pocl requires fork()

[Andreas Klöckner]

pocl currently seems to assume that it's safe to fork other processes during kernel build. Unfortunately, in the high-performance compute use case, that's not true. In particular, MPI processes with Infiniband buffers mapped into their address space do not seem to take forking very well, often leading to a crash.

[Andreas Klöckner (inform)]

Obvious follow-up remakr: It would be ideal if the entire compilation could be done in-process, e.g. using clang as a shared library, if that's possible.

[Pekka Jääskeläinen (pekka-jaaskelainen)]

Yes, it should be possible and actually was discussed recently in #pocl as a way to get rid of the requirement of pocl scripts in PATH. Getting rid of the scripts would make the code base more easily portable e.g. to Windows. Clang and LLVM both should have library APIs we need.

[Kalle Raiskila (kraiskil)]

lp:~kraiskil/+junk/pocl_llvm_api
Just a bruteforce removal of one system() call so far.