[llvm] [RFC][AMDGPU] Add vulkan:private/nonprivate MMRAs support (PR #78573)

[Sameer Sahasrabuddhe via llvm-commits]

================
@@ -10,6 +10,20 @@
 /// Memory legalizer - implements memory model. More information can be
 /// found here:
 ///   http://llvm.org/docs/AMDGPUUsage.html#memory-model
+///
+/// Notes on the Vulkan Memory Model, which is controlled by the
+/// `amdgpu.vulkan.memory-model` module flag. When that flag
+/// is provided:
+/// - vulkan:private operations cannot be atomic, and as such
+///   their codegen is not affected.
+/// - vulkan:nonprivate bypasses all caches not coherent for GFXIP.
+///   - vulkan:nonprivate does not need to use invalidates or writeback
----------------
ssahasra wrote:

I think you mean fences do not need to invalidate or write back the cache?

https://github.com/llvm/llvm-project/pull/78573