From mboxrd@z Thu Jan 1 00:00:00 1970 From: Thomas Petazzoni Date: Tue, 28 Sep 2010 22:36:52 +0200 Subject: [Buildroot] Libtool work: a tentative summary Message-ID: <20100928223652.6dc72083@surf> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: buildroot@busybox.net Hello, There has recently been quite some work done around libtool handling in Buildroot, in the form of two proposals done by Martin Banky and Lionel Landwerlin. Those two proposals are quite different and we need to make a choice between them. This e-mail tries to summarize the libtool problem, how it is currently handled in Buildroot, the technical proposals made by Martin and Lionel with their respective advantages and drawbacks. Martin and Lionel are obviously invited to comment on those (and the rest of this e-mail as well). Libtool problems ================ Libtool problems in cross-compilation mode are well-known. The famous http://www.metastatic.org/text/libtool.html lists most of the problems, and some hacks to solve them. Basically, Libtool does not understand the fact that we may be building a root filesystem in a different place than "/", and keeps looking in /usr/lib for libraries. It is worth noting that libtool 2.4 has finally introduced a --sysroot argument which should finally solve those problems. However, until libtool 2.4 is picked up by most packages, we will have to live with the libtool problems for quite some time. See http://lists.gnu.org/archive/html/libtool/2010-09/msg00082.html. Current Buildroot solution ========================== The current Buildroot solution consists in two mechanisms : *) Apply the package/buildroot-libtool.patch to packages when _LIBTOOL_PATCH is YES. This is the default value for autotools-based packages. This patch directly modifies the ltmain.sh script, which will be used to generate the libtool script in the package as configure time. This patch only works with libtool 1.5.x, but a number of packages are now using libtool 2.2.x, so more and more packages are doing _LIBTOOL_PATCH=NO to disable the patch. This works for most packages but not for others (example: I'm bumping GTK to 2.20 and it needs a patch for the ltmain.sh script which has been generated by libtool 2.2.x) *) Modify the .la files installed by Buildroot in order to add $(STAGING_DIR) at the beginning of the "libdir" paths. Martin Banky proposal ===================== Martin Banky has sent a proposal to improve libtool handling in Buildroot on September 16th. See http://lists.busybox.net/pipermail/buildroot/2010-September/037505.html and the following 4 e-mails from Martin. Basically, Martin's proposal is to continue with the current solution to the libtool problem, but to clean it up a bit and to extend it to libtool 2.2. This patch set has 3 main components : *) Simplify the existing package/buildroot-libtool.patch by removing useless chunks *) Rename package/buildroot-libtool.patch to package/buildroot-libtool-v1.5.patch, add a new package/buildroot-libtool-v2.2.patch. Those two patches respectively modify the ltmain.sh generated by libtool 1.5 and 2.2. Then the Autotools infrastructure is modified to apply one patch or the other depending on which libtool version the package is using. *) The remaining modifications remove libtool-related patches in various packages that have become useless thanks to the previous patches. Advantages of this approach : *) We're keeping the current solution and only cleanup/extend it to work with recent libtool versions. This is good from a "conservative" point of view. *) As today, we don't need to autoreconf all autotools-based packages just to fix the libtool issue. Drawbacks of this approach : *) We may have similar issues with future versions of libtool, even if the addition of --sysroot to libtool 2.4 probably means that libtool 1.5 and 2.2 will remain the only two versions of libtool causing problems in widespread use. *) It keeps the current solution, which is a bit ugly: modify scripts that are generated, and hack the .la files. *) It modifies the .la by adding informations relative to $(STAGING_DIR), which may cause problems if we want to generate packages (.ipkg) or if we would like to allow users to relocate the Buildroot staging directory. Lionel Landwerlin proposal ========================== Lionel Landwerlin has sent a proposal to improve libtool handling in Buildroot on September, 18th. See http://lists.busybox.net/pipermail/buildroot/2010-September/037663.html for the mailing list post. Basically, Lionel's proposal is to patch libtool itself to make it sysroot-capable and then autoreconfigure all autotools-based packages to make them use the fixed libtool version. Lionel's proposal is relatively complicated because in order for the patch on libtool to take effect, libtool itself needs to be autoreconfigured, which in turn requires libtool. So Lionel has introduced a host-libtool-host package, which is an unmodified libtool, installed in a new $(HOST_HOST_DIR) directory. He later uses that host-libtool-host to autoreconfigure the host-libtool (which has been previously patched to understand sysroot related things). host-libtool-host will be kept for the autoreconfiguration of host packages, while host-libtool is used for the autoreconfiguration of target packages. (If your head hasn't exploded yet, keep reading, otherwise, take a break and come back in 10 minutes). This is what the first patch in Lionel patch set does. The other patches are only minor related changes. Advantages of this approach : *) We no longer arbitrarly patch ltmain.sh in every package, with the need to carry libtool-version-specific patches. *) We no longer need to modify the .la files. *) The approach seems cleaner. *) It is closer to what we could have with libtool 2.4, except that libtool 2.4 wouldn't need to be patch as it already supports --sysroot by default. Therefore, a large part of the complexity of this patch set could be removed (since we wouldn't have to autoreconfigure libtool itself). However, we'd still need to autoreconfigure all autotools-based packages (see below). I know Lionel has started experimenting with libtool 2.4, as can be seen on http://lists.gnu.org/archive/html/bug-libtool/2010-09/msg00064.html Drawbacks of this approach : *) We need to autoreconfigure *ALL* autotools-based packages. This may be an issue for two reasons: 1/ speed, 2/ packages for which autoreconfiguration does not work properly (yes those packages should be fixed, but anyway, this may add some work to us). Obviously, as packages will start using libtool 2.4 by default, this shouldn't be a problem, but we all know that it's going to take years before everybody moves to that newer libtool version. *) The HOST_HOST_DIR thing and the host-libtool-host as it is implemented today is pretty ugly. I'd rather have a either single libtool installed in $(HOST_DIR), which work for both the autoreconfiguration of host and target packages, or have two libtools installed in $(HOST_DIR), on for host packages (just named 'libtool') and another for target packages (named 'cross-libtool' or something similar). Conclusion ========== Therefore, the choice seems to be between : * A conservative solution, only extending what we have today, with limited changes and impact, but keeping the existing hacks in place. * A radically different solution, with wider changes, more in line with libtool future, but requiring an autoreconfiguration of all packages that will not have libtool 2.4 by default. In my opinion, we should settle on a solution before mid-october, in order to merge it before the end of october and give us enough time for testing before the release at the end of november. Considering how much time Martin and Lionel have dedicated to this issue, I think that their respective work deserve some attention. Thoughts ? Comments ? Thomas -- Thomas Petazzoni, Free Electrons Kernel, drivers, real-time and embedded Linux development, consulting, training and support. http://free-electrons.com