[llvm] [AArch64] Optimize memcpy for non-power of two sizes (PR #168890)

[David Green via llvm-commits]

https://github.com/davemgreen commented:

The problem with this existing code in getMemcpyLoadsAndStores is that it doesn't work very well in some cases - it performs an optimization that SDAG in goes ahead and undo's if it can tell that the pointers do not alias.

Your code sounds fine but can we see if we can stop it from producing loads/stores that count down instead of up (or random, as it tends to be)?

https://github.com/llvm/llvm-project/pull/168890