1 //===- lib/Linker/LinkModules.cpp - Module Linker Implementation ----------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the LLVM module linker.
11 //
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Linker/Linker.h"
15 #include "llvm-c/Linker.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallString.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/DiagnosticInfo.h"
21 #include "llvm/IR/DiagnosticPrinter.h"
22 #include "llvm/IR/LLVMContext.h"
23 #include "llvm/IR/Module.h"
24 #include "llvm/IR/TypeFinder.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Transforms/Utils/Cloning.h"
29 #include <cctype>
30 #include <tuple>
31 using namespace llvm;
34 //===----------------------------------------------------------------------===//
35 // TypeMap implementation.
36 //===----------------------------------------------------------------------===//
38 namespace {
39 typedef SmallPtrSet<StructType *, 32> TypeSet;
41 class TypeMapTy : public ValueMapTypeRemapper {
42 /// This is a mapping from a source type to a destination type to use.
43 DenseMap<Type*, Type*> MappedTypes;
45 /// When checking to see if two subgraphs are isomorphic, we speculatively
46 /// add types to MappedTypes, but keep track of them here in case we need to
47 /// roll back.
48 SmallVector<Type*, 16> SpeculativeTypes;
50 /// This is a list of non-opaque structs in the source module that are mapped
51 /// to an opaque struct in the destination module.
52 SmallVector<StructType*, 16> SrcDefinitionsToResolve;
54 /// This is the set of opaque types in the destination modules who are
55 /// getting a body from the source module.
56 SmallPtrSet<StructType*, 16> DstResolvedOpaqueTypes;
58 public:
59 TypeMapTy(TypeSet &Set) : DstStructTypesSet(Set) {}
61 TypeSet &DstStructTypesSet;
62 /// Indicate that the specified type in the destination module is conceptually
63 /// equivalent to the specified type in the source module.
64 void addTypeMapping(Type *DstTy, Type *SrcTy);
66 /// Produce a body for an opaque type in the dest module from a type
67 /// definition in the source module.
68 void linkDefinedTypeBodies();
70 /// Return the mapped type to use for the specified input type from the
71 /// source module.
72 Type *get(Type *SrcTy);
74 FunctionType *get(FunctionType *T) {
75 return cast<FunctionType>(get((Type *)T));
76 }
78 /// Dump out the type map for debugging purposes.
79 void dump() const {
80 for (auto &Pair : MappedTypes) {
81 dbgs() << "TypeMap: ";
82 Pair.first->print(dbgs());
83 dbgs() << " => ";
84 Pair.second->print(dbgs());
85 dbgs() << '\n';
86 }
87 }
89 private:
90 Type *remapType(Type *SrcTy) override { return get(SrcTy); }
92 bool areTypesIsomorphic(Type *DstTy, Type *SrcTy);
93 };
94 }
96 void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) {
97 // Check to see if these types are recursively isomorphic and establish a
98 // mapping between them if so.
99 if (areTypesIsomorphic(DstTy, SrcTy)) {
100 SpeculativeTypes.clear();
101 return;
102 }
104 // Oops, they aren't isomorphic. Just discard this request by rolling out
105 // any speculative mappings we've established.
106 unsigned Removed = 0;
107 for (unsigned I = 0, E = SpeculativeTypes.size(); I != E; ++I) {
108 Type *SrcTy = SpeculativeTypes[I];
109 auto Iter = MappedTypes.find(SrcTy);
110 auto *DstTy = dyn_cast<StructType>(Iter->second);
111 if (DstTy && DstResolvedOpaqueTypes.erase(DstTy))
112 Removed++;
113 MappedTypes.erase(Iter);
114 }
115 SrcDefinitionsToResolve.resize(SrcDefinitionsToResolve.size() - Removed);
116 SpeculativeTypes.clear();
117 }
119 /// Recursively walk this pair of types, returning true if they are isomorphic,
120 /// false if they are not.
121 bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) {
122 // Two types with differing kinds are clearly not isomorphic.
123 if (DstTy->getTypeID() != SrcTy->getTypeID())
124 return false;
126 // If we have an entry in the MappedTypes table, then we have our answer.
127 Type *&Entry = MappedTypes[SrcTy];
128 if (Entry)
129 return Entry == DstTy;
131 // Two identical types are clearly isomorphic. Remember this
132 // non-speculatively.
133 if (DstTy == SrcTy) {
134 Entry = DstTy;
135 return true;
136 }
138 // Okay, we have two types with identical kinds that we haven't seen before.
140 // If this is an opaque struct type, special case it.
141 if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) {
142 // Mapping an opaque type to any struct, just keep the dest struct.
143 if (SSTy->isOpaque()) {
144 Entry = DstTy;
145 SpeculativeTypes.push_back(SrcTy);
146 return true;
147 }
149 // Mapping a non-opaque source type to an opaque dest. If this is the first
150 // type that we're mapping onto this destination type then we succeed. Keep
151 // the dest, but fill it in later. If this is the second (different) type
152 // that we're trying to map onto the same opaque type then we fail.
153 if (cast<StructType>(DstTy)->isOpaque()) {
154 // We can only map one source type onto the opaque destination type.
155 if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)).second)
156 return false;
157 SrcDefinitionsToResolve.push_back(SSTy);
158 SpeculativeTypes.push_back(SrcTy);
159 Entry = DstTy;
160 return true;
161 }
162 }
164 // If the number of subtypes disagree between the two types, then we fail.
165 if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes())
166 return false;
168 // Fail if any of the extra properties (e.g. array size) of the type disagree.
169 if (isa<IntegerType>(DstTy))
170 return false; // bitwidth disagrees.
171 if (PointerType *PT = dyn_cast<PointerType>(DstTy)) {
172 if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace())
173 return false;
175 } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) {
176 if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg())
177 return false;
178 } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) {
179 StructType *SSTy = cast<StructType>(SrcTy);
180 if (DSTy->isLiteral() != SSTy->isLiteral() ||
181 DSTy->isPacked() != SSTy->isPacked())
182 return false;
183 } else if (ArrayType *DATy = dyn_cast<ArrayType>(DstTy)) {
184 if (DATy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements())
185 return false;
186 } else if (VectorType *DVTy = dyn_cast<VectorType>(DstTy)) {
187 if (DVTy->getNumElements() != cast<VectorType>(SrcTy)->getNumElements())
188 return false;
189 }
191 // Otherwise, we speculate that these two types will line up and recursively
192 // check the subelements.
193 Entry = DstTy;
194 SpeculativeTypes.push_back(SrcTy);
196 for (unsigned I = 0, E = SrcTy->getNumContainedTypes(); I != E; ++I)
197 if (!areTypesIsomorphic(DstTy->getContainedType(I),
198 SrcTy->getContainedType(I)))
199 return false;
201 // If everything seems to have lined up, then everything is great.
202 return true;
203 }
205 void TypeMapTy::linkDefinedTypeBodies() {
206 SmallVector<Type*, 16> Elements;
207 for (StructType *SrcSTy : SrcDefinitionsToResolve) {
208 StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]);
209 assert(DstSTy->isOpaque());
211 // Map the body of the source type over to a new body for the dest type.
212 Elements.resize(SrcSTy->getNumElements());
213 for (unsigned I = 0, E = Elements.size(); I != E; ++I)
214 Elements[I] = get(SrcSTy->getElementType(I));
216 DstSTy->setBody(Elements, SrcSTy->isPacked());
217 }
218 SrcDefinitionsToResolve.clear();
219 DstResolvedOpaqueTypes.clear();
220 }
222 Type *TypeMapTy::get(Type *Ty) {
223 // If we already have an entry for this type, return it.
224 Type **Entry = &MappedTypes[Ty];
225 if (*Entry)
226 return *Entry;
228 // If this is not a named struct type, then just map all of the elements and
229 // then rebuild the type from inside out.
230 if (!isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral()) {
231 // If there are no element types to map, then the type is itself. This is
232 // true for the anonymous {} struct, things like 'float', integers, etc.
233 if (Ty->getNumContainedTypes() == 0)
234 return *Entry = Ty;
236 // Remap all of the elements, keeping track of whether any of them change.
237 bool AnyChange = false;
238 SmallVector<Type*, 4> ElementTypes;
239 ElementTypes.resize(Ty->getNumContainedTypes());
240 for (unsigned I = 0, E = Ty->getNumContainedTypes(); I != E; ++I) {
241 ElementTypes[I] = get(Ty->getContainedType(I));
242 AnyChange |= ElementTypes[I] != Ty->getContainedType(I);
243 }
245 // If we found our type while recursively processing stuff, just use it.
246 Entry = &MappedTypes[Ty];
247 if (*Entry)
248 return *Entry;
250 // If all of the element types mapped directly over, then the type is usable
251 // as-is.
252 if (!AnyChange)
253 return *Entry = Ty;
255 // Otherwise, rebuild a modified type.
256 switch (Ty->getTypeID()) {
257 default:
258 llvm_unreachable("unknown derived type to remap");
259 case Type::ArrayTyID:
260 return *Entry = ArrayType::get(ElementTypes[0],
261 cast<ArrayType>(Ty)->getNumElements());
262 case Type::VectorTyID:
263 return *Entry = VectorType::get(ElementTypes[0],
264 cast<VectorType>(Ty)->getNumElements());
265 case Type::PointerTyID:
266 return *Entry = PointerType::get(
267 ElementTypes[0], cast<PointerType>(Ty)->getAddressSpace());
268 case Type::FunctionTyID:
269 return *Entry = FunctionType::get(ElementTypes[0],
270 makeArrayRef(ElementTypes).slice(1),
271 cast<FunctionType>(Ty)->isVarArg());
272 case Type::StructTyID:
273 // Note that this is only reached for anonymous structs.
274 return *Entry = StructType::get(Ty->getContext(), ElementTypes,
275 cast<StructType>(Ty)->isPacked());
276 }
277 }
279 // Otherwise, this is an unmapped named struct. If the struct can be directly
280 // mapped over, just use it as-is. This happens in a case when the linked-in
281 // module has something like:
282 // %T = type {%T*, i32}
283 // @GV = global %T* null
284 // where T does not exist at all in the destination module.
285 //
286 // The other case we watch for is when the type is not in the destination
287 // module, but that it has to be rebuilt because it refers to something that
288 // is already mapped. For example, if the destination module has:
289 // %A = type { i32 }
290 // and the source module has something like
291 // %A' = type { i32 }
292 // %B = type { %A'* }
293 // @GV = global %B* null
294 // then we want to create a new type: "%B = type { %A*}" and have it take the
295 // pristine "%B" name from the source module.
296 //
297 // To determine which case this is, we have to recursively walk the type graph
298 // speculating that we'll be able to reuse it unmodified. Only if this is
299 // safe would we map the entire thing over. Because this is an optimization,
300 // and is not required for the prettiness of the linked module, we just skip
301 // it and always rebuild a type here.
302 StructType *STy = cast<StructType>(Ty);
304 // If the type is opaque, we can just use it directly.
305 if (STy->isOpaque()) {
306 // A named structure type from src module is used. Add it to the Set of
307 // identified structs in the destination module.
308 DstStructTypesSet.insert(STy);
309 return *Entry = STy;
310 }
312 // Otherwise we create a new type.
313 StructType *DTy = StructType::create(STy->getContext());
314 // A new identified structure type was created. Add it to the set of
315 // identified structs in the destination module.
316 DstStructTypesSet.insert(DTy);
317 *Entry = DTy;
319 SmallVector<Type*, 4> ElementTypes;
320 ElementTypes.resize(STy->getNumElements());
321 for (unsigned I = 0, E = ElementTypes.size(); I != E; ++I)
322 ElementTypes[I] = get(STy->getElementType(I));
323 DTy->setBody(ElementTypes, STy->isPacked());
325 // Steal STy's name.
326 if (STy->hasName()) {
327 SmallString<16> TmpName = STy->getName();
328 STy->setName("");
329 DTy->setName(TmpName);
330 }
332 return DTy;
333 }
335 //===----------------------------------------------------------------------===//
336 // ModuleLinker implementation.
337 //===----------------------------------------------------------------------===//
339 namespace {
340 class ModuleLinker;
342 /// Creates prototypes for functions that are lazily linked on the fly. This
343 /// speeds up linking for modules with many/ lazily linked functions of which
344 /// few get used.
345 class ValueMaterializerTy : public ValueMaterializer {
346 TypeMapTy &TypeMap;
347 Module *DstM;
348 std::vector<Function *> &LazilyLinkFunctions;
350 public:
351 ValueMaterializerTy(TypeMapTy &TypeMap, Module *DstM,
352 std::vector<Function *> &LazilyLinkFunctions)
353 : ValueMaterializer(), TypeMap(TypeMap), DstM(DstM),
354 LazilyLinkFunctions(LazilyLinkFunctions) {}
356 Value *materializeValueFor(Value *V) override;
357 };
359 class LinkDiagnosticInfo : public DiagnosticInfo {
360 const Twine &Msg;
362 public:
363 LinkDiagnosticInfo(DiagnosticSeverity Severity, const Twine &Msg);
364 void print(DiagnosticPrinter &DP) const override;
365 };
366 LinkDiagnosticInfo::LinkDiagnosticInfo(DiagnosticSeverity Severity,
367 const Twine &Msg)
368 : DiagnosticInfo(DK_Linker, Severity), Msg(Msg) {}
369 void LinkDiagnosticInfo::print(DiagnosticPrinter &DP) const { DP << Msg; }
371 /// This is an implementation class for the LinkModules function, which is the
372 /// entrypoint for this file.
373 class ModuleLinker {
374 Module *DstM, *SrcM;
376 TypeMapTy TypeMap;
377 ValueMaterializerTy ValMaterializer;
379 /// Mapping of values from what they used to be in Src, to what they are now
380 /// in DstM. ValueToValueMapTy is a ValueMap, which involves some overhead
381 /// due to the use of Value handles which the Linker doesn't actually need,
382 /// but this allows us to reuse the ValueMapper code.
383 ValueToValueMapTy ValueMap;
385 struct AppendingVarInfo {
386 GlobalVariable *NewGV; // New aggregate global in dest module.
387 const Constant *DstInit; // Old initializer from dest module.
388 const Constant *SrcInit; // Old initializer from src module.
389 };
391 std::vector<AppendingVarInfo> AppendingVars;
393 // Set of items not to link in from source.
394 SmallPtrSet<const Value *, 16> DoNotLinkFromSource;
396 // Vector of functions to lazily link in.
397 std::vector<Function *> LazilyLinkFunctions;
399 Linker::DiagnosticHandlerFunction DiagnosticHandler;
401 public:
402 ModuleLinker(Module *dstM, TypeSet &Set, Module *srcM,
403 Linker::DiagnosticHandlerFunction DiagnosticHandler)
404 : DstM(dstM), SrcM(srcM), TypeMap(Set),
405 ValMaterializer(TypeMap, DstM, LazilyLinkFunctions),
406 DiagnosticHandler(DiagnosticHandler) {}
408 bool run();
410 private:
411 bool shouldLinkFromSource(bool &LinkFromSrc, const GlobalValue &Dest,
412 const GlobalValue &Src);
414 /// Helper method for setting a message and returning an error code.
415 bool emitError(const Twine &Message) {
416 DiagnosticHandler(LinkDiagnosticInfo(DS_Error, Message));
417 return true;
418 }
420 void emitWarning(const Twine &Message) {
421 DiagnosticHandler(LinkDiagnosticInfo(DS_Warning, Message));
422 }
424 bool getComdatLeader(Module *M, StringRef ComdatName,
425 const GlobalVariable *&GVar);
426 bool computeResultingSelectionKind(StringRef ComdatName,
427 Comdat::SelectionKind Src,
428 Comdat::SelectionKind Dst,
429 Comdat::SelectionKind &Result,
430 bool &LinkFromSrc);
431 std::map<const Comdat *, std::pair<Comdat::SelectionKind, bool>>
432 ComdatsChosen;
433 bool getComdatResult(const Comdat *SrcC, Comdat::SelectionKind &SK,
434 bool &LinkFromSrc);
436 /// Given a global in the source module, return the global in the
437 /// destination module that is being linked to, if any.
438 GlobalValue *getLinkedToGlobal(const GlobalValue *SrcGV) {
439 // If the source has no name it can't link. If it has local linkage,
440 // there is no name match-up going on.
441 if (!SrcGV->hasName() || SrcGV->hasLocalLinkage())
442 return nullptr;
444 // Otherwise see if we have a match in the destination module's symtab.
445 GlobalValue *DGV = DstM->getNamedValue(SrcGV->getName());
446 if (!DGV)
447 return nullptr;
449 // If we found a global with the same name in the dest module, but it has
450 // internal linkage, we are really not doing any linkage here.
451 if (DGV->hasLocalLinkage())
452 return nullptr;
454 // Otherwise, we do in fact link to the destination global.
455 return DGV;
456 }
458 void computeTypeMapping();
460 void upgradeMismatchedGlobalArray(StringRef Name);
461 void upgradeMismatchedGlobals();
463 bool linkAppendingVarProto(GlobalVariable *DstGV,
464 const GlobalVariable *SrcGV);
466 bool linkGlobalValueProto(GlobalValue *GV);
467 GlobalValue *linkGlobalVariableProto(const GlobalVariable *SGVar,
468 GlobalValue *DGV, bool LinkFromSrc);
469 GlobalValue *linkFunctionProto(const Function *SF, GlobalValue *DGV,
470 bool LinkFromSrc);
471 GlobalValue *linkGlobalAliasProto(const GlobalAlias *SGA, GlobalValue *DGV,
472 bool LinkFromSrc);
474 bool linkModuleFlagsMetadata();
476 void linkAppendingVarInit(const AppendingVarInfo &AVI);
477 void linkGlobalInits();
478 void linkFunctionBody(Function *Dst, Function *Src);
479 void linkAliasBodies();
480 void linkNamedMDNodes();
481 };
482 }
484 /// The LLVM SymbolTable class autorenames globals that conflict in the symbol
485 /// table. This is good for all clients except for us. Go through the trouble
486 /// to force this back.
487 static void forceRenaming(GlobalValue *GV, StringRef Name) {
488 // If the global doesn't force its name or if it already has the right name,
489 // there is nothing for us to do.
490 if (GV->hasLocalLinkage() || GV->getName() == Name)
491 return;
493 Module *M = GV->getParent();
495 // If there is a conflict, rename the conflict.
496 if (GlobalValue *ConflictGV = M->getNamedValue(Name)) {
497 GV->takeName(ConflictGV);
498 ConflictGV->setName(Name); // This will cause ConflictGV to get renamed
499 assert(ConflictGV->getName() != Name && "forceRenaming didn't work");
500 } else {
501 GV->setName(Name); // Force the name back
502 }
503 }
505 /// copy additional attributes (those not needed to construct a GlobalValue)
506 /// from the SrcGV to the DestGV.
507 static void copyGVAttributes(GlobalValue *DestGV, const GlobalValue *SrcGV) {
508 // Use the maximum alignment, rather than just copying the alignment of SrcGV.
509 auto *DestGO = dyn_cast<GlobalObject>(DestGV);
510 unsigned Alignment;
511 if (DestGO)
512 Alignment = std::max(DestGO->getAlignment(), SrcGV->getAlignment());
514 DestGV->copyAttributesFrom(SrcGV);
516 if (DestGO)
517 DestGO->setAlignment(Alignment);
519 forceRenaming(DestGV, SrcGV->getName());
520 }
522 static bool isLessConstraining(GlobalValue::VisibilityTypes a,
523 GlobalValue::VisibilityTypes b) {
524 if (a == GlobalValue::HiddenVisibility)
525 return false;
526 if (b == GlobalValue::HiddenVisibility)
527 return true;
528 if (a == GlobalValue::ProtectedVisibility)
529 return false;
530 if (b == GlobalValue::ProtectedVisibility)
531 return true;
532 return false;
533 }
535 Value *ValueMaterializerTy::materializeValueFor(Value *V) {
536 Function *SF = dyn_cast<Function>(V);
537 if (!SF)
538 return nullptr;
540 Function *DF = Function::Create(TypeMap.get(SF->getFunctionType()),
541 SF->getLinkage(), SF->getName(), DstM);
542 copyGVAttributes(DF, SF);
544 if (Comdat *SC = SF->getComdat()) {
545 Comdat *DC = DstM->getOrInsertComdat(SC->getName());
546 DF->setComdat(DC);
547 }
549 LazilyLinkFunctions.push_back(SF);
550 return DF;
551 }
553 bool ModuleLinker::getComdatLeader(Module *M, StringRef ComdatName,
554 const GlobalVariable *&GVar) {
555 const GlobalValue *GVal = M->getNamedValue(ComdatName);
556 if (const auto *GA = dyn_cast_or_null<GlobalAlias>(GVal)) {
557 GVal = GA->getBaseObject();
558 if (!GVal)
559 // We cannot resolve the size of the aliasee yet.
560 return emitError("Linking COMDATs named '" + ComdatName +
561 "': COMDAT key involves incomputable alias size.");
562 }
564 GVar = dyn_cast_or_null<GlobalVariable>(GVal);
565 if (!GVar)
566 return emitError(
567 "Linking COMDATs named '" + ComdatName +
568 "': GlobalVariable required for data dependent selection!");
570 return false;
571 }
573 bool ModuleLinker::computeResultingSelectionKind(StringRef ComdatName,
574 Comdat::SelectionKind Src,
575 Comdat::SelectionKind Dst,
576 Comdat::SelectionKind &Result,
577 bool &LinkFromSrc) {
578 // The ability to mix Comdat::SelectionKind::Any with
579 // Comdat::SelectionKind::Largest is a behavior that comes from COFF.
580 bool DstAnyOrLargest = Dst == Comdat::SelectionKind::Any ||
581 Dst == Comdat::SelectionKind::Largest;
582 bool SrcAnyOrLargest = Src == Comdat::SelectionKind::Any ||
583 Src == Comdat::SelectionKind::Largest;
584 if (DstAnyOrLargest && SrcAnyOrLargest) {
585 if (Dst == Comdat::SelectionKind::Largest ||
586 Src == Comdat::SelectionKind::Largest)
587 Result = Comdat::SelectionKind::Largest;
588 else
589 Result = Comdat::SelectionKind::Any;
590 } else if (Src == Dst) {
591 Result = Dst;
592 } else {
593 return emitError("Linking COMDATs named '" + ComdatName +
594 "': invalid selection kinds!");
595 }
597 switch (Result) {
598 case Comdat::SelectionKind::Any:
599 // Go with Dst.
600 LinkFromSrc = false;
601 break;
602 case Comdat::SelectionKind::NoDuplicates:
603 return emitError("Linking COMDATs named '" + ComdatName +
604 "': noduplicates has been violated!");
605 case Comdat::SelectionKind::ExactMatch:
606 case Comdat::SelectionKind::Largest:
607 case Comdat::SelectionKind::SameSize: {
608 const GlobalVariable *DstGV;
609 const GlobalVariable *SrcGV;
610 if (getComdatLeader(DstM, ComdatName, DstGV) ||
611 getComdatLeader(SrcM, ComdatName, SrcGV))
612 return true;
614 const DataLayout *DstDL = DstM->getDataLayout();
615 const DataLayout *SrcDL = SrcM->getDataLayout();
616 if (!DstDL || !SrcDL) {
617 return emitError(
618 "Linking COMDATs named '" + ComdatName +
619 "': can't do size dependent selection without DataLayout!");
620 }
621 uint64_t DstSize =
622 DstDL->getTypeAllocSize(DstGV->getType()->getPointerElementType());
623 uint64_t SrcSize =
624 SrcDL->getTypeAllocSize(SrcGV->getType()->getPointerElementType());
625 if (Result == Comdat::SelectionKind::ExactMatch) {
626 if (SrcGV->getInitializer() != DstGV->getInitializer())
627 return emitError("Linking COMDATs named '" + ComdatName +
628 "': ExactMatch violated!");
629 LinkFromSrc = false;
630 } else if (Result == Comdat::SelectionKind::Largest) {
631 LinkFromSrc = SrcSize > DstSize;
632 } else if (Result == Comdat::SelectionKind::SameSize) {
633 if (SrcSize != DstSize)
634 return emitError("Linking COMDATs named '" + ComdatName +
635 "': SameSize violated!");
636 LinkFromSrc = false;
637 } else {
638 llvm_unreachable("unknown selection kind");
639 }
640 break;
641 }
642 }
644 return false;
645 }
647 bool ModuleLinker::getComdatResult(const Comdat *SrcC,
648 Comdat::SelectionKind &Result,
649 bool &LinkFromSrc) {
650 Comdat::SelectionKind SSK = SrcC->getSelectionKind();
651 StringRef ComdatName = SrcC->getName();
652 Module::ComdatSymTabType &ComdatSymTab = DstM->getComdatSymbolTable();
653 Module::ComdatSymTabType::iterator DstCI = ComdatSymTab.find(ComdatName);
655 if (DstCI == ComdatSymTab.end()) {
656 // Use the comdat if it is only available in one of the modules.
657 LinkFromSrc = true;
658 Result = SSK;
659 return false;
660 }
662 const Comdat *DstC = &DstCI->second;
663 Comdat::SelectionKind DSK = DstC->getSelectionKind();
664 return computeResultingSelectionKind(ComdatName, SSK, DSK, Result,
665 LinkFromSrc);
666 }
668 bool ModuleLinker::shouldLinkFromSource(bool &LinkFromSrc,
669 const GlobalValue &Dest,
670 const GlobalValue &Src) {
671 // We always have to add Src if it has appending linkage.
672 if (Src.hasAppendingLinkage()) {
673 LinkFromSrc = true;
674 return false;
675 }
677 bool SrcIsDeclaration = Src.isDeclarationForLinker();
678 bool DestIsDeclaration = Dest.isDeclarationForLinker();
680 if (SrcIsDeclaration) {
681 // If Src is external or if both Src & Dest are external.. Just link the
682 // external globals, we aren't adding anything.
683 if (Src.hasDLLImportStorageClass()) {
684 // If one of GVs is marked as DLLImport, result should be dllimport'ed.
685 LinkFromSrc = DestIsDeclaration;
686 return false;
687 }
688 // If the Dest is weak, use the source linkage.
689 LinkFromSrc = Dest.hasExternalWeakLinkage();
690 return false;
691 }
693 if (DestIsDeclaration) {
694 // If Dest is external but Src is not:
695 LinkFromSrc = true;
696 return false;
697 }
699 if (Src.hasCommonLinkage()) {
700 if (Dest.hasLinkOnceLinkage() || Dest.hasWeakLinkage()) {
701 LinkFromSrc = true;
702 return false;
703 }
705 if (!Dest.hasCommonLinkage()) {
706 LinkFromSrc = false;
707 return false;
708 }
710 // FIXME: Make datalayout mandatory and just use getDataLayout().
711 DataLayout DL(Dest.getParent());
713 uint64_t DestSize = DL.getTypeAllocSize(Dest.getType()->getElementType());
714 uint64_t SrcSize = DL.getTypeAllocSize(Src.getType()->getElementType());
715 LinkFromSrc = SrcSize > DestSize;
716 return false;
717 }
719 if (Src.isWeakForLinker()) {
720 assert(!Dest.hasExternalWeakLinkage());
721 assert(!Dest.hasAvailableExternallyLinkage());
723 if (Dest.hasLinkOnceLinkage() && Src.hasWeakLinkage()) {
724 LinkFromSrc = true;
725 return false;
726 }
728 LinkFromSrc = false;
729 return false;
730 }
732 if (Dest.isWeakForLinker()) {
733 assert(Src.hasExternalLinkage());
734 LinkFromSrc = true;
735 return false;
736 }
738 assert(!Src.hasExternalWeakLinkage());
739 assert(!Dest.hasExternalWeakLinkage());
740 assert(Dest.hasExternalLinkage() && Src.hasExternalLinkage() &&
741 "Unexpected linkage type!");
742 return emitError("Linking globals named '" + Src.getName() +
743 "': symbol multiply defined!");
744 }
746 /// Loop over all of the linked values to compute type mappings. For example,
747 /// if we link "extern Foo *x" and "Foo *x = NULL", then we have two struct
748 /// types 'Foo' but one got renamed when the module was loaded into the same
749 /// LLVMContext.
750 void ModuleLinker::computeTypeMapping() {
751 for (GlobalValue &SGV : SrcM->globals()) {
752 GlobalValue *DGV = getLinkedToGlobal(&SGV);
753 if (!DGV)
754 continue;
756 if (!DGV->hasAppendingLinkage() || !SGV.hasAppendingLinkage()) {
757 TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
758 continue;
759 }
761 // Unify the element type of appending arrays.
762 ArrayType *DAT = cast<ArrayType>(DGV->getType()->getElementType());
763 ArrayType *SAT = cast<ArrayType>(SGV.getType()->getElementType());
764 TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
765 }
767 for (GlobalValue &SGV : *SrcM) {
768 if (GlobalValue *DGV = getLinkedToGlobal(&SGV))
769 TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
770 }
772 for (GlobalValue &SGV : SrcM->aliases()) {
773 if (GlobalValue *DGV = getLinkedToGlobal(&SGV))
774 TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
775 }
777 // Incorporate types by name, scanning all the types in the source module.
778 // At this point, the destination module may have a type "%foo = { i32 }" for
779 // example. When the source module got loaded into the same LLVMContext, if
780 // it had the same type, it would have been renamed to "%foo.42 = { i32 }".
781 TypeFinder SrcStructTypes;
782 SrcStructTypes.run(*SrcM, true);
783 SmallPtrSet<StructType*, 32> SrcStructTypesSet(SrcStructTypes.begin(),
784 SrcStructTypes.end());
786 for (unsigned i = 0, e = SrcStructTypes.size(); i != e; ++i) {
787 StructType *ST = SrcStructTypes[i];
788 if (!ST->hasName()) continue;
790 // Check to see if there is a dot in the name followed by a digit.
791 size_t DotPos = ST->getName().rfind('.');
792 if (DotPos == 0 || DotPos == StringRef::npos ||
793 ST->getName().back() == '.' ||
794 !isdigit(static_cast<unsigned char>(ST->getName()[DotPos+1])))
795 continue;
797 // Check to see if the destination module has a struct with the prefix name.
798 if (StructType *DST = DstM->getTypeByName(ST->getName().substr(0, DotPos)))
799 // Don't use it if this actually came from the source module. They're in
800 // the same LLVMContext after all. Also don't use it unless the type is
801 // actually used in the destination module. This can happen in situations
802 // like this:
803 //
804 // Module A Module B
805 // -------- --------
806 // %Z = type { %A } %B = type { %C.1 }
807 // %A = type { %B.1, [7 x i8] } %C.1 = type { i8* }
808 // %B.1 = type { %C } %A.2 = type { %B.3, [5 x i8] }
809 // %C = type { i8* } %B.3 = type { %C.1 }
810 //
811 // When we link Module B with Module A, the '%B' in Module B is
812 // used. However, that would then use '%C.1'. But when we process '%C.1',
813 // we prefer to take the '%C' version. So we are then left with both
814 // '%C.1' and '%C' being used for the same types. This leads to some
815 // variables using one type and some using the other.
816 if (!SrcStructTypesSet.count(DST) && TypeMap.DstStructTypesSet.count(DST))
817 TypeMap.addTypeMapping(DST, ST);
818 }
820 // Now that we have discovered all of the type equivalences, get a body for
821 // any 'opaque' types in the dest module that are now resolved.
822 TypeMap.linkDefinedTypeBodies();
823 }
825 static void upgradeGlobalArray(GlobalVariable *GV) {
826 ArrayType *ATy = cast<ArrayType>(GV->getType()->getElementType());
827 StructType *OldTy = cast<StructType>(ATy->getElementType());
828 assert(OldTy->getNumElements() == 2 && "Expected to upgrade from 2 elements");
830 // Get the upgraded 3 element type.
831 PointerType *VoidPtrTy = Type::getInt8Ty(GV->getContext())->getPointerTo();
832 Type *Tys[3] = {OldTy->getElementType(0), OldTy->getElementType(1),
833 VoidPtrTy};
834 StructType *NewTy = StructType::get(GV->getContext(), Tys, false);
836 // Build new constants with a null third field filled in.
837 Constant *OldInitC = GV->getInitializer();
838 ConstantArray *OldInit = dyn_cast<ConstantArray>(OldInitC);
839 if (!OldInit && !isa<ConstantAggregateZero>(OldInitC))
840 // Invalid initializer; give up.
841 return;
842 std::vector<Constant *> Initializers;
843 if (OldInit && OldInit->getNumOperands()) {
844 Value *Null = Constant::getNullValue(VoidPtrTy);
845 for (Use &U : OldInit->operands()) {
846 ConstantStruct *Init = cast<ConstantStruct>(U.get());
847 Initializers.push_back(ConstantStruct::get(
848 NewTy, Init->getOperand(0), Init->getOperand(1), Null, nullptr));
849 }
850 }
851 assert(Initializers.size() == ATy->getNumElements() &&
852 "Failed to copy all array elements");
854 // Replace the old GV with a new one.
855 ATy = ArrayType::get(NewTy, Initializers.size());
856 Constant *NewInit = ConstantArray::get(ATy, Initializers);
857 GlobalVariable *NewGV = new GlobalVariable(
858 *GV->getParent(), ATy, GV->isConstant(), GV->getLinkage(), NewInit, "",
859 GV, GV->getThreadLocalMode(), GV->getType()->getAddressSpace(),
860 GV->isExternallyInitialized());
861 NewGV->copyAttributesFrom(GV);
862 NewGV->takeName(GV);
863 assert(GV->use_empty() && "program cannot use initializer list");
864 GV->eraseFromParent();
865 }
867 void ModuleLinker::upgradeMismatchedGlobalArray(StringRef Name) {
868 // Look for the global arrays.
869 auto *DstGV = dyn_cast_or_null<GlobalVariable>(DstM->getNamedValue(Name));
870 if (!DstGV)
871 return;
872 auto *SrcGV = dyn_cast_or_null<GlobalVariable>(SrcM->getNamedValue(Name));
873 if (!SrcGV)
874 return;
876 // Check if the types already match.
877 auto *DstTy = cast<ArrayType>(DstGV->getType()->getElementType());
878 auto *SrcTy =
879 cast<ArrayType>(TypeMap.get(SrcGV->getType()->getElementType()));
880 if (DstTy == SrcTy)
881 return;
883 // Grab the element types. We can only upgrade an array of a two-field
884 // struct. Only bother if the other one has three-fields.
885 auto *DstEltTy = cast<StructType>(DstTy->getElementType());
886 auto *SrcEltTy = cast<StructType>(SrcTy->getElementType());
887 if (DstEltTy->getNumElements() == 2 && SrcEltTy->getNumElements() == 3) {
888 upgradeGlobalArray(DstGV);
889 return;
890 }
891 if (DstEltTy->getNumElements() == 3 && SrcEltTy->getNumElements() == 2)
892 upgradeGlobalArray(SrcGV);
894 // We can't upgrade any other differences.
895 }
897 void ModuleLinker::upgradeMismatchedGlobals() {
898 upgradeMismatchedGlobalArray("llvm.global_ctors");
899 upgradeMismatchedGlobalArray("llvm.global_dtors");
900 }
902 /// If there were any appending global variables, link them together now.
903 /// Return true on error.
904 bool ModuleLinker::linkAppendingVarProto(GlobalVariable *DstGV,
905 const GlobalVariable *SrcGV) {
907 if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage())
908 return emitError("Linking globals named '" + SrcGV->getName() +
909 "': can only link appending global with another appending global!");
911 ArrayType *DstTy = cast<ArrayType>(DstGV->getType()->getElementType());
912 ArrayType *SrcTy =
913 cast<ArrayType>(TypeMap.get(SrcGV->getType()->getElementType()));
914 Type *EltTy = DstTy->getElementType();
916 // Check to see that they two arrays agree on type.
917 if (EltTy != SrcTy->getElementType())
918 return emitError("Appending variables with different element types!");
919 if (DstGV->isConstant() != SrcGV->isConstant())
920 return emitError("Appending variables linked with different const'ness!");
922 if (DstGV->getAlignment() != SrcGV->getAlignment())
923 return emitError(
924 "Appending variables with different alignment need to be linked!");
926 if (DstGV->getVisibility() != SrcGV->getVisibility())
927 return emitError(
928 "Appending variables with different visibility need to be linked!");
930 if (DstGV->hasUnnamedAddr() != SrcGV->hasUnnamedAddr())
931 return emitError(
932 "Appending variables with different unnamed_addr need to be linked!");
934 if (StringRef(DstGV->getSection()) != SrcGV->getSection())
935 return emitError(
936 "Appending variables with different section name need to be linked!");
938 uint64_t NewSize = DstTy->getNumElements() + SrcTy->getNumElements();
939 ArrayType *NewType = ArrayType::get(EltTy, NewSize);
941 // Create the new global variable.
942 GlobalVariable *NG =
943 new GlobalVariable(*DstGV->getParent(), NewType, SrcGV->isConstant(),
944 DstGV->getLinkage(), /*init*/nullptr, /*name*/"", DstGV,
945 DstGV->getThreadLocalMode(),
946 DstGV->getType()->getAddressSpace());
948 // Propagate alignment, visibility and section info.
949 copyGVAttributes(NG, DstGV);
951 AppendingVarInfo AVI;
952 AVI.NewGV = NG;
953 AVI.DstInit = DstGV->getInitializer();
954 AVI.SrcInit = SrcGV->getInitializer();
955 AppendingVars.push_back(AVI);
957 // Replace any uses of the two global variables with uses of the new
958 // global.
959 ValueMap[SrcGV] = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType()));
961 DstGV->replaceAllUsesWith(ConstantExpr::getBitCast(NG, DstGV->getType()));
962 DstGV->eraseFromParent();
964 // Track the source variable so we don't try to link it.
965 DoNotLinkFromSource.insert(SrcGV);
967 return false;
968 }
970 bool ModuleLinker::linkGlobalValueProto(GlobalValue *SGV) {
971 GlobalValue *DGV = getLinkedToGlobal(SGV);
973 // Handle the ultra special appending linkage case first.
974 if (DGV && DGV->hasAppendingLinkage())
975 return linkAppendingVarProto(cast<GlobalVariable>(DGV),
976 cast<GlobalVariable>(SGV));
978 bool LinkFromSrc = true;
979 Comdat *C = nullptr;
980 GlobalValue::VisibilityTypes Visibility = SGV->getVisibility();
981 bool HasUnnamedAddr = SGV->hasUnnamedAddr();
983 if (const Comdat *SC = SGV->getComdat()) {
984 Comdat::SelectionKind SK;
985 std::tie(SK, LinkFromSrc) = ComdatsChosen[SC];
986 C = DstM->getOrInsertComdat(SC->getName());
987 C->setSelectionKind(SK);
988 } else if (DGV) {
989 if (shouldLinkFromSource(LinkFromSrc, *DGV, *SGV))
990 return true;
991 }
993 if (!LinkFromSrc) {
994 // Track the source global so that we don't attempt to copy it over when
995 // processing global initializers.
996 DoNotLinkFromSource.insert(SGV);
998 if (DGV)
999 // Make sure to remember this mapping.
1000 ValueMap[SGV] =
1001 ConstantExpr::getBitCast(DGV, TypeMap.get(SGV->getType()));
1002 }
1004 if (DGV) {
1005 Visibility = isLessConstraining(Visibility, DGV->getVisibility())
1006 ? DGV->getVisibility()
1007 : Visibility;
1008 HasUnnamedAddr = HasUnnamedAddr && DGV->hasUnnamedAddr();
1009 }
1011 if (!LinkFromSrc && !DGV)
1012 return false;
1014 GlobalValue *NewGV;
1015 if (auto *SGVar = dyn_cast<GlobalVariable>(SGV)) {
1016 NewGV = linkGlobalVariableProto(SGVar, DGV, LinkFromSrc);
1017 if (!NewGV)
1018 return true;
1019 } else if (auto *SF = dyn_cast<Function>(SGV)) {
1020 NewGV = linkFunctionProto(SF, DGV, LinkFromSrc);
1021 } else {
1022 NewGV = linkGlobalAliasProto(cast<GlobalAlias>(SGV), DGV, LinkFromSrc);
1023 }
1025 if (NewGV) {
1026 if (NewGV != DGV)
1027 copyGVAttributes(NewGV, SGV);
1029 NewGV->setUnnamedAddr(HasUnnamedAddr);
1030 NewGV->setVisibility(Visibility);
1032 if (auto *NewGO = dyn_cast<GlobalObject>(NewGV)) {
1033 if (C)
1034 NewGO->setComdat(C);
1035 }
1037 // Make sure to remember this mapping.
1038 if (NewGV != DGV) {
1039 if (DGV) {
1040 DGV->replaceAllUsesWith(
1041 ConstantExpr::getBitCast(NewGV, DGV->getType()));
1042 DGV->eraseFromParent();
1043 }
1044 ValueMap[SGV] = NewGV;
1045 }
1046 }
1048 return false;
1049 }
1051 /// Loop through the global variables in the src module and merge them into the
1052 /// dest module.
1053 GlobalValue *ModuleLinker::linkGlobalVariableProto(const GlobalVariable *SGVar,
1054 GlobalValue *DGV,
1055 bool LinkFromSrc) {
1056 unsigned Alignment = 0;
1057 bool ClearConstant = false;
1059 if (DGV) {
1060 if (DGV->hasCommonLinkage() && SGVar->hasCommonLinkage())
1061 Alignment = std::max(SGVar->getAlignment(), DGV->getAlignment());
1063 auto *DGVar = dyn_cast<GlobalVariable>(DGV);
1064 if (!SGVar->isConstant() || (DGVar && !DGVar->isConstant()))
1065 ClearConstant = true;
1066 }
1068 if (!LinkFromSrc) {
1069 if (auto *NewGVar = dyn_cast<GlobalVariable>(DGV)) {
1070 if (Alignment)
1071 NewGVar->setAlignment(Alignment);
1072 if (NewGVar->isDeclaration() && ClearConstant)
1073 NewGVar->setConstant(false);
1074 }
1075 return DGV;
1076 }
1078 // No linking to be performed or linking from the source: simply create an
1079 // identical version of the symbol over in the dest module... the
1080 // initializer will be filled in later by LinkGlobalInits.
1081 GlobalVariable *NewDGV = new GlobalVariable(
1082 *DstM, TypeMap.get(SGVar->getType()->getElementType()),
1083 SGVar->isConstant(), SGVar->getLinkage(), /*init*/ nullptr,
1084 SGVar->getName(), /*insertbefore*/ nullptr, SGVar->getThreadLocalMode(),
1085 SGVar->getType()->getAddressSpace());
1087 if (Alignment)
1088 NewDGV->setAlignment(Alignment);
1090 return NewDGV;
1091 }
1093 /// Link the function in the source module into the destination module if
1094 /// needed, setting up mapping information.
1095 GlobalValue *ModuleLinker::linkFunctionProto(const Function *SF,
1096 GlobalValue *DGV,
1097 bool LinkFromSrc) {
1098 if (!LinkFromSrc)
1099 return DGV;
1101 // If the function is to be lazily linked, don't create it just yet.
1102 // The ValueMaterializerTy will deal with creating it if it's used.
1103 if (!DGV && (SF->hasLocalLinkage() || SF->hasLinkOnceLinkage() ||
1104 SF->hasAvailableExternallyLinkage())) {
1105 DoNotLinkFromSource.insert(SF);
1106 return nullptr;
1107 }
1109 // If there is no linkage to be performed or we are linking from the source,
1110 // bring SF over.
1111 return Function::Create(TypeMap.get(SF->getFunctionType()), SF->getLinkage(),
1112 SF->getName(), DstM);
1113 }
1115 /// Set up prototypes for any aliases that come over from the source module.
1116 GlobalValue *ModuleLinker::linkGlobalAliasProto(const GlobalAlias *SGA,
1117 GlobalValue *DGV,
1118 bool LinkFromSrc) {
1119 if (!LinkFromSrc)
1120 return DGV;
1122 // If there is no linkage to be performed or we're linking from the source,
1123 // bring over SGA.
1124 auto *PTy = cast<PointerType>(TypeMap.get(SGA->getType()));
1125 return GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
1126 SGA->getLinkage(), SGA->getName(), DstM);
1127 }
1129 static void getArrayElements(const Constant *C,
1130 SmallVectorImpl<Constant *> &Dest) {
1131 unsigned NumElements = cast<ArrayType>(C->getType())->getNumElements();
1133 for (unsigned i = 0; i != NumElements; ++i)
1134 Dest.push_back(C->getAggregateElement(i));
1135 }
1137 void ModuleLinker::linkAppendingVarInit(const AppendingVarInfo &AVI) {
1138 // Merge the initializer.
1139 SmallVector<Constant *, 16> DstElements;
1140 getArrayElements(AVI.DstInit, DstElements);
1142 SmallVector<Constant *, 16> SrcElements;
1143 getArrayElements(AVI.SrcInit, SrcElements);
1145 ArrayType *NewType = cast<ArrayType>(AVI.NewGV->getType()->getElementType());
1147 StringRef Name = AVI.NewGV->getName();
1148 bool IsNewStructor =
1149 (Name == "llvm.global_ctors" || Name == "llvm.global_dtors") &&
1150 cast<StructType>(NewType->getElementType())->getNumElements() == 3;
1152 for (auto *V : SrcElements) {
1153 if (IsNewStructor) {
1154 Constant *Key = V->getAggregateElement(2);
1155 if (DoNotLinkFromSource.count(Key))
1156 continue;
1157 }
1158 DstElements.push_back(
1159 MapValue(V, ValueMap, RF_None, &TypeMap, &ValMaterializer));
1160 }
1161 if (IsNewStructor) {
1162 NewType = ArrayType::get(NewType->getElementType(), DstElements.size());
1163 AVI.NewGV->mutateType(PointerType::get(NewType, 0));
1164 }
1166 AVI.NewGV->setInitializer(ConstantArray::get(NewType, DstElements));
1167 }
1169 /// Update the initializers in the Dest module now that all globals that may be
1170 /// referenced are in Dest.
1171 void ModuleLinker::linkGlobalInits() {
1172 // Loop over all of the globals in the src module, mapping them over as we go
1173 for (Module::const_global_iterator I = SrcM->global_begin(),
1174 E = SrcM->global_end(); I != E; ++I) {
1176 // Only process initialized GV's or ones not already in dest.
1177 if (!I->hasInitializer() || DoNotLinkFromSource.count(I)) continue;
1179 // Grab destination global variable.
1180 GlobalVariable *DGV = cast<GlobalVariable>(ValueMap[I]);
1181 // Figure out what the initializer looks like in the dest module.
1182 DGV->setInitializer(MapValue(I->getInitializer(), ValueMap,
1183 RF_None, &TypeMap, &ValMaterializer));
1184 }
1185 }
1187 /// Copy the source function over into the dest function and fix up references
1188 /// to values. At this point we know that Dest is an external function, and
1189 /// that Src is not.
1190 void ModuleLinker::linkFunctionBody(Function *Dst, Function *Src) {
1191 assert(Src && Dst && Dst->isDeclaration() && !Src->isDeclaration());
1193 // Go through and convert function arguments over, remembering the mapping.
1194 Function::arg_iterator DI = Dst->arg_begin();
1195 for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
1196 I != E; ++I, ++DI) {
1197 DI->setName(I->getName()); // Copy the name over.
1199 // Add a mapping to our mapping.
1200 ValueMap[I] = DI;
1201 }
1203 // Splice the body of the source function into the dest function.
1204 Dst->getBasicBlockList().splice(Dst->end(), Src->getBasicBlockList());
1206 // At this point, all of the instructions and values of the function are now
1207 // copied over. The only problem is that they are still referencing values in
1208 // the Source function as operands. Loop through all of the operands of the
1209 // functions and patch them up to point to the local versions.
1210 for (Function::iterator BB = Dst->begin(), BE = Dst->end(); BB != BE; ++BB)
1211 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1212 RemapInstruction(I, ValueMap, RF_IgnoreMissingEntries, &TypeMap,
1213 &ValMaterializer);
1215 // There is no need to map the arguments anymore.
1216 for (Function::arg_iterator I = Src->arg_begin(), E = Src->arg_end();
1217 I != E; ++I)
1218 ValueMap.erase(I);
1220 }
1222 /// Insert all of the aliases in Src into the Dest module.
1223 void ModuleLinker::linkAliasBodies() {
1224 for (Module::alias_iterator I = SrcM->alias_begin(), E = SrcM->alias_end();
1225 I != E; ++I) {
1226 if (DoNotLinkFromSource.count(I))
1227 continue;
1228 if (Constant *Aliasee = I->getAliasee()) {
1229 GlobalAlias *DA = cast<GlobalAlias>(ValueMap[I]);
1230 Constant *Val =
1231 MapValue(Aliasee, ValueMap, RF_None, &TypeMap, &ValMaterializer);
1232 DA->setAliasee(Val);
1233 }
1234 }
1235 }
1237 /// Insert all of the named MDNodes in Src into the Dest module.
1238 void ModuleLinker::linkNamedMDNodes() {
1239 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1240 for (Module::const_named_metadata_iterator I = SrcM->named_metadata_begin(),
1241 E = SrcM->named_metadata_end(); I != E; ++I) {
1242 // Don't link module flags here. Do them separately.
1243 if (&*I == SrcModFlags) continue;
1244 NamedMDNode *DestNMD = DstM->getOrInsertNamedMetadata(I->getName());
1245 // Add Src elements into Dest node.
1246 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1247 DestNMD->addOperand(MapValue(I->getOperand(i), ValueMap,
1248 RF_None, &TypeMap, &ValMaterializer));
1249 }
1250 }
1252 /// Merge the linker flags in Src into the Dest module.
1253 bool ModuleLinker::linkModuleFlagsMetadata() {
1254 // If the source module has no module flags, we are done.
1255 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1256 if (!SrcModFlags) return false;
1258 // If the destination module doesn't have module flags yet, then just copy
1259 // over the source module's flags.
1260 NamedMDNode *DstModFlags = DstM->getOrInsertModuleFlagsMetadata();
1261 if (DstModFlags->getNumOperands() == 0) {
1262 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I)
1263 DstModFlags->addOperand(SrcModFlags->getOperand(I));
1265 return false;
1266 }
1268 // First build a map of the existing module flags and requirements.
1269 DenseMap<MDString*, MDNode*> Flags;
1270 SmallSetVector<MDNode*, 16> Requirements;
1271 for (unsigned I = 0, E = DstModFlags->getNumOperands(); I != E; ++I) {
1272 MDNode *Op = DstModFlags->getOperand(I);
1273 ConstantInt *Behavior = cast<ConstantInt>(Op->getOperand(0));
1274 MDString *ID = cast<MDString>(Op->getOperand(1));
1276 if (Behavior->getZExtValue() == Module::Require) {
1277 Requirements.insert(cast<MDNode>(Op->getOperand(2)));
1278 } else {
1279 Flags[ID] = Op;
1280 }
1281 }
1283 // Merge in the flags from the source module, and also collect its set of
1284 // requirements.
1285 bool HasErr = false;
1286 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) {
1287 MDNode *SrcOp = SrcModFlags->getOperand(I);
1288 ConstantInt *SrcBehavior = cast<ConstantInt>(SrcOp->getOperand(0));
1289 MDString *ID = cast<MDString>(SrcOp->getOperand(1));
1290 MDNode *DstOp = Flags.lookup(ID);
1291 unsigned SrcBehaviorValue = SrcBehavior->getZExtValue();
1293 // If this is a requirement, add it and continue.
1294 if (SrcBehaviorValue == Module::Require) {
1295 // If the destination module does not already have this requirement, add
1296 // it.
1297 if (Requirements.insert(cast<MDNode>(SrcOp->getOperand(2)))) {
1298 DstModFlags->addOperand(SrcOp);
1299 }
1300 continue;
1301 }
1303 // If there is no existing flag with this ID, just add it.
1304 if (!DstOp) {
1305 Flags[ID] = SrcOp;
1306 DstModFlags->addOperand(SrcOp);
1307 continue;
1308 }
1310 // Otherwise, perform a merge.
1311 ConstantInt *DstBehavior = cast<ConstantInt>(DstOp->getOperand(0));
1312 unsigned DstBehaviorValue = DstBehavior->getZExtValue();
1314 // If either flag has override behavior, handle it first.
1315 if (DstBehaviorValue == Module::Override) {
1316 // Diagnose inconsistent flags which both have override behavior.
1317 if (SrcBehaviorValue == Module::Override &&
1318 SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1319 HasErr |= emitError("linking module flags '" + ID->getString() +
1320 "': IDs have conflicting override values");
1321 }
1322 continue;
1323 } else if (SrcBehaviorValue == Module::Override) {
1324 // Update the destination flag to that of the source.
1325 DstOp->replaceOperandWith(0, SrcBehavior);
1326 DstOp->replaceOperandWith(2, SrcOp->getOperand(2));
1327 continue;
1328 }
1330 // Diagnose inconsistent merge behavior types.
1331 if (SrcBehaviorValue != DstBehaviorValue) {
1332 HasErr |= emitError("linking module flags '" + ID->getString() +
1333 "': IDs have conflicting behaviors");
1334 continue;
1335 }
1337 // Perform the merge for standard behavior types.
1338 switch (SrcBehaviorValue) {
1339 case Module::Require:
1340 case Module::Override: llvm_unreachable("not possible");
1341 case Module::Error: {
1342 // Emit an error if the values differ.
1343 if (SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1344 HasErr |= emitError("linking module flags '" + ID->getString() +
1345 "': IDs have conflicting values");
1346 }
1347 continue;
1348 }
1349 case Module::Warning: {
1350 // Emit a warning if the values differ.
1351 if (SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1352 emitWarning("linking module flags '" + ID->getString() +
1353 "': IDs have conflicting values");
1354 }
1355 continue;
1356 }
1357 case Module::Append: {
1358 MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2));
1359 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1360 unsigned NumOps = DstValue->getNumOperands() + SrcValue->getNumOperands();
1361 Value **VP, **Values = VP = new Value*[NumOps];
1362 for (unsigned i = 0, e = DstValue->getNumOperands(); i != e; ++i, ++VP)
1363 *VP = DstValue->getOperand(i);
1364 for (unsigned i = 0, e = SrcValue->getNumOperands(); i != e; ++i, ++VP)
1365 *VP = SrcValue->getOperand(i);
1366 DstOp->replaceOperandWith(2, MDNode::get(DstM->getContext(),
1367 ArrayRef<Value*>(Values,
1368 NumOps)));
1369 delete[] Values;
1370 break;
1371 }
1372 case Module::AppendUnique: {
1373 SmallSetVector<Value*, 16> Elts;
1374 MDNode *DstValue = cast<MDNode>(DstOp->getOperand(2));
1375 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1376 for (unsigned i = 0, e = DstValue->getNumOperands(); i != e; ++i)
1377 Elts.insert(DstValue->getOperand(i));
1378 for (unsigned i = 0, e = SrcValue->getNumOperands(); i != e; ++i)
1379 Elts.insert(SrcValue->getOperand(i));
1380 DstOp->replaceOperandWith(2, MDNode::get(DstM->getContext(),
1381 ArrayRef<Value*>(Elts.begin(),
1382 Elts.end())));
1383 break;
1384 }
1385 }
1386 }
1388 // Check all of the requirements.
1389 for (unsigned I = 0, E = Requirements.size(); I != E; ++I) {
1390 MDNode *Requirement = Requirements[I];
1391 MDString *Flag = cast<MDString>(Requirement->getOperand(0));
1392 Value *ReqValue = Requirement->getOperand(1);
1394 MDNode *Op = Flags[Flag];
1395 if (!Op || Op->getOperand(2) != ReqValue) {
1396 HasErr |= emitError("linking module flags '" + Flag->getString() +
1397 "': does not have the required value");
1398 continue;
1399 }
1400 }
1402 return HasErr;
1403 }
1405 bool ModuleLinker::run() {
1406 assert(DstM && "Null destination module");
1407 assert(SrcM && "Null source module");
1409 // Inherit the target data from the source module if the destination module
1410 // doesn't have one already.
1411 if (!DstM->getDataLayout() && SrcM->getDataLayout())
1412 DstM->setDataLayout(SrcM->getDataLayout());
1414 // Copy the target triple from the source to dest if the dest's is empty.
1415 if (DstM->getTargetTriple().empty() && !SrcM->getTargetTriple().empty())
1416 DstM->setTargetTriple(SrcM->getTargetTriple());
1418 if (SrcM->getDataLayout() && DstM->getDataLayout() &&
1419 *SrcM->getDataLayout() != *DstM->getDataLayout()) {
1420 emitWarning("Linking two modules of different data layouts: '" +
1421 SrcM->getModuleIdentifier() + "' is '" +
1422 SrcM->getDataLayoutStr() + "' whereas '" +
1423 DstM->getModuleIdentifier() + "' is '" +
1424 DstM->getDataLayoutStr() + "'\n");
1425 }
1426 if (!SrcM->getTargetTriple().empty() &&
1427 DstM->getTargetTriple() != SrcM->getTargetTriple()) {
1428 emitWarning("Linking two modules of different target triples: " +
1429 SrcM->getModuleIdentifier() + "' is '" +
1430 SrcM->getTargetTriple() + "' whereas '" +
1431 DstM->getModuleIdentifier() + "' is '" +
1432 DstM->getTargetTriple() + "'\n");
1433 }
1435 // Append the module inline asm string.
1436 if (!SrcM->getModuleInlineAsm().empty()) {
1437 if (DstM->getModuleInlineAsm().empty())
1438 DstM->setModuleInlineAsm(SrcM->getModuleInlineAsm());
1439 else
1440 DstM->setModuleInlineAsm(DstM->getModuleInlineAsm()+"\n"+
1441 SrcM->getModuleInlineAsm());
1442 }
1444 // Loop over all of the linked values to compute type mappings.
1445 computeTypeMapping();
1447 ComdatsChosen.clear();
1448 for (const auto &SMEC : SrcM->getComdatSymbolTable()) {
1449 const Comdat &C = SMEC.getValue();
1450 if (ComdatsChosen.count(&C))
1451 continue;
1452 Comdat::SelectionKind SK;
1453 bool LinkFromSrc;
1454 if (getComdatResult(&C, SK, LinkFromSrc))
1455 return true;
1456 ComdatsChosen[&C] = std::make_pair(SK, LinkFromSrc);
1457 }
1459 // Upgrade mismatched global arrays.
1460 upgradeMismatchedGlobals();
1462 // Insert all of the globals in src into the DstM module... without linking
1463 // initializers (which could refer to functions not yet mapped over).
1464 for (Module::global_iterator I = SrcM->global_begin(),
1465 E = SrcM->global_end(); I != E; ++I)
1466 if (linkGlobalValueProto(I))
1467 return true;
1469 // Link the functions together between the two modules, without doing function
1470 // bodies... this just adds external function prototypes to the DstM
1471 // function... We do this so that when we begin processing function bodies,
1472 // all of the global values that may be referenced are available in our
1473 // ValueMap.
1474 for (Module::iterator I = SrcM->begin(), E = SrcM->end(); I != E; ++I)
1475 if (linkGlobalValueProto(I))
1476 return true;
1478 // If there were any aliases, link them now.
1479 for (Module::alias_iterator I = SrcM->alias_begin(),
1480 E = SrcM->alias_end(); I != E; ++I)
1481 if (linkGlobalValueProto(I))
1482 return true;
1484 for (unsigned i = 0, e = AppendingVars.size(); i != e; ++i)
1485 linkAppendingVarInit(AppendingVars[i]);
1487 // Link in the function bodies that are defined in the source module into
1488 // DstM.
1489 for (Module::iterator SF = SrcM->begin(), E = SrcM->end(); SF != E; ++SF) {
1490 // Skip if not linking from source.
1491 if (DoNotLinkFromSource.count(SF)) continue;
1493 Function *DF = cast<Function>(ValueMap[SF]);
1494 if (SF->hasPrefixData()) {
1495 // Link in the prefix data.
1496 DF->setPrefixData(MapValue(
1497 SF->getPrefixData(), ValueMap, RF_None, &TypeMap, &ValMaterializer));
1498 }
1500 // Materialize if needed.
1501 if (std::error_code EC = SF->materialize())
1502 return emitError(EC.message());
1504 // Skip if no body (function is external).
1505 if (SF->isDeclaration())
1506 continue;
1508 linkFunctionBody(DF, SF);
1509 SF->Dematerialize();
1510 }
1512 // Resolve all uses of aliases with aliasees.
1513 linkAliasBodies();
1515 // Remap all of the named MDNodes in Src into the DstM module. We do this
1516 // after linking GlobalValues so that MDNodes that reference GlobalValues
1517 // are properly remapped.
1518 linkNamedMDNodes();
1520 // Merge the module flags into the DstM module.
1521 if (linkModuleFlagsMetadata())
1522 return true;
1524 // Update the initializers in the DstM module now that all globals that may
1525 // be referenced are in DstM.
1526 linkGlobalInits();
1528 // Process vector of lazily linked in functions.
1529 bool LinkedInAnyFunctions;
1530 do {
1531 LinkedInAnyFunctions = false;
1533 for(std::vector<Function*>::iterator I = LazilyLinkFunctions.begin(),
1534 E = LazilyLinkFunctions.end(); I != E; ++I) {
1535 Function *SF = *I;
1536 if (!SF)
1537 continue;
1539 Function *DF = cast<Function>(ValueMap[SF]);
1540 if (SF->hasPrefixData()) {
1541 // Link in the prefix data.
1542 DF->setPrefixData(MapValue(SF->getPrefixData(),
1543 ValueMap,
1544 RF_None,
1545 &TypeMap,
1546 &ValMaterializer));
1547 }
1549 // Materialize if needed.
1550 if (std::error_code EC = SF->materialize())
1551 return emitError(EC.message());
1553 // Skip if no body (function is external).
1554 if (SF->isDeclaration())
1555 continue;
1557 // Erase from vector *before* the function body is linked - linkFunctionBody could
1558 // invalidate I.
1559 LazilyLinkFunctions.erase(I);
1561 // Link in function body.
1562 linkFunctionBody(DF, SF);
1563 SF->Dematerialize();
1565 // Set flag to indicate we may have more functions to lazily link in
1566 // since we linked in a function.
1567 LinkedInAnyFunctions = true;
1568 break;
1569 }
1570 } while (LinkedInAnyFunctions);
1572 return false;
1573 }
1575 void Linker::init(Module *M, DiagnosticHandlerFunction DiagnosticHandler) {
1576 this->Composite = M;
1577 this->DiagnosticHandler = DiagnosticHandler;
1579 TypeFinder StructTypes;
1580 StructTypes.run(*M, true);
1581 IdentifiedStructTypes.insert(StructTypes.begin(), StructTypes.end());
1582 }
1584 Linker::Linker(Module *M, DiagnosticHandlerFunction DiagnosticHandler) {
1585 init(M, DiagnosticHandler);
1586 }
1588 Linker::Linker(Module *M) {
1589 init(M, [this](const DiagnosticInfo &DI) {
1590 Composite->getContext().diagnose(DI);
1591 });
1592 }
1594 Linker::~Linker() {
1595 }
1597 void Linker::deleteModule() {
1598 delete Composite;
1599 Composite = nullptr;
1600 }
1602 bool Linker::linkInModule(Module *Src) {
1603 ModuleLinker TheLinker(Composite, IdentifiedStructTypes, Src,
1604 DiagnosticHandler);
1605 return TheLinker.run();
1606 }
1608 //===----------------------------------------------------------------------===//
1609 // LinkModules entrypoint.
1610 //===----------------------------------------------------------------------===//
1612 /// This function links two modules together, with the resulting Dest module
1613 /// modified to be the composite of the two input modules. If an error occurs,
1614 /// true is returned and ErrorMsg (if not null) is set to indicate the problem.
1615 /// Upon failure, the Dest module could be in a modified state, and shouldn't be
1616 /// relied on to be consistent.
1617 bool Linker::LinkModules(Module *Dest, Module *Src,
1618 DiagnosticHandlerFunction DiagnosticHandler) {
1619 Linker L(Dest, DiagnosticHandler);
1620 return L.linkInModule(Src);
1621 }
1623 bool Linker::LinkModules(Module *Dest, Module *Src) {
1624 Linker L(Dest);
1625 return L.linkInModule(Src);
1626 }
1628 //===----------------------------------------------------------------------===//
1629 // C API.
1630 //===----------------------------------------------------------------------===//
1632 LLVMBool LLVMLinkModules(LLVMModuleRef Dest, LLVMModuleRef Src,
1633 LLVMLinkerMode Mode, char **OutMessages) {
1634 Module *D = unwrap(Dest);
1635 std::string Message;
1636 raw_string_ostream Stream(Message);
1637 DiagnosticPrinterRawOStream DP(Stream);
1639 LLVMBool Result = Linker::LinkModules(
1640 D, unwrap(Src), [&](const DiagnosticInfo &DI) { DI.print(DP); });
1642 if (OutMessages && Result)
1643 *OutMessages = strdup(Message.c_str());
1644 return Result;
1645 }