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

[Jon Chesterfield via llvm-commits]

JonChesterfield wrote:

This doesn't sound right.

This patch fixes miscompilation of programs with complicated (indirect function themed) call graphs, which were not really expected to exist outside of O0. rocRAND is supposed to be performance sensitive so really shouldn't be full of indirect calls.

It might mean we've regressed codegen on things that were previously compiled correctly in which case we've made a mistake here, it should have been a no-op. Hopefully the internal bug report has some IR associated, it would be useful to see which case we have here.

No to reverting though, this fixed a straightforward miscompilation that was blocking flang O0. We either accept that rocRAND should stop using indirect calls or work out where we've accidentally regressed.

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