projects / ceph.git / blame
| Commit | Line | Data |
|---|---|---|
| a4b75251 TL |
1 | // SPDX-License-Identifier: BSD-3-Clause |
| 2 | /* Copyright 2017-2020, Intel Corporation */ | |
| 3 | ||
| 4 | #include <immintrin.h> | |
| 5 | #include <stddef.h> | |
| 6 | #include <stdint.h> | |
| 7 | ||
| 8 | #include "pmem2_arch.h" | |
| 9 | #include "flush.h" | |
| 10 | #include "memcpy_memset.h" | |
| 11 | #include "memcpy_sse2.h" | |
| 12 | #include "valgrind_internal.h" | |
| 13 | ||
| 14 | static force_inline __m128i | |
| 15 | mm_loadu_si128(const char *src, unsigned idx) | |
| 16 | { | |
| 17 | return _mm_loadu_si128((const __m128i *)src + idx); | |
| 18 | } | |
| 19 | ||
| 20 | static force_inline void | |
| 21 | mm_stream_si128(char *dest, unsigned idx, __m128i src) | |
| 22 | { | |
| 23 | _mm_stream_si128((__m128i *)dest + idx, src); | |
| 24 | barrier(); | |
| 25 | } | |
| 26 | ||
| 27 | static force_inline void | |
| 28 | memmove_movnt4x64b(char *dest, const char *src) | |
| 29 | { | |
| 30 | __m128i xmm0 = mm_loadu_si128(src, 0); | |
| 31 | __m128i xmm1 = mm_loadu_si128(src, 1); | |
| 32 | __m128i xmm2 = mm_loadu_si128(src, 2); | |
| 33 | __m128i xmm3 = mm_loadu_si128(src, 3); | |
| 34 | __m128i xmm4 = mm_loadu_si128(src, 4); | |
| 35 | __m128i xmm5 = mm_loadu_si128(src, 5); | |
| 36 | __m128i xmm6 = mm_loadu_si128(src, 6); | |
| 37 | __m128i xmm7 = mm_loadu_si128(src, 7); | |
| 38 | __m128i xmm8 = mm_loadu_si128(src, 8); | |
| 39 | __m128i xmm9 = mm_loadu_si128(src, 9); | |
| 40 | __m128i xmm10 = mm_loadu_si128(src, 10); | |
| 41 | __m128i xmm11 = mm_loadu_si128(src, 11); | |
| 42 | __m128i xmm12 = mm_loadu_si128(src, 12); | |
| 43 | __m128i xmm13 = mm_loadu_si128(src, 13); | |
| 44 | __m128i xmm14 = mm_loadu_si128(src, 14); | |
| 45 | __m128i xmm15 = mm_loadu_si128(src, 15); | |
| 46 | ||
| 47 | mm_stream_si128(dest, 0, xmm0); | |
| 48 | mm_stream_si128(dest, 1, xmm1); | |
| 49 | mm_stream_si128(dest, 2, xmm2); | |
| 50 | mm_stream_si128(dest, 3, xmm3); | |
| 51 | mm_stream_si128(dest, 4, xmm4); | |
| 52 | mm_stream_si128(dest, 5, xmm5); | |
| 53 | mm_stream_si128(dest, 6, xmm6); | |
| 54 | mm_stream_si128(dest, 7, xmm7); | |
| 55 | mm_stream_si128(dest, 8, xmm8); | |
| 56 | mm_stream_si128(dest, 9, xmm9); | |
| 57 | mm_stream_si128(dest, 10, xmm10); | |
| 58 | mm_stream_si128(dest, 11, xmm11); | |
| 59 | mm_stream_si128(dest, 12, xmm12); | |
| 60 | mm_stream_si128(dest, 13, xmm13); | |
| 61 | mm_stream_si128(dest, 14, xmm14); | |
| 62 | mm_stream_si128(dest, 15, xmm15); | |
| 63 | } | |
| 64 | ||
| 65 | static force_inline void | |
| 66 | memmove_movnt2x64b(char *dest, const char *src) | |
| 67 | { | |
| 68 | __m128i xmm0 = mm_loadu_si128(src, 0); | |
| 69 | __m128i xmm1 = mm_loadu_si128(src, 1); | |
| 70 | __m128i xmm2 = mm_loadu_si128(src, 2); | |
| 71 | __m128i xmm3 = mm_loadu_si128(src, 3); | |
| 72 | __m128i xmm4 = mm_loadu_si128(src, 4); | |
| 73 | __m128i xmm5 = mm_loadu_si128(src, 5); | |
| 74 | __m128i xmm6 = mm_loadu_si128(src, 6); | |
| 75 | __m128i xmm7 = mm_loadu_si128(src, 7); | |
| 76 | ||
| 77 | mm_stream_si128(dest, 0, xmm0); | |
| 78 | mm_stream_si128(dest, 1, xmm1); | |
| 79 | mm_stream_si128(dest, 2, xmm2); | |
| 80 | mm_stream_si128(dest, 3, xmm3); | |
| 81 | mm_stream_si128(dest, 4, xmm4); | |
| 82 | mm_stream_si128(dest, 5, xmm5); | |
| 83 | mm_stream_si128(dest, 6, xmm6); | |
| 84 | mm_stream_si128(dest, 7, xmm7); | |
| 85 | } | |
| 86 | ||
| 87 | static force_inline void | |
| 88 | memmove_movnt1x64b(char *dest, const char *src) | |
| 89 | { | |
| 90 | __m128i xmm0 = mm_loadu_si128(src, 0); | |
| 91 | __m128i xmm1 = mm_loadu_si128(src, 1); | |
| 92 | __m128i xmm2 = mm_loadu_si128(src, 2); | |
| 93 | __m128i xmm3 = mm_loadu_si128(src, 3); | |
| 94 | ||
| 95 | mm_stream_si128(dest, 0, xmm0); | |
| 96 | mm_stream_si128(dest, 1, xmm1); | |
| 97 | mm_stream_si128(dest, 2, xmm2); | |
| 98 | mm_stream_si128(dest, 3, xmm3); | |
| 99 | } | |
| 100 | ||
| 101 | static force_inline void | |
| 102 | memmove_movnt1x32b(char *dest, const char *src) | |
| 103 | { | |
| 104 | __m128i xmm0 = mm_loadu_si128(src, 0); | |
| 105 | __m128i xmm1 = mm_loadu_si128(src, 1); | |
| 106 | ||
| 107 | mm_stream_si128(dest, 0, xmm0); | |
| 108 | mm_stream_si128(dest, 1, xmm1); | |
| 109 | } | |
| 110 | ||
| 111 | static force_inline void | |
| 112 | memmove_movnt1x16b(char *dest, const char *src) | |
| 113 | { | |
| 114 | __m128i xmm0 = mm_loadu_si128(src, 0); | |
| 115 | ||
| 116 | mm_stream_si128(dest, 0, xmm0); | |
| 117 | } | |
| 118 | ||
| 119 | static force_inline void | |
| 120 | memmove_movnt1x8b(char *dest, const char *src) | |
| 121 | { | |
| 122 | _mm_stream_si64((long long *)dest, *(long long *)src); | |
| 123 | } | |
| 124 | ||
| 125 | static force_inline void | |
| 126 | memmove_movnt1x4b(char *dest, const char *src) | |
| 127 | { | |
| 128 | _mm_stream_si32((int *)dest, *(int *)src); | |
| 129 | } | |
| 130 | ||
| 131 | static force_inline void | |
| 132 | memmove_movnt_sse_fw(char *dest, const char *src, size_t len, flush_fn flush, | |
| 133 | perf_barrier_fn perf_barrier) | |
| 134 | { | |
| 135 | size_t cnt = (uint64_t)dest & 63; | |
| 136 | if (cnt > 0) { | |
| 137 | cnt = 64 - cnt; | |
| 138 | ||
| 139 | if (cnt > len) | |
| 140 | cnt = len; | |
| 141 | ||
| 142 | memmove_small_sse2(dest, src, cnt, flush); | |
| 143 | ||
| 144 | dest += cnt; | |
| 145 | src += cnt; | |
| 146 | len -= cnt; | |
| 147 | } | |
| 148 | ||
| 149 | const char *srcend = src + len; | |
| 150 | prefetch_ini_fw(src, len); | |
| 151 | ||
| 152 | while (len >= PERF_BARRIER_SIZE) { | |
| 153 | prefetch_next_fw(src, srcend); | |
| 154 | ||
| 155 | memmove_movnt4x64b(dest, src); | |
| 156 | dest += 4 * 64; | |
| 157 | src += 4 * 64; | |
| 158 | len -= 4 * 64; | |
| 159 | ||
| 160 | memmove_movnt4x64b(dest, src); | |
| 161 | dest += 4 * 64; | |
| 162 | src += 4 * 64; | |
| 163 | len -= 4 * 64; | |
| 164 | ||
| 165 | memmove_movnt4x64b(dest, src); | |
| 166 | dest += 4 * 64; | |
| 167 | src += 4 * 64; | |
| 168 | len -= 4 * 64; | |
| 169 | ||
| 170 | COMPILE_ERROR_ON(PERF_BARRIER_SIZE != (4 + 4 + 4) * 64); | |
| 171 | ||
| 172 | if (len) | |
| 173 | perf_barrier(); | |
| 174 | } | |
| 175 | ||
| 176 | while (len >= 4 * 64) { | |
| 177 | memmove_movnt4x64b(dest, src); | |
| 178 | dest += 4 * 64; | |
| 179 | src += 4 * 64; | |
| 180 | len -= 4 * 64; | |
| 181 | } | |
| 182 | ||
| 183 | if (len >= 2 * 64) { | |
| 184 | memmove_movnt2x64b(dest, src); | |
| 185 | dest += 2 * 64; | |
| 186 | src += 2 * 64; | |
| 187 | len -= 2 * 64; | |
| 188 | } | |
| 189 | ||
| 190 | if (len >= 1 * 64) { | |
| 191 | memmove_movnt1x64b(dest, src); | |
| 192 | ||
| 193 | dest += 1 * 64; | |
| 194 | src += 1 * 64; | |
| 195 | len -= 1 * 64; | |
| 196 | } | |
| 197 | ||
| 198 | if (len == 0) | |
| 199 | return; | |
| 200 | ||
| 201 | /* There's no point in using more than 1 nt store for 1 cache line. */ | |
| 202 | if (util_is_pow2(len)) { | |
| 203 | if (len == 32) | |
| 204 | memmove_movnt1x32b(dest, src); | |
| 205 | else if (len == 16) | |
| 206 | memmove_movnt1x16b(dest, src); | |
| 207 | else if (len == 8) | |
| 208 | memmove_movnt1x8b(dest, src); | |
| 209 | else if (len == 4) | |
| 210 | memmove_movnt1x4b(dest, src); | |
| 211 | else | |
| 212 | goto nonnt; | |
| 213 | ||
| 214 | return; | |
| 215 | } | |
| 216 | ||
| 217 | nonnt: | |
| 218 | memmove_small_sse2(dest, src, len, flush); | |
| 219 | } | |
| 220 | ||
| 221 | static force_inline void | |
| 222 | memmove_movnt_sse_bw(char *dest, const char *src, size_t len, flush_fn flush, | |
| 223 | perf_barrier_fn perf_barrier) | |
| 224 | { | |
| 225 | dest += len; | |
| 226 | src += len; | |
| 227 | ||
| 228 | size_t cnt = (uint64_t)dest & 63; | |
| 229 | if (cnt > 0) { | |
| 230 | if (cnt > len) | |
| 231 | cnt = len; | |
| 232 | ||
| 233 | dest -= cnt; | |
| 234 | src -= cnt; | |
| 235 | len -= cnt; | |
| 236 | ||
| 237 | memmove_small_sse2(dest, src, cnt, flush); | |
| 238 | } | |
| 239 | ||
| 240 | const char *srcbegin = src - len; | |
| 241 | prefetch_ini_bw(src, len); | |
| 242 | ||
| 243 | while (len >= PERF_BARRIER_SIZE) { | |
| 244 | prefetch_next_bw(src, srcbegin); | |
| 245 | ||
| 246 | dest -= 4 * 64; | |
| 247 | src -= 4 * 64; | |
| 248 | len -= 4 * 64; | |
| 249 | memmove_movnt4x64b(dest, src); | |
| 250 | ||
| 251 | dest -= 4 * 64; | |
| 252 | src -= 4 * 64; | |
| 253 | len -= 4 * 64; | |
| 254 | memmove_movnt4x64b(dest, src); | |
| 255 | ||
| 256 | dest -= 4 * 64; | |
| 257 | src -= 4 * 64; | |
| 258 | len -= 4 * 64; | |
| 259 | memmove_movnt4x64b(dest, src); | |
| 260 | ||
| 261 | COMPILE_ERROR_ON(PERF_BARRIER_SIZE != (4 + 4 + 4) * 64); | |
| 262 | ||
| 263 | if (len) | |
| 264 | perf_barrier(); | |
| 265 | } | |
| 266 | ||
| 267 | while (len >= 4 * 64) { | |
| 268 | dest -= 4 * 64; | |
| 269 | src -= 4 * 64; | |
| 270 | len -= 4 * 64; | |
| 271 | memmove_movnt4x64b(dest, src); | |
| 272 | } | |
| 273 | ||
| 274 | if (len >= 2 * 64) { | |
| 275 | dest -= 2 * 64; | |
| 276 | src -= 2 * 64; | |
| 277 | len -= 2 * 64; | |
| 278 | memmove_movnt2x64b(dest, src); | |
| 279 | } | |
| 280 | ||
| 281 | if (len >= 1 * 64) { | |
| 282 | dest -= 1 * 64; | |
| 283 | src -= 1 * 64; | |
| 284 | len -= 1 * 64; | |
| 285 | memmove_movnt1x64b(dest, src); | |
| 286 | } | |
| 287 | ||
| 288 | if (len == 0) | |
| 289 | return; | |
| 290 | ||
| 291 | /* There's no point in using more than 1 nt store for 1 cache line. */ | |
| 292 | if (util_is_pow2(len)) { | |
| 293 | if (len == 32) { | |
| 294 | dest -= 32; | |
| 295 | src -= 32; | |
| 296 | memmove_movnt1x32b(dest, src); | |
| 297 | } else if (len == 16) { | |
| 298 | dest -= 16; | |
| 299 | src -= 16; | |
| 300 | memmove_movnt1x16b(dest, src); | |
| 301 | } else if (len == 8) { | |
| 302 | dest -= 8; | |
| 303 | src -= 8; | |
| 304 | memmove_movnt1x8b(dest, src); | |
| 305 | } else if (len == 4) { | |
| 306 | dest -= 4; | |
| 307 | src -= 4; | |
| 308 | memmove_movnt1x4b(dest, src); | |
| 309 | } else { | |
| 310 | goto nonnt; | |
| 311 | } | |
| 312 | ||
| 313 | return; | |
| 314 | } | |
| 315 | ||
| 316 | nonnt: | |
| 317 | dest -= len; | |
| 318 | src -= len; | |
| 319 | memmove_small_sse2(dest, src, len, flush); | |
| 320 | } | |
| 321 | ||
| 322 | static force_inline void | |
| 323 | memmove_movnt_sse2(char *dest, const char *src, size_t len, flush_fn flush, | |
| 324 | barrier_fn barrier, perf_barrier_fn perf_barrier) | |
| 325 | { | |
| 326 | if ((uintptr_t)dest - (uintptr_t)src >= len) | |
| 327 | memmove_movnt_sse_fw(dest, src, len, flush, perf_barrier); | |
| 328 | else | |
| 329 | memmove_movnt_sse_bw(dest, src, len, flush, perf_barrier); | |
| 330 | ||
| 331 | barrier(); | |
| 332 | ||
| 333 | VALGRIND_DO_FLUSH(dest, len); | |
| 334 | } | |
| 335 | ||
| 336 | /* variants without perf_barrier */ | |
| 337 | ||
| 338 | void | |
| 339 | memmove_movnt_sse2_noflush_nobarrier(char *dest, const char *src, size_t len) | |
| 340 | { | |
| 341 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 342 | ||
| 343 | memmove_movnt_sse2(dest, src, len, noflush, barrier_after_ntstores, | |
| 344 | no_barrier); | |
| 345 | } | |
| 346 | ||
| 347 | void | |
| 348 | memmove_movnt_sse2_empty_nobarrier(char *dest, const char *src, size_t len) | |
| 349 | { | |
| 350 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 351 | ||
| 352 | memmove_movnt_sse2(dest, src, len, flush_empty_nolog, | |
| 353 | barrier_after_ntstores, no_barrier); | |
| 354 | } | |
| 355 | ||
| 356 | void | |
| 357 | memmove_movnt_sse2_clflush_nobarrier(char *dest, const char *src, size_t len) | |
| 358 | { | |
| 359 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 360 | ||
| 361 | memmove_movnt_sse2(dest, src, len, flush_clflush_nolog, | |
| 362 | barrier_after_ntstores, no_barrier); | |
| 363 | } | |
| 364 | ||
| 365 | void | |
| 366 | memmove_movnt_sse2_clflushopt_nobarrier(char *dest, const char *src, size_t len) | |
| 367 | { | |
| 368 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 369 | ||
| 370 | memmove_movnt_sse2(dest, src, len, flush_clflushopt_nolog, | |
| 371 | no_barrier_after_ntstores, no_barrier); | |
| 372 | } | |
| 373 | ||
| 374 | void | |
| 375 | memmove_movnt_sse2_clwb_nobarrier(char *dest, const char *src, size_t len) | |
| 376 | { | |
| 377 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 378 | ||
| 379 | memmove_movnt_sse2(dest, src, len, flush_clwb_nolog, | |
| 380 | no_barrier_after_ntstores, no_barrier); | |
| 381 | } | |
| 382 | ||
| 383 | /* variants with perf_barrier */ | |
| 384 | ||
| 385 | void | |
| 386 | memmove_movnt_sse2_noflush_wcbarrier(char *dest, const char *src, size_t len) | |
| 387 | { | |
| 388 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 389 | ||
| 390 | memmove_movnt_sse2(dest, src, len, noflush, barrier_after_ntstores, | |
| 391 | wc_barrier); | |
| 392 | } | |
| 393 | ||
| 394 | void | |
| 395 | memmove_movnt_sse2_empty_wcbarrier(char *dest, const char *src, size_t len) | |
| 396 | { | |
| 397 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 398 | ||
| 399 | memmove_movnt_sse2(dest, src, len, flush_empty_nolog, | |
| 400 | barrier_after_ntstores, wc_barrier); | |
| 401 | } | |
| 402 | ||
| 403 | void | |
| 404 | memmove_movnt_sse2_clflush_wcbarrier(char *dest, const char *src, size_t len) | |
| 405 | { | |
| 406 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 407 | ||
| 408 | memmove_movnt_sse2(dest, src, len, flush_clflush_nolog, | |
| 409 | barrier_after_ntstores, wc_barrier); | |
| 410 | } | |
| 411 | ||
| 412 | void | |
| 413 | memmove_movnt_sse2_clflushopt_wcbarrier(char *dest, const char *src, size_t len) | |
| 414 | { | |
| 415 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 416 | ||
| 417 | memmove_movnt_sse2(dest, src, len, flush_clflushopt_nolog, | |
| 418 | no_barrier_after_ntstores, wc_barrier); | |
| 419 | } | |
| 420 | ||
| 421 | void | |
| 422 | memmove_movnt_sse2_clwb_wcbarrier(char *dest, const char *src, size_t len) | |
| 423 | { | |
| 424 | LOG(15, "dest %p src %p len %zu", dest, src, len); | |
| 425 | ||
| 426 | memmove_movnt_sse2(dest, src, len, flush_clwb_nolog, | |
| 427 | no_barrier_after_ntstores, wc_barrier); | |
| 428 | } |