This is something I've come across while working on #1734. Compiling the following

\documentclass{article}
\begin{filecontents*}[overwrite]{\jobname.latexml}
DefMacro('\dumpArg {}', sub {
  my ($gullet, $arg) = @_;
  print STDERR "args:\n";
  foreach ( @{$arg} ) {
    print STDERR $_->getCatcode . ': ' . $_->toString . "\n"; }
  return $_[1]; });
1;
\end{filecontents*}
\providecommand{\dumpArg}[1]{#1}
\begin{document}
previous line
\dumpArg{1}
%previous line
\dumpArg{2}
done
\end{document}

with LaTeX yields the pdf "previous line 1 2 done" (and creates the latexml file). Compiling with latexml --nocomments creates the xml

    <p>previous line
1
2
done</p>

and stderr

args:
12: 1
args:
12: 2

Compiling with comments, however, creates the xml

    <p><!--  %previous line -->previous line
1
2
done</p>

so that the comment has moved to an unexpected place. Even worse, stderr shows

args:
12: 1
args:
14: %previous line
12: 2

so that the "previous line" comment became the first argument to the subsequent macro.

I'm not sure why this is happening. I don't think it generally affects things, since it appears you have to be accessing the arguments in a non-standard way. But siunitx accesses the arguments in a non-standard way, so I wanted to ask if this is expected and we need to modify siunitx. (And I assume --nocomments is why we've not noticed this before, since that removes the behavior. Otherwise, t/complex/si.tex doesn't compile.)