| 1 | /* memcopy.h -- inline functions to copy small data chunks. |
| 2 | * For conditions of distribution and use, see copyright notice in zlib.h |
| 3 | */ |
| 4 | #ifndef MEMCOPY_H_ |
| 5 | #define MEMCOPY_H_ |
| 6 | |
| 7 | #include "zendian.h" |
| 8 | |
| 9 | /* Load 64 bits from IN and place the bytes at offset BITS in the result. */ |
| 10 | static inline uint64_t load_64_bits(const unsigned char *in, unsigned bits) { |
| 11 | uint64_t chunk; |
| 12 | memcpy(&chunk, in, sizeof(chunk)); |
| 13 | |
| 14 | #if BYTE_ORDER == LITTLE_ENDIAN |
| 15 | return chunk << bits; |
| 16 | #else |
| 17 | return ZSWAP64(chunk) << bits; |
| 18 | #endif |
| 19 | } |
| 20 | |
| 21 | #if defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 22 | #include <arm_neon.h> |
| 23 | typedef uint8x16_t inffast_chunk_t; |
| 24 | #define INFFAST_CHUNKSIZE sizeof(inffast_chunk_t) |
| 25 | #endif |
| 26 | |
| 27 | #if defined(X86_SSE2) |
| 28 | #include <immintrin.h> |
| 29 | typedef __m128i inffast_chunk_t; |
| 30 | #define INFFAST_CHUNKSIZE sizeof(inffast_chunk_t) |
| 31 | #endif |
| 32 | |
| 33 | #ifdef INFFAST_CHUNKSIZE |
| 34 | /* |
| 35 | Ask the compiler to perform a wide, unaligned load with an machine |
| 36 | instruction appropriate for the inffast_chunk_t type. |
| 37 | */ |
| 38 | static inline inffast_chunk_t loadchunk(unsigned char const* s) { |
| 39 | inffast_chunk_t c; |
| 40 | memcpy(&c, s, sizeof(c)); |
| 41 | return c; |
| 42 | } |
| 43 | |
| 44 | /* |
| 45 | Ask the compiler to perform a wide, unaligned store with an machine |
| 46 | instruction appropriate for the inffast_chunk_t type. |
| 47 | */ |
| 48 | static inline void storechunk(unsigned char* d, inffast_chunk_t c) { |
| 49 | memcpy(d, &c, sizeof(c)); |
| 50 | } |
| 51 | |
| 52 | /* |
| 53 | Behave like memcpy, but assume that it's OK to overwrite at least |
| 54 | INFFAST_CHUNKSIZE bytes of output even if the length is shorter than this, |
| 55 | that the length is non-zero, and that `from` lags `out` by at least |
| 56 | INFFAST_CHUNKSIZE bytes (or that they don't overlap at all or simply that |
| 57 | the distance is less than the length of the copy). |
| 58 | |
| 59 | Aside from better memory bus utilisation, this means that short copies |
| 60 | (INFFAST_CHUNKSIZE bytes or fewer) will fall straight through the loop |
| 61 | without iteration, which will hopefully make the branch prediction more |
| 62 | reliable. |
| 63 | */ |
| 64 | static inline unsigned char* chunkcopy(unsigned char *out, unsigned char const *from, unsigned len) { |
| 65 | --len; |
| 66 | storechunk(out, loadchunk(from)); |
| 67 | out += (len % INFFAST_CHUNKSIZE) + 1; |
| 68 | from += (len % INFFAST_CHUNKSIZE) + 1; |
| 69 | len /= INFFAST_CHUNKSIZE; |
| 70 | while (len > 0) { |
| 71 | storechunk(out, loadchunk(from)); |
| 72 | out += INFFAST_CHUNKSIZE; |
| 73 | from += INFFAST_CHUNKSIZE; |
| 74 | --len; |
| 75 | } |
| 76 | return out; |
| 77 | } |
| 78 | |
| 79 | /* |
| 80 | Behave like chunkcopy, but avoid writing beyond of legal output. |
| 81 | */ |
| 82 | static inline unsigned char* chunkcopysafe(unsigned char *out, unsigned char const *from, unsigned len, |
| 83 | unsigned char *safe) { |
| 84 | if ((safe - out) < (ptrdiff_t)INFFAST_CHUNKSIZE) { |
| 85 | if (len & 8) { |
| 86 | memcpy(out, from, 8); |
| 87 | out += 8; |
| 88 | from += 8; |
| 89 | } |
| 90 | if (len & 4) { |
| 91 | memcpy(out, from, 4); |
| 92 | out += 4; |
| 93 | from += 4; |
| 94 | } |
| 95 | if (len & 2) { |
| 96 | memcpy(out, from, 2); |
| 97 | out += 2; |
| 98 | from += 2; |
| 99 | } |
| 100 | if (len & 1) { |
| 101 | *out++ = *from++; |
| 102 | } |
| 103 | return out; |
| 104 | } |
| 105 | return chunkcopy(out, from, len); |
| 106 | } |
| 107 | |
| 108 | /* |
| 109 | Perform short copies until distance can be rewritten as being at least |
| 110 | INFFAST_CHUNKSIZE. |
| 111 | |
| 112 | This assumes that it's OK to overwrite at least the first |
| 113 | 2*INFFAST_CHUNKSIZE bytes of output even if the copy is shorter than this. |
| 114 | This assumption holds because inflate_fast() starts every iteration with at |
| 115 | least 258 bytes of output space available (258 being the maximum length |
| 116 | output from a single token; see inflate_fast()'s assumptions below). |
| 117 | */ |
| 118 | static inline unsigned char* chunkunroll(unsigned char *out, unsigned *dist, unsigned *len) { |
| 119 | unsigned char const *from = out - *dist; |
| 120 | while (*dist < *len && *dist < INFFAST_CHUNKSIZE) { |
| 121 | storechunk(out, loadchunk(from)); |
| 122 | out += *dist; |
| 123 | *len -= *dist; |
| 124 | *dist += *dist; |
| 125 | } |
| 126 | return out; |
| 127 | } |
| 128 | |
| 129 | static inline inffast_chunk_t chunkmemset_1(unsigned char *from) { |
| 130 | #if defined(X86_SSE2) |
| 131 | int8_t c; |
| 132 | memcpy(&c, from, sizeof(c)); |
| 133 | return _mm_set1_epi8(c); |
| 134 | #elif defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 135 | return vld1q_dup_u8(from); |
| 136 | #endif |
| 137 | } |
| 138 | |
| 139 | static inline inffast_chunk_t chunkmemset_2(unsigned char *from) { |
| 140 | int16_t c; |
| 141 | memcpy(&c, from, sizeof(c)); |
| 142 | #if defined(X86_SSE2) |
| 143 | return _mm_set1_epi16(c); |
| 144 | #elif defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 145 | return vreinterpretq_u8_s16(vdupq_n_s16(c)); |
| 146 | #endif |
| 147 | } |
| 148 | |
| 149 | static inline inffast_chunk_t chunkmemset_4(unsigned char *from) { |
| 150 | int32_t c; |
| 151 | memcpy(&c, from, sizeof(c)); |
| 152 | #if defined(X86_SSE2) |
| 153 | return _mm_set1_epi32(c); |
| 154 | #elif defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 155 | return vreinterpretq_u8_s32(vdupq_n_s32(c)); |
| 156 | #endif |
| 157 | } |
| 158 | |
| 159 | static inline inffast_chunk_t chunkmemset_8(unsigned char *from) { |
| 160 | #if defined(X86_SSE2) |
| 161 | int64_t c; |
| 162 | memcpy(&c, from, sizeof(c)); |
| 163 | return _mm_set1_epi64x(c); |
| 164 | #elif defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 165 | return vcombine_u8(vld1_u8(from), vld1_u8(from)); |
| 166 | #endif |
| 167 | } |
| 168 | |
| 169 | #if defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 170 | static inline unsigned char *chunkmemset_3(unsigned char *out, unsigned char *from, unsigned dist, unsigned len) { |
| 171 | uint8x8x3_t chunks; |
| 172 | unsigned sz = sizeof(chunks); |
| 173 | if (len < sz) { |
| 174 | out = chunkunroll(out, &dist, &len); |
| 175 | return chunkcopy(out, out - dist, len); |
| 176 | } |
| 177 | |
| 178 | /* Load 3 bytes 'a,b,c' from FROM and duplicate across all lanes: |
| 179 | chunks[0] = {a,a,a,a,a,a,a,a} |
| 180 | chunks[1] = {b,b,b,b,b,b,b,b} |
| 181 | chunks[2] = {c,c,c,c,c,c,c,c}. */ |
| 182 | chunks = vld3_dup_u8(from); |
| 183 | |
| 184 | unsigned rem = len % sz; |
| 185 | len -= rem; |
| 186 | while (len) { |
| 187 | /* Store "a,b,c, ..., a,b,c". */ |
| 188 | vst3_u8(out, chunks); |
| 189 | out += sz; |
| 190 | len -= sz; |
| 191 | } |
| 192 | |
| 193 | if (!rem) |
| 194 | return out; |
| 195 | |
| 196 | /* Last, deal with the case when LEN is not a multiple of SZ. */ |
| 197 | out = chunkunroll(out, &dist, &rem); |
| 198 | return chunkcopy(out, out - dist, rem); |
| 199 | } |
| 200 | #endif |
| 201 | |
| 202 | #if defined(__aarch64__) || defined(_M_ARM64) |
| 203 | static inline unsigned char *chunkmemset_6(unsigned char *out, unsigned char *from, unsigned dist, unsigned len) { |
| 204 | uint16x8x3_t chunks; |
| 205 | unsigned sz = sizeof(chunks); |
| 206 | if (len < sz) { |
| 207 | out = chunkunroll(out, &dist, &len); |
| 208 | return chunkcopy(out, out - dist, len); |
| 209 | } |
| 210 | |
| 211 | /* Load 6 bytes 'ab,cd,ef' from FROM and duplicate across all lanes: |
| 212 | chunks[0] = {ab,ab,ab,ab,ab,ab,ab,ab} |
| 213 | chunks[1] = {cd,cd,cd,cd,cd,cd,cd,cd} |
| 214 | chunks[2] = {ef,ef,ef,ef,ef,ef,ef,ef}. */ |
| 215 | chunks = vld3q_dup_u16((unsigned short *)from); |
| 216 | |
| 217 | unsigned rem = len % sz; |
| 218 | len -= rem; |
| 219 | while (len) { |
| 220 | /* Store "ab,cd,ef, ..., ab,cd,ef". */ |
| 221 | vst3q_u16((unsigned short *)out, chunks); |
| 222 | out += sz; |
| 223 | len -= sz; |
| 224 | } |
| 225 | |
| 226 | if (!rem) |
| 227 | return out; |
| 228 | |
| 229 | /* Last, deal with the case when LEN is not a multiple of SZ. */ |
| 230 | out = chunkunroll(out, &dist, &rem); |
| 231 | return chunkcopy(out, out - dist, rem); |
| 232 | } |
| 233 | #endif |
| 234 | |
| 235 | /* Copy DIST bytes from OUT - DIST into OUT + DIST * k, for 0 <= k < LEN/DIST. Return OUT + LEN. */ |
| 236 | static inline unsigned char *chunkmemset(unsigned char *out, unsigned dist, unsigned len) { |
| 237 | /* Debug performance related issues when len < sizeof(uint64_t): |
| 238 | Assert(len >= sizeof(uint64_t), "chunkmemset should be called on larger chunks"); */ |
| 239 | Assert(dist > 0, "cannot have a distance 0" ); |
| 240 | |
| 241 | unsigned char *from = out - dist; |
| 242 | inffast_chunk_t chunk; |
| 243 | unsigned sz = sizeof(chunk); |
| 244 | if (len < sz) { |
| 245 | do { |
| 246 | *out++ = *from++; |
| 247 | --len; |
| 248 | } while (len != 0); |
| 249 | return out; |
| 250 | } |
| 251 | |
| 252 | switch (dist) { |
| 253 | case 1: { |
| 254 | chunk = chunkmemset_1(from); |
| 255 | break; |
| 256 | } |
| 257 | case 2: { |
| 258 | chunk = chunkmemset_2(from); |
| 259 | break; |
| 260 | } |
| 261 | #if defined(__ARM_NEON__) || defined(__ARM_NEON) |
| 262 | case 3: |
| 263 | return chunkmemset_3(out, from, dist, len); |
| 264 | #endif |
| 265 | case 4: { |
| 266 | chunk = chunkmemset_4(from); |
| 267 | break; |
| 268 | } |
| 269 | #if defined(__aarch64__) || defined(_M_ARM64) |
| 270 | case 6: |
| 271 | return chunkmemset_6(out, from, dist, len); |
| 272 | #endif |
| 273 | case 8: { |
| 274 | chunk = chunkmemset_8(from); |
| 275 | break; |
| 276 | } |
| 277 | case 16: |
| 278 | memcpy(&chunk, from, sz); |
| 279 | break; |
| 280 | |
| 281 | default: |
| 282 | out = chunkunroll(out, &dist, &len); |
| 283 | return chunkcopy(out, out - dist, len); |
| 284 | } |
| 285 | |
| 286 | unsigned rem = len % sz; |
| 287 | len -= rem; |
| 288 | while (len) { |
| 289 | memcpy(out, &chunk, sz); |
| 290 | out += sz; |
| 291 | len -= sz; |
| 292 | } |
| 293 | |
| 294 | /* Last, deal with the case when LEN is not a multiple of SZ. */ |
| 295 | if (rem) |
| 296 | memcpy(out, &chunk, rem); |
| 297 | out += rem; |
| 298 | return out; |
| 299 | } |
| 300 | |
| 301 | static inline unsigned char* chunkmemsetsafe(unsigned char *out, unsigned dist, unsigned len, unsigned left) { |
| 302 | if (left < (unsigned)(3 * INFFAST_CHUNKSIZE)) { |
| 303 | while (len > 0) { |
| 304 | *out = *(out - dist); |
| 305 | out++; |
| 306 | --len; |
| 307 | } |
| 308 | return out; |
| 309 | } |
| 310 | |
| 311 | return chunkmemset(out, dist, len); |
| 312 | } |
| 313 | |
| 314 | #else /* INFFAST_CHUNKSIZE */ |
| 315 | |
| 316 | static inline unsigned char *copy_1_bytes(unsigned char *out, unsigned char *from) { |
| 317 | *out++ = *from; |
| 318 | return out; |
| 319 | } |
| 320 | |
| 321 | static inline unsigned char *copy_2_bytes(unsigned char *out, unsigned char *from) { |
| 322 | uint16_t chunk; |
| 323 | unsigned sz = sizeof(chunk); |
| 324 | memcpy(&chunk, from, sz); |
| 325 | memcpy(out, &chunk, sz); |
| 326 | return out + sz; |
| 327 | } |
| 328 | |
| 329 | static inline unsigned char *copy_3_bytes(unsigned char *out, unsigned char *from) { |
| 330 | out = copy_1_bytes(out, from); |
| 331 | return copy_2_bytes(out, from + 1); |
| 332 | } |
| 333 | |
| 334 | static inline unsigned char *copy_4_bytes(unsigned char *out, unsigned char *from) { |
| 335 | uint32_t chunk; |
| 336 | unsigned sz = sizeof(chunk); |
| 337 | memcpy(&chunk, from, sz); |
| 338 | memcpy(out, &chunk, sz); |
| 339 | return out + sz; |
| 340 | } |
| 341 | |
| 342 | static inline unsigned char *copy_5_bytes(unsigned char *out, unsigned char *from) { |
| 343 | out = copy_1_bytes(out, from); |
| 344 | return copy_4_bytes(out, from + 1); |
| 345 | } |
| 346 | |
| 347 | static inline unsigned char *copy_6_bytes(unsigned char *out, unsigned char *from) { |
| 348 | out = copy_2_bytes(out, from); |
| 349 | return copy_4_bytes(out, from + 2); |
| 350 | } |
| 351 | |
| 352 | static inline unsigned char *copy_7_bytes(unsigned char *out, unsigned char *from) { |
| 353 | out = copy_3_bytes(out, from); |
| 354 | return copy_4_bytes(out, from + 3); |
| 355 | } |
| 356 | |
| 357 | static inline unsigned char *copy_8_bytes(unsigned char *out, unsigned char *from) { |
| 358 | uint64_t chunk; |
| 359 | unsigned sz = sizeof(chunk); |
| 360 | memcpy(&chunk, from, sz); |
| 361 | memcpy(out, &chunk, sz); |
| 362 | return out + sz; |
| 363 | } |
| 364 | |
| 365 | /* Copy LEN bytes (7 or fewer) from FROM into OUT. Return OUT + LEN. */ |
| 366 | static inline unsigned char *copy_bytes(unsigned char *out, unsigned char *from, unsigned len) { |
| 367 | Assert(len < 8, "copy_bytes should be called with less than 8 bytes" ); |
| 368 | |
| 369 | #ifndef UNALIGNED_OK |
| 370 | while (len--) { |
| 371 | *out++ = *from++; |
| 372 | } |
| 373 | return out; |
| 374 | #else |
| 375 | switch (len) { |
| 376 | case 7: |
| 377 | return copy_7_bytes(out, from); |
| 378 | case 6: |
| 379 | return copy_6_bytes(out, from); |
| 380 | case 5: |
| 381 | return copy_5_bytes(out, from); |
| 382 | case 4: |
| 383 | return copy_4_bytes(out, from); |
| 384 | case 3: |
| 385 | return copy_3_bytes(out, from); |
| 386 | case 2: |
| 387 | return copy_2_bytes(out, from); |
| 388 | case 1: |
| 389 | return copy_1_bytes(out, from); |
| 390 | case 0: |
| 391 | return out; |
| 392 | default: |
| 393 | Assert(0, "should not happen" ); |
| 394 | } |
| 395 | |
| 396 | return out; |
| 397 | #endif /* UNALIGNED_OK */ |
| 398 | } |
| 399 | |
| 400 | /* Copy LEN bytes (7 or fewer) from FROM into OUT. Return OUT + LEN. */ |
| 401 | static inline unsigned char *set_bytes(unsigned char *out, unsigned char *from, unsigned dist, unsigned len) { |
| 402 | Assert(len < 8, "set_bytes should be called with less than 8 bytes" ); |
| 403 | |
| 404 | #ifndef UNALIGNED_OK |
| 405 | (void)dist; |
| 406 | while (len--) { |
| 407 | *out++ = *from++; |
| 408 | } |
| 409 | return out; |
| 410 | #else |
| 411 | if (dist >= len) |
| 412 | return copy_bytes(out, from, len); |
| 413 | |
| 414 | switch (dist) { |
| 415 | case 6: |
| 416 | Assert(len == 7, "len should be exactly 7" ); |
| 417 | out = copy_6_bytes(out, from); |
| 418 | return copy_1_bytes(out, from); |
| 419 | |
| 420 | case 5: |
| 421 | Assert(len == 6 || len == 7, "len should be either 6 or 7" ); |
| 422 | out = copy_5_bytes(out, from); |
| 423 | return copy_bytes(out, from, len - 5); |
| 424 | |
| 425 | case 4: |
| 426 | Assert(len == 5 || len == 6 || len == 7, "len should be either 5, 6, or 7" ); |
| 427 | out = copy_4_bytes(out, from); |
| 428 | return copy_bytes(out, from, len - 4); |
| 429 | |
| 430 | case 3: |
| 431 | Assert(4 <= len && len <= 7, "len should be between 4 and 7" ); |
| 432 | out = copy_3_bytes(out, from); |
| 433 | switch (len) { |
| 434 | case 7: |
| 435 | return copy_4_bytes(out, from); |
| 436 | case 6: |
| 437 | return copy_3_bytes(out, from); |
| 438 | case 5: |
| 439 | return copy_2_bytes(out, from); |
| 440 | case 4: |
| 441 | return copy_1_bytes(out, from); |
| 442 | default: |
| 443 | Assert(0, "should not happen" ); |
| 444 | break; |
| 445 | } |
| 446 | |
| 447 | case 2: |
| 448 | Assert(3 <= len && len <= 7, "len should be between 3 and 7" ); |
| 449 | out = copy_2_bytes(out, from); |
| 450 | switch (len) { |
| 451 | case 7: |
| 452 | out = copy_4_bytes(out, from); |
| 453 | out = copy_1_bytes(out, from); |
| 454 | return out; |
| 455 | case 6: |
| 456 | out = copy_4_bytes(out, from); |
| 457 | return out; |
| 458 | case 5: |
| 459 | out = copy_2_bytes(out, from); |
| 460 | out = copy_1_bytes(out, from); |
| 461 | return out; |
| 462 | case 4: |
| 463 | out = copy_2_bytes(out, from); |
| 464 | return out; |
| 465 | case 3: |
| 466 | out = copy_1_bytes(out, from); |
| 467 | return out; |
| 468 | default: |
| 469 | Assert(0, "should not happen" ); |
| 470 | break; |
| 471 | } |
| 472 | |
| 473 | case 1: |
| 474 | Assert(2 <= len && len <= 7, "len should be between 2 and 7" ); |
| 475 | unsigned char c = *from; |
| 476 | switch (len) { |
| 477 | case 7: |
| 478 | memset(out, c, 7); |
| 479 | return out + 7; |
| 480 | case 6: |
| 481 | memset(out, c, 6); |
| 482 | return out + 6; |
| 483 | case 5: |
| 484 | memset(out, c, 5); |
| 485 | return out + 5; |
| 486 | case 4: |
| 487 | memset(out, c, 4); |
| 488 | return out + 4; |
| 489 | case 3: |
| 490 | memset(out, c, 3); |
| 491 | return out + 3; |
| 492 | case 2: |
| 493 | memset(out, c, 2); |
| 494 | return out + 2; |
| 495 | default: |
| 496 | Assert(0, "should not happen" ); |
| 497 | break; |
| 498 | } |
| 499 | } |
| 500 | return out; |
| 501 | #endif /* UNALIGNED_OK */ |
| 502 | } |
| 503 | |
| 504 | /* Byte by byte semantics: copy LEN bytes from OUT + DIST and write them to OUT. Return OUT + LEN. */ |
| 505 | static inline unsigned char *chunk_memcpy(unsigned char *out, unsigned char *from, unsigned len) { |
| 506 | unsigned sz = sizeof(uint64_t); |
| 507 | Assert(len >= sz, "chunk_memcpy should be called on larger chunks" ); |
| 508 | |
| 509 | /* Copy a few bytes to make sure the loop below has a multiple of SZ bytes to be copied. */ |
| 510 | copy_8_bytes(out, from); |
| 511 | |
| 512 | unsigned rem = len % sz; |
| 513 | len /= sz; |
| 514 | out += rem; |
| 515 | from += rem; |
| 516 | |
| 517 | unsigned by8 = len % sz; |
| 518 | len -= by8; |
| 519 | switch (by8) { |
| 520 | case 7: |
| 521 | out = copy_8_bytes(out, from); |
| 522 | from += sz; |
| 523 | case 6: |
| 524 | out = copy_8_bytes(out, from); |
| 525 | from += sz; |
| 526 | case 5: |
| 527 | out = copy_8_bytes(out, from); |
| 528 | from += sz; |
| 529 | case 4: |
| 530 | out = copy_8_bytes(out, from); |
| 531 | from += sz; |
| 532 | case 3: |
| 533 | out = copy_8_bytes(out, from); |
| 534 | from += sz; |
| 535 | case 2: |
| 536 | out = copy_8_bytes(out, from); |
| 537 | from += sz; |
| 538 | case 1: |
| 539 | out = copy_8_bytes(out, from); |
| 540 | from += sz; |
| 541 | } |
| 542 | |
| 543 | while (len) { |
| 544 | out = copy_8_bytes(out, from); |
| 545 | from += sz; |
| 546 | out = copy_8_bytes(out, from); |
| 547 | from += sz; |
| 548 | out = copy_8_bytes(out, from); |
| 549 | from += sz; |
| 550 | out = copy_8_bytes(out, from); |
| 551 | from += sz; |
| 552 | out = copy_8_bytes(out, from); |
| 553 | from += sz; |
| 554 | out = copy_8_bytes(out, from); |
| 555 | from += sz; |
| 556 | out = copy_8_bytes(out, from); |
| 557 | from += sz; |
| 558 | out = copy_8_bytes(out, from); |
| 559 | from += sz; |
| 560 | |
| 561 | len -= 8; |
| 562 | } |
| 563 | |
| 564 | return out; |
| 565 | } |
| 566 | |
| 567 | /* Memset LEN bytes in OUT with the value at OUT - 1. Return OUT + LEN. */ |
| 568 | static inline unsigned char *byte_memset(unsigned char *out, unsigned len) { |
| 569 | unsigned sz = sizeof(uint64_t); |
| 570 | Assert(len >= sz, "byte_memset should be called on larger chunks" ); |
| 571 | |
| 572 | unsigned char *from = out - 1; |
| 573 | unsigned char c = *from; |
| 574 | |
| 575 | /* First, deal with the case when LEN is not a multiple of SZ. */ |
| 576 | memset(out, c, sz); |
| 577 | unsigned rem = len % sz; |
| 578 | len /= sz; |
| 579 | out += rem; |
| 580 | |
| 581 | unsigned by8 = len % 8; |
| 582 | len -= by8; |
| 583 | switch (by8) { |
| 584 | case 7: |
| 585 | memset(out, c, sz); |
| 586 | out += sz; |
| 587 | case 6: |
| 588 | memset(out, c, sz); |
| 589 | out += sz; |
| 590 | case 5: |
| 591 | memset(out, c, sz); |
| 592 | out += sz; |
| 593 | case 4: |
| 594 | memset(out, c, sz); |
| 595 | out += sz; |
| 596 | case 3: |
| 597 | memset(out, c, sz); |
| 598 | out += sz; |
| 599 | case 2: |
| 600 | memset(out, c, sz); |
| 601 | out += sz; |
| 602 | case 1: |
| 603 | memset(out, c, sz); |
| 604 | out += sz; |
| 605 | } |
| 606 | |
| 607 | while (len) { |
| 608 | /* When sz is a constant, the compiler replaces __builtin_memset with an |
| 609 | inline version that does not incur a function call overhead. */ |
| 610 | memset(out, c, sz); |
| 611 | out += sz; |
| 612 | memset(out, c, sz); |
| 613 | out += sz; |
| 614 | memset(out, c, sz); |
| 615 | out += sz; |
| 616 | memset(out, c, sz); |
| 617 | out += sz; |
| 618 | memset(out, c, sz); |
| 619 | out += sz; |
| 620 | memset(out, c, sz); |
| 621 | out += sz; |
| 622 | memset(out, c, sz); |
| 623 | out += sz; |
| 624 | memset(out, c, sz); |
| 625 | out += sz; |
| 626 | len -= 8; |
| 627 | } |
| 628 | |
| 629 | return out; |
| 630 | } |
| 631 | |
| 632 | /* Copy DIST bytes from OUT - DIST into OUT + DIST * k, for 0 <= k < LEN/DIST. Return OUT + LEN. */ |
| 633 | static inline unsigned char *chunk_memset(unsigned char *out, unsigned char *from, unsigned dist, unsigned len) { |
| 634 | if (dist >= len) |
| 635 | return chunk_memcpy(out, from, len); |
| 636 | |
| 637 | Assert(len >= sizeof(uint64_t), "chunk_memset should be called on larger chunks" ); |
| 638 | |
| 639 | /* Double up the size of the memset pattern until reaching the largest pattern of size less than SZ. */ |
| 640 | unsigned sz = sizeof(uint64_t); |
| 641 | while (dist < len && dist < sz) { |
| 642 | copy_8_bytes(out, from); |
| 643 | |
| 644 | out += dist; |
| 645 | len -= dist; |
| 646 | dist += dist; |
| 647 | |
| 648 | /* Make sure the next memcpy has at least SZ bytes to be copied. */ |
| 649 | if (len < sz) |
| 650 | /* Finish up byte by byte when there are not enough bytes left. */ |
| 651 | return set_bytes(out, from, dist, len); |
| 652 | } |
| 653 | |
| 654 | return chunk_memcpy(out, from, len); |
| 655 | } |
| 656 | |
| 657 | /* Byte by byte semantics: copy LEN bytes from FROM and write them to OUT. Return OUT + LEN. */ |
| 658 | static inline unsigned char *chunk_copy(unsigned char *out, unsigned char *from, int dist, unsigned len) { |
| 659 | if (len < sizeof(uint64_t)) { |
| 660 | if (dist > 0) |
| 661 | return set_bytes(out, from, dist, len); |
| 662 | |
| 663 | return copy_bytes(out, from, len); |
| 664 | } |
| 665 | |
| 666 | if (dist == 1) |
| 667 | return byte_memset(out, len); |
| 668 | |
| 669 | if (dist > 0) |
| 670 | return chunk_memset(out, from, dist, len); |
| 671 | |
| 672 | return chunk_memcpy(out, from, len); |
| 673 | } |
| 674 | #endif /* INFFAST_CHUNKSIZE */ |
| 675 | #endif /* MEMCOPY_H_ */ |
| 676 | |