| 1 | /* |
| 2 | * Copyright (c) 2015-2016, Intel Corporation |
| 3 | * |
| 4 | * Redistribution and use in source and binary forms, with or without |
| 5 | * modification, are permitted provided that the following conditions are met: |
| 6 | * |
| 7 | * * Redistributions of source code must retain the above copyright notice, |
| 8 | * this list of conditions and the following disclaimer. |
| 9 | * * Redistributions in binary form must reproduce the above copyright |
| 10 | * notice, this list of conditions and the following disclaimer in the |
| 11 | * documentation and/or other materials provided with the distribution. |
| 12 | * * Neither the name of Intel Corporation nor the names of its contributors |
| 13 | * may be used to endorse or promote products derived from this software |
| 14 | * without specific prior written permission. |
| 15 | * |
| 16 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 17 | * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 18 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 19 | * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE |
| 20 | * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 21 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 22 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 23 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 24 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 25 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 26 | * POSSIBILITY OF SUCH DAMAGE. |
| 27 | */ |
| 28 | |
| 29 | /** \file |
| 30 | * \brief Vermicelli: single-byte and double-byte acceleration. |
| 31 | */ |
| 32 | |
| 33 | #ifndef VERMICELLI_H |
| 34 | #define VERMICELLI_H |
| 35 | |
| 36 | #include "util/bitutils.h" |
| 37 | #include "util/simd_utils.h" |
| 38 | #include "util/unaligned.h" |
| 39 | |
| 40 | #include "vermicelli_sse.h" |
| 41 | |
| 42 | static really_inline |
| 43 | const u8 *vermicelliExec(char c, char nocase, const u8 *buf, |
| 44 | const u8 *buf_end) { |
| 45 | DEBUG_PRINTF("verm scan %s\\x%02hhx over %zu bytes\n" , |
| 46 | nocase ? "nocase " : "" , c, (size_t)(buf_end - buf)); |
| 47 | assert(buf < buf_end); |
| 48 | |
| 49 | // Handle small scans. |
| 50 | if (buf_end - buf < VERM_BOUNDARY) { |
| 51 | for (; buf < buf_end; buf++) { |
| 52 | char cur = (char)*buf; |
| 53 | if (nocase) { |
| 54 | cur &= CASE_CLEAR; |
| 55 | } |
| 56 | if (cur == c) { |
| 57 | break; |
| 58 | } |
| 59 | } |
| 60 | return buf; |
| 61 | } |
| 62 | |
| 63 | VERM_TYPE chars = VERM_SET_FN(c); /* nocase already uppercase */ |
| 64 | uintptr_t min = (uintptr_t)buf % VERM_BOUNDARY; |
| 65 | if (min) { |
| 66 | // Input isn't aligned, so we need to run one iteration with an |
| 67 | // unaligned load, then skip buf forward to the next aligned address. |
| 68 | // There's some small overlap here, but we don't mind scanning it twice |
| 69 | // if we can do it quickly, do we? |
| 70 | const u8 *ptr = nocase ? vermUnalignNocase(chars, buf, 0) |
| 71 | : vermUnalign(chars, buf, 0); |
| 72 | if (ptr) { |
| 73 | return ptr; |
| 74 | } |
| 75 | |
| 76 | buf += VERM_BOUNDARY - min; |
| 77 | assert(buf < buf_end); |
| 78 | } |
| 79 | |
| 80 | // Aligned loops from here on in |
| 81 | const u8 *ptr = nocase ? vermSearchAlignedNocase(chars, buf, buf_end - 1, 0) |
| 82 | : vermSearchAligned(chars, buf, buf_end - 1, 0); |
| 83 | if (ptr) { |
| 84 | return ptr; |
| 85 | } |
| 86 | |
| 87 | // Tidy up the mess at the end |
| 88 | ptr = nocase ? vermUnalignNocase(chars, buf_end - VERM_BOUNDARY, 0) |
| 89 | : vermUnalign(chars, buf_end - VERM_BOUNDARY, 0); |
| 90 | return ptr ? ptr : buf_end; |
| 91 | } |
| 92 | |
| 93 | /* like vermicelliExec except returns the address of the first character which |
| 94 | * is not c */ |
| 95 | static really_inline |
| 96 | const u8 *nvermicelliExec(char c, char nocase, const u8 *buf, |
| 97 | const u8 *buf_end) { |
| 98 | DEBUG_PRINTF("nverm scan %s\\x%02hhx over %zu bytes\n" , |
| 99 | nocase ? "nocase " : "" , c, (size_t)(buf_end - buf)); |
| 100 | assert(buf < buf_end); |
| 101 | |
| 102 | // Handle small scans. |
| 103 | if (buf_end - buf < VERM_BOUNDARY) { |
| 104 | for (; buf < buf_end; buf++) { |
| 105 | char cur = (char)*buf; |
| 106 | if (nocase) { |
| 107 | cur &= CASE_CLEAR; |
| 108 | } |
| 109 | if (cur != c) { |
| 110 | break; |
| 111 | } |
| 112 | } |
| 113 | return buf; |
| 114 | } |
| 115 | |
| 116 | VERM_TYPE chars = VERM_SET_FN(c); /* nocase already uppercase */ |
| 117 | size_t min = (size_t)buf % VERM_BOUNDARY; |
| 118 | if (min) { |
| 119 | // Input isn't aligned, so we need to run one iteration with an |
| 120 | // unaligned load, then skip buf forward to the next aligned address. |
| 121 | // There's some small overlap here, but we don't mind scanning it twice |
| 122 | // if we can do it quickly, do we? |
| 123 | const u8 *ptr = nocase ? vermUnalignNocase(chars, buf, 1) |
| 124 | : vermUnalign(chars, buf, 1); |
| 125 | if (ptr) { |
| 126 | return ptr; |
| 127 | } |
| 128 | |
| 129 | buf += VERM_BOUNDARY - min; |
| 130 | assert(buf < buf_end); |
| 131 | } |
| 132 | |
| 133 | // Aligned loops from here on in |
| 134 | const u8 *ptr = nocase ? vermSearchAlignedNocase(chars, buf, buf_end - 1, 1) |
| 135 | : vermSearchAligned(chars, buf, buf_end - 1, 1); |
| 136 | if (ptr) { |
| 137 | return ptr; |
| 138 | } |
| 139 | |
| 140 | // Tidy up the mess at the end |
| 141 | ptr = nocase ? vermUnalignNocase(chars, buf_end - VERM_BOUNDARY, 1) |
| 142 | : vermUnalign(chars, buf_end - VERM_BOUNDARY, 1); |
| 143 | return ptr ? ptr : buf_end; |
| 144 | } |
| 145 | |
| 146 | static really_inline |
| 147 | const u8 *vermicelliDoubleExec(char c1, char c2, char nocase, const u8 *buf, |
| 148 | const u8 *buf_end) { |
| 149 | DEBUG_PRINTF("double verm scan %s\\x%02hhx%02hhx over %zu bytes\n" , |
| 150 | nocase ? "nocase " : "" , c1, c2, (size_t)(buf_end - buf)); |
| 151 | assert(buf < buf_end); |
| 152 | assert((buf_end - buf) >= VERM_BOUNDARY); |
| 153 | |
| 154 | uintptr_t min = (uintptr_t)buf % VERM_BOUNDARY; |
| 155 | VERM_TYPE chars1 = VERM_SET_FN(c1); /* nocase already uppercase */ |
| 156 | VERM_TYPE chars2 = VERM_SET_FN(c2); /* nocase already uppercase */ |
| 157 | |
| 158 | if (min) { |
| 159 | // Input isn't aligned, so we need to run one iteration with an |
| 160 | // unaligned load, then skip buf forward to the next aligned address. |
| 161 | // There's some small overlap here, but we don't mind scanning it twice |
| 162 | // if we can do it quickly, do we? |
| 163 | const u8 *ptr = nocase |
| 164 | ? dvermPreconditionNocase(chars1, chars2, buf) |
| 165 | : dvermPrecondition(chars1, chars2, buf); |
| 166 | if (ptr) { |
| 167 | return ptr; |
| 168 | } |
| 169 | |
| 170 | buf += VERM_BOUNDARY - min; |
| 171 | assert(buf < buf_end); |
| 172 | } |
| 173 | |
| 174 | // Aligned loops from here on in |
| 175 | const u8 *ptr = nocase ? dvermSearchAlignedNocase(chars1, chars2, c1, c2, |
| 176 | buf, buf_end) |
| 177 | : dvermSearchAligned(chars1, chars2, c1, c2, buf, |
| 178 | buf_end); |
| 179 | if (ptr) { |
| 180 | return ptr; |
| 181 | } |
| 182 | |
| 183 | // Tidy up the mess at the end |
| 184 | ptr = nocase ? dvermPreconditionNocase(chars1, chars2, |
| 185 | buf_end - VERM_BOUNDARY) |
| 186 | : dvermPrecondition(chars1, chars2, buf_end - VERM_BOUNDARY); |
| 187 | |
| 188 | if (ptr) { |
| 189 | return ptr; |
| 190 | } |
| 191 | |
| 192 | /* check for partial match at end */ |
| 193 | u8 mask = nocase ? CASE_CLEAR : 0xff; |
| 194 | if ((buf_end[-1] & mask) == (u8)c1) { |
| 195 | DEBUG_PRINTF("partial!!!\n" ); |
| 196 | return buf_end - 1; |
| 197 | } |
| 198 | |
| 199 | return buf_end; |
| 200 | } |
| 201 | |
| 202 | static really_inline |
| 203 | const u8 *vermicelliDoubleMaskedExec(char c1, char c2, char m1, char m2, |
| 204 | const u8 *buf, const u8 *buf_end) { |
| 205 | DEBUG_PRINTF("double verm scan (\\x%02hhx&\\x%02hhx)(\\x%02hhx&\\x%02hhx) " |
| 206 | "over %zu bytes\n" , c1, m1, c2, m2, (size_t)(buf_end - buf)); |
| 207 | assert(buf < buf_end); |
| 208 | assert((buf_end - buf) >= VERM_BOUNDARY); |
| 209 | |
| 210 | uintptr_t min = (uintptr_t)buf % VERM_BOUNDARY; |
| 211 | VERM_TYPE chars1 = VERM_SET_FN(c1); |
| 212 | VERM_TYPE chars2 = VERM_SET_FN(c2); |
| 213 | VERM_TYPE mask1 = VERM_SET_FN(m1); |
| 214 | VERM_TYPE mask2 = VERM_SET_FN(m2); |
| 215 | |
| 216 | if (min) { |
| 217 | // Input isn't aligned, so we need to run one iteration with an |
| 218 | // unaligned load, then skip buf forward to the next aligned address. |
| 219 | // There's some small overlap here, but we don't mind scanning it twice |
| 220 | // if we can do it quickly, do we? |
| 221 | const u8 *p = dvermPreconditionMasked(chars1, chars2, mask1, mask2, buf); |
| 222 | if (p) { |
| 223 | return p; |
| 224 | } |
| 225 | |
| 226 | buf += VERM_BOUNDARY - min; |
| 227 | assert(buf < buf_end); |
| 228 | } |
| 229 | |
| 230 | // Aligned loops from here on in |
| 231 | const u8 *ptr = dvermSearchAlignedMasked(chars1, chars2, mask1, mask2, c1, |
| 232 | c2, m1, m2, buf, buf_end); |
| 233 | if (ptr) { |
| 234 | return ptr; |
| 235 | } |
| 236 | |
| 237 | // Tidy up the mess at the end |
| 238 | ptr = dvermPreconditionMasked(chars1, chars2, mask1, mask2, |
| 239 | buf_end - VERM_BOUNDARY); |
| 240 | |
| 241 | if (ptr) { |
| 242 | return ptr; |
| 243 | } |
| 244 | |
| 245 | /* check for partial match at end */ |
| 246 | if ((buf_end[-1] & m1) == (u8)c1) { |
| 247 | return buf_end - 1; |
| 248 | } |
| 249 | |
| 250 | return buf_end; |
| 251 | } |
| 252 | |
| 253 | // Reverse vermicelli scan. Provides exact semantics and returns (buf - 1) if |
| 254 | // character not found. |
| 255 | static really_inline |
| 256 | const u8 *rvermicelliExec(char c, char nocase, const u8 *buf, |
| 257 | const u8 *buf_end) { |
| 258 | DEBUG_PRINTF("rev verm scan %s\\x%02hhx over %zu bytes\n" , |
| 259 | nocase ? "nocase " : "" , c, (size_t)(buf_end - buf)); |
| 260 | assert(buf < buf_end); |
| 261 | |
| 262 | // Handle small scans. |
| 263 | if (buf_end - buf < VERM_BOUNDARY) { |
| 264 | for (buf_end--; buf_end >= buf; buf_end--) { |
| 265 | char cur = (char)*buf_end; |
| 266 | if (nocase) { |
| 267 | cur &= CASE_CLEAR; |
| 268 | } |
| 269 | if (cur == c) { |
| 270 | break; |
| 271 | } |
| 272 | } |
| 273 | return buf_end; |
| 274 | } |
| 275 | |
| 276 | VERM_TYPE chars = VERM_SET_FN(c); /* nocase already uppercase */ |
| 277 | size_t min = (size_t)buf_end % VERM_BOUNDARY; |
| 278 | |
| 279 | if (min) { |
| 280 | // Input isn't aligned, so we need to run one iteration with an |
| 281 | // unaligned load, then skip buf backward to the next aligned address. |
| 282 | // There's some small overlap here, but we don't mind scanning it twice |
| 283 | // if we can do it quickly, do we? |
| 284 | if (nocase) { |
| 285 | const u8 *ptr = |
| 286 | rvermUnalignNocase(chars, buf_end - VERM_BOUNDARY, 0); |
| 287 | if (ptr) { |
| 288 | return ptr; |
| 289 | } |
| 290 | } else { |
| 291 | const u8 *ptr = rvermUnalign(chars, buf_end - VERM_BOUNDARY, 0); |
| 292 | if (ptr) { |
| 293 | return ptr; |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | buf_end -= min; |
| 298 | if (buf >= buf_end) { |
| 299 | return buf_end; |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | // Aligned loops from here on in. |
| 304 | const u8 *ptr = nocase ? rvermSearchAlignedNocase(chars, buf, buf_end, 0) |
| 305 | : rvermSearchAligned(chars, buf, buf_end, 0); |
| 306 | if (ptr) { |
| 307 | return ptr; |
| 308 | } |
| 309 | |
| 310 | // Tidy up the mess at the end, return buf - 1 if not found. |
| 311 | ptr = nocase ? rvermUnalignNocase(chars, buf, 0) |
| 312 | : rvermUnalign(chars, buf, 0); |
| 313 | return ptr ? ptr : buf - 1; |
| 314 | } |
| 315 | |
| 316 | /* like rvermicelliExec except returns the address of the last character which |
| 317 | * is not c */ |
| 318 | static really_inline |
| 319 | const u8 *rnvermicelliExec(char c, char nocase, const u8 *buf, |
| 320 | const u8 *buf_end) { |
| 321 | DEBUG_PRINTF("rev verm scan %s\\x%02hhx over %zu bytes\n" , |
| 322 | nocase ? "nocase " : "" , c, (size_t)(buf_end - buf)); |
| 323 | assert(buf < buf_end); |
| 324 | |
| 325 | // Handle small scans. |
| 326 | if (buf_end - buf < VERM_BOUNDARY) { |
| 327 | for (buf_end--; buf_end >= buf; buf_end--) { |
| 328 | char cur = (char)*buf_end; |
| 329 | if (nocase) { |
| 330 | cur &= CASE_CLEAR; |
| 331 | } |
| 332 | if (cur != c) { |
| 333 | break; |
| 334 | } |
| 335 | } |
| 336 | return buf_end; |
| 337 | } |
| 338 | |
| 339 | VERM_TYPE chars = VERM_SET_FN(c); /* nocase already uppercase */ |
| 340 | size_t min = (size_t)buf_end % VERM_BOUNDARY; |
| 341 | |
| 342 | if (min) { |
| 343 | // Input isn't aligned, so we need to run one iteration with an |
| 344 | // unaligned load, then skip buf backward to the next aligned address. |
| 345 | // There's some small overlap here, but we don't mind scanning it twice |
| 346 | // if we can do it quickly, do we? |
| 347 | if (nocase) { |
| 348 | const u8 *ptr = |
| 349 | rvermUnalignNocase(chars, buf_end - VERM_BOUNDARY, 1); |
| 350 | if (ptr) { |
| 351 | return ptr; |
| 352 | } |
| 353 | } else { |
| 354 | const u8 *ptr = rvermUnalign(chars, buf_end - VERM_BOUNDARY, 1); |
| 355 | if (ptr) { |
| 356 | return ptr; |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | buf_end -= min; |
| 361 | if (buf >= buf_end) { |
| 362 | return buf_end; |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | // Aligned loops from here on in. |
| 367 | const u8 *ptr = nocase ? rvermSearchAlignedNocase(chars, buf, buf_end, 1) |
| 368 | : rvermSearchAligned(chars, buf, buf_end, 1); |
| 369 | if (ptr) { |
| 370 | return ptr; |
| 371 | } |
| 372 | |
| 373 | // Tidy up the mess at the end, return buf - 1 if not found. |
| 374 | ptr = nocase ? rvermUnalignNocase(chars, buf, 1) |
| 375 | : rvermUnalign(chars, buf, 1); |
| 376 | return ptr ? ptr : buf - 1; |
| 377 | } |
| 378 | |
| 379 | /* returns highest offset of c2 (NOTE: not c1) */ |
| 380 | static really_inline |
| 381 | const u8 *rvermicelliDoubleExec(char c1, char c2, char nocase, const u8 *buf, |
| 382 | const u8 *buf_end) { |
| 383 | DEBUG_PRINTF("rev double verm scan %s\\x%02hhx%02hhx over %zu bytes\n" , |
| 384 | nocase ? "nocase " : "" , c1, c2, (size_t)(buf_end - buf)); |
| 385 | assert(buf < buf_end); |
| 386 | assert((buf_end - buf) >= VERM_BOUNDARY); |
| 387 | |
| 388 | size_t min = (size_t)buf_end % VERM_BOUNDARY; |
| 389 | VERM_TYPE chars1 = VERM_SET_FN(c1); /* nocase already uppercase */ |
| 390 | VERM_TYPE chars2 = VERM_SET_FN(c2); /* nocase already uppercase */ |
| 391 | |
| 392 | if (min) { |
| 393 | // input not aligned, so we need to run one iteration with an unaligned |
| 394 | // load, then skip buf forward to the next aligned address. There's |
| 395 | // some small overlap here, but we don't mind scanning it twice if we |
| 396 | // can do it quickly, do we? |
| 397 | const u8 *ptr; |
| 398 | if (nocase) { |
| 399 | ptr = rdvermPreconditionNocase(chars1, chars2, |
| 400 | buf_end - VERM_BOUNDARY); |
| 401 | } else { |
| 402 | ptr = rdvermPrecondition(chars1, chars2, buf_end - VERM_BOUNDARY); |
| 403 | } |
| 404 | |
| 405 | if (ptr) { |
| 406 | return ptr; |
| 407 | } |
| 408 | |
| 409 | buf_end -= min; |
| 410 | if (buf >= buf_end) { |
| 411 | return buf_end; |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | // Aligned loops from here on in |
| 416 | if (nocase) { |
| 417 | return rdvermSearchAlignedNocase(chars1, chars2, c1, c2, buf, buf_end); |
| 418 | } else { |
| 419 | return rdvermSearchAligned(chars1, chars2, c1, c2, buf, buf_end); |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | #endif /* VERMICELLI_H */ |
| 424 | |