The history of UTF-8 as told by Rob Pike

Rob Pike explains how Ken Thompson invented UTF-8 in one evening and how they together built the first system-wide implementation in less than a week.

Subject: UTF-8 history
From: "Rob 'Commander' Pike" <r (at) google.com>
Date: Wed, 30 Apr 2003 22:32:32 -0700 (Thu 06:32 BST)
To: mkuhn (at) acm.org, henry (at) spsystems.net
Cc: ken (at) entrisphere.com

Looking around at some UTF-8 background, I see the same incorrect
story being repeated over and over.  The incorrect version is:
    1. IBM designed UTF-8.
    2. Plan 9 implemented it.
That's not true.  UTF-8 was designed, in front of my eyes, on a
placemat in a New Jersey diner one night in September or so 1992.

What happened was this.  We had used the original UTF from ISO 10646
to make Plan 9 support 16-bit characters, but we hated it.  We were
close to shipping the system when, late one afternoon, I received a
call from some folks, I think at IBM - I remember them being in Austin
- who were in an X/Open committee meeting.  They wanted Ken and me to
vet their FSS/UTF design.  We understood why they were introducing a
new design, and Ken and I suddenly realized there was an opportunity
to use our experience to design a really good standard and get the
X/Open guys to push it out.  We suggested this and the deal was, if we
could do it fast, OK.  So we went to dinner, Ken figured out the
bit-packing, and when we came back to the lab after dinner we called
the X/Open guys and explained our scheme.  We mailed them an outline
of our spec, and they replied saying that it was better than theirs (I
don't believe I ever actually saw their proposal; I know I don't
remember it) and how fast could we implement it?  I think this was a
Wednesday night and we promised a complete running system by Monday,
which I think was when their big vote was.

So that night Ken wrote packing and unpacking code and I started
tearing into the C and graphics libraries.  The next day all the code
was done and we started converting the text files on the system
itself.  By Friday some time Plan 9 was running, and only running,
what would be called UTF-8.  We called X/Open and the rest, as they
say, is slightly rewritten history.

Why didn't we just use their FSS/UTF?  As I remember, it was because
in that first phone call I sang out a list of desiderata for any such
encoding, and FSS/UTF was lacking at least one - the ability to
synchronize a byte stream picked up mid-run, with less that one
character being consumed before synchronization.  Becuase that was
lacking, we felt free - and were given freedom - to roll our own.

I think the "IBM designed it, Plan 9 implemented it" story originates
in RFC2279.  At the time, we were so happy UTF-8 was catching on we
didn't say anything about the bungled history.  Neither of us is at
the Labs any more, but I bet there's an e-mail thread in the archive
there that would support our story and I might be able to get someone
to dig it out.

So, full kudos to the X/Open and IBM folks for making the opportunity
happen and for pushing it forward, but Ken designed it with me
cheering him on, whatever the history books say.

-rob

Date: Sat, 07 Jun 2003 18:44:05 -0700
From: "Rob `Commander' Pike" <r@google.com>
To: Markus Kuhn <Markus.Kuhn@cl.cam.ac.uk>
cc: henry@spsystems.net, ken@entrisphere.com,
   Greger Leijonhufvud <greger@friherr.com>
Subject: Re: UTF-8 history

I asked Russ Cox to dig through the archives. I have attached his message.
I think you'll agree it supports the story I sent earlier. The mail we
sent to X/Open (I believe Ken did the editing and mailing of that document)
includes a new desideratum #6 about discovering character boundaries.
We'll never know how much the original X/Open proposal influenced us;
the two proposals are very different but do share some characteristics.
I don't remember looking at it in detail, but it was a long time ago.
I very clearly remember Ken writing on the placemat and wished we had
kept it!

-rob

From: Russ Cox <rsc@plan9.bell-labs.com>
To: r@google.com
Subject: utf digging
Date-Sent: Saturday, June 07, 2003 7:46 PM -0400


bootes's /sys/src/libc/port/rune.c changed from the
division-heavy old utf on sep 4 1992.
the version that made it into the dump
is dated 19:51:55.  it was commented
the next day but otherwise remained unchanged
until nov 14 1996, when runelen was sped up by
inspecting the rune explicitly rather than
using runetochar's return value.  may 26 2001
was the next and last change, to add runenlen.

here are some mails from your mail boxes
that turn up by grepping for utf.  the first
refers to utf.c, which is a copy of a wctomb and mbtowc
that handle the full 6-byte utf-8 encoding of 32-bit runes.
it's quite ugly, with all the logic in control flow.
i assume it became the code in the proposal
as a result of that first mail.

in /usr/ken/utf/xutf i found a copy of what
appears to be the original not-self-synchronizing
encoding proposal, with the utf-8 scheme tacked
onto the end (starting at "We define 7 byte types").
that's also below.  the version below is the first,
dated sep 2 23:44:10.  it went through a number of
edits to become the second mail below by the
morning of Sep 8.

the mail log shows that second mail going out
as well as taking a while to come back to ken.

helix: Sep  8 03:22:13: ken: upas/sendmail: remote inet!xopen.co.uk!xojig 
>From ken Tue Sep  8 03:22:07 EDT 1992 (xojig@xopen.co.uk) 6833
helix: Sep  8 03:22:13: ken: upas/sendmail: delivered rob From ken Tue Sep 
8 03:22:07 EDT 1992 6833
helix: Sep  8 03:22:16: ken: upas/sendmail: remote pyxis!andrew From ken 
Tue Sep  8 03:22:07 EDT 1992 (andrew) 6833
helix: Sep  8 03:22:19: ken: upas/sendmail: remote coma!dmr From ken Tue 
Sep  8 03:22:07 EDT 1992 (dmr) 6833
helix: Sep  8 03:25:52: ken: upas/sendmail: delivered rob From ken Tue Sep 
8 03:24:58 EDT 1992 141
helix: Sep  8 03:36:13: ken: upas/sendmail: delivered ken From ken Tue Sep 
8 03:36:12 EDT 1992 6833

enjoy.



>From ken Fri Sep  4 03:37:39 EDT 1992
you might want to look at
    /usr/ken/utf/utf.c
and see if you can make it prettier.

>From ken Tue Sep  8 03:22:07 EDT 1992
Here is our modified FSS-UTF proposal.  The words are the same as on
the previous proposal.  My apologies to the author.  The code has been
tested to some degree and should be pretty good shape.  We have
converted Plan 9 to use this encoding and are about to issue a
distribution to an initial set of university users.

File System Safe Universal Character Set Transformation Format (FSS-UTF)
--------------------------------------------------------------------------

With the approval of ISO/IEC 10646 (Unicode) as an international
standard and the anticipated wide spread use of this universal coded
character set (UCS), it is necessary for historically ASCII based
operating systems to devise ways to cope with representation and
handling of the large number of characters that are possible to be
encoded by this new standard.

There are several challenges presented by UCS which must be dealt with
by historical operating systems and the C-language programming
environment.  The most significant of these challenges is the encoding
scheme used by UCS. More precisely, the challenge is the marrying of
the UCS standard with existing programming languages and existing
operating systems and utilities.

The challenges of the programming languages and the UCS standard are
being dealt with by other activities in the industry.  However, we are
still faced with the handling of UCS by historical operating systems
and utilities.  Prominent among the operating system UCS handling
concerns is the representation of the data within the file system.  An
underlying assumption is that there is an absolute requirement to
maintain the existing operating system software investment while at
the same time taking advantage of the use the large number of
characters provided by the UCS.

UCS provides the capability to encode multi-lingual text within a
single coded character set.  However, UCS and its UTF variant do not
protect null bytes and/or the ASCII slash ("/") making these character
encodings incompatible with existing Unix implementations.  The
following proposal provides a Unix compatible transformation format of
UCS such that Unix systems can support multi-lingual text in a single
encoding.  This transformation format encoding is intended to be used
as a file code.  This transformation format encoding of UCS is
intended as an intermediate step towards full UCS support.  However,
since nearly all Unix implementations face the same obstacles in
supporting UCS, this proposal is intended to provide a common and
compatible encoding during this transition stage.


Goal/Objective
--------------

With the assumption that most, if not all, of the issues surrounding
the handling and storing of UCS in historical operating system file
systems are understood, the objective is to define a UCS
transformation format which also meets the requirement of being usable
on a historical operating system file system in a non-disruptive
manner.  The intent is that UCS will be the process code for the
transformation format, which is usable as a file code.

Criteria for the Transformation Format
--------------------------------------

Below are the guidelines that were used in defining the UCS
transformation format:

    1) Compatibility with historical file systems:

    Historical file systems disallow the null byte and the ASCII
    slash character as a part of the file name.

    2) Compatibility with existing programs:

    The existing model for multibyte processing is that ASCII does
    not occur anywhere in a multibyte encoding.  There should be
    no ASCII code values for any part of a transformation format
    representation of a character that was not in the ASCII
    character set in the UCS representation of the character.

    3) Ease of conversion from/to UCS.

    4) The first byte should indicate the number of bytes to
    follow in a multibyte sequence.

    5) The transformation format should not be extravagant in
    terms of number of bytes used for encoding.

    6) It should be possible to find the start of a character
    efficiently starting from an arbitrary location in a byte
    stream.


Proposed FSS-UTF
----------------

The proposed UCS transformation format encodes UCS values in the range
[0,0x7fffffff] using multibyte characters of lengths 1, 2, 3, 4, 5,
and 6 bytes.  For all encodings of more than one byte, the initial
byte determines the number of bytes used and the high-order bit in
each byte is set.  Every byte that does not start 10xxxxxx is the
start of a UCS character sequence.

An easy way to remember this transformation format is to note that the
number of high-order 1's in the first byte signifies the number of
bytes in the multibyte character:

   Bits  Hex Min  Hex Max  Byte Sequence in Binary
1    7  00000000 0000007f 0vvvvvvv
2   11  00000080 000007FF 110vvvvv 10vvvvvv
3   16  00000800 0000FFFF 1110vvvv 10vvvvvv 10vvvvvv
4   21  00010000 001FFFFF 11110vvv 10vvvvvv 10vvvvvv 10vvvvvv
5   26  00200000 03FFFFFF 111110vv 10vvvvvv 10vvvvvv 10vvvvvv 10vvvvvv
6   31  04000000 7FFFFFFF 1111110v 10vvvvvv 10vvvvvv 10vvvvvv 10vvvvvv 
10vvvvvv

The UCS value is just the concatenation of the v bits in the multibyte
encoding.  When there are multiple ways to encode a value, for example
UCS 0, only the shortest encoding is legal.

Below are sample implementations of the C standard wctomb() and
mbtowc() functions which demonstrate the algorithms for converting
from UCS to the transformation format and converting from the
transformation format to UCS. The sample implementations include error
checks, some of which may not be necessary for conformance:

typedef
struct
{
    int cmask;
    int cval;
    int shift;
    long    lmask;
    long    lval;
} Tab;

static
Tab tab[] =
{
    0x80,   0x00,   0*6,    0x7F,       0,      /* 1 byte sequence */
    0xE0,   0xC0,   1*6,    0x7FF,      0x80,       /* 2 byte sequence */
    0xF0,   0xE0,   2*6,    0xFFFF,     0x800,      /* 3 byte sequence */
    0xF8,   0xF0,   3*6,    0x1FFFFF,   0x10000,    /* 4 byte sequence */
    0xFC,   0xF8,   4*6,    0x3FFFFFF,  0x200000,   /* 5 byte sequence */
    0xFE,   0xFC,   5*6,    0x7FFFFFFF, 0x4000000,  /* 6 byte sequence */
    0,                          /* end of table */
};

int
mbtowc(wchar_t *p, char *s, size_t n)
{
    long l;
    int c0, c, nc;
    Tab *t;

    if(s == 0)
        return 0;

    nc = 0;
    if(n <= nc)
        return -1;
    c0 = *s & 0xff;
    l = c0;
    for(t=tab; t->cmask; t++) {
        nc++;
        if((c0 & t->cmask) == t->cval) {
            l &= t->lmask;
            if(l < t->lval)
                return -1;
            *p = l;
            return nc;
        }
        if(n <= nc)
            return -1;
        s++;
        c = (*s ^ 0x80) & 0xFF;
        if(c & 0xC0)
            return -1;
        l = (l<<6) | c;
    }
    return -1;
}

int
wctomb(char *s, wchar_t wc)
{
    long l;
    int c, nc;
    Tab *t;

    if(s == 0)
        return 0;

    l = wc;
    nc = 0;
    for(t=tab; t->cmask; t++) {
        nc++;
        if(l <= t->lmask) {
            c = t->shift;
            *s = t->cval | (l>>c);
            while(c > 0) {
                c -= 6;
                s++;
                *s = 0x80 | ((l>>c) & 0x3F);
            }
            return nc;
        }
    }
    return -1;
}

>From ken Tue Sep  8 03:24:58 EDT 1992
i mailed it out, but it went into a black hole.
i didnt get my copy. it must be hung up on the
internat address with coma down or something.

>From ken Tue Sep  8 03:42:43 EDT 1992
i finally got my copy.



--- /usr/ken/utf/xutf from dump of Sep 2 1992 ---

 File System Safe Universal Character Set Transformation Format (FSS-UTF)
 --------------------------------------------------------------------------

 With the approval of ISO/IEC 10646 (Unicode) as an international
 standard and the anticipated wide spread use of this universal coded
 character set (UCS), it is necessary for historically ASCII based
 operating systems to devise ways to cope with representation and
 handling of the large number of characters that are possible to be
 encoded by this new standard.

 There are several challenges presented by UCS which must be dealt with
 by historical operating systems and the C-language programming
 environment. The most significant of these challenges is the encoding
 scheme used by UCS.  More precisely, the challenge is the marrying of
 the UCS standard with existing programming languages and existing
 operating systems and utilities.

 The challenges of the programming languages and the UCS standard are
 being dealt with by other activities in the industry.  However, we are
 still faced with the handling of UCS by historical operating systems and
 utilities. Prominent among the operating system UCS handling concerns is
 the representation of the data within the file system. An underlying
 assumption is that there is an absolute requirement to maintain the
 existing operating system software investment while at the same time
 taking advantage of the use the large number of characters provided by
 the UCS.

 UCS provides the capability to encode multi-lingual text within a single
 coded character set.  However, UCS and its UTF variant do not protect
 null bytes and/or the ASCII slash ("/") making these character encodings
 incompatible with existing Unix implementations.  The following proposal
 provides a Unix compatible transformation format of UCS such that Unix
 systems can support multi-lingual text in a single encoding.  This
 transformation format encoding is intended to be used as a file code.
 This transformation format encoding of UCS is intended as an
 intermediate step towards full UCS support.  However, since nearly all
 Unix implementations face the same obstacles in supporting UCS, this
 proposal is intended to provide a common and compatible encoding during
 this transition stage.


 Goal/Objective
 --------------

 With the assumption that most, if not all, of the issues surrounding the
 handling and storing of UCS in historical operating system file systems
 are understood, the objective is to define a UCS transformation format
 which also meets the requirement of being usable on a historical
 operating system file system in a non-disruptive manner. The intent is
 that UCS will be the process code for the transformation format, which
 is usable as a file code.

 Criteria for the Transformation Format
 --------------------------------------

 Below are the guidelines that were used in defining the UCS
 transformation format:

     1) Compatibility with historical file systems:

    Historical file systems disallow the null byte and the ASCII
    slash character as a part of the file name.

     2) Compatibility with existing programs:

    The existing model for multibyte processing is that ASCII does
    not occur anywhere in a multibyte encoding.  There should be no
    ASCII code values for any part of a transformation format
    representation of a character that was not in the ASCII character
    set in the UCS representation of the character.

     3) Ease of conversion from/to UCS.

     4) The first byte should indicate the number of bytes to follow in a
    multibyte sequence.

     5) The transformation format should not be extravagant in terms of
    number of bytes used for encoding.


 Proposed FSS-UTF
 ----------------

 The proposed UCS transformation format encodes UCS values in the range
 [0,0x7fffffff] using multibyte characters of lengths 1, 2, 3, 4, and 5
 bytes.  For all encodings of more than one byte, the initial byte
 determines the number of bytes used and the high-order bit in each byte
 is set.

 An easy way to remember this transformation format is to note that the
 number of high-order 1's in the first byte is the same as the number of
 subsequent bytes in the multibyte character:

    Bits  Hex Min  Hex Max         Byte Sequence in Binary
 1    7  00000000 0000007f 0zzzzzzz
 2   13  00000080 0000207f 10zzzzzz 1yyyyyyy
 3   19  00002080 0008207f 110zzzzz 1yyyyyyy 1xxxxxxx
 4   25  00082080 0208207f 1110zzzz 1yyyyyyy 1xxxxxxx 1wwwwwww
 5   31  02082080 7fffffff 11110zzz 1yyyyyyy 1xxxxxxx 1wwwwwww 1vvvvvvv

 The bits included in the byte sequence is biased by the minimum value
 so that if all the z's, y's, x's, w's, and v's are zero, the minimum
 value is represented.  In the byte sequences, the lowest-order encoded
 bits are in the last byte; the high-order bits (the z's) are in the
 first byte.

 This transformation format uses the byte values in the entire range of
 0x80 to 0xff, inclusive, as part of multibyte sequences.  Given the
 assumption that at most there are seven (7) useful bits per byte, this
 transformation format is close to minimal in its number of bytes used.

 Below are sample implementations of the C standard wctomb() and
 mbtowc() functions which demonstrate the algorithms for converting from
 UCS to the transformation format and converting from the transformation
 format to UCS.  The sample implementations include error checks, some
 of which may not be necessary for conformance:

#define OFF1   0x0000080
#define OFF2   0x0002080
#define OFF3   0x0082080
#define OFF4   0x2082080

int wctomb(char *s, wchar_t wc)
{
       if (s == 0)
           return 0;       /* no shift states */
#ifdef wchar_t_is_signed
       if (wc < 0)
           goto bad;
#endif
       if (wc <= 0x7f)         /* fits in 7 bits */
       {
           s[0] = wc;
           return 1;
       }
       if (wc <= 0x1fff + OFF1)        /* fits in 13 bits */
       {
           wc -= OFF1;
           s[0] = 0x80 | (wc >> 7);
           s[1] = 0x80 | (wc & 0x7f);
           return 2;
       }
       if (wc <= 0x7ffff + OFF2)       /* fits in 19 bits */
       {
           wc -= OFF2;
           s[0] = 0xc0 | (wc >> 14);
           s[1] = 0x80 | ((wc >> 7) & 0x7f);
           s[2] = 0x80 | (wc & 0x7f);
           return 3;
       }
       if (wc <= 0x1ffffff + OFF3)     /* fits in 25 bits */
       {
           wc -= OFF3;
           s[0] = 0xe0 | (wc >> 21);
           s[1] = 0x80 | ((wc >> 14) & 0x7f);
           s[2] = 0x80 | ((wc >> 7) & 0x7f);
           s[3] = 0x80 | (wc & 0x7f);
           return 4;
       }
#if !defined(wchar_t_is_signed) || defined(wchar_t_is_more_than_32_bits)
       if (wc > 0x7fffffff)
           goto bad;
#endif
       wc -= OFF4;
       s[0] = 0xf0 | (wc >> 28);
       s[1] = 0x80 | ((wc >> 21) & 0x7f);
       s[2] = 0x80 | ((wc >> 14) & 0x7f);
       s[3] = 0x80 | ((wc >> 7) & 0x7f);
       s[4] = 0x80 | (wc & 0x7f);
       return 5;
bad:;
       errno = EILSEQ;
       return -1;
}


int mbtowc(wchar_t *p, const char *s, size_t n)
{
       unsigned char *uc;      /* so that all bytes are nonnegative */

       if ((uc = (unsigned char *)s) == 0)
           return 0;               /* no shift states */
       if (n == 0)
           return -1;
       if ((*p = uc[0]) < 0x80)
           return uc[0] != '\0';   /* return 0 for '\0', else 1 */
       if (uc[0] < 0xc0)
       {
           if (n < 2)
               return -1;
           if (uc[1] < 0x80)
               goto bad;
           *p &= 0x3f;
           *p <<= 7;
           *p |= uc[1] & 0x7f;
           *p += OFF1;
           return 2;
       }
       if (uc[0] < 0xe0)
       {
           if (n < 3)
               return -1;
           if (uc[1] < 0x80 || uc[2] < 0x80)
               goto bad;
           *p &= 0x1f;
           *p <<= 14;
           *p |= (uc[1] & 0x7f) << 7;
           *p |= uc[2] & 0x7f;
           *p += OFF2;
           return 3;
       }
       if (uc[0] < 0xf0)
       {
           if (n < 4)
               return -1;
           if (uc[1] < 0x80 || uc[2] < 0x80 || uc[3] < 0x80)
               goto bad;
           *p &= 0x0f;
           *p <<= 21;
           *p |= (uc[1] & 0x7f) << 14;
           *p |= (uc[2] & 0x7f) << 7;
           *p |= uc[3] & 0x7f;
           *p += OFF3;
           return 4;
       }
       if (uc[0] < 0xf8)
       {
           if (n < 5)
               return -1;
           if (uc[1] < 0x80 || uc[2] < 0x80 || uc[3] < 0x80 || uc[4] < 0x80)
               goto bad;
           *p &= 0x07;
           *p <<= 28;
           *p |= (uc[1] & 0x7f) << 21;
           *p |= (uc[2] & 0x7f) << 14;
           *p |= (uc[3] & 0x7f) << 7;
           *p |= uc[4] & 0x7f;
           if (((*p += OFF4) & ~(wchar_t)0x7fffffff) == 0)
               return 5;
       }
bad:;
       errno = EILSEQ;
       return -1;
}

We define 7 byte types:
T0  0xxxxxxx    7 free bits
Tx  10xxxxxx    6 free bits
T1  110xxxxx    5 free bits
T2  1110xxxx    4 free bits
T3  11110xxx    3 free bits
T4  111110xx    2 free bits
T5  111111xx    2 free bits

Encoding is as follows.
>From hex   Thru hex    Sequence        Bits
00000000    0000007f    T0          7
00000080    000007FF    T1 Tx           11
00000800    0000FFFF    T2 Tx Tx        16
00010000    001FFFFF    T3 Tx Tx Tx     21
00200000    03FFFFFF    T4 Tx Tx Tx Tx      26
04000000    FFFFFFFF    T5 Tx Tx Tx Tx Tx   32

Some notes:

1. The 2 byte sequence has 2^11 codes, yet only 2^11-2^7
are allowed. The codes in the range 0-7f are illegal.
I think this is preferable to a pile of magic additive
constants for no real benefit. Similar comment applies
to all of the longer sequences.

2. The 4, 5, and 6 byte sequences are only there for
political reasons. I would prefer to delete these.

3. The 6 byte sequence covers 32 bits, the FSS-UTF
proposal only covers 31.

4. All of the sequences synchronize on any byte that is
not a Tx byte.