Conan: linking problems with start-group and end-group

Created on 8 Apr 2019 · 7Comments · Source: conan-io/conan

Hello,

Take a look a the abseil's conanfile.py here:
https://github.com/abseil/abseil-cpp/blob/master/conanfile.py

Its package info looks like this:

        if self.settings.os == "Linux":
            self.cpp_info.libs = ["-Wl,--start-group"]
        self.cpp_info.libs.extend(tools.collect_libs(self))
        if self.settings.os == "Linux":
            self.cpp_info.libs.extend(["-Wl,--end-group", "pthread"])

Now, I am packaging my codebase with conan as well and rely on start-group/end-group to remove circular dependencies amongst my libs. My package info looks like this:

            self.cpp_info.libs = ["-Wl,--start-group"]
            self.cpp_info.libs.extend(['foo', 'bar'])
            self.cpp_info.libs.extend(["-Wl,--end-group"])

Now here is the surprising bit. The final linker command removes start-group/end-group from my libraries if I also depend on abseil! (If I do not depend on abseil, then everything works fine).

linker command looks like this:

... foo.a bar.a -Wl,--start-group absl.a -Wl,--end-group -lpthread -Wl,--start-group -Wl,--end-group

As you can see, it created an empty start/end group at the end, and didn't encapsulate my libs in it. My suspicion is that since everything is a list of strings, GNU Linker (or cmake) changes the ordering since it thinks that start-group/end-group are just few more libs. Not sure what is the best course of action here, but any help is appreciated. Thanks!

triaging question

Source

skgbanga

All 7 comments

If you run twice self.cpp_info.libs = xxx you are overwritting the values, so use self.cpp_info.lib.extend(["foo", "bar"]).

Said that including using like that -Wl,--start-group and -Wl,--end-group in the libs list sounds awful and should be avoided if possible. I really don't understand how it could work, honestly.

lasote on 8 Apr 2019

@lasote Apologies, I wrongly typed that. Corrected the original question.

I agree with you that start/end group is awful. I raised this in #2387 (https://github.com/conan-io/conan/issues/2387#issuecomment-474156925).

Right now, cmake knows the dependency between my archives, but there is no way to export that to conan. So as a roundabout, I added start/end group to future-proof myself in case I add more libraries.

skgbanga on 8 Apr 2019

Thanks. So, to be honest, I don't know what to do with this issue, I'm not surprised that using these groups fail somehow, internally Conan is considering each element as a library name, adding -l prefixes and so on, so I can bet some awful situation is happening there.

Is it not possible for you to just declare the right order in the library list? I talked with @uilianries and he is going to suggest to do that for the abseil recipe.

lasote on 9 Apr 2019

@lasote that is what I am currently doing. (i.e. specifying the order manually).

skgbanga on 9 Apr 2019

The bast we can do is forcing the order list by cpp_info.libs and set a specific version. Since Conan 1.14, the method tools.collect_libs sorts the libraries by name, which doesn't work for Abseil case.