git.sesse.net Git - plocate/blob - plocate-build.cpp

   1 #include "db.h"
   2 #include "dprintf.h"
   3 #include "turbopfor-encode.h"
   4
   5 #include <algorithm>
   6 #include <assert.h>
   7 #include <chrono>
   8 #include <getopt.h>
   9 #include <iosfwd>
  10 #include <math.h>
  11 #include <memory>
  12 #include <random>
  13 #include <stdint.h>
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17 #include <string>
  18 #include <string_view>
  19 #include <sys/stat.h>
  20 #include <utility>
  21 #include <vector>
  22 #include <zdict.h>
  23 #include <zstd.h>
  24
  25 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  26
  27 #define NUM_TRIGRAMS 16777216
  28
  29 using namespace std;
  30 using namespace std::chrono;
  31
  32 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
  33
  34 constexpr unsigned num_overflow_slots = 16;
  35 bool use_debug = false;
  36
  37 static inline uint32_t read_unigram(const string_view s, size_t idx)
  38 {
  39         if (idx < s.size()) {
  40                 return (unsigned char)s[idx];
  41         } else {
  42                 return 0;
  43         }
  44 }
  45
  46 static inline uint32_t read_trigram(const string_view s, size_t start)
  47 {
  48         return read_unigram(s, start) |
  49                 (read_unigram(s, start + 1) << 8) |
  50                 (read_unigram(s, start + 2) << 16);
  51 }
  52
  53 enum {
  54         DBE_NORMAL = 0, /* A non-directory file */
  55         DBE_DIRECTORY = 1, /* A directory */
  56         DBE_END = 2 /* End of directory contents; contains no name */
  57 };
  58
  59 // From mlocate.
  60 struct db_header {
  61         uint8_t magic[8];
  62         uint32_t conf_size;
  63         uint8_t version;
  64         uint8_t check_visibility;
  65         uint8_t pad[2];
  66 };
  67
  68 // From mlocate.
  69 struct db_directory {
  70         uint64_t time_sec;
  71         uint32_t time_nsec;
  72         uint8_t pad[4];
  73 };
  74
  75 class PostingListBuilder {
  76 public:
  77         inline void add_docid(uint32_t docid);
  78         void finish();
  79
  80         string encoded;
  81         size_t num_docids = 0;
  82
  83 private:
  84         void write_header(uint32_t docid);
  85         void append_block();
  86
  87         vector<uint32_t> pending_deltas;
  88
  89         uint32_t last_block_end, last_docid = -1;
  90 };
  91
  92 void PostingListBuilder::add_docid(uint32_t docid)
  93 {
  94         // Deduplicate against the last inserted value, if any.
  95         if (docid == last_docid) {
  96                 return;
  97         }
  98
  99         if (num_docids == 0) {
 100                 // Very first docid.
 101                 write_header(docid);
 102                 ++num_docids;
 103                 last_block_end = last_docid = docid;
 104                 return;
 105         }
 106
 107         pending_deltas.push_back(docid - last_docid - 1);
 108         last_docid = docid;
 109         if (pending_deltas.size() == 128) {
 110                 append_block();
 111                 pending_deltas.clear();
 112                 last_block_end = docid;
 113         }
 114         ++num_docids;
 115 }
 116
 117 void PostingListBuilder::finish()
 118 {
 119         if (pending_deltas.empty()) {
 120                 return;
 121         }
 122
 123         assert(!encoded.empty());  // write_header() should already have run.
 124
 125         // No interleaving for partial blocks.
 126         unsigned char buf[P4NENC_BOUND(128)];
 127         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
 128         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 129 }
 130
 131 void PostingListBuilder::append_block()
 132 {
 133         unsigned char buf[P4NENC_BOUND(128)];
 134         assert(pending_deltas.size() == 128);
 135         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
 136         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 137 }
 138
 139 void PostingListBuilder::write_header(uint32_t docid)
 140 {
 141         unsigned char buf[P4NENC_BOUND(1)];
 142         unsigned char *end = write_baseval(docid, buf);
 143         encoded.append(reinterpret_cast<char *>(buf), end - buf);
 144 }
 145
 146 class DatabaseReceiver {
 147 public:
 148         virtual ~DatabaseReceiver() = default;
 149         virtual void add_file(string filename) = 0;
 150         virtual void flush_block() = 0;
 151 };
 152
 153 class DictionaryBuilder : public DatabaseReceiver {
 154 public:
 155         DictionaryBuilder(size_t blocks_to_keep, size_t block_size)
 156                 : blocks_to_keep(blocks_to_keep), block_size(block_size) {}
 157         void add_file(string filename) override;
 158         void flush_block() override;
 159         string train(size_t buf_size);
 160
 161 private:
 162         const size_t blocks_to_keep, block_size;
 163         string current_block;
 164         uint64_t block_num = 0;
 165         size_t num_files_in_block = 0;
 166
 167         std::mt19937 reservoir_rand{ 1234 };  // Fixed seed for reproducibility.
 168         bool keep_current_block = true;
 169         int64_t slot_for_current_block = -1;
 170
 171         vector<string> sampled_blocks;
 172         vector<size_t> lengths;
 173 };
 174
 175 void DictionaryBuilder::add_file(string filename)
 176 {
 177         if (keep_current_block) {  // Only bother saving the filenames if we're actually keeping the block.
 178                 if (!current_block.empty()) {
 179                         current_block.push_back('\0');
 180                 }
 181                 current_block += filename;
 182         }
 183         if (++num_files_in_block == block_size) {
 184                 flush_block();
 185         }
 186 }
 187
 188 void DictionaryBuilder::flush_block()
 189 {
 190         if (keep_current_block) {
 191                 if (slot_for_current_block == -1) {
 192                         lengths.push_back(current_block.size());
 193                         sampled_blocks.push_back(move(current_block));
 194                 } else {
 195                         lengths[slot_for_current_block] = current_block.size();
 196                         sampled_blocks[slot_for_current_block] = move(current_block);
 197                 }
 198         }
 199         current_block.clear();
 200         num_files_in_block = 0;
 201         ++block_num;
 202
 203         if (block_num < blocks_to_keep) {
 204                 keep_current_block = true;
 205                 slot_for_current_block = -1;
 206         } else {
 207                 // Keep every block with equal probability (reservoir sampling).
 208                 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
 209                 keep_current_block = (idx < blocks_to_keep);
 210                 slot_for_current_block = idx;
 211         }
 212 }
 213
 214 string DictionaryBuilder::train(size_t buf_size)
 215 {
 216         string dictionary_buf;
 217         sort(sampled_blocks.begin(), sampled_blocks.end());  // Seemingly important for decompression speed.
 218         for (const string &block : sampled_blocks) {
 219                 dictionary_buf += block;
 220         }
 221
 222         string buf;
 223         buf.resize(buf_size);
 224         size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
 225         if (ret == size_t(-1)) {
 226                 return "";
 227         }
 228         dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
 229         buf.resize(ret);
 230
 231         sampled_blocks.clear();
 232         lengths.clear();
 233
 234         return buf;
 235 }
 236
 237 class Corpus : public DatabaseReceiver {
 238 public:
 239         Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict)
 240                 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), cdict(cdict)
 241         {
 242                 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 243         }
 244         ~Corpus() override
 245         {
 246                 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 247                         delete invindex[i];
 248                 }
 249         }
 250
 251         void add_file(string filename) override;
 252         void flush_block() override;
 253
 254         vector<uint64_t> filename_blocks;
 255         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 256         bool seen_trigram(uint32_t trgm)
 257         {
 258                 return invindex[trgm] != nullptr;
 259         }
 260         PostingListBuilder &get_pl_builder(uint32_t trgm)
 261         {
 262                 if (invindex[trgm] == nullptr) {
 263                         invindex[trgm] = new PostingListBuilder;
 264                 }
 265                 return *invindex[trgm];
 266         }
 267         size_t num_trigrams() const;
 268
 269 private:
 270         unique_ptr<PostingListBuilder *[]> invindex;
 271         FILE *outfp;
 272         string current_block;
 273         string tempbuf;
 274         const size_t block_size;
 275         ZSTD_CDict *cdict;
 276 };
 277
 278 void Corpus::add_file(string filename)
 279 {
 280         ++num_files;
 281         if (!current_block.empty()) {
 282                 current_block.push_back('\0');
 283         }
 284         current_block += filename;
 285         if (++num_files_in_block == block_size) {
 286                 flush_block();
 287         }
 288 }
 289
 290 void Corpus::flush_block()
 291 {
 292         if (current_block.empty()) {
 293                 return;
 294         }
 295
 296         uint32_t docid = num_blocks;
 297
 298         // Create trigrams.
 299         const char *ptr = current_block.c_str();
 300         while (ptr < current_block.c_str() + current_block.size()) {
 301                 string_view s(ptr);
 302                 if (s.size() >= 3) {
 303                         for (size_t j = 0; j < s.size() - 2; ++j) {
 304                                 uint32_t trgm = read_trigram(s, j);
 305                                 get_pl_builder(trgm).add_docid(docid);
 306                         }
 307                 }
 308                 ptr += s.size() + 1;
 309         }
 310
 311         // Compress and add the filename block.
 312         filename_blocks.push_back(ftell(outfp));
 313         string compressed = zstd_compress(current_block, cdict, &tempbuf);
 314         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 315                 perror("fwrite()");
 316                 exit(1);
 317         }
 318
 319         current_block.clear();
 320         num_files_in_block = 0;
 321         ++num_blocks;
 322 }
 323
 324 size_t Corpus::num_trigrams() const
 325 {
 326         size_t num = 0;
 327         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 328                 if (invindex[trgm] != nullptr) {
 329                         ++num;
 330                 }
 331         }
 332         return num;
 333 }
 334
 335 string read_cstr(FILE *fp)
 336 {
 337         string ret;
 338         for (;;) {
 339                 int ch = getc(fp);
 340                 if (ch == -1) {
 341                         perror("getc");
 342                         exit(1);
 343                 }
 344                 if (ch == 0) {
 345                         return ret;
 346                 }
 347                 ret.push_back(ch);
 348         }
 349 }
 350
 351 void handle_directory(FILE *fp, DatabaseReceiver *receiver)
 352 {
 353         db_directory dummy;
 354         if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
 355                 if (feof(fp)) {
 356                         return;
 357                 } else {
 358                         perror("fread");
 359                 }
 360         }
 361
 362         string dir_path = read_cstr(fp);
 363         if (dir_path == "/") {
 364                 dir_path = "";
 365         }
 366
 367         for (;;) {
 368                 int type = getc(fp);
 369                 if (type == DBE_NORMAL) {
 370                         string filename = read_cstr(fp);
 371                         receiver->add_file(dir_path + "/" + filename);
 372                 } else if (type == DBE_DIRECTORY) {
 373                         string dirname = read_cstr(fp);
 374                         receiver->add_file(dir_path + "/" + dirname);
 375                 } else {
 376                         return;  // Probably end.
 377                 }
 378         }
 379 }
 380
 381 void read_plaintext(FILE *fp, DatabaseReceiver *receiver)
 382 {
 383         if (fseek(fp, 0, SEEK_SET) != 0) {
 384                 perror("fseek");
 385                 exit(1);
 386         }
 387
 388         while (!feof(fp)) {
 389                 char buf[1024];
 390                 if (fgets(buf, sizeof(buf), fp) == nullptr) {
 391                         break;
 392                 }
 393                 string s(buf);
 394                 assert(!s.empty());
 395                 while (s.back() != '\n' && !feof(fp)) {
 396                         // The string was longer than the buffer, so read again.
 397                         if (fgets(buf, sizeof(buf), fp) == nullptr) {
 398                                 break;
 399                         }
 400                         s += buf;
 401                 }
 402                 if (!s.empty() && s.back() == '\n')
 403                         s.pop_back();
 404                 receiver->add_file(move(s));
 405         }
 406 }
 407
 408 void read_mlocate(FILE *fp, DatabaseReceiver *receiver)
 409 {
 410         if (fseek(fp, 0, SEEK_SET) != 0) {
 411                 perror("fseek");
 412                 exit(1);
 413         }
 414
 415         db_header hdr;
 416         if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
 417                 perror("short read");
 418                 exit(1);
 419         }
 420
 421         // TODO: Care about the base path.
 422         string path = read_cstr(fp);
 423         while (!feof(fp)) {
 424                 handle_directory(fp, receiver);
 425         }
 426 }
 427
 428 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
 429 {
 430         static ZSTD_CCtx *ctx = nullptr;
 431         if (ctx == nullptr) {
 432                 ctx = ZSTD_createCCtx();
 433         }
 434
 435         size_t max_size = ZSTD_compressBound(src.size());
 436         if (tempbuf->size() < max_size) {
 437                 tempbuf->resize(max_size);
 438         }
 439         size_t size;
 440         if (cdict == nullptr) {
 441                 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 442         } else {
 443                 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
 444         }
 445         return string(tempbuf->data(), size);
 446 }
 447
 448 bool is_prime(uint32_t x)
 449 {
 450         if ((x % 2) == 0 || (x % 3) == 0) {
 451                 return false;
 452         }
 453         uint32_t limit = ceil(sqrt(x));
 454         for (uint32_t factor = 5; factor <= limit; ++factor) {
 455                 if ((x % factor) == 0) {
 456                         return false;
 457                 }
 458         }
 459         return true;
 460 }
 461
 462 uint32_t next_prime(uint32_t x)
 463 {
 464         if ((x % 2) == 0) {
 465                 ++x;
 466         }
 467         while (!is_prime(x)) {
 468                 x += 2;
 469         }
 470         return x;
 471 }
 472
 473 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 474 {
 475         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 476         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 477                 ht[i].trgm = uint32_t(-1);
 478                 ht[i].num_docids = 0;
 479                 ht[i].offset = 0;
 480         }
 481         for (uint32_t trgm : all_trigrams) {
 482                 // We don't know offset yet, so set it to zero.
 483                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 484
 485                 uint32_t bucket = hash_trigram(trgm, ht_size);
 486                 unsigned distance = 0;
 487                 while (ht[bucket].num_docids != 0) {
 488                         // Robin Hood hashing; reduces the longest distance by a lot.
 489                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 490                         if (distance > other_distance) {
 491                                 swap(to_insert, ht[bucket]);
 492                                 distance = other_distance;
 493                         }
 494
 495                         ++bucket, ++distance;
 496                         if (distance > num_overflow_slots) {
 497                                 return nullptr;
 498                         }
 499                 }
 500                 ht[bucket] = to_insert;
 501         }
 502         return ht;
 503 }
 504
 505 class DatabaseBuilder {
 506 public:
 507         DatabaseBuilder(const char *outfile, int block_size, string dictionary);
 508         Corpus *start_corpus();
 509         void finish_corpus();
 510
 511 private:
 512         FILE *outfp;
 513         Header hdr;
 514         const int block_size;
 515         steady_clock::time_point corpus_start;
 516         Corpus *corpus = nullptr;
 517         ZSTD_CDict *cdict = nullptr;
 518 };
 519
 520 DatabaseBuilder::DatabaseBuilder(const char *outfile, int block_size, string dictionary)
 521         : block_size(block_size)
 522 {
 523         umask(0027);
 524         outfp = fopen(outfile, "wb");
 525         if (outfp == nullptr) {
 526                 perror(outfile);
 527                 exit(1);
 528         }
 529
 530         // Write the header.
 531         memcpy(hdr.magic, "\0plocate", 8);
 532         hdr.version = -1;  // Mark as broken.
 533         hdr.hashtable_size = 0;  // Not known yet.
 534         hdr.extra_ht_slots = num_overflow_slots;
 535         hdr.num_docids = 0;
 536         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 537         hdr.max_version = 1;
 538         hdr.filename_index_offset_bytes = -1;
 539         hdr.zstd_dictionary_length_bytes = -1;
 540         fwrite(&hdr, sizeof(hdr), 1, outfp);
 541
 542         if (dictionary.empty()) {
 543                 hdr.zstd_dictionary_offset_bytes = 0;
 544                 hdr.zstd_dictionary_length_bytes = 0;
 545         } else {
 546                 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
 547                 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
 548                 hdr.zstd_dictionary_length_bytes = dictionary.size();
 549                 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
 550         }
 551 }
 552
 553 Corpus *DatabaseBuilder::start_corpus()
 554 {
 555         corpus_start = steady_clock::now();
 556         corpus = new Corpus(outfp, block_size, cdict);
 557         return corpus;
 558 }
 559
 560 void DatabaseBuilder::finish_corpus()
 561 {
 562         corpus->flush_block();
 563         hdr.num_docids = corpus->filename_blocks.size();
 564
 565         // Stick an empty block at the end as sentinel.
 566         corpus->filename_blocks.push_back(ftell(outfp));
 567         const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
 568
 569         // Write the offsets to the filenames.
 570         hdr.filename_index_offset_bytes = ftell(outfp);
 571         const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
 572         fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
 573         corpus->filename_blocks.clear();
 574         corpus->filename_blocks.shrink_to_fit();
 575
 576         // Finish up encoding the posting lists.
 577         size_t trigrams = 0, longest_posting_list = 0;
 578         size_t bytes_for_posting_lists = 0;
 579         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 580                 if (!corpus->seen_trigram(trgm))
 581                         continue;
 582                 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
 583                 pl_builder.finish();
 584                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 585                 trigrams += pl_builder.num_docids;
 586                 bytes_for_posting_lists += pl_builder.encoded.size();
 587         }
 588         size_t num_trigrams = corpus->num_trigrams();
 589         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 590                 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
 591         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 592
 593         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
 594
 595         // Find the used trigrams.
 596         vector<uint32_t> all_trigrams;
 597         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 598                 if (corpus->seen_trigram(trgm)) {
 599                         all_trigrams.push_back(trgm);
 600                 }
 601         }
 602
 603         // Create the hash table.
 604         unique_ptr<Trigram[]> hashtable;
 605         uint32_t ht_size = next_prime(all_trigrams.size());
 606         for (;;) {
 607                 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
 608                 if (hashtable == nullptr) {
 609                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 610                         ht_size = next_prime(ht_size * 1.05);
 611                 } else {
 612                         dprintf("Created hash table of size %u.\n\n", ht_size);
 613                         break;
 614                 }
 615         }
 616
 617         // Find the offsets for each posting list.
 618         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 619         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 620         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 621                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 622                 if (hashtable[i].num_docids == 0) {
 623                         continue;
 624                 }
 625
 626                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 627                 offset += encoded.size();
 628         }
 629
 630         // Write the hash table.
 631         hdr.hash_table_offset_bytes = ftell(outfp);
 632         hdr.hashtable_size = ht_size;
 633         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 634
 635         // Write the actual posting lists.
 636         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 637                 if (hashtable[i].num_docids == 0) {
 638                         continue;
 639                 }
 640                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 641                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 642         }
 643
 644         // Rewind, and write the updated header.
 645         hdr.version = 1;
 646         fseek(outfp, 0, SEEK_SET);
 647         fwrite(&hdr, sizeof(hdr), 1, outfp);
 648         fclose(outfp);
 649
 650         size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 651
 652         dprintf("Block size:     %7d files\n", block_size);
 653         dprintf("Dictionary:     %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
 654         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 655         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 656         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 657         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 658         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 659         dprintf("\n");
 660 }
 661
 662 void do_build(const char *infile, const char *outfile, int block_size, bool plaintext)
 663 {
 664         FILE *infp = fopen(infile, "rb");
 665         if (infp == nullptr) {
 666                 perror(infile);
 667                 exit(1);
 668         }
 669
 670         // Train the dictionary by sampling real blocks.
 671         // The documentation for ZDICT_trainFromBuffer() claims that a reasonable
 672         // dictionary size is ~100 kB, but 1 kB seems to actually compress better for us,
 673         // and decompress just as fast.
 674         DictionaryBuilder builder(/*blocks_to_keep=*/1000, block_size);
 675         if (plaintext) {
 676                 read_plaintext(infp, &builder);
 677         } else {
 678                 read_mlocate(infp, &builder);
 679         }
 680         string dictionary = builder.train(1024);
 681
 682         DatabaseBuilder db(outfile, block_size, dictionary);
 683         Corpus *corpus = db.start_corpus();
 684         if (plaintext) {
 685                 read_plaintext(infp, corpus);
 686         } else {
 687                 read_mlocate(infp, corpus);
 688         }
 689         fclose(infp);
 690
 691         dprintf("Read %zu files from %s\n", corpus->num_files, infile);
 692         db.finish_corpus();
 693 }
 694
 695 void usage()
 696 {
 697         printf(
 698                 "Usage: plocate-build MLOCATE_DB PLOCATE_DB\n"
 699                 "\n"
 700                 "Generate plocate index from mlocate.db, typically /var/lib/mlocate/mlocate.db.\n"
 701                 "Normally, the destination should be /var/lib/mlocate/plocate.db.\n"
 702                 "\n"
 703                 "  -b, --block-size SIZE  number of filenames to store in each block (default 32)\n"
 704                 "  -p, --plaintext        input is a plaintext file, not an mlocate database\n"
 705                 "      --help             print this help\n"
 706                 "      --version          print version information\n");
 707 }
 708
 709 void version()
 710 {
 711         printf("plocate-build %s\n", PLOCATE_VERSION);
 712         printf("Copyright 2020 Steinar H. Gunderson\n");
 713         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 714         printf("This is free software: you are free to change and redistribute it.\n");
 715         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 716 }
 717
 718 int main(int argc, char **argv)
 719 {
 720         static const struct option long_options[] = {
 721                 { "block-size", required_argument, 0, 'b' },
 722                 { "plaintext", no_argument, 0, 'p' },
 723                 { "help", no_argument, 0, 'h' },
 724                 { "version", no_argument, 0, 'V' },
 725                 { "debug", no_argument, 0, 'D' },  // Not documented.
 726                 { 0, 0, 0, 0 }
 727         };
 728
 729         int block_size = 32;
 730         bool plaintext = false;
 731
 732         setlocale(LC_ALL, "");
 733         for (;;) {
 734                 int option_index = 0;
 735                 int c = getopt_long(argc, argv, "b:hpVD", long_options, &option_index);
 736                 if (c == -1) {
 737                         break;
 738                 }
 739                 switch (c) {
 740                 case 'b':
 741                         block_size = atoi(optarg);
 742                         break;
 743                 case 'p':
 744                         plaintext = true;
 745                         break;
 746                 case 'h':
 747                         usage();
 748                         exit(0);
 749                 case 'V':
 750                         version();
 751                         exit(0);
 752                 case 'D':
 753                         use_debug = true;
 754                         break;
 755                 default:
 756                         exit(1);
 757                 }
 758         }
 759
 760         if (argc - optind != 2) {
 761                 usage();
 762                 exit(1);
 763         }
 764
 765         do_build(argv[optind], argv[optind + 1], block_size, plaintext);
 766         exit(EXIT_SUCCESS);
 767 }