[llvm] [AArch64] Use INDEX for constant Neon step vectors (PR #113424)

[Paul Walker via llvm-commits]

paulwalker-arm wrote:

1. We've tried to be conservative when using all the different forms of `INDEX`.  At least on Neoverse cores their bandwidth can be limiting and so we generally prefer the immediate only forms that can be hoisted out of loops leaving only high bandwidth arithmetic instructions in loop rather than an in-loop register variant of `INDEX`.

I have some work that I need to extract into a pull request to better enforce this by making the decision explicit (i.e. lowering code can choose exactly the variant it wants) rather than the current refcount method we use during instruction selection.

2. Agreed.

https://github.com/llvm/llvm-project/pull/113424