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/Analysis/ValueTracking.h"
53 #include "llvm/IR/Dominators.h"
54 #include "llvm/IR/IRBuilder.h"
55 #include "llvm/IR/InlineAsm.h"
56 #include "llvm/IR/InstVisitor.h"
57 #include "llvm/IR/LLVMContext.h"
58 #include "llvm/IR/MDBuilder.h"
59 #include "llvm/IR/Type.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/Pass.h"
62 #include "llvm/Support/CommandLine.h"
63 #include "llvm/Support/SpecialCaseList.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/Transforms/Utils/Local.h"
73 // The -dfsan-preserve-alignment flag controls whether this pass assumes that
74 // alignment requirements provided by the input IR are correct. For example,
75 // if the input IR contains a load with alignment 8, this flag will cause
76 // the shadow load to have alignment 16. This flag is disabled by default as
77 // we have unfortunately encountered too much code (including Clang itself;
78 // see PR14291) which performs misaligned access.
79 static cl::opt<bool> ClPreserveAlignment(
80 "dfsan-preserve-alignment",
81 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
84 // The ABI list file controls how shadow parameters are passed. The pass treats
85 // every function labelled "uninstrumented" in the ABI list file as conforming
86 // to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains
87 // additional annotations for those functions, a call to one of those functions
88 // will produce a warning message, as the labelling behaviour of the function is
89 // unknown. The other supported annotations are "functional" and "discard",
90 // which are described below under DataFlowSanitizer::WrapperKind.
91 static cl::opt<std::string> ClABIListFile(
93 cl::desc("File listing native ABI functions and how the pass treats them"),
96 // Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented
97 // functions (see DataFlowSanitizer::InstrumentedABI below).
98 static cl::opt<bool> ClArgsABI(
100 cl::desc("Use the argument ABI rather than the TLS ABI"),
103 // Controls whether the pass includes or ignores the labels of pointers in load
105 static cl::opt<bool> ClCombinePointerLabelsOnLoad(
106 "dfsan-combine-pointer-labels-on-load",
107 cl::desc("Combine the label of the pointer with the label of the data when "
108 "loading from memory."),
109 cl::Hidden, cl::init(true));
111 // Controls whether the pass includes or ignores the labels of pointers in
112 // stores instructions.
113 static cl::opt<bool> ClCombinePointerLabelsOnStore(
114 "dfsan-combine-pointer-labels-on-store",
115 cl::desc("Combine the label of the pointer with the label of the data when "
116 "storing in memory."),
117 cl::Hidden, cl::init(false));
119 static cl::opt<bool> ClDebugNonzeroLabels(
120 "dfsan-debug-nonzero-labels",
121 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
122 "load or return with a nonzero label"),
127 StringRef GetGlobalTypeString(const GlobalValue &G) {
128 // Types of GlobalVariables are always pointer types.
129 Type *GType = G.getType()->getElementType();
130 // For now we support blacklisting struct types only.
131 if (StructType *SGType = dyn_cast<StructType>(GType)) {
132 if (!SGType->isLiteral())
133 return SGType->getName();
135 return "<unknown type>";
139 std::unique_ptr<SpecialCaseList> SCL;
142 DFSanABIList(SpecialCaseList *SCL) : SCL(SCL) {}
144 /// Returns whether either this function or its source file are listed in the
146 bool isIn(const Function &F, const StringRef Category) const {
147 return isIn(*F.getParent(), Category) ||
148 SCL->inSection("fun", F.getName(), Category);
151 /// Returns whether this global alias is listed in the given category.
153 /// If GA aliases a function, the alias's name is matched as a function name
154 /// would be. Similarly, aliases of globals are matched like globals.
155 bool isIn(const GlobalAlias &GA, const StringRef Category) const {
156 if (isIn(*GA.getParent(), Category))
159 if (isa<FunctionType>(GA.getType()->getElementType()))
160 return SCL->inSection("fun", GA.getName(), Category);
162 return SCL->inSection("global", GA.getName(), Category) ||
163 SCL->inSection("type", GetGlobalTypeString(GA), Category);
166 /// Returns whether this module is listed in the given category.
167 bool isIn(const Module &M, const StringRef Category) const {
168 return SCL->inSection("src", M.getModuleIdentifier(), Category);
172 class DataFlowSanitizer : public ModulePass {
173 friend struct DFSanFunction;
174 friend class DFSanVisitor;
180 /// Which ABI should be used for instrumented functions?
181 enum InstrumentedABI {
182 /// Argument and return value labels are passed through additional
183 /// arguments and by modifying the return type.
186 /// Argument and return value labels are passed through TLS variables
187 /// __dfsan_arg_tls and __dfsan_retval_tls.
191 /// How should calls to uninstrumented functions be handled?
193 /// This function is present in an uninstrumented form but we don't know
194 /// how it should be handled. Print a warning and call the function anyway.
195 /// Don't label the return value.
198 /// This function does not write to (user-accessible) memory, and its return
199 /// value is unlabelled.
202 /// This function does not write to (user-accessible) memory, and the label
203 /// of its return value is the union of the label of its arguments.
206 /// Instead of calling the function, a custom wrapper __dfsw_F is called,
207 /// where F is the name of the function. This function may wrap the
208 /// original function or provide its own implementation. This is similar to
209 /// the IA_Args ABI, except that IA_Args uses a struct return type to
210 /// pass the return value shadow in a register, while WK_Custom uses an
211 /// extra pointer argument to return the shadow. This allows the wrapped
212 /// form of the function type to be expressed in C.
216 const DataLayout *DL;
219 IntegerType *ShadowTy;
220 PointerType *ShadowPtrTy;
221 IntegerType *IntptrTy;
222 ConstantInt *ZeroShadow;
223 ConstantInt *ShadowPtrMask;
224 ConstantInt *ShadowPtrMul;
227 void *(*GetArgTLSPtr)();
228 void *(*GetRetvalTLSPtr)();
230 Constant *GetRetvalTLS;
231 FunctionType *DFSanUnionFnTy;
232 FunctionType *DFSanUnionLoadFnTy;
233 FunctionType *DFSanUnimplementedFnTy;
234 FunctionType *DFSanSetLabelFnTy;
235 FunctionType *DFSanNonzeroLabelFnTy;
236 Constant *DFSanUnionFn;
237 Constant *DFSanCheckedUnionFn;
238 Constant *DFSanUnionLoadFn;
239 Constant *DFSanUnimplementedFn;
240 Constant *DFSanSetLabelFn;
241 Constant *DFSanNonzeroLabelFn;
242 MDNode *ColdCallWeights;
243 DFSanABIList ABIList;
244 DenseMap<Value *, Function *> UnwrappedFnMap;
245 AttributeSet ReadOnlyNoneAttrs;
247 Value *getShadowAddress(Value *Addr, Instruction *Pos);
248 bool isInstrumented(const Function *F);
249 bool isInstrumented(const GlobalAlias *GA);
250 FunctionType *getArgsFunctionType(FunctionType *T);
251 FunctionType *getTrampolineFunctionType(FunctionType *T);
252 FunctionType *getCustomFunctionType(FunctionType *T);
253 InstrumentedABI getInstrumentedABI();
254 WrapperKind getWrapperKind(Function *F);
255 void addGlobalNamePrefix(GlobalValue *GV);
256 Function *buildWrapperFunction(Function *F, StringRef NewFName,
257 GlobalValue::LinkageTypes NewFLink,
258 FunctionType *NewFT);
259 Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName);
262 DataFlowSanitizer(StringRef ABIListFile = StringRef(),
263 void *(*getArgTLS)() = nullptr,
264 void *(*getRetValTLS)() = nullptr);
266 bool doInitialization(Module &M) override;
267 bool runOnModule(Module &M) override;
270 struct DFSanFunction {
271 DataFlowSanitizer &DFS;
274 DataFlowSanitizer::InstrumentedABI IA;
278 AllocaInst *LabelReturnAlloca;
279 DenseMap<Value *, Value *> ValShadowMap;
280 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
281 std::vector<std::pair<PHINode *, PHINode *> > PHIFixups;
282 DenseSet<Instruction *> SkipInsts;
283 DenseSet<Value *> NonZeroChecks;
286 struct CachedCombinedShadow {
290 DenseMap<std::pair<Value *, Value *>, CachedCombinedShadow>
291 CachedCombinedShadows;
292 DenseMap<Value *, std::set<Value *>> ShadowElements;
294 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI)
295 : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()),
296 IsNativeABI(IsNativeABI), ArgTLSPtr(nullptr), RetvalTLSPtr(nullptr),
297 LabelReturnAlloca(nullptr) {
299 // FIXME: Need to track down the register allocator issue which causes poor
300 // performance in pathological cases with large numbers of basic blocks.
301 AvoidNewBlocks = F->size() > 1000;
303 Value *getArgTLSPtr();
304 Value *getArgTLS(unsigned Index, Instruction *Pos);
305 Value *getRetvalTLS();
306 Value *getShadow(Value *V);
307 void setShadow(Instruction *I, Value *Shadow);
308 Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
309 Value *combineOperandShadows(Instruction *Inst);
310 Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align,
312 void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow,
316 class DFSanVisitor : public InstVisitor<DFSanVisitor> {
319 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
321 void visitOperandShadowInst(Instruction &I);
323 void visitBinaryOperator(BinaryOperator &BO);
324 void visitCastInst(CastInst &CI);
325 void visitCmpInst(CmpInst &CI);
326 void visitGetElementPtrInst(GetElementPtrInst &GEPI);
327 void visitLoadInst(LoadInst &LI);
328 void visitStoreInst(StoreInst &SI);
329 void visitReturnInst(ReturnInst &RI);
330 void visitCallSite(CallSite CS);
331 void visitPHINode(PHINode &PN);
332 void visitExtractElementInst(ExtractElementInst &I);
333 void visitInsertElementInst(InsertElementInst &I);
334 void visitShuffleVectorInst(ShuffleVectorInst &I);
335 void visitExtractValueInst(ExtractValueInst &I);
336 void visitInsertValueInst(InsertValueInst &I);
337 void visitAllocaInst(AllocaInst &I);
338 void visitSelectInst(SelectInst &I);
339 void visitMemSetInst(MemSetInst &I);
340 void visitMemTransferInst(MemTransferInst &I);
345 char DataFlowSanitizer::ID;
346 INITIALIZE_PASS(DataFlowSanitizer, "dfsan",
347 "DataFlowSanitizer: dynamic data flow analysis.", false, false)
349 ModulePass *llvm::createDataFlowSanitizerPass(StringRef ABIListFile,
350 void *(*getArgTLS)(),
351 void *(*getRetValTLS)()) {
352 return new DataFlowSanitizer(ABIListFile, getArgTLS, getRetValTLS);
355 DataFlowSanitizer::DataFlowSanitizer(StringRef ABIListFile,
356 void *(*getArgTLS)(),
357 void *(*getRetValTLS)())
358 : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS),
359 ABIList(SpecialCaseList::createOrDie(ABIListFile.empty() ? ClABIListFile
363 FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) {
364 llvm::SmallVector<Type *, 4> ArgTypes;
365 std::copy(T->param_begin(), T->param_end(), std::back_inserter(ArgTypes));
366 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
367 ArgTypes.push_back(ShadowTy);
369 ArgTypes.push_back(ShadowPtrTy);
370 Type *RetType = T->getReturnType();
371 if (!RetType->isVoidTy())
372 RetType = StructType::get(RetType, ShadowTy, (Type *)nullptr);
373 return FunctionType::get(RetType, ArgTypes, T->isVarArg());
376 FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) {
377 assert(!T->isVarArg());
378 llvm::SmallVector<Type *, 4> ArgTypes;
379 ArgTypes.push_back(T->getPointerTo());
380 std::copy(T->param_begin(), T->param_end(), std::back_inserter(ArgTypes));
381 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
382 ArgTypes.push_back(ShadowTy);
383 Type *RetType = T->getReturnType();
384 if (!RetType->isVoidTy())
385 ArgTypes.push_back(ShadowPtrTy);
386 return FunctionType::get(T->getReturnType(), ArgTypes, false);
389 FunctionType *DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
390 assert(!T->isVarArg());
391 llvm::SmallVector<Type *, 4> ArgTypes;
392 for (FunctionType::param_iterator i = T->param_begin(), e = T->param_end();
395 if (isa<PointerType>(*i) && (FT = dyn_cast<FunctionType>(cast<PointerType>(
396 *i)->getElementType()))) {
397 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo());
398 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx));
400 ArgTypes.push_back(*i);
403 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
404 ArgTypes.push_back(ShadowTy);
405 Type *RetType = T->getReturnType();
406 if (!RetType->isVoidTy())
407 ArgTypes.push_back(ShadowPtrTy);
408 return FunctionType::get(T->getReturnType(), ArgTypes, false);
411 bool DataFlowSanitizer::doInitialization(Module &M) {
412 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
414 report_fatal_error("data layout missing");
415 DL = &DLP->getDataLayout();
418 Ctx = &M.getContext();
419 ShadowTy = IntegerType::get(*Ctx, ShadowWidth);
420 ShadowPtrTy = PointerType::getUnqual(ShadowTy);
421 IntptrTy = DL->getIntPtrType(*Ctx);
422 ZeroShadow = ConstantInt::getSigned(ShadowTy, 0);
423 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL);
424 ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8);
426 Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy };
428 FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false);
429 Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy };
431 FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false);
432 DFSanUnimplementedFnTy = FunctionType::get(
433 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
434 Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy };
435 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx),
436 DFSanSetLabelArgs, /*isVarArg=*/false);
437 DFSanNonzeroLabelFnTy = FunctionType::get(
438 Type::getVoidTy(*Ctx), ArrayRef<Type *>(), /*isVarArg=*/false);
441 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
443 GetArgTLS = ConstantExpr::getIntToPtr(
444 ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)),
445 PointerType::getUnqual(
446 FunctionType::get(PointerType::getUnqual(ArgTLSTy),
449 if (GetRetvalTLSPtr) {
451 GetRetvalTLS = ConstantExpr::getIntToPtr(
452 ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)),
453 PointerType::getUnqual(
454 FunctionType::get(PointerType::getUnqual(ShadowTy),
458 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
462 bool DataFlowSanitizer::isInstrumented(const Function *F) {
463 return !ABIList.isIn(*F, "uninstrumented");
466 bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
467 return !ABIList.isIn(*GA, "uninstrumented");
470 DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() {
471 return ClArgsABI ? IA_Args : IA_TLS;
474 DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
475 if (ABIList.isIn(*F, "functional"))
476 return WK_Functional;
477 if (ABIList.isIn(*F, "discard"))
479 if (ABIList.isIn(*F, "custom") && !F->isVarArg())
485 void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) {
486 std::string GVName = GV->getName(), Prefix = "dfs$";
487 GV->setName(Prefix + GVName);
489 // Try to change the name of the function in module inline asm. We only do
490 // this for specific asm directives, currently only ".symver", to try to avoid
491 // corrupting asm which happens to contain the symbol name as a substring.
492 // Note that the substitution for .symver assumes that the versioned symbol
493 // also has an instrumented name.
494 std::string Asm = GV->getParent()->getModuleInlineAsm();
495 std::string SearchStr = ".symver " + GVName + ",";
496 size_t Pos = Asm.find(SearchStr);
497 if (Pos != std::string::npos) {
498 Asm.replace(Pos, SearchStr.size(),
499 ".symver " + Prefix + GVName + "," + Prefix);
500 GV->getParent()->setModuleInlineAsm(Asm);
505 DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
506 GlobalValue::LinkageTypes NewFLink,
507 FunctionType *NewFT) {
508 FunctionType *FT = F->getFunctionType();
509 Function *NewF = Function::Create(NewFT, NewFLink, NewFName,
511 NewF->copyAttributesFrom(F);
512 NewF->removeAttributes(
513 AttributeSet::ReturnIndex,
514 AttributeFuncs::typeIncompatible(NewFT->getReturnType(),
515 AttributeSet::ReturnIndex));
517 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF);
518 std::vector<Value *> Args;
519 unsigned n = FT->getNumParams();
520 for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n)
521 Args.push_back(&*ai);
522 CallInst *CI = CallInst::Create(F, Args, "", BB);
523 if (FT->getReturnType()->isVoidTy())
524 ReturnInst::Create(*Ctx, BB);
526 ReturnInst::Create(*Ctx, CI, BB);
531 Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT,
533 FunctionType *FTT = getTrampolineFunctionType(FT);
534 Constant *C = Mod->getOrInsertFunction(FName, FTT);
535 Function *F = dyn_cast<Function>(C);
536 if (F && F->isDeclaration()) {
537 F->setLinkage(GlobalValue::LinkOnceODRLinkage);
538 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
539 std::vector<Value *> Args;
540 Function::arg_iterator AI = F->arg_begin(); ++AI;
541 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N)
542 Args.push_back(&*AI);
544 CallInst::Create(&F->getArgumentList().front(), Args, "", BB);
546 if (FT->getReturnType()->isVoidTy())
547 RI = ReturnInst::Create(*Ctx, BB);
549 RI = ReturnInst::Create(*Ctx, CI, BB);
551 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true);
552 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI;
553 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N)
554 DFSF.ValShadowMap[ValAI] = ShadowAI;
555 DFSanVisitor(DFSF).visitCallInst(*CI);
556 if (!FT->getReturnType()->isVoidTy())
557 new StoreInst(DFSF.getShadow(RI->getReturnValue()),
558 &F->getArgumentList().back(), RI);
564 bool DataFlowSanitizer::runOnModule(Module &M) {
568 if (ABIList.isIn(M, "skip"))
572 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
573 ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy);
574 if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS))
575 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
577 if (!GetRetvalTLSPtr) {
578 RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy);
579 if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS))
580 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
583 DFSanUnionFn = Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy);
584 if (Function *F = dyn_cast<Function>(DFSanUnionFn)) {
585 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
586 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone);
587 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
588 F->addAttribute(1, Attribute::ZExt);
589 F->addAttribute(2, Attribute::ZExt);
591 DFSanCheckedUnionFn = Mod->getOrInsertFunction("dfsan_union", DFSanUnionFnTy);
592 if (Function *F = dyn_cast<Function>(DFSanCheckedUnionFn)) {
593 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
594 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone);
595 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
596 F->addAttribute(1, Attribute::ZExt);
597 F->addAttribute(2, Attribute::ZExt);
600 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy);
601 if (Function *F = dyn_cast<Function>(DFSanUnionLoadFn)) {
602 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
603 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadOnly);
604 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
606 DFSanUnimplementedFn =
607 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy);
609 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy);
610 if (Function *F = dyn_cast<Function>(DFSanSetLabelFn)) {
611 F->addAttribute(1, Attribute::ZExt);
613 DFSanNonzeroLabelFn =
614 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy);
616 std::vector<Function *> FnsToInstrument;
617 llvm::SmallPtrSet<Function *, 2> FnsWithNativeABI;
618 for (Module::iterator i = M.begin(), e = M.end(); i != e; ++i) {
619 if (!i->isIntrinsic() &&
621 i != DFSanCheckedUnionFn &&
622 i != DFSanUnionLoadFn &&
623 i != DFSanUnimplementedFn &&
624 i != DFSanSetLabelFn &&
625 i != DFSanNonzeroLabelFn)
626 FnsToInstrument.push_back(&*i);
629 // Give function aliases prefixes when necessary, and build wrappers where the
630 // instrumentedness is inconsistent.
631 for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) {
632 GlobalAlias *GA = &*i;
634 // Don't stop on weak. We assume people aren't playing games with the
635 // instrumentedness of overridden weak aliases.
636 if (auto F = dyn_cast<Function>(GA->getBaseObject())) {
637 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F);
638 if (GAInst && FInst) {
639 addGlobalNamePrefix(GA);
640 } else if (GAInst != FInst) {
641 // Non-instrumented alias of an instrumented function, or vice versa.
642 // Replace the alias with a native-ABI wrapper of the aliasee. The pass
643 // below will take care of instrumenting it.
645 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType());
646 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType()));
648 GA->eraseFromParent();
649 FnsToInstrument.push_back(NewF);
655 B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
656 ReadOnlyNoneAttrs = AttributeSet::get(*Ctx, AttributeSet::FunctionIndex, B);
658 // First, change the ABI of every function in the module. ABI-listed
659 // functions keep their original ABI and get a wrapper function.
660 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
661 e = FnsToInstrument.end();
664 FunctionType *FT = F.getFunctionType();
666 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
667 FT->getReturnType()->isVoidTy());
669 if (isInstrumented(&F)) {
670 // Instrumented functions get a 'dfs$' prefix. This allows us to more
671 // easily identify cases of mismatching ABIs.
672 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) {
673 FunctionType *NewFT = getArgsFunctionType(FT);
674 Function *NewF = Function::Create(NewFT, F.getLinkage(), "", &M);
675 NewF->copyAttributesFrom(&F);
676 NewF->removeAttributes(
677 AttributeSet::ReturnIndex,
678 AttributeFuncs::typeIncompatible(NewFT->getReturnType(),
679 AttributeSet::ReturnIndex));
680 for (Function::arg_iterator FArg = F.arg_begin(),
681 NewFArg = NewF->arg_begin(),
682 FArgEnd = F.arg_end();
683 FArg != FArgEnd; ++FArg, ++NewFArg) {
684 FArg->replaceAllUsesWith(NewFArg);
686 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList());
688 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end();
690 BlockAddress *BA = dyn_cast<BlockAddress>(*UI);
693 BA->replaceAllUsesWith(
694 BlockAddress::get(NewF, BA->getBasicBlock()));
698 F.replaceAllUsesWith(
699 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)));
703 addGlobalNamePrefix(NewF);
705 addGlobalNamePrefix(&F);
707 // Hopefully, nobody will try to indirectly call a vararg
709 } else if (FT->isVarArg()) {
710 UnwrappedFnMap[&F] = &F;
712 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) {
713 // Build a wrapper function for F. The wrapper simply calls F, and is
714 // added to FnsToInstrument so that any instrumentation according to its
715 // WrapperKind is done in the second pass below.
716 FunctionType *NewFT = getInstrumentedABI() == IA_Args
717 ? getArgsFunctionType(FT)
719 Function *NewF = buildWrapperFunction(
720 &F, std::string("dfsw$") + std::string(F.getName()),
721 GlobalValue::LinkOnceODRLinkage, NewFT);
722 if (getInstrumentedABI() == IA_TLS)
723 NewF->removeAttributes(AttributeSet::FunctionIndex, ReadOnlyNoneAttrs);
725 Value *WrappedFnCst =
726 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT));
727 F.replaceAllUsesWith(WrappedFnCst);
728 UnwrappedFnMap[WrappedFnCst] = &F;
731 if (!F.isDeclaration()) {
732 // This function is probably defining an interposition of an
733 // uninstrumented function and hence needs to keep the original ABI.
734 // But any functions it may call need to use the instrumented ABI, so
735 // we instrument it in a mode which preserves the original ABI.
736 FnsWithNativeABI.insert(&F);
738 // This code needs to rebuild the iterators, as they may be invalidated
739 // by the push_back, taking care that the new range does not include
740 // any functions added by this code.
741 size_t N = i - FnsToInstrument.begin(),
742 Count = e - FnsToInstrument.begin();
743 FnsToInstrument.push_back(&F);
744 i = FnsToInstrument.begin() + N;
745 e = FnsToInstrument.begin() + Count;
750 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
751 e = FnsToInstrument.end();
753 if (!*i || (*i)->isDeclaration())
756 removeUnreachableBlocks(**i);
758 DFSanFunction DFSF(*this, *i, FnsWithNativeABI.count(*i));
760 // DFSanVisitor may create new basic blocks, which confuses df_iterator.
761 // Build a copy of the list before iterating over it.
762 llvm::SmallVector<BasicBlock *, 4> BBList(
763 depth_first(&(*i)->getEntryBlock()));
765 for (llvm::SmallVector<BasicBlock *, 4>::iterator i = BBList.begin(),
768 Instruction *Inst = &(*i)->front();
770 // DFSanVisitor may split the current basic block, changing the current
771 // instruction's next pointer and moving the next instruction to the
772 // tail block from which we should continue.
773 Instruction *Next = Inst->getNextNode();
774 // DFSanVisitor may delete Inst, so keep track of whether it was a
776 bool IsTerminator = isa<TerminatorInst>(Inst);
777 if (!DFSF.SkipInsts.count(Inst))
778 DFSanVisitor(DFSF).visit(Inst);
785 // We will not necessarily be able to compute the shadow for every phi node
786 // until we have visited every block. Therefore, the code that handles phi
787 // nodes adds them to the PHIFixups list so that they can be properly
789 for (std::vector<std::pair<PHINode *, PHINode *> >::iterator
790 i = DFSF.PHIFixups.begin(),
791 e = DFSF.PHIFixups.end();
793 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n;
795 i->second->setIncomingValue(
796 val, DFSF.getShadow(i->first->getIncomingValue(val)));
800 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
801 // places (i.e. instructions in basic blocks we haven't even begun visiting
802 // yet). To make our life easier, do this work in a pass after the main
804 if (ClDebugNonzeroLabels) {
805 for (DenseSet<Value *>::iterator i = DFSF.NonZeroChecks.begin(),
806 e = DFSF.NonZeroChecks.end();
809 if (Instruction *I = dyn_cast<Instruction>(*i))
810 Pos = I->getNextNode();
812 Pos = DFSF.F->getEntryBlock().begin();
813 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos))
814 Pos = Pos->getNextNode();
815 IRBuilder<> IRB(Pos);
816 Value *Ne = IRB.CreateICmpNE(*i, DFSF.DFS.ZeroShadow);
817 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
818 Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
819 IRBuilder<> ThenIRB(BI);
820 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn);
828 Value *DFSanFunction::getArgTLSPtr() {
832 return ArgTLSPtr = DFS.ArgTLS;
834 IRBuilder<> IRB(F->getEntryBlock().begin());
835 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS);
838 Value *DFSanFunction::getRetvalTLS() {
842 return RetvalTLSPtr = DFS.RetvalTLS;
844 IRBuilder<> IRB(F->getEntryBlock().begin());
845 return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS);
848 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) {
849 IRBuilder<> IRB(Pos);
850 return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx);
853 Value *DFSanFunction::getShadow(Value *V) {
854 if (!isa<Argument>(V) && !isa<Instruction>(V))
855 return DFS.ZeroShadow;
856 Value *&Shadow = ValShadowMap[V];
858 if (Argument *A = dyn_cast<Argument>(V)) {
860 return DFS.ZeroShadow;
862 case DataFlowSanitizer::IA_TLS: {
863 Value *ArgTLSPtr = getArgTLSPtr();
864 Instruction *ArgTLSPos =
865 DFS.ArgTLS ? &*F->getEntryBlock().begin()
866 : cast<Instruction>(ArgTLSPtr)->getNextNode();
867 IRBuilder<> IRB(ArgTLSPos);
868 Shadow = IRB.CreateLoad(getArgTLS(A->getArgNo(), ArgTLSPos));
871 case DataFlowSanitizer::IA_Args: {
872 unsigned ArgIdx = A->getArgNo() + F->getArgumentList().size() / 2;
873 Function::arg_iterator i = F->arg_begin();
877 assert(Shadow->getType() == DFS.ShadowTy);
881 NonZeroChecks.insert(Shadow);
883 Shadow = DFS.ZeroShadow;
889 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
890 assert(!ValShadowMap.count(I));
891 assert(Shadow->getType() == DFS.ShadowTy);
892 ValShadowMap[I] = Shadow;
895 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
896 assert(Addr != RetvalTLS && "Reinstrumenting?");
897 IRBuilder<> IRB(Pos);
898 return IRB.CreateIntToPtr(
900 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy), ShadowPtrMask),
905 // Generates IR to compute the union of the two given shadows, inserting it
906 // before Pos. Returns the computed union Value.
907 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) {
908 if (V1 == DFS.ZeroShadow)
910 if (V2 == DFS.ZeroShadow)
915 auto V1Elems = ShadowElements.find(V1);
916 auto V2Elems = ShadowElements.find(V2);
917 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) {
918 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(),
919 V2Elems->second.begin(), V2Elems->second.end())) {
921 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(),
922 V1Elems->second.begin(), V1Elems->second.end())) {
925 } else if (V1Elems != ShadowElements.end()) {
926 if (V1Elems->second.count(V2))
928 } else if (V2Elems != ShadowElements.end()) {
929 if (V2Elems->second.count(V1))
933 auto Key = std::make_pair(V1, V2);
935 std::swap(Key.first, Key.second);
936 CachedCombinedShadow &CCS = CachedCombinedShadows[Key];
937 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent()))
940 IRBuilder<> IRB(Pos);
941 if (AvoidNewBlocks) {
942 CallInst *Call = IRB.CreateCall2(DFS.DFSanCheckedUnionFn, V1, V2);
943 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
944 Call->addAttribute(1, Attribute::ZExt);
945 Call->addAttribute(2, Attribute::ZExt);
947 CCS.Block = Pos->getParent();
950 BasicBlock *Head = Pos->getParent();
951 Value *Ne = IRB.CreateICmpNE(V1, V2);
952 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
953 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT));
954 IRBuilder<> ThenIRB(BI);
955 CallInst *Call = ThenIRB.CreateCall2(DFS.DFSanUnionFn, V1, V2);
956 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
957 Call->addAttribute(1, Attribute::ZExt);
958 Call->addAttribute(2, Attribute::ZExt);
960 BasicBlock *Tail = BI->getSuccessor(0);
961 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", Tail->begin());
962 Phi->addIncoming(Call, Call->getParent());
963 Phi->addIncoming(V1, Head);
969 std::set<Value *> UnionElems;
970 if (V1Elems != ShadowElements.end()) {
971 UnionElems = V1Elems->second;
973 UnionElems.insert(V1);
975 if (V2Elems != ShadowElements.end()) {
976 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end());
978 UnionElems.insert(V2);
980 ShadowElements[CCS.Shadow] = std::move(UnionElems);
985 // A convenience function which folds the shadows of each of the operands
986 // of the provided instruction Inst, inserting the IR before Inst. Returns
987 // the computed union Value.
988 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
989 if (Inst->getNumOperands() == 0)
990 return DFS.ZeroShadow;
992 Value *Shadow = getShadow(Inst->getOperand(0));
993 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) {
994 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst);
999 void DFSanVisitor::visitOperandShadowInst(Instruction &I) {
1000 Value *CombinedShadow = DFSF.combineOperandShadows(&I);
1001 DFSF.setShadow(&I, CombinedShadow);
1004 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where
1005 // Addr has alignment Align, and take the union of each of those shadows.
1006 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align,
1008 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1009 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1010 AllocaShadowMap.find(AI);
1011 if (i != AllocaShadowMap.end()) {
1012 IRBuilder<> IRB(Pos);
1013 return IRB.CreateLoad(i->second);
1017 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1018 SmallVector<Value *, 2> Objs;
1019 GetUnderlyingObjects(Addr, Objs, DFS.DL);
1020 bool AllConstants = true;
1021 for (SmallVector<Value *, 2>::iterator i = Objs.begin(), e = Objs.end();
1023 if (isa<Function>(*i) || isa<BlockAddress>(*i))
1025 if (isa<GlobalVariable>(*i) && cast<GlobalVariable>(*i)->isConstant())
1028 AllConstants = false;
1032 return DFS.ZeroShadow;
1034 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1037 return DFS.ZeroShadow;
1039 LoadInst *LI = new LoadInst(ShadowAddr, "", Pos);
1040 LI->setAlignment(ShadowAlign);
1044 IRBuilder<> IRB(Pos);
1045 Value *ShadowAddr1 =
1046 IRB.CreateGEP(ShadowAddr, ConstantInt::get(DFS.IntptrTy, 1));
1047 return combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign),
1048 IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign), Pos);
1051 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) {
1052 // Fast path for the common case where each byte has identical shadow: load
1053 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any
1054 // shadow is non-equal.
1055 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F);
1056 IRBuilder<> FallbackIRB(FallbackBB);
1057 CallInst *FallbackCall = FallbackIRB.CreateCall2(
1058 DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size));
1059 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
1061 // Compare each of the shadows stored in the loaded 64 bits to each other,
1062 // by computing (WideShadow rotl ShadowWidth) == WideShadow.
1063 IRBuilder<> IRB(Pos);
1065 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx));
1066 Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1067 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy);
1068 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth);
1069 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth);
1070 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow);
1071 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow);
1073 BasicBlock *Head = Pos->getParent();
1074 BasicBlock *Tail = Head->splitBasicBlock(Pos);
1076 if (DomTreeNode *OldNode = DT.getNode(Head)) {
1077 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end());
1079 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head);
1080 for (auto Child : Children)
1081 DT.changeImmediateDominator(Child, NewNode);
1084 // In the following code LastBr will refer to the previous basic block's
1085 // conditional branch instruction, whose true successor is fixed up to point
1086 // to the next block during the loop below or to the tail after the final
1088 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq);
1089 ReplaceInstWithInst(Head->getTerminator(), LastBr);
1090 DT.addNewBlock(FallbackBB, Head);
1092 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size;
1093 Ofs += 64 / DFS.ShadowWidth) {
1094 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F);
1095 DT.addNewBlock(NextBB, LastBr->getParent());
1096 IRBuilder<> NextIRB(NextBB);
1097 WideAddr = NextIRB.CreateGEP(WideAddr, ConstantInt::get(DFS.IntptrTy, 1));
1098 Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1099 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow);
1100 LastBr->setSuccessor(0, NextBB);
1101 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB);
1104 LastBr->setSuccessor(0, Tail);
1105 FallbackIRB.CreateBr(Tail);
1106 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1107 Shadow->addIncoming(FallbackCall, FallbackBB);
1108 Shadow->addIncoming(TruncShadow, LastBr->getParent());
1112 IRBuilder<> IRB(Pos);
1113 CallInst *FallbackCall = IRB.CreateCall2(
1114 DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size));
1115 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt);
1116 return FallbackCall;
1119 void DFSanVisitor::visitLoadInst(LoadInst &LI) {
1120 uint64_t Size = DFSF.DFS.DL->getTypeStoreSize(LI.getType());
1122 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow);
1127 if (ClPreserveAlignment) {
1128 Align = LI.getAlignment();
1130 Align = DFSF.DFS.DL->getABITypeAlignment(LI.getType());
1134 IRBuilder<> IRB(&LI);
1135 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI);
1136 if (ClCombinePointerLabelsOnLoad) {
1137 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
1138 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI);
1140 if (Shadow != DFSF.DFS.ZeroShadow)
1141 DFSF.NonZeroChecks.insert(Shadow);
1143 DFSF.setShadow(&LI, Shadow);
1146 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align,
1147 Value *Shadow, Instruction *Pos) {
1148 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1149 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1150 AllocaShadowMap.find(AI);
1151 if (i != AllocaShadowMap.end()) {
1152 IRBuilder<> IRB(Pos);
1153 IRB.CreateStore(Shadow, i->second);
1158 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1159 IRBuilder<> IRB(Pos);
1160 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1161 if (Shadow == DFS.ZeroShadow) {
1162 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth);
1163 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0);
1164 Value *ExtShadowAddr =
1165 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy));
1166 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign);
1170 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth;
1171 uint64_t Offset = 0;
1172 if (Size >= ShadowVecSize) {
1173 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize);
1174 Value *ShadowVec = UndefValue::get(ShadowVecTy);
1175 for (unsigned i = 0; i != ShadowVecSize; ++i) {
1176 ShadowVec = IRB.CreateInsertElement(
1177 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i));
1179 Value *ShadowVecAddr =
1180 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy));
1182 Value *CurShadowVecAddr = IRB.CreateConstGEP1_32(ShadowVecAddr, Offset);
1183 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign);
1184 Size -= ShadowVecSize;
1186 } while (Size >= ShadowVecSize);
1187 Offset *= ShadowVecSize;
1190 Value *CurShadowAddr = IRB.CreateConstGEP1_32(ShadowAddr, Offset);
1191 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign);
1197 void DFSanVisitor::visitStoreInst(StoreInst &SI) {
1199 DFSF.DFS.DL->getTypeStoreSize(SI.getValueOperand()->getType());
1204 if (ClPreserveAlignment) {
1205 Align = SI.getAlignment();
1207 Align = DFSF.DFS.DL->getABITypeAlignment(SI.getValueOperand()->getType());
1212 Value* Shadow = DFSF.getShadow(SI.getValueOperand());
1213 if (ClCombinePointerLabelsOnStore) {
1214 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand());
1215 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI);
1217 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI);
1220 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
1221 visitOperandShadowInst(BO);
1224 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); }
1226 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); }
1228 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
1229 visitOperandShadowInst(GEPI);
1232 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
1233 visitOperandShadowInst(I);
1236 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
1237 visitOperandShadowInst(I);
1240 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
1241 visitOperandShadowInst(I);
1244 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
1245 visitOperandShadowInst(I);
1248 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
1249 visitOperandShadowInst(I);
1252 void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
1253 bool AllLoadsStores = true;
1254 for (User *U : I.users()) {
1255 if (isa<LoadInst>(U))
1258 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1259 if (SI->getPointerOperand() == &I)
1263 AllLoadsStores = false;
1266 if (AllLoadsStores) {
1267 IRBuilder<> IRB(&I);
1268 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy);
1270 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow);
1273 void DFSanVisitor::visitSelectInst(SelectInst &I) {
1274 Value *CondShadow = DFSF.getShadow(I.getCondition());
1275 Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
1276 Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
1278 if (isa<VectorType>(I.getCondition()->getType())) {
1281 DFSF.combineShadows(
1282 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I));
1285 if (TrueShadow == FalseShadow) {
1286 ShadowSel = TrueShadow;
1289 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I);
1291 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I));
1295 void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
1296 IRBuilder<> IRB(&I);
1297 Value *ValShadow = DFSF.getShadow(I.getValue());
1299 DFSF.DFS.DFSanSetLabelFn, ValShadow,
1300 IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(*DFSF.DFS.Ctx)),
1301 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy));
1304 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
1305 IRBuilder<> IRB(&I);
1306 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I);
1307 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I);
1308 Value *LenShadow = IRB.CreateMul(
1310 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8));
1312 if (ClPreserveAlignment) {
1313 AlignShadow = IRB.CreateMul(I.getAlignmentCst(),
1314 ConstantInt::get(I.getAlignmentCst()->getType(),
1315 DFSF.DFS.ShadowWidth / 8));
1317 AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(),
1318 DFSF.DFS.ShadowWidth / 8);
1320 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx);
1321 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr);
1322 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr);
1323 IRB.CreateCall5(I.getCalledValue(), DestShadow, SrcShadow, LenShadow,
1324 AlignShadow, I.getVolatileCst());
1327 void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
1328 if (!DFSF.IsNativeABI && RI.getReturnValue()) {
1330 case DataFlowSanitizer::IA_TLS: {
1331 Value *S = DFSF.getShadow(RI.getReturnValue());
1332 IRBuilder<> IRB(&RI);
1333 IRB.CreateStore(S, DFSF.getRetvalTLS());
1336 case DataFlowSanitizer::IA_Args: {
1337 IRBuilder<> IRB(&RI);
1338 Type *RT = DFSF.F->getFunctionType()->getReturnType();
1340 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0);
1342 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1);
1343 RI.setOperand(0, InsShadow);
1350 void DFSanVisitor::visitCallSite(CallSite CS) {
1351 Function *F = CS.getCalledFunction();
1352 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) {
1353 visitOperandShadowInst(*CS.getInstruction());
1357 IRBuilder<> IRB(CS.getInstruction());
1359 DenseMap<Value *, Function *>::iterator i =
1360 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue());
1361 if (i != DFSF.DFS.UnwrappedFnMap.end()) {
1362 Function *F = i->second;
1363 switch (DFSF.DFS.getWrapperKind(F)) {
1364 case DataFlowSanitizer::WK_Warning: {
1365 CS.setCalledFunction(F);
1366 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn,
1367 IRB.CreateGlobalStringPtr(F->getName()));
1368 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1371 case DataFlowSanitizer::WK_Discard: {
1372 CS.setCalledFunction(F);
1373 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1376 case DataFlowSanitizer::WK_Functional: {
1377 CS.setCalledFunction(F);
1378 visitOperandShadowInst(*CS.getInstruction());
1381 case DataFlowSanitizer::WK_Custom: {
1382 // Don't try to handle invokes of custom functions, it's too complicated.
1383 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
1385 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1386 FunctionType *FT = F->getFunctionType();
1387 FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT);
1388 std::string CustomFName = "__dfsw_";
1389 CustomFName += F->getName();
1391 DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT);
1392 if (Function *CustomFn = dyn_cast<Function>(CustomF)) {
1393 CustomFn->copyAttributesFrom(F);
1395 // Custom functions returning non-void will write to the return label.
1396 if (!FT->getReturnType()->isVoidTy()) {
1397 CustomFn->removeAttributes(AttributeSet::FunctionIndex,
1398 DFSF.DFS.ReadOnlyNoneAttrs);
1402 std::vector<Value *> Args;
1404 CallSite::arg_iterator i = CS.arg_begin();
1405 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) {
1406 Type *T = (*i)->getType();
1407 FunctionType *ParamFT;
1408 if (isa<PointerType>(T) &&
1409 (ParamFT = dyn_cast<FunctionType>(
1410 cast<PointerType>(T)->getElementType()))) {
1411 std::string TName = "dfst";
1412 TName += utostr(FT->getNumParams() - n);
1414 TName += F->getName();
1415 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName);
1418 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx)));
1425 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1426 Args.push_back(DFSF.getShadow(*i));
1428 if (!FT->getReturnType()->isVoidTy()) {
1429 if (!DFSF.LabelReturnAlloca) {
1430 DFSF.LabelReturnAlloca =
1431 new AllocaInst(DFSF.DFS.ShadowTy, "labelreturn",
1432 DFSF.F->getEntryBlock().begin());
1434 Args.push_back(DFSF.LabelReturnAlloca);
1437 CallInst *CustomCI = IRB.CreateCall(CustomF, Args);
1438 CustomCI->setCallingConv(CI->getCallingConv());
1439 CustomCI->setAttributes(CI->getAttributes());
1441 if (!FT->getReturnType()->isVoidTy()) {
1442 LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca);
1443 DFSF.setShadow(CustomCI, LabelLoad);
1446 CI->replaceAllUsesWith(CustomCI);
1447 CI->eraseFromParent();
1455 FunctionType *FT = cast<FunctionType>(
1456 CS.getCalledValue()->getType()->getPointerElementType());
1457 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1458 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) {
1459 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)),
1460 DFSF.getArgTLS(i, CS.getInstruction()));
1464 Instruction *Next = nullptr;
1465 if (!CS.getType()->isVoidTy()) {
1466 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1467 if (II->getNormalDest()->getSinglePredecessor()) {
1468 Next = II->getNormalDest()->begin();
1471 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DFS);
1472 Next = NewBB->begin();
1475 Next = CS->getNextNode();
1478 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1479 IRBuilder<> NextIRB(Next);
1480 LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS());
1481 DFSF.SkipInsts.insert(LI);
1482 DFSF.setShadow(CS.getInstruction(), LI);
1483 DFSF.NonZeroChecks.insert(LI);
1487 // Do all instrumentation for IA_Args down here to defer tampering with the
1488 // CFG in a way that SplitEdge may be able to detect.
1489 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) {
1490 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT);
1492 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT));
1493 std::vector<Value *> Args;
1495 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1496 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1500 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1501 Args.push_back(DFSF.getShadow(*i));
1503 if (FT->isVarArg()) {
1504 unsigned VarArgSize = CS.arg_size() - FT->getNumParams();
1505 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize);
1506 AllocaInst *VarArgShadow =
1507 new AllocaInst(VarArgArrayTy, "", DFSF.F->getEntryBlock().begin());
1508 Args.push_back(IRB.CreateConstGEP2_32(VarArgShadow, 0, 0));
1509 for (unsigned n = 0; i != e; ++i, ++n) {
1510 IRB.CreateStore(DFSF.getShadow(*i),
1511 IRB.CreateConstGEP2_32(VarArgShadow, 0, n));
1517 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1518 NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(),
1521 NewCS = IRB.CreateCall(Func, Args);
1523 NewCS.setCallingConv(CS.getCallingConv());
1524 NewCS.setAttributes(CS.getAttributes().removeAttributes(
1525 *DFSF.DFS.Ctx, AttributeSet::ReturnIndex,
1526 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType(),
1527 AttributeSet::ReturnIndex)));
1530 ExtractValueInst *ExVal =
1531 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next);
1532 DFSF.SkipInsts.insert(ExVal);
1533 ExtractValueInst *ExShadow =
1534 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next);
1535 DFSF.SkipInsts.insert(ExShadow);
1536 DFSF.setShadow(ExVal, ExShadow);
1537 DFSF.NonZeroChecks.insert(ExShadow);
1539 CS.getInstruction()->replaceAllUsesWith(ExVal);
1542 CS.getInstruction()->eraseFromParent();
1546 void DFSanVisitor::visitPHINode(PHINode &PN) {
1548 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN);
1550 // Give the shadow phi node valid predecessors to fool SplitEdge into working.
1551 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy);
1552 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e;
1554 ShadowPN->addIncoming(UndefShadow, *i);
1557 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN));
1558 DFSF.setShadow(&PN, ShadowPN);