git.sesse.net Git - plocate/blob - database-builder.cpp

   1 #include "database-builder.h"
   2
   3 #include "dprintf.h"
   4 #include "turbopfor-encode.h"
   5
   6 #include <algorithm>
   7 #include <assert.h>
   8 #include <fcntl.h>
   9 #include <string.h>
  10 #include <string_view>
  11 #include <sys/stat.h>
  12 #include <sys/time.h>
  13 #include <sys/types.h>
  14 #include <unistd.h>
  15 #include <zdict.h>
  16 #include <zstd.h>
  17
  18 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  19
  20 #define NUM_TRIGRAMS 16777216
  21
  22 using namespace std;
  23 using namespace std::chrono;
  24
  25 constexpr unsigned num_overflow_slots = 16;
  26
  27 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
  28
  29 static inline uint32_t read_unigram(const string_view s, size_t idx)
  30 {
  31         if (idx < s.size()) {
  32                 return (unsigned char)s[idx];
  33         } else {
  34                 return 0;
  35         }
  36 }
  37
  38 static inline uint32_t read_trigram(const string_view s, size_t start)
  39 {
  40         return read_unigram(s, start) |
  41                 (read_unigram(s, start + 1) << 8) |
  42                 (read_unigram(s, start + 2) << 16);
  43 }
  44
  45 class PostingListBuilder {
  46 public:
  47         inline void add_docid(uint32_t docid);
  48         void finish();
  49
  50         string encoded;
  51         size_t num_docids = 0;
  52
  53 private:
  54         void write_header(uint32_t docid);
  55         void append_block();
  56
  57         vector<uint32_t> pending_deltas;
  58
  59         uint32_t last_block_end, last_docid = -1;
  60 };
  61
  62 void PostingListBuilder::add_docid(uint32_t docid)
  63 {
  64         // Deduplicate against the last inserted value, if any.
  65         if (docid == last_docid) {
  66                 return;
  67         }
  68
  69         if (num_docids == 0) {
  70                 // Very first docid.
  71                 write_header(docid);
  72                 ++num_docids;
  73                 last_block_end = last_docid = docid;
  74                 return;
  75         }
  76
  77         pending_deltas.push_back(docid - last_docid - 1);
  78         last_docid = docid;
  79         if (pending_deltas.size() == 128) {
  80                 append_block();
  81                 pending_deltas.clear();
  82                 last_block_end = docid;
  83         }
  84         ++num_docids;
  85 }
  86
  87 void PostingListBuilder::finish()
  88 {
  89         if (pending_deltas.empty()) {
  90                 return;
  91         }
  92
  93         assert(!encoded.empty());  // write_header() should already have run.
  94
  95         // No interleaving for partial blocks.
  96         unsigned char buf[P4NENC_BOUND(128)];
  97         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
  98         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
  99 }
 100
 101 void PostingListBuilder::append_block()
 102 {
 103         unsigned char buf[P4NENC_BOUND(128)];
 104         assert(pending_deltas.size() == 128);
 105         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
 106         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 107 }
 108
 109 void PostingListBuilder::write_header(uint32_t docid)
 110 {
 111         unsigned char buf[P4NENC_BOUND(1)];
 112         unsigned char *end = write_baseval(docid, buf);
 113         encoded.append(reinterpret_cast<char *>(buf), end - buf);
 114 }
 115
 116 void DictionaryBuilder::add_file(string filename, dir_time)
 117 {
 118         if (keep_current_block) {  // Only bother saving the filenames if we're actually keeping the block.
 119                 if (!current_block.empty()) {
 120                         current_block.push_back('\0');
 121                 }
 122                 current_block += filename;
 123         }
 124         if (++num_files_in_block == block_size) {
 125                 flush_block();
 126         }
 127 }
 128
 129 void DictionaryBuilder::flush_block()
 130 {
 131         if (keep_current_block) {
 132                 if (slot_for_current_block == -1) {
 133                         lengths.push_back(current_block.size());
 134                         sampled_blocks.push_back(move(current_block));
 135                 } else {
 136                         lengths[slot_for_current_block] = current_block.size();
 137                         sampled_blocks[slot_for_current_block] = move(current_block);
 138                 }
 139         }
 140         current_block.clear();
 141         num_files_in_block = 0;
 142         ++block_num;
 143
 144         if (block_num < blocks_to_keep) {
 145                 keep_current_block = true;
 146                 slot_for_current_block = -1;
 147         } else {
 148                 // Keep every block with equal probability (reservoir sampling).
 149                 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
 150                 keep_current_block = (idx < blocks_to_keep);
 151                 slot_for_current_block = idx;
 152         }
 153 }
 154
 155 string DictionaryBuilder::train(size_t buf_size)
 156 {
 157         string dictionary_buf;
 158         sort(sampled_blocks.begin(), sampled_blocks.end());  // Seemingly important for decompression speed.
 159         for (const string &block : sampled_blocks) {
 160                 dictionary_buf += block;
 161         }
 162
 163         string buf;
 164         buf.resize(buf_size);
 165         size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
 166         if (ret == size_t(-1)) {
 167                 return "";
 168         }
 169         dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
 170         buf.resize(ret);
 171
 172         sampled_blocks.clear();
 173         lengths.clear();
 174
 175         return buf;
 176 }
 177
 178 Corpus::Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
 179         : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
 180 {
 181         fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 182         if (store_dir_times) {
 183                 dir_time_ctx = ZSTD_createCStream();
 184                 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
 185         }
 186 }
 187
 188 Corpus::~Corpus()
 189 {
 190         for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 191                 delete invindex[i];
 192         }
 193 }
 194
 195 PostingListBuilder &Corpus::get_pl_builder(uint32_t trgm)
 196 {
 197         if (invindex[trgm] == nullptr) {
 198                 invindex[trgm] = new PostingListBuilder;
 199         }
 200         return *invindex[trgm];
 201 }
 202
 203 void Corpus::add_file(string filename, dir_time dt)
 204 {
 205         ++num_files;
 206         if (!current_block.empty()) {
 207                 current_block.push_back('\0');
 208         }
 209         current_block += filename;
 210         if (++num_files_in_block == block_size) {
 211                 flush_block();
 212         }
 213
 214         if (store_dir_times) {
 215                 if (dt.sec == -1) {
 216                         // Not a directory.
 217                         dir_times.push_back('\0');
 218                 } else {
 219                         dir_times.push_back('\1');
 220                         dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
 221                         dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
 222                 }
 223                 compress_dir_times(/*allowed_slop=*/4096);
 224         }
 225 }
 226
 227 void Corpus::compress_dir_times(size_t allowed_slop)
 228 {
 229         while (dir_times.size() >= allowed_slop) {
 230                 size_t old_size = dir_times_compressed.size();
 231                 dir_times_compressed.resize(old_size + 4096);
 232
 233                 ZSTD_outBuffer outbuf;
 234                 outbuf.dst = dir_times_compressed.data() + old_size;
 235                 outbuf.size = 4096;
 236                 outbuf.pos = 0;
 237
 238                 ZSTD_inBuffer inbuf;
 239                 inbuf.src = dir_times.data();
 240                 inbuf.size = dir_times.size();
 241                 inbuf.pos = 0;
 242
 243                 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
 244                 if (ret < 0) {
 245                         fprintf(stderr, "ZSTD_compressStream() failed\n");
 246                         exit(1);
 247                 }
 248
 249                 dir_times_compressed.resize(old_size + outbuf.pos);
 250                 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
 251
 252                 if (outbuf.pos == 0 && inbuf.pos == 0) {
 253                         // Nothing happened (not enough data?), try again later.
 254                         return;
 255                 }
 256         }
 257 }
 258
 259 void Corpus::flush_block()
 260 {
 261         if (current_block.empty()) {
 262                 return;
 263         }
 264
 265         uint32_t docid = num_blocks;
 266
 267         // Create trigrams.
 268         const char *ptr = current_block.c_str();
 269         while (ptr < current_block.c_str() + current_block.size()) {
 270                 string_view s(ptr);
 271                 if (s.size() >= 3) {
 272                         for (size_t j = 0; j < s.size() - 2; ++j) {
 273                                 uint32_t trgm = read_trigram(s, j);
 274                                 get_pl_builder(trgm).add_docid(docid);
 275                         }
 276                 }
 277                 ptr += s.size() + 1;
 278         }
 279
 280         // Compress and add the filename block.
 281         filename_blocks.push_back(ftell(outfp));
 282         string compressed = zstd_compress(current_block, cdict, &tempbuf);
 283         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 284                 perror("fwrite()");
 285                 exit(1);
 286         }
 287
 288         current_block.clear();
 289         num_files_in_block = 0;
 290         ++num_blocks;
 291 }
 292
 293 void Corpus::finish()
 294 {
 295         flush_block();
 296 }
 297
 298 size_t Corpus::num_trigrams() const
 299 {
 300         size_t num = 0;
 301         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 302                 if (invindex[trgm] != nullptr) {
 303                         ++num;
 304                 }
 305         }
 306         return num;
 307 }
 308
 309 string Corpus::get_compressed_dir_times()
 310 {
 311         if (!store_dir_times) {
 312                 return "";
 313         }
 314         compress_dir_times(/*allowed_slop=*/0);
 315         assert(dir_times.empty());
 316
 317         for (;;) {
 318                 size_t old_size = dir_times_compressed.size();
 319                 dir_times_compressed.resize(old_size + 4096);
 320
 321                 ZSTD_outBuffer outbuf;
 322                 outbuf.dst = dir_times_compressed.data() + old_size;
 323                 outbuf.size = 4096;
 324                 outbuf.pos = 0;
 325
 326                 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
 327                 if (ret < 0) {
 328                         fprintf(stderr, "ZSTD_compressStream() failed\n");
 329                         exit(1);
 330                 }
 331
 332                 dir_times_compressed.resize(old_size + outbuf.pos);
 333
 334                 if (ret == 0) {
 335                         // All done.
 336                         break;
 337                 }
 338         }
 339
 340         return dir_times_compressed;
 341 }
 342
 343 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
 344 {
 345         static ZSTD_CCtx *ctx = nullptr;
 346         if (ctx == nullptr) {
 347                 ctx = ZSTD_createCCtx();
 348         }
 349
 350         size_t max_size = ZSTD_compressBound(src.size());
 351         if (tempbuf->size() < max_size) {
 352                 tempbuf->resize(max_size);
 353         }
 354         size_t size;
 355         if (cdict == nullptr) {
 356                 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 357         } else {
 358                 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
 359         }
 360         return string(tempbuf->data(), size);
 361 }
 362
 363 bool is_prime(uint32_t x)
 364 {
 365         if ((x % 2) == 0 || (x % 3) == 0) {
 366                 return false;
 367         }
 368         uint32_t limit = ceil(sqrt(x));
 369         for (uint32_t factor = 5; factor <= limit; ++factor) {
 370                 if ((x % factor) == 0) {
 371                         return false;
 372                 }
 373         }
 374         return true;
 375 }
 376
 377 uint32_t next_prime(uint32_t x)
 378 {
 379         if ((x % 2) == 0) {
 380                 ++x;
 381         }
 382         while (!is_prime(x)) {
 383                 x += 2;
 384         }
 385         return x;
 386 }
 387
 388 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 389 {
 390         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 391         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 392                 ht[i].trgm = uint32_t(-1);
 393                 ht[i].num_docids = 0;
 394                 ht[i].offset = 0;
 395         }
 396         for (uint32_t trgm : all_trigrams) {
 397                 // We don't know offset yet, so set it to zero.
 398                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 399
 400                 uint32_t bucket = hash_trigram(trgm, ht_size);
 401                 unsigned distance = 0;
 402                 while (ht[bucket].num_docids != 0) {
 403                         // Robin Hood hashing; reduces the longest distance by a lot.
 404                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 405                         if (distance > other_distance) {
 406                                 swap(to_insert, ht[bucket]);
 407                                 distance = other_distance;
 408                         }
 409
 410                         ++bucket, ++distance;
 411                         if (distance > num_overflow_slots) {
 412                                 return nullptr;
 413                         }
 414                 }
 415                 ht[bucket] = to_insert;
 416         }
 417         return ht;
 418 }
 419
 420 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
 421         : outfile(outfile), block_size(block_size)
 422 {
 423         umask(0027);
 424
 425         string path = outfile;
 426         path.resize(path.find_last_of('/') + 1);
 427         if (path.empty()) {
 428                 path = ".";
 429         }
 430         int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
 431         if (fd == -1) {
 432                 perror(path.c_str());
 433                 exit(1);
 434         }
 435
 436         if (owner != (gid_t)-1) {
 437                 if (fchown(fd, (uid_t)-1, owner) == -1) {
 438                         perror("fchown");
 439                         exit(1);
 440                 }
 441         }
 442
 443         outfp = fdopen(fd, "wb");
 444         if (outfp == nullptr) {
 445                 perror(outfile);
 446                 exit(1);
 447         }
 448
 449         // Write the header.
 450         memcpy(hdr.magic, "\0plocate", 8);
 451         hdr.version = -1;  // Mark as broken.
 452         hdr.hashtable_size = 0;  // Not known yet.
 453         hdr.extra_ht_slots = num_overflow_slots;
 454         hdr.num_docids = 0;
 455         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 456         hdr.max_version = 2;
 457         hdr.filename_index_offset_bytes = -1;
 458         hdr.zstd_dictionary_length_bytes = -1;
 459         hdr.check_visibility = check_visibility;
 460         fwrite(&hdr, sizeof(hdr), 1, outfp);
 461
 462         if (dictionary.empty()) {
 463                 hdr.zstd_dictionary_offset_bytes = 0;
 464                 hdr.zstd_dictionary_length_bytes = 0;
 465         } else {
 466                 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
 467                 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
 468                 hdr.zstd_dictionary_length_bytes = dictionary.size();
 469                 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
 470         }
 471
 472         hdr.directory_data_length_bytes = 0;
 473         hdr.directory_data_offset_bytes = 0;
 474         hdr.next_zstd_dictionary_length_bytes = 0;
 475         hdr.next_zstd_dictionary_offset_bytes = 0;
 476         hdr.conf_block_length_bytes = 0;
 477         hdr.conf_block_offset_bytes = 0;
 478 }
 479
 480 Corpus *DatabaseBuilder::start_corpus(bool store_dir_times)
 481 {
 482         corpus_start = steady_clock::now();
 483         corpus = new Corpus(outfp, block_size, cdict, store_dir_times);
 484         return corpus;
 485 }
 486
 487 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
 488 {
 489         this->next_dictionary = move(next_dictionary);
 490 }
 491
 492 void DatabaseBuilder::set_conf_block(std::string conf_block)
 493 {
 494         this->conf_block = move(conf_block);
 495 }
 496
 497 void DatabaseBuilder::finish_corpus()
 498 {
 499         corpus->finish();
 500         hdr.num_docids = corpus->filename_blocks.size();
 501
 502         // Stick an empty block at the end as sentinel.
 503         corpus->filename_blocks.push_back(ftell(outfp));
 504         const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
 505
 506         // Write the offsets to the filenames.
 507         hdr.filename_index_offset_bytes = ftell(outfp);
 508         const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
 509         fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
 510         corpus->filename_blocks.clear();
 511         corpus->filename_blocks.shrink_to_fit();
 512
 513         // Finish up encoding the posting lists.
 514         size_t trigrams = 0, longest_posting_list = 0;
 515         size_t bytes_for_posting_lists = 0;
 516         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 517                 if (!corpus->seen_trigram(trgm))
 518                         continue;
 519                 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
 520                 pl_builder.finish();
 521                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 522                 trigrams += pl_builder.num_docids;
 523                 bytes_for_posting_lists += pl_builder.encoded.size();
 524         }
 525         size_t num_trigrams = corpus->num_trigrams();
 526         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 527                 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
 528         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 529
 530         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
 531
 532         // Find the used trigrams.
 533         vector<uint32_t> all_trigrams;
 534         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 535                 if (corpus->seen_trigram(trgm)) {
 536                         all_trigrams.push_back(trgm);
 537                 }
 538         }
 539
 540         // Create the hash table.
 541         unique_ptr<Trigram[]> hashtable;
 542         uint32_t ht_size = next_prime(all_trigrams.size());
 543         for (;;) {
 544                 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
 545                 if (hashtable == nullptr) {
 546                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 547                         ht_size = next_prime(ht_size * 1.05);
 548                 } else {
 549                         dprintf("Created hash table of size %u.\n\n", ht_size);
 550                         break;
 551                 }
 552         }
 553
 554         // Find the offsets for each posting list.
 555         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 556         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 557         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 558                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 559                 if (hashtable[i].num_docids == 0) {
 560                         continue;
 561                 }
 562
 563                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 564                 offset += encoded.size();
 565         }
 566
 567         // Write the hash table.
 568         hdr.hash_table_offset_bytes = ftell(outfp);
 569         hdr.hashtable_size = ht_size;
 570         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 571
 572         // Write the actual posting lists.
 573         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 574                 if (hashtable[i].num_docids == 0) {
 575                         continue;
 576                 }
 577                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 578                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 579         }
 580
 581         // Finally, write the directory times (for updatedb).
 582         string compressed_dir_times = corpus->get_compressed_dir_times();
 583         size_t bytes_for_compressed_dir_times = 0;
 584         if (!compressed_dir_times.empty()) {
 585                 hdr.directory_data_offset_bytes = ftell(outfp);
 586                 hdr.directory_data_length_bytes = compressed_dir_times.size();
 587                 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
 588                 bytes_for_compressed_dir_times = compressed_dir_times.size();
 589                 compressed_dir_times.clear();
 590         }
 591
 592         // Write the recommended dictionary for next update.
 593         if (!next_dictionary.empty()) {
 594                 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
 595                 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
 596                 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
 597         }
 598
 599         // And the configuration block.
 600         if (!conf_block.empty()) {
 601                 hdr.conf_block_offset_bytes = ftell(outfp);
 602                 hdr.conf_block_length_bytes = conf_block.size();
 603                 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
 604         }
 605
 606         // Rewind, and write the updated header.
 607         hdr.version = 1;
 608         fseek(outfp, 0, SEEK_SET);
 609         fwrite(&hdr, sizeof(hdr), 1, outfp);
 610
 611         // Give the file a proper name, making it visible in the file system.
 612         // TODO: It would be nice to be able to do this atomically, like with rename.
 613         unlink(outfile.c_str());
 614         char procpath[256];
 615         snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
 616         if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
 617                 perror("linkat");
 618                 exit(1);
 619         }
 620
 621         fclose(outfp);
 622
 623         size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
 624
 625         dprintf("Block size:     %7d files\n", block_size);
 626         dprintf("Dictionary:     %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
 627         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 628         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 629         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 630         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 631         if (bytes_for_compressed_dir_times != 0) {
 632                 dprintf("Modify times:   %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
 633         }
 634         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 635         dprintf("\n");
 636 }