture can be even more effective (and even slower).

An example of valgrind output is

==12539== Invalid read of size 4
==12539==    at 0x1CDF6CBE: csc_compTr (Mutils.c:273)
==12539==    by 0x1CE07E1E: tsc_transpose (dtCMatrix.c:25)
==12539==    by 0x80A67A7: do_dotcall (dotcode.c:858)
==12539==    by 0x80CACE2: Rf_eval (eval.c:400)
==12539==    by 0x80CB5AF: R_execClosure (eval.c:658)
==12539==    by 0x80CB98E: R_execMethod (eval.c:760)
==12539==    by 0x1B93DEFA: R_standardGeneric (methods_list_dispatch.c:624)
==12539==    by 0x810262E: do_standardGeneric (objects.c:1012)
==12539==    by 0x80CAD23: Rf_eval (eval.c:403)
==12539==    by 0x80CB2F0: Rf_applyClosure (eval.c:573)
==12539==    by 0x80CADCC: Rf_eval (eval.c:414)
==12539==    by 0x80CAA03: Rf_eval (eval.c:362)
==12539==  Address 0x1C0D2EA8 is 280 bytes inside a block of size 1996 alloc'd
==12539==    at 0x1B9008D1: malloc (vg_replace_malloc.c:149)
==12539==    by 0x80F1B34: GetNewPage (memory.c:610)
==12539==    by 0x80F7515: Rf_allocVector (memory.c:1915)
...

This example is from an instrumented version of R, while tracking down a bug in the Matrix package in 2006. The first line indicates that R has tried to read 4 bytes from a memory address that it does not have access to. This is followed by a C stack trace showing where the error occurred. Next is a description of the memory that was accessed. It is inside a block allocated by malloc, called from GetNewPage, that is, in the internal R heap. Since this memory all belongs to R, valgrind would not (and did not) detect the problem in an uninstrumented build of R. In this example the stack trace was enough to isolate and fix the bug, which was in tsc_transpose, and in this example running under gctorture() did not provide any additional information.

valgrind is good at spotting the use of uninitialized values: use option --track-origins=yes to show where these originated from. What it cannot detect is the misuse of arrays allocated on the stack: this includes C automatic variables and some133 Fortran arrays.

It is possible to run all the examples, tests and vignettes covered by R CMD check under valgrind by using the option --use-valgrind. If you do this you will need to select the valgrind options some other way, for example by having a ~/.valgrindrc file containing

--leak-check=full
--track-origins=yes

or setting the environment variable VALGRIND_OPTS. As from R 4.2.0, --use-valgrind also uses valgrind when re-building the vignettes.

This section has described the use of memtest, the default (and most useful) of valgrind’s tools. There are others described in its documentation: helgrind can be useful for threaded programs.


4.3.3 Using the Address Sanitizer

AddressSanitizer (‘ASan’) is a tool with similar aims to the memory checker in valgrind. It is available with suitable builds134 of gcc and clang on common Linux and macOS platforms. See https://clang.llvm.org/docs/UsersManual.html#controlling-code-generation, https://clang.llvm.org/docs/AddressSanitizer.html and https://github.com/google/sanitizers.

More thorough checks of C++ code are done if the C++ library has been ‘annotated’: at the time of writing this applied to std::vector in libc++ for use with clang and gives rise to ‘container-overflow135 reports.

It requires code to have been compiled and linked with -fsanitize=address and compiling with -fno-omit-frame-pointer will give more legible reports. It has a runtime penalty of 2–3x, extended compilation times and uses substantially more memory, often 1–2GB, at run time. On 64-bit platforms it reserves (but does not allocate) 16–20TB of virtual memory: restrictive shell settings can cause problems. It can be helpful to increase the stack size, for example to 40MB.

By comparison with valgrind, ASan can detect misuse of stack and global variables but not the use of uninitialized memory.

Recent versions return symbolic addresses for the location of the error provided llvm-symbolizer136 is on the path: if it is available but not on the path or has been renamed137, one can use an environment variable, e.g.

ASAN_SYMBOLIZER_PATH=/path/to/llvm-symbolizer

An alternative is to pipe the output through asan_symbolize.py138 and perhaps then (for compiled C++ code) c++filt. (On macOS, you may need to run dsymutil to get line-number reports.)

The simplest way to make use of this is to build a version of R with something like

CC="gcc -std=gnu99 -fsanitize=address"
CFLAGS="-fno-omit-frame-pointer -g -O2 -Wall -pedantic -mtune=native"

which will ensure that the libasan run-time library is compiled into the R executable. However this check can be enabled on a per-package basis by using a ~/.R/Makevars file like

CC = gcc -std=gnu99 -fsanitize=address -fno-omit-frame-pointer
CXX = g++ -fsanitize=address -fno-omit-frame-pointer
FC = gfortran -fsanitize=address

(Note that -fsanitize=address has to be part of the compiler specification to ensure it is used for linking. These settings will not be honoured by packages which ignore ~/.R/Makevars.) It will be necessary to build R with

MAIN_LDFLAGS = -fsanitize=address

to link the runtime libraries into the R executable if it was not specified as part of ‘CC’ when R was built. (For some builds without OpenMP, -pthread is also required.)

For options available via the environment variable ASAN_OPTIONS see https://github.com/google/sanitizers/wiki/AddressSanitizerFlags. With gcc additional control is available via the --param flag: see its man page.

For more detailed information on an error, R can be run under a debugger with a breakpoint set before the address sanitizer report is produced: for gdb or lldb you could use

break __asan_report_error

(See https://github.com/google/sanitizers/wiki/AddressSanitizerAndDebugger.)

More recent versions139 added the flag -fsanitize-address-use-after-scope: see https://github.com/google/sanitizers/wiki/AddressSanitizerUseAfterScope.

One of the checks done by ASan is that malloc/free and in C++ new/delete and new[]/delete[] are used consistently (rather than say free being used to deallocate memory allocated by new[]). This matters on some systems but not all: unfortunately on some of those where it does not matter, system libraries140 are not consistent. The check can be suppressed by including ‘alloc_dealloc_mismatch=0’ in ASAN_OPTIONS.

ASan also checks system calls and sometimes reports can refer to problems in the system software and not the package nor R. A couple of reports have been of ‘heap-use-after-free’ errors in the X11 libraries called from Tcl/Tk.


4.3.3.1 Using the Leak Sanitizer

For x86_64 Linux there is a leak sanitizer, ‘LSan’: see https://github.com/google/sanitizers/wiki/AddressSanitizerLeakSanitizer. This is available on recent versions of gcc and clang, and where available is compiled in as part of ASan.

One way to invoke this from an ASan-enabled build is by the environment variable

ASAN_OPTIONS='detect_leaks=1'

However, this was made the default as from clang 3.5 and gcc 5.1.0.

When LSan is enabled, leaks give the process a failure error status (by default 23). For an R package this means the R process, and as the parser retains some memory to the end of the process, if R itself was built against ASan all runs will have a failure error status (which may include running R as part of building R itself).

To disable this, allocation-mismatch checking and some strict C++ checking use

setenv ASAN_OPTIONS 'alloc_dealloc_mismatch=0:detect_leaks=0:detect_odr_violation=0'

LSan also has a ‘stand-alone’ mode where it is compiled in using -fsanitize=leak and avoids the run-time overhead of ASan.


4.3.4 Using the Undefined Behaviour Sanitizer

‘Undefined behaviour’ is where the language standard does not require particular behaviour from the compiler. Examples include division by zero (where for doubles R requires the ISO/IEC 60559 behaviour but C/C++ do not), use of zero-length arrays, shifts too far for signed types (e.g. int x, y; y = x << 31;), out-of-range coercion, invalid C++ casts and mis-alignment. Not uncommon examples of out-of-range coercion in R packages are attempts to coerce a NaN or infinity to type int or NA_INTEGER to an unsigned type such as size_t. Also common is y[x - 1] forgetting that x might be NA_INTEGER.

‘UBSanitizer’ is a tool for C/C++ source code selected by -fsanitize=undefined in suitable builds141 of clang and GCC. Its (main) runtime library is linked into each package’s DLL, so it is less often needed to be included in MAIN_LDFLAGS. Platforms supported by clang are listed at https://clang.llvm.org/docs/UndefinedBehaviorSanitizer.html#supported-platforms: CRAN uses it for C/C++ with both GCC and clang on ‘x86_64’ Linux: the two toolchains often highlight different things with more reports from clang than GCC.

This sanitizer may be combined with the Address Sanitizer by -fsanitize=undefined,address (where both are supported, and we have seen library conflicts for clang 17 and later).

Finer control of what is checked can be achieved by other options.

For clang see https://clang.llvm.org/docs/UndefinedBehaviorSanitizer.html#ubsan-checks. The current set is (on a single line):

-fsanitize=alignment,bool,bounds,builtin,enum,float-cast-overflow,
float-divide-by-zero,function,implicit-unsigned-integer-truncation,
implicit-signed-integer-truncation,implicit-integer-sign-change,
integer-divide-by-zero,nonnull-attribute,null,nullability-arg,
nullability-assign,nullability-return,object-size,
pointer-overflow,return,returns-nonnull-attribute,shift,
signed-integer-overflow,unreachable,unsigned-integer-overflow,
unsigned-shift-base,vla-bound,vptr

(plus the more specific versions array-bounsds, local-bounds, shift-base and shift-exponent), or use something like

-fsanitize=undefined -fno-sanitize=float-divide-by-zero

where in recent versions -fno-sanitize=float-divide-by-zero is the default.

Options return and vptr apply only to C++: to use vptr its run-time library needs to be linked into the main R executable by building the latter with something like

MAIN_LD="clang++ -fsanitize=undefined"

Option float-divide-by-zero is undesirable for use with R which allow such divisions as part of IEC 60559 arithmetic, and in versions of clang since June 2019 it is no longer part of -fsanitize=undefined.

There are also groups of options implicit-integer-truncation, mplicit-integer-arithmetic-value-change, implicit-conversion, integer and nullability.

For GCC see https://gcc.gnu.org/onlinedocs/gcc/Instrumentation-Options.html (or the manual for your version of GCC, installed or via https://gcc.gnu.org/onlinedocs/: look for ‘Program Instrumentation Options’) for the options supported by GCC: versions 13.x supported

-fsanitize=alignment,bool,bounds,builtin,enum,integer-divide-by-zero,
nonnull-attribute,null,object-size,pointer-overflow,return,
returns-nonnull-attribute,shift,signed-integer-overflow,
unreachable,vla-bound,vptr

plus the more specific versions shift-base and shift-exponent and non-default options

bounds-strict,float-cast-overflow,float-divide-by-zero

where float-divide-by-zero is not desirable for R uses and bounds-strict is an extension of bounds.

Other useful flags include

-no-fsanitize-recover

which causes the first report to be fatal (it always is for the unreachable and return suboptions). For more detailed information on where the runtime error occurs, using

setenv UBSAN_OPTIONS 'print_stacktrace=1'

will include a traceback in the report. Beyond that, R can be run under a debugger with a breakpoint set before the sanitizer report is produced: for gdb or lldb you could use

break __ubsan_handle_float_cast_overflow
break __ubsan_handle_float_cast_overflow_abort

or similar (there are handlers for each type of undefined behaviour).

There are also the compiler flags -fcatch-undefined-behavior and -ftrapv, said to be more reliable in clang than gcc.

For more details on the topic see https://blog.regehr.org/archives/213 and https://blog.llvm.org/2011/05/what-every-c-programmer-should-know.html (which has 3 parts).

It may or may not be possible to build R itself with -fsanitize=undefined: problems have in the past been seen with OpenMP-using code with gcc but there has been success with clang up to version 16.. However, problems have been seen with clang 17 and later, including missing entry points and R builds hanging. What has succeeded is to use UBSAN just for the package under test (and not in combination with ASAN). To do so, check with an unaltered R, using a custom Makevars file something like

CC = clang -fsanitize=undefined -fno-sanitize=float-divide-by-zero -fno-omit-frame-pointer
CXX = clang++ -fsanitize=undefined -fno-sanitize=float-divide-by-zero -fno-omit-frame-pointer -frtti

UBSAN_DIR = /path/to/LLVM18/lib/clang/18/lib/x86_64-unknown-linux-gnu
SAN_LIBS = $(UBSAN_DIR)/libclang_rt.ubsan_standalone.a $(UBSAN_DIR)/libclang_rt.ubsan_standalone_cxx.a

which links the UBSAN libraries statically into the package-under-test’s DSO. It is also possible to use the dynamic library via

SAN_LIBS = -L$(UBSAN_DIR) -Wl,-rpath,$(UBSAN_DIR) -lclang_rt.ubsan_standalone

provided UBSAN_DIR is added to the runtime library path (as shown or using LD_LIBRARY_PATH). N.B.: The details, especially the paths used, have changed several times recently.


4.3.5 Other analyses with ‘clang’

Recent versions of clang on ‘x86_64’ Linux have ‘ThreadSanitizer’ (https://github.com/google/sanitizers/wiki#threadsanitizer), a ‘data race detector for C/C++ programs’, and ‘MemorySanitizer’ (https://clang.llvm.org/docs/MemorySanitizer.html, https://github.com/google/sanitizers) for the detection of uninitialized memory. Both are based on and provide similar functionality to tools in valgrind.

clang has a ‘Static Analyzer’ which can be run on the source files during compilation: see https://clang-analyzer.llvm.org/.


4.3.6 Other analyses with ‘gcc’

GCC 10 introduced a new flag -fanalyzer which does static analysis during compilation, currently for C code. It is regarded as experimental and it may slow down computation considerably when problems are found (and use many GB of resident memory). There is some overlap with problems detected by the Undefined Behaviour sanitizer, but some issues are only reported by this tool and as it is a static analysis, it does not rely on code paths being exercised.

See https://gcc.gnu.org/onlinedocs/gcc-10.1.0/gcc/Static-Analyzer-Options.html (or the documentation for your version of gcc if later) and https://developers.redhat.com/blog/2020/03/26/static-analysis-in-gcc-10


4.3.7 Using ‘Dr. Memory’

‘Dr. Memory’ from https://drmemory.org/ is a memory checker for (currently) Windows, Linux and macOS with similar aims to valgrind. It works with unmodified executables142 and detects memory access errors, uninitialized reads and memory leaks.


4.3.8 Fortran array bounds checking

Most of the Fortran compilers used with R allow code to be compiled with checking of array bounds: for example gfortran has option -fbounds-check. This will give an error when the upper or lower bound is exceeded, e.g.

At line 97 of file .../src/appl/dqrdc2.f
Fortran runtime error: Index '1' of dimension 1 of array 'x' above upper bound of 0

One does need to be aware that lazy programmers often specify Fortran dimensions as 1 rather than * or a real bound and these will be reported (as may * dimensions)

It is easy to arrange to use this check on just the code in your package: add to ~/.R/Makevars something like (for gfortran)

FFLAGS = -g -O2 -mtune=native -fbounds-check

when you run R CMD check.

This may report errors with the way that Fortran character variables are passed, particularly when Fortran subroutines are called from C code and character lengths are not passed (see Fortran character strings).


4.4 Debugging compiled code

Sooner or later programmers will be faced with the need to debug compiled code loaded into R. This section is geared to platforms using gdb with code compiled by gcc, but similar things are possible with other debuggers such as lldb (https://lldb.llvm.org/, used on macOS) and Sun’s dbx: some debuggers have graphical front-ends available.

Consider first ‘crashes’, that is when R terminated unexpectedly with an illegal memory access (a ‘segfault’ or ‘bus error’), illegal instruction or similar. Unix-alike versions of R use a signal handler which aims to give some basic information. For example

 *** caught segfault ***
address 0x20000028, cause 'memory not mapped'

Traceback:
 1: .identC(class1[[1]], class2)
 2: possibleExtends(class(sloti), classi, ClassDef2 = getClassDef(classi,
where = where))
 3: validObject(t(cu))
 4: stopifnot(validObject(cu <- as(tu, "dtCMatrix")), validObject(t(cu)),
validObject(t(tu)))

Possible actions:
1: abort (with core dump)
2: normal R exit
3: exit R without saving workspace
4: exit R saving workspace
Selection: 3

Since the R process may be damaged, the only really safe options are the first or third. (Note that a core dump is only produced where enabled: a common default in a shell is to limit its size to 0, thereby disabling it.)

A fairly common cause of such crashes is a package which uses .C or .Fortran and writes beyond (at either end) one of the arguments it is passed. There is a good way to detect this: using options(CBoundsCheck = TRUE) (which can be selected via the environment variable R_C_BOUNDS_CHECK=yes) changes the way .C and .Fortran work to check if the compiled code writes in the 64 bytes at either end of an argument.

Another cause of a ‘crash’ is to overrun the C stack. R tries to track that in its own code, but it may happen in third-party compiled code. For modern POSIX-compliant OSes R can safely catch that and return to the top-level prompt, so one gets something like

> .C("aaa")
Error: segfault from C stack overflow
>

However, C stack overflows are fatal under Windows and normally defeat attempts at debugging on that platform. Further, the size of the stack is set when R is compiled on Windows, whereas on POSIX OSes it can be set in the shell from which R is launched.

If you have a crash which gives a core dump you can use something like

gdb /path/to/R/bin/exec/R core.12345

to examine the core dump. If core dumps are disabled or to catch errors that do not generate a dump one can run R directly under a debugger by for example

$ R -d gdb --vanilla
...
gdb> run

at which point R will run normally, and hopefully the debugger will catch the error and return to its prompt. This can also be used to catch infinite loops or interrupt very long-running code. For a simple example

> for(i in 1:1e7) x <- rnorm(100)
[hit Ctrl-C]
Program received signal SIGINT, Interrupt.
0x00397682 in _int_free () from /lib/tls/libc.so.6
(gdb) where
#0  0x00397682 in _int_free () from /lib/tls/libc.so.6
#1  0x00397eba in free () from /lib/tls/libc.so.6
#2  0xb7cf2551 in R_gc_internal (size_needed=313)
    at /users/ripley/R/svn/R-devel/src/main/memory.c:743
#3  0xb7cf3617 in Rf_allocVector (type=13, length=626)
    at /users/ripley/R/svn/R-devel/src/main/memory.c:1906
#4  0xb7c3f6d3 in PutRNGstate ()
    at /users/ripley/R/svn/R-devel/src/main/RNG.c:351
#5  0xb7d6c0a5 in do_random2 (call=0x94bf7d4, op=0x92580e8, args=0x9698f98,
    rho=0x9698f28) at /users/ripley/R/svn/R-devel/src/main/random.c:183
...

In many cases it is possible to attach a debugger to a running process: this is helpful if an alternative front-end is in use or to investigate a task that seems to be taking far too long. This is done by something like

gdb -p pid

where pid is the id of the R executable or front-end process and can be found from within a running R process by calling Sys.getpid() or from a process monitor. This stops the process so its state can be examined: use continue to resume execution.

Some “tricks” worth knowing follow:


4.4.1 Finding entry points in dynamically loaded code

Under most compilation environments, compiled code dynamically loaded into R cannot have breakpoints set within it until it is loaded. To use a symbolic debugger on such dynamically loaded code under Unix-alikes use

  • Call the debugger on the R executable, for example by R -d gdb.
  • Start R.
  • At the R prompt, use dyn.load or library to load your shared object.
  • Send an interrupt signal. This will put you back to the debugger prompt.
  • Set the breakpoints in your code.
  • Continue execution of R by typing signal 0 and hitting return (RET).

Under Windows signals may not be able to be used, and if so the procedure is more complicated. See the rw-FAQ.


4.4.2 Inspecting R objects when debugging

The key to inspecting R objects from compiled code is the function PrintValue(SEXP s) which uses the normal R printing mechanisms to print the R object pointed to by s, or the safer version R_PV(SEXP s) which will only print ‘objects’.

One way to make use of PrintValue is to insert suitable calls into the code to be debugged.

Another way is to call R_PV from the symbolic debugger. (PrintValue is hidden as Rf_PrintValue.) For example, from gdb we can use

(gdb) p R_PV(ab)

using the object ab from the convolution example, if we have placed a suitable breakpoint in the convolution C code.

To examine an arbitrary R object we need to work a little harder. For example, let

R> DF <- data.frame(a = 1:3, b = 4:6)

By setting a breakpoint at do_get and typing get("DF") at the R prompt, one can find out the address in memory of DF, for example

Value returned is $1 = (SEXPREC *) 0x40583e1c
(gdb) p *$1
$2 = {
  sxpinfo = {type = 19, obj = 1, named = 1, gp = 0,
    mark = 0, debug = 0, trace = 0, = 0},
  attrib = 0x40583e80,
  u = {
    vecsxp = {
      length = 2,
      type = {c = 0x40634700 "0>X@D>X@0>X@", i = 0x40634700,
        f = 0x40634700, z = 0x40634700, s = 0x40634700},
      truelength = 1075851272,
    },
    primsxp = {offset = 2},
    symsxp = {pname = 0x2, value = 0x40634700, internal = 0x40203008},
    listsxp = {carval = 0x2, cdrval = 0x40634700, tagval = 0x40203008},
    envsxp = {frame = 0x2, enclos = 0x40634700},
    closxp = {formals = 0x2, body = 0x40634700, env = 0x40203008},
    promsxp = {value = 0x2, expr = 0x40634700, env = 0x40203008}
  }
}

(Debugger output reformatted for better legibility).

Using R_PV() one can “inspect” the values of the various elements of the SEXP, for example,

(gdb) p R_PV($1->attrib)
$names
[1] "a" "b"

$row.names
[1] "1" "2" "3"

$class
[1] "data.frame"

$3 = void

To find out where exactly the corresponding information is stored, one needs to go “deeper”:

(gdb) set $a = $1->attrib
(gdb) p $a->u.listsxp.tagval->u.symsxp.pname->u.vecsxp.type.c
$4 = 0x405d40e8 "names"
(gdb) p $a->u.listsxp.carval->u.vecsxp.type.s[1]->u.vecsxp.type.c
$5 = 0x40634378 "b"
(gdb) p $1->u.vecsxp.type.s[0]->u.vecsxp.type.i[0]
$6 = 1
(gdb) p $1->u.vecsxp.type.s[1]->u.vecsxp.type.i[1]
$7 = 5

Another alternative is the R_inspect function which shows the low-level structure of the objects recursively (addresses differ from the above as this example is created on another machine):

(gdb) p R_inspect($1)
@100954d18 19 VECSXP g0c2 [OBJ,NAM(2),ATT] (len=2, tl=0)
  @100954d50 13 INTSXP g0c2 [NAM(2)] (len=3, tl=0) 1,2,3
  @100954d88 13 INTSXP g0c2 [NAM(2)] (len=3, tl=0) 4,5,6
ATTRIB:
  @102a70140 02 LISTSXP g0c0 []
    TAG: @10083c478 01 SYMSXP g0c0 [MARK,NAM(2),gp=0x4000] "names"
    @100954dc0 16 STRSXP g0c2 [NAM(2)] (len=2, tl=0)
      @10099df28 09 CHARSXP g0c1 [MARK,gp=0x21] "a"
      @10095e518 09 CHARSXP g0c1 [MARK,gp=0x21] "b"
    TAG: @100859e60 01 SYMSXP g0c0 [MARK,NAM(2),gp=0x4000] "row.names"
    @102a6f868 13 INTSXP g0c1 [NAM(2)] (len=2, tl=1) -2147483648,-3
    TAG: @10083c948 01 SYMSXP g0c0 [MARK,gp=0x4000] "class"
    @102a6f838 16 STRSXP g0c1 [NAM(2)] (len=1, tl=1)
      @1008c6d48 09 CHARSXP g0c2 [MARK,gp=0x21,ATT] "data.frame"

In general the representation of each object follows the format:

@<address> <type-nr> <type-name> <gc-info> [<flags>] ...

For a more fine-grained control over the depth of the recursion and the output of vectors R_inspect3 takes additional two character() parameters: maximum depth and the maximal number of elements that will be printed for scalar vectors. The defaults in R_inspect are currently -1 (no limit) and 5 respectively.


4.4.3 Debugging on macOS

To debug code in a package it is easiest to unpack it in a directory and install it with

R CMD INSTALL --dsym pkgname

as macOS does not store debugging symbols in the .so file. (It is not necessary to have R built with debugging symbols, although compiling the package should be done including -g in CFLAGS / CXXFLAGS / FFLAGS / FCFLAGS as appropriate.)

Security measures may prevent running a CRAN binary distribution of R under lldb or attaching this as a debugger (https://cran.r-project.org/bin/macosx/RMacOSX-FAQ.html#I-cannot-attach-debugger-to-R), although both were possible on High Sierra and are again from R 4.2.0. This can also affect locally compiled builds, where attaching to an interactive R session under Big Sur or Monterey worked in 2022 after giving administrator permission via a popup-up. (To debug in what Apple deems a non-interactive session, e.g. logged in remotely, see man DevToolsSecurity.)

Debugging a local build of R on macOS can raise additional hurdles as environment variables such as DYLD_FALLBACK_LIBRARY_PATH are not usually passed through143 the lldb process, resulting in messages like

R -d lldb
...
(lldb) run
Process 16828 launched: '/path/to/bin/exec/R' (x86_64)
dyld: Library not loaded: libR.dylib
  Referenced from: /path/to/bin/exec/R

A quick workaround is to symlink the dylibs under R_HOME/lib to somewhere where they will be found such as the current working directory. It would be possible to do as the distribution does144 and use install_name_tool, but that would have to be done for all the dylibs including those in packages.

It may be simplest to attach the debugger to a running process (see above). Specifically, run R and when it is at the prompt just before a command that is to be debugged, at a terminal

ps -ef | grep exec/R
# identify the PID pid for the next command: it is the second item
lldb -p pid
(lldb) continue

and then return to the R console.

For non-interactive use, one may need lldb --batch.


5 System and foreign language interfaces

Many of the functions described here have entry-point names with a Rf_ prefix: if they are called from C code (but not C++ code as from R 4.5.0) that prefix can be omitted. Users are encouraged to use the prefix when writing new C code.


5.1 Operating system access

Access to operating system functions is via the R functions system and system2. The details will differ by platform (see the on-line help), and about all that can safely be assumed is that the first argument will be a string command that will be passed for execution (not necessarily by a shell) and the second argument to system will be internal which if true will collect the output of the command into an R character vector.

On POSIX-compliant OSes these commands pass a command-line to a shell: Windows is not POSIX-compliant and there is a separate function shell to do so.

The function system.time is available for timing. Timing on child processes is only available on Unix-alikes, and may not be reliable there.


5.2 Interface functions .C and .Fortran

These two functions provide an interface to compiled code that has been linked into R, either at build time or via dyn.load (see dyn.load and dyn.unload). They are primarily intended for compiled C and Fortran code respectively, but the .C function can be used with other languages which can generate C interfaces, for example C++ (see Interfacing C++ code).

The first argument to each function is a character string specifying the symbol name as known145 to C or Fortran, that is the function or subroutine name. (That the symbol is loaded can be tested by, for example, is.loaded("cg"). Use the name you pass to .C or .Fortran rather than the translated symbol name.)

There can be up to 65 further arguments giving R objects to be passed to compiled code. Normally these are copied before being passed in, and copied again to an R list object when the compiled code returns. If the arguments are given names, these are used as names for the components in the returned list object (but not passed to the compiled code).

The following table gives the mapping between the modes of R atomic vectors and the types of arguments to a C function or Fortran subroutine.

R storage modeC typeFortran type
logicalint *INTEGER
integerint *INTEGER
doubledouble *DOUBLE PRECISION
complexRcomplex *DOUBLE COMPLEX
characterchar **CHARACTER(255)
rawunsigned char *none

On all R platforms int and INTEGER are 32-bit. Code ported from S-PLUS (which uses long * for logical and integer) will not work on all 64-bit platforms (although it may appear to work on some, including ‘x86_64’ Windows). Note also that if your compiled code is a mixture of C functions and Fortran subprograms the argument types must match as given in the table above.

C type Rcomplex is a structure with double members r and i defined in the header file R_ext/Complex.h.146 (On most platforms this is stored in a way compatible with the C99 double complex type: however, it may not be possible to pass Rcomplex to a C99 function expecting a double complex argument. Nor need it be compatible with a C++ complex type. Moreover, the compatibility can depend on the optimization level set for the compiler.)

Only a single character string of fixed length can be passed to or from Fortran (the length is not passed), and the success of this is compiler-dependent: its use was formally deprecated in 2019. Other R objects can be passed to .C, but it is much better to use one of the other interfaces.

It is possible to pass numeric vectors of storage mode double to C as float * or to Fortran as REAL by setting the attribute Csingle, most conveniently by using the R functions as.single, single or mode. This is intended only to be used to aid interfacing existing C or Fortran code.

Logical values are sent as 0 (FALSE), 1 (TRUE) or INT_MIN = -2147483648 (NA, but only if NAOK is true), and the compiled code should return one of these three values. (Non-zero values other than INT_MIN are mapped to TRUE.) Note that the use of int * for Fortran logical is not guaranteed to be portable (although people have gotten away with it for many years): it is better to pass integers and convert to/from Fortran logical in a Fortran wrapper.

Unless formal argument NAOK is true, all the other arguments are checked for missing values NA and for the IEEE special values NaN, Inf and -Inf, and the presence of any of these generates an error. If it is true, these values are passed unchecked.

Argument PACKAGE confines the search for the symbol name to a specific shared object (or use "base" for code compiled into R). Its use is highly desirable, as there is no way to avoid two package writers using the same symbol name, and such name clashes are normally sufficient to cause R to crash. (If it is not present and the call is from the body of a function defined in a package namespace, the shared object loaded by the first (if any) useDynLib directive will be used.)

Note that the compiled code should not return anything except through its arguments: C functions should be of type void and Fortran subprograms should be subroutines.

To fix ideas, let us consider a very simple example which convolves two finite sequences. (This is hard to do fast in interpreted R code, but easy in C code.) We could do this using .C by

void convolve(double *a, int *na, double *b, int *nb, double *ab)
{
    int nab = *na + *nb - 1;

    for(int i = 0; i < nab; i++)
        ab[i] = 0.0;
    for(int i = 0; i < *na; i++)
        for(int j = 0; j < *nb; j++)
            ab[i + j] += a[i] * b[j];
}

called from R by

conv <- function(a, b)
    .C("convolve",
       as.double(a),
       as.integer(length(a)),
       as.double(b),
       as.integer(length(b)),
       ab = double(length(a) + length(b) - 1))$ab

Note that we take care to coerce all the arguments to the correct R storage mode before calling .C; mistakes in matching the types can lead to wrong results or hard-to-catch errors.

Special care is needed in handling character vector arguments in C (or C++). On entry the contents of the elements are duplicated and assigned to the elements of a char ** array, and on exit the elements of the C array are copied to create new elements of a character vector. This means that the contents of the character strings of the char ** array can be changed, including to \0 to shorten the string, but the strings cannot be lengthened. It is possible147 to allocate a new string via R_alloc and replace an entry in the char ** array by the new string. However, when character vectors are used other than in a read-only way, the .Call interface is much to be preferred.

Passing character strings to Fortran code needs even more care, is deprecated and should be avoided where possible. Only the first element of the character vector is passed in, as a fixed-length (255) character array. Up to 255 characters are passed back to a length-one character vector. How well this works (or even if it works at all) depends on the C and Fortran compilers on each platform (including on their options). Often what is being passed to Fortran is one of a small set of possible values (a factor in R terms) which could alternatively be passed as an integer code: similarly Fortran code that wants to generate diagnostic messages could pass an integer code to a C or R wrapper which would convert it to a character string.

It is possible to pass some R objects other than atomic vectors via .C, but this is only supported for historical compatibility: use the .Call or .External interfaces for such objects. Any C/C++ code that includes Rinternals.h should be called via .Call or .External.

.Fortran is primarily intended for Fortran 77 code, and long precedes any support for ‘modern’ Fortran. Nowadays implementations of Fortran support the Fortran 2003 module iso_c_binding, a better way to interface modern Fortran code to R is to use .C and write a C interface using use iso_c_binding.


5.3 dyn.load and dyn.unload

Compiled code to be used with R is loaded as a shared object (Unix-alikes including macOS, see Creating shared objects for more information) or DLL (Windows).

The shared object/DLL is loaded by dyn.load and unloaded by dyn.unload. Unloading is not normally necessary and is not safe in general, but it is needed to allow the DLL to be re-built on some platforms, including Windows. Unloading a DLL and then re-loading a DLL of the same name may not work: Solaris used the first version loaded. A DLL that registers C finalizers, but fails to unregister them when unloaded, may cause R to crash after unloading.

The first argument to both functions is a character string giving the path to the object. Programmers should not assume a specific file extension for the object/DLL (such as .so) but use a construction like

file.path(path1, path2, paste0("mylib", .Platform$dynlib.ext))

for platform independence. On Unix-alike systems the path supplied to dyn.load can be an absolute path, one relative to the current directory or, if it starts with ‘~’, relative to the user’s home directory.

Loading is most often done automatically based on the useDynLib() declaration in the NAMESPACE file, but may be done explicitly via a call to library.dynam. This has the form

library.dynam("libname", package, lib.loc)

where libname is the object/DLL name with the extension omitted. Note that the first argument, chname, should not be package since this will not work if the package is installed under another name.

Under some Unix-alike systems there is a choice of how the symbols are resolved when the object is loaded, governed by the arguments local and now. Only use these if really necessary: in particular using now=FALSE and then calling an unresolved symbol will terminate R unceremoniously.

R provides a way of executing some code automatically when a object/DLL is either loaded or unloaded. This can be used, for example, to register native routines with R’s dynamic symbol mechanism, initialize some data in the native code, or initialize a third party library. On loading a DLL, R will look for a routine within that DLL named R_init_lib where lib is the name of the DLL file with the extension removed. For example, in the command

library.dynam("mylib", package, lib.loc)

R looks for the symbol named R_init_mylib. Similarly, when unloading the object, R looks for a routine named R_unload_lib, e.g., R_unload_mylib. In either case, if the routine is present, R will invoke it and pass it a single argument describing the DLL. This is a value of type DllInfo which is defined in the Rdynload.h file in the R_ext directory.

Note that there are some implicit restrictions on this mechanism as the basename of the DLL needs to be both a valid file name and valid as part of a C entry point (e.g. it cannot contain ‘.’): for portable code it is best to confine DLL names to be ASCII alphanumeric plus underscore. If entry point R_init_lib is not found it is also looked for with ‘.’ replaced by ‘_’.

The following example shows templates for the initialization and unload routines for the mylib DLL.

#include <R_ext/Rdynload.h>

void
R_init_mylib(DllInfo *info)
{
  /* Register routines,
     allocate resources. */
}

void
R_unload_mylib(DllInfo *info)
{
  /* Release resources. */
}

If a shared object/DLL is loaded more than once the most recent version is used.148 More generally, if the same symbol name appears in several shared objects, the most recently loaded occurrence is used. The PACKAGE argument and registration (see the next section) provide good ways to avoid any ambiguity in which occurrence is meant.

On Unix-alikes the paths used to resolve dynamically-linked dependent libraries are fixed (for security reasons) when the process is launched, so dyn.load will only look for such libraries in the locations set by the R shell script (via etc/ldpaths) and in the OS-specific defaults.

Windows allows more control (and less security) over where dependent DLLs are looked for. On all versions this includes the PATH environment variable, but with lowest priority: note that it does not include the directory from which the DLL was loaded. It is possible to add a single path with quite high priority via the DLLpath argument to dyn.load. This is (by default) used by library.dynam to include the package’s libs/x64 directory (on Intel) in the DLL search path.


5.4 Registering native routines

By ‘native’ routine, we mean an entry point in compiled code.

In calls to .C, .Call, .Fortran and .External, R must locate the specified native routine by looking in the appropriate shared object/DLL. By default, R uses the operating-system-specific dynamic loader to lookup the symbol in all149 loaded DLLs and the R executable or libraries it is linked to. Alternatively, the author of the DLL can explicitly register routines with R and use a single, platform-independent mechanism for finding the routines in the DLL. One can use this registration mechanism to provide additional information about a routine, including the number and type of the arguments, and also make it available to R programmers under a different name.

Registering routines has two main advantages: it provides a faster150 way to find the address of the entry point via tables stored in the DLL at compilation time, and it provides a run-time check that the entry point is called with the right number of arguments and, optionally, the right argument types.

To register routines with R, one calls the C routine R_registerRoutines. This is typically done when the DLL is first loaded within the initialization routine R_init_dll name described in dyn.load and dyn.unload. R_registerRoutines takes 5 arguments. The first is the DllInfo object passed by R to the initialization routine. This is where R stores the information about the methods. The remaining 4 arguments are arrays describing the routines for each of the 4 different interfaces: .C, .Call, .Fortran and .External. Each argument is a NULL-terminated array of the element types given in the following table:

.CR_CMethodDef
.CallR_CallMethodDef
.FortranR_FortranMethodDef
.ExternalR_ExternalMethodDef

Currently, the R_ExternalMethodDef type is the same as R_CallMethodDef type and contains fields for the name of the routine by which it can be accessed in R, a pointer to the actual native symbol (i.e., the routine itself), and the number of arguments the routine expects to be passed from R. For example, if we had a routine named myCall defined as

SEXP myCall(SEXP a, SEXP b, SEXP c);

we would describe this as

static const R_CallMethodDef callMethods[]  = {
  {"myCall", (DL_FUNC) &myCall, 3},
  {NULL, NULL, 0}
};

along with any other routines for the .Call interface. For routines with a variable number of arguments invoked via the .External interface, one specifies -1 for the number of arguments which tells R not to check the actual number passed.

Routines for use with the .C and .Fortran interfaces are described with similar data structures, which have one optional additional field for describing the type of each argument. If specified, this field should be an array with the SEXP types describing the expected type of each argument of the routine. (Technically, the elements of the types array are of type R_NativePrimitiveArgType which is just an unsigned integer.) The R types and corresponding type identifiers are provided in the following table:

numericREALSXP
integerINTSXP
logicalLGLSXP
singleSINGLESXP
characterSTRSXP
listVECSXP

Consider a C routine, myC, declared as

void myC(double *x, int *n, char **names, int *status);

We would register it as

static R_NativePrimitiveArgType myC_type[] = {
    REALSXP, INTSXP, STRSXP, LGLSXP
};

static const R_CMethodDef cMethods[] = {
   {"myC", (DL_FUNC) &myC, 4, myC_type},
   {NULL, NULL, 0, NULL}
};

If registering types, check carefully that the number of types matches the number of arguments: as the type array (here myC_type) is passed as a pointer in C, the registration mechanism cannot check this for you.

Note that .Fortran entry points are mapped to lowercase, so registration should use lowercase only.

Having created the arrays describing each routine, the last step is to actually register them with R. We do this by calling R_registerRoutines. For example, if we have the descriptions above for the routines accessed by the .C and .Call we would use the following code:

void
R_init_myLib(DllInfo *info)
{
   R_registerRoutines(info, cMethods, callMethods, NULL, NULL);
}

This routine will be invoked when R loads the shared object/DLL named myLib. The last two arguments in the call to R_registerRoutines are for the routines accessed by .Fortran and .External interfaces. In our example, these are given as NULL since we have no routines of these types.

When R unloads a shared object/DLL, its registrations are removed. There is no other facility for unregistering a symbol.

Examples of registering routines can be found in the different packages in the R source tree (e.g., stats and graphics). Also, there is a brief, high-level introduction in R News (volume 1/3, September 2001, pages 20–23, https://www.r-project.org/doc/Rnews/Rnews_2001-3.pdf).

Once routines are registered, they can be referred to as R objects if this is arranged in the useDynLib call in the package’s NAMESPACE file (see useDynLib). So for example the stats package has

# Refer to all C/Fortran routines by their name prefixed by C_
useDynLib(stats, .registration = TRUE, .fixes = "C_")

in its NAMESPACE file, and then ansari.test’s default methods can contain

        pansari <- function(q, m, n)
            .C(C_pansari, as.integer(length(q)), p = as.double(q),
                as.integer(m), as.integer(n))$p

This avoids the overhead of looking up an entry point each time it is used, and ensures that the entry point in the package is the one used (without a PACKAGE = "pkg" argument).

R_init_ routines are often of the form

void attribute_visible R_init_mypkg(DllInfo *dll)
{
    R_registerRoutines(dll, CEntries, CallEntries, FortEntries,
                       ExternalEntries);
    R_useDynamicSymbols(dll, FALSE);
    R_forceSymbols(dll, TRUE);
...
}

The R_useDynamicSymbols call says the DLL is not to be searched for entry points specified by character strings so .C etc calls will only find registered symbols: the R_forceSymbols call only allows .C etc calls which specify entry points by R objects such as C_pansari (and not by character strings). Each provides some protection against accidentally finding your entry points when people supply a character string without a package, and avoids slowing down such searches. (For the visibility attribute see Controlling visibility.)

In more detail, if a package mypkg contains entry points reg and unreg and the first is registered as a 0-argument .Call routine, we could use (from code in the package)

.Call("reg")
.Call("unreg")

Without or with registration, these will both work. If R_init_mypkg calls R_useDynamicSymbols(dll, FALSE), only the first will work. If in addition to registration the NAMESPACE file contains

useDynLib(mypkg, .registration = TRUE, .fixes = "C_")

then we can call .Call(C_reg). Finally, if R_init_mypkg also calls R_forceSymbols(dll, TRUE), only .Call(C_reg) will work (and not .Call("reg")). This is usually what we want: it ensures that all of our own .Call calls go directly to the intended code in our package and that no one else accidentally finds our entry points. (Should someone need to call our code from outside the package, for example for debugging, they can use .Call(mypkg:::C_reg).)


5.4.1 Speed considerations

Sometimes registering native routines or using a PACKAGE argument can make a large difference. The results can depend quite markedly on the OS (and even if it is 32- or 64-bit), on the version of R and what else is loaded into R at the time.

To fix ideas, first consider x86_64 OS 10.7 and R 2.15.2. A simple .Call function might be

foo <- function(x) .Call("foo", x)

with C code

#include <Rinternals.h>

SEXP foo(SEXP x)
{
    return x;
}

If we compile with by R CMD SHLIB foo.c, load the code by dyn.load("foo.so") and run foo(pi) it took around 22 microseconds (us). Specifying the DLL by

foo2 <- function(x) .Call("foo", x, PACKAGE = "foo")

reduced the time to 1.7 us.

Now consider making these functions part of a package whose NAMESPACE file uses useDynlib(foo). This immediately reduces the running time as "foo" will be preferentially looked for foo.dll. Without specifying PACKAGE it took about 5 us (it needs to fathom out the appropriate DLL each time it is invoked but it does not need to search all DLLs), and with the PACKAGE argument it is again about 1.7 us.

Next suppose the package has registered the native routine foo. Then foo() still has to find the appropriate DLL but can get to the entry point in the DLL faster, in about 4.2 us. And foo2() now takes about 1 us. If we register the symbols in the NAMESPACE file and use

foo3 <- function(x) .Call(C_foo, x)

then the address for the native routine is looked up just once when the package is loaded, and foo3(pi) takes about 0.8 us.

Versions using .C() rather than .Call() took about 0.2 us longer.

These are all quite small differences, but C routines are not uncommonly invoked millions of times for run times of a few microseconds each, and those doing such things may wish to be aware of the differences.

On Linux and Solaris there is a smaller overhead in looking up symbols.

Symbol lookup on Windows used to be far slower, so R maintains a small cache. If the cache is currently empty enough that the symbol can be stored in the cache then the performance is similar to Linux and Solaris: if not it may be slower. R’s own code always uses registered symbols and so these never contribute to the cache: however many other packages do rely on symbol lookup.

In more recent versions of R all the standard packages register native symbols and do not allow symbol search, so in a new session foo() can only look in foo.so and may be as fast as foo2(). This will no longer apply when many contributed packages are loaded, and generally those last loaded are searched first. For example, consider R 3.3.2 on x86_64 Linux. In an empty R session, both foo() and foo2() took about 0.75 us; however after packages igraph and spatstat had been loaded (which loaded another 12 DLLs), foo() took 3.6 us but foo2() still took about 0.80 us. Using registration in a package reduced this to 0.55 us and foo3() took 0.40 us, times which were unchanged when further packages were loaded.


5.4.2 Example: converting a package to use registration

The splines package was converted to use symbol registration in 2001, but we can use it as an example151 of what needs to be done for a small package.

  • Find the relevant entry points. This is somewhat OS-specific, but something like the following should be possible at the OS command-line
    nm -g /path/to/splines.so | grep " T "
    0000000000002670 T _spline_basis
    0000000000001ec0 T _spline_value
    

    This indicates that there are two relevant entry points. (They may or may not have a leading underscore, as here. Fortran entry points will have a trailing underscore on all current platforms.) Check in the R code that they are called by the package and how: in this case they are used by .Call.

    Alternatively, examine the package’s R code for all .C, .Fortran, .Call and .External calls.

  • Construct the registration table. First write skeleton registration code, conventionally in file src/init.c (or at the end of the only C source file in the package: if included in a C++ file the ‘R_init’ function would need to be declared extern "C"):
    #include <stdlib.h> // for NULL
    #include <R_ext/Rdynload.h>
    
    #define CALLDEF(name, n)  {#name, (DL_FUNC) &name, n}
    
    static const R_CallMethodDef R_CallDef[] = {
       CALLDEF(spline_basis, ?),
       CALLDEF(spline_value, ?),
       {NULL, NULL, 0}
    };
    
    void R_init_splines(DllInfo *dll)
    {
        R_registerRoutines(dll, NULL, R_CallDef, NULL, NULL);
    }
    

    and then replace the ? in the skeleton with the actual numbers of arguments. You will need to add declarations (also known as ‘prototypes’) of the functions unless appending to the only C source file. Some packages will already have these in a header file, or you could create one and include it in init.c, for example splines.h containing

    #include <Rinternals.h> // for SEXP
    extern SEXP spline_basis(SEXP knots, SEXP order, SEXP xvals, SEXP derivs);
    extern SEXP spline_value(SEXP knots, SEXP coeff, SEXP order, SEXP x, SEXP deriv);
    

    Tools are available to extract declarations, at least for C and C++ code: see the help file for package_native_routine_registration_skeleton in package tools. Here we could have used

    cproto -I/path/to/R/include -e splines.c
    

    For examples of registering other types of calls, see packages graphics and stats. In particular, when registering entry points for .Fortran one needs declarations as if called from C, such as

    #include <R_ext/RS.h>
    void F77_NAME(supsmu)(int *n, double *x, double *y,
                          double *w, int *iper, double *span, double *alpha,
                          double *smo, double *sc, double *edf);
    

    gfortran 8.4, 9.2 and later can help generate such prototypes with its flag -fc-prototypes-external (although one will need to replace the hard-coded trailing underscore with the F77_NAME macro).

    One can get away with inaccurate argument lists in the declarations: it is easy to specify the arguments for .Call (all SEXP) and .External (one SEXP) and as the arguments for .C and .Fortran are all pointers, specifying them as void * suffices. (For most platforms one can omit all the arguments, although link-time optimization will warn, as will compilers set up to warn on strict prototypes – and C23 requires correct arguments.)

    Using -fc-prototypes-external will give a prototype using int_least32_t *lgl for Fortran LOGICAL LGL, but this is not portable and traditionally it has been assumed that the C/C++ equivalent was int *lgl. If adding a declaration just to register a .Fortran call, the most portable version is void *lgl.

  • (Optional but highly recommended.) Restrict .Call etc to use the symbols you chose to register by editing src/init.c to contain
    void R_init_splines(DllInfo *dll)
    {
        R_registerRoutines(dll, NULL, R_CallDef, NULL, NULL);
        R_useDynamicSymbols(dll, FALSE);
    }
    

A skeleton for the steps so far can be made using package_native_routine_registration_skeleton in package tools. This will optionally create declarations based on the usage in the R code.

The remaining steps are optional but recommended.

  • Edit the NAMESPACE file to create R objects for the registered symbols:
    useDynLib(splines, .registration = TRUE, .fixes = "C_")
    
  • Find all the relevant calls in the R code and edit them to use the R objects. This entailed changing the lines
    temp <- .Call("spline_basis", knots, ord, x, derivs, PACKAGE = "splines")
    y[accept] <- .Call("spline_value", knots, coeff, ord, x[accept], deriv, PACKAGE = "splines")
    y = .Call("spline_value", knots, coef(object), ord, x, deriv, PACKAGE = "splines")
    

    to

    temp <- .Call(C_spline_basis, knots, ord, x, derivs)
    y[accept] <- .Call(C_spline_value, knots, coeff, ord, x[accept], deriv)
    y = .Call(C_spline_value, knots, coef(object), ord, x, deriv)
    

    Check that there is no exportPattern directive which unintentionally exports the newly created R objects.

  • Restrict .Call to use the R symbols by editing src/init.c to contain
    void R_init_splines(DllInfo *dll)
    {
        R_registerRoutines(dll, NULL, R_CallDef, NULL, NULL);
        R_useDynamicSymbols(dll, FALSE);
        R_forceSymbols(dll, TRUE);
    }
    
  • Consider visibility. On some OSes we can hide entry points from the loader, which precludes any possible name clashes and calling them accidentally (usually with incorrect arguments and crashing the R process). If we repeat the first step we now see
    nm -g /path/to/splines.so | grep " T "
    0000000000002e00 T _R_init_splines
    00000000000025e0 T _spline_basis
    0000000000001e20 T _spline_value
    

    If there were any entry points not intended to be used by the package we should try to avoid exporting them, for example by making them static. Now that the two relevant entry points are only accessed via the registration table, we can hide them. There are two ways to do so on some152 Unix-alikes. We can hide individual entry points via

    #include <R_ext/Visibility.h>
    
    SEXP attribute_hidden
    spline_basis(SEXP knots, SEXP order, SEXP xvals, SEXP derivs)
    ...
    
    SEXP attribute_hidden
    spline_value(SEXP knots, SEXP coeff, SEXP order, SEXP x, SEXP deriv)
    ...
    

    Alternatively, we can change the default visibility for all C symbols by including

    PKG_CFLAGS = $(C_VISIBILITY)
    

    in src/Makevars, and then we need to allow registration by declaring R_init_splines to be visible:

    #include <R_ext/Visibility.h>
    
    void attribute_visible
    R_init_splines(DllInfo *dll)
    ...
    

    See Controlling visibility for more details, including using Fortran code and ways to restrict visibility on Windows.

  • We end up with a file src/init.c containing
    #include <stdlib.h>
    #include <R_ext/Rdynload.h>
    #include <R_ext/Visibility.h>  // optional
    
    #include "splines.h"
    
    #define CALLDEF(name, n)  {#name, (DL_FUNC) &name, n}
    
    static const R_CallMethodDef R_CallDef[] = {
        CALLDEF(spline_basis, 4),
        CALLDEF(spline_value, 5),
        {NULL, NULL, 0}
    };
    
    void
    attribute_visible  // optional
    R_init_splines(DllInfo *dll)
    {
        R_registerRoutines(dll, NULL, R_CallDef, NULL, NULL);
        R_useDynamicSymbols(dll, FALSE);
        R_forceSymbols(dll, TRUE);
    }
    

5.4.3 Linking to native routines in other packages

In addition to registering C routines to be called by R, it can at times be useful for one package to make some of its C routines available to be called by C code in another package. The interface consists of two routines declared in header R_ext/Rdynload.h as

void R_RegisterCCallable(const char *package, const char *name,
                         DL_FUNC fptr);
DL_FUNC R_GetCCallable(const char *package, const char *name);

A package packA that wants to make a C routine myCfun available to C code in other packages would include the call

R_RegisterCCallable("packA", "myCfun", myCfun);

in its initialization function R_init_packA. A package packB that wants to use this routine would retrieve the function pointer with a call of the form

p_myCfun = R_GetCCallable("packA", "myCfun");

As the type DL_FUNC is only appropriate for functions with no arguments, other users will need to cast to an appropriate type. For example

typedef SEXP (*na_omit_xts_func) (SEXP x);
...
  na_omit_xts_func fun = (na_omit_xts_func) R_GetCCallable("xts", "na_omit_xts");
  return fun(x);

The author of packB is responsible for ensuring that p_myCfun has an appropriate declaration. In the future R may provide some automated tools to simplify exporting larger numbers of routines.

A package that wishes to make use of header files in other packages needs to declare them as a comma-separated list in the field ‘LinkingTo’ in the DESCRIPTION file. This then arranges for the include directories in the installed linked-to packages to be added to the include paths for C and C++ code.

It must specify153Imports’ or ‘Depends’ of those packages, for they have to be loaded154 prior to this one (so the path to their compiled code has been registered).

CRAN examples of the use of this mechanism include coxme linking to bdsmatrix and xts linking to zoo.

NB: this mechanism is fragile, as changes to the interface provided by packA have to be recognised by packB. The consequences of not doing so have included serious corruption to the memory pool of the R session. Either packB has to depend on the exact version of packA or there needs to be a mechanism for packB to test at runtime the version of packA it is linked to matches that it was compiled against.

On rare occasions in can be useful for C code in one package to dynamically look up the address in another package. This can be done using R_FindSymbol:

DL_FUNC R_FindSymbol(char const *name, char const *pkg,
                     R_RegisteredNativeSymbol *symbol);

5.5 Creating shared objects

Shared objects for loading into R can be created using R CMD SHLIB. This accepts as arguments a list of files which must be object files (with extension .o) or sources for C, C++, Fortran, Objective C or Objective C++ (with extensions .c, .cc or .cpp, .f (fixed-form Fortran), .f90 or .f95 (free-form), .m, and .mm or .M, respectively), or commands to be passed to the linker. See R CMD SHLIB --help (or the R help for SHLIB) for usage information. Note that files intended for the Fortran pre-processor with extension .F are not accepted.

If compiling the source files does not work “out of the box”, you can specify additional flags by setting some of the variables PKG_CPPFLAGS (for the C/C++ preprocessor, mainly ‘-I’, ‘-D’ and ‘-U’ flags), PKG_CFLAGS, PKG_CXXFLAGS, PKG_FFLAGS, PKG_OBJCFLAGS, and PKG_OBJCXXFLAGS (for the C, C++, Fortran, Objective C, and Objective C++ compilers, respectively) in the file Makevars in the compilation directory (or, of course, create the object files directly from the command line). Similarly, variable PKG_LIBS in Makevars can be used for additional ‘-l’ and ‘-L’ flags to be passed to the linker when building the shared object. (Supplying linker commands as arguments to R CMD SHLIB will take precedence over PKG_LIBS in Makevars.)

It is possible to arrange to include compiled code from other languages by setting the macro ‘OBJECTS’ in file Makevars, together with suitable rules to make the objects.

Flags that are already set (for example in file etcR_ARCH/Makeconf) can be overridden by the environment variable MAKEFLAGS (at least for systems using a POSIX-compliant make), as in (Bourne shell syntax)

MAKEFLAGS="CFLAGS=-O3" R CMD SHLIB *.c

It is also possible to set such variables in personal Makevars files, which are read after the local Makevars and the system makefiles or in a site-wide Makevars.site file. See Customizing package compilation in R Installation and Administration for more information.

Note that as R CMD SHLIB uses Make, it will not remake a shared object just because the flags have changed, and if test.c and test.f both exist in the current directory

R CMD SHLIB test.f

will compile test.c!

If the src subdirectory of an add-on package contains source code with one of the extensions listed above or a file Makevars but not a file Makefile, R CMD INSTALL creates a shared object (for loading into R through useDynlib in the NAMESPACE, or in the .onLoad function of the package) using the R CMD SHLIB mechanism. If file Makevars exists it is read first, then the system makefile and then any personal Makevars files.

If the src subdirectory of package contains a file Makefile, this is used by R CMD INSTALL in place of the R CMD SHLIB mechanism. make is called with makefiles R_HOME/etcR_ARCH/Makeconf, src/Makefile and any personal Makevars files (in that order). The first target found in src/Makefile is used.

It is better to make use of a Makevars file rather than a Makefile: the latter should be needed only exceptionally.

Under Windows the same commands work, but Makevars.win will be used in preference to Makevars, and only src/Makefile.win will be used by R CMD INSTALL with src/Makefile being ignored. Since R 4.2.0, Makevars.ucrt will be used in preference to Makevars.win and src/Makefile.ucrt will be used in preference to src/Makefile.win. For past experiences of building DLLs with a variety of compilers, see file ‘README.packages’. Under Windows you can supply an exports definitions file called dllname-win.def: otherwise all entry points in objects (but not libraries) supplied to R CMD SHLIB will be exported from the DLL. An example is stats-win.def for the stats package: a CRAN example in package fastICA.

If you feel tempted to read the source code and subvert these mechanisms, please resist. Far too much developer time has been wasted in chasing down errors caused by failures to follow this documentation, and even more by package authors demanding explanations as to why their packages no longer work. In particular, undocumented environment or make variables are not for use by package writers and are subject to change without notice.


5.6 Interfacing C++ code

Suppose we have the following hypothetical C++ library, consisting of the two files X.h and X.cpp, and implementing the two classes X and Y which we want to use in R.

// X.h

class X {
public: X (); ~X ();
};

class Y {
public: Y (); ~Y ();
};
// X.cpp

#include <R.h>
#include "X.h"

static Y y;

X::X()  { REprintf("constructor X\n"); }
X::~X() { REprintf("destructor X\n");  }
Y::Y()  { REprintf("constructor Y\n"); }
Y::~Y() { REprintf("destructor Y\n");  }

To use with R, the only thing we have to do is writing a wrapper function and ensuring that the function is enclosed in

extern "C" {

}

For example,

// X_main.cpp:

#include "X.h"

extern "C" {

void X_main () {
  X x;
}

} // extern "C"

Compiling and linking should be done with the C++ compiler-linker (rather than the C compiler-linker or the linker itself); otherwise, the C++ initialization code (and hence the constructor of the static variable Y) are not called. On a properly configured system, one can simply use

R CMD SHLIB X.cpp X_main.cpp

to create the shared object, typically X.so (the file name extension may be different on your platform). Now starting R yields

R version 2.14.1 Patched (2012-01-16 r58124)
Copyright (C) 2012 The R Foundation for Statistical Computing
...
Type    "q()" to quit R.

R> dyn.load(paste("X", .Platform$dynlib.ext, sep = ""))
constructor Y
R> .C("X_main")
constructor X
destructor X
list()
R> q()
Save workspace image? [y/n/c]: y
destructor Y

The R for Windows FAQ (rw-FAQ) contains details of how to compile this example under Windows.

Earlier versions of this example used C++ iostreams: this is best avoided. There is no guarantee that the output will appear in the R console, and indeed it will not on the R for Windows console. Use R code or the C entry points (see Printing) for all I/O if at all possible. Examples have been seen where merely loading a DLL that contained calls to C++ I/O upset R’s own C I/O (for example by resetting buffers on open files).

Most R header files can be included within C++ programs but they should not be included within an extern "C" block (as they include system headers155).

5.6.1 External C++ code

Quite a lot of external C++ software is header-only (e.g. most of the Boost ‘libraries’ including all those supplied by package BH, and most of Armadillo as supplied by package RcppArmadillo) and so is compiled when an R package which uses it is installed. This causes few problems.

A small number of external libraries used in R packages have a C++ interface to a library of compiled code, e.g. packages sf and rjags. This raises many more problems! The C++ interface uses name-mangling and the ABI156 may depend on the compiler, version and even C++ defines157, so requires the package C++ code to be compiled in exactly the same way as the library (and what that was is often undocumented).

Even fewer external libraries use C++ internally but present a C interface, such as GEOS used by sf and other packages.. These require the C++ runtime library to be linked into the package’s shared object/DLL, and this is best done by including a dummy C++ file in the package sources.

There is a trend to link to the C++ interfaces offered by C software such as hdf5, pcre and ImageMagick. Their C interfaces are much preferred for portability (and can be used from C++ code). Also, the C++ interfaces are often optional in the software build or packaged separately and so users installing from package sources are less likely to already have them installed.


5.7 Fortran I/O

We have already warned against the use of C++ iostreams not least because output is not guaranteed to appear on the R console, and this warning applies equally to Fortran output to units * and 6. See Printing from Fortran, which describes workarounds.

When R was first developed, most Fortran compilers implemented I/O on top of the C I/O system and so the two interworked successfully. This was true of g77, but no longer of gfortran as used in gcc 4 and later. In particular, any package that makes use of Fortran I/O will when compiled on Windows interfere with C I/O: when the Fortran I/O support code is initialized (typically when the package is loaded) the C stdout and stderr are switched to LF line endings. (Function init in file src/modules/lapack/init_win.c shows how to mitigate this. In a package this would look something like

#ifdef _WIN32
# include <fcntl.h>
#endif

void R_init_mypkgname(DllInfo *dll)
{
    // Native symbol registration calls

#ifdef _WIN32
    // gfortran I/O initialization sets these to _O_BINARY
    setmode(1, _O_TEXT); /* stdout */
    setmode(2, _O_TEXT); /* stderr */
#endif
}

in the file used for native symbol registration.)


5.8 Linking to other packages

It is not in general possible to link a DLL in package packA to a DLL provided by package packB (for the security reasons mentioned in dyn.load and dyn.unload, and also because some platforms distinguish between shared objects and dynamic libraries), but it is on Windows.

Note that there can be tricky versioning issues here, as package packB could be re-installed after package packA — it is desirable that the API provided by package packB remains backwards-compatible.

Shipping a static library in package packB for other packages to link to avoids most of the difficulties.


5.8.1 Unix-alikes

It is possible to link a shared object in package packA to a library provided by package packB under limited circumstances on a Unix-alike OS. There are severe portability issues, so this is not recommended for a distributed package.

This is easiest if packB provides a static library packB/lib/libpackB.a. (Note using directory lib rather than libs is conventional, and architecture-specific sub-directories may be needed and are assumed in the sample code below. The code in the static library will need to be compiled with PIC flags on platforms where it matters.) Then as the code from package packB is incorporated when package packA is installed, we only need to find the static library at install time for package packA. The only issue is to find package packB, and for that we can ask R by something like (long lines broken for display here)

PKGB_PATH=`echo 'library(packB);
  cat(system.file("lib",  package="packB", mustWork=TRUE))' \
 | "${R_HOME}/bin/R" --vanilla --no-echo`
PKG_LIBS="$(PKGB_PATH)$(R_ARCH)/libpackB.a"

For a dynamic library packB/lib/libpackB.so (packB/lib/libpackB.dylib on macOS: note that you cannot link to a shared object, .so, on that platform) we could use

PKGB_PATH=`echo 'library(packB);
  cat(system.file("lib", package="packB", mustWork=TRUE))' \
 | "${R_HOME}/bin/R" --vanilla --no-echo`
PKG_LIBS=-L"$(PKGB_PATH)$(R_ARCH)" -lpackB

This will work for installation, but very likely not when package packB is loaded, as the path to package packB’s lib directory is not in the ld.so158 search path. You can arrange to put it there before R is launched by setting (on some platforms) LD_RUN_PATH or LD_LIBRARY_PATH or adding to the ld.so cache (see man ldconfig). On platforms that support it, the path to the directory containing the dynamic library can be hardcoded at install time (which assumes that the location of package packB will not be changed nor the package updated to a changed API). On systems with the gcc or clang and the GNU linker (e.g. Linux) and some others this can be done by e.g.

PKGB_PATH=`echo 'library(packB);
  cat(system.file("lib", package="packB", mustWork=TRUE)))' \
 | "${R_HOME}/bin/R" --vanilla --no-echo`
PKG_LIBS=-L"$(PKGB_PATH)$(R_ARCH)" -Wl,-rpath,"$(PKGB_PATH)$(R_ARCH)" -lpackB

Some other systems (e.g. Solaris with its native linker) use -Rdir rather than -rpath,dir (and this is accepted by the compiler as well as the linker).

It may be possible to figure out what is required semi-automatically from the result of R CMD libtool --config (look for ‘hardcode’).

Making headers provided by package packB available to the code to be compiled in package packA can be done by the LinkingTo mechanism (see Registering native routines).


5.8.2 Windows

Suppose package packA wants to make use of compiled code provided by packB in DLL packB/libs/exB.dll, possibly the package’s DLL packB/libs/packB.dll. (This can be extended to linking to more than one package in a similar way.) There are three issues to be addressed:

  • Making headers provided by package packB available to the code to be compiled in package packA.

    This is done by the LinkingTo mechanism (see Registering native routines).

  • preparing packA.dll to link to packB/libs/exB.dll.

    This needs an entry in Makevars.win or Makevars.ucrt of the form

    PKG_LIBS= -L<something> -lexB
    

    and one possibility is that <something> is the path to the installed pkgB/libs directory. To find that we need to ask R where it is by something like

    PKGB_PATH=`echo 'library(packB);
      cat(system.file("libs", package="packB", mustWork=TRUE))' \
     | rterm --vanilla --no-echo`
    PKG_LIBS= -L"$(PKGB_PATH)$(R_ARCH)" -lexB
    

    Another possibility is to use an import library, shipping with package packA an exports file exB.def. Then Makevars.win (or Makevars.ucrt) could contain

    PKG_LIBS= -L. -lexB
    
    all: $(SHLIB) before
    
    before: libexB.dll.a
    libexB.dll.a: exB.def
    

    and then installing package packA will make and use the import library for exB.dll. (One way to prepare the exports file is to use pexports.exe.)

  • loading packA.dll which depends on exB.dll.

    If exB.dll was used by package packB (because it is in fact packB.dll or packB.dll depends on it) and packB has been loaded before packA, then nothing more needs to be done as exB.dll will already be loaded into the R executable. (This is the most common scenario.)

    More generally, we can use the DLLpath argument to library.dynam to ensure that exB.dll is found, for example by setting

    library.dynam("packA", pkg, lib,
                  DLLpath = system.file("libs", package="packB"))
    

    Note that DLLpath can only set one path, and so for linking to two or more packages you would need to resort to setting environment variable PATH.


5.9 Handling R objects in C

Using C code to speed up the execution of an R function is often very fruitful. Traditionally this has been done via the .C function in R. However, if a user wants to write C code using internal R data structures, then that can be done using the .Call and .External functions. The syntax for the calling function in R in each case is similar to that of .C, but the two functions have different C interfaces. Generally the .Call interface is simpler to use, but .External is a little more general.

A call to .Call is very similar to .C, for example

.Call("convolve2", a, b)

The first argument should be a character string giving a C symbol name of code that has already been loaded into R. Up to 65 R objects can passed as arguments. The C side of the interface is

#include <R.h>
#include <Rinternals.h>

SEXP convolve2(SEXP a, SEXP b)
 ...

A call to .External is almost identical

.External("convolveE", a, b)

but the C side of the interface is different, having only one argument

#include <R.h>
#include <Rinternals.h>

SEXP convolveE(SEXP args)
 ...

Here args is a LISTSXP, a Lisp-style pairlist from which the arguments can be extracted.

In each case the R objects are available for manipulation via a set of functions and macros defined in the header file Rinternals.h or some S-compatibility macros159 See Interface functions .Call and .External for details on .Call and .External.

Before you decide to use .Call or .External, you should look at other alternatives. First, consider working in interpreted R code; if this is fast enough, this is normally the best option. You should also see if using .C is enough. If the task to be performed in C is simple enough involving only atomic vectors and requiring no call to R, .C suffices. A great deal of useful code was written using just .C before .Call and .External were available. These interfaces allow much more control, but they also impose much greater responsibilities so need to be used with care. Neither .Call nor .External copy their arguments: you should treat arguments you receive through these interfaces as read-only.

To handle R objects from within C code we use the macros and functions that have been used to implement the core parts of R. A public160 subset of these is defined in the header file Rinternals.h in the directory R_INCLUDE_DIR (default R_HOME/include) that should be available on any R installation.

A substantial amount of R, including the standard packages, is implemented using the functions and macros described here, so the R source code provides a rich source of examples and “how to do it”: do make use of the source code for inspirational examples.

It is necessary to know something about how R objects are handled in C code. All the R objects you will deal with will be handled with the type SEXP161, which is a pointer to a structure with typedef SEXPREC. Think of this structure as a variant type that can handle all the usual types of R objects, that is vectors of various modes, functions, environments, language objects and so on. The details are given later in this section and in R Internal Structures in R Internals, but for most purposes the programmer does not need to know them. Think rather of a model such as that used by Visual Basic, in which R objects are handed around in C code (as they are in interpreted R code) as the variant type, and the appropriate part is extracted for, for example, numerical calculations, only when it is needed. As in interpreted R code, much use is made of coercion to force the variant object to the right type.


5.9.1 Handling the effects of garbage collection

We need to know a little about the way R handles memory allocation. The memory allocated for R objects is not freed by the user; instead, the memory is from time to time garbage collected. That is, some or all of the allocated memory not being used is freed or marked as re-usable.

The R object types are represented by a C structure defined by a typedef SEXPREC in Rinternals.h. It contains several things among which are pointers to data blocks and to other SEXPRECs. A SEXP is simply a pointer to a SEXPREC.

If you create an R object in your C code, you must tell R that you are using the object by using the PROTECT macro on a pointer to the object. This tells R that the object is in use so it is not destroyed during garbage collection. Notice that it is the object which is protected, not the pointer variable. It is a common mistake to believe that if you invoked PROTECT(p) at some point then p is protected from then on, but that is not true once a new object is assigned to p.

Protecting an R object automatically protects all the R objects pointed to in the corresponding SEXPREC, for example all elements of a protected list are automatically protected.

The programmer is solely responsible for housekeeping the calls to PROTECT. There is a corresponding macro UNPROTECT that takes as argument an int giving the number of objects to unprotect when they are no longer needed. The protection mechanism is stack-based, so UNPROTECT(n) unprotects the last n objects which were protected. The calls to PROTECT and UNPROTECT must balance when the user’s code returns and should balance in all functions. R will warn about "stack imbalance in .Call" (or .External) if the housekeeping is wrong.

Here is a small example of creating an R numeric vector in C code:

#include <R.h>
#include <Rinternals.h>

    SEXP ab;
      ....
    ab = PROTECT(RF_allocVector(REALSXP, 2));
    REAL(ab)[0] = 123.45;
    REAL(ab)[1] = 67.89;
    UNPROTECT(1);

Now, the reader may ask how the R object could possibly get removed during those manipulations, as it is just our C code that is running. As it happens, we can do without the protection in this example, but in general we do not know (nor want to know) what is hiding behind the R macros and functions we use, and any of them might cause memory to be allocated, hence garbage collection and hence our object ab to be removed. It is usually wise to err on the side of caution and assume that any of the R macros and functions might remove the object.

In some cases it is necessary to keep better track of whether protection is really needed. Be particularly aware of situations where a large number of objects are generated. The pointer protection stack has a fixed size (default 10,000) and can become full. It is not a good idea then to just PROTECT everything in sight and UNPROTECT several thousand objects at the end. It will almost invariably be possible to either assign the objects as part of another object (which automatically protects them) or unprotect them immediately after use.

There is a less-used macro UNPROTECT_PTR(s) that unprotects the object pointed to by the SEXP s, even if it is not the top item on the pointer protection stack. This macro was introduced for use in the parser, where the code interfacing with the R heap is generated and the generator cannot be configured to insert proper calls to PROTECT and UNPROTECT. However, UNPROTECT_PTR is dangerous to use in combination with UNPROTECT when the same object has been protected multiple times. It has been superseded by multi-set based functions R_PreserveInMSet and R_ReleaseFromMSet, which protect objects in a multi-set created by R_NewPreciousMSet and typically itself protected using PROTECT. These functions should not be needed outside parsers.

Sometimes an object is changed (for example duplicated, coerced or grown) yet the current value needs to be protected. For these cases PROTECT_WITH_INDEX saves an index of the protection location that can be used to replace the protected value using REPROTECT. For example (from the internal code for optim)

    PROTECT_INDEX ipx;

    ....
    PROTECT_WITH_INDEX(s = Rf_eval(OS->R_fcall, OS->R_env), &ipx);
    REPROTECT(s = Rf_coerceVector(s, REALSXP), ipx);

Note that it is dangerous to mix UNPROTECT_PTR also with PROTECT_WITH_INDEX, as the former changes the protection locations of objects that were protected after the one being unprotected.

There is another way to avoid the effects of garbage collection: a call to R_PreserveObject adds an object to an internal list of objects not to be collected, and a subsequent call to R_ReleaseObject removes it from that list. This provides a way for objects which are not returned as part of R objects to be protected across calls to compiled code: on the other hand it becomes the user’s responsibility to release them when they are no longer needed (and this often requires the use of a finalizer). It is less efficient than the normal protection mechanism, and should be used sparingly.

For functions from packages as well as R to safely co-operate in protecting objects, certain rules have to be followed:

  • Pointer-protection balance. Calls to PROTECT and UNPROTECT should balance in each function. A function may only call UNPROTECT or REPROTECT on objects it has itself protected. Note that the pointer protection stack balance is restored automatically on non-local transfer of control (See Condition handling and cleanup code.), as if a call to UNPROTECT was invoked with the right argument.
  • Caller protection. It is the responsibility of the caller that all arguments passed to a function are protected and will stay protected for the whole execution of the callee. Typically this is achieved by PROTECT and UNPROTECT calls.
  • Protecting return values. Any R objects returned from a function are unprotected (the callee must maintain pointer-protection balance), and hence should be protected immediately by the caller. To be safe against future code changes, assume that any R object returned from any function may need protection. Note that even when conceptually returning an existing protected object, that object may be duplicated.
  • All functions/macros allocate. To be safe against future code changes, assume that any function or macro may allocate and hence garbage collector may run and destroy unprotected objects.

It is always safe and recommended to follow those rules. In fact, several R functions and macros protect their own arguments and some functions do not allocate or do not allocate when used in a certain way, but that is subject to change, so relying on that may be fragile. PROTECT and PROTECT_WITH_INDEX can be safely called with unprotected arguments and UNPROTECT does not allocate.


5.9.2 Allocating storage

For many purposes it is sufficient to allocate R objects and manipulate those. There are quite a few Rf_allocXxx functions defined in Rinternals.h—you may want to explore them.

One that is commonly used is Rf_allocVector, the C-level equivalent of R-level vector() and its wrappers such as integer() and character(). One distinction is that whereas the R functions always initialize the elements of the vector, Rf_allocVector only does so for lists, expressions and character vectors (the cases where the elements are themselves R objects). Other useful allocation functions are Rf_alloc3DArray, Rf_allocArray, and Rf_allocMatrix.

At times it can be useful to allocate a larger initial result vector and resize it to a shorter length if that is sufficient. The functions Rf_lengthgets and Rf_xlengthgets accomplish this; they are analogous to using length(x) <- n in R. Typically these functions return a freshly allocated object, but in some cases they may re-use the supplied object.

When creating new result objects it can be useful to fill them in with values from an existing object. The functions Rf_copyVector and Rf_copyMatrix can be used for this. Rf_copyMostAttributes can also simplify setting up a result object; it is used internally for results of arithmetic operations.

If storage is required for C objects during the calculations this is best allocated by calling R_alloc; see Memory allocation. All of these memory allocation routines do their own error-checking, so the programmer may assume that they will raise an error and not return if the memory cannot be allocated.


5.9.3 Details of R types

Users of the Rinternals.h macros will need to know how the R types are known internally. The different R data types are represented in C by SEXPTYPE. Some of these are familiar from R and some are internal data types. The usual R object modes are given in the table.

SEXPTYPER equivalent
REALSXPnumeric with storage mode double
INTSXPinteger
CPLXSXPcomplex
LGLSXPlogical
STRSXPcharacter
VECSXPlist (generic vector)
LISTSXPpairlist
DOTSXPa ‘’ object
NILSXPNULL
SYMSXPname/symbol
CLOSXPfunction or function closure
ENVSXPenvironment

Among the important internal SEXPTYPEs are LANGSXP, CHARSXP, PROMSXP, etc. (N.B.: although it is possible to return objects of internal types, it is unsafe to do so as assumptions are made about how they are handled which may be violated at user-level evaluation.) More details are given in R Internal Structures in R Internals.

Unless you are very sure about the type of the arguments, the code should check the data types. Sometimes it may also be necessary to check data types of objects created by evaluating an R expression in the C code. You can use functions like Rf_isReal, Rf_isInteger and Rf_isString to do type checking. Other such functions declared in the header file Rinternals.h include Rf_iisNull, Rf_iisSymbol, Rf_iisLogical, Rf_iisComplex, Rf_iisExpression, and Rf_iisEnvironment. All of these take a SEXP as argument and return 1 or 0 to indicate TRUE or FALSE.

What happens if the SEXP is not of the correct type? Sometimes you have no other option except to generate an error. You can use the function Rf_error for this. It is usually better to coerce the object to the correct type. For example, if you find that an SEXP is of the type INTEGER, but you need a REAL object, you can change the type by using

newSexp = PROTECT(Rf_coerceVector(oldSexp, REALSXP));

Protection is needed as a new object is created; the object formerly pointed to by the SEXP is still protected but now unused.162

All the coercion functions do their own error-checking, and generate NAs with a warning or stop with an error as appropriate.

Note that these coercion functions are not the same as calling as.numeric (and so on) in R code, as they do not dispatch on the class of the object. Thus it is normally preferable to do the coercion in the calling R code.

So far we have only seen how to create and coerce R objects from C code, and how to extract the numeric data from numeric R vectors. These can suffice to take us a long way in interfacing R objects to numerical algorithms, but we may need to know a little more to create useful return objects.


5.9.4 Attributes

Many R objects have attributes: some of the most useful are classes and the dim and dimnames that mark objects as matrices or arrays. It can also be helpful to work with the names attribute of vectors.

To illustrate this, let us write code to take the outer product of two vectors (which outer and %o% already do). As usual the R code is simple

out <- function(x, y)
{
    storage.mode(x) <- storage.mode(y) <- "double"
    .Call("out", x, y)
}

where we expect x and y to be numeric vectors (possibly integer), possibly with names. This time we do the coercion in the calling R code.

C code to do the computations is

#include <R.h>
#include <Rinternals.h>

SEXP out(SEXP x, SEXP y)
{
    int nx = Rf_length(x), ny = Rf_length(y);
    SEXP ans = PROTECT(Rf_allocMatrix(REALSXP, nx, ny));
    double *rx = REAL(x), *ry = REAL(y), *rans = REAL(ans);
    for(int i = 0; i < nx; i++) {
        double tmp = rx[i];
        for(int j = 0; j < ny; j++)
            rans[i + nx*j] = tmp * ry[j];
    }
    UNPROTECT(1);
    return ans;
}

Note the way REAL is used: as it is a function call it can be considerably faster to store the result and index that.

However, we would like to set the dimnames of the result. We can use

#include <R.h>
#include <Rinternals.h>

SEXP out(SEXP x, SEXP y)
{
    int nx = Rf_length(x), ny = Rf_length(y);
    SEXP ans = PROTECT(Rf_allocMatrix(REALSXP, nx, ny));
    double *rx = REAL(x), *ry = REAL(y), *rans = REAL(ans);

    for(int i = 0; i < nx; i++) {
      double tmp = rx[i];
      for(int j = 0; j < ny; j++)
        rans[i + nx*j] = tmp * ry[j];
    }

    SEXP dimnames = PROTECT(Rf_allocVector(VECSXP, 2));
    SET_VECTOR_ELT(dimnames, 0, Rf_getAttrib(x, R_NamesSymbol));
    SET_VECTOR_ELT(dimnames, 1, Rf_getAttrib(y, R_NamesSymbol));
    Rf_setAttrib(ans, R_DimNamesSymbol, dimnames);

    UNPROTECT(2);
    return ans;
}

This example introduces several new features. The Rf_getAttrib and Rf_setAttrib functions get and set individual attributes. Their second argument is a SEXP defining the name in the symbol table of the attribute we want; these and many such symbols are defined in the header file Rinternals.h.

There are shortcuts here too: the functions namesgets, dimgets and dimnamesgets are the internal versions of the default methods of names<-, dim<- and dimnames<- (for vectors and arrays), and there are functions such as Rf_GetColNames, Rf_GetRowNames, Rf_GetMatrixDimnames and Rf_GetArrayDimnames.

What happens if we want to add an attribute that is not pre-defined? We need to add a symbol for it via a call to Rf_install. Suppose for illustration we wanted to add an attribute "version" with value 3.0. We could use

SEXP version;
version = PROTECT(Rf_allocVector(REALSXP, 1));
REAL(version)[0] = 3.0;
Rf_setAttrib(ans, Rf_install("version"), version);
UNPROTECT(1);

Using Rf_install when it is not needed is harmless and provides a simple way to retrieve the symbol from the symbol table if it is already installed. However, the lookup takes a non-trivial amount of time, so consider code such as

static SEXP VerSymbol = NULL;
...
if (VerSymbol == NULL) VerSymbol = Rf_install("version");

if it is to be done frequently.

This example can be simplified by another convenience function:

SEXP version = PROTECT(Rf_ScalarReal(3.0));
Rf_setAttrib(ans, Rf_install("version"), version);
UNPROTECT(1);

If a result is to be a vector with all elements named, then Rf_mkNamed can be used to allocate a vector of a specified type. Names are provided as a C vector of strings terminated by an empty string:

const char *nms[] = {"xi", "yi", "zi", ""};
Rf_mkNamed(VECSXP, nms);

Symbols can also be installed or retrieved based on a name in a CHARSXP object using either Rf_installChar or Rf_installTrChar. These used to differ in handling character encoding but have been identical since R 4.0.0.


5.9.5 Classes

In R the class is just the attribute named "class" so it can be handled as such, but there is a shortcut Rf_classgets. Suppose we want to give the return value in our example the class "mat". We can use

#include <R.h>
#include <Rinternals.h>
      ....
    SEXP ans, dim, dimnames, class;
      ....
    class = PROTECT(Rf_allocVector(STRSXP, 1));
    SET_STRING_ELT(class, 0, Rf_mkChar("mat"));
    Rf_classgets(ans, class);
    UNPROTECT(4);
    return ans;
}

As the value is a character vector, we have to know how to create that from a C character array, which we do using the function mkChar.


5.9.6 S4 objects

Several functions are available for working with S4 objects and classes in C, including:

SEXP Rf_allocS4Object(void);
SEXP Rf_asS4(SEXP, Rboolean, int);
int R_check_class_etc(SEXP x, const char **valid);
SEXP R_do_MAKE_CLASS(const char *what);
SEXP R_do_new_object(SEXP class_def);
SEXP R_do_slot(SEXP obj, SEXP name);
SEXP R_do_slot_assign(SEXP obj, SEXP name, SEXP value);
SEXP R_getClassDef  (const char *what);
int R_has_slot(SEXP obj, SEXP name);

5.9.7 Handling lists

Some care is needed with lists, as R moved early on from using LISP-like lists (now called “pairlists”) to S-like generic vectors. As a result, the appropriate test for an object of mode list is Rf_isNewList, and we need Rf_allocVector(VECSXP, n) and not Rf_allocList(n).

List elements can be retrieved or set by direct access to the elements of the generic vector. Suppose we have a list object

a <- list(f = 1, g = 2, h = 3)

Then we can access a$g as a[[2]] by

    double g;
      ....
    g = REAL(VECTOR_ELT(a, 1))[0];

This can rapidly become tedious, and the following function (based on one in package stats) is very useful:

/* get the list element named str (ASCII), or return NULL */

SEXP getListElement(SEXP list, const char *str)
{
    SEXP elmt = R_NilValue, names = Rf_getAttrib(list, R_NamesSymbol);

    for (int i = 0; i < Rf_length(list); i++)
        if(strcmp(CHAR(STRING_ELT(names, i)), str) == 0) {
           /* ASCII only */
           elmt = VECTOR_ELT(list, i);
           break;
        }
    return elmt;
}

and enables us to say

  double g;
  g = REAL(getListElement(a, "g"))[0];

This code only works for names that are ASCII (see Character encoding issues).


5.9.8 Handling character data

R character vectors are stored as STRSXPs, a vector type like VECSXP where every element is of type CHARSXP. The CHARSXP elements of STRSXPs are accessed using STRING_ELT and SET_STRING_ELT.

CHARSXPs are read-only objects and must never be modified. In particular, the C-style string contained in a CHARSXP should be treated as read-only and for this reason the CHAR function used to access the character data of a CHARSXP returns (const char *) (this also allows compilers to issue warnings about improper use). Since CHARSXPs are immutable, the same CHARSXP can be shared by any STRSXP needing an element representing the same string. R maintains a global cache of CHARSXPs so that there is only ever one CHARSXP representing a given string in memory. It most cases it is easier to use Rf_translateChar or Rf_translateCharUTF8 to obtain the C string and it is safer against potential future changes in R (see Character encoding issues).

You can obtain a CHARSXP by calling Rf_mkChar and providing a NUL-terminated C-style string. This function will return a pre-existing CHARSXP if one with a matching string already exists, otherwise it will create a new one and add it to the cache before returning it to you. The variant Rf_mkCharLen can be used to create a CHARSXP from part of a buffer and will ensure null-termination.

Note that R character strings are restricted to 2^31 - 1 bytes, and hence so should the input to Rf_mkChar be (C allows longer strings on 64-bit platforms).


5.9.9 Working with closures

New function closure objects can be created with R_mkClosure:

SEXP R_mkClosure(SEXP formals, SEXP body, SEXP rho);

The components of a closure can be extracted with R_ClosureFormals, R_ClosureBody, and R_ClosureEnv. For a byte compiled closure R_ClosureBody returns the compiled body. R_ClosureExpr returns the body expression for both compiled and uncompiled closures. The expression for a compiled object can be obtained with R_BytecodeExpr.


5.9.10 Finding and setting variables

It will be usual that all the R objects needed in our C computations are passed as arguments to .Call or .External, but it is possible to find the values of R objects from within the C given their names. The following code is the equivalent of get(name, envir = rho).

SEXP getvar(SEXP name, SEXP rho)
{
    SEXP ans;

    if (!Rf_isString(name) || Rf_length(name) != 1)
        Rf_error("name is not a single string");
    if (!Rf_isEnvironment(rho))
        Rf_error("rho should be an environment");
    ans = R_getVar(Rf_installChar(STRING_ELT(name, 0)), rho, TRUE);
    if (TYPEOF(ans) != REALSXP || Rf_length(ans) == 0)
        Rf_error("value is not a numeric vector with at least one element");
    Rprintf("first value is %f\n", REAL(ans)[0]);
    return R_NilValue;
}

The main work is done by R_getVar, but to use it we need to install name as a name in the symbol table. As we wanted the value for internal use, we return NULL.

R_getVar is similar to the R function get. It signals an error if there is no binding for the variable in the environment. R_getVarEx can be used to return a default value if no binding is found; this corresponds to the R function get0. The third argument to R_getVar and R_getVarEx corresponds to the inherits argument to the R function get.

Functions with syntax

void Rf_defineVar(SEXP symbol, SEXP value, SEXP rho)
void Rf_setVar(SEXP symbol, SEXP value, SEXP rho)

can be used to assign values to R variables. defineVar creates a new binding or changes the value of an existing binding in the specified environment frame; it is the analogue of assign(symbol, value, envir = rho, inherits = FALSE), but unlike assign, defineVar does not make a copy of the object value.163 setVar searches for an existing binding for symbol in rho or its enclosing environments. If a binding is found, its value is changed to value. Otherwise, a new binding with the specified value is created in the global environment. This corresponds to assign(symbol, value, envir = rho, inherits = TRUE).

At times it may also be useful to create a new environment frame in C code. R_NewEnv is a C version of the R function new.env:

SEXP R_NewEnv(SEXP enclos, int hash, int size)

5.9.11 Some convenience functions

Some operations are done so frequently that there are convenience functions to handle them. (All these are provided via the header file Rinternals.h.)

Suppose we wanted to pass a single logical argument ignore_quotes: we could use

    int ign = Rf_asLogical(ignore_quotes);
    if(ign == NA_LOGICAL) Rf_error("'ignore_quotes' must be TRUE or FALSE");

which will do any coercion needed (at least from a vector argument), and return NA_LOGICAL if the value passed was NA or coercion failed. There are also Rf_asInteger, Rf_asReal and Rf_asComplex. The function Rf_asChar returns a CHARSXP. All of these functions ignore any elements of an input vector after the first. The function Rf_asCharacterFactor converts a factor to a character vector.

To return a length-one real vector we can use

    double x;

    ...
    return Rf_ScalarReal(x);

and there are versions of this for all the atomic vector types (those for a length-one character vector being Rf_ScalarString with argument a CHARSXP and Rf_mkString with argument const char *).

SEXP Rf_ScalarReal(double);
SEXP Rf_ScalarInteger(int);
SEXP Rf_ScalarLogical(int)
SEXP Rf_ScalarRaw(Rbyte);
SEXP Rf_ScalarComplex(Rcomplex);
SEXP Rf_ScalarString(SEXP);
SEXP Rf_mkString(const char *);

Some of the Rf_isXXXX functions differ from their apparent R-level counterparts: for example Rf_isVector is true for any atomic vector type (Rf_isVectorAtomic) and for lists and expressions (Rf_isVectorList) (with no check on attributes). Rf_isMatrix is a test of a length-2 "dim" attribute.

Rboolean Rf_isVector(SEXP);
Rboolean Rf_isVectorAtomic(SEXP);
Rboolean Rf_isVectorList(SEXP);
Rboolean Rf_isMatrix(SEXP);
Rboolean Rf_isPairList(SEXP);
Rboolean Rf_isPrimitive(SEXP);
Rboolean Rf_isTs(SEXP);
Rboolean Rf_isNumeric(SEXP);
Rboolean Rf_isArray(SEXP);
Rboolean Rf_isFactor(SEXP);
Rboolean Rf_isObject(SEXP);
Rboolean Rf_isFunction(SEXP);
Rboolean Rf_isLanguage(SEXP);
Rboolean Rf_isNewList(SEXP);
Rboolean Rf_isList(SEXP);
Rboolean Rf_isOrdered(SEXP);
Rboolean Rf_isUnordered(SEXP);
Rboolean Rf_isS4(SEXP);
Rboolean Rf_isNumber(SEXP);
Rboolean Rf_isDataFrame (SEXP);

Some additional predicates:

Rboolean Rf_isBlankString(const char *);
Rboolean Rf_StringBlank(SEXP);
Rboolean Rf_StringFalse(const char *);
Rboolean Rf_StringTrue(const char *);
int IS_LONG_VEC(SEXP);
int IS_SCALAR(SEXP, int);

There are a series of small macros/functions to help construct pairlists and language objects (whose internal structures just differ by SEXPTYPE). Function CONS(u, v) is the basic building block: it constructs a pairlist from u followed by v (which is a pairlist or R_NilValue). LCONS is a variant that constructs a language object. Functions Rf_list1 to Rf_list6 construct a pairlist from one to six items, and Rf_lang1 to Rf_lang6 do the same for a language object (a function to call plus zero to five arguments). Functions Rf_elt and Rf_lastElt find the i-th element and the last element of a pairlist, and Rf_nthcdr returns a pointer to the n-th position in the pairlist (whose CAR is the n-th item).

Functions Rf_str2type and Rf_type2str map R length-one character strings to and from SEXPTYPE numbers, and Rf_type2char maps numbers to C character strings. Rf_type2str_nowarn does not issue a warning if the SEXPTYPE is invalid.


5.9.11.1 Semi-internal convenience functions

There is quite a collection of functions that may be used in your C code if you are willing to adapt to rare API changes. These typically contain the “workhorses” of their R counterparts.

Functions Rf_any_duplicated and Rf_any_duplicated3 are fast versions of R’s any(duplicated(.)).

Function R_compute_identical corresponds to R’s identical function. Function R_BindingIsLocked corresponds to R’s bindingIsLocked function. Function R_ParentEnv corresponds to R’s parent.env.

The C functions Rf_inherits and Rf_topenv correspond to the R functions of the same base name. The C function Rf_GetOption1 corresponds to the R function getOption without specifying a default. Rf_GetOptionWidth returns the value of the width option as an int. The C function Rf_nlevels returns the number of levels of a factor. Unlike its R counterpart it always returns zero for non-factors.

For vectors the C function Rf_duplicated returns a logical vector indicating for each element whether it is duplicated or not. A second argument specifies the direction of the search.

The C function R_lsInternal3 returns a character vector of the names of variables in an environment. The second and third arguments specify whether all names are desired and whether the result should be sorted.

Some convenience functions for working with pairlist objects include Rf_copyListMatrix, Rf_listAppend, Rf_isVectorizable, Rf_VectorToPairList, and Rf_PairToVectorList

Some convenience functions for working with name spaces and environments include R_existsVarInFrame, R_removeVarFromFrame, R_PackageEnvName, R_IsPackageEnv, R_FindNamespace, R_IsNamespaceEnv, and R_NamespaceEnvSpec.

The C functions Rf_match and Rf_pmatch correspond to the R functions of the same base name. The C-level workhorse for partial matching is provided by Rf_psmatch.

The C functions R_forceAndCall and Rf_isUnsorted correspond to the R functions forceAndCall and is.unsorted.


5.9.12 Named objects and copying

[The NAMED mechanism has been replaced by reference counting.]

When assignments are done in R such as

x <- 1:10
y <- x

the named object is not necessarily copied, so after those two assignments y and x are bound to the same SEXPREC (the structure a SEXP points to). This means that any code which alters one of them has to make a copy before modifying the copy if the usual R semantics are to apply. Note that whereas .C and .Fortran do copy their arguments, .Call and .External do not. So Rf_duplicate is commonly called on arguments to .Call before modifying them. If only the top level is modified it may suffice to call Rf_shallow_duplicate.

At times it may be necessary to copy attributes from one object to another. This can be done using DUPLICATE_ATTRIB or SHALLOW_DUPLICATE_ATTRIB ANY_ATTRIB checks whether there are any attributes and CLEAR_ATTRIB removes all attributes.

However, at least some of this copying is unneeded. In the first assignment shown, x <- 1:10, R first creates an object with value 1:10 and then assigns it to x but if x is modified no copy is necessary as the temporary object with value 1:10 cannot be referred to again. R distinguishes between named and unnamed objects via a field in a SEXPREC that can be accessed via the macros NAMED and SET_NAMED. This can take values

0

The object is not bound to any symbol

1

The object has been bound to exactly one symbol

>= 2

The object has potentially been bound to two or more symbols, and one should act as if another variable is currently bound to this value. The maximal value is NAMEDMAX.

Note the past tenses: R does not do currently do full reference counting and there may currently be fewer bindings.

It is safe to modify the value of any SEXP for which NAMED(foo) is zero, and if NAMED(foo) is two or more, the value should be duplicated (via a call to duplicate) before any modification. Note that it is the responsibility of the author of the code making the modification to do the duplication, even if it is x whose value is being modified after y <- x.

The case NAMED(foo) == 1 allows some optimization, but it can be ignored (and duplication done whenever NAMED(foo) > 0). (This optimization is not currently usable in user code.) It is intended for use within replacement functions. Suppose we used

x <- 1:10
foo(x) <- 3

which is computed as

x <- 1:10
x <- "foo<-"(x, 3)

Then inside "foo<-" the object pointing to the current value of x will have NAMED(foo) as one, and it would be safe to modify it as the only symbol bound to it is x and that will be rebound immediately. (Provided the remaining code in "foo<-" make no reference to x, and no one is going to attempt a direct call such as y <- "foo<-"(x).)

This mechanism was replaced in R 4.0.0. To support future changes, package code should use NO_REFERENCES, MAYBE_REFERENCED, NOT_SHARED, MAYBE_SHARED, and MARK_NOT_MUTABLE.


5.10 Interface functions .Call and .External

In this section we consider the details of the R/C interfaces.

These two interfaces have almost the same functionality. .Call is based on the interface of the same name in S version 4, and .External is based on R’s .Internal. .External is more complex but allows a variable number of arguments.


5.10.1 Calling .Call

Let us convert our finite convolution example to use .Call. The calling function in R is

conv <- function(a, b) .Call("convolve2", a, b)

which could hardly be simpler, but as we shall see all the type coercion is transferred to the C code, which is

#include <R.h>
#include <Rinternals.h>

SEXP convolve2(SEXP a, SEXP b)
{
    int na, nb, nab;
    double *xa, *xb, *xab;
    SEXP ab;

    a = PROTECT(Rf_coerceVector(a, REALSXP));
    b = PROTECT(Rf_coerceVector(b, REALSXP));
    na = Rf_length(a); nb = Rf_length(b); nab = na + nb - 1;
    ab = PROTECT(Rf_allocVector(REALSXP, nab));
    xa = REAL(a); xb = REAL(b); xab = REAL(ab);
    for(int i = 0; i < nab; i++) xab[i] = 0.0;
    for(int i = 0; i < na; i++)
        for(int j = 0; j < nb; j++) xab[i + j] += xa[i] * xb[j];
    UNPROTECT(3);
    return ab;
}

5.10.2 Calling .External

We can use the same example to illustrate .External. The R code changes only by replacing .Call by .External

conv <- function(a, b) .External("convolveE", a, b)

but the main change is how the arguments are passed to the C code, this time as a single SEXP. The only change to the C code is how we handle the arguments.

#include <R.h>
#include <Rinternals.h>

SEXP convolveE(SEXP args)
{
    int i, j, na, nb, nab;
    double *xa, *xb, *xab;
    SEXP a, b, ab;

    a = PROTECT(Rf_coerceVector(CADR(args), REALSXP));
    b = PROTECT(Rf_coerceVector(CADDR(args), REALSXP));
    ...
}

Once again we do not need to protect the arguments, as in the R side of the interface they are objects that are already in use. The macros

  first = CADR(args);
  second = CADDR(args);
  third = CADDDR(args);
  fourth = CAD4R(args);
  fifth  = CAD5R(args);

provide convenient ways to access the first five arguments. More generally we can use the CDR and CAR macros as in

  args = CDR(args); a = CAR(args);
  args = CDR(args); b = CAR(args);

which clearly allows us to extract an unlimited number of arguments (whereas .Call has a limit, albeit at 65 not a small one).

More usefully, the .External interface provides an easy way to handle calls with a variable number of arguments, as length(args) will give the number of arguments supplied (of which the first is ignored). We may need to know the names (‘tags’) given to the actual arguments, which we can by using the TAG macro and using something like the following example, that prints the names and the first value of its arguments if they are vector types.

SEXP showArgs(SEXP args)
{
    void *vmax = vmaxget();
    args = CDR(args); /* skip 'name' */
    for(int i = 0; args != R_NilValue; i++, args = CDR(args)) {
        const char *name =
            Rf_isNull(TAG(args)) ? "" : Rf_translateChar(PRINTNAME(TAG(args)));
        SEXP el = CAR(args);
        if (length(el) == 0) {
            Rprintf("[%d] '%s' R type, length 0\n", i+1, name);
           continue;
        }
        switch(TYPEOF(el)) {
        case REALSXP:
            Rprintf("[%d] '%s' %f\n", i+1, name, REAL(el)[0]);
            break;
        case LGLSXP:
        case INTSXP:
            Rprintf("[%d] '%s' %d\n", i+1, name, INTEGER(el)[0]);
            break;
        case CPLXSXP:
        {
            Rcomplex cpl = COMPLEX(el)[0];
            Rprintf("[%d] '%s' %f + %fi\n", i+1, name, cpl.r, cpl.i);
        }
            break;
        case STRSXP:
            Rprintf("[%d] '%s' %s\n", i+1, name,
                   Rf_translateChar(STRING_ELT(el, 0)));
           break;
        default:
            Rprintf("[%d] '%s' R type\n", i+1, name);
       }
    }
    vmaxset(vmax);
    return R_NilValue;
}

This can be called by the wrapper function

showArgs <- function(...) invisible(.External("showArgs", ...))

Note that this style of programming is convenient but not necessary, as an alternative style is

showArgs1 <- function(...) invisible(.Call("showArgs1", list(...)))

The (very similar) C code is in the scripts.

Additional functions for accessing pairlist components are CAAR, CDAR, CDDR, and CDDDR. These components can be modified with SETCAR, SETCDR, SETCADR, SETCADDR, SETCADDDR, and SETCAD4R.


5.10.3 Missing and special values

One piece of error-checking the .C call does (unless NAOK is true) is to check for missing (NA) and IEEE special values (Inf, -Inf and NaN) and give an error if any are found. With the .Call interface these will be passed to our code. In this example the special values are no problem, as IEC 60559 arithmetic will handle them correctly. In the current implementation this is also true of NA as it is a type of NaN, but it is unwise to rely on such details. Thus we will re-write the code to handle NAs using macros defined in R_ext/Arith.h included by R.h.

The code changes are the same in any of the versions of convolve2 or convolveE:

    ...
  for(int i = 0; i < na; i++)
    for(int j = 0; j < nb; j++)
        if(ISNA(xa[i]) || ISNA(xb[j]) || ISNA(xab[i + j]))
            xab[i + j] = NA_REAL;
        else
            xab[i + j] += xa[i] * xb[j];
    ...

Note that the ISNA macro, and the similar macros ISNAN (which checks for NaN or NA) and R_FINITE (which is false for NA and all the special values), only apply to numeric values of type double. Missingness of integers, logicals and character strings can be tested by equality to the constants NA_INTEGER, NA_LOGICAL and NA_STRING. These and NA_REAL can be used to set elements of R vectors to NA.

The constants R_NaN, R_PosInf and R_NegInf can be used to set doubles to the special values.


5.11 Evaluating R expressions from C

The main function we will use is

SEXP Rf_eval(SEXP expr, SEXP rho);

the equivalent of the interpreted R code eval(expr, envir = rho) (so rho must be an environment), although we can also make use of Rf_findVar, Rf_defineVar and Rf_findFun (which restricts the search to functions).

To see how this might be applied, here is a simplified internal version of lapply for expressions, used as

a <- list(a = 1:5, b = rnorm(10), test = runif(100))
.Call("lapply", a, quote(sum(x)), new.env())

with C code

SEXP lapply(SEXP list, SEXP expr, SEXP rho)
{
    int n = Rf_length(list);
    SEXP ans;

    if(!Rf_isNewList(list)) Rf_error("'list' must be a list");
    if(!Rf_isEnvironment(rho)) Rf_error("'rho' should be an environment");
    ans = PROTECT(Rf_allocVector(VECSXP, n));
    for(int i = 0; i < n; i++) {
        Rf_defineVar(Rf_install("x"), VECTOR_ELT(list, i), rho);
        SET_VECTOR_ELT(ans, i, Rf_eval(expr, rho));
    }
    Rf_setAttrib(ans, R_NamesSymbol, Rf_getAttrib(list, R_NamesSymbol));
    UNPROTECT(1);
    return ans;
}

It would be closer to lapply if we could pass in a function rather than an expression. One way to do this is via interpreted R code as in the next example, but it is possible (if somewhat obscure) to do this in C code. The following is based on the code in src/main/optimize.c.

SEXP lapply2(SEXP list, SEXP fn, SEXP rho)
{
    int n = length(list);
    SEXP R_fcall, ans;

    if(!Rf_isNewList(list)) Rf_error("'list' must be a list");
    if(!Rf_isFunction(fn)) Rf_error("'fn' must be a function");
    if(!Rf_isEnvironment(rho)) Rf_error("'rho' should be an environment");
    R_fcall = PROTECT(Rf_lang2(fn, R_NilValue));
    ans = PROTECT(Rf_allocVector(VECSXP, n));
    for(int i = 0; i < n; i++) {
        SETCADR(R_fcall, VECTOR_ELT(list, i));
        SET_VECTOR_ELT(ans, i, Rf_eval(R_fcall, rho));
    }
    Rf_setAttrib(ans, R_NamesSymbol, Rf_getAttrib(list, R_NamesSymbol));
    UNPROTECT(2);
    return ans;
}

used by

.Call("lapply2", a, sum, new.env())

Function Rf_lang2 creates an executable pairlist of two elements, but this will only be clear to those with a knowledge of a LISP-like language.

As a more comprehensive example of constructing an R call in C code and evaluating, consider the following fragment. Similar code appears in the definition of do_docall in src/main/coerce.c.

    SEXP s, t;
    t = s = PROTECT(RF_allocLang(3));
    SETCAR(t, Rf_install("print")); t = CDR(t);
    SETCAR(t,  CAR(a)); t = CDR(t);
    SETCAR(t, Rf_ScalarInteger(digits));
    SET_TAG(t, Rf_install("digits"));
    Rf_eval(s, env);
    UNPROTECT(1);

The function Rf_allocLang is available as of R 4.4.1; for older versions replace Rf_allocLang(3) with

LCONS(R_NilValue, Rf_allocList(2))

At this point CAR(a) is the R object to be printed, the current attribute. There are three steps: the call is constructed as a pairlist of length 3, the list is filled in, and the expression represented by the pairlist is evaluated.

A pairlist is quite distinct from a generic vector list, the only user-visible form of list in R. A pairlist is a linked list (with CDR(t) computing the next entry), with items (accessed by CAR(t)) and names or tags (set by SET_TAG). In this call there are to be three items, a symbol (pointing to the function to be called) and two argument values, the first unnamed and the second named. Setting the type to LANGSXP makes this a call which can be evaluated.

Customarily, the evaluation environment is passed from the calling R code (see rho above). In special cases it is possible that the C code may need to obtain the current evaluation environment which can be done via R_GetCurrentEnv() function.


5.11.1 Zero-finding

In this section we re-work the example of Becker, Chambers & Wilks (1988, pp.~205–10) on finding a zero of a univariate function. The R code and an example are

zero <- function(f, guesses, tol = 1e-7) {
    f.check <- function(x) {
        x <- f(x)
        if(!is.numeric(x)) stop("Need a numeric result")
        as.double(x)
    }
    .Call("zero", body(f.check), as.double(guesses), as.double(tol),
          new.env())
}

cube1 <- function(x) (x^2 + 1) * (x - 1.5)
zero(cube1, c(0, 5))

where this time we do the coercion and error-checking in the R code. The C code is

SEXP mkans(double x)
{
    // no need for PROTECT() here, as REAL(.) does not allocate:
    SEXP ans = Rf_allocVector(REALSXP, 1);
    REAL(ans)[0] = x;
    return ans;
}

double feval(double x, SEXP f, SEXP rho)
{
    // a version with (too) much PROTECT()ion .. "better safe than sorry"
    SEXP symbol, value;
    PROTECT(symbol = Rf_install("x"));
    PROTECT(value = mkans(x));
    Rf_defineVar(symbol, value, rho);
    UNPROTECT(2);
    return(REAL(Rf_eval(f, rho))[0]);
}

SEXP zero(SEXP f, SEXP guesses, SEXP stol, SEXP rho)
{
    double x0 = REAL(guesses)[0], x1 = REAL(guesses)[1],
           tol = REAL(stol)[0];
    double f0, f1, fc, xc;

    if(tol <= 0.0) Rf_error("non-positive tol value");
    f0 = feval(x0, f, rho); f1 = feval(x1, f, rho);
    if(f0 == 0.0) return mkans(x0);
    if(f1 == 0.0) return mkans(x1);
    if(f0*f1 > 0.0) error("x[0] and x[1] have the same sign");

    for(;;) {
        xc = 0.5*(x0+x1);
        if(fabs(x0-x1) < tol) return  mkans(xc);
        fc = feval(xc, f, rho);
        if(fc == 0) return  mkans(xc);
        if(f0*fc > 0.0) {
            x0 = xc; f0 = fc;
        } else {
            x1 = xc; f1 = fc;
        }
    }
}

5.11.2 Calculating numerical derivatives

We will use a longer example (by Saikat DebRoy) to illustrate the use of evaluation and .External. This calculates numerical derivatives, something that could be done as effectively in interpreted R code but may be needed as part of a larger C calculation.

An interpreted R version and an example are

numeric.deriv <- function(expr, theta, rho=sys.frame(sys.parent()))
{
    eps <- sqrt(.Machine$double.eps)
    ans <- eval(substitute(expr), rho)
    grad <- matrix(, length(ans), length(theta),
                   dimnames=list(NULL, theta))
    for (i in seq_along(theta)) {
        old <- get(theta[i], envir=rho)
        delta <- eps * max(1, abs(old))
        assign(theta[i], old+delta, envir=rho)
        ans1 <- eval(substitute(expr), rho)
        assign(theta[i], old, envir=rho)
        grad[, i] <- (ans1 - ans)/delta
    }
    attr(ans, "gradient") <- grad
    ans
}
omega <- 1:5; x <- 1; y <- 2
numeric.deriv(sin(omega*x*y), c("x", "y"))

where expr is an expression, theta a character vector of variable names and rho the environment to be used.

For the compiled version the call from R will be

.External("numeric_deriv", expr, theta, rho)

with example usage

.External("numeric_deriv", quote(sin(omega*x*y)),
          c("x", "y"), .GlobalEnv)

Note the need to quote the expression to stop it being evaluated in the caller.

Here is the complete C code which we will explain section by section.

#include <R.h>
#include <Rinternals.h>
#include <float.h> /* for DBL_EPSILON */

SEXP numeric_deriv(SEXP args)
{
    SEXP theta, expr, rho, ans, ans1, gradient, par, dimnames;
    double tt, xx, delta, eps = sqrt(DBL_EPSILON), *rgr, *rans;
    int i, start;

    expr = CADR(args);
    if(!Rf_isString(theta = CADDR(args)))
        Rf_error("theta should be of type character");
    if(!Rf_isEnvironment(rho = CADDDR(args)))
        Rf_error("rho should be an environment");

    ans = PROTECT(Rf_coerceVector(eval(expr, rho), REALSXP));
    gradient = PROTECT(Rf_allocMatrix(REALSXP, LENGTH(ans), LENGTH(theta)));
    rgr = REAL(gradient); rans = REAL(ans);

    for(i = 0, start = 0; i < LENGTH(theta); i++, start += LENGTH(ans)) {
        par = PROTECT(Rf_findVar(Rf_installChar(STRING_ELT(theta, i)), rho));
        tt = REAL(par)[0];
        xx = fabs(tt);
        delta = (xx < 1) ? eps : xx*eps;
        REAL(par)[0] += delta;
        ans1 = PROTECT(Rf_coerceVector(Rf_eval(expr, rho), REALSXP));
        for(int j = 0; j < LENGTH(ans); j++)
            rgr[j + start] = (REAL(ans1)[j] - rans[j])/delta;
        REAL(par)[0] = tt;
        UNPROTECT(2); /* par, ans1 */
    }

    dimnames = PROTECT(Rf_allocVector(VECSXP, 2));
    SET_VECTOR_ELT(dimnames, 1,  theta);
    Rf_dimnamesgets(gradient, dimnames);
    Rf_setAttrib(ans, Rf_install("gradient"), gradient);
    UNPROTECT(3); /* ans  gradient  dimnames */
    return ans;
}

The code to handle the arguments is

    expr = CADR(args);
    if(!Rf_isString(theta = CADDR(args)))
        Rf_error("theta should be of type character");
    if(!Rf_isEnvironment(rho = CADDDR(args)))
        Rf_error("rho should be an environment");

Note that we check for correct types of theta and rho but do not check the type of expr. That is because eval can handle many types of R objects other than EXPRSXP. There is no useful coercion we can do, so we stop with an error message if the arguments are not of the correct mode.

The first step in the code is to evaluate the expression in the environment rho, by

    ans = PROTECT(Rf_coerceVector(eval(expr, rho), REALSXP));

We then allocate space for the calculated derivative by

    gradient = PROTECT(Rf_allocMatrix(REALSXP, LENGTH(ans), LENGTH(theta)));

The first argument to Rf_allocMatrix gives the SEXPTYPE of the matrix: here we want it to be REALSXP. The other two arguments are the numbers of rows and columns. (Note that LENGTH is intended to be used for vectors: Rf_length is more generally applicable.)

    for(i = 0, start = 0; i < LENGTH(theta); i++, start += LENGTH(ans)) {
        par = PROTECT(Rf_findVar(Rf_installChar(STRING_ELT(theta, i)), rho));

Here, we are entering a for loop. We loop through each of the variables. In the for loop, we first create a symbol corresponding to the i-th element of the STRSXP theta. Here, STRING_ELT(theta, i) accesses the i-th element of the STRSXP theta. installChar() installs the element as a name and Rf_findVar finds its value.

        tt = REAL(par)[0];
        xx = fabs(tt);
        delta = (xx < 1) ? eps : xx*eps;
        REAL(par)[0] += delta;
        ans1 = PROTECT(Rf_coerceVector(eval(expr, rho), REALSXP));

We first extract the real value of the parameter, then calculate delta, the increment to be used for approximating the numerical derivative. Then we change the value stored in par (in environment rho) by delta and evaluate expr in environment rho again. Because we are directly dealing with original R memory locations here, R does the evaluation for the changed parameter value.

        for(int j = 0; j < LENGTH(ans); j++)
            rgr[j + start] = (REAL(ans1)[j] - rans[j])/delta;
        REAL(par)[0] = tt;
        UNPROTECT(2);
    }

Now, we compute the i-th column of the gradient matrix. Note how it is accessed: R stores matrices by column (like Fortran).

    dimnames = PROTECT(Rf_allocVector(VECSXP, 2));
    SET_VECTOR_ELT(dimnames, 1, theta);
    Rf_dimnamesgets(gradient, dimnames);
    Rf_setAttrib(ans, install("gradient"), gradient);
    UNPROTECT(3);
    return ans;
}

First we add column names to the gradient matrix. This is done by allocating a list (a VECSXP) whose first element, the row names, is NULL (the default) and the second element, the column names, is set as theta. This list is then assigned as the attribute having the symbol R_DimNamesSymbol. Finally we set the gradient matrix as the gradient attribute of ans, unprotect the remaining protected locations and return the answer ans.


5.12 Parsing R code from C

Suppose an R extension wants to accept an R expression from the user and evaluate it. The previous section covered evaluation, but the expression will be entered as text and needs to be parsed first. A small part of R’s parse interface is declared in header file R_ext/Parse.h164.

An example of the usage can be found in the (example) Windows package windlgs included in the R source tree. The essential part is

#include <R.h>
#include <Rinternals.h>
#include <R_ext/Parse.h>

SEXP menu_ttest3()
{
    char cmd[256];
    SEXP cmdSexp, cmdexpr, ans = R_NilValue;
    ParseStatus status;
   ...
    if(done == 1) {
        cmdSexp = PROTECT(Rf_allocVector(STRSXP, 1));
        SET_STRING_ELT(cmdSexp, 0, Rf_mkChar(cmd));
        cmdexpr = PROTECT(R_ParseVector(cmdSexp, -1, &status, R_NilValue));
        if (status != PARSE_OK) {
            UNPROTECT(2);
            Rf_error("invalid call %s", cmd);
        }
        /* Loop is needed here as EXPSEXP will be of length > 1 */
        for(int i = 0; i < Rf_length(cmdexpr); i++)
            ans = Rf_eval(VECTOR_ELT(cmdexpr, i), R_GlobalEnv);
        UNPROTECT(2);
    }
    return ans;
}

Note that a single line of text may give rise to more than one R expression.

R_ParseVector is essentially the code used to implement parse(text=) at R level. The first argument is a character vector (corresponding to text) and the second the maximal number of expressions to parse (corresponding to n). The third argument is a pointer to a variable of an enumeration type, and it is normal (as parse does) to regard all values other than PARSE_OK as an error. Other values which might be returned are PARSE_INCOMPLETE (an incomplete expression was found) and PARSE_ERROR (a syntax error), in both cases the value returned being R_NilValue. The fourth argument is a length one character vector to be used as a filename in error messages, a srcfile object or the R NULL object (as in the example above). If a srcfile object was used, a srcref attribute would be attached to the result, containing a list of srcref objects of the same length as the expression, to allow it to be echoed with its original formatting.

Two higher-level alternatives are R_ParseString and R_ParseEvalString:

Function: SEXP R_ParseString (const char *str)
Function: SEXP R_ParseEvalString (const char *str, SEXP env)

R_ParseString Parses the code in str and returns the resulting expression. An error is signaled if parsing str produces more than one R expression. R_ParseEvalString first parses str, then evaluates the expression in the environment env, and returns the result.

An example from src/main/objects.c:

call = R_ParseString("base::nameOfClass(X)");

5.12.1 Accessing source references

The source references added by the parser are recorded by R’s evaluator as it evaluates code. Two functions make these available to debuggers running C code:

SEXP R_GetCurrentSrcref(int skip);

This function checks R_Srcref and the current evaluation stack for entries that contain source reference information. The skip argument tells how many source references to skip before returning the SEXP of the srcref object, counting from the top of the stack. If skip < 0, abs(skip) locations are counted up from the bottom of the stack. If too few or no source references are found, NULL is returned.

SEXP R_GetSrcFilename(SEXP srcref);

This function extracts the filename from the source reference for display, returning a length 1 character vector containing the filename. If no name is found, "" is returned.


5.13 External pointers and weak references

The SEXPTYPEs EXTPTRSXP and WEAKREFSXP can be encountered at R level, but are created in C code.

External pointer SEXPs are intended to handle references to C structures such as ‘handles’, and are used for this purpose in package RODBC for example. They are unusual in their copying semantics in that when an R object is copied, the external pointer object is not duplicated. (For this reason external pointers should only be used as part of an object with normal semantics, for example an attribute or an element of a list.)

An external pointer is created by

SEXP R_MakeExternalPtr(void *p, SEXP tag, SEXP prot);

where p is the pointer (and hence this cannot portably be a function pointer), and tag and prot are references to ordinary R objects which will remain in existence (be protected from garbage collection) for the lifetime of the external pointer object. A useful convention is to use the tag field for some form of type identification and the prot field for protecting the memory that the external pointer represents, if that memory is allocated from the R heap. Both tag and prot can be R_NilValue, and often are.

An alternative way to create an external pointer from a function pointer is

typedef void * (*R_DL_FUNC)();
SEXP R_MakeExternalPtrFn(R_DL_FUNC p, SEXP tag, SEXP prot);

The elements of an external pointer can be accessed and set via

void *R_ExternalPtrAddr(SEXP s);
DL_FUNC R_ExternalPtrAddrFn(SEXP s);
SEXP R_ExternalPtrTag(SEXP s);
SEXP R_ExternalPtrProtected(SEXP s);
void R_ClearExternalPtr(SEXP s);
void R_SetExternalPtrAddr(SEXP s, void *p);
void R_SetExternalPtrTag(SEXP s, SEXP tag);
void R_SetExternalPtrProtected(SEXP s, SEXP p);

Clearing a pointer sets its value to the C NULL pointer.

An external pointer object can have a finalizer, a piece of code to be run when the object is garbage collected. This can be R code or C code, and the various interfaces are, respectively.

void R_RegisterFinalizer(SEXP s, SEXP fun);
void R_RegisterFinalizerEx(SEXP s, SEXP fun, Rboolean onexit);

typedef void (*R_CFinalizer_t)(SEXP);
void R_RegisterCFinalizer(SEXP s, R_CFinalizer_t fun);
void R_RegisterCFinalizerEx(SEXP s, R_CFinalizer_t fun, Rboolean onexit);

The R function indicated by fun should be a function of a single argument, the object to be finalized. R does not perform a garbage collection when shutting down, and the onexit argument of the extended forms can be used to ask that the finalizer be run during a normal shutdown of the R session. It is suggested that it is good practice to clear the pointer on finalization.

The only R level function for interacting with external pointers is reg.finalizer which can be used to set a finalizer.

It is probably not a good idea to allow an external pointer to be saved and then reloaded, but if this happens the pointer will be set to the C NULL pointer.

Finalizers can be run at many places in the code base and much of it, including the R interpreter, is not re-entrant. So great care is needed in choosing the code to be run in a finalizer. Finalizers are marked to be run at garbage collection but only run at a somewhat safe point thereafter.

Weak references are used to allow the programmer to maintain information on entities without preventing the garbage collection of the entities once they become unreachable.

A weak reference contains a key and a value. The value is reachable if it is either reachable directly or via weak references with reachable keys. Once a value is determined to be unreachable during garbage collection, the key and value are set to R_NilValue and the finalizer will be run later in the garbage collection.

Weak reference objects are created by one of

SEXP R_MakeWeakRef(SEXP key, SEXP val, SEXP fin, Rboolean onexit);
SEXP R_MakeWeakRefC(SEXP key, SEXP val, R_CFinalizer_t fin,
                    Rboolean onexit);

where the R or C finalizer are specified in exactly the same way as for an external pointer object (whose finalization interface is implemented via weak references).

The parts can be accessed via

SEXP R_WeakRefKey(SEXP w);
SEXP R_WeakRefValue(SEXP w);
void R_RunWeakRefFinalizer(SEXP w);

A toy example of the use of weak references can be found at https://homepage.stat.uiowa.edu/~luke/R/references/weakfinex.html, but that is used to add finalizers to external pointers which can now be done more directly. At the time of writing no CRAN or Bioconductor package used weak references.


5.13.1 An example

Package RODBC uses external pointers to maintain its channels, connections to databases. There can be several connections open at once, and the status information for each is stored in a C structure (pointed to by thisHandle in the code extract below) that is returned via an external pointer as part of the RODBC ‘channel’ (as the "handle_ptr" attribute). The external pointer is created by

    SEXP ans, ptr;
    ans = PROTECT(Rf_allocVector(INTSXP, 1));
    ptr = R_MakeExternalPtr(thisHandle, Rf_install("RODBC_channel"), R_NilValue);
    PROTECT(ptr);
    R_RegisterCFinalizerEx(ptr, chanFinalizer, TRUE);
            ...
    /* return the channel no */
    INTEGER(ans)[0] = nChannels;
    /* and the connection string as an attribute */
    Rf_setAttrib(ans, Rf_install("connection.string"), constr);
    Rf_setAttrib(ans, Rf_install("handle_ptr"), ptr);
    UNPROTECT(3);
    return ans;

Note the symbol given to identify the usage of the external pointer, and the use of the finalizer. Since the final argument when registering the finalizer is TRUE, the finalizer will be run at the end of the R session (unless it crashes). This is used to close and clean up the connection to the database. The finalizer code is simply

static void chanFinalizer(SEXP ptr)
{
    if(!R_ExternalPtrAddr(ptr)) return;
    inRODBCClose(R_ExternalPtrAddr(ptr));
    R_ClearExternalPtr(ptr); /* not really needed */
}

Clearing the pointer and checking for a NULL pointer avoids any possibility of attempting to close an already-closed channel.

R’s connections provide another example of using external pointers, in that case purely to be able to use a finalizer to close and destroy the connection if it is no longer is use.


5.14 Vector accessor functions

The vector accessors like REAL, INTEGER, LOGICAL, RAW, COMPLEX, and VECTOR_ELT are functions when used in R extensions. (For efficiency they may be macros or inline functions when used in the R source code, apart from SET_STRING_ELT and SET_VECTOR_ELT which are always functions. When used outside the R source code all vector accessors are functions.) There are also read-only versions that return a const data pointer. For example, the return type of REAL_RO is const double *. These accessor functions check that they are being used on an appropriate type of SEXP. For VECSXP and STRSXP objects only read-only pointers are available as modifying their data directly would violate assumptions the memory manager depends on. DATAPTR_RO returns a generic read-only data pointer for any vector object.

Formerly it was possible for packages to obtain internal versions of some accessors by defining ‘USE_RINTERNALS’ before including Rinternals.h. This is no longer the case. Defining ‘USE_RINTERNALS’ now has no effect.

Atomic vector elements can also be accessed and set using element-wise operations like INTEGER_ELT and SET_INTEGER_ELT. For objects with a compact representation using these may avoid fully materializing the object. In contrast, obtaining a data pointer will have to fully materialize the object.


5.15 Character encoding issues

CHARSXPs can be marked as coming from a known encoding (Latin-1 or UTF-8). This is mainly intended for human-readable output, and most packages can just treat such CHARSXPs as a whole. However, if they need to be interpreted as characters or output at C level then it would normally be correct to ensure that they are converted to the encoding of the current locale: this can be done by accessing the data in the CHARSXP by Rf_translateChar rather than by CHAR. If re-encoding is needed this allocates memory with R_alloc which thus persists to the end of the .Call/.External call unless vmaxset is used (see Transient storage allocation).

There is a similar function Rf_translateCharUTF8 which converts to UTF-8: this has the advantage that a faithful translation is almost always possible (whereas only a few languages can be represented in the encoding of the current locale unless that is UTF-8).

Both Rf_translateChar and Rf_translateCharUTF8 will translate any input, using escapes such as ‘<A9>’ and ‘<U+0093>’ to represent untranslatable parts of the input.

There is a public interface to the encoding marked on CHARSXPs via

typedef enum {CE_NATIVE, CE_UTF8, CE_LATIN1, CE_BYTES, CE_SYMBOL, CE_ANY} cetype_t;
cetype_t Rf_getCharCE(SEXP);
SEXP Rf_mkCharCE(const char *, cetype_t);

Only CE_UTF8 and CE_LATIN1 are marked on CHARSXPs (and so Rf_getCharCE will only return one of the first three), and these should only be used on non-ASCII strings. Value CE_BYTES is used to make CHARSXPs which should be regarded as a set of bytes and not translated. Value CE_SYMBOL is used internally to indicate Adobe Symbol encoding. Value CE_ANY is used to indicate a character string that will not need re-encoding – this is used for character strings known to be in ASCII, and can also be used as an input parameter where the intention is that the string is treated as a series of bytes. (See the comments under Rf_mkChar about the length of input allowed.)

Function

Rboolean Rf_charIsASCII(SEXP);

can be used to detect whether a given CHARSXP represents an ASCII string. The implementation is equivalent to checking individual characters, but may be faster.

Function

Rboolean Rf_charIsUTF8(SEXP);

can be used to detect whether the internal representation of a given CHARSXP accessed via CHAR is UTF-8 (including ASCII). This function is rarely needed and specifically is not needed with Rf_translateCharUTF8, because such check is already included. However, when needed, it is better to use it in preference of Rf_getCharCE, as it is safer against future changes in the semantics of encoding marks and covers strings internally represented in the native encoding. Note that charIsUTF8() is not equivalent to getCharCE() == CE_UTF8.

Similarly, function

Rboolean Rf_charIsLatin1(SEXP);

can be used to detect whether the internal representation of a given CHARSXP accessed via CHAR is latin1 (including ASCII). It is not equivalent to Rf_getCharCE() == CE_LATIN1.

Function

const char *Rf_reEnc(const char *x, cetype_t ce_in, cetype_t ce_out,
                     int subst);

can be used to re-encode character strings: like Rf_translateChar it returns a string allocated by R_alloc. This can translate from CE_SYMBOL to CE_UTF8, but not conversely. Argument subst controls what to do with untranslatable characters or invalid input: this is done byte-by-byte with 1 indicates to output hex of the form <a0>, and 2 to replace by ., with any other value causing the byte to produce no output.

There is also

SEXP Rf_mkCharLenCE(const char *, int, cetype_t);

to create marked character strings of a given length.


5.16 Writing compact-representation-friendly code

A simple way to iterate in C over the elements of an atomic vector is to obtain a data pointer and index into that pointer with standard C indexing. However, if the object has a compact representation, then obtaining the data pointer will force the object to be fully materialized. An alternative is to use one of the following functions to query whether a data pointer is available.

Function: const int * LOGICAL_OR_NULL (SEXP x)
Function: const int * INTEGER_OR_NULL (SEXP x)
Function: const double * REAL_OR_NULL (SEXP x)
Function: const Rcomplex * COMPLEX_OR_NULL (SEXP x)
Function: const Rbyte * RAW_OR_NULL (SEXP x)
Function: const void * DATAPTR_OR_NULL (SEXP x)

These functions will return a data pointer if one is available. For vectors with a compact representation these functions will return NULL.

If a data pointer is not available, then code can access elements one at a time with functions like REAL_ELT. This is often sufficient, but in some cases can be inefficient. An alternative is to request data for contiguous blocks of elements. For a good choice of block size this can be nearly as efficient as direct pointer access.

Function: R_xlen_t INTEGER_GET_REGION (SEXP sx, R_xlen_t i, R_xlen_t n, int *buf)
Function: R_xlen_t LOGICAL_GET_REGION (SEXP sx, R_xlen_t i, R_xlen_t n, int *buf)
Function: R_xlen_t REAL_GET_REGION (SEXP sx, R_xlen_t i, R_xlen_t n, double *buf)
Function: R_xlen_t COMPLEX_GET_REGION (SEXP sx, R_xlen_t i, R_xlen_t n, Rcomplex *buf)
Function: R_xlen_t RAW_GET_REGION (SEXP sx, R_xlen_t i, R_xlen_t n, Rbyte *buf)

These functions copy a contiguous set of up to n elements starting with element i into a buffer buf. The return value is the actual number of elements copied, which may be less than n.

Macros in R_ext/Itermacros.h may help in implementing an iteration strategy.

Some functions useful in implementing new alternate representation classes, beyond those defined in R_ext/Altrep.h, include ALTREP, ALTREP_CLASS, R_altrep_data1, R_set_altrep_data1, R_altrep_data2, and R_set_altrep_data2.

For some objects it may be possible to very efficiently determine whether the object is sorted or contains no NA values. These functions can be used to query this information:

Function: int LOGICAL_NO_NA (SEXP x)
Function: int INTEGER_NO_NA (SEXP x)
Function: int REAL_NO_NA (SEXP x)
Function: int STRING_NO_NA (SEXP x)

A TRUE result means it is known that there are no NA values. A FALSE result means it is not known whether there are any NA values.

Function: int INTEGER_IS_SORTED (SEXP x)
Function: int REAL_IS_SORTED (SEXP x)
Function: int STRING_IS_SORTED (SEXP x)

These functions return one of SORTED_DECR, SORTED_INCR, or UNKNOWN_SORTEDNESS.


6 The R API: entry points for C code

There are a large number of entry points in the R executable/DLL that can be called from C code (and a few that can be called from Fortran code). Only those documented here are stable enough that they will only be changed with considerable notice.

The recommended procedure to use these is to include the header file R.h in your C code by

#include <R.h>

This will include several other header files from the directory R_INCLUDE_DIR/R_ext, and there are other header files there that can be included too, but many of the features they contain should be regarded as undocumented and unstable.

Most of these header files, including all those included by R.h, can be used from C++ code. (However, they cannot safely be included in a extern "C" { } block as they may include C++ headers when included from C++ code—and whether this succeeds is system-specific).

Note: Because R re-maps many of its external names to avoid clashes with system or user code, it is essential to include the appropriate header files when using these entry points.

This remapping can cause problems165, and can be eliminated by defining R_NO_REMAP (before including any R headers) and prepending ‘Rf_’ to all the function names used from Rinternals.h and R_ext/Error.h. These problems can usually be avoided by including other headers (such as system headers and those for external software used by the package) before any R headers. (Headers from other packages may include R headers directly or via inclusion from further packages, and may define R_NO_REMAP with or without including Rinternals.h.)

As from R 4.5.0, R_NO_REMAP is always defined when the R headers are included from C++ code.

If you decide to define R_NO_REMAP in your code, do use something like

#ifndef R_NO_REMAP
# define R_NO_REMAP
#endif

to avoid distracting compiler warnings.

Some of these entry points are declared in header Rmath.h, most of which are remapped there. That remapping can be eliminated by defining R_NO_REMAP_RMATH (before including any R headers) and prepending ‘Rf_’ to the function names used from that header except

exp_rand norm_rand unif_rand signrank_free wilcox_free

We can classify the entry points as

API

Entry points which are documented in this manual and declared in an installed header file. These can be used in distributed packages and ideally will only be changed after deprecation. See API index.

public

Entry points declared in an installed header file that are exported on all R platforms but are not documented and subject to change without notice. Do not use these in distributed code. Their declarations will eventually be moved out of installed header files.

private

Entry points that are used when building R and exported on all R platforms but are not declared in the installed header files. Do not use these in distributed code.

hidden

Entry points that are where possible (Windows and some modern Unix-alike compilers/loaders when using R as a shared library) not exported.

experimental

Entry points declared in an installed header file that are part of an experimental API, such as R_ext/Altrep.h. These are subject to change, so package authors wishing to use these should be prepared to adapt. See Experimental API index.

embedding

Entry points intended primarily for embedding and creating new front-ends. It is not clear that this needs to be a separate category but it may be useful to keep it separate for now. See Embedding API index.

If you would like to use an entry point or variable that is not identified as part of the API in this document, or is currently hidden, you can make a request for it to be made available. Entry points or variables not identified as in the API may be changed or removed with no notice as part of efforts to improve aspects of R.

Work in progress: Currently Entry points in the API are identified in the source for this document with @apifun, @eapifun, and @embfun entries. Similarly, @apivar, @eapivar, and @embvar identify variables, and @apihdr, @eapihdr, and @embhdr identify headers in the API. @forfun identifies entry points to be called as Fortran subroutines. This could be used for programmatic extraction, but the specific format is work in progress and even the way this document is produced is subject to change.


6.1 Memory allocation

There are two types of memory allocation available to the C programmer, one in which R manages the clean-up and the other in which users have full control (and responsibility).

These functions are declared in header R_ext/RS.h which is included by R.h.


6.1.1 Transient storage allocation

Here R will reclaim the memory at the end of the call to .C, .Call or .External. Use

char *R_alloc(size_t n, int size)

which allocates n units of size bytes each. A typical usage (from package stats) is

x = (int *) R_alloc(nrows(merge)+2, sizeof(int));

(size_t is defined in stddef.h which the header defining R_alloc includes.)

There is a similar call, S_alloc (named for compatibility with older versions of S) which zeroes the memory allocated,

char *S_alloc(long n, int size)

and

char *S_realloc(char *p, long new, long old, int size)

which (for new > old) changes the allocation size from old to new units, and zeroes the additional units. NB: these calls are best avoided as long is insufficient for large memory allocations on 64-bit Windows (where it is limited to 2^31-1 bytes).

This memory is taken from the heap, and released at the end of the .C, .Call or .External call. Users can also manage it, by noting the current position with a call to vmaxget and subsequently clearing memory allocated by a call to vmaxset. An example might be

void *vmax = vmaxget()
// a loop involving the use of R_alloc at each iteration
vmaxset(vmax)

This is only recommended for experts.

Note that this memory will be freed on error or user interrupt (if allowed: see Allowing interrupts).

The memory returned is only guaranteed to be aligned as required for double pointers: take precautions if casting to a pointer which needs more. There is also

long double *R_allocLD(size_t n)

which is guaranteed to have the 16-byte alignment needed for long double pointers on some platforms.

These functions should only be used in code called by .C etc, never from front-ends. They are not thread-safe.


6.1.2 User-controlled memory

The other form of memory allocation is an interface to malloc, the interface providing R error signaling. This memory lasts until freed by the user and is additional to the memory allocated for the R workspace.

The interface macros are

type* R_Calloc(size_t n, type)
type* R_Realloc(any *p, size_t n, type)
void R_Free(any *p)

providing analogues of calloc, realloc and free. If there is an error during allocation it is handled by R, so if these return the memory has been successfully allocated or freed. R_Free will set the pointer p to NULL.

Users should arrange to R_Free this memory when no longer needed, including on error or user interrupt. This can often be done most conveniently from an on.exit action in the calling R function – see pwilcox for an example.

Do not assume that memory allocated by R_Calloc/R_Realloc comes from the same pool as used by malloc:166 in particular do not use free or strdup with it.

Memory obtained by these macros should be aligned in the same way as malloc, that is ‘suitably aligned for any kind of variable’.

Historically the macros Calloc, Free and Realloc were used but have been removed in \R 4.5.0.

R_Calloc, R_Realloc, and R_Free are currently implemented as macros expanding to calls to R_chk_calloc, R_chk_realloc, and R_chk_free, respectively. These should not be called directly as they may be removed in the future.

char * CallocCharBuf(size_t n)
void * Memcpy(q, p, n)
void * Memzero(p, n)

CallocCharBuf(n) is shorthand for R_Calloc(n+1, char) to allow for the nul terminator. Memcpy and Memzero take n items from array p and copy them to array q or zero them respectively.


6.2 Error signaling

The basic error signaling routines are the equivalents of stop and warning in R code, and use the same interface.

void Rf_error(const char * format, ...);
void Rf_warning(const char * format, ...);
void Rf_errorcall(SEXP call, const char * format, ...);
void Rf_warningcall(SEXP call, const char * format, ...);
void Rf_warningcall_immediate(SEXP call, const char * format, ...);

These have the same call sequences as calls to printf, but in the simplest case can be called with a single character string argument giving the error message. (Don’t do this if the string contains ‘%’ or might otherwise be interpreted as a format.)

These are defined in header R_ext/Error.h included by R.h. NB: when R_NO_REMAP is defined (as is done for C++ code), Rf_error etc must be used.


6.2.1 Error signaling from Fortran

There are two interface function provided to call Rf_error and Rf_warning from Fortran code, in each case with a simple character string argument. They are defined as

subroutine rexit(message)
subroutine rwarn(message)

Messages of more than 255 characters are truncated, with a warning.


6.3 Random number generation

The interface to R’s internal random number generation routines is

double unif_rand();
double norm_rand();
double exp_rand();
double R_unif_index(double);

giving one uniform, normal or exponential pseudo-random variate. However, before these are used, the user must call

GetRNGstate();

and after all the required variates have been generated, call

PutRNGstate();

These essentially read in (or create) .Random.seed and write it out after use.

These are defined in header R_ext/Random.h. These functions are never remapped.

The random number generator is private to R; there is no way to select the kind of RNG nor set the seed except by evaluating calls to the R functions which do so.

The C code behind R’s rxxx functions can be accessed by including the header file Rmath.h; See Distribution functions. Those calls should also be preceded and followed by calls to GetRNGstate and PutRNGstate.


6.3.1 Random-number generation from Fortran

It was explained earlier that Fortran random-number generators should not be used in R packages, not least as packages cannot safely initialize them. Rather a package should call R’s built-in generators: one way to do so is to use C wrappers like

#include <R_ext/RS.h>
#include <R_ext/Random.h>

void F77_SUB(getRNGseed)(void) {
    GetRNGstate();
}
void F77_SUB(putRNGseed)(void) {
    PutRNGstate();
}
double F77_SUB(unifRand)(void) {
    return(unif_rand());
}

called from Fortran code like

      ...
      double precision X
      call getRNGseed()
      X = unifRand()
      ...
      call putRNGseed()

Alternatively one could use Fortran 2003’s iso_c_binding module by something like (fixed-form Fortran 90 code):

      module rngfuncs
        use iso_c_binding
        interface
          double precision
     *      function unifRand() bind(C, name = "unif_rand")
          end function unifRand

          subroutine getRNGseed() bind(C, name = "GetRNGstate")
          end subroutine getRNGseed

          subroutine putRNGseed() bind(C, name = "PutRNGstate")
          end subroutine putRNGseed
        end interface
      end module rngfuncs

      subroutine testit
      use rngfuncs
      double precision X
      call getRNGseed()
      X = unifRand()
      print *, X
      call putRNGSeed()
      end subroutine testit

6.4 Missing and IEEE special values

A set of functions is provided to test for NA, Inf, -Inf and NaN. These functions are accessed via macros:

ISNA(x)        True for R’s NA only
ISNAN(x)       True for R’s NA and IEEE NaN
R_FINITE(x)    False for Inf, -Inf, NA, NaN

and via function R_IsNaN which is true for NaN but not NA.

Do use R_FINITE rather than isfinite or finite; the latter is often mendacious and isfinite is only available on a some platforms, on which R_FINITE is a macro expanding to isfinite.

Currently in C code ISNAN is a macro calling isnan. (Since this gives problems on some C++ systems, if the R headers are called from C++ code a function call is used.)

You can check for Inf or -Inf by testing equality to R_PosInf or R_NegInf, and set (but not test) an NA as NA_REAL.

All of the above apply to double variables only. For integer variables there is a variable accessed by the macro NA_INTEGER which can used to set or test for missingness.

These are defined in header R_ext/Arith.h included by R.h.


6.5 Printing

The most useful function for printing from a C routine compiled into R is Rprintf. This is used in exactly the same way as printf, but is guaranteed to write to R’s output (which might be a GUI console rather than a file, and can be re-directed by sink). It is wise to write complete lines (including the "\n") before returning to R. It is defined in R_ext/Print.h.

The function REprintf is similar but writes on the error stream (stderr) which may or may not be different from the standard output stream.

Functions Rvprintf and REvprintf are analogues using the vprintf interface. Because that is a C99167 interface, they are only defined by R_ext/Print.h in C++ code if the macro R_USE_C99_IN_CXX is defined before it is included or (as from R 4.0.0) a C++11 compiler is used.

Another circumstance when it may be important to use these functions is when using parallel computation on a cluster of computational nodes, as their output will be re-directed/logged appropriately.


6.5.1 Printing from Fortran

On many systems Fortran write and print statements can be used, but the output may not interleave well with that of C, and may be invisible on GUI interfaces. They are not portable and best avoided.

Some subroutines are provided to ease the output of information from Fortran code.

subroutine dblepr(label, nchar, data, ndata)
subroutine realpr(label, nchar, data, ndata)
subroutine intpr (label, nchar, data, ndata)

and from R 4.0.0,

subroutine labelpr(label, nchar)
subroutine dblepr1(label, nchar, var)
subroutine realpr1(label, nchar, var)
subroutine intpr1 (label, nchar, var)

Here label is a character label of up to 255 characters, nchar is its length (which can be -1 if the whole label is to be used), data is an array of length at least ndata of the appropriate type (double precision, real and integer respectively) and var is a (scalar) variable. These routines print the label on one line and then print data or var as if it were an R vector on subsequent line(s). Note that some compilers will give an error or warning unless data is an array: others will accept a scalar when ndata has value one or zero. NB: There is no check on the type of data or var, so using real (including a real constant) instead of double precision will give incorrect answers.

intpr works with zero ndata so can be used to print a label in earlier versions of R.


6.6 Calling C from Fortran and vice versa

Naming conventions for symbols generated by Fortran differ by platform: it is not safe to assume that Fortran names appear to C with a trailing underscore. To help cover up the platform-specific differences there is a set of macros168 that should be used.

F77_SUB(name)

to define a function in C to be called from Fortran

F77_NAME(name)

to declare a Fortran routine in C before use

F77_CALL(name)

to call a Fortran routine from C

On current platforms these are the same, but it is unwise to rely on this. Note that names containing underscores were not legal in Fortran 77, and are not portably handled by the above macros. (Also, all Fortran names for use by R are lower case, but this is not enforced by the macros.)

For example, suppose we want to call R’s normal random numbers from Fortran. We need a C wrapper along the lines of

#include <R.h>

void F77_SUB(rndstart)(void) { GetRNGstate(); }
void F77_SUB(rndend)(void) { PutRNGstate(); }
double F77_SUB(normrnd)(void) { return norm_rand(); }

to be called from Fortran as in

      subroutine testit()
      double precision normrnd, x
      call rndstart()
      x = normrnd()
      call dblepr("X was", 5, x, 1)
      call rndend()
      end

Note that this is not guaranteed to be portable, for the return conventions might not be compatible between the C and Fortran compilers used. (Passing values via arguments is safer.)

The standard packages, for example stats, are a rich source of further examples.

Where supported, link time optimization provides a reliable way to check the consistency of calls to C from Fortran or vice versa. See Using Link-time Optimization. One place where this occurs is the registration of .Fortran calls in C code (see Registering native routines). For example

init.c:10:13: warning: type of 'vsom_' does not match original
 declaration [-Wlto-type-mismatch]
  extern void F77_NAME(vsom)(void *, void *, void *, void *,
    void *, void *, void *, void *, void *);
vsom.f90:20:33: note: type mismatch in parameter 9
   subroutine vsom(neurons,dt,dtrows,dtcols,xdim,ydim,alpha,train)
vsom.f90:20:33: note: 'vsom' was previously declared here

shows that a subroutine has been registered with 9 arguments (as that is what the .Fortran call used) but only has 8.


6.6.1 Fortran character strings

Passing character strings from C to Fortran or vice versa is not portable, but can be done with care. The internal representations are different: a character array in C (or C++) is NUL-terminated so its length can be computed by strlen. Fortran character arrays are typically stored as an array of bytes and a length. This matters when passing strings from C to Fortran or vice versa: in many cases one has been able to get away with passing the string but not the length. However, in 2019 this changed for gfortran, starting with version 9 but backported to versions 7 and 8. Several months later, gfortran 9.2 introduced an option

-ftail-call-workaround

and made it the current default but said it might be withdrawn in future.

Suppose we want a function to report a message from Fortran to R’s console (one could use labelpr, or intpr with dummy data, but this might be the basis of a custom reporting function). Suppose the equivalent in Fortran would be

      subroutine rmsg(msg)
      character*(*) msg
      print *.msg
      end

in file rmsg.f. Using gfortran 9.2 and later we can extract the C view by

gfortran -c -fc-prototypes-external rmsg.f

which gives

void rmsg_ (char *msg, size_t msg_len);

(where size_t applies to version 8 and later). We could re-write that portably in C as

#ifndef USE_FC_LEN_T
# define USE_FC_LEN_T
#endif
#include <Rconfig.h> // included by R.h, so define USE_FC_LEN_T early

void F77_NAME(rmsg)(char *msg, FC_LEN_T msg_len)
{
    char cmsg[msg_len+1];
    strncpy(cmsg, msg, msg_len);
    cmsg[msg_len] = '\0'; // nul-terminate the string, to be sure
    // do something with 'cmsg'
}

in code depending on R(>= 3.6.2). For earlier versions of R we could just assume that msg is NUL-terminated (not guaranteed, but people have been getting away with it for many years), so the complete C side might be

#ifndef USE_FC_LEN_T
# define USE_FC_LEN_T
#endif
#include <Rconfig.h>

#ifdef FC_LEN_T
void F77_NAME(rmsg)(char *msg, FC_LEN_T msg_len)
{
    char cmsg[msg_len+1];
    strncpy(cmsg, msg, msg_len);
    cmsg[msg_len] = '\0';
    // do something with 'cmsg'
}
#else
void F77_NAME(rmsg)(char *msg)
{
    // do something with 'msg'
}
#endif

(USE_FC_LEN_T is the default as from R 4.3.0.)

An alternative is to use Fortran 2003 features to set up the Fortran routine to pass a C-compatible character string. We could use something like

      module cfuncs
        use iso_c_binding, only: c_char, c_null_char
        interface
          subroutine cmsg(msg) bind(C, name = 'cmsg')
            use iso_c_binding, only: c_char
            character(kind = c_char):: msg(*)
          end subroutine cmsg
        end interface
      end module

      subroutine rmsg(msg)
        use cfuncs
        character(*) msg
        call cmsg(msg//c_null_char) ! need to concatenate a nul terminator
      end subroutine rmsg

where the C side is simply

void cmsg(const char *msg)
{
    // do something with nul-terminated string 'msg'
}

If you use bind to a C function as here, the only way to check that the bound definition is correct is to compile the package with LTO (which requires compatible C and Fortran compilers, usually gcc and gfortran).

Passing a variable-length string from C to Fortran is trickier, but https://www.intel.com/content/www/us/en/docs/fortran-compiler/developer-guide-reference/2023-0/bind-c.html provides a recipe. However, all the uses in BLAS and LAPACK are of a single character, and for these we can write a wrapper in Fortran along the lines of

      subroutine c_dgemm(transa, transb, m, n, k, alpha,
     +     a, lda, b, ldb, beta, c, ldc)
     +     bind(C, name = 'Cdgemm')
        use iso_c_binding, only : c_char, c_int, c_double
        character(c_char), intent(in) :: transa, transb
        integer(c_int), intent(in) :: m, n, k, lda, ldb, ldc
        real(c_double), intent(in) :: alpha, beta, a(lda, *), b(ldb, *)
        real(c_double), intent(out) ::  c(ldc, *)
        call dgemm(transa, transb, m, n, k, alpha,
     +             a, lda, b, ldb, beta, c, ldc)
      end subroutine c_dgemm

which is then called from C with declaration

void
Cdgemm(const char *transa, const char *transb, const int *m,
       const int *n, const int *k, const double *alpha,
       const double *a, const int *lda, const double *b, const int *ldb,
       const double *beta, double *c, const int *ldc);

Alternatively, do as R does as from version 3.6.2 and pass the character length(s) from C to Fortran. A portable way to do this is

// before any R headers, or define in PKG_CPPFLAGS
#ifndef  USE_FC_LEN_T
# define USE_FC_LEN_T
#endif
#include <Rconfig.h>
#include <R_ext/BLAS.h>
#ifndef FCONE
# define FCONE
#endif
...
        F77_CALL(dgemm)("N", "T", &nrx, &ncy, &ncx, &one, x,
                        &nrx, y, &nry, &zero, z, &nrx FCONE FCONE);

(Note there is no comma before or between the FCONE invocations.) It is strongly recommended that packages which call from C/C++ BLAS/LAPACK routines with character arguments adopt this approach: packages not using it will fail to install as from R 4.3.0.


6.6.2 Fortran LOGICAL

Passing Fortran LOGICAL variables to/from C/C++ is potentially compiler-dependent. Fortran compilers have long used a 32-bit integer type so it is pretty portable to use int * on the C/C++ side. However, recent versions of gfortran via the option -fc-prototypes-external say the C equivalent is int_least32_t *: ‘Link-Time Optimization’ will report int * as a mismatch. It is possible to use iso_c_binding in Fortran 2003 to map LOGICAL variables to the C99 type _Bool, but it is usually simpler to pass integers.


6.6.3 Passing functions

A number of packages call C functions passed as arguments to Fortran code along the lines of

c         subroutine fcn(m,n,x,fvec,iflag)
c         integer m,n,iflag
c         double precision x(n),fvec(m)
...
      subroutine lmdif(fcn, ...

where the C declaration and call are

void fcn_lmdif(int *m, int *n, double *par, double *fvec, int *iflag);

void F77_NAME(lmdif)(void (*fcn_lmdif)(int *m, int *n, double *par,
                                       double *fvec, int *iflag), ...

F77_CALL(lmdif)(&fcn_lmdif, ...

This works on most platforms but depends on the C and Fortran compilers agreeing on calling conventions: this have been seen to fail. The most portable solution seems to be to convert the Fortran code to C, perhaps using f2c.


6.7 Numerical analysis subroutines

R contains a large number of mathematical functions for its own use, for example numerical linear algebra computations and special functions.

The header files R_ext/BLAS.h, R_ext/Lapack.h and R_ext/Linpack.h contain declarations of the BLAS, LAPACK and LINPACK linear algebra functions included in R. These are expressed as calls to Fortran subroutines, and they will also be usable from users’ Fortran code. Although not part of the official API, this set of subroutines is unlikely to change (but might be supplemented).

The header file Rmath.h lists many other functions that are available and documented in the following subsections. Many of these are C interfaces to the code behind R functions, so the R function documentation may give further details.

If R_NO_REMAP_RMATH most of these will need to be prefixed by Rf_: see the header file for which ones.


6.7.1 Distribution functions

The routines used to calculate densities, cumulative distribution functions and quantile functions for the standard statistical distributions are available as entry points.

The arguments for the entry points follow the pattern of those for the normal distribution:

double dnorm(double x, double mu, double sigma, int give_log);
double pnorm(double x, double mu, double sigma, int lower_tail,
             int give_log);
double qnorm(double p, double mu, double sigma, int lower_tail,
             int log_p);
double rnorm(double mu, double sigma);

That is, the first argument gives the position for the density and CDF and probability for the quantile function, followed by the distribution’s parameters. Argument lower_tail should be TRUE (or 1) for normal use, but can be FALSE (or 0) if the probability of the upper tail is desired or specified.

Finally, give_log should be non-zero if the result is required on log scale, and log_p should be non-zero if p has been specified on log scale.

Note that you directly get the cumulative (or “integrated”) hazard function, H(t) = - log(1 - F(t)), by using

- pdist(t, ..., /*lower_tail = */ FALSE, /* give_log = */ TRUE)

or shorter (and more cryptic) - pdist(t, ..., 0, 1).

The random-variate generation routine rnorm returns one normal variate. See Random number generation, for the protocol in using the random-variate routines.

Note that these argument sequences are (apart from the names and that rnorm has no n) mainly the same as the corresponding R functions of the same name, so the documentation of the R functions can be used. Note that the exponential and gamma distributions are parametrized by scale rather than rate.

For reference, the following table gives the basic name (to be prefixed by ‘d’, ‘p’, ‘q’ or ‘r’ apart from the exceptions noted) and distribution-specific arguments for the complete set of distributions.

betabetaa, b
non-central betanbetaa, b, ncp
binomialbinomn, p
Cauchycauchylocation, scale
chi-squaredchisqdf
non-central chi-squarednchisqdf, ncp
exponentialexpscale (and not rate)
Ffn1, n2
non-central Fnfn1, n2, ncp
gammagammashape, scale
geometricgeomp
hypergeometrichyperNR, NB, n
logisticlogislocation, scale
lognormallnormlogmean, logsd
negative binomialnbinomsize, prob
normalnormmu, sigma
Poissonpoislambda
Student’s ttn
non-central tntdf, delta
Studentized rangetukey (*)rr, cc, df
uniformunifa, b
Weibullweibullshape, scale
Wilcoxon rank sumwilcoxm, n
Wilcoxon signed ranksignrankn

Entries marked with an asterisk only have ‘p’ and ‘q’ functions available, and none of the non-central distributions have ‘r’ functions.

(If remapping is suppressed, the Normal distribution names are Rf_dnorm4, Rf_pnorm5 and Rf_qnorm5.)

Additionally, a multivariate RNG for the multinomial distribution is

void Rf_rmultinom(int n, double* prob, int K, int* rN)

where K = length(prob), sum(prob[.]) == 1 and rN must point to a length-K integer vector n1 n2 .. nK where each entry nj=rN[j] is “filled” by a random binomial from Bin(n; prob[j]), constrained to sum(rN[.]) == n.

After calls to dwilcox, pwilcox or qwilcox the function wilcox_free() should be called, and similarly signrank_free() for the signed rank functions. Since wilcox_free() and signrank_free() were only added to Rmath.h in R  4.2.0, their use requires something like

#include "Rmath.h"
#include "Rversion.h"

#if R_VERSION < R_Version(4, 2, 0)
extern void wilcox_free(void);
extern void signrank_free(void);
#endif

For the negative binomial distribution (‘nbinom’), in addition to the (size, prob) parametrization, the alternative (size, mu) parametrization is provided as well by functions ‘[dpqr]nbinom_mu()’, see ?NegBinomial in R.

Functions dpois_raw(x, *) and dbinom_raw(x, *) are versions of the Poisson and binomial probability mass functions which work continuously in x, whereas dbinom(x,*) and dpois(x,*) only return non zero values for integer x.

double dbinom_raw(double x, double n, double p, double q, int give_log)
double dpois_raw (double x, double lambda, int give_log)

Note that dbinom_raw() returns both p and q = 1-p which may be advantageous when one of them is close to 1.


6.7.2 Mathematical functions

Function: double gammafn (double x)
Function: double lgammafn (double x)
Function: double digamma (double x)
Function: double trigamma (double x)
Function: double tetragamma (double x)
Function: double pentagamma (double x)
Function: double psigamma (double x, double deriv)
Function: void dpsifn (double x, int n, int kode, int m, double* ans, int* nz, int* ierr)

The Gamma function, the natural logarithm of its absolute value and first four derivatives and the n-th derivative of Psi, the digamma function, which is the derivative of lgammafn. In other words, digamma(x) is the same as psigamma(x,0), trigamma(x) == psigamma(x,1), etc. The underlying workhorse, dpsifn(), is useful, e.g., when several derivatives of log Gamma=lgammafn are desired. It computes and returns in ans[] the length-m sequence (-1)^(k+1) / gamma(k+1) * psi(k;x) for k = n ... n+m-1, where psi(k;x) is the k-th derivative of Psi(x), i.e., psigamma(x,k). For more details, see the comments in src/nmath/polygamma.c.

Function: double beta (double a, double b)
Function: double lbeta (double a, double b)

The (complete) Beta function and its natural logarithm.

Function: double choose (double n, double k)
Function: double lchoose (double n, double k)

The number of combinations of k items chosen from n and the natural logarithm of its absolute value, generalized to arbitrary real n. k is rounded to the nearest integer (with a warning if needed).

Function: double bessel_i (double x, double nu, double expo)
Function: double bessel_j (double x, double nu)
Function: double bessel_k (double x, double nu, double expo)
Function: double bessel_y (double x, double nu)

Bessel functions of types I, J, K and Y with index nu. For bessel_i and bessel_k there is the option to return exp(-x) I(xnu) or exp(x) K(xnu) if expo is 2. (Use expo == 1 for unscaled values.)


6.7.3 Numerical Utilities

There are a few other numerical utility functions available as entry points.

Function: double R_pow (double x, double y)
Function: double R_pow_di (double x, int i)
Function: double pow1p (double x, double y)

R_pow(x, y) and R_pow_di(x, i) compute x^y and x^i, respectively using R_FINITE checks and returning the proper result (the same as R) for the cases where x, y or i are 0 or missing or infinite or NaN.

pow1p(x, y) computes (1 + x)^y, accurately even for small x, i.e., |x| << 1.

Function: double log1p (double x)

Computes log(1 + x) (log 1 plus x), accurately even for small x, i.e., |x| << 1.

This should be provided by your platform, in which case it is not included in Rmath.h, but is (probably) in math.h which Rmath.h includes (except under C++, so it may not be declared for C++98).

Function: double log1pmx (double x)

Computes log(1 + x) - x (log 1 plus x minus x), accurately even for small x, i.e., |x| << 1.

Function: double log1pexp (double x)

Computes log(1 + exp(x)) (log 1 plus exp), accurately, notably for large x, e.g., x > 720.

Function: double log1mexp (double x)

Computes log(1 - exp(-x)) (log 1 minus exp), accurately, carefully for two regions of x, optimally cutting off at log 2 (= 0.693147..), using ((-x) > -M_LN2 ? log(-expm1(-x)) : log1p(-exp(-x))).

Function: double expm1 (double x)

Computes exp(x) - 1 (exp x minus 1), accurately even for small x, i.e., |x| << 1.

This should be provided by your platform, in which case it is not included in Rmath.h, but is (probably) in math.h which Rmath.h includes (except under C++, so it may not be declared for C++98).

Function: double lgamma1p (double x)

Computes log(gamma(x + 1)) (log(gamma(1 plus x))), accurately even for small x, i.e., 0 < x < 0.5.

Function: double cospi (double x)

Computes cos(pi * x) (where pi is 3.14159...), accurately, notably for half integer x.

This might be provided by your platform169, in which case it is not included in Rmath.h, but is in math.h which Rmath.h includes. (Ensure that neither math.h nor cmath is included before Rmath.h or define

#define __STDC_WANT_IEC_60559_FUNCS_EXT__ 1

before the first inclusion.)

Function: double sinpi (double x)

Computes sin(pi * x) accurately, notably for (half) integer x.

This might be provided by your platform, in which case it is not included in Rmath.h, but is in math.h which Rmath.h includes (but see the comments for cospi).

Function: double Rtanpi (double x)

Computes tan(pi * x) accurately, notably for integer x, giving NaN for half integer x and exactly +1 or -1 for (non half) quarter integers.

Function: double tanpi (double x)

Computes tan(pi * x) accurately for integer x with possibly platform dependent behavior for half (and quarter) integers. This might be provided by your platform, in which case it is not included in Rmath.h, but is in math.h which Rmath.h includes (but see the comments for cospi).

Function: double logspace_add (double logx, double logy)
Function: double logspace_sub (double logx, double logy)
Function: double logspace_sum (const double* logx, int n)

Compute the log of a sum or difference from logs of terms, i.e., “x + y” as log (exp(logx) + exp(logy)) and “x - y” as log (exp(logx) - exp(logy)), and “sum_i x[i]” as log (sum[i = 1:n exp(logx[i])] ) without causing unnecessary overflows or throwing away too much accuracy.

Function: int imax2 (int x, int y)
Function: int imin2 (int x, int y)
Function: double fmax2 (double x, double y)
Function: double fmin2 (double x, double y)

Return the larger (max) or smaller (min) of two integer or double numbers, respectively. Note that fmax2 and fmin2 differ from C99/C++11’s fmax and fmin when one of the arguments is a NaN: these versions return NaN.

Function: double sign (double x)

Compute the signum function, where sign(x) is 1, 0, or -1, when x is positive, 0, or negative, respectively, and NaN if x is a NaN.

Function: double fsign (double x, double y)

Performs “transfer of sign” and is defined as |x| * sign(y).

Function: double fprec (double x, double digits)

Returns the value of x rounded to digits significant decimal digits.

This is the function used by R’s signif().

Function: double fround (double x, double digits)

Returns the value of x rounded to digits decimal digits (after the decimal point).

This is the function used by R’s round(). (Note that C99/C++11 provide a round function but C++98 need not.)

Function: double ftrunc (double x)

Returns the value of x truncated (to an integer value) towards zero.


6.7.4 Mathematical constants

R has a set of commonly used mathematical constants encompassing constants defined by POSIX and usually found in headers math.h and cmath, as well as further ones that are used in statistical computations. These are defined to (at least) 30 digits accuracy in Rmath.h. The following definitions use ln(x) for the natural logarithm (log(x) in R).

NameDefinition (ln = log)round(value, 7)
M_Ee2.7182818
M_LOG2Elog2(e)1.4426950
M_LOG10Elog10(e)0.4342945
M_LN2ln(2)0.6931472
M_LN10ln(10)2.3025851
M_PIpi3.1415927
M_PI_2pi/21.5707963
M_PI_4pi/40.7853982
M_1_PI1/pi0.3183099
M_2_PI2/pi0.6366198
M_2_SQRTPI2/sqrt(pi)1.1283792
M_SQRT2sqrt(2)1.4142136
M_SQRT1_21/sqrt(2)0.7071068
M_SQRT_3sqrt(3)1.7320508
M_SQRT_32sqrt(32)5.6568542
M_LOG10_2log10(2)0.3010300
M_2PI2*pi6.2831853
M_SQRT_PIsqrt(pi)1.7724539
M_1_SQRT_2PI1/sqrt(2*pi)0.3989423
M_SQRT_2dPIsqrt(2/pi)0.7978846
M_LN_SQRT_PIln(sqrt(pi))0.5723649
M_LN_SQRT_2PIln(sqrt(2*pi))0.9189385
M_LN_SQRT_PId2ln(sqrt(pi/2))0.2257914

For compatibility with S this file used to define the constant PI this is defunct and should be replaced by M_PI. Header Constants.h includes either C header float.h or C++ header cfloat, which provide constants such as DBL_MAX.

Further, the included header R_ext/Boolean.h has enumeration constants TRUE and FALSE of type Rboolean in order to provide a way of using “logical” variables in C consistently. This can conflict with other software: for example it conflicts with the headers in IJG’s jpeg-9 (but not earlier versions).


6.8 Optimization

The C code underlying optim can be accessed directly. The user needs to supply a function to compute the function to be minimized, of the type

typedef double optimfn(int n, double *par, void *ex);

where the first argument is the number of parameters in the second argument. The third argument is a pointer passed down from the calling routine, normally used to carry auxiliary information.

Some of the methods also require a gradient function

typedef void optimgr(int n, double *par, double *gr, void *ex);

which passes back the gradient in the gr argument. No function is provided for finite-differencing, nor for approximating the Hessian at the result.

The interfaces (defined in header R_ext/Applic.h) are

  • Nelder Mead:
    void nmmin(int n, double *xin, double *x, double *Fmin, optimfn fn,
               int *fail, double abstol, double intol, void *ex,
               double alpha, double beta, double gamma, int trace,
               int *fncount, int maxit);
    
  • BFGS:
    void vmmin(int n, double *x, double *Fmin,
               optimfn fn, optimgr gr, int maxit, int trace,
               int *mask, double abstol, double reltol, int nREPORT,
               void *ex, int *fncount, int *grcount, int *fail);
    
  • Conjugate gradients:
    void cgmin(int n, double *xin, double *x, double *Fmin,
               optimfn fn, optimgr gr, int *fail, double abstol,
               double intol, void *ex, int type, int trace,
               int *fncount, int *grcount, int maxit);
    
  • Limited-memory BFGS with bounds:
    void lbfgsb(int n, int lmm, double *x, double *lower,
                double *upper, int *nbd, double *Fmin, optimfn fn,
                optimgr gr, int *fail, void *ex, double factr,
                double pgtol, int *fncount, int *grcount,
                int maxit, char *msg, int trace, int nREPORT);
    
  • Simulated annealing:
    void samin(int n, double *x, double *Fmin, optimfn fn, int maxit,
               int tmax, double temp, int trace, void *ex);
    

Many of the arguments are common to the various methods. n is the number of parameters, x or xin is the starting parameters on entry and x the final parameters on exit, with final value returned in Fmin. Most of the other parameters can be found from the help page for optim: see the source code src/appl/lbfgsb.c for the values of nbd, which specifies which bounds are to be used.


6.9 Integration

The C code underlying integrate can be accessed directly. The user needs to supply a vectorizing C function to compute the function to be integrated, of the type

typedef void integr_fn(double *x, int n, void *ex);

where x[] is both input and output and has length n, i.e., a C function, say fn, of type integr_fn must basically do for(i in 1:n) x[i] := f(x[i], ex). The vectorization requirement can be used to speed up the integrand instead of calling it n times. Note that in the current implementation built on QUADPACK, n will be either 15 or 21. The ex argument is a pointer passed down from the calling routine, normally used to carry auxiliary information.

There are interfaces (defined in header R_ext/Applic.h) for integrals over finite and infinite intervals (or “ranges” or “integration boundaries”).

  • Finite:
    void Rdqags(integr_fn f, void *ex, double *a, double *b,
                double *epsabs, double *epsrel,
                double *result, double *abserr, int *neval, int *ier,
                int *limit, int *lenw, int *last,
                int *iwork, double *work);
    
  • Infinite:
    void Rdqagi(integr_fn f, void *ex, double *bound, int *inf,
                double *epsabs, double *epsrel,
                double *result, double *abserr, int *neval, int *ier,
                int *limit, int *lenw, int *last,
                int *iwork, double *work);
    

Only the 3rd and 4th argument differ for the two integrators; for the finite range integral using Rdqags, a and b are the integration interval bounds, whereas for an infinite range integral using Rdqagi, bound is the finite bound of the integration (if the integral is not doubly-infinite) and inf is a code indicating the kind of integration range,

inf = 1

corresponds to (bound, +Inf),

inf = -1

corresponds to (-Inf, bound),

inf = 2

corresponds to (-Inf, +Inf),

f and ex define the integrand function, see above; epsabs and epsrel specify the absolute and relative accuracy requested, result, abserr and last are the output components value, abs.err and subdivisions of the R function integrate, where neval gives the number of integrand function evaluations, and the error code ier is translated to R’s integrate() $ message, look at that function definition. limit corresponds to integrate(..., subdivisions = *). It seems you should always define the two work arrays and the length of the second one as

    lenw = 4 * limit;
    iwork =   (int *) R_alloc(limit, sizeof(int));
    work = (double *) R_alloc(lenw,  sizeof(double));

The comments in the source code in src/appl/integrate.c give more details, particularly about reasons for failure (ier >= 1).


6.10 Utility functions

R has a fairly comprehensive set of sort routines which are made available to users’ C code. The following is declared in header file Rinternals.h.

Function: void R_orderVector (int* indx, int n, SEXP arglist, Rboolean nalast, Rboolean decreasing)
Function: void R_orderVector1 (int* indx, int n, SEXP x, Rboolean nalast, Rboolean decreasing)

R_orderVector() corresponds to R’s order(..., na.last, decreasing). More specifically, indx <- order(x, y, na.last, decreasing) corresponds to R_orderVector(indx, n, Rf_lang2(x, y), nalast, decreasing) and for three vectors, Rf_lang3(x,y,z) is used as arglist.

Both R_orderVector and R_orderVector1 assume the vector indx to be allocated to length >= n. On return, indx[] contains a permutation of 0:(n-1), i.e., 0-based C indices (and not 1-based R indices, as R’s order()).

When ordering only one vector, R_orderVector1 is faster and corresponds (but is 0-based) to R’s indx <- order(x, na.last, decreasing). It was added in R 3.3.0.

All other sort routines are declared in header file R_ext/Utils.h (included by R.h) and include the following.

Function: void R_isort (int* x, int n)
Function: void R_rsort (double* x, int n)
Function: void R_csort (Rcomplex* x, int n)
Function: void rsort_with_index (double* x, int* index, int n)

The first three sort integer, real (double) and complex data respectively. (Complex numbers are sorted by the real part first then the imaginary part.) NAs are sorted last.

rsort_with_index sorts on x, and applies the same permutation to index. NAs are sorted last.

Function: void Rf_revsort (double* x, int* index, int n)

Is similar to rsort_with_index but sorts into decreasing order, and NAs are not handled.

Function: void Rf_iPsort (int* x, int n, int k)
Function: void Rf_rPsort (double* x, int n, int k)
Function: void Rf_cPsort (Rcomplex* x, int n, int k)

These all provide (very) partial sorting: they permute x so that x[k] is in the correct place with smaller values to the left, larger ones to the right.

Function: void R_qsort (double *v, size_t i, size_t j)
Function: void R_qsort_I (double *v, int *I, int i, int j)
Function: void R_qsort_int (int *iv, size_t i, size_t j)
Function: void R_qsort_int_I (int *iv, int *I, int i, int j)

These routines sort v[i:j] or iv[i:j] (using 1-indexing, i.e., v[1] is the first element) calling the quicksort algorithm as used by R’s sort(v, method = "quick") and documented on the help page for the R function sort. The ..._I() versions also return the sort.index() vector in I. Note that the ordering is not stable, so tied values may be permuted.

Note that NAs are not handled (explicitly) and you should use different sorting functions if NAs can be present.

Function: subroutine qsort4 (double precision v, integer indx, integer ii, integer jj)
Function: subroutine qsort3 (double precision v, integer ii, integer jj)

The Fortran interface routines for sorting double precision vectors are qsort3 and qsort4, equivalent to R_qsort and R_qsort_I, respectively.