> On Oct. 14, 2014, 9:06 p.m., Timothy St. Clair wrote: > > configure.ac, line 281 > > <https://reviews.apache.org/r/26426/diff/1/?file=714874#file714874line281> > > > > Is there a reason you want to leave debug symbols out of optimized > > builds? > > > > cmake has the pattern correct imho: > > Release > > Debug > > ReleaseWithDebug > > > > A ReleaseWithDebug allows packagers, such as myself, to build > > w/debugsymbols that are stripped out into a .debuginfo package which can be > > used by developers for tracing "When bears attack". Granted that it is > > tenuous debugging at best, but it's better then nothing. > > > > So I think we want all three modes, stripping all debug information is > > not really idea. > > Cody Maloney wrote: > My main motivation is to shrink the size of libmesos. Yesterday sugis in > #mesos had one which was 213M. For the buildbot internally, full debian > packages (which are compressed) of mesos weign in at 165M a piece (Yes, > stripping post-build would help a lot, but why build it to begin with?). Most > of this is debug info. Also, we build a bunch of different ways, and when > libmesos is as big as it is, a decent amount of time ends up being spent on > disk I/O reading / writing all the debug info when we are really just trying > to ensure it builds on all the different platforms (Not to mention storage > and file size shipping things around the network to centralized repositories). > > The simple toggle between debug and release, removing the legacy logic > gets us most of the benefit. > > If you have a good place to point me for what is needed to get a > 'ReleaseWithDebug' info build up and running I can definitely work on adding > that as well. > > Cody Maloney wrote: > It is also entirely possible to specify custom CXXFLAGS + CFLAGS with > this patch to get the old "optimized debug" build. The flags set by > --enable-debug (or not having it), to get back the "optimized debug" build > from before. > > Timothy St. Clair wrote: > I suppose rpm builds default CXXFLAGS to '-O2 -g ...', so I can buy the > argument of just making it easier on the average person. > > Ben Mahler wrote: > Re-opening because conflating debug and optimization under a single flag > seems a bit confusing to me. > > In practice, we run mesos **with optimizations** (for free performance > win), and **with debugging symbols** (for meaningful backtraces and core > dumps), which was previously the default but now requires the setting of > CXXFLAGS to obtain. :( > > For development, I think by default we need debugging symbols turned on. > Otherwise most people will forget to configure with `--enable-debug` and > consequently will need to recompile everything when they encounter a > backtrace or need to debug. CI jobs will encounter this issue as well. > > If there are use cases that merit no debugging symbols, seems nice to > make that explicit given it runs the risk of making debugging difficult. > > Thoughts? > > Cody Maloney wrote: > Note that the two were conflated under one flag previously. This just > changes how we do it to be a little more standard way of conflating them. > Generally people talk about `Debug` or `Relase` builds. Not the specific > compiler flags that get switched. The general switch is -O2 -> -O0 + -g and > back (Possibly with a couple extra warning flags). Autotools projects which > implement a debug vs. release build almost always do '--enable-debug' / the > default is 'optimized' build. > > https://gcc.gnu.org/wiki/DebugFission - that artical has some good > explanations why debug info gets problematic as programs get larger and > larger, which is only going to get worse over time. Already mesos debug info > is over 100MB. As the codebase grows, this will get worse than it already is, > increasing the minimum machine specs you need to develop / compile Mesos. > > Compiler implementors generally give that '-O2' should be used to > generate an optimal binary. '-g' should be used when you are planning to > attach GDB and step through a program line by line. General development > generally people want nice backtraces, which is inbetween those. See > **Backtraces**. The section **Coredumps** talks about how we can give all the > info in coredumps without running full debug builds ('-g') everywhere. > > **Backtraces** should generally be reasonably meaningful as long as you > don't strip the binary. Most of the symbols are still there (You just lose > inlined things). You don't get file names and line numbers, but all the > functions in mesos should be uniquely named (C++'s ODR rule), so you get out > a mangled C++ name which you can demangle with `c++filt` (Shipped with GCC) > and you will get the full namespace prefixed name of the C++ function. Find > the function with your favorite text editor search, life is happy. > > Pretty much every binary which is shipped on your machine via a > distribution is sent like this. The backtraces + coredumps that are gathered > come from this. > > [Clang has a option > -gline-tables-only](http://clang.llvm.org/docs/UsersManual.html#cmdoption-gline-tables-only) > which adds just the file name and line numbers to all functions in the debug > info (Which can be done with minimal space overhead. [Gcc has > '-g1'](https://gcc.gnu.org/onlinedocs/gcc/Debugging-Options.html#Debugging-Options), > which would mean that you would get line number + file names in back traces > with less overhead. > > Those options would make it so what people see printed by the build bot > is exactly what they see now. I think that should be an easily accessible > level, but not default build. > > > **Coredumps**: It is possible to make coredumps just as inspectable > without including debug information in all the shipping binaries. In fact, > you generally probably don't want to run gdb on the binary on the production > host anyways. For this, what typically happens is that > > 1. People actually only look at the backtrace to figure out what went > wrong > 2. You can generate debug information which lives in seperate files from > the target executable (-gsplit-dwarf on new versions of GCC, requires a > relatively new toolchain. I believe there is a way to do it on older tooling > where you split it out later, but am not familiar with it). This debug info > would generally live in a '-dbg' package. Which you can install / run on your > local machine to inspect whatever variables remain directly, as well as > global state. > > > **OVERALL** > I don't see anywhere where this makes debugging considerably more > difficult than it is already. Yes, you can't attach a GDB to any build of > mesos without thinking about it. I think most of the time most developers > aren't attaching GDB. > > CI jobs I think can be easily switched over where they care. There aren't > that many of them around. I agree it is something people need to be aware of, > but not a hard level of change. Most CI systems don't let people connect in > and step through the program with GDB to where it failed a test. It just > gives a backtrace. In the case of a buildbot it isn't hard to set CXXFLAGS to > just the level you want to get that ('-g1'). No major loss there with this > change. > > The only case this makes considerably more difficult is that if you want > to attach GDB to inspect variabels you need to have specified --enable-debug > at configure time. I don't think that is a major loss, or something which is > in the standard compile/test loop of most most Mesos contributors / > developers (Although we can collect evidence if desired) > > Ben Mahler wrote: > Thanks for being thorough! Having `-g1` and `-gline-tables-only` sound > good to me for release builds, any reason not to include those? > > As for local development, I would _personally_ be fine with optimization > turned off and `-g1` and `-gline-tables-only`, if the smaller debug > information has a measurable improvement on compile speed or memory > consumption (let's measure!). I tend to only need backtrace, or at most gdb > thread backtraces from a dump. > > These two use-cases seem to suggest leaving the existing > `--disable-optimize` and merely tweaking the debug flags to avoid dumping all > debug information into release builds. > How about for this change, we tackle the following: > > Default is Release: -O2 -g1 (or -gline-tables-only) > --disable-optimize: -O0 -g1 (or -gline-tables-only) > > In a subsequent patch, we can add an orthogonal debug flag for those who > want to the full set of debug information: > > --enable-debug: turns on -g > > Thoughts? It seems we can acheive Debug / Release builds even if we don't > map the distinction to a single binary configure flag. > > Cody Maloney wrote: > I'm going to run some tests: > CXXFLAGS=-O2 > CXXFLAGS=-O0 -g1 > CXXFLAGS=-O0 -g > CXXFLAGS=-O2 -g > > I'm going to measure compile time resulting size, of the build directory, > and size of the install directory. Will provide some feedback after that. > > Note '-g' and '-g1' probably override eachother (they are actually the > same flag), So if we want to do that, we need to make it swap that flag which > will take a little bit more complex logic. The simplest/cleanest thing to do > if people want a debug is to say 'Specify your own CFLAGS'
Ran the tests, results posted to mesos dev mailing list: http://www.mail-archive.com/[email protected]/msg22097.html I think we should use just -O0 for default dev build, don't touch CXXFLAGS at all if they are passed (which means we get the right behavior in packaging systems for various distros). I can add a --optimize flag to set CXXFLAGS to -O2. I can add a --debug flag to set CXXFLAGS to -g2. Those are useful if you are doing a full build and then afterwards turning it into a distro package / tarball which you then ship around (If you use a distro pacakging system, those specify the packaging guideliene flags to configure via environment variables) I think people are better off specifying those manually then having flags which wrap them in. Those flags aren't independent toggles how autotools projects are setup, so doing it right is really really hard / there will be several edge cases just not properly handled when people hand-specify some CXXFLAGS. In the longer run a different build system (CMake?) would be the cleaner if we want explicit specification. - Cody ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/26426/#review56573 ----------------------------------------------------------- On Oct. 14, 2014, 11:07 p.m., Cody Maloney wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/26426/ > ----------------------------------------------------------- > > (Updated Oct. 14, 2014, 11:07 p.m.) > > > Review request for mesos, Benjamin Hindman and Timothy St. Clair. > > > Repository: mesos-git > > > Description > ------- > > Reworks building mesos in "debug" vs. "release". By default, mesos is now > built in release (no debug info, optimized build). If '--enable-debug' is > specified to configure, than optimization will be turned off, and debug info > will be turned on. > > This also adds a variable 'DEBUG' to the build environment, which people can > use in code to see if mesos is built with debugging to enable extra > assertions / checks. For release builds we may want to set 'NDEBUG' which > removes assert()'s, but that is a seperate discussion. > > Main benefits: > 1) Getting a build to include/exclude debug information at will is feasible. > Before some things like using clang would forcibly enable debug info in all > cases > 2) libmesos.so and the other binaries which get packaged up for use in > distributions shrink considerably without manually stripping post-build > (Improves build time, makes packaging cleaner) > > > Diffs > ----- > > configure.ac 2b372e06006250b5230956ef096473e98f3fa590 > > Diff: https://reviews.apache.org/r/26426/diff/ > > > Testing > ------- > > Built with both --enable-debug and without, checking that the flags get > passed through correctly. > > > Thanks, > > Cody Maloney > >
