[llvm] [AMDGPU] Fix missing cache bypass scope on seq_cst buffer fat pointer atomics (PR #213585)

[Arseniy Obolenskiy via llvm-commits]

aobolensk wrote:

> Wouldn't adding a scope operand to the intrinsic be a less hacky way to get this done? Using a new custom metadata might be dropped by other passes because it's not recognized, provoking a correctness issue. At the very least, you'd want this documented in llvm-project/llvm/docs/AMDGPUUsage.rst
> 
> Not that this would be a solution to this specific bug, but It's also worth noting that the amdgpu memory model effort will eventually get to buffer intrinsics and it would include scoping info (e.g. see #191246 and related stack).

Thanks for the suggestion, yes, that'd be cleaner. Mentioned it in the docs

https://github.com/llvm/llvm-project/pull/213585