xref: /llvm-project-15.0.7/lld/wasm/Writer.cpp (revision c98ec609)
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 "InputEvent.h"
13 #include "InputGlobal.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/ErrorHandler.h"
21 #include "lld/Common/Memory.h"
22 #include "lld/Common/Strings.h"
23 #include "lld/Common/Threads.h"
24 #include "llvm/ADT/DenseSet.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/StringMap.h"
28 #include "llvm/BinaryFormat/Wasm.h"
29 #include "llvm/Object/WasmTraits.h"
30 #include "llvm/Support/FileOutputBuffer.h"
31 #include "llvm/Support/Format.h"
32 #include "llvm/Support/FormatVariadic.h"
33 #include "llvm/Support/LEB128.h"
34 
35 #include <cstdarg>
36 #include <map>
37 
38 #define DEBUG_TYPE "lld"
39 
40 using namespace llvm;
41 using namespace llvm::wasm;
42 using namespace lld;
43 using namespace lld::wasm;
44 
45 static constexpr int stackAlignment = 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   void createInitMemoryFunction();
58   void createApplyRelocationsFunction();
59   void createCallCtorsFunction();
60   void createInitTLSFunction();
61 
62   void assignIndexes();
63   void populateSymtab();
64   void populateProducers();
65   void populateTargetFeatures();
66   void calculateInitFunctions();
67   void calculateImports();
68   void calculateExports();
69   void calculateCustomSections();
70   void calculateTypes();
71   void createOutputSegments();
72   void layoutMemory();
73   void createHeader();
74 
75   void addSection(OutputSection *sec);
76 
77   void addSections();
78 
79   void createCustomSections();
80   void createSyntheticSections();
81   void finalizeSections();
82 
83   // Custom sections
84   void createRelocSections();
85 
86   void writeHeader();
87   void writeSections();
88 
89   uint64_t fileSize = 0;
90 
91   std::vector<WasmInitEntry> initFunctions;
92   llvm::StringMap<std::vector<InputSection *>> customSectionMapping;
93 
94   // Elements that are used to construct the final output
95   std::string header;
96   std::vector<OutputSection *> outputSections;
97 
98   std::unique_ptr<FileOutputBuffer> buffer;
99 
100   std::vector<OutputSegment *> segments;
101   llvm::SmallDenseMap<StringRef, OutputSegment *> segmentMap;
102 };
103 
104 } // anonymous namespace
105 
106 void Writer::calculateCustomSections() {
107   log("calculateCustomSections");
108   bool stripDebug = config->stripDebug || config->stripAll;
109   for (ObjFile *file : symtab->objectFiles) {
110     for (InputSection *section : file->customSections) {
111       StringRef name = section->getName();
112       // These custom sections are known the linker and synthesized rather than
113       // blindly copied
114       if (name == "linking" || name == "name" || name == "producers" ||
115           name == "target_features" || name.startswith("reloc."))
116         continue;
117       // .. or it is a debug section
118       if (stripDebug && name.startswith(".debug_"))
119         continue;
120       customSectionMapping[name].push_back(section);
121     }
122   }
123 }
124 
125 void Writer::createCustomSections() {
126   log("createCustomSections");
127   for (auto &pair : customSectionMapping) {
128     StringRef name = pair.first();
129     LLVM_DEBUG(dbgs() << "createCustomSection: " << name << "\n");
130 
131     OutputSection *sec = make<CustomSection>(name, pair.second);
132     if (config->relocatable || config->emitRelocs) {
133       auto *sym = make<OutputSectionSymbol>(sec);
134       out.linkingSec->addToSymtab(sym);
135       sec->sectionSym = sym;
136     }
137     addSection(sec);
138   }
139 }
140 
141 // Create relocations sections in the final output.
142 // These are only created when relocatable output is requested.
143 void Writer::createRelocSections() {
144   log("createRelocSections");
145   // Don't use iterator here since we are adding to OutputSection
146   size_t origSize = outputSections.size();
147   for (size_t i = 0; i < origSize; i++) {
148     LLVM_DEBUG(dbgs() << "check section " << i << "\n");
149     OutputSection *sec = outputSections[i];
150 
151     // Count the number of needed sections.
152     uint32_t count = sec->getNumRelocations();
153     if (!count)
154       continue;
155 
156     StringRef name;
157     if (sec->type == WASM_SEC_DATA)
158       name = "reloc.DATA";
159     else if (sec->type == WASM_SEC_CODE)
160       name = "reloc.CODE";
161     else if (sec->type == WASM_SEC_CUSTOM)
162       name = saver.save("reloc." + sec->name);
163     else
164       llvm_unreachable(
165           "relocations only supported for code, data, or custom sections");
166 
167     addSection(make<RelocSection>(name, sec));
168   }
169 }
170 
171 void Writer::populateProducers() {
172   for (ObjFile *file : symtab->objectFiles) {
173     const WasmProducerInfo &info = file->getWasmObj()->getProducerInfo();
174     out.producersSec->addInfo(info);
175   }
176 }
177 
178 void Writer::writeHeader() {
179   memcpy(buffer->getBufferStart(), header.data(), header.size());
180 }
181 
182 void Writer::writeSections() {
183   uint8_t *buf = buffer->getBufferStart();
184   parallelForEach(outputSections, [buf](OutputSection *s) {
185     assert(s->isNeeded());
186     s->writeTo(buf);
187   });
188 }
189 
190 // Fix the memory layout of the output binary.  This assigns memory offsets
191 // to each of the input data sections as well as the explicit stack region.
192 // The default memory layout is as follows, from low to high.
193 //
194 //  - initialized data (starting at Config->globalBase)
195 //  - BSS data (not currently implemented in llvm)
196 //  - explicit stack (Config->ZStackSize)
197 //  - heap start / unallocated
198 //
199 // The --stack-first option means that stack is placed before any static data.
200 // This can be useful since it means that stack overflow traps immediately
201 // rather than overwriting global data, but also increases code size since all
202 // static data loads and stores requires larger offsets.
203 void Writer::layoutMemory() {
204   uint32_t memoryPtr = 0;
205 
206   auto placeStack = [&]() {
207     if (config->relocatable || config->isPic)
208       return;
209     memoryPtr = alignTo(memoryPtr, stackAlignment);
210     if (config->zStackSize != alignTo(config->zStackSize, stackAlignment))
211       error("stack size must be " + Twine(stackAlignment) + "-byte aligned");
212     log("mem: stack size  = " + Twine(config->zStackSize));
213     log("mem: stack base  = " + Twine(memoryPtr));
214     memoryPtr += config->zStackSize;
215     auto *sp = cast<DefinedGlobal>(WasmSym::stackPointer);
216     sp->global->global.InitExpr.Value.Int32 = memoryPtr;
217     log("mem: stack top   = " + Twine(memoryPtr));
218   };
219 
220   if (config->stackFirst) {
221     placeStack();
222   } else {
223     memoryPtr = config->globalBase;
224     log("mem: global base = " + Twine(config->globalBase));
225   }
226 
227   if (WasmSym::globalBase)
228     WasmSym::globalBase->setVirtualAddress(memoryPtr);
229   if (WasmSym::definedMemoryBase)
230     WasmSym::definedMemoryBase->setVirtualAddress(memoryPtr);
231 
232   uint32_t dataStart = memoryPtr;
233 
234   // Arbitrarily set __dso_handle handle to point to the start of the data
235   // segments.
236   if (WasmSym::dsoHandle)
237     WasmSym::dsoHandle->setVirtualAddress(dataStart);
238 
239   out.dylinkSec->memAlign = 0;
240   for (OutputSegment *seg : segments) {
241     out.dylinkSec->memAlign = std::max(out.dylinkSec->memAlign, seg->alignment);
242     memoryPtr = alignTo(memoryPtr, 1ULL << seg->alignment);
243     seg->startVA = memoryPtr;
244     log(formatv("mem: {0,-15} offset={1,-8} size={2,-8} align={3}", seg->name,
245                 memoryPtr, seg->size, seg->alignment));
246     memoryPtr += seg->size;
247 
248     if (WasmSym::tlsSize && seg->name == ".tdata") {
249       auto *tlsSize = cast<DefinedGlobal>(WasmSym::tlsSize);
250       tlsSize->global->global.InitExpr.Value.Int32 = seg->size;
251 
252       auto *tlsAlign = cast<DefinedGlobal>(WasmSym::tlsAlign);
253       tlsAlign->global->global.InitExpr.Value.Int32 = 1U << seg->alignment;
254     }
255   }
256 
257   // Make space for the memory initialization flag
258   if (WasmSym::initMemoryFlag) {
259     memoryPtr = alignTo(memoryPtr, 4);
260     WasmSym::initMemoryFlag->setVirtualAddress(memoryPtr);
261     log(formatv("mem: {0,-15} offset={1,-8} size={2,-8} align={3}",
262                 "__wasm_init_memory_flag", memoryPtr, 4, 4));
263     memoryPtr += 4;
264   }
265 
266   // TODO: Add .bss space here.
267   if (WasmSym::dataEnd)
268     WasmSym::dataEnd->setVirtualAddress(memoryPtr);
269 
270   log("mem: static data = " + Twine(memoryPtr - dataStart));
271 
272   if (config->shared) {
273     out.dylinkSec->memSize = memoryPtr;
274     return;
275   }
276 
277   if (!config->stackFirst)
278     placeStack();
279 
280   // Set `__heap_base` to directly follow the end of the stack or global data.
281   // The fact that this comes last means that a malloc/brk implementation
282   // can grow the heap at runtime.
283   log("mem: heap base   = " + Twine(memoryPtr));
284   if (WasmSym::heapBase)
285     WasmSym::heapBase->setVirtualAddress(memoryPtr);
286 
287   if (config->initialMemory != 0) {
288     if (config->initialMemory != alignTo(config->initialMemory, WasmPageSize))
289       error("initial memory must be " + Twine(WasmPageSize) + "-byte aligned");
290     if (memoryPtr > config->initialMemory)
291       error("initial memory too small, " + Twine(memoryPtr) + " bytes needed");
292     else
293       memoryPtr = config->initialMemory;
294   }
295   out.dylinkSec->memSize = memoryPtr;
296   out.memorySec->numMemoryPages =
297       alignTo(memoryPtr, WasmPageSize) / WasmPageSize;
298   log("mem: total pages = " + Twine(out.memorySec->numMemoryPages));
299 
300   // Check max if explicitly supplied or required by shared memory
301   if (config->maxMemory != 0 || config->sharedMemory) {
302     if (config->maxMemory != alignTo(config->maxMemory, WasmPageSize))
303       error("maximum memory must be " + Twine(WasmPageSize) + "-byte aligned");
304     if (memoryPtr > config->maxMemory)
305       error("maximum memory too small, " + Twine(memoryPtr) + " bytes needed");
306     out.memorySec->maxMemoryPages = config->maxMemory / WasmPageSize;
307     log("mem: max pages   = " + Twine(out.memorySec->maxMemoryPages));
308   }
309 }
310 
311 void Writer::addSection(OutputSection *sec) {
312   if (!sec->isNeeded())
313     return;
314   log("addSection: " + toString(*sec));
315   sec->sectionIndex = outputSections.size();
316   outputSections.push_back(sec);
317 }
318 
319 // If a section name is valid as a C identifier (which is rare because of
320 // the leading '.'), linkers are expected to define __start_<secname> and
321 // __stop_<secname> symbols. They are at beginning and end of the section,
322 // respectively. This is not requested by the ELF standard, but GNU ld and
323 // gold provide the feature, and used by many programs.
324 static void addStartStopSymbols(const OutputSegment *seg) {
325   StringRef name = seg->name;
326   if (!isValidCIdentifier(name))
327     return;
328   LLVM_DEBUG(dbgs() << "addStartStopSymbols: " << name << "\n");
329   uint32_t start = seg->startVA;
330   uint32_t stop = start + seg->size;
331   symtab->addOptionalDataSymbol(saver.save("__start_" + name), start);
332   symtab->addOptionalDataSymbol(saver.save("__stop_" + name), stop);
333 }
334 
335 void Writer::addSections() {
336   addSection(out.dylinkSec);
337   addSection(out.typeSec);
338   addSection(out.importSec);
339   addSection(out.functionSec);
340   addSection(out.tableSec);
341   addSection(out.memorySec);
342   addSection(out.globalSec);
343   addSection(out.eventSec);
344   addSection(out.exportSec);
345   addSection(out.startSec);
346   addSection(out.elemSec);
347   addSection(out.dataCountSec);
348 
349   addSection(make<CodeSection>(out.functionSec->inputFunctions));
350   addSection(make<DataSection>(segments));
351 
352   createCustomSections();
353 
354   addSection(out.linkingSec);
355   if (config->emitRelocs || config->relocatable) {
356     createRelocSections();
357   }
358 
359   addSection(out.nameSec);
360   addSection(out.producersSec);
361   addSection(out.targetFeaturesSec);
362 }
363 
364 void Writer::finalizeSections() {
365   for (OutputSection *s : outputSections) {
366     s->setOffset(fileSize);
367     s->finalizeContents();
368     fileSize += s->getSize();
369   }
370 }
371 
372 void Writer::populateTargetFeatures() {
373   StringMap<std::string> used;
374   StringMap<std::string> required;
375   StringMap<std::string> disallowed;
376   SmallSet<std::string, 8> &allowed = out.targetFeaturesSec->features;
377   bool tlsUsed = false;
378 
379   // Only infer used features if user did not specify features
380   bool inferFeatures = !config->features.hasValue();
381 
382   if (!inferFeatures) {
383     auto &explicitFeatures = config->features.getValue();
384     allowed.insert(explicitFeatures.begin(), explicitFeatures.end());
385     if (!config->checkFeatures)
386       return;
387   }
388 
389   // Find the sets of used, required, and disallowed features
390   for (ObjFile *file : symtab->objectFiles) {
391     StringRef fileName(file->getName());
392     for (auto &feature : file->getWasmObj()->getTargetFeatures()) {
393       switch (feature.Prefix) {
394       case WASM_FEATURE_PREFIX_USED:
395         used.insert({feature.Name, fileName});
396         break;
397       case WASM_FEATURE_PREFIX_REQUIRED:
398         used.insert({feature.Name, fileName});
399         required.insert({feature.Name, fileName});
400         break;
401       case WASM_FEATURE_PREFIX_DISALLOWED:
402         disallowed.insert({feature.Name, fileName});
403         break;
404       default:
405         error("Unrecognized feature policy prefix " +
406               std::to_string(feature.Prefix));
407       }
408     }
409 
410     // Find TLS data segments
411     auto isTLS = [](InputSegment *segment) {
412       StringRef name = segment->getName();
413       return segment->live &&
414              (name.startswith(".tdata") || name.startswith(".tbss"));
415     };
416     tlsUsed = tlsUsed ||
417               std::any_of(file->segments.begin(), file->segments.end(), isTLS);
418   }
419 
420   if (inferFeatures)
421     allowed.insert(used.keys().begin(), used.keys().end());
422 
423   if (allowed.count("atomics") && !config->sharedMemory) {
424     if (inferFeatures)
425       error(Twine("'atomics' feature is used by ") + used["atomics"] +
426             ", so --shared-memory must be used");
427     else
428       error("'atomics' feature is used, so --shared-memory must be used");
429   }
430 
431   if (!config->checkFeatures)
432     return;
433 
434   if (disallowed.count("atomics") && config->sharedMemory)
435     error("'atomics' feature is disallowed by " + disallowed["atomics"] +
436           ", so --shared-memory must not be used");
437 
438   if (!allowed.count("atomics") && config->sharedMemory)
439     error("'atomics' feature must be used in order to use shared "
440           "memory");
441 
442   if (!allowed.count("bulk-memory") && config->sharedMemory)
443     error("'bulk-memory' feature must be used in order to use shared "
444           "memory");
445 
446   if (!allowed.count("bulk-memory") && tlsUsed)
447     error("'bulk-memory' feature must be used in order to use thread-local "
448           "storage");
449 
450   // Validate that used features are allowed in output
451   if (!inferFeatures) {
452     for (auto &feature : used.keys()) {
453       if (!allowed.count(feature))
454         error(Twine("Target feature '") + feature + "' used by " +
455               used[feature] + " is not allowed.");
456     }
457   }
458 
459   // Validate the required and disallowed constraints for each file
460   for (ObjFile *file : symtab->objectFiles) {
461     StringRef fileName(file->getName());
462     SmallSet<std::string, 8> objectFeatures;
463     for (auto &feature : file->getWasmObj()->getTargetFeatures()) {
464       if (feature.Prefix == WASM_FEATURE_PREFIX_DISALLOWED)
465         continue;
466       objectFeatures.insert(feature.Name);
467       if (disallowed.count(feature.Name))
468         error(Twine("Target feature '") + feature.Name + "' used in " +
469               fileName + " is disallowed by " + disallowed[feature.Name] +
470               ". Use --no-check-features to suppress.");
471     }
472     for (auto &feature : required.keys()) {
473       if (!objectFeatures.count(feature))
474         error(Twine("Missing target feature '") + feature + "' in " + fileName +
475               ", required by " + required[feature] +
476               ". Use --no-check-features to suppress.");
477     }
478   }
479 }
480 
481 void Writer::calculateImports() {
482   for (Symbol *sym : symtab->getSymbols()) {
483     if (!sym->isUndefined())
484       continue;
485     if (sym->isWeak() && !config->relocatable)
486       continue;
487     if (!sym->isLive())
488       continue;
489     if (!sym->isUsedInRegularObj)
490       continue;
491     // We don't generate imports for data symbols. They however can be imported
492     // as GOT entries.
493     if (isa<DataSymbol>(sym))
494       continue;
495 
496     LLVM_DEBUG(dbgs() << "import: " << sym->getName() << "\n");
497     out.importSec->addImport(sym);
498   }
499 }
500 
501 void Writer::calculateExports() {
502   if (config->relocatable)
503     return;
504 
505   if (!config->relocatable && !config->importMemory)
506     out.exportSec->exports.push_back(
507         WasmExport{"memory", WASM_EXTERNAL_MEMORY, 0});
508 
509   if (!config->relocatable && config->exportTable)
510     out.exportSec->exports.push_back(
511         WasmExport{functionTableName, WASM_EXTERNAL_TABLE, 0});
512 
513   unsigned fakeGlobalIndex = out.importSec->getNumImportedGlobals() +
514                              out.globalSec->inputGlobals.size();
515 
516   for (Symbol *sym : symtab->getSymbols()) {
517     if (!sym->isExported())
518       continue;
519     if (!sym->isLive())
520       continue;
521 
522     StringRef name = sym->getName();
523     WasmExport export_;
524     if (auto *f = dyn_cast<DefinedFunction>(sym)) {
525       export_ = {name, WASM_EXTERNAL_FUNCTION, f->getFunctionIndex()};
526     } else if (auto *g = dyn_cast<DefinedGlobal>(sym)) {
527       // TODO(sbc): Remove this check once to mutable global proposal is
528       // implement in all major browsers.
529       // See: https://github.com/WebAssembly/mutable-global
530       if (g->getGlobalType()->Mutable) {
531         // Only __stack_pointer and __tls_base should ever be create as mutable.
532         assert(g == WasmSym::stackPointer || g == WasmSym::tlsBase);
533         continue;
534       }
535       export_ = {name, WASM_EXTERNAL_GLOBAL, g->getGlobalIndex()};
536     } else if (auto *e = dyn_cast<DefinedEvent>(sym)) {
537       export_ = {name, WASM_EXTERNAL_EVENT, e->getEventIndex()};
538     } else {
539       auto *d = cast<DefinedData>(sym);
540       out.globalSec->definedFakeGlobals.emplace_back(d);
541       export_ = {name, WASM_EXTERNAL_GLOBAL, fakeGlobalIndex++};
542     }
543 
544     LLVM_DEBUG(dbgs() << "Export: " << name << "\n");
545     out.exportSec->exports.push_back(export_);
546   }
547 }
548 
549 void Writer::populateSymtab() {
550   if (!config->relocatable && !config->emitRelocs)
551     return;
552 
553   for (Symbol *sym : symtab->getSymbols())
554     if (sym->isUsedInRegularObj && sym->isLive())
555       out.linkingSec->addToSymtab(sym);
556 
557   for (ObjFile *file : symtab->objectFiles) {
558     LLVM_DEBUG(dbgs() << "Local symtab entries: " << file->getName() << "\n");
559     for (Symbol *sym : file->getSymbols())
560       if (sym->isLocal() && !isa<SectionSymbol>(sym) && sym->isLive())
561         out.linkingSec->addToSymtab(sym);
562   }
563 }
564 
565 void Writer::calculateTypes() {
566   // The output type section is the union of the following sets:
567   // 1. Any signature used in the TYPE relocation
568   // 2. The signatures of all imported functions
569   // 3. The signatures of all defined functions
570   // 4. The signatures of all imported events
571   // 5. The signatures of all defined events
572 
573   for (ObjFile *file : symtab->objectFiles) {
574     ArrayRef<WasmSignature> types = file->getWasmObj()->types();
575     for (uint32_t i = 0; i < types.size(); i++)
576       if (file->typeIsUsed[i])
577         file->typeMap[i] = out.typeSec->registerType(types[i]);
578   }
579 
580   for (const Symbol *sym : out.importSec->importedSymbols) {
581     if (auto *f = dyn_cast<FunctionSymbol>(sym))
582       out.typeSec->registerType(*f->signature);
583     else if (auto *e = dyn_cast<EventSymbol>(sym))
584       out.typeSec->registerType(*e->signature);
585   }
586 
587   for (const InputFunction *f : out.functionSec->inputFunctions)
588     out.typeSec->registerType(f->signature);
589 
590   for (const InputEvent *e : out.eventSec->inputEvents)
591     out.typeSec->registerType(e->signature);
592 }
593 
594 static void scanRelocations() {
595   for (ObjFile *file : symtab->objectFiles) {
596     LLVM_DEBUG(dbgs() << "scanRelocations: " << file->getName() << "\n");
597     for (InputChunk *chunk : file->functions)
598       scanRelocations(chunk);
599     for (InputChunk *chunk : file->segments)
600       scanRelocations(chunk);
601     for (auto &p : file->customSections)
602       scanRelocations(p);
603   }
604 }
605 
606 void Writer::assignIndexes() {
607   // Seal the import section, since other index spaces such as function and
608   // global are effected by the number of imports.
609   out.importSec->seal();
610 
611   for (InputFunction *func : symtab->syntheticFunctions)
612     out.functionSec->addFunction(func);
613 
614   for (ObjFile *file : symtab->objectFiles) {
615     LLVM_DEBUG(dbgs() << "Functions: " << file->getName() << "\n");
616     for (InputFunction *func : file->functions)
617       out.functionSec->addFunction(func);
618   }
619 
620   for (InputGlobal *global : symtab->syntheticGlobals)
621     out.globalSec->addGlobal(global);
622 
623   for (ObjFile *file : symtab->objectFiles) {
624     LLVM_DEBUG(dbgs() << "Globals: " << file->getName() << "\n");
625     for (InputGlobal *global : file->globals)
626       out.globalSec->addGlobal(global);
627   }
628 
629   for (ObjFile *file : symtab->objectFiles) {
630     LLVM_DEBUG(dbgs() << "Events: " << file->getName() << "\n");
631     for (InputEvent *event : file->events)
632       out.eventSec->addEvent(event);
633   }
634 
635   out.globalSec->assignIndexes();
636 }
637 
638 static StringRef getOutputDataSegmentName(StringRef name) {
639   // With PIC code we currently only support a single data segment since
640   // we only have a single __memory_base to use as our base address.
641   if (config->isPic)
642     return ".data";
643   // We only support one thread-local segment, so we must merge the segments
644   // despite --no-merge-data-segments.
645   // We also need to merge .tbss into .tdata so they share the same offsets.
646   if (name.startswith(".tdata") || name.startswith(".tbss"))
647     return ".tdata";
648   if (!config->mergeDataSegments)
649     return name;
650   if (name.startswith(".text."))
651     return ".text";
652   if (name.startswith(".data."))
653     return ".data";
654   if (name.startswith(".bss."))
655     return ".bss";
656   if (name.startswith(".rodata."))
657     return ".rodata";
658   return name;
659 }
660 
661 void Writer::createOutputSegments() {
662   for (ObjFile *file : symtab->objectFiles) {
663     for (InputSegment *segment : file->segments) {
664       if (!segment->live)
665         continue;
666       StringRef name = getOutputDataSegmentName(segment->getName());
667       OutputSegment *&s = segmentMap[name];
668       if (s == nullptr) {
669         LLVM_DEBUG(dbgs() << "new segment: " << name << "\n");
670         s = make<OutputSegment>(name, segments.size());
671         if (config->sharedMemory || name == ".tdata")
672           s->initFlags = WASM_SEGMENT_IS_PASSIVE;
673         segments.push_back(s);
674       }
675       s->addInputSegment(segment);
676       LLVM_DEBUG(dbgs() << "added data: " << name << ": " << s->size << "\n");
677     }
678   }
679 }
680 
681 static void createFunction(DefinedFunction *func, StringRef bodyContent) {
682   std::string functionBody;
683   {
684     raw_string_ostream os(functionBody);
685     writeUleb128(os, bodyContent.size(), "function size");
686     os << bodyContent;
687   }
688   ArrayRef<uint8_t> body = arrayRefFromStringRef(saver.save(functionBody));
689   cast<SyntheticFunction>(func->function)->setBody(body);
690 }
691 
692 void Writer::createInitMemoryFunction() {
693   LLVM_DEBUG(dbgs() << "createInitMemoryFunction\n");
694   assert(WasmSym::initMemoryFlag);
695   uint32_t flagAddress = WasmSym::initMemoryFlag->getVirtualAddress();
696   std::string bodyContent;
697   {
698     raw_string_ostream os(bodyContent);
699     writeUleb128(os, 0, "num locals");
700 
701     if (segments.size()) {
702       // Initialize memory in a thread-safe manner. The thread that successfully
703       // increments the flag from 0 to 1 is is responsible for performing the
704       // memory initialization. Other threads go sleep on the flag until the
705       // first thread finishing initializing memory, increments the flag to 2,
706       // and wakes all the other threads. Once the flag has been set to 2,
707       // subsequently started threads will skip the sleep. All threads
708       // unconditionally drop their passive data segments once memory has been
709       // initialized. The generated code is as follows:
710       //
711       // (func $__wasm_init_memory
712       //  (if
713       //   (i32.atomic.rmw.cmpxchg align=2 offset=0
714       //    (i32.const $__init_memory_flag)
715       //    (i32.const 0)
716       //    (i32.const 1)
717       //   )
718       //   (then
719       //    (drop
720       //     (i32.atomic.wait align=2 offset=0
721       //      (i32.const $__init_memory_flag)
722       //      (i32.const 1)
723       //      (i32.const -1)
724       //     )
725       //    )
726       //   )
727       //   (else
728       //    ( ... initialize data segments ... )
729       //    (i32.atomic.store align=2 offset=0
730       //     (i32.const $__init_memory_flag)
731       //     (i32.const 2)
732       //    )
733       //    (drop
734       //     (i32.atomic.notify align=2 offset=0
735       //      (i32.const $__init_memory_flag)
736       //      (i32.const -1u)
737       //     )
738       //    )
739       //   )
740       //  )
741       //  ( ... drop data segments ... )
742       // )
743 
744       // Atomically check whether this is the main thread.
745       writeI32Const(os, flagAddress, "flag address");
746       writeI32Const(os, 0, "expected flag value");
747       writeI32Const(os, 1, "flag value");
748       writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix");
749       writeUleb128(os, WASM_OPCODE_I32_RMW_CMPXCHG, "i32.atomic.rmw.cmpxchg");
750       writeMemArg(os, 2, 0);
751       writeU8(os, WASM_OPCODE_IF, "IF");
752       writeU8(os, WASM_TYPE_NORESULT, "blocktype");
753 
754       // Did not increment 0, so wait for main thread to initialize memory
755       writeI32Const(os, flagAddress, "flag address");
756       writeI32Const(os, 1, "expected flag value");
757       writeI64Const(os, -1, "timeout");
758       writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix");
759       writeUleb128(os, WASM_OPCODE_I32_ATOMIC_WAIT, "i32.atomic.wait");
760       writeMemArg(os, 2, 0);
761       writeU8(os, WASM_OPCODE_DROP, "drop");
762 
763       writeU8(os, WASM_OPCODE_ELSE, "ELSE");
764 
765       // Did increment 0, so conditionally initialize passive data segments
766       for (const OutputSegment *s : segments) {
767         if (s->initFlags & WASM_SEGMENT_IS_PASSIVE && s->name != ".tdata") {
768           // destination address
769           writeI32Const(os, s->startVA, "destination address");
770           // source segment offset
771           writeI32Const(os, 0, "segment offset");
772           // memory region size
773           writeI32Const(os, s->size, "memory region size");
774           // memory.init instruction
775           writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix");
776           writeUleb128(os, WASM_OPCODE_MEMORY_INIT, "memory.init");
777           writeUleb128(os, s->index, "segment index immediate");
778           writeU8(os, 0, "memory index immediate");
779         }
780       }
781 
782       // Set flag to 2 to mark end of initialization
783       writeI32Const(os, flagAddress, "flag address");
784       writeI32Const(os, 2, "flag value");
785       writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix");
786       writeUleb128(os, WASM_OPCODE_I32_ATOMIC_STORE, "i32.atomic.store");
787       writeMemArg(os, 2, 0);
788 
789       // Notify any waiters that memory initialization is complete
790       writeI32Const(os, flagAddress, "flag address");
791       writeI32Const(os, -1, "number of waiters");
792       writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix");
793       writeUleb128(os, WASM_OPCODE_ATOMIC_NOTIFY, "atomic.notify");
794       writeMemArg(os, 2, 0);
795       writeU8(os, WASM_OPCODE_DROP, "drop");
796 
797       writeU8(os, WASM_OPCODE_END, "END");
798 
799       // Unconditionally drop passive data segments
800       for (const OutputSegment *s : segments) {
801         if (s->initFlags & WASM_SEGMENT_IS_PASSIVE && s->name != ".tdata") {
802           // data.drop instruction
803           writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix");
804           writeUleb128(os, WASM_OPCODE_DATA_DROP, "data.drop");
805           writeUleb128(os, s->index, "segment index immediate");
806         }
807       }
808     }
809     writeU8(os, WASM_OPCODE_END, "END");
810   }
811 
812   createFunction(WasmSym::initMemory, bodyContent);
813 }
814 
815 // For -shared (PIC) output, we create create a synthetic function which will
816 // apply any relocations to the data segments on startup.  This function is
817 // called __wasm_apply_relocs and is added at the beginning of __wasm_call_ctors
818 // before any of the constructors run.
819 void Writer::createApplyRelocationsFunction() {
820   LLVM_DEBUG(dbgs() << "createApplyRelocationsFunction\n");
821   // First write the body's contents to a string.
822   std::string bodyContent;
823   {
824     raw_string_ostream os(bodyContent);
825     writeUleb128(os, 0, "num locals");
826     for (const OutputSegment *seg : segments)
827       for (const InputSegment *inSeg : seg->inputSegments)
828         inSeg->generateRelocationCode(os);
829     writeU8(os, WASM_OPCODE_END, "END");
830   }
831 
832   createFunction(WasmSym::applyRelocs, bodyContent);
833 }
834 
835 // Create synthetic "__wasm_call_ctors" function based on ctor functions
836 // in input object.
837 void Writer::createCallCtorsFunction() {
838   if (!WasmSym::callCtors->isLive())
839     return;
840 
841   // First write the body's contents to a string.
842   std::string bodyContent;
843   {
844     raw_string_ostream os(bodyContent);
845     writeUleb128(os, 0, "num locals");
846 
847     if (config->isPic) {
848       writeU8(os, WASM_OPCODE_CALL, "CALL");
849       writeUleb128(os, WasmSym::applyRelocs->getFunctionIndex(),
850                    "function index");
851     }
852 
853     // Call constructors
854     for (const WasmInitEntry &f : initFunctions) {
855       writeU8(os, WASM_OPCODE_CALL, "CALL");
856       writeUleb128(os, f.sym->getFunctionIndex(), "function index");
857     }
858     writeU8(os, WASM_OPCODE_END, "END");
859   }
860 
861   createFunction(WasmSym::callCtors, bodyContent);
862 }
863 
864 void Writer::createInitTLSFunction() {
865   if (!WasmSym::initTLS->isLive())
866     return;
867 
868   std::string bodyContent;
869   {
870     raw_string_ostream os(bodyContent);
871 
872     OutputSegment *tlsSeg = nullptr;
873     for (auto *seg : segments) {
874       if (seg->name == ".tdata") {
875         tlsSeg = seg;
876         break;
877       }
878     }
879 
880     writeUleb128(os, 0, "num locals");
881     if (tlsSeg) {
882       writeU8(os, WASM_OPCODE_LOCAL_GET, "local.get");
883       writeUleb128(os, 0, "local index");
884 
885       writeU8(os, WASM_OPCODE_GLOBAL_SET, "global.set");
886       writeUleb128(os, WasmSym::tlsBase->getGlobalIndex(), "global index");
887 
888       writeU8(os, WASM_OPCODE_LOCAL_GET, "local.get");
889       writeUleb128(os, 0, "local index");
890 
891       writeI32Const(os, 0, "segment offset");
892 
893       writeI32Const(os, tlsSeg->size, "memory region size");
894 
895       writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix");
896       writeUleb128(os, WASM_OPCODE_MEMORY_INIT, "MEMORY.INIT");
897       writeUleb128(os, tlsSeg->index, "segment index immediate");
898       writeU8(os, 0, "memory index immediate");
899     }
900     writeU8(os, WASM_OPCODE_END, "end function");
901   }
902 
903   createFunction(WasmSym::initTLS, bodyContent);
904 }
905 
906 // Populate InitFunctions vector with init functions from all input objects.
907 // This is then used either when creating the output linking section or to
908 // synthesize the "__wasm_call_ctors" function.
909 void Writer::calculateInitFunctions() {
910   if (!config->relocatable && !WasmSym::callCtors->isLive())
911     return;
912 
913   for (ObjFile *file : symtab->objectFiles) {
914     const WasmLinkingData &l = file->getWasmObj()->linkingData();
915     for (const WasmInitFunc &f : l.InitFunctions) {
916       FunctionSymbol *sym = file->getFunctionSymbol(f.Symbol);
917       // comdat exclusions can cause init functions be discarded.
918       if (sym->isDiscarded())
919         continue;
920       assert(sym->isLive());
921       if (*sym->signature != WasmSignature{{}, {}})
922         error("invalid signature for init func: " + toString(*sym));
923       LLVM_DEBUG(dbgs() << "initFunctions: " << toString(*sym) << "\n");
924       initFunctions.emplace_back(WasmInitEntry{sym, f.Priority});
925     }
926   }
927 
928   // Sort in order of priority (lowest first) so that they are called
929   // in the correct order.
930   llvm::stable_sort(initFunctions,
931                     [](const WasmInitEntry &l, const WasmInitEntry &r) {
932                       return l.priority < r.priority;
933                     });
934 }
935 
936 void Writer::createSyntheticSections() {
937   out.dylinkSec = make<DylinkSection>();
938   out.typeSec = make<TypeSection>();
939   out.importSec = make<ImportSection>();
940   out.functionSec = make<FunctionSection>();
941   out.tableSec = make<TableSection>();
942   out.memorySec = make<MemorySection>();
943   out.globalSec = make<GlobalSection>();
944   out.eventSec = make<EventSection>();
945   out.exportSec = make<ExportSection>();
946   out.startSec = make<StartSection>(segments.size());
947   out.elemSec = make<ElemSection>();
948   out.dataCountSec = make<DataCountSection>(segments.size());
949   out.linkingSec = make<LinkingSection>(initFunctions, segments);
950   out.nameSec = make<NameSection>();
951   out.producersSec = make<ProducersSection>();
952   out.targetFeaturesSec = make<TargetFeaturesSection>();
953 }
954 
955 void Writer::run() {
956   if (config->relocatable || config->isPic)
957     config->globalBase = 0;
958 
959   // For PIC code the table base is assigned dynamically by the loader.
960   // For non-PIC, we start at 1 so that accessing table index 0 always traps.
961   if (!config->isPic) {
962     config->tableBase = 1;
963     if (WasmSym::definedTableBase)
964       WasmSym::definedTableBase->setVirtualAddress(config->tableBase);
965   }
966 
967   log("-- createOutputSegments");
968   createOutputSegments();
969   log("-- createSyntheticSections");
970   createSyntheticSections();
971   log("-- populateProducers");
972   populateProducers();
973   log("-- populateTargetFeatures");
974   populateTargetFeatures();
975   log("-- calculateImports");
976   calculateImports();
977   log("-- layoutMemory");
978   layoutMemory();
979 
980   if (!config->relocatable) {
981     // Create linker synthesized __start_SECNAME/__stop_SECNAME symbols
982     // This has to be done after memory layout is performed.
983     for (const OutputSegment *seg : segments)
984       addStartStopSymbols(seg);
985   }
986 
987   log("-- scanRelocations");
988   scanRelocations();
989   log("-- assignIndexes");
990   assignIndexes();
991   log("-- calculateInitFunctions");
992   calculateInitFunctions();
993 
994   if (!config->relocatable) {
995     // Create linker synthesized functions
996     if (config->sharedMemory)
997       createInitMemoryFunction();
998     if (config->isPic)
999       createApplyRelocationsFunction();
1000     createCallCtorsFunction();
1001   }
1002 
1003   if (!config->relocatable && config->sharedMemory && !config->shared)
1004     createInitTLSFunction();
1005 
1006   if (errorCount())
1007     return;
1008 
1009   log("-- calculateTypes");
1010   calculateTypes();
1011   log("-- calculateExports");
1012   calculateExports();
1013   log("-- calculateCustomSections");
1014   calculateCustomSections();
1015   log("-- populateSymtab");
1016   populateSymtab();
1017   log("-- addSections");
1018   addSections();
1019 
1020   if (errorHandler().verbose) {
1021     log("Defined Functions: " + Twine(out.functionSec->inputFunctions.size()));
1022     log("Defined Globals  : " + Twine(out.globalSec->inputGlobals.size()));
1023     log("Defined Events   : " + Twine(out.eventSec->inputEvents.size()));
1024     log("Function Imports : " +
1025         Twine(out.importSec->getNumImportedFunctions()));
1026     log("Global Imports   : " + Twine(out.importSec->getNumImportedGlobals()));
1027     log("Event Imports    : " + Twine(out.importSec->getNumImportedEvents()));
1028     for (ObjFile *file : symtab->objectFiles)
1029       file->dumpInfo();
1030   }
1031 
1032   createHeader();
1033   log("-- finalizeSections");
1034   finalizeSections();
1035 
1036   log("-- openFile");
1037   openFile();
1038   if (errorCount())
1039     return;
1040 
1041   writeHeader();
1042 
1043   log("-- writeSections");
1044   writeSections();
1045   if (errorCount())
1046     return;
1047 
1048   if (Error e = buffer->commit())
1049     fatal("failed to write the output file: " + toString(std::move(e)));
1050 }
1051 
1052 // Open a result file.
1053 void Writer::openFile() {
1054   log("writing: " + config->outputFile);
1055 
1056   Expected<std::unique_ptr<FileOutputBuffer>> bufferOrErr =
1057       FileOutputBuffer::create(config->outputFile, fileSize,
1058                                FileOutputBuffer::F_executable);
1059 
1060   if (!bufferOrErr)
1061     error("failed to open " + config->outputFile + ": " +
1062           toString(bufferOrErr.takeError()));
1063   else
1064     buffer = std::move(*bufferOrErr);
1065 }
1066 
1067 void Writer::createHeader() {
1068   raw_string_ostream os(header);
1069   writeBytes(os, WasmMagic, sizeof(WasmMagic), "wasm magic");
1070   writeU32(os, WasmVersion, "wasm version");
1071   os.flush();
1072   fileSize += header.size();
1073 }
1074 
1075 void lld::wasm::writeResult() { Writer().run(); }
1076