Re: Fix makedepend depmode to cope with VPATH builds

[Stefano Lattarini]

On Saturday 09 April 2011, Ralf Wildenhues wrote:
> Hi Stefano,
> 
> * Stefano Lattarini wrote on Sat, Apr 09, 2011 at 11:38:47AM CEST:
> > On Friday 08 April 2011, Ralf Wildenhues wrote:
> > > * Stefano Lattarini wrote on Fri, Apr 08, 2011 at 01:05:53PM CEST:
> > > > On Wednesday 06 April 2011, Ralf Wildenhues wrote:
> 
> > > > > +  # makedepend may prepend the VPATH from the source file name to 
> > > > > the object.
> > > > > +  sed_object_re=`echo $object | sed 's/[].[^$\\*|]/\\\\&/g'`
> > > > >
> > > > A couple of questions/observations about the sed command:
> > > >  1. I wasn't sure whether `.', `$', `^' and `*' are to be escaped when 
> > > > used in
> > > >     a bracket expression; however, the POSIX standard at:
> > > >       <http://pubs.opengroup.org/onlinepubs/007908799/xbd/re.html>
> > > >     says they shouldn't be.  Also, I couldn't find any reference in the
> > > >     Autoconf manual pointing to a sed implementation that doesn't 
> > > > respect
> > > >     this part of the standard.
> > > 
> > > Ah, cool.  Then I recon we don't need this escaping step at all.
> > >
> > Why not?  The sed command below:
> >   sed "s|^.*\($sed_object_re *:\)|\1|" "$tmpdepfile" > "$depfile"
> > doesn't use any bracket, does it?
> 
> I'm not sure I understand the question.  Can you rephrase it?
>
> > but I don't see how it's related to my observation above.
> > What am I missing?
> 
> I guess then I don't understand your observation then.  Your observation
> tells me that object file names of the form foo^bar.o and foo$baz.o are
> not problematic.
>
No.  It just tells that sed regular expressions like [ab.^$0-9] are not
problematic, because `.' and `$' are *not* recognized as metacharacters
in *bracket expression*, and `^' is recognized as a metacharacter in a
bracket expression only when used as *first* character there.  Which
simply means that your original command:
  sed_object_re=`echo $object | sed 's/[].[^$\\*|]/\\\\&/g'`
was fully correct, and didn't lack any required escaping (while, on the
first impression, I had *erronously* thought it did).

But re-reading what I wrote in my first reply, I must admit it sounded
like I was adding something new to the discussion, while in fact I was
just saying "hey, at first I wasn't sure that this change of yours was
correct, but by reading the standards/manuals more carefully, I realized
you are indeed completely right".  Miscommunication on my part -- sorry
for the noise and the time I made us loose.

> > > The only character I was worried about was $, as it can occur in
> > > object names now and then (ok, mostly java, and that doesn't use
> > > depcomp).  Any others in the list are pretty much excluded anyway,
> > > because they will evoke shell expansion errors when used in a
> > > makefile unquoted.  Except for '.', but I'm not afraid of matching
> > > a bit too broadly here, the final 'o' will take care that we don't
> > > match any header files.
> > >
> > OK, this makes sense (even if I'd err on the safe side w.r.t. the `.'
> > character),
> 
> OK, can you come up with a practical example (i.e., a valid object file
> name) where we would match something wrongly here?
>
Pratical examples?  No, I'm just being a little paranoid (as usual).
Well, too much paranoid this time: see below.

> Remember the input
> file only contains lines of the form
>   $object: header header...
>   $object: header \
>    header header...
> While somebody might arguably name their header foo.o, I do think they
> deserve what they get.
>
Agreed.  Especially because, to cause a spurious match, they would have to
*also* have a `:' character in the header name, i.e., have an header named
like "foo.o:bar.h" or "foo.obj:bar.h" -- in which case they probably do
deserve to suffer a bit.

> I observe that no one in their right mind names their
> object foo[bar.o or foo]bar.o or foo\bar.o (see the various instspc*
> tests for why they would not work in practice anyway).  So I conclude
> that it is not needed to do regex-escaping for the object file name.
>
OK, fully agreed now (well, once the possible problems with `$' and `^'
have been cleared up).

> > > @@ -503,7 +503,8 @@ makedepend)
> > >    touch "$tmpdepfile"
> > >    ${MAKEDEPEND-makedepend} -o"$obj_suffix" -f"$tmpdepfile" "$@"
> > >    rm -f "$depfile"
> > > -  cat < "$tmpdepfile" > "$depfile"
> > > +  # makedepend may prepend the VPATH from the source file name to the 
> > > object.
> > >
> > I'd add a comment here telling why there's no need to escape $object for the
> > sed commmand just below.
> 
> I will, after I've understood your remaining comments.
>
Hope this reply is clearer than the previous ones.

Thanks,
  Stefano