16 #include <sys/types.h>
18 #include <unordered_map>
22 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
24 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
27 using namespace std::chrono;
29 string zstd_compress(const string &src, string *tempbuf);
31 constexpr unsigned num_overflow_slots = 16;
33 static inline uint32_t read_unigram(const string_view s, size_t idx)
36 return (unsigned char)s[idx];
42 static inline uint32_t read_trigram(const string_view s, size_t start)
44 return read_unigram(s, start) |
45 (read_unigram(s, start + 1) << 8) |
46 (read_unigram(s, start + 2) << 16);
50 DBE_NORMAL = 0, /* A non-directory file */
51 DBE_DIRECTORY = 1, /* A directory */
52 DBE_END = 2 /* End of directory contents; contains no name */
60 uint8_t check_visibility;
71 class PostingListBuilder {
73 void add_docid(uint32_t docid);
77 size_t num_docids = 0;
80 void write_header(uint32_t docid);
83 vector<uint32_t> pending_docids;
85 uint32_t last_block_end, last_docid = -1;
88 void PostingListBuilder::add_docid(uint32_t docid)
90 // Deduplicate against the last inserted value, if any.
91 if (docid == last_docid) {
95 if (num_docids == 0) {
99 last_block_end = last_docid = docid;
104 pending_docids.push_back(docid);
105 if (pending_docids.size() == 128) {
107 pending_docids.clear();
108 last_block_end = docid;
113 void PostingListBuilder::finish()
115 if (pending_docids.empty()) {
119 assert(!encoded.empty()); // write_header() should already have run.
121 // No interleaving for partial blocks.
122 unsigned char buf[P4NENC_BOUND(128)];
123 unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
124 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
127 void PostingListBuilder::append_block()
129 unsigned char buf[P4NENC_BOUND(128)];
130 assert(pending_docids.size() == 128);
131 unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
132 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
135 void PostingListBuilder::write_header(uint32_t docid)
137 unsigned char buf[P4NENC_BOUND(1)];
138 size_t bytes = p4nd1enc128v32(&docid, 1, buf);
139 encoded.append(reinterpret_cast<char *>(buf), bytes);
144 Corpus(FILE *outfp, size_t block_size)
145 : outfp(outfp), block_size(block_size) {}
146 void add_file(string filename);
149 vector<uint64_t> filename_blocks;
150 unordered_map<uint32_t, PostingListBuilder> invindex;
151 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
155 string current_block;
157 const size_t block_size;
160 void Corpus::add_file(string filename)
163 if (!current_block.empty()) {
164 current_block.push_back('\0');
166 current_block += filename;
167 if (++num_files_in_block == block_size) {
172 void Corpus::flush_block()
174 if (current_block.empty()) {
178 uint32_t docid = num_blocks;
181 const char *ptr = current_block.c_str();
182 while (ptr < current_block.c_str() + current_block.size()) {
185 for (size_t j = 0; j < s.size() - 2; ++j) {
186 uint32_t trgm = read_trigram(s, j);
187 invindex[trgm].add_docid(docid);
193 // Compress and add the filename block.
194 filename_blocks.push_back(ftell(outfp));
195 string compressed = zstd_compress(current_block, &tempbuf);
196 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
201 current_block.clear();
202 num_files_in_block = 0;
206 string read_cstr(FILE *fp)
222 void handle_directory(FILE *fp, Corpus *corpus)
225 if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
233 string dir_path = read_cstr(fp);
234 if (dir_path == "/") {
240 if (type == DBE_NORMAL) {
241 string filename = read_cstr(fp);
242 corpus->add_file(dir_path + "/" + filename);
243 } else if (type == DBE_DIRECTORY) {
244 string dirname = read_cstr(fp);
245 corpus->add_file(dir_path + "/" + dirname);
247 return; // Probably end.
252 void read_mlocate(const char *filename, Corpus *corpus)
254 FILE *fp = fopen(filename, "rb");
261 if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
262 perror("short read");
266 // TODO: Care about the base path.
267 string path = read_cstr(fp);
269 handle_directory(fp, corpus);
274 string zstd_compress(const string &src, string *tempbuf)
276 size_t max_size = ZSTD_compressBound(src.size());
277 if (tempbuf->size() < max_size) {
278 tempbuf->resize(max_size);
280 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
281 return string(tempbuf->data(), size);
284 bool is_prime(uint32_t x)
286 if ((x % 2) == 0 || (x % 3) == 0) {
289 uint32_t limit = ceil(sqrt(x));
290 for (uint32_t factor = 5; factor <= limit; ++factor) {
291 if ((x % factor) == 0) {
298 uint32_t next_prime(uint32_t x)
303 while (!is_prime(x)) {
309 unique_ptr<Trigram[]> create_hashtable(const Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
311 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
312 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
313 ht[i].trgm = uint32_t(-1);
314 ht[i].num_docids = 0;
317 for (uint32_t trgm : all_trigrams) {
318 // We don't know offset yet, so set it to zero.
319 Trigram to_insert{ trgm, uint32_t(corpus.invindex.find(trgm)->second.num_docids), 0 };
321 uint32_t bucket = hash_trigram(trgm, ht_size);
322 unsigned distance = 0;
323 while (ht[bucket].num_docids != 0) {
324 // Robin Hood hashing; reduces the longest distance by a lot.
325 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
326 if (distance > other_distance) {
327 swap(to_insert, ht[bucket]);
328 distance = other_distance;
331 ++bucket, ++distance;
332 if (distance > num_overflow_slots) {
336 ht[bucket] = to_insert;
341 void do_build(const char *infile, const char *outfile, int block_size)
343 steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
346 FILE *outfp = fopen(outfile, "wb");
350 memcpy(hdr.magic, "\0plocate", 8);
351 hdr.version = -1; // Mark as broken.
352 hdr.hashtable_size = 0; // Not known yet.
353 hdr.extra_ht_slots = num_overflow_slots;
354 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
355 hdr.filename_index_offset_bytes = -1;
356 fwrite(&hdr, sizeof(hdr), 1, outfp);
358 Corpus corpus(outfp, block_size);
360 read_mlocate(infile, &corpus);
361 if (false) { // To read a plain text file.
362 FILE *fp = fopen(infile, "r");
365 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
369 if (s.back() == '\n')
371 corpus.add_file(move(s));
375 corpus.flush_block();
376 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
377 hdr.num_docids = corpus.filename_blocks.size();
379 // Stick an empty block at the end as sentinel.
380 corpus.filename_blocks.push_back(ftell(outfp));
381 const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
383 // Write the offsets to the filenames.
384 hdr.filename_index_offset_bytes = ftell(outfp);
385 const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
386 fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
387 corpus.filename_blocks.clear();
388 corpus.filename_blocks.shrink_to_fit();
390 // Finish up encoding the posting lists.
391 size_t trigrams = 0, longest_posting_list = 0;
392 size_t bytes_for_posting_lists = 0;
393 for (auto &[trigram, pl_builder] : corpus.invindex) {
395 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
396 trigrams += pl_builder.num_docids;
397 bytes_for_posting_lists += pl_builder.encoded.size();
399 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
400 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
401 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
403 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
405 // Sort the trigrams, mostly to get a consistent result every time
406 // (the hash table will put things in random order anyway).
407 vector<uint32_t> all_trigrams;
408 for (auto &[trigram, pl_builder] : corpus.invindex) {
409 all_trigrams.push_back(trigram);
411 sort(all_trigrams.begin(), all_trigrams.end());
413 // Create the hash table.
414 unique_ptr<Trigram[]> hashtable;
415 uint32_t ht_size = next_prime(all_trigrams.size());
417 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
418 if (hashtable == nullptr) {
419 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
420 ht_size = next_prime(ht_size * 1.05);
422 dprintf("Created hash table of size %u.\n\n", ht_size);
427 // Find the offsets for each posting list.
428 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
429 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
430 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
431 hashtable[i].offset = offset; // Needs to be there even for empty slots.
432 if (hashtable[i].num_docids == 0) {
436 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
437 offset += encoded.size();
440 // Write the hash table.
441 hdr.hash_table_offset_bytes = ftell(outfp);
442 hdr.hashtable_size = ht_size;
443 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
445 // Write the actual posting lists.
446 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
447 if (hashtable[i].num_docids == 0) {
450 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
451 fwrite(encoded.data(), encoded.size(), 1, outfp);
454 // Rewind, and write the updated header.
456 fseek(outfp, 0, SEEK_SET);
457 fwrite(&hdr, sizeof(hdr), 1, outfp);
460 size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
462 dprintf("Block size: %7d files\n", block_size);
463 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
464 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
465 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
466 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
467 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
471 int main(int argc, char **argv)
473 do_build(argv[1], argv[2], 32);