[llvm] [AArch64] Add MATCH loops to LoopIdiomVectorizePass (PR #101976)

[David Sherwood via llvm-commits]

david-arm wrote:

Alternatively, if you can prove at compile time the vector loop won't fault you can also do that as a first step, then have a follow-on patch that enables loop versioning. For example, if you have some IR where the pointers all come from an `alloca` and you know the trip counts for both loops you can use `llvm::isDereferenceableAndAlignedInLoop` to ask the question if the load pointer can be unconditionally dereferenced.

https://github.com/llvm/llvm-project/pull/101976