News for hackers. Stuff that matters.
Re:LearnCPP: Website devoted to teaching you how to program ... by galangalalgol on Tuesday January 03, 2023 @07:44AM (#34230692)
For the build system, I like: http://cliutils.gitlab.io/modern-cmake/

In addition to building, it will show how to add cppcheck and clang-tidy to your cmake files. That part is pretty easy. Just don't turn on every check in clang-tidy. Many of them are for specific code bases. Modernization checks are my favorite ones. Keeps you doing stuff the modern way.

My link has you doing dependency management with git submodules. That is a method still in use and even advocated for, but conan is often touted as the more modern way to do it. I've used it on some teams, and I'm undecided. I also have no good resources for it.

Re:LearnCPP: Website devoted to teaching you how to program ... by geokon on Tuesday January 03, 2023 @10:12AM (#34232338)
It's been quite a few years since I've really written any C++ and I'm surprised the landscape hasn't evolved at all. People are making the same mistakes...

Modern CMake should be using toolchain files to specify the toolchain/flags. That should be included in ever CMake intro. (Hacking those into your CMakeLists.txt is not okay)

gitsubmodules are also a hack

1: Many dependencies are not setup to be used as a CMake subdirectory. Variables such as the project-version-number get overwritten by the parent project and more dangerously the submodule can start to change variables in the overall project.

2: Dependencies will often reuse target names. For instance, libraries often have a target called `uninstall`. Because redeclaring a target in a different subdirectory is a no-no - CMake crashes

Proper dependency management can be done directly through CMake without repeating yourself by using Hunter: https://hunter.readthedocs.io/

This solved the diamond problem, gives you proper namespaces in CMake and with Polly you can have sane toolchain files (https://github.com/ruslo/polly - but you can also just write them manually). Everything is built on CMake and Git (and git hashes). You get nice dependency build caches and forking dependencies is also trivial

EDIT: The creator of Hunter has his own CMake Intro. I haven't gone through it while-learning, but it looks comprehensive and informative: https://cgold.readthedocs.io/en/latest/

Re:LearnCPP: Website devoted to teaching you how to program ... by jcelerier on Tuesday January 03, 2023 @10:57AM (#34233080)
> Modern CMake should be using toolchain files to specify the toolchain/flags. That should be included in ever CMake intro. (Hacking those into your CMakeLists.txt is not okay)

this so much. I made a toolchain generator to make myself toolchains with common set of flags (asan, lto, etc specified with a single simple word on the command line for each "feature" or at worst a "key=value" pair, e.g. "compiler=gcc", "linker=lld") and really it made my life so much better: https://github.com/jcelerier/cninja

Re:LearnCPP: Website devoted to teaching you how to program ... by galangalalgol on Tuesday January 03, 2023 @12:20PM (#34234432)
Git submodules are definitely a hack. My experience with conan has been unpleasant for several reasons, but mostly in regards to writing dependencies for code bases that use submodules and aren't interested in my team's suggestion they use conan. Vcpkg doesn't seem better, and probably a bit worse. Does hunter make this any easier?
Re:LearnCPP: Website devoted to teaching you how to program ... by geokon on Tuesday January 03, 2023 @12:53PM (#34235016)
I'm not super clear what your question is.

If the dependency you want to use has no dependencies of its own - then it can be used directly with no changes (as long it's CMakeLists.txt is sane and it can use a passed-in toolchain file)

https://hunter.readthedocs.io/en/latest/creating-new/create/...

If the dependency has its own dependencies, then you need to "Hunterize" it - specify library versions, tweak the targets to use namespaces, etc. They're usually small changes. They have a huge list of Hunterized forks:

https://hunter.readthedocs.io/en/latest/packages.html

Finally, a CMakeLists.txt that's been made to work with Hunter can always revert back and be run without Hunter. So the change isn't intrusive. https://hunter.readthedocs.io/en/latest/overview/compatibili...

A minimal CMakelists.txt will look something like:

    HunterGate(
        URL "https://github.com/cpp-pm/hunter/archive/v0.23.297.tar.gz"
        SHA1 "3319fe6a3b08090df7df98dee75134d68e2ef5a3"
    )
    project(Foo)
    hunter_add_package(Boost COMPONENTS regex system filesystem)
    find_package(Boost CONFIG REQUIRED regex system filesystem)
    add_executable(foo foo.cpp)
    target_link_libraries(foo PUBLIC Boost::regex Boost::system Boost::filesystem)
The first block specifies the available package list (you can fork and host your own). If you disable Hunter this isn't used

The line that reads `hunter_add_package` is the Hunter part that downloads and build the dependency (using the current toolchain file). But if you disable Hunter then that line would simply get skipped.

In the third line `find_package` will run.. and if you had Hunter build the dependency then it'd use that as the target - otherwise it just runs like "normal CMake" and it tries to find the package elsewhere (like from your system package manager or whatever)

If your dependency has its own dependencies as git submodules.. then I'm not super sure you have a very clean migration path b/c they will not be using `find_package` and will instead be using `add_subdirectory`. Maybe you could patch their CMakeLists.txt to use Hunter when explicitely enabled, and then drop down to `add_subdirectory` otherwise. It'd be a bit ugly but they would keep their workflow I guess. You'd need to check that their targets end up with the same namespaced names (with the :: notation) when seen from your project... I don't know off the top of my head how Hunter handles a subdirectory with a project name of it's own

Re:LearnCPP: Website devoted to teaching you how to program ... by galangalalgol on Tuesday January 03, 2023 @07:09PM (#34239989)
Sorry I wasn't clear. I was responsible for the dependency that another team was using as a plugin. They weren't going to switch to conan, and our code was already using Conan for itself and its dependencies. They wanted to bring us in as a submodule, and that was just never going to work. I think what I'm pulling out of this thread is that c++ package management is so fragmented, there isn't really a safe choice to ensure you can use or be used by people who made other choices.
Re:LearnCPP: Website devoted to teaching you how to program ... by geokon on Wednesday January 04, 2023 @02:09AM (#34242916)
Oh okay - I think in that case it's possible Hunter would solve your problems. If you rewrite your library to use Hunter, then when it gets pulled in with add_subdirectory(), it should just build and get it's dependencies with Hunter - all from within CMake itself. No external tool necessary (the Hunter functions are defined in the project root `.cmake` file)

I don't see any reason that wouldn't work - but.. you'd have to test it. I've never tried running a Hunterless top-level project with a Hunter-ized subdirectory. The top level project would still get some benefits like the namespaced targets (so you won't clobber anything in top-level)

The only small catch is that Hunter inevitably does build a cache of build artifacts/targets for your dependencies. As long as you keep using the same toolchain files these keep getting reused each time you rebuild your project. That does however mean you will get some user folder like ~/.hunter with those build artifacts (look at the Hunter config options.. maybe this can be somehow hidden in the build directory?). I dunno if that would irritate the project users. If it's company internal that's probably an okay ask. If it's random devs on github then that's a bit not-nice :)

Re:LearnCPP: Website devoted to teaching you how to program ... by galangalalgol on Wednesday January 04, 2023 @11:26AM (#34247600)
Thanks! That looks like it might be fine. The conan cache works the same way but all our builds and deployments are containerized so the ~/.conan wasn't the issue, it was that conan was in control, not cmake.
Re:LearnCPP: Website devoted to teaching you how to program ... by dongping on Tuesday January 03, 2023 @10:58AM (#34233094)
Honestly, I find the usage of the adjective "modern" deceptive and insincere in the marketing term "Modern CMake". The "modern" scripting language doesn't even support return values and the variables lack basic type-safety. It breaks for anything beyond a simple Hello World.

I would recommend any other C++ build tools that use a proper programming language for scripting, be it Meson or Bazel.

Re:LearnCPP: Website devoted to teaching you how to program ... by galangalalgol on Tuesday January 03, 2023 @12:14PM (#34234334)
I dislike cmake, so when a sibling post mentioned meson I went looking. It doesn't seem to have good support for cuda, and neither does bazel, to the point I don't immediately find anyone getting it to work. These days the only reason I start a project at work in c++ instead of rust is to use cuda, so that would be a strong consideration. Is this easier than my brief glance would indicate?
Re:LearnCPP: Website devoted to teaching you how to program ... by dongping on Tuesday January 03, 2023 @12:29PM (#34234616)
I have yet to encounter problems with Cuda in our Bazel environment, but I don't have to maintain the Bazel rules, since our build infra team does that.

I would think that rules like https://github.com/liuliu/rules_cuda should work for you in Bazel. (Tensorflow also uses Bazel, so I don't think that this should be a problem.)