As described in this proposal: https://groups.google.com/forum/#!topic/llvm-dev/GtWfCc5j-4U
In OpenCL 2.x, two atomic operations on the same atomic object need to have the same scope to prevent a data race. This derives from the definition of "inclusive scope" in OpenCL 2.x. Encoding OpenCL 2.x scope as metadata in LLVM IR would be a problem because there cannot be a "safe default value" to be used when the metadata is dropped. If the "largest" scope is used as the default, then the optimizer must guarantee that the metadata is dropped from every atomic operation in the whole program, or not dropped at all.
We cannot use the metadata approach since this metadata can be dropped during the processing of one module but not dropped in the processing of a second module, potentially resulting in inconsistent scopes for synchronizing operations leading to data races and subsequently leading to correctness issues.