← more articles

GNU Coreutils - Multibyte/unicode support

Random notes and pointers regarding the on-going effort to add multibyte and unicode support in GNU Coreutils.

If you’re considering working on multibyte/unicode/utf8 support in GNU coreutils (or other packages) - reading these should bring you up to speed (and hopefully save some time, too).

NOTE: multibyte, multibyte-sequences, unicode, utf-8 are sometimes used interchangeably throughout the document, but the intent is to support all multibyte locales, not just UTF8 encodings.


  1. Relevant Discussions on coreutils’ mailing lists - long and technical discussions with lots of relevant information.
  2. Related Bug Reports - various bugs reports from past years. Linked here if they contain useful points and/or detailed replies.
  3. Useful websites - good readings
  4. Online tools - websites providing useful conversions and information.
  5. Low-level command-line conversion - using printf,od,uconv,perl.
  6. invalid sequences - invalid sequences.
  7. Unicode glyph rendering - interplay between libc and xterm.
  8. cygwin and 16-bit wchar_t - special handling for systems where wchar_t is 16-bit, and UTF-16 surrogates (cygwin/OpenSolaris/AIX).
  9. glyph width and wcwidth issues - issues relating to incorrect wcwidth(3) results and different glyph rendering.
  10. expand - expand topics.
  11. wc - wc topics.
  12. cut - cut topics.
  13. head/tail - head/tail topics.
  14. tr - tr topics.
  15. fold/fmt - fold and fmt topics.
  16. od - od topics.
  17. unorm - unorm topics.
  18. Unicode Explained: The Book - pointers to relevant pages.

Relevant Discussions on Coreutils’ mailing lists





Other programs

Useful websites

Online tools

Low-level command-line conversion

Print unicode with perl:

perl -e 'print "\x{3A3}\n"'
perl -e 'print "\N{U+03A3}\n"'
perl -e 'print "\N{GREEK CAPITAL LETTER SIGMA}\n"'

Using coreutils’ printf (note that other printfs such as FreeBSD do not support \u):

printf '\u03A3\n'
printf '\U000003A3\n'

print octets with od (i.e. the binary encoding in the corrent locale, ‘\316\250’ is the octal representation of UTF-8 encoding for ):

$ printf 'Ψ\n' |  LC_ALL=C od -c -An
   316 250  \n

For maximum portability, always use LC_ALL=C (because FreeBSD’s od does support multibyte input and will not display the octet’s octal value). Print in octal, as every POSIX-compliant printf can handle octal values:

$ printf '\316\250\n'

Displaying hex unicode codepoints:

$ printf "Σημ" | iconv -t UTF-16LE | od -tx2 -An
 03a3 03b7 03bc

The above assumes little-endian (e.g. intel) CPU. Change to UTF-16BE for big-endian machines. It also assumes all code-points in the input fit into 16-bits (which is not a safe assumption). If some characters require more than 16-bit, a safer option is to use 32-bits for every code point:

$ printf "Σημ" | iconv -t UTF-32LE | od -tx4 -An
 000003a3 000003b7 000003bc

Checking for invalid multibyte-sequences with GNU sed:

The following example works in UTF-8 locale, and relies on the fact the GNU sed’s regular expression will not match invalid sequences (i.e. anything that was not replaces by the regex is invalid octet). If the output isn’t empty, the input had invalid multibyte sequenes:

$ printf 'a\xCEc\n' | sed 's/.*//g' | od -tx1c -An
  ce 0a
  c  \n

# Same detection for an input file:
$ printf 'ab\nc\n\xCE\xCEde\n\xCE\xA3f\n' > invalid.txt

$ sed -n 's/.//g ; H ; $@{x;s/\n//g;l@}' invalid.txt

# With few more commands, the offending line can be printed as well:
$ sed -n 's/.//g;=;l' invalid.txt | paste - -  | awk '$2!="$"'
3       \316\316$

# GNU sed in C locale can edit octets directly:
$ LC_ALL=C sed '3s/\o316\o316//' invalid.txt > fixed.txt

Converting UTF to \u sequences (using ‘uconv’ from ICU package):

$ printf "Σημεῖόν" | uconv -x 'hex-any ; any-hex'

Converting UTF-8 to named unicode charactes:

$ printf "Σημεῖόν" | uconv -x 'hex-any ; any-name'

List of possible transliterations in ‘uconv’/ICU:

uconv -L | tr ' ' '\n' | grep -i any | sort -f |  less

# such as:
uconv -x 'hex-any ; any-hex/perl'
uconv -x 'hex-any ; any-hex/java'
uconv -x 'hex-any ; any-hex/c'

ICU’s uconv supports several methods to handle invalid data (called ‘callbacks’ in their man page). This is part of the inspiration for unorm (the proposed coreutils program). Examples:

$ printf 'ab\342cdef' | uconv
Conversion to Unicode from codepage failed at input byte position 2. Bytes: e2 Error: Illegal character found
$ printf 'ab\342cdef' | uconv --callback substitute
$ printf 'ab\342cdef' | uconv --callback escape-c

Invalid Sequences

Modified-UTF-8 allows encoding NULL as 0xC0 0x80. This allows the byte with the value of zero, which is now not used for any character, to be used as a string terminator. https://en.wikipedia.org/wiki/Null_character#Encoding.

Invalid Sequences: https://en.wikipedia.org/wiki/UTF-8#Invalid_byte_sequences https://en.wikipedia.org/wiki/UTF-8#Invalid_code_points

Unicode glyph rendering

Show different unicode font implementation/support in terminals:

  1. Easy case: ‘e’ + combining mark (where a pre-combined ‘e’ exists):

    $ printf 'e\u0301\n'

    Works on gnome-terminal, mac-os-x-terminal, xterm. doesn’t work on ‘st’ (simple-terminal from st.suckless.org), prints ‘e’ followed by empty ‘grave’.

  2. Advanced support: any letter (regardless of pre-combined letter support:

    $ printf 'x\u0301e\n'

    On gnome-terminal,mac-os-terminal, prints ‘x’ with grave (nonsensical, but graphically correct) followed by ‘b’.

    On xterm, simple-term: prints ‘xe’

Cygwin (and other systems with 16-bit wchar_t)

Cygwin UTF-16 problems: https://cygwin.com/ml/cygwin/2011-02/msg00037.html - long and interesting discussion. First mention of possibility of wwchar_t and abstratction layer.

Cygwin Internationalization: https://cygwin.com/cygwin-ug-net/setup-locale.html - keeps recommending UTF8 everywhere.

How cygwin deals internally with windows filenames (which are UTF16): https://cygwin.com/cygwin-ug-net/using-specialnames.html#pathnames-unusual

But iswalpha takes wint_t which IS int32_t - perhaps do conversion manually? see https://cygwin.com/ml/cygwin/2011-02/msg00039.html and https://cygwin.com/ml/cygwin/2011-02/msg00044.html.

  1. printf can generate them, but on wchar_t/64-bit systems, mbrtowc(3) can’t decode them:

    $ printf '\ud800\n' | iconv -f utf-8 iconv:
    illegal input sequence at position 0
  2. A file containing:

    printf '\ud800\udc000\n' > 1.txt

    will be interpreted as 6 invalid octets on 64bit systems, and as either ‘U+100000’ or ‘U+D800 U+DC00’ on cygwin. which is correct ?

  3. On Cygwin, this input can be detected (and rejected to maintain consistency) by checking mbstate_t.__count==4 . What about other systems ?

whcar_t is NOT always UCS4: https://www.gnu.org/software/libunistring/manual/html_node/The-wchar_005ft-mess.html


In OpenSolaris, only under unicode locales, “wchar_t” is UTF-32 (good enough?). from https://docs.oracle.com/cd/E36784_01/html/E39536/gmwkm.html:

The ISO/IEC 9899 standard does not specify the form or the
encoding of the contents for the wchar_t data type. Because it is
an implementation-specific data type, it is not portable.

Although many implementations use some Unicode encoding forms for
the contents of the wchar_t data type, do not assume that the
contents ofwchar_t are Unicode. Some platforms use UCS-4 or UCS-2
for their wide-character encoding.

In Oracle Solaris, the internal form of wchar_t is specific to a

In the Oracle Solaris Unicode locales, wchar_t has the UTF-32
Unicode encoding form, and other locales have different


From https://www.ibm.com/support/knowledgecenter/en/ssw_aix_53/com.ibm.aix.nls/doc/nlsgdrf/codeset_over.htm:

On AIX 5.1 and later, the wchar_t datatype is 32–bit in the 64–bit
environment and 16–bit in the 32–bit environment.

The locale methods have been standardized such that in most
locales, the value stored in the wchar_t for a particular
character will always be its Unicode data value. [...]  All
locales use Unicode for their wide character code values (process
code), except the IBM-eucTW codeset.

glyph width and wcwidth issues

Expand, pr, fold, fmt will have glyph-width related issues. Some glyphs’ widths can not be determined by libc - but only by the graphical program that will render them on screen.

In other cases, the glyphs are optionally zero-width combining characters, or stand-alone visible characters. Example: Skin-tone modifiers (not zero width, but optionally is if it follows an face/hand emoji). See http://unicode.org/reports/tr51/#Diversity.


# With space (non-emoji) before the modifier, it is rendered as a normal character:
$ printf '\U0001F466 \U0001F3FB\n'
👦 🏻

# with an emoji preceeding it, it is combined:
$ printf '\U0001F466\U0001F3FB\n'

# NOTE for readers: whether the above is rendered as a single
# face depends on your web-browser or text editor.

The mbbuffer-debug (from this patch) is used below to examine multibyte input. The W column shows the result of wcwidth() of the character.

When checking with wcwidth, it gives width of 1 on MAC OS X (‘1’ is not always accurate, depending on later combining rendering):

$ printf '\U0001F3FB' | ./src/mbbuffer-debug -r
ofs  line colB colC V wc(dec) wc(hex) Ch  W n octets
0    1    1    1    y  127995 0x1f3fb =   1 4 0xf0 0x9f 0x8f 0xbb

Yet on glibc, wcwidth returns -1 for all SMP codepoints (as wcwidth returns -1 for all non-printables):

$ printf '\U0001F3FB' | ./src/mbbuffer-debug -r
ofs  line colB colC V wc(dec) wc(hex) Ch  W n octets
0    1    1    1    y  127995 0x1f3fb =  -1 4 0xf0 0x9f 0x8f 0xbb

It is not at all clear if there’s a “correct” width, as the visualization results differ based on the rendering environment (the following looks different between Mac OSX terminal and gnome-terminal 3.6.2 and safari and chrome):

$ printf 'a\U0001F466\U0001F3FBaa\tb\naaaa\tb\n'
a👦🏻aa   b
aaaa    b

Should we use ‘wcswidth’, or alternatively, process “EmojiModifiers” propery? (see http://unicode.org/reports/tr51/#Data_Files but then, the list of possible specific properies is endless).

Other “Modifier Symbols” (Category Sk): https://codepoints.net/search?gc=Sk

See also: Multi-Person Grouping: http://unicode.org/reports/tr51/#Multi_Person_Groupings Can be rendered as multiple icons or one combined icon (taking one or more characters).

More examples

joiner such as: U+20E0 COMBINING ENCLOSING CIRCLE BACKSLASH which as an overlaid glyph, to indicate a prohibition or “NO”

$ printf '\U0001F52b\u20E0\n'  # no guns

$ printf '\U0001F399\u20E0\n'  # no microphones

Some characters are ‘combining’, and wcwidth does indicate they have zero width (which is good for expand/pr/fold/fmt), but when rendered they actually consume visual space on the screen, messing up alignment. This is COMBINING ENCLOSING KEYCAP (U+20E3):

$ printf 'a\u20E3aa\tb\naaaa\tb\n'
a⃣aa     b
aaaa    b

# this is multibyte-aware expand, still incorrect output:
$  printf 'a\u20E3aa\tb\naaaa\tb\n'  | ./src/expand
a⃣aa     b
aaaa    b

# Despite wcwidth giving W=2, the character is rendered
# wider than a single column on the screen:
$ printf 'X\u20E3\n' | ./src/mbbuffer-debug -r
ofs  line colB colC V wc(dec) wc(hex) Ch  W n octets
0    1    1    1    y      88 0x00058 X   1 1 X
1    1    2    2    y    8419 0x020e3  ⃣   0 3 0xe2 0x83 0xa3
4    1    5    3    y      10 0x0000a =  -1 1 0x0a

# NOTE to readers:
# On Mac OS X terminal, Safari and Firefox,
# the above is rendered as 'a' surrounded by a square. YMMV.


  1. How many characters are NON-PRINTABLE (i.e. wcwidth()==-1), but in expand we do not treat them properly? when adding columns in ‘expand’, ensure wcwidth>0 ?? Do all “word_break” or “line_break” characters have wcwidth()==-1 ?
  2. check wcwidth() on CJK idograms (does it return 2?)

Not bugs, but worth knowing

Combined letters

In Laten Extended-B: Some glyphs are 2-letters squeezed into a width of 1. wcwidth on glibc and macos seems to handle it correctly (return width==1), and on Xterm,Mac Terminal they are indeed rendered in width of 1 (there seems to be a problem on gnome-terminal/ubuntu-14.04, but that’s a font issue).

Example: :

$ printf '\u01c4a\nbc\n'

Full Width glyphs

Some glyphs are designated as full-width, meaning the consume a width of 2 characters, and can be used for easy alignment with CJK characters. See https://en.wikipedia.org/wiki/Halfwidth_and_fullwidth_forms#In_Unicode

This is Full-width capital B U+FF22:

$ printf 'a\uFF22c\nabc\n'

Both glibc and MacOS-X wcwidth gives width==2 for these (which is good for expand/pr/fmt/fold):

$ printf '\uFF22' | ./src/mbbuffer-debug -r
ofs  line colB colC V wc(dec) wc(hex) Ch  W n octets
0    1    1    1    y   65314 0x0ff22 B  2 3 0xef 0xbc 0xa2

From Cygwin’s internationalization page:

There's a class of characters in the Unicode character set, called
the "CJK Ambiguous Width" characters. For these characters, the
width returned by the wcwidth/wcswidth functions is
usually 1. This can be a problem with East-Asian languages, which
historically use character sets where these characters have a
width of 2. Therefore, wcwidth/wcswidth return 2 as the width of
these characters when an East-Asian charset such as GBK or SJIS is
selected, or when UTF-8 is selected and the language is specified
as "zh" (Chinese), "ja" (Japanese), or "ko" (Korean). This is not
correct in all circumstances, hence the locale modifier
"@cjknarrow" can be used to force wcwidth/wcswidth to return 1 for
the ambiguous width characters.


See width/wcwidth issues above.


  1. Add a special option in ‘wc’ to count non-zero width characters?? (But then, what about optional modifiers, e.g. skin-color and family-joiner?)
  2. What about counting SMP characters (which gives wcwidth()==-1).


option to never cut in a combining-mark? (or technically, only cut in clear graphmeme? e.g. never before ZWJ, BIDI mark, etc.)?


how to treat zero-width-joiners, how to treat combined characters ? based on width ?


Multibyte sequence case conversion (the following works on Mac/FreeBSD, not yet in coreutils):

$ printf '\u0103\n'

$ printf '\u0103\n' | /usr/bin/tr '[:lower:]' '[:upper:]'

Pádraig Brady wrote (privately):

I’ve also noticed interesting chars like the titlecase letter ‘Lj’ (U+01C8) which is neither upper or lower but does have an upper case (U+01C7), or the fact the there are only 2 code points Ⱥ (U+023A) and Ⱦ (U+023E) that increase in length (2 to 3 bytes) when lower-cased.

Deleting multibyte-sequences:

$ printf '\u0103b\u0106d\n'

$ printf '\u0103b\u0106d\n' | LC_ALL=C od -tc -An
 304 203   b 304 206   d  \n

# On Mac/FreeBSD, it works as expected:
$ printf '\u0103b\u0106d\n' | tr -d 'ă' | LC_ALL=C od -tc -An
   b 304 206   d  \n

# Coreutils 'tr' treats input as two independant octets, delete
# both instances of \304 resulting in invalid output:
$ printf '\u0103b\u0106d\n' | tr -d 'ă' | LC_ALL=C od -tc -An
   b 206   d  \n

POSIX tr(1) says (in “Extended Description”):

“\octal - […] Multi-byte characters require multiple, concatenated escape sequences of this type, including the leading for each byte."

Based on my understanding of the above, the following ‘should work’:

# It doesn't work at all on Mac/FreeBSD:
$ printf '\u0103b\u0106d\n' | tr -d '\304\203' | LC_ALL=C od -tc -An
 304 203   b 304 206   d  \n

# It doesn't work on GNU tr (since as above, it treats the octets indepedently):
$ printf '\u0103b\u0106d\n' | tr -d '\304\203' | LC_ALL=C od -tc -An
   b 206   d  \n
  1. Should uppercase mapping of ligatures turns into two letters?

  2. German Capital Sharp S is a similar issue.

     $ printf '\uFB01' | tr '[:lower:]' '[:upper:]'
  3. POSIX says Chracter Ranges are UNDEFINED in non-posix locale:

    > "c-c"
    > "In locales other than the POSIX locale, this construct has unspecified behavior."

    How to handle these when we do implement multibyte support?

  4. What does the -C (upper-case C) exactly do according to POSIX ?

     > "The ISO POSIX-2:1993 standard had a -c option that behaved
     > similarly to the -C option, but did not supply functionality
     > equivalent to the -c option specified in POSIX.1-2008."  The
     > earlier version also said that octal sequences referred to
     > collating elements and could be placed adjacent to each
     > other to specify multi-byte characters. However, it was
     > noted that this caused ambiguities because tr would not be
     > able to tell whether adjacent octal sequences were intending
     > to specify multi-byte characters or multiple single byte
     > characters. POSIX.1-2008 specifies that octal sequences
     > always refer to single byte binary values when used to
     > specify an endpoint of a range of collating elements.  "
  5. Equivalent classes: The following is supposed to work (i.e. replace also the umlaut-a into X), but does not work on FreeBSD-10.3/OpenBSD-6 which supposed to support it (TODO: check on musl-libc):

     $ printf 'abc \303\244\303\202 def\n' | LC_ALL=en_US.UTF-8 tr '[=a=]' X
     Xbc ä def

    There is no portable for an application to determine ‘equivalence class’ without knowledge of libc internals. FreeBSD’s tr is supposed to be able to do it by assuming it knows the internals of its libc: https://github.com/freebsd/freebsd/blob/master/usr.bin/tr/str.c#L212.

    A lot depends on the system’s libc. For example, the following works on glibc but not on Mac (in both cases using GNU sed):

     # works on glibc with gnu sed:
     $ printf 'abc \303\244\303\202 def\n' | LC_ALL=en_US.UTF-8 sed 's/[[=a=]]/X/g' 
     Xbc XX def

Implementation issues: The critical strucutres in the tr code.

Other places also assume only 256 different values (e.g. enum { N_CHARS = UCHAR_MAX + 1 };).


character ‘WJ’ (word-joiner) - special treatment in ‘fold / fmt’?

Does any ‘space’ character is space, or ‘iswspace’, or only ASCII 0x20,0x09,0x0d ?


FreeBSD’s join bails out on invalid sequences: see function ‘mbssep()’ in https://github.com/freebsd/freebsd/blob/master/usr.bin/join/join.c#L362.

Currently join DOES support some locale-comparison, as fields are compared with gnulib’s memcoll (which uses strcoll(3) internally).

Two things that are not supported (and are partially implemented in redhat’s i18n patch):

  1. multibyte field delimiters - but the patch turns the global delimiter variable into a string, making processing slower in all cases.

  2. Case-insensitive comparison - the patch allocates new buffers for every key (in every line) and iterates with mbrtowc+towupper.

Also, risk of collating into same order (cf. Karl Berry surprised results from sort in bug#23677.



$ printf "\u03a8\n" | od -tx1c -An
   ce  a8  0a
  316 250  \n

In Mac/FreeBSD:

$ printf  "\u03A8\n" | od -t x1c -An
  ce  a8  0a
  Ψ   **  \n

in Mac/FreeBSD: invalid mb-seqeuences:

$ printf  "\xce\xce\n" | od -t x1c
  ce  ce  0a
 316 316  \n

Implementation problem: POSIX says the FIRST character of a valid multibyte sequence should display the character, and the following octets should show ‘’. But the first octet might appear on the LAST character of the line, and the ‘’ should be displayed on the following line.

In FreeBSD, it ‘just works’:

$ printf "aaaaaaaaaaaaaaa\316\250bb\n" | od -An -tc
       a   a   a   a   a   a   a   a   a   a   a   a   a   a   a   Ψ
      **   b   b  \n

They (FreeBSD) have implemented a ‘peek’ option following a multibyte octet: https://github.com/freebsd/freebsd/blob/master/usr.bin/hexdump/conv.c#L98.

On GNU coreutils’ od, it seems (IIUC) that the implementation reads exactly the (known) amounts of octets needed to display each line, and adding ‘peeking’ feature will be tricky: http://git.savannah.gnu.org/cgit/coreutils.git/tree/src/od.c#n1360


TODO: organize this mess…

Check normalization according to NormalizationTest.txt

Check compatiblity of:







Unexpected?? U+00E6 LATIN SMALL LETTER AE (æ) is NOT decomposed:

“Originally a ligature representing a Latin diphthong, it has been promoted to the full status of a letter in the alphabets of some languages, including Danish, Norwegian, Icelandic and Faroese.”

  • https://codepoints.net/U+00E6

However this is decomposable:


Check decomposition of: wcwidth() on LATIN-EXtended-B characters: \u01c4 Latin Capital Letter DZ with caron \u01c5 Latin Capital Letter D with Small Letter Z with caron \u01c6 Latin Small Letter DZ with caron \u01c7 Latin Capital Letter LJ \u01c8 Latin Capital Letter L with Small Letter J \u01c9 Latin Small Letter LJ \u01ca Latin Capital Letter NJ \u01cb Latin Capital Letter N with Small Letter J \u01cc Latin Small Letter NJ

  \u01f1 Latin Capital Letter DZ
  \u01f2 Latin Capital Letter D with Small Letter
  \u01f3 Latin Small Letter DZ

  \u1f6 Latin Capital Letter Hwair (seems to work through with wcwidth==1)

  \u01fC Latin Capital Letter AE with acute Ǽ (seems to woth with wcwidth==1)

  \u0238 Latin Small Letter DB Digraph ȸ (seems ok) - despite being called digraphs
  \u0239 Latin Small Letter QP Digraph ȹ (seems ok)

ESPECIALLY the “DZ with Caron” - is it decomposed to D,Z,Caron or D,z-with-caron ?

and these two: does decompistion results in O,dot,macron ?

\u0230 Latin Capital Letter O with dot above and macron Ȱ -
\u0231 Latin Small Letter O with dot above and macron ȱ

Check decomposition/compatiblity of IPA block, e.g.

\u2A3 Latin Small Letter DZ Digraph ʣ - does this translates to DZ, or
      to the Latin-Extended-B 'DZ' latter?
up to and including:
\u2AB Latin Small Letter LZ Digraph

Check compatability and decomposition of ‘fullwidth’ characters, see https://en.wikipedia.org/wiki/Halfwidth_and_fullwidth_forms#In_Unicode e.g. does ‘\uFF21’ (full-width A) decomposes to ascii ‘A’ ?

Check compatibility/decomposition of entire block: https://en.wikipedia.org/wiki/Alphabetic_Presentation_Forms


TODO: organize this mess…

Modified NULL (\xC0\x80)

(Unicode book page 282): Unicode conformance

 UNAssigned Code Points (C4)
 Test unassigned codes (don't generate, don't change) in all programs.
 Test non-characters (U+FFFE, U+FFFF)
 Test surrogate codes

Surrogate codepoints treated as invalid on “normal” unixes:

$ printf '\uD800\n' | ./src/mbbuffer-test -r
ofs  line colB colC V wc(dec) wc(hex) Ch w n octets
0    1    1    1    n       *       * *  * 1 0xed
1    1    2    2    n       *       * *  * 1 0xa0
2    1    3    3    n       *       * *  * 1 0x80
3    1    4    4    y      10 0x0000a =  -1 1 0x0a

but on Cygwin:

Administrator@WIN-9FFSHRJAFVN ~/coreutils-8.25.71-1437c
$  printf '\uD800\n' | ./src/mbbuffer-test -r
ofs  line colB colC V wc(dec) wc(hex) Ch  W n octets
0    1    1    1    y   55296 0x0d800 =  -1 3 0xed 0xa0 0x80
3    1    4    2    y      10 0x0000a =  -1 1 0x0a

page 437: Check all special characters and their effects in various programs.

TODO for Book: Show examples of conversion cases in page 501/502.

 sed (and grep/gawk) NEVER match regular expressions to invalid
   multibyte sequences. To Force matching, use LC_ALL=C.

 $ printf '\xe1\xbc\x11' | LC_ALL=C ./sed/sed 's/./X/g' | od -tx1
 0000000 58 58 58

TODO: Special handling for “modified UTF-8” with NULL as

  UTF-8 "\xC0\x80" ?
  system's native mbrtowc does not handle it,
  and will return -1 .

TODO: prepare for all types of invalid sequences:

  https://en.wikipedia.org/wiki/UTF-8#Invalid_byte_sequences :
  Not all sequences of bytes are valid UTF-8. A UTF-8 decoder should be prepared for:
  * the red invalid bytes in the above table
  * an unexpected continuation byte
  * a leading byte not followed by enough continuation bytes (can happen in
    simple string truncation, when a string is too long to fit when copying it)
  * an overlong encoding as described above
  * a sequence that decodes to an invalid code point as described below

  https://en.wikipedia.org/wiki/UTF-8#Invalid_code_points :
  Since RFC 3629 (November 2003), the high and low
  surrogate halves used by UTF-16 (U+D800 through U+DFFF) and code
  points not encodable by UTF-16 (those after U+10FFFF) are not
  legal Unicode values, and their UTF-8 encoding must be treated
  as an invalid byte sequence.

  Not decoding surrogate halves makes it impossible to store
  invalid UTF-16, such as Windows filenames, as UTF-8. Therefore,
  detecting these as errors is often not implemented and there are
  attempts to define this behavior formally (see WTF-8 and CESU

from https://www.cl.cam.ac.uk/~mgk25/ucs/examples/UTF-8-test.txt:

Section3.3 Sequences with last continuation byte missing:

All bytes of an incomplete sequence should be signalled as a single
malformed sequence, i.e., you should see only a single replacement
character in each of the next 10 tests. (Characters as in section 2).

Mbbuffer currently reports EACH invalid octet instead of just one per incomplete sequence.

TODO: Does incomplete sequence in the middle of the file reported as incomplete (mbrtowc==-2) or invalid (mbrtowc==-1) ?

If we report on the FIRST octet (including line,byte/char offset), the user (needing low-level processing) won’t be able to tell the differences without further processing. By reporting all octets, we provide easier work-arounds (but we also ‘pollute’ stdout with more “invalid char” markers than needed). Perhaps add this as an option?

On Ubuntu 14.04 with xterm 322, terminal prints only one “invalid char”:

$ printf '\ud800\n'

On Ubuntu 14.04 with gnome-terminal 3.6.2, nothing is printf.

Mac OS X terminal prints 3 question marks:

$ printf '\ud800\n'

Web browsers print 3 characters: visit https://www.cl.cam.ac.uk/~mgk25/ucs/examples/UTF-8-test.txt and view section 5.1.1.

On Linux: Google Chrom 51.0, Firefox 48.0 On Mac: Safari 8.0.7, Chrome 47.0, IceCat 37.0

$ printf '\367\277\277\n' | ./src/mbbuffer-test.exe -r
ofs  line colB colC V wc(dec) wc(hex) Ch  W n octets
0    1    1    1    n       *       * *   * 1 0xf7
1    1    2    2    n       *       * *   * 1 0xbf
2    1    3    3    n       *       * *   * 1 0xbf
3    1    4    4    y      10 0x0000a =  -1 1 0x0a

Also see seciont 5.1 (Single UTF-16 surrogates) where he claims each invalid sequence should result in ONE “invalid character” output: e.g.

5.1.1  U+D800 = ed a0 80 = "���"

In such cases ‘mbrtowc’ return -1 THREE times (or is it because we reset mbstate_t after each failure?)


"[...] U+FFFE and U+FFFF must not occur in normalUTF-8 or UCS-4
data. UTF-8 decoders should treat them like malformedor overlong
sequences for safety reasons."

Unicode Book



Unicode Explained

Internationalize Documents, Programs, and Web Sites

By Jukka K. Korpela


The following are pointers and notes from the above book which seemd (IMHO) relevant for coreutils’ multibyte implementation (or for testing).

page 11: Finish har har har - change with char-classes, regex, normalization, upper/lower cases

page 12: ligature “fi”: change for normalization, char classes

page 23: German lower-case “strasse” (sharp-S?) becomes “SS” in upper-case (two characters). Also differ from greek “beta” glyph. ‘\u00DF’

page23: 0-with-cross is diameter in mechanical writing, or a letter in Nordic languages?

Length of BIDI markers (zero width?)

page 29 (2nd paragraph from bottom): Greek Sigma in middle vs final form. If there’s no equivalence between them, how about sort order ?

page 29 (top): initial/middle/final/separate contextual forms (e.g hebrew/arabic) Sort order ?

Page 143: Transcoding tools http://www.unicode.org/Public/MAPPINGS TODO: Download these for offline processing.

Page 145: Repertiore requirements Characters in each language: http://www.eki.ee/letter

Page 169: Named Sequences http://www.unicode.org/Public/UNIDATA/NamedSequences.txt

Page 178: Table 4.3: Code-point Classification TODO: Test unassigned, Surrogate, Private-Use input. Ensure no bugs, should be passed as it. What about “wc” and “cut” ?

Page 182: DiGraph e.g “ll” in Spanish, “Ch” in some others - two distinct characters logically treated as one by native language speakers. VS æ (\u00E6) which is one character for “ae”. ij (\u0133) which is small latin ligature “ij”. TODO: Check unicode-normalization-decomposition.

Page 185: unicode standard - chapters Chapter 5: Implementation Guidelines

Page 194: Varient Selectors Unicode markers affecting the precending code-point, ∩︀ (\u2229 - “intersection” symbol) followed by \uFE00 (“variant selector” VS1). Affect font in applicaiton ? IS this Zero width character ? TODO: check with ‘expand’, ‘cut’, ‘wc’.

Page 195: Ligatures. in Danish/Norwegian, æ (\u00E6) is an independent letter, vs just a ligature of two letters “ae” in other languages. TODO: Test sort order with such input in Danish-vs-other locales. TODO: in Danish locale, should unicode normaliztion NOT decompose it?? Unicode “Alphabetic PResentation Form” block (U+FB00..U+FB4F).

 TODO: Test decomposition of ligatures in that block (e.g. hebrew ligatures?)
 $ printf '\ufb00\n'
 $ printf '\ufb03\n'
 $ printf '\ufb4a\n' תּ

 ZWJ (\u200d) should instruct the application to join the
 characters before/afer into legature.
 Doesn't seem to work (on Mac OS):
    $ printf 'f\u200Di\n'
    f i

 Similarly, ZWNJ (\u200C) should prevent joining.
 TODO: test ZWJ,ZWNJ (zero width or "invisible control" chars?)
 in cut/expand/wc.

Page 196: Vowels vs Marks Hebrew+Arabic: Nikud.

 Hindi (Devanagaris script):
 \u092A (pa) followed by \u0942 (uu) appears as one glyph (puu).

Page 211, Table 5-1: General Category VAlues.

Page 216: Character Property ‘ea’ = Asian Width Full,Half,narrow. Affects ‘expand’ ?

Page 216: Grapheme Clusters? for ‘fold/fmt/cut’ ?

Page 219,220: Use ‘WB’ (WordBreak)’ or ‘WS’ (Whitespace) ‘SB’ (Sentense-break) properties for counding-words in ‘wc’ ?

 gfdafda d dfsa fdsa fdsa
 fdafda fdsafdsa

Page 220: Property ‘SFC’ (Simple-Case-Folding): Upper/Lower case are simple.

Page 227: Canonical vs Compatability mapping Canonical: different encoding for SAME symbol. Compatibility: fundamentally similar characters, differ in rendering/ usage (and sometimes in meaning)

 Examples in Book
 \u2126 = \u03a9

Page 231: Iterative decompisition:

 ANGSTRAM (U+212b):
   $ printf '\u212b\n'
 Is canonical-mapped to 'A-with-Ring U+00C5':
   $ printf '\u00c5\n'
 Which is canonical mapped to 'A + combining mark ring (U+030a)':
   $ printf 'A\u30a\n'

Page 233:

 Decomposition of 'VULGAR HALF', 'MICRO SIGN', 'E WITH GRAVE':

   $ printf '\u00BD\u00b5\u00e8\n'

    MICRO SIGN  => greek mu

With decomposition (E + combing grave mark):
  $ printf '\u00BD\u00b5\u00e8\n'| ./src/unorm -n nfkd  | iconv -t ucs-2le | od -tx2 -An
  0031 2044 0032 03bc 0065 0300 000a

Without decomposition ('E WITH GRAVE' stays as-is):
  $ printf '\u00BD\u00b5\u00e8\n'| ./src/unorm -n nfkc  | iconv -t ucs-2le | od -tx2
  0031 2044 0032 03bc 00e8 000a

TODO: for 'sort', 'uniq', 'join':
   Test the above strings as 'equivalent' (strxfrm/strcoll) ?

Page 249: Collation order no official collation order. Unicode Technical STandard #10 http://www.unicode.org/reports/tr30/tr30-4.html

Page 256: Text Boundaries See the files in /Users/gordon/projects/unicode-mapping/www.unicode.org/Public/9.0.0/ucd/auxiliary like WordBreakProprty.txt includes test files TODO: for wc,fold,fmt,cut ? TODO: instead of ‘iswspace’ is unicode ‘‘Alphabetic’ Property?

 For Book: document exceptions for Thai/Lao/Hiragana ?

Page 276: Line-BReaking rules for fold/fmt ?

Page 282: Unicode Conformance requirements TODO: Test unassigned codes (don’t generate, don’t change) in all programs.

Page 285: Conformance: C12a: unorm is conformant.

Page 286: Conformance: C14,15,16 (normalization): unorm is conformant.

page 287: Conformance: When mentioning normalization, use proper terms (for unorm)

Page 299: UTF-8 vs ISO-8859-1 For Book

Page 300: Duplicate Octet Range rable, add octal

Page 392: Duplicate table of control characters, add octal mention sed, printf, od for book/ /website

Page 414: Fixed with charachers (e.g. em/en dashes) TODO: How to treat in ‘expand’ ?

page 426: Line-break chracters in unicode for fold/fmt , what about ‘wc’ ? LS (U+2028) Line Separator PS (U+2029) Paragraph separator For Book/website

Page 426: mathenatical and technical symbols For Book/website: canocnical compatiblity with other chars.

Page 438: ‘other’ non alphabetical markers, should they be counted as words?

 $ env printf '\ufff9assaf\ufffagordon\ufffb\n' |  wc
 1       1      21

 in HTML this would be rendered as two words, assaf/gordon.

Page 468: Invisible characters ?

Page 469: MArkup vs plaintext: Table 9-2: should these characters be counted in ‘wc’, skipped in ‘expand’, non-break with ‘cut’ ?

 Are these considerd "word break" properties?
 should SED's "\b \B \< \>" regex operators support them?

Page 592: Patterns, regex patterns. TODO: ensure tests cases according to page 594.

page 597: “Basic Unicode Support” TODO: Check which coreutils fall under the requirements, and whether they comply.

Last but not least

Support the cause: Adopt a Unicode Character! http://www.unicode.org/consortium/adopted-characters.html

Unorganized (yet)

OpenBSD removes non-utf8 locales: http://marc.info/?l=openbsd-cvs&m=143956261214725&w=2


TODO: learn from grep’s multibyte-white-space test

← more articles