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

[Jeremy Morse via llvm-commits]

jmorse wrote:

I suppose with non-x86 architectures not supporting instruction referencing, this means they'll be slower than the previous implementation uploaded. Un-necessary load elimination happens here and is disabled, instead of happening really late in code emission like the previous implementation. But this is fine -- when everyone supports instruction referencing everything will be faster!

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