[llvm] [AMDGPU] SplitModule: Better handling of aliases & inline assembly (PR #106528)

[Pierre van Houtryve via llvm-commits]

Pierre-vh wrote:

> Other optimizations already redirect calls to aliases to direct calls, so I don't think this pass should have to put in the effort to also do the same. I think GlobalOpt?

Now that I remember, this is indeed an artificial case.
I'll check if just adding an opt run line also does the same thing, if it does, then I'll just eliminate that part of the alias handling.

https://github.com/llvm/llvm-project/pull/106528