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"), 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.

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 on a Unix-alike to indicate if source code in the package 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). 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.

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 preventing⁷ 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, latin2 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.⁸ 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/developers⁹) 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 usually¹⁰ 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)

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 patchlevel (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 example¹¹ 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 needed¹² 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).

• Suggested packages

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

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 extensions¹³ .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 ‘SysDataCompression’ DESCRIPTION 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, formfeed, LF and CR) should be used in code files. Other characters are accepted in comments¹⁴, but then the comments may not be readable in e.g. a UTF-8 locale. Non-ASCII characters in object names will normally¹⁵ 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 means¹⁶ 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. 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++¹⁷ with associated extensions .c, .cc or .cpp, .f, .f90 or .f95, .m, and .mm, respectively. We recommend using .h for headers, also for C++¹⁸ or Fortran 9x 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 used in C or C++. On all Windows builds (even 64-bit ones) ‘_WIN32’ will be defined: on 64-bit Windows builds also ‘_WIN64’. On macOS ‘__APPLE__’ is defined¹⁹; for an ‘Apple Silicon’ platform, test for both ‘__APPLE__’ and ‘__arm64__’.

The default rules can be tweaked by setting macros²⁰ 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 for UCRT or Rtools42, 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).

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.

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 corresponding²¹ .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 (including Windows) 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 standard²² 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 .RData²³ 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 event²⁴ 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 Windows²⁵ 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’.

Other characters may benefit from being encoded: 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’.

1.2.1 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 are 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,

There are some macros³³ 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 be³⁴ included following it, as most BLAS libraries are written at least partially in Fortran.

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: 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.

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 32- and 64-bit code, for example (GNU make syntax, allowed on Windows)

ifeq "$(WIN)" "64"
PKG_LIBS = value for 64-bit Windows
else
PKG_LIBS = value for 32-bit Windows
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 is 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.

• OpenMP support
• Using pthreads
• Compiling in sub-directories

SHLIB_OPENMP_CFLAGS
SHLIB_OPENMP_CXXFLAGS
SHLIB_OPENMP_FFLAGS

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

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

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

  undefined symbol: __atomic_compare_exchange
  undefined symbol: __atomic_load

#pragma omp parallel for num_threads(nthreads) …

PKG_CPPFLAGS = -pthread
PKG_LIBS = -pthread

PKG_CPPFLAGS = -D_REENTRANT
PKG_LIBS = -lpthread

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)

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)

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)")

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 F9x 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’. On known platforms the Fortran compiler will also accept free-form Fortran 90/95 code with extension .f90 or .f95, but those are not used by R itself so this is not required.

The same compiler is used⁴³ 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 widely supported.

Some compilers specified by ‘FC’ will accept Fortran 2003, 2008 or 2018 code: such code should still use file extension .f90 or .f95. Most platforms use gfortran where you may need to include -std=f2003, -std=f2008 or (from version 8) -std=f2018 in PKG_FFLAGS or PKG_FCFLAGS: the default is ‘GNU Fortran’, Fortran 95 with non-standard extensions. The Oracle f95 compiler ‘accepts some Fortran 2003/8 features’ (search for ‘Oracle Developer Studio 12.6: Fortran User’s Guide’ and look for §4.6). Intel Fortran has full Fortran 2008 support from version 17.0, and some 2018 support in version 16.0 and more in version 19.0.

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

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 update⁴⁴ (‘C++14’) was published in December 2014 and will be 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 2023: R will support that from version 4.3.0.

What standard a C++ compiler aims to support can be hard to determine: the value⁴⁵ 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 unary_function (deprecated in C++11, removed in C++17).

Different versions of R have specified different default C++ standards, so for maximal portability a package should specify the standard it requires. In order to specify C++11 code in a package, Makevars file (or Makevars.win or Makevars.ucrt on Windows) should include the line

CXX_STD = CXX11

Compilation and linking will then be done with the C++11 compiler (if any).

Packages without a src/Makevars or src/Makefile file may specify that they require C++11 for code in the src directory by including ‘C++11’ in the ‘SystemRequirements’ field of the DESCRIPTION file, e.g.

SystemRequirements: C++11

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

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

CXX11=`"${R_HOME}/bin/R" CMD config CXX11`
if test -z "$CXX11"; then
  AC_MSG_ERROR([No C++11 compiler is available])
fi
CXX11STD=`"${R_HOME}/bin/R" CMD config CXX11STD`
CXX="${CXX11} ${CXX11STD}"
CXXFLAGS=`"${R_HOME}/bin/R" CMD config CXX11FLAGS`
AC_LANG(C++)

if C++11 was specified, but using CXX instead of CXX11 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 && $(MAKE) \
            CXX="$(CXX11) $(CXX11STD)" CXXFLAGS="$(CXX11FLAGS) $(CXX11PICFLAGS)")

Note that the mechanisms described here specify C++11 for code compiled by R CMD SHLIB as used by default by R CMD INSTALL. They do not necessarily apply if there is a src/Makefile file, nor to compilation done in vignettes or via other packages.

Support for a C++14 compiler (where available) has been in R since version 3.4.0. Similar considerations to C++11 apply, with the variables associated with the C++14 compiler using the prefix ‘CXX14’ instead of ‘CXX11’. For example, to use C++14 code in a package, the package’s Makevars file (or Makevars.win or Makevars.ucrt on Windows) should include the line

CXX_STD = CXX14

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’⁴⁶. 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

Note that Windows builds prior to R 4.0.0 used g++ 4.9.x which had only partial C++14 support, and the flag to obtain that support was not included in the default Windows build of R — one could try something like

CXX14="$(BINPREF)g++ $(M_ARCH)"
CXX14FLAGS="-O2 -Wall" 
CXX14STD=-std=gnu1y 

in HOME/.R/Makevars.win. The g++ version used from R 4.0.0 supports C++14 with flag -std=gnu14 and for back-compatibility -std=gnu1y.

C++17 (as from R 3.4.0) and C++20 (as from R 4.0.0) can be specified in an analogous way (replacing 14 by 17 or 20) 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 and the OpenCSW compilers for Solaris.

Note that C++17 or C++20 ‘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.

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.

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.

The discussion above is about the standard R ways of compiling C++: it will not apply to packages using src/Makefile or configure or building in a subdirectory that do not set the C++ standard. And the default C++ standard varies widely and gets changed frequently by vendors – for example Apple clang 14 defaults to C++98, LLVM clang 14 to C++14 and g++ 12 to C++17.

1.2.5 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 ../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.

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 available⁴⁸. (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 files⁴⁹ 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. Windows⁵⁰ 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 code in package vignettes (see Writing package vignettes) is executed, and the vignette PDFs 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 error⁵¹ in executing the R code in vignette foo.ext, a log file foo.ext.log is created in the check directory. The vignette PDFs 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.
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 version⁵² 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 declarations⁵³in 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.

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 Windows and macOS 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 systems⁵⁴ or from eclipse⁵⁵, directories with names check, chm, or ending .Rcheck or Old or old and files GNUMakefile⁵⁶, Read-and-delete-me or with base names starting with ‘.#’, or starting and ending with ‘#’, or ending in ‘~’, ‘.bak’ or ‘.swp’, are excluded by default⁵⁷. 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 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.

1.4.1 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.

1.5.1 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.

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 arguments⁶³ 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.⁶⁴ The directive

useDynLib(foo)

registers the shared object foo⁶⁵ 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: methods 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 exportClassPattern⁶⁶ 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 package⁶⁷, 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.

1.6.1 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 has binaries for Windows and 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 Preview⁸⁹ 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 and from 9.55 to 9.56).

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.⁹⁰. 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 available⁹¹ 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.

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.

C++ standards: From version 3.6.0 (3.6.2 on Windows), R defaulted to C++11 where available⁹²; R 4.1.0 and later default to C++14 (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 and C++14. (Specifying C++14 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.)

As noted elsewhere, the C++17 standard was finalized in Dec 2017 and support for it is patchy or absent on several platforms used for R. So portable code should not require C++17 (let alone C++20).

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,⁹³ 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 combinations⁹⁴ they are not declared otherwise. So you may need to include something like one of ⁹⁵

#define _XOPEN_SOURCE 600

or

#ifdef __GLIBC__
# define _POSIX_C_SOURCE 200809L
#endif

before any headers. (strdup, strncasecmp and strnlen are such functions – there are several older platforms which do 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, macOS and Solaris, and the FreeBSD manual pages at https://www.freebsd.org/cgi/man.cgi allow you to select many OSes⁹⁶) 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, clang, 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.h⁹⁷ 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 use⁹⁸.
• 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 NAN⁹⁹ 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.

  The GNU Fortran compiler also supports a large number of non-portable extensions, the most commonly encountered one being ISNAN¹⁰⁰. Some are listed at https://gcc.gnu.org/onlinedocs/gfortran/Extensions-implemented-in-GNU-Fortran.html. 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.

  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.

• Including C-style headers in C++ code is not portable. Including the legacy header¹⁰¹ 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 nor Solaris.) 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.

  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 OSs 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.

• 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. The most common issue involves the names of the Intel CPU registers such as CS, DS, ES, FS, GS and SS (and more with longer abbreviations¹⁰²) defined on i586/x64 Solaris in <sys/regset.h> and often included indirectly by <stdlib.h> and other core headers. Further examples are ERR, VERSION, LITTLE_ENDIAN, zero and I (which is defined in Solaris’ <complex.h> as a compiler intrinsic for the imaginary unit). 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 include ulong on several OSes, index_t and single on Solaris and thread using LLVM clang++ from version 9. (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 POSIX¹⁰³ 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 is 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. Examples include bind1st/bind2nd (use std::bind or lambdas¹⁰⁴) 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. So avoid common words and keywords in other programming languages.
• Be careful about including C headers in C++ code. Issues include
  • Use of the register storage class specifier (see the previous item).
  • The C99 keyword restrict is not part of¹⁰⁵ 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).

• 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
  #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_t and recent compilers are warning about passing 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 (a function without arguments should have prototype fn(void)) and is likely to become the default in future compilers. (It already is for LLVM clang in C23 mode.)
• 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 warn¹⁰⁶ 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.)

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

• Common symbols

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

1.6.5 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 does.
• (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).

1.8.1 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.

1.10.1 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.

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 typically¹¹¹ 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 ‘@’, and to avoid possible problems with the HTML help system it should not contain ‘/’ nor a space. (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. Entries in the package manual will be in alphabetic¹¹² 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.) 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. Optional text may precede¹¹³ this list (see for example the help for rle). Note that \value is implicitly a \describe environment, so that environment should not be used for listing components, just individual \item{}{} entries.

\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 between comments

## IGNORE_RDIFF_BEGIN
## IGNORE_RDIFF_END

is ignored when comparing check output to reference output (a -Ex.Rout.save file). This 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 objects that are not part of the package’s API. If the help page for object foo has keyword ‘internal’, then help(foo) gives this help page, but foo is excluded from several object indices, including the alphabetical list of objects 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 = LIBS argument is used, then the Rd file will be generated in final form, containing the information that would be produced up to library(help = pkgname). 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.

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.

3.4.1 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.

• sprof
• oprofile and operf

% setenv LD_PROFILE /path/to/R_HOME/library/stats/libs/stats.so
R
... run the boot example
% 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

rm /var/tmp/path/to/R_HOME/library/stats/libs/stats.so.profile
... to clean up ...

library(parallel)
N <- 1e6
dates <- sprintf('%04d-%02d-%02d', as.integer(2000+rnorm(N)),
                 as.integer(runif(N, 1, 12)), as.integer(runif(N, 1, 28)))
system.time(a <- as.POSIXct(dates, format = "%Y-%m-%d"))

   user  system elapsed
  0.371   0.237   0.612

                     self.time self.pct total.time total.pct
"strptime"                1.70    41.06       1.70     41.06
"as.POSIXct.POSIXlt"      1.40    33.82       1.42     34.30
"sprintf"                 0.74    17.87       0.98     23.67
...

operf R -f pvec.R
opreport
opreport -l /path/to/R_HOME/bin/exec/R
opannotate --source /path/to/R_HOME/bin/exec/R
## And for the system time
opreport -l /lib64/libc.so.6

CPU_CLK_UNHALT...|
  samples|      %|
------------------
   166761 99.9161 Rdev
        CPU_CLK_UNHALT...|
          samples|      %|
        ------------------
            70586 42.3276 no-vmlinux
            56963 34.1585 libc-2.16.so
            36922 22.1407 R
             1584  0.9499 stats.so
              624  0.3742 libm-2.16.so
...

samples  %        image name symbol name
10397    28.5123  R           R_gc_internal
5683     15.5848  R           do_sprintf
3036      8.3258  R           do_asPOSIXct
2427      6.6557  R           do_strptime
2421      6.6392  R           Rf_mkCharLenCE
1480      4.0587  R           w_strptime_internal
1202      3.2963  R           Rf_qnorm5
1165      3.1948  R           unif_rand
675       1.8511  R           mktime0
617       1.6920  R           makelt
617       1.6920  R           validate_tm
584       1.6015  R           day_of_the_week
...

system.time(example(sm.variogram))
...
   user  system elapsed
  5.543   3.202   8.785

  samples|      %|
------------------
   381845 99.9885 R
        CPU_CLK_UNHALT...|
          samples|      %|
        ------------------
           187484 49.0995 sm.so
           169627 44.4230 no-vmlinux
            12636  3.3092 libgfortran.so.3.0.0
             6455  1.6905 R

% su
% opcontrol --no-vmlinux
% (optional, some platforms) opcontrol --callgraph=5
% opcontrol --start
% exit

% R
... run the boot example
% opcontrol --dump
% opreport -l /path/to/R_HOME/library/stats/libs/stats.so
...
samples  %        symbol name
1623     75.5939  anonymous symbol from section .plt
349      16.2552  numeric_deriv
113       5.2632  nls_iter
62        2.8878  getListElement
% opreport -l /path/to/R_HOME/bin/exec/R
...
samples  %        symbol name
76052    11.9912  Rf_eval
54670     8.6198  Rf_findVarInFrame3
37814     5.9622  Rf_allocVector
31489     4.9649  Rf_duplicate
28221     4.4496  Rf_protect
26485     4.1759  Rf_cons
23650     3.7289  Rf_matchArgs
21088     3.3250  Rf_findFun
19995     3.1526  findVarLocInFrame
14871     2.3447  Rf_evalList
13794     2.1749  R_Newhashpjw
13522     2.1320  R_gc_internal
...

3.4.2 macOS

Developers have recommended sample (or Sampler.app, which is a GUI version), Shark (in version of Xcode up to those for Snow Leopard), and Instruments (part of Xcode, see https://help.apple.com/instruments/mac/current/).

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 Solaris¹²² 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 BLASes that use CPU-specific instructions so you may need to build a version of R specifically to use with valgrind.

On platforms where valgrind is 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 uses¹²³ of uninitialised memory and has little impact on speed (compared to level 0). Level 2 will detect many other memory-use bugs¹²⁴ 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 some¹²⁵ 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.

On macOS you may need to ensure that debugging symbols are made available (so valgrind reports line numbers in files). This can usually be done with the valgrind option --dsymutil=yes to ask for the symbols to be dumped when the .so file is loaded. This will not work where packages are installed into a system area (such as the R.framework) and can be slow. Installing packages with R CMD INSTALL --dsym installs the dumped symbols. (This can also be done by setting environment variable PKG_MAKE_DSYM to a non-empty value before the INSTALL.)

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 builds¹²⁶ 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-overflow’¹²⁷ 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-symbolizer¹²⁸ is on the path: if it is available but not on the path or has been renamed¹²⁹, 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.py¹³⁰ 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 versions¹³¹ 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 dealloc memory allocated by new[]). This matters on some systems but not all: unfortunately on some of those where it does not matter, system libraries¹³² 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.

• Using the Leak Sanitizer

ASAN_OPTIONS='detect_leaks=1'

setenv ASAN_OPTIONS ‘alloc_dealloc_mismatch=0:detect_leaks=0:detect_odr_violation=0’

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 builds¹³³ 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 can be combined with the Address Sanitizer by -fsanitize=undefined,address (where both are supported).

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,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 shift-base and shift-exponent) a subset of which could be combined with address, or use something like

-fsanitize=undefined -fno-sanitize=float-divide-by-zero

Options function, 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.

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: 10 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

bound-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 been seen with OpenMP-using code with gcc but there has been success with clang.

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) 32-bit Windows, Linux and macOS with similar aims to valgrind. It works with unmodified executables¹³⁴ 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 and Oracle Developer Studio has -C. 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).

4.4.1 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 0RET.

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.)