[BUG] -Werror=catch-value made warnings as errors, then build failed.

[kaigai]

Describe the bug

During the cuDF module build from the source code, it generates terrible amount of compilation errors, like:

/home/kaigai/anaconda3/envs/cudf_dev/include/rmm/rmm.hpp: In instantiation of 'rmmError_t rmm::alloc(T**, size_t, cudaStream_t, const char*, unsigned int) [with T = char; size_t = long unsigned int; cudaStream_t = CUstream_st*]':
/home/kaigai/repo/cudf/cpp/custrings/strings/../regex/../util.inl:96:29:   required from 'T* device_alloc(size_t, cudaStream_t) [with T = char; size_t = long unsigned int; cudaStream_t = CUstream_st*]'
/home/kaigai/repo/cudf/cpp/custrings/strings/replace_backref.cu:155:93:   required from here
/home/kaigai/anaconda3/envs/cudf_dev/include/rmm/rmm.hpp:139:3: error: catching polymorphic type 'class std::exception' by value [-Werror=catch-value=]
   } catch(std::exception e) {

Although I could not check all the error message (build script had been generated error message for 20 minutes...), it looks to me c/c++ compiler detected a warning related to catch exceptions, and a compiler option: -Werror=catch-value treated it as error.

Steps/Code to reproduce bug

https://github.com/rapidsai/cudf/blob/branch-0.12/CONTRIBUTING.md#script-to-build-cudf-from-source

We followed the steps described above.
make -j kicks a series of build chains, then raised this problem.

Environment overview (please complete the following information)

• OS: CentOS 8 with latest packages
• gcc-8.2.1-3.5.el8.x86_64
• cuda-nvcc-10-1-10.1.243-1.x86_64
• cmake-3.14.5-1.fc29.x86_64
  • RHEL8/CentOS8 distributed cmake-3.11, so we picked up from Fedora29
• boost-devel-1.66.0-6.el8.x86_64
• boost169-devel-1.69.0-4.el8.x86_64

Additional comment

I have been trying to build/install cuDF packages for more than a week, however, it does not work well.
Is it possible to distribute the RPM packages for RHEL/CentOS installation bases?

[harrism]

This looks like an RMM issue rather than a libcudf issue.

[kaigai]

OK, I try to build https://github.com/rapidsai/rmm individually.
Let me report later.

[harrism]

Even if it builds OK, it's in a header so I think we should fix it (it's a 1-character fix. :). If you can file the same issue text above as an RMM issue that would help us assign and track it.

[harrism]

Fixed by rapidsai/rmm#215

[trevorsm7]

I think this issue is blocked on NVIDIA/jitify#36 as well as a few changes in cudf

[trevorsm7]

I checked a similar fix into NVIDIA/jitify, but it needs to be merged into the cudf branch: NVIDIA/jitify#37

[harrism]

@trevorsm7 Can you submit a PR to the cudf fork of Jitify to merge the latest NVIDIA/jitify?

[trevorsm7]

@trxcllnt suggested waiting on the binops PR to merge jitify (cudf is currently ~30 commits behind)