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

[David Green via llvm-commits]

https://github.com/davemgreen commented:

I don't believe this code has ever worked very well. The DAG often just undoes it as an optimisation.

This seems like it would an improvement baring those issues, I would guess that the existing code did it the way it did to try and produce the resulting loads/stores more in order.

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