Re: Breakage in out-of-tree builds

[Hans-Bernhard Bröker]

Am 16.04.2017 um 12:24 schrieb Gavin Smith:
> On Sat, Apr 15, 2017 at 11:12:05PM +0200, Hans-Bernhard Bröker wrote:
> > the recent round of updates to the gnulib stuff triggered what must be
> > a rather old bug in the tp/maintain/regenerate_docstr.pl script.
>
> I suspect it is the update to the translation files, rather than the
> gnulib update.

You're probably right.

> > completely fails to work in out-of-tree, a.k.a. VPATH builds.  The problem
> > is that this perl script writes makefile rules into tp/Makefile.docstr which
> > look like this:
> >
> > $(srcdir)/../po_document/'"$lingua.gmo"':
> > $(srcdir)/../po_document/'"$lingua.po"'
> >         cd $(srcdir)/../po_document/ && $(MAKE) $(AM_MAKEFLAGS) '"$lingua.gmo"'
> >
> > But in a VPATH build, there is no makefile in $(srcdir)/../po_document. The
> > Makefile to be used is in $(builddir)/../po_document instead, or the shorter
> > version: ../po_document.  So the above rule really should become
> >
> > $(srcdir)/../po_document/'"$lingua.gmo"':
> > $(srcdir)/../po_document/'"$lingua.po"'
> >         cd ../po_document/ && $(MAKE) $(AM_MAKEFLAGS) '"$lingua.gmo"'
>
> I think that the *.gmo files should be built in the srcdir, because that
> is where they are in a distribution archive.

And they will be: both target and dependency of that rule are explicitly 
in $(srcdir), and I was not proposing to change that.  But the 
_Makefile_ that actually does the job is in $(builddir).

> This is the old problem with generated files being or not being tracked
> in version control.

Not really, IMHO.

> The *.gmo files are not tracked at the moment. If
> they had been tracked, they would not have been out-of-date in your SVN
> checkout.

The problem is not that I had to build them.  The problem is that the 
build didn't _work_, because the Makefile rule is broken.

> I can think of two possible fixes:
> * Track the generated *.gmo files under po_document in SVN.

Please, no.

> * Remove the makefile rules in tp/Makefile.docstr (via the script you
> mentioned) in order not to regenerate the *.gmo files.  They would have
> to be regenerated at some other time.

That would risk breaking the option to run 'make' directly in 
$(top_builddir)/tp.  The rules that copy stuff from po_documents to 
tp/LocaleData need those files for input, so it's the responsibility of 
that Makefile to procure them.

> > Or maybe it should just run a single sub-make in po_document from tp's
> > $(builddir) and be done with that, instead of one sub-make per out-of-date
> > .po file target.
>
> "make update-gmo" seems to be the commmand to do this.

But without digging even further into GNU-make-only feature space, I 
don't see an easy way to make each invocation of make in 'tp' run "make 
update-gmo" exactly once, and doing that _before_ any of the .gmo files 
are actually needed.