[clang] [llvm] [AMDGPU] Add OpenCL-specific fence address space masks (PR #78572)

[Sameer Sahasrabuddhe via llvm-commits]

ssahasra wrote:

> Should we also rename the MMRA to `amdgpu-fence-as` (remove OpenCL from the name) ?

Even the "fence" prefix is not entirely correct. The same tags also make sense on a load-acquire or store-release, which are "fence like" instructions, or "operations with implicit fences". Why not just "as:global", "as:local", etc? The fact that they are used as !mmra on a fence-like instruction makes it clear that they represent the address spaces that are caused to be synchronized by that instruction, and not incidentally the address space that the load-acquire or store-release itself wants to access.

https://github.com/llvm/llvm-project/pull/78572