Content-Type: text/plain; format=flowed; delsp=yes; charset=iso-8859-1
MIME-Version: 1.0
References: <44B57B36.2030603@digi21.net>
Content-Transfer-Encoding: 8bit
User-Agent: Opera Mail/9.00 (MacPPC)
Message-ID:  <op.tcut7fuass64je@morten.local>
Date:         Tue, 18 Jul 2006 00:31:53 +0200
Reply-To: Mailing list for the LaTeX3 project
              <LATEX-L@LISTSERV.UNI-HEIDELBERG.DE>
Sender: Mailing list for the LaTeX3 project <LATEX-L@LISTSERV.UNI-HEIDELBERG.DE>
From: =?iso-8859-1?Q?Morten_H=F8gholm?=
              <morten.hoegholm@LATEX-PROJECT.ORG>
Subject: Re: Ifs
To: LATEX-L@LISTSERV.UNI-HEIDELBERG.DE
In-Reply-To:  <44B57B36.2030603@digi21.net>
Precedence: list
Status: R

On Thu, 13 Jul 2006 00:44:06 +0200, Javier Múgica de Rivera  
<javier@DIGI21.NET> wrote:

Hi Javier,

> There is a control secuence named \chk_if_exist_cs:N with the following  
> definition:
>
> \def:Npn \chk_if_exist_cs:N #1 {
>   \if:w \cs_if_exist_p:N #1
>   \else:
>     \err_latex_bug:x{Command~ `\token_to_string:N #1'~
>                      not~ yet~ defined!}
>   \fi:}
>
> i.e., if the control sequence already exists it does nothing, but \else  
> an error is issued. The opposite to \chk_if_exist_cs:N is, amazingly,  
> \chk_new_cs:N,
>
> \def:Npn \chk_new_cs:N #1{
>   \if_meaning:NN #1\c_undefined
>   \else:
>     \if_meaning:NN #1\scan_stop:
>     \else:
>       \err_latex_bug:x {Command~name~`\token_to_string:N #1'~
>                             already~defined!~
>                         Current~meaning:~\token_to_meaning:N #1
>                       }
>     \fi:
>   \fi:
> %<*trace>
> some code
> %</trace>
> }
>
> \chk_if_exist_cs:N does the check via \cs_if_exist_p:N, which reverts  
> \cs_if_free_p:N, which in turn checks that the cs is undefined and not  
> literally \c_undefined. But \chk_new_cs:N does not!, and it is the one  
> actually used to test for free commands when new-defining something, so  
> that something like \def_new:Npn\c_undefined{whatever} would work, and  
> it shouldn't.

It certainly shouldn't! Thanks for pointing this out. I have changed the  
function \chk_new_cs:N to follow the logic of only allowing to define a  
control sequence if it is either undefined, \relax, and also textually  
different from \c_undefined and \scan_stop:.

> It also happens that all the tests purposed to see if a cs is free, they  
> test for equality with both \c_undefined and \scan_stop: (\relax). The  
> logical equivalence of a cs equal to \scan_stop: to a undefined cs is  
> due to the fact that \expandafter\ifx\csname foo\endcsname creates the  
> control sequence, and lets it equal to \relax, as we all know, and also
> it was common to ``undefine'' a cs by \let\foo\relax. Now we can  
> undefine via \let\foo\c_undefined (the control sequence \c_undefined is  
> the most tricky trick I've ever seen. It would have been nice if we had  
> known it from the begining) and test for the inexistence of a cs with  
> etex \ifdefined and \ifcsname without creating it, so there is no reason  
> why to think about someting equal to \scan_stop: as a free cs.

Unfortunately it is not quite that simple. How can \def_new:Npn or any  
other function for that matter know how the control sequence to be defined  
has been passed on to it? Answer is it can't. It would be very impractical  
if
   \def_new:Npn \foo{}
and
   \def_new:cpn {foo}
were not identical.

There is only one sure-fire way to make to prevent a \cs:w ... \cs_end:  
operation from turning the control sequence into \scan_stop: and that is  
by introducing grouping which can then restore the meaning to undefined  
after ending the group. For example you could do

   \def:Npn \exp_args:Nc #1#2{
     {\exp_after:NN } \exp_after:NN#1  \cs:w #2\cs_end:
   }

This of course renders the whole operation non-expandable and so we are  
faced with choosing the lesser of two evils. This is also why IMO the  
\ifdefined primitive is of limited use because the token may very well  
have been created by \cs:w ... \cs_end: at an earlier step in the  
processing so you always have to do an additional check for \relax. With  
the expansion mechanism used in the expl3 code base this is particularly  
important.

On the other hand the \ifcsname primitive can be used in situations where  
you know the control sequence in question is guaranteed to not have been  
defined by accident elsewhere. An example from l3xref, where the property  
list storing all information for a label is never entered directly as a  
single token so not reason to create an extra entry in the hash table:

\def_new:Npn \xref_get_value:nn #1#2 {
   \cs_if_really_free:cTF{g_xref_#2_plist}
   {??}
   {
%    \end{macrocode}
% This next expansion may look a little weird but it isn't if you
% think about it!
%    \begin{macrocode}
     \exp_args:Ncc \exp_after:NN
     {xref_get_value_#1_aux:w}{g_xref_#2_plist}
     \q_nil
   }
}

> There exist \if_really_free... that do these tests. In my opinion these  
> should be the normal \cs_if_free..., and nothing with the extrange names  
> \if_really_free... should exist. So my point of view is:
> 1. All tests that finally use tex \ifx for free-cs testing should be  
> replaced by \if_cs_exist:N and \if_cs_exist:w (\ifdefined and \ifcsname)
> 2. A control sequence equal to \relax being equivalent to a free control  
> sequence is an accident of the past we should commpletely forget about.

I disagree on both points for the reasons given above. I'm well aware that  
the "really_free" name is bad but I couldn't think of anything better at  
the time... I'm open to suggestions! :-)

> p.s. I also thik that names for testing if a certain condition is true  
> or false should always contain ``if'' to avoid confusion.

I think I wrote a similar comment in the sources somewhere a while ago.  
Anything that tests a certain condition should probably have a name  
containing "if".
-- 
Morten