XL: Support linking objects compiled with clang++ and xlc or xlf that use OpenMP GPU offloading

SalesForce Case TS001968579
PMR 82483,227,000

XLF compiler: Errors linking objects compiled with clang++ and xlc or xlf that use OpenMP GPU offloading.

LLNL is having problems linking objects that were compiled using the clang++ (LLVM) and XLF (or XLC) compilers. Errors occur during the link process when either clang++ or xlf is used to perform the link.

John Gyllenhaal, from LLNL, has already reached out to the clang compiler developers Doru Bercea <gheorghe-teod.bercea@ibm.com> and Bardia Mahjour <bmahjour@ca.ibm.com> and CC'd Wang Chen <wdchen@ca.ibm.com>, the XL compiler release manager.

We can compile our host-only OpenMP c++ with clang++ (either the IBM trunk version or the clang 3.8 version) and our GPU OpenMP FORTRAN with xlf if we link with xlf, but we run into problems as soon as we add GPU OpenMP into the c++ or try to link with clang++.

Is it possible to link xlf GPU code and clang++ GPU code in the future (or now if there is a magic formula)? Extra steps are ok that does the GPU linking per language, for example.

The same errors with code tests using a small saxpy kernel whether I mix XLC or XLF compiled OpenMP4.5 kernels with CLANG OpenMP4.5 kernels. It's not FORTRAN specific.

My first guess was maybe the libomptarget libraries installed in the CLANG vs XL compiler areas are using some incompatible header versions ( like omptarget-nvptx.h ). One of the errors is NVLINK complaining that the omptarget_nvptx_device_State is a different size than expected.