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

[David Sherwood via llvm-commits]

david-arm wrote:

The memory checks for the inner loop only need to be done once I assume, since I'd expect the start value for the needle to be loop invariant? If not, you'll need to calculate the entire needle range for the whole outer loop. I'd expect a minor hit to performance, but given the fantastic improvements you're reporting I'd still expect it to be a nice win. :) For the mismatch case the trip counts were typically quite small and the memory checks reduced the performance from around 8% -> 7% or something like that.

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