[AArch64] Suboptimal `memcpy` for non-power of two sizes

[Kmeakin]

Split off from #165625

https://godbolt.org/z/Yfq56nfTP

LLVM assembly:

copy65:
        ldr     q0, [x1]
        str     q0, [x0]
        ldp     q2, q0, [x1, #32]
        ldrb    w8, [x1, #64]
        ldr     q1, [x1, #16]
        strb    w8, [x0, #64]
        stp     q2, q0, [x0, #32]
        str     q1, [x0, #16]
        ret

GCC assembly:

copy65:
        ldp     q29, q28, [x1]
        ldp     q31, q30, [x1, 32]
        ldrb    w1, [x1, 64]
        stp     q29, q28, [x0]
        stp     q31, q30, [x0, 32]
        strb    w1, [x0, 64]
        ret

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Karl Meakin (Kmeakin)

Split off from https://github.com//issues/165625

https://godbolt.org/z/Yfq56nfTP

LLVM assembly:

copy65:
        ldr     q0, [x1]
        str     q0, [x0]
        ldp     q2, q0, [x1, #<!-- -->32]
        ldrb    w8, [x1, #<!-- -->64]
        ldr     q1, [x1, #<!-- -->16]
        strb    w8, [x0, #<!-- -->64]
        stp     q2, q0, [x0, #<!-- -->32]
        str     q1, [x0, #<!-- -->16]
        ret

GCC assembly:

copy65:
        ldp     q29, q28, [x1]
        ldp     q31, q30, [x1, 32]
        ldrb    w1, [x1, 64]
        stp     q29, q28, [x0]
        stp     q31, q30, [x0, 32]
        strb    w1, [x0, 64]
        ret

[clingfei]

I am glad to work on it.