[llvm] [llvm][amdgpu] Handle indirect refs to LDS GVs during LDS lowering (PR #124089)

[Jon Chesterfield via llvm-commits]

JonChesterfield wrote:

1. Downloaded the docker containing the compiler that claims a performance regression
2. Downloaded rocRAND and checked out the specific commit
3. Set save-temps in environment variable, built the benchmark
4. Copied benchmark_rocrand_kernel-hip-amdgcn-amd-amdhsa-gfx942.bc out
5. Ran it through amd-staging llc with -print-before=amdgpu-lower-module-lds, extracted the IR
6. Passed that IR through the amd-staging pass with and without this patch, no difference
7. Ran the benchmark through the staging llc, extracted different IR
8. Passed that IR through the amd-staging pass with and without this patch, no difference

I claim that the regression is elsewhere.

https://github.com/llvm/llvm-project/pull/124089