1 //===-- DataFlowSanitizer.cpp - dynamic data flow analysis ----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file 10 /// This file is a part of DataFlowSanitizer, a generalised dynamic data flow 11 /// analysis. 12 /// 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. 17 /// 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: 22 /// 23 /// +--------------------+ 0x800000000000 (top of memory) 24 /// | application memory | 25 /// +--------------------+ 0x700000008000 (kAppAddr) 26 /// | | 27 /// | unused | 28 /// | | 29 /// +--------------------+ 0x200200000000 (kUnusedAddr) 30 /// | union table | 31 /// +--------------------+ 0x200000000000 (kUnionTableAddr) 32 /// | shadow memory | 33 /// +--------------------+ 0x000000010000 (kShadowAddr) 34 /// | reserved by kernel | 35 /// +--------------------+ 0x000000000000 36 /// 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. 43 /// 44 /// For more information, please refer to the design document: 45 /// http://clang.llvm.org/docs/DataFlowSanitizerDesign.html 46 47 #include "llvm/Transforms/Instrumentation.h" 48 #include "llvm/ADT/DenseMap.h" 49 #include "llvm/ADT/DenseSet.h" 50 #include "llvm/ADT/DepthFirstIterator.h" 51 #include "llvm/ADT/StringExtras.h" 52 #include "llvm/ADT/Triple.h" 53 #include "llvm/Analysis/ValueTracking.h" 54 #include "llvm/IR/Dominators.h" 55 #include "llvm/IR/DebugInfo.h" 56 #include "llvm/IR/IRBuilder.h" 57 #include "llvm/IR/InlineAsm.h" 58 #include "llvm/IR/InstVisitor.h" 59 #include "llvm/IR/LLVMContext.h" 60 #include "llvm/IR/MDBuilder.h" 61 #include "llvm/IR/Type.h" 62 #include "llvm/IR/Value.h" 63 #include "llvm/Pass.h" 64 #include "llvm/Support/CommandLine.h" 65 #include "llvm/Support/SpecialCaseList.h" 66 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 67 #include "llvm/Transforms/Utils/Local.h" 68 #include <algorithm> 69 #include <iterator> 70 #include <set> 71 #include <utility> 72 73 using namespace llvm; 74 75 // The -dfsan-preserve-alignment flag controls whether this pass assumes that 76 // alignment requirements provided by the input IR are correct. For example, 77 // if the input IR contains a load with alignment 8, this flag will cause 78 // the shadow load to have alignment 16. This flag is disabled by default as 79 // we have unfortunately encountered too much code (including Clang itself; 80 // see PR14291) which performs misaligned access. 81 static cl::opt<bool> ClPreserveAlignment( 82 "dfsan-preserve-alignment", 83 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden, 84 cl::init(false)); 85 86 // The ABI list files control how shadow parameters are passed. The pass treats 87 // every function labelled "uninstrumented" in the ABI list file as conforming 88 // to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains 89 // additional annotations for those functions, a call to one of those functions 90 // will produce a warning message, as the labelling behaviour of the function is 91 // unknown. The other supported annotations are "functional" and "discard", 92 // which are described below under DataFlowSanitizer::WrapperKind. 93 static cl::list<std::string> ClABIListFiles( 94 "dfsan-abilist", 95 cl::desc("File listing native ABI functions and how the pass treats them"), 96 cl::Hidden); 97 98 // Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented 99 // functions (see DataFlowSanitizer::InstrumentedABI below). 100 static cl::opt<bool> ClArgsABI( 101 "dfsan-args-abi", 102 cl::desc("Use the argument ABI rather than the TLS ABI"), 103 cl::Hidden); 104 105 // Controls whether the pass includes or ignores the labels of pointers in load 106 // instructions. 107 static cl::opt<bool> ClCombinePointerLabelsOnLoad( 108 "dfsan-combine-pointer-labels-on-load", 109 cl::desc("Combine the label of the pointer with the label of the data when " 110 "loading from memory."), 111 cl::Hidden, cl::init(true)); 112 113 // Controls whether the pass includes or ignores the labels of pointers in 114 // stores instructions. 115 static cl::opt<bool> ClCombinePointerLabelsOnStore( 116 "dfsan-combine-pointer-labels-on-store", 117 cl::desc("Combine the label of the pointer with the label of the data when " 118 "storing in memory."), 119 cl::Hidden, cl::init(false)); 120 121 static cl::opt<bool> ClDebugNonzeroLabels( 122 "dfsan-debug-nonzero-labels", 123 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, " 124 "load or return with a nonzero label"), 125 cl::Hidden); 126 127 namespace { 128 129 StringRef GetGlobalTypeString(const GlobalValue &G) { 130 // Types of GlobalVariables are always pointer types. 131 Type *GType = G.getType()->getElementType(); 132 // For now we support blacklisting struct types only. 133 if (StructType *SGType = dyn_cast<StructType>(GType)) { 134 if (!SGType->isLiteral()) 135 return SGType->getName(); 136 } 137 return "<unknown type>"; 138 } 139 140 class DFSanABIList { 141 std::unique_ptr<SpecialCaseList> SCL; 142 143 public: 144 DFSanABIList() {} 145 146 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); } 147 148 /// Returns whether either this function or its source file are listed in the 149 /// given category. 150 bool isIn(const Function &F, StringRef Category) const { 151 return isIn(*F.getParent(), Category) || 152 SCL->inSection("fun", F.getName(), Category); 153 } 154 155 /// Returns whether this global alias is listed in the given category. 156 /// 157 /// If GA aliases a function, the alias's name is matched as a function name 158 /// would be. Similarly, aliases of globals are matched like globals. 159 bool isIn(const GlobalAlias &GA, StringRef Category) const { 160 if (isIn(*GA.getParent(), Category)) 161 return true; 162 163 if (isa<FunctionType>(GA.getType()->getElementType())) 164 return SCL->inSection("fun", GA.getName(), Category); 165 166 return SCL->inSection("global", GA.getName(), Category) || 167 SCL->inSection("type", GetGlobalTypeString(GA), Category); 168 } 169 170 /// Returns whether this module is listed in the given category. 171 bool isIn(const Module &M, StringRef Category) const { 172 return SCL->inSection("src", M.getModuleIdentifier(), Category); 173 } 174 }; 175 176 class DataFlowSanitizer : public ModulePass { 177 friend struct DFSanFunction; 178 friend class DFSanVisitor; 179 180 enum { 181 ShadowWidth = 16 182 }; 183 184 /// Which ABI should be used for instrumented functions? 185 enum InstrumentedABI { 186 /// Argument and return value labels are passed through additional 187 /// arguments and by modifying the return type. 188 IA_Args, 189 190 /// Argument and return value labels are passed through TLS variables 191 /// __dfsan_arg_tls and __dfsan_retval_tls. 192 IA_TLS 193 }; 194 195 /// How should calls to uninstrumented functions be handled? 196 enum WrapperKind { 197 /// This function is present in an uninstrumented form but we don't know 198 /// how it should be handled. Print a warning and call the function anyway. 199 /// Don't label the return value. 200 WK_Warning, 201 202 /// This function does not write to (user-accessible) memory, and its return 203 /// value is unlabelled. 204 WK_Discard, 205 206 /// This function does not write to (user-accessible) memory, and the label 207 /// of its return value is the union of the label of its arguments. 208 WK_Functional, 209 210 /// Instead of calling the function, a custom wrapper __dfsw_F is called, 211 /// where F is the name of the function. This function may wrap the 212 /// original function or provide its own implementation. This is similar to 213 /// the IA_Args ABI, except that IA_Args uses a struct return type to 214 /// pass the return value shadow in a register, while WK_Custom uses an 215 /// extra pointer argument to return the shadow. This allows the wrapped 216 /// form of the function type to be expressed in C. 217 WK_Custom 218 }; 219 220 Module *Mod; 221 LLVMContext *Ctx; 222 IntegerType *ShadowTy; 223 PointerType *ShadowPtrTy; 224 IntegerType *IntptrTy; 225 ConstantInt *ZeroShadow; 226 ConstantInt *ShadowPtrMask; 227 ConstantInt *ShadowPtrMul; 228 Constant *ArgTLS; 229 Constant *RetvalTLS; 230 void *(*GetArgTLSPtr)(); 231 void *(*GetRetvalTLSPtr)(); 232 Constant *GetArgTLS; 233 Constant *GetRetvalTLS; 234 FunctionType *DFSanUnionFnTy; 235 FunctionType *DFSanUnionLoadFnTy; 236 FunctionType *DFSanUnimplementedFnTy; 237 FunctionType *DFSanSetLabelFnTy; 238 FunctionType *DFSanNonzeroLabelFnTy; 239 FunctionType *DFSanVarargWrapperFnTy; 240 Constant *DFSanUnionFn; 241 Constant *DFSanCheckedUnionFn; 242 Constant *DFSanUnionLoadFn; 243 Constant *DFSanUnimplementedFn; 244 Constant *DFSanSetLabelFn; 245 Constant *DFSanNonzeroLabelFn; 246 Constant *DFSanVarargWrapperFn; 247 MDNode *ColdCallWeights; 248 DFSanABIList ABIList; 249 DenseMap<Value *, Function *> UnwrappedFnMap; 250 AttributeSet ReadOnlyNoneAttrs; 251 DenseMap<const Function *, DISubprogram> FunctionDIs; 252 253 Value *getShadowAddress(Value *Addr, Instruction *Pos); 254 bool isInstrumented(const Function *F); 255 bool isInstrumented(const GlobalAlias *GA); 256 FunctionType *getArgsFunctionType(FunctionType *T); 257 FunctionType *getTrampolineFunctionType(FunctionType *T); 258 FunctionType *getCustomFunctionType(FunctionType *T); 259 InstrumentedABI getInstrumentedABI(); 260 WrapperKind getWrapperKind(Function *F); 261 void addGlobalNamePrefix(GlobalValue *GV); 262 Function *buildWrapperFunction(Function *F, StringRef NewFName, 263 GlobalValue::LinkageTypes NewFLink, 264 FunctionType *NewFT); 265 Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName); 266 267 public: 268 DataFlowSanitizer( 269 const std::vector<std::string> &ABIListFiles = std::vector<std::string>(), 270 void *(*getArgTLS)() = nullptr, void *(*getRetValTLS)() = nullptr); 271 static char ID; 272 bool doInitialization(Module &M) override; 273 bool runOnModule(Module &M) override; 274 }; 275 276 struct DFSanFunction { 277 DataFlowSanitizer &DFS; 278 Function *F; 279 DominatorTree DT; 280 DataFlowSanitizer::InstrumentedABI IA; 281 bool IsNativeABI; 282 Value *ArgTLSPtr; 283 Value *RetvalTLSPtr; 284 AllocaInst *LabelReturnAlloca; 285 DenseMap<Value *, Value *> ValShadowMap; 286 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap; 287 std::vector<std::pair<PHINode *, PHINode *> > PHIFixups; 288 DenseSet<Instruction *> SkipInsts; 289 std::vector<Value *> NonZeroChecks; 290 bool AvoidNewBlocks; 291 292 struct CachedCombinedShadow { 293 BasicBlock *Block; 294 Value *Shadow; 295 }; 296 DenseMap<std::pair<Value *, Value *>, CachedCombinedShadow> 297 CachedCombinedShadows; 298 DenseMap<Value *, std::set<Value *>> ShadowElements; 299 300 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI) 301 : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()), 302 IsNativeABI(IsNativeABI), ArgTLSPtr(nullptr), RetvalTLSPtr(nullptr), 303 LabelReturnAlloca(nullptr) { 304 DT.recalculate(*F); 305 // FIXME: Need to track down the register allocator issue which causes poor 306 // performance in pathological cases with large numbers of basic blocks. 307 AvoidNewBlocks = F->size() > 1000; 308 } 309 Value *getArgTLSPtr(); 310 Value *getArgTLS(unsigned Index, Instruction *Pos); 311 Value *getRetvalTLS(); 312 Value *getShadow(Value *V); 313 void setShadow(Instruction *I, Value *Shadow); 314 Value *combineShadows(Value *V1, Value *V2, Instruction *Pos); 315 Value *combineOperandShadows(Instruction *Inst); 316 Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align, 317 Instruction *Pos); 318 void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow, 319 Instruction *Pos); 320 }; 321 322 class DFSanVisitor : public InstVisitor<DFSanVisitor> { 323 public: 324 DFSanFunction &DFSF; 325 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {} 326 327 void visitOperandShadowInst(Instruction &I); 328 329 void visitBinaryOperator(BinaryOperator &BO); 330 void visitCastInst(CastInst &CI); 331 void visitCmpInst(CmpInst &CI); 332 void visitGetElementPtrInst(GetElementPtrInst &GEPI); 333 void visitLoadInst(LoadInst &LI); 334 void visitStoreInst(StoreInst &SI); 335 void visitReturnInst(ReturnInst &RI); 336 void visitCallSite(CallSite CS); 337 void visitPHINode(PHINode &PN); 338 void visitExtractElementInst(ExtractElementInst &I); 339 void visitInsertElementInst(InsertElementInst &I); 340 void visitShuffleVectorInst(ShuffleVectorInst &I); 341 void visitExtractValueInst(ExtractValueInst &I); 342 void visitInsertValueInst(InsertValueInst &I); 343 void visitAllocaInst(AllocaInst &I); 344 void visitSelectInst(SelectInst &I); 345 void visitMemSetInst(MemSetInst &I); 346 void visitMemTransferInst(MemTransferInst &I); 347 }; 348 349 } 350 351 char DataFlowSanitizer::ID; 352 INITIALIZE_PASS(DataFlowSanitizer, "dfsan", 353 "DataFlowSanitizer: dynamic data flow analysis.", false, false) 354 355 ModulePass * 356 llvm::createDataFlowSanitizerPass(const std::vector<std::string> &ABIListFiles, 357 void *(*getArgTLS)(), 358 void *(*getRetValTLS)()) { 359 return new DataFlowSanitizer(ABIListFiles, getArgTLS, getRetValTLS); 360 } 361 362 DataFlowSanitizer::DataFlowSanitizer( 363 const std::vector<std::string> &ABIListFiles, void *(*getArgTLS)(), 364 void *(*getRetValTLS)()) 365 : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS) { 366 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles)); 367 AllABIListFiles.insert(AllABIListFiles.end(), ClABIListFiles.begin(), 368 ClABIListFiles.end()); 369 ABIList.set(SpecialCaseList::createOrDie(AllABIListFiles)); 370 } 371 372 FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) { 373 llvm::SmallVector<Type *, 4> ArgTypes(T->param_begin(), T->param_end()); 374 ArgTypes.append(T->getNumParams(), ShadowTy); 375 if (T->isVarArg()) 376 ArgTypes.push_back(ShadowPtrTy); 377 Type *RetType = T->getReturnType(); 378 if (!RetType->isVoidTy()) 379 RetType = StructType::get(RetType, ShadowTy, (Type *)nullptr); 380 return FunctionType::get(RetType, ArgTypes, T->isVarArg()); 381 } 382 383 FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) { 384 assert(!T->isVarArg()); 385 llvm::SmallVector<Type *, 4> ArgTypes; 386 ArgTypes.push_back(T->getPointerTo()); 387 ArgTypes.append(T->param_begin(), T->param_end()); 388 ArgTypes.append(T->getNumParams(), ShadowTy); 389 Type *RetType = T->getReturnType(); 390 if (!RetType->isVoidTy()) 391 ArgTypes.push_back(ShadowPtrTy); 392 return FunctionType::get(T->getReturnType(), ArgTypes, false); 393 } 394 395 FunctionType *DataFlowSanitizer::getCustomFunctionType(FunctionType *T) { 396 llvm::SmallVector<Type *, 4> ArgTypes; 397 for (FunctionType::param_iterator i = T->param_begin(), e = T->param_end(); 398 i != e; ++i) { 399 FunctionType *FT; 400 if (isa<PointerType>(*i) && (FT = dyn_cast<FunctionType>(cast<PointerType>( 401 *i)->getElementType()))) { 402 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo()); 403 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx)); 404 } else { 405 ArgTypes.push_back(*i); 406 } 407 } 408 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i) 409 ArgTypes.push_back(ShadowTy); 410 if (T->isVarArg()) 411 ArgTypes.push_back(ShadowPtrTy); 412 Type *RetType = T->getReturnType(); 413 if (!RetType->isVoidTy()) 414 ArgTypes.push_back(ShadowPtrTy); 415 return FunctionType::get(T->getReturnType(), ArgTypes, T->isVarArg()); 416 } 417 418 bool DataFlowSanitizer::doInitialization(Module &M) { 419 llvm::Triple TargetTriple(M.getTargetTriple()); 420 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64; 421 bool IsMIPS64 = TargetTriple.getArch() == llvm::Triple::mips64 || 422 TargetTriple.getArch() == llvm::Triple::mips64el; 423 424 const DataLayout &DL = M.getDataLayout(); 425 426 Mod = &M; 427 Ctx = &M.getContext(); 428 ShadowTy = IntegerType::get(*Ctx, ShadowWidth); 429 ShadowPtrTy = PointerType::getUnqual(ShadowTy); 430 IntptrTy = DL.getIntPtrType(*Ctx); 431 ZeroShadow = ConstantInt::getSigned(ShadowTy, 0); 432 ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8); 433 if (IsX86_64) 434 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL); 435 else if (IsMIPS64) 436 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0xF000000000LL); 437 else 438 report_fatal_error("unsupported triple"); 439 440 Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy }; 441 DFSanUnionFnTy = 442 FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false); 443 Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy }; 444 DFSanUnionLoadFnTy = 445 FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false); 446 DFSanUnimplementedFnTy = FunctionType::get( 447 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false); 448 Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy }; 449 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx), 450 DFSanSetLabelArgs, /*isVarArg=*/false); 451 DFSanNonzeroLabelFnTy = FunctionType::get( 452 Type::getVoidTy(*Ctx), None, /*isVarArg=*/false); 453 DFSanVarargWrapperFnTy = FunctionType::get( 454 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false); 455 456 if (GetArgTLSPtr) { 457 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64); 458 ArgTLS = nullptr; 459 GetArgTLS = ConstantExpr::getIntToPtr( 460 ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)), 461 PointerType::getUnqual( 462 FunctionType::get(PointerType::getUnqual(ArgTLSTy), 463 (Type *)nullptr))); 464 } 465 if (GetRetvalTLSPtr) { 466 RetvalTLS = nullptr; 467 GetRetvalTLS = ConstantExpr::getIntToPtr( 468 ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)), 469 PointerType::getUnqual( 470 FunctionType::get(PointerType::getUnqual(ShadowTy), 471 (Type *)nullptr))); 472 } 473 474 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000); 475 return true; 476 } 477 478 bool DataFlowSanitizer::isInstrumented(const Function *F) { 479 return !ABIList.isIn(*F, "uninstrumented"); 480 } 481 482 bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) { 483 return !ABIList.isIn(*GA, "uninstrumented"); 484 } 485 486 DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() { 487 return ClArgsABI ? IA_Args : IA_TLS; 488 } 489 490 DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) { 491 if (ABIList.isIn(*F, "functional")) 492 return WK_Functional; 493 if (ABIList.isIn(*F, "discard")) 494 return WK_Discard; 495 if (ABIList.isIn(*F, "custom")) 496 return WK_Custom; 497 498 return WK_Warning; 499 } 500 501 void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) { 502 std::string GVName = GV->getName(), Prefix = "dfs$"; 503 GV->setName(Prefix + GVName); 504 505 // Try to change the name of the function in module inline asm. We only do 506 // this for specific asm directives, currently only ".symver", to try to avoid 507 // corrupting asm which happens to contain the symbol name as a substring. 508 // Note that the substitution for .symver assumes that the versioned symbol 509 // also has an instrumented name. 510 std::string Asm = GV->getParent()->getModuleInlineAsm(); 511 std::string SearchStr = ".symver " + GVName + ","; 512 size_t Pos = Asm.find(SearchStr); 513 if (Pos != std::string::npos) { 514 Asm.replace(Pos, SearchStr.size(), 515 ".symver " + Prefix + GVName + "," + Prefix); 516 GV->getParent()->setModuleInlineAsm(Asm); 517 } 518 } 519 520 Function * 521 DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName, 522 GlobalValue::LinkageTypes NewFLink, 523 FunctionType *NewFT) { 524 FunctionType *FT = F->getFunctionType(); 525 Function *NewF = Function::Create(NewFT, NewFLink, NewFName, 526 F->getParent()); 527 NewF->copyAttributesFrom(F); 528 NewF->removeAttributes( 529 AttributeSet::ReturnIndex, 530 AttributeFuncs::typeIncompatible(NewFT->getReturnType(), 531 AttributeSet::ReturnIndex)); 532 533 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF); 534 if (F->isVarArg()) { 535 NewF->removeAttributes( 536 AttributeSet::FunctionIndex, 537 AttributeSet().addAttribute(*Ctx, AttributeSet::FunctionIndex, 538 "split-stack")); 539 CallInst::Create(DFSanVarargWrapperFn, 540 IRBuilder<>(BB).CreateGlobalStringPtr(F->getName()), "", 541 BB); 542 new UnreachableInst(*Ctx, BB); 543 } else { 544 std::vector<Value *> Args; 545 unsigned n = FT->getNumParams(); 546 for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n) 547 Args.push_back(&*ai); 548 CallInst *CI = CallInst::Create(F, Args, "", BB); 549 if (FT->getReturnType()->isVoidTy()) 550 ReturnInst::Create(*Ctx, BB); 551 else 552 ReturnInst::Create(*Ctx, CI, BB); 553 } 554 555 return NewF; 556 } 557 558 Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT, 559 StringRef FName) { 560 FunctionType *FTT = getTrampolineFunctionType(FT); 561 Constant *C = Mod->getOrInsertFunction(FName, FTT); 562 Function *F = dyn_cast<Function>(C); 563 if (F && F->isDeclaration()) { 564 F->setLinkage(GlobalValue::LinkOnceODRLinkage); 565 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F); 566 std::vector<Value *> Args; 567 Function::arg_iterator AI = F->arg_begin(); ++AI; 568 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N) 569 Args.push_back(&*AI); 570 CallInst *CI = 571 CallInst::Create(&F->getArgumentList().front(), Args, "", BB); 572 ReturnInst *RI; 573 if (FT->getReturnType()->isVoidTy()) 574 RI = ReturnInst::Create(*Ctx, BB); 575 else 576 RI = ReturnInst::Create(*Ctx, CI, BB); 577 578 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true); 579 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI; 580 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N) 581 DFSF.ValShadowMap[ValAI] = ShadowAI; 582 DFSanVisitor(DFSF).visitCallInst(*CI); 583 if (!FT->getReturnType()->isVoidTy()) 584 new StoreInst(DFSF.getShadow(RI->getReturnValue()), 585 &F->getArgumentList().back(), RI); 586 } 587 588 return C; 589 } 590 591 bool DataFlowSanitizer::runOnModule(Module &M) { 592 if (ABIList.isIn(M, "skip")) 593 return false; 594 595 FunctionDIs = makeSubprogramMap(M); 596 597 if (!GetArgTLSPtr) { 598 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64); 599 ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy); 600 if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS)) 601 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel); 602 } 603 if (!GetRetvalTLSPtr) { 604 RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy); 605 if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS)) 606 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel); 607 } 608 609 DFSanUnionFn = Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy); 610 if (Function *F = dyn_cast<Function>(DFSanUnionFn)) { 611 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind); 612 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone); 613 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 614 F->addAttribute(1, Attribute::ZExt); 615 F->addAttribute(2, Attribute::ZExt); 616 } 617 DFSanCheckedUnionFn = Mod->getOrInsertFunction("dfsan_union", DFSanUnionFnTy); 618 if (Function *F = dyn_cast<Function>(DFSanCheckedUnionFn)) { 619 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind); 620 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone); 621 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 622 F->addAttribute(1, Attribute::ZExt); 623 F->addAttribute(2, Attribute::ZExt); 624 } 625 DFSanUnionLoadFn = 626 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy); 627 if (Function *F = dyn_cast<Function>(DFSanUnionLoadFn)) { 628 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind); 629 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadOnly); 630 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 631 } 632 DFSanUnimplementedFn = 633 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy); 634 DFSanSetLabelFn = 635 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy); 636 if (Function *F = dyn_cast<Function>(DFSanSetLabelFn)) { 637 F->addAttribute(1, Attribute::ZExt); 638 } 639 DFSanNonzeroLabelFn = 640 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy); 641 DFSanVarargWrapperFn = Mod->getOrInsertFunction("__dfsan_vararg_wrapper", 642 DFSanVarargWrapperFnTy); 643 644 std::vector<Function *> FnsToInstrument; 645 llvm::SmallPtrSet<Function *, 2> FnsWithNativeABI; 646 for (Module::iterator i = M.begin(), e = M.end(); i != e; ++i) { 647 if (!i->isIntrinsic() && 648 i != DFSanUnionFn && 649 i != DFSanCheckedUnionFn && 650 i != DFSanUnionLoadFn && 651 i != DFSanUnimplementedFn && 652 i != DFSanSetLabelFn && 653 i != DFSanNonzeroLabelFn && 654 i != DFSanVarargWrapperFn) 655 FnsToInstrument.push_back(&*i); 656 } 657 658 // Give function aliases prefixes when necessary, and build wrappers where the 659 // instrumentedness is inconsistent. 660 for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) { 661 GlobalAlias *GA = &*i; 662 ++i; 663 // Don't stop on weak. We assume people aren't playing games with the 664 // instrumentedness of overridden weak aliases. 665 if (auto F = dyn_cast<Function>(GA->getBaseObject())) { 666 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F); 667 if (GAInst && FInst) { 668 addGlobalNamePrefix(GA); 669 } else if (GAInst != FInst) { 670 // Non-instrumented alias of an instrumented function, or vice versa. 671 // Replace the alias with a native-ABI wrapper of the aliasee. The pass 672 // below will take care of instrumenting it. 673 Function *NewF = 674 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType()); 675 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType())); 676 NewF->takeName(GA); 677 GA->eraseFromParent(); 678 FnsToInstrument.push_back(NewF); 679 } 680 } 681 } 682 683 AttrBuilder B; 684 B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone); 685 ReadOnlyNoneAttrs = AttributeSet::get(*Ctx, AttributeSet::FunctionIndex, B); 686 687 // First, change the ABI of every function in the module. ABI-listed 688 // functions keep their original ABI and get a wrapper function. 689 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(), 690 e = FnsToInstrument.end(); 691 i != e; ++i) { 692 Function &F = **i; 693 FunctionType *FT = F.getFunctionType(); 694 695 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() && 696 FT->getReturnType()->isVoidTy()); 697 698 if (isInstrumented(&F)) { 699 // Instrumented functions get a 'dfs$' prefix. This allows us to more 700 // easily identify cases of mismatching ABIs. 701 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) { 702 FunctionType *NewFT = getArgsFunctionType(FT); 703 Function *NewF = Function::Create(NewFT, F.getLinkage(), "", &M); 704 NewF->copyAttributesFrom(&F); 705 NewF->removeAttributes( 706 AttributeSet::ReturnIndex, 707 AttributeFuncs::typeIncompatible(NewFT->getReturnType(), 708 AttributeSet::ReturnIndex)); 709 for (Function::arg_iterator FArg = F.arg_begin(), 710 NewFArg = NewF->arg_begin(), 711 FArgEnd = F.arg_end(); 712 FArg != FArgEnd; ++FArg, ++NewFArg) { 713 FArg->replaceAllUsesWith(NewFArg); 714 } 715 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList()); 716 717 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end(); 718 UI != UE;) { 719 BlockAddress *BA = dyn_cast<BlockAddress>(*UI); 720 ++UI; 721 if (BA) { 722 BA->replaceAllUsesWith( 723 BlockAddress::get(NewF, BA->getBasicBlock())); 724 delete BA; 725 } 726 } 727 F.replaceAllUsesWith( 728 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT))); 729 NewF->takeName(&F); 730 F.eraseFromParent(); 731 *i = NewF; 732 addGlobalNamePrefix(NewF); 733 } else { 734 addGlobalNamePrefix(&F); 735 } 736 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) { 737 // Build a wrapper function for F. The wrapper simply calls F, and is 738 // added to FnsToInstrument so that any instrumentation according to its 739 // WrapperKind is done in the second pass below. 740 FunctionType *NewFT = getInstrumentedABI() == IA_Args 741 ? getArgsFunctionType(FT) 742 : FT; 743 Function *NewF = buildWrapperFunction( 744 &F, std::string("dfsw$") + std::string(F.getName()), 745 GlobalValue::LinkOnceODRLinkage, NewFT); 746 if (getInstrumentedABI() == IA_TLS) 747 NewF->removeAttributes(AttributeSet::FunctionIndex, ReadOnlyNoneAttrs); 748 749 Value *WrappedFnCst = 750 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)); 751 F.replaceAllUsesWith(WrappedFnCst); 752 753 // Patch the pointer to LLVM function in debug info descriptor. 754 auto DI = FunctionDIs.find(&F); 755 if (DI != FunctionDIs.end()) 756 DI->second->replaceFunction(&F); 757 758 UnwrappedFnMap[WrappedFnCst] = &F; 759 *i = NewF; 760 761 if (!F.isDeclaration()) { 762 // This function is probably defining an interposition of an 763 // uninstrumented function and hence needs to keep the original ABI. 764 // But any functions it may call need to use the instrumented ABI, so 765 // we instrument it in a mode which preserves the original ABI. 766 FnsWithNativeABI.insert(&F); 767 768 // This code needs to rebuild the iterators, as they may be invalidated 769 // by the push_back, taking care that the new range does not include 770 // any functions added by this code. 771 size_t N = i - FnsToInstrument.begin(), 772 Count = e - FnsToInstrument.begin(); 773 FnsToInstrument.push_back(&F); 774 i = FnsToInstrument.begin() + N; 775 e = FnsToInstrument.begin() + Count; 776 } 777 // Hopefully, nobody will try to indirectly call a vararg 778 // function... yet. 779 } else if (FT->isVarArg()) { 780 UnwrappedFnMap[&F] = &F; 781 *i = nullptr; 782 } 783 } 784 785 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(), 786 e = FnsToInstrument.end(); 787 i != e; ++i) { 788 if (!*i || (*i)->isDeclaration()) 789 continue; 790 791 removeUnreachableBlocks(**i); 792 793 DFSanFunction DFSF(*this, *i, FnsWithNativeABI.count(*i)); 794 795 // DFSanVisitor may create new basic blocks, which confuses df_iterator. 796 // Build a copy of the list before iterating over it. 797 llvm::SmallVector<BasicBlock *, 4> BBList( 798 depth_first(&(*i)->getEntryBlock())); 799 800 for (llvm::SmallVector<BasicBlock *, 4>::iterator i = BBList.begin(), 801 e = BBList.end(); 802 i != e; ++i) { 803 Instruction *Inst = &(*i)->front(); 804 while (1) { 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 810 // terminator. 811 bool IsTerminator = isa<TerminatorInst>(Inst); 812 if (!DFSF.SkipInsts.count(Inst)) 813 DFSanVisitor(DFSF).visit(Inst); 814 if (IsTerminator) 815 break; 816 Inst = Next; 817 } 818 } 819 820 // 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 823 // handled here. 824 for (std::vector<std::pair<PHINode *, PHINode *> >::iterator 825 i = DFSF.PHIFixups.begin(), 826 e = DFSF.PHIFixups.end(); 827 i != e; ++i) { 828 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n; 829 ++val) { 830 i->second->setIncomingValue( 831 val, DFSF.getShadow(i->first->getIncomingValue(val))); 832 } 833 } 834 835 // -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 838 // instrumentation. 839 if (ClDebugNonzeroLabels) { 840 for (Value *V : DFSF.NonZeroChecks) { 841 Instruction *Pos; 842 if (Instruction *I = dyn_cast<Instruction>(V)) 843 Pos = I->getNextNode(); 844 else 845 Pos = DFSF.F->getEntryBlock().begin(); 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); 854 } 855 } 856 } 857 858 return false; 859 } 860 861 Value *DFSanFunction::getArgTLSPtr() { 862 if (ArgTLSPtr) 863 return ArgTLSPtr; 864 if (DFS.ArgTLS) 865 return ArgTLSPtr = DFS.ArgTLS; 866 867 IRBuilder<> IRB(F->getEntryBlock().begin()); 868 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS); 869 } 870 871 Value *DFSanFunction::getRetvalTLS() { 872 if (RetvalTLSPtr) 873 return RetvalTLSPtr; 874 if (DFS.RetvalTLS) 875 return RetvalTLSPtr = DFS.RetvalTLS; 876 877 IRBuilder<> IRB(F->getEntryBlock().begin()); 878 return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS); 879 } 880 881 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) { 882 IRBuilder<> IRB(Pos); 883 return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx); 884 } 885 886 Value *DFSanFunction::getShadow(Value *V) { 887 if (!isa<Argument>(V) && !isa<Instruction>(V)) 888 return DFS.ZeroShadow; 889 Value *&Shadow = ValShadowMap[V]; 890 if (!Shadow) { 891 if (Argument *A = dyn_cast<Argument>(V)) { 892 if (IsNativeABI) 893 return DFS.ZeroShadow; 894 switch (IA) { 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)); 902 break; 903 } 904 case DataFlowSanitizer::IA_Args: { 905 unsigned ArgIdx = A->getArgNo() + F->getArgumentList().size() / 2; 906 Function::arg_iterator i = F->arg_begin(); 907 while (ArgIdx--) 908 ++i; 909 Shadow = i; 910 assert(Shadow->getType() == DFS.ShadowTy); 911 break; 912 } 913 } 914 NonZeroChecks.push_back(Shadow); 915 } else { 916 Shadow = DFS.ZeroShadow; 917 } 918 } 919 return Shadow; 920 } 921 922 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) { 923 assert(!ValShadowMap.count(I)); 924 assert(Shadow->getType() == DFS.ShadowTy); 925 ValShadowMap[I] = Shadow; 926 } 927 928 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) { 929 assert(Addr != RetvalTLS && "Reinstrumenting?"); 930 IRBuilder<> IRB(Pos); 931 return IRB.CreateIntToPtr( 932 IRB.CreateMul( 933 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy), ShadowPtrMask), 934 ShadowPtrMul), 935 ShadowPtrTy); 936 } 937 938 // Generates IR to compute the union of the two given shadows, inserting it 939 // before Pos. Returns the computed union Value. 940 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) { 941 if (V1 == DFS.ZeroShadow) 942 return V2; 943 if (V2 == DFS.ZeroShadow) 944 return V1; 945 if (V1 == V2) 946 return V1; 947 948 auto V1Elems = ShadowElements.find(V1); 949 auto V2Elems = ShadowElements.find(V2); 950 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) { 951 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(), 952 V2Elems->second.begin(), V2Elems->second.end())) { 953 return V1; 954 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(), 955 V1Elems->second.begin(), V1Elems->second.end())) { 956 return V2; 957 } 958 } else if (V1Elems != ShadowElements.end()) { 959 if (V1Elems->second.count(V2)) 960 return V1; 961 } else if (V2Elems != ShadowElements.end()) { 962 if (V2Elems->second.count(V1)) 963 return V2; 964 } 965 966 auto Key = std::make_pair(V1, V2); 967 if (V1 > V2) 968 std::swap(Key.first, Key.second); 969 CachedCombinedShadow &CCS = CachedCombinedShadows[Key]; 970 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent())) 971 return CCS.Shadow; 972 973 IRBuilder<> IRB(Pos); 974 if (AvoidNewBlocks) { 975 CallInst *Call = IRB.CreateCall2(DFS.DFSanCheckedUnionFn, V1, V2); 976 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 977 Call->addAttribute(1, Attribute::ZExt); 978 Call->addAttribute(2, Attribute::ZExt); 979 980 CCS.Block = Pos->getParent(); 981 CCS.Shadow = Call; 982 } else { 983 BasicBlock *Head = Pos->getParent(); 984 Value *Ne = IRB.CreateICmpNE(V1, V2); 985 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen( 986 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT)); 987 IRBuilder<> ThenIRB(BI); 988 CallInst *Call = ThenIRB.CreateCall2(DFS.DFSanUnionFn, V1, V2); 989 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 990 Call->addAttribute(1, Attribute::ZExt); 991 Call->addAttribute(2, Attribute::ZExt); 992 993 BasicBlock *Tail = BI->getSuccessor(0); 994 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", Tail->begin()); 995 Phi->addIncoming(Call, Call->getParent()); 996 Phi->addIncoming(V1, Head); 997 998 CCS.Block = Tail; 999 CCS.Shadow = Phi; 1000 } 1001 1002 std::set<Value *> UnionElems; 1003 if (V1Elems != ShadowElements.end()) { 1004 UnionElems = V1Elems->second; 1005 } else { 1006 UnionElems.insert(V1); 1007 } 1008 if (V2Elems != ShadowElements.end()) { 1009 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end()); 1010 } else { 1011 UnionElems.insert(V2); 1012 } 1013 ShadowElements[CCS.Shadow] = std::move(UnionElems); 1014 1015 return CCS.Shadow; 1016 } 1017 1018 // A convenience function which folds the shadows of each of the operands 1019 // of the provided instruction Inst, inserting the IR before Inst. Returns 1020 // the computed union Value. 1021 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) { 1022 if (Inst->getNumOperands() == 0) 1023 return DFS.ZeroShadow; 1024 1025 Value *Shadow = getShadow(Inst->getOperand(0)); 1026 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) { 1027 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst); 1028 } 1029 return Shadow; 1030 } 1031 1032 void DFSanVisitor::visitOperandShadowInst(Instruction &I) { 1033 Value *CombinedShadow = DFSF.combineOperandShadows(&I); 1034 DFSF.setShadow(&I, CombinedShadow); 1035 } 1036 1037 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where 1038 // Addr has alignment Align, and take the union of each of those shadows. 1039 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align, 1040 Instruction *Pos) { 1041 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) { 1042 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i = 1043 AllocaShadowMap.find(AI); 1044 if (i != AllocaShadowMap.end()) { 1045 IRBuilder<> IRB(Pos); 1046 return IRB.CreateLoad(i->second); 1047 } 1048 } 1049 1050 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8; 1051 SmallVector<Value *, 2> Objs; 1052 GetUnderlyingObjects(Addr, Objs, Pos->getModule()->getDataLayout()); 1053 bool AllConstants = true; 1054 for (SmallVector<Value *, 2>::iterator i = Objs.begin(), e = Objs.end(); 1055 i != e; ++i) { 1056 if (isa<Function>(*i) || isa<BlockAddress>(*i)) 1057 continue; 1058 if (isa<GlobalVariable>(*i) && cast<GlobalVariable>(*i)->isConstant()) 1059 continue; 1060 1061 AllConstants = false; 1062 break; 1063 } 1064 if (AllConstants) 1065 return DFS.ZeroShadow; 1066 1067 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos); 1068 switch (Size) { 1069 case 0: 1070 return DFS.ZeroShadow; 1071 case 1: { 1072 LoadInst *LI = new LoadInst(ShadowAddr, "", Pos); 1073 LI->setAlignment(ShadowAlign); 1074 return LI; 1075 } 1076 case 2: { 1077 IRBuilder<> IRB(Pos); 1078 Value *ShadowAddr1 = IRB.CreateGEP(DFS.ShadowTy, ShadowAddr, 1079 ConstantInt::get(DFS.IntptrTy, 1)); 1080 return combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign), 1081 IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign), Pos); 1082 } 1083 } 1084 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) { 1085 // Fast path for the common case where each byte has identical shadow: load 1086 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any 1087 // shadow is non-equal. 1088 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F); 1089 IRBuilder<> FallbackIRB(FallbackBB); 1090 CallInst *FallbackCall = FallbackIRB.CreateCall2( 1091 DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)); 1092 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1093 1094 // Compare each of the shadows stored in the loaded 64 bits to each other, 1095 // by computing (WideShadow rotl ShadowWidth) == WideShadow. 1096 IRBuilder<> IRB(Pos); 1097 Value *WideAddr = 1098 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx)); 1099 Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign); 1100 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy); 1101 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth); 1102 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth); 1103 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow); 1104 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow); 1105 1106 BasicBlock *Head = Pos->getParent(); 1107 BasicBlock *Tail = Head->splitBasicBlock(Pos); 1108 1109 if (DomTreeNode *OldNode = DT.getNode(Head)) { 1110 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end()); 1111 1112 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head); 1113 for (auto Child : Children) 1114 DT.changeImmediateDominator(Child, NewNode); 1115 } 1116 1117 // In the following code LastBr will refer to the previous basic block's 1118 // conditional branch instruction, whose true successor is fixed up to point 1119 // to the next block during the loop below or to the tail after the final 1120 // iteration. 1121 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq); 1122 ReplaceInstWithInst(Head->getTerminator(), LastBr); 1123 DT.addNewBlock(FallbackBB, Head); 1124 1125 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size; 1126 Ofs += 64 / DFS.ShadowWidth) { 1127 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F); 1128 DT.addNewBlock(NextBB, LastBr->getParent()); 1129 IRBuilder<> NextIRB(NextBB); 1130 WideAddr = NextIRB.CreateGEP(Type::getInt64Ty(*DFS.Ctx), WideAddr, 1131 ConstantInt::get(DFS.IntptrTy, 1)); 1132 Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign); 1133 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow); 1134 LastBr->setSuccessor(0, NextBB); 1135 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB); 1136 } 1137 1138 LastBr->setSuccessor(0, Tail); 1139 FallbackIRB.CreateBr(Tail); 1140 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front()); 1141 Shadow->addIncoming(FallbackCall, FallbackBB); 1142 Shadow->addIncoming(TruncShadow, LastBr->getParent()); 1143 return Shadow; 1144 } 1145 1146 IRBuilder<> IRB(Pos); 1147 CallInst *FallbackCall = IRB.CreateCall2( 1148 DFS.DFSanUnionLoadFn, ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)); 1149 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1150 return FallbackCall; 1151 } 1152 1153 void DFSanVisitor::visitLoadInst(LoadInst &LI) { 1154 auto &DL = LI.getModule()->getDataLayout(); 1155 uint64_t Size = DL.getTypeStoreSize(LI.getType()); 1156 if (Size == 0) { 1157 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow); 1158 return; 1159 } 1160 1161 uint64_t Align; 1162 if (ClPreserveAlignment) { 1163 Align = LI.getAlignment(); 1164 if (Align == 0) 1165 Align = DL.getABITypeAlignment(LI.getType()); 1166 } else { 1167 Align = 1; 1168 } 1169 IRBuilder<> IRB(&LI); 1170 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI); 1171 if (ClCombinePointerLabelsOnLoad) { 1172 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand()); 1173 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI); 1174 } 1175 if (Shadow != DFSF.DFS.ZeroShadow) 1176 DFSF.NonZeroChecks.push_back(Shadow); 1177 1178 DFSF.setShadow(&LI, Shadow); 1179 } 1180 1181 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align, 1182 Value *Shadow, Instruction *Pos) { 1183 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) { 1184 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i = 1185 AllocaShadowMap.find(AI); 1186 if (i != AllocaShadowMap.end()) { 1187 IRBuilder<> IRB(Pos); 1188 IRB.CreateStore(Shadow, i->second); 1189 return; 1190 } 1191 } 1192 1193 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8; 1194 IRBuilder<> IRB(Pos); 1195 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos); 1196 if (Shadow == DFS.ZeroShadow) { 1197 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth); 1198 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0); 1199 Value *ExtShadowAddr = 1200 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy)); 1201 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign); 1202 return; 1203 } 1204 1205 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth; 1206 uint64_t Offset = 0; 1207 if (Size >= ShadowVecSize) { 1208 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize); 1209 Value *ShadowVec = UndefValue::get(ShadowVecTy); 1210 for (unsigned i = 0; i != ShadowVecSize; ++i) { 1211 ShadowVec = IRB.CreateInsertElement( 1212 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i)); 1213 } 1214 Value *ShadowVecAddr = 1215 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy)); 1216 do { 1217 Value *CurShadowVecAddr = 1218 IRB.CreateConstGEP1_32(ShadowVecTy, ShadowVecAddr, Offset); 1219 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign); 1220 Size -= ShadowVecSize; 1221 ++Offset; 1222 } while (Size >= ShadowVecSize); 1223 Offset *= ShadowVecSize; 1224 } 1225 while (Size > 0) { 1226 Value *CurShadowAddr = 1227 IRB.CreateConstGEP1_32(DFS.ShadowTy, ShadowAddr, Offset); 1228 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign); 1229 --Size; 1230 ++Offset; 1231 } 1232 } 1233 1234 void DFSanVisitor::visitStoreInst(StoreInst &SI) { 1235 auto &DL = SI.getModule()->getDataLayout(); 1236 uint64_t Size = DL.getTypeStoreSize(SI.getValueOperand()->getType()); 1237 if (Size == 0) 1238 return; 1239 1240 uint64_t Align; 1241 if (ClPreserveAlignment) { 1242 Align = SI.getAlignment(); 1243 if (Align == 0) 1244 Align = DL.getABITypeAlignment(SI.getValueOperand()->getType()); 1245 } else { 1246 Align = 1; 1247 } 1248 1249 Value* Shadow = DFSF.getShadow(SI.getValueOperand()); 1250 if (ClCombinePointerLabelsOnStore) { 1251 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand()); 1252 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI); 1253 } 1254 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI); 1255 } 1256 1257 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) { 1258 visitOperandShadowInst(BO); 1259 } 1260 1261 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); } 1262 1263 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); } 1264 1265 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) { 1266 visitOperandShadowInst(GEPI); 1267 } 1268 1269 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) { 1270 visitOperandShadowInst(I); 1271 } 1272 1273 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) { 1274 visitOperandShadowInst(I); 1275 } 1276 1277 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) { 1278 visitOperandShadowInst(I); 1279 } 1280 1281 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) { 1282 visitOperandShadowInst(I); 1283 } 1284 1285 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) { 1286 visitOperandShadowInst(I); 1287 } 1288 1289 void DFSanVisitor::visitAllocaInst(AllocaInst &I) { 1290 bool AllLoadsStores = true; 1291 for (User *U : I.users()) { 1292 if (isa<LoadInst>(U)) 1293 continue; 1294 1295 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1296 if (SI->getPointerOperand() == &I) 1297 continue; 1298 } 1299 1300 AllLoadsStores = false; 1301 break; 1302 } 1303 if (AllLoadsStores) { 1304 IRBuilder<> IRB(&I); 1305 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy); 1306 } 1307 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow); 1308 } 1309 1310 void DFSanVisitor::visitSelectInst(SelectInst &I) { 1311 Value *CondShadow = DFSF.getShadow(I.getCondition()); 1312 Value *TrueShadow = DFSF.getShadow(I.getTrueValue()); 1313 Value *FalseShadow = DFSF.getShadow(I.getFalseValue()); 1314 1315 if (isa<VectorType>(I.getCondition()->getType())) { 1316 DFSF.setShadow( 1317 &I, 1318 DFSF.combineShadows( 1319 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I)); 1320 } else { 1321 Value *ShadowSel; 1322 if (TrueShadow == FalseShadow) { 1323 ShadowSel = TrueShadow; 1324 } else { 1325 ShadowSel = 1326 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I); 1327 } 1328 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I)); 1329 } 1330 } 1331 1332 void DFSanVisitor::visitMemSetInst(MemSetInst &I) { 1333 IRBuilder<> IRB(&I); 1334 Value *ValShadow = DFSF.getShadow(I.getValue()); 1335 IRB.CreateCall3( 1336 DFSF.DFS.DFSanSetLabelFn, ValShadow, 1337 IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(*DFSF.DFS.Ctx)), 1338 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)); 1339 } 1340 1341 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) { 1342 IRBuilder<> IRB(&I); 1343 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I); 1344 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I); 1345 Value *LenShadow = IRB.CreateMul( 1346 I.getLength(), 1347 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8)); 1348 Value *AlignShadow; 1349 if (ClPreserveAlignment) { 1350 AlignShadow = IRB.CreateMul(I.getAlignmentCst(), 1351 ConstantInt::get(I.getAlignmentCst()->getType(), 1352 DFSF.DFS.ShadowWidth / 8)); 1353 } else { 1354 AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(), 1355 DFSF.DFS.ShadowWidth / 8); 1356 } 1357 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx); 1358 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr); 1359 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr); 1360 IRB.CreateCall5(I.getCalledValue(), DestShadow, SrcShadow, LenShadow, 1361 AlignShadow, I.getVolatileCst()); 1362 } 1363 1364 void DFSanVisitor::visitReturnInst(ReturnInst &RI) { 1365 if (!DFSF.IsNativeABI && RI.getReturnValue()) { 1366 switch (DFSF.IA) { 1367 case DataFlowSanitizer::IA_TLS: { 1368 Value *S = DFSF.getShadow(RI.getReturnValue()); 1369 IRBuilder<> IRB(&RI); 1370 IRB.CreateStore(S, DFSF.getRetvalTLS()); 1371 break; 1372 } 1373 case DataFlowSanitizer::IA_Args: { 1374 IRBuilder<> IRB(&RI); 1375 Type *RT = DFSF.F->getFunctionType()->getReturnType(); 1376 Value *InsVal = 1377 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0); 1378 Value *InsShadow = 1379 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1); 1380 RI.setOperand(0, InsShadow); 1381 break; 1382 } 1383 } 1384 } 1385 } 1386 1387 void DFSanVisitor::visitCallSite(CallSite CS) { 1388 Function *F = CS.getCalledFunction(); 1389 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) { 1390 visitOperandShadowInst(*CS.getInstruction()); 1391 return; 1392 } 1393 1394 // Calls to this function are synthesized in wrappers, and we shouldn't 1395 // instrument them. 1396 if (F == DFSF.DFS.DFSanVarargWrapperFn) 1397 return; 1398 1399 assert(!(cast<FunctionType>( 1400 CS.getCalledValue()->getType()->getPointerElementType())->isVarArg() && 1401 dyn_cast<InvokeInst>(CS.getInstruction()))); 1402 1403 IRBuilder<> IRB(CS.getInstruction()); 1404 1405 DenseMap<Value *, Function *>::iterator i = 1406 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue()); 1407 if (i != DFSF.DFS.UnwrappedFnMap.end()) { 1408 Function *F = i->second; 1409 switch (DFSF.DFS.getWrapperKind(F)) { 1410 case DataFlowSanitizer::WK_Warning: { 1411 CS.setCalledFunction(F); 1412 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn, 1413 IRB.CreateGlobalStringPtr(F->getName())); 1414 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow); 1415 return; 1416 } 1417 case DataFlowSanitizer::WK_Discard: { 1418 CS.setCalledFunction(F); 1419 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow); 1420 return; 1421 } 1422 case DataFlowSanitizer::WK_Functional: { 1423 CS.setCalledFunction(F); 1424 visitOperandShadowInst(*CS.getInstruction()); 1425 return; 1426 } 1427 case DataFlowSanitizer::WK_Custom: { 1428 // Don't try to handle invokes of custom functions, it's too complicated. 1429 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_ 1430 // wrapper. 1431 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 1432 FunctionType *FT = F->getFunctionType(); 1433 FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT); 1434 std::string CustomFName = "__dfsw_"; 1435 CustomFName += F->getName(); 1436 Constant *CustomF = 1437 DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT); 1438 if (Function *CustomFn = dyn_cast<Function>(CustomF)) { 1439 CustomFn->copyAttributesFrom(F); 1440 1441 // Custom functions returning non-void will write to the return label. 1442 if (!FT->getReturnType()->isVoidTy()) { 1443 CustomFn->removeAttributes(AttributeSet::FunctionIndex, 1444 DFSF.DFS.ReadOnlyNoneAttrs); 1445 } 1446 } 1447 1448 std::vector<Value *> Args; 1449 1450 CallSite::arg_iterator i = CS.arg_begin(); 1451 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) { 1452 Type *T = (*i)->getType(); 1453 FunctionType *ParamFT; 1454 if (isa<PointerType>(T) && 1455 (ParamFT = dyn_cast<FunctionType>( 1456 cast<PointerType>(T)->getElementType()))) { 1457 std::string TName = "dfst"; 1458 TName += utostr(FT->getNumParams() - n); 1459 TName += "$"; 1460 TName += F->getName(); 1461 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName); 1462 Args.push_back(T); 1463 Args.push_back( 1464 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx))); 1465 } else { 1466 Args.push_back(*i); 1467 } 1468 } 1469 1470 i = CS.arg_begin(); 1471 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1472 Args.push_back(DFSF.getShadow(*i)); 1473 1474 if (FT->isVarArg()) { 1475 auto *LabelVATy = ArrayType::get(DFSF.DFS.ShadowTy, 1476 CS.arg_size() - FT->getNumParams()); 1477 auto *LabelVAAlloca = new AllocaInst(LabelVATy, "labelva", 1478 DFSF.F->getEntryBlock().begin()); 1479 1480 for (unsigned n = 0; i != CS.arg_end(); ++i, ++n) { 1481 auto LabelVAPtr = IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, n); 1482 IRB.CreateStore(DFSF.getShadow(*i), LabelVAPtr); 1483 } 1484 1485 Args.push_back(IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, 0)); 1486 } 1487 1488 if (!FT->getReturnType()->isVoidTy()) { 1489 if (!DFSF.LabelReturnAlloca) { 1490 DFSF.LabelReturnAlloca = 1491 new AllocaInst(DFSF.DFS.ShadowTy, "labelreturn", 1492 DFSF.F->getEntryBlock().begin()); 1493 } 1494 Args.push_back(DFSF.LabelReturnAlloca); 1495 } 1496 1497 for (i = CS.arg_begin() + FT->getNumParams(); i != CS.arg_end(); ++i) 1498 Args.push_back(*i); 1499 1500 CallInst *CustomCI = IRB.CreateCall(CustomF, Args); 1501 CustomCI->setCallingConv(CI->getCallingConv()); 1502 CustomCI->setAttributes(CI->getAttributes()); 1503 1504 if (!FT->getReturnType()->isVoidTy()) { 1505 LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca); 1506 DFSF.setShadow(CustomCI, LabelLoad); 1507 } 1508 1509 CI->replaceAllUsesWith(CustomCI); 1510 CI->eraseFromParent(); 1511 return; 1512 } 1513 break; 1514 } 1515 } 1516 } 1517 1518 FunctionType *FT = cast<FunctionType>( 1519 CS.getCalledValue()->getType()->getPointerElementType()); 1520 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) { 1521 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) { 1522 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)), 1523 DFSF.getArgTLS(i, CS.getInstruction())); 1524 } 1525 } 1526 1527 Instruction *Next = nullptr; 1528 if (!CS.getType()->isVoidTy()) { 1529 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1530 if (II->getNormalDest()->getSinglePredecessor()) { 1531 Next = II->getNormalDest()->begin(); 1532 } else { 1533 BasicBlock *NewBB = 1534 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT); 1535 Next = NewBB->begin(); 1536 } 1537 } else { 1538 Next = CS->getNextNode(); 1539 } 1540 1541 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) { 1542 IRBuilder<> NextIRB(Next); 1543 LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS()); 1544 DFSF.SkipInsts.insert(LI); 1545 DFSF.setShadow(CS.getInstruction(), LI); 1546 DFSF.NonZeroChecks.push_back(LI); 1547 } 1548 } 1549 1550 // Do all instrumentation for IA_Args down here to defer tampering with the 1551 // CFG in a way that SplitEdge may be able to detect. 1552 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) { 1553 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT); 1554 Value *Func = 1555 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT)); 1556 std::vector<Value *> Args; 1557 1558 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 1559 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1560 Args.push_back(*i); 1561 1562 i = CS.arg_begin(); 1563 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1564 Args.push_back(DFSF.getShadow(*i)); 1565 1566 if (FT->isVarArg()) { 1567 unsigned VarArgSize = CS.arg_size() - FT->getNumParams(); 1568 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize); 1569 AllocaInst *VarArgShadow = 1570 new AllocaInst(VarArgArrayTy, "", DFSF.F->getEntryBlock().begin()); 1571 Args.push_back(IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, 0)); 1572 for (unsigned n = 0; i != e; ++i, ++n) { 1573 IRB.CreateStore( 1574 DFSF.getShadow(*i), 1575 IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, n)); 1576 Args.push_back(*i); 1577 } 1578 } 1579 1580 CallSite NewCS; 1581 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1582 NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(), 1583 Args); 1584 } else { 1585 NewCS = IRB.CreateCall(Func, Args); 1586 } 1587 NewCS.setCallingConv(CS.getCallingConv()); 1588 NewCS.setAttributes(CS.getAttributes().removeAttributes( 1589 *DFSF.DFS.Ctx, AttributeSet::ReturnIndex, 1590 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType(), 1591 AttributeSet::ReturnIndex))); 1592 1593 if (Next) { 1594 ExtractValueInst *ExVal = 1595 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next); 1596 DFSF.SkipInsts.insert(ExVal); 1597 ExtractValueInst *ExShadow = 1598 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next); 1599 DFSF.SkipInsts.insert(ExShadow); 1600 DFSF.setShadow(ExVal, ExShadow); 1601 DFSF.NonZeroChecks.push_back(ExShadow); 1602 1603 CS.getInstruction()->replaceAllUsesWith(ExVal); 1604 } 1605 1606 CS.getInstruction()->eraseFromParent(); 1607 } 1608 } 1609 1610 void DFSanVisitor::visitPHINode(PHINode &PN) { 1611 PHINode *ShadowPN = 1612 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN); 1613 1614 // Give the shadow phi node valid predecessors to fool SplitEdge into working. 1615 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy); 1616 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e; 1617 ++i) { 1618 ShadowPN->addIncoming(UndefShadow, *i); 1619 } 1620 1621 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN)); 1622 DFSF.setShadow(&PN, ShadowPN); 1623 } 1624