## LATEX-L@LISTSERV.UNI-HEIDELBERG.DE

 Options: Use Forum View Use Monospaced Font Show Text Part by Default Show All Mail Headers Message: [<< First] [< Prev] [Next >] [Last >>] Topic: [<< First] [< Prev] [Next >] [Last >>] Author: [<< First] [< Prev] [Next >] [Last >>]

 Subject: Re: LaTeX's internal char prepresentation (UTF8 or Unicode?) From: Frank Mittelbach <[log in to unmask]> Reply To: Mailing list for the LaTeX3 project <[log in to unmask]> Date: Sat, 17 Feb 2001 14:20:55 +0100 Content-Type: text/plain Parts/Attachments: text/plain (246 lines)
A few people will unfortunately get this posting twice since it is both sent
to LATEX-L as well as to the Omega developers (several of which are on
LaTeX-L), sorry for that.

We thought this advisable as we make a number of suggestions regarding
extensions/changes to Omega's character token processing. (Any technical
discusion of these suggestions should probably be confined to the omega
developers list though)

In the disucssion below LICR stands for LaTeX Internal Character
Representation.

---------------------------------------------

Javier,

I use 'we' a lot here because Chris is looking over my shoulder as I type and
I am pretending that he agress with me:-). [meta remark: this was typed by
Chris -Frank]

> Which is the purpose of the LICR? Apparently, it's only an
> intermediate step before creating the final output. That

that is not at all the way we understand this process.

what i'm refering to is the "only" in your statement. clearly it is a step in
the sequence from source to final output.

The LICR is the representation that is to be used when, as Chris put it, LaTeX
reasons about the character data and manipulates it. Part of this process is
collection of a TOC is something we think should happen while LaTeX keeps all
its data in LICR form.

As a consequence of this we consider writing to files with the purpose of
rereading that file back in is something that has to happen within the LICR
context since only then can LaTeX reprocess this data properly. (There are
other forms of writing to files, or to the terminal, where LaTeX should
(conceptually) leave the LICR and convert the data to a suitable output
representation).

So our model is something like this

- trans C (eg Uppercasing)
|      |
|      |
(eg 8bit,utf8)           V      |
é    --- trans A  -->    LICR +------>  trans B  --> ^^e9
^      |
|      |
|      |
trans D (eg generating TOC)

with current LATeX based on TeX we have

trans A = the inputenc method
LICR    = LaTeX Internal Character Representation
is a unique representation of characters by 7bit charcater tokens
plus expansion invariant cs-names (which in extenal files are also
represented by 7bit strings)
trans B = fontenc translations when typesetting (producing hlists)
trans C = \MakeUppercase etc
trans D = writing to .aux files and reading them back in
puttings things into marks and manipulating them,
etc

so entering the LICR is done via one process (trans A) and then all the
reasoning and data manipulation happens within the LICR context and only the
at the final stage do we leave the LICR, eg you typeset something (aka spots
on paper :-) or you display a message on the terminal

due to TeX limitations some of the system output (eg log file) can't be fully
controlled (eg overfull hbox  is displaying data from after trans B instead of
displaying that with a transformation suitable to the target device).

> can be true in TeX, but not in Omega because the LICR can
> be processed by external tools (spelling, syntax, etc.)
> There are lots of tools using Unicode and very likely there
> will be more in a future. However, there are only a handful
> of tools understanding the current LICR and it's unlikely

it is true that a) Omega does offer more general support for manipulating
the data b) external tools that directly understand the LICR will be few.

> there will be more (they are eventually expanded and therefore
> cannot be processed anyway, the very fact that unicode chars
> are actual letter' chars is critical).

that is not our understanding however. transformation of LICR is supposed to
happen only when leaving its domain eg for the final typesetting step. Having
an LICR that is unicode chars clearly makes it simpler for an external tool to
manipulate data and send it back to the system; but there is nothing generally
restricting about the LICR (in current LaTeX) being not just unicode
characters. All that is needed is to provide the external tool with a
translation to understand the data.

> So, having true
> Unicode text (perhaps with tags, which can be removed if
> necessary) at some part of the internal processing is imo
> an essential feature in future extensions to TeX.

agreed

> And indeed
> Omega is an extension which can cope with that; I wouldn't like
> renounce that.

we think not (yet) at least not the code that is currently available (to us
--- from the texlive 5d CD)

you wrote:

> Omega can represent internally non ascii chars and hence
> actual chars are used instead of macros (with a few exceptions).
> Trivial as it can seem, this difference is in fact a HUGE
> difference. For example, the path followed by é will be:
>
>  é --an encoding ocp-|           |-- T1 font ocp-->  ^^e9
>                      +-> U+00E9 -+
>  \'e -fontenc (!)----|           |- OT1 font ocp -> \OT1\'{e}

what you are describing there is, in our understanding, effectively a
replacement for trans B in our above diagram, ie our understanding of what is
possible in Omega currently is roughly looking like this:

-------                   --trans C
|      |                  |       |
|      |                  |       |
(8 bit number)           V      | (produce hlist)  V       |
é    --- trans A --->   OICR1 +--- trans B ------> OICR2 +--> trans E
^      |
|      |
|      |
trans D (eg generating TOC)

trans A = tokenising 8 bit numbers as the corresponding 16bit numbers
Example:
if é was in the cp437 code page (German DOS) it would
be the 8bit char "82; that would become the 16bit token with
number "0082 (which is NOT é in unicode = "00E9)

if on the other hand  é was in latin1 (where it is "E9) we
get "00E9

if the input was in utf8 you would not get unicode chars as
the result but sequences of 16bit chars all starting with "00
--- so unicode charcater that is multibyte in utf8 would not
become the corect unicode 16-bit token but would become a sequnce
of tokens each of the form "00

OICR1   = Omega Internal Character Representation 1:
16bit representation of characters for which without additional
external information one can't tell which character is refered to.

trans B = process when Omega is producing hlists, ie only when it forms
paragraphs or hboxes. Only after that point, or rather while doing
that, can ocps be used to transform the OICR1 further. To turn it
into OICR2 one would at this stage apply what you called "an
encoding ocp" and/or transform commands from the LICR, eg \'e to
OICR2 ie a unicode char. But to be able to transform OICR1 into
OICR2 you need to have the original encoding information still
present

OICR2   = Omega Internal Character Representation 2:
16bit representation of characters as unicode positions (or so we
hope if the transformation from OICR1 to OICR2 worked)

trans C = As an example, \MakeUppercase now works on the OICR2 representation
as an ocp; it is an interesting question whether it should happen
this late in the process (on typeset stuff) (By the way, the
primitive \uppercase would as we think work on the tokens before
producing the hlist, ie on OICR1 tokens)

trans D = writing to .aux files and reading them back in; puttings things
into marks and manipulating them, etc. All these transformations
work on OICR1, see below.

Discussion:
===========

The problem really is transforms of type D which are using OICR1 and are thus
likely to break in the sense that their encoding information is lost in the
process.

So we think that the translation process from external source data in some
encoding to the OICR should happen not at trans B via encoding ocps but at
trans A so that OICR1 = OICR2.

Note that this translation process from external encoding to OICR would work
on streams and not on finite (token) lists so it should have slightly
different characteristics compared with ocps.

We are sure it will be difficult to provide control of such a translation
process at trans A if the control should be from within the source document
and usable by authors and or packages, eg changing the input encoding midway
in an argument could have similar restrictions to say, \catcode changes in
such places. E.g, you couldn't do

\def\french#1{{\inputencoding{latin1}#1}}

because then for

bla bla \french{foo} bla bla

the input encoding change would not be noticed until after "foo" has already
been tokenised (incorrectly) --- yes, we know that this example could be made
to work using Don's \footnote trick but as with LaTeX's \verb there will be
situations in which even more elaborate implementations will still fail due to
tokenisation happening before any macro expansion is possible.

Another problem of the current model seems to be that, even if trans A did the
encoding transformation to Unicode ie we have only a single OICR,
transformations of type D (ie transformation of character token strings) can't
be controlled by a mechanism similar to the one that is available for
transformations of type C, ie in one case we have ocps and in the other area,
when we work on structural issues like building TOC or arranging data for page
representation no such mechanism is available. Thus is seems interesting to
think about whether or not a similar concept (not necessarily the same!)
should be made available for this part of the process.

In other words the concept of ocps makes perfect sense for character string
manipulation but one has to [pretend] to typeset something to have them
available in current Omega, but a large amount of document processing is
concerned with character string manipulation not related to typesetting at
all. As a small example, when displaying an error message and error message
(transform of type D) one should transform character data from OICR back to
the encoding used by the (OS interface to the) display device.

I hope this explains a bit more about our understanding of the LICR and how we
think it could be generalised for a system that internally uses Unicode
characters and string transformation processes.

cheers
Frank (with Chris editing and criticising:-)
`