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

[David Green via llvm-commits]

davemgreen wrote:

I agree that this seems perfectly sensible to me. I tried it on a few other machines (a mixture of emulation and real hardware) and it looked profitable to pick neon on a tie. (Even the Neoverse-V1 with a 256bit SVE length came out as a every-so-slight, meaninglessly small performance gain apparently. Obviously there were a lot of things getting better and worse).

The fact that Neon has LDP and postinc would feel like enough of a reason to favour Neon on a tie. Cost modelling in the mid-end is always going to be fairly high level and there will always be ties. Favouring Neon sounds like the most reasonable option between the two, providing if we have performance to back it up.

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