UTF-8 Support on Windows



R internally allows strings to be represented in the current native encoding, in UTF-8 and in Latin 1. When interacting with the operating system or external libraries, all these representations have to be converted to native encoding. On Linux and macOS today this is not a problem, because the native encoding is UTF-8, so all Unicode characters are supported. On Windows, the native encoding cannot be UTF-8 nor any other that could represent all Unicode characters.

Windows sometimes replaces characters by similarly looking representable ones (“best-fit”), which often works well but sometimes has surprising results, e.g. alpha character becomes letter a. In other cases, Windows may substitute non-representable characters by question marks or other and R may substitute by \uxxx, \UXXXXXXXX or other escapes. A number of functions accessing the OS consequently have complicated semantics and implementation on Windows. For example, normalizePath for a valid path tries to return also a valid path, which is a path to the same file. In a naive implementation, the normalized path could be non-existent or point to a different file due to best-fit, even when the original path is perfectly representable and valid.

This limitation of R on Windows is a source of pain for users who need to work with characters not representable in their native encoding. R provides “shortcuts” that sometimes bypass the conversion, e.g. when reading UTF-8 text files via readLines, but these work only for selected cases, when external software is not involved and their use is difficult.

Finally, the latest Windows 10 allows to set UTF-8 as the native encoding. R has been modified to allow this setting, which wasn’t hard as this has been supported on Unix/macOS for years.

The bad news is that the UTF-8 support on Windows requires Universal C Runtime (UCRT), a new C runtime. We need a new compiler toolchain and have to rebuild all external libraries for R and packages: no object files built using the older toolchains (Rtools 4 and older) can be re-used.

UCRT can be installed on older versions of Windows, but UTF-8 support will only work on Windows 10 (November 2019 update) and newer.

The rest of this text explains in more detail what native UTF-8 support would bring to Windows users of R. This text is simplifying out a number of details in order to be accessible to R users who are not package developers. An additional text for package developers and maintainers of infrastructures to build R on Windows is provided here, with details on how to build R using different infrastructures and how to build R with UCRT.

A demo binary installer for R and recommended packages is available (a link appears later in this text) as well as a demo toolchain, which has a number of libraries and headers for many but not all CRAN/BIOC packages.

Implications for RGui

RGui (RStudio is similar as it uses the same interface to R) is a Windows-only application implemented using Windows API and UTF-16LE. In R 4.0 and earlier, RGui can already work with all Unicode characters.

RGui can print UTF-8 R strings. When running with RGui, R escapes UTF-8 strings and embeds them into strings otherwise in native encoding on output. RGui understands this proprietary encoding and converts to UTF-16LE before printing. This is intended to be used in all outputs R produces for RGui, but the approach has its limits: it becomes complicated when formatting the output and R does not know yet where it will be printed. Many corner cases have been fixed, some recently, but likely some are remaining.

RGui can already pass Unicode strings to R. This is implemented by another semi-proprietary embedding, RGui converts UTF-16LE strings to the native encoding, but replaces the non-representable characters by \u and \U escapes that are understood by the parser. The parser will then turn these into R UTF-8 strings. This means that non-representable characters can be used only where \u and \U escapes are allowed by R, which includes R string literals where it is most important, but such characters cannot be even in comments.

As a side note here, I believe that to keep international collaboration on software development, all code should be in ASCII, definitely all symbols, and I would say even in English, including comments. But still, R is used also interactively and this is a technical limitation, not an intentionally enforced requirement.

For example, one can paste these Czech characters into Rgui: ěščřžýáíé. On Windows running in a English locale:

> x <- "ěščřžýáíé"
> Encoding(x)
[1] "UTF-8"
> x
[1] "ěščřžýáíé"

This works fine. But, a comment is already a problem:

> f <- function() {
+ x # ěščřžýáíé
+ }
> f
function() {
x # \u11bš\u10d\u159žýáíé
}

Some characters are fine, some are not.

In the experimental build of R, UTF-8 is the native encoding, so RGui will not use any \u, \U escapes when sending text to R and R will not embed any UTF-8 strings, because the native encoding is already UTF-8. The example above then works fine:

> f <- function() {
+ x # ěščřžýáíé
+ }
> f
function() {
x # ěščřžýáíé
}

UTF-8 is selected automatically as the encoding for the current locale in the experimental build:

> Sys.getlocale()
[1] "LC_COLLATE=English_United States.utf8;LC_CTYPE=English_United States.utf8;LC_MONETARY=English_United States.utf8;LC_NUMERIC=C;LC_TIME=English_United States.utf8"
> 

Note that RGui still needs to use fonts that can correctly represent the characters. Similarly, not all fonts are expected to correctly display examples in this text.

Implications for RTerm

RTerm is a Windows application not using Unicode, like most of R it is implemented using the standard C library assuming that the encoding-specific operations will work according to the C locale. In R 4.0 and earlier, RTerm cannot handle non-representable characters.

We cannot even paste non-representable characters to R. They will be converted automatically to the native encoding. Pasting “ěščřžýáíé” results in

> escrzyáíé

For the Czech text on Windows running in English locale, this is not so bad (only some diacritics marks are removed), but still not the exact representation. For Asian languages on Windows running in English locale, the result is unusable.

In principle, we can use the \u and \U sequences manually to input strings, but they still cannot be printed correctly:

> x <- "\u11b\u161\u10d\u159\u17e\u0fd\u0e1\u0ed\u0e9"
> Encoding(x)
[1] "UTF-8"
> x
[1] "escrzyáíé"
> as.hexmode(utf8ToInt(x))
[1] "11b" "161" "10d" "159" "17e" "0fd" "0e1" "0ed" "0e9"

The output shows that the string is correct inside R, it just cannot be printed on RTerm.

In the experimental build of R, if we run cmd.exe and then change the code page to UTF-8 via “chcp 65001” before running RTerm, this works as it should

> x <- "ěščřžýáíé"
> x
[1] "ěščřžýáíé"
> x <- "ěščřžýáíé"
> Encoding(x)
[1] "UTF-8"
> x
[1] "ěščřžýáíé"

This text is not going into details about where the characters exactly get converted/best-fitted, but the key thing is that with the UTF-8 build and when running cmd.exe in the UTF-8 code page (65001), without any modification of RTerm code, RTerm works with Unicode characters.

As with RGui, the terminal also needs apropriate fonts.. The same example with a Japanese text: こんにちは, 今日は

> x <- "こんにちは, 今日は"
> Encoding(x)
[1] "UTF-8"
> x
[1] "こんにちは, 今日は"

This example works fine with the experimental build on my system, but with the default font (Consolas), the characters are replaced by a question mark in a square. Still, just switching to another font, e.g. FangSong, in the cmd.exe menu, the characters appear correct in already printed text. The characters will also be correct when one pastes them to an application that uses the right font.

Implications for interaction with the OS

R on Windows already uses the Windows API in many cases instead of the standard C library to avoid the conversion or to get access to Windows-specific functionality. More specifically, R tries to always do it when passing strings to the OS, e.g. creating a file with a non-representable name already works. R converts UTF-8 strings to UTF16-LE, which Windows understands. However, R packages or external libraries often would not have such Windows specific code and would not be able to do that. With the experimental build, these problems disappear because the standard C functions, which in turn usually call the non-unicode Windows API, will use UTF-8.

A different situation is when getting strings from the operating system, for example listing files in a directory. R on Windows in such cases uses the C, non-unicode API or converts to the native encoding, unless this is a direct transformation of inputs that are already UTF-8. Please see R documentation for details; this text provides a simplification of the technical details.

In principle, R could also have used Windows-specific UTF-16LE calls and convert the strings to UTF-8, which R can represent. It would not be that much more work given how much effort has been spent on the functions passing strings to Windows.

However, R has been careful not to introduce UTF-8 strings for things the user has not already intentionally made UTF-8, because of problems that this would cause for packages not handling encodings correctly. Such packages will mysteriously start failing when incorrectly using strings in UTF-8 but thinking they were in native encoding. Such problems will not be found by automated testing, because tests don’t use such unusual inputs and are often run in English or similar locales.

This precaution came at a price of increased complexity. For example, the normalizePath implementation could be half the code size or less if we allowed introducing UTF-8 strings. R instead normalizes “less”, e.g. does not follow a symlink if it helps, but produces a representable path name for one that is in native encoding.

With UTF-8 as the native encoding, these considerations are no longer needed. Listing files in a directory when not-representable is no longer an issue (when valid Unicode) and it works in the experimental build without any code change.

Another issue is with external libraries that already started solving this problem their way, long before Windows 10. Some libraries bypass any external code and the C library for strings and perform string operations in UTF-8 or UTF-16LE, sometimes with the help of external libraries, typically ICU.

When R interacts with such libraries, it needs to know which encoding those libraries expect, and that sometimes changes from native encoding to UTF-8 as the libraries evolve. For example, Cairo switched to UTF-8, so R had to notice, and had to convert strings for newer Cairo versions to UTF-8 but for older versions to native encoding.

Such change is sometimes hard to notice, because the type remains the same, char *. Also handling these situations increases code complexity. One has to read carefully the change logs for external libraries, otherwise running into bugs that are hard to debug and almost impossible to detect by tests, as they don’t use unusual characters. Such transitions of external libraries will no longer be an issue with UTF-8 being the native encoding.

Implications for internal functionality

R allows multiple encoding of strings in R character objects, with a flag whether it is UTF-8, Latin 1 or native. But, eventually strings have to be converted to char * when interacting with the C library, the operating system and other external libraries, or with external code incorporated into R.

Historically, the assumption was that once typed char *, the strings are always in one encoding, and then it needs to be the native encoding. This makes a lot of sense as otherwise maintaining the code becomes difficult, but R made a number of exceptions e.g. for the shortcut in readLines, and sometimes it helps to keep strings longer as R character objects. Still, sometimes the conversion to native encoding is done just to have a char * representation of the string, even though not yet interacting with C/external code. All these conversions disappear when UTF-8 becomes the native encoding.

One related example is R symbols. They need to have a unique representation defined by a pointer stored in the R symbol table. For any effective implementation, they need to be in the same encoding, which now is the native encoding. A logical improvement would be converting to UTF-8, instead, but that would have potentially non-trivial performance overhead. These concerns are no longer necessary when UTF-8 becomes the native encoding.

In R 4.0, this limitation has as undesirable impact on hash maps:

e <- new.env(hash=TRUE)
assign("a", "letter a", envir=e)
assign("\u3b1", "letter alpha", envir=e)
ls(e)

On Windows, this produces a hash map with just a single element named “a”, because \u3b1 (α) gets best-fitted by Windows to letter “a”. With the experimental build, this works fine as it does on Unix/macOS, adding two elements to the hash map. Even though using non-ASCII variable names is probably not the right thing to do, a hash map really should be able to support arbitrary Unicode keys.

Demo

The experimental build of R can be downloaded from here. It has base and recommended packages, but will not work with other packages that use native code. The experimental toolchain allows to test more packages (but not all CRAN/BIOC), more information is available here and may be updated without notice (there is always SVN history of it). Not for production use.