Kamp: A Generation Lost in the Bazaar [LWN.net]

Kamp: A Generation Lost in the Bazaar

Posted Aug 21, 2012 2:04 UTC (Tue) by hummassa (subscriber, #307) [Link]

That it needs pthreads?

Kamp: A Generation Lost in the Bazaar

Posted Aug 21, 2012 5:38 UTC (Tue) by pbonzini (subscriber, #60935) [Link]

"test x == y" is unportable.

But it's otherwise a very well-written configure.ac

Kamp: A Generation Lost in the Bazaar

Posted Aug 21, 2012 20:40 UTC (Tue) by cmccabe (guest, #60281) [Link]

Good find. I hope this puts to rest the idea that you need to write "significant shell content" to run into portability problems with autotools. You only need to write one line, or half a line.

By the way, gtest has since switched to CMake.

Kamp: A Generation Lost in the Bazaar

Posted Aug 21, 2012 22:36 UTC (Tue) by nix (subscriber, #2304) [Link]

For that project, who cares? Ceph is intrinsically nonportable, IIRC.

(Arguing that Autoconf is awful because it includes shell fragments is also vitiated in the presence of a project that uses any manually-written shell scripts at all. As most do.)

Kamp: A Generation Lost in the Bazaar

Posted Aug 22, 2012 6:15 UTC (Wed) by cmccabe (guest, #60281) [Link]

As a sometime contributor to Ceph, I can definitely say that no, it is NOT "intrinsically nonportable." The servers which implement the filesystem itself are all userspace processes, which should, in theory, be portable to any POSIX OS. In fact, you can see a few places in the source where someone added #ifdef DARWIN and similar things. There was never a serious effort to get things working well on non-Linux platforms, just due to limited resources and interest. But that may change in the future.

And yes, the kernel component is Linux-only, but there are other ways to access the filesystem.

> (Arguing that Autoconf is awful because it includes shell
> fragments is also vitiated in the presence of a project
> that uses any manually-written shell scripts at all. As
> most do.)

All the standalone shell scripts that are currently in the Ceph server repo are used for testing purposes. That's why they're in directories with names like 'qa', 'test', and so forth. Later on, the project started moving more towards Python for testing... but I digress.

Would you be ok with a build system that required you to write Makefiles in C? Oh, but your project includes .c files, so your arguments are vitiated. Come on, this is absurd, and you know it.

Kamp: A Generation Lost in the Bazaar

Posted Aug 22, 2012 7:40 UTC (Wed) by cmccabe (guest, #60281) [Link]

In case it's not clear, the opinions about autotools here are my own, not necessarily those of other Ceph developers. I'm not claiming to speak for them (although if you are reading this, guys, consider ditching autotools).

Kamp: A Generation Lost in the Bazaar

Posted Aug 22, 2012 13:38 UTC (Wed) by nix (subscriber, #2304) [Link]

I'd be unhappy with a build system that required me to write makefiles in C because C is a bad language for writing makefiles in. However, I note that makefiles are themselves little more than dependencies tied to shell script fragments, so clearly the shell and make are relatively tightly associated (indeed, make knows internally what characters are common shell metacharacters, and stuff like that). configure scripts can in any case accept C fragments too (and Pascal fragments, C++ fragments and the like). They are used where they are suitable, i.e. when trying to figure out properties of the compilation environment.

What I was trying to point out is that the nonportability of shell script is a bad argument against autoconf if you use any other shell script in your source tree, because you are just as likely to introduce nonportabilities there as in the configure.ac (and they would be just as hard to detect). The portability or otherwise of C is irrelevant to this argument: your objection is a non sequitur.

configure scripts depend on the shell environment in any case, so allowing you to introduce your own shell fragments introduces no new dependencies, which was very important to the Autoconf designers. (Again, the manual *says* all this. This is not secret hidden knowledge.)

Kamp: A Generation Lost in the Bazaar

Posted Aug 22, 2012 18:27 UTC (Wed) by cmccabe (guest, #60281) [Link]

> However, I note that makefiles are themselves little more than
> dependencies tied to shell script fragments, so clearly the shell and make
> are relatively tightly associated (indeed, make knows internally what
> characters are common shell metacharacters, and stuff like that).

Again, this is not a binary choice between plain old Makefiles and autotools. My choice is CMake.

CMake has its own scripting language which is portable across platforms (including Windows), versioned, and specifically designed for this use-case. Using the policy mechanism you can tell CMake "act as if you were CMake 2.6." Anyone with a version of CMake higher than or equal to 2.6 can then compile your project exactly as you intended. Shell scripts don't have any of these things.

Yes, shell scripts are appropriate for some use cases. But a build system is just not one of them. Let's use the right tool for the job rather than trying to explain how with a little duct tape, a plunger can be used as a hairbrush.

At the end of the day, nothing can make your code portable but you. But the build system can help you or hurt you. If it closes off platforms to you, like autotools closes off Windows to you, that hurts. If it forces you to write shell scripts to do everything, that hurts portability too. I've spent a long time fixing build breakages that resulted from autotools' poor choices. The question of what kind of Makefile.am or configure.ac a superintelligent Zen master who has studied AIX, IRIX, csh, tcsh, bash, zsh, dash, and every other shell out there might write is not interesting to me. In the real world, autotools = not portable.

Kamp: A Generation Lost in the Bazaar

Posted Aug 22, 2012 19:21 UTC (Wed) by hummassa (subscriber, #307) [Link]

That is really not my experience. I prefer CMake to autotools for new code, but CMake was not really crosscompiler-friendly the last time I checked. Autotools is really much more complicated, but it will even work in MinGW/CygWin without a lot of problems... and it will suport crosscompiling without a lot of EXTRA effort. Yes, it is possible that CMake works well with MSVC, but I wouldn't know about it.

Kamp: A Generation Lost in the Bazaar

Posted Aug 22, 2012 20:17 UTC (Wed) by boudewijn (subscriber, #14185) [Link]

I know that it does work well with msvc. I also know that it works well when cross-compiling from linux to windows. Maybe the setup is a bit more involved than with autotools, maybe it's not -- maybe the people feel it is more involved just don't remember the effort they expended on the autotools setup when they had to work with it for the first time. And yes, cmake works very well with mingw on windows itself.

Really, people who have an opinion on cmake and autotools in relation to windows, whether it's with mingw, msvc, or icc, but haven't checked out the way the kde-windows project uses cmake should check it out. Doing that will make their life _much_ easier.

Kamp: A Generation Lost in the Bazaar

Posted Aug 24, 2012 1:16 UTC (Fri) by cmccabe (guest, #60281) [Link]

CMake has supported cross-compiling for a long time. http://www.vtk.org/Wiki/CMake_Cross_Compiling

Various projects may have trouble with cross-compiling, because their authors did something problematic (like assuming that they could build a binary with the target compiler and then run it, or using hard-coded paths in a dumb way). However, this is just as true with autotools. Fix the bugs and move on.