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/Target/TargetMachine.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BlackList.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Transforms/Utils/ModuleUtils.h"
54 static const uint64_t kDefaultShadowScale = 3;
55 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
56 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
57 static const uint64_t kDefaultShort64bitShadowOffset = 0x7FFF8000; // < 2G.
58 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 41;
60 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
61 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
62 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
64 static const char *kAsanModuleCtorName = "asan.module_ctor";
65 static const char *kAsanModuleDtorName = "asan.module_dtor";
66 static const int kAsanCtorAndCtorPriority = 1;
67 static const char *kAsanReportErrorTemplate = "__asan_report_";
68 static const char *kAsanReportLoadN = "__asan_report_load_n";
69 static const char *kAsanReportStoreN = "__asan_report_store_n";
70 static const char *kAsanRegisterGlobalsName = "__asan_register_globals";
71 static const char *kAsanUnregisterGlobalsName = "__asan_unregister_globals";
72 static const char *kAsanPoisonGlobalsName = "__asan_before_dynamic_init";
73 static const char *kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init";
74 static const char *kAsanInitName = "__asan_init_v3";
75 static const char *kAsanHandleNoReturnName = "__asan_handle_no_return";
76 static const char *kAsanMappingOffsetName = "__asan_mapping_offset";
77 static const char *kAsanMappingScaleName = "__asan_mapping_scale";
78 static const char *kAsanStackMallocName = "__asan_stack_malloc";
79 static const char *kAsanStackFreeName = "__asan_stack_free";
80 static const char *kAsanGenPrefix = "__asan_gen_";
81 static const char *kAsanPoisonStackMemoryName = "__asan_poison_stack_memory";
82 static const char *kAsanUnpoisonStackMemoryName =
83 "__asan_unpoison_stack_memory";
85 static const int kAsanStackLeftRedzoneMagic = 0xf1;
86 static const int kAsanStackMidRedzoneMagic = 0xf2;
87 static const int kAsanStackRightRedzoneMagic = 0xf3;
88 static const int kAsanStackPartialRedzoneMagic = 0xf4;
90 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
91 static const size_t kNumberOfAccessSizes = 5;
93 // Command-line flags.
95 // This flag may need to be replaced with -f[no-]asan-reads.
96 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
97 cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
98 static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
99 cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
100 static cl::opt<bool> ClInstrumentAtomics("asan-instrument-atomics",
101 cl::desc("instrument atomic instructions (rmw, cmpxchg)"),
102 cl::Hidden, cl::init(true));
103 static cl::opt<bool> ClAlwaysSlowPath("asan-always-slow-path",
104 cl::desc("use instrumentation with slow path for all accesses"),
105 cl::Hidden, cl::init(false));
106 // This flag limits the number of instructions to be instrumented
107 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
108 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
110 static cl::opt<int> ClMaxInsnsToInstrumentPerBB("asan-max-ins-per-bb",
112 cl::desc("maximal number of instructions to instrument in any given BB"),
114 // This flag may need to be replaced with -f[no]asan-stack.
115 static cl::opt<bool> ClStack("asan-stack",
116 cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
117 // This flag may need to be replaced with -f[no]asan-use-after-return.
118 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
119 cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
120 // This flag may need to be replaced with -f[no]asan-globals.
121 static cl::opt<bool> ClGlobals("asan-globals",
122 cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
123 static cl::opt<bool> ClInitializers("asan-initialization-order",
124 cl::desc("Handle C++ initializer order"), cl::Hidden, cl::init(false));
125 static cl::opt<bool> ClMemIntrin("asan-memintrin",
126 cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
127 static cl::opt<bool> ClRealignStack("asan-realign-stack",
128 cl::desc("Realign stack to 32"), cl::Hidden, cl::init(true));
129 static cl::opt<std::string> ClBlacklistFile("asan-blacklist",
130 cl::desc("File containing the list of objects to ignore "
131 "during instrumentation"), cl::Hidden);
133 // These flags allow to change the shadow mapping.
134 // The shadow mapping looks like
135 // Shadow = (Mem >> scale) + (1 << offset_log)
136 static cl::opt<int> ClMappingScale("asan-mapping-scale",
137 cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
138 static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
139 cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
140 static cl::opt<bool> ClShort64BitOffset("asan-short-64bit-mapping-offset",
141 cl::desc("Use short immediate constant as the mapping offset for 64bit"),
142 cl::Hidden, cl::init(true));
144 // Optimization flags. Not user visible, used mostly for testing
145 // and benchmarking the tool.
146 static cl::opt<bool> ClOpt("asan-opt",
147 cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
148 static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
149 cl::desc("Instrument the same temp just once"), cl::Hidden,
151 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
152 cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
154 static cl::opt<bool> ClCheckLifetime("asan-check-lifetime",
155 cl::desc("Use llvm.lifetime intrinsics to insert extra checks"),
156 cl::Hidden, cl::init(false));
159 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
161 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
162 cl::Hidden, cl::init(0));
163 static cl::opt<std::string> ClDebugFunc("asan-debug-func",
164 cl::Hidden, cl::desc("Debug func"));
165 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
166 cl::Hidden, cl::init(-1));
167 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
168 cl::Hidden, cl::init(-1));
171 /// A set of dynamically initialized globals extracted from metadata.
172 class SetOfDynamicallyInitializedGlobals {
174 void Init(Module& M) {
175 // Clang generates metadata identifying all dynamically initialized globals.
176 NamedMDNode *DynamicGlobals =
177 M.getNamedMetadata("llvm.asan.dynamically_initialized_globals");
180 for (int i = 0, n = DynamicGlobals->getNumOperands(); i < n; ++i) {
181 MDNode *MDN = DynamicGlobals->getOperand(i);
182 assert(MDN->getNumOperands() == 1);
183 Value *VG = MDN->getOperand(0);
184 // The optimizer may optimize away a global entirely, in which case we
185 // cannot instrument access to it.
188 DynInitGlobals.insert(cast<GlobalVariable>(VG));
191 bool Contains(GlobalVariable *G) { return DynInitGlobals.count(G) != 0; }
193 SmallSet<GlobalValue*, 32> DynInitGlobals;
196 /// This struct defines the shadow mapping using the rule:
197 /// shadow = (mem >> Scale) ADD-or-OR Offset.
198 struct ShadowMapping {
204 static ShadowMapping getShadowMapping(const Module &M, int LongSize,
205 bool ZeroBaseShadow) {
206 llvm::Triple TargetTriple(M.getTargetTriple());
207 bool IsAndroid = TargetTriple.getEnvironment() == llvm::Triple::Android;
208 bool IsMacOSX = TargetTriple.getOS() == llvm::Triple::MacOSX;
209 bool IsPPC64 = TargetTriple.getArch() == llvm::Triple::ppc64;
210 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
212 ShadowMapping Mapping;
214 // OR-ing shadow offset if more efficient (at least on x86),
215 // but on ppc64 we have to use add since the shadow offset is not neccesary
216 // 1/8-th of the address space.
217 Mapping.OrShadowOffset = !IsPPC64 && !ClShort64BitOffset;
219 Mapping.Offset = (IsAndroid || ZeroBaseShadow) ? 0 :
220 (LongSize == 32 ? kDefaultShadowOffset32 :
221 IsPPC64 ? kPPC64_ShadowOffset64 : kDefaultShadowOffset64);
222 if (!ZeroBaseShadow && ClShort64BitOffset && IsX86_64 && !IsMacOSX) {
223 assert(LongSize == 64);
224 Mapping.Offset = kDefaultShort64bitShadowOffset;
226 if (!ZeroBaseShadow && ClMappingOffsetLog >= 0) {
227 // Zero offset log is the special case.
228 Mapping.Offset = (ClMappingOffsetLog == 0) ? 0 : 1ULL << ClMappingOffsetLog;
231 Mapping.Scale = kDefaultShadowScale;
232 if (ClMappingScale) {
233 Mapping.Scale = ClMappingScale;
239 static size_t RedzoneSizeForScale(int MappingScale) {
240 // Redzone used for stack and globals is at least 32 bytes.
241 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
242 return std::max(32U, 1U << MappingScale);
245 /// AddressSanitizer: instrument the code in module to find memory bugs.
246 struct AddressSanitizer : public FunctionPass {
247 AddressSanitizer(bool CheckInitOrder = true,
248 bool CheckUseAfterReturn = false,
249 bool CheckLifetime = false,
250 StringRef BlacklistFile = StringRef(),
251 bool ZeroBaseShadow = false)
253 CheckInitOrder(CheckInitOrder || ClInitializers),
254 CheckUseAfterReturn(CheckUseAfterReturn || ClUseAfterReturn),
255 CheckLifetime(CheckLifetime || ClCheckLifetime),
256 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
258 ZeroBaseShadow(ZeroBaseShadow) {}
259 virtual const char *getPassName() const {
260 return "AddressSanitizerFunctionPass";
262 void instrumentMop(Instruction *I);
263 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
264 Value *Addr, uint32_t TypeSize, bool IsWrite,
265 Value *SizeArgument);
266 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
267 Value *ShadowValue, uint32_t TypeSize);
268 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
269 bool IsWrite, size_t AccessSizeIndex,
270 Value *SizeArgument);
271 bool instrumentMemIntrinsic(MemIntrinsic *MI);
272 void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
274 Instruction *InsertBefore, bool IsWrite);
275 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
276 bool runOnFunction(Function &F);
277 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
278 void emitShadowMapping(Module &M, IRBuilder<> &IRB) const;
279 virtual bool doInitialization(Module &M);
280 static char ID; // Pass identification, replacement for typeid
283 void initializeCallbacks(Module &M);
285 bool ShouldInstrumentGlobal(GlobalVariable *G);
286 bool LooksLikeCodeInBug11395(Instruction *I);
287 void FindDynamicInitializers(Module &M);
290 bool CheckUseAfterReturn;
292 SmallString<64> BlacklistFile;
299 ShadowMapping Mapping;
300 Function *AsanCtorFunction;
301 Function *AsanInitFunction;
302 Function *AsanHandleNoReturnFunc;
303 OwningPtr<BlackList> BL;
304 // This array is indexed by AccessIsWrite and log2(AccessSize).
305 Function *AsanErrorCallback[2][kNumberOfAccessSizes];
306 // This array is indexed by AccessIsWrite.
307 Function *AsanErrorCallbackSized[2];
309 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
311 friend struct FunctionStackPoisoner;
314 class AddressSanitizerModule : public ModulePass {
316 AddressSanitizerModule(bool CheckInitOrder = true,
317 StringRef BlacklistFile = StringRef(),
318 bool ZeroBaseShadow = false)
320 CheckInitOrder(CheckInitOrder || ClInitializers),
321 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
323 ZeroBaseShadow(ZeroBaseShadow) {}
324 bool runOnModule(Module &M);
325 static char ID; // Pass identification, replacement for typeid
326 virtual const char *getPassName() const {
327 return "AddressSanitizerModule";
331 void initializeCallbacks(Module &M);
333 bool ShouldInstrumentGlobal(GlobalVariable *G);
334 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
335 size_t RedzoneSize() const {
336 return RedzoneSizeForScale(Mapping.Scale);
340 SmallString<64> BlacklistFile;
343 OwningPtr<BlackList> BL;
344 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
348 ShadowMapping Mapping;
349 Function *AsanPoisonGlobals;
350 Function *AsanUnpoisonGlobals;
351 Function *AsanRegisterGlobals;
352 Function *AsanUnregisterGlobals;
355 // Stack poisoning does not play well with exception handling.
356 // When an exception is thrown, we essentially bypass the code
357 // that unpoisones the stack. This is why the run-time library has
358 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
359 // stack in the interceptor. This however does not work inside the
360 // actual function which catches the exception. Most likely because the
361 // compiler hoists the load of the shadow value somewhere too high.
362 // This causes asan to report a non-existing bug on 453.povray.
363 // It sounds like an LLVM bug.
364 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
366 AddressSanitizer &ASan;
371 ShadowMapping Mapping;
373 SmallVector<AllocaInst*, 16> AllocaVec;
374 SmallVector<Instruction*, 8> RetVec;
375 uint64_t TotalStackSize;
376 unsigned StackAlignment;
378 Function *AsanStackMallocFunc, *AsanStackFreeFunc;
379 Function *AsanPoisonStackMemoryFunc, *AsanUnpoisonStackMemoryFunc;
381 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
382 struct AllocaPoisonCall {
383 IntrinsicInst *InsBefore;
387 SmallVector<AllocaPoisonCall, 8> AllocaPoisonCallVec;
389 // Maps Value to an AllocaInst from which the Value is originated.
390 typedef DenseMap<Value*, AllocaInst*> AllocaForValueMapTy;
391 AllocaForValueMapTy AllocaForValue;
393 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
394 : F(F), ASan(ASan), DIB(*F.getParent()), C(ASan.C),
395 IntptrTy(ASan.IntptrTy), IntptrPtrTy(PointerType::get(IntptrTy, 0)),
396 Mapping(ASan.Mapping),
397 TotalStackSize(0), StackAlignment(1 << Mapping.Scale) {}
399 bool runOnFunction() {
400 if (!ClStack) return false;
401 // Collect alloca, ret, lifetime instructions etc.
402 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
403 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
404 BasicBlock *BB = *DI;
407 if (AllocaVec.empty()) return false;
409 initializeCallbacks(*F.getParent());
419 // Finds all static Alloca instructions and puts
420 // poisoned red zones around all of them.
421 // Then unpoison everything back before the function returns.
424 // ----------------------- Visitors.
425 /// \brief Collect all Ret instructions.
426 void visitReturnInst(ReturnInst &RI) {
427 RetVec.push_back(&RI);
430 /// \brief Collect Alloca instructions we want (and can) handle.
431 void visitAllocaInst(AllocaInst &AI) {
432 if (!isInterestingAlloca(AI)) return;
434 StackAlignment = std::max(StackAlignment, AI.getAlignment());
435 AllocaVec.push_back(&AI);
436 uint64_t AlignedSize = getAlignedAllocaSize(&AI);
437 TotalStackSize += AlignedSize;
440 /// \brief Collect lifetime intrinsic calls to check for use-after-scope
442 void visitIntrinsicInst(IntrinsicInst &II) {
443 if (!ASan.CheckLifetime) return;
444 Intrinsic::ID ID = II.getIntrinsicID();
445 if (ID != Intrinsic::lifetime_start &&
446 ID != Intrinsic::lifetime_end)
448 // Found lifetime intrinsic, add ASan instrumentation if necessary.
449 ConstantInt *Size = dyn_cast<ConstantInt>(II.getArgOperand(0));
450 // If size argument is undefined, don't do anything.
451 if (Size->isMinusOne()) return;
452 // Check that size doesn't saturate uint64_t and can
453 // be stored in IntptrTy.
454 const uint64_t SizeValue = Size->getValue().getLimitedValue();
455 if (SizeValue == ~0ULL ||
456 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
458 // Find alloca instruction that corresponds to llvm.lifetime argument.
459 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1));
461 bool DoPoison = (ID == Intrinsic::lifetime_end);
462 AllocaPoisonCall APC = {&II, SizeValue, DoPoison};
463 AllocaPoisonCallVec.push_back(APC);
466 // ---------------------- Helpers.
467 void initializeCallbacks(Module &M);
469 // Check if we want (and can) handle this alloca.
470 bool isInterestingAlloca(AllocaInst &AI) {
471 return (!AI.isArrayAllocation() &&
472 AI.isStaticAlloca() &&
473 AI.getAllocatedType()->isSized());
476 size_t RedzoneSize() const {
477 return RedzoneSizeForScale(Mapping.Scale);
479 uint64_t getAllocaSizeInBytes(AllocaInst *AI) {
480 Type *Ty = AI->getAllocatedType();
481 uint64_t SizeInBytes = ASan.TD->getTypeAllocSize(Ty);
484 uint64_t getAlignedSize(uint64_t SizeInBytes) {
485 size_t RZ = RedzoneSize();
486 return ((SizeInBytes + RZ - 1) / RZ) * RZ;
488 uint64_t getAlignedAllocaSize(AllocaInst *AI) {
489 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
490 return getAlignedSize(SizeInBytes);
492 /// Finds alloca where the value comes from.
493 AllocaInst *findAllocaForValue(Value *V);
494 void poisonRedZones(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB,
495 Value *ShadowBase, bool DoPoison);
496 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> IRB, bool DoPoison);
501 char AddressSanitizer::ID = 0;
502 INITIALIZE_PASS(AddressSanitizer, "asan",
503 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
505 FunctionPass *llvm::createAddressSanitizerFunctionPass(
506 bool CheckInitOrder, bool CheckUseAfterReturn, bool CheckLifetime,
507 StringRef BlacklistFile, bool ZeroBaseShadow) {
508 return new AddressSanitizer(CheckInitOrder, CheckUseAfterReturn,
509 CheckLifetime, BlacklistFile, ZeroBaseShadow);
512 char AddressSanitizerModule::ID = 0;
513 INITIALIZE_PASS(AddressSanitizerModule, "asan-module",
514 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
515 "ModulePass", false, false)
516 ModulePass *llvm::createAddressSanitizerModulePass(
517 bool CheckInitOrder, StringRef BlacklistFile, bool ZeroBaseShadow) {
518 return new AddressSanitizerModule(CheckInitOrder, BlacklistFile,
522 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
523 size_t Res = CountTrailingZeros_32(TypeSize / 8);
524 assert(Res < kNumberOfAccessSizes);
528 // Create a constant for Str so that we can pass it to the run-time lib.
529 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
530 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
531 GlobalVariable *GV = new GlobalVariable(M, StrConst->getType(), true,
532 GlobalValue::PrivateLinkage, StrConst,
534 GV->setUnnamedAddr(true); // Ok to merge these.
535 GV->setAlignment(1); // Strings may not be merged w/o setting align 1.
539 static bool GlobalWasGeneratedByAsan(GlobalVariable *G) {
540 return G->getName().find(kAsanGenPrefix) == 0;
543 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
545 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
546 if (Mapping.Offset == 0)
548 // (Shadow >> scale) | offset
549 if (Mapping.OrShadowOffset)
550 return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
552 return IRB.CreateAdd(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
555 void AddressSanitizer::instrumentMemIntrinsicParam(
556 Instruction *OrigIns,
557 Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
558 IRBuilder<> IRB(InsertBefore);
559 if (Size->getType() != IntptrTy)
560 Size = IRB.CreateIntCast(Size, IntptrTy, false);
561 // Check the first byte.
562 instrumentAddress(OrigIns, InsertBefore, Addr, 8, IsWrite, Size);
563 // Check the last byte.
564 IRB.SetInsertPoint(InsertBefore);
565 Value *SizeMinusOne = IRB.CreateSub(Size, ConstantInt::get(IntptrTy, 1));
566 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
567 Value *AddrLast = IRB.CreateAdd(AddrLong, SizeMinusOne);
568 instrumentAddress(OrigIns, InsertBefore, AddrLast, 8, IsWrite, Size);
571 // Instrument memset/memmove/memcpy
572 bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
573 Value *Dst = MI->getDest();
574 MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
575 Value *Src = MemTran ? MemTran->getSource() : 0;
576 Value *Length = MI->getLength();
578 Constant *ConstLength = dyn_cast<Constant>(Length);
579 Instruction *InsertBefore = MI;
581 if (ConstLength->isNullValue()) return false;
583 // The size is not a constant so it could be zero -- check at run-time.
584 IRBuilder<> IRB(InsertBefore);
586 Value *Cmp = IRB.CreateICmpNE(Length,
587 Constant::getNullValue(Length->getType()));
588 InsertBefore = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
591 instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
593 instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
597 // If I is an interesting memory access, return the PointerOperand
598 // and set IsWrite. Otherwise return NULL.
599 static Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite) {
600 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
601 if (!ClInstrumentReads) return NULL;
603 return LI->getPointerOperand();
605 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
606 if (!ClInstrumentWrites) return NULL;
608 return SI->getPointerOperand();
610 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
611 if (!ClInstrumentAtomics) return NULL;
613 return RMW->getPointerOperand();
615 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
616 if (!ClInstrumentAtomics) return NULL;
618 return XCHG->getPointerOperand();
623 void AddressSanitizer::instrumentMop(Instruction *I) {
624 bool IsWrite = false;
625 Value *Addr = isInterestingMemoryAccess(I, &IsWrite);
627 if (ClOpt && ClOptGlobals) {
628 if (GlobalVariable *G = dyn_cast<GlobalVariable>(Addr)) {
629 // If initialization order checking is disabled, a simple access to a
630 // dynamically initialized global is always valid.
633 // If a global variable does not have dynamic initialization we don't
634 // have to instrument it. However, if a global does not have initailizer
635 // at all, we assume it has dynamic initializer (in other TU).
636 if (G->hasInitializer() && !DynamicallyInitializedGlobals.Contains(G))
641 Type *OrigPtrTy = Addr->getType();
642 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
644 assert(OrigTy->isSized());
645 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
647 assert((TypeSize % 8) == 0);
649 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check.
650 if (TypeSize == 8 || TypeSize == 16 ||
651 TypeSize == 32 || TypeSize == 64 || TypeSize == 128)
652 return instrumentAddress(I, I, Addr, TypeSize, IsWrite, 0);
653 // Instrument unusual size (but still multiple of 8).
654 // We can not do it with a single check, so we do 1-byte check for the first
655 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
656 // to report the actual access size.
658 Value *LastByte = IRB.CreateIntToPtr(
659 IRB.CreateAdd(IRB.CreatePointerCast(Addr, IntptrTy),
660 ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
662 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
663 instrumentAddress(I, I, Addr, 8, IsWrite, Size);
664 instrumentAddress(I, I, LastByte, 8, IsWrite, Size);
667 // Validate the result of Module::getOrInsertFunction called for an interface
668 // function of AddressSanitizer. If the instrumented module defines a function
669 // with the same name, their prototypes must match, otherwise
670 // getOrInsertFunction returns a bitcast.
671 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
672 if (isa<Function>(FuncOrBitcast)) return cast<Function>(FuncOrBitcast);
673 FuncOrBitcast->dump();
674 report_fatal_error("trying to redefine an AddressSanitizer "
675 "interface function");
678 Instruction *AddressSanitizer::generateCrashCode(
679 Instruction *InsertBefore, Value *Addr,
680 bool IsWrite, size_t AccessSizeIndex, Value *SizeArgument) {
681 IRBuilder<> IRB(InsertBefore);
682 CallInst *Call = SizeArgument
683 ? IRB.CreateCall2(AsanErrorCallbackSized[IsWrite], Addr, SizeArgument)
684 : IRB.CreateCall(AsanErrorCallback[IsWrite][AccessSizeIndex], Addr);
686 // We don't do Call->setDoesNotReturn() because the BB already has
687 // UnreachableInst at the end.
688 // This EmptyAsm is required to avoid callback merge.
689 IRB.CreateCall(EmptyAsm);
693 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
696 size_t Granularity = 1 << Mapping.Scale;
697 // Addr & (Granularity - 1)
698 Value *LastAccessedByte = IRB.CreateAnd(
699 AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
700 // (Addr & (Granularity - 1)) + size - 1
701 if (TypeSize / 8 > 1)
702 LastAccessedByte = IRB.CreateAdd(
703 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
704 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
705 LastAccessedByte = IRB.CreateIntCast(
706 LastAccessedByte, ShadowValue->getType(), false);
707 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
708 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
711 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
712 Instruction *InsertBefore,
713 Value *Addr, uint32_t TypeSize,
714 bool IsWrite, Value *SizeArgument) {
715 IRBuilder<> IRB(InsertBefore);
716 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
718 Type *ShadowTy = IntegerType::get(
719 *C, std::max(8U, TypeSize >> Mapping.Scale));
720 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
721 Value *ShadowPtr = memToShadow(AddrLong, IRB);
722 Value *CmpVal = Constant::getNullValue(ShadowTy);
723 Value *ShadowValue = IRB.CreateLoad(
724 IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
726 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
727 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
728 size_t Granularity = 1 << Mapping.Scale;
729 TerminatorInst *CrashTerm = 0;
731 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
732 TerminatorInst *CheckTerm =
733 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
734 assert(dyn_cast<BranchInst>(CheckTerm)->isUnconditional());
735 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
736 IRB.SetInsertPoint(CheckTerm);
737 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
738 BasicBlock *CrashBlock =
739 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
740 CrashTerm = new UnreachableInst(*C, CrashBlock);
741 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
742 ReplaceInstWithInst(CheckTerm, NewTerm);
744 CrashTerm = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), true);
747 Instruction *Crash = generateCrashCode(
748 CrashTerm, AddrLong, IsWrite, AccessSizeIndex, SizeArgument);
749 Crash->setDebugLoc(OrigIns->getDebugLoc());
752 void AddressSanitizerModule::createInitializerPoisonCalls(
753 Module &M, GlobalValue *ModuleName) {
754 // We do all of our poisoning and unpoisoning within _GLOBAL__I_a.
755 Function *GlobalInit = M.getFunction("_GLOBAL__I_a");
756 // If that function is not present, this TU contains no globals, or they have
757 // all been optimized away
761 // Set up the arguments to our poison/unpoison functions.
762 IRBuilder<> IRB(GlobalInit->begin()->getFirstInsertionPt());
764 // Add a call to poison all external globals before the given function starts.
765 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
766 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
768 // Add calls to unpoison all globals before each return instruction.
769 for (Function::iterator I = GlobalInit->begin(), E = GlobalInit->end();
771 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator())) {
772 CallInst::Create(AsanUnpoisonGlobals, "", RI);
777 bool AddressSanitizerModule::ShouldInstrumentGlobal(GlobalVariable *G) {
778 Type *Ty = cast<PointerType>(G->getType())->getElementType();
779 DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
781 if (BL->isIn(*G)) return false;
782 if (!Ty->isSized()) return false;
783 if (!G->hasInitializer()) return false;
784 if (GlobalWasGeneratedByAsan(G)) return false; // Our own global.
785 // Touch only those globals that will not be defined in other modules.
786 // Don't handle ODR type linkages since other modules may be built w/o asan.
787 if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
788 G->getLinkage() != GlobalVariable::PrivateLinkage &&
789 G->getLinkage() != GlobalVariable::InternalLinkage)
791 // Two problems with thread-locals:
792 // - The address of the main thread's copy can't be computed at link-time.
793 // - Need to poison all copies, not just the main thread's one.
794 if (G->isThreadLocal())
796 // For now, just ignore this Alloca if the alignment is large.
797 if (G->getAlignment() > RedzoneSize()) return false;
799 // Ignore all the globals with the names starting with "\01L_OBJC_".
800 // Many of those are put into the .cstring section. The linker compresses
801 // that section by removing the spare \0s after the string terminator, so
802 // our redzones get broken.
803 if ((G->getName().find("\01L_OBJC_") == 0) ||
804 (G->getName().find("\01l_OBJC_") == 0)) {
805 DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
809 if (G->hasSection()) {
810 StringRef Section(G->getSection());
811 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
812 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
814 if ((Section.find("__OBJC,") == 0) ||
815 (Section.find("__DATA, __objc_") == 0)) {
816 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
819 // See http://code.google.com/p/address-sanitizer/issues/detail?id=32
820 // Constant CFString instances are compiled in the following way:
821 // -- the string buffer is emitted into
822 // __TEXT,__cstring,cstring_literals
823 // -- the constant NSConstantString structure referencing that buffer
824 // is placed into __DATA,__cfstring
825 // Therefore there's no point in placing redzones into __DATA,__cfstring.
826 // Moreover, it causes the linker to crash on OS X 10.7
827 if (Section.find("__DATA,__cfstring") == 0) {
828 DEBUG(dbgs() << "Ignoring CFString: " << *G);
836 void AddressSanitizerModule::initializeCallbacks(Module &M) {
838 // Declare our poisoning and unpoisoning functions.
839 AsanPoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
840 kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, NULL));
841 AsanPoisonGlobals->setLinkage(Function::ExternalLinkage);
842 AsanUnpoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
843 kAsanUnpoisonGlobalsName, IRB.getVoidTy(), NULL));
844 AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage);
845 // Declare functions that register/unregister globals.
846 AsanRegisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
847 kAsanRegisterGlobalsName, IRB.getVoidTy(),
848 IntptrTy, IntptrTy, NULL));
849 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
850 AsanUnregisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
851 kAsanUnregisterGlobalsName,
852 IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
853 AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage);
856 // This function replaces all global variables with new variables that have
857 // trailing redzones. It also creates a function that poisons
858 // redzones and inserts this function into llvm.global_ctors.
859 bool AddressSanitizerModule::runOnModule(Module &M) {
860 if (!ClGlobals) return false;
861 TD = getAnalysisIfAvailable<DataLayout>();
864 BL.reset(new BlackList(BlacklistFile));
865 if (BL->isIn(M)) return false;
866 C = &(M.getContext());
867 int LongSize = TD->getPointerSizeInBits();
868 IntptrTy = Type::getIntNTy(*C, LongSize);
869 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
870 initializeCallbacks(M);
871 DynamicallyInitializedGlobals.Init(M);
873 SmallVector<GlobalVariable *, 16> GlobalsToChange;
875 for (Module::GlobalListType::iterator G = M.global_begin(),
876 E = M.global_end(); G != E; ++G) {
877 if (ShouldInstrumentGlobal(G))
878 GlobalsToChange.push_back(G);
881 size_t n = GlobalsToChange.size();
882 if (n == 0) return false;
884 // A global is described by a structure
887 // size_t size_with_redzone;
889 // const char *module_name;
890 // size_t has_dynamic_init;
891 // We initialize an array of such structures and pass it to a run-time call.
892 StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
894 IntptrTy, IntptrTy, NULL);
895 SmallVector<Constant *, 16> Initializers(n), DynamicInit;
898 Function *CtorFunc = M.getFunction(kAsanModuleCtorName);
900 IRBuilder<> IRB(CtorFunc->getEntryBlock().getTerminator());
902 bool HasDynamicallyInitializedGlobals = false;
904 GlobalVariable *ModuleName = createPrivateGlobalForString(
905 M, M.getModuleIdentifier());
906 // We shouldn't merge same module names, as this string serves as unique
907 // module ID in runtime.
908 ModuleName->setUnnamedAddr(false);
910 for (size_t i = 0; i < n; i++) {
911 static const uint64_t kMaxGlobalRedzone = 1 << 18;
912 GlobalVariable *G = GlobalsToChange[i];
913 PointerType *PtrTy = cast<PointerType>(G->getType());
914 Type *Ty = PtrTy->getElementType();
915 uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
916 uint64_t MinRZ = RedzoneSize();
917 // MinRZ <= RZ <= kMaxGlobalRedzone
918 // and trying to make RZ to be ~ 1/4 of SizeInBytes.
919 uint64_t RZ = std::max(MinRZ,
920 std::min(kMaxGlobalRedzone,
921 (SizeInBytes / MinRZ / 4) * MinRZ));
922 uint64_t RightRedzoneSize = RZ;
924 if (SizeInBytes % MinRZ)
925 RightRedzoneSize += MinRZ - (SizeInBytes % MinRZ);
926 assert(((RightRedzoneSize + SizeInBytes) % MinRZ) == 0);
927 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
928 // Determine whether this global should be poisoned in initialization.
929 bool GlobalHasDynamicInitializer =
930 DynamicallyInitializedGlobals.Contains(G);
931 // Don't check initialization order if this global is blacklisted.
932 GlobalHasDynamicInitializer &= !BL->isInInit(*G);
934 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
935 Constant *NewInitializer = ConstantStruct::get(
936 NewTy, G->getInitializer(),
937 Constant::getNullValue(RightRedZoneTy), NULL);
939 GlobalVariable *Name = createPrivateGlobalForString(M, G->getName());
941 // Create a new global variable with enough space for a redzone.
942 GlobalVariable *NewGlobal = new GlobalVariable(
943 M, NewTy, G->isConstant(), G->getLinkage(),
944 NewInitializer, "", G, G->getThreadLocalMode());
945 NewGlobal->copyAttributesFrom(G);
946 NewGlobal->setAlignment(MinRZ);
949 Indices2[0] = IRB.getInt32(0);
950 Indices2[1] = IRB.getInt32(0);
952 G->replaceAllUsesWith(
953 ConstantExpr::getGetElementPtr(NewGlobal, Indices2, true));
954 NewGlobal->takeName(G);
955 G->eraseFromParent();
957 Initializers[i] = ConstantStruct::get(
959 ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
960 ConstantInt::get(IntptrTy, SizeInBytes),
961 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
962 ConstantExpr::getPointerCast(Name, IntptrTy),
963 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
964 ConstantInt::get(IntptrTy, GlobalHasDynamicInitializer),
967 // Populate the first and last globals declared in this TU.
968 if (CheckInitOrder && GlobalHasDynamicInitializer)
969 HasDynamicallyInitializedGlobals = true;
971 DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
974 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
975 GlobalVariable *AllGlobals = new GlobalVariable(
976 M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage,
977 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
979 // Create calls for poisoning before initializers run and unpoisoning after.
980 if (CheckInitOrder && HasDynamicallyInitializedGlobals)
981 createInitializerPoisonCalls(M, ModuleName);
982 IRB.CreateCall2(AsanRegisterGlobals,
983 IRB.CreatePointerCast(AllGlobals, IntptrTy),
984 ConstantInt::get(IntptrTy, n));
986 // We also need to unregister globals at the end, e.g. when a shared library
988 Function *AsanDtorFunction = Function::Create(
989 FunctionType::get(Type::getVoidTy(*C), false),
990 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
991 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
992 IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB));
993 IRB_Dtor.CreateCall2(AsanUnregisterGlobals,
994 IRB.CreatePointerCast(AllGlobals, IntptrTy),
995 ConstantInt::get(IntptrTy, n));
996 appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndCtorPriority);
1002 void AddressSanitizer::initializeCallbacks(Module &M) {
1003 IRBuilder<> IRB(*C);
1004 // Create __asan_report* callbacks.
1005 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
1006 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
1007 AccessSizeIndex++) {
1008 // IsWrite and TypeSize are encoded in the function name.
1009 std::string FunctionName = std::string(kAsanReportErrorTemplate) +
1010 (AccessIsWrite ? "store" : "load") + itostr(1 << AccessSizeIndex);
1011 // If we are merging crash callbacks, they have two parameters.
1012 AsanErrorCallback[AccessIsWrite][AccessSizeIndex] =
1013 checkInterfaceFunction(M.getOrInsertFunction(
1014 FunctionName, IRB.getVoidTy(), IntptrTy, NULL));
1017 AsanErrorCallbackSized[0] = checkInterfaceFunction(M.getOrInsertFunction(
1018 kAsanReportLoadN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1019 AsanErrorCallbackSized[1] = checkInterfaceFunction(M.getOrInsertFunction(
1020 kAsanReportStoreN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1022 AsanHandleNoReturnFunc = checkInterfaceFunction(M.getOrInsertFunction(
1023 kAsanHandleNoReturnName, IRB.getVoidTy(), NULL));
1024 // We insert an empty inline asm after __asan_report* to avoid callback merge.
1025 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
1026 StringRef(""), StringRef(""),
1027 /*hasSideEffects=*/true);
1030 void AddressSanitizer::emitShadowMapping(Module &M, IRBuilder<> &IRB) const {
1031 // Tell the values of mapping offset and scale to the run-time.
1032 GlobalValue *asan_mapping_offset =
1033 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1034 ConstantInt::get(IntptrTy, Mapping.Offset),
1035 kAsanMappingOffsetName);
1036 // Read the global, otherwise it may be optimized away.
1037 IRB.CreateLoad(asan_mapping_offset, true);
1039 GlobalValue *asan_mapping_scale =
1040 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1041 ConstantInt::get(IntptrTy, Mapping.Scale),
1042 kAsanMappingScaleName);
1043 // Read the global, otherwise it may be optimized away.
1044 IRB.CreateLoad(asan_mapping_scale, true);
1048 bool AddressSanitizer::doInitialization(Module &M) {
1049 // Initialize the private fields. No one has accessed them before.
1050 TD = getAnalysisIfAvailable<DataLayout>();
1054 BL.reset(new BlackList(BlacklistFile));
1055 DynamicallyInitializedGlobals.Init(M);
1057 C = &(M.getContext());
1058 LongSize = TD->getPointerSizeInBits();
1059 IntptrTy = Type::getIntNTy(*C, LongSize);
1061 AsanCtorFunction = Function::Create(
1062 FunctionType::get(Type::getVoidTy(*C), false),
1063 GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
1064 BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
1065 // call __asan_init in the module ctor.
1066 IRBuilder<> IRB(ReturnInst::Create(*C, AsanCtorBB));
1067 AsanInitFunction = checkInterfaceFunction(
1068 M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
1069 AsanInitFunction->setLinkage(Function::ExternalLinkage);
1070 IRB.CreateCall(AsanInitFunction);
1072 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
1073 emitShadowMapping(M, IRB);
1075 appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority);
1079 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
1080 // For each NSObject descendant having a +load method, this method is invoked
1081 // by the ObjC runtime before any of the static constructors is called.
1082 // Therefore we need to instrument such methods with a call to __asan_init
1083 // at the beginning in order to initialize our runtime before any access to
1084 // the shadow memory.
1085 // We cannot just ignore these methods, because they may call other
1086 // instrumented functions.
1087 if (F.getName().find(" load]") != std::string::npos) {
1088 IRBuilder<> IRB(F.begin()->begin());
1089 IRB.CreateCall(AsanInitFunction);
1095 bool AddressSanitizer::runOnFunction(Function &F) {
1096 if (BL->isIn(F)) return false;
1097 if (&F == AsanCtorFunction) return false;
1098 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
1099 DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
1100 initializeCallbacks(*F.getParent());
1102 // If needed, insert __asan_init before checking for SanitizeAddress attr.
1103 maybeInsertAsanInitAtFunctionEntry(F);
1105 if (!F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1106 Attribute::SanitizeAddress))
1109 if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
1112 // We want to instrument every address only once per basic block (unless there
1113 // are calls between uses).
1114 SmallSet<Value*, 16> TempsToInstrument;
1115 SmallVector<Instruction*, 16> ToInstrument;
1116 SmallVector<Instruction*, 8> NoReturnCalls;
1119 // Fill the set of memory operations to instrument.
1120 for (Function::iterator FI = F.begin(), FE = F.end();
1122 TempsToInstrument.clear();
1123 int NumInsnsPerBB = 0;
1124 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1126 if (LooksLikeCodeInBug11395(BI)) return false;
1127 if (Value *Addr = isInterestingMemoryAccess(BI, &IsWrite)) {
1128 if (ClOpt && ClOptSameTemp) {
1129 if (!TempsToInstrument.insert(Addr))
1130 continue; // We've seen this temp in the current BB.
1132 } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
1137 // A call inside BB.
1138 TempsToInstrument.clear();
1139 if (CS.doesNotReturn())
1140 NoReturnCalls.push_back(CS.getInstruction());
1144 ToInstrument.push_back(BI);
1146 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB)
1152 int NumInstrumented = 0;
1153 for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
1154 Instruction *Inst = ToInstrument[i];
1155 if (ClDebugMin < 0 || ClDebugMax < 0 ||
1156 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
1157 if (isInterestingMemoryAccess(Inst, &IsWrite))
1158 instrumentMop(Inst);
1160 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
1165 FunctionStackPoisoner FSP(F, *this);
1166 bool ChangedStack = FSP.runOnFunction();
1168 // We must unpoison the stack before every NoReturn call (throw, _exit, etc).
1169 // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37
1170 for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) {
1171 Instruction *CI = NoReturnCalls[i];
1172 IRBuilder<> IRB(CI);
1173 IRB.CreateCall(AsanHandleNoReturnFunc);
1175 DEBUG(dbgs() << "ASAN done instrumenting:\n" << F << "\n");
1177 return NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty();
1180 static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
1181 if (ShadowRedzoneSize == 1) return PoisonByte;
1182 if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
1183 if (ShadowRedzoneSize == 4)
1184 return (PoisonByte << 24) + (PoisonByte << 16) +
1185 (PoisonByte << 8) + (PoisonByte);
1186 llvm_unreachable("ShadowRedzoneSize is either 1, 2 or 4");
1189 static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
1192 size_t ShadowGranularity,
1194 for (size_t i = 0; i < RZSize;
1195 i+= ShadowGranularity, Shadow++) {
1196 if (i + ShadowGranularity <= Size) {
1197 *Shadow = 0; // fully addressable
1198 } else if (i >= Size) {
1199 *Shadow = Magic; // unaddressable
1201 *Shadow = Size - i; // first Size-i bytes are addressable
1206 // Workaround for bug 11395: we don't want to instrument stack in functions
1207 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
1208 // FIXME: remove once the bug 11395 is fixed.
1209 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
1210 if (LongSize != 32) return false;
1211 CallInst *CI = dyn_cast<CallInst>(I);
1212 if (!CI || !CI->isInlineAsm()) return false;
1213 if (CI->getNumArgOperands() <= 5) return false;
1214 // We have inline assembly with quite a few arguments.
1218 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
1219 IRBuilder<> IRB(*C);
1220 AsanStackMallocFunc = checkInterfaceFunction(M.getOrInsertFunction(
1221 kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL));
1222 AsanStackFreeFunc = checkInterfaceFunction(M.getOrInsertFunction(
1223 kAsanStackFreeName, IRB.getVoidTy(),
1224 IntptrTy, IntptrTy, IntptrTy, NULL));
1225 AsanPoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1226 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1227 AsanUnpoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1228 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1231 void FunctionStackPoisoner::poisonRedZones(
1232 const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB, Value *ShadowBase,
1234 size_t ShadowRZSize = RedzoneSize() >> Mapping.Scale;
1235 assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
1236 Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
1237 Type *RZPtrTy = PointerType::get(RZTy, 0);
1239 Value *PoisonLeft = ConstantInt::get(RZTy,
1240 ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
1241 Value *PoisonMid = ConstantInt::get(RZTy,
1242 ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
1243 Value *PoisonRight = ConstantInt::get(RZTy,
1244 ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
1246 // poison the first red zone.
1247 IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
1249 // poison all other red zones.
1250 uint64_t Pos = RedzoneSize();
1251 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1252 AllocaInst *AI = AllocaVec[i];
1253 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1254 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1255 assert(AlignedSize - SizeInBytes < RedzoneSize());
1260 assert(ShadowBase->getType() == IntptrTy);
1261 if (SizeInBytes < AlignedSize) {
1262 // Poison the partial redzone at right
1263 Ptr = IRB.CreateAdd(
1264 ShadowBase, ConstantInt::get(IntptrTy,
1265 (Pos >> Mapping.Scale) - ShadowRZSize));
1266 size_t AddressableBytes = RedzoneSize() - (AlignedSize - SizeInBytes);
1267 uint32_t Poison = 0;
1269 PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
1271 1ULL << Mapping.Scale,
1272 kAsanStackPartialRedzoneMagic);
1274 Value *PartialPoison = ConstantInt::get(RZTy, Poison);
1275 IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1278 // Poison the full redzone at right.
1279 Ptr = IRB.CreateAdd(ShadowBase,
1280 ConstantInt::get(IntptrTy, Pos >> Mapping.Scale));
1281 bool LastAlloca = (i == AllocaVec.size() - 1);
1282 Value *Poison = LastAlloca ? PoisonRight : PoisonMid;
1283 IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1285 Pos += RedzoneSize();
1289 void FunctionStackPoisoner::poisonStack() {
1290 uint64_t LocalStackSize = TotalStackSize +
1291 (AllocaVec.size() + 1) * RedzoneSize();
1293 bool DoStackMalloc = ASan.CheckUseAfterReturn
1294 && LocalStackSize <= kMaxStackMallocSize;
1296 assert(AllocaVec.size() > 0);
1297 Instruction *InsBefore = AllocaVec[0];
1298 IRBuilder<> IRB(InsBefore);
1301 Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
1302 AllocaInst *MyAlloca =
1303 new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
1304 if (ClRealignStack && StackAlignment < RedzoneSize())
1305 StackAlignment = RedzoneSize();
1306 MyAlloca->setAlignment(StackAlignment);
1307 assert(MyAlloca->isStaticAlloca());
1308 Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
1309 Value *LocalStackBase = OrigStackBase;
1311 if (DoStackMalloc) {
1312 LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc,
1313 ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
1316 // This string will be parsed by the run-time (DescribeAddressIfStack).
1317 SmallString<2048> StackDescriptionStorage;
1318 raw_svector_ostream StackDescription(StackDescriptionStorage);
1319 StackDescription << AllocaVec.size() << " ";
1321 // Insert poison calls for lifetime intrinsics for alloca.
1322 bool HavePoisonedAllocas = false;
1323 for (size_t i = 0, n = AllocaPoisonCallVec.size(); i < n; i++) {
1324 const AllocaPoisonCall &APC = AllocaPoisonCallVec[i];
1325 IntrinsicInst *II = APC.InsBefore;
1326 AllocaInst *AI = findAllocaForValue(II->getArgOperand(1));
1328 IRBuilder<> IRB(II);
1329 poisonAlloca(AI, APC.Size, IRB, APC.DoPoison);
1330 HavePoisonedAllocas |= APC.DoPoison;
1333 uint64_t Pos = RedzoneSize();
1334 // Replace Alloca instructions with base+offset.
1335 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1336 AllocaInst *AI = AllocaVec[i];
1337 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1338 StringRef Name = AI->getName();
1339 StackDescription << Pos << " " << SizeInBytes << " "
1340 << Name.size() << " " << Name << " ";
1341 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1342 assert((AlignedSize % RedzoneSize()) == 0);
1343 Value *NewAllocaPtr = IRB.CreateIntToPtr(
1344 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
1346 replaceDbgDeclareForAlloca(AI, NewAllocaPtr, DIB);
1347 AI->replaceAllUsesWith(NewAllocaPtr);
1348 Pos += AlignedSize + RedzoneSize();
1350 assert(Pos == LocalStackSize);
1352 // The left-most redzone has enough space for at least 4 pointers.
1353 // Write the Magic value to redzone[0].
1354 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
1355 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
1357 // Write the frame description constant to redzone[1].
1358 Value *BasePlus1 = IRB.CreateIntToPtr(
1359 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, ASan.LongSize/8)),
1361 GlobalVariable *StackDescriptionGlobal =
1362 createPrivateGlobalForString(*F.getParent(), StackDescription.str());
1363 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal,
1365 IRB.CreateStore(Description, BasePlus1);
1366 // Write the PC to redzone[2].
1367 Value *BasePlus2 = IRB.CreateIntToPtr(
1368 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy,
1369 2 * ASan.LongSize/8)),
1371 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
1373 // Poison the stack redzones at the entry.
1374 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
1375 poisonRedZones(AllocaVec, IRB, ShadowBase, true);
1377 // Unpoison the stack before all ret instructions.
1378 for (size_t i = 0, n = RetVec.size(); i < n; i++) {
1379 Instruction *Ret = RetVec[i];
1380 IRBuilder<> IRBRet(Ret);
1381 // Mark the current frame as retired.
1382 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
1384 // Unpoison the stack.
1385 poisonRedZones(AllocaVec, IRBRet, ShadowBase, false);
1386 if (DoStackMalloc) {
1387 // In use-after-return mode, mark the whole stack frame unaddressable.
1388 IRBRet.CreateCall3(AsanStackFreeFunc, LocalStackBase,
1389 ConstantInt::get(IntptrTy, LocalStackSize),
1391 } else if (HavePoisonedAllocas) {
1392 // If we poisoned some allocas in llvm.lifetime analysis,
1393 // unpoison whole stack frame now.
1394 assert(LocalStackBase == OrigStackBase);
1395 poisonAlloca(LocalStackBase, LocalStackSize, IRBRet, false);
1399 // We are done. Remove the old unused alloca instructions.
1400 for (size_t i = 0, n = AllocaVec.size(); i < n; i++)
1401 AllocaVec[i]->eraseFromParent();
1404 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
1405 IRBuilder<> IRB, bool DoPoison) {
1406 // For now just insert the call to ASan runtime.
1407 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
1408 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
1409 IRB.CreateCall2(DoPoison ? AsanPoisonStackMemoryFunc
1410 : AsanUnpoisonStackMemoryFunc,
1414 // Handling llvm.lifetime intrinsics for a given %alloca:
1415 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
1416 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
1417 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
1418 // could be poisoned by previous llvm.lifetime.end instruction, as the
1419 // variable may go in and out of scope several times, e.g. in loops).
1420 // (3) if we poisoned at least one %alloca in a function,
1421 // unpoison the whole stack frame at function exit.
1423 AllocaInst *FunctionStackPoisoner::findAllocaForValue(Value *V) {
1424 if (AllocaInst *AI = dyn_cast<AllocaInst>(V))
1425 // We're intested only in allocas we can handle.
1426 return isInterestingAlloca(*AI) ? AI : 0;
1427 // See if we've already calculated (or started to calculate) alloca for a
1429 AllocaForValueMapTy::iterator I = AllocaForValue.find(V);
1430 if (I != AllocaForValue.end())
1432 // Store 0 while we're calculating alloca for value V to avoid
1433 // infinite recursion if the value references itself.
1434 AllocaForValue[V] = 0;
1435 AllocaInst *Res = 0;
1436 if (CastInst *CI = dyn_cast<CastInst>(V))
1437 Res = findAllocaForValue(CI->getOperand(0));
1438 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1439 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1440 Value *IncValue = PN->getIncomingValue(i);
1441 // Allow self-referencing phi-nodes.
1442 if (IncValue == PN) continue;
1443 AllocaInst *IncValueAI = findAllocaForValue(IncValue);
1444 // AI for incoming values should exist and should all be equal.
1445 if (IncValueAI == 0 || (Res != 0 && IncValueAI != Res))
1451 AllocaForValue[V] = Res;