4 #include <unordered_map>
11 #include <arpa/inet.h>
13 #include <sys/types.h>
19 #define P4NENC_BOUND(n) ((n+127)/128+(n+32)*sizeof(uint32_t))
21 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
24 using namespace std::chrono;
26 string zstd_compress(const string &src, string *tempbuf);
28 static inline uint32_t read_unigram(const string_view s, size_t idx)
31 return (unsigned char)s[idx];
37 static inline uint32_t read_trigram(const string_view s, size_t start)
39 return read_unigram(s, start) |
40 (read_unigram(s, start + 1) << 8) |
41 (read_unigram(s, start + 2) << 16);
46 DBE_NORMAL = 0, /* A non-directory file */
47 DBE_DIRECTORY = 1, /* A directory */
48 DBE_END = 2 /* End of directory contents; contains no name */
57 uint8_t check_visibility;
71 Corpus(size_t block_size) : block_size(block_size) {}
72 void add_file(string filename);
75 vector<string> filename_blocks;
76 unordered_map<uint32_t, vector<uint32_t>> invindex;
77 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
82 const size_t block_size;
85 void Corpus::add_file(string filename)
88 if (!current_block.empty()) {
89 current_block.push_back('\0');
91 current_block += filename;
92 if (++num_files_in_block == block_size) {
97 void Corpus::flush_block()
99 if (current_block.empty()) {
103 uint32_t docid = num_blocks;
106 const char *ptr = current_block.c_str();
107 while (ptr < current_block.c_str() + current_block.size()) {
110 for (size_t j = 0; j < s.size() - 2; ++j) {
111 uint32_t trgm = read_trigram(s, j);
112 vector<uint32_t> &docids = invindex[trgm];
113 if (docids.empty() || docids.back() != docid) {
114 docids.push_back(docid);
121 // Compress and add the filename block.
122 filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
124 current_block.clear();
125 num_files_in_block = 0;
129 const char *handle_directory(const char *ptr, Corpus *corpus)
131 ptr += sizeof(db_directory);
133 string dir_path = ptr;
134 ptr += dir_path.size() + 1;
135 if (dir_path == "/") {
140 uint8_t type = *ptr++;
141 if (type == DBE_NORMAL) {
142 string filename = ptr;
143 corpus->add_file(dir_path + "/" + filename);
144 ptr += filename.size() + 1;
145 } else if (type == DBE_DIRECTORY) {
146 string dirname = ptr;
147 corpus->add_file(dir_path + "/" + dirname);
148 ptr += dirname.size() + 1;
155 void read_mlocate(const char *filename, Corpus *corpus)
157 int fd = open(filename, O_RDONLY);
162 off_t len = lseek(fd, 0, SEEK_END);
167 const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
168 if (data == MAP_FAILED) {
173 const db_header *hdr = (const db_header *)data;
175 // TODO: Care about the base path.
176 string path = data + sizeof(db_header);
177 uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
179 const char *ptr = data + offset;
180 while (ptr < data + len) {
181 ptr = handle_directory(ptr, corpus);
184 munmap((void *)data, len);
188 string zstd_compress(const string &src, string *tempbuf)
190 size_t max_size = ZSTD_compressBound(src.size());
191 if (tempbuf->size() < max_size) {
192 tempbuf->resize(max_size);
194 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
195 return string(tempbuf->data(), size);
198 void do_build(const char *infile, const char *outfile, int block_size)
200 //steady_clock::time_point start = steady_clock::now();
202 Corpus corpus(block_size);
204 read_mlocate(infile, &corpus);
205 if (false) { // To read a plain text file.
206 FILE *fp = fopen(infile, "r");
209 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
213 if (s.back() == '\n') s.pop_back();
214 corpus.add_file(move(s));
218 corpus.flush_block();
219 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
221 // Compress posting lists.
223 size_t trigrams = 0, longest_posting_list = 0;
224 size_t bytes_used = 0;
225 unordered_map<uint32_t, string> pl;
226 for (auto &[trigram, docids] : corpus.invindex) {
227 longest_posting_list = max(longest_posting_list, docids.size());
228 trigrams += docids.size();
230 size_t bytes_needed = P4NENC_BOUND(docids.size());
231 if (buf.size() < bytes_needed) buf.resize(bytes_needed);
232 size_t bytes = p4nd1enc128v32(&docids[0], docids.size(), reinterpret_cast<unsigned char *>(&buf[0]));
233 pl[trigram] = string(buf.data(), bytes);
236 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
237 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
239 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
240 //steady_clock::time_point end = steady_clock::now();
241 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
243 vector<uint32_t> all_trigrams;
244 for (auto &[trigram, docids] : corpus.invindex) {
245 all_trigrams.push_back(trigram);
247 sort(all_trigrams.begin(), all_trigrams.end());
250 FILE *outfp = fopen(outfile, "wb");
253 uint64_t num_trigrams = corpus.invindex.size(); // 64 to get alignment.
254 fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
256 // Find out where the posting lists will end.
257 uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
258 uint64_t filename_index_offset = offset;
259 for (uint32_t trgm : all_trigrams) {
260 filename_index_offset += pl[trgm].size();
262 fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
264 size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
266 for (uint32_t trgm : all_trigrams) {
267 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
268 fwrite(&trgm, sizeof(trgm), 1, outfp); // One byte wasted, but OK.
269 uint32_t num_docids = corpus.invindex[trgm].size();
270 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
271 fwrite(&offset, sizeof(offset), 1, outfp);
272 offset += pl[trgm].size();
273 bytes_for_trigrams += 16;
276 // Write the actual posting lists.
277 for (uint32_t trgm : all_trigrams) {
278 fwrite(pl[trgm].data(), pl[trgm].size(), 1, outfp);
279 bytes_for_posting_lists += pl[trgm].size();
282 // Stick an empty block at the end as sentinel.
283 corpus.filename_blocks.push_back("");
285 // Write the offsets to the filenames.
286 offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
287 for (const string &filename : corpus.filename_blocks) {
288 fwrite(&offset, sizeof(offset), 1, outfp);
289 offset += filename.size();
290 bytes_for_filename_index += sizeof(offset);
291 bytes_for_filenames += filename.size();
294 // Write the actual filenames.
295 for (const string &filename : corpus.filename_blocks) {
296 fwrite(filename.data(), filename.size(), 1, outfp);
301 //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
303 dprintf("Block size: %7d files\n", block_size);
304 dprintf("Trigrams: %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
305 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
306 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
307 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
308 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
312 int main(int argc, char **argv)
314 do_build(argv[1], argv[2], 32);