17 #include <sys/types.h>
19 #include <unordered_map>
23 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
25 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
28 using namespace std::chrono;
30 string zstd_compress(const string &src, string *tempbuf);
32 constexpr unsigned num_overflow_slots = 16;
34 static inline uint32_t read_unigram(const string_view s, size_t idx)
37 return (unsigned char)s[idx];
43 static inline uint32_t read_trigram(const string_view s, size_t start)
45 return read_unigram(s, start) |
46 (read_unigram(s, start + 1) << 8) |
47 (read_unigram(s, start + 2) << 16);
51 DBE_NORMAL = 0, /* A non-directory file */
52 DBE_DIRECTORY = 1, /* A directory */
53 DBE_END = 2 /* End of directory contents; contains no name */
61 uint8_t check_visibility;
72 class PostingListBuilder {
74 void add_docid(uint32_t docid);
78 size_t num_docids = 0;
81 void write_header(uint32_t docid);
84 vector<uint32_t> pending_docids;
86 uint32_t last_block_end;
89 void PostingListBuilder::add_docid(uint32_t docid)
91 // Deduplicate against the last inserted value, if any.
92 if (pending_docids.empty()) {
93 if (encoded.empty()) {
97 last_block_end = docid;
99 } else if (docid == last_block_end) {
103 if (docid == pending_docids.back()) {
108 pending_docids.push_back(docid);
109 if (pending_docids.size() == 128) {
111 pending_docids.clear();
112 last_block_end = docid;
117 void PostingListBuilder::finish()
119 if (pending_docids.empty()) {
123 assert(!encoded.empty()); // write_header() should already have run.
125 // No interleaving for partial blocks.
126 unsigned char buf[P4NENC_BOUND(128)];
127 unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
128 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
131 void PostingListBuilder::append_block()
133 unsigned char buf[P4NENC_BOUND(128)];
134 assert(pending_docids.size() == 128);
135 unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
136 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
139 void PostingListBuilder::write_header(uint32_t docid)
141 unsigned char buf[P4NENC_BOUND(1)];
142 size_t bytes = p4nd1enc128v32(&docid, 1, buf);
143 encoded.append(reinterpret_cast<char *>(buf), bytes);
148 Corpus(FILE *outfp, size_t block_size)
149 : outfp(outfp), block_size(block_size) {}
150 void add_file(string filename);
153 vector<uint64_t> filename_blocks;
154 unordered_map<uint32_t, PostingListBuilder> invindex;
155 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
159 string current_block;
161 const size_t block_size;
164 void Corpus::add_file(string filename)
167 if (!current_block.empty()) {
168 current_block.push_back('\0');
170 current_block += filename;
171 if (++num_files_in_block == block_size) {
176 void Corpus::flush_block()
178 if (current_block.empty()) {
182 uint32_t docid = num_blocks;
185 const char *ptr = current_block.c_str();
186 while (ptr < current_block.c_str() + current_block.size()) {
189 for (size_t j = 0; j < s.size() - 2; ++j) {
190 uint32_t trgm = read_trigram(s, j);
191 invindex[trgm].add_docid(docid);
197 // Compress and add the filename block.
198 filename_blocks.push_back(ftell(outfp));
199 string compressed = zstd_compress(current_block, &tempbuf);
200 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
205 current_block.clear();
206 num_files_in_block = 0;
210 const char *handle_directory(const char *ptr, Corpus *corpus)
212 ptr += sizeof(db_directory);
214 string dir_path = ptr;
215 ptr += dir_path.size() + 1;
216 if (dir_path == "/") {
221 uint8_t type = *ptr++;
222 if (type == DBE_NORMAL) {
223 string filename = ptr;
224 corpus->add_file(dir_path + "/" + filename);
225 ptr += filename.size() + 1;
226 } else if (type == DBE_DIRECTORY) {
227 string dirname = ptr;
228 corpus->add_file(dir_path + "/" + dirname);
229 ptr += dirname.size() + 1;
236 void read_mlocate(const char *filename, Corpus *corpus)
238 int fd = open(filename, O_RDONLY);
243 off_t len = lseek(fd, 0, SEEK_END);
248 const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
249 if (data == MAP_FAILED) {
254 const db_header *hdr = (const db_header *)data;
256 // TODO: Care about the base path.
257 string path = data + sizeof(db_header);
258 uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
260 const char *ptr = data + offset;
261 while (ptr < data + len) {
262 ptr = handle_directory(ptr, corpus);
265 munmap((void *)data, len);
269 string zstd_compress(const string &src, string *tempbuf)
271 size_t max_size = ZSTD_compressBound(src.size());
272 if (tempbuf->size() < max_size) {
273 tempbuf->resize(max_size);
275 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
276 return string(tempbuf->data(), size);
279 bool is_prime(uint32_t x)
281 if ((x % 2) == 0 || (x % 3) == 0) {
284 uint32_t limit = ceil(sqrt(x));
285 for (uint32_t factor = 5; factor <= limit; ++factor) {
286 if ((x % factor) == 0) {
293 uint32_t next_prime(uint32_t x)
298 while (!is_prime(x)) {
304 unique_ptr<Trigram[]> create_hashtable(const Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
306 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
307 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
308 ht[i].trgm = uint32_t(-1);
309 ht[i].num_docids = 0;
312 for (uint32_t trgm : all_trigrams) {
313 // We don't know offset yet, so set it to zero.
314 Trigram to_insert{ trgm, uint32_t(corpus.invindex.find(trgm)->second.num_docids), 0 };
316 uint32_t bucket = hash_trigram(trgm, ht_size);
317 unsigned distance = 0;
318 while (ht[bucket].num_docids != 0) {
319 // Robin Hood hashing; reduces the longest distance by a lot.
320 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
321 if (distance > other_distance) {
322 swap(to_insert, ht[bucket]);
323 distance = other_distance;
326 ++bucket, ++distance;
327 if (distance > num_overflow_slots) {
331 ht[bucket] = to_insert;
336 void do_build(const char *infile, const char *outfile, int block_size)
338 steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
341 FILE *outfp = fopen(outfile, "wb");
345 memcpy(hdr.magic, "\0plocate", 8);
346 hdr.version = -1; // Mark as broken.
347 hdr.hashtable_size = 0; // Not known yet.
348 hdr.extra_ht_slots = num_overflow_slots;
349 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
350 hdr.filename_index_offset_bytes = -1;
351 fwrite(&hdr, sizeof(hdr), 1, outfp);
353 Corpus corpus(outfp, block_size);
355 read_mlocate(infile, &corpus);
356 if (false) { // To read a plain text file.
357 FILE *fp = fopen(infile, "r");
360 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
364 if (s.back() == '\n')
366 corpus.add_file(move(s));
370 corpus.flush_block();
371 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
373 // Stick an empty block at the end as sentinel.
374 corpus.filename_blocks.push_back(ftell(outfp));
375 const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
377 // Write the offsets to the filenames.
378 hdr.filename_index_offset_bytes = ftell(outfp);
379 const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
380 fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
381 corpus.filename_blocks.clear();
382 corpus.filename_blocks.shrink_to_fit();
384 // Finish up encoding the posting lists.
385 size_t trigrams = 0, longest_posting_list = 0;
386 size_t bytes_for_posting_lists = 0;
387 for (auto &[trigram, pl_builder] : corpus.invindex) {
389 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
390 trigrams += pl_builder.num_docids;
391 bytes_for_posting_lists += pl_builder.encoded.size();
393 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
394 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
395 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
397 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
399 // Sort the trigrams, mostly to get a consistent result every time
400 // (the hash table will put things in random order anyway).
401 vector<uint32_t> all_trigrams;
402 for (auto &[trigram, pl_builder] : corpus.invindex) {
403 all_trigrams.push_back(trigram);
405 sort(all_trigrams.begin(), all_trigrams.end());
407 // Create the hash table.
408 unique_ptr<Trigram[]> hashtable;
409 uint32_t ht_size = next_prime(all_trigrams.size());
411 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
412 if (hashtable == nullptr) {
413 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
414 ht_size = next_prime(ht_size * 1.05);
416 dprintf("Created hash table of size %u.\n\n", ht_size);
421 // Find the offsets for each posting list.
422 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
423 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
424 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
425 hashtable[i].offset = offset; // Needs to be there even for empty slots.
426 if (hashtable[i].num_docids == 0) {
430 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
431 offset += encoded.size();
434 // Write the hash table.
435 hdr.hash_table_offset_bytes = ftell(outfp);
436 hdr.hashtable_size = ht_size;
437 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
439 // Write the actual posting lists.
440 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
441 if (hashtable[i].num_docids == 0) {
444 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
445 fwrite(encoded.data(), encoded.size(), 1, outfp);
448 // Rewind, and write the updated header.
450 fseek(outfp, 0, SEEK_SET);
451 fwrite(&hdr, sizeof(hdr), 1, outfp);
454 size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
456 dprintf("Block size: %7d files\n", block_size);
457 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
458 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
459 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
460 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
461 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
465 int main(int argc, char **argv)
467 do_build(argv[1], argv[2], 32);