[clang] [llvm] [AMDGPU] Add async variants of tensor load/store LDS intrinsics. (PR #200775)

[Krzysztof Drewniak via cfe-commits]

krzysz00 wrote:

> There's no reason why @llvm.av.global.load.b128() cannot be implemented as a simple library function that contains a GLOBAL_LOAD instruction with the proper flags, if such a library existed to cover all subtargets.

I think that'd also require something like the proposed "allow operand bundles that don't have 'you can delete this' semantics on load/store" that came out of Nvidia a while back

https://github.com/llvm/llvm-project/pull/200775