1 //===-- AddressSanitizer.cpp - memory error detector ------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file is a part of AddressSanitizer, an address sanity checker.
11 // Details of the algorithm:
12 // http://code.google.com/p/address-sanitizer/wiki/AddressSanitizerAlgorithm
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "asan"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/OwningPtr.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallString.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/StringExtras.h"
24 #include "llvm/Function.h"
25 #include "llvm/InlineAsm.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/LLVMContext.h"
28 #include "llvm/Module.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/DataTypes.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/IRBuilder.h"
33 #include "llvm/Support/MemoryBuffer.h"
34 #include "llvm/Support/Regex.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Support/system_error.h"
37 #include "llvm/Target/TargetData.h"
38 #include "llvm/Target/TargetMachine.h"
39 #include "llvm/Transforms/Instrumentation.h"
40 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
41 #include "llvm/Transforms/Utils/ModuleUtils.h"
42 #include "llvm/Type.h"
49 static const uint64_t kDefaultShadowScale = 3;
50 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
51 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
53 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
54 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
55 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
57 static const char *kAsanModuleCtorName = "asan.module_ctor";
58 static const char *kAsanReportErrorTemplate = "__asan_report_";
59 static const char *kAsanRegisterGlobalsName = "__asan_register_globals";
60 static const char *kAsanInitName = "__asan_init";
61 static const char *kAsanMappingOffsetName = "__asan_mapping_offset";
62 static const char *kAsanMappingScaleName = "__asan_mapping_scale";
63 static const char *kAsanStackMallocName = "__asan_stack_malloc";
64 static const char *kAsanStackFreeName = "__asan_stack_free";
66 static const int kAsanStackLeftRedzoneMagic = 0xf1;
67 static const int kAsanStackMidRedzoneMagic = 0xf2;
68 static const int kAsanStackRightRedzoneMagic = 0xf3;
69 static const int kAsanStackPartialRedzoneMagic = 0xf4;
71 // Command-line flags.
73 // This flag may need to be replaced with -f[no-]asan-reads.
74 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
75 cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
76 static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
77 cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
78 // This flag may need to be replaced with -f[no]asan-stack.
79 static cl::opt<bool> ClStack("asan-stack",
80 cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
81 // This flag may need to be replaced with -f[no]asan-use-after-return.
82 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
83 cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
84 // This flag may need to be replaced with -f[no]asan-globals.
85 static cl::opt<bool> ClGlobals("asan-globals",
86 cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
87 static cl::opt<bool> ClMemIntrin("asan-memintrin",
88 cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
89 // This flag may need to be replaced with -fasan-blacklist.
90 static cl::opt<std::string> ClBlackListFile("asan-blacklist",
91 cl::desc("File containing the list of functions to ignore "
92 "during instrumentation"), cl::Hidden);
93 static cl::opt<bool> ClUseCall("asan-use-call",
94 cl::desc("Use function call to generate a crash"), cl::Hidden,
97 // These flags allow to change the shadow mapping.
98 // The shadow mapping looks like
99 // Shadow = (Mem >> scale) + (1 << offset_log)
100 static cl::opt<int> ClMappingScale("asan-mapping-scale",
101 cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
102 static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
103 cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
105 // Optimization flags. Not user visible, used mostly for testing
106 // and benchmarking the tool.
107 static cl::opt<bool> ClOpt("asan-opt",
108 cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
109 static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
110 cl::desc("Instrument the same temp just once"), cl::Hidden,
112 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
113 cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
116 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
118 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
119 cl::Hidden, cl::init(0));
120 static cl::opt<std::string> ClDebugFunc("asan-debug-func",
121 cl::Hidden, cl::desc("Debug func"));
122 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
123 cl::Hidden, cl::init(-1));
124 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
125 cl::Hidden, cl::init(-1));
129 // Blacklisted functions are not instrumented.
130 // The blacklist file contains one or more lines like this:
132 // fun:FunctionWildCard
134 // This is similar to the "ignore" feature of ThreadSanitizer.
135 // http://code.google.com/p/data-race-test/wiki/ThreadSanitizerIgnores
138 BlackList(const std::string &Path);
139 bool isIn(const Function &F);
144 /// AddressSanitizer: instrument the code in module to find memory bugs.
145 struct AddressSanitizer : public ModulePass {
147 void instrumentMop(Instruction *I);
148 void instrumentAddress(Instruction *OrigIns, IRBuilder<> &IRB,
149 Value *Addr, uint32_t TypeSize, bool IsWrite);
150 Instruction *generateCrashCode(IRBuilder<> &IRB, Value *Addr,
151 bool IsWrite, uint32_t TypeSize);
152 bool instrumentMemIntrinsic(MemIntrinsic *MI);
153 void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
155 Instruction *InsertBefore, bool IsWrite);
156 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
157 bool handleFunction(Module &M, Function &F);
158 bool poisonStackInFunction(Module &M, Function &F);
159 virtual bool runOnModule(Module &M);
160 bool insertGlobalRedzones(Module &M);
161 BranchInst *splitBlockAndInsertIfThen(Instruction *SplitBefore, Value *Cmp);
162 static char ID; // Pass identification, replacement for typeid
166 uint64_t getAllocaSizeInBytes(AllocaInst *AI) {
167 Type *Ty = AI->getAllocatedType();
168 uint64_t SizeInBytes = TD->getTypeStoreSizeInBits(Ty) / 8;
171 uint64_t getAlignedSize(uint64_t SizeInBytes) {
172 return ((SizeInBytes + RedzoneSize - 1)
173 / RedzoneSize) * RedzoneSize;
175 uint64_t getAlignedAllocaSize(AllocaInst *AI) {
176 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
177 return getAlignedSize(SizeInBytes);
180 void PoisonStack(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB,
181 Value *ShadowBase, bool DoPoison);
182 bool LooksLikeCodeInBug11395(Instruction *I);
184 Module *CurrentModule;
187 uint64_t MappingOffset;
193 Function *AsanCtorFunction;
194 Function *AsanInitFunction;
195 Instruction *CtorInsertBefore;
196 OwningPtr<BlackList> BL;
200 char AddressSanitizer::ID = 0;
201 INITIALIZE_PASS(AddressSanitizer, "asan",
202 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
204 AddressSanitizer::AddressSanitizer() : ModulePass(ID) { }
205 ModulePass *llvm::createAddressSanitizerPass() {
206 return new AddressSanitizer();
209 // Create a constant for Str so that we can pass it to the run-time lib.
210 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
211 Constant *StrConst = ConstantArray::get(M.getContext(), Str);
212 return new GlobalVariable(M, StrConst->getType(), true,
213 GlobalValue::PrivateLinkage, StrConst, "");
216 // Split the basic block and insert an if-then code.
228 // Returns the NewBasicBlock's terminator.
229 BranchInst *AddressSanitizer::splitBlockAndInsertIfThen(
230 Instruction *SplitBefore, Value *Cmp) {
231 BasicBlock *Head = SplitBefore->getParent();
232 BasicBlock *Tail = Head->splitBasicBlock(SplitBefore);
233 TerminatorInst *HeadOldTerm = Head->getTerminator();
234 BasicBlock *NewBasicBlock =
235 BasicBlock::Create(*C, "", Head->getParent());
236 BranchInst *HeadNewTerm = BranchInst::Create(/*ifTrue*/NewBasicBlock,
239 ReplaceInstWithInst(HeadOldTerm, HeadNewTerm);
241 BranchInst *CheckTerm = BranchInst::Create(Tail, NewBasicBlock);
245 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
247 Shadow = IRB.CreateLShr(Shadow, MappingScale);
248 if (MappingOffset == 0)
250 // (Shadow >> scale) | offset
251 return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy,
255 void AddressSanitizer::instrumentMemIntrinsicParam(Instruction *OrigIns,
256 Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
257 // Check the first byte.
259 IRBuilder<> IRB(InsertBefore);
260 instrumentAddress(OrigIns, IRB, Addr, 8, IsWrite);
262 // Check the last byte.
264 IRBuilder<> IRB(InsertBefore);
265 Value *SizeMinusOne = IRB.CreateSub(
266 Size, ConstantInt::get(Size->getType(), 1));
267 SizeMinusOne = IRB.CreateIntCast(SizeMinusOne, IntptrTy, false);
268 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
269 Value *AddrPlusSizeMinisOne = IRB.CreateAdd(AddrLong, SizeMinusOne);
270 instrumentAddress(OrigIns, IRB, AddrPlusSizeMinisOne, 8, IsWrite);
274 // Instrument memset/memmove/memcpy
275 bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
276 Value *Dst = MI->getDest();
277 MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
278 Value *Src = MemTran ? MemTran->getSource() : NULL;
279 Value *Length = MI->getLength();
281 Constant *ConstLength = dyn_cast<Constant>(Length);
282 Instruction *InsertBefore = MI;
284 if (ConstLength->isNullValue()) return false;
286 // The size is not a constant so it could be zero -- check at run-time.
287 IRBuilder<> IRB(InsertBefore);
289 Value *Cmp = IRB.CreateICmpNE(Length,
290 Constant::getNullValue(Length->getType()));
291 InsertBefore = splitBlockAndInsertIfThen(InsertBefore, Cmp);
294 instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
296 instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
300 static Value *getLDSTOperand(Instruction *I) {
301 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
302 return LI->getPointerOperand();
304 return cast<StoreInst>(*I).getPointerOperand();
307 void AddressSanitizer::instrumentMop(Instruction *I) {
308 int IsWrite = isa<StoreInst>(*I);
309 Value *Addr = getLDSTOperand(I);
310 if (ClOpt && ClOptGlobals && isa<GlobalVariable>(Addr)) {
311 // We are accessing a global scalar variable. Nothing to catch here.
314 Type *OrigPtrTy = Addr->getType();
315 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
317 assert(OrigTy->isSized());
318 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
320 if (TypeSize != 8 && TypeSize != 16 &&
321 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
322 // Ignore all unusual sizes.
327 instrumentAddress(I, IRB, Addr, TypeSize, IsWrite);
330 Instruction *AddressSanitizer::generateCrashCode(
331 IRBuilder<> &IRB, Value *Addr, bool IsWrite, uint32_t TypeSize) {
334 // Here we use a call instead of arch-specific asm to report an error.
335 // This is almost always slower (because the codegen needs to generate
336 // prologue/epilogue for otherwise leaf functions) and generates more code.
337 // This mode could be useful if we can not use SIGILL for some reason.
339 // IsWrite and TypeSize are encoded in the function name.
340 std::string FunctionName = std::string(kAsanReportErrorTemplate) +
341 (IsWrite ? "store" : "load") + itostr(TypeSize / 8);
342 Value *ReportWarningFunc = CurrentModule->getOrInsertFunction(
343 FunctionName, IRB.getVoidTy(), IntptrTy, NULL);
344 CallInst *Call = IRB.CreateCall(ReportWarningFunc, Addr);
345 Call->setDoesNotReturn();
349 uint32_t LogOfSizeInBytes = CountTrailingZeros_32(TypeSize / 8);
350 assert(8U * (1 << LogOfSizeInBytes) == TypeSize);
351 uint8_t TelltaleValue = IsWrite * 8 + LogOfSizeInBytes;
352 assert(TelltaleValue < 16);
354 // Move the failing address to %rax/%eax
355 FunctionType *Fn1Ty = FunctionType::get(
356 IRB.getVoidTy(), ArrayRef<Type*>(IntptrTy), false);
357 const char *MovStr = LongSize == 32
358 ? "mov $0, %eax" : "mov $0, %rax";
359 Value *AsmMov = InlineAsm::get(
360 Fn1Ty, StringRef(MovStr), StringRef("r"), true);
361 IRB.CreateCall(AsmMov, Addr);
363 // crash with ud2; could use int3, but it is less friendly to gdb.
364 // after ud2 put a 1-byte instruction that encodes the access type and size.
366 const char *TelltaleInsns[16] = {
385 std::string AsmStr = "ud2;";
386 AsmStr += TelltaleInsns[TelltaleValue];
387 Value *MyAsm = InlineAsm::get(FunctionType::get(Type::getVoidTy(*C), false),
388 StringRef(AsmStr), StringRef(""), true);
389 CallInst *AsmCall = IRB.CreateCall(MyAsm);
391 // This saves us one jump, but triggers a bug in RA (or somewhere else):
392 // while building 483.xalancbmk the compiler goes into infinite loop in
393 // llvm::SpillPlacement::iterate() / RAGreedy::growRegion
394 // AsmCall->setDoesNotReturn();
398 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
399 IRBuilder<> &IRB, Value *Addr,
400 uint32_t TypeSize, bool IsWrite) {
401 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
403 Type *ShadowTy = IntegerType::get(
404 *C, std::max(8U, TypeSize >> MappingScale));
405 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
406 Value *ShadowPtr = memToShadow(AddrLong, IRB);
407 Value *CmpVal = Constant::getNullValue(ShadowTy);
408 Value *ShadowValue = IRB.CreateLoad(
409 IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
411 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
413 Instruction *CheckTerm = splitBlockAndInsertIfThen(
414 cast<Instruction>(Cmp)->getNextNode(), Cmp);
415 IRBuilder<> IRB2(CheckTerm);
417 size_t Granularity = 1 << MappingScale;
418 if (TypeSize < 8 * Granularity) {
419 // Addr & (Granularity - 1)
420 Value *Lower3Bits = IRB2.CreateAnd(
421 AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
422 // (Addr & (Granularity - 1)) + size - 1
423 Value *LastAccessedByte = IRB2.CreateAdd(
424 Lower3Bits, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
425 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
426 LastAccessedByte = IRB2.CreateIntCast(
427 LastAccessedByte, IRB.getInt8Ty(), false);
428 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
429 Value *Cmp2 = IRB2.CreateICmpSGE(LastAccessedByte, ShadowValue);
431 CheckTerm = splitBlockAndInsertIfThen(CheckTerm, Cmp2);
434 IRBuilder<> IRB1(CheckTerm);
435 Instruction *Crash = generateCrashCode(IRB1, AddrLong, IsWrite, TypeSize);
436 Crash->setDebugLoc(OrigIns->getDebugLoc());
439 // This function replaces all global variables with new variables that have
440 // trailing redzones. It also creates a function that poisons
441 // redzones and inserts this function into llvm.global_ctors.
442 bool AddressSanitizer::insertGlobalRedzones(Module &M) {
443 SmallVector<GlobalVariable *, 16> GlobalsToChange;
445 for (Module::GlobalListType::iterator G = M.getGlobalList().begin(),
446 E = M.getGlobalList().end(); G != E; ++G) {
447 Type *Ty = cast<PointerType>(G->getType())->getElementType();
448 DEBUG(dbgs() << "GLOBAL: " << *G);
450 if (!Ty->isSized()) continue;
451 if (!G->hasInitializer()) continue;
452 // Touch only those globals that will not be defined in other modules.
453 // Don't handle ODR type linkages since other modules may be built w/o asan.
454 if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
455 G->getLinkage() != GlobalVariable::PrivateLinkage &&
456 G->getLinkage() != GlobalVariable::InternalLinkage)
458 // For now, just ignore this Alloca if the alignment is large.
459 if (G->getAlignment() > RedzoneSize) continue;
461 // Ignore all the globals with the names starting with "\01L_OBJC_".
462 // Many of those are put into the .cstring section. The linker compresses
463 // that section by removing the spare \0s after the string terminator, so
464 // our redzones get broken.
465 if ((G->getName().find("\01L_OBJC_") == 0) ||
466 (G->getName().find("\01l_OBJC_") == 0)) {
467 DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
471 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
472 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
474 if (G->hasSection()) {
475 StringRef Section(G->getSection());
476 if ((Section.find("__OBJC,") == 0) ||
477 (Section.find("__DATA, __objc_") == 0)) {
478 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
483 GlobalsToChange.push_back(G);
486 size_t n = GlobalsToChange.size();
487 if (n == 0) return false;
489 // A global is described by a structure
492 // size_t size_with_redzone;
494 // We initialize an array of such structures and pass it to a run-time call.
495 StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
496 IntptrTy, IntptrTy, NULL);
497 SmallVector<Constant *, 16> Initializers(n);
499 IRBuilder<> IRB(CtorInsertBefore);
501 for (size_t i = 0; i < n; i++) {
502 GlobalVariable *G = GlobalsToChange[i];
503 PointerType *PtrTy = cast<PointerType>(G->getType());
504 Type *Ty = PtrTy->getElementType();
505 uint64_t SizeInBytes = TD->getTypeStoreSizeInBits(Ty) / 8;
506 uint64_t RightRedzoneSize = RedzoneSize +
507 (RedzoneSize - (SizeInBytes % RedzoneSize));
508 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
510 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
511 Constant *NewInitializer = ConstantStruct::get(
512 NewTy, G->getInitializer(),
513 Constant::getNullValue(RightRedZoneTy), NULL);
515 GlobalVariable *Name = createPrivateGlobalForString(M, G->getName());
517 // Create a new global variable with enough space for a redzone.
518 GlobalVariable *NewGlobal = new GlobalVariable(
519 M, NewTy, G->isConstant(), G->getLinkage(),
520 NewInitializer, "", G, G->isThreadLocal());
521 NewGlobal->copyAttributesFrom(G);
522 NewGlobal->setAlignment(RedzoneSize);
525 Indices2[0] = IRB.getInt32(0);
526 Indices2[1] = IRB.getInt32(0);
528 G->replaceAllUsesWith(
529 ConstantExpr::getGetElementPtr(NewGlobal, Indices2, 2));
530 NewGlobal->takeName(G);
531 G->eraseFromParent();
533 Initializers[i] = ConstantStruct::get(
535 ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
536 ConstantInt::get(IntptrTy, SizeInBytes),
537 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
538 ConstantExpr::getPointerCast(Name, IntptrTy),
540 DEBUG(dbgs() << "NEW GLOBAL:\n" << *NewGlobal);
543 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
544 GlobalVariable *AllGlobals = new GlobalVariable(
545 M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage,
546 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
548 Function *AsanRegisterGlobals = cast<Function>(M.getOrInsertFunction(
549 kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
550 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
552 IRB.CreateCall2(AsanRegisterGlobals,
553 IRB.CreatePointerCast(AllGlobals, IntptrTy),
554 ConstantInt::get(IntptrTy, n));
561 bool AddressSanitizer::runOnModule(Module &M) {
562 // Initialize the private fields. No one has accessed them before.
563 TD = getAnalysisIfAvailable<TargetData>();
566 BL.reset(new BlackList(ClBlackListFile));
569 C = &(M.getContext());
570 LongSize = TD->getPointerSizeInBits();
571 IntptrTy = Type::getIntNTy(*C, LongSize);
572 IntptrPtrTy = PointerType::get(IntptrTy, 0);
574 AsanCtorFunction = Function::Create(
575 FunctionType::get(Type::getVoidTy(*C), false),
576 GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
577 BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
578 CtorInsertBefore = ReturnInst::Create(*C, AsanCtorBB);
580 // call __asan_init in the module ctor.
581 IRBuilder<> IRB(CtorInsertBefore);
582 AsanInitFunction = cast<Function>(
583 M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
584 AsanInitFunction->setLinkage(Function::ExternalLinkage);
585 IRB.CreateCall(AsanInitFunction);
587 MappingOffset = LongSize == 32
588 ? kDefaultShadowOffset32 : kDefaultShadowOffset64;
589 if (ClMappingOffsetLog >= 0) {
590 if (ClMappingOffsetLog == 0) {
594 MappingOffset = 1ULL << ClMappingOffsetLog;
597 MappingScale = kDefaultShadowScale;
598 if (ClMappingScale) {
599 MappingScale = ClMappingScale;
601 // Redzone used for stack and globals is at least 32 bytes.
602 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
603 RedzoneSize = std::max(32, (int)(1 << MappingScale));
608 Res |= insertGlobalRedzones(M);
610 // Tell the run-time the current values of mapping offset and scale.
611 GlobalValue *asan_mapping_offset =
612 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
613 ConstantInt::get(IntptrTy, MappingOffset),
614 kAsanMappingOffsetName);
615 GlobalValue *asan_mapping_scale =
616 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
617 ConstantInt::get(IntptrTy, MappingScale),
618 kAsanMappingScaleName);
619 // Read these globals, otherwise they may be optimized away.
620 IRB.CreateLoad(asan_mapping_scale, true);
621 IRB.CreateLoad(asan_mapping_offset, true);
624 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
625 if (F->isDeclaration()) continue;
626 Res |= handleFunction(M, *F);
629 appendToGlobalCtors(M, AsanCtorFunction, 1 /*high priority*/);
634 bool AddressSanitizer::handleFunction(Module &M, Function &F) {
635 if (BL->isIn(F)) return false;
636 if (&F == AsanCtorFunction) return false;
638 if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
640 // We want to instrument every address only once per basic block
641 // (unless there are calls between uses).
642 SmallSet<Value*, 16> TempsToInstrument;
643 SmallVector<Instruction*, 16> ToInstrument;
645 // Fill the set of memory operations to instrument.
646 for (Function::iterator FI = F.begin(), FE = F.end();
648 TempsToInstrument.clear();
649 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
651 if ((isa<LoadInst>(BI) && ClInstrumentReads) ||
652 (isa<StoreInst>(BI) && ClInstrumentWrites)) {
653 Value *Addr = getLDSTOperand(BI);
654 if (ClOpt && ClOptSameTemp) {
655 if (!TempsToInstrument.insert(Addr))
656 continue; // We've seen this temp in the current BB.
658 } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
661 if (isa<CallInst>(BI)) {
663 TempsToInstrument.clear();
667 ToInstrument.push_back(BI);
672 int NumInstrumented = 0;
673 for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
674 Instruction *Inst = ToInstrument[i];
675 if (ClDebugMin < 0 || ClDebugMax < 0 ||
676 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
677 if (isa<StoreInst>(Inst) || isa<LoadInst>(Inst))
680 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
687 bool ChangedStack = poisonStackInFunction(M, F);
689 // For each NSObject descendant having a +load method, this method is invoked
690 // by the ObjC runtime before any of the static constructors is called.
691 // Therefore we need to instrument such methods with a call to __asan_init
692 // at the beginning in order to initialize our runtime before any access to
693 // the shadow memory.
694 // We cannot just ignore these methods, because they may call other
695 // instrumented functions.
696 if (F.getName().find(" load]") != std::string::npos) {
697 IRBuilder<> IRB(F.begin()->begin());
698 IRB.CreateCall(AsanInitFunction);
701 return NumInstrumented > 0 || ChangedStack;
704 static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
705 if (ShadowRedzoneSize == 1) return PoisonByte;
706 if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
707 if (ShadowRedzoneSize == 4)
708 return (PoisonByte << 24) + (PoisonByte << 16) +
709 (PoisonByte << 8) + (PoisonByte);
710 assert(0 && "ShadowRedzoneSize is either 1, 2 or 4");
714 static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
717 size_t ShadowGranularity,
719 for (size_t i = 0; i < RedzoneSize;
720 i+= ShadowGranularity, Shadow++) {
721 if (i + ShadowGranularity <= Size) {
722 *Shadow = 0; // fully addressable
723 } else if (i >= Size) {
724 *Shadow = Magic; // unaddressable
726 *Shadow = Size - i; // first Size-i bytes are addressable
731 void AddressSanitizer::PoisonStack(const ArrayRef<AllocaInst*> &AllocaVec,
733 Value *ShadowBase, bool DoPoison) {
734 size_t ShadowRZSize = RedzoneSize >> MappingScale;
735 assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
736 Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
737 Type *RZPtrTy = PointerType::get(RZTy, 0);
739 Value *PoisonLeft = ConstantInt::get(RZTy,
740 ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
741 Value *PoisonMid = ConstantInt::get(RZTy,
742 ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
743 Value *PoisonRight = ConstantInt::get(RZTy,
744 ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
746 // poison the first red zone.
747 IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
749 // poison all other red zones.
750 uint64_t Pos = RedzoneSize;
751 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
752 AllocaInst *AI = AllocaVec[i];
753 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
754 uint64_t AlignedSize = getAlignedAllocaSize(AI);
755 assert(AlignedSize - SizeInBytes < RedzoneSize);
760 assert(ShadowBase->getType() == IntptrTy);
761 if (SizeInBytes < AlignedSize) {
762 // Poison the partial redzone at right
764 ShadowBase, ConstantInt::get(IntptrTy,
765 (Pos >> MappingScale) - ShadowRZSize));
766 size_t AddressableBytes = RedzoneSize - (AlignedSize - SizeInBytes);
769 PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
771 1ULL << MappingScale,
772 kAsanStackPartialRedzoneMagic);
774 Value *PartialPoison = ConstantInt::get(RZTy, Poison);
775 IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
778 // Poison the full redzone at right.
779 Ptr = IRB.CreateAdd(ShadowBase,
780 ConstantInt::get(IntptrTy, Pos >> MappingScale));
781 Value *Poison = i == AllocaVec.size() - 1 ? PoisonRight : PoisonMid;
782 IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
788 // Workaround for bug 11395: we don't want to instrument stack in functions
789 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
790 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
791 if (LongSize != 32) return false;
792 CallInst *CI = dyn_cast<CallInst>(I);
793 if (!CI || !CI->isInlineAsm()) return false;
794 if (CI->getNumArgOperands() <= 5) return false;
795 // We have inline assembly with quite a few arguments.
799 // Find all static Alloca instructions and put
800 // poisoned red zones around all of them.
801 // Then unpoison everything back before the function returns.
803 // Stack poisoning does not play well with exception handling.
804 // When an exception is thrown, we essentially bypass the code
805 // that unpoisones the stack. This is why the run-time library has
806 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
807 // stack in the interceptor. This however does not work inside the
808 // actual function which catches the exception. Most likely because the
809 // compiler hoists the load of the shadow value somewhere too high.
810 // This causes asan to report a non-existing bug on 453.povray.
811 // It sounds like an LLVM bug.
812 bool AddressSanitizer::poisonStackInFunction(Module &M, Function &F) {
813 if (!ClStack) return false;
814 SmallVector<AllocaInst*, 16> AllocaVec;
815 SmallVector<Instruction*, 8> RetVec;
816 uint64_t TotalSize = 0;
818 // Filter out Alloca instructions we want (and can) handle.
819 // Collect Ret instructions.
820 for (Function::iterator FI = F.begin(), FE = F.end();
822 BasicBlock &BB = *FI;
823 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
825 if (LooksLikeCodeInBug11395(BI)) return false;
826 if (isa<ReturnInst>(BI)) {
827 RetVec.push_back(BI);
831 AllocaInst *AI = dyn_cast<AllocaInst>(BI);
833 if (AI->isArrayAllocation()) continue;
834 if (!AI->isStaticAlloca()) continue;
835 if (!AI->getAllocatedType()->isSized()) continue;
836 if (AI->getAlignment() > RedzoneSize) continue;
837 AllocaVec.push_back(AI);
838 uint64_t AlignedSize = getAlignedAllocaSize(AI);
839 TotalSize += AlignedSize;
843 if (AllocaVec.empty()) return false;
845 uint64_t LocalStackSize = TotalSize + (AllocaVec.size() + 1) * RedzoneSize;
847 bool DoStackMalloc = ClUseAfterReturn
848 && LocalStackSize <= kMaxStackMallocSize;
850 Instruction *InsBefore = AllocaVec[0];
851 IRBuilder<> IRB(InsBefore);
854 Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
855 AllocaInst *MyAlloca =
856 new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
857 MyAlloca->setAlignment(RedzoneSize);
858 assert(MyAlloca->isStaticAlloca());
859 Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
860 Value *LocalStackBase = OrigStackBase;
863 Value *AsanStackMallocFunc = M.getOrInsertFunction(
864 kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL);
865 LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc,
866 ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
869 // This string will be parsed by the run-time (DescribeStackAddress).
870 SmallString<2048> StackDescriptionStorage;
871 raw_svector_ostream StackDescription(StackDescriptionStorage);
872 StackDescription << F.getName() << " " << AllocaVec.size() << " ";
874 uint64_t Pos = RedzoneSize;
875 // Replace Alloca instructions with base+offset.
876 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
877 AllocaInst *AI = AllocaVec[i];
878 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
879 StringRef Name = AI->getName();
880 StackDescription << Pos << " " << SizeInBytes << " "
881 << Name.size() << " " << Name << " ";
882 uint64_t AlignedSize = getAlignedAllocaSize(AI);
883 assert((AlignedSize % RedzoneSize) == 0);
884 AI->replaceAllUsesWith(
886 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
888 Pos += AlignedSize + RedzoneSize;
890 assert(Pos == LocalStackSize);
892 // Write the Magic value and the frame description constant to the redzone.
893 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
894 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
896 Value *BasePlus1 = IRB.CreateAdd(LocalStackBase,
897 ConstantInt::get(IntptrTy, LongSize/8));
898 BasePlus1 = IRB.CreateIntToPtr(BasePlus1, IntptrPtrTy);
899 Value *Description = IRB.CreatePointerCast(
900 createPrivateGlobalForString(M, StackDescription.str()),
902 IRB.CreateStore(Description, BasePlus1);
904 // Poison the stack redzones at the entry.
905 Value *ShadowBase = memToShadow(LocalStackBase, IRB);
906 PoisonStack(ArrayRef<AllocaInst*>(AllocaVec), IRB, ShadowBase, true);
908 Value *AsanStackFreeFunc = NULL;
910 AsanStackFreeFunc = M.getOrInsertFunction(
911 kAsanStackFreeName, IRB.getVoidTy(),
912 IntptrTy, IntptrTy, IntptrTy, NULL);
915 // Unpoison the stack before all ret instructions.
916 for (size_t i = 0, n = RetVec.size(); i < n; i++) {
917 Instruction *Ret = RetVec[i];
918 IRBuilder<> IRBRet(Ret);
920 // Mark the current frame as retired.
921 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
923 // Unpoison the stack.
924 PoisonStack(ArrayRef<AllocaInst*>(AllocaVec), IRBRet, ShadowBase, false);
927 IRBRet.CreateCall3(AsanStackFreeFunc, LocalStackBase,
928 ConstantInt::get(IntptrTy, LocalStackSize),
940 BlackList::BlackList(const std::string &Path) {
942 const char *kFunPrefix = "fun:";
943 if (!ClBlackListFile.size()) return;
946 OwningPtr<MemoryBuffer> File;
947 if (error_code EC = MemoryBuffer::getFile(ClBlackListFile.c_str(), File)) {
948 errs() << EC.message();
951 MemoryBuffer *Buff = File.take();
952 const char *Data = Buff->getBufferStart();
953 size_t DataLen = Buff->getBufferSize();
954 SmallVector<StringRef, 16> Lines;
955 SplitString(StringRef(Data, DataLen), Lines, "\n\r");
956 for (size_t i = 0, numLines = Lines.size(); i < numLines; i++) {
957 if (Lines[i].startswith(kFunPrefix)) {
958 std::string ThisFunc = Lines[i].substr(strlen(kFunPrefix));
962 // add ThisFunc replacing * with .*
963 for (size_t j = 0, n = ThisFunc.size(); j < n; j++) {
964 if (ThisFunc[j] == '*')
971 Functions = new Regex(Fun);
975 bool BlackList::isIn(const Function &F) {
977 bool Res = Functions->match(F.getName());