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/ADT/DenseMap.h"
48 #include "llvm/ADT/DenseSet.h"
49 #include "llvm/ADT/DepthFirstIterator.h"
50 #include "llvm/ADT/StringExtras.h"
51 #include "llvm/ADT/Triple.h"
52 #include "llvm/Analysis/ValueTracking.h"
53 #include "llvm/IR/DebugInfo.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/InlineAsm.h"
57 #include "llvm/IR/InstVisitor.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/MDBuilder.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/Value.h"
62 #include "llvm/Pass.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/SpecialCaseList.h"
65 #include "llvm/Transforms/Instrumentation.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
75 // External symbol to be used when generating the shadow address for
76 // architectures with multiple VMAs. Instead of using a constant integer
77 // the runtime will set the external mask based on the VMA range.
78 static const char *const kDFSanExternShadowPtrMask = "__dfsan_shadow_ptr_mask";
80 // The -dfsan-preserve-alignment flag controls whether this pass assumes that
81 // alignment requirements provided by the input IR are correct. For example,
82 // if the input IR contains a load with alignment 8, this flag will cause
83 // the shadow load to have alignment 16. This flag is disabled by default as
84 // we have unfortunately encountered too much code (including Clang itself;
85 // see PR14291) which performs misaligned access.
86 static cl::opt<bool> ClPreserveAlignment(
87 "dfsan-preserve-alignment",
88 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
91 // The ABI list files control how shadow parameters are passed. The pass treats
92 // every function labelled "uninstrumented" in the ABI list file as conforming
93 // to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains
94 // additional annotations for those functions, a call to one of those functions
95 // will produce a warning message, as the labelling behaviour of the function is
96 // unknown. The other supported annotations are "functional" and "discard",
97 // which are described below under DataFlowSanitizer::WrapperKind.
98 static cl::list<std::string> ClABIListFiles(
100 cl::desc("File listing native ABI functions and how the pass treats them"),
103 // Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented
104 // functions (see DataFlowSanitizer::InstrumentedABI below).
105 static cl::opt<bool> ClArgsABI(
107 cl::desc("Use the argument ABI rather than the TLS ABI"),
110 // Controls whether the pass includes or ignores the labels of pointers in load
112 static cl::opt<bool> ClCombinePointerLabelsOnLoad(
113 "dfsan-combine-pointer-labels-on-load",
114 cl::desc("Combine the label of the pointer with the label of the data when "
115 "loading from memory."),
116 cl::Hidden, cl::init(true));
118 // Controls whether the pass includes or ignores the labels of pointers in
119 // stores instructions.
120 static cl::opt<bool> ClCombinePointerLabelsOnStore(
121 "dfsan-combine-pointer-labels-on-store",
122 cl::desc("Combine the label of the pointer with the label of the data when "
123 "storing in memory."),
124 cl::Hidden, cl::init(false));
126 static cl::opt<bool> ClDebugNonzeroLabels(
127 "dfsan-debug-nonzero-labels",
128 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
129 "load or return with a nonzero label"),
135 StringRef GetGlobalTypeString(const GlobalValue &G) {
136 // Types of GlobalVariables are always pointer types.
137 Type *GType = G.getValueType();
138 // For now we support blacklisting struct types only.
139 if (StructType *SGType = dyn_cast<StructType>(GType)) {
140 if (!SGType->isLiteral())
141 return SGType->getName();
143 return "<unknown type>";
147 std::unique_ptr<SpecialCaseList> SCL;
152 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); }
154 /// Returns whether either this function or its source file are listed in the
156 bool isIn(const Function &F, StringRef Category) const {
157 return isIn(*F.getParent(), Category) ||
158 SCL->inSection("fun", F.getName(), Category);
161 /// Returns whether this global alias is listed in the given category.
163 /// If GA aliases a function, the alias's name is matched as a function name
164 /// would be. Similarly, aliases of globals are matched like globals.
165 bool isIn(const GlobalAlias &GA, StringRef Category) const {
166 if (isIn(*GA.getParent(), Category))
169 if (isa<FunctionType>(GA.getValueType()))
170 return SCL->inSection("fun", GA.getName(), Category);
172 return SCL->inSection("global", GA.getName(), Category) ||
173 SCL->inSection("type", GetGlobalTypeString(GA), Category);
176 /// Returns whether this module is listed in the given category.
177 bool isIn(const Module &M, StringRef Category) const {
178 return SCL->inSection("src", M.getModuleIdentifier(), Category);
182 class DataFlowSanitizer : public ModulePass {
183 friend struct DFSanFunction;
184 friend class DFSanVisitor;
190 /// Which ABI should be used for instrumented functions?
191 enum InstrumentedABI {
192 /// Argument and return value labels are passed through additional
193 /// arguments and by modifying the return type.
196 /// Argument and return value labels are passed through TLS variables
197 /// __dfsan_arg_tls and __dfsan_retval_tls.
201 /// How should calls to uninstrumented functions be handled?
203 /// This function is present in an uninstrumented form but we don't know
204 /// how it should be handled. Print a warning and call the function anyway.
205 /// Don't label the return value.
208 /// This function does not write to (user-accessible) memory, and its return
209 /// value is unlabelled.
212 /// This function does not write to (user-accessible) memory, and the label
213 /// of its return value is the union of the label of its arguments.
216 /// Instead of calling the function, a custom wrapper __dfsw_F is called,
217 /// where F is the name of the function. This function may wrap the
218 /// original function or provide its own implementation. This is similar to
219 /// the IA_Args ABI, except that IA_Args uses a struct return type to
220 /// pass the return value shadow in a register, while WK_Custom uses an
221 /// extra pointer argument to return the shadow. This allows the wrapped
222 /// form of the function type to be expressed in C.
228 IntegerType *ShadowTy;
229 PointerType *ShadowPtrTy;
230 IntegerType *IntptrTy;
231 ConstantInt *ZeroShadow;
232 ConstantInt *ShadowPtrMask;
233 ConstantInt *ShadowPtrMul;
236 void *(*GetArgTLSPtr)();
237 void *(*GetRetvalTLSPtr)();
239 Constant *GetRetvalTLS;
240 Constant *ExternalShadowMask;
241 FunctionType *DFSanUnionFnTy;
242 FunctionType *DFSanUnionLoadFnTy;
243 FunctionType *DFSanUnimplementedFnTy;
244 FunctionType *DFSanSetLabelFnTy;
245 FunctionType *DFSanNonzeroLabelFnTy;
246 FunctionType *DFSanVarargWrapperFnTy;
247 Constant *DFSanUnionFn;
248 Constant *DFSanCheckedUnionFn;
249 Constant *DFSanUnionLoadFn;
250 Constant *DFSanUnimplementedFn;
251 Constant *DFSanSetLabelFn;
252 Constant *DFSanNonzeroLabelFn;
253 Constant *DFSanVarargWrapperFn;
254 MDNode *ColdCallWeights;
255 DFSanABIList ABIList;
256 DenseMap<Value *, Function *> UnwrappedFnMap;
257 AttrBuilder ReadOnlyNoneAttrs;
258 bool DFSanRuntimeShadowMask;
260 Value *getShadowAddress(Value *Addr, Instruction *Pos);
261 bool isInstrumented(const Function *F);
262 bool isInstrumented(const GlobalAlias *GA);
263 FunctionType *getArgsFunctionType(FunctionType *T);
264 FunctionType *getTrampolineFunctionType(FunctionType *T);
265 FunctionType *getCustomFunctionType(FunctionType *T);
266 InstrumentedABI getInstrumentedABI();
267 WrapperKind getWrapperKind(Function *F);
268 void addGlobalNamePrefix(GlobalValue *GV);
269 Function *buildWrapperFunction(Function *F, StringRef NewFName,
270 GlobalValue::LinkageTypes NewFLink,
271 FunctionType *NewFT);
272 Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName);
276 const std::vector<std::string> &ABIListFiles = std::vector<std::string>(),
277 void *(*getArgTLS)() = nullptr, void *(*getRetValTLS)() = nullptr);
279 bool doInitialization(Module &M) override;
280 bool runOnModule(Module &M) override;
283 struct DFSanFunction {
284 DataFlowSanitizer &DFS;
287 DataFlowSanitizer::InstrumentedABI IA;
291 AllocaInst *LabelReturnAlloca;
292 DenseMap<Value *, Value *> ValShadowMap;
293 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
294 std::vector<std::pair<PHINode *, PHINode *> > PHIFixups;
295 DenseSet<Instruction *> SkipInsts;
296 std::vector<Value *> NonZeroChecks;
299 struct CachedCombinedShadow {
303 DenseMap<std::pair<Value *, Value *>, CachedCombinedShadow>
304 CachedCombinedShadows;
305 DenseMap<Value *, std::set<Value *>> ShadowElements;
307 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI)
308 : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()),
309 IsNativeABI(IsNativeABI), ArgTLSPtr(nullptr), RetvalTLSPtr(nullptr),
310 LabelReturnAlloca(nullptr) {
312 // FIXME: Need to track down the register allocator issue which causes poor
313 // performance in pathological cases with large numbers of basic blocks.
314 AvoidNewBlocks = F->size() > 1000;
316 Value *getArgTLSPtr();
317 Value *getArgTLS(unsigned Index, Instruction *Pos);
318 Value *getRetvalTLS();
319 Value *getShadow(Value *V);
320 void setShadow(Instruction *I, Value *Shadow);
321 Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
322 Value *combineOperandShadows(Instruction *Inst);
323 Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align,
325 void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow,
329 class DFSanVisitor : public InstVisitor<DFSanVisitor> {
332 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
334 const DataLayout &getDataLayout() const {
335 return DFSF.F->getParent()->getDataLayout();
338 void visitOperandShadowInst(Instruction &I);
340 void visitBinaryOperator(BinaryOperator &BO);
341 void visitCastInst(CastInst &CI);
342 void visitCmpInst(CmpInst &CI);
343 void visitGetElementPtrInst(GetElementPtrInst &GEPI);
344 void visitLoadInst(LoadInst &LI);
345 void visitStoreInst(StoreInst &SI);
346 void visitReturnInst(ReturnInst &RI);
347 void visitCallSite(CallSite CS);
348 void visitPHINode(PHINode &PN);
349 void visitExtractElementInst(ExtractElementInst &I);
350 void visitInsertElementInst(InsertElementInst &I);
351 void visitShuffleVectorInst(ShuffleVectorInst &I);
352 void visitExtractValueInst(ExtractValueInst &I);
353 void visitInsertValueInst(InsertValueInst &I);
354 void visitAllocaInst(AllocaInst &I);
355 void visitSelectInst(SelectInst &I);
356 void visitMemSetInst(MemSetInst &I);
357 void visitMemTransferInst(MemTransferInst &I);
362 char DataFlowSanitizer::ID;
363 INITIALIZE_PASS(DataFlowSanitizer, "dfsan",
364 "DataFlowSanitizer: dynamic data flow analysis.", false, false)
367 llvm::createDataFlowSanitizerPass(const std::vector<std::string> &ABIListFiles,
368 void *(*getArgTLS)(),
369 void *(*getRetValTLS)()) {
370 return new DataFlowSanitizer(ABIListFiles, getArgTLS, getRetValTLS);
373 DataFlowSanitizer::DataFlowSanitizer(
374 const std::vector<std::string> &ABIListFiles, void *(*getArgTLS)(),
375 void *(*getRetValTLS)())
376 : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS),
377 DFSanRuntimeShadowMask(false) {
378 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles));
379 AllABIListFiles.insert(AllABIListFiles.end(), ClABIListFiles.begin(),
380 ClABIListFiles.end());
381 ABIList.set(SpecialCaseList::createOrDie(AllABIListFiles));
384 FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) {
385 llvm::SmallVector<Type *, 4> ArgTypes(T->param_begin(), T->param_end());
386 ArgTypes.append(T->getNumParams(), ShadowTy);
388 ArgTypes.push_back(ShadowPtrTy);
389 Type *RetType = T->getReturnType();
390 if (!RetType->isVoidTy())
391 RetType = StructType::get(RetType, ShadowTy);
392 return FunctionType::get(RetType, ArgTypes, T->isVarArg());
395 FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) {
396 assert(!T->isVarArg());
397 llvm::SmallVector<Type *, 4> ArgTypes;
398 ArgTypes.push_back(T->getPointerTo());
399 ArgTypes.append(T->param_begin(), T->param_end());
400 ArgTypes.append(T->getNumParams(), ShadowTy);
401 Type *RetType = T->getReturnType();
402 if (!RetType->isVoidTy())
403 ArgTypes.push_back(ShadowPtrTy);
404 return FunctionType::get(T->getReturnType(), ArgTypes, false);
407 FunctionType *DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
408 llvm::SmallVector<Type *, 4> ArgTypes;
409 for (FunctionType::param_iterator i = T->param_begin(), e = T->param_end();
412 if (isa<PointerType>(*i) && (FT = dyn_cast<FunctionType>(cast<PointerType>(
413 *i)->getElementType()))) {
414 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo());
415 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx));
417 ArgTypes.push_back(*i);
420 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
421 ArgTypes.push_back(ShadowTy);
423 ArgTypes.push_back(ShadowPtrTy);
424 Type *RetType = T->getReturnType();
425 if (!RetType->isVoidTy())
426 ArgTypes.push_back(ShadowPtrTy);
427 return FunctionType::get(T->getReturnType(), ArgTypes, T->isVarArg());
430 bool DataFlowSanitizer::doInitialization(Module &M) {
431 llvm::Triple TargetTriple(M.getTargetTriple());
432 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
433 bool IsMIPS64 = TargetTriple.getArch() == llvm::Triple::mips64 ||
434 TargetTriple.getArch() == llvm::Triple::mips64el;
435 bool IsAArch64 = TargetTriple.getArch() == llvm::Triple::aarch64 ||
436 TargetTriple.getArch() == llvm::Triple::aarch64_be;
438 const DataLayout &DL = M.getDataLayout();
441 Ctx = &M.getContext();
442 ShadowTy = IntegerType::get(*Ctx, ShadowWidth);
443 ShadowPtrTy = PointerType::getUnqual(ShadowTy);
444 IntptrTy = DL.getIntPtrType(*Ctx);
445 ZeroShadow = ConstantInt::getSigned(ShadowTy, 0);
446 ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8);
448 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL);
450 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0xF000000000LL);
451 // AArch64 supports multiple VMAs and the shadow mask is set at runtime.
453 DFSanRuntimeShadowMask = true;
455 report_fatal_error("unsupported triple");
457 Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy };
459 FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false);
460 Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy };
462 FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false);
463 DFSanUnimplementedFnTy = FunctionType::get(
464 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
465 Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy };
466 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx),
467 DFSanSetLabelArgs, /*isVarArg=*/false);
468 DFSanNonzeroLabelFnTy = FunctionType::get(
469 Type::getVoidTy(*Ctx), None, /*isVarArg=*/false);
470 DFSanVarargWrapperFnTy = FunctionType::get(
471 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
474 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
476 GetArgTLS = ConstantExpr::getIntToPtr(
477 ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)),
478 PointerType::getUnqual(
479 FunctionType::get(PointerType::getUnqual(ArgTLSTy), false)));
481 if (GetRetvalTLSPtr) {
483 GetRetvalTLS = ConstantExpr::getIntToPtr(
484 ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)),
485 PointerType::getUnqual(
486 FunctionType::get(PointerType::getUnqual(ShadowTy), false)));
489 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
493 bool DataFlowSanitizer::isInstrumented(const Function *F) {
494 return !ABIList.isIn(*F, "uninstrumented");
497 bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
498 return !ABIList.isIn(*GA, "uninstrumented");
501 DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() {
502 return ClArgsABI ? IA_Args : IA_TLS;
505 DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
506 if (ABIList.isIn(*F, "functional"))
507 return WK_Functional;
508 if (ABIList.isIn(*F, "discard"))
510 if (ABIList.isIn(*F, "custom"))
516 void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) {
517 std::string GVName = GV->getName(), Prefix = "dfs$";
518 GV->setName(Prefix + GVName);
520 // Try to change the name of the function in module inline asm. We only do
521 // this for specific asm directives, currently only ".symver", to try to avoid
522 // corrupting asm which happens to contain the symbol name as a substring.
523 // Note that the substitution for .symver assumes that the versioned symbol
524 // also has an instrumented name.
525 std::string Asm = GV->getParent()->getModuleInlineAsm();
526 std::string SearchStr = ".symver " + GVName + ",";
527 size_t Pos = Asm.find(SearchStr);
528 if (Pos != std::string::npos) {
529 Asm.replace(Pos, SearchStr.size(),
530 ".symver " + Prefix + GVName + "," + Prefix);
531 GV->getParent()->setModuleInlineAsm(Asm);
536 DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
537 GlobalValue::LinkageTypes NewFLink,
538 FunctionType *NewFT) {
539 FunctionType *FT = F->getFunctionType();
540 Function *NewF = Function::Create(NewFT, NewFLink, NewFName,
542 NewF->copyAttributesFrom(F);
543 NewF->removeAttributes(
544 AttributeList::ReturnIndex,
545 AttributeFuncs::typeIncompatible(NewFT->getReturnType()));
547 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF);
549 NewF->removeAttributes(AttributeList::FunctionIndex,
550 AttrBuilder().addAttribute("split-stack"));
551 CallInst::Create(DFSanVarargWrapperFn,
552 IRBuilder<>(BB).CreateGlobalStringPtr(F->getName()), "",
554 new UnreachableInst(*Ctx, BB);
556 std::vector<Value *> Args;
557 unsigned n = FT->getNumParams();
558 for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n)
559 Args.push_back(&*ai);
560 CallInst *CI = CallInst::Create(F, Args, "", BB);
561 if (FT->getReturnType()->isVoidTy())
562 ReturnInst::Create(*Ctx, BB);
564 ReturnInst::Create(*Ctx, CI, BB);
570 Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT,
572 FunctionType *FTT = getTrampolineFunctionType(FT);
573 Constant *C = Mod->getOrInsertFunction(FName, FTT);
574 Function *F = dyn_cast<Function>(C);
575 if (F && F->isDeclaration()) {
576 F->setLinkage(GlobalValue::LinkOnceODRLinkage);
577 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
578 std::vector<Value *> Args;
579 Function::arg_iterator AI = F->arg_begin(); ++AI;
580 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N)
581 Args.push_back(&*AI);
582 CallInst *CI = CallInst::Create(&*F->arg_begin(), Args, "", BB);
584 if (FT->getReturnType()->isVoidTy())
585 RI = ReturnInst::Create(*Ctx, BB);
587 RI = ReturnInst::Create(*Ctx, CI, BB);
589 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true);
590 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI;
591 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N)
592 DFSF.ValShadowMap[&*ValAI] = &*ShadowAI;
593 DFSanVisitor(DFSF).visitCallInst(*CI);
594 if (!FT->getReturnType()->isVoidTy())
595 new StoreInst(DFSF.getShadow(RI->getReturnValue()),
596 &*std::prev(F->arg_end()), RI);
602 bool DataFlowSanitizer::runOnModule(Module &M) {
603 if (ABIList.isIn(M, "skip"))
607 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
608 ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy);
609 if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS))
610 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
612 if (!GetRetvalTLSPtr) {
613 RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy);
614 if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS))
615 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
619 Mod->getOrInsertGlobal(kDFSanExternShadowPtrMask, IntptrTy);
621 DFSanUnionFn = Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy);
622 if (Function *F = dyn_cast<Function>(DFSanUnionFn)) {
623 F->addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind);
624 F->addAttribute(AttributeList::FunctionIndex, Attribute::ReadNone);
625 F->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
626 F->addParamAttr(0, Attribute::ZExt);
627 F->addParamAttr(1, Attribute::ZExt);
629 DFSanCheckedUnionFn = Mod->getOrInsertFunction("dfsan_union", DFSanUnionFnTy);
630 if (Function *F = dyn_cast<Function>(DFSanCheckedUnionFn)) {
631 F->addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind);
632 F->addAttribute(AttributeList::FunctionIndex, Attribute::ReadNone);
633 F->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
634 F->addParamAttr(0, Attribute::ZExt);
635 F->addParamAttr(1, Attribute::ZExt);
638 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy);
639 if (Function *F = dyn_cast<Function>(DFSanUnionLoadFn)) {
640 F->addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind);
641 F->addAttribute(AttributeList::FunctionIndex, Attribute::ReadOnly);
642 F->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
644 DFSanUnimplementedFn =
645 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy);
647 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy);
648 if (Function *F = dyn_cast<Function>(DFSanSetLabelFn)) {
649 F->addParamAttr(0, Attribute::ZExt);
651 DFSanNonzeroLabelFn =
652 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy);
653 DFSanVarargWrapperFn = Mod->getOrInsertFunction("__dfsan_vararg_wrapper",
654 DFSanVarargWrapperFnTy);
656 std::vector<Function *> FnsToInstrument;
657 llvm::SmallPtrSet<Function *, 2> FnsWithNativeABI;
658 for (Function &i : M) {
659 if (!i.isIntrinsic() &&
660 &i != DFSanUnionFn &&
661 &i != DFSanCheckedUnionFn &&
662 &i != DFSanUnionLoadFn &&
663 &i != DFSanUnimplementedFn &&
664 &i != DFSanSetLabelFn &&
665 &i != DFSanNonzeroLabelFn &&
666 &i != DFSanVarargWrapperFn)
667 FnsToInstrument.push_back(&i);
670 // Give function aliases prefixes when necessary, and build wrappers where the
671 // instrumentedness is inconsistent.
672 for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) {
673 GlobalAlias *GA = &*i;
675 // Don't stop on weak. We assume people aren't playing games with the
676 // instrumentedness of overridden weak aliases.
677 if (auto F = dyn_cast<Function>(GA->getBaseObject())) {
678 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F);
679 if (GAInst && FInst) {
680 addGlobalNamePrefix(GA);
681 } else if (GAInst != FInst) {
682 // Non-instrumented alias of an instrumented function, or vice versa.
683 // Replace the alias with a native-ABI wrapper of the aliasee. The pass
684 // below will take care of instrumenting it.
686 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType());
687 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType()));
689 GA->eraseFromParent();
690 FnsToInstrument.push_back(NewF);
695 ReadOnlyNoneAttrs.addAttribute(Attribute::ReadOnly)
696 .addAttribute(Attribute::ReadNone);
698 // First, change the ABI of every function in the module. ABI-listed
699 // functions keep their original ABI and get a wrapper function.
700 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
701 e = FnsToInstrument.end();
704 FunctionType *FT = F.getFunctionType();
706 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
707 FT->getReturnType()->isVoidTy());
709 if (isInstrumented(&F)) {
710 // Instrumented functions get a 'dfs$' prefix. This allows us to more
711 // easily identify cases of mismatching ABIs.
712 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) {
713 FunctionType *NewFT = getArgsFunctionType(FT);
714 Function *NewF = Function::Create(NewFT, F.getLinkage(), "", &M);
715 NewF->copyAttributesFrom(&F);
716 NewF->removeAttributes(
717 AttributeList::ReturnIndex,
718 AttributeFuncs::typeIncompatible(NewFT->getReturnType()));
719 for (Function::arg_iterator FArg = F.arg_begin(),
720 NewFArg = NewF->arg_begin(),
721 FArgEnd = F.arg_end();
722 FArg != FArgEnd; ++FArg, ++NewFArg) {
723 FArg->replaceAllUsesWith(&*NewFArg);
725 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList());
727 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end();
729 BlockAddress *BA = dyn_cast<BlockAddress>(*UI);
732 BA->replaceAllUsesWith(
733 BlockAddress::get(NewF, BA->getBasicBlock()));
737 F.replaceAllUsesWith(
738 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)));
742 addGlobalNamePrefix(NewF);
744 addGlobalNamePrefix(&F);
746 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) {
747 // Build a wrapper function for F. The wrapper simply calls F, and is
748 // added to FnsToInstrument so that any instrumentation according to its
749 // WrapperKind is done in the second pass below.
750 FunctionType *NewFT = getInstrumentedABI() == IA_Args
751 ? getArgsFunctionType(FT)
753 Function *NewF = buildWrapperFunction(
754 &F, std::string("dfsw$") + std::string(F.getName()),
755 GlobalValue::LinkOnceODRLinkage, NewFT);
756 if (getInstrumentedABI() == IA_TLS)
757 NewF->removeAttributes(AttributeList::FunctionIndex, ReadOnlyNoneAttrs);
759 Value *WrappedFnCst =
760 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT));
761 F.replaceAllUsesWith(WrappedFnCst);
763 UnwrappedFnMap[WrappedFnCst] = &F;
766 if (!F.isDeclaration()) {
767 // This function is probably defining an interposition of an
768 // uninstrumented function and hence needs to keep the original ABI.
769 // But any functions it may call need to use the instrumented ABI, so
770 // we instrument it in a mode which preserves the original ABI.
771 FnsWithNativeABI.insert(&F);
773 // This code needs to rebuild the iterators, as they may be invalidated
774 // by the push_back, taking care that the new range does not include
775 // any functions added by this code.
776 size_t N = i - FnsToInstrument.begin(),
777 Count = e - FnsToInstrument.begin();
778 FnsToInstrument.push_back(&F);
779 i = FnsToInstrument.begin() + N;
780 e = FnsToInstrument.begin() + Count;
782 // Hopefully, nobody will try to indirectly call a vararg
784 } else if (FT->isVarArg()) {
785 UnwrappedFnMap[&F] = &F;
790 for (Function *i : FnsToInstrument) {
791 if (!i || i->isDeclaration())
794 removeUnreachableBlocks(*i);
796 DFSanFunction DFSF(*this, i, FnsWithNativeABI.count(i));
798 // DFSanVisitor may create new basic blocks, which confuses df_iterator.
799 // Build a copy of the list before iterating over it.
800 llvm::SmallVector<BasicBlock *, 4> BBList(depth_first(&i->getEntryBlock()));
802 for (BasicBlock *i : BBList) {
803 Instruction *Inst = &i->front();
805 // DFSanVisitor may split the current basic block, changing the current
806 // instruction's next pointer and moving the next instruction to the
807 // tail block from which we should continue.
808 Instruction *Next = Inst->getNextNode();
809 // DFSanVisitor may delete Inst, so keep track of whether it was a
811 bool IsTerminator = isa<TerminatorInst>(Inst);
812 if (!DFSF.SkipInsts.count(Inst))
813 DFSanVisitor(DFSF).visit(Inst);
820 // We will not necessarily be able to compute the shadow for every phi node
821 // until we have visited every block. Therefore, the code that handles phi
822 // nodes adds them to the PHIFixups list so that they can be properly
824 for (std::vector<std::pair<PHINode *, PHINode *> >::iterator
825 i = DFSF.PHIFixups.begin(),
826 e = DFSF.PHIFixups.end();
828 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n;
830 i->second->setIncomingValue(
831 val, DFSF.getShadow(i->first->getIncomingValue(val)));
835 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
836 // places (i.e. instructions in basic blocks we haven't even begun visiting
837 // yet). To make our life easier, do this work in a pass after the main
839 if (ClDebugNonzeroLabels) {
840 for (Value *V : DFSF.NonZeroChecks) {
842 if (Instruction *I = dyn_cast<Instruction>(V))
843 Pos = I->getNextNode();
845 Pos = &DFSF.F->getEntryBlock().front();
846 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos))
847 Pos = Pos->getNextNode();
848 IRBuilder<> IRB(Pos);
849 Value *Ne = IRB.CreateICmpNE(V, DFSF.DFS.ZeroShadow);
850 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
851 Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
852 IRBuilder<> ThenIRB(BI);
853 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn, {});
861 Value *DFSanFunction::getArgTLSPtr() {
865 return ArgTLSPtr = DFS.ArgTLS;
867 IRBuilder<> IRB(&F->getEntryBlock().front());
868 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS, {});
871 Value *DFSanFunction::getRetvalTLS() {
875 return RetvalTLSPtr = DFS.RetvalTLS;
877 IRBuilder<> IRB(&F->getEntryBlock().front());
878 return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS, {});
881 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) {
882 IRBuilder<> IRB(Pos);
883 return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx);
886 Value *DFSanFunction::getShadow(Value *V) {
887 if (!isa<Argument>(V) && !isa<Instruction>(V))
888 return DFS.ZeroShadow;
889 Value *&Shadow = ValShadowMap[V];
891 if (Argument *A = dyn_cast<Argument>(V)) {
893 return DFS.ZeroShadow;
895 case DataFlowSanitizer::IA_TLS: {
896 Value *ArgTLSPtr = getArgTLSPtr();
897 Instruction *ArgTLSPos =
898 DFS.ArgTLS ? &*F->getEntryBlock().begin()
899 : cast<Instruction>(ArgTLSPtr)->getNextNode();
900 IRBuilder<> IRB(ArgTLSPos);
901 Shadow = IRB.CreateLoad(getArgTLS(A->getArgNo(), ArgTLSPos));
904 case DataFlowSanitizer::IA_Args: {
905 unsigned ArgIdx = A->getArgNo() + F->arg_size() / 2;
906 Function::arg_iterator i = F->arg_begin();
910 assert(Shadow->getType() == DFS.ShadowTy);
914 NonZeroChecks.push_back(Shadow);
916 Shadow = DFS.ZeroShadow;
922 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
923 assert(!ValShadowMap.count(I));
924 assert(Shadow->getType() == DFS.ShadowTy);
925 ValShadowMap[I] = Shadow;
928 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
929 assert(Addr != RetvalTLS && "Reinstrumenting?");
930 IRBuilder<> IRB(Pos);
931 Value *ShadowPtrMaskValue;
932 if (DFSanRuntimeShadowMask)
933 ShadowPtrMaskValue = IRB.CreateLoad(IntptrTy, ExternalShadowMask);
935 ShadowPtrMaskValue = ShadowPtrMask;
936 return IRB.CreateIntToPtr(
938 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy),
939 IRB.CreatePtrToInt(ShadowPtrMaskValue, IntptrTy)),
944 // Generates IR to compute the union of the two given shadows, inserting it
945 // before Pos. Returns the computed union Value.
946 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) {
947 if (V1 == DFS.ZeroShadow)
949 if (V2 == DFS.ZeroShadow)
954 auto V1Elems = ShadowElements.find(V1);
955 auto V2Elems = ShadowElements.find(V2);
956 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) {
957 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(),
958 V2Elems->second.begin(), V2Elems->second.end())) {
960 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(),
961 V1Elems->second.begin(), V1Elems->second.end())) {
964 } else if (V1Elems != ShadowElements.end()) {
965 if (V1Elems->second.count(V2))
967 } else if (V2Elems != ShadowElements.end()) {
968 if (V2Elems->second.count(V1))
972 auto Key = std::make_pair(V1, V2);
974 std::swap(Key.first, Key.second);
975 CachedCombinedShadow &CCS = CachedCombinedShadows[Key];
976 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent()))
979 IRBuilder<> IRB(Pos);
980 if (AvoidNewBlocks) {
981 CallInst *Call = IRB.CreateCall(DFS.DFSanCheckedUnionFn, {V1, V2});
982 Call->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
983 Call->addParamAttr(0, Attribute::ZExt);
984 Call->addParamAttr(1, Attribute::ZExt);
986 CCS.Block = Pos->getParent();
989 BasicBlock *Head = Pos->getParent();
990 Value *Ne = IRB.CreateICmpNE(V1, V2);
991 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
992 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT));
993 IRBuilder<> ThenIRB(BI);
994 CallInst *Call = ThenIRB.CreateCall(DFS.DFSanUnionFn, {V1, V2});
995 Call->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
996 Call->addParamAttr(0, Attribute::ZExt);
997 Call->addParamAttr(1, Attribute::ZExt);
999 BasicBlock *Tail = BI->getSuccessor(0);
1000 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1001 Phi->addIncoming(Call, Call->getParent());
1002 Phi->addIncoming(V1, Head);
1008 std::set<Value *> UnionElems;
1009 if (V1Elems != ShadowElements.end()) {
1010 UnionElems = V1Elems->second;
1012 UnionElems.insert(V1);
1014 if (V2Elems != ShadowElements.end()) {
1015 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end());
1017 UnionElems.insert(V2);
1019 ShadowElements[CCS.Shadow] = std::move(UnionElems);
1024 // A convenience function which folds the shadows of each of the operands
1025 // of the provided instruction Inst, inserting the IR before Inst. Returns
1026 // the computed union Value.
1027 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
1028 if (Inst->getNumOperands() == 0)
1029 return DFS.ZeroShadow;
1031 Value *Shadow = getShadow(Inst->getOperand(0));
1032 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) {
1033 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst);
1038 void DFSanVisitor::visitOperandShadowInst(Instruction &I) {
1039 Value *CombinedShadow = DFSF.combineOperandShadows(&I);
1040 DFSF.setShadow(&I, CombinedShadow);
1043 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where
1044 // Addr has alignment Align, and take the union of each of those shadows.
1045 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align,
1047 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1048 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1049 AllocaShadowMap.find(AI);
1050 if (i != AllocaShadowMap.end()) {
1051 IRBuilder<> IRB(Pos);
1052 return IRB.CreateLoad(i->second);
1056 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1057 SmallVector<Value *, 2> Objs;
1058 GetUnderlyingObjects(Addr, Objs, Pos->getModule()->getDataLayout());
1059 bool AllConstants = true;
1060 for (Value *Obj : Objs) {
1061 if (isa<Function>(Obj) || isa<BlockAddress>(Obj))
1063 if (isa<GlobalVariable>(Obj) && cast<GlobalVariable>(Obj)->isConstant())
1066 AllConstants = false;
1070 return DFS.ZeroShadow;
1072 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1075 return DFS.ZeroShadow;
1077 LoadInst *LI = new LoadInst(ShadowAddr, "", Pos);
1078 LI->setAlignment(ShadowAlign);
1082 IRBuilder<> IRB(Pos);
1083 Value *ShadowAddr1 = IRB.CreateGEP(DFS.ShadowTy, ShadowAddr,
1084 ConstantInt::get(DFS.IntptrTy, 1));
1085 return combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign),
1086 IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign), Pos);
1089 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) {
1090 // Fast path for the common case where each byte has identical shadow: load
1091 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any
1092 // shadow is non-equal.
1093 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F);
1094 IRBuilder<> FallbackIRB(FallbackBB);
1095 CallInst *FallbackCall = FallbackIRB.CreateCall(
1096 DFS.DFSanUnionLoadFn,
1097 {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)});
1098 FallbackCall->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
1100 // Compare each of the shadows stored in the loaded 64 bits to each other,
1101 // by computing (WideShadow rotl ShadowWidth) == WideShadow.
1102 IRBuilder<> IRB(Pos);
1104 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx));
1105 Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1106 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy);
1107 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth);
1108 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth);
1109 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow);
1110 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow);
1112 BasicBlock *Head = Pos->getParent();
1113 BasicBlock *Tail = Head->splitBasicBlock(Pos->getIterator());
1115 if (DomTreeNode *OldNode = DT.getNode(Head)) {
1116 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end());
1118 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head);
1119 for (auto Child : Children)
1120 DT.changeImmediateDominator(Child, NewNode);
1123 // In the following code LastBr will refer to the previous basic block's
1124 // conditional branch instruction, whose true successor is fixed up to point
1125 // to the next block during the loop below or to the tail after the final
1127 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq);
1128 ReplaceInstWithInst(Head->getTerminator(), LastBr);
1129 DT.addNewBlock(FallbackBB, Head);
1131 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size;
1132 Ofs += 64 / DFS.ShadowWidth) {
1133 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F);
1134 DT.addNewBlock(NextBB, LastBr->getParent());
1135 IRBuilder<> NextIRB(NextBB);
1136 WideAddr = NextIRB.CreateGEP(Type::getInt64Ty(*DFS.Ctx), WideAddr,
1137 ConstantInt::get(DFS.IntptrTy, 1));
1138 Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign);
1139 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow);
1140 LastBr->setSuccessor(0, NextBB);
1141 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB);
1144 LastBr->setSuccessor(0, Tail);
1145 FallbackIRB.CreateBr(Tail);
1146 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1147 Shadow->addIncoming(FallbackCall, FallbackBB);
1148 Shadow->addIncoming(TruncShadow, LastBr->getParent());
1152 IRBuilder<> IRB(Pos);
1153 CallInst *FallbackCall = IRB.CreateCall(
1154 DFS.DFSanUnionLoadFn, {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)});
1155 FallbackCall->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
1156 return FallbackCall;
1159 void DFSanVisitor::visitLoadInst(LoadInst &LI) {
1160 auto &DL = LI.getModule()->getDataLayout();
1161 uint64_t Size = DL.getTypeStoreSize(LI.getType());
1163 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow);
1168 if (ClPreserveAlignment) {
1169 Align = LI.getAlignment();
1171 Align = DL.getABITypeAlignment(LI.getType());
1175 IRBuilder<> IRB(&LI);
1176 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI);
1177 if (ClCombinePointerLabelsOnLoad) {
1178 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
1179 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI);
1181 if (Shadow != DFSF.DFS.ZeroShadow)
1182 DFSF.NonZeroChecks.push_back(Shadow);
1184 DFSF.setShadow(&LI, Shadow);
1187 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align,
1188 Value *Shadow, Instruction *Pos) {
1189 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1190 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i =
1191 AllocaShadowMap.find(AI);
1192 if (i != AllocaShadowMap.end()) {
1193 IRBuilder<> IRB(Pos);
1194 IRB.CreateStore(Shadow, i->second);
1199 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1200 IRBuilder<> IRB(Pos);
1201 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1202 if (Shadow == DFS.ZeroShadow) {
1203 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth);
1204 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0);
1205 Value *ExtShadowAddr =
1206 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy));
1207 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign);
1211 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth;
1212 uint64_t Offset = 0;
1213 if (Size >= ShadowVecSize) {
1214 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize);
1215 Value *ShadowVec = UndefValue::get(ShadowVecTy);
1216 for (unsigned i = 0; i != ShadowVecSize; ++i) {
1217 ShadowVec = IRB.CreateInsertElement(
1218 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i));
1220 Value *ShadowVecAddr =
1221 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy));
1223 Value *CurShadowVecAddr =
1224 IRB.CreateConstGEP1_32(ShadowVecTy, ShadowVecAddr, Offset);
1225 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign);
1226 Size -= ShadowVecSize;
1228 } while (Size >= ShadowVecSize);
1229 Offset *= ShadowVecSize;
1232 Value *CurShadowAddr =
1233 IRB.CreateConstGEP1_32(DFS.ShadowTy, ShadowAddr, Offset);
1234 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign);
1240 void DFSanVisitor::visitStoreInst(StoreInst &SI) {
1241 auto &DL = SI.getModule()->getDataLayout();
1242 uint64_t Size = DL.getTypeStoreSize(SI.getValueOperand()->getType());
1247 if (ClPreserveAlignment) {
1248 Align = SI.getAlignment();
1250 Align = DL.getABITypeAlignment(SI.getValueOperand()->getType());
1255 Value* Shadow = DFSF.getShadow(SI.getValueOperand());
1256 if (ClCombinePointerLabelsOnStore) {
1257 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand());
1258 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI);
1260 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI);
1263 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
1264 visitOperandShadowInst(BO);
1267 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); }
1269 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); }
1271 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
1272 visitOperandShadowInst(GEPI);
1275 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
1276 visitOperandShadowInst(I);
1279 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
1280 visitOperandShadowInst(I);
1283 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
1284 visitOperandShadowInst(I);
1287 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
1288 visitOperandShadowInst(I);
1291 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
1292 visitOperandShadowInst(I);
1295 void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
1296 bool AllLoadsStores = true;
1297 for (User *U : I.users()) {
1298 if (isa<LoadInst>(U))
1301 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1302 if (SI->getPointerOperand() == &I)
1306 AllLoadsStores = false;
1309 if (AllLoadsStores) {
1310 IRBuilder<> IRB(&I);
1311 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy);
1313 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow);
1316 void DFSanVisitor::visitSelectInst(SelectInst &I) {
1317 Value *CondShadow = DFSF.getShadow(I.getCondition());
1318 Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
1319 Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
1321 if (isa<VectorType>(I.getCondition()->getType())) {
1324 DFSF.combineShadows(
1325 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I));
1328 if (TrueShadow == FalseShadow) {
1329 ShadowSel = TrueShadow;
1332 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I);
1334 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I));
1338 void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
1339 IRBuilder<> IRB(&I);
1340 Value *ValShadow = DFSF.getShadow(I.getValue());
1341 IRB.CreateCall(DFSF.DFS.DFSanSetLabelFn,
1342 {ValShadow, IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(
1344 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)});
1347 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
1348 IRBuilder<> IRB(&I);
1349 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I);
1350 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I);
1351 Value *LenShadow = IRB.CreateMul(
1353 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8));
1355 if (ClPreserveAlignment) {
1356 AlignShadow = IRB.CreateMul(I.getAlignmentCst(),
1357 ConstantInt::get(I.getAlignmentCst()->getType(),
1358 DFSF.DFS.ShadowWidth / 8));
1360 AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(),
1361 DFSF.DFS.ShadowWidth / 8);
1363 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx);
1364 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr);
1365 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr);
1366 IRB.CreateCall(I.getCalledValue(), {DestShadow, SrcShadow, LenShadow,
1367 AlignShadow, I.getVolatileCst()});
1370 void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
1371 if (!DFSF.IsNativeABI && RI.getReturnValue()) {
1373 case DataFlowSanitizer::IA_TLS: {
1374 Value *S = DFSF.getShadow(RI.getReturnValue());
1375 IRBuilder<> IRB(&RI);
1376 IRB.CreateStore(S, DFSF.getRetvalTLS());
1379 case DataFlowSanitizer::IA_Args: {
1380 IRBuilder<> IRB(&RI);
1381 Type *RT = DFSF.F->getFunctionType()->getReturnType();
1383 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0);
1385 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1);
1386 RI.setOperand(0, InsShadow);
1393 void DFSanVisitor::visitCallSite(CallSite CS) {
1394 Function *F = CS.getCalledFunction();
1395 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) {
1396 visitOperandShadowInst(*CS.getInstruction());
1400 // Calls to this function are synthesized in wrappers, and we shouldn't
1402 if (F == DFSF.DFS.DFSanVarargWrapperFn)
1405 IRBuilder<> IRB(CS.getInstruction());
1407 DenseMap<Value *, Function *>::iterator i =
1408 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue());
1409 if (i != DFSF.DFS.UnwrappedFnMap.end()) {
1410 Function *F = i->second;
1411 switch (DFSF.DFS.getWrapperKind(F)) {
1412 case DataFlowSanitizer::WK_Warning: {
1413 CS.setCalledFunction(F);
1414 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn,
1415 IRB.CreateGlobalStringPtr(F->getName()));
1416 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1419 case DataFlowSanitizer::WK_Discard: {
1420 CS.setCalledFunction(F);
1421 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1424 case DataFlowSanitizer::WK_Functional: {
1425 CS.setCalledFunction(F);
1426 visitOperandShadowInst(*CS.getInstruction());
1429 case DataFlowSanitizer::WK_Custom: {
1430 // Don't try to handle invokes of custom functions, it's too complicated.
1431 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
1433 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1434 FunctionType *FT = F->getFunctionType();
1435 FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT);
1436 std::string CustomFName = "__dfsw_";
1437 CustomFName += F->getName();
1439 DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT);
1440 if (Function *CustomFn = dyn_cast<Function>(CustomF)) {
1441 CustomFn->copyAttributesFrom(F);
1443 // Custom functions returning non-void will write to the return label.
1444 if (!FT->getReturnType()->isVoidTy()) {
1445 CustomFn->removeAttributes(AttributeList::FunctionIndex,
1446 DFSF.DFS.ReadOnlyNoneAttrs);
1450 std::vector<Value *> Args;
1452 CallSite::arg_iterator i = CS.arg_begin();
1453 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) {
1454 Type *T = (*i)->getType();
1455 FunctionType *ParamFT;
1456 if (isa<PointerType>(T) &&
1457 (ParamFT = dyn_cast<FunctionType>(
1458 cast<PointerType>(T)->getElementType()))) {
1459 std::string TName = "dfst";
1460 TName += utostr(FT->getNumParams() - n);
1462 TName += F->getName();
1463 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName);
1466 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx)));
1473 const unsigned ShadowArgStart = Args.size();
1474 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1475 Args.push_back(DFSF.getShadow(*i));
1477 if (FT->isVarArg()) {
1478 auto *LabelVATy = ArrayType::get(DFSF.DFS.ShadowTy,
1479 CS.arg_size() - FT->getNumParams());
1480 auto *LabelVAAlloca = new AllocaInst(
1481 LabelVATy, getDataLayout().getAllocaAddrSpace(),
1482 "labelva", &DFSF.F->getEntryBlock().front());
1484 for (unsigned n = 0; i != CS.arg_end(); ++i, ++n) {
1485 auto LabelVAPtr = IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, n);
1486 IRB.CreateStore(DFSF.getShadow(*i), LabelVAPtr);
1489 Args.push_back(IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, 0));
1492 if (!FT->getReturnType()->isVoidTy()) {
1493 if (!DFSF.LabelReturnAlloca) {
1494 DFSF.LabelReturnAlloca =
1495 new AllocaInst(DFSF.DFS.ShadowTy,
1496 getDataLayout().getAllocaAddrSpace(),
1497 "labelreturn", &DFSF.F->getEntryBlock().front());
1499 Args.push_back(DFSF.LabelReturnAlloca);
1502 for (i = CS.arg_begin() + FT->getNumParams(); i != CS.arg_end(); ++i)
1505 CallInst *CustomCI = IRB.CreateCall(CustomF, Args);
1506 CustomCI->setCallingConv(CI->getCallingConv());
1507 CustomCI->setAttributes(CI->getAttributes());
1509 // Update the parameter attributes of the custom call instruction to
1510 // zero extend the shadow parameters. This is required for targets
1511 // which consider ShadowTy an illegal type.
1512 for (unsigned n = 0; n < FT->getNumParams(); n++) {
1513 const unsigned ArgNo = ShadowArgStart + n;
1514 if (CustomCI->getArgOperand(ArgNo)->getType() == DFSF.DFS.ShadowTy)
1515 CustomCI->addParamAttr(ArgNo, Attribute::ZExt);
1518 if (!FT->getReturnType()->isVoidTy()) {
1519 LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca);
1520 DFSF.setShadow(CustomCI, LabelLoad);
1523 CI->replaceAllUsesWith(CustomCI);
1524 CI->eraseFromParent();
1532 FunctionType *FT = cast<FunctionType>(
1533 CS.getCalledValue()->getType()->getPointerElementType());
1534 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1535 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) {
1536 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)),
1537 DFSF.getArgTLS(i, CS.getInstruction()));
1541 Instruction *Next = nullptr;
1542 if (!CS.getType()->isVoidTy()) {
1543 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1544 if (II->getNormalDest()->getSinglePredecessor()) {
1545 Next = &II->getNormalDest()->front();
1548 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT);
1549 Next = &NewBB->front();
1552 assert(CS->getIterator() != CS->getParent()->end());
1553 Next = CS->getNextNode();
1556 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1557 IRBuilder<> NextIRB(Next);
1558 LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS());
1559 DFSF.SkipInsts.insert(LI);
1560 DFSF.setShadow(CS.getInstruction(), LI);
1561 DFSF.NonZeroChecks.push_back(LI);
1565 // Do all instrumentation for IA_Args down here to defer tampering with the
1566 // CFG in a way that SplitEdge may be able to detect.
1567 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) {
1568 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT);
1570 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT));
1571 std::vector<Value *> Args;
1573 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1574 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1578 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1579 Args.push_back(DFSF.getShadow(*i));
1581 if (FT->isVarArg()) {
1582 unsigned VarArgSize = CS.arg_size() - FT->getNumParams();
1583 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize);
1584 AllocaInst *VarArgShadow =
1585 new AllocaInst(VarArgArrayTy, getDataLayout().getAllocaAddrSpace(),
1586 "", &DFSF.F->getEntryBlock().front());
1587 Args.push_back(IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, 0));
1588 for (unsigned n = 0; i != e; ++i, ++n) {
1591 IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, n));
1597 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1598 NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(),
1601 NewCS = IRB.CreateCall(Func, Args);
1603 NewCS.setCallingConv(CS.getCallingConv());
1604 NewCS.setAttributes(CS.getAttributes().removeAttributes(
1605 *DFSF.DFS.Ctx, AttributeList::ReturnIndex,
1606 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType())));
1609 ExtractValueInst *ExVal =
1610 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next);
1611 DFSF.SkipInsts.insert(ExVal);
1612 ExtractValueInst *ExShadow =
1613 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next);
1614 DFSF.SkipInsts.insert(ExShadow);
1615 DFSF.setShadow(ExVal, ExShadow);
1616 DFSF.NonZeroChecks.push_back(ExShadow);
1618 CS.getInstruction()->replaceAllUsesWith(ExVal);
1621 CS.getInstruction()->eraseFromParent();
1625 void DFSanVisitor::visitPHINode(PHINode &PN) {
1627 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN);
1629 // Give the shadow phi node valid predecessors to fool SplitEdge into working.
1630 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy);
1631 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e;
1633 ShadowPN->addIncoming(UndefShadow, *i);
1636 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN));
1637 DFSF.setShadow(&PN, ShadowPN);