[llvm] [LV][AArch64] Prefer Fixed over Scalable if cost-model is equal (Neoverse V2) (PR #95819)

[David Sherwood via llvm-commits]

david-arm wrote:

Hi @sjoerdmeijer, in the loops that you looked at where SVE performed worse did the SVE code itself look optimal, i.e. as good as it can be? The only thing I'm nervous about here is that we're switching to NEON due to poor codegen or a bad cost model, rather than due to a fundamental problem with SVE. For example, if we're generating more code in the SVE loop then that suggests the costs for the SVE loop are just too low so by hiking the costs we would favour NEON anyway. We should definitely fix any cost model issues like that where we can. Also, if you have any examples where SVE codegen is actually worse can you create issues for them? Thanks!

https://github.com/llvm/llvm-project/pull/95819