Package recipes using modern CMake (required 3.8)

[Bo-Yuan-Huang]

• Target-based recipe for in-source embedding
• Exporting the package (coreir library)

[rdaly525]

@cdonovick, you are a bit more CMAKE savvy than I am. Could you do a review?

[Kuree]

This will 100% break pycoreir...

[rdaly525]

@Kuree, what issues do you foresee?

[Kuree]

Per PEP 513, the binary wheel has to satisfy the following conditions:

GLIBC <= 2.5
CXXABI <= 3.4.8
GLIBCXX <= 3.4.9

Due to the minimum requirement of gcc-4.9, we have to statically link the glibc++. This new PR removes the hacks that allow static linking in pycoreir.

[Bo-Yuan-Huang]

Not an expert on Python binding, but it should be fine if it is about building a static library.

https://github.com/Bo-Yuan-Huang/coreir/blob/e12b74d659a479e4763de50afe70132d2d7100a9/CMakeLists.txt#L23
One can enable static linking by turning the standard CMake build-in option ‘BUILD_SHARED_LIBS’ to OFF/FALSE.

@Kuree, is there other issue that will break the binding? I’d be glad to help!

[Kuree]

Unfortunately that won't cut it. As I mentioned earlier, you need to statically link against glibc and glibc++ to do so. You also need to be careful about the D_GLIBCXX_USE_CXX11_ABI flag. In my experience you need to turn that off to be compatible with modern C++ compilers.

One way to try out is to follow the build steps in pycoreir here. If you can produce a manylinux tagged wheel it means you're good to go. You need to run that in a docker container that's been modified from the official manylinux container. I think the easiest way is to fork pycoreir and change the remote where it clones coreir from and let it run in travis. It can be difficult to debug inside travis though.

Please let me know if you have any problems setting up the pycoreir build environment.