[llvm] [AMDGPU] add SROA before InferAddressSpace (PR #137909)

[Gang Chen via llvm-commits]

cmc-rep wrote:

> > we probably should try harder to avoid the additional run with reordering
> 
> +1

Any suggestion here? I do not see other obvious solution here: I cannot find usage of SROA-like optimization in our llc-pipeline.ll. Seems this is needed after inlining.

https://github.com/llvm/llvm-project/pull/137909