[NTG-context] The odd semantics of \begincsname

Hans Hagen j.hagen at xs4all.nl
Sat Aug 17 12:15:49 CEST 2019

On 8/17/2019 10:56 AM, Henri Menke wrote:
> On 17/08/19 8:48 PM, Hans Hagen wrote:
>> On 8/17/2019 9:19 AM, Henri Menke wrote:
>>> Dear list,
>>>
>>> According to the LuaTeX documentation:
>>>
>>>       “The \begincsname primitive is like \csname but doesn’t create a
>>>       relaxed equivalent when there is no such name.”
>>>
>>> I thought it would be possible to use this fact to skip the \relax-ed
>>> definition when \def-ining a new control sequence, but the following MWE
>>> fails with \inaccessible:
>>>
>>>       \expandafter\gdef\csname yes\endcsname{}
>>>       \expandafter\gdef\begincsname no\endcsname{}
>>>       \bye
>>>
>>> Is this a bug or is this behaviour intended?  Could this be fixed by
>>> making manufacture_csname aware whether it is in a def_cmd context or
>>> not?
>> [sorry to those who are not interested in these low level issues, just skip]
>>
>> intended ... it expands to basically nothing so you get no token
>> representing a 'name' after the gdef .. the expansion is pushed in from
>> of whatever comes next (which could be another \expandafter for instance)
>>
>> you suggest that if \begincsname could behave differently when it's
>> after a \def, \gdef, (and then quite some more definition related
>> commands), it could behave differently but it not an option
>>
>> for instance (as mentioned) there can be more than one expansion going
>> on after these define commands, like expanding a macro that itself
>> expands to \csname so one has several \expandafters before the gdef
>> then); there is actually no looking back in scanning tokens unless a
>> token has been scanned already and looking forward would involve
>> expansion so a circular mess
>>
>> an option could be not to push something on the save stack (a side
>> effect of creating the csname, which has a little impact on performance
>> and nesting) but removing that bit might give other side effects (e.g.
>> for successive reassignments inside a group, maybe even mixed local and
>> global); i did a quick test with that and it gives quite incompatible
>> output in ConTeXt so that's definitely a no-go (adding all kind fo
>> saveguards and checks in the engine doesn't pay off, especially not for
>> something that never was a problem)
>>
>> some time ago i considered a convenience command \[e]defcsname, as it
>> saves a few tokens (no gain in performance as all the related things
>> still need to happen); but even that one would probably create the name
>> in the same way
>>
>> so ... this is the way it is ... (i must admit that it never gave me any
>> issues so whatever triggered the question, there's probbaly a way around
>> it)
>
> I can accept this answer.  Just for a little context, the question was
> triggered by this:
>
>      https://tex.stackexchange.com/questions/504501/global-variant-of-csname-endcsname
>
> In short: Having thousands of
>
>      \expandafter\gdef\csname foo\endcsname{}
>
> inside a group (as happens for xmltex), can lead to a save_stack
> overflow.  One way around it is to do
>
>      \begingroup\expandafter\endgroup\expandafter\gdef\csname foo\endcsname{}

Sure, just group. But actually, if one needs that many csnames one can
wonder about the approach. One can bump the save stack just like one
also might have to bump the hash (extra) size (either of them can
overflow).

Also, probably a bit of extra grouping can happen at a different level,
not for each csname but for in this case an xml element, which is also
more efficient

> The \expandafter inside the group will pull the evaluation of \csname
> into the group which will discard the save_stack at the \endgroup, thus
> avoiding the build-up.  However, this construction is a bit hard to
> understand so I was wondering whether

well, instead of this:

\begingroup\expandafter\endgroup\expandafter\gdef\csname foo\endcsname{}

one can just use this:

\begingroup\expandafter\gdef\csname foo\endcsname{}\endgroup

which is less tokens, less pushing/poping and therefore a litle faster
(but often neglectable compared to other things that tex/macros do in
most cases) but of course it looks less 'cool' and 'expert' and creates
less 'awe' .. so let's add another one:

{\expandafter}\expandafter\gdef\csname foo\endcsname{}

this one is performance wise close to the second case (normal grouping)
but it might look more puzzling which is why i should wrap it:

\def\defcsname {{\expandafter}\expandafter\def \csname}
\def\gdefcsname{{\expandafter}\expandafter\gdef\csname}

which then is about as efficient as the first alternative with two
\expandafter usage using \begingroup\endgroup (okay, efficiency depends
of course on the engine too, and probably on the cpu as well)

(you can argue that \expandafter and \noexpand and \futurelet ... were
added to tex so that one could boost his resume ... the more you use in
sequence the more expert you are; but you can also argue that they add
some charm to tex, a nice playground and such)

>      \expandafter\gdef\begincsname foo\endcsname{}
>
> could be used instead to elide the save_stack (which doesn't work
> because \begincsname does not actually build a \csname).
it does when it's known and then it puts something in the input (a
token), but when unknown it doesn't so you effectively get \def{} which
is not what you want (ok, maybe some weird usage where { is defined as
macro does, which actually can make sense when one handles xml with
active characters).

Hans

-----------------------------------------------------------------
Hans Hagen | PRAGMA ADE
Ridderstraat 27 | 8061 GH Hasselt | The Netherlands
tel: 038 477 53 69 | www.pragma-ade.nl | www.pragma-pod.nl
-----------------------------------------------------------------