[llvm] [AArch64] Increase scatter overhead on Neoverse-V2 (PR #101296)

[David Green via llvm-commits]

https://github.com/davemgreen commented:

I'm pretty sure we have had reports recently of cases where the gather/scatter cost should be lower in order to get the best performance!

@david-arm is away at the moment. I don't love it when we have a load of random factors that don't have much meaning for each different CPU. I prefer it when they work the same, if we can, but gather / scatter can be hard to come up with a single value for especially with how LLVM models costs at the moment.

Generally we try and avoid getProcFamily nowadays and add subtarget features instead. I think in the long run we need a better cost model for gather/scatters. In the meantime can we move the Overhead values into the subtarget?

https://github.com/llvm/llvm-project/pull/101296