[llvm] [CodeGen] Correctly handle non-standard cases in RemoveLoadsIntoFakeUses (PR #111551)

[Stephen Tozer via llvm-commits]

SLTozer wrote:

> Spills for these still feels like a situation we should have avoided in the first place. Is it possible to teach foldMemoryOperand to absorb the spill reference directly into the instruction, or otherwise avoid inserting the restore in the first place?

I'm making a first pass attempt at this; if it turns out to be simple without affecting value liveness, then that solution could replace this pass, which would simplify matters! I've not worked much in regalloc though, so if there are complications I'd prefer to land this patch as a stopgap-measure to ensure that fake uses aren't handled incorrectly (potentially causing crashes) in the meanwhile.

https://github.com/llvm/llvm-project/pull/111551