On a Kabini system using kernel 4.8-rc8, Mesa and LLVM git, running GROMACS
OpenCL example 1536 from [1] results in

radeon: Failed to allocate virtual address for buffer:
radeon:    size      : 4096 bytes
radeon:    alignment : 4096 bytes
radeon:    domains   : 2
radeon:    va        : 0x000000000948b000
radeon: Failed to deallocate virtual address for buffer:
radeon:    size      : 4096 bytes
radeon:    va        : 0x948b000
radeon: Failed to allocate virtual address for buffer:
radeon:    size      : 4096 bytes
radeon:    alignment : 4096 bytes
radeon:    domains   : 2
radeon:    va        : 0x000000000948b000
radeon: Failed to deallocate virtual address for buffer:
radeon:    size      : 4096 bytes
radeon:    va        : 0x948b000

This is one of the larger examples, so it requires more memory than the
previous ones. I will test the largest example (3072) soon, and I expect it to
crash the same way.

This is a regression. It will be very hard to bisect LLVM and kernel on this
machine as it is really slow, but Mesa is likely bisectable. Let me know if it
is worth it.

[1] ftp://ftp.gromacs.org/pub/benchmarks/water_GMX50_bare.tar.gz