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 // 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"; 79 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, 89 cl::init(false)); 90 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( 99 "dfsan-abilist", 100 cl::desc("File listing native ABI functions and how the pass treats them"), 101 cl::Hidden); 102 103 // 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( 106 "dfsan-args-abi", 107 cl::desc("Use the argument ABI rather than the TLS ABI"), 108 cl::Hidden); 109 110 // Controls whether the pass includes or ignores the labels of pointers in load 111 // instructions. 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)); 117 118 // 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)); 125 126 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"), 130 cl::Hidden); 131 132 133 namespace { 134 135 StringRef GetGlobalTypeString(const GlobalValue &G) { 136 // Types of GlobalVariables are always pointer types. 137 Type *GType = G.getType()->getElementType(); 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(); 142 } 143 return "<unknown type>"; 144 } 145 146 class DFSanABIList { 147 std::unique_ptr<SpecialCaseList> SCL; 148 149 public: 150 DFSanABIList() {} 151 152 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); } 153 154 /// Returns whether either this function or its source file are listed in the 155 /// given category. 156 bool isIn(const Function &F, StringRef Category) const { 157 return isIn(*F.getParent(), Category) || 158 SCL->inSection("fun", F.getName(), Category); 159 } 160 161 /// Returns whether this global alias is listed in the given category. 162 /// 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)) 167 return true; 168 169 if (isa<FunctionType>(GA.getType()->getElementType())) 170 return SCL->inSection("fun", GA.getName(), Category); 171 172 return SCL->inSection("global", GA.getName(), Category) || 173 SCL->inSection("type", GetGlobalTypeString(GA), Category); 174 } 175 176 /// 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); 179 } 180 }; 181 182 class DataFlowSanitizer : public ModulePass { 183 friend struct DFSanFunction; 184 friend class DFSanVisitor; 185 186 enum { 187 ShadowWidth = 16 188 }; 189 190 /// 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. 194 IA_Args, 195 196 /// Argument and return value labels are passed through TLS variables 197 /// __dfsan_arg_tls and __dfsan_retval_tls. 198 IA_TLS 199 }; 200 201 /// How should calls to uninstrumented functions be handled? 202 enum WrapperKind { 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. 206 WK_Warning, 207 208 /// This function does not write to (user-accessible) memory, and its return 209 /// value is unlabelled. 210 WK_Discard, 211 212 /// 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. 214 WK_Functional, 215 216 /// 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. 223 WK_Custom 224 }; 225 226 Module *Mod; 227 LLVMContext *Ctx; 228 IntegerType *ShadowTy; 229 PointerType *ShadowPtrTy; 230 IntegerType *IntptrTy; 231 ConstantInt *ZeroShadow; 232 ConstantInt *ShadowPtrMask; 233 ConstantInt *ShadowPtrMul; 234 Constant *ArgTLS; 235 Constant *RetvalTLS; 236 void *(*GetArgTLSPtr)(); 237 void *(*GetRetvalTLSPtr)(); 238 Constant *GetArgTLS; 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 AttributeSet ReadOnlyNoneAttrs; 258 bool DFSanRuntimeShadowMask; 259 260 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); 273 274 public: 275 DataFlowSanitizer( 276 const std::vector<std::string> &ABIListFiles = std::vector<std::string>(), 277 void *(*getArgTLS)() = nullptr, void *(*getRetValTLS)() = nullptr); 278 static char ID; 279 bool doInitialization(Module &M) override; 280 bool runOnModule(Module &M) override; 281 }; 282 283 struct DFSanFunction { 284 DataFlowSanitizer &DFS; 285 Function *F; 286 DominatorTree DT; 287 DataFlowSanitizer::InstrumentedABI IA; 288 bool IsNativeABI; 289 Value *ArgTLSPtr; 290 Value *RetvalTLSPtr; 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; 297 bool AvoidNewBlocks; 298 299 struct CachedCombinedShadow { 300 BasicBlock *Block; 301 Value *Shadow; 302 }; 303 DenseMap<std::pair<Value *, Value *>, CachedCombinedShadow> 304 CachedCombinedShadows; 305 DenseMap<Value *, std::set<Value *>> ShadowElements; 306 307 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) { 311 DT.recalculate(*F); 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; 315 } 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, 324 Instruction *Pos); 325 void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow, 326 Instruction *Pos); 327 }; 328 329 class DFSanVisitor : public InstVisitor<DFSanVisitor> { 330 public: 331 DFSanFunction &DFSF; 332 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {} 333 334 void visitOperandShadowInst(Instruction &I); 335 336 void visitBinaryOperator(BinaryOperator &BO); 337 void visitCastInst(CastInst &CI); 338 void visitCmpInst(CmpInst &CI); 339 void visitGetElementPtrInst(GetElementPtrInst &GEPI); 340 void visitLoadInst(LoadInst &LI); 341 void visitStoreInst(StoreInst &SI); 342 void visitReturnInst(ReturnInst &RI); 343 void visitCallSite(CallSite CS); 344 void visitPHINode(PHINode &PN); 345 void visitExtractElementInst(ExtractElementInst &I); 346 void visitInsertElementInst(InsertElementInst &I); 347 void visitShuffleVectorInst(ShuffleVectorInst &I); 348 void visitExtractValueInst(ExtractValueInst &I); 349 void visitInsertValueInst(InsertValueInst &I); 350 void visitAllocaInst(AllocaInst &I); 351 void visitSelectInst(SelectInst &I); 352 void visitMemSetInst(MemSetInst &I); 353 void visitMemTransferInst(MemTransferInst &I); 354 }; 355 356 } 357 358 char DataFlowSanitizer::ID; 359 INITIALIZE_PASS(DataFlowSanitizer, "dfsan", 360 "DataFlowSanitizer: dynamic data flow analysis.", false, false) 361 362 ModulePass * 363 llvm::createDataFlowSanitizerPass(const std::vector<std::string> &ABIListFiles, 364 void *(*getArgTLS)(), 365 void *(*getRetValTLS)()) { 366 return new DataFlowSanitizer(ABIListFiles, getArgTLS, getRetValTLS); 367 } 368 369 DataFlowSanitizer::DataFlowSanitizer( 370 const std::vector<std::string> &ABIListFiles, void *(*getArgTLS)(), 371 void *(*getRetValTLS)()) 372 : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS), 373 DFSanRuntimeShadowMask(false) { 374 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles)); 375 AllABIListFiles.insert(AllABIListFiles.end(), ClABIListFiles.begin(), 376 ClABIListFiles.end()); 377 ABIList.set(SpecialCaseList::createOrDie(AllABIListFiles)); 378 } 379 380 FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) { 381 llvm::SmallVector<Type *, 4> ArgTypes(T->param_begin(), T->param_end()); 382 ArgTypes.append(T->getNumParams(), ShadowTy); 383 if (T->isVarArg()) 384 ArgTypes.push_back(ShadowPtrTy); 385 Type *RetType = T->getReturnType(); 386 if (!RetType->isVoidTy()) 387 RetType = StructType::get(RetType, ShadowTy, (Type *)nullptr); 388 return FunctionType::get(RetType, ArgTypes, T->isVarArg()); 389 } 390 391 FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) { 392 assert(!T->isVarArg()); 393 llvm::SmallVector<Type *, 4> ArgTypes; 394 ArgTypes.push_back(T->getPointerTo()); 395 ArgTypes.append(T->param_begin(), T->param_end()); 396 ArgTypes.append(T->getNumParams(), ShadowTy); 397 Type *RetType = T->getReturnType(); 398 if (!RetType->isVoidTy()) 399 ArgTypes.push_back(ShadowPtrTy); 400 return FunctionType::get(T->getReturnType(), ArgTypes, false); 401 } 402 403 FunctionType *DataFlowSanitizer::getCustomFunctionType(FunctionType *T) { 404 llvm::SmallVector<Type *, 4> ArgTypes; 405 for (FunctionType::param_iterator i = T->param_begin(), e = T->param_end(); 406 i != e; ++i) { 407 FunctionType *FT; 408 if (isa<PointerType>(*i) && (FT = dyn_cast<FunctionType>(cast<PointerType>( 409 *i)->getElementType()))) { 410 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo()); 411 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx)); 412 } else { 413 ArgTypes.push_back(*i); 414 } 415 } 416 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i) 417 ArgTypes.push_back(ShadowTy); 418 if (T->isVarArg()) 419 ArgTypes.push_back(ShadowPtrTy); 420 Type *RetType = T->getReturnType(); 421 if (!RetType->isVoidTy()) 422 ArgTypes.push_back(ShadowPtrTy); 423 return FunctionType::get(T->getReturnType(), ArgTypes, T->isVarArg()); 424 } 425 426 bool DataFlowSanitizer::doInitialization(Module &M) { 427 llvm::Triple TargetTriple(M.getTargetTriple()); 428 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64; 429 bool IsMIPS64 = TargetTriple.getArch() == llvm::Triple::mips64 || 430 TargetTriple.getArch() == llvm::Triple::mips64el; 431 bool IsAArch64 = TargetTriple.getArch() == llvm::Triple::aarch64 || 432 TargetTriple.getArch() == llvm::Triple::aarch64_be; 433 434 const DataLayout &DL = M.getDataLayout(); 435 436 Mod = &M; 437 Ctx = &M.getContext(); 438 ShadowTy = IntegerType::get(*Ctx, ShadowWidth); 439 ShadowPtrTy = PointerType::getUnqual(ShadowTy); 440 IntptrTy = DL.getIntPtrType(*Ctx); 441 ZeroShadow = ConstantInt::getSigned(ShadowTy, 0); 442 ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8); 443 if (IsX86_64) 444 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL); 445 else if (IsMIPS64) 446 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0xF000000000LL); 447 // AArch64 supports multiple VMAs and the shadow mask is set at runtime. 448 else if (IsAArch64) 449 DFSanRuntimeShadowMask = true; 450 else 451 report_fatal_error("unsupported triple"); 452 453 Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy }; 454 DFSanUnionFnTy = 455 FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false); 456 Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy }; 457 DFSanUnionLoadFnTy = 458 FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false); 459 DFSanUnimplementedFnTy = FunctionType::get( 460 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false); 461 Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy }; 462 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx), 463 DFSanSetLabelArgs, /*isVarArg=*/false); 464 DFSanNonzeroLabelFnTy = FunctionType::get( 465 Type::getVoidTy(*Ctx), None, /*isVarArg=*/false); 466 DFSanVarargWrapperFnTy = FunctionType::get( 467 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false); 468 469 if (GetArgTLSPtr) { 470 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64); 471 ArgTLS = nullptr; 472 GetArgTLS = ConstantExpr::getIntToPtr( 473 ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)), 474 PointerType::getUnqual( 475 FunctionType::get(PointerType::getUnqual(ArgTLSTy), 476 (Type *)nullptr))); 477 } 478 if (GetRetvalTLSPtr) { 479 RetvalTLS = nullptr; 480 GetRetvalTLS = ConstantExpr::getIntToPtr( 481 ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)), 482 PointerType::getUnqual( 483 FunctionType::get(PointerType::getUnqual(ShadowTy), 484 (Type *)nullptr))); 485 } 486 487 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000); 488 return true; 489 } 490 491 bool DataFlowSanitizer::isInstrumented(const Function *F) { 492 return !ABIList.isIn(*F, "uninstrumented"); 493 } 494 495 bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) { 496 return !ABIList.isIn(*GA, "uninstrumented"); 497 } 498 499 DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() { 500 return ClArgsABI ? IA_Args : IA_TLS; 501 } 502 503 DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) { 504 if (ABIList.isIn(*F, "functional")) 505 return WK_Functional; 506 if (ABIList.isIn(*F, "discard")) 507 return WK_Discard; 508 if (ABIList.isIn(*F, "custom")) 509 return WK_Custom; 510 511 return WK_Warning; 512 } 513 514 void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) { 515 std::string GVName = GV->getName(), Prefix = "dfs$"; 516 GV->setName(Prefix + GVName); 517 518 // Try to change the name of the function in module inline asm. We only do 519 // this for specific asm directives, currently only ".symver", to try to avoid 520 // corrupting asm which happens to contain the symbol name as a substring. 521 // Note that the substitution for .symver assumes that the versioned symbol 522 // also has an instrumented name. 523 std::string Asm = GV->getParent()->getModuleInlineAsm(); 524 std::string SearchStr = ".symver " + GVName + ","; 525 size_t Pos = Asm.find(SearchStr); 526 if (Pos != std::string::npos) { 527 Asm.replace(Pos, SearchStr.size(), 528 ".symver " + Prefix + GVName + "," + Prefix); 529 GV->getParent()->setModuleInlineAsm(Asm); 530 } 531 } 532 533 Function * 534 DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName, 535 GlobalValue::LinkageTypes NewFLink, 536 FunctionType *NewFT) { 537 FunctionType *FT = F->getFunctionType(); 538 Function *NewF = Function::Create(NewFT, NewFLink, NewFName, 539 F->getParent()); 540 NewF->copyAttributesFrom(F); 541 NewF->removeAttributes( 542 AttributeSet::ReturnIndex, 543 AttributeSet::get(F->getContext(), AttributeSet::ReturnIndex, 544 AttributeFuncs::typeIncompatible(NewFT->getReturnType()))); 545 546 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF); 547 if (F->isVarArg()) { 548 NewF->removeAttributes( 549 AttributeSet::FunctionIndex, 550 AttributeSet().addAttribute(*Ctx, AttributeSet::FunctionIndex, 551 "split-stack")); 552 CallInst::Create(DFSanVarargWrapperFn, 553 IRBuilder<>(BB).CreateGlobalStringPtr(F->getName()), "", 554 BB); 555 new UnreachableInst(*Ctx, BB); 556 } else { 557 std::vector<Value *> Args; 558 unsigned n = FT->getNumParams(); 559 for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n) 560 Args.push_back(&*ai); 561 CallInst *CI = CallInst::Create(F, Args, "", BB); 562 if (FT->getReturnType()->isVoidTy()) 563 ReturnInst::Create(*Ctx, BB); 564 else 565 ReturnInst::Create(*Ctx, CI, BB); 566 } 567 568 return NewF; 569 } 570 571 Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT, 572 StringRef FName) { 573 FunctionType *FTT = getTrampolineFunctionType(FT); 574 Constant *C = Mod->getOrInsertFunction(FName, FTT); 575 Function *F = dyn_cast<Function>(C); 576 if (F && F->isDeclaration()) { 577 F->setLinkage(GlobalValue::LinkOnceODRLinkage); 578 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F); 579 std::vector<Value *> Args; 580 Function::arg_iterator AI = F->arg_begin(); ++AI; 581 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N) 582 Args.push_back(&*AI); 583 CallInst *CI = 584 CallInst::Create(&F->getArgumentList().front(), Args, "", BB); 585 ReturnInst *RI; 586 if (FT->getReturnType()->isVoidTy()) 587 RI = ReturnInst::Create(*Ctx, BB); 588 else 589 RI = ReturnInst::Create(*Ctx, CI, BB); 590 591 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true); 592 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI; 593 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N) 594 DFSF.ValShadowMap[&*ValAI] = &*ShadowAI; 595 DFSanVisitor(DFSF).visitCallInst(*CI); 596 if (!FT->getReturnType()->isVoidTy()) 597 new StoreInst(DFSF.getShadow(RI->getReturnValue()), 598 &F->getArgumentList().back(), RI); 599 } 600 601 return C; 602 } 603 604 bool DataFlowSanitizer::runOnModule(Module &M) { 605 if (ABIList.isIn(M, "skip")) 606 return false; 607 608 if (!GetArgTLSPtr) { 609 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64); 610 ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy); 611 if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS)) 612 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel); 613 } 614 if (!GetRetvalTLSPtr) { 615 RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy); 616 if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS)) 617 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel); 618 } 619 620 ExternalShadowMask = 621 Mod->getOrInsertGlobal(kDFSanExternShadowPtrMask, IntptrTy); 622 623 DFSanUnionFn = Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy); 624 if (Function *F = dyn_cast<Function>(DFSanUnionFn)) { 625 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind); 626 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone); 627 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 628 F->addAttribute(1, Attribute::ZExt); 629 F->addAttribute(2, Attribute::ZExt); 630 } 631 DFSanCheckedUnionFn = Mod->getOrInsertFunction("dfsan_union", DFSanUnionFnTy); 632 if (Function *F = dyn_cast<Function>(DFSanCheckedUnionFn)) { 633 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind); 634 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone); 635 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 636 F->addAttribute(1, Attribute::ZExt); 637 F->addAttribute(2, Attribute::ZExt); 638 } 639 DFSanUnionLoadFn = 640 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy); 641 if (Function *F = dyn_cast<Function>(DFSanUnionLoadFn)) { 642 F->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind); 643 F->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadOnly); 644 F->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 645 } 646 DFSanUnimplementedFn = 647 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy); 648 DFSanSetLabelFn = 649 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy); 650 if (Function *F = dyn_cast<Function>(DFSanSetLabelFn)) { 651 F->addAttribute(1, Attribute::ZExt); 652 } 653 DFSanNonzeroLabelFn = 654 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy); 655 DFSanVarargWrapperFn = Mod->getOrInsertFunction("__dfsan_vararg_wrapper", 656 DFSanVarargWrapperFnTy); 657 658 std::vector<Function *> FnsToInstrument; 659 llvm::SmallPtrSet<Function *, 2> FnsWithNativeABI; 660 for (Function &i : M) { 661 if (!i.isIntrinsic() && 662 &i != DFSanUnionFn && 663 &i != DFSanCheckedUnionFn && 664 &i != DFSanUnionLoadFn && 665 &i != DFSanUnimplementedFn && 666 &i != DFSanSetLabelFn && 667 &i != DFSanNonzeroLabelFn && 668 &i != DFSanVarargWrapperFn) 669 FnsToInstrument.push_back(&i); 670 } 671 672 // Give function aliases prefixes when necessary, and build wrappers where the 673 // instrumentedness is inconsistent. 674 for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) { 675 GlobalAlias *GA = &*i; 676 ++i; 677 // Don't stop on weak. We assume people aren't playing games with the 678 // instrumentedness of overridden weak aliases. 679 if (auto F = dyn_cast<Function>(GA->getBaseObject())) { 680 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F); 681 if (GAInst && FInst) { 682 addGlobalNamePrefix(GA); 683 } else if (GAInst != FInst) { 684 // Non-instrumented alias of an instrumented function, or vice versa. 685 // Replace the alias with a native-ABI wrapper of the aliasee. The pass 686 // below will take care of instrumenting it. 687 Function *NewF = 688 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType()); 689 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType())); 690 NewF->takeName(GA); 691 GA->eraseFromParent(); 692 FnsToInstrument.push_back(NewF); 693 } 694 } 695 } 696 697 AttrBuilder B; 698 B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone); 699 ReadOnlyNoneAttrs = AttributeSet::get(*Ctx, AttributeSet::FunctionIndex, B); 700 701 // First, change the ABI of every function in the module. ABI-listed 702 // functions keep their original ABI and get a wrapper function. 703 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(), 704 e = FnsToInstrument.end(); 705 i != e; ++i) { 706 Function &F = **i; 707 FunctionType *FT = F.getFunctionType(); 708 709 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() && 710 FT->getReturnType()->isVoidTy()); 711 712 if (isInstrumented(&F)) { 713 // Instrumented functions get a 'dfs$' prefix. This allows us to more 714 // easily identify cases of mismatching ABIs. 715 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) { 716 FunctionType *NewFT = getArgsFunctionType(FT); 717 Function *NewF = Function::Create(NewFT, F.getLinkage(), "", &M); 718 NewF->copyAttributesFrom(&F); 719 NewF->removeAttributes( 720 AttributeSet::ReturnIndex, 721 AttributeSet::get(NewF->getContext(), AttributeSet::ReturnIndex, 722 AttributeFuncs::typeIncompatible(NewFT->getReturnType()))); 723 for (Function::arg_iterator FArg = F.arg_begin(), 724 NewFArg = NewF->arg_begin(), 725 FArgEnd = F.arg_end(); 726 FArg != FArgEnd; ++FArg, ++NewFArg) { 727 FArg->replaceAllUsesWith(&*NewFArg); 728 } 729 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList()); 730 731 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end(); 732 UI != UE;) { 733 BlockAddress *BA = dyn_cast<BlockAddress>(*UI); 734 ++UI; 735 if (BA) { 736 BA->replaceAllUsesWith( 737 BlockAddress::get(NewF, BA->getBasicBlock())); 738 delete BA; 739 } 740 } 741 F.replaceAllUsesWith( 742 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT))); 743 NewF->takeName(&F); 744 F.eraseFromParent(); 745 *i = NewF; 746 addGlobalNamePrefix(NewF); 747 } else { 748 addGlobalNamePrefix(&F); 749 } 750 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) { 751 // Build a wrapper function for F. The wrapper simply calls F, and is 752 // added to FnsToInstrument so that any instrumentation according to its 753 // WrapperKind is done in the second pass below. 754 FunctionType *NewFT = getInstrumentedABI() == IA_Args 755 ? getArgsFunctionType(FT) 756 : FT; 757 Function *NewF = buildWrapperFunction( 758 &F, std::string("dfsw$") + std::string(F.getName()), 759 GlobalValue::LinkOnceODRLinkage, NewFT); 760 if (getInstrumentedABI() == IA_TLS) 761 NewF->removeAttributes(AttributeSet::FunctionIndex, ReadOnlyNoneAttrs); 762 763 Value *WrappedFnCst = 764 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)); 765 F.replaceAllUsesWith(WrappedFnCst); 766 767 UnwrappedFnMap[WrappedFnCst] = &F; 768 *i = NewF; 769 770 if (!F.isDeclaration()) { 771 // This function is probably defining an interposition of an 772 // uninstrumented function and hence needs to keep the original ABI. 773 // But any functions it may call need to use the instrumented ABI, so 774 // we instrument it in a mode which preserves the original ABI. 775 FnsWithNativeABI.insert(&F); 776 777 // This code needs to rebuild the iterators, as they may be invalidated 778 // by the push_back, taking care that the new range does not include 779 // any functions added by this code. 780 size_t N = i - FnsToInstrument.begin(), 781 Count = e - FnsToInstrument.begin(); 782 FnsToInstrument.push_back(&F); 783 i = FnsToInstrument.begin() + N; 784 e = FnsToInstrument.begin() + Count; 785 } 786 // Hopefully, nobody will try to indirectly call a vararg 787 // function... yet. 788 } else if (FT->isVarArg()) { 789 UnwrappedFnMap[&F] = &F; 790 *i = nullptr; 791 } 792 } 793 794 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(), 795 e = FnsToInstrument.end(); 796 i != e; ++i) { 797 if (!*i || (*i)->isDeclaration()) 798 continue; 799 800 removeUnreachableBlocks(**i); 801 802 DFSanFunction DFSF(*this, *i, FnsWithNativeABI.count(*i)); 803 804 // DFSanVisitor may create new basic blocks, which confuses df_iterator. 805 // Build a copy of the list before iterating over it. 806 llvm::SmallVector<BasicBlock *, 4> BBList( 807 depth_first(&(*i)->getEntryBlock())); 808 809 for (llvm::SmallVector<BasicBlock *, 4>::iterator i = BBList.begin(), 810 e = BBList.end(); 811 i != e; ++i) { 812 Instruction *Inst = &(*i)->front(); 813 while (1) { 814 // DFSanVisitor may split the current basic block, changing the current 815 // instruction's next pointer and moving the next instruction to the 816 // tail block from which we should continue. 817 Instruction *Next = Inst->getNextNode(); 818 // DFSanVisitor may delete Inst, so keep track of whether it was a 819 // terminator. 820 bool IsTerminator = isa<TerminatorInst>(Inst); 821 if (!DFSF.SkipInsts.count(Inst)) 822 DFSanVisitor(DFSF).visit(Inst); 823 if (IsTerminator) 824 break; 825 Inst = Next; 826 } 827 } 828 829 // We will not necessarily be able to compute the shadow for every phi node 830 // until we have visited every block. Therefore, the code that handles phi 831 // nodes adds them to the PHIFixups list so that they can be properly 832 // handled here. 833 for (std::vector<std::pair<PHINode *, PHINode *> >::iterator 834 i = DFSF.PHIFixups.begin(), 835 e = DFSF.PHIFixups.end(); 836 i != e; ++i) { 837 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n; 838 ++val) { 839 i->second->setIncomingValue( 840 val, DFSF.getShadow(i->first->getIncomingValue(val))); 841 } 842 } 843 844 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy 845 // places (i.e. instructions in basic blocks we haven't even begun visiting 846 // yet). To make our life easier, do this work in a pass after the main 847 // instrumentation. 848 if (ClDebugNonzeroLabels) { 849 for (Value *V : DFSF.NonZeroChecks) { 850 Instruction *Pos; 851 if (Instruction *I = dyn_cast<Instruction>(V)) 852 Pos = I->getNextNode(); 853 else 854 Pos = &DFSF.F->getEntryBlock().front(); 855 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos)) 856 Pos = Pos->getNextNode(); 857 IRBuilder<> IRB(Pos); 858 Value *Ne = IRB.CreateICmpNE(V, DFSF.DFS.ZeroShadow); 859 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen( 860 Ne, Pos, /*Unreachable=*/false, ColdCallWeights)); 861 IRBuilder<> ThenIRB(BI); 862 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn, {}); 863 } 864 } 865 } 866 867 return false; 868 } 869 870 Value *DFSanFunction::getArgTLSPtr() { 871 if (ArgTLSPtr) 872 return ArgTLSPtr; 873 if (DFS.ArgTLS) 874 return ArgTLSPtr = DFS.ArgTLS; 875 876 IRBuilder<> IRB(&F->getEntryBlock().front()); 877 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS, {}); 878 } 879 880 Value *DFSanFunction::getRetvalTLS() { 881 if (RetvalTLSPtr) 882 return RetvalTLSPtr; 883 if (DFS.RetvalTLS) 884 return RetvalTLSPtr = DFS.RetvalTLS; 885 886 IRBuilder<> IRB(&F->getEntryBlock().front()); 887 return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS, {}); 888 } 889 890 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) { 891 IRBuilder<> IRB(Pos); 892 return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx); 893 } 894 895 Value *DFSanFunction::getShadow(Value *V) { 896 if (!isa<Argument>(V) && !isa<Instruction>(V)) 897 return DFS.ZeroShadow; 898 Value *&Shadow = ValShadowMap[V]; 899 if (!Shadow) { 900 if (Argument *A = dyn_cast<Argument>(V)) { 901 if (IsNativeABI) 902 return DFS.ZeroShadow; 903 switch (IA) { 904 case DataFlowSanitizer::IA_TLS: { 905 Value *ArgTLSPtr = getArgTLSPtr(); 906 Instruction *ArgTLSPos = 907 DFS.ArgTLS ? &*F->getEntryBlock().begin() 908 : cast<Instruction>(ArgTLSPtr)->getNextNode(); 909 IRBuilder<> IRB(ArgTLSPos); 910 Shadow = IRB.CreateLoad(getArgTLS(A->getArgNo(), ArgTLSPos)); 911 break; 912 } 913 case DataFlowSanitizer::IA_Args: { 914 unsigned ArgIdx = A->getArgNo() + F->getArgumentList().size() / 2; 915 Function::arg_iterator i = F->arg_begin(); 916 while (ArgIdx--) 917 ++i; 918 Shadow = &*i; 919 assert(Shadow->getType() == DFS.ShadowTy); 920 break; 921 } 922 } 923 NonZeroChecks.push_back(Shadow); 924 } else { 925 Shadow = DFS.ZeroShadow; 926 } 927 } 928 return Shadow; 929 } 930 931 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) { 932 assert(!ValShadowMap.count(I)); 933 assert(Shadow->getType() == DFS.ShadowTy); 934 ValShadowMap[I] = Shadow; 935 } 936 937 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) { 938 assert(Addr != RetvalTLS && "Reinstrumenting?"); 939 IRBuilder<> IRB(Pos); 940 Value *ShadowPtrMaskValue; 941 if (DFSanRuntimeShadowMask) 942 ShadowPtrMaskValue = IRB.CreateLoad(IntptrTy, ExternalShadowMask); 943 else 944 ShadowPtrMaskValue = ShadowPtrMask; 945 return IRB.CreateIntToPtr( 946 IRB.CreateMul( 947 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy), 948 IRB.CreatePtrToInt(ShadowPtrMaskValue, IntptrTy)), 949 ShadowPtrMul), 950 ShadowPtrTy); 951 } 952 953 // Generates IR to compute the union of the two given shadows, inserting it 954 // before Pos. Returns the computed union Value. 955 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) { 956 if (V1 == DFS.ZeroShadow) 957 return V2; 958 if (V2 == DFS.ZeroShadow) 959 return V1; 960 if (V1 == V2) 961 return V1; 962 963 auto V1Elems = ShadowElements.find(V1); 964 auto V2Elems = ShadowElements.find(V2); 965 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) { 966 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(), 967 V2Elems->second.begin(), V2Elems->second.end())) { 968 return V1; 969 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(), 970 V1Elems->second.begin(), V1Elems->second.end())) { 971 return V2; 972 } 973 } else if (V1Elems != ShadowElements.end()) { 974 if (V1Elems->second.count(V2)) 975 return V1; 976 } else if (V2Elems != ShadowElements.end()) { 977 if (V2Elems->second.count(V1)) 978 return V2; 979 } 980 981 auto Key = std::make_pair(V1, V2); 982 if (V1 > V2) 983 std::swap(Key.first, Key.second); 984 CachedCombinedShadow &CCS = CachedCombinedShadows[Key]; 985 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent())) 986 return CCS.Shadow; 987 988 IRBuilder<> IRB(Pos); 989 if (AvoidNewBlocks) { 990 CallInst *Call = IRB.CreateCall(DFS.DFSanCheckedUnionFn, {V1, V2}); 991 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 992 Call->addAttribute(1, Attribute::ZExt); 993 Call->addAttribute(2, Attribute::ZExt); 994 995 CCS.Block = Pos->getParent(); 996 CCS.Shadow = Call; 997 } else { 998 BasicBlock *Head = Pos->getParent(); 999 Value *Ne = IRB.CreateICmpNE(V1, V2); 1000 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen( 1001 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT)); 1002 IRBuilder<> ThenIRB(BI); 1003 CallInst *Call = ThenIRB.CreateCall(DFS.DFSanUnionFn, {V1, V2}); 1004 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1005 Call->addAttribute(1, Attribute::ZExt); 1006 Call->addAttribute(2, Attribute::ZExt); 1007 1008 BasicBlock *Tail = BI->getSuccessor(0); 1009 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front()); 1010 Phi->addIncoming(Call, Call->getParent()); 1011 Phi->addIncoming(V1, Head); 1012 1013 CCS.Block = Tail; 1014 CCS.Shadow = Phi; 1015 } 1016 1017 std::set<Value *> UnionElems; 1018 if (V1Elems != ShadowElements.end()) { 1019 UnionElems = V1Elems->second; 1020 } else { 1021 UnionElems.insert(V1); 1022 } 1023 if (V2Elems != ShadowElements.end()) { 1024 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end()); 1025 } else { 1026 UnionElems.insert(V2); 1027 } 1028 ShadowElements[CCS.Shadow] = std::move(UnionElems); 1029 1030 return CCS.Shadow; 1031 } 1032 1033 // A convenience function which folds the shadows of each of the operands 1034 // of the provided instruction Inst, inserting the IR before Inst. Returns 1035 // the computed union Value. 1036 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) { 1037 if (Inst->getNumOperands() == 0) 1038 return DFS.ZeroShadow; 1039 1040 Value *Shadow = getShadow(Inst->getOperand(0)); 1041 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) { 1042 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst); 1043 } 1044 return Shadow; 1045 } 1046 1047 void DFSanVisitor::visitOperandShadowInst(Instruction &I) { 1048 Value *CombinedShadow = DFSF.combineOperandShadows(&I); 1049 DFSF.setShadow(&I, CombinedShadow); 1050 } 1051 1052 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where 1053 // Addr has alignment Align, and take the union of each of those shadows. 1054 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align, 1055 Instruction *Pos) { 1056 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) { 1057 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i = 1058 AllocaShadowMap.find(AI); 1059 if (i != AllocaShadowMap.end()) { 1060 IRBuilder<> IRB(Pos); 1061 return IRB.CreateLoad(i->second); 1062 } 1063 } 1064 1065 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8; 1066 SmallVector<Value *, 2> Objs; 1067 GetUnderlyingObjects(Addr, Objs, Pos->getModule()->getDataLayout()); 1068 bool AllConstants = true; 1069 for (SmallVector<Value *, 2>::iterator i = Objs.begin(), e = Objs.end(); 1070 i != e; ++i) { 1071 if (isa<Function>(*i) || isa<BlockAddress>(*i)) 1072 continue; 1073 if (isa<GlobalVariable>(*i) && cast<GlobalVariable>(*i)->isConstant()) 1074 continue; 1075 1076 AllConstants = false; 1077 break; 1078 } 1079 if (AllConstants) 1080 return DFS.ZeroShadow; 1081 1082 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos); 1083 switch (Size) { 1084 case 0: 1085 return DFS.ZeroShadow; 1086 case 1: { 1087 LoadInst *LI = new LoadInst(ShadowAddr, "", Pos); 1088 LI->setAlignment(ShadowAlign); 1089 return LI; 1090 } 1091 case 2: { 1092 IRBuilder<> IRB(Pos); 1093 Value *ShadowAddr1 = IRB.CreateGEP(DFS.ShadowTy, ShadowAddr, 1094 ConstantInt::get(DFS.IntptrTy, 1)); 1095 return combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign), 1096 IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign), Pos); 1097 } 1098 } 1099 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) { 1100 // Fast path for the common case where each byte has identical shadow: load 1101 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any 1102 // shadow is non-equal. 1103 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F); 1104 IRBuilder<> FallbackIRB(FallbackBB); 1105 CallInst *FallbackCall = FallbackIRB.CreateCall( 1106 DFS.DFSanUnionLoadFn, 1107 {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)}); 1108 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1109 1110 // Compare each of the shadows stored in the loaded 64 bits to each other, 1111 // by computing (WideShadow rotl ShadowWidth) == WideShadow. 1112 IRBuilder<> IRB(Pos); 1113 Value *WideAddr = 1114 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx)); 1115 Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign); 1116 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy); 1117 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth); 1118 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth); 1119 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow); 1120 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow); 1121 1122 BasicBlock *Head = Pos->getParent(); 1123 BasicBlock *Tail = Head->splitBasicBlock(Pos->getIterator()); 1124 1125 if (DomTreeNode *OldNode = DT.getNode(Head)) { 1126 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end()); 1127 1128 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head); 1129 for (auto Child : Children) 1130 DT.changeImmediateDominator(Child, NewNode); 1131 } 1132 1133 // In the following code LastBr will refer to the previous basic block's 1134 // conditional branch instruction, whose true successor is fixed up to point 1135 // to the next block during the loop below or to the tail after the final 1136 // iteration. 1137 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq); 1138 ReplaceInstWithInst(Head->getTerminator(), LastBr); 1139 DT.addNewBlock(FallbackBB, Head); 1140 1141 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size; 1142 Ofs += 64 / DFS.ShadowWidth) { 1143 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F); 1144 DT.addNewBlock(NextBB, LastBr->getParent()); 1145 IRBuilder<> NextIRB(NextBB); 1146 WideAddr = NextIRB.CreateGEP(Type::getInt64Ty(*DFS.Ctx), WideAddr, 1147 ConstantInt::get(DFS.IntptrTy, 1)); 1148 Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign); 1149 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow); 1150 LastBr->setSuccessor(0, NextBB); 1151 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB); 1152 } 1153 1154 LastBr->setSuccessor(0, Tail); 1155 FallbackIRB.CreateBr(Tail); 1156 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front()); 1157 Shadow->addIncoming(FallbackCall, FallbackBB); 1158 Shadow->addIncoming(TruncShadow, LastBr->getParent()); 1159 return Shadow; 1160 } 1161 1162 IRBuilder<> IRB(Pos); 1163 CallInst *FallbackCall = IRB.CreateCall( 1164 DFS.DFSanUnionLoadFn, {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)}); 1165 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1166 return FallbackCall; 1167 } 1168 1169 void DFSanVisitor::visitLoadInst(LoadInst &LI) { 1170 auto &DL = LI.getModule()->getDataLayout(); 1171 uint64_t Size = DL.getTypeStoreSize(LI.getType()); 1172 if (Size == 0) { 1173 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow); 1174 return; 1175 } 1176 1177 uint64_t Align; 1178 if (ClPreserveAlignment) { 1179 Align = LI.getAlignment(); 1180 if (Align == 0) 1181 Align = DL.getABITypeAlignment(LI.getType()); 1182 } else { 1183 Align = 1; 1184 } 1185 IRBuilder<> IRB(&LI); 1186 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI); 1187 if (ClCombinePointerLabelsOnLoad) { 1188 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand()); 1189 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI); 1190 } 1191 if (Shadow != DFSF.DFS.ZeroShadow) 1192 DFSF.NonZeroChecks.push_back(Shadow); 1193 1194 DFSF.setShadow(&LI, Shadow); 1195 } 1196 1197 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align, 1198 Value *Shadow, Instruction *Pos) { 1199 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) { 1200 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i = 1201 AllocaShadowMap.find(AI); 1202 if (i != AllocaShadowMap.end()) { 1203 IRBuilder<> IRB(Pos); 1204 IRB.CreateStore(Shadow, i->second); 1205 return; 1206 } 1207 } 1208 1209 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8; 1210 IRBuilder<> IRB(Pos); 1211 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos); 1212 if (Shadow == DFS.ZeroShadow) { 1213 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth); 1214 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0); 1215 Value *ExtShadowAddr = 1216 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy)); 1217 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign); 1218 return; 1219 } 1220 1221 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth; 1222 uint64_t Offset = 0; 1223 if (Size >= ShadowVecSize) { 1224 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize); 1225 Value *ShadowVec = UndefValue::get(ShadowVecTy); 1226 for (unsigned i = 0; i != ShadowVecSize; ++i) { 1227 ShadowVec = IRB.CreateInsertElement( 1228 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i)); 1229 } 1230 Value *ShadowVecAddr = 1231 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy)); 1232 do { 1233 Value *CurShadowVecAddr = 1234 IRB.CreateConstGEP1_32(ShadowVecTy, ShadowVecAddr, Offset); 1235 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign); 1236 Size -= ShadowVecSize; 1237 ++Offset; 1238 } while (Size >= ShadowVecSize); 1239 Offset *= ShadowVecSize; 1240 } 1241 while (Size > 0) { 1242 Value *CurShadowAddr = 1243 IRB.CreateConstGEP1_32(DFS.ShadowTy, ShadowAddr, Offset); 1244 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign); 1245 --Size; 1246 ++Offset; 1247 } 1248 } 1249 1250 void DFSanVisitor::visitStoreInst(StoreInst &SI) { 1251 auto &DL = SI.getModule()->getDataLayout(); 1252 uint64_t Size = DL.getTypeStoreSize(SI.getValueOperand()->getType()); 1253 if (Size == 0) 1254 return; 1255 1256 uint64_t Align; 1257 if (ClPreserveAlignment) { 1258 Align = SI.getAlignment(); 1259 if (Align == 0) 1260 Align = DL.getABITypeAlignment(SI.getValueOperand()->getType()); 1261 } else { 1262 Align = 1; 1263 } 1264 1265 Value* Shadow = DFSF.getShadow(SI.getValueOperand()); 1266 if (ClCombinePointerLabelsOnStore) { 1267 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand()); 1268 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI); 1269 } 1270 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI); 1271 } 1272 1273 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) { 1274 visitOperandShadowInst(BO); 1275 } 1276 1277 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); } 1278 1279 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); } 1280 1281 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) { 1282 visitOperandShadowInst(GEPI); 1283 } 1284 1285 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) { 1286 visitOperandShadowInst(I); 1287 } 1288 1289 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) { 1290 visitOperandShadowInst(I); 1291 } 1292 1293 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) { 1294 visitOperandShadowInst(I); 1295 } 1296 1297 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) { 1298 visitOperandShadowInst(I); 1299 } 1300 1301 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) { 1302 visitOperandShadowInst(I); 1303 } 1304 1305 void DFSanVisitor::visitAllocaInst(AllocaInst &I) { 1306 bool AllLoadsStores = true; 1307 for (User *U : I.users()) { 1308 if (isa<LoadInst>(U)) 1309 continue; 1310 1311 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1312 if (SI->getPointerOperand() == &I) 1313 continue; 1314 } 1315 1316 AllLoadsStores = false; 1317 break; 1318 } 1319 if (AllLoadsStores) { 1320 IRBuilder<> IRB(&I); 1321 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy); 1322 } 1323 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow); 1324 } 1325 1326 void DFSanVisitor::visitSelectInst(SelectInst &I) { 1327 Value *CondShadow = DFSF.getShadow(I.getCondition()); 1328 Value *TrueShadow = DFSF.getShadow(I.getTrueValue()); 1329 Value *FalseShadow = DFSF.getShadow(I.getFalseValue()); 1330 1331 if (isa<VectorType>(I.getCondition()->getType())) { 1332 DFSF.setShadow( 1333 &I, 1334 DFSF.combineShadows( 1335 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I)); 1336 } else { 1337 Value *ShadowSel; 1338 if (TrueShadow == FalseShadow) { 1339 ShadowSel = TrueShadow; 1340 } else { 1341 ShadowSel = 1342 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I); 1343 } 1344 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I)); 1345 } 1346 } 1347 1348 void DFSanVisitor::visitMemSetInst(MemSetInst &I) { 1349 IRBuilder<> IRB(&I); 1350 Value *ValShadow = DFSF.getShadow(I.getValue()); 1351 IRB.CreateCall(DFSF.DFS.DFSanSetLabelFn, 1352 {ValShadow, IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy( 1353 *DFSF.DFS.Ctx)), 1354 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)}); 1355 } 1356 1357 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) { 1358 IRBuilder<> IRB(&I); 1359 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I); 1360 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I); 1361 Value *LenShadow = IRB.CreateMul( 1362 I.getLength(), 1363 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8)); 1364 Value *AlignShadow; 1365 if (ClPreserveAlignment) { 1366 AlignShadow = IRB.CreateMul(I.getAlignmentCst(), 1367 ConstantInt::get(I.getAlignmentCst()->getType(), 1368 DFSF.DFS.ShadowWidth / 8)); 1369 } else { 1370 AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(), 1371 DFSF.DFS.ShadowWidth / 8); 1372 } 1373 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx); 1374 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr); 1375 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr); 1376 IRB.CreateCall(I.getCalledValue(), {DestShadow, SrcShadow, LenShadow, 1377 AlignShadow, I.getVolatileCst()}); 1378 } 1379 1380 void DFSanVisitor::visitReturnInst(ReturnInst &RI) { 1381 if (!DFSF.IsNativeABI && RI.getReturnValue()) { 1382 switch (DFSF.IA) { 1383 case DataFlowSanitizer::IA_TLS: { 1384 Value *S = DFSF.getShadow(RI.getReturnValue()); 1385 IRBuilder<> IRB(&RI); 1386 IRB.CreateStore(S, DFSF.getRetvalTLS()); 1387 break; 1388 } 1389 case DataFlowSanitizer::IA_Args: { 1390 IRBuilder<> IRB(&RI); 1391 Type *RT = DFSF.F->getFunctionType()->getReturnType(); 1392 Value *InsVal = 1393 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0); 1394 Value *InsShadow = 1395 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1); 1396 RI.setOperand(0, InsShadow); 1397 break; 1398 } 1399 } 1400 } 1401 } 1402 1403 void DFSanVisitor::visitCallSite(CallSite CS) { 1404 Function *F = CS.getCalledFunction(); 1405 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) { 1406 visitOperandShadowInst(*CS.getInstruction()); 1407 return; 1408 } 1409 1410 // Calls to this function are synthesized in wrappers, and we shouldn't 1411 // instrument them. 1412 if (F == DFSF.DFS.DFSanVarargWrapperFn) 1413 return; 1414 1415 assert(!(cast<FunctionType>( 1416 CS.getCalledValue()->getType()->getPointerElementType())->isVarArg() && 1417 dyn_cast<InvokeInst>(CS.getInstruction()))); 1418 1419 IRBuilder<> IRB(CS.getInstruction()); 1420 1421 DenseMap<Value *, Function *>::iterator i = 1422 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue()); 1423 if (i != DFSF.DFS.UnwrappedFnMap.end()) { 1424 Function *F = i->second; 1425 switch (DFSF.DFS.getWrapperKind(F)) { 1426 case DataFlowSanitizer::WK_Warning: { 1427 CS.setCalledFunction(F); 1428 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn, 1429 IRB.CreateGlobalStringPtr(F->getName())); 1430 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow); 1431 return; 1432 } 1433 case DataFlowSanitizer::WK_Discard: { 1434 CS.setCalledFunction(F); 1435 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow); 1436 return; 1437 } 1438 case DataFlowSanitizer::WK_Functional: { 1439 CS.setCalledFunction(F); 1440 visitOperandShadowInst(*CS.getInstruction()); 1441 return; 1442 } 1443 case DataFlowSanitizer::WK_Custom: { 1444 // Don't try to handle invokes of custom functions, it's too complicated. 1445 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_ 1446 // wrapper. 1447 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 1448 FunctionType *FT = F->getFunctionType(); 1449 FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT); 1450 std::string CustomFName = "__dfsw_"; 1451 CustomFName += F->getName(); 1452 Constant *CustomF = 1453 DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT); 1454 if (Function *CustomFn = dyn_cast<Function>(CustomF)) { 1455 CustomFn->copyAttributesFrom(F); 1456 1457 // Custom functions returning non-void will write to the return label. 1458 if (!FT->getReturnType()->isVoidTy()) { 1459 CustomFn->removeAttributes(AttributeSet::FunctionIndex, 1460 DFSF.DFS.ReadOnlyNoneAttrs); 1461 } 1462 } 1463 1464 std::vector<Value *> Args; 1465 1466 CallSite::arg_iterator i = CS.arg_begin(); 1467 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) { 1468 Type *T = (*i)->getType(); 1469 FunctionType *ParamFT; 1470 if (isa<PointerType>(T) && 1471 (ParamFT = dyn_cast<FunctionType>( 1472 cast<PointerType>(T)->getElementType()))) { 1473 std::string TName = "dfst"; 1474 TName += utostr(FT->getNumParams() - n); 1475 TName += "$"; 1476 TName += F->getName(); 1477 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName); 1478 Args.push_back(T); 1479 Args.push_back( 1480 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx))); 1481 } else { 1482 Args.push_back(*i); 1483 } 1484 } 1485 1486 i = CS.arg_begin(); 1487 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1488 Args.push_back(DFSF.getShadow(*i)); 1489 1490 if (FT->isVarArg()) { 1491 auto *LabelVATy = ArrayType::get(DFSF.DFS.ShadowTy, 1492 CS.arg_size() - FT->getNumParams()); 1493 auto *LabelVAAlloca = new AllocaInst( 1494 LabelVATy, "labelva", &DFSF.F->getEntryBlock().front()); 1495 1496 for (unsigned n = 0; i != CS.arg_end(); ++i, ++n) { 1497 auto LabelVAPtr = IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, n); 1498 IRB.CreateStore(DFSF.getShadow(*i), LabelVAPtr); 1499 } 1500 1501 Args.push_back(IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, 0)); 1502 } 1503 1504 if (!FT->getReturnType()->isVoidTy()) { 1505 if (!DFSF.LabelReturnAlloca) { 1506 DFSF.LabelReturnAlloca = 1507 new AllocaInst(DFSF.DFS.ShadowTy, "labelreturn", 1508 &DFSF.F->getEntryBlock().front()); 1509 } 1510 Args.push_back(DFSF.LabelReturnAlloca); 1511 } 1512 1513 for (i = CS.arg_begin() + FT->getNumParams(); i != CS.arg_end(); ++i) 1514 Args.push_back(*i); 1515 1516 CallInst *CustomCI = IRB.CreateCall(CustomF, Args); 1517 CustomCI->setCallingConv(CI->getCallingConv()); 1518 CustomCI->setAttributes(CI->getAttributes()); 1519 1520 if (!FT->getReturnType()->isVoidTy()) { 1521 LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca); 1522 DFSF.setShadow(CustomCI, LabelLoad); 1523 } 1524 1525 CI->replaceAllUsesWith(CustomCI); 1526 CI->eraseFromParent(); 1527 return; 1528 } 1529 break; 1530 } 1531 } 1532 } 1533 1534 FunctionType *FT = cast<FunctionType>( 1535 CS.getCalledValue()->getType()->getPointerElementType()); 1536 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) { 1537 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) { 1538 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)), 1539 DFSF.getArgTLS(i, CS.getInstruction())); 1540 } 1541 } 1542 1543 Instruction *Next = nullptr; 1544 if (!CS.getType()->isVoidTy()) { 1545 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1546 if (II->getNormalDest()->getSinglePredecessor()) { 1547 Next = &II->getNormalDest()->front(); 1548 } else { 1549 BasicBlock *NewBB = 1550 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT); 1551 Next = &NewBB->front(); 1552 } 1553 } else { 1554 assert(CS->getIterator() != CS->getParent()->end()); 1555 Next = CS->getNextNode(); 1556 } 1557 1558 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) { 1559 IRBuilder<> NextIRB(Next); 1560 LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS()); 1561 DFSF.SkipInsts.insert(LI); 1562 DFSF.setShadow(CS.getInstruction(), LI); 1563 DFSF.NonZeroChecks.push_back(LI); 1564 } 1565 } 1566 1567 // Do all instrumentation for IA_Args down here to defer tampering with the 1568 // CFG in a way that SplitEdge may be able to detect. 1569 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) { 1570 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT); 1571 Value *Func = 1572 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT)); 1573 std::vector<Value *> Args; 1574 1575 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 1576 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1577 Args.push_back(*i); 1578 1579 i = CS.arg_begin(); 1580 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1581 Args.push_back(DFSF.getShadow(*i)); 1582 1583 if (FT->isVarArg()) { 1584 unsigned VarArgSize = CS.arg_size() - FT->getNumParams(); 1585 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize); 1586 AllocaInst *VarArgShadow = 1587 new AllocaInst(VarArgArrayTy, "", &DFSF.F->getEntryBlock().front()); 1588 Args.push_back(IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, 0)); 1589 for (unsigned n = 0; i != e; ++i, ++n) { 1590 IRB.CreateStore( 1591 DFSF.getShadow(*i), 1592 IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, n)); 1593 Args.push_back(*i); 1594 } 1595 } 1596 1597 CallSite NewCS; 1598 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1599 NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(), 1600 Args); 1601 } else { 1602 NewCS = IRB.CreateCall(Func, Args); 1603 } 1604 NewCS.setCallingConv(CS.getCallingConv()); 1605 NewCS.setAttributes(CS.getAttributes().removeAttributes( 1606 *DFSF.DFS.Ctx, AttributeSet::ReturnIndex, 1607 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType()))); 1608 1609 if (Next) { 1610 ExtractValueInst *ExVal = 1611 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next); 1612 DFSF.SkipInsts.insert(ExVal); 1613 ExtractValueInst *ExShadow = 1614 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next); 1615 DFSF.SkipInsts.insert(ExShadow); 1616 DFSF.setShadow(ExVal, ExShadow); 1617 DFSF.NonZeroChecks.push_back(ExShadow); 1618 1619 CS.getInstruction()->replaceAllUsesWith(ExVal); 1620 } 1621 1622 CS.getInstruction()->eraseFromParent(); 1623 } 1624 } 1625 1626 void DFSanVisitor::visitPHINode(PHINode &PN) { 1627 PHINode *ShadowPN = 1628 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN); 1629 1630 // Give the shadow phi node valid predecessors to fool SplitEdge into working. 1631 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy); 1632 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e; 1633 ++i) { 1634 ShadowPN->addIncoming(UndefShadow, *i); 1635 } 1636 1637 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN)); 1638 DFSF.setShadow(&PN, ShadowPN); 1639 } 1640