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

[Kmeakin]

memcpy

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

memmove

https://godbolt.org/z/TafacnzY3

LLVM assembly:

move7:
        ldrb    w8, [x1, #6]
        ldrh    w9, [x1, #4]
        ldr     w10, [x1]
        strb    w8, [x0, #6]
        strh    w9, [x0, #4]
        str     w10, [x0]
        ret

GCC assembly:

move7:
        ldr     w2, [x1]
        ldr     w1, [x1, 3]
        str     w2, [x0]
        str     w1, [x0, 3]
        ret

memset

https://godbolt.org/z/n3qq3EYrT

LLVM assembly:

set4:
        mov     w8, #16843009
        and     w9, w1, #0xff
        mul     w8, w9, w8
        str     w8, [x0]
        ret

GCC assembly:

set4:
        dup     v31.16b, w1
        str     s31, [x0]
        ret

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Karl Meakin (Kmeakin)

https://godbolt.org/z/9E1bPo9Eb ```c #define COPY(size) \ void copy##size(void* dst, void* src) { __builtin_memcpy(dst, src, size); }

COPY(65)
COPY(66)
COPY(67)

LLVM assembly:
```asm
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

copy66:
        ldr     q0, [x1]
        str     q0, [x0]
        ldp     q2, q0, [x1, #<!-- -->32]
        ldrh    w8, [x1, #<!-- -->64]
        ldr     q1, [x1, #<!-- -->16]
        strh    w8, [x0, #<!-- -->64]
        stp     q2, q0, [x0, #<!-- -->32]
        str     q1, [x0, #<!-- -->16]
        ret

copy67:
        ldr     q0, [x1]
        str     q0, [x0]
        ldp     q2, q0, [x1, #<!-- -->32]
        ldur    w8, [x1, #<!-- -->63]
        ldr     q1, [x1, #<!-- -->16]
        stur    w8, [x0, #<!-- -->63]
        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
copy66:
        ldp     q29, q28, [x1]
        ldp     q31, q30, [x1, 32]
        ldrh    w1, [x1, 64]
        stp     q29, q28, [x0]
        stp     q31, q30, [x0, 32]
        strh    w1, [x0, 64]
        ret
copy67:
        ldp     q29, q28, [x1, 32]
        ldp     q31, q30, [x1]
        ldr     w1, [x1, 63]
        stp     q29, q28, [x0, 32]
        stp     q31, q30, [x0]
        str     w1, [x0, 63]
        ret

[efriedma-quic]

Please split into separate issues; these aren't really connected, despite them all being "string functions", and it'll be really confusing if someone fixes one of them.