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

[Gang Chen via llvm-commits]

cmc-rep wrote:

> > I can see mulitple LIT test changed after adding SROA, for example simple_indirect_call.ll, I checked, load and store get eliminated. This was originally from Nicolai. He observed this issue with some HIP code he wrote.
> 
> The codegen pipeline is not responsible for most of the heavy optimizations. Ideally we would not run SROA or InferAddressSpaces at all in the codegen pipeline. We need a phase ordering test, which runs opt -O2/O3 which demonstrates the issue. Nothing in llvm/test/CodeGen will show if this is useful or not

I kind of see what you mean. I have limited the change to the callback part only. That should be sufficient for Nicolai's case.
I am not sure what a phase ordering test is. Could you give me a reference?

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