Thursday, January 28, 2021

Strengths, weaknesses, opportunities, and threats facing the GNU Autotools

I’ve been a contributor to GNU projects for many years, notably both GCC and GNU libc, and recently I led the effort to make the first release of Autoconf since 2012 (release announcement for Autoconf 2.70). For background and context, see the LWN article my colleague Sumana Harihareswara of Changeset Consulting wrote.

Autoconf not having made a release in eight years is a symptom of a deeper problem. Many GNU projects, including all of the other components of the Autotools (Automake, Libtool, Gnulib, etc.) and the software they depend upon (GNU M4, GNU Make, etc.) have seen a steady decline in both contributor enthusiasm and user base over the past decade. I include myself in the group of declining enthusiasts; I would not have done the work leading up to the Autoconf 2.70 release if I had not been paid to do it. (I would like to say thank you to the project funders: Bloomberg, Keith Bostic, and the GNU Toolchain Fund of the FSF.)

The Autotools are in particularly bad shape due to the decline in contributor enthusiasm. Preparation for the Autoconf 2.70 release took almost twice as long as anticipated; I made five beta releases between July and December 2020, and merged 157 patches, most of them bugfixes. On more than one occasion I was asked why I was going to the trouble—isn’t Autoconf (and the rest of the tools by implication) thoroughly obsolete? Why doesn’t everyone switch to something newer, like CMake or Meson? (See the comments on Sumana’s LWN article for examples.)

I personally don’t think that the Autotools are obsolete, or even all that much more difficult to work with than some of the alternatives, but it is a fair question. Should development of the Autotools continue? If they are to continue, we need to find people who have the time and the inclination (and perhaps also the funding) to maintain them steadily, rather than in six-month release sprints every eight years. We also need a proper roadmap for where further development should take these projects. As a starting point for the conversation about whether the projects should continue, and what the roadmap should be, I was inspired by Sumana’s book in progress on open source project management (sample chapters are available from her website) to write up a strengths, weaknesses, opportunities, and threats analysis of Autotools.

This inventory can help us figure out how to build on new opportunities, using the Autotools’ substantial strengths, and where to invest to guard against threats and shore up current weaknesses.

Followup discussion should go to the Autoconf mailing list.

Strengths

In summary: as the category leader for decades, the Autotools benefit from their architectural approach, interoperability, edge case coverage, standards adherence, user trust, and existing install base.

  • Autoconf’s feature-based approach to compiled-code portability scales better than lists of system quirks.
  • The Autotools carry 30+ years’ worth of embedded knowledge about portability traps for C programs and shell-based build scripting on Unix (and to a lesser extent Windows and others), including variants of Unix that no other comparable configuration tool supports.
  • Autoconf and Automake support cross-compilation better than competing build systems.
  • Autoconf and Automake support software written in multiple languages better than some competing build systems (but see below).
  • Autoconf is very extensible, and there are lots of third-party macros available.
  • Tarball releases produced by Autotools have fewer build dependencies than tarball releases produced by competing tools.
  • Tarball releases produced by Autotools have a predictable, standardized (literally; it’s a key aspect of the GNU Coding Standards) interface for setting build-time options, building them, testing them, and installing them.
  • Automake tries very hard to generate Makefiles that will work with any Make implementation, not just GNU make, and not even just (GNU or BSD) make.
  • The Autotools have excellent reference-level documentation (better than CMake and Meson’s).
  • As they are GNU projects, users can have confidence that Autotools are and will always remain Free Software.
  • Relatedly, users can trust that architectural decisions are not driven by the needs of particular large corporations.
  • There is a large installed base, and switching to a competing build system is a lot of work.

Weaknesses

In summary: Autoconf’s core function is to solve a problem that software developers, working primarily in C, had in the 1990s/early 2000s (during the Unix wars). System programming interfaces have become much more standardized since then, and the shell environment, much less buggy. Developers of new code, today, looking at existing configure scripts and documentation, cannot easily determine which of the portability traps Autoconf knows about are still relevant to them. Similarly, maintainers of older programs have a hard time knowing which of their existing portability checks are still necessary. And weak coordination with other Autotools compounds the issue.

Autoconf

  • Autoconf (and the rest of the Autotools) are written in a combination of four old and difficult programming languages: Bourne shell, the portable subset of Make, Perl, and M4. Competing build systems tend to use newer, more ergonomic languages, which both makes it easier for them to get things done, and makes it easier for them to attract new developers.
  • All the supported languages except C and C++ are second-class citizens.
  • The set of languages that are supported has no particular rationale. Several new and increasingly popular compiled-code languages (e.g. Swift and Rust) are not supported, while oddities like Erlang are.
  • Much of that 30 years’ worth of embedded knowledge about portability traps is obsolete. There’s no systematic policy for deciding when some problem is too obsolete to worry about anymore.
  • Support for newer platforms, C standard editions, etc. is weaker than support for older things.
  • Autoconf’s extensibility is unsystematic; many of those third-party macros reach into its guts, and do things that create awkward compatibility constraints on core development. Same for existing configure.acs.
  • The code quality of third-party macros varies widely; bad third-party macros reflect poorly on Autoconf proper.
  • Some of the ancillary tools distributed with Autoconf don’t work well; most importantly, autoupdate (which is supposed to patch a configure.ac to bring it in line with current Autoconf’s recommendations) is so limited and unreliable that it might be better not to have it at all.
  • Feature gaps in GNU M4 hold back development of Autoconf.
  • There are few active developers and no continuing funders.
  • GNU project status discourages new contributors because of the paperwork requirements and the perceived lack of executive-level leadership.
  • There is no continuous integration and no culture of code review. Test suites exist but are not comprehensive enough (and at the same time they’re very slow).
  • Bugs, feature requests, and submitted patches are not tracked systematically. (This is partially dependent on FSF/GNU infrastructure improvements which are indefinitely delayed.)
  • There’s a history of releases breaking compatibility, and thus people are hesitant to upgrade. At the same time, Linux distributions actively want to force-upgrade everything they ship to ensure architecture support, leading to upstream/downstream friction.
  • Guide-level documentation is superficial and outdated.
  • Building an Autotools-based project directly from its VCS checkout is often significantly harder than building it from a tarball release, and may involve tracking down and installing any number of unusual tools.
  • The Autotools depend on other GNU software that is not actively maintained, most importantly GNU M4, and to a lesser extent GNU Make.
  • Coordination among the Autotools is weak, even though the tools are tightly coupled to each other. There are portions of codebases that exist solely for interoperability with other tools in the toolchain, which leads to overlapping maintainer and reviewer responsibility, slow code review and inconvenient copyright assignment processes multiplying, and causing confusion and dropped balls. For instance, there is code shared among Autoconf, Automake, and/or Gnulib by copying files between source repositories; changes to these files are extra inconvenient. The lack of coordination also makes it harder for tool maintainers to deprecate old functionality, or to decouple interfaces to make things more extensible; maintainers do not negotiate policies with each other to help. For instance, Autoconf has trouble knowing when it is safe to remove internal kludges that old versions of Automake depend on, and certain shell commands (e.g. aclocal) are distributed with one package but abstractly belong to another.
  • Division of labor among the Autotools, and the sources of third-party macros, is ad-hoc and unclear. (Which macros should be part of Autoconf proper? Which should be part of Gnulib? Which should be part of the Autoconf Macro Archive? Which should be shipped with Automake? Which tools should autoreconf know how to run? Etc.)
  • Automake and Libtool are not nearly as extensible as Autoconf is.
  • Unlike several competitors, Automake only works with Make, not with newer build drivers (e.g. Ninja).
  • Because Automake tries to generate Makefiles that will work with any Make implementation, the Makefiles it generates are much more complicated and slow than they would be if they took advantage of GNU and/or BSD extensions.
  • Libtool is notoriously slow, brittle, and difficult to modify (even worse than Autoconf proper). This is partially due to technical debt and partially due to maintaining support for completely obsolete platforms (e.g. old versions of AIX).
  • Libtool has opinions about the proper way to manage shared libraries that Linux distributions actively disagree with, forcing them to kludge around its code during package builds.
  • Alternatives to Libtool have all failed to gain traction, largely because Automake only supports building shared libraries using Libtool or an exact drop-in replacement.

Opportunities

Because of its extensible architecture, install base, and wellspring of user trust, Autotools can react to these industry changes and thus spur increases in usage, investment, and developer contribution.

  • Renewed interest in Autotools due to the Autoconf 2.70 release.
  • Renewed interest in systems programming due to the new generation of systems programming languages (Go, Rust, D, Swift(?), Dart(?), etc. may create an opportunity for a build system that handles them well particularly if it handles polyglot projects well (see below).
  • Cross-compilation is experiencing new appeal because of the increasing popularity of ARM and RISC-V CPUs, and of small devices (too small to compile their own code) based on these chips.
  • The Free software ecosystem as a whole would benefit from a reconciliation between the traditional model of software distribution (compiled code with stable interfaces, released as tarballs at regular intervals, installed once on any given computer and depended on as shared libraries and/or binaries) and the newer depend directly on VCS checkouts and bundle everything model described below. Autotools contributors have the experience and knowledge to lead this effort.
  • Funding may be available for projects targeting the weaknesses listed above.

Threats

These threats may lead to a further decrease in Autotools developer contribution, funding, and momentum.

  • Increasing mindshare of competing projects (CMake, Meson, Generate-Ninja, …).
  • Increasing mindshare of programming languages that come with a build system that works out of the box, as long as you only use that one language in your project. (These systems typically cannot handle a polyglot project at all, hence the above opportunity for a third-party system that handles polyglot projects well.)
  • Increasing preference for building software from VCS checkouts (perhaps at a specific tag, perhaps not) rather than via tarballs.
  • Increasing mindshare of the software distribution model originated by Node.js, Ruby, etc. where each application bundles all of its dependencies. While this is considered a profoundly bad idea by Linux distribution maintainers in particular (because it makes it much harder to find and patch a buggy dependency) and makes it harder for end-users to modify the software (because out-of-date dependencies may be very different from what their own documentation—describing the latest version—says), it is significantly more convenient for upstream developers. Competing build systems handle this model much better than Autoconf does.

Thanks to Sumana Harihareswara for inspiration and editing.

Followup discussion should go to the Autoconf mailing list.



from Hacker News https://ift.tt/39fhG5E

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.