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/Transforms/Instrumentation.h"
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DepthFirstIterator.h"
22 #include "llvm/ADT/OwningPtr.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallString.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/Triple.h"
28 #include "llvm/DIBuilder.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/InstVisitor.h"
38 #include "llvm/Support/CallSite.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/DataTypes.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Support/system_error.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/BlackList.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ModuleUtils.h"
53 static const uint64_t kDefaultShadowScale = 3;
54 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
55 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
56 static const uint64_t kDefaultShort64bitShadowOffset = 0x7FFF8000; // < 2G.
57 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 41;
59 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
60 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
61 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
63 static const char *kAsanModuleCtorName = "asan.module_ctor";
64 static const char *kAsanModuleDtorName = "asan.module_dtor";
65 static const int kAsanCtorAndCtorPriority = 1;
66 static const char *kAsanReportErrorTemplate = "__asan_report_";
67 static const char *kAsanReportLoadN = "__asan_report_load_n";
68 static const char *kAsanReportStoreN = "__asan_report_store_n";
69 static const char *kAsanRegisterGlobalsName = "__asan_register_globals";
70 static const char *kAsanUnregisterGlobalsName = "__asan_unregister_globals";
71 static const char *kAsanPoisonGlobalsName = "__asan_before_dynamic_init";
72 static const char *kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init";
73 static const char *kAsanInitName = "__asan_init_v3";
74 static const char *kAsanHandleNoReturnName = "__asan_handle_no_return";
75 static const char *kAsanMappingOffsetName = "__asan_mapping_offset";
76 static const char *kAsanMappingScaleName = "__asan_mapping_scale";
77 static const char *kAsanStackMallocName = "__asan_stack_malloc";
78 static const char *kAsanStackFreeName = "__asan_stack_free";
79 static const char *kAsanGenPrefix = "__asan_gen_";
80 static const char *kAsanPoisonStackMemoryName = "__asan_poison_stack_memory";
81 static const char *kAsanUnpoisonStackMemoryName =
82 "__asan_unpoison_stack_memory";
84 static const int kAsanStackLeftRedzoneMagic = 0xf1;
85 static const int kAsanStackMidRedzoneMagic = 0xf2;
86 static const int kAsanStackRightRedzoneMagic = 0xf3;
87 static const int kAsanStackPartialRedzoneMagic = 0xf4;
89 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
90 static const size_t kNumberOfAccessSizes = 5;
92 // Command-line flags.
94 // This flag may need to be replaced with -f[no-]asan-reads.
95 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
96 cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
97 static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
98 cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
99 static cl::opt<bool> ClInstrumentAtomics("asan-instrument-atomics",
100 cl::desc("instrument atomic instructions (rmw, cmpxchg)"),
101 cl::Hidden, cl::init(true));
102 static cl::opt<bool> ClAlwaysSlowPath("asan-always-slow-path",
103 cl::desc("use instrumentation with slow path for all accesses"),
104 cl::Hidden, cl::init(false));
105 // This flag limits the number of instructions to be instrumented
106 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
107 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
109 static cl::opt<int> ClMaxInsnsToInstrumentPerBB("asan-max-ins-per-bb",
111 cl::desc("maximal number of instructions to instrument in any given BB"),
113 // This flag may need to be replaced with -f[no]asan-stack.
114 static cl::opt<bool> ClStack("asan-stack",
115 cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
116 // This flag may need to be replaced with -f[no]asan-use-after-return.
117 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
118 cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
119 // This flag may need to be replaced with -f[no]asan-globals.
120 static cl::opt<bool> ClGlobals("asan-globals",
121 cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
122 static cl::opt<bool> ClInitializers("asan-initialization-order",
123 cl::desc("Handle C++ initializer order"), cl::Hidden, cl::init(false));
124 static cl::opt<bool> ClMemIntrin("asan-memintrin",
125 cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
126 static cl::opt<bool> ClRealignStack("asan-realign-stack",
127 cl::desc("Realign stack to 32"), cl::Hidden, cl::init(true));
128 static cl::opt<std::string> ClBlacklistFile("asan-blacklist",
129 cl::desc("File containing the list of objects to ignore "
130 "during instrumentation"), cl::Hidden);
132 // These flags allow to change the shadow mapping.
133 // The shadow mapping looks like
134 // Shadow = (Mem >> scale) + (1 << offset_log)
135 static cl::opt<int> ClMappingScale("asan-mapping-scale",
136 cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
137 static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
138 cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
139 static cl::opt<bool> ClShort64BitOffset("asan-short-64bit-mapping-offset",
140 cl::desc("Use short immediate constant as the mapping offset for 64bit"),
141 cl::Hidden, cl::init(true));
143 // Optimization flags. Not user visible, used mostly for testing
144 // and benchmarking the tool.
145 static cl::opt<bool> ClOpt("asan-opt",
146 cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
147 static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
148 cl::desc("Instrument the same temp just once"), cl::Hidden,
150 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
151 cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
153 static cl::opt<bool> ClCheckLifetime("asan-check-lifetime",
154 cl::desc("Use llvm.lifetime intrinsics to insert extra checks"),
155 cl::Hidden, cl::init(false));
158 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
160 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
161 cl::Hidden, cl::init(0));
162 static cl::opt<std::string> ClDebugFunc("asan-debug-func",
163 cl::Hidden, cl::desc("Debug func"));
164 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
165 cl::Hidden, cl::init(-1));
166 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
167 cl::Hidden, cl::init(-1));
170 /// A set of dynamically initialized globals extracted from metadata.
171 class SetOfDynamicallyInitializedGlobals {
173 void Init(Module& M) {
174 // Clang generates metadata identifying all dynamically initialized globals.
175 NamedMDNode *DynamicGlobals =
176 M.getNamedMetadata("llvm.asan.dynamically_initialized_globals");
179 for (int i = 0, n = DynamicGlobals->getNumOperands(); i < n; ++i) {
180 MDNode *MDN = DynamicGlobals->getOperand(i);
181 assert(MDN->getNumOperands() == 1);
182 Value *VG = MDN->getOperand(0);
183 // The optimizer may optimize away a global entirely, in which case we
184 // cannot instrument access to it.
187 DynInitGlobals.insert(cast<GlobalVariable>(VG));
190 bool Contains(GlobalVariable *G) { return DynInitGlobals.count(G) != 0; }
192 SmallSet<GlobalValue*, 32> DynInitGlobals;
195 /// This struct defines the shadow mapping using the rule:
196 /// shadow = (mem >> Scale) ADD-or-OR Offset.
197 struct ShadowMapping {
203 static ShadowMapping getShadowMapping(const Module &M, int LongSize,
204 bool ZeroBaseShadow) {
205 llvm::Triple TargetTriple(M.getTargetTriple());
206 bool IsAndroid = TargetTriple.getEnvironment() == llvm::Triple::Android;
207 bool IsMacOSX = TargetTriple.getOS() == llvm::Triple::MacOSX;
208 bool IsPPC64 = TargetTriple.getArch() == llvm::Triple::ppc64;
209 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
211 ShadowMapping Mapping;
213 // OR-ing shadow offset if more efficient (at least on x86),
214 // but on ppc64 we have to use add since the shadow offset is not neccesary
215 // 1/8-th of the address space.
216 Mapping.OrShadowOffset = !IsPPC64 && !ClShort64BitOffset;
218 Mapping.Offset = (IsAndroid || ZeroBaseShadow) ? 0 :
219 (LongSize == 32 ? kDefaultShadowOffset32 :
220 IsPPC64 ? kPPC64_ShadowOffset64 : kDefaultShadowOffset64);
221 if (!ZeroBaseShadow && ClShort64BitOffset && IsX86_64 && !IsMacOSX) {
222 assert(LongSize == 64);
223 Mapping.Offset = kDefaultShort64bitShadowOffset;
225 if (!ZeroBaseShadow && ClMappingOffsetLog >= 0) {
226 // Zero offset log is the special case.
227 Mapping.Offset = (ClMappingOffsetLog == 0) ? 0 : 1ULL << ClMappingOffsetLog;
230 Mapping.Scale = kDefaultShadowScale;
231 if (ClMappingScale) {
232 Mapping.Scale = ClMappingScale;
238 static size_t RedzoneSizeForScale(int MappingScale) {
239 // Redzone used for stack and globals is at least 32 bytes.
240 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
241 return std::max(32U, 1U << MappingScale);
244 /// AddressSanitizer: instrument the code in module to find memory bugs.
245 struct AddressSanitizer : public FunctionPass {
246 AddressSanitizer(bool CheckInitOrder = true,
247 bool CheckUseAfterReturn = false,
248 bool CheckLifetime = false,
249 StringRef BlacklistFile = StringRef(),
250 bool ZeroBaseShadow = false)
252 CheckInitOrder(CheckInitOrder || ClInitializers),
253 CheckUseAfterReturn(CheckUseAfterReturn || ClUseAfterReturn),
254 CheckLifetime(CheckLifetime || ClCheckLifetime),
255 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
257 ZeroBaseShadow(ZeroBaseShadow) {}
258 virtual const char *getPassName() const {
259 return "AddressSanitizerFunctionPass";
261 void instrumentMop(Instruction *I);
262 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
263 Value *Addr, uint32_t TypeSize, bool IsWrite,
264 Value *SizeArgument);
265 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
266 Value *ShadowValue, uint32_t TypeSize);
267 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
268 bool IsWrite, size_t AccessSizeIndex,
269 Value *SizeArgument);
270 bool instrumentMemIntrinsic(MemIntrinsic *MI);
271 void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
273 Instruction *InsertBefore, bool IsWrite);
274 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
275 bool runOnFunction(Function &F);
276 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
277 void emitShadowMapping(Module &M, IRBuilder<> &IRB) const;
278 virtual bool doInitialization(Module &M);
279 static char ID; // Pass identification, replacement for typeid
282 void initializeCallbacks(Module &M);
284 bool ShouldInstrumentGlobal(GlobalVariable *G);
285 bool LooksLikeCodeInBug11395(Instruction *I);
286 void FindDynamicInitializers(Module &M);
289 bool CheckUseAfterReturn;
291 SmallString<64> BlacklistFile;
298 ShadowMapping Mapping;
299 Function *AsanCtorFunction;
300 Function *AsanInitFunction;
301 Function *AsanHandleNoReturnFunc;
302 OwningPtr<BlackList> BL;
303 // This array is indexed by AccessIsWrite and log2(AccessSize).
304 Function *AsanErrorCallback[2][kNumberOfAccessSizes];
305 // This array is indexed by AccessIsWrite.
306 Function *AsanErrorCallbackSized[2];
308 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
310 friend struct FunctionStackPoisoner;
313 class AddressSanitizerModule : public ModulePass {
315 AddressSanitizerModule(bool CheckInitOrder = true,
316 StringRef BlacklistFile = StringRef(),
317 bool ZeroBaseShadow = false)
319 CheckInitOrder(CheckInitOrder || ClInitializers),
320 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
322 ZeroBaseShadow(ZeroBaseShadow) {}
323 bool runOnModule(Module &M);
324 static char ID; // Pass identification, replacement for typeid
325 virtual const char *getPassName() const {
326 return "AddressSanitizerModule";
330 void initializeCallbacks(Module &M);
332 bool ShouldInstrumentGlobal(GlobalVariable *G);
333 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
334 size_t RedzoneSize() const {
335 return RedzoneSizeForScale(Mapping.Scale);
339 SmallString<64> BlacklistFile;
342 OwningPtr<BlackList> BL;
343 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
347 ShadowMapping Mapping;
348 Function *AsanPoisonGlobals;
349 Function *AsanUnpoisonGlobals;
350 Function *AsanRegisterGlobals;
351 Function *AsanUnregisterGlobals;
354 // Stack poisoning does not play well with exception handling.
355 // When an exception is thrown, we essentially bypass the code
356 // that unpoisones the stack. This is why the run-time library has
357 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
358 // stack in the interceptor. This however does not work inside the
359 // actual function which catches the exception. Most likely because the
360 // compiler hoists the load of the shadow value somewhere too high.
361 // This causes asan to report a non-existing bug on 453.povray.
362 // It sounds like an LLVM bug.
363 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
365 AddressSanitizer &ASan;
370 ShadowMapping Mapping;
372 SmallVector<AllocaInst*, 16> AllocaVec;
373 SmallVector<Instruction*, 8> RetVec;
374 uint64_t TotalStackSize;
375 unsigned StackAlignment;
377 Function *AsanStackMallocFunc, *AsanStackFreeFunc;
378 Function *AsanPoisonStackMemoryFunc, *AsanUnpoisonStackMemoryFunc;
380 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
381 struct AllocaPoisonCall {
382 IntrinsicInst *InsBefore;
386 SmallVector<AllocaPoisonCall, 8> AllocaPoisonCallVec;
388 // Maps Value to an AllocaInst from which the Value is originated.
389 typedef DenseMap<Value*, AllocaInst*> AllocaForValueMapTy;
390 AllocaForValueMapTy AllocaForValue;
392 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
393 : F(F), ASan(ASan), DIB(*F.getParent()), C(ASan.C),
394 IntptrTy(ASan.IntptrTy), IntptrPtrTy(PointerType::get(IntptrTy, 0)),
395 Mapping(ASan.Mapping),
396 TotalStackSize(0), StackAlignment(1 << Mapping.Scale) {}
398 bool runOnFunction() {
399 if (!ClStack) return false;
400 // Collect alloca, ret, lifetime instructions etc.
401 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
402 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
403 BasicBlock *BB = *DI;
406 if (AllocaVec.empty()) return false;
408 initializeCallbacks(*F.getParent());
418 // Finds all static Alloca instructions and puts
419 // poisoned red zones around all of them.
420 // Then unpoison everything back before the function returns.
423 // ----------------------- Visitors.
424 /// \brief Collect all Ret instructions.
425 void visitReturnInst(ReturnInst &RI) {
426 RetVec.push_back(&RI);
429 /// \brief Collect Alloca instructions we want (and can) handle.
430 void visitAllocaInst(AllocaInst &AI) {
431 if (!isInterestingAlloca(AI)) return;
433 StackAlignment = std::max(StackAlignment, AI.getAlignment());
434 AllocaVec.push_back(&AI);
435 uint64_t AlignedSize = getAlignedAllocaSize(&AI);
436 TotalStackSize += AlignedSize;
439 /// \brief Collect lifetime intrinsic calls to check for use-after-scope
441 void visitIntrinsicInst(IntrinsicInst &II) {
442 if (!ASan.CheckLifetime) return;
443 Intrinsic::ID ID = II.getIntrinsicID();
444 if (ID != Intrinsic::lifetime_start &&
445 ID != Intrinsic::lifetime_end)
447 // Found lifetime intrinsic, add ASan instrumentation if necessary.
448 ConstantInt *Size = dyn_cast<ConstantInt>(II.getArgOperand(0));
449 // If size argument is undefined, don't do anything.
450 if (Size->isMinusOne()) return;
451 // Check that size doesn't saturate uint64_t and can
452 // be stored in IntptrTy.
453 const uint64_t SizeValue = Size->getValue().getLimitedValue();
454 if (SizeValue == ~0ULL ||
455 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
457 // Find alloca instruction that corresponds to llvm.lifetime argument.
458 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1));
460 bool DoPoison = (ID == Intrinsic::lifetime_end);
461 AllocaPoisonCall APC = {&II, SizeValue, DoPoison};
462 AllocaPoisonCallVec.push_back(APC);
465 // ---------------------- Helpers.
466 void initializeCallbacks(Module &M);
468 // Check if we want (and can) handle this alloca.
469 bool isInterestingAlloca(AllocaInst &AI) {
470 return (!AI.isArrayAllocation() &&
471 AI.isStaticAlloca() &&
472 AI.getAllocatedType()->isSized());
475 size_t RedzoneSize() const {
476 return RedzoneSizeForScale(Mapping.Scale);
478 uint64_t getAllocaSizeInBytes(AllocaInst *AI) {
479 Type *Ty = AI->getAllocatedType();
480 uint64_t SizeInBytes = ASan.TD->getTypeAllocSize(Ty);
483 uint64_t getAlignedSize(uint64_t SizeInBytes) {
484 size_t RZ = RedzoneSize();
485 return ((SizeInBytes + RZ - 1) / RZ) * RZ;
487 uint64_t getAlignedAllocaSize(AllocaInst *AI) {
488 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
489 return getAlignedSize(SizeInBytes);
491 /// Finds alloca where the value comes from.
492 AllocaInst *findAllocaForValue(Value *V);
493 void poisonRedZones(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB,
494 Value *ShadowBase, bool DoPoison);
495 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> IRB, bool DoPoison);
500 char AddressSanitizer::ID = 0;
501 INITIALIZE_PASS(AddressSanitizer, "asan",
502 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
504 FunctionPass *llvm::createAddressSanitizerFunctionPass(
505 bool CheckInitOrder, bool CheckUseAfterReturn, bool CheckLifetime,
506 StringRef BlacklistFile, bool ZeroBaseShadow) {
507 return new AddressSanitizer(CheckInitOrder, CheckUseAfterReturn,
508 CheckLifetime, BlacklistFile, ZeroBaseShadow);
511 char AddressSanitizerModule::ID = 0;
512 INITIALIZE_PASS(AddressSanitizerModule, "asan-module",
513 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
514 "ModulePass", false, false)
515 ModulePass *llvm::createAddressSanitizerModulePass(
516 bool CheckInitOrder, StringRef BlacklistFile, bool ZeroBaseShadow) {
517 return new AddressSanitizerModule(CheckInitOrder, BlacklistFile,
521 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
522 size_t Res = CountTrailingZeros_32(TypeSize / 8);
523 assert(Res < kNumberOfAccessSizes);
527 // Create a constant for Str so that we can pass it to the run-time lib.
528 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
529 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
530 GlobalVariable *GV = new GlobalVariable(M, StrConst->getType(), true,
531 GlobalValue::PrivateLinkage, StrConst,
533 GV->setUnnamedAddr(true); // Ok to merge these.
534 GV->setAlignment(1); // Strings may not be merged w/o setting align 1.
538 static bool GlobalWasGeneratedByAsan(GlobalVariable *G) {
539 return G->getName().find(kAsanGenPrefix) == 0;
542 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
544 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
545 if (Mapping.Offset == 0)
547 // (Shadow >> scale) | offset
548 if (Mapping.OrShadowOffset)
549 return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
551 return IRB.CreateAdd(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
554 void AddressSanitizer::instrumentMemIntrinsicParam(
555 Instruction *OrigIns,
556 Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
557 IRBuilder<> IRB(InsertBefore);
558 if (Size->getType() != IntptrTy)
559 Size = IRB.CreateIntCast(Size, IntptrTy, false);
560 // Check the first byte.
561 instrumentAddress(OrigIns, InsertBefore, Addr, 8, IsWrite, Size);
562 // Check the last byte.
563 IRB.SetInsertPoint(InsertBefore);
564 Value *SizeMinusOne = IRB.CreateSub(Size, ConstantInt::get(IntptrTy, 1));
565 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
566 Value *AddrLast = IRB.CreateAdd(AddrLong, SizeMinusOne);
567 instrumentAddress(OrigIns, InsertBefore, AddrLast, 8, IsWrite, Size);
570 // Instrument memset/memmove/memcpy
571 bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
572 Value *Dst = MI->getDest();
573 MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
574 Value *Src = MemTran ? MemTran->getSource() : 0;
575 Value *Length = MI->getLength();
577 Constant *ConstLength = dyn_cast<Constant>(Length);
578 Instruction *InsertBefore = MI;
580 if (ConstLength->isNullValue()) return false;
582 // The size is not a constant so it could be zero -- check at run-time.
583 IRBuilder<> IRB(InsertBefore);
585 Value *Cmp = IRB.CreateICmpNE(Length,
586 Constant::getNullValue(Length->getType()));
587 InsertBefore = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
590 instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
592 instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
596 // If I is an interesting memory access, return the PointerOperand
597 // and set IsWrite. Otherwise return NULL.
598 static Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite) {
599 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
600 if (!ClInstrumentReads) return NULL;
602 return LI->getPointerOperand();
604 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
605 if (!ClInstrumentWrites) return NULL;
607 return SI->getPointerOperand();
609 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
610 if (!ClInstrumentAtomics) return NULL;
612 return RMW->getPointerOperand();
614 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
615 if (!ClInstrumentAtomics) return NULL;
617 return XCHG->getPointerOperand();
622 void AddressSanitizer::instrumentMop(Instruction *I) {
623 bool IsWrite = false;
624 Value *Addr = isInterestingMemoryAccess(I, &IsWrite);
626 if (ClOpt && ClOptGlobals) {
627 if (GlobalVariable *G = dyn_cast<GlobalVariable>(Addr)) {
628 // If initialization order checking is disabled, a simple access to a
629 // dynamically initialized global is always valid.
632 // If a global variable does not have dynamic initialization we don't
633 // have to instrument it. However, if a global does not have initailizer
634 // at all, we assume it has dynamic initializer (in other TU).
635 if (G->hasInitializer() && !DynamicallyInitializedGlobals.Contains(G))
640 Type *OrigPtrTy = Addr->getType();
641 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
643 assert(OrigTy->isSized());
644 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
646 assert((TypeSize % 8) == 0);
648 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check.
649 if (TypeSize == 8 || TypeSize == 16 ||
650 TypeSize == 32 || TypeSize == 64 || TypeSize == 128)
651 return instrumentAddress(I, I, Addr, TypeSize, IsWrite, 0);
652 // Instrument unusual size (but still multiple of 8).
653 // We can not do it with a single check, so we do 1-byte check for the first
654 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
655 // to report the actual access size.
657 Value *LastByte = IRB.CreateIntToPtr(
658 IRB.CreateAdd(IRB.CreatePointerCast(Addr, IntptrTy),
659 ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
661 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
662 instrumentAddress(I, I, Addr, 8, IsWrite, Size);
663 instrumentAddress(I, I, LastByte, 8, IsWrite, Size);
666 // Validate the result of Module::getOrInsertFunction called for an interface
667 // function of AddressSanitizer. If the instrumented module defines a function
668 // with the same name, their prototypes must match, otherwise
669 // getOrInsertFunction returns a bitcast.
670 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
671 if (isa<Function>(FuncOrBitcast)) return cast<Function>(FuncOrBitcast);
672 FuncOrBitcast->dump();
673 report_fatal_error("trying to redefine an AddressSanitizer "
674 "interface function");
677 Instruction *AddressSanitizer::generateCrashCode(
678 Instruction *InsertBefore, Value *Addr,
679 bool IsWrite, size_t AccessSizeIndex, Value *SizeArgument) {
680 IRBuilder<> IRB(InsertBefore);
681 CallInst *Call = SizeArgument
682 ? IRB.CreateCall2(AsanErrorCallbackSized[IsWrite], Addr, SizeArgument)
683 : IRB.CreateCall(AsanErrorCallback[IsWrite][AccessSizeIndex], Addr);
685 // We don't do Call->setDoesNotReturn() because the BB already has
686 // UnreachableInst at the end.
687 // This EmptyAsm is required to avoid callback merge.
688 IRB.CreateCall(EmptyAsm);
692 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
695 size_t Granularity = 1 << Mapping.Scale;
696 // Addr & (Granularity - 1)
697 Value *LastAccessedByte = IRB.CreateAnd(
698 AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
699 // (Addr & (Granularity - 1)) + size - 1
700 if (TypeSize / 8 > 1)
701 LastAccessedByte = IRB.CreateAdd(
702 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
703 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
704 LastAccessedByte = IRB.CreateIntCast(
705 LastAccessedByte, ShadowValue->getType(), false);
706 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
707 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
710 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
711 Instruction *InsertBefore,
712 Value *Addr, uint32_t TypeSize,
713 bool IsWrite, Value *SizeArgument) {
714 IRBuilder<> IRB(InsertBefore);
715 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
717 Type *ShadowTy = IntegerType::get(
718 *C, std::max(8U, TypeSize >> Mapping.Scale));
719 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
720 Value *ShadowPtr = memToShadow(AddrLong, IRB);
721 Value *CmpVal = Constant::getNullValue(ShadowTy);
722 Value *ShadowValue = IRB.CreateLoad(
723 IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
725 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
726 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
727 size_t Granularity = 1 << Mapping.Scale;
728 TerminatorInst *CrashTerm = 0;
730 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
731 TerminatorInst *CheckTerm =
732 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
733 assert(dyn_cast<BranchInst>(CheckTerm)->isUnconditional());
734 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
735 IRB.SetInsertPoint(CheckTerm);
736 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
737 BasicBlock *CrashBlock =
738 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
739 CrashTerm = new UnreachableInst(*C, CrashBlock);
740 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
741 ReplaceInstWithInst(CheckTerm, NewTerm);
743 CrashTerm = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), true);
746 Instruction *Crash = generateCrashCode(
747 CrashTerm, AddrLong, IsWrite, AccessSizeIndex, SizeArgument);
748 Crash->setDebugLoc(OrigIns->getDebugLoc());
751 void AddressSanitizerModule::createInitializerPoisonCalls(
752 Module &M, GlobalValue *ModuleName) {
753 // We do all of our poisoning and unpoisoning within _GLOBAL__I_a.
754 Function *GlobalInit = M.getFunction("_GLOBAL__I_a");
755 // If that function is not present, this TU contains no globals, or they have
756 // all been optimized away
760 // Set up the arguments to our poison/unpoison functions.
761 IRBuilder<> IRB(GlobalInit->begin()->getFirstInsertionPt());
763 // Add a call to poison all external globals before the given function starts.
764 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
765 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
767 // Add calls to unpoison all globals before each return instruction.
768 for (Function::iterator I = GlobalInit->begin(), E = GlobalInit->end();
770 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator())) {
771 CallInst::Create(AsanUnpoisonGlobals, "", RI);
776 bool AddressSanitizerModule::ShouldInstrumentGlobal(GlobalVariable *G) {
777 Type *Ty = cast<PointerType>(G->getType())->getElementType();
778 DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
780 if (BL->isIn(*G)) return false;
781 if (!Ty->isSized()) return false;
782 if (!G->hasInitializer()) return false;
783 if (GlobalWasGeneratedByAsan(G)) return false; // Our own global.
784 // Touch only those globals that will not be defined in other modules.
785 // Don't handle ODR type linkages since other modules may be built w/o asan.
786 if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
787 G->getLinkage() != GlobalVariable::PrivateLinkage &&
788 G->getLinkage() != GlobalVariable::InternalLinkage)
790 // Two problems with thread-locals:
791 // - The address of the main thread's copy can't be computed at link-time.
792 // - Need to poison all copies, not just the main thread's one.
793 if (G->isThreadLocal())
795 // For now, just ignore this Alloca if the alignment is large.
796 if (G->getAlignment() > RedzoneSize()) return false;
798 // Ignore all the globals with the names starting with "\01L_OBJC_".
799 // Many of those are put into the .cstring section. The linker compresses
800 // that section by removing the spare \0s after the string terminator, so
801 // our redzones get broken.
802 if ((G->getName().find("\01L_OBJC_") == 0) ||
803 (G->getName().find("\01l_OBJC_") == 0)) {
804 DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
808 if (G->hasSection()) {
809 StringRef Section(G->getSection());
810 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
811 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
813 if ((Section.find("__OBJC,") == 0) ||
814 (Section.find("__DATA, __objc_") == 0)) {
815 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
818 // See http://code.google.com/p/address-sanitizer/issues/detail?id=32
819 // Constant CFString instances are compiled in the following way:
820 // -- the string buffer is emitted into
821 // __TEXT,__cstring,cstring_literals
822 // -- the constant NSConstantString structure referencing that buffer
823 // is placed into __DATA,__cfstring
824 // Therefore there's no point in placing redzones into __DATA,__cfstring.
825 // Moreover, it causes the linker to crash on OS X 10.7
826 if (Section.find("__DATA,__cfstring") == 0) {
827 DEBUG(dbgs() << "Ignoring CFString: " << *G);
835 void AddressSanitizerModule::initializeCallbacks(Module &M) {
837 // Declare our poisoning and unpoisoning functions.
838 AsanPoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
839 kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, NULL));
840 AsanPoisonGlobals->setLinkage(Function::ExternalLinkage);
841 AsanUnpoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
842 kAsanUnpoisonGlobalsName, IRB.getVoidTy(), NULL));
843 AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage);
844 // Declare functions that register/unregister globals.
845 AsanRegisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
846 kAsanRegisterGlobalsName, IRB.getVoidTy(),
847 IntptrTy, IntptrTy, NULL));
848 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
849 AsanUnregisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
850 kAsanUnregisterGlobalsName,
851 IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
852 AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage);
855 // This function replaces all global variables with new variables that have
856 // trailing redzones. It also creates a function that poisons
857 // redzones and inserts this function into llvm.global_ctors.
858 bool AddressSanitizerModule::runOnModule(Module &M) {
859 if (!ClGlobals) return false;
860 TD = getAnalysisIfAvailable<DataLayout>();
863 BL.reset(new BlackList(BlacklistFile));
864 if (BL->isIn(M)) return false;
865 C = &(M.getContext());
866 int LongSize = TD->getPointerSizeInBits();
867 IntptrTy = Type::getIntNTy(*C, LongSize);
868 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
869 initializeCallbacks(M);
870 DynamicallyInitializedGlobals.Init(M);
872 SmallVector<GlobalVariable *, 16> GlobalsToChange;
874 for (Module::GlobalListType::iterator G = M.global_begin(),
875 E = M.global_end(); G != E; ++G) {
876 if (ShouldInstrumentGlobal(G))
877 GlobalsToChange.push_back(G);
880 size_t n = GlobalsToChange.size();
881 if (n == 0) return false;
883 // A global is described by a structure
886 // size_t size_with_redzone;
888 // const char *module_name;
889 // size_t has_dynamic_init;
890 // We initialize an array of such structures and pass it to a run-time call.
891 StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
893 IntptrTy, IntptrTy, NULL);
894 SmallVector<Constant *, 16> Initializers(n), DynamicInit;
897 Function *CtorFunc = M.getFunction(kAsanModuleCtorName);
899 IRBuilder<> IRB(CtorFunc->getEntryBlock().getTerminator());
901 bool HasDynamicallyInitializedGlobals = false;
903 GlobalVariable *ModuleName = createPrivateGlobalForString(
904 M, M.getModuleIdentifier());
905 // We shouldn't merge same module names, as this string serves as unique
906 // module ID in runtime.
907 ModuleName->setUnnamedAddr(false);
909 for (size_t i = 0; i < n; i++) {
910 static const uint64_t kMaxGlobalRedzone = 1 << 18;
911 GlobalVariable *G = GlobalsToChange[i];
912 PointerType *PtrTy = cast<PointerType>(G->getType());
913 Type *Ty = PtrTy->getElementType();
914 uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
915 uint64_t MinRZ = RedzoneSize();
916 // MinRZ <= RZ <= kMaxGlobalRedzone
917 // and trying to make RZ to be ~ 1/4 of SizeInBytes.
918 uint64_t RZ = std::max(MinRZ,
919 std::min(kMaxGlobalRedzone,
920 (SizeInBytes / MinRZ / 4) * MinRZ));
921 uint64_t RightRedzoneSize = RZ;
923 if (SizeInBytes % MinRZ)
924 RightRedzoneSize += MinRZ - (SizeInBytes % MinRZ);
925 assert(((RightRedzoneSize + SizeInBytes) % MinRZ) == 0);
926 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
927 // Determine whether this global should be poisoned in initialization.
928 bool GlobalHasDynamicInitializer =
929 DynamicallyInitializedGlobals.Contains(G);
930 // Don't check initialization order if this global is blacklisted.
931 GlobalHasDynamicInitializer &= !BL->isInInit(*G);
933 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
934 Constant *NewInitializer = ConstantStruct::get(
935 NewTy, G->getInitializer(),
936 Constant::getNullValue(RightRedZoneTy), NULL);
938 GlobalVariable *Name = createPrivateGlobalForString(M, G->getName());
940 // Create a new global variable with enough space for a redzone.
941 GlobalVariable *NewGlobal = new GlobalVariable(
942 M, NewTy, G->isConstant(), G->getLinkage(),
943 NewInitializer, "", G, G->getThreadLocalMode());
944 NewGlobal->copyAttributesFrom(G);
945 NewGlobal->setAlignment(MinRZ);
948 Indices2[0] = IRB.getInt32(0);
949 Indices2[1] = IRB.getInt32(0);
951 G->replaceAllUsesWith(
952 ConstantExpr::getGetElementPtr(NewGlobal, Indices2, true));
953 NewGlobal->takeName(G);
954 G->eraseFromParent();
956 Initializers[i] = ConstantStruct::get(
958 ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
959 ConstantInt::get(IntptrTy, SizeInBytes),
960 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
961 ConstantExpr::getPointerCast(Name, IntptrTy),
962 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
963 ConstantInt::get(IntptrTy, GlobalHasDynamicInitializer),
966 // Populate the first and last globals declared in this TU.
967 if (CheckInitOrder && GlobalHasDynamicInitializer)
968 HasDynamicallyInitializedGlobals = true;
970 DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
973 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
974 GlobalVariable *AllGlobals = new GlobalVariable(
975 M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage,
976 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
978 // Create calls for poisoning before initializers run and unpoisoning after.
979 if (CheckInitOrder && HasDynamicallyInitializedGlobals)
980 createInitializerPoisonCalls(M, ModuleName);
981 IRB.CreateCall2(AsanRegisterGlobals,
982 IRB.CreatePointerCast(AllGlobals, IntptrTy),
983 ConstantInt::get(IntptrTy, n));
985 // We also need to unregister globals at the end, e.g. when a shared library
987 Function *AsanDtorFunction = Function::Create(
988 FunctionType::get(Type::getVoidTy(*C), false),
989 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
990 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
991 IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB));
992 IRB_Dtor.CreateCall2(AsanUnregisterGlobals,
993 IRB.CreatePointerCast(AllGlobals, IntptrTy),
994 ConstantInt::get(IntptrTy, n));
995 appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndCtorPriority);
1001 void AddressSanitizer::initializeCallbacks(Module &M) {
1002 IRBuilder<> IRB(*C);
1003 // Create __asan_report* callbacks.
1004 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
1005 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
1006 AccessSizeIndex++) {
1007 // IsWrite and TypeSize are encoded in the function name.
1008 std::string FunctionName = std::string(kAsanReportErrorTemplate) +
1009 (AccessIsWrite ? "store" : "load") + itostr(1 << AccessSizeIndex);
1010 // If we are merging crash callbacks, they have two parameters.
1011 AsanErrorCallback[AccessIsWrite][AccessSizeIndex] =
1012 checkInterfaceFunction(M.getOrInsertFunction(
1013 FunctionName, IRB.getVoidTy(), IntptrTy, NULL));
1016 AsanErrorCallbackSized[0] = checkInterfaceFunction(M.getOrInsertFunction(
1017 kAsanReportLoadN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1018 AsanErrorCallbackSized[1] = checkInterfaceFunction(M.getOrInsertFunction(
1019 kAsanReportStoreN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1021 AsanHandleNoReturnFunc = checkInterfaceFunction(M.getOrInsertFunction(
1022 kAsanHandleNoReturnName, IRB.getVoidTy(), NULL));
1023 // We insert an empty inline asm after __asan_report* to avoid callback merge.
1024 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
1025 StringRef(""), StringRef(""),
1026 /*hasSideEffects=*/true);
1029 void AddressSanitizer::emitShadowMapping(Module &M, IRBuilder<> &IRB) const {
1030 // Tell the values of mapping offset and scale to the run-time.
1031 GlobalValue *asan_mapping_offset =
1032 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1033 ConstantInt::get(IntptrTy, Mapping.Offset),
1034 kAsanMappingOffsetName);
1035 // Read the global, otherwise it may be optimized away.
1036 IRB.CreateLoad(asan_mapping_offset, true);
1038 GlobalValue *asan_mapping_scale =
1039 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1040 ConstantInt::get(IntptrTy, Mapping.Scale),
1041 kAsanMappingScaleName);
1042 // Read the global, otherwise it may be optimized away.
1043 IRB.CreateLoad(asan_mapping_scale, true);
1047 bool AddressSanitizer::doInitialization(Module &M) {
1048 // Initialize the private fields. No one has accessed them before.
1049 TD = getAnalysisIfAvailable<DataLayout>();
1053 BL.reset(new BlackList(BlacklistFile));
1054 DynamicallyInitializedGlobals.Init(M);
1056 C = &(M.getContext());
1057 LongSize = TD->getPointerSizeInBits();
1058 IntptrTy = Type::getIntNTy(*C, LongSize);
1060 AsanCtorFunction = Function::Create(
1061 FunctionType::get(Type::getVoidTy(*C), false),
1062 GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
1063 BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
1064 // call __asan_init in the module ctor.
1065 IRBuilder<> IRB(ReturnInst::Create(*C, AsanCtorBB));
1066 AsanInitFunction = checkInterfaceFunction(
1067 M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
1068 AsanInitFunction->setLinkage(Function::ExternalLinkage);
1069 IRB.CreateCall(AsanInitFunction);
1071 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
1072 emitShadowMapping(M, IRB);
1074 appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority);
1078 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
1079 // For each NSObject descendant having a +load method, this method is invoked
1080 // by the ObjC runtime before any of the static constructors is called.
1081 // Therefore we need to instrument such methods with a call to __asan_init
1082 // at the beginning in order to initialize our runtime before any access to
1083 // the shadow memory.
1084 // We cannot just ignore these methods, because they may call other
1085 // instrumented functions.
1086 if (F.getName().find(" load]") != std::string::npos) {
1087 IRBuilder<> IRB(F.begin()->begin());
1088 IRB.CreateCall(AsanInitFunction);
1094 bool AddressSanitizer::runOnFunction(Function &F) {
1095 if (BL->isIn(F)) return false;
1096 if (&F == AsanCtorFunction) return false;
1097 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
1098 DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
1099 initializeCallbacks(*F.getParent());
1101 // If needed, insert __asan_init before checking for SanitizeAddress attr.
1102 maybeInsertAsanInitAtFunctionEntry(F);
1104 if (!F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1105 Attribute::SanitizeAddress))
1108 if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
1111 // We want to instrument every address only once per basic block (unless there
1112 // are calls between uses).
1113 SmallSet<Value*, 16> TempsToInstrument;
1114 SmallVector<Instruction*, 16> ToInstrument;
1115 SmallVector<Instruction*, 8> NoReturnCalls;
1118 // Fill the set of memory operations to instrument.
1119 for (Function::iterator FI = F.begin(), FE = F.end();
1121 TempsToInstrument.clear();
1122 int NumInsnsPerBB = 0;
1123 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1125 if (LooksLikeCodeInBug11395(BI)) return false;
1126 if (Value *Addr = isInterestingMemoryAccess(BI, &IsWrite)) {
1127 if (ClOpt && ClOptSameTemp) {
1128 if (!TempsToInstrument.insert(Addr))
1129 continue; // We've seen this temp in the current BB.
1131 } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
1136 // A call inside BB.
1137 TempsToInstrument.clear();
1138 if (CS.doesNotReturn())
1139 NoReturnCalls.push_back(CS.getInstruction());
1143 ToInstrument.push_back(BI);
1145 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB)
1151 int NumInstrumented = 0;
1152 for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
1153 Instruction *Inst = ToInstrument[i];
1154 if (ClDebugMin < 0 || ClDebugMax < 0 ||
1155 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
1156 if (isInterestingMemoryAccess(Inst, &IsWrite))
1157 instrumentMop(Inst);
1159 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
1164 FunctionStackPoisoner FSP(F, *this);
1165 bool ChangedStack = FSP.runOnFunction();
1167 // We must unpoison the stack before every NoReturn call (throw, _exit, etc).
1168 // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37
1169 for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) {
1170 Instruction *CI = NoReturnCalls[i];
1171 IRBuilder<> IRB(CI);
1172 IRB.CreateCall(AsanHandleNoReturnFunc);
1174 DEBUG(dbgs() << "ASAN done instrumenting:\n" << F << "\n");
1176 return NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty();
1179 static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
1180 if (ShadowRedzoneSize == 1) return PoisonByte;
1181 if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
1182 if (ShadowRedzoneSize == 4)
1183 return (PoisonByte << 24) + (PoisonByte << 16) +
1184 (PoisonByte << 8) + (PoisonByte);
1185 llvm_unreachable("ShadowRedzoneSize is either 1, 2 or 4");
1188 static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
1191 size_t ShadowGranularity,
1193 for (size_t i = 0; i < RZSize;
1194 i+= ShadowGranularity, Shadow++) {
1195 if (i + ShadowGranularity <= Size) {
1196 *Shadow = 0; // fully addressable
1197 } else if (i >= Size) {
1198 *Shadow = Magic; // unaddressable
1200 *Shadow = Size - i; // first Size-i bytes are addressable
1205 // Workaround for bug 11395: we don't want to instrument stack in functions
1206 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
1207 // FIXME: remove once the bug 11395 is fixed.
1208 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
1209 if (LongSize != 32) return false;
1210 CallInst *CI = dyn_cast<CallInst>(I);
1211 if (!CI || !CI->isInlineAsm()) return false;
1212 if (CI->getNumArgOperands() <= 5) return false;
1213 // We have inline assembly with quite a few arguments.
1217 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
1218 IRBuilder<> IRB(*C);
1219 AsanStackMallocFunc = checkInterfaceFunction(M.getOrInsertFunction(
1220 kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL));
1221 AsanStackFreeFunc = checkInterfaceFunction(M.getOrInsertFunction(
1222 kAsanStackFreeName, IRB.getVoidTy(),
1223 IntptrTy, IntptrTy, IntptrTy, NULL));
1224 AsanPoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1225 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1226 AsanUnpoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1227 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1230 void FunctionStackPoisoner::poisonRedZones(
1231 const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB, Value *ShadowBase,
1233 size_t ShadowRZSize = RedzoneSize() >> Mapping.Scale;
1234 assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
1235 Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
1236 Type *RZPtrTy = PointerType::get(RZTy, 0);
1238 Value *PoisonLeft = ConstantInt::get(RZTy,
1239 ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
1240 Value *PoisonMid = ConstantInt::get(RZTy,
1241 ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
1242 Value *PoisonRight = ConstantInt::get(RZTy,
1243 ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
1245 // poison the first red zone.
1246 IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
1248 // poison all other red zones.
1249 uint64_t Pos = RedzoneSize();
1250 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1251 AllocaInst *AI = AllocaVec[i];
1252 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1253 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1254 assert(AlignedSize - SizeInBytes < RedzoneSize());
1259 assert(ShadowBase->getType() == IntptrTy);
1260 if (SizeInBytes < AlignedSize) {
1261 // Poison the partial redzone at right
1262 Ptr = IRB.CreateAdd(
1263 ShadowBase, ConstantInt::get(IntptrTy,
1264 (Pos >> Mapping.Scale) - ShadowRZSize));
1265 size_t AddressableBytes = RedzoneSize() - (AlignedSize - SizeInBytes);
1266 uint32_t Poison = 0;
1268 PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
1270 1ULL << Mapping.Scale,
1271 kAsanStackPartialRedzoneMagic);
1273 Value *PartialPoison = ConstantInt::get(RZTy, Poison);
1274 IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1277 // Poison the full redzone at right.
1278 Ptr = IRB.CreateAdd(ShadowBase,
1279 ConstantInt::get(IntptrTy, Pos >> Mapping.Scale));
1280 bool LastAlloca = (i == AllocaVec.size() - 1);
1281 Value *Poison = LastAlloca ? PoisonRight : PoisonMid;
1282 IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1284 Pos += RedzoneSize();
1288 void FunctionStackPoisoner::poisonStack() {
1289 uint64_t LocalStackSize = TotalStackSize +
1290 (AllocaVec.size() + 1) * RedzoneSize();
1292 bool DoStackMalloc = ASan.CheckUseAfterReturn
1293 && LocalStackSize <= kMaxStackMallocSize;
1295 assert(AllocaVec.size() > 0);
1296 Instruction *InsBefore = AllocaVec[0];
1297 IRBuilder<> IRB(InsBefore);
1300 Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
1301 AllocaInst *MyAlloca =
1302 new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
1303 if (ClRealignStack && StackAlignment < RedzoneSize())
1304 StackAlignment = RedzoneSize();
1305 MyAlloca->setAlignment(StackAlignment);
1306 assert(MyAlloca->isStaticAlloca());
1307 Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
1308 Value *LocalStackBase = OrigStackBase;
1310 if (DoStackMalloc) {
1311 LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc,
1312 ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
1315 // This string will be parsed by the run-time (DescribeAddressIfStack).
1316 SmallString<2048> StackDescriptionStorage;
1317 raw_svector_ostream StackDescription(StackDescriptionStorage);
1318 StackDescription << AllocaVec.size() << " ";
1320 // Insert poison calls for lifetime intrinsics for alloca.
1321 bool HavePoisonedAllocas = false;
1322 for (size_t i = 0, n = AllocaPoisonCallVec.size(); i < n; i++) {
1323 const AllocaPoisonCall &APC = AllocaPoisonCallVec[i];
1324 IntrinsicInst *II = APC.InsBefore;
1325 AllocaInst *AI = findAllocaForValue(II->getArgOperand(1));
1327 IRBuilder<> IRB(II);
1328 poisonAlloca(AI, APC.Size, IRB, APC.DoPoison);
1329 HavePoisonedAllocas |= APC.DoPoison;
1332 uint64_t Pos = RedzoneSize();
1333 // Replace Alloca instructions with base+offset.
1334 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1335 AllocaInst *AI = AllocaVec[i];
1336 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1337 StringRef Name = AI->getName();
1338 StackDescription << Pos << " " << SizeInBytes << " "
1339 << Name.size() << " " << Name << " ";
1340 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1341 assert((AlignedSize % RedzoneSize()) == 0);
1342 Value *NewAllocaPtr = IRB.CreateIntToPtr(
1343 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
1345 replaceDbgDeclareForAlloca(AI, NewAllocaPtr, DIB);
1346 AI->replaceAllUsesWith(NewAllocaPtr);
1347 Pos += AlignedSize + RedzoneSize();
1349 assert(Pos == LocalStackSize);
1351 // The left-most redzone has enough space for at least 4 pointers.
1352 // Write the Magic value to redzone[0].
1353 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
1354 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
1356 // Write the frame description constant to redzone[1].
1357 Value *BasePlus1 = IRB.CreateIntToPtr(
1358 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, ASan.LongSize/8)),
1360 GlobalVariable *StackDescriptionGlobal =
1361 createPrivateGlobalForString(*F.getParent(), StackDescription.str());
1362 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal,
1364 IRB.CreateStore(Description, BasePlus1);
1365 // Write the PC to redzone[2].
1366 Value *BasePlus2 = IRB.CreateIntToPtr(
1367 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy,
1368 2 * ASan.LongSize/8)),
1370 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
1372 // Poison the stack redzones at the entry.
1373 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
1374 poisonRedZones(AllocaVec, IRB, ShadowBase, true);
1376 // Unpoison the stack before all ret instructions.
1377 for (size_t i = 0, n = RetVec.size(); i < n; i++) {
1378 Instruction *Ret = RetVec[i];
1379 IRBuilder<> IRBRet(Ret);
1380 // Mark the current frame as retired.
1381 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
1383 // Unpoison the stack.
1384 poisonRedZones(AllocaVec, IRBRet, ShadowBase, false);
1385 if (DoStackMalloc) {
1386 // In use-after-return mode, mark the whole stack frame unaddressable.
1387 IRBRet.CreateCall3(AsanStackFreeFunc, LocalStackBase,
1388 ConstantInt::get(IntptrTy, LocalStackSize),
1390 } else if (HavePoisonedAllocas) {
1391 // If we poisoned some allocas in llvm.lifetime analysis,
1392 // unpoison whole stack frame now.
1393 assert(LocalStackBase == OrigStackBase);
1394 poisonAlloca(LocalStackBase, LocalStackSize, IRBRet, false);
1398 // We are done. Remove the old unused alloca instructions.
1399 for (size_t i = 0, n = AllocaVec.size(); i < n; i++)
1400 AllocaVec[i]->eraseFromParent();
1403 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
1404 IRBuilder<> IRB, bool DoPoison) {
1405 // For now just insert the call to ASan runtime.
1406 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
1407 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
1408 IRB.CreateCall2(DoPoison ? AsanPoisonStackMemoryFunc
1409 : AsanUnpoisonStackMemoryFunc,
1413 // Handling llvm.lifetime intrinsics for a given %alloca:
1414 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
1415 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
1416 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
1417 // could be poisoned by previous llvm.lifetime.end instruction, as the
1418 // variable may go in and out of scope several times, e.g. in loops).
1419 // (3) if we poisoned at least one %alloca in a function,
1420 // unpoison the whole stack frame at function exit.
1422 AllocaInst *FunctionStackPoisoner::findAllocaForValue(Value *V) {
1423 if (AllocaInst *AI = dyn_cast<AllocaInst>(V))
1424 // We're intested only in allocas we can handle.
1425 return isInterestingAlloca(*AI) ? AI : 0;
1426 // See if we've already calculated (or started to calculate) alloca for a
1428 AllocaForValueMapTy::iterator I = AllocaForValue.find(V);
1429 if (I != AllocaForValue.end())
1431 // Store 0 while we're calculating alloca for value V to avoid
1432 // infinite recursion if the value references itself.
1433 AllocaForValue[V] = 0;
1434 AllocaInst *Res = 0;
1435 if (CastInst *CI = dyn_cast<CastInst>(V))
1436 Res = findAllocaForValue(CI->getOperand(0));
1437 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1438 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1439 Value *IncValue = PN->getIncomingValue(i);
1440 // Allow self-referencing phi-nodes.
1441 if (IncValue == PN) continue;
1442 AllocaInst *IncValueAI = findAllocaForValue(IncValue);
1443 // AI for incoming values should exist and should all be equal.
1444 if (IncValueAI == 0 || (Res != 0 && IncValueAI != Res))
1450 AllocaForValue[V] = Res;