Latest development build 2021-02-12 12:35 UTC

Fix translation of cmpxchg instruction

Previous implementation optimized produced code assuming a particular
pattern in input LLVM IR. But that assumption is not always correct.
Now we always insert extra instructions to preserve semantics of cmpxchg.
Those extra instruction can be optimized later.

Signed-off-by: Alexey Sotkin <alexey.sotkin@intel.com>