| 1 | #include <cinttypes> |
|---|
| 2 | #include <cstdint> |
|---|
| 3 | #include <cstdio> |
|---|
| 4 | |
|---|
| 5 | union alignas(16) xmm_t { |
|---|
| 6 | uint64_t as_uint64[2]; |
|---|
| 7 | uint8_t as_uint8[16]; |
|---|
| 8 | }; |
|---|
| 9 | |
|---|
| 10 | int main() { |
|---|
| 11 | constexpr xmm_t xmm_fill = { |
|---|
| 12 | .as_uint64 = { 0, 0 } |
|---|
| 13 | }; |
|---|
| 14 | |
|---|
| 15 | uint64_t mm[8]; |
|---|
| 16 | xmm_t xmm[8]; |
|---|
| 17 | |
|---|
| 18 | asm volatile( |
|---|
| 19 | "movq %2, %%mm0\n\t" |
|---|
| 20 | "movq %2, %%mm1\n\t" |
|---|
| 21 | "movq %2, %%mm2\n\t" |
|---|
| 22 | "movq %2, %%mm3\n\t" |
|---|
| 23 | "movq %2, %%mm4\n\t" |
|---|
| 24 | "movq %2, %%mm5\n\t" |
|---|
| 25 | "movq %2, %%mm6\n\t" |
|---|
| 26 | "movq %2, %%mm7\n\t" |
|---|
| 27 | "\n\t" |
|---|
| 28 | "movaps %2, %%xmm0\n\t" |
|---|
| 29 | "movaps %2, %%xmm1\n\t" |
|---|
| 30 | "movaps %2, %%xmm2\n\t" |
|---|
| 31 | "movaps %2, %%xmm3\n\t" |
|---|
| 32 | "movaps %2, %%xmm4\n\t" |
|---|
| 33 | "movaps %2, %%xmm5\n\t" |
|---|
| 34 | "movaps %2, %%xmm6\n\t" |
|---|
| 35 | "movaps %2, %%xmm7\n\t" |
|---|
| 36 | "\n\t" |
|---|
| 37 | "int3\n\t" |
|---|
| 38 | "\n\t" |
|---|
| 39 | "movq %%mm0, 0x00(%0)\n\t" |
|---|
| 40 | "movq %%mm1, 0x08(%0)\n\t" |
|---|
| 41 | "movq %%mm2, 0x10(%0)\n\t" |
|---|
| 42 | "movq %%mm3, 0x18(%0)\n\t" |
|---|
| 43 | "movq %%mm4, 0x20(%0)\n\t" |
|---|
| 44 | "movq %%mm5, 0x28(%0)\n\t" |
|---|
| 45 | "movq %%mm6, 0x30(%0)\n\t" |
|---|
| 46 | "movq %%mm7, 0x38(%0)\n\t" |
|---|
| 47 | "\n\t" |
|---|
| 48 | "movaps %%xmm0, 0x00(%1)\n\t" |
|---|
| 49 | "movaps %%xmm1, 0x10(%1)\n\t" |
|---|
| 50 | "movaps %%xmm2, 0x20(%1)\n\t" |
|---|
| 51 | "movaps %%xmm3, 0x30(%1)\n\t" |
|---|
| 52 | "movaps %%xmm4, 0x40(%1)\n\t" |
|---|
| 53 | "movaps %%xmm5, 0x50(%1)\n\t" |
|---|
| 54 | "movaps %%xmm6, 0x60(%1)\n\t" |
|---|
| 55 | "movaps %%xmm7, 0x70(%1)\n\t" |
|---|
| 56 | : |
|---|
| 57 | : "a" (mm), "b" (xmm), "m" (xmm_fill) |
|---|
| 58 | : "%mm0" , "%mm1" , "%mm2" , "%mm3" , "%mm4" , "%mm5" , "%mm6" , "%mm7" , "%xmm0" , |
|---|
| 59 | "%xmm1" , "%xmm2" , "%xmm3" , "%xmm4" , "%xmm5" , "%xmm6" , "%xmm7" |
|---|
| 60 | ); |
|---|
| 61 | |
|---|
| 62 | for (int i = 0; i < 8; ++i) |
|---|
| 63 | printf(format: "mm%d = 0x%016" PRIx64 "\n" , i, mm[i]); |
|---|
| 64 | for (int i = 0; i < 8; ++i) { |
|---|
| 65 | printf(format: "xmm%d = { " , i); |
|---|
| 66 | for (int j = 0; j < sizeof(xmm->as_uint8); ++j) |
|---|
| 67 | printf(format: "0x%02x " , xmm[i].as_uint8[j]); |
|---|
| 68 | printf(format: "}\n" ); |
|---|
| 69 | } |
|---|
| 70 | |
|---|
| 71 | return 0; |
|---|
| 72 | } |
|---|
| 73 | |
|---|
