1 //===- Writer.cpp ---------------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "Writer.h" 10 #include "Config.h" 11 #include "InputChunks.h" 12 #include "InputElement.h" 13 #include "MapFile.h" 14 #include "OutputSections.h" 15 #include "OutputSegment.h" 16 #include "Relocations.h" 17 #include "SymbolTable.h" 18 #include "SyntheticSections.h" 19 #include "WriterUtils.h" 20 #include "lld/Common/CommonLinkerContext.h" 21 #include "lld/Common/Strings.h" 22 #include "llvm/ADT/DenseSet.h" 23 #include "llvm/ADT/SmallSet.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringMap.h" 26 #include "llvm/BinaryFormat/Wasm.h" 27 #include "llvm/BinaryFormat/WasmTraits.h" 28 #include "llvm/Support/FileOutputBuffer.h" 29 #include "llvm/Support/Format.h" 30 #include "llvm/Support/FormatVariadic.h" 31 #include "llvm/Support/LEB128.h" 32 #include "llvm/Support/Parallel.h" 33 34 #include <cstdarg> 35 #include <map> 36 37 #define DEBUG_TYPE "lld" 38 39 using namespace llvm; 40 using namespace llvm::wasm; 41 42 namespace lld { 43 namespace wasm { 44 static constexpr int stackAlignment = 16; 45 static constexpr int heapAlignment = 16; 46 47 namespace { 48 49 // The writer writes a SymbolTable result to a file. 50 class Writer { 51 public: 52 void run(); 53 54 private: 55 void openFile(); 56 57 bool needsPassiveInitialization(const OutputSegment *segment); 58 bool hasPassiveInitializedSegments(); 59 60 void createSyntheticInitFunctions(); 61 void createInitMemoryFunction(); 62 void createStartFunction(); 63 void createApplyDataRelocationsFunction(); 64 void createApplyGlobalRelocationsFunction(); 65 void createApplyGlobalTLSRelocationsFunction(); 66 void createCallCtorsFunction(); 67 void createInitTLSFunction(); 68 void createCommandExportWrappers(); 69 void createCommandExportWrapper(uint32_t functionIndex, DefinedFunction *f); 70 71 void assignIndexes(); 72 void populateSymtab(); 73 void populateProducers(); 74 void populateTargetFeatures(); 75 // populateTargetFeatures happens early on so some checks are delayed 76 // until imports and exports are finalized. There are run unstead 77 // in checkImportExportTargetFeatures 78 void checkImportExportTargetFeatures(); 79 void calculateInitFunctions(); 80 void calculateImports(); 81 void calculateExports(); 82 void calculateCustomSections(); 83 void calculateTypes(); 84 void createOutputSegments(); 85 OutputSegment *createOutputSegment(StringRef name); 86 void combineOutputSegments(); 87 void layoutMemory(); 88 void createHeader(); 89 90 void addSection(OutputSection *sec); 91 92 void addSections(); 93 94 void createCustomSections(); 95 void createSyntheticSections(); 96 void createSyntheticSectionsPostLayout(); 97 void finalizeSections(); 98 99 // Custom sections 100 void createRelocSections(); 101 102 void writeHeader(); 103 void writeSections(); 104 105 uint64_t fileSize = 0; 106 107 std::vector<WasmInitEntry> initFunctions; 108 llvm::StringMap<std::vector<InputChunk *>> customSectionMapping; 109 110 // Stable storage for command export wrapper function name strings. 111 std::list<std::string> commandExportWrapperNames; 112 113 // Elements that are used to construct the final output 114 std::string header; 115 std::vector<OutputSection *> outputSections; 116 117 std::unique_ptr<FileOutputBuffer> buffer; 118 119 std::vector<OutputSegment *> segments; 120 llvm::SmallDenseMap<StringRef, OutputSegment *> segmentMap; 121 }; 122 123 } // anonymous namespace 124 125 void Writer::calculateCustomSections() { 126 log("calculateCustomSections"); 127 bool stripDebug = config->stripDebug || config->stripAll; 128 for (ObjFile *file : symtab->objectFiles) { 129 for (InputChunk *section : file->customSections) { 130 // Exclude COMDAT sections that are not selected for inclusion 131 if (section->discarded) 132 continue; 133 StringRef name = section->getName(); 134 // These custom sections are known the linker and synthesized rather than 135 // blindly copied. 136 if (name == "linking" || name == "name" || name == "producers" || 137 name == "target_features" || name.startswith("reloc.")) 138 continue; 139 // These custom sections are generated by `clang -fembed-bitcode`. 140 // These are used by the rust toolchain to ship LTO data along with 141 // compiled object code, but they don't want this included in the linker 142 // output. 143 if (name == ".llvmbc" || name == ".llvmcmd") 144 continue; 145 // Strip debug section in that option was specified. 146 if (stripDebug && name.startswith(".debug_")) 147 continue; 148 // Otherwise include custom sections by default and concatenate their 149 // contents. 150 customSectionMapping[name].push_back(section); 151 } 152 } 153 } 154 155 void Writer::createCustomSections() { 156 log("createCustomSections"); 157 for (auto &pair : customSectionMapping) { 158 StringRef name = pair.first(); 159 LLVM_DEBUG(dbgs() << "createCustomSection: " << name << "\n"); 160 161 OutputSection *sec = make<CustomSection>(std::string(name), pair.second); 162 if (config->relocatable || config->emitRelocs) { 163 auto *sym = make<OutputSectionSymbol>(sec); 164 out.linkingSec->addToSymtab(sym); 165 sec->sectionSym = sym; 166 } 167 addSection(sec); 168 } 169 } 170 171 // Create relocations sections in the final output. 172 // These are only created when relocatable output is requested. 173 void Writer::createRelocSections() { 174 log("createRelocSections"); 175 // Don't use iterator here since we are adding to OutputSection 176 size_t origSize = outputSections.size(); 177 for (size_t i = 0; i < origSize; i++) { 178 LLVM_DEBUG(dbgs() << "check section " << i << "\n"); 179 OutputSection *sec = outputSections[i]; 180 181 // Count the number of needed sections. 182 uint32_t count = sec->getNumRelocations(); 183 if (!count) 184 continue; 185 186 StringRef name; 187 if (sec->type == WASM_SEC_DATA) 188 name = "reloc.DATA"; 189 else if (sec->type == WASM_SEC_CODE) 190 name = "reloc.CODE"; 191 else if (sec->type == WASM_SEC_CUSTOM) 192 name = saver().save("reloc." + sec->name); 193 else 194 llvm_unreachable( 195 "relocations only supported for code, data, or custom sections"); 196 197 addSection(make<RelocSection>(name, sec)); 198 } 199 } 200 201 void Writer::populateProducers() { 202 for (ObjFile *file : symtab->objectFiles) { 203 const WasmProducerInfo &info = file->getWasmObj()->getProducerInfo(); 204 out.producersSec->addInfo(info); 205 } 206 } 207 208 void Writer::writeHeader() { 209 memcpy(buffer->getBufferStart(), header.data(), header.size()); 210 } 211 212 void Writer::writeSections() { 213 uint8_t *buf = buffer->getBufferStart(); 214 parallelForEach(outputSections, [buf](OutputSection *s) { 215 assert(s->isNeeded()); 216 s->writeTo(buf); 217 }); 218 } 219 220 static void setGlobalPtr(DefinedGlobal *g, uint64_t memoryPtr) { 221 g->global->setPointerValue(memoryPtr); 222 } 223 224 // Fix the memory layout of the output binary. This assigns memory offsets 225 // to each of the input data sections as well as the explicit stack region. 226 // The default memory layout is as follows, from low to high. 227 // 228 // - initialized data (starting at Config->globalBase) 229 // - BSS data (not currently implemented in llvm) 230 // - explicit stack (Config->ZStackSize) 231 // - heap start / unallocated 232 // 233 // The --stack-first option means that stack is placed before any static data. 234 // This can be useful since it means that stack overflow traps immediately 235 // rather than overwriting global data, but also increases code size since all 236 // static data loads and stores requires larger offsets. 237 void Writer::layoutMemory() { 238 uint64_t memoryPtr = 0; 239 240 auto placeStack = [&]() { 241 if (config->relocatable || config->isPic) 242 return; 243 memoryPtr = alignTo(memoryPtr, stackAlignment); 244 if (config->zStackSize != alignTo(config->zStackSize, stackAlignment)) 245 error("stack size must be " + Twine(stackAlignment) + "-byte aligned"); 246 log("mem: stack size = " + Twine(config->zStackSize)); 247 log("mem: stack base = " + Twine(memoryPtr)); 248 memoryPtr += config->zStackSize; 249 setGlobalPtr(cast<DefinedGlobal>(WasmSym::stackPointer), memoryPtr); 250 log("mem: stack top = " + Twine(memoryPtr)); 251 }; 252 253 if (config->stackFirst) { 254 placeStack(); 255 } else { 256 memoryPtr = config->globalBase; 257 log("mem: global base = " + Twine(config->globalBase)); 258 } 259 260 if (WasmSym::globalBase) 261 WasmSym::globalBase->setVA(memoryPtr); 262 263 uint64_t dataStart = memoryPtr; 264 265 // Arbitrarily set __dso_handle handle to point to the start of the data 266 // segments. 267 if (WasmSym::dsoHandle) 268 WasmSym::dsoHandle->setVA(dataStart); 269 270 out.dylinkSec->memAlign = 0; 271 for (OutputSegment *seg : segments) { 272 out.dylinkSec->memAlign = std::max(out.dylinkSec->memAlign, seg->alignment); 273 memoryPtr = alignTo(memoryPtr, 1ULL << seg->alignment); 274 seg->startVA = memoryPtr; 275 log(formatv("mem: {0,-15} offset={1,-8} size={2,-8} align={3}", seg->name, 276 memoryPtr, seg->size, seg->alignment)); 277 278 if (!config->relocatable && seg->isTLS()) { 279 if (config->sharedMemory) { 280 auto *tlsSize = cast<DefinedGlobal>(WasmSym::tlsSize); 281 setGlobalPtr(tlsSize, seg->size); 282 283 auto *tlsAlign = cast<DefinedGlobal>(WasmSym::tlsAlign); 284 setGlobalPtr(tlsAlign, int64_t{1} << seg->alignment); 285 } else if (WasmSym::tlsBase) { 286 auto *tlsBase = cast<DefinedGlobal>(WasmSym::tlsBase); 287 setGlobalPtr(tlsBase, memoryPtr); 288 } 289 } 290 291 memoryPtr += seg->size; 292 } 293 294 // Make space for the memory initialization flag 295 if (config->sharedMemory && hasPassiveInitializedSegments()) { 296 memoryPtr = alignTo(memoryPtr, 4); 297 WasmSym::initMemoryFlag = symtab->addSyntheticDataSymbol( 298 "__wasm_init_memory_flag", WASM_SYMBOL_VISIBILITY_HIDDEN); 299 WasmSym::initMemoryFlag->markLive(); 300 WasmSym::initMemoryFlag->setVA(memoryPtr); 301 log(formatv("mem: {0,-15} offset={1,-8} size={2,-8} align={3}", 302 "__wasm_init_memory_flag", memoryPtr, 4, 4)); 303 memoryPtr += 4; 304 } 305 306 if (WasmSym::dataEnd) 307 WasmSym::dataEnd->setVA(memoryPtr); 308 309 uint64_t staticDataSize = memoryPtr - dataStart; 310 log("mem: static data = " + Twine(staticDataSize)); 311 if (config->isPic) 312 out.dylinkSec->memSize = staticDataSize; 313 314 if (!config->stackFirst) 315 placeStack(); 316 317 if (WasmSym::heapBase) { 318 // Set `__heap_base` to follow the end of the stack or global data. The 319 // fact that this comes last means that a malloc/brk implementation can 320 // grow the heap at runtime. 321 // We'll align the heap base here because memory allocators might expect 322 // __heap_base to be aligned already. 323 memoryPtr = alignTo(memoryPtr, heapAlignment); 324 log("mem: heap base = " + Twine(memoryPtr)); 325 WasmSym::heapBase->setVA(memoryPtr); 326 } 327 328 uint64_t maxMemorySetting = 1ULL 329 << (config->is64.getValueOr(false) ? 48 : 32); 330 331 if (config->initialMemory != 0) { 332 if (config->initialMemory != alignTo(config->initialMemory, WasmPageSize)) 333 error("initial memory must be " + Twine(WasmPageSize) + "-byte aligned"); 334 if (memoryPtr > config->initialMemory) 335 error("initial memory too small, " + Twine(memoryPtr) + " bytes needed"); 336 if (config->initialMemory > maxMemorySetting) 337 error("initial memory too large, cannot be greater than " + 338 Twine(maxMemorySetting)); 339 memoryPtr = config->initialMemory; 340 } 341 out.memorySec->numMemoryPages = 342 alignTo(memoryPtr, WasmPageSize) / WasmPageSize; 343 log("mem: total pages = " + Twine(out.memorySec->numMemoryPages)); 344 345 if (config->maxMemory != 0) { 346 if (config->maxMemory != alignTo(config->maxMemory, WasmPageSize)) 347 error("maximum memory must be " + Twine(WasmPageSize) + "-byte aligned"); 348 if (memoryPtr > config->maxMemory) 349 error("maximum memory too small, " + Twine(memoryPtr) + " bytes needed"); 350 if (config->maxMemory > maxMemorySetting) 351 error("maximum memory too large, cannot be greater than " + 352 Twine(maxMemorySetting)); 353 } 354 355 // Check max if explicitly supplied or required by shared memory 356 if (config->maxMemory != 0 || config->sharedMemory) { 357 uint64_t max = config->maxMemory; 358 if (max == 0) { 359 // If no maxMemory config was supplied but we are building with 360 // shared memory, we need to pick a sensible upper limit. 361 if (config->isPic) 362 max = maxMemorySetting; 363 else 364 max = alignTo(memoryPtr, WasmPageSize); 365 } 366 out.memorySec->maxMemoryPages = max / WasmPageSize; 367 log("mem: max pages = " + Twine(out.memorySec->maxMemoryPages)); 368 } 369 } 370 371 void Writer::addSection(OutputSection *sec) { 372 if (!sec->isNeeded()) 373 return; 374 log("addSection: " + toString(*sec)); 375 sec->sectionIndex = outputSections.size(); 376 outputSections.push_back(sec); 377 } 378 379 // If a section name is valid as a C identifier (which is rare because of 380 // the leading '.'), linkers are expected to define __start_<secname> and 381 // __stop_<secname> symbols. They are at beginning and end of the section, 382 // respectively. This is not requested by the ELF standard, but GNU ld and 383 // gold provide the feature, and used by many programs. 384 static void addStartStopSymbols(const OutputSegment *seg) { 385 StringRef name = seg->name; 386 if (!isValidCIdentifier(name)) 387 return; 388 LLVM_DEBUG(dbgs() << "addStartStopSymbols: " << name << "\n"); 389 uint64_t start = seg->startVA; 390 uint64_t stop = start + seg->size; 391 symtab->addOptionalDataSymbol(saver().save("__start_" + name), start); 392 symtab->addOptionalDataSymbol(saver().save("__stop_" + name), stop); 393 } 394 395 void Writer::addSections() { 396 addSection(out.dylinkSec); 397 addSection(out.typeSec); 398 addSection(out.importSec); 399 addSection(out.functionSec); 400 addSection(out.tableSec); 401 addSection(out.memorySec); 402 addSection(out.tagSec); 403 addSection(out.globalSec); 404 addSection(out.exportSec); 405 addSection(out.startSec); 406 addSection(out.elemSec); 407 addSection(out.dataCountSec); 408 409 addSection(make<CodeSection>(out.functionSec->inputFunctions)); 410 addSection(make<DataSection>(segments)); 411 412 createCustomSections(); 413 414 addSection(out.linkingSec); 415 if (config->emitRelocs || config->relocatable) { 416 createRelocSections(); 417 } 418 419 addSection(out.nameSec); 420 addSection(out.producersSec); 421 addSection(out.targetFeaturesSec); 422 } 423 424 void Writer::finalizeSections() { 425 for (OutputSection *s : outputSections) { 426 s->setOffset(fileSize); 427 s->finalizeContents(); 428 fileSize += s->getSize(); 429 } 430 } 431 432 void Writer::populateTargetFeatures() { 433 StringMap<std::string> used; 434 StringMap<std::string> required; 435 StringMap<std::string> disallowed; 436 SmallSet<std::string, 8> &allowed = out.targetFeaturesSec->features; 437 bool tlsUsed = false; 438 439 if (config->isPic) { 440 // This should not be necessary because all PIC objects should 441 // contain the mutable-globals feature. 442 // TODO(https://bugs.llvm.org/show_bug.cgi?id=52339) 443 allowed.insert("mutable-globals"); 444 } 445 446 // Only infer used features if user did not specify features 447 bool inferFeatures = !config->features.hasValue(); 448 449 if (!inferFeatures) { 450 auto &explicitFeatures = config->features.getValue(); 451 allowed.insert(explicitFeatures.begin(), explicitFeatures.end()); 452 if (!config->checkFeatures) 453 return; 454 } 455 456 // Find the sets of used, required, and disallowed features 457 for (ObjFile *file : symtab->objectFiles) { 458 StringRef fileName(file->getName()); 459 for (auto &feature : file->getWasmObj()->getTargetFeatures()) { 460 switch (feature.Prefix) { 461 case WASM_FEATURE_PREFIX_USED: 462 used.insert({feature.Name, std::string(fileName)}); 463 break; 464 case WASM_FEATURE_PREFIX_REQUIRED: 465 used.insert({feature.Name, std::string(fileName)}); 466 required.insert({feature.Name, std::string(fileName)}); 467 break; 468 case WASM_FEATURE_PREFIX_DISALLOWED: 469 disallowed.insert({feature.Name, std::string(fileName)}); 470 break; 471 default: 472 error("Unrecognized feature policy prefix " + 473 std::to_string(feature.Prefix)); 474 } 475 } 476 477 // Find TLS data segments 478 auto isTLS = [](InputChunk *segment) { 479 return segment->live && segment->isTLS(); 480 }; 481 tlsUsed = tlsUsed || llvm::any_of(file->segments, isTLS); 482 } 483 484 if (inferFeatures) 485 for (const auto &key : used.keys()) 486 allowed.insert(std::string(key)); 487 488 if (!config->checkFeatures) 489 return; 490 491 if (config->sharedMemory) { 492 if (disallowed.count("shared-mem")) 493 error("--shared-memory is disallowed by " + disallowed["shared-mem"] + 494 " because it was not compiled with 'atomics' or 'bulk-memory' " 495 "features."); 496 497 for (auto feature : {"atomics", "bulk-memory"}) 498 if (!allowed.count(feature)) 499 error(StringRef("'") + feature + 500 "' feature must be used in order to use shared memory"); 501 } 502 503 if (tlsUsed) { 504 for (auto feature : {"atomics", "bulk-memory"}) 505 if (!allowed.count(feature)) 506 error(StringRef("'") + feature + 507 "' feature must be used in order to use thread-local storage"); 508 } 509 510 // Validate that used features are allowed in output 511 if (!inferFeatures) { 512 for (const auto &feature : used.keys()) { 513 if (!allowed.count(std::string(feature))) 514 error(Twine("Target feature '") + feature + "' used by " + 515 used[feature] + " is not allowed."); 516 } 517 } 518 519 // Validate the required and disallowed constraints for each file 520 for (ObjFile *file : symtab->objectFiles) { 521 StringRef fileName(file->getName()); 522 SmallSet<std::string, 8> objectFeatures; 523 for (const auto &feature : file->getWasmObj()->getTargetFeatures()) { 524 if (feature.Prefix == WASM_FEATURE_PREFIX_DISALLOWED) 525 continue; 526 objectFeatures.insert(feature.Name); 527 if (disallowed.count(feature.Name)) 528 error(Twine("Target feature '") + feature.Name + "' used in " + 529 fileName + " is disallowed by " + disallowed[feature.Name] + 530 ". Use --no-check-features to suppress."); 531 } 532 for (const auto &feature : required.keys()) { 533 if (!objectFeatures.count(std::string(feature))) 534 error(Twine("Missing target feature '") + feature + "' in " + fileName + 535 ", required by " + required[feature] + 536 ". Use --no-check-features to suppress."); 537 } 538 } 539 540 // Normally we don't include bss segments in the binary. In particular if 541 // memory is not being imported then we can assume its zero initialized. 542 // In the case the memory is imported, we and we can use the memory.fill 543 // instrction than we can also avoid inluding the segments. 544 if (config->importMemory && !allowed.count("bulk-memory")) 545 config->emitBssSegments = true; 546 } 547 548 void Writer::checkImportExportTargetFeatures() { 549 if (config->relocatable || !config->checkFeatures) 550 return; 551 552 if (out.targetFeaturesSec->features.count("mutable-globals") == 0) { 553 for (const Symbol *sym : out.importSec->importedSymbols) { 554 if (auto *global = dyn_cast<GlobalSymbol>(sym)) { 555 if (global->getGlobalType()->Mutable) { 556 error(Twine("mutable global imported but 'mutable-globals' feature " 557 "not present in inputs: `") + 558 toString(*sym) + "`. Use --no-check-features to suppress."); 559 } 560 } 561 } 562 for (const Symbol *sym : out.exportSec->exportedSymbols) { 563 if (isa<GlobalSymbol>(sym)) { 564 error(Twine("mutable global exported but 'mutable-globals' feature " 565 "not present in inputs: `") + 566 toString(*sym) + "`. Use --no-check-features to suppress."); 567 } 568 } 569 } 570 } 571 572 static bool shouldImport(Symbol *sym) { 573 // We don't generate imports for data symbols. They however can be imported 574 // as GOT entries. 575 if (isa<DataSymbol>(sym)) 576 return false; 577 if (!sym->isLive()) 578 return false; 579 if (!sym->isUsedInRegularObj) 580 return false; 581 582 // When a symbol is weakly defined in a shared library we need to allow 583 // it to be overridden by another module so need to both import 584 // and export the symbol. 585 if (config->shared && sym->isDefined() && sym->isWeak()) 586 return true; 587 if (!sym->isUndefined()) 588 return false; 589 if (sym->isWeak() && !config->relocatable && !config->isPic) 590 return false; 591 592 // In PIC mode we only need to import functions when they are called directly. 593 // Indirect usage all goes via GOT imports. 594 if (config->isPic) { 595 if (auto *f = dyn_cast<UndefinedFunction>(sym)) 596 if (!f->isCalledDirectly) 597 return false; 598 } 599 600 if (config->isPic || config->relocatable || config->importUndefined) 601 return true; 602 if (config->allowUndefinedSymbols.count(sym->getName()) != 0) 603 return true; 604 605 return sym->importName.hasValue(); 606 } 607 608 void Writer::calculateImports() { 609 // Some inputs require that the indirect function table be assigned to table 610 // number 0, so if it is present and is an import, allocate it before any 611 // other tables. 612 if (WasmSym::indirectFunctionTable && 613 shouldImport(WasmSym::indirectFunctionTable)) 614 out.importSec->addImport(WasmSym::indirectFunctionTable); 615 616 for (Symbol *sym : symtab->getSymbols()) { 617 if (!shouldImport(sym)) 618 continue; 619 if (sym == WasmSym::indirectFunctionTable) 620 continue; 621 LLVM_DEBUG(dbgs() << "import: " << sym->getName() << "\n"); 622 out.importSec->addImport(sym); 623 } 624 } 625 626 void Writer::calculateExports() { 627 if (config->relocatable) 628 return; 629 630 if (!config->relocatable && !config->importMemory) 631 out.exportSec->exports.push_back( 632 WasmExport{"memory", WASM_EXTERNAL_MEMORY, 0}); 633 634 unsigned globalIndex = 635 out.importSec->getNumImportedGlobals() + out.globalSec->numGlobals(); 636 637 for (Symbol *sym : symtab->getSymbols()) { 638 if (!sym->isExported()) 639 continue; 640 if (!sym->isLive()) 641 continue; 642 643 StringRef name = sym->getName(); 644 WasmExport export_; 645 if (auto *f = dyn_cast<DefinedFunction>(sym)) { 646 if (Optional<StringRef> exportName = f->function->getExportName()) { 647 name = *exportName; 648 } 649 export_ = {name, WASM_EXTERNAL_FUNCTION, f->getExportedFunctionIndex()}; 650 } else if (auto *g = dyn_cast<DefinedGlobal>(sym)) { 651 if (g->getGlobalType()->Mutable && !g->getFile() && !g->forceExport) { 652 // Avoid exporting mutable globals are linker synthesized (e.g. 653 // __stack_pointer or __tls_base) unless they are explicitly exported 654 // from the command line. 655 // Without this check `--export-all` would cause any program using the 656 // stack pointer to export a mutable global even if none of the input 657 // files were built with the `mutable-globals` feature. 658 continue; 659 } 660 export_ = {name, WASM_EXTERNAL_GLOBAL, g->getGlobalIndex()}; 661 } else if (auto *t = dyn_cast<DefinedTag>(sym)) { 662 export_ = {name, WASM_EXTERNAL_TAG, t->getTagIndex()}; 663 } else if (auto *d = dyn_cast<DefinedData>(sym)) { 664 out.globalSec->dataAddressGlobals.push_back(d); 665 export_ = {name, WASM_EXTERNAL_GLOBAL, globalIndex++}; 666 } else { 667 auto *t = cast<DefinedTable>(sym); 668 export_ = {name, WASM_EXTERNAL_TABLE, t->getTableNumber()}; 669 } 670 671 LLVM_DEBUG(dbgs() << "Export: " << name << "\n"); 672 out.exportSec->exports.push_back(export_); 673 out.exportSec->exportedSymbols.push_back(sym); 674 } 675 } 676 677 void Writer::populateSymtab() { 678 if (!config->relocatable && !config->emitRelocs) 679 return; 680 681 for (Symbol *sym : symtab->getSymbols()) 682 if (sym->isUsedInRegularObj && sym->isLive()) 683 out.linkingSec->addToSymtab(sym); 684 685 for (ObjFile *file : symtab->objectFiles) { 686 LLVM_DEBUG(dbgs() << "Local symtab entries: " << file->getName() << "\n"); 687 for (Symbol *sym : file->getSymbols()) 688 if (sym->isLocal() && !isa<SectionSymbol>(sym) && sym->isLive()) 689 out.linkingSec->addToSymtab(sym); 690 } 691 } 692 693 void Writer::calculateTypes() { 694 // The output type section is the union of the following sets: 695 // 1. Any signature used in the TYPE relocation 696 // 2. The signatures of all imported functions 697 // 3. The signatures of all defined functions 698 // 4. The signatures of all imported tags 699 // 5. The signatures of all defined tags 700 701 for (ObjFile *file : symtab->objectFiles) { 702 ArrayRef<WasmSignature> types = file->getWasmObj()->types(); 703 for (uint32_t i = 0; i < types.size(); i++) 704 if (file->typeIsUsed[i]) 705 file->typeMap[i] = out.typeSec->registerType(types[i]); 706 } 707 708 for (const Symbol *sym : out.importSec->importedSymbols) { 709 if (auto *f = dyn_cast<FunctionSymbol>(sym)) 710 out.typeSec->registerType(*f->signature); 711 else if (auto *t = dyn_cast<TagSymbol>(sym)) 712 out.typeSec->registerType(*t->signature); 713 } 714 715 for (const InputFunction *f : out.functionSec->inputFunctions) 716 out.typeSec->registerType(f->signature); 717 718 for (const InputTag *t : out.tagSec->inputTags) 719 out.typeSec->registerType(t->signature); 720 } 721 722 // In a command-style link, create a wrapper for each exported symbol 723 // which calls the constructors and destructors. 724 void Writer::createCommandExportWrappers() { 725 // This logic doesn't currently support Emscripten-style PIC mode. 726 assert(!config->isPic); 727 728 // If there are no ctors and there's no libc `__wasm_call_dtors` to 729 // call, don't wrap the exports. 730 if (initFunctions.empty() && WasmSym::callDtors == nullptr) 731 return; 732 733 std::vector<DefinedFunction *> toWrap; 734 735 for (Symbol *sym : symtab->getSymbols()) 736 if (sym->isExported()) 737 if (auto *f = dyn_cast<DefinedFunction>(sym)) 738 toWrap.push_back(f); 739 740 for (auto *f : toWrap) { 741 auto funcNameStr = (f->getName() + ".command_export").str(); 742 commandExportWrapperNames.push_back(funcNameStr); 743 const std::string &funcName = commandExportWrapperNames.back(); 744 745 auto func = make<SyntheticFunction>(*f->getSignature(), funcName); 746 if (f->function->getExportName().hasValue()) 747 func->setExportName(f->function->getExportName()->str()); 748 else 749 func->setExportName(f->getName().str()); 750 751 DefinedFunction *def = 752 symtab->addSyntheticFunction(funcName, f->flags, func); 753 def->markLive(); 754 755 def->flags |= WASM_SYMBOL_EXPORTED; 756 def->flags &= ~WASM_SYMBOL_VISIBILITY_HIDDEN; 757 def->forceExport = f->forceExport; 758 759 f->flags |= WASM_SYMBOL_VISIBILITY_HIDDEN; 760 f->flags &= ~WASM_SYMBOL_EXPORTED; 761 f->forceExport = false; 762 763 out.functionSec->addFunction(func); 764 765 createCommandExportWrapper(f->getFunctionIndex(), def); 766 } 767 } 768 769 static void finalizeIndirectFunctionTable() { 770 if (!WasmSym::indirectFunctionTable) 771 return; 772 773 if (shouldImport(WasmSym::indirectFunctionTable) && 774 !WasmSym::indirectFunctionTable->hasTableNumber()) { 775 // Processing -Bsymbolic relocations resulted in a late requirement that the 776 // indirect function table be present, and we are running in --import-table 777 // mode. Add the table now to the imports section. Otherwise it will be 778 // added to the tables section later in assignIndexes. 779 out.importSec->addImport(WasmSym::indirectFunctionTable); 780 } 781 782 uint32_t tableSize = config->tableBase + out.elemSec->numEntries(); 783 WasmLimits limits = {0, tableSize, 0}; 784 if (WasmSym::indirectFunctionTable->isDefined() && !config->growableTable) { 785 limits.Flags |= WASM_LIMITS_FLAG_HAS_MAX; 786 limits.Maximum = limits.Minimum; 787 } 788 WasmSym::indirectFunctionTable->setLimits(limits); 789 } 790 791 static void scanRelocations() { 792 for (ObjFile *file : symtab->objectFiles) { 793 LLVM_DEBUG(dbgs() << "scanRelocations: " << file->getName() << "\n"); 794 for (InputChunk *chunk : file->functions) 795 scanRelocations(chunk); 796 for (InputChunk *chunk : file->segments) 797 scanRelocations(chunk); 798 for (auto &p : file->customSections) 799 scanRelocations(p); 800 } 801 } 802 803 void Writer::assignIndexes() { 804 // Seal the import section, since other index spaces such as function and 805 // global are effected by the number of imports. 806 out.importSec->seal(); 807 808 for (InputFunction *func : symtab->syntheticFunctions) 809 out.functionSec->addFunction(func); 810 811 for (ObjFile *file : symtab->objectFiles) { 812 LLVM_DEBUG(dbgs() << "Functions: " << file->getName() << "\n"); 813 for (InputFunction *func : file->functions) 814 out.functionSec->addFunction(func); 815 } 816 817 for (InputGlobal *global : symtab->syntheticGlobals) 818 out.globalSec->addGlobal(global); 819 820 for (ObjFile *file : symtab->objectFiles) { 821 LLVM_DEBUG(dbgs() << "Globals: " << file->getName() << "\n"); 822 for (InputGlobal *global : file->globals) 823 out.globalSec->addGlobal(global); 824 } 825 826 for (ObjFile *file : symtab->objectFiles) { 827 LLVM_DEBUG(dbgs() << "Tags: " << file->getName() << "\n"); 828 for (InputTag *tag : file->tags) 829 out.tagSec->addTag(tag); 830 } 831 832 for (ObjFile *file : symtab->objectFiles) { 833 LLVM_DEBUG(dbgs() << "Tables: " << file->getName() << "\n"); 834 for (InputTable *table : file->tables) 835 out.tableSec->addTable(table); 836 } 837 838 for (InputTable *table : symtab->syntheticTables) 839 out.tableSec->addTable(table); 840 841 out.globalSec->assignIndexes(); 842 out.tableSec->assignIndexes(); 843 } 844 845 static StringRef getOutputDataSegmentName(const InputChunk &seg) { 846 // We always merge .tbss and .tdata into a single TLS segment so all TLS 847 // symbols are be relative to single __tls_base. 848 if (seg.isTLS()) 849 return ".tdata"; 850 StringRef name = seg.getName(); 851 if (!config->mergeDataSegments) 852 return name; 853 if (name.startswith(".text.")) 854 return ".text"; 855 if (name.startswith(".data.")) 856 return ".data"; 857 if (name.startswith(".bss.")) 858 return ".bss"; 859 if (name.startswith(".rodata.")) 860 return ".rodata"; 861 return name; 862 } 863 864 OutputSegment *Writer::createOutputSegment(StringRef name) { 865 LLVM_DEBUG(dbgs() << "new segment: " << name << "\n"); 866 OutputSegment *s = make<OutputSegment>(name); 867 if (config->sharedMemory) 868 s->initFlags = WASM_DATA_SEGMENT_IS_PASSIVE; 869 if (!config->relocatable && name.startswith(".bss")) 870 s->isBss = true; 871 segments.push_back(s); 872 return s; 873 } 874 875 void Writer::createOutputSegments() { 876 for (ObjFile *file : symtab->objectFiles) { 877 for (InputChunk *segment : file->segments) { 878 if (!segment->live) 879 continue; 880 StringRef name = getOutputDataSegmentName(*segment); 881 OutputSegment *s = nullptr; 882 // When running in relocatable mode we can't merge segments that are part 883 // of comdat groups since the ultimate linker needs to be able exclude or 884 // include them individually. 885 if (config->relocatable && !segment->getComdatName().empty()) { 886 s = createOutputSegment(name); 887 } else { 888 if (segmentMap.count(name) == 0) 889 segmentMap[name] = createOutputSegment(name); 890 s = segmentMap[name]; 891 } 892 s->addInputSegment(segment); 893 } 894 } 895 896 // Sort segments by type, placing .bss last 897 std::stable_sort(segments.begin(), segments.end(), 898 [](const OutputSegment *a, const OutputSegment *b) { 899 auto order = [](StringRef name) { 900 return StringSwitch<int>(name) 901 .StartsWith(".tdata", 0) 902 .StartsWith(".rodata", 1) 903 .StartsWith(".data", 2) 904 .StartsWith(".bss", 4) 905 .Default(3); 906 }; 907 return order(a->name) < order(b->name); 908 }); 909 910 for (size_t i = 0; i < segments.size(); ++i) 911 segments[i]->index = i; 912 913 // Merge MergeInputSections into a single MergeSyntheticSection. 914 LLVM_DEBUG(dbgs() << "-- finalize input semgments\n"); 915 for (OutputSegment *seg : segments) 916 seg->finalizeInputSegments(); 917 } 918 919 void Writer::combineOutputSegments() { 920 // With PIC code we currently only support a single active data segment since 921 // we only have a single __memory_base to use as our base address. This pass 922 // combines all data segments into a single .data segment. 923 // This restructions can be relaxed once we have extended constant 924 // expressions available: 925 // https://github.com/WebAssembly/extended-const 926 assert(config->isPic && !config->sharedMemory); 927 if (segments.size() <= 1) 928 return; 929 OutputSegment *combined = make<OutputSegment>(".data"); 930 combined->startVA = segments[0]->startVA; 931 for (OutputSegment *s : segments) { 932 bool first = true; 933 for (InputChunk *inSeg : s->inputSegments) { 934 if (first) 935 inSeg->alignment = std::max(inSeg->alignment, s->alignment); 936 first = false; 937 #ifndef NDEBUG 938 uint64_t oldVA = inSeg->getVA(); 939 #endif 940 combined->addInputSegment(inSeg); 941 #ifndef NDEBUG 942 uint64_t newVA = inSeg->getVA(); 943 LLVM_DEBUG(dbgs() << "added input segment. name=" << inSeg->getName() 944 << " oldVA=" << oldVA << " newVA=" << newVA << "\n"); 945 assert(oldVA == newVA); 946 #endif 947 } 948 } 949 950 segments = {combined}; 951 } 952 953 static void createFunction(DefinedFunction *func, StringRef bodyContent) { 954 std::string functionBody; 955 { 956 raw_string_ostream os(functionBody); 957 writeUleb128(os, bodyContent.size(), "function size"); 958 os << bodyContent; 959 } 960 ArrayRef<uint8_t> body = arrayRefFromStringRef(saver().save(functionBody)); 961 cast<SyntheticFunction>(func->function)->setBody(body); 962 } 963 964 bool Writer::needsPassiveInitialization(const OutputSegment *segment) { 965 // TLS segments are initialized separately 966 if (segment->isTLS()) 967 return false; 968 // If bulk memory features is supported then we can perform bss initialization 969 // (via memory.fill) during `__wasm_init_memory`. 970 if (config->importMemory && !segment->requiredInBinary()) 971 return true; 972 return segment->initFlags & WASM_DATA_SEGMENT_IS_PASSIVE; 973 } 974 975 bool Writer::hasPassiveInitializedSegments() { 976 return llvm::any_of(segments, [this](const OutputSegment *s) { 977 return this->needsPassiveInitialization(s); 978 }); 979 } 980 981 void Writer::createSyntheticInitFunctions() { 982 if (config->relocatable) 983 return; 984 985 static WasmSignature nullSignature = {{}, {}}; 986 987 // Passive segments are used to avoid memory being reinitialized on each 988 // thread's instantiation. These passive segments are initialized and 989 // dropped in __wasm_init_memory, which is registered as the start function 990 // We also initialize bss segments (using memory.fill) as part of this 991 // function. 992 if (hasPassiveInitializedSegments()) { 993 WasmSym::initMemory = symtab->addSyntheticFunction( 994 "__wasm_init_memory", WASM_SYMBOL_VISIBILITY_HIDDEN, 995 make<SyntheticFunction>(nullSignature, "__wasm_init_memory")); 996 WasmSym::initMemory->markLive(); 997 } 998 999 if (config->sharedMemory && out.globalSec->needsTLSRelocations()) { 1000 WasmSym::applyGlobalTLSRelocs = symtab->addSyntheticFunction( 1001 "__wasm_apply_global_tls_relocs", WASM_SYMBOL_VISIBILITY_HIDDEN, 1002 make<SyntheticFunction>(nullSignature, 1003 "__wasm_apply_global_tls_relocs")); 1004 WasmSym::applyGlobalTLSRelocs->markLive(); 1005 } 1006 1007 if (config->isPic) { 1008 // For PIC code we create synthetic functions that apply relocations. 1009 // These get called from __wasm_call_ctors before the user-level 1010 // constructors. 1011 WasmSym::applyDataRelocs = symtab->addSyntheticFunction( 1012 "__wasm_apply_data_relocs", WASM_SYMBOL_VISIBILITY_HIDDEN, 1013 make<SyntheticFunction>(nullSignature, "__wasm_apply_data_relocs")); 1014 WasmSym::applyDataRelocs->markLive(); 1015 1016 if (out.globalSec->needsRelocations()) { 1017 WasmSym::applyGlobalRelocs = symtab->addSyntheticFunction( 1018 "__wasm_apply_global_relocs", WASM_SYMBOL_VISIBILITY_HIDDEN, 1019 make<SyntheticFunction>(nullSignature, "__wasm_apply_global_relocs")); 1020 WasmSym::applyGlobalRelocs->markLive(); 1021 } 1022 } 1023 1024 int startCount = 0; 1025 if (WasmSym::applyGlobalRelocs) 1026 startCount++; 1027 if (WasmSym::WasmSym::initMemory || WasmSym::applyDataRelocs) 1028 startCount++; 1029 1030 // If there is only one start function we can just use that function 1031 // itself as the Wasm start function, otherwise we need to synthesize 1032 // a new function to call them in sequence. 1033 if (startCount > 1) { 1034 WasmSym::startFunction = symtab->addSyntheticFunction( 1035 "__wasm_start", WASM_SYMBOL_VISIBILITY_HIDDEN, 1036 make<SyntheticFunction>(nullSignature, "__wasm_start")); 1037 WasmSym::startFunction->markLive(); 1038 } 1039 } 1040 1041 void Writer::createInitMemoryFunction() { 1042 LLVM_DEBUG(dbgs() << "createInitMemoryFunction\n"); 1043 assert(WasmSym::initMemory); 1044 assert(hasPassiveInitializedSegments()); 1045 uint64_t flagAddress; 1046 if (config->sharedMemory) { 1047 assert(WasmSym::initMemoryFlag); 1048 flagAddress = WasmSym::initMemoryFlag->getVA(); 1049 } 1050 bool is64 = config->is64.getValueOr(false); 1051 std::string bodyContent; 1052 { 1053 raw_string_ostream os(bodyContent); 1054 // Initialize memory in a thread-safe manner. The thread that successfully 1055 // increments the flag from 0 to 1 is is responsible for performing the 1056 // memory initialization. Other threads go sleep on the flag until the 1057 // first thread finishing initializing memory, increments the flag to 2, 1058 // and wakes all the other threads. Once the flag has been set to 2, 1059 // subsequently started threads will skip the sleep. All threads 1060 // unconditionally drop their passive data segments once memory has been 1061 // initialized. The generated code is as follows: 1062 // 1063 // (func $__wasm_init_memory 1064 // (block $drop 1065 // (block $wait 1066 // (block $init 1067 // (br_table $init $wait $drop 1068 // (i32.atomic.rmw.cmpxchg align=2 offset=0 1069 // (i32.const $__init_memory_flag) 1070 // (i32.const 0) 1071 // (i32.const 1) 1072 // ) 1073 // ) 1074 // ) ;; $init 1075 // ( ... initialize data segments ... ) 1076 // (i32.atomic.store align=2 offset=0 1077 // (i32.const $__init_memory_flag) 1078 // (i32.const 2) 1079 // ) 1080 // (drop 1081 // (i32.atomic.notify align=2 offset=0 1082 // (i32.const $__init_memory_flag) 1083 // (i32.const -1u) 1084 // ) 1085 // ) 1086 // (br $drop) 1087 // ) ;; $wait 1088 // (drop 1089 // (i32.atomic.wait align=2 offset=0 1090 // (i32.const $__init_memory_flag) 1091 // (i32.const 1) 1092 // (i32.const -1) 1093 // ) 1094 // ) 1095 // ) ;; $drop 1096 // ( ... drop data segments ... ) 1097 // ) 1098 // 1099 // When we are building with PIC, calculate the flag location using: 1100 // 1101 // (global.get $__memory_base) 1102 // (i32.const $__init_memory_flag) 1103 // (i32.const 1) 1104 1105 auto writeGetFlagAddress = [&]() { 1106 if (config->isPic) { 1107 writeU8(os, WASM_OPCODE_LOCAL_GET, "local.get"); 1108 writeUleb128(os, 0, "local 0"); 1109 } else { 1110 writePtrConst(os, flagAddress, is64, "flag address"); 1111 } 1112 }; 1113 1114 if (config->sharedMemory) { 1115 // With PIC code we cache the flag address in local 0 1116 if (config->isPic) { 1117 writeUleb128(os, 1, "num local decls"); 1118 writeUleb128(os, 1, "local count"); 1119 writeU8(os, is64 ? WASM_TYPE_I64 : WASM_TYPE_I32, "address type"); 1120 writeU8(os, WASM_OPCODE_GLOBAL_GET, "GLOBAL_GET"); 1121 writeUleb128(os, WasmSym::memoryBase->getGlobalIndex(), "memory_base"); 1122 writePtrConst(os, flagAddress, is64, "flag address"); 1123 writeU8(os, is64 ? WASM_OPCODE_I64_ADD : WASM_OPCODE_I32_ADD, "add"); 1124 writeU8(os, WASM_OPCODE_LOCAL_SET, "local.set"); 1125 writeUleb128(os, 0, "local 0"); 1126 } else { 1127 writeUleb128(os, 0, "num locals"); 1128 } 1129 1130 // Set up destination blocks 1131 writeU8(os, WASM_OPCODE_BLOCK, "block $drop"); 1132 writeU8(os, WASM_TYPE_NORESULT, "block type"); 1133 writeU8(os, WASM_OPCODE_BLOCK, "block $wait"); 1134 writeU8(os, WASM_TYPE_NORESULT, "block type"); 1135 writeU8(os, WASM_OPCODE_BLOCK, "block $init"); 1136 writeU8(os, WASM_TYPE_NORESULT, "block type"); 1137 1138 // Atomically check whether we win the race. 1139 writeGetFlagAddress(); 1140 writeI32Const(os, 0, "expected flag value"); 1141 writeI32Const(os, 1, "new flag value"); 1142 writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix"); 1143 writeUleb128(os, WASM_OPCODE_I32_RMW_CMPXCHG, "i32.atomic.rmw.cmpxchg"); 1144 writeMemArg(os, 2, 0); 1145 1146 // Based on the value, decide what to do next. 1147 writeU8(os, WASM_OPCODE_BR_TABLE, "br_table"); 1148 writeUleb128(os, 2, "label vector length"); 1149 writeUleb128(os, 0, "label $init"); 1150 writeUleb128(os, 1, "label $wait"); 1151 writeUleb128(os, 2, "default label $drop"); 1152 1153 // Initialize passive data segments 1154 writeU8(os, WASM_OPCODE_END, "end $init"); 1155 } else { 1156 writeUleb128(os, 0, "num local decls"); 1157 } 1158 1159 for (const OutputSegment *s : segments) { 1160 if (needsPassiveInitialization(s)) { 1161 // For passive BSS segments we can simple issue a memory.fill(0). 1162 // For non-BSS segments we do a memory.init. Both these 1163 // instructions take as thier first argument the destination 1164 // address. 1165 writePtrConst(os, s->startVA, is64, "destination address"); 1166 if (config->isPic) { 1167 writeU8(os, WASM_OPCODE_GLOBAL_GET, "GLOBAL_GET"); 1168 writeUleb128(os, WasmSym::memoryBase->getGlobalIndex(), 1169 "memory_base"); 1170 writeU8(os, is64 ? WASM_OPCODE_I64_ADD : WASM_OPCODE_I32_ADD, 1171 "i32.add"); 1172 } 1173 if (s->isBss) { 1174 writeI32Const(os, 0, "fill value"); 1175 writeI32Const(os, s->size, "memory region size"); 1176 writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix"); 1177 writeUleb128(os, WASM_OPCODE_MEMORY_FILL, "memory.fill"); 1178 writeU8(os, 0, "memory index immediate"); 1179 } else { 1180 writeI32Const(os, 0, "source segment offset"); 1181 writeI32Const(os, s->size, "memory region size"); 1182 writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix"); 1183 writeUleb128(os, WASM_OPCODE_MEMORY_INIT, "memory.init"); 1184 writeUleb128(os, s->index, "segment index immediate"); 1185 writeU8(os, 0, "memory index immediate"); 1186 } 1187 } 1188 } 1189 1190 // Memory init is now complete. Apply data relocation if there 1191 // are any. 1192 if (WasmSym::applyDataRelocs) { 1193 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1194 writeUleb128(os, WasmSym::applyDataRelocs->getFunctionIndex(), 1195 "function index"); 1196 } 1197 1198 if (config->sharedMemory) { 1199 // Set flag to 2 to mark end of initialization 1200 writeGetFlagAddress(); 1201 writeI32Const(os, 2, "flag value"); 1202 writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix"); 1203 writeUleb128(os, WASM_OPCODE_I32_ATOMIC_STORE, "i32.atomic.store"); 1204 writeMemArg(os, 2, 0); 1205 1206 // Notify any waiters that memory initialization is complete 1207 writeGetFlagAddress(); 1208 writeI32Const(os, -1, "number of waiters"); 1209 writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix"); 1210 writeUleb128(os, WASM_OPCODE_ATOMIC_NOTIFY, "atomic.notify"); 1211 writeMemArg(os, 2, 0); 1212 writeU8(os, WASM_OPCODE_DROP, "drop"); 1213 1214 // Branch to drop the segments 1215 writeU8(os, WASM_OPCODE_BR, "br"); 1216 writeUleb128(os, 1, "label $drop"); 1217 1218 // Wait for the winning thread to initialize memory 1219 writeU8(os, WASM_OPCODE_END, "end $wait"); 1220 writeGetFlagAddress(); 1221 writeI32Const(os, 1, "expected flag value"); 1222 writeI64Const(os, -1, "timeout"); 1223 1224 writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix"); 1225 writeUleb128(os, WASM_OPCODE_I32_ATOMIC_WAIT, "i32.atomic.wait"); 1226 writeMemArg(os, 2, 0); 1227 writeU8(os, WASM_OPCODE_DROP, "drop"); 1228 1229 // Unconditionally drop passive data segments 1230 writeU8(os, WASM_OPCODE_END, "end $drop"); 1231 } 1232 1233 for (const OutputSegment *s : segments) { 1234 if (needsPassiveInitialization(s) && !s->isBss) { 1235 // data.drop instruction 1236 writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix"); 1237 writeUleb128(os, WASM_OPCODE_DATA_DROP, "data.drop"); 1238 writeUleb128(os, s->index, "segment index immediate"); 1239 } 1240 } 1241 1242 // End the function 1243 writeU8(os, WASM_OPCODE_END, "END"); 1244 } 1245 1246 createFunction(WasmSym::initMemory, bodyContent); 1247 } 1248 1249 void Writer::createStartFunction() { 1250 // If the start function exists when we have more than one function to call. 1251 if (WasmSym::startFunction) { 1252 std::string bodyContent; 1253 { 1254 raw_string_ostream os(bodyContent); 1255 writeUleb128(os, 0, "num locals"); 1256 if (WasmSym::initMemory) { 1257 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1258 writeUleb128(os, WasmSym::initMemory->getFunctionIndex(), 1259 "function index"); 1260 } else if (WasmSym::applyDataRelocs) { 1261 // When initMemory is present it calls applyDataRelocs. If not, 1262 // we must call it directly. 1263 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1264 writeUleb128(os, WasmSym::applyDataRelocs->getFunctionIndex(), 1265 "function index"); 1266 } 1267 if (WasmSym::applyGlobalRelocs) { 1268 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1269 writeUleb128(os, WasmSym::applyGlobalRelocs->getFunctionIndex(), 1270 "function index"); 1271 } 1272 writeU8(os, WASM_OPCODE_END, "END"); 1273 } 1274 createFunction(WasmSym::startFunction, bodyContent); 1275 } else if (WasmSym::initMemory) { 1276 WasmSym::startFunction = WasmSym::initMemory; 1277 } else if (WasmSym::applyGlobalRelocs) { 1278 WasmSym::startFunction = WasmSym::applyGlobalRelocs; 1279 } else if (WasmSym::applyDataRelocs) { 1280 WasmSym::startFunction = WasmSym::applyDataRelocs; 1281 } 1282 } 1283 1284 // For -shared (PIC) output, we create create a synthetic function which will 1285 // apply any relocations to the data segments on startup. This function is 1286 // called `__wasm_apply_data_relocs` and is added at the beginning of 1287 // `__wasm_call_ctors` before any of the constructors run. 1288 void Writer::createApplyDataRelocationsFunction() { 1289 LLVM_DEBUG(dbgs() << "createApplyDataRelocationsFunction\n"); 1290 // First write the body's contents to a string. 1291 std::string bodyContent; 1292 { 1293 raw_string_ostream os(bodyContent); 1294 writeUleb128(os, 0, "num locals"); 1295 for (const OutputSegment *seg : segments) 1296 for (const InputChunk *inSeg : seg->inputSegments) 1297 inSeg->generateRelocationCode(os); 1298 1299 writeU8(os, WASM_OPCODE_END, "END"); 1300 } 1301 1302 createFunction(WasmSym::applyDataRelocs, bodyContent); 1303 } 1304 1305 // Similar to createApplyDataRelocationsFunction but generates relocation code 1306 // for WebAssembly globals. Because these globals are not shared between threads 1307 // these relocation need to run on every thread. 1308 void Writer::createApplyGlobalRelocationsFunction() { 1309 // First write the body's contents to a string. 1310 std::string bodyContent; 1311 { 1312 raw_string_ostream os(bodyContent); 1313 writeUleb128(os, 0, "num locals"); 1314 out.globalSec->generateRelocationCode(os, false); 1315 writeU8(os, WASM_OPCODE_END, "END"); 1316 } 1317 1318 createFunction(WasmSym::applyGlobalRelocs, bodyContent); 1319 } 1320 1321 // Similar to createApplyGlobalRelocationsFunction but for 1322 // TLS symbols. This cannot be run during the start function 1323 // but must be delayed until __wasm_init_tls is called. 1324 void Writer::createApplyGlobalTLSRelocationsFunction() { 1325 // First write the body's contents to a string. 1326 std::string bodyContent; 1327 { 1328 raw_string_ostream os(bodyContent); 1329 writeUleb128(os, 0, "num locals"); 1330 out.globalSec->generateRelocationCode(os, true); 1331 writeU8(os, WASM_OPCODE_END, "END"); 1332 } 1333 1334 createFunction(WasmSym::applyGlobalTLSRelocs, bodyContent); 1335 } 1336 1337 // Create synthetic "__wasm_call_ctors" function based on ctor functions 1338 // in input object. 1339 void Writer::createCallCtorsFunction() { 1340 // If __wasm_call_ctors isn't referenced, there aren't any ctors, and we 1341 // aren't calling `__wasm_apply_data_relocs` for Emscripten-style PIC, don't 1342 // define the `__wasm_call_ctors` function. 1343 if (!WasmSym::callCtors->isLive() && initFunctions.empty()) 1344 return; 1345 1346 // First write the body's contents to a string. 1347 std::string bodyContent; 1348 { 1349 raw_string_ostream os(bodyContent); 1350 writeUleb128(os, 0, "num locals"); 1351 1352 if (WasmSym::applyDataRelocs && !WasmSym::initMemory) { 1353 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1354 writeUleb128(os, WasmSym::applyDataRelocs->getFunctionIndex(), 1355 "function index"); 1356 } 1357 1358 // Call constructors 1359 for (const WasmInitEntry &f : initFunctions) { 1360 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1361 writeUleb128(os, f.sym->getFunctionIndex(), "function index"); 1362 for (size_t i = 0; i < f.sym->signature->Returns.size(); i++) { 1363 writeU8(os, WASM_OPCODE_DROP, "DROP"); 1364 } 1365 } 1366 1367 writeU8(os, WASM_OPCODE_END, "END"); 1368 } 1369 1370 createFunction(WasmSym::callCtors, bodyContent); 1371 } 1372 1373 // Create a wrapper around a function export which calls the 1374 // static constructors and destructors. 1375 void Writer::createCommandExportWrapper(uint32_t functionIndex, 1376 DefinedFunction *f) { 1377 // First write the body's contents to a string. 1378 std::string bodyContent; 1379 { 1380 raw_string_ostream os(bodyContent); 1381 writeUleb128(os, 0, "num locals"); 1382 1383 // Call `__wasm_call_ctors` which call static constructors (and 1384 // applies any runtime relocations in Emscripten-style PIC mode) 1385 if (WasmSym::callCtors->isLive()) { 1386 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1387 writeUleb128(os, WasmSym::callCtors->getFunctionIndex(), 1388 "function index"); 1389 } 1390 1391 // Call the user's code, leaving any return values on the operand stack. 1392 for (size_t i = 0; i < f->signature->Params.size(); ++i) { 1393 writeU8(os, WASM_OPCODE_LOCAL_GET, "local.get"); 1394 writeUleb128(os, i, "local index"); 1395 } 1396 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1397 writeUleb128(os, functionIndex, "function index"); 1398 1399 // Call the function that calls the destructors. 1400 if (DefinedFunction *callDtors = WasmSym::callDtors) { 1401 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1402 writeUleb128(os, callDtors->getFunctionIndex(), "function index"); 1403 } 1404 1405 // End the function, returning the return values from the user's code. 1406 writeU8(os, WASM_OPCODE_END, "END"); 1407 } 1408 1409 createFunction(f, bodyContent); 1410 } 1411 1412 void Writer::createInitTLSFunction() { 1413 std::string bodyContent; 1414 { 1415 raw_string_ostream os(bodyContent); 1416 1417 OutputSegment *tlsSeg = nullptr; 1418 for (auto *seg : segments) { 1419 if (seg->name == ".tdata") { 1420 tlsSeg = seg; 1421 break; 1422 } 1423 } 1424 1425 writeUleb128(os, 0, "num locals"); 1426 if (tlsSeg) { 1427 writeU8(os, WASM_OPCODE_LOCAL_GET, "local.get"); 1428 writeUleb128(os, 0, "local index"); 1429 1430 writeU8(os, WASM_OPCODE_GLOBAL_SET, "global.set"); 1431 writeUleb128(os, WasmSym::tlsBase->getGlobalIndex(), "global index"); 1432 1433 // FIXME(wvo): this local needs to be I64 in wasm64, or we need an extend op. 1434 writeU8(os, WASM_OPCODE_LOCAL_GET, "local.get"); 1435 writeUleb128(os, 0, "local index"); 1436 1437 writeI32Const(os, 0, "segment offset"); 1438 1439 writeI32Const(os, tlsSeg->size, "memory region size"); 1440 1441 writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix"); 1442 writeUleb128(os, WASM_OPCODE_MEMORY_INIT, "MEMORY.INIT"); 1443 writeUleb128(os, tlsSeg->index, "segment index immediate"); 1444 writeU8(os, 0, "memory index immediate"); 1445 } 1446 1447 if (WasmSym::applyGlobalTLSRelocs) { 1448 writeU8(os, WASM_OPCODE_CALL, "CALL"); 1449 writeUleb128(os, WasmSym::applyGlobalTLSRelocs->getFunctionIndex(), 1450 "function index"); 1451 } 1452 writeU8(os, WASM_OPCODE_END, "end function"); 1453 } 1454 1455 createFunction(WasmSym::initTLS, bodyContent); 1456 } 1457 1458 // Populate InitFunctions vector with init functions from all input objects. 1459 // This is then used either when creating the output linking section or to 1460 // synthesize the "__wasm_call_ctors" function. 1461 void Writer::calculateInitFunctions() { 1462 if (!config->relocatable && !WasmSym::callCtors->isLive()) 1463 return; 1464 1465 for (ObjFile *file : symtab->objectFiles) { 1466 const WasmLinkingData &l = file->getWasmObj()->linkingData(); 1467 for (const WasmInitFunc &f : l.InitFunctions) { 1468 FunctionSymbol *sym = file->getFunctionSymbol(f.Symbol); 1469 // comdat exclusions can cause init functions be discarded. 1470 if (sym->isDiscarded() || !sym->isLive()) 1471 continue; 1472 if (sym->signature->Params.size() != 0) 1473 error("constructor functions cannot take arguments: " + toString(*sym)); 1474 LLVM_DEBUG(dbgs() << "initFunctions: " << toString(*sym) << "\n"); 1475 initFunctions.emplace_back(WasmInitEntry{sym, f.Priority}); 1476 } 1477 } 1478 1479 // Sort in order of priority (lowest first) so that they are called 1480 // in the correct order. 1481 llvm::stable_sort(initFunctions, 1482 [](const WasmInitEntry &l, const WasmInitEntry &r) { 1483 return l.priority < r.priority; 1484 }); 1485 } 1486 1487 void Writer::createSyntheticSections() { 1488 out.dylinkSec = make<DylinkSection>(); 1489 out.typeSec = make<TypeSection>(); 1490 out.importSec = make<ImportSection>(); 1491 out.functionSec = make<FunctionSection>(); 1492 out.tableSec = make<TableSection>(); 1493 out.memorySec = make<MemorySection>(); 1494 out.tagSec = make<TagSection>(); 1495 out.globalSec = make<GlobalSection>(); 1496 out.exportSec = make<ExportSection>(); 1497 out.startSec = make<StartSection>(); 1498 out.elemSec = make<ElemSection>(); 1499 out.producersSec = make<ProducersSection>(); 1500 out.targetFeaturesSec = make<TargetFeaturesSection>(); 1501 } 1502 1503 void Writer::createSyntheticSectionsPostLayout() { 1504 out.dataCountSec = make<DataCountSection>(segments); 1505 out.linkingSec = make<LinkingSection>(initFunctions, segments); 1506 out.nameSec = make<NameSection>(segments); 1507 } 1508 1509 void Writer::run() { 1510 if (config->relocatable || config->isPic) 1511 config->globalBase = 0; 1512 1513 // For PIC code the table base is assigned dynamically by the loader. 1514 // For non-PIC, we start at 1 so that accessing table index 0 always traps. 1515 if (!config->isPic) { 1516 config->tableBase = 1; 1517 if (WasmSym::definedTableBase) 1518 WasmSym::definedTableBase->setVA(config->tableBase); 1519 if (WasmSym::definedTableBase32) 1520 WasmSym::definedTableBase32->setVA(config->tableBase); 1521 } 1522 1523 log("-- createOutputSegments"); 1524 createOutputSegments(); 1525 log("-- createSyntheticSections"); 1526 createSyntheticSections(); 1527 log("-- layoutMemory"); 1528 layoutMemory(); 1529 1530 if (!config->relocatable) { 1531 // Create linker synthesized __start_SECNAME/__stop_SECNAME symbols 1532 // This has to be done after memory layout is performed. 1533 for (const OutputSegment *seg : segments) { 1534 addStartStopSymbols(seg); 1535 } 1536 } 1537 1538 for (auto &pair : config->exportedSymbols) { 1539 Symbol *sym = symtab->find(pair.first()); 1540 if (sym && sym->isDefined()) 1541 sym->forceExport = true; 1542 } 1543 1544 // Delay reporting error about explicit exports until after 1545 // addStartStopSymbols which can create optional symbols. 1546 for (auto &name : config->requiredExports) { 1547 Symbol *sym = symtab->find(name); 1548 if (!sym || !sym->isDefined()) { 1549 if (config->unresolvedSymbols == UnresolvedPolicy::ReportError) 1550 error(Twine("symbol exported via --export not found: ") + name); 1551 if (config->unresolvedSymbols == UnresolvedPolicy::Warn) 1552 warn(Twine("symbol exported via --export not found: ") + name); 1553 } 1554 } 1555 1556 if (config->isPic && !config->sharedMemory) { 1557 // In shared memory mode all data segments are passive and initialized 1558 // via __wasm_init_memory. 1559 log("-- combineOutputSegments"); 1560 combineOutputSegments(); 1561 } 1562 1563 log("-- createSyntheticSectionsPostLayout"); 1564 createSyntheticSectionsPostLayout(); 1565 log("-- populateProducers"); 1566 populateProducers(); 1567 log("-- calculateImports"); 1568 calculateImports(); 1569 log("-- scanRelocations"); 1570 scanRelocations(); 1571 log("-- finalizeIndirectFunctionTable"); 1572 finalizeIndirectFunctionTable(); 1573 log("-- populateTargetFeatures"); 1574 populateTargetFeatures(); 1575 log("-- createSyntheticInitFunctions"); 1576 createSyntheticInitFunctions(); 1577 log("-- assignIndexes"); 1578 assignIndexes(); 1579 log("-- calculateInitFunctions"); 1580 calculateInitFunctions(); 1581 1582 if (!config->relocatable) { 1583 // Create linker synthesized functions 1584 if (WasmSym::applyDataRelocs) 1585 createApplyDataRelocationsFunction(); 1586 if (WasmSym::applyGlobalRelocs) 1587 createApplyGlobalRelocationsFunction(); 1588 if (WasmSym::applyGlobalTLSRelocs) 1589 createApplyGlobalTLSRelocationsFunction(); 1590 if (WasmSym::initMemory) 1591 createInitMemoryFunction(); 1592 createStartFunction(); 1593 1594 createCallCtorsFunction(); 1595 1596 // Create export wrappers for commands if needed. 1597 // 1598 // If the input contains a call to `__wasm_call_ctors`, either in one of 1599 // the input objects or an explicit export from the command-line, we 1600 // assume ctors and dtors are taken care of already. 1601 if (!config->relocatable && !config->isPic && 1602 !WasmSym::callCtors->isUsedInRegularObj && 1603 !WasmSym::callCtors->isExported()) { 1604 log("-- createCommandExportWrappers"); 1605 createCommandExportWrappers(); 1606 } 1607 } 1608 1609 if (WasmSym::initTLS && WasmSym::initTLS->isLive()) 1610 createInitTLSFunction(); 1611 1612 if (errorCount()) 1613 return; 1614 1615 log("-- calculateTypes"); 1616 calculateTypes(); 1617 log("-- calculateExports"); 1618 calculateExports(); 1619 log("-- calculateCustomSections"); 1620 calculateCustomSections(); 1621 log("-- populateSymtab"); 1622 populateSymtab(); 1623 log("-- checkImportExportTargetFeatures"); 1624 checkImportExportTargetFeatures(); 1625 log("-- addSections"); 1626 addSections(); 1627 1628 if (errorHandler().verbose) { 1629 log("Defined Functions: " + Twine(out.functionSec->inputFunctions.size())); 1630 log("Defined Globals : " + Twine(out.globalSec->numGlobals())); 1631 log("Defined Tags : " + Twine(out.tagSec->inputTags.size())); 1632 log("Defined Tables : " + Twine(out.tableSec->inputTables.size())); 1633 log("Function Imports : " + 1634 Twine(out.importSec->getNumImportedFunctions())); 1635 log("Global Imports : " + Twine(out.importSec->getNumImportedGlobals())); 1636 log("Tag Imports : " + Twine(out.importSec->getNumImportedTags())); 1637 log("Table Imports : " + Twine(out.importSec->getNumImportedTables())); 1638 } 1639 1640 createHeader(); 1641 log("-- finalizeSections"); 1642 finalizeSections(); 1643 1644 log("-- writeMapFile"); 1645 writeMapFile(outputSections); 1646 1647 log("-- openFile"); 1648 openFile(); 1649 if (errorCount()) 1650 return; 1651 1652 writeHeader(); 1653 1654 log("-- writeSections"); 1655 writeSections(); 1656 if (errorCount()) 1657 return; 1658 1659 if (Error e = buffer->commit()) 1660 fatal("failed to write the output file: " + toString(std::move(e))); 1661 } 1662 1663 // Open a result file. 1664 void Writer::openFile() { 1665 log("writing: " + config->outputFile); 1666 1667 Expected<std::unique_ptr<FileOutputBuffer>> bufferOrErr = 1668 FileOutputBuffer::create(config->outputFile, fileSize, 1669 FileOutputBuffer::F_executable); 1670 1671 if (!bufferOrErr) 1672 error("failed to open " + config->outputFile + ": " + 1673 toString(bufferOrErr.takeError())); 1674 else 1675 buffer = std::move(*bufferOrErr); 1676 } 1677 1678 void Writer::createHeader() { 1679 raw_string_ostream os(header); 1680 writeBytes(os, WasmMagic, sizeof(WasmMagic), "wasm magic"); 1681 writeU32(os, WasmVersion, "wasm version"); 1682 os.flush(); 1683 fileSize += header.size(); 1684 } 1685 1686 void writeResult() { Writer().run(); } 1687 1688 } // namespace wasm 1689 } // namespace lld 1690