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         dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
 226         buf.resize(ret);
 227
 228         sampled_blocks.clear();
 229         lengths.clear();
 230
 231         return buf;
 232 }
 233
 234 class Corpus : public DatabaseReceiver {
 235 public:
 236         Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict)
 237                 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), cdict(cdict)
 238         {
 239                 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 240         }
 241         ~Corpus() override
 242         {
 243                 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 244                         delete invindex[i];
 245                 }
 246         }
 247
 248         void add_file(string filename) override;
 249         void flush_block() override;
 250
 251         vector<uint64_t> filename_blocks;
 252         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 253         bool seen_trigram(uint32_t trgm)
 254         {
 255                 return invindex[trgm] != nullptr;
 256         }
 257         PostingListBuilder &get_pl_builder(uint32_t trgm)
 258         {
 259                 if (invindex[trgm] == nullptr) {
 260                         invindex[trgm] = new PostingListBuilder;
 261                 }
 262                 return *invindex[trgm];
 263         }
 264         size_t num_trigrams() const;
 265
 266 private:
 267         unique_ptr<PostingListBuilder *[]> invindex;
 268         FILE *outfp;
 269         string current_block;
 270         string tempbuf;
 271         const size_t block_size;
 272         ZSTD_CDict *cdict;
 273 };
 274
 275 void Corpus::add_file(string filename)
 276 {
 277         ++num_files;
 278         if (!current_block.empty()) {
 279                 current_block.push_back('\0');
 280         }
 281         current_block += filename;
 282         if (++num_files_in_block == block_size) {
 283                 flush_block();
 284         }
 285 }
 286
 287 void Corpus::flush_block()
 288 {
 289         if (current_block.empty()) {
 290                 return;
 291         }
 292
 293         uint32_t docid = num_blocks;
 294
 295         // Create trigrams.
 296         const char *ptr = current_block.c_str();
 297         while (ptr < current_block.c_str() + current_block.size()) {
 298                 string_view s(ptr);
 299                 if (s.size() >= 3) {
 300                         for (size_t j = 0; j < s.size() - 2; ++j) {
 301                                 uint32_t trgm = read_trigram(s, j);
 302                                 get_pl_builder(trgm).add_docid(docid);
 303                         }
 304                 }
 305                 ptr += s.size() + 1;
 306         }
 307
 308         // Compress and add the filename block.
 309         filename_blocks.push_back(ftell(outfp));
 310         string compressed = zstd_compress(current_block, cdict, &tempbuf);
 311         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 312                 perror("fwrite()");
 313                 exit(1);
 314         }
 315
 316         current_block.clear();
 317         num_files_in_block = 0;
 318         ++num_blocks;
 319 }
 320
 321 size_t Corpus::num_trigrams() const
 322 {
 323         size_t num = 0;
 324         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 325                 if (invindex[trgm] != nullptr) {
 326                         ++num;
 327                 }
 328         }
 329         return num;
 330 }
 331
 332 string read_cstr(FILE *fp)
 333 {
 334         string ret;
 335         for (;;) {
 336                 int ch = getc(fp);
 337                 if (ch == -1) {
 338                         perror("getc");
 339                         exit(1);
 340                 }
 341                 if (ch == 0) {
 342                         return ret;
 343                 }
 344                 ret.push_back(ch);
 345         }
 346 }
 347
 348 void handle_directory(FILE *fp, DatabaseReceiver *receiver)
 349 {
 350         db_directory dummy;
 351         if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
 352                 if (feof(fp)) {
 353                         return;
 354                 } else {
 355                         perror("fread");
 356                 }
 357         }
 358
 359         string dir_path = read_cstr(fp);
 360         if (dir_path == "/") {
 361                 dir_path = "";
 362         }
 363
 364         for (;;) {
 365                 int type = getc(fp);
 366                 if (type == DBE_NORMAL) {
 367                         string filename = read_cstr(fp);
 368                         receiver->add_file(dir_path + "/" + filename);
 369                 } else if (type == DBE_DIRECTORY) {
 370                         string dirname = read_cstr(fp);
 371                         receiver->add_file(dir_path + "/" + dirname);
 372                 } else {
 373                         return;  // Probably end.
 374                 }
 375         }
 376 }
 377
 378 void read_mlocate(const char *filename, DatabaseReceiver *receiver)
 379 {
 380         FILE *fp = fopen(filename, "rb");
 381         if (fp == nullptr) {
 382                 perror(filename);
 383                 exit(1);
 384         }
 385
 386         db_header hdr;
 387         if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
 388                 perror("short read");
 389                 exit(1);
 390         }
 391
 392         // TODO: Care about the base path.
 393         string path = read_cstr(fp);
 394         while (!feof(fp)) {
 395                 handle_directory(fp, receiver);
 396         }
 397         fclose(fp);
 398 }
 399
 400 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
 401 {
 402         static ZSTD_CCtx *ctx = nullptr;
 403         if (ctx == nullptr) {
 404                 ctx = ZSTD_createCCtx();
 405         }
 406
 407         size_t max_size = ZSTD_compressBound(src.size());
 408         if (tempbuf->size() < max_size) {
 409                 tempbuf->resize(max_size);
 410         }
 411         size_t size;
 412         if (cdict == nullptr) {
 413                 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 414         } else {
 415                 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
 416         }
 417         return string(tempbuf->data(), size);
 418 }
 419
 420 bool is_prime(uint32_t x)
 421 {
 422         if ((x % 2) == 0 || (x % 3) == 0) {
 423                 return false;
 424         }
 425         uint32_t limit = ceil(sqrt(x));
 426         for (uint32_t factor = 5; factor <= limit; ++factor) {
 427                 if ((x % factor) == 0) {
 428                         return false;
 429                 }
 430         }
 431         return true;
 432 }
 433
 434 uint32_t next_prime(uint32_t x)
 435 {
 436         if ((x % 2) == 0) {
 437                 ++x;
 438         }
 439         while (!is_prime(x)) {
 440                 x += 2;
 441         }
 442         return x;
 443 }
 444
 445 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 446 {
 447         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 448         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 449                 ht[i].trgm = uint32_t(-1);
 450                 ht[i].num_docids = 0;
 451                 ht[i].offset = 0;
 452         }
 453         for (uint32_t trgm : all_trigrams) {
 454                 // We don't know offset yet, so set it to zero.
 455                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 456
 457                 uint32_t bucket = hash_trigram(trgm, ht_size);
 458                 unsigned distance = 0;
 459                 while (ht[bucket].num_docids != 0) {
 460                         // Robin Hood hashing; reduces the longest distance by a lot.
 461                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 462                         if (distance > other_distance) {
 463                                 swap(to_insert, ht[bucket]);
 464                                 distance = other_distance;
 465                         }
 466
 467                         ++bucket, ++distance;
 468                         if (distance > num_overflow_slots) {
 469                                 return nullptr;
 470                         }
 471                 }
 472                 ht[bucket] = to_insert;
 473         }
 474         return ht;
 475 }
 476
 477 void do_build(const char *infile, const char *outfile, int block_size)
 478 {
 479         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 480
 481         umask(0027);
 482         FILE *outfp = fopen(outfile, "wb");
 483         if (outfp == nullptr) {
 484                 perror(outfile);
 485                 exit(1);
 486         }
 487
 488         // Write the header.
 489         Header hdr;
 490         memcpy(hdr.magic, "\0plocate", 8);
 491         hdr.version = -1;  // Mark as broken.
 492         hdr.hashtable_size = 0;  // Not known yet.
 493         hdr.extra_ht_slots = num_overflow_slots;
 494         hdr.num_docids = 0;
 495         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 496         hdr.max_version = 1;
 497         hdr.filename_index_offset_bytes = -1;
 498         hdr.zstd_dictionary_length_bytes = -1;
 499         fwrite(&hdr, sizeof(hdr), 1, outfp);
 500
 501         // Train the dictionary by sampling real blocks.
 502         // The documentation for ZDICT_trainFromBuffer() claims that a reasonable
 503         // dictionary size is ~100 kB, but 1 kB seems to actually compress better for us,
 504         // and decompress just as fast.
 505         DictionaryBuilder builder(/*blocks_to_keep=*/1000, block_size);
 506         read_mlocate(infile, &builder);
 507         string dictionary = builder.train(1024);
 508         ZSTD_CDict *cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
 509
 510         hdr.zstd_dictionary_offset_bytes = ftell(outfp);
 511         fwrite(dictionary.data(), dictionary.size(), 1, outfp);
 512         hdr.zstd_dictionary_length_bytes = dictionary.size();
 513
 514         Corpus corpus(outfp, block_size, cdict);
 515         read_mlocate(infile, &corpus);
 516         if (false) {  // To read a plain text file.
 517                 FILE *fp = fopen(infile, "r");
 518                 while (!feof(fp)) {
 519                         char buf[1024];
 520                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 521                                 break;
 522                         }
 523                         string s(buf);
 524                         if (s.back() == '\n')
 525                                 s.pop_back();
 526                         corpus.add_file(move(s));
 527                 }
 528                 fclose(fp);
 529         }
 530         corpus.flush_block();
 531         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 532         hdr.num_docids = corpus.filename_blocks.size();
 533
 534         // Stick an empty block at the end as sentinel.
 535         corpus.filename_blocks.push_back(ftell(outfp));
 536         const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
 537
 538         // Write the offsets to the filenames.
 539         hdr.filename_index_offset_bytes = ftell(outfp);
 540         const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
 541         fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
 542         corpus.filename_blocks.clear();
 543         corpus.filename_blocks.shrink_to_fit();
 544
 545         // Finish up encoding the posting lists.
 546         size_t trigrams = 0, longest_posting_list = 0;
 547         size_t bytes_for_posting_lists = 0;
 548         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 549                 if (!corpus.seen_trigram(trgm))
 550                         continue;
 551                 PostingListBuilder &pl_builder = corpus.get_pl_builder(trgm);
 552                 pl_builder.finish();
 553                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 554                 trigrams += pl_builder.num_docids;
 555                 bytes_for_posting_lists += pl_builder.encoded.size();
 556         }
 557         size_t num_trigrams = corpus.num_trigrams();
 558         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 559                 corpus.num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
 560         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 561
 562         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 563
 564         // Find the used trigrams.
 565         vector<uint32_t> all_trigrams;
 566         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 567                 if (corpus.seen_trigram(trgm)) {
 568                         all_trigrams.push_back(trgm);
 569                 }
 570         }
 571
 572         // Create the hash table.
 573         unique_ptr<Trigram[]> hashtable;
 574         uint32_t ht_size = next_prime(all_trigrams.size());
 575         for (;;) {
 576                 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
 577                 if (hashtable == nullptr) {
 578                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 579                         ht_size = next_prime(ht_size * 1.05);
 580                 } else {
 581                         dprintf("Created hash table of size %u.\n\n", ht_size);
 582                         break;
 583                 }
 584         }
 585
 586         // Find the offsets for each posting list.
 587         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 588         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 589         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 590                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 591                 if (hashtable[i].num_docids == 0) {
 592                         continue;
 593                 }
 594
 595                 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 596                 offset += encoded.size();
 597         }
 598
 599         // Write the hash table.
 600         hdr.hash_table_offset_bytes = ftell(outfp);
 601         hdr.hashtable_size = ht_size;
 602         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 603
 604         // Write the actual posting lists.
 605         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 606                 if (hashtable[i].num_docids == 0) {
 607                         continue;
 608                 }
 609                 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 610                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 611         }
 612
 613         // Rewind, and write the updated header.
 614         hdr.version = 1;
 615         fseek(outfp, 0, SEEK_SET);
 616         fwrite(&hdr, sizeof(hdr), 1, outfp);
 617         fclose(outfp);
 618
 619         size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 620
 621         dprintf("Block size:     %7d files\n", block_size);
 622         dprintf("Dictionary:     %'7.1f MB\n", dictionary.size() / 1048576.0);
 623         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 624         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 625         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 626         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 627         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 628         dprintf("\n");
 629 }
 630
 631 void usage()
 632 {
 633         printf(
 634                 "Usage: plocate-build MLOCATE_DB PLOCATE_DB\n"
 635                 "\n"
 636                 "Generate plocate index from mlocate.db, typically /var/lib/mlocate/mlocate.db.\n"
 637                 "Normally, the destination should be /var/lib/mlocate/plocate.db.\n"
 638                 "\n"
 639                 "  -b, --block-size SIZE  number of filenames to store in each block (default 32)\n"
 640                 "      --help             print this help\n"
 641                 "      --version          print version information\n");
 642 }
 643
 644 void version()
 645 {
 646         printf("plocate-build %s\n", PLOCATE_VERSION);
 647         printf("Copyright 2020 Steinar H. Gunderson\n");
 648         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 649         printf("This is free software: you are free to change and redistribute it.\n");
 650         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 651 }
 652
 653 int main(int argc, char **argv)
 654 {
 655         static const struct option long_options[] = {
 656                 { "block-size", required_argument, 0, 'b' },
 657                 { "help", no_argument, 0, 'h' },
 658                 { "version", no_argument, 0, 'V' },
 659                 { "debug", no_argument, 0, 'D' },  // Not documented.
 660                 { 0, 0, 0, 0 }
 661         };
 662
 663         int block_size = 32;
 664
 665         setlocale(LC_ALL, "");
 666         for (;;) {
 667                 int option_index = 0;
 668                 int c = getopt_long(argc, argv, "b:hVD", long_options, &option_index);
 669                 if (c == -1) {
 670                         break;
 671                 }
 672                 switch (c) {
 673                 case 'b':
 674                         block_size = atoi(optarg);
 675                         break;
 676                 case 'h':
 677                         usage();
 678                         exit(0);
 679                 case 'V':
 680                         version();
 681                         exit(0);
 682                 case 'D':
 683                         use_debug = true;
 684                         break;
 685                 default:
 686                         exit(1);
 687                 }
 688         }
 689
 690         if (argc - optind != 2) {
 691                 usage();
 692                 exit(1);
 693         }
 694
 695         do_build(argv[optind], argv[optind + 1], block_size);
 696         exit(EXIT_SUCCESS);
 697 }