Re: [PATCH] {master} missing: do not touch timestamps; only warn for out-of-date files

[Dave Hart]

On Wed, Jun 20, 2012 at 9:30 PM, Stefano Lattarini
<address@hidden> wrote:
> Before this change, the missing script had a twofold role:
>
>  - it warned the user if some required maintainer tools was missing,
>    or too old;
>
>  - in such a case, it tried to "fix" the timestamp of the files that
>    should have been rebuilt by that tool (without actually updating
>    the file contents, of course), to allow the build to continue.
>
> The second capability used to be quite useful in the days when most
> projects committed files generated by maintainer-only tools in their
> VCS repository (today the trend is not to keep such generated files
> VCS-committed anymore).  In such a setup, the "timestamp-fixing"
> capability of 'missing' was quite useful, in that it allowed users
> lacking some required maintainer tool to build from a VCS checkout
> in the face of skewed timestamps (as could have been caused by
> "cvs update" or "git checkout").
>
> But then, when the automatic remake rules kicked in due to the
> generated files being *actually out-of-date* (e.g., because the user
> had modified 'configure.ac' but lacked a modern-enough autoconf to
> rebuild it), that behaviour of 'missing' caused the same problem that
> plagued AM_MAINTAINER_MODE; i.e., the user would get non-dependable
> builds and inconsistent statuses of the build tree -- changes to
> source files don't reflect on generated files, and this can be very
> confusing and cause hard-to-spot errors).
>
> So we now believe that the best approach to deal with timestamp-related
> issues is not to have 'missing' to "automagically" try to resolve
> them (with all the risk and brittleness entailed), but rather to
> suggest those projects still keeping generated files committed in their
> VCS to provide a proper (say) 'fix-timestamp.sh' script that touches
> the timestamp of the checked-out files, to ensure no spurious rebuild
> will be triggered.  As a bonus, such a script can be more aware of the
> particularities, nooks and corner cases of a project, and thus more
> reliable than the old 'missing' script.
>
> An example of this approach is offered by GNU awk (release 4.0.1, Git
> tag 'gawk-4.0.1', commit b85b04e8).  The GNU awk maintainers commit
> the Autotools-generated files (configure, Makefile.in, etc.) and other
> generated in the project's Git repository, but offer a useful (albeit
> IMHO poorly named) 'bootstrap.sh' script that fixes the timestamps of
> those files, to ensure no useless remake is triggered in a freshly
> clones repository:
>
>    #! /bin/sh
>    # bootstrap.sh --- touch relevant files to avoid out-of-date issues
>    #                  in Git sandboxes
>    touch aclocal.m4
>    find awklib -type f -print | xargs touch
>    sleep 1
>    touch configure
>    sleep 2
>    touch configh.in
>    sleep 1
>    touch test/Maketests
>    find . -name Makefile.in -print | xargs touch
>    touch doc/*.info
>    touch po/*.gmo
>    touch po/stamp-po
>    touch awkgram.c
>    touch command.c
>    touch version.c
>
> A similar, simplified script is also reported as an example in the
> manual.
>
> * NEWS: Update.
> * doc/automake.texi (Auxiliary Programs, maintainer-mod): Update.
> (CVS): Do not suggest that 'missing' can "automagically" fix botched
> timestamp due to a "cvs update"; this isn't true anymore.  Give an
> example of a custom 'fix-timestamp.sh' script that can be used to
> obtain the same effect (in a slightly more laborious but also more
> explicit and less brittle way).  Fix some minor typos and improper
> wordings while we are at it.

I had no idea missing would monkey with timestamps like that.  I'm
wholeheartedly in favor of removing that capability in the name of
predictably correct results.

While I agree fix-timestamps.sh might be a clearer name, using the
name bootstrap for the clone-from-VCS before-first-configure step.
NTP tarballs carry a "bootstrap" script, though it is not needed nor
recommended when building from a tarball without changing sources.
Naming it fix-timestamps.sh would be misleading, as it also invokes
{$AUTORECONF-autoreconf} -v -i.  I suspect that nomenclature of
bootstrap being the step after clone and before autotools or make.
The other naming I've seen for such scripts is even less appealing,
autoconf.sh.

Cheers,
Dave Hart