LLVMFuzzerTestOneInput: 29| 5.26k|extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) { 30| 5.26k| FuzzedDataProvider fdp(data, size); 31| 5.26k| constexpr int kMaxStringSize = 1024; 32| 5.26k| std::string source = fdp.ConsumeRandomLengthString(kMaxStringSize); 33| 21.0k| for (auto& e : simdutf::get_available_implementations()) { ------------------ | Branch (33:16): [True: 21.0k, False: 5.26k] ------------------ 34| 21.0k| if (!e->supported_by_runtime_system()) { ------------------ | Branch (34:9): [True: 5.26k, False: 15.8k] ------------------ 35| 5.26k| continue; 36| 5.26k| } 37| | /** 38| | * Transcoding from UTF-8 to UTF-16LE. 39| | */ 40| 15.8k| bool validutf8 = e->validate_utf8(source.c_str(), source.size()); 41| 15.8k| auto rutf8 = e->validate_utf8_with_errors(source.c_str(), source.size()); 42| 15.8k| if (validutf8 != (rutf8.error == simdutf::SUCCESS)) { // they should agree ------------------ | Branch (42:9): [True: 0, False: 15.8k] ------------------ 43| 0| print_input(source, e); 44| 0| abort(); 45| 0| } 46| 15.8k| if (validutf8) { ------------------ | Branch (46:9): [True: 14.5k, False: 1.20k] ------------------ 47| | // We need a buffer of size where to write the UTF-16LE words. 48| 14.5k| size_t expected_utf16words = 49| 14.5k| e->utf16_length_from_utf8(source.c_str(), source.size()); 50| 14.5k| std::unique_ptr utf16_output{ 51| 14.5k| new char16_t[expected_utf16words]}; 52| | // convert to UTF-16LE 53| 14.5k| size_t utf16words = e->convert_utf8_to_utf16le( 54| 14.5k| source.c_str(), source.size(), utf16_output.get()); 55| | // It wrote utf16words * sizeof(char16_t) bytes. 56| 14.5k| bool validutf16 = e->validate_utf16le(utf16_output.get(), utf16words); 57| 14.5k| if (!validutf16) { ------------------ | Branch (57:11): [True: 0, False: 14.5k] ------------------ 58| 0| print_input(source, e); 59| 0| abort(); 60| 0| } 61| | // convert it back: 62| | // We need a buffer of size where to write the UTF-8 words. 63| 14.5k| size_t expected_utf8words = 64| 14.5k| e->utf8_length_from_utf16le(utf16_output.get(), utf16words); 65| 14.5k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 66| | // convert to UTF-8 67| 14.5k| size_t utf8words = e->convert_utf16le_to_utf8( 68| 14.5k| utf16_output.get(), utf16words, utf8_output.get()); 69| 14.5k| std::string final_string(utf8_output.get(), utf8words); 70| 14.5k| if (final_string != source) { ------------------ | Branch (70:11): [True: 0, False: 14.5k] ------------------ 71| 0| print_input(source, e); 72| 0| abort(); 73| 0| } 74| 14.5k| } else { 75| | // invalid input!!! 76| | // We need a buffer of size where to write the UTF-16LE words. 77| 1.20k| size_t expected_utf16words = 78| 1.20k| e->utf16_length_from_utf8(source.c_str(), source.size()); 79| 1.20k| std::unique_ptr utf16_output{ 80| 1.20k| new char16_t[expected_utf16words]}; 81| | // convert to UTF-16LE 82| 1.20k| size_t utf16words = e->convert_utf8_to_utf16le( 83| 1.20k| source.c_str(), source.size(), utf16_output.get()); 84| 1.20k| if (utf16words != 0) { ------------------ | Branch (84:11): [True: 0, False: 1.20k] ------------------ 85| 0| print_input(source, e); 86| 0| abort(); 87| 0| } 88| 1.20k| } 89| | 90| | /** 91| | * Transcoding from UTF-8 to UTF-16BE. 92| | */ 93| 15.8k| if (validutf8) { ------------------ | Branch (93:9): [True: 14.5k, False: 1.20k] ------------------ 94| | // We need a buffer of size where to write the UTF-16BE words. 95| 14.5k| size_t expected_utf16words = 96| 14.5k| e->utf16_length_from_utf8(source.c_str(), source.size()); 97| 14.5k| std::unique_ptr utf16_output{ 98| 14.5k| new char16_t[expected_utf16words]}; 99| | // convert to UTF-16BE 100| 14.5k| size_t utf16words = e->convert_utf8_to_utf16be( 101| 14.5k| source.c_str(), source.size(), utf16_output.get()); 102| | // It wrote utf16words * sizeof(char16_t) bytes. 103| 14.5k| bool validutf16 = e->validate_utf16be(utf16_output.get(), utf16words); 104| 14.5k| if (!validutf16) { ------------------ | Branch (104:11): [True: 0, False: 14.5k] ------------------ 105| 0| print_input(source, e); 106| 0| abort(); 107| 0| } 108| | // convert it back: 109| | // We need a buffer of size where to write the UTF-8 words. 110| 14.5k| size_t expected_utf8words = 111| 14.5k| e->utf8_length_from_utf16be(utf16_output.get(), utf16words); 112| 14.5k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 113| | // convert to UTF-8 114| 14.5k| size_t utf8words = e->convert_utf16be_to_utf8( 115| 14.5k| utf16_output.get(), utf16words, utf8_output.get()); 116| 14.5k| std::string final_string(utf8_output.get(), utf8words); 117| 14.5k| if (final_string != source) { ------------------ | Branch (117:11): [True: 0, False: 14.5k] ------------------ 118| 0| print_input(source, e); 119| 0| abort(); 120| 0| } 121| 14.5k| } else { 122| | // invalid input!!! 123| | // We need a buffer of size where to write the UTF-16BE words. 124| 1.20k| size_t expected_utf16words = 125| 1.20k| e->utf16_length_from_utf8(source.c_str(), source.size()); 126| 1.20k| std::unique_ptr utf16_output{ 127| 1.20k| new char16_t[expected_utf16words]}; 128| | // convert to UTF-16BE 129| 1.20k| size_t utf16words = e->convert_utf8_to_utf16be( 130| 1.20k| source.c_str(), source.size(), utf16_output.get()); 131| 1.20k| if (utf16words != 0) { ------------------ | Branch (131:11): [True: 0, False: 1.20k] ------------------ 132| 0| print_input(source, e); 133| 0| abort(); 134| 0| } 135| 1.20k| } 136| | /** 137| | * Transcoding from UTF-8 to UTF-32. 138| | */ 139| 15.8k| if (validutf8) { ------------------ | Branch (139:9): [True: 14.5k, False: 1.20k] ------------------ 140| | // We need a buffer of size where to write the UTF-32 words. 141| 14.5k| size_t expected_utf32words = 142| 14.5k| e->utf32_length_from_utf8(source.c_str(), source.size()); 143| 14.5k| std::unique_ptr utf32_output{ 144| 14.5k| new char32_t[expected_utf32words]}; 145| | // convert to UTF-32 146| 14.5k| size_t utf32words = e->convert_utf8_to_utf32( 147| 14.5k| source.c_str(), source.size(), utf32_output.get()); 148| | // It wrote utf32words * sizeof(char32_t) bytes. 149| 14.5k| bool validutf32 = e->validate_utf32(utf32_output.get(), utf32words); 150| 14.5k| if (!validutf32) { ------------------ | Branch (150:11): [True: 0, False: 14.5k] ------------------ 151| 0| return -1; 152| 0| } 153| | // convert it back: 154| | // We need a buffer of size where to write the UTF-8 words. 155| 14.5k| size_t expected_utf8words = 156| 14.5k| e->utf8_length_from_utf32(utf32_output.get(), utf32words); 157| 14.5k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 158| | // convert to UTF-8 159| 14.5k| size_t utf8words = e->convert_utf32_to_utf8( 160| 14.5k| utf32_output.get(), utf32words, utf8_output.get()); 161| 14.5k| std::string final_string(utf8_output.get(), utf8words); 162| 14.5k| if (source != final_string) { ------------------ | Branch (162:11): [True: 0, False: 14.5k] ------------------ 163| 0| print_input(source, e); 164| 0| abort(); 165| 0| } 166| 14.5k| } else { 167| | // invalid input!!! 168| 1.20k| size_t expected_utf32words = 169| 1.20k| e->utf32_length_from_utf8(source.c_str(), source.size()); 170| 1.20k| std::unique_ptr utf32_output{ 171| 1.20k| new char32_t[expected_utf32words]}; 172| | // convert to UTF-32 173| 1.20k| size_t utf32words = e->convert_utf8_to_utf32( 174| 1.20k| source.c_str(), source.size(), utf32_output.get()); 175| 1.20k| if (utf32words != 0) { ------------------ | Branch (175:11): [True: 0, False: 1.20k] ------------------ 176| 0| print_input(source, e); 177| 0| abort(); 178| 0| } 179| 1.20k| } 180| | 181| | /** 182| | * Transcoding from UTF-8 to Latin 1 183| | */ 184| 15.8k| if (validutf8) { ------------------ | Branch (184:9): [True: 14.5k, False: 1.20k] ------------------ 185| | // We need a buffer of size where to write the UTF-16LE words. 186| 14.5k| size_t expected_latin1words = 187| 14.5k| e->latin1_length_from_utf8(source.c_str(), source.size()); 188| 14.5k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 189| | // convert to latin1 190| 14.5k| size_t latin1words = e->convert_utf8_to_latin1( 191| 14.5k| source.c_str(), source.size(), latin1_output.get()); 192| 14.5k| if (latin1words != 0) { ------------------ | Branch (192:11): [True: 717, False: 13.8k] ------------------ 193| | // convert it back: 194| | // We need a buffer of size where to write the UTF-8 words. 195| 717| size_t expected_utf8words = 196| 717| e->utf8_length_from_latin1(latin1_output.get(), latin1words); 197| 717| std::unique_ptr utf8_output{new char[expected_utf8words]}; 198| | // convert to UTF-8 199| 717| size_t utf8words = e->convert_latin1_to_utf8( 200| 717| latin1_output.get(), latin1words, utf8_output.get()); 201| 717| std::string final_string(utf8_output.get(), utf8words); 202| 717| if (final_string != source) { ------------------ | Branch (202:13): [True: 0, False: 717] ------------------ 203| 0| print_input(source, e); 204| 0| abort(); 205| 0| } 206| 717| } 207| 14.5k| } else { 208| | // invalid input!!! 209| | // We need a buffer of size where to write the Latin 1 words. 210| 1.20k| size_t expected_latin1words = 211| 1.20k| e->latin1_length_from_utf8(source.c_str(), source.size()); 212| 1.20k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 213| | // convert to Latin 1 214| 1.20k| size_t latin1words = e->convert_utf8_to_latin1( 215| 1.20k| source.c_str(), source.size(), latin1_output.get()); 216| 1.20k| if (latin1words != 0) { ------------------ | Branch (216:11): [True: 0, False: 1.20k] ------------------ 217| 0| print_input(source, e); 218| 0| abort(); 219| 0| } 220| 1.20k| } 221| | /** 222| | * Transcoding from UTF-16LE to UTF-8. 223| | */ 224| 15.8k| { 225| | // Get new source data here as this will allow the fuzzer to optimize it's 226| | // input for UTF16-LE. 227| 15.8k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 228| | // We copy to avoid alignment issues. 229| 15.8k| std::unique_ptr utf16_source{new char16_t[source.size() / 2]}; 230| 15.8k| if (source.data() != nullptr) { ------------------ | Branch (230:11): [True: 15.8k, False: 0] ------------------ 231| 15.8k| std::memcpy(utf16_source.get(), source.data(), source.size() / 2 * 2); 232| 15.8k| } 233| 15.8k| bool validutf16le = 234| 15.8k| e->validate_utf16le(utf16_source.get(), source.size() / 2); 235| 15.8k| auto rutf16le = e->validate_utf16le_with_errors(utf16_source.get(), 236| 15.8k| source.size() / 2); 237| 15.8k| if (validutf16le != ------------------ | Branch (237:11): [True: 0, False: 15.8k] ------------------ 238| 15.8k| (rutf16le.error == simdutf::SUCCESS)) { // they should agree 239| 0| print_input(source, e); 240| 0| abort(); 241| 0| } 242| 15.8k| if (validutf16le) { ------------------ | Branch (242:11): [True: 15.5k, False: 258] ------------------ 243| | // We need a buffer of size where to write the UTF-16 words. 244| 15.5k| size_t expected_utf8words = 245| 15.5k| e->utf8_length_from_utf16le(utf16_source.get(), source.size() / 2); 246| 15.5k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 247| 15.5k| size_t utf8words = e->convert_utf16le_to_utf8( 248| 15.5k| utf16_source.get(), source.size() / 2, utf8_output.get()); 249| | // It wrote utf16words * sizeof(char16_t) bytes. 250| 15.5k| bool validutf8 = e->validate_utf8(utf8_output.get(), utf8words); 251| 15.5k| if (!validutf8) { ------------------ | Branch (251:13): [True: 0, False: 15.5k] ------------------ 252| 0| print_input(source, e); 253| 0| abort(); 254| 0| } 255| | // convert it back: 256| | // We need a buffer of size where to write the UTF-16 words. 257| 15.5k| size_t expected_utf16words = 258| 15.5k| e->utf16_length_from_utf8(utf8_output.get(), utf8words); 259| 15.5k| std::unique_ptr utf16_output{ 260| 15.5k| new char16_t[expected_utf16words]}; 261| | // convert to UTF-8 262| 15.5k| size_t utf16words = e->convert_utf8_to_utf16le( 263| 15.5k| utf8_output.get(), utf8words, utf16_output.get()); 264| 100k| for (size_t i = 0; i < source.size() / 2; i++) { ------------------ | Branch (264:28): [True: 84.8k, False: 15.5k] ------------------ 265| 84.8k| if (utf16_output.get()[i] != (utf16_source.get())[i]) { ------------------ | Branch (265:15): [True: 0, False: 84.8k] ------------------ 266| 0| print_input(source, e); 267| 0| abort(); 268| 0| } 269| 84.8k| } 270| 15.5k| } else { 271| | // invalid input!!! 272| | // We need a buffer of size where to write the UTF-16 words. 273| 258| size_t expected_utf8words = 274| 258| e->utf8_length_from_utf16le(utf16_source.get(), source.size() / 2); 275| 258| std::unique_ptr utf8_output{new char[expected_utf8words]}; 276| 258| size_t utf8words = e->convert_utf16le_to_utf8( 277| 258| utf16_source.get(), source.size() / 2, utf8_output.get()); 278| 258| if (utf8words != 0) { ------------------ | Branch (278:13): [True: 0, False: 258] ------------------ 279| 0| print_input(source, e); 280| 0| abort(); 281| 0| } 282| 258| } 283| 15.8k| } 284| | 285| | /** 286| | * Transcoding from UTF-16BE to UTF-8. 287| | */ 288| 15.8k| { 289| | // Get new source data here as this will allow the fuzzer to optimize it's 290| | // input for UTF16-BE. 291| 15.8k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 292| 15.8k| std::unique_ptr utf16_source{new char16_t[source.size() / 2]}; 293| 15.8k| if (source.data() != nullptr) { ------------------ | Branch (293:11): [True: 15.8k, False: 0] ------------------ 294| 15.8k| std::memcpy(utf16_source.get(), source.data(), source.size() / 2 * 2); 295| 15.8k| } 296| 15.8k| bool validutf16be = 297| 15.8k| e->validate_utf16be(utf16_source.get(), source.size() / 2); 298| 15.8k| auto rutf16be = e->validate_utf16be_with_errors(utf16_source.get(), 299| 15.8k| source.size() / 2); 300| 15.8k| if (validutf16be != ------------------ | Branch (300:11): [True: 0, False: 15.8k] ------------------ 301| 15.8k| (rutf16be.error == simdutf::SUCCESS)) { // they should agree 302| 0| print_input(source, e); 303| 0| abort(); 304| 0| } 305| 15.8k| if (validutf16be) { ------------------ | Branch (305:11): [True: 15.5k, False: 236] ------------------ 306| | // We need a buffer of size where to write the UTF-16 words. 307| 15.5k| size_t expected_utf8words = 308| 15.5k| e->utf8_length_from_utf16be(utf16_source.get(), source.size() / 2); 309| 15.5k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 310| 15.5k| size_t utf8words = e->convert_utf16be_to_utf8( 311| 15.5k| utf16_source.get(), source.size() / 2, utf8_output.get()); 312| | // It wrote utf16words * sizeof(char16_t) bytes. 313| 15.5k| bool validutf8 = e->validate_utf8(utf8_output.get(), utf8words); 314| 15.5k| if (!validutf8) { ------------------ | Branch (314:13): [True: 0, False: 15.5k] ------------------ 315| 0| print_input(source, e); 316| 0| abort(); 317| 0| } 318| | // convert it back: 319| | // We need a buffer of size where to write the UTF-16 words. 320| 15.5k| size_t expected_utf16words = 321| 15.5k| e->utf16_length_from_utf8(utf8_output.get(), utf8words); 322| 15.5k| std::unique_ptr utf16_output{ 323| 15.5k| new char16_t[expected_utf16words]}; 324| | // convert to UTF-8 325| 15.5k| size_t utf16words = e->convert_utf8_to_utf16be( 326| 15.5k| utf8_output.get(), utf8words, utf16_output.get()); 327| 103k| for (size_t i = 0; i < source.size() / 2; i++) { ------------------ | Branch (327:28): [True: 88.0k, False: 15.5k] ------------------ 328| 88.0k| if (utf16_output.get()[i] != (utf16_source.get())[i]) { ------------------ | Branch (328:15): [True: 0, False: 88.0k] ------------------ 329| 0| print_input(source, e); 330| 0| abort(); 331| 0| } 332| 88.0k| } 333| 15.5k| } else { 334| | // invalid input!!! 335| | // We need a buffer of size where to write the UTF-16 words. 336| 236| size_t expected_utf8words = 337| 236| e->utf8_length_from_utf16be(utf16_source.get(), source.size() / 2); 338| 236| std::unique_ptr utf8_output{new char[expected_utf8words]}; 339| 236| size_t utf8words = e->convert_utf16be_to_utf8( 340| 236| utf16_source.get(), source.size() / 2, utf8_output.get()); 341| 236| if (utf8words != 0) { ------------------ | Branch (341:13): [True: 0, False: 236] ------------------ 342| 0| print_input(source, e); 343| 0| abort(); 344| 0| } 345| 236| } 346| 15.8k| } 347| | 348| | /** 349| | * Transcoding from latin1 to UTF-8. 350| | */ 351| | // Get new source data here as this will allow the fuzzer to optimize it's 352| | // input for latin1. 353| 15.8k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 354| 15.8k| bool validlatin1 = true; // has to be 355| 15.8k| if (validlatin1) { ------------------ | Branch (355:9): [True: 15.8k, False: 0] ------------------ 356| | // We need a buffer of size where to write the UTF-8 words. 357| 15.8k| size_t expected_utf8words = 358| 15.8k| e->utf8_length_from_latin1(source.c_str(), source.size()); 359| 15.8k| std::unique_ptr utf8_output{new char[expected_utf8words]}; 360| 15.8k| size_t utf8words = e->convert_latin1_to_utf8( 361| 15.8k| source.c_str(), source.size(), utf8_output.get()); 362| | // It wrote utf8words * sizeof(char) bytes. 363| 15.8k| bool validutf8 = e->validate_utf8(utf8_output.get(), utf8words); 364| 15.8k| if (!validutf8) { ------------------ | Branch (364:11): [True: 0, False: 15.8k] ------------------ 365| 0| print_input(source, e); 366| 0| abort(); 367| 0| } 368| | // convert it back: 369| | // We need a buffer of size where to write the latin1 words. 370| 15.8k| size_t expected_latin1words = 371| 15.8k| e->latin1_length_from_utf8(utf8_output.get(), utf8words); 372| 15.8k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 373| | // convert to latin1 374| 15.8k| size_t latin1words = e->convert_utf8_to_latin1( 375| 15.8k| utf8_output.get(), utf8words, latin1_output.get()); 376| 421k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (376:26): [True: 405k, False: 15.8k] ------------------ 377| 405k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (377:13): [True: 0, False: 405k] ------------------ 378| 0| print_input(source, e); 379| 0| abort(); 380| 0| } 381| 405k| } 382| 15.8k| } 383| 15.8k| if (validlatin1) { ------------------ | Branch (383:9): [True: 15.8k, False: 0] ------------------ 384| | // We need a buffer of size where to write the UTF-16 words. 385| 15.8k| size_t expected_utf16words = e->utf16_length_from_latin1(source.size()); 386| 15.8k| std::unique_ptr utf16_output{ 387| 15.8k| new char16_t[expected_utf16words]}; 388| 15.8k| size_t utf16words = e->convert_latin1_to_utf16le( 389| 15.8k| source.c_str(), source.size(), utf16_output.get()); 390| | // It wrote utf16words * sizeof(char16_t) bytes. 391| 15.8k| bool validutf16 = e->validate_utf16le(utf16_output.get(), utf16words); 392| 15.8k| if (!validutf16) { ------------------ | Branch (392:11): [True: 0, False: 15.8k] ------------------ 393| 0| print_input(source, e); 394| 0| abort(); 395| 0| } 396| | // convert it back: 397| | // We need a buffer of size where to write the latin1 words. 398| 15.8k| size_t expected_latin1words = e->latin1_length_from_utf16(utf16words); 399| 15.8k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 400| | // convert to latin1 401| 15.8k| size_t latin1words = e->convert_utf16le_to_latin1( 402| 15.8k| utf16_output.get(), utf16words, latin1_output.get()); 403| 421k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (403:26): [True: 405k, False: 15.8k] ------------------ 404| 405k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (404:13): [True: 0, False: 405k] ------------------ 405| 0| print_input(source, e); 406| 0| abort(); 407| 0| } 408| 405k| } 409| 15.8k| } 410| 15.8k| if (validlatin1) { ------------------ | Branch (410:9): [True: 15.8k, False: 0] ------------------ 411| | // We need a buffer of size where to write the UTF-16 words. 412| 15.8k| size_t expected_utf16words = e->utf16_length_from_latin1(source.size()); 413| 15.8k| std::unique_ptr utf16_output{ 414| 15.8k| new char16_t[expected_utf16words]}; 415| 15.8k| size_t utf16words = e->convert_latin1_to_utf16be( 416| 15.8k| source.c_str(), source.size(), utf16_output.get()); 417| | // It wrote utf16words * sizeof(char16_t) bytes. 418| 15.8k| bool validutf16 = e->validate_utf16be(utf16_output.get(), utf16words); 419| 15.8k| if (!validutf16) { ------------------ | Branch (419:11): [True: 0, False: 15.8k] ------------------ 420| 0| print_input(source, e); 421| 0| abort(); 422| 0| } 423| | // convert it back: 424| | // We need a buffer of size where to write the latin1 words. 425| 15.8k| size_t expected_latin1words = e->latin1_length_from_utf16(utf16words); 426| 15.8k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 427| | // convert to latin1 428| 15.8k| size_t latin1words = e->convert_utf16be_to_latin1( 429| 15.8k| utf16_output.get(), utf16words, latin1_output.get()); 430| 421k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (430:26): [True: 405k, False: 15.8k] ------------------ 431| 405k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (431:13): [True: 0, False: 405k] ------------------ 432| 0| print_input(source, e); 433| 0| abort(); 434| 0| } 435| 405k| } 436| 15.8k| } 437| | 438| 15.8k| if (validlatin1) { ------------------ | Branch (438:9): [True: 15.8k, False: 0] ------------------ 439| | // We need a buffer of size where to write the UTF-16 words. 440| 15.8k| size_t expected_utf32words = e->utf32_length_from_latin1(source.size()); 441| 15.8k| std::unique_ptr utf32_output{ 442| 15.8k| new char32_t[expected_utf32words]}; 443| 15.8k| size_t utf32words = e->convert_latin1_to_utf32( 444| 15.8k| source.c_str(), source.size(), utf32_output.get()); 445| | // It wrote utf16words * sizeof(char16_t) bytes. 446| 15.8k| bool validutf32 = e->validate_utf32(utf32_output.get(), utf32words); 447| 15.8k| if (!validutf32) { ------------------ | Branch (447:11): [True: 0, False: 15.8k] ------------------ 448| 0| print_input(source, e); 449| 0| abort(); 450| 0| } 451| | // convert it back: 452| | // We need a buffer of size where to write the latin1 words. 453| 15.8k| size_t expected_latin1words = e->latin1_length_from_utf32(utf32words); 454| 15.8k| std::unique_ptr latin1_output{new char[expected_latin1words]}; 455| | // convert to latin1 456| 15.8k| size_t latin1words = e->convert_utf32_to_latin1( 457| 15.8k| utf32_output.get(), utf32words, latin1_output.get()); 458| 421k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (458:26): [True: 405k, False: 15.8k] ------------------ 459| 405k| if (latin1_output.get()[i] != (source.c_str())[i]) { ------------------ | Branch (459:13): [True: 0, False: 405k] ------------------ 460| 0| print_input(source, e); 461| 0| abort(); 462| 0| } 463| 405k| } 464| 15.8k| } 465| | 466| | /// Base64 tests. We begin by trying to decode the input, even if we 467| | /// expect it to fail. 468| 15.8k| { 469| 15.8k| size_t max_length_needed = 470| 15.8k| e->maximal_binary_length_from_base64(source.data(), source.size()); 471| 15.8k| std::vector back(max_length_needed); 472| 15.8k| simdutf::result r = 473| 15.8k| e->base64_to_binary(source.data(), source.size(), back.data()); 474| 15.8k| if (r.error == simdutf::error_code::SUCCESS) { ------------------ | Branch (474:11): [True: 14.4k, False: 1.38k] ------------------ 475| | // We expect failure but if we succeed, then we should have a roundtrip. 476| 14.4k| back.resize(r.count); 477| 14.4k| std::vector back2(e->base64_length_from_binary(back.size())); 478| 14.4k| size_t base64size = 479| 14.4k| e->binary_to_base64(back.data(), back.size(), back2.data()); 480| 14.4k| back2.resize(base64size); 481| 14.4k| std::vector back3( 482| 14.4k| e->maximal_binary_length_from_base64(back2.data(), back2.size())); 483| 14.4k| simdutf::result r2 = 484| 14.4k| e->base64_to_binary(back2.data(), back2.size(), back3.data()); 485| 14.4k| if (r2.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (485:13): [True: 0, False: 14.4k] ------------------ 486| 0| print_input(source, e); 487| 0| return false; 488| 0| } 489| 14.4k| if (r2.count != back.size()) { ------------------ | Branch (489:13): [True: 0, False: 14.4k] ------------------ 490| 0| print_input(source, e); 491| 0| return false; 492| 0| } 493| 14.4k| if (back3.size() != back.size()) { ------------------ | Branch (493:13): [True: 0, False: 14.4k] ------------------ 494| 0| print_input(source, e); 495| 0| return false; 496| 0| } 497| 14.4k| } 498| 15.8k| } 499| | 500| | // Same as above, but we use the safe decoder version. 501| 15.8k| { 502| 15.8k| size_t max_length_needed = 503| 15.8k| e->maximal_binary_length_from_base64(source.data(), source.size()); 504| 15.8k| std::vector back(max_length_needed); 505| 15.8k| simdutf::result r = simdutf::base64_to_binary_safe( 506| 15.8k| source.data(), source.size(), back.data(), max_length_needed); 507| 15.8k| if (r.error == simdutf::error_code::SUCCESS) { ------------------ | Branch (507:11): [True: 14.4k, False: 1.38k] ------------------ 508| | // We expect failure but if we succeed, then we should have a roundtrip. 509| 14.4k| back.resize(max_length_needed); 510| 14.4k| std::vector back2(e->base64_length_from_binary(back.size())); 511| 14.4k| size_t base64size = 512| 14.4k| e->binary_to_base64(back.data(), back.size(), back2.data()); 513| 14.4k| back2.resize(base64size); 514| 14.4k| size_t max_length_needed2 = 515| 14.4k| e->maximal_binary_length_from_base64(back2.data(), back2.size()); 516| 14.4k| std::vector back3(max_length_needed2); 517| 14.4k| simdutf::result r2 = simdutf::base64_to_binary_safe( 518| 14.4k| back2.data(), back2.size(), back3.data(), max_length_needed2); 519| 14.4k| if (r2.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (519:13): [True: 0, False: 14.4k] ------------------ 520| 0| print_input(source, e); 521| 0| return false; 522| 0| } 523| 14.4k| if (max_length_needed != back.size()) { ------------------ | Branch (523:13): [True: 0, False: 14.4k] ------------------ 524| 0| print_input(source, e); 525| 0| return false; 526| 0| } 527| 14.4k| if (back3.size() != back.size()) { ------------------ | Branch (527:13): [True: 0, False: 14.4k] ------------------ 528| 0| print_input(source, e); 529| 0| return false; 530| 0| } 531| 14.4k| } 532| 15.8k| } 533| | /// Base64 tests. We encode the content as binary in base64 and we decode 534| | /// it, it should always succeed. 535| 15.8k| { 536| 15.8k| source = fdp.ConsumeRandomLengthString(kMaxStringSize); 537| 15.8k| std::vector base64buffer( 538| 15.8k| e->base64_length_from_binary(source.size())); 539| 15.8k| size_t base64size = e->binary_to_base64(source.data(), source.size(), 540| 15.8k| base64buffer.data()); 541| 15.8k| if (base64size != base64buffer.size()) { ------------------ | Branch (541:11): [True: 0, False: 15.8k] ------------------ 542| 0| print_input(source, e); 543| 0| abort(); 544| 0| } 545| 15.8k| std::vector back(e->maximal_binary_length_from_base64( 546| 15.8k| base64buffer.data(), base64buffer.size())); 547| 15.8k| simdutf::result r = e->base64_to_binary(base64buffer.data(), 548| 15.8k| base64buffer.size(), back.data()); 549| 15.8k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (549:11): [True: 0, False: 15.8k] ------------------ 550| 0| print_input(source, e); 551| 0| abort(); 552| 0| } 553| 15.8k| if (r.count != source.size()) { ------------------ | Branch (553:11): [True: 0, False: 15.8k] ------------------ 554| 0| print_input(source, e); 555| 0| abort(); 556| 0| } 557| 362k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (557:26): [True: 346k, False: 15.8k] ------------------ 558| 346k| if (back[i] != (source.c_str())[i]) { ------------------ | Branch (558:13): [True: 0, False: 346k] ------------------ 559| 0| print_input(source, e); 560| 0| abort(); 561| 0| } 562| 346k| } 563| 15.8k| size_t max_length = back.size(); 564| 15.8k| r = simdutf::base64_to_binary_safe( 565| 15.8k| base64buffer.data(), base64buffer.size(), back.data(), max_length); 566| 15.8k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (566:11): [True: 0, False: 15.8k] ------------------ 567| 0| printf("base64 round trip failed, error code %d\n", r.error); 568| 0| print_input(source, e); 569| 0| return false; 570| 0| } 571| 15.8k| if (max_length != source.size()) { ------------------ | Branch (571:11): [True: 0, False: 15.8k] ------------------ 572| 0| printf("base64 safe round trip failed, not the same size %zu %zu\n", 573| 0| max_length, source.size()); 574| 0| print_input(source, e); 575| 0| return false; 576| 0| } 577| 362k| for (size_t i = 0; i < source.size(); i++) { ------------------ | Branch (577:26): [True: 346k, False: 15.8k] ------------------ 578| 346k| if (back[i] != (source.c_str())[i]) { ------------------ | Branch (578:13): [True: 0, False: 346k] ------------------ 579| 0| printf("base64 round trip failed, same size, different content\n"); 580| 0| print_input(source, e); 581| 0| return false; 582| 0| } 583| 346k| } 584| 15.8k| } 585| | 586| 15.8k| } // for (auto &e : simdutf::get_available_implementations()) { 587| | 588| 5.26k| return 0; 589| 5.26k|} // 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| 46.0k| bool decode_up_to_bad_char) noexcept { 64| 46.0k| static_assert(std::is_same::value || 65| 46.0k| std::is_same::value, 66| 46.0k| "Only char and char16_t are supported."); 67| 46.0k| size_t remaining_input_length = length; 68| 46.0k| size_t remaining_output_length = outlen; 69| 46.0k| size_t input_position = 0; 70| 46.0k| size_t output_position = 0; 71| | 72| | // We also do a first pass using the fast path to decode as much as possible 73| 46.0k| size_t safe_input = detail::min( 74| 46.0k| remaining_input_length, 75| 46.0k| base64_length_from_binary(remaining_output_length / 3 * 3, options)); 76| 46.0k| bool done_with_partial = (safe_input == remaining_input_length); 77| 46.0k| 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| 46.0k| { 89| 46.0k| r = get_active_implementation()->base64_to_binary_details( 90| 46.0k| input + input_position, safe_input, output + output_position, options, 91| 46.0k| done_with_partial ------------------ | Branch (91:9): [True: 43.2k, False: 2.78k] ------------------ 92| 46.0k| ? last_chunk_handling_options 93| 46.0k| : simdutf::last_chunk_handling_options::only_full_chunks); 94| 46.0k| } 95| 46.0k| simdutf_log_assert(r.input_count <= safe_input, 96| 46.0k| "You should not read more than safe_input"); 97| 46.0k| simdutf_log_assert(r.output_count <= remaining_output_length, 98| 46.0k| "You should not write more than remaining_output_length"); 99| | // Technically redundant, but we want to be explicit about it. 100| 46.0k| input_position += r.input_count; 101| 46.0k| output_position += r.output_count; 102| 46.0k| remaining_input_length -= r.input_count; 103| 46.0k| remaining_output_length -= r.output_count; 104| 46.0k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (104:7): [True: 1.08k, False: 44.9k] ------------------ 105| | // There is an error. We return. 106| 1.08k| if (decode_up_to_bad_char && ------------------ | Branch (106:9): [True: 0, False: 1.08k] ------------------ 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| 1.08k| outlen = output_position; 112| 1.08k| return {r.error, input_position}; 113| 1.08k| } 114| | 115| 44.9k| if (done_with_partial) { ------------------ | Branch (115:7): [True: 42.6k, False: 2.28k] ------------------ 116| | // We are done. We have decoded everything. 117| 42.6k| outlen = output_position; 118| 42.6k| return {simdutf::error_code::SUCCESS, input_position}; 119| 42.6k| } 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| 2.28k| r = simdutf::scalar::base64::base64_to_binary_details_safe_impl( 123| 2.28k| input + input_position, remaining_input_length, output + output_position, 124| 2.28k| remaining_output_length, options, last_chunk_handling_options); 125| 2.28k| input_position += r.input_count; 126| 2.28k| output_position += r.output_count; 127| 2.28k| remaining_input_length -= r.input_count; 128| 2.28k| remaining_output_length -= r.output_count; 129| | 130| 2.28k| if (r.error != simdutf::error_code::SUCCESS) { ------------------ | Branch (130:7): [True: 304, False: 1.98k] ------------------ 131| | // There is an error. We return. 132| 304| if (decode_up_to_bad_char && ------------------ | Branch (132:9): [True: 0, False: 304] ------------------ 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| 304| outlen = output_position; 138| 304| return {r.error, input_position}; 139| 304| } 140| 1.98k| if (input_position < length) { ------------------ | Branch (140:7): [True: 0, False: 1.98k] ------------------ 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| 1.98k| outlen = output_position; 156| 1.98k| return {simdutf::error_code::SUCCESS, input_position}; 157| 2.28k|} _ZN7simdutf6resultC2ENS_10error_codeEm: 87| 150k| : error{err}, count{pos} {} _ZNK7simdutf6result6is_errEv: 93| 8.06k| simdutf_really_inline simdutf_constexpr23 bool is_err() const noexcept { 94| 8.06k| return error != error_code::SUCCESS; 95| 8.06k| } _ZN7simdutf11full_resultC2Ev: 106| 46.0k| : error{error_code::SUCCESS}, input_count{0}, output_count{0} {} _ZN7simdutf11full_resultC2ENS_10error_codeEmm: 111| 94.2k| : error{err}, input_count{pos_in}, output_count{pos_out} {} _ZN7simdutf11full_resultC2ENS_10error_codeEmmb: 114| 125| : error{err}, input_count{pos_in}, output_count{pos_out}, 115| 125| padding_error{padding_err} {} _ZNK7simdutf11full_resultcvNS_6resultEEv: 117| 46.0k| simdutf_really_inline simdutf_constexpr23 operator result() const noexcept { 118| 46.0k| if (error == error_code::SUCCESS) { ------------------ | Branch (118:9): [True: 44.6k, False: 1.38k] ------------------ 119| 44.6k| return result{error, output_count}; 120| 44.6k| } else { 121| 1.38k| return result{error, input_count}; 122| 1.38k| } 123| 46.0k| } _ZN7simdutf10is_partialENS_27last_chunk_handling_optionsE: 4136| 19.2k|is_partial(last_chunk_handling_options options) { 4137| 19.2k| return (options == stop_before_partial) || (options == only_full_chunks); ------------------ | Branch (4137:10): [True: 0, False: 19.2k] | Branch (4137:46): [True: 3.55k, False: 15.6k] ------------------ 4138| 19.2k|} _ZNK7simdutf14implementation25required_instruction_setsEv: 5157| 21.0k| virtual uint32_t required_instruction_sets() const { 5158| 21.0k| return _required_instruction_sets; 5159| 21.0k| } _ZNK7simdutf14implementation24utf16_length_from_latin1Em: 6347| 31.6k| utf16_length_from_latin1(size_t length) const noexcept { 6348| 31.6k| return length; 6349| 31.6k| } _ZNK7simdutf14implementation24latin1_length_from_utf32Em: 6535| 15.8k| latin1_length_from_utf32(size_t length) const noexcept { 6536| 15.8k| return length; 6537| 15.8k| } _ZNK7simdutf14implementation24latin1_length_from_utf16Em: 6572| 31.6k| latin1_length_from_utf16(size_t length) const noexcept { 6573| 31.6k| return length; 6574| 31.6k| } _ZNK7simdutf14implementation24utf32_length_from_latin1Em: 6605| 15.8k| utf32_length_from_latin1(size_t length) const noexcept { 6606| 15.8k| return length; 6607| 15.8k| } _ZN7simdutf14implementationC2EPKcS2_j: 7043| 5| : _name(name), _description(description), 7044| 5| _required_instruction_sets(required_instruction_sets) {} _ZN7simdutf8internal29available_implementation_listC2Ev: 7075| 1| simdutf_really_inline available_implementation_list() {} _ZN7simdutf25base64_length_from_binaryEmNS_14base64_optionsE: 4447| 46.0k| size_t length, base64_options options = base64_default) noexcept { 4448| 46.0k| return scalar::base64::base64_length_from_binary(length, options); 4449| 46.0k|} _ZN7simdutf16base64_ignorableEcNS_14base64_optionsE: 4862| 4.02k|base64_ignorable(char input, base64_options options = base64_default) noexcept { 4863| 4.02k| return scalar::base64::is_ignorable(input, options); 4864| 4.02k|} simdutf.cpp:_ZN7simdutf6detail12_GLOBAL__N_13minEmm: 59| 46.0k|constexpr std::size_t min(std::size_t a, std::size_t b) { 60| 46.0k| return a < b ? a : b; ------------------ | Branch (60:10): [True: 0, False: 46.0k] ------------------ 61| 46.0k|} _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEaSEPS3_: 7146| 1| atomic_ptr &operator=(T *_ptr) { 7147| 1| ptr = _ptr; 7148| 1| return *this; 7149| 1| } _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEcvPS3_Ev: 7143| 1| operator T *() { return ptr.load(); } _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEC2EPS3_: 7123| 1| atomic_ptr(T *_ptr) : ptr{_ptr} {} _ZN7simdutf8internal10atomic_ptrIKNS_14implementationEEptEv: 7145| 46.0k| T *operator->() { return ptr.load(); } simdutf.cpp:_ZN7simdutf8internalL30detect_supported_architecturesEv: 236| 21.0k|static inline uint32_t detect_supported_architectures() { 237| 21.0k| uint32_t eax; 238| 21.0k| uint32_t ebx = 0; 239| 21.0k| uint32_t ecx = 0; 240| 21.0k| uint32_t edx = 0; 241| 21.0k| uint32_t host_isa = 0x0; 242| | 243| | // EBX for EAX=0x1 244| 21.0k| eax = 0x1; 245| 21.0k| cpuid(&eax, &ebx, &ecx, &edx); 246| | 247| 21.0k| if (ecx & cpuid_bit::sse42) { ------------------ | Branch (247:7): [True: 21.0k, False: 0] ------------------ 248| 21.0k| host_isa |= instruction_set::SSE42; 249| 21.0k| } 250| | 251| 21.0k| if (ecx & cpuid_bit::pclmulqdq) { ------------------ | Branch (251:7): [True: 21.0k, False: 0] ------------------ 252| 21.0k| host_isa |= instruction_set::PCLMULQDQ; 253| 21.0k| } 254| | 255| 21.0k| if ((ecx & cpuid_bit::osxsave) != cpuid_bit::osxsave) { ------------------ | Branch (255:7): [True: 0, False: 21.0k] ------------------ 256| 0| return host_isa; 257| 0| } 258| | 259| | // xgetbv for checking if the OS saves registers 260| 21.0k| uint64_t xcr0 = xgetbv(); 261| | 262| 21.0k| if ((xcr0 & cpuid_bit::xcr0_bit::avx256_saved) == 0) { ------------------ | Branch (262:7): [True: 0, False: 21.0k] ------------------ 263| 0| return host_isa; 264| 0| } 265| | // ECX for EAX=0x7 266| 21.0k| eax = 0x7; 267| 21.0k| ecx = 0x0; // Sub-leaf = 0 268| 21.0k| cpuid(&eax, &ebx, &ecx, &edx); 269| 21.0k| if (ebx & cpuid_bit::ebx::avx2) { ------------------ | Branch (269:7): [True: 21.0k, False: 0] ------------------ 270| 21.0k| host_isa |= instruction_set::AVX2; 271| 21.0k| } 272| 21.0k| if (ebx & cpuid_bit::ebx::bmi1) { ------------------ | Branch (272:7): [True: 21.0k, False: 0] ------------------ 273| 21.0k| host_isa |= instruction_set::BMI1; 274| 21.0k| } 275| 21.0k| if (ebx & cpuid_bit::ebx::bmi2) { ------------------ | Branch (275:7): [True: 21.0k, False: 0] ------------------ 276| 21.0k| host_isa |= instruction_set::BMI2; 277| 21.0k| } 278| 21.0k| if (!((xcr0 & cpuid_bit::xcr0_bit::avx512_saved) == ------------------ | Branch (278:7): [True: 21.0k, False: 0] ------------------ 279| 21.0k| cpuid_bit::xcr0_bit::avx512_saved)) { 280| 21.0k| return host_isa; 281| 21.0k| } 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| 21.0k|} simdutf.cpp:_ZN7simdutf8internalL5cpuidEPjS1_S1_S1_: 202| 42.1k| 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| 42.1k| uint32_t a = *eax, b, c = *ecx, d; 216| 42.1k| asm volatile("cpuid\n\t" : "+a"(a), "=b"(b), "+c"(c), "=d"(d)); 217| 42.1k| *eax = a; 218| 42.1k| *ebx = b; 219| 42.1k| *ecx = c; 220| 42.1k| *edx = d; 221| 42.1k| #endif 222| 42.1k|} simdutf.cpp:_ZN7simdutf8internalL6xgetbvEv: 224| 21.0k|static inline uint64_t xgetbv() { 225| | #if defined(_MSC_VER) 226| | return _xgetbv(0); 227| | #elif defined(__FILC__) 228| | return zxgetbv(); 229| | #else 230| 21.0k| uint32_t xcr0_lo, xcr0_hi; 231| 21.0k| asm volatile("xgetbv\n\t" : "=a"(xcr0_lo), "=d"(xcr0_hi) : "c"(0)); 232| 21.0k| return xcr0_lo | ((uint64_t)xcr0_hi << 32); 233| 21.0k| #endif 234| 21.0k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6425base64_length_from_binaryEmNS_14base64_optionsE: 836| 90.6k|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| 90.6k| bool use_padding = 846| 90.6k| ((options & base64_url) == 0) ^ 847| 90.6k| ((options & base64_reverse_padding) == base64_reverse_padding); 848| 90.6k| if (!use_padding) { ------------------ | Branch (848:7): [True: 0, False: 90.6k] ------------------ 849| 0| return length / 3 * 4 + ((length % 3) ? (length % 3) + 1 : 0); ------------------ | Branch (849:30): [True: 0, False: 0] ------------------ 850| 0| } 851| 90.6k| return (length + 2) / 3 * 852| 90.6k| 4; // We use padding to make the length a multiple of 4. 853| 90.6k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6418tail_encode_base64EPcPKcmNS_14base64_optionsE: 648| 15.4k| char *dst, const char *src, size_t srclen, base64_options options) { 649| 15.4k| return tail_encode_base64_impl(dst, src, srclen, options); 650| 15.4k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423tail_encode_base64_implILb0EEEmPcPKcmNS_14base64_optionsEmm: 418| 44.6k| 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| 44.6k| bool use_padding = 437| 44.6k| ((options & base64_url) == 0) ^ 438| 44.6k| ((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| 44.6k| const char *e0 = (options & base64_url) ? tables::base64::base64_url::e0 ------------------ | Branch (441:20): [True: 0, False: 44.6k] ------------------ 442| 44.6k| : tables::base64::base64_default::e0; 443| 44.6k| const char *e1 = (options & base64_url) ? tables::base64::base64_url::e1 ------------------ | Branch (443:20): [True: 0, False: 44.6k] ------------------ 444| 44.6k| : tables::base64::base64_default::e1; 445| 44.6k| const char *e2 = (options & base64_url) ? tables::base64::base64_url::e2 ------------------ | Branch (445:20): [True: 0, False: 44.6k] ------------------ 446| 44.6k| : tables::base64::base64_default::e2; 447| 44.6k| char *out = dst; 448| 44.6k| size_t i = 0; 449| 44.6k| uint8_t t1, t2, t3; 450| 85.4k| for (; i + 2 < srclen; i += 3) { ------------------ | Branch (450:10): [True: 40.7k, False: 44.6k] ------------------ 451| 40.7k| t1 = uint8_t(src[i]); 452| 40.7k| t2 = uint8_t(src[i + 1]); 453| 40.7k| 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| 40.7k| } else { 493| 40.7k| *out++ = e0[t1]; 494| 40.7k| *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 495| 40.7k| *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)]; 496| 40.7k| *out++ = e2[t3]; 497| 40.7k| } 498| 40.7k| } 499| 44.6k| switch (srclen - i) { 500| 42.6k| case 0: ------------------ | Branch (500:3): [True: 42.6k, False: 2.03k] ------------------ 501| 42.6k| break; 502| 1.27k| case 1: ------------------ | Branch (502:3): [True: 1.27k, False: 43.3k] ------------------ 503| 1.27k| 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| 1.27k| } else { 559| 1.27k| *out++ = e0[t1]; 560| 1.27k| *out++ = e1[(t1 & 0x03) << 4]; 561| 1.27k| if (use_padding) { ------------------ | Branch (561:11): [True: 1.27k, False: 0] ------------------ 562| 1.27k| *out++ = '='; 563| 1.27k| *out++ = '='; 564| 1.27k| } 565| 1.27k| } 566| 1.27k| break; 567| 760| default: /* case 2 */ ------------------ | Branch (567:3): [True: 760, False: 43.8k] ------------------ 568| 760| t1 = uint8_t(src[i]); 569| 760| 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| 760| } else { 634| 760| *out++ = e0[t1]; 635| 760| *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)]; 636| 760| *out++ = e2[(t2 & 0x0F) << 2]; 637| 760| if (use_padding) { ------------------ | Branch (637:11): [True: 760, False: 0] ------------------ 638| 760| *out++ = '='; 639| 760| } 640| 760| } 641| 44.6k| } 642| 44.6k| return (size_t)(out - dst); 643| 44.6k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6413is_eight_byteIcEEbT_: 19| 321k|template simdutf_constexpr23 bool is_eight_byte(char_type c) { 20| 321k| if constexpr (sizeof(char_type) == 1) { 21| 321k| return true; 22| 321k| } 23| 0| return uint8_t(c) == c; 24| 321k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6433maximal_binary_length_from_base64IPKcEEmT_m: 654| 76.2k|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| 76.2k| size_t padding = 0; 659| 76.2k| if (length > 0) { ------------------ | Branch (659:7): [True: 7.30k, False: 68.9k] ------------------ 660| 7.30k| if (input[length - 1] == '=') { ------------------ | Branch (660:9): [True: 2.22k, False: 5.08k] ------------------ 661| 2.22k| padding++; 662| 2.22k| if (length > 1 && input[length - 2] == '=') { ------------------ | Branch (662:11): [True: 2.19k, False: 32] | Branch (662:25): [True: 1.32k, False: 864] ------------------ 663| 1.32k| padding++; 664| 1.32k| } 665| 2.22k| } 666| 7.30k| } 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| 76.2k| size_t actual_length = length - padding; 686| 76.2k| if (actual_length % 4 <= 1) { ------------------ | Branch (686:7): [True: 73.2k, False: 2.98k] ------------------ 687| 73.2k| return actual_length / 4 * 3; 688| 73.2k| } 689| | // if we have a valid input, then the remainder must be 2 or 3 adding one or 690| | // two extra bytes. 691| 2.98k| return actual_length / 4 * 3 + (actual_length % 4) - 1; 692| 76.2k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6412is_ignorableIcEEbT_NS_14base64_optionsE: 28| 4.02k| simdutf::base64_options options) { 29| 4.02k| const uint8_t *to_base64 = 30| 4.02k| (options & base64_default_or_url) ------------------ | Branch (30:7): [True: 0, False: 4.02k] ------------------ 31| 4.02k| ? tables::base64::to_base64_default_or_url_value 32| 4.02k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (32:14): [True: 0, False: 4.02k] ------------------ 33| 4.02k| : tables::base64::to_base64_value); 34| 4.02k| const bool ignore_garbage = 35| 4.02k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (35:7): [True: 0, False: 4.02k] ------------------ 36| 4.02k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (36:7): [True: 0, False: 4.02k] ------------------ 37| 4.02k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (37:7): [True: 0, False: 4.02k] ------------------ 38| 4.02k| uint8_t code = to_base64[uint8_t(c)]; 39| 4.02k| if (is_eight_byte(c) && code <= 63) { ------------------ | Branch (39:7): [True: 4.02k, False: 0] | Branch (39:27): [True: 469, False: 3.55k] ------------------ 40| 469| return false; 41| 469| } 42| 3.55k| if (is_eight_byte(c) && code == 64) { ------------------ | Branch (42:7): [True: 3.55k, False: 0] | Branch (42:27): [True: 3.55k, False: 0] ------------------ 43| 3.55k| return true; 44| 3.55k| } 45| 0| return ignore_garbage; 46| 3.55k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base648find_endIcEENS2_13reduced_inputEPKT_mNS_14base64_optionsE: 100| 94.3k| simdutf::base64_options options) { 101| 94.3k| const uint8_t *to_base64 = 102| 94.3k| (options & base64_default_or_url) ------------------ | Branch (102:7): [True: 0, False: 94.3k] ------------------ 103| 94.3k| ? tables::base64::to_base64_default_or_url_value 104| 94.3k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (104:14): [True: 0, False: 94.3k] ------------------ 105| 94.3k| : tables::base64::to_base64_value); 106| 94.3k| const bool ignore_garbage = 107| 94.3k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (107:7): [True: 0, False: 94.3k] ------------------ 108| 94.3k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (108:7): [True: 0, False: 94.3k] ------------------ 109| 94.3k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (109:7): [True: 0, False: 94.3k] ------------------ 110| | 111| 94.3k| 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| 94.3k| size_t full_input_length = srclen; 115| | // skip trailing spaces 116| 103k| while (!ignore_garbage && srclen > 0 && ------------------ | Branch (116:10): [True: 103k, False: 0] | Branch (116:29): [True: 19.3k, False: 83.7k] ------------------ 117| 19.3k| scalar::base64::is_eight_byte(src[srclen - 1]) && ------------------ | Branch (117:10): [True: 19.3k, False: 0] ------------------ 118| 19.3k| to_base64[uint8_t(src[srclen - 1])] == 64) { ------------------ | Branch (118:10): [True: 8.69k, False: 10.6k] ------------------ 119| 8.69k| srclen--; 120| 8.69k| } 121| 94.3k| size_t equallocation = 122| 94.3k| srclen; // location of the first padding character if any 123| 94.3k| if (ignore_garbage) { ------------------ | Branch (123:7): [True: 0, False: 94.3k] ------------------ 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| 94.3k| if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') { ------------------ | Branch (135:7): [True: 94.3k, False: 0] | Branch (135:26): [True: 10.6k, False: 83.7k] | Branch (135:40): [True: 3.58k, False: 7.04k] ------------------ 136| | // This is the last '=' sign. 137| 3.58k| equallocation = srclen - 1; 138| 3.58k| srclen--; 139| 3.58k| equalsigns = 1; 140| | // skip trailing spaces 141| 4.02k| while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) && ------------------ | Branch (141:12): [True: 3.97k, False: 51] | Branch (141:26): [True: 3.97k, False: 0] ------------------ 142| 3.97k| to_base64[uint8_t(src[srclen - 1])] == 64) { ------------------ | Branch (142:12): [True: 441, False: 3.53k] ------------------ 143| 441| srclen--; 144| 441| } 145| 3.58k| if (srclen > 0 && src[srclen - 1] == '=') { ------------------ | Branch (145:9): [True: 3.53k, False: 51] | Branch (145:23): [True: 2.23k, False: 1.30k] ------------------ 146| | // This is the second '=' sign. 147| 2.23k| equallocation = srclen - 1; 148| 2.23k| srclen--; 149| 2.23k| equalsigns = 2; 150| 2.23k| } 151| 3.58k| } 152| 94.3k| return {equalsigns, equallocation, srclen, full_input_length}; 153| 94.3k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423base64_tail_decode_safeIcEENS_11full_resultEPcmPKT_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 385| 2.23k| base64_options options, last_chunk_handling_options last_chunk_options) { 386| 2.23k| return base64_tail_decode_impl(dst, outlen, src, length, 387| 2.23k| padding_characters, options, 388| 2.23k| last_chunk_options); 389| 2.23k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423base64_tail_decode_implILb1EcEENS_11full_resultEPcmPKT0_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 164| 2.23k| base64_options options, last_chunk_handling_options last_chunk_options) { 165| 2.23k| char *dstend = dst + outlen; 166| 2.23k| (void)dstend; 167| | // This looks like 10 branches, but we expect the compiler to resolve this to 168| | // two branches (easily predicted): 169| 2.23k| const uint8_t *to_base64 = 170| 2.23k| (options & base64_default_or_url) ------------------ | Branch (170:7): [True: 0, False: 2.23k] ------------------ 171| 2.23k| ? tables::base64::to_base64_default_or_url_value 172| 2.23k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (172:14): [True: 0, False: 2.23k] ------------------ 173| 2.23k| : tables::base64::to_base64_value); 174| 2.23k| const uint32_t *d0 = 175| 2.23k| (options & base64_default_or_url) ------------------ | Branch (175:7): [True: 0, False: 2.23k] ------------------ 176| 2.23k| ? tables::base64::base64_default_or_url::d0 177| 2.23k| : ((options & base64_url) ? tables::base64::base64_url::d0 ------------------ | Branch (177:14): [True: 0, False: 2.23k] ------------------ 178| 2.23k| : tables::base64::base64_default::d0); 179| 2.23k| const uint32_t *d1 = 180| 2.23k| (options & base64_default_or_url) ------------------ | Branch (180:7): [True: 0, False: 2.23k] ------------------ 181| 2.23k| ? tables::base64::base64_default_or_url::d1 182| 2.23k| : ((options & base64_url) ? tables::base64::base64_url::d1 ------------------ | Branch (182:14): [True: 0, False: 2.23k] ------------------ 183| 2.23k| : tables::base64::base64_default::d1); 184| 2.23k| const uint32_t *d2 = 185| 2.23k| (options & base64_default_or_url) ------------------ | Branch (185:7): [True: 0, False: 2.23k] ------------------ 186| 2.23k| ? tables::base64::base64_default_or_url::d2 187| 2.23k| : ((options & base64_url) ? tables::base64::base64_url::d2 ------------------ | Branch (187:14): [True: 0, False: 2.23k] ------------------ 188| 2.23k| : tables::base64::base64_default::d2); 189| 2.23k| const uint32_t *d3 = 190| 2.23k| (options & base64_default_or_url) ------------------ | Branch (190:7): [True: 0, False: 2.23k] ------------------ 191| 2.23k| ? tables::base64::base64_default_or_url::d3 192| 2.23k| : ((options & base64_url) ? tables::base64::base64_url::d3 ------------------ | Branch (192:14): [True: 0, False: 2.23k] ------------------ 193| 2.23k| : tables::base64::base64_default::d3); 194| 2.23k| const bool ignore_garbage = 195| 2.23k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (195:7): [True: 0, False: 2.23k] ------------------ 196| 2.23k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (196:7): [True: 0, False: 2.23k] ------------------ 197| 2.23k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (197:7): [True: 0, False: 2.23k] ------------------ 198| | 199| 2.23k| const char_type *srcend = src + length; 200| 2.23k| const char_type *srcinit = src; 201| 2.23k| const char *dstinit = dst; 202| | 203| 2.23k| uint32_t x; 204| 2.23k| size_t idx; 205| 2.23k| uint8_t buffer[4]; 206| 2.27k| while (true) { ------------------ | Branch (206:10): [True: 2.27k, Folded] ------------------ 207| 2.34k| while (srcend - src >= 4 && is_eight_byte(src[0]) && ------------------ | Branch (207:12): [True: 249, False: 2.09k] | 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: 69, False: 180] ------------------ 211| 249| d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) { 212| 69| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (212:11): [True: 69, Folded] | Branch (212:29): [True: 0, False: 69] ------------------ 213| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(src - srcinit), 214| 0| size_t(dst - dstinit)}; 215| 0| } 216| 69| *dst++ = static_cast(x & 0xFF); 217| 69| *dst++ = static_cast((x >> 8) & 0xFF); 218| 69| *dst++ = static_cast((x >> 16) & 0xFF); 219| 69| src += 4; 220| 69| } 221| 2.27k| const char_type *srccur = src; 222| 2.27k| idx = 0; 223| | // we need at least four characters. 224| 2.27k|#ifdef __clang__ 225| | // If possible, we read four characters at a time. (It is an optimization.) 226| 2.27k| if (ignore_garbage && src + 4 <= srcend) { ------------------ | Branch (226:9): [True: 0, False: 2.27k] | 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| 2.27k|#endif 248| 12.9k| while ((idx < 4) && (src < srcend)) { ------------------ | Branch (248:12): [True: 12.9k, False: 38] | Branch (248:25): [True: 10.8k, False: 2.05k] ------------------ 249| 10.8k| char_type c = *src; 250| | 251| 10.8k| uint8_t code = to_base64[uint8_t(c)]; 252| 10.8k| buffer[idx] = uint8_t(code); 253| 10.8k| if (is_eight_byte(c) && code <= 63) { ------------------ | Branch (253:11): [True: 10.8k, False: 0] | Branch (253:31): [True: 4.76k, False: 6.12k] ------------------ 254| 4.76k| idx++; 255| 6.12k| } else if (!ignore_garbage && ------------------ | Branch (255:18): [True: 6.12k, False: 0] ------------------ 256| 6.12k| (code > 64 || !scalar::base64::is_eight_byte(c))) { ------------------ | Branch (256:19): [True: 175, False: 5.94k] | Branch (256:32): [True: 0, False: 5.94k] ------------------ 257| 175| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 258| 175| size_t(dst - dstinit)}; 259| 5.94k| } else { 260| | // We have a space or a newline or garbage. We ignore it. 261| 5.94k| } 262| 10.7k| src++; 263| 10.7k| } 264| 2.09k| if (idx != 4) { ------------------ | Branch (264:9): [True: 2.05k, False: 38] ------------------ 265| 2.05k| 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| 2.05k| if (!ignore_garbage && (idx + padding_characters > 4)) { ------------------ | Branch (268:11): [True: 2.05k, False: 0] | Branch (268:30): [True: 7, False: 2.05k] ------------------ 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| 2.05k| if (!ignore_garbage && ------------------ | Branch (277:11): [True: 2.05k, False: 0] ------------------ 278| 2.05k| last_chunk_options == last_chunk_handling_options::loose && ------------------ | Branch (278:11): [True: 2.05k, False: 0] ------------------ 279| 2.05k| (idx >= 2) && padding_characters > 0 && ------------------ | Branch (279:11): [True: 1.87k, False: 179] | Branch (279:25): [True: 1.69k, False: 181] ------------------ 280| 1.69k| ((idx + padding_characters) & 3) != 0) { ------------------ | Branch (280:11): [True: 13, False: 1.67k] ------------------ 281| 13| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 282| 13| size_t(dst - dstinit), true}; 283| 13| } 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| 2.03k| if (!ignore_garbage && ------------------ | Branch (288:13): [True: 2.03k, False: 0] ------------------ 289| 2.03k| last_chunk_options == last_chunk_handling_options::strict && ------------------ | Branch (289:13): [True: 0, False: 2.03k] ------------------ 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| 2.03k| if ((last_chunk_options == ------------------ | Branch (298:16): [True: 0, False: 2.03k] ------------------ 299| 2.03k| 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| 2.03k| (last_chunk_options == ------------------ | Branch (302:16): [True: 0, False: 2.03k] ------------------ 303| 2.03k| 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| 2.03k| } else { 310| 2.03k| if (idx == 2) { ------------------ | Branch (310:17): [True: 1.20k, False: 834] ------------------ 311| 1.20k| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 312| 1.20k| (uint32_t(buffer[1]) << 2 * 6); 313| 1.20k| if (!ignore_garbage && ------------------ | Branch (313:19): [True: 1.20k, False: 0] ------------------ 314| 1.20k| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (314:19): [True: 0, False: 1.20k] ------------------ 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| 1.20k| if (check_capacity && dstend - dst < 1) { ------------------ | Branch (319:19): [True: 1.20k, Folded] | Branch (319:37): [True: 0, False: 1.20k] ------------------ 320| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 321| 0| size_t(dst - dstinit)}; 322| 0| } 323| 1.20k| *dst++ = static_cast((triple >> 16) & 0xFF); 324| 1.20k| } else if (idx == 3) { ------------------ | Branch (324:24): [True: 655, False: 179] ------------------ 325| 655| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 326| 655| (uint32_t(buffer[1]) << 2 * 6) + 327| 655| (uint32_t(buffer[2]) << 1 * 6); 328| 655| if (!ignore_garbage && ------------------ | Branch (328:19): [True: 655, False: 0] ------------------ 329| 655| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (329:19): [True: 0, False: 655] ------------------ 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| 655| if (check_capacity && dstend - dst < 2) { ------------------ | Branch (334:19): [True: 655, Folded] | Branch (334:37): [True: 0, False: 655] ------------------ 335| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 336| 0| size_t(dst - dstinit)}; 337| 0| } 338| 655| *dst++ = static_cast((triple >> 16) & 0xFF); 339| 655| *dst++ = static_cast((triple >> 8) & 0xFF); 340| 655| } else if (!ignore_garbage && idx == 1 && ------------------ | Branch (340:24): [True: 179, False: 0] | Branch (340:43): [True: 82, False: 97] ------------------ 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| 97| } else if (!ignore_garbage && idx == 0 && padding_characters > 0) { ------------------ | Branch (346:24): [True: 97, False: 0] | Branch (346:43): [True: 97, False: 0] | Branch (346:55): [True: 7, False: 90] ------------------ 347| 7| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 348| 7| size_t(dst - dstinit), true}; 349| 7| } 350| 1.94k| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 351| 2.03k| } 352| 2.05k| } 353| 38| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (353:9): [True: 38, Folded] | Branch (353:27): [True: 0, False: 38] ------------------ 354| 0| return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit), 355| 0| size_t(dst - dstinit)}; 356| 0| } 357| 38| uint32_t triple = 358| 38| (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6) + 359| 38| (uint32_t(buffer[2]) << 1 * 6) + (uint32_t(buffer[3]) << 0 * 6); 360| 38| *dst++ = static_cast((triple >> 16) & 0xFF); 361| 38| *dst++ = static_cast((triple >> 8) & 0xFF); 362| 38| *dst++ = static_cast(triple & 0xFF); 363| 38| } 364| 2.23k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6417patch_tail_resultENS_11full_resultEmmmmNS_27last_chunk_handling_optionsE: 394| 8.97k| last_chunk_handling_options last_chunk_options) { 395| 8.97k| r.input_count += previous_input; 396| 8.97k| r.output_count += previous_output; 397| 8.97k| if (r.padding_error) { ------------------ | Branch (397:7): [True: 92, False: 8.87k] ------------------ 398| 92| r.input_count = equallocation; 399| 92| } 400| | 401| 8.97k| if (r.error == error_code::SUCCESS) { ------------------ | Branch (401:7): [True: 7.32k, False: 1.65k] ------------------ 402| 7.32k| if (!is_partial(last_chunk_options)) { ------------------ | Branch (402:9): [True: 5.69k, False: 1.62k] ------------------ 403| | // A success when we are not in stop_before_partial mode. 404| | // means that we have consumed the whole input buffer. 405| 5.69k| r.input_count = full_input_length; 406| 5.69k| } else if (r.output_count % 3 != 0) { ------------------ | Branch (406:16): [True: 0, False: 1.62k] ------------------ 407| 0| r.input_count = full_input_length; 408| 0| } 409| 7.32k| } 410| 8.97k| return r; 411| 8.97k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6434base64_to_binary_details_safe_implIcEENS_11full_resultEPKT_mPcmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 783| 2.28k| last_chunk_handling_options last_chunk_options) noexcept { 784| 2.28k| const bool ignore_garbage = 785| 2.28k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (785:7): [True: 0, False: 2.28k] ------------------ 786| 2.28k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (786:7): [True: 0, False: 2.28k] ------------------ 787| 2.28k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (787:7): [True: 0, False: 2.28k] ------------------ 788| 2.28k| auto ri = simdutf::scalar::base64::find_end(input, length, options); 789| 2.28k| size_t equallocation = ri.equallocation; 790| 2.28k| size_t equalsigns = ri.equalsigns; 791| 2.28k| length = ri.srclen; 792| 2.28k| size_t full_input_length = ri.full_input_length; 793| 2.28k| if (length == 0) { ------------------ | Branch (793:7): [True: 52, False: 2.23k] ------------------ 794| 52| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (794:9): [True: 52, False: 0] | Branch (794:28): [True: 20, False: 32] ------------------ 795| 20| return {INVALID_BASE64_CHARACTER, equallocation, 0}; 796| 20| } 797| 32| return {SUCCESS, full_input_length, 0}; 798| 52| } 799| 2.23k| full_result r = scalar::base64::base64_tail_decode_safe( 800| 2.23k| output, outlen, input, length, equalsigns, options, last_chunk_options); 801| 2.23k| r = scalar::base64::patch_tail_result(r, 0, 0, equallocation, 802| 2.23k| full_input_length, last_chunk_options); 803| 2.23k| if (!is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (803:7): [True: 2.23k, False: 0] | Branch (803:42): [True: 1.94k, False: 284] ------------------ 804| 1.94k| equalsigns > 0 && !ignore_garbage) { ------------------ | Branch (804:7): [True: 1.67k, False: 271] | Branch (804:25): [True: 1.67k, False: 0] ------------------ 805| | // additional checks 806| 1.67k| if ((r.output_count % 3 == 0) || ------------------ | Branch (806:9): [True: 0, False: 1.67k] ------------------ 807| 1.67k| ((r.output_count % 3) + 1 + equalsigns != 4)) { ------------------ | Branch (807:9): [True: 0, False: 1.67k] ------------------ 808| 0| return {INVALID_BASE64_CHARACTER, equallocation, r.output_count}; 809| 0| } 810| 1.67k| } 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| 2.23k| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (816:7): [True: 0, False: 2.23k] | 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| 2.23k| return r; 833| 2.23k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6429base64_to_binary_details_implIcEENS_11full_resultEPKT_mPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 727| 15.6k| last_chunk_handling_options last_chunk_options) noexcept { 728| 15.6k| const bool ignore_garbage = 729| 15.6k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (729:7): [True: 0, False: 15.6k] ------------------ 730| 15.6k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (730:7): [True: 0, False: 15.6k] ------------------ 731| 15.6k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (731:7): [True: 0, False: 15.6k] ------------------ 732| 15.6k| auto ri = simdutf::scalar::base64::find_end(input, length, options); 733| 15.6k| size_t equallocation = ri.equallocation; 734| 15.6k| size_t equalsigns = ri.equalsigns; 735| 15.6k| length = ri.srclen; 736| 15.6k| size_t full_input_length = ri.full_input_length; 737| 15.6k| if (length == 0) { ------------------ | Branch (737:7): [True: 15.2k, False: 393] ------------------ 738| 15.2k| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (738:9): [True: 15.2k, False: 0] | Branch (738:28): [True: 5, False: 15.2k] ------------------ 739| 5| return {INVALID_BASE64_CHARACTER, equallocation, 0, true}; 740| 5| } 741| 15.2k| return {SUCCESS, full_input_length, 0}; 742| 15.2k| } 743| 393| full_result r = scalar::base64::base64_tail_decode( 744| 393| output, input, length, equalsigns, options, last_chunk_options); 745| 393| r = scalar::base64::patch_tail_result(r, 0, 0, equallocation, 746| 393| full_input_length, last_chunk_options); 747| 393| if (!is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (747:7): [True: 393, False: 0] | Branch (747:42): [True: 219, False: 174] ------------------ 748| 219| equalsigns > 0 && !ignore_garbage) { ------------------ | Branch (748:7): [True: 117, False: 102] | Branch (748:25): [True: 117, False: 0] ------------------ 749| | // additional checks 750| 117| if ((r.output_count % 3 == 0) || ------------------ | Branch (750:9): [True: 0, False: 117] ------------------ 751| 117| ((r.output_count % 3) + 1 + equalsigns != 4)) { ------------------ | Branch (751:9): [True: 0, False: 117] ------------------ 752| 0| return {INVALID_BASE64_CHARACTER, equallocation, r.output_count, true}; 753| 0| } 754| 117| } 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| 393| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (759:7): [True: 0, False: 393] | 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| 393| return r; 776| 393|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6418base64_tail_decodeIcEENS_11full_resultEPcPKT_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 371| 6.73k| base64_options options, last_chunk_handling_options last_chunk_options) { 372| 6.73k| return base64_tail_decode_impl(dst, 0, src, length, padding_characters, 373| 6.73k| options, last_chunk_options); 374| 6.73k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16base6423base64_tail_decode_implILb0EcEENS_11full_resultEPcmPKT0_mmNS_14base64_optionsENS_27last_chunk_handling_optionsE: 164| 6.73k| base64_options options, last_chunk_handling_options last_chunk_options) { 165| 6.73k| char *dstend = dst + outlen; 166| 6.73k| (void)dstend; 167| | // This looks like 10 branches, but we expect the compiler to resolve this to 168| | // two branches (easily predicted): 169| 6.73k| const uint8_t *to_base64 = 170| 6.73k| (options & base64_default_or_url) ------------------ | Branch (170:7): [True: 0, False: 6.73k] ------------------ 171| 6.73k| ? tables::base64::to_base64_default_or_url_value 172| 6.73k| : ((options & base64_url) ? tables::base64::to_base64_url_value ------------------ | Branch (172:14): [True: 0, False: 6.73k] ------------------ 173| 6.73k| : tables::base64::to_base64_value); 174| 6.73k| const uint32_t *d0 = 175| 6.73k| (options & base64_default_or_url) ------------------ | Branch (175:7): [True: 0, False: 6.73k] ------------------ 176| 6.73k| ? tables::base64::base64_default_or_url::d0 177| 6.73k| : ((options & base64_url) ? tables::base64::base64_url::d0 ------------------ | Branch (177:14): [True: 0, False: 6.73k] ------------------ 178| 6.73k| : tables::base64::base64_default::d0); 179| 6.73k| const uint32_t *d1 = 180| 6.73k| (options & base64_default_or_url) ------------------ | Branch (180:7): [True: 0, False: 6.73k] ------------------ 181| 6.73k| ? tables::base64::base64_default_or_url::d1 182| 6.73k| : ((options & base64_url) ? tables::base64::base64_url::d1 ------------------ | Branch (182:14): [True: 0, False: 6.73k] ------------------ 183| 6.73k| : tables::base64::base64_default::d1); 184| 6.73k| const uint32_t *d2 = 185| 6.73k| (options & base64_default_or_url) ------------------ | Branch (185:7): [True: 0, False: 6.73k] ------------------ 186| 6.73k| ? tables::base64::base64_default_or_url::d2 187| 6.73k| : ((options & base64_url) ? tables::base64::base64_url::d2 ------------------ | Branch (187:14): [True: 0, False: 6.73k] ------------------ 188| 6.73k| : tables::base64::base64_default::d2); 189| 6.73k| const uint32_t *d3 = 190| 6.73k| (options & base64_default_or_url) ------------------ | Branch (190:7): [True: 0, False: 6.73k] ------------------ 191| 6.73k| ? tables::base64::base64_default_or_url::d3 192| 6.73k| : ((options & base64_url) ? tables::base64::base64_url::d3 ------------------ | Branch (192:14): [True: 0, False: 6.73k] ------------------ 193| 6.73k| : tables::base64::base64_default::d3); 194| 6.73k| const bool ignore_garbage = 195| 6.73k| (options == base64_options::base64_url_accept_garbage) || ------------------ | Branch (195:7): [True: 0, False: 6.73k] ------------------ 196| 6.73k| (options == base64_options::base64_default_accept_garbage) || ------------------ | Branch (196:7): [True: 0, False: 6.73k] ------------------ 197| 6.73k| (options == base64_options::base64_default_or_url_accept_garbage); ------------------ | Branch (197:7): [True: 0, False: 6.73k] ------------------ 198| | 199| 6.73k| const char_type *srcend = src + length; 200| 6.73k| const char_type *srcinit = src; 201| 6.73k| const char *dstinit = dst; 202| | 203| 6.73k| uint32_t x; 204| 6.73k| size_t idx; 205| 6.73k| uint8_t buffer[4]; 206| 7.46k| while (true) { ------------------ | Branch (206:10): [True: 7.46k, Folded] ------------------ 207| 65.6k| while (srcend - src >= 4 && is_eight_byte(src[0]) && ------------------ | Branch (207:12): [True: 60.1k, False: 5.45k] | Branch (207:33): [True: 60.1k, False: 0] ------------------ 208| 60.1k| is_eight_byte(src[1]) && is_eight_byte(src[2]) && ------------------ | Branch (208:12): [True: 60.1k, False: 0] | Branch (208:37): [True: 60.1k, False: 0] ------------------ 209| 60.1k| is_eight_byte(src[3]) && ------------------ | Branch (209:12): [True: 60.1k, False: 0] ------------------ 210| 60.1k| (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] | ------------------ | Branch (210:12): [True: 58.1k, False: 2.00k] ------------------ 211| 60.1k| d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) { 212| 58.1k| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (212:11): [Folded, False: 58.1k] | 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| 58.1k| *dst++ = static_cast(x & 0xFF); 217| 58.1k| *dst++ = static_cast((x >> 8) & 0xFF); 218| 58.1k| *dst++ = static_cast((x >> 16) & 0xFF); 219| 58.1k| src += 4; 220| 58.1k| } 221| 7.46k| const char_type *srccur = src; 222| 7.46k| idx = 0; 223| | // we need at least four characters. 224| 7.46k|#ifdef __clang__ 225| | // If possible, we read four characters at a time. (It is an optimization.) 226| 7.46k| if (ignore_garbage && src + 4 <= srcend) { ------------------ | Branch (226:9): [True: 0, False: 7.46k] | 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| 7.46k|#endif 248| 26.8k| while ((idx < 4) && (src < srcend)) { ------------------ | Branch (248:12): [True: 26.1k, False: 727] | Branch (248:25): [True: 20.4k, False: 5.64k] ------------------ 249| 20.4k| char_type c = *src; 250| | 251| 20.4k| uint8_t code = to_base64[uint8_t(c)]; 252| 20.4k| buffer[idx] = uint8_t(code); 253| 20.4k| if (is_eight_byte(c) && code <= 63) { ------------------ | Branch (253:11): [True: 20.4k, False: 0] | Branch (253:31): [True: 9.42k, False: 11.0k] ------------------ 254| 9.42k| idx++; 255| 11.0k| } else if (!ignore_garbage && ------------------ | Branch (255:18): [True: 11.0k, False: 0] ------------------ 256| 11.0k| (code > 64 || !scalar::base64::is_eight_byte(c))) { ------------------ | Branch (256:19): [True: 1.08k, False: 9.96k] | Branch (256:32): [True: 0, False: 9.96k] ------------------ 257| 1.08k| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 258| 1.08k| size_t(dst - dstinit)}; 259| 9.96k| } else { 260| | // We have a space or a newline or garbage. We ignore it. 261| 9.96k| } 262| 19.3k| src++; 263| 19.3k| } 264| 6.37k| if (idx != 4) { ------------------ | Branch (264:9): [True: 5.64k, False: 727] ------------------ 265| 5.64k| 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| 5.64k| if (!ignore_garbage && (idx + padding_characters > 4)) { ------------------ | Branch (268:11): [True: 5.64k, False: 0] | Branch (268:30): [True: 15, False: 5.63k] ------------------ 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| 5.63k| if (!ignore_garbage && ------------------ | Branch (277:11): [True: 5.63k, False: 0] ------------------ 278| 5.63k| last_chunk_options == last_chunk_handling_options::loose && ------------------ | Branch (278:11): [True: 4.00k, False: 1.62k] ------------------ 279| 4.00k| (idx >= 2) && padding_characters > 0 && ------------------ | Branch (279:11): [True: 2.24k, False: 1.75k] | Branch (279:25): [True: 1.71k, False: 535] ------------------ 280| 1.71k| ((idx + padding_characters) & 3) != 0) { ------------------ | Branch (280:11): [True: 27, False: 1.68k] ------------------ 281| 27| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 282| 27| size_t(dst - dstinit), true}; 283| 27| } 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| 5.60k| if (!ignore_garbage && ------------------ | Branch (288:13): [True: 5.60k, False: 0] ------------------ 289| 5.60k| last_chunk_options == last_chunk_handling_options::strict && ------------------ | Branch (289:13): [True: 0, False: 5.60k] ------------------ 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| 5.60k| if ((last_chunk_options == ------------------ | Branch (298:16): [True: 0, False: 5.60k] ------------------ 299| 5.60k| 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| 5.60k| (last_chunk_options == ------------------ | Branch (302:16): [True: 1.62k, False: 3.97k] ------------------ 303| 5.60k| last_chunk_handling_options::only_full_chunks && 304| 1.62k| (idx >= 2 || padding_characters == 0))) { ------------------ | Branch (304:17): [True: 197, False: 1.43k] | Branch (304:29): [True: 1.42k, False: 6] ------------------ 305| | // partial means that we are *not* going to consume the read 306| | // characters. We need to rewind the src pointer. 307| 1.62k| src = srccur; 308| 1.62k| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 309| 3.98k| } else { 310| 3.98k| if (idx == 2) { ------------------ | Branch (310:17): [True: 1.35k, False: 2.63k] ------------------ 311| 1.35k| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 312| 1.35k| (uint32_t(buffer[1]) << 2 * 6); 313| 1.35k| if (!ignore_garbage && ------------------ | Branch (313:19): [True: 1.35k, False: 0] ------------------ 314| 1.35k| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (314:19): [True: 0, False: 1.35k] ------------------ 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| 1.35k| if (check_capacity && dstend - dst < 1) { ------------------ | Branch (319:19): [Folded, False: 1.35k] | 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| 1.35k| *dst++ = static_cast((triple >> 16) & 0xFF); 324| 2.63k| } else if (idx == 3) { ------------------ | Branch (324:24): [True: 865, False: 1.76k] ------------------ 325| 865| uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) + 326| 865| (uint32_t(buffer[1]) << 2 * 6) + 327| 865| (uint32_t(buffer[2]) << 1 * 6); 328| 865| if (!ignore_garbage && ------------------ | Branch (328:19): [True: 865, False: 0] ------------------ 329| 865| (last_chunk_options == last_chunk_handling_options::strict) && ------------------ | Branch (329:19): [True: 0, False: 865] ------------------ 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| 865| if (check_capacity && dstend - dst < 2) { ------------------ | Branch (334:19): [Folded, False: 865] | 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| 865| *dst++ = static_cast((triple >> 16) & 0xFF); 339| 865| *dst++ = static_cast((triple >> 8) & 0xFF); 340| 1.76k| } else if (!ignore_garbage && idx == 1 && ------------------ | Branch (340:24): [True: 1.76k, False: 0] | Branch (340:43): [True: 212, False: 1.55k] ------------------ 341| 212| (!is_partial(last_chunk_options) || ------------------ | Branch (341:25): [True: 206, False: 6] ------------------ 342| 6| (is_partial(last_chunk_options) && ------------------ | Branch (342:26): [True: 6, False: 0] ------------------ 343| 212| padding_characters > 0))) { ------------------ | Branch (343:26): [True: 6, False: 0] ------------------ 344| 212| return {BASE64_INPUT_REMAINDER, size_t(src - srcinit), 345| 212| size_t(dst - dstinit)}; 346| 1.55k| } else if (!ignore_garbage && idx == 0 && padding_characters > 0) { ------------------ | Branch (346:24): [True: 1.55k, False: 0] | Branch (346:43): [True: 1.55k, False: 0] | Branch (346:55): [True: 23, False: 1.53k] ------------------ 347| 23| return {INVALID_BASE64_CHARACTER, size_t(src - srcinit), 348| 23| size_t(dst - dstinit), true}; 349| 23| } 350| 3.74k| return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)}; 351| 3.98k| } 352| 5.63k| } 353| 727| if (check_capacity && dstend - dst < 3) { ------------------ | Branch (353:9): [Folded, False: 727] | 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| 727| uint32_t triple = 358| 727| (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6) + 359| 727| (uint32_t(buffer[2]) << 1 * 6) + (uint32_t(buffer[3]) << 0 * 6); 360| 727| *dst++ = static_cast((triple >> 16) & 0xFF); 361| 727| *dst++ = static_cast((triple >> 8) & 0xFF); 362| 727| *dst++ = static_cast(triple & 0xFF); 363| 727| } 364| 6.73k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_16latin123utf8_length_from_latin1EPKcm: 10| 11.0k| size_t len) { 11| 11.0k| const uint8_t *c = reinterpret_cast(buf); 12| 11.0k| size_t answer = 0; 13| 25.2k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (13:22): [True: 14.1k, False: 11.0k] ------------------ 14| 14.1k| if ((c[i] >> 7)) { ------------------ | Branch (14:9): [True: 3.58k, False: 10.5k] ------------------ 15| 3.58k| answer++; 16| 3.58k| } 17| 14.1k| } 18| 11.0k| return answer + len; 19| 11.0k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115latin1_to_utf167convertILNS_10endiannessE0EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 14| 6.51k| char16_t *utf16_output) { 15| 6.51k| size_t pos = 0; 16| 6.51k| char16_t *start{utf16_output}; 17| | 18| 101k| while (pos < len) { ------------------ | Branch (18:10): [True: 95.4k, False: 6.51k] ------------------ 19| 95.4k| uint16_t word = 20| 95.4k| uint8_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point 21| 95.4k| *utf16_output++ = 22| 95.4k| char16_t(match_system(big_endian) ? word : u16_swap_bytes(word)); ------------------ | Branch (22:18): [True: 95.4k, Folded] ------------------ 23| 95.4k| pos++; 24| 95.4k| } 25| | 26| 6.51k| return utf16_output - start; 27| 6.51k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115latin1_to_utf167convertILNS_10endiannessE1EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 14| 6.51k| char16_t *utf16_output) { 15| 6.51k| size_t pos = 0; 16| 6.51k| char16_t *start{utf16_output}; 17| | 18| 101k| while (pos < len) { ------------------ | Branch (18:10): [True: 95.4k, False: 6.51k] ------------------ 19| 95.4k| uint16_t word = 20| 95.4k| uint8_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point 21| 95.4k| *utf16_output++ = 22| 95.4k| char16_t(match_system(big_endian) ? word : u16_swap_bytes(word)); ------------------ | Branch (22:18): [Folded, False: 95.4k] ------------------ 23| 95.4k| pos++; 24| 95.4k| } 25| | 26| 6.51k| return utf16_output - start; 27| 6.51k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115latin1_to_utf327convertIPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_mPDi: 14| 6.47k| char32_t *utf32_output) { 15| 6.47k| char32_t *start{utf32_output}; 16| 100k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (16:22): [True: 93.8k, False: 6.47k] ------------------ 17| 93.8k| *utf32_output++ = uint8_t(data[i]); 18| 93.8k| } 19| 6.47k| return utf32_output - start; 20| 6.47k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf87convertEPKcmPc: 66| 7.60k| char *utf8_output) { 67| 7.60k| return convert(reinterpret_cast(buf), len, 68| 7.60k| utf8_output); 69| 7.60k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf87convertIPKhPcQaasr7simdutf6detailE22indexes_into_byte_likeIT_Esr7simdutf6detailE26index_assignable_from_charIT0_EEEmS7_mS8_: 17| 7.60k| OutputPtr utf8_output) { 18| | // const unsigned char *data = reinterpret_cast(buf); 19| 7.60k| size_t pos = 0; 20| 7.60k| size_t utf8_pos = 0; 21| | 22| 79.0k| while (pos < len) { ------------------ | Branch (22:10): [True: 71.4k, False: 7.60k] ------------------ 23| |#if SIMDUTF_CPLUSPLUS23 24| | if !consteval 25| |#endif 26| 71.4k| { 27| | // try to convert the next block of 16 ASCII bytes 28| 71.4k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (28:11): [True: 53.9k, False: 17.5k] ------------------ 29| | // they are ascii 30| 53.9k| uint64_t v1; 31| 53.9k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 32| 53.9k| uint64_t v2; 33| 53.9k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 34| 53.9k| uint64_t v{v1 | 35| 53.9k| v2}; // We are only interested in these bits: 1000 1000 1000 36| | // 1000, so it makes sense to concatenate everything 37| 53.9k| if ((v & 0x8080808080808080) == ------------------ | Branch (37:13): [True: 5.20k, False: 48.7k] ------------------ 38| 53.9k| 0) { // if NONE of these are set, e.g. all of them are zero, then 39| | // everything is ASCII 40| 5.20k| size_t final_pos = pos + 16; 41| 88.5k| while (pos < final_pos) { ------------------ | Branch (41:18): [True: 83.2k, False: 5.20k] ------------------ 42| 83.2k| utf8_output[utf8_pos++] = char(data[pos]); 43| 83.2k| pos++; 44| 83.2k| } 45| 5.20k| continue; 46| 5.20k| } 47| 53.9k| } // if (pos + 16 <= len) 48| 71.4k| } // !consteval scope 49| | 50| 66.2k| unsigned char byte = data[pos]; 51| 66.2k| if ((byte & 0x80) == 0) { // if ASCII ------------------ | Branch (51:9): [True: 29.6k, False: 36.5k] ------------------ 52| | // will generate one UTF-8 bytes 53| 29.6k| utf8_output[utf8_pos++] = char(byte); 54| 29.6k| pos++; 55| 36.5k| } else { 56| | // will generate two UTF-8 bytes 57| 36.5k| utf8_output[utf8_pos++] = char((byte >> 6) | 0b11000000); 58| 36.5k| utf8_output[utf8_pos++] = char((byte & 0b111111) | 0b10000000); 59| 36.5k| pos++; 60| 36.5k| } 61| 66.2k| } // while 62| 7.60k| return utf8_pos; 63| 7.60k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf823utf8_length_from_latin1IPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_m: 152| 5.42k|utf8_length_from_latin1(InputPtr input, size_t length) noexcept { 153| 5.42k| size_t answer = length; 154| 5.42k| size_t i = 0; 155| | 156| |#if SIMDUTF_CPLUSPLUS23 157| | if !consteval 158| |#endif 159| 5.42k| { 160| 5.42k| auto pop = [](uint64_t v) { 161| 5.42k| return (size_t)(((v >> 7) & UINT64_C(0x0101010101010101)) * 162| 5.42k| UINT64_C(0x0101010101010101) >> 163| 5.42k| 56); 164| 5.42k| }; 165| 9.24k| for (; i + 32 <= length; i += 32) { ------------------ | Branch (165:12): [True: 3.81k, False: 5.42k] ------------------ 166| 3.81k| uint64_t v; 167| 3.81k| memcpy(&v, input + i, 8); 168| 3.81k| answer += pop(v); 169| 3.81k| memcpy(&v, input + i + 8, sizeof(v)); 170| 3.81k| answer += pop(v); 171| 3.81k| memcpy(&v, input + i + 16, sizeof(v)); 172| 3.81k| answer += pop(v); 173| 3.81k| memcpy(&v, input + i + 24, sizeof(v)); 174| 3.81k| answer += pop(v); 175| 3.81k| } 176| 5.72k| for (; i + 8 <= length; i += 8) { ------------------ | Branch (176:12): [True: 296, False: 5.42k] ------------------ 177| 296| uint64_t v; 178| 296| memcpy(&v, input + i, sizeof(v)); 179| 296| answer += pop(v); 180| 296| } 181| 5.42k| } // !consteval scope 182| 6.24k| for (; i + 1 <= length; i += 1) { ------------------ | Branch (182:10): [True: 817, False: 5.42k] ------------------ 183| 817| answer += static_cast(input[i]) >> 7; 184| 817| } 185| 5.42k| return answer; 186| 5.42k|} simdutf.cpp:_ZZN7simdutf6scalar12_GLOBAL__N_114latin1_to_utf823utf8_length_from_latin1IPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_mENKUlmE_clEm: 160| 15.5k| auto pop = [](uint64_t v) { 161| 15.5k| return (size_t)(((v >> 7) & UINT64_C(0x0101010101010101)) * 162| | UINT64_C(0x0101010101010101) >> 163| 15.5k| 56); 164| 15.5k| }; _ZN7simdutf6scalar14u16_swap_bytesEt: 8| 864k|u16_swap_bytes(const uint16_t word) { 9| 864k| return uint16_t((word >> 8) | (word << 8)); 10| 864k|} _ZN7simdutf6scalar14u32_swap_bytesEj: 13| 6.45k|u32_swap_bytes(const uint32_t word) { 14| 6.45k| return ((word >> 24) & 0xff) | // move byte 3 to byte 0 15| 6.45k| ((word << 8) & 0xff0000) | // move byte 1 to byte 2 16| 6.45k| ((word >> 8) & 0xff00) | // move byte 2 to byte 1 17| 6.45k| ((word << 24) & 0xff000000); // byte 0 to byte 3 18| 6.45k|} _ZN7simdutf6scalar5utf1614swap_if_neededILNS_10endiannessE1EEEtt: 27| 442k|template constexpr uint16_t swap_if_needed(uint16_t c) { 28| 442k| return !match_system(big_endian) ? scalar::u16_swap_bytes(c) : c; ------------------ | Branch (28:10): [True: 442k, Folded] ------------------ 29| 442k|} _ZN7simdutf6scalar5utf1614swap_if_neededILNS_10endiannessE0EEEtt: 27| 447k|template constexpr uint16_t swap_if_needed(uint16_t c) { 28| 447k| return !match_system(big_endian) ? scalar::u16_swap_bytes(c) : c; ------------------ | Branch (28:10): [Folded, False: 447k] ------------------ 29| 447k|} _ZN7simdutf6scalar5utf168validateILNS_10endiannessE0EEEbPKDsm: 22| 19.5k|validate(const char16_t *data, size_t len) noexcept { 23| 19.5k| uint64_t pos = 0; 24| 289k| while (pos < len) { ------------------ | Branch (24:10): [True: 270k, False: 19.3k] ------------------ 25| 270k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 26| 270k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (26:9): [True: 2.83k, False: 267k] ------------------ 27| 2.83k| if (pos + 1 >= len) { ------------------ | Branch (27:11): [True: 22, False: 2.80k] ------------------ 28| 22| return false; 29| 22| } 30| 2.80k| char16_t diff = char16_t(word - 0xD800); 31| 2.80k| if (diff > 0x3FF) { ------------------ | Branch (31:11): [True: 61, False: 2.74k] ------------------ 32| 61| return false; 33| 61| } 34| 2.74k| char16_t next_word = !match_system(big_endian) ------------------ | Branch (34:28): [Folded, False: 2.74k] ------------------ 35| 2.74k| ? u16_swap_bytes(data[pos + 1]) 36| 2.74k| : data[pos + 1]; 37| 2.74k| char16_t diff2 = char16_t(next_word - 0xDC00); 38| 2.74k| if (diff2 > 0x3FF) { ------------------ | Branch (38:11): [True: 65, False: 2.68k] ------------------ 39| 65| return false; 40| 65| } 41| 2.68k| pos += 2; 42| 267k| } else { 43| 267k| pos++; 44| 267k| } 45| 270k| } 46| 19.3k| return true; 47| 19.5k|} _ZN7simdutf6scalar5utf168validateILNS_10endiannessE1EEEbPKDsm: 22| 19.5k|validate(const char16_t *data, size_t len) noexcept { 23| 19.5k| uint64_t pos = 0; 24| 288k| while (pos < len) { ------------------ | Branch (24:10): [True: 269k, False: 19.4k] ------------------ 25| 269k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 26| 269k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (26:9): [True: 2.60k, False: 266k] ------------------ 27| 2.60k| if (pos + 1 >= len) { ------------------ | Branch (27:11): [True: 20, False: 2.58k] ------------------ 28| 20| return false; 29| 20| } 30| 2.58k| char16_t diff = char16_t(word - 0xD800); 31| 2.58k| if (diff > 0x3FF) { ------------------ | Branch (31:11): [True: 35, False: 2.55k] ------------------ 32| 35| return false; 33| 35| } 34| 2.55k| char16_t next_word = !match_system(big_endian) ------------------ | Branch (34:28): [True: 2.55k, Folded] ------------------ 35| 2.55k| ? u16_swap_bytes(data[pos + 1]) 36| 2.55k| : data[pos + 1]; 37| 2.55k| char16_t diff2 = char16_t(next_word - 0xDC00); 38| 2.55k| if (diff2 > 0x3FF) { ------------------ | Branch (38:11): [True: 76, False: 2.47k] ------------------ 39| 76| return false; 40| 76| } 41| 2.47k| pos += 2; 42| 266k| } else { 43| 266k| pos++; 44| 266k| } 45| 269k| } 46| 19.4k| return true; 47| 19.5k|} _ZN7simdutf6scalar5utf1620validate_with_errorsILNS_10endiannessE0EEENS_6resultEPKDsm: 51| 5.94k|validate_with_errors(const char16_t *data, size_t len) noexcept { 52| 5.94k| size_t pos = 0; 53| 51.4k| while (pos < len) { ------------------ | Branch (53:10): [True: 45.7k, False: 5.68k] ------------------ 54| 45.7k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 55| 45.7k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (55:9): [True: 1.24k, False: 44.5k] ------------------ 56| 1.24k| if (pos + 1 >= len) { ------------------ | Branch (56:11): [True: 22, False: 1.22k] ------------------ 57| 22| return result(error_code::SURROGATE, pos); 58| 22| } 59| 1.22k| char16_t diff = char16_t(word - 0xD800); 60| 1.22k| if (diff > 0x3FF) { ------------------ | Branch (60:11): [True: 113, False: 1.10k] ------------------ 61| 113| return result(error_code::SURROGATE, pos); 62| 113| } 63| 1.10k| char16_t next_word = !match_system(big_endian) ------------------ | Branch (63:28): [Folded, False: 1.10k] ------------------ 64| 1.10k| ? u16_swap_bytes(data[pos + 1]) 65| 1.10k| : data[pos + 1]; 66| 1.10k| char16_t diff2 = uint16_t(next_word - 0xDC00); 67| 1.10k| if (diff2 > 0x3FF) { ------------------ | Branch (67:11): [True: 123, False: 985] ------------------ 68| 123| return result(error_code::SURROGATE, pos); 69| 123| } 70| 985| pos += 2; 71| 44.5k| } else { 72| 44.5k| pos++; 73| 44.5k| } 74| 45.7k| } 75| 5.68k| return result(error_code::SUCCESS, pos); 76| 5.94k|} _ZN7simdutf6scalar5utf1620validate_with_errorsILNS_10endiannessE1EEENS_6resultEPKDsm: 51| 5.94k|validate_with_errors(const char16_t *data, size_t len) noexcept { 52| 5.94k| size_t pos = 0; 53| 50.5k| while (pos < len) { ------------------ | Branch (53:10): [True: 44.8k, False: 5.70k] ------------------ 54| 44.8k| char16_t word = scalar::utf16::swap_if_needed(data[pos]); 55| 44.8k| if ((word & 0xF800) == 0xD800) { ------------------ | Branch (55:9): [True: 1.01k, False: 43.8k] ------------------ 56| 1.01k| if (pos + 1 >= len) { ------------------ | Branch (56:11): [True: 20, False: 990] ------------------ 57| 20| return result(error_code::SURROGATE, pos); 58| 20| } 59| 990| char16_t diff = char16_t(word - 0xD800); 60| 990| if (diff > 0x3FF) { ------------------ | Branch (60:11): [True: 75, False: 915] ------------------ 61| 75| return result(error_code::SURROGATE, pos); 62| 75| } 63| 915| char16_t next_word = !match_system(big_endian) ------------------ | Branch (63:28): [True: 915, Folded] ------------------ 64| 915| ? u16_swap_bytes(data[pos + 1]) 65| 915| : data[pos + 1]; 66| 915| char16_t diff2 = uint16_t(next_word - 0xDC00); 67| 915| if (diff2 > 0x3FF) { ------------------ | Branch (67:11): [True: 141, False: 774] ------------------ 68| 141| return result(error_code::SURROGATE, pos); 69| 141| } 70| 774| pos += 2; 71| 43.8k| } else { 72| 43.8k| pos++; 73| 43.8k| } 74| 44.8k| } 75| 5.70k| return result(error_code::SUCCESS, pos); 76| 5.94k|} _ZN7simdutf6scalar5utf1622utf8_length_from_utf16ILNS_10endiannessE0EEEmPKDsm: 91| 30.4k| size_t len) { 92| | // We are not BOM aware. 93| 30.4k| size_t counter{0}; 94| 127k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (94:22): [True: 96.8k, False: 30.4k] ------------------ 95| 96.8k| char16_t word = scalar::utf16::swap_if_needed(p[i]); 96| 96.8k| counter++; // ASCII 97| 96.8k| counter += static_cast( 98| 96.8k| word > 99| 96.8k| 0x7F); // non-ASCII is at least 2 bytes, surrogates are 2*2 == 4 bytes 100| 96.8k| counter += static_cast((word > 0x7FF && word <= 0xD7FF) || ------------------ | Branch (100:37): [True: 40.3k, False: 56.4k] | Branch (100:53): [True: 32.6k, False: 7.71k] ------------------ 101| 64.1k| (word >= 0xE000)); // three-byte ------------------ | Branch (101:36): [True: 4.04k, False: 60.1k] ------------------ 102| 96.8k| } 103| 30.4k| return counter; 104| 30.4k|} _ZN7simdutf6scalar5utf1622utf8_length_from_utf16ILNS_10endiannessE1EEEmPKDsm: 91| 30.4k| size_t len) { 92| | // We are not BOM aware. 93| 30.4k| size_t counter{0}; 94| 125k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (94:22): [True: 95.0k, False: 30.4k] ------------------ 95| 95.0k| char16_t word = scalar::utf16::swap_if_needed(p[i]); 96| 95.0k| counter++; // ASCII 97| 95.0k| counter += static_cast( 98| 95.0k| word > 99| 95.0k| 0x7F); // non-ASCII is at least 2 bytes, surrogates are 2*2 == 4 bytes 100| 95.0k| counter += static_cast((word > 0x7FF && word <= 0xD7FF) || ------------------ | Branch (100:37): [True: 39.6k, False: 55.3k] | Branch (100:53): [True: 33.2k, False: 6.41k] ------------------ 101| 61.7k| (word >= 0xE000)); // three-byte ------------------ | Branch (101:36): [True: 3.57k, False: 58.2k] ------------------ 102| 95.0k| } 103| 30.4k| return counter; 104| 30.4k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115utf16_to_latin17convertILNS_10endiannessE0EPKDsPcQaasr7simdutf6detailE18indexes_into_utf16IT0_Esr7simdutf6detailE26index_assignable_from_charIT1_EEEmS8_mS9_: 17| 6.50k| OutputPtr latin_output) { 18| 6.50k| if (len == 0) { ------------------ | Branch (18:7): [True: 5.03k, False: 1.47k] ------------------ 19| 5.03k| return 0; 20| 5.03k| } 21| 1.47k| size_t pos = 0; 22| 1.47k| const auto latin_output_start = latin_output; 23| 1.47k| uint16_t word = 0; 24| 1.47k| uint16_t too_large = 0; 25| | 26| 99.6k| while (pos < len) { ------------------ | Branch (26:10): [True: 98.1k, False: 1.47k] ------------------ 27| 98.1k| word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (27:12): [Folded, False: 98.1k] ------------------ 28| 98.1k| too_large |= word; 29| 98.1k| *latin_output++ = char(word & 0xFF); 30| 98.1k| pos++; 31| 98.1k| } 32| 1.47k| if ((too_large & 0xFF00) != 0) { ------------------ | Branch (32:7): [True: 0, False: 1.47k] ------------------ 33| 0| return 0; 34| 0| } 35| | 36| 1.47k| return latin_output - latin_output_start; 37| 1.47k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115utf16_to_latin17convertILNS_10endiannessE1EPKDsPcQaasr7simdutf6detailE18indexes_into_utf16IT0_Esr7simdutf6detailE26index_assignable_from_charIT1_EEEmS8_mS9_: 17| 6.50k| OutputPtr latin_output) { 18| 6.50k| if (len == 0) { ------------------ | Branch (18:7): [True: 5.03k, False: 1.47k] ------------------ 19| 5.03k| return 0; 20| 5.03k| } 21| 1.47k| size_t pos = 0; 22| 1.47k| const auto latin_output_start = latin_output; 23| 1.47k| uint16_t word = 0; 24| 1.47k| uint16_t too_large = 0; 25| | 26| 99.6k| while (pos < len) { ------------------ | Branch (26:10): [True: 98.1k, False: 1.47k] ------------------ 27| 98.1k| word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (27:12): [True: 98.1k, Folded] ------------------ 28| 98.1k| too_large |= word; 29| 98.1k| *latin_output++ = char(word & 0xFF); 30| 98.1k| pos++; 31| 98.1k| } 32| 1.47k| if ((too_large & 0xFF00) != 0) { ------------------ | Branch (32:7): [True: 0, False: 1.47k] ------------------ 33| 0| return 0; 34| 0| } 35| | 36| 1.47k| return latin_output - latin_output_start; 37| 1.47k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf16_to_utf87convertILNS_10endiannessE0EPKDsPcQsr7simdutf6detailE18indexes_into_utf16IT0_EEEmS8_mT1_: 17| 12.3k| OutputPtr utf8_output) { 18| 12.3k| size_t pos = 0; 19| 12.3k| const auto start = utf8_output; 20| 70.6k| while (pos < len) { ------------------ | Branch (20:10): [True: 58.4k, False: 12.2k] ------------------ 21| |#if SIMDUTF_CPLUSPLUS23 22| | if !consteval 23| |#endif 24| 58.4k| { 25| | // try to convert the next block of 8 bytes 26| 58.4k| if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that ------------------ | Branch (26:11): [True: 54.4k, False: 3.93k] ------------------ 27| | // they are ascii 28| 54.4k| uint64_t v; 29| 54.4k| ::memcpy(&v, data + pos, sizeof(uint64_t)); 30| | if constexpr (!match_system(big_endian)) { 31| | v = (v >> 8) | (v << (64 - 8)); 32| | } 33| 54.4k| if ((v & 0xFF80FF80FF80FF80) == 0) { ------------------ | Branch (33:13): [True: 14.6k, False: 39.8k] ------------------ 34| 14.6k| size_t final_pos = pos + 4; 35| 73.0k| while (pos < final_pos) { ------------------ | Branch (35:18): [True: 58.4k, False: 14.6k] ------------------ 36| 58.4k| *utf8_output++ = !match_system(big_endian) ------------------ | Branch (36:30): [Folded, False: 58.4k] ------------------ 37| 58.4k| ? char(u16_swap_bytes(data[pos])) 38| 58.4k| : char(data[pos]); 39| 58.4k| pos++; 40| 58.4k| } 41| 14.6k| continue; 42| 14.6k| } 43| 54.4k| } 44| 58.4k| } 45| 43.8k| uint16_t word = 46| 43.8k| !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (46:9): [Folded, False: 43.8k] ------------------ 47| 43.8k| if ((word & 0xFF80) == 0) { ------------------ | Branch (47:9): [True: 4.67k, False: 39.1k] ------------------ 48| | // will generate one UTF-8 bytes 49| 4.67k| *utf8_output++ = char(word); 50| 4.67k| pos++; 51| 39.1k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (51:16): [True: 3.24k, False: 35.9k] ------------------ 52| | // will generate two UTF-8 bytes 53| | // we have 0b110XXXXX 0b10XXXXXX 54| 3.24k| *utf8_output++ = char((word >> 6) | 0b11000000); 55| 3.24k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 56| 3.24k| pos++; 57| 35.9k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (57:16): [True: 34.1k, False: 1.80k] ------------------ 58| | // will generate three UTF-8 bytes 59| | // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX 60| 34.1k| *utf8_output++ = char((word >> 12) | 0b11100000); 61| 34.1k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 62| 34.1k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 63| 34.1k| pos++; 64| 34.1k| } else { 65| | // must be a surrogate pair 66| 1.80k| if (pos + 1 >= len) { ------------------ | Branch (66:11): [True: 22, False: 1.78k] ------------------ 67| 22| return 0; 68| 22| } 69| 1.78k| uint16_t diff = uint16_t(word - 0xD800); 70| 1.78k| if (diff > 0x3FF) { ------------------ | Branch (70:11): [True: 50, False: 1.73k] ------------------ 71| 50| return 0; 72| 50| } 73| 1.73k| uint16_t next_word = !match_system(big_endian) ------------------ | Branch (73:28): [Folded, False: 1.73k] ------------------ 74| 1.73k| ? u16_swap_bytes(data[pos + 1]) 75| 1.73k| : data[pos + 1]; 76| 1.73k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 77| 1.73k| if (diff2 > 0x3FF) { ------------------ | Branch (77:11): [True: 51, False: 1.68k] ------------------ 78| 51| return 0; 79| 51| } 80| 1.68k| uint32_t value = (diff << 10) + diff2 + 0x10000; 81| | // will generate four UTF-8 bytes 82| | // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX 83| 1.68k| *utf8_output++ = char((value >> 18) | 0b11110000); 84| 1.68k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 85| 1.68k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 86| 1.68k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 87| 1.68k| pos += 2; 88| 1.68k| } 89| 43.8k| } 90| 12.2k| return utf8_output - start; 91| 12.3k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf16_to_utf87convertILNS_10endiannessE1EPKDsPcQsr7simdutf6detailE18indexes_into_utf16IT0_EEEmS8_mT1_: 17| 12.3k| OutputPtr utf8_output) { 18| 12.3k| size_t pos = 0; 19| 12.3k| const auto start = utf8_output; 20| 69.5k| while (pos < len) { ------------------ | Branch (20:10): [True: 57.3k, False: 12.2k] ------------------ 21| |#if SIMDUTF_CPLUSPLUS23 22| | if !consteval 23| |#endif 24| 57.3k| { 25| | // try to convert the next block of 8 bytes 26| 57.3k| if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that ------------------ | Branch (26:11): [True: 53.4k, False: 3.91k] ------------------ 27| | // they are ascii 28| 53.4k| uint64_t v; 29| 53.4k| ::memcpy(&v, data + pos, sizeof(uint64_t)); 30| 53.4k| if constexpr (!match_system(big_endian)) { 31| 53.4k| v = (v >> 8) | (v << (64 - 8)); 32| 53.4k| } 33| 53.4k| if ((v & 0xFF80FF80FF80FF80) == 0) { ------------------ | Branch (33:13): [True: 14.5k, False: 38.8k] ------------------ 34| 14.5k| size_t final_pos = pos + 4; 35| 72.6k| while (pos < final_pos) { ------------------ | Branch (35:18): [True: 58.1k, False: 14.5k] ------------------ 36| 58.1k| *utf8_output++ = !match_system(big_endian) ------------------ | Branch (36:30): [True: 58.1k, Folded] ------------------ 37| 58.1k| ? char(u16_swap_bytes(data[pos])) 38| 58.1k| : char(data[pos]); 39| 58.1k| pos++; 40| 58.1k| } 41| 14.5k| continue; 42| 14.5k| } 43| 53.4k| } 44| 57.3k| } 45| 42.7k| uint16_t word = 46| 42.7k| !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos]; ------------------ | Branch (46:9): [True: 42.7k, Folded] ------------------ 47| 42.7k| if ((word & 0xFF80) == 0) { ------------------ | Branch (47:9): [True: 4.64k, False: 38.1k] ------------------ 48| | // will generate one UTF-8 bytes 49| 4.64k| *utf8_output++ = char(word); 50| 4.64k| pos++; 51| 38.1k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (51:16): [True: 2.08k, False: 36.0k] ------------------ 52| | // will generate two UTF-8 bytes 53| | // we have 0b110XXXXX 0b10XXXXXX 54| 2.08k| *utf8_output++ = char((word >> 6) | 0b11000000); 55| 2.08k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 56| 2.08k| pos++; 57| 36.0k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (57:16): [True: 34.4k, False: 1.59k] ------------------ 58| | // will generate three UTF-8 bytes 59| | // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX 60| 34.4k| *utf8_output++ = char((word >> 12) | 0b11100000); 61| 34.4k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 62| 34.4k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 63| 34.4k| pos++; 64| 34.4k| } else { 65| | // must be a surrogate pair 66| 1.59k| if (pos + 1 >= len) { ------------------ | Branch (66:11): [True: 20, False: 1.57k] ------------------ 67| 20| return 0; 68| 20| } 69| 1.57k| uint16_t diff = uint16_t(word - 0xD800); 70| 1.57k| if (diff > 0x3FF) { ------------------ | Branch (70:11): [True: 27, False: 1.54k] ------------------ 71| 27| return 0; 72| 27| } 73| 1.54k| uint16_t next_word = !match_system(big_endian) ------------------ | Branch (73:28): [True: 1.54k, Folded] ------------------ 74| 1.54k| ? u16_swap_bytes(data[pos + 1]) 75| 1.54k| : data[pos + 1]; 76| 1.54k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 77| 1.54k| if (diff2 > 0x3FF) { ------------------ | Branch (77:11): [True: 63, False: 1.48k] ------------------ 78| 63| return 0; 79| 63| } 80| 1.48k| uint32_t value = (diff << 10) + diff2 + 0x10000; 81| | // will generate four UTF-8 bytes 82| | // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX 83| 1.48k| *utf8_output++ = char((value >> 18) | 0b11110000); 84| 1.48k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 85| 1.48k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 86| 1.48k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 87| 1.48k| pos += 2; 88| 1.48k| } 89| 42.7k| } 90| 12.2k| return utf8_output - start; 91| 12.3k|} _ZN7simdutf6scalar5utf328validateEPKDim: 25| 13.6k| size_t len) noexcept { 26| 13.6k| return validate(reinterpret_cast(buf), len); 27| 13.6k|} _ZN7simdutf6scalar5utf3222utf8_length_from_utf32EPKDim: 54| 14.5k| size_t len) { 55| | // We are not BOM aware. 56| 14.5k| size_t counter{0}; 57| 62.6k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (57:22): [True: 48.0k, False: 14.5k] ------------------ 58| | // credit: @ttsugriy for the vectorizable approach 59| 48.0k| counter++; // ASCII 60| 48.0k| counter += static_cast(p[i] > 0x7F); // two-byte 61| 48.0k| counter += static_cast(p[i] > 0x7FF); // three-byte 62| 48.0k| counter += static_cast(p[i] > 0xFFFF); // four-bytes 63| 48.0k| } 64| 14.5k| return counter; 65| 14.5k|} _ZN7simdutf6scalar5utf328validateIPKjQsr7simdutf6detailE19indexes_into_uint32IT_EEEbS5_m: 13| 13.6k| size_t len) noexcept { 14| 13.6k| uint64_t pos = 0; 15| 160k| for (; pos < len; pos++) { ------------------ | Branch (15:10): [True: 146k, False: 13.6k] ------------------ 16| 146k| uint32_t word = data[pos]; 17| 146k| if (word > 0x10FFFF || (word >= 0xD800 && word <= 0xDFFF)) { ------------------ | Branch (17:9): [True: 0, False: 146k] | Branch (17:29): [True: 1.32k, False: 145k] | Branch (17:47): [True: 0, False: 1.32k] ------------------ 18| 0| return false; 19| 0| } 20| 146k| } 21| 13.6k| return true; 22| 13.6k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_115utf32_to_latin17convertEPKDimPc: 12| 6.53k| char *latin1_output) { 13| 6.53k| char *start = latin1_output; 14| 6.53k| uint32_t utf32_char; 15| 6.53k| size_t pos = 0; 16| 6.53k| uint32_t too_large = 0; 17| | 18| 105k| while (pos < len) { ------------------ | Branch (18:10): [True: 99.1k, False: 6.53k] ------------------ 19| 99.1k| utf32_char = (uint32_t)data[pos]; 20| 99.1k| too_large |= utf32_char; 21| 99.1k| *latin1_output++ = (char)(utf32_char & 0xFF); 22| 99.1k| pos++; 23| 99.1k| } 24| 6.53k| if ((too_large & 0xFFFFFF00) != 0) { ------------------ | Branch (24:7): [True: 0, False: 6.53k] ------------------ 25| 0| return 0; 26| 0| } 27| 6.53k| return latin1_output - start; 28| 6.53k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf32_to_utf87convertIPKDiPcQaasr7simdutf6detailE18indexes_into_utf32IT_Esr7simdutf6detailE26index_assignable_from_charIT0_EEEmS7_mS8_: 17| 6.55k| OutputPtr utf8_output) { 18| 6.55k| size_t pos = 0; 19| 6.55k| auto start = utf8_output; 20| 40.6k| while (pos < len) { ------------------ | Branch (20:10): [True: 34.0k, False: 6.55k] ------------------ 21| |#if SIMDUTF_CPLUSPLUS23 22| | if !consteval 23| |#endif 24| 34.0k| { // try to convert the next block of 2 ASCII characters 25| 34.0k| if (pos + 2 <= len) { // if it is safe to read 8 more bytes, check that ------------------ | Branch (25:11): [True: 32.8k, False: 1.22k] ------------------ 26| | // they are ascii 27| 32.8k| uint64_t v; 28| 32.8k| ::memcpy(&v, data + pos, sizeof(uint64_t)); 29| 32.8k| if ((v & 0xFFFFFF80FFFFFF80) == 0) { ------------------ | Branch (29:13): [True: 28.9k, False: 3.91k] ------------------ 30| 28.9k| *utf8_output++ = char(data[pos]); 31| 28.9k| *utf8_output++ = char(data[pos + 1]); 32| 28.9k| pos += 2; 33| 28.9k| continue; 34| 28.9k| } 35| 32.8k| } 36| 34.0k| } 37| | 38| 5.13k| uint32_t word = data[pos]; 39| 5.13k| if ((word & 0xFFFFFF80) == 0) { ------------------ | Branch (39:9): [True: 1.55k, False: 3.58k] ------------------ 40| | // will generate one UTF-8 bytes 41| 1.55k| *utf8_output++ = char(word); 42| 1.55k| pos++; 43| 3.58k| } else if ((word & 0xFFFFF800) == 0) { ------------------ | Branch (43:16): [True: 1.17k, False: 2.40k] ------------------ 44| | // will generate two UTF-8 bytes 45| | // we have 0b110XXXXX 0b10XXXXXX 46| 1.17k| *utf8_output++ = char((word >> 6) | 0b11000000); 47| 1.17k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 48| 1.17k| pos++; 49| 2.40k| } else if ((word & 0xFFFF0000) == 0) { ------------------ | Branch (49:16): [True: 1.19k, False: 1.21k] ------------------ 50| | // will generate three UTF-8 bytes 51| | // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX 52| 1.19k| if (word >= 0xD800 && word <= 0xDFFF) { ------------------ | Branch (52:11): [True: 731, False: 468] | Branch (52:29): [True: 0, False: 731] ------------------ 53| 0| return 0; 54| 0| } 55| 1.19k| *utf8_output++ = char((word >> 12) | 0b11100000); 56| 1.19k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 57| 1.19k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 58| 1.19k| pos++; 59| 1.21k| } else { 60| | // will generate four UTF-8 bytes 61| | // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX 62| 1.21k| if (word > 0x10FFFF) { ------------------ | Branch (62:11): [True: 0, False: 1.21k] ------------------ 63| 0| return 0; 64| 0| } 65| 1.21k| *utf8_output++ = char((word >> 18) | 0b11110000); 66| 1.21k| *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000); 67| 1.21k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 68| 1.21k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 69| 1.21k| pos++; 70| 1.21k| } 71| 5.13k| } 72| 6.55k| return utf8_output - start; 73| 6.55k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf88validateEPKcm: 113| 20.9k| size_t len) noexcept { 114| 20.9k| return validate(reinterpret_cast(buf), len); 115| 20.9k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf88validateIPKhEEbT_m: 14| 20.9k| size_t len) noexcept { 15| 20.9k| static_assert( 16| 20.9k| std::is_same::type, uint8_t>::value, 17| 20.9k| "dereferencing the data pointer must result in a uint8_t"); 18| 20.9k| uint64_t pos = 0; 19| 20.9k| uint32_t code_point = 0; 20| 110k| while (pos < len) { ------------------ | Branch (20:10): [True: 89.9k, False: 20.4k] ------------------ 21| 89.9k| uint64_t next_pos; 22| |#if SIMDUTF_CPLUSPLUS23 23| | if !consteval 24| |#endif 25| 89.9k| { // check if the next 16 bytes are ascii. 26| 89.9k| next_pos = pos + 16; 27| 89.9k| if (next_pos <= len) { // if it is safe to read 16 more bytes, check ------------------ | Branch (27:11): [True: 87.6k, False: 2.22k] ------------------ 28| | // that they are ascii 29| 87.6k| uint64_t v1{}; 30| 87.6k| std::memcpy(&v1, data + pos, sizeof(uint64_t)); 31| 87.6k| uint64_t v2{}; 32| 87.6k| std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 33| 87.6k| uint64_t v{v1 | v2}; 34| 87.6k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (34:13): [True: 5.49k, False: 82.2k] ------------------ 35| 5.49k| pos = next_pos; 36| 5.49k| continue; 37| 5.49k| } 38| 87.6k| } 39| 89.9k| } 40| | 41| 84.4k| unsigned char byte = data[pos]; 42| | 43| 110k| while (byte < 0b10000000) { ------------------ | Branch (43:12): [True: 26.5k, False: 84.1k] ------------------ 44| 26.5k| if (++pos == len) { ------------------ | Branch (44:11): [True: 298, False: 26.2k] ------------------ 45| 298| return true; 46| 298| } 47| 26.2k| byte = data[pos]; 48| 26.2k| } 49| | 50| 84.1k| if ((byte & 0b11100000) == 0b11000000) { ------------------ | Branch (50:9): [True: 32.9k, False: 51.1k] ------------------ 51| 32.9k| next_pos = pos + 2; 52| 32.9k| if (next_pos > len) { ------------------ | Branch (52:11): [True: 10, False: 32.9k] ------------------ 53| 10| return false; 54| 10| } 55| 32.9k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (55:11): [True: 25, False: 32.9k] ------------------ 56| 25| return false; 57| 25| } 58| | // range check 59| 32.9k| code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 60| 32.9k| if (code_point < 0x80) { ------------------ | Branch (60:11): [True: 5, False: 32.9k] ------------------ 61| 5| return false; 62| 5| } 63| 51.1k| } else if ((byte & 0b11110000) == 0b11100000) { ------------------ | Branch (63:16): [True: 49.7k, False: 1.41k] ------------------ 64| 49.7k| next_pos = pos + 3; 65| 49.7k| if (next_pos > len) { ------------------ | Branch (65:11): [True: 5, False: 49.7k] ------------------ 66| 5| return false; 67| 5| } 68| 49.7k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (68:11): [True: 10, False: 49.7k] ------------------ 69| 10| return false; 70| 10| } 71| 49.7k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (71:11): [True: 8, False: 49.7k] ------------------ 72| 8| return false; 73| 8| } 74| | // range check 75| 49.7k| code_point = (byte & 0b00001111) << 12 | 76| 49.7k| (data[pos + 1] & 0b00111111) << 6 | 77| 49.7k| (data[pos + 2] & 0b00111111); 78| 49.7k| if ((code_point < 0x800) || ------------------ | Branch (78:11): [True: 9, False: 49.7k] ------------------ 79| 49.7k| (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (79:12): [True: 4.61k, False: 45.1k] | Branch (79:35): [True: 3, False: 4.61k] ------------------ 80| 12| return false; 81| 12| } 82| 49.7k| } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (82:16): [True: 1.32k, False: 87] ------------------ 83| 1.32k| next_pos = pos + 4; 84| 1.32k| if (next_pos > len) { ------------------ | Branch (84:11): [True: 4, False: 1.32k] ------------------ 85| 4| return false; 86| 4| } 87| 1.32k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (87:11): [True: 8, False: 1.31k] ------------------ 88| 8| return false; 89| 8| } 90| 1.31k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (90:11): [True: 7, False: 1.30k] ------------------ 91| 7| return false; 92| 7| } 93| 1.30k| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (93:11): [True: 8, False: 1.29k] ------------------ 94| 8| return false; 95| 8| } 96| | // range check 97| 1.29k| code_point = 98| 1.29k| (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 | 99| 1.29k| (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111); 100| 1.29k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (100:11): [True: 7, False: 1.29k] | Branch (100:35): [True: 3, False: 1.28k] ------------------ 101| 10| return false; 102| 10| } 103| 1.29k| } else { 104| | // we may have a continuation 105| 87| return false; 106| 87| } 107| 83.9k| pos = next_pos; 108| 83.9k| } 109| 20.4k| return true; 110| 20.9k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf820validate_with_errorsEPKcm: 221| 6.19k|validate_with_errors(const char *buf, size_t len) noexcept { 222| 6.19k| return validate_with_errors(reinterpret_cast(buf), len); 223| 6.19k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf820validate_with_errorsIPKhEENS_6resultET_m: 119| 6.19k|validate_with_errors(BytePtr data, size_t len) noexcept { 120| 6.19k| static_assert( 121| 6.19k| std::is_same::type, uint8_t>::value, 122| 6.19k| "dereferencing the data pointer must result in a uint8_t"); 123| 6.19k| size_t pos = 0; 124| 6.19k| uint32_t code_point = 0; 125| 12.6k| while (pos < len) { ------------------ | Branch (125:10): [True: 7.69k, False: 4.93k] ------------------ 126| | // check of the next 16 bytes are ascii. 127| 7.69k| size_t next_pos = pos + 16; 128| 7.69k| if (next_pos <= ------------------ | Branch (128:9): [True: 6.64k, False: 1.04k] ------------------ 129| 7.69k| len) { // if it is safe to read 16 more bytes, check that they are ascii 130| 6.64k| uint64_t v1; 131| 6.64k| std::memcpy(&v1, data + pos, sizeof(uint64_t)); 132| 6.64k| uint64_t v2; 133| 6.64k| std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 134| 6.64k| uint64_t v{v1 | v2}; 135| 6.64k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (135:11): [True: 2.77k, False: 3.87k] ------------------ 136| 2.77k| pos = next_pos; 137| 2.77k| continue; 138| 2.77k| } 139| 6.64k| } 140| 4.91k| unsigned char byte = data[pos]; 141| | 142| 13.4k| while (byte < 0b10000000) { ------------------ | Branch (142:12): [True: 8.70k, False: 4.78k] ------------------ 143| 8.70k| if (++pos == len) { ------------------ | Branch (143:11): [True: 133, False: 8.57k] ------------------ 144| 133| return result(error_code::SUCCESS, len); 145| 133| } 146| 8.57k| byte = data[pos]; 147| 8.57k| } 148| | 149| 4.78k| if ((byte & 0b11100000) == 0b11000000) { ------------------ | Branch (149:9): [True: 1.75k, False: 3.02k] ------------------ 150| 1.75k| next_pos = pos + 2; 151| 1.75k| if (next_pos > len) { ------------------ | Branch (151:11): [True: 40, False: 1.71k] ------------------ 152| 40| return result(error_code::TOO_SHORT, pos); 153| 40| } 154| 1.71k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (154:11): [True: 123, False: 1.59k] ------------------ 155| 123| return result(error_code::TOO_SHORT, pos); 156| 123| } 157| | // range check 158| 1.59k| code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 159| 1.59k| if (code_point < 0x80) { ------------------ | Branch (159:11): [True: 19, False: 1.57k] ------------------ 160| 19| return result(error_code::OVERLONG, pos); 161| 19| } 162| 3.02k| } else if ((byte & 0b11110000) == 0b11100000) { ------------------ | Branch (162:16): [True: 1.53k, False: 1.49k] ------------------ 163| 1.53k| next_pos = pos + 3; 164| 1.53k| if (next_pos > len) { ------------------ | Branch (164:11): [True: 49, False: 1.48k] ------------------ 165| 49| return result(error_code::TOO_SHORT, pos); 166| 49| } 167| 1.48k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (167:11): [True: 68, False: 1.41k] ------------------ 168| 68| return result(error_code::TOO_SHORT, pos); 169| 68| } 170| 1.41k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (170:11): [True: 42, False: 1.37k] ------------------ 171| 42| return result(error_code::TOO_SHORT, pos); 172| 42| } 173| | // range check 174| 1.37k| code_point = (byte & 0b00001111) << 12 | 175| 1.37k| (data[pos + 1] & 0b00111111) << 6 | 176| 1.37k| (data[pos + 2] & 0b00111111); 177| 1.37k| if (code_point < 0x800) { ------------------ | Branch (177:11): [True: 21, False: 1.35k] ------------------ 178| 21| return result(error_code::OVERLONG, pos); 179| 21| } 180| 1.35k| if (0xd7ff < code_point && code_point < 0xe000) { ------------------ | Branch (180:11): [True: 705, False: 648] | Branch (180:34): [True: 13, False: 692] ------------------ 181| 13| return result(error_code::SURROGATE, pos); 182| 13| } 183| 1.49k| } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (183:16): [True: 982, False: 511] ------------------ 184| 982| next_pos = pos + 4; 185| 982| if (next_pos > len) { ------------------ | Branch (185:11): [True: 30, False: 952] ------------------ 186| 30| return result(error_code::TOO_SHORT, pos); 187| 30| } 188| 952| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (188:11): [True: 64, False: 888] ------------------ 189| 64| return result(error_code::TOO_SHORT, pos); 190| 64| } 191| 888| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (191:11): [True: 50, False: 838] ------------------ 192| 50| return result(error_code::TOO_SHORT, pos); 193| 50| } 194| 838| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (194:11): [True: 51, False: 787] ------------------ 195| 51| return result(error_code::TOO_SHORT, pos); 196| 51| } 197| | // range check 198| 787| code_point = 199| 787| (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 | 200| 787| (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111); 201| 787| if (code_point <= 0xffff) { ------------------ | Branch (201:11): [True: 19, False: 768] ------------------ 202| 19| return result(error_code::OVERLONG, pos); 203| 19| } 204| 768| if (0x10ffff < code_point) { ------------------ | Branch (204:11): [True: 24, False: 744] ------------------ 205| 24| return result(error_code::TOO_LARGE, pos); 206| 24| } 207| 768| } else { 208| | // we either have too many continuation bytes or an invalid leading byte 209| 511| if ((byte & 0b11000000) == 0b10000000) { ------------------ | Branch (209:11): [True: 327, False: 184] ------------------ 210| 327| return result(error_code::TOO_LONG, pos); 211| 327| } else { 212| 184| return result(error_code::HEADER_BITS, pos); 213| 184| } 214| 511| } 215| 3.66k| pos = next_pos; 216| 3.66k| } 217| 4.93k| return result(error_code::SUCCESS, len); 218| 6.19k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf831rewind_and_validate_with_errorsEPKcS4_m: 232| 1.00k| const char *start, const char *buf, size_t len) noexcept { 233| | // First check that we start with a leading byte 234| 1.00k| if ((*start & 0b11000000) == 0b10000000) { ------------------ | Branch (234:7): [True: 81, False: 925] ------------------ 235| 81| return result(error_code::TOO_LONG, 0); 236| 81| } 237| 925| size_t extra_len{0}; 238| | // A leading byte cannot be further than 4 bytes away 239| 958| for (int i = 0; i < 5; i++) { ------------------ | Branch (239:19): [True: 958, False: 0] ------------------ 240| 958| unsigned char byte = *buf; 241| 958| if ((byte & 0b11000000) != 0b10000000) { ------------------ | Branch (241:9): [True: 925, False: 33] ------------------ 242| 925| break; 243| 925| } else { 244| 33| buf--; 245| 33| extra_len++; 246| 33| } 247| 958| } 248| | 249| 925| result res = validate_with_errors(buf, len + extra_len); 250| 925| res.count -= extra_len; // Might underflow 251| 925| return res; 252| 1.00k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf817count_code_pointsIPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_m: 258| 47.4k|simdutf_constexpr23 size_t count_code_points(InputPtr data, size_t len) { 259| 47.4k| size_t counter{0}; 260| 524k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (260:22): [True: 477k, False: 47.4k] ------------------ 261| | // -65 is 0b10111111, anything larger in two-complement's should start a new 262| | // code point. 263| 477k| if (int8_t(data[i]) > -65) { ------------------ | Branch (263:9): [True: 392k, False: 84.5k] ------------------ 264| 392k| counter++; 265| 392k| } 266| 477k| } 267| 47.4k| return counter; 268| 47.4k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_14utf822utf16_length_from_utf8IPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_m: 274| 62.7k|simdutf_constexpr23 size_t utf16_length_from_utf8(InputPtr data, size_t len) { 275| 62.7k| size_t counter{0}; 276| 462k| for (size_t i = 0; i < len; i++) { ------------------ | Branch (276:22): [True: 400k, False: 62.7k] ------------------ 277| 400k| if (int8_t(data[i]) > -65) { ------------------ | Branch (277:9): [True: 266k, False: 133k] ------------------ 278| 266k| counter++; 279| 266k| } 280| 400k| if (uint8_t(data[i]) >= 240) { ------------------ | Branch (280:9): [True: 11.7k, False: 388k] ------------------ 281| 11.7k| counter++; 282| 11.7k| } 283| 400k| } 284| 62.7k| return counter; 285| 62.7k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_114utf8_to_latin17convertIPKcPcQaasr7simdutf6detailE22indexes_into_byte_likeIT_Esr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mS8_: 17| 14.4k| OutputPtr latin_output) { 18| 14.4k| size_t pos = 0; 19| 14.4k| auto start = latin_output; 20| | 21| 105k| while (pos < len) { ------------------ | Branch (21:10): [True: 92.6k, False: 12.6k] ------------------ 22| |#if SIMDUTF_CPLUSPLUS23 23| | if !consteval 24| |#endif 25| 92.6k| { 26| | // try to convert the next block of 16 ASCII bytes 27| 92.6k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (27:11): [True: 76.3k, False: 16.2k] ------------------ 28| | // they are ascii 29| 76.3k| uint64_t v1; 30| 76.3k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 31| 76.3k| uint64_t v2; 32| 76.3k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 33| 76.3k| uint64_t v{v1 | v2}; // We are only interested in these bits: 1000 1000 34| | // 1000 1000 .... etc 35| 76.3k| if ((v & 0x8080808080808080) == ------------------ | Branch (35:13): [True: 8.62k, False: 67.7k] ------------------ 36| 76.3k| 0) { // if NONE of these are set, e.g. all of them are zero, then 37| | // everything is ASCII 38| 8.62k| size_t final_pos = pos + 16; 39| 146k| while (pos < final_pos) { ------------------ | Branch (39:18): [True: 138k, False: 8.62k] ------------------ 40| 138k| *latin_output++ = char(data[pos]); 41| 138k| pos++; 42| 138k| } 43| 8.62k| continue; 44| 8.62k| } 45| 76.3k| } 46| 92.6k| } 47| | 48| | // suppose it is not an all ASCII byte sequence 49| 83.9k| uint8_t leading_byte = data[pos]; // leading byte 50| 83.9k| if (leading_byte < 0b10000000) { ------------------ | Branch (50:9): [True: 39.1k, False: 44.8k] ------------------ 51| | // converting one ASCII byte !!! 52| 39.1k| *latin_output++ = char(leading_byte); 53| 39.1k| pos++; 54| 44.8k| } else if ((leading_byte & 0b11100000) == ------------------ | Branch (54:16): [True: 43.2k, False: 1.58k] ------------------ 55| 44.8k| 0b11000000) { // the first three bits indicate: 56| | // We have a two-byte UTF-8 57| 43.2k| if (pos + 1 >= len) { ------------------ | Branch (57:11): [True: 29, False: 43.1k] ------------------ 58| 29| return 0; 59| 29| } // minimal bound checking 60| 43.1k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (60:11): [True: 86, False: 43.1k] ------------------ 61| 86| return 0; 62| 86| } // checks if the next byte is a valid continuation byte in UTF-8. A 63| | // valid continuation byte starts with 10. 64| | // range check - 65| 43.1k| uint32_t code_point = 66| 43.1k| (leading_byte & 0b00011111) << 6 | 67| 43.1k| (data[pos + 1] & 68| 43.1k| 0b00111111); // assembles the Unicode code point from the two bytes. 69| | // It does this by discarding the leading 110 and 10 70| | // bits from the two bytes, shifting the remaining bits 71| | // of the first byte, and then combining the results 72| | // with a bitwise OR operation. 73| 43.1k| if (code_point < 0x80 || 0xFF < code_point) { ------------------ | Branch (73:11): [True: 17, False: 43.0k] | Branch (73:32): [True: 155, False: 42.9k] ------------------ 74| 172| return 0; // We only care about the range 129-255 which is Non-ASCII 75| | // latin1 characters. A code_point beneath 0x80 is invalid as 76| | // it is already covered by bytes whose leading bit is zero. 77| 172| } 78| 42.9k| *latin_output++ = char(code_point); 79| 42.9k| pos += 2; 80| 42.9k| } else { 81| 1.58k| return 0; 82| 1.58k| } 83| 83.9k| } 84| 12.6k| return latin_output - start; 85| 14.4k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf8_to_utf167convertILNS_10endiannessE0EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 16| 13.0k| char16_t *utf16_output) { 17| 13.0k| size_t pos = 0; 18| 13.0k| char16_t *start{utf16_output}; 19| 75.8k| while (pos < len) { ------------------ | Branch (19:10): [True: 63.5k, False: 12.2k] ------------------ 20| |#if SIMDUTF_CPLUSPLUS23 21| | if !consteval 22| |#endif 23| | // try to convert the next block of 16 ASCII bytes 24| 63.5k| { 25| 63.5k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (25:11): [True: 49.9k, False: 13.6k] ------------------ 26| | // they are ascii 27| 49.9k| uint64_t v1; 28| 49.9k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 29| 49.9k| uint64_t v2; 30| 49.9k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 31| 49.9k| uint64_t v{v1 | v2}; 32| 49.9k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (32:13): [True: 3.94k, False: 45.9k] ------------------ 33| 3.94k| size_t final_pos = pos + 16; 34| 67.0k| while (pos < final_pos) { ------------------ | Branch (34:18): [True: 63.1k, False: 3.94k] ------------------ 35| 63.1k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (35:31): [Folded, False: 63.1k] ------------------ 36| 63.1k| ? char16_t(u16_swap_bytes(data[pos])) 37| 63.1k| : char16_t(data[pos]); 38| 63.1k| pos++; 39| 63.1k| } 40| 3.94k| continue; 41| 3.94k| } 42| 49.9k| } 43| 63.5k| } 44| | 45| 59.6k| uint8_t leading_byte = data[pos]; // leading byte 46| 59.6k| if (leading_byte < 0b10000000) { ------------------ | Branch (46:9): [True: 21.0k, False: 38.5k] ------------------ 47| | // converting one ASCII byte !!! 48| 21.0k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (48:25): [Folded, False: 21.0k] ------------------ 49| 21.0k| ? char16_t(u16_swap_bytes(leading_byte)) 50| 21.0k| : char16_t(leading_byte); 51| 21.0k| pos++; 52| 38.5k| } else if ((leading_byte & 0b11100000) == 0b11000000) { ------------------ | Branch (52:16): [True: 4.42k, False: 34.1k] ------------------ 53| | // We have a two-byte UTF-8, it should become 54| | // a single UTF-16 word. 55| 4.42k| if (pos + 1 >= len) { ------------------ | Branch (55:11): [True: 40, False: 4.38k] ------------------ 56| 40| return 0; 57| 40| } // minimal bound checking 58| 4.38k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (58:11): [True: 81, False: 4.30k] ------------------ 59| 81| return 0; 60| 81| } 61| | // range check 62| 4.30k| uint32_t code_point = 63| 4.30k| (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 64| 4.30k| if (code_point < 0x80) { ------------------ | Branch (64:11): [True: 18, False: 4.28k] ------------------ 65| 18| return 0; 66| 18| } 67| | if constexpr (!match_system(big_endian)) { 68| | code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 69| | } 70| 4.28k| *utf16_output++ = char16_t(code_point); 71| 4.28k| pos += 2; 72| 34.1k| } else if ((leading_byte & 0b11110000) == 0b11100000) { ------------------ | Branch (72:16): [True: 31.1k, False: 2.98k] ------------------ 73| | // We have a three-byte UTF-8, it should become 74| | // a single UTF-16 word. 75| 31.1k| if (pos + 2 >= len) { ------------------ | Branch (75:11): [True: 49, False: 31.0k] ------------------ 76| 49| return 0; 77| 49| } // minimal bound checking 78| | 79| 31.0k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (79:11): [True: 34, False: 31.0k] ------------------ 80| 34| return 0; 81| 34| } 82| 31.0k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (82:11): [True: 28, False: 31.0k] ------------------ 83| 28| return 0; 84| 28| } 85| | // range check 86| 31.0k| uint32_t code_point = (leading_byte & 0b00001111) << 12 | 87| 31.0k| (data[pos + 1] & 0b00111111) << 6 | 88| 31.0k| (data[pos + 2] & 0b00111111); 89| 31.0k| if (code_point < 0x800 || (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (89:11): [True: 20, False: 30.9k] | Branch (89:34): [True: 4.33k, False: 26.6k] | Branch (89:57): [True: 13, False: 4.32k] ------------------ 90| 33| return 0; 91| 33| } 92| | if constexpr (!match_system(big_endian)) { 93| | code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 94| | } 95| 30.9k| *utf16_output++ = char16_t(code_point); 96| 30.9k| pos += 3; 97| 30.9k| } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (97:16): [True: 2.63k, False: 356] ------------------ 98| | // we have a 4-byte UTF-8 word. 99| 2.63k| if (pos + 3 >= len) { ------------------ | Branch (99:11): [True: 30, False: 2.60k] ------------------ 100| 30| return 0; 101| 30| } // minimal bound checking 102| 2.60k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (102:11): [True: 34, False: 2.56k] ------------------ 103| 34| return 0; 104| 34| } 105| 2.56k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (105:11): [True: 26, False: 2.54k] ------------------ 106| 26| return 0; 107| 26| } 108| 2.54k| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (108:11): [True: 27, False: 2.51k] ------------------ 109| 27| return 0; 110| 27| } 111| | 112| | // range check 113| 2.51k| uint32_t code_point = (leading_byte & 0b00000111) << 18 | 114| 2.51k| (data[pos + 1] & 0b00111111) << 12 | 115| 2.51k| (data[pos + 2] & 0b00111111) << 6 | 116| 2.51k| (data[pos + 3] & 0b00111111); 117| 2.51k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (117:11): [True: 17, False: 2.49k] | Branch (117:35): [True: 21, False: 2.47k] ------------------ 118| 38| return 0; 119| 38| } 120| 2.47k| code_point -= 0x10000; 121| 2.47k| uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10)); 122| 2.47k| uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF)); 123| | if constexpr (!match_system(big_endian)) { 124| | high_surrogate = u16_swap_bytes(high_surrogate); 125| | low_surrogate = u16_swap_bytes(low_surrogate); 126| | } 127| 2.47k| *utf16_output++ = char16_t(high_surrogate); 128| 2.47k| *utf16_output++ = char16_t(low_surrogate); 129| 2.47k| pos += 4; 130| 2.47k| } else { 131| 356| return 0; 132| 356| } 133| 59.6k| } 134| 12.2k| return utf16_output - start; 135| 13.0k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf8_to_utf167convertILNS_10endiannessE1EPKcQsr7simdutf6detailE22indexes_into_byte_likeIT0_EEEmS7_mPDs: 16| 13.0k| char16_t *utf16_output) { 17| 13.0k| size_t pos = 0; 18| 13.0k| char16_t *start{utf16_output}; 19| 74.3k| while (pos < len) { ------------------ | Branch (19:10): [True: 62.1k, False: 12.2k] ------------------ 20| |#if SIMDUTF_CPLUSPLUS23 21| | if !consteval 22| |#endif 23| | // try to convert the next block of 16 ASCII bytes 24| 62.1k| { 25| 62.1k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (25:11): [True: 48.5k, False: 13.5k] ------------------ 26| | // they are ascii 27| 48.5k| uint64_t v1; 28| 48.5k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 29| 48.5k| uint64_t v2; 30| 48.5k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 31| 48.5k| uint64_t v{v1 | v2}; 32| 48.5k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (32:13): [True: 3.94k, False: 44.5k] ------------------ 33| 3.94k| size_t final_pos = pos + 16; 34| 67.0k| while (pos < final_pos) { ------------------ | Branch (34:18): [True: 63.1k, False: 3.94k] ------------------ 35| 63.1k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (35:31): [True: 63.1k, Folded] ------------------ 36| 63.1k| ? char16_t(u16_swap_bytes(data[pos])) 37| 63.1k| : char16_t(data[pos]); 38| 63.1k| pos++; 39| 63.1k| } 40| 3.94k| continue; 41| 3.94k| } 42| 48.5k| } 43| 62.1k| } 44| | 45| 58.1k| uint8_t leading_byte = data[pos]; // leading byte 46| 58.1k| if (leading_byte < 0b10000000) { ------------------ | Branch (46:9): [True: 21.0k, False: 37.1k] ------------------ 47| | // converting one ASCII byte !!! 48| 21.0k| *utf16_output++ = !match_system(big_endian) ------------------ | Branch (48:25): [True: 21.0k, Folded] ------------------ 49| 21.0k| ? char16_t(u16_swap_bytes(leading_byte)) 50| 21.0k| : char16_t(leading_byte); 51| 21.0k| pos++; 52| 37.1k| } else if ((leading_byte & 0b11100000) == 0b11000000) { ------------------ | Branch (52:16): [True: 3.33k, False: 33.8k] ------------------ 53| | // We have a two-byte UTF-8, it should become 54| | // a single UTF-16 word. 55| 3.33k| if (pos + 1 >= len) { ------------------ | Branch (55:11): [True: 40, False: 3.29k] ------------------ 56| 40| return 0; 57| 40| } // minimal bound checking 58| 3.29k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (58:11): [True: 81, False: 3.21k] ------------------ 59| 81| return 0; 60| 81| } 61| | // range check 62| 3.21k| uint32_t code_point = 63| 3.21k| (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111); 64| 3.21k| if (code_point < 0x80) { ------------------ | Branch (64:11): [True: 18, False: 3.19k] ------------------ 65| 18| return 0; 66| 18| } 67| 3.19k| if constexpr (!match_system(big_endian)) { 68| 3.19k| code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 69| 3.19k| } 70| 3.19k| *utf16_output++ = char16_t(code_point); 71| 3.19k| pos += 2; 72| 33.8k| } else if ((leading_byte & 0b11110000) == 0b11100000) { ------------------ | Branch (72:16): [True: 31.1k, False: 2.68k] ------------------ 73| | // We have a three-byte UTF-8, it should become 74| | // a single UTF-16 word. 75| 31.1k| if (pos + 2 >= len) { ------------------ | Branch (75:11): [True: 49, False: 31.0k] ------------------ 76| 49| return 0; 77| 49| } // minimal bound checking 78| | 79| 31.0k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (79:11): [True: 34, False: 31.0k] ------------------ 80| 34| return 0; 81| 34| } 82| 31.0k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (82:11): [True: 28, False: 31.0k] ------------------ 83| 28| return 0; 84| 28| } 85| | // range check 86| 31.0k| uint32_t code_point = (leading_byte & 0b00001111) << 12 | 87| 31.0k| (data[pos + 1] & 0b00111111) << 6 | 88| 31.0k| (data[pos + 2] & 0b00111111); 89| 31.0k| if (code_point < 0x800 || (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (89:11): [True: 20, False: 31.0k] | Branch (89:34): [True: 4.00k, False: 26.9k] | Branch (89:57): [True: 13, False: 3.99k] ------------------ 90| 33| return 0; 91| 33| } 92| 30.9k| if constexpr (!match_system(big_endian)) { 93| 30.9k| code_point = uint32_t(u16_swap_bytes(uint16_t(code_point))); 94| 30.9k| } 95| 30.9k| *utf16_output++ = char16_t(code_point); 96| 30.9k| pos += 3; 97| 30.9k| } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (97:16): [True: 2.32k, False: 356] ------------------ 98| | // we have a 4-byte UTF-8 word. 99| 2.32k| if (pos + 3 >= len) { ------------------ | Branch (99:11): [True: 30, False: 2.29k] ------------------ 100| 30| return 0; 101| 30| } // minimal bound checking 102| 2.29k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (102:11): [True: 34, False: 2.26k] ------------------ 103| 34| return 0; 104| 34| } 105| 2.26k| if ((data[pos + 2] & 0b11000000) != 0b10000000) { ------------------ | Branch (105:11): [True: 26, False: 2.23k] ------------------ 106| 26| return 0; 107| 26| } 108| 2.23k| if ((data[pos + 3] & 0b11000000) != 0b10000000) { ------------------ | Branch (108:11): [True: 27, False: 2.20k] ------------------ 109| 27| return 0; 110| 27| } 111| | 112| | // range check 113| 2.20k| uint32_t code_point = (leading_byte & 0b00000111) << 18 | 114| 2.20k| (data[pos + 1] & 0b00111111) << 12 | 115| 2.20k| (data[pos + 2] & 0b00111111) << 6 | 116| 2.20k| (data[pos + 3] & 0b00111111); 117| 2.20k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (117:11): [True: 17, False: 2.19k] | Branch (117:35): [True: 21, False: 2.16k] ------------------ 118| 38| return 0; 119| 38| } 120| 2.16k| code_point -= 0x10000; 121| 2.16k| uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10)); 122| 2.16k| uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF)); 123| 2.16k| if constexpr (!match_system(big_endian)) { 124| 2.16k| high_surrogate = u16_swap_bytes(high_surrogate); 125| 2.16k| low_surrogate = u16_swap_bytes(low_surrogate); 126| 2.16k| } 127| 2.16k| *utf16_output++ = char16_t(high_surrogate); 128| 2.16k| *utf16_output++ = char16_t(low_surrogate); 129| 2.16k| pos += 4; 130| 2.16k| } else { 131| 356| return 0; 132| 356| } 133| 58.1k| } 134| 12.2k| return utf16_output - start; 135| 13.0k|} simdutf.cpp:_ZN7simdutf6scalar12_GLOBAL__N_113utf8_to_utf327convertIPKcQsr7simdutf6detailE22indexes_into_byte_likeIT_EEEmS6_mPDi: 16| 7.35k| char32_t *utf32_output) { 17| 7.35k| size_t pos = 0; 18| 7.35k| char32_t *start{utf32_output}; 19| 36.6k| while (pos < len) { ------------------ | Branch (19:10): [True: 30.0k, False: 6.55k] ------------------ 20| |#if SIMDUTF_CPLUSPLUS23 21| | if !consteval 22| |#endif 23| 30.0k| { 24| | // try to convert the next block of 16 ASCII bytes 25| 30.0k| if (pos + 16 <= len) { // if it is safe to read 16 more bytes, check that ------------------ | Branch (25:11): [True: 19.2k, False: 10.8k] ------------------ 26| | // they are ascii 27| 19.2k| uint64_t v1; 28| 19.2k| ::memcpy(&v1, data + pos, sizeof(uint64_t)); 29| 19.2k| uint64_t v2; 30| 19.2k| ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t)); 31| 19.2k| uint64_t v{v1 | v2}; 32| 19.2k| if ((v & 0x8080808080808080) == 0) { ------------------ | Branch (32:13): [True: 3.78k, False: 15.4k] ------------------ 33| 3.78k| size_t final_pos = pos + 16; 34| 64.2k| while (pos < final_pos) { ------------------ | Branch (34:18): [True: 60.4k, False: 3.78k] ------------------ 35| 60.4k| *utf32_output++ = uint8_t(data[pos]); 36| 60.4k| pos++; 37| 60.4k| } 38| 3.78k| continue; 39| 3.78k| } 40| 19.2k| } 41| 30.0k| } 42| 26.2k| auto leading_byte = uint8_t(data[pos]); // leading byte 43| 26.2k| if (leading_byte < 0b10000000) { ------------------ | Branch (43:9): [True: 19.2k, False: 7.04k] ------------------ 44| | // converting one ASCII byte !!! 45| 19.2k| *utf32_output++ = char32_t(leading_byte); 46| 19.2k| pos++; 47| 19.2k| } else if ((leading_byte & 0b11100000) == 0b11000000) { ------------------ | Branch (47:16): [True: 2.43k, False: 4.60k] ------------------ 48| | // We have a two-byte UTF-8 49| 2.43k| if (pos + 1 >= len) { ------------------ | Branch (49:11): [True: 40, False: 2.39k] ------------------ 50| 40| return 0; 51| 40| } // minimal bound checking 52| 2.39k| if ((data[pos + 1] & 0b11000000) != 0b10000000) { ------------------ | Branch (52:11): [True: 81, False: 2.31k] ------------------ 53| 81| return 0; 54| 81| } 55| | // range check 56| 2.31k| uint32_t code_point = (leading_byte & 0b00011111) << 6 | 57| 2.31k| (uint8_t(data[pos + 1]) & 0b00111111); 58| 2.31k| if (code_point < 0x80) { ------------------ | Branch (58:11): [True: 18, False: 2.29k] ------------------ 59| 18| return 0; 60| 18| } 61| 2.29k| *utf32_output++ = char32_t(code_point); 62| 2.29k| pos += 2; 63| 4.60k| } else if ((leading_byte & 0b11110000) == 0b11100000) { ------------------ | Branch (63:16): [True: 2.28k, False: 2.32k] ------------------ 64| | // We have a three-byte UTF-8 65| 2.28k| if (pos + 2 >= len) { ------------------ | Branch (65:11): [True: 49, False: 2.23k] ------------------ 66| 49| return 0; 67| 49| } // minimal bound checking 68| | 69| 2.23k| if ((uint8_t(data[pos + 1]) & 0b11000000) != 0b10000000) { ------------------ | Branch (69:11): [True: 34, False: 2.19k] ------------------ 70| 34| return 0; 71| 34| } 72| 2.19k| if ((uint8_t(data[pos + 2]) & 0b11000000) != 0b10000000) { ------------------ | Branch (72:11): [True: 28, False: 2.17k] ------------------ 73| 28| return 0; 74| 28| } 75| | // range check 76| 2.17k| uint32_t code_point = (leading_byte & 0b00001111) << 12 | 77| 2.17k| (uint8_t(data[pos + 1]) & 0b00111111) << 6 | 78| 2.17k| (uint8_t(data[pos + 2]) & 0b00111111); 79| 2.17k| if (code_point < 0x800 || (0xd7ff < code_point && code_point < 0xe000)) { ------------------ | Branch (79:11): [True: 20, False: 2.15k] | Branch (79:34): [True: 1.12k, False: 1.02k] | Branch (79:57): [True: 13, False: 1.11k] ------------------ 80| 33| return 0; 81| 33| } 82| 2.13k| *utf32_output++ = char32_t(code_point); 83| 2.13k| pos += 3; 84| 2.32k| } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000 ------------------ | Branch (84:16): [True: 1.96k, False: 356] ------------------ 85| | // we have a 4-byte UTF-8 word. 86| 1.96k| if (pos + 3 >= len) { ------------------ | Branch (86:11): [True: 30, False: 1.93k] ------------------ 87| 30| return 0; 88| 30| } // minimal bound checking 89| 1.93k| if ((uint8_t(data[pos + 1]) & 0b11000000) != 0b10000000) { ------------------ | Branch (89:11): [True: 34, False: 1.90k] ------------------ 90| 34| return 0; 91| 34| } 92| 1.90k| if ((uint8_t(data[pos + 2]) & 0b11000000) != 0b10000000) { ------------------ | Branch (92:11): [True: 26, False: 1.87k] ------------------ 93| 26| return 0; 94| 26| } 95| 1.87k| if ((uint8_t(data[pos + 3]) & 0b11000000) != 0b10000000) { ------------------ | Branch (95:11): [True: 27, False: 1.85k] ------------------ 96| 27| return 0; 97| 27| } 98| | 99| | // range check 100| 1.85k| uint32_t code_point = (leading_byte & 0b00000111) << 18 | 101| 1.85k| (uint8_t(data[pos + 1]) & 0b00111111) << 12 | 102| 1.85k| (uint8_t(data[pos + 2]) & 0b00111111) << 6 | 103| 1.85k| (uint8_t(data[pos + 3]) & 0b00111111); 104| 1.85k| if (code_point <= 0xffff || 0x10ffff < code_point) { ------------------ | Branch (104:11): [True: 17, False: 1.83k] | Branch (104:35): [True: 21, False: 1.81k] ------------------ 105| 38| return 0; 106| 38| } 107| 1.81k| *utf32_output++ = char32_t(code_point); 108| 1.81k| pos += 4; 109| 1.81k| } else { 110| 356| return 0; 111| 356| } 112| 26.2k| } 113| 6.55k| return utf32_output - start; 114| 7.35k|} _ZNK7simdutf8fallback14implementation13validate_utf8EPKcm: 37| 20.9k|implementation::validate_utf8(const char *buf, size_t len) const noexcept { 38| 20.9k| return scalar::utf8::validate(buf, len); 39| 20.9k|} _ZNK7simdutf8fallback14implementation25validate_utf8_with_errorsEPKcm: 44| 5.26k| const char *buf, size_t len) const noexcept { 45| 5.26k| return scalar::utf8::validate_with_errors(buf, len); 46| 5.26k|} _ZNK7simdutf8fallback14implementation16validate_utf16leEPKDsm: 77| 15.6k| size_t len) const noexcept { 78| 15.6k| return scalar::utf16::validate(buf, len); 79| 15.6k|} _ZNK7simdutf8fallback14implementation16validate_utf16beEPKDsm: 85| 15.6k| size_t len) const noexcept { 86| 15.6k| return scalar::utf16::validate(buf, len); 87| 15.6k|} _ZNK7simdutf8fallback14implementation28validate_utf16le_with_errorsEPKDsm: 90| 5.26k| const char16_t *buf, size_t len) const noexcept { 91| 5.26k| return scalar::utf16::validate_with_errors(buf, len); 92| 5.26k|} _ZNK7simdutf8fallback14implementation28validate_utf16be_with_errorsEPKDsm: 95| 5.26k| const char16_t *buf, size_t len) const noexcept { 96| 5.26k| return scalar::utf16::validate_with_errors(buf, len); 97| 5.26k|} _ZNK7simdutf8fallback14implementation14validate_utf32EPKDim: 114| 10.3k|implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept { 115| 10.3k| return scalar::utf32::validate(buf, len); 116| 10.3k|} _ZNK7simdutf8fallback14implementation22convert_latin1_to_utf8EPKcmPc: 128| 5.42k| const char *buf, size_t len, char *utf8_output) const noexcept { 129| 5.42k| return scalar::latin1_to_utf8::convert(buf, len, utf8_output); 130| 5.42k|} _ZNK7simdutf8fallback14implementation25convert_latin1_to_utf16leEPKcmPDs: 135| 5.26k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 136| 5.26k| return scalar::latin1_to_utf16::convert(buf, len, 137| 5.26k| utf16_output); 138| 5.26k|} _ZNK7simdutf8fallback14implementation25convert_latin1_to_utf16beEPKcmPDs: 141| 5.26k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 142| 5.26k| return scalar::latin1_to_utf16::convert(buf, len, 143| 5.26k| utf16_output); 144| 5.26k|} _ZNK7simdutf8fallback14implementation23convert_latin1_to_utf32EPKcmPDi: 149| 5.26k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 150| 5.26k| return scalar::latin1_to_utf32::convert(buf, len, utf32_output); 151| 5.26k|} _ZNK7simdutf8fallback14implementation22convert_utf8_to_latin1EPKcmPc: 156| 10.5k| const char *buf, size_t len, char *latin1_output) const noexcept { 157| 10.5k| return scalar::utf8_to_latin1::convert(buf, len, latin1_output); 158| 10.5k|} _ZNK7simdutf8fallback14implementation23convert_utf8_to_utf16leEPKcmPDs: 173| 10.4k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 174| 10.4k| return scalar::utf8_to_utf16::convert(buf, len, 175| 10.4k| utf16_output); 176| 10.4k|} _ZNK7simdutf8fallback14implementation23convert_utf8_to_utf16beEPKcmPDs: 179| 10.4k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 180| 10.4k| return scalar::utf8_to_utf16::convert(buf, len, 181| 10.4k| utf16_output); 182| 10.4k|} _ZNK7simdutf8fallback14implementation21convert_utf8_to_utf32EPKcmPDi: 211| 5.26k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 212| 5.26k| return scalar::utf8_to_utf32::convert(buf, len, utf32_output); 213| 5.26k|} _ZNK7simdutf8fallback14implementation25convert_utf16le_to_latin1EPKDsmPc: 228| 5.26k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 229| 5.26k| return scalar::utf16_to_latin1::convert(buf, len, 230| 5.26k| latin1_output); 231| 5.26k|} _ZNK7simdutf8fallback14implementation25convert_utf16be_to_latin1EPKDsmPc: 234| 5.26k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 235| 5.26k| return scalar::utf16_to_latin1::convert(buf, len, 236| 5.26k| latin1_output); 237| 5.26k|} _ZNK7simdutf8fallback14implementation23convert_utf16le_to_utf8EPKDsmPc: 268| 10.3k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 269| 10.3k| return scalar::utf16_to_utf8::convert(buf, len, 270| 10.3k| utf8_output); 271| 10.3k|} _ZNK7simdutf8fallback14implementation23convert_utf16be_to_utf8EPKDsmPc: 274| 10.3k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 275| 10.3k| return scalar::utf16_to_utf8::convert(buf, len, utf8_output); 276| 10.3k|} _ZNK7simdutf8fallback14implementation23convert_utf32_to_latin1EPKDimPc: 305| 5.26k| const char32_t *buf, size_t len, char *latin1_output) const noexcept { 306| 5.26k| return scalar::utf32_to_latin1::convert(buf, len, latin1_output); 307| 5.26k|} _ZNK7simdutf8fallback14implementation21convert_utf32_to_utf8EPKDimPc: 322| 5.06k| const char32_t *buf, size_t len, char *utf8_output) const noexcept { 323| 5.06k| return scalar::utf32_to_utf8::convert(buf, len, utf8_output); 324| 5.06k|} _ZNK7simdutf8fallback14implementation23latin1_length_from_utf8EPKcm: 438| 10.5k| const char *buf, size_t len) const noexcept { 439| 10.5k| return scalar::utf8::count_code_points(buf, len); 440| 10.5k|} _ZNK7simdutf8fallback14implementation23utf8_length_from_latin1EPKcm: 445| 5.42k| const char *input, size_t length) const noexcept { 446| 5.42k| return scalar::latin1_to_utf8::utf8_length_from_latin1(input, length); 447| 5.42k|} _ZNK7simdutf8fallback14implementation24utf8_length_from_utf16leEPKDsm: 452| 10.3k| const char16_t *input, size_t length) const noexcept { 453| 10.3k| return scalar::utf16::utf8_length_from_utf16(input, 454| 10.3k| length); 455| 10.3k|} _ZNK7simdutf8fallback14implementation24utf8_length_from_utf16beEPKDsm: 458| 10.3k| const char16_t *input, size_t length) const noexcept { 459| 10.3k| return scalar::utf16::utf8_length_from_utf16(input, length); 460| 10.3k|} _ZNK7simdutf8fallback14implementation22utf16_length_from_utf8EPKcm: 478| 20.9k| const char *input, size_t length) const noexcept { 479| 20.9k| return scalar::utf8::utf16_length_from_utf8(input, length); 480| 20.9k|} _ZNK7simdutf8fallback14implementation22utf8_length_from_utf32EPKDim: 513| 5.06k| const char32_t *input, size_t length) const noexcept { 514| 5.06k| return scalar::utf32::utf8_length_from_utf32(input, length); 515| 5.06k|} _ZNK7simdutf8fallback14implementation22utf32_length_from_utf8EPKcm: 527| 5.26k| const char *input, size_t length) const noexcept { 528| 5.26k| return scalar::utf8::count_code_points(input, length); 529| 5.26k|} _ZNK7simdutf8fallback14implementation16base64_to_binaryEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 536| 15.6k| last_chunk_handling_options last_chunk_options) const noexcept { 537| 15.6k| return simdutf::scalar::base64::base64_to_binary_details_impl( 538| 15.6k| input, length, output, options, last_chunk_options); 539| 15.6k|} _ZNK7simdutf8fallback14implementation16binary_to_base64EPKcmPcNS_14base64_optionsE: 564| 15.4k| base64_options options) const noexcept { 565| 15.4k| return scalar::base64::tail_encode_base64(output, input, length, options); 566| 15.4k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_16base6422compress_decode_base64ILb0ELb0ELb0EcEENS_11full_resultEPcPKT2_mNS_14base64_optionsENS_27last_chunk_handling_optionsE: 45| 61.1k| last_chunk_handling_options last_chunk_options) { 46| 61.1k| const uint8_t *to_base64 = 47| 61.1k| default_or_url ? tables::base64::to_base64_default_or_url_value ------------------ | Branch (47:7): [Folded, False: 61.1k] ------------------ 48| 61.1k| : (base64_url ? tables::base64::to_base64_url_value ------------------ | Branch (48:25): [Folded, False: 61.1k] ------------------ 49| 61.1k| : tables::base64::to_base64_value); 50| 61.1k| auto ri = simdutf::scalar::base64::find_end(src, srclen, options); 51| 61.1k| size_t equallocation = ri.equallocation; 52| 61.1k| size_t equalsigns = ri.equalsigns; 53| 61.1k| srclen = ri.srclen; 54| 61.1k| size_t full_input_length = ri.full_input_length; 55| 61.1k| if (srclen == 0) { ------------------ | Branch (55:7): [True: 54.6k, False: 6.43k] ------------------ 56| 54.6k| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (56:9): [True: 54.6k, Folded] | Branch (56:28): [True: 19, False: 54.6k] ------------------ 57| 19| return {INVALID_BASE64_CHARACTER, equallocation, 0, true}; 58| 19| } 59| 54.6k| return {SUCCESS, full_input_length, 0}; 60| 54.6k| } 61| 6.43k| char *end_of_safe_64byte_zone = 62| 6.43k| dst == nullptr ------------------ | Branch (62:7): [True: 73, False: 6.36k] ------------------ 63| 6.43k| ? nullptr 64| 6.43k| : ((srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 ------------------ | Branch (64:14): [True: 2.31k, False: 4.04k] ------------------ 65| 6.36k| : dst); 66| | 67| 6.43k| const chartype *const srcinit = src; 68| 6.43k| const char *const dstinit = dst; 69| 6.43k| const chartype *const srcend = src + srclen; 70| | 71| 6.43k| constexpr size_t block_size = 6; 72| 6.43k| static_assert(block_size >= 2, "block_size must be at least two"); 73| 6.43k| char buffer[block_size * 64]; 74| 6.43k| char *bufferptr = buffer; 75| 6.43k| if (srclen >= 64) { ------------------ | Branch (75:7): [True: 3.19k, False: 3.23k] ------------------ 76| 3.19k| const chartype *const srcend64 = src + srclen - 64; 77| 18.9k| while (src <= srcend64) { ------------------ | Branch (77:12): [True: 16.5k, False: 2.38k] ------------------ 78| 16.5k| block64 b(src); 79| 16.5k| src += 64; 80| 16.5k| uint64_t error = 0; 81| 16.5k| const uint64_t badcharmask = 82| 16.5k| b.to_base64_mask(&error); 83| 16.5k| if (!ignore_garbage && error) { ------------------ | Branch (83:11): [True: 16.5k, Folded] | Branch (83:30): [True: 810, False: 15.7k] ------------------ 84| 810| src -= 64; 85| 810| const size_t error_offset = trailing_zeroes(error); 86| 810| return {error_code::INVALID_BASE64_CHARACTER, 87| 810| size_t(src - srcinit + error_offset), size_t(dst - dstinit)}; 88| 810| } 89| 15.7k| if (badcharmask != 0) { ------------------ | Branch (89:11): [True: 2.23k, False: 13.5k] ------------------ 90| 2.23k| bufferptr += b.compress_block(badcharmask, bufferptr); 91| 13.5k| } else if (bufferptr != buffer) { ------------------ | Branch (91:18): [True: 395, False: 13.1k] ------------------ 92| 395| b.copy_block(bufferptr); 93| 395| bufferptr += 64; 94| 13.1k| } else { 95| 13.1k| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (95:13): [True: 812, False: 12.3k] ------------------ 96| 812| b.base64_decode_block_safe(dst); 97| 12.3k| } else { 98| 12.3k| b.base64_decode_block(dst); 99| 12.3k| } 100| 13.1k| dst += 48; 101| 13.1k| } 102| 15.7k| if (bufferptr >= (block_size - 1) * 64 + buffer) { ------------------ | Branch (102:11): [True: 240, False: 15.5k] ------------------ 103| 1.20k| for (size_t i = 0; i < (block_size - 2); i++) { ------------------ | Branch (103:28): [True: 960, False: 240] ------------------ 104| 960| base64_decode_block(dst, buffer + i * 64); 105| 960| dst += 48; 106| 960| } 107| 240| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (107:13): [True: 0, False: 240] ------------------ 108| 0| base64_decode_block_safe(dst, buffer + (block_size - 2) * 64); 109| 240| } else { 110| 240| base64_decode_block(dst, buffer + (block_size - 2) * 64); 111| 240| } 112| 240| dst += 48; 113| 240| std::memcpy(buffer, buffer + (block_size - 1) * 64, 114| 240| 64); // 64 might be too much 115| 240| bufferptr -= (block_size - 1) * 64; 116| 240| } 117| 15.7k| } 118| 3.19k| } 119| | 120| 5.62k| 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| 5.62k| int last_block = (int)((bufferptr - buffer_start) % 64); 124| 5.62k| if (last_block != 0 && srcend - src + last_block >= 64) { ------------------ | Branch (124:7): [True: 668, False: 4.95k] | Branch (124:26): [True: 134, False: 534] ------------------ 125| | 126| 1.82k| while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) { ------------------ | Branch (126:12): [True: 1.72k, False: 100] | Branch (126:52): [True: 1.70k, False: 20] ------------------ 127| 1.70k| uint8_t val = to_base64[uint8_t(*src)]; 128| 1.70k| *bufferptr = char(val); 129| 1.70k| if (!ignore_garbage && ------------------ | Branch (129:11): [True: 1.70k, Folded] ------------------ 130| 1.70k| (!scalar::base64::is_eight_byte(*src) || val > 64)) { ------------------ | Branch (130:12): [True: 0, False: 1.70k] | Branch (130:52): [True: 14, False: 1.68k] ------------------ 131| 14| return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit), 132| 14| size_t(dst - dstinit)}; 133| 14| } 134| 1.68k| bufferptr += (val <= 63); 135| 1.68k| src++; 136| 1.68k| } 137| 134| } 138| | 139| 6.14k| for (; buffer_start + 64 <= bufferptr; buffer_start += 64) { ------------------ | Branch (139:10): [True: 532, False: 5.60k] ------------------ 140| 532| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (140:9): [True: 50, False: 482] ------------------ 141| 50| base64_decode_block_safe(dst, buffer_start); 142| 482| } else { 143| 482| base64_decode_block(dst, buffer_start); 144| 482| } 145| 532| dst += 48; 146| 532| } 147| 5.60k| if ((bufferptr - buffer_start) % 64 != 0) { ------------------ | Branch (147:7): [True: 554, False: 5.05k] ------------------ 148| 5.29k| while (buffer_start + 4 < bufferptr) { ------------------ | Branch (148:12): [True: 4.73k, False: 554] ------------------ 149| 4.73k| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 150| 4.73k| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 151| 4.73k| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 152| 4.73k| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 153| 4.73k| << 8; 154| 4.73k|#if !SIMDUTF_IS_BIG_ENDIAN 155| 4.73k| triple = scalar::u32_swap_bytes(triple); 156| 4.73k|#endif 157| 4.73k| std::memcpy(dst, &triple, 3); 158| | 159| 4.73k| dst += 3; 160| 4.73k| buffer_start += 4; 161| 4.73k| } 162| 554| if (buffer_start + 4 <= bufferptr) { ------------------ | Branch (162:9): [True: 150, False: 404] ------------------ 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| 554| int leftover = int(bufferptr - buffer_start); 179| 1.43k| while (leftover > 0) { ------------------ | Branch (179:12): [True: 880, False: 554] ------------------ 180| 880| if (!ignore_garbage) { ------------------ | Branch (180:11): [True: 880, Folded] ------------------ 181| 4.51k| while (to_base64[uint8_t(*(src - 1))] == 64) { ------------------ | Branch (181:16): [True: 3.63k, False: 880] ------------------ 182| 3.63k| src--; 183| 3.63k| } 184| 880| } 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| 880| src--; 190| 880| leftover--; 191| 880| } 192| 554| } 193| 5.60k| if (src < srcend + equalsigns) { ------------------ | Branch (193:7): [True: 5.20k, False: 409] ------------------ 194| 5.20k| full_result r = scalar::base64::base64_tail_decode( 195| 5.20k| dst, src, srcend - src, equalsigns, options, last_chunk_options); 196| 5.20k| r = scalar::base64::patch_tail_result( 197| 5.20k| r, size_t(src - srcinit), size_t(dst - dstinit), equallocation, 198| 5.20k| 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| 5.20k| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (203:9): [True: 1.91k, False: 3.28k] | Branch (203:43): [True: 1.62k, False: 296] ------------------ 204| 1.62k| r.input_count < full_input_length) { ------------------ | Branch (204:9): [True: 317, False: 1.30k] ------------------ 205| | // First check if we can extend the input to the end of the stream 206| 2.06k| while (r.input_count < full_input_length && ------------------ | Branch (206:14): [True: 2.02k, False: 39] ------------------ 207| 2.02k| base64_ignorable(*(srcinit + r.input_count), options)) { ------------------ | Branch (207:14): [True: 1.74k, False: 278] ------------------ 208| 1.74k| r.input_count++; 209| 1.74k| } 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: 278, False: 39] ------------------ 213| 2.08k| while (r.input_count > 0 && ------------------ | Branch (213:16): [True: 1.99k, False: 87] ------------------ 214| 1.99k| base64_ignorable(*(srcinit + r.input_count - 1), options)) { ------------------ | Branch (214:16): [True: 1.80k, False: 191] ------------------ 215| 1.80k| r.input_count--; 216| 1.80k| } 217| 278| } 218| 317| } 219| 5.20k| return r; 220| 5.20k| } 221| 409| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (221:7): [True: 409, Folded] | Branch (221:26): [True: 0, False: 409] ------------------ 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| 409| return {SUCCESS, srclen, size_t(dst - dstinit)}; 229| 409|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_16base6422compress_decode_base64ILb0ELb0ELb0EcEENS_11full_resultEPcPKT2_mNS_14base64_optionsENS_27last_chunk_handling_optionsE: 45| 15.3k| last_chunk_handling_options last_chunk_options) { 46| 15.3k| const uint8_t *to_base64 = 47| 15.3k| default_or_url ? tables::base64::to_base64_default_or_url_value ------------------ | Branch (47:7): [Folded, False: 15.3k] ------------------ 48| 15.3k| : (base64_url ? tables::base64::to_base64_url_value ------------------ | Branch (48:25): [Folded, False: 15.3k] ------------------ 49| 15.3k| : tables::base64::to_base64_value); 50| 15.3k| auto ri = simdutf::scalar::base64::find_end(src, srclen, options); 51| 15.3k| size_t equallocation = ri.equallocation; 52| 15.3k| size_t equalsigns = ri.equalsigns; 53| 15.3k| srclen = ri.srclen; 54| 15.3k| size_t full_input_length = ri.full_input_length; 55| 15.3k| if (srclen == 0) { ------------------ | Branch (55:7): [True: 13.7k, False: 1.51k] ------------------ 56| 13.7k| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (56:9): [True: 13.7k, Folded] | Branch (56:28): [True: 9, False: 13.7k] ------------------ 57| 9| return {INVALID_BASE64_CHARACTER, equallocation, 0, true}; 58| 9| } 59| 13.7k| return {SUCCESS, full_input_length, 0}; 60| 13.7k| } 61| 1.51k| char *end_of_safe_64byte_zone = 62| 1.51k| dst == nullptr ------------------ | Branch (62:7): [True: 30, False: 1.48k] ------------------ 63| 1.51k| ? nullptr 64| 1.51k| : ((srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 ------------------ | Branch (64:14): [True: 488, False: 994] ------------------ 65| 1.48k| : dst); 66| | 67| 1.51k| const chartype *const srcinit = src; 68| 1.51k| const char *const dstinit = dst; 69| 1.51k| const chartype *const srcend = src + srclen; 70| | 71| 1.51k| constexpr size_t block_size = 6; 72| 1.51k| static_assert(block_size >= 2, "block_size must be at least two"); 73| 1.51k| char buffer[block_size * 64]; 74| 1.51k| char *bufferptr = buffer; 75| 1.51k| if (srclen >= 64) { ------------------ | Branch (75:7): [True: 786, False: 726] ------------------ 76| 786| const chartype *const srcend64 = src + srclen - 64; 77| 4.40k| while (src <= srcend64) { ------------------ | Branch (77:12): [True: 3.86k, False: 547] ------------------ 78| 3.86k| block64 b(src); 79| 3.86k| src += 64; 80| 3.86k| uint64_t error = 0; 81| 3.86k| const uint64_t badcharmask = 82| 3.86k| b.to_base64_mask(&error); 83| 3.86k| if (!ignore_garbage && error) { ------------------ | Branch (83:11): [True: 3.86k, Folded] | Branch (83:30): [True: 239, False: 3.62k] ------------------ 84| 239| src -= 64; 85| 239| const size_t error_offset = trailing_zeroes(error); 86| 239| return {error_code::INVALID_BASE64_CHARACTER, 87| 239| size_t(src - srcinit + error_offset), size_t(dst - dstinit)}; 88| 239| } 89| 3.62k| if (badcharmask != 0) { ------------------ | Branch (89:11): [True: 724, False: 2.89k] ------------------ 90| 724| bufferptr += b.compress_block(badcharmask, bufferptr); 91| 2.89k| } else if (bufferptr != buffer) { ------------------ | Branch (91:18): [True: 94, False: 2.80k] ------------------ 92| 94| b.copy_block(bufferptr); 93| 94| bufferptr += 64; 94| 2.80k| } else { 95| 2.80k| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (95:13): [True: 131, False: 2.67k] ------------------ 96| 131| b.base64_decode_block_safe(dst); 97| 2.67k| } else { 98| 2.67k| b.base64_decode_block(dst); 99| 2.67k| } 100| 2.80k| dst += 48; 101| 2.80k| } 102| 3.62k| if (bufferptr >= (block_size - 1) * 64 + buffer) { ------------------ | Branch (102:11): [True: 70, False: 3.55k] ------------------ 103| 350| for (size_t i = 0; i < (block_size - 2); i++) { ------------------ | Branch (103:28): [True: 280, False: 70] ------------------ 104| 280| base64_decode_block(dst, buffer + i * 64); 105| 280| dst += 48; 106| 280| } 107| 70| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (107:13): [True: 0, False: 70] ------------------ 108| 0| base64_decode_block_safe(dst, buffer + (block_size - 2) * 64); 109| 70| } else { 110| 70| base64_decode_block(dst, buffer + (block_size - 2) * 64); 111| 70| } 112| 70| dst += 48; 113| 70| std::memcpy(buffer, buffer + (block_size - 1) * 64, 114| 70| 64); // 64 might be too much 115| 70| bufferptr -= (block_size - 1) * 64; 116| 70| } 117| 3.62k| } 118| 786| } 119| | 120| 1.27k| 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| 1.27k| int last_block = (int)((bufferptr - buffer_start) % 64); 124| 1.27k| if (last_block != 0 && srcend - src + last_block >= 64) { ------------------ | Branch (124:7): [True: 231, False: 1.04k] | Branch (124:26): [True: 45, False: 186] ------------------ 125| | 126| 450| while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) { ------------------ | Branch (126:12): [True: 420, False: 30] | Branch (126:52): [True: 411, False: 9] ------------------ 127| 411| uint8_t val = to_base64[uint8_t(*src)]; 128| 411| *bufferptr = char(val); 129| 411| if (!ignore_garbage && ------------------ | Branch (129:11): [True: 411, Folded] ------------------ 130| 411| (!scalar::base64::is_eight_byte(*src) || val > 64)) { ------------------ | Branch (130:12): [True: 0, False: 411] | Branch (130:52): [True: 6, False: 405] ------------------ 131| 6| return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit), 132| 6| size_t(dst - dstinit)}; 133| 6| } 134| 405| bufferptr += (val <= 63); 135| 405| src++; 136| 405| } 137| 45| } 138| | 139| 1.42k| for (; buffer_start + 64 <= bufferptr; buffer_start += 64) { ------------------ | Branch (139:10): [True: 159, False: 1.26k] ------------------ 140| 159| if (dst >= end_of_safe_64byte_zone) { ------------------ | Branch (140:9): [True: 20, False: 139] ------------------ 141| 20| base64_decode_block_safe(dst, buffer_start); 142| 139| } else { 143| 139| base64_decode_block(dst, buffer_start); 144| 139| } 145| 159| dst += 48; 146| 159| } 147| 1.26k| if ((bufferptr - buffer_start) % 64 != 0) { ------------------ | Branch (147:7): [True: 195, False: 1.07k] ------------------ 148| 1.71k| while (buffer_start + 4 < bufferptr) { ------------------ | Branch (148:12): [True: 1.52k, False: 195] ------------------ 149| 1.52k| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 150| 1.52k| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 151| 1.52k| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 152| 1.52k| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 153| 1.52k| << 8; 154| 1.52k|#if !SIMDUTF_IS_BIG_ENDIAN 155| 1.52k| triple = scalar::u32_swap_bytes(triple); 156| 1.52k|#endif 157| 1.52k| std::memcpy(dst, &triple, 3); 158| | 159| 1.52k| dst += 3; 160| 1.52k| buffer_start += 4; 161| 1.52k| } 162| 195| if (buffer_start + 4 <= bufferptr) { ------------------ | Branch (162:9): [True: 49, False: 146] ------------------ 163| 49| uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) + 164| 49| (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) + 165| 49| (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) + 166| 49| (uint32_t(uint8_t(buffer_start[3])) << 0 * 6)) 167| 49| << 8; 168| 49|#if !SIMDUTF_IS_BIG_ENDIAN 169| 49| triple = scalar::u32_swap_bytes(triple); 170| 49|#endif 171| 49| std::memcpy(dst, &triple, 3); 172| | 173| 49| dst += 3; 174| 49| buffer_start += 4; 175| 49| } 176| | // we may have 1, 2 or 3 bytes left and we need to decode them so let us 177| | // backtrack 178| 195| int leftover = int(bufferptr - buffer_start); 179| 509| while (leftover > 0) { ------------------ | Branch (179:12): [True: 314, False: 195] ------------------ 180| 314| if (!ignore_garbage) { ------------------ | Branch (180:11): [True: 314, Folded] ------------------ 181| 2.16k| while (to_base64[uint8_t(*(src - 1))] == 64) { ------------------ | Branch (181:16): [True: 1.85k, False: 314] ------------------ 182| 1.85k| src--; 183| 1.85k| } 184| 314| } 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| 314| src--; 190| 314| leftover--; 191| 314| } 192| 195| } 193| 1.26k| if (src < srcend + equalsigns) { ------------------ | Branch (193:7): [True: 1.14k, False: 123] ------------------ 194| 1.14k| full_result r = scalar::base64::base64_tail_decode( 195| 1.14k| dst, src, srcend - src, equalsigns, options, last_chunk_options); 196| 1.14k| r = scalar::base64::patch_tail_result( 197| 1.14k| r, size_t(src - srcinit), size_t(dst - dstinit), equallocation, 198| 1.14k| 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| 1.14k| if (is_partial(last_chunk_options) && r.error == error_code::SUCCESS && ------------------ | Branch (203:9): [True: 0, False: 1.14k] | 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| 1.14k| return r; 220| 1.14k| } 221| 123| if (!ignore_garbage && equalsigns > 0) { ------------------ | Branch (221:7): [True: 123, Folded] | Branch (221:26): [True: 0, False: 123] ------------------ 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| 123| return {SUCCESS, srclen, size_t(dst - dstinit)}; 229| 123|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE13get_remainderEPh: 57| 25.9k|buf_block_reader::get_remainder(uint8_t *dst) const { 58| 25.9k| if (len == idx) { ------------------ | Branch (58:7): [True: 21.7k, False: 4.14k] ------------------ 59| 21.7k| return 0; 60| 21.7k| } // memcpy(dst, null, 0) will trigger an error with some sanitizers 61| 4.14k| std::memset(dst, 0x20, 62| 4.14k| STEP_SIZE); // std::memset STEP_SIZE because it is more efficient 63| | // to write out 8 or 16 bytes at once. 64| 4.14k| std::memcpy(dst, buf + idx, len - idx); 65| 4.14k| return len - idx; 66| 25.9k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE13get_remainderEPh: 57| 25.9k|buf_block_reader::get_remainder(uint8_t *dst) const { 58| 25.9k| if (len == idx) { ------------------ | Branch (58:7): [True: 22.7k, False: 3.19k] ------------------ 59| 22.7k| return 0; 60| 22.7k| } // memcpy(dst, null, 0) will trigger an error with some sanitizers 61| 3.19k| std::memset(dst, 0x20, 62| 3.19k| STEP_SIZE); // std::memset STEP_SIZE because it is more efficient 63| | // to write out 8 or 16 bytes at once. 64| 3.19k| std::memcpy(dst, buf + idx, len - idx); 65| 3.19k| return len - idx; 66| 25.9k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EEC2EPKhm: 36| 26.1k| : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, ------------------ | Branch (36:42): [True: 24.2k, False: 1.89k] ------------------ 37| 26.1k| idx{0} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE14has_full_blockEv: 45| 33.0k|simdutf_really_inline bool buf_block_reader::has_full_block() const { 46| 33.0k| return idx < lenminusstep; 47| 33.0k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE10full_blockEv: 51| 7.19k|buf_block_reader::full_block() const { 52| 7.19k| return &buf[idx]; 53| 7.19k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_116buf_block_readerILm64EE7advanceEv: 69| 32.8k|simdutf_really_inline void buf_block_reader::advance() { 70| 32.8k| idx += STEP_SIZE; 71| 32.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EEC2EPKhm: 36| 26.1k| : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, ------------------ | Branch (36:42): [True: 24.3k, False: 1.78k] ------------------ 37| 26.1k| idx{0} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE14has_full_blockEv: 45| 35.4k|simdutf_really_inline bool buf_block_reader::has_full_block() const { 46| 35.4k| return idx < lenminusstep; 47| 35.4k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE10full_blockEv: 51| 9.50k|buf_block_reader::full_block() const { 52| 9.50k| return &buf[idx]; 53| 9.50k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_116buf_block_readerILm64EE7advanceEv: 69| 35.1k|simdutf_really_inline void buf_block_reader::advance() { 70| 35.1k| idx += STEP_SIZE; 71| 35.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE0EEEmPKDsm: 10| 9.93k| size_t size) { 11| 9.93k| size_t pos = 0; 12| | 13| 9.93k| using vector_u16 = simd16; 14| 9.93k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 9.93k| const auto one = vector_u16::splat(1); 17| | 18| 9.93k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 9.93k| 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| 9.93k| constexpr size_t max_iterations = 65535 / 2; 26| 9.93k| size_t iteration = max_iterations; 27| | 28| 15.3k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 5.40k, False: 9.93k] ------------------ 29| 5.40k| 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| 5.40k| 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| 5.40k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 5.40k| 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| 5.40k| v_count += c0; 68| 5.40k| v_count += c1; 69| 5.40k| v_count += vector_u16(is_surrogate); 70| | 71| 5.40k| iteration -= 1; 72| 5.40k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 5.40k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 5.40k| } 78| | 79| 9.93k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 9.93k, False: 0] ------------------ 80| 9.93k| count += v_count.sum(); 81| 9.93k| } 82| | 83| 9.93k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 9.93k| size - pos); 85| 9.93k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE1EEEmPKDsm: 10| 9.93k| size_t size) { 11| 9.93k| size_t pos = 0; 12| | 13| 9.93k| using vector_u16 = simd16; 14| 9.93k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 9.93k| const auto one = vector_u16::splat(1); 17| | 18| 9.93k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 9.93k| 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| 9.93k| constexpr size_t max_iterations = 65535 / 2; 26| 9.93k| size_t iteration = max_iterations; 27| | 28| 15.5k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 5.56k, False: 9.93k] ------------------ 29| 5.56k| auto input = vector_u16::load(reinterpret_cast(in + pos)); 30| 5.56k| if constexpr (!match_system(big_endian)) { 31| 5.56k| input = input.swap_bytes(); 32| 5.56k| } 33| | // 0xd800 .. 0xdbff - low surrogate 34| | // 0xdc00 .. 0xdfff - high surrogate 35| 5.56k| 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| 5.56k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 5.56k| 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| 5.56k| v_count += c0; 68| 5.56k| v_count += c1; 69| 5.56k| v_count += vector_u16(is_surrogate); 70| | 71| 5.56k| iteration -= 1; 72| 5.56k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 5.56k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 5.56k| } 78| | 79| 9.93k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 9.93k, False: 0] ------------------ 80| 9.93k| count += v_count.sum(); 81| 9.93k| } 82| | 83| 9.93k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 9.93k| size - pos); 85| 9.93k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE0EEEmPKDsm: 10| 10.1k| size_t size) { 11| 10.1k| size_t pos = 0; 12| | 13| 10.1k| using vector_u16 = simd16; 14| 10.1k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 10.1k| const auto one = vector_u16::splat(1); 17| | 18| 10.1k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 10.1k| 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| 10.1k| constexpr size_t max_iterations = 65535 / 2; 26| 10.1k| size_t iteration = max_iterations; 27| | 28| 26.2k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 16.1k, False: 10.1k] ------------------ 29| 16.1k| 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| 16.1k| 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| 16.1k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 16.1k| 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| 16.1k| v_count += c0; 68| 16.1k| v_count += c1; 69| 16.1k| v_count += vector_u16(is_surrogate); 70| | 71| 16.1k| iteration -= 1; 72| 16.1k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 16.1k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 16.1k| } 78| | 79| 10.1k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 10.1k, False: 0] ------------------ 80| 10.1k| count += v_count.sum(); 81| 10.1k| } 82| | 83| 10.1k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 10.1k| size - pos); 85| 10.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1631utf8_length_from_utf16_bytemaskILNS_10endiannessE1EEEmPKDsm: 10| 10.1k| size_t size) { 11| 10.1k| size_t pos = 0; 12| | 13| 10.1k| using vector_u16 = simd16; 14| 10.1k| constexpr size_t N = vector_u16::ELEMENTS; 15| | 16| 10.1k| const auto one = vector_u16::splat(1); 17| | 18| 10.1k| auto v_count = vector_u16::zero(); 19| | 20| | // each char16 yields at least one byte 21| 10.1k| 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| 10.1k| constexpr size_t max_iterations = 65535 / 2; 26| 10.1k| size_t iteration = max_iterations; 27| | 28| 26.4k| for (; pos < size / N * N; pos += N) { ------------------ | Branch (28:10): [True: 16.3k, False: 10.1k] ------------------ 29| 16.3k| auto input = vector_u16::load(reinterpret_cast(in + pos)); 30| 16.3k| if constexpr (!match_system(big_endian)) { 31| 16.3k| input = input.swap_bytes(); 32| 16.3k| } 33| | // 0xd800 .. 0xdbff - low surrogate 34| | // 0xdc00 .. 0xdfff - high surrogate 35| 16.3k| 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| 16.3k| const auto c0 = min(input & uint16_t(0xff80), one); 39| | 40| | // c1 - chars that yield 3-byte UTF-8 codes (including surrogates) 41| 16.3k| 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| 16.3k| v_count += c0; 68| 16.3k| v_count += c1; 69| 16.3k| v_count += vector_u16(is_surrogate); 70| | 71| 16.3k| iteration -= 1; 72| 16.3k| if (iteration == 0) { ------------------ | Branch (72:9): [True: 0, False: 16.3k] ------------------ 73| 0| count += v_count.sum(); 74| 0| v_count = vector_u16::zero(); 75| 0| iteration = max_iterations; 76| 0| } 77| 16.3k| } 78| | 79| 10.1k| if (iteration > 0) { ------------------ | Branch (79:7): [True: 10.1k, False: 0] ------------------ 80| 10.1k| count += v_count.sum(); 81| 10.1k| } 82| | 83| 10.1k| return count + scalar::utf16::utf8_length_from_utf16(in + pos, 84| 10.1k| size - pos); 85| 10.1k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf3222utf8_length_from_utf32EPKDim: 11| 4.66k| size_t length) { 12| 4.66k| using vector_u32 = simd32; 13| | 14| 4.66k| 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| 4.66k| const size_t max_increment = 3; 19| | 20| 4.66k| 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| 4.66k| const auto v_ffffff80 = vector_u32::splat(0xffffff80); 28| 4.66k| const auto v_fffff800 = vector_u32::splat(0xfffff800); 29| 4.66k| const auto v_ffff0000 = vector_u32::splat(0xffff0000); 30| 4.66k| const auto one = vector_u32::splat(1); 31| 4.66k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 32| | 33| 4.66k| size_t counter = 0; 34| | 35| | // 1. vectorized loop unrolled 4 times 36| 4.66k| { 37| | // we use vector of uint32 counters, this is why this limit is used 38| 4.66k| const size_t max_iterations = 39| 4.66k| std::numeric_limits::max() / (max_increment * 4); 40| 4.66k| size_t blocks = length / (N * 4); 41| 4.66k| length -= blocks * (N * 4); 42| 4.89k| while (blocks != 0) { ------------------ | Branch (42:12): [True: 227, False: 4.66k] ------------------ 43| 227| const size_t iterations = min(blocks, max_iterations); 44| 227| blocks -= iterations; 45| | 46| 227| simd32 acc = vector_u32::zero(); 47| 1.85k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (47:26): [True: 1.62k, False: 227] ------------------ 48| 1.62k| const auto in0 = vector_u32(input + 0 * N); 49| 1.62k| const auto in1 = vector_u32(input + 1 * N); 50| 1.62k| const auto in2 = vector_u32(input + 2 * N); 51| 1.62k| 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| 1.62k| acc += min(one, in0 & v_ffffff80); 70| 1.62k| acc += min(one, in1 & v_ffffff80); 71| 1.62k| acc += min(one, in2 & v_ffffff80); 72| 1.62k| acc += min(one, in3 & v_ffffff80); 73| | 74| 1.62k| acc += min(one, in0 & v_fffff800); 75| 1.62k| acc += min(one, in1 & v_fffff800); 76| 1.62k| acc += min(one, in2 & v_fffff800); 77| 1.62k| acc += min(one, in3 & v_fffff800); 78| | 79| 1.62k| acc += min(one, in0 & v_ffff0000); 80| 1.62k| acc += min(one, in1 & v_ffff0000); 81| 1.62k| acc += min(one, in2 & v_ffff0000); 82| 1.62k| acc += min(one, in3 & v_ffff0000); 83| 1.62k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 84| | 85| 1.62k| input += 4 * N; 86| 1.62k| } 87| | 88| 227| counter += acc.sum(); 89| 227| } 90| 4.66k| } 91| | 92| | // 2. vectorized loop for tail 93| 4.66k| { 94| 4.66k| const size_t max_iterations = 95| 4.66k| std::numeric_limits::max() / max_increment; 96| 4.66k| size_t blocks = length / N; 97| 4.66k| length -= blocks * N; 98| 5.08k| while (blocks != 0) { ------------------ | Branch (98:12): [True: 419, False: 4.66k] ------------------ 99| 419| const size_t iterations = min(blocks, max_iterations); 100| 419| blocks -= iterations; 101| | 102| 419| auto acc = vector_u32::zero(); 103| 1.38k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (103:26): [True: 966, False: 419] ------------------ 104| 966| 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| 966| acc += min(one, in & v_ffffff80); 112| 966| acc += min(one, in & v_fffff800); 113| 966| acc += min(one, in & v_ffff0000); 114| 966|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 115| | 116| 966| input += N; 117| 966| } 118| | 119| 419| counter += acc.sum(); 120| 419| } 121| 4.66k| } 122| | 123| 4.66k| const size_t consumed = input - start; 124| 4.66k| if (consumed != 0) { ------------------ | Branch (124:7): [True: 488, False: 4.18k] ------------------ 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| 488| counter += consumed; 128| 488| } 129| | 130| 4.66k| return counter + scalar::utf32::utf8_length_from_utf32(input, length); 131| 4.66k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf323minImEET_S4_S4_: 8| 646|template T min(T a, T b) { return a <= b ? a : b; } ------------------ | Branch (8:48): [True: 646, False: 0] ------------------ simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf3222utf8_length_from_utf32EPKDim: 11| 4.86k| size_t length) { 12| 4.86k| using vector_u32 = simd32; 13| | 14| 4.86k| 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| 4.86k| const size_t max_increment = 3; 19| | 20| 4.86k| 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| 4.86k| const auto v_ffffff80 = vector_u32::splat(0xffffff80); 28| 4.86k| const auto v_fffff800 = vector_u32::splat(0xfffff800); 29| 4.86k| const auto v_ffff0000 = vector_u32::splat(0xffff0000); 30| 4.86k| const auto one = vector_u32::splat(1); 31| 4.86k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 32| | 33| 4.86k| size_t counter = 0; 34| | 35| | // 1. vectorized loop unrolled 4 times 36| 4.86k| { 37| | // we use vector of uint32 counters, this is why this limit is used 38| 4.86k| const size_t max_iterations = 39| 4.86k| std::numeric_limits::max() / (max_increment * 4); 40| 4.86k| size_t blocks = length / (N * 4); 41| 4.86k| length -= blocks * (N * 4); 42| 5.34k| while (blocks != 0) { ------------------ | Branch (42:12): [True: 484, False: 4.86k] ------------------ 43| 484| const size_t iterations = min(blocks, max_iterations); 44| 484| blocks -= iterations; 45| | 46| 484| simd32 acc = vector_u32::zero(); 47| 5.48k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (47:26): [True: 5.00k, False: 484] ------------------ 48| 5.00k| const auto in0 = vector_u32(input + 0 * N); 49| 5.00k| const auto in1 = vector_u32(input + 1 * N); 50| 5.00k| const auto in2 = vector_u32(input + 2 * N); 51| 5.00k| 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| 5.00k| acc += min(one, in0 & v_ffffff80); 70| 5.00k| acc += min(one, in1 & v_ffffff80); 71| 5.00k| acc += min(one, in2 & v_ffffff80); 72| 5.00k| acc += min(one, in3 & v_ffffff80); 73| | 74| 5.00k| acc += min(one, in0 & v_fffff800); 75| 5.00k| acc += min(one, in1 & v_fffff800); 76| 5.00k| acc += min(one, in2 & v_fffff800); 77| 5.00k| acc += min(one, in3 & v_fffff800); 78| | 79| 5.00k| acc += min(one, in0 & v_ffff0000); 80| 5.00k| acc += min(one, in1 & v_ffff0000); 81| 5.00k| acc += min(one, in2 & v_ffff0000); 82| 5.00k| acc += min(one, in3 & v_ffff0000); 83| 5.00k|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 84| | 85| 5.00k| input += 4 * N; 86| 5.00k| } 87| | 88| 484| counter += acc.sum(); 89| 484| } 90| 4.86k| } 91| | 92| | // 2. vectorized loop for tail 93| 4.86k| { 94| 4.86k| const size_t max_iterations = 95| 4.86k| std::numeric_limits::max() / max_increment; 96| 4.86k| size_t blocks = length / N; 97| 4.86k| length -= blocks * N; 98| 5.28k| while (blocks != 0) { ------------------ | Branch (98:12): [True: 426, False: 4.86k] ------------------ 99| 426| const size_t iterations = min(blocks, max_iterations); 100| 426| blocks -= iterations; 101| | 102| 426| auto acc = vector_u32::zero(); 103| 1.41k| for (size_t i = 0; i < iterations; i++) { ------------------ | Branch (103:26): [True: 990, False: 426] ------------------ 104| 990| 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| 990| acc += min(one, in & v_ffffff80); 112| 990| acc += min(one, in & v_fffff800); 113| 990| acc += min(one, in & v_ffff0000); 114| 990|#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP 115| | 116| 990| input += N; 117| 990| } 118| | 119| 426| counter += acc.sum(); 120| 426| } 121| 4.86k| } 122| | 123| 4.86k| const size_t consumed = input - start; 124| 4.86k| if (consumed != 0) { ------------------ | Branch (124:7): [True: 524, False: 4.33k] ------------------ 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| 524| counter += consumed; 128| 524| } 129| | 130| 4.86k| return counter + scalar::utf32::utf8_length_from_utf32(input, length); 131| 4.86k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf323minImEET_S4_S4_: 8| 910|template T min(T a, T b) { return a <= b ? a : b; } ------------------ | Branch (8:48): [True: 910, False: 0] ------------------ simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14utf826count_code_points_bytemaskEPKcm: 21| 10.5k|count_code_points_bytemask(const char *in, size_t size) { 22| 10.5k| using vector_i8 = simd8; 23| 10.5k| using vector_u8 = simd8; 24| 10.5k| using vector_u64 = simd64; 25| | 26| 10.5k| constexpr size_t N = vector_i8::SIZE; 27| 10.5k| constexpr size_t max_iterations = 255 / 4; 28| | 29| 10.5k| size_t pos = 0; 30| 10.5k| size_t count = 0; 31| | 32| 10.5k| auto counters = vector_u64::zero(); 33| 10.5k| auto local = vector_u8::zero(); 34| 10.5k| size_t iterations = 0; 35| 12.3k| for (; pos + 4 * N <= size; pos += 4 * N) { ------------------ | Branch (35:10): [True: 1.84k, False: 10.5k] ------------------ 36| 1.84k| const auto input0 = 37| 1.84k| simd8::load(reinterpret_cast(in + pos + 0 * N)); 38| 1.84k| const auto input1 = 39| 1.84k| simd8::load(reinterpret_cast(in + pos + 1 * N)); 40| 1.84k| const auto input2 = 41| 1.84k| simd8::load(reinterpret_cast(in + pos + 2 * N)); 42| 1.84k| const auto input3 = 43| 1.84k| simd8::load(reinterpret_cast(in + pos + 3 * N)); 44| 1.84k| const auto mask0 = input0 > int8_t(-65); 45| 1.84k| const auto mask1 = input1 > int8_t(-65); 46| 1.84k| const auto mask2 = input2 > int8_t(-65); 47| 1.84k| const auto mask3 = input3 > int8_t(-65); 48| | 49| 1.84k| local -= vector_u8(mask0); 50| 1.84k| local -= vector_u8(mask1); 51| 1.84k| local -= vector_u8(mask2); 52| 1.84k| local -= vector_u8(mask3); 53| | 54| 1.84k| iterations += 1; 55| 1.84k| if (iterations == max_iterations) { ------------------ | Branch (55:9): [True: 0, False: 1.84k] ------------------ 56| 0| counters += sum_8bytes(local); 57| 0| local = vector_u8::zero(); 58| 0| iterations = 0; 59| 0| } 60| 1.84k| } 61| | 62| 10.5k| if (iterations > 0) { ------------------ | Branch (62:7): [True: 454, False: 10.0k] ------------------ 63| 454| count += local.sum_bytes(); 64| 454| } 65| | 66| 10.5k| count += counters.sum(); 67| | 68| 10.5k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 69| 10.5k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14utf817count_code_pointsEPKcm: 8| 5.26k|simdutf_really_inline size_t count_code_points(const char *in, size_t size) { 9| 5.26k| size_t pos = 0; 10| 5.26k| size_t count = 0; 11| 7.03k| for (; pos + 64 <= size; pos += 64) { ------------------ | Branch (11:10): [True: 1.76k, False: 5.26k] ------------------ 12| 1.76k| simd8x64 input(reinterpret_cast(in + pos)); 13| 1.76k| uint64_t utf8_continuation_mask = input.gt(-65); 14| 1.76k| count += count_ones(utf8_continuation_mask); 15| 1.76k| } 16| 5.26k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 17| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14utf826count_code_points_bytemaskEPKcm: 21| 10.5k|count_code_points_bytemask(const char *in, size_t size) { 22| 10.5k| using vector_i8 = simd8; 23| 10.5k| using vector_u8 = simd8; 24| 10.5k| using vector_u64 = simd64; 25| | 26| 10.5k| constexpr size_t N = vector_i8::SIZE; 27| 10.5k| constexpr size_t max_iterations = 255 / 4; 28| | 29| 10.5k| size_t pos = 0; 30| 10.5k| size_t count = 0; 31| | 32| 10.5k| auto counters = vector_u64::zero(); 33| 10.5k| auto local = vector_u8::zero(); 34| 10.5k| size_t iterations = 0; 35| 16.0k| for (; pos + 4 * N <= size; pos += 4 * N) { ------------------ | Branch (35:10): [True: 5.51k, False: 10.5k] ------------------ 36| 5.51k| const auto input0 = 37| 5.51k| simd8::load(reinterpret_cast(in + pos + 0 * N)); 38| 5.51k| const auto input1 = 39| 5.51k| simd8::load(reinterpret_cast(in + pos + 1 * N)); 40| 5.51k| const auto input2 = 41| 5.51k| simd8::load(reinterpret_cast(in + pos + 2 * N)); 42| 5.51k| const auto input3 = 43| 5.51k| simd8::load(reinterpret_cast(in + pos + 3 * N)); 44| 5.51k| const auto mask0 = input0 > int8_t(-65); 45| 5.51k| const auto mask1 = input1 > int8_t(-65); 46| 5.51k| const auto mask2 = input2 > int8_t(-65); 47| 5.51k| const auto mask3 = input3 > int8_t(-65); 48| | 49| 5.51k| local -= vector_u8(mask0); 50| 5.51k| local -= vector_u8(mask1); 51| 5.51k| local -= vector_u8(mask2); 52| 5.51k| local -= vector_u8(mask3); 53| | 54| 5.51k| iterations += 1; 55| 5.51k| if (iterations == max_iterations) { ------------------ | Branch (55:9): [True: 0, False: 5.51k] ------------------ 56| 0| counters += sum_8bytes(local); 57| 0| local = vector_u8::zero(); 58| 0| iterations = 0; 59| 0| } 60| 5.51k| } 61| | 62| 10.5k| if (iterations > 0) { ------------------ | Branch (62:7): [True: 1.02k, False: 9.50k] ------------------ 63| 1.02k| count += local.sum_bytes(); 64| 1.02k| } 65| | 66| 10.5k| count += counters.sum(); 67| | 68| 10.5k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 69| 10.5k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14utf817count_code_pointsEPKcm: 8| 5.26k|simdutf_really_inline size_t count_code_points(const char *in, size_t size) { 9| 5.26k| size_t pos = 0; 10| 5.26k| size_t count = 0; 11| 7.81k| for (; pos + 64 <= size; pos += 64) { ------------------ | Branch (11:10): [True: 2.54k, False: 5.26k] ------------------ 12| 2.54k| simd8x64 input(reinterpret_cast(in + pos)); 13| 2.54k| uint64_t utf8_continuation_mask = input.gt(-65); 14| 2.54k| count += count_ones(utf8_continuation_mask); 15| 2.54k| } 16| 5.26k| return count + scalar::utf8::count_code_points(in + pos, size - pos); 17| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14utf831utf16_length_from_utf8_bytemaskEPKcm: 9| 20.8k| size_t size) { 10| 20.8k| using vector_i8 = simd8; 11| 20.8k| using vector_u8 = simd8; 12| 20.8k| using vector_u64 = simd64; 13| | 14| 20.8k| constexpr size_t N = vector_i8::SIZE; 15| 20.8k| constexpr size_t max_iterations = 255 / 2; 16| | 17| 20.8k| auto counters = vector_u64::zero(); 18| 20.8k| auto local = vector_u8::zero(); 19| | 20| 20.8k| size_t iterations = 0; 21| 20.8k| size_t pos = 0; 22| 20.8k| size_t count = 0; 23| 32.2k| for (; pos + N <= size; pos += N) { ------------------ | Branch (23:10): [True: 11.3k, False: 20.8k] ------------------ 24| 11.3k| const auto input = 25| 11.3k| vector_i8::load(reinterpret_cast(in + pos)); 26| | 27| 11.3k| const auto continuation = input > int8_t(-65); 28| 11.3k| const auto utf_4bytes = vector_u8(input.value) >= uint8_t(240); 29| | 30| 11.3k| local -= vector_u8(continuation); 31| 11.3k| local -= vector_u8(utf_4bytes); 32| | 33| 11.3k| iterations += 1; 34| 11.3k| if (iterations == max_iterations) { ------------------ | Branch (34:9): [True: 0, False: 11.3k] ------------------ 35| 0| counters += sum_8bytes(local); 36| 0| local = vector_u8::zero(); 37| 0| iterations = 0; 38| 0| } 39| 11.3k| } 40| | 41| 20.8k| if (iterations > 0) { ------------------ | Branch (41:7): [True: 1.62k, False: 19.2k] ------------------ 42| 1.62k| count += local.sum_bytes(); 43| 1.62k| } 44| | 45| 20.8k| count += counters.sum(); 46| | 47| 20.8k| return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos); 48| 20.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14utf831utf16_length_from_utf8_bytemaskEPKcm: 9| 20.8k| size_t size) { 10| 20.8k| using vector_i8 = simd8; 11| 20.8k| using vector_u8 = simd8; 12| 20.8k| using vector_u64 = simd64; 13| | 14| 20.8k| constexpr size_t N = vector_i8::SIZE; 15| 20.8k| constexpr size_t max_iterations = 255 / 2; 16| | 17| 20.8k| auto counters = vector_u64::zero(); 18| 20.8k| auto local = vector_u8::zero(); 19| | 20| 20.8k| size_t iterations = 0; 21| 20.8k| size_t pos = 0; 22| 20.8k| size_t count = 0; 23| 54.1k| for (; pos + N <= size; pos += N) { ------------------ | Branch (23:10): [True: 33.3k, False: 20.8k] ------------------ 24| 33.3k| const auto input = 25| 33.3k| vector_i8::load(reinterpret_cast(in + pos)); 26| | 27| 33.3k| const auto continuation = input > int8_t(-65); 28| 33.3k| const auto utf_4bytes = vector_u8(input.value) >= uint8_t(240); 29| | 30| 33.3k| local -= vector_u8(continuation); 31| 33.3k| local -= vector_u8(utf_4bytes); 32| | 33| 33.3k| iterations += 1; 34| 33.3k| if (iterations == max_iterations) { ------------------ | Branch (34:9): [True: 0, False: 33.3k] ------------------ 35| 0| counters += sum_8bytes(local); 36| 0| local = vector_u8::zero(); 37| 0| iterations = 0; 38| 0| } 39| 33.3k| } 40| | 41| 20.8k| if (iterations > 0) { ------------------ | Branch (41:7): [True: 1.90k, False: 18.9k] ------------------ 42| 1.90k| count += local.sum_bytes(); 43| 1.90k| } 44| | 45| 20.8k| count += counters.sum(); 46| | 47| 20.8k| return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos); 48| 20.8k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoderC2Ev: 101| 10.5k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoder7convertEPKcmPc: 115| 10.5k| char *latin1_output) { 116| 10.5k| size_t pos = 0; 117| 10.5k| 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| 10.5k| size_t leading_byte = 0; 125| 10.5k| size_t margin = size; 126| 49.4k| for (; margin > 0 && leading_byte < 16; margin--) { ------------------ | Branch (126:12): [True: 40.3k, False: 9.00k] | Branch (126:26): [True: 38.8k, False: 1.52k] ------------------ 127| 38.8k| leading_byte += (int8_t(in[margin - 1]) > 128| 38.8k| -65); // twos complement of -65 is 1011 1111 ... 129| 38.8k| } 130| | // If the input is long enough, then we have that margin-1 is the eight last 131| | // leading byte. 132| 10.5k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 133| 14.5k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (133:12): [True: 3.96k, False: 10.5k] ------------------ 134| 3.96k| simd8x64 input(reinterpret_cast(in + pos)); 135| 3.96k| if (input.is_ascii()) { ------------------ | Branch (135:11): [True: 1.68k, False: 2.28k] ------------------ 136| 1.68k| input.store((int8_t *)latin1_output); 137| 1.68k| latin1_output += 64; 138| 1.68k| pos += 64; 139| 2.28k| } else { 140| | // you might think that a for-loop would work, but under Visual Studio, 141| | // it is not good enough. 142| 2.28k| static_assert( 143| 2.28k| (simd8x64::NUM_CHUNKS == 2) || 144| 2.28k| (simd8x64::NUM_CHUNKS == 4), 145| 2.28k| "We support either two or four chunks per 64-byte block."); 146| 2.28k| auto zero = simd8{uint8_t(0)}; 147| 2.28k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 148| 2.28k| this->check_utf8_bytes(input.chunks[0], zero); 149| 2.28k| 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| 2.28k| uint64_t utf8_continuation_mask = 157| 2.28k| input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in 158| | // this case, we also have ASCII to account for. 159| 2.28k| if (utf8_continuation_mask & 1) { ------------------ | Branch (159:13): [True: 5, False: 2.27k] ------------------ 160| 5| return 0; // error 161| 5| } 162| 2.27k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 163| 2.27k| 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| 2.27k| size_t max_starting_point = (pos + 64) - 12; 168| | // Next loop is going to run at least five times. 169| 18.8k| while (pos < max_starting_point) { ------------------ | Branch (169:16): [True: 16.5k, False: 2.27k] ------------------ 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| 16.5k| size_t consumed = convert_masked_utf8_to_latin1( 180| 16.5k| in + pos, utf8_end_of_code_point_mask, latin1_output); 181| 16.5k| pos += consumed; 182| 16.5k| utf8_end_of_code_point_mask >>= consumed; 183| 16.5k| } 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| 2.27k| } 189| 3.96k| } 190| 10.5k| if (errors()) { ------------------ | Branch (190:9): [True: 175, False: 10.3k] ------------------ 191| 175| return 0; 192| 175| } 193| 10.3k| if (pos < size) { ------------------ | Branch (193:9): [True: 2.30k, False: 8.04k] ------------------ 194| 2.30k| size_t howmany = 195| 2.30k| scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output); 196| 2.30k| if (howmany == 0) { ------------------ | Branch (196:11): [True: 940, False: 1.36k] ------------------ 197| 940| return 0; 198| 940| } 199| 1.36k| latin1_output += howmany; 200| 1.36k| } 201| 9.41k| return latin1_output - start; 202| 10.3k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 106| 4.56k| 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| 4.56k| simd8 prev1 = input.prev<1>(prev_input); 111| 4.56k| this->error |= check_special_cases(input, prev1); 112| 4.56k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_114utf8_to_latin119check_special_casesENS1_4simd5simd8IhEES5_: 8| 4.56k|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| 4.56k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 20| | // 11______ 11______ 21| 4.56k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 22| 4.56k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 23| 4.56k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 24| 4.56k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 25| 4.56k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 26| 4.56k| 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| 4.56k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 35| | // 11110101 1000____ 36| | // 1111011_ 1000____ 37| | // 11111___ 1000____ 38| 4.56k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 39| 4.56k| constexpr const uint8_t FORBIDDEN = 0xff; 40| | 41| 4.56k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 42| | // 0_______ ________ 43| 4.56k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 44| 4.56k| TOO_LONG, 45| | // 10______ ________ 46| 4.56k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 47| | // 1100____ ________ 48| 4.56k| TOO_SHORT | OVERLONG_2, 49| | // 1101____ ________ 50| 4.56k| FORBIDDEN, 51| | // 1110____ ________ 52| 4.56k| FORBIDDEN, 53| | // 1111____ ________ 54| 4.56k| FORBIDDEN); 55| 4.56k| constexpr const uint8_t CARRY = 56| 4.56k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 57| 4.56k| const simd8 byte_1_low = 58| 4.56k| (prev1 & 0x0F) 59| 4.56k| .lookup_16( 60| | // ____0000 ________ 61| 4.56k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 62| | // ____0001 ________ 63| 4.56k| CARRY | OVERLONG_2, 64| | // ____001_ ________ 65| 4.56k| CARRY, CARRY, 66| | 67| | // ____0100 ________ 68| 4.56k| FORBIDDEN, 69| | // ____0101 ________ 70| 4.56k| FORBIDDEN, 71| | // ____011_ ________ 72| 4.56k| FORBIDDEN, FORBIDDEN, 73| | 74| | // ____1___ ________ 75| 4.56k| FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, 76| | // ____1101 ________ 77| 4.56k| FORBIDDEN, FORBIDDEN, FORBIDDEN); 78| 4.56k| const simd8 byte_2_high = input.shr<4>().lookup_16( 79| | // ________ 0_______ 80| 4.56k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 81| 4.56k| TOO_SHORT, TOO_SHORT, 82| | 83| | // ________ 1000____ 84| 4.56k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 85| 4.56k| OVERLONG_4, 86| | // ________ 1001____ 87| 4.56k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 88| | // ________ 101_____ 89| 4.56k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 90| 4.56k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 91| | 92| | // ________ 11______ 93| 4.56k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 94| 4.56k| return (byte_1_high & byte_1_low & byte_2_high); 95| 4.56k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_114utf8_to_latin121validating_transcoder6errorsEv: 308| 10.5k| simdutf_really_inline bool errors() const { 309| 10.5k| return this->error.any_bits_set_anywhere(); 310| 10.5k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoderC2Ev: 101| 10.5k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoder7convertEPKcmPc: 115| 10.5k| char *latin1_output) { 116| 10.5k| size_t pos = 0; 117| 10.5k| 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| 10.5k| size_t leading_byte = 0; 125| 10.5k| size_t margin = size; 126| 40.5k| for (; margin > 0 && leading_byte < 16; margin--) { ------------------ | Branch (126:12): [True: 31.4k, False: 9.19k] | Branch (126:26): [True: 30.0k, False: 1.34k] ------------------ 127| 30.0k| leading_byte += (int8_t(in[margin - 1]) > 128| 30.0k| -65); // twos complement of -65 is 1011 1111 ... 129| 30.0k| } 130| | // If the input is long enough, then we have that margin-1 is the eight last 131| | // leading byte. 132| 10.5k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 133| 15.6k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (133:12): [True: 5.12k, False: 10.5k] ------------------ 134| 5.12k| simd8x64 input(reinterpret_cast(in + pos)); 135| 5.12k| if (input.is_ascii()) { ------------------ | Branch (135:11): [True: 2.03k, False: 3.08k] ------------------ 136| 2.03k| input.store((int8_t *)latin1_output); 137| 2.03k| latin1_output += 64; 138| 2.03k| pos += 64; 139| 3.08k| } else { 140| | // you might think that a for-loop would work, but under Visual Studio, 141| | // it is not good enough. 142| 3.08k| static_assert( 143| 3.08k| (simd8x64::NUM_CHUNKS == 2) || 144| 3.08k| (simd8x64::NUM_CHUNKS == 4), 145| 3.08k| "We support either two or four chunks per 64-byte block."); 146| 3.08k| 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| 3.08k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 151| 3.08k| this->check_utf8_bytes(input.chunks[0], zero); 152| 3.08k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 153| 3.08k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 154| 3.08k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 155| 3.08k| } 156| 3.08k| uint64_t utf8_continuation_mask = 157| 3.08k| input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in 158| | // this case, we also have ASCII to account for. 159| 3.08k| if (utf8_continuation_mask & 1) { ------------------ | Branch (159:13): [True: 11, False: 3.07k] ------------------ 160| 11| return 0; // error 161| 11| } 162| 3.07k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 163| 3.07k| 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| 3.07k| size_t max_starting_point = (pos + 64) - 12; 168| | // Next loop is going to run at least five times. 169| 25.6k| while (pos < max_starting_point) { ------------------ | Branch (169:16): [True: 22.5k, False: 3.07k] ------------------ 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| 22.5k| size_t consumed = convert_masked_utf8_to_latin1( 180| 22.5k| in + pos, utf8_end_of_code_point_mask, latin1_output); 181| 22.5k| pos += consumed; 182| 22.5k| utf8_end_of_code_point_mask >>= consumed; 183| 22.5k| } 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| 3.07k| } 189| 5.12k| } 190| 10.5k| if (errors()) { ------------------ | Branch (190:9): [True: 218, False: 10.3k] ------------------ 191| 218| return 0; 192| 218| } 193| 10.3k| if (pos < size) { ------------------ | Branch (193:9): [True: 1.64k, False: 8.66k] ------------------ 194| 1.64k| size_t howmany = 195| 1.64k| scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output); 196| 1.64k| if (howmany == 0) { ------------------ | Branch (196:11): [True: 588, False: 1.05k] ------------------ 197| 588| return 0; 198| 588| } 199| 1.05k| latin1_output += howmany; 200| 1.05k| } 201| 9.71k| return latin1_output - start; 202| 10.3k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 106| 12.3k| 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| 12.3k| simd8 prev1 = input.prev<1>(prev_input); 111| 12.3k| this->error |= check_special_cases(input, prev1); 112| 12.3k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_114utf8_to_latin119check_special_casesENS1_4simd5simd8IhEES5_: 8| 12.3k|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| 12.3k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 20| | // 11______ 11______ 21| 12.3k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 22| 12.3k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 23| 12.3k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 24| 12.3k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 25| 12.3k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 26| 12.3k| 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| 12.3k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 35| | // 11110101 1000____ 36| | // 1111011_ 1000____ 37| | // 11111___ 1000____ 38| 12.3k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 39| 12.3k| constexpr const uint8_t FORBIDDEN = 0xff; 40| | 41| 12.3k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 42| | // 0_______ ________ 43| 12.3k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 44| 12.3k| TOO_LONG, 45| | // 10______ ________ 46| 12.3k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 47| | // 1100____ ________ 48| 12.3k| TOO_SHORT | OVERLONG_2, 49| | // 1101____ ________ 50| 12.3k| FORBIDDEN, 51| | // 1110____ ________ 52| 12.3k| FORBIDDEN, 53| | // 1111____ ________ 54| 12.3k| FORBIDDEN); 55| 12.3k| constexpr const uint8_t CARRY = 56| 12.3k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 57| 12.3k| const simd8 byte_1_low = 58| 12.3k| (prev1 & 0x0F) 59| 12.3k| .lookup_16( 60| | // ____0000 ________ 61| 12.3k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 62| | // ____0001 ________ 63| 12.3k| CARRY | OVERLONG_2, 64| | // ____001_ ________ 65| 12.3k| CARRY, CARRY, 66| | 67| | // ____0100 ________ 68| 12.3k| FORBIDDEN, 69| | // ____0101 ________ 70| 12.3k| FORBIDDEN, 71| | // ____011_ ________ 72| 12.3k| FORBIDDEN, FORBIDDEN, 73| | 74| | // ____1___ ________ 75| 12.3k| FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, 76| | // ____1101 ________ 77| 12.3k| FORBIDDEN, FORBIDDEN, FORBIDDEN); 78| 12.3k| const simd8 byte_2_high = input.shr<4>().lookup_16( 79| | // ________ 0_______ 80| 12.3k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 81| 12.3k| TOO_SHORT, TOO_SHORT, 82| | 83| | // ________ 1000____ 84| 12.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 85| 12.3k| OVERLONG_4, 86| | // ________ 1001____ 87| 12.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 88| | // ________ 101_____ 89| 12.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 90| 12.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 91| | 92| | // ________ 11______ 93| 12.3k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 94| 12.3k| return (byte_1_high & byte_1_low & byte_2_high); 95| 12.3k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_114utf8_to_latin121validating_transcoder6errorsEv: 308| 10.5k| simdutf_really_inline bool errors() const { 309| 10.5k| return this->error.any_bits_set_anywhere(); 310| 10.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoderC2Ev: 114| 20.8k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoderC2Ev: 114| 20.8k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE0EEEmPKcmPDs: 130| 10.4k| char16_t *utf16_output) { 131| 10.4k| size_t pos = 0; 132| 10.4k| 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| 10.4k| size_t leading_byte = 0; 140| 10.4k| size_t margin = size; 141| 27.3k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 17.8k, False: 9.42k] | Branch (141:26): [True: 16.8k, False: 1.01k] ------------------ 142| 16.8k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 16.8k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 10.4k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 13.0k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 2.64k, False: 10.4k] ------------------ 148| 2.64k| simd8x64 input(reinterpret_cast(in + pos)); 149| 2.64k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 578, False: 2.06k] ------------------ 150| 578| input.store_ascii_as_utf16(utf16_output); 151| 578| utf16_output += 64; 152| 578| pos += 64; 153| 2.06k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 2.06k| static_assert( 157| 2.06k| (simd8x64::NUM_CHUNKS == 2) || 158| 2.06k| (simd8x64::NUM_CHUNKS == 4), 159| 2.06k| "We support either two or four chunks per 64-byte block."); 160| 2.06k| auto zero = simd8{uint8_t(0)}; 161| 2.06k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 162| 2.06k| this->check_utf8_bytes(input.chunks[0], zero); 163| 2.06k| 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| 2.06k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 2.06k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 7, False: 2.06k] ------------------ 172| 7| return 0; // error 173| 7| } 174| 2.06k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 2.06k| 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| 2.06k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 17.2k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 15.1k, False: 2.06k] ------------------ 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| 15.1k| size_t consumed = convert_masked_utf8_to_utf16( 192| 15.1k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 15.1k| pos += consumed; 194| 15.1k| utf8_end_of_code_point_mask >>= consumed; 195| 15.1k| } 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| 2.06k| } 201| 2.64k| } 202| 10.4k| if (errors()) { ------------------ | Branch (202:9): [True: 193, False: 10.2k] ------------------ 203| 193| return 0; 204| 193| } 205| 10.2k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.44k, False: 8.79k] ------------------ 206| 1.44k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.44k| in + pos, size - pos, utf16_output); 208| 1.44k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 400, False: 1.04k] ------------------ 209| 400| return 0; 210| 400| } 211| 1.04k| utf16_output += howmany; 212| 1.04k| } 213| 9.83k| return utf16_output - start; 214| 10.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 8.46k| 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| 8.46k| simd8 prev1 = input.prev<1>(prev_input); 124| 8.46k| simd8 sc = check_special_cases(input, prev1); 125| 8.46k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 8.46k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1619check_special_casesENS1_4simd5simd8IhEES5_: 8| 8.46k|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| 8.46k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 8.46k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 8.46k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 8.46k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 8.46k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 8.46k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 8.46k| 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| 8.46k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 8.46k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 8.46k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 8.46k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 8.46k| TOO_LONG, 40| | // 10______ ________ 41| 8.46k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 8.46k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 8.46k| TOO_SHORT, 46| | // 1110____ ________ 47| 8.46k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 8.46k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 8.46k| constexpr const uint8_t CARRY = 51| 8.46k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 8.46k| const simd8 byte_1_low = 53| 8.46k| (prev1 & 0x0F) 54| 8.46k| .lookup_16( 55| | // ____0000 ________ 56| 8.46k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 8.46k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 8.46k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 8.46k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 8.46k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 8.46k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 8.46k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 8.46k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 8.46k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 8.46k| OVERLONG_4, 88| | // ________ 1001____ 89| 8.46k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 8.46k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 8.46k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 8.46k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 8.46k| return (byte_1_high & byte_1_low & byte_2_high); 97| 8.46k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1623check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 8.46k| const simd8 sc) { 102| 8.46k| simd8 prev2 = input.prev<2>(prev_input); 103| 8.46k| simd8 prev3 = input.prev<3>(prev_input); 104| 8.46k| simd8 must23 = 105| 8.46k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 8.46k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 8.46k| return must23_80 ^ sc; 108| 8.46k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder6errorsEv: 324| 20.8k| simdutf_really_inline bool errors() const { 325| 20.8k| return this->error.any_bits_set_anywhere(); 326| 20.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE1EEEmPKcmPDs: 130| 10.4k| char16_t *utf16_output) { 131| 10.4k| size_t pos = 0; 132| 10.4k| 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| 10.4k| size_t leading_byte = 0; 140| 10.4k| size_t margin = size; 141| 27.6k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 18.2k, False: 9.42k] | Branch (141:26): [True: 17.1k, False: 1.02k] ------------------ 142| 17.1k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 17.1k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 10.4k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 13.1k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 2.74k, False: 10.4k] ------------------ 148| 2.74k| simd8x64 input(reinterpret_cast(in + pos)); 149| 2.74k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 584, False: 2.16k] ------------------ 150| 584| input.store_ascii_as_utf16(utf16_output); 151| 584| utf16_output += 64; 152| 584| pos += 64; 153| 2.16k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 2.16k| static_assert( 157| 2.16k| (simd8x64::NUM_CHUNKS == 2) || 158| 2.16k| (simd8x64::NUM_CHUNKS == 4), 159| 2.16k| "We support either two or four chunks per 64-byte block."); 160| 2.16k| auto zero = simd8{uint8_t(0)}; 161| 2.16k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 162| 2.16k| this->check_utf8_bytes(input.chunks[0], zero); 163| 2.16k| 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| 2.16k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 2.16k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 7, False: 2.15k] ------------------ 172| 7| return 0; // error 173| 7| } 174| 2.15k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 2.15k| 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| 2.15k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 17.7k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 15.6k, False: 2.15k] ------------------ 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| 2.15k| } 201| 2.74k| } 202| 10.4k| if (errors()) { ------------------ | Branch (202:9): [True: 193, False: 10.2k] ------------------ 203| 193| return 0; 204| 193| } 205| 10.2k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.45k, False: 8.79k] ------------------ 206| 1.45k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.45k| in + pos, size - pos, utf16_output); 208| 1.45k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 400, False: 1.05k] ------------------ 209| 400| return 0; 210| 400| } 211| 1.05k| utf16_output += howmany; 212| 1.05k| } 213| 9.84k| return utf16_output - start; 214| 10.2k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE0EEEmPKcmPDs: 130| 10.4k| char16_t *utf16_output) { 131| 10.4k| size_t pos = 0; 132| 10.4k| 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| 10.4k| size_t leading_byte = 0; 140| 10.4k| size_t margin = size; 141| 23.5k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 14.0k, False: 9.47k] | Branch (141:26): [True: 13.1k, False: 950] ------------------ 142| 13.1k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 13.1k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 10.4k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 14.3k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 3.89k, False: 10.4k] ------------------ 148| 3.89k| simd8x64 input(reinterpret_cast(in + pos)); 149| 3.89k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 916, False: 2.97k] ------------------ 150| 916| input.store_ascii_as_utf16(utf16_output); 151| 916| utf16_output += 64; 152| 916| pos += 64; 153| 2.97k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 2.97k| static_assert( 157| 2.97k| (simd8x64::NUM_CHUNKS == 2) || 158| 2.97k| (simd8x64::NUM_CHUNKS == 4), 159| 2.97k| "We support either two or four chunks per 64-byte block."); 160| 2.97k| 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| 2.97k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 165| 2.97k| this->check_utf8_bytes(input.chunks[0], zero); 166| 2.97k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 167| 2.97k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 168| 2.97k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 169| 2.97k| } 170| 2.97k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 2.97k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 13, False: 2.96k] ------------------ 172| 13| return 0; // error 173| 13| } 174| 2.96k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 2.96k| 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| 2.96k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 23.8k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 20.8k, False: 2.96k] ------------------ 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| 20.8k| size_t consumed = convert_masked_utf8_to_utf16( 192| 20.8k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 20.8k| pos += consumed; 194| 20.8k| utf8_end_of_code_point_mask >>= consumed; 195| 20.8k| } 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| 2.96k| } 201| 3.89k| } 202| 10.4k| if (errors()) { ------------------ | Branch (202:9): [True: 198, False: 10.2k] ------------------ 203| 198| return 0; 204| 198| } 205| 10.2k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.11k, False: 9.10k] ------------------ 206| 1.11k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.11k| in + pos, size - pos, utf16_output); 208| 1.11k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 195, False: 923] ------------------ 209| 195| return 0; 210| 195| } 211| 923| utf16_output += howmany; 212| 923| } 213| 10.0k| return utf16_output - start; 214| 10.2k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 24.4k| 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| 24.4k| simd8 prev1 = input.prev<1>(prev_input); 124| 24.4k| simd8 sc = check_special_cases(input, prev1); 125| 24.4k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 24.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1619check_special_casesENS1_4simd5simd8IhEES5_: 8| 24.4k|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| 24.4k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 24.4k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 24.4k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 24.4k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 24.4k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 24.4k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 24.4k| 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| 24.4k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 24.4k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 24.4k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 24.4k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 24.4k| TOO_LONG, 40| | // 10______ ________ 41| 24.4k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 24.4k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 24.4k| TOO_SHORT, 46| | // 1110____ ________ 47| 24.4k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 24.4k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 24.4k| constexpr const uint8_t CARRY = 51| 24.4k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 24.4k| const simd8 byte_1_low = 53| 24.4k| (prev1 & 0x0F) 54| 24.4k| .lookup_16( 55| | // ____0000 ________ 56| 24.4k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 24.4k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 24.4k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 24.4k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 24.4k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 24.4k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 24.4k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 24.4k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 24.4k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 24.4k| OVERLONG_4, 88| | // ________ 1001____ 89| 24.4k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 24.4k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 24.4k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 24.4k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 24.4k| return (byte_1_high & byte_1_low & byte_2_high); 97| 24.4k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1623check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 24.4k| const simd8 sc) { 102| 24.4k| simd8 prev2 = input.prev<2>(prev_input); 103| 24.4k| simd8 prev3 = input.prev<3>(prev_input); 104| 24.4k| simd8 must23 = 105| 24.4k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 24.4k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 24.4k| return must23_80 ^ sc; 108| 24.4k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder6errorsEv: 324| 20.8k| simdutf_really_inline bool errors() const { 325| 20.8k| return this->error.any_bits_set_anywhere(); 326| 20.8k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf1621validating_transcoder7convertILNS_10endiannessE1EEEmPKcmPDs: 130| 10.4k| char16_t *utf16_output) { 131| 10.4k| size_t pos = 0; 132| 10.4k| 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| 10.4k| size_t leading_byte = 0; 140| 10.4k| size_t margin = size; 141| 23.6k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (141:12): [True: 14.1k, False: 9.48k] | Branch (141:26): [True: 13.2k, False: 952] ------------------ 142| 13.2k| leading_byte += (int8_t(in[margin - 1]) > -65); 143| 13.2k| } 144| | // If the input is long enough, then we have that margin-1 is the eight last 145| | // leading byte. 146| 10.4k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 147| 14.4k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (147:12): [True: 4.05k, False: 10.4k] ------------------ 148| 4.05k| simd8x64 input(reinterpret_cast(in + pos)); 149| 4.05k| if (input.is_ascii()) { ------------------ | Branch (149:11): [True: 915, False: 3.13k] ------------------ 150| 915| input.store_ascii_as_utf16(utf16_output); 151| 915| utf16_output += 64; 152| 915| pos += 64; 153| 3.13k| } else { 154| | // you might think that a for-loop would work, but under Visual Studio, 155| | // it is not good enough. 156| 3.13k| static_assert( 157| 3.13k| (simd8x64::NUM_CHUNKS == 2) || 158| 3.13k| (simd8x64::NUM_CHUNKS == 4), 159| 3.13k| "We support either two or four chunks per 64-byte block."); 160| 3.13k| 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| 3.13k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 165| 3.13k| this->check_utf8_bytes(input.chunks[0], zero); 166| 3.13k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 167| 3.13k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 168| 3.13k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 169| 3.13k| } 170| 3.13k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 171| 3.13k| if (utf8_continuation_mask & 1) { ------------------ | Branch (171:13): [True: 13, False: 3.12k] ------------------ 172| 13| return 0; // error 173| 13| } 174| 3.12k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 175| 3.12k| 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| 3.12k| size_t max_starting_point = (pos + 64) - 12; 180| | // Next loop is going to run at least five times. 181| 24.7k| while (pos < max_starting_point) { ------------------ | Branch (181:16): [True: 21.6k, False: 3.12k] ------------------ 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| 21.6k| size_t consumed = convert_masked_utf8_to_utf16( 192| 21.6k| in + pos, utf8_end_of_code_point_mask, utf16_output); 193| 21.6k| pos += consumed; 194| 21.6k| utf8_end_of_code_point_mask >>= consumed; 195| 21.6k| } 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| 3.12k| } 201| 4.05k| } 202| 10.4k| if (errors()) { ------------------ | Branch (202:9): [True: 198, False: 10.2k] ------------------ 203| 198| return 0; 204| 198| } 205| 10.2k| if (pos < size) { ------------------ | Branch (205:9): [True: 1.11k, False: 9.11k] ------------------ 206| 1.11k| size_t howmany = scalar::utf8_to_utf16::convert( 207| 1.11k| in + pos, size - pos, utf16_output); 208| 1.11k| if (howmany == 0) { ------------------ | Branch (208:11): [True: 195, False: 920] ------------------ 209| 195| return 0; 210| 195| } 211| 920| utf16_output += howmany; 212| 920| } 213| 10.0k| return utf16_output - start; 214| 10.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoderC2Ev: 114| 5.26k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoder7convertEPKcmPDi: 129| 5.26k| char32_t *utf32_output) { 130| 5.26k| size_t pos = 0; 131| 5.26k| 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| 5.26k| size_t leading_byte = 0; 139| 5.26k| size_t margin = size; 140| 17.6k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (140:12): [True: 13.2k, False: 4.42k] | Branch (140:26): [True: 12.3k, False: 843] ------------------ 141| 12.3k| leading_byte += (int8_t(in[margin - 1]) > -65); 142| 12.3k| } 143| | // If the input is long enough, then we have that margin-1 is the fourth 144| | // last leading byte. 145| 5.26k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 146| 7.03k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (146:12): [True: 1.76k, False: 5.26k] ------------------ 147| 1.76k| simd8x64 input(reinterpret_cast(in + pos)); 148| 1.76k| if (input.is_ascii()) { ------------------ | Branch (148:11): [True: 569, False: 1.20k] ------------------ 149| 569| input.store_ascii_as_utf32(utf32_output); 150| 569| utf32_output += 64; 151| 569| pos += 64; 152| 1.20k| } else { 153| | // you might think that a for-loop would work, but under Visual Studio, 154| | // it is not good enough. 155| 1.20k| static_assert( 156| 1.20k| (simd8x64::NUM_CHUNKS == 2) || 157| 1.20k| (simd8x64::NUM_CHUNKS == 4), 158| 1.20k| "We support either two or four chunks per 64-byte block."); 159| 1.20k| auto zero = simd8{uint8_t(0)}; 160| 1.20k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 161| 1.20k| this->check_utf8_bytes(input.chunks[0], zero); 162| 1.20k| 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| 1.20k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 170| 1.20k| if (utf8_continuation_mask & 1) { ------------------ | Branch (170:13): [True: 7, False: 1.19k] ------------------ 171| 7| return 0; // we have an error 172| 7| } 173| 1.19k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 174| 1.19k| 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| 1.19k| size_t max_starting_point = (pos + 64) - 12; 179| | // Next loop is going to run at least five times. 180| 11.8k| while (pos < max_starting_point) { ------------------ | Branch (180:16): [True: 10.6k, False: 1.19k] ------------------ 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| 10.6k| size_t consumed = convert_masked_utf8_to_utf32( 191| 10.6k| in + pos, utf8_end_of_code_point_mask, utf32_output); 192| 10.6k| pos += consumed; 193| 10.6k| utf8_end_of_code_point_mask >>= consumed; 194| 10.6k| } 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| 1.19k| } 200| 1.76k| } 201| 5.26k| if (errors()) { ------------------ | Branch (201:9): [True: 193, False: 5.06k] ------------------ 202| 193| return 0; 203| 193| } 204| 5.06k| if (pos < size) { ------------------ | Branch (204:9): [True: 1.18k, False: 3.88k] ------------------ 205| 1.18k| size_t howmany = 206| 1.18k| scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output); 207| 1.18k| if (howmany == 0) { ------------------ | Branch (207:11): [True: 400, False: 788] ------------------ 208| 400| return 0; 209| 400| } 210| 788| utf32_output += howmany; 211| 788| } 212| 4.66k| return utf32_output - start; 213| 5.06k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 2.40k| 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| 2.40k| simd8 prev1 = input.prev<1>(prev_input); 124| 2.40k| simd8 sc = check_special_cases(input, prev1); 125| 2.40k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 2.40k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3219check_special_casesENS1_4simd5simd8IhEES5_: 8| 2.40k|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| 2.40k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 2.40k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 2.40k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 2.40k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 2.40k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 2.40k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 2.40k| 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| 2.40k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 2.40k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 2.40k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 2.40k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 2.40k| TOO_LONG, 40| | // 10______ ________ 41| 2.40k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 2.40k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 2.40k| TOO_SHORT, 46| | // 1110____ ________ 47| 2.40k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 2.40k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 2.40k| constexpr const uint8_t CARRY = 51| 2.40k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 2.40k| const simd8 byte_1_low = 53| 2.40k| (prev1 & 0x0F) 54| 2.40k| .lookup_16( 55| | // ____0000 ________ 56| 2.40k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 2.40k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 2.40k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 2.40k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 2.40k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 2.40k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 2.40k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 2.40k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 2.40k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 2.40k| OVERLONG_4, 88| | // ________ 1001____ 89| 2.40k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 2.40k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 2.40k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 2.40k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 2.40k| return (byte_1_high & byte_1_low & byte_2_high); 97| 2.40k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3223check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 2.40k| const simd8 sc) { 102| 2.40k| simd8 prev2 = input.prev<2>(prev_input); 103| 2.40k| simd8 prev3 = input.prev<3>(prev_input); 104| 2.40k| simd8 must23 = 105| 2.40k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 2.40k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 2.40k| return must23_80 ^ sc; 108| 2.40k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_113utf8_to_utf3221validating_transcoder6errorsEv: 310| 5.26k| simdutf_really_inline bool errors() const { 311| 5.26k| return this->error.any_bits_set_anywhere(); 312| 5.26k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoderC2Ev: 114| 5.26k| validating_transcoder() : error(uint8_t(0)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoder7convertEPKcmPDi: 129| 5.26k| char32_t *utf32_output) { 130| 5.26k| size_t pos = 0; 131| 5.26k| 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| 5.26k| size_t leading_byte = 0; 139| 5.26k| size_t margin = size; 140| 14.2k| for (; margin > 0 && leading_byte < 8; margin--) { ------------------ | Branch (140:12): [True: 9.81k, False: 4.48k] | Branch (140:26): [True: 9.03k, False: 779] ------------------ 141| 9.03k| leading_byte += (int8_t(in[margin - 1]) > -65); 142| 9.03k| } 143| | // If the input is long enough, then we have that margin-1 is the fourth 144| | // last leading byte. 145| 5.26k| const size_t safety_margin = size - margin + 1; // to avoid overruns! 146| 7.83k| while (pos + 64 + safety_margin <= size) { ------------------ | Branch (146:12): [True: 2.58k, False: 5.25k] ------------------ 147| 2.58k| simd8x64 input(reinterpret_cast(in + pos)); 148| 2.58k| if (input.is_ascii()) { ------------------ | Branch (148:11): [True: 886, False: 1.69k] ------------------ 149| 886| input.store_ascii_as_utf32(utf32_output); 150| 886| utf32_output += 64; 151| 886| pos += 64; 152| 1.69k| } else { 153| | // you might think that a for-loop would work, but under Visual Studio, 154| | // it is not good enough. 155| 1.69k| static_assert( 156| 1.69k| (simd8x64::NUM_CHUNKS == 2) || 157| 1.69k| (simd8x64::NUM_CHUNKS == 4), 158| 1.69k| "We support either two or four chunks per 64-byte block."); 159| 1.69k| 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| 1.69k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 164| 1.69k| this->check_utf8_bytes(input.chunks[0], zero); 165| 1.69k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 166| 1.69k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 167| 1.69k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 168| 1.69k| } 169| 1.69k| uint64_t utf8_continuation_mask = input.lt(-65 + 1); 170| 1.69k| if (utf8_continuation_mask & 1) { ------------------ | Branch (170:13): [True: 13, False: 1.68k] ------------------ 171| 13| return 0; // we have an error 172| 13| } 173| 1.68k| uint64_t utf8_leading_mask = ~utf8_continuation_mask; 174| 1.68k| 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| 1.68k| size_t max_starting_point = (pos + 64) - 12; 179| | // Next loop is going to run at least five times. 180| 16.0k| while (pos < max_starting_point) { ------------------ | Branch (180:16): [True: 14.3k, False: 1.68k] ------------------ 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| 14.3k| size_t consumed = convert_masked_utf8_to_utf32( 191| 14.3k| in + pos, utf8_end_of_code_point_mask, utf32_output); 192| 14.3k| pos += consumed; 193| 14.3k| utf8_end_of_code_point_mask >>= consumed; 194| 14.3k| } 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| 1.68k| } 200| 2.58k| } 201| 5.25k| if (errors()) { ------------------ | Branch (201:9): [True: 198, False: 5.05k] ------------------ 202| 198| return 0; 203| 198| } 204| 5.05k| if (pos < size) { ------------------ | Branch (204:9): [True: 896, False: 4.16k] ------------------ 205| 896| size_t howmany = 206| 896| scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output); 207| 896| if (howmany == 0) { ------------------ | Branch (207:11): [True: 195, False: 701] ------------------ 208| 195| return 0; 209| 195| } 210| 701| utf32_output += howmany; 211| 701| } 212| 4.86k| return utf32_output - start; 213| 5.05k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoder16check_utf8_bytesENS1_4simd5simd8IhEES6_: 119| 6.77k| 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| 6.77k| simd8 prev1 = input.prev<1>(prev_input); 124| 6.77k| simd8 sc = check_special_cases(input, prev1); 125| 6.77k| this->error |= check_multibyte_lengths(input, prev_input, sc); 126| 6.77k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3219check_special_casesENS1_4simd5simd8IhEES5_: 8| 6.77k|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| 6.77k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 16| | // 11______ 11______ 17| 6.77k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 18| 6.77k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 19| 6.77k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 20| 6.77k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 21| 6.77k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 22| 6.77k| 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| 6.77k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 31| | // 11110101 1000____ 32| | // 1111011_ 1000____ 33| | // 11111___ 1000____ 34| 6.77k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 35| | 36| 6.77k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 37| | // 0_______ ________ 38| 6.77k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 39| 6.77k| TOO_LONG, 40| | // 10______ ________ 41| 6.77k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 42| | // 1100____ ________ 43| 6.77k| TOO_SHORT | OVERLONG_2, 44| | // 1101____ ________ 45| 6.77k| TOO_SHORT, 46| | // 1110____ ________ 47| 6.77k| TOO_SHORT | OVERLONG_3 | SURROGATE, 48| | // 1111____ ________ 49| 6.77k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 50| 6.77k| constexpr const uint8_t CARRY = 51| 6.77k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 52| 6.77k| const simd8 byte_1_low = 53| 6.77k| (prev1 & 0x0F) 54| 6.77k| .lookup_16( 55| | // ____0000 ________ 56| 6.77k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 57| | // ____0001 ________ 58| 6.77k| CARRY | OVERLONG_2, 59| | // ____001_ ________ 60| 6.77k| CARRY, CARRY, 61| | 62| | // ____0100 ________ 63| 6.77k| CARRY | TOO_LARGE, 64| | // ____0101 ________ 65| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 66| | // ____011_ ________ 67| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 68| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| | 70| | // ____1___ ________ 71| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 72| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| | // ____1101 ________ 77| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 78| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000, 79| 6.77k| CARRY | TOO_LARGE | TOO_LARGE_1000); 80| 6.77k| const simd8 byte_2_high = input.shr<4>().lookup_16( 81| | // ________ 0_______ 82| 6.77k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 83| 6.77k| TOO_SHORT, TOO_SHORT, 84| | 85| | // ________ 1000____ 86| 6.77k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 87| 6.77k| OVERLONG_4, 88| | // ________ 1001____ 89| 6.77k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 90| | // ________ 101_____ 91| 6.77k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 92| 6.77k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| | 94| | // ________ 11______ 95| 6.77k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 96| 6.77k| return (byte_1_high & byte_1_low & byte_2_high); 97| 6.77k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3223check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 101| 6.77k| const simd8 sc) { 102| 6.77k| simd8 prev2 = input.prev<2>(prev_input); 103| 6.77k| simd8 prev3 = input.prev<3>(prev_input); 104| 6.77k| simd8 must23 = 105| 6.77k| simd8(must_be_2_3_continuation(prev2, prev3)); 106| 6.77k| simd8 must23_80 = must23 & uint8_t(0x80); 107| 6.77k| return must23_80 ^ sc; 108| 6.77k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_113utf8_to_utf3221validating_transcoder6errorsEv: 310| 5.25k| simdutf_really_inline bool errors() const { 311| 5.25k| return this->error.any_bits_set_anywhere(); 312| 5.25k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker16check_next_inputERKNS1_4simd8simd8x64IhEE: 187| 33.0k| simdutf_really_inline void check_next_input(const simd8x64 &input) { 188| 33.0k| if (simdutf_likely(is_ascii(input))) { ------------------ | | 90| 33.0k| #define simdutf_likely(x) __builtin_expect(!!(x), 1) | | ------------------ | | | Branch (90:31): [True: 25.4k, False: 7.60k] | | ------------------ ------------------ 189| 25.4k| this->error |= this->prev_incomplete; 190| 25.4k| } else { 191| | // you might think that a for-loop would work, but under Visual Studio, it 192| | // is not good enough. 193| 7.60k| static_assert((simd8x64::NUM_CHUNKS == 2) || 194| 7.60k| (simd8x64::NUM_CHUNKS == 4), 195| 7.60k| "We support either two or four chunks per 64-byte block."); 196| 7.60k| if constexpr (simd8x64::NUM_CHUNKS == 2) { 197| 7.60k| this->check_utf8_bytes(input.chunks[0], this->prev_input_block); 198| 7.60k| 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| 7.60k| this->prev_incomplete = 206| 7.60k| is_incomplete(input.chunks[simd8x64::NUM_CHUNKS - 1]); 207| 7.60k| this->prev_input_block = input.chunks[simd8x64::NUM_CHUNKS - 1]; 208| 7.60k| } 209| 33.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker16check_utf8_bytesENS1_4simd5simd8IhEES6_: 169| 15.2k| 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| 15.2k| simd8 prev1 = input.prev<1>(prev_input); 174| 15.2k| simd8 sc = check_special_cases(input, prev1); 175| 15.2k| this->error |= check_multibyte_lengths(input, prev_input, sc); 176| 15.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation19check_special_casesENS1_4simd5simd8IhEES5_: 9| 15.2k|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| 15.2k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 17| | // 11______ 11______ 18| 15.2k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 19| 15.2k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 20| 15.2k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 21| 15.2k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 22| 15.2k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 23| 15.2k| 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| 15.2k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 32| | // 11110101 1000____ 33| | // 1111011_ 1000____ 34| | // 11111___ 1000____ 35| 15.2k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 36| | 37| 15.2k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 38| | // 0_______ ________ 39| 15.2k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 40| 15.2k| TOO_LONG, 41| | // 10______ ________ 42| 15.2k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 43| | // 1100____ ________ 44| 15.2k| TOO_SHORT | OVERLONG_2, 45| | // 1101____ ________ 46| 15.2k| TOO_SHORT, 47| | // 1110____ ________ 48| 15.2k| TOO_SHORT | OVERLONG_3 | SURROGATE, 49| | // 1111____ ________ 50| 15.2k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 51| 15.2k| constexpr const uint8_t CARRY = 52| 15.2k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 53| 15.2k| const simd8 byte_1_low = 54| 15.2k| (prev1 & 0x0F) 55| 15.2k| .lookup_16( 56| | // ____0000 ________ 57| 15.2k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 58| | // ____0001 ________ 59| 15.2k| CARRY | OVERLONG_2, 60| | // ____001_ ________ 61| 15.2k| CARRY, CARRY, 62| | 63| | // ____0100 ________ 64| 15.2k| CARRY | TOO_LARGE, 65| | // ____0101 ________ 66| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 67| | // ____011_ ________ 68| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 70| | 71| | // ____1___ ________ 72| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 77| | // ____1101 ________ 78| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 79| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000, 80| 15.2k| CARRY | TOO_LARGE | TOO_LARGE_1000); 81| 15.2k| const simd8 byte_2_high = input.shr<4>().lookup_16( 82| | // ________ 0_______ 83| 15.2k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 84| 15.2k| TOO_SHORT, TOO_SHORT, 85| | 86| | // ________ 1000____ 87| 15.2k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 88| 15.2k| OVERLONG_4, 89| | // ________ 1001____ 90| 15.2k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 91| | // ________ 101_____ 92| 15.2k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| 15.2k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 94| | 95| | // ________ 11______ 96| 15.2k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 97| 15.2k| return (byte_1_high & byte_1_low & byte_2_high); 98| 15.2k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation23check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 102| 15.2k| const simd8 sc) { 103| 15.2k| simd8 prev2 = input.prev<2>(prev_input); 104| 15.2k| simd8 prev3 = input.prev<3>(prev_input); 105| 15.2k| simd8 must23 = 106| 15.2k| simd8(must_be_2_3_continuation(prev2, prev3)); 107| 15.2k| simd8 must23_80 = must23 & uint8_t(0x80); 108| 15.2k| return must23_80 ^ sc; 109| 15.2k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation13is_incompleteENS1_4simd5simd8IhEE: 115| 7.60k|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| 7.60k| static const uint8_t max_array[32] = {255, 120| 7.60k| 255, 121| 7.60k| 255, 122| 7.60k| 255, 123| 7.60k| 255, 124| 7.60k| 255, 125| 7.60k| 255, 126| 7.60k| 255, 127| 7.60k| 255, 128| 7.60k| 255, 129| 7.60k| 255, 130| 7.60k| 255, 131| 7.60k| 255, 132| 7.60k| 255, 133| 7.60k| 255, 134| 7.60k| 255, 135| 7.60k| 255, 136| 7.60k| 255, 137| 7.60k| 255, 138| 7.60k| 255, 139| 7.60k| 255, 140| 7.60k| 255, 141| 7.60k| 255, 142| 7.60k| 255, 143| 7.60k| 255, 144| 7.60k| 255, 145| 7.60k| 255, 146| 7.60k| 255, 147| 7.60k| 255, 148| 7.60k| 0b11110000u - 1, 149| 7.60k| 0b11100000u - 1, 150| 7.60k| 0b11000000u - 1}; 151| 7.60k| const simd8 max_value( 152| 7.60k| &max_array[sizeof(max_array) - sizeof(simd8)]); 153| 7.60k| return input.gt_bits(max_value); 154| 7.60k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker9check_eofEv: 181| 25.9k| 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| 25.9k| this->error |= this->prev_incomplete; 185| 25.9k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_115utf8_validation12utf8_checker6errorsEv: 212| 27.0k| simdutf_really_inline bool errors() const { 213| 27.0k| return this->error.any_bits_set_anywhere(); 214| 27.0k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker16check_next_inputERKNS1_4simd8simd8x64IhEE: 187| 35.4k| simdutf_really_inline void check_next_input(const simd8x64 &input) { 188| 35.4k| if (simdutf_likely(is_ascii(input))) { ------------------ | | 90| 35.4k| #define simdutf_likely(x) __builtin_expect(!!(x), 1) | | ------------------ | | | Branch (90:31): [True: 26.8k, False: 8.59k] | | ------------------ ------------------ 189| 26.8k| this->error |= this->prev_incomplete; 190| 26.8k| } else { 191| | // you might think that a for-loop would work, but under Visual Studio, it 192| | // is not good enough. 193| 8.59k| static_assert((simd8x64::NUM_CHUNKS == 2) || 194| 8.59k| (simd8x64::NUM_CHUNKS == 4), 195| 8.59k| "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| 8.59k| } else if constexpr (simd8x64::NUM_CHUNKS == 4) { 200| 8.59k| this->check_utf8_bytes(input.chunks[0], this->prev_input_block); 201| 8.59k| this->check_utf8_bytes(input.chunks[1], input.chunks[0]); 202| 8.59k| this->check_utf8_bytes(input.chunks[2], input.chunks[1]); 203| 8.59k| this->check_utf8_bytes(input.chunks[3], input.chunks[2]); 204| 8.59k| } 205| 8.59k| this->prev_incomplete = 206| 8.59k| is_incomplete(input.chunks[simd8x64::NUM_CHUNKS - 1]); 207| 8.59k| this->prev_input_block = input.chunks[simd8x64::NUM_CHUNKS - 1]; 208| 8.59k| } 209| 35.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker16check_utf8_bytesENS1_4simd5simd8IhEES6_: 169| 34.3k| 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| 34.3k| simd8 prev1 = input.prev<1>(prev_input); 174| 34.3k| simd8 sc = check_special_cases(input, prev1); 175| 34.3k| this->error |= check_multibyte_lengths(input, prev_input, sc); 176| 34.3k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation19check_special_casesENS1_4simd5simd8IhEES5_: 9| 34.3k|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| 34.3k| constexpr const uint8_t TOO_SHORT = 1 << 0; // 11______ 0_______ 17| | // 11______ 11______ 18| 34.3k| constexpr const uint8_t TOO_LONG = 1 << 1; // 0_______ 10______ 19| 34.3k| constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____ 20| 34.3k| constexpr const uint8_t SURROGATE = 1 << 4; // 11101101 101_____ 21| 34.3k| constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______ 22| 34.3k| constexpr const uint8_t TWO_CONTS = 1 << 7; // 10______ 10______ 23| 34.3k| 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| 34.3k| constexpr const uint8_t TOO_LARGE_1000 = 1 << 6; 32| | // 11110101 1000____ 33| | // 1111011_ 1000____ 34| | // 11111___ 1000____ 35| 34.3k| constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____ 36| | 37| 34.3k| const simd8 byte_1_high = prev1.shr<4>().lookup_16( 38| | // 0_______ ________ 39| 34.3k| TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, 40| 34.3k| TOO_LONG, 41| | // 10______ ________ 42| 34.3k| TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS, 43| | // 1100____ ________ 44| 34.3k| TOO_SHORT | OVERLONG_2, 45| | // 1101____ ________ 46| 34.3k| TOO_SHORT, 47| | // 1110____ ________ 48| 34.3k| TOO_SHORT | OVERLONG_3 | SURROGATE, 49| | // 1111____ ________ 50| 34.3k| TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4); 51| 34.3k| constexpr const uint8_t CARRY = 52| 34.3k| TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 . 53| 34.3k| const simd8 byte_1_low = 54| 34.3k| (prev1 & 0x0F) 55| 34.3k| .lookup_16( 56| | // ____0000 ________ 57| 34.3k| CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4, 58| | // ____0001 ________ 59| 34.3k| CARRY | OVERLONG_2, 60| | // ____001_ ________ 61| 34.3k| CARRY, CARRY, 62| | 63| | // ____0100 ________ 64| 34.3k| CARRY | TOO_LARGE, 65| | // ____0101 ________ 66| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 67| | // ____011_ ________ 68| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 69| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 70| | 71| | // ____1___ ________ 72| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 73| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 74| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 75| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 76| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 77| | // ____1101 ________ 78| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE, 79| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000, 80| 34.3k| CARRY | TOO_LARGE | TOO_LARGE_1000); 81| 34.3k| const simd8 byte_2_high = input.shr<4>().lookup_16( 82| | // ________ 0_______ 83| 34.3k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, 84| 34.3k| TOO_SHORT, TOO_SHORT, 85| | 86| | // ________ 1000____ 87| 34.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | 88| 34.3k| OVERLONG_4, 89| | // ________ 1001____ 90| 34.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE, 91| | // ________ 101_____ 92| 34.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 93| 34.3k| TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE, 94| | 95| | // ________ 11______ 96| 34.3k| TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT); 97| 34.3k| return (byte_1_high & byte_1_low & byte_2_high); 98| 34.3k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation23check_multibyte_lengthsENS1_4simd5simd8IhEES5_S5_: 102| 34.3k| const simd8 sc) { 103| 34.3k| simd8 prev2 = input.prev<2>(prev_input); 104| 34.3k| simd8 prev3 = input.prev<3>(prev_input); 105| 34.3k| simd8 must23 = 106| 34.3k| simd8(must_be_2_3_continuation(prev2, prev3)); 107| 34.3k| simd8 must23_80 = must23 & uint8_t(0x80); 108| 34.3k| return must23_80 ^ sc; 109| 34.3k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation13is_incompleteENS1_4simd5simd8IhEE: 115| 8.59k|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| 8.59k| static const uint8_t max_array[32] = {255, 120| 8.59k| 255, 121| 8.59k| 255, 122| 8.59k| 255, 123| 8.59k| 255, 124| 8.59k| 255, 125| 8.59k| 255, 126| 8.59k| 255, 127| 8.59k| 255, 128| 8.59k| 255, 129| 8.59k| 255, 130| 8.59k| 255, 131| 8.59k| 255, 132| 8.59k| 255, 133| 8.59k| 255, 134| 8.59k| 255, 135| 8.59k| 255, 136| 8.59k| 255, 137| 8.59k| 255, 138| 8.59k| 255, 139| 8.59k| 255, 140| 8.59k| 255, 141| 8.59k| 255, 142| 8.59k| 255, 143| 8.59k| 255, 144| 8.59k| 255, 145| 8.59k| 255, 146| 8.59k| 255, 147| 8.59k| 255, 148| 8.59k| 0b11110000u - 1, 149| 8.59k| 0b11100000u - 1, 150| 8.59k| 0b11000000u - 1}; 151| 8.59k| const simd8 max_value( 152| 8.59k| &max_array[sizeof(max_array) - sizeof(simd8)]); 153| 8.59k| return input.gt_bits(max_value); 154| 8.59k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker9check_eofEv: 181| 25.9k| 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| 25.9k| this->error |= this->prev_incomplete; 185| 25.9k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_115utf8_validation12utf8_checker6errorsEv: 212| 27.5k| simdutf_really_inline bool errors() const { 213| 27.5k| return this->error.any_bits_set_anywhere(); 214| 27.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation21generic_validate_utf8EPKcm: 27| 20.8k|bool generic_validate_utf8(const char *input, size_t length) { 28| 20.8k| return generic_validate_utf8( 29| 20.8k| reinterpret_cast(input), length); 30| 20.8k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation21generic_validate_utf8INS2_12utf8_checkerEEEbPKhm: 10| 20.8k|bool generic_validate_utf8(const uint8_t *input, size_t length) { 11| 20.8k| checker c{}; 12| 20.8k| buf_block_reader<64> reader(input, length); 13| 26.9k| while (reader.has_full_block()) { ------------------ | Branch (13:10): [True: 6.02k, False: 20.8k] ------------------ 14| 6.02k| simd::simd8x64 in(reader.full_block()); 15| 6.02k| c.check_next_input(in); 16| 6.02k| reader.advance(); 17| 6.02k| } 18| 20.8k| uint8_t block[64]{}; 19| 20.8k| reader.get_remainder(block); 20| 20.8k| simd::simd8x64 in(block); 21| 20.8k| c.check_next_input(in); 22| 20.8k| reader.advance(); 23| 20.8k| c.check_eof(); 24| 20.8k| return !c.errors(); 25| 20.8k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsEPKcm: 76| 5.26k|result generic_validate_utf8_with_errors(const char *input, size_t length) { 77| 5.26k| return generic_validate_utf8_with_errors( 78| 5.26k| reinterpret_cast(input), length); 79| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsINS2_12utf8_checkerEEENS_6resultEPKhm: 36| 5.26k|result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) { 37| 5.26k| checker c{}; 38| 5.26k| buf_block_reader<64> reader(input, length); 39| 5.26k| size_t count{0}; 40| 6.18k| while (reader.has_full_block()) { ------------------ | Branch (40:10): [True: 1.17k, False: 5.01k] ------------------ 41| 1.17k| simd::simd8x64 in(reader.full_block()); 42| 1.17k| c.check_next_input(in); 43| 1.17k| if (c.errors()) { ------------------ | Branch (43:9): [True: 252, False: 920] ------------------ 44| 252| if (count != 0) { ------------------ | Branch (44:11): [True: 14, False: 238] ------------------ 45| 14| count--; 46| 14| } // Sometimes the error is only detected in the next chunk 47| 252| result res = scalar::utf8::rewind_and_validate_with_errors( 48| 252| reinterpret_cast(input), 49| 252| reinterpret_cast(input + count), length - count); 50| 252| res.count += count; 51| 252| return res; 52| 252| } 53| 920| reader.advance(); 54| 920| count += 64; 55| 920| } 56| 5.01k| uint8_t block[64]{}; 57| 5.01k| reader.get_remainder(block); 58| 5.01k| simd::simd8x64 in(block); 59| 5.01k| c.check_next_input(in); 60| 5.01k| reader.advance(); 61| 5.01k| c.check_eof(); 62| 5.01k| if (c.errors()) { ------------------ | Branch (62:7): [True: 348, False: 4.66k] ------------------ 63| 348| if (count != 0) { ------------------ | Branch (63:9): [True: 15, False: 333] ------------------ 64| 15| count--; 65| 15| } // Sometimes the error is only detected in the next chunk 66| 348| result res = scalar::utf8::rewind_and_validate_with_errors( 67| 348| reinterpret_cast(input), 68| 348| reinterpret_cast(input) + count, length - count); 69| 348| res.count += count; 70| 348| return res; 71| 4.66k| } else { 72| 4.66k| return result(error_code::SUCCESS, length); 73| 4.66k| } 74| 5.01k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation21generic_validate_utf8EPKcm: 27| 20.8k|bool generic_validate_utf8(const char *input, size_t length) { 28| 20.8k| return generic_validate_utf8( 29| 20.8k| reinterpret_cast(input), length); 30| 20.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation21generic_validate_utf8INS2_12utf8_checkerEEEbPKhm: 10| 20.8k|bool generic_validate_utf8(const uint8_t *input, size_t length) { 11| 20.8k| checker c{}; 12| 20.8k| buf_block_reader<64> reader(input, length); 13| 28.7k| while (reader.has_full_block()) { ------------------ | Branch (13:10): [True: 7.86k, False: 20.8k] ------------------ 14| 7.86k| simd::simd8x64 in(reader.full_block()); 15| 7.86k| c.check_next_input(in); 16| 7.86k| reader.advance(); 17| 7.86k| } 18| 20.8k| uint8_t block[64]{}; 19| 20.8k| reader.get_remainder(block); 20| 20.8k| simd::simd8x64 in(block); 21| 20.8k| c.check_next_input(in); 22| 20.8k| reader.advance(); 23| 20.8k| c.check_eof(); 24| 20.8k| return !c.errors(); 25| 20.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsEPKcm: 76| 5.26k|result generic_validate_utf8_with_errors(const char *input, size_t length) { 77| 5.26k| return generic_validate_utf8_with_errors( 78| 5.26k| reinterpret_cast(input), length); 79| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115utf8_validation33generic_validate_utf8_with_errorsINS2_12utf8_checkerEEENS_6resultEPKhm: 36| 5.26k|result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) { 37| 5.26k| checker c{}; 38| 5.26k| buf_block_reader<64> reader(input, length); 39| 5.26k| size_t count{0}; 40| 6.68k| while (reader.has_full_block()) { ------------------ | Branch (40:10): [True: 1.64k, False: 5.04k] ------------------ 41| 1.64k| simd::simd8x64 in(reader.full_block()); 42| 1.64k| c.check_next_input(in); 43| 1.64k| if (c.errors()) { ------------------ | Branch (43:9): [True: 226, False: 1.41k] ------------------ 44| 226| if (count != 0) { ------------------ | Branch (44:11): [True: 34, False: 192] ------------------ 45| 34| count--; 46| 34| } // Sometimes the error is only detected in the next chunk 47| 226| result res = scalar::utf8::rewind_and_validate_with_errors( 48| 226| reinterpret_cast(input), 49| 226| reinterpret_cast(input + count), length - count); 50| 226| res.count += count; 51| 226| return res; 52| 226| } 53| 1.41k| reader.advance(); 54| 1.41k| count += 64; 55| 1.41k| } 56| 5.04k| uint8_t block[64]{}; 57| 5.04k| reader.get_remainder(block); 58| 5.04k| simd::simd8x64 in(block); 59| 5.04k| c.check_next_input(in); 60| 5.04k| reader.advance(); 61| 5.04k| c.check_eof(); 62| 5.04k| if (c.errors()) { ------------------ | Branch (62:7): [True: 180, False: 4.86k] ------------------ 63| 180| if (count != 0) { ------------------ | Branch (63:9): [True: 13, False: 167] ------------------ 64| 13| count--; 65| 13| } // Sometimes the error is only detected in the next chunk 66| 180| result res = scalar::utf8::rewind_and_validate_with_errors( 67| 180| reinterpret_cast(input), 68| 180| reinterpret_cast(input) + count, length - count); 69| 180| res.count += count; 70| 180| return res; 71| 4.86k| } else { 72| 4.86k| return result(error_code::SUCCESS, length); 73| 4.86k| } 74| 5.04k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE0EEEKNS_6resultEPKDsm: 49| 7.50k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.50k| if (simdutf_unlikely(size == 0)) { ------------------ | | 93| 7.50k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 4.91k, False: 2.58k] | | ------------------ ------------------ 51| 4.91k| return result(error_code::SUCCESS, 0); 52| 4.91k| } 53| | 54| 2.58k| const char16_t *start = input; 55| 2.58k| const char16_t *end = input + size; 56| | 57| 2.58k| const auto v_d8 = simd8::splat(0xd8); 58| 2.58k| const auto v_f8 = simd8::splat(0xf8); 59| 2.58k| const auto v_fc = simd8::splat(0xfc); 60| 2.58k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 14.3k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 11.8k, False: 2.49k] ------------------ 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| 11.8k| auto in0 = simd16(input); 67| 11.8k| auto in1 = 68| 11.8k| 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| 11.8k| 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| 11.8k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 11.8k| const uint16_t surrogates_bitmask = 77| 11.8k| static_cast(surrogates_wordmask.to_bitmask()); 78| 11.8k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 10.6k, False: 1.17k] ------------------ 79| 10.6k| input += 16; 80| 10.6k| } 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| 1.17k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.17k| const auto vH = (in & v_fc) == v_dc; 93| 1.17k| 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| 1.17k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.17k| const uint16_t a = static_cast( 100| 1.17k| 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| 1.17k| const uint16_t b = static_cast( 104| 1.17k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.17k| const uint16_t c = static_cast( 107| 1.17k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.17k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 782, False: 392] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 782| input += 16; 113| 782| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 298, False: 94] ------------------ 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| 298| input += 15; 119| 298| } else { 120| 94| return result(error_code::SURROGATE, input - start); 121| 94| } 122| 1.17k| } 123| 11.8k| } 124| | 125| 2.49k| return result(error_code::SUCCESS, input - start); 126| 2.58k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE1EEEKNS_6resultEPKDsm: 49| 7.51k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.51k| if (simdutf_unlikely(size == 0)) { ------------------ | | 93| 7.51k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 4.91k, False: 2.59k] | | ------------------ ------------------ 51| 4.91k| return result(error_code::SUCCESS, 0); 52| 4.91k| } 53| | 54| 2.59k| const char16_t *start = input; 55| 2.59k| const char16_t *end = input + size; 56| | 57| 2.59k| const auto v_d8 = simd8::splat(0xd8); 58| 2.59k| const auto v_f8 = simd8::splat(0xf8); 59| 2.59k| const auto v_fc = simd8::splat(0xfc); 60| 2.59k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 14.6k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 12.0k, False: 2.51k] ------------------ 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| 12.0k| auto in0 = simd16(input); 67| 12.0k| auto in1 = 68| 12.0k| 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| 12.0k| 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| 12.0k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 12.0k| const uint16_t surrogates_bitmask = 77| 12.0k| static_cast(surrogates_wordmask.to_bitmask()); 78| 12.0k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 11.0k, False: 1.08k] ------------------ 79| 11.0k| input += 16; 80| 11.0k| } 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| 1.08k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.08k| const auto vH = (in & v_fc) == v_dc; 93| 1.08k| 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| 1.08k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.08k| const uint16_t a = static_cast( 100| 1.08k| 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| 1.08k| const uint16_t b = static_cast( 104| 1.08k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.08k| const uint16_t c = static_cast( 107| 1.08k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.08k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 666, False: 414] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 666| input += 16; 113| 666| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 328, False: 86] ------------------ 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| 328| input += 15; 119| 328| } else { 120| 86| return result(error_code::SURROGATE, input - start); 121| 86| } 122| 1.08k| } 123| 12.0k| } 124| | 125| 2.51k| return result(error_code::SUCCESS, input - start); 126| 2.59k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE0EEEKNS_6resultEPKDsm: 49| 7.06k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.06k| if (simdutf_unlikely(size == 0)) { ------------------ | | 93| 7.06k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 4.93k, False: 2.12k] | | ------------------ ------------------ 51| 4.93k| return result(error_code::SUCCESS, 0); 52| 4.93k| } 53| | 54| 2.12k| const char16_t *start = input; 55| 2.12k| const char16_t *end = input + size; 56| | 57| 2.12k| const auto v_d8 = simd8::splat(0xd8); 58| 2.12k| const auto v_f8 = simd8::splat(0xf8); 59| 2.12k| const auto v_fc = simd8::splat(0xfc); 60| 2.12k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 19.6k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 17.6k, False: 2.00k] ------------------ 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| 17.6k| auto in0 = simd16(input); 67| 17.6k| auto in1 = 68| 17.6k| 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| 17.6k| 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| 17.6k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 17.6k| const uint16_t surrogates_bitmask = 77| 17.6k| static_cast(surrogates_wordmask.to_bitmask()); 78| 17.6k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 16.4k, False: 1.27k] ------------------ 79| 16.4k| input += 16; 80| 16.4k| } 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| 1.27k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.27k| const auto vH = (in & v_fc) == v_dc; 93| 1.27k| 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| 1.27k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.27k| const uint16_t a = static_cast( 100| 1.27k| 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| 1.27k| const uint16_t b = static_cast( 104| 1.27k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.27k| const uint16_t c = static_cast( 107| 1.27k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.27k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 800, False: 472] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 800| input += 16; 113| 800| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 346, False: 126] ------------------ 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| 126| return result(error_code::SURROGATE, input - start); 121| 126| } 122| 1.27k| } 123| 17.6k| } 124| | 125| 2.00k| return result(error_code::SUCCESS, input - start); 126| 2.12k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf1626validate_utf16_with_errorsILNS_10endiannessE1EEEKNS_6resultEPKDsm: 49| 7.05k|const result validate_utf16_with_errors(const char16_t *input, size_t size) { 50| 7.05k| if (simdutf_unlikely(size == 0)) { ------------------ | | 93| 7.05k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 4.94k, False: 2.10k] | | ------------------ ------------------ 51| 4.94k| return result(error_code::SUCCESS, 0); 52| 4.94k| } 53| | 54| 2.10k| const char16_t *start = input; 55| 2.10k| const char16_t *end = input + size; 56| | 57| 2.10k| const auto v_d8 = simd8::splat(0xd8); 58| 2.10k| const auto v_f8 = simd8::splat(0xf8); 59| 2.10k| const auto v_fc = simd8::splat(0xfc); 60| 2.10k| const auto v_dc = simd8::splat(0xdc); 61| | 62| 19.8k| while (input + simd16::SIZE * 2 < end) { ------------------ | Branch (62:10): [True: 17.8k, False: 1.98k] ------------------ 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| 17.8k| auto in0 = simd16(input); 67| 17.8k| auto in1 = 68| 17.8k| 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| 17.8k| 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| 17.8k| const auto surrogates_wordmask = (in & v_f8) == v_d8; 76| 17.8k| const uint16_t surrogates_bitmask = 77| 17.8k| static_cast(surrogates_wordmask.to_bitmask()); 78| 17.8k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (78:9): [True: 16.6k, False: 1.20k] ------------------ 79| 16.6k| input += 16; 80| 16.6k| } 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| 1.20k| const uint16_t V = static_cast(~surrogates_bitmask); 90| | 91| | // H - word-mask for high surrogates: the six highest bits are 0b1101'11 92| 1.20k| const auto vH = (in & v_fc) == v_dc; 93| 1.20k| 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| 1.20k| const uint16_t L = static_cast(~H & surrogates_bitmask); 98| | 99| 1.20k| const uint16_t a = static_cast( 100| 1.20k| 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| 1.20k| const uint16_t b = static_cast( 104| 1.20k| a << 1); // Just mark that the opinput - startite fact is hold, 105| | // thanks to that we have only two masks for valid case. 106| 1.20k| const uint16_t c = static_cast( 107| 1.20k| V | a | b); // Combine all the masks into the final one. 108| | 109| 1.20k| if (c == 0xffff) { ------------------ | Branch (109:11): [True: 726, False: 480] ------------------ 110| | // The whole input register contains valid UTF-16, i.e., 111| | // either single code units or proper surrogate pairs. 112| 726| input += 16; 113| 726| } else if (c == 0x7fff) { ------------------ | Branch (113:18): [True: 356, False: 124] ------------------ 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| 124| return result(error_code::SURROGATE, input - start); 121| 124| } 122| 1.20k| } 123| 17.8k| } 124| | 125| 1.98k| return result(error_code::SUCCESS, input - start); 126| 2.10k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf328validateEPKDim: 6| 9.93k|simdutf_really_inline bool validate(const char32_t *input, size_t size) { 7| 9.93k| if (simdutf_unlikely(size == 0)) { ------------------ | | 93| 9.93k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 8.04k, False: 1.88k] | | ------------------ ------------------ 8| | // empty input is valid UTF-32. protect the implementation from 9| | // handling nullptr 10| 8.04k| return true; 11| 8.04k| } 12| | 13| 1.88k| const char32_t *end = input + size; 14| | 15| 1.88k| using vector_u32 = simd32; 16| | 17| 1.88k| const auto standardmax = vector_u32::splat(0x10ffff); 18| 1.88k| const auto offset = vector_u32::splat(0xffff2000); 19| 1.88k| const auto standardoffsetmax = vector_u32::splat(0xfffff7ff); 20| 1.88k| auto currentmax = vector_u32::zero(); 21| 1.88k| auto currentoffsetmax = vector_u32::zero(); 22| | 23| 1.88k| constexpr size_t N = vector_u32::ELEMENTS; 24| | 25| 27.9k| while (input + N < end) { ------------------ | Branch (25:10): [True: 26.0k, False: 1.88k] ------------------ 26| 26.0k| auto in = vector_u32(input); 27| 26.0k| if constexpr (!match_system(endianness::BIG)) { 28| 26.0k| in.swap_bytes(); 29| 26.0k| } 30| | 31| 26.0k| currentmax = max(currentmax, in); 32| 26.0k| currentoffsetmax = max(currentoffsetmax, in + offset); 33| 26.0k| input += N; 34| 26.0k| } 35| | 36| 1.88k| const auto too_large = currentmax > standardmax; 37| 1.88k| if (too_large.any()) { ------------------ | Branch (37:7): [True: 0, False: 1.88k] ------------------ 38| 0| return false; 39| 0| } 40| | 41| 1.88k| const auto surrogate = currentoffsetmax > standardoffsetmax; 42| 1.88k| if (surrogate.any()) { ------------------ | Branch (42:7): [True: 0, False: 1.88k] ------------------ 43| 0| return false; 44| 0| } 45| | 46| 1.88k| return scalar::utf32::validate(input, end - input); 47| 1.88k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_15utf328validateEPKDim: 6| 10.1k|simdutf_really_inline bool validate(const char32_t *input, size_t size) { 7| 10.1k| if (simdutf_unlikely(size == 0)) { ------------------ | | 93| 10.1k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 8.66k, False: 1.47k] | | ------------------ ------------------ 8| | // empty input is valid UTF-32. protect the implementation from 9| | // handling nullptr 10| 8.66k| return true; 11| 8.66k| } 12| | 13| 1.47k| const char32_t *end = input + size; 14| | 15| 1.47k| using vector_u32 = simd32; 16| | 17| 1.47k| const auto standardmax = vector_u32::splat(0x10ffff); 18| 1.47k| const auto offset = vector_u32::splat(0xffff2000); 19| 1.47k| const auto standardoffsetmax = vector_u32::splat(0xfffff7ff); 20| 1.47k| auto currentmax = vector_u32::zero(); 21| 1.47k| auto currentoffsetmax = vector_u32::zero(); 22| | 23| 1.47k| constexpr size_t N = vector_u32::ELEMENTS; 24| | 25| 62.1k| while (input + N < end) { ------------------ | Branch (25:10): [True: 60.6k, False: 1.47k] ------------------ 26| 60.6k| auto in = vector_u32(input); 27| 60.6k| if constexpr (!match_system(endianness::BIG)) { 28| 60.6k| in.swap_bytes(); 29| 60.6k| } 30| | 31| 60.6k| currentmax = max(currentmax, in); 32| 60.6k| currentoffsetmax = max(currentoffsetmax, in + offset); 33| 60.6k| input += N; 34| 60.6k| } 35| | 36| 1.47k| const auto too_large = currentmax > standardmax; 37| 1.47k| if (too_large.any()) { ------------------ | Branch (37:7): [True: 0, False: 1.47k] ------------------ 38| 0| return false; 39| 0| } 40| | 41| 1.47k| const auto surrogate = currentoffsetmax > standardoffsetmax; 42| 1.47k| if (surrogate.any()) { ------------------ | Branch (42:7): [True: 0, False: 1.47k] ------------------ 43| 0| return false; 44| 0| } 45| | 46| 1.47k| return scalar::utf32::validate(input, end - input); 47| 1.47k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block64C2EPKc: 450| 16.5k| simdutf_really_inline block64(const char *src) { 451| 16.5k| chunks[0] = _mm256_loadu_si256(reinterpret_cast(src)); 452| 16.5k| chunks[1] = _mm256_loadu_si256(reinterpret_cast(src + 32)); 453| 16.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6414compress_blockEmPc: 630| 2.23k| simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) { 631| 2.23k| if (is_power_of_two(mask)) { ------------------ | Branch (631:9): [True: 849, False: 1.38k] ------------------ 632| 849| return compress_block_single(mask, output); 633| 849| } 634| | 635| 1.38k| uint64_t nmask = ~mask; 636| 1.38k| compress(chunks[0], uint32_t(mask), output); 637| 1.38k| compress(chunks[1], uint32_t(mask >> 32), 638| 1.38k| output + count_ones(nmask & 0xFFFFFFFF)); 639| 1.38k| return count_ones(nmask); 640| 2.23k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6421compress_block_singleEmPc: 643| 849| char *output) { 644| 849| const size_t pos64 = trailing_zeroes(mask); 645| 849| const int8_t pos = pos64 & 0xf; 646| 849| switch (pos64 >> 4) { ------------------ | Branch (646:13): [True: 849, False: 0] ------------------ 647| 241| case 0b00: { ------------------ | Branch (647:5): [True: 241, False: 608] ------------------ 648| 241| const __m128i lane0 = _mm256_extracti128_si256(chunks[0], 0); 649| 241| const __m128i lane1 = _mm256_extracti128_si256(chunks[0], 1); 650| | 651| 241| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 652| 241| const __m128i v1 = 653| 241| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 654| 241| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 655| 241| const __m128i sh = _mm_sub_epi8(v1, v2); 656| 241| const __m128i compressed = _mm_shuffle_epi8(lane0, sh); 657| | 658| 241| _mm_storeu_si128((__m128i *)(output + 0 * 16), compressed); 659| 241| _mm_storeu_si128((__m128i *)(output + 1 * 16 - 1), lane1); 660| 241| _mm256_storeu_si256((__m256i *)(output + 2 * 16 - 1), chunks[1]); 661| 241| } break; 662| 207| case 0b01: { ------------------ | Branch (662:5): [True: 207, False: 642] ------------------ 663| 207| const __m128i lane0 = _mm256_extracti128_si256(chunks[0], 0); 664| 207| const __m128i lane1 = _mm256_extracti128_si256(chunks[0], 1); 665| 207| _mm_storeu_si128((__m128i *)(output + 0 * 16), lane0); 666| | 667| 207| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 668| 207| const __m128i v1 = 669| 207| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 670| 207| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 671| 207| const __m128i sh = _mm_sub_epi8(v1, v2); 672| 207| const __m128i compressed = _mm_shuffle_epi8(lane1, sh); 673| | 674| 207| _mm_storeu_si128((__m128i *)(output + 1 * 16), compressed); 675| 207| _mm256_storeu_si256((__m256i *)(output + 2 * 16 - 1), chunks[1]); 676| 207| } break; 677| 222| case 0b10: { ------------------ | Branch (677:5): [True: 222, False: 627] ------------------ 678| 222| const __m128i lane2 = _mm256_extracti128_si256(chunks[1], 0); 679| 222| const __m128i lane3 = _mm256_extracti128_si256(chunks[1], 1); 680| | 681| 222| _mm256_storeu_si256((__m256i *)(output + 0 * 16), chunks[0]); 682| | 683| 222| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 684| 222| const __m128i v1 = 685| 222| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 686| 222| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 687| 222| const __m128i sh = _mm_sub_epi8(v1, v2); 688| 222| const __m128i compressed = _mm_shuffle_epi8(lane2, sh); 689| | 690| 222| _mm_storeu_si128((__m128i *)(output + 2 * 16), compressed); 691| 222| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), lane3); 692| 222| } break; 693| 179| case 0b11: { ------------------ | Branch (693:5): [True: 179, False: 670] ------------------ 694| 179| const __m128i lane2 = _mm256_extracti128_si256(chunks[1], 0); 695| 179| const __m128i lane3 = _mm256_extracti128_si256(chunks[1], 1); 696| | 697| 179| _mm256_storeu_si256((__m256i *)(output + 0 * 16), chunks[0]); 698| 179| _mm_storeu_si128((__m128i *)(output + 2 * 16), lane2); 699| | 700| 179| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 701| 179| const __m128i v1 = 702| 179| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 703| 179| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 704| 179| const __m128i sh = _mm_sub_epi8(v1, v2); 705| 179| const __m128i compressed = _mm_shuffle_epi8(lane3, sh); 706| | 707| 179| _mm_storeu_si128((__m128i *)(output + 3 * 16), compressed); 708| 179| } break; 709| 849| } 710| | 711| 849| return 63; 712| 849| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_18compressIvEEvDv4_xjPc: 398| 2.77k|simdutf_really_inline void compress(__m256i data, uint32_t mask, char *output) { 399| 2.77k| if (mask == 0) { ------------------ | Branch (399:7): [True: 464, False: 2.31k] ------------------ 400| 464| _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), data); 401| 464| return; 402| 464| } 403| 2.31k| compress(_mm256_castsi256_si128(data), uint16_t(mask), output); 404| | compress(_mm256_extracti128_si256(data, 1), uint16_t(mask >> 16), 405| 2.31k| output + count_ones(~mask & 0xFFFF)); 406| 2.31k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_18compressEDv2_xtPc: 361| 4.62k|static inline void compress(__m128i data, uint16_t mask, char *output) { 362| 4.62k| if (mask == 0) { ------------------ | Branch (362:7): [True: 1.09k, False: 3.53k] ------------------ 363| 1.09k| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data); 364| 1.09k| return; 365| 1.09k| } 366| | // this particular implementation was inspired by work done by @animetosho 367| | // we do it in two steps, first 8 bytes and then second 8 bytes 368| 3.53k| uint8_t mask1 = uint8_t(mask); // least significant 8 bits 369| 3.53k| uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits 370| | // next line just loads the 64-bit values thintable_epi8[mask1] and 371| | // thintable_epi8[mask2] into a 128-bit register, using only 372| | // two instructions on most compilers. 373| | 374| 3.53k| __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2], 375| 3.53k| tables::base64::thintable_epi8[mask1]); 376| | // we increment by 0x08 the second half of the mask 377| 3.53k| shufmask = 378| 3.53k| _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0)); 379| | // this is the version "nearly pruned" 380| 3.53k| __m128i pruned = _mm_shuffle_epi8(data, shufmask); 381| | // we still need to put the two halves together. 382| | // we compute the popcount of the first half: 383| 3.53k| int pop1 = tables::base64::BitsSetTable256mul2[mask1]; 384| | // then load the corresponding mask, what it does is to write 385| | // only the first pop1 bytes from the first 8 bytes, and then 386| | // it fills in with the bytes from the second 8 bytes + some filling 387| | // at the end. 388| 3.53k| __m128i compactmask = _mm_loadu_si128(reinterpret_cast( 389| 3.53k| tables::base64::pshufb_combine_table + pop1 * 8)); 390| 3.53k| __m128i answer = _mm_shuffle_epi8(pruned, compactmask); 391| | 392| 3.53k| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer); 393| 3.53k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6410copy_blockEPc: 475| 395| simdutf_really_inline void copy_block(char *output) { 476| 395| _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), chunks[0]); 477| 395| _mm256_storeu_si256(reinterpret_cast<__m256i *>(output + 32), chunks[1]); 478| 395| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6424base64_decode_block_safeEPc: 486| 812| simdutf_really_inline void base64_decode_block_safe(char *out) { 487| 812| base64_decode(out, chunks[0]); 488| 812| alignas(32) char buffer[32]; // We enforce safety with a buffer. 489| 812| base64_decode(buffer, chunks[1]); 490| 812| std::memcpy(out + 24, buffer, 24); 491| 812| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113base64_decodeIvEEvPcDv4_x: 409| 29.7k|simdutf_really_inline void base64_decode(char *out, __m256i str) { 410| | // credit: aqrit 411| 29.7k| const __m256i pack_shuffle = 412| 29.7k| _mm256_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1, 413| 29.7k| 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1); 414| 29.7k| const __m256i t0 = _mm256_maddubs_epi16(str, _mm256_set1_epi32(0x01400140)); 415| 29.7k| const __m256i t1 = _mm256_madd_epi16(t0, _mm256_set1_epi32(0x00011000)); 416| 29.7k| const __m256i t2 = _mm256_shuffle_epi8(t1, pack_shuffle); 417| | 418| | // Store the output: 419| 29.7k| _mm_storeu_si128((__m128i *)out, _mm256_castsi256_si128(t2)); 420| | _mm_storeu_si128((__m128i *)(out + 12), _mm256_extracti128_si256(t2, 1)); 421| 29.7k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6419base64_decode_blockEPc: 481| 12.3k| simdutf_really_inline void base64_decode_block(char *out) { 482| 12.3k| base64_decode(out, chunks[0]); 483| 12.3k| base64_decode(out + 24, chunks[1]); 484| 12.3k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_119base64_decode_blockIvEEvPcPKc: 424| 1.68k|simdutf_really_inline void base64_decode_block(char *out, const char *src) { 425| 1.68k| base64_decode(out, 426| 1.68k| _mm256_loadu_si256(reinterpret_cast(src))); 427| 1.68k| base64_decode(out + 24, _mm256_loadu_si256( 428| 1.68k| reinterpret_cast(src + 32))); 429| 1.68k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_124base64_decode_block_safeIvEEvPcPKc: 433| 50| const char *src) { 434| 50| base64_decode(out, 435| 50| _mm256_loadu_si256(reinterpret_cast(src))); 436| 50| alignas(32) char buffer[32]; // We enforce safety with a buffer. 437| 50| base64_decode( 438| 50| buffer, _mm256_loadu_si256(reinterpret_cast(src + 32))); 439| 50| std::memcpy(out + 24, buffer, 24); 440| 50|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEmPm: 494| 16.5k| simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) { 495| 16.5k| uint32_t err0 = 0; 496| 16.5k| uint32_t err1 = 0; 497| 16.5k| uint64_t m0 = to_base64_mask( 498| 16.5k| &chunks[0], &err0); 499| 16.5k| uint64_t m1 = to_base64_mask( 500| 16.5k| &chunks[1], &err1); 501| 16.5k| if (!ignore_garbage) { ------------------ | Branch (501:9): [True: 16.5k, Folded] ------------------ 502| 16.5k| *error = err0 | ((uint64_t)err1 << 32); 503| 16.5k| } 504| 16.5k| return m0 | (m1 << 32); 505| 16.5k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEjPDv4_xPj: 508| 33.1k| simdutf_really_inline uint32_t to_base64_mask(__m256i *src, uint32_t *error) { 509| 33.1k| const __m256i ascii_space_tbl = 510| 33.1k| _mm256_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa, 511| 33.1k| 0x0, 0xc, 0xd, 0x0, 0x0, 0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 512| 33.1k| 0x0, 0x0, 0x0, 0x9, 0xa, 0x0, 0xc, 0xd, 0x0, 0x0); 513| | // credit: aqrit 514| 33.1k| __m256i delta_asso; 515| 33.1k| if (default_or_url) { ------------------ | Branch (515:9): [Folded, False: 33.1k] ------------------ 516| 0| delta_asso = _mm256_setr_epi8( 517| 0| 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 518| 0| 0x00, 0x00, 0x11, 0x00, 0x16, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 519| 0| 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x16); 520| 33.1k| } else if (base64_url) { ------------------ | Branch (520:16): [Folded, False: 33.1k] ------------------ 521| 0| delta_asso = _mm256_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 522| 0| 0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF, 0x1, 0x1, 523| 0| 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 0x0, 0x0, 524| 0| 0x0, 0x0, 0xF, 0x0, 0xF); 525| 33.1k| } else { 526| 33.1k| delta_asso = _mm256_setr_epi8( 527| 33.1k| 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 528| 33.1k| 0x00, 0x00, 0x0F, 0x00, 0x0F, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 529| 33.1k| 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F); 530| 33.1k| } 531| | 532| 33.1k| __m256i delta_values; 533| 33.1k| if (default_or_url) { ------------------ | Branch (533:9): [Folded, False: 33.1k] ------------------ 534| 0| delta_values = _mm256_setr_epi8( 535| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0x13), 536| 0| uint8_t(0x04), uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 537| 0| uint8_t(0xB9), uint8_t(0x00), uint8_t(0xFF), uint8_t(0x11), 538| 0| uint8_t(0xFF), uint8_t(0xBF), uint8_t(0x10), uint8_t(0xB9), 539| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0x13), 540| 0| uint8_t(0x04), uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 541| 0| uint8_t(0xB9), uint8_t(0x00), uint8_t(0xFF), uint8_t(0x11), 542| 0| uint8_t(0xFF), uint8_t(0xBF), uint8_t(0x10), uint8_t(0xB9)); 543| 33.1k| } else if (base64_url) { ------------------ | Branch (543:16): [Folded, False: 33.1k] ------------------ 544| 0| delta_values = _mm256_setr_epi8( 545| 0| 0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 546| 0| uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF), uint8_t(0xE0), 547| 0| uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), 548| 0| uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3), 549| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9)); 550| 33.1k| } else { 551| 33.1k| delta_values = _mm256_setr_epi8( 552| 33.1k| int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), int8_t(0x04), 553| 33.1k| int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9), int8_t(0x00), 554| 33.1k| int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), 555| 33.1k| int8_t(0xB9), int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), 556| 33.1k| int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9), 557| 33.1k| int8_t(0x00), int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF), 558| 33.1k| int8_t(0xB9), int8_t(0xB9)); 559| 33.1k| } 560| | 561| 33.1k| __m256i check_asso; 562| 33.1k| if (default_or_url) { ------------------ | Branch (562:9): [Folded, False: 33.1k] ------------------ 563| 0| check_asso = _mm256_setr_epi8( 564| 0| 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x03, 565| 0| 0x07, 0x0B, 0x0E, 0x0B, 0x06, 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 566| 0| 0x01, 0x01, 0x01, 0x01, 0x03, 0x07, 0x0B, 0x0E, 0x0B, 0x06); 567| 33.1k| } else if (base64_url) { ------------------ | Branch (567:16): [Folded, False: 33.1k] ------------------ 568| 0| check_asso = _mm256_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 569| 0| 0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6, 0xD, 0x1, 570| 0| 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x3, 571| 0| 0x7, 0xB, 0xE, 0xB, 0x6); 572| 33.1k| } else { 573| 33.1k| check_asso = _mm256_setr_epi8( 574| 33.1k| 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x03, 575| 33.1k| 0x07, 0x0B, 0x0B, 0x0B, 0x0F, 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 576| 33.1k| 0x01, 0x01, 0x01, 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F); 577| 33.1k| } 578| 33.1k| __m256i check_values; 579| 33.1k| if (default_or_url) { ------------------ | Branch (579:9): [Folded, False: 33.1k] ------------------ 580| 0| check_values = _mm256_setr_epi8( 581| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 582| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xD5), uint8_t(0xA6), 583| 0| uint8_t(0xB5), uint8_t(0xA1), uint8_t(0x00), uint8_t(0x80), 584| 0| uint8_t(0x00), uint8_t(0x80), uint8_t(0x00), uint8_t(0x80), 585| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 586| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xD5), uint8_t(0xA6), 587| 0| uint8_t(0xB5), uint8_t(0xA1), uint8_t(0x00), uint8_t(0x80), 588| 0| uint8_t(0x00), uint8_t(0x80), uint8_t(0x00), uint8_t(0x80)); 589| 33.1k| } else if (base64_url) { ------------------ | Branch (589:16): [Folded, False: 33.1k] ------------------ 590| 0| check_values = _mm256_setr_epi8( 591| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 592| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6), uint8_t(0xA6), 593| 0| uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, uint8_t(0x80), 594| 0| 0x0, uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 595| 0| uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6), 596| 0| uint8_t(0xA6), uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, 597| 0| uint8_t(0x80), 0x0, uint8_t(0x80)); 598| 33.1k| } else { 599| 33.1k| check_values = _mm256_setr_epi8( 600| 33.1k| int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0xCF), 601| 33.1k| int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5), int8_t(0x86), 602| 33.1k| int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80), int8_t(0x91), 603| 33.1k| int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), 604| 33.1k| int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5), 605| 33.1k| int8_t(0x86), int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80), 606| 33.1k| int8_t(0x91), int8_t(0x80)); 607| 33.1k| } 608| 33.1k| const __m256i shifted = _mm256_srli_epi32(*src, 3); 609| 33.1k| __m256i delta_hash = 610| 33.1k| _mm256_avg_epu8(_mm256_shuffle_epi8(delta_asso, *src), shifted); 611| 33.1k| if (default_or_url) { ------------------ | Branch (611:9): [Folded, False: 33.1k] ------------------ 612| 0| delta_hash = _mm256_and_si256(delta_hash, _mm256_set1_epi8(0xf)); 613| 0| } 614| 33.1k| const __m256i check_hash = 615| 33.1k| _mm256_avg_epu8(_mm256_shuffle_epi8(check_asso, *src), shifted); 616| 33.1k| const __m256i out = 617| 33.1k| _mm256_adds_epi8(_mm256_shuffle_epi8(delta_values, delta_hash), *src); 618| 33.1k| const __m256i chk = 619| 33.1k| _mm256_adds_epi8(_mm256_shuffle_epi8(check_values, check_hash), *src); 620| 33.1k| const int mask = _mm256_movemask_epi8(chk); 621| 33.1k| if (!ignore_garbage && mask) { ------------------ | Branch (621:9): [True: 33.1k, Folded] | Branch (621:28): [True: 4.71k, False: 28.4k] ------------------ 622| 4.71k| __m256i ascii_space = 623| 4.71k| _mm256_cmpeq_epi8(_mm256_shuffle_epi8(ascii_space_tbl, *src), *src); 624| 4.71k| *error = (mask ^ _mm256_movemask_epi8(ascii_space)); 625| 4.71k| } 626| 33.1k| *src = out; 627| 33.1k| return (uint32_t)mask; 628| 33.1k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_113encode_base64ILb0EEEmPcPKcmNS_14base64_optionsE: 357| 14.3k| base64_options options) { 358| 14.3k| return avx2_encode_base64_impl(dst, src, srclen, options); 359| 14.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_123avx2_encode_base64_implILb0ELb0EEEmPcPKcmNS_14base64_optionsEm: 99| 14.3k| size_t line_length = simdutf::default_line_length) { 100| 14.3k| size_t offset = 0; 101| | 102| 14.3k| if (line_length < 4) { ------------------ | Branch (102:7): [True: 0, False: 14.3k] ------------------ 103| 0| line_length = 4; // We do not support line_length less than 4 104| 0| } 105| | // credit: Wojciech Muła 106| 14.3k| const uint8_t *input = (const uint8_t *)src; 107| | 108| 14.3k| uint8_t *out = (uint8_t *)dst; 109| 14.3k| const __m256i shuf = 110| 14.3k| _mm256_set_epi8(10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1, 111| | 112| 14.3k| 10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1); 113| 14.3k| size_t i = 0; 114| 16.4k| for (; i + 100 <= srclen; i += 96) { ------------------ | Branch (114:10): [True: 2.01k, False: 14.3k] ------------------ 115| 2.01k| const __m128i lo0 = _mm_loadu_si128( 116| 2.01k| reinterpret_cast(input + i + 4 * 3 * 0)); 117| 2.01k| const __m128i hi0 = _mm_loadu_si128( 118| 2.01k| reinterpret_cast(input + i + 4 * 3 * 1)); 119| 2.01k| const __m128i lo1 = _mm_loadu_si128( 120| 2.01k| reinterpret_cast(input + i + 4 * 3 * 2)); 121| 2.01k| const __m128i hi1 = _mm_loadu_si128( 122| 2.01k| reinterpret_cast(input + i + 4 * 3 * 3)); 123| 2.01k| const __m128i lo2 = _mm_loadu_si128( 124| 2.01k| reinterpret_cast(input + i + 4 * 3 * 4)); 125| 2.01k| const __m128i hi2 = _mm_loadu_si128( 126| 2.01k| reinterpret_cast(input + i + 4 * 3 * 5)); 127| 2.01k| const __m128i lo3 = _mm_loadu_si128( 128| 2.01k| reinterpret_cast(input + i + 4 * 3 * 6)); 129| 2.01k| const __m128i hi3 = _mm_loadu_si128( 130| 2.01k| reinterpret_cast(input + i + 4 * 3 * 7)); 131| | 132| 2.01k| __m256i in0 = _mm256_shuffle_epi8(_mm256_set_m128i(hi0, lo0), shuf); ------------------ | | 4| 2.01k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.01k| _mm256_castsi128_si256(xmm2), 2) ------------------ 133| 2.01k| __m256i in1 = _mm256_shuffle_epi8(_mm256_set_m128i(hi1, lo1), shuf); ------------------ | | 4| 2.01k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.01k| _mm256_castsi128_si256(xmm2), 2) ------------------ 134| 2.01k| __m256i in2 = _mm256_shuffle_epi8(_mm256_set_m128i(hi2, lo2), shuf); ------------------ | | 4| 2.01k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.01k| _mm256_castsi128_si256(xmm2), 2) ------------------ 135| 2.01k| __m256i in3 = _mm256_shuffle_epi8(_mm256_set_m128i(hi3, lo3), shuf); ------------------ | | 4| 2.01k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 2.01k| _mm256_castsi128_si256(xmm2), 2) ------------------ 136| | 137| 2.01k| const __m256i t0_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x0fc0fc00)); 138| 2.01k| const __m256i t0_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x0fc0fc00)); 139| 2.01k| const __m256i t0_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x0fc0fc00)); 140| 2.01k| const __m256i t0_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x0fc0fc00)); 141| | 142| 2.01k| const __m256i t1_0 = 143| 2.01k| _mm256_mulhi_epu16(t0_0, _mm256_set1_epi32(0x04000040)); 144| 2.01k| const __m256i t1_1 = 145| 2.01k| _mm256_mulhi_epu16(t0_1, _mm256_set1_epi32(0x04000040)); 146| 2.01k| const __m256i t1_2 = 147| 2.01k| _mm256_mulhi_epu16(t0_2, _mm256_set1_epi32(0x04000040)); 148| 2.01k| const __m256i t1_3 = 149| 2.01k| _mm256_mulhi_epu16(t0_3, _mm256_set1_epi32(0x04000040)); 150| | 151| 2.01k| const __m256i t2_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x003f03f0)); 152| 2.01k| const __m256i t2_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x003f03f0)); 153| 2.01k| const __m256i t2_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x003f03f0)); 154| 2.01k| const __m256i t2_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x003f03f0)); 155| | 156| 2.01k| const __m256i t3_0 = 157| 2.01k| _mm256_mullo_epi16(t2_0, _mm256_set1_epi32(0x01000010)); 158| 2.01k| const __m256i t3_1 = 159| 2.01k| _mm256_mullo_epi16(t2_1, _mm256_set1_epi32(0x01000010)); 160| 2.01k| const __m256i t3_2 = 161| 2.01k| _mm256_mullo_epi16(t2_2, _mm256_set1_epi32(0x01000010)); 162| 2.01k| const __m256i t3_3 = 163| 2.01k| _mm256_mullo_epi16(t2_3, _mm256_set1_epi32(0x01000010)); 164| | 165| 2.01k| const __m256i input0 = _mm256_or_si256(t1_0, t3_0); 166| 2.01k| const __m256i input1 = _mm256_or_si256(t1_1, t3_1); 167| 2.01k| const __m256i input2 = _mm256_or_si256(t1_2, t3_2); 168| 2.01k| const __m256i input3 = _mm256_or_si256(t1_3, t3_3); 169| | 170| 2.01k| if (use_lines) { ------------------ | Branch (170:9): [Folded, False: 2.01k] ------------------ 171| 0| if (line_length >= 32) { // fast path ------------------ | Branch (171:11): [True: 0, False: 0] ------------------ 172| 0| __m256i result; 173| 0| result = lookup_pshufb_improved(input0); 174| 0| if (offset + 32 > line_length) { ------------------ | Branch (174:13): [True: 0, False: 0] ------------------ 175| 0| size_t location_end = line_length - offset; 176| 0| size_t to_move = 32 - location_end; 177| | // We could do this, or extract instead. 178| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 179| 0| _mm256_storeu_si256( 180| 0| reinterpret_cast<__m256i *>(out), 181| 0| insert_line_feed32(result, static_cast(location_end))); 182| 0| offset = to_move; 183| 0| out += 32 + 1; 184| 0| } else { 185| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 186| 0| offset += 32; 187| 0| out += 32; 188| 0| } 189| 0| result = lookup_pshufb_improved(input1); 190| | 191| 0| if (offset + 32 > line_length) { ------------------ | Branch (191:13): [True: 0, False: 0] ------------------ 192| 0| size_t location_end = line_length - offset; 193| 0| size_t to_move = 32 - location_end; 194| | 195| | // We could do this, or extract instead. 196| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 197| 0| _mm256_storeu_si256( 198| 0| reinterpret_cast<__m256i *>(out), 199| 0| insert_line_feed32(result, static_cast(location_end))); 200| | // see above. 201| | // out[32] = static_cast(_mm256_extract_epi8(result, 31)); 202| 0| offset = to_move; 203| 0| out += 32 + 1; 204| 0| } else { 205| | 206| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 207| | 208| 0| offset += 32; 209| 0| out += 32; 210| 0| } 211| 0| result = lookup_pshufb_improved(input2); 212| | 213| 0| if (offset + 32 > line_length) { ------------------ | Branch (213:13): [True: 0, False: 0] ------------------ 214| 0| size_t location_end = line_length - offset; 215| 0| size_t to_move = 32 - location_end; 216| | 217| | // We could do this, or extract instead. 218| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 219| 0| _mm256_storeu_si256( 220| 0| reinterpret_cast<__m256i *>(out), 221| 0| insert_line_feed32(result, static_cast(location_end))); 222| | // see above. 223| | // out[32] = static_cast(_mm256_extract_epi8(result, 31)); 224| 0| offset = to_move; 225| 0| out += 32 + 1; 226| 0| } else { 227| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 228| 0| offset += 32; 229| 0| out += 32; 230| 0| } 231| 0| result = lookup_pshufb_improved(input3); 232| | 233| 0| if (offset + 32 > line_length) { ------------------ | Branch (233:13): [True: 0, False: 0] ------------------ 234| 0| size_t location_end = line_length - offset; 235| 0| size_t to_move = 32 - location_end; 236| | 237| | // We could do this, or extract instead. 238| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 1), result); 239| 0| _mm256_storeu_si256( 240| 0| reinterpret_cast<__m256i *>(out), 241| 0| insert_line_feed32(result, static_cast(location_end))); 242| | // see above. 243| | // out[32] = static_cast(_mm256_extract_epi8(result, 31)); 244| 0| offset = to_move; 245| 0| out += 32 + 1; 246| 0| } else { 247| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), result); 248| 0| offset += 32; 249| 0| out += 32; 250| 0| } 251| 0| } else { // slow path 252| | // could be optimized 253| 0| uint8_t buffer[128]; 254| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer), 255| 0| lookup_pshufb_improved(input0)); 256| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer + 32), 257| 0| lookup_pshufb_improved(input1)); 258| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer + 64), 259| 0| lookup_pshufb_improved(input2)); 260| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer + 96), 261| 0| lookup_pshufb_improved(input3)); 262| 0| size_t out_pos = 0; 263| 0| size_t local_offset = offset; 264| 0| for (size_t j = 0; j < 128;) { ------------------ | Branch (264:28): [True: 0, False: 0] ------------------ 265| 0| if (local_offset == line_length) { ------------------ | Branch (265:15): [True: 0, False: 0] ------------------ 266| 0| out[out_pos++] = '\n'; 267| 0| local_offset = 0; 268| 0| } 269| 0| out[out_pos++] = buffer[j++]; 270| 0| local_offset++; 271| 0| } 272| 0| offset = local_offset; 273| 0| out += out_pos; 274| 0| } 275| 2.01k| } else { 276| 2.01k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), 277| 2.01k| lookup_pshufb_improved(input0)); 278| 2.01k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 32), 279| 2.01k| lookup_pshufb_improved(input1)); 280| 2.01k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 64), 281| 2.01k| lookup_pshufb_improved(input2)); 282| 2.01k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out + 96), 283| 2.01k| lookup_pshufb_improved(input3)); 284| | 285| 2.01k| out += 128; 286| 2.01k| } 287| 2.01k| } 288| 16.0k| for (; i + 28 <= srclen; i += 24) { ------------------ | Branch (288:10): [True: 1.63k, False: 14.3k] ------------------ 289| | // lo = [xxxx|DDDC|CCBB|BAAA] 290| | // hi = [xxxx|HHHG|GGFF|FEEE] 291| 1.63k| const __m128i lo = 292| 1.63k| _mm_loadu_si128(reinterpret_cast(input + i)); 293| 1.63k| const __m128i hi = 294| 1.63k| _mm_loadu_si128(reinterpret_cast(input + i + 4 * 3)); 295| | 296| | // bytes from groups A, B and C are needed in separate 32-bit lanes 297| | // in = [0HHH|0GGG|0FFF|0EEE[0DDD|0CCC|0BBB|0AAA] 298| 1.63k| __m256i in = _mm256_shuffle_epi8(_mm256_set_m128i(hi, lo), shuf); ------------------ | | 4| 1.63k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 1.63k| _mm256_castsi128_si256(xmm2), 2) ------------------ 299| | 300| | // this part is well commented in encode.sse.cpp 301| | 302| 1.63k| const __m256i t0 = _mm256_and_si256(in, _mm256_set1_epi32(0x0fc0fc00)); 303| 1.63k| const __m256i t1 = _mm256_mulhi_epu16(t0, _mm256_set1_epi32(0x04000040)); 304| 1.63k| const __m256i t2 = _mm256_and_si256(in, _mm256_set1_epi32(0x003f03f0)); 305| 1.63k| const __m256i t3 = _mm256_mullo_epi16(t2, _mm256_set1_epi32(0x01000010)); 306| 1.63k| const __m256i indices = _mm256_or_si256(t1, t3); 307| | 308| 1.63k| if (use_lines) { ------------------ | Branch (308:9): [Folded, False: 1.63k] ------------------ 309| 0| if (line_length >= 32) { // fast path ------------------ | Branch (309:11): [True: 0, False: 0] ------------------ 310| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), 311| 0| lookup_pshufb_improved(indices)); 312| | 313| 0| if (offset + 32 > line_length) { ------------------ | Branch (313:13): [True: 0, False: 0] ------------------ 314| 0| size_t location_end = line_length - offset; 315| 0| size_t to_move = 32 - location_end; 316| 0| std::memmove(out + location_end + 1, out + location_end, to_move); 317| 0| out[location_end] = '\n'; 318| 0| offset = to_move; 319| 0| out += 32 + 1; 320| 0| } else { 321| 0| offset += 32; 322| 0| out += 32; 323| 0| } 324| 0| } else { // slow path 325| | // could be optimized 326| 0| alignas(32) uint8_t buffer[32]; 327| 0| _mm256_storeu_si256(reinterpret_cast<__m256i *>(buffer), 328| 0| lookup_pshufb_improved(indices)); 329| 0| std::memcpy(out, buffer, 32); 330| 0| size_t out_pos = 0; 331| 0| size_t local_offset = offset; 332| 0| for (size_t j = 0; j < 32;) { ------------------ | Branch (332:28): [True: 0, False: 0] ------------------ 333| 0| if (local_offset == line_length) { ------------------ | Branch (333:15): [True: 0, False: 0] ------------------ 334| 0| out[out_pos++] = '\n'; 335| 0| local_offset = 0; 336| 0| } 337| 0| out[out_pos++] = buffer[j++]; 338| 0| local_offset++; 339| 0| } 340| 0| offset = local_offset; 341| 0| out += out_pos; 342| 0| } 343| 1.63k| } else { 344| 1.63k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(out), 345| 1.63k| lookup_pshufb_improved(indices)); 346| | 347| 1.63k| out += 32; 348| 1.63k| } 349| 1.63k| } 350| 14.3k| return ((char *)out - (char *)dst) + 351| 14.3k| scalar::base64::tail_encode_base64_impl( 352| 14.3k| (char *)out, src + i, srclen - i, options, line_length, offset); 353| 14.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_122lookup_pshufb_improvedILb0EEEDv4_xS3_: 30| 9.69k|simdutf_really_inline __m256i lookup_pshufb_improved(const __m256i input) { 31| | // Precomputed shuffle masks for K = 1 to 16 32| | // credit: Wojciech Muła 33| 9.69k| __m256i result = _mm256_subs_epu8(input, _mm256_set1_epi8(51)); 34| 9.69k| const __m256i less = _mm256_cmpgt_epi8(_mm256_set1_epi8(26), input); 35| 9.69k| result = 36| 9.69k| _mm256_or_si256(result, _mm256_and_si256(less, _mm256_set1_epi8(13))); 37| 9.69k| __m256i shift_LUT; 38| 9.69k| if (base64_url) { ------------------ | Branch (38:7): [Folded, False: 9.69k] ------------------ 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| 9.69k| } else { 46| 9.69k| shift_LUT = _mm256_setr_epi8( 47| 9.69k| 'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 48| 9.69k| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0, 49| | 50| 9.69k| 'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 51| 9.69k| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0); 52| 9.69k| } 53| | 54| 9.69k| result = _mm256_shuffle_epi8(shift_LUT, result); 55| 9.69k| return _mm256_add_epi8(result, input); 56| 9.69k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_latin1_to_utf16ILNS_10endiannessE0EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.26k| char16_t *utf16_output) { 5| 5.26k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| | 7| 5.26k| size_t i = 0; 8| 14.6k| for (; i < rounded_len; i += 16) { ------------------ | Branch (8:10): [True: 9.39k, False: 5.26k] ------------------ 9| | // Load 16 bytes from the address (input + i) into a xmm register 10| 9.39k| const __m128i latin1 = 11| 9.39k| _mm_loadu_si128(reinterpret_cast(latin1_input + i)); 12| | 13| | // Zero extend each byte in `in` to word 14| 9.39k| __m256i utf16 = _mm256_cvtepu8_epi16(latin1); 15| | 16| 9.39k| if (big_endian) { ------------------ | Branch (16:9): [Folded, False: 9.39k] ------------------ 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| 9.39k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output + i), utf16); 25| 9.39k| } 26| | 27| 5.26k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 28| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_latin1_to_utf16ILNS_10endiannessE1EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.26k| char16_t *utf16_output) { 5| 5.26k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| | 7| 5.26k| size_t i = 0; 8| 14.6k| for (; i < rounded_len; i += 16) { ------------------ | Branch (8:10): [True: 9.39k, False: 5.26k] ------------------ 9| | // Load 16 bytes from the address (input + i) into a xmm register 10| 9.39k| const __m128i latin1 = 11| 9.39k| _mm_loadu_si128(reinterpret_cast(latin1_input + i)); 12| | 13| | // Zero extend each byte in `in` to word 14| 9.39k| __m256i utf16 = _mm256_cvtepu8_epi16(latin1); 15| | 16| 9.39k| if (big_endian) { ------------------ | Branch (16:9): [True: 9.39k, Folded] ------------------ 17| 9.39k| const __m128i swap128 = 18| 9.39k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 19| 9.39k| const __m256i swap = _mm256_set_m128i(swap128, swap128); ------------------ | | 4| 9.39k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 9.39k| _mm256_castsi128_si256(xmm2), 2) ------------------ 20| 9.39k| utf16 = _mm256_shuffle_epi8(utf16, swap); 21| 9.39k| } 22| | 23| | // Store the contents of xmm1 into the address pointed by (output + i) 24| 9.39k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output + i), utf16); 25| 9.39k| } 26| | 27| 5.26k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 28| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_latin1_to_utf32EPKcmPDi: 3| 5.26k| char32_t *utf32_output) { 4| 5.26k| size_t rounded_len = ((len | 7) ^ 7); // Round down to nearest multiple of 8 5| | 6| 24.2k| for (size_t i = 0; i < rounded_len; i += 8) { ------------------ | Branch (6:22): [True: 18.9k, False: 5.26k] ------------------ 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| 5.26k| return std::make_pair(buf + rounded_len, utf32_output + rounded_len); 20| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_127avx2_convert_latin1_to_utf8EPKcmPc: 3| 5.53k| char *utf8_output) { 4| 5.53k| const char *end = latin1_input + len; 5| 5.53k| const __m256i v_0000 = _mm256_setzero_si256(); 6| 5.53k| const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080); 7| 5.53k| const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80); 8| 5.53k| const size_t safety_margin = 12; 9| | 10| 16.1k| while (end - latin1_input >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (10:10): [True: 10.6k, False: 5.53k] ------------------ 11| 10.6k| __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| 10.6k| const __m128i v_80 = _mm_set1_epi8((char)0x80); 14| 10.6k| if (_mm_testz_si128(in8, v_80)) { // ASCII fast path!!!! ------------------ | Branch (14:9): [True: 7.88k, False: 2.76k] ------------------ 15| | // 1. store (16 bytes) 16| 7.88k| _mm_storeu_si128((__m128i *)utf8_output, in8); 17| | // 2. adjust pointers 18| 7.88k| latin1_input += 16; 19| 7.88k| utf8_output += 16; 20| 7.88k| continue; // we are done for this round! 21| 7.88k| } 22| | // We proceed only with the first 16 bytes. 23| 2.76k| 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| 2.76k| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 29| 2.76k| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 30| | 31| | // t0 = [0000|00aa|bbbb|bb00] 32| 2.76k| const __m256i t0 = _mm256_slli_epi16(in, 2); 33| | // t1 = [0000|00aa|0000|0000] 34| 2.76k| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 35| | // t2 = [0000|0000|00bb|bbbb] 36| 2.76k| const __m256i t2 = _mm256_and_si256(in, v_003f); 37| | // t3 = [000a|aaaa|00bb|bbbb] 38| 2.76k| const __m256i t3 = _mm256_or_si256(t1, t2); 39| | // t4 = [1100|00aa|10bb|bbbb] 40| 2.76k| 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| 2.76k| const __m256i one_byte_bytemask = 46| 2.76k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000); 47| 2.76k| const uint32_t one_byte_bitmask = 48| 2.76k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 49| | 50| 2.76k| const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in, one_byte_bytemask); 51| | 52| | // 3. prepare bitmask for 8-bit lookup 53| 2.76k| const uint32_t M0 = one_byte_bitmask & 0x55555555; 54| 2.76k| const uint32_t M1 = M0 >> 7; 55| 2.76k| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 56| | // 4. pack the bytes 57| | 58| 2.76k| const uint8_t *row = 59| 2.76k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 60| 2.76k| const uint8_t *row_2 = 61| 2.76k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 16)] 62| 2.76k| [0]; 63| | 64| 2.76k| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 65| 2.76k| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 66| | 67| 2.76k| const __m256i utf8_packed = _mm256_shuffle_epi8( 68| 2.76k| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 2.76k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 2.76k| _mm256_castsi128_si256(xmm2), 2) ------------------ 69| | // 5. store bytes 70| 2.76k| _mm_storeu_si128((__m128i *)utf8_output, 71| 2.76k| _mm256_castsi256_si128(utf8_packed)); 72| 2.76k| utf8_output += row[0]; 73| 2.76k| _mm_storeu_si128((__m128i *)utf8_output, 74| 2.76k| _mm256_extractf128_si256(utf8_packed, 1)); 75| 2.76k| utf8_output += row_2[0]; 76| | 77| | // 6. adjust pointers 78| 2.76k| latin1_input += 16; 79| 2.76k| continue; 80| | 81| 10.6k| } // while 82| 5.53k| return std::make_pair(latin1_input, utf8_output); 83| 5.53k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_utf16_to_latin1ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.26k| char *latin1_output) { 5| 5.26k| const char16_t *end = buf + len; 6| 9.84k| while (end - buf >= 32) { ------------------ | Branch (6:10): [True: 4.58k, False: 5.26k] ------------------ 7| | // Load 16 UTF-16 characters into 256-bit AVX2 register 8| 4.58k| __m256i in0 = _mm256_loadu_si256(reinterpret_cast(buf)); 9| 4.58k| __m256i in1 = 10| 4.58k| _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| 4.58k| __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00); 21| 4.58k| if (_mm256_testz_si256(_mm256_or_si256(in0, in1), high_byte_mask)) { ------------------ | Branch (21:9): [True: 4.58k, False: 0] ------------------ 22| | // Pack 16-bit characters into 8-bit and store in latin1_output 23| 4.58k| const __m256i packed = _mm256_packus_epi16(in0, in1); 24| | 25| 4.58k| const __m256i result = _mm256_permute4x64_epi64(packed, 0b11011000); 26| | 27| 4.58k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(latin1_output), result); 28| | // Adjust pointers for the next iteration 29| 4.58k| buf += 32; 30| 4.58k| latin1_output += 32; 31| 4.58k| } else { 32| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 33| 0| } 34| 4.58k| } // while 35| 5.26k| return std::make_pair(buf, latin1_output); 36| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_utf16_to_latin1ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.26k| char *latin1_output) { 5| 5.26k| const char16_t *end = buf + len; 6| 9.84k| while (end - buf >= 32) { ------------------ | Branch (6:10): [True: 4.58k, False: 5.26k] ------------------ 7| | // Load 16 UTF-16 characters into 256-bit AVX2 register 8| 4.58k| __m256i in0 = _mm256_loadu_si256(reinterpret_cast(buf)); 9| 4.58k| __m256i in1 = 10| 4.58k| _mm256_loadu_si256(reinterpret_cast(buf + 16)); 11| | 12| 4.58k| if constexpr (!match_system(big_endian)) { 13| 4.58k| const __m256i swap = _mm256_setr_epi8( 14| 4.58k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 15| 4.58k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 16| 4.58k| in0 = _mm256_shuffle_epi8(in0, swap); 17| 4.58k| in1 = _mm256_shuffle_epi8(in1, swap); 18| 4.58k| } 19| | 20| 4.58k| __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00); 21| 4.58k| if (_mm256_testz_si256(_mm256_or_si256(in0, in1), high_byte_mask)) { ------------------ | Branch (21:9): [True: 4.58k, False: 0] ------------------ 22| | // Pack 16-bit characters into 8-bit and store in latin1_output 23| 4.58k| const __m256i packed = _mm256_packus_epi16(in0, in1); 24| | 25| 4.58k| const __m256i result = _mm256_permute4x64_epi64(packed, 0b11011000); 26| | 27| 4.58k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(latin1_output), result); 28| | // Adjust pointers for the next iteration 29| 4.58k| buf += 32; 30| 4.58k| latin1_output += 32; 31| 4.58k| } else { 32| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 33| 0| } 34| 4.58k| } // while 35| 5.26k| return std::make_pair(buf, latin1_output); 36| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_126avx2_convert_utf16_to_utf8ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 9.93k|avx2_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| 9.93k| const char16_t *end = buf + len; 58| 9.93k| const __m256i v_0000 = _mm256_setzero_si256(); 59| 9.93k| const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800); 60| 9.93k| const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800); 61| 9.93k| const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080); 62| 9.93k| const size_t safety_margin = 63| 9.93k| 12; // to avoid overruns, see issue 64| | // https://github.com/simdutf/simdutf/issues/92 65| | 66| 14.6k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (66:10): [True: 4.82k, False: 9.87k] ------------------ 67| 4.82k| __m256i in = _mm256_loadu_si256((__m256i *)buf); 68| 4.82k| if (big_endian) { ------------------ | Branch (68:9): [Folded, False: 4.82k] ------------------ 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| 4.82k| const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80); 76| 4.82k| if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 2.38k, False: 2.43k] ------------------ 77| | // 1. pack the bytes 78| 2.38k| const __m128i utf8_packed = _mm_packus_epi16( 79| 2.38k| _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1)); 80| | // 2. store (16 bytes) 81| 2.38k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 82| | // 3. adjust pointers 83| 2.38k| buf += 16; 84| 2.38k| utf8_output += 16; 85| 2.38k| continue; // we are done for this round! 86| 2.38k| } 87| | // no bits set above 7th bit 88| 2.43k| const __m256i one_byte_bytemask = 89| 2.43k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000); 90| 2.43k| const uint32_t one_byte_bitmask = 91| 2.43k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 92| | 93| | // no bits set above 11th bit 94| 2.43k| const __m256i one_or_two_bytes_bytemask = 95| 2.43k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000); 96| 2.43k| const uint32_t one_or_two_bytes_bitmask = 97| 2.43k| static_cast(_mm256_movemask_epi8(one_or_two_bytes_bytemask)); 98| 2.43k| if (one_or_two_bytes_bitmask == 0xffffffff) { ------------------ | Branch (98:9): [True: 182, 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| 182| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 104| 182| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 105| | 106| | // t0 = [000a|aaaa|bbbb|bb00] 107| 182| const __m256i t0 = _mm256_slli_epi16(in, 2); 108| | // t1 = [000a|aaaa|0000|0000] 109| 182| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 110| | // t2 = [0000|0000|00bb|bbbb] 111| 182| const __m256i t2 = _mm256_and_si256(in, v_003f); 112| | // t3 = [000a|aaaa|00bb|bbbb] 113| 182| const __m256i t3 = _mm256_or_si256(t1, t2); 114| | // t4 = [110a|aaaa|10bb|bbbb] 115| 182| const __m256i t4 = _mm256_or_si256(t3, v_c080); 116| | 117| | // 2. merge ASCII and 2-byte codewords 118| 182| const __m256i utf8_unpacked = 119| 182| _mm256_blendv_epi8(t4, in, one_byte_bytemask); 120| | 121| | // 3. prepare bitmask for 8-bit lookup 122| 182| const uint32_t M0 = one_byte_bitmask & 0x55555555; 123| 182| const uint32_t M1 = M0 >> 7; 124| 182| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 125| | // 4. pack the bytes 126| | 127| 182| const uint8_t *row = 128| 182| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 129| 182| const uint8_t *row_2 = 130| 182| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 131| 182| 16)][0]; 132| | 133| 182| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 134| 182| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 135| | 136| 182| const __m256i utf8_packed = _mm256_shuffle_epi8( 137| 182| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 182| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 182| _mm256_castsi128_si256(xmm2), 2) ------------------ 138| | // 5. store bytes 139| 182| _mm_storeu_si128((__m128i *)utf8_output, 140| 182| _mm256_castsi256_si128(utf8_packed)); 141| 182| utf8_output += row[0]; 142| 182| _mm_storeu_si128((__m128i *)utf8_output, 143| 182| _mm256_extractf128_si256(utf8_packed, 1)); 144| 182| utf8_output += row_2[0]; 145| | 146| | // 6. adjust pointers 147| 182| buf += 16; 148| 182| continue; 149| 182| } 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: 1.02k, False: 1.23k] ------------------ 163| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 164| 1.02k| const __m256i dup_even = _mm256_setr_epi16( 165| 1.02k| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e, 166| 1.02k| 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| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) 196| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 197| 1.02k| const __m256i t0 = _mm256_shuffle_epi8(in, dup_even); 198| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 199| 1.02k| const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 195| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 200| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 201| 1.02k| const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 195| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 202| | 203| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 204| 1.02k| const __m256i s0 = _mm256_srli_epi16(in, 4); 205| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 206| 1.02k| const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 195| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 207| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 208| 1.02k| const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 195| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 209| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 210| 1.02k| const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 195| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 211| 1.02k| const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, 212| 1.02k| simdutf_vec(0b0100000000000000)); ------------------ | | 195| 1.02k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 213| 1.02k| const __m256i s4 = _mm256_xor_si256(s3, m0); 214| 1.02k|#undef simdutf_vec 215| | 216| | // 4. expand code units 16-bit => 32-bit 217| 1.02k| const __m256i out0 = _mm256_unpacklo_epi16(t2, s4); 218| 1.02k| 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| 1.02k| const uint32_t mask = (one_byte_bitmask & 0x55555555) | 222| 1.02k| (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| 1.02k| const uint8_t mask0 = uint8_t(mask); 243| 1.02k| const uint8_t *row0 = 244| 1.02k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 245| 1.02k| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 246| 1.02k| const __m128i utf8_0 = 247| 1.02k| _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0); 248| | 249| 1.02k| const uint8_t mask1 = static_cast(mask >> 8); 250| 1.02k| const uint8_t *row1 = 251| 1.02k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 252| 1.02k| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 253| 1.02k| const __m128i utf8_1 = 254| 1.02k| _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1); 255| | 256| 1.02k| const uint8_t mask2 = static_cast(mask >> 16); 257| 1.02k| const uint8_t *row2 = 258| 1.02k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0]; 259| 1.02k| const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1)); 260| 1.02k| const __m128i utf8_2 = 261| 1.02k| _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2); 262| | 263| 1.02k| const uint8_t mask3 = static_cast(mask >> 24); 264| 1.02k| const uint8_t *row3 = 265| 1.02k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0]; 266| 1.02k| const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1)); 267| 1.02k| const __m128i utf8_3 = 268| 1.02k| _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3); 269| | 270| 1.02k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 271| 1.02k| utf8_output += row0[0]; 272| 1.02k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 273| 1.02k| utf8_output += row1[0]; 274| 1.02k| _mm_storeu_si128((__m128i *)utf8_output, utf8_2); 275| 1.02k| utf8_output += row2[0]; 276| 1.02k| _mm_storeu_si128((__m128i *)utf8_output, utf8_3); 277| 1.02k| utf8_output += row3[0]; 278| 1.02k| buf += 16; 279| | // surrogate pair(s) in a register 280| 1.23k| } 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| 1.23k| size_t forward = 15; 285| 1.23k| size_t k = 0; 286| 1.23k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (286:11): [True: 0, False: 1.23k] ------------------ 287| 0| forward = size_t(end - buf - 1); 288| 0| } 289| 17.1k| for (; k < forward; k++) { ------------------ | Branch (289:14): [True: 15.9k, False: 1.16k] ------------------ 290| 15.9k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 291| 15.9k| if ((word & 0xFF80) == 0) { ------------------ | Branch (291:13): [True: 9.51k, False: 6.46k] ------------------ 292| 9.51k| *utf8_output++ = char(word); 293| 9.51k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (293:20): [True: 531, False: 5.93k] ------------------ 294| 531| *utf8_output++ = char((word >> 6) | 0b11000000); 295| 531| *utf8_output++ = char((word & 0b111111) | 0b10000000); 296| 5.93k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (296:20): [True: 3.92k, False: 2.01k] ------------------ 297| 3.92k| *utf8_output++ = char((word >> 12) | 0b11100000); 298| 3.92k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 299| 3.92k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 300| 3.92k| } else { 301| | // must be a surrogate pair 302| 2.01k| uint16_t diff = uint16_t(word - 0xD800); 303| 2.01k| uint16_t next_word = 304| 2.01k| scalar::utf16::swap_if_needed(buf[k + 1]); 305| 2.01k| k++; 306| 2.01k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 307| 2.01k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (307:15): [True: 65, False: 1.95k] ------------------ 308| 65| return std::make_pair(nullptr, utf8_output); 309| 65| } 310| 1.95k| uint32_t value = (diff << 10) + diff2 + 0x10000; 311| 1.95k| *utf8_output++ = char((value >> 18) | 0b11110000); 312| 1.95k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 313| 1.95k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 314| 1.95k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 315| 1.95k| } 316| 15.9k| } 317| 1.16k| buf += k; 318| 1.16k| } 319| 2.25k| } // while 320| 9.87k| return std::make_pair(buf, utf8_output); 321| 9.93k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_126avx2_convert_utf16_to_utf8ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 9.93k|avx2_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| 9.93k| const char16_t *end = buf + len; 58| 9.93k| const __m256i v_0000 = _mm256_setzero_si256(); 59| 9.93k| const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800); 60| 9.93k| const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800); 61| 9.93k| const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080); 62| 9.93k| const size_t safety_margin = 63| 9.93k| 12; // to avoid overruns, see issue 64| | // https://github.com/simdutf/simdutf/issues/92 65| | 66| 14.8k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (66:10): [True: 4.94k, False: 9.87k] ------------------ 67| 4.94k| __m256i in = _mm256_loadu_si256((__m256i *)buf); 68| 4.94k| if (big_endian) { ------------------ | Branch (68:9): [True: 4.94k, Folded] ------------------ 69| 4.94k| const __m256i swap = _mm256_setr_epi8( 70| 4.94k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 71| 4.94k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 72| 4.94k| in = _mm256_shuffle_epi8(in, swap); 73| 4.94k| } 74| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 75| 4.94k| const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80); 76| 4.94k| if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 2.40k, False: 2.54k] ------------------ 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| 2.54k| const __m256i one_byte_bytemask = 89| 2.54k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000); 90| 2.54k| const uint32_t one_byte_bitmask = 91| 2.54k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 92| | 93| | // no bits set above 11th bit 94| 2.54k| const __m256i one_or_two_bytes_bytemask = 95| 2.54k| _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000); 96| 2.54k| const uint32_t one_or_two_bytes_bitmask = 97| 2.54k| static_cast(_mm256_movemask_epi8(one_or_two_bytes_bytemask)); 98| 2.54k| if (one_or_two_bytes_bitmask == 0xffffffff) { ------------------ | Branch (98:9): [True: 164, False: 2.38k] ------------------ 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| 164| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 104| 164| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 105| | 106| | // t0 = [000a|aaaa|bbbb|bb00] 107| 164| const __m256i t0 = _mm256_slli_epi16(in, 2); 108| | // t1 = [000a|aaaa|0000|0000] 109| 164| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 110| | // t2 = [0000|0000|00bb|bbbb] 111| 164| const __m256i t2 = _mm256_and_si256(in, v_003f); 112| | // t3 = [000a|aaaa|00bb|bbbb] 113| 164| const __m256i t3 = _mm256_or_si256(t1, t2); 114| | // t4 = [110a|aaaa|10bb|bbbb] 115| 164| const __m256i t4 = _mm256_or_si256(t3, v_c080); 116| | 117| | // 2. merge ASCII and 2-byte codewords 118| 164| const __m256i utf8_unpacked = 119| 164| _mm256_blendv_epi8(t4, in, one_byte_bytemask); 120| | 121| | // 3. prepare bitmask for 8-bit lookup 122| 164| const uint32_t M0 = one_byte_bitmask & 0x55555555; 123| 164| const uint32_t M1 = M0 >> 7; 124| 164| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 125| | // 4. pack the bytes 126| | 127| 164| const uint8_t *row = 128| 164| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 129| 164| const uint8_t *row_2 = 130| 164| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 131| 164| 16)][0]; 132| | 133| 164| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 134| 164| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 135| | 136| 164| const __m256i utf8_packed = _mm256_shuffle_epi8( 137| 164| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 164| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 164| _mm256_castsi128_si256(xmm2), 2) ------------------ 138| | // 5. store bytes 139| 164| _mm_storeu_si128((__m128i *)utf8_output, 140| 164| _mm256_castsi256_si128(utf8_packed)); 141| 164| utf8_output += row[0]; 142| 164| _mm_storeu_si128((__m128i *)utf8_output, 143| 164| _mm256_extractf128_si256(utf8_packed, 1)); 144| 164| utf8_output += row_2[0]; 145| | 146| | // 6. adjust pointers 147| 164| buf += 16; 148| 164| continue; 149| 164| } 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| 2.38k| const __m256i surrogates_bytemask = 154| 2.38k| _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| 2.38k| const uint32_t surrogates_bitmask = 159| 2.38k| 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| 2.38k| if (surrogates_bitmask == 0x00000000) { ------------------ | Branch (162:9): [True: 1.21k, False: 1.17k] ------------------ 163| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 164| 1.21k| const __m256i dup_even = _mm256_setr_epi16( 165| 1.21k| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e, 166| 1.21k| 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| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) 196| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 197| 1.21k| const __m256i t0 = _mm256_shuffle_epi8(in, dup_even); 198| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 199| 1.21k| const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 195| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 200| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 201| 1.21k| const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 195| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 202| | 203| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 204| 1.21k| const __m256i s0 = _mm256_srli_epi16(in, 4); 205| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 206| 1.21k| const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 195| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 207| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 208| 1.21k| const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 195| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 209| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 210| 1.21k| const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 195| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 211| 1.21k| const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, 212| 1.21k| simdutf_vec(0b0100000000000000)); ------------------ | | 195| 1.21k|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 213| 1.21k| const __m256i s4 = _mm256_xor_si256(s3, m0); 214| 1.21k|#undef simdutf_vec 215| | 216| | // 4. expand code units 16-bit => 32-bit 217| 1.21k| const __m256i out0 = _mm256_unpacklo_epi16(t2, s4); 218| 1.21k| 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| 1.21k| const uint32_t mask = (one_byte_bitmask & 0x55555555) | 222| 1.21k| (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| 1.21k| const uint8_t mask0 = uint8_t(mask); 243| 1.21k| const uint8_t *row0 = 244| 1.21k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 245| 1.21k| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 246| 1.21k| const __m128i utf8_0 = 247| 1.21k| _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0); 248| | 249| 1.21k| const uint8_t mask1 = static_cast(mask >> 8); 250| 1.21k| const uint8_t *row1 = 251| 1.21k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 252| 1.21k| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 253| 1.21k| const __m128i utf8_1 = 254| 1.21k| _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1); 255| | 256| 1.21k| const uint8_t mask2 = static_cast(mask >> 16); 257| 1.21k| const uint8_t *row2 = 258| 1.21k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0]; 259| 1.21k| const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1)); 260| 1.21k| const __m128i utf8_2 = 261| 1.21k| _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2); 262| | 263| 1.21k| const uint8_t mask3 = static_cast(mask >> 24); 264| 1.21k| const uint8_t *row3 = 265| 1.21k| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0]; 266| 1.21k| const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1)); 267| 1.21k| const __m128i utf8_3 = 268| 1.21k| _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3); 269| | 270| 1.21k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 271| 1.21k| utf8_output += row0[0]; 272| 1.21k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 273| 1.21k| utf8_output += row1[0]; 274| 1.21k| _mm_storeu_si128((__m128i *)utf8_output, utf8_2); 275| 1.21k| utf8_output += row2[0]; 276| 1.21k| _mm_storeu_si128((__m128i *)utf8_output, utf8_3); 277| 1.21k| utf8_output += row3[0]; 278| 1.21k| buf += 16; 279| | // surrogate pair(s) in a register 280| 1.21k| } 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| 1.17k| size_t forward = 15; 285| 1.17k| size_t k = 0; 286| 1.17k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (286:11): [True: 0, False: 1.17k] ------------------ 287| 0| forward = size_t(end - buf - 1); 288| 0| } 289| 16.3k| for (; k < forward; k++) { ------------------ | Branch (289:14): [True: 15.2k, False: 1.11k] ------------------ 290| 15.2k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 291| 15.2k| if ((word & 0xFF80) == 0) { ------------------ | Branch (291:13): [True: 9.46k, False: 5.81k] ------------------ 292| 9.46k| *utf8_output++ = char(word); 293| 9.46k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (293:20): [True: 626, False: 5.18k] ------------------ 294| 626| *utf8_output++ = char((word >> 6) | 0b11000000); 295| 626| *utf8_output++ = char((word & 0b111111) | 0b10000000); 296| 5.18k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (296:20): [True: 3.31k, False: 1.87k] ------------------ 297| 3.31k| *utf8_output++ = char((word >> 12) | 0b11100000); 298| 3.31k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 299| 3.31k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 300| 3.31k| } else { 301| | // must be a surrogate pair 302| 1.87k| uint16_t diff = uint16_t(word - 0xD800); 303| 1.87k| uint16_t next_word = 304| 1.87k| scalar::utf16::swap_if_needed(buf[k + 1]); 305| 1.87k| k++; 306| 1.87k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 307| 1.87k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (307:15): [True: 57, False: 1.82k] ------------------ 308| 57| return std::make_pair(nullptr, utf8_output); 309| 57| } 310| 1.82k| uint32_t value = (diff << 10) + diff2 + 0x10000; 311| 1.82k| *utf8_output++ = char((value >> 18) | 0b11110000); 312| 1.82k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 313| 1.82k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 314| 1.82k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 315| 1.82k| } 316| 15.2k| } 317| 1.11k| buf += k; 318| 1.11k| } 319| 2.38k| } // while 320| 9.87k| return std::make_pair(buf, utf8_output); 321| 9.93k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128avx2_convert_utf32_to_latin1EPKDimPc: 3| 5.26k| char *latin1_output) { 4| 5.26k| const size_t rounded_len = 5| 5.26k| len & ~0x1F; // Round down to nearest multiple of 32 6| | 7| 5.26k| const __m256i high_bytes_mask = _mm256_set1_epi32(0xFFFFFF00); 8| | 9| 9.84k| for (size_t i = 0; i < rounded_len; i += 4 * 8) { ------------------ | Branch (9:22): [True: 4.58k, False: 5.26k] ------------------ 10| 4.58k| __m256i a = _mm256_loadu_si256((__m256i *)(buf + 0 * 8)); 11| 4.58k| __m256i b = _mm256_loadu_si256((__m256i *)(buf + 1 * 8)); 12| 4.58k| __m256i c = _mm256_loadu_si256((__m256i *)(buf + 2 * 8)); 13| 4.58k| __m256i d = _mm256_loadu_si256((__m256i *)(buf + 3 * 8)); 14| | 15| 4.58k| const __m256i check_combined = 16| 4.58k| _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d)); 17| | 18| 4.58k| if (!_mm256_testz_si256(check_combined, high_bytes_mask)) { ------------------ | Branch (18:9): [True: 0, False: 4.58k] ------------------ 19| 0| return std::make_pair(nullptr, latin1_output); 20| 0| } 21| | 22| 4.58k| b = _mm256_slli_epi32(b, 1 * 8); 23| 4.58k| c = _mm256_slli_epi32(c, 2 * 8); 24| 4.58k| 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| 4.58k| const __m256i t0 = 35| 4.58k| _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| 4.58k| const __m256i shuffle_bytes = 40| 4.58k| _mm256_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15, 41| 4.58k| 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15); 42| | 43| 4.58k| 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| 4.58k| const __m256i shuffle_dwords = _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7); 48| 4.58k| const __m256i t2 = _mm256_permutevar8x32_epi32(t1, shuffle_dwords); 49| |// clang format on 50| | 51| 4.58k| _mm256_storeu_si256((__m256i *)latin1_output, t2); 52| | 53| 4.58k| latin1_output += 32; 54| 4.58k| buf += 32; 55| 4.58k| } 56| | 57| 5.26k| return std::make_pair(buf, latin1_output); 58| 5.26k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_126avx2_convert_utf32_to_utf8EPKDimPc: 2| 4.66k|avx2_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) { 3| 4.66k| const char32_t *end = buf + len; 4| 4.66k| const __m256i v_0000 = _mm256_setzero_si256(); 5| 4.66k| const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000); 6| 4.66k| const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80); 7| 4.66k| const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800); 8| 4.66k| const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080); 9| 4.66k| const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff); 10| 4.66k| __m256i running_max = _mm256_setzero_si256(); 11| 4.66k| __m256i forbidden_bytemask = _mm256_setzero_si256(); 12| | 13| 4.66k| const size_t safety_margin = 14| 4.66k| 12; // to avoid overruns, see issue 15| | // https://github.com/simdutf/simdutf/issues/92 16| | 17| 8.02k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (17:10): [True: 3.35k, False: 4.66k] ------------------ 18| 3.35k| __m256i in = _mm256_loadu_si256((__m256i *)buf); 19| 3.35k| __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1); 20| 3.35k| 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| 3.35k| __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff), 25| 3.35k| _mm256_and_si256(nextin, v_7fffffff)); 26| 3.35k| 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| 3.35k| if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (31:9): [True: 2.26k, False: 1.08k] ------------------ 32| | // 1. pack the bytes 33| 2.26k| const __m128i utf8_packed = _mm_packus_epi16( 34| 2.26k| _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1)); 35| | // 2. store (16 bytes) 36| 2.26k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 37| | // 3. adjust pointers 38| 2.26k| buf += 16; 39| 2.26k| utf8_output += 16; 40| 2.26k| continue; // we are done for this round! 41| 2.26k| } 42| | // no bits set above 7th bit 43| 1.08k| const __m256i one_byte_bytemask = 44| 1.08k| _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000); 45| 1.08k| const uint32_t one_byte_bitmask = 46| 1.08k| static_cast(_mm256_movemask_epi8(one_byte_bytemask)); 47| | 48| | // no bits set above 11th bit 49| 1.08k| const __m256i one_or_two_bytes_bytemask = 50| 1.08k| _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000); 51| 1.08k| const uint32_t one_or_two_bytes_bitmask = 52| 1.08k| static_cast(_mm256_movemask_epi8(one_or_two_bytes_bytemask)); 53| 1.08k| if (one_or_two_bytes_bitmask == 0xffffffff) { ------------------ | Branch (53:9): [True: 153, False: 932] ------------------ 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| 153| const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00); 58| 153| const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f); 59| | 60| | // t0 = [000a|aaaa|bbbb|bb00] 61| 153| const __m256i t0 = _mm256_slli_epi16(in_16, 2); 62| | // t1 = [000a|aaaa|0000|0000] 63| 153| const __m256i t1 = _mm256_and_si256(t0, v_1f00); 64| | // t2 = [0000|0000|00bb|bbbb] 65| 153| const __m256i t2 = _mm256_and_si256(in_16, v_003f); 66| | // t3 = [000a|aaaa|00bb|bbbb] 67| 153| const __m256i t3 = _mm256_or_si256(t1, t2); 68| | // t4 = [110a|aaaa|10bb|bbbb] 69| 153| const __m256i t4 = _mm256_or_si256(t3, v_c080); 70| | 71| | // 2. merge ASCII and 2-byte codewords 72| 153| const __m256i utf8_unpacked = 73| 153| _mm256_blendv_epi8(t4, in_16, one_byte_bytemask); 74| | 75| | // 3. prepare bitmask for 8-bit lookup 76| 153| const uint32_t M0 = one_byte_bitmask & 0x55555555; 77| 153| const uint32_t M1 = M0 >> 7; 78| 153| const uint32_t M2 = (M1 | M0) & 0x00ff00ff; 79| | // 4. pack the bytes 80| | 81| 153| const uint8_t *row = 82| 153| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0]; 83| 153| const uint8_t *row_2 = 84| 153| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 85| 153| 16)][0]; 86| | 87| 153| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 88| 153| const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1)); 89| | 90| 153| const __m256i utf8_packed = _mm256_shuffle_epi8( 91| 153| utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2)); ------------------ | | 7| 153| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 8| 153| _mm256_castsi128_si256(xmm2), 2) ------------------ 92| | // 5. store bytes 93| 153| _mm_storeu_si128((__m128i *)utf8_output, 94| 153| _mm256_castsi256_si128(utf8_packed)); 95| 153| utf8_output += row[0]; 96| 153| _mm_storeu_si128((__m128i *)utf8_output, 97| 153| _mm256_extractf128_si256(utf8_packed, 1)); 98| 153| utf8_output += row_2[0]; 99| | 100| | // 6. adjust pointers 101| 153| buf += 16; 102| 153| continue; 103| 153| } 104| | // Must check for overflow in packing 105| 932| const __m256i saturation_bytemask = _mm256_cmpeq_epi32( 106| 932| _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000); 107| 932| const uint32_t saturation_bitmask = 108| 932| static_cast(_mm256_movemask_epi8(saturation_bytemask)); 109| 932| if (saturation_bitmask == 0xffffffff) { ------------------ | Branch (109:9): [True: 124, False: 808] ------------------ 110| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 111| 124| const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800); 112| 124| forbidden_bytemask = _mm256_or_si256( 113| 124| forbidden_bytemask, 114| 124| _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800)); 115| | 116| 124| const __m256i dup_even = _mm256_setr_epi16( 117| 124| 0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e, 118| 124| 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| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) 148| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 149| 124| const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even); 150| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 151| 124| const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 147| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 152| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 153| 124| const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 147| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 154| | 155| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 156| 124| const __m256i s0 = _mm256_srli_epi16(in_16, 4); 157| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 158| 124| const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 147| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 159| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 160| 124| const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 147| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 161| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 162| 124| const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 147| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 163| 124| const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, 164| 124| simdutf_vec(0b0100000000000000)); ------------------ | | 147| 124|#define simdutf_vec(x) _mm256_set1_epi16(static_cast(x)) ------------------ 165| 124| const __m256i s4 = _mm256_xor_si256(s3, m0); 166| 124|#undef simdutf_vec 167| | 168| | // 4. expand code units 16-bit => 32-bit 169| 124| const __m256i out0 = _mm256_unpacklo_epi16(t2, s4); 170| 124| 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| 124| const uint32_t mask = (one_byte_bitmask & 0x55555555) | 174| 124| (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| 124| const uint8_t mask0 = uint8_t(mask); 195| 124| const uint8_t *row0 = 196| 124| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 197| 124| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 198| 124| const __m128i utf8_0 = 199| 124| _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0); 200| | 201| 124| const uint8_t mask1 = static_cast(mask >> 8); 202| 124| const uint8_t *row1 = 203| 124| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 204| 124| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 205| 124| const __m128i utf8_1 = 206| 124| _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1); 207| | 208| 124| const uint8_t mask2 = static_cast(mask >> 16); 209| 124| const uint8_t *row2 = 210| 124| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0]; 211| 124| const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1)); 212| 124| const __m128i utf8_2 = 213| 124| _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2); 214| | 215| 124| const uint8_t mask3 = static_cast(mask >> 24); 216| 124| const uint8_t *row3 = 217| 124| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0]; 218| 124| const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1)); 219| 124| const __m128i utf8_3 = 220| 124| _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3); 221| | 222| 124| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 223| 124| utf8_output += row0[0]; 224| 124| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 225| 124| utf8_output += row1[0]; 226| 124| _mm_storeu_si128((__m128i *)utf8_output, utf8_2); 227| 124| utf8_output += row2[0]; 228| 124| _mm_storeu_si128((__m128i *)utf8_output, utf8_3); 229| 124| utf8_output += row3[0]; 230| 124| buf += 16; 231| 808| } 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| 808| size_t forward = 15; 237| 808| size_t k = 0; 238| 808| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (238:11): [True: 0, False: 808] ------------------ 239| 0| forward = size_t(end - buf - 1); 240| 0| } 241| 12.9k| for (; k < forward; k++) { ------------------ | Branch (241:14): [True: 12.1k, False: 808] ------------------ 242| 12.1k| uint32_t word = buf[k]; 243| 12.1k| if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII) ------------------ | Branch (243:13): [True: 9.74k, False: 2.37k] ------------------ 244| 9.74k| *utf8_output++ = char(word); 245| 9.74k| } else if ((word & 0xFFFFF800) == 0) { // 2-byte ------------------ | Branch (245:20): [True: 480, False: 1.89k] ------------------ 246| 480| *utf8_output++ = char((word >> 6) | 0b11000000); 247| 480| *utf8_output++ = char((word & 0b111111) | 0b10000000); 248| 1.89k| } else if ((word & 0xFFFF0000) == 0) { // 3-byte ------------------ | Branch (248:20): [True: 636, False: 1.25k] ------------------ 249| 636| if (word >= 0xD800 && word <= 0xDFFF) { ------------------ | Branch (249:15): [True: 311, False: 325] | Branch (249:33): [True: 0, False: 311] ------------------ 250| 0| return std::make_pair(nullptr, utf8_output); 251| 0| } 252| 636| *utf8_output++ = char((word >> 12) | 0b11100000); 253| 636| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 254| 636| *utf8_output++ = char((word & 0b111111) | 0b10000000); 255| 1.25k| } else { // 4-byte 256| 1.25k| if (word > 0x10FFFF) { ------------------ | Branch (256:15): [True: 0, False: 1.25k] ------------------ 257| 0| return std::make_pair(nullptr, utf8_output); 258| 0| } 259| 1.25k| *utf8_output++ = char((word >> 18) | 0b11110000); 260| 1.25k| *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000); 261| 1.25k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 262| 1.25k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 263| 1.25k| } 264| 12.1k| } 265| 808| buf += k; 266| 808| } 267| 932| } // while 268| | 269| | // check for invalid input 270| 4.66k| const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff); 271| 4.66k| if (static_cast(_mm256_movemask_epi8(_mm256_cmpeq_epi32( ------------------ | Branch (271:7): [True: 0, False: 4.66k] ------------------ 272| 4.66k| _mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) { 273| 0| return std::make_pair(nullptr, utf8_output); 274| 0| } 275| | 276| 4.66k| if (static_cast(_mm256_movemask_epi8(forbidden_bytemask)) != 0) { ------------------ | Branch (276:7): [True: 0, False: 4.66k] ------------------ 277| 0| return std::make_pair(nullptr, utf8_output); 278| 0| } 279| | 280| 4.66k| return std::make_pair(buf, utf8_output); 281| 4.66k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_129convert_masked_utf8_to_latin1EPKcmRPc: 9| 16.5k| 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| 16.5k| const __m128i in = _mm_loadu_si128((__m128i *)input); 21| | 22| 16.5k| const uint16_t input_utf8_end_of_code_point_mask = 23| 16.5k| utf8_end_of_code_point_mask & 24| 16.5k| 0xfff; // we are only processing 12 bytes in case it is not all ASCII 25| | 26| 16.5k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (26:7): [True: 124, False: 16.4k] ------------------ 27| | // We process the data in chunks of 12 bytes. 28| 124| _mm_storeu_si128(reinterpret_cast<__m128i *>(latin1_output), in); 29| 124| latin1_output += 12; // We wrote 12 characters. 30| 124| return 12; // We consumed 1 bytes. 31| 124| } 32| | /// We do not have a fast path available, so we fallback. 33| 16.4k| const uint8_t idx = 34| 16.4k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 35| 16.4k| const uint8_t consumed = 36| 16.4k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 37| | // this indicates an invalid input: 38| 16.4k| if (idx >= 64) { ------------------ | Branch (38:7): [True: 5.22k, False: 11.1k] ------------------ 39| 5.22k| return consumed; 40| 5.22k| } 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| 11.1k| const __m128i sh = 48| 11.1k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 49| 11.1k| const __m128i perm = _mm_shuffle_epi8(in, sh); 50| 11.1k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 51| 11.1k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 52| 11.1k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 53| 11.1k| 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| 11.1k| _mm_storel_epi64((__m128i *)latin1_output, latin1_packed); 58| 11.1k| latin1_output += 6; // We wrote 6 bytes. 59| 11.1k| return consumed; 60| 16.4k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE0EEEmPKcmRPDs: 10| 15.1k| 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| 15.1k| const __m128i swap = 23| 15.1k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 15.1k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 15.1k| const uint16_t input_utf8_end_of_code_point_mask = 26| 15.1k| utf8_end_of_code_point_mask & 0xfff; 27| 15.1k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 95, False: 15.0k] ------------------ 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): [Folded, False: 95] ------------------ 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| 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| 15.0k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (40:7): [True: 142, False: 14.9k] ------------------ 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| 142| const __m128i sh = 45| 142| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 46| 142| const __m128i perm = _mm_shuffle_epi8(in, sh); 47| 142| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 48| 142| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 49| 142| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 50| 142| if (big_endian) ------------------ | Branch (50:9): [Folded, False: 142] ------------------ 51| 0| composed = _mm_shuffle_epi8(composed, swap); 52| 142| _mm_storeu_si128((__m128i *)utf16_output, composed); 53| 142| utf16_output += 8; // We wrote 16 bytes, 8 code points. 54| 142| return 16; 55| 142| } 56| 14.9k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (56:7): [True: 3.51k, False: 11.4k] ------------------ 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| 3.51k| const __m128i sh = 61| 3.51k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 62| 3.51k| const __m128i perm = _mm_shuffle_epi8(in, sh); 63| 3.51k| const __m128i ascii = 64| 3.51k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 65| 3.51k| const __m128i middlebyte = 66| 3.51k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 67| 3.51k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 68| 3.51k| const __m128i highbyte = 69| 3.51k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 70| 3.51k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 71| 3.51k| const __m128i composed = 72| 3.51k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 73| 3.51k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 74| 3.51k| if (big_endian) ------------------ | Branch (74:9): [Folded, False: 3.51k] ------------------ 75| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 76| 3.51k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 77| 3.51k| utf16_output += 4; 78| 3.51k| return 12; 79| 3.51k| } 80| | 81| 11.4k| const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex 82| 11.4k| [input_utf8_end_of_code_point_mask][0]; 83| 11.4k| const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex 84| 11.4k| [input_utf8_end_of_code_point_mask][1]; 85| 11.4k| if (idx < 64) { ------------------ | Branch (85:7): [True: 4.65k, False: 6.76k] ------------------ 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| 4.65k| const __m128i sh = _mm_loadu_si128( 93| 4.65k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 94| 4.65k| const __m128i perm = _mm_shuffle_epi8(in, sh); 95| 4.65k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 96| 4.65k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 97| 4.65k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 98| 4.65k| if (big_endian) ------------------ | Branch (98:9): [Folded, False: 4.65k] ------------------ 99| 0| composed = _mm_shuffle_epi8(composed, swap); 100| 4.65k| _mm_storeu_si128((__m128i *)utf16_output, composed); 101| 4.65k| utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential 102| | // overflow of 4 bytes. 103| 6.76k| } else if (idx < 145) { ------------------ | Branch (103:14): [True: 1.69k, False: 5.06k] ------------------ 104| | // FOUR (4) input code-code units 105| 1.69k| const __m128i sh = _mm_loadu_si128( 106| 1.69k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 107| 1.69k| const __m128i perm = _mm_shuffle_epi8(in, sh); 108| 1.69k| const __m128i ascii = 109| 1.69k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 110| 1.69k| const __m128i middlebyte = 111| 1.69k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 112| 1.69k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 113| 1.69k| const __m128i highbyte = 114| 1.69k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 115| 1.69k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 116| 1.69k| const __m128i composed = 117| 1.69k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 118| 1.69k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 119| 1.69k| if (big_endian) ------------------ | Branch (119:9): [Folded, False: 1.69k] ------------------ 120| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 121| 1.69k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 122| 1.69k| utf16_output += 4; // Here we overflow by 8 bytes. 123| 5.06k| } else if (idx < 209) { ------------------ | Branch (123:14): [True: 4.86k, False: 200] ------------------ 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| 4.86k| const __m128i sh = _mm_loadu_si128( 135| 4.86k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 136| 4.86k| const __m128i perm = _mm_shuffle_epi8(in, sh); 137| 4.86k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 138| 4.86k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 139| 4.86k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 140| 4.86k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 141| | // correct for spurious high bit 142| 4.86k| const __m128i correct = 143| 4.86k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 144| 4.86k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 145| 4.86k| 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| 4.86k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 149| 4.86k| 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| 4.86k| const __m128i composed = 155| 4.86k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 156| 4.86k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 157| 4.86k| const __m128i composedminus = 158| 4.86k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 159| 4.86k| const __m128i lowtenbits = 160| 4.86k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 161| | // Notice the 0x3ff mask: 162| 4.86k| const __m128i hightenbits = 163| 4.86k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 164| 4.86k| const __m128i lowtenbitsadd = 165| 4.86k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 166| 4.86k| const __m128i hightenbitsadd = 167| 4.86k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 168| 4.86k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 169| 4.86k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 170| 4.86k| uint32_t basic_buffer[4]; 171| 4.86k| uint32_t basic_buffer_swap[4]; 172| 4.86k| if (big_endian) { ------------------ | Branch (172:9): [Folded, False: 4.86k] ------------------ 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| 4.86k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 178| 4.86k| uint32_t surrogate_buffer[4]; 179| 4.86k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 180| 19.4k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (180:24): [True: 14.6k, False: 4.86k] ------------------ 181| 14.6k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (181:11): [True: 3.00k, False: 11.5k] ------------------ 182| 3.00k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 183| 3.00k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 184| 3.00k| utf16_output += 2; 185| 11.5k| } else { 186| 11.5k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (186:27): [Folded, False: 11.5k] ------------------ 187| 11.5k| : uint16_t(basic_buffer[i]); 188| 11.5k| utf16_output++; 189| 11.5k| } 190| 14.6k| } 191| 4.86k| } else { 192| | // here we know that there is an error but we do not handle errors 193| 200| } 194| 11.4k| return consumed; 195| 14.9k|} 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: 96, False: 15.5k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| 96| __m256i ascii = _mm256_cvtepu8_epi16(in); 30| 96| if (big_endian) { ------------------ | Branch (30:9): [True: 96, Folded] ------------------ 31| 96| const __m256i swap256 = _mm256_setr_epi8( 32| 96| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 33| 96| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 34| 96| ascii = _mm256_shuffle_epi8(ascii, swap256); 35| 96| } 36| 96| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output), ascii); 37| 96| utf16_output += 12; // We wrote 12 16-bit characters. 38| 96| return 12; // We consumed 12 bytes. 39| 96| } 40| 15.5k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (40:7): [True: 121, 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| 121| const __m128i sh = 45| 121| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 46| 121| const __m128i perm = _mm_shuffle_epi8(in, sh); 47| 121| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 48| 121| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 49| 121| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 50| 121| if (big_endian) ------------------ | Branch (50:9): [True: 121, Folded] ------------------ 51| 121| composed = _mm_shuffle_epi8(composed, swap); 52| 121| _mm_storeu_si128((__m128i *)utf16_output, composed); 53| 121| utf16_output += 8; // We wrote 16 bytes, 8 code points. 54| 121| return 16; 55| 121| } 56| 15.3k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (56:7): [True: 4.01k, False: 11.3k] ------------------ 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| 4.01k| const __m128i sh = 61| 4.01k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 62| 4.01k| const __m128i perm = _mm_shuffle_epi8(in, sh); 63| 4.01k| const __m128i ascii = 64| 4.01k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 65| 4.01k| const __m128i middlebyte = 66| 4.01k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 67| 4.01k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 68| 4.01k| const __m128i highbyte = 69| 4.01k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 70| 4.01k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 71| 4.01k| const __m128i composed = 72| 4.01k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 73| 4.01k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 74| 4.01k| if (big_endian) ------------------ | Branch (74:9): [True: 4.01k, Folded] ------------------ 75| 4.01k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 76| 4.01k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 77| 4.01k| utf16_output += 4; 78| 4.01k| return 12; 79| 4.01k| } 80| | 81| 11.3k| const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex 82| 11.3k| [input_utf8_end_of_code_point_mask][0]; 83| 11.3k| const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex 84| 11.3k| [input_utf8_end_of_code_point_mask][1]; 85| 11.3k| if (idx < 64) { ------------------ | Branch (85:7): [True: 4.66k, False: 6.71k] ------------------ 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| 4.66k| const __m128i sh = _mm_loadu_si128( 93| 4.66k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 94| 4.66k| const __m128i perm = _mm_shuffle_epi8(in, sh); 95| 4.66k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 96| 4.66k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 97| 4.66k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 98| 4.66k| if (big_endian) ------------------ | Branch (98:9): [True: 4.66k, Folded] ------------------ 99| 4.66k| composed = _mm_shuffle_epi8(composed, swap); 100| 4.66k| _mm_storeu_si128((__m128i *)utf16_output, composed); 101| 4.66k| utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential 102| | // overflow of 4 bytes. 103| 6.71k| } else if (idx < 145) { ------------------ | Branch (103:14): [True: 1.70k, False: 5.00k] ------------------ 104| | // FOUR (4) input code-code units 105| 1.70k| const __m128i sh = _mm_loadu_si128( 106| 1.70k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 107| 1.70k| const __m128i perm = _mm_shuffle_epi8(in, sh); 108| 1.70k| const __m128i ascii = 109| 1.70k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 110| 1.70k| const __m128i middlebyte = 111| 1.70k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 112| 1.70k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 113| 1.70k| const __m128i highbyte = 114| 1.70k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 115| 1.70k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 116| 1.70k| const __m128i composed = 117| 1.70k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 118| 1.70k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 119| 1.70k| if (big_endian) ------------------ | Branch (119:9): [True: 1.70k, Folded] ------------------ 120| 1.70k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 121| 1.70k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 122| 1.70k| utf16_output += 4; // Here we overflow by 8 bytes. 123| 5.00k| } else if (idx < 209) { ------------------ | Branch (123:14): [True: 4.80k, False: 200] ------------------ 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| 4.80k| const __m128i sh = _mm_loadu_si128( 135| 4.80k| (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]); 136| 4.80k| const __m128i perm = _mm_shuffle_epi8(in, sh); 137| 4.80k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 138| 4.80k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 139| 4.80k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 140| 4.80k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 141| | // correct for spurious high bit 142| 4.80k| const __m128i correct = 143| 4.80k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 144| 4.80k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 145| 4.80k| 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| 4.80k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 149| 4.80k| 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| 4.80k| const __m128i composed = 155| 4.80k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 156| 4.80k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 157| 4.80k| const __m128i composedminus = 158| 4.80k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 159| 4.80k| const __m128i lowtenbits = 160| 4.80k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 161| | // Notice the 0x3ff mask: 162| 4.80k| const __m128i hightenbits = 163| 4.80k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 164| 4.80k| const __m128i lowtenbitsadd = 165| 4.80k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 166| 4.80k| const __m128i hightenbitsadd = 167| 4.80k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 168| 4.80k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 169| 4.80k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 170| 4.80k| uint32_t basic_buffer[4]; 171| 4.80k| uint32_t basic_buffer_swap[4]; 172| 4.80k| if (big_endian) { ------------------ | Branch (172:9): [True: 4.80k, Folded] ------------------ 173| 4.80k| _mm_storeu_si128((__m128i *)basic_buffer_swap, 174| 4.80k| _mm_shuffle_epi8(composed, swap)); 175| 4.80k| surrogates = _mm_shuffle_epi8(surrogates, swap); 176| 4.80k| } 177| 4.80k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 178| 4.80k| uint32_t surrogate_buffer[4]; 179| 4.80k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 180| 19.2k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (180:24): [True: 14.4k, False: 4.80k] ------------------ 181| 14.4k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (181:11): [True: 2.97k, False: 11.4k] ------------------ 182| 2.97k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 183| 2.97k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 184| 2.97k| utf16_output += 2; 185| 11.4k| } else { 186| 11.4k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (186:27): [True: 11.4k, Folded] ------------------ 187| 11.4k| : uint16_t(basic_buffer[i]); 188| 11.4k| utf16_output++; 189| 11.4k| } 190| 14.4k| } 191| 4.80k| } else { 192| | // here we know that there is an error but we do not handle errors 193| 200| } 194| 11.3k| return consumed; 195| 15.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_128convert_masked_utf8_to_utf32EPKcmRPDi: 9| 10.6k| 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| 10.6k| const __m128i in = _mm_loadu_si128((__m128i *)input); 22| 10.6k| const uint16_t input_utf8_end_of_code_point_mask = 23| 10.6k| utf8_end_of_code_point_mask & 0xfff; 24| 10.6k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (24:7): [True: 95, False: 10.5k] ------------------ 25| | // We process the data in chunks of 12 bytes. 26| 95| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output), 27| 95| _mm256_cvtepu8_epi32(in)); 28| 95| _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output + 8), 29| 95| _mm256_cvtepu8_epi32(_mm_srli_si128(in, 8))); 30| 95| utf32_output += 12; // We wrote 12 32-bit characters. 31| 95| return 12; // We consumed 12 bytes. 32| 95| } 33| 10.5k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (33:7): [True: 92, False: 10.4k] ------------------ 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| 92| const __m128i sh = 38| 92| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 39| 92| const __m128i perm = _mm_shuffle_epi8(in, sh); 40| 92| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 41| 92| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 42| 92| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 43| 92| _mm256_storeu_si256((__m256i *)utf32_output, 44| 92| _mm256_cvtepu16_epi32(composed)); 45| 92| utf32_output += 8; // We wrote 16 bytes, 8 code points. 46| 92| return 16; 47| 92| } 48| 10.4k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (48:7): [True: 203, False: 10.2k] ------------------ 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| 203| const __m128i sh = 53| 203| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 54| 203| const __m128i perm = _mm_shuffle_epi8(in, sh); 55| 203| const __m128i ascii = 56| 203| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 57| 203| const __m128i middlebyte = 58| 203| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 59| 203| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 60| 203| const __m128i highbyte = 61| 203| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 62| 203| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 63| 203| const __m128i composed = 64| 203| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 65| 203| _mm_storeu_si128((__m128i *)utf32_output, composed); 66| 203| utf32_output += 4; 67| 203| return 12; 68| 203| } 69| | /// We do not have a fast path available, so we fallback. 70| | 71| 10.2k| const uint8_t idx = 72| 10.2k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 73| 10.2k| const uint8_t consumed = 74| 10.2k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 75| 10.2k| if (idx < 64) { ------------------ | Branch (75:7): [True: 4.53k, False: 5.75k] ------------------ 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| 4.53k| const __m128i sh = 83| 4.53k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 84| 4.53k| const __m128i perm = _mm_shuffle_epi8(in, sh); 85| 4.53k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 86| 4.53k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 87| 4.53k| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 88| 4.53k| _mm256_storeu_si256((__m256i *)utf32_output, 89| 4.53k| _mm256_cvtepu16_epi32(composed)); 90| 4.53k| utf32_output += 6; // We wrote 24 bytes, 6 code points. There is a potential 91| | // overflow of 32 - 24 = 8 bytes. 92| 5.75k| } else if (idx < 145) { ------------------ | Branch (92:14): [True: 1.19k, False: 4.56k] ------------------ 93| | // FOUR (4) input code-code units 94| 1.19k| const __m128i sh = 95| 1.19k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 96| 1.19k| const __m128i perm = _mm_shuffle_epi8(in, sh); 97| 1.19k| const __m128i ascii = 98| 1.19k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 99| 1.19k| const __m128i middlebyte = 100| 1.19k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 101| 1.19k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 102| 1.19k| const __m128i highbyte = 103| 1.19k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 104| 1.19k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 105| 1.19k| const __m128i composed = 106| 1.19k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 107| 1.19k| _mm_storeu_si128((__m128i *)utf32_output, composed); 108| 1.19k| utf32_output += 4; 109| 4.56k| } else if (idx < 209) { ------------------ | Branch (109:14): [True: 4.36k, False: 200] ------------------ 110| | // TWO (2) input code-code units 111| 4.36k| const __m128i sh = 112| 4.36k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 113| 4.36k| const __m128i perm = _mm_shuffle_epi8(in, sh); 114| 4.36k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 115| 4.36k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 116| 4.36k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 117| 4.36k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 118| | // correct for spurious high bit 119| 4.36k| const __m128i correct = 120| 4.36k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 121| 4.36k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 122| 4.36k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 123| 4.36k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000)); 124| 4.36k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 125| 4.36k| const __m128i composed = 126| 4.36k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 127| 4.36k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 128| 4.36k| _mm_storeu_si128((__m128i *)utf32_output, composed); 129| 4.36k| utf32_output += 130| 4.36k| 3; // We wrote 3 * 4 bytes, there is a potential overflow of 4 bytes. 131| 4.36k| } else { 132| | // here we know that there is an error but we do not handle errors 133| 200| } 134| 10.2k| return consumed; 135| 10.4k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1623utf16_gather_high_bytesILNS_10endiannessE0EEENS1_4simd5simd8IhEERKNS5_6simd16ItEESB_: 3| 11.8k| const simd16 &in1) { 4| 11.8k| if (big_endian) { ------------------ | Branch (4:7): [Folded, False: 11.8k] ------------------ 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| 11.8k| } else { 12| 11.8k| const auto t0 = in0.shr<8>(); 13| 11.8k| const auto t1 = in1.shr<8>(); 14| | 15| 11.8k| return simd16::pack(t0, t1); 16| 11.8k| } 17| 11.8k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_15utf1623utf16_gather_high_bytesILNS_10endiannessE1EEENS1_4simd5simd8IhEERKNS5_6simd16ItEESB_: 3| 12.0k| const simd16 &in1) { 4| 12.0k| if (big_endian) { ------------------ | Branch (4:7): [True: 12.0k, Folded] ------------------ 5| | // we want lower bytes 6| 12.0k| const auto mask = simd16(0x00ff); 7| 12.0k| const auto t0 = in0 & mask; 8| 12.0k| const auto t1 = in1 & mask; 9| | 10| 12.0k| return simd16::pack(t0, t1); 11| 12.0k| } 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| 12.0k|} _ZNK7simdutf7haswell14implementation13validate_utf8EPKcm: 266| 20.8k|implementation::validate_utf8(const char *buf, size_t len) const noexcept { 267| 20.8k| return haswell::utf8_validation::generic_validate_utf8(buf, len); 268| 20.8k|} _ZNK7simdutf7haswell14implementation25validate_utf8_with_errorsEPKcm: 273| 5.26k| const char *buf, size_t len) const noexcept { 274| 5.26k| return haswell::utf8_validation::generic_validate_utf8_with_errors(buf, len); 275| 5.26k|} _ZNK7simdutf7haswell14implementation16validate_utf16leEPKDsm: 312| 15.2k| size_t len) const noexcept { 313| 15.2k| if (simdutf_unlikely(len == 0)) { ------------------ | | 93| 15.2k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 12.9k, False: 2.23k] | | ------------------ ------------------ 314| | // empty input is valid UTF-16. protect the implementation from 315| | // handling nullptr 316| 12.9k| return true; 317| 12.9k| } 318| 2.23k| const auto res = 319| 2.23k| haswell::utf16::validate_utf16_with_errors(buf, len); 320| 2.23k| if (res.is_err()) { ------------------ | Branch (320:7): [True: 47, False: 2.19k] ------------------ 321| 47| return false; 322| 47| } 323| | 324| 2.19k| if (res.count == len) { ------------------ | Branch (324:7): [True: 0, False: 2.19k] ------------------ 325| 0| return true; 326| 0| } 327| | 328| 2.19k| return scalar::utf16::validate(buf + res.count, 329| 2.19k| len - res.count); 330| 2.19k|} _ZNK7simdutf7haswell14implementation16validate_utf16beEPKDsm: 336| 15.2k| size_t len) const noexcept { 337| 15.2k| if (simdutf_unlikely(len == 0)) { ------------------ | | 93| 15.2k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 12.9k, False: 2.24k] | | ------------------ ------------------ 338| | // empty input is valid UTF-16. protect the implementation from 339| | // handling nullptr 340| 12.9k| return true; 341| 12.9k| } 342| 2.24k| const auto res = 343| 2.24k| haswell::utf16::validate_utf16_with_errors(buf, len); 344| 2.24k| if (res.is_err()) { ------------------ | Branch (344:7): [True: 43, False: 2.19k] ------------------ 345| 43| return false; 346| 43| } 347| | 348| 2.19k| if (res.count == len) { ------------------ | Branch (348:7): [True: 0, False: 2.19k] ------------------ 349| 0| return true; 350| 0| } 351| | 352| 2.19k| return scalar::utf16::validate(buf + res.count, 353| 2.19k| len - res.count); 354| 2.19k|} _ZNK7simdutf7haswell14implementation28validate_utf16le_with_errorsEPKDsm: 357| 5.26k| const char16_t *buf, size_t len) const noexcept { 358| | 359| 5.26k| const result res = 360| 5.26k| haswell::utf16::validate_utf16_with_errors(buf, len); 361| 5.26k| if (res.count != len) { ------------------ | Branch (361:7): [True: 349, False: 4.91k] ------------------ 362| 349| const result scalar_res = 363| 349| scalar::utf16::validate_with_errors( 364| 349| buf + res.count, len - res.count); 365| 349| return result(scalar_res.error, res.count + scalar_res.count); 366| 4.91k| } else { 367| 4.91k| return res; 368| 4.91k| } 369| 5.26k|} _ZNK7simdutf7haswell14implementation28validate_utf16be_with_errorsEPKDsm: 372| 5.26k| const char16_t *buf, size_t len) const noexcept { 373| 5.26k| const result res = 374| 5.26k| haswell::utf16::validate_utf16_with_errors(buf, len); 375| 5.26k| if (res.count != len) { ------------------ | Branch (375:7): [True: 354, False: 4.91k] ------------------ 376| 354| const result scalar_res = 377| 354| scalar::utf16::validate_with_errors(buf + res.count, 378| 354| len - res.count); 379| 354| return result(scalar_res.error, res.count + scalar_res.count); 380| 4.91k| } else { 381| 4.91k| return res; 382| 4.91k| } 383| 5.26k|} _ZNK7simdutf7haswell14implementation14validate_utf32EPKDim: 398| 9.93k|implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept { 399| 9.93k| return utf32::validate(buf, len); 400| 9.93k|} _ZNK7simdutf7haswell14implementation22convert_latin1_to_utf8EPKcmPc: 412| 5.53k| const char *buf, size_t len, char *utf8_output) const noexcept { 413| 5.53k| std::pair ret = 414| 5.53k| avx2_convert_latin1_to_utf8(buf, len, utf8_output); 415| 5.53k| size_t converted_chars = ret.second - utf8_output; 416| | 417| 5.53k| if (ret.first != buf + len) { ------------------ | Branch (417:7): [True: 1.36k, False: 4.16k] ------------------ 418| 1.36k| const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert( 419| 1.36k| ret.first, len - (ret.first - buf), ret.second); 420| 1.36k| converted_chars += scalar_converted_chars; 421| 1.36k| } 422| | 423| 5.53k| return converted_chars; 424| 5.53k|} _ZNK7simdutf7haswell14implementation25convert_latin1_to_utf16leEPKcmPDs: 429| 5.26k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 430| 5.26k| std::pair ret = 431| 5.26k| avx2_convert_latin1_to_utf16(buf, len, utf16_output); 432| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (432:7): [True: 0, False: 5.26k] ------------------ 433| 0| return 0; 434| 0| } 435| 5.26k| size_t converted_chars = ret.second - utf16_output; 436| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (436:7): [True: 787, False: 4.48k] ------------------ 437| 787| const size_t scalar_converted_chars = 438| 787| scalar::latin1_to_utf16::convert( 439| 787| ret.first, len - (ret.first - buf), ret.second); 440| 787| if (scalar_converted_chars == 0) { ------------------ | Branch (440:9): [True: 0, False: 787] ------------------ 441| 0| return 0; 442| 0| } 443| 787| converted_chars += scalar_converted_chars; 444| 787| } 445| 5.26k| return converted_chars; 446| 5.26k|} _ZNK7simdutf7haswell14implementation25convert_latin1_to_utf16beEPKcmPDs: 449| 5.26k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 450| 5.26k| std::pair ret = 451| 5.26k| avx2_convert_latin1_to_utf16(buf, len, utf16_output); 452| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (452:7): [True: 0, False: 5.26k] ------------------ 453| 0| return 0; 454| 0| } 455| 5.26k| size_t converted_chars = ret.second - utf16_output; 456| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (456:7): [True: 787, False: 4.48k] ------------------ 457| 787| const size_t scalar_converted_chars = 458| 787| scalar::latin1_to_utf16::convert( 459| 787| ret.first, len - (ret.first - buf), ret.second); 460| 787| if (scalar_converted_chars == 0) { ------------------ | Branch (460:9): [True: 0, False: 787] ------------------ 461| 0| return 0; 462| 0| } 463| 787| converted_chars += scalar_converted_chars; 464| 787| } 465| 5.26k| return converted_chars; 466| 5.26k|} _ZNK7simdutf7haswell14implementation23convert_latin1_to_utf32EPKcmPDi: 471| 5.26k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 472| 5.26k| std::pair ret = 473| 5.26k| avx2_convert_latin1_to_utf32(buf, len, utf32_output); 474| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (474:7): [True: 0, False: 5.26k] ------------------ 475| 0| return 0; 476| 0| } 477| 5.26k| size_t converted_chars = ret.second - utf32_output; 478| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (478:7): [True: 746, False: 4.52k] ------------------ 479| 746| const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert( 480| 746| ret.first, len - (ret.first - buf), ret.second); 481| 746| if (scalar_converted_chars == 0) { ------------------ | Branch (481:9): [True: 0, False: 746] ------------------ 482| 0| return 0; 483| 0| } 484| 746| converted_chars += scalar_converted_chars; 485| 746| } 486| 5.26k| return converted_chars; 487| 5.26k|} _ZNK7simdutf7haswell14implementation22convert_utf8_to_latin1EPKcmPc: 492| 10.5k| const char *buf, size_t len, char *latin1_output) const noexcept { 493| 10.5k| utf8_to_latin1::validating_transcoder converter; 494| 10.5k| return converter.convert(buf, len, latin1_output); 495| 10.5k|} _ZNK7simdutf7haswell14implementation23convert_utf8_to_utf16leEPKcmPDs: 511| 10.4k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 512| 10.4k| utf8_to_utf16::validating_transcoder converter; 513| 10.4k| return converter.convert(buf, len, utf16_output); 514| 10.4k|} _ZNK7simdutf7haswell14implementation23convert_utf8_to_utf16beEPKcmPDs: 517| 10.4k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 518| 10.4k| utf8_to_utf16::validating_transcoder converter; 519| 10.4k| return converter.convert(buf, len, utf16_output); 520| 10.4k|} _ZNK7simdutf7haswell14implementation21convert_utf8_to_utf32EPKcmPDi: 550| 5.26k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 551| 5.26k| utf8_to_utf32::validating_transcoder converter; 552| 5.26k| return converter.convert(buf, len, utf32_output); 553| 5.26k|} _ZNK7simdutf7haswell14implementation25convert_utf16le_to_latin1EPKDsmPc: 569| 5.26k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 570| 5.26k| std::pair ret = 571| 5.26k| haswell::avx2_convert_utf16_to_latin1(buf, len, 572| 5.26k| latin1_output); 573| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (573:7): [True: 0, False: 5.26k] ------------------ 574| 0| return 0; 575| 0| } 576| 5.26k| size_t saved_bytes = ret.second - latin1_output; 577| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (577:7): [True: 807, False: 4.46k] ------------------ 578| 807| const size_t scalar_saved_bytes = 579| 807| scalar::utf16_to_latin1::convert( 580| 807| ret.first, len - (ret.first - buf), ret.second); 581| 807| if (scalar_saved_bytes == 0) { ------------------ | Branch (581:9): [True: 0, False: 807] ------------------ 582| 0| return 0; 583| 0| } 584| 807| saved_bytes += scalar_saved_bytes; 585| 807| } 586| 5.26k| return saved_bytes; 587| 5.26k|} _ZNK7simdutf7haswell14implementation25convert_utf16be_to_latin1EPKDsmPc: 590| 5.26k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 591| 5.26k| std::pair ret = 592| 5.26k| haswell::avx2_convert_utf16_to_latin1(buf, len, 593| 5.26k| latin1_output); 594| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (594:7): [True: 0, False: 5.26k] ------------------ 595| 0| return 0; 596| 0| } 597| 5.26k| size_t saved_bytes = ret.second - latin1_output; 598| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (598:7): [True: 807, False: 4.46k] ------------------ 599| 807| const size_t scalar_saved_bytes = 600| 807| scalar::utf16_to_latin1::convert( 601| 807| ret.first, len - (ret.first - buf), ret.second); 602| 807| if (scalar_saved_bytes == 0) { ------------------ | Branch (602:9): [True: 0, False: 807] ------------------ 603| 0| return 0; 604| 0| } 605| 807| saved_bytes += scalar_saved_bytes; 606| 807| } 607| 5.26k| return saved_bytes; 608| 5.26k|} _ZNK7simdutf7haswell14implementation23convert_utf16le_to_utf8EPKDsmPc: 679| 9.93k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 680| 9.93k| std::pair ret = 681| 9.93k| haswell::avx2_convert_utf16_to_utf8(buf, len, 682| 9.93k| utf8_output); 683| 9.93k| if (ret.first == nullptr) { ------------------ | Branch (683:7): [True: 65, False: 9.87k] ------------------ 684| 65| return 0; 685| 65| } 686| 9.87k| size_t saved_bytes = ret.second - utf8_output; 687| 9.87k| if (ret.first != buf + len) { ------------------ | Branch (687:7): [True: 1.07k, False: 8.79k] ------------------ 688| 1.07k| const size_t scalar_saved_bytes = 689| 1.07k| scalar::utf16_to_utf8::convert( 690| 1.07k| ret.first, len - (ret.first - buf), ret.second); 691| 1.07k| if (scalar_saved_bytes == 0) { ------------------ | Branch (691:9): [True: 32, False: 1.04k] ------------------ 692| 32| return 0; 693| 32| } 694| 1.04k| saved_bytes += scalar_saved_bytes; 695| 1.04k| } 696| 9.83k| return saved_bytes; 697| 9.87k|} _ZNK7simdutf7haswell14implementation23convert_utf16be_to_utf8EPKDsmPc: 700| 9.93k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 701| 9.93k| std::pair ret = 702| 9.93k| haswell::avx2_convert_utf16_to_utf8(buf, len, 703| 9.93k| utf8_output); 704| 9.93k| if (ret.first == nullptr) { ------------------ | Branch (704:7): [True: 57, False: 9.87k] ------------------ 705| 57| return 0; 706| 57| } 707| 9.87k| size_t saved_bytes = ret.second - utf8_output; 708| 9.87k| if (ret.first != buf + len) { ------------------ | Branch (708:7): [True: 1.08k, False: 8.79k] ------------------ 709| 1.08k| const size_t scalar_saved_bytes = 710| 1.08k| scalar::utf16_to_utf8::convert( 711| 1.08k| ret.first, len - (ret.first - buf), ret.second); 712| 1.08k| if (scalar_saved_bytes == 0) { ------------------ | Branch (712:9): [True: 30, False: 1.05k] ------------------ 713| 30| return 0; 714| 30| } 715| 1.05k| saved_bytes += scalar_saved_bytes; 716| 1.05k| } 717| 9.84k| return saved_bytes; 718| 9.87k|} _ZNK7simdutf7haswell14implementation21convert_utf32_to_utf8EPKDimPc: 789| 4.66k| const char32_t *buf, size_t len, char *utf8_output) const noexcept { 790| 4.66k| std::pair ret = 791| 4.66k| avx2_convert_utf32_to_utf8(buf, len, utf8_output); 792| 4.66k| if (ret.first == nullptr) { ------------------ | Branch (792:7): [True: 0, False: 4.66k] ------------------ 793| 0| return 0; 794| 0| } 795| 4.66k| size_t saved_bytes = ret.second - utf8_output; 796| 4.66k| if (ret.first != buf + len) { ------------------ | Branch (796:7): [True: 788, False: 3.88k] ------------------ 797| 788| const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert( 798| 788| ret.first, len - (ret.first - buf), ret.second); 799| 788| if (scalar_saved_bytes == 0) { ------------------ | Branch (799:9): [True: 0, False: 788] ------------------ 800| 0| return 0; 801| 0| } 802| 788| saved_bytes += scalar_saved_bytes; 803| 788| } 804| 4.66k| return saved_bytes; 805| 4.66k|} _ZNK7simdutf7haswell14implementation23convert_utf32_to_latin1EPKDimPc: 810| 5.26k| const char32_t *buf, size_t len, char *latin1_output) const noexcept { 811| 5.26k| std::pair ret = 812| 5.26k| avx2_convert_utf32_to_latin1(buf, len, latin1_output); 813| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (813:7): [True: 0, False: 5.26k] ------------------ 814| 0| return 0; 815| 0| } 816| 5.26k| size_t saved_bytes = ret.second - latin1_output; 817| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (817:7): [True: 807, False: 4.46k] ------------------ 818| 807| const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert( 819| 807| ret.first, len - (ret.first - buf), ret.second); 820| 807| if (scalar_saved_bytes == 0) { ------------------ | Branch (820:9): [True: 0, False: 807] ------------------ 821| 0| return 0; 822| 0| } 823| 807| saved_bytes += scalar_saved_bytes; 824| 807| } 825| 5.26k| return saved_bytes; 826| 5.26k|} _ZNK7simdutf7haswell14implementation10count_utf8EPKcm: 1117| 10.5k|implementation::count_utf8(const char *in, size_t size) const noexcept { 1118| 10.5k| return utf8::count_code_points_bytemask(in, size); 1119| 10.5k|} _ZNK7simdutf7haswell14implementation23latin1_length_from_utf8EPKcm: 1124| 10.5k| const char *buf, size_t len) const noexcept { 1125| 10.5k| return count_utf8(buf, len); 1126| 10.5k|} _ZNK7simdutf7haswell14implementation24utf8_length_from_utf16leEPKDsm: 1131| 9.93k| const char16_t *input, size_t length) const noexcept { 1132| 9.93k| return utf16::utf8_length_from_utf16_bytemask(input, 1133| 9.93k| length); 1134| 9.93k|} _ZNK7simdutf7haswell14implementation24utf8_length_from_utf16beEPKDsm: 1137| 9.93k| const char16_t *input, size_t length) const noexcept { 1138| 9.93k| return utf16::utf8_length_from_utf16_bytemask(input, length); 1139| 9.93k|} _ZNK7simdutf7haswell14implementation22utf16_length_from_utf8EPKcm: 1156| 20.8k| const char *input, size_t length) const noexcept { 1157| 20.8k| return utf8::utf16_length_from_utf8_bytemask(input, length); 1158| 20.8k|} _ZNK7simdutf7haswell14implementation23utf8_length_from_latin1EPKcm: 1203| 5.53k| const char *input, size_t len) const noexcept { 1204| 5.53k| const uint8_t *data = reinterpret_cast(input); 1205| 5.53k| size_t answer = len / sizeof(__m256i) * sizeof(__m256i); 1206| 5.53k| size_t i = 0; 1207| 5.53k| if (answer >= 2048) { // long strings optimization ------------------ | Branch (1207:7): [True: 0, False: 5.53k] ------------------ 1208| 0| __m256i four_64bits = _mm256_setzero_si256(); 1209| 0| while (i + sizeof(__m256i) <= len) { ------------------ | Branch (1209:12): [True: 0, False: 0] ------------------ 1210| 0| __m256i runner = _mm256_setzero_si256(); 1211| | // We can do up to 255 loops without overflow. 1212| 0| size_t iterations = (len - i) / sizeof(__m256i); 1213| 0| if (iterations > 255) { ------------------ | Branch (1213:11): [True: 0, False: 0] ------------------ 1214| 0| iterations = 255; 1215| 0| } 1216| 0| size_t max_i = i + iterations * sizeof(__m256i) - sizeof(__m256i); 1217| 0| for (; i + 4 * sizeof(__m256i) <= max_i; i += 4 * sizeof(__m256i)) { ------------------ | Branch (1217:14): [True: 0, False: 0] ------------------ 1218| 0| __m256i input1 = _mm256_loadu_si256((const __m256i *)(data + i)); 1219| 0| __m256i input2 = 1220| 0| _mm256_loadu_si256((const __m256i *)(data + i + sizeof(__m256i))); 1221| 0| __m256i input3 = _mm256_loadu_si256( 1222| 0| (const __m256i *)(data + i + 2 * sizeof(__m256i))); 1223| 0| __m256i input4 = _mm256_loadu_si256( 1224| 0| (const __m256i *)(data + i + 3 * sizeof(__m256i))); 1225| 0| __m256i input12 = 1226| 0| _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input1), 1227| 0| _mm256_cmpgt_epi8(_mm256_setzero_si256(), input2)); 1228| 0| __m256i input23 = 1229| 0| _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input3), 1230| 0| _mm256_cmpgt_epi8(_mm256_setzero_si256(), input4)); 1231| 0| __m256i input1234 = _mm256_add_epi8(input12, input23); 1232| 0| runner = _mm256_sub_epi8(runner, input1234); 1233| 0| } 1234| 0| for (; i <= max_i; i += sizeof(__m256i)) { ------------------ | Branch (1234:14): [True: 0, False: 0] ------------------ 1235| 0| __m256i input_256_chunk = 1236| 0| _mm256_loadu_si256((const __m256i *)(data + i)); 1237| 0| runner = _mm256_sub_epi8( 1238| 0| runner, _mm256_cmpgt_epi8(_mm256_setzero_si256(), input_256_chunk)); 1239| 0| } 1240| 0| four_64bits = _mm256_add_epi64( 1241| 0| four_64bits, _mm256_sad_epu8(runner, _mm256_setzero_si256())); 1242| 0| } 1243| 0| answer += _mm256_extract_epi64(four_64bits, 0) + 1244| 0| _mm256_extract_epi64(four_64bits, 1) + 1245| 0| _mm256_extract_epi64(four_64bits, 2) + 1246| 0| _mm256_extract_epi64(four_64bits, 3); 1247| 5.53k| } else if (answer > 0) { ------------------ | Branch (1247:14): [True: 804, False: 4.73k] ------------------ 1248| 6.39k| for (; i + sizeof(__m256i) <= len; i += sizeof(__m256i)) { ------------------ | Branch (1248:12): [True: 5.58k, False: 804] ------------------ 1249| 5.58k| __m256i latin = _mm256_loadu_si256((const __m256i *)(data + i)); 1250| 5.58k| uint32_t non_ascii = _mm256_movemask_epi8(latin); 1251| 5.58k| answer += count_ones(non_ascii); 1252| 5.58k| } 1253| 804| } 1254| 5.53k| return answer + scalar::latin1::utf8_length_from_latin1( 1255| 5.53k| reinterpret_cast(data + i), len - i); 1256| 5.53k|} _ZNK7simdutf7haswell14implementation22utf8_length_from_utf32EPKDim: 1261| 4.66k| const char32_t *input, size_t length) const noexcept { 1262| 4.66k| return utf32::utf8_length_from_utf32(input, length); 1263| 4.66k|} _ZNK7simdutf7haswell14implementation22utf32_length_from_utf8EPKcm: 1289| 5.26k| const char *input, size_t length) const noexcept { 1290| 5.26k| return utf8::count_code_points(input, length); 1291| 5.26k|} _ZNK7simdutf7haswell14implementation16base64_to_binaryEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 1297| 15.0k| last_chunk_handling_options last_chunk_options) const noexcept { 1298| 15.0k| if (options & base64_default_or_url) { ------------------ | Branch (1298:7): [True: 0, False: 15.0k] ------------------ 1299| 0| if (options == base64_options::base64_default_or_url_accept_garbage) { ------------------ | Branch (1299:9): [True: 0, False: 0] ------------------ 1300| 0| return base64::compress_decode_base64( 1301| 0| output, input, length, options, last_chunk_options); 1302| 0| } else { 1303| 0| return base64::compress_decode_base64( 1304| 0| output, input, length, options, last_chunk_options); 1305| 0| } 1306| 15.0k| } else if (options & base64_url) { ------------------ | Branch (1306:14): [True: 0, False: 15.0k] ------------------ 1307| 0| if (options == base64_options::base64_url_accept_garbage) { ------------------ | Branch (1307:9): [True: 0, False: 0] ------------------ 1308| 0| return base64::compress_decode_base64( 1309| 0| output, input, length, options, last_chunk_options); 1310| 0| } else { 1311| 0| return base64::compress_decode_base64( 1312| 0| output, input, length, options, last_chunk_options); 1313| 0| } 1314| 15.0k| } else { 1315| 15.0k| if (options == base64_options::base64_default_accept_garbage) { ------------------ | Branch (1315:9): [True: 0, False: 15.0k] ------------------ 1316| 0| return base64::compress_decode_base64( 1317| 0| output, input, length, options, last_chunk_options); 1318| 15.0k| } else { 1319| 15.0k| return base64::compress_decode_base64( 1320| 15.0k| output, input, length, options, last_chunk_options); 1321| 15.0k| } 1322| 15.0k| } 1323| 15.0k|} _ZNK7simdutf7haswell14implementation24base64_to_binary_detailsEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 1327| 46.0k| last_chunk_handling_options last_chunk_options) const noexcept { 1328| 46.0k| if (options & base64_default_or_url) { ------------------ | Branch (1328:7): [True: 0, False: 46.0k] ------------------ 1329| 0| if (options == base64_options::base64_default_or_url_accept_garbage) { ------------------ | Branch (1329:9): [True: 0, False: 0] ------------------ 1330| 0| return base64::compress_decode_base64( 1331| 0| output, input, length, options, last_chunk_options); 1332| 0| } else { 1333| 0| return base64::compress_decode_base64( 1334| 0| output, input, length, options, last_chunk_options); 1335| 0| } 1336| 46.0k| } else if (options & base64_url) { ------------------ | Branch (1336:14): [True: 0, False: 46.0k] ------------------ 1337| 0| if (options == base64_options::base64_url_accept_garbage) { ------------------ | Branch (1337:9): [True: 0, False: 0] ------------------ 1338| 0| return base64::compress_decode_base64( 1339| 0| output, input, length, options, last_chunk_options); 1340| 0| } else { 1341| 0| return base64::compress_decode_base64( 1342| 0| output, input, length, options, last_chunk_options); 1343| 0| } 1344| 46.0k| } else { 1345| 46.0k| if (options == base64_options::base64_default_accept_garbage) { ------------------ | Branch (1345:9): [True: 0, False: 46.0k] ------------------ 1346| 0| return base64::compress_decode_base64( 1347| 0| output, input, length, options, last_chunk_options); 1348| 46.0k| } else { 1349| 46.0k| return base64::compress_decode_base64( 1350| 46.0k| output, input, length, options, last_chunk_options); 1351| 46.0k| } 1352| 46.0k| } 1353| 46.0k|} _ZNK7simdutf7haswell14implementation16binary_to_base64EPKcmPcNS_14base64_optionsE: 1417| 14.3k| base64_options options) const noexcept { 1418| 14.3k| if (options & base64_url) { ------------------ | Branch (1418:7): [True: 0, False: 14.3k] ------------------ 1419| 0| return encode_base64(output, input, length, options); 1420| 14.3k| } else { 1421| 14.3k| return encode_base64(output, input, length, options); 1422| 14.3k| } 1423| 14.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_18is_asciiERKNS1_4simd8simd8x64IhEE: 13| 33.0k|simdutf_really_inline bool is_ascii(const simd8x64 &input) { 14| 33.0k| return input.reduce_or().is_ascii(); 15| 33.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_124must_be_2_3_continuationENS1_4simd5simd8IhEES4_: 22| 26.0k| const simd8 prev3) { 23| 26.0k| simd8 is_third_byte = 24| 26.0k| prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be > 0x80 25| 26.0k| simd8 is_fourth_byte = 26| 26.0k| prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be > 0x80 27| 26.0k| return simd8(is_third_byte | is_fourth_byte); 28| 26.0k|} _ZNK7simdutf14implementation27supported_by_runtime_systemEv: 64| 21.0k|bool implementation::supported_by_runtime_system() const { 65| 21.0k| uint32_t required_instruction_sets = this->required_instruction_sets(); 66| 21.0k| uint32_t supported_instruction_sets = 67| 21.0k| internal::detect_supported_architectures(); 68| 21.0k| return ((supported_instruction_sets & required_instruction_sets) == 69| 21.0k| required_instruction_sets); 70| 21.0k|} _ZNK7simdutf14implementation33maximal_binary_length_from_base64EPKcm: 117| 76.2k| const char *input, size_t length) const noexcept { 118| 76.2k| return scalar::base64::maximal_binary_length_from_base64(input, length); 119| 76.2k|} _ZNK7simdutf14implementation25base64_length_from_binaryEmNS_14base64_optionsE: 137| 44.6k| size_t length, base64_options options) const noexcept { 138| 44.6k| return scalar::base64::base64_length_from_binary(length, options); 139| 44.6k|} _ZNK7simdutf8internal29available_implementation_list5beginEv: 1503| 5.26k|available_implementation_list::begin() const noexcept { 1504| 5.26k| return internal::get_available_implementation_pointers().begin(); 1505| 5.26k|} _ZNK7simdutf8internal29available_implementation_list3endEv: 1507| 5.26k|available_implementation_list::end() const noexcept { 1508| 5.26k| return internal::get_available_implementation_pointers().end(); 1509| 5.26k|} _ZNK7simdutf8internal29available_implementation_list21detect_best_supportedEv: 1511| 1|available_implementation_list::detect_best_supported() const noexcept { 1512| | // They are prelisted in priority order, so we just go down the list 1513| 1| uint32_t supported_instruction_sets = 1514| 1| internal::detect_supported_architectures(); 1515| 1| for (const implementation *impl : ------------------ | Branch (1515:35): [True: 2, False: 0] ------------------ 1516| 2| internal::get_available_implementation_pointers()) { 1517| 2| uint32_t required_instruction_sets = impl->required_instruction_sets(); 1518| 2| if ((supported_instruction_sets & required_instruction_sets) == ------------------ | Branch (1518:9): [True: 1, False: 1] ------------------ 1519| 2| required_instruction_sets) { 1520| 1| return impl; 1521| 1| } 1522| 2| } 1523| 0| return get_unsupported_singleton(); // this should never happen? 1524| 1|} _ZNK7simdutf8internal49detect_best_supported_implementation_on_first_use8set_bestEv: 1527| 1|detect_best_supported_implementation_on_first_use::set_best() const noexcept { 1528| 1| SIMDUTF_PUSH_DISABLE_WARNINGS 1529| | SIMDUTF_DISABLE_DEPRECATED_WARNING // Disable CRT_SECURE warning on MSVC: 1530| | // manually verified this is safe 1531| 1| char *force_implementation_name = getenv("SIMDUTF_FORCE_IMPLEMENTATION"); 1532| 1| SIMDUTF_POP_DISABLE_WARNINGS 1533| | 1534| 1| if (force_implementation_name) { ------------------ | Branch (1534:7): [True: 0, False: 1] ------------------ 1535| 0| auto force_implementation = 1536| 0| get_available_implementations()[force_implementation_name]; 1537| 0| if (force_implementation) { ------------------ | Branch (1537:9): [True: 0, False: 0] ------------------ 1538| 0| return get_active_implementation() = force_implementation; 1539| 0| } else { 1540| | // Note: abort() and stderr usage within the library is forbidden. 1541| 0| return get_active_implementation() = get_unsupported_singleton(); 1542| 0| } 1543| 0| } 1544| 1| return get_active_implementation() = 1545| 1| get_available_implementations().detect_best_supported(); 1546| 1|} _ZN7simdutf29get_available_implementationsEv: 1558| 5.26k|get_available_implementations() { 1559| 5.26k|#if !SIMDUTF_USE_STATIC_INITIALIZATION 1560| 5.26k| static const internal::available_implementation_list 1561| 5.26k| available_implementations_instance{}; 1562| 5.26k|#endif 1563| 5.26k| return available_implementations_instance; 1564| 5.26k|} _ZN7simdutf25get_active_implementationEv: 1586| 46.0k|get_active_implementation() { 1587| 46.0k|#if !SIMDUTF_USE_STATIC_INITIALIZATION 1588| 46.0k| #if !SIMDUTF_SINGLE_IMPLEMENTATION 1589| 46.0k| static const internal::detect_best_supported_implementation_on_first_use 1590| 46.0k| detect_best_supported_implementation_on_first_use_singleton; 1591| 46.0k| #endif 1592| 46.0k| static internal::atomic_ptr 1593| 46.0k| active_implementation_instance{ 1594| | #if SIMDUTF_SINGLE_IMPLEMENTATION 1595| | internal::get_single_implementation() 1596| | #else 1597| 46.0k| &detect_best_supported_implementation_on_first_use_singleton 1598| 46.0k| #endif 1599| 46.0k| }; 1600| 46.0k|#endif 1601| 46.0k| return active_implementation_instance; 1602| 46.0k|} _ZN7simdutf21base64_to_binary_safeEPKcmPcRmNS_14base64_optionsENS_27last_chunk_handling_optionsEb: 2599| 46.0k| bool decode_up_to_bad_char) noexcept { 2600| 46.0k| return base64_to_binary_safe_impl(input, length, output, outlen, 2601| 46.0k| options, last_chunk_handling_options, 2602| 46.0k| decode_up_to_bad_char); 2603| 46.0k|} simdutf.cpp:_ZN7simdutf8internalL37get_available_implementation_pointersEv: 933| 10.5k|get_available_implementation_pointers() { 934| 10.5k|#if !SIMDUTF_USE_STATIC_INITIALIZATION 935| 10.5k| static const std::initializer_list 936| 10.5k| available_implementation_pointers{ 937| 10.5k| #if SIMDUTF_IMPLEMENTATION_ICELAKE 938| 10.5k| get_icelake_singleton(), 939| 10.5k| #endif 940| 10.5k| #if SIMDUTF_IMPLEMENTATION_HASWELL 941| 10.5k| get_haswell_singleton(), 942| 10.5k| #endif 943| 10.5k| #if SIMDUTF_IMPLEMENTATION_WESTMERE 944| 10.5k| get_westmere_singleton(), 945| 10.5k| #endif 946| | #if SIMDUTF_IMPLEMENTATION_ARM64 947| | get_arm64_singleton(), 948| | #endif 949| | #if SIMDUTF_IMPLEMENTATION_PPC64 950| | get_ppc64_singleton(), 951| | #endif 952| | #if SIMDUTF_IMPLEMENTATION_RVV 953| | get_rvv_singleton(), 954| | #endif 955| | #if SIMDUTF_IMPLEMENTATION_LASX 956| | get_lasx_singleton(), 957| | #endif 958| | #if SIMDUTF_IMPLEMENTATION_LSX 959| | get_lsx_singleton(), 960| | #endif 961| 10.5k| #if SIMDUTF_IMPLEMENTATION_FALLBACK 962| 10.5k| get_fallback_singleton(), 963| 10.5k| #endif 964| 10.5k| }; 965| 10.5k|#endif 966| 10.5k| return available_implementation_pointers; 967| 10.5k|} simdutf.cpp:_ZN7simdutf8internalL21get_icelake_singletonEv: 157| 1|static const icelake::implementation *get_icelake_singleton() { 158| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 159| 1| static const icelake::implementation icelake_singleton{}; 160| 1| #endif 161| 1| return &icelake_singleton; 162| 1|} simdutf.cpp:_ZN7simdutf8internalL21get_haswell_singletonEv: 168| 1|static const haswell::implementation *get_haswell_singleton() { 169| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 170| 1| static const haswell::implementation haswell_singleton{}; 171| 1| #endif 172| 1| return &haswell_singleton; 173| 1|} simdutf.cpp:_ZN7simdutf8internalL22get_westmere_singletonEv: 179| 1|static const westmere::implementation *get_westmere_singleton() { 180| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 181| 1| static const westmere::implementation westmere_singleton{}; 182| 1| #endif 183| 1| return &westmere_singleton; 184| 1|} simdutf.cpp:_ZN7simdutf8internalL22get_fallback_singletonEv: 245| 1|static const fallback::implementation *get_fallback_singleton() { 246| 1| #if !SIMDUTF_USE_STATIC_INITIALIZATION 247| 1| static const fallback::implementation fallback_singleton{}; 248| 1| #endif 249| 1| return &fallback_singleton; 250| 1|} _ZN7simdutf8internal49detect_best_supported_implementation_on_first_useC2Ev: 887| 1| : implementation("best_supported_detector", 888| 1| "Detects the best supported implementation and sets it", 889| 1| 0) {} _ZNK7simdutf8internal49detect_best_supported_implementation_on_first_use24base64_to_binary_detailsEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 828| 1| last_chunk_handling_options::loose) const noexcept override { 829| 1| return set_best()->base64_to_binary_details(input, length, output, options, 830| 1| last_chunk_handling_options); 831| 1| } _ZN7simdutf8fallback14implementationC2Ev: 16| 1| : simdutf::implementation("fallback", "Generic fallback implementation", 17| 1| 0) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_110count_onesEm: 14| 12.4k|simdutf_really_inline long long int count_ones(uint64_t input_num) { 15| | return _popcnt64(input_num); 16| 12.4k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115trailing_zeroesEm: 20| 1.65k|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| 1.65k| return __builtin_ctzll(input_num); 25| 1.65k| #endif // SIMDUTF_REGULAR_VISUAL_STUDIO 26| 1.65k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_115is_power_of_twoImEEbT_: 29| 2.23k|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| 78.4k| simdutf_really_inline simd8() : base8_numeric() {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhEC2Ev: 124| 78.4k| simdutf_really_inline base8_numeric() : base8() {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2Ev: 71| 78.4k| simdutf_really_inline base8() : base>() {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2Ev: 14| 78.4k| simdutf_really_inline base() : value{__m256i()} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IhE9reduce_orEv: 294| 33.0k| simdutf_really_inline simd8 reduce_or() const { 295| 33.0k| return this->chunks[0] | this->chunks[1]; 296| 33.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEorES5_: 50| 141k| simdutf_really_inline Child operator|(const Child other) const { 51| 141k| return _mm256_or_si256(*this, other); 52| 141k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2EDv4_x: 193| 1.26M| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhEC2EDv4_x: 126| 1.26M| : base8(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2EDv4_x: 73| 1.26M| simdutf_really_inline base8(const __m256i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2EDv4_x: 17| 1.26M| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE8is_asciiEv: 234| 33.0k| simdutf_really_inline bool is_ascii() const { 235| 33.0k| return _mm256_movemask_epi8(*this) == 0; 236| 33.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEoRES5_: 59| 82.0k| simdutf_really_inline Child &operator|=(const Child other) { 60| 82.0k| auto this_cast = static_cast(this); 61| 82.0k| *this_cast = *this_cast | other; 62| 82.0k| return *this_cast; 63| 82.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi1EEENS4_IhEES8_: 83| 30.6k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 84| | return _mm256_alignr_epi8( 85| 30.6k| *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N); 86| 30.6k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE3shrILi4EEES4_v: 245| 61.2k| template simdutf_really_inline simd8 shr() const { 246| 61.2k| return simd8(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); 247| 61.2k| } 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| 91.8k| L replace15) const { 158| 91.8k| return lookup_16(simd8::repeat_16( 159| 91.8k| replace0, replace1, replace2, replace3, replace4, replace5, replace6, 160| 91.8k| replace7, replace8, replace9, replace10, replace11, replace12, 161| 91.8k| replace13, replace14, replace15)); 162| 91.8k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE9lookup_16IhEENS2_5simd8IT_EES8_: 148| 91.8k| simdutf_really_inline simd8 lookup_16(simd8 lookup_table) const { 149| 91.8k| return _mm256_shuffle_epi8(lookup_table, *this); 150| 91.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE9repeat_16Ehhhhhhhhhhhhhhhh: 118| 91.8k| T v14, T v15) { 119| 91.8k| return simd8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, 120| 91.8k| v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, 121| 91.8k| v12, v13, v14, v15); 122| 91.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2Ehhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh: 207| 91.8k| : simd8(_mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, 208| 91.8k| v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, 209| 91.8k| v22, v23, v24, v25, v26, v27, v28, v29, v30, 210| 91.8k| v31)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2Eh: 195| 225k| simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi2EEENS4_IhEES8_: 83| 26.0k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 84| | return _mm256_alignr_epi8( 85| 26.0k| *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N); 86| 26.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi3EEENS4_IhEES8_: 83| 26.0k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 84| | return _mm256_alignr_epi8( 85| 26.0k| *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N); 86| 26.0k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE14saturating_subES4_: 214| 59.7k| saturating_sub(const simd8 other) const { 215| 59.7k| return _mm256_subs_epu8(*this, other); 216| 59.7k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IbEC2EDv4_x: 95| 101k| simdutf_really_inline simd8(const __m256i _value) : base8(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IbNS2_5simd8IbEEEC2EDv4_x: 73| 101k| simdutf_really_inline base8(const __m256i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEC2EDv4_x: 17| 101k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEcvRKDv4_xEv: 19| 56.0k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEeoES5_: 56| 26.0k| simdutf_really_inline Child operator^(const Child other) const { 57| 26.0k| return _mm256_xor_si256(*this, other); 58| 26.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEC2EPKh: 197| 7.60k| simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE4loadEPKh: 111| 73.7k| static simdutf_really_inline simd8 load(const T values[32]) { 112| 73.7k| return _mm256_loadu_si256(reinterpret_cast(values)); 113| 73.7k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE7gt_bitsES4_: 225| 7.60k| gt_bits(const simd8 other) const { 226| 7.60k| return this->saturating_sub(other); 227| 7.60k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEC2EDv4_x: 17| 306k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEcvRKDv4_xEv: 19| 327k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IbE10to_bitmaskEv: 99| 45.1k| simdutf_really_inline uint32_t to_bitmask() const { 100| 45.1k| return uint32_t(_mm256_movemask_epi8(value)); 101| 45.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE21any_bits_set_anywhereEv: 241| 63.7k| simdutf_really_inline bool any_bits_set_anywhere() const { 242| 63.7k| return !bits_not_set_anywhere(); 243| 63.7k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE21bits_not_set_anywhereEv: 237| 63.7k| simdutf_really_inline bool bits_not_set_anywhere() const { 238| 63.7k| return _mm256_testz_si256(*this, *this); 239| 63.7k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaEC2EPKa: 274| 12.8k| : chunks{simd8::load(ptr), 275| 12.8k| simd8::load(ptr + sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaE4loadEPKa: 111| 44.4k| static simdutf_really_inline simd8 load(const T values[32]) { 112| 44.4k| return _mm256_loadu_si256(reinterpret_cast(values)); 113| 44.4k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEC2EDv4_x: 169| 83.7k| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaEC2EDv4_x: 126| 83.7k| : base8(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5base8IaNS2_5simd8IbEEEC2EDv4_x: 73| 83.7k| simdutf_really_inline base8(const __m256i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEC2EDv4_x: 17| 83.7k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE8is_asciiEv: 298| 11.1k| simdutf_really_inline bool is_ascii() const { 299| 11.1k| return this->reduce_or().is_ascii(); 300| 11.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE9reduce_orEv: 294| 11.1k| simdutf_really_inline simd8 reduce_or() const { 295| 11.1k| return this->chunks[0] | this->chunks[1]; 296| 11.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEorES5_: 50| 11.1k| simdutf_really_inline Child operator|(const Child other) const { 51| 11.1k| return _mm256_or_si256(*this, other); 52| 11.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEcvRKDv4_xEv: 19| 121k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaE8is_asciiEv: 177| 11.1k| simdutf_really_inline bool is_ascii() const { 178| 11.1k| return _mm256_movemask_epi8(*this) == 0; 179| 11.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE5storeEPa: 277| 1.68k| simdutf_really_inline void store(T *ptr) const { 278| 1.68k| this->chunks[0].store(ptr + sizeof(simd8) * 0 / sizeof(T)); 279| 1.68k| this->chunks[1].store(ptr + sizeof(simd8) * 1 / sizeof(T)); 280| 1.68k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaE5storeEPa: 129| 3.37k| simdutf_really_inline void store(T dst[32]) const { 130| 3.37k| return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); 131| 3.37k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEcvNS3_IhEEEv: 256| 23.1k|simdutf_really_inline simd8::operator simd8() const { 257| 23.1k| return this->value; 258| 23.1k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE2ltEa: 325| 7.71k| simdutf_really_inline uint64_t lt(const T m) const { 326| 7.71k| const simd8 mask = simd8::splat(m); 327| 7.71k| return simd8x64(this->chunks[0] < mask, this->chunks[1] < mask) 328| 7.71k| .to_bitmask(); 329| 7.71k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIaE5splatEa: 105| 28.1k| static simdutf_really_inline simd8 splat(T _value) { 106| 28.1k| return _mm256_set1_epi8(_value); 107| 28.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEltES4_: 184| 15.4k| simdutf_really_inline simd8 operator<(const simd8 other) const { 185| 15.4k| return _mm256_cmpgt_epi8(other, *this); 186| 15.4k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IbEC2ENS2_5simd8IbEES6_: 272| 9.47k| : chunks{chunk0, chunk1} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IbE10to_bitmaskEv: 282| 9.47k| simdutf_really_inline uint64_t to_bitmask() const { 283| 9.47k| uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask()); 284| 9.47k| uint64_t r_hi = this->chunks[1].to_bitmask(); 285| 9.47k| return r_lo | (r_hi << 32); 286| 9.47k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf32EPDi: 310| 569| simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const { 311| 569| this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8) * 0); 312| 569| this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8) * 1); 313| 569| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf32EPDi: 36| 1.13k| simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const { 37| 1.13k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), 38| 1.13k| _mm256_cvtepu8_epi32(_mm256_castsi256_si128(*this))); 39| 1.13k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 8), 40| 1.13k| _mm256_cvtepu8_epi32(_mm256_castsi256_si128( 41| 1.13k| _mm256_srli_si256(*this, 8)))); 42| 1.13k| _mm256_storeu_si256( 43| 1.13k| reinterpret_cast<__m256i *>(ptr + 16), 44| 1.13k| _mm256_cvtepu8_epi32(_mm256_extractf128_si256(*this, 1))); 45| 1.13k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 24), 46| | _mm256_cvtepu8_epi32(_mm_srli_si128( 47| 1.13k| _mm256_extractf128_si256(*this, 1), 8))); 48| 1.13k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE4zeroEv: 108| 31.4k| static simdutf_really_inline simd8 zero() { 109| 31.4k| return _mm256_setzero_si256(); 110| 31.4k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEgtES4_: 181| 22.2k| simdutf_really_inline simd8 operator>(const simd8 other) const { 182| 22.2k| return _mm256_cmpgt_epi8(*this, other); 183| 22.2k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd5simd8IaEC2Ea: 172| 18.7k| 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| 2.07k| simdutf_really_inline uint64_t sum_bytes() const { 250| 2.07k| const auto tmp = _mm256_sad_epu8(value, _mm256_setzero_si256()); 251| | 252| 2.07k| return _mm256_extract_epi64(tmp, 0) + _mm256_extract_epi64(tmp, 1) + 253| 2.07k| _mm256_extract_epi64(tmp, 2) + _mm256_extract_epi64(tmp, 3); 254| 2.07k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhEgeES4_: 229| 11.3k| operator>=(const simd8 other) const { 230| 11.3k| return other.min_val(*this) == other; 231| 11.3k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd5simd8IhE7min_valES4_: 220| 11.3k| min_val(const simd8 other) const { 221| 11.3k| return _mm256_min_epu8(other, *this); 222| 11.3k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE2gtEa: 331| 1.76k| simdutf_really_inline uint64_t gt(const T m) const { 332| 1.76k| const simd8 mask = simd8::splat(m); 333| 1.76k| return simd8x64(this->chunks[0] > mask, this->chunks[1] > mask) 334| 1.76k| .to_bitmask(); 335| 1.76k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEC2EDv4_x: 17| 10.9k| simdutf_really_inline base(const __m256i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEcvRKDv4_xEv: 19| 10.9k| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd13base8_numericIhE5splatEh: 105| 246k| static simdutf_really_inline simd8 splat(T _value) { 106| 246k| return _mm256_set1_epi8(_value); 107| 246k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEcvRKDv4_xEv: 19| 1.70M| simdutf_really_inline operator const __m256i &() const { return this->value; } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdeqENS2_5simd8IhEES4_: 76| 37.5k| const simd8 rhs) { 77| 37.5k| return _mm256_cmpeq_epi8(lhs, rhs); 78| 37.5k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEanES5_: 53| 205k| simdutf_really_inline Child operator&(const Child other) const { 54| 205k| return _mm256_and_si256(*this, other); 55| 205k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IhEC2EPKh: 274| 33.0k| : chunks{simd8::load(ptr), 275| 33.0k| simd8::load(ptr + sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEanES5_: 53| 57.1k| simdutf_really_inline Child operator&(const Child other) const { 54| 57.1k| return _mm256_and_si256(*this, other); 55| 57.1k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 303| 578| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 304| 578| this->chunks[0].template store_ascii_as_utf16(ptr + 305| 578| sizeof(simd8) * 0); 306| 578| this->chunks[1].template store_ascii_as_utf16(ptr + 307| 578| sizeof(simd8) * 1); 308| 578| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 22| 1.15k| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 23| 1.15k| __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this)); 24| 1.15k| __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this, 1)); 25| 1.15k| if (big_endian) { ------------------ | Branch (25:9): [Folded, False: 1.15k] ------------------ 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| 1.15k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first); 33| 1.15k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second); 34| 1.15k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 303| 584| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 304| 584| this->chunks[0].template store_ascii_as_utf16(ptr + 305| 584| sizeof(simd8) * 0); 306| 584| this->chunks[1].template store_ascii_as_utf16(ptr + 307| 584| sizeof(simd8) * 1); 308| 584| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 22| 1.16k| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 23| 1.16k| __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this)); 24| 1.16k| __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this, 1)); 25| 1.16k| if (big_endian) { ------------------ | Branch (25:9): [True: 1.16k, Folded] ------------------ 26| 1.16k| const __m256i swap = _mm256_setr_epi8( 27| 1.16k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 28| 1.16k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 29| 1.16k| first = _mm256_shuffle_epi8(first, swap); 30| 1.16k| second = _mm256_shuffle_epi8(second, swap); 31| 1.16k| } 32| 1.16k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first); 33| 1.16k| _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second); 34| 1.16k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItEC2EDv4_x: 97| 306k| : base16_numeric(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItEC2EDv4_x: 73| 306k| : base16(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6base16ItNS2_6simd16IbEEEC2EDv4_x: 20| 306k| : base>(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE4packERKS4_S6_: 149| 23.9k| 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| 23.9k| const __m128i lo_0 = _mm256_extracti128_si256(v0, 0); 156| 23.9k| const __m128i lo_1 = _mm256_extracti128_si256(v1, 0); 157| | 158| | // get the 1st lanes 159| 23.9k| const __m128i hi_0 = _mm256_extracti128_si256(v0, 1); 160| 23.9k| const __m128i hi_1 = _mm256_extracti128_si256(v1, 1); 161| | 162| | // build new vectors (shuffle lanes) 163| 23.9k| const __m256i t0 = _mm256_set_m128i(lo_1, lo_0); ------------------ | | 4| 23.9k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 23.9k| _mm256_castsi128_si256(xmm2), 2) ------------------ 164| 23.9k| const __m256i t1 = _mm256_set_m128i(hi_1, hi_0); ------------------ | | 4| 23.9k| _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), \ | | 5| 23.9k| _mm256_castsi128_si256(xmm2), 2) ------------------ 165| | 166| | // pack code units in linear order from v0 and v1 167| 23.9k| return _mm256_packus_epi16(t0, t1); 168| 23.9k| } 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| 5.56k| simdutf_really_inline simd16 swap_bytes() const { 140| 5.56k| const __m256i swap = _mm256_setr_epi8( 141| 5.56k| 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18, 142| 5.56k| 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30); 143| 5.56k| return _mm256_shuffle_epi8(*this, swap); 144| 5.56k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItE5splatEt: 58| 75.8k| static simdutf_really_inline simd16 splat(T _value) { 59| 75.8k| return _mm256_set1_epi16(_value); 60| 75.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdeqENS2_6simd16ItEES4_: 26| 10.9k| const simd16 rhs) { 27| 10.9k| return _mm256_cmpeq_epi16(lhs, rhs); 28| 10.9k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16IbEC2EDv4_x: 45| 10.9k| simdutf_really_inline simd16(const __m256i _value) : base16(_value) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6base16IbNS2_6simd16IbEEEC2EDv4_x: 20| 10.9k| : base>(_value) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE3shrILi8EEES4_v: 134| 23.7k| template simdutf_really_inline simd16 shr() const { 135| 23.7k| return simd16(_mm256_srli_epi16(*this, N)); 136| 23.7k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItEC2EPKDs: 104| 47.9k| : simd16(load(reinterpret_cast(values))) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd16ItEC2Et: 100| 55.9k| simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItE4zeroEv: 62| 19.8k| static simdutf_really_inline simd16 zero() { 63| 19.8k| return _mm256_setzero_si256(); 64| 19.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd3minENS2_6simd16ItEES4_: 259| 21.9k|simd16 min(const simd16 a, simd16 b) { 260| 21.9k| return _mm256_min_epu16(a.value, b.value); 261| 21.9k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItEpLENS2_6simd16ItEE: 87| 32.9k| simdutf_really_inline simd16 &operator+=(const simd16 other) { 88| 32.9k| *this = *this + other; 89| 32.9k| return *static_cast *>(this); 90| 32.9k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd14base16_numericItEplENS2_6simd16ItEE: 84| 32.9k| simdutf_really_inline simd16 operator+(const simd16 other) const { 85| 32.9k| return _mm256_add_epi16(*this, other); 86| 32.9k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd16ItE3sumEv: 170| 19.8k| simdutf_really_inline uint64_t sum() const { 171| 19.8k| const auto lo_u16 = _mm256_and_si256(value, _mm256_set1_epi32(0x0000ffff)); 172| 19.8k| const auto hi_u16 = _mm256_srli_epi32(value, 16); 173| 19.8k| const auto sum_u32 = _mm256_add_epi32(lo_u16, hi_u16); 174| | 175| 19.8k| const auto lo_u32 = 176| 19.8k| _mm256_and_si256(sum_u32, _mm256_set1_epi64x(0xffffffff)); 177| 19.8k| const auto hi_u32 = _mm256_srli_epi64(sum_u32, 32); 178| 19.8k| const auto sum_u64 = _mm256_add_epi64(lo_u32, hi_u32); 179| | 180| 19.8k| return uint64_t(_mm256_extract_epi64(sum_u64, 0)) + 181| 19.8k| uint64_t(_mm256_extract_epi64(sum_u64, 1)) + 182| 19.8k| uint64_t(_mm256_extract_epi64(sum_u64, 2)) + 183| | uint64_t(_mm256_extract_epi64(sum_u64, 3)); 184| 19.8k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE5splatEj: 46| 24.3k| simdutf_really_inline static simd32 splat(uint32_t v) { 47| 24.3k| return _mm256_set1_epi32(v); 48| 24.3k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjEC2EDv4_x: 9| 177k| simdutf_really_inline simd32(const __m256i v) : value(v) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE4zeroEv: 42| 4.42k| simdutf_really_inline static simd32 zero() { 43| 4.42k| return _mm256_setzero_si256(); 44| 4.42k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjEC2IDiEEPKT_: 13| 33.4k| : value(_mm256_loadu_si256(reinterpret_cast(ptr))) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE10swap_bytesEv: 27| 26.0k| simdutf_really_inline simd32 swap_bytes() const { 28| 26.0k| const __m256i shuffle = 29| 26.0k| _mm256_setr_epi8(3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12, 30| 26.0k| 3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12); 31| | 32| 26.0k| return _mm256_shuffle_epi8(value, shuffle); 33| 26.0k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd3maxENS2_6simd32IjEES4_: 80| 52.0k| const simd32 b) { 81| 52.0k| return _mm256_max_epu32(a.value, b.value); 82| 52.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdplENS2_6simd32IjEES4_: 90| 26.0k| const simd32 b) { 91| 26.0k| return _mm256_add_epi32(a.value, b.value); 92| 26.0k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdgtENS2_6simd32IjEES4_: 109| 3.77k| const simd32 b) { 110| 3.77k| return !(b >= a); 111| 3.77k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdntENS2_6simd32IbEE: 104| 3.77k|simdutf_really_inline simd32 operator!(const simd32 v) { 105| 3.77k| return _mm256_xor_si256(v.value, _mm256_set1_epi8(-1)); 106| 3.77k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IbEC2EDv4_x: 59| 7.55k| simdutf_really_inline simd32(const __m256i v) : value(v) {} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdgeENS2_6simd32IjEES4_: 100| 3.77k| const simd32 b) { 101| 3.77k| return _mm256_cmpeq_epi32(_mm256_max_epu32(a.value, b.value), a.value); 102| 3.77k|} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd32IbE3anyEv: 61| 3.77k| simdutf_really_inline bool any() const { 62| 3.77k| return _mm256_movemask_epi8(value) != 0; 63| 3.77k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simdanENS2_6simd32IjEES4_: 85| 22.3k| const simd32 a) { 86| 22.3k| return _mm256_and_si256(a.value, b.value); 87| 22.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd3minENS2_6simd32IjEES4_: 75| 22.3k| const simd32 a) { 76| 22.3k| return _mm256_min_epu32(a.value, b.value); 77| 22.3k|} simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd32IjEpLES4_: 36| 22.3k| simdutf_really_inline simd32 &operator+=(const simd32 other) { 37| 22.3k| value = _mm256_add_epi32(value, other.value); 38| 22.3k| return *this; 39| 22.3k| } simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd32IjE3sumEv: 15| 646| simdutf_really_inline uint64_t sum() const { 16| 646| const __m256i mask = _mm256_set1_epi64x(0xffffffff); 17| 646| const __m256i t0 = _mm256_and_si256(value, mask); 18| 646| const __m256i t1 = _mm256_srli_epi64(value, 32); 19| 646| const __m256i t2 = _mm256_add_epi64(t0, t1); 20| | 21| 646| return uint64_t(_mm256_extract_epi64(t2, 0)) + 22| 646| uint64_t(_mm256_extract_epi64(t2, 1)) + 23| 646| uint64_t(_mm256_extract_epi64(t2, 2)) + 24| | uint64_t(_mm256_extract_epi64(t2, 3)); 25| 646| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd64ImE4zeroEv: 27| 31.4k| simdutf_really_inline static simd64 zero() { 28| 31.4k| return _mm256_setzero_si256(); 29| 31.4k| } simdutf.cpp:_ZN7simdutf7haswell12_GLOBAL__N_14simd6simd64ImEC2EDv4_x: 9| 31.4k| simdutf_really_inline simd64(const __m256i v) : value(v) {} simdutf.cpp:_ZNK7simdutf7haswell12_GLOBAL__N_14simd6simd64ImE3sumEv: 15| 31.4k| simdutf_really_inline uint64_t sum() const { 16| 31.4k| return _mm256_extract_epi64(value, 0) + _mm256_extract_epi64(value, 1) + 17| 31.4k| _mm256_extract_epi64(value, 2) + _mm256_extract_epi64(value, 3); 18| 31.4k| } _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| 17.3k|simdutf_really_inline long long int count_ones(uint64_t input_num) { 15| | return _popcnt64(input_num); 16| 17.3k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115trailing_zeroesEm: 20| 475|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| 475| return __builtin_ctzll(input_num); 27| 475| #endif // SIMDUTF_REGULAR_VISUAL_STUDIO 28| 475|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_115is_power_of_twoImEEbT_: 31| 724|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| 78.3k| simdutf_really_inline simd8() : base8_numeric() {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEC2Ev: 118| 78.3k| simdutf_really_inline base8_numeric() : base8() {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2Ev: 68| 78.3k| simdutf_really_inline base8() : base>() {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2Ev: 13| 78.3k| simdutf_really_inline base() : value{__m128i()} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IhE9reduce_orEv: 290| 35.4k| simdutf_really_inline simd8 reduce_or() const { 291| 35.4k| return (this->chunks[0] | this->chunks[1]) | 292| 35.4k| (this->chunks[2] | this->chunks[3]); 293| 35.4k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEorES5_: 42| 302k| simdutf_really_inline Child operator|(const Child other) const { 43| 302k| return _mm_or_si128(*this, other); 44| 302k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2EDv2_x: 185| 2.96M| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEC2EDv2_x: 120| 2.96M| : base8(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEEC2EDv2_x: 69| 2.96M| simdutf_really_inline base8(const __m128i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEC2EDv2_x: 16| 2.96M| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE8is_asciiEv: 228| 35.4k| simdutf_really_inline bool is_ascii() const { 229| 35.4k| return _mm_movemask_epi8(*this) == 0; 230| 35.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEoRES5_: 51| 130k| simdutf_really_inline Child &operator|=(const Child other) { 52| 130k| auto this_cast = static_cast(this); 53| 130k| *this_cast = *this_cast | other; 54| 130k| return *this_cast; 55| 130k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi1EEENS4_IhEES8_: 79| 77.9k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 80| | return _mm_alignr_epi8(*this, prev_chunk, 16 - N); 81| 77.9k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE3shrILi4EEES4_v: 238| 155k| template simdutf_really_inline simd8 shr() const { 239| 155k| return simd8(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); 240| 155k| } 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| 233k| L replace15) const { 152| 233k| return lookup_16(simd8::repeat_16( 153| 233k| replace0, replace1, replace2, replace3, replace4, replace5, replace6, 154| 233k| replace7, replace8, replace9, replace10, replace11, replace12, 155| 233k| replace13, replace14, replace15)); 156| 233k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE9lookup_16IhEENS2_5simd8IT_EES8_: 142| 233k| simdutf_really_inline simd8 lookup_16(simd8 lookup_table) const { 143| 233k| return _mm_shuffle_epi8(lookup_table, *this); 144| 233k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE9repeat_16Ehhhhhhhhhhhhhhhh: 113| 233k| T v14, T v15) { 114| 233k| return simd8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, 115| 233k| v14, v15); 116| 233k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2Ehhhhhhhhhhhhhhhh: 196| 233k| : simd8(_mm_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, 197| 233k| v12, v13, v14, v15)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2Eh: 188| 511k| simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi2EEENS4_IhEES8_: 79| 65.6k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 80| | return _mm_alignr_epi8(*this, prev_chunk, 16 - N); 81| 65.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5base8IhNS2_5simd8IbEEE4prevILi3EEENS4_IhEES8_: 79| 65.6k| simdutf_really_inline simd8 prev(const simd8 prev_chunk) const { 80| | return _mm_alignr_epi8(*this, prev_chunk, 16 - N); 81| 65.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE14saturating_subES4_: 201| 139k| saturating_sub(const simd8 other) const { 202| 139k| return _mm_subs_epu8(*this, other); 203| 139k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IbEC2EDv2_x: 91| 246k| simdutf_really_inline simd8(const __m128i _value) : base8(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IbNS2_5simd8IbEEEC2EDv2_x: 69| 246k| simdutf_really_inline base8(const __m128i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEC2EDv2_x: 16| 246k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IbEEEcvRKDv2_xEv: 18| 246k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEeoES5_: 48| 65.6k| simdutf_really_inline Child operator^(const Child other) const { 49| 65.6k| return _mm_xor_si128(*this, other); 50| 65.6k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEC2EPKh: 190| 8.59k| simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE4loadEPKh: 106| 150k| static simdutf_really_inline simd8 load(const T values[16]) { 107| 150k| return _mm_loadu_si128(reinterpret_cast(values)); 108| 150k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE7gt_bitsES4_: 212| 8.59k| gt_bits(const simd8 other) const { 213| 8.59k| return this->saturating_sub(other); 214| 8.59k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEC2EDv2_x: 16| 671k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEcvRKDv2_xEv: 18| 649k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IbE10to_bitmaskEv: 95| 91.7k| simdutf_really_inline int to_bitmask() const { 96| 91.7k| return _mm_movemask_epi8(*this); 97| 91.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE21any_bits_set_anywhereEv: 235| 64.1k| simdutf_really_inline bool any_bits_set_anywhere() const { 236| 64.1k| return !bits_not_set_anywhere(); 237| 64.1k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE21bits_not_set_anywhereEv: 232| 64.1k| simdutf_really_inline bool bits_not_set_anywhere() const { 233| 64.1k| return _mm_testz_si128(*this, *this); 234| 64.1k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaEC2EPKa: 270| 18.1k| : chunks{simd8::load(ptr), 271| 18.1k| simd8::load(ptr + sizeof(simd8) / sizeof(T)), 272| 18.1k| simd8::load(ptr + 2 * sizeof(simd8) / sizeof(T)), 273| 18.1k| simd8::load(ptr + 3 * sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaE4loadEPKa: 106| 128k| static simdutf_really_inline simd8 load(const T values[16]) { 107| 128k| return _mm_loadu_si128(reinterpret_cast(values)); 108| 128k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEC2EDv2_x: 163| 243k| : base8_numeric(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaEC2EDv2_x: 120| 243k| : base8(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5base8IaNS2_5simd8IbEEEC2EDv2_x: 69| 243k| simdutf_really_inline base8(const __m128i _value) : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEC2EDv2_x: 16| 243k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE8is_asciiEv: 295| 15.6k| simdutf_really_inline bool is_ascii() const { 296| 15.6k| return this->reduce_or().is_ascii(); 297| 15.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE9reduce_orEv: 290| 15.6k| simdutf_really_inline simd8 reduce_or() const { 291| 15.6k| return (this->chunks[0] | this->chunks[1]) | 292| 15.6k| (this->chunks[2] | this->chunks[3]); 293| 15.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEorES5_: 42| 46.9k| simdutf_really_inline Child operator|(const Child other) const { 43| 46.9k| return _mm_or_si128(*this, other); 44| 46.9k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEEcvRKDv2_xEv: 18| 364k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaE8is_asciiEv: 168| 15.6k| simdutf_really_inline bool is_ascii() const { 169| 15.6k| return _mm_movemask_epi8(*this) == 0; 170| 15.6k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE5storeEPa: 275| 2.03k| simdutf_really_inline void store(T *ptr) const { 276| 2.03k| this->chunks[0].store(ptr + sizeof(simd8) * 0 / sizeof(T)); 277| 2.03k| this->chunks[1].store(ptr + sizeof(simd8) * 1 / sizeof(T)); 278| 2.03k| this->chunks[2].store(ptr + sizeof(simd8) * 2 / sizeof(T)); 279| 2.03k| this->chunks[3].store(ptr + sizeof(simd8) * 3 / sizeof(T)); 280| 2.03k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaE5storeEPa: 123| 8.15k| simdutf_really_inline void store(T dst[16]) const { 124| 8.15k| return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); 125| 8.15k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEcvNS3_IhEEEv: 251| 76.2k|simdutf_really_inline simd8::operator simd8() const { 252| 76.2k| return this->value; 253| 76.2k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE2ltEa: 326| 10.8k| simdutf_really_inline uint64_t lt(const T m) const { 327| 10.8k| const simd8 mask = simd8::splat(m); 328| 10.8k| return simd8x64(this->chunks[0] < mask, this->chunks[1] < mask, 329| 10.8k| this->chunks[2] < mask, this->chunks[3] < mask) 330| 10.8k| .to_bitmask(); 331| 10.8k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIaE5splatEa: 102| 68.8k| static simdutf_really_inline simd8 splat(T _value) { 103| 68.8k| return _mm_set1_epi8(_value); 104| 68.8k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEltES4_: 176| 43.5k| simdutf_really_inline simd8 operator<(const simd8 other) const { 177| 43.5k| return _mm_cmpgt_epi8(other, *this); 178| 43.5k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IbEC2ENS2_5simd8IbEES6_S6_S6_: 268| 13.4k| : chunks{chunk0, chunk1, chunk2, chunk3} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IbE10to_bitmaskEv: 318| 13.4k| simdutf_really_inline uint64_t to_bitmask() const { 319| 13.4k| uint64_t r0 = uint32_t(this->chunks[0].to_bitmask()); 320| 13.4k| uint64_t r1 = this->chunks[1].to_bitmask(); 321| 13.4k| uint64_t r2 = this->chunks[2].to_bitmask(); 322| 13.4k| uint64_t r3 = this->chunks[3].to_bitmask(); 323| 13.4k| return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48); 324| 13.4k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf32EPDi: 311| 886| simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const { 312| 886| this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8) * 0); 313| 886| this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8) * 1); 314| 886| this->chunks[2].store_ascii_as_utf32(ptr + sizeof(simd8) * 2); 315| 886| this->chunks[3].store_ascii_as_utf32(ptr + sizeof(simd8) * 3); 316| 886| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf32EPDi: 32| 3.54k| simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const { 33| 3.54k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p), _mm_cvtepu8_epi32(*this)); 34| 3.54k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 4), 35| 3.54k| _mm_cvtepu8_epi32(_mm_srli_si128(*this, 4))); 36| 3.54k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), 37| 3.54k| _mm_cvtepu8_epi32(_mm_srli_si128(*this, 8))); 38| 3.54k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 12), 39| | _mm_cvtepu8_epi32(_mm_srli_si128(*this, 12))); 40| 3.54k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE4zeroEv: 105| 31.4k| static simdutf_really_inline simd8 zero() { return _mm_setzero_si128(); } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEgtES4_: 173| 65.5k| simdutf_really_inline simd8 operator>(const simd8 other) const { 174| 65.5k| return _mm_cmpgt_epi8(*this, other); 175| 65.5k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd5simd8IaEC2Ea: 165| 55.3k| simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEmIENS2_5simd8IhEE: 134| 88.7k| simdutf_really_inline simd8 &operator-=(const simd8 other) { 135| 88.7k| *this = *this - other; 136| 88.7k| return *static_cast *>(this); 137| 88.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhEmiENS2_5simd8IhEE: 131| 88.7k| simdutf_really_inline simd8 operator-(const simd8 other) const { 132| 88.7k| return _mm_sub_epi8(*this, other); 133| 88.7k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE9sum_bytesEv: 245| 2.93k| simdutf_really_inline uint64_t sum_bytes() const { 246| 2.93k| const auto tmp = _mm_sad_epu8(value, _mm_setzero_si128()); 247| 2.93k| return _mm_extract_epi64(tmp, 0) + _mm_extract_epi64(tmp, 1); 248| 2.93k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhEgeES4_: 217| 33.3k| operator>=(const simd8 other) const { 218| 33.3k| return other.min_val(*this) == other; 219| 33.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd5simd8IhE7min_valES4_: 207| 33.3k| min_val(const simd8 other) const { 208| 33.3k| return _mm_min_epu8(*this, other); 209| 33.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE2gtEa: 333| 2.54k| simdutf_really_inline uint64_t gt(const T m) const { 334| 2.54k| const simd8 mask = simd8::splat(m); 335| 2.54k| return simd8x64(this->chunks[0] > mask, this->chunks[1] > mask, 336| 2.54k| this->chunks[2] > mask, this->chunks[3] > mask) 337| 2.54k| .to_bitmask(); 338| 2.54k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEC2EDv2_x: 16| 32.4k| simdutf_really_inline base(const __m128i _value) : value(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16IbEEEcvRKDv2_xEv: 18| 32.4k| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd13base8_numericIhE5splatEh: 102| 528k| static simdutf_really_inline simd8 splat(T _value) { 103| 528k| return _mm_set1_epi8(_value); 104| 528k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEcvRKDv2_xEv: 18| 3.66M| simdutf_really_inline operator const __m128i &() const { return this->value; } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdeqENS2_5simd8IhEES4_: 72| 71.3k| const simd8 rhs) { 73| 71.3k| return _mm_cmpeq_epi8(lhs, rhs); 74| 71.3k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IhEEEanES5_: 45| 493k| simdutf_really_inline Child operator&(const Child other) const { 46| 493k| return _mm_and_si128(*this, other); 47| 493k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IhEC2EPKh: 270| 35.4k| : chunks{simd8::load(ptr), 271| 35.4k| simd8::load(ptr + sizeof(simd8) / sizeof(T)), 272| 35.4k| simd8::load(ptr + 2 * sizeof(simd8) / sizeof(T)), 273| 35.4k| simd8::load(ptr + 3 * sizeof(simd8) / sizeof(T))} {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_6simd16ItEEEanES5_: 45| 133k| simdutf_really_inline Child operator&(const Child other) const { 46| 133k| return _mm_and_si128(*this, other); 47| 133k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 300| 916| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 301| 916| this->chunks[0].template store_ascii_as_utf16(ptr + 302| 916| sizeof(simd8) * 0); 303| 916| this->chunks[1].template store_ascii_as_utf16(ptr + 304| 916| sizeof(simd8) * 1); 305| 916| this->chunks[2].template store_ascii_as_utf16(ptr + 306| 916| sizeof(simd8) * 2); 307| 916| this->chunks[3].template store_ascii_as_utf16(ptr + 308| 916| sizeof(simd8) * 3); 309| 916| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE0EEEvPDs: 20| 3.66k| simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const { 21| 3.66k| __m128i first = _mm_cvtepu8_epi16(*this); 22| 3.66k| __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this, 8)); 23| 3.66k| if (big_endian) { ------------------ | Branch (23:9): [Folded, False: 3.66k] ------------------ 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| 3.66k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first); 30| 3.66k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), second); 31| 3.66k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd8simd8x64IaE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 300| 915| simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const { 301| 915| this->chunks[0].template store_ascii_as_utf16(ptr + 302| 915| sizeof(simd8) * 0); 303| 915| this->chunks[1].template store_ascii_as_utf16(ptr + 304| 915| sizeof(simd8) * 1); 305| 915| this->chunks[2].template store_ascii_as_utf16(ptr + 306| 915| sizeof(simd8) * 2); 307| 915| this->chunks[3].template store_ascii_as_utf16(ptr + 308| 915| sizeof(simd8) * 3); 309| 915| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd4baseINS2_5simd8IaEEE20store_ascii_as_utf16ILNS_10endiannessE1EEEvPDs: 20| 3.66k| simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const { 21| 3.66k| __m128i first = _mm_cvtepu8_epi16(*this); 22| 3.66k| __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this, 8)); 23| 3.66k| if (big_endian) { ------------------ | Branch (23:9): [True: 3.66k, Folded] ------------------ 24| 3.66k| const __m128i swap = 25| 3.66k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 26| 3.66k| first = _mm_shuffle_epi8(first, swap); 27| 3.66k| second = _mm_shuffle_epi8(second, swap); 28| 3.66k| } 29| 3.66k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first); 30| 3.66k| _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), second); 31| 3.66k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItEC2EDv2_x: 79| 671k| : base16_numeric(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItEC2EDv2_x: 54| 671k| : base16(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6base16ItNS2_6simd16IbEEEC2EDv2_x: 8| 671k| : base>(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE4packERKS4_S6_: 130| 35.5k| const simd16 &v1) { 131| 35.5k| return _mm_packus_epi16(v0, v1); 132| 35.5k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItE4loadEPKt: 47| 103k| static simdutf_really_inline simd16 load(const T values[8]) { 48| 103k| return _mm_loadu_si128(reinterpret_cast(values)); 49| 103k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE10swap_bytesEv: 121| 16.3k| simdutf_really_inline simd16 swap_bytes() const { 122| 16.3k| const __m128i swap = 123| 16.3k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 124| 16.3k| return _mm_shuffle_epi8(*this, swap); 125| 16.3k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItE5splatEt: 41| 168k| static simdutf_really_inline simd16 splat(T _value) { 42| 168k| return _mm_set1_epi16(_value); 43| 168k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdeqENS2_6simd16ItEES4_: 11| 32.4k| const simd16 rhs) { 12| 32.4k| return _mm_cmpeq_epi16(lhs, rhs); 13| 32.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16IbEC2EDv2_x: 28| 32.4k| simdutf_really_inline simd16(const __m128i _value) : base16(_value) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6base16IbNS2_6simd16IbEEEC2EDv2_x: 8| 32.4k| : base>(_value) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE3shrILi8EEES4_v: 116| 35.3k| template simdutf_really_inline simd16 shr() const { 117| 35.3k| return simd16(_mm_srli_epi16(*this, N)); 118| 35.3k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItEC2EPKDs: 86| 71.1k| : simd16(load(reinterpret_cast(values))) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd16ItEC2Et: 82| 147k| simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItE4zeroEv: 45| 20.2k| static simdutf_really_inline simd16 zero() { return _mm_setzero_si128(); } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd3minENS2_6simd16ItEES4_: 240| 64.9k|simd16 min(const simd16 a, simd16 b) { 241| 64.9k| return _mm_min_epu16(a.value, b.value); 242| 64.9k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItEpLENS2_6simd16ItEE: 68| 97.4k| simdutf_really_inline simd16 &operator+=(const simd16 other) { 69| 97.4k| *this = *this + other; 70| 97.4k| return *static_cast *>(this); 71| 97.4k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd14base16_numericItEplENS2_6simd16ItEE: 65| 97.4k| simdutf_really_inline simd16 operator+(const simd16 other) const { 66| 97.4k| return _mm_add_epi16(*this, other); 67| 97.4k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd16ItE3sumEv: 134| 20.2k| simdutf_really_inline uint64_t sum() const { 135| 20.2k| const auto lo_u16 = _mm_and_si128(value, _mm_set1_epi32(0x0000ffff)); 136| 20.2k| const auto hi_u16 = _mm_srli_epi32(value, 16); 137| 20.2k| const auto sum_u32 = _mm_add_epi32(lo_u16, hi_u16); 138| | 139| 20.2k| const auto lo_u32 = _mm_and_si128(sum_u32, _mm_set1_epi64x(0xffffffff)); 140| 20.2k| const auto hi_u32 = _mm_srli_epi64(sum_u32, 32); 141| 20.2k| const auto sum_u64 = _mm_add_epi64(lo_u32, hi_u32); 142| | 143| 20.2k| return uint64_t(_mm_extract_epi64(sum_u64, 0)) + 144| | uint64_t(_mm_extract_epi64(sum_u64, 1)); 145| 20.2k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE5splatEj: 56| 23.8k| simdutf_really_inline static simd32 splat(uint32_t v) { 57| 23.8k| return _mm_set1_epi32(v); 58| 23.8k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjEC2EDv2_x: 9| 396k| simdutf_really_inline simd32(const __m128i v) : value(v) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE4zeroEv: 52| 3.85k| simdutf_really_inline static simd32 zero() { 53| 3.85k| return _mm_setzero_si128(); 54| 3.85k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjEC2IDiEEPKT_: 13| 81.6k| : value(_mm_loadu_si128(reinterpret_cast(ptr))) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE10swap_bytesEv: 22| 60.6k| simdutf_really_inline simd32 swap_bytes() const { 23| 60.6k| const __m128i shuffle = 24| 60.6k| _mm_setr_epi8(3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12); 25| | 26| 60.6k| return _mm_shuffle_epi8(value, shuffle); 27| 60.6k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd3maxENS2_6simd32IjEES4_: 94| 121k| const simd32 b) { 95| 121k| return _mm_max_epu32(a.value, b.value); 96| 121k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdplENS2_6simd32IjEES4_: 119| 60.6k| const simd32 b) { 120| 60.6k| return _mm_add_epi32(a.value, b.value); 121| 60.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdgtENS2_6simd32IjEES4_: 143| 2.94k| const simd32 b) { 144| 2.94k| return !(b >= a); 145| 2.94k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdntENS2_6simd32IbEE: 138| 2.94k|simdutf_really_inline simd32 operator!(const simd32 v) { 139| 2.94k| return _mm_xor_si128(v.value, _mm_set1_epi8(-1)); 140| 2.94k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IbEC2EDv2_x: 69| 5.88k| simdutf_really_inline simd32(const __m128i v) : value(v) {} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdgeENS2_6simd32IjEES4_: 134| 2.94k| const simd32 b) { 135| 2.94k| return _mm_cmpeq_epi32(_mm_max_epu32(a.value, b.value), a.value); 136| 2.94k|} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd32IbE3anyEv: 71| 2.94k| simdutf_really_inline bool any() const { 72| 2.94k| return _mm_movemask_epi8(value) != 0; 73| 2.94k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simdanENS2_6simd32IjEES4_: 104| 62.9k| const simd32 b) { 105| 62.9k| return _mm_and_si128(a.value, b.value); 106| 62.9k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd3minENS2_6simd32IjEES4_: 89| 62.9k| const simd32 b) { 90| 62.9k| return _mm_min_epu32(a.value, b.value); 91| 62.9k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd32IjEpLES4_: 46| 62.9k| simdutf_really_inline simd32 &operator+=(const simd32 other) { 47| 62.9k| value = _mm_add_epi32(value, other.value); 48| 62.9k| return *this; 49| 62.9k| } simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd32IjE3sumEv: 15| 910| simdutf_really_inline uint64_t sum() const { 16| 910| return uint64_t(_mm_extract_epi32(value, 0)) + 17| 910| uint64_t(_mm_extract_epi32(value, 1)) + 18| 910| uint64_t(_mm_extract_epi32(value, 2)) + 19| | uint64_t(_mm_extract_epi32(value, 3)); 20| 910| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd64ImE4zeroEv: 26| 31.4k| simdutf_really_inline static simd64 zero() { 27| 31.4k| return _mm_setzero_si128(); 28| 31.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_14simd6simd64ImEC2EDv2_x: 9| 31.4k| simdutf_really_inline simd64(const __m128i v) : value(v) {} simdutf.cpp:_ZNK7simdutf8westmere12_GLOBAL__N_14simd6simd64ImE3sumEv: 15| 31.4k| simdutf_really_inline uint64_t sum() const { 16| 31.4k| return _mm_extract_epi64(value, 0) + _mm_extract_epi64(value, 1); 17| 31.4k| } _ZNK7simdutf8westmere14implementation13validate_utf8EPKcm: 299| 20.8k|implementation::validate_utf8(const char *buf, size_t len) const noexcept { 300| 20.8k| return westmere::utf8_validation::generic_validate_utf8(buf, len); 301| 20.8k|} _ZNK7simdutf8westmere14implementation25validate_utf8_with_errorsEPKcm: 306| 5.26k| const char *buf, size_t len) const noexcept { 307| 5.26k| return westmere::utf8_validation::generic_validate_utf8_with_errors(buf, len); 308| 5.26k|} _ZNK7simdutf8westmere14implementation16validate_utf16leEPKDsm: 345| 15.3k| size_t len) const noexcept { 346| 15.3k| if (simdutf_unlikely(len == 0)) { ------------------ | | 93| 15.3k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 13.5k, False: 1.79k] | | ------------------ ------------------ 347| | // empty input is valid UTF-16. protect the implementation from 348| | // handling nullptr 349| 13.5k| return true; 350| 13.5k| } 351| 1.79k| const auto res = 352| 1.79k| westmere::utf16::validate_utf16_with_errors(buf, len); 353| 1.79k| if (res.is_err()) { ------------------ | Branch (353:7): [True: 63, False: 1.73k] ------------------ 354| 63| return false; 355| 63| } 356| | 357| 1.73k| if (res.count == len) ------------------ | Branch (357:7): [True: 0, False: 1.73k] ------------------ 358| 0| return true; 359| | 360| 1.73k| return scalar::utf16::validate(buf + res.count, 361| 1.73k| len - res.count); 362| 1.73k|} _ZNK7simdutf8westmere14implementation16validate_utf16beEPKDsm: 368| 15.3k| size_t len) const noexcept { 369| 15.3k| if (simdutf_unlikely(len == 0)) { ------------------ | | 93| 15.3k| #define simdutf_unlikely(x) __builtin_expect(!!(x), 0) | | ------------------ | | | Branch (93:33): [True: 13.6k, False: 1.78k] | | ------------------ ------------------ 370| | // empty input is valid UTF-16. protect the implementation from 371| | // handling nullptr 372| 13.6k| return true; 373| 13.6k| } 374| 1.78k| const auto res = 375| 1.78k| westmere::utf16::validate_utf16_with_errors(buf, len); 376| 1.78k| if (res.is_err()) { ------------------ | Branch (376:7): [True: 62, False: 1.72k] ------------------ 377| 62| return false; 378| 62| } 379| | 380| 1.72k| if (res.count == len) ------------------ | Branch (380:7): [True: 0, False: 1.72k] ------------------ 381| 0| return true; 382| | 383| 1.72k| return scalar::utf16::validate(buf + res.count, 384| 1.72k| len - res.count); 385| 1.72k|} _ZNK7simdutf8westmere14implementation28validate_utf16le_with_errorsEPKDsm: 388| 5.26k| const char16_t *buf, size_t len) const noexcept { 389| 5.26k| const result res = 390| 5.26k| westmere::utf16::validate_utf16_with_errors(buf, len); 391| 5.26k| if (res.count != len) { ------------------ | Branch (391:7): [True: 329, False: 4.93k] ------------------ 392| 329| const result scalar_res = 393| 329| scalar::utf16::validate_with_errors( 394| 329| buf + res.count, len - res.count); 395| 329| return result(scalar_res.error, res.count + scalar_res.count); 396| 4.93k| } else { 397| 4.93k| return res; 398| 4.93k| } 399| 5.26k|} _ZNK7simdutf8westmere14implementation28validate_utf16be_with_errorsEPKDsm: 402| 5.26k| const char16_t *buf, size_t len) const noexcept { 403| 5.26k| const result res = 404| 5.26k| westmere::utf16::validate_utf16_with_errors(buf, len); 405| 5.26k| if (res.count != len) { ------------------ | Branch (405:7): [True: 319, False: 4.94k] ------------------ 406| 319| result scalar_res = scalar::utf16::validate_with_errors( 407| 319| buf + res.count, len - res.count); 408| 319| return result(scalar_res.error, res.count + scalar_res.count); 409| 4.94k| } else { 410| 4.94k| return res; 411| 4.94k| } 412| 5.26k|} _ZNK7simdutf8westmere14implementation14validate_utf32EPKDim: 427| 10.1k|implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept { 428| 10.1k| return utf32::validate(buf, len); 429| 10.1k|} _ZNK7simdutf8westmere14implementation22convert_latin1_to_utf8EPKcmPc: 441| 5.55k| const char *buf, size_t len, char *utf8_output) const noexcept { 442| | 443| 5.55k| std::pair ret = 444| 5.55k| sse_convert_latin1_to_utf8(buf, len, utf8_output); 445| 5.55k| size_t converted_chars = ret.second - utf8_output; 446| | 447| 5.55k| if (ret.first != buf + len) { ------------------ | Branch (447:7): [True: 812, False: 4.74k] ------------------ 448| 812| const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert( 449| 812| ret.first, len - (ret.first - buf), ret.second); 450| 812| converted_chars += scalar_converted_chars; 451| 812| } 452| | 453| 5.55k| return converted_chars; 454| 5.55k|} _ZNK7simdutf8westmere14implementation25convert_latin1_to_utf16leEPKcmPDs: 459| 5.26k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 460| 5.26k| std::pair ret = 461| 5.26k| sse_convert_latin1_to_utf16(buf, len, utf16_output); 462| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (462:7): [True: 0, False: 5.26k] ------------------ 463| 0| return 0; 464| 0| } 465| 5.26k| size_t converted_chars = ret.second - utf16_output; 466| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (466:7): [True: 460, False: 4.80k] ------------------ 467| 460| const size_t scalar_converted_chars = 468| 460| scalar::latin1_to_utf16::convert( 469| 460| ret.first, len - (ret.first - buf), ret.second); 470| 460| if (scalar_converted_chars == 0) { ------------------ | Branch (470:9): [True: 0, False: 460] ------------------ 471| 0| return 0; 472| 0| } 473| 460| converted_chars += scalar_converted_chars; 474| 460| } 475| 5.26k| return converted_chars; 476| 5.26k|} _ZNK7simdutf8westmere14implementation25convert_latin1_to_utf16beEPKcmPDs: 479| 5.26k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 480| 5.26k| std::pair ret = 481| 5.26k| sse_convert_latin1_to_utf16(buf, len, utf16_output); 482| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (482:7): [True: 0, False: 5.26k] ------------------ 483| 0| return 0; 484| 0| } 485| 5.26k| size_t converted_chars = ret.second - utf16_output; 486| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (486:7): [True: 460, False: 4.80k] ------------------ 487| 460| const size_t scalar_converted_chars = 488| 460| scalar::latin1_to_utf16::convert( 489| 460| ret.first, len - (ret.first - buf), ret.second); 490| 460| if (scalar_converted_chars == 0) { ------------------ | Branch (490:9): [True: 0, False: 460] ------------------ 491| 0| return 0; 492| 0| } 493| 460| converted_chars += scalar_converted_chars; 494| 460| } 495| 5.26k| return converted_chars; 496| 5.26k|} _ZNK7simdutf8westmere14implementation23convert_latin1_to_utf32EPKcmPDi: 501| 5.26k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 502| 5.26k| std::pair ret = 503| 5.26k| sse_convert_latin1_to_utf32(buf, len, utf32_output); 504| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (504:7): [True: 0, False: 5.26k] ------------------ 505| 0| return 0; 506| 0| } 507| 5.26k| size_t converted_chars = ret.second - utf32_output; 508| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (508:7): [True: 460, False: 4.80k] ------------------ 509| 460| const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert( 510| 460| ret.first, len - (ret.first - buf), ret.second); 511| 460| if (scalar_converted_chars == 0) { ------------------ | Branch (511:9): [True: 0, False: 460] ------------------ 512| 0| return 0; 513| 0| } 514| 460| converted_chars += scalar_converted_chars; 515| 460| } 516| 5.26k| return converted_chars; 517| 5.26k|} _ZNK7simdutf8westmere14implementation22convert_utf8_to_latin1EPKcmPc: 522| 10.5k| const char *buf, size_t len, char *latin1_output) const noexcept { 523| 10.5k| utf8_to_latin1::validating_transcoder converter; 524| 10.5k| return converter.convert(buf, len, latin1_output); 525| 10.5k|} _ZNK7simdutf8westmere14implementation23convert_utf8_to_utf16leEPKcmPDs: 541| 10.4k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 542| 10.4k| utf8_to_utf16::validating_transcoder converter; 543| 10.4k| return converter.convert(buf, len, utf16_output); 544| 10.4k|} _ZNK7simdutf8westmere14implementation23convert_utf8_to_utf16beEPKcmPDs: 547| 10.4k| const char *buf, size_t len, char16_t *utf16_output) const noexcept { 548| 10.4k| utf8_to_utf16::validating_transcoder converter; 549| 10.4k| return converter.convert(buf, len, utf16_output); 550| 10.4k|} _ZNK7simdutf8westmere14implementation21convert_utf8_to_utf32EPKcmPDi: 580| 5.26k| const char *buf, size_t len, char32_t *utf32_output) const noexcept { 581| 5.26k| utf8_to_utf32::validating_transcoder converter; 582| 5.26k| return converter.convert(buf, len, utf32_output); 583| 5.26k|} _ZNK7simdutf8westmere14implementation25convert_utf16le_to_latin1EPKDsmPc: 599| 5.26k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 600| 5.26k| std::pair ret = 601| 5.26k| sse_convert_utf16_to_latin1(buf, len, latin1_output); 602| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (602:7): [True: 0, False: 5.26k] ------------------ 603| 0| return 0; 604| 0| } 605| 5.26k| size_t saved_bytes = ret.second - latin1_output; 606| | 607| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (607:7): [True: 430, False: 4.83k] ------------------ 608| 430| const size_t scalar_saved_bytes = 609| 430| scalar::utf16_to_latin1::convert( 610| 430| ret.first, len - (ret.first - buf), ret.second); 611| 430| if (scalar_saved_bytes == 0) { ------------------ | Branch (611:9): [True: 0, False: 430] ------------------ 612| 0| return 0; 613| 0| } 614| 430| saved_bytes += scalar_saved_bytes; 615| 430| } 616| 5.26k| return saved_bytes; 617| 5.26k|} _ZNK7simdutf8westmere14implementation25convert_utf16be_to_latin1EPKDsmPc: 620| 5.26k| const char16_t *buf, size_t len, char *latin1_output) const noexcept { 621| 5.26k| std::pair ret = 622| 5.26k| sse_convert_utf16_to_latin1(buf, len, latin1_output); 623| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (623:7): [True: 0, False: 5.26k] ------------------ 624| 0| return 0; 625| 0| } 626| 5.26k| size_t saved_bytes = ret.second - latin1_output; 627| | 628| 5.26k| if (ret.first != buf + len) { ------------------ | Branch (628:7): [True: 430, False: 4.83k] ------------------ 629| 430| const size_t scalar_saved_bytes = 630| 430| scalar::utf16_to_latin1::convert( 631| 430| ret.first, len - (ret.first - buf), ret.second); 632| 430| if (scalar_saved_bytes == 0) { ------------------ | Branch (632:9): [True: 0, False: 430] ------------------ 633| 0| return 0; 634| 0| } 635| 430| saved_bytes += scalar_saved_bytes; 636| 430| } 637| 5.26k| return saved_bytes; 638| 5.26k|} _ZNK7simdutf8westmere14implementation23convert_utf16le_to_utf8EPKDsmPc: 709| 10.1k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 710| 10.1k| std::pair ret = 711| 10.1k| sse_convert_utf16_to_utf8(buf, len, utf8_output); 712| 10.1k| if (ret.first == nullptr) { ------------------ | Branch (712:7): [True: 70, False: 10.0k] ------------------ 713| 70| return 0; 714| 70| } 715| 10.0k| size_t saved_bytes = ret.second - utf8_output; 716| 10.0k| if (ret.first != buf + len) { ------------------ | Branch (716:7): [True: 960, False: 9.10k] ------------------ 717| 960| const size_t scalar_saved_bytes = 718| 960| scalar::utf16_to_utf8::convert( 719| 960| ret.first, len - (ret.first - buf), ret.second); 720| 960| if (scalar_saved_bytes == 0) { ------------------ | Branch (720:9): [True: 37, False: 923] ------------------ 721| 37| return 0; 722| 37| } 723| 923| saved_bytes += scalar_saved_bytes; 724| 923| } 725| 10.0k| return saved_bytes; 726| 10.0k|} _ZNK7simdutf8westmere14implementation23convert_utf16be_to_utf8EPKDsmPc: 729| 10.1k| const char16_t *buf, size_t len, char *utf8_output) const noexcept { 730| 10.1k| std::pair ret = 731| 10.1k| sse_convert_utf16_to_utf8(buf, len, utf8_output); 732| 10.1k| if (ret.first == nullptr) { ------------------ | Branch (732:7): [True: 69, False: 10.0k] ------------------ 733| 69| return 0; 734| 69| } 735| 10.0k| size_t saved_bytes = ret.second - utf8_output; 736| 10.0k| if (ret.first != buf + len) { ------------------ | Branch (736:7): [True: 951, False: 9.11k] ------------------ 737| 951| const size_t scalar_saved_bytes = 738| 951| scalar::utf16_to_utf8::convert( 739| 951| ret.first, len - (ret.first - buf), ret.second); 740| 951| if (scalar_saved_bytes == 0) { ------------------ | Branch (740:9): [True: 31, False: 920] ------------------ 741| 31| return 0; 742| 31| } 743| 920| saved_bytes += scalar_saved_bytes; 744| 920| } 745| 10.0k| return saved_bytes; 746| 10.0k|} _ZNK7simdutf8westmere14implementation23convert_utf32_to_latin1EPKDimPc: 817| 5.26k| const char32_t *buf, size_t len, char *latin1_output) const noexcept { 818| 5.26k| std::pair ret = 819| 5.26k| sse_convert_utf32_to_latin1(buf, len, latin1_output); 820| 5.26k| if (ret.first == nullptr) { ------------------ | Branch (820:7): [True: 0, False: 5.26k] ------------------ 821| 0| return 0; 822| 0| } 823| 5.26k| size_t saved_bytes = ret.second - latin1_output; 824| | // if (ret.first != buf + len) { 825| 5.26k| if (ret.first < buf + len) { ------------------ | Branch (825:7): [True: 460, False: 4.80k] ------------------ 826| 460| const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert( 827| 460| ret.first, len - (ret.first - buf), ret.second); 828| 460| if (scalar_saved_bytes == 0) { ------------------ | Branch (828:9): [True: 0, False: 460] ------------------ 829| 0| return 0; 830| 0| } 831| 460| saved_bytes += scalar_saved_bytes; 832| 460| } 833| 5.26k| return saved_bytes; 834| 5.26k|} _ZNK7simdutf8westmere14implementation21convert_utf32_to_utf8EPKDimPc: 868| 4.86k| const char32_t *buf, size_t len, char *utf8_output) const noexcept { 869| 4.86k| std::pair ret = 870| 4.86k| sse_convert_utf32_to_utf8(buf, len, utf8_output); 871| 4.86k| if (ret.first == nullptr) { ------------------ | Branch (871:7): [True: 0, False: 4.86k] ------------------ 872| 0| return 0; 873| 0| } 874| 4.86k| size_t saved_bytes = ret.second - utf8_output; 875| 4.86k| if (ret.first != buf + len) { ------------------ | Branch (875:7): [True: 701, False: 4.16k] ------------------ 876| 701| const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert( 877| 701| ret.first, len - (ret.first - buf), ret.second); 878| 701| if (scalar_saved_bytes == 0) { ------------------ | Branch (878:9): [True: 0, False: 701] ------------------ 879| 0| return 0; 880| 0| } 881| 701| saved_bytes += scalar_saved_bytes; 882| 701| } 883| 4.86k| return saved_bytes; 884| 4.86k|} _ZNK7simdutf8westmere14implementation10count_utf8EPKcm: 1144| 10.5k|implementation::count_utf8(const char *input, size_t length) const noexcept { 1145| 10.5k| return utf8::count_code_points_bytemask(input, length); 1146| 10.5k|} _ZNK7simdutf8westmere14implementation23latin1_length_from_utf8EPKcm: 1151| 10.5k| const char *buf, size_t len) const noexcept { 1152| 10.5k| return count_utf8(buf, len); 1153| 10.5k|} _ZNK7simdutf8westmere14implementation24utf8_length_from_utf16leEPKDsm: 1158| 10.1k| const char16_t *input, size_t length) const noexcept { 1159| 10.1k| return utf16::utf8_length_from_utf16_bytemask(input, 1160| 10.1k| length); 1161| 10.1k|} _ZNK7simdutf8westmere14implementation24utf8_length_from_utf16beEPKDsm: 1164| 10.1k| const char16_t *input, size_t length) const noexcept { 1165| 10.1k| return utf16::utf8_length_from_utf16_bytemask(input, length); 1166| 10.1k|} _ZNK7simdutf8westmere14implementation23utf8_length_from_latin1EPKcm: 1171| 5.55k| const char *input, size_t len) const noexcept { 1172| 5.55k| const uint8_t *str = reinterpret_cast(input); 1173| 5.55k| size_t answer = len / sizeof(__m128i) * sizeof(__m128i); 1174| 5.55k| size_t i = 0; 1175| 5.55k| if (answer >= 2048) { // long strings optimization ------------------ | Branch (1175:7): [True: 0, False: 5.55k] ------------------ 1176| 0| __m128i two_64bits = _mm_setzero_si128(); 1177| 0| while (i + sizeof(__m128i) <= len) { ------------------ | Branch (1177:12): [True: 0, False: 0] ------------------ 1178| 0| __m128i runner = _mm_setzero_si128(); 1179| 0| size_t iterations = (len - i) / sizeof(__m128i); 1180| 0| if (iterations > 255) { ------------------ | Branch (1180:11): [True: 0, False: 0] ------------------ 1181| 0| iterations = 255; 1182| 0| } 1183| 0| size_t max_i = i + iterations * sizeof(__m128i) - sizeof(__m128i); 1184| 0| for (; i + 4 * sizeof(__m128i) <= max_i; i += 4 * sizeof(__m128i)) { ------------------ | Branch (1184:14): [True: 0, False: 0] ------------------ 1185| 0| __m128i input1 = _mm_loadu_si128((const __m128i *)(str + i)); 1186| 0| __m128i input2 = 1187| 0| _mm_loadu_si128((const __m128i *)(str + i + sizeof(__m128i))); 1188| 0| __m128i input3 = 1189| 0| _mm_loadu_si128((const __m128i *)(str + i + 2 * sizeof(__m128i))); 1190| 0| __m128i input4 = 1191| 0| _mm_loadu_si128((const __m128i *)(str + i + 3 * sizeof(__m128i))); 1192| 0| __m128i input12 = 1193| 0| _mm_add_epi8(_mm_cmpgt_epi8(_mm_setzero_si128(), input1), 1194| 0| _mm_cmpgt_epi8(_mm_setzero_si128(), input2)); 1195| 0| __m128i input34 = 1196| 0| _mm_add_epi8(_mm_cmpgt_epi8(_mm_setzero_si128(), input3), 1197| 0| _mm_cmpgt_epi8(_mm_setzero_si128(), input4)); 1198| 0| __m128i input1234 = _mm_add_epi8(input12, input34); 1199| 0| runner = _mm_sub_epi8(runner, input1234); 1200| 0| } 1201| 0| for (; i <= max_i; i += sizeof(__m128i)) { ------------------ | Branch (1201:14): [True: 0, False: 0] ------------------ 1202| 0| __m128i more_input = _mm_loadu_si128((const __m128i *)(str + i)); 1203| 0| runner = _mm_sub_epi8(runner, 1204| 0| _mm_cmpgt_epi8(_mm_setzero_si128(), more_input)); 1205| 0| } 1206| 0| two_64bits = 1207| 0| _mm_add_epi64(two_64bits, _mm_sad_epu8(runner, _mm_setzero_si128())); 1208| 0| } 1209| 0| answer += 1210| 0| _mm_extract_epi64(two_64bits, 0) + _mm_extract_epi64(two_64bits, 1); 1211| 5.55k| } else if (answer > 0) { // short string optimization ------------------ | Branch (1211:14): [True: 792, False: 4.76k] ------------------ 1212| 7.13k| for (; i + 2 * sizeof(__m128i) <= len; i += 2 * sizeof(__m128i)) { ------------------ | Branch (1212:12): [True: 6.34k, False: 792] ------------------ 1213| 6.34k| __m128i latin = _mm_loadu_si128((const __m128i *)(input + i)); 1214| 6.34k| uint16_t non_ascii = (uint16_t)_mm_movemask_epi8(latin); 1215| 6.34k| answer += count_ones(non_ascii); 1216| 6.34k| latin = _mm_loadu_si128((const __m128i *)(input + i) + 1); 1217| 6.34k| non_ascii = (uint16_t)_mm_movemask_epi8(latin); 1218| 6.34k| answer += count_ones(non_ascii); 1219| 6.34k| } 1220| 1.01k| for (; i + sizeof(__m128i) <= len; i += sizeof(__m128i)) { ------------------ | Branch (1220:12): [True: 221, False: 792] ------------------ 1221| 221| __m128i latin = _mm_loadu_si128((const __m128i *)(input + i)); 1222| 221| uint16_t non_ascii = (uint16_t)_mm_movemask_epi8(latin); 1223| 221| answer += count_ones(non_ascii); 1224| 221| } 1225| 792| } 1226| 5.55k| return answer + scalar::latin1::utf8_length_from_latin1( 1227| 5.55k| reinterpret_cast(str + i), len - i); 1228| 5.55k|} _ZNK7simdutf8westmere14implementation22utf16_length_from_utf8EPKcm: 1245| 20.8k| const char *input, size_t length) const noexcept { 1246| 20.8k| return utf8::utf16_length_from_utf8_bytemask(input, length); 1247| 20.8k|} _ZNK7simdutf8westmere14implementation22utf8_length_from_utf32EPKDim: 1292| 4.86k| const char32_t *input, size_t length) const noexcept { 1293| 4.86k| return utf32::utf8_length_from_utf32(input, length); 1294| 4.86k|} _ZNK7simdutf8westmere14implementation22utf32_length_from_utf8EPKcm: 1320| 5.26k| const char *input, size_t length) const noexcept { 1321| 5.26k| return utf8::count_code_points(input, length); 1322| 5.26k|} _ZNK7simdutf8westmere14implementation16base64_to_binaryEPKcmPcNS_14base64_optionsENS_27last_chunk_handling_optionsE: 1328| 15.3k| last_chunk_handling_options last_chunk_options) const noexcept { 1329| 15.3k| if (options & base64_default_or_url) { ------------------ | Branch (1329:7): [True: 0, False: 15.3k] ------------------ 1330| 0| if (options == base64_options::base64_default_or_url_accept_garbage) { ------------------ | Branch (1330:9): [True: 0, False: 0] ------------------ 1331| 0| return base64::compress_decode_base64( 1332| 0| output, input, length, options, last_chunk_options); 1333| 0| } else { 1334| 0| return base64::compress_decode_base64( 1335| 0| output, input, length, options, last_chunk_options); 1336| 0| } 1337| 15.3k| } else if (options & base64_url) { ------------------ | Branch (1337:14): [True: 0, False: 15.3k] ------------------ 1338| 0| if (options == base64_options::base64_url_accept_garbage) { ------------------ | Branch (1338:9): [True: 0, False: 0] ------------------ 1339| 0| return base64::compress_decode_base64( 1340| 0| output, input, length, options, last_chunk_options); 1341| 0| } else { 1342| 0| return base64::compress_decode_base64( 1343| 0| output, input, length, options, last_chunk_options); 1344| 0| } 1345| 15.3k| } else { 1346| 15.3k| if (options == base64_options::base64_default_accept_garbage) { ------------------ | Branch (1346:9): [True: 0, False: 15.3k] ------------------ 1347| 0| return base64::compress_decode_base64( 1348| 0| output, input, length, options, last_chunk_options); 1349| 15.3k| } else { 1350| 15.3k| return base64::compress_decode_base64( 1351| 15.3k| output, input, length, options, last_chunk_options); 1352| 15.3k| } 1353| 15.3k| } 1354| 15.3k|} _ZNK7simdutf8westmere14implementation16binary_to_base64EPKcmPcNS_14base64_optionsE: 1448| 14.8k| base64_options options) const noexcept { 1449| 14.8k| if (options & base64_url) { ------------------ | Branch (1449:7): [True: 0, False: 14.8k] ------------------ 1450| 0| return encode_base64(output, input, length, options); 1451| 14.8k| } else { 1452| 14.8k| return encode_base64(output, input, length, options); 1453| 14.8k| } 1454| 14.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18is_asciiERKNS1_4simd8simd8x64IhEE: 13| 35.4k|simdutf_really_inline bool is_ascii(const simd8x64 &input) { 14| 35.4k| return input.reduce_or().is_ascii(); 15| 35.4k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_124must_be_2_3_continuationENS1_4simd5simd8IhEES4_: 22| 65.6k| const simd8 prev3) { 23| 65.6k| simd8 is_third_byte = 24| 65.6k| prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be >= 0x80 25| 65.6k| simd8 is_fourth_byte = 26| 65.6k| prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be >= 0x80 27| 65.6k| return simd8(is_third_byte | is_fourth_byte); 28| 65.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18internal8westmere26write_v_u16_11bits_to_utf8EDv2_xRPcS4_S4_: 57| 7.89k| const __m128i v_ff80) { 58| | // no bits set above 7th bit 59| 7.89k| const __m128i one_byte_bytemask = 60| 7.89k| _mm_cmpeq_epi16(_mm_and_si128(v_u16, v_ff80), v_0000); 61| 7.89k| const uint16_t one_byte_bitmask = 62| 7.89k| static_cast(_mm_movemask_epi8(one_byte_bytemask)); 63| | 64| 7.89k| write_v_u16_11bits_to_utf8(v_u16, utf8_output, one_byte_bytemask, 65| 7.89k| one_byte_bitmask); 66| 7.89k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18internal8westmere26write_v_u16_11bits_to_utf8EDv2_xRPcS4_t: 10| 11.8k| const uint16_t one_byte_bitmask) { 11| | // 0b1100_0000_1000_0000 12| 11.8k| const __m128i v_c080 = _mm_set1_epi16((int16_t)0xc080); 13| | // 0b0001_1111_0000_0000 14| 11.8k| const __m128i v_1f00 = _mm_set1_epi16((int16_t)0x1f00); 15| | // 0b0000_0000_0011_1111 16| 11.8k| 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| 11.8k| const __m128i t0 = _mm_slli_epi16(v_u16, 2); 24| | // t1 = [000a|aaaa|0000|0000] 25| 11.8k| const __m128i t1 = _mm_and_si128(t0, v_1f00); 26| | // t2 = [0000|0000|00bb|bbbb] 27| 11.8k| const __m128i t2 = _mm_and_si128(v_u16, v_003f); 28| | // t3 = [000a|aaaa|00bb|bbbb] 29| 11.8k| const __m128i t3 = _mm_or_si128(t1, t2); 30| | // t4 = [110a|aaaa|10bb|bbbb] 31| 11.8k| const __m128i t4 = _mm_or_si128(t3, v_c080); 32| | 33| | // 2. merge ASCII and 2-byte codewords 34| 11.8k| 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| 11.8k| const uint16_t m0 = one_byte_bitmask & 0x5555; // m0 = 0h0g0f0e0d0c0b0a 40| 11.8k| const uint16_t m1 = static_cast(m0 >> 7); // m1 = 00000000h0g0f0e0 41| 11.8k| const uint8_t m2 = static_cast((m0 | m1) & 0xff); // m2 = hdgcfbea 42| | // 4. pack the bytes 43| 11.8k| const uint8_t *row = 44| 11.8k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0]; 45| 11.8k| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 46| 11.8k| const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle); 47| | 48| | // 5. store bytes 49| 11.8k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 50| | 51| | // 6. adjust pointers 52| 11.8k| utf8_output += row[0]; 53| 11.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block64C2EPKc: 408| 3.86k| simdutf_really_inline block64(const char *src) { 409| 3.86k| chunks[0] = _mm_loadu_si128(reinterpret_cast(src)); 410| 3.86k| chunks[1] = _mm_loadu_si128(reinterpret_cast(src + 16)); 411| 3.86k| chunks[2] = _mm_loadu_si128(reinterpret_cast(src + 32)); 412| 3.86k| chunks[3] = _mm_loadu_si128(reinterpret_cast(src + 48)); 413| 3.86k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6414compress_blockEmPc: 449| 724| simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) { 450| 724| if (is_power_of_two(mask)) { ------------------ | Branch (450:9): [True: 236, False: 488] ------------------ 451| 236| return compress_block_single(mask, output); 452| 236| } 453| | 454| 488| uint64_t nmask = ~mask; 455| 488| compress(chunks[0], uint16_t(mask), output); 456| 488| compress(chunks[1], uint16_t(mask >> 16), 457| 488| output + count_ones(nmask & 0xFFFF)); 458| 488| compress(chunks[2], uint16_t(mask >> 32), 459| 488| output + count_ones(nmask & 0xFFFFFFFF)); 460| 488| compress(chunks[3], uint16_t(mask >> 48), 461| 488| output + count_ones(nmask & 0xFFFFFFFFFFFFULL)); 462| 488| return count_ones(nmask); 463| 724| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6421compress_block_singleEmPc: 467| 236| char *output) { 468| 236| const size_t pos64 = trailing_zeroes(mask); 469| 236| const int8_t pos = pos64 & 0xf; 470| 236| switch (pos64 >> 4) { ------------------ | Branch (470:13): [True: 236, False: 0] ------------------ 471| 67| case 0b00: { ------------------ | Branch (471:5): [True: 67, False: 169] ------------------ 472| 67| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 473| 67| const __m128i v1 = 474| 67| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 475| 67| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 476| 67| const __m128i sh = _mm_sub_epi8(v1, v2); 477| 67| const __m128i compressed = _mm_shuffle_epi8(chunks[0], sh); 478| | 479| 67| _mm_storeu_si128((__m128i *)(output + 0 * 16), compressed); 480| 67| _mm_storeu_si128((__m128i *)(output + 1 * 16 - 1), chunks[1]); 481| 67| _mm_storeu_si128((__m128i *)(output + 2 * 16 - 1), chunks[2]); 482| 67| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]); 483| 67| } break; 484| 65| case 0b01: { ------------------ | Branch (484:5): [True: 65, False: 171] ------------------ 485| 65| _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]); 486| | 487| 65| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 488| 65| const __m128i v1 = 489| 65| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 490| 65| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 491| 65| const __m128i sh = _mm_sub_epi8(v1, v2); 492| 65| const __m128i compressed = _mm_shuffle_epi8(chunks[1], sh); 493| | 494| 65| _mm_storeu_si128((__m128i *)(output + 1 * 16), compressed); 495| 65| _mm_storeu_si128((__m128i *)(output + 2 * 16 - 1), chunks[2]); 496| 65| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]); 497| 65| } break; 498| 55| case 0b10: { ------------------ | Branch (498:5): [True: 55, False: 181] ------------------ 499| 55| _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]); 500| 55| _mm_storeu_si128((__m128i *)(output + 1 * 16), chunks[1]); 501| | 502| 55| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 503| 55| const __m128i v1 = 504| 55| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 505| 55| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 506| 55| const __m128i sh = _mm_sub_epi8(v1, v2); 507| 55| const __m128i compressed = _mm_shuffle_epi8(chunks[2], sh); 508| | 509| 55| _mm_storeu_si128((__m128i *)(output + 2 * 16), compressed); 510| 55| _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]); 511| 55| } break; 512| 49| case 0b11: { ------------------ | Branch (512:5): [True: 49, False: 187] ------------------ 513| 49| _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]); 514| 49| _mm_storeu_si128((__m128i *)(output + 1 * 16), chunks[1]); 515| 49| _mm_storeu_si128((__m128i *)(output + 2 * 16), chunks[2]); 516| | 517| 49| const __m128i v0 = _mm_set1_epi8(char(pos - 1)); 518| 49| const __m128i v1 = 519| 49| _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); 520| 49| const __m128i v2 = _mm_cmpgt_epi8(v1, v0); 521| 49| const __m128i sh = _mm_sub_epi8(v1, v2); 522| 49| const __m128i compressed = _mm_shuffle_epi8(chunks[3], sh); 523| | 524| 49| _mm_storeu_si128((__m128i *)(output + 3 * 16), compressed); 525| 49| } break; 526| 236| } 527| | 528| 236| return 63; 529| 236| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_18compressEDv2_xtPc: 329| 1.95k| char *output) { 330| 1.95k| if (mask == 0) { ------------------ | Branch (330:7): [True: 718, False: 1.23k] ------------------ 331| 718| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data); 332| 718| return; 333| 718| } 334| | 335| | // this particular implementation was inspired by work done by @animetosho 336| | // we do it in two steps, first 8 bytes and then second 8 bytes 337| 1.23k| uint8_t mask1 = uint8_t(mask); // least significant 8 bits 338| 1.23k| uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits 339| | // next line just loads the 64-bit values thintable_epi8[mask1] and 340| | // thintable_epi8[mask2] into a 128-bit register, using only 341| | // two instructions on most compilers. 342| | 343| 1.23k| __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2], 344| 1.23k| tables::base64::thintable_epi8[mask1]); 345| | // we increment by 0x08 the second half of the mask 346| 1.23k| shufmask = 347| 1.23k| _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0)); 348| | // this is the version "nearly pruned" 349| 1.23k| __m128i pruned = _mm_shuffle_epi8(data, shufmask); 350| | // we still need to put the two halves together. 351| | // we compute the popcount of the first half: 352| 1.23k| int pop1 = tables::base64::BitsSetTable256mul2[mask1]; 353| | // then load the corresponding mask, what it does is to write 354| | // only the first pop1 bytes from the first 8 bytes, and then 355| | // it fills in with the bytes from the second 8 bytes + some filling 356| | // at the end. 357| 1.23k| __m128i compactmask = _mm_loadu_si128(reinterpret_cast( 358| 1.23k| tables::base64::pshufb_combine_table + pop1 * 8)); 359| 1.23k| __m128i answer = _mm_shuffle_epi8(pruned, compactmask); 360| 1.23k| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer); 361| 1.23k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6410copy_blockEPc: 441| 94| simdutf_really_inline void copy_block(char *output) { 442| 94| _mm_storeu_si128(reinterpret_cast<__m128i *>(output), chunks[0]); 443| 94| _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 16), chunks[1]); 444| 94| _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 32), chunks[2]); 445| 94| _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 48), chunks[3]); 446| 94| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6424base64_decode_block_safeEPc: 659| 131| simdutf_really_inline void base64_decode_block_safe(char *out) { 660| 131| base64_decode(out, chunks[0]); 661| 131| base64_decode(out + 12, chunks[1]); 662| 131| base64_decode(out + 24, chunks[2]); 663| 131| char buffer[16]; 664| 131| base64_decode(buffer, chunks[3]); 665| 131| std::memcpy(out + 36, buffer, 12); 666| 131| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113base64_decodeEPcDv2_x: 363| 13.2k|static simdutf_really_inline void base64_decode(char *out, __m128i str) { 364| | // credit: aqrit 365| | 366| 13.2k| const __m128i pack_shuffle = 367| 13.2k| _mm_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1); 368| | 369| 13.2k| const __m128i t0 = _mm_maddubs_epi16(str, _mm_set1_epi32(0x01400140)); 370| 13.2k| const __m128i t1 = _mm_madd_epi16(t0, _mm_set1_epi32(0x00011000)); 371| 13.2k| const __m128i t2 = _mm_shuffle_epi8(t1, pack_shuffle); 372| | // Store the output: 373| | // this writes 16 bytes, but we only need 12. 374| 13.2k| _mm_storeu_si128((__m128i *)out, t2); 375| 13.2k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6419base64_decode_blockEPc: 651| 2.67k| simdutf_really_inline void base64_decode_block(char *out) { 652| 2.67k| base64_decode(out, chunks[0]); 653| 2.67k| base64_decode(out + 12, chunks[1]); 654| 2.67k| base64_decode(out + 24, chunks[2]); 655| 2.67k| base64_decode(out + 36, chunks[3]); 656| 2.67k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_119base64_decode_blockEPcPKc: 378| 489|static inline void base64_decode_block(char *out, const char *src) { 379| 489| base64_decode(out, _mm_loadu_si128(reinterpret_cast(src))); 380| 489| base64_decode(out + 12, 381| 489| _mm_loadu_si128(reinterpret_cast(src + 16))); 382| 489| base64_decode(out + 24, 383| 489| _mm_loadu_si128(reinterpret_cast(src + 32))); 384| 489| base64_decode(out + 36, 385| 489| _mm_loadu_si128(reinterpret_cast(src + 48))); 386| 489|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_124base64_decode_block_safeEPcPKc: 388| 20|static inline void base64_decode_block_safe(char *out, const char *src) { 389| 20| base64_decode(out, _mm_loadu_si128(reinterpret_cast(src))); 390| 20| base64_decode(out + 12, 391| 20| _mm_loadu_si128(reinterpret_cast(src + 16))); 392| 20| base64_decode(out + 24, 393| 20| _mm_loadu_si128(reinterpret_cast(src + 32))); 394| 20| char buffer[16]; 395| 20| base64_decode(buffer, 396| 20| _mm_loadu_si128(reinterpret_cast(src + 48))); 397| 20| std::memcpy(out + 36, buffer, 12); 398| 20|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEmPm: 533| 3.86k| simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) { 534| 3.86k| uint32_t err0 = 0; 535| 3.86k| uint32_t err1 = 0; 536| 3.86k| uint32_t err2 = 0; 537| 3.86k| uint32_t err3 = 0; 538| 3.86k| uint64_t m0 = to_base64_mask( 539| 3.86k| &chunks[0], &err0); 540| 3.86k| uint64_t m1 = to_base64_mask( 541| 3.86k| &chunks[1], &err1); 542| 3.86k| uint64_t m2 = to_base64_mask( 543| 3.86k| &chunks[2], &err2); 544| 3.86k| uint64_t m3 = to_base64_mask( 545| 3.86k| &chunks[3], &err3); 546| 3.86k| if (!ignore_garbage) { ------------------ | Branch (546:9): [True: 3.86k, Folded] ------------------ 547| 3.86k| *error = (err0) | ((uint64_t)err1 << 16) | ((uint64_t)err2 << 32) | 548| 3.86k| ((uint64_t)err3 << 48); 549| 3.86k| } 550| 3.86k| return m0 | (m1 << 16) | (m2 << 32) | (m3 << 48); 551| 3.86k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_17block6414to_base64_maskILb0ELb0ELb0EEEtPDv2_xPj: 555| 15.4k| simdutf_really_inline uint16_t to_base64_mask(__m128i *src, uint32_t *error) { 556| 15.4k| const __m128i ascii_space_tbl = 557| 15.4k| _mm_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa, 558| 15.4k| 0x0, 0xc, 0xd, 0x0, 0x0); 559| | // credit: aqrit 560| 15.4k| __m128i delta_asso; 561| 15.4k| if (default_or_url) { ------------------ | Branch (561:9): [Folded, False: 15.4k] ------------------ 562| 0| delta_asso = 563| 0| _mm_setr_epi8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 564| 0| 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x16); 565| 15.4k| } else if (base64_url) { ------------------ | Branch (565:16): [Folded, False: 15.4k] ------------------ 566| 0| delta_asso = _mm_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 567| 0| 0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF); 568| 15.4k| } else { 569| 15.4k| delta_asso = 570| 15.4k| _mm_setr_epi8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 571| 15.4k| 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F); 572| 15.4k| } 573| 15.4k| __m128i delta_values; 574| 15.4k| if (default_or_url) { ------------------ | Branch (574:9): [Folded, False: 15.4k] ------------------ 575| 0| delta_values = _mm_setr_epi8( 576| 0| uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0x13), 577| 0| uint8_t(0x04), uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9), 578| 0| uint8_t(0xB9), uint8_t(0x00), uint8_t(0xFF), uint8_t(0x11), 579| 0| uint8_t(0xFF), uint8_t(0xBF), uint8_t(0x10), uint8_t(0xB9)); 580| | 581| 15.4k| } else if (base64_url) { ------------------ | Branch (581:16): [Folded, False: 15.4k] ------------------ 582| 0| delta_values = _mm_setr_epi8(0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), 583| 0| uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9), 584| 0| 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF), 585| 0| uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9)); 586| 15.4k| } else { 587| 15.4k| delta_values = 588| 15.4k| _mm_setr_epi8(int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), 589| 15.4k| int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), 590| 15.4k| int8_t(0xB9), int8_t(0x00), int8_t(0x10), int8_t(0xC3), 591| 15.4k| int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9)); 592| 15.4k| } 593| 15.4k| __m128i check_asso; 594| 15.4k| if (default_or_url) { ------------------ | Branch (594:9): [Folded, False: 15.4k] ------------------ 595| 0| check_asso = 596| 0| _mm_setr_epi8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 597| 0| 0x01, 0x03, 0x07, 0x0B, 0x0E, 0x0B, 0x06); 598| 15.4k| } else if (base64_url) { ------------------ | Branch (598:16): [Folded, False: 15.4k] ------------------ 599| 0| check_asso = _mm_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 600| 0| 0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6); 601| 15.4k| } else { 602| 15.4k| check_asso = 603| 15.4k| _mm_setr_epi8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 604| 15.4k| 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F); 605| 15.4k| } 606| 15.4k| __m128i check_values; 607| 15.4k| if (default_or_url) { ------------------ | Branch (607:9): [Folded, False: 15.4k] ------------------ 608| 0| check_values = _mm_setr_epi8( 609| 0| uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 610| 0| uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xD5), uint8_t(0xA6), 611| 0| uint8_t(0xB5), uint8_t(0xA1), uint8_t(0x00), uint8_t(0x80), 612| 0| uint8_t(0x00), uint8_t(0x80), uint8_t(0x00), uint8_t(0x80)); 613| 15.4k| } else if (base64_url) { ------------------ | Branch (613:16): [Folded, False: 15.4k] ------------------ 614| 0| check_values = _mm_setr_epi8(uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), 615| 0| uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF), 616| 0| uint8_t(0xB6), uint8_t(0xA6), uint8_t(0xB5), 617| 0| uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, 618| 0| uint8_t(0x80), 0x0, uint8_t(0x80)); 619| 15.4k| } else { 620| 15.4k| check_values = 621| 15.4k| _mm_setr_epi8(int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), 622| 15.4k| int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), 623| 15.4k| int8_t(0xB5), int8_t(0x86), int8_t(0xD1), int8_t(0x80), 624| 15.4k| int8_t(0xB1), int8_t(0x80), int8_t(0x91), int8_t(0x80)); 625| 15.4k| } 626| 15.4k| const __m128i shifted = _mm_srli_epi32(*src, 3); 627| | 628| 15.4k| __m128i delta_hash = 629| 15.4k| _mm_avg_epu8(_mm_shuffle_epi8(delta_asso, *src), shifted); 630| 15.4k| if (default_or_url) { ------------------ | Branch (630:9): [Folded, False: 15.4k] ------------------ 631| 0| delta_hash = _mm_and_si128(delta_hash, _mm_set1_epi8(0xf)); 632| 0| } 633| 15.4k| const __m128i check_hash = 634| 15.4k| _mm_avg_epu8(_mm_shuffle_epi8(check_asso, *src), shifted); 635| | 636| 15.4k| const __m128i out = 637| 15.4k| _mm_adds_epi8(_mm_shuffle_epi8(delta_values, delta_hash), *src); 638| 15.4k| const __m128i chk = 639| 15.4k| _mm_adds_epi8(_mm_shuffle_epi8(check_values, check_hash), *src); 640| 15.4k| const int mask = _mm_movemask_epi8(chk); 641| 15.4k| if (!ignore_garbage && mask) { ------------------ | Branch (641:9): [True: 15.4k, Folded] | Branch (641:28): [True: 2.36k, False: 13.0k] ------------------ 642| 2.36k| __m128i ascii_space = 643| 2.36k| _mm_cmpeq_epi8(_mm_shuffle_epi8(ascii_space_tbl, *src), *src); 644| 2.36k| *error = (mask ^ _mm_movemask_epi8(ascii_space)); 645| 2.36k| } 646| 15.4k| *src = out; 647| 15.4k| return (uint16_t)mask; 648| 15.4k| } simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_113encode_base64ILb0EEEmPcPKcmNS_14base64_optionsE: 322| 14.8k| base64_options options) { 323| 14.8k| return encode_base64_impl(dst, src, srclen, options); 324| 14.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_118encode_base64_implILb0ELb0EEEmPcPKcmNS_14base64_optionsEm: 92| 14.8k| size_t line_length = simdutf::default_line_length) { 93| 14.8k| size_t offset = 0; 94| 14.8k| if (line_length < 4) { ------------------ | Branch (94:7): [True: 0, False: 14.8k] ------------------ 95| 0| line_length = 4; // We do not support line_length less than 4 96| 0| } 97| | // credit: Wojciech Muła 98| | // SSE (lookup: pshufb improved unrolled) 99| 14.8k| const uint8_t *input = (const uint8_t *)src; 100| | 101| 14.8k| uint8_t *out = (uint8_t *)dst; 102| 14.8k| const __m128i shuf = 103| 14.8k| _mm_set_epi8(10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1); 104| | 105| 14.8k| size_t i = 0; 106| 17.8k| for (; i + 52 <= srclen; i += 48) { ------------------ | Branch (106:10): [True: 3.03k, False: 14.8k] ------------------ 107| 3.03k| __m128i in0 = _mm_loadu_si128( 108| 3.03k| reinterpret_cast(input + i + 4 * 3 * 0)); 109| 3.03k| __m128i in1 = _mm_loadu_si128( 110| 3.03k| reinterpret_cast(input + i + 4 * 3 * 1)); 111| 3.03k| __m128i in2 = _mm_loadu_si128( 112| 3.03k| reinterpret_cast(input + i + 4 * 3 * 2)); 113| 3.03k| __m128i in3 = _mm_loadu_si128( 114| 3.03k| reinterpret_cast(input + i + 4 * 3 * 3)); 115| | 116| 3.03k| in0 = _mm_shuffle_epi8(in0, shuf); 117| 3.03k| in1 = _mm_shuffle_epi8(in1, shuf); 118| 3.03k| in2 = _mm_shuffle_epi8(in2, shuf); 119| 3.03k| in3 = _mm_shuffle_epi8(in3, shuf); 120| | 121| 3.03k| const __m128i t0_0 = _mm_and_si128(in0, _mm_set1_epi32(0x0fc0fc00)); 122| 3.03k| const __m128i t0_1 = _mm_and_si128(in1, _mm_set1_epi32(0x0fc0fc00)); 123| 3.03k| const __m128i t0_2 = _mm_and_si128(in2, _mm_set1_epi32(0x0fc0fc00)); 124| 3.03k| const __m128i t0_3 = _mm_and_si128(in3, _mm_set1_epi32(0x0fc0fc00)); 125| | 126| 3.03k| const __m128i t1_0 = _mm_mulhi_epu16(t0_0, _mm_set1_epi32(0x04000040)); 127| 3.03k| const __m128i t1_1 = _mm_mulhi_epu16(t0_1, _mm_set1_epi32(0x04000040)); 128| 3.03k| const __m128i t1_2 = _mm_mulhi_epu16(t0_2, _mm_set1_epi32(0x04000040)); 129| 3.03k| const __m128i t1_3 = _mm_mulhi_epu16(t0_3, _mm_set1_epi32(0x04000040)); 130| | 131| 3.03k| const __m128i t2_0 = _mm_and_si128(in0, _mm_set1_epi32(0x003f03f0)); 132| 3.03k| const __m128i t2_1 = _mm_and_si128(in1, _mm_set1_epi32(0x003f03f0)); 133| 3.03k| const __m128i t2_2 = _mm_and_si128(in2, _mm_set1_epi32(0x003f03f0)); 134| 3.03k| const __m128i t2_3 = _mm_and_si128(in3, _mm_set1_epi32(0x003f03f0)); 135| | 136| 3.03k| const __m128i t3_0 = _mm_mullo_epi16(t2_0, _mm_set1_epi32(0x01000010)); 137| 3.03k| const __m128i t3_1 = _mm_mullo_epi16(t2_1, _mm_set1_epi32(0x01000010)); 138| 3.03k| const __m128i t3_2 = _mm_mullo_epi16(t2_2, _mm_set1_epi32(0x01000010)); 139| 3.03k| const __m128i t3_3 = _mm_mullo_epi16(t2_3, _mm_set1_epi32(0x01000010)); 140| | 141| 3.03k| const __m128i input0 = _mm_or_si128(t1_0, t3_0); 142| 3.03k| const __m128i input1 = _mm_or_si128(t1_1, t3_1); 143| 3.03k| const __m128i input2 = _mm_or_si128(t1_2, t3_2); 144| 3.03k| const __m128i input3 = _mm_or_si128(t1_3, t3_3); 145| | 146| 3.03k| const __m128i t0 = lookup_pshufb_improved(input0); 147| 3.03k| const __m128i t1 = lookup_pshufb_improved(input1); 148| 3.03k| const __m128i t2 = lookup_pshufb_improved(input2); 149| 3.03k| const __m128i t3 = lookup_pshufb_improved(input3); 150| | 151| 3.03k| if (use_lines) { ------------------ | Branch (151:9): [Folded, False: 3.03k] ------------------ 152| 0| if (line_length >= 64) { // fast path ------------------ | Branch (152:11): [True: 0, False: 0] ------------------ 153| 0| if (offset + 64 > line_length) { ------------------ | Branch (153:13): [True: 0, False: 0] ------------------ 154| 0| size_t location_end = line_length - offset; 155| 0| size_t to_move = 64 - location_end; 156| 0| if (location_end < 16) { ------------------ | Branch (156:15): [True: 0, False: 0] ------------------ 157| | // We can store or extract store. See below. 158| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t0); 159| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 160| 0| insert_line_feed16(t0, location_end)); 161| 0| out[16] = static_cast(_mm_extract_epi8(t0, 15)); 162| 0| out += 17; 163| 0| } else { 164| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t0); 165| 0| out += 16; 166| 0| } 167| 0| if (location_end >= 16 && location_end < 32) { ------------------ | Branch (167:15): [True: 0, False: 0] | Branch (167:37): [True: 0, False: 0] ------------------ 168| | // We can store or extract store. See below. 169| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t1); 170| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 171| 0| insert_line_feed16(t1, location_end - 16)); 172| 0| out[16] = static_cast(_mm_extract_epi8(t1, 15)); 173| 0| out += 17; 174| 0| } else { 175| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t1); 176| 0| out += 16; 177| 0| } 178| 0| if (location_end >= 32 && location_end < 48) { ------------------ | Branch (178:15): [True: 0, False: 0] | Branch (178:37): [True: 0, False: 0] ------------------ 179| | // We can store or extract store. See below. 180| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t2); 181| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 182| 0| insert_line_feed16(t2, location_end - 32)); 183| 0| out[16] = static_cast(_mm_extract_epi8(t2, 15)); 184| 0| out += 17; 185| 0| } else { 186| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t2); 187| 0| out += 16; 188| 0| } 189| 0| if (location_end >= 48) { ------------------ | Branch (189:15): [True: 0, False: 0] ------------------ 190| | // We can store or extract store. See below. 191| | //_mm_storeu_si128(reinterpret_cast<__m128i *>(out+1), t3); 192| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), 193| 0| insert_line_feed16(t3, location_end - 48)); 194| 0| out[16] = static_cast(_mm_extract_epi8(t3, 15)); 195| 0| out += 17; 196| 0| } else { 197| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t3); 198| 0| out += 16; 199| 0| } 200| 0| offset = to_move; 201| 0| } else { 202| | 203| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t0); 204| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 16), t1); 205| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 32), t2); 206| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 48), t3); 207| 0| offset += 64; 208| 0| out += 64; 209| 0| } 210| 0| } else { // slow path 211| | // could be optimized 212| 0| alignas(64) uint8_t buffer[64]; 213| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer), t0); 214| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + 16), t1); 215| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + 32), t2); 216| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + 48), t3); 217| 0| std::memcpy(out, buffer, 64); 218| 0| size_t out_pos = 0; 219| 0| size_t local_offset = offset; 220| 0| for (size_t j = 0; j < 64;) { ------------------ | Branch (220:28): [True: 0, False: 0] ------------------ 221| 0| if (local_offset == line_length) { ------------------ | Branch (221:15): [True: 0, False: 0] ------------------ 222| 0| out[out_pos++] = '\n'; 223| 0| local_offset = 0; 224| 0| } 225| 0| out[out_pos++] = buffer[j++]; 226| 0| local_offset++; 227| 0| } 228| 0| offset = local_offset; 229| 0| out += out_pos; 230| 0| } 231| 3.03k| } else { 232| 3.03k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), t0); 233| 3.03k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 16), t1); 234| 3.03k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 32), t2); 235| 3.03k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out + 48), t3); 236| 3.03k| out += 64; 237| 3.03k| } 238| 3.03k| } 239| 15.8k| for (; i + 16 <= srclen; i += 12) { ------------------ | Branch (239:10): [True: 1.04k, False: 14.8k] ------------------ 240| | 241| 1.04k| __m128i in = _mm_loadu_si128(reinterpret_cast(input + i)); 242| | 243| | // bytes from groups A, B and C are needed in separate 32-bit lanes 244| | // in = [DDDD|CCCC|BBBB|AAAA] 245| | // 246| | // an input triplet has layout 247| | // [????????|ccdddddd|bbbbcccc|aaaaaabb] 248| | // byte 3 byte 2 byte 1 byte 0 -- byte 3 comes from the next 249| | // triplet 250| | // 251| | // shuffling changes the order of bytes: 1, 0, 2, 1 252| | // [bbbbcccc|ccdddddd|aaaaaabb|bbbbcccc] 253| | // ^^^^ ^^^^^^^^ ^^^^^^^^ ^^^^ 254| | // processed bits 255| 1.04k| in = _mm_shuffle_epi8(in, shuf); 256| | 257| | // unpacking 258| | 259| | // t0 = [0000cccc|cc000000|aaaaaa00|00000000] 260| 1.04k| const __m128i t0 = _mm_and_si128(in, _mm_set1_epi32(0x0fc0fc00)); 261| | // t1 = [00000000|00cccccc|00000000|00aaaaaa] 262| | // (c * (1 << 10), a * (1 << 6)) >> 16 (note: an unsigned 263| | // multiplication) 264| 1.04k| const __m128i t1 = _mm_mulhi_epu16(t0, _mm_set1_epi32(0x04000040)); 265| | 266| | // t2 = [00000000|00dddddd|000000bb|bbbb0000] 267| 1.04k| const __m128i t2 = _mm_and_si128(in, _mm_set1_epi32(0x003f03f0)); 268| | // t3 = [00dddddd|00000000|00bbbbbb|00000000]( 269| | // (d * (1 << 8), b * (1 << 4)) 270| 1.04k| const __m128i t3 = _mm_mullo_epi16(t2, _mm_set1_epi32(0x01000010)); 271| | 272| | // res = [00dddddd|00cccccc|00bbbbbb|00aaaaaa] = t1 | t3 273| 1.04k| const __m128i indices = _mm_or_si128(t1, t3); 274| | 275| 1.04k| const __m128i T0 = lookup_pshufb_improved(indices); 276| | 277| 1.04k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), T0); 278| | 279| 1.04k| if (use_lines) { ------------------ | Branch (279:9): [Folded, False: 1.04k] ------------------ 280| 0| if (line_length >= 16) { // fast path ------------------ | Branch (280:11): [True: 0, False: 0] ------------------ 281| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), T0); 282| 0| if (offset + 16 > line_length) { ------------------ | Branch (282:13): [True: 0, False: 0] ------------------ 283| 0| size_t location_end = line_length - offset; 284| 0| size_t to_move = 16 - location_end; 285| 0| std::memmove(out + location_end + 1, out + location_end, to_move); 286| 0| out[location_end] = '\n'; 287| 0| offset = to_move; 288| 0| out += 16 + 1; 289| 0| } else { 290| 0| offset += 16; 291| 0| out += 16; 292| 0| } 293| 0| } else { // slow path 294| | // could be optimized 295| 0| uint8_t buffer[16]; 296| 0| _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer), T0); 297| 0| size_t out_pos = 0; 298| 0| size_t local_offset = offset; 299| 0| for (size_t j = 0; j < 16;) { ------------------ | Branch (299:28): [True: 0, False: 0] ------------------ 300| 0| if (local_offset == line_length) { ------------------ | Branch (300:15): [True: 0, False: 0] ------------------ 301| 0| out[out_pos++] = '\n'; 302| 0| local_offset = 0; 303| 0| } 304| 0| out[out_pos++] = buffer[j++]; 305| 0| local_offset++; 306| 0| } 307| 0| offset = local_offset; 308| 0| out += out_pos; 309| 0| } 310| 1.04k| } else { 311| 1.04k| _mm_storeu_si128(reinterpret_cast<__m128i *>(out), T0); 312| 1.04k| out += 16; 313| 1.04k| } 314| 1.04k| } 315| 14.8k| return ((char *)out - (char *)dst) + 316| 14.8k| scalar::base64::tail_encode_base64_impl( 317| 14.8k| (char *)out, src + i, srclen - i, options, line_length, offset); 318| 14.8k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_122lookup_pshufb_improvedILb0EEEDv2_xS3_: 30| 13.1k|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| 13.1k| __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| 13.1k| const __m128i less = _mm_cmpgt_epi8(_mm_set1_epi8(26), input); 42| 13.1k| result = _mm_or_si128(result, _mm_and_si128(less, _mm_set1_epi8(13))); 43| | 44| 13.1k| __m128i shift_LUT; 45| 13.1k| if (base64_url) { ------------------ | Branch (45:7): [Folded, False: 13.1k] ------------------ 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| 13.1k| } else { 50| 13.1k| shift_LUT = _mm_setr_epi8('a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 51| 13.1k| '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, 52| 13.1k| '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0); 53| 13.1k| } 54| | 55| | // read shift 56| 13.1k| result = _mm_shuffle_epi8(shift_LUT, result); 57| | 58| 13.1k| return _mm_add_epi8(result, input); 59| 13.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_latin1_to_utf16ILNS_10endiannessE0EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.26k| char16_t *utf16_output) { 5| 5.26k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| 15.2k| for (size_t i = 0; i < rounded_len; i += 16) { ------------------ | Branch (6:22): [True: 9.97k, False: 5.26k] ------------------ 7| | // Load 16 Latin1 characters into a 128-bit register 8| 9.97k| __m128i in = 9| 9.97k| _mm_loadu_si128(reinterpret_cast(&latin1_input[i])); 10| 9.97k| __m128i out1 = big_endian ? _mm_unpacklo_epi8(_mm_setzero_si128(), in) ------------------ | Branch (10:20): [Folded, False: 9.97k] ------------------ 11| 9.97k| : _mm_unpacklo_epi8(in, _mm_setzero_si128()); 12| 9.97k| __m128i out2 = big_endian ? _mm_unpackhi_epi8(_mm_setzero_si128(), in) ------------------ | Branch (12:20): [Folded, False: 9.97k] ------------------ 13| 9.97k| : _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| 9.97k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i]), out1); 17| 9.97k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i + 8]), out2); 18| 9.97k| } 19| | // return pointers pointing to where we left off 20| 5.26k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 21| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_latin1_to_utf16ILNS_10endiannessE1EEENSt3__14pairIPKcPDsEES7_mS8_: 4| 5.26k| char16_t *utf16_output) { 5| 5.26k| size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 6| 15.2k| for (size_t i = 0; i < rounded_len; i += 16) { ------------------ | Branch (6:22): [True: 9.97k, False: 5.26k] ------------------ 7| | // Load 16 Latin1 characters into a 128-bit register 8| 9.97k| __m128i in = 9| 9.97k| _mm_loadu_si128(reinterpret_cast(&latin1_input[i])); 10| 9.97k| __m128i out1 = big_endian ? _mm_unpacklo_epi8(_mm_setzero_si128(), in) ------------------ | Branch (10:20): [True: 9.97k, Folded] ------------------ 11| 9.97k| : _mm_unpacklo_epi8(in, _mm_setzero_si128()); 12| 9.97k| __m128i out2 = big_endian ? _mm_unpackhi_epi8(_mm_setzero_si128(), in) ------------------ | Branch (12:20): [True: 9.97k, Folded] ------------------ 13| 9.97k| : _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| 9.97k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i]), out1); 17| 9.97k| _mm_storeu_si128(reinterpret_cast<__m128i *>(&utf16_output[i + 8]), out2); 18| 9.97k| } 19| | // return pointers pointing to where we left off 20| 5.26k| return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len); 21| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_latin1_to_utf32EPKcmPDi: 3| 5.26k| char32_t *utf32_output) { 4| 5.26k| const char *end = buf + len; 5| | 6| 15.2k| while (end - buf >= 16) { ------------------ | Branch (6:10): [True: 9.97k, False: 5.26k] ------------------ 7| | // Load 16 Latin1 characters (16 bytes) into a 128-bit register 8| 9.97k| __m128i in = _mm_loadu_si128((__m128i *)buf); 9| | 10| | // Shift input to process next 4 bytes 11| 9.97k| __m128i in_shifted1 = _mm_srli_si128(in, 4); 12| 9.97k| __m128i in_shifted2 = _mm_srli_si128(in, 8); 13| 9.97k| __m128i in_shifted3 = _mm_srli_si128(in, 12); 14| | 15| | // expand 8-bit to 32-bit unit 16| 9.97k| __m128i out1 = _mm_cvtepu8_epi32(in); 17| 9.97k| __m128i out2 = _mm_cvtepu8_epi32(in_shifted1); 18| 9.97k| __m128i out3 = _mm_cvtepu8_epi32(in_shifted2); 19| 9.97k| __m128i out4 = _mm_cvtepu8_epi32(in_shifted3); 20| | 21| 9.97k| _mm_storeu_si128((__m128i *)utf32_output, out1); 22| 9.97k| _mm_storeu_si128((__m128i *)(utf32_output + 4), out2); 23| 9.97k| _mm_storeu_si128((__m128i *)(utf32_output + 8), out3); 24| 9.97k| _mm_storeu_si128((__m128i *)(utf32_output + 12), out4); 25| | 26| 9.97k| utf32_output += 16; 27| 9.97k| buf += 16; 28| 9.97k| } 29| | 30| 5.26k| return std::make_pair(buf, utf32_output); 31| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_126sse_convert_latin1_to_utf8EPKcmPc: 3| 5.55k| const size_t latin_input_length, char *utf8_output) { 4| 5.55k| const char *end = latin_input + latin_input_length; 5| | 6| 5.55k| const __m128i v_0000 = _mm_setzero_si128(); 7| | // 0b1000_0000 8| 5.55k| const __m128i v_80 = _mm_set1_epi8((uint8_t)0x80); 9| | // 0b1111_1111_1000_0000 10| 5.55k| const __m128i v_ff80 = _mm_set1_epi16((uint16_t)0xff80); 11| | 12| 5.55k| const __m128i latin_1_half_into_u16_byte_mask = 13| 5.55k| _mm_setr_epi8(0, '\x80', 1, '\x80', 2, '\x80', 3, '\x80', 4, '\x80', 5, 14| 5.55k| '\x80', 6, '\x80', 7, '\x80'); 15| | 16| 5.55k| const __m128i latin_2_half_into_u16_byte_mask = 17| 5.55k| _mm_setr_epi8(8, '\x80', 9, '\x80', 10, '\x80', 11, '\x80', 12, '\x80', 18| 5.55k| 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| 17.8k| while (end - latin_input >= 16 + 8) { ------------------ | Branch (25:10): [True: 12.2k, False: 5.55k] ------------------ 26| | // Load 16 Latin1 characters (16 bytes) into a 128-bit register 27| 12.2k| __m128i v_latin = _mm_loadu_si128((__m128i *)latin_input); 28| | 29| 12.2k| if (_mm_testz_si128(v_latin, v_80)) { // ASCII fast path!!!! ------------------ | Branch (29:9): [True: 8.44k, False: 3.84k] ------------------ 30| 8.44k| _mm_storeu_si128((__m128i *)utf8_output, v_latin); 31| 8.44k| latin_input += 16; 32| 8.44k| utf8_output += 16; 33| 8.44k| continue; 34| 8.44k| } 35| | 36| | // assuming a/b are bytes and A/B are uint16 of the same value 37| | // aaaa_aaaa_bbbb_bbbb -> AAAA_AAAA 38| 3.84k| __m128i v_u16_latin_1_half = 39| 3.84k| _mm_shuffle_epi8(v_latin, latin_1_half_into_u16_byte_mask); 40| | // aaaa_aaaa_bbbb_bbbb -> BBBB_BBBB 41| 3.84k| __m128i v_u16_latin_2_half = 42| 3.84k| _mm_shuffle_epi8(v_latin, latin_2_half_into_u16_byte_mask); 43| | 44| 3.84k| internal::westmere::write_v_u16_11bits_to_utf8(v_u16_latin_1_half, 45| 3.84k| utf8_output, v_0000, v_ff80); 46| 3.84k| internal::westmere::write_v_u16_11bits_to_utf8(v_u16_latin_2_half, 47| 3.84k| utf8_output, v_0000, v_ff80); 48| 3.84k| latin_input += 16; 49| 3.84k| } 50| | 51| 5.55k| if (end - latin_input >= 16) { ------------------ | Branch (51:7): [True: 615, False: 4.94k] ------------------ 52| | // Load 16 Latin1 characters (16 bytes) into a 128-bit register 53| 615| __m128i v_latin = _mm_loadu_si128((__m128i *)latin_input); 54| | 55| 615| if (_mm_testz_si128(v_latin, v_80)) { // ASCII fast path!!!! ------------------ | Branch (55:9): [True: 404, False: 211] ------------------ 56| 404| _mm_storeu_si128((__m128i *)utf8_output, v_latin); 57| 404| latin_input += 16; 58| 404| utf8_output += 16; 59| 404| } else { 60| | // assuming a/b are bytes and A/B are uint16 of the same value 61| | // aaaa_aaaa_bbbb_bbbb -> AAAA_AAAA 62| 211| __m128i v_u16_latin_1_half = 63| 211| _mm_shuffle_epi8(v_latin, latin_1_half_into_u16_byte_mask); 64| 211| internal::westmere::write_v_u16_11bits_to_utf8( 65| 211| v_u16_latin_1_half, utf8_output, v_0000, v_ff80); 66| 211| latin_input += 8; 67| 211| } 68| 615| } 69| | 70| 5.55k| return std::make_pair(latin_input, utf8_output); 71| 5.55k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_utf16_to_latin1ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.26k| char *latin1_output) { 5| 5.26k| const char16_t *end = buf + len; 6| 25.3k| while (end - buf >= 8) { ------------------ | Branch (6:10): [True: 20.0k, False: 5.26k] ------------------ 7| | // Load 8 UTF-16 characters into 128-bit SSE register 8| 20.0k| __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| 20.0k| __m128i high_byte_mask = _mm_set1_epi16((int16_t)0xFF00); 17| 20.0k| if (_mm_testz_si128(in, high_byte_mask)) { ------------------ | Branch (17:9): [True: 20.0k, False: 0] ------------------ 18| | // Pack 16-bit characters into 8-bit and store in latin1_output 19| 20.0k| __m128i latin1_packed = _mm_packus_epi16(in, in); 20| 20.0k| _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output), 21| 20.0k| latin1_packed); 22| | // Adjust pointers for next iteration 23| 20.0k| buf += 8; 24| 20.0k| latin1_output += 8; 25| 20.0k| } else { 26| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 27| 0| } 28| 20.0k| } // while 29| 5.26k| return std::make_pair(buf, latin1_output); 30| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_utf16_to_latin1ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 4| 5.26k| char *latin1_output) { 5| 5.26k| const char16_t *end = buf + len; 6| 25.3k| while (end - buf >= 8) { ------------------ | Branch (6:10): [True: 20.0k, False: 5.26k] ------------------ 7| | // Load 8 UTF-16 characters into 128-bit SSE register 8| 20.0k| __m128i in = _mm_loadu_si128(reinterpret_cast(buf)); 9| | 10| 20.0k| if constexpr (!match_system(big_endian)) { 11| 20.0k| const __m128i swap = 12| 20.0k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 13| 20.0k| in = _mm_shuffle_epi8(in, swap); 14| 20.0k| } 15| | 16| 20.0k| __m128i high_byte_mask = _mm_set1_epi16((int16_t)0xFF00); 17| 20.0k| if (_mm_testz_si128(in, high_byte_mask)) { ------------------ | Branch (17:9): [True: 20.0k, False: 0] ------------------ 18| | // Pack 16-bit characters into 8-bit and store in latin1_output 19| 20.0k| __m128i latin1_packed = _mm_packus_epi16(in, in); 20| 20.0k| _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output), 21| 20.0k| latin1_packed); 22| | // Adjust pointers for next iteration 23| 20.0k| buf += 8; 24| 20.0k| latin1_output += 8; 25| 20.0k| } else { 26| 0| return std::make_pair(nullptr, reinterpret_cast(latin1_output)); 27| 0| } 28| 20.0k| } // while 29| 5.26k| return std::make_pair(buf, latin1_output); 30| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_125sse_convert_utf16_to_utf8ILNS_10endiannessE0EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 10.1k|sse_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| | 58| 10.1k| const char16_t *end = buf + len; 59| | 60| 10.1k| const __m128i v_0000 = _mm_setzero_si128(); 61| 10.1k| const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800); 62| 10.1k| const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800); 63| 10.1k| const size_t safety_margin = 64| 10.1k| 12; // to avoid overruns, see issue 65| | // https://github.com/simdutf/simdutf/issues/92 66| | 67| 19.5k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (67:10): [True: 9.46k, False: 10.0k] ------------------ 68| 9.46k| __m128i in = _mm_loadu_si128((__m128i *)buf); 69| 9.46k| if (big_endian) { ------------------ | Branch (69:9): [Folded, False: 9.46k] ------------------ 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| 9.46k| const __m128i v_ff80 = _mm_set1_epi16((int16_t)0xff80); 76| 9.46k| if (_mm_testz_si128(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 3.43k, False: 6.03k] ------------------ 77| 3.43k| __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1); 78| 3.43k| if (big_endian) { ------------------ | Branch (78:11): [Folded, False: 3.43k] ------------------ 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| 3.43k| if (!_mm_testz_si128(nextin, v_ff80)) { ------------------ | Branch (83:11): [True: 368, False: 3.06k] ------------------ 84| | // 1. pack the bytes 85| | // obviously suboptimal. 86| 368| const __m128i utf8_packed = _mm_packus_epi16(in, in); 87| | // 2. store (16 bytes) 88| 368| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 89| | // 3. adjust pointers 90| 368| buf += 8; 91| 368| utf8_output += 8; 92| 368| in = nextin; 93| 3.06k| } else { 94| | // 1. pack the bytes 95| | // obviously suboptimal. 96| 3.06k| const __m128i utf8_packed = _mm_packus_epi16(in, nextin); 97| | // 2. store (16 bytes) 98| 3.06k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 99| | // 3. adjust pointers 100| 3.06k| buf += 16; 101| 3.06k| utf8_output += 16; 102| 3.06k| continue; // we are done for this round! 103| 3.06k| } 104| 3.43k| } 105| | 106| | // no bits set above 7th bit 107| 6.39k| const __m128i one_byte_bytemask = 108| 6.39k| _mm_cmpeq_epi16(_mm_and_si128(in, v_ff80), v_0000); 109| 6.39k| const uint16_t one_byte_bitmask = 110| 6.39k| static_cast(_mm_movemask_epi8(one_byte_bytemask)); 111| | 112| | // no bits set above 11th bit 113| 6.39k| const __m128i one_or_two_bytes_bytemask = 114| 6.39k| _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_0000); 115| 6.39k| const uint16_t one_or_two_bytes_bitmask = 116| 6.39k| static_cast(_mm_movemask_epi8(one_or_two_bytes_bytemask)); 117| | 118| 6.39k| if (one_or_two_bytes_bitmask == 0xffff) { ------------------ | Branch (118:9): [True: 1.99k, False: 4.40k] ------------------ 119| 1.99k| internal::westmere::write_v_u16_11bits_to_utf8( 120| 1.99k| in, utf8_output, one_byte_bytemask, one_byte_bitmask); 121| 1.99k| buf += 8; 122| 1.99k| continue; 123| 1.99k| } 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| 4.40k| const __m128i surrogates_bytemask = 129| 4.40k| _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| 4.40k| const uint16_t surrogates_bitmask = 134| 4.40k| 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| 4.40k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (137:9): [True: 3.29k, False: 1.11k] ------------------ 138| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 139| 3.29k| const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606, 140| 3.29k| 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| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) 170| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 171| 3.29k| const __m128i t0 = _mm_shuffle_epi8(in, dup_even); 172| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 173| 3.29k| const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 169| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 174| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 175| 3.29k| const __m128i t2 = _mm_or_si128(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 169| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 176| | 177| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 178| 3.29k| const __m128i s0 = _mm_srli_epi16(in, 4); 179| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 180| 3.29k| const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 169| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 181| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 182| 3.29k| const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 169| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 183| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 184| 3.29k| const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 169| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 185| 3.29k| const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, 186| 3.29k| simdutf_vec(0b0100000000000000)); ------------------ | | 169| 3.29k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 187| 3.29k| const __m128i s4 = _mm_xor_si128(s3, m0); 188| 3.29k|#undef simdutf_vec 189| | 190| | // 4. expand code units 16-bit => 32-bit 191| 3.29k| const __m128i out0 = _mm_unpacklo_epi16(t2, s4); 192| 3.29k| 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| 3.29k| const uint16_t mask = 196| 3.29k| (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa); 197| 3.29k| if (mask == 0) { ------------------ | Branch (197:11): [True: 2.48k, False: 811] ------------------ 198| | // We only have three-byte code units. Use fast path. 199| 2.48k| const __m128i shuffle = _mm_setr_epi8(2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 200| 2.48k| 15, 13, -1, -1, -1, -1); 201| 2.48k| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle); 202| 2.48k| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle); 203| 2.48k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 204| 2.48k| utf8_output += 12; 205| 2.48k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 206| 2.48k| utf8_output += 12; 207| 2.48k| buf += 8; 208| 2.48k| continue; 209| 2.48k| } 210| 811| const uint8_t mask0 = uint8_t(mask); 211| | 212| 811| const uint8_t *row0 = 213| 811| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 214| 811| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 215| 811| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0); 216| | 217| 811| const uint8_t mask1 = static_cast(mask >> 8); 218| | 219| 811| const uint8_t *row1 = 220| 811| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 221| 811| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 222| 811| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1); 223| | 224| 811| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 225| 811| utf8_output += row0[0]; 226| 811| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 227| 811| utf8_output += row1[0]; 228| | 229| 811| buf += 8; 230| | // surrogate pair(s) in a register 231| 1.11k| } 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| 1.11k| size_t forward = 15; 236| 1.11k| size_t k = 0; 237| 1.11k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (237:11): [True: 0, False: 1.11k] ------------------ 238| 0| forward = size_t(end - buf - 1); 239| 0| } 240| 15.3k| for (; k < forward; k++) { ------------------ | Branch (240:14): [True: 14.2k, False: 1.04k] ------------------ 241| 14.2k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 242| 14.2k| if ((word & 0xFF80) == 0) { ------------------ | Branch (242:13): [True: 8.37k, False: 5.88k] ------------------ 243| 8.37k| *utf8_output++ = char(word); 244| 8.37k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (244:20): [True: 638, False: 5.25k] ------------------ 245| 638| *utf8_output++ = char((word >> 6) | 0b11000000); 246| 638| *utf8_output++ = char((word & 0b111111) | 0b10000000); 247| 5.25k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (247:20): [True: 3.50k, False: 1.74k] ------------------ 248| 3.50k| *utf8_output++ = char((word >> 12) | 0b11100000); 249| 3.50k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 250| 3.50k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 251| 3.50k| } else { 252| | // must be a surrogate pair 253| 1.74k| uint16_t diff = uint16_t(word - 0xD800); 254| 1.74k| uint16_t next_word = 255| 1.74k| scalar::utf16::swap_if_needed(buf[k + 1]); 256| 1.74k| k++; 257| 1.74k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 258| 1.74k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (258:15): [True: 70, False: 1.67k] ------------------ 259| 70| return std::make_pair(nullptr, utf8_output); 260| 70| } 261| 1.67k| uint32_t value = (diff << 10) + diff2 + 0x10000; 262| 1.67k| *utf8_output++ = char((value >> 18) | 0b11110000); 263| 1.67k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 264| 1.67k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 265| 1.67k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 266| 1.67k| } 267| 14.2k| } 268| 1.04k| buf += k; 269| 1.04k| } 270| 4.40k| } // while 271| | 272| 10.0k| return std::make_pair(buf, utf8_output); 273| 10.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_125sse_convert_utf16_to_utf8ILNS_10endiannessE1EEENSt3__14pairIPKDsPcEES7_mS8_: 56| 10.1k|sse_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) { 57| | 58| 10.1k| const char16_t *end = buf + len; 59| | 60| 10.1k| const __m128i v_0000 = _mm_setzero_si128(); 61| 10.1k| const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800); 62| 10.1k| const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800); 63| 10.1k| const size_t safety_margin = 64| 10.1k| 12; // to avoid overruns, see issue 65| | // https://github.com/simdutf/simdutf/issues/92 66| | 67| 19.7k| while (end - buf >= std::ptrdiff_t(16 + safety_margin)) { ------------------ | Branch (67:10): [True: 9.68k, False: 10.0k] ------------------ 68| 9.68k| __m128i in = _mm_loadu_si128((__m128i *)buf); 69| 9.68k| if (big_endian) { ------------------ | Branch (69:9): [True: 9.68k, Folded] ------------------ 70| 9.68k| const __m128i swap = 71| 9.68k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 72| 9.68k| in = _mm_shuffle_epi8(in, swap); 73| 9.68k| } 74| | // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes 75| 9.68k| const __m128i v_ff80 = _mm_set1_epi16((int16_t)0xff80); 76| 9.68k| if (_mm_testz_si128(in, v_ff80)) { // ASCII fast path!!!! ------------------ | Branch (76:9): [True: 3.42k, False: 6.25k] ------------------ 77| 3.42k| __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1); 78| 3.42k| if (big_endian) { ------------------ | Branch (78:11): [True: 3.42k, Folded] ------------------ 79| 3.42k| const __m128i swap = 80| 3.42k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 81| 3.42k| nextin = _mm_shuffle_epi8(nextin, swap); 82| 3.42k| } 83| 3.42k| if (!_mm_testz_si128(nextin, v_ff80)) { ------------------ | Branch (83:11): [True: 362, False: 3.06k] ------------------ 84| | // 1. pack the bytes 85| | // obviously suboptimal. 86| 362| const __m128i utf8_packed = _mm_packus_epi16(in, in); 87| | // 2. store (16 bytes) 88| 362| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 89| | // 3. adjust pointers 90| 362| buf += 8; 91| 362| utf8_output += 8; 92| 362| in = nextin; 93| 3.06k| } else { 94| | // 1. pack the bytes 95| | // obviously suboptimal. 96| 3.06k| const __m128i utf8_packed = _mm_packus_epi16(in, nextin); 97| | // 2. store (16 bytes) 98| 3.06k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 99| | // 3. adjust pointers 100| 3.06k| buf += 16; 101| 3.06k| utf8_output += 16; 102| 3.06k| continue; // we are done for this round! 103| 3.06k| } 104| 3.42k| } 105| | 106| | // no bits set above 7th bit 107| 6.62k| const __m128i one_byte_bytemask = 108| 6.62k| _mm_cmpeq_epi16(_mm_and_si128(in, v_ff80), v_0000); 109| 6.62k| const uint16_t one_byte_bitmask = 110| 6.62k| static_cast(_mm_movemask_epi8(one_byte_bytemask)); 111| | 112| | // no bits set above 11th bit 113| 6.62k| const __m128i one_or_two_bytes_bytemask = 114| 6.62k| _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_0000); 115| 6.62k| const uint16_t one_or_two_bytes_bitmask = 116| 6.62k| static_cast(_mm_movemask_epi8(one_or_two_bytes_bytemask)); 117| | 118| 6.62k| if (one_or_two_bytes_bitmask == 0xffff) { ------------------ | Branch (118:9): [True: 1.97k, False: 4.64k] ------------------ 119| 1.97k| internal::westmere::write_v_u16_11bits_to_utf8( 120| 1.97k| in, utf8_output, one_byte_bytemask, one_byte_bitmask); 121| 1.97k| buf += 8; 122| 1.97k| continue; 123| 1.97k| } 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| 4.64k| const __m128i surrogates_bytemask = 129| 4.64k| _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| 4.64k| const uint16_t surrogates_bitmask = 134| 4.64k| 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| 4.64k| if (surrogates_bitmask == 0x0000) { ------------------ | Branch (137:9): [True: 3.57k, False: 1.07k] ------------------ 138| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 139| 3.57k| const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606, 140| 3.57k| 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| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) 170| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 171| 3.57k| const __m128i t0 = _mm_shuffle_epi8(in, dup_even); 172| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 173| 3.57k| const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 169| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 174| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 175| 3.57k| const __m128i t2 = _mm_or_si128(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 169| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 176| | 177| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 178| 3.57k| const __m128i s0 = _mm_srli_epi16(in, 4); 179| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 180| 3.57k| const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 169| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 181| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 182| 3.57k| const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 169| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 183| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 184| 3.57k| const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 169| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 185| 3.57k| const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, 186| 3.57k| simdutf_vec(0b0100000000000000)); ------------------ | | 169| 3.57k|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 187| 3.57k| const __m128i s4 = _mm_xor_si128(s3, m0); 188| 3.57k|#undef simdutf_vec 189| | 190| | // 4. expand code units 16-bit => 32-bit 191| 3.57k| const __m128i out0 = _mm_unpacklo_epi16(t2, s4); 192| 3.57k| 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| 3.57k| const uint16_t mask = 196| 3.57k| (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa); 197| 3.57k| if (mask == 0) { ------------------ | Branch (197:11): [True: 2.69k, False: 875] ------------------ 198| | // We only have three-byte code units. Use fast path. 199| 2.69k| const __m128i shuffle = _mm_setr_epi8(2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 200| 2.69k| 15, 13, -1, -1, -1, -1); 201| 2.69k| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle); 202| 2.69k| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle); 203| 2.69k| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 204| 2.69k| utf8_output += 12; 205| 2.69k| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 206| 2.69k| utf8_output += 12; 207| 2.69k| buf += 8; 208| 2.69k| continue; 209| 2.69k| } 210| 875| const uint8_t mask0 = uint8_t(mask); 211| | 212| 875| const uint8_t *row0 = 213| 875| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 214| 875| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 215| 875| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0); 216| | 217| 875| const uint8_t mask1 = static_cast(mask >> 8); 218| | 219| 875| const uint8_t *row1 = 220| 875| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 221| 875| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 222| 875| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1); 223| | 224| 875| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 225| 875| utf8_output += row0[0]; 226| 875| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 227| 875| utf8_output += row1[0]; 228| | 229| 875| buf += 8; 230| | // surrogate pair(s) in a register 231| 1.07k| } 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| 1.07k| size_t forward = 15; 236| 1.07k| size_t k = 0; 237| 1.07k| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (237:11): [True: 0, False: 1.07k] ------------------ 238| 0| forward = size_t(end - buf - 1); 239| 0| } 240| 14.9k| for (; k < forward; k++) { ------------------ | Branch (240:14): [True: 13.9k, False: 1.01k] ------------------ 241| 13.9k| uint16_t word = scalar::utf16::swap_if_needed(buf[k]); 242| 13.9k| if ((word & 0xFF80) == 0) { ------------------ | Branch (242:13): [True: 8.37k, False: 5.59k] ------------------ 243| 8.37k| *utf8_output++ = char(word); 244| 8.37k| } else if ((word & 0xF800) == 0) { ------------------ | Branch (244:20): [True: 689, False: 4.91k] ------------------ 245| 689| *utf8_output++ = char((word >> 6) | 0b11000000); 246| 689| *utf8_output++ = char((word & 0b111111) | 0b10000000); 247| 4.91k| } else if ((word & 0xF800) != 0xD800) { ------------------ | Branch (247:20): [True: 3.27k, False: 1.63k] ------------------ 248| 3.27k| *utf8_output++ = char((word >> 12) | 0b11100000); 249| 3.27k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 250| 3.27k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 251| 3.27k| } else { 252| | // must be a surrogate pair 253| 1.63k| uint16_t diff = uint16_t(word - 0xD800); 254| 1.63k| uint16_t next_word = 255| 1.63k| scalar::utf16::swap_if_needed(buf[k + 1]); 256| 1.63k| k++; 257| 1.63k| uint16_t diff2 = uint16_t(next_word - 0xDC00); 258| 1.63k| if ((diff | diff2) > 0x3FF) { ------------------ | Branch (258:15): [True: 69, False: 1.57k] ------------------ 259| 69| return std::make_pair(nullptr, utf8_output); 260| 69| } 261| 1.57k| uint32_t value = (diff << 10) + diff2 + 0x10000; 262| 1.57k| *utf8_output++ = char((value >> 18) | 0b11110000); 263| 1.57k| *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000); 264| 1.57k| *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000); 265| 1.57k| *utf8_output++ = char((value & 0b111111) | 0b10000000); 266| 1.57k| } 267| 13.9k| } 268| 1.01k| buf += k; 269| 1.01k| } 270| 4.64k| } // while 271| | 272| 10.0k| return std::make_pair(buf, utf8_output); 273| 10.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_127sse_convert_utf32_to_latin1EPKDimPc: 3| 5.26k| char *latin1_output) { 4| 5.26k| const size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16 5| | 6| 5.26k| __m128i high_bytes_mask = _mm_set1_epi32(0xFFFFFF00); 7| 5.26k| __m128i shufmask = 8| 5.26k| _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 12, 8, 4, 0); 9| | 10| 15.2k| for (size_t i = 0; i < rounded_len; i += 16) { ------------------ | Branch (10:22): [True: 9.97k, False: 5.26k] ------------------ 11| 9.97k| __m128i in1 = _mm_loadu_si128((__m128i *)buf); 12| 9.97k| __m128i in2 = _mm_loadu_si128((__m128i *)(buf + 4)); 13| 9.97k| __m128i in3 = _mm_loadu_si128((__m128i *)(buf + 8)); 14| 9.97k| __m128i in4 = _mm_loadu_si128((__m128i *)(buf + 12)); 15| | 16| 9.97k| __m128i check_combined = _mm_or_si128(in1, in2); 17| 9.97k| check_combined = _mm_or_si128(check_combined, in3); 18| 9.97k| check_combined = _mm_or_si128(check_combined, in4); 19| | 20| 9.97k| if (!_mm_testz_si128(check_combined, high_bytes_mask)) { ------------------ | Branch (20:9): [True: 0, False: 9.97k] ------------------ 21| 0| return std::make_pair(nullptr, latin1_output); 22| 0| } 23| 9.97k| __m128i pack1 = _mm_unpacklo_epi32(_mm_shuffle_epi8(in1, shufmask), 24| 9.97k| _mm_shuffle_epi8(in2, shufmask)); 25| 9.97k| __m128i pack2 = _mm_unpacklo_epi32(_mm_shuffle_epi8(in3, shufmask), 26| 9.97k| _mm_shuffle_epi8(in4, shufmask)); 27| 9.97k| __m128i pack = _mm_unpacklo_epi64(pack1, pack2); 28| 9.97k| _mm_storeu_si128((__m128i *)latin1_output, pack); 29| 9.97k| latin1_output += 16; 30| 9.97k| buf += 16; 31| 9.97k| } 32| | 33| 5.26k| return std::make_pair(buf, latin1_output); 34| 5.26k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_125sse_convert_utf32_to_utf8EPKDimPc: 2| 4.86k|sse_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) { 3| 4.86k| const char32_t *end = buf + len; 4| | 5| 4.86k| const __m128i v_0000 = _mm_setzero_si128(); //__m128 = 128 bits 6| 4.86k| const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800); // 1111 1000 0000 7| | // 0000 8| 4.86k| const __m128i v_c080 = _mm_set1_epi16((uint16_t)0xc080); // 1100 0000 1000 9| | // 0000 10| 4.86k| const __m128i v_ff80 = _mm_set1_epi16((uint16_t)0xff80); // 1111 1111 1000 11| | // 0000 12| 4.86k| const __m128i v_ffff0000 = _mm_set1_epi32( 13| 4.86k| (uint32_t)0xffff0000); // 1111 1111 1111 1111 0000 0000 0000 0000 14| 4.86k| const __m128i v_7fffffff = _mm_set1_epi32( 15| 4.86k| (uint32_t)0x7fffffff); // 0111 1111 1111 1111 1111 1111 1111 1111 16| 4.86k| __m128i running_max = _mm_setzero_si128(); 17| 4.86k| __m128i forbidden_bytemask = _mm_setzero_si128(); 18| 4.86k| const size_t safety_margin = 19| 4.86k| 12; // to avoid overruns, see issue 20| | // https://github.com/simdutf/simdutf/issues/92 21| | 22| 10.4k| while (end - buf >= ------------------ | Branch (22:10): [True: 5.63k, False: 4.86k] ------------------ 23| 10.4k| std::ptrdiff_t( 24| 10.4k| 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| 5.63k| __m128i in = _mm_loadu_si128((__m128i *)buf); 29| 5.63k| __m128i nextin = _mm_loadu_si128( 30| 5.63k| (__m128i *)buf + 1); // These two values can hold only 8 UTF32 chars 31| 5.63k| running_max = _mm_max_epu32( 32| 5.63k| _mm_max_epu32(in, running_max), // take element-wise max char32_t from 33| | // in and running_max vector 34| 5.63k| 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| 5.63k| __m128i in_16 = _mm_packus_epi32( 40| 5.63k| _mm_and_si128(in, v_7fffffff), 41| 5.63k| _mm_and_si128( 42| 5.63k| nextin, 43| 5.63k| 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| 5.63k| if (_mm_testz_si128(in_16, v_ff80)) { // if the first two blocks are ASCII ------------------ | Branch (60:9): [True: 3.26k, False: 2.37k] ------------------ 61| 3.26k| __m128i thirdin = _mm_loadu_si128((__m128i *)buf + 2); 62| 3.26k| __m128i fourthin = _mm_loadu_si128((__m128i *)buf + 3); 63| 3.26k| running_max = _mm_max_epu32( 64| 3.26k| _mm_max_epu32(thirdin, running_max), 65| 3.26k| fourthin); // take the running max of all 4 vectors thus far 66| 3.26k| __m128i nextin_16 = _mm_packus_epi32( 67| 3.26k| _mm_and_si128(thirdin, v_7fffffff), 68| 3.26k| _mm_and_si128(fourthin, 69| 3.26k| v_7fffffff)); // pack into 1 vector, now you have two 70| 3.26k| if (!_mm_testz_si128( ------------------ | Branch (70:11): [True: 335, False: 2.93k] ------------------ 71| 3.26k| nextin_16, 72| 3.26k| v_ff80)) { // checks if the second packed vector is ASCII, if not: 73| | // 1. pack the bytes 74| | // obviously suboptimal. 75| 335| const __m128i utf8_packed = _mm_packus_epi16( 76| 335| in_16, in_16); // creates two copy of in_16 in 1 vector 77| | // 2. store (16 bytes) 78| 335| _mm_storeu_si128((__m128i *)utf8_output, 79| 335| utf8_packed); // put them into the output 80| | // 3. adjust pointers 81| 335| buf += 8; // the char32_t buffer pointer goes up 8 char32_t chars* 32 82| | // bits = 256 bits 83| 335| utf8_output += 84| 335| 8; // same with output, e.g. lift the first two blocks alone. 85| | // Proceed with next input 86| 335| in_16 = nextin_16; 87| | // We need to update in and nextin because they are used later. 88| 335| in = thirdin; 89| 335| nextin = fourthin; 90| 2.93k| } else { 91| | // 1. pack the bytes 92| 2.93k| const __m128i utf8_packed = _mm_packus_epi16(in_16, nextin_16); 93| | // 2. store (16 bytes) 94| 2.93k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 95| | // 3. adjust pointers 96| 2.93k| buf += 16; 97| 2.93k| utf8_output += 16; 98| 2.93k| continue; // we are done for this round! 99| 2.93k| } 100| 3.26k| } 101| | 102| | // no bits set above 7th bit -- find out all the ASCII characters 103| 2.70k| const __m128i one_byte_bytemask = 104| 2.70k| _mm_cmpeq_epi16( // this takes four bytes at a time and compares: 105| 2.70k| _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| 2.70k| v_0000 // 108| 2.70k| ); // 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| 2.70k| const uint16_t one_byte_bitmask = static_cast(_mm_movemask_epi8( 113| 2.70k| 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| 2.70k| const __m128i one_or_two_bytes_bytemask = 118| 2.70k| _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_0000); 119| 2.70k| const uint16_t one_or_two_bytes_bitmask = 120| 2.70k| static_cast(_mm_movemask_epi8(one_or_two_bytes_bytemask)); 121| | 122| 2.70k| if (one_or_two_bytes_bitmask == 0xffff) { ------------------ | Branch (122:9): [True: 1.59k, False: 1.10k] ------------------ 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| 1.59k| const __m128i v_1f00 = 129| 1.59k| _mm_set1_epi16((int16_t)0x1f00); // 0001 1111 0000 0000 130| 1.59k| const __m128i v_003f = 131| 1.59k| _mm_set1_epi16((int16_t)0x003f); // 0000 0000 0011 1111 132| | 133| | // t0 = [000a|aaaa|bbbb|bb00] 134| 1.59k| const __m128i t0 = _mm_slli_epi16(in_16, 2); // shift packed vector by two 135| | // t1 = [000a|aaaa|0000|0000] 136| 1.59k| const __m128i t1 = _mm_and_si128(t0, v_1f00); // potential first utf8 byte 137| | // t2 = [0000|0000|00bb|bbbb] 138| 1.59k| const __m128i t2 = 139| 1.59k| _mm_and_si128(in_16, v_003f); // potential second utf8 byte 140| | // t3 = [000a|aaaa|00bb|bbbb] 141| 1.59k| const __m128i t3 = 142| 1.59k| _mm_or_si128(t1, t2); // first and second potential utf8 byte together 143| | // t4 = [110a|aaaa|10bb|bbbb] 144| 1.59k| const __m128i t4 = _mm_or_si128( 145| 1.59k| t3, 146| 1.59k| v_c080); // t3 | 1100 0000 1000 0000 = full potential 2-byte utf8 unit 147| | 148| | // 2. merge ASCII and 2-byte codewords 149| 1.59k| const __m128i utf8_unpacked = 150| 1.59k| _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| 1.59k| const uint16_t m0 = one_byte_bitmask & 0x5555; // m0 = 0h0g0f0e0d0c0b0a 156| 1.59k| const uint16_t m1 = 157| 1.59k| static_cast(m0 >> 7); // m1 = 00000000h0g0f0e0 158| 1.59k| const uint8_t m2 = 159| 1.59k| static_cast((m0 | m1) & 0xff); // m2 = hdgcfbea 160| | // 4. pack the bytes 161| 1.59k| const uint8_t *row = 162| 1.59k| &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0]; 163| 1.59k| const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1)); 164| 1.59k| const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle); 165| | 166| | // 5. store bytes 167| 1.59k| _mm_storeu_si128((__m128i *)utf8_output, utf8_packed); 168| | 169| | // 6. adjust pointers 170| 1.59k| buf += 8; 171| 1.59k| utf8_output += row[0]; 172| 1.59k| continue; 173| 1.59k| } 174| | 175| | // Check for overflow in packing 176| | 177| 1.10k| const __m128i saturation_bytemask = _mm_cmpeq_epi32( 178| 1.10k| _mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000); 179| 1.10k| const uint32_t saturation_bitmask = 180| 1.10k| static_cast(_mm_movemask_epi8(saturation_bytemask)); 181| 1.10k| if (saturation_bitmask == 0xffff) { ------------------ | Branch (181:9): [True: 435, False: 672] ------------------ 182| | // case: code units from register produce either 1, 2 or 3 UTF-8 bytes 183| 435| const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800); 184| 435| forbidden_bytemask = 185| 435| _mm_or_si128(forbidden_bytemask, 186| 435| _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_d800)); 187| | 188| 435| const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606, 189| 435| 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| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) 219| | // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc] 220| 435| const __m128i t0 = _mm_shuffle_epi8(in_16, dup_even); 221| | // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc] 222| 435| const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111)); ------------------ | | 218| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 223| | // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc] 224| 435| const __m128i t2 = _mm_or_si128(t1, simdutf_vec(0b1000000000000000)); ------------------ | | 218| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 225| | 226| | // [aaaa|bbbb|bbcc|cccc] => [0000|aaaa|bbbb|bbcc] 227| 435| const __m128i s0 = _mm_srli_epi16(in_16, 4); 228| | // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00] 229| 435| const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100)); ------------------ | | 218| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 230| | // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa] 231| 435| const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140)); ------------------ | | 218| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 232| | // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa] 233| 435| const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000)); ------------------ | | 218| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 234| 435| const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, 235| 435| simdutf_vec(0b0100000000000000)); ------------------ | | 218| 435|#define simdutf_vec(x) _mm_set1_epi16(static_cast(x)) ------------------ 236| 435| const __m128i s4 = _mm_xor_si128(s3, m0); 237| 435|#undef simdutf_vec 238| | 239| | // 4. expand code units 16-bit => 32-bit 240| 435| const __m128i out0 = _mm_unpacklo_epi16(t2, s4); 241| 435| 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| 435| const uint16_t mask = 245| 435| (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa); 246| 435| if (mask == 0) { ------------------ | Branch (246:11): [True: 74, False: 361] ------------------ 247| | // We only have three-byte code units. Use fast path. 248| 74| const __m128i shuffle = _mm_setr_epi8(2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 249| 74| 15, 13, -1, -1, -1, -1); 250| 74| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle); 251| 74| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle); 252| 74| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 253| 74| utf8_output += 12; 254| 74| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 255| 74| utf8_output += 12; 256| 74| buf += 8; 257| 74| continue; 258| 74| } 259| 361| const uint8_t mask0 = uint8_t(mask); 260| | 261| 361| const uint8_t *row0 = 262| 361| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0]; 263| 361| const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1)); 264| 361| const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0); 265| | 266| 361| const uint8_t mask1 = static_cast(mask >> 8); 267| | 268| 361| const uint8_t *row1 = 269| 361| &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0]; 270| 361| const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1)); 271| 361| const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1); 272| | 273| 361| _mm_storeu_si128((__m128i *)utf8_output, utf8_0); 274| 361| utf8_output += row0[0]; 275| 361| _mm_storeu_si128((__m128i *)utf8_output, utf8_1); 276| 361| utf8_output += row1[0]; 277| | 278| 361| buf += 8; 279| 672| } 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| 672| size_t forward = 15; 285| 672| size_t k = 0; 286| 672| if (size_t(end - buf) < forward + 1) { ------------------ | Branch (286:11): [True: 0, False: 672] ------------------ 287| 0| forward = size_t(end - buf - 1); 288| 0| } 289| 10.7k| for (; k < forward; k++) { ------------------ | Branch (289:14): [True: 10.0k, False: 672] ------------------ 290| 10.0k| uint32_t word = buf[k]; 291| 10.0k| if ((word & 0xFFFFFF80) == 0) { ------------------ | Branch (291:13): [True: 8.01k, False: 2.06k] ------------------ 292| 8.01k| *utf8_output++ = char(word); 293| 8.01k| } else if ((word & 0xFFFFF800) == 0) { ------------------ | Branch (293:20): [True: 474, False: 1.59k] ------------------ 294| 474| *utf8_output++ = char((word >> 6) | 0b11000000); 295| 474| *utf8_output++ = char((word & 0b111111) | 0b10000000); 296| 1.59k| } else if ((word & 0xFFFF0000) == 0) { ------------------ | Branch (296:20): [True: 526, False: 1.06k] ------------------ 297| 526| if (word >= 0xD800 && word <= 0xDFFF) { ------------------ | Branch (297:15): [True: 245, False: 281] | Branch (297:33): [True: 0, False: 245] ------------------ 298| 0| return std::make_pair(nullptr, utf8_output); 299| 0| } 300| 526| *utf8_output++ = char((word >> 12) | 0b11100000); 301| 526| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 302| 526| *utf8_output++ = char((word & 0b111111) | 0b10000000); 303| 1.06k| } else { 304| 1.06k| if (word > 0x10FFFF) { ------------------ | Branch (304:15): [True: 0, False: 1.06k] ------------------ 305| 0| return std::make_pair(nullptr, utf8_output); 306| 0| } 307| 1.06k| *utf8_output++ = char((word >> 18) | 0b11110000); 308| 1.06k| *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000); 309| 1.06k| *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000); 310| 1.06k| *utf8_output++ = char((word & 0b111111) | 0b10000000); 311| 1.06k| } 312| 10.0k| } 313| 672| buf += k; 314| 672| } 315| 1.10k| } // while 316| | 317| | // check for invalid input 318| 4.86k| const __m128i v_10ffff = _mm_set1_epi32((uint32_t)0x10ffff); 319| 4.86k| if (static_cast(_mm_movemask_epi8(_mm_cmpeq_epi32( ------------------ | Branch (319:7): [True: 0, False: 4.86k] ------------------ 320| 4.86k| _mm_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffff) { 321| 0| return std::make_pair(nullptr, utf8_output); 322| 0| } 323| | 324| 4.86k| if (static_cast(_mm_movemask_epi8(forbidden_bytemask)) != 0) { ------------------ | Branch (324:7): [True: 0, False: 4.86k] ------------------ 325| 0| return std::make_pair(nullptr, utf8_output); 326| 0| } 327| | 328| 4.86k| return std::make_pair(buf, utf8_output); 329| 4.86k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_129convert_masked_utf8_to_latin1EPKcmRPc: 9| 22.5k| 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| 22.5k| const __m128i in = _mm_loadu_si128((__m128i *)input); 21| 22.5k| const uint16_t input_utf8_end_of_code_point_mask = 22| 22.5k| utf8_end_of_code_point_mask & 23| 22.5k| 0xfff; // we are only processing 12 bytes in case it is not all ASCII 24| 22.5k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (24:7): [True: 128, False: 22.4k] ------------------ 25| | // We process the data in chunks of 12 bytes. 26| 128| _mm_storeu_si128(reinterpret_cast<__m128i *>(latin1_output), in); 27| 128| latin1_output += 12; // We wrote 12 characters. 28| 128| return 12; // We consumed 12 bytes. 29| 128| } 30| | /// We do not have a fast path available, so we fallback. 31| 22.4k| const uint8_t idx = 32| 22.4k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 33| 22.4k| const uint8_t consumed = 34| 22.4k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 35| | // this indicates an invalid input: 36| 22.4k| if (idx >= 64) { ------------------ | Branch (36:7): [True: 5.56k, False: 16.8k] ------------------ 37| 5.56k| return consumed; 38| 5.56k| } 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| 16.8k| const __m128i sh = 46| 16.8k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 47| 16.8k| const __m128i perm = _mm_shuffle_epi8(in, sh); 48| 16.8k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 49| 16.8k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 50| 16.8k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 51| 16.8k| 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| 16.8k| _mm_storel_epi64((__m128i *)latin1_output, latin1_packed); 56| 16.8k| latin1_output += 6; // We wrote 6 bytes. 57| 16.8k| return consumed; 58| 22.4k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE0EEEmPKcmRPDs: 10| 20.8k| 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| 20.8k| const __m128i swap = 23| 20.8k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 20.8k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 20.8k| const uint16_t input_utf8_end_of_code_point_mask = 26| 20.8k| utf8_end_of_code_point_mask & 0xfff; 27| 20.8k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 85, False: 20.7k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| | // Note: using 16 bytes is unsafe, see issue_ossfuzz_71218 30| 85| __m128i ascii_first = _mm_cvtepu8_epi16(in); 31| 85| __m128i ascii_second = _mm_cvtepu8_epi16(_mm_srli_si128(in, 8)); 32| 85| if (big_endian) { ------------------ | Branch (32:9): [Folded, False: 85] ------------------ 33| 0| ascii_first = _mm_shuffle_epi8(ascii_first, swap); 34| 0| ascii_second = _mm_shuffle_epi8(ascii_second, swap); 35| 0| } 36| 85| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output), ascii_first); 37| 85| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + 8), 38| 85| ascii_second); 39| 85| utf16_output += 12; // We wrote 12 16-bit characters. 40| 85| return 12; // We consumed 12 bytes. 41| 85| } 42| 20.7k| if (((utf8_end_of_code_point_mask & 0xFFFF) == 0xaaaa)) { ------------------ | Branch (42:7): [True: 373, False: 20.3k] ------------------ 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| 373| const __m128i sh = 47| 373| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 48| 373| const __m128i perm = _mm_shuffle_epi8(in, sh); 49| 373| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 50| 373| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 51| 373| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 52| 373| if (big_endian) ------------------ | Branch (52:9): [Folded, False: 373] ------------------ 53| 0| composed = _mm_shuffle_epi8(composed, swap); 54| 373| _mm_storeu_si128((__m128i *)utf16_output, composed); 55| 373| utf16_output += 8; // We wrote 16 bytes, 8 code points. 56| 373| return 16; 57| 373| } 58| 20.3k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (58:7): [True: 5.29k, False: 15.0k] ------------------ 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| 5.29k| const __m128i sh = 63| 5.29k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 64| 5.29k| const __m128i perm = _mm_shuffle_epi8(in, sh); 65| 5.29k| const __m128i ascii = 66| 5.29k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 67| 5.29k| const __m128i middlebyte = 68| 5.29k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 69| 5.29k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 70| 5.29k| const __m128i highbyte = 71| 5.29k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 72| 5.29k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 73| 5.29k| const __m128i composed = 74| 5.29k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 75| 5.29k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 76| 5.29k| if (big_endian) ------------------ | Branch (76:9): [Folded, False: 5.29k] ------------------ 77| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 78| 5.29k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 79| 5.29k| utf16_output += 4; 80| 5.29k| return 12; 81| 5.29k| } 82| | /// We do not have a fast path available, so we fallback. 83| | 84| 15.0k| const uint8_t idx = 85| 15.0k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 86| 15.0k| const uint8_t consumed = 87| 15.0k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 88| 15.0k| if (idx < 64) { ------------------ | Branch (88:7): [True: 8.18k, False: 6.91k] ------------------ 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| 8.18k| const __m128i sh = 96| 8.18k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 97| 8.18k| const __m128i perm = _mm_shuffle_epi8(in, sh); 98| 8.18k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 99| 8.18k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 100| 8.18k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 101| 8.18k| if (big_endian) ------------------ | Branch (101:9): [Folded, False: 8.18k] ------------------ 102| 0| composed = _mm_shuffle_epi8(composed, swap); 103| 8.18k| _mm_storeu_si128((__m128i *)utf16_output, composed); 104| 8.18k| utf16_output += 6; // We wrote 12 bytes, 6 code points. 105| 8.18k| } else if (idx < 145) { ------------------ | Branch (105:14): [True: 1.93k, False: 4.97k] ------------------ 106| | // FOUR (4) input code-code units 107| 1.93k| const __m128i sh = 108| 1.93k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 109| 1.93k| const __m128i perm = _mm_shuffle_epi8(in, sh); 110| 1.93k| const __m128i ascii = 111| 1.93k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 112| 1.93k| const __m128i middlebyte = 113| 1.93k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 114| 1.93k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 115| 1.93k| const __m128i highbyte = 116| 1.93k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 117| 1.93k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 118| 1.93k| const __m128i composed = 119| 1.93k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 120| 1.93k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 121| 1.93k| if (big_endian) ------------------ | Branch (121:9): [Folded, False: 1.93k] ------------------ 122| 0| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 123| 1.93k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 124| 1.93k| utf16_output += 4; 125| 4.97k| } else if (idx < 209) { ------------------ | Branch (125:14): [True: 4.68k, False: 289] ------------------ 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| 4.68k| const __m128i sh = 137| 4.68k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 138| 4.68k| const __m128i perm = _mm_shuffle_epi8(in, sh); 139| 4.68k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 140| 4.68k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 141| 4.68k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 142| 4.68k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 143| | // correct for spurious high bit 144| 4.68k| const __m128i correct = 145| 4.68k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 146| 4.68k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 147| 4.68k| 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| 4.68k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 151| 4.68k| 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| 4.68k| const __m128i composed = 157| 4.68k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 158| 4.68k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 159| 4.68k| const __m128i composedminus = 160| 4.68k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 161| 4.68k| const __m128i lowtenbits = 162| 4.68k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 163| | // Notice the 0x3ff mask: 164| 4.68k| const __m128i hightenbits = 165| 4.68k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 166| 4.68k| const __m128i lowtenbitsadd = 167| 4.68k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 168| 4.68k| const __m128i hightenbitsadd = 169| 4.68k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 170| 4.68k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 171| 4.68k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 172| 4.68k| uint32_t basic_buffer[4]; 173| 4.68k| uint32_t basic_buffer_swap[4]; 174| 4.68k| if (big_endian) { ------------------ | Branch (174:9): [Folded, False: 4.68k] ------------------ 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| 4.68k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 180| 4.68k| uint32_t surrogate_buffer[4]; 181| 4.68k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 182| 18.7k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (182:24): [True: 14.0k, False: 4.68k] ------------------ 183| 14.0k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (183:11): [True: 3.08k, False: 10.9k] ------------------ 184| 3.08k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 185| 3.08k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 186| 3.08k| utf16_output += 2; 187| 10.9k| } else { 188| 10.9k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (188:27): [Folded, False: 10.9k] ------------------ 189| 10.9k| : uint16_t(basic_buffer[i]); 190| 10.9k| utf16_output++; 191| 10.9k| } 192| 14.0k| } 193| 4.68k| } else { 194| | // here we know that there is an error but we do not handle errors 195| 289| } 196| 15.0k| return consumed; 197| 20.3k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_128convert_masked_utf8_to_utf16ILNS_10endiannessE1EEEmPKcmRPDs: 10| 21.6k| 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| 21.6k| const __m128i swap = 23| 21.6k| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 24| 21.6k| const __m128i in = _mm_loadu_si128((__m128i *)input); 25| 21.6k| const uint16_t input_utf8_end_of_code_point_mask = 26| 21.6k| utf8_end_of_code_point_mask & 0xfff; 27| 21.6k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (27:7): [True: 86, False: 21.5k] ------------------ 28| | // We process the data in chunks of 12 bytes. 29| | // Note: using 16 bytes is unsafe, see issue_ossfuzz_71218 30| 86| __m128i ascii_first = _mm_cvtepu8_epi16(in); 31| 86| __m128i ascii_second = _mm_cvtepu8_epi16(_mm_srli_si128(in, 8)); 32| 86| if (big_endian) { ------------------ | Branch (32:9): [True: 86, Folded] ------------------ 33| 86| ascii_first = _mm_shuffle_epi8(ascii_first, swap); 34| 86| ascii_second = _mm_shuffle_epi8(ascii_second, swap); 35| 86| } 36| 86| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output), ascii_first); 37| 86| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + 8), 38| 86| ascii_second); 39| 86| utf16_output += 12; // We wrote 12 16-bit characters. 40| 86| return 12; // We consumed 12 bytes. 41| 86| } 42| 21.5k| if (((utf8_end_of_code_point_mask & 0xFFFF) == 0xaaaa)) { ------------------ | Branch (42:7): [True: 409, False: 21.1k] ------------------ 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| 409| const __m128i sh = 47| 409| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 48| 409| const __m128i perm = _mm_shuffle_epi8(in, sh); 49| 409| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 50| 409| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 51| 409| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 52| 409| if (big_endian) ------------------ | Branch (52:9): [True: 409, Folded] ------------------ 53| 409| composed = _mm_shuffle_epi8(composed, swap); 54| 409| _mm_storeu_si128((__m128i *)utf16_output, composed); 55| 409| utf16_output += 8; // We wrote 16 bytes, 8 code points. 56| 409| return 16; 57| 409| } 58| 21.1k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (58:7): [True: 5.97k, False: 15.1k] ------------------ 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| 5.97k| const __m128i sh = 63| 5.97k| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 64| 5.97k| const __m128i perm = _mm_shuffle_epi8(in, sh); 65| 5.97k| const __m128i ascii = 66| 5.97k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 67| 5.97k| const __m128i middlebyte = 68| 5.97k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 69| 5.97k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 70| 5.97k| const __m128i highbyte = 71| 5.97k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 72| 5.97k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 73| 5.97k| const __m128i composed = 74| 5.97k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 75| 5.97k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 76| 5.97k| if (big_endian) ------------------ | Branch (76:9): [True: 5.97k, Folded] ------------------ 77| 5.97k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 78| 5.97k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 79| 5.97k| utf16_output += 4; 80| 5.97k| return 12; 81| 5.97k| } 82| | /// We do not have a fast path available, so we fallback. 83| | 84| 15.1k| const uint8_t idx = 85| 15.1k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 86| 15.1k| const uint8_t consumed = 87| 15.1k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 88| 15.1k| if (idx < 64) { ------------------ | Branch (88:7): [True: 8.07k, False: 7.08k] ------------------ 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| 8.07k| const __m128i sh = 96| 8.07k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 97| 8.07k| const __m128i perm = _mm_shuffle_epi8(in, sh); 98| 8.07k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 99| 8.07k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 100| 8.07k| __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 101| 8.07k| if (big_endian) ------------------ | Branch (101:9): [True: 8.07k, Folded] ------------------ 102| 8.07k| composed = _mm_shuffle_epi8(composed, swap); 103| 8.07k| _mm_storeu_si128((__m128i *)utf16_output, composed); 104| 8.07k| utf16_output += 6; // We wrote 12 bytes, 6 code points. 105| 8.07k| } else if (idx < 145) { ------------------ | Branch (105:14): [True: 2.13k, False: 4.94k] ------------------ 106| | // FOUR (4) input code-code units 107| 2.13k| const __m128i sh = 108| 2.13k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 109| 2.13k| const __m128i perm = _mm_shuffle_epi8(in, sh); 110| 2.13k| const __m128i ascii = 111| 2.13k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 112| 2.13k| const __m128i middlebyte = 113| 2.13k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 114| 2.13k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 115| 2.13k| const __m128i highbyte = 116| 2.13k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 117| 2.13k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 118| 2.13k| const __m128i composed = 119| 2.13k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 120| 2.13k| __m128i composed_repacked = _mm_packus_epi32(composed, composed); 121| 2.13k| if (big_endian) ------------------ | Branch (121:9): [True: 2.13k, Folded] ------------------ 122| 2.13k| composed_repacked = _mm_shuffle_epi8(composed_repacked, swap); 123| 2.13k| _mm_storeu_si128((__m128i *)utf16_output, composed_repacked); 124| 2.13k| utf16_output += 4; 125| 4.94k| } else if (idx < 209) { ------------------ | Branch (125:14): [True: 4.66k, False: 289] ------------------ 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| 4.66k| const __m128i sh = 137| 4.66k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 138| 4.66k| const __m128i perm = _mm_shuffle_epi8(in, sh); 139| 4.66k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 140| 4.66k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 141| 4.66k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 142| 4.66k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 143| | // correct for spurious high bit 144| 4.66k| const __m128i correct = 145| 4.66k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 146| 4.66k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 147| 4.66k| 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| 4.66k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000)); 151| 4.66k| 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| 4.66k| const __m128i composed = 157| 4.66k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 158| 4.66k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 159| 4.66k| const __m128i composedminus = 160| 4.66k| _mm_sub_epi32(composed, _mm_set1_epi32(0x10000)); 161| 4.66k| const __m128i lowtenbits = 162| 4.66k| _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff)); 163| | // Notice the 0x3ff mask: 164| 4.66k| const __m128i hightenbits = 165| 4.66k| _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff)); 166| 4.66k| const __m128i lowtenbitsadd = 167| 4.66k| _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00)); 168| 4.66k| const __m128i hightenbitsadd = 169| 4.66k| _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800)); 170| 4.66k| const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16); 171| 4.66k| __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted); 172| 4.66k| uint32_t basic_buffer[4]; 173| 4.66k| uint32_t basic_buffer_swap[4]; 174| 4.66k| if (big_endian) { ------------------ | Branch (174:9): [True: 4.66k, Folded] ------------------ 175| 4.66k| _mm_storeu_si128((__m128i *)basic_buffer_swap, 176| 4.66k| _mm_shuffle_epi8(composed, swap)); 177| 4.66k| surrogates = _mm_shuffle_epi8(surrogates, swap); 178| 4.66k| } 179| 4.66k| _mm_storeu_si128((__m128i *)basic_buffer, composed); 180| 4.66k| uint32_t surrogate_buffer[4]; 181| 4.66k| _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates); 182| 18.6k| for (size_t i = 0; i < 3; i++) { ------------------ | Branch (182:24): [True: 13.9k, False: 4.66k] ------------------ 183| 13.9k| if (basic_buffer[i] > 0x3c00000) { ------------------ | Branch (183:11): [True: 3.01k, False: 10.9k] ------------------ 184| 3.01k| utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff); 185| 3.01k| utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16); 186| 3.01k| utf16_output += 2; 187| 10.9k| } else { 188| 10.9k| utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) ------------------ | Branch (188:27): [True: 10.9k, Folded] ------------------ 189| 10.9k| : uint16_t(basic_buffer[i]); 190| 10.9k| utf16_output++; 191| 10.9k| } 192| 13.9k| } 193| 4.66k| } else { 194| | // here we know that there is an error but we do not handle errors 195| 289| } 196| 15.1k| return consumed; 197| 21.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_128convert_masked_utf8_to_utf32EPKcmRPDi: 9| 14.3k| 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| 14.3k| const __m128i in = _mm_loadu_si128((__m128i *)input); 22| 14.3k| const uint16_t input_utf8_end_of_code_point_mask = 23| 14.3k| utf8_end_of_code_point_mask & 0xfff; 24| 14.3k| if (utf8_end_of_code_point_mask == 0xfff) { ------------------ | Branch (24:7): [True: 85, False: 14.2k] ------------------ 25| | // We process the data in chunks of 12 bytes. 26| 85| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), 27| 85| _mm_cvtepu8_epi32(in)); 28| 85| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 4), 29| 85| _mm_cvtepu8_epi32(_mm_srli_si128(in, 4))); 30| 85| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 8), 31| 85| _mm_cvtepu8_epi32(_mm_srli_si128(in, 8))); 32| 85| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 12), 33| 85| _mm_cvtepu8_epi32(_mm_srli_si128(in, 12))); 34| 85| utf32_output += 12; // We wrote 12 32-bit characters. 35| 85| return 12; // We consumed 12 bytes. 36| 85| } 37| 14.2k| if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) { ------------------ | Branch (37:7): [True: 117, False: 14.1k] ------------------ 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| 117| const __m128i sh = 42| 117| _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14); 43| 117| const __m128i perm = _mm_shuffle_epi8(in, sh); 44| 117| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 45| 117| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 46| 117| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 47| 117| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), 48| 117| _mm_cvtepu16_epi32(composed)); 49| 117| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 4), 50| 117| _mm_cvtepu16_epi32(_mm_srli_si128(composed, 8))); 51| 117| utf32_output += 8; // We wrote 32 bytes, 8 code points. 52| 117| return 16; 53| 117| } 54| 14.1k| if (input_utf8_end_of_code_point_mask == 0x924) { ------------------ | Branch (54:7): [True: 272, False: 13.8k] ------------------ 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| 272| const __m128i sh = 59| 272| _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1); 60| 272| const __m128i perm = _mm_shuffle_epi8(in, sh); 61| 272| const __m128i ascii = 62| 272| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 63| 272| const __m128i middlebyte = 64| 272| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 65| 272| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 66| 272| const __m128i highbyte = 67| 272| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 68| 272| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 69| 272| const __m128i composed = 70| 272| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 71| 272| _mm_storeu_si128((__m128i *)utf32_output, composed); 72| 272| utf32_output += 4; 73| 272| return 12; 74| 272| } 75| | /// We do not have a fast path available, so we fallback. 76| | 77| 13.8k| const uint8_t idx = 78| 13.8k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0]; 79| 13.8k| const uint8_t consumed = 80| 13.8k| tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1]; 81| 13.8k| if (idx < 64) { ------------------ | Branch (81:7): [True: 7.91k, False: 5.94k] ------------------ 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| 7.91k| const __m128i sh = 89| 7.91k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 90| 7.91k| const __m128i perm = _mm_shuffle_epi8(in, sh); 91| 7.91k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f)); 92| 7.91k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00)); 93| 7.91k| const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2)); 94| 7.91k| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), 95| 7.91k| _mm_cvtepu16_epi32(composed)); 96| 7.91k| _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output + 4), 97| 7.91k| _mm_cvtepu16_epi32(_mm_srli_si128(composed, 8))); 98| 7.91k| utf32_output += 6; // We wrote 12 bytes, 6 code points. 99| 7.91k| } else if (idx < 145) { ------------------ | Branch (99:14): [True: 1.43k, False: 4.50k] ------------------ 100| | // FOUR (4) input code-code units 101| 1.43k| const __m128i sh = 102| 1.43k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 103| 1.43k| const __m128i perm = _mm_shuffle_epi8(in, sh); 104| 1.43k| const __m128i ascii = 105| 1.43k| _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits 106| 1.43k| const __m128i middlebyte = 107| 1.43k| _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits 108| 1.43k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 109| 1.43k| const __m128i highbyte = 110| 1.43k| _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits 111| 1.43k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4); 112| 1.43k| const __m128i composed = 113| 1.43k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted); 114| 1.43k| _mm_storeu_si128((__m128i *)utf32_output, composed); 115| 1.43k| utf32_output += 4; 116| 4.50k| } else if (idx < 209) { ------------------ | Branch (116:14): [True: 4.22k, False: 289] ------------------ 117| | // TWO (2) input code-code units 118| 4.22k| const __m128i sh = 119| 4.22k| _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]); 120| 4.22k| const __m128i perm = _mm_shuffle_epi8(in, sh); 121| 4.22k| const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f)); 122| 4.22k| const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); 123| 4.22k| const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2); 124| 4.22k| __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000)); 125| | // correct for spurious high bit 126| 4.22k| const __m128i correct = 127| 4.22k| _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1); 128| 4.22k| middlehighbyte = _mm_xor_si128(correct, middlehighbyte); 129| 4.22k| const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4); 130| 4.22k| const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000)); 131| 4.22k| const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6); 132| 4.22k| const __m128i composed = 133| 4.22k| _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), 134| 4.22k| _mm_or_si128(highbyte_shifted, middlehighbyte_shifted)); 135| 4.22k| _mm_storeu_si128((__m128i *)utf32_output, composed); 136| 4.22k| utf32_output += 3; 137| 4.22k| } else { 138| | // here we know that there is an error but we do not handle errors 139| 289| } 140| 13.8k| return consumed; 141| 14.1k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_123utf16_gather_high_bytesILNS_10endiannessE0EEENS1_4simd5simd8IhEENS4_6simd16ItEES8_: 3| 17.6k| const simd16 in1) { 4| 17.6k| if (big_endian) { ------------------ | Branch (4:7): [Folded, False: 17.6k] ------------------ 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| 17.6k| } else { 12| 17.6k| const auto t0 = in0.shr<8>(); 13| 17.6k| const auto t1 = in1.shr<8>(); 14| | 15| 17.6k| return simd16::pack(t0, t1); 16| 17.6k| } 17| 17.6k|} simdutf.cpp:_ZN7simdutf8westmere12_GLOBAL__N_123utf16_gather_high_bytesILNS_10endiannessE1EEENS1_4simd5simd8IhEENS4_6simd16ItEES8_: 3| 17.8k| const simd16 in1) { 4| 17.8k| if (big_endian) { ------------------ | Branch (4:7): [True: 17.8k, Folded] ------------------ 5| | // we want lower bytes 6| 17.8k| const auto mask = simd16(0x00ff); 7| 17.8k| const auto t0 = in0 & mask; 8| 17.8k| const auto t1 = in1 & mask; 9| | 10| 17.8k| return simd16::pack(t0, t1); 11| 17.8k| } 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| 17.8k|}