In the original TeX program, the `tally` variable is increased whenever
`print_char` is called (except when character `s` is newline), regardless of
`selector`.
However, in `tprint` function of LuaTeX, `tally` is only called when `selector`
is `pseudo`.
This leads to incorrect indentation of the second line, for example the
following TeX file
```tex
\errorcontextlines=10000
\def\c{1 \errorerror 2}
\def\b #1 {7 #1 8}
\def\a{5 \b {3 \c 4} 6}
\hbox{7 \a 8}
\stop
```
would lead to the following error (observe that the `<argument>` line is wrong):
```
\c ->1 \errorerror
2
<argument> 3 \c
4
\b #1 ->7 #1
8
\a ->5 \b {3 \c 4}
6
l.5 \hbox{7 \a
8}
```
I have two ideas how to fix it.
One is to add `strlen(sss)` to `tally` regardless.
The other is the following. This would still be incorrect if `selector` is
`no_print` or `term_only`, however I think these cases will not happen.
```diff
diff --git a/source/texk/web2c/luatexdir/tex/printing.c
b/source/texk/web2c/luatexdir/tex/printing.c
index 8a52700..b628280 100644
--- a/source/texk/web2c/luatexdir/tex/printing.c
+++ b/source/texk/web2c/luatexdir/tex/printing.c
@@ -498,6 +498,7 @@ void tprint(const char *sss)
t_flush_buffer(log_file,file_offset);
}
}
+ tally++;
}
if (*buffer) {
buffer[i++] = '\0';
```
I'm not sure which option is faster (on one hand the `strlen()` option need a
separate loop, on the other hand it's a built-in library function so it should
be fast?)
Regardless, if performance doesn't matter that much, I think `strlen()` option
is better.
------
On an academical note, there is still a small difference between LuaTeX and
PDFTeX with the following patch. With the following TeX code
```tex
\errorcontextlines=10000
\newlinechar=`u
\def\c{1 \errorerror 2}
\def\b #1 {7 #1 8}
\def\a{5 \b {3 \c 4} 6}
\hbox{7 \a 8}
\stop
```
then the output of LuaTeX would contains
```
<arg
ment> 3 \c
4
```
with 16 spaces before `4`, while PDFTeX only has 15 spaces before `4`.
In practice, chances are nobody cares. For performance reason it's fine to
ignore this case.
