This mail discusses details of the implementation of the xdoc2l3
\NewDocumentCommand command, and issues with these.
First a bit of context: In my old code, all argument grabbers were
being expanded in the context
<grabber> <subsequent grabbers> \XD@endgrab <next argument> ...
A grabber would do whatever it had to do to grab the <next argument>,
and was then supposed to append it to \toks@; thereafter the
\XD@endgrab macro expanded to \the\[log in to unmask] Xparse does roughly the same
thing, but has an explicit \toks_use:N\l_xparse_grabbed_args_toks
instead of the \XD@endgrab macro and uses \l_xparse_grabbed_args_toks
as the token marking the beginning of arguments.
Implementing processors in this context was actually very easy, as all
they had to do was to grab the <next argument>, do something to it, and
then put the result back as a mandatory argument after \XD@endgrab,
where the next processor would expect it. I'm a bit worried that the
xparse construction, exposing the implementation detail of
\l_xparse_grabbed_args_toks, in the long run will be restrictive;
processors don't need to know the name of the token register (or even
that there is one), but they do need to know the name of the mark
token.
*Example.* The x{<bool>}{<pre>}{<post>} specifier compiles into the
processor \XD@gp@expanded{<bool>}{<pre>}{<post>}, which is simply
defined as
\def\XD@gp@expanded#1#2#3#4\XD@endgrab#5{%
\protected@edef\@tempa{{%
\IfBooleanTF{#1}{#2{#5}#3}{#2#5#3}%
}}%
\expandafter\XD@gh@put \expandafter{\@tempa}{#4}%
}
\def\XD@gh@put#1#2{#2\XD@endgrab#1} % Generic helper macro
---shovelling things around a bit, but quite straightforward. The
(possibly less roundabout) approach of not throwing a processed
argument back to the other side of the \XD@endgrab marker has the
disadvantage that it forces a distinction between different types of
processors and thus increases the complexity of the system.
Two things in xparse which I didn't implement are the P and W
specifiers. In the case of W this is mostly because the LaTeX2e
\@ifnextchar gobbles spaces, and I didn't feel like rolling my own; I
have however encountered cases where I would have wanted an "o"
specifier that didn't gobble spaces, so I certainly see the need for
it. In the case of P, I was simply lazy; I don't think xdoc will need
\long arguments anyway, and they increase the complexity quite
considerably.
However, while thinking about the matter, it occurred to me that the
status for some grabber/processor of being non-\long implies that one
cannot use \par in the <default> for an O argument, even if that
argument itself is \long; the reason being that such a \par token then
appears within the <subsequent grabbers>. This issue should at least be
remarked upon in xparse.dtx.
One *new* thing I did in this implementation is that I used groups to
isolate processing of different arguments and also the argument
grabbing procedure as a whole. Concretely the latter was just the
matter of changing the defitions
\def\XD@grab@arguments#1#2{%
\toks@={#1}%
#2\XD@endgrab
}
\def\XD@endgrab{\the\toks@}
(as found in current xdoc2) to
\def\XD@grab@arguments#1#2{%
\bgroup
\toks@={#1}%
#2\XD@endgrab
}
\def\XD@endgrab{\expandafter\egroup \the\toks@}
As for the per-argument groups, if you look at xdoc2l3test.dvi you'll
see that each argument grabber begins with an explicit \bgroup, whereas
the matching \egroup is hidden in the ending \XD@ge@parameter or
\XD@[log in to unmask] At least that the g{<assignment>} processors should
affect tokenisation and be possible to combine with optional-argument
type processors requires that things are done this way. There are
however at least two problems with doing it in general:
1. If there is an error when TeX is grabbing arguments, then the
macro simply disappears, along with all text grabbed so far. Such
errors arise easily as typos, and would in this case cause the
closing \egroups to disappear, throwing off the nesting of groups.
Seeking to limit the damage caused by this, I chose to use \bgroup and
\egroup so that at least the nest end of an environment would get
things back on track, but that causes a new problem.
2. In math mode, \bgroup\egroup is far from invisible.
For xdoc this is acceptable, since the commands defined are anyway
mostly for use in vertical or horizontal mode, but xparse should be
able to define general commands and thus cannot accept this.
\begin{TeX-extension-idea}
It occurs to me that it would be useful to have groups that nest even
weaker than \bgroup-\egroup, and which do not contribute anything to
math lists.
\end{TeX-extension-idea}
The existing alternative is to use \begingroup and \endgroup, but then
one must guard against getting these mismatched. I have one idea for
this, but I haven't tried it out yet:
\def\XD@gh@insure#1{%
\let\reserved@a=#1%
\expandafter\futurelet \expandafter\@let@token
\expandafter\XD@gh@insure@
}
\def\XD@gh@insure@{%
\ifx \reserved@a\@let@token \else \egroup\egroup \fi
}
The idea is to instead of a naked macro (like \XD@gp@expanded) that
processes material after \XD@endgrab write
\XD@gh@insure{\protected@edef}\XD@gp@expanded
where \protected@edef is the first token in the expected expansion of
\XD@[log in to unmask] If that is not what lies ahead, then apparently
expansion of \XD@gp@expanded failed, and it is necessary to insert some
emergency \egroups to get the nesting back to normal.
3. Adding groups makes the command non-mouth-expandable.
In particular, it will throw TeX out of the inner loop
(disrupt kerning etc.)
Again, this is something I can live with in xdoc, but if the LaTeX3
\NewDocumentCommand is going to replace \newcommand, then there will
probably be cases where one wants to define a command that is really
just a macro. This is probably doable, if \NewDocumentCommand only
introduces the groups if they're really needed -- using a simpler
implementation for the nice cases, similar to that for the tail of m
arguments in current xparse -- but could lead to a rather complex
codebase. Still, making the distinction be between "commands with only
`m' arguments" and "all other commands" wouldn't need to make the code
any more complicated than it is today. The only disruption that one
would then get by starting and ending groups as part of argument
grabbing is between an \accent and its base character, but perhaps that
is acceptable.
However, I couldn't help dreaming a bit... If one was allowed to add
one new TeX primitive, for the purpose of solving the issues above,
then what should that be? Much of our problems stem from the
restriction that TeX does not allow assignments while expanding things
(even though the mouth is Turing-complete), so my choice would be to
offer an escape from this restriction.
\begin{TeX-extension-idea}
Add an \emph{expandable} primitive that allows for some commands
(primarily assignments) to be \emph{executed} and affect the
expansion.
The syntax could be
\expandedwith{<text>}{<commands>}
and the semantics could be as follows:
\begin{enumerate}
\item
A new group is begun.
\item
The <commands> are executed.
\item
The <text> is expanded (as discussed regarding
new \expanded primitive). The result of this will
be the one-step expansion of the entire \expandedwith.
\item
The new group is closed.
\end{enumerate}
With this, a non-spacebypassing \@ifnextchar could be conservatively
implemented as
\begin{verbatim}
\long\def\@ifnextchar#1#2#3{%
\expandedwith{%
\ifx\@let@token\reserved@d
\noexpand\@firstoftwo
\else
\noexpand\@secondoftwo
\fi
\the\toks@
}{%
\let\reserved@d=#1%
\toks@={{#2}{#3}}%
\futurelet\@let@token
}%
}
\end{verbatim}
(it would look as if the \@ifnextchar<char> is replaced by
\@firstoftwo or \@secondoftwo in one expansion step).
More relevant here is however that \XD@grab@arguments could be
implemented as
\begin{verbatim}
\def\XD@grab@arguments#1#2{%
\expandedwith{\the\toks@}{\iffalse}\fi
\toks@={#1}%
#2\XD@endgrab
}
\def\XD@endgrab{\iffalse{\fi}}
\end{verbatim}
---certainly a braintwister, but everything in this area tends
to be. \emoticon{:-)}
Implementation-wise, I would try making a procedure which is
like a simplified main_control and have \expandedwith work
primarily by calling this to execute the <commands>. Since most
of the complexity of main_control seems to lie in the "inner loop"
spaghetti, this might actually be straightforward. In terms of
the traditional TeX anatomy, this would (sensitive readers beware)
amount to fitting TeX's mouth with an auxiliary stomach that's
severely ostomized; anything solid (such as ordinary text characters)
mistakenly fed down it should throw errors, but invisible stuff not
involving typesetting should be OK.
\end{TeX-extension-idea}
Such dreaming aside, I think the Big Issue is whether argument grabbers
should operate in a group or not.
Lars Hellström
|