[clang] [llvm] [clang-tools-extra] [AArch64] Add an AArch64 pass for loop idiom transformations (PR #72273)

[David Sherwood via cfe-commits]

david-arm wrote:

> Hi! I wonder that have you conducted any tests to determine the potential performance increase of this pass in the SPEC2017 557xz benchmark? I attempted to apply it to the xz benchmark, but only one copy(--copies=1) demonstrated a significant increase(about 3%), but there was no increase when I set --copies=128 or higher. Do you have any suggestions or test results that you could share?

The most significant gains with xz have already been achieved when https://github.com/llvm/llvm-project/pull/77480 and https://github.com/llvm/llvm-project/pull/77480 landed, which improved performance by 6-7% for neoverse-v1. This PR is a NFC refactoring patch so it won't improve performance further. My follow-on patch (not yet posted) will trigger more cases in xz, but I don't expect any substantial performance gains for xz. The main goal of extending this pass further is to improve code coverage and testing, and hopefully there will be other applications besides xz that will benefit too.

https://github.com/llvm/llvm-project/pull/72273