[llvm] [ScheduleDAG] Dirty height/depth in addPred/removePred even for latency zero (PR #102915)

[Björn Pettersson via llvm-commits]

bjope wrote:

> Correctness issue or enhancement?

I think that kind of depends on how you see it.
I don't know about any case where this has resulted in miscompile (specially since it has been incorrect in-tree for 15 years). So it is not a correctness issue in that sense.

The goal with the patch is to serve as an optimization/enhancement. It should make the scheduling DAG more _correct_, allowing optimizations based on the scheduling DAG to pick a better schedule when looking at height/depth.

https://github.com/llvm/llvm-project/pull/102915