[llvm] [MachineScheduler] Don't add redundant dependencies (PR #205613)

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

> When I read the description I got a feeling that this was NFC (just skipping redundant deps).
> 
> But it seems like this impacts codegen. Is that just a side-effect due to the scheduler taking different actions based on in which order deps are found etc when analysing the scheduling DAG. So when being lucky codegen improves, or if being unlucky there could be regressions. Or should I fear that something is wrong when seing codegen diffs?

yes this is just different processing order which may be lucky/unlucky

https://github.com/llvm/llvm-project/pull/205613