On Sep 11, 2016, at 16:17, Xavier de Gaye xdegaye@gmail.com wrote:

Changeset 816ae3abd928 regenerated the configure script with 'runstatedir' again. AFAIK Autoconf 2.70 has still not yet been released. Please let us stick with 2.69.

As far as I can tell, the spurious "runstatedir" doesn't affect Python builds. But I'll make sure to rerun autoconf and remove it before tagging for b1.

-- Ned Deily nad@python.org -- []

2016-09-12 8:27 GMT+02:00 Benjamin Peterson benjamin@python.org:

The correct way to solve this is probably to stop checking in the generated configure

Please keep it, it's convenient :-)

and generate it with a "blessed" autoconf version in the release tarballs.

+1 for that: we should modify the release PEP for that?

Victor

On Nov 22, 2016, at 11:06, Xavier de Gaye xdegaye@gmail.com wrote:

The configure file on the default and 3.6 branches have been generated with autoconf 2.70 once again. This is annoying when you have to maintain patches to this configure file in order to build on a non supported platform.

I'm sorry about that. I did promise to rerun with autoconf 2.69 before tagging the release so committers didn't have to worry about it but I didn't notice my note to do so until after 3.6.0b4 had already been tagged. I'll try to do better for rc1.

Perhaps another solution to the problem might be to not include the autoconf-generated changes in the patches and just always run autoconf before doing a build? That's what we suggest for patches submitted to the tracker.

And this might also be a candidate for handling in our upcoming new development workflow, i.e. something like having autoconf automatically be run as part of checkins. If it hasn't already been discussed there, it might be worth bringing up on the core-workflow mailing list.

-- Ned Deily nad@python.org -- []

I do not think we should require individual developers committing changes to configure.ac to use a particular version of autoconf when regenerating configure. That is a burden.

The solution to your problem is to maintain your patches _only_ against configure.ac and rerun autoconf using whatever version you need yourself.

If a release manager decides they want configure generated with a specific version of autoconf at release time, they should rerun that specific version of autoconf and commit the result _for the release_.

otherwise the checked in configure script exists entirely as a convenience. configure.ac is always the source of truth.

-gps

On Tue, Nov 22, 2016 at 11:38 PM Xavier de Gaye xdegaye@gmail.com wrote:

On 11/22/2016 08:16 PM, Ned Deily wrote:

...

On Nov 22, 2016, at 11:06, Xavier de Gaye xdegaye@gmail.com wrote:

...

The configure file on the default and 3.6 branches have been generated with autoconf 2.70 once again. This is annoying when you have to maintain patches to this configure file in order to build on a non supported platform.

I'm sorry about that. I did promise to rerun with autoconf 2.69 before tagging the release so committers didn't have to worry about it but I didn't notice my note to do so until after 3.6.0b4 had already been tagged. I'll try to do better for rc1.

Perhaps another solution to the problem might be to not include the autoconf-generated changes in the patches and just always run autoconf before doing a build? That's what we suggest for patches submitted to the tracker.

And this might also be a candidate for handling in our upcoming new development workflow, i.e. something like having autoconf automatically be run as part of checkins. If it hasn't already been discussed there, it might be worth bringing up on the core-workflow mailing list.

-- Ned Deily nad@python.org -- []

From the configure logs since last july, it seems that Benjamin and Serhiy are the only one using autoconf 2.70:

 changeset 102530:b04560c3ce69 - author Benjamin Peterson
 changeset 103648:816ae3abd928 - author Serhiy Storchaka

If it is not too difficult to build autoconf 2.69 from source, then the solution could be that they switch to autoconf 2.69.

Xavier

---

python-committers mailing list python-committers@python.org https://mail.python.org/mailman/listinfo/python-committers Code of Conduct: https://www.python.org/psf/codeofconduct/

...

On 11/22/2016 08:16 PM, Ned Deily wrote:

...

On Nov 22, 2016, at 11:06, Xavier de Gaye xdegaye@gmail.com wrote:

...

The configure file on the default and 3.6 branches have been generated with autoconf 2.70 once again. This is annoying when you have to maintain patches to this configure file in order to build on a non supported platform.

Perhaps another solution to the problem might be to not include the autoconf-generated changes in the patches and just always run autoconf before doing a build? That's what we suggest for patches submitted to the tracker.

On 23 November 2016 at 22:49, Gregory P. Smith greg@krypto.org wrote:

The solution to your problem is to maintain your patches _only_ against configure.ac and rerun autoconf using whatever version you need yourself.

Would this solution (forget about regerating the files until build time) be enough Xavier?

FWIW I make heavy use of the Mercurial interactive patch mode. I use it to filter out any unnecessary generated changes, while selecting other generated changes relevant to a patch. I.e.

hg commit --interactive hg qrefresh --interactive

On 11/23/2016 11:49 PM, Gregory P. Smith wrote:

I do not think we should require individual developers committing changes to configure.ac http://configure.ac to use a particular version of autoconf when regenerating configure. That is a burden.

I do not agree, configure is a file tracked in the mercurial repository and commit changes to this file must be correct and done with the right tool. If this is a burden for developers to use a particular version of autoconf, then a solution to our problem might be to stop tracking configure in the repository and update the Makefile to run autoconf whenever configure.ac is modified.

On Thu, 24 Nov 2016 12:22:00 +0100, Victor Stinner victor.stinner@gmail.com wrote:

I know that tracking generated files is not pure, but it's very convenient, so please keep them: configure, Python/importlib.h, Python/importlib_external.h, etc.!

When testing Python on some "custom" operating systems, I already had enough issues to compile Python :-) For example, on OpenIndiana, Python wanted to rebuild the grammar but the system Python was 2.6 which didn't work (I don't recall the details). Moreover, the "make touch" command didn't work neither :-)

Right, tracking those artifacts is a long standing policy and exists for good reasons. Our policy is that the committed changes should be done with the "right" version of the tool to minimize churn, and I think we should maintain that policy even if we sometimes screw up.

(I thought we had actually introduced a check for it in the Makefile, but I guess not...that would make it inconvenient for someone to intentionally use a different version for a custom build.)

--David

On Nov 24, 2016, at 10:05, Eric V. Smith eric@trueblade.com wrote:

On Nov 24, 2016, at 10:00 AM, R. David Murray rdmurray@bitdance.com wrote:

...

(I thought we had actually introduced a check for it in the Makefile, but I guess not...that would make it inconvenient for someone to intentionally use a different version for a custom build.)

I'm quite sure we did have such a check back some years ago back when there were less compatible versions of autoconf in wide use. It could be restored if someone wants to write a patch.

Would it be possible to add a commit hook? It seems like that would be the correct place to catch it.

It could but I'm sure you'd agree this is not something we need to be spending time on with our soon-to-be-retired hg-based workflow. As I suggested earlier in this thread, if there is interest in doing something, it should be considered for the new git-based workflow and, "if it hasn't already been discussed there, it might be worth bringing up on the core-workflow mailing list".

-- Ned Deily nad@python.org -- []