1 //===-- DataFlowSanitizer.cpp - dynamic data flow analysis ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file is a part of DataFlowSanitizer, a generalised dynamic data flow
13 /// Unlike other Sanitizer tools, this tool is not designed to detect a specific
14 /// class of bugs on its own. Instead, it provides a generic dynamic data flow
15 /// analysis framework to be used by clients to help detect application-specific
16 /// issues within their own code.
18 /// The analysis is based on automatic propagation of data flow labels (also
19 /// known as taint labels) through a program as it performs computation. Each
20 /// byte of application memory is backed by two bytes of shadow memory which
21 /// hold the label. On Linux/x86_64, memory is laid out as follows:
23 /// +--------------------+ 0x800000000000 (top of memory)
24 /// | application memory |
25 /// +--------------------+ 0x700000008000 (kAppAddr)
29 /// +--------------------+ 0x200200000000 (kUnusedAddr)
31 /// +--------------------+ 0x200000000000 (kUnionTableAddr)
33 /// +--------------------+ 0x000000010000 (kShadowAddr)
34 /// | reserved by kernel |
35 /// +--------------------+ 0x000000000000
37 /// To derive a shadow memory address from an application memory address,
38 /// bits 44-46 are cleared to bring the address into the range
39 /// [0x000000008000,0x100000000000). Then the address is shifted left by 1 to
40 /// account for the double byte representation of shadow labels and move the
41 /// address into the shadow memory range. See the function
42 /// DataFlowSanitizer::getShadowAddress below.
44 /// For more information, please refer to the design document:
45 /// http://clang.llvm.org/docs/DataFlowSanitizerDesign.html
47 #include "llvm/Transforms/Instrumentation.h"
48 #include "llvm/ADT/DenseMap.h"
49 #include "llvm/ADT/DenseSet.h"
50 #include "llvm/ADT/DepthFirstIterator.h"
51 #include "llvm/ADT/StringExtras.h"
52 #include "llvm/ADT/Triple.h"
53 #include "llvm/Analysis/ValueTracking.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/DebugInfo.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InlineAsm.h"
58 #include "llvm/IR/InstVisitor.h"
59 #include "llvm/IR/LLVMContext.h"
60 #include "llvm/IR/MDBuilder.h"
61 #include "llvm/IR/Type.h"
62 #include "llvm/IR/Value.h"
63 #include "llvm/Pass.h"
64 #include "llvm/Support/CommandLine.h"
65 #include "llvm/Support/SpecialCaseList.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
75 // The -dfsan-preserve-alignment flag controls whether this pass assumes that
76 // alignment requirements provided by the input IR are correct. For example,
77 // if the input IR contains a load with alignment 8, this flag will cause
78 // the shadow load to have alignment 16. This flag is disabled by default as
79 // we have unfortunately encountered too much code (including Clang itself;
80 // see PR14291) which performs misaligned access.
81 static cl::opt<bool> ClPreserveAlignment(
82 "dfsan-preserve-alignment",
83 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
86 // The ABI list files control how shadow parameters are passed. The pass treats
87 // every function labelled "uninstrumented" in the ABI list file as conforming
88 // to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains
89 // additional annotations for those functions, a call to one of those functions
90 // will produce a warning message, as the labelling behaviour of the function is
91 // unknown. The other supported annotations are "functional" and "discard",
92 // which are described below under DataFlowSanitizer::WrapperKind.
93 static cl::list<std::string> ClABIListFiles(
95 cl::desc("File listing native ABI functions and how the pass treats them"),
98 // Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented
99 // functions (see DataFlowSanitizer::InstrumentedABI below).
100 static cl::opt<bool> ClArgsABI(
102 cl::desc("Use the argument ABI rather than the TLS ABI"),
105 // Controls whether the pass includes or ignores the labels of pointers in load
107 static cl::opt<bool> ClCombinePointerLabelsOnLoad(
108 "dfsan-combine-pointer-labels-on-load",
109 cl::desc("Combine the label of the pointer with the label of the data when "
110 "loading from memory."),
111 cl::Hidden, cl::init(true));
113 // Controls whether the pass includes or ignores the labels of pointers in
114 // stores instructions.
115 static cl::opt<bool> ClCombinePointerLabelsOnStore(
116 "dfsan-combine-pointer-labels-on-store",
117 cl::desc("Combine the label of the pointer with the label of the data when "
118 "storing in memory."),
119 cl::Hidden, cl::init(false));
121 static cl::opt<bool> ClDebugNonzeroLabels(
122 "dfsan-debug-nonzero-labels",
123 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
124 "load or return with a nonzero label"),
129 StringRef GetGlobalTypeString(const GlobalValue &G) {
130 // Types of GlobalVariables are always pointer types.
131 Type *GType = G.getType()->getElementType();
132 // For now we support blacklisting struct types only.
133 if (StructType *SGType = dyn_cast<StructType>(GType)) {
134 if (!SGType->isLiteral())
135 return SGType->getName();
137 return "<unknown type>";
141 std::unique_ptr<SpecialCaseList> SCL;
146 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); }
148 /// Returns whether either this function or its source file are listed in the
150 bool isIn(const Function &F, StringRef Category) const {
151 return isIn(*F.getParent(), Category) ||
152 SCL->inSection("fun", F.getName(), Category);
155 /// Returns whether this global alias is listed in the given category.
157 /// If GA aliases a function, the alias's name is matched as a function name
158 /// would be. Similarly, aliases of globals are matched like globals.
159 bool isIn(const GlobalAlias &GA, StringRef Category) const {
160 if (isIn(*GA.getParent(), Category))
163 if (isa<FunctionType>(GA.getType()->getElementType()))
164 return SCL->inSection("fun", GA.getName(), Category);
166 return SCL->inSection("global", GA.getName(), Category) ||
167 SCL->inSection("type", GetGlobalTypeString(GA), Category);
170 /// Returns whether this module is listed in the given category.
171 bool isIn(const Module &M, StringRef Category) const {
172 return SCL->inSection("src", M.getModuleIdentifier(), Category);
176 class DataFlowSanitizer : public ModulePass {
177 friend struct DFSanFunction;
178 friend class DFSanVisitor;
184 /// Which ABI should be used for instrumented functions?
185 enum InstrumentedABI {
186 /// Argument and return value labels are passed through additional
187 /// arguments and by modifying the return type.
190 /// Argument and return value labels are passed through TLS variables
191 /// __dfsan_arg_tls and __dfsan_retval_tls.
195 /// How should calls to uninstrumented functions be handled?
197 /// This function is present in an uninstrumented form but we don't know
198 /// how it should be handled. Print a warning and call the function anyway.
199 /// Don't label the return value.
202 /// This function does not write to (user-accessible) memory, and its return
203 /// value is unlabelled.
206 /// This function does not write to (user-accessible) memory, and the label
207 /// of its return value is the union of the label of its arguments.
210 /// Instead of calling the function, a custom wrapper __dfsw_F is called,
211 /// where F is the name of the function. This function may wrap the
212 /// original function or provide its own implementation. This is similar to
213 /// the IA_Args ABI, except that IA_Args uses a struct return type to
214 /// pass the return value shadow in a register, while WK_Custom uses an
215 /// extra pointer argument to return the shadow. This allows the wrapped
216 /// form of the function type to be expressed in C.
220 const DataLayout *DL;
223 IntegerType *ShadowTy;
224 PointerType *ShadowPtrTy;
225 IntegerType *IntptrTy;
226 ConstantInt *ZeroShadow;
227 ConstantInt *ShadowPtrMask;
228 ConstantInt *ShadowPtrMul;
231 void *(*GetArgTLSPtr)();
232 void *(*GetRetvalTLSPtr)();
234 Constant *GetRetvalTLS;
235 FunctionType *DFSanUnionFnTy;
236 FunctionType *DFSanUnionLoadFnTy;
237 FunctionType *DFSanUnimplementedFnTy;
238 FunctionType *DFSanSetLabelFnTy;
239 FunctionType *DFSanNonzeroLabelFnTy;
240 FunctionType *DFSanVarargWrapperFnTy;
241 Constant *DFSanUnionFn;
242 Constant *DFSanCheckedUnionFn;
243 Constant *DFSanUnionLoadFn;
244 Constant *DFSanUnimplementedFn;
245 Constant *DFSanSetLabelFn;
246 Constant *DFSanNonzeroLabelFn;
247 Constant *DFSanVarargWrapperFn;
248 MDNode *ColdCallWeights;
249 DFSanABIList ABIList;
250 DenseMap<Value *, Function *> UnwrappedFnMap;
251 AttributeSet ReadOnlyNoneAttrs;
252 DenseMap<const Function *, DISubprogram> FunctionDIs;
254 Value *getShadowAddress(Value *Addr, Instruction *Pos);
255 bool isInstrumented(const Function *F);
256 bool isInstrumented(const GlobalAlias *GA);
257 FunctionType *getArgsFunctionType(FunctionType *T);
258 FunctionType *getTrampolineFunctionType(FunctionType *T);
259 FunctionType *getCustomFunctionType(FunctionType *T);
260 InstrumentedABI getInstrumentedABI();
261 WrapperKind getWrapperKind(Function *F);
262 void addGlobalNamePrefix(GlobalValue *GV);
263 Function *buildWrapperFunction(Function *F, StringRef NewFName,
264 GlobalValue::LinkageTypes NewFLink,
265 FunctionType *NewFT);
266 Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName);
270 const std::vector<std::string> &ABIListFiles = std::vector<std::string>(),
271 void *(*getArgTLS)() = nullptr, void *(*getRetValTLS)() = nullptr);
273 bool doInitialization(Module &M) override;
274 bool runOnModule(Module &M) override;
277 struct DFSanFunction {
278 DataFlowSanitizer &DFS;
281 DataFlowSanitizer::InstrumentedABI IA;
285 AllocaInst *LabelReturnAlloca;
286 DenseMap<Value *, Value *> ValShadowMap;
287 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
288 std::vector<std::pair<PHINode *, PHINode *> > PHIFixups;
289 DenseSet<Instruction *> SkipInsts;
290 std::vector<Value *> NonZeroChecks;
293 struct CachedCombinedShadow {
297 DenseMap<std::pair<Value *, Value *>, CachedCombinedShadow>
298 CachedCombinedShadows;
299 DenseMap<Value *, std::set<Value *>> ShadowElements;
301 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI)
302 : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()),
303 IsNativeABI(IsNativeABI), ArgTLSPtr(nullptr), RetvalTLSPtr(nullptr),
304 LabelReturnAlloca(nullptr) {
306 // FIXME: Need to track down the register allocator issue which causes poor
307 // performance in pathological cases with large numbers of basic blocks.
308 AvoidNewBlocks = F->size() > 1000;
310 Value *getArgTLSPtr();
311 Value *getArgTLS(unsigned Index, Instruction *Pos);
312 Value *getRetvalTLS();
313 Value *getShadow(Value *V);
314 void setShadow(Instruction *I, Value *Shadow);
315 Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
316 Value *combineOperandShadows(Instruction *Inst);
317 Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align,
319 void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow,
323 class DFSanVisitor : public InstVisitor<DFSanVisitor> {
326 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
328 void visitOperandShadowInst(Instruction &I);
330 void visitBinaryOperator(BinaryOperator &BO);
331 void visitCastInst(CastInst &CI);
332 void visitCmpInst(CmpInst &CI);
333 void visitGetElementPtrInst(GetElementPtrInst &GEPI);
334 void visitLoadInst(LoadInst &LI);
335 void visitStoreInst(StoreInst &SI);
336 void visitReturnInst(ReturnInst &RI);
337 void visitCallSite(CallSite CS);
338 void visitPHINode(PHINode &PN);
339 void visitExtractElementInst(ExtractElementInst &I);
340 void visitInsertElementInst(InsertElementInst &I);
341 void visitShuffleVectorInst(ShuffleVectorInst &I);
342 void visitExtractValueInst(ExtractValueInst &I);
343 void visitInsertValueInst(InsertValueInst &I);
344 void visitAllocaInst(AllocaInst &I);
345 void visitSelectInst(SelectInst &I);
346 void visitMemSetInst(MemSetInst &I);
347 void visitMemTransferInst(MemTransferInst &I);
352 char DataFlowSanitizer::ID;
353 INITIALIZE_PASS(DataFlowSanitizer, "dfsan",
354 "DataFlowSanitizer: dynamic data flow analysis.", false, false)
357 llvm::createDataFlowSanitizerPass(const std::vector<std::string> &ABIListFiles,
358 void *(*getArgTLS)(),
359 void *(*getRetValTLS)()) {
360 return new DataFlowSanitizer(ABIListFiles, getArgTLS, getRetValTLS);
363 DataFlowSanitizer::DataFlowSanitizer(
364 const std::vector<std::string> &ABIListFiles, void *(*getArgTLS)(),
365 void *(*getRetValTLS)())
366 : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS) {
367 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles));
368 AllABIListFiles.insert(AllABIListFiles.end(), ClABIListFiles.begin(),
369 ClABIListFiles.end());
370 ABIList.set(SpecialCaseList::createOrDie(AllABIListFiles));
373 FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) {
374 llvm::SmallVector<Type *, 4> ArgTypes(T->param_begin(), T->param_end());
375 ArgTypes.append(T->getNumParams(), ShadowTy);
377 ArgTypes.push_back(ShadowPtrTy);
378 Type *RetType = T->getReturnType();
379 if (!RetType->isVoidTy())
380 RetType = StructType::get(RetType, ShadowTy, (Type *)nullptr);
381 return FunctionType::get(RetType, ArgTypes, T->isVarArg());
384 FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) {
385 assert(!T->isVarArg());
386 llvm::SmallVector<Type *, 4> ArgTypes;
387 ArgTypes.push_back(T->getPointerTo());
388 ArgTypes.append(T->param_begin(), T->param_end());
389 ArgTypes.append(T->getNumParams(), ShadowTy);
390 Type *RetType = T->getReturnType();
391 if (!RetType->isVoidTy())
392 ArgTypes.push_back(ShadowPtrTy);
393 return FunctionType::get(T->getReturnType(), ArgTypes, false);
396 FunctionType *DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
397 llvm::SmallVector<Type *, 4> ArgTypes;
398 for (FunctionType::param_iterator i = T->param_begin(), e = T->param_end();
401 if (isa<PointerType>(*i) && (FT = dyn_cast<FunctionType>(cast<PointerType>(
402 *i)->getElementType()))) {
403 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo());
404 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx));
406 ArgTypes.push_back(*i);
409 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
410 ArgTypes.push_back(ShadowTy);
412 ArgTypes.push_back(ShadowPtrTy);
413 Type *RetType = T->getReturnType();
414 if (!RetType->isVoidTy())
415 ArgTypes.push_back(ShadowPtrTy);
416 return FunctionType::get(T->getReturnType(), ArgTypes, T->isVarArg());
419 bool DataFlowSanitizer::doInitialization(Module &M) {
420 llvm::Triple TargetTriple(M.getTargetTriple());
421 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
422 bool IsMIPS64 = TargetTriple.getArch() == llvm::Triple::mips64 ||
423 TargetTriple.getArch() == llvm::Triple::mips64el;
425 DL = &M.getDataLayout();
428 Ctx = &M.getContext();
429 ShadowTy = IntegerType::get(*Ctx, ShadowWidth);
430 ShadowPtrTy = PointerType::getUnqual(ShadowTy);
431 IntptrTy = DL->getIntPtrType(*Ctx);
432 ZeroShadow = ConstantInt::getSigned(ShadowTy, 0);
433 ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8);
435 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL);
437 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0xF000000000LL);
439 report_fatal_error("unsupported triple");
441 Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy };
443 FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false);
444 Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy };
446 FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false);
447 DFSanUnimplementedFnTy = FunctionType::get(
448 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
449 Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy };
450 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx),
451 DFSanSetLabelArgs, /*isVarArg=*/false);
452 DFSanNonzeroLabelFnTy = FunctionType::get(
453 Type::getVoidTy(*Ctx), None, /*isVarArg=*/false);
454 DFSanVarargWrapperFnTy = FunctionType::get(
455 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
458 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
460 GetArgTLS = ConstantExpr::getIntToPtr(
461 ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)),
462 PointerType::getUnqual(
463 FunctionType::get(PointerType::getUnqual(ArgTLSTy),
466 if (GetRetvalTLSPtr) {
468 GetRetvalTLS = ConstantExpr::getIntToPtr(
469 ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)),
470 PointerType::getUnqual(
471 FunctionType::get(PointerType::getUnqual(ShadowTy),
475 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
479 bool DataFlowSanitizer::isInstrumented(const Function *F) {
480 return !ABIList.isIn(*F, "uninstrumented");
483 bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
484 return !ABIList.isIn(*GA, "uninstrumented");
487 DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() {
488 return ClArgsABI ? IA_Args : IA_TLS;
491 DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
492 if (ABIList.isIn(*F, "functional"))
493 return WK_Functional;
494 if (ABIList.isIn(*F, "discard"))
496 if (ABIList.isIn(*F, "custom"))
502 void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) {
503 std::string GVName = GV->getName(), Prefix = "dfs$";
504 GV->setName(Prefix + GVName);
506 // Try to change the name of the function in module inline asm. We only do
507 // this for specific asm directives, currently only ".symver", to try to avoid
508 // corrupting asm which happens to contain the symbol name as a substring.
509 // Note that the substitution for .symver assumes that the versioned symbol
510 // also has an instrumented name.
511 std::string Asm = GV->getParent()->getModuleInlineAsm();
512 std::string SearchStr = ".symver " + GVName + ",";
513 size_t Pos = Asm.find(SearchStr);
514 if (Pos != std::string::npos) {
515 Asm.replace(Pos, SearchStr.size(),
516 ".symver " + Prefix + GVName + "," + Prefix);
517 GV->getParent()->setModuleInlineAsm(Asm);
522 DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
523 GlobalValue::LinkageTypes NewFLink,
524 FunctionType *NewFT) {
525 FunctionType *FT = F->getFunctionType();
526 Function *NewF = Function::Create(NewFT, NewFLink, NewFName,
528 NewF->copyAttributesFrom(F);
529 NewF->removeAttributes(
530 AttributeSet::ReturnIndex,
531 AttributeFuncs::typeIncompatible(NewFT->getReturnType(),
532 AttributeSet::ReturnIndex));
534 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF);
536 NewF->removeAttributes(
537 AttributeSet::FunctionIndex,
538 AttributeSet().addAttribute(*Ctx, AttributeSet::FunctionIndex,
540 CallInst::Create(DFSanVarargWrapperFn,
541 IRBuilder<>(BB).CreateGlobalStringPtr(F->getName()), "",
543 new UnreachableInst(*Ctx, BB);
545 std::vector<Value *> Args;
546 unsigned n = FT->getNumParams();
547 for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n)
548 Args.push_back(&*ai);
549 CallInst *CI = CallInst::Create(F, Args, "", BB);
550 if (FT->getReturnType()->isVoidTy())
551 ReturnInst::Create(*Ctx, BB);
553 ReturnInst::Create(*Ctx, CI, BB);
559 Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT,
561 FunctionType *FTT = getTrampolineFunctionType(FT);
562 Constant *C = Mod->getOrInsertFunction(FName, FTT);
563 Function *F = dyn_cast<Function>(C);
564 if (F && F->isDeclaration()) {
565 F->setLinkage(GlobalValue::LinkOnceODRLinkage);
566 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
567 std::vector<Value *> Args;
568 Function::arg_iterator AI = F->arg_begin(); ++AI;
569 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N)
570 Args.push_back(&*AI);
572 CallInst::Create(&F->getArgumentList().front(), Args, "", BB);
574 if (FT->getReturnType()->isVoidTy())
575 RI = ReturnInst::Create(*Ctx, BB);
577 RI = ReturnInst::Create(*Ctx, CI, BB);
579 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true);
580 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI;
581 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N)
582 DFSF.ValShadowMap[ValAI] = ShadowAI;
583 DFSanVisitor(DFSF).visitCallInst(*CI);
584 if (!FT->getReturnType()->isVoidTy())
585 new StoreInst(DFSF.getShadow(RI->getReturnValue()),
586 &F->getArgumentList().back(), RI);
592 bool DataFlowSanitizer::runOnModule(Module &M) {
594 if (ABIList.isIn(M, "skip"))
597 FunctionDIs = makeSubprogramMap(M);
600 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
601 ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy);
602 if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS))
603 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
605 if (!GetRetvalTLSPtr) {
606 RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy);
607 if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS))
608 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
611 DFSanUnionFn = Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy);
612 if (Function *F = dyn_cast<Function>(DFSanUnionFn)) {
613 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
614 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone);
615 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
616 F->addAttribute(1, Attribute::ZExt);
617 F->addAttribute(2, Attribute::ZExt);
619 DFSanCheckedUnionFn = Mod->getOrInsertFunction("dfsan_union", DFSanUnionFnTy);
620 if (Function *F = dyn_cast<Function>(DFSanCheckedUnionFn)) {
621 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
622 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone);
623 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
624 F->addAttribute(1, Attribute::ZExt);
625 F->addAttribute(2, Attribute::ZExt);
628 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy);
629 if (Function *F = dyn_cast<Function>(DFSanUnionLoadFn)) {
630 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
631 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadOnly);
632 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
634 DFSanUnimplementedFn =
635 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy);
637 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy);
638 if (Function *F = dyn_cast<Function>(DFSanSetLabelFn)) {
639 F->addAttribute(1, Attribute::ZExt);
641 DFSanNonzeroLabelFn =
642 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy);
643 DFSanVarargWrapperFn = Mod->getOrInsertFunction("__dfsan_vararg_wrapper",
644 DFSanVarargWrapperFnTy);
646 std::vector<Function *> FnsToInstrument;
647 llvm::SmallPtrSet<Function *, 2> FnsWithNativeABI;
648 for (Module::iterator i = M.begin(), e = M.end(); i != e; ++i) {
649 if (!i->isIntrinsic() &&
651 i != DFSanCheckedUnionFn &&
652 i != DFSanUnionLoadFn &&
653 i != DFSanUnimplementedFn &&
654 i != DFSanSetLabelFn &&
655 i != DFSanNonzeroLabelFn &&
656 i != DFSanVarargWrapperFn)
657 FnsToInstrument.push_back(&*i);
660 // Give function aliases prefixes when necessary, and build wrappers where the
661 // instrumentedness is inconsistent.
662 for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) {
663 GlobalAlias *GA = &*i;
665 // Don't stop on weak. We assume people aren't playing games with the
666 // instrumentedness of overridden weak aliases.
667 if (auto F = dyn_cast<Function>(GA->getBaseObject())) {
668 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F);
669 if (GAInst && FInst) {
670 addGlobalNamePrefix(GA);
671 } else if (GAInst != FInst) {
672 // Non-instrumented alias of an instrumented function, or vice versa.
673 // Replace the alias with a native-ABI wrapper of the aliasee. The pass
674 // below will take care of instrumenting it.
676 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType());
677 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType()));
679 GA->eraseFromParent();
680 FnsToInstrument.push_back(NewF);
686 B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
687 ReadOnlyNoneAttrs = AttributeSet::get(*Ctx, AttributeSet::FunctionIndex, B);
689 // First, change the ABI of every function in the module. ABI-listed
690 // functions keep their original ABI and get a wrapper function.
691 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
692 e = FnsToInstrument.end();
695 FunctionType *FT = F.getFunctionType();
697 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
698 FT->getReturnType()->isVoidTy());
700 if (isInstrumented(&F)) {
701 // Instrumented functions get a 'dfs$' prefix. This allows us to more
702 // easily identify cases of mismatching ABIs.
703 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) {
704 FunctionType *NewFT = getArgsFunctionType(FT);
705 Function *NewF = Function::Create(NewFT, F.getLinkage(), "", &M);
706 NewF->copyAttributesFrom(&F);
707 NewF->removeAttributes(
708 AttributeSet::ReturnIndex,
709 AttributeFuncs::typeIncompatible(NewFT->getReturnType(),
710 AttributeSet::ReturnIndex));
711 for (Function::arg_iterator FArg = F.arg_begin(),
712 NewFArg = NewF->arg_begin(),
713 FArgEnd = F.arg_end();
714 FArg != FArgEnd; ++FArg, ++NewFArg) {
715 FArg->replaceAllUsesWith(NewFArg);
717 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList());
719 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end();
721 BlockAddress *BA = dyn_cast<BlockAddress>(*UI);
724 BA->replaceAllUsesWith(
725 BlockAddress::get(NewF, BA->getBasicBlock()));
729 F.replaceAllUsesWith(
730 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)));
734 addGlobalNamePrefix(NewF);
736 addGlobalNamePrefix(&F);
738 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) {
739 // Build a wrapper function for F. The wrapper simply calls F, and is
740 // added to FnsToInstrument so that any instrumentation according to its
741 // WrapperKind is done in the second pass below.
742 FunctionType *NewFT = getInstrumentedABI() == IA_Args
743 ? getArgsFunctionType(FT)
745 Function *NewF = buildWrapperFunction(
746 &F, std::string("dfsw$") + std::string(F.getName()),
747 GlobalValue::LinkOnceODRLinkage, NewFT);
748 if (getInstrumentedABI() == IA_TLS)
749 NewF->removeAttributes(AttributeSet::FunctionIndex, ReadOnlyNoneAttrs);
751 Value *WrappedFnCst =
752 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT));
753 F.replaceAllUsesWith(WrappedFnCst);
755 // Patch the pointer to LLVM function in debug info descriptor.
756 auto DI = FunctionDIs.find(&F);
757 if (DI != FunctionDIs.end())
758 DI->second.replaceFunction(&F);
760 UnwrappedFnMap[WrappedFnCst] = &F;
763 if (!F.isDeclaration()) {
764 // This function is probably defining an interposition of an
765 // uninstrumented function and hence needs to keep the original ABI.
766 // But any functions it may call need to use the instrumented ABI, so
767 // we instrument it in a mode which preserves the original ABI.
768 FnsWithNativeABI.insert(&F);
770 // This code needs to rebuild the iterators, as they may be invalidated
771 // by the push_back, taking care that the new range does not include
772 // any functions added by this code.
773 size_t N = i - FnsToInstrument.begin(),
774 Count = e - FnsToInstrument.begin();
775 FnsToInstrument.push_back(&F);
776 i = FnsToInstrument.begin() + N;
777 e = FnsToInstrument.begin() + Count;
779 // Hopefully, nobody will try to indirectly call a vararg
781 } else if (FT->isVarArg()) {
782 UnwrappedFnMap[&F] = &F;
787 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
788 e = FnsToInstrument.end();
790 if (!*i || (*i)->isDeclaration())
793 removeUnreachableBlocks(**i);
795 DFSanFunction DFSF(*this, *i, FnsWithNativeABI.count(*i));
797 // DFSanVisitor may create new basic blocks, which confuses df_iterator.
798 // Build a copy of the list before iterating over it.
799 llvm::SmallVector<BasicBlock *, 4> BBList(
800 depth_first(&(*i)->getEntryBlock()));
802 for (llvm::SmallVector<BasicBlock *, 4>::iterator i = BBList.begin(),
805 Instruction *Inst = &(*i)->front();
807 // DFSanVisitor may split the current basic block, changing the current
808 // instruction's next pointer and moving the next instruction to the
809 // tail block from which we should continue.
810 Instruction *Next = Inst->getNextNode();
811 // DFSanVisitor may delete Inst, so keep track of whether it was a
813 bool IsTerminator = isa<TerminatorInst>(Inst);
814 if (!DFSF.SkipInsts.count(Inst))
815 DFSanVisitor(DFSF).visit(Inst);
822 // We will not necessarily be able to compute the shadow for every phi node
823 // until we have visited every block. Therefore, the code that handles phi
824 // nodes adds them to the PHIFixups list so that they can be properly
826 for (std::vector<std::pair<PHINode *, PHINode *> >::iterator
827 i = DFSF.PHIFixups.begin(),
828 e = DFSF.PHIFixups.end();
830 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n;
832 i->second->setIncomingValue(
833 val, DFSF.getShadow(i->first->getIncomingValue(val)));
837 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
838 // places (i.e. instructions in basic blocks we haven't even begun visiting
839 // yet). To make our life easier, do this work in a pass after the main
841 if (ClDebugNonzeroLabels) {
842 for (Value *V : DFSF.NonZeroChecks) {
844 if (Instruction *I = dyn_cast<Instruction>(V))
845 Pos = I->getNextNode();
847 Pos = DFSF.F->getEntryBlock().begin();
848 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos))
849 Pos = Pos->getNextNode();
850 IRBuilder<> IRB(Pos);
851 Value *Ne = IRB.CreateICmpNE(V, DFSF.DFS.ZeroShadow);
852 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
853 Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
854 IRBuilder<> ThenIRB(BI);
855 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn);
863 Value *DFSanFunction::getArgTLSPtr() {
867 return ArgTLSPtr = DFS.ArgTLS;
869 IRBuilder<> IRB(F->getEntryBlock().begin());
870 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS);
873 Value *DFSanFunction::getRetvalTLS() {
877 return RetvalTLSPtr = DFS.RetvalTLS;
879 IRBuilder<> IRB(F->getEntryBlock().begin());
880 return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS);
883 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) {
884 IRBuilder<> IRB(Pos);
885 return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx);
888 Value *DFSanFunction::getShadow(Value *V) {
889 if (!isa<Argument>(V) && !isa<Instruction>(V))
890 return DFS.ZeroShadow;
891 Value *&Shadow = ValShadowMap[V];
893 if (Argument *A = dyn_cast<Argument>(V)) {
895 return DFS.ZeroShadow;
897 case DataFlowSanitizer::IA_TLS: {
898 Value *ArgTLSPtr = getArgTLSPtr();
899 Instruction *ArgTLSPos =
900 DFS.ArgTLS ? &*F->getEntryBlock().begin()
901 : cast<Instruction>(ArgTLSPtr)->getNextNode();
902 IRBuilder<> IRB(ArgTLSPos);
903 Shadow = IRB.CreateLoad(getArgTLS(A->getArgNo(), ArgTLSPos));
906 case DataFlowSanitizer::IA_Args: {
907 unsigned ArgIdx = A->getArgNo() + F->getArgumentList().size() / 2;
908 Function::arg_iterator i = F->arg_begin();
912 assert(Shadow->getType() == DFS.ShadowTy);
916 NonZeroChecks.push_back(Shadow);
918 Shadow = DFS.ZeroShadow;
924 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
925 assert(!ValShadowMap.count(I));
926 assert(Shadow->getType() == DFS.ShadowTy);
927 ValShadowMap[I] = Shadow;
930 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
931 assert(Addr != RetvalTLS && "Reinstrumenting?");
932 IRBuilder<> IRB(Pos);
933 return IRB.CreateIntToPtr(
935 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy), ShadowPtrMask),
940 // Generates IR to compute the union of the two given shadows, inserting it
941 // before Pos. Returns the computed union Value.
942 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) {
943 if (V1 == DFS.ZeroShadow)
945 if (V2 == DFS.ZeroShadow)
950 auto V1Elems = ShadowElements.find(V1);
951 auto V2Elems = ShadowElements.find(V2);
952 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) {
953 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(),
954 V2Elems->second.begin(), V2Elems->second.end())) {
956 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(),
957 V1Elems->second.begin(), V1Elems->second.end())) {
960 } else if (V1Elems != ShadowElements.end()) {
961 if (V1Elems->second.count(V2))
963 } else if (V2Elems != ShadowElements.end()) {
964 if (V2Elems->second.count(V1))
968 auto Key = std::make_pair(V1, V2);
970 std::swap(Key.first, Key.second);
971 CachedCombinedShadow &CCS = CachedCombinedShadows[Key];
972 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent()))
975 IRBuilder<> IRB(Pos);
976 if (AvoidNewBlocks) {
977 CallInst *Call = IRB.CreateCall2(DFS.DFSanCheckedUnionFn, V1, V2);
978 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
979 Call->addAttribute(1, Attribute::ZExt);
980 Call->addAttribute(2, Attribute::ZExt);
982 CCS.Block = Pos->getParent();
985 BasicBlock *Head = Pos->getParent();
986 Value *Ne = IRB.CreateICmpNE(V1, V2);
987 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
988 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT));
989 IRBuilder<> ThenIRB(BI);
990 CallInst *Call = ThenIRB.CreateCall2(DFS.DFSanUnionFn, V1, V2);
991 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
992 Call->addAttribute(1, Attribute::ZExt);
993 Call->addAttribute(2, Attribute::ZExt);
995 BasicBlock *Tail = BI->getSuccessor(0);
996 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", Tail->begin());
997 Phi->addIncoming(Call, Call->getParent());
998 Phi->addIncoming(V1, Head);
1004 std::set<Value *> UnionElems;
1005 if (V1Elems != ShadowElements.end()) {
1006 UnionElems = V1Elems->second;
1008 UnionElems.insert(V1);
1010 if (V2Elems != ShadowElements.end()) {
1011 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end());
1013 UnionElems.insert(V2);
1015 ShadowElements[CCS.Shadow] = std::move(UnionElems);
1020 // A convenience function which folds the shadows of each of the operands
1021 // of the provided instruction Inst, inserting the IR before Inst. Returns
1022 // the computed union Value.
1023 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
1024 if (Inst->getNumOperands() == 0)
1025 return DFS.ZeroShadow;
1027 Value *Shadow = getShadow(Inst->getOperand(0));
1028 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) {
1029 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst);
1034 void DFSanVisitor::visitOperandShadowInst(Instruction &I) {
1035 Value *CombinedShadow = DFSF.combineOperandShadows(&I);
1036 DFSF.setShadow(&I, CombinedShadow);
1039 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where
1040 // Addr has alignment Align, and take the union of each of those shadows.
1041 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align,
1043 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1044 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1045 AllocaShadowMap.find(AI);
1046 if (i != AllocaShadowMap.end()) {
1047 IRBuilder<> IRB(Pos);
1048 return IRB.CreateLoad(i->second);
1052 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1053 SmallVector<Value *, 2> Objs;
1054 GetUnderlyingObjects(Addr, Objs, DFS.DL);
1055 bool AllConstants = true;
1056 for (SmallVector<Value *, 2>::iterator i = Objs.begin(), e = Objs.end();
1058 if (isa<Function>(*i) || isa<BlockAddress>(*i))
1060 if (isa<GlobalVariable>(*i) && cast<GlobalVariable>(*i)->isConstant())
1063 AllConstants = false;
1067 return DFS.ZeroShadow;
1069 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1072 return DFS.ZeroShadow;
1074 LoadInst *LI = new LoadInst(ShadowAddr, "", Pos);
1075 LI->setAlignment(ShadowAlign);
1079 IRBuilder<> IRB(Pos);
1080 Value *ShadowAddr1 =
1081 IRB.CreateGEP(ShadowAddr, ConstantInt::get(DFS.IntptrTy, 1));
1082 return combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign),
1083 IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign), Pos);
1086 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) {
1087 // Fast path for the common case where each byte has identical shadow: load
1088 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any
1089 // shadow is non-equal.
1090 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F);
1091 IRBuilder<> FallbackIRB(FallbackBB);
1092 CallInst *FallbackCall = FallbackIRB.CreateCall2(
1093 DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size));
1094 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
1096 // Compare each of the shadows stored in the loaded 64 bits to each other,
1097 // by computing (WideShadow rotl ShadowWidth) == WideShadow.
1098 IRBuilder<> IRB(Pos);
1100 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx));
1101 Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1102 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy);
1103 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth);
1104 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth);
1105 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow);
1106 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow);
1108 BasicBlock *Head = Pos->getParent();
1109 BasicBlock *Tail = Head->splitBasicBlock(Pos);
1111 if (DomTreeNode *OldNode = DT.getNode(Head)) {
1112 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end());
1114 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head);
1115 for (auto Child : Children)
1116 DT.changeImmediateDominator(Child, NewNode);
1119 // In the following code LastBr will refer to the previous basic block's
1120 // conditional branch instruction, whose true successor is fixed up to point
1121 // to the next block during the loop below or to the tail after the final
1123 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq);
1124 ReplaceInstWithInst(Head->getTerminator(), LastBr);
1125 DT.addNewBlock(FallbackBB, Head);
1127 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size;
1128 Ofs += 64 / DFS.ShadowWidth) {
1129 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F);
1130 DT.addNewBlock(NextBB, LastBr->getParent());
1131 IRBuilder<> NextIRB(NextBB);
1132 WideAddr = NextIRB.CreateGEP(WideAddr, ConstantInt::get(DFS.IntptrTy, 1));
1133 Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1134 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow);
1135 LastBr->setSuccessor(0, NextBB);
1136 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB);
1139 LastBr->setSuccessor(0, Tail);
1140 FallbackIRB.CreateBr(Tail);
1141 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1142 Shadow->addIncoming(FallbackCall, FallbackBB);
1143 Shadow->addIncoming(TruncShadow, LastBr->getParent());
1147 IRBuilder<> IRB(Pos);
1148 CallInst *FallbackCall = IRB.CreateCall2(
1149 DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size));
1150 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
1151 return FallbackCall;
1154 void DFSanVisitor::visitLoadInst(LoadInst &LI) {
1155 uint64_t Size = DFSF.DFS.DL->getTypeStoreSize(LI.getType());
1157 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow);
1162 if (ClPreserveAlignment) {
1163 Align = LI.getAlignment();
1165 Align = DFSF.DFS.DL->getABITypeAlignment(LI.getType());
1169 IRBuilder<> IRB(&LI);
1170 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI);
1171 if (ClCombinePointerLabelsOnLoad) {
1172 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
1173 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI);
1175 if (Shadow != DFSF.DFS.ZeroShadow)
1176 DFSF.NonZeroChecks.push_back(Shadow);
1178 DFSF.setShadow(&LI, Shadow);
1181 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align,
1182 Value *Shadow, Instruction *Pos) {
1183 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1184 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1185 AllocaShadowMap.find(AI);
1186 if (i != AllocaShadowMap.end()) {
1187 IRBuilder<> IRB(Pos);
1188 IRB.CreateStore(Shadow, i->second);
1193 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1194 IRBuilder<> IRB(Pos);
1195 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1196 if (Shadow == DFS.ZeroShadow) {
1197 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth);
1198 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0);
1199 Value *ExtShadowAddr =
1200 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy));
1201 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign);
1205 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth;
1206 uint64_t Offset = 0;
1207 if (Size >= ShadowVecSize) {
1208 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize);
1209 Value *ShadowVec = UndefValue::get(ShadowVecTy);
1210 for (unsigned i = 0; i != ShadowVecSize; ++i) {
1211 ShadowVec = IRB.CreateInsertElement(
1212 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i));
1214 Value *ShadowVecAddr =
1215 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy));
1217 Value *CurShadowVecAddr = IRB.CreateConstGEP1_32(ShadowVecAddr, Offset);
1218 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign);
1219 Size -= ShadowVecSize;
1221 } while (Size >= ShadowVecSize);
1222 Offset *= ShadowVecSize;
1225 Value *CurShadowAddr = IRB.CreateConstGEP1_32(ShadowAddr, Offset);
1226 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign);
1232 void DFSanVisitor::visitStoreInst(StoreInst &SI) {
1234 DFSF.DFS.DL->getTypeStoreSize(SI.getValueOperand()->getType());
1239 if (ClPreserveAlignment) {
1240 Align = SI.getAlignment();
1242 Align = DFSF.DFS.DL->getABITypeAlignment(SI.getValueOperand()->getType());
1247 Value* Shadow = DFSF.getShadow(SI.getValueOperand());
1248 if (ClCombinePointerLabelsOnStore) {
1249 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand());
1250 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI);
1252 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI);
1255 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
1256 visitOperandShadowInst(BO);
1259 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); }
1261 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); }
1263 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
1264 visitOperandShadowInst(GEPI);
1267 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
1268 visitOperandShadowInst(I);
1271 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
1272 visitOperandShadowInst(I);
1275 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
1276 visitOperandShadowInst(I);
1279 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
1280 visitOperandShadowInst(I);
1283 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
1284 visitOperandShadowInst(I);
1287 void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
1288 bool AllLoadsStores = true;
1289 for (User *U : I.users()) {
1290 if (isa<LoadInst>(U))
1293 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1294 if (SI->getPointerOperand() == &I)
1298 AllLoadsStores = false;
1301 if (AllLoadsStores) {
1302 IRBuilder<> IRB(&I);
1303 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy);
1305 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow);
1308 void DFSanVisitor::visitSelectInst(SelectInst &I) {
1309 Value *CondShadow = DFSF.getShadow(I.getCondition());
1310 Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
1311 Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
1313 if (isa<VectorType>(I.getCondition()->getType())) {
1316 DFSF.combineShadows(
1317 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I));
1320 if (TrueShadow == FalseShadow) {
1321 ShadowSel = TrueShadow;
1324 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I);
1326 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I));
1330 void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
1331 IRBuilder<> IRB(&I);
1332 Value *ValShadow = DFSF.getShadow(I.getValue());
1334 DFSF.DFS.DFSanSetLabelFn, ValShadow,
1335 IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(*DFSF.DFS.Ctx)),
1336 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy));
1339 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
1340 IRBuilder<> IRB(&I);
1341 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I);
1342 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I);
1343 Value *LenShadow = IRB.CreateMul(
1345 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8));
1347 if (ClPreserveAlignment) {
1348 AlignShadow = IRB.CreateMul(I.getAlignmentCst(),
1349 ConstantInt::get(I.getAlignmentCst()->getType(),
1350 DFSF.DFS.ShadowWidth / 8));
1352 AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(),
1353 DFSF.DFS.ShadowWidth / 8);
1355 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx);
1356 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr);
1357 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr);
1358 IRB.CreateCall5(I.getCalledValue(), DestShadow, SrcShadow, LenShadow,
1359 AlignShadow, I.getVolatileCst());
1362 void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
1363 if (!DFSF.IsNativeABI && RI.getReturnValue()) {
1365 case DataFlowSanitizer::IA_TLS: {
1366 Value *S = DFSF.getShadow(RI.getReturnValue());
1367 IRBuilder<> IRB(&RI);
1368 IRB.CreateStore(S, DFSF.getRetvalTLS());
1371 case DataFlowSanitizer::IA_Args: {
1372 IRBuilder<> IRB(&RI);
1373 Type *RT = DFSF.F->getFunctionType()->getReturnType();
1375 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0);
1377 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1);
1378 RI.setOperand(0, InsShadow);
1385 void DFSanVisitor::visitCallSite(CallSite CS) {
1386 Function *F = CS.getCalledFunction();
1387 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) {
1388 visitOperandShadowInst(*CS.getInstruction());
1392 // Calls to this function are synthesized in wrappers, and we shouldn't
1394 if (F == DFSF.DFS.DFSanVarargWrapperFn)
1397 assert(!(cast<FunctionType>(
1398 CS.getCalledValue()->getType()->getPointerElementType())->isVarArg() &&
1399 dyn_cast<InvokeInst>(CS.getInstruction())));
1401 IRBuilder<> IRB(CS.getInstruction());
1403 DenseMap<Value *, Function *>::iterator i =
1404 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue());
1405 if (i != DFSF.DFS.UnwrappedFnMap.end()) {
1406 Function *F = i->second;
1407 switch (DFSF.DFS.getWrapperKind(F)) {
1408 case DataFlowSanitizer::WK_Warning: {
1409 CS.setCalledFunction(F);
1410 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn,
1411 IRB.CreateGlobalStringPtr(F->getName()));
1412 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1415 case DataFlowSanitizer::WK_Discard: {
1416 CS.setCalledFunction(F);
1417 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1420 case DataFlowSanitizer::WK_Functional: {
1421 CS.setCalledFunction(F);
1422 visitOperandShadowInst(*CS.getInstruction());
1425 case DataFlowSanitizer::WK_Custom: {
1426 // Don't try to handle invokes of custom functions, it's too complicated.
1427 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
1429 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1430 FunctionType *FT = F->getFunctionType();
1431 FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT);
1432 std::string CustomFName = "__dfsw_";
1433 CustomFName += F->getName();
1435 DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT);
1436 if (Function *CustomFn = dyn_cast<Function>(CustomF)) {
1437 CustomFn->copyAttributesFrom(F);
1439 // Custom functions returning non-void will write to the return label.
1440 if (!FT->getReturnType()->isVoidTy()) {
1441 CustomFn->removeAttributes(AttributeSet::FunctionIndex,
1442 DFSF.DFS.ReadOnlyNoneAttrs);
1446 std::vector<Value *> Args;
1448 CallSite::arg_iterator i = CS.arg_begin();
1449 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) {
1450 Type *T = (*i)->getType();
1451 FunctionType *ParamFT;
1452 if (isa<PointerType>(T) &&
1453 (ParamFT = dyn_cast<FunctionType>(
1454 cast<PointerType>(T)->getElementType()))) {
1455 std::string TName = "dfst";
1456 TName += utostr(FT->getNumParams() - n);
1458 TName += F->getName();
1459 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName);
1462 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx)));
1469 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1470 Args.push_back(DFSF.getShadow(*i));
1472 if (FT->isVarArg()) {
1473 auto LabelVAAlloca =
1474 new AllocaInst(ArrayType::get(DFSF.DFS.ShadowTy,
1475 CS.arg_size() - FT->getNumParams()),
1476 "labelva", DFSF.F->getEntryBlock().begin());
1478 for (unsigned n = 0; i != CS.arg_end(); ++i, ++n) {
1479 auto LabelVAPtr = IRB.CreateStructGEP(LabelVAAlloca, n);
1480 IRB.CreateStore(DFSF.getShadow(*i), LabelVAPtr);
1483 Args.push_back(IRB.CreateStructGEP(LabelVAAlloca, 0));
1486 if (!FT->getReturnType()->isVoidTy()) {
1487 if (!DFSF.LabelReturnAlloca) {
1488 DFSF.LabelReturnAlloca =
1489 new AllocaInst(DFSF.DFS.ShadowTy, "labelreturn",
1490 DFSF.F->getEntryBlock().begin());
1492 Args.push_back(DFSF.LabelReturnAlloca);
1495 for (i = CS.arg_begin() + FT->getNumParams(); i != CS.arg_end(); ++i)
1498 CallInst *CustomCI = IRB.CreateCall(CustomF, Args);
1499 CustomCI->setCallingConv(CI->getCallingConv());
1500 CustomCI->setAttributes(CI->getAttributes());
1502 if (!FT->getReturnType()->isVoidTy()) {
1503 LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca);
1504 DFSF.setShadow(CustomCI, LabelLoad);
1507 CI->replaceAllUsesWith(CustomCI);
1508 CI->eraseFromParent();
1516 FunctionType *FT = cast<FunctionType>(
1517 CS.getCalledValue()->getType()->getPointerElementType());
1518 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1519 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) {
1520 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)),
1521 DFSF.getArgTLS(i, CS.getInstruction()));
1525 Instruction *Next = nullptr;
1526 if (!CS.getType()->isVoidTy()) {
1527 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1528 if (II->getNormalDest()->getSinglePredecessor()) {
1529 Next = II->getNormalDest()->begin();
1532 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT);
1533 Next = NewBB->begin();
1536 Next = CS->getNextNode();
1539 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1540 IRBuilder<> NextIRB(Next);
1541 LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS());
1542 DFSF.SkipInsts.insert(LI);
1543 DFSF.setShadow(CS.getInstruction(), LI);
1544 DFSF.NonZeroChecks.push_back(LI);
1548 // Do all instrumentation for IA_Args down here to defer tampering with the
1549 // CFG in a way that SplitEdge may be able to detect.
1550 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) {
1551 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT);
1553 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT));
1554 std::vector<Value *> Args;
1556 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1557 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1561 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1562 Args.push_back(DFSF.getShadow(*i));
1564 if (FT->isVarArg()) {
1565 unsigned VarArgSize = CS.arg_size() - FT->getNumParams();
1566 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize);
1567 AllocaInst *VarArgShadow =
1568 new AllocaInst(VarArgArrayTy, "", DFSF.F->getEntryBlock().begin());
1569 Args.push_back(IRB.CreateConstGEP2_32(VarArgShadow, 0, 0));
1570 for (unsigned n = 0; i != e; ++i, ++n) {
1571 IRB.CreateStore(DFSF.getShadow(*i),
1572 IRB.CreateConstGEP2_32(VarArgShadow, 0, n));
1578 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1579 NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(),
1582 NewCS = IRB.CreateCall(Func, Args);
1584 NewCS.setCallingConv(CS.getCallingConv());
1585 NewCS.setAttributes(CS.getAttributes().removeAttributes(
1586 *DFSF.DFS.Ctx, AttributeSet::ReturnIndex,
1587 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType(),
1588 AttributeSet::ReturnIndex)));
1591 ExtractValueInst *ExVal =
1592 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next);
1593 DFSF.SkipInsts.insert(ExVal);
1594 ExtractValueInst *ExShadow =
1595 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next);
1596 DFSF.SkipInsts.insert(ExShadow);
1597 DFSF.setShadow(ExVal, ExShadow);
1598 DFSF.NonZeroChecks.push_back(ExShadow);
1600 CS.getInstruction()->replaceAllUsesWith(ExVal);
1603 CS.getInstruction()->eraseFromParent();
1607 void DFSanVisitor::visitPHINode(PHINode &PN) {
1609 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN);
1611 // Give the shadow phi node valid predecessors to fool SplitEdge into working.
1612 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy);
1613 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e;
1615 ShadowPN->addIncoming(UndefShadow, *i);
1618 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN));
1619 DFSF.setShadow(&PN, ShadowPN);