1 //===- lib/Linker/IRMover.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 "llvm/Linker/IRMover.h"
10 #include "LinkDiagnosticInfo.h"
11 #include "llvm/ADT/ScopeExit.h"
12 #include "llvm/ADT/SetVector.h"
13 #include "llvm/ADT/SmallPtrSet.h"
14 #include "llvm/ADT/SmallString.h"
15 #include "llvm/IR/AutoUpgrade.h"
16 #include "llvm/IR/Constants.h"
17 #include "llvm/IR/DebugInfoMetadata.h"
18 #include "llvm/IR/DiagnosticPrinter.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/GVMaterializer.h"
21 #include "llvm/IR/GlobalValue.h"
22 #include "llvm/IR/Instruction.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/Intrinsics.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/IR/PseudoProbe.h"
27 #include "llvm/IR/TypeFinder.h"
28 #include "llvm/Object/ModuleSymbolTable.h"
29 #include "llvm/Support/Error.h"
30 #include "llvm/Support/Path.h"
31 #include "llvm/TargetParser/Triple.h"
32 #include "llvm/Transforms/Utils/ValueMapper.h"
33 #include <optional>
34 #include <utility>
35 using namespace llvm;
36
37 //===----------------------------------------------------------------------===//
38 // TypeMap implementation.
39 //===----------------------------------------------------------------------===//
40
41 namespace {
42 class TypeMapTy : public ValueMapTypeRemapper {
43 /// This is a mapping from a source type to a destination type to use.
44 DenseMap<Type *, Type *> MappedTypes;
45
46 /// When checking to see if two subgraphs are isomorphic, we speculatively
47 /// add types to MappedTypes, but keep track of them here in case we need to
48 /// roll back.
49 SmallVector<Type *, 16> SpeculativeTypes;
50
51 SmallVector<StructType *, 16> SpeculativeDstOpaqueTypes;
52
53 /// This is a list of non-opaque structs in the source module that are mapped
54 /// to an opaque struct in the destination module.
55 SmallVector<StructType *, 16> SrcDefinitionsToResolve;
56
57 /// This is the set of opaque types in the destination modules who are
58 /// getting a body from the source module.
59 SmallPtrSet<StructType *, 16> DstResolvedOpaqueTypes;
60
61 public:
TypeMapTy(IRMover::IdentifiedStructTypeSet & DstStructTypesSet)62 TypeMapTy(IRMover::IdentifiedStructTypeSet &DstStructTypesSet)
63 : DstStructTypesSet(DstStructTypesSet) {}
64
65 IRMover::IdentifiedStructTypeSet &DstStructTypesSet;
66 /// Indicate that the specified type in the destination module is conceptually
67 /// equivalent to the specified type in the source module.
68 void addTypeMapping(Type *DstTy, Type *SrcTy);
69
70 /// Produce a body for an opaque type in the dest module from a type
71 /// definition in the source module.
72 void linkDefinedTypeBodies();
73
74 /// Return the mapped type to use for the specified input type from the
75 /// source module.
76 Type *get(Type *SrcTy);
77 Type *get(Type *SrcTy, SmallPtrSet<StructType *, 8> &Visited);
78
79 void finishType(StructType *DTy, StructType *STy, ArrayRef<Type *> ETypes);
80
get(FunctionType * T)81 FunctionType *get(FunctionType *T) {
82 return cast<FunctionType>(get((Type *)T));
83 }
84
85 private:
remapType(Type * SrcTy)86 Type *remapType(Type *SrcTy) override { return get(SrcTy); }
87
88 bool areTypesIsomorphic(Type *DstTy, Type *SrcTy);
89 };
90 }
91
addTypeMapping(Type * DstTy,Type * SrcTy)92 void TypeMapTy::addTypeMapping(Type *DstTy, Type *SrcTy) {
93 assert(SpeculativeTypes.empty());
94 assert(SpeculativeDstOpaqueTypes.empty());
95
96 // Check to see if these types are recursively isomorphic and establish a
97 // mapping between them if so.
98 if (!areTypesIsomorphic(DstTy, SrcTy)) {
99 // Oops, they aren't isomorphic. Just discard this request by rolling out
100 // any speculative mappings we've established.
101 for (Type *Ty : SpeculativeTypes)
102 MappedTypes.erase(Ty);
103
104 SrcDefinitionsToResolve.resize(SrcDefinitionsToResolve.size() -
105 SpeculativeDstOpaqueTypes.size());
106 for (StructType *Ty : SpeculativeDstOpaqueTypes)
107 DstResolvedOpaqueTypes.erase(Ty);
108 } else {
109 // SrcTy and DstTy are recursively ismorphic. We clear names of SrcTy
110 // and all its descendants to lower amount of renaming in LLVM context
111 // Renaming occurs because we load all source modules to the same context
112 // and declaration with existing name gets renamed (i.e Foo -> Foo.42).
113 // As a result we may get several different types in the destination
114 // module, which are in fact the same.
115 for (Type *Ty : SpeculativeTypes)
116 if (auto *STy = dyn_cast<StructType>(Ty))
117 if (STy->hasName())
118 STy->setName("");
119 }
120 SpeculativeTypes.clear();
121 SpeculativeDstOpaqueTypes.clear();
122 }
123
124 /// Recursively walk this pair of types, returning true if they are isomorphic,
125 /// false if they are not.
areTypesIsomorphic(Type * DstTy,Type * SrcTy)126 bool TypeMapTy::areTypesIsomorphic(Type *DstTy, Type *SrcTy) {
127 // Two types with differing kinds are clearly not isomorphic.
128 if (DstTy->getTypeID() != SrcTy->getTypeID())
129 return false;
130
131 // If we have an entry in the MappedTypes table, then we have our answer.
132 Type *&Entry = MappedTypes[SrcTy];
133 if (Entry)
134 return Entry == DstTy;
135
136 // Two identical types are clearly isomorphic. Remember this
137 // non-speculatively.
138 if (DstTy == SrcTy) {
139 Entry = DstTy;
140 return true;
141 }
142
143 // Okay, we have two types with identical kinds that we haven't seen before.
144
145 // If this is an opaque struct type, special case it.
146 if (StructType *SSTy = dyn_cast<StructType>(SrcTy)) {
147 // Mapping an opaque type to any struct, just keep the dest struct.
148 if (SSTy->isOpaque()) {
149 Entry = DstTy;
150 SpeculativeTypes.push_back(SrcTy);
151 return true;
152 }
153
154 // Mapping a non-opaque source type to an opaque dest. If this is the first
155 // type that we're mapping onto this destination type then we succeed. Keep
156 // the dest, but fill it in later. If this is the second (different) type
157 // that we're trying to map onto the same opaque type then we fail.
158 if (cast<StructType>(DstTy)->isOpaque()) {
159 // We can only map one source type onto the opaque destination type.
160 if (!DstResolvedOpaqueTypes.insert(cast<StructType>(DstTy)).second)
161 return false;
162 SrcDefinitionsToResolve.push_back(SSTy);
163 SpeculativeTypes.push_back(SrcTy);
164 SpeculativeDstOpaqueTypes.push_back(cast<StructType>(DstTy));
165 Entry = DstTy;
166 return true;
167 }
168 }
169
170 // If the number of subtypes disagree between the two types, then we fail.
171 if (SrcTy->getNumContainedTypes() != DstTy->getNumContainedTypes())
172 return false;
173
174 // Fail if any of the extra properties (e.g. array size) of the type disagree.
175 if (isa<IntegerType>(DstTy))
176 return false; // bitwidth disagrees.
177 if (PointerType *PT = dyn_cast<PointerType>(DstTy)) {
178 if (PT->getAddressSpace() != cast<PointerType>(SrcTy)->getAddressSpace())
179 return false;
180 } else if (FunctionType *FT = dyn_cast<FunctionType>(DstTy)) {
181 if (FT->isVarArg() != cast<FunctionType>(SrcTy)->isVarArg())
182 return false;
183 } else if (StructType *DSTy = dyn_cast<StructType>(DstTy)) {
184 StructType *SSTy = cast<StructType>(SrcTy);
185 if (DSTy->isLiteral() != SSTy->isLiteral() ||
186 DSTy->isPacked() != SSTy->isPacked())
187 return false;
188 } else if (auto *DArrTy = dyn_cast<ArrayType>(DstTy)) {
189 if (DArrTy->getNumElements() != cast<ArrayType>(SrcTy)->getNumElements())
190 return false;
191 } else if (auto *DVecTy = dyn_cast<VectorType>(DstTy)) {
192 if (DVecTy->getElementCount() != cast<VectorType>(SrcTy)->getElementCount())
193 return false;
194 }
195
196 // Otherwise, we speculate that these two types will line up and recursively
197 // check the subelements.
198 Entry = DstTy;
199 SpeculativeTypes.push_back(SrcTy);
200
201 for (unsigned I = 0, E = SrcTy->getNumContainedTypes(); I != E; ++I)
202 if (!areTypesIsomorphic(DstTy->getContainedType(I),
203 SrcTy->getContainedType(I)))
204 return false;
205
206 // If everything seems to have lined up, then everything is great.
207 return true;
208 }
209
linkDefinedTypeBodies()210 void TypeMapTy::linkDefinedTypeBodies() {
211 SmallVector<Type *, 16> Elements;
212 for (StructType *SrcSTy : SrcDefinitionsToResolve) {
213 StructType *DstSTy = cast<StructType>(MappedTypes[SrcSTy]);
214 assert(DstSTy->isOpaque());
215
216 // Map the body of the source type over to a new body for the dest type.
217 Elements.resize(SrcSTy->getNumElements());
218 for (unsigned I = 0, E = Elements.size(); I != E; ++I)
219 Elements[I] = get(SrcSTy->getElementType(I));
220
221 DstSTy->setBody(Elements, SrcSTy->isPacked());
222 DstStructTypesSet.switchToNonOpaque(DstSTy);
223 }
224 SrcDefinitionsToResolve.clear();
225 DstResolvedOpaqueTypes.clear();
226 }
227
finishType(StructType * DTy,StructType * STy,ArrayRef<Type * > ETypes)228 void TypeMapTy::finishType(StructType *DTy, StructType *STy,
229 ArrayRef<Type *> ETypes) {
230 DTy->setBody(ETypes, STy->isPacked());
231
232 // Steal STy's name.
233 if (STy->hasName()) {
234 SmallString<16> TmpName = STy->getName();
235 STy->setName("");
236 DTy->setName(TmpName);
237 }
238
239 DstStructTypesSet.addNonOpaque(DTy);
240 }
241
get(Type * Ty)242 Type *TypeMapTy::get(Type *Ty) {
243 SmallPtrSet<StructType *, 8> Visited;
244 return get(Ty, Visited);
245 }
246
get(Type * Ty,SmallPtrSet<StructType *,8> & Visited)247 Type *TypeMapTy::get(Type *Ty, SmallPtrSet<StructType *, 8> &Visited) {
248 // If we already have an entry for this type, return it.
249 Type **Entry = &MappedTypes[Ty];
250 if (*Entry)
251 return *Entry;
252
253 // These are types that LLVM itself will unique.
254 bool IsUniqued = !isa<StructType>(Ty) || cast<StructType>(Ty)->isLiteral();
255
256 if (!IsUniqued) {
257 #ifndef NDEBUG
258 for (auto &Pair : MappedTypes) {
259 assert(!(Pair.first != Ty && Pair.second == Ty) &&
260 "mapping to a source type");
261 }
262 #endif
263
264 if (!Visited.insert(cast<StructType>(Ty)).second) {
265 StructType *DTy = StructType::create(Ty->getContext());
266 return *Entry = DTy;
267 }
268 }
269
270 // If this is not a recursive type, then just map all of the elements and
271 // then rebuild the type from inside out.
272 SmallVector<Type *, 4> ElementTypes;
273
274 // If there are no element types to map, then the type is itself. This is
275 // true for the anonymous {} struct, things like 'float', integers, etc.
276 if (Ty->getNumContainedTypes() == 0 && IsUniqued)
277 return *Entry = Ty;
278
279 // Remap all of the elements, keeping track of whether any of them change.
280 bool AnyChange = false;
281 ElementTypes.resize(Ty->getNumContainedTypes());
282 for (unsigned I = 0, E = Ty->getNumContainedTypes(); I != E; ++I) {
283 ElementTypes[I] = get(Ty->getContainedType(I), Visited);
284 AnyChange |= ElementTypes[I] != Ty->getContainedType(I);
285 }
286
287 // If we found our type while recursively processing stuff, just use it.
288 Entry = &MappedTypes[Ty];
289 if (*Entry) {
290 if (auto *DTy = dyn_cast<StructType>(*Entry)) {
291 if (DTy->isOpaque()) {
292 auto *STy = cast<StructType>(Ty);
293 finishType(DTy, STy, ElementTypes);
294 }
295 }
296 return *Entry;
297 }
298
299 // If all of the element types mapped directly over and the type is not
300 // a named struct, then the type is usable as-is.
301 if (!AnyChange && IsUniqued)
302 return *Entry = Ty;
303
304 // Otherwise, rebuild a modified type.
305 switch (Ty->getTypeID()) {
306 default:
307 llvm_unreachable("unknown derived type to remap");
308 case Type::ArrayTyID:
309 return *Entry = ArrayType::get(ElementTypes[0],
310 cast<ArrayType>(Ty)->getNumElements());
311 case Type::ScalableVectorTyID:
312 case Type::FixedVectorTyID:
313 return *Entry = VectorType::get(ElementTypes[0],
314 cast<VectorType>(Ty)->getElementCount());
315 case Type::PointerTyID:
316 return *Entry = PointerType::get(ElementTypes[0],
317 cast<PointerType>(Ty)->getAddressSpace());
318 case Type::FunctionTyID:
319 return *Entry = FunctionType::get(ElementTypes[0],
320 ArrayRef(ElementTypes).slice(1),
321 cast<FunctionType>(Ty)->isVarArg());
322 case Type::StructTyID: {
323 auto *STy = cast<StructType>(Ty);
324 bool IsPacked = STy->isPacked();
325 if (IsUniqued)
326 return *Entry = StructType::get(Ty->getContext(), ElementTypes, IsPacked);
327
328 // If the type is opaque, we can just use it directly.
329 if (STy->isOpaque()) {
330 DstStructTypesSet.addOpaque(STy);
331 return *Entry = Ty;
332 }
333
334 if (StructType *OldT =
335 DstStructTypesSet.findNonOpaque(ElementTypes, IsPacked)) {
336 STy->setName("");
337 return *Entry = OldT;
338 }
339
340 if (!AnyChange) {
341 DstStructTypesSet.addNonOpaque(STy);
342 return *Entry = Ty;
343 }
344
345 StructType *DTy = StructType::create(Ty->getContext());
346 finishType(DTy, STy, ElementTypes);
347 return *Entry = DTy;
348 }
349 }
350 }
351
LinkDiagnosticInfo(DiagnosticSeverity Severity,const Twine & Msg)352 LinkDiagnosticInfo::LinkDiagnosticInfo(DiagnosticSeverity Severity,
353 const Twine &Msg)
354 : DiagnosticInfo(DK_Linker, Severity), Msg(Msg) {}
print(DiagnosticPrinter & DP) const355 void LinkDiagnosticInfo::print(DiagnosticPrinter &DP) const { DP << Msg; }
356
357 //===----------------------------------------------------------------------===//
358 // IRLinker implementation.
359 //===----------------------------------------------------------------------===//
360
361 namespace {
362 class IRLinker;
363
364 /// Creates prototypes for functions that are lazily linked on the fly. This
365 /// speeds up linking for modules with many/ lazily linked functions of which
366 /// few get used.
367 class GlobalValueMaterializer final : public ValueMaterializer {
368 IRLinker &TheIRLinker;
369
370 public:
GlobalValueMaterializer(IRLinker & TheIRLinker)371 GlobalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {}
372 Value *materialize(Value *V) override;
373 };
374
375 class LocalValueMaterializer final : public ValueMaterializer {
376 IRLinker &TheIRLinker;
377
378 public:
LocalValueMaterializer(IRLinker & TheIRLinker)379 LocalValueMaterializer(IRLinker &TheIRLinker) : TheIRLinker(TheIRLinker) {}
380 Value *materialize(Value *V) override;
381 };
382
383 /// Type of the Metadata map in \a ValueToValueMapTy.
384 typedef DenseMap<const Metadata *, TrackingMDRef> MDMapT;
385
386 /// This is responsible for keeping track of the state used for moving data
387 /// from SrcM to DstM.
388 class IRLinker {
389 Module &DstM;
390 std::unique_ptr<Module> SrcM;
391
392 /// See IRMover::move().
393 IRMover::LazyCallback AddLazyFor;
394
395 TypeMapTy TypeMap;
396 GlobalValueMaterializer GValMaterializer;
397 LocalValueMaterializer LValMaterializer;
398
399 /// A metadata map that's shared between IRLinker instances.
400 MDMapT &SharedMDs;
401
402 /// Mapping of values from what they used to be in Src, to what they are now
403 /// in DstM. ValueToValueMapTy is a ValueMap, which involves some overhead
404 /// due to the use of Value handles which the Linker doesn't actually need,
405 /// but this allows us to reuse the ValueMapper code.
406 ValueToValueMapTy ValueMap;
407 ValueToValueMapTy IndirectSymbolValueMap;
408
409 DenseSet<GlobalValue *> ValuesToLink;
410 std::vector<GlobalValue *> Worklist;
411 std::vector<std::pair<GlobalValue *, Value*>> RAUWWorklist;
412
413 /// Set of globals with eagerly copied metadata that may require remapping.
414 /// This remapping is performed after metadata linking.
415 DenseSet<GlobalObject *> UnmappedMetadata;
416
maybeAdd(GlobalValue * GV)417 void maybeAdd(GlobalValue *GV) {
418 if (ValuesToLink.insert(GV).second)
419 Worklist.push_back(GV);
420 }
421
422 /// Whether we are importing globals for ThinLTO, as opposed to linking the
423 /// source module. If this flag is set, it means that we can rely on some
424 /// other object file to define any non-GlobalValue entities defined by the
425 /// source module. This currently causes us to not link retained types in
426 /// debug info metadata and module inline asm.
427 bool IsPerformingImport;
428
429 /// Set to true when all global value body linking is complete (including
430 /// lazy linking). Used to prevent metadata linking from creating new
431 /// references.
432 bool DoneLinkingBodies = false;
433
434 /// The Error encountered during materialization. We use an Optional here to
435 /// avoid needing to manage an unconsumed success value.
436 std::optional<Error> FoundError;
setError(Error E)437 void setError(Error E) {
438 if (E)
439 FoundError = std::move(E);
440 }
441
442 /// Most of the errors produced by this module are inconvertible StringErrors.
443 /// This convenience function lets us return one of those more easily.
stringErr(const Twine & T)444 Error stringErr(const Twine &T) {
445 return make_error<StringError>(T, inconvertibleErrorCode());
446 }
447
448 /// Entry point for mapping values and alternate context for mapping aliases.
449 ValueMapper Mapper;
450 unsigned IndirectSymbolMCID;
451
452 /// Handles cloning of a global values from the source module into
453 /// the destination module, including setting the attributes and visibility.
454 GlobalValue *copyGlobalValueProto(const GlobalValue *SGV, bool ForDefinition);
455
emitWarning(const Twine & Message)456 void emitWarning(const Twine &Message) {
457 SrcM->getContext().diagnose(LinkDiagnosticInfo(DS_Warning, Message));
458 }
459
460 /// Given a global in the source module, return the global in the
461 /// destination module that is being linked to, if any.
getLinkedToGlobal(const GlobalValue * SrcGV)462 GlobalValue *getLinkedToGlobal(const GlobalValue *SrcGV) {
463 // If the source has no name it can't link. If it has local linkage,
464 // there is no name match-up going on.
465 if (!SrcGV->hasName() || SrcGV->hasLocalLinkage())
466 return nullptr;
467
468 // Otherwise see if we have a match in the destination module's symtab.
469 GlobalValue *DGV = DstM.getNamedValue(SrcGV->getName());
470 if (!DGV)
471 return nullptr;
472
473 // If we found a global with the same name in the dest module, but it has
474 // internal linkage, we are really not doing any linkage here.
475 if (DGV->hasLocalLinkage())
476 return nullptr;
477
478 // If we found an intrinsic declaration with mismatching prototypes, we
479 // probably had a nameclash. Don't use that version.
480 if (auto *FDGV = dyn_cast<Function>(DGV))
481 if (FDGV->isIntrinsic())
482 if (const auto *FSrcGV = dyn_cast<Function>(SrcGV))
483 if (FDGV->getFunctionType() != TypeMap.get(FSrcGV->getFunctionType()))
484 return nullptr;
485
486 // Otherwise, we do in fact link to the destination global.
487 return DGV;
488 }
489
490 void computeTypeMapping();
491
492 Expected<Constant *> linkAppendingVarProto(GlobalVariable *DstGV,
493 const GlobalVariable *SrcGV);
494
495 /// Given the GlobaValue \p SGV in the source module, and the matching
496 /// GlobalValue \p DGV (if any), return true if the linker will pull \p SGV
497 /// into the destination module.
498 ///
499 /// Note this code may call the client-provided \p AddLazyFor.
500 bool shouldLink(GlobalValue *DGV, GlobalValue &SGV);
501 Expected<Constant *> linkGlobalValueProto(GlobalValue *GV,
502 bool ForIndirectSymbol);
503
504 Error linkModuleFlagsMetadata();
505
506 void linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src);
507 Error linkFunctionBody(Function &Dst, Function &Src);
508 void linkAliasAliasee(GlobalAlias &Dst, GlobalAlias &Src);
509 void linkIFuncResolver(GlobalIFunc &Dst, GlobalIFunc &Src);
510 Error linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src);
511
512 /// Replace all types in the source AttributeList with the
513 /// corresponding destination type.
514 AttributeList mapAttributeTypes(LLVMContext &C, AttributeList Attrs);
515
516 /// Functions that take care of cloning a specific global value type
517 /// into the destination module.
518 GlobalVariable *copyGlobalVariableProto(const GlobalVariable *SGVar);
519 Function *copyFunctionProto(const Function *SF);
520 GlobalValue *copyIndirectSymbolProto(const GlobalValue *SGV);
521
522 /// Perform "replace all uses with" operations. These work items need to be
523 /// performed as part of materialization, but we postpone them to happen after
524 /// materialization is done. The materializer called by ValueMapper is not
525 /// expected to delete constants, as ValueMapper is holding pointers to some
526 /// of them, but constant destruction may be indirectly triggered by RAUW.
527 /// Hence, the need to move this out of the materialization call chain.
528 void flushRAUWWorklist();
529
530 /// When importing for ThinLTO, prevent importing of types listed on
531 /// the DICompileUnit that we don't need a copy of in the importing
532 /// module.
533 void prepareCompileUnitsForImport();
534 void linkNamedMDNodes();
535
536 /// Update attributes while linking.
537 void updateAttributes(GlobalValue &GV);
538
539 public:
IRLinker(Module & DstM,MDMapT & SharedMDs,IRMover::IdentifiedStructTypeSet & Set,std::unique_ptr<Module> SrcM,ArrayRef<GlobalValue * > ValuesToLink,IRMover::LazyCallback AddLazyFor,bool IsPerformingImport)540 IRLinker(Module &DstM, MDMapT &SharedMDs,
541 IRMover::IdentifiedStructTypeSet &Set, std::unique_ptr<Module> SrcM,
542 ArrayRef<GlobalValue *> ValuesToLink,
543 IRMover::LazyCallback AddLazyFor, bool IsPerformingImport)
544 : DstM(DstM), SrcM(std::move(SrcM)), AddLazyFor(std::move(AddLazyFor)),
545 TypeMap(Set), GValMaterializer(*this), LValMaterializer(*this),
546 SharedMDs(SharedMDs), IsPerformingImport(IsPerformingImport),
547 Mapper(ValueMap, RF_ReuseAndMutateDistinctMDs | RF_IgnoreMissingLocals,
548 &TypeMap, &GValMaterializer),
549 IndirectSymbolMCID(Mapper.registerAlternateMappingContext(
550 IndirectSymbolValueMap, &LValMaterializer)) {
551 ValueMap.getMDMap() = std::move(SharedMDs);
552 for (GlobalValue *GV : ValuesToLink)
553 maybeAdd(GV);
554 if (IsPerformingImport)
555 prepareCompileUnitsForImport();
556 }
~IRLinker()557 ~IRLinker() { SharedMDs = std::move(*ValueMap.getMDMap()); }
558
559 Error run();
560 Value *materialize(Value *V, bool ForIndirectSymbol);
561 };
562 }
563
564 /// The LLVM SymbolTable class autorenames globals that conflict in the symbol
565 /// table. This is good for all clients except for us. Go through the trouble
566 /// to force this back.
forceRenaming(GlobalValue * GV,StringRef Name)567 static void forceRenaming(GlobalValue *GV, StringRef Name) {
568 // If the global doesn't force its name or if it already has the right name,
569 // there is nothing for us to do.
570 if (GV->hasLocalLinkage() || GV->getName() == Name)
571 return;
572
573 Module *M = GV->getParent();
574
575 // If there is a conflict, rename the conflict.
576 if (GlobalValue *ConflictGV = M->getNamedValue(Name)) {
577 GV->takeName(ConflictGV);
578 ConflictGV->setName(Name); // This will cause ConflictGV to get renamed
579 assert(ConflictGV->getName() != Name && "forceRenaming didn't work");
580 } else {
581 GV->setName(Name); // Force the name back
582 }
583 }
584
materialize(Value * SGV)585 Value *GlobalValueMaterializer::materialize(Value *SGV) {
586 return TheIRLinker.materialize(SGV, false);
587 }
588
materialize(Value * SGV)589 Value *LocalValueMaterializer::materialize(Value *SGV) {
590 return TheIRLinker.materialize(SGV, true);
591 }
592
materialize(Value * V,bool ForIndirectSymbol)593 Value *IRLinker::materialize(Value *V, bool ForIndirectSymbol) {
594 auto *SGV = dyn_cast<GlobalValue>(V);
595 if (!SGV)
596 return nullptr;
597
598 // When linking a global from other modules than source & dest, skip
599 // materializing it because it would be mapped later when its containing
600 // module is linked. Linking it now would potentially pull in many types that
601 // may not be mapped properly.
602 if (SGV->getParent() != &DstM && SGV->getParent() != SrcM.get())
603 return nullptr;
604
605 Expected<Constant *> NewProto = linkGlobalValueProto(SGV, ForIndirectSymbol);
606 if (!NewProto) {
607 setError(NewProto.takeError());
608 return nullptr;
609 }
610 if (!*NewProto)
611 return nullptr;
612
613 GlobalValue *New = dyn_cast<GlobalValue>(*NewProto);
614 if (!New)
615 return *NewProto;
616
617 // If we already created the body, just return.
618 if (auto *F = dyn_cast<Function>(New)) {
619 if (!F->isDeclaration())
620 return New;
621 } else if (auto *V = dyn_cast<GlobalVariable>(New)) {
622 if (V->hasInitializer() || V->hasAppendingLinkage())
623 return New;
624 } else if (auto *GA = dyn_cast<GlobalAlias>(New)) {
625 if (GA->getAliasee())
626 return New;
627 } else if (auto *GI = dyn_cast<GlobalIFunc>(New)) {
628 if (GI->getResolver())
629 return New;
630 } else {
631 llvm_unreachable("Invalid GlobalValue type");
632 }
633
634 // If the global is being linked for an indirect symbol, it may have already
635 // been scheduled to satisfy a regular symbol. Similarly, a global being linked
636 // for a regular symbol may have already been scheduled for an indirect
637 // symbol. Check for these cases by looking in the other value map and
638 // confirming the same value has been scheduled. If there is an entry in the
639 // ValueMap but the value is different, it means that the value already had a
640 // definition in the destination module (linkonce for instance), but we need a
641 // new definition for the indirect symbol ("New" will be different).
642 if ((ForIndirectSymbol && ValueMap.lookup(SGV) == New) ||
643 (!ForIndirectSymbol && IndirectSymbolValueMap.lookup(SGV) == New))
644 return New;
645
646 if (ForIndirectSymbol || shouldLink(New, *SGV))
647 setError(linkGlobalValueBody(*New, *SGV));
648
649 updateAttributes(*New);
650 return New;
651 }
652
653 /// Loop through the global variables in the src module and merge them into the
654 /// dest module.
copyGlobalVariableProto(const GlobalVariable * SGVar)655 GlobalVariable *IRLinker::copyGlobalVariableProto(const GlobalVariable *SGVar) {
656 // No linking to be performed or linking from the source: simply create an
657 // identical version of the symbol over in the dest module... the
658 // initializer will be filled in later by LinkGlobalInits.
659 GlobalVariable *NewDGV =
660 new GlobalVariable(DstM, TypeMap.get(SGVar->getValueType()),
661 SGVar->isConstant(), GlobalValue::ExternalLinkage,
662 /*init*/ nullptr, SGVar->getName(),
663 /*insertbefore*/ nullptr, SGVar->getThreadLocalMode(),
664 SGVar->getAddressSpace());
665 NewDGV->setAlignment(SGVar->getAlign());
666 NewDGV->copyAttributesFrom(SGVar);
667 return NewDGV;
668 }
669
mapAttributeTypes(LLVMContext & C,AttributeList Attrs)670 AttributeList IRLinker::mapAttributeTypes(LLVMContext &C, AttributeList Attrs) {
671 for (unsigned i = 0; i < Attrs.getNumAttrSets(); ++i) {
672 for (int AttrIdx = Attribute::FirstTypeAttr;
673 AttrIdx <= Attribute::LastTypeAttr; AttrIdx++) {
674 Attribute::AttrKind TypedAttr = (Attribute::AttrKind)AttrIdx;
675 if (Attrs.hasAttributeAtIndex(i, TypedAttr)) {
676 if (Type *Ty =
677 Attrs.getAttributeAtIndex(i, TypedAttr).getValueAsType()) {
678 Attrs = Attrs.replaceAttributeTypeAtIndex(C, i, TypedAttr,
679 TypeMap.get(Ty));
680 break;
681 }
682 }
683 }
684 }
685 return Attrs;
686 }
687
688 /// Link the function in the source module into the destination module if
689 /// needed, setting up mapping information.
copyFunctionProto(const Function * SF)690 Function *IRLinker::copyFunctionProto(const Function *SF) {
691 // If there is no linkage to be performed or we are linking from the source,
692 // bring SF over.
693 auto *F = Function::Create(TypeMap.get(SF->getFunctionType()),
694 GlobalValue::ExternalLinkage,
695 SF->getAddressSpace(), SF->getName(), &DstM);
696 F->copyAttributesFrom(SF);
697 F->setAttributes(mapAttributeTypes(F->getContext(), F->getAttributes()));
698 F->IsNewDbgInfoFormat = SF->IsNewDbgInfoFormat;
699 return F;
700 }
701
702 /// Set up prototypes for any indirect symbols that come over from the source
703 /// module.
copyIndirectSymbolProto(const GlobalValue * SGV)704 GlobalValue *IRLinker::copyIndirectSymbolProto(const GlobalValue *SGV) {
705 // If there is no linkage to be performed or we're linking from the source,
706 // bring over SGA.
707 auto *Ty = TypeMap.get(SGV->getValueType());
708
709 if (auto *GA = dyn_cast<GlobalAlias>(SGV)) {
710 auto *DGA = GlobalAlias::create(Ty, SGV->getAddressSpace(),
711 GlobalValue::ExternalLinkage,
712 SGV->getName(), &DstM);
713 DGA->copyAttributesFrom(GA);
714 return DGA;
715 }
716
717 if (auto *GI = dyn_cast<GlobalIFunc>(SGV)) {
718 auto *DGI = GlobalIFunc::create(Ty, SGV->getAddressSpace(),
719 GlobalValue::ExternalLinkage,
720 SGV->getName(), nullptr, &DstM);
721 DGI->copyAttributesFrom(GI);
722 return DGI;
723 }
724
725 llvm_unreachable("Invalid source global value type");
726 }
727
copyGlobalValueProto(const GlobalValue * SGV,bool ForDefinition)728 GlobalValue *IRLinker::copyGlobalValueProto(const GlobalValue *SGV,
729 bool ForDefinition) {
730 GlobalValue *NewGV;
731 if (auto *SGVar = dyn_cast<GlobalVariable>(SGV)) {
732 NewGV = copyGlobalVariableProto(SGVar);
733 } else if (auto *SF = dyn_cast<Function>(SGV)) {
734 NewGV = copyFunctionProto(SF);
735 } else {
736 if (ForDefinition)
737 NewGV = copyIndirectSymbolProto(SGV);
738 else if (SGV->getValueType()->isFunctionTy())
739 NewGV =
740 Function::Create(cast<FunctionType>(TypeMap.get(SGV->getValueType())),
741 GlobalValue::ExternalLinkage, SGV->getAddressSpace(),
742 SGV->getName(), &DstM);
743 else
744 NewGV =
745 new GlobalVariable(DstM, TypeMap.get(SGV->getValueType()),
746 /*isConstant*/ false, GlobalValue::ExternalLinkage,
747 /*init*/ nullptr, SGV->getName(),
748 /*insertbefore*/ nullptr,
749 SGV->getThreadLocalMode(), SGV->getAddressSpace());
750 }
751
752 if (ForDefinition)
753 NewGV->setLinkage(SGV->getLinkage());
754 else if (SGV->hasExternalWeakLinkage())
755 NewGV->setLinkage(GlobalValue::ExternalWeakLinkage);
756
757 if (auto *NewGO = dyn_cast<GlobalObject>(NewGV)) {
758 // Metadata for global variables and function declarations is copied eagerly.
759 if (isa<GlobalVariable>(SGV) || SGV->isDeclaration()) {
760 NewGO->copyMetadata(cast<GlobalObject>(SGV), 0);
761 if (SGV->isDeclaration() && NewGO->hasMetadata())
762 UnmappedMetadata.insert(NewGO);
763 }
764 }
765
766 // Remove these copied constants in case this stays a declaration, since
767 // they point to the source module. If the def is linked the values will
768 // be mapped in during linkFunctionBody.
769 if (auto *NewF = dyn_cast<Function>(NewGV)) {
770 NewF->setPersonalityFn(nullptr);
771 NewF->setPrefixData(nullptr);
772 NewF->setPrologueData(nullptr);
773 }
774
775 return NewGV;
776 }
777
getTypeNamePrefix(StringRef Name)778 static StringRef getTypeNamePrefix(StringRef Name) {
779 size_t DotPos = Name.rfind('.');
780 return (DotPos == 0 || DotPos == StringRef::npos || Name.back() == '.' ||
781 !isdigit(static_cast<unsigned char>(Name[DotPos + 1])))
782 ? Name
783 : Name.substr(0, DotPos);
784 }
785
786 /// Loop over all of the linked values to compute type mappings. For example,
787 /// if we link "extern Foo *x" and "Foo *x = NULL", then we have two struct
788 /// types 'Foo' but one got renamed when the module was loaded into the same
789 /// LLVMContext.
computeTypeMapping()790 void IRLinker::computeTypeMapping() {
791 for (GlobalValue &SGV : SrcM->globals()) {
792 GlobalValue *DGV = getLinkedToGlobal(&SGV);
793 if (!DGV)
794 continue;
795
796 if (!DGV->hasAppendingLinkage() || !SGV.hasAppendingLinkage()) {
797 TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
798 continue;
799 }
800
801 // Unify the element type of appending arrays.
802 ArrayType *DAT = cast<ArrayType>(DGV->getValueType());
803 ArrayType *SAT = cast<ArrayType>(SGV.getValueType());
804 TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
805 }
806
807 for (GlobalValue &SGV : *SrcM)
808 if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) {
809 if (DGV->getType() == SGV.getType()) {
810 // If the types of DGV and SGV are the same, it means that DGV is from
811 // the source module and got added to DstM from a shared metadata. We
812 // shouldn't map this type to itself in case the type's components get
813 // remapped to a new type from DstM (for instance, during the loop over
814 // SrcM->getIdentifiedStructTypes() below).
815 continue;
816 }
817
818 TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
819 }
820
821 for (GlobalValue &SGV : SrcM->aliases())
822 if (GlobalValue *DGV = getLinkedToGlobal(&SGV))
823 TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
824
825 // Incorporate types by name, scanning all the types in the source module.
826 // At this point, the destination module may have a type "%foo = { i32 }" for
827 // example. When the source module got loaded into the same LLVMContext, if
828 // it had the same type, it would have been renamed to "%foo.42 = { i32 }".
829 std::vector<StructType *> Types = SrcM->getIdentifiedStructTypes();
830 for (StructType *ST : Types) {
831 if (!ST->hasName())
832 continue;
833
834 if (TypeMap.DstStructTypesSet.hasType(ST)) {
835 // This is actually a type from the destination module.
836 // getIdentifiedStructTypes() can have found it by walking debug info
837 // metadata nodes, some of which get linked by name when ODR Type Uniquing
838 // is enabled on the Context, from the source to the destination module.
839 continue;
840 }
841
842 auto STTypePrefix = getTypeNamePrefix(ST->getName());
843 if (STTypePrefix.size() == ST->getName().size())
844 continue;
845
846 // Check to see if the destination module has a struct with the prefix name.
847 StructType *DST = StructType::getTypeByName(ST->getContext(), STTypePrefix);
848 if (!DST)
849 continue;
850
851 // Don't use it if this actually came from the source module. They're in
852 // the same LLVMContext after all. Also don't use it unless the type is
853 // actually used in the destination module. This can happen in situations
854 // like this:
855 //
856 // Module A Module B
857 // -------- --------
858 // %Z = type { %A } %B = type { %C.1 }
859 // %A = type { %B.1, [7 x i8] } %C.1 = type { i8* }
860 // %B.1 = type { %C } %A.2 = type { %B.3, [5 x i8] }
861 // %C = type { i8* } %B.3 = type { %C.1 }
862 //
863 // When we link Module B with Module A, the '%B' in Module B is
864 // used. However, that would then use '%C.1'. But when we process '%C.1',
865 // we prefer to take the '%C' version. So we are then left with both
866 // '%C.1' and '%C' being used for the same types. This leads to some
867 // variables using one type and some using the other.
868 if (TypeMap.DstStructTypesSet.hasType(DST))
869 TypeMap.addTypeMapping(DST, ST);
870 }
871
872 // Now that we have discovered all of the type equivalences, get a body for
873 // any 'opaque' types in the dest module that are now resolved.
874 TypeMap.linkDefinedTypeBodies();
875 }
876
getArrayElements(const Constant * C,SmallVectorImpl<Constant * > & Dest)877 static void getArrayElements(const Constant *C,
878 SmallVectorImpl<Constant *> &Dest) {
879 unsigned NumElements = cast<ArrayType>(C->getType())->getNumElements();
880
881 for (unsigned i = 0; i != NumElements; ++i)
882 Dest.push_back(C->getAggregateElement(i));
883 }
884
885 /// If there were any appending global variables, link them together now.
886 Expected<Constant *>
linkAppendingVarProto(GlobalVariable * DstGV,const GlobalVariable * SrcGV)887 IRLinker::linkAppendingVarProto(GlobalVariable *DstGV,
888 const GlobalVariable *SrcGV) {
889 // Check that both variables have compatible properties.
890 if (DstGV && !DstGV->isDeclaration() && !SrcGV->isDeclaration()) {
891 if (!SrcGV->hasAppendingLinkage() || !DstGV->hasAppendingLinkage())
892 return stringErr(
893 "Linking globals named '" + SrcGV->getName() +
894 "': can only link appending global with another appending "
895 "global!");
896
897 if (DstGV->isConstant() != SrcGV->isConstant())
898 return stringErr("Appending variables linked with different const'ness!");
899
900 if (DstGV->getAlign() != SrcGV->getAlign())
901 return stringErr(
902 "Appending variables with different alignment need to be linked!");
903
904 if (DstGV->getVisibility() != SrcGV->getVisibility())
905 return stringErr(
906 "Appending variables with different visibility need to be linked!");
907
908 if (DstGV->hasGlobalUnnamedAddr() != SrcGV->hasGlobalUnnamedAddr())
909 return stringErr(
910 "Appending variables with different unnamed_addr need to be linked!");
911
912 if (DstGV->getSection() != SrcGV->getSection())
913 return stringErr(
914 "Appending variables with different section name need to be linked!");
915
916 if (DstGV->getAddressSpace() != SrcGV->getAddressSpace())
917 return stringErr("Appending variables with different address spaces need "
918 "to be linked!");
919 }
920
921 // Do not need to do anything if source is a declaration.
922 if (SrcGV->isDeclaration())
923 return DstGV;
924
925 Type *EltTy = cast<ArrayType>(TypeMap.get(SrcGV->getValueType()))
926 ->getElementType();
927
928 // FIXME: This upgrade is done during linking to support the C API. Once the
929 // old form is deprecated, we should move this upgrade to
930 // llvm::UpgradeGlobalVariable() and simplify the logic here and in
931 // Mapper::mapAppendingVariable() in ValueMapper.cpp.
932 StringRef Name = SrcGV->getName();
933 bool IsNewStructor = false;
934 bool IsOldStructor = false;
935 if (Name == "llvm.global_ctors" || Name == "llvm.global_dtors") {
936 if (cast<StructType>(EltTy)->getNumElements() == 3)
937 IsNewStructor = true;
938 else
939 IsOldStructor = true;
940 }
941
942 PointerType *VoidPtrTy = PointerType::get(SrcGV->getContext(), 0);
943 if (IsOldStructor) {
944 auto &ST = *cast<StructType>(EltTy);
945 Type *Tys[3] = {ST.getElementType(0), ST.getElementType(1), VoidPtrTy};
946 EltTy = StructType::get(SrcGV->getContext(), Tys, false);
947 }
948
949 uint64_t DstNumElements = 0;
950 if (DstGV && !DstGV->isDeclaration()) {
951 ArrayType *DstTy = cast<ArrayType>(DstGV->getValueType());
952 DstNumElements = DstTy->getNumElements();
953
954 // Check to see that they two arrays agree on type.
955 if (EltTy != DstTy->getElementType())
956 return stringErr("Appending variables with different element types!");
957 }
958
959 SmallVector<Constant *, 16> SrcElements;
960 getArrayElements(SrcGV->getInitializer(), SrcElements);
961
962 if (IsNewStructor) {
963 erase_if(SrcElements, [this](Constant *E) {
964 auto *Key =
965 dyn_cast<GlobalValue>(E->getAggregateElement(2)->stripPointerCasts());
966 if (!Key)
967 return false;
968 GlobalValue *DGV = getLinkedToGlobal(Key);
969 return !shouldLink(DGV, *Key);
970 });
971 }
972 uint64_t NewSize = DstNumElements + SrcElements.size();
973 ArrayType *NewType = ArrayType::get(EltTy, NewSize);
974
975 // Create the new global variable.
976 GlobalVariable *NG = new GlobalVariable(
977 DstM, NewType, SrcGV->isConstant(), SrcGV->getLinkage(),
978 /*init*/ nullptr, /*name*/ "", DstGV, SrcGV->getThreadLocalMode(),
979 SrcGV->getAddressSpace());
980
981 NG->copyAttributesFrom(SrcGV);
982 forceRenaming(NG, SrcGV->getName());
983
984 Constant *Ret = ConstantExpr::getBitCast(NG, TypeMap.get(SrcGV->getType()));
985
986 Mapper.scheduleMapAppendingVariable(
987 *NG,
988 (DstGV && !DstGV->isDeclaration()) ? DstGV->getInitializer() : nullptr,
989 IsOldStructor, SrcElements);
990
991 // Replace any uses of the two global variables with uses of the new
992 // global.
993 if (DstGV) {
994 RAUWWorklist.push_back(std::make_pair(DstGV, NG));
995 }
996
997 return Ret;
998 }
999
shouldLink(GlobalValue * DGV,GlobalValue & SGV)1000 bool IRLinker::shouldLink(GlobalValue *DGV, GlobalValue &SGV) {
1001 if (ValuesToLink.count(&SGV) || SGV.hasLocalLinkage())
1002 return true;
1003
1004 if (DGV && !DGV->isDeclarationForLinker())
1005 return false;
1006
1007 if (SGV.isDeclaration() || DoneLinkingBodies)
1008 return false;
1009
1010 // Callback to the client to give a chance to lazily add the Global to the
1011 // list of value to link.
1012 bool LazilyAdded = false;
1013 if (AddLazyFor)
1014 AddLazyFor(SGV, [this, &LazilyAdded](GlobalValue &GV) {
1015 maybeAdd(&GV);
1016 LazilyAdded = true;
1017 });
1018 return LazilyAdded;
1019 }
1020
linkGlobalValueProto(GlobalValue * SGV,bool ForIndirectSymbol)1021 Expected<Constant *> IRLinker::linkGlobalValueProto(GlobalValue *SGV,
1022 bool ForIndirectSymbol) {
1023 GlobalValue *DGV = getLinkedToGlobal(SGV);
1024
1025 bool ShouldLink = shouldLink(DGV, *SGV);
1026
1027 // just missing from map
1028 if (ShouldLink) {
1029 auto I = ValueMap.find(SGV);
1030 if (I != ValueMap.end())
1031 return cast<Constant>(I->second);
1032
1033 I = IndirectSymbolValueMap.find(SGV);
1034 if (I != IndirectSymbolValueMap.end())
1035 return cast<Constant>(I->second);
1036 }
1037
1038 if (!ShouldLink && ForIndirectSymbol)
1039 DGV = nullptr;
1040
1041 // Handle the ultra special appending linkage case first.
1042 if (SGV->hasAppendingLinkage() || (DGV && DGV->hasAppendingLinkage()))
1043 return linkAppendingVarProto(cast_or_null<GlobalVariable>(DGV),
1044 cast<GlobalVariable>(SGV));
1045
1046 bool NeedsRenaming = false;
1047 GlobalValue *NewGV;
1048 if (DGV && !ShouldLink) {
1049 NewGV = DGV;
1050 } else {
1051 // If we are done linking global value bodies (i.e. we are performing
1052 // metadata linking), don't link in the global value due to this
1053 // reference, simply map it to null.
1054 if (DoneLinkingBodies)
1055 return nullptr;
1056
1057 NewGV = copyGlobalValueProto(SGV, ShouldLink || ForIndirectSymbol);
1058 if (ShouldLink || !ForIndirectSymbol)
1059 NeedsRenaming = true;
1060 }
1061
1062 // Overloaded intrinsics have overloaded types names as part of their
1063 // names. If we renamed overloaded types we should rename the intrinsic
1064 // as well.
1065 if (Function *F = dyn_cast<Function>(NewGV))
1066 if (auto Remangled = Intrinsic::remangleIntrinsicFunction(F)) {
1067 // Note: remangleIntrinsicFunction does not copy metadata and as such
1068 // F should not occur in the set of objects with unmapped metadata.
1069 // If this assertion fails then remangleIntrinsicFunction needs updating.
1070 assert(!UnmappedMetadata.count(F) && "intrinsic has unmapped metadata");
1071 NewGV->eraseFromParent();
1072 NewGV = *Remangled;
1073 NeedsRenaming = false;
1074 }
1075
1076 if (NeedsRenaming)
1077 forceRenaming(NewGV, SGV->getName());
1078
1079 if (ShouldLink || ForIndirectSymbol) {
1080 if (const Comdat *SC = SGV->getComdat()) {
1081 if (auto *GO = dyn_cast<GlobalObject>(NewGV)) {
1082 Comdat *DC = DstM.getOrInsertComdat(SC->getName());
1083 DC->setSelectionKind(SC->getSelectionKind());
1084 GO->setComdat(DC);
1085 }
1086 }
1087 }
1088
1089 if (!ShouldLink && ForIndirectSymbol)
1090 NewGV->setLinkage(GlobalValue::InternalLinkage);
1091
1092 Constant *C = NewGV;
1093 // Only create a bitcast if necessary. In particular, with
1094 // DebugTypeODRUniquing we may reach metadata in the destination module
1095 // containing a GV from the source module, in which case SGV will be
1096 // the same as DGV and NewGV, and TypeMap.get() will assert since it
1097 // assumes it is being invoked on a type in the source module.
1098 if (DGV && NewGV != SGV) {
1099 C = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
1100 NewGV, TypeMap.get(SGV->getType()));
1101 }
1102
1103 if (DGV && NewGV != DGV) {
1104 // Schedule "replace all uses with" to happen after materializing is
1105 // done. It is not safe to do it now, since ValueMapper may be holding
1106 // pointers to constants that will get deleted if RAUW runs.
1107 RAUWWorklist.push_back(std::make_pair(
1108 DGV,
1109 ConstantExpr::getPointerBitCastOrAddrSpaceCast(NewGV, DGV->getType())));
1110 }
1111
1112 return C;
1113 }
1114
1115 /// Update the initializers in the Dest module now that all globals that may be
1116 /// referenced are in Dest.
linkGlobalVariable(GlobalVariable & Dst,GlobalVariable & Src)1117 void IRLinker::linkGlobalVariable(GlobalVariable &Dst, GlobalVariable &Src) {
1118 // Figure out what the initializer looks like in the dest module.
1119 Mapper.scheduleMapGlobalInitializer(Dst, *Src.getInitializer());
1120 }
1121
1122 /// Copy the source function over into the dest function and fix up references
1123 /// to values. At this point we know that Dest is an external function, and
1124 /// that Src is not.
linkFunctionBody(Function & Dst,Function & Src)1125 Error IRLinker::linkFunctionBody(Function &Dst, Function &Src) {
1126 assert(Dst.isDeclaration() && !Src.isDeclaration());
1127
1128 // Materialize if needed.
1129 if (Error Err = Src.materialize())
1130 return Err;
1131
1132 // Link in the operands without remapping.
1133 if (Src.hasPrefixData())
1134 Dst.setPrefixData(Src.getPrefixData());
1135 if (Src.hasPrologueData())
1136 Dst.setPrologueData(Src.getPrologueData());
1137 if (Src.hasPersonalityFn())
1138 Dst.setPersonalityFn(Src.getPersonalityFn());
1139 assert(Src.IsNewDbgInfoFormat == Dst.IsNewDbgInfoFormat);
1140
1141 // Copy over the metadata attachments without remapping.
1142 Dst.copyMetadata(&Src, 0);
1143
1144 // Steal arguments and splice the body of Src into Dst.
1145 Dst.stealArgumentListFrom(Src);
1146 Dst.splice(Dst.end(), &Src);
1147
1148 // Everything has been moved over. Remap it.
1149 Mapper.scheduleRemapFunction(Dst);
1150 return Error::success();
1151 }
1152
linkAliasAliasee(GlobalAlias & Dst,GlobalAlias & Src)1153 void IRLinker::linkAliasAliasee(GlobalAlias &Dst, GlobalAlias &Src) {
1154 Mapper.scheduleMapGlobalAlias(Dst, *Src.getAliasee(), IndirectSymbolMCID);
1155 }
1156
linkIFuncResolver(GlobalIFunc & Dst,GlobalIFunc & Src)1157 void IRLinker::linkIFuncResolver(GlobalIFunc &Dst, GlobalIFunc &Src) {
1158 Mapper.scheduleMapGlobalIFunc(Dst, *Src.getResolver(), IndirectSymbolMCID);
1159 }
1160
linkGlobalValueBody(GlobalValue & Dst,GlobalValue & Src)1161 Error IRLinker::linkGlobalValueBody(GlobalValue &Dst, GlobalValue &Src) {
1162 if (auto *F = dyn_cast<Function>(&Src))
1163 return linkFunctionBody(cast<Function>(Dst), *F);
1164 if (auto *GVar = dyn_cast<GlobalVariable>(&Src)) {
1165 linkGlobalVariable(cast<GlobalVariable>(Dst), *GVar);
1166 return Error::success();
1167 }
1168 if (auto *GA = dyn_cast<GlobalAlias>(&Src)) {
1169 linkAliasAliasee(cast<GlobalAlias>(Dst), *GA);
1170 return Error::success();
1171 }
1172 linkIFuncResolver(cast<GlobalIFunc>(Dst), cast<GlobalIFunc>(Src));
1173 return Error::success();
1174 }
1175
flushRAUWWorklist()1176 void IRLinker::flushRAUWWorklist() {
1177 for (const auto &Elem : RAUWWorklist) {
1178 GlobalValue *Old;
1179 Value *New;
1180 std::tie(Old, New) = Elem;
1181
1182 Old->replaceAllUsesWith(New);
1183 Old->eraseFromParent();
1184 }
1185 RAUWWorklist.clear();
1186 }
1187
prepareCompileUnitsForImport()1188 void IRLinker::prepareCompileUnitsForImport() {
1189 NamedMDNode *SrcCompileUnits = SrcM->getNamedMetadata("llvm.dbg.cu");
1190 if (!SrcCompileUnits)
1191 return;
1192 // When importing for ThinLTO, prevent importing of types listed on
1193 // the DICompileUnit that we don't need a copy of in the importing
1194 // module. They will be emitted by the originating module.
1195 for (MDNode *N : SrcCompileUnits->operands()) {
1196 auto *CU = cast<DICompileUnit>(N);
1197 assert(CU && "Expected valid compile unit");
1198 // Enums, macros, and retained types don't need to be listed on the
1199 // imported DICompileUnit. This means they will only be imported
1200 // if reached from the mapped IR.
1201 CU->replaceEnumTypes(nullptr);
1202 CU->replaceMacros(nullptr);
1203 CU->replaceRetainedTypes(nullptr);
1204
1205 // The original definition (or at least its debug info - if the variable is
1206 // internalized and optimized away) will remain in the source module, so
1207 // there's no need to import them.
1208 // If LLVM ever does more advanced optimizations on global variables
1209 // (removing/localizing write operations, for instance) that can track
1210 // through debug info, this decision may need to be revisited - but do so
1211 // with care when it comes to debug info size. Emitting small CUs containing
1212 // only a few imported entities into every destination module may be very
1213 // size inefficient.
1214 CU->replaceGlobalVariables(nullptr);
1215
1216 CU->replaceImportedEntities(nullptr);
1217 }
1218 }
1219
1220 /// Insert all of the named MDNodes in Src into the Dest module.
linkNamedMDNodes()1221 void IRLinker::linkNamedMDNodes() {
1222 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1223 for (const NamedMDNode &NMD : SrcM->named_metadata()) {
1224 // Don't link module flags here. Do them separately.
1225 if (&NMD == SrcModFlags)
1226 continue;
1227 // Don't import pseudo probe descriptors here for thinLTO. They will be
1228 // emitted by the originating module.
1229 if (IsPerformingImport && NMD.getName() == PseudoProbeDescMetadataName) {
1230 if (!DstM.getNamedMetadata(NMD.getName()))
1231 emitWarning("Pseudo-probe ignored: source module '" +
1232 SrcM->getModuleIdentifier() +
1233 "' is compiled with -fpseudo-probe-for-profiling while "
1234 "destination module '" +
1235 DstM.getModuleIdentifier() + "' is not\n");
1236 continue;
1237 }
1238 // The stats are computed per module and will all be merged in the binary.
1239 // Importing the metadata will cause duplication of the stats.
1240 if (IsPerformingImport && NMD.getName() == "llvm.stats")
1241 continue;
1242
1243 NamedMDNode *DestNMD = DstM.getOrInsertNamedMetadata(NMD.getName());
1244 // Add Src elements into Dest node.
1245 for (const MDNode *Op : NMD.operands())
1246 DestNMD->addOperand(Mapper.mapMDNode(*Op));
1247 }
1248 }
1249
1250 /// Merge the linker flags in Src into the Dest module.
linkModuleFlagsMetadata()1251 Error IRLinker::linkModuleFlagsMetadata() {
1252 // If the source module has no module flags, we are done.
1253 const NamedMDNode *SrcModFlags = SrcM->getModuleFlagsMetadata();
1254 if (!SrcModFlags)
1255 return Error::success();
1256
1257 // Check for module flag for updates before do anything.
1258 UpgradeModuleFlags(*SrcM);
1259
1260 // If the destination module doesn't have module flags yet, then just copy
1261 // over the source module's flags.
1262 NamedMDNode *DstModFlags = DstM.getOrInsertModuleFlagsMetadata();
1263 if (DstModFlags->getNumOperands() == 0) {
1264 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I)
1265 DstModFlags->addOperand(SrcModFlags->getOperand(I));
1266
1267 return Error::success();
1268 }
1269
1270 // First build a map of the existing module flags and requirements.
1271 DenseMap<MDString *, std::pair<MDNode *, unsigned>> Flags;
1272 SmallSetVector<MDNode *, 16> Requirements;
1273 SmallVector<unsigned, 0> Mins;
1274 DenseSet<MDString *> SeenMin;
1275 for (unsigned I = 0, E = DstModFlags->getNumOperands(); I != E; ++I) {
1276 MDNode *Op = DstModFlags->getOperand(I);
1277 uint64_t Behavior =
1278 mdconst::extract<ConstantInt>(Op->getOperand(0))->getZExtValue();
1279 MDString *ID = cast<MDString>(Op->getOperand(1));
1280
1281 if (Behavior == Module::Require) {
1282 Requirements.insert(cast<MDNode>(Op->getOperand(2)));
1283 } else {
1284 if (Behavior == Module::Min)
1285 Mins.push_back(I);
1286 Flags[ID] = std::make_pair(Op, I);
1287 }
1288 }
1289
1290 // Merge in the flags from the source module, and also collect its set of
1291 // requirements.
1292 for (unsigned I = 0, E = SrcModFlags->getNumOperands(); I != E; ++I) {
1293 MDNode *SrcOp = SrcModFlags->getOperand(I);
1294 ConstantInt *SrcBehavior =
1295 mdconst::extract<ConstantInt>(SrcOp->getOperand(0));
1296 MDString *ID = cast<MDString>(SrcOp->getOperand(1));
1297 MDNode *DstOp;
1298 unsigned DstIndex;
1299 std::tie(DstOp, DstIndex) = Flags.lookup(ID);
1300 unsigned SrcBehaviorValue = SrcBehavior->getZExtValue();
1301 SeenMin.insert(ID);
1302
1303 // If this is a requirement, add it and continue.
1304 if (SrcBehaviorValue == Module::Require) {
1305 // If the destination module does not already have this requirement, add
1306 // it.
1307 if (Requirements.insert(cast<MDNode>(SrcOp->getOperand(2)))) {
1308 DstModFlags->addOperand(SrcOp);
1309 }
1310 continue;
1311 }
1312
1313 // If there is no existing flag with this ID, just add it.
1314 if (!DstOp) {
1315 if (SrcBehaviorValue == Module::Min) {
1316 Mins.push_back(DstModFlags->getNumOperands());
1317 SeenMin.erase(ID);
1318 }
1319 Flags[ID] = std::make_pair(SrcOp, DstModFlags->getNumOperands());
1320 DstModFlags->addOperand(SrcOp);
1321 continue;
1322 }
1323
1324 // Otherwise, perform a merge.
1325 ConstantInt *DstBehavior =
1326 mdconst::extract<ConstantInt>(DstOp->getOperand(0));
1327 unsigned DstBehaviorValue = DstBehavior->getZExtValue();
1328
1329 auto overrideDstValue = [&]() {
1330 DstModFlags->setOperand(DstIndex, SrcOp);
1331 Flags[ID].first = SrcOp;
1332 };
1333
1334 // If either flag has override behavior, handle it first.
1335 if (DstBehaviorValue == Module::Override) {
1336 // Diagnose inconsistent flags which both have override behavior.
1337 if (SrcBehaviorValue == Module::Override &&
1338 SrcOp->getOperand(2) != DstOp->getOperand(2))
1339 return stringErr("linking module flags '" + ID->getString() +
1340 "': IDs have conflicting override values in '" +
1341 SrcM->getModuleIdentifier() + "' and '" +
1342 DstM.getModuleIdentifier() + "'");
1343 continue;
1344 } else if (SrcBehaviorValue == Module::Override) {
1345 // Update the destination flag to that of the source.
1346 overrideDstValue();
1347 continue;
1348 }
1349
1350 // Diagnose inconsistent merge behavior types.
1351 if (SrcBehaviorValue != DstBehaviorValue) {
1352 bool MinAndWarn = (SrcBehaviorValue == Module::Min &&
1353 DstBehaviorValue == Module::Warning) ||
1354 (DstBehaviorValue == Module::Min &&
1355 SrcBehaviorValue == Module::Warning);
1356 bool MaxAndWarn = (SrcBehaviorValue == Module::Max &&
1357 DstBehaviorValue == Module::Warning) ||
1358 (DstBehaviorValue == Module::Max &&
1359 SrcBehaviorValue == Module::Warning);
1360 if (!(MaxAndWarn || MinAndWarn))
1361 return stringErr("linking module flags '" + ID->getString() +
1362 "': IDs have conflicting behaviors in '" +
1363 SrcM->getModuleIdentifier() + "' and '" +
1364 DstM.getModuleIdentifier() + "'");
1365 }
1366
1367 auto ensureDistinctOp = [&](MDNode *DstValue) {
1368 assert(isa<MDTuple>(DstValue) &&
1369 "Expected MDTuple when appending module flags");
1370 if (DstValue->isDistinct())
1371 return dyn_cast<MDTuple>(DstValue);
1372 ArrayRef<MDOperand> DstOperands = DstValue->operands();
1373 MDTuple *New = MDTuple::getDistinct(
1374 DstM.getContext(),
1375 SmallVector<Metadata *, 4>(DstOperands.begin(), DstOperands.end()));
1376 Metadata *FlagOps[] = {DstOp->getOperand(0), ID, New};
1377 MDNode *Flag = MDTuple::getDistinct(DstM.getContext(), FlagOps);
1378 DstModFlags->setOperand(DstIndex, Flag);
1379 Flags[ID].first = Flag;
1380 return New;
1381 };
1382
1383 // Emit a warning if the values differ and either source or destination
1384 // request Warning behavior.
1385 if ((DstBehaviorValue == Module::Warning ||
1386 SrcBehaviorValue == Module::Warning) &&
1387 SrcOp->getOperand(2) != DstOp->getOperand(2)) {
1388 std::string Str;
1389 raw_string_ostream(Str)
1390 << "linking module flags '" << ID->getString()
1391 << "': IDs have conflicting values ('" << *SrcOp->getOperand(2)
1392 << "' from " << SrcM->getModuleIdentifier() << " with '"
1393 << *DstOp->getOperand(2) << "' from " << DstM.getModuleIdentifier()
1394 << ')';
1395 emitWarning(Str);
1396 }
1397
1398 // Choose the minimum if either source or destination request Min behavior.
1399 if (DstBehaviorValue == Module::Min || SrcBehaviorValue == Module::Min) {
1400 ConstantInt *DstValue =
1401 mdconst::extract<ConstantInt>(DstOp->getOperand(2));
1402 ConstantInt *SrcValue =
1403 mdconst::extract<ConstantInt>(SrcOp->getOperand(2));
1404
1405 // The resulting flag should have a Min behavior, and contain the minimum
1406 // value from between the source and destination values.
1407 Metadata *FlagOps[] = {
1408 (DstBehaviorValue != Module::Min ? SrcOp : DstOp)->getOperand(0), ID,
1409 (SrcValue->getZExtValue() < DstValue->getZExtValue() ? SrcOp : DstOp)
1410 ->getOperand(2)};
1411 MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps);
1412 DstModFlags->setOperand(DstIndex, Flag);
1413 Flags[ID].first = Flag;
1414 continue;
1415 }
1416
1417 // Choose the maximum if either source or destination request Max behavior.
1418 if (DstBehaviorValue == Module::Max || SrcBehaviorValue == Module::Max) {
1419 ConstantInt *DstValue =
1420 mdconst::extract<ConstantInt>(DstOp->getOperand(2));
1421 ConstantInt *SrcValue =
1422 mdconst::extract<ConstantInt>(SrcOp->getOperand(2));
1423
1424 // The resulting flag should have a Max behavior, and contain the maximum
1425 // value from between the source and destination values.
1426 Metadata *FlagOps[] = {
1427 (DstBehaviorValue != Module::Max ? SrcOp : DstOp)->getOperand(0), ID,
1428 (SrcValue->getZExtValue() > DstValue->getZExtValue() ? SrcOp : DstOp)
1429 ->getOperand(2)};
1430 MDNode *Flag = MDNode::get(DstM.getContext(), FlagOps);
1431 DstModFlags->setOperand(DstIndex, Flag);
1432 Flags[ID].first = Flag;
1433 continue;
1434 }
1435
1436 // Perform the merge for standard behavior types.
1437 switch (SrcBehaviorValue) {
1438 case Module::Require:
1439 case Module::Override:
1440 llvm_unreachable("not possible");
1441 case Module::Error: {
1442 // Emit an error if the values differ.
1443 if (SrcOp->getOperand(2) != DstOp->getOperand(2))
1444 return stringErr("linking module flags '" + ID->getString() +
1445 "': IDs have conflicting values in '" +
1446 SrcM->getModuleIdentifier() + "' and '" +
1447 DstM.getModuleIdentifier() + "'");
1448 continue;
1449 }
1450 case Module::Warning: {
1451 break;
1452 }
1453 case Module::Max: {
1454 break;
1455 }
1456 case Module::Append: {
1457 MDTuple *DstValue = ensureDistinctOp(cast<MDNode>(DstOp->getOperand(2)));
1458 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1459 for (const auto &O : SrcValue->operands())
1460 DstValue->push_back(O);
1461 break;
1462 }
1463 case Module::AppendUnique: {
1464 SmallSetVector<Metadata *, 16> Elts;
1465 MDTuple *DstValue = ensureDistinctOp(cast<MDNode>(DstOp->getOperand(2)));
1466 MDNode *SrcValue = cast<MDNode>(SrcOp->getOperand(2));
1467 Elts.insert(DstValue->op_begin(), DstValue->op_end());
1468 Elts.insert(SrcValue->op_begin(), SrcValue->op_end());
1469 for (auto I = DstValue->getNumOperands(); I < Elts.size(); I++)
1470 DstValue->push_back(Elts[I]);
1471 break;
1472 }
1473 }
1474
1475 }
1476
1477 // For the Min behavior, set the value to 0 if either module does not have the
1478 // flag.
1479 for (auto Idx : Mins) {
1480 MDNode *Op = DstModFlags->getOperand(Idx);
1481 MDString *ID = cast<MDString>(Op->getOperand(1));
1482 if (!SeenMin.count(ID)) {
1483 ConstantInt *V = mdconst::extract<ConstantInt>(Op->getOperand(2));
1484 Metadata *FlagOps[] = {
1485 Op->getOperand(0), ID,
1486 ConstantAsMetadata::get(ConstantInt::get(V->getType(), 0))};
1487 DstModFlags->setOperand(Idx, MDNode::get(DstM.getContext(), FlagOps));
1488 }
1489 }
1490
1491 // Check all of the requirements.
1492 for (MDNode *Requirement : Requirements) {
1493 MDString *Flag = cast<MDString>(Requirement->getOperand(0));
1494 Metadata *ReqValue = Requirement->getOperand(1);
1495
1496 MDNode *Op = Flags[Flag].first;
1497 if (!Op || Op->getOperand(2) != ReqValue)
1498 return stringErr("linking module flags '" + Flag->getString() +
1499 "': does not have the required value");
1500 }
1501 return Error::success();
1502 }
1503
1504 /// Return InlineAsm adjusted with target-specific directives if required.
1505 /// For ARM and Thumb, we have to add directives to select the appropriate ISA
1506 /// to support mixing module-level inline assembly from ARM and Thumb modules.
adjustInlineAsm(const std::string & InlineAsm,const Triple & Triple)1507 static std::string adjustInlineAsm(const std::string &InlineAsm,
1508 const Triple &Triple) {
1509 if (Triple.getArch() == Triple::thumb || Triple.getArch() == Triple::thumbeb)
1510 return ".text\n.balign 2\n.thumb\n" + InlineAsm;
1511 if (Triple.getArch() == Triple::arm || Triple.getArch() == Triple::armeb)
1512 return ".text\n.balign 4\n.arm\n" + InlineAsm;
1513 return InlineAsm;
1514 }
1515
updateAttributes(GlobalValue & GV)1516 void IRLinker::updateAttributes(GlobalValue &GV) {
1517 /// Remove nocallback attribute while linking, because nocallback attribute
1518 /// indicates that the function is only allowed to jump back into caller's
1519 /// module only by a return or an exception. When modules are linked, this
1520 /// property cannot be guaranteed anymore. For example, the nocallback
1521 /// function may contain a call to another module. But if we merge its caller
1522 /// and callee module here, and not the module containing the nocallback
1523 /// function definition itself, the nocallback property will be violated
1524 /// (since the nocallback function will call back into the newly merged module
1525 /// containing both its caller and callee). This could happen if the module
1526 /// containing the nocallback function definition is native code, so it does
1527 /// not participate in the LTO link. Note if the nocallback function does
1528 /// participate in the LTO link, and thus ends up in the merged module
1529 /// containing its caller and callee, removing the attribute doesn't hurt as
1530 /// it has no effect on definitions in the same module.
1531 if (auto *F = dyn_cast<Function>(&GV)) {
1532 if (!F->isIntrinsic())
1533 F->removeFnAttr(llvm::Attribute::NoCallback);
1534
1535 // Remove nocallback attribute when it is on a call-site.
1536 for (BasicBlock &BB : *F)
1537 for (Instruction &I : BB)
1538 if (CallBase *CI = dyn_cast<CallBase>(&I))
1539 CI->removeFnAttr(Attribute::NoCallback);
1540 }
1541 }
1542
run()1543 Error IRLinker::run() {
1544 // Ensure metadata materialized before value mapping.
1545 if (SrcM->getMaterializer())
1546 if (Error Err = SrcM->getMaterializer()->materializeMetadata())
1547 return Err;
1548
1549 // Convert source module to match dest for the duration of the link.
1550 ScopedDbgInfoFormatSetter FormatSetter(*SrcM, DstM.IsNewDbgInfoFormat);
1551
1552 // Inherit the target data from the source module if the destination
1553 // module doesn't have one already.
1554 if (DstM.getDataLayout().isDefault())
1555 DstM.setDataLayout(SrcM->getDataLayout());
1556
1557 // Copy the target triple from the source to dest if the dest's is empty.
1558 if (DstM.getTargetTriple().empty() && !SrcM->getTargetTriple().empty())
1559 DstM.setTargetTriple(SrcM->getTargetTriple());
1560
1561 Triple SrcTriple(SrcM->getTargetTriple()), DstTriple(DstM.getTargetTriple());
1562
1563 // During CUDA compilation we have to link with the bitcode supplied with
1564 // CUDA. libdevice bitcode either has no data layout set (pre-CUDA-11), or has
1565 // the layout that is different from the one used by LLVM/clang (it does not
1566 // include i128). Issuing a warning is not very helpful as there's not much
1567 // the user can do about it.
1568 bool EnableDLWarning = true;
1569 bool EnableTripleWarning = true;
1570 if (SrcTriple.isNVPTX() && DstTriple.isNVPTX()) {
1571 std::string ModuleId = SrcM->getModuleIdentifier();
1572 StringRef FileName = llvm::sys::path::filename(ModuleId);
1573 bool SrcIsLibDevice =
1574 FileName.starts_with("libdevice") && FileName.ends_with(".10.bc");
1575 bool SrcHasLibDeviceDL =
1576 (SrcM->getDataLayoutStr().empty() ||
1577 SrcM->getDataLayoutStr() == "e-i64:64-v16:16-v32:32-n16:32:64");
1578 // libdevice bitcode uses nvptx64-nvidia-gpulibs or just
1579 // 'nvptx-unknown-unknown' triple (before CUDA-10.x) and is compatible with
1580 // all NVPTX variants.
1581 bool SrcHasLibDeviceTriple = (SrcTriple.getVendor() == Triple::NVIDIA &&
1582 SrcTriple.getOSName() == "gpulibs") ||
1583 (SrcTriple.getVendorName() == "unknown" &&
1584 SrcTriple.getOSName() == "unknown");
1585 EnableTripleWarning = !(SrcIsLibDevice && SrcHasLibDeviceTriple);
1586 EnableDLWarning = !(SrcIsLibDevice && SrcHasLibDeviceDL);
1587 }
1588
1589 if (EnableDLWarning && (SrcM->getDataLayout() != DstM.getDataLayout())) {
1590 emitWarning("Linking two modules of different data layouts: '" +
1591 SrcM->getModuleIdentifier() + "' is '" +
1592 SrcM->getDataLayoutStr() + "' whereas '" +
1593 DstM.getModuleIdentifier() + "' is '" +
1594 DstM.getDataLayoutStr() + "'\n");
1595 }
1596
1597 if (EnableTripleWarning && !SrcM->getTargetTriple().empty() &&
1598 !SrcTriple.isCompatibleWith(DstTriple))
1599 emitWarning("Linking two modules of different target triples: '" +
1600 SrcM->getModuleIdentifier() + "' is '" +
1601 SrcM->getTargetTriple() + "' whereas '" +
1602 DstM.getModuleIdentifier() + "' is '" + DstM.getTargetTriple() +
1603 "'\n");
1604
1605 DstM.setTargetTriple(SrcTriple.merge(DstTriple));
1606
1607 // Loop over all of the linked values to compute type mappings.
1608 computeTypeMapping();
1609
1610 std::reverse(Worklist.begin(), Worklist.end());
1611 while (!Worklist.empty()) {
1612 GlobalValue *GV = Worklist.back();
1613 Worklist.pop_back();
1614
1615 // Already mapped.
1616 if (ValueMap.find(GV) != ValueMap.end() ||
1617 IndirectSymbolValueMap.find(GV) != IndirectSymbolValueMap.end())
1618 continue;
1619
1620 assert(!GV->isDeclaration());
1621 Mapper.mapValue(*GV);
1622 if (FoundError)
1623 return std::move(*FoundError);
1624 flushRAUWWorklist();
1625 }
1626
1627 // Note that we are done linking global value bodies. This prevents
1628 // metadata linking from creating new references.
1629 DoneLinkingBodies = true;
1630 Mapper.addFlags(RF_NullMapMissingGlobalValues);
1631
1632 // Remap all of the named MDNodes in Src into the DstM module. We do this
1633 // after linking GlobalValues so that MDNodes that reference GlobalValues
1634 // are properly remapped.
1635 linkNamedMDNodes();
1636
1637 // Clean up any global objects with potentially unmapped metadata.
1638 // Specifically declarations which did not become definitions.
1639 for (GlobalObject *NGO : UnmappedMetadata) {
1640 if (NGO->isDeclaration())
1641 Mapper.remapGlobalObjectMetadata(*NGO);
1642 }
1643
1644 if (!IsPerformingImport && !SrcM->getModuleInlineAsm().empty()) {
1645 // Append the module inline asm string.
1646 DstM.appendModuleInlineAsm(adjustInlineAsm(SrcM->getModuleInlineAsm(),
1647 SrcTriple));
1648 } else if (IsPerformingImport) {
1649 // Import any symver directives for symbols in DstM.
1650 ModuleSymbolTable::CollectAsmSymvers(*SrcM,
1651 [&](StringRef Name, StringRef Alias) {
1652 if (DstM.getNamedValue(Name)) {
1653 SmallString<256> S(".symver ");
1654 S += Name;
1655 S += ", ";
1656 S += Alias;
1657 DstM.appendModuleInlineAsm(S);
1658 }
1659 });
1660 }
1661
1662 // Reorder the globals just added to the destination module to match their
1663 // original order in the source module.
1664 for (GlobalVariable &GV : SrcM->globals()) {
1665 if (GV.hasAppendingLinkage())
1666 continue;
1667 Value *NewValue = Mapper.mapValue(GV);
1668 if (NewValue) {
1669 auto *NewGV = dyn_cast<GlobalVariable>(NewValue->stripPointerCasts());
1670 if (NewGV) {
1671 NewGV->removeFromParent();
1672 DstM.insertGlobalVariable(NewGV);
1673 }
1674 }
1675 }
1676
1677 // Merge the module flags into the DstM module.
1678 return linkModuleFlagsMetadata();
1679 }
1680
KeyTy(ArrayRef<Type * > E,bool P)1681 IRMover::StructTypeKeyInfo::KeyTy::KeyTy(ArrayRef<Type *> E, bool P)
1682 : ETypes(E), IsPacked(P) {}
1683
KeyTy(const StructType * ST)1684 IRMover::StructTypeKeyInfo::KeyTy::KeyTy(const StructType *ST)
1685 : ETypes(ST->elements()), IsPacked(ST->isPacked()) {}
1686
operator ==(const KeyTy & That) const1687 bool IRMover::StructTypeKeyInfo::KeyTy::operator==(const KeyTy &That) const {
1688 return IsPacked == That.IsPacked && ETypes == That.ETypes;
1689 }
1690
operator !=(const KeyTy & That) const1691 bool IRMover::StructTypeKeyInfo::KeyTy::operator!=(const KeyTy &That) const {
1692 return !this->operator==(That);
1693 }
1694
getEmptyKey()1695 StructType *IRMover::StructTypeKeyInfo::getEmptyKey() {
1696 return DenseMapInfo<StructType *>::getEmptyKey();
1697 }
1698
getTombstoneKey()1699 StructType *IRMover::StructTypeKeyInfo::getTombstoneKey() {
1700 return DenseMapInfo<StructType *>::getTombstoneKey();
1701 }
1702
getHashValue(const KeyTy & Key)1703 unsigned IRMover::StructTypeKeyInfo::getHashValue(const KeyTy &Key) {
1704 return hash_combine(hash_combine_range(Key.ETypes.begin(), Key.ETypes.end()),
1705 Key.IsPacked);
1706 }
1707
getHashValue(const StructType * ST)1708 unsigned IRMover::StructTypeKeyInfo::getHashValue(const StructType *ST) {
1709 return getHashValue(KeyTy(ST));
1710 }
1711
isEqual(const KeyTy & LHS,const StructType * RHS)1712 bool IRMover::StructTypeKeyInfo::isEqual(const KeyTy &LHS,
1713 const StructType *RHS) {
1714 if (RHS == getEmptyKey() || RHS == getTombstoneKey())
1715 return false;
1716 return LHS == KeyTy(RHS);
1717 }
1718
isEqual(const StructType * LHS,const StructType * RHS)1719 bool IRMover::StructTypeKeyInfo::isEqual(const StructType *LHS,
1720 const StructType *RHS) {
1721 if (RHS == getEmptyKey() || RHS == getTombstoneKey())
1722 return LHS == RHS;
1723 return KeyTy(LHS) == KeyTy(RHS);
1724 }
1725
addNonOpaque(StructType * Ty)1726 void IRMover::IdentifiedStructTypeSet::addNonOpaque(StructType *Ty) {
1727 assert(!Ty->isOpaque());
1728 NonOpaqueStructTypes.insert(Ty);
1729 }
1730
switchToNonOpaque(StructType * Ty)1731 void IRMover::IdentifiedStructTypeSet::switchToNonOpaque(StructType *Ty) {
1732 assert(!Ty->isOpaque());
1733 NonOpaqueStructTypes.insert(Ty);
1734 bool Removed = OpaqueStructTypes.erase(Ty);
1735 (void)Removed;
1736 assert(Removed);
1737 }
1738
addOpaque(StructType * Ty)1739 void IRMover::IdentifiedStructTypeSet::addOpaque(StructType *Ty) {
1740 assert(Ty->isOpaque());
1741 OpaqueStructTypes.insert(Ty);
1742 }
1743
1744 StructType *
findNonOpaque(ArrayRef<Type * > ETypes,bool IsPacked)1745 IRMover::IdentifiedStructTypeSet::findNonOpaque(ArrayRef<Type *> ETypes,
1746 bool IsPacked) {
1747 IRMover::StructTypeKeyInfo::KeyTy Key(ETypes, IsPacked);
1748 auto I = NonOpaqueStructTypes.find_as(Key);
1749 return I == NonOpaqueStructTypes.end() ? nullptr : *I;
1750 }
1751
hasType(StructType * Ty)1752 bool IRMover::IdentifiedStructTypeSet::hasType(StructType *Ty) {
1753 if (Ty->isOpaque())
1754 return OpaqueStructTypes.count(Ty);
1755 auto I = NonOpaqueStructTypes.find(Ty);
1756 return I == NonOpaqueStructTypes.end() ? false : *I == Ty;
1757 }
1758
IRMover(Module & M)1759 IRMover::IRMover(Module &M) : Composite(M) {
1760 TypeFinder StructTypes;
1761 StructTypes.run(M, /* OnlyNamed */ false);
1762 for (StructType *Ty : StructTypes) {
1763 if (Ty->isOpaque())
1764 IdentifiedStructTypes.addOpaque(Ty);
1765 else
1766 IdentifiedStructTypes.addNonOpaque(Ty);
1767 }
1768 // Self-map metadatas in the destination module. This is needed when
1769 // DebugTypeODRUniquing is enabled on the LLVMContext, since metadata in the
1770 // destination module may be reached from the source module.
1771 for (const auto *MD : StructTypes.getVisitedMetadata()) {
1772 SharedMDs[MD].reset(const_cast<MDNode *>(MD));
1773 }
1774 }
1775
move(std::unique_ptr<Module> Src,ArrayRef<GlobalValue * > ValuesToLink,LazyCallback AddLazyFor,bool IsPerformingImport)1776 Error IRMover::move(std::unique_ptr<Module> Src,
1777 ArrayRef<GlobalValue *> ValuesToLink,
1778 LazyCallback AddLazyFor, bool IsPerformingImport) {
1779 IRLinker TheIRLinker(Composite, SharedMDs, IdentifiedStructTypes,
1780 std::move(Src), ValuesToLink, std::move(AddLazyFor),
1781 IsPerformingImport);
1782 Error E = TheIRLinker.run();
1783 Composite.dropTriviallyDeadConstantArrays();
1784 return E;
1785 }
1786