OpenGL(mesa)+OpenCL applications segfault due to multiple LLVMs

[kraiskil]

This is one of the problems already recorded in issue #44 and partially fixed in d86c335.

When an application links against mesa and pocl (even via ICD), and the mesa backend used is LLVMpipe, mesa will jump to the LLVM symbols introduced when LLVM was linked to pocl.

Commit d86c335, added -Bsymbolic flag to linker (on GNU systems). This fixed the inverse issue, i.e. libpocl jumped to the LLVM symbols found in mesa.

Sample backtrace:
#0 0x00007ffff6079475 in *__GI_raise (sig=) at ../nptl/sysdeps/unix/sysv/linux/raise.c:64
#1 0x00007ffff607c6f0 in *__GI_abort () at abort.c:92
#2 0x00007ffff6072621 in *__GI___assert_fail (assertion=0x7ffff1923ca0 "findOption(Name) == Values.size() && "Option already exists!"", file=, line=675, function=0x7

ffff1f77660 "void llvm::cl::parser::addLiteralOption(const char_, const DT&, const char_) [with DT = llvm::ScheduleDAGSDNodes* ()(llvm::SelectionDAGISel, llvm::CodeGenOpt:
:Level); DataType = llvm::Sche"...) at assert.c:81
#3 0x00007ffff07cab7c in void llvm::cl::parser<llvm::ScheduleDAGSDNodes* ()(llvm::SelectionDAGISel, llvm::CodeGenOpt::Level)>::addLiteralOption<llvm::ScheduleDAGSDNodes* (*)(llvm::

SelectionDAGISel_, llvm::CodeGenOpt::Level)>(char const_, llvm::ScheduleDAGSDNodes* (* const&)(llvm::SelectionDAGISel_, llvm::CodeGenOpt::Level), char const_) clone .part.609 fro
m /opt/llvm/3.4//lib/libLLVM-3.4.so
#4 0x00007ffff159618e in llvm::RegisterPassParserllvm::RegisterScheduler::NotifyAdd(char const_, void_ ()(), char const) () from /opt/llvm/3.4//lib/libLLVM-3.4.so
#5 0x00007fffeeff8236 in llvm::MachinePassRegistry::Remove(llvm::MachinePassRegistryNode*) () from /usr/lib/x86_64-linux-gnu/dri/swrast_dri.so

[kraiskil]

The above is only a work-around, a proper fix would be to detect this sort of mesa-GLX lib at configure time.

[pjaaskel]

One can argue whether this is really a pocl or LLVMpipe bug ;)

[kraiskil]

Mesa bug, or even LLVM? If mesa uses the LLVM C++ API (which is known to change), then mesa should link with -Bsymbolic. If mesa uses the LLVM C API, then this is an issue with LLVM-C (IIRC, the llvm libs expose both C and C++ API in the same libraries, so there is no solution for this). Or then it is a distro bug, in that they build some package wrong.

LLVM discourages using the shared libraries (http://llvm.org/docs/CMake.html#frequently-used-cmake-variables). Probably just because of these sort of issues.

So this is still a pocl issue (marking this issue 'enhancement', not 'bug' :)).
Especially since this issue popped up with the API mode, so it shows in 0.9 for the first time, and therefore a lot of OpenGL-OpenCL mixed users will end up in a situation where "it works with ${VENDOR}-OCL, segfaults with pocl".

[pjaaskel]

The old multiple LLVM issue(s). There's a newer bug with a link to the LLVM bug.