Why don't we always use \long\def instead of \def?

Question

Why don't we always use \long\def instead of \def? It seems that \long\def is more complete than \def.
By the way, I know the difference between these two commands.

Answer 1

We do. \newcommand is \long by default; if you want un-long macros, you have to use \newcommand*.

---

A more philosophical answer: \long is one of the several error-catching mechanisms that Knuth built into TeX. Or rather, not using \long allows TeX to catch missing-brace errors where the argument of a macro would appear to span multiple paragraphs; that's where the "runaway argument" errors come from. So using \long by default means that you may use up TeX's memory before finding an error.

This is similar to my answer to this question about \outer, which has a similar purpose in buck-stopping out-of-control code. My conclusion there was that \outer reflects an obsolete prioritization of computer resources versus human resources, and I believe that I will make the same judgment of \long here. It appears that in both cases, the authors of LaTeX take this position as well.

Answer 2

Let me put my LaTeX3 hat on before Joseph does. [But Joseph, feel free to edit.]

The current plan for LaTeX3 is to make almost all macros long. The first aspect of that is at the level of defining macros: by default, macros are \long, and you need to type the extra _nopar to make them non-long "no-\par". For instance,

\cs_set:Npn                  = \long \def
\cs_set_nopar:Npn            =       \def
\cs_gset_protected:Npx       = \long \protected \xdef
\cs_gset_protected_nopar:Npx =       \protected \xdef

The second aspect is the convention that we have recently (i.e., perhaps not yet on CTAN?) enforced in the kernel about which macros are long and which are not. Clearly, you never want to do \newcount\par (in LaTeX3 coding conventions, \int_new:N\par), so until recently we had defined \int_new:N to be short. However, the error that you get in that case is not very enlightening anyways, and other cases are not so clear cut. Thus, we decided that every macro that takes parameters would be \long, and parameterless macros should be _nopar.

The third aspect is the user interface, currently covered by xparse. Only there does it make sense to try and catch missing braces and other "runaway" arguments. At this level, we keep \long and non-\long arguments, adding + before \long arguments in the specification: for instance,

\NewDocumentCommand{\foo}{ m +m }
  { \section{#1} {\bfseries #2} }

creates a macro \foo with a short first argument and a long second argument.