On 25 April 2017 at 09:35, Nikunj A Dadhania <nik...@linux.vnet.ibm.com> wrote: > Travis builds failure was reported for powernv boot-serial test with > qemu built with clang. > > Debugging revealed that CONFIG_ATOMIC64 wasnt getting set for the clang > build because of that atomic operations weren't being used and was > resulting in MTTCG failure in the powernv boot-serial test. > > libatomic is required to successfully test atomic64 and atomic128 for > clang. Introduced newer checks for the same. And on failure default to > single threaded tcg support in PPC64. > > Signed-off-by: Nikunj A Dadhania <nik...@linux.vnet.ibm.com>
Do we really need libatomic? I thought the intention here was that atomic operations were only done on types of width of the pointer or less, which should all be doable by the compiler inline, and thus don't need the external library. In the past "doesn't work without libatomic" usually meant "accidentally tried to do an atomic operation on a type that is too wide". thanks -- PMM
