LLVMFuzzerTestOneInput: 28| 5.56k|extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) { 29| 5.56k| FuzzedDataProvider fdp(data, size); 30| 5.56k| constexpr int kMaxStringSize = 1024; 31| 5.56k| std::string source = fdp.ConsumeRandomLengthString(kMaxStringSize); 32| 22.2k| for (auto& e : simdutf::get_available_implementations()) { ------------------ | Branch (32:16): [True: 22.2k, False: 5.56k] ------------------ 33| 22.2k| if (!e->supported_by_runtime_system()) { ------------------ | Branch (33:9): [True: 5.56k, False: 16.6k] ------------------ 34| 5.56k| continue; 35| 5.56k| } 36| | /** 37| | * Transcoding from UTF-8 to UTF-16LE. 38| | */ 39| 16.6k| bool validutf8 = e->validate_utf8(source.c_str(), source.size()); 40| 16.6k| auto rutf8 = e->validate_utf8_with_errors(source.c_str(), source.size()); 41| 16.6k| if (validutf8 != (rutf8.error == simdutf::SUCCESS)) { // they should agree ------------------ | Branch (41:9): [True: 0, False: 16.6k] ------------------ 42| 0| print_input(source, e); 43| 0| abort(); 44| 0| } 45| 16.6k| if (validutf8) { ------------------ | Branch (45:9): [True: 15.4k, False: 1.23k] ------------------ 46| | // We need a buffer of size where to write the UTF-16LE words. 47| 15.4k| size_t expected_utf16words = 48| 15.4k| e->utf16_length_from_utf8(source.c_str(), source.size()); 49| 15.4k| std::unique_ptr utf16_output{ 50| 15.4k| new char16_t[expected_utf16words]}; 51| | // convert to UTF-16LE 52| 15.4k| size_t utf16words = e->convert_utf8_to_utf16le( 53| 15.4k| source.c_str(), source.size(), utf16_output.get()); 54| | // It wrote utf16words * sizeof(char16_t) bytes. 55| 15.4k| bool validutf16 = e->validate_utf16le(utf16_output.get(), utf16words); 56| 15.4k| if (!validutf16) { ------------------ | Branch (56:11): [True: 0, False: 15.4k] ------------------ 57| 0| print_input(source, e); 58| 0| abort(); 59| 0| } 60| | // convert it back: 61| | // We need a buffer of size where to write the UTF-8 words. 62| 15.4k| size_t expected_utf8words = 63| 15.4k| e->utf8_length_from_utf16le(utf16_output.get(), utf16words); 64| 15.4k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 65| | // convert to UTF-8 66| 15.4k| size_t utf8words = e->convert_utf16le_to_utf8( 67| 15.4k| utf16_output.get(), utf16words, utf8_output.get()); 68| 15.4k| std::string final_string(utf8_output.get(), utf8words); 69| 15.4k| if (final_string != source) { ------------------ | Branch (69:11): [True: 0, False: 15.4k] ------------------ 70| 0| print_input(source, e); 71| 0| abort(); 72| 0| } 73| 15.4k| } else { 74| | // invalid input!!! 75| | // We need a buffer of size where to write the UTF-16LE words. 76| 1.23k| size_t expected_utf16words = 77| 1.23k| e->utf16_length_from_utf8(source.c_str(), source.size()); 78| 1.23k| std::unique_ptr utf16_output{ 79| 1.23k| new char16_t[expected_utf16words]}; 80| | // convert to UTF-16LE 81| 1.23k| size_t utf16words = e->convert_utf8_to_utf16le( 82| 1.23k| source.c_str(), source.size(), utf16_output.get()); 83| 1.23k| if (utf16words != 0) { ------------------ | Branch (83:11): [True: 0, False: 1.23k] ------------------ 84| 0| print_input(source, e); 85| 0| abort(); 86| 0| } 87| 1.23k| } 88| | 89| | /** 90| | * Transcoding from UTF-8 to UTF-16BE. 91| | */ 92| 16.6k| if (validutf8) { ------------------ | Branch (92:9): [True: 15.4k, False: 1.23k] ------------------ 93| | // We need a buffer of size where to write the UTF-16BE words. 94| 15.4k| size_t expected_utf16words = 95| 15.4k| e->utf16_length_from_utf8(source.c_str(), source.size()); 96| 15.4k| std::unique_ptr utf16_output{ 97| 15.4k| new char16_t[expected_utf16words]}; 98| | // convert to UTF-16BE 99| 15.4k| size_t utf16words = e->convert_utf8_to_utf16be( 100| 15.4k| source.c_str(), source.size(), utf16_output.get()); 101| | // It wrote utf16words * sizeof(char16_t) bytes. 102| 15.4k| bool validutf16 = e->validate_utf16be(utf16_output.get(), utf16words); 103| 15.4k| if (!validutf16) { ------------------ | Branch (103:11): [True: 0, False: 15.4k] ------------------ 104| 0| print_input(source, e); 105| 0| abort(); 106| 0| } 107| | // convert it back: 108| | // We need a buffer of size where to write the UTF-8 words. 109| 15.4k| size_t expected_utf8words = 110| 15.4k| e->utf8_length_from_utf16be(utf16_output.get(), utf16words); 111| 15.4k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 112| | // convert to UTF-8 113| 15.4k| size_t utf8words = e->convert_utf16be_to_utf8( 114| 15.4k| utf16_output.get(), utf16words, utf8_output.get()); 115| 15.4k| std::string final_string(utf8_output.get(), utf8words); 116| 15.4k| if (final_string != source) { ------------------ | Branch (116:11): [True: 0, False: 15.4k] ------------------ 117| 0| print_input(source, e); 118| 0| abort(); 119| 0| } 120| 15.4k| } else { 121| | // invalid input!!! 122| | // We need a buffer of size where to write the UTF-16BE words. 123| 1.23k| size_t expected_utf16words = 124| 1.23k| e->utf16_length_from_utf8(source.c_str(), source.size()); 125| 1.23k| std::unique_ptr utf16_output{ 126| 1.23k| new char16_t[expected_utf16words]}; 127| | // convert to UTF-16BE 128| 1.23k| size_t utf16words = e->convert_utf8_to_utf16be( 129| 1.23k| source.c_str(), source.size(), utf16_output.get()); 130| 1.23k| if (utf16words != 0) { ------------------ | Branch (130:11): [True: 0, False: 1.23k] ------------------ 131| 0| print_input(source, e); 132| 0| abort(); 133| 0| } 134| 1.23k| } 135| | /** 136| | * Transcoding from UTF-8 to UTF-32. 137| | */ 138| 16.6k| if (validutf8) { ------------------ | Branch (138:9): [True: 15.4k, False: 1.23k] ------------------ 139| | // We need a buffer of size where to write the UTF-32 words. 140| 15.4k| size_t expected_utf32words = 141| 15.4k| e->utf32_length_from_utf8(source.c_str(), source.size()); 142| 15.4k| std::unique_ptr utf32_output{ 143| 15.4k| new char32_t[expected_utf32words]}; 144| | // convert to UTF-32 145| 15.4k| size_t utf32words = e->convert_utf8_to_utf32( 146| 15.4k| source.c_str(), source.size(), utf32_output.get()); 147| | // It wrote utf32words * sizeof(char32_t) bytes. 148| 15.4k| bool validutf32 = e->validate_utf32(utf32_output.get(), utf32words); 149| 15.4k| if (!validutf32) { ------------------ | Branch (149:11): [True: 0, False: 15.4k] ------------------ 150| 0| return -1; 151| 0| } 152| | // convert it back: 153| | // We need a buffer of size where to write the UTF-8 words. 154| 15.4k| size_t expected_utf8words = 155| 15.4k| e->utf8_length_from_utf32(utf32_output.get(), utf32words); 156| 15.4k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 157| | // convert to UTF-8 158| 15.4k| size_t utf8words = e->convert_utf32_to_utf8( 159| 15.4k| utf32_output.get(), utf32words, utf8_output.get()); 160| 15.4k| std::string final_string(utf8_output.get(), utf8words); 161| 15.4k| if (source != final_string) { ------------------ | Branch (161:11): [True: 0, False: 15.4k] ------------------ 162| 0| print_input(source, e); 163| 0| abort(); 164| 0| } 165| 15.4k| } else { 166| | // invalid input!!! 167| 1.23k| size_t expected_utf32words = 168| 1.23k| e->utf32_length_from_utf8(source.c_str(), source.size()); 169| 1.23k| std::unique_ptr utf32_output{ 170| 1.23k| new char32_t[expected_utf32words]}; 171| | // convert to UTF-32 172| 1.23k| size_t utf32words = e->convert_utf8_to_utf32( 173| 1.23k| source.c_str(), source.size(), utf32_output.get()); 174| 1.23k| if (utf32words != 0) { ------------------ | Branch (174:11): [True: 0, False: 1.23k] ------------------ 175| 0| print_input(source, e); 176| 0| abort(); 177| 0| } 178| 1.23k| } 179| | 180| | /** 181| | * Transcoding from UTF-8 to Latin 1 182| | */ 183| 16.6k| if (validutf8) { ------------------ | Branch (183:9): [True: 15.4k, False: 1.23k] ------------------ 184| | // We need a buffer of size where to write the UTF-16LE words. 185| 15.4k| size_t expected_latin1words = 186| 15.4k| e->latin1_length_from_utf8(source.c_str(), source.size()); 187| 15.4k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 188| | // convert to latin1 189| 15.4k| size_t latin1words = e->convert_utf8_to_latin1( 190| 15.4k| source.c_str(), source.size(), latin1_output.get()); 191| 15.4k| if (latin1words != 0) { ------------------ | Branch (191:11): [True: 745, False: 14.7k] ------------------ 192| | // convert it back: 193| | // We need a buffer of size where to write the UTF-8 words. 194| 745| size_t expected_utf8words = 195| 745| e->utf8_length_from_latin1(latin1_output.get(), latin1words); 196| 745| std::unique_ptr utf8_output{new char[expected_utf8words]}; 197| | // convert to UTF-8 198| 745| size_t utf8words = e->convert_latin1_to_utf8( 199| 745| latin1_output.get(), latin1words, utf8_output.get()); 200| 745| std::string final_string(utf8_output.get(), utf8words); 201| 745| if (final_string != source) { ------------------ | Branch (201:13): [True: 0, False: 745] ------------------ 202| 0| print_input(source, e); 203| 0| abort(); 204| 0| } 205| 745| } 206| 15.4k| } else { 207| | // invalid input!!! 208| | // We need a buffer of size where to write the Latin 1 words. 209| 1.23k| size_t expected_latin1words = 210| 1.23k| e->latin1_length_from_utf8(source.c_str(), source.size()); 211| 1.23k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 212| | // convert to Latin 1 213| 1.23k| size_t latin1words = e->convert_utf8_to_latin1( 214| 1.23k| source.c_str(), source.size(), latin1_output.get()); 215| 1.23k| if (latin1words != 0) { ------------------ | Branch (215:11): [True: 0, False: 1.23k] ------------------ 216| 0| print_input(source, e); 217| 0| abort(); 218| 0| } 219| 1.23k| } 220| | /** 221| | * Transcoding from UTF-16LE to UTF-8. 222| | */ 223| 16.6k| { 224| | // Get new source data here as this will allow the fuzzer to optimize it's 225| | // input for UTF16-LE. 226| 16.6k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 227| | // We copy to avoid alignment issues. 228| 16.6k| std::unique_ptr utf16_source{new char16_t[source.size() / 2]}; 229| 16.6k| if (source.data() != nullptr) { ------------------ | Branch (229:11): [True: 16.6k, False: 0] ------------------ 230| 16.6k| std::memcpy(utf16_source.get(), source.data(), source.size() / 2 * 2); 231| 16.6k| } 232| 16.6k| bool validutf16le = 233| 16.6k| e->validate_utf16le(utf16_source.get(), source.size() / 2); 234| 16.6k| auto rutf16le = e->validate_utf16le_with_errors(utf16_source.get(), 235| 16.6k| source.size() / 2); 236| 16.6k| if (validutf16le != ------------------ | Branch (236:11): [True: 0, False: 16.6k] ------------------ 237| 16.6k| (rutf16le.error == simdutf::SUCCESS)) { // they should agree 238| 0| print_input(source, e); 239| 0| abort(); 240| 0| } 241| 16.6k| if (validutf16le) { ------------------ | Branch (241:11): [True: 16.4k, False: 242] ------------------ 242| | // We need a buffer of size where to write the UTF-16 words. 243| 16.4k| size_t expected_utf8words = 244| 16.4k| e->utf8_length_from_utf16le(utf16_source.get(), source.size() / 2); 245| 16.4k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 246| 16.4k| size_t utf8words = e->convert_utf16le_to_utf8( 247| 16.4k| utf16_source.get(), source.size() / 2, utf8_output.get()); 248| | // It wrote utf16words * sizeof(char16_t) bytes. 249| 16.4k| bool validutf8 = e->validate_utf8(utf8_output.get(), utf8words); 250| 16.4k| if (!validutf8) { ------------------ | Branch (250:13): [True: 0, False: 16.4k] ------------------ 251| 0| print_input(source, e); 252| 0| abort(); 253| 0| } 254| | // convert it back: 255| | // We need a buffer of size where to write the UTF-16 words. 256| 16.4k| size_t expected_utf16words = 257| 16.4k| e->utf16_length_from_utf8(utf8_output.get(), utf8words); 258| 16.4k| std::unique_ptr utf16_output{ 259| 16.4k| new char16_t[expected_utf16words]}; 260| | // convert to UTF-8 261| 16.4k| size_t utf16words = e->convert_utf8_to_utf16le( 262| 16.4k| utf8_output.get(), utf8words, utf16_output.get()); 263| 106k| for (size_t i = 0; i < source.size() / 2; i++) { ------------------ | Branch (263:28): [True: 89.9k, False: 16.4k] ------------------ 264| 89.9k| if (utf16_output.get()[i] != (utf16_source.get())[i]) { ------------------ | Branch (264:15): [True: 0, False: 89.9k] ------------------ 265| 0| print_input(source, e); 266| 0| abort(); 267| 0| } 268| 89.9k| } 269| 16.4k| } else { 270| | // invalid input!!! 271| | // We need a buffer of size where to write the UTF-16 words. 272| 242| size_t expected_utf8words = 273| 242| e->utf8_length_from_utf16le(utf16_source.get(), source.size() / 2); 274| 242| std::unique_ptr utf8_output{new char[expected_utf8words]}; 275| 242| size_t utf8words = e->convert_utf16le_to_utf8( 276| 242| utf16_source.get(), source.size() / 2, utf8_output.get()); 277| 242| if (utf8words != 0) { ------------------ | Branch (277:13): [True: 0, False: 242] ------------------ 278| 0| print_input(source, e); 279| 0| abort(); 280| 0| } 281| 242| } 282| 16.6k| } 283| | 284| | /** 285| | * Transcoding from UTF-16BE to UTF-8. 286| | */ 287| 16.6k| { 288| | // Get new source data here as this will allow the fuzzer to optimize it's 289| | // input for UTF16-BE. 290| 16.6k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 291| 16.6k| std::unique_ptr utf16_source{new char16_t[source.size() / 2]}; 292| 16.6k| if (source.data() != nullptr) { ------------------ | Branch (292:11): [True: 16.6k, False: 0] ------------------ 293| 16.6k| std::memcpy(utf16_source.get(), source.data(), source.size() / 2 * 2); 294| 16.6k| } 295| 16.6k| bool validutf16be = 296| 16.6k| e->validate_utf16be(utf16_source.get(), source.size() / 2); 297| 16.6k| auto rutf16be = e->validate_utf16be_with_errors(utf16_source.get(), 298| 16.6k| source.size() / 2); 299| 16.6k| if (validutf16be != ------------------ | Branch (299:11): [True: 0, False: 16.6k] ------------------ 300| 16.6k| (rutf16be.error == simdutf::SUCCESS)) { // they should agree 301| 0| print_input(source, e); 302| 0| abort(); 303| 0| } 304| 16.6k| if (validutf16be) { ------------------ | Branch (304:11): [True: 16.4k, False: 221] ------------------ 305| | // We need a buffer of size where to write the UTF-16 words. 306| 16.4k| size_t expected_utf8words = 307| 16.4k| e->utf8_length_from_utf16be(utf16_source.get(), source.size() / 2); 308| 16.4k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 309| 16.4k| size_t utf8words = e->convert_utf16be_to_utf8( 310| 16.4k| utf16_source.get(), source.size() / 2, utf8_output.get()); 311| | // It wrote utf16words * sizeof(char16_t) bytes. 312| 16.4k| bool validutf8 = e->validate_utf8(utf8_output.get(), utf8words); 313| 16.4k| if (!validutf8) { ------------------ | Branch (313:13): [True: 0, False: 16.4k] ------------------ 314| 0| print_input(source, e); 315| 0| abort(); 316| 0| } 317| | // convert it back: 318| | // We need a buffer of size where to write the UTF-16 words. 319| 16.4k| size_t expected_utf16words = 320| 16.4k| e->utf16_length_from_utf8(utf8_output.get(), utf8words); 321| 16.4k| std::unique_ptr utf16_output{ 322| 16.4k| new char16_t[expected_utf16words]}; 323| | // convert to UTF-8 324| 16.4k| size_t utf16words = e->convert_utf8_to_utf16be( 325| 16.4k| utf8_output.get(), utf8words, utf16_output.get()); 326| 109k| for (size_t i = 0; i < source.size() / 2; i++) { ------------------ | Branch (326:28): [True: 92.9k, False: 16.4k] ------------------ 327| 92.9k| if (utf16_output.get()[i] != (utf16_source.get())[i]) { ------------------ | Branch (327:15): [True: 0, False: 92.9k] ------------------ 328| 0| print_input(source, e); 329| 0| abort(); 330| 0| } 331| 92.9k| } 332| 16.4k| } else { 333| | // invalid input!!! 334| | // We need a buffer of size where to write the UTF-16 words. 335| 221| size_t expected_utf8words = 336| 221| e->utf8_length_from_utf16be(utf16_source.get(), source.size() / 2); 337| 221| std::unique_ptr utf8_output{new char[expected_utf8words]}; 338| 221| size_t utf8words = e->convert_utf16be_to_utf8( 339| 221| utf16_source.get(), source.size() / 2, utf8_output.get()); 340| 221| if (utf8words != 0) { ------------------ | Branch (340:13): [True: 0, False: 221] ------------------ 341| 0| print_input(source, e); 342| 0| abort(); 343| 0| } 344| 221| } 345| 16.6k| } 346| | 347| | /** 348| | * Transcoding from latin1 to UTF-8. 349| | */ 350| | // Get new source data here as this will allow the fuzzer to optimize it's 351| | // input for latin1. 352| 16.6k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 353| 16.6k| bool validlatin1 = true; // has to be 354| 16.6k| if (validlatin1) { ------------------ | Branch (354:9): [True: 16.6k, False: 0] ------------------ 355| | // We need a buffer of size where to write the UTF-8 words. 356| 16.6k| size_t expected_utf8words = 357| 16.6k| e->utf8_length_from_latin1(source.c_str(), source.size()); 358| 16.6k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 359| 16.6k| size_t utf8words = e->convert_latin1_to_utf8( 360| 16.6k| source.c_str(), source.size(), utf8_output.get()); 361| | // It wrote utf8words * sizeof(char) bytes. 362| 16.6k| bool validutf8 = e->validate_utf8(utf8_output.get(), utf8words); 363| 16.6k| if (!validutf8) { ------------------ | Branch (363:11): [True: 0, False: 16.6k] ------------------ 364| 0| print_input(source, e); 365| 0| abort(); 366| 0| } 367| | // convert it back: 368| | // We need a buffer of size where to write the latin1 words. 369| 16.6k| size_t expected_latin1words = 370| 16.6k| e->latin1_length_from_utf8(utf8_output.get(), utf8words); 371| 16.6k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 372| | // convert to latin1 373| 16.6k| size_t latin1words = e->convert_utf8_to_latin1( 374| 16.6k| utf8_output.get(), utf8words, latin1_output.get()); 375| 430k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (375:26): [True: 414k, False: 16.6k] ------------------ 376| 414k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (376:13): [True: 0, False: 414k] ------------------ 377| 0| print_input(source, e); 378| 0| abort(); 379| 0| } 380| 414k| } 381| 16.6k| } 382| 16.6k| if (validlatin1) { ------------------ | Branch (382:9): [True: 16.6k, False: 0] ------------------ 383| | // We need a buffer of size where to write the UTF-16 words. 384| 16.6k| size_t expected_utf16words = e->utf16_length_from_latin1(source.size()); 385| 16.6k| std::unique_ptr utf16_output{ 386| 16.6k| new char16_t[expected_utf16words]}; 387| 16.6k| size_t utf16words = e->convert_latin1_to_utf16le( 388| 16.6k| source.c_str(), source.size(), utf16_output.get()); 389| | // It wrote utf16words * sizeof(char16_t) bytes. 390| 16.6k| bool validutf16 = e->validate_utf16le(utf16_output.get(), utf16words); 391| 16.6k| if (!validutf16) { ------------------ | Branch (391:11): [True: 0, False: 16.6k] ------------------ 392| 0| print_input(source, e); 393| 0| abort(); 394| 0| } 395| | // convert it back: 396| | // We need a buffer of size where to write the latin1 words. 397| 16.6k| size_t expected_latin1words = e->latin1_length_from_utf16(utf16words); 398| 16.6k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 399| | // convert to latin1 400| 16.6k| size_t latin1words = e->convert_utf16le_to_latin1( 401| 16.6k| utf16_output.get(), utf16words, latin1_output.get()); 402| 430k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (402:26): [True: 414k, False: 16.6k] ------------------ 403| 414k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (403:13): [True: 0, False: 414k] ------------------ 404| 0| print_input(source, e); 405| 0| abort(); 406| 0| } 407| 414k| } 408| 16.6k| } 409| 16.6k| if (validlatin1) { ------------------ | Branch (409:9): [True: 16.6k, False: 0] ------------------ 410| | // We need a buffer of size where to write the UTF-16 words. 411| 16.6k| size_t expected_utf16words = e->utf16_length_from_latin1(source.size()); 412| 16.6k| std::unique_ptr utf16_output{ 413| 16.6k| new char16_t[expected_utf16words]}; 414| 16.6k| size_t utf16words = e->convert_latin1_to_utf16be( 415| 16.6k| source.c_str(), source.size(), utf16_output.get()); 416| | // It wrote utf16words * sizeof(char16_t) bytes. 417| 16.6k| bool validutf16 = e->validate_utf16be(utf16_output.get(), utf16words); 418| 16.6k| if (!validutf16) { ------------------ | Branch (418:11): [True: 0, False: 16.6k] ------------------ 419| 0| print_input(source, e); 420| 0| abort(); 421| 0| } 422| | // convert it back: 423| | // We need a buffer of size where to write the latin1 words. 424| 16.6k| size_t expected_latin1words = e->latin1_length_from_utf16(utf16words); 425| 16.6k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 426| | // convert to latin1 427| 16.6k| size_t latin1words = e->convert_utf16be_to_latin1( 428| 16.6k| utf16_output.get(), utf16words, latin1_output.get()); 429| 430k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (429:26): [True: 414k, False: 16.6k] ------------------ 430| 414k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (430:13): [True: 0, False: 414k] ------------------ 431| 0| print_input(source, e); 432| 0| abort(); 433| 0| } 434| 414k| } 435| 16.6k| } 436| | 437| 16.6k| if (validlatin1) { ------------------ | Branch (437:9): [True: 16.6k, False: 0] ------------------ 438| | // We need a buffer of size where to write the UTF-16 words. 439| 16.6k| size_t expected_utf32words = e->utf32_length_from_latin1(source.size()); 440| 16.6k| std::unique_ptr utf32_output{ 441| 16.6k| new char32_t[expected_utf32words]}; 442| 16.6k| size_t utf32words = e->convert_latin1_to_utf32( 443| 16.6k| source.c_str(), source.size(), utf32_output.get()); 444| | // It wrote utf16words * sizeof(char16_t) bytes. 445| 16.6k| bool validutf32 = e->validate_utf32(utf32_output.get(), utf32words); 446| 16.6k| if (!validutf32) { ------------------ | Branch (446:11): [True: 0, False: 16.6k] ------------------ 447| 0| print_input(source, e); 448| 0| abort(); 449| 0| } 450| | // convert it back: 451| | // We need a buffer of size where to write the latin1 words. 452| 16.6k| size_t expected_latin1words = e->latin1_length_from_utf32(utf32words); 453| 16.6k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 454| | // convert to latin1 455| 16.6k| size_t latin1words = e->convert_utf32_to_latin1( 456| 16.6k| utf32_output.get(), utf32words, latin1_output.get()); 457| 430k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (457:26): [True: 414k, False: 16.6k] ------------------ 458| 414k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (458:13): [True: 0, False: 414k] ------------------ 459| 0| print_input(source, e); 460| 0| abort(); 461| 0| } 462| 414k| } 463| 16.6k| } 464| | 465| | /// Base64 tests. We begin by trying to decode the input, even if we 466| | /// expect it to fail. 467| 16.6k| { 468| 16.6k| size_t max_length_needed = 469| 16.6k| e->maximal_binary_length_from_base64(source.data(), source.size()); 470| 16.6k| std::vector back(max_length_needed); 471| 16.6k| simdutf::result r = 472| 16.6k| e->base64_to_binary(source.data(), source.size(), back.data()); 473| 16.6k| if (r.error == simdutf::error_code::SUCCESS) { ------------------ | Branch (473:11): [True: 15.2k, False: 1.46k] ------------------ 474| | // We expect failure but if we succeed, then we should have a roundtrip. 475| 15.2k| back.resize(r.count); 476| 15.2k| std::vector back2(e->base64_length_from_binary(back.size())); 477| 15.2k| size_t base64size = 478| 15.2k| e->binary_to_base64(back.data(), back.size(), back2.data()); 479| 15.2k| back2.resize(base64size); 480| 15.2k| std::vector back3( 481| 15.2k| e->maximal_binary_length_from_base64(back2.data(), back2.size())); 482| 15.2k| simdutf::result r2 = 483| 15.2k| e->base64_to_binary(back2.data(), back2.size(), back3.data()); 484| 15.2k| if (r2.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (484:13): [True: 0, False: 15.2k] ------------------ 485| 0| print_input(source, e); 486| 0| return false; 487| 0| } 488| 15.2k| if (r2.count != back.size()) { ------------------ | Branch (488:13): [True: 0, False: 15.2k] ------------------ 489| 0| print_input(source, e); 490| 0| return false; 491| 0| } 492| 15.2k| if (back3.size() != back.size()) { ------------------ | Branch (492:13): [True: 0, False: 15.2k] ------------------ 493| 0| print_input(source, e); 494| 0| return false; 495| 0| } 496| 15.2k| } 497| 16.6k| } 498| | 499| | // Same as above, but we use the safe decoder version. 500| 16.6k| { 501| 16.6k| size_t max_length_needed = 502| 16.6k| e->maximal_binary_length_from_base64(source.data(), source.size()); 503| 16.6k| std::vector back(max_length_needed); 504| 16.6k| simdutf::result r = simdutf::base64_to_binary_safe( 505| 16.6k| source.data(), source.size(), back.data(), max_length_needed); 506| 16.6k| if (r.error == simdutf::error_code::SUCCESS) { ------------------ | Branch (506:11): [True: 15.2k, False: 1.46k] ------------------ 507| | // We expect failure but if we succeed, then we should have a roundtrip. 508| 15.2k| back.resize(max_length_needed); 509| 15.2k| std::vector back2(e->base64_length_from_binary(back.size())); 510| 15.2k| size_t base64size = 511| 15.2k| e->binary_to_base64(back.data(), back.size(), back2.data()); 512| 15.2k| back2.resize(base64size); 513| 15.2k| size_t max_length_needed2 = 514| 15.2k| e->maximal_binary_length_from_base64(back2.data(), back2.size()); 515| 15.2k| std::vector back3(max_length_needed2); 516| 15.2k| simdutf::result r2 = simdutf::base64_to_binary_safe( 517| 15.2k| back2.data(), back2.size(), back3.data(), max_length_needed2); 518| 15.2k| if (r2.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (518:13): [True: 0, False: 15.2k] ------------------ 519| 0| print_input(source, e); 520| 0| return false; 521| 0| } 522| 15.2k| if (max_length_needed != back.size()) { ------------------ | Branch (522:13): [True: 0, False: 15.2k] ------------------ 523| 0| print_input(source, e); 524| 0| return false; 525| 0| } 526| 15.2k| if (back3.size() != back.size()) { ------------------ | Branch (526:13): [True: 0, False: 15.2k] ------------------ 527| 0| print_input(source, e); 528| 0| return false; 529| 0| } 530| 15.2k| } 531| 16.6k| } 532| | /// Base64 tests. We encode the content as binary in base64 and we decode 533| | /// it, it should always succeed. 534| 16.6k| { 535| 16.6k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 536| 16.6k| std::vector base64buffer( 537| 16.6k| e->base64_length_from_binary(source.size())); 538| 16.6k| size_t base64size = e->binary_to_base64(source.data(), source.size(), 539| 16.6k| base64buffer.data()); 540| 16.6k| if (base64size != base64buffer.size()) { ------------------ | Branch (540:11): [True: 0, False: 16.6k] ------------------ 541| 0| print_input(source, e); 542| 0| abort(); 543| 0| } 544| 16.6k| std::vector back(e->maximal_binary_length_from_base64( 545| 16.6k| base64buffer.data(), base64buffer.size())); 546| 16.6k| simdutf::result r = e->base64_to_binary(base64buffer.data(), 547| 16.6k| base64buffer.size(), back.data()); 548| 16.6k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (548:11): [True: 0, False: 16.6k] ------------------ 549| 0| print_input(source, e); 550| 0| abort(); 551| 0| } 552| 16.6k| if (r.count != source.size()) { ------------------ | Branch (552:11): [True: 0, False: 16.6k] ------------------ 553| 0| print_input(source, e); 554| 0| abort(); 555| 0| } 556| 376k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (556:26): [True: 360k, False: 16.6k] ------------------ 557| 360k| if (back[i] != (source.c_str())[i]) { ------------------ | Branch (557:13): [True: 0, False: 360k] ------------------ 558| 0| print_input(source, e); 559| 0| abort(); 560| 0| } 561| 360k| } 562| 16.6k| size_t max_length = back.size(); 563| 16.6k| r = simdutf::base64_to_binary_safe( 564| 16.6k| base64buffer.data(), base64buffer.size(), back.data(), max_length); 565| 16.6k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (565:11): [True: 0, False: 16.6k] ------------------ 566| 0| printf("base64 round trip failed, error code %d\n", r.error); 567| 0| print_input(source, e); 568| 0| return false; 569| 0| } 570| 16.6k| if (max_length != source.size()) { ------------------ | Branch (570:11): [True: 0, False: 16.6k] ------------------ 571| 0| printf("base64 safe round trip failed, not the same size %zu %zu\n", 572| 0| max_length, source.size()); 573| 0| print_input(source, e); 574| 0| return false; 575| 0| } 576| 376k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (576:26): [True: 360k, False: 16.6k] ------------------ 577| 360k| if (back[i] != (source.c_str())[i]) { ------------------ | Branch (577:13): [True: 0, False: 360k] ------------------ 578| 0| printf("base64 round trip failed, same size, different content\n"); 579| 0| print_input(source, e); 580| 0| return false; 581| 0| } 582| 360k| } 583| 16.6k| } 584| | 585| 16.6k| } // for (auto &e : simdutf::get_available_implementations()) { 586| | 587| 5.56k| return 0; 588| 5.56k|} // extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) { _ZN7simdutf26base64_to_binary_safe_implIcEENS_6resultEPKT_mPcRmNS_14base64_optionsENS_27last_chunk_handling_optionsEb: 63| 48.5k| bool decode_up_to_bad_char) noexcept { 64| 48.5k| static_assert(std::is_same::value || 65| 48.5k| std::is_same::value, 66| 48.5k| "Only char and char16_t are supported."); 67| 48.5k| size_t remaining_input_length = length; 68| 48.5k| size_t remaining_output_length = outlen; 69| 48.5k| size_t input_position = 0; 70| 48.5k| size_t output_position = 0; 71| | 72| | // We also do a first pass using the fast path to decode as much as possible 73| 48.5k| size_t safe_input = (std::min)( 74| 48.5k| remaining_input_length, 75| 48.5k| base64_length_from_binary(remaining_output_length / 3 * 3, options)); 76| 48.5k| bool done_with_partial = (safe_input == remaining_input_length); 77| 48.5k| simdutf::full_result r; 78| | 79| |#if SIMDUTF_CPLUSPLUS23 80| | if consteval { 81| | r = scalar::base64::base64_to_binary_details_impl( 82| | input + input_position, safe_input, output + output_position, options, 83| | done_with_partial 84| | ? last_chunk_handling_options 85| | : simdutf::last_chunk_handling_options::only_full_chunks); 86| | } else 87| |#endif 88| 48.5k| { 89| 48.5k| r = get_active_implementation()->base64_to_binary_details( 90| 48.5k| input + input_position, safe_input, output + output_position, options, 91| 48.5k| done_with_partial ------------------ | Branch (91:9): [True: 45.7k, False: 2.87k] ------------------ 92| 48.5k| ? last_chunk_handling_options 93| 48.5k| : simdutf::last_chunk_handling_options::only_full_chunks); 94| 48.5k| } 95| 48.5k| simdutf_log_assert(r.input_count <= safe_input, 96| 48.5k| "You should not read more than safe_input"); 97| 48.5k| simdutf_log_assert(r.output_count <= remaining_output_length, 98| 48.5k| "You should not write more than remaining_output_length"); 99| | // Technically redundant, but we want to be explicit about it. 100| 48.5k| input_position += r.input_count; 101| 48.5k| output_position += r.output_count; 102| 48.5k| remaining_input_length -= r.input_count; 103| 48.5k| remaining_output_length -= r.output_count; 104| 48.5k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (104:7): [True: 1.14k, False: 47.4k] ------------------ 105| | // There is an error. We return. 106| 1.14k| if (decode_up_to_bad_char && ------------------ | Branch (106:9): [True: 0, False: 1.14k] ------------------ 107| 0| r.error == error_code::INVALID_BASE64_CHARACTER) { ------------------ | Branch (107:9): [True: 0, False: 0] ------------------ 108| 0| return slow_base64_to_binary_safe_impl( 109| 0| input, length, output, outlen, options, last_chunk_handling_options); 110| 0| } 111| 1.14k| outlen = output_position; 112| 1.14k| return {r.error, input_position}; 113| 1.14k| } 114| | 115| 47.4k| if (done_with_partial) { ------------------ | Branch (115:7): [True: 45.0k, False: 2.37k] ------------------ 116| | // We are done. We have decoded everything. 117| 45.0k| outlen = output_position; 118| 45.0k| return {simdutf::error_code::SUCCESS, input_position}; 119| 45.0k| } 120| | // We have decoded some data, but we still have some data to decode. 121| | // We need to decode the rest of the input buffer. 122| 2.37k| r = simdutf::scalar::base64::base64_to_binary_details_safe_impl( 123| 2.37k| input + input_position, remaining_input_length, output + output_position, 124| 2.37k| remaining_output_length, options, last_chunk_handling_options); 125| 2.37k| input_position += r.input_count; 126| 2.37k| output_position += r.output_count; 127| 2.37k| remaining_input_length -= r.input_count; 128| 2.37k| remaining_output_length -= r.output_count; 129| | 130| 2.37k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (130:7): [True: 324, False: 2.05k] ------------------ 131| | // There is an error. We return. 132| 324| if (decode_up_to_bad_char && ------------------ | Branch (132:9): [True: 0, False: 324] ------------------ 133| 0| r.error == error_code::INVALID_BASE64_CHARACTER) { ------------------ | Branch (133:9): [True: 0, False: 0] ------------------ 134| 0| return slow_base64_to_binary_safe_impl( 135| 0| input, length, output, outlen, options, last_chunk_handling_options); 136| 0| } 137| 324| outlen = output_position; 138| 324| return {r.error, input_position}; 139| 324| } 140| 2.05k| if (input_position < length) { ------------------ | Branch (140:7): [True: 0, False: 2.05k] ------------------ 141| | // We cannot process the entire input in one go, so we need to 142| | // process it in two steps: first the fast path, then the slow path. 143| | // In some cases, the processing might 'eat up' trailing ignorable 144| | // characters in the fast path, but that can be a problem. 145| | // suppose we have just white space followed by a single base64 character. 146| | // If we first process the white space with the fast path, it will 147| | // eat all of it. But, by the JavaScript standard, we should consume 148| | // no character. See 149| | // https://tc39.es/proposal-arraybuffer-base64/spec/#sec-frombase64 150| 0| while (input_position > 0 && ------------------ | Branch (150:12): [True: 0, False: 0] ------------------ 151| 0| base64_ignorable(input[input_position - 1], options)) { ------------------ | Branch (151:12): [True: 0, False: 0] ------------------ 152| 0| input_position--; 153| 0| } 154| 0| } 155| 2.05k| outlen = output_position; 156| 2.05k| return {simdutf::error_code::SUCCESS, input_position}; 157| 2.37k|} _ZN7simdutf6resultC2ENS_10error_codeEm: 87| 158k| : error{err}, count{pos} {} _ZNK7simdutf6result6is_errEv: 93| 8.50k| simdutf_really_inline simdutf_constexpr23 bool is_err() const noexcept { 94| 8.50k| return error != error_code::SUCCESS; 95| 8.50k| } _ZN7simdutf11full_resultC2Ev: 106| 48.5k| : error{error_code::SUCCESS}, input_count{0}, output_count{0} {} _ZN7simdutf11full_resultC2ENS_10error_codeEmm: 111| 99.4k| : error{err}, input_count{pos_in}, output_count{pos_out} {} _ZN7simdutf11full_resultC2ENS_10error_codeEmmb: 114| 127| : error{err}, input_count{pos_in}, output_count{pos_out}, 115| 127| padding_error{padding_err} {} _ZNK7simdutf11full_resultcvNS_6resultEEv: 117| 48.5k| simdutf_really_inline simdutf_constexpr23 operator result() const noexcept { 118| 48.5k| if (error == error_code::SUCCESS) { ------------------ | Branch (118:9): [True: 47.1k, False: 1.46k] ------------------ 119| 47.1k| return result{error, output_count}; 120| 47.1k| } else { 121| 1.46k| return result{error, input_count}; 122| 1.46k| } 123| 48.5k| } _ZN7simdutf10is_partialENS_27last_chunk_handling_optionsE: 4122| 20.0k|is_partial(last_chunk_handling_options options) { 4123| 20.0k| return (options == stop_before_partial) || (options == only_full_chunks); ------------------ | Branch (4123:10): [True: 0, False: 20.0k] | Branch (4123:46): [True: 3.67k, False: 16.3k] ------------------ 4124| 20.0k|} _ZNK7simdutf14implementation25required_instruction_setsEv: 5143| 22.2k| virtual uint32_t required_instruction_sets() const { 5144| 22.2k| return _required_instruction_sets; 5145| 22.2k| } _ZNK7simdutf14implementation24utf16_length_from_latin1Em: 6333| 33.3k| utf16_length_from_latin1(size_t length) const noexcept { 6334| 33.3k| return length; 6335| 33.3k| } _ZNK7simdutf14implementation24latin1_length_from_utf32Em: 6521| 16.6k| latin1_length_from_utf32(size_t length) const noexcept { 6522| 16.6k| return length; 6523| 16.6k| } _ZNK7simdutf14implementation24latin1_length_from_utf16Em: 6558| 33.3k| latin1_length_from_utf16(size_t length) const noexcept { 6559| 33.3k| return length; 6560| 33.3k| } _ZNK7simdutf14implementation24utf32_length_from_latin1Em: 6591| 16.6k| utf32_length_from_latin1(size_t length) const noexcept { 6592| 16.6k| return length; 6593| 16.6k| } _ZN7simdutf14implementationC2EPKcS2_j: 7029| 5| : _name(name), _description(description), 7030| 5| _required_instruction_sets(required_instruction_sets) {} _ZN7simdutf8internal29available_implementation_listC2Ev: 7061| 1| simdutf_really_inline available_implementation_list() {} _ZN7simdutf25base64_length_from_binaryEmNS_14base64_optionsE: 4433| 48.5k| size_t length, base64_options options = base64_default) noexcept { 4434| 48.5k| return scalar::base64::base64_length_from_binary(length, options); 4435| 48.5k|} _ZN7simdutf16base64_ignorableEcNS_14base64_optionsE: 4848| 3.48k|base64_ignorable(char input, base64_options options = base64_default) noexcept { 4849| 3.48k| return scalar::base64::is_ignorable(input, options); 4850| 3.48k|} _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEaSEPS3_: 7132| 1| atomic_ptr &operator=(T *_ptr) { 7133| 1| ptr = _ptr; 7134| 1| return *this; 7135| 1| } _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEcvPS3_Ev: 7129| 1| operator T *() { return ptr.load(); } _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEC2EPS3_: 7109| 1| atomic_ptr(T *_ptr) : ptr{_ptr} {} _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEptEv: 7131| 48.5k| T *operator->() { return ptr.load(); } simdutf.cpp:_ZN7simdutf8internalL30detect_supported_architecturesEv: 236| 22.2k|static inline uint32_t detect_supported_architectures() { 237| 22.2k| uint32_t eax; 238| 22.2k| uint32_t ebx = 0; 239| 22.2k| uint32_t ecx = 0; 240| 22.2k| uint32_t edx = 0; 241| 22.2k| uint32_t host_isa = 0x0; 242| | 243| | // EBX for EAX=0x1 244| 22.2k| eax = 0x1; 245| 22.2k| cpuid(&eax, &ebx, &ecx, &edx); 246| | 247| 22.2k| if (ecx & cpuid_bit::sse42) { ------------------ | Branch (247:7): [True: 22.2k, False: 0] ------------------ 248| 22.2k| host_isa |= instruction_set::SSE42; 249| 22.2k| } 250| | 251| 22.2k| if (ecx & cpuid_bit::pclmulqdq) { ------------------ | Branch (251:7): [True: 22.2k, False: 0] ------------------ 252| 22.2k| host_isa |= instruction_set::PCLMULQDQ; 253| 22.2k| } 254| | 255| 22.2k| if ((ecx & cpuid_bit::osxsave) != cpuid_bit::osxsave) { ------------------ | Branch (255:7): [True: 0, False: 22.2k] ------------------ 256| 0| return host_isa; 257| 0| } 258| | 259| | // xgetbv for checking if the OS saves registers 260| 22.2k| uint64_t xcr0 = xgetbv(); 261| | 262| 22.2k| if ((xcr0 & cpuid_bit::xcr0_bit::avx256_saved) == 0) { ------------------ | Branch (262:7): [True: 0, False: 22.2k] ------------------ 263| 0| return host_isa; 264| 0| } 265| | // ECX for EAX=0x7 266| 22.2k| eax = 0x7; 267| 22.2k| ecx = 0x0; // Sub-leaf = 0 268| 22.2k| cpuid(&eax, &ebx, &ecx, &edx); 269| 22.2k| if (ebx & cpuid_bit::ebx::avx2) { ------------------ | Branch (269:7): [True: 22.2k, False: 0] ------------------ 270| 22.2k| host_isa |= instruction_set::AVX2; 271| 22.2k| } 272| 22.2k| if (ebx & cpuid_bit::ebx::bmi1) { ------------------ | Branch (272:7): [True: 22.2k, False: 0] ------------------ 273| 22.2k| host_isa |= instruction_set::BMI1; 274| 22.2k| } 275| 22.2k| if (ebx & cpuid_bit::ebx::bmi2) { ------------------ | Branch (275:7): [True: 22.2k, False: 0] ------------------ 276| 22.2k| host_isa |= instruction_set::BMI2; 277| 22.2k| } 278| 22.2k| if (!((xcr0 & cpuid_bit::xcr0_bit::avx512_saved) == ------------------ | Branch (278:7): [True: 22.2k, False: 0] ------------------ 279| 22.2k| cpuid_bit::xcr0_bit::avx512_saved)) { 280| 22.2k| return host_isa; 281| 22.2k| } 282| 0| if (ebx & cpuid_bit::ebx::avx512f) { ------------------ | Branch (282:7): [True: 0, False: 0] ------------------ 283| 0| host_isa |= instruction_set::AVX512F; 284| 0| } 285| 0| if (ebx & cpuid_bit::ebx::avx512bw) { ------------------ | Branch (285:7): [True: 0, False: 0] ------------------ 286| 0| host_isa |= instruction_set::AVX512BW; 287| 0| } 288| 0| if (ebx & cpuid_bit::ebx::avx512cd) { ------------------ | Branch (288:7): [True: 0, False: 0] ------------------ 289| 0| host_isa |= instruction_set::AVX512CD; 290| 0| } 291| 0| if (ebx & cpuid_bit::ebx::avx512dq) { ------------------ | Branch (291:7): [True: 0, False: 0] ------------------ 292| 0| host_isa |= instruction_set::AVX512DQ; 293| 0| } 294| 0| if (ebx & cpuid_bit::ebx::avx512vl) { ------------------ | Branch (294:7): [True: 0, False: 0] ------------------ 295| 0| host_isa |= instruction_set::AVX512VL; 296| 0| } 297| 0| if (ecx & cpuid_bit::ecx::avx512vbmi2) { ------------------ | Branch (297:7): [True: 0, False: 0] ------------------ 298| 0| host_isa |= instruction_set::AVX512VBMI2; 299| 0| } 300| 0| if (ecx & cpuid_bit::ecx::avx512vpopcnt) { ------------------ | Branch (300:7): [True: 0, False: 0] ------------------ 301| 0| host_isa |= instruction_set::AVX512VPOPCNTDQ; 302| 0| } 303| 0| return host_isa; 304| 22.2k|} simdutf.cpp:_ZN7simdutf8internalL5cpuidEPjS1_S1_S1_: 202| 44.4k| uint32_t *edx) { 203| | #if defined(_MSC_VER) 204| | int cpu_info[4]; 205| | __cpuidex(cpu_info, *eax, *ecx); 206| | *eax = cpu_info[0]; 207| | *ebx = cpu_info[1]; 208| | *ecx = cpu_info[2]; 209| | *edx = cpu_info[3]; 210| | #elif (defined(HAVE_GCC_GET_CPUID) && defined(USE_GCC_GET_CPUID)) || \ 211| | defined(__FILC__) 212| | uint32_t level = *eax; 213| | __get_cpuid(level, eax, ebx, ecx, edx); 214| | #else 215| 44.4k| uint32_t a = *eax, b, c = *ecx, d; 216| 44.4k| asm volatile("cpuid\n\t" : "+a"(a), "=b"(b), "+c"(c), "=d"(d)); 217| 44.4k| *eax = a; 218| 44.4k| *ebx = b; 219| 44.4k| *ecx = c; 220| 44.4k| *edx = d; 221| 44.4k| #endif 222| 44.4k|} simdutf.cpp:_ZN7simdutf8internalL6xgetbvEv: 224| 22.2k|static inline uint64_t xgetbv() { 225| | #if defined(_MSC_VER) 226| | return _xgetbv(0); 227| | #elif defined(__FILC__) 228| | return zxgetbv(); 229| | #else 230| 22.2k| uint32_t xcr0_lo, xcr0_hi; 231| 22.2k| asm volatile("xgetbv\n\t" : "=a"(xcr0_lo), "=d"(xcr0_hi) : "c"(0)); 232| 22.2k| return xcr0_lo | ((uint64_t)xcr0_hi << 32); 233| 22.2k| #endif 234| 22.2k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6425base64_length_from_binaryEmNS_14base64_optionsE: 836| 95.7k|base64_length_from_binary(size_t length, base64_options options) noexcept { 837| | // By default, we use padding if we are not using the URL variant. 838| | // This is check with ((options & base64_url) == 0) which returns true if we 839| | // are not using the URL variant. However, we also allow 'inversion' of the 840| | // convention with the base64_reverse_padding option. If the 841| | // base64_reverse_padding option is set, we use padding if we are using the 842| | // URL variant, and we omit it if we are not using the URL variant. This is 843| | // checked with 844| | // ((options & base64_reverse_padding) == base64_reverse_padding). 845| 95.7k| bool use_padding = 846| 95.7k| ((options & base64_url) == 0) ^ 847| 95.7k| ((options & base64_reverse_padding) == base64_reverse_padding); 848| 95.7k| if (!use_padding) { ------------------ | Branch (848:7): [True: 0, False: 95.7k] ------------------ 849| 0| return length / 3 * 4 + ((length % 3) ? (length % 3) + 1 : 0); ------------------ | Branch (849:30): [True: 0, False: 0] ------------------ 850| 0| } 851| 95.7k| return (length + 2) / 3 * 852| 95.7k| 4; // We use padding to make the length a multiple of 4. 853| 95.7k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6418tail_encode_base64EPcPKcmNS_14base64_optionsE: 648| 16.3k| char *dst, const char *src, size_t srclen, base64_options options) { 649| 16.3k| return tail_encode_base64_impl(dst, src, srclen, options); 650| 16.3k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423tail_encode_base64_implILb0EEEmPcPKcmNS_14base64_optionsEmm: 418| 47.1k| size_t line_length = simdutf::default_line_length, size_t line_offset = 0) { 419| | if constexpr (use_lines) { 420| | // sanitize line_length and starting_line_offset. 421| | // line_length must be greater than 3. 422| | if (line_length < 4) { 423| | line_length = 4; 424| | } 425| | simdutf_log_assert(line_offset <= line_length, 426| | "line_offset should be less than line_length"); 427| | } 428| | // By default, we use padding if we are not using the URL variant. 429| | // This is check with ((options & base64_url) == 0) which returns true if we 430| | // are not using the URL variant. However, we also allow 'inversion' of the 431| | // convention with the base64_reverse_padding option. If the 432| | // base64_reverse_padding option is set, we use padding if we are using the 433| | // URL variant, and we omit it if we are not using the URL variant. This is 434| | // checked with 435| | // ((options & base64_reverse_padding) == base64_reverse_padding). 436| 47.1k| bool use_padding = 437| 47.1k| ((options & base64_url) == 0) ^ 438| 47.1k| ((options & base64_reverse_padding) == base64_reverse_padding); 439| | // This looks like 3 branches, but we expect the compiler to resolve this to 440| | // a single branch: 441| 47.1k| const char *e0 = (options & base64_url) ? tables::base64::base64_url::e0 ------------------ | Branch (441:20): [True: 0, False: 47.1k] ------------------ 442| 47.1k| : tables::base64::base64_default::e0; 443| 47.1k| const char *e1 = (options & base64_url) ? tables::base64::base64_url::e1 ------------------ | Branch (443:20): [True: 0, False: 47.1k] ------------------ 444| 47.1k| : tables::base64::base64_default::e1; 445| 47.1k| const char *e2 = (options & base64_url) ? tables::base64::base64_url::e2 ------------------ | Branch (445:20): [True: 0, False: 47.1k] ------------------ 446| 47.1k| : tables::base64::base64_default::e2; 447| 47.1k| char *out = dst; 448| 47.1k| size_t i = 0; 449| 47.1k| uint8_t t1, t2, t3; 450| 92.4k| for (; i + 2 < srclen; i += 3) { ------------------ | Branch (450:10): [True: 45.3k, False: 47.1k] ------------------ 451| 45.3k| t1 = uint8_t(src[i]); 452| 45.3k| t2 = uint8_t(src[i + 1]); 453| 45.3k| t3 = uint8_t(src[i + 2]); 454| | if constexpr (use_lines) { 455| | if (line_offset + 3 >= line_length) { 456| | if (line_offset == line_length) { 457| | *out++ = '\n'; 458| | *out++ = e0[t1]; 459| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 460| | *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 461| | *out++ = e2[t3]; 462| | line_offset = 4; 463| | } else if (line_offset + 1 == line_length) { 464| | *out++ = e0[t1]; 465| | *out++ = '\n'; 466| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 467| | *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 468| | *out++ = e2[t3]; 469| | line_offset = 3; 470| | } else if (line_offset + 2 == line_length) { 471| | *out++ = e0[t1]; 472| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 473| | *out++ = '\n'; 474| | *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 475| | *out++ = e2[t3]; 476| | line_offset = 2; 477| | } else if (line_offset + 3 == line_length) { 478| | *out++ = e0[t1]; 479| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 480| | *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 481| | *out++ = '\n'; 482| | *out++ = e2[t3]; 483| | line_offset = 1; 484| | } 485| | } else { 486| | *out++ = e0[t1]; 487| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 488| | *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 489| | *out++ = e2[t3]; 490| | line_offset += 4; 491| | } 492| 45.3k| } else { 493| 45.3k| *out++ = e0[t1]; 494| 45.3k| *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 495| 45.3k| *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 496| 45.3k| *out++ = e2[t3]; 497| 45.3k| } 498| 45.3k| } 499| 47.1k| switch (srclen - i) { 500| 44.9k| case 0: ------------------ | Branch (500:3): [True: 44.9k, False: 2.13k] ------------------ 501| 44.9k| break; 502| 1.35k| case 1: ------------------ | Branch (502:3): [True: 1.35k, False: 45.7k] ------------------ 503| 1.35k| t1 = uint8_t(src[i]); 504| | if constexpr (use_lines) { 505| | if (use_padding) { 506| | if (line_offset + 3 >= line_length) { 507| | if (line_offset == line_length) { 508| | *out++ = '\n'; 509| | *out++ = e0[t1]; 510| | *out++ = e1[(t1 & 0x03) << 4]; 511| | *out++ = '='; 512| | *out++ = '='; 513| | } else if (line_offset + 1 == line_length) { 514| | *out++ = e0[t1]; 515| | *out++ = '\n'; 516| | *out++ = e1[(t1 & 0x03) << 4]; 517| | *out++ = '='; 518| | *out++ = '='; 519| | } else if (line_offset + 2 == line_length) { 520| | *out++ = e0[t1]; 521| | *out++ = e1[(t1 & 0x03) << 4]; 522| | *out++ = '\n'; 523| | *out++ = '='; 524| | *out++ = '='; 525| | } else if (line_offset + 3 == line_length) { 526| | *out++ = e0[t1]; 527| | *out++ = e1[(t1 & 0x03) << 4]; 528| | *out++ = '='; 529| | *out++ = '\n'; 530| | *out++ = '='; 531| | } 532| | } else { 533| | *out++ = e0[t1]; 534| | *out++ = e1[(t1 & 0x03) << 4]; 535| | *out++ = '='; 536| | *out++ = '='; 537| | } 538| | } else { 539| | if (line_offset + 2 >= line_length) { 540| | if (line_offset == line_length) { 541| | *out++ = '\n'; 542| | *out++ = e0[uint8_t(src[i])]; 543| | *out++ = e1[(uint8_t(src[i]) & 0x03) << 4]; 544| | } else if (line_offset + 1 == line_length) { 545| | *out++ = e0[uint8_t(src[i])]; 546| | *out++ = '\n'; 547| | *out++ = e1[(uint8_t(src[i]) & 0x03) << 4]; 548| | } else { 549| | *out++ = e0[uint8_t(src[i])]; 550| | *out++ = e1[(uint8_t(src[i]) & 0x03) << 4]; 551| | // *out++ = '\n'; ==> no newline at the end of the output 552| | } 553| | } else { 554| | *out++ = e0[uint8_t(src[i])]; 555| | *out++ = e1[(uint8_t(src[i]) & 0x03) << 4]; 556| | } 557| | } 558| 1.35k| } else { 559| 1.35k| *out++ = e0[t1]; 560| 1.35k| *out++ = e1[(t1 & 0x03) << 4]; 561| 1.35k| if (use_padding) { ------------------ | Branch (561:11): [True: 1.35k, False: 0] ------------------ 562| 1.35k| *out++ = '='; 563| 1.35k| *out++ = '='; 564| 1.35k| } 565| 1.35k| } 566| 1.35k| break; 567| 779| default: /* case 2 */ ------------------ | Branch (567:3): [True: 779, False: 46.3k] ------------------ 568| 779| t1 = uint8_t(src[i]); 569| 779| t2 = uint8_t(src[i + 1]); 570| | if constexpr (use_lines) { 571| | if (use_padding) { 572| | if (line_offset + 3 >= line_length) { 573| | if (line_offset == line_length) { 574| | *out++ = '\n'; 575| | *out++ = e0[t1]; 576| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 577| | *out++ = e2[(t2 & 0x0F) << 2]; 578| | *out++ = '='; 579| | } else if (line_offset + 1 == line_length) { 580| | *out++ = e0[t1]; 581| | *out++ = '\n'; 582| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 583| | *out++ = e2[(t2 & 0x0F) << 2]; 584| | *out++ = '='; 585| | } else if (line_offset + 2 == line_length) { 586| | *out++ = e0[t1]; 587| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 588| | *out++ = '\n'; 589| | *out++ = e2[(t2 & 0x0F) << 2]; 590| | *out++ = '='; 591| | } else if (line_offset + 3 == line_length) { 592| | *out++ = e0[t1]; 593| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 594| | *out++ = e2[(t2 & 0x0F) << 2]; 595| | *out++ = '\n'; 596| | *out++ = '='; 597| | } 598| | } else { 599| | *out++ = e0[t1]; 600| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 601| | *out++ = e2[(t2 & 0x0F) << 2]; 602| | *out++ = '='; 603| | } 604| | } else { 605| | if (line_offset + 3 >= line_length) { 606| | if (line_offset == line_length) { 607| | *out++ = '\n'; 608| | *out++ = e0[t1]; 609| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 610| | *out++ = e2[(t2 & 0x0F) << 2]; 611| | } else if (line_offset + 1 == line_length) { 612| | *out++ = e0[t1]; 613| | *out++ = '\n'; 614| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 615| | *out++ = e2[(t2 & 0x0F) << 2]; 616| | } else if (line_offset + 2 == line_length) { 617| | *out++ = e0[t1]; 618| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 619| | *out++ = '\n'; 620| | *out++ = e2[(t2 & 0x0F) << 2]; 621| | } else { 622| | *out++ = e0[t1]; 623| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 624| | *out++ = e2[(t2 & 0x0F) << 2]; 625| | // *out++ = '\n'; ==> no newline at the end of the output 626| | } 627| | } else { 628| | *out++ = e0[t1]; 629| | *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 630| | *out++ = e2[(t2 & 0x0F) << 2]; 631| | } 632| | } 633| 779| } else { 634| 779| *out++ = e0[t1]; 635| 779| *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 636| 779| *out++ = e2[(t2 & 0x0F) << 2]; 637| 779| if (use_padding) { ------------------ | Branch (637:11): [True: 779, False: 0] ------------------ 638| 779| *out++ = '='; 639| 779| } 640| 779| } 641| 47.1k| } 642| 47.1k| return (size_t)(out - dst); 643| 47.1k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6413is_eight_byteIcEEbT_: 19| 354k|template simdutf_constexpr23 bool is_eight_byte(char_type c) { 20| 354k| if constexpr (sizeof(char_type) == 1) { 21| 354k| return true; 22| 354k| } 23| 0| return uint8_t(c) == c; 24| 354k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6433maximal_binary_length_from_base64IPKcEEmT_m: 654| 80.5k|maximal_binary_length_from_base64(InputPtr input, size_t length) noexcept { 655| | // We process the padding characters ('=') at the end to make sure 656| | // that we return an exact result when the input has no ignorable characters 657| | // (e.g., spaces). 658| 80.5k| size_t padding = 0; 659| 80.5k| if (length > 0) { ------------------ | Branch (659:7): [True: 7.62k, False: 72.8k] ------------------ 660| 7.62k| if (input[length - 1] == '=') { ------------------ | Branch (660:9): [True: 2.31k, False: 5.30k] ------------------ 661| 2.31k| padding++; 662| 2.31k| if (length > 1 && input[length - 2] == '=') { ------------------ | Branch (662:11): [True: 2.28k, False: 38] | Branch (662:25): [True: 1.40k, False: 875] ------------------ 663| 1.40k| padding++; 664| 1.40k| } 665| 2.31k| } 666| 7.62k| } 667| | // The input is not otherwise processed for ignorable characters or 668| | // validation, so that the function runs in constant time (very fast). In 669| | // practice, base64 inputs without ignorable characters are common and the 670| | // common case are line separated inputs with relatively long lines (e.g., 76 671| | // characters) which leads this function to a slight (1%) overestimation of 672| | // the output size. 673| | // 674| | // Of course, some inputs might contain an arbitrary number of spaces or 675| | // newlines, which would make this function return a very pessimistic output 676| | // size but systems that produce base64 outputs typically do not do that and 677| | // if they do, they do not care much about minimizing memory usage. 678| | // 679| | // In specialized applications, users may know that their input is line 680| | // separated, which can be checked very quickly by by iterating (e.g., over 76 681| | // character chunks, looking for the linefeed characters only). We could 682| | // provide a specialized function for that, but it is not clear that the added 683| | // complexity is worth it for us. 684| | // 685| 80.5k| size_t actual_length = length - padding; 686| 80.5k| if (actual_length % 4 <= 1) { ------------------ | Branch (686:7): [True: 77.4k, False: 3.09k] ------------------ 687| 77.4k| return actual_length / 4 * 3; 688| 77.4k| } 689| | // if we have a valid input, then the remainder must be 2 or 3 adding one or 690| | // two extra bytes. 691| 3.09k| return actual_length / 4 * 3 + (actual_length % 4) - 1; 692| 80.5k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6412is_ignorableIcEEbT_NS_14base64_optionsE: 28| 3.48k| simdutf::base64_options options) { 29| 3.48k| const uint8_t *to_base64 = 30| 3.48k| (options & base64_default_or_url) ------------------ | Branch (30:7): [True: 0, False: 3.48k] ------------------ 31| 3.48k| ? tables::base64::to_base64_default_or_url_value 32| 3.48k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (32:14): [True: 0, False: 3.48k] ------------------ 33| 3.48k| : tables::base64::to_base64_value); 34| 3.48k| const bool ignore_garbage = 35| 3.48k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (35:7): [True: 0, False: 3.48k] ------------------ 36| 3.48k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (36:7): [True: 0, False: 3.48k] ------------------ 37| 3.48k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (37:7): [True: 0, False: 3.48k] ------------------ 38| 3.48k| uint8_t code = to_base64[uint8_t(c)]; 39| 3.48k| if (is_eight_byte(c) && code <= 63) { ------------------ | Branch (39:7): [True: 3.48k, False: 0] | Branch (39:27): [True: 460, False: 3.02k] ------------------ 40| 460| return false; 41| 460| } 42| 3.02k| if (is_eight_byte(c) && code == 64) { ------------------ | Branch (42:7): [True: 3.02k, False: 0] | Branch (42:27): [True: 3.02k, False: 0] ------------------ 43| 3.02k| return true; 44| 3.02k| } 45| 0| return ignore_garbage; 46| 3.02k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base648find_endIcEENS2_13reduced_inputEPKT_mNS_14base64_optionsE: 100| 99.5k| simdutf::base64_options options) { 101| 99.5k| const uint8_t *to_base64 = 102| 99.5k| (options & base64_default_or_url) ------------------ | Branch (102:7): [True: 0, False: 99.5k] ------------------ 103| 99.5k| ? tables::base64::to_base64_default_or_url_value 104| 99.5k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (104:14): [True: 0, False: 99.5k] ------------------ 105| 99.5k| : tables::base64::to_base64_value); 106| 99.5k| const bool ignore_garbage = 107| 99.5k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (107:7): [True: 0, False: 99.5k] ------------------ 108| 99.5k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (108:7): [True: 0, False: 99.5k] ------------------ 109| 99.5k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (109:7): [True: 0, False: 99.5k] ------------------ 110| | 111| 99.5k| size_t equalsigns = 0; 112| | // We intentionally include trailing spaces in the full input length. 113| | // See https://github.com/simdutf/simdutf/issues/824 114| 99.5k| size_t full_input_length = srclen; 115| | // skip trailing spaces 116| 108k| while (!ignore_garbage && srclen > 0 && ------------------ | Branch (116:10): [True: 108k, False: 0] | Branch (116:29): [True: 19.7k, False: 88.4k] ------------------ 117| 19.7k| scalar::base64::is_eight_byte(src[srclen - 1]) && ------------------ | Branch (117:10): [True: 19.7k, False: 0] ------------------ 118| 19.7k| to_base64[uint8_t(src[srclen - 1])] == 64) { ------------------ | Branch (118:10): [True: 8.59k, False: 11.1k] ------------------ 119| 8.59k| srclen--; 120| 8.59k| } 121| 99.5k| size_t equallocation = 122| 99.5k| srclen; // location of the first padding character if any 123| 99.5k| if (ignore_garbage) { ------------------ | Branch (123:7): [True: 0, False: 99.5k] ------------------ 124| | // Technically, we don't need to find the first padding character, we can 125| | // just change our algorithms, but it adds substantial complexity. 126| 0| auto it = simdutf::find(src, src + srclen, '='); 127| 0| if (it != src + srclen) { ------------------ | Branch (127:9): [True: 0, False: 0] ------------------ 128| 0| equallocation = it - src; 129| 0| equalsigns = 1; 130| 0| srclen = equallocation; 131| 0| full_input_length = equallocation + 1; 132| 0| } 133| 0| return {equalsigns, equallocation, srclen, full_input_length}; 134| 0| } 135| 99.5k| if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') { ------------------ | Branch (135:7): [True: 99.5k, False: 0] | Branch (135:26): [True: 11.1k, False: 88.4k] | Branch (135:40): [True: 3.73k, False: 7.37k] ------------------ 136| | // This is the last '=' sign. 137| 3.73k| equallocation = srclen - 1; 138| 3.73k| srclen--; 139| 3.73k| equalsigns = 1; 140| | // skip trailing spaces 141| 4.10k| while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) && ------------------ | Branch (141:12): [True: 4.04k, False: 61] | Branch (141:26): [True: 4.04k, False: 0] ------------------ 142| 4.04k| to_base64[uint8_t(src[srclen - 1])] == 64) { ------------------ | Branch (142:12): [True: 371, False: 3.66k] ------------------ 143| 371| srclen--; 144| 371| } 145| 3.73k| if (srclen > 0 && src[srclen - 1] == '=') { ------------------ | Branch (145:9): [True: 3.66k, False: 61] | Branch (145:23): [True: 2.34k, False: 1.32k] ------------------ 146| | // This is the second '=' sign. 147| 2.34k| equallocation = srclen - 1; 148| 2.34k| srclen--; 149| 2.34k| equalsigns = 2; 150| 2.34k| } 151| 3.73k| } 152| 99.5k| return {equalsigns, equallocation, srclen, full_input_length}; 153| 99.5k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423base64_tail_decode_safeIcEENS_11full_resultEPcmPKT_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 385| 2.32k| base64_options options, last_chunk_handling_options last_chunk_options) { 386| 2.32k| return base64_tail_decode_impl(dst, outlen, src, length, 387| 2.32k| padding_characters, options, 388| 2.32k| last_chunk_options); 389| 2.32k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423base64_tail_decode_implILb1EcEENS_11full_resultEPcmPKT0_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 164| 2.32k| base64_options options, last_chunk_handling_options last_chunk_options) { 165| 2.32k| char *dstend = dst + outlen; 166| 2.32k| (void)dstend; 167| | // This looks like 10 branches, but we expect the compiler to resolve this to 168| | // two branches (easily predicted): 169| 2.32k| const uint8_t *to_base64 = 170| 2.32k| (options & base64_default_or_url) ------------------ | Branch (170:7): [True: 0, False: 2.32k] ------------------ 171| 2.32k| ? tables::base64::to_base64_default_or_url_value 172| 2.32k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (172:14): [True: 0, False: 2.32k] ------------------ 173| 2.32k| : tables::base64::to_base64_value); 174| 2.32k| const uint32_t *d0 = 175| 2.32k| (options & base64_default_or_url) ------------------ | Branch (175:7): [True: 0, False: 2.32k] ------------------ 176| 2.32k| ? tables::base64::base64_default_or_url::d0 177| 2.32k| : ((options & base64_url) ? tables::base64::base64_url::d0 ------------------ | Branch (177:14): [True: 0, False: 2.32k] ------------------ 178| 2.32k| : tables::base64::base64_default::d0); 179| 2.32k| const uint32_t *d1 = 180| 2.32k| (options & base64_default_or_url) ------------------ | Branch (180:7): [True: 0, False: 2.32k] ------------------ 181| 2.32k| ? tables::base64::base64_default_or_url::d1 182| 2.32k| : ((options & base64_url) ? tables::base64::base64_url::d1 ------------------ | Branch (182:14): [True: 0, False: 2.32k] ------------------ 183| 2.32k| : tables::base64::base64_default::d1); 184| 2.32k| const uint32_t *d2 = 185| 2.32k| (options & base64_default_or_url) ------------------ | Branch (185:7): [True: 0, False: 2.32k] ------------------ 186| 2.32k| ? tables::base64::base64_default_or_url::d2 187| 2.32k| : ((options & base64_url) ? tables::base64::base64_url::d2 ------------------ | Branch (187:14): [True: 0, False: 2.32k] ------------------ 188| 2.32k| : tables::base64::base64_default::d2); 189| 2.32k| const uint32_t *d3 = 190| 2.32k| (options & base64_default_or_url) ------------------ | Branch (190:7): [True: 0, False: 2.32k] ------------------ 191| 2.32k| ? tables::base64::base64_default_or_url::d3 192| 2.32k| : ((options & base64_url) ? tables::base64::base64_url::d3 ------------------ | Branch (192:14): [True: 0, False: 2.32k] ------------------ 193| 2.32k| : tables::base64::base64_default::d3); 194| 2.32k| const bool ignore_garbage = 195| 2.32k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (195:7): [True: 0, False: 2.32k] ------------------ 196| 2.32k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (196:7): [True: 0, False: 2.32k] ------------------ 197| 2.32k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (197:7): [True: 0, False: 2.32k] ------------------ 198| | 199| 2.32k| const char_type *srcend = src + length; 200| 2.32k| const char_type *srcinit = src; 201| 2.32k| const char *dstinit = dst; 202| | 203| 2.32k| uint32_t x; 204| 2.32k| size_t idx; 205| 2.32k| uint8_t buffer[4]; 206| 2.36k| while (true) { ------------------ | Branch (206:10): [True: 2.36k, Folded] ------------------ 207| 2.42k| while (srcend - src >= 4 && is_eight_byte(src[0]) && ------------------ | Branch (207:12): [True: 249, False: 2.17k] | Branch (207:33): [True: 249, False: 0] ------------------ 208| 249| is_eight_byte(src[1]) && is_eight_byte(src[2]) && ------------------ | Branch (208:12): [True: 249, False: 0] | Branch (208:37): [True: 249, False: 0] ------------------ 209| 249| is_eight_byte(src[3]) && ------------------ | Branch (209:12): [True: 249, False: 0] ------------------ 210| 249| (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] | ------------------ | Branch (210:12): [True: 62, False: 187] ------------------ 211| 249| d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) { 212| 62| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (212:11): [True: 62, Folded] | Branch (212:29): [True: 0, False: 62] ------------------ 213| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(src - srcinit), 214| 0| size_t(dst - dstinit)}; 215| 0| } 216| 62| *dst++ = static_cast(x & 0xFF); 217| 62| *dst++ = static_cast((x >> 8) & 0xFF); 218| 62| *dst++ = static_cast((x >> 16) & 0xFF); 219| 62| src += 4; 220| 62| } 221| 2.36k| const char_type *srccur = src; 222| 2.36k| idx = 0; 223| | // we need at least four characters. 224| 2.36k|#ifdef __clang__ 225| | // If possible, we read four characters at a time. (It is an optimization.) 226| 2.36k| if (ignore_garbage && src + 4 <= srcend) { ------------------ | Branch (226:9): [True: 0, False: 2.36k] | Branch (226:27): [True: 0, False: 0] ------------------ 227| 0| char_type c0 = src[0]; 228| 0| char_type c1 = src[1]; 229| 0| char_type c2 = src[2]; 230| 0| char_type c3 = src[3]; 231| | 232| 0| uint8_t code0 = to_base64[uint8_t(c0)]; 233| 0| uint8_t code1 = to_base64[uint8_t(c1)]; 234| 0| uint8_t code2 = to_base64[uint8_t(c2)]; 235| 0| uint8_t code3 = to_base64[uint8_t(c3)]; 236| | 237| 0| buffer[idx] = code0; 238| 0| idx += (is_eight_byte(c0) && code0 <= 63); ------------------ | Branch (238:15): [True: 0, False: 0] | Branch (238:36): [True: 0, False: 0] ------------------ 239| 0| buffer[idx] = code1; 240| 0| idx += (is_eight_byte(c1) && code1 <= 63); ------------------ | Branch (240:15): [True: 0, False: 0] | Branch (240:36): [True: 0, False: 0] ------------------ 241| 0| buffer[idx] = code2; 242| 0| idx += (is_eight_byte(c2) && code2 <= 63); ------------------ | Branch (242:15): [True: 0, False: 0] | Branch (242:36): [True: 0, False: 0] ------------------ 243| 0| buffer[idx] = code3; 244| 0| idx += (is_eight_byte(c3) && code3 <= 63); ------------------ | Branch (244:15): [True: 0, False: 0] | Branch (244:36): [True: 0, False: 0] ------------------ 245| 0| src += 4; 246| 0| } 247| 2.36k|#endif 248| 13.6k| while ((idx < 4) && (src < srcend)) { ------------------ | Branch (248:12): [True: 13.5k, False: 41] | Branch (248:25): [True: 11.4k, False: 2.13k] ------------------ 249| 11.4k| char_type c = *src; 250| | 251| 11.4k| uint8_t code = to_base64[uint8_t(c)]; 252| 11.4k| buffer[idx] = uint8_t(code); 253| 11.4k| if (is_eight_byte(c) && code <= 63) { ------------------ | Branch (253:11): [True: 11.4k, False: 0] | Branch (253:31): [True: 4.93k, False: 6.51k] ------------------ 254| 4.93k| idx++; 255| 6.51k| } else if (!ignore_garbage && ------------------ | Branch (255:18): [True: 6.51k, False: 0] ------------------ 256| 6.51k| (code > 64 || !scalar::base64::is_eight_byte(c))) { ------------------ | Branch (256:19): [True: 191, False: 6.32k] | Branch (256:32): [True: 0, False: 6.32k] ------------------ 257| 191| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 258| 191| size_t(dst - dstinit)}; 259| 6.32k| } else { 260| | // We have a space or a newline or garbage. We ignore it. 261| 6.32k| } 262| 11.2k| src++; 263| 11.2k| } 264| 2.17k| if (idx != 4) { ------------------ | Branch (264:9): [True: 2.13k, False: 41] ------------------ 265| 2.13k| simdutf_log_assert(idx < 4, "idx should be less than 4"); 266| | // We never should have that the number of base64 characters + the 267| | // number of padding characters is more than 4. 268| 2.13k| if (!ignore_garbage && (idx + padding_characters > 4)) { ------------------ | Branch (268:11): [True: 2.13k, False: 0] | Branch (268:30): [True: 7, False: 2.12k] ------------------ 269| 7| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 270| 7| size_t(dst - dstinit), true}; 271| 7| } 272| | 273| | // The idea here is that in loose mode, 274| | // if there is padding at all, it must be used 275| | // to form 4-wise chunk. However, in loose mode, 276| | // we do accept no padding at all. 277| 2.12k| if (!ignore_garbage && ------------------ | Branch (277:11): [True: 2.12k, False: 0] ------------------ 278| 2.12k| last_chunk_options == last_chunk_handling_options::loose && ------------------ | Branch (278:11): [True: 2.12k, False: 0] ------------------ 279| 2.12k| (idx >= 2) && padding_characters > 0 && ------------------ | Branch (279:11): [True: 1.94k, False: 177] | Branch (279:25): [True: 1.76k, False: 184] ------------------ 280| 1.76k| ((idx + padding_characters) & 3) != 0) { ------------------ | Branch (280:11): [True: 11, False: 1.75k] ------------------ 281| 11| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 282| 11| size_t(dst - dstinit), true}; 283| 11| } else 284| | 285| | // The idea here is that in strict mode, we do not want to accept 286| | // incomplete base64 chunks. So if the chunk was otherwise valid, we 287| | // return BASE64_INPUT_REMAINDER. 288| 2.11k| if (!ignore_garbage && ------------------ | Branch (288:13): [True: 2.11k, False: 0] ------------------ 289| 2.11k| last_chunk_options == last_chunk_handling_options::strict && ------------------ | Branch (289:13): [True: 0, False: 2.11k] ------------------ 290| 0| (idx >= 2) && ((idx + padding_characters) & 3) != 0) { ------------------ | Branch (290:13): [True: 0, False: 0] | Branch (290:27): [True: 0, False: 0] ------------------ 291| | // The partial chunk was at src - idx 292| 0| return {BASE64_INPUT_REMAINDER, size_t(src - srcinit), 293| 0| size_t(dst - dstinit), true}; 294| 0| } else 295| | // If there is a partial chunk with insufficient padding, with 296| | // stop_before_partial, we need to just ignore it. In "only full" 297| | // mode, skip the minute there are padding characters. 298| 2.11k| if ((last_chunk_options == ------------------ | Branch (298:16): [True: 0, False: 2.11k] ------------------ 299| 2.11k| last_chunk_handling_options::stop_before_partial && 300| 0| (padding_characters + idx < 4) && (idx != 0) && ------------------ | Branch (300:16): [True: 0, False: 0] | Branch (300:50): [True: 0, False: 0] ------------------ 301| 0| (idx >= 2 || padding_characters == 0)) || ------------------ | Branch (301:17): [True: 0, False: 0] | Branch (301:29): [True: 0, False: 0] ------------------ 302| 2.11k| (last_chunk_options == ------------------ | Branch (302:16): [True: 0, False: 2.11k] ------------------ 303| 2.11k| last_chunk_handling_options::only_full_chunks && 304| 0| (idx >= 2 || padding_characters == 0))) { ------------------ | Branch (304:17): [True: 0, False: 0] | Branch (304:29): [True: 0, False: 0] ------------------ 305| | // partial means that we are *not* going to consume the read 306| | // characters. We need to rewind the src pointer. 307| 0| src = srccur; 308| 0| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 309| 2.11k| } else { 310| 2.11k| if (idx == 2) { ------------------ | Branch (310:17): [True: 1.27k, False: 841] ------------------ 311| 1.27k| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 312| 1.27k| (uint32_t(buffer[1]) << 2 * 6); 313| 1.27k| if (!ignore_garbage && ------------------ | Branch (313:19): [True: 1.27k, False: 0] ------------------ 314| 1.27k| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (314:19): [True: 0, False: 1.27k] ------------------ 315| 0| (triple & 0xffff)) { ------------------ | Branch (315:19): [True: 0, False: 0] ------------------ 316| 0| return {BASE64_EXTRA_BITS, size_t(src - srcinit), 317| 0| size_t(dst - dstinit)}; 318| 0| } 319| 1.27k| if (check_capacity && dstend - dst < 1) { ------------------ | Branch (319:19): [True: 1.27k, Folded] | Branch (319:37): [True: 0, False: 1.27k] ------------------ 320| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 321| 0| size_t(dst - dstinit)}; 322| 0| } 323| 1.27k| *dst++ = static_cast((triple >> 16) & 0xFF); 324| 1.27k| } else if (idx == 3) { ------------------ | Branch (324:24): [True: 664, False: 177] ------------------ 325| 664| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 326| 664| (uint32_t(buffer[1]) << 2 * 6) + 327| 664| (uint32_t(buffer[2]) << 1 * 6); 328| 664| if (!ignore_garbage && ------------------ | Branch (328:19): [True: 664, False: 0] ------------------ 329| 664| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (329:19): [True: 0, False: 664] ------------------ 330| 0| (triple & 0xff)) { ------------------ | Branch (330:19): [True: 0, False: 0] ------------------ 331| 0| return {BASE64_EXTRA_BITS, size_t(src - srcinit), 332| 0| size_t(dst - dstinit)}; 333| 0| } 334| 664| if (check_capacity && dstend - dst < 2) { ------------------ | Branch (334:19): [True: 664, Folded] | Branch (334:37): [True: 0, False: 664] ------------------ 335| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 336| 0| size_t(dst - dstinit)}; 337| 0| } 338| 664| *dst++ = static_cast((triple >> 16) & 0xFF); 339| 664| *dst++ = static_cast((triple >> 8) & 0xFF); 340| 664| } else if (!ignore_garbage && idx == 1 && ------------------ | Branch (340:24): [True: 177, False: 0] | Branch (340:43): [True: 82, False: 95] ------------------ 341| 82| (!is_partial(last_chunk_options) || ------------------ | Branch (341:25): [True: 82, False: 0] ------------------ 342| 0| (is_partial(last_chunk_options) && ------------------ | Branch (342:26): [True: 0, False: 0] ------------------ 343| 82| padding_characters > 0))) { ------------------ | Branch (343:26): [True: 0, False: 0] ------------------ 344| 82| return {BASE64_INPUT_REMAINDER, size_t(src - srcinit), 345| 82| size_t(dst - dstinit)}; 346| 95| } else if (!ignore_garbage && idx == 0 && padding_characters > 0) { ------------------ | Branch (346:24): [True: 95, False: 0] | Branch (346:43): [True: 95, False: 0] | Branch (346:55): [True: 7, False: 88] ------------------ 347| 7| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 348| 7| size_t(dst - dstinit), true}; 349| 7| } 350| 2.02k| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 351| 2.11k| } 352| 2.12k| } 353| 41| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (353:9): [True: 41, Folded] | Branch (353:27): [True: 0, False: 41] ------------------ 354| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 355| 0| size_t(dst - dstinit)}; 356| 0| } 357| 41| uint32_t triple = 358| 41| (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6) + 359| 41| (uint32_t(buffer[2]) << 1 * 6) + (uint32_t(buffer[3]) << 0 * 6); 360| 41| *dst++ = static_cast((triple >> 16) & 0xFF); 361| 41| *dst++ = static_cast((triple >> 8) & 0xFF); 362| 41| *dst++ = static_cast(triple & 0xFF); 363| 41| } 364| 2.32k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6417patch_tail_resultENS_11full_resultEmmmmNS_27last_chunk_handling_optionsE: 394| 9.36k| last_chunk_handling_options last_chunk_options) { 395| 9.36k| r.input_count += previous_input; 396| 9.36k| r.output_count += previous_output; 397| 9.36k| if (r.padding_error) { ------------------ | Branch (397:7): [True: 90, False: 9.27k] ------------------ 398| 90| r.input_count = equallocation; 399| 90| } 400| | 401| 9.36k| if (r.error == error_code::SUCCESS) { ------------------ | Branch (401:7): [True: 7.65k, False: 1.71k] ------------------ 402| 7.65k| if (!is_partial(last_chunk_options)) { ------------------ | Branch (402:9): [True: 5.97k, False: 1.67k] ------------------ 403| | // A success when we are not in stop_before_partial mode. 404| | // means that we have consumed the whole input buffer. 405| 5.97k| r.input_count = full_input_length; 406| 5.97k| } else if (r.output_count % 3 != 0) { ------------------ | Branch (406:16): [True: 0, False: 1.67k] ------------------ 407| 0| r.input_count = full_input_length; 408| 0| } 409| 7.65k| } 410| 9.36k| return r; 411| 9.36k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6434base64_to_binary_details_safe_implIcEENS_11full_resultEPKT_mPcmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 783| 2.37k| last_chunk_handling_options last_chunk_options) noexcept { 784| 2.37k| const bool ignore_garbage = 785| 2.37k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (785:7): [True: 0, False: 2.37k] ------------------ 786| 2.37k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (786:7): [True: 0, False: 2.37k] ------------------ 787| 2.37k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (787:7): [True: 0, False: 2.37k] ------------------ 788| 2.37k| auto ri = simdutf::scalar::base64::find_end(input, length, options); 789| 2.37k| size_t equallocation = ri.equallocation; 790| 2.37k| size_t equalsigns = ri.equalsigns; 791| 2.37k| length = ri.srclen; 792| 2.37k| size_t full_input_length = ri.full_input_length; 793| 2.37k| if (length == 0) { ------------------ | Branch (793:7): [True: 56, False: 2.32k] ------------------ 794| 56| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (794:9): [True: 56, False: 0] | Branch (794:28): [True: 26, False: 30] ------------------ 795| 26| return {INVALID_BASE64_CHARACTER, equallocation, 0}; 796| 26| } 797| 30| return {SUCCESS, full_input_length, 0}; 798| 56| } 799| 2.32k| full_result r = scalar::base64::base64_tail_decode_safe( 800| 2.32k| output, outlen, input, length, equalsigns, options, last_chunk_options); 801| 2.32k| r = scalar::base64::patch_tail_result(r, 0, 0, equallocation, 802| 2.32k| full_input_length, last_chunk_options); 803| 2.32k| if (!is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (803:7): [True: 2.32k, False: 0] | Branch (803:42): [True: 2.02k, False: 298] ------------------ 804| 2.02k| equalsigns > 0 && !ignore_garbage) { ------------------ | Branch (804:7): [True: 1.75k, False: 272] | Branch (804:25): [True: 1.75k, False: 0] ------------------ 805| | // additional checks 806| 1.75k| if ((r.output_count % 3 == 0) || ------------------ | Branch (806:9): [True: 0, False: 1.75k] ------------------ 807| 1.75k| ((r.output_count % 3) + 1 + equalsigns != 4)) { ------------------ | Branch (807:9): [True: 0, False: 1.75k] ------------------ 808| 0| return {INVALID_BASE64_CHARACTER, equallocation, r.output_count}; 809| 0| } 810| 1.75k| } 811| | 812| | // When is_partial(last_chunk_options) is true, we must either end with 813| | // the end of the stream (beyond whitespace) or right after a non-ignorable 814| | // character or at the very beginning of the stream. 815| | // See https://tc39.es/proposal-arraybuffer-base64/spec/#sec-frombase64 816| 2.32k| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (816:7): [True: 0, False: 2.32k] | Branch (816:41): [True: 0, False: 0] ------------------ 817| 0| r.input_count < full_input_length) { ------------------ | Branch (817:7): [True: 0, False: 0] ------------------ 818| | // First check if we can extend the input to the end of the stream 819| 0| while (r.input_count < full_input_length && ------------------ | Branch (819:12): [True: 0, False: 0] ------------------ 820| 0| base64_ignorable(*(input + r.input_count), options)) { ------------------ | Branch (820:12): [True: 0, False: 0] ------------------ 821| 0| r.input_count++; 822| 0| } 823| | // If we are still not at the end of the stream, then we must backtrack 824| | // to the last non-ignorable character. 825| 0| if (r.input_count < full_input_length) { ------------------ | Branch (825:9): [True: 0, False: 0] ------------------ 826| 0| while (r.input_count > 0 && ------------------ | Branch (826:14): [True: 0, False: 0] ------------------ 827| 0| base64_ignorable(*(input + r.input_count - 1), options)) { ------------------ | Branch (827:14): [True: 0, False: 0] ------------------ 828| 0| r.input_count--; 829| 0| } 830| 0| } 831| 0| } 832| 2.32k| return r; 833| 2.32k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6429base64_to_binary_details_implIcEENS_11full_resultEPKT_mPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 727| 16.4k| last_chunk_handling_options last_chunk_options) noexcept { 728| 16.4k| const bool ignore_garbage = 729| 16.4k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (729:7): [True: 0, False: 16.4k] ------------------ 730| 16.4k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (730:7): [True: 0, False: 16.4k] ------------------ 731| 16.4k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (731:7): [True: 0, False: 16.4k] ------------------ 732| 16.4k| auto ri = simdutf::scalar::base64::find_end(input, length, options); 733| 16.4k| size_t equallocation = ri.equallocation; 734| 16.4k| size_t equalsigns = ri.equalsigns; 735| 16.4k| length = ri.srclen; 736| 16.4k| size_t full_input_length = ri.full_input_length; 737| 16.4k| if (length == 0) { ------------------ | Branch (737:7): [True: 16.0k, False: 415] ------------------ 738| 16.0k| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (738:9): [True: 16.0k, False: 0] | Branch (738:28): [True: 6, False: 16.0k] ------------------ 739| 6| return {INVALID_BASE64_CHARACTER, equallocation, 0, true}; 740| 6| } 741| 16.0k| return {SUCCESS, full_input_length, 0}; 742| 16.0k| } 743| 415| full_result r = scalar::base64::base64_tail_decode( 744| 415| output, input, length, equalsigns, options, last_chunk_options); 745| 415| r = scalar::base64::patch_tail_result(r, 0, 0, equallocation, 746| 415| full_input_length, last_chunk_options); 747| 415| if (!is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (747:7): [True: 415, False: 0] | Branch (747:42): [True: 233, False: 182] ------------------ 748| 233| equalsigns > 0 && !ignore_garbage) { ------------------ | Branch (748:7): [True: 122, False: 111] | Branch (748:25): [True: 122, False: 0] ------------------ 749| | // additional checks 750| 122| if ((r.output_count % 3 == 0) || ------------------ | Branch (750:9): [True: 0, False: 122] ------------------ 751| 122| ((r.output_count % 3) + 1 + equalsigns != 4)) { ------------------ | Branch (751:9): [True: 0, False: 122] ------------------ 752| 0| return {INVALID_BASE64_CHARACTER, equallocation, r.output_count, true}; 753| 0| } 754| 122| } 755| | // When is_partial(last_chunk_options) is true, we must either end with 756| | // the end of the stream (beyond whitespace) or right after a non-ignorable 757| | // character or at the very beginning of the stream. 758| | // See https://tc39.es/proposal-arraybuffer-base64/spec/#sec-frombase64 759| 415| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (759:7): [True: 0, False: 415] | Branch (759:41): [True: 0, False: 0] ------------------ 760| 0| r.input_count < full_input_length) { ------------------ | Branch (760:7): [True: 0, False: 0] ------------------ 761| | // First check if we can extend the input to the end of the stream 762| 0| while (r.input_count < full_input_length && ------------------ | Branch (762:12): [True: 0, False: 0] ------------------ 763| 0| base64_ignorable(*(input + r.input_count), options)) { ------------------ | Branch (763:12): [True: 0, False: 0] ------------------ 764| 0| r.input_count++; 765| 0| } 766| | // If we are still not at the end of the stream, then we must backtrack 767| | // to the last non-ignorable character. 768| 0| if (r.input_count < full_input_length) { ------------------ | Branch (768:9): [True: 0, False: 0] ------------------ 769| 0| while (r.input_count > 0 && ------------------ | Branch (769:14): [True: 0, False: 0] ------------------ 770| 0| base64_ignorable(*(input + r.input_count - 1), options)) { ------------------ | Branch (770:14): [True: 0, False: 0] ------------------ 771| 0| r.input_count--; 772| 0| } 773| 0| } 774| 0| } 775| 415| return r; 776| 415|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6418base64_tail_decodeIcEENS_11full_resultEPcPKT_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 371| 7.04k| base64_options options, last_chunk_handling_options last_chunk_options) { 372| 7.04k| return base64_tail_decode_impl(dst, 0, src, length, padding_characters, 373| 7.04k| options, last_chunk_options); 374| 7.04k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423base64_tail_decode_implILb0EcEENS_11full_resultEPcmPKT0_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 164| 7.04k| base64_options options, last_chunk_handling_options last_chunk_options) { 165| 7.04k| char *dstend = dst + outlen; 166| 7.04k| (void)dstend; 167| | // This looks like 10 branches, but we expect the compiler to resolve this to 168| | // two branches (easily predicted): 169| 7.04k| const uint8_t *to_base64 = 170| 7.04k| (options & base64_default_or_url) ------------------ | Branch (170:7): [True: 0, False: 7.04k] ------------------ 171| 7.04k| ? tables::base64::to_base64_default_or_url_value 172| 7.04k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (172:14): [True: 0, False: 7.04k] ------------------ 173| 7.04k| : tables::base64::to_base64_value); 174| 7.04k| const uint32_t *d0 = 175| 7.04k| (options & base64_default_or_url) ------------------ | Branch (175:7): [True: 0, False: 7.04k] ------------------ 176| 7.04k| ? tables::base64::base64_default_or_url::d0 177| 7.04k| : ((options & base64_url) ? tables::base64::base64_url::d0 ------------------ | Branch (177:14): [True: 0, False: 7.04k] ------------------ 178| 7.04k| : tables::base64::base64_default::d0); 179| 7.04k| const uint32_t *d1 = 180| 7.04k| (options & base64_default_or_url) ------------------ | Branch (180:7): [True: 0, False: 7.04k] ------------------ 181| 7.04k| ? tables::base64::base64_default_or_url::d1 182| 7.04k| : ((options & base64_url) ? tables::base64::base64_url::d1 ------------------ | Branch (182:14): [True: 0, False: 7.04k] ------------------ 183| 7.04k| : tables::base64::base64_default::d1); 184| 7.04k| const uint32_t *d2 = 185| 7.04k| (options & base64_default_or_url) ------------------ | Branch (185:7): [True: 0, False: 7.04k] ------------------ 186| 7.04k| ? tables::base64::base64_default_or_url::d2 187| 7.04k| : ((options & base64_url) ? tables::base64::base64_url::d2 ------------------ | Branch (187:14): [True: 0, False: 7.04k] ------------------ 188| 7.04k| : tables::base64::base64_default::d2); 189| 7.04k| const uint32_t *d3 = 190| 7.04k| (options & base64_default_or_url) ------------------ | Branch (190:7): [True: 0, False: 7.04k] ------------------ 191| 7.04k| ? tables::base64::base64_default_or_url::d3 192| 7.04k| : ((options & base64_url) ? tables::base64::base64_url::d3 ------------------ | Branch (192:14): [True: 0, False: 7.04k] ------------------ 193| 7.04k| : tables::base64::base64_default::d3); 194| 7.04k| const bool ignore_garbage = 195| 7.04k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (195:7): [True: 0, False: 7.04k] ------------------ 196| 7.04k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (196:7): [True: 0, False: 7.04k] ------------------ 197| 7.04k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (197:7): [True: 0, False: 7.04k] ------------------ 198| | 199| 7.04k| const char_type *srcend = src + length; 200| 7.04k| const char_type *srcinit = src; 201| 7.04k| const char *dstinit = dst; 202| | 203| 7.04k| uint32_t x; 204| 7.04k| size_t idx; 205| 7.04k| uint8_t buffer[4]; 206| 7.82k| while (true) { ------------------ | Branch (206:10): [True: 7.82k, Folded] ------------------ 207| 71.1k| while (srcend - src >= 4 && is_eight_byte(src[0]) && ------------------ | Branch (207:12): [True: 65.4k, False: 5.71k] | Branch (207:33): [True: 65.4k, False: 0] ------------------ 208| 65.4k| is_eight_byte(src[1]) && is_eight_byte(src[2]) && ------------------ | Branch (208:12): [True: 65.4k, False: 0] | Branch (208:37): [True: 65.4k, False: 0] ------------------ 209| 65.4k| is_eight_byte(src[3]) && ------------------ | Branch (209:12): [True: 65.4k, False: 0] ------------------ 210| 65.4k| (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] | ------------------ | Branch (210:12): [True: 63.3k, False: 2.11k] ------------------ 211| 65.4k| d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) { 212| 63.3k| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (212:11): [Folded, False: 63.3k] | Branch (212:29): [True: 0, False: 0] ------------------ 213| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(src - srcinit), 214| 0| size_t(dst - dstinit)}; 215| 0| } 216| 63.3k| *dst++ = static_cast(x & 0xFF); 217| 63.3k| *dst++ = static_cast((x >> 8) & 0xFF); 218| 63.3k| *dst++ = static_cast((x >> 16) & 0xFF); 219| 63.3k| src += 4; 220| 63.3k| } 221| 7.82k| const char_type *srccur = src; 222| 7.82k| idx = 0; 223| | // we need at least four characters. 224| 7.82k|#ifdef __clang__ 225| | // If possible, we read four characters at a time. (It is an optimization.) 226| 7.82k| if (ignore_garbage && src + 4 <= srcend) { ------------------ | Branch (226:9): [True: 0, False: 7.82k] | Branch (226:27): [True: 0, False: 0] ------------------ 227| 0| char_type c0 = src[0]; 228| 0| char_type c1 = src[1]; 229| 0| char_type c2 = src[2]; 230| 0| char_type c3 = src[3]; 231| | 232| 0| uint8_t code0 = to_base64[uint8_t(c0)]; 233| 0| uint8_t code1 = to_base64[uint8_t(c1)]; 234| 0| uint8_t code2 = to_base64[uint8_t(c2)]; 235| 0| uint8_t code3 = to_base64[uint8_t(c3)]; 236| | 237| 0| buffer[idx] = code0; 238| 0| idx += (is_eight_byte(c0) && code0 <= 63); ------------------ | Branch (238:15): [True: 0, False: 0] | Branch (238:36): [True: 0, False: 0] ------------------ 239| 0| buffer[idx] = code1; 240| 0| idx += (is_eight_byte(c1) && code1 <= 63); ------------------ | Branch (240:15): [True: 0, False: 0] | Branch (240:36): [True: 0, False: 0] ------------------ 241| 0| buffer[idx] = code2; 242| 0| idx += (is_eight_byte(c2) && code2 <= 63); ------------------ | Branch (242:15): [True: 0, False: 0] | Branch (242:36): [True: 0, False: 0] ------------------ 243| 0| buffer[idx] = code3; 244| 0| idx += (is_eight_byte(c3) && code3 <= 63); ------------------ | Branch (244:15): [True: 0, False: 0] | Branch (244:36): [True: 0, False: 0] ------------------ 245| 0| src += 4; 246| 0| } 247| 7.82k|#endif 248| 33.1k| while ((idx < 4) && (src < srcend)) { ------------------ | Branch (248:12): [True: 32.4k, False: 786] | Branch (248:25): [True: 26.4k, False: 5.91k] ------------------ 249| 26.4k| char_type c = *src; 250| | 251| 26.4k| uint8_t code = to_base64[uint8_t(c)]; 252| 26.4k| buffer[idx] = uint8_t(code); 253| 26.4k| if (is_eight_byte(c) && code <= 63) { ------------------ | Branch (253:11): [True: 26.4k, False: 0] | Branch (253:31): [True: 9.98k, False: 16.4k] ------------------ 254| 9.98k| idx++; 255| 16.4k| } else if (!ignore_garbage && ------------------ | Branch (255:18): [True: 16.4k, False: 0] ------------------ 256| 16.4k| (code > 64 || !scalar::base64::is_eight_byte(c))) { ------------------ | Branch (256:19): [True: 1.12k, False: 15.3k] | Branch (256:32): [True: 0, False: 15.3k] ------------------ 257| 1.12k| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 258| 1.12k| size_t(dst - dstinit)}; 259| 15.3k| } else { 260| | // We have a space or a newline or garbage. We ignore it. 261| 15.3k| } 262| 25.3k| src++; 263| 25.3k| } 264| 6.70k| if (idx != 4) { ------------------ | Branch (264:9): [True: 5.91k, False: 786] ------------------ 265| 5.91k| simdutf_log_assert(idx < 4, "idx should be less than 4"); 266| | // We never should have that the number of base64 characters + the 267| | // number of padding characters is more than 4. 268| 5.91k| if (!ignore_garbage && (idx + padding_characters > 4)) { ------------------ | Branch (268:11): [True: 5.91k, False: 0] | Branch (268:30): [True: 15, False: 5.90k] ------------------ 269| 15| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 270| 15| size_t(dst - dstinit), true}; 271| 15| } 272| | 273| | // The idea here is that in loose mode, 274| | // if there is padding at all, it must be used 275| | // to form 4-wise chunk. However, in loose mode, 276| | // we do accept no padding at all. 277| 5.90k| if (!ignore_garbage && ------------------ | Branch (277:11): [True: 5.90k, False: 0] ------------------ 278| 5.90k| last_chunk_options == last_chunk_handling_options::loose && ------------------ | Branch (278:11): [True: 4.21k, False: 1.68k] ------------------ 279| 4.21k| (idx >= 2) && padding_characters > 0 && ------------------ | Branch (279:11): [True: 2.36k, False: 1.85k] | Branch (279:25): [True: 1.78k, False: 582] ------------------ 280| 1.78k| ((idx + padding_characters) & 3) != 0) { ------------------ | Branch (280:11): [True: 29, False: 1.75k] ------------------ 281| 29| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 282| 29| size_t(dst - dstinit), true}; 283| 29| } else 284| | 285| | // The idea here is that in strict mode, we do not want to accept 286| | // incomplete base64 chunks. So if the chunk was otherwise valid, we 287| | // return BASE64_INPUT_REMAINDER. 288| 5.87k| if (!ignore_garbage && ------------------ | Branch (288:13): [True: 5.87k, False: 0] ------------------ 289| 5.87k| last_chunk_options == last_chunk_handling_options::strict && ------------------ | Branch (289:13): [True: 0, False: 5.87k] ------------------ 290| 0| (idx >= 2) && ((idx + padding_characters) & 3) != 0) { ------------------ | Branch (290:13): [True: 0, False: 0] | Branch (290:27): [True: 0, False: 0] ------------------ 291| | // The partial chunk was at src - idx 292| 0| return {BASE64_INPUT_REMAINDER, size_t(src - srcinit), 293| 0| size_t(dst - dstinit), true}; 294| 0| } else 295| | // If there is a partial chunk with insufficient padding, with 296| | // stop_before_partial, we need to just ignore it. In "only full" 297| | // mode, skip the minute there are padding characters. 298| 5.87k| if ((last_chunk_options == ------------------ | Branch (298:16): [True: 0, False: 5.87k] ------------------ 299| 5.87k| last_chunk_handling_options::stop_before_partial && 300| 0| (padding_characters + idx < 4) && (idx != 0) && ------------------ | Branch (300:16): [True: 0, False: 0] | Branch (300:50): [True: 0, False: 0] ------------------ 301| 0| (idx >= 2 || padding_characters == 0)) || ------------------ | Branch (301:17): [True: 0, False: 0] | Branch (301:29): [True: 0, False: 0] ------------------ 302| 5.87k| (last_chunk_options == ------------------ | Branch (302:16): [True: 1.68k, False: 4.18k] ------------------ 303| 5.87k| last_chunk_handling_options::only_full_chunks && 304| 1.68k| (idx >= 2 || padding_characters == 0))) { ------------------ | Branch (304:17): [True: 187, False: 1.49k] | Branch (304:29): [True: 1.49k, False: 8] ------------------ 305| | // partial means that we are *not* going to consume the read 306| | // characters. We need to rewind the src pointer. 307| 1.67k| src = srccur; 308| 1.67k| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 309| 4.19k| } else { 310| 4.19k| if (idx == 2) { ------------------ | Branch (310:17): [True: 1.44k, False: 2.75k] ------------------ 311| 1.44k| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 312| 1.44k| (uint32_t(buffer[1]) << 2 * 6); 313| 1.44k| if (!ignore_garbage && ------------------ | Branch (313:19): [True: 1.44k, False: 0] ------------------ 314| 1.44k| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (314:19): [True: 0, False: 1.44k] ------------------ 315| 0| (triple & 0xffff)) { ------------------ | Branch (315:19): [True: 0, False: 0] ------------------ 316| 0| return {BASE64_EXTRA_BITS, size_t(src - srcinit), 317| 0| size_t(dst - dstinit)}; 318| 0| } 319| 1.44k| if (check_capacity && dstend - dst < 1) { ------------------ | Branch (319:19): [Folded, False: 1.44k] | Branch (319:37): [True: 0, False: 0] ------------------ 320| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 321| 0| size_t(dst - dstinit)}; 322| 0| } 323| 1.44k| *dst++ = static_cast((triple >> 16) & 0xFF); 324| 2.75k| } else if (idx == 3) { ------------------ | Branch (324:24): [True: 894, False: 1.85k] ------------------ 325| 894| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 326| 894| (uint32_t(buffer[1]) << 2 * 6) + 327| 894| (uint32_t(buffer[2]) << 1 * 6); 328| 894| if (!ignore_garbage && ------------------ | Branch (328:19): [True: 894, False: 0] ------------------ 329| 894| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (329:19): [True: 0, False: 894] ------------------ 330| 0| (triple & 0xff)) { ------------------ | Branch (330:19): [True: 0, False: 0] ------------------ 331| 0| return {BASE64_EXTRA_BITS, size_t(src - srcinit), 332| 0| size_t(dst - dstinit)}; 333| 0| } 334| 894| if (check_capacity && dstend - dst < 2) { ------------------ | Branch (334:19): [Folded, False: 894] | Branch (334:37): [True: 0, False: 0] ------------------ 335| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 336| 0| size_t(dst - dstinit)}; 337| 0| } 338| 894| *dst++ = static_cast((triple >> 16) & 0xFF); 339| 894| *dst++ = static_cast((triple >> 8) & 0xFF); 340| 1.85k| } else if (!ignore_garbage && idx == 1 && ------------------ | Branch (340:24): [True: 1.85k, False: 0] | Branch (340:43): [True: 227, False: 1.63k] ------------------ 341| 227| (!is_partial(last_chunk_options) || ------------------ | Branch (341:25): [True: 219, False: 8] ------------------ 342| 8| (is_partial(last_chunk_options) && ------------------ | Branch (342:26): [True: 8, False: 0] ------------------ 343| 227| padding_characters > 0))) { ------------------ | Branch (343:26): [True: 8, False: 0] ------------------ 344| 227| return {BASE64_INPUT_REMAINDER, size_t(src - srcinit), 345| 227| size_t(dst - dstinit)}; 346| 1.63k| } else if (!ignore_garbage && idx == 0 && padding_characters > 0) { ------------------ | Branch (346:24): [True: 1.63k, False: 0] | Branch (346:43): [True: 1.63k, False: 0] | Branch (346:55): [True: 21, False: 1.61k] ------------------ 347| 21| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 348| 21| size_t(dst - dstinit), true}; 349| 21| } 350| 3.94k| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 351| 4.19k| } 352| 5.90k| } 353| 786| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (353:9): [Folded, False: 786] | Branch (353:27): [True: 0, False: 0] ------------------ 354| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 355| 0| size_t(dst - dstinit)}; 356| 0| } 357| 786| uint32_t triple = 358| 786| (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6) + 359| 786| (uint32_t(buffer[2]) << 1 * 6) + (uint32_t(buffer[3]) << 0 * 6); 360| 786| *dst++ = static_cast((triple >> 16) & 0xFF); 361| 786| *dst++ = static_cast((triple >> 8) & 0xFF); 362| 786| *dst++ = static_cast(triple & 0xFF); 363| 786| } 364| 7.04k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16latin123utf8_length_from_latin1EPKcm: 10| 11.6k| size_t len) { 11| 11.6k| const uint8_t *c = reinterpret_cast(buf); 12| 11.6k| size_t answer = 0; 13| 26.3k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (13:22): [True: 14.6k, False: 11.6k] ------------------ 14| 14.6k| if ((c[i] >> 7)) { ------------------ | Branch (14:9): [True: 3.63k, False: 10.9k] ------------------ 15| 3.63k| answer++; 16| 3.63k| } 17| 14.6k| } 18| 11.6k| return answer + len; 19| 11.6k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115latin1_to_utf167convertILNS_10endiannessE0EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 14| 6.82k| char16_t *utf16_output) { 15| 6.82k| size_t pos = 0; 16| 6.82k| char16_t *start{utf16_output}; 17| | 18| 106k| while (pos < len) { ------------------ | Branch (18:10): [True: 99.6k, False: 6.82k] ------------------ 19| 99.6k| uint16_t word = 20| 99.6k| uint8_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point 21| 99.6k| *utf16_output++ = 22| 99.6k| char16_t(match_system(big_endian) ? word : u16_swap_bytes(word)); ------------------ | Branch (22:18): [True: 99.6k, Folded] ------------------ 23| 99.6k| pos++; 24| 99.6k| } 25| | 26| 6.82k| return utf16_output - start; 27| 6.82k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115latin1_to_utf167convertILNS_10endiannessE1EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 14| 6.82k| char16_t *utf16_output) { 15| 6.82k| size_t pos = 0; 16| 6.82k| char16_t *start{utf16_output}; 17| | 18| 106k| while (pos < len) { ------------------ | Branch (18:10): [True: 99.6k, False: 6.82k] ------------------ 19| 99.6k| uint16_t word = 20| 99.6k| uint8_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point 21| 99.6k| *utf16_output++ = 22| 99.6k| char16_t(match_system(big_endian) ? word : u16_swap_bytes(word)); ------------------ | Branch (22:18): [Folded, False: 99.6k] ------------------ 23| 99.6k| pos++; 24| 99.6k| } 25| | 26| 6.82k| return utf16_output - start; 27| 6.82k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115latin1_to_utf327convertIPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_mPDi: 14| 6.77k| char32_t *utf32_output) { 15| 6.77k| char32_t *start{utf32_output}; 16| 104k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (16:22): [True: 98.0k, False: 6.77k] ------------------ 17| 98.0k| *utf32_output++ = uint8_t(data[i]); 18| 98.0k| } 19| 6.77k| return utf32_output - start; 20| 6.77k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf87convertEPKcmPc: 64| 8.00k| char *utf8_output) { 65| 8.00k| return convert(reinterpret_cast(buf), len, 66| 8.00k| utf8_output); 67| 8.00k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf87convertIPKhPcQaasr7simdutf6detailE22indexes_into_byte_likeIT_Esr7simdutf6detailE26index_assignable_from_charIT0_EEEmS7_mS8_: 15| 8.00k| OutputPtr utf8_output) { 16| | // const unsigned char *data = reinterpret_cast(buf); 17| 8.00k| size_t pos = 0; 18| 8.00k| size_t utf8_pos = 0; 19| | 20| 78.2k| while (pos < len) { ------------------ | Branch (20:10): [True: 70.2k, False: 8.00k] ------------------ 21| |#if SIMDUTF_CPLUSPLUS23 22| | if !consteval 23| |#endif 24| 70.2k| { 25| | // try to convert the next block of 16 ASCII bytes 26| 70.2k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (26:11): [True: 52.0k, False: 18.2k] ------------------ 27| | // they are ascii 28| 52.0k| uint64_t v1; 29| 52.0k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 30| 52.0k| uint64_t v2; 31| 52.0k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 32| 52.0k| uint64_t v{v1 | 33| 52.0k| v2}; // We are only interested in these bits: 1000 1000 1000 34| | // 1000, so it makes sense to concatenate everything 35| 52.0k| if ((v & 0x8080808080808080) == ------------------ | Branch (35:13): [True: 6.05k, False: 45.9k] ------------------ 36| 52.0k| 0) { // if NONE of these are set, e.g. all of them are zero, then 37| | // everything is ASCII 38| 6.05k| size_t final_pos = pos + 16; 39| 102k| while (pos < final_pos) { ------------------ | Branch (39:18): [True: 96.9k, False: 6.05k] ------------------ 40| 96.9k| utf8_output[utf8_pos++] = char(data[pos]); 41| 96.9k| pos++; 42| 96.9k| } 43| 6.05k| continue; 44| 6.05k| } 45| 52.0k| } // if (pos + 16 <= len) 46| 70.2k| } // !consteval scope 47| | 48| 64.2k| unsigned char byte = data[pos]; 49| 64.2k| if ((byte & 0x80) == 0) { // if ASCII ------------------ | Branch (49:9): [True: 28.0k, False: 36.2k] ------------------ 50| | // will generate one UTF-8 bytes 51| 28.0k| utf8_output[utf8_pos++] = char(byte); 52| 28.0k| pos++; 53| 36.2k| } else { 54| | // will generate two UTF-8 bytes 55| 36.2k| utf8_output[utf8_pos++] = char((byte >> 6) | 0b11000000); 56| 36.2k| utf8_output[utf8_pos++] = char((byte & 0b111111) | 0b10000000); 57| 36.2k| pos++; 58| 36.2k| } 59| 64.2k| } // while 60| 8.00k| return utf8_pos; 61| 8.00k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf823utf8_length_from_latin1IPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_m: 150| 5.73k|utf8_length_from_latin1(InputPtr input, size_t length) noexcept { 151| 5.73k| size_t answer = length; 152| 5.73k| size_t i = 0; 153| | 154| |#if SIMDUTF_CPLUSPLUS23 155| | if !consteval 156| |#endif 157| 5.73k| { 158| 5.73k| auto pop = [](uint64_t v) { 159| 5.73k| return (size_t)(((v >> 7) & UINT64_C(0x0101010101010101)) * 160| 5.73k| UINT64_C(0x0101010101010101) >> 161| 5.73k| 56); 162| 5.73k| }; 163| 9.87k| for (; i + 32 <= length; i += 32) { ------------------ | Branch (163:12): [True: 4.14k, False: 5.73k] ------------------ 164| 4.14k| uint64_t v; 165| 4.14k| memcpy(&v, input + i, 8); 166| 4.14k| answer += pop(v); 167| 4.14k| memcpy(&v, input + i + 8, sizeof(v)); 168| 4.14k| answer += pop(v); 169| 4.14k| memcpy(&v, input + i + 16, sizeof(v)); 170| 4.14k| answer += pop(v); 171| 4.14k| memcpy(&v, input + i + 24, sizeof(v)); 172| 4.14k| answer += pop(v); 173| 4.14k| } 174| 6.02k| for (; i + 8 <= length; i += 8) { ------------------ | Branch (174:12): [True: 295, False: 5.73k] ------------------ 175| 295| uint64_t v; 176| 295| memcpy(&v, input + i, sizeof(v)); 177| 295| answer += pop(v); 178| 295| } 179| 5.73k| } // !consteval scope 180| 6.57k| for (; i + 1 <= length; i += 1) { ------------------ | Branch (180:10): [True: 846, False: 5.73k] ------------------ 181| 846| answer += static_cast(input[i]) >> 7; 182| 846| } 183| 5.73k| return answer; 184| 5.73k|} simdutf.cpp:_ZZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf823utf8_length_from_latin1IPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_mENKUlmE_clEm: 158| 16.8k| auto pop = [](uint64_t v) { 159| 16.8k| return (size_t)(((v >> 7) & UINT64_C(0x0101010101010101)) * 160| | UINT64_C(0x0101010101010101) >> 161| 16.8k| 56); 162| 16.8k| }; _ZN7simdutf6scalar14u16_swap_bytesEt: 8| 910k|u16_swap_bytes(const uint16_t word) { 9| 910k| return uint16_t((word >> 8) | (word << 8)); 10| 910k|} _ZN7simdutf6scalar14u32_swap_bytesEj: 13| 7.04k|u32_swap_bytes(const uint32_t word) { 14| 7.04k| return ((word >> 24) & 0xff) | // move byte 3 to byte 0 15| 7.04k| ((word << 8) & 0xff0000) | // move byte 1 to byte 2 16| 7.04k| ((word >> 8) & 0xff00) | // move byte 2 to byte 1 17| 7.04k| ((word << 24) & 0xff000000); // byte 0 to byte 3 18| 7.04k|} _ZN7simdutf6scalar5utf1614swap_if_neededILNS_10endiannessE1EEEtt: 27| 464k|template constexpr uint16_t swap_if_needed(uint16_t c) { 28| 464k| return !match_system(big_endian) ? scalar::u16_swap_bytes(c) : c; ------------------ | Branch (28:10): [True: 464k, Folded] ------------------ 29| 464k|} _ZN7simdutf6scalar5utf1614swap_if_neededILNS_10endiannessE0EEEtt: 27| 475k|template constexpr uint16_t swap_if_needed(uint16_t c) { 28| 475k| return !match_system(big_endian) ? scalar::u16_swap_bytes(c) : c; ------------------ | Branch (28:10): [Folded, False: 475k] ------------------ 29| 475k|} _ZN7simdutf6scalar5utf168validateILNS_10endiannessE0EEEbPKDsm: 22| 20.6k|validate(const char16_t *data, size_t len) noexcept { 23| 20.6k| uint64_t pos = 0; 24| 309k| while (pos < len) { ------------------ | Branch (24:10): [True: 288k, False: 20.4k] ------------------ 25| 288k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 26| 288k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (26:9): [True: 2.57k, False: 286k] ------------------ 27| 2.57k| if (pos + 1 >= len) { ------------------ | Branch (27:11): [True: 24, False: 2.54k] ------------------ 28| 24| return false; 29| 24| } 30| 2.54k| char16_t diff = char16_t(word - 0xD800); 31| 2.54k| if (diff > 0x3FF) { ------------------ | Branch (31:11): [True: 48, False: 2.50k] ------------------ 32| 48| return false; 33| 48| } 34| 2.50k| char16_t next_word = !match_system(big_endian) ------------------ | Branch (34:28): [Folded, False: 2.50k] ------------------ 35| 2.50k| ? u16_swap_bytes(data[pos + 1]) 36| 2.50k| : data[pos + 1]; 37| 2.50k| char16_t diff2 = char16_t(next_word - 0xDC00); 38| 2.50k| if (diff2 > 0x3FF) { ------------------ | Branch (38:11): [True: 75, False: 2.42k] ------------------ 39| 75| return false; 40| 75| } 41| 2.42k| pos += 2; 42| 286k| } else { 43| 286k| pos++; 44| 286k| } 45| 288k| } 46| 20.4k| return true; 47| 20.6k|} _ZN7simdutf6scalar5utf168validateILNS_10endiannessE1EEEbPKDsm: 22| 20.6k|validate(const char16_t *data, size_t len) noexcept { 23| 20.6k| uint64_t pos = 0; 24| 305k| while (pos < len) { ------------------ | Branch (24:10): [True: 285k, False: 20.4k] ------------------ 25| 285k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 26| 285k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (26:9): [True: 2.51k, False: 282k] ------------------ 27| 2.51k| if (pos + 1 >= len) { ------------------ | Branch (27:11): [True: 20, False: 2.49k] ------------------ 28| 20| return false; 29| 20| } 30| 2.49k| char16_t diff = char16_t(word - 0xD800); 31| 2.49k| if (diff > 0x3FF) { ------------------ | Branch (31:11): [True: 41, False: 2.45k] ------------------ 32| 41| return false; 33| 41| } 34| 2.45k| char16_t next_word = !match_system(big_endian) ------------------ | Branch (34:28): [True: 2.45k, Folded] ------------------ 35| 2.45k| ? u16_swap_bytes(data[pos + 1]) 36| 2.45k| : data[pos + 1]; 37| 2.45k| char16_t diff2 = char16_t(next_word - 0xDC00); 38| 2.45k| if (diff2 > 0x3FF) { ------------------ | Branch (38:11): [True: 72, False: 2.38k] ------------------ 39| 72| return false; 40| 72| } 41| 2.38k| pos += 2; 42| 282k| } else { 43| 282k| pos++; 44| 282k| } 45| 285k| } 46| 20.4k| return true; 47| 20.6k|} _ZN7simdutf6scalar5utf1620validate_with_errorsILNS_10endiannessE0EEENS_6resultEPKDsm: 51| 6.26k|validate_with_errors(const char16_t *data, size_t len) noexcept { 52| 6.26k| size_t pos = 0; 53| 54.6k| while (pos < len) { ------------------ | Branch (53:10): [True: 48.5k, False: 6.02k] ------------------ 54| 48.5k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 55| 48.5k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (55:9): [True: 931, False: 47.6k] ------------------ 56| 931| if (pos + 1 >= len) { ------------------ | Branch (56:11): [True: 24, False: 907] ------------------ 57| 24| return result(error_code::SURROGATE, pos); 58| 24| } 59| 907| char16_t diff = char16_t(word - 0xD800); 60| 907| if (diff > 0x3FF) { ------------------ | Branch (60:11): [True: 100, False: 807] ------------------ 61| 100| return result(error_code::SURROGATE, pos); 62| 100| } 63| 807| char16_t next_word = !match_system(big_endian) ------------------ | Branch (63:28): [Folded, False: 807] ------------------ 64| 807| ? u16_swap_bytes(data[pos + 1]) 65| 807| : data[pos + 1]; 66| 807| char16_t diff2 = uint16_t(next_word - 0xDC00); 67| 807| if (diff2 > 0x3FF) { ------------------ | Branch (67:11): [True: 118, False: 689] ------------------ 68| 118| return result(error_code::SURROGATE, pos); 69| 118| } 70| 689| pos += 2; 71| 47.6k| } else { 72| 47.6k| pos++; 73| 47.6k| } 74| 48.5k| } 75| 6.02k| return result(error_code::SUCCESS, pos); 76| 6.26k|} _ZN7simdutf6scalar5utf1620validate_with_errorsILNS_10endiannessE1EEENS_6resultEPKDsm: 51| 6.25k|validate_with_errors(const char16_t *data, size_t len) noexcept { 52| 6.25k| size_t pos = 0; 53| 51.1k| while (pos < len) { ------------------ | Branch (53:10): [True: 45.0k, False: 6.03k] ------------------ 54| 45.0k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 55| 45.0k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (55:9): [True: 880, False: 44.1k] ------------------ 56| 880| if (pos + 1 >= len) { ------------------ | Branch (56:11): [True: 20, False: 860] ------------------ 57| 20| return result(error_code::SURROGATE, pos); 58| 20| } 59| 860| char16_t diff = char16_t(word - 0xD800); 60| 860| if (diff > 0x3FF) { ------------------ | Branch (60:11): [True: 88, False: 772] ------------------ 61| 88| return result(error_code::SURROGATE, pos); 62| 88| } 63| 772| char16_t next_word = !match_system(big_endian) ------------------ | Branch (63:28): [True: 772, Folded] ------------------ 64| 772| ? u16_swap_bytes(data[pos + 1]) 65| 772| : data[pos + 1]; 66| 772| char16_t diff2 = uint16_t(next_word - 0xDC00); 67| 772| if (diff2 > 0x3FF) { ------------------ | Branch (67:11): [True: 113, False: 659] ------------------ 68| 113| return result(error_code::SURROGATE, pos); 69| 113| } 70| 659| pos += 2; 71| 44.1k| } else { 72| 44.1k| pos++; 73| 44.1k| } 74| 45.0k| } 75| 6.03k| return result(error_code::SUCCESS, pos); 76| 6.25k|} _ZN7simdutf6scalar5utf1622utf8_length_from_utf16ILNS_10endiannessE0EEEmPKDsm: 91| 32.1k| size_t len) { 92| | // We are not BOM aware. 93| 32.1k| size_t counter{0}; 94| 135k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (94:22): [True: 103k, False: 32.1k] ------------------ 95| 103k| char16_t word = scalar::utf16::swap_if_needed(p[i]); 96| 103k| counter++; // ASCII 97| 103k| counter += static_cast( 98| 103k| word > 99| 103k| 0x7F); // non-ASCII is at least 2 bytes, surrogates are 2*2 == 4 bytes 100| 103k| counter += static_cast((word > 0x7FF && word <= 0xD7FF) || ------------------ | Branch (100:37): [True: 41.5k, False: 62.1k] | Branch (100:53): [True: 35.6k, False: 5.83k] ------------------ 101| 67.9k| (word >= 0xE000)); // three-byte ------------------ | Branch (101:36): [True: 3.39k, False: 64.5k] ------------------ 102| 103k| } 103| 32.1k| return counter; 104| 32.1k|} _ZN7simdutf6scalar5utf1622utf8_length_from_utf16ILNS_10endiannessE1EEEmPKDsm: 91| 32.1k| size_t len) { 92| | // We are not BOM aware. 93| 32.1k| size_t counter{0}; 94| 133k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (94:22): [True: 101k, False: 32.1k] ------------------ 95| 101k| char16_t word = scalar::utf16::swap_if_needed(p[i]); 96| 101k| counter++; // ASCII 97| 101k| counter += static_cast( 98| 101k| word > 99| 101k| 0x7F); // non-ASCII is at least 2 bytes, surrogates are 2*2 == 4 bytes 100| 101k| counter += static_cast((word > 0x7FF && word <= 0xD7FF) || ------------------ | Branch (100:37): [True: 39.2k, False: 62.1k] | Branch (100:53): [True: 33.4k, False: 5.87k] ------------------ 101| 67.9k| (word >= 0xE000)); // three-byte ------------------ | Branch (101:36): [True: 3.13k, False: 64.8k] ------------------ 102| 101k| } 103| 32.1k| return counter; 104| 32.1k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115utf16_to_latin17convertILNS_10endiannessE0EPKDsPcQaasr7simdutf6detailE18indexes_into_utf16IT0_Esr7simdutf6detailE26index_assignable_from_charIT1_EEEmS8_mS9_: 17| 6.81k| OutputPtr latin_output) { 18| 6.81k| if (len == 0) { ------------------ | Branch (18:7): [True: 5.31k, False: 1.50k] ------------------ 19| 5.31k| return 0; 20| 5.31k| } 21| 1.50k| size_t pos = 0; 22| 1.50k| const auto latin_output_start = latin_output; 23| 1.50k| uint16_t word = 0; 24| 1.50k| uint16_t too_large = 0; 25| | 26| 103k| while (pos < len) { ------------------ | Branch (26:10): [True: 102k, False: 1.50k] ------------------ 27| 102k| word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (27:12): [Folded, False: 102k] ------------------ 28| 102k| too_large |= word; 29| 102k| *latin_output++ = char(word & 0xFF); 30| 102k| pos++; 31| 102k| } 32| 1.50k| if ((too_large & 0xFF00) != 0) { ------------------ | Branch (32:7): [True: 0, False: 1.50k] ------------------ 33| 0| return 0; 34| 0| } 35| | 36| 1.50k| return latin_output - latin_output_start; 37| 1.50k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115utf16_to_latin17convertILNS_10endiannessE1EPKDsPcQaasr7simdutf6detailE18indexes_into_utf16IT0_Esr7simdutf6detailE26index_assignable_from_charIT1_EEEmS8_mS9_: 17| 6.81k| OutputPtr latin_output) { 18| 6.81k| if (len == 0) { ------------------ | Branch (18:7): [True: 5.31k, False: 1.50k] ------------------ 19| 5.31k| return 0; 20| 5.31k| } 21| 1.50k| size_t pos = 0; 22| 1.50k| const auto latin_output_start = latin_output; 23| 1.50k| uint16_t word = 0; 24| 1.50k| uint16_t too_large = 0; 25| | 26| 103k| while (pos < len) { ------------------ | Branch (26:10): [True: 102k, False: 1.50k] ------------------ 27| 102k| word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (27:12): [True: 102k, Folded] ------------------ 28| 102k| too_large |= word; 29| 102k| *latin_output++ = char(word & 0xFF); 30| 102k| pos++; 31| 102k| } 32| 1.50k| if ((too_large & 0xFF00) != 0) { ------------------ | Branch (32:7): [True: 0, False: 1.50k] ------------------ 33| 0| return 0; 34| 0| } 35| | 36| 1.50k| return latin_output - latin_output_start; 37| 1.50k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf16_to_utf87convertILNS_10endiannessE0EPKDsPcQsr7simdutf6detailE18indexes_into_utf16IT0_EEEmS8_mT1_: 15| 13.1k| OutputPtr utf8_output) { 16| 13.1k| size_t pos = 0; 17| 13.1k| const auto start = utf8_output; 18| 75.9k| while (pos < len) { ------------------ | Branch (18:10): [True: 62.9k, False: 12.9k] ------------------ 19| |#if SIMDUTF_CPLUSPLUS23 20| | if !consteval 21| |#endif 22| 62.9k| { 23| | // try to convert the next block of 8 bytes 24| 62.9k| if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that ------------------ | Branch (24:11): [True: 58.6k, False: 4.26k] ------------------ 25| | // they are ascii 26| 58.6k| uint64_t v; 27| 58.6k| ::memcpy(&v, data + pos, sizeof(uint64_t)); 28| | if constexpr (!match_system(big_endian)) { 29| | v = (v >> 8) | (v << (64 - 8)); 30| | } 31| 58.6k| if ((v & 0xFF80FF80FF80FF80) == 0) { ------------------ | Branch (31:13): [True: 16.0k, False: 42.5k] ------------------ 32| 16.0k| size_t final_pos = pos + 4; 33| 80.3k| while (pos < final_pos) { ------------------ | Branch (33:18): [True: 64.2k, False: 16.0k] ------------------ 34| 64.2k| *utf8_output++ = !match_system(big_endian) ------------------ | Branch (34:30): [Folded, False: 64.2k] ------------------ 35| 64.2k| ? char(u16_swap_bytes(data[pos])) 36| 64.2k| : char(data[pos]); 37| 64.2k| pos++; 38| 64.2k| } 39| 16.0k| continue; 40| 16.0k| } 41| 58.6k| } 42| 62.9k| } 43| 46.8k| uint16_t word = 44| 46.8k| !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (44:9): [Folded, False: 46.8k] ------------------ 45| 46.8k| if ((word & 0xFF80) == 0) { ------------------ | Branch (45:9): [True: 4.89k, False: 41.9k] ------------------ 46| | // will generate one UTF-8 bytes 47| 4.89k| *utf8_output++ = char(word); 48| 4.89k| pos++; 49| 41.9k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (49:16): [True: 3.04k, False: 38.9k] ------------------ 50| | // will generate two UTF-8 bytes 51| | // we have 0b110XXXXX 0b10XXXXXX 52| 3.04k| *utf8_output++ = char((word >> 6) | 0b11000000); 53| 3.04k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 54| 3.04k| pos++; 55| 38.9k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (55:16): [True: 37.4k, False: 1.48k] ------------------ 56| | // will generate three UTF-8 bytes 57| | // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX 58| 37.4k| *utf8_output++ = char((word >> 12) | 0b11100000); 59| 37.4k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 60| 37.4k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 61| 37.4k| pos++; 62| 37.4k| } else { 63| | // must be a surrogate pair 64| 1.48k| if (pos + 1 >= len) { ------------------ | Branch (64:11): [True: 24, False: 1.46k] ------------------ 65| 24| return 0; 66| 24| } 67| 1.46k| uint16_t diff = uint16_t(word - 0xD800); 68| 1.46k| if (diff > 0x3FF) { ------------------ | Branch (68:11): [True: 42, False: 1.41k] ------------------ 69| 42| return 0; 70| 42| } 71| 1.41k| uint16_t next_word = !match_system(big_endian) ------------------ | Branch (71:28): [Folded, False: 1.41k] ------------------ 72| 1.41k| ? u16_swap_bytes(data[pos + 1]) 73| 1.41k| : data[pos + 1]; 74| 1.41k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 75| 1.41k| if (diff2 > 0x3FF) { ------------------ | Branch (75:11): [True: 62, False: 1.35k] ------------------ 76| 62| return 0; 77| 62| } 78| 1.35k| uint32_t value = (diff << 10) + diff2 + 0x10000; 79| | // will generate four UTF-8 bytes 80| | // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX 81| 1.35k| *utf8_output++ = char((value >> 18) | 0b11110000); 82| 1.35k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 83| 1.35k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 84| 1.35k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 85| 1.35k| pos += 2; 86| 1.35k| } 87| 46.8k| } 88| 12.9k| return utf8_output - start; 89| 13.1k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf16_to_utf87convertILNS_10endiannessE1EPKDsPcQsr7simdutf6detailE18indexes_into_utf16IT0_EEEmS8_mT1_: 15| 13.0k| OutputPtr utf8_output) { 16| 13.0k| size_t pos = 0; 17| 13.0k| const auto start = utf8_output; 18| 71.2k| while (pos < len) { ------------------ | Branch (18:10): [True: 58.3k, False: 12.9k] ------------------ 19| |#if SIMDUTF_CPLUSPLUS23 20| | if !consteval 21| |#endif 22| 58.3k| { 23| | // try to convert the next block of 8 bytes 24| 58.3k| if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that ------------------ | Branch (24:11): [True: 54.0k, False: 4.21k] ------------------ 25| | // they are ascii 26| 54.0k| uint64_t v; 27| 54.0k| ::memcpy(&v, data + pos, sizeof(uint64_t)); 28| 54.0k| if constexpr (!match_system(big_endian)) { 29| 54.0k| v = (v >> 8) | (v << (64 - 8)); 30| 54.0k| } 31| 54.0k| if ((v & 0xFF80FF80FF80FF80) == 0) { ------------------ | Branch (31:13): [True: 16.1k, False: 37.9k] ------------------ 32| 16.1k| size_t final_pos = pos + 4; 33| 80.8k| while (pos < final_pos) { ------------------ | Branch (33:18): [True: 64.7k, False: 16.1k] ------------------ 34| 64.7k| *utf8_output++ = !match_system(big_endian) ------------------ | Branch (34:30): [True: 64.7k, Folded] ------------------ 35| 64.7k| ? char(u16_swap_bytes(data[pos])) 36| 64.7k| : char(data[pos]); 37| 64.7k| pos++; 38| 64.7k| } 39| 16.1k| continue; 40| 16.1k| } 41| 54.0k| } 42| 58.3k| } 43| 42.1k| uint16_t word = 44| 42.1k| !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (44:9): [True: 42.1k, Folded] ------------------ 45| 42.1k| if ((word & 0xFF80) == 0) { ------------------ | Branch (45:9): [True: 4.77k, False: 37.3k] ------------------ 46| | // will generate one UTF-8 bytes 47| 4.77k| *utf8_output++ = char(word); 48| 4.77k| pos++; 49| 37.3k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (49:16): [True: 2.26k, False: 35.0k] ------------------ 50| | // will generate two UTF-8 bytes 51| | // we have 0b110XXXXX 0b10XXXXXX 52| 2.26k| *utf8_output++ = char((word >> 6) | 0b11000000); 53| 2.26k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 54| 2.26k| pos++; 55| 35.0k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (55:16): [True: 33.6k, False: 1.43k] ------------------ 56| | // will generate three UTF-8 bytes 57| | // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX 58| 33.6k| *utf8_output++ = char((word >> 12) | 0b11100000); 59| 33.6k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 60| 33.6k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 61| 33.6k| pos++; 62| 33.6k| } else { 63| | // must be a surrogate pair 64| 1.43k| if (pos + 1 >= len) { ------------------ | Branch (64:11): [True: 20, False: 1.41k] ------------------ 65| 20| return 0; 66| 20| } 67| 1.41k| uint16_t diff = uint16_t(word - 0xD800); 68| 1.41k| if (diff > 0x3FF) { ------------------ | Branch (68:11): [True: 28, False: 1.38k] ------------------ 69| 28| return 0; 70| 28| } 71| 1.38k| uint16_t next_word = !match_system(big_endian) ------------------ | Branch (71:28): [True: 1.38k, Folded] ------------------ 72| 1.38k| ? u16_swap_bytes(data[pos + 1]) 73| 1.38k| : data[pos + 1]; 74| 1.38k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 75| 1.38k| if (diff2 > 0x3FF) { ------------------ | Branch (75:11): [True: 65, False: 1.32k] ------------------ 76| 65| return 0; 77| 65| } 78| 1.32k| uint32_t value = (diff << 10) + diff2 + 0x10000; 79| | // will generate four UTF-8 bytes 80| | // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX 81| 1.32k| *utf8_output++ = char((value >> 18) | 0b11110000); 82| 1.32k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 83| 1.32k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 84| 1.32k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 85| 1.32k| pos += 2; 86| 1.32k| } 87| 42.1k| } 88| 12.9k| return utf8_output - start; 89| 13.0k|} _ZN7simdutf6scalar5utf328validateEPKDim: 25| 14.4k| size_t len) noexcept { 26| 14.4k| return validate(reinterpret_cast(buf), len); 27| 14.4k|} _ZN7simdutf6scalar5utf3222utf8_length_from_utf32EPKDim: 54| 15.4k| size_t len) { 55| | // We are not BOM aware. 56| 15.4k| size_t counter{0}; 57| 68.8k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (57:22): [True: 53.4k, False: 15.4k] ------------------ 58| | // credit: @ttsugriy for the vectorizable approach 59| 53.4k| counter++; // ASCII 60| 53.4k| counter += static_cast(p[i] > 0x7F); // two-byte 61| 53.4k| counter += static_cast(p[i] > 0x7FF); // three-byte 62| 53.4k| counter += static_cast(p[i] > 0xFFFF); // four-bytes 63| 53.4k| } 64| 15.4k| return counter; 65| 15.4k|} _ZN7simdutf6scalar5utf328validateIPKjQsr7simdutf6detailE19indexes_into_uint32IT_EEEbS5_m: 13| 14.4k| size_t len) noexcept { 14| 14.4k| uint64_t pos = 0; 15| 171k| for (; pos < len; pos++) { ------------------ | Branch (15:10): [True: 156k, False: 14.4k] ------------------ 16| 156k| uint32_t word = data[pos]; 17| 156k| if (word > 0x10FFFF || (word >= 0xD800 && word <= 0xDFFF)) { ------------------ | Branch (17:9): [True: 0, False: 156k] | Branch (17:29): [True: 1.04k, False: 155k] | Branch (17:47): [True: 0, False: 1.04k] ------------------ 18| 0| return false; 19| 0| } 20| 156k| } 21| 14.4k| return true; 22| 14.4k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115utf32_to_latin17convertEPKDimPc: 10| 6.84k| char *latin1_output) { 11| 6.84k| char *start = latin1_output; 12| 6.84k| uint32_t utf32_char; 13| 6.84k| size_t pos = 0; 14| 6.84k| uint32_t too_large = 0; 15| | 16| 110k| while (pos < len) { ------------------ | Branch (16:10): [True: 103k, False: 6.84k] ------------------ 17| 103k| utf32_char = (uint32_t)data[pos]; 18| 103k| too_large |= utf32_char; 19| 103k| *latin1_output++ = (char)(utf32_char & 0xFF); 20| 103k| pos++; 21| 103k| } 22| 6.84k| if ((too_large & 0xFFFFFF00) != 0) { ------------------ | Branch (22:7): [True: 0, False: 6.84k] ------------------ 23| 0| return 0; 24| 0| } 25| 6.84k| return latin1_output - start; 26| 6.84k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf32_to_utf87convertIPKDiPcQaasr7simdutf6detailE18indexes_into_utf32IT_Esr7simdutf6detailE26index_assignable_from_charIT0_EEEmS7_mS8_: 15| 6.95k| OutputPtr utf8_output) { 16| 6.95k| size_t pos = 0; 17| 6.95k| auto start = utf8_output; 18| 44.1k| while (pos < len) { ------------------ | Branch (18:10): [True: 37.2k, False: 6.95k] ------------------ 19| |#if SIMDUTF_CPLUSPLUS23 20| | if !consteval 21| |#endif 22| 37.2k| { // try to convert the next block of 2 ASCII characters 23| 37.2k| if (pos + 2 <= len) { // if it is safe to read 8 more bytes, check that ------------------ | Branch (23:11): [True: 35.9k, False: 1.32k] ------------------ 24| | // they are ascii 25| 35.9k| uint64_t v; 26| 35.9k| ::memcpy(&v, data + pos, sizeof(uint64_t)); 27| 35.9k| if ((v & 0xFFFFFF80FFFFFF80) == 0) { ------------------ | Branch (27:13): [True: 32.0k, False: 3.87k] ------------------ 28| 32.0k| *utf8_output++ = char(data[pos]); 29| 32.0k| *utf8_output++ = char(data[pos + 1]); 30| 32.0k| pos += 2; 31| 32.0k| continue; 32| 32.0k| } 33| 35.9k| } 34| 37.2k| } 35| | 36| 5.19k| uint32_t word = data[pos]; 37| 5.19k| if ((word & 0xFFFFFF80) == 0) { ------------------ | Branch (37:9): [True: 1.67k, False: 3.52k] ------------------ 38| | // will generate one UTF-8 bytes 39| 1.67k| *utf8_output++ = char(word); 40| 1.67k| pos++; 41| 3.52k| } else if ((word & 0xFFFFF800) == 0) { ------------------ | Branch (41:16): [True: 1.38k, False: 2.14k] ------------------ 42| | // will generate two UTF-8 bytes 43| | // we have 0b110XXXXX 0b10XXXXXX 44| 1.38k| *utf8_output++ = char((word >> 6) | 0b11000000); 45| 1.38k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 46| 1.38k| pos++; 47| 2.14k| } else if ((word & 0xFFFF0000) == 0) { ------------------ | Branch (47:16): [True: 952, False: 1.18k] ------------------ 48| | // will generate three UTF-8 bytes 49| | // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX 50| 952| if (word >= 0xD800 && word <= 0xDFFF) { ------------------ | Branch (50:11): [True: 494, False: 458] | Branch (50:29): [True: 0, False: 494] ------------------ 51| 0| return 0; 52| 0| } 53| 952| *utf8_output++ = char((word >> 12) | 0b11100000); 54| 952| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 55| 952| *utf8_output++ = char((word & 0b111111) | 0b10000000); 56| 952| pos++; 57| 1.18k| } else { 58| | // will generate four UTF-8 bytes 59| | // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX 60| 1.18k| if (word > 0x10FFFF) { ------------------ | Branch (60:11): [True: 0, False: 1.18k] ------------------ 61| 0| return 0; 62| 0| } 63| 1.18k| *utf8_output++ = char((word >> 18) | 0b11110000); 64| 1.18k| *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000); 65| 1.18k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 66| 1.18k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 67| 1.18k| pos++; 68| 1.18k| } 69| 5.19k| } 70| 6.95k| return utf8_output - start; 71| 6.95k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf88validateEPKcm: 111| 22.1k| size_t len) noexcept { 112| 22.1k| return validate(reinterpret_cast(buf), len); 113| 22.1k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf88validateIPKhEEbT_m: 12| 22.1k| size_t len) noexcept { 13| 22.1k| static_assert( 14| 22.1k| std::is_same::type, uint8_t>::value, 15| 22.1k| "dereferencing the data pointer must result in a uint8_t"); 16| 22.1k| uint64_t pos = 0; 17| 22.1k| uint32_t code_point = 0; 18| 113k| while (pos < len) { ------------------ | Branch (18:10): [True: 91.4k, False: 21.6k] ------------------ 19| 91.4k| uint64_t next_pos; 20| |#if SIMDUTF_CPLUSPLUS23 21| | if !consteval 22| |#endif 23| 91.4k| { // check if the next 16 bytes are ascii. 24| 91.4k| next_pos = pos + 16; 25| 91.4k| if (next_pos <= len) { // if it is safe to read 16 more bytes, check ------------------ | Branch (25:11): [True: 89.0k, False: 2.33k] ------------------ 26| | // that they are ascii 27| 89.0k| uint64_t v1{}; 28| 89.0k| std::memcpy(&v1, data + pos, sizeof(uint64_t)); 29| 89.0k| uint64_t v2{}; 30| 89.0k| std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 31| 89.0k| uint64_t v{v1 | v2}; 32| 89.0k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (32:13): [True: 6.34k, False: 82.7k] ------------------ 33| 6.34k| pos = next_pos; 34| 6.34k| continue; 35| 6.34k| } 36| 89.0k| } 37| 91.4k| } 38| | 39| 85.0k| unsigned char byte = data[pos]; 40| | 41| 110k| while (byte < 0b10000000) { ------------------ | Branch (41:12): [True: 25.6k, False: 84.7k] ------------------ 42| 25.6k| if (++pos == len) { ------------------ | Branch (42:11): [True: 290, False: 25.3k] ------------------ 43| 290| return true; 44| 290| } 45| 25.3k| byte = data[pos]; 46| 25.3k| } 47| | 48| 84.7k| if ((byte & 0b11100000) == 0b11000000) { ------------------ | Branch (48:9): [True: 32.4k, False: 52.2k] ------------------ 49| 32.4k| next_pos = pos + 2; 50| 32.4k| if (next_pos > len) { ------------------ | Branch (50:11): [True: 11, False: 32.4k] ------------------ 51| 11| return false; 52| 11| } 53| 32.4k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (53:11): [True: 31, False: 32.4k] ------------------ 54| 31| return false; 55| 31| } 56| | // range check 57| 32.4k| code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 58| 32.4k| if (code_point < 0x80) { ------------------ | Branch (58:11): [True: 4, False: 32.4k] ------------------ 59| 4| return false; 60| 4| } 61| 52.2k| } else if ((byte & 0b11110000) == 0b11100000) { ------------------ | Branch (61:16): [True: 51.0k, False: 1.20k] ------------------ 62| 51.0k| next_pos = pos + 3; 63| 51.0k| if (next_pos > len) { ------------------ | Branch (63:11): [True: 10, False: 51.0k] ------------------ 64| 10| return false; 65| 10| } 66| 51.0k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (66:11): [True: 8, False: 51.0k] ------------------ 67| 8| return false; 68| 8| } 69| 51.0k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (69:11): [True: 4, False: 51.0k] ------------------ 70| 4| return false; 71| 4| } 72| | // range check 73| 51.0k| code_point = (byte & 0b00001111) << 12 | 74| 51.0k| (data[pos + 1] & 0b00111111) << 6 | 75| 51.0k| (data[pos + 2] & 0b00111111); 76| 51.0k| if ((code_point < 0x800) || ------------------ | Branch (76:11): [True: 8, False: 51.0k] ------------------ 77| 51.0k| (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (77:12): [True: 4.02k, False: 47.0k] | Branch (77:35): [True: 4, False: 4.01k] ------------------ 78| 12| return false; 79| 12| } 80| 51.0k| } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (80:16): [True: 1.11k, False: 89] ------------------ 81| 1.11k| next_pos = pos + 4; 82| 1.11k| if (next_pos > len) { ------------------ | Branch (82:11): [True: 7, False: 1.11k] ------------------ 83| 7| return false; 84| 7| } 85| 1.11k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (85:11): [True: 6, False: 1.10k] ------------------ 86| 6| return false; 87| 6| } 88| 1.10k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (88:11): [True: 12, False: 1.09k] ------------------ 89| 12| return false; 90| 12| } 91| 1.09k| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (91:11): [True: 4, False: 1.08k] ------------------ 92| 4| return false; 93| 4| } 94| | // range check 95| 1.08k| code_point = 96| 1.08k| (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 | 97| 1.08k| (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111); 98| 1.08k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (98:11): [True: 7, False: 1.08k] | Branch (98:35): [True: 5, False: 1.07k] ------------------ 99| 12| return false; 100| 12| } 101| 1.08k| } else { 102| | // we may have a continuation 103| 89| return false; 104| 89| } 105| 84.5k| pos = next_pos; 106| 84.5k| } 107| 21.6k| return true; 108| 22.1k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf820validate_with_errorsEPKcm: 219| 6.50k|validate_with_errors(const char *buf, size_t len) noexcept { 220| 6.50k| return validate_with_errors(reinterpret_cast(buf), len); 221| 6.50k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf820validate_with_errorsIPKhEENS_6resultET_m: 117| 6.50k|validate_with_errors(BytePtr data, size_t len) noexcept { 118| 6.50k| static_assert( 119| 6.50k| std::is_same::type, uint8_t>::value, 120| 6.50k| "dereferencing the data pointer must result in a uint8_t"); 121| 6.50k| size_t pos = 0; 122| 6.50k| uint32_t code_point = 0; 123| 13.2k| while (pos < len) { ------------------ | Branch (123:10): [True: 8.04k, False: 5.22k] ------------------ 124| | // check of the next 16 bytes are ascii. 125| 8.04k| size_t next_pos = pos + 16; 126| 8.04k| if (next_pos <= ------------------ | Branch (126:9): [True: 6.96k, False: 1.08k] ------------------ 127| 8.04k| len) { // if it is safe to read 16 more bytes, check that they are ascii 128| 6.96k| uint64_t v1; 129| 6.96k| std::memcpy(&v1, data + pos, sizeof(uint64_t)); 130| 6.96k| uint64_t v2; 131| 6.96k| std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 132| 6.96k| uint64_t v{v1 | v2}; 133| 6.96k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (133:11): [True: 3.16k, False: 3.79k] ------------------ 134| 3.16k| pos = next_pos; 135| 3.16k| continue; 136| 3.16k| } 137| 6.96k| } 138| 4.87k| unsigned char byte = data[pos]; 139| | 140| 14.5k| while (byte < 0b10000000) { ------------------ | Branch (140:12): [True: 9.84k, False: 4.74k] ------------------ 141| 9.84k| if (++pos == len) { ------------------ | Branch (141:11): [True: 132, False: 9.71k] ------------------ 142| 132| return result(error_code::SUCCESS, len); 143| 132| } 144| 9.71k| byte = data[pos]; 145| 9.71k| } 146| | 147| 4.74k| if ((byte & 0b11100000) == 0b11000000) { ------------------ | Branch (147:9): [True: 2.14k, False: 2.60k] ------------------ 148| 2.14k| next_pos = pos + 2; 149| 2.14k| if (next_pos > len) { ------------------ | Branch (149:11): [True: 45, False: 2.09k] ------------------ 150| 45| return result(error_code::TOO_SHORT, pos); 151| 45| } 152| 2.09k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (152:11): [True: 123, False: 1.97k] ------------------ 153| 123| return result(error_code::TOO_SHORT, pos); 154| 123| } 155| | // range check 156| 1.97k| code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 157| 1.97k| if (code_point < 0x80) { ------------------ | Branch (157:11): [True: 21, False: 1.95k] ------------------ 158| 21| return result(error_code::OVERLONG, pos); 159| 21| } 160| 2.60k| } else if ((byte & 0b11110000) == 0b11100000) { ------------------ | Branch (160:16): [True: 1.11k, False: 1.48k] ------------------ 161| 1.11k| next_pos = pos + 3; 162| 1.11k| if (next_pos > len) { ------------------ | Branch (162:11): [True: 57, False: 1.05k] ------------------ 163| 57| return result(error_code::TOO_SHORT, pos); 164| 57| } 165| 1.05k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (165:11): [True: 74, False: 985] ------------------ 166| 74| return result(error_code::TOO_SHORT, pos); 167| 74| } 168| 985| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (168:11): [True: 45, False: 940] ------------------ 169| 45| return result(error_code::TOO_SHORT, pos); 170| 45| } 171| | // range check 172| 940| code_point = (byte & 0b00001111) << 12 | 173| 940| (data[pos + 1] & 0b00111111) << 6 | 174| 940| (data[pos + 2] & 0b00111111); 175| 940| if (code_point < 0x800) { ------------------ | Branch (175:11): [True: 25, False: 915] ------------------ 176| 25| return result(error_code::OVERLONG, pos); 177| 25| } 178| 915| if (0xd7ff < code_point && code_point < 0xe000) { ------------------ | Branch (178:11): [True: 394, False: 521] | Branch (178:34): [True: 16, False: 378] ------------------ 179| 16| return result(error_code::SURROGATE, pos); 180| 16| } 181| 1.48k| } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (181:16): [True: 980, False: 509] ------------------ 182| 980| next_pos = pos + 4; 183| 980| if (next_pos > len) { ------------------ | Branch (183:11): [True: 38, False: 942] ------------------ 184| 38| return result(error_code::TOO_SHORT, pos); 185| 38| } 186| 942| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (186:11): [True: 71, False: 871] ------------------ 187| 71| return result(error_code::TOO_SHORT, pos); 188| 71| } 189| 871| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (189:11): [True: 49, False: 822] ------------------ 190| 49| return result(error_code::TOO_SHORT, pos); 191| 49| } 192| 822| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (192:11): [True: 44, False: 778] ------------------ 193| 44| return result(error_code::TOO_SHORT, pos); 194| 44| } 195| | // range check 196| 778| code_point = 197| 778| (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 | 198| 778| (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111); 199| 778| if (code_point <= 0xffff) { ------------------ | Branch (199:11): [True: 18, False: 760] ------------------ 200| 18| return result(error_code::OVERLONG, pos); 201| 18| } 202| 760| if (0x10ffff < code_point) { ------------------ | Branch (202:11): [True: 19, False: 741] ------------------ 203| 19| return result(error_code::TOO_LARGE, pos); 204| 19| } 205| 760| } else { 206| | // we either have too many continuation bytes or an invalid leading byte 207| 509| if ((byte & 0b11000000) == 0b10000000) { ------------------ | Branch (207:11): [True: 302, False: 207] ------------------ 208| 302| return result(error_code::TOO_LONG, pos); 209| 302| } else { 210| 207| return result(error_code::HEADER_BITS, pos); 211| 207| } 212| 509| } 213| 3.59k| pos = next_pos; 214| 3.59k| } 215| 5.22k| return result(error_code::SUCCESS, len); 216| 6.50k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf831rewind_and_validate_with_errorsEPKcS4_m: 230| 1.02k| const char *start, const char *buf, size_t len) noexcept { 231| | // First check that we start with a leading byte 232| 1.02k| if ((*start & 0b11000000) == 0b10000000) { ------------------ | Branch (232:7): [True: 76, False: 944] ------------------ 233| 76| return result(error_code::TOO_LONG, 0); 234| 76| } 235| 944| size_t extra_len{0}; 236| | // A leading byte cannot be further than 4 bytes away 237| 982| for (int i = 0; i < 5; i++) { ------------------ | Branch (237:19): [True: 982, False: 0] ------------------ 238| 982| unsigned char byte = *buf; 239| 982| if ((byte & 0b11000000) != 0b10000000) { ------------------ | Branch (239:9): [True: 944, False: 38] ------------------ 240| 944| break; 241| 944| } else { 242| 38| buf--; 243| 38| extra_len++; 244| 38| } 245| 982| } 246| | 247| 944| result res = validate_with_errors(buf, len + extra_len); 248| 944| res.count -= extra_len; // Might underflow 249| 944| return res; 250| 1.02k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf817count_code_pointsIPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_m: 256| 50.0k|simdutf_constexpr23 size_t count_code_points(InputPtr data, size_t len) { 257| 50.0k| size_t counter{0}; 258| 546k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (258:22): [True: 496k, False: 50.0k] ------------------ 259| | // -65 is 0b10111111, anything larger in two-complement's should start a new 260| | // code point. 261| 496k| if (int8_t(data[i]) > -65) { ------------------ | Branch (261:9): [True: 412k, False: 84.5k] ------------------ 262| 412k| counter++; 263| 412k| } 264| 496k| } 265| 50.0k| return counter; 266| 50.0k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf822utf16_length_from_utf8IPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_m: 272| 66.2k|simdutf_constexpr23 size_t utf16_length_from_utf8(InputPtr data, size_t len) { 273| 66.2k| size_t counter{0}; 274| 479k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (274:22): [True: 412k, False: 66.2k] ------------------ 275| 412k| if (int8_t(data[i]) > -65) { ------------------ | Branch (275:9): [True: 276k, False: 136k] ------------------ 276| 276k| counter++; 277| 276k| } 278| 412k| if (uint8_t(data[i]) >= 240) { ------------------ | Branch (278:9): [True: 10.1k, False: 402k] ------------------ 279| 10.1k| counter++; 280| 10.1k| } 281| 412k| } 282| 66.2k| return counter; 283| 66.2k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114utf8_to_latin17convertIPKcPcQaasr7simdutf6detailE22indexes_into_byte_likeIT_Esr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mS8_: 15| 15.3k| OutputPtr latin_output) { 16| 15.3k| size_t pos = 0; 17| 15.3k| auto start = latin_output; 18| | 19| 107k| while (pos < len) { ------------------ | Branch (19:10): [True: 93.9k, False: 13.2k] ------------------ 20| |#if SIMDUTF_CPLUSPLUS23 21| | if !consteval 22| |#endif 23| 93.9k| { 24| | // try to convert the next block of 16 ASCII bytes 25| 93.9k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (25:11): [True: 76.6k, False: 17.2k] ------------------ 26| | // they are ascii 27| 76.6k| uint64_t v1; 28| 76.6k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 29| 76.6k| uint64_t v2; 30| 76.6k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 31| 76.6k| uint64_t v{v1 | v2}; // We are only interested in these bits: 1000 1000 32| | // 1000 1000 .... etc 33| 76.6k| if ((v & 0x8080808080808080) == ------------------ | Branch (33:13): [True: 9.69k, False: 67.0k] ------------------ 34| 76.6k| 0) { // if NONE of these are set, e.g. all of them are zero, then 35| | // everything is ASCII 36| 9.69k| size_t final_pos = pos + 16; 37| 164k| while (pos < final_pos) { ------------------ | Branch (37:18): [True: 155k, False: 9.69k] ------------------ 38| 155k| *latin_output++ = char(data[pos]); 39| 155k| pos++; 40| 155k| } 41| 9.69k| continue; 42| 9.69k| } 43| 76.6k| } 44| 93.9k| } 45| | 46| | // suppose it is not an all ASCII byte sequence 47| 84.2k| uint8_t leading_byte = data[pos]; // leading byte 48| 84.2k| if (leading_byte < 0b10000000) { ------------------ | Branch (48:9): [True: 39.1k, False: 45.0k] ------------------ 49| | // converting one ASCII byte !!! 50| 39.1k| *latin_output++ = char(leading_byte); 51| 39.1k| pos++; 52| 45.0k| } else if ((leading_byte & 0b11100000) == ------------------ | Branch (52:16): [True: 43.2k, False: 1.72k] ------------------ 53| 45.0k| 0b11000000) { // the first three bits indicate: 54| | // We have a two-byte UTF-8 55| 43.2k| if (pos + 1 >= len) { ------------------ | Branch (55:11): [True: 24, False: 43.2k] ------------------ 56| 24| return 0; 57| 24| } // minimal bound checking 58| 43.2k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (58:11): [True: 86, False: 43.1k] ------------------ 59| 86| return 0; 60| 86| } // checks if the next byte is a valid continuation byte in UTF-8. A 61| | // valid continuation byte starts with 10. 62| | // range check - 63| 43.1k| uint32_t code_point = 64| 43.1k| (leading_byte & 0b00011111) << 6 | 65| 43.1k| (data[pos + 1] & 66| 43.1k| 0b00111111); // assembles the Unicode code point from the two bytes. 67| | // It does this by discarding the leading 110 and 10 68| | // bits from the two bytes, shifting the remaining bits 69| | // of the first byte, and then combining the results 70| | // with a bitwise OR operation. 71| 43.1k| if (code_point < 0x80 || 0xFF < code_point) { ------------------ | Branch (71:11): [True: 21, False: 43.1k] | Branch (71:32): [True: 171, False: 42.9k] ------------------ 72| 192| return 0; // We only care about the range 129-255 which is Non-ASCII 73| | // latin1 characters. A code_point beneath 0x80 is invalid as 74| | // it is already covered by bytes whose leading bit is zero. 75| 192| } 76| 42.9k| *latin_output++ = char(code_point); 77| 42.9k| pos += 2; 78| 42.9k| } else { 79| 1.72k| return 0; 80| 1.72k| } 81| 84.2k| } 82| 13.2k| return latin_output - start; 83| 15.3k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf8_to_utf167convertILNS_10endiannessE0EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 14| 13.7k| char16_t *utf16_output) { 15| 13.7k| size_t pos = 0; 16| 13.7k| char16_t *start{utf16_output}; 17| 82.2k| while (pos < len) { ------------------ | Branch (17:10): [True: 69.2k, False: 12.9k] ------------------ 18| |#if SIMDUTF_CPLUSPLUS23 19| | if !consteval 20| |#endif 21| | // try to convert the next block of 16 ASCII bytes 22| 69.2k| { 23| 69.2k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (23:11): [True: 54.6k, False: 14.6k] ------------------ 24| | // they are ascii 25| 54.6k| uint64_t v1; 26| 54.6k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 27| 54.6k| uint64_t v2; 28| 54.6k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 29| 54.6k| uint64_t v{v1 | v2}; 30| 54.6k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (30:13): [True: 4.43k, False: 50.1k] ------------------ 31| 4.43k| size_t final_pos = pos + 16; 32| 75.4k| while (pos < final_pos) { ------------------ | Branch (32:18): [True: 71.0k, False: 4.43k] ------------------ 33| 71.0k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (33:31): [Folded, False: 71.0k] ------------------ 34| 71.0k| ? char16_t(u16_swap_bytes(data[pos])) 35| 71.0k| : char16_t(data[pos]); 36| 71.0k| pos++; 37| 71.0k| } 38| 4.43k| continue; 39| 4.43k| } 40| 54.6k| } 41| 69.2k| } 42| | 43| 64.8k| uint8_t leading_byte = data[pos]; // leading byte 44| 64.8k| if (leading_byte < 0b10000000) { ------------------ | Branch (44:9): [True: 23.5k, False: 41.2k] ------------------ 45| | // converting one ASCII byte !!! 46| 23.5k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (46:25): [Folded, False: 23.5k] ------------------ 47| 23.5k| ? char16_t(u16_swap_bytes(leading_byte)) 48| 23.5k| : char16_t(leading_byte); 49| 23.5k| pos++; 50| 41.2k| } else if ((leading_byte & 0b11100000) == 0b11000000) { ------------------ | Branch (50:16): [True: 4.40k, False: 36.8k] ------------------ 51| | // We have a two-byte UTF-8, it should become 52| | // a single UTF-16 word. 53| 4.40k| if (pos + 1 >= len) { ------------------ | Branch (53:11): [True: 45, False: 4.35k] ------------------ 54| 45| return 0; 55| 45| } // minimal bound checking 56| 4.35k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (56:11): [True: 83, False: 4.27k] ------------------ 57| 83| return 0; 58| 83| } 59| | // range check 60| 4.27k| uint32_t code_point = 61| 4.27k| (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 62| 4.27k| if (code_point < 0x80) { ------------------ | Branch (62:11): [True: 21, False: 4.25k] ------------------ 63| 21| return 0; 64| 21| } 65| | if constexpr (!match_system(big_endian)) { 66| | code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 67| | } 68| 4.25k| *utf16_output++ = char16_t(code_point); 69| 4.25k| pos += 2; 70| 36.8k| } else if ((leading_byte & 0b11110000) == 0b11100000) { ------------------ | Branch (70:16): [True: 33.8k, False: 3.00k] ------------------ 71| | // We have a three-byte UTF-8, it should become 72| | // a single UTF-16 word. 73| 33.8k| if (pos + 2 >= len) { ------------------ | Branch (73:11): [True: 57, False: 33.8k] ------------------ 74| 57| return 0; 75| 57| } // minimal bound checking 76| | 77| 33.8k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (77:11): [True: 34, False: 33.7k] ------------------ 78| 34| return 0; 79| 34| } 80| 33.7k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (80:11): [True: 22, False: 33.7k] ------------------ 81| 22| return 0; 82| 22| } 83| | // range check 84| 33.7k| uint32_t code_point = (leading_byte & 0b00001111) << 12 | 85| 33.7k| (data[pos + 1] & 0b00111111) << 6 | 86| 33.7k| (data[pos + 2] & 0b00111111); 87| 33.7k| if (code_point < 0x800 || (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (87:11): [True: 25, False: 33.7k] | Branch (87:34): [True: 3.94k, False: 29.7k] | Branch (87:57): [True: 15, False: 3.92k] ------------------ 88| 40| return 0; 89| 40| } 90| | if constexpr (!match_system(big_endian)) { 91| | code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 92| | } 93| 33.7k| *utf16_output++ = char16_t(code_point); 94| 33.7k| pos += 3; 95| 33.7k| } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (95:16): [True: 2.65k, False: 353] ------------------ 96| | // we have a 4-byte UTF-8 word. 97| 2.65k| if (pos + 3 >= len) { ------------------ | Branch (97:11): [True: 38, False: 2.61k] ------------------ 98| 38| return 0; 99| 38| } // minimal bound checking 100| 2.61k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (100:11): [True: 31, False: 2.58k] ------------------ 101| 31| return 0; 102| 31| } 103| 2.58k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (103:11): [True: 27, False: 2.55k] ------------------ 104| 27| return 0; 105| 27| } 106| 2.55k| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (106:11): [True: 24, False: 2.53k] ------------------ 107| 24| return 0; 108| 24| } 109| | 110| | // range check 111| 2.53k| uint32_t code_point = (leading_byte & 0b00000111) << 18 | 112| 2.53k| (data[pos + 1] & 0b00111111) << 12 | 113| 2.53k| (data[pos + 2] & 0b00111111) << 6 | 114| 2.53k| (data[pos + 3] & 0b00111111); 115| 2.53k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (115:11): [True: 16, False: 2.51k] | Branch (115:35): [True: 19, False: 2.50k] ------------------ 116| 35| return 0; 117| 35| } 118| 2.50k| code_point -= 0x10000; 119| 2.50k| uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10)); 120| 2.50k| uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF)); 121| | if constexpr (!match_system(big_endian)) { 122| | high_surrogate = u16_swap_bytes(high_surrogate); 123| | low_surrogate = u16_swap_bytes(low_surrogate); 124| | } 125| 2.50k| *utf16_output++ = char16_t(high_surrogate); 126| 2.50k| *utf16_output++ = char16_t(low_surrogate); 127| 2.50k| pos += 4; 128| 2.50k| } else { 129| 353| return 0; 130| 353| } 131| 64.8k| } 132| 12.9k| return utf16_output - start; 133| 13.7k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf8_to_utf167convertILNS_10endiannessE1EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 14| 13.7k| char16_t *utf16_output) { 15| 13.7k| size_t pos = 0; 16| 13.7k| char16_t *start{utf16_output}; 17| 77.8k| while (pos < len) { ------------------ | Branch (17:10): [True: 64.8k, False: 12.9k] ------------------ 18| |#if SIMDUTF_CPLUSPLUS23 19| | if !consteval 20| |#endif 21| | // try to convert the next block of 16 ASCII bytes 22| 64.8k| { 23| 64.8k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (23:11): [True: 50.2k, False: 14.5k] ------------------ 24| | // they are ascii 25| 50.2k| uint64_t v1; 26| 50.2k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 27| 50.2k| uint64_t v2; 28| 50.2k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 29| 50.2k| uint64_t v{v1 | v2}; 30| 50.2k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (30:13): [True: 4.44k, False: 45.8k] ------------------ 31| 4.44k| size_t final_pos = pos + 16; 32| 75.4k| while (pos < final_pos) { ------------------ | Branch (32:18): [True: 71.0k, False: 4.44k] ------------------ 33| 71.0k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (33:31): [True: 71.0k, Folded] ------------------ 34| 71.0k| ? char16_t(u16_swap_bytes(data[pos])) 35| 71.0k| : char16_t(data[pos]); 36| 71.0k| pos++; 37| 71.0k| } 38| 4.44k| continue; 39| 4.44k| } 40| 50.2k| } 41| 64.8k| } 42| | 43| 60.3k| uint8_t leading_byte = data[pos]; // leading byte 44| 60.3k| if (leading_byte < 0b10000000) { ------------------ | Branch (44:9): [True: 23.6k, False: 36.6k] ------------------ 45| | // converting one ASCII byte !!! 46| 23.6k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (46:25): [True: 23.6k, Folded] ------------------ 47| 23.6k| ? char16_t(u16_swap_bytes(leading_byte)) 48| 23.6k| : char16_t(leading_byte); 49| 23.6k| pos++; 50| 36.6k| } else if ((leading_byte & 0b11100000) == 0b11000000) { ------------------ | Branch (50:16): [True: 3.74k, False: 32.9k] ------------------ 51| | // We have a two-byte UTF-8, it should become 52| | // a single UTF-16 word. 53| 3.74k| if (pos + 1 >= len) { ------------------ | Branch (53:11): [True: 45, False: 3.70k] ------------------ 54| 45| return 0; 55| 45| } // minimal bound checking 56| 3.70k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (56:11): [True: 83, False: 3.61k] ------------------ 57| 83| return 0; 58| 83| } 59| | // range check 60| 3.61k| uint32_t code_point = 61| 3.61k| (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 62| 3.61k| if (code_point < 0x80) { ------------------ | Branch (62:11): [True: 21, False: 3.59k] ------------------ 63| 21| return 0; 64| 21| } 65| 3.59k| if constexpr (!match_system(big_endian)) { 66| 3.59k| code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 67| 3.59k| } 68| 3.59k| *utf16_output++ = char16_t(code_point); 69| 3.59k| pos += 2; 70| 32.9k| } else if ((leading_byte & 0b11110000) == 0b11100000) { ------------------ | Branch (70:16): [True: 30.2k, False: 2.69k] ------------------ 71| | // We have a three-byte UTF-8, it should become 72| | // a single UTF-16 word. 73| 30.2k| if (pos + 2 >= len) { ------------------ | Branch (73:11): [True: 57, False: 30.2k] ------------------ 74| 57| return 0; 75| 57| } // minimal bound checking 76| | 77| 30.2k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (77:11): [True: 34, False: 30.1k] ------------------ 78| 34| return 0; 79| 34| } 80| 30.1k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (80:11): [True: 22, False: 30.1k] ------------------ 81| 22| return 0; 82| 22| } 83| | // range check 84| 30.1k| uint32_t code_point = (leading_byte & 0b00001111) << 12 | 85| 30.1k| (data[pos + 1] & 0b00111111) << 6 | 86| 30.1k| (data[pos + 2] & 0b00111111); 87| 30.1k| if (code_point < 0x800 || (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (87:11): [True: 25, False: 30.1k] | Branch (87:34): [True: 3.82k, False: 26.2k] | Branch (87:57): [True: 15, False: 3.81k] ------------------ 88| 40| return 0; 89| 40| } 90| 30.1k| if constexpr (!match_system(big_endian)) { 91| 30.1k| code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 92| 30.1k| } 93| 30.1k| *utf16_output++ = char16_t(code_point); 94| 30.1k| pos += 3; 95| 30.1k| } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (95:16): [True: 2.33k, False: 353] ------------------ 96| | // we have a 4-byte UTF-8 word. 97| 2.33k| if (pos + 3 >= len) { ------------------ | Branch (97:11): [True: 38, False: 2.30k] ------------------ 98| 38| return 0; 99| 38| } // minimal bound checking 100| 2.30k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (100:11): [True: 31, False: 2.26k] ------------------ 101| 31| return 0; 102| 31| } 103| 2.26k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (103:11): [True: 27, False: 2.24k] ------------------ 104| 27| return 0; 105| 27| } 106| 2.24k| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (106:11): [True: 24, False: 2.21k] ------------------ 107| 24| return 0; 108| 24| } 109| | 110| | // range check 111| 2.21k| uint32_t code_point = (leading_byte & 0b00000111) << 18 | 112| 2.21k| (data[pos + 1] & 0b00111111) << 12 | 113| 2.21k| (data[pos + 2] & 0b00111111) << 6 | 114| 2.21k| (data[pos + 3] & 0b00111111); 115| 2.21k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (115:11): [True: 16, False: 2.20k] | Branch (115:35): [True: 19, False: 2.18k] ------------------ 116| 35| return 0; 117| 35| } 118| 2.18k| code_point -= 0x10000; 119| 2.18k| uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10)); 120| 2.18k| uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF)); 121| 2.18k| if constexpr (!match_system(big_endian)) { 122| 2.18k| high_surrogate = u16_swap_bytes(high_surrogate); 123| 2.18k| low_surrogate = u16_swap_bytes(low_surrogate); 124| 2.18k| } 125| 2.18k| *utf16_output++ = char16_t(high_surrogate); 126| 2.18k| *utf16_output++ = char16_t(low_surrogate); 127| 2.18k| pos += 4; 128| 2.18k| } else { 129| 353| return 0; 130| 353| } 131| 60.3k| } 132| 12.9k| return utf16_output - start; 133| 13.7k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf8_to_utf327convertIPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_mPDi: 14| 7.76k| char32_t *utf32_output) { 15| 7.76k| size_t pos = 0; 16| 7.76k| char32_t *start{utf32_output}; 17| 40.0k| while (pos < len) { ------------------ | Branch (17:10): [True: 33.0k, False: 6.95k] ------------------ 18| |#if SIMDUTF_CPLUSPLUS23 19| | if !consteval 20| |#endif 21| 33.0k| { 22| | // try to convert the next block of 16 ASCII bytes 23| 33.0k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (23:11): [True: 21.4k, False: 11.6k] ------------------ 24| | // they are ascii 25| 21.4k| uint64_t v1; 26| 21.4k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 27| 21.4k| uint64_t v2; 28| 21.4k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 29| 21.4k| uint64_t v{v1 | v2}; 30| 21.4k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (30:13): [True: 4.26k, False: 17.1k] ------------------ 31| 4.26k| size_t final_pos = pos + 16; 32| 72.4k| while (pos < final_pos) { ------------------ | Branch (32:18): [True: 68.1k, False: 4.26k] ------------------ 33| 68.1k| *utf32_output++ = uint8_t(data[pos]); 34| 68.1k| pos++; 35| 68.1k| } 36| 4.26k| continue; 37| 4.26k| } 38| 21.4k| } 39| 33.0k| } 40| 28.8k| auto leading_byte = uint8_t(data[pos]); // leading byte 41| 28.8k| if (leading_byte < 0b10000000) { ------------------ | Branch (41:9): [True: 21.5k, False: 7.22k] ------------------ 42| | // converting one ASCII byte !!! 43| 21.5k| *utf32_output++ = char32_t(leading_byte); 44| 21.5k| pos++; 45| 21.5k| } else if ((leading_byte & 0b11100000) == 0b11000000) { ------------------ | Branch (45:16): [True: 2.83k, False: 4.38k] ------------------ 46| | // We have a two-byte UTF-8 47| 2.83k| if (pos + 1 >= len) { ------------------ | Branch (47:11): [True: 45, False: 2.79k] ------------------ 48| 45| return 0; 49| 45| } // minimal bound checking 50| 2.79k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (50:11): [True: 83, False: 2.71k] ------------------ 51| 83| return 0; 52| 83| } 53| | // range check 54| 2.71k| uint32_t code_point = (leading_byte & 0b00011111) << 6 | 55| 2.71k| (uint8_t(data[pos + 1]) & 0b00111111); 56| 2.71k| if (code_point < 0x80) { ------------------ | Branch (56:11): [True: 21, False: 2.69k] ------------------ 57| 21| return 0; 58| 21| } 59| 2.69k| *utf32_output++ = char32_t(code_point); 60| 2.69k| pos += 2; 61| 4.38k| } else if ((leading_byte & 0b11110000) == 0b11100000) { ------------------ | Branch (61:16): [True: 1.92k, False: 2.46k] ------------------ 62| | // We have a three-byte UTF-8 63| 1.92k| if (pos + 2 >= len) { ------------------ | Branch (63:11): [True: 57, False: 1.86k] ------------------ 64| 57| return 0; 65| 57| } // minimal bound checking 66| | 67| 1.86k| if ((uint8_t(data[pos + 1]) & 0b11000000) != 0b10000000) { ------------------ | Branch (67:11): [True: 34, False: 1.83k] ------------------ 68| 34| return 0; 69| 34| } 70| 1.83k| if ((uint8_t(data[pos + 2]) & 0b11000000) != 0b10000000) { ------------------ | Branch (70:11): [True: 22, False: 1.81k] ------------------ 71| 22| return 0; 72| 22| } 73| | // range check 74| 1.81k| uint32_t code_point = (leading_byte & 0b00001111) << 12 | 75| 1.81k| (uint8_t(data[pos + 1]) & 0b00111111) << 6 | 76| 1.81k| (uint8_t(data[pos + 2]) & 0b00111111); 77| 1.81k| if (code_point < 0x800 || (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (77:11): [True: 25, False: 1.78k] | Branch (77:34): [True: 879, False: 907] | Branch (77:57): [True: 15, False: 864] ------------------ 78| 40| return 0; 79| 40| } 80| 1.77k| *utf32_output++ = char32_t(code_point); 81| 1.77k| pos += 3; 82| 2.46k| } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (82:16): [True: 2.11k, False: 353] ------------------ 83| | // we have a 4-byte UTF-8 word. 84| 2.11k| if (pos + 3 >= len) { ------------------ | Branch (84:11): [True: 38, False: 2.07k] ------------------ 85| 38| return 0; 86| 38| } // minimal bound checking 87| 2.07k| if ((uint8_t(data[pos + 1]) & 0b11000000) != 0b10000000) { ------------------ | Branch (87:11): [True: 31, False: 2.04k] ------------------ 88| 31| return 0; 89| 31| } 90| 2.04k| if ((uint8_t(data[pos + 2]) & 0b11000000) != 0b10000000) { ------------------ | Branch (90:11): [True: 27, False: 2.01k] ------------------ 91| 27| return 0; 92| 27| } 93| 2.01k| if ((uint8_t(data[pos + 3]) & 0b11000000) != 0b10000000) { ------------------ | Branch (93:11): [True: 24, False: 1.99k] ------------------ 94| 24| return 0; 95| 24| } 96| | 97| | // range check 98| 1.99k| uint32_t code_point = (leading_byte & 0b00000111) << 18 | 99| 1.99k| (uint8_t(data[pos + 1]) & 0b00111111) << 12 | 100| 1.99k| (uint8_t(data[pos + 2]) & 0b00111111) << 6 | 101| 1.99k| (uint8_t(data[pos + 3]) & 0b00111111); 102| 1.99k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (102:11): [True: 16, False: 1.97k] | Branch (102:35): [True: 19, False: 1.95k] ------------------ 103| 35| return 0; 104| 35| } 105| 1.95k| *utf32_output++ = char32_t(code_point); 106| 1.95k| pos += 4; 107| 1.95k| } else { 108| 353| return 0; 109| 353| } 110| 28.8k| } 111| 6.95k| return utf32_output - start; 112| 7.76k|} _ZNK7simdutf8fallback14implementation13validate_utf8EPKcm: 37| 22.1k|implementation::validate_utf8(const char *buf, size_t len) const noexcept { 38| 22.1k| return scalar::utf8::validate(buf, len); 39| 22.1k|} _ZNK7simdutf8fallback14implementation25validate_utf8_with_errorsEPKcm: 44| 5.56k| const char *buf, size_t len) const noexcept { 45| 5.56k| return scalar::utf8::validate_with_errors(buf, len); 46| 5.56k|} _ZNK7simdutf8fallback14implementation16validate_utf16leEPKDsm: 77| 16.4k| size_t len) const noexcept { 78| 16.4k| return scalar::utf16::validate(buf, len); 79| 16.4k|} _ZNK7simdutf8fallback14implementation16validate_utf16beEPKDsm: 85| 16.4k| size_t len) const noexcept { 86| 16.4k| return scalar::utf16::validate(buf, len); 87| 16.4k|} _ZNK7simdutf8fallback14implementation28validate_utf16le_with_errorsEPKDsm: 90| 5.56k| const char16_t *buf, size_t len) const noexcept { 91| 5.56k| return scalar::utf16::validate_with_errors(buf, len); 92| 5.56k|} _ZNK7simdutf8fallback14implementation28validate_utf16be_with_errorsEPKDsm: 95| 5.56k| const char16_t *buf, size_t len) const noexcept { 96| 5.56k| return scalar::utf16::validate_with_errors(buf, len); 97| 5.56k|} _ZNK7simdutf8fallback14implementation14validate_utf32EPKDim: 114| 10.9k|implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept { 115| 10.9k| return scalar::utf32::validate(buf, len); 116| 10.9k|} _ZNK7simdutf8fallback14implementation22convert_latin1_to_utf8EPKcmPc: 128| 5.73k| const char *buf, size_t len, char *utf8_output) const noexcept { 129| 5.73k| return scalar::latin1_to_utf8::convert(buf, len, utf8_output); 130| 5.73k|} _ZNK7simdutf8fallback14implementation25convert_latin1_to_utf16leEPKcmPDs: 135| 5.56k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 136| 5.56k| return scalar::latin1_to_utf16::convert(buf, len, 137| 5.56k| utf16_output); 138| 5.56k|} _ZNK7simdutf8fallback14implementation25convert_latin1_to_utf16beEPKcmPDs: 141| 5.56k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 142| 5.56k| return scalar::latin1_to_utf16::convert(buf, len, 143| 5.56k| utf16_output); 144| 5.56k|} _ZNK7simdutf8fallback14implementation23convert_latin1_to_utf32EPKcmPDi: 149| 5.56k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 150| 5.56k| return scalar::latin1_to_utf32::convert(buf, len, utf32_output); 151| 5.56k|} _ZNK7simdutf8fallback14implementation22convert_utf8_to_latin1EPKcmPc: 156| 11.1k| const char *buf, size_t len, char *latin1_output) const noexcept { 157| 11.1k| return scalar::utf8_to_latin1::convert(buf, len, latin1_output); 158| 11.1k|} _ZNK7simdutf8fallback14implementation23convert_utf8_to_utf16leEPKcmPDs: 173| 11.0k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 174| 11.0k| return scalar::utf8_to_utf16::convert(buf, len, 175| 11.0k| utf16_output); 176| 11.0k|} _ZNK7simdutf8fallback14implementation23convert_utf8_to_utf16beEPKcmPDs: 179| 11.0k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 180| 11.0k| return scalar::utf8_to_utf16::convert(buf, len, 181| 11.0k| utf16_output); 182| 11.0k|} _ZNK7simdutf8fallback14implementation21convert_utf8_to_utf32EPKcmPDi: 211| 5.56k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 212| 5.56k| return scalar::utf8_to_utf32::convert(buf, len, utf32_output); 213| 5.56k|} _ZNK7simdutf8fallback14implementation25convert_utf16le_to_latin1EPKDsmPc: 228| 5.56k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 229| 5.56k| return scalar::utf16_to_latin1::convert(buf, len, 230| 5.56k| latin1_output); 231| 5.56k|} _ZNK7simdutf8fallback14implementation25convert_utf16be_to_latin1EPKDsmPc: 234| 5.56k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 235| 5.56k| return scalar::utf16_to_latin1::convert(buf, len, 236| 5.56k| latin1_output); 237| 5.56k|} _ZNK7simdutf8fallback14implementation23convert_utf16le_to_utf8EPKDsmPc: 268| 10.9k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 269| 10.9k| return scalar::utf16_to_utf8::convert(buf, len, 270| 10.9k| utf8_output); 271| 10.9k|} _ZNK7simdutf8fallback14implementation23convert_utf16be_to_utf8EPKDsmPc: 274| 10.9k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 275| 10.9k| return scalar::utf16_to_utf8::convert(buf, len, utf8_output); 276| 10.9k|} _ZNK7simdutf8fallback14implementation23convert_utf32_to_latin1EPKDimPc: 305| 5.56k| const char32_t *buf, size_t len, char *latin1_output) const noexcept { 306| 5.56k| return scalar::utf32_to_latin1::convert(buf, len, latin1_output); 307| 5.56k|} _ZNK7simdutf8fallback14implementation21convert_utf32_to_utf8EPKDimPc: 322| 5.35k| const char32_t *buf, size_t len, char *utf8_output) const noexcept { 323| 5.35k| return scalar::utf32_to_utf8::convert(buf, len, utf8_output); 324| 5.35k|} _ZNK7simdutf8fallback14implementation23latin1_length_from_utf8EPKcm: 438| 11.1k| const char *buf, size_t len) const noexcept { 439| 11.1k| return scalar::utf8::count_code_points(buf, len); 440| 11.1k|} _ZNK7simdutf8fallback14implementation23utf8_length_from_latin1EPKcm: 445| 5.73k| const char *input, size_t length) const noexcept { 446| 5.73k| return scalar::latin1_to_utf8::utf8_length_from_latin1(input, length); 447| 5.73k|} _ZNK7simdutf8fallback14implementation24utf8_length_from_utf16leEPKDsm: 452| 10.9k| const char16_t *input, size_t length) const noexcept { 453| 10.9k| return scalar::utf16::utf8_length_from_utf16(input, 454| 10.9k| length); 455| 10.9k|} _ZNK7simdutf8fallback14implementation24utf8_length_from_utf16beEPKDsm: 458| 10.9k| const char16_t *input, size_t length) const noexcept { 459| 10.9k| return scalar::utf16::utf8_length_from_utf16(input, length); 460| 10.9k|} _ZNK7simdutf8fallback14implementation22utf16_length_from_utf8EPKcm: 478| 22.1k| const char *input, size_t length) const noexcept { 479| 22.1k| return scalar::utf8::utf16_length_from_utf8(input, length); 480| 22.1k|} _ZNK7simdutf8fallback14implementation22utf8_length_from_utf32EPKDim: 513| 5.35k| const char32_t *input, size_t length) const noexcept { 514| 5.35k| return scalar::utf32::utf8_length_from_utf32(input, length); 515| 5.35k|} _ZNK7simdutf8fallback14implementation22utf32_length_from_utf8EPKcm: 527| 5.56k| const char *input, size_t length) const noexcept { 528| 5.56k| return scalar::utf8::count_code_points(input, length); 529| 5.56k|} _ZNK7simdutf8fallback14implementation16base64_to_binaryEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 536| 16.4k| last_chunk_handling_options last_chunk_options) const noexcept { 537| 16.4k| return simdutf::scalar::base64::base64_to_binary_details_impl( 538| 16.4k| input, length, output, options, last_chunk_options); 539| 16.4k|} _ZNK7simdutf8fallback14implementation16binary_to_base64EPKcmPcNS_14base64_optionsE: 564| 16.3k| base64_options options) const noexcept { 565| 16.3k| return scalar::base64::tail_encode_base64(output, input, length, options); 566| 16.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_16base6422compress_decode_base64ILb0ELb0ELb0EcEENS_11full_resultEPcPKT2_mNS_14base64_optionsENS_27last_chunk_handling_optionsE: 45| 64.5k| last_chunk_handling_options last_chunk_options) { 46| 64.5k| const uint8_t *to_base64 = 47| 64.5k| default_or_url ? tables::base64::to_base64_default_or_url_value ------------------ | Branch (47:7): [Folded, False: 64.5k] ------------------ 48| 64.5k| : (base64_url ? tables::base64::to_base64_url_value ------------------ | Branch (48:25): [Folded, False: 64.5k] ------------------ 49| 64.5k| : tables::base64::to_base64_value); 50| 64.5k| auto ri = simdutf::scalar::base64::find_end(src, srclen, options); 51| 64.5k| size_t equallocation = ri.equallocation; 52| 64.5k| size_t equalsigns = ri.equalsigns; 53| 64.5k| srclen = ri.srclen; 54| 64.5k| size_t full_input_length = ri.full_input_length; 55| 64.5k| if (srclen == 0) { ------------------ | Branch (55:7): [True: 57.7k, False: 6.74k] ------------------ 56| 57.7k| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (56:9): [True: 57.7k, Folded] | Branch (56:28): [True: 20, False: 57.7k] ------------------ 57| 20| return {INVALID_BASE64_CHARACTER, equallocation, 0, true}; 58| 20| } 59| 57.7k| return {SUCCESS, full_input_length, 0}; 60| 57.7k| } 61| 6.74k| char *end_of_safe_64byte_zone = 62| 6.74k| dst == nullptr ------------------ | Branch (62:7): [True: 85, False: 6.66k] ------------------ 63| 6.74k| ? nullptr 64| 6.74k| : ((srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 ------------------ | Branch (64:14): [True: 2.38k, False: 4.27k] ------------------ 65| 6.66k| : dst); 66| | 67| 6.74k| const chartype *const srcinit = src; 68| 6.74k| const char *const dstinit = dst; 69| 6.74k| const chartype *const srcend = src + srclen; 70| | 71| 6.74k| constexpr size_t block_size = 6; 72| 6.74k| static_assert(block_size >= 2, "block_size must be at least two"); 73| 6.74k| char buffer[block_size * 64]; 74| 6.74k| char *bufferptr = buffer; 75| 6.74k| if (srclen >= 64) { ------------------ | Branch (75:7): [True: 3.36k, False: 3.38k] ------------------ 76| 3.36k| const chartype *const srcend64 = src + srclen - 64; 77| 19.7k| while (src <= srcend64) { ------------------ | Branch (77:12): [True: 17.2k, False: 2.48k] ------------------ 78| 17.2k| block64 b(src); 79| 17.2k| src += 64; 80| 17.2k| uint64_t error = 0; 81| 17.2k| const uint64_t badcharmask = 82| 17.2k| b.to_base64_mask(&error); 83| 17.2k| if (!ignore_garbage && error) { ------------------ | Branch (83:11): [True: 17.2k, Folded] | Branch (83:30): [True: 885, False: 16.3k] ------------------ 84| 885| src -= 64; 85| 885| const size_t error_offset = trailing_zeroes(error); 86| 885| return {error_code::INVALID_BASE64_CHARACTER, 87| 885| size_t(src - srcinit + error_offset), size_t(dst - dstinit)}; 88| 885| } 89| 16.3k| if (badcharmask != 0) { ------------------ | Branch (89:11): [True: 2.32k, False: 14.0k] ------------------ 90| 2.32k| bufferptr += b.compress_block(badcharmask, bufferptr); 91| 14.0k| } else if (bufferptr != buffer) { ------------------ | Branch (91:18): [True: 326, False: 13.7k] ------------------ 92| 326| b.copy_block(bufferptr); 93| 326| bufferptr += 64; 94| 13.7k| } else { 95| 13.7k| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (95:13): [True: 836, False: 12.8k] ------------------ 96| 836| b.base64_decode_block_safe(dst); 97| 12.8k| } else { 98| 12.8k| b.base64_decode_block(dst); 99| 12.8k| } 100| 13.7k| dst += 48; 101| 13.7k| } 102| 16.3k| if (bufferptr >= (block_size - 1) * 64 + buffer) { ------------------ | Branch (102:11): [True: 233, False: 16.1k] ------------------ 103| 1.16k| for (size_t i = 0; i < (block_size - 2); i++) { ------------------ | Branch (103:28): [True: 932, False: 233] ------------------ 104| 932| base64_decode_block(dst, buffer + i * 64); 105| 932| dst += 48; 106| 932| } 107| 233| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (107:13): [True: 0, False: 233] ------------------ 108| 0| base64_decode_block_safe(dst, buffer + (block_size - 2) * 64); 109| 233| } else { 110| 233| base64_decode_block(dst, buffer + (block_size - 2) * 64); 111| 233| } 112| 233| dst += 48; 113| 233| std::memcpy(buffer, buffer + (block_size - 1) * 64, 114| 233| 64); // 64 might be too much 115| 233| bufferptr -= (block_size - 1) * 64; 116| 233| } 117| 16.3k| } 118| 3.36k| } 119| | 120| 5.86k| char *buffer_start = buffer; 121| | // Optimization note: if this is almost full, then it is worth our 122| | // time, otherwise, we should just decode directly. 123| 5.86k| int last_block = (int)((bufferptr - buffer_start) % 64); 124| 5.86k| if (last_block != 0 && srcend - src + last_block >= 64) { ------------------ | Branch (124:7): [True: 710, False: 5.15k] | Branch (124:26): [True: 125, False: 585] ------------------ 125| | 126| 1.56k| while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) { ------------------ | Branch (126:12): [True: 1.48k, False: 84] | Branch (126:52): [True: 1.45k, False: 29] ------------------ 127| 1.45k| uint8_t val = to_base64[uint8_t(*src)]; 128| 1.45k| *bufferptr = char(val); 129| 1.45k| if (!ignore_garbage && ------------------ | Branch (129:11): [True: 1.45k, Folded] ------------------ 130| 1.45k| (!scalar::base64::is_eight_byte(*src) || val > 64)) { ------------------ | Branch (130:12): [True: 0, False: 1.45k] | Branch (130:52): [True: 12, False: 1.44k] ------------------ 131| 12| return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit), 132| 12| size_t(dst - dstinit)}; 133| 12| } 134| 1.44k| bufferptr += (val <= 63); 135| 1.44k| src++; 136| 1.44k| } 137| 125| } 138| | 139| 6.37k| for (; buffer_start + 64 <= bufferptr; buffer_start += 64) { ------------------ | Branch (139:10): [True: 525, False: 5.85k] ------------------ 140| 525| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (140:9): [True: 50, False: 475] ------------------ 141| 50| base64_decode_block_safe(dst, buffer_start); 142| 475| } else { 143| 475| base64_decode_block(dst, buffer_start); 144| 475| } 145| 525| dst += 48; 146| 525| } 147| 5.85k| if ((bufferptr - buffer_start) % 64 != 0) { ------------------ | Branch (147:7): [True: 614, False: 5.23k] ------------------ 148| 5.81k| while (buffer_start + 4 < bufferptr) { ------------------ | Branch (148:12): [True: 5.20k, False: 614] ------------------ 149| 5.20k| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 150| 5.20k| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 151| 5.20k| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 152| 5.20k| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 153| 5.20k| << 8; 154| 5.20k|#if !SIMDUTF_IS_BIG_ENDIAN 155| 5.20k| triple = scalar::u32_swap_bytes(triple); 156| 5.20k|#endif 157| 5.20k| std::memcpy(dst, &triple, 3); 158| | 159| 5.20k| dst += 3; 160| 5.20k| buffer_start += 4; 161| 5.20k| } 162| 614| if (buffer_start + 4 <= bufferptr) { ------------------ | Branch (162:9): [True: 150, False: 464] ------------------ 163| 150| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 164| 150| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 165| 150| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 166| 150| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 167| 150| << 8; 168| 150|#if !SIMDUTF_IS_BIG_ENDIAN 169| 150| triple = scalar::u32_swap_bytes(triple); 170| 150|#endif 171| 150| std::memcpy(dst, &triple, 3); 172| | 173| 150| dst += 3; 174| 150| buffer_start += 4; 175| 150| } 176| | // we may have 1, 2 or 3 bytes left and we need to decode them so let us 177| | // backtrack 178| 614| int leftover = int(bufferptr - buffer_start); 179| 1.61k| while (leftover > 0) { ------------------ | Branch (179:12): [True: 1.00k, False: 614] ------------------ 180| 1.00k| if (!ignore_garbage) { ------------------ | Branch (180:11): [True: 1.00k, Folded] ------------------ 181| 8.63k| while (to_base64[uint8_t(*(src - 1))] == 64) { ------------------ | Branch (181:16): [True: 7.63k, False: 1.00k] ------------------ 182| 7.63k| src--; 183| 7.63k| } 184| 1.00k| } else { 185| 0| while (to_base64[uint8_t(*(src - 1))] >= 64) { ------------------ | Branch (185:16): [True: 0, False: 0] ------------------ 186| 0| src--; 187| 0| } 188| 0| } 189| 1.00k| src--; 190| 1.00k| leftover--; 191| 1.00k| } 192| 614| } 193| 5.85k| if (src < srcend + equalsigns) { ------------------ | Branch (193:7): [True: 5.44k, False: 403] ------------------ 194| 5.44k| full_result r = scalar::base64::base64_tail_decode( 195| 5.44k| dst, src, srcend - src, equalsigns, options, last_chunk_options); 196| 5.44k| r = scalar::base64::patch_tail_result( 197| 5.44k| r, size_t(src - srcinit), size_t(dst - dstinit), equallocation, 198| 5.44k| full_input_length, last_chunk_options); 199| | // When is_partial(last_chunk_options) is true, we must either end with 200| | // the end of the stream (beyond whitespace) or right after a non-ignorable 201| | // character or at the very beginning of the stream. 202| | // See https://tc39.es/proposal-arraybuffer-base64/spec/#sec-frombase64 203| 5.44k| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (203:9): [True: 1.97k, False: 3.47k] | Branch (203:43): [True: 1.67k, False: 298] ------------------ 204| 1.67k| r.input_count < full_input_length) { ------------------ | Branch (204:9): [True: 317, False: 1.36k] ------------------ 205| | // First check if we can extend the input to the end of the stream 206| 2.25k| while (r.input_count < full_input_length && ------------------ | Branch (206:14): [True: 2.21k, False: 42] ------------------ 207| 2.21k| base64_ignorable(*(srcinit + r.input_count), options)) { ------------------ | Branch (207:14): [True: 1.94k, False: 275] ------------------ 208| 1.94k| r.input_count++; 209| 1.94k| } 210| | // If we are still not at the end of the stream, then we must backtrack 211| | // to the last non-ignorable character. 212| 317| if (r.input_count < full_input_length) { ------------------ | Branch (212:11): [True: 275, False: 42] ------------------ 213| 1.35k| while (r.input_count > 0 && ------------------ | Branch (213:16): [True: 1.26k, False: 90] ------------------ 214| 1.26k| base64_ignorable(*(srcinit + r.input_count - 1), options)) { ------------------ | Branch (214:16): [True: 1.08k, False: 185] ------------------ 215| 1.08k| r.input_count--; 216| 1.08k| } 217| 275| } 218| 317| } 219| 5.44k| return r; 220| 5.44k| } 221| 403| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (221:7): [True: 403, Folded] | Branch (221:26): [True: 0, False: 403] ------------------ 222| 0| if ((size_t(dst - dstinit) % 3 == 0) || ------------------ | Branch (222:9): [True: 0, False: 0] ------------------ 223| 0| ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) { ------------------ | Branch (223:9): [True: 0, False: 0] ------------------ 224| 0| return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit), 225| 0| true}; 226| 0| } 227| 0| } 228| 403| return {SUCCESS, srclen, size_t(dst - dstinit)}; 229| 403|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_16base6422compress_decode_base64ILb0ELb0ELb0EcEENS_11full_resultEPcPKT2_mNS_14base64_optionsENS_27last_chunk_handling_optionsE: 45| 16.1k| last_chunk_handling_options last_chunk_options) { 46| 16.1k| const uint8_t *to_base64 = 47| 16.1k| default_or_url ? tables::base64::to_base64_default_or_url_value ------------------ | Branch (47:7): [Folded, False: 16.1k] ------------------ 48| 16.1k| : (base64_url ? tables::base64::to_base64_url_value ------------------ | Branch (48:25): [Folded, False: 16.1k] ------------------ 49| 16.1k| : tables::base64::to_base64_value); 50| 16.1k| auto ri = simdutf::scalar::base64::find_end(src, srclen, options); 51| 16.1k| size_t equallocation = ri.equallocation; 52| 16.1k| size_t equalsigns = ri.equalsigns; 53| 16.1k| srclen = ri.srclen; 54| 16.1k| size_t full_input_length = ri.full_input_length; 55| 16.1k| if (srclen == 0) { ------------------ | Branch (55:7): [True: 14.6k, False: 1.55k] ------------------ 56| 14.6k| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (56:9): [True: 14.6k, Folded] | Branch (56:28): [True: 11, False: 14.5k] ------------------ 57| 11| return {INVALID_BASE64_CHARACTER, equallocation, 0, true}; 58| 11| } 59| 14.5k| return {SUCCESS, full_input_length, 0}; 60| 14.6k| } 61| 1.55k| char *end_of_safe_64byte_zone = 62| 1.55k| dst == nullptr ------------------ | Branch (62:7): [True: 32, False: 1.52k] ------------------ 63| 1.55k| ? nullptr 64| 1.55k| : ((srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 ------------------ | Branch (64:14): [True: 508, False: 1.01k] ------------------ 65| 1.52k| : dst); 66| | 67| 1.55k| const chartype *const srcinit = src; 68| 1.55k| const char *const dstinit = dst; 69| 1.55k| const chartype *const srcend = src + srclen; 70| | 71| 1.55k| constexpr size_t block_size = 6; 72| 1.55k| static_assert(block_size >= 2, "block_size must be at least two"); 73| 1.55k| char buffer[block_size * 64]; 74| 1.55k| char *bufferptr = buffer; 75| 1.55k| if (srclen >= 64) { ------------------ | Branch (75:7): [True: 809, False: 745] ------------------ 76| 809| const chartype *const srcend64 = src + srclen - 64; 77| 4.52k| while (src <= srcend64) { ------------------ | Branch (77:12): [True: 3.97k, False: 556] ------------------ 78| 3.97k| block64 b(src); 79| 3.97k| src += 64; 80| 3.97k| uint64_t error = 0; 81| 3.97k| const uint64_t badcharmask = 82| 3.97k| b.to_base64_mask(&error); 83| 3.97k| if (!ignore_garbage && error) { ------------------ | Branch (83:11): [True: 3.97k, Folded] | Branch (83:30): [True: 253, False: 3.71k] ------------------ 84| 253| src -= 64; 85| 253| const size_t error_offset = trailing_zeroes(error); 86| 253| return {error_code::INVALID_BASE64_CHARACTER, 87| 253| size_t(src - srcinit + error_offset), size_t(dst - dstinit)}; 88| 253| } 89| 3.71k| if (badcharmask != 0) { ------------------ | Branch (89:11): [True: 752, False: 2.96k] ------------------ 90| 752| bufferptr += b.compress_block(badcharmask, bufferptr); 91| 2.96k| } else if (bufferptr != buffer) { ------------------ | Branch (91:18): [True: 68, False: 2.89k] ------------------ 92| 68| b.copy_block(bufferptr); 93| 68| bufferptr += 64; 94| 2.89k| } else { 95| 2.89k| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (95:13): [True: 120, False: 2.77k] ------------------ 96| 120| b.base64_decode_block_safe(dst); 97| 2.77k| } else { 98| 2.77k| b.base64_decode_block(dst); 99| 2.77k| } 100| 2.89k| dst += 48; 101| 2.89k| } 102| 3.71k| if (bufferptr >= (block_size - 1) * 64 + buffer) { ------------------ | Branch (102:11): [True: 68, False: 3.65k] ------------------ 103| 340| for (size_t i = 0; i < (block_size - 2); i++) { ------------------ | Branch (103:28): [True: 272, False: 68] ------------------ 104| 272| base64_decode_block(dst, buffer + i * 64); 105| 272| dst += 48; 106| 272| } 107| 68| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (107:13): [True: 0, False: 68] ------------------ 108| 0| base64_decode_block_safe(dst, buffer + (block_size - 2) * 64); 109| 68| } else { 110| 68| base64_decode_block(dst, buffer + (block_size - 2) * 64); 111| 68| } 112| 68| dst += 48; 113| 68| std::memcpy(buffer, buffer + (block_size - 1) * 64, 114| 68| 64); // 64 might be too much 115| 68| bufferptr -= (block_size - 1) * 64; 116| 68| } 117| 3.71k| } 118| 809| } 119| | 120| 1.30k| char *buffer_start = buffer; 121| | // Optimization note: if this is almost full, then it is worth our 122| | // time, otherwise, we should just decode directly. 123| 1.30k| int last_block = (int)((bufferptr - buffer_start) % 64); 124| 1.30k| if (last_block != 0 && srcend - src + last_block >= 64) { ------------------ | Branch (124:7): [True: 238, False: 1.06k] | Branch (124:26): [True: 41, False: 197] ------------------ 125| | 126| 351| while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) { ------------------ | Branch (126:12): [True: 324, False: 27] | Branch (126:52): [True: 314, False: 10] ------------------ 127| 314| uint8_t val = to_base64[uint8_t(*src)]; 128| 314| *bufferptr = char(val); 129| 314| if (!ignore_garbage && ------------------ | Branch (129:11): [True: 314, Folded] ------------------ 130| 314| (!scalar::base64::is_eight_byte(*src) || val > 64)) { ------------------ | Branch (130:12): [True: 0, False: 314] | Branch (130:52): [True: 4, False: 310] ------------------ 131| 4| return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit), 132| 4| size_t(dst - dstinit)}; 133| 4| } 134| 310| bufferptr += (val <= 63); 135| 310| src++; 136| 310| } 137| 41| } 138| | 139| 1.45k| for (; buffer_start + 64 <= bufferptr; buffer_start += 64) { ------------------ | Branch (139:10): [True: 154, False: 1.29k] ------------------ 140| 154| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (140:9): [True: 21, False: 133] ------------------ 141| 21| base64_decode_block_safe(dst, buffer_start); 142| 133| } else { 143| 133| base64_decode_block(dst, buffer_start); 144| 133| } 145| 154| dst += 48; 146| 154| } 147| 1.29k| if ((bufferptr - buffer_start) % 64 != 0) { ------------------ | Branch (147:7): [True: 207, False: 1.09k] ------------------ 148| 1.84k| while (buffer_start + 4 < bufferptr) { ------------------ | Branch (148:12): [True: 1.63k, False: 207] ------------------ 149| 1.63k| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 150| 1.63k| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 151| 1.63k| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 152| 1.63k| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 153| 1.63k| << 8; 154| 1.63k|#if !SIMDUTF_IS_BIG_ENDIAN 155| 1.63k| triple = scalar::u32_swap_bytes(triple); 156| 1.63k|#endif 157| 1.63k| std::memcpy(dst, &triple, 3); 158| | 159| 1.63k| dst += 3; 160| 1.63k| buffer_start += 4; 161| 1.63k| } 162| 207| if (buffer_start + 4 <= bufferptr) { ------------------ | Branch (162:9): [True: 53, False: 154] ------------------ 163| 53| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 164| 53| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 165| 53| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 166| 53| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 167| 53| << 8; 168| 53|#if !SIMDUTF_IS_BIG_ENDIAN 169| 53| triple = scalar::u32_swap_bytes(triple); 170| 53|#endif 171| 53| std::memcpy(dst, &triple, 3); 172| | 173| 53| dst += 3; 174| 53| buffer_start += 4; 175| 53| } 176| | // we may have 1, 2 or 3 bytes left and we need to decode them so let us 177| | // backtrack 178| 207| int leftover = int(bufferptr - buffer_start); 179| 524| while (leftover > 0) { ------------------ | Branch (179:12): [True: 317, False: 207] ------------------ 180| 317| if (!ignore_garbage) { ------------------ | Branch (180:11): [True: 317, Folded] ------------------ 181| 2.74k| while (to_base64[uint8_t(*(src - 1))] == 64) { ------------------ | Branch (181:16): [True: 2.42k, False: 317] ------------------ 182| 2.42k| src--; 183| 2.42k| } 184| 317| } else { 185| 0| while (to_base64[uint8_t(*(src - 1))] >= 64) { ------------------ | Branch (185:16): [True: 0, False: 0] ------------------ 186| 0| src--; 187| 0| } 188| 0| } 189| 317| src--; 190| 317| leftover--; 191| 317| } 192| 207| } 193| 1.29k| if (src < srcend + equalsigns) { ------------------ | Branch (193:7): [True: 1.17k, False: 119] ------------------ 194| 1.17k| full_result r = scalar::base64::base64_tail_decode( 195| 1.17k| dst, src, srcend - src, equalsigns, options, last_chunk_options); 196| 1.17k| r = scalar::base64::patch_tail_result( 197| 1.17k| r, size_t(src - srcinit), size_t(dst - dstinit), equallocation, 198| 1.17k| full_input_length, last_chunk_options); 199| | // When is_partial(last_chunk_options) is true, we must either end with 200| | // the end of the stream (beyond whitespace) or right after a non-ignorable 201| | // character or at the very beginning of the stream. 202| | // See https://tc39.es/proposal-arraybuffer-base64/spec/#sec-frombase64 203| 1.17k| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (203:9): [True: 0, False: 1.17k] | Branch (203:43): [True: 0, False: 0] ------------------ 204| 0| r.input_count < full_input_length) { ------------------ | Branch (204:9): [True: 0, False: 0] ------------------ 205| | // First check if we can extend the input to the end of the stream 206| 0| while (r.input_count < full_input_length && ------------------ | Branch (206:14): [True: 0, False: 0] ------------------ 207| 0| base64_ignorable(*(srcinit + r.input_count), options)) { ------------------ | Branch (207:14): [True: 0, False: 0] ------------------ 208| 0| r.input_count++; 209| 0| } 210| | // If we are still not at the end of the stream, then we must backtrack 211| | // to the last non-ignorable character. 212| 0| if (r.input_count < full_input_length) { ------------------ | Branch (212:11): [True: 0, False: 0] ------------------ 213| 0| while (r.input_count > 0 && ------------------ | Branch (213:16): [True: 0, False: 0] ------------------ 214| 0| base64_ignorable(*(srcinit + r.input_count - 1), options)) { ------------------ | Branch (214:16): [True: 0, False: 0] ------------------ 215| 0| r.input_count--; 216| 0| } 217| 0| } 218| 0| } 219| 1.17k| return r; 220| 1.17k| } 221| 119| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (221:7): [True: 119, Folded] | Branch (221:26): [True: 0, False: 119] ------------------ 222| 0| if ((size_t(dst - dstinit) % 3 == 0) || ------------------ | Branch (222:9): [True: 0, False: 0] ------------------ 223| 0| ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) { ------------------ | Branch (223:9): [True: 0, False: 0] ------------------ 224| 0| return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit), 225| 0| true}; 226| 0| } 227| 0| } 228| 119| return {SUCCESS, srclen, size_t(dst - dstinit)}; 229| 119|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE13get_remainderEPh: 57| 27.3k|buf_block_reader::get_remainder(uint8_t *dst) const { 58| 27.3k| if (len == idx) { ------------------ | Branch (58:7): [True: 23.0k, False: 4.38k] ------------------ 59| 23.0k| return 0; 60| 23.0k| } // memcpy(dst, null, 0) will trigger an error with some sanitizers 61| 4.38k| std::memset(dst, 0x20, 62| 4.38k| STEP_SIZE); // std::memset STEP_SIZE because it is more efficient 63| | // to write out 8 or 16 bytes at once. 64| 4.38k| std::memcpy(dst, buf + idx, len - idx); 65| 4.38k| return len - idx; 66| 27.3k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE13get_remainderEPh: 57| 27.3k|buf_block_reader::get_remainder(uint8_t *dst) const { 58| 27.3k| if (len == idx) { ------------------ | Branch (58:7): [True: 24.0k, False: 3.35k] ------------------ 59| 24.0k| return 0; 60| 24.0k| } // memcpy(dst, null, 0) will trigger an error with some sanitizers 61| 3.35k| std::memset(dst, 0x20, 62| 3.35k| STEP_SIZE); // std::memset STEP_SIZE because it is more efficient 63| | // to write out 8 or 16 bytes at once. 64| 3.35k| std::memcpy(dst, buf + idx, len - idx); 65| 3.35k| return len - idx; 66| 27.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EEC2EPKhm: 36| 27.6k| : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, ------------------ | Branch (36:42): [True: 25.6k, False: 1.95k] ------------------ 37| 27.6k| idx{0} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE14has_full_blockEv: 45| 34.5k|simdutf_really_inline bool buf_block_reader::has_full_block() const { 46| 34.5k| return idx < lenminusstep; 47| 34.5k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE10full_blockEv: 51| 7.21k|buf_block_reader::full_block() const { 52| 7.21k| return &buf[idx]; 53| 7.21k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE7advanceEv: 69| 34.3k|simdutf_really_inline void buf_block_reader::advance() { 70| 34.3k| idx += STEP_SIZE; 71| 34.3k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EEC2EPKhm: 36| 27.6k| : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, ------------------ | Branch (36:42): [True: 25.7k, False: 1.87k] ------------------ 37| 27.6k| idx{0} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE14has_full_blockEv: 45| 37.2k|simdutf_really_inline bool buf_block_reader::has_full_block() const { 46| 37.2k| return idx < lenminusstep; 47| 37.2k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE10full_blockEv: 51| 9.90k|buf_block_reader::full_block() const { 52| 9.90k| return &buf[idx]; 53| 9.90k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE7advanceEv: 69| 37.0k|simdutf_really_inline void buf_block_reader::advance() { 70| 37.0k| idx += STEP_SIZE; 71| 37.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE0EEEmPKDsm: 10| 10.5k| size_t size) { 11| 10.5k| size_t pos = 0; 12| | 13| 10.5k| using vector_u16 = simd16; 14| 10.5k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 10.5k| const auto one = vector_u16::splat(1); 17| | 18| 10.5k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 10.5k| size_t count = size / N * N; 22| | 23| | // in a single iteration the increment is 0, 1 or 2, despite we have 24| | // three additions 25| 10.5k| constexpr size_t max_iterations = 65535 / 2; 26| 10.5k| size_t iteration = max_iterations; 27| | 28| 15.9k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 5.39k, False: 10.5k] ------------------ 29| 5.39k| auto input = vector_u16::load(reinterpret_cast(in + pos)); 30| | if constexpr (!match_system(big_endian)) { 31| | input = input.swap_bytes(); 32| | } 33| | // 0xd800 .. 0xdbff - low surrogate 34| | // 0xdc00 .. 0xdfff - high surrogate 35| 5.39k| const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800)); 36| | 37| | // c0 - chars that yield 2- or 3-byte UTF-8 codes 38| 5.39k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 5.39k| const auto c1 = min(input & uint16_t(0xf800), one); 42| | 43| | /* 44| | Explanation how the counting works. 45| | 46| | In the case of a non-surrogate character we count: 47| | * always 1 -- see how `count` is initialized above; 48| | * c0 = 1 if the current char yields 2 or 3 bytes; 49| | * c1 = 1 if the current char yields 3 bytes. 50| | 51| | Thus, we always have correct count for the current char: 52| | from 1, 2 or 3 bytes. 53| | 54| | A trickier part is how we count surrogate pairs. Whether 55| | we encounter a surrogate (low or high), we count it as 56| | 3 chars and then minus 1 (`is_surrogate` is -1 or 0). 57| | Each surrogate char yields 2. A surrogate pair, that 58| | is a low surrogate followed by a high one, yields 59| | the expected 4 bytes. 60| | 61| | It also correctly handles cases when low surrogate is 62| | processed by the this loop, but high surrogate is counted 63| | by the scalar procedure. The scalar procedure uses exactly 64| | the described approach, thanks to that for valid UTF-16 65| | strings it always count correctly. 66| | */ 67| 5.39k| v_count += c0; 68| 5.39k| v_count += c1; 69| 5.39k| v_count += vector_u16(is_surrogate); 70| | 71| 5.39k| iteration -= 1; 72| 5.39k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 5.39k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 5.39k| } 78| | 79| 10.5k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 10.5k, False: 0] ------------------ 80| 10.5k| count += v_count.sum(); 81| 10.5k| } 82| | 83| 10.5k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 10.5k| size - pos); 85| 10.5k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE1EEEmPKDsm: 10| 10.5k| size_t size) { 11| 10.5k| size_t pos = 0; 12| | 13| 10.5k| using vector_u16 = simd16; 14| 10.5k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 10.5k| const auto one = vector_u16::splat(1); 17| | 18| 10.5k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 10.5k| size_t count = size / N * N; 22| | 23| | // in a single iteration the increment is 0, 1 or 2, despite we have 24| | // three additions 25| 10.5k| constexpr size_t max_iterations = 65535 / 2; 26| 10.5k| size_t iteration = max_iterations; 27| | 28| 16.2k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 5.77k, False: 10.5k] ------------------ 29| 5.77k| auto input = vector_u16::load(reinterpret_cast(in + pos)); 30| 5.77k| if constexpr (!match_system(big_endian)) { 31| 5.77k| input = input.swap_bytes(); 32| 5.77k| } 33| | // 0xd800 .. 0xdbff - low surrogate 34| | // 0xdc00 .. 0xdfff - high surrogate 35| 5.77k| const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800)); 36| | 37| | // c0 - chars that yield 2- or 3-byte UTF-8 codes 38| 5.77k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 5.77k| const auto c1 = min(input & uint16_t(0xf800), one); 42| | 43| | /* 44| | Explanation how the counting works. 45| | 46| | In the case of a non-surrogate character we count: 47| | * always 1 -- see how `count` is initialized above; 48| | * c0 = 1 if the current char yields 2 or 3 bytes; 49| | * c1 = 1 if the current char yields 3 bytes. 50| | 51| | Thus, we always have correct count for the current char: 52| | from 1, 2 or 3 bytes. 53| | 54| | A trickier part is how we count surrogate pairs. Whether 55| | we encounter a surrogate (low or high), we count it as 56| | 3 chars and then minus 1 (`is_surrogate` is -1 or 0). 57| | Each surrogate char yields 2. A surrogate pair, that 58| | is a low surrogate followed by a high one, yields 59| | the expected 4 bytes. 60| | 61| | It also correctly handles cases when low surrogate is 62| | processed by the this loop, but high surrogate is counted 63| | by the scalar procedure. The scalar procedure uses exactly 64| | the described approach, thanks to that for valid UTF-16 65| | strings it always count correctly. 66| | */ 67| 5.77k| v_count += c0; 68| 5.77k| v_count += c1; 69| 5.77k| v_count += vector_u16(is_surrogate); 70| | 71| 5.77k| iteration -= 1; 72| 5.77k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 5.77k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 5.77k| } 78| | 79| 10.5k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 10.5k, False: 0] ------------------ 80| 10.5k| count += v_count.sum(); 81| 10.5k| } 82| | 83| 10.5k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 10.5k| size - pos); 85| 10.5k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE0EEEmPKDsm: 10| 10.7k| size_t size) { 11| 10.7k| size_t pos = 0; 12| | 13| 10.7k| using vector_u16 = simd16; 14| 10.7k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 10.7k| const auto one = vector_u16::splat(1); 17| | 18| 10.7k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 10.7k| size_t count = size / N * N; 22| | 23| | // in a single iteration the increment is 0, 1 or 2, despite we have 24| | // three additions 25| 10.7k| constexpr size_t max_iterations = 65535 / 2; 26| 10.7k| size_t iteration = max_iterations; 27| | 28| 27.2k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 16.5k, False: 10.7k] ------------------ 29| 16.5k| auto input = vector_u16::load(reinterpret_cast(in + pos)); 30| | if constexpr (!match_system(big_endian)) { 31| | input = input.swap_bytes(); 32| | } 33| | // 0xd800 .. 0xdbff - low surrogate 34| | // 0xdc00 .. 0xdfff - high surrogate 35| 16.5k| const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800)); 36| | 37| | // c0 - chars that yield 2- or 3-byte UTF-8 codes 38| 16.5k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 16.5k| const auto c1 = min(input & uint16_t(0xf800), one); 42| | 43| | /* 44| | Explanation how the counting works. 45| | 46| | In the case of a non-surrogate character we count: 47| | * always 1 -- see how `count` is initialized above; 48| | * c0 = 1 if the current char yields 2 or 3 bytes; 49| | * c1 = 1 if the current char yields 3 bytes. 50| | 51| | Thus, we always have correct count for the current char: 52| | from 1, 2 or 3 bytes. 53| | 54| | A trickier part is how we count surrogate pairs. Whether 55| | we encounter a surrogate (low or high), we count it as 56| | 3 chars and then minus 1 (`is_surrogate` is -1 or 0). 57| | Each surrogate char yields 2. A surrogate pair, that 58| | is a low surrogate followed by a high one, yields 59| | the expected 4 bytes. 60| | 61| | It also correctly handles cases when low surrogate is 62| | processed by the this loop, but high surrogate is counted 63| | by the scalar procedure. The scalar procedure uses exactly 64| | the described approach, thanks to that for valid UTF-16 65| | strings it always count correctly. 66| | */ 67| 16.5k| v_count += c0; 68| 16.5k| v_count += c1; 69| 16.5k| v_count += vector_u16(is_surrogate); 70| | 71| 16.5k| iteration -= 1; 72| 16.5k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 16.5k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 16.5k| } 78| | 79| 10.7k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 10.7k, False: 0] ------------------ 80| 10.7k| count += v_count.sum(); 81| 10.7k| } 82| | 83| 10.7k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 10.7k| size - pos); 85| 10.7k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE1EEEmPKDsm: 10| 10.7k| size_t size) { 11| 10.7k| size_t pos = 0; 12| | 13| 10.7k| using vector_u16 = simd16; 14| 10.7k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 10.7k| const auto one = vector_u16::splat(1); 17| | 18| 10.7k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 10.7k| size_t count = size / N * N; 22| | 23| | // in a single iteration the increment is 0, 1 or 2, despite we have 24| | // three additions 25| 10.7k| constexpr size_t max_iterations = 65535 / 2; 26| 10.7k| size_t iteration = max_iterations; 27| | 28| 27.5k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 16.7k, False: 10.7k] ------------------ 29| 16.7k| auto input = vector_u16::load(reinterpret_cast(in + pos)); 30| 16.7k| if constexpr (!match_system(big_endian)) { 31| 16.7k| input = input.swap_bytes(); 32| 16.7k| } 33| | // 0xd800 .. 0xdbff - low surrogate 34| | // 0xdc00 .. 0xdfff - high surrogate 35| 16.7k| const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800)); 36| | 37| | // c0 - chars that yield 2- or 3-byte UTF-8 codes 38| 16.7k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 16.7k| const auto c1 = min(input & uint16_t(0xf800), one); 42| | 43| | /* 44| | Explanation how the counting works. 45| | 46| | In the case of a non-surrogate character we count: 47| | * always 1 -- see how `count` is initialized above; 48| | * c0 = 1 if the current char yields 2 or 3 bytes; 49| | * c1 = 1 if the current char yields 3 bytes. 50| | 51| | Thus, we always have correct count for the current char: 52| | from 1, 2 or 3 bytes. 53| | 54| | A trickier part is how we count surrogate pairs. Whether 55| | we encounter a surrogate (low or high), we count it as 56| | 3 chars and then minus 1 (`is_surrogate` is -1 or 0). 57| | Each surrogate char yields 2. A surrogate pair, that 58| | is a low surrogate followed by a high one, yields 59| | the expected 4 bytes. 60| | 61| | It also correctly handles cases when low surrogate is 62| | processed by the this loop, but high surrogate is counted 63| | by the scalar procedure. The scalar procedure uses exactly 64| | the described approach, thanks to that for valid UTF-16 65| | strings it always count correctly. 66| | */ 67| 16.7k| v_count += c0; 68| 16.7k| v_count += c1; 69| 16.7k| v_count += vector_u16(is_surrogate); 70| | 71| 16.7k| iteration -= 1; 72| 16.7k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 16.7k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 16.7k| } 78| | 79| 10.7k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 10.7k, False: 0] ------------------ 80| 10.7k| count += v_count.sum(); 81| 10.7k| } 82| | 83| 10.7k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 10.7k| size - pos); 85| 10.7k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf3222utf8_length_from_utf32EPKDim: 11| 4.96k| size_t length) { 12| 4.96k| using vector_u32 = simd32; 13| | 14| 4.96k| const char32_t *start = input; 15| | 16| | // we add up to three ones in a single iteration (see the vectorized loop in 17| | // section #2 below) 18| 4.96k| const size_t max_increment = 3; 19| | 20| 4.96k| const size_t N = vector_u32::ELEMENTS; 21| | 22| |#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP 23| | const auto v_0000007f = vector_u32::splat(0x0000007f); 24| | const auto v_000007ff = vector_u32::splat(0x000007ff); 25| | const auto v_0000ffff = vector_u32::splat(0x0000ffff); 26| |#else 27| 4.96k| const auto v_ffffff80 = vector_u32::splat(0xffffff80); 28| 4.96k| const auto v_fffff800 = vector_u32::splat(0xfffff800); 29| 4.96k| const auto v_ffff0000 = vector_u32::splat(0xffff0000); 30| 4.96k| const auto one = vector_u32::splat(1); 31| 4.96k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 32| | 33| 4.96k| size_t counter = 0; 34| | 35| | // 1. vectorized loop unrolled 4 times 36| 4.96k| { 37| | // we use vector of uint32 counters, this is why this limit is used 38| 4.96k| const size_t max_iterations = 39| 4.96k| std::numeric_limits::max() / (max_increment * 4); 40| 4.96k| size_t blocks = length / (N * 4); 41| 4.96k| length -= blocks * (N * 4); 42| 5.18k| while (blocks != 0) { ------------------ | Branch (42:12): [True: 218, False: 4.96k] ------------------ 43| 218| const size_t iterations = min(blocks, max_iterations); 44| 218| blocks -= iterations; 45| | 46| 218| simd32 acc = vector_u32::zero(); 47| 1.76k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (47:26): [True: 1.54k, False: 218] ------------------ 48| 1.54k| const auto in0 = vector_u32(input + 0 * N); 49| 1.54k| const auto in1 = vector_u32(input + 1 * N); 50| 1.54k| const auto in2 = vector_u32(input + 2 * N); 51| 1.54k| const auto in3 = vector_u32(input + 3 * N); 52| | 53| |#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP 54| | acc -= as_vector_u32(in0 > v_0000007f); 55| | acc -= as_vector_u32(in1 > v_0000007f); 56| | acc -= as_vector_u32(in2 > v_0000007f); 57| | acc -= as_vector_u32(in3 > v_0000007f); 58| | 59| | acc -= as_vector_u32(in0 > v_000007ff); 60| | acc -= as_vector_u32(in1 > v_000007ff); 61| | acc -= as_vector_u32(in2 > v_000007ff); 62| | acc -= as_vector_u32(in3 > v_000007ff); 63| | 64| | acc -= as_vector_u32(in0 > v_0000ffff); 65| | acc -= as_vector_u32(in1 > v_0000ffff); 66| | acc -= as_vector_u32(in2 > v_0000ffff); 67| | acc -= as_vector_u32(in3 > v_0000ffff); 68| |#else 69| 1.54k| acc += min(one, in0 & v_ffffff80); 70| 1.54k| acc += min(one, in1 & v_ffffff80); 71| 1.54k| acc += min(one, in2 & v_ffffff80); 72| 1.54k| acc += min(one, in3 & v_ffffff80); 73| | 74| 1.54k| acc += min(one, in0 & v_fffff800); 75| 1.54k| acc += min(one, in1 & v_fffff800); 76| 1.54k| acc += min(one, in2 & v_fffff800); 77| 1.54k| acc += min(one, in3 & v_fffff800); 78| | 79| 1.54k| acc += min(one, in0 & v_ffff0000); 80| 1.54k| acc += min(one, in1 & v_ffff0000); 81| 1.54k| acc += min(one, in2 & v_ffff0000); 82| 1.54k| acc += min(one, in3 & v_ffff0000); 83| 1.54k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 84| | 85| 1.54k| input += 4 * N; 86| 1.54k| } 87| | 88| 218| counter += acc.sum(); 89| 218| } 90| 4.96k| } 91| | 92| | // 2. vectorized loop for tail 93| 4.96k| { 94| 4.96k| const size_t max_iterations = 95| 4.96k| std::numeric_limits::max() / max_increment; 96| 4.96k| size_t blocks = length / N; 97| 4.96k| length -= blocks * N; 98| 5.45k| while (blocks != 0) { ------------------ | Branch (98:12): [True: 494, False: 4.96k] ------------------ 99| 494| const size_t iterations = min(blocks, max_iterations); 100| 494| blocks -= iterations; 101| | 102| 494| auto acc = vector_u32::zero(); 103| 1.66k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (103:26): [True: 1.16k, False: 494] ------------------ 104| 1.16k| const auto in = vector_u32(input); 105| | 106| |#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP 107| | acc -= as_vector_u32(in > v_0000007f); 108| | acc -= as_vector_u32(in > v_000007ff); 109| | acc -= as_vector_u32(in > v_0000ffff); 110| |#else 111| 1.16k| acc += min(one, in & v_ffffff80); 112| 1.16k| acc += min(one, in & v_fffff800); 113| 1.16k| acc += min(one, in & v_ffff0000); 114| 1.16k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 115| | 116| 1.16k| input += N; 117| 1.16k| } 118| | 119| 494| counter += acc.sum(); 120| 494| } 121| 4.96k| } 122| | 123| 4.96k| const size_t consumed = input - start; 124| 4.96k| if (consumed != 0) { ------------------ | Branch (124:7): [True: 555, False: 4.40k] ------------------ 125| | // We don't count 0th bytes in the vectorized loops above, this 126| | // is why we need to count them in the end. 127| 555| counter += consumed; 128| 555| } 129| | 130| 4.96k| return counter + scalar::utf32::utf8_length_from_utf32(input, length); 131| 4.96k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf323minImEET_S4_S4_: 8| 712|template T min(T a, T b) { return a <= b ? a : b; } ------------------ | Branch (8:48): [True: 712, False: 0] ------------------ simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf3222utf8_length_from_utf32EPKDim: 11| 5.14k| size_t length) { 12| 5.14k| using vector_u32 = simd32; 13| | 14| 5.14k| const char32_t *start = input; 15| | 16| | // we add up to three ones in a single iteration (see the vectorized loop in 17| | // section #2 below) 18| 5.14k| const size_t max_increment = 3; 19| | 20| 5.14k| const size_t N = vector_u32::ELEMENTS; 21| | 22| |#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP 23| | const auto v_0000007f = vector_u32::splat(0x0000007f); 24| | const auto v_000007ff = vector_u32::splat(0x000007ff); 25| | const auto v_0000ffff = vector_u32::splat(0x0000ffff); 26| |#else 27| 5.14k| const auto v_ffffff80 = vector_u32::splat(0xffffff80); 28| 5.14k| const auto v_fffff800 = vector_u32::splat(0xfffff800); 29| 5.14k| const auto v_ffff0000 = vector_u32::splat(0xffff0000); 30| 5.14k| const auto one = vector_u32::splat(1); 31| 5.14k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 32| | 33| 5.14k| size_t counter = 0; 34| | 35| | // 1. vectorized loop unrolled 4 times 36| 5.14k| { 37| | // we use vector of uint32 counters, this is why this limit is used 38| 5.14k| const size_t max_iterations = 39| 5.14k| std::numeric_limits::max() / (max_increment * 4); 40| 5.14k| size_t blocks = length / (N * 4); 41| 5.14k| length -= blocks * (N * 4); 42| 5.64k| while (blocks != 0) { ------------------ | Branch (42:12): [True: 505, False: 5.14k] ------------------ 43| 505| const size_t iterations = min(blocks, max_iterations); 44| 505| blocks -= iterations; 45| | 46| 505| simd32 acc = vector_u32::zero(); 47| 5.66k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (47:26): [True: 5.15k, False: 505] ------------------ 48| 5.15k| const auto in0 = vector_u32(input + 0 * N); 49| 5.15k| const auto in1 = vector_u32(input + 1 * N); 50| 5.15k| const auto in2 = vector_u32(input + 2 * N); 51| 5.15k| const auto in3 = vector_u32(input + 3 * N); 52| | 53| |#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP 54| | acc -= as_vector_u32(in0 > v_0000007f); 55| | acc -= as_vector_u32(in1 > v_0000007f); 56| | acc -= as_vector_u32(in2 > v_0000007f); 57| | acc -= as_vector_u32(in3 > v_0000007f); 58| | 59| | acc -= as_vector_u32(in0 > v_000007ff); 60| | acc -= as_vector_u32(in1 > v_000007ff); 61| | acc -= as_vector_u32(in2 > v_000007ff); 62| | acc -= as_vector_u32(in3 > v_000007ff); 63| | 64| | acc -= as_vector_u32(in0 > v_0000ffff); 65| | acc -= as_vector_u32(in1 > v_0000ffff); 66| | acc -= as_vector_u32(in2 > v_0000ffff); 67| | acc -= as_vector_u32(in3 > v_0000ffff); 68| |#else 69| 5.15k| acc += min(one, in0 & v_ffffff80); 70| 5.15k| acc += min(one, in1 & v_ffffff80); 71| 5.15k| acc += min(one, in2 & v_ffffff80); 72| 5.15k| acc += min(one, in3 & v_ffffff80); 73| | 74| 5.15k| acc += min(one, in0 & v_fffff800); 75| 5.15k| acc += min(one, in1 & v_fffff800); 76| 5.15k| acc += min(one, in2 & v_fffff800); 77| 5.15k| acc += min(one, in3 & v_fffff800); 78| | 79| 5.15k| acc += min(one, in0 & v_ffff0000); 80| 5.15k| acc += min(one, in1 & v_ffff0000); 81| 5.15k| acc += min(one, in2 & v_ffff0000); 82| 5.15k| acc += min(one, in3 & v_ffff0000); 83| 5.15k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 84| | 85| 5.15k| input += 4 * N; 86| 5.15k| } 87| | 88| 505| counter += acc.sum(); 89| 505| } 90| 5.14k| } 91| | 92| | // 2. vectorized loop for tail 93| 5.14k| { 94| 5.14k| const size_t max_iterations = 95| 5.14k| std::numeric_limits::max() / max_increment; 96| 5.14k| size_t blocks = length / N; 97| 5.14k| length -= blocks * N; 98| 5.58k| while (blocks != 0) { ------------------ | Branch (98:12): [True: 445, False: 5.14k] ------------------ 99| 445| const size_t iterations = min(blocks, max_iterations); 100| 445| blocks -= iterations; 101| | 102| 445| auto acc = vector_u32::zero(); 103| 1.48k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (103:26): [True: 1.03k, False: 445] ------------------ 104| 1.03k| const auto in = vector_u32(input); 105| | 106| |#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP 107| | acc -= as_vector_u32(in > v_0000007f); 108| | acc -= as_vector_u32(in > v_000007ff); 109| | acc -= as_vector_u32(in > v_0000ffff); 110| |#else 111| 1.03k| acc += min(one, in & v_ffffff80); 112| 1.03k| acc += min(one, in & v_fffff800); 113| 1.03k| acc += min(one, in & v_ffff0000); 114| 1.03k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 115| | 116| 1.03k| input += N; 117| 1.03k| } 118| | 119| 445| counter += acc.sum(); 120| 445| } 121| 5.14k| } 122| | 123| 5.14k| const size_t consumed = input - start; 124| 5.14k| if (consumed != 0) { ------------------ | Branch (124:7): [True: 547, False: 4.59k] ------------------ 125| | // We don't count 0th bytes in the vectorized loops above, this 126| | // is why we need to count them in the end. 127| 547| counter += consumed; 128| 547| } 129| | 130| 5.14k| return counter + scalar::utf32::utf8_length_from_utf32(input, length); 131| 5.14k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf323minImEET_S4_S4_: 8| 950|template T min(T a, T b) { return a <= b ? a : b; } ------------------ | Branch (8:48): [True: 950, False: 0] ------------------ simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14utf826count_code_points_bytemaskEPKcm: 21| 11.1k|count_code_points_bytemask(const char *in, size_t size) { 22| 11.1k| using vector_i8 = simd8; 23| 11.1k| using vector_u8 = simd8; 24| 11.1k| using vector_u64 = simd64; 25| | 26| 11.1k| constexpr size_t N = vector_i8::SIZE; 27| 11.1k| constexpr size_t max_iterations = 255 / 4; 28| | 29| 11.1k| size_t pos = 0; 30| 11.1k| size_t count = 0; 31| | 32| 11.1k| auto counters = vector_u64::zero(); 33| 11.1k| auto local = vector_u8::zero(); 34| 11.1k| size_t iterations = 0; 35| 12.9k| for (; pos + 4 * N <= size; pos += 4 * N) { ------------------ | Branch (35:10): [True: 1.77k, False: 11.1k] ------------------ 36| 1.77k| const auto input0 = 37| 1.77k| simd8::load(reinterpret_cast(in + pos + 0 * N)); 38| 1.77k| const auto input1 = 39| 1.77k| simd8::load(reinterpret_cast(in + pos + 1 * N)); 40| 1.77k| const auto input2 = 41| 1.77k| simd8::load(reinterpret_cast(in + pos + 2 * N)); 42| 1.77k| const auto input3 = 43| 1.77k| simd8::load(reinterpret_cast(in + pos + 3 * N)); 44| 1.77k| const auto mask0 = input0 > int8_t(-65); 45| 1.77k| const auto mask1 = input1 > int8_t(-65); 46| 1.77k| const auto mask2 = input2 > int8_t(-65); 47| 1.77k| const auto mask3 = input3 > int8_t(-65); 48| | 49| 1.77k| local -= vector_u8(mask0); 50| 1.77k| local -= vector_u8(mask1); 51| 1.77k| local -= vector_u8(mask2); 52| 1.77k| local -= vector_u8(mask3); 53| | 54| 1.77k| iterations += 1; 55| 1.77k| if (iterations == max_iterations) { ------------------ | Branch (55:9): [True: 0, False: 1.77k] ------------------ 56| 0| counters += sum_8bytes(local); 57| 0| local = vector_u8::zero(); 58| 0| iterations = 0; 59| 0| } 60| 1.77k| } 61| | 62| 11.1k| if (iterations > 0) { ------------------ | Branch (62:7): [True: 440, False: 10.6k] ------------------ 63| 440| count += local.sum_bytes(); 64| 440| } 65| | 66| 11.1k| count += counters.sum(); 67| | 68| 11.1k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 69| 11.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14utf817count_code_pointsEPKcm: 8| 5.56k|simdutf_really_inline size_t count_code_points(const char *in, size_t size) { 9| 5.56k| size_t pos = 0; 10| 5.56k| size_t count = 0; 11| 7.25k| for (; pos + 64 <= size; pos += 64) { ------------------ | Branch (11:10): [True: 1.69k, False: 5.56k] ------------------ 12| 1.69k| simd8x64 input(reinterpret_cast(in + pos)); 13| 1.69k| uint64_t utf8_continuation_mask = input.gt(-65); 14| 1.69k| count += count_ones(utf8_continuation_mask); 15| 1.69k| } 16| 5.56k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 17| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14utf826count_code_points_bytemaskEPKcm: 21| 11.1k|count_code_points_bytemask(const char *in, size_t size) { 22| 11.1k| using vector_i8 = simd8; 23| 11.1k| using vector_u8 = simd8; 24| 11.1k| using vector_u64 = simd64; 25| | 26| 11.1k| constexpr size_t N = vector_i8::SIZE; 27| 11.1k| constexpr size_t max_iterations = 255 / 4; 28| | 29| 11.1k| size_t pos = 0; 30| 11.1k| size_t count = 0; 31| | 32| 11.1k| auto counters = vector_u64::zero(); 33| 11.1k| auto local = vector_u8::zero(); 34| 11.1k| size_t iterations = 0; 35| 16.8k| for (; pos + 4 * N <= size; pos += 4 * N) { ------------------ | Branch (35:10): [True: 5.69k, False: 11.1k] ------------------ 36| 5.69k| const auto input0 = 37| 5.69k| simd8::load(reinterpret_cast(in + pos + 0 * N)); 38| 5.69k| const auto input1 = 39| 5.69k| simd8::load(reinterpret_cast(in + pos + 1 * N)); 40| 5.69k| const auto input2 = 41| 5.69k| simd8::load(reinterpret_cast(in + pos + 2 * N)); 42| 5.69k| const auto input3 = 43| 5.69k| simd8::load(reinterpret_cast(in + pos + 3 * N)); 44| 5.69k| const auto mask0 = input0 > int8_t(-65); 45| 5.69k| const auto mask1 = input1 > int8_t(-65); 46| 5.69k| const auto mask2 = input2 > int8_t(-65); 47| 5.69k| const auto mask3 = input3 > int8_t(-65); 48| | 49| 5.69k| local -= vector_u8(mask0); 50| 5.69k| local -= vector_u8(mask1); 51| 5.69k| local -= vector_u8(mask2); 52| 5.69k| local -= vector_u8(mask3); 53| | 54| 5.69k| iterations += 1; 55| 5.69k| if (iterations == max_iterations) { ------------------ | Branch (55:9): [True: 0, False: 5.69k] ------------------ 56| 0| counters += sum_8bytes(local); 57| 0| local = vector_u8::zero(); 58| 0| iterations = 0; 59| 0| } 60| 5.69k| } 61| | 62| 11.1k| if (iterations > 0) { ------------------ | Branch (62:7): [True: 1.06k, False: 10.0k] ------------------ 63| 1.06k| count += local.sum_bytes(); 64| 1.06k| } 65| | 66| 11.1k| count += counters.sum(); 67| | 68| 11.1k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 69| 11.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14utf817count_code_pointsEPKcm: 8| 5.56k|simdutf_really_inline size_t count_code_points(const char *in, size_t size) { 9| 5.56k| size_t pos = 0; 10| 5.56k| size_t count = 0; 11| 8.19k| for (; pos + 64 <= size; pos += 64) { ------------------ | Branch (11:10): [True: 2.63k, False: 5.56k] ------------------ 12| 2.63k| simd8x64 input(reinterpret_cast(in + pos)); 13| 2.63k| uint64_t utf8_continuation_mask = input.gt(-65); 14| 2.63k| count += count_ones(utf8_continuation_mask); 15| 2.63k| } 16| 5.56k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 17| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14utf831utf16_length_from_utf8_bytemaskEPKcm: 9| 22.0k| size_t size) { 10| 22.0k| using vector_i8 = simd8; 11| 22.0k| using vector_u8 = simd8; 12| 22.0k| using vector_u64 = simd64; 13| | 14| 22.0k| constexpr size_t N = vector_i8::SIZE; 15| 22.0k| constexpr size_t max_iterations = 255 / 2; 16| | 17| 22.0k| auto counters = vector_u64::zero(); 18| 22.0k| auto local = vector_u8::zero(); 19| | 20| 22.0k| size_t iterations = 0; 21| 22.0k| size_t pos = 0; 22| 22.0k| size_t count = 0; 23| 33.5k| for (; pos + N <= size; pos += N) { ------------------ | Branch (23:10): [True: 11.4k, False: 22.0k] ------------------ 24| 11.4k| const auto input = 25| 11.4k| vector_i8::load(reinterpret_cast(in + pos)); 26| | 27| 11.4k| const auto continuation = input > int8_t(-65); 28| 11.4k| const auto utf_4bytes = vector_u8(input.value) >= uint8_t(240); 29| | 30| 11.4k| local -= vector_u8(continuation); 31| 11.4k| local -= vector_u8(utf_4bytes); 32| | 33| 11.4k| iterations += 1; 34| 11.4k| if (iterations == max_iterations) { ------------------ | Branch (34:9): [True: 0, False: 11.4k] ------------------ 35| 0| counters += sum_8bytes(local); 36| 0| local = vector_u8::zero(); 37| 0| iterations = 0; 38| 0| } 39| 11.4k| } 40| | 41| 22.0k| if (iterations > 0) { ------------------ | Branch (41:7): [True: 1.73k, False: 20.3k] ------------------ 42| 1.73k| count += local.sum_bytes(); 43| 1.73k| } 44| | 45| 22.0k| count += counters.sum(); 46| | 47| 22.0k| return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos); 48| 22.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14utf831utf16_length_from_utf8_bytemaskEPKcm: 9| 22.0k| size_t size) { 10| 22.0k| using vector_i8 = simd8; 11| 22.0k| using vector_u8 = simd8; 12| 22.0k| using vector_u64 = simd64; 13| | 14| 22.0k| constexpr size_t N = vector_i8::SIZE; 15| 22.0k| constexpr size_t max_iterations = 255 / 2; 16| | 17| 22.0k| auto counters = vector_u64::zero(); 18| 22.0k| auto local = vector_u8::zero(); 19| | 20| 22.0k| size_t iterations = 0; 21| 22.0k| size_t pos = 0; 22| 22.0k| size_t count = 0; 23| 56.7k| for (; pos + N <= size; pos += N) { ------------------ | Branch (23:10): [True: 34.7k, False: 22.0k] ------------------ 24| 34.7k| const auto input = 25| 34.7k| vector_i8::load(reinterpret_cast(in + pos)); 26| | 27| 34.7k| const auto continuation = input > int8_t(-65); 28| 34.7k| const auto utf_4bytes = vector_u8(input.value) >= uint8_t(240); 29| | 30| 34.7k| local -= vector_u8(continuation); 31| 34.7k| local -= vector_u8(utf_4bytes); 32| | 33| 34.7k| iterations += 1; 34| 34.7k| if (iterations == max_iterations) { ------------------ | Branch (34:9): [True: 0, False: 34.7k] ------------------ 35| 0| counters += sum_8bytes(local); 36| 0| local = vector_u8::zero(); 37| 0| iterations = 0; 38| 0| } 39| 34.7k| } 40| | 41| 22.0k| if (iterations > 0) { ------------------ | Branch (41:7): [True: 1.99k, False: 20.0k] ------------------ 42| 1.99k| count += local.sum_bytes(); 43| 1.99k| } 44| | 45| 22.0k| count += counters.sum(); 46| | 47| 22.0k| return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos); 48| 22.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoderC2Ev: 101| 11.1k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoder7convertEPKcmPc: 115| 11.1k| char *latin1_output) { 116| 11.1k| size_t pos = 0; 117| 11.1k| char *start{latin1_output}; 118| | // In the worst case, we have the haswell kernel which can cause an overflow 119| | // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the 120| | // last 16 bytes, and if the data is valid, then it is entirely safe because 121| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 122| | // generally assume that you have valid UTF-8 input, so we are going to go 123| | // back from the end counting 16 leading bytes, to give us a good margin. 124| 11.1k| size_t leading_byte = 0; 125| 11.1k| size_t margin = size; 126| 52.2k| for (; margin > 0 && leading_byte < 16; margin--) { ------------------ | Branch (126:12): [True: 42.7k, False: 9.48k] | Branch (126:26): [True: 41.1k, False: 1.63k] ------------------ 127| 41.1k| leading_byte += (int8_t(in[margin - 1]) > 128| 41.1k| -65); // twos complement of -65 is 1011 1111 ... 129| 41.1k| } 130| | // If the input is long enough, then we have that margin-1 is the eight last 131| | // leading byte. 132| 11.1k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 133| 14.9k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (133:12): [True: 3.81k, False: 11.1k] ------------------ 134| 3.81k| simd8x64 input(reinterpret_cast(in + pos)); 135| 3.81k| if (input.is_ascii()) { ------------------ | Branch (135:11): [True: 1.66k, False: 2.15k] ------------------ 136| 1.66k| input.store((int8_t *)latin1_output); 137| 1.66k| latin1_output += 64; 138| 1.66k| pos += 64; 139| 2.15k| } else { 140| | // you might think that a for-loop would work, but under Visual Studio, 141| | // it is not good enough. 142| 2.15k| static_assert( 143| 2.15k| (simd8x64::NUM_CHUNKS == 2) || 144| 2.15k| (simd8x64::NUM_CHUNKS == 4), 145| 2.15k| "We support either two or four chunks per 64-byte block."); 146| 2.15k| auto zero = simd8{uint8_t(0)}; 147| 2.15k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 148| 2.15k| this->check_utf8_bytes(input.chunks[0], zero); 149| 2.15k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 150| | } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 151| | this->check_utf8_bytes(input.chunks[0], zero); 152| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 153| | this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 154| | this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 155| | } 156| 2.15k| uint64_t utf8_continuation_mask = 157| 2.15k| input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in 158| | // this case, we also have ASCII to account for. 159| 2.15k| if (utf8_continuation_mask & 1) { ------------------ | Branch (159:13): [True: 8, False: 2.14k] ------------------ 160| 8| return 0; // error 161| 8| } 162| 2.14k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 163| 2.14k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 164| | // We process in blocks of up to 12 bytes except possibly 165| | // for fast paths which may process up to 16 bytes. For the 166| | // slow path to work, we should have at least 12 input bytes left. 167| 2.14k| size_t max_starting_point = (pos + 64) - 12; 168| | // Next loop is going to run at least five times. 169| 17.7k| while (pos < max_starting_point) { ------------------ | Branch (169:16): [True: 15.5k, False: 2.14k] ------------------ 170| | // Performance note: our ability to compute 'consumed' and 171| | // then shift and recompute is critical. If there is a 172| | // latency of, say, 4 cycles on getting 'consumed', then 173| | // the inner loop might have a total latency of about 6 cycles. 174| | // Yet we process between 6 to 12 inputs bytes, thus we get 175| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 176| | // for this section of the code. Hence, there is a limit 177| | // to how much we can further increase this latency before 178| | // it seriously harms performance. 179| 15.5k| size_t consumed = convert_masked_utf8_to_latin1( 180| 15.5k| in + pos, utf8_end_of_code_point_mask, latin1_output); 181| 15.5k| pos += consumed; 182| 15.5k| utf8_end_of_code_point_mask >>= consumed; 183| 15.5k| } 184| | // At this point there may remain between 0 and 12 bytes in the 185| | // 64-byte block. These bytes will be processed again. So we have an 186| | // 80% efficiency (in the worst case). In practice we expect an 187| | // 85% to 90% efficiency. 188| 2.14k| } 189| 3.81k| } 190| 11.1k| if (errors()) { ------------------ | Branch (190:9): [True: 156, False: 10.9k] ------------------ 191| 156| return 0; 192| 156| } 193| 10.9k| if (pos < size) { ------------------ | Branch (193:9): [True: 2.46k, False: 8.49k] ------------------ 194| 2.46k| size_t howmany = 195| 2.46k| scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output); 196| 2.46k| if (howmany == 0) { ------------------ | Branch (196:11): [True: 1.02k, False: 1.44k] ------------------ 197| 1.02k| return 0; 198| 1.02k| } 199| 1.44k| latin1_output += howmany; 200| 1.44k| } 201| 9.93k| return latin1_output - start; 202| 10.9k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 106| 4.30k| const simd8 prev_input) { 107| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 108| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 109| | // small negative numbers) 110| 4.30k| simd8 prev1 = input.prev<1>(prev_input); 111| 4.30k| this->error |= check_special_cases(input, prev1); 112| 4.30k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin119check_special_casesENS1_4simd5simd8IhEES5_: 8| 4.30k|check_special_cases(const simd8 input, const simd8 prev1) { 9| | // For UTF-8 to Latin 1, we can allow any ASCII character, and any 10| | // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or 11| | // 0b11000010 and nothing else. 12| | // 13| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 14| | // Bit 1 = Too Long (ASCII followed by continuation) 15| | // Bit 2 = Overlong 3-byte 16| | // Bit 4 = Surrogate 17| | // Bit 5 = Overlong 2-byte 18| | // Bit 7 = Two Continuations 19| 4.30k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 20| | // 11______ 11______ 21| 4.30k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 22| 4.30k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 23| 4.30k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 24| 4.30k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 25| 4.30k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 26| 4.30k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 27| | // 11110100 101_____ 28| | // 11110101 1001____ 29| | // 11110101 101_____ 30| | // 1111011_ 1001____ 31| | // 1111011_ 101_____ 32| | // 11111___ 1001____ 33| | // 11111___ 101_____ 34| 4.30k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 35| | // 11110101 1000____ 36| | // 1111011_ 1000____ 37| | // 11111___ 1000____ 38| 4.30k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 39| 4.30k| constexpr const uint8_t FORBIDDEN = 0xff; 40| | 41| 4.30k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 42| | // 0_______ ________ 43| 4.30k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 44| 4.30k| TOO_LONG, 45| | // 10______ ________ 46| 4.30k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 47| | // 1100____ ________ 48| 4.30k| TOO_SHORT | OVERLONG_2, 49| | // 1101____ ________ 50| 4.30k| FORBIDDEN, 51| | // 1110____ ________ 52| 4.30k| FORBIDDEN, 53| | // 1111____ ________ 54| 4.30k| FORBIDDEN); 55| 4.30k| constexpr const uint8_t CARRY = 56| 4.30k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 57| 4.30k| const simd8 byte_1_low = 58| 4.30k| (prev1 & 0x0F) 59| 4.30k| .lookup_16( 60| | // ____0000 ________ 61| 4.30k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 62| | // ____0001 ________ 63| 4.30k| CARRY | OVERLONG_2, 64| | // ____001_ ________ 65| 4.30k| CARRY, CARRY, 66| | 67| | // ____0100 ________ 68| 4.30k| FORBIDDEN, 69| | // ____0101 ________ 70| 4.30k| FORBIDDEN, 71| | // ____011_ ________ 72| 4.30k| FORBIDDEN, FORBIDDEN, 73| | 74| | // ____1___ ________ 75| 4.30k| FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, 76| | // ____1101 ________ 77| 4.30k| FORBIDDEN, FORBIDDEN, FORBIDDEN); 78| 4.30k| const simd8 byte_2_high = input.shr<4>().lookup_16( 79| | // ________ 0_______ 80| 4.30k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 81| 4.30k| TOO_SHORT, TOO_SHORT, 82| | 83| | // ________ 1000____ 84| 4.30k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 85| 4.30k| OVERLONG_4, 86| | // ________ 1001____ 87| 4.30k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 88| | // ________ 101_____ 89| 4.30k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 90| 4.30k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 91| | 92| | // ________ 11______ 93| 4.30k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 94| 4.30k| return (byte_1_high & byte_1_low & byte_2_high); 95| 4.30k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoder6errorsEv: 308| 11.1k| simdutf_really_inline bool errors() const { 309| 11.1k| return this->error.any_bits_set_anywhere(); 310| 11.1k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoderC2Ev: 101| 11.1k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoder7convertEPKcmPc: 115| 11.1k| char *latin1_output) { 116| 11.1k| size_t pos = 0; 117| 11.1k| char *start{latin1_output}; 118| | // In the worst case, we have the haswell kernel which can cause an overflow 119| | // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the 120| | // last 16 bytes, and if the data is valid, then it is entirely safe because 121| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 122| | // generally assume that you have valid UTF-8 input, so we are going to go 123| | // back from the end counting 16 leading bytes, to give us a good margin. 124| 11.1k| size_t leading_byte = 0; 125| 11.1k| size_t margin = size; 126| 41.4k| for (; margin > 0 && leading_byte < 16; margin--) { ------------------ | Branch (126:12): [True: 31.7k, False: 9.72k] | Branch (126:26): [True: 30.3k, False: 1.39k] ------------------ 127| 30.3k| leading_byte += (int8_t(in[margin - 1]) > 128| 30.3k| -65); // twos complement of -65 is 1011 1111 ... 129| 30.3k| } 130| | // If the input is long enough, then we have that margin-1 is the eight last 131| | // leading byte. 132| 11.1k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 133| 16.4k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (133:12): [True: 5.32k, False: 11.1k] ------------------ 134| 5.32k| simd8x64 input(reinterpret_cast(in + pos)); 135| 5.32k| if (input.is_ascii()) { ------------------ | Branch (135:11): [True: 2.08k, False: 3.24k] ------------------ 136| 2.08k| input.store((int8_t *)latin1_output); 137| 2.08k| latin1_output += 64; 138| 2.08k| pos += 64; 139| 3.24k| } else { 140| | // you might think that a for-loop would work, but under Visual Studio, 141| | // it is not good enough. 142| 3.24k| static_assert( 143| 3.24k| (simd8x64::NUM_CHUNKS == 2) || 144| 3.24k| (simd8x64::NUM_CHUNKS == 4), 145| 3.24k| "We support either two or four chunks per 64-byte block."); 146| 3.24k| auto zero = simd8{uint8_t(0)}; 147| | if constexpr (simd8x64::NUM_CHUNKS == 2) { 148| | this->check_utf8_bytes(input.chunks[0], zero); 149| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 150| 3.24k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 151| 3.24k| this->check_utf8_bytes(input.chunks[0], zero); 152| 3.24k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 153| 3.24k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 154| 3.24k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 155| 3.24k| } 156| 3.24k| uint64_t utf8_continuation_mask = 157| 3.24k| input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in 158| | // this case, we also have ASCII to account for. 159| 3.24k| if (utf8_continuation_mask & 1) { ------------------ | Branch (159:13): [True: 9, False: 3.23k] ------------------ 160| 9| return 0; // error 161| 9| } 162| 3.23k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 163| 3.23k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 164| | // We process in blocks of up to 12 bytes except possibly 165| | // for fast paths which may process up to 16 bytes. For the 166| | // slow path to work, we should have at least 12 input bytes left. 167| 3.23k| size_t max_starting_point = (pos + 64) - 12; 168| | // Next loop is going to run at least five times. 169| 26.9k| while (pos < max_starting_point) { ------------------ | Branch (169:16): [True: 23.7k, False: 3.23k] ------------------ 170| | // Performance note: our ability to compute 'consumed' and 171| | // then shift and recompute is critical. If there is a 172| | // latency of, say, 4 cycles on getting 'consumed', then 173| | // the inner loop might have a total latency of about 6 cycles. 174| | // Yet we process between 6 to 12 inputs bytes, thus we get 175| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 176| | // for this section of the code. Hence, there is a limit 177| | // to how much we can further increase this latency before 178| | // it seriously harms performance. 179| 23.7k| size_t consumed = convert_masked_utf8_to_latin1( 180| 23.7k| in + pos, utf8_end_of_code_point_mask, latin1_output); 181| 23.7k| pos += consumed; 182| 23.7k| utf8_end_of_code_point_mask >>= consumed; 183| 23.7k| } 184| | // At this point there may remain between 0 and 12 bytes in the 185| | // 64-byte block. These bytes will be processed again. So we have an 186| | // 80% efficiency (in the worst case). In practice we expect an 187| | // 85% to 90% efficiency. 188| 3.23k| } 189| 5.32k| } 190| 11.1k| if (errors()) { ------------------ | Branch (190:9): [True: 215, False: 10.9k] ------------------ 191| 215| return 0; 192| 215| } 193| 10.9k| if (pos < size) { ------------------ | Branch (193:9): [True: 1.72k, False: 9.17k] ------------------ 194| 1.72k| size_t howmany = 195| 1.72k| scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output); 196| 1.72k| if (howmany == 0) { ------------------ | Branch (196:11): [True: 641, False: 1.08k] ------------------ 197| 641| return 0; 198| 641| } 199| 1.08k| latin1_output += howmany; 200| 1.08k| } 201| 10.2k| return latin1_output - start; 202| 10.9k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 106| 12.9k| const simd8 prev_input) { 107| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 108| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 109| | // small negative numbers) 110| 12.9k| simd8 prev1 = input.prev<1>(prev_input); 111| 12.9k| this->error |= check_special_cases(input, prev1); 112| 12.9k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin119check_special_casesENS1_4simd5simd8IhEES5_: 8| 12.9k|check_special_cases(const simd8 input, const simd8 prev1) { 9| | // For UTF-8 to Latin 1, we can allow any ASCII character, and any 10| | // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or 11| | // 0b11000010 and nothing else. 12| | // 13| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 14| | // Bit 1 = Too Long (ASCII followed by continuation) 15| | // Bit 2 = Overlong 3-byte 16| | // Bit 4 = Surrogate 17| | // Bit 5 = Overlong 2-byte 18| | // Bit 7 = Two Continuations 19| 12.9k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 20| | // 11______ 11______ 21| 12.9k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 22| 12.9k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 23| 12.9k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 24| 12.9k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 25| 12.9k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 26| 12.9k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 27| | // 11110100 101_____ 28| | // 11110101 1001____ 29| | // 11110101 101_____ 30| | // 1111011_ 1001____ 31| | // 1111011_ 101_____ 32| | // 11111___ 1001____ 33| | // 11111___ 101_____ 34| 12.9k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 35| | // 11110101 1000____ 36| | // 1111011_ 1000____ 37| | // 11111___ 1000____ 38| 12.9k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 39| 12.9k| constexpr const uint8_t FORBIDDEN = 0xff; 40| | 41| 12.9k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 42| | // 0_______ ________ 43| 12.9k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 44| 12.9k| TOO_LONG, 45| | // 10______ ________ 46| 12.9k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 47| | // 1100____ ________ 48| 12.9k| TOO_SHORT | OVERLONG_2, 49| | // 1101____ ________ 50| 12.9k| FORBIDDEN, 51| | // 1110____ ________ 52| 12.9k| FORBIDDEN, 53| | // 1111____ ________ 54| 12.9k| FORBIDDEN); 55| 12.9k| constexpr const uint8_t CARRY = 56| 12.9k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 57| 12.9k| const simd8 byte_1_low = 58| 12.9k| (prev1 & 0x0F) 59| 12.9k| .lookup_16( 60| | // ____0000 ________ 61| 12.9k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 62| | // ____0001 ________ 63| 12.9k| CARRY | OVERLONG_2, 64| | // ____001_ ________ 65| 12.9k| CARRY, CARRY, 66| | 67| | // ____0100 ________ 68| 12.9k| FORBIDDEN, 69| | // ____0101 ________ 70| 12.9k| FORBIDDEN, 71| | // ____011_ ________ 72| 12.9k| FORBIDDEN, FORBIDDEN, 73| | 74| | // ____1___ ________ 75| 12.9k| FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, 76| | // ____1101 ________ 77| 12.9k| FORBIDDEN, FORBIDDEN, FORBIDDEN); 78| 12.9k| const simd8 byte_2_high = input.shr<4>().lookup_16( 79| | // ________ 0_______ 80| 12.9k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 81| 12.9k| TOO_SHORT, TOO_SHORT, 82| | 83| | // ________ 1000____ 84| 12.9k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 85| 12.9k| OVERLONG_4, 86| | // ________ 1001____ 87| 12.9k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 88| | // ________ 101_____ 89| 12.9k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 90| 12.9k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 91| | 92| | // ________ 11______ 93| 12.9k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 94| 12.9k| return (byte_1_high & byte_1_low & byte_2_high); 95| 12.9k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoder6errorsEv: 308| 11.1k| simdutf_really_inline bool errors() const { 309| 11.1k| return this->error.any_bits_set_anywhere(); 310| 11.1k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoderC2Ev: 114| 22.0k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoderC2Ev: 114| 22.0k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE0EEEmPKcmPDs: 130| 11.0k| char16_t *utf16_output) { 131| 11.0k| size_t pos = 0; 132| 11.0k| char16_t *start{utf16_output}; 133| | // In the worst case, we have the haswell kernel which can cause an overflow 134| | // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the 135| | // last 16 bytes, and if the data is valid, then it is entirely safe because 136| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 137| | // generally assume that you have valid UTF-8 input, so we are going to go 138| | // back from the end counting 8 leading bytes, to give us a good margin. 139| 11.0k| size_t leading_byte = 0; 140| 11.0k| size_t margin = size; 141| 29.0k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 19.1k, False: 9.94k] | Branch (141:26): [True: 18.0k, False: 1.08k] ------------------ 142| 18.0k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 18.0k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 11.0k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 13.5k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 2.56k, False: 11.0k] ------------------ 148| 2.56k| simd8x64 input(reinterpret_cast(in + pos)); 149| 2.56k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 599, False: 1.97k] ------------------ 150| 599| input.store_ascii_as_utf16(utf16_output); 151| 599| utf16_output += 64; 152| 599| pos += 64; 153| 1.97k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 1.97k| static_assert( 157| 1.97k| (simd8x64::NUM_CHUNKS == 2) || 158| 1.97k| (simd8x64::NUM_CHUNKS == 4), 159| 1.97k| "We support either two or four chunks per 64-byte block."); 160| 1.97k| auto zero = simd8{uint8_t(0)}; 161| 1.97k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 162| 1.97k| this->check_utf8_bytes(input.chunks[0], zero); 163| 1.97k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 164| | } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 165| | this->check_utf8_bytes(input.chunks[0], zero); 166| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 167| | this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 168| | this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 169| | } 170| 1.97k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 1.97k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 9, False: 1.96k] ------------------ 172| 9| return 0; // error 173| 9| } 174| 1.96k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 1.96k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 176| | // We process in blocks of up to 12 bytes except possibly 177| | // for fast paths which may process up to 16 bytes. For the 178| | // slow path to work, we should have at least 12 input bytes left. 179| 1.96k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 16.2k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 14.3k, False: 1.96k] ------------------ 182| | // Performance note: our ability to compute 'consumed' and 183| | // then shift and recompute is critical. If there is a 184| | // latency of, say, 4 cycles on getting 'consumed', then 185| | // the inner loop might have a total latency of about 6 cycles. 186| | // Yet we process between 6 to 12 inputs bytes, thus we get 187| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 188| | // for this section of the code. Hence, there is a limit 189| | // to how much we can further increase this latency before 190| | // it seriously harms performance. 191| 14.3k| size_t consumed = convert_masked_utf8_to_utf16( 192| 14.3k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 14.3k| pos += consumed; 194| 14.3k| utf8_end_of_code_point_mask >>= consumed; 195| 14.3k| } 196| | // At this point there may remain between 0 and 12 bytes in the 197| | // 64-byte block. These bytes will be processed again. So we have an 198| | // 80% efficiency (in the worst case). In practice we expect an 199| | // 85% to 90% efficiency. 200| 1.96k| } 201| 2.56k| } 202| 11.0k| if (errors()) { ------------------ | Branch (202:9): [True: 188, False: 10.8k] ------------------ 203| 188| return 0; 204| 188| } 205| 10.8k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.52k, False: 9.30k] ------------------ 206| 1.52k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.52k| in + pos, size - pos, utf16_output); 208| 1.52k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 403, False: 1.12k] ------------------ 209| 403| return 0; 210| 403| } 211| 1.12k| utf16_output += howmany; 212| 1.12k| } 213| 10.4k| return utf16_output - start; 214| 10.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 8.36k| const simd8 prev_input) { 120| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 121| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 122| | // small negative numbers) 123| 8.36k| simd8 prev1 = input.prev<1>(prev_input); 124| 8.36k| simd8 sc = check_special_cases(input, prev1); 125| 8.36k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 8.36k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1619check_special_casesENS1_4simd5simd8IhEES5_: 8| 8.36k|check_special_cases(const simd8 input, const simd8 prev1) { 9| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 10| | // Bit 1 = Too Long (ASCII followed by continuation) 11| | // Bit 2 = Overlong 3-byte 12| | // Bit 4 = Surrogate 13| | // Bit 5 = Overlong 2-byte 14| | // Bit 7 = Two Continuations 15| 8.36k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 8.36k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 8.36k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 8.36k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 8.36k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 8.36k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 8.36k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 23| | // 11110100 101_____ 24| | // 11110101 1001____ 25| | // 11110101 101_____ 26| | // 1111011_ 1001____ 27| | // 1111011_ 101_____ 28| | // 11111___ 1001____ 29| | // 11111___ 101_____ 30| 8.36k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 8.36k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 8.36k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 8.36k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 8.36k| TOO_LONG, 40| | // 10______ ________ 41| 8.36k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 8.36k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 8.36k| TOO_SHORT, 46| | // 1110____ ________ 47| 8.36k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 8.36k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 8.36k| constexpr const uint8_t CARRY = 51| 8.36k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 8.36k| const simd8 byte_1_low = 53| 8.36k| (prev1 & 0x0F) 54| 8.36k| .lookup_16( 55| | // ____0000 ________ 56| 8.36k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 8.36k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 8.36k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 8.36k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 8.36k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 8.36k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 8.36k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 8.36k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 8.36k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 8.36k| OVERLONG_4, 88| | // ________ 1001____ 89| 8.36k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 8.36k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 8.36k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 8.36k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 8.36k| return (byte_1_high & byte_1_low & byte_2_high); 97| 8.36k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1623check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 8.36k| const simd8 sc) { 102| 8.36k| simd8 prev2 = input.prev<2>(prev_input); 103| 8.36k| simd8 prev3 = input.prev<3>(prev_input); 104| 8.36k| simd8 must23 = 105| 8.36k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 8.36k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 8.36k| return must23_80 ^ sc; 108| 8.36k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder6errorsEv: 324| 22.0k| simdutf_really_inline bool errors() const { 325| 22.0k| return this->error.any_bits_set_anywhere(); 326| 22.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE1EEEmPKcmPDs: 130| 11.0k| char16_t *utf16_output) { 131| 11.0k| size_t pos = 0; 132| 11.0k| char16_t *start{utf16_output}; 133| | // In the worst case, we have the haswell kernel which can cause an overflow 134| | // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the 135| | // last 16 bytes, and if the data is valid, then it is entirely safe because 136| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 137| | // generally assume that you have valid UTF-8 input, so we are going to go 138| | // back from the end counting 8 leading bytes, to give us a good margin. 139| 11.0k| size_t leading_byte = 0; 140| 11.0k| size_t margin = size; 141| 29.5k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 19.6k, False: 9.93k] | Branch (141:26): [True: 18.5k, False: 1.10k] ------------------ 142| 18.5k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 18.5k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 11.0k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 13.8k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 2.81k, False: 11.0k] ------------------ 148| 2.81k| simd8x64 input(reinterpret_cast(in + pos)); 149| 2.81k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 598, False: 2.21k] ------------------ 150| 598| input.store_ascii_as_utf16(utf16_output); 151| 598| utf16_output += 64; 152| 598| pos += 64; 153| 2.21k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 2.21k| static_assert( 157| 2.21k| (simd8x64::NUM_CHUNKS == 2) || 158| 2.21k| (simd8x64::NUM_CHUNKS == 4), 159| 2.21k| "We support either two or four chunks per 64-byte block."); 160| 2.21k| auto zero = simd8{uint8_t(0)}; 161| 2.21k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 162| 2.21k| this->check_utf8_bytes(input.chunks[0], zero); 163| 2.21k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 164| | } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 165| | this->check_utf8_bytes(input.chunks[0], zero); 166| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 167| | this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 168| | this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 169| | } 170| 2.21k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 2.21k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 9, False: 2.20k] ------------------ 172| 9| return 0; // error 173| 9| } 174| 2.20k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 2.20k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 176| | // We process in blocks of up to 12 bytes except possibly 177| | // for fast paths which may process up to 16 bytes. For the 178| | // slow path to work, we should have at least 12 input bytes left. 179| 2.20k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 17.8k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 15.6k, False: 2.20k] ------------------ 182| | // Performance note: our ability to compute 'consumed' and 183| | // then shift and recompute is critical. If there is a 184| | // latency of, say, 4 cycles on getting 'consumed', then 185| | // the inner loop might have a total latency of about 6 cycles. 186| | // Yet we process between 6 to 12 inputs bytes, thus we get 187| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 188| | // for this section of the code. Hence, there is a limit 189| | // to how much we can further increase this latency before 190| | // it seriously harms performance. 191| 15.6k| size_t consumed = convert_masked_utf8_to_utf16( 192| 15.6k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 15.6k| pos += consumed; 194| 15.6k| utf8_end_of_code_point_mask >>= consumed; 195| 15.6k| } 196| | // At this point there may remain between 0 and 12 bytes in the 197| | // 64-byte block. These bytes will be processed again. So we have an 198| | // 80% efficiency (in the worst case). In practice we expect an 199| | // 85% to 90% efficiency. 200| 2.20k| } 201| 2.81k| } 202| 11.0k| if (errors()) { ------------------ | Branch (202:9): [True: 188, False: 10.8k] ------------------ 203| 188| return 0; 204| 188| } 205| 10.8k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.54k, False: 9.30k] ------------------ 206| 1.54k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.54k| in + pos, size - pos, utf16_output); 208| 1.54k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 403, False: 1.14k] ------------------ 209| 403| return 0; 210| 403| } 211| 1.14k| utf16_output += howmany; 212| 1.14k| } 213| 10.4k| return utf16_output - start; 214| 10.8k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE0EEEmPKcmPDs: 130| 11.0k| char16_t *utf16_output) { 131| 11.0k| size_t pos = 0; 132| 11.0k| char16_t *start{utf16_output}; 133| | // In the worst case, we have the haswell kernel which can cause an overflow 134| | // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the 135| | // last 16 bytes, and if the data is valid, then it is entirely safe because 136| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 137| | // generally assume that you have valid UTF-8 input, so we are going to go 138| | // back from the end counting 8 leading bytes, to give us a good margin. 139| 11.0k| size_t leading_byte = 0; 140| 11.0k| size_t margin = size; 141| 24.3k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 14.2k, False: 10.0k] | Branch (141:26): [True: 13.3k, False: 998] ------------------ 142| 13.3k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 13.3k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 11.0k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 15.1k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 4.10k, False: 11.0k] ------------------ 148| 4.10k| simd8x64 input(reinterpret_cast(in + pos)); 149| 4.10k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 974, False: 3.12k] ------------------ 150| 974| input.store_ascii_as_utf16(utf16_output); 151| 974| utf16_output += 64; 152| 974| pos += 64; 153| 3.12k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 3.12k| static_assert( 157| 3.12k| (simd8x64::NUM_CHUNKS == 2) || 158| 3.12k| (simd8x64::NUM_CHUNKS == 4), 159| 3.12k| "We support either two or four chunks per 64-byte block."); 160| 3.12k| auto zero = simd8{uint8_t(0)}; 161| | if constexpr (simd8x64::NUM_CHUNKS == 2) { 162| | this->check_utf8_bytes(input.chunks[0], zero); 163| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 164| 3.12k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 165| 3.12k| this->check_utf8_bytes(input.chunks[0], zero); 166| 3.12k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 167| 3.12k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 168| 3.12k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 169| 3.12k| } 170| 3.12k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 3.12k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 10, False: 3.11k] ------------------ 172| 10| return 0; // error 173| 10| } 174| 3.11k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 3.11k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 176| | // We process in blocks of up to 12 bytes except possibly 177| | // for fast paths which may process up to 16 bytes. For the 178| | // slow path to work, we should have at least 12 input bytes left. 179| 3.11k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 25.0k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 21.9k, False: 3.11k] ------------------ 182| | // Performance note: our ability to compute 'consumed' and 183| | // then shift and recompute is critical. If there is a 184| | // latency of, say, 4 cycles on getting 'consumed', then 185| | // the inner loop might have a total latency of about 6 cycles. 186| | // Yet we process between 6 to 12 inputs bytes, thus we get 187| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 188| | // for this section of the code. Hence, there is a limit 189| | // to how much we can further increase this latency before 190| | // it seriously harms performance. 191| 21.9k| size_t consumed = convert_masked_utf8_to_utf16( 192| 21.9k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 21.9k| pos += consumed; 194| 21.9k| utf8_end_of_code_point_mask >>= consumed; 195| 21.9k| } 196| | // At this point there may remain between 0 and 12 bytes in the 197| | // 64-byte block. These bytes will be processed again. So we have an 198| | // 80% efficiency (in the worst case). In practice we expect an 199| | // 85% to 90% efficiency. 200| 3.11k| } 201| 4.10k| } 202| 11.0k| if (errors()) { ------------------ | Branch (202:9): [True: 213, False: 10.7k] ------------------ 203| 213| return 0; 204| 213| } 205| 10.7k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.18k, False: 9.61k] ------------------ 206| 1.18k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.18k| in + pos, size - pos, utf16_output); 208| 1.18k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 197, False: 987] ------------------ 209| 197| return 0; 210| 197| } 211| 987| utf16_output += howmany; 212| 987| } 213| 10.6k| return utf16_output - start; 214| 10.7k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 25.6k| const simd8 prev_input) { 120| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 121| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 122| | // small negative numbers) 123| 25.6k| simd8 prev1 = input.prev<1>(prev_input); 124| 25.6k| simd8 sc = check_special_cases(input, prev1); 125| 25.6k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 25.6k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1619check_special_casesENS1_4simd5simd8IhEES5_: 8| 25.6k|check_special_cases(const simd8 input, const simd8 prev1) { 9| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 10| | // Bit 1 = Too Long (ASCII followed by continuation) 11| | // Bit 2 = Overlong 3-byte 12| | // Bit 4 = Surrogate 13| | // Bit 5 = Overlong 2-byte 14| | // Bit 7 = Two Continuations 15| 25.6k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 25.6k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 25.6k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 25.6k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 25.6k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 25.6k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 25.6k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 23| | // 11110100 101_____ 24| | // 11110101 1001____ 25| | // 11110101 101_____ 26| | // 1111011_ 1001____ 27| | // 1111011_ 101_____ 28| | // 11111___ 1001____ 29| | // 11111___ 101_____ 30| 25.6k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 25.6k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 25.6k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 25.6k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 25.6k| TOO_LONG, 40| | // 10______ ________ 41| 25.6k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 25.6k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 25.6k| TOO_SHORT, 46| | // 1110____ ________ 47| 25.6k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 25.6k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 25.6k| constexpr const uint8_t CARRY = 51| 25.6k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 25.6k| const simd8 byte_1_low = 53| 25.6k| (prev1 & 0x0F) 54| 25.6k| .lookup_16( 55| | // ____0000 ________ 56| 25.6k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 25.6k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 25.6k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 25.6k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 25.6k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 25.6k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 25.6k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 25.6k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 25.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 25.6k| OVERLONG_4, 88| | // ________ 1001____ 89| 25.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 25.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 25.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 25.6k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 25.6k| return (byte_1_high & byte_1_low & byte_2_high); 97| 25.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1623check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 25.6k| const simd8 sc) { 102| 25.6k| simd8 prev2 = input.prev<2>(prev_input); 103| 25.6k| simd8 prev3 = input.prev<3>(prev_input); 104| 25.6k| simd8 must23 = 105| 25.6k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 25.6k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 25.6k| return must23_80 ^ sc; 108| 25.6k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder6errorsEv: 324| 22.0k| simdutf_really_inline bool errors() const { 325| 22.0k| return this->error.any_bits_set_anywhere(); 326| 22.0k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE1EEEmPKcmPDs: 130| 11.0k| char16_t *utf16_output) { 131| 11.0k| size_t pos = 0; 132| 11.0k| char16_t *start{utf16_output}; 133| | // In the worst case, we have the haswell kernel which can cause an overflow 134| | // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the 135| | // last 16 bytes, and if the data is valid, then it is entirely safe because 136| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 137| | // generally assume that you have valid UTF-8 input, so we are going to go 138| | // back from the end counting 8 leading bytes, to give us a good margin. 139| 11.0k| size_t leading_byte = 0; 140| 11.0k| size_t margin = size; 141| 24.3k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 14.3k, False: 10.0k] | Branch (141:26): [True: 13.3k, False: 995] ------------------ 142| 13.3k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 13.3k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 11.0k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 15.2k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 4.24k, False: 11.0k] ------------------ 148| 4.24k| simd8x64 input(reinterpret_cast(in + pos)); 149| 4.24k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 973, False: 3.27k] ------------------ 150| 973| input.store_ascii_as_utf16(utf16_output); 151| 973| utf16_output += 64; 152| 973| pos += 64; 153| 3.27k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 3.27k| static_assert( 157| 3.27k| (simd8x64::NUM_CHUNKS == 2) || 158| 3.27k| (simd8x64::NUM_CHUNKS == 4), 159| 3.27k| "We support either two or four chunks per 64-byte block."); 160| 3.27k| auto zero = simd8{uint8_t(0)}; 161| | if constexpr (simd8x64::NUM_CHUNKS == 2) { 162| | this->check_utf8_bytes(input.chunks[0], zero); 163| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 164| 3.27k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 165| 3.27k| this->check_utf8_bytes(input.chunks[0], zero); 166| 3.27k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 167| 3.27k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 168| 3.27k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 169| 3.27k| } 170| 3.27k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 3.27k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 10, False: 3.26k] ------------------ 172| 10| return 0; // error 173| 10| } 174| 3.26k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 3.26k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 176| | // We process in blocks of up to 12 bytes except possibly 177| | // for fast paths which may process up to 16 bytes. For the 178| | // slow path to work, we should have at least 12 input bytes left. 179| 3.26k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 25.8k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 22.6k, False: 3.26k] ------------------ 182| | // Performance note: our ability to compute 'consumed' and 183| | // then shift and recompute is critical. If there is a 184| | // latency of, say, 4 cycles on getting 'consumed', then 185| | // the inner loop might have a total latency of about 6 cycles. 186| | // Yet we process between 6 to 12 inputs bytes, thus we get 187| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 188| | // for this section of the code. Hence, there is a limit 189| | // to how much we can further increase this latency before 190| | // it seriously harms performance. 191| 22.6k| size_t consumed = convert_masked_utf8_to_utf16( 192| 22.6k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 22.6k| pos += consumed; 194| 22.6k| utf8_end_of_code_point_mask >>= consumed; 195| 22.6k| } 196| | // At this point there may remain between 0 and 12 bytes in the 197| | // 64-byte block. These bytes will be processed again. So we have an 198| | // 80% efficiency (in the worst case). In practice we expect an 199| | // 85% to 90% efficiency. 200| 3.26k| } 201| 4.24k| } 202| 11.0k| if (errors()) { ------------------ | Branch (202:9): [True: 213, False: 10.8k] ------------------ 203| 213| return 0; 204| 213| } 205| 10.8k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.17k, False: 9.63k] ------------------ 206| 1.17k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.17k| in + pos, size - pos, utf16_output); 208| 1.17k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 197, False: 980] ------------------ 209| 197| return 0; 210| 197| } 211| 980| utf16_output += howmany; 212| 980| } 213| 10.6k| return utf16_output - start; 214| 10.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoderC2Ev: 114| 5.56k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoder7convertEPKcmPDi: 129| 5.56k| char32_t *utf32_output) { 130| 5.56k| size_t pos = 0; 131| 5.56k| char32_t *start{utf32_output}; 132| | // In the worst case, we have the haswell kernel which can cause an overflow 133| | // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the 134| | // last 16 bytes, and if the data is valid, then it is entirely safe because 135| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 136| | // generally assume that you have valid UTF-8 input, so we are going to go 137| | // back from the end counting 16 leading bytes, to give us a good margin. 138| 5.56k| size_t leading_byte = 0; 139| 5.56k| size_t margin = size; 140| 18.9k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (140:12): [True: 14.2k, False: 4.65k] | Branch (140:26): [True: 13.3k, False: 903] ------------------ 141| 13.3k| leading_byte += (int8_t(in[margin - 1]) > -65); 142| 13.3k| } 143| | // If the input is long enough, then we have that margin-1 is the fourth 144| | // last leading byte. 145| 5.56k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 146| 7.22k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (146:12): [True: 1.67k, False: 5.55k] ------------------ 147| 1.67k| simd8x64 input(reinterpret_cast(in + pos)); 148| 1.67k| if (input.is_ascii()) { ------------------ | Branch (148:11): [True: 583, False: 1.09k] ------------------ 149| 583| input.store_ascii_as_utf32(utf32_output); 150| 583| utf32_output += 64; 151| 583| pos += 64; 152| 1.09k| } else { 153| | // you might think that a for-loop would work, but under Visual Studio, 154| | // it is not good enough. 155| 1.09k| static_assert( 156| 1.09k| (simd8x64::NUM_CHUNKS == 2) || 157| 1.09k| (simd8x64::NUM_CHUNKS == 4), 158| 1.09k| "We support either two or four chunks per 64-byte block."); 159| 1.09k| auto zero = simd8{uint8_t(0)}; 160| 1.09k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 161| 1.09k| this->check_utf8_bytes(input.chunks[0], zero); 162| 1.09k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 163| | } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 164| | this->check_utf8_bytes(input.chunks[0], zero); 165| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 166| | this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 167| | this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 168| | } 169| 1.09k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 170| 1.09k| if (utf8_continuation_mask & 1) { ------------------ | Branch (170:13): [True: 9, False: 1.08k] ------------------ 171| 9| return 0; // we have an error 172| 9| } 173| 1.08k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 174| 1.08k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 175| | // We process in blocks of up to 12 bytes except possibly 176| | // for fast paths which may process up to 16 bytes. For the 177| | // slow path to work, we should have at least 12 input bytes left. 178| 1.08k| size_t max_starting_point = (pos + 64) - 12; 179| | // Next loop is going to run at least five times. 180| 10.8k| while (pos < max_starting_point) { ------------------ | Branch (180:16): [True: 9.80k, False: 1.08k] ------------------ 181| | // Performance note: our ability to compute 'consumed' and 182| | // then shift and recompute is critical. If there is a 183| | // latency of, say, 4 cycles on getting 'consumed', then 184| | // the inner loop might have a total latency of about 6 cycles. 185| | // Yet we process between 6 to 12 inputs bytes, thus we get 186| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 187| | // for this section of the code. Hence, there is a limit 188| | // to how much we can further increase this latency before 189| | // it seriously harms performance. 190| 9.80k| size_t consumed = convert_masked_utf8_to_utf32( 191| 9.80k| in + pos, utf8_end_of_code_point_mask, utf32_output); 192| 9.80k| pos += consumed; 193| 9.80k| utf8_end_of_code_point_mask >>= consumed; 194| 9.80k| } 195| | // At this point there may remain between 0 and 12 bytes in the 196| | // 64-byte block. These bytes will be processed again. So we have an 197| | // 80% efficiency (in the worst case). In practice we expect an 198| | // 85% to 90% efficiency. 199| 1.08k| } 200| 1.67k| } 201| 5.55k| if (errors()) { ------------------ | Branch (201:9): [True: 188, False: 5.36k] ------------------ 202| 188| return 0; 203| 188| } 204| 5.36k| if (pos < size) { ------------------ | Branch (204:9): [True: 1.26k, False: 4.10k] ------------------ 205| 1.26k| size_t howmany = 206| 1.26k| scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output); 207| 1.26k| if (howmany == 0) { ------------------ | Branch (207:11): [True: 403, False: 861] ------------------ 208| 403| return 0; 209| 403| } 210| 861| utf32_output += howmany; 211| 861| } 212| 4.96k| return utf32_output - start; 213| 5.36k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 2.18k| const simd8 prev_input) { 120| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 121| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 122| | // small negative numbers) 123| 2.18k| simd8 prev1 = input.prev<1>(prev_input); 124| 2.18k| simd8 sc = check_special_cases(input, prev1); 125| 2.18k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 2.18k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3219check_special_casesENS1_4simd5simd8IhEES5_: 8| 2.18k|check_special_cases(const simd8 input, const simd8 prev1) { 9| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 10| | // Bit 1 = Too Long (ASCII followed by continuation) 11| | // Bit 2 = Overlong 3-byte 12| | // Bit 4 = Surrogate 13| | // Bit 5 = Overlong 2-byte 14| | // Bit 7 = Two Continuations 15| 2.18k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 2.18k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 2.18k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 2.18k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 2.18k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 2.18k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 2.18k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 23| | // 11110100 101_____ 24| | // 11110101 1001____ 25| | // 11110101 101_____ 26| | // 1111011_ 1001____ 27| | // 1111011_ 101_____ 28| | // 11111___ 1001____ 29| | // 11111___ 101_____ 30| 2.18k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 2.18k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 2.18k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 2.18k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 2.18k| TOO_LONG, 40| | // 10______ ________ 41| 2.18k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 2.18k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 2.18k| TOO_SHORT, 46| | // 1110____ ________ 47| 2.18k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 2.18k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 2.18k| constexpr const uint8_t CARRY = 51| 2.18k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 2.18k| const simd8 byte_1_low = 53| 2.18k| (prev1 & 0x0F) 54| 2.18k| .lookup_16( 55| | // ____0000 ________ 56| 2.18k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 2.18k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 2.18k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 2.18k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 2.18k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 2.18k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 2.18k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 2.18k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 2.18k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 2.18k| OVERLONG_4, 88| | // ________ 1001____ 89| 2.18k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 2.18k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 2.18k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 2.18k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 2.18k| return (byte_1_high & byte_1_low & byte_2_high); 97| 2.18k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3223check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 2.18k| const simd8 sc) { 102| 2.18k| simd8 prev2 = input.prev<2>(prev_input); 103| 2.18k| simd8 prev3 = input.prev<3>(prev_input); 104| 2.18k| simd8 must23 = 105| 2.18k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 2.18k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 2.18k| return must23_80 ^ sc; 108| 2.18k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoder6errorsEv: 310| 5.55k| simdutf_really_inline bool errors() const { 311| 5.55k| return this->error.any_bits_set_anywhere(); 312| 5.55k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoderC2Ev: 114| 5.56k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoder7convertEPKcmPDi: 129| 5.56k| char32_t *utf32_output) { 130| 5.56k| size_t pos = 0; 131| 5.56k| char32_t *start{utf32_output}; 132| | // In the worst case, we have the haswell kernel which can cause an overflow 133| | // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the 134| | // last 16 bytes, and if the data is valid, then it is entirely safe because 135| | // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot 136| | // generally assume that you have valid UTF-8 input, so we are going to go 137| | // back from the end counting 16 leading bytes, to give us a good margin. 138| 5.56k| size_t leading_byte = 0; 139| 5.56k| size_t margin = size; 140| 14.4k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (140:12): [True: 9.71k, False: 4.75k] | Branch (140:26): [True: 8.90k, False: 811] ------------------ 141| 8.90k| leading_byte += (int8_t(in[margin - 1]) > -65); 142| 8.90k| } 143| | // If the input is long enough, then we have that margin-1 is the fourth 144| | // last leading byte. 145| 5.56k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 146| 8.26k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (146:12): [True: 2.71k, False: 5.55k] ------------------ 147| 2.71k| simd8x64 input(reinterpret_cast(in + pos)); 148| 2.71k| if (input.is_ascii()) { ------------------ | Branch (148:11): [True: 944, False: 1.76k] ------------------ 149| 944| input.store_ascii_as_utf32(utf32_output); 150| 944| utf32_output += 64; 151| 944| pos += 64; 152| 1.76k| } else { 153| | // you might think that a for-loop would work, but under Visual Studio, 154| | // it is not good enough. 155| 1.76k| static_assert( 156| 1.76k| (simd8x64::NUM_CHUNKS == 2) || 157| 1.76k| (simd8x64::NUM_CHUNKS == 4), 158| 1.76k| "We support either two or four chunks per 64-byte block."); 159| 1.76k| auto zero = simd8{uint8_t(0)}; 160| | if constexpr (simd8x64::NUM_CHUNKS == 2) { 161| | this->check_utf8_bytes(input.chunks[0], zero); 162| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 163| 1.76k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 164| 1.76k| this->check_utf8_bytes(input.chunks[0], zero); 165| 1.76k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 166| 1.76k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 167| 1.76k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 168| 1.76k| } 169| 1.76k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 170| 1.76k| if (utf8_continuation_mask & 1) { ------------------ | Branch (170:13): [True: 10, False: 1.75k] ------------------ 171| 10| return 0; // we have an error 172| 10| } 173| 1.75k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 174| 1.75k| uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1; 175| | // We process in blocks of up to 12 bytes except possibly 176| | // for fast paths which may process up to 16 bytes. For the 177| | // slow path to work, we should have at least 12 input bytes left. 178| 1.75k| size_t max_starting_point = (pos + 64) - 12; 179| | // Next loop is going to run at least five times. 180| 16.6k| while (pos < max_starting_point) { ------------------ | Branch (180:16): [True: 14.9k, False: 1.75k] ------------------ 181| | // Performance note: our ability to compute 'consumed' and 182| | // then shift and recompute is critical. If there is a 183| | // latency of, say, 4 cycles on getting 'consumed', then 184| | // the inner loop might have a total latency of about 6 cycles. 185| | // Yet we process between 6 to 12 inputs bytes, thus we get 186| | // a speed limit between 1 cycle/byte and 0.5 cycle/byte 187| | // for this section of the code. Hence, there is a limit 188| | // to how much we can further increase this latency before 189| | // it seriously harms performance. 190| 14.9k| size_t consumed = convert_masked_utf8_to_utf32( 191| 14.9k| in + pos, utf8_end_of_code_point_mask, utf32_output); 192| 14.9k| pos += consumed; 193| 14.9k| utf8_end_of_code_point_mask >>= consumed; 194| 14.9k| } 195| | // At this point there may remain between 0 and 12 bytes in the 196| | // 64-byte block. These bytes will be processed again. So we have an 197| | // 80% efficiency (in the worst case). In practice we expect an 198| | // 85% to 90% efficiency. 199| 1.75k| } 200| 2.71k| } 201| 5.55k| if (errors()) { ------------------ | Branch (201:9): [True: 213, False: 5.33k] ------------------ 202| 213| return 0; 203| 213| } 204| 5.33k| if (pos < size) { ------------------ | Branch (204:9): [True: 940, False: 4.39k] ------------------ 205| 940| size_t howmany = 206| 940| scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output); 207| 940| if (howmany == 0) { ------------------ | Branch (207:11): [True: 197, False: 743] ------------------ 208| 197| return 0; 209| 197| } 210| 743| utf32_output += howmany; 211| 743| } 212| 5.14k| return utf32_output - start; 213| 5.33k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 7.06k| const simd8 prev_input) { 120| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 121| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 122| | // small negative numbers) 123| 7.06k| simd8 prev1 = input.prev<1>(prev_input); 124| 7.06k| simd8 sc = check_special_cases(input, prev1); 125| 7.06k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 7.06k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3219check_special_casesENS1_4simd5simd8IhEES5_: 8| 7.06k|check_special_cases(const simd8 input, const simd8 prev1) { 9| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 10| | // Bit 1 = Too Long (ASCII followed by continuation) 11| | // Bit 2 = Overlong 3-byte 12| | // Bit 4 = Surrogate 13| | // Bit 5 = Overlong 2-byte 14| | // Bit 7 = Two Continuations 15| 7.06k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 7.06k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 7.06k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 7.06k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 7.06k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 7.06k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 7.06k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 23| | // 11110100 101_____ 24| | // 11110101 1001____ 25| | // 11110101 101_____ 26| | // 1111011_ 1001____ 27| | // 1111011_ 101_____ 28| | // 11111___ 1001____ 29| | // 11111___ 101_____ 30| 7.06k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 7.06k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 7.06k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 7.06k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 7.06k| TOO_LONG, 40| | // 10______ ________ 41| 7.06k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 7.06k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 7.06k| TOO_SHORT, 46| | // 1110____ ________ 47| 7.06k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 7.06k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 7.06k| constexpr const uint8_t CARRY = 51| 7.06k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 7.06k| const simd8 byte_1_low = 53| 7.06k| (prev1 & 0x0F) 54| 7.06k| .lookup_16( 55| | // ____0000 ________ 56| 7.06k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 7.06k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 7.06k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 7.06k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 7.06k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 7.06k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 7.06k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 7.06k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 7.06k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 7.06k| OVERLONG_4, 88| | // ________ 1001____ 89| 7.06k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 7.06k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 7.06k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 7.06k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 7.06k| return (byte_1_high & byte_1_low & byte_2_high); 97| 7.06k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3223check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 7.06k| const simd8 sc) { 102| 7.06k| simd8 prev2 = input.prev<2>(prev_input); 103| 7.06k| simd8 prev3 = input.prev<3>(prev_input); 104| 7.06k| simd8 must23 = 105| 7.06k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 7.06k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 7.06k| return must23_80 ^ sc; 108| 7.06k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoder6errorsEv: 310| 5.55k| simdutf_really_inline bool errors() const { 311| 5.55k| return this->error.any_bits_set_anywhere(); 312| 5.55k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker16check_next_inputERKNS1_4simd8simd8x64IhEE: 187| 34.5k| simdutf_really_inline void check_next_input(const simd8x64 &input) { 188| 34.5k| if (simdutf_likely(is_ascii(input))) { ------------------ | | 89| 34.5k| #define simdutf_likely(x) __builtin_expect(!!(x), 1) | | ------------------ | | | Branch (89:31): [True: 26.7k, False: 7.82k] | | ------------------ ------------------ 189| 26.7k| this->error |= this->prev_incomplete; 190| 26.7k| } else { 191| | // you might think that a for-loop would work, but under Visual Studio, it 192| | // is not good enough. 193| 7.82k| static_assert((simd8x64::NUM_CHUNKS == 2) || 194| 7.82k| (simd8x64::NUM_CHUNKS == 4), 195| 7.82k| "We support either two or four chunks per 64-byte block."); 196| 7.82k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 197| 7.82k| this->check_utf8_bytes(input.chunks[0], this->prev_input_block); 198| 7.82k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 199| | } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 200| | this->check_utf8_bytes(input.chunks[0], this->prev_input_block); 201| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 202| | this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 203| | this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 204| | } 205| 7.82k| this->prev_incomplete = 206| 7.82k| is_incomplete(input.chunks[simd8x64::NUM_CHUNKS - 1]); 207| 7.82k| this->prev_input_block = input.chunks[simd8x64::NUM_CHUNKS - 1]; 208| 7.82k| } 209| 34.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker16check_utf8_bytesENS1_4simd5simd8IhEES6_: 169| 15.6k| const simd8 prev_input) { 170| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 171| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 172| | // small negative numbers) 173| 15.6k| simd8 prev1 = input.prev<1>(prev_input); 174| 15.6k| simd8 sc = check_special_cases(input, prev1); 175| 15.6k| this->error |= check_multibyte_lengths(input, prev_input, sc); 176| 15.6k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation19check_special_casesENS1_4simd5simd8IhEES5_: 9| 15.6k|check_special_cases(const simd8 input, const simd8 prev1) { 10| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 11| | // Bit 1 = Too Long (ASCII followed by continuation) 12| | // Bit 2 = Overlong 3-byte 13| | // Bit 4 = Surrogate 14| | // Bit 5 = Overlong 2-byte 15| | // Bit 7 = Two Continuations 16| 15.6k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 17| | // 11______ 11______ 18| 15.6k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 19| 15.6k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 20| 15.6k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 21| 15.6k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 22| 15.6k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 23| 15.6k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 24| | // 11110100 101_____ 25| | // 11110101 1001____ 26| | // 11110101 101_____ 27| | // 1111011_ 1001____ 28| | // 1111011_ 101_____ 29| | // 11111___ 1001____ 30| | // 11111___ 101_____ 31| 15.6k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 32| | // 11110101 1000____ 33| | // 1111011_ 1000____ 34| | // 11111___ 1000____ 35| 15.6k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 36| | 37| 15.6k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 38| | // 0_______ ________ 39| 15.6k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 40| 15.6k| TOO_LONG, 41| | // 10______ ________ 42| 15.6k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 43| | // 1100____ ________ 44| 15.6k| TOO_SHORT | OVERLONG_2, 45| | // 1101____ ________ 46| 15.6k| TOO_SHORT, 47| | // 1110____ ________ 48| 15.6k| TOO_SHORT | OVERLONG_3 | SURROGATE, 49| | // 1111____ ________ 50| 15.6k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 51| 15.6k| constexpr const uint8_t CARRY = 52| 15.6k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 53| 15.6k| const simd8 byte_1_low = 54| 15.6k| (prev1 & 0x0F) 55| 15.6k| .lookup_16( 56| | // ____0000 ________ 57| 15.6k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 58| | // ____0001 ________ 59| 15.6k| CARRY | OVERLONG_2, 60| | // ____001_ ________ 61| 15.6k| CARRY, CARRY, 62| | 63| | // ____0100 ________ 64| 15.6k| CARRY | TOO_LARGE, 65| | // ____0101 ________ 66| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 67| | // ____011_ ________ 68| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 70| | 71| | // ____1___ ________ 72| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 77| | // ____1101 ________ 78| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 79| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000, 80| 15.6k| CARRY | TOO_LARGE | TOO_LARGE_1000); 81| 15.6k| const simd8 byte_2_high = input.shr<4>().lookup_16( 82| | // ________ 0_______ 83| 15.6k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 84| 15.6k| TOO_SHORT, TOO_SHORT, 85| | 86| | // ________ 1000____ 87| 15.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 88| 15.6k| OVERLONG_4, 89| | // ________ 1001____ 90| 15.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 91| | // ________ 101_____ 92| 15.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| 15.6k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 94| | 95| | // ________ 11______ 96| 15.6k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 97| 15.6k| return (byte_1_high & byte_1_low & byte_2_high); 98| 15.6k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation23check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 102| 15.6k| const simd8 sc) { 103| 15.6k| simd8 prev2 = input.prev<2>(prev_input); 104| 15.6k| simd8 prev3 = input.prev<3>(prev_input); 105| 15.6k| simd8 must23 = 106| 15.6k| simd8(must_be_2_3_continuation(prev2, prev3)); 107| 15.6k| simd8 must23_80 = must23 & uint8_t(0x80); 108| 15.6k| return must23_80 ^ sc; 109| 15.6k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation13is_incompleteENS1_4simd5simd8IhEE: 115| 7.82k|simdutf_really_inline simd8 is_incomplete(const simd8 input) { 116| | // If the previous input's last 3 bytes match this, they're too short (they 117| | // ended at EOF): 118| | // ... 1111____ 111_____ 11______ 119| 7.82k| static const uint8_t max_array[32] = {255, 120| 7.82k| 255, 121| 7.82k| 255, 122| 7.82k| 255, 123| 7.82k| 255, 124| 7.82k| 255, 125| 7.82k| 255, 126| 7.82k| 255, 127| 7.82k| 255, 128| 7.82k| 255, 129| 7.82k| 255, 130| 7.82k| 255, 131| 7.82k| 255, 132| 7.82k| 255, 133| 7.82k| 255, 134| 7.82k| 255, 135| 7.82k| 255, 136| 7.82k| 255, 137| 7.82k| 255, 138| 7.82k| 255, 139| 7.82k| 255, 140| 7.82k| 255, 141| 7.82k| 255, 142| 7.82k| 255, 143| 7.82k| 255, 144| 7.82k| 255, 145| 7.82k| 255, 146| 7.82k| 255, 147| 7.82k| 255, 148| 7.82k| 0b11110000u - 1, 149| 7.82k| 0b11100000u - 1, 150| 7.82k| 0b11000000u - 1}; 151| 7.82k| const simd8 max_value( 152| 7.82k| &max_array[sizeof(max_array) - sizeof(simd8)]); 153| 7.82k| return input.gt_bits(max_value); 154| 7.82k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker9check_eofEv: 181| 27.3k| simdutf_really_inline void check_eof() { 182| | // If the previous block had incomplete UTF-8 characters at the end, an 183| | // ASCII block can't possibly finish them. 184| 27.3k| this->error |= this->prev_incomplete; 185| 27.3k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker6errorsEv: 212| 28.5k| simdutf_really_inline bool errors() const { 213| 28.5k| return this->error.any_bits_set_anywhere(); 214| 28.5k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker16check_next_inputERKNS1_4simd8simd8x64IhEE: 187| 37.2k| simdutf_really_inline void check_next_input(const simd8x64 &input) { 188| 37.2k| if (simdutf_likely(is_ascii(input))) { ------------------ | | 89| 37.2k| #define simdutf_likely(x) __builtin_expect(!!(x), 1) | | ------------------ | | | Branch (89:31): [True: 28.2k, False: 9.01k] | | ------------------ ------------------ 189| 28.2k| this->error |= this->prev_incomplete; 190| 28.2k| } else { 191| | // you might think that a for-loop would work, but under Visual Studio, it 192| | // is not good enough. 193| 9.01k| static_assert((simd8x64::NUM_CHUNKS == 2) || 194| 9.01k| (simd8x64::NUM_CHUNKS == 4), 195| 9.01k| "We support either two or four chunks per 64-byte block."); 196| | if constexpr (simd8x64::NUM_CHUNKS == 2) { 197| | this->check_utf8_bytes(input.chunks[0], this->prev_input_block); 198| | this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 199| 9.01k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 200| 9.01k| this->check_utf8_bytes(input.chunks[0], this->prev_input_block); 201| 9.01k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 202| 9.01k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 203| 9.01k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 204| 9.01k| } 205| 9.01k| this->prev_incomplete = 206| 9.01k| is_incomplete(input.chunks[simd8x64::NUM_CHUNKS - 1]); 207| 9.01k| this->prev_input_block = input.chunks[simd8x64::NUM_CHUNKS - 1]; 208| 9.01k| } 209| 37.2k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker16check_utf8_bytesENS1_4simd5simd8IhEES6_: 169| 36.0k| const simd8 prev_input) { 170| | // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ 171| | // lead bytes (2, 3, 4-byte leads become large positive numbers instead of 172| | // small negative numbers) 173| 36.0k| simd8 prev1 = input.prev<1>(prev_input); 174| 36.0k| simd8 sc = check_special_cases(input, prev1); 175| 36.0k| this->error |= check_multibyte_lengths(input, prev_input, sc); 176| 36.0k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation19check_special_casesENS1_4simd5simd8IhEES5_: 9| 36.0k|check_special_cases(const simd8 input, const simd8 prev1) { 10| | // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII) 11| | // Bit 1 = Too Long (ASCII followed by continuation) 12| | // Bit 2 = Overlong 3-byte 13| | // Bit 4 = Surrogate 14| | // Bit 5 = Overlong 2-byte 15| | // Bit 7 = Two Continuations 16| 36.0k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 17| | // 11______ 11______ 18| 36.0k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 19| 36.0k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 20| 36.0k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 21| 36.0k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 22| 36.0k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 23| 36.0k| constexpr const uint8_t TOO_LARGE = 1 << 3; // 11110100 1001____ 24| | // 11110100 101_____ 25| | // 11110101 1001____ 26| | // 11110101 101_____ 27| | // 1111011_ 1001____ 28| | // 1111011_ 101_____ 29| | // 11111___ 1001____ 30| | // 11111___ 101_____ 31| 36.0k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 32| | // 11110101 1000____ 33| | // 1111011_ 1000____ 34| | // 11111___ 1000____ 35| 36.0k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 36| | 37| 36.0k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 38| | // 0_______ ________ 39| 36.0k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 40| 36.0k| TOO_LONG, 41| | // 10______ ________ 42| 36.0k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 43| | // 1100____ ________ 44| 36.0k| TOO_SHORT | OVERLONG_2, 45| | // 1101____ ________ 46| 36.0k| TOO_SHORT, 47| | // 1110____ ________ 48| 36.0k| TOO_SHORT | OVERLONG_3 | SURROGATE, 49| | // 1111____ ________ 50| 36.0k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 51| 36.0k| constexpr const uint8_t CARRY = 52| 36.0k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 53| 36.0k| const simd8 byte_1_low = 54| 36.0k| (prev1 & 0x0F) 55| 36.0k| .lookup_16( 56| | // ____0000 ________ 57| 36.0k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 58| | // ____0001 ________ 59| 36.0k| CARRY | OVERLONG_2, 60| | // ____001_ ________ 61| 36.0k| CARRY, CARRY, 62| | 63| | // ____0100 ________ 64| 36.0k| CARRY | TOO_LARGE, 65| | // ____0101 ________ 66| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 67| | // ____011_ ________ 68| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 70| | 71| | // ____1___ ________ 72| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 77| | // ____1101 ________ 78| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 79| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000, 80| 36.0k| CARRY | TOO_LARGE | TOO_LARGE_1000); 81| 36.0k| const simd8 byte_2_high = input.shr<4>().lookup_16( 82| | // ________ 0_______ 83| 36.0k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 84| 36.0k| TOO_SHORT, TOO_SHORT, 85| | 86| | // ________ 1000____ 87| 36.0k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 88| 36.0k| OVERLONG_4, 89| | // ________ 1001____ 90| 36.0k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 91| | // ________ 101_____ 92| 36.0k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| 36.0k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 94| | 95| | // ________ 11______ 96| 36.0k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 97| 36.0k| return (byte_1_high & byte_1_low & byte_2_high); 98| 36.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation23check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 102| 36.0k| const simd8 sc) { 103| 36.0k| simd8 prev2 = input.prev<2>(prev_input); 104| 36.0k| simd8 prev3 = input.prev<3>(prev_input); 105| 36.0k| simd8 must23 = 106| 36.0k| simd8(must_be_2_3_continuation(prev2, prev3)); 107| 36.0k| simd8 must23_80 = must23 & uint8_t(0x80); 108| 36.0k| return must23_80 ^ sc; 109| 36.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation13is_incompleteENS1_4simd5simd8IhEE: 115| 9.01k|simdutf_really_inline simd8 is_incomplete(const simd8 input) { 116| | // If the previous input's last 3 bytes match this, they're too short (they 117| | // ended at EOF): 118| | // ... 1111____ 111_____ 11______ 119| 9.01k| static const uint8_t max_array[32] = {255, 120| 9.01k| 255, 121| 9.01k| 255, 122| 9.01k| 255, 123| 9.01k| 255, 124| 9.01k| 255, 125| 9.01k| 255, 126| 9.01k| 255, 127| 9.01k| 255, 128| 9.01k| 255, 129| 9.01k| 255, 130| 9.01k| 255, 131| 9.01k| 255, 132| 9.01k| 255, 133| 9.01k| 255, 134| 9.01k| 255, 135| 9.01k| 255, 136| 9.01k| 255, 137| 9.01k| 255, 138| 9.01k| 255, 139| 9.01k| 255, 140| 9.01k| 255, 141| 9.01k| 255, 142| 9.01k| 255, 143| 9.01k| 255, 144| 9.01k| 255, 145| 9.01k| 255, 146| 9.01k| 255, 147| 9.01k| 255, 148| 9.01k| 0b11110000u - 1, 149| 9.01k| 0b11100000u - 1, 150| 9.01k| 0b11000000u - 1}; 151| 9.01k| const simd8 max_value( 152| 9.01k| &max_array[sizeof(max_array) - sizeof(simd8)]); 153| 9.01k| return input.gt_bits(max_value); 154| 9.01k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker9check_eofEv: 181| 27.3k| simdutf_really_inline void check_eof() { 182| | // If the previous block had incomplete UTF-8 characters at the end, an 183| | // ASCII block can't possibly finish them. 184| 27.3k| this->error |= this->prev_incomplete; 185| 27.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker6errorsEv: 212| 29.0k| simdutf_really_inline bool errors() const { 213| 29.0k| return this->error.any_bits_set_anywhere(); 214| 29.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation21generic_validate_utf8EPKcm: 27| 22.0k|bool generic_validate_utf8(const char *input, size_t length) { 28| 22.0k| return generic_validate_utf8( 29| 22.0k| reinterpret_cast(input), length); 30| 22.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation21generic_validate_utf8INS2_12utf8_checkerEEEbPKhm: 10| 22.0k|bool generic_validate_utf8(const uint8_t *input, size_t length) { 11| 22.0k| checker c{}; 12| 22.0k| buf_block_reader<64> reader(input, length); 13| 28.1k| while (reader.has_full_block()) { ------------------ | Branch (13:10): [True: 6.08k, False: 22.0k] ------------------ 14| 6.08k| simd::simd8x64 in(reader.full_block()); 15| 6.08k| c.check_next_input(in); 16| 6.08k| reader.advance(); 17| 6.08k| } 18| 22.0k| uint8_t block[64]{}; 19| 22.0k| reader.get_remainder(block); 20| 22.0k| simd::simd8x64 in(block); 21| 22.0k| c.check_next_input(in); 22| 22.0k| reader.advance(); 23| 22.0k| c.check_eof(); 24| 22.0k| return !c.errors(); 25| 22.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsEPKcm: 76| 5.56k|result generic_validate_utf8_with_errors(const char *input, size_t length) { 77| 5.56k| return generic_validate_utf8_with_errors( 78| 5.56k| reinterpret_cast(input), length); 79| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsINS2_12utf8_checkerEEENS_6resultEPKhm: 36| 5.56k|result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) { 37| 5.56k| checker c{}; 38| 5.56k| buf_block_reader<64> reader(input, length); 39| 5.56k| size_t count{0}; 40| 6.43k| while (reader.has_full_block()) { ------------------ | Branch (40:10): [True: 1.12k, False: 5.31k] ------------------ 41| 1.12k| simd::simd8x64 in(reader.full_block()); 42| 1.12k| c.check_next_input(in); 43| 1.12k| if (c.errors()) { ------------------ | Branch (43:9): [True: 250, False: 871] ------------------ 44| 250| if (count != 0) { ------------------ | Branch (44:11): [True: 13, False: 237] ------------------ 45| 13| count--; 46| 13| } // Sometimes the error is only detected in the next chunk 47| 250| result res = scalar::utf8::rewind_and_validate_with_errors( 48| 250| reinterpret_cast(input), 49| 250| reinterpret_cast(input + count), length - count); 50| 250| res.count += count; 51| 250| return res; 52| 250| } 53| 871| reader.advance(); 54| 871| count += 64; 55| 871| } 56| 5.31k| uint8_t block[64]{}; 57| 5.31k| reader.get_remainder(block); 58| 5.31k| simd::simd8x64 in(block); 59| 5.31k| c.check_next_input(in); 60| 5.31k| reader.advance(); 61| 5.31k| c.check_eof(); 62| 5.31k| if (c.errors()) { ------------------ | Branch (62:7): [True: 350, False: 4.96k] ------------------ 63| 350| if (count != 0) { ------------------ | Branch (63:9): [True: 13, False: 337] ------------------ 64| 13| count--; 65| 13| } // Sometimes the error is only detected in the next chunk 66| 350| result res = scalar::utf8::rewind_and_validate_with_errors( 67| 350| reinterpret_cast(input), 68| 350| reinterpret_cast(input) + count, length - count); 69| 350| res.count += count; 70| 350| return res; 71| 4.96k| } else { 72| 4.96k| return result(error_code::SUCCESS, length); 73| 4.96k| } 74| 5.31k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation21generic_validate_utf8EPKcm: 27| 22.0k|bool generic_validate_utf8(const char *input, size_t length) { 28| 22.0k| return generic_validate_utf8( 29| 22.0k| reinterpret_cast(input), length); 30| 22.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation21generic_validate_utf8INS2_12utf8_checkerEEEbPKhm: 10| 22.0k|bool generic_validate_utf8(const uint8_t *input, size_t length) { 11| 22.0k| checker c{}; 12| 22.0k| buf_block_reader<64> reader(input, length); 13| 30.2k| while (reader.has_full_block()) { ------------------ | Branch (13:10): [True: 8.19k, False: 22.0k] ------------------ 14| 8.19k| simd::simd8x64 in(reader.full_block()); 15| 8.19k| c.check_next_input(in); 16| 8.19k| reader.advance(); 17| 8.19k| } 18| 22.0k| uint8_t block[64]{}; 19| 22.0k| reader.get_remainder(block); 20| 22.0k| simd::simd8x64 in(block); 21| 22.0k| c.check_next_input(in); 22| 22.0k| reader.advance(); 23| 22.0k| c.check_eof(); 24| 22.0k| return !c.errors(); 25| 22.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsEPKcm: 76| 5.56k|result generic_validate_utf8_with_errors(const char *input, size_t length) { 77| 5.56k| return generic_validate_utf8_with_errors( 78| 5.56k| reinterpret_cast(input), length); 79| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsINS2_12utf8_checkerEEENS_6resultEPKhm: 36| 5.56k|result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) { 37| 5.56k| checker c{}; 38| 5.56k| buf_block_reader<64> reader(input, length); 39| 5.56k| size_t count{0}; 40| 7.03k| while (reader.has_full_block()) { ------------------ | Branch (40:10): [True: 1.70k, False: 5.32k] ------------------ 41| 1.70k| simd::simd8x64 in(reader.full_block()); 42| 1.70k| c.check_next_input(in); 43| 1.70k| if (c.errors()) { ------------------ | Branch (43:9): [True: 238, False: 1.46k] ------------------ 44| 238| if (count != 0) { ------------------ | Branch (44:11): [True: 36, False: 202] ------------------ 45| 36| count--; 46| 36| } // Sometimes the error is only detected in the next chunk 47| 238| result res = scalar::utf8::rewind_and_validate_with_errors( 48| 238| reinterpret_cast(input), 49| 238| reinterpret_cast(input + count), length - count); 50| 238| res.count += count; 51| 238| return res; 52| 238| } 53| 1.46k| reader.advance(); 54| 1.46k| count += 64; 55| 1.46k| } 56| 5.32k| uint8_t block[64]{}; 57| 5.32k| reader.get_remainder(block); 58| 5.32k| simd::simd8x64 in(block); 59| 5.32k| c.check_next_input(in); 60| 5.32k| reader.advance(); 61| 5.32k| c.check_eof(); 62| 5.32k| if (c.errors()) { ------------------ | Branch (62:7): [True: 182, False: 5.14k] ------------------ 63| 182| if (count != 0) { ------------------ | Branch (63:9): [True: 14, False: 168] ------------------ 64| 14| count--; 65| 14| } // Sometimes the error is only detected in the next chunk 66| 182| result res = scalar::utf8::rewind_and_validate_with_errors( 67| 182| reinterpret_cast(input), 68| 182| reinterpret_cast(input) + count, length - count); 69| 182| res.count += count; 70| 182| return res; 71| 5.14k| } else { 72| 5.14k| return result(error_code::SUCCESS, length); 73| 5.14k| } 74| 5.32k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE0EEEKNS_6resultEPKDsm: 49| 7.94k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.94k| if (simdutf_unlikely(size == 0)) { ------------------ | | 92| 7.94k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 5.20k, False: 2.74k] | | ------------------ ------------------ 51| 5.20k| return result(error_code::SUCCESS, 0); 52| 5.20k| } 53| | 54| 2.74k| const char16_t *start = input; 55| 2.74k| const char16_t *end = input + size; 56| | 57| 2.74k| const auto v_d8 = simd8::splat(0xd8); 58| 2.74k| const auto v_f8 = simd8::splat(0xf8); 59| 2.74k| const auto v_fc = simd8::splat(0xfc); 60| 2.74k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 14.2k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 11.6k, False: 2.65k] ------------------ 63| | // 0. Load data: since the validation takes into account only higher 64| | // byte of each word, we compress the two vectors into one which 65| | // consists only the higher bytes. 66| 11.6k| auto in0 = simd16(input); 67| 11.6k| auto in1 = 68| 11.6k| simd16(input + simd16::SIZE / sizeof(char16_t)); 69| | 70| | // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16 71| | // and yields a single vector having only higher bytes of characters. 72| 11.6k| const auto in = utf16_gather_high_bytes(in0, in1); 73| | 74| | // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy). 75| 11.6k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 11.6k| const uint16_t surrogates_bitmask = 77| 11.6k| static_cast(surrogates_wordmask.to_bitmask()); 78| 11.6k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 10.5k, False: 1.09k] ------------------ 79| 10.5k| input += 16; 80| 10.5k| } else { 81| | // 2. We have some surrogates that have to be distinguished: 82| | // - low surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF) 83| | // - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF) 84| | // 85| | // Fact: high surrogate has 11th bit set (3rd bit in the higher byte) 86| | 87| | // V - non-surrogate code units 88| | // V = not surrogates_wordmask 89| 1.09k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.09k| const auto vH = (in & v_fc) == v_dc; 93| 1.09k| const uint16_t H = static_cast(vH.to_bitmask()); 94| | 95| | // L - word mask for low surrogates 96| | // L = not H and surrogates_wordmask 97| 1.09k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.09k| const uint16_t a = static_cast( 100| 1.09k| L & (H >> 1)); // A low surrogate must be followed by high one. 101| | // (A low surrogate placed in the 7th register's word 102| | // is an exception we handle.) 103| 1.09k| const uint16_t b = static_cast( 104| 1.09k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.09k| const uint16_t c = static_cast( 107| 1.09k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.09k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 679, False: 419] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 679| input += 16; 113| 679| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 335, False: 84] ------------------ 114| | // The 15 lower code units of the input register contains valid UTF-16. 115| | // The 15th word may be either a low or high surrogate. It the next 116| | // iteration we 1) check if the low surrogate is followed by a high 117| | // one, 2) reject sole high surrogate. 118| 335| input += 15; 119| 335| } else { 120| 84| return result(error_code::SURROGATE, input - start); 121| 84| } 122| 1.09k| } 123| 11.6k| } 124| | 125| 2.65k| return result(error_code::SUCCESS, input - start); 126| 2.74k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE1EEEKNS_6resultEPKDsm: 49| 7.95k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.95k| if (simdutf_unlikely(size == 0)) { ------------------ | | 92| 7.95k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 5.19k, False: 2.75k] | | ------------------ ------------------ 51| 5.19k| return result(error_code::SUCCESS, 0); 52| 5.19k| } 53| | 54| 2.75k| const char16_t *start = input; 55| 2.75k| const char16_t *end = input + size; 56| | 57| 2.75k| const auto v_d8 = simd8::splat(0xd8); 58| 2.75k| const auto v_f8 = simd8::splat(0xf8); 59| 2.75k| const auto v_fc = simd8::splat(0xfc); 60| 2.75k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 14.8k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 12.1k, False: 2.67k] ------------------ 63| | // 0. Load data: since the validation takes into account only higher 64| | // byte of each word, we compress the two vectors into one which 65| | // consists only the higher bytes. 66| 12.1k| auto in0 = simd16(input); 67| 12.1k| auto in1 = 68| 12.1k| simd16(input + simd16::SIZE / sizeof(char16_t)); 69| | 70| | // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16 71| | // and yields a single vector having only higher bytes of characters. 72| 12.1k| const auto in = utf16_gather_high_bytes(in0, in1); 73| | 74| | // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy). 75| 12.1k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 12.1k| const uint16_t surrogates_bitmask = 77| 12.1k| static_cast(surrogates_wordmask.to_bitmask()); 78| 12.1k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 11.2k, False: 988] ------------------ 79| 11.2k| input += 16; 80| 11.2k| } else { 81| | // 2. We have some surrogates that have to be distinguished: 82| | // - low surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF) 83| | // - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF) 84| | // 85| | // Fact: high surrogate has 11th bit set (3rd bit in the higher byte) 86| | 87| | // V - non-surrogate code units 88| | // V = not surrogates_wordmask 89| 988| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 988| const auto vH = (in & v_fc) == v_dc; 93| 988| const uint16_t H = static_cast(vH.to_bitmask()); 94| | 95| | // L - word mask for low surrogates 96| | // L = not H and surrogates_wordmask 97| 988| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 988| const uint16_t a = static_cast( 100| 988| L & (H >> 1)); // A low surrogate must be followed by high one. 101| | // (A low surrogate placed in the 7th register's word 102| | // is an exception we handle.) 103| 988| const uint16_t b = static_cast( 104| 988| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 988| const uint16_t c = static_cast( 107| 988| V | a | b); // Combine all the masks into the final one. 108| | 109| 988| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 553, False: 435] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 553| input += 16; 113| 553| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 357, False: 78] ------------------ 114| | // The 15 lower code units of the input register contains valid UTF-16. 115| | // The 15th word may be either a low or high surrogate. It the next 116| | // iteration we 1) check if the low surrogate is followed by a high 117| | // one, 2) reject sole high surrogate. 118| 357| input += 15; 119| 357| } else { 120| 78| return result(error_code::SURROGATE, input - start); 121| 78| } 122| 988| } 123| 12.1k| } 124| | 125| 2.67k| return result(error_code::SUCCESS, input - start); 126| 2.75k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE0EEEKNS_6resultEPKDsm: 49| 7.43k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.43k| if (simdutf_unlikely(size == 0)) { ------------------ | | 92| 7.43k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 5.21k, False: 2.21k] | | ------------------ ------------------ 51| 5.21k| return result(error_code::SUCCESS, 0); 52| 5.21k| } 53| | 54| 2.21k| const char16_t *start = input; 55| 2.21k| const char16_t *end = input + size; 56| | 57| 2.21k| const auto v_d8 = simd8::splat(0xd8); 58| 2.21k| const auto v_f8 = simd8::splat(0xf8); 59| 2.21k| const auto v_fc = simd8::splat(0xfc); 60| 2.21k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 20.5k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 18.4k, False: 2.11k] ------------------ 63| | // 0. Load data: since the validation takes into account only higher 64| | // byte of each word, we compress the two vectors into one which 65| | // consists only the higher bytes. 66| 18.4k| auto in0 = simd16(input); 67| 18.4k| auto in1 = 68| 18.4k| simd16(input + simd16::SIZE / sizeof(char16_t)); 69| | 70| | // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16 71| | // and yields a single vector having only higher bytes of characters. 72| 18.4k| const auto in = utf16_gather_high_bytes(in0, in1); 73| | 74| | // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy). 75| 18.4k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 18.4k| const uint16_t surrogates_bitmask = 77| 18.4k| static_cast(surrogates_wordmask.to_bitmask()); 78| 18.4k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 17.1k, False: 1.29k] ------------------ 79| 17.1k| input += 16; 80| 17.1k| } else { 81| | // 2. We have some surrogates that have to be distinguished: 82| | // - low surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF) 83| | // - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF) 84| | // 85| | // Fact: high surrogate has 11th bit set (3rd bit in the higher byte) 86| | 87| | // V - non-surrogate code units 88| | // V = not surrogates_wordmask 89| 1.29k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.29k| const auto vH = (in & v_fc) == v_dc; 93| 1.29k| const uint16_t H = static_cast(vH.to_bitmask()); 94| | 95| | // L - word mask for low surrogates 96| | // L = not H and surrogates_wordmask 97| 1.29k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.29k| const uint16_t a = static_cast( 100| 1.29k| L & (H >> 1)); // A low surrogate must be followed by high one. 101| | // (A low surrogate placed in the 7th register's word 102| | // is an exception we handle.) 103| 1.29k| const uint16_t b = static_cast( 104| 1.29k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.29k| const uint16_t c = static_cast( 107| 1.29k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.29k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 845, False: 452] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 845| input += 16; 113| 845| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 346, False: 106] ------------------ 114| | // The 15 lower code units of the input register contains valid UTF-16. 115| | // The 15th word may be either a low or high surrogate. It the next 116| | // iteration we 1) check if the low surrogate is followed by a high 117| | // one, 2) reject sole high surrogate. 118| 346| input += 15; 119| 346| } else { 120| 106| return result(error_code::SURROGATE, input - start); 121| 106| } 122| 1.29k| } 123| 18.4k| } 124| | 125| 2.11k| return result(error_code::SUCCESS, input - start); 126| 2.21k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE1EEEKNS_6resultEPKDsm: 49| 7.41k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.41k| if (simdutf_unlikely(size == 0)) { ------------------ | | 92| 7.41k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 5.23k, False: 2.17k] | | ------------------ ------------------ 51| 5.23k| return result(error_code::SUCCESS, 0); 52| 5.23k| } 53| | 54| 2.17k| const char16_t *start = input; 55| 2.17k| const char16_t *end = input + size; 56| | 57| 2.17k| const auto v_d8 = simd8::splat(0xd8); 58| 2.17k| const auto v_f8 = simd8::splat(0xf8); 59| 2.17k| const auto v_fc = simd8::splat(0xfc); 60| 2.17k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 20.6k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 18.5k, False: 2.07k] ------------------ 63| | // 0. Load data: since the validation takes into account only higher 64| | // byte of each word, we compress the two vectors into one which 65| | // consists only the higher bytes. 66| 18.5k| auto in0 = simd16(input); 67| 18.5k| auto in1 = 68| 18.5k| simd16(input + simd16::SIZE / sizeof(char16_t)); 69| | 70| | // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16 71| | // and yields a single vector having only higher bytes of characters. 72| 18.5k| const auto in = utf16_gather_high_bytes(in0, in1); 73| | 74| | // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy). 75| 18.5k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 18.5k| const uint16_t surrogates_bitmask = 77| 18.5k| static_cast(surrogates_wordmask.to_bitmask()); 78| 18.5k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 17.3k, False: 1.21k] ------------------ 79| 17.3k| input += 16; 80| 17.3k| } else { 81| | // 2. We have some surrogates that have to be distinguished: 82| | // - low surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF) 83| | // - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF) 84| | // 85| | // Fact: high surrogate has 11th bit set (3rd bit in the higher byte) 86| | 87| | // V - non-surrogate code units 88| | // V = not surrogates_wordmask 89| 1.21k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.21k| const auto vH = (in & v_fc) == v_dc; 93| 1.21k| const uint16_t H = static_cast(vH.to_bitmask()); 94| | 95| | // L - word mask for low surrogates 96| | // L = not H and surrogates_wordmask 97| 1.21k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.21k| const uint16_t a = static_cast( 100| 1.21k| L & (H >> 1)); // A low surrogate must be followed by high one. 101| | // (A low surrogate placed in the 7th register's word 102| | // is an exception we handle.) 103| 1.21k| const uint16_t b = static_cast( 104| 1.21k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.21k| const uint16_t c = static_cast( 107| 1.21k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.21k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 761, False: 454] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 761| input += 16; 113| 761| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 356, False: 98] ------------------ 114| | // The 15 lower code units of the input register contains valid UTF-16. 115| | // The 15th word may be either a low or high surrogate. It the next 116| | // iteration we 1) check if the low surrogate is followed by a high 117| | // one, 2) reject sole high surrogate. 118| 356| input += 15; 119| 356| } else { 120| 98| return result(error_code::SURROGATE, input - start); 121| 98| } 122| 1.21k| } 123| 18.5k| } 124| | 125| 2.07k| return result(error_code::SUCCESS, input - start); 126| 2.17k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf328validateEPKDim: 6| 10.5k|simdutf_really_inline bool validate(const char32_t *input, size_t size) { 7| 10.5k| if (simdutf_unlikely(size == 0)) { ------------------ | | 92| 10.5k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 8.49k, False: 2.03k] | | ------------------ ------------------ 8| | // empty input is valid UTF-32. protect the implementation from 9| | // handling nullptr 10| 8.49k| return true; 11| 8.49k| } 12| | 13| 2.03k| const char32_t *end = input + size; 14| | 15| 2.03k| using vector_u32 = simd32; 16| | 17| 2.03k| const auto standardmax = vector_u32::splat(0x10ffff); 18| 2.03k| const auto offset = vector_u32::splat(0xffff2000); 19| 2.03k| const auto standardoffsetmax = vector_u32::splat(0xfffff7ff); 20| 2.03k| auto currentmax = vector_u32::zero(); 21| 2.03k| auto currentoffsetmax = vector_u32::zero(); 22| | 23| 2.03k| constexpr size_t N = vector_u32::ELEMENTS; 24| | 25| 27.8k| while (input + N < end) { ------------------ | Branch (25:10): [True: 25.8k, False: 2.03k] ------------------ 26| 25.8k| auto in = vector_u32(input); 27| 25.8k| if constexpr (!match_system(endianness::BIG)) { 28| 25.8k| in.swap_bytes(); 29| 25.8k| } 30| | 31| 25.8k| currentmax = max(currentmax, in); 32| 25.8k| currentoffsetmax = max(currentoffsetmax, in + offset); 33| 25.8k| input += N; 34| 25.8k| } 35| | 36| 2.03k| const auto too_large = currentmax > standardmax; 37| 2.03k| if (too_large.any()) { ------------------ | Branch (37:7): [True: 0, False: 2.03k] ------------------ 38| 0| return false; 39| 0| } 40| | 41| 2.03k| const auto surrogate = currentoffsetmax > standardoffsetmax; 42| 2.03k| if (surrogate.any()) { ------------------ | Branch (42:7): [True: 0, False: 2.03k] ------------------ 43| 0| return false; 44| 0| } 45| | 46| 2.03k| return scalar::utf32::validate(input, end - input); 47| 2.03k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf328validateEPKDim: 6| 10.7k|simdutf_really_inline bool validate(const char32_t *input, size_t size) { 7| 10.7k| if (simdutf_unlikely(size == 0)) { ------------------ | | 92| 10.7k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 9.17k, False: 1.52k] | | ------------------ ------------------ 8| | // empty input is valid UTF-32. protect the implementation from 9| | // handling nullptr 10| 9.17k| return true; 11| 9.17k| } 12| | 13| 1.52k| const char32_t *end = input + size; 14| | 15| 1.52k| using vector_u32 = simd32; 16| | 17| 1.52k| const auto standardmax = vector_u32::splat(0x10ffff); 18| 1.52k| const auto offset = vector_u32::splat(0xffff2000); 19| 1.52k| const auto standardoffsetmax = vector_u32::splat(0xfffff7ff); 20| 1.52k| auto currentmax = vector_u32::zero(); 21| 1.52k| auto currentoffsetmax = vector_u32::zero(); 22| | 23| 1.52k| constexpr size_t N = vector_u32::ELEMENTS; 24| | 25| 64.0k| while (input + N < end) { ------------------ | Branch (25:10): [True: 62.5k, False: 1.52k] ------------------ 26| 62.5k| auto in = vector_u32(input); 27| 62.5k| if constexpr (!match_system(endianness::BIG)) { 28| 62.5k| in.swap_bytes(); 29| 62.5k| } 30| | 31| 62.5k| currentmax = max(currentmax, in); 32| 62.5k| currentoffsetmax = max(currentoffsetmax, in + offset); 33| 62.5k| input += N; 34| 62.5k| } 35| | 36| 1.52k| const auto too_large = currentmax > standardmax; 37| 1.52k| if (too_large.any()) { ------------------ | Branch (37:7): [True: 0, False: 1.52k] ------------------ 38| 0| return false; 39| 0| } 40| | 41| 1.52k| const auto surrogate = currentoffsetmax > standardoffsetmax; 42| 1.52k| if (surrogate.any()) { ------------------ | Branch (42:7): [True: 0, False: 1.52k] ------------------ 43| 0| return false; 44| 0| } 45| | 46| 1.52k| return scalar::utf32::validate(input, end - input); 47| 1.52k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block64C2EPKc: 507| 17.2k| simdutf_really_inline block64(const char *src) { 508| 17.2k| chunks[0] = _mm256_loadu_si256(reinterpret_cast(src)); 509| 17.2k| chunks[1] = _mm256_loadu_si256(reinterpret_cast(src + 32)); 510| 17.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6414compress_blockEmPc: 687| 2.32k| simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) { 688| 2.32k| if (is_power_of_two(mask)) { ------------------ | Branch (688:9): [True: 675, False: 1.65k] ------------------ 689| 675| return compress_block_single(mask, output); 690| 675| } 691| | 692| 1.65k| uint64_t nmask = ~mask; 693| 1.65k| compress(chunks[0], uint32_t(mask), output); 694| 1.65k| compress(chunks[1], uint32_t(mask >> 32), 695| 1.65k| output + count_ones(nmask & 0xFFFFFFFF)); 696| 1.65k| return count_ones(nmask); 697| 2.32k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6421compress_block_singleEmPc: 700| 675| char *output) { 701| 675| const size_t pos64 = trailing_zeroes(mask); 702| 675| const int8_t pos = pos64 & 0xf; 703| 675| switch (pos64 >> 4) { ------------------ | Branch (703:13): [True: 675, False: 0] ------------------ 704| 196| case 0b00: { ------------------ | Branch (704:5): [True: 196, False: 479] ------------------ 705| 196| const __m128i lane0 = _mm256_extracti128_si256(chunks[0], 0); 706| 196| const __m128i lane1 = _mm256_extracti128_si256(chunks[0], 1); 707| | 708| 196| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 709| 196| const __m128i v1 = 710| 196| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 711| 196| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 712| 196| const __m128i sh = _mm_sub_epi8(v1, v2); 713| 196| const __m128i compressed = _mm_shuffle_epi8(lane0, sh); 714| | 715| 196| _mm_storeu_si128((__m128i *)(output + 0 * 16), compressed); 716| 196| _mm_storeu_si128((__m128i *)(output + 1 * 16 - 1), lane1); 717| 196| _mm256_storeu_si256((__m256i *)(output + 2 * 16 - 1), chunks[1]); 718| 196| } break; 719| 176| case 0b01: { ------------------ | Branch (719:5): [True: 176, False: 499] ------------------ 720| 176| const __m128i lane0 = _mm256_extracti128_si256(chunks[0], 0); 721| 176| const __m128i lane1 = _mm256_extracti128_si256(chunks[0], 1); 722| 176| _mm_storeu_si128((__m128i *)(output + 0 * 16), lane0); 723| | 724| 176| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 725| 176| const __m128i v1 = 726| 176| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 727| 176| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 728| 176| const __m128i sh = _mm_sub_epi8(v1, v2); 729| 176| const __m128i compressed = _mm_shuffle_epi8(lane1, sh); 730| | 731| 176| _mm_storeu_si128((__m128i *)(output + 1 * 16), compressed); 732| 176| _mm256_storeu_si256((__m256i *)(output + 2 * 16 - 1), chunks[1]); 733| 176| } break; 734| 150| case 0b10: { ------------------ | Branch (734:5): [True: 150, False: 525] ------------------ 735| 150| const __m128i lane2 = _mm256_extracti128_si256(chunks[1], 0); 736| 150| const __m128i lane3 = _mm256_extracti128_si256(chunks[1], 1); 737| | 738| 150| _mm256_storeu_si256((__m256i *)(output + 0 * 16), chunks[0]); 739| | 740| 150| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 741| 150| const __m128i v1 = 742| 150| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 743| 150| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 744| 150| const __m128i sh = _mm_sub_epi8(v1, v2); 745| 150| const __m128i compressed = _mm_shuffle_epi8(lane2, sh); 746| | 747| 150| _mm_storeu_si128((__m128i *)(output + 2 * 16), compressed); 748| 150| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), lane3); 749| 150| } break; 750| 153| case 0b11: { ------------------ | Branch (750:5): [True: 153, False: 522] ------------------ 751| 153| const __m128i lane2 = _mm256_extracti128_si256(chunks[1], 0); 752| 153| const __m128i lane3 = _mm256_extracti128_si256(chunks[1], 1); 753| | 754| 153| _mm256_storeu_si256((__m256i *)(output + 0 * 16), chunks[0]); 755| 153| _mm_storeu_si128((__m128i *)(output + 2 * 16), lane2); 756| | 757| 153| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 758| 153| const __m128i v1 = 759| 153| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 760| 153| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 761| 153| const __m128i sh = _mm_sub_epi8(v1, v2); 762| 153| const __m128i compressed = _mm_shuffle_epi8(lane3, sh); 763| | 764| 153| _mm_storeu_si128((__m128i *)(output + 3 * 16), compressed); 765| 153| } break; 766| 675| } 767| | 768| 675| return 63; 769| 675| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_18compressIvEEvDv4_xjPc: 455| 3.30k|simdutf_really_inline void compress(__m256i data, uint32_t mask, char *output) { 456| 3.30k| if (mask == 0) { ------------------ | Branch (456:7): [True: 639, False: 2.66k] ------------------ 457| 639| _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), data); 458| 639| return; 459| 639| } 460| 2.66k| compress(_mm256_castsi256_si128(data), uint16_t(mask), output); 461| | compress(_mm256_extracti128_si256(data, 1), uint16_t(mask >> 16), 462| 2.66k| output + count_ones(~mask & 0xFFFF)); 463| 2.66k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_18compressEDv2_xtPc: 418| 5.33k|static inline void compress(__m128i data, uint16_t mask, char *output) { 419| 5.33k| if (mask == 0) { ------------------ | Branch (419:7): [True: 1.36k, False: 3.96k] ------------------ 420| 1.36k| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data); 421| 1.36k| return; 422| 1.36k| } 423| | // this particular implementation was inspired by work done by @animetosho 424| | // we do it in two steps, first 8 bytes and then second 8 bytes 425| 3.96k| uint8_t mask1 = uint8_t(mask); // least significant 8 bits 426| 3.96k| uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits 427| | // next line just loads the 64-bit values thintable_epi8[mask1] and 428| | // thintable_epi8[mask2] into a 128-bit register, using only 429| | // two instructions on most compilers. 430| | 431| 3.96k| __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2], 432| 3.96k| tables::base64::thintable_epi8[mask1]); 433| | // we increment by 0x08 the second half of the mask 434| 3.96k| shufmask = 435| 3.96k| _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0)); 436| | // this is the version "nearly pruned" 437| 3.96k| __m128i pruned = _mm_shuffle_epi8(data, shufmask); 438| | // we still need to put the two halves together. 439| | // we compute the popcount of the first half: 440| 3.96k| int pop1 = tables::base64::BitsSetTable256mul2[mask1]; 441| | // then load the corresponding mask, what it does is to write 442| | // only the first pop1 bytes from the first 8 bytes, and then 443| | // it fills in with the bytes from the second 8 bytes + some filling 444| | // at the end. 445| 3.96k| __m128i compactmask = _mm_loadu_si128(reinterpret_cast( 446| 3.96k| tables::base64::pshufb_combine_table + pop1 * 8)); 447| 3.96k| __m128i answer = _mm_shuffle_epi8(pruned, compactmask); 448| | 449| 3.96k| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer); 450| 3.96k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6410copy_blockEPc: 532| 326| simdutf_really_inline void copy_block(char *output) { 533| 326| _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), chunks[0]); 534| 326| _mm256_storeu_si256(reinterpret_cast<__m256i *>(output + 32), chunks[1]); 535| 326| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6424base64_decode_block_safeEPc: 543| 836| simdutf_really_inline void base64_decode_block_safe(char *out) { 544| 836| base64_decode(out, chunks[0]); 545| 836| alignas(32) char buffer[32]; // We enforce safety with a buffer. 546| 836| base64_decode(buffer, chunks[1]); 547| 836| std::memcpy(out + 24, buffer, 24); 548| 836| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113base64_decodeIvEEvPcDv4_x: 466| 30.7k|simdutf_really_inline void base64_decode(char *out, __m256i str) { 467| | // credit: aqrit 468| 30.7k| const __m256i pack_shuffle = 469| 30.7k| _mm256_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1, 470| 30.7k| 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1); 471| 30.7k| const __m256i t0 = _mm256_maddubs_epi16(str, _mm256_set1_epi32(0x01400140)); 472| 30.7k| const __m256i t1 = _mm256_madd_epi16(t0, _mm256_set1_epi32(0x00011000)); 473| 30.7k| const __m256i t2 = _mm256_shuffle_epi8(t1, pack_shuffle); 474| | 475| | // Store the output: 476| 30.7k| _mm_storeu_si128((__m128i *)out, _mm256_castsi256_si128(t2)); 477| | _mm_storeu_si128((__m128i *)(out + 12), _mm256_extracti128_si256(t2, 1)); 478| 30.7k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6419base64_decode_blockEPc: 538| 12.8k| simdutf_really_inline void base64_decode_block(char *out) { 539| 12.8k| base64_decode(out, chunks[0]); 540| 12.8k| base64_decode(out + 24, chunks[1]); 541| 12.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_119base64_decode_blockIvEEvPcPKc: 481| 1.64k|simdutf_really_inline void base64_decode_block(char *out, const char *src) { 482| 1.64k| base64_decode(out, 483| 1.64k| _mm256_loadu_si256(reinterpret_cast(src))); 484| 1.64k| base64_decode(out + 24, _mm256_loadu_si256( 485| 1.64k| reinterpret_cast(src + 32))); 486| 1.64k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_124base64_decode_block_safeIvEEvPcPKc: 490| 50| const char *src) { 491| 50| base64_decode(out, 492| 50| _mm256_loadu_si256(reinterpret_cast(src))); 493| 50| alignas(32) char buffer[32]; // We enforce safety with a buffer. 494| 50| base64_decode( 495| 50| buffer, _mm256_loadu_si256(reinterpret_cast(src + 32))); 496| 50| std::memcpy(out + 24, buffer, 24); 497| 50|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEmPm: 551| 17.2k| simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) { 552| 17.2k| uint32_t err0 = 0; 553| 17.2k| uint32_t err1 = 0; 554| 17.2k| uint64_t m0 = to_base64_mask( 555| 17.2k| &chunks[0], &err0); 556| 17.2k| uint64_t m1 = to_base64_mask( 557| 17.2k| &chunks[1], &err1); 558| 17.2k| if (!ignore_garbage) { ------------------ | Branch (558:9): [True: 17.2k, Folded] ------------------ 559| 17.2k| *error = err0 | ((uint64_t)err1 << 32); 560| 17.2k| } 561| 17.2k| return m0 | (m1 << 32); 562| 17.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEjPDv4_xPj: 565| 34.4k| simdutf_really_inline uint32_t to_base64_mask(__m256i *src, uint32_t *error) { 566| 34.4k| const __m256i ascii_space_tbl = 567| 34.4k| _mm256_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa, 568| 34.4k| 0x0, 0xc, 0xd, 0x0, 0x0, 0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 569| 34.4k| 0x0, 0x0, 0x0, 0x9, 0xa, 0x0, 0xc, 0xd, 0x0, 0x0); 570| | // credit: aqrit 571| 34.4k| __m256i delta_asso; 572| 34.4k| if (default_or_url) { ------------------ | Branch (572:9): [Folded, False: 34.4k] ------------------ 573| 0| delta_asso = _mm256_setr_epi8( 574| 0| 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 575| 0| 0x00, 0x00, 0x11, 0x00, 0x16, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 576| 0| 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x16); 577| 34.4k| } else if (base64_url) { ------------------ | Branch (577:16): [Folded, False: 34.4k] ------------------ 578| 0| delta_asso = _mm256_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 579| 0| 0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF, 0x1, 0x1, 580| 0| 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 0x0, 0x0, 581| 0| 0x0, 0x0, 0xF, 0x0, 0xF); 582| 34.4k| } else { 583| 34.4k| delta_asso = _mm256_setr_epi8( 584| 34.4k| 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 585| 34.4k| 0x00, 0x00, 0x0F, 0x00, 0x0F, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 586| 34.4k| 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F); 587| 34.4k| } 588| | 589| 34.4k| __m256i delta_values; 590| 34.4k| if (default_or_url) { ------------------ | Branch (590:9): [Folded, False: 34.4k] ------------------ 591| 0| delta_values = _mm256_setr_epi8( 592| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0x13), 593| 0| uint8_t(0x04), uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 594| 0| uint8_t(0xB9), uint8_t(0x00), uint8_t(0xFF), uint8_t(0x11), 595| 0| uint8_t(0xFF), uint8_t(0xBF), uint8_t(0x10), uint8_t(0xB9), 596| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0x13), 597| 0| uint8_t(0x04), uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 598| 0| uint8_t(0xB9), uint8_t(0x00), uint8_t(0xFF), uint8_t(0x11), 599| 0| uint8_t(0xFF), uint8_t(0xBF), uint8_t(0x10), uint8_t(0xB9)); 600| 34.4k| } else if (base64_url) { ------------------ | Branch (600:16): [Folded, False: 34.4k] ------------------ 601| 0| delta_values = _mm256_setr_epi8( 602| 0| 0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 603| 0| uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF), uint8_t(0xE0), 604| 0| uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), 605| 0| uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3), 606| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9)); 607| 34.4k| } else { 608| 34.4k| delta_values = _mm256_setr_epi8( 609| 34.4k| int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), int8_t(0x04), 610| 34.4k| int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9), int8_t(0x00), 611| 34.4k| int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), 612| 34.4k| int8_t(0xB9), int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), 613| 34.4k| int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9), 614| 34.4k| int8_t(0x00), int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF), 615| 34.4k| int8_t(0xB9), int8_t(0xB9)); 616| 34.4k| } 617| | 618| 34.4k| __m256i check_asso; 619| 34.4k| if (default_or_url) { ------------------ | Branch (619:9): [Folded, False: 34.4k] ------------------ 620| 0| check_asso = _mm256_setr_epi8( 621| 0| 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x03, 622| 0| 0x07, 0x0B, 0x0E, 0x0B, 0x06, 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 623| 0| 0x01, 0x01, 0x01, 0x01, 0x03, 0x07, 0x0B, 0x0E, 0x0B, 0x06); 624| 34.4k| } else if (base64_url) { ------------------ | Branch (624:16): [Folded, False: 34.4k] ------------------ 625| 0| check_asso = _mm256_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 626| 0| 0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6, 0xD, 0x1, 627| 0| 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x3, 628| 0| 0x7, 0xB, 0xE, 0xB, 0x6); 629| 34.4k| } else { 630| 34.4k| check_asso = _mm256_setr_epi8( 631| 34.4k| 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x03, 632| 34.4k| 0x07, 0x0B, 0x0B, 0x0B, 0x0F, 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 633| 34.4k| 0x01, 0x01, 0x01, 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F); 634| 34.4k| } 635| 34.4k| __m256i check_values; 636| 34.4k| if (default_or_url) { ------------------ | Branch (636:9): [Folded, False: 34.4k] ------------------ 637| 0| check_values = _mm256_setr_epi8( 638| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 639| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xD5), uint8_t(0xA6), 640| 0| uint8_t(0xB5), uint8_t(0xA1), uint8_t(0x00), uint8_t(0x80), 641| 0| uint8_t(0x00), uint8_t(0x80), uint8_t(0x00), uint8_t(0x80), 642| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 643| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xD5), uint8_t(0xA6), 644| 0| uint8_t(0xB5), uint8_t(0xA1), uint8_t(0x00), uint8_t(0x80), 645| 0| uint8_t(0x00), uint8_t(0x80), uint8_t(0x00), uint8_t(0x80)); 646| 34.4k| } else if (base64_url) { ------------------ | Branch (646:16): [Folded, False: 34.4k] ------------------ 647| 0| check_values = _mm256_setr_epi8( 648| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 649| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6), uint8_t(0xA6), 650| 0| uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, uint8_t(0x80), 651| 0| 0x0, uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 652| 0| uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6), 653| 0| uint8_t(0xA6), uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, 654| 0| uint8_t(0x80), 0x0, uint8_t(0x80)); 655| 34.4k| } else { 656| 34.4k| check_values = _mm256_setr_epi8( 657| 34.4k| int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0xCF), 658| 34.4k| int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5), int8_t(0x86), 659| 34.4k| int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80), int8_t(0x91), 660| 34.4k| int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), 661| 34.4k| int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5), 662| 34.4k| int8_t(0x86), int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80), 663| 34.4k| int8_t(0x91), int8_t(0x80)); 664| 34.4k| } 665| 34.4k| const __m256i shifted = _mm256_srli_epi32(*src, 3); 666| 34.4k| __m256i delta_hash = 667| 34.4k| _mm256_avg_epu8(_mm256_shuffle_epi8(delta_asso, *src), shifted); 668| 34.4k| if (default_or_url) { ------------------ | Branch (668:9): [Folded, False: 34.4k] ------------------ 669| 0| delta_hash = _mm256_and_si256(delta_hash, _mm256_set1_epi8(0xf)); 670| 0| } 671| 34.4k| const __m256i check_hash = 672| 34.4k| _mm256_avg_epu8(_mm256_shuffle_epi8(check_asso, *src), shifted); 673| 34.4k| const __m256i out = 674| 34.4k| _mm256_adds_epi8(_mm256_shuffle_epi8(delta_values, delta_hash), *src); 675| 34.4k| const __m256i chk = 676| 34.4k| _mm256_adds_epi8(_mm256_shuffle_epi8(check_values, check_hash), *src); 677| 34.4k| const int mask = _mm256_movemask_epi8(chk); 678| 34.4k| if (!ignore_garbage && mask) { ------------------ | Branch (678:9): [True: 34.4k, Folded] | Branch (678:28): [True: 5.05k, False: 29.4k] ------------------ 679| 5.05k| __m256i ascii_space = 680| 5.05k| _mm256_cmpeq_epi8(_mm256_shuffle_epi8(ascii_space_tbl, *src), *src); 681| 5.05k| *error = (mask ^ _mm256_movemask_epi8(ascii_space)); 682| 5.05k| } 683| 34.4k| *src = out; 684| 34.4k| return (uint32_t)mask; 685| 34.4k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113encode_base64ILb0EEEmPcPKcmNS_14base64_optionsE: 414| 15.1k| base64_options options) { 415| 15.1k| return avx2_encode_base64_impl(dst, src, srclen, options); 416| 15.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_123avx2_encode_base64_implILb0ELb0EEEmPcPKcmNS_14base64_optionsEm: 156| 15.1k| size_t line_length = simdutf::default_line_length) { 157| 15.1k| size_t offset = 0; 158| | 159| 15.1k| if (line_length < 4) { ------------------ | Branch (159:7): [True: 0, False: 15.1k] ------------------ 160| 0| line_length = 4; // We do not support line_length less than 4 161| 0| } 162| | // credit: Wojciech Muła 163| 15.1k| const uint8_t *input = (const uint8_t *)src; 164| | 165| 15.1k| uint8_t *out = (uint8_t *)dst; 166| 15.1k| const __m256i shuf = 167| 15.1k| _mm256_set_epi8(10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1, 168| | 169| 15.1k| 10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1); 170| 15.1k| size_t i = 0; 171| 17.2k| for (; i + 100 <= srclen; i += 96) { ------------------ | Branch (171:10): [True: 2.08k, False: 15.1k] ------------------ 172| 2.08k| const __m128i lo0 = _mm_loadu_si128( 173| 2.08k| reinterpret_cast(input + i + 4 * 3 * 0)); 174| 2.08k| const __m128i hi0 = _mm_loadu_si128( 175| 2.08k| reinterpret_cast(input + i + 4 * 3 * 1)); 176| 2.08k| const __m128i lo1 = _mm_loadu_si128( 177| 2.08k| reinterpret_cast(input + i + 4 * 3 * 2)); 178| 2.08k| const __m128i hi1 = _mm_loadu_si128( 179| 2.08k| reinterpret_cast(input + i + 4 * 3 * 3)); 180| 2.08k| const __m128i lo2 = _mm_loadu_si128( 181| 2.08k| reinterpret_cast(input + i + 4 * 3 * 4)); 182| 2.08k| const __m128i hi2 = _mm_loadu_si128( 183| 2.08k| reinterpret_cast(input + i + 4 * 3 * 5)); 184| 2.08k| const __m128i lo3 = _mm_loadu_si128( 185| 2.08k| reinterpret_cast(input + i + 4 * 3 * 6)); 186| 2.08k| const __m128i hi3 = _mm_loadu_si128( 187| 2.08k| reinterpret_cast(input + i + 4 * 3 * 7)); 188| | 189| 2.08k| __m256i in0 = _mm256_shuffle_epi8(_mm256_set_m128i(hi0, lo0), shuf); ------------------ | | 4| 2.08k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.08k| _mm256_castsi128_si256(xmm2), 2) ------------------ 190| 2.08k| __m256i in1 = _mm256_shuffle_epi8(_mm256_set_m128i(hi1, lo1), shuf); ------------------ | | 4| 2.08k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.08k| _mm256_castsi128_si256(xmm2), 2) ------------------ 191| 2.08k| __m256i in2 = _mm256_shuffle_epi8(_mm256_set_m128i(hi2, lo2), shuf); ------------------ | | 4| 2.08k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.08k| _mm256_castsi128_si256(xmm2), 2) ------------------ 192| 2.08k| __m256i in3 = _mm256_shuffle_epi8(_mm256_set_m128i(hi3, lo3), shuf); ------------------ | | 4| 2.08k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.08k| _mm256_castsi128_si256(xmm2), 2) ------------------ 193| | 194| 2.08k| const __m256i t0_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x0fc0fc00)); 195| 2.08k| const __m256i t0_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x0fc0fc00)); 196| 2.08k| const __m256i t0_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x0fc0fc00)); 197| 2.08k| const __m256i t0_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x0fc0fc00)); 198| | 199| 2.08k| const __m256i t1_0 = 200| 2.08k| _mm256_mulhi_epu16(t0_0, _mm256_set1_epi32(0x04000040)); 201| 2.08k| const __m256i t1_1 = 202| 2.08k| _mm256_mulhi_epu16(t0_1, _mm256_set1_epi32(0x04000040)); 203| 2.08k| const __m256i t1_2 = 204| 2.08k| _mm256_mulhi_epu16(t0_2, _mm256_set1_epi32(0x04000040)); 205| 2.08k| const __m256i t1_3 = 206| 2.08k| _mm256_mulhi_epu16(t0_3, _mm256_set1_epi32(0x04000040)); 207| | 208| 2.08k| const __m256i t2_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x003f03f0)); 209| 2.08k| const __m256i t2_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x003f03f0)); 210| 2.08k| const __m256i t2_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x003f03f0)); 211| 2.08k| const __m256i t2_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x003f03f0)); 212| | 213| 2.08k| const __m256i t3_0 = 214| 2.08k| _mm256_mullo_epi16(t2_0, _mm256_set1_epi32(0x01000010)); 215| 2.08k| const __m256i t3_1 = 216| 2.08k| _mm256_mullo_epi16(t2_1, _mm256_set1_epi32(0x01000010)); 217| 2.08k| const __m256i t3_2 = 218| 2.08k| _mm256_mullo_epi16(t2_2, _mm256_set1_epi32(0x01000010)); 219| 2.08k| const __m256i t3_3 = 220| 2.08k| _mm256_mullo_epi16(t2_3, _mm256_set1_epi32(0x01000010)); 221| | 222| 2.08k| const __m256i input0 = _mm256_or_si256(t1_0, t3_0); 223| 2.08k| const __m256i input1 = _mm256_or_si256(t1_1, t3_1); 224| 2.08k| const __m256i input2 = _mm256_or_si256(t1_2, t3_2); 225| 2.08k| const __m256i input3 = _mm256_or_si256(t1_3, t3_3); 226| | 227| 2.08k| if (use_lines) { ------------------ | Branch (227:9): [Folded, False: 2.08k] ------------------ 228| 0| if (line_length >= 32) { // fast path ------------------ | Branch (228:11): [True: 0, False: 0] ------------------ 229| 0| __m256i result; 230| 0| result = lookup_pshufb_improved(input0); 231| 0| if (offset + 32 > line_length) { ------------------ | Branch (231:13): [True: 0, False: 0] ------------------ 232| 0| size_t location_end = line_length - offset; 233| 0| size_t to_move = 32 - location_end; 234| | // We could do this, or extract instead. 235| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 236| 0| _mm256_storeu_si256( 237| 0| reinterpret_cast<__m256i *>(out), 238| 0| insert_line_feed32(result, static_cast(location_end))); 239| 0| offset = to_move; 240| 0| out += 32 + 1; 241| 0| } else { 242| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 243| 0| offset += 32; 244| 0| out += 32; 245| 0| } 246| 0| result = lookup_pshufb_improved(input1); 247| | 248| 0| if (offset + 32 > line_length) { ------------------ | Branch (248:13): [True: 0, False: 0] ------------------ 249| 0| size_t location_end = line_length - offset; 250| 0| size_t to_move = 32 - location_end; 251| | 252| | // We could do this, or extract instead. 253| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 254| 0| _mm256_storeu_si256( 255| 0| reinterpret_cast<__m256i *>(out), 256| 0| insert_line_feed32(result, static_cast(location_end))); 257| | // see above. 258| | // out[32] = static_cast(_mm256_extract_epi8(result, 31)); 259| 0| offset = to_move; 260| 0| out += 32 + 1; 261| 0| } else { 262| | 263| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 264| | 265| 0| offset += 32; 266| 0| out += 32; 267| 0| } 268| 0| result = lookup_pshufb_improved(input2); 269| | 270| 0| if (offset + 32 > line_length) { ------------------ | Branch (270:13): [True: 0, False: 0] ------------------ 271| 0| size_t location_end = line_length - offset; 272| 0| size_t to_move = 32 - location_end; 273| | 274| | // We could do this, or extract instead. 275| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 276| 0| _mm256_storeu_si256( 277| 0| reinterpret_cast<__m256i *>(out), 278| 0| insert_line_feed32(result, static_cast(location_end))); 279| | // see above. 280| | // out[32] = static_cast(_mm256_extract_epi8(result, 31)); 281| 0| offset = to_move; 282| 0| out += 32 + 1; 283| 0| } else { 284| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 285| 0| offset += 32; 286| 0| out += 32; 287| 0| } 288| 0| result = lookup_pshufb_improved(input3); 289| | 290| 0| if (offset + 32 > line_length) { ------------------ | Branch (290:13): [True: 0, False: 0] ------------------ 291| 0| size_t location_end = line_length - offset; 292| 0| size_t to_move = 32 - location_end; 293| | 294| | // We could do this, or extract instead. 295| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 296| 0| _mm256_storeu_si256( 297| 0| reinterpret_cast<__m256i *>(out), 298| 0| insert_line_feed32(result, static_cast(location_end))); 299| | // see above. 300| | // out[32] = static_cast(_mm256_extract_epi8(result, 31)); 301| 0| offset = to_move; 302| 0| out += 32 + 1; 303| 0| } else { 304| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 305| 0| offset += 32; 306| 0| out += 32; 307| 0| } 308| 0| } else { // slow path 309| | // could be optimized 310| 0| uint8_t buffer[128]; 311| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer), 312| 0| lookup_pshufb_improved(input0)); 313| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer + 32), 314| 0| lookup_pshufb_improved(input1)); 315| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer + 64), 316| 0| lookup_pshufb_improved(input2)); 317| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer + 96), 318| 0| lookup_pshufb_improved(input3)); 319| 0| size_t out_pos = 0; 320| 0| size_t local_offset = offset; 321| 0| for (size_t j = 0; j < 128;) { ------------------ | Branch (321:28): [True: 0, False: 0] ------------------ 322| 0| if (local_offset == line_length) { ------------------ | Branch (322:15): [True: 0, False: 0] ------------------ 323| 0| out[out_pos++] = '\n'; 324| 0| local_offset = 0; 325| 0| } 326| 0| out[out_pos++] = buffer[j++]; 327| 0| local_offset++; 328| 0| } 329| 0| offset = local_offset; 330| 0| out += out_pos; 331| 0| } 332| 2.08k| } else { 333| 2.08k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), 334| 2.08k| lookup_pshufb_improved(input0)); 335| 2.08k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 32), 336| 2.08k| lookup_pshufb_improved(input1)); 337| 2.08k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 64), 338| 2.08k| lookup_pshufb_improved(input2)); 339| 2.08k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 96), 340| 2.08k| lookup_pshufb_improved(input3)); 341| | 342| 2.08k| out += 128; 343| 2.08k| } 344| 2.08k| } 345| 16.8k| for (; i + 28 <= srclen; i += 24) { ------------------ | Branch (345:10): [True: 1.66k, False: 15.1k] ------------------ 346| | // lo = [xxxx|DDDC|CCBB|BAAA] 347| | // hi = [xxxx|HHHG|GGFF|FEEE] 348| 1.66k| const __m128i lo = 349| 1.66k| _mm_loadu_si128(reinterpret_cast(input + i)); 350| 1.66k| const __m128i hi = 351| 1.66k| _mm_loadu_si128(reinterpret_cast(input + i + 4 * 3)); 352| | 353| | // bytes from groups A, B and C are needed in separate 32-bit lanes 354| | // in = [0HHH|0GGG|0FFF|0EEE[0DDD|0CCC|0BBB|0AAA] 355| 1.66k| __m256i in = _mm256_shuffle_epi8(_mm256_set_m128i(hi, lo), shuf); ------------------ | | 4| 1.66k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 1.66k| _mm256_castsi128_si256(xmm2), 2) ------------------ 356| | 357| | // this part is well commented in encode.sse.cpp 358| | 359| 1.66k| const __m256i t0 = _mm256_and_si256(in, _mm256_set1_epi32(0x0fc0fc00)); 360| 1.66k| const __m256i t1 = _mm256_mulhi_epu16(t0, _mm256_set1_epi32(0x04000040)); 361| 1.66k| const __m256i t2 = _mm256_and_si256(in, _mm256_set1_epi32(0x003f03f0)); 362| 1.66k| const __m256i t3 = _mm256_mullo_epi16(t2, _mm256_set1_epi32(0x01000010)); 363| 1.66k| const __m256i indices = _mm256_or_si256(t1, t3); 364| | 365| 1.66k| if (use_lines) { ------------------ | Branch (365:9): [Folded, False: 1.66k] ------------------ 366| 0| if (line_length >= 32) { // fast path ------------------ | Branch (366:11): [True: 0, False: 0] ------------------ 367| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), 368| 0| lookup_pshufb_improved(indices)); 369| | 370| 0| if (offset + 32 > line_length) { ------------------ | Branch (370:13): [True: 0, False: 0] ------------------ 371| 0| size_t location_end = line_length - offset; 372| 0| size_t to_move = 32 - location_end; 373| 0| std::memmove(out + location_end + 1, out + location_end, to_move); 374| 0| out[location_end] = '\n'; 375| 0| offset = to_move; 376| 0| out += 32 + 1; 377| 0| } else { 378| 0| offset += 32; 379| 0| out += 32; 380| 0| } 381| 0| } else { // slow path 382| | // could be optimized 383| 0| alignas(32) uint8_t buffer[32]; 384| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer), 385| 0| lookup_pshufb_improved(indices)); 386| 0| std::memcpy(out, buffer, 32); 387| 0| size_t out_pos = 0; 388| 0| size_t local_offset = offset; 389| 0| for (size_t j = 0; j < 32;) { ------------------ | Branch (389:28): [True: 0, False: 0] ------------------ 390| 0| if (local_offset == line_length) { ------------------ | Branch (390:15): [True: 0, False: 0] ------------------ 391| 0| out[out_pos++] = '\n'; 392| 0| local_offset = 0; 393| 0| } 394| 0| out[out_pos++] = buffer[j++]; 395| 0| local_offset++; 396| 0| } 397| 0| offset = local_offset; 398| 0| out += out_pos; 399| 0| } 400| 1.66k| } else { 401| 1.66k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), 402| 1.66k| lookup_pshufb_improved(indices)); 403| | 404| 1.66k| out += 32; 405| 1.66k| } 406| 1.66k| } 407| 15.1k| return ((char *)out - (char *)dst) + 408| 15.1k| scalar::base64::tail_encode_base64_impl( 409| 15.1k| (char *)out, src + i, srclen - i, options, line_length, offset); 410| 15.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_122lookup_pshufb_improvedILb0EEEDv4_xS3_: 30| 10.0k|simdutf_really_inline __m256i lookup_pshufb_improved(const __m256i input) { 31| | // Precomputed shuffle masks for K = 1 to 16 32| | // credit: Wojciech Muła 33| 10.0k| __m256i result = _mm256_subs_epu8(input, _mm256_set1_epi8(51)); 34| 10.0k| const __m256i less = _mm256_cmpgt_epi8(_mm256_set1_epi8(26), input); 35| 10.0k| result = 36| 10.0k| _mm256_or_si256(result, _mm256_and_si256(less, _mm256_set1_epi8(13))); 37| 10.0k| __m256i shift_LUT; 38| 10.0k| if (base64_url) { ------------------ | Branch (38:7): [Folded, False: 10.0k] ------------------ 39| 0| shift_LUT = _mm256_setr_epi8( 40| 0| 'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 41| 0| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0, 42| | 43| 0| 'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 44| 0| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0); 45| 10.0k| } else { 46| 10.0k| shift_LUT = _mm256_setr_epi8( 47| 10.0k| 'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 48| 10.0k| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0, 49| | 50| 10.0k| 'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 51| 10.0k| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0); 52| 10.0k| } 53| | 54| 10.0k| result = _mm256_shuffle_epi8(shift_LUT, result); 55| 10.0k| return _mm256_add_epi8(result, input); 56| 10.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_latin1_to_utf16ILNS_10endiannessE0EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.56k| char16_t *utf16_output) { 5| 5.56k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| | 7| 5.56k| size_t i = 0; 8| 14.9k| for (; i < rounded_len; i += 16) { ------------------ | Branch (8:10): [True: 9.36k, False: 5.56k] ------------------ 9| | // Load 16 bytes from the address (input + i) into a xmm register 10| 9.36k| const __m128i latin1 = 11| 9.36k| _mm_loadu_si128(reinterpret_cast(latin1_input + i)); 12| | 13| | // Zero extend each byte in `in` to word 14| 9.36k| __m256i utf16 = _mm256_cvtepu8_epi16(latin1); 15| | 16| 9.36k| if (big_endian) { ------------------ | Branch (16:9): [Folded, False: 9.36k] ------------------ 17| 0| const __m128i swap128 = 18| 0| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 19| 0| const __m256i swap = _mm256_set_m128i(swap128, swap128); ------------------ | | 4| 0| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 0| _mm256_castsi128_si256(xmm2), 2) ------------------ 20| 0| utf16 = _mm256_shuffle_epi8(utf16, swap); 21| 0| } 22| | 23| | // Store the contents of xmm1 into the address pointed by (output + i) 24| 9.36k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output + i), utf16); 25| 9.36k| } 26| | 27| 5.56k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 28| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_latin1_to_utf16ILNS_10endiannessE1EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.56k| char16_t *utf16_output) { 5| 5.56k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| | 7| 5.56k| size_t i = 0; 8| 14.9k| for (; i < rounded_len; i += 16) { ------------------ | Branch (8:10): [True: 9.36k, False: 5.56k] ------------------ 9| | // Load 16 bytes from the address (input + i) into a xmm register 10| 9.36k| const __m128i latin1 = 11| 9.36k| _mm_loadu_si128(reinterpret_cast(latin1_input + i)); 12| | 13| | // Zero extend each byte in `in` to word 14| 9.36k| __m256i utf16 = _mm256_cvtepu8_epi16(latin1); 15| | 16| 9.36k| if (big_endian) { ------------------ | Branch (16:9): [True: 9.36k, Folded] ------------------ 17| 9.36k| const __m128i swap128 = 18| 9.36k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 19| 9.36k| const __m256i swap = _mm256_set_m128i(swap128, swap128); ------------------ | | 4| 9.36k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 9.36k| _mm256_castsi128_si256(xmm2), 2) ------------------ 20| 9.36k| utf16 = _mm256_shuffle_epi8(utf16, swap); 21| 9.36k| } 22| | 23| | // Store the contents of xmm1 into the address pointed by (output + i) 24| 9.36k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output + i), utf16); 25| 9.36k| } 26| | 27| 5.56k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 28| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_latin1_to_utf32EPKcmPDi: 3| 5.56k| char32_t *utf32_output) { 4| 5.56k| size_t rounded_len = ((len | 7) ^ 7); // Round down to nearest multiple of 8 5| | 6| 24.5k| for (size_t i = 0; i < rounded_len; i += 8) { ------------------ | Branch (6:22): [True: 18.9k, False: 5.56k] ------------------ 7| | // Load 8 Latin1 characters into a 64-bit register 8| 18.9k| __m128i in = _mm_loadl_epi64((__m128i *)&buf[i]); 9| | 10| | // Zero extend each set of 8 Latin1 characters to 8 32-bit integers using 11| | // vpmovzxbd 12| 18.9k| __m256i out = _mm256_cvtepu8_epi32(in); 13| | 14| | // Store the results back to memory 15| 18.9k| _mm256_storeu_si256((__m256i *)&utf32_output[i], out); 16| 18.9k| } 17| | 18| | // return pointers pointing to where we left off 19| 5.56k| return std::make_pair(buf + rounded_len, utf32_output + rounded_len); 20| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_127avx2_convert_latin1_to_utf8EPKcmPc: 3| 5.83k| char *utf8_output) { 4| 5.83k| const char *end = latin1_input + len; 5| 5.83k| const __m256i v_0000 = _mm256_setzero_si256(); 6| 5.83k| const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080); 7| 5.83k| const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80); 8| 5.83k| const size_t safety_margin = 12; 9| | 10| 16.5k| while (end - latin1_input >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (10:10): [True: 10.7k, False: 5.83k] ------------------ 11| 10.7k| __m128i in8 = _mm_loadu_si128((__m128i *)latin1_input); 12| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 13| 10.7k| const __m128i v_80 = _mm_set1_epi8((char)0x80); 14| 10.7k| if (_mm_testz_si128(in8, v_80)) { // ASCII fast path!!!! ------------------ | Branch (14:9): [True: 7.98k, False: 2.74k] ------------------ 15| | // 1. store (16 bytes) 16| 7.98k| _mm_storeu_si128((__m128i *)utf8_output, in8); 17| | // 2. adjust pointers 18| 7.98k| latin1_input += 16; 19| 7.98k| utf8_output += 16; 20| 7.98k| continue; // we are done for this round! 21| 7.98k| } 22| | // We proceed only with the first 16 bytes. 23| 2.74k| const __m256i in = _mm256_cvtepu8_epi16((in8)); 24| | 25| | // 1. prepare 2-byte values 26| | // input 16-bit word : [0000|0000|aabb|bbbb] x 8 27| | // expected output : [1100|00aa|10bb|bbbb] x 8 28| 2.74k| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 29| 2.74k| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 30| | 31| | // t0 = [0000|00aa|bbbb|bb00] 32| 2.74k| const __m256i t0 = _mm256_slli_epi16(in, 2); 33| | // t1 = [0000|00aa|0000|0000] 34| 2.74k| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 35| | // t2 = [0000|0000|00bb|bbbb] 36| 2.74k| const __m256i t2 = _mm256_and_si256(in, v_003f); 37| | // t3 = [000a|aaaa|00bb|bbbb] 38| 2.74k| const __m256i t3 = _mm256_or_si256(t1, t2); 39| | // t4 = [1100|00aa|10bb|bbbb] 40| 2.74k| const __m256i t4 = _mm256_or_si256(t3, v_c080); 41| | 42| | // 2. merge ASCII and 2-byte codewords 43| | 44| | // no bits set above 7th bit 45| 2.74k| const __m256i one_byte_bytemask = 46| 2.74k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000); 47| 2.74k| const uint32_t one_byte_bitmask = 48| 2.74k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 49| | 50| 2.74k| const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in, one_byte_bytemask); 51| | 52| | // 3. prepare bitmask for 8-bit lookup 53| 2.74k| const uint32_t M0 = one_byte_bitmask & 0x55555555; 54| 2.74k| const uint32_t M1 = M0 >> 7; 55| 2.74k| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 56| | // 4. pack the bytes 57| | 58| 2.74k| const uint8_t *row = 59| 2.74k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 60| 2.74k| const uint8_t *row_2 = 61| 2.74k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 16)] 62| 2.74k| [0]; 63| | 64| 2.74k| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 65| 2.74k| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 66| | 67| 2.74k| const __m256i utf8_packed = _mm256_shuffle_epi8( 68| 2.74k| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 2.74k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 2.74k| _mm256_castsi128_si256(xmm2), 2) ------------------ 69| | // 5. store bytes 70| 2.74k| _mm_storeu_si128((__m128i *)utf8_output, 71| 2.74k| _mm256_castsi256_si128(utf8_packed)); 72| 2.74k| utf8_output += row[0]; 73| 2.74k| _mm_storeu_si128((__m128i *)utf8_output, 74| 2.74k| _mm256_extractf128_si256(utf8_packed, 1)); 75| 2.74k| utf8_output += row_2[0]; 76| | 77| | // 6. adjust pointers 78| 2.74k| latin1_input += 16; 79| 2.74k| continue; 80| | 81| 10.7k| } // while 82| 5.83k| return std::make_pair(latin1_input, utf8_output); 83| 5.83k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_utf16_to_latin1ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.56k| char *latin1_output) { 5| 5.56k| const char16_t *end = buf + len; 6| 10.1k| while (end - buf >= 32) { ------------------ | Branch (6:10): [True: 4.56k, False: 5.56k] ------------------ 7| | // Load 16 UTF-16 characters into 256-bit AVX2 register 8| 4.56k| __m256i in0 = _mm256_loadu_si256(reinterpret_cast(buf)); 9| 4.56k| __m256i in1 = 10| 4.56k| _mm256_loadu_si256(reinterpret_cast(buf + 16)); 11| | 12| | if constexpr (!match_system(big_endian)) { 13| | const __m256i swap = _mm256_setr_epi8( 14| | 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 15| | 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 16| | in0 = _mm256_shuffle_epi8(in0, swap); 17| | in1 = _mm256_shuffle_epi8(in1, swap); 18| | } 19| | 20| 4.56k| __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00); 21| 4.56k| if (_mm256_testz_si256(_mm256_or_si256(in0, in1), high_byte_mask)) { ------------------ | Branch (21:9): [True: 4.56k, False: 0] ------------------ 22| | // Pack 16-bit characters into 8-bit and store in latin1_output 23| 4.56k| const __m256i packed = _mm256_packus_epi16(in0, in1); 24| | 25| 4.56k| const __m256i result = _mm256_permute4x64_epi64(packed, 0b11011000); 26| | 27| 4.56k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(latin1_output), result); 28| | // Adjust pointers for the next iteration 29| 4.56k| buf += 32; 30| 4.56k| latin1_output += 32; 31| 4.56k| } else { 32| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 33| 0| } 34| 4.56k| } // while 35| 5.56k| return std::make_pair(buf, latin1_output); 36| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_utf16_to_latin1ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.56k| char *latin1_output) { 5| 5.56k| const char16_t *end = buf + len; 6| 10.1k| while (end - buf >= 32) { ------------------ | Branch (6:10): [True: 4.56k, False: 5.56k] ------------------ 7| | // Load 16 UTF-16 characters into 256-bit AVX2 register 8| 4.56k| __m256i in0 = _mm256_loadu_si256(reinterpret_cast(buf)); 9| 4.56k| __m256i in1 = 10| 4.56k| _mm256_loadu_si256(reinterpret_cast(buf + 16)); 11| | 12| 4.56k| if constexpr (!match_system(big_endian)) { 13| 4.56k| const __m256i swap = _mm256_setr_epi8( 14| 4.56k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 15| 4.56k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 16| 4.56k| in0 = _mm256_shuffle_epi8(in0, swap); 17| 4.56k| in1 = _mm256_shuffle_epi8(in1, swap); 18| 4.56k| } 19| | 20| 4.56k| __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00); 21| 4.56k| if (_mm256_testz_si256(_mm256_or_si256(in0, in1), high_byte_mask)) { ------------------ | Branch (21:9): [True: 4.56k, False: 0] ------------------ 22| | // Pack 16-bit characters into 8-bit and store in latin1_output 23| 4.56k| const __m256i packed = _mm256_packus_epi16(in0, in1); 24| | 25| 4.56k| const __m256i result = _mm256_permute4x64_epi64(packed, 0b11011000); 26| | 27| 4.56k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(latin1_output), result); 28| | // Adjust pointers for the next iteration 29| 4.56k| buf += 32; 30| 4.56k| latin1_output += 32; 31| 4.56k| } else { 32| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 33| 0| } 34| 4.56k| } // while 35| 5.56k| return std::make_pair(buf, latin1_output); 36| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_126avx2_convert_utf16_to_utf8ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 10.5k|avx2_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| 10.5k| const char16_t *end = buf + len; 58| 10.5k| const __m256i v_0000 = _mm256_setzero_si256(); 59| 10.5k| const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800); 60| 10.5k| const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800); 61| 10.5k| const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080); 62| 10.5k| const size_t safety_margin = 63| 10.5k| 12; // to avoid overruns, see issue 64| | // https://github.com/simdutf/simdutf/issues/92 65| | 66| 15.2k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (66:10): [True: 4.80k, False: 10.4k] ------------------ 67| 4.80k| __m256i in = _mm256_loadu_si256((__m256i *)buf); 68| 4.80k| if (big_endian) { ------------------ | Branch (68:9): [Folded, False: 4.80k] ------------------ 69| 0| const __m256i swap = _mm256_setr_epi8( 70| 0| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 71| 0| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 72| 0| in = _mm256_shuffle_epi8(in, swap); 73| 0| } 74| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 75| 4.80k| const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80); 76| 4.80k| if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 2.40k, False: 2.40k] ------------------ 77| | // 1. pack the bytes 78| 2.40k| const __m128i utf8_packed = _mm_packus_epi16( 79| 2.40k| _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1)); 80| | // 2. store (16 bytes) 81| 2.40k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 82| | // 3. adjust pointers 83| 2.40k| buf += 16; 84| 2.40k| utf8_output += 16; 85| 2.40k| continue; // we are done for this round! 86| 2.40k| } 87| | // no bits set above 7th bit 88| 2.40k| const __m256i one_byte_bytemask = 89| 2.40k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000); 90| 2.40k| const uint32_t one_byte_bitmask = 91| 2.40k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 92| | 93| | // no bits set above 11th bit 94| 2.40k| const __m256i one_or_two_bytes_bytemask = 95| 2.40k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000); 96| 2.40k| const uint32_t one_or_two_bytes_bitmask = 97| 2.40k| static_cast(_mm256_movemask_epi8(one_or_two_bytes_bytemask)); 98| 2.40k| if (one_or_two_bytes_bitmask == 0xffffffff) { ------------------ | Branch (98:9): [True: 148, False: 2.25k] ------------------ 99| | 100| | // 1. prepare 2-byte values 101| | // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8 102| | // expected output : [110a|aaaa|10bb|bbbb] x 8 103| 148| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 104| 148| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 105| | 106| | // t0 = [000a|aaaa|bbbb|bb00] 107| 148| const __m256i t0 = _mm256_slli_epi16(in, 2); 108| | // t1 = [000a|aaaa|0000|0000] 109| 148| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 110| | // t2 = [0000|0000|00bb|bbbb] 111| 148| const __m256i t2 = _mm256_and_si256(in, v_003f); 112| | // t3 = [000a|aaaa|00bb|bbbb] 113| 148| const __m256i t3 = _mm256_or_si256(t1, t2); 114| | // t4 = [110a|aaaa|10bb|bbbb] 115| 148| const __m256i t4 = _mm256_or_si256(t3, v_c080); 116| | 117| | // 2. merge ASCII and 2-byte codewords 118| 148| const __m256i utf8_unpacked = 119| 148| _mm256_blendv_epi8(t4, in, one_byte_bytemask); 120| | 121| | // 3. prepare bitmask for 8-bit lookup 122| 148| const uint32_t M0 = one_byte_bitmask & 0x55555555; 123| 148| const uint32_t M1 = M0 >> 7; 124| 148| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 125| | // 4. pack the bytes 126| | 127| 148| const uint8_t *row = 128| 148| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 129| 148| const uint8_t *row_2 = 130| 148| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 131| 148| 16)][0]; 132| | 133| 148| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 134| 148| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 135| | 136| 148| const __m256i utf8_packed = _mm256_shuffle_epi8( 137| 148| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 148| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 148| _mm256_castsi128_si256(xmm2), 2) ------------------ 138| | // 5. store bytes 139| 148| _mm_storeu_si128((__m128i *)utf8_output, 140| 148| _mm256_castsi256_si128(utf8_packed)); 141| 148| utf8_output += row[0]; 142| 148| _mm_storeu_si128((__m128i *)utf8_output, 143| 148| _mm256_extractf128_si256(utf8_packed, 1)); 144| 148| utf8_output += row_2[0]; 145| | 146| | // 6. adjust pointers 147| 148| buf += 16; 148| 148| continue; 149| 148| } 150| | // 1. Check if there are any surrogate word in the input chunk. 151| | // We have also deal with situation when there is a surrogate word 152| | // at the end of a chunk. 153| 2.25k| const __m256i surrogates_bytemask = 154| 2.25k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800); 155| | 156| | // bitmask = 0x0000 if there are no surrogates 157| | // = 0xc000 if the last word is a surrogate 158| 2.25k| const uint32_t surrogates_bitmask = 159| 2.25k| static_cast(_mm256_movemask_epi8(surrogates_bytemask)); 160| | // It might seem like checking for surrogates_bitmask == 0xc000 could help. 161| | // However, it is likely an uncommon occurrence. 162| 2.25k| if (surrogates_bitmask == 0x00000000) { ------------------ | Branch (162:9): [True: 1.04k, False: 1.20k] ------------------ 163| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 164| 1.04k| const __m256i dup_even = _mm256_setr_epi16( 165| 1.04k| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e, 166| 1.04k| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e); 167| | 168| | /* In this branch we handle three cases: 169| | 1. [0000|0000|0ccc|cccc] => [0ccc|cccc] - 170| | single UFT-8 byte 171| | 2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc] - two 172| | UTF-8 bytes 173| | 3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - 174| | three UTF-8 bytes 175| | 176| | We expand the input word (16-bit) into two code units (32-bit), thus 177| | we have room for four bytes. However, we need five distinct bit 178| | layouts. Note that the last byte in cases #2 and #3 is the same. 179| | 180| | We precompute byte 1 for case #1 and the common byte for cases #2 & #3 181| | in register t2. 182| | 183| | We precompute byte 1 for case #3 and -- **conditionally** -- precompute 184| | either byte 1 for case #2 or byte 2 for case #3. Note that they 185| | differ by exactly one bit. 186| | 187| | Finally from these two code units we build proper UTF-8 sequence, taking 188| | into account the case (i.e, the number of bytes to write). 189| | */ 190| | /** 191| | * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce: 192| | * t2 => [0ccc|cccc] [10cc|cccc] 193| | * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb]) 194| | */ 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) 196| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 197| 1.04k| const __m256i t0 = _mm256_shuffle_epi8(in, dup_even); 198| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 199| 1.04k| const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 200| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 201| 1.04k| const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 202| | 203| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 204| 1.04k| const __m256i s0 = _mm256_srli_epi16(in, 4); 205| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 206| 1.04k| const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 207| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 208| 1.04k| const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 209| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 210| 1.04k| const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 211| 1.04k| const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, 212| 1.04k| simdutf_vec(0b0100000000000000)); ------------------ | | 195| 1.04k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 213| 1.04k| const __m256i s4 = _mm256_xor_si256(s3, m0); 214| 1.04k|#undef simdutf_vec 215| | 216| | // 4. expand code units 16-bit => 32-bit 217| 1.04k| const __m256i out0 = _mm256_unpacklo_epi16(t2, s4); 218| 1.04k| const __m256i out1 = _mm256_unpackhi_epi16(t2, s4); 219| | 220| | // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle 221| 1.04k| const uint32_t mask = (one_byte_bitmask & 0x55555555) | 222| 1.04k| (one_or_two_bytes_bitmask & 0xaaaaaaaa); 223| | // Due to the wider registers, the following path is less likely to be 224| | // useful. 225| | /*if(mask == 0) { 226| | // We only have three-byte code units. Use fast path. 227| | const __m256i shuffle = 228| | _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 229| | 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 = 230| | _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 = 231| | _mm256_shuffle_epi8(out1, shuffle); 232| | _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0)); 233| | utf8_output += 12; 234| | _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1)); 235| | utf8_output += 12; 236| | _mm_storeu_si128((__m128i*)utf8_output, 237| | _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12; 238| | _mm_storeu_si128((__m128i*)utf8_output, 239| | _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16; 240| | continue; 241| | }*/ 242| 1.04k| const uint8_t mask0 = uint8_t(mask); 243| 1.04k| const uint8_t *row0 = 244| 1.04k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 245| 1.04k| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 246| 1.04k| const __m128i utf8_0 = 247| 1.04k| _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0); 248| | 249| 1.04k| const uint8_t mask1 = static_cast(mask >> 8); 250| 1.04k| const uint8_t *row1 = 251| 1.04k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 252| 1.04k| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 253| 1.04k| const __m128i utf8_1 = 254| 1.04k| _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1); 255| | 256| 1.04k| const uint8_t mask2 = static_cast(mask >> 16); 257| 1.04k| const uint8_t *row2 = 258| 1.04k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0]; 259| 1.04k| const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1)); 260| 1.04k| const __m128i utf8_2 = 261| 1.04k| _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2); 262| | 263| 1.04k| const uint8_t mask3 = static_cast(mask >> 24); 264| 1.04k| const uint8_t *row3 = 265| 1.04k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0]; 266| 1.04k| const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1)); 267| 1.04k| const __m128i utf8_3 = 268| 1.04k| _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3); 269| | 270| 1.04k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 271| 1.04k| utf8_output += row0[0]; 272| 1.04k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 273| 1.04k| utf8_output += row1[0]; 274| 1.04k| _mm_storeu_si128((__m128i *)utf8_output, utf8_2); 275| 1.04k| utf8_output += row2[0]; 276| 1.04k| _mm_storeu_si128((__m128i *)utf8_output, utf8_3); 277| 1.04k| utf8_output += row3[0]; 278| 1.04k| buf += 16; 279| | // surrogate pair(s) in a register 280| 1.20k| } else { 281| | // Let us do a scalar fallback. 282| | // It may seem wasteful to use scalar code, but being efficient with SIMD 283| | // in the presence of surrogate pairs may require non-trivial tables. 284| 1.20k| size_t forward = 15; 285| 1.20k| size_t k = 0; 286| 1.20k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (286:11): [True: 0, False: 1.20k] ------------------ 287| 0| forward = size_t(end - buf - 1); 288| 0| } 289| 16.9k| for (; k < forward; k++) { ------------------ | Branch (289:14): [True: 15.8k, False: 1.15k] ------------------ 290| 15.8k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 291| 15.8k| if ((word & 0xFF80) == 0) { ------------------ | Branch (291:13): [True: 9.41k, False: 6.43k] ------------------ 292| 9.41k| *utf8_output++ = char(word); 293| 9.41k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (293:20): [True: 508, False: 5.92k] ------------------ 294| 508| *utf8_output++ = char((word >> 6) | 0b11000000); 295| 508| *utf8_output++ = char((word & 0b111111) | 0b10000000); 296| 5.92k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (296:20): [True: 3.99k, False: 1.93k] ------------------ 297| 3.99k| *utf8_output++ = char((word >> 12) | 0b11100000); 298| 3.99k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 299| 3.99k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 300| 3.99k| } else { 301| | // must be a surrogate pair 302| 1.93k| uint16_t diff = uint16_t(word - 0xD800); 303| 1.93k| uint16_t next_word = 304| 1.93k| scalar::utf16::swap_if_needed(buf[k + 1]); 305| 1.93k| k++; 306| 1.93k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 307| 1.93k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (307:15): [True: 55, False: 1.87k] ------------------ 308| 55| return std::make_pair(nullptr, utf8_output); 309| 55| } 310| 1.87k| uint32_t value = (diff << 10) + diff2 + 0x10000; 311| 1.87k| *utf8_output++ = char((value >> 18) | 0b11110000); 312| 1.87k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 313| 1.87k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 314| 1.87k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 315| 1.87k| } 316| 15.8k| } 317| 1.15k| buf += k; 318| 1.15k| } 319| 2.25k| } // while 320| 10.4k| return std::make_pair(buf, utf8_output); 321| 10.5k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_126avx2_convert_utf16_to_utf8ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 10.5k|avx2_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| 10.5k| const char16_t *end = buf + len; 58| 10.5k| const __m256i v_0000 = _mm256_setzero_si256(); 59| 10.5k| const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800); 60| 10.5k| const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800); 61| 10.5k| const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080); 62| 10.5k| const size_t safety_margin = 63| 10.5k| 12; // to avoid overruns, see issue 64| | // https://github.com/simdutf/simdutf/issues/92 65| | 66| 15.5k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (66:10): [True: 5.12k, False: 10.4k] ------------------ 67| 5.12k| __m256i in = _mm256_loadu_si256((__m256i *)buf); 68| 5.12k| if (big_endian) { ------------------ | Branch (68:9): [True: 5.12k, Folded] ------------------ 69| 5.12k| const __m256i swap = _mm256_setr_epi8( 70| 5.12k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 71| 5.12k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 72| 5.12k| in = _mm256_shuffle_epi8(in, swap); 73| 5.12k| } 74| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 75| 5.12k| const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80); 76| 5.12k| if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 2.40k, False: 2.72k] ------------------ 77| | // 1. pack the bytes 78| 2.40k| const __m128i utf8_packed = _mm_packus_epi16( 79| 2.40k| _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1)); 80| | // 2. store (16 bytes) 81| 2.40k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 82| | // 3. adjust pointers 83| 2.40k| buf += 16; 84| 2.40k| utf8_output += 16; 85| 2.40k| continue; // we are done for this round! 86| 2.40k| } 87| | // no bits set above 7th bit 88| 2.72k| const __m256i one_byte_bytemask = 89| 2.72k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000); 90| 2.72k| const uint32_t one_byte_bitmask = 91| 2.72k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 92| | 93| | // no bits set above 11th bit 94| 2.72k| const __m256i one_or_two_bytes_bytemask = 95| 2.72k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000); 96| 2.72k| const uint32_t one_or_two_bytes_bitmask = 97| 2.72k| static_cast(_mm256_movemask_epi8(one_or_two_bytes_bytemask)); 98| 2.72k| if (one_or_two_bytes_bitmask == 0xffffffff) { ------------------ | Branch (98:9): [True: 145, False: 2.58k] ------------------ 99| | 100| | // 1. prepare 2-byte values 101| | // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8 102| | // expected output : [110a|aaaa|10bb|bbbb] x 8 103| 145| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 104| 145| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 105| | 106| | // t0 = [000a|aaaa|bbbb|bb00] 107| 145| const __m256i t0 = _mm256_slli_epi16(in, 2); 108| | // t1 = [000a|aaaa|0000|0000] 109| 145| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 110| | // t2 = [0000|0000|00bb|bbbb] 111| 145| const __m256i t2 = _mm256_and_si256(in, v_003f); 112| | // t3 = [000a|aaaa|00bb|bbbb] 113| 145| const __m256i t3 = _mm256_or_si256(t1, t2); 114| | // t4 = [110a|aaaa|10bb|bbbb] 115| 145| const __m256i t4 = _mm256_or_si256(t3, v_c080); 116| | 117| | // 2. merge ASCII and 2-byte codewords 118| 145| const __m256i utf8_unpacked = 119| 145| _mm256_blendv_epi8(t4, in, one_byte_bytemask); 120| | 121| | // 3. prepare bitmask for 8-bit lookup 122| 145| const uint32_t M0 = one_byte_bitmask & 0x55555555; 123| 145| const uint32_t M1 = M0 >> 7; 124| 145| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 125| | // 4. pack the bytes 126| | 127| 145| const uint8_t *row = 128| 145| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 129| 145| const uint8_t *row_2 = 130| 145| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 131| 145| 16)][0]; 132| | 133| 145| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 134| 145| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 135| | 136| 145| const __m256i utf8_packed = _mm256_shuffle_epi8( 137| 145| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 145| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 145| _mm256_castsi128_si256(xmm2), 2) ------------------ 138| | // 5. store bytes 139| 145| _mm_storeu_si128((__m128i *)utf8_output, 140| 145| _mm256_castsi256_si128(utf8_packed)); 141| 145| utf8_output += row[0]; 142| 145| _mm_storeu_si128((__m128i *)utf8_output, 143| 145| _mm256_extractf128_si256(utf8_packed, 1)); 144| 145| utf8_output += row_2[0]; 145| | 146| | // 6. adjust pointers 147| 145| buf += 16; 148| 145| continue; 149| 145| } 150| | // 1. Check if there are any surrogate word in the input chunk. 151| | // We have also deal with situation when there is a surrogate word 152| | // at the end of a chunk. 153| 2.58k| const __m256i surrogates_bytemask = 154| 2.58k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800); 155| | 156| | // bitmask = 0x0000 if there are no surrogates 157| | // = 0xc000 if the last word is a surrogate 158| 2.58k| const uint32_t surrogates_bitmask = 159| 2.58k| static_cast(_mm256_movemask_epi8(surrogates_bytemask)); 160| | // It might seem like checking for surrogates_bitmask == 0xc000 could help. 161| | // However, it is likely an uncommon occurrence. 162| 2.58k| if (surrogates_bitmask == 0x00000000) { ------------------ | Branch (162:9): [True: 1.44k, False: 1.13k] ------------------ 163| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 164| 1.44k| const __m256i dup_even = _mm256_setr_epi16( 165| 1.44k| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e, 166| 1.44k| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e); 167| | 168| | /* In this branch we handle three cases: 169| | 1. [0000|0000|0ccc|cccc] => [0ccc|cccc] - 170| | single UFT-8 byte 171| | 2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc] - two 172| | UTF-8 bytes 173| | 3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - 174| | three UTF-8 bytes 175| | 176| | We expand the input word (16-bit) into two code units (32-bit), thus 177| | we have room for four bytes. However, we need five distinct bit 178| | layouts. Note that the last byte in cases #2 and #3 is the same. 179| | 180| | We precompute byte 1 for case #1 and the common byte for cases #2 & #3 181| | in register t2. 182| | 183| | We precompute byte 1 for case #3 and -- **conditionally** -- precompute 184| | either byte 1 for case #2 or byte 2 for case #3. Note that they 185| | differ by exactly one bit. 186| | 187| | Finally from these two code units we build proper UTF-8 sequence, taking 188| | into account the case (i.e, the number of bytes to write). 189| | */ 190| | /** 191| | * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce: 192| | * t2 => [0ccc|cccc] [10cc|cccc] 193| | * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb]) 194| | */ 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) 196| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 197| 1.44k| const __m256i t0 = _mm256_shuffle_epi8(in, dup_even); 198| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 199| 1.44k| const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 200| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 201| 1.44k| const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 202| | 203| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 204| 1.44k| const __m256i s0 = _mm256_srli_epi16(in, 4); 205| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 206| 1.44k| const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 207| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 208| 1.44k| const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 209| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 210| 1.44k| const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 211| 1.44k| const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, 212| 1.44k| simdutf_vec(0b0100000000000000)); ------------------ | | 195| 1.44k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 213| 1.44k| const __m256i s4 = _mm256_xor_si256(s3, m0); 214| 1.44k|#undef simdutf_vec 215| | 216| | // 4. expand code units 16-bit => 32-bit 217| 1.44k| const __m256i out0 = _mm256_unpacklo_epi16(t2, s4); 218| 1.44k| const __m256i out1 = _mm256_unpackhi_epi16(t2, s4); 219| | 220| | // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle 221| 1.44k| const uint32_t mask = (one_byte_bitmask & 0x55555555) | 222| 1.44k| (one_or_two_bytes_bitmask & 0xaaaaaaaa); 223| | // Due to the wider registers, the following path is less likely to be 224| | // useful. 225| | /*if(mask == 0) { 226| | // We only have three-byte code units. Use fast path. 227| | const __m256i shuffle = 228| | _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 229| | 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 = 230| | _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 = 231| | _mm256_shuffle_epi8(out1, shuffle); 232| | _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0)); 233| | utf8_output += 12; 234| | _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1)); 235| | utf8_output += 12; 236| | _mm_storeu_si128((__m128i*)utf8_output, 237| | _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12; 238| | _mm_storeu_si128((__m128i*)utf8_output, 239| | _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16; 240| | continue; 241| | }*/ 242| 1.44k| const uint8_t mask0 = uint8_t(mask); 243| 1.44k| const uint8_t *row0 = 244| 1.44k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 245| 1.44k| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 246| 1.44k| const __m128i utf8_0 = 247| 1.44k| _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0); 248| | 249| 1.44k| const uint8_t mask1 = static_cast(mask >> 8); 250| 1.44k| const uint8_t *row1 = 251| 1.44k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 252| 1.44k| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 253| 1.44k| const __m128i utf8_1 = 254| 1.44k| _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1); 255| | 256| 1.44k| const uint8_t mask2 = static_cast(mask >> 16); 257| 1.44k| const uint8_t *row2 = 258| 1.44k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0]; 259| 1.44k| const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1)); 260| 1.44k| const __m128i utf8_2 = 261| 1.44k| _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2); 262| | 263| 1.44k| const uint8_t mask3 = static_cast(mask >> 24); 264| 1.44k| const uint8_t *row3 = 265| 1.44k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0]; 266| 1.44k| const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1)); 267| 1.44k| const __m128i utf8_3 = 268| 1.44k| _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3); 269| | 270| 1.44k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 271| 1.44k| utf8_output += row0[0]; 272| 1.44k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 273| 1.44k| utf8_output += row1[0]; 274| 1.44k| _mm_storeu_si128((__m128i *)utf8_output, utf8_2); 275| 1.44k| utf8_output += row2[0]; 276| 1.44k| _mm_storeu_si128((__m128i *)utf8_output, utf8_3); 277| 1.44k| utf8_output += row3[0]; 278| 1.44k| buf += 16; 279| | // surrogate pair(s) in a register 280| 1.44k| } else { 281| | // Let us do a scalar fallback. 282| | // It may seem wasteful to use scalar code, but being efficient with SIMD 283| | // in the presence of surrogate pairs may require non-trivial tables. 284| 1.13k| size_t forward = 15; 285| 1.13k| size_t k = 0; 286| 1.13k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (286:11): [True: 0, False: 1.13k] ------------------ 287| 0| forward = size_t(end - buf - 1); 288| 0| } 289| 16.0k| for (; k < forward; k++) { ------------------ | Branch (289:14): [True: 14.9k, False: 1.08k] ------------------ 290| 14.9k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 291| 14.9k| if ((word & 0xFF80) == 0) { ------------------ | Branch (291:13): [True: 9.30k, False: 5.63k] ------------------ 292| 9.30k| *utf8_output++ = char(word); 293| 9.30k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (293:20): [True: 682, False: 4.95k] ------------------ 294| 682| *utf8_output++ = char((word >> 6) | 0b11000000); 295| 682| *utf8_output++ = char((word & 0b111111) | 0b10000000); 296| 4.95k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (296:20): [True: 3.16k, False: 1.79k] ------------------ 297| 3.16k| *utf8_output++ = char((word >> 12) | 0b11100000); 298| 3.16k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 299| 3.16k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 300| 3.16k| } else { 301| | // must be a surrogate pair 302| 1.79k| uint16_t diff = uint16_t(word - 0xD800); 303| 1.79k| uint16_t next_word = 304| 1.79k| scalar::utf16::swap_if_needed(buf[k + 1]); 305| 1.79k| k++; 306| 1.79k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 307| 1.79k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (307:15): [True: 52, False: 1.73k] ------------------ 308| 52| return std::make_pair(nullptr, utf8_output); 309| 52| } 310| 1.73k| uint32_t value = (diff << 10) + diff2 + 0x10000; 311| 1.73k| *utf8_output++ = char((value >> 18) | 0b11110000); 312| 1.73k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 313| 1.73k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 314| 1.73k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 315| 1.73k| } 316| 14.9k| } 317| 1.08k| buf += k; 318| 1.08k| } 319| 2.58k| } // while 320| 10.4k| return std::make_pair(buf, utf8_output); 321| 10.5k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_utf32_to_latin1EPKDimPc: 3| 5.56k| char *latin1_output) { 4| 5.56k| const size_t rounded_len = 5| 5.56k| len & ~0x1F; // Round down to nearest multiple of 32 6| | 7| 5.56k| const __m256i high_bytes_mask = _mm256_set1_epi32(0xFFFFFF00); 8| | 9| 10.1k| for (size_t i = 0; i < rounded_len; i += 4 * 8) { ------------------ | Branch (9:22): [True: 4.56k, False: 5.56k] ------------------ 10| 4.56k| __m256i a = _mm256_loadu_si256((__m256i *)(buf + 0 * 8)); 11| 4.56k| __m256i b = _mm256_loadu_si256((__m256i *)(buf + 1 * 8)); 12| 4.56k| __m256i c = _mm256_loadu_si256((__m256i *)(buf + 2 * 8)); 13| 4.56k| __m256i d = _mm256_loadu_si256((__m256i *)(buf + 3 * 8)); 14| | 15| 4.56k| const __m256i check_combined = 16| 4.56k| _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d)); 17| | 18| 4.56k| if (!_mm256_testz_si256(check_combined, high_bytes_mask)) { ------------------ | Branch (18:9): [True: 0, False: 4.56k] ------------------ 19| 0| return std::make_pair(nullptr, latin1_output); 20| 0| } 21| | 22| 4.56k| b = _mm256_slli_epi32(b, 1 * 8); 23| 4.56k| c = _mm256_slli_epi32(c, 2 * 8); 24| 4.56k| d = _mm256_slli_epi32(d, 3 * 8); 25| | 26| | // clang-format off 27| | 28| | // a = [.. .. .. a7|.. .. .. a6|.. .. .. a5|.. .. .. a4||.. .. .. a3|.. .. .. a2|.. .. .. a1|.. .. .. a0] 29| | // b = [.. .. b7 ..|.. .. b6 ..|.. .. b5 ..|.. .. b4 ..||.. .. b3 ..|.. .. b2 ..|.. .. b1 ..|.. .. b0 ..] 30| | // c = [.. c7 .. ..|.. c6 .. ..|.. c5 .. ..|.. c4 .. ..||.. c3 .. ..|.. c2 .. ..|.. c1 .. ..|.. c0 .. ..] 31| | // d = [d7 .. .. ..|d6 .. .. ..|d5 .. .. ..|d4 .. .. ..||d3 .. .. ..|d2 .. .. ..|d1 .. .. ..|d0 .. .. ..] 32| | 33| | // t0 = [d7 c7 b7 a7|d6 c6 b6 a6|d5 c5 b5 a5|d4 c4 b4 a4||d3 c3 b3 a3|d2 c2 b2 a2|d1 c1 b1 a1|d0 c0 b0 a0] 34| 4.56k| const __m256i t0 = 35| 4.56k| _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d)); 36| | 37| | // shuffle bytes within 128-bit lanes 38| | // t1 = [d7 d6 d5 d4|c7 c6 c5 c4|b7 b6 b5 b4|a7 a6 a5 a4||d3 d2 d1 d0|c3 c2 c1 c0|b3 b2 b1 b0|a3 a2 a1 a0] 39| 4.56k| const __m256i shuffle_bytes = 40| 4.56k| _mm256_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15, 41| 4.56k| 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15); 42| | 43| 4.56k| const __m256i t1 = _mm256_shuffle_epi8(t0, shuffle_bytes); 44| | 45| | // reshuffle dwords 46| | // t2 = [d7 d6 d5 d4|d3 d2 d1 d0|c7 c6 c5 c4|c3 c2 c1 c0||b7 b6 b5 b4|b3 b2 b1 b0|a7 a6 a5 a4|a3 a2 a1 a0] 47| 4.56k| const __m256i shuffle_dwords = _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7); 48| 4.56k| const __m256i t2 = _mm256_permutevar8x32_epi32(t1, shuffle_dwords); 49| |// clang format on 50| | 51| 4.56k| _mm256_storeu_si256((__m256i *)latin1_output, t2); 52| | 53| 4.56k| latin1_output += 32; 54| 4.56k| buf += 32; 55| 4.56k| } 56| | 57| 5.56k| return std::make_pair(buf, latin1_output); 58| 5.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_126avx2_convert_utf32_to_utf8EPKDimPc: 2| 4.96k|avx2_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) { 3| 4.96k| const char32_t *end = buf + len; 4| 4.96k| const __m256i v_0000 = _mm256_setzero_si256(); 5| 4.96k| const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000); 6| 4.96k| const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80); 7| 4.96k| const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800); 8| 4.96k| const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080); 9| 4.96k| const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff); 10| 4.96k| __m256i running_max = _mm256_setzero_si256(); 11| 4.96k| __m256i forbidden_bytemask = _mm256_setzero_si256(); 12| | 13| 4.96k| const size_t safety_margin = 14| 4.96k| 12; // to avoid overruns, see issue 15| | // https://github.com/simdutf/simdutf/issues/92 16| | 17| 8.22k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (17:10): [True: 3.26k, False: 4.96k] ------------------ 18| 3.26k| __m256i in = _mm256_loadu_si256((__m256i *)buf); 19| 3.26k| __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1); 20| 3.26k| running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin); 21| | 22| | // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned 23| | // saturation 24| 3.26k| __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff), 25| 3.26k| _mm256_and_si256(nextin, v_7fffffff)); 26| 3.26k| in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000); 27| | 28| | // Try to apply UTF-16 => UTF-8 routine on 256 bits 29| | // (haswell/avx2_convert_utf16_to_utf8.cpp) 30| | 31| 3.26k| if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (31:9): [True: 2.27k, False: 990] ------------------ 32| | // 1. pack the bytes 33| 2.27k| const __m128i utf8_packed = _mm_packus_epi16( 34| 2.27k| _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1)); 35| | // 2. store (16 bytes) 36| 2.27k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 37| | // 3. adjust pointers 38| 2.27k| buf += 16; 39| 2.27k| utf8_output += 16; 40| 2.27k| continue; // we are done for this round! 41| 2.27k| } 42| | // no bits set above 7th bit 43| 990| const __m256i one_byte_bytemask = 44| 990| _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000); 45| 990| const uint32_t one_byte_bitmask = 46| 990| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 47| | 48| | // no bits set above 11th bit 49| 990| const __m256i one_or_two_bytes_bytemask = 50| 990| _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000); 51| 990| const uint32_t one_or_two_bytes_bitmask = 52| 990| static_cast(_mm256_movemask_epi8(one_or_two_bytes_bytemask)); 53| 990| if (one_or_two_bytes_bitmask == 0xffffffff) { ------------------ | Branch (53:9): [True: 123, False: 867] ------------------ 54| | // 1. prepare 2-byte values 55| | // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8 56| | // expected output : [110a|aaaa|10bb|bbbb] x 8 57| 123| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 58| 123| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 59| | 60| | // t0 = [000a|aaaa|bbbb|bb00] 61| 123| const __m256i t0 = _mm256_slli_epi16(in_16, 2); 62| | // t1 = [000a|aaaa|0000|0000] 63| 123| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 64| | // t2 = [0000|0000|00bb|bbbb] 65| 123| const __m256i t2 = _mm256_and_si256(in_16, v_003f); 66| | // t3 = [000a|aaaa|00bb|bbbb] 67| 123| const __m256i t3 = _mm256_or_si256(t1, t2); 68| | // t4 = [110a|aaaa|10bb|bbbb] 69| 123| const __m256i t4 = _mm256_or_si256(t3, v_c080); 70| | 71| | // 2. merge ASCII and 2-byte codewords 72| 123| const __m256i utf8_unpacked = 73| 123| _mm256_blendv_epi8(t4, in_16, one_byte_bytemask); 74| | 75| | // 3. prepare bitmask for 8-bit lookup 76| 123| const uint32_t M0 = one_byte_bitmask & 0x55555555; 77| 123| const uint32_t M1 = M0 >> 7; 78| 123| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 79| | // 4. pack the bytes 80| | 81| 123| const uint8_t *row = 82| 123| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 83| 123| const uint8_t *row_2 = 84| 123| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 85| 123| 16)][0]; 86| | 87| 123| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 88| 123| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 89| | 90| 123| const __m256i utf8_packed = _mm256_shuffle_epi8( 91| 123| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 123| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 123| _mm256_castsi128_si256(xmm2), 2) ------------------ 92| | // 5. store bytes 93| 123| _mm_storeu_si128((__m128i *)utf8_output, 94| 123| _mm256_castsi256_si128(utf8_packed)); 95| 123| utf8_output += row[0]; 96| 123| _mm_storeu_si128((__m128i *)utf8_output, 97| 123| _mm256_extractf128_si256(utf8_packed, 1)); 98| 123| utf8_output += row_2[0]; 99| | 100| | // 6. adjust pointers 101| 123| buf += 16; 102| 123| continue; 103| 123| } 104| | // Must check for overflow in packing 105| 867| const __m256i saturation_bytemask = _mm256_cmpeq_epi32( 106| 867| _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000); 107| 867| const uint32_t saturation_bitmask = 108| 867| static_cast(_mm256_movemask_epi8(saturation_bytemask)); 109| 867| if (saturation_bitmask == 0xffffffff) { ------------------ | Branch (109:9): [True: 107, False: 760] ------------------ 110| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 111| 107| const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800); 112| 107| forbidden_bytemask = _mm256_or_si256( 113| 107| forbidden_bytemask, 114| 107| _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800)); 115| | 116| 107| const __m256i dup_even = _mm256_setr_epi16( 117| 107| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e, 118| 107| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e); 119| | 120| | /* In this branch we handle three cases: 121| | 1. [0000|0000|0ccc|cccc] => [0ccc|cccc] - 122| | single UFT-8 byte 123| | 2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc] - two 124| | UTF-8 bytes 125| | 3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - 126| | three UTF-8 bytes 127| | 128| | We expand the input word (16-bit) into two code units (32-bit), thus 129| | we have room for four bytes. However, we need five distinct bit 130| | layouts. Note that the last byte in cases #2 and #3 is the same. 131| | 132| | We precompute byte 1 for case #1 and the common byte for cases #2 & #3 133| | in register t2. 134| | 135| | We precompute byte 1 for case #3 and -- **conditionally** -- precompute 136| | either byte 1 for case #2 or byte 2 for case #3. Note that they 137| | differ by exactly one bit. 138| | 139| | Finally from these two code units we build proper UTF-8 sequence, taking 140| | into account the case (i.e, the number of bytes to write). 141| | */ 142| | /** 143| | * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce: 144| | * t2 => [0ccc|cccc] [10cc|cccc] 145| | * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb]) 146| | */ 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) 148| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 149| 107| const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even); 150| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 151| 107| const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 152| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 153| 107| const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 154| | 155| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 156| 107| const __m256i s0 = _mm256_srli_epi16(in_16, 4); 157| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 158| 107| const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 159| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 160| 107| const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 161| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 162| 107| const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 163| 107| const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, 164| 107| simdutf_vec(0b0100000000000000)); ------------------ | | 147| 107|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 165| 107| const __m256i s4 = _mm256_xor_si256(s3, m0); 166| 107|#undef simdutf_vec 167| | 168| | // 4. expand code units 16-bit => 32-bit 169| 107| const __m256i out0 = _mm256_unpacklo_epi16(t2, s4); 170| 107| const __m256i out1 = _mm256_unpackhi_epi16(t2, s4); 171| | 172| | // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle 173| 107| const uint32_t mask = (one_byte_bitmask & 0x55555555) | 174| 107| (one_or_two_bytes_bitmask & 0xaaaaaaaa); 175| | // Due to the wider registers, the following path is less likely to be 176| | // useful. 177| | /*if(mask == 0) { 178| | // We only have three-byte code units. Use fast path. 179| | const __m256i shuffle = 180| | _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 181| | 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 = 182| | _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 = 183| | _mm256_shuffle_epi8(out1, shuffle); 184| | _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0)); 185| | utf8_output += 12; 186| | _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1)); 187| | utf8_output += 12; 188| | _mm_storeu_si128((__m128i*)utf8_output, 189| | _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12; 190| | _mm_storeu_si128((__m128i*)utf8_output, 191| | _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16; 192| | continue; 193| | }*/ 194| 107| const uint8_t mask0 = uint8_t(mask); 195| 107| const uint8_t *row0 = 196| 107| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 197| 107| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 198| 107| const __m128i utf8_0 = 199| 107| _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0); 200| | 201| 107| const uint8_t mask1 = static_cast(mask >> 8); 202| 107| const uint8_t *row1 = 203| 107| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 204| 107| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 205| 107| const __m128i utf8_1 = 206| 107| _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1); 207| | 208| 107| const uint8_t mask2 = static_cast(mask >> 16); 209| 107| const uint8_t *row2 = 210| 107| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0]; 211| 107| const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1)); 212| 107| const __m128i utf8_2 = 213| 107| _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2); 214| | 215| 107| const uint8_t mask3 = static_cast(mask >> 24); 216| 107| const uint8_t *row3 = 217| 107| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0]; 218| 107| const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1)); 219| 107| const __m128i utf8_3 = 220| 107| _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3); 221| | 222| 107| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 223| 107| utf8_output += row0[0]; 224| 107| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 225| 107| utf8_output += row1[0]; 226| 107| _mm_storeu_si128((__m128i *)utf8_output, utf8_2); 227| 107| utf8_output += row2[0]; 228| 107| _mm_storeu_si128((__m128i *)utf8_output, utf8_3); 229| 107| utf8_output += row3[0]; 230| 107| buf += 16; 231| 760| } else { 232| | // case: at least one 32-bit word is larger than 0xFFFF <=> it will 233| | // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem 234| | // wasteful to use scalar code, but being efficient with SIMD may require 235| | // large, non-trivial tables? 236| 760| size_t forward = 15; 237| 760| size_t k = 0; 238| 760| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (238:11): [True: 0, False: 760] ------------------ 239| 0| forward = size_t(end - buf - 1); 240| 0| } 241| 12.1k| for (; k < forward; k++) { ------------------ | Branch (241:14): [True: 11.4k, False: 760] ------------------ 242| 11.4k| uint32_t word = buf[k]; 243| 11.4k| if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII) ------------------ | Branch (243:13): [True: 9.29k, False: 2.10k] ------------------ 244| 9.29k| *utf8_output++ = char(word); 245| 9.29k| } else if ((word & 0xFFFFF800) == 0) { // 2-byte ------------------ | Branch (245:20): [True: 450, False: 1.65k] ------------------ 246| 450| *utf8_output++ = char((word >> 6) | 0b11000000); 247| 450| *utf8_output++ = char((word & 0b111111) | 0b10000000); 248| 1.65k| } else if ((word & 0xFFFF0000) == 0) { // 3-byte ------------------ | Branch (248:20): [True: 494, False: 1.16k] ------------------ 249| 494| if (word >= 0xD800 && word <= 0xDFFF) { ------------------ | Branch (249:15): [True: 230, False: 264] | Branch (249:33): [True: 0, False: 230] ------------------ 250| 0| return std::make_pair(nullptr, utf8_output); 251| 0| } 252| 494| *utf8_output++ = char((word >> 12) | 0b11100000); 253| 494| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 254| 494| *utf8_output++ = char((word & 0b111111) | 0b10000000); 255| 1.16k| } else { // 4-byte 256| 1.16k| if (word > 0x10FFFF) { ------------------ | Branch (256:15): [True: 0, False: 1.16k] ------------------ 257| 0| return std::make_pair(nullptr, utf8_output); 258| 0| } 259| 1.16k| *utf8_output++ = char((word >> 18) | 0b11110000); 260| 1.16k| *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000); 261| 1.16k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 262| 1.16k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 263| 1.16k| } 264| 11.4k| } 265| 760| buf += k; 266| 760| } 267| 867| } // while 268| | 269| | // check for invalid input 270| 4.96k| const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff); 271| 4.96k| if (static_cast(_mm256_movemask_epi8(_mm256_cmpeq_epi32( ------------------ | Branch (271:7): [True: 0, False: 4.96k] ------------------ 272| 4.96k| _mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) { 273| 0| return std::make_pair(nullptr, utf8_output); 274| 0| } 275| | 276| 4.96k| if (static_cast(_mm256_movemask_epi8(forbidden_bytemask)) != 0) { ------------------ | Branch (276:7): [True: 0, False: 4.96k] ------------------ 277| 0| return std::make_pair(nullptr, utf8_output); 278| 0| } 279| | 280| 4.96k| return std::make_pair(buf, utf8_output); 281| 4.96k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_129convert_masked_utf8_to_latin1EPKcmRPc: 9| 15.5k| char *&latin1_output) { 10| | // we use an approach where we try to process up to 12 input bytes. 11| | // Why 12 input bytes and not 16? Because we are concerned with the size of 12| | // the lookup tables. Also 12 is nicely divisible by two and three. 13| | // 14| | // 15| | // Optimization note: our main path below is load-latency dependent. Thus it 16| | // is maybe beneficial to have fast paths that depend on branch prediction but 17| | // have less latency. This results in more instructions but, potentially, also 18| | // higher speeds. 19| | // 20| 15.5k| const __m128i in = _mm_loadu_si128((__m128i *)input); 21| | 22| 15.5k| const uint16_t input_utf8_end_of_code_point_mask = 23| 15.5k| utf8_end_of_code_point_mask & 24| 15.5k| 0xfff; // we are only processing 12 bytes in case it is not all ASCII 25| | 26| 15.5k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (26:7): [True: 120, False: 15.4k] ------------------ 27| | // We process the data in chunks of 12 bytes. 28| 120| _mm_storeu_si128(reinterpret_cast<__m128i *>(latin1_output), in); 29| 120| latin1_output += 12; // We wrote 12 characters. 30| 120| return 12; // We consumed 1 bytes. 31| 120| } 32| | /// We do not have a fast path available, so we fallback. 33| 15.4k| const uint8_t idx = 34| 15.4k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 35| 15.4k| const uint8_t consumed = 36| 15.4k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 37| | // this indicates an invalid input: 38| 15.4k| if (idx >= 64) { ------------------ | Branch (38:7): [True: 5.03k, False: 10.4k] ------------------ 39| 5.03k| return consumed; 40| 5.03k| } 41| | // Here we should have (idx < 64), if not, there is a bug in the validation or 42| | // elsewhere. SIX (6) input code-code units this is a relatively easy scenario 43| | // we process SIX (6) input code-code units. The max length in bytes of six 44| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 45| | // processors where pdep/pext is fast, we might be able to use a small lookup 46| | // table. 47| 10.4k| const __m128i sh = 48| 10.4k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 49| 10.4k| const __m128i perm = _mm_shuffle_epi8(in, sh); 50| 10.4k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 51| 10.4k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 52| 10.4k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 53| 10.4k| const __m128i latin1_packed = _mm_packus_epi16(composed, composed); 54| | // writing 8 bytes even though we only care about the first 6 bytes. 55| | // performance note: it would be faster to use _mm_storeu_si128, we should 56| | // investigate. 57| 10.4k| _mm_storel_epi64((__m128i *)latin1_output, latin1_packed); 58| 10.4k| latin1_output += 6; // We wrote 6 bytes. 59| 10.4k| return consumed; 60| 15.4k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE0EEEmPKcmRPDs: 10| 14.3k| char16_t *&utf16_output) { 11| | // we use an approach where we try to process up to 12 input bytes. 12| | // Why 12 input bytes and not 16? Because we are concerned with the size of 13| | // the lookup tables. Also 12 is nicely divisible by two and three. 14| | // 15| | // 16| | // Optimization note: our main path below is load-latency dependent. Thus it 17| | // is maybe beneficial to have fast paths that depend on branch prediction but 18| | // have less latency. This results in more instructions but, potentially, also 19| | // higher speeds. 20| | // 21| | // We first try a few fast paths. 22| 14.3k| const __m128i swap = 23| 14.3k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 14.3k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 14.3k| const uint16_t input_utf8_end_of_code_point_mask = 26| 14.3k| utf8_end_of_code_point_mask & 0xfff; 27| 14.3k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 92, False: 14.2k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| 92| __m256i ascii = _mm256_cvtepu8_epi16(in); 30| 92| if (big_endian) { ------------------ | Branch (30:9): [Folded, False: 92] ------------------ 31| 0| const __m256i swap256 = _mm256_setr_epi8( 32| 0| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 33| 0| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 34| 0| ascii = _mm256_shuffle_epi8(ascii, swap256); 35| 0| } 36| 92| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output), ascii); 37| 92| utf16_output += 12; // We wrote 12 16-bit characters. 38| 92| return 12; // We consumed 12 bytes. 39| 92| } 40| 14.2k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (40:7): [True: 129, False: 14.1k] ------------------ 41| | // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte 42| | // UTF-16 code units. There is probably a more efficient sequence, but the 43| | // following might do. 44| 129| const __m128i sh = 45| 129| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 46| 129| const __m128i perm = _mm_shuffle_epi8(in, sh); 47| 129| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 48| 129| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 49| 129| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 50| 129| if (big_endian) ------------------ | Branch (50:9): [Folded, False: 129] ------------------ 51| 0| composed = _mm_shuffle_epi8(composed, swap); 52| 129| _mm_storeu_si128((__m128i *)utf16_output, composed); 53| 129| utf16_output += 8; // We wrote 16 bytes, 8 code points. 54| 129| return 16; 55| 129| } 56| 14.1k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (56:7): [True: 3.61k, False: 10.4k] ------------------ 57| | // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte 58| | // UTF-16 code units. There is probably a more efficient sequence, but the 59| | // following might do. 60| 3.61k| const __m128i sh = 61| 3.61k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 62| 3.61k| const __m128i perm = _mm_shuffle_epi8(in, sh); 63| 3.61k| const __m128i ascii = 64| 3.61k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 65| 3.61k| const __m128i middlebyte = 66| 3.61k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 67| 3.61k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 68| 3.61k| const __m128i highbyte = 69| 3.61k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 70| 3.61k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 71| 3.61k| const __m128i composed = 72| 3.61k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 73| 3.61k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 74| 3.61k| if (big_endian) ------------------ | Branch (74:9): [Folded, False: 3.61k] ------------------ 75| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 76| 3.61k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 77| 3.61k| utf16_output += 4; 78| 3.61k| return 12; 79| 3.61k| } 80| | 81| 10.4k| const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex 82| 10.4k| [input_utf8_end_of_code_point_mask][0]; 83| 10.4k| const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex 84| 10.4k| [input_utf8_end_of_code_point_mask][1]; 85| 10.4k| if (idx < 64) { ------------------ | Branch (85:7): [True: 3.94k, False: 6.54k] ------------------ 86| | // SIX (6) input code-code units 87| | // this is a relatively easy scenario 88| | // we process SIX (6) input code-code units. The max length in bytes of six 89| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 90| | // processors where pdep/pext is fast, we might be able to use a small 91| | // lookup table. 92| 3.94k| const __m128i sh = _mm_loadu_si128( 93| 3.94k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 94| 3.94k| const __m128i perm = _mm_shuffle_epi8(in, sh); 95| 3.94k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 96| 3.94k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 97| 3.94k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 98| 3.94k| if (big_endian) ------------------ | Branch (98:9): [Folded, False: 3.94k] ------------------ 99| 0| composed = _mm_shuffle_epi8(composed, swap); 100| 3.94k| _mm_storeu_si128((__m128i *)utf16_output, composed); 101| 3.94k| utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential 102| | // overflow of 4 bytes. 103| 6.54k| } else if (idx < 145) { ------------------ | Branch (103:14): [True: 1.64k, False: 4.90k] ------------------ 104| | // FOUR (4) input code-code units 105| 1.64k| const __m128i sh = _mm_loadu_si128( 106| 1.64k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 107| 1.64k| const __m128i perm = _mm_shuffle_epi8(in, sh); 108| 1.64k| const __m128i ascii = 109| 1.64k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 110| 1.64k| const __m128i middlebyte = 111| 1.64k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 112| 1.64k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 113| 1.64k| const __m128i highbyte = 114| 1.64k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 115| 1.64k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 116| 1.64k| const __m128i composed = 117| 1.64k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 118| 1.64k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 119| 1.64k| if (big_endian) ------------------ | Branch (119:9): [Folded, False: 1.64k] ------------------ 120| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 121| 1.64k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 122| 1.64k| utf16_output += 4; // Here we overflow by 8 bytes. 123| 4.90k| } else if (idx < 209) { ------------------ | Branch (123:14): [True: 4.71k, False: 191] ------------------ 124| | // TWO (2) input code-code units 125| | ////////////// 126| | // There might be garbage inputs where a leading byte mascarades as a 127| | // four-byte leading byte (by being followed by 3 continuation byte), but is 128| | // not greater than 0xf0. This could trigger a buffer overflow if we only 129| | // counted leading bytes of the form 0xf0 as generating surrogate pairs, 130| | // without further UTF-8 validation. Thus we must be careful to ensure that 131| | // only leading bytes at least as large as 0xf0 generate surrogate pairs. We 132| | // do as at the cost of an extra mask. 133| | ///////////// 134| 4.71k| const __m128i sh = _mm_loadu_si128( 135| 4.71k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 136| 4.71k| const __m128i perm = _mm_shuffle_epi8(in, sh); 137| 4.71k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 138| 4.71k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 139| 4.71k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 140| 4.71k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 141| | // correct for spurious high bit 142| 4.71k| const __m128i correct = 143| 4.71k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 144| 4.71k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 145| 4.71k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 146| | // We deliberately carry the leading four bits in highbyte if they are 147| | // present, we remove them later when computing hightenbits. 148| 4.71k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 149| 4.71k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 150| | // When we need to generate a surrogate pair (leading byte > 0xF0), then 151| | // the corresponding 32-bit value in 'composed' will be greater than 152| | // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the 153| | // location of the surrogate pairs. 154| 4.71k| const __m128i composed = 155| 4.71k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 156| 4.71k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 157| 4.71k| const __m128i composedminus = 158| 4.71k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 159| 4.71k| const __m128i lowtenbits = 160| 4.71k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 161| | // Notice the 0x3ff mask: 162| 4.71k| const __m128i hightenbits = 163| 4.71k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 164| 4.71k| const __m128i lowtenbitsadd = 165| 4.71k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 166| 4.71k| const __m128i hightenbitsadd = 167| 4.71k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 168| 4.71k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 169| 4.71k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 170| 4.71k| uint32_t basic_buffer[4]; 171| 4.71k| uint32_t basic_buffer_swap[4]; 172| 4.71k| if (big_endian) { ------------------ | Branch (172:9): [Folded, False: 4.71k] ------------------ 173| 0| _mm_storeu_si128((__m128i *)basic_buffer_swap, 174| 0| _mm_shuffle_epi8(composed, swap)); 175| 0| surrogates = _mm_shuffle_epi8(surrogates, swap); 176| 0| } 177| 4.71k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 178| 4.71k| uint32_t surrogate_buffer[4]; 179| 4.71k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 180| 18.8k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (180:24): [True: 14.1k, False: 4.71k] ------------------ 181| 14.1k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (181:11): [True: 2.82k, False: 11.3k] ------------------ 182| 2.82k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 183| 2.82k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 184| 2.82k| utf16_output += 2; 185| 11.3k| } else { 186| 11.3k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (186:27): [Folded, False: 11.3k] ------------------ 187| 11.3k| : uint16_t(basic_buffer[i]); 188| 11.3k| utf16_output++; 189| 11.3k| } 190| 14.1k| } 191| 4.71k| } else { 192| | // here we know that there is an error but we do not handle errors 193| 191| } 194| 10.4k| return consumed; 195| 14.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE1EEEmPKcmRPDs: 10| 15.6k| char16_t *&utf16_output) { 11| | // we use an approach where we try to process up to 12 input bytes. 12| | // Why 12 input bytes and not 16? Because we are concerned with the size of 13| | // the lookup tables. Also 12 is nicely divisible by two and three. 14| | // 15| | // 16| | // Optimization note: our main path below is load-latency dependent. Thus it 17| | // is maybe beneficial to have fast paths that depend on branch prediction but 18| | // have less latency. This results in more instructions but, potentially, also 19| | // higher speeds. 20| | // 21| | // We first try a few fast paths. 22| 15.6k| const __m128i swap = 23| 15.6k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 15.6k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 15.6k| const uint16_t input_utf8_end_of_code_point_mask = 26| 15.6k| utf8_end_of_code_point_mask & 0xfff; 27| 15.6k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 95, False: 15.5k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| 95| __m256i ascii = _mm256_cvtepu8_epi16(in); 30| 95| if (big_endian) { ------------------ | Branch (30:9): [True: 95, Folded] ------------------ 31| 95| const __m256i swap256 = _mm256_setr_epi8( 32| 95| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 33| 95| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 34| 95| ascii = _mm256_shuffle_epi8(ascii, swap256); 35| 95| } 36| 95| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output), ascii); 37| 95| utf16_output += 12; // We wrote 12 16-bit characters. 38| 95| return 12; // We consumed 12 bytes. 39| 95| } 40| 15.5k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (40:7): [True: 118, False: 15.3k] ------------------ 41| | // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte 42| | // UTF-16 code units. There is probably a more efficient sequence, but the 43| | // following might do. 44| 118| const __m128i sh = 45| 118| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 46| 118| const __m128i perm = _mm_shuffle_epi8(in, sh); 47| 118| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 48| 118| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 49| 118| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 50| 118| if (big_endian) ------------------ | Branch (50:9): [True: 118, Folded] ------------------ 51| 118| composed = _mm_shuffle_epi8(composed, swap); 52| 118| _mm_storeu_si128((__m128i *)utf16_output, composed); 53| 118| utf16_output += 8; // We wrote 16 bytes, 8 code points. 54| 118| return 16; 55| 118| } 56| 15.3k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (56:7): [True: 4.63k, False: 10.7k] ------------------ 57| | // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte 58| | // UTF-16 code units. There is probably a more efficient sequence, but the 59| | // following might do. 60| 4.63k| const __m128i sh = 61| 4.63k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 62| 4.63k| const __m128i perm = _mm_shuffle_epi8(in, sh); 63| 4.63k| const __m128i ascii = 64| 4.63k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 65| 4.63k| const __m128i middlebyte = 66| 4.63k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 67| 4.63k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 68| 4.63k| const __m128i highbyte = 69| 4.63k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 70| 4.63k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 71| 4.63k| const __m128i composed = 72| 4.63k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 73| 4.63k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 74| 4.63k| if (big_endian) ------------------ | Branch (74:9): [True: 4.63k, Folded] ------------------ 75| 4.63k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 76| 4.63k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 77| 4.63k| utf16_output += 4; 78| 4.63k| return 12; 79| 4.63k| } 80| | 81| 10.7k| const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex 82| 10.7k| [input_utf8_end_of_code_point_mask][0]; 83| 10.7k| const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex 84| 10.7k| [input_utf8_end_of_code_point_mask][1]; 85| 10.7k| if (idx < 64) { ------------------ | Branch (85:7): [True: 4.05k, False: 6.70k] ------------------ 86| | // SIX (6) input code-code units 87| | // this is a relatively easy scenario 88| | // we process SIX (6) input code-code units. The max length in bytes of six 89| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 90| | // processors where pdep/pext is fast, we might be able to use a small 91| | // lookup table. 92| 4.05k| const __m128i sh = _mm_loadu_si128( 93| 4.05k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 94| 4.05k| const __m128i perm = _mm_shuffle_epi8(in, sh); 95| 4.05k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 96| 4.05k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 97| 4.05k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 98| 4.05k| if (big_endian) ------------------ | Branch (98:9): [True: 4.05k, Folded] ------------------ 99| 4.05k| composed = _mm_shuffle_epi8(composed, swap); 100| 4.05k| _mm_storeu_si128((__m128i *)utf16_output, composed); 101| 4.05k| utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential 102| | // overflow of 4 bytes. 103| 6.70k| } else if (idx < 145) { ------------------ | Branch (103:14): [True: 1.88k, False: 4.81k] ------------------ 104| | // FOUR (4) input code-code units 105| 1.88k| const __m128i sh = _mm_loadu_si128( 106| 1.88k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 107| 1.88k| const __m128i perm = _mm_shuffle_epi8(in, sh); 108| 1.88k| const __m128i ascii = 109| 1.88k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 110| 1.88k| const __m128i middlebyte = 111| 1.88k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 112| 1.88k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 113| 1.88k| const __m128i highbyte = 114| 1.88k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 115| 1.88k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 116| 1.88k| const __m128i composed = 117| 1.88k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 118| 1.88k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 119| 1.88k| if (big_endian) ------------------ | Branch (119:9): [True: 1.88k, Folded] ------------------ 120| 1.88k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 121| 1.88k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 122| 1.88k| utf16_output += 4; // Here we overflow by 8 bytes. 123| 4.81k| } else if (idx < 209) { ------------------ | Branch (123:14): [True: 4.62k, False: 191] ------------------ 124| | // TWO (2) input code-code units 125| | ////////////// 126| | // There might be garbage inputs where a leading byte mascarades as a 127| | // four-byte leading byte (by being followed by 3 continuation byte), but is 128| | // not greater than 0xf0. This could trigger a buffer overflow if we only 129| | // counted leading bytes of the form 0xf0 as generating surrogate pairs, 130| | // without further UTF-8 validation. Thus we must be careful to ensure that 131| | // only leading bytes at least as large as 0xf0 generate surrogate pairs. We 132| | // do as at the cost of an extra mask. 133| | ///////////// 134| 4.62k| const __m128i sh = _mm_loadu_si128( 135| 4.62k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 136| 4.62k| const __m128i perm = _mm_shuffle_epi8(in, sh); 137| 4.62k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 138| 4.62k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 139| 4.62k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 140| 4.62k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 141| | // correct for spurious high bit 142| 4.62k| const __m128i correct = 143| 4.62k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 144| 4.62k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 145| 4.62k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 146| | // We deliberately carry the leading four bits in highbyte if they are 147| | // present, we remove them later when computing hightenbits. 148| 4.62k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 149| 4.62k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 150| | // When we need to generate a surrogate pair (leading byte > 0xF0), then 151| | // the corresponding 32-bit value in 'composed' will be greater than 152| | // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the 153| | // location of the surrogate pairs. 154| 4.62k| const __m128i composed = 155| 4.62k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 156| 4.62k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 157| 4.62k| const __m128i composedminus = 158| 4.62k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 159| 4.62k| const __m128i lowtenbits = 160| 4.62k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 161| | // Notice the 0x3ff mask: 162| 4.62k| const __m128i hightenbits = 163| 4.62k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 164| 4.62k| const __m128i lowtenbitsadd = 165| 4.62k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 166| 4.62k| const __m128i hightenbitsadd = 167| 4.62k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 168| 4.62k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 169| 4.62k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 170| 4.62k| uint32_t basic_buffer[4]; 171| 4.62k| uint32_t basic_buffer_swap[4]; 172| 4.62k| if (big_endian) { ------------------ | Branch (172:9): [True: 4.62k, Folded] ------------------ 173| 4.62k| _mm_storeu_si128((__m128i *)basic_buffer_swap, 174| 4.62k| _mm_shuffle_epi8(composed, swap)); 175| 4.62k| surrogates = _mm_shuffle_epi8(surrogates, swap); 176| 4.62k| } 177| 4.62k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 178| 4.62k| uint32_t surrogate_buffer[4]; 179| 4.62k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 180| 18.5k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (180:24): [True: 13.8k, False: 4.62k] ------------------ 181| 13.8k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (181:11): [True: 2.77k, False: 11.1k] ------------------ 182| 2.77k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 183| 2.77k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 184| 2.77k| utf16_output += 2; 185| 11.1k| } else { 186| 11.1k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (186:27): [True: 11.1k, Folded] ------------------ 187| 11.1k| : uint16_t(basic_buffer[i]); 188| 11.1k| utf16_output++; 189| 11.1k| } 190| 13.8k| } 191| 4.62k| } else { 192| | // here we know that there is an error but we do not handle errors 193| 191| } 194| 10.7k| return consumed; 195| 15.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128convert_masked_utf8_to_utf32EPKcmRPDi: 9| 9.80k| char32_t *&utf32_output) { 10| | // we use an approach where we try to process up to 12 input bytes. 11| | // Why 12 input bytes and not 16? Because we are concerned with the size of 12| | // the lookup tables. Also 12 is nicely divisible by two and three. 13| | // 14| | // 15| | // Optimization note: our main path below is load-latency dependent. Thus it 16| | // is maybe beneficial to have fast paths that depend on branch prediction but 17| | // have less latency. This results in more instructions but, potentially, also 18| | // higher speeds. 19| | // 20| | // We first try a few fast paths. 21| 9.80k| const __m128i in = _mm_loadu_si128((__m128i *)input); 22| 9.80k| const uint16_t input_utf8_end_of_code_point_mask = 23| 9.80k| utf8_end_of_code_point_mask & 0xfff; 24| 9.80k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (24:7): [True: 92, False: 9.71k] ------------------ 25| | // We process the data in chunks of 12 bytes. 26| 92| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output), 27| 92| _mm256_cvtepu8_epi32(in)); 28| 92| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output + 8), 29| 92| _mm256_cvtepu8_epi32(_mm_srli_si128(in, 8))); 30| 92| utf32_output += 12; // We wrote 12 32-bit characters. 31| 92| return 12; // We consumed 12 bytes. 32| 92| } 33| 9.71k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (33:7): [True: 103, False: 9.61k] ------------------ 34| | // We want to take 8 2-byte UTF-8 code units and turn them into 8 4-byte 35| | // UTF-32 code units. There is probably a more efficient sequence, but the 36| | // following might do. 37| 103| const __m128i sh = 38| 103| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 39| 103| const __m128i perm = _mm_shuffle_epi8(in, sh); 40| 103| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 41| 103| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 42| 103| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 43| 103| _mm256_storeu_si256((__m256i *)utf32_output, 44| 103| _mm256_cvtepu16_epi32(composed)); 45| 103| utf32_output += 8; // We wrote 16 bytes, 8 code points. 46| 103| return 16; 47| 103| } 48| 9.61k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (48:7): [True: 169, False: 9.44k] ------------------ 49| | // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte 50| | // UTF-32 code units. There is probably a more efficient sequence, but the 51| | // following might do. 52| 169| const __m128i sh = 53| 169| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 54| 169| const __m128i perm = _mm_shuffle_epi8(in, sh); 55| 169| const __m128i ascii = 56| 169| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 57| 169| const __m128i middlebyte = 58| 169| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 59| 169| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 60| 169| const __m128i highbyte = 61| 169| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 62| 169| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 63| 169| const __m128i composed = 64| 169| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 65| 169| _mm_storeu_si128((__m128i *)utf32_output, composed); 66| 169| utf32_output += 4; 67| 169| return 12; 68| 169| } 69| | /// We do not have a fast path available, so we fallback. 70| | 71| 9.44k| const uint8_t idx = 72| 9.44k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 73| 9.44k| const uint8_t consumed = 74| 9.44k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 75| 9.44k| if (idx < 64) { ------------------ | Branch (75:7): [True: 3.86k, False: 5.57k] ------------------ 76| | // SIX (6) input code-code units 77| | // this is a relatively easy scenario 78| | // we process SIX (6) input code-code units. The max length in bytes of six 79| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 80| | // processors where pdep/pext is fast, we might be able to use a small 81| | // lookup table. 82| 3.86k| const __m128i sh = 83| 3.86k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 84| 3.86k| const __m128i perm = _mm_shuffle_epi8(in, sh); 85| 3.86k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 86| 3.86k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 87| 3.86k| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 88| 3.86k| _mm256_storeu_si256((__m256i *)utf32_output, 89| 3.86k| _mm256_cvtepu16_epi32(composed)); 90| 3.86k| utf32_output += 6; // We wrote 24 bytes, 6 code points. There is a potential 91| | // overflow of 32 - 24 = 8 bytes. 92| 5.57k| } else if (idx < 145) { ------------------ | Branch (92:14): [True: 1.17k, False: 4.40k] ------------------ 93| | // FOUR (4) input code-code units 94| 1.17k| const __m128i sh = 95| 1.17k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 96| 1.17k| const __m128i perm = _mm_shuffle_epi8(in, sh); 97| 1.17k| const __m128i ascii = 98| 1.17k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 99| 1.17k| const __m128i middlebyte = 100| 1.17k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 101| 1.17k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 102| 1.17k| const __m128i highbyte = 103| 1.17k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 104| 1.17k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 105| 1.17k| const __m128i composed = 106| 1.17k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 107| 1.17k| _mm_storeu_si128((__m128i *)utf32_output, composed); 108| 1.17k| utf32_output += 4; 109| 4.40k| } else if (idx < 209) { ------------------ | Branch (109:14): [True: 4.20k, False: 191] ------------------ 110| | // TWO (2) input code-code units 111| 4.20k| const __m128i sh = 112| 4.20k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 113| 4.20k| const __m128i perm = _mm_shuffle_epi8(in, sh); 114| 4.20k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 115| 4.20k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 116| 4.20k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 117| 4.20k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 118| | // correct for spurious high bit 119| 4.20k| const __m128i correct = 120| 4.20k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 121| 4.20k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 122| 4.20k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 123| 4.20k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000)); 124| 4.20k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 125| 4.20k| const __m128i composed = 126| 4.20k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 127| 4.20k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 128| 4.20k| _mm_storeu_si128((__m128i *)utf32_output, composed); 129| 4.20k| utf32_output += 130| 4.20k| 3; // We wrote 3 * 4 bytes, there is a potential overflow of 4 bytes. 131| 4.20k| } else { 132| | // here we know that there is an error but we do not handle errors 133| 191| } 134| 9.44k| return consumed; 135| 9.61k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1623utf16_gather_high_bytesILNS_10endiannessE0EEENS1_4simd5simd8IhEERKNS5_6simd16ItEESB_: 3| 11.6k| const simd16 &in1) { 4| 11.6k| if (big_endian) { ------------------ | Branch (4:7): [Folded, False: 11.6k] ------------------ 5| | // we want lower bytes 6| 0| const auto mask = simd16(0x00ff); 7| 0| const auto t0 = in0 & mask; 8| 0| const auto t1 = in1 & mask; 9| | 10| 0| return simd16::pack(t0, t1); 11| 11.6k| } else { 12| 11.6k| const auto t0 = in0.shr<8>(); 13| 11.6k| const auto t1 = in1.shr<8>(); 14| | 15| 11.6k| return simd16::pack(t0, t1); 16| 11.6k| } 17| 11.6k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1623utf16_gather_high_bytesILNS_10endiannessE1EEENS1_4simd5simd8IhEERKNS5_6simd16ItEESB_: 3| 12.1k| const simd16 &in1) { 4| 12.1k| if (big_endian) { ------------------ | Branch (4:7): [True: 12.1k, Folded] ------------------ 5| | // we want lower bytes 6| 12.1k| const auto mask = simd16(0x00ff); 7| 12.1k| const auto t0 = in0 & mask; 8| 12.1k| const auto t1 = in1 & mask; 9| | 10| 12.1k| return simd16::pack(t0, t1); 11| 12.1k| } else { 12| 0| const auto t0 = in0.shr<8>(); 13| 0| const auto t1 = in1.shr<8>(); 14| | 15| 0| return simd16::pack(t0, t1); 16| 0| } 17| 12.1k|} _ZNK7simdutf7haswell14implementation13validate_utf8EPKcm: 264| 22.0k|implementation::validate_utf8(const char *buf, size_t len) const noexcept { 265| 22.0k| return haswell::utf8_validation::generic_validate_utf8(buf, len); 266| 22.0k|} _ZNK7simdutf7haswell14implementation25validate_utf8_with_errorsEPKcm: 271| 5.56k| const char *buf, size_t len) const noexcept { 272| 5.56k| return haswell::utf8_validation::generic_validate_utf8_with_errors(buf, len); 273| 5.56k|} _ZNK7simdutf7haswell14implementation16validate_utf16leEPKDsm: 310| 16.0k| size_t len) const noexcept { 311| 16.0k| if (simdutf_unlikely(len == 0)) { ------------------ | | 92| 16.0k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 13.7k, False: 2.38k] | | ------------------ ------------------ 312| | // empty input is valid UTF-16. protect the implementation from 313| | // handling nullptr 314| 13.7k| return true; 315| 13.7k| } 316| 2.38k| const auto res = 317| 2.38k| haswell::utf16::validate_utf16_with_errors(buf, len); 318| 2.38k| if (res.is_err()) { ------------------ | Branch (318:7): [True: 42, False: 2.34k] ------------------ 319| 42| return false; 320| 42| } 321| | 322| 2.34k| if (res.count == len) { ------------------ | Branch (322:7): [True: 0, False: 2.34k] ------------------ 323| 0| return true; 324| 0| } 325| | 326| 2.34k| return scalar::utf16::validate(buf + res.count, 327| 2.34k| len - res.count); 328| 2.34k|} _ZNK7simdutf7haswell14implementation16validate_utf16beEPKDsm: 334| 16.0k| size_t len) const noexcept { 335| 16.0k| if (simdutf_unlikely(len == 0)) { ------------------ | | 92| 16.0k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 13.6k, False: 2.39k] | | ------------------ ------------------ 336| | // empty input is valid UTF-16. protect the implementation from 337| | // handling nullptr 338| 13.6k| return true; 339| 13.6k| } 340| 2.39k| const auto res = 341| 2.39k| haswell::utf16::validate_utf16_with_errors(buf, len); 342| 2.39k| if (res.is_err()) { ------------------ | Branch (342:7): [True: 39, False: 2.35k] ------------------ 343| 39| return false; 344| 39| } 345| | 346| 2.35k| if (res.count == len) { ------------------ | Branch (346:7): [True: 0, False: 2.35k] ------------------ 347| 0| return true; 348| 0| } 349| | 350| 2.35k| return scalar::utf16::validate(buf + res.count, 351| 2.35k| len - res.count); 352| 2.35k|} _ZNK7simdutf7haswell14implementation28validate_utf16le_with_errorsEPKDsm: 355| 5.56k| const char16_t *buf, size_t len) const noexcept { 356| | 357| 5.56k| const result res = 358| 5.56k| haswell::utf16::validate_utf16_with_errors(buf, len); 359| 5.56k| if (res.count != len) { ------------------ | Branch (359:7): [True: 355, False: 5.20k] ------------------ 360| 355| const result scalar_res = 361| 355| scalar::utf16::validate_with_errors( 362| 355| buf + res.count, len - res.count); 363| 355| return result(scalar_res.error, res.count + scalar_res.count); 364| 5.20k| } else { 365| 5.20k| return res; 366| 5.20k| } 367| 5.56k|} _ZNK7simdutf7haswell14implementation28validate_utf16be_with_errorsEPKDsm: 370| 5.56k| const char16_t *buf, size_t len) const noexcept { 371| 5.56k| const result res = 372| 5.56k| haswell::utf16::validate_utf16_with_errors(buf, len); 373| 5.56k| if (res.count != len) { ------------------ | Branch (373:7): [True: 363, False: 5.19k] ------------------ 374| 363| const result scalar_res = 375| 363| scalar::utf16::validate_with_errors(buf + res.count, 376| 363| len - res.count); 377| 363| return result(scalar_res.error, res.count + scalar_res.count); 378| 5.19k| } else { 379| 5.19k| return res; 380| 5.19k| } 381| 5.56k|} _ZNK7simdutf7haswell14implementation14validate_utf32EPKDim: 396| 10.5k|implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept { 397| 10.5k| return utf32::validate(buf, len); 398| 10.5k|} _ZNK7simdutf7haswell14implementation22convert_latin1_to_utf8EPKcmPc: 410| 5.83k| const char *buf, size_t len, char *utf8_output) const noexcept { 411| 5.83k| std::pair ret = 412| 5.83k| avx2_convert_latin1_to_utf8(buf, len, utf8_output); 413| 5.83k| size_t converted_chars = ret.second - utf8_output; 414| | 415| 5.83k| if (ret.first != buf + len) { ------------------ | Branch (415:7): [True: 1.44k, False: 4.39k] ------------------ 416| 1.44k| const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert( 417| 1.44k| ret.first, len - (ret.first - buf), ret.second); 418| 1.44k| converted_chars += scalar_converted_chars; 419| 1.44k| } 420| | 421| 5.83k| return converted_chars; 422| 5.83k|} _ZNK7simdutf7haswell14implementation25convert_latin1_to_utf16leEPKcmPDs: 427| 5.56k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 428| 5.56k| std::pair ret = 429| 5.56k| avx2_convert_latin1_to_utf16(buf, len, utf16_output); 430| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (430:7): [True: 0, False: 5.56k] ------------------ 431| 0| return 0; 432| 0| } 433| 5.56k| size_t converted_chars = ret.second - utf16_output; 434| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (434:7): [True: 808, False: 4.75k] ------------------ 435| 808| const size_t scalar_converted_chars = 436| 808| scalar::latin1_to_utf16::convert( 437| 808| ret.first, len - (ret.first - buf), ret.second); 438| 808| if (scalar_converted_chars == 0) { ------------------ | Branch (438:9): [True: 0, False: 808] ------------------ 439| 0| return 0; 440| 0| } 441| 808| converted_chars += scalar_converted_chars; 442| 808| } 443| 5.56k| return converted_chars; 444| 5.56k|} _ZNK7simdutf7haswell14implementation25convert_latin1_to_utf16beEPKcmPDs: 447| 5.56k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 448| 5.56k| std::pair ret = 449| 5.56k| avx2_convert_latin1_to_utf16(buf, len, utf16_output); 450| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (450:7): [True: 0, False: 5.56k] ------------------ 451| 0| return 0; 452| 0| } 453| 5.56k| size_t converted_chars = ret.second - utf16_output; 454| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (454:7): [True: 808, False: 4.75k] ------------------ 455| 808| const size_t scalar_converted_chars = 456| 808| scalar::latin1_to_utf16::convert( 457| 808| ret.first, len - (ret.first - buf), ret.second); 458| 808| if (scalar_converted_chars == 0) { ------------------ | Branch (458:9): [True: 0, False: 808] ------------------ 459| 0| return 0; 460| 0| } 461| 808| converted_chars += scalar_converted_chars; 462| 808| } 463| 5.56k| return converted_chars; 464| 5.56k|} _ZNK7simdutf7haswell14implementation23convert_latin1_to_utf32EPKcmPDi: 469| 5.56k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 470| 5.56k| std::pair ret = 471| 5.56k| avx2_convert_latin1_to_utf32(buf, len, utf32_output); 472| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (472:7): [True: 0, False: 5.56k] ------------------ 473| 0| return 0; 474| 0| } 475| 5.56k| size_t converted_chars = ret.second - utf32_output; 476| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (476:7): [True: 759, False: 4.80k] ------------------ 477| 759| const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert( 478| 759| ret.first, len - (ret.first - buf), ret.second); 479| 759| if (scalar_converted_chars == 0) { ------------------ | Branch (479:9): [True: 0, False: 759] ------------------ 480| 0| return 0; 481| 0| } 482| 759| converted_chars += scalar_converted_chars; 483| 759| } 484| 5.56k| return converted_chars; 485| 5.56k|} _ZNK7simdutf7haswell14implementation22convert_utf8_to_latin1EPKcmPc: 490| 11.1k| const char *buf, size_t len, char *latin1_output) const noexcept { 491| 11.1k| utf8_to_latin1::validating_transcoder converter; 492| 11.1k| return converter.convert(buf, len, latin1_output); 493| 11.1k|} _ZNK7simdutf7haswell14implementation23convert_utf8_to_utf16leEPKcmPDs: 509| 11.0k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 510| 11.0k| utf8_to_utf16::validating_transcoder converter; 511| 11.0k| return converter.convert(buf, len, utf16_output); 512| 11.0k|} _ZNK7simdutf7haswell14implementation23convert_utf8_to_utf16beEPKcmPDs: 515| 11.0k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 516| 11.0k| utf8_to_utf16::validating_transcoder converter; 517| 11.0k| return converter.convert(buf, len, utf16_output); 518| 11.0k|} _ZNK7simdutf7haswell14implementation21convert_utf8_to_utf32EPKcmPDi: 548| 5.56k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 549| 5.56k| utf8_to_utf32::validating_transcoder converter; 550| 5.56k| return converter.convert(buf, len, utf32_output); 551| 5.56k|} _ZNK7simdutf7haswell14implementation25convert_utf16le_to_latin1EPKDsmPc: 567| 5.56k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 568| 5.56k| std::pair ret = 569| 5.56k| haswell::avx2_convert_utf16_to_latin1(buf, len, 570| 5.56k| latin1_output); 571| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (571:7): [True: 0, False: 5.56k] ------------------ 572| 0| return 0; 573| 0| } 574| 5.56k| size_t saved_bytes = ret.second - latin1_output; 575| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (575:7): [True: 828, False: 4.73k] ------------------ 576| 828| const size_t scalar_saved_bytes = 577| 828| scalar::utf16_to_latin1::convert( 578| 828| ret.first, len - (ret.first - buf), ret.second); 579| 828| if (scalar_saved_bytes == 0) { ------------------ | Branch (579:9): [True: 0, False: 828] ------------------ 580| 0| return 0; 581| 0| } 582| 828| saved_bytes += scalar_saved_bytes; 583| 828| } 584| 5.56k| return saved_bytes; 585| 5.56k|} _ZNK7simdutf7haswell14implementation25convert_utf16be_to_latin1EPKDsmPc: 588| 5.56k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 589| 5.56k| std::pair ret = 590| 5.56k| haswell::avx2_convert_utf16_to_latin1(buf, len, 591| 5.56k| latin1_output); 592| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (592:7): [True: 0, False: 5.56k] ------------------ 593| 0| return 0; 594| 0| } 595| 5.56k| size_t saved_bytes = ret.second - latin1_output; 596| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (596:7): [True: 828, False: 4.73k] ------------------ 597| 828| const size_t scalar_saved_bytes = 598| 828| scalar::utf16_to_latin1::convert( 599| 828| ret.first, len - (ret.first - buf), ret.second); 600| 828| if (scalar_saved_bytes == 0) { ------------------ | Branch (600:9): [True: 0, False: 828] ------------------ 601| 0| return 0; 602| 0| } 603| 828| saved_bytes += scalar_saved_bytes; 604| 828| } 605| 5.56k| return saved_bytes; 606| 5.56k|} _ZNK7simdutf7haswell14implementation23convert_utf16le_to_utf8EPKDsmPc: 677| 10.5k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 678| 10.5k| std::pair ret = 679| 10.5k| haswell::avx2_convert_utf16_to_utf8(buf, len, 680| 10.5k| utf8_output); 681| 10.5k| if (ret.first == nullptr) { ------------------ | Branch (681:7): [True: 55, False: 10.4k] ------------------ 682| 55| return 0; 683| 55| } 684| 10.4k| size_t saved_bytes = ret.second - utf8_output; 685| 10.4k| if (ret.first != buf + len) { ------------------ | Branch (685:7): [True: 1.16k, False: 9.30k] ------------------ 686| 1.16k| const size_t scalar_saved_bytes = 687| 1.16k| scalar::utf16_to_utf8::convert( 688| 1.16k| ret.first, len - (ret.first - buf), ret.second); 689| 1.16k| if (scalar_saved_bytes == 0) { ------------------ | Branch (689:9): [True: 37, False: 1.12k] ------------------ 690| 37| return 0; 691| 37| } 692| 1.12k| saved_bytes += scalar_saved_bytes; 693| 1.12k| } 694| 10.4k| return saved_bytes; 695| 10.4k|} _ZNK7simdutf7haswell14implementation23convert_utf16be_to_utf8EPKDsmPc: 698| 10.5k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 699| 10.5k| std::pair ret = 700| 10.5k| haswell::avx2_convert_utf16_to_utf8(buf, len, 701| 10.5k| utf8_output); 702| 10.5k| if (ret.first == nullptr) { ------------------ | Branch (702:7): [True: 52, False: 10.4k] ------------------ 703| 52| return 0; 704| 52| } 705| 10.4k| size_t saved_bytes = ret.second - utf8_output; 706| 10.4k| if (ret.first != buf + len) { ------------------ | Branch (706:7): [True: 1.17k, False: 9.30k] ------------------ 707| 1.17k| const size_t scalar_saved_bytes = 708| 1.17k| scalar::utf16_to_utf8::convert( 709| 1.17k| ret.first, len - (ret.first - buf), ret.second); 710| 1.17k| if (scalar_saved_bytes == 0) { ------------------ | Branch (710:9): [True: 32, False: 1.14k] ------------------ 711| 32| return 0; 712| 32| } 713| 1.14k| saved_bytes += scalar_saved_bytes; 714| 1.14k| } 715| 10.4k| return saved_bytes; 716| 10.4k|} _ZNK7simdutf7haswell14implementation21convert_utf32_to_utf8EPKDimPc: 787| 4.96k| const char32_t *buf, size_t len, char *utf8_output) const noexcept { 788| 4.96k| std::pair ret = 789| 4.96k| avx2_convert_utf32_to_utf8(buf, len, utf8_output); 790| 4.96k| if (ret.first == nullptr) { ------------------ | Branch (790:7): [True: 0, False: 4.96k] ------------------ 791| 0| return 0; 792| 0| } 793| 4.96k| size_t saved_bytes = ret.second - utf8_output; 794| 4.96k| if (ret.first != buf + len) { ------------------ | Branch (794:7): [True: 861, False: 4.10k] ------------------ 795| 861| const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert( 796| 861| ret.first, len - (ret.first - buf), ret.second); 797| 861| if (scalar_saved_bytes == 0) { ------------------ | Branch (797:9): [True: 0, False: 861] ------------------ 798| 0| return 0; 799| 0| } 800| 861| saved_bytes += scalar_saved_bytes; 801| 861| } 802| 4.96k| return saved_bytes; 803| 4.96k|} _ZNK7simdutf7haswell14implementation23convert_utf32_to_latin1EPKDimPc: 808| 5.56k| const char32_t *buf, size_t len, char *latin1_output) const noexcept { 809| 5.56k| std::pair ret = 810| 5.56k| avx2_convert_utf32_to_latin1(buf, len, latin1_output); 811| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (811:7): [True: 0, False: 5.56k] ------------------ 812| 0| return 0; 813| 0| } 814| 5.56k| size_t saved_bytes = ret.second - latin1_output; 815| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (815:7): [True: 828, False: 4.73k] ------------------ 816| 828| const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert( 817| 828| ret.first, len - (ret.first - buf), ret.second); 818| 828| if (scalar_saved_bytes == 0) { ------------------ | Branch (818:9): [True: 0, False: 828] ------------------ 819| 0| return 0; 820| 0| } 821| 828| saved_bytes += scalar_saved_bytes; 822| 828| } 823| 5.56k| return saved_bytes; 824| 5.56k|} _ZNK7simdutf7haswell14implementation10count_utf8EPKcm: 1115| 11.1k|implementation::count_utf8(const char *in, size_t size) const noexcept { 1116| 11.1k| return utf8::count_code_points_bytemask(in, size); 1117| 11.1k|} _ZNK7simdutf7haswell14implementation23latin1_length_from_utf8EPKcm: 1122| 11.1k| const char *buf, size_t len) const noexcept { 1123| 11.1k| return count_utf8(buf, len); 1124| 11.1k|} _ZNK7simdutf7haswell14implementation24utf8_length_from_utf16leEPKDsm: 1129| 10.5k| const char16_t *input, size_t length) const noexcept { 1130| 10.5k| return utf16::utf8_length_from_utf16_bytemask(input, 1131| 10.5k| length); 1132| 10.5k|} _ZNK7simdutf7haswell14implementation24utf8_length_from_utf16beEPKDsm: 1135| 10.5k| const char16_t *input, size_t length) const noexcept { 1136| 10.5k| return utf16::utf8_length_from_utf16_bytemask(input, length); 1137| 10.5k|} _ZNK7simdutf7haswell14implementation22utf16_length_from_utf8EPKcm: 1154| 22.0k| const char *input, size_t length) const noexcept { 1155| 22.0k| return utf8::utf16_length_from_utf8_bytemask(input, length); 1156| 22.0k|} _ZNK7simdutf7haswell14implementation23utf8_length_from_latin1EPKcm: 1189| 5.83k| const char *input, size_t len) const noexcept { 1190| 5.83k| const uint8_t *data = reinterpret_cast(input); 1191| 5.83k| size_t answer = len / sizeof(__m256i) * sizeof(__m256i); 1192| 5.83k| size_t i = 0; 1193| 5.83k| if (answer >= 2048) { // long strings optimization ------------------ | Branch (1193:7): [True: 0, False: 5.83k] ------------------ 1194| 0| __m256i four_64bits = _mm256_setzero_si256(); 1195| 0| while (i + sizeof(__m256i) <= len) { ------------------ | Branch (1195:12): [True: 0, False: 0] ------------------ 1196| 0| __m256i runner = _mm256_setzero_si256(); 1197| | // We can do up to 255 loops without overflow. 1198| 0| size_t iterations = (len - i) / sizeof(__m256i); 1199| 0| if (iterations > 255) { ------------------ | Branch (1199:11): [True: 0, False: 0] ------------------ 1200| 0| iterations = 255; 1201| 0| } 1202| 0| size_t max_i = i + iterations * sizeof(__m256i) - sizeof(__m256i); 1203| 0| for (; i + 4 * sizeof(__m256i) <= max_i; i += 4 * sizeof(__m256i)) { ------------------ | Branch (1203:14): [True: 0, False: 0] ------------------ 1204| 0| __m256i input1 = _mm256_loadu_si256((const __m256i *)(data + i)); 1205| 0| __m256i input2 = 1206| 0| _mm256_loadu_si256((const __m256i *)(data + i + sizeof(__m256i))); 1207| 0| __m256i input3 = _mm256_loadu_si256( 1208| 0| (const __m256i *)(data + i + 2 * sizeof(__m256i))); 1209| 0| __m256i input4 = _mm256_loadu_si256( 1210| 0| (const __m256i *)(data + i + 3 * sizeof(__m256i))); 1211| 0| __m256i input12 = 1212| 0| _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input1), 1213| 0| _mm256_cmpgt_epi8(_mm256_setzero_si256(), input2)); 1214| 0| __m256i input23 = 1215| 0| _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input3), 1216| 0| _mm256_cmpgt_epi8(_mm256_setzero_si256(), input4)); 1217| 0| __m256i input1234 = _mm256_add_epi8(input12, input23); 1218| 0| runner = _mm256_sub_epi8(runner, input1234); 1219| 0| } 1220| 0| for (; i <= max_i; i += sizeof(__m256i)) { ------------------ | Branch (1220:14): [True: 0, False: 0] ------------------ 1221| 0| __m256i input_256_chunk = 1222| 0| _mm256_loadu_si256((const __m256i *)(data + i)); 1223| 0| runner = _mm256_sub_epi8( 1224| 0| runner, _mm256_cmpgt_epi8(_mm256_setzero_si256(), input_256_chunk)); 1225| 0| } 1226| 0| four_64bits = _mm256_add_epi64( 1227| 0| four_64bits, _mm256_sad_epu8(runner, _mm256_setzero_si256())); 1228| 0| } 1229| 0| answer += _mm256_extract_epi64(four_64bits, 0) + 1230| 0| _mm256_extract_epi64(four_64bits, 1) + 1231| 0| _mm256_extract_epi64(four_64bits, 2) + 1232| 0| _mm256_extract_epi64(four_64bits, 3); 1233| 5.83k| } else if (answer > 0) { ------------------ | Branch (1233:14): [True: 865, False: 4.97k] ------------------ 1234| 6.51k| for (; i + sizeof(__m256i) <= len; i += sizeof(__m256i)) { ------------------ | Branch (1234:12): [True: 5.64k, False: 865] ------------------ 1235| 5.64k| __m256i latin = _mm256_loadu_si256((const __m256i *)(data + i)); 1236| 5.64k| uint32_t non_ascii = _mm256_movemask_epi8(latin); 1237| 5.64k| answer += count_ones(non_ascii); 1238| 5.64k| } 1239| 865| } 1240| 5.83k| return answer + scalar::latin1::utf8_length_from_latin1( 1241| 5.83k| reinterpret_cast(data + i), len - i); 1242| 5.83k|} _ZNK7simdutf7haswell14implementation22utf8_length_from_utf32EPKDim: 1247| 4.96k| const char32_t *input, size_t length) const noexcept { 1248| 4.96k| return utf32::utf8_length_from_utf32(input, length); 1249| 4.96k|} _ZNK7simdutf7haswell14implementation22utf32_length_from_utf8EPKcm: 1275| 5.56k| const char *input, size_t length) const noexcept { 1276| 5.56k| return utf8::count_code_points(input, length); 1277| 5.56k|} _ZNK7simdutf7haswell14implementation16base64_to_binaryEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 1283| 15.9k| last_chunk_handling_options last_chunk_options) const noexcept { 1284| 15.9k| if (options & base64_default_or_url) { ------------------ | Branch (1284:7): [True: 0, False: 15.9k] ------------------ 1285| 0| if (options == base64_options::base64_default_or_url_accept_garbage) { ------------------ | Branch (1285:9): [True: 0, False: 0] ------------------ 1286| 0| return base64::compress_decode_base64( 1287| 0| output, input, length, options, last_chunk_options); 1288| 0| } else { 1289| 0| return base64::compress_decode_base64( 1290| 0| output, input, length, options, last_chunk_options); 1291| 0| } 1292| 15.9k| } else if (options & base64_url) { ------------------ | Branch (1292:14): [True: 0, False: 15.9k] ------------------ 1293| 0| if (options == base64_options::base64_url_accept_garbage) { ------------------ | Branch (1293:9): [True: 0, False: 0] ------------------ 1294| 0| return base64::compress_decode_base64( 1295| 0| output, input, length, options, last_chunk_options); 1296| 0| } else { 1297| 0| return base64::compress_decode_base64( 1298| 0| output, input, length, options, last_chunk_options); 1299| 0| } 1300| 15.9k| } else { 1301| 15.9k| if (options == base64_options::base64_default_accept_garbage) { ------------------ | Branch (1301:9): [True: 0, False: 15.9k] ------------------ 1302| 0| return base64::compress_decode_base64( 1303| 0| output, input, length, options, last_chunk_options); 1304| 15.9k| } else { 1305| 15.9k| return base64::compress_decode_base64( 1306| 15.9k| output, input, length, options, last_chunk_options); 1307| 15.9k| } 1308| 15.9k| } 1309| 15.9k|} _ZNK7simdutf7haswell14implementation24base64_to_binary_detailsEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 1313| 48.5k| last_chunk_handling_options last_chunk_options) const noexcept { 1314| 48.5k| if (options & base64_default_or_url) { ------------------ | Branch (1314:7): [True: 0, False: 48.5k] ------------------ 1315| 0| if (options == base64_options::base64_default_or_url_accept_garbage) { ------------------ | Branch (1315:9): [True: 0, False: 0] ------------------ 1316| 0| return base64::compress_decode_base64( 1317| 0| output, input, length, options, last_chunk_options); 1318| 0| } else { 1319| 0| return base64::compress_decode_base64( 1320| 0| output, input, length, options, last_chunk_options); 1321| 0| } 1322| 48.5k| } else if (options & base64_url) { ------------------ | Branch (1322:14): [True: 0, False: 48.5k] ------------------ 1323| 0| if (options == base64_options::base64_url_accept_garbage) { ------------------ | Branch (1323:9): [True: 0, False: 0] ------------------ 1324| 0| return base64::compress_decode_base64( 1325| 0| output, input, length, options, last_chunk_options); 1326| 0| } else { 1327| 0| return base64::compress_decode_base64( 1328| 0| output, input, length, options, last_chunk_options); 1329| 0| } 1330| 48.5k| } else { 1331| 48.5k| if (options == base64_options::base64_default_accept_garbage) { ------------------ | Branch (1331:9): [True: 0, False: 48.5k] ------------------ 1332| 0| return base64::compress_decode_base64( 1333| 0| output, input, length, options, last_chunk_options); 1334| 48.5k| } else { 1335| 48.5k| return base64::compress_decode_base64( 1336| 48.5k| output, input, length, options, last_chunk_options); 1337| 48.5k| } 1338| 48.5k| } 1339| 48.5k|} _ZNK7simdutf7haswell14implementation16binary_to_base64EPKcmPcNS_14base64_optionsE: 1403| 15.1k| base64_options options) const noexcept { 1404| 15.1k| if (options & base64_url) { ------------------ | Branch (1404:7): [True: 0, False: 15.1k] ------------------ 1405| 0| return encode_base64(output, input, length, options); 1406| 15.1k| } else { 1407| 15.1k| return encode_base64(output, input, length, options); 1408| 15.1k| } 1409| 15.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_18is_asciiERKNS1_4simd8simd8x64IhEE: 13| 34.5k|simdutf_really_inline bool is_ascii(const simd8x64 &input) { 14| 34.5k| return input.reduce_or().is_ascii(); 15| 34.5k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_124must_be_2_3_continuationENS1_4simd5simd8IhEES4_: 22| 26.1k| const simd8 prev3) { 23| 26.1k| simd8 is_third_byte = 24| 26.1k| prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be > 0x80 25| 26.1k| simd8 is_fourth_byte = 26| 26.1k| prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be > 0x80 27| 26.1k| return simd8(is_third_byte | is_fourth_byte); 28| 26.1k|} _ZNK7simdutf14implementation27supported_by_runtime_systemEv: 65| 22.2k|bool implementation::supported_by_runtime_system() const { 66| 22.2k| uint32_t required_instruction_sets = this->required_instruction_sets(); 67| 22.2k| uint32_t supported_instruction_sets = 68| 22.2k| internal::detect_supported_architectures(); 69| 22.2k| return ((supported_instruction_sets & required_instruction_sets) == 70| 22.2k| required_instruction_sets); 71| 22.2k|} _ZNK7simdutf14implementation33maximal_binary_length_from_base64EPKcm: 118| 80.5k| const char *input, size_t length) const noexcept { 119| 80.5k| return scalar::base64::maximal_binary_length_from_base64(input, length); 120| 80.5k|} _ZNK7simdutf14implementation25base64_length_from_binaryEmNS_14base64_optionsE: 138| 47.1k| size_t length, base64_options options) const noexcept { 139| 47.1k| return scalar::base64::base64_length_from_binary(length, options); 140| 47.1k|} _ZNK7simdutf8internal29available_implementation_list5beginEv: 1504| 5.56k|available_implementation_list::begin() const noexcept { 1505| 5.56k| return internal::get_available_implementation_pointers().begin(); 1506| 5.56k|} _ZNK7simdutf8internal29available_implementation_list3endEv: 1508| 5.56k|available_implementation_list::end() const noexcept { 1509| 5.56k| return internal::get_available_implementation_pointers().end(); 1510| 5.56k|} _ZNK7simdutf8internal29available_implementation_list21detect_best_supportedEv: 1512| 1|available_implementation_list::detect_best_supported() const noexcept { 1513| | // They are prelisted in priority order, so we just go down the list 1514| 1| uint32_t supported_instruction_sets = 1515| 1| internal::detect_supported_architectures(); 1516| 1| for (const implementation *impl : ------------------ | Branch (1516:35): [True: 2, False: 0] ------------------ 1517| 2| internal::get_available_implementation_pointers()) { 1518| 2| uint32_t required_instruction_sets = impl->required_instruction_sets(); 1519| 2| if ((supported_instruction_sets & required_instruction_sets) == ------------------ | Branch (1519:9): [True: 1, False: 1] ------------------ 1520| 2| required_instruction_sets) { 1521| 1| return impl; 1522| 1| } 1523| 2| } 1524| 0| return get_unsupported_singleton(); // this should never happen? 1525| 1|} _ZNK7simdutf8internal49detect_best_supported_implementation_on_first_use8set_bestEv: 1528| 1|detect_best_supported_implementation_on_first_use::set_best() const noexcept { 1529| 1| SIMDUTF_PUSH_DISABLE_WARNINGS 1530| | SIMDUTF_DISABLE_DEPRECATED_WARNING // Disable CRT_SECURE warning on MSVC: 1531| | // manually verified this is safe 1532| 1| char *force_implementation_name = getenv("SIMDUTF_FORCE_IMPLEMENTATION"); 1533| 1| SIMDUTF_POP_DISABLE_WARNINGS 1534| | 1535| 1| if (force_implementation_name) { ------------------ | Branch (1535:7): [True: 0, False: 1] ------------------ 1536| 0| auto force_implementation = 1537| 0| get_available_implementations()[force_implementation_name]; 1538| 0| if (force_implementation) { ------------------ | Branch (1538:9): [True: 0, False: 0] ------------------ 1539| 0| return get_active_implementation() = force_implementation; 1540| 0| } else { 1541| | // Note: abort() and stderr usage within the library is forbidden. 1542| 0| return get_active_implementation() = get_unsupported_singleton(); 1543| 0| } 1544| 0| } 1545| 1| return get_active_implementation() = 1546| 1| get_available_implementations().detect_best_supported(); 1547| 1|} _ZN7simdutf29get_available_implementationsEv: 1559| 5.56k|get_available_implementations() { 1560| 5.56k|#if !SIMDUTF_USE_STATIC_INITIALIZATION 1561| 5.56k| static const internal::available_implementation_list 1562| 5.56k| available_implementations_instance{}; 1563| 5.56k|#endif 1564| 5.56k| return available_implementations_instance; 1565| 5.56k|} _ZN7simdutf25get_active_implementationEv: 1587| 48.5k|get_active_implementation() { 1588| 48.5k|#if !SIMDUTF_USE_STATIC_INITIALIZATION 1589| 48.5k| #if !SIMDUTF_SINGLE_IMPLEMENTATION 1590| 48.5k| static const internal::detect_best_supported_implementation_on_first_use 1591| 48.5k| detect_best_supported_implementation_on_first_use_singleton; 1592| 48.5k| #endif 1593| 48.5k| static internal::atomic_ptr 1594| 48.5k| active_implementation_instance{ 1595| | #if SIMDUTF_SINGLE_IMPLEMENTATION 1596| | internal::get_single_implementation() 1597| | #else 1598| 48.5k| &detect_best_supported_implementation_on_first_use_singleton 1599| 48.5k| #endif 1600| 48.5k| }; 1601| 48.5k|#endif 1602| 48.5k| return active_implementation_instance; 1603| 48.5k|} _ZN7simdutf21base64_to_binary_safeEPKcmPcRmNS_14base64_optionsENS_27last_chunk_handling_optionsEb: 2600| 48.5k| bool decode_up_to_bad_char) noexcept { 2601| 48.5k| return base64_to_binary_safe_impl(input, length, output, outlen, 2602| 48.5k| options, last_chunk_handling_options, 2603| 48.5k| decode_up_to_bad_char); 2604| 48.5k|} simdutf.cpp:_ZN7simdutf8internalL37get_available_implementation_pointersEv: 934| 11.1k|get_available_implementation_pointers() { 935| 11.1k|#if !SIMDUTF_USE_STATIC_INITIALIZATION 936| 11.1k| static const std::initializer_list 937| 11.1k| available_implementation_pointers{ 938| 11.1k| #if SIMDUTF_IMPLEMENTATION_ICELAKE 939| 11.1k| get_icelake_singleton(), 940| 11.1k| #endif 941| 11.1k| #if SIMDUTF_IMPLEMENTATION_HASWELL 942| 11.1k| get_haswell_singleton(), 943| 11.1k| #endif 944| 11.1k| #if SIMDUTF_IMPLEMENTATION_WESTMERE 945| 11.1k| get_westmere_singleton(), 946| 11.1k| #endif 947| | #if SIMDUTF_IMPLEMENTATION_ARM64 948| | get_arm64_singleton(), 949| | #endif 950| | #if SIMDUTF_IMPLEMENTATION_PPC64 951| | get_ppc64_singleton(), 952| | #endif 953| | #if SIMDUTF_IMPLEMENTATION_RVV 954| | get_rvv_singleton(), 955| | #endif 956| | #if SIMDUTF_IMPLEMENTATION_LASX 957| | get_lasx_singleton(), 958| | #endif 959| | #if SIMDUTF_IMPLEMENTATION_LSX 960| | get_lsx_singleton(), 961| | #endif 962| 11.1k| #if SIMDUTF_IMPLEMENTATION_FALLBACK 963| 11.1k| get_fallback_singleton(), 964| 11.1k| #endif 965| 11.1k| }; 966| 11.1k|#endif 967| 11.1k| return available_implementation_pointers; 968| 11.1k|} simdutf.cpp:_ZN7simdutf8internalL21get_icelake_singletonEv: 158| 1|static const icelake::implementation *get_icelake_singleton() { 159| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 160| 1| static const icelake::implementation icelake_singleton{}; 161| 1| #endif 162| 1| return &icelake_singleton; 163| 1|} simdutf.cpp:_ZN7simdutf8internalL21get_haswell_singletonEv: 169| 1|static const haswell::implementation *get_haswell_singleton() { 170| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 171| 1| static const haswell::implementation haswell_singleton{}; 172| 1| #endif 173| 1| return &haswell_singleton; 174| 1|} simdutf.cpp:_ZN7simdutf8internalL22get_westmere_singletonEv: 180| 1|static const westmere::implementation *get_westmere_singleton() { 181| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 182| 1| static const westmere::implementation westmere_singleton{}; 183| 1| #endif 184| 1| return &westmere_singleton; 185| 1|} simdutf.cpp:_ZN7simdutf8internalL22get_fallback_singletonEv: 246| 1|static const fallback::implementation *get_fallback_singleton() { 247| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 248| 1| static const fallback::implementation fallback_singleton{}; 249| 1| #endif 250| 1| return &fallback_singleton; 251| 1|} _ZN7simdutf8internal49detect_best_supported_implementation_on_first_useC2Ev: 888| 1| : implementation("best_supported_detector", 889| 1| "Detects the best supported implementation and sets it", 890| 1| 0) {} _ZNK7simdutf8internal49detect_best_supported_implementation_on_first_use24base64_to_binary_detailsEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 829| 1| last_chunk_handling_options::loose) const noexcept override { 830| 1| return set_best()->base64_to_binary_details(input, length, output, options, 831| 1| last_chunk_handling_options); 832| 1| } _ZN7simdutf8fallback14implementationC2Ev: 16| 1| : simdutf::implementation("fallback", "Generic fallback implementation", 17| 1| 0) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_110count_onesEm: 14| 13.3k|simdutf_really_inline long long int count_ones(uint64_t input_num) { 15| | return _popcnt64(input_num); 16| 13.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115trailing_zeroesEm: 20| 1.56k|simdutf_really_inline int trailing_zeroes(uint64_t input_num) { 21| | #if SIMDUTF_REGULAR_VISUAL_STUDIO 22| | return (int)_tzcnt_u64(input_num); 23| | #else // SIMDUTF_REGULAR_VISUAL_STUDIO 24| 1.56k| return __builtin_ctzll(input_num); 25| 1.56k| #endif // SIMDUTF_REGULAR_VISUAL_STUDIO 26| 1.56k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115is_power_of_twoImEEbT_: 29| 2.32k|template bool is_power_of_two(T x) { return (x & (x - 1)) == 0; } _ZN7simdutf7haswell14implementationC2Ev: 16| 1| : simdutf::implementation("haswell", "Intel/AMD AVX2", 17| 1| internal::instruction_set::AVX2 | 18| 1| internal::instruction_set::BMI1 | 19| 1| internal::instruction_set::BMI2) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2Ev: 191| 82.9k| simdutf_really_inline simd8() : base8_numeric() {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhEC2Ev: 124| 82.9k| simdutf_really_inline base8_numeric() : base8() {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2Ev: 71| 82.9k| simdutf_really_inline base8() : base>() {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2Ev: 14| 82.9k| simdutf_really_inline base() : value{__m256i()} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IhE9reduce_orEv: 294| 34.5k| simdutf_really_inline simd8 reduce_or() const { 295| 34.5k| return this->chunks[0] | this->chunks[1]; 296| 34.5k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEorES5_: 50| 145k| simdutf_really_inline Child operator|(const Child other) const { 51| 145k| return _mm256_or_si256(*this, other); 52| 145k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2EDv4_x: 193| 1.27M| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhEC2EDv4_x: 126| 1.27M| : base8(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2EDv4_x: 73| 1.27M| simdutf_really_inline base8(const __m256i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2EDv4_x: 17| 1.27M| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE8is_asciiEv: 234| 34.5k| simdutf_really_inline bool is_ascii() const { 235| 34.5k| return _mm256_movemask_epi8(*this) == 0; 236| 34.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEoRES5_: 59| 84.6k| simdutf_really_inline Child &operator|=(const Child other) { 60| 84.6k| auto this_cast = static_cast(this); 61| 84.6k| *this_cast = *this_cast | other; 62| 84.6k| return *this_cast; 63| 84.6k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi1EEENS4_IhEES8_: 83| 30.5k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 84| | return _mm256_alignr_epi8( 85| 30.5k| *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N); 86| 30.5k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE3shrILi4EEES4_v: 245| 61.0k| template simdutf_really_inline simd8 shr() const { 246| 61.0k| return simd8(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); 247| 61.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE9lookup_16IhEENS2_5simd8IT_EES7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_: 157| 91.5k| L replace15) const { 158| 91.5k| return lookup_16(simd8::repeat_16( 159| 91.5k| replace0, replace1, replace2, replace3, replace4, replace5, replace6, 160| 91.5k| replace7, replace8, replace9, replace10, replace11, replace12, 161| 91.5k| replace13, replace14, replace15)); 162| 91.5k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE9lookup_16IhEENS2_5simd8IT_EES8_: 148| 91.5k| simdutf_really_inline simd8 lookup_16(simd8 lookup_table) const { 149| 91.5k| return _mm256_shuffle_epi8(lookup_table, *this); 150| 91.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE9repeat_16Ehhhhhhhhhhhhhhhh: 118| 91.5k| T v14, T v15) { 119| 91.5k| return simd8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, 120| 91.5k| v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, 121| 91.5k| v12, v13, v14, v15); 122| 91.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2Ehhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh: 207| 91.5k| : simd8(_mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, 208| 91.5k| v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, 209| 91.5k| v22, v23, v24, v25, v26, v27, v28, v29, v30, 210| 91.5k| v31)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2Eh: 195| 227k| simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi2EEENS4_IhEES8_: 83| 26.1k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 84| | return _mm256_alignr_epi8( 85| 26.1k| *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N); 86| 26.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi3EEENS4_IhEES8_: 83| 26.1k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 84| | return _mm256_alignr_epi8( 85| 26.1k| *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N); 86| 26.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE14saturating_subES4_: 214| 60.2k| saturating_sub(const simd8 other) const { 215| 60.2k| return _mm256_subs_epu8(*this, other); 216| 60.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IbEC2EDv4_x: 95| 100k| simdutf_really_inline simd8(const __m256i _value) : base8(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IbNS2_5simd8IbEEEC2EDv4_x: 73| 100k| simdutf_really_inline base8(const __m256i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEC2EDv4_x: 17| 100k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEcvRKDv4_xEv: 19| 56.2k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEeoES5_: 56| 26.1k| simdutf_really_inline Child operator^(const Child other) const { 57| 26.1k| return _mm256_xor_si256(*this, other); 58| 26.1k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2EPKh: 197| 7.82k| simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE4loadEPKh: 111| 77.0k| static simdutf_really_inline simd8 load(const T values[32]) { 112| 77.0k| return _mm256_loadu_si256(reinterpret_cast(values)); 113| 77.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE7gt_bitsES4_: 225| 7.82k| gt_bits(const simd8 other) const { 226| 7.82k| return this->saturating_sub(other); 227| 7.82k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEC2EDv4_x: 17| 311k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEcvRKDv4_xEv: 19| 329k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IbE10to_bitmaskEv: 99| 44.1k| simdutf_really_inline uint32_t to_bitmask() const { 100| 44.1k| return uint32_t(_mm256_movemask_epi8(value)); 101| 44.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE21any_bits_set_anywhereEv: 241| 67.2k| simdutf_really_inline bool any_bits_set_anywhere() const { 242| 67.2k| return !bits_not_set_anywhere(); 243| 67.2k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE21bits_not_set_anywhereEv: 237| 67.2k| simdutf_really_inline bool bits_not_set_anywhere() const { 238| 67.2k| return _mm256_testz_si256(*this, *this); 239| 67.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaEC2EPKa: 274| 12.5k| : chunks{simd8::load(ptr), 275| 12.5k| simd8::load(ptr + sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaE4loadEPKa: 111| 43.6k| static simdutf_really_inline simd8 load(const T values[32]) { 112| 43.6k| return _mm256_loadu_si256(reinterpret_cast(values)); 113| 43.6k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEC2EDv4_x: 169| 82.2k| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaEC2EDv4_x: 126| 82.2k| : base8(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IaNS2_5simd8IbEEEC2EDv4_x: 73| 82.2k| simdutf_really_inline base8(const __m256i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEC2EDv4_x: 17| 82.2k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE8is_asciiEv: 298| 10.8k| simdutf_really_inline bool is_ascii() const { 299| 10.8k| return this->reduce_or().is_ascii(); 300| 10.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE9reduce_orEv: 294| 10.8k| simdutf_really_inline simd8 reduce_or() const { 295| 10.8k| return this->chunks[0] | this->chunks[1]; 296| 10.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEorES5_: 50| 10.8k| simdutf_really_inline Child operator|(const Child other) const { 51| 10.8k| return _mm256_or_si256(*this, other); 52| 10.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEcvRKDv4_xEv: 19| 119k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaE8is_asciiEv: 177| 10.8k| simdutf_really_inline bool is_ascii() const { 178| 10.8k| return _mm256_movemask_epi8(*this) == 0; 179| 10.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE5storeEPa: 277| 1.66k| simdutf_really_inline void store(T *ptr) const { 278| 1.66k| this->chunks[0].store(ptr + sizeof(simd8) * 0 / sizeof(T)); 279| 1.66k| this->chunks[1].store(ptr + sizeof(simd8) * 1 / sizeof(T)); 280| 1.66k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaE5storeEPa: 129| 3.32k| simdutf_really_inline void store(T dst[32]) const { 130| 3.32k| return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); 131| 3.32k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEcvNS3_IhEEEv: 256| 22.2k|simdutf_really_inline simd8::operator simd8() const { 257| 22.2k| return this->value; 258| 22.2k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE2ltEa: 325| 7.42k| simdutf_really_inline uint64_t lt(const T m) const { 326| 7.42k| const simd8 mask = simd8::splat(m); 327| 7.42k| return simd8x64(this->chunks[0] < mask, this->chunks[1] < mask) 328| 7.42k| .to_bitmask(); 329| 7.42k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaE5splatEa: 105| 27.6k| static simdutf_really_inline simd8 splat(T _value) { 106| 27.6k| return _mm256_set1_epi8(_value); 107| 27.6k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEltES4_: 184| 14.8k| simdutf_really_inline simd8 operator<(const simd8 other) const { 185| 14.8k| return _mm256_cmpgt_epi8(other, *this); 186| 14.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IbEC2ENS2_5simd8IbEES6_: 272| 9.11k| : chunks{chunk0, chunk1} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IbE10to_bitmaskEv: 282| 9.11k| simdutf_really_inline uint64_t to_bitmask() const { 283| 9.11k| uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask()); 284| 9.11k| uint64_t r_hi = this->chunks[1].to_bitmask(); 285| 9.11k| return r_lo | (r_hi << 32); 286| 9.11k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf32EPDi: 310| 583| simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const { 311| 583| this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8) * 0); 312| 583| this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8) * 1); 313| 583| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf32EPDi: 36| 1.16k| simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const { 37| 1.16k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), 38| 1.16k| _mm256_cvtepu8_epi32(_mm256_castsi256_si128(*this))); 39| 1.16k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 8), 40| 1.16k| _mm256_cvtepu8_epi32(_mm256_castsi256_si128( 41| 1.16k| _mm256_srli_si256(*this, 8)))); 42| 1.16k| _mm256_storeu_si256( 43| 1.16k| reinterpret_cast<__m256i *>(ptr + 16), 44| 1.16k| _mm256_cvtepu8_epi32(_mm256_extractf128_si256(*this, 1))); 45| 1.16k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 24), 46| | _mm256_cvtepu8_epi32(_mm_srli_si128( 47| 1.16k| _mm256_extractf128_si256(*this, 1), 8))); 48| 1.16k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE4zeroEv: 108| 33.1k| static simdutf_really_inline simd8 zero() { 109| 33.1k| return _mm256_setzero_si256(); 110| 33.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEgtES4_: 181| 21.9k| simdutf_really_inline simd8 operator>(const simd8 other) const { 182| 21.9k| return _mm256_cmpgt_epi8(*this, other); 183| 21.9k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEC2Ea: 172| 18.5k| simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhEmIENS2_5simd8IhEE: 137| 30.0k| simdutf_really_inline simd8 &operator-=(const simd8 other) { 138| 30.0k| *this = *this - other; 139| 30.0k| return *static_cast *>(this); 140| 30.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhEmiENS2_5simd8IhEE: 134| 30.0k| simdutf_really_inline simd8 operator-(const simd8 other) const { 135| 30.0k| return _mm256_sub_epi8(*this, other); 136| 30.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE9sum_bytesEv: 249| 2.17k| simdutf_really_inline uint64_t sum_bytes() const { 250| 2.17k| const auto tmp = _mm256_sad_epu8(value, _mm256_setzero_si256()); 251| | 252| 2.17k| return _mm256_extract_epi64(tmp, 0) + _mm256_extract_epi64(tmp, 1) + 253| 2.17k| _mm256_extract_epi64(tmp, 2) + _mm256_extract_epi64(tmp, 3); 254| 2.17k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEgeES4_: 229| 11.4k| operator>=(const simd8 other) const { 230| 11.4k| return other.min_val(*this) == other; 231| 11.4k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE7min_valES4_: 220| 11.4k| min_val(const simd8 other) const { 221| 11.4k| return _mm256_min_epu8(other, *this); 222| 11.4k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE2gtEa: 331| 1.69k| simdutf_really_inline uint64_t gt(const T m) const { 332| 1.69k| const simd8 mask = simd8::splat(m); 333| 1.69k| return simd8x64(this->chunks[0] > mask, this->chunks[1] > mask) 334| 1.69k| .to_bitmask(); 335| 1.69k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEC2EDv4_x: 17| 11.1k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEcvRKDv4_xEv: 19| 11.1k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE5splatEh: 105| 249k| static simdutf_really_inline simd8 splat(T _value) { 106| 249k| return _mm256_set1_epi8(_value); 107| 249k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEcvRKDv4_xEv: 19| 1.71M| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdeqENS2_5simd8IhEES4_: 76| 37.3k| const simd8 rhs) { 77| 37.3k| return _mm256_cmpeq_epi8(lhs, rhs); 78| 37.3k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEanES5_: 53| 204k| simdutf_really_inline Child operator&(const Child other) const { 54| 204k| return _mm256_and_si256(*this, other); 55| 204k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IhEC2EPKh: 274| 34.5k| : chunks{simd8::load(ptr), 275| 34.5k| simd8::load(ptr + sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEanES5_: 53| 57.8k| simdutf_really_inline Child operator&(const Child other) const { 54| 57.8k| return _mm256_and_si256(*this, other); 55| 57.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 303| 599| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 304| 599| this->chunks[0].template store_ascii_as_utf16(ptr + 305| 599| sizeof(simd8) * 0); 306| 599| this->chunks[1].template store_ascii_as_utf16(ptr + 307| 599| sizeof(simd8) * 1); 308| 599| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 22| 1.19k| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 23| 1.19k| __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this)); 24| 1.19k| __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this, 1)); 25| 1.19k| if (big_endian) { ------------------ | Branch (25:9): [Folded, False: 1.19k] ------------------ 26| 0| const __m256i swap = _mm256_setr_epi8( 27| 0| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 28| 0| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 29| 0| first = _mm256_shuffle_epi8(first, swap); 30| 0| second = _mm256_shuffle_epi8(second, swap); 31| 0| } 32| 1.19k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first); 33| 1.19k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second); 34| 1.19k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 303| 598| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 304| 598| this->chunks[0].template store_ascii_as_utf16(ptr + 305| 598| sizeof(simd8) * 0); 306| 598| this->chunks[1].template store_ascii_as_utf16(ptr + 307| 598| sizeof(simd8) * 1); 308| 598| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 22| 1.19k| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 23| 1.19k| __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this)); 24| 1.19k| __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this, 1)); 25| 1.19k| if (big_endian) { ------------------ | Branch (25:9): [True: 1.19k, Folded] ------------------ 26| 1.19k| const __m256i swap = _mm256_setr_epi8( 27| 1.19k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 28| 1.19k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 29| 1.19k| first = _mm256_shuffle_epi8(first, swap); 30| 1.19k| second = _mm256_shuffle_epi8(second, swap); 31| 1.19k| } 32| 1.19k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first); 33| 1.19k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second); 34| 1.19k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItEC2EDv4_x: 97| 311k| : base16_numeric(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItEC2EDv4_x: 73| 311k| : base16(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6base16ItNS2_6simd16IbEEEC2EDv4_x: 20| 311k| : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE4packERKS4_S6_: 149| 23.8k| const simd16 &v1) { 150| | // Note: the AVX2 variant of pack operates on 128-bit lanes, thus 151| | // we have to shuffle lanes in order to produce bytes in the 152| | // correct order. 153| | 154| | // get the 0th lanes 155| 23.8k| const __m128i lo_0 = _mm256_extracti128_si256(v0, 0); 156| 23.8k| const __m128i lo_1 = _mm256_extracti128_si256(v1, 0); 157| | 158| | // get the 1st lanes 159| 23.8k| const __m128i hi_0 = _mm256_extracti128_si256(v0, 1); 160| 23.8k| const __m128i hi_1 = _mm256_extracti128_si256(v1, 1); 161| | 162| | // build new vectors (shuffle lanes) 163| 23.8k| const __m256i t0 = _mm256_set_m128i(lo_1, lo_0); ------------------ | | 4| 23.8k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 23.8k| _mm256_castsi128_si256(xmm2), 2) ------------------ 164| 23.8k| const __m256i t1 = _mm256_set_m128i(hi_1, hi_0); ------------------ | | 4| 23.8k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 23.8k| _mm256_castsi128_si256(xmm2), 2) ------------------ 165| | 166| | // pack code units in linear order from v0 and v1 167| 23.8k| return _mm256_packus_epi16(t0, t1); 168| 23.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItE4loadEPKt: 66| 58.8k| static simdutf_really_inline simd16 load(const T values[8]) { 67| 58.8k| return _mm256_loadu_si256(reinterpret_cast(values)); 68| 58.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE10swap_bytesEv: 139| 5.77k| simdutf_really_inline simd16 swap_bytes() const { 140| 5.77k| const __m256i swap = _mm256_setr_epi8( 141| 5.77k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 142| 5.77k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 143| 5.77k| return _mm256_shuffle_epi8(*this, swap); 144| 5.77k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItE5splatEt: 58| 77.9k| static simdutf_really_inline simd16 splat(T _value) { 59| 77.9k| return _mm256_set1_epi16(_value); 60| 77.9k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdeqENS2_6simd16ItEES4_: 26| 11.1k| const simd16 rhs) { 27| 11.1k| return _mm256_cmpeq_epi16(lhs, rhs); 28| 11.1k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16IbEC2EDv4_x: 45| 11.1k| simdutf_really_inline simd16(const __m256i _value) : base16(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6base16IbNS2_6simd16IbEEEC2EDv4_x: 20| 11.1k| : base>(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE3shrILi8EEES4_v: 134| 23.2k| template simdutf_really_inline simd16 shr() const { 135| 23.2k| return simd16(_mm256_srli_epi16(*this, N)); 136| 23.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItEC2EPKDs: 104| 47.6k| : simd16(load(reinterpret_cast(values))) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItEC2Et: 100| 56.8k| simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItE4zeroEv: 62| 21.0k| static simdutf_really_inline simd16 zero() { 63| 21.0k| return _mm256_setzero_si256(); 64| 21.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd3minENS2_6simd16ItEES4_: 259| 22.3k|simd16 min(const simd16 a, simd16 b) { 260| 22.3k| return _mm256_min_epu16(a.value, b.value); 261| 22.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItEpLENS2_6simd16ItEE: 87| 33.5k| simdutf_really_inline simd16 &operator+=(const simd16 other) { 88| 33.5k| *this = *this + other; 89| 33.5k| return *static_cast *>(this); 90| 33.5k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItEplENS2_6simd16ItEE: 84| 33.5k| simdutf_really_inline simd16 operator+(const simd16 other) const { 85| 33.5k| return _mm256_add_epi16(*this, other); 86| 33.5k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE3sumEv: 170| 21.0k| simdutf_really_inline uint64_t sum() const { 171| 21.0k| const auto lo_u16 = _mm256_and_si256(value, _mm256_set1_epi32(0x0000ffff)); 172| 21.0k| const auto hi_u16 = _mm256_srli_epi32(value, 16); 173| 21.0k| const auto sum_u32 = _mm256_add_epi32(lo_u16, hi_u16); 174| | 175| 21.0k| const auto lo_u32 = 176| 21.0k| _mm256_and_si256(sum_u32, _mm256_set1_epi64x(0xffffffff)); 177| 21.0k| const auto hi_u32 = _mm256_srli_epi64(sum_u32, 32); 178| 21.0k| const auto sum_u64 = _mm256_add_epi64(lo_u32, hi_u32); 179| | 180| 21.0k| return uint64_t(_mm256_extract_epi64(sum_u64, 0)) + 181| 21.0k| uint64_t(_mm256_extract_epi64(sum_u64, 1)) + 182| 21.0k| uint64_t(_mm256_extract_epi64(sum_u64, 2)) + 183| | uint64_t(_mm256_extract_epi64(sum_u64, 3)); 184| 21.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE5splatEj: 46| 25.9k| simdutf_really_inline static simd32 splat(uint32_t v) { 47| 25.9k| return _mm256_set1_epi32(v); 48| 25.9k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjEC2EDv4_x: 9| 178k| simdutf_really_inline simd32(const __m256i v) : value(v) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE4zeroEv: 42| 4.77k| simdutf_really_inline static simd32 zero() { 43| 4.77k| return _mm256_setzero_si256(); 44| 4.77k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjEC2IDiEEPKT_: 13| 33.1k| : value(_mm256_loadu_si256(reinterpret_cast(ptr))) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE10swap_bytesEv: 27| 25.8k| simdutf_really_inline simd32 swap_bytes() const { 28| 25.8k| const __m256i shuffle = 29| 25.8k| _mm256_setr_epi8(3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12, 30| 25.8k| 3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12); 31| | 32| 25.8k| return _mm256_shuffle_epi8(value, shuffle); 33| 25.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd3maxENS2_6simd32IjEES4_: 80| 51.6k| const simd32 b) { 81| 51.6k| return _mm256_max_epu32(a.value, b.value); 82| 51.6k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdplENS2_6simd32IjEES4_: 90| 25.8k| const simd32 b) { 91| 25.8k| return _mm256_add_epi32(a.value, b.value); 92| 25.8k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdgtENS2_6simd32IjEES4_: 109| 4.06k| const simd32 b) { 110| 4.06k| return !(b >= a); 111| 4.06k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdntENS2_6simd32IbEE: 104| 4.06k|simdutf_really_inline simd32 operator!(const simd32 v) { 105| 4.06k| return _mm256_xor_si256(v.value, _mm256_set1_epi8(-1)); 106| 4.06k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IbEC2EDv4_x: 59| 8.12k| simdutf_really_inline simd32(const __m256i v) : value(v) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdgeENS2_6simd32IjEES4_: 100| 4.06k| const simd32 b) { 101| 4.06k| return _mm256_cmpeq_epi32(_mm256_max_epu32(a.value, b.value), a.value); 102| 4.06k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd32IbE3anyEv: 61| 4.06k| simdutf_really_inline bool any() const { 62| 4.06k| return _mm256_movemask_epi8(value) != 0; 63| 4.06k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdanENS2_6simd32IjEES4_: 85| 22.0k| const simd32 a) { 86| 22.0k| return _mm256_and_si256(a.value, b.value); 87| 22.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd3minENS2_6simd32IjEES4_: 75| 22.0k| const simd32 a) { 76| 22.0k| return _mm256_min_epu32(a.value, b.value); 77| 22.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjEpLES4_: 36| 22.0k| simdutf_really_inline simd32 &operator+=(const simd32 other) { 37| 22.0k| value = _mm256_add_epi32(value, other.value); 38| 22.0k| return *this; 39| 22.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE3sumEv: 15| 712| simdutf_really_inline uint64_t sum() const { 16| 712| const __m256i mask = _mm256_set1_epi64x(0xffffffff); 17| 712| const __m256i t0 = _mm256_and_si256(value, mask); 18| 712| const __m256i t1 = _mm256_srli_epi64(value, 32); 19| 712| const __m256i t2 = _mm256_add_epi64(t0, t1); 20| | 21| 712| return uint64_t(_mm256_extract_epi64(t2, 0)) + 22| 712| uint64_t(_mm256_extract_epi64(t2, 1)) + 23| 712| uint64_t(_mm256_extract_epi64(t2, 2)) + 24| | uint64_t(_mm256_extract_epi64(t2, 3)); 25| 712| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd64ImE4zeroEv: 27| 33.1k| simdutf_really_inline static simd64 zero() { 28| 33.1k| return _mm256_setzero_si256(); 29| 33.1k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd64ImEC2EDv4_x: 9| 33.1k| simdutf_really_inline simd64(const __m256i v) : value(v) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd64ImE3sumEv: 15| 33.1k| simdutf_really_inline uint64_t sum() const { 16| 33.1k| return _mm256_extract_epi64(value, 0) + _mm256_extract_epi64(value, 1) + 17| 33.1k| _mm256_extract_epi64(value, 2) + _mm256_extract_epi64(value, 3); 18| 33.1k| } _ZN7simdutf7icelake14implementationC2Ev: 16| 1| : simdutf::implementation( 17| 1| "icelake", 18| 1| "Intel AVX512 (AVX-512BW, AVX-512CD, AVX-512VL, AVX-512VBMI2 " 19| 1| "extensions)", 20| 1| internal::instruction_set::AVX2 | internal::instruction_set::BMI1 | 21| 1| internal::instruction_set::BMI2 | 22| 1| internal::instruction_set::AVX512BW | 23| 1| internal::instruction_set::AVX512CD | 24| 1| internal::instruction_set::AVX512VL | 25| 1| internal::instruction_set::AVX512VBMI2 | 26| 1| internal::instruction_set::AVX512VPOPCNTDQ) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_110count_onesEm: 14| 18.3k|simdutf_really_inline long long int count_ones(uint64_t input_num) { 15| | return _popcnt64(input_num); 16| 18.3k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115trailing_zeroesEm: 20| 466|simdutf_really_inline int trailing_zeroes(uint64_t input_num) { 21| | #if SIMDUTF_REGULAR_VISUAL_STUDIO 22| | unsigned long ret; 23| | _BitScanForward64(&ret, input_num); 24| | return (int)ret; 25| | #else // SIMDUTF_REGULAR_VISUAL_STUDIO 26| 466| return __builtin_ctzll(input_num); 27| 466| #endif // SIMDUTF_REGULAR_VISUAL_STUDIO 28| 466|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115is_power_of_twoImEEbT_: 31| 752|template bool is_power_of_two(T x) { return (x & (x - 1)) == 0; } _ZN7simdutf8westmere14implementationC2Ev: 18| 1| : simdutf::implementation("westmere", "Intel/AMD SSE4.2", 19| 1| internal::instruction_set::SSE42) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2Ev: 183| 82.8k| simdutf_really_inline simd8() : base8_numeric() {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEC2Ev: 118| 82.8k| simdutf_really_inline base8_numeric() : base8() {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2Ev: 68| 82.8k| simdutf_really_inline base8() : base>() {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2Ev: 13| 82.8k| simdutf_really_inline base() : value{__m128i()} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IhE9reduce_orEv: 290| 37.2k| simdutf_really_inline simd8 reduce_or() const { 291| 37.2k| return (this->chunks[0] | this->chunks[1]) | 292| 37.2k| (this->chunks[2] | this->chunks[3]); 293| 37.2k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEorES5_: 42| 317k| simdutf_really_inline Child operator|(const Child other) const { 43| 317k| return _mm_or_si128(*this, other); 44| 317k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2EDv2_x: 185| 3.10M| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEC2EDv2_x: 120| 3.10M| : base8(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2EDv2_x: 69| 3.10M| simdutf_really_inline base8(const __m128i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2EDv2_x: 16| 3.10M| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE8is_asciiEv: 228| 37.2k| simdutf_really_inline bool is_ascii() const { 229| 37.2k| return _mm_movemask_epi8(*this) == 0; 230| 37.2k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEoRES5_: 51| 137k| simdutf_really_inline Child &operator|=(const Child other) { 52| 137k| auto this_cast = static_cast(this); 53| 137k| *this_cast = *this_cast | other; 54| 137k| return *this_cast; 55| 137k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi1EEENS4_IhEES8_: 79| 81.6k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 80| | return _mm_alignr_epi8(*this, prev_chunk, 16 - N); 81| 81.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE3shrILi4EEES4_v: 238| 163k| template simdutf_really_inline simd8 shr() const { 239| 163k| return simd8(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); 240| 163k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE9lookup_16IhEENS2_5simd8IT_EES7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_S7_: 151| 245k| L replace15) const { 152| 245k| return lookup_16(simd8::repeat_16( 153| 245k| replace0, replace1, replace2, replace3, replace4, replace5, replace6, 154| 245k| replace7, replace8, replace9, replace10, replace11, replace12, 155| 245k| replace13, replace14, replace15)); 156| 245k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE9lookup_16IhEENS2_5simd8IT_EES8_: 142| 245k| simdutf_really_inline simd8 lookup_16(simd8 lookup_table) const { 143| 245k| return _mm_shuffle_epi8(lookup_table, *this); 144| 245k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE9repeat_16Ehhhhhhhhhhhhhhhh: 113| 245k| T v14, T v15) { 114| 245k| return simd8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, 115| 245k| v14, v15); 116| 245k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2Ehhhhhhhhhhhhhhhh: 196| 245k| : simd8(_mm_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, 197| 245k| v12, v13, v14, v15)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2Eh: 188| 536k| simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi2EEENS4_IhEES8_: 79| 68.7k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 80| | return _mm_alignr_epi8(*this, prev_chunk, 16 - N); 81| 68.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi3EEENS4_IhEES8_: 79| 68.7k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 80| | return _mm_alignr_epi8(*this, prev_chunk, 16 - N); 81| 68.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE14saturating_subES4_: 201| 146k| saturating_sub(const simd8 other) const { 202| 146k| return _mm_subs_epu8(*this, other); 203| 146k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IbEC2EDv2_x: 91| 256k| simdutf_really_inline simd8(const __m128i _value) : base8(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IbNS2_5simd8IbEEEC2EDv2_x: 69| 256k| simdutf_really_inline base8(const __m128i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEC2EDv2_x: 16| 256k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEcvRKDv2_xEv: 18| 256k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEeoES5_: 48| 68.7k| simdutf_really_inline Child operator^(const Child other) const { 49| 68.7k| return _mm_xor_si128(*this, other); 50| 68.7k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2EPKh: 190| 9.01k| simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE4loadEPKh: 106| 158k| static simdutf_really_inline simd8 load(const T values[16]) { 107| 158k| return _mm_loadu_si128(reinterpret_cast(values)); 108| 158k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE7gt_bitsES4_: 212| 9.01k| gt_bits(const simd8 other) const { 213| 9.01k| return this->saturating_sub(other); 214| 9.01k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEC2EDv2_x: 16| 693k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEcvRKDv2_xEv: 18| 669k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IbE10to_bitmaskEv: 95| 95.6k| simdutf_really_inline int to_bitmask() const { 96| 95.6k| return _mm_movemask_epi8(*this); 97| 95.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE21any_bits_set_anywhereEv: 235| 67.7k| simdutf_really_inline bool any_bits_set_anywhere() const { 236| 67.7k| return !bits_not_set_anywhere(); 237| 67.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE21bits_not_set_anywhereEv: 232| 67.7k| simdutf_really_inline bool bits_not_set_anywhere() const { 233| 67.7k| return _mm_testz_si128(*this, *this); 234| 67.7k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaEC2EPKa: 270| 19.0k| : chunks{simd8::load(ptr), 271| 19.0k| simd8::load(ptr + sizeof(simd8) / sizeof(T)), 272| 19.0k| simd8::load(ptr + 2 * sizeof(simd8) / sizeof(T)), 273| 19.0k| simd8::load(ptr + 3 * sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaE4loadEPKa: 106| 133k| static simdutf_really_inline simd8 load(const T values[16]) { 107| 133k| return _mm_loadu_si128(reinterpret_cast(values)); 108| 133k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEC2EDv2_x: 163| 254k| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaEC2EDv2_x: 120| 254k| : base8(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IaNS2_5simd8IbEEEC2EDv2_x: 69| 254k| simdutf_really_inline base8(const __m128i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEC2EDv2_x: 16| 254k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE8is_asciiEv: 295| 16.3k| simdutf_really_inline bool is_ascii() const { 296| 16.3k| return this->reduce_or().is_ascii(); 297| 16.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE9reduce_orEv: 290| 16.3k| simdutf_really_inline simd8 reduce_or() const { 291| 16.3k| return (this->chunks[0] | this->chunks[1]) | 292| 16.3k| (this->chunks[2] | this->chunks[3]); 293| 16.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEorES5_: 42| 49.1k| simdutf_really_inline Child operator|(const Child other) const { 43| 49.1k| return _mm_or_si128(*this, other); 44| 49.1k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEcvRKDv2_xEv: 18| 381k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaE8is_asciiEv: 168| 16.3k| simdutf_really_inline bool is_ascii() const { 169| 16.3k| return _mm_movemask_epi8(*this) == 0; 170| 16.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE5storeEPa: 275| 2.08k| simdutf_really_inline void store(T *ptr) const { 276| 2.08k| this->chunks[0].store(ptr + sizeof(simd8) * 0 / sizeof(T)); 277| 2.08k| this->chunks[1].store(ptr + sizeof(simd8) * 1 / sizeof(T)); 278| 2.08k| this->chunks[2].store(ptr + sizeof(simd8) * 2 / sizeof(T)); 279| 2.08k| this->chunks[3].store(ptr + sizeof(simd8) * 3 / sizeof(T)); 280| 2.08k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaE5storeEPa: 123| 8.33k| simdutf_really_inline void store(T dst[16]) const { 124| 8.33k| return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); 125| 8.33k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEcvNS3_IhEEEv: 251| 79.8k|simdutf_really_inline simd8::operator simd8() const { 252| 79.8k| return this->value; 253| 79.8k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE2ltEa: 326| 11.4k| simdutf_really_inline uint64_t lt(const T m) const { 327| 11.4k| const simd8 mask = simd8::splat(m); 328| 11.4k| return simd8x64(this->chunks[0] < mask, this->chunks[1] < mask, 329| 11.4k| this->chunks[2] < mask, this->chunks[3] < mask) 330| 11.4k| .to_bitmask(); 331| 11.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaE5splatEa: 102| 71.5k| static simdutf_really_inline simd8 splat(T _value) { 103| 71.5k| return _mm_set1_epi8(_value); 104| 71.5k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEltES4_: 176| 45.6k| simdutf_really_inline simd8 operator<(const simd8 other) const { 177| 45.6k| return _mm_cmpgt_epi8(other, *this); 178| 45.6k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IbEC2ENS2_5simd8IbEES6_S6_S6_: 268| 14.0k| : chunks{chunk0, chunk1, chunk2, chunk3} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IbE10to_bitmaskEv: 318| 14.0k| simdutf_really_inline uint64_t to_bitmask() const { 319| 14.0k| uint64_t r0 = uint32_t(this->chunks[0].to_bitmask()); 320| 14.0k| uint64_t r1 = this->chunks[1].to_bitmask(); 321| 14.0k| uint64_t r2 = this->chunks[2].to_bitmask(); 322| 14.0k| uint64_t r3 = this->chunks[3].to_bitmask(); 323| 14.0k| return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48); 324| 14.0k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf32EPDi: 311| 944| simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const { 312| 944| this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8) * 0); 313| 944| this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8) * 1); 314| 944| this->chunks[2].store_ascii_as_utf32(ptr + sizeof(simd8) * 2); 315| 944| this->chunks[3].store_ascii_as_utf32(ptr + sizeof(simd8) * 3); 316| 944| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf32EPDi: 32| 3.77k| simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const { 33| 3.77k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p), _mm_cvtepu8_epi32(*this)); 34| 3.77k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 4), 35| 3.77k| _mm_cvtepu8_epi32(_mm_srli_si128(*this, 4))); 36| 3.77k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), 37| 3.77k| _mm_cvtepu8_epi32(_mm_srli_si128(*this, 8))); 38| 3.77k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 12), 39| | _mm_cvtepu8_epi32(_mm_srli_si128(*this, 12))); 40| 3.77k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE4zeroEv: 105| 33.1k| static simdutf_really_inline simd8 zero() { return _mm_setzero_si128(); } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEgtES4_: 173| 68.0k| simdutf_really_inline simd8 operator>(const simd8 other) const { 174| 68.0k| return _mm_cmpgt_epi8(*this, other); 175| 68.0k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEC2Ea: 165| 57.4k| simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEmIENS2_5simd8IhEE: 134| 92.2k| simdutf_really_inline simd8 &operator-=(const simd8 other) { 135| 92.2k| *this = *this - other; 136| 92.2k| return *static_cast *>(this); 137| 92.2k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEmiENS2_5simd8IhEE: 131| 92.2k| simdutf_really_inline simd8 operator-(const simd8 other) const { 132| 92.2k| return _mm_sub_epi8(*this, other); 133| 92.2k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE9sum_bytesEv: 245| 3.05k| simdutf_really_inline uint64_t sum_bytes() const { 246| 3.05k| const auto tmp = _mm_sad_epu8(value, _mm_setzero_si128()); 247| 3.05k| return _mm_extract_epi64(tmp, 0) + _mm_extract_epi64(tmp, 1); 248| 3.05k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEgeES4_: 217| 34.7k| operator>=(const simd8 other) const { 218| 34.7k| return other.min_val(*this) == other; 219| 34.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE7min_valES4_: 207| 34.7k| min_val(const simd8 other) const { 208| 34.7k| return _mm_min_epu8(*this, other); 209| 34.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE2gtEa: 333| 2.63k| simdutf_really_inline uint64_t gt(const T m) const { 334| 2.63k| const simd8 mask = simd8::splat(m); 335| 2.63k| return simd8x64(this->chunks[0] > mask, this->chunks[1] > mask, 336| 2.63k| this->chunks[2] > mask, this->chunks[3] > mask) 337| 2.63k| .to_bitmask(); 338| 2.63k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEC2EDv2_x: 16| 33.3k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEcvRKDv2_xEv: 18| 33.3k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE5splatEh: 102| 553k| static simdutf_really_inline simd8 splat(T _value) { 103| 553k| return _mm_set1_epi8(_value); 104| 553k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEcvRKDv2_xEv: 18| 3.83M| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdeqENS2_5simd8IhEES4_: 72| 74.1k| const simd8 rhs) { 73| 74.1k| return _mm_cmpeq_epi8(lhs, rhs); 74| 74.1k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEanES5_: 45| 516k| simdutf_really_inline Child operator&(const Child other) const { 46| 516k| return _mm_and_si128(*this, other); 47| 516k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IhEC2EPKh: 270| 37.2k| : chunks{simd8::load(ptr), 271| 37.2k| simd8::load(ptr + sizeof(simd8) / sizeof(T)), 272| 37.2k| simd8::load(ptr + 2 * sizeof(simd8) / sizeof(T)), 273| 37.2k| simd8::load(ptr + 3 * sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEanES5_: 45| 137k| simdutf_really_inline Child operator&(const Child other) const { 46| 137k| return _mm_and_si128(*this, other); 47| 137k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 300| 974| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 301| 974| this->chunks[0].template store_ascii_as_utf16(ptr + 302| 974| sizeof(simd8) * 0); 303| 974| this->chunks[1].template store_ascii_as_utf16(ptr + 304| 974| sizeof(simd8) * 1); 305| 974| this->chunks[2].template store_ascii_as_utf16(ptr + 306| 974| sizeof(simd8) * 2); 307| 974| this->chunks[3].template store_ascii_as_utf16(ptr + 308| 974| sizeof(simd8) * 3); 309| 974| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 20| 3.89k| simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const { 21| 3.89k| __m128i first = _mm_cvtepu8_epi16(*this); 22| 3.89k| __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this, 8)); 23| 3.89k| if (big_endian) { ------------------ | Branch (23:9): [Folded, False: 3.89k] ------------------ 24| 0| const __m128i swap = 25| 0| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 26| 0| first = _mm_shuffle_epi8(first, swap); 27| 0| second = _mm_shuffle_epi8(second, swap); 28| 0| } 29| 3.89k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first); 30| 3.89k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), second); 31| 3.89k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 300| 973| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 301| 973| this->chunks[0].template store_ascii_as_utf16(ptr + 302| 973| sizeof(simd8) * 0); 303| 973| this->chunks[1].template store_ascii_as_utf16(ptr + 304| 973| sizeof(simd8) * 1); 305| 973| this->chunks[2].template store_ascii_as_utf16(ptr + 306| 973| sizeof(simd8) * 2); 307| 973| this->chunks[3].template store_ascii_as_utf16(ptr + 308| 973| sizeof(simd8) * 3); 309| 973| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 20| 3.89k| simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const { 21| 3.89k| __m128i first = _mm_cvtepu8_epi16(*this); 22| 3.89k| __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this, 8)); 23| 3.89k| if (big_endian) { ------------------ | Branch (23:9): [True: 3.89k, Folded] ------------------ 24| 3.89k| const __m128i swap = 25| 3.89k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 26| 3.89k| first = _mm_shuffle_epi8(first, swap); 27| 3.89k| second = _mm_shuffle_epi8(second, swap); 28| 3.89k| } 29| 3.89k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first); 30| 3.89k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), second); 31| 3.89k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItEC2EDv2_x: 79| 693k| : base16_numeric(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItEC2EDv2_x: 54| 693k| : base16(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6base16ItNS2_6simd16IbEEEC2EDv2_x: 8| 693k| : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE4packERKS4_S6_: 130| 36.9k| const simd16 &v1) { 131| 36.9k| return _mm_packus_epi16(v0, v1); 132| 36.9k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItE4loadEPKt: 47| 107k| static simdutf_really_inline simd16 load(const T values[8]) { 48| 107k| return _mm_loadu_si128(reinterpret_cast(values)); 49| 107k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE10swap_bytesEv: 121| 16.7k| simdutf_really_inline simd16 swap_bytes() const { 122| 16.7k| const __m128i swap = 123| 16.7k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 124| 16.7k| return _mm_shuffle_epi8(*this, swap); 125| 16.7k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItE5splatEt: 41| 173k| static simdutf_really_inline simd16 splat(T _value) { 42| 173k| return _mm_set1_epi16(_value); 43| 173k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdeqENS2_6simd16ItEES4_: 11| 33.3k| const simd16 rhs) { 12| 33.3k| return _mm_cmpeq_epi16(lhs, rhs); 13| 33.3k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16IbEC2EDv2_x: 28| 33.3k| simdutf_really_inline simd16(const __m128i _value) : base16(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6base16IbNS2_6simd16IbEEEC2EDv2_x: 8| 33.3k| : base>(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE3shrILi8EEES4_v: 116| 36.8k| template simdutf_really_inline simd16 shr() const { 117| 36.8k| return simd16(_mm_srli_epi16(*this, N)); 118| 36.8k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItEC2EPKDs: 86| 73.9k| : simd16(load(reinterpret_cast(values))) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItEC2Et: 82| 152k| simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItE4zeroEv: 45| 21.4k| static simdutf_really_inline simd16 zero() { return _mm_setzero_si128(); } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd3minENS2_6simd16ItEES4_: 240| 66.7k|simd16 min(const simd16 a, simd16 b) { 241| 66.7k| return _mm_min_epu16(a.value, b.value); 242| 66.7k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItEpLENS2_6simd16ItEE: 68| 100k| simdutf_really_inline simd16 &operator+=(const simd16 other) { 69| 100k| *this = *this + other; 70| 100k| return *static_cast *>(this); 71| 100k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItEplENS2_6simd16ItEE: 65| 100k| simdutf_really_inline simd16 operator+(const simd16 other) const { 66| 100k| return _mm_add_epi16(*this, other); 67| 100k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE3sumEv: 134| 21.4k| simdutf_really_inline uint64_t sum() const { 135| 21.4k| const auto lo_u16 = _mm_and_si128(value, _mm_set1_epi32(0x0000ffff)); 136| 21.4k| const auto hi_u16 = _mm_srli_epi32(value, 16); 137| 21.4k| const auto sum_u32 = _mm_add_epi32(lo_u16, hi_u16); 138| | 139| 21.4k| const auto lo_u32 = _mm_and_si128(sum_u32, _mm_set1_epi64x(0xffffffff)); 140| 21.4k| const auto hi_u32 = _mm_srli_epi64(sum_u32, 32); 141| 21.4k| const auto sum_u64 = _mm_add_epi64(lo_u32, hi_u32); 142| | 143| 21.4k| return uint64_t(_mm_extract_epi64(sum_u64, 0)) + 144| | uint64_t(_mm_extract_epi64(sum_u64, 1)); 145| 21.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE5splatEj: 56| 25.1k| simdutf_really_inline static simd32 splat(uint32_t v) { 57| 25.1k| return _mm_set1_epi32(v); 58| 25.1k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjEC2EDv2_x: 9| 409k| simdutf_really_inline simd32(const __m128i v) : value(v) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE4zeroEv: 52| 4.00k| simdutf_really_inline static simd32 zero() { 53| 4.00k| return _mm_setzero_si128(); 54| 4.00k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjEC2IDiEEPKT_: 13| 84.2k| : value(_mm_loadu_si128(reinterpret_cast(ptr))) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE10swap_bytesEv: 22| 62.5k| simdutf_really_inline simd32 swap_bytes() const { 23| 62.5k| const __m128i shuffle = 24| 62.5k| _mm_setr_epi8(3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12); 25| | 26| 62.5k| return _mm_shuffle_epi8(value, shuffle); 27| 62.5k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd3maxENS2_6simd32IjEES4_: 94| 125k| const simd32 b) { 95| 125k| return _mm_max_epu32(a.value, b.value); 96| 125k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdplENS2_6simd32IjEES4_: 119| 62.5k| const simd32 b) { 120| 62.5k| return _mm_add_epi32(a.value, b.value); 121| 62.5k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdgtENS2_6simd32IjEES4_: 143| 3.05k| const simd32 b) { 144| 3.05k| return !(b >= a); 145| 3.05k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdntENS2_6simd32IbEE: 138| 3.05k|simdutf_really_inline simd32 operator!(const simd32 v) { 139| 3.05k| return _mm_xor_si128(v.value, _mm_set1_epi8(-1)); 140| 3.05k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IbEC2EDv2_x: 69| 6.10k| simdutf_really_inline simd32(const __m128i v) : value(v) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdgeENS2_6simd32IjEES4_: 134| 3.05k| const simd32 b) { 135| 3.05k| return _mm_cmpeq_epi32(_mm_max_epu32(a.value, b.value), a.value); 136| 3.05k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd32IbE3anyEv: 71| 3.05k| simdutf_really_inline bool any() const { 72| 3.05k| return _mm_movemask_epi8(value) != 0; 73| 3.05k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdanENS2_6simd32IjEES4_: 104| 65.0k| const simd32 b) { 105| 65.0k| return _mm_and_si128(a.value, b.value); 106| 65.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd3minENS2_6simd32IjEES4_: 89| 65.0k| const simd32 b) { 90| 65.0k| return _mm_min_epu32(a.value, b.value); 91| 65.0k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjEpLES4_: 46| 65.0k| simdutf_really_inline simd32 &operator+=(const simd32 other) { 47| 65.0k| value = _mm_add_epi32(value, other.value); 48| 65.0k| return *this; 49| 65.0k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE3sumEv: 15| 950| simdutf_really_inline uint64_t sum() const { 16| 950| return uint64_t(_mm_extract_epi32(value, 0)) + 17| 950| uint64_t(_mm_extract_epi32(value, 1)) + 18| 950| uint64_t(_mm_extract_epi32(value, 2)) + 19| | uint64_t(_mm_extract_epi32(value, 3)); 20| 950| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd64ImE4zeroEv: 26| 33.1k| simdutf_really_inline static simd64 zero() { 27| 33.1k| return _mm_setzero_si128(); 28| 33.1k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd64ImEC2EDv2_x: 9| 33.1k| simdutf_really_inline simd64(const __m128i v) : value(v) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd64ImE3sumEv: 15| 33.1k| simdutf_really_inline uint64_t sum() const { 16| 33.1k| return _mm_extract_epi64(value, 0) + _mm_extract_epi64(value, 1); 17| 33.1k| } _ZNK7simdutf8westmere14implementation13validate_utf8EPKcm: 297| 22.0k|implementation::validate_utf8(const char *buf, size_t len) const noexcept { 298| 22.0k| return westmere::utf8_validation::generic_validate_utf8(buf, len); 299| 22.0k|} _ZNK7simdutf8westmere14implementation25validate_utf8_with_errorsEPKcm: 304| 5.56k| const char *buf, size_t len) const noexcept { 305| 5.56k| return westmere::utf8_validation::generic_validate_utf8_with_errors(buf, len); 306| 5.56k|} _ZNK7simdutf8westmere14implementation16validate_utf16leEPKDsm: 343| 16.2k| size_t len) const noexcept { 344| 16.2k| if (simdutf_unlikely(len == 0)) { ------------------ | | 92| 16.2k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 14.3k, False: 1.87k] | | ------------------ ------------------ 345| | // empty input is valid UTF-16. protect the implementation from 346| | // handling nullptr 347| 14.3k| return true; 348| 14.3k| } 349| 1.87k| const auto res = 350| 1.87k| westmere::utf16::validate_utf16_with_errors(buf, len); 351| 1.87k| if (res.is_err()) { ------------------ | Branch (351:7): [True: 53, False: 1.81k] ------------------ 352| 53| return false; 353| 53| } 354| | 355| 1.81k| if (res.count == len) ------------------ | Branch (355:7): [True: 0, False: 1.81k] ------------------ 356| 0| return true; 357| | 358| 1.81k| return scalar::utf16::validate(buf + res.count, 359| 1.81k| len - res.count); 360| 1.81k|} _ZNK7simdutf8westmere14implementation16validate_utf16beEPKDsm: 366| 16.2k| size_t len) const noexcept { 367| 16.2k| if (simdutf_unlikely(len == 0)) { ------------------ | | 92| 16.2k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (92:33): [True: 14.4k, False: 1.85k] | | ------------------ ------------------ 368| | // empty input is valid UTF-16. protect the implementation from 369| | // handling nullptr 370| 14.4k| return true; 371| 14.4k| } 372| 1.85k| const auto res = 373| 1.85k| westmere::utf16::validate_utf16_with_errors(buf, len); 374| 1.85k| if (res.is_err()) { ------------------ | Branch (374:7): [True: 49, False: 1.80k] ------------------ 375| 49| return false; 376| 49| } 377| | 378| 1.80k| if (res.count == len) ------------------ | Branch (378:7): [True: 0, False: 1.80k] ------------------ 379| 0| return true; 380| | 381| 1.80k| return scalar::utf16::validate(buf + res.count, 382| 1.80k| len - res.count); 383| 1.80k|} _ZNK7simdutf8westmere14implementation28validate_utf16le_with_errorsEPKDsm: 386| 5.56k| const char16_t *buf, size_t len) const noexcept { 387| 5.56k| const result res = 388| 5.56k| westmere::utf16::validate_utf16_with_errors(buf, len); 389| 5.56k| if (res.count != len) { ------------------ | Branch (389:7): [True: 347, False: 5.21k] ------------------ 390| 347| const result scalar_res = 391| 347| scalar::utf16::validate_with_errors( 392| 347| buf + res.count, len - res.count); 393| 347| return result(scalar_res.error, res.count + scalar_res.count); 394| 5.21k| } else { 395| 5.21k| return res; 396| 5.21k| } 397| 5.56k|} _ZNK7simdutf8westmere14implementation28validate_utf16be_with_errorsEPKDsm: 400| 5.56k| const char16_t *buf, size_t len) const noexcept { 401| 5.56k| const result res = 402| 5.56k| westmere::utf16::validate_utf16_with_errors(buf, len); 403| 5.56k| if (res.count != len) { ------------------ | Branch (403:7): [True: 326, False: 5.23k] ------------------ 404| 326| result scalar_res = scalar::utf16::validate_with_errors( 405| 326| buf + res.count, len - res.count); 406| 326| return result(scalar_res.error, res.count + scalar_res.count); 407| 5.23k| } else { 408| 5.23k| return res; 409| 5.23k| } 410| 5.56k|} _ZNK7simdutf8westmere14implementation14validate_utf32EPKDim: 425| 10.7k|implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept { 426| 10.7k| return utf32::validate(buf, len); 427| 10.7k|} _ZNK7simdutf8westmere14implementation22convert_latin1_to_utf8EPKcmPc: 439| 5.86k| const char *buf, size_t len, char *utf8_output) const noexcept { 440| | 441| 5.86k| std::pair ret = 442| 5.86k| sse_convert_latin1_to_utf8(buf, len, utf8_output); 443| 5.86k| size_t converted_chars = ret.second - utf8_output; 444| | 445| 5.86k| if (ret.first != buf + len) { ------------------ | Branch (445:7): [True: 830, False: 5.03k] ------------------ 446| 830| const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert( 447| 830| ret.first, len - (ret.first - buf), ret.second); 448| 830| converted_chars += scalar_converted_chars; 449| 830| } 450| | 451| 5.86k| return converted_chars; 452| 5.86k|} _ZNK7simdutf8westmere14implementation25convert_latin1_to_utf16leEPKcmPDs: 457| 5.56k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 458| 5.56k| std::pair ret = 459| 5.56k| sse_convert_latin1_to_utf16(buf, len, utf16_output); 460| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (460:7): [True: 0, False: 5.56k] ------------------ 461| 0| return 0; 462| 0| } 463| 5.56k| size_t converted_chars = ret.second - utf16_output; 464| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (464:7): [True: 458, False: 5.10k] ------------------ 465| 458| const size_t scalar_converted_chars = 466| 458| scalar::latin1_to_utf16::convert( 467| 458| ret.first, len - (ret.first - buf), ret.second); 468| 458| if (scalar_converted_chars == 0) { ------------------ | Branch (468:9): [True: 0, False: 458] ------------------ 469| 0| return 0; 470| 0| } 471| 458| converted_chars += scalar_converted_chars; 472| 458| } 473| 5.56k| return converted_chars; 474| 5.56k|} _ZNK7simdutf8westmere14implementation25convert_latin1_to_utf16beEPKcmPDs: 477| 5.56k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 478| 5.56k| std::pair ret = 479| 5.56k| sse_convert_latin1_to_utf16(buf, len, utf16_output); 480| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (480:7): [True: 0, False: 5.56k] ------------------ 481| 0| return 0; 482| 0| } 483| 5.56k| size_t converted_chars = ret.second - utf16_output; 484| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (484:7): [True: 458, False: 5.10k] ------------------ 485| 458| const size_t scalar_converted_chars = 486| 458| scalar::latin1_to_utf16::convert( 487| 458| ret.first, len - (ret.first - buf), ret.second); 488| 458| if (scalar_converted_chars == 0) { ------------------ | Branch (488:9): [True: 0, False: 458] ------------------ 489| 0| return 0; 490| 0| } 491| 458| converted_chars += scalar_converted_chars; 492| 458| } 493| 5.56k| return converted_chars; 494| 5.56k|} _ZNK7simdutf8westmere14implementation23convert_latin1_to_utf32EPKcmPDi: 499| 5.56k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 500| 5.56k| std::pair ret = 501| 5.56k| sse_convert_latin1_to_utf32(buf, len, utf32_output); 502| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (502:7): [True: 0, False: 5.56k] ------------------ 503| 0| return 0; 504| 0| } 505| 5.56k| size_t converted_chars = ret.second - utf32_output; 506| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (506:7): [True: 458, False: 5.10k] ------------------ 507| 458| const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert( 508| 458| ret.first, len - (ret.first - buf), ret.second); 509| 458| if (scalar_converted_chars == 0) { ------------------ | Branch (509:9): [True: 0, False: 458] ------------------ 510| 0| return 0; 511| 0| } 512| 458| converted_chars += scalar_converted_chars; 513| 458| } 514| 5.56k| return converted_chars; 515| 5.56k|} _ZNK7simdutf8westmere14implementation22convert_utf8_to_latin1EPKcmPc: 520| 11.1k| const char *buf, size_t len, char *latin1_output) const noexcept { 521| 11.1k| utf8_to_latin1::validating_transcoder converter; 522| 11.1k| return converter.convert(buf, len, latin1_output); 523| 11.1k|} _ZNK7simdutf8westmere14implementation23convert_utf8_to_utf16leEPKcmPDs: 539| 11.0k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 540| 11.0k| utf8_to_utf16::validating_transcoder converter; 541| 11.0k| return converter.convert(buf, len, utf16_output); 542| 11.0k|} _ZNK7simdutf8westmere14implementation23convert_utf8_to_utf16beEPKcmPDs: 545| 11.0k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 546| 11.0k| utf8_to_utf16::validating_transcoder converter; 547| 11.0k| return converter.convert(buf, len, utf16_output); 548| 11.0k|} _ZNK7simdutf8westmere14implementation21convert_utf8_to_utf32EPKcmPDi: 578| 5.56k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 579| 5.56k| utf8_to_utf32::validating_transcoder converter; 580| 5.56k| return converter.convert(buf, len, utf32_output); 581| 5.56k|} _ZNK7simdutf8westmere14implementation25convert_utf16le_to_latin1EPKDsmPc: 597| 5.56k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 598| 5.56k| std::pair ret = 599| 5.56k| sse_convert_utf16_to_latin1(buf, len, latin1_output); 600| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (600:7): [True: 0, False: 5.56k] ------------------ 601| 0| return 0; 602| 0| } 603| 5.56k| size_t saved_bytes = ret.second - latin1_output; 604| | 605| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (605:7): [True: 428, False: 5.13k] ------------------ 606| 428| const size_t scalar_saved_bytes = 607| 428| scalar::utf16_to_latin1::convert( 608| 428| ret.first, len - (ret.first - buf), ret.second); 609| 428| if (scalar_saved_bytes == 0) { ------------------ | Branch (609:9): [True: 0, False: 428] ------------------ 610| 0| return 0; 611| 0| } 612| 428| saved_bytes += scalar_saved_bytes; 613| 428| } 614| 5.56k| return saved_bytes; 615| 5.56k|} _ZNK7simdutf8westmere14implementation25convert_utf16be_to_latin1EPKDsmPc: 618| 5.56k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 619| 5.56k| std::pair ret = 620| 5.56k| sse_convert_utf16_to_latin1(buf, len, latin1_output); 621| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (621:7): [True: 0, False: 5.56k] ------------------ 622| 0| return 0; 623| 0| } 624| 5.56k| size_t saved_bytes = ret.second - latin1_output; 625| | 626| 5.56k| if (ret.first != buf + len) { ------------------ | Branch (626:7): [True: 428, False: 5.13k] ------------------ 627| 428| const size_t scalar_saved_bytes = 628| 428| scalar::utf16_to_latin1::convert( 629| 428| ret.first, len - (ret.first - buf), ret.second); 630| 428| if (scalar_saved_bytes == 0) { ------------------ | Branch (630:9): [True: 0, False: 428] ------------------ 631| 0| return 0; 632| 0| } 633| 428| saved_bytes += scalar_saved_bytes; 634| 428| } 635| 5.56k| return saved_bytes; 636| 5.56k|} _ZNK7simdutf8westmere14implementation23convert_utf16le_to_utf8EPKDsmPc: 707| 10.7k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 708| 10.7k| std::pair ret = 709| 10.7k| sse_convert_utf16_to_utf8(buf, len, utf8_output); 710| 10.7k| if (ret.first == nullptr) { ------------------ | Branch (710:7): [True: 59, False: 10.6k] ------------------ 711| 59| return 0; 712| 59| } 713| 10.6k| size_t saved_bytes = ret.second - utf8_output; 714| 10.6k| if (ret.first != buf + len) { ------------------ | Branch (714:7): [True: 1.03k, False: 9.61k] ------------------ 715| 1.03k| const size_t scalar_saved_bytes = 716| 1.03k| scalar::utf16_to_utf8::convert( 717| 1.03k| ret.first, len - (ret.first - buf), ret.second); 718| 1.03k| if (scalar_saved_bytes == 0) { ------------------ | Branch (718:9): [True: 44, False: 987] ------------------ 719| 44| return 0; 720| 44| } 721| 987| saved_bytes += scalar_saved_bytes; 722| 987| } 723| 10.6k| return saved_bytes; 724| 10.6k|} _ZNK7simdutf8westmere14implementation23convert_utf16be_to_utf8EPKDsmPc: 727| 10.7k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 728| 10.7k| std::pair ret = 729| 10.7k| sse_convert_utf16_to_utf8(buf, len, utf8_output); 730| 10.7k| if (ret.first == nullptr) { ------------------ | Branch (730:7): [True: 56, False: 10.6k] ------------------ 731| 56| return 0; 732| 56| } 733| 10.6k| size_t saved_bytes = ret.second - utf8_output; 734| 10.6k| if (ret.first != buf + len) { ------------------ | Branch (734:7): [True: 1.01k, False: 9.63k] ------------------ 735| 1.01k| const size_t scalar_saved_bytes = 736| 1.01k| scalar::utf16_to_utf8::convert( 737| 1.01k| ret.first, len - (ret.first - buf), ret.second); 738| 1.01k| if (scalar_saved_bytes == 0) { ------------------ | Branch (738:9): [True: 33, False: 980] ------------------ 739| 33| return 0; 740| 33| } 741| 980| saved_bytes += scalar_saved_bytes; 742| 980| } 743| 10.6k| return saved_bytes; 744| 10.6k|} _ZNK7simdutf8westmere14implementation23convert_utf32_to_latin1EPKDimPc: 815| 5.56k| const char32_t *buf, size_t len, char *latin1_output) const noexcept { 816| 5.56k| std::pair ret = 817| 5.56k| sse_convert_utf32_to_latin1(buf, len, latin1_output); 818| 5.56k| if (ret.first == nullptr) { ------------------ | Branch (818:7): [True: 0, False: 5.56k] ------------------ 819| 0| return 0; 820| 0| } 821| 5.56k| size_t saved_bytes = ret.second - latin1_output; 822| | // if (ret.first != buf + len) { 823| 5.56k| if (ret.first < buf + len) { ------------------ | Branch (823:7): [True: 458, False: 5.10k] ------------------ 824| 458| const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert( 825| 458| ret.first, len - (ret.first - buf), ret.second); 826| 458| if (scalar_saved_bytes == 0) { ------------------ | Branch (826:9): [True: 0, False: 458] ------------------ 827| 0| return 0; 828| 0| } 829| 458| saved_bytes += scalar_saved_bytes; 830| 458| } 831| 5.56k| return saved_bytes; 832| 5.56k|} _ZNK7simdutf8westmere14implementation21convert_utf32_to_utf8EPKDimPc: 866| 5.14k| const char32_t *buf, size_t len, char *utf8_output) const noexcept { 867| 5.14k| std::pair ret = 868| 5.14k| sse_convert_utf32_to_utf8(buf, len, utf8_output); 869| 5.14k| if (ret.first == nullptr) { ------------------ | Branch (869:7): [True: 0, False: 5.14k] ------------------ 870| 0| return 0; 871| 0| } 872| 5.14k| size_t saved_bytes = ret.second - utf8_output; 873| 5.14k| if (ret.first != buf + len) { ------------------ | Branch (873:7): [True: 743, False: 4.39k] ------------------ 874| 743| const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert( 875| 743| ret.first, len - (ret.first - buf), ret.second); 876| 743| if (scalar_saved_bytes == 0) { ------------------ | Branch (876:9): [True: 0, False: 743] ------------------ 877| 0| return 0; 878| 0| } 879| 743| saved_bytes += scalar_saved_bytes; 880| 743| } 881| 5.14k| return saved_bytes; 882| 5.14k|} _ZNK7simdutf8westmere14implementation10count_utf8EPKcm: 1142| 11.1k|implementation::count_utf8(const char *input, size_t length) const noexcept { 1143| 11.1k| return utf8::count_code_points_bytemask(input, length); 1144| 11.1k|} _ZNK7simdutf8westmere14implementation23latin1_length_from_utf8EPKcm: 1149| 11.1k| const char *buf, size_t len) const noexcept { 1150| 11.1k| return count_utf8(buf, len); 1151| 11.1k|} _ZNK7simdutf8westmere14implementation24utf8_length_from_utf16leEPKDsm: 1156| 10.7k| const char16_t *input, size_t length) const noexcept { 1157| 10.7k| return utf16::utf8_length_from_utf16_bytemask(input, 1158| 10.7k| length); 1159| 10.7k|} _ZNK7simdutf8westmere14implementation24utf8_length_from_utf16beEPKDsm: 1162| 10.7k| const char16_t *input, size_t length) const noexcept { 1163| 10.7k| return utf16::utf8_length_from_utf16_bytemask(input, length); 1164| 10.7k|} _ZNK7simdutf8westmere14implementation23utf8_length_from_latin1EPKcm: 1169| 5.86k| const char *input, size_t len) const noexcept { 1170| 5.86k| const uint8_t *str = reinterpret_cast(input); 1171| 5.86k| size_t answer = len / sizeof(__m128i) * sizeof(__m128i); 1172| 5.86k| size_t i = 0; 1173| 5.86k| if (answer >= 2048) { // long strings optimization ------------------ | Branch (1173:7): [True: 0, False: 5.86k] ------------------ 1174| 0| __m128i two_64bits = _mm_setzero_si128(); 1175| 0| while (i + sizeof(__m128i) <= len) { ------------------ | Branch (1175:12): [True: 0, False: 0] ------------------ 1176| 0| __m128i runner = _mm_setzero_si128(); 1177| 0| size_t iterations = (len - i) / sizeof(__m128i); 1178| 0| if (iterations > 255) { ------------------ | Branch (1178:11): [True: 0, False: 0] ------------------ 1179| 0| iterations = 255; 1180| 0| } 1181| 0| size_t max_i = i + iterations * sizeof(__m128i) - sizeof(__m128i); 1182| 0| for (; i + 4 * sizeof(__m128i) <= max_i; i += 4 * sizeof(__m128i)) { ------------------ | Branch (1182:14): [True: 0, False: 0] ------------------ 1183| 0| __m128i input1 = _mm_loadu_si128((const __m128i *)(str + i)); 1184| 0| __m128i input2 = 1185| 0| _mm_loadu_si128((const __m128i *)(str + i + sizeof(__m128i))); 1186| 0| __m128i input3 = 1187| 0| _mm_loadu_si128((const __m128i *)(str + i + 2 * sizeof(__m128i))); 1188| 0| __m128i input4 = 1189| 0| _mm_loadu_si128((const __m128i *)(str + i + 3 * sizeof(__m128i))); 1190| 0| __m128i input12 = 1191| 0| _mm_add_epi8(_mm_cmpgt_epi8(_mm_setzero_si128(), input1), 1192| 0| _mm_cmpgt_epi8(_mm_setzero_si128(), input2)); 1193| 0| __m128i input34 = 1194| 0| _mm_add_epi8(_mm_cmpgt_epi8(_mm_setzero_si128(), input3), 1195| 0| _mm_cmpgt_epi8(_mm_setzero_si128(), input4)); 1196| 0| __m128i input1234 = _mm_add_epi8(input12, input34); 1197| 0| runner = _mm_sub_epi8(runner, input1234); 1198| 0| } 1199| 0| for (; i <= max_i; i += sizeof(__m128i)) { ------------------ | Branch (1199:14): [True: 0, False: 0] ------------------ 1200| 0| __m128i more_input = _mm_loadu_si128((const __m128i *)(str + i)); 1201| 0| runner = _mm_sub_epi8(runner, 1202| 0| _mm_cmpgt_epi8(_mm_setzero_si128(), more_input)); 1203| 0| } 1204| 0| two_64bits = 1205| 0| _mm_add_epi64(two_64bits, _mm_sad_epu8(runner, _mm_setzero_si128())); 1206| 0| } 1207| 0| answer += 1208| 0| _mm_extract_epi64(two_64bits, 0) + _mm_extract_epi64(two_64bits, 1); 1209| 5.86k| } else if (answer > 0) { // short string optimization ------------------ | Branch (1209:14): [True: 821, False: 5.03k] ------------------ 1210| 7.48k| for (; i + 2 * sizeof(__m128i) <= len; i += 2 * sizeof(__m128i)) { ------------------ | Branch (1210:12): [True: 6.66k, False: 821] ------------------ 1211| 6.66k| __m128i latin = _mm_loadu_si128((const __m128i *)(input + i)); 1212| 6.66k| uint16_t non_ascii = (uint16_t)_mm_movemask_epi8(latin); 1213| 6.66k| answer += count_ones(non_ascii); 1214| 6.66k| latin = _mm_loadu_si128((const __m128i *)(input + i) + 1); 1215| 6.66k| non_ascii = (uint16_t)_mm_movemask_epi8(latin); 1216| 6.66k| answer += count_ones(non_ascii); 1217| 6.66k| } 1218| 1.05k| for (; i + sizeof(__m128i) <= len; i += sizeof(__m128i)) { ------------------ | Branch (1218:12): [True: 230, False: 821] ------------------ 1219| 230| __m128i latin = _mm_loadu_si128((const __m128i *)(input + i)); 1220| 230| uint16_t non_ascii = (uint16_t)_mm_movemask_epi8(latin); 1221| 230| answer += count_ones(non_ascii); 1222| 230| } 1223| 821| } 1224| 5.86k| return answer + scalar::latin1::utf8_length_from_latin1( 1225| 5.86k| reinterpret_cast(str + i), len - i); 1226| 5.86k|} _ZNK7simdutf8westmere14implementation22utf16_length_from_utf8EPKcm: 1243| 22.0k| const char *input, size_t length) const noexcept { 1244| 22.0k| return utf8::utf16_length_from_utf8_bytemask(input, length); 1245| 22.0k|} _ZNK7simdutf8westmere14implementation22utf8_length_from_utf32EPKDim: 1278| 5.14k| const char32_t *input, size_t length) const noexcept { 1279| 5.14k| return utf32::utf8_length_from_utf32(input, length); 1280| 5.14k|} _ZNK7simdutf8westmere14implementation22utf32_length_from_utf8EPKcm: 1306| 5.56k| const char *input, size_t length) const noexcept { 1307| 5.56k| return utf8::count_code_points(input, length); 1308| 5.56k|} _ZNK7simdutf8westmere14implementation16base64_to_binaryEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 1314| 16.1k| last_chunk_handling_options last_chunk_options) const noexcept { 1315| 16.1k| if (options & base64_default_or_url) { ------------------ | Branch (1315:7): [True: 0, False: 16.1k] ------------------ 1316| 0| if (options == base64_options::base64_default_or_url_accept_garbage) { ------------------ | Branch (1316:9): [True: 0, False: 0] ------------------ 1317| 0| return base64::compress_decode_base64( 1318| 0| output, input, length, options, last_chunk_options); 1319| 0| } else { 1320| 0| return base64::compress_decode_base64( 1321| 0| output, input, length, options, last_chunk_options); 1322| 0| } 1323| 16.1k| } else if (options & base64_url) { ------------------ | Branch (1323:14): [True: 0, False: 16.1k] ------------------ 1324| 0| if (options == base64_options::base64_url_accept_garbage) { ------------------ | Branch (1324:9): [True: 0, False: 0] ------------------ 1325| 0| return base64::compress_decode_base64( 1326| 0| output, input, length, options, last_chunk_options); 1327| 0| } else { 1328| 0| return base64::compress_decode_base64( 1329| 0| output, input, length, options, last_chunk_options); 1330| 0| } 1331| 16.1k| } else { 1332| 16.1k| if (options == base64_options::base64_default_accept_garbage) { ------------------ | Branch (1332:9): [True: 0, False: 16.1k] ------------------ 1333| 0| return base64::compress_decode_base64( 1334| 0| output, input, length, options, last_chunk_options); 1335| 16.1k| } else { 1336| 16.1k| return base64::compress_decode_base64( 1337| 16.1k| output, input, length, options, last_chunk_options); 1338| 16.1k| } 1339| 16.1k| } 1340| 16.1k|} _ZNK7simdutf8westmere14implementation16binary_to_base64EPKcmPcNS_14base64_optionsE: 1434| 15.6k| base64_options options) const noexcept { 1435| 15.6k| if (options & base64_url) { ------------------ | Branch (1435:7): [True: 0, False: 15.6k] ------------------ 1436| 0| return encode_base64(output, input, length, options); 1437| 15.6k| } else { 1438| 15.6k| return encode_base64(output, input, length, options); 1439| 15.6k| } 1440| 15.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18is_asciiERKNS1_4simd8simd8x64IhEE: 13| 37.2k|simdutf_really_inline bool is_ascii(const simd8x64 &input) { 14| 37.2k| return input.reduce_or().is_ascii(); 15| 37.2k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_124must_be_2_3_continuationENS1_4simd5simd8IhEES4_: 22| 68.7k| const simd8 prev3) { 23| 68.7k| simd8 is_third_byte = 24| 68.7k| prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be >= 0x80 25| 68.7k| simd8 is_fourth_byte = 26| 68.7k| prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be >= 0x80 27| 68.7k| return simd8(is_third_byte | is_fourth_byte); 28| 68.7k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18internal8westmere26write_v_u16_11bits_to_utf8EDv2_xRPcS4_S4_: 57| 8.44k| const __m128i v_ff80) { 58| | // no bits set above 7th bit 59| 8.44k| const __m128i one_byte_bytemask = 60| 8.44k| _mm_cmpeq_epi16(_mm_and_si128(v_u16, v_ff80), v_0000); 61| 8.44k| const uint16_t one_byte_bitmask = 62| 8.44k| static_cast(_mm_movemask_epi8(one_byte_bytemask)); 63| | 64| 8.44k| write_v_u16_11bits_to_utf8(v_u16, utf8_output, one_byte_bytemask, 65| 8.44k| one_byte_bitmask); 66| 8.44k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18internal8westmere26write_v_u16_11bits_to_utf8EDv2_xRPcS4_t: 10| 12.2k| const uint16_t one_byte_bitmask) { 11| | // 0b1100_0000_1000_0000 12| 12.2k| const __m128i v_c080 = _mm_set1_epi16((int16_t)0xc080); 13| | // 0b0001_1111_0000_0000 14| 12.2k| const __m128i v_1f00 = _mm_set1_epi16((int16_t)0x1f00); 15| | // 0b0000_0000_0011_1111 16| 12.2k| const __m128i v_003f = _mm_set1_epi16((int16_t)0x003f); 17| | 18| | // 1. prepare 2-byte values 19| | // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8 20| | // expected output : [110a|aaaa|10bb|bbbb] x 8 21| | 22| | // t0 = [000a|aaaa|bbbb|bb00] 23| 12.2k| const __m128i t0 = _mm_slli_epi16(v_u16, 2); 24| | // t1 = [000a|aaaa|0000|0000] 25| 12.2k| const __m128i t1 = _mm_and_si128(t0, v_1f00); 26| | // t2 = [0000|0000|00bb|bbbb] 27| 12.2k| const __m128i t2 = _mm_and_si128(v_u16, v_003f); 28| | // t3 = [000a|aaaa|00bb|bbbb] 29| 12.2k| const __m128i t3 = _mm_or_si128(t1, t2); 30| | // t4 = [110a|aaaa|10bb|bbbb] 31| 12.2k| const __m128i t4 = _mm_or_si128(t3, v_c080); 32| | 33| | // 2. merge ASCII and 2-byte codewords 34| 12.2k| const __m128i utf8_unpacked = _mm_blendv_epi8(t4, v_u16, one_byte_bytemask); 35| | 36| | // 3. prepare bitmask for 8-bit lookup 37| | // one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - MSB, a 38| | // - LSB) 39| 12.2k| const uint16_t m0 = one_byte_bitmask & 0x5555; // m0 = 0h0g0f0e0d0c0b0a 40| 12.2k| const uint16_t m1 = static_cast(m0 >> 7); // m1 = 00000000h0g0f0e0 41| 12.2k| const uint8_t m2 = static_cast((m0 | m1) & 0xff); // m2 = hdgcfbea 42| | // 4. pack the bytes 43| 12.2k| const uint8_t *row = 44| 12.2k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0]; 45| 12.2k| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 46| 12.2k| const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle); 47| | 48| | // 5. store bytes 49| 12.2k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 50| | 51| | // 6. adjust pointers 52| 12.2k| utf8_output += row[0]; 53| 12.2k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block64C2EPKc: 413| 3.97k| simdutf_really_inline block64(const char *src) { 414| 3.97k| chunks[0] = _mm_loadu_si128(reinterpret_cast(src)); 415| 3.97k| chunks[1] = _mm_loadu_si128(reinterpret_cast(src + 16)); 416| 3.97k| chunks[2] = _mm_loadu_si128(reinterpret_cast(src + 32)); 417| 3.97k| chunks[3] = _mm_loadu_si128(reinterpret_cast(src + 48)); 418| 3.97k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6414compress_blockEmPc: 454| 752| simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) { 455| 752| if (is_power_of_two(mask)) { ------------------ | Branch (455:9): [True: 213, False: 539] ------------------ 456| 213| return compress_block_single(mask, output); 457| 213| } 458| | 459| 539| uint64_t nmask = ~mask; 460| 539| compress(chunks[0], uint16_t(mask), output); 461| 539| compress(chunks[1], uint16_t(mask >> 16), 462| 539| output + count_ones(nmask & 0xFFFF)); 463| 539| compress(chunks[2], uint16_t(mask >> 32), 464| 539| output + count_ones(nmask & 0xFFFFFFFF)); 465| 539| compress(chunks[3], uint16_t(mask >> 48), 466| 539| output + count_ones(nmask & 0xFFFFFFFFFFFFULL)); 467| 539| return count_ones(nmask); 468| 752| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6421compress_block_singleEmPc: 472| 213| char *output) { 473| 213| const size_t pos64 = trailing_zeroes(mask); 474| 213| const int8_t pos = pos64 & 0xf; 475| 213| switch (pos64 >> 4) { ------------------ | Branch (475:13): [True: 213, False: 0] ------------------ 476| 48| case 0b00: { ------------------ | Branch (476:5): [True: 48, False: 165] ------------------ 477| 48| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 478| 48| const __m128i v1 = 479| 48| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 480| 48| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 481| 48| const __m128i sh = _mm_sub_epi8(v1, v2); 482| 48| const __m128i compressed = _mm_shuffle_epi8(chunks[0], sh); 483| | 484| 48| _mm_storeu_si128((__m128i *)(output + 0 * 16), compressed); 485| 48| _mm_storeu_si128((__m128i *)(output + 1 * 16 - 1), chunks[1]); 486| 48| _mm_storeu_si128((__m128i *)(output + 2 * 16 - 1), chunks[2]); 487| 48| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]); 488| 48| } break; 489| 66| case 0b01: { ------------------ | Branch (489:5): [True: 66, False: 147] ------------------ 490| 66| _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]); 491| | 492| 66| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 493| 66| const __m128i v1 = 494| 66| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 495| 66| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 496| 66| const __m128i sh = _mm_sub_epi8(v1, v2); 497| 66| const __m128i compressed = _mm_shuffle_epi8(chunks[1], sh); 498| | 499| 66| _mm_storeu_si128((__m128i *)(output + 1 * 16), compressed); 500| 66| _mm_storeu_si128((__m128i *)(output + 2 * 16 - 1), chunks[2]); 501| 66| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]); 502| 66| } break; 503| 50| case 0b10: { ------------------ | Branch (503:5): [True: 50, False: 163] ------------------ 504| 50| _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]); 505| 50| _mm_storeu_si128((__m128i *)(output + 1 * 16), chunks[1]); 506| | 507| 50| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 508| 50| const __m128i v1 = 509| 50| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 510| 50| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 511| 50| const __m128i sh = _mm_sub_epi8(v1, v2); 512| 50| const __m128i compressed = _mm_shuffle_epi8(chunks[2], sh); 513| | 514| 50| _mm_storeu_si128((__m128i *)(output + 2 * 16), compressed); 515| 50| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]); 516| 50| } break; 517| 49| case 0b11: { ------------------ | Branch (517:5): [True: 49, False: 164] ------------------ 518| 49| _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]); 519| 49| _mm_storeu_si128((__m128i *)(output + 1 * 16), chunks[1]); 520| 49| _mm_storeu_si128((__m128i *)(output + 2 * 16), chunks[2]); 521| | 522| 49| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 523| 49| const __m128i v1 = 524| 49| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 525| 49| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 526| 49| const __m128i sh = _mm_sub_epi8(v1, v2); 527| 49| const __m128i compressed = _mm_shuffle_epi8(chunks[3], sh); 528| | 529| 49| _mm_storeu_si128((__m128i *)(output + 3 * 16), compressed); 530| 49| } break; 531| 213| } 532| | 533| 213| return 63; 534| 213| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18compressEDv2_xtPc: 334| 2.15k| char *output) { 335| 2.15k| if (mask == 0) { ------------------ | Branch (335:7): [True: 832, False: 1.32k] ------------------ 336| 832| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data); 337| 832| return; 338| 832| } 339| | 340| | // this particular implementation was inspired by work done by @animetosho 341| | // we do it in two steps, first 8 bytes and then second 8 bytes 342| 1.32k| uint8_t mask1 = uint8_t(mask); // least significant 8 bits 343| 1.32k| uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits 344| | // next line just loads the 64-bit values thintable_epi8[mask1] and 345| | // thintable_epi8[mask2] into a 128-bit register, using only 346| | // two instructions on most compilers. 347| | 348| 1.32k| __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2], 349| 1.32k| tables::base64::thintable_epi8[mask1]); 350| | // we increment by 0x08 the second half of the mask 351| 1.32k| shufmask = 352| 1.32k| _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0)); 353| | // this is the version "nearly pruned" 354| 1.32k| __m128i pruned = _mm_shuffle_epi8(data, shufmask); 355| | // we still need to put the two halves together. 356| | // we compute the popcount of the first half: 357| 1.32k| int pop1 = tables::base64::BitsSetTable256mul2[mask1]; 358| | // then load the corresponding mask, what it does is to write 359| | // only the first pop1 bytes from the first 8 bytes, and then 360| | // it fills in with the bytes from the second 8 bytes + some filling 361| | // at the end. 362| 1.32k| __m128i compactmask = _mm_loadu_si128(reinterpret_cast( 363| 1.32k| tables::base64::pshufb_combine_table + pop1 * 8)); 364| 1.32k| __m128i answer = _mm_shuffle_epi8(pruned, compactmask); 365| 1.32k| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer); 366| 1.32k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6410copy_blockEPc: 446| 68| simdutf_really_inline void copy_block(char *output) { 447| 68| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), chunks[0]); 448| 68| _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 16), chunks[1]); 449| 68| _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 32), chunks[2]); 450| 68| _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 48), chunks[3]); 451| 68| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6424base64_decode_block_safeEPc: 664| 120| simdutf_really_inline void base64_decode_block_safe(char *out) { 665| 120| base64_decode(out, chunks[0]); 666| 120| base64_decode(out + 12, chunks[1]); 667| 120| base64_decode(out + 24, chunks[2]); 668| 120| char buffer[16]; 669| 120| base64_decode(buffer, chunks[3]); 670| 120| std::memcpy(out + 36, buffer, 12); 671| 120| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113base64_decodeEPcDv2_x: 368| 13.5k|static simdutf_really_inline void base64_decode(char *out, __m128i str) { 369| | // credit: aqrit 370| | 371| 13.5k| const __m128i pack_shuffle = 372| 13.5k| _mm_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1); 373| | 374| 13.5k| const __m128i t0 = _mm_maddubs_epi16(str, _mm_set1_epi32(0x01400140)); 375| 13.5k| const __m128i t1 = _mm_madd_epi16(t0, _mm_set1_epi32(0x00011000)); 376| 13.5k| const __m128i t2 = _mm_shuffle_epi8(t1, pack_shuffle); 377| | // Store the output: 378| | // this writes 16 bytes, but we only need 12. 379| 13.5k| _mm_storeu_si128((__m128i *)out, t2); 380| 13.5k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6419base64_decode_blockEPc: 656| 2.77k| simdutf_really_inline void base64_decode_block(char *out) { 657| 2.77k| base64_decode(out, chunks[0]); 658| 2.77k| base64_decode(out + 12, chunks[1]); 659| 2.77k| base64_decode(out + 24, chunks[2]); 660| 2.77k| base64_decode(out + 36, chunks[3]); 661| 2.77k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_119base64_decode_blockEPcPKc: 383| 473|static inline void base64_decode_block(char *out, const char *src) { 384| 473| base64_decode(out, _mm_loadu_si128(reinterpret_cast(src))); 385| 473| base64_decode(out + 12, 386| 473| _mm_loadu_si128(reinterpret_cast(src + 16))); 387| 473| base64_decode(out + 24, 388| 473| _mm_loadu_si128(reinterpret_cast(src + 32))); 389| 473| base64_decode(out + 36, 390| 473| _mm_loadu_si128(reinterpret_cast(src + 48))); 391| 473|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_124base64_decode_block_safeEPcPKc: 393| 21|static inline void base64_decode_block_safe(char *out, const char *src) { 394| 21| base64_decode(out, _mm_loadu_si128(reinterpret_cast(src))); 395| 21| base64_decode(out + 12, 396| 21| _mm_loadu_si128(reinterpret_cast(src + 16))); 397| 21| base64_decode(out + 24, 398| 21| _mm_loadu_si128(reinterpret_cast(src + 32))); 399| 21| char buffer[16]; 400| 21| base64_decode(buffer, 401| 21| _mm_loadu_si128(reinterpret_cast(src + 48))); 402| 21| std::memcpy(out + 36, buffer, 12); 403| 21|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEmPm: 538| 3.97k| simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) { 539| 3.97k| uint32_t err0 = 0; 540| 3.97k| uint32_t err1 = 0; 541| 3.97k| uint32_t err2 = 0; 542| 3.97k| uint32_t err3 = 0; 543| 3.97k| uint64_t m0 = to_base64_mask( 544| 3.97k| &chunks[0], &err0); 545| 3.97k| uint64_t m1 = to_base64_mask( 546| 3.97k| &chunks[1], &err1); 547| 3.97k| uint64_t m2 = to_base64_mask( 548| 3.97k| &chunks[2], &err2); 549| 3.97k| uint64_t m3 = to_base64_mask( 550| 3.97k| &chunks[3], &err3); 551| 3.97k| if (!ignore_garbage) { ------------------ | Branch (551:9): [True: 3.97k, Folded] ------------------ 552| 3.97k| *error = (err0) | ((uint64_t)err1 << 16) | ((uint64_t)err2 << 32) | 553| 3.97k| ((uint64_t)err3 << 48); 554| 3.97k| } 555| 3.97k| return m0 | (m1 << 16) | (m2 << 32) | (m3 << 48); 556| 3.97k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEtPDv2_xPj: 560| 15.8k| simdutf_really_inline uint16_t to_base64_mask(__m128i *src, uint32_t *error) { 561| 15.8k| const __m128i ascii_space_tbl = 562| 15.8k| _mm_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa, 563| 15.8k| 0x0, 0xc, 0xd, 0x0, 0x0); 564| | // credit: aqrit 565| 15.8k| __m128i delta_asso; 566| 15.8k| if (default_or_url) { ------------------ | Branch (566:9): [Folded, False: 15.8k] ------------------ 567| 0| delta_asso = 568| 0| _mm_setr_epi8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 569| 0| 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x16); 570| 15.8k| } else if (base64_url) { ------------------ | Branch (570:16): [Folded, False: 15.8k] ------------------ 571| 0| delta_asso = _mm_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 572| 0| 0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF); 573| 15.8k| } else { 574| 15.8k| delta_asso = 575| 15.8k| _mm_setr_epi8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 576| 15.8k| 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F); 577| 15.8k| } 578| 15.8k| __m128i delta_values; 579| 15.8k| if (default_or_url) { ------------------ | Branch (579:9): [Folded, False: 15.8k] ------------------ 580| 0| delta_values = _mm_setr_epi8( 581| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0x13), 582| 0| uint8_t(0x04), uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 583| 0| uint8_t(0xB9), uint8_t(0x00), uint8_t(0xFF), uint8_t(0x11), 584| 0| uint8_t(0xFF), uint8_t(0xBF), uint8_t(0x10), uint8_t(0xB9)); 585| | 586| 15.8k| } else if (base64_url) { ------------------ | Branch (586:16): [Folded, False: 15.8k] ------------------ 587| 0| delta_values = _mm_setr_epi8(0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), 588| 0| uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9), 589| 0| 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF), 590| 0| uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9)); 591| 15.8k| } else { 592| 15.8k| delta_values = 593| 15.8k| _mm_setr_epi8(int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), 594| 15.8k| int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), 595| 15.8k| int8_t(0xB9), int8_t(0x00), int8_t(0x10), int8_t(0xC3), 596| 15.8k| int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9)); 597| 15.8k| } 598| 15.8k| __m128i check_asso; 599| 15.8k| if (default_or_url) { ------------------ | Branch (599:9): [Folded, False: 15.8k] ------------------ 600| 0| check_asso = 601| 0| _mm_setr_epi8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 602| 0| 0x01, 0x03, 0x07, 0x0B, 0x0E, 0x0B, 0x06); 603| 15.8k| } else if (base64_url) { ------------------ | Branch (603:16): [Folded, False: 15.8k] ------------------ 604| 0| check_asso = _mm_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 605| 0| 0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6); 606| 15.8k| } else { 607| 15.8k| check_asso = 608| 15.8k| _mm_setr_epi8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 609| 15.8k| 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F); 610| 15.8k| } 611| 15.8k| __m128i check_values; 612| 15.8k| if (default_or_url) { ------------------ | Branch (612:9): [Folded, False: 15.8k] ------------------ 613| 0| check_values = _mm_setr_epi8( 614| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 615| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xD5), uint8_t(0xA6), 616| 0| uint8_t(0xB5), uint8_t(0xA1), uint8_t(0x00), uint8_t(0x80), 617| 0| uint8_t(0x00), uint8_t(0x80), uint8_t(0x00), uint8_t(0x80)); 618| 15.8k| } else if (base64_url) { ------------------ | Branch (618:16): [Folded, False: 15.8k] ------------------ 619| 0| check_values = _mm_setr_epi8(uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 620| 0| uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF), 621| 0| uint8_t(0xB6), uint8_t(0xA6), uint8_t(0xB5), 622| 0| uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, 623| 0| uint8_t(0x80), 0x0, uint8_t(0x80)); 624| 15.8k| } else { 625| 15.8k| check_values = 626| 15.8k| _mm_setr_epi8(int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), 627| 15.8k| int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), 628| 15.8k| int8_t(0xB5), int8_t(0x86), int8_t(0xD1), int8_t(0x80), 629| 15.8k| int8_t(0xB1), int8_t(0x80), int8_t(0x91), int8_t(0x80)); 630| 15.8k| } 631| 15.8k| const __m128i shifted = _mm_srli_epi32(*src, 3); 632| | 633| 15.8k| __m128i delta_hash = 634| 15.8k| _mm_avg_epu8(_mm_shuffle_epi8(delta_asso, *src), shifted); 635| 15.8k| if (default_or_url) { ------------------ | Branch (635:9): [Folded, False: 15.8k] ------------------ 636| 0| delta_hash = _mm_and_si128(delta_hash, _mm_set1_epi8(0xf)); 637| 0| } 638| 15.8k| const __m128i check_hash = 639| 15.8k| _mm_avg_epu8(_mm_shuffle_epi8(check_asso, *src), shifted); 640| | 641| 15.8k| const __m128i out = 642| 15.8k| _mm_adds_epi8(_mm_shuffle_epi8(delta_values, delta_hash), *src); 643| 15.8k| const __m128i chk = 644| 15.8k| _mm_adds_epi8(_mm_shuffle_epi8(check_values, check_hash), *src); 645| 15.8k| const int mask = _mm_movemask_epi8(chk); 646| 15.8k| if (!ignore_garbage && mask) { ------------------ | Branch (646:9): [True: 15.8k, Folded] | Branch (646:28): [True: 2.47k, False: 13.4k] ------------------ 647| 2.47k| __m128i ascii_space = 648| 2.47k| _mm_cmpeq_epi8(_mm_shuffle_epi8(ascii_space_tbl, *src), *src); 649| 2.47k| *error = (mask ^ _mm_movemask_epi8(ascii_space)); 650| 2.47k| } 651| 15.8k| *src = out; 652| 15.8k| return (uint16_t)mask; 653| 15.8k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113encode_base64ILb0EEEmPcPKcmNS_14base64_optionsE: 327| 15.6k| base64_options options) { 328| 15.6k| return encode_base64_impl(dst, src, srclen, options); 329| 15.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_118encode_base64_implILb0ELb0EEEmPcPKcmNS_14base64_optionsEm: 97| 15.6k| size_t line_length = simdutf::default_line_length) { 98| 15.6k| size_t offset = 0; 99| 15.6k| if (line_length < 4) { ------------------ | Branch (99:7): [True: 0, False: 15.6k] ------------------ 100| 0| line_length = 4; // We do not support line_length less than 4 101| 0| } 102| | // credit: Wojciech Muła 103| | // SSE (lookup: pshufb improved unrolled) 104| 15.6k| const uint8_t *input = (const uint8_t *)src; 105| | 106| 15.6k| uint8_t *out = (uint8_t *)dst; 107| 15.6k| const __m128i shuf = 108| 15.6k| _mm_set_epi8(10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1); 109| | 110| 15.6k| size_t i = 0; 111| 18.7k| for (; i + 52 <= srclen; i += 48) { ------------------ | Branch (111:10): [True: 3.13k, False: 15.6k] ------------------ 112| 3.13k| __m128i in0 = _mm_loadu_si128( 113| 3.13k| reinterpret_cast(input + i + 4 * 3 * 0)); 114| 3.13k| __m128i in1 = _mm_loadu_si128( 115| 3.13k| reinterpret_cast(input + i + 4 * 3 * 1)); 116| 3.13k| __m128i in2 = _mm_loadu_si128( 117| 3.13k| reinterpret_cast(input + i + 4 * 3 * 2)); 118| 3.13k| __m128i in3 = _mm_loadu_si128( 119| 3.13k| reinterpret_cast(input + i + 4 * 3 * 3)); 120| | 121| 3.13k| in0 = _mm_shuffle_epi8(in0, shuf); 122| 3.13k| in1 = _mm_shuffle_epi8(in1, shuf); 123| 3.13k| in2 = _mm_shuffle_epi8(in2, shuf); 124| 3.13k| in3 = _mm_shuffle_epi8(in3, shuf); 125| | 126| 3.13k| const __m128i t0_0 = _mm_and_si128(in0, _mm_set1_epi32(0x0fc0fc00)); 127| 3.13k| const __m128i t0_1 = _mm_and_si128(in1, _mm_set1_epi32(0x0fc0fc00)); 128| 3.13k| const __m128i t0_2 = _mm_and_si128(in2, _mm_set1_epi32(0x0fc0fc00)); 129| 3.13k| const __m128i t0_3 = _mm_and_si128(in3, _mm_set1_epi32(0x0fc0fc00)); 130| | 131| 3.13k| const __m128i t1_0 = _mm_mulhi_epu16(t0_0, _mm_set1_epi32(0x04000040)); 132| 3.13k| const __m128i t1_1 = _mm_mulhi_epu16(t0_1, _mm_set1_epi32(0x04000040)); 133| 3.13k| const __m128i t1_2 = _mm_mulhi_epu16(t0_2, _mm_set1_epi32(0x04000040)); 134| 3.13k| const __m128i t1_3 = _mm_mulhi_epu16(t0_3, _mm_set1_epi32(0x04000040)); 135| | 136| 3.13k| const __m128i t2_0 = _mm_and_si128(in0, _mm_set1_epi32(0x003f03f0)); 137| 3.13k| const __m128i t2_1 = _mm_and_si128(in1, _mm_set1_epi32(0x003f03f0)); 138| 3.13k| const __m128i t2_2 = _mm_and_si128(in2, _mm_set1_epi32(0x003f03f0)); 139| 3.13k| const __m128i t2_3 = _mm_and_si128(in3, _mm_set1_epi32(0x003f03f0)); 140| | 141| 3.13k| const __m128i t3_0 = _mm_mullo_epi16(t2_0, _mm_set1_epi32(0x01000010)); 142| 3.13k| const __m128i t3_1 = _mm_mullo_epi16(t2_1, _mm_set1_epi32(0x01000010)); 143| 3.13k| const __m128i t3_2 = _mm_mullo_epi16(t2_2, _mm_set1_epi32(0x01000010)); 144| 3.13k| const __m128i t3_3 = _mm_mullo_epi16(t2_3, _mm_set1_epi32(0x01000010)); 145| | 146| 3.13k| const __m128i input0 = _mm_or_si128(t1_0, t3_0); 147| 3.13k| const __m128i input1 = _mm_or_si128(t1_1, t3_1); 148| 3.13k| const __m128i input2 = _mm_or_si128(t1_2, t3_2); 149| 3.13k| const __m128i input3 = _mm_or_si128(t1_3, t3_3); 150| | 151| 3.13k| const __m128i t0 = lookup_pshufb_improved(input0); 152| 3.13k| const __m128i t1 = lookup_pshufb_improved(input1); 153| 3.13k| const __m128i t2 = lookup_pshufb_improved(input2); 154| 3.13k| const __m128i t3 = lookup_pshufb_improved(input3); 155| | 156| 3.13k| if (use_lines) { ------------------ | Branch (156:9): [Folded, False: 3.13k] ------------------ 157| 0| if (line_length >= 64) { // fast path ------------------ | Branch (157:11): [True: 0, False: 0] ------------------ 158| 0| if (offset + 64 > line_length) { ------------------ | Branch (158:13): [True: 0, False: 0] ------------------ 159| 0| size_t location_end = line_length - offset; 160| 0| size_t to_move = 64 - location_end; 161| 0| if (location_end < 16) { ------------------ | Branch (161:15): [True: 0, False: 0] ------------------ 162| | // We can store or extract store. See below. 163| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t0); 164| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 165| 0| insert_line_feed16(t0, location_end)); 166| 0| out[16] = static_cast(_mm_extract_epi8(t0, 15)); 167| 0| out += 17; 168| 0| } else { 169| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t0); 170| 0| out += 16; 171| 0| } 172| 0| if (location_end >= 16 && location_end < 32) { ------------------ | Branch (172:15): [True: 0, False: 0] | Branch (172:37): [True: 0, False: 0] ------------------ 173| | // We can store or extract store. See below. 174| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t1); 175| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 176| 0| insert_line_feed16(t1, location_end - 16)); 177| 0| out[16] = static_cast(_mm_extract_epi8(t1, 15)); 178| 0| out += 17; 179| 0| } else { 180| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t1); 181| 0| out += 16; 182| 0| } 183| 0| if (location_end >= 32 && location_end < 48) { ------------------ | Branch (183:15): [True: 0, False: 0] | Branch (183:37): [True: 0, False: 0] ------------------ 184| | // We can store or extract store. See below. 185| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t2); 186| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 187| 0| insert_line_feed16(t2, location_end - 32)); 188| 0| out[16] = static_cast(_mm_extract_epi8(t2, 15)); 189| 0| out += 17; 190| 0| } else { 191| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t2); 192| 0| out += 16; 193| 0| } 194| 0| if (location_end >= 48) { ------------------ | Branch (194:15): [True: 0, False: 0] ------------------ 195| | // We can store or extract store. See below. 196| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t3); 197| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 198| 0| insert_line_feed16(t3, location_end - 48)); 199| 0| out[16] = static_cast(_mm_extract_epi8(t3, 15)); 200| 0| out += 17; 201| 0| } else { 202| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t3); 203| 0| out += 16; 204| 0| } 205| 0| offset = to_move; 206| 0| } else { 207| | 208| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t0); 209| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 16), t1); 210| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 32), t2); 211| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 48), t3); 212| 0| offset += 64; 213| 0| out += 64; 214| 0| } 215| 0| } else { // slow path 216| | // could be optimized 217| 0| alignas(64) uint8_t buffer[64]; 218| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer), t0); 219| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + 16), t1); 220| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + 32), t2); 221| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + 48), t3); 222| 0| std::memcpy(out, buffer, 64); 223| 0| size_t out_pos = 0; 224| 0| size_t local_offset = offset; 225| 0| for (size_t j = 0; j < 64;) { ------------------ | Branch (225:28): [True: 0, False: 0] ------------------ 226| 0| if (local_offset == line_length) { ------------------ | Branch (226:15): [True: 0, False: 0] ------------------ 227| 0| out[out_pos++] = '\n'; 228| 0| local_offset = 0; 229| 0| } 230| 0| out[out_pos++] = buffer[j++]; 231| 0| local_offset++; 232| 0| } 233| 0| offset = local_offset; 234| 0| out += out_pos; 235| 0| } 236| 3.13k| } else { 237| 3.13k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t0); 238| 3.13k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 16), t1); 239| 3.13k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 32), t2); 240| 3.13k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 48), t3); 241| 3.13k| out += 64; 242| 3.13k| } 243| 3.13k| } 244| 16.7k| for (; i + 16 <= srclen; i += 12) { ------------------ | Branch (244:10): [True: 1.09k, False: 15.6k] ------------------ 245| | 246| 1.09k| __m128i in = _mm_loadu_si128(reinterpret_cast(input + i)); 247| | 248| | // bytes from groups A, B and C are needed in separate 32-bit lanes 249| | // in = [DDDD|CCCC|BBBB|AAAA] 250| | // 251| | // an input triplet has layout 252| | // [????????|ccdddddd|bbbbcccc|aaaaaabb] 253| | // byte 3 byte 2 byte 1 byte 0 -- byte 3 comes from the next 254| | // triplet 255| | // 256| | // shuffling changes the order of bytes: 1, 0, 2, 1 257| | // [bbbbcccc|ccdddddd|aaaaaabb|bbbbcccc] 258| | // ^^^^ ^^^^^^^^ ^^^^^^^^ ^^^^ 259| | // processed bits 260| 1.09k| in = _mm_shuffle_epi8(in, shuf); 261| | 262| | // unpacking 263| | 264| | // t0 = [0000cccc|cc000000|aaaaaa00|00000000] 265| 1.09k| const __m128i t0 = _mm_and_si128(in, _mm_set1_epi32(0x0fc0fc00)); 266| | // t1 = [00000000|00cccccc|00000000|00aaaaaa] 267| | // (c * (1 << 10), a * (1 << 6)) >> 16 (note: an unsigned 268| | // multiplication) 269| 1.09k| const __m128i t1 = _mm_mulhi_epu16(t0, _mm_set1_epi32(0x04000040)); 270| | 271| | // t2 = [00000000|00dddddd|000000bb|bbbb0000] 272| 1.09k| const __m128i t2 = _mm_and_si128(in, _mm_set1_epi32(0x003f03f0)); 273| | // t3 = [00dddddd|00000000|00bbbbbb|00000000]( 274| | // (d * (1 << 8), b * (1 << 4)) 275| 1.09k| const __m128i t3 = _mm_mullo_epi16(t2, _mm_set1_epi32(0x01000010)); 276| | 277| | // res = [00dddddd|00cccccc|00bbbbbb|00aaaaaa] = t1 | t3 278| 1.09k| const __m128i indices = _mm_or_si128(t1, t3); 279| | 280| 1.09k| const __m128i T0 = lookup_pshufb_improved(indices); 281| | 282| 1.09k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), T0); 283| | 284| 1.09k| if (use_lines) { ------------------ | Branch (284:9): [Folded, False: 1.09k] ------------------ 285| 0| if (line_length >= 16) { // fast path ------------------ | Branch (285:11): [True: 0, False: 0] ------------------ 286| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), T0); 287| 0| if (offset + 16 > line_length) { ------------------ | Branch (287:13): [True: 0, False: 0] ------------------ 288| 0| size_t location_end = line_length - offset; 289| 0| size_t to_move = 16 - location_end; 290| 0| std::memmove(out + location_end + 1, out + location_end, to_move); 291| 0| out[location_end] = '\n'; 292| 0| offset = to_move; 293| 0| out += 16 + 1; 294| 0| } else { 295| 0| offset += 16; 296| 0| out += 16; 297| 0| } 298| 0| } else { // slow path 299| | // could be optimized 300| 0| uint8_t buffer[16]; 301| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer), T0); 302| 0| size_t out_pos = 0; 303| 0| size_t local_offset = offset; 304| 0| for (size_t j = 0; j < 16;) { ------------------ | Branch (304:28): [True: 0, False: 0] ------------------ 305| 0| if (local_offset == line_length) { ------------------ | Branch (305:15): [True: 0, False: 0] ------------------ 306| 0| out[out_pos++] = '\n'; 307| 0| local_offset = 0; 308| 0| } 309| 0| out[out_pos++] = buffer[j++]; 310| 0| local_offset++; 311| 0| } 312| 0| offset = local_offset; 313| 0| out += out_pos; 314| 0| } 315| 1.09k| } else { 316| 1.09k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), T0); 317| 1.09k| out += 16; 318| 1.09k| } 319| 1.09k| } 320| 15.6k| return ((char *)out - (char *)dst) + 321| 15.6k| scalar::base64::tail_encode_base64_impl( 322| 15.6k| (char *)out, src + i, srclen - i, options, line_length, offset); 323| 15.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_122lookup_pshufb_improvedILb0EEEDv2_xS3_: 30| 13.6k|template __m128i lookup_pshufb_improved(const __m128i input) { 31| | // credit: Wojciech Muła 32| | // reduce 0..51 -> 0 33| | // 52..61 -> 1 .. 10 34| | // 62 -> 11 35| | // 63 -> 12 36| 13.6k| __m128i result = _mm_subs_epu8(input, _mm_set1_epi8(51)); 37| | 38| | // distinguish between ranges 0..25 and 26..51: 39| | // 0 .. 25 -> remains 0 40| | // 26 .. 51 -> becomes 13 41| 13.6k| const __m128i less = _mm_cmpgt_epi8(_mm_set1_epi8(26), input); 42| 13.6k| result = _mm_or_si128(result, _mm_and_si128(less, _mm_set1_epi8(13))); 43| | 44| 13.6k| __m128i shift_LUT; 45| 13.6k| if (base64_url) { ------------------ | Branch (45:7): [Folded, False: 13.6k] ------------------ 46| 0| shift_LUT = _mm_setr_epi8('a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 47| 0| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 48| 0| '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0); 49| 13.6k| } else { 50| 13.6k| shift_LUT = _mm_setr_epi8('a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 51| 13.6k| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 52| 13.6k| '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0); 53| 13.6k| } 54| | 55| | // read shift 56| 13.6k| result = _mm_shuffle_epi8(shift_LUT, result); 57| | 58| 13.6k| return _mm_add_epi8(result, input); 59| 13.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_latin1_to_utf16ILNS_10endiannessE0EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.56k| char16_t *utf16_output) { 5| 5.56k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| 15.8k| for (size_t i = 0; i < rounded_len; i += 16) { ------------------ | Branch (6:22): [True: 10.2k, False: 5.56k] ------------------ 7| | // Load 16 Latin1 characters into a 128-bit register 8| 10.2k| __m128i in = 9| 10.2k| _mm_loadu_si128(reinterpret_cast(&latin1_input[i])); 10| 10.2k| __m128i out1 = big_endian ? _mm_unpacklo_epi8(_mm_setzero_si128(), in) ------------------ | Branch (10:20): [Folded, False: 10.2k] ------------------ 11| 10.2k| : _mm_unpacklo_epi8(in, _mm_setzero_si128()); 12| 10.2k| __m128i out2 = big_endian ? _mm_unpackhi_epi8(_mm_setzero_si128(), in) ------------------ | Branch (12:20): [Folded, False: 10.2k] ------------------ 13| 10.2k| : _mm_unpackhi_epi8(in, _mm_setzero_si128()); 14| | // Zero extend each Latin1 character to 16-bit integers and store the 15| | // results back to memory 16| 10.2k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i]), out1); 17| 10.2k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i + 8]), out2); 18| 10.2k| } 19| | // return pointers pointing to where we left off 20| 5.56k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 21| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_latin1_to_utf16ILNS_10endiannessE1EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.56k| char16_t *utf16_output) { 5| 5.56k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| 15.8k| for (size_t i = 0; i < rounded_len; i += 16) { ------------------ | Branch (6:22): [True: 10.2k, False: 5.56k] ------------------ 7| | // Load 16 Latin1 characters into a 128-bit register 8| 10.2k| __m128i in = 9| 10.2k| _mm_loadu_si128(reinterpret_cast(&latin1_input[i])); 10| 10.2k| __m128i out1 = big_endian ? _mm_unpacklo_epi8(_mm_setzero_si128(), in) ------------------ | Branch (10:20): [True: 10.2k, Folded] ------------------ 11| 10.2k| : _mm_unpacklo_epi8(in, _mm_setzero_si128()); 12| 10.2k| __m128i out2 = big_endian ? _mm_unpackhi_epi8(_mm_setzero_si128(), in) ------------------ | Branch (12:20): [True: 10.2k, Folded] ------------------ 13| 10.2k| : _mm_unpackhi_epi8(in, _mm_setzero_si128()); 14| | // Zero extend each Latin1 character to 16-bit integers and store the 15| | // results back to memory 16| 10.2k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i]), out1); 17| 10.2k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i + 8]), out2); 18| 10.2k| } 19| | // return pointers pointing to where we left off 20| 5.56k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 21| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_latin1_to_utf32EPKcmPDi: 3| 5.56k| char32_t *utf32_output) { 4| 5.56k| const char *end = buf + len; 5| | 6| 15.8k| while (end - buf >= 16) { ------------------ | Branch (6:10): [True: 10.2k, False: 5.56k] ------------------ 7| | // Load 16 Latin1 characters (16 bytes) into a 128-bit register 8| 10.2k| __m128i in = _mm_loadu_si128((__m128i *)buf); 9| | 10| | // Shift input to process next 4 bytes 11| 10.2k| __m128i in_shifted1 = _mm_srli_si128(in, 4); 12| 10.2k| __m128i in_shifted2 = _mm_srli_si128(in, 8); 13| 10.2k| __m128i in_shifted3 = _mm_srli_si128(in, 12); 14| | 15| | // expand 8-bit to 32-bit unit 16| 10.2k| __m128i out1 = _mm_cvtepu8_epi32(in); 17| 10.2k| __m128i out2 = _mm_cvtepu8_epi32(in_shifted1); 18| 10.2k| __m128i out3 = _mm_cvtepu8_epi32(in_shifted2); 19| 10.2k| __m128i out4 = _mm_cvtepu8_epi32(in_shifted3); 20| | 21| 10.2k| _mm_storeu_si128((__m128i *)utf32_output, out1); 22| 10.2k| _mm_storeu_si128((__m128i *)(utf32_output + 4), out2); 23| 10.2k| _mm_storeu_si128((__m128i *)(utf32_output + 8), out3); 24| 10.2k| _mm_storeu_si128((__m128i *)(utf32_output + 12), out4); 25| | 26| 10.2k| utf32_output += 16; 27| 10.2k| buf += 16; 28| 10.2k| } 29| | 30| 5.56k| return std::make_pair(buf, utf32_output); 31| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_126sse_convert_latin1_to_utf8EPKcmPc: 3| 5.86k| const size_t latin_input_length, char *utf8_output) { 4| 5.86k| const char *end = latin_input + latin_input_length; 5| | 6| 5.86k| const __m128i v_0000 = _mm_setzero_si128(); 7| | // 0b1000_0000 8| 5.86k| const __m128i v_80 = _mm_set1_epi8((uint8_t)0x80); 9| | // 0b1111_1111_1000_0000 10| 5.86k| const __m128i v_ff80 = _mm_set1_epi16((uint16_t)0xff80); 11| | 12| 5.86k| const __m128i latin_1_half_into_u16_byte_mask = 13| 5.86k| _mm_setr_epi8(0, '\x80', 1, '\x80', 2, '\x80', 3, '\x80', 4, '\x80', 5, 14| 5.86k| '\x80', 6, '\x80', 7, '\x80'); 15| | 16| 5.86k| const __m128i latin_2_half_into_u16_byte_mask = 17| 5.86k| _mm_setr_epi8(8, '\x80', 9, '\x80', 10, '\x80', 11, '\x80', 12, '\x80', 18| 5.86k| 13, '\x80', 14, '\x80', 15, '\x80'); 19| | 20| | // each latin1 takes 1-2 utf8 bytes 21| | // slow path writes useful 8-15 bytes twice (eagerly writes 16 bytes and then 22| | // adjust the pointer) so the last write can exceed the utf8_output size by 23| | // 8-1 bytes by reserving 8 extra input bytes, we expect the output to have 24| | // 8-16 bytes free 25| 18.7k| while (end - latin_input >= 16 + 8) { ------------------ | Branch (25:10): [True: 12.9k, False: 5.86k] ------------------ 26| | // Load 16 Latin1 characters (16 bytes) into a 128-bit register 27| 12.9k| __m128i v_latin = _mm_loadu_si128((__m128i *)latin_input); 28| | 29| 12.9k| if (_mm_testz_si128(v_latin, v_80)) { // ASCII fast path!!!! ------------------ | Branch (29:9): [True: 8.81k, False: 4.11k] ------------------ 30| 8.81k| _mm_storeu_si128((__m128i *)utf8_output, v_latin); 31| 8.81k| latin_input += 16; 32| 8.81k| utf8_output += 16; 33| 8.81k| continue; 34| 8.81k| } 35| | 36| | // assuming a/b are bytes and A/B are uint16 of the same value 37| | // aaaa_aaaa_bbbb_bbbb -> AAAA_AAAA 38| 4.11k| __m128i v_u16_latin_1_half = 39| 4.11k| _mm_shuffle_epi8(v_latin, latin_1_half_into_u16_byte_mask); 40| | // aaaa_aaaa_bbbb_bbbb -> BBBB_BBBB 41| 4.11k| __m128i v_u16_latin_2_half = 42| 4.11k| _mm_shuffle_epi8(v_latin, latin_2_half_into_u16_byte_mask); 43| | 44| 4.11k| internal::westmere::write_v_u16_11bits_to_utf8(v_u16_latin_1_half, 45| 4.11k| utf8_output, v_0000, v_ff80); 46| 4.11k| internal::westmere::write_v_u16_11bits_to_utf8(v_u16_latin_2_half, 47| 4.11k| utf8_output, v_0000, v_ff80); 48| 4.11k| latin_input += 16; 49| 4.11k| } 50| | 51| 5.86k| if (end - latin_input >= 16) { ------------------ | Branch (51:7): [True: 638, False: 5.22k] ------------------ 52| | // Load 16 Latin1 characters (16 bytes) into a 128-bit register 53| 638| __m128i v_latin = _mm_loadu_si128((__m128i *)latin_input); 54| | 55| 638| if (_mm_testz_si128(v_latin, v_80)) { // ASCII fast path!!!! ------------------ | Branch (55:9): [True: 419, False: 219] ------------------ 56| 419| _mm_storeu_si128((__m128i *)utf8_output, v_latin); 57| 419| latin_input += 16; 58| 419| utf8_output += 16; 59| 419| } else { 60| | // assuming a/b are bytes and A/B are uint16 of the same value 61| | // aaaa_aaaa_bbbb_bbbb -> AAAA_AAAA 62| 219| __m128i v_u16_latin_1_half = 63| 219| _mm_shuffle_epi8(v_latin, latin_1_half_into_u16_byte_mask); 64| 219| internal::westmere::write_v_u16_11bits_to_utf8( 65| 219| v_u16_latin_1_half, utf8_output, v_0000, v_ff80); 66| 219| latin_input += 8; 67| 219| } 68| 638| } 69| | 70| 5.86k| return std::make_pair(latin_input, utf8_output); 71| 5.86k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_utf16_to_latin1ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.56k| char *latin1_output) { 5| 5.56k| const char16_t *end = buf + len; 6| 26.2k| while (end - buf >= 8) { ------------------ | Branch (6:10): [True: 20.7k, False: 5.56k] ------------------ 7| | // Load 8 UTF-16 characters into 128-bit SSE register 8| 20.7k| __m128i in = _mm_loadu_si128(reinterpret_cast(buf)); 9| | 10| | if constexpr (!match_system(big_endian)) { 11| | const __m128i swap = 12| | _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 13| | in = _mm_shuffle_epi8(in, swap); 14| | } 15| | 16| 20.7k| __m128i high_byte_mask = _mm_set1_epi16((int16_t)0xFF00); 17| 20.7k| if (_mm_testz_si128(in, high_byte_mask)) { ------------------ | Branch (17:9): [True: 20.7k, False: 0] ------------------ 18| | // Pack 16-bit characters into 8-bit and store in latin1_output 19| 20.7k| __m128i latin1_packed = _mm_packus_epi16(in, in); 20| 20.7k| _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output), 21| 20.7k| latin1_packed); 22| | // Adjust pointers for next iteration 23| 20.7k| buf += 8; 24| 20.7k| latin1_output += 8; 25| 20.7k| } else { 26| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 27| 0| } 28| 20.7k| } // while 29| 5.56k| return std::make_pair(buf, latin1_output); 30| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_utf16_to_latin1ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.56k| char *latin1_output) { 5| 5.56k| const char16_t *end = buf + len; 6| 26.2k| while (end - buf >= 8) { ------------------ | Branch (6:10): [True: 20.7k, False: 5.56k] ------------------ 7| | // Load 8 UTF-16 characters into 128-bit SSE register 8| 20.7k| __m128i in = _mm_loadu_si128(reinterpret_cast(buf)); 9| | 10| 20.7k| if constexpr (!match_system(big_endian)) { 11| 20.7k| const __m128i swap = 12| 20.7k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 13| 20.7k| in = _mm_shuffle_epi8(in, swap); 14| 20.7k| } 15| | 16| 20.7k| __m128i high_byte_mask = _mm_set1_epi16((int16_t)0xFF00); 17| 20.7k| if (_mm_testz_si128(in, high_byte_mask)) { ------------------ | Branch (17:9): [True: 20.7k, False: 0] ------------------ 18| | // Pack 16-bit characters into 8-bit and store in latin1_output 19| 20.7k| __m128i latin1_packed = _mm_packus_epi16(in, in); 20| 20.7k| _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output), 21| 20.7k| latin1_packed); 22| | // Adjust pointers for next iteration 23| 20.7k| buf += 8; 24| 20.7k| latin1_output += 8; 25| 20.7k| } else { 26| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 27| 0| } 28| 20.7k| } // while 29| 5.56k| return std::make_pair(buf, latin1_output); 30| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_125sse_convert_utf16_to_utf8ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 10.7k|sse_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| | 58| 10.7k| const char16_t *end = buf + len; 59| | 60| 10.7k| const __m128i v_0000 = _mm_setzero_si128(); 61| 10.7k| const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800); 62| 10.7k| const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800); 63| 10.7k| const size_t safety_margin = 64| 10.7k| 12; // to avoid overruns, see issue 65| | // https://github.com/simdutf/simdutf/issues/92 66| | 67| 20.4k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (67:10): [True: 9.85k, False: 10.6k] ------------------ 68| 9.85k| __m128i in = _mm_loadu_si128((__m128i *)buf); 69| 9.85k| if (big_endian) { ------------------ | Branch (69:9): [Folded, False: 9.85k] ------------------ 70| 0| const __m128i swap = 71| 0| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 72| 0| in = _mm_shuffle_epi8(in, swap); 73| 0| } 74| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 75| 9.85k| const __m128i v_ff80 = _mm_set1_epi16((int16_t)0xff80); 76| 9.85k| if (_mm_testz_si128(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 3.64k, False: 6.20k] ------------------ 77| 3.64k| __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1); 78| 3.64k| if (big_endian) { ------------------ | Branch (78:11): [Folded, False: 3.64k] ------------------ 79| 0| const __m128i swap = 80| 0| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 81| 0| nextin = _mm_shuffle_epi8(nextin, swap); 82| 0| } 83| 3.64k| if (!_mm_testz_si128(nextin, v_ff80)) { ------------------ | Branch (83:11): [True: 422, False: 3.22k] ------------------ 84| | // 1. pack the bytes 85| | // obviously suboptimal. 86| 422| const __m128i utf8_packed = _mm_packus_epi16(in, in); 87| | // 2. store (16 bytes) 88| 422| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 89| | // 3. adjust pointers 90| 422| buf += 8; 91| 422| utf8_output += 8; 92| 422| in = nextin; 93| 3.22k| } else { 94| | // 1. pack the bytes 95| | // obviously suboptimal. 96| 3.22k| const __m128i utf8_packed = _mm_packus_epi16(in, nextin); 97| | // 2. store (16 bytes) 98| 3.22k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 99| | // 3. adjust pointers 100| 3.22k| buf += 16; 101| 3.22k| utf8_output += 16; 102| 3.22k| continue; // we are done for this round! 103| 3.22k| } 104| 3.64k| } 105| | 106| | // no bits set above 7th bit 107| 6.63k| const __m128i one_byte_bytemask = 108| 6.63k| _mm_cmpeq_epi16(_mm_and_si128(in, v_ff80), v_0000); 109| 6.63k| const uint16_t one_byte_bitmask = 110| 6.63k| static_cast(_mm_movemask_epi8(one_byte_bytemask)); 111| | 112| | // no bits set above 11th bit 113| 6.63k| const __m128i one_or_two_bytes_bytemask = 114| 6.63k| _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_0000); 115| 6.63k| const uint16_t one_or_two_bytes_bitmask = 116| 6.63k| static_cast(_mm_movemask_epi8(one_or_two_bytes_bytemask)); 117| | 118| 6.63k| if (one_or_two_bytes_bitmask == 0xffff) { ------------------ | Branch (118:9): [True: 1.91k, False: 4.71k] ------------------ 119| 1.91k| internal::westmere::write_v_u16_11bits_to_utf8( 120| 1.91k| in, utf8_output, one_byte_bytemask, one_byte_bitmask); 121| 1.91k| buf += 8; 122| 1.91k| continue; 123| 1.91k| } 124| | 125| | // 1. Check if there are any surrogate word in the input chunk. 126| | // We have also deal with situation when there is a surrogate word 127| | // at the end of a chunk. 128| 4.71k| const __m128i surrogates_bytemask = 129| 4.71k| _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_d800); 130| | 131| | // bitmask = 0x0000 if there are no surrogates 132| | // = 0xc000 if the last word is a surrogate 133| 4.71k| const uint16_t surrogates_bitmask = 134| 4.71k| static_cast(_mm_movemask_epi8(surrogates_bytemask)); 135| | // It might seem like checking for surrogates_bitmask == 0xc000 could help. 136| | // However, it is likely an uncommon occurrence. 137| 4.71k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (137:9): [True: 3.56k, False: 1.14k] ------------------ 138| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 139| 3.56k| const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606, 140| 3.56k| 0x0808, 0x0a0a, 0x0c0c, 0x0e0e); 141| | 142| | /* In this branch we handle three cases: 143| | 1. [0000|0000|0ccc|cccc] => [0ccc|cccc] - 144| | single UFT-8 byte 145| | 2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc] - two 146| | UTF-8 bytes 147| | 3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - 148| | three UTF-8 bytes 149| | 150| | We expand the input word (16-bit) into two code units (32-bit), thus 151| | we have room for four bytes. However, we need five distinct bit 152| | layouts. Note that the last byte in cases #2 and #3 is the same. 153| | 154| | We precompute byte 1 for case #1 and the common byte for cases #2 & #3 155| | in register t2. 156| | 157| | We precompute byte 1 for case #3 and -- **conditionally** -- precompute 158| | either byte 1 for case #2 or byte 2 for case #3. Note that they 159| | differ by exactly one bit. 160| | 161| | Finally from these two code units we build proper UTF-8 sequence, taking 162| | into account the case (i.e, the number of bytes to write). 163| | */ 164| | /** 165| | * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce: 166| | * t2 => [0ccc|cccc] [10cc|cccc] 167| | * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb]) 168| | */ 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) 170| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 171| 3.56k| const __m128i t0 = _mm_shuffle_epi8(in, dup_even); 172| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 173| 3.56k| const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 174| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 175| 3.56k| const __m128i t2 = _mm_or_si128(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 176| | 177| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 178| 3.56k| const __m128i s0 = _mm_srli_epi16(in, 4); 179| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 180| 3.56k| const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 181| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 182| 3.56k| const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 183| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 184| 3.56k| const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 185| 3.56k| const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, 186| 3.56k| simdutf_vec(0b0100000000000000)); ------------------ | | 169| 3.56k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 187| 3.56k| const __m128i s4 = _mm_xor_si128(s3, m0); 188| 3.56k|#undef simdutf_vec 189| | 190| | // 4. expand code units 16-bit => 32-bit 191| 3.56k| const __m128i out0 = _mm_unpacklo_epi16(t2, s4); 192| 3.56k| const __m128i out1 = _mm_unpackhi_epi16(t2, s4); 193| | 194| | // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle 195| 3.56k| const uint16_t mask = 196| 3.56k| (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa); 197| 3.56k| if (mask == 0) { ------------------ | Branch (197:11): [True: 2.63k, False: 931] ------------------ 198| | // We only have three-byte code units. Use fast path. 199| 2.63k| const __m128i shuffle = _mm_setr_epi8(2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 200| 2.63k| 15, 13, -1, -1, -1, -1); 201| 2.63k| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle); 202| 2.63k| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle); 203| 2.63k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 204| 2.63k| utf8_output += 12; 205| 2.63k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 206| 2.63k| utf8_output += 12; 207| 2.63k| buf += 8; 208| 2.63k| continue; 209| 2.63k| } 210| 931| const uint8_t mask0 = uint8_t(mask); 211| | 212| 931| const uint8_t *row0 = 213| 931| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 214| 931| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 215| 931| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0); 216| | 217| 931| const uint8_t mask1 = static_cast(mask >> 8); 218| | 219| 931| const uint8_t *row1 = 220| 931| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 221| 931| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 222| 931| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1); 223| | 224| 931| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 225| 931| utf8_output += row0[0]; 226| 931| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 227| 931| utf8_output += row1[0]; 228| | 229| 931| buf += 8; 230| | // surrogate pair(s) in a register 231| 1.14k| } else { 232| | // Let us do a scalar fallback. 233| | // It may seem wasteful to use scalar code, but being efficient with SIMD 234| | // in the presence of surrogate pairs may require non-trivial tables. 235| 1.14k| size_t forward = 15; 236| 1.14k| size_t k = 0; 237| 1.14k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (237:11): [True: 0, False: 1.14k] ------------------ 238| 0| forward = size_t(end - buf - 1); 239| 0| } 240| 15.8k| for (; k < forward; k++) { ------------------ | Branch (240:14): [True: 14.8k, False: 1.08k] ------------------ 241| 14.8k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 242| 14.8k| if ((word & 0xFF80) == 0) { ------------------ | Branch (242:13): [True: 8.79k, False: 6.00k] ------------------ 243| 8.79k| *utf8_output++ = char(word); 244| 8.79k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (244:20): [True: 560, False: 5.44k] ------------------ 245| 560| *utf8_output++ = char((word >> 6) | 0b11000000); 246| 560| *utf8_output++ = char((word & 0b111111) | 0b10000000); 247| 5.44k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (247:20): [True: 3.61k, False: 1.82k] ------------------ 248| 3.61k| *utf8_output++ = char((word >> 12) | 0b11100000); 249| 3.61k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 250| 3.61k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 251| 3.61k| } else { 252| | // must be a surrogate pair 253| 1.82k| uint16_t diff = uint16_t(word - 0xD800); 254| 1.82k| uint16_t next_word = 255| 1.82k| scalar::utf16::swap_if_needed(buf[k + 1]); 256| 1.82k| k++; 257| 1.82k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 258| 1.82k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (258:15): [True: 59, False: 1.77k] ------------------ 259| 59| return std::make_pair(nullptr, utf8_output); 260| 59| } 261| 1.77k| uint32_t value = (diff << 10) + diff2 + 0x10000; 262| 1.77k| *utf8_output++ = char((value >> 18) | 0b11110000); 263| 1.77k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 264| 1.77k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 265| 1.77k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 266| 1.77k| } 267| 14.8k| } 268| 1.08k| buf += k; 269| 1.08k| } 270| 4.71k| } // while 271| | 272| 10.6k| return std::make_pair(buf, utf8_output); 273| 10.7k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_125sse_convert_utf16_to_utf8ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 10.7k|sse_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| | 58| 10.7k| const char16_t *end = buf + len; 59| | 60| 10.7k| const __m128i v_0000 = _mm_setzero_si128(); 61| 10.7k| const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800); 62| 10.7k| const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800); 63| 10.7k| const size_t safety_margin = 64| 10.7k| 12; // to avoid overruns, see issue 65| | // https://github.com/simdutf/simdutf/issues/92 66| | 67| 20.7k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (67:10): [True: 10.0k, False: 10.6k] ------------------ 68| 10.0k| __m128i in = _mm_loadu_si128((__m128i *)buf); 69| 10.0k| if (big_endian) { ------------------ | Branch (69:9): [True: 10.0k, Folded] ------------------ 70| 10.0k| const __m128i swap = 71| 10.0k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 72| 10.0k| in = _mm_shuffle_epi8(in, swap); 73| 10.0k| } 74| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 75| 10.0k| const __m128i v_ff80 = _mm_set1_epi16((int16_t)0xff80); 76| 10.0k| if (_mm_testz_si128(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 3.61k, False: 6.43k] ------------------ 77| 3.61k| __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1); 78| 3.61k| if (big_endian) { ------------------ | Branch (78:11): [True: 3.61k, Folded] ------------------ 79| 3.61k| const __m128i swap = 80| 3.61k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 81| 3.61k| nextin = _mm_shuffle_epi8(nextin, swap); 82| 3.61k| } 83| 3.61k| if (!_mm_testz_si128(nextin, v_ff80)) { ------------------ | Branch (83:11): [True: 407, False: 3.21k] ------------------ 84| | // 1. pack the bytes 85| | // obviously suboptimal. 86| 407| const __m128i utf8_packed = _mm_packus_epi16(in, in); 87| | // 2. store (16 bytes) 88| 407| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 89| | // 3. adjust pointers 90| 407| buf += 8; 91| 407| utf8_output += 8; 92| 407| in = nextin; 93| 3.21k| } else { 94| | // 1. pack the bytes 95| | // obviously suboptimal. 96| 3.21k| const __m128i utf8_packed = _mm_packus_epi16(in, nextin); 97| | // 2. store (16 bytes) 98| 3.21k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 99| | // 3. adjust pointers 100| 3.21k| buf += 16; 101| 3.21k| utf8_output += 16; 102| 3.21k| continue; // we are done for this round! 103| 3.21k| } 104| 3.61k| } 105| | 106| | // no bits set above 7th bit 107| 6.84k| const __m128i one_byte_bytemask = 108| 6.84k| _mm_cmpeq_epi16(_mm_and_si128(in, v_ff80), v_0000); 109| 6.84k| const uint16_t one_byte_bitmask = 110| 6.84k| static_cast(_mm_movemask_epi8(one_byte_bytemask)); 111| | 112| | // no bits set above 11th bit 113| 6.84k| const __m128i one_or_two_bytes_bytemask = 114| 6.84k| _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_0000); 115| 6.84k| const uint16_t one_or_two_bytes_bitmask = 116| 6.84k| static_cast(_mm_movemask_epi8(one_or_two_bytes_bytemask)); 117| | 118| 6.84k| if (one_or_two_bytes_bitmask == 0xffff) { ------------------ | Branch (118:9): [True: 1.91k, False: 4.92k] ------------------ 119| 1.91k| internal::westmere::write_v_u16_11bits_to_utf8( 120| 1.91k| in, utf8_output, one_byte_bytemask, one_byte_bitmask); 121| 1.91k| buf += 8; 122| 1.91k| continue; 123| 1.91k| } 124| | 125| | // 1. Check if there are any surrogate word in the input chunk. 126| | // We have also deal with situation when there is a surrogate word 127| | // at the end of a chunk. 128| 4.92k| const __m128i surrogates_bytemask = 129| 4.92k| _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_d800); 130| | 131| | // bitmask = 0x0000 if there are no surrogates 132| | // = 0xc000 if the last word is a surrogate 133| 4.92k| const uint16_t surrogates_bitmask = 134| 4.92k| static_cast(_mm_movemask_epi8(surrogates_bytemask)); 135| | // It might seem like checking for surrogates_bitmask == 0xc000 could help. 136| | // However, it is likely an uncommon occurrence. 137| 4.92k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (137:9): [True: 3.81k, False: 1.10k] ------------------ 138| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 139| 3.81k| const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606, 140| 3.81k| 0x0808, 0x0a0a, 0x0c0c, 0x0e0e); 141| | 142| | /* In this branch we handle three cases: 143| | 1. [0000|0000|0ccc|cccc] => [0ccc|cccc] - 144| | single UFT-8 byte 145| | 2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc] - two 146| | UTF-8 bytes 147| | 3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - 148| | three UTF-8 bytes 149| | 150| | We expand the input word (16-bit) into two code units (32-bit), thus 151| | we have room for four bytes. However, we need five distinct bit 152| | layouts. Note that the last byte in cases #2 and #3 is the same. 153| | 154| | We precompute byte 1 for case #1 and the common byte for cases #2 & #3 155| | in register t2. 156| | 157| | We precompute byte 1 for case #3 and -- **conditionally** -- precompute 158| | either byte 1 for case #2 or byte 2 for case #3. Note that they 159| | differ by exactly one bit. 160| | 161| | Finally from these two code units we build proper UTF-8 sequence, taking 162| | into account the case (i.e, the number of bytes to write). 163| | */ 164| | /** 165| | * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce: 166| | * t2 => [0ccc|cccc] [10cc|cccc] 167| | * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb]) 168| | */ 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) 170| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 171| 3.81k| const __m128i t0 = _mm_shuffle_epi8(in, dup_even); 172| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 173| 3.81k| const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 174| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 175| 3.81k| const __m128i t2 = _mm_or_si128(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 176| | 177| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 178| 3.81k| const __m128i s0 = _mm_srli_epi16(in, 4); 179| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 180| 3.81k| const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 181| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 182| 3.81k| const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 183| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 184| 3.81k| const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 185| 3.81k| const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, 186| 3.81k| simdutf_vec(0b0100000000000000)); ------------------ | | 169| 3.81k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 187| 3.81k| const __m128i s4 = _mm_xor_si128(s3, m0); 188| 3.81k|#undef simdutf_vec 189| | 190| | // 4. expand code units 16-bit => 32-bit 191| 3.81k| const __m128i out0 = _mm_unpacklo_epi16(t2, s4); 192| 3.81k| const __m128i out1 = _mm_unpackhi_epi16(t2, s4); 193| | 194| | // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle 195| 3.81k| const uint16_t mask = 196| 3.81k| (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa); 197| 3.81k| if (mask == 0) { ------------------ | Branch (197:11): [True: 2.91k, False: 904] ------------------ 198| | // We only have three-byte code units. Use fast path. 199| 2.91k| const __m128i shuffle = _mm_setr_epi8(2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 200| 2.91k| 15, 13, -1, -1, -1, -1); 201| 2.91k| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle); 202| 2.91k| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle); 203| 2.91k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 204| 2.91k| utf8_output += 12; 205| 2.91k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 206| 2.91k| utf8_output += 12; 207| 2.91k| buf += 8; 208| 2.91k| continue; 209| 2.91k| } 210| 904| const uint8_t mask0 = uint8_t(mask); 211| | 212| 904| const uint8_t *row0 = 213| 904| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 214| 904| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 215| 904| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0); 216| | 217| 904| const uint8_t mask1 = static_cast(mask >> 8); 218| | 219| 904| const uint8_t *row1 = 220| 904| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 221| 904| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 222| 904| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1); 223| | 224| 904| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 225| 904| utf8_output += row0[0]; 226| 904| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 227| 904| utf8_output += row1[0]; 228| | 229| 904| buf += 8; 230| | // surrogate pair(s) in a register 231| 1.10k| } else { 232| | // Let us do a scalar fallback. 233| | // It may seem wasteful to use scalar code, but being efficient with SIMD 234| | // in the presence of surrogate pairs may require non-trivial tables. 235| 1.10k| size_t forward = 15; 236| 1.10k| size_t k = 0; 237| 1.10k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (237:11): [True: 0, False: 1.10k] ------------------ 238| 0| forward = size_t(end - buf - 1); 239| 0| } 240| 15.4k| for (; k < forward; k++) { ------------------ | Branch (240:14): [True: 14.4k, False: 1.05k] ------------------ 241| 14.4k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 242| 14.4k| if ((word & 0xFF80) == 0) { ------------------ | Branch (242:13): [True: 8.77k, False: 5.65k] ------------------ 243| 8.77k| *utf8_output++ = char(word); 244| 8.77k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (244:20): [True: 612, False: 5.04k] ------------------ 245| 612| *utf8_output++ = char((word >> 6) | 0b11000000); 246| 612| *utf8_output++ = char((word & 0b111111) | 0b10000000); 247| 5.04k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (247:20): [True: 3.34k, False: 1.69k] ------------------ 248| 3.34k| *utf8_output++ = char((word >> 12) | 0b11100000); 249| 3.34k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 250| 3.34k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 251| 3.34k| } else { 252| | // must be a surrogate pair 253| 1.69k| uint16_t diff = uint16_t(word - 0xD800); 254| 1.69k| uint16_t next_word = 255| 1.69k| scalar::utf16::swap_if_needed(buf[k + 1]); 256| 1.69k| k++; 257| 1.69k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 258| 1.69k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (258:15): [True: 56, False: 1.63k] ------------------ 259| 56| return std::make_pair(nullptr, utf8_output); 260| 56| } 261| 1.63k| uint32_t value = (diff << 10) + diff2 + 0x10000; 262| 1.63k| *utf8_output++ = char((value >> 18) | 0b11110000); 263| 1.63k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 264| 1.63k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 265| 1.63k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 266| 1.63k| } 267| 14.4k| } 268| 1.05k| buf += k; 269| 1.05k| } 270| 4.92k| } // while 271| | 272| 10.6k| return std::make_pair(buf, utf8_output); 273| 10.7k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_utf32_to_latin1EPKDimPc: 3| 5.56k| char *latin1_output) { 4| 5.56k| const size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 5| | 6| 5.56k| __m128i high_bytes_mask = _mm_set1_epi32(0xFFFFFF00); 7| 5.56k| __m128i shufmask = 8| 5.56k| _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 12, 8, 4, 0); 9| | 10| 15.8k| for (size_t i = 0; i < rounded_len; i += 16) { ------------------ | Branch (10:22): [True: 10.2k, False: 5.56k] ------------------ 11| 10.2k| __m128i in1 = _mm_loadu_si128((__m128i *)buf); 12| 10.2k| __m128i in2 = _mm_loadu_si128((__m128i *)(buf + 4)); 13| 10.2k| __m128i in3 = _mm_loadu_si128((__m128i *)(buf + 8)); 14| 10.2k| __m128i in4 = _mm_loadu_si128((__m128i *)(buf + 12)); 15| | 16| 10.2k| __m128i check_combined = _mm_or_si128(in1, in2); 17| 10.2k| check_combined = _mm_or_si128(check_combined, in3); 18| 10.2k| check_combined = _mm_or_si128(check_combined, in4); 19| | 20| 10.2k| if (!_mm_testz_si128(check_combined, high_bytes_mask)) { ------------------ | Branch (20:9): [True: 0, False: 10.2k] ------------------ 21| 0| return std::make_pair(nullptr, latin1_output); 22| 0| } 23| 10.2k| __m128i pack1 = _mm_unpacklo_epi32(_mm_shuffle_epi8(in1, shufmask), 24| 10.2k| _mm_shuffle_epi8(in2, shufmask)); 25| 10.2k| __m128i pack2 = _mm_unpacklo_epi32(_mm_shuffle_epi8(in3, shufmask), 26| 10.2k| _mm_shuffle_epi8(in4, shufmask)); 27| 10.2k| __m128i pack = _mm_unpacklo_epi64(pack1, pack2); 28| 10.2k| _mm_storeu_si128((__m128i *)latin1_output, pack); 29| 10.2k| latin1_output += 16; 30| 10.2k| buf += 16; 31| 10.2k| } 32| | 33| 5.56k| return std::make_pair(buf, latin1_output); 34| 5.56k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_125sse_convert_utf32_to_utf8EPKDimPc: 2| 5.14k|sse_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) { 3| 5.14k| const char32_t *end = buf + len; 4| | 5| 5.14k| const __m128i v_0000 = _mm_setzero_si128(); //__m128 = 128 bits 6| 5.14k| const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800); // 1111 1000 0000 7| | // 0000 8| 5.14k| const __m128i v_c080 = _mm_set1_epi16((uint16_t)0xc080); // 1100 0000 1000 9| | // 0000 10| 5.14k| const __m128i v_ff80 = _mm_set1_epi16((uint16_t)0xff80); // 1111 1111 1000 11| | // 0000 12| 5.14k| const __m128i v_ffff0000 = _mm_set1_epi32( 13| 5.14k| (uint32_t)0xffff0000); // 1111 1111 1111 1111 0000 0000 0000 0000 14| 5.14k| const __m128i v_7fffffff = _mm_set1_epi32( 15| 5.14k| (uint32_t)0x7fffffff); // 0111 1111 1111 1111 1111 1111 1111 1111 16| 5.14k| __m128i running_max = _mm_setzero_si128(); 17| 5.14k| __m128i forbidden_bytemask = _mm_setzero_si128(); 18| 5.14k| const size_t safety_margin = 19| 5.14k| 12; // to avoid overruns, see issue 20| | // https://github.com/simdutf/simdutf/issues/92 21| | 22| 10.8k| while (end - buf >= ------------------ | Branch (22:10): [True: 5.71k, False: 5.14k] ------------------ 23| 10.8k| std::ptrdiff_t( 24| 10.8k| 16 + safety_margin)) { // buf is a char32_t pointer, each char32_t 25| | // has 4 bytes or 32 bits, thus buf + 16 * 26| | // char_32t = 512 bits = 64 bytes 27| | // We load two 16 bytes registers for a total of 32 bytes or 16 characters. 28| 5.71k| __m128i in = _mm_loadu_si128((__m128i *)buf); 29| 5.71k| __m128i nextin = _mm_loadu_si128( 30| 5.71k| (__m128i *)buf + 1); // These two values can hold only 8 UTF32 chars 31| 5.71k| running_max = _mm_max_epu32( 32| 5.71k| _mm_max_epu32(in, running_max), // take element-wise max char32_t from 33| | // in and running_max vector 34| 5.71k| nextin); // and take element-wise max element from nextin and 35| | // running_max vector 36| | 37| | // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned 38| | // saturation 39| 5.71k| __m128i in_16 = _mm_packus_epi32( 40| 5.71k| _mm_and_si128(in, v_7fffffff), 41| 5.71k| _mm_and_si128( 42| 5.71k| nextin, 43| 5.71k| v_7fffffff)); // in this context pack the two __m128 into a single 44| | // By ensuring the highest bit is set to 0(&v_7fffffff), we are making sure 45| | // all values are interpreted as non-negative, or specifically, the values 46| | // are within the range of valid Unicode code points. remember : having 47| | // leading byte 0 means a positive number by the two complements system. 48| | // Unicode is well beneath the range where you'll start getting issues so 49| | // that's OK. 50| | 51| | // Try to apply UTF-16 => UTF-8 from ./sse_convert_utf16_to_utf8.cpp 52| | 53| | // Check for ASCII fast path 54| | 55| | // ASCII fast path!!!! 56| | // We eagerly load another 32 bytes, hoping that they will be ASCII too. 57| | // The intuition is that we try to collect 16 ASCII characters which 58| | // requires a total of 64 bytes of input. If we fail, we just pass thirdin 59| | // and fourthin as our new inputs. 60| 5.71k| if (_mm_testz_si128(in_16, v_ff80)) { // if the first two blocks are ASCII ------------------ | Branch (60:9): [True: 3.45k, False: 2.26k] ------------------ 61| 3.45k| __m128i thirdin = _mm_loadu_si128((__m128i *)buf + 2); 62| 3.45k| __m128i fourthin = _mm_loadu_si128((__m128i *)buf + 3); 63| 3.45k| running_max = _mm_max_epu32( 64| 3.45k| _mm_max_epu32(thirdin, running_max), 65| 3.45k| fourthin); // take the running max of all 4 vectors thus far 66| 3.45k| __m128i nextin_16 = _mm_packus_epi32( 67| 3.45k| _mm_and_si128(thirdin, v_7fffffff), 68| 3.45k| _mm_and_si128(fourthin, 69| 3.45k| v_7fffffff)); // pack into 1 vector, now you have two 70| 3.45k| if (!_mm_testz_si128( ------------------ | Branch (70:11): [True: 377, False: 3.07k] ------------------ 71| 3.45k| nextin_16, 72| 3.45k| v_ff80)) { // checks if the second packed vector is ASCII, if not: 73| | // 1. pack the bytes 74| | // obviously suboptimal. 75| 377| const __m128i utf8_packed = _mm_packus_epi16( 76| 377| in_16, in_16); // creates two copy of in_16 in 1 vector 77| | // 2. store (16 bytes) 78| 377| _mm_storeu_si128((__m128i *)utf8_output, 79| 377| utf8_packed); // put them into the output 80| | // 3. adjust pointers 81| 377| buf += 8; // the char32_t buffer pointer goes up 8 char32_t chars* 32 82| | // bits = 256 bits 83| 377| utf8_output += 84| 377| 8; // same with output, e.g. lift the first two blocks alone. 85| | // Proceed with next input 86| 377| in_16 = nextin_16; 87| | // We need to update in and nextin because they are used later. 88| 377| in = thirdin; 89| 377| nextin = fourthin; 90| 3.07k| } else { 91| | // 1. pack the bytes 92| 3.07k| const __m128i utf8_packed = _mm_packus_epi16(in_16, nextin_16); 93| | // 2. store (16 bytes) 94| 3.07k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 95| | // 3. adjust pointers 96| 3.07k| buf += 16; 97| 3.07k| utf8_output += 16; 98| 3.07k| continue; // we are done for this round! 99| 3.07k| } 100| 3.45k| } 101| | 102| | // no bits set above 7th bit -- find out all the ASCII characters 103| 2.64k| const __m128i one_byte_bytemask = 104| 2.64k| _mm_cmpeq_epi16( // this takes four bytes at a time and compares: 105| 2.64k| _mm_and_si128(in_16, v_ff80), // the vector that get only the first 106| | // 9 bits of each 16-bit/2-byte units 107| 2.64k| v_0000 // 108| 2.64k| ); // they should be all zero if they are ASCII. E.g. ASCII in UTF32 is 109| | // of format 0000 0000 0000 0XXX XXXX 110| | // _mm_cmpeq_epi16 should now return a 1111 1111 1111 1111 for equals, and 111| | // 0000 0000 0000 0000 if not for each 16-bit/2-byte units 112| 2.64k| const uint16_t one_byte_bitmask = static_cast(_mm_movemask_epi8( 113| 2.64k| one_byte_bytemask)); // collect the MSB from previous vector and put 114| | // them into uint16_t mas 115| | 116| | // no bits set above 11th bit 117| 2.64k| const __m128i one_or_two_bytes_bytemask = 118| 2.64k| _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_0000); 119| 2.64k| const uint16_t one_or_two_bytes_bitmask = 120| 2.64k| static_cast(_mm_movemask_epi8(one_or_two_bytes_bytemask)); 121| | 122| 2.64k| if (one_or_two_bytes_bitmask == 0xffff) { ------------------ | Branch (122:9): [True: 1.52k, False: 1.11k] ------------------ 123| | // case: all code units either produce 1 or 2 UTF-8 bytes (at least one 124| | // produces 2 bytes) 125| | // 1. prepare 2-byte values 126| | // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8 127| | // expected output : [110a|aaaa|10bb|bbbb] x 8 128| 1.52k| const __m128i v_1f00 = 129| 1.52k| _mm_set1_epi16((int16_t)0x1f00); // 0001 1111 0000 0000 130| 1.52k| const __m128i v_003f = 131| 1.52k| _mm_set1_epi16((int16_t)0x003f); // 0000 0000 0011 1111 132| | 133| | // t0 = [000a|aaaa|bbbb|bb00] 134| 1.52k| const __m128i t0 = _mm_slli_epi16(in_16, 2); // shift packed vector by two 135| | // t1 = [000a|aaaa|0000|0000] 136| 1.52k| const __m128i t1 = _mm_and_si128(t0, v_1f00); // potential first utf8 byte 137| | // t2 = [0000|0000|00bb|bbbb] 138| 1.52k| const __m128i t2 = 139| 1.52k| _mm_and_si128(in_16, v_003f); // potential second utf8 byte 140| | // t3 = [000a|aaaa|00bb|bbbb] 141| 1.52k| const __m128i t3 = 142| 1.52k| _mm_or_si128(t1, t2); // first and second potential utf8 byte together 143| | // t4 = [110a|aaaa|10bb|bbbb] 144| 1.52k| const __m128i t4 = _mm_or_si128( 145| 1.52k| t3, 146| 1.52k| v_c080); // t3 | 1100 0000 1000 0000 = full potential 2-byte utf8 unit 147| | 148| | // 2. merge ASCII and 2-byte codewords 149| 1.52k| const __m128i utf8_unpacked = 150| 1.52k| _mm_blendv_epi8(t4, in_16, one_byte_bytemask); 151| | 152| | // 3. prepare bitmask for 8-bit lookup 153| | // one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - 154| | // MSB, a - LSB) 155| 1.52k| const uint16_t m0 = one_byte_bitmask & 0x5555; // m0 = 0h0g0f0e0d0c0b0a 156| 1.52k| const uint16_t m1 = 157| 1.52k| static_cast(m0 >> 7); // m1 = 00000000h0g0f0e0 158| 1.52k| const uint8_t m2 = 159| 1.52k| static_cast((m0 | m1) & 0xff); // m2 = hdgcfbea 160| | // 4. pack the bytes 161| 1.52k| const uint8_t *row = 162| 1.52k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0]; 163| 1.52k| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 164| 1.52k| const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle); 165| | 166| | // 5. store bytes 167| 1.52k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 168| | 169| | // 6. adjust pointers 170| 1.52k| buf += 8; 171| 1.52k| utf8_output += row[0]; 172| 1.52k| continue; 173| 1.52k| } 174| | 175| | // Check for overflow in packing 176| | 177| 1.11k| const __m128i saturation_bytemask = _mm_cmpeq_epi32( 178| 1.11k| _mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000); 179| 1.11k| const uint32_t saturation_bitmask = 180| 1.11k| static_cast(_mm_movemask_epi8(saturation_bytemask)); 181| 1.11k| if (saturation_bitmask == 0xffff) { ------------------ | Branch (181:9): [True: 407, False: 711] ------------------ 182| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 183| 407| const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800); 184| 407| forbidden_bytemask = 185| 407| _mm_or_si128(forbidden_bytemask, 186| 407| _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_d800)); 187| | 188| 407| const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606, 189| 407| 0x0808, 0x0a0a, 0x0c0c, 0x0e0e); 190| | 191| | /* In this branch we handle three cases: 192| | 1. [0000|0000|0ccc|cccc] => [0ccc|cccc] - 193| | single UFT-8 byte 194| | 2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc] - 195| | two UTF-8 bytes 196| | 3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - 197| | three UTF-8 bytes 198| | 199| | We expand the input word (16-bit) into two code units (32-bit), thus 200| | we have room for four bytes. However, we need five distinct bit 201| | layouts. Note that the last byte in cases #2 and #3 is the same. 202| | 203| | We precompute byte 1 for case #1 and the common byte for cases #2 & #3 204| | in register t2. 205| | 206| | We precompute byte 1 for case #3 and -- **conditionally** -- precompute 207| | either byte 1 for case #2 or byte 2 for case #3. Note that they 208| | differ by exactly one bit. 209| | 210| | Finally from these two code units we build proper UTF-8 sequence, taking 211| | into account the case (i.e, the number of bytes to write). 212| | */ 213| | /** 214| | * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce: 215| | * t2 => [0ccc|cccc] [10cc|cccc] 216| | * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb]) 217| | */ 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) 219| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 220| 407| const __m128i t0 = _mm_shuffle_epi8(in_16, dup_even); 221| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 222| 407| const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 223| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 224| 407| const __m128i t2 = _mm_or_si128(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 225| | 226| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 227| 407| const __m128i s0 = _mm_srli_epi16(in_16, 4); 228| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 229| 407| const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 230| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 231| 407| const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 232| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 233| 407| const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 234| 407| const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, 235| 407| simdutf_vec(0b0100000000000000)); ------------------ | | 218| 407|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 236| 407| const __m128i s4 = _mm_xor_si128(s3, m0); 237| 407|#undef simdutf_vec 238| | 239| | // 4. expand code units 16-bit => 32-bit 240| 407| const __m128i out0 = _mm_unpacklo_epi16(t2, s4); 241| 407| const __m128i out1 = _mm_unpackhi_epi16(t2, s4); 242| | 243| | // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle 244| 407| const uint16_t mask = 245| 407| (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa); 246| 407| if (mask == 0) { ------------------ | Branch (246:11): [True: 82, False: 325] ------------------ 247| | // We only have three-byte code units. Use fast path. 248| 82| const __m128i shuffle = _mm_setr_epi8(2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 249| 82| 15, 13, -1, -1, -1, -1); 250| 82| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle); 251| 82| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle); 252| 82| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 253| 82| utf8_output += 12; 254| 82| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 255| 82| utf8_output += 12; 256| 82| buf += 8; 257| 82| continue; 258| 82| } 259| 325| const uint8_t mask0 = uint8_t(mask); 260| | 261| 325| const uint8_t *row0 = 262| 325| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 263| 325| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 264| 325| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0); 265| | 266| 325| const uint8_t mask1 = static_cast(mask >> 8); 267| | 268| 325| const uint8_t *row1 = 269| 325| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 270| 325| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 271| 325| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1); 272| | 273| 325| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 274| 325| utf8_output += row0[0]; 275| 325| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 276| 325| utf8_output += row1[0]; 277| | 278| 325| buf += 8; 279| 711| } else { 280| | // case: at least one 32-bit word produce a surrogate pair in UTF-16 <=> 281| | // will produce four UTF-8 bytes Let us do a scalar fallback. It may seem 282| | // wasteful to use scalar code, but being efficient with SIMD in the 283| | // presence of surrogate pairs may require non-trivial tables. 284| 711| size_t forward = 15; 285| 711| size_t k = 0; 286| 711| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (286:11): [True: 0, False: 711] ------------------ 287| 0| forward = size_t(end - buf - 1); 288| 0| } 289| 11.3k| for (; k < forward; k++) { ------------------ | Branch (289:14): [True: 10.6k, False: 711] ------------------ 290| 10.6k| uint32_t word = buf[k]; 291| 10.6k| if ((word & 0xFFFFFF80) == 0) { ------------------ | Branch (291:13): [True: 8.43k, False: 2.22k] ------------------ 292| 8.43k| *utf8_output++ = char(word); 293| 8.43k| } else if ((word & 0xFFFFF800) == 0) { ------------------ | Branch (293:20): [True: 432, False: 1.79k] ------------------ 294| 432| *utf8_output++ = char((word >> 6) | 0b11000000); 295| 432| *utf8_output++ = char((word & 0b111111) | 0b10000000); 296| 1.79k| } else if ((word & 0xFFFF0000) == 0) { ------------------ | Branch (296:20): [True: 646, False: 1.14k] ------------------ 297| 646| if (word >= 0xD800 && word <= 0xDFFF) { ------------------ | Branch (297:15): [True: 235, False: 411] | Branch (297:33): [True: 0, False: 235] ------------------ 298| 0| return std::make_pair(nullptr, utf8_output); 299| 0| } 300| 646| *utf8_output++ = char((word >> 12) | 0b11100000); 301| 646| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 302| 646| *utf8_output++ = char((word & 0b111111) | 0b10000000); 303| 1.14k| } else { 304| 1.14k| if (word > 0x10FFFF) { ------------------ | Branch (304:15): [True: 0, False: 1.14k] ------------------ 305| 0| return std::make_pair(nullptr, utf8_output); 306| 0| } 307| 1.14k| *utf8_output++ = char((word >> 18) | 0b11110000); 308| 1.14k| *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000); 309| 1.14k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 310| 1.14k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 311| 1.14k| } 312| 10.6k| } 313| 711| buf += k; 314| 711| } 315| 1.11k| } // while 316| | 317| | // check for invalid input 318| 5.14k| const __m128i v_10ffff = _mm_set1_epi32((uint32_t)0x10ffff); 319| 5.14k| if (static_cast(_mm_movemask_epi8(_mm_cmpeq_epi32( ------------------ | Branch (319:7): [True: 0, False: 5.14k] ------------------ 320| 5.14k| _mm_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffff) { 321| 0| return std::make_pair(nullptr, utf8_output); 322| 0| } 323| | 324| 5.14k| if (static_cast(_mm_movemask_epi8(forbidden_bytemask)) != 0) { ------------------ | Branch (324:7): [True: 0, False: 5.14k] ------------------ 325| 0| return std::make_pair(nullptr, utf8_output); 326| 0| } 327| | 328| 5.14k| return std::make_pair(buf, utf8_output); 329| 5.14k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_129convert_masked_utf8_to_latin1EPKcmRPc: 9| 23.7k| char *&latin1_output) { 10| | // we use an approach where we try to process up to 12 input bytes. 11| | // Why 12 input bytes and not 16? Because we are concerned with the size of 12| | // the lookup tables. Also 12 is nicely divisible by two and three. 13| | // 14| | // 15| | // Optimization note: our main path below is load-latency dependent. Thus it 16| | // is maybe beneficial to have fast paths that depend on branch prediction but 17| | // have less latency. This results in more instructions but, potentially, also 18| | // higher speeds. 19| | // 20| 23.7k| const __m128i in = _mm_loadu_si128((__m128i *)input); 21| 23.7k| const uint16_t input_utf8_end_of_code_point_mask = 22| 23.7k| utf8_end_of_code_point_mask & 23| 23.7k| 0xfff; // we are only processing 12 bytes in case it is not all ASCII 24| 23.7k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (24:7): [True: 141, False: 23.6k] ------------------ 25| | // We process the data in chunks of 12 bytes. 26| 141| _mm_storeu_si128(reinterpret_cast<__m128i *>(latin1_output), in); 27| 141| latin1_output += 12; // We wrote 12 characters. 28| 141| return 12; // We consumed 12 bytes. 29| 141| } 30| | /// We do not have a fast path available, so we fallback. 31| 23.6k| const uint8_t idx = 32| 23.6k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 33| 23.6k| const uint8_t consumed = 34| 23.6k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 35| | // this indicates an invalid input: 36| 23.6k| if (idx >= 64) { ------------------ | Branch (36:7): [True: 5.95k, False: 17.6k] ------------------ 37| 5.95k| return consumed; 38| 5.95k| } 39| | // Here we should have (idx < 64), if not, there is a bug in the validation or 40| | // elsewhere. SIX (6) input code-code units this is a relatively easy scenario 41| | // we process SIX (6) input code-code units. The max length in bytes of six 42| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 43| | // processors where pdep/pext is fast, we might be able to use a small lookup 44| | // table. 45| 17.6k| const __m128i sh = 46| 17.6k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 47| 17.6k| const __m128i perm = _mm_shuffle_epi8(in, sh); 48| 17.6k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 49| 17.6k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 50| 17.6k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 51| 17.6k| const __m128i latin1_packed = _mm_packus_epi16(composed, composed); 52| | // writing 8 bytes even though we only care about the first 6 bytes. 53| | // performance note: it would be faster to use _mm_storeu_si128, we should 54| | // investigate. 55| 17.6k| _mm_storel_epi64((__m128i *)latin1_output, latin1_packed); 56| 17.6k| latin1_output += 6; // We wrote 6 bytes. 57| 17.6k| return consumed; 58| 23.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE0EEEmPKcmRPDs: 10| 21.9k| char16_t *&utf16_output) { 11| | // we use an approach where we try to process up to 12 input bytes. 12| | // Why 12 input bytes and not 16? Because we are concerned with the size of 13| | // the lookup tables. Also 12 is nicely divisible by two and three. 14| | // 15| | // 16| | // Optimization note: our main path below is load-latency dependent. Thus it 17| | // is maybe beneficial to have fast paths that depend on branch prediction but 18| | // have less latency. This results in more instructions but, potentially, also 19| | // higher speeds. 20| | // 21| | // We first try a few fast paths. 22| 21.9k| const __m128i swap = 23| 21.9k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 21.9k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 21.9k| const uint16_t input_utf8_end_of_code_point_mask = 26| 21.9k| utf8_end_of_code_point_mask & 0xfff; 27| 21.9k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 92, False: 21.8k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| | // Note: using 16 bytes is unsafe, see issue_ossfuzz_71218 30| 92| __m128i ascii_first = _mm_cvtepu8_epi16(in); 31| 92| __m128i ascii_second = _mm_cvtepu8_epi16(_mm_srli_si128(in, 8)); 32| 92| if (big_endian) { ------------------ | Branch (32:9): [Folded, False: 92] ------------------ 33| 0| ascii_first = _mm_shuffle_epi8(ascii_first, swap); 34| 0| ascii_second = _mm_shuffle_epi8(ascii_second, swap); 35| 0| } 36| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output), ascii_first); 37| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + 8), 38| 92| ascii_second); 39| 92| utf16_output += 12; // We wrote 12 16-bit characters. 40| 92| return 12; // We consumed 12 bytes. 41| 92| } 42| 21.8k| if (((utf8_end_of_code_point_mask & 0xFFFF) == 0xaaaa)) { ------------------ | Branch (42:7): [True: 332, False: 21.5k] ------------------ 43| | // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte 44| | // UTF-16 code units. There is probably a more efficient sequence, but the 45| | // following might do. 46| 332| const __m128i sh = 47| 332| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 48| 332| const __m128i perm = _mm_shuffle_epi8(in, sh); 49| 332| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 50| 332| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 51| 332| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 52| 332| if (big_endian) ------------------ | Branch (52:9): [Folded, False: 332] ------------------ 53| 0| composed = _mm_shuffle_epi8(composed, swap); 54| 332| _mm_storeu_si128((__m128i *)utf16_output, composed); 55| 332| utf16_output += 8; // We wrote 16 bytes, 8 code points. 56| 332| return 16; 57| 332| } 58| 21.5k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (58:7): [True: 5.78k, False: 15.7k] ------------------ 59| | // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte 60| | // UTF-16 code units. There is probably a more efficient sequence, but the 61| | // following might do. 62| 5.78k| const __m128i sh = 63| 5.78k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 64| 5.78k| const __m128i perm = _mm_shuffle_epi8(in, sh); 65| 5.78k| const __m128i ascii = 66| 5.78k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 67| 5.78k| const __m128i middlebyte = 68| 5.78k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 69| 5.78k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 70| 5.78k| const __m128i highbyte = 71| 5.78k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 72| 5.78k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 73| 5.78k| const __m128i composed = 74| 5.78k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 75| 5.78k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 76| 5.78k| if (big_endian) ------------------ | Branch (76:9): [Folded, False: 5.78k] ------------------ 77| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 78| 5.78k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 79| 5.78k| utf16_output += 4; 80| 5.78k| return 12; 81| 5.78k| } 82| | /// We do not have a fast path available, so we fallback. 83| | 84| 15.7k| const uint8_t idx = 85| 15.7k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 86| 15.7k| const uint8_t consumed = 87| 15.7k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 88| 15.7k| if (idx < 64) { ------------------ | Branch (88:7): [True: 8.40k, False: 7.34k] ------------------ 89| | // SIX (6) input code-code units 90| | // this is a relatively easy scenario 91| | // we process SIX (6) input code-code units. The max length in bytes of six 92| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 93| | // processors where pdep/pext is fast, we might be able to use a small 94| | // lookup table. 95| 8.40k| const __m128i sh = 96| 8.40k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 97| 8.40k| const __m128i perm = _mm_shuffle_epi8(in, sh); 98| 8.40k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 99| 8.40k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 100| 8.40k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 101| 8.40k| if (big_endian) ------------------ | Branch (101:9): [Folded, False: 8.40k] ------------------ 102| 0| composed = _mm_shuffle_epi8(composed, swap); 103| 8.40k| _mm_storeu_si128((__m128i *)utf16_output, composed); 104| 8.40k| utf16_output += 6; // We wrote 12 bytes, 6 code points. 105| 8.40k| } else if (idx < 145) { ------------------ | Branch (105:14): [True: 2.22k, False: 5.12k] ------------------ 106| | // FOUR (4) input code-code units 107| 2.22k| const __m128i sh = 108| 2.22k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 109| 2.22k| const __m128i perm = _mm_shuffle_epi8(in, sh); 110| 2.22k| const __m128i ascii = 111| 2.22k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 112| 2.22k| const __m128i middlebyte = 113| 2.22k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 114| 2.22k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 115| 2.22k| const __m128i highbyte = 116| 2.22k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 117| 2.22k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 118| 2.22k| const __m128i composed = 119| 2.22k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 120| 2.22k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 121| 2.22k| if (big_endian) ------------------ | Branch (121:9): [Folded, False: 2.22k] ------------------ 122| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 123| 2.22k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 124| 2.22k| utf16_output += 4; 125| 5.12k| } else if (idx < 209) { ------------------ | Branch (125:14): [True: 4.89k, False: 222] ------------------ 126| | // TWO (2) input code-code units 127| | ////////////// 128| | // There might be garbage inputs where a leading byte mascarades as a 129| | // four-byte leading byte (by being followed by 3 continuation byte), but is 130| | // not greater than 0xf0. This could trigger a buffer overflow if we only 131| | // counted leading bytes of the form 0xf0 as generating surrogate pairs, 132| | // without further UTF-8 validation. Thus we must be careful to ensure that 133| | // only leading bytes at least as large as 0xf0 generate surrogate pairs. We 134| | // do as at the cost of an extra mask. 135| | ///////////// 136| 4.89k| const __m128i sh = 137| 4.89k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 138| 4.89k| const __m128i perm = _mm_shuffle_epi8(in, sh); 139| 4.89k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 140| 4.89k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 141| 4.89k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 142| 4.89k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 143| | // correct for spurious high bit 144| 4.89k| const __m128i correct = 145| 4.89k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 146| 4.89k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 147| 4.89k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 148| | // We deliberately carry the leading four bits in highbyte if they are 149| | // present, we remove them later when computing hightenbits. 150| 4.89k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 151| 4.89k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 152| | // When we need to generate a surrogate pair (leading byte > 0xF0), then 153| | // the corresponding 32-bit value in 'composed' will be greater than 154| | // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the 155| | // location of the surrogate pairs. 156| 4.89k| const __m128i composed = 157| 4.89k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 158| 4.89k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 159| 4.89k| const __m128i composedminus = 160| 4.89k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 161| 4.89k| const __m128i lowtenbits = 162| 4.89k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 163| | // Notice the 0x3ff mask: 164| 4.89k| const __m128i hightenbits = 165| 4.89k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 166| 4.89k| const __m128i lowtenbitsadd = 167| 4.89k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 168| 4.89k| const __m128i hightenbitsadd = 169| 4.89k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 170| 4.89k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 171| 4.89k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 172| 4.89k| uint32_t basic_buffer[4]; 173| 4.89k| uint32_t basic_buffer_swap[4]; 174| 4.89k| if (big_endian) { ------------------ | Branch (174:9): [Folded, False: 4.89k] ------------------ 175| 0| _mm_storeu_si128((__m128i *)basic_buffer_swap, 176| 0| _mm_shuffle_epi8(composed, swap)); 177| 0| surrogates = _mm_shuffle_epi8(surrogates, swap); 178| 0| } 179| 4.89k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 180| 4.89k| uint32_t surrogate_buffer[4]; 181| 4.89k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 182| 19.5k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (182:24): [True: 14.6k, False: 4.89k] ------------------ 183| 14.6k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (183:11): [True: 3.23k, False: 11.4k] ------------------ 184| 3.23k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 185| 3.23k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 186| 3.23k| utf16_output += 2; 187| 11.4k| } else { 188| 11.4k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (188:27): [Folded, False: 11.4k] ------------------ 189| 11.4k| : uint16_t(basic_buffer[i]); 190| 11.4k| utf16_output++; 191| 11.4k| } 192| 14.6k| } 193| 4.89k| } else { 194| | // here we know that there is an error but we do not handle errors 195| 222| } 196| 15.7k| return consumed; 197| 21.5k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE1EEEmPKcmRPDs: 10| 22.6k| char16_t *&utf16_output) { 11| | // we use an approach where we try to process up to 12 input bytes. 12| | // Why 12 input bytes and not 16? Because we are concerned with the size of 13| | // the lookup tables. Also 12 is nicely divisible by two and three. 14| | // 15| | // 16| | // Optimization note: our main path below is load-latency dependent. Thus it 17| | // is maybe beneficial to have fast paths that depend on branch prediction but 18| | // have less latency. This results in more instructions but, potentially, also 19| | // higher speeds. 20| | // 21| | // We first try a few fast paths. 22| 22.6k| const __m128i swap = 23| 22.6k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 22.6k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 22.6k| const uint16_t input_utf8_end_of_code_point_mask = 26| 22.6k| utf8_end_of_code_point_mask & 0xfff; 27| 22.6k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 92, False: 22.5k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| | // Note: using 16 bytes is unsafe, see issue_ossfuzz_71218 30| 92| __m128i ascii_first = _mm_cvtepu8_epi16(in); 31| 92| __m128i ascii_second = _mm_cvtepu8_epi16(_mm_srli_si128(in, 8)); 32| 92| if (big_endian) { ------------------ | Branch (32:9): [True: 92, Folded] ------------------ 33| 92| ascii_first = _mm_shuffle_epi8(ascii_first, swap); 34| 92| ascii_second = _mm_shuffle_epi8(ascii_second, swap); 35| 92| } 36| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output), ascii_first); 37| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + 8), 38| 92| ascii_second); 39| 92| utf16_output += 12; // We wrote 12 16-bit characters. 40| 92| return 12; // We consumed 12 bytes. 41| 92| } 42| 22.5k| if (((utf8_end_of_code_point_mask & 0xFFFF) == 0xaaaa)) { ------------------ | Branch (42:7): [True: 354, False: 22.1k] ------------------ 43| | // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte 44| | // UTF-16 code units. There is probably a more efficient sequence, but the 45| | // following might do. 46| 354| const __m128i sh = 47| 354| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 48| 354| const __m128i perm = _mm_shuffle_epi8(in, sh); 49| 354| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 50| 354| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 51| 354| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 52| 354| if (big_endian) ------------------ | Branch (52:9): [True: 354, Folded] ------------------ 53| 354| composed = _mm_shuffle_epi8(composed, swap); 54| 354| _mm_storeu_si128((__m128i *)utf16_output, composed); 55| 354| utf16_output += 8; // We wrote 16 bytes, 8 code points. 56| 354| return 16; 57| 354| } 58| 22.1k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (58:7): [True: 6.47k, False: 15.6k] ------------------ 59| | // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte 60| | // UTF-16 code units. There is probably a more efficient sequence, but the 61| | // following might do. 62| 6.47k| const __m128i sh = 63| 6.47k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 64| 6.47k| const __m128i perm = _mm_shuffle_epi8(in, sh); 65| 6.47k| const __m128i ascii = 66| 6.47k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 67| 6.47k| const __m128i middlebyte = 68| 6.47k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 69| 6.47k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 70| 6.47k| const __m128i highbyte = 71| 6.47k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 72| 6.47k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 73| 6.47k| const __m128i composed = 74| 6.47k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 75| 6.47k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 76| 6.47k| if (big_endian) ------------------ | Branch (76:9): [True: 6.47k, Folded] ------------------ 77| 6.47k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 78| 6.47k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 79| 6.47k| utf16_output += 4; 80| 6.47k| return 12; 81| 6.47k| } 82| | /// We do not have a fast path available, so we fallback. 83| | 84| 15.6k| const uint8_t idx = 85| 15.6k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 86| 15.6k| const uint8_t consumed = 87| 15.6k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 88| 15.6k| if (idx < 64) { ------------------ | Branch (88:7): [True: 8.31k, False: 7.38k] ------------------ 89| | // SIX (6) input code-code units 90| | // this is a relatively easy scenario 91| | // we process SIX (6) input code-code units. The max length in bytes of six 92| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 93| | // processors where pdep/pext is fast, we might be able to use a small 94| | // lookup table. 95| 8.31k| const __m128i sh = 96| 8.31k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 97| 8.31k| const __m128i perm = _mm_shuffle_epi8(in, sh); 98| 8.31k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 99| 8.31k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 100| 8.31k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 101| 8.31k| if (big_endian) ------------------ | Branch (101:9): [True: 8.31k, Folded] ------------------ 102| 8.31k| composed = _mm_shuffle_epi8(composed, swap); 103| 8.31k| _mm_storeu_si128((__m128i *)utf16_output, composed); 104| 8.31k| utf16_output += 6; // We wrote 12 bytes, 6 code points. 105| 8.31k| } else if (idx < 145) { ------------------ | Branch (105:14): [True: 2.27k, False: 5.11k] ------------------ 106| | // FOUR (4) input code-code units 107| 2.27k| const __m128i sh = 108| 2.27k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 109| 2.27k| const __m128i perm = _mm_shuffle_epi8(in, sh); 110| 2.27k| const __m128i ascii = 111| 2.27k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 112| 2.27k| const __m128i middlebyte = 113| 2.27k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 114| 2.27k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 115| 2.27k| const __m128i highbyte = 116| 2.27k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 117| 2.27k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 118| 2.27k| const __m128i composed = 119| 2.27k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 120| 2.27k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 121| 2.27k| if (big_endian) ------------------ | Branch (121:9): [True: 2.27k, Folded] ------------------ 122| 2.27k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 123| 2.27k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 124| 2.27k| utf16_output += 4; 125| 5.11k| } else if (idx < 209) { ------------------ | Branch (125:14): [True: 4.89k, False: 222] ------------------ 126| | // TWO (2) input code-code units 127| | ////////////// 128| | // There might be garbage inputs where a leading byte mascarades as a 129| | // four-byte leading byte (by being followed by 3 continuation byte), but is 130| | // not greater than 0xf0. This could trigger a buffer overflow if we only 131| | // counted leading bytes of the form 0xf0 as generating surrogate pairs, 132| | // without further UTF-8 validation. Thus we must be careful to ensure that 133| | // only leading bytes at least as large as 0xf0 generate surrogate pairs. We 134| | // do as at the cost of an extra mask. 135| | ///////////// 136| 4.89k| const __m128i sh = 137| 4.89k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 138| 4.89k| const __m128i perm = _mm_shuffle_epi8(in, sh); 139| 4.89k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 140| 4.89k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 141| 4.89k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 142| 4.89k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 143| | // correct for spurious high bit 144| 4.89k| const __m128i correct = 145| 4.89k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 146| 4.89k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 147| 4.89k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 148| | // We deliberately carry the leading four bits in highbyte if they are 149| | // present, we remove them later when computing hightenbits. 150| 4.89k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 151| 4.89k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 152| | // When we need to generate a surrogate pair (leading byte > 0xF0), then 153| | // the corresponding 32-bit value in 'composed' will be greater than 154| | // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the 155| | // location of the surrogate pairs. 156| 4.89k| const __m128i composed = 157| 4.89k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 158| 4.89k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 159| 4.89k| const __m128i composedminus = 160| 4.89k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 161| 4.89k| const __m128i lowtenbits = 162| 4.89k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 163| | // Notice the 0x3ff mask: 164| 4.89k| const __m128i hightenbits = 165| 4.89k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 166| 4.89k| const __m128i lowtenbitsadd = 167| 4.89k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 168| 4.89k| const __m128i hightenbitsadd = 169| 4.89k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 170| 4.89k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 171| 4.89k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 172| 4.89k| uint32_t basic_buffer[4]; 173| 4.89k| uint32_t basic_buffer_swap[4]; 174| 4.89k| if (big_endian) { ------------------ | Branch (174:9): [True: 4.89k, Folded] ------------------ 175| 4.89k| _mm_storeu_si128((__m128i *)basic_buffer_swap, 176| 4.89k| _mm_shuffle_epi8(composed, swap)); 177| 4.89k| surrogates = _mm_shuffle_epi8(surrogates, swap); 178| 4.89k| } 179| 4.89k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 180| 4.89k| uint32_t surrogate_buffer[4]; 181| 4.89k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 182| 19.5k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (182:24): [True: 14.6k, False: 4.89k] ------------------ 183| 14.6k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (183:11): [True: 3.14k, False: 11.5k] ------------------ 184| 3.14k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 185| 3.14k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 186| 3.14k| utf16_output += 2; 187| 11.5k| } else { 188| 11.5k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (188:27): [True: 11.5k, Folded] ------------------ 189| 11.5k| : uint16_t(basic_buffer[i]); 190| 11.5k| utf16_output++; 191| 11.5k| } 192| 14.6k| } 193| 4.89k| } else { 194| | // here we know that there is an error but we do not handle errors 195| 222| } 196| 15.6k| return consumed; 197| 22.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_128convert_masked_utf8_to_utf32EPKcmRPDi: 9| 14.9k| char32_t *&utf32_output) { 10| | // we use an approach where we try to process up to 12 input bytes. 11| | // Why 12 input bytes and not 16? Because we are concerned with the size of 12| | // the lookup tables. Also 12 is nicely divisible by two and three. 13| | // 14| | // 15| | // Optimization note: our main path below is load-latency dependent. Thus it 16| | // is maybe beneficial to have fast paths that depend on branch prediction but 17| | // have less latency. This results in more instructions but, potentially, also 18| | // higher speeds. 19| | // 20| | // We first try a few fast paths. 21| 14.9k| const __m128i in = _mm_loadu_si128((__m128i *)input); 22| 14.9k| const uint16_t input_utf8_end_of_code_point_mask = 23| 14.9k| utf8_end_of_code_point_mask & 0xfff; 24| 14.9k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (24:7): [True: 92, False: 14.8k] ------------------ 25| | // We process the data in chunks of 12 bytes. 26| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), 27| 92| _mm_cvtepu8_epi32(in)); 28| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 4), 29| 92| _mm_cvtepu8_epi32(_mm_srli_si128(in, 4))); 30| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 8), 31| 92| _mm_cvtepu8_epi32(_mm_srli_si128(in, 8))); 32| 92| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 12), 33| 92| _mm_cvtepu8_epi32(_mm_srli_si128(in, 12))); 34| 92| utf32_output += 12; // We wrote 12 32-bit characters. 35| 92| return 12; // We consumed 12 bytes. 36| 92| } 37| 14.8k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (37:7): [True: 131, False: 14.7k] ------------------ 38| | // We want to take 8 2-byte UTF-8 code units and turn them into 8 4-byte 39| | // UTF-32 code units. There is probably a more efficient sequence, but the 40| | // following might do. 41| 131| const __m128i sh = 42| 131| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 43| 131| const __m128i perm = _mm_shuffle_epi8(in, sh); 44| 131| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 45| 131| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 46| 131| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 47| 131| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), 48| 131| _mm_cvtepu16_epi32(composed)); 49| 131| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 4), 50| 131| _mm_cvtepu16_epi32(_mm_srli_si128(composed, 8))); 51| 131| utf32_output += 8; // We wrote 32 bytes, 8 code points. 52| 131| return 16; 53| 131| } 54| 14.7k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (54:7): [True: 461, False: 14.2k] ------------------ 55| | // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte 56| | // UTF-32 code units. There is probably a more efficient sequence, but the 57| | // following might do. 58| 461| const __m128i sh = 59| 461| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 60| 461| const __m128i perm = _mm_shuffle_epi8(in, sh); 61| 461| const __m128i ascii = 62| 461| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 63| 461| const __m128i middlebyte = 64| 461| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 65| 461| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 66| 461| const __m128i highbyte = 67| 461| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 68| 461| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 69| 461| const __m128i composed = 70| 461| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 71| 461| _mm_storeu_si128((__m128i *)utf32_output, composed); 72| 461| utf32_output += 4; 73| 461| return 12; 74| 461| } 75| | /// We do not have a fast path available, so we fallback. 76| | 77| 14.2k| const uint8_t idx = 78| 14.2k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 79| 14.2k| const uint8_t consumed = 80| 14.2k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 81| 14.2k| if (idx < 64) { ------------------ | Branch (81:7): [True: 8.04k, False: 6.20k] ------------------ 82| | // SIX (6) input code-code units 83| | // this is a relatively easy scenario 84| | // we process SIX (6) input code-code units. The max length in bytes of six 85| | // code code units spanning between 1 and 2 bytes each is 12 bytes. On 86| | // processors where pdep/pext is fast, we might be able to use a small 87| | // lookup table. 88| 8.04k| const __m128i sh = 89| 8.04k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 90| 8.04k| const __m128i perm = _mm_shuffle_epi8(in, sh); 91| 8.04k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 92| 8.04k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 93| 8.04k| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 94| 8.04k| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), 95| 8.04k| _mm_cvtepu16_epi32(composed)); 96| 8.04k| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 4), 97| 8.04k| _mm_cvtepu16_epi32(_mm_srli_si128(composed, 8))); 98| 8.04k| utf32_output += 6; // We wrote 12 bytes, 6 code points. 99| 8.04k| } else if (idx < 145) { ------------------ | Branch (99:14): [True: 1.53k, False: 4.67k] ------------------ 100| | // FOUR (4) input code-code units 101| 1.53k| const __m128i sh = 102| 1.53k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 103| 1.53k| const __m128i perm = _mm_shuffle_epi8(in, sh); 104| 1.53k| const __m128i ascii = 105| 1.53k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 106| 1.53k| const __m128i middlebyte = 107| 1.53k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 108| 1.53k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 109| 1.53k| const __m128i highbyte = 110| 1.53k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 111| 1.53k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 112| 1.53k| const __m128i composed = 113| 1.53k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 114| 1.53k| _mm_storeu_si128((__m128i *)utf32_output, composed); 115| 1.53k| utf32_output += 4; 116| 4.67k| } else if (idx < 209) { ------------------ | Branch (116:14): [True: 4.45k, False: 222] ------------------ 117| | // TWO (2) input code-code units 118| 4.45k| const __m128i sh = 119| 4.45k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 120| 4.45k| const __m128i perm = _mm_shuffle_epi8(in, sh); 121| 4.45k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 122| 4.45k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 123| 4.45k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 124| 4.45k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 125| | // correct for spurious high bit 126| 4.45k| const __m128i correct = 127| 4.45k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 128| 4.45k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 129| 4.45k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 130| 4.45k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000)); 131| 4.45k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 132| 4.45k| const __m128i composed = 133| 4.45k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 134| 4.45k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 135| 4.45k| _mm_storeu_si128((__m128i *)utf32_output, composed); 136| 4.45k| utf32_output += 3; 137| 4.45k| } else { 138| | // here we know that there is an error but we do not handle errors 139| 222| } 140| 14.2k| return consumed; 141| 14.7k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_123utf16_gather_high_bytesILNS_10endiannessE0EEENS1_4simd5simd8IhEENS4_6simd16ItEES8_: 3| 18.4k| const simd16 in1) { 4| 18.4k| if (big_endian) { ------------------ | Branch (4:7): [Folded, False: 18.4k] ------------------ 5| | // we want lower bytes 6| 0| const auto mask = simd16(0x00ff); 7| 0| const auto t0 = in0 & mask; 8| 0| const auto t1 = in1 & mask; 9| | 10| 0| return simd16::pack(t0, t1); 11| 18.4k| } else { 12| 18.4k| const auto t0 = in0.shr<8>(); 13| 18.4k| const auto t1 = in1.shr<8>(); 14| | 15| 18.4k| return simd16::pack(t0, t1); 16| 18.4k| } 17| 18.4k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_123utf16_gather_high_bytesILNS_10endiannessE1EEENS1_4simd5simd8IhEENS4_6simd16ItEES8_: 3| 18.5k| const simd16 in1) { 4| 18.5k| if (big_endian) { ------------------ | Branch (4:7): [True: 18.5k, Folded] ------------------ 5| | // we want lower bytes 6| 18.5k| const auto mask = simd16(0x00ff); 7| 18.5k| const auto t0 = in0 & mask; 8| 18.5k| const auto t1 = in1 & mask; 9| | 10| 18.5k| return simd16::pack(t0, t1); 11| 18.5k| } else { 12| 0| const auto t0 = in0.shr<8>(); 13| 0| const auto t1 = in1.shr<8>(); 14| | 15| 0| return simd16::pack(t0, t1); 16| 0| } 17| 18.5k|}