[llvm] [AArch64] Utilize `XAR` for certain vector rotates (PR #137629)

[David Green via llvm-commits]

davemgreen wrote:

MOVIv2d_ns will actually set all the upper bits too, so it can use MOVIv2d_ns and an SUBREG_TO_REG. Something like `SUBREG_TO_REG 0, Vn, zsub` will tell the compiler that the fp128 is extended to a zreg and that the top bits are zeroes.

https://github.com/llvm/llvm-project/pull/137629