1 #include <cstdint>
2 
3 struct alignas(16) vec_t {
4   uint64_t a, b;
5 };
6 
7 int main() {
8   constexpr uint64_t gprs[] = {
9     0x0001020304050607,
10     0x1011121314151617,
11     0x2021222324252627,
12     0x3031323334353637,
13     0x4041424344454647,
14     0x5051525354555657,
15     0x6061626364656667,
16     0x7071727374757677,
17   };
18 
19   constexpr vec_t vecs[] = {
20     { 0x0F0E0D0C0B0A0908, 0x1716151413121110, },
21     { 0x100F0E0D0C0B0A09, 0x1817161514131211, },
22     { 0x11100F0E0D0C0B0A, 0x1918171615141312, },
23     { 0x1211100F0E0D0C0B, 0x1A19181716151413, },
24     { 0x131211100F0E0D0C, 0x1B1A191817161514, },
25     { 0x14131211100F0E0D, 0x1C1B1A1918171615, },
26     { 0x1514131211100F0E, 0x1D1C1B1A19181716, },
27     { 0x161514131211100F, 0x1E1D1C1B1A191817, },
28   };
29 
30   asm volatile(
31     "ldp      x0,  x1,  [%0]\n\t"
32     "ldp      x2,  x3,  [%0, #16]\n\t"
33     "ldp      x4,  x5,  [%0, #32]\n\t"
34     "ldp      x6,  x7,  [%0, #48]\n\t"
35     "\n\t"
36     "ld1      {v0.2d, v1.2d, v2.2d, v3.2d}, [%1], #64\n\t"
37     "ld1      {v4.2d, v5.2d, v6.2d, v7.2d}, [%1], #64\n\t"
38     "\n\t"
39     "brk      #0\n\t"
40     :
41     : "r"(gprs), "r"(vecs)
42     : "x0", "x1", "x2", "x3", "x4", "x5", "x6", "x7",
43       "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
44   );
45 
46   return 0;
47 }
48