[llvm] [AArch64] Increase fixed-length SVE scalarization costs (PR #226120)

[David Green via llvm-commits]

https://github.com/davemgreen commented:

My understanding is that getScalarizationOverhead was added because the default cost model used a 0 lane cost for inserts/extract to lane 0 at the time, which isn't something that should apply to integer types. It has always been difficult to change these costs without causing regressions.

With that no long being the case, it might be best to remove getScalarizationOverhead if the performance remains OK. I would expect the cost to come through getVectorInstrCost for the most part, unless there is a need to accurately get the cost of parts that are shared when scalarizing (which might be the case for extracting from scalable vectors, but doesn't seem to be at the moment).

https://github.com/llvm/llvm-project/pull/226120