4 #include <unordered_map>
11 #include <arpa/inet.h>
13 #include <sys/types.h>
18 #define P4NENC_BOUND(n) ((n+127)/128+(n+32)*sizeof(uint32_t))
20 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
23 using namespace std::chrono;
25 static inline uint32_t read_unigram(const string &s, size_t idx)
28 return (unsigned char)s[idx];
34 static inline uint32_t read_trigram(const string &s, size_t start)
36 return read_unigram(s, start) |
37 (read_unigram(s, start + 1) << 8) |
38 (read_unigram(s, start + 2) << 16);
43 DBE_NORMAL = 0, /* A non-directory file */
44 DBE_DIRECTORY = 1, /* A directory */
45 DBE_END = 2 /* End of directory contents; contains no name */
54 uint8_t check_visibility;
66 const char *handle_directory(const char *ptr, vector<string> *files)
68 ptr += sizeof(db_directory);
70 string dir_path = ptr;
71 ptr += dir_path.size() + 1;
72 if (dir_path == "/") {
77 uint8_t type = *ptr++;
78 if (type == DBE_NORMAL) {
79 string filename = ptr;
80 files->push_back(dir_path + "/" + filename);
81 ptr += filename.size() + 1;
82 } else if (type == DBE_DIRECTORY) {
84 files->push_back(dir_path + "/" + dirname);
85 ptr += dirname.size() + 1;
92 void read_mlocate(const char *filename, vector<string> *files)
94 int fd = open(filename, O_RDONLY);
99 off_t len = lseek(fd, 0, SEEK_END);
104 const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
105 if (data == MAP_FAILED) {
110 const db_header *hdr = (const db_header *)data;
112 // TODO: Care about the base path.
113 string path = data + sizeof(db_header);
114 uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
116 const char *ptr = data + offset;
117 while (ptr < data + len) {
118 ptr = handle_directory(ptr, files);
121 munmap((void *)data, len);
125 void do_build(const char *infile, const char *outfile)
127 //steady_clock::time_point start = steady_clock::now();
129 vector<string> files;
130 read_mlocate(infile, &files);
131 if (false) { // To read a plain text file.
132 FILE *fp = fopen(infile, "r");
135 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
139 if (s.back() == '\n') s.pop_back();
140 files.push_back(move(s));
144 dprintf("Read %zu files from %s\n", files.size(), infile);
146 unordered_map<uint32_t, string> pl;
147 size_t trigrams = 0, longest_posting_list = 0;
148 unordered_map<uint32_t, vector<uint32_t>> invindex;
149 for (size_t i = 0; i < files.size(); ++i) {
150 const string &s = files[i];
152 for (size_t j = 0; j < s.size() - 2; ++j) {
153 uint32_t trgm = read_trigram(s, j);
154 invindex[trgm].push_back(i);
159 size_t bytes_used = 0;
160 for (auto &[trigram, docids] : invindex) {
161 auto last = unique(docids.begin(), docids.end());
162 docids.erase(last, docids.end());
163 longest_posting_list = max(longest_posting_list, docids.size());
164 trigrams += docids.size();
166 size_t bytes_needed = P4NENC_BOUND(docids.size());
167 if (buf.size() < bytes_needed) buf.resize(bytes_needed);
168 size_t bytes = p4nd1enc128v32(&docids[0], docids.size(), reinterpret_cast<unsigned char *>(&buf[0]));
169 pl[trigram] = string(buf.data(), bytes);
172 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
173 files.size(), invindex.size(), trigrams, double(trigrams) / invindex.size(), longest_posting_list);
175 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
176 //steady_clock::time_point end = steady_clock::now();
177 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
179 vector<uint32_t> all_trigrams;
180 for (auto &[trigram, docids] : invindex) {
181 all_trigrams.push_back(trigram);
183 sort(all_trigrams.begin(), all_trigrams.end());
186 FILE *outfp = fopen(outfile, "wb");
189 uint64_t num_trigrams = invindex.size(); // 64 to get alignment.
190 fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
192 // Find out where the posting lists will end.
193 uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
194 uint64_t filename_index_offset = offset;
195 for (uint32_t trgm : all_trigrams) {
196 filename_index_offset += pl[trgm].size();
198 fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
200 size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
202 for (uint32_t trgm : all_trigrams) {
203 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
204 fwrite(&trgm, sizeof(trgm), 1, outfp); // One byte wasted, but OK.
205 uint32_t num_docids = invindex[trgm].size();
206 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
207 fwrite(&offset, sizeof(offset), 1, outfp);
208 offset += pl[trgm].size();
209 bytes_for_trigrams += 16;
212 // Write the actual posting lists.
213 for (uint32_t trgm : all_trigrams) {
214 fwrite(pl[trgm].data(), pl[trgm].size(), 1, outfp);
215 bytes_for_posting_lists += pl[trgm].size();
218 // Write the offsets to the filenames.
219 offset = filename_index_offset + files.size() * sizeof(offset);
220 for (const string &filename : files) {
221 fwrite(&offset, sizeof(offset), 1, outfp);
222 offset += filename.size() + 1;
223 bytes_for_filename_index += sizeof(offset);
224 bytes_for_filenames += filename.size() + 1;
227 // Write the actual filenames.
228 for (const string &filename : files) {
229 fwrite(filename.c_str(), filename.size() + 1, 1, outfp);
234 size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
236 dprintf("Trigrams: %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
237 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
238 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
239 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
240 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
244 int main(int argc, char **argv)
246 do_build(argv[1], argv[2]);