Slashdot Mirror
Alternatives to Autoconf?
Despairing Developer queries: "Once autoconf was a great way to make widely portable programs. But now when you have to spend more time sorting out incompatibilities between autoconf versions, breaking battles between autoconf compatibility wrappers and configure.in compatibility functions that tries to outsmart each other, and on top of that see the list of dependencies increase (it's not that fun to compile perl on unicos) and performance diving rapidly, what is your escape plan? Is there a drop in replacement for autoconf? Is there something else out there that is as portable as autoconf to use instead?"
QMake (from Trolltech) is a possible replacement, though not free (according to the FSF)!
i wish there was a way to moderate articles up, because you've hit on one of my major (*major*) pet psychotic hatreds regarding developing software. auto(conf|make) sucks badly. it's bearable if you're developing from scratch (not depending on other libraries) or require that your bundled versions of libraries be used. but when your software depends on, say, 123098123871237 other packages (i.e. you're writing for gnome or kde), you're boned.
unfortunately, there are no reasonable replacements that i know of, which is probably a testament to the nastiness inherent in solving this problem. a pity, really -- auto(conf|make) and company are a really good idea (in theory). unfortunately, there seems to be some really bad crack smoke involved in designing these tools. first (and probably foremost) in my mind is why isn't there a database of some sort which would at least allow the option of keeping track of which versions of what applications have been configured how and installed where.
These are both sort of combined configuration and build systems (which is the way it should be IMHO). Scons requires Python (>=1.5.2, IIRC), so it is as "only" as portable as Python itself (which is to say "very") while cmake doesn't require anything except a C++ compiler. The actual "language" in Scons is just regular python while cmake uses a hideous custom language.
HAND.
I feel your pain, but this isn't just autoconf. Its a general theme of the way we compute. Version nightmares. I think the problem is unavoidable because of the way we currently compute: 1) Competition and the enormous diversity today will always leads to heterogenous systems, no matter how good their intentions initially 2) The semantics of software and its environment are not embedded in the data, which of course means when a version changes, something somewhere breaks, and someone somewhere has to fix it Theres only two solutions, either by decreasing diversity through standardisation (oh heck the version of that keeps changing too), or real autonomic computing operating at a higher level of abstraction. Roll on autonomic computing, real self-configuring and self-adapting systems. Until then, we can only attempt to minimise the problems, and can only ever solve them in a limited scope for a limited period of time.
they may be a drop in replacement for developers, but for packagers and people trying to track changes and new versions, both cmake and scons (blender!) are horrible. They cost us (A group of 10 people working on a distro) enormous amount of extra time (blender's upgrade to
all in all autoconf maybe a problem for developers, but for packagers it is still *by* *far* the best.
SCons isn't the solution either. SCons relies way too heavily on python and doesn't make better distribution tarballs. Most developers using SCons rooll out horrible tarballs that cannot even detect the proper Python version! (blender!!!).
SCons makes you lazy, do the work, then your application builds BETTER on MORE platforms with autoconf
You can find it at pmk.sourceforge.net
Or else, you can have a look at A-A-P, by nobody else than Bram Moolenaar, the author of the One True Editor, a.k.a. ViM :-)
There is also Package-framework, by Tom Lord, the author of the infamous Arch SCM.
I was about to mention SCons, too, but other people already did (it always pay to check other comments just before posting, especially on /. :-)
To sum it up : there is no shortage of alternatives to the incredibly hairy Autoconf/Automake nightmare. The problem is, people are still using them for the very same reason they use CVS instead of Arch/Subversion, or Sendmail instead of Postfix/Exim : because they're considered ``standard'' tools, and people feel more comfortable with software they know to be used by plenty of other people (millions of programmers can't all be wrong. Can they ?). I really hope they'll stop making this kind of mistakes soon, so I won't need to curse them everytime I have to debug some Autoconf breakage...
Xenu brings order!
pkg-config singlehandedly is creating a sea of calm in autoconf land while still keeping the strength of autoconf. Take Xfce4 for instance. Within a relatively short timeframe this small group (4-5 devs) have written an entire framework that compiles on tens of platforms by using PKG_CONFIG extensively together with autoconf
> Is there something else out there that is as
> portable as autoconf to use instead?
Yeah. It's called GNU Make.
Seriously, if you write your makefiles and your
code in a responsibly portable manner, there's
absolutely no reason for autoconf or automake.
And it's not hard. I've done it repeatedly.
The auto* tools are an antipattern virus.
-I like my women like I like my tea: green-
Chances are the other packages people here are talking about haven't ever been built on a cray, too. A Cray is not exactly common-place.
;) I'd think that the autoconf developers would appreciate a good bug report, just like any other project would.
:)
Usually, the simplest way to deal with broken scripts/automake/autoconf tests, is to use a better shell. Take the bash, tcsh, or whichever shell works on Unicos, and run your autoconf tests throught that. If you think you've found a problem in autoconf itself, run the regression test suite and submit a bug report.
A quick google search over the autoconf mailing list archive shows that there were 0 posts on Unicos in the last two years. So chances are, that any bugs in autoconf itself in the last two years have not been tested, or that it has no bugs on Unicos
If your problem is autoconf's performance, then a) use a faster shell, b) use --cache-file, c) use ccache to run those myriads of little C tests faster. Or fix the tests in question.
A classic on slow platforms is the maximum shell length test. That one comes from libtool, not from autoconf. It's been vastly improved in libtool 1.5.6, for example. If your upstream doesn't use a recent version of libtool, convince them to do so.
I was pretty fed up with autoconf myself, and wrote a little C app to emulate it. Initial ./configure time dropped to a second and reconfigure is instantaneous. I would recommend it for your simpler projects:
bsconf.c
and
bsconf.h,
the latter being the configuration file.
... Those who do not understand autoconf are doomed to reinvent it. Poorly.
What kind of ${CRAP} are you writing that need ${YOURLANGSUCKS} anyway?
${MYLANGROXORS}
${FSCKYOU}
Case in point: I wanted to compile cairo on my Mandrake 9.1 system. I couldn't until I edited the autoconf file to remove "new commands" and added phony files to make pkgconfig happy. Then it compiled just fine and worked. I tried to compile the demos and was completely frustrated and eventually hand-wrote a trivial makefile and they all compiled just fine and worked (except for the GTK one...). I am now trying to compile the "Glitz" OpenGL backend, and I am running into the same troubles: I can't prove it yet, but I strongly suspect it will compile just fine on my machine, if I can just get around the mysteries and complaints of autoconf/automake/pkgconfig, probably by wasting a great deal of time and divining the basic, and probably simple, Makefile that would compile it.
The incredibly frustrating, and dare I say stupid thing is that the only thing I need to "update" on my machine is the damn autoconf tools! I actually have all the libraries and headers these things need. That is completely backwards! In fact due to autoconf they have pretty much said "this only compiles on the very newest experimental Linux system". Well in that case, you have eliminated the need for "autoconf" and you could send out Makefiles that work only on a new Linux system. That would probably be easier for somebody like me to edit and get working on an older system.
When you do try to fix this, you run into the horrifyingly bad syntax and rules of M4 and GMake. Supposedly this is because they want to be back-compatable and run on older systems. But they lie when they freely add new "autoconf" commands so that the newest version is needed. Why not scrap the whole thing and try a modern syntax?
My proposed solution: make ONE compiled program that does "configure" and "make" and "make depend" (and "install" and "uninstall" and "clean" and "dist" and all the other special-cased targets...). This program can use the existing automake/conf stuff so it can be compiled for multiple platforms. The program then reads one file, in editable text (no XML, and it should be trivial to add/remove a source file by adding/deleting a line). This file should be parsed in a procedural way, with "if" and "for" statements and functions (ie it is perl or something) and the result should be the dependencies tree and it can then run the programs (the result is extremely static and has actual filenames and commands, not templates or "rules"). Make it really easy to add and remove switches to the compiler. Make it save user answers to questions in a file so it can provide those answers again, and make a gui program that provides panels and checkboxes to change those user answers. Make it automatically check for dependencies in any C style source files by looking for "#include" without running the compiler, and save the results in a binary database with date stamps so it can run this instantly as needed. Any package dependenciies should be checked with "if" statements in the file, the program would have enough commands to do typical file system things like look for files or grep a file for something, and it is then trivial to write a file that checks for something in multiple places by using "if-else" statements.
"checking for stdlib.h..." indeed. If you don't have an ISO-C environment, you're so screwed that there's nothing autoconf can do to save you: you don't know how the language works, so whether or not you're missing a few library functions isn't going to make any difference.
I once worked on a fairly large process control system, of which less than 1% was portability code. It ran on Solaris, AIX, HP-UX (this was pre-Linux). It also ran on VMS and WinNT. Most of the portability issues dealt with the entirely different models for process management and memory-mapped files, not with pipsqueek stuff like autoconf.
It's real simple: you read the docs. You determine what the standard actually requires, not what your development system happens to do. And you program to that, then test.
As for the autotools proclaimed ability to port to systems you weren't planning for, that's so much BS. If the system is sufficiently different that ANSI C and POSIX aren't sufficient, then you're going to have to update your code anyway.
Replacing autoconf with some other crutch isn't going to help. Just ditch it entirely.
autoconf has two purposes:
Converting from a plain Makefile to configure.in+Makefile.in is straightforward if your Makefile already uses variables for binaries and directories.
The main reason to use autoconf is the second point above; when you write code that uses system calls (eg read(2), write(2)) and not libc calls (eg fread(3C), fwrite(3C)), that differ between operating systems (eg BSD vs SysV vs POSIX), and need to detect which taste you have.
To answer your questions: autoconf is something I hook in when:
A new program: autoconfconf
I am struck by the recursive nature of the situation.
Unknown host pong.
Indeed. We develop server-side code on Windows (2k and XP), and deploy on Linux, Windows, and Solaris. With three major products representing millions of lines of code developed over a period of approximately 6 years, we have never had a single instance of platform incompatibility.