| 1 | #include <stdio.h> |
| 2 | #include <stdint.h> |
| 3 | #include <stdlib.h> |
| 4 | #include <string.h> |
| 5 | |
| 6 | #ifndef TEST_FILE |
| 7 | #define TEST_FILE "test-mmx.h" |
| 8 | #endif |
| 9 | #ifndef EMMS |
| 10 | #define EMMS "emms" |
| 11 | #endif |
| 12 | |
| 13 | typedef void (*testfn)(void); |
| 14 | |
| 15 | typedef struct { |
| 16 | uint64_t q0, q1; |
| 17 | } __attribute__((aligned(16))) v2di; |
| 18 | |
| 19 | typedef struct { |
| 20 | uint64_t mm[8]; |
| 21 | v2di xmm[8]; |
| 22 | uint64_t r[16]; |
| 23 | uint64_t flags; |
| 24 | uint32_t ff; |
| 25 | uint64_t pad; |
| 26 | v2di mem[4]; |
| 27 | v2di mem0[4]; |
| 28 | } reg_state; |
| 29 | |
| 30 | typedef struct { |
| 31 | int n; |
| 32 | testfn fn; |
| 33 | const char *s; |
| 34 | reg_state *init; |
| 35 | } TestDef; |
| 36 | |
| 37 | reg_state initI; |
| 38 | reg_state initF32; |
| 39 | reg_state initF64; |
| 40 | |
| 41 | static void dump_mmx(int n, const uint64_t *r, int ff) |
| 42 | { |
| 43 | if (ff == 32) { |
| 44 | float v[2]; |
| 45 | memcpy(v, r, sizeof(v)); |
| 46 | printf("MM%d = %016lx %8g %8g\n", n, *r, v[1], v[0]); |
| 47 | } else { |
| 48 | printf("MM%d = %016lx\n", n, *r); |
| 49 | } |
| 50 | } |
| 51 | |
| 52 | static void dump_xmm(const char *name, int n, const v2di *r, int ff) |
| 53 | { |
| 54 | printf("%s%d = %016lx %016lx\n", |
| 55 | name, n, r->q1, r->q0); |
| 56 | if (ff == 32) { |
| 57 | float v[4]; |
| 58 | memcpy(v, r, sizeof(v)); |
| 59 | printf(" %8g %8g %8g %8g\n", |
| 60 | v[3], v[2], v[1], v[0]); |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | static void dump_regs(reg_state *s, int ff) |
| 65 | { |
| 66 | int i; |
| 67 | |
| 68 | for (i = 0; i < 8; i++) { |
| 69 | dump_mmx(i, &s->mm[i], ff); |
| 70 | } |
| 71 | for (i = 0; i < 4; i++) { |
| 72 | dump_xmm("mem", i, &s->mem0[i], 0); |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | static void compare_state(const reg_state *a, const reg_state *b) |
| 77 | { |
| 78 | int i; |
| 79 | for (i = 0; i < 8; i++) { |
| 80 | if (a->mm[i] != b->mm[i]) { |
| 81 | printf("MM%d = %016lx\n", i, b->mm[i]); |
| 82 | } |
| 83 | } |
| 84 | for (i = 0; i < 16; i++) { |
| 85 | if (a->r[i] != b->r[i]) { |
| 86 | printf("r%d = %016lx\n", i, b->r[i]); |
| 87 | } |
| 88 | } |
| 89 | for (i = 0; i < 8; i++) { |
| 90 | if (memcmp(&a->xmm[i], &b->xmm[i], 8)) { |
| 91 | dump_xmm("xmm", i, &b->xmm[i], a->ff); |
| 92 | } |
| 93 | } |
| 94 | for (i = 0; i < 4; i++) { |
| 95 | if (memcmp(&a->mem0[i], &a->mem[i], 16)) { |
| 96 | dump_xmm("mem", i, &a->mem[i], a->ff); |
| 97 | } |
| 98 | } |
| 99 | if (a->flags != b->flags) { |
| 100 | printf("FLAGS = %016lx\n", b->flags); |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | #define LOADMM(r, o) "movq " #r ", " #o "[%0]\n\t" |
| 105 | #define LOADXMM(r, o) "movdqa " #r ", " #o "[%0]\n\t" |
| 106 | #define STOREMM(r, o) "movq " #o "[%1], " #r "\n\t" |
| 107 | #define STOREXMM(r, o) "movdqa " #o "[%1], " #r "\n\t" |
| 108 | #define MMREG(F) \ |
| 109 | F(mm0, 0x00) \ |
| 110 | F(mm1, 0x08) \ |
| 111 | F(mm2, 0x10) \ |
| 112 | F(mm3, 0x18) \ |
| 113 | F(mm4, 0x20) \ |
| 114 | F(mm5, 0x28) \ |
| 115 | F(mm6, 0x30) \ |
| 116 | F(mm7, 0x38) |
| 117 | #define XMMREG(F) \ |
| 118 | F(xmm0, 0x040) \ |
| 119 | F(xmm1, 0x050) \ |
| 120 | F(xmm2, 0x060) \ |
| 121 | F(xmm3, 0x070) \ |
| 122 | F(xmm4, 0x080) \ |
| 123 | F(xmm5, 0x090) \ |
| 124 | F(xmm6, 0x0a0) \ |
| 125 | F(xmm7, 0x0b0) |
| 126 | #define LOADREG(r, o) "mov " #r ", " #o "[rax]\n\t" |
| 127 | #define STOREREG(r, o) "mov " #o "[rax], " #r "\n\t" |
| 128 | #define REG(F) \ |
| 129 | F(rbx, 0xc8) \ |
| 130 | F(rcx, 0xd0) \ |
| 131 | F(rdx, 0xd8) \ |
| 132 | F(rsi, 0xe0) \ |
| 133 | F(rdi, 0xe8) \ |
| 134 | F(r8, 0x100) \ |
| 135 | F(r9, 0x108) \ |
| 136 | F(r10, 0x110) \ |
| 137 | F(r11, 0x118) \ |
| 138 | F(r12, 0x120) \ |
| 139 | F(r13, 0x128) \ |
| 140 | F(r14, 0x130) \ |
| 141 | F(r15, 0x138) \ |
| 142 | |
| 143 | static void run_test(const TestDef *t) |
| 144 | { |
| 145 | reg_state result; |
| 146 | reg_state *init = t->init; |
| 147 | memcpy(init->mem, init->mem0, sizeof(init->mem)); |
| 148 | printf("%5d %s\n", t->n, t->s); |
| 149 | asm volatile( |
| 150 | MMREG(LOADMM) |
| 151 | XMMREG(LOADXMM) |
| 152 | "sub rsp, 128\n\t" |
| 153 | "push rax\n\t" |
| 154 | "push rbx\n\t" |
| 155 | "push rcx\n\t" |
| 156 | "push rdx\n\t" |
| 157 | "push %1\n\t" |
| 158 | "push %2\n\t" |
| 159 | "mov rax, %0\n\t" |
| 160 | "pushf\n\t" |
| 161 | "pop rbx\n\t" |
| 162 | "shr rbx, 8\n\t" |
| 163 | "shl rbx, 8\n\t" |
| 164 | "mov rcx, 0x140[rax]\n\t" |
| 165 | "and rcx, 0xff\n\t" |
| 166 | "or rbx, rcx\n\t" |
| 167 | "push rbx\n\t" |
| 168 | "popf\n\t" |
| 169 | REG(LOADREG) |
| 170 | "mov rax, 0xc0[rax]\n\t" |
| 171 | "call [rsp]\n\t" |
| 172 | "mov [rsp], rax\n\t" |
| 173 | "mov rax, 8[rsp]\n\t" |
| 174 | REG(STOREREG) |
| 175 | "mov rbx, [rsp]\n\t" |
| 176 | "mov 0xc0[rax], rbx\n\t" |
| 177 | "mov rbx, 0\n\t" |
| 178 | "mov 0xf0[rax], rbx\n\t" |
| 179 | "mov 0xf8[rax], rbx\n\t" |
| 180 | "pushf\n\t" |
| 181 | "pop rbx\n\t" |
| 182 | "and rbx, 0xff\n\t" |
| 183 | "mov 0x140[rax], rbx\n\t" |
| 184 | "add rsp, 16\n\t" |
| 185 | "pop rdx\n\t" |
| 186 | "pop rcx\n\t" |
| 187 | "pop rbx\n\t" |
| 188 | "pop rax\n\t" |
| 189 | "add rsp, 128\n\t" |
| 190 | MMREG(STOREMM) |
| 191 | EMMS "\n\t" |
| 192 | XMMREG(STOREXMM) |
| 193 | : : "r"(init), "r"(&result), "r"(t->fn) |
| 194 | : "memory", "cc", |
| 195 | "rsi", "rdi", |
| 196 | "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", |
| 197 | "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7", |
| 198 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", |
| 199 | "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", |
| 200 | "xmm12", "xmm13", "xmm14", "xmm15" |
| 201 | ); |
| 202 | compare_state(init, &result); |
| 203 | } |
| 204 | |
| 205 | #define TEST(n, cmd, type) \ |
| 206 | static void __attribute__((naked)) test_##n(void) \ |
| 207 | { \ |
| 208 | asm volatile(cmd); \ |
| 209 | asm volatile("ret"); \ |
| 210 | } |
| 211 | #include TEST_FILE |
| 212 | |
| 213 | |
| 214 | static const TestDef test_table[] = { |
| 215 | #define TEST(n, cmd, type) {n, test_##n, cmd, &init##type}, |
| 216 | #include TEST_FILE |
| 217 | {-1, NULL, "", NULL} |
| 218 | }; |
| 219 | |
| 220 | static void run_all(void) |
| 221 | { |
| 222 | const TestDef *t; |
| 223 | for (t = test_table; t->fn; t++) { |
| 224 | run_test(t); |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | #define ARRAY_LEN(x) (sizeof(x) / sizeof(x[0])) |
| 229 | |
| 230 | float val_f32[] = {2.0, -1.0, 4.8, 0.8, 3, -42.0, 5e6, 7.5, 8.3}; |
| 231 | uint64_t val_i64[] = { |
| 232 | 0x3d6b3b6a9e4118f2lu, 0x355ae76d2774d78clu, |
| 233 | 0xd851c54a56bf1f29lu, 0x4a84d1d50bf4c4fflu, |
| 234 | 0x5826475e2c5fd799lu, 0xfd32edc01243f5e9lu, |
| 235 | }; |
| 236 | |
| 237 | v2di deadbeef = {0xa5a5a5a5deadbeefull, 0xa5a5a5a5deadbeefull}; |
| 238 | |
| 239 | void init_f32reg(uint64_t *r) |
| 240 | { |
| 241 | static int n; |
| 242 | float v[2]; |
| 243 | int i; |
| 244 | for (i = 0; i < 2; i++) { |
| 245 | v[i] = val_f32[n++]; |
| 246 | if (n == ARRAY_LEN(val_f32)) { |
| 247 | n = 0; |
| 248 | } |
| 249 | } |
| 250 | memcpy(r, v, sizeof(*r)); |
| 251 | } |
| 252 | |
| 253 | void init_intreg(uint64_t *r) |
| 254 | { |
| 255 | static uint64_t mask; |
| 256 | static int n; |
| 257 | |
| 258 | *r = val_i64[n] ^ mask; |
| 259 | n++; |
| 260 | if (n == ARRAY_LEN(val_i64)) { |
| 261 | n = 0; |
| 262 | mask *= 0x104C11DB7; |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | static void init_all(reg_state *s) |
| 267 | { |
| 268 | int i; |
| 269 | |
| 270 | for (i = 0; i < 16; i++) { |
| 271 | init_intreg(&s->r[i]); |
| 272 | } |
| 273 | s->r[3] = (uint64_t)&s->mem[0]; /* rdx */ |
| 274 | s->r[5] = (uint64_t)&s->mem[2]; /* rdi */ |
| 275 | s->r[6] = 0; |
| 276 | s->r[7] = 0; |
| 277 | s->flags = 2; |
| 278 | for (i = 0; i < 8; i++) { |
| 279 | s->xmm[i] = deadbeef; |
| 280 | memcpy(&s->mm[i], &s->xmm[i], sizeof(s->mm[i])); |
| 281 | } |
| 282 | for (i = 0; i < 2; i++) { |
| 283 | s->mem0[i] = deadbeef; |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | int main(int argc, char *argv[]) |
| 288 | { |
| 289 | init_all(&initI); |
| 290 | init_intreg(&initI.mm[5]); |
| 291 | init_intreg(&initI.mm[6]); |
| 292 | init_intreg(&initI.mm[7]); |
| 293 | init_intreg(&initI.mem0[1].q0); |
| 294 | init_intreg(&initI.mem0[1].q1); |
| 295 | printf("Int:\n"); |
| 296 | dump_regs(&initI, 0); |
| 297 | |
| 298 | init_all(&initF32); |
| 299 | init_f32reg(&initF32.mm[5]); |
| 300 | init_f32reg(&initF32.mm[6]); |
| 301 | init_f32reg(&initF32.mm[7]); |
| 302 | init_f32reg(&initF32.mem0[1].q0); |
| 303 | init_f32reg(&initF32.mem0[1].q1); |
| 304 | initF32.ff = 32; |
| 305 | printf("F32:\n"); |
| 306 | dump_regs(&initF32, 32); |
| 307 | |
| 308 | if (argc > 1) { |
| 309 | int n = atoi(argv[1]); |
| 310 | run_test(&test_table[n]); |
| 311 | } else { |
| 312 | run_all(); |
| 313 | } |
| 314 | return 0; |
| 315 | } |