Varnish guy's hatred for autoconf
varnish-cache.org
varnish-cache.org
I wrote a major part of a complex build system in autoconf[1,2]. I gotta say, just about ANYTHING would be better. There's no abstraction; at no point can you forget you're writing m4sugar (implemented on top of m4) that is writing bourne shell code. There's no indication if you screw up in subtle ways except an sh parse error when you run configure. I've had to run configure with sh -x more times than I can count. EVERY SINGLE autoconf release changes the interface, and there's no reasonable way to detect the autoconf version to work around it. Every autoconf release requires hours of tinkering. Autoconf 2.65 changed the quoting rules for m4_upper and m4_lower; making those work reliably on old and new versions is obnoxious.
I've spent many, many hours working on autoconf code. I consider myself an expert at it. And yet, I'm willing to throw it all away and replace it with ANY build system that isn't a pile of junk. I just haven't found one yet :(
The only ones that rise above the noise are scons, waf, and cmake. (No, I'm not using jam/bjam or ant. DIAF.) CMake does most of what I want, except they invented a completely idiotic new programming language. Waf and CMake both provide me with a system to make a build system instead of a build system itself (screw that). SCons reinvents make... poorly and slowly. Of these, only cmake handles windows without forcing me to sit in front of a windows machine for hours. (Autoconf, obviously, falls over hardcore on windows unless you set up a Unixy build environment on it. Not The Point.)
Anyway, that was a cathartic rant. (No, but seriously. I'm up for more suggestions, or you can try to convince me that my opinions of the above tools are wrong. I really do want to switch away, I just haven't found anything that's a big enough improvement to justify it.)
1) http://github.com/visionworkbench/visionworkbench/tree/maste...
2) http://github.com/visionworkbench/visionworkbench/blob/maste...
It's fairly easy to search and link to third-party libraries. Really, yes, the language is a tad weird, but very simple and frankly it's a small price to pay to have a real cross platform C++ makefile system.
As you said the fact that CMake supports Windows and Visual Studio out of the box is a big plus as well.
CMake also enables you to automatically run unit tests and package your release into a tar.gz or a setup.
Really, we're very, very happy with CMake and the effort we put in it paid off a lot. I really encourage anyone needing cross platform makefiles to check it out.
To sum it up, CMake is a makefile compiler. You write your makefile in "CMake", and it compiles it into a classic Makefile, Visual Studio project, etc.
I've played around with CMake also, and I've found it to be the most reasonable. This is especially true if you have a diverse team with members that prefer Emacs/VI or Visual Studio and a project that needs to work natively on multiple platforms. Different people work better in different environments, and there's nothing wrong with that.
It's not perfect, but I find myself saying 'WTF' less when using CMake. I have no idea why they decided to make their own little language, but its syntax is super-simple, and very well documented.
No, I'm not using jam/bjam or ant. DIAF.
I've used it in several projects so far & although it did require the occasional tinkering it wasn't even close to being as big of a time sink as autotools were.
Well, not usually.
Their conditional expressions are goofy, it's never clear whether mixing variables and lists will do the right thing, and good luck doing something complex using only documentation. So yes, the CMake language is bad.
But once you write a file, it Just Works. I don't usually need to revisit a CMake file to correct something, unless I'm adding/removing files or dependencies. I never had that feeling with bjam or autoconf - small changes to the code always had a major impact on these build systems. We compile code on both Windows and Linux, so it's the only viable option, and thankfully it's not too bad :)
It's been coded in C, is scriptable with lua and that would get around the CMake specific language. It generates the project files for the platforms you want, so it may save you from writing the autoconf files yourself.
Autotools are exceptionally bad for several reasons:
- They are difficult and not at all intuitive to learn and use. As another poster noted, there are no abstractions whatsoever, so what's stopping me from just writing a shell script that will check for the location of libparadise v3.1.6? This may be a subjective issue, so please check out gtkpod's CMakeLists.txt and configure.in and tell me with a straight face that you'd pick autotools.
- They are well-documented and easily extendable...Oops, no they're not!
- They aren't backward or forward compatible, and almost every other major version has serious regressions. This is a big whoop if you claim to want to produce portable software.
- Configuration is scattered around the build tree instead of one file. Seriously, I don't need this. I have enough of an issue with my current project's 400+ modules, thank you very much.
- Slow. They're beautified shell scripts, after all. But also, they check for unneeded dependencies and antiquated features (checking for an ANSI-C conforming const...OK).
- And last but not least, its "caching" is the most braindead half-implemented feature I've ever encountered since 1984. It's literally cost me hours to discover problems caused by caching the config files.
On the other hand, there aren't any better alternatives, so I'll just go cry for a bit.
Indeed. As the person who cross-compiled a ton of libraries for Windows (http://fedoraproject.org/wiki/SIGs/MinGW) I can tell you that autoconf and CMake are the only build environments which handle cross-compilation properly and transparently.
BTW CMake is not better or worse than autoconf, just different.
I have never seen "configure" work particularly badly, and I've had to build some pretty hairy, dependency-ridden crap on Linux, Solaris and Mac OS X. It has held up remarkably well, and many projects are good about providing useful "configure" options. Usually fixing something requires an environment variable or configure option, and not a makefile hack.
There is no question the generated code is practically indecipherable. But you have to treat it like a compiler; no one goes leafing through the ".o" that GCC generates on a regular basis, so why worry about what "autoconf" and "automake" produce? Like any widely-used tool (such as a compiler), you can put a fair amount of trust in the maintainers to not make anything too broken; and if you see a problem, you can file a bug report.
It is also somewhat reassuring that even if the generated result is wrong, it is at least theoretically possible to fix it; just try fixing something that goes wrong in Visual Studio.
I will allow that "m4" is showing its age, and these days it's a heck of a lot easier to Google some examples and adapt them, than it is to figure out how to add a custom rule from scratch. The big challenge for the GNU Build System is to create a modern version that makes new assumptions (e.g. "perl" always exists), and update their methods accordingly.
Actually, object files are nothing but undecipherable. They are machine code, well documented by the architecture manuals and the ELF format. There are also tools to introspect them (objdump, ld, etc.), which, believe you me, is done on a regular basis by a lot of people.
Try the -S flag to the compiler, that way you don't need to 'leaf though the .o file' (you can't really leaf through it, it's not a text file, but an assembly source file is.
It probably does, but most infrastructure software in both RHEL/Fedora and Ubuntu has been created in Python for nearly a decade now.
Git can optionally use autoconf. To do this you can run make configure && ./configure which'll use this configure.ac file to generate the configure script: http://github.com/git/git/blob/pu/configure.ac#L1
Autoconf will then write out a config.mak file, which is generated from this template: http://github.com/git/git/blob/pu/config.mak.in#L1
That's where autoconf's involvement ends. All it's doing is turning simple macros like this:
# Define NO_GETTEXT if you don't have libintl.h
AC_CHECK_HEADER([libintl.h],
[NO_GETTEXT=],
[NO_GETTEXT=YesPlease])
AC_SUBST(NO_GETTEXT)
NO_GETTEXT=
NO_GETTEXT=YesPlease
Autoconf can certainly get hairy, but in most cases using it is a lot easier than not doing it. Consider this definition:
AC_CHECK_LIB([curl], [curl_global_init],
[NO_CURL=],
[NO_CURL=YesPlease])
Also note that the autotools are more than just autoconf. You can stop here, but you can also generate your makefiles with automake, and create your libraries with libtool.
On the bright side, there aren't no triple negatives neither.
autoconf is a bit of a Faustian bargain though - it promises a quick and easy solution to the labour-intensive and mindlessly boring task of managing all the microscopic differences between (and within!) Unix variants. Unfortunately it brings its own complexity, fragility and bugs along with it. It also encourages a programming style best described as "#ifdef hell" which is just poison to good maintainability.
Autocrap is horrible in two ways: First, it's horrible to work with; and second, it's responsible for making any attempt to avoid it horribly painful.
If autocrap didn't exist, people would write operating systems without so many gratuitous incompatibilities. What possible excuse is there for the fact that many linux distributions need a special compiler flag in order to support files greater than 4 GB?
- bsd and derivates (including OS/X)
- linux (64 and 32 bit sub-flavors)
- QnX (though very different under the hood)
And lots of others. So 'it's an 80's thing' but it will be very hard to get rid of, 'just a few different linux distributions' does not describe the situation accurately.
It also isn't rare to find older boxes still chugging away quietly at some task in a corner somewhere running any one of those older operating systems and those sometimes need software compiled for them as well.
For instance Git which is a fairly popular package created in 2005 runs on at least: Linux, GNU, GNU/Hurd, GNU/Linux, Tru64 UNIX, GNU/kFreeBSD, SCO UnixWare, SCO UNIX SV, Darwin, various versions of Solaris, Cygwin, *BSD, AIX, IRIX, IRIX64, HP-UX, Windows, MINGW.
Older packages like Perl and Emacs compile on a much wider range of systems.
But times have changed. Windows is now more compatible with Linux than a random pair of mid-90s UNIXes were with each other. Now that the major incompatibilities gone, all autocrap is doing is keeping the lesser incompatibilities around.
Windows doesn't have forks (needs emulation via threads), second class support for things like symlinks, and isn't POSIX compatible (yes, I know you can get that as a separate install).
Porting a Linux program to Windows is like porting it to Novell, or some random mainfram OS. Not another OS in the Unix family.
Varnish guy also left HP-UX off his list of still-kicking Unix implementations.
The continuing exponential growth of computer use hides the fact for many that there are still an enormous number of "non-mainstream" systems still in essential use. Within the last couple weeks, someone on the autoconf list just asked for help supporting AS400.
Kids these days. Sheesh. ;-)
http://people.freebsd.org/~phk/
Varnish is a pretty good example of how far you can push a unix box in terms of network performance, one site I work for/with serves up billions of images daily using varnish without a single hickup.
If phk criticizes something it's usually worth listening to, he really knows his stuff.
What do you think it is that makes Varnish so fast? I've read the article (http://varnish-cache.org/wiki/ArchitectNotes) but I'm curious to hear about it from people's experience. Sometimes the stated reasons for something being great are not the same as the real reasons.
It's a textbook example of how to do a thing like this.
The major point I took from the articles is that you should let the kernel manage what's in RAM -- it's going to do it anyway, so don't fight it -- and instead of building complicated paging and caching mechanisms, focus on laying out your data to take advantage of how the kernel already does it.
What I want to know is, what are the other such major points?
p.s. Good luck tomorrow :)
Especially interesting to me because I spent 5 years on memory issues in grad school, steeped in the idea that number of comparisons or number of instructions per unit input is often a red herring. Unfortunately the alternative is a much more complex model to juggle in your head.
Can you expand on this? What do you think of the article's argument and how does it relate to what you studied? Perhaps this should be in its own thread, but I'm really curious about this because, right or wrong, what he's advocating is not a common design style.
I spent a good couple of years thinking up layout optimizations like the one in Figure 6. I think that's fairly common design style, when people choose to think about this stuff. The challenge is coming up with consistently better layouts for a specific application. One benefit of java, for example, is that you can move things around without worrying about pointer arithmetic like in C.
And his rant about Figure 7 resonates. Again, it's brain-dead obvious that you should take the memory hierarchy of your system into account because you waste millions of cycles waiting on a disk read, but it's really hard to do. Our education system doesn't help, and neither do mainstream tools.
I think of it as concentric revolutions. The pentium's algorithm for processing instructions out-of-order first showed up in a mainframe 30 years before (http://en.wikipedia.org/wiki/Tomasulo_algorithm). The internet gradually rediscovers what desktop devs gradually rediscovered what the minicomputer devs originally dreamt up. Most of the time it's not common yet simply because people haven't gotten around to it.
It's basically YAGNI on the scale of industry segments. A successful startup will gradually make its infrastructure more stable and higher-performance after it's ensured it'll be around. And a successful new technology will first validate its reason for being before gradually trying to use what previous technologies have learned.
Sorry this is poorly written. I'm not sure what parts of this are non-obvious, so feel free to ask more questions.
The selling point of the autoconf tools is that writing your own shell script is tedious and error prone. This was not my experience.
http://github.com/chkoreff/Clump/blob/master/src/README
People have been downloading it consistently ever since, and I know some of them like it a lot.
So with this program you just cd into your code directory, type "clump", and it builds your executable program in ../bin.
One current limitation: it expects all the .h and .c files to be in a single directory. I could enhance it to walk through a directory structure, but I haven't gotten around to it. Yeah I know, even after 8 years. :)
Besides - most the replacements that have been written eventually end up even more complicated than autotools..
The whole point of autotools is that it allows you to have a different configuration for every system the software is built on. However, this is a very undesirable property for modern software; instead, we want our builds to be exactly repeatable. If you download my software and build it, the results should be bit-for-bit identical to the results I get. If they're not, then how do you know if your build is correct?
It is funny how most software today is auto-customized using autotools and then a package maintainer for the software for some OS comes in and un-does all that auto-customization to get repeatable results. It would be much better to just assume there's going to be a package manager configuring everything exactly as he wants it, and avoid autotools completely.
It also gives users a familiar way to optionally enable features at build time, such as choosing between MySQL and PostgreSQL support, as opposed to requiring the user to edit some Makefiles and twiddle some #define's manually.
POSIX says that it's <sys/stat.h>. Problem solved.
determining the platform's word size
sizeof(size_t). Problem solved.
Other than that, the real world isn't black and white either. Although many unixen claim to be POSIX compliant, based on my experience they are mostly POSIX-ish.
And there's also no guarantee for sizeof(size_t) being a platform's word size.
Sadly what POSIX says is cold comfort when you're trying to compile something and the computer in front of you has put stat.h somewhere else.
Pluses:
Fast, probably the fastest build system there is after a hand-coded Makefile. MD5 sums are clearly not the bottleneck in SCons, and it is a fundamental design problem there.
Waf uses Python. This means that you can use all the right idioms for lists of things, loops, etc.
Not just a Make-clone, waf has build-system features. For example, it can check for headers, libraries, use pkg-config, has cross-compilation support, can create distributions, etc. Waf has separate configure and build steps too - in SCons if you use the autoconf-like features, they are checked every time you build (albeit cached).
Waf by default does an out of source build. This is also a minus though, as it is only available as a sub-directory of the source directory. You cannot build a single source from read-only media on various platforms, as waf wants to write files below the source tree whatever you do.
Intra-project dependencies are easy to use. Convenience-style libraries add the right -L and -I values to executables that use them.
File globbing works well. It isn't like CMake where using FILE(GLOB *.c) has a bunch of caveats.
Minuses
Waf uses Python. This means you have to quote everything. In make or CMake, it's main.c. In Waf (and SCons) you have to write "main.c". Similarly for command line arguments and so on. Being a proper programming language means you can "over-engineer" the build system, ending up with something that is perhaps well-written, but far more difficult to follow that the equivalent almost-declarative CMake input file.
Writing your own tools to process types of inputs that are not supported by default seemed tougher to me than with SCons or with CMake. Waf uses Python decorators to add methods to the default build context, which are then executed depending on the decorator-filters for the types of file in question. It is an elegant solution, but in practice too magical and it is tricky to debug.
Waf is really unstable (in the sense that it changes regularly, not that it crashes). The API changes every release, features are removed and deprecation warnings are rare. So if you hit a bug that gets fixed in a newer version and want to upgrade, chances are that your wscript file will no longer work. You are encouraged to ship a Waf binary-blob version with your source, similar to the configure "shell-blob" I guess, so if you can work around any bugs or live without newer features and stick to a specific version then this may be less of an issue.
There are warts in the API - sub_options vs sub_config vs add_subdirs to recurse directories, variable and function names randomly shortened (blddir, env, add_post_fun, uselib), having to pass an argument 'features="cc cprogram"' to compile an executable vs a dedicated cprogram method.... Version 1.6 will address most of these, but again you face the upgrade issue then.
Conclusion
I would use CMake for new stuff. It is not as cool, doesn't have an uninstall target, but is more compatible with people's expectations (the way an out of source build works, for example), has code coverage, unit tests, as well as being tested on more platforms. It's language is easy to learn - there is only COMMAND(arguments) - and the build scripts you write are generally "brain dead", it's impossible to write code that is difficult to follow. You can tell it has been used by more people and has a lot of the wonky parts hammered out.
Also multiple incompatible compilers, and cross-compiling… and Darwin doesn't have uname -o, and Linux's uname -o has a / in the output.
Hmm, what isn't wrong with it?
http://www.thinkgeek.com/tshirts-apparel/unisex/frustrations...
Much of what he says is fair. Many decisions made way back when would be different now. But the autotools are well-documented, well-maintained, and they do work.
The way forward (for the autotool-friendly audience anyway) is likely something like Tromey's quagmire -- http://code.google.com/p/quagmire/.
The problem is that most people have to extend things with their own tests and disregard cross-compiling (by assuming that the build machine is the same as the target machine).
(The VCL domain-specific language is translated to C, compiled to shared lib, and dlopen(2)'ed, that eliminated the runtime overhead of an interpreter)
/phk
Granted, we have an automated build system that requires hand tweaking from time to time, but that's something we introduced ourselves. If I wanted I could still hit shift-ctrl-b on any box and have a working version of the project. (And in all the flavor-of-the-month web techs it's usually just a matter of pointing a browser at the source.)
Yeah, all of them (serious answer).
Granted, we have an automated build system that requires hand tweaking from time to time, but that's something we introduced ourselves.
An automated build system that A) does not require someone to sit in front of a machine and press a button to start it. B) doesn't just produce executables based on a working copy, but is aware of SCM (including branches), runs unit tests (and sends alert mails when they fail), does code analysis, and can be arbitrarily extended to do whatever else you need is NOT a luxury, but an essential requirement. If you don't have it, you're not developing, you're toying around.
I had read the comments here to mean that people still have to toy with scripts just to get their project to compile.
So you're saying that the source -> executable problem is indeed solved for every language. It's the (source control) -> (source code) + (db change scripts) -> (QA Website) + (QA Database) problem that we're talking about. I didn't gather that from the article.
At that point I understood why Microsoft still has a positive balance sheet.
Most of the libs I've used under VC had their own build files for just that environment exactly because you don't really need all the autoconf / make crap in there. This doesn't have anything to do with MS's balance sheets imho, simply because you'd have a much better experience compiling gnu make-based ODE libs under a ...nix system. The things you have to fight with so much create a build system, which you can take for granted on a target system (as in `apt-get install build-essentials`).
Your issue seems to me orthogonal to what the article is about, on many levels. Spending more time in ...nix environments, I could just as well wonder: "Why is it so hard to have a sane POSIX-friendly, command-line build environment in MS systems? At that point I don't understood why Microsoft still has a positive balance sheet."
If the project maintainer doesn't include binaries, and nobody else posts them, then the user doesn't have a choice.
You'll find that ODE can be downright Windows-hostile in pieces of its implementation that have nothing to do with GNU toolchain issues. Take a look at http://www.ode.org/old_list_archives/2003-October/010046.htm... for the most common problem; they don't care that the stack has different restrictions in Windows and Linux. You'll also find ODE to be downright thread-hostile; the entire library is non-reentrant (they built their own non-threadsafe layer on top of malloc inside the engine; you can't trust any two functions to be callable from separate threads, even if they have logically nothing to do with each other).
ODE is the first physics package most people seem to suggest, but I think it's more trouble than it's worth.
He's the reason that everyone's [hero|enemy] Zed F. Shaw picked sheddingbikes.org.