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

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

> Hi @sjoerdmeijer, thanks for putting up the new patch! I realise I'm backtracking on this at the risk of you being annoyed :), but I now wonder if it's worth reverting to the original version for now that always favours fixed-width vectorisation in the event of a cost-model tie specifically for neoverse-v2. I do absolutely think that where we find issues with the cost model or codegen using scalable vectors we should try to fix those, but I can buy the argument that for many simple loops in practice when using scalable vectorisation the SVE predication does not come entirely for free and that there is a slight reason to favour fixed-width.

Absolutely no problem. :) Thanks for your patience and help with this, @david-arm. I have reverted this to the previous version, which should have all minor comments addressed.

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