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.getValueType(); 138 // For now we support blacklisting struct types only. 139 if (StructType *SGType = dyn_cast<StructType>(GType)) { 140 if (!SGType->isLiteral()) 141 return SGType->getName(); 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.getValueType())) 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 (Function *i : FnsToInstrument) { 795 if (!i || i->isDeclaration()) 796 continue; 797 798 removeUnreachableBlocks(*i); 799 800 DFSanFunction DFSF(*this, i, FnsWithNativeABI.count(i)); 801 802 // DFSanVisitor may create new basic blocks, which confuses df_iterator. 803 // Build a copy of the list before iterating over it. 804 llvm::SmallVector<BasicBlock *, 4> BBList(depth_first(&i->getEntryBlock())); 805 806 for (BasicBlock *i : BBList) { 807 Instruction *Inst = &i->front(); 808 while (1) { 809 // DFSanVisitor may split the current basic block, changing the current 810 // instruction's next pointer and moving the next instruction to the 811 // tail block from which we should continue. 812 Instruction *Next = Inst->getNextNode(); 813 // DFSanVisitor may delete Inst, so keep track of whether it was a 814 // terminator. 815 bool IsTerminator = isa<TerminatorInst>(Inst); 816 if (!DFSF.SkipInsts.count(Inst)) 817 DFSanVisitor(DFSF).visit(Inst); 818 if (IsTerminator) 819 break; 820 Inst = Next; 821 } 822 } 823 824 // We will not necessarily be able to compute the shadow for every phi node 825 // until we have visited every block. Therefore, the code that handles phi 826 // nodes adds them to the PHIFixups list so that they can be properly 827 // handled here. 828 for (std::vector<std::pair<PHINode *, PHINode *> >::iterator 829 i = DFSF.PHIFixups.begin(), 830 e = DFSF.PHIFixups.end(); 831 i != e; ++i) { 832 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n; 833 ++val) { 834 i->second->setIncomingValue( 835 val, DFSF.getShadow(i->first->getIncomingValue(val))); 836 } 837 } 838 839 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy 840 // places (i.e. instructions in basic blocks we haven't even begun visiting 841 // yet). To make our life easier, do this work in a pass after the main 842 // instrumentation. 843 if (ClDebugNonzeroLabels) { 844 for (Value *V : DFSF.NonZeroChecks) { 845 Instruction *Pos; 846 if (Instruction *I = dyn_cast<Instruction>(V)) 847 Pos = I->getNextNode(); 848 else 849 Pos = &DFSF.F->getEntryBlock().front(); 850 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos)) 851 Pos = Pos->getNextNode(); 852 IRBuilder<> IRB(Pos); 853 Value *Ne = IRB.CreateICmpNE(V, DFSF.DFS.ZeroShadow); 854 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen( 855 Ne, Pos, /*Unreachable=*/false, ColdCallWeights)); 856 IRBuilder<> ThenIRB(BI); 857 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn, {}); 858 } 859 } 860 } 861 862 return false; 863 } 864 865 Value *DFSanFunction::getArgTLSPtr() { 866 if (ArgTLSPtr) 867 return ArgTLSPtr; 868 if (DFS.ArgTLS) 869 return ArgTLSPtr = DFS.ArgTLS; 870 871 IRBuilder<> IRB(&F->getEntryBlock().front()); 872 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLS, {}); 873 } 874 875 Value *DFSanFunction::getRetvalTLS() { 876 if (RetvalTLSPtr) 877 return RetvalTLSPtr; 878 if (DFS.RetvalTLS) 879 return RetvalTLSPtr = DFS.RetvalTLS; 880 881 IRBuilder<> IRB(&F->getEntryBlock().front()); 882 return RetvalTLSPtr = IRB.CreateCall(DFS.GetRetvalTLS, {}); 883 } 884 885 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) { 886 IRBuilder<> IRB(Pos); 887 return IRB.CreateConstGEP2_64(getArgTLSPtr(), 0, Idx); 888 } 889 890 Value *DFSanFunction::getShadow(Value *V) { 891 if (!isa<Argument>(V) && !isa<Instruction>(V)) 892 return DFS.ZeroShadow; 893 Value *&Shadow = ValShadowMap[V]; 894 if (!Shadow) { 895 if (Argument *A = dyn_cast<Argument>(V)) { 896 if (IsNativeABI) 897 return DFS.ZeroShadow; 898 switch (IA) { 899 case DataFlowSanitizer::IA_TLS: { 900 Value *ArgTLSPtr = getArgTLSPtr(); 901 Instruction *ArgTLSPos = 902 DFS.ArgTLS ? &*F->getEntryBlock().begin() 903 : cast<Instruction>(ArgTLSPtr)->getNextNode(); 904 IRBuilder<> IRB(ArgTLSPos); 905 Shadow = IRB.CreateLoad(getArgTLS(A->getArgNo(), ArgTLSPos)); 906 break; 907 } 908 case DataFlowSanitizer::IA_Args: { 909 unsigned ArgIdx = A->getArgNo() + F->getArgumentList().size() / 2; 910 Function::arg_iterator i = F->arg_begin(); 911 while (ArgIdx--) 912 ++i; 913 Shadow = &*i; 914 assert(Shadow->getType() == DFS.ShadowTy); 915 break; 916 } 917 } 918 NonZeroChecks.push_back(Shadow); 919 } else { 920 Shadow = DFS.ZeroShadow; 921 } 922 } 923 return Shadow; 924 } 925 926 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) { 927 assert(!ValShadowMap.count(I)); 928 assert(Shadow->getType() == DFS.ShadowTy); 929 ValShadowMap[I] = Shadow; 930 } 931 932 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) { 933 assert(Addr != RetvalTLS && "Reinstrumenting?"); 934 IRBuilder<> IRB(Pos); 935 Value *ShadowPtrMaskValue; 936 if (DFSanRuntimeShadowMask) 937 ShadowPtrMaskValue = IRB.CreateLoad(IntptrTy, ExternalShadowMask); 938 else 939 ShadowPtrMaskValue = ShadowPtrMask; 940 return IRB.CreateIntToPtr( 941 IRB.CreateMul( 942 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy), 943 IRB.CreatePtrToInt(ShadowPtrMaskValue, IntptrTy)), 944 ShadowPtrMul), 945 ShadowPtrTy); 946 } 947 948 // Generates IR to compute the union of the two given shadows, inserting it 949 // before Pos. Returns the computed union Value. 950 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) { 951 if (V1 == DFS.ZeroShadow) 952 return V2; 953 if (V2 == DFS.ZeroShadow) 954 return V1; 955 if (V1 == V2) 956 return V1; 957 958 auto V1Elems = ShadowElements.find(V1); 959 auto V2Elems = ShadowElements.find(V2); 960 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) { 961 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(), 962 V2Elems->second.begin(), V2Elems->second.end())) { 963 return V1; 964 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(), 965 V1Elems->second.begin(), V1Elems->second.end())) { 966 return V2; 967 } 968 } else if (V1Elems != ShadowElements.end()) { 969 if (V1Elems->second.count(V2)) 970 return V1; 971 } else if (V2Elems != ShadowElements.end()) { 972 if (V2Elems->second.count(V1)) 973 return V2; 974 } 975 976 auto Key = std::make_pair(V1, V2); 977 if (V1 > V2) 978 std::swap(Key.first, Key.second); 979 CachedCombinedShadow &CCS = CachedCombinedShadows[Key]; 980 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent())) 981 return CCS.Shadow; 982 983 IRBuilder<> IRB(Pos); 984 if (AvoidNewBlocks) { 985 CallInst *Call = IRB.CreateCall(DFS.DFSanCheckedUnionFn, {V1, V2}); 986 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 987 Call->addAttribute(1, Attribute::ZExt); 988 Call->addAttribute(2, Attribute::ZExt); 989 990 CCS.Block = Pos->getParent(); 991 CCS.Shadow = Call; 992 } else { 993 BasicBlock *Head = Pos->getParent(); 994 Value *Ne = IRB.CreateICmpNE(V1, V2); 995 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen( 996 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT)); 997 IRBuilder<> ThenIRB(BI); 998 CallInst *Call = ThenIRB.CreateCall(DFS.DFSanUnionFn, {V1, V2}); 999 Call->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1000 Call->addAttribute(1, Attribute::ZExt); 1001 Call->addAttribute(2, Attribute::ZExt); 1002 1003 BasicBlock *Tail = BI->getSuccessor(0); 1004 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front()); 1005 Phi->addIncoming(Call, Call->getParent()); 1006 Phi->addIncoming(V1, Head); 1007 1008 CCS.Block = Tail; 1009 CCS.Shadow = Phi; 1010 } 1011 1012 std::set<Value *> UnionElems; 1013 if (V1Elems != ShadowElements.end()) { 1014 UnionElems = V1Elems->second; 1015 } else { 1016 UnionElems.insert(V1); 1017 } 1018 if (V2Elems != ShadowElements.end()) { 1019 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end()); 1020 } else { 1021 UnionElems.insert(V2); 1022 } 1023 ShadowElements[CCS.Shadow] = std::move(UnionElems); 1024 1025 return CCS.Shadow; 1026 } 1027 1028 // A convenience function which folds the shadows of each of the operands 1029 // of the provided instruction Inst, inserting the IR before Inst. Returns 1030 // the computed union Value. 1031 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) { 1032 if (Inst->getNumOperands() == 0) 1033 return DFS.ZeroShadow; 1034 1035 Value *Shadow = getShadow(Inst->getOperand(0)); 1036 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) { 1037 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst); 1038 } 1039 return Shadow; 1040 } 1041 1042 void DFSanVisitor::visitOperandShadowInst(Instruction &I) { 1043 Value *CombinedShadow = DFSF.combineOperandShadows(&I); 1044 DFSF.setShadow(&I, CombinedShadow); 1045 } 1046 1047 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where 1048 // Addr has alignment Align, and take the union of each of those shadows. 1049 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align, 1050 Instruction *Pos) { 1051 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) { 1052 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i = 1053 AllocaShadowMap.find(AI); 1054 if (i != AllocaShadowMap.end()) { 1055 IRBuilder<> IRB(Pos); 1056 return IRB.CreateLoad(i->second); 1057 } 1058 } 1059 1060 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8; 1061 SmallVector<Value *, 2> Objs; 1062 GetUnderlyingObjects(Addr, Objs, Pos->getModule()->getDataLayout()); 1063 bool AllConstants = true; 1064 for (Value *Obj : Objs) { 1065 if (isa<Function>(Obj) || isa<BlockAddress>(Obj)) 1066 continue; 1067 if (isa<GlobalVariable>(Obj) && cast<GlobalVariable>(Obj)->isConstant()) 1068 continue; 1069 1070 AllConstants = false; 1071 break; 1072 } 1073 if (AllConstants) 1074 return DFS.ZeroShadow; 1075 1076 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos); 1077 switch (Size) { 1078 case 0: 1079 return DFS.ZeroShadow; 1080 case 1: { 1081 LoadInst *LI = new LoadInst(ShadowAddr, "", Pos); 1082 LI->setAlignment(ShadowAlign); 1083 return LI; 1084 } 1085 case 2: { 1086 IRBuilder<> IRB(Pos); 1087 Value *ShadowAddr1 = IRB.CreateGEP(DFS.ShadowTy, ShadowAddr, 1088 ConstantInt::get(DFS.IntptrTy, 1)); 1089 return combineShadows(IRB.CreateAlignedLoad(ShadowAddr, ShadowAlign), 1090 IRB.CreateAlignedLoad(ShadowAddr1, ShadowAlign), Pos); 1091 } 1092 } 1093 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) { 1094 // Fast path for the common case where each byte has identical shadow: load 1095 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any 1096 // shadow is non-equal. 1097 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F); 1098 IRBuilder<> FallbackIRB(FallbackBB); 1099 CallInst *FallbackCall = FallbackIRB.CreateCall( 1100 DFS.DFSanUnionLoadFn, 1101 {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)}); 1102 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1103 1104 // Compare each of the shadows stored in the loaded 64 bits to each other, 1105 // by computing (WideShadow rotl ShadowWidth) == WideShadow. 1106 IRBuilder<> IRB(Pos); 1107 Value *WideAddr = 1108 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx)); 1109 Value *WideShadow = IRB.CreateAlignedLoad(WideAddr, ShadowAlign); 1110 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy); 1111 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth); 1112 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth); 1113 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow); 1114 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow); 1115 1116 BasicBlock *Head = Pos->getParent(); 1117 BasicBlock *Tail = Head->splitBasicBlock(Pos->getIterator()); 1118 1119 if (DomTreeNode *OldNode = DT.getNode(Head)) { 1120 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end()); 1121 1122 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head); 1123 for (auto Child : Children) 1124 DT.changeImmediateDominator(Child, NewNode); 1125 } 1126 1127 // In the following code LastBr will refer to the previous basic block's 1128 // conditional branch instruction, whose true successor is fixed up to point 1129 // to the next block during the loop below or to the tail after the final 1130 // iteration. 1131 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq); 1132 ReplaceInstWithInst(Head->getTerminator(), LastBr); 1133 DT.addNewBlock(FallbackBB, Head); 1134 1135 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size; 1136 Ofs += 64 / DFS.ShadowWidth) { 1137 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F); 1138 DT.addNewBlock(NextBB, LastBr->getParent()); 1139 IRBuilder<> NextIRB(NextBB); 1140 WideAddr = NextIRB.CreateGEP(Type::getInt64Ty(*DFS.Ctx), WideAddr, 1141 ConstantInt::get(DFS.IntptrTy, 1)); 1142 Value *NextWideShadow = NextIRB.CreateAlignedLoad(WideAddr, ShadowAlign); 1143 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow); 1144 LastBr->setSuccessor(0, NextBB); 1145 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB); 1146 } 1147 1148 LastBr->setSuccessor(0, Tail); 1149 FallbackIRB.CreateBr(Tail); 1150 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front()); 1151 Shadow->addIncoming(FallbackCall, FallbackBB); 1152 Shadow->addIncoming(TruncShadow, LastBr->getParent()); 1153 return Shadow; 1154 } 1155 1156 IRBuilder<> IRB(Pos); 1157 CallInst *FallbackCall = IRB.CreateCall( 1158 DFS.DFSanUnionLoadFn, {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)}); 1159 FallbackCall->addAttribute(AttributeSet::ReturnIndex, Attribute::ZExt); 1160 return FallbackCall; 1161 } 1162 1163 void DFSanVisitor::visitLoadInst(LoadInst &LI) { 1164 auto &DL = LI.getModule()->getDataLayout(); 1165 uint64_t Size = DL.getTypeStoreSize(LI.getType()); 1166 if (Size == 0) { 1167 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow); 1168 return; 1169 } 1170 1171 uint64_t Align; 1172 if (ClPreserveAlignment) { 1173 Align = LI.getAlignment(); 1174 if (Align == 0) 1175 Align = DL.getABITypeAlignment(LI.getType()); 1176 } else { 1177 Align = 1; 1178 } 1179 IRBuilder<> IRB(&LI); 1180 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI); 1181 if (ClCombinePointerLabelsOnLoad) { 1182 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand()); 1183 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI); 1184 } 1185 if (Shadow != DFSF.DFS.ZeroShadow) 1186 DFSF.NonZeroChecks.push_back(Shadow); 1187 1188 DFSF.setShadow(&LI, Shadow); 1189 } 1190 1191 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align, 1192 Value *Shadow, Instruction *Pos) { 1193 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) { 1194 llvm::DenseMap<AllocaInst *, AllocaInst *>::iterator i = 1195 AllocaShadowMap.find(AI); 1196 if (i != AllocaShadowMap.end()) { 1197 IRBuilder<> IRB(Pos); 1198 IRB.CreateStore(Shadow, i->second); 1199 return; 1200 } 1201 } 1202 1203 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8; 1204 IRBuilder<> IRB(Pos); 1205 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos); 1206 if (Shadow == DFS.ZeroShadow) { 1207 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth); 1208 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0); 1209 Value *ExtShadowAddr = 1210 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy)); 1211 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign); 1212 return; 1213 } 1214 1215 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth; 1216 uint64_t Offset = 0; 1217 if (Size >= ShadowVecSize) { 1218 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize); 1219 Value *ShadowVec = UndefValue::get(ShadowVecTy); 1220 for (unsigned i = 0; i != ShadowVecSize; ++i) { 1221 ShadowVec = IRB.CreateInsertElement( 1222 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i)); 1223 } 1224 Value *ShadowVecAddr = 1225 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy)); 1226 do { 1227 Value *CurShadowVecAddr = 1228 IRB.CreateConstGEP1_32(ShadowVecTy, ShadowVecAddr, Offset); 1229 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign); 1230 Size -= ShadowVecSize; 1231 ++Offset; 1232 } while (Size >= ShadowVecSize); 1233 Offset *= ShadowVecSize; 1234 } 1235 while (Size > 0) { 1236 Value *CurShadowAddr = 1237 IRB.CreateConstGEP1_32(DFS.ShadowTy, ShadowAddr, Offset); 1238 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign); 1239 --Size; 1240 ++Offset; 1241 } 1242 } 1243 1244 void DFSanVisitor::visitStoreInst(StoreInst &SI) { 1245 auto &DL = SI.getModule()->getDataLayout(); 1246 uint64_t Size = DL.getTypeStoreSize(SI.getValueOperand()->getType()); 1247 if (Size == 0) 1248 return; 1249 1250 uint64_t Align; 1251 if (ClPreserveAlignment) { 1252 Align = SI.getAlignment(); 1253 if (Align == 0) 1254 Align = DL.getABITypeAlignment(SI.getValueOperand()->getType()); 1255 } else { 1256 Align = 1; 1257 } 1258 1259 Value* Shadow = DFSF.getShadow(SI.getValueOperand()); 1260 if (ClCombinePointerLabelsOnStore) { 1261 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand()); 1262 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI); 1263 } 1264 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI); 1265 } 1266 1267 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) { 1268 visitOperandShadowInst(BO); 1269 } 1270 1271 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); } 1272 1273 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); } 1274 1275 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) { 1276 visitOperandShadowInst(GEPI); 1277 } 1278 1279 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) { 1280 visitOperandShadowInst(I); 1281 } 1282 1283 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) { 1284 visitOperandShadowInst(I); 1285 } 1286 1287 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) { 1288 visitOperandShadowInst(I); 1289 } 1290 1291 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) { 1292 visitOperandShadowInst(I); 1293 } 1294 1295 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) { 1296 visitOperandShadowInst(I); 1297 } 1298 1299 void DFSanVisitor::visitAllocaInst(AllocaInst &I) { 1300 bool AllLoadsStores = true; 1301 for (User *U : I.users()) { 1302 if (isa<LoadInst>(U)) 1303 continue; 1304 1305 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1306 if (SI->getPointerOperand() == &I) 1307 continue; 1308 } 1309 1310 AllLoadsStores = false; 1311 break; 1312 } 1313 if (AllLoadsStores) { 1314 IRBuilder<> IRB(&I); 1315 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy); 1316 } 1317 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow); 1318 } 1319 1320 void DFSanVisitor::visitSelectInst(SelectInst &I) { 1321 Value *CondShadow = DFSF.getShadow(I.getCondition()); 1322 Value *TrueShadow = DFSF.getShadow(I.getTrueValue()); 1323 Value *FalseShadow = DFSF.getShadow(I.getFalseValue()); 1324 1325 if (isa<VectorType>(I.getCondition()->getType())) { 1326 DFSF.setShadow( 1327 &I, 1328 DFSF.combineShadows( 1329 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I)); 1330 } else { 1331 Value *ShadowSel; 1332 if (TrueShadow == FalseShadow) { 1333 ShadowSel = TrueShadow; 1334 } else { 1335 ShadowSel = 1336 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I); 1337 } 1338 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I)); 1339 } 1340 } 1341 1342 void DFSanVisitor::visitMemSetInst(MemSetInst &I) { 1343 IRBuilder<> IRB(&I); 1344 Value *ValShadow = DFSF.getShadow(I.getValue()); 1345 IRB.CreateCall(DFSF.DFS.DFSanSetLabelFn, 1346 {ValShadow, IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy( 1347 *DFSF.DFS.Ctx)), 1348 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)}); 1349 } 1350 1351 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) { 1352 IRBuilder<> IRB(&I); 1353 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I); 1354 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I); 1355 Value *LenShadow = IRB.CreateMul( 1356 I.getLength(), 1357 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8)); 1358 Value *AlignShadow; 1359 if (ClPreserveAlignment) { 1360 AlignShadow = IRB.CreateMul(I.getAlignmentCst(), 1361 ConstantInt::get(I.getAlignmentCst()->getType(), 1362 DFSF.DFS.ShadowWidth / 8)); 1363 } else { 1364 AlignShadow = ConstantInt::get(I.getAlignmentCst()->getType(), 1365 DFSF.DFS.ShadowWidth / 8); 1366 } 1367 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx); 1368 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr); 1369 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr); 1370 IRB.CreateCall(I.getCalledValue(), {DestShadow, SrcShadow, LenShadow, 1371 AlignShadow, I.getVolatileCst()}); 1372 } 1373 1374 void DFSanVisitor::visitReturnInst(ReturnInst &RI) { 1375 if (!DFSF.IsNativeABI && RI.getReturnValue()) { 1376 switch (DFSF.IA) { 1377 case DataFlowSanitizer::IA_TLS: { 1378 Value *S = DFSF.getShadow(RI.getReturnValue()); 1379 IRBuilder<> IRB(&RI); 1380 IRB.CreateStore(S, DFSF.getRetvalTLS()); 1381 break; 1382 } 1383 case DataFlowSanitizer::IA_Args: { 1384 IRBuilder<> IRB(&RI); 1385 Type *RT = DFSF.F->getFunctionType()->getReturnType(); 1386 Value *InsVal = 1387 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0); 1388 Value *InsShadow = 1389 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1); 1390 RI.setOperand(0, InsShadow); 1391 break; 1392 } 1393 } 1394 } 1395 } 1396 1397 void DFSanVisitor::visitCallSite(CallSite CS) { 1398 Function *F = CS.getCalledFunction(); 1399 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) { 1400 visitOperandShadowInst(*CS.getInstruction()); 1401 return; 1402 } 1403 1404 // Calls to this function are synthesized in wrappers, and we shouldn't 1405 // instrument them. 1406 if (F == DFSF.DFS.DFSanVarargWrapperFn) 1407 return; 1408 1409 IRBuilder<> IRB(CS.getInstruction()); 1410 1411 DenseMap<Value *, Function *>::iterator i = 1412 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue()); 1413 if (i != DFSF.DFS.UnwrappedFnMap.end()) { 1414 Function *F = i->second; 1415 switch (DFSF.DFS.getWrapperKind(F)) { 1416 case DataFlowSanitizer::WK_Warning: { 1417 CS.setCalledFunction(F); 1418 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn, 1419 IRB.CreateGlobalStringPtr(F->getName())); 1420 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow); 1421 return; 1422 } 1423 case DataFlowSanitizer::WK_Discard: { 1424 CS.setCalledFunction(F); 1425 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow); 1426 return; 1427 } 1428 case DataFlowSanitizer::WK_Functional: { 1429 CS.setCalledFunction(F); 1430 visitOperandShadowInst(*CS.getInstruction()); 1431 return; 1432 } 1433 case DataFlowSanitizer::WK_Custom: { 1434 // Don't try to handle invokes of custom functions, it's too complicated. 1435 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_ 1436 // wrapper. 1437 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 1438 FunctionType *FT = F->getFunctionType(); 1439 FunctionType *CustomFT = DFSF.DFS.getCustomFunctionType(FT); 1440 std::string CustomFName = "__dfsw_"; 1441 CustomFName += F->getName(); 1442 Constant *CustomF = 1443 DFSF.DFS.Mod->getOrInsertFunction(CustomFName, CustomFT); 1444 if (Function *CustomFn = dyn_cast<Function>(CustomF)) { 1445 CustomFn->copyAttributesFrom(F); 1446 1447 // Custom functions returning non-void will write to the return label. 1448 if (!FT->getReturnType()->isVoidTy()) { 1449 CustomFn->removeAttributes(AttributeSet::FunctionIndex, 1450 DFSF.DFS.ReadOnlyNoneAttrs); 1451 } 1452 } 1453 1454 std::vector<Value *> Args; 1455 1456 CallSite::arg_iterator i = CS.arg_begin(); 1457 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) { 1458 Type *T = (*i)->getType(); 1459 FunctionType *ParamFT; 1460 if (isa<PointerType>(T) && 1461 (ParamFT = dyn_cast<FunctionType>( 1462 cast<PointerType>(T)->getElementType()))) { 1463 std::string TName = "dfst"; 1464 TName += utostr(FT->getNumParams() - n); 1465 TName += "$"; 1466 TName += F->getName(); 1467 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName); 1468 Args.push_back(T); 1469 Args.push_back( 1470 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx))); 1471 } else { 1472 Args.push_back(*i); 1473 } 1474 } 1475 1476 i = CS.arg_begin(); 1477 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1478 Args.push_back(DFSF.getShadow(*i)); 1479 1480 if (FT->isVarArg()) { 1481 auto *LabelVATy = ArrayType::get(DFSF.DFS.ShadowTy, 1482 CS.arg_size() - FT->getNumParams()); 1483 auto *LabelVAAlloca = new AllocaInst( 1484 LabelVATy, "labelva", &DFSF.F->getEntryBlock().front()); 1485 1486 for (unsigned n = 0; i != CS.arg_end(); ++i, ++n) { 1487 auto LabelVAPtr = IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, n); 1488 IRB.CreateStore(DFSF.getShadow(*i), LabelVAPtr); 1489 } 1490 1491 Args.push_back(IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, 0)); 1492 } 1493 1494 if (!FT->getReturnType()->isVoidTy()) { 1495 if (!DFSF.LabelReturnAlloca) { 1496 DFSF.LabelReturnAlloca = 1497 new AllocaInst(DFSF.DFS.ShadowTy, "labelreturn", 1498 &DFSF.F->getEntryBlock().front()); 1499 } 1500 Args.push_back(DFSF.LabelReturnAlloca); 1501 } 1502 1503 for (i = CS.arg_begin() + FT->getNumParams(); i != CS.arg_end(); ++i) 1504 Args.push_back(*i); 1505 1506 CallInst *CustomCI = IRB.CreateCall(CustomF, Args); 1507 CustomCI->setCallingConv(CI->getCallingConv()); 1508 CustomCI->setAttributes(CI->getAttributes()); 1509 1510 if (!FT->getReturnType()->isVoidTy()) { 1511 LoadInst *LabelLoad = IRB.CreateLoad(DFSF.LabelReturnAlloca); 1512 DFSF.setShadow(CustomCI, LabelLoad); 1513 } 1514 1515 CI->replaceAllUsesWith(CustomCI); 1516 CI->eraseFromParent(); 1517 return; 1518 } 1519 break; 1520 } 1521 } 1522 } 1523 1524 FunctionType *FT = cast<FunctionType>( 1525 CS.getCalledValue()->getType()->getPointerElementType()); 1526 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) { 1527 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) { 1528 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)), 1529 DFSF.getArgTLS(i, CS.getInstruction())); 1530 } 1531 } 1532 1533 Instruction *Next = nullptr; 1534 if (!CS.getType()->isVoidTy()) { 1535 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1536 if (II->getNormalDest()->getSinglePredecessor()) { 1537 Next = &II->getNormalDest()->front(); 1538 } else { 1539 BasicBlock *NewBB = 1540 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT); 1541 Next = &NewBB->front(); 1542 } 1543 } else { 1544 assert(CS->getIterator() != CS->getParent()->end()); 1545 Next = CS->getNextNode(); 1546 } 1547 1548 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) { 1549 IRBuilder<> NextIRB(Next); 1550 LoadInst *LI = NextIRB.CreateLoad(DFSF.getRetvalTLS()); 1551 DFSF.SkipInsts.insert(LI); 1552 DFSF.setShadow(CS.getInstruction(), LI); 1553 DFSF.NonZeroChecks.push_back(LI); 1554 } 1555 } 1556 1557 // Do all instrumentation for IA_Args down here to defer tampering with the 1558 // CFG in a way that SplitEdge may be able to detect. 1559 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) { 1560 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT); 1561 Value *Func = 1562 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT)); 1563 std::vector<Value *> Args; 1564 1565 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 1566 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1567 Args.push_back(*i); 1568 1569 i = CS.arg_begin(); 1570 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) 1571 Args.push_back(DFSF.getShadow(*i)); 1572 1573 if (FT->isVarArg()) { 1574 unsigned VarArgSize = CS.arg_size() - FT->getNumParams(); 1575 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize); 1576 AllocaInst *VarArgShadow = 1577 new AllocaInst(VarArgArrayTy, "", &DFSF.F->getEntryBlock().front()); 1578 Args.push_back(IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, 0)); 1579 for (unsigned n = 0; i != e; ++i, ++n) { 1580 IRB.CreateStore( 1581 DFSF.getShadow(*i), 1582 IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, n)); 1583 Args.push_back(*i); 1584 } 1585 } 1586 1587 CallSite NewCS; 1588 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1589 NewCS = IRB.CreateInvoke(Func, II->getNormalDest(), II->getUnwindDest(), 1590 Args); 1591 } else { 1592 NewCS = IRB.CreateCall(Func, Args); 1593 } 1594 NewCS.setCallingConv(CS.getCallingConv()); 1595 NewCS.setAttributes(CS.getAttributes().removeAttributes( 1596 *DFSF.DFS.Ctx, AttributeSet::ReturnIndex, 1597 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType()))); 1598 1599 if (Next) { 1600 ExtractValueInst *ExVal = 1601 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next); 1602 DFSF.SkipInsts.insert(ExVal); 1603 ExtractValueInst *ExShadow = 1604 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next); 1605 DFSF.SkipInsts.insert(ExShadow); 1606 DFSF.setShadow(ExVal, ExShadow); 1607 DFSF.NonZeroChecks.push_back(ExShadow); 1608 1609 CS.getInstruction()->replaceAllUsesWith(ExVal); 1610 } 1611 1612 CS.getInstruction()->eraseFromParent(); 1613 } 1614 } 1615 1616 void DFSanVisitor::visitPHINode(PHINode &PN) { 1617 PHINode *ShadowPN = 1618 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN); 1619 1620 // Give the shadow phi node valid predecessors to fool SplitEdge into working. 1621 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy); 1622 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e; 1623 ++i) { 1624 ShadowPN->addIncoming(UndefShadow, *i); 1625 } 1626 1627 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN)); 1628 DFSF.setShadow(&PN, ShadowPN); 1629 } 1630