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/Endian.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Support/system_error.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BlackList.h"
47 #include "llvm/Transforms/Utils/Cloning.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/ModuleUtils.h"
55 static const uint64_t kDefaultShadowScale = 3;
56 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
57 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
58 static const uint64_t kDefaultShort64bitShadowOffset = 0x7FFF8000; // < 2G.
59 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 41;
60 static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa8000;
62 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
63 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
64 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
66 static const char *kAsanModuleCtorName = "asan.module_ctor";
67 static const char *kAsanModuleDtorName = "asan.module_dtor";
68 static const int kAsanCtorAndCtorPriority = 1;
69 static const char *kAsanReportErrorTemplate = "__asan_report_";
70 static const char *kAsanReportLoadN = "__asan_report_load_n";
71 static const char *kAsanReportStoreN = "__asan_report_store_n";
72 static const char *kAsanRegisterGlobalsName = "__asan_register_globals";
73 static const char *kAsanUnregisterGlobalsName = "__asan_unregister_globals";
74 static const char *kAsanPoisonGlobalsName = "__asan_before_dynamic_init";
75 static const char *kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init";
76 static const char *kAsanInitName = "__asan_init_v3";
77 static const char *kAsanHandleNoReturnName = "__asan_handle_no_return";
78 static const char *kAsanMappingOffsetName = "__asan_mapping_offset";
79 static const char *kAsanMappingScaleName = "__asan_mapping_scale";
80 static const char *kAsanStackMallocName = "__asan_stack_malloc";
81 static const char *kAsanStackFreeName = "__asan_stack_free";
82 static const char *kAsanGenPrefix = "__asan_gen_";
83 static const char *kAsanPoisonStackMemoryName = "__asan_poison_stack_memory";
84 static const char *kAsanUnpoisonStackMemoryName =
85 "__asan_unpoison_stack_memory";
87 static const int kAsanStackLeftRedzoneMagic = 0xf1;
88 static const int kAsanStackMidRedzoneMagic = 0xf2;
89 static const int kAsanStackRightRedzoneMagic = 0xf3;
90 static const int kAsanStackPartialRedzoneMagic = 0xf4;
92 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
93 static const size_t kNumberOfAccessSizes = 5;
95 // Command-line flags.
97 // This flag may need to be replaced with -f[no-]asan-reads.
98 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
99 cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
100 static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
101 cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
102 static cl::opt<bool> ClInstrumentAtomics("asan-instrument-atomics",
103 cl::desc("instrument atomic instructions (rmw, cmpxchg)"),
104 cl::Hidden, cl::init(true));
105 static cl::opt<bool> ClAlwaysSlowPath("asan-always-slow-path",
106 cl::desc("use instrumentation with slow path for all accesses"),
107 cl::Hidden, cl::init(false));
108 // This flag limits the number of instructions to be instrumented
109 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
110 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
112 static cl::opt<int> ClMaxInsnsToInstrumentPerBB("asan-max-ins-per-bb",
114 cl::desc("maximal number of instructions to instrument in any given BB"),
116 // This flag may need to be replaced with -f[no]asan-stack.
117 static cl::opt<bool> ClStack("asan-stack",
118 cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
119 // This flag may need to be replaced with -f[no]asan-use-after-return.
120 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
121 cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
122 // This flag may need to be replaced with -f[no]asan-globals.
123 static cl::opt<bool> ClGlobals("asan-globals",
124 cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
125 static cl::opt<bool> ClInitializers("asan-initialization-order",
126 cl::desc("Handle C++ initializer order"), cl::Hidden, cl::init(false));
127 static cl::opt<bool> ClMemIntrin("asan-memintrin",
128 cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
129 static cl::opt<bool> ClRealignStack("asan-realign-stack",
130 cl::desc("Realign stack to 32"), cl::Hidden, cl::init(true));
131 static cl::opt<std::string> ClBlacklistFile("asan-blacklist",
132 cl::desc("File containing the list of objects to ignore "
133 "during instrumentation"), cl::Hidden);
135 // This is an experimental feature that will allow to choose between
136 // instrumented and non-instrumented code at link-time.
137 // If this option is on, just before instrumenting a function we create its
138 // clone; if the function is not changed by asan the clone is deleted.
139 // If we end up with a clone, we put the instrumented function into a section
140 // called "ASAN" and the uninstrumented function into a section called "NOASAN".
142 // This is still a prototype, we need to figure out a way to keep two copies of
143 // a function so that the linker can easily choose one of them.
144 static cl::opt<bool> ClKeepUninstrumented("asan-keep-uninstrumented-functions",
145 cl::desc("Keep uninstrumented copies of functions"),
146 cl::Hidden, cl::init(false));
148 // These flags allow to change the shadow mapping.
149 // The shadow mapping looks like
150 // Shadow = (Mem >> scale) + (1 << offset_log)
151 static cl::opt<int> ClMappingScale("asan-mapping-scale",
152 cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
153 static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
154 cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
155 static cl::opt<bool> ClShort64BitOffset("asan-short-64bit-mapping-offset",
156 cl::desc("Use short immediate constant as the mapping offset for 64bit"),
157 cl::Hidden, cl::init(true));
159 // Optimization flags. Not user visible, used mostly for testing
160 // and benchmarking the tool.
161 static cl::opt<bool> ClOpt("asan-opt",
162 cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
163 static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
164 cl::desc("Instrument the same temp just once"), cl::Hidden,
166 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
167 cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
169 static cl::opt<bool> ClCheckLifetime("asan-check-lifetime",
170 cl::desc("Use llvm.lifetime intrinsics to insert extra checks"),
171 cl::Hidden, cl::init(false));
174 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
176 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
177 cl::Hidden, cl::init(0));
178 static cl::opt<std::string> ClDebugFunc("asan-debug-func",
179 cl::Hidden, cl::desc("Debug func"));
180 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
181 cl::Hidden, cl::init(-1));
182 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
183 cl::Hidden, cl::init(-1));
186 /// A set of dynamically initialized globals extracted from metadata.
187 class SetOfDynamicallyInitializedGlobals {
189 void Init(Module& M) {
190 // Clang generates metadata identifying all dynamically initialized globals.
191 NamedMDNode *DynamicGlobals =
192 M.getNamedMetadata("llvm.asan.dynamically_initialized_globals");
195 for (int i = 0, n = DynamicGlobals->getNumOperands(); i < n; ++i) {
196 MDNode *MDN = DynamicGlobals->getOperand(i);
197 assert(MDN->getNumOperands() == 1);
198 Value *VG = MDN->getOperand(0);
199 // The optimizer may optimize away a global entirely, in which case we
200 // cannot instrument access to it.
203 DynInitGlobals.insert(cast<GlobalVariable>(VG));
206 bool Contains(GlobalVariable *G) { return DynInitGlobals.count(G) != 0; }
208 SmallSet<GlobalValue*, 32> DynInitGlobals;
211 /// This struct defines the shadow mapping using the rule:
212 /// shadow = (mem >> Scale) ADD-or-OR Offset.
213 struct ShadowMapping {
219 static ShadowMapping getShadowMapping(const Module &M, int LongSize,
220 bool ZeroBaseShadow) {
221 llvm::Triple TargetTriple(M.getTargetTriple());
222 bool IsAndroid = TargetTriple.getEnvironment() == llvm::Triple::Android;
223 bool IsMacOSX = TargetTriple.getOS() == llvm::Triple::MacOSX;
224 bool IsPPC64 = TargetTriple.getArch() == llvm::Triple::ppc64;
225 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
226 bool IsMIPS32 = TargetTriple.getArch() == llvm::Triple::mips ||
227 TargetTriple.getArch() == llvm::Triple::mipsel;
229 ShadowMapping Mapping;
231 // OR-ing shadow offset if more efficient (at least on x86),
232 // but on ppc64 we have to use add since the shadow offset is not neccesary
233 // 1/8-th of the address space.
234 Mapping.OrShadowOffset = !IsPPC64 && !ClShort64BitOffset;
236 Mapping.Offset = (IsAndroid || ZeroBaseShadow) ? 0 :
238 (IsMIPS32 ? kMIPS32_ShadowOffset32 : kDefaultShadowOffset32) :
239 IsPPC64 ? kPPC64_ShadowOffset64 : kDefaultShadowOffset64);
240 if (!ZeroBaseShadow && ClShort64BitOffset && IsX86_64 && !IsMacOSX) {
241 assert(LongSize == 64);
242 Mapping.Offset = kDefaultShort64bitShadowOffset;
244 if (!ZeroBaseShadow && ClMappingOffsetLog >= 0) {
245 // Zero offset log is the special case.
246 Mapping.Offset = (ClMappingOffsetLog == 0) ? 0 : 1ULL << ClMappingOffsetLog;
249 Mapping.Scale = kDefaultShadowScale;
250 if (ClMappingScale) {
251 Mapping.Scale = ClMappingScale;
257 static size_t RedzoneSizeForScale(int MappingScale) {
258 // Redzone used for stack and globals is at least 32 bytes.
259 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
260 return std::max(32U, 1U << MappingScale);
263 /// AddressSanitizer: instrument the code in module to find memory bugs.
264 struct AddressSanitizer : public FunctionPass {
265 AddressSanitizer(bool CheckInitOrder = true,
266 bool CheckUseAfterReturn = false,
267 bool CheckLifetime = false,
268 StringRef BlacklistFile = StringRef(),
269 bool ZeroBaseShadow = false)
271 CheckInitOrder(CheckInitOrder || ClInitializers),
272 CheckUseAfterReturn(CheckUseAfterReturn || ClUseAfterReturn),
273 CheckLifetime(CheckLifetime || ClCheckLifetime),
274 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
276 ZeroBaseShadow(ZeroBaseShadow) {}
277 virtual const char *getPassName() const {
278 return "AddressSanitizerFunctionPass";
280 void instrumentMop(Instruction *I);
281 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
282 Value *Addr, uint32_t TypeSize, bool IsWrite,
283 Value *SizeArgument);
284 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
285 Value *ShadowValue, uint32_t TypeSize);
286 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
287 bool IsWrite, size_t AccessSizeIndex,
288 Value *SizeArgument);
289 bool instrumentMemIntrinsic(MemIntrinsic *MI);
290 void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
292 Instruction *InsertBefore, bool IsWrite);
293 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
294 bool runOnFunction(Function &F);
295 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
296 void emitShadowMapping(Module &M, IRBuilder<> &IRB) const;
297 virtual bool doInitialization(Module &M);
298 static char ID; // Pass identification, replacement for typeid
301 void initializeCallbacks(Module &M);
303 bool ShouldInstrumentGlobal(GlobalVariable *G);
304 bool LooksLikeCodeInBug11395(Instruction *I);
305 void FindDynamicInitializers(Module &M);
308 bool CheckUseAfterReturn;
310 SmallString<64> BlacklistFile;
317 ShadowMapping Mapping;
318 Function *AsanCtorFunction;
319 Function *AsanInitFunction;
320 Function *AsanHandleNoReturnFunc;
321 OwningPtr<BlackList> BL;
322 // This array is indexed by AccessIsWrite and log2(AccessSize).
323 Function *AsanErrorCallback[2][kNumberOfAccessSizes];
324 // This array is indexed by AccessIsWrite.
325 Function *AsanErrorCallbackSized[2];
327 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
329 friend struct FunctionStackPoisoner;
332 class AddressSanitizerModule : public ModulePass {
334 AddressSanitizerModule(bool CheckInitOrder = true,
335 StringRef BlacklistFile = StringRef(),
336 bool ZeroBaseShadow = false)
338 CheckInitOrder(CheckInitOrder || ClInitializers),
339 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
341 ZeroBaseShadow(ZeroBaseShadow) {}
342 bool runOnModule(Module &M);
343 static char ID; // Pass identification, replacement for typeid
344 virtual const char *getPassName() const {
345 return "AddressSanitizerModule";
349 void initializeCallbacks(Module &M);
351 bool ShouldInstrumentGlobal(GlobalVariable *G);
352 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
353 size_t RedzoneSize() const {
354 return RedzoneSizeForScale(Mapping.Scale);
358 SmallString<64> BlacklistFile;
361 OwningPtr<BlackList> BL;
362 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
366 ShadowMapping Mapping;
367 Function *AsanPoisonGlobals;
368 Function *AsanUnpoisonGlobals;
369 Function *AsanRegisterGlobals;
370 Function *AsanUnregisterGlobals;
373 // Stack poisoning does not play well with exception handling.
374 // When an exception is thrown, we essentially bypass the code
375 // that unpoisones the stack. This is why the run-time library has
376 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
377 // stack in the interceptor. This however does not work inside the
378 // actual function which catches the exception. Most likely because the
379 // compiler hoists the load of the shadow value somewhere too high.
380 // This causes asan to report a non-existing bug on 453.povray.
381 // It sounds like an LLVM bug.
382 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
384 AddressSanitizer &ASan;
389 ShadowMapping Mapping;
391 SmallVector<AllocaInst*, 16> AllocaVec;
392 SmallVector<Instruction*, 8> RetVec;
393 uint64_t TotalStackSize;
394 unsigned StackAlignment;
396 Function *AsanStackMallocFunc, *AsanStackFreeFunc;
397 Function *AsanPoisonStackMemoryFunc, *AsanUnpoisonStackMemoryFunc;
399 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
400 struct AllocaPoisonCall {
401 IntrinsicInst *InsBefore;
405 SmallVector<AllocaPoisonCall, 8> AllocaPoisonCallVec;
407 // Maps Value to an AllocaInst from which the Value is originated.
408 typedef DenseMap<Value*, AllocaInst*> AllocaForValueMapTy;
409 AllocaForValueMapTy AllocaForValue;
411 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
412 : F(F), ASan(ASan), DIB(*F.getParent()), C(ASan.C),
413 IntptrTy(ASan.IntptrTy), IntptrPtrTy(PointerType::get(IntptrTy, 0)),
414 Mapping(ASan.Mapping),
415 TotalStackSize(0), StackAlignment(1 << Mapping.Scale) {}
417 bool runOnFunction() {
418 if (!ClStack) return false;
419 // Collect alloca, ret, lifetime instructions etc.
420 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
421 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
422 BasicBlock *BB = *DI;
425 if (AllocaVec.empty()) return false;
427 initializeCallbacks(*F.getParent());
437 // Finds all static Alloca instructions and puts
438 // poisoned red zones around all of them.
439 // Then unpoison everything back before the function returns.
442 // ----------------------- Visitors.
443 /// \brief Collect all Ret instructions.
444 void visitReturnInst(ReturnInst &RI) {
445 RetVec.push_back(&RI);
448 /// \brief Collect Alloca instructions we want (and can) handle.
449 void visitAllocaInst(AllocaInst &AI) {
450 if (!isInterestingAlloca(AI)) return;
452 StackAlignment = std::max(StackAlignment, AI.getAlignment());
453 AllocaVec.push_back(&AI);
454 uint64_t AlignedSize = getAlignedAllocaSize(&AI);
455 TotalStackSize += AlignedSize;
458 /// \brief Collect lifetime intrinsic calls to check for use-after-scope
460 void visitIntrinsicInst(IntrinsicInst &II) {
461 if (!ASan.CheckLifetime) return;
462 Intrinsic::ID ID = II.getIntrinsicID();
463 if (ID != Intrinsic::lifetime_start &&
464 ID != Intrinsic::lifetime_end)
466 // Found lifetime intrinsic, add ASan instrumentation if necessary.
467 ConstantInt *Size = dyn_cast<ConstantInt>(II.getArgOperand(0));
468 // If size argument is undefined, don't do anything.
469 if (Size->isMinusOne()) return;
470 // Check that size doesn't saturate uint64_t and can
471 // be stored in IntptrTy.
472 const uint64_t SizeValue = Size->getValue().getLimitedValue();
473 if (SizeValue == ~0ULL ||
474 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
476 // Find alloca instruction that corresponds to llvm.lifetime argument.
477 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1));
479 bool DoPoison = (ID == Intrinsic::lifetime_end);
480 AllocaPoisonCall APC = {&II, SizeValue, DoPoison};
481 AllocaPoisonCallVec.push_back(APC);
484 // ---------------------- Helpers.
485 void initializeCallbacks(Module &M);
487 // Check if we want (and can) handle this alloca.
488 bool isInterestingAlloca(AllocaInst &AI) {
489 return (!AI.isArrayAllocation() &&
490 AI.isStaticAlloca() &&
491 AI.getAlignment() <= RedzoneSize() &&
492 AI.getAllocatedType()->isSized());
495 size_t RedzoneSize() const {
496 return RedzoneSizeForScale(Mapping.Scale);
498 uint64_t getAllocaSizeInBytes(AllocaInst *AI) {
499 Type *Ty = AI->getAllocatedType();
500 uint64_t SizeInBytes = ASan.TD->getTypeAllocSize(Ty);
503 uint64_t getAlignedSize(uint64_t SizeInBytes) {
504 size_t RZ = RedzoneSize();
505 return ((SizeInBytes + RZ - 1) / RZ) * RZ;
507 uint64_t getAlignedAllocaSize(AllocaInst *AI) {
508 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
509 return getAlignedSize(SizeInBytes);
511 /// Finds alloca where the value comes from.
512 AllocaInst *findAllocaForValue(Value *V);
513 void poisonRedZones(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB,
514 Value *ShadowBase, bool DoPoison);
515 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> IRB, bool DoPoison);
520 char AddressSanitizer::ID = 0;
521 INITIALIZE_PASS(AddressSanitizer, "asan",
522 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
524 FunctionPass *llvm::createAddressSanitizerFunctionPass(
525 bool CheckInitOrder, bool CheckUseAfterReturn, bool CheckLifetime,
526 StringRef BlacklistFile, bool ZeroBaseShadow) {
527 return new AddressSanitizer(CheckInitOrder, CheckUseAfterReturn,
528 CheckLifetime, BlacklistFile, ZeroBaseShadow);
531 char AddressSanitizerModule::ID = 0;
532 INITIALIZE_PASS(AddressSanitizerModule, "asan-module",
533 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
534 "ModulePass", false, false)
535 ModulePass *llvm::createAddressSanitizerModulePass(
536 bool CheckInitOrder, StringRef BlacklistFile, bool ZeroBaseShadow) {
537 return new AddressSanitizerModule(CheckInitOrder, BlacklistFile,
541 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
542 size_t Res = countTrailingZeros(TypeSize / 8);
543 assert(Res < kNumberOfAccessSizes);
547 // Create a constant for Str so that we can pass it to the run-time lib.
548 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
549 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
550 GlobalVariable *GV = new GlobalVariable(M, StrConst->getType(), true,
551 GlobalValue::PrivateLinkage, StrConst,
553 GV->setUnnamedAddr(true); // Ok to merge these.
554 GV->setAlignment(1); // Strings may not be merged w/o setting align 1.
558 static bool GlobalWasGeneratedByAsan(GlobalVariable *G) {
559 return G->getName().find(kAsanGenPrefix) == 0;
562 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
564 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
565 if (Mapping.Offset == 0)
567 // (Shadow >> scale) | offset
568 if (Mapping.OrShadowOffset)
569 return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
571 return IRB.CreateAdd(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
574 void AddressSanitizer::instrumentMemIntrinsicParam(
575 Instruction *OrigIns,
576 Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
577 IRBuilder<> IRB(InsertBefore);
578 if (Size->getType() != IntptrTy)
579 Size = IRB.CreateIntCast(Size, IntptrTy, false);
580 // Check the first byte.
581 instrumentAddress(OrigIns, InsertBefore, Addr, 8, IsWrite, Size);
582 // Check the last byte.
583 IRB.SetInsertPoint(InsertBefore);
584 Value *SizeMinusOne = IRB.CreateSub(Size, ConstantInt::get(IntptrTy, 1));
585 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
586 Value *AddrLast = IRB.CreateAdd(AddrLong, SizeMinusOne);
587 instrumentAddress(OrigIns, InsertBefore, AddrLast, 8, IsWrite, Size);
590 // Instrument memset/memmove/memcpy
591 bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
592 Value *Dst = MI->getDest();
593 MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
594 Value *Src = MemTran ? MemTran->getSource() : 0;
595 Value *Length = MI->getLength();
597 Constant *ConstLength = dyn_cast<Constant>(Length);
598 Instruction *InsertBefore = MI;
600 if (ConstLength->isNullValue()) return false;
602 // The size is not a constant so it could be zero -- check at run-time.
603 IRBuilder<> IRB(InsertBefore);
605 Value *Cmp = IRB.CreateICmpNE(Length,
606 Constant::getNullValue(Length->getType()));
607 InsertBefore = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
610 instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
612 instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
616 // If I is an interesting memory access, return the PointerOperand
617 // and set IsWrite. Otherwise return NULL.
618 static Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite) {
619 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
620 if (!ClInstrumentReads) return NULL;
622 return LI->getPointerOperand();
624 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
625 if (!ClInstrumentWrites) return NULL;
627 return SI->getPointerOperand();
629 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
630 if (!ClInstrumentAtomics) return NULL;
632 return RMW->getPointerOperand();
634 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
635 if (!ClInstrumentAtomics) return NULL;
637 return XCHG->getPointerOperand();
642 void AddressSanitizer::instrumentMop(Instruction *I) {
643 bool IsWrite = false;
644 Value *Addr = isInterestingMemoryAccess(I, &IsWrite);
646 if (ClOpt && ClOptGlobals) {
647 if (GlobalVariable *G = dyn_cast<GlobalVariable>(Addr)) {
648 // If initialization order checking is disabled, a simple access to a
649 // dynamically initialized global is always valid.
652 // If a global variable does not have dynamic initialization we don't
653 // have to instrument it. However, if a global does not have initailizer
654 // at all, we assume it has dynamic initializer (in other TU).
655 if (G->hasInitializer() && !DynamicallyInitializedGlobals.Contains(G))
660 Type *OrigPtrTy = Addr->getType();
661 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
663 assert(OrigTy->isSized());
664 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
666 assert((TypeSize % 8) == 0);
668 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check.
669 if (TypeSize == 8 || TypeSize == 16 ||
670 TypeSize == 32 || TypeSize == 64 || TypeSize == 128)
671 return instrumentAddress(I, I, Addr, TypeSize, IsWrite, 0);
672 // Instrument unusual size (but still multiple of 8).
673 // We can not do it with a single check, so we do 1-byte check for the first
674 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
675 // to report the actual access size.
677 Value *LastByte = IRB.CreateIntToPtr(
678 IRB.CreateAdd(IRB.CreatePointerCast(Addr, IntptrTy),
679 ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
681 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
682 instrumentAddress(I, I, Addr, 8, IsWrite, Size);
683 instrumentAddress(I, I, LastByte, 8, IsWrite, Size);
686 // Validate the result of Module::getOrInsertFunction called for an interface
687 // function of AddressSanitizer. If the instrumented module defines a function
688 // with the same name, their prototypes must match, otherwise
689 // getOrInsertFunction returns a bitcast.
690 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
691 if (isa<Function>(FuncOrBitcast)) return cast<Function>(FuncOrBitcast);
692 FuncOrBitcast->dump();
693 report_fatal_error("trying to redefine an AddressSanitizer "
694 "interface function");
697 Instruction *AddressSanitizer::generateCrashCode(
698 Instruction *InsertBefore, Value *Addr,
699 bool IsWrite, size_t AccessSizeIndex, Value *SizeArgument) {
700 IRBuilder<> IRB(InsertBefore);
701 CallInst *Call = SizeArgument
702 ? IRB.CreateCall2(AsanErrorCallbackSized[IsWrite], Addr, SizeArgument)
703 : IRB.CreateCall(AsanErrorCallback[IsWrite][AccessSizeIndex], Addr);
705 // We don't do Call->setDoesNotReturn() because the BB already has
706 // UnreachableInst at the end.
707 // This EmptyAsm is required to avoid callback merge.
708 IRB.CreateCall(EmptyAsm);
712 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
715 size_t Granularity = 1 << Mapping.Scale;
716 // Addr & (Granularity - 1)
717 Value *LastAccessedByte = IRB.CreateAnd(
718 AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
719 // (Addr & (Granularity - 1)) + size - 1
720 if (TypeSize / 8 > 1)
721 LastAccessedByte = IRB.CreateAdd(
722 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
723 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
724 LastAccessedByte = IRB.CreateIntCast(
725 LastAccessedByte, ShadowValue->getType(), false);
726 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
727 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
730 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
731 Instruction *InsertBefore,
732 Value *Addr, uint32_t TypeSize,
733 bool IsWrite, Value *SizeArgument) {
734 IRBuilder<> IRB(InsertBefore);
735 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
737 Type *ShadowTy = IntegerType::get(
738 *C, std::max(8U, TypeSize >> Mapping.Scale));
739 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
740 Value *ShadowPtr = memToShadow(AddrLong, IRB);
741 Value *CmpVal = Constant::getNullValue(ShadowTy);
742 Value *ShadowValue = IRB.CreateLoad(
743 IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
745 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
746 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
747 size_t Granularity = 1 << Mapping.Scale;
748 TerminatorInst *CrashTerm = 0;
750 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
751 TerminatorInst *CheckTerm =
752 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
753 assert(dyn_cast<BranchInst>(CheckTerm)->isUnconditional());
754 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
755 IRB.SetInsertPoint(CheckTerm);
756 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
757 BasicBlock *CrashBlock =
758 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
759 CrashTerm = new UnreachableInst(*C, CrashBlock);
760 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
761 ReplaceInstWithInst(CheckTerm, NewTerm);
763 CrashTerm = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), true);
766 Instruction *Crash = generateCrashCode(
767 CrashTerm, AddrLong, IsWrite, AccessSizeIndex, SizeArgument);
768 Crash->setDebugLoc(OrigIns->getDebugLoc());
771 void AddressSanitizerModule::createInitializerPoisonCalls(
772 Module &M, GlobalValue *ModuleName) {
773 // We do all of our poisoning and unpoisoning within _GLOBAL__I_a.
774 Function *GlobalInit = M.getFunction("_GLOBAL__I_a");
775 // If that function is not present, this TU contains no globals, or they have
776 // all been optimized away
780 // Set up the arguments to our poison/unpoison functions.
781 IRBuilder<> IRB(GlobalInit->begin()->getFirstInsertionPt());
783 // Add a call to poison all external globals before the given function starts.
784 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
785 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
787 // Add calls to unpoison all globals before each return instruction.
788 for (Function::iterator I = GlobalInit->begin(), E = GlobalInit->end();
790 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator())) {
791 CallInst::Create(AsanUnpoisonGlobals, "", RI);
796 bool AddressSanitizerModule::ShouldInstrumentGlobal(GlobalVariable *G) {
797 Type *Ty = cast<PointerType>(G->getType())->getElementType();
798 DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
800 if (BL->isIn(*G)) return false;
801 if (!Ty->isSized()) return false;
802 if (!G->hasInitializer()) return false;
803 if (GlobalWasGeneratedByAsan(G)) return false; // Our own global.
804 // Touch only those globals that will not be defined in other modules.
805 // Don't handle ODR type linkages since other modules may be built w/o asan.
806 if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
807 G->getLinkage() != GlobalVariable::PrivateLinkage &&
808 G->getLinkage() != GlobalVariable::InternalLinkage)
810 // Two problems with thread-locals:
811 // - The address of the main thread's copy can't be computed at link-time.
812 // - Need to poison all copies, not just the main thread's one.
813 if (G->isThreadLocal())
815 // For now, just ignore this Alloca if the alignment is large.
816 if (G->getAlignment() > RedzoneSize()) return false;
818 // Ignore all the globals with the names starting with "\01L_OBJC_".
819 // Many of those are put into the .cstring section. The linker compresses
820 // that section by removing the spare \0s after the string terminator, so
821 // our redzones get broken.
822 if ((G->getName().find("\01L_OBJC_") == 0) ||
823 (G->getName().find("\01l_OBJC_") == 0)) {
824 DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
828 if (G->hasSection()) {
829 StringRef Section(G->getSection());
830 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
831 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
833 if ((Section.find("__OBJC,") == 0) ||
834 (Section.find("__DATA, __objc_") == 0)) {
835 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
838 // See http://code.google.com/p/address-sanitizer/issues/detail?id=32
839 // Constant CFString instances are compiled in the following way:
840 // -- the string buffer is emitted into
841 // __TEXT,__cstring,cstring_literals
842 // -- the constant NSConstantString structure referencing that buffer
843 // is placed into __DATA,__cfstring
844 // Therefore there's no point in placing redzones into __DATA,__cfstring.
845 // Moreover, it causes the linker to crash on OS X 10.7
846 if (Section.find("__DATA,__cfstring") == 0) {
847 DEBUG(dbgs() << "Ignoring CFString: " << *G);
855 void AddressSanitizerModule::initializeCallbacks(Module &M) {
857 // Declare our poisoning and unpoisoning functions.
858 AsanPoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
859 kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, NULL));
860 AsanPoisonGlobals->setLinkage(Function::ExternalLinkage);
861 AsanUnpoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
862 kAsanUnpoisonGlobalsName, IRB.getVoidTy(), NULL));
863 AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage);
864 // Declare functions that register/unregister globals.
865 AsanRegisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
866 kAsanRegisterGlobalsName, IRB.getVoidTy(),
867 IntptrTy, IntptrTy, NULL));
868 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
869 AsanUnregisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
870 kAsanUnregisterGlobalsName,
871 IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
872 AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage);
875 // This function replaces all global variables with new variables that have
876 // trailing redzones. It also creates a function that poisons
877 // redzones and inserts this function into llvm.global_ctors.
878 bool AddressSanitizerModule::runOnModule(Module &M) {
879 if (!ClGlobals) return false;
880 TD = getAnalysisIfAvailable<DataLayout>();
883 BL.reset(new BlackList(BlacklistFile));
884 if (BL->isIn(M)) return false;
885 C = &(M.getContext());
886 int LongSize = TD->getPointerSizeInBits();
887 IntptrTy = Type::getIntNTy(*C, LongSize);
888 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
889 initializeCallbacks(M);
890 DynamicallyInitializedGlobals.Init(M);
892 SmallVector<GlobalVariable *, 16> GlobalsToChange;
894 for (Module::GlobalListType::iterator G = M.global_begin(),
895 E = M.global_end(); G != E; ++G) {
896 if (ShouldInstrumentGlobal(G))
897 GlobalsToChange.push_back(G);
900 size_t n = GlobalsToChange.size();
901 if (n == 0) return false;
903 // A global is described by a structure
906 // size_t size_with_redzone;
908 // const char *module_name;
909 // size_t has_dynamic_init;
910 // We initialize an array of such structures and pass it to a run-time call.
911 StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
913 IntptrTy, IntptrTy, NULL);
914 SmallVector<Constant *, 16> Initializers(n), DynamicInit;
917 Function *CtorFunc = M.getFunction(kAsanModuleCtorName);
919 IRBuilder<> IRB(CtorFunc->getEntryBlock().getTerminator());
921 bool HasDynamicallyInitializedGlobals = false;
923 GlobalVariable *ModuleName = createPrivateGlobalForString(
924 M, M.getModuleIdentifier());
925 // We shouldn't merge same module names, as this string serves as unique
926 // module ID in runtime.
927 ModuleName->setUnnamedAddr(false);
929 for (size_t i = 0; i < n; i++) {
930 static const uint64_t kMaxGlobalRedzone = 1 << 18;
931 GlobalVariable *G = GlobalsToChange[i];
932 PointerType *PtrTy = cast<PointerType>(G->getType());
933 Type *Ty = PtrTy->getElementType();
934 uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
935 uint64_t MinRZ = RedzoneSize();
936 // MinRZ <= RZ <= kMaxGlobalRedzone
937 // and trying to make RZ to be ~ 1/4 of SizeInBytes.
938 uint64_t RZ = std::max(MinRZ,
939 std::min(kMaxGlobalRedzone,
940 (SizeInBytes / MinRZ / 4) * MinRZ));
941 uint64_t RightRedzoneSize = RZ;
943 if (SizeInBytes % MinRZ)
944 RightRedzoneSize += MinRZ - (SizeInBytes % MinRZ);
945 assert(((RightRedzoneSize + SizeInBytes) % MinRZ) == 0);
946 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
947 // Determine whether this global should be poisoned in initialization.
948 bool GlobalHasDynamicInitializer =
949 DynamicallyInitializedGlobals.Contains(G);
950 // Don't check initialization order if this global is blacklisted.
951 GlobalHasDynamicInitializer &= !BL->isInInit(*G);
953 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
954 Constant *NewInitializer = ConstantStruct::get(
955 NewTy, G->getInitializer(),
956 Constant::getNullValue(RightRedZoneTy), NULL);
958 GlobalVariable *Name = createPrivateGlobalForString(M, G->getName());
960 // Create a new global variable with enough space for a redzone.
961 GlobalVariable *NewGlobal = new GlobalVariable(
962 M, NewTy, G->isConstant(), G->getLinkage(),
963 NewInitializer, "", G, G->getThreadLocalMode());
964 NewGlobal->copyAttributesFrom(G);
965 NewGlobal->setAlignment(MinRZ);
968 Indices2[0] = IRB.getInt32(0);
969 Indices2[1] = IRB.getInt32(0);
971 G->replaceAllUsesWith(
972 ConstantExpr::getGetElementPtr(NewGlobal, Indices2, true));
973 NewGlobal->takeName(G);
974 G->eraseFromParent();
976 Initializers[i] = ConstantStruct::get(
978 ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
979 ConstantInt::get(IntptrTy, SizeInBytes),
980 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
981 ConstantExpr::getPointerCast(Name, IntptrTy),
982 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
983 ConstantInt::get(IntptrTy, GlobalHasDynamicInitializer),
986 // Populate the first and last globals declared in this TU.
987 if (CheckInitOrder && GlobalHasDynamicInitializer)
988 HasDynamicallyInitializedGlobals = true;
990 DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
993 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
994 GlobalVariable *AllGlobals = new GlobalVariable(
995 M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage,
996 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
998 // Create calls for poisoning before initializers run and unpoisoning after.
999 if (CheckInitOrder && HasDynamicallyInitializedGlobals)
1000 createInitializerPoisonCalls(M, ModuleName);
1001 IRB.CreateCall2(AsanRegisterGlobals,
1002 IRB.CreatePointerCast(AllGlobals, IntptrTy),
1003 ConstantInt::get(IntptrTy, n));
1005 // We also need to unregister globals at the end, e.g. when a shared library
1007 Function *AsanDtorFunction = Function::Create(
1008 FunctionType::get(Type::getVoidTy(*C), false),
1009 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
1010 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
1011 IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB));
1012 IRB_Dtor.CreateCall2(AsanUnregisterGlobals,
1013 IRB.CreatePointerCast(AllGlobals, IntptrTy),
1014 ConstantInt::get(IntptrTy, n));
1015 appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndCtorPriority);
1021 void AddressSanitizer::initializeCallbacks(Module &M) {
1022 IRBuilder<> IRB(*C);
1023 // Create __asan_report* callbacks.
1024 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
1025 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
1026 AccessSizeIndex++) {
1027 // IsWrite and TypeSize are encoded in the function name.
1028 std::string FunctionName = std::string(kAsanReportErrorTemplate) +
1029 (AccessIsWrite ? "store" : "load") + itostr(1 << AccessSizeIndex);
1030 // If we are merging crash callbacks, they have two parameters.
1031 AsanErrorCallback[AccessIsWrite][AccessSizeIndex] =
1032 checkInterfaceFunction(M.getOrInsertFunction(
1033 FunctionName, IRB.getVoidTy(), IntptrTy, NULL));
1036 AsanErrorCallbackSized[0] = checkInterfaceFunction(M.getOrInsertFunction(
1037 kAsanReportLoadN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1038 AsanErrorCallbackSized[1] = checkInterfaceFunction(M.getOrInsertFunction(
1039 kAsanReportStoreN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1041 AsanHandleNoReturnFunc = checkInterfaceFunction(M.getOrInsertFunction(
1042 kAsanHandleNoReturnName, IRB.getVoidTy(), NULL));
1043 // We insert an empty inline asm after __asan_report* to avoid callback merge.
1044 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
1045 StringRef(""), StringRef(""),
1046 /*hasSideEffects=*/true);
1049 void AddressSanitizer::emitShadowMapping(Module &M, IRBuilder<> &IRB) const {
1050 // Tell the values of mapping offset and scale to the run-time.
1051 GlobalValue *asan_mapping_offset =
1052 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1053 ConstantInt::get(IntptrTy, Mapping.Offset),
1054 kAsanMappingOffsetName);
1055 // Read the global, otherwise it may be optimized away.
1056 IRB.CreateLoad(asan_mapping_offset, true);
1058 GlobalValue *asan_mapping_scale =
1059 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1060 ConstantInt::get(IntptrTy, Mapping.Scale),
1061 kAsanMappingScaleName);
1062 // Read the global, otherwise it may be optimized away.
1063 IRB.CreateLoad(asan_mapping_scale, true);
1067 bool AddressSanitizer::doInitialization(Module &M) {
1068 // Initialize the private fields. No one has accessed them before.
1069 TD = getAnalysisIfAvailable<DataLayout>();
1073 BL.reset(new BlackList(BlacklistFile));
1074 DynamicallyInitializedGlobals.Init(M);
1076 C = &(M.getContext());
1077 LongSize = TD->getPointerSizeInBits();
1078 IntptrTy = Type::getIntNTy(*C, LongSize);
1080 AsanCtorFunction = Function::Create(
1081 FunctionType::get(Type::getVoidTy(*C), false),
1082 GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
1083 BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
1084 // call __asan_init in the module ctor.
1085 IRBuilder<> IRB(ReturnInst::Create(*C, AsanCtorBB));
1086 AsanInitFunction = checkInterfaceFunction(
1087 M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
1088 AsanInitFunction->setLinkage(Function::ExternalLinkage);
1089 IRB.CreateCall(AsanInitFunction);
1091 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
1092 emitShadowMapping(M, IRB);
1094 appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority);
1098 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
1099 // For each NSObject descendant having a +load method, this method is invoked
1100 // by the ObjC runtime before any of the static constructors is called.
1101 // Therefore we need to instrument such methods with a call to __asan_init
1102 // at the beginning in order to initialize our runtime before any access to
1103 // the shadow memory.
1104 // We cannot just ignore these methods, because they may call other
1105 // instrumented functions.
1106 if (F.getName().find(" load]") != std::string::npos) {
1107 IRBuilder<> IRB(F.begin()->begin());
1108 IRB.CreateCall(AsanInitFunction);
1114 bool AddressSanitizer::runOnFunction(Function &F) {
1115 if (BL->isIn(F)) return false;
1116 if (&F == AsanCtorFunction) return false;
1117 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
1118 DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
1119 initializeCallbacks(*F.getParent());
1121 // If needed, insert __asan_init before checking for SanitizeAddress attr.
1122 maybeInsertAsanInitAtFunctionEntry(F);
1124 if (!F.hasFnAttribute(Attribute::SanitizeAddress))
1127 if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
1130 // We want to instrument every address only once per basic block (unless there
1131 // are calls between uses).
1132 SmallSet<Value*, 16> TempsToInstrument;
1133 SmallVector<Instruction*, 16> ToInstrument;
1134 SmallVector<Instruction*, 8> NoReturnCalls;
1138 // Fill the set of memory operations to instrument.
1139 for (Function::iterator FI = F.begin(), FE = F.end();
1141 TempsToInstrument.clear();
1142 int NumInsnsPerBB = 0;
1143 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1145 if (LooksLikeCodeInBug11395(BI)) return false;
1146 if (Value *Addr = isInterestingMemoryAccess(BI, &IsWrite)) {
1147 if (ClOpt && ClOptSameTemp) {
1148 if (!TempsToInstrument.insert(Addr))
1149 continue; // We've seen this temp in the current BB.
1151 } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
1154 if (isa<AllocaInst>(BI))
1158 // A call inside BB.
1159 TempsToInstrument.clear();
1160 if (CS.doesNotReturn())
1161 NoReturnCalls.push_back(CS.getInstruction());
1165 ToInstrument.push_back(BI);
1167 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB)
1172 Function *UninstrumentedDuplicate = 0;
1173 bool LikelyToInstrument =
1174 !NoReturnCalls.empty() || !ToInstrument.empty() || (NumAllocas > 0);
1175 if (ClKeepUninstrumented && LikelyToInstrument) {
1176 ValueToValueMapTy VMap;
1177 UninstrumentedDuplicate = CloneFunction(&F, VMap, false);
1178 UninstrumentedDuplicate->removeFnAttr(Attribute::SanitizeAddress);
1179 UninstrumentedDuplicate->setName("NOASAN_" + F.getName());
1180 F.getParent()->getFunctionList().push_back(UninstrumentedDuplicate);
1184 int NumInstrumented = 0;
1185 for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
1186 Instruction *Inst = ToInstrument[i];
1187 if (ClDebugMin < 0 || ClDebugMax < 0 ||
1188 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
1189 if (isInterestingMemoryAccess(Inst, &IsWrite))
1190 instrumentMop(Inst);
1192 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
1197 FunctionStackPoisoner FSP(F, *this);
1198 bool ChangedStack = FSP.runOnFunction();
1200 // We must unpoison the stack before every NoReturn call (throw, _exit, etc).
1201 // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37
1202 for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) {
1203 Instruction *CI = NoReturnCalls[i];
1204 IRBuilder<> IRB(CI);
1205 IRB.CreateCall(AsanHandleNoReturnFunc);
1208 bool res = NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty();
1209 DEBUG(dbgs() << "ASAN done instrumenting: " << res << " " << F << "\n");
1211 if (ClKeepUninstrumented) {
1213 // No instrumentation is done, no need for the duplicate.
1214 if (UninstrumentedDuplicate)
1215 UninstrumentedDuplicate->eraseFromParent();
1217 // The function was instrumented. We must have the duplicate.
1218 assert(UninstrumentedDuplicate);
1219 UninstrumentedDuplicate->setSection("NOASAN");
1220 assert(!F.hasSection());
1221 F.setSection("ASAN");
1228 static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
1229 if (ShadowRedzoneSize == 1) return PoisonByte;
1230 if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
1231 if (ShadowRedzoneSize == 4)
1232 return (PoisonByte << 24) + (PoisonByte << 16) +
1233 (PoisonByte << 8) + (PoisonByte);
1234 llvm_unreachable("ShadowRedzoneSize is either 1, 2 or 4");
1237 static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
1240 size_t ShadowGranularity,
1242 for (size_t i = 0; i < RZSize;
1243 i+= ShadowGranularity, Shadow++) {
1244 if (i + ShadowGranularity <= Size) {
1245 *Shadow = 0; // fully addressable
1246 } else if (i >= Size) {
1247 *Shadow = Magic; // unaddressable
1249 *Shadow = Size - i; // first Size-i bytes are addressable
1254 // Workaround for bug 11395: we don't want to instrument stack in functions
1255 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
1256 // FIXME: remove once the bug 11395 is fixed.
1257 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
1258 if (LongSize != 32) return false;
1259 CallInst *CI = dyn_cast<CallInst>(I);
1260 if (!CI || !CI->isInlineAsm()) return false;
1261 if (CI->getNumArgOperands() <= 5) return false;
1262 // We have inline assembly with quite a few arguments.
1266 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
1267 IRBuilder<> IRB(*C);
1268 AsanStackMallocFunc = checkInterfaceFunction(M.getOrInsertFunction(
1269 kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL));
1270 AsanStackFreeFunc = checkInterfaceFunction(M.getOrInsertFunction(
1271 kAsanStackFreeName, IRB.getVoidTy(),
1272 IntptrTy, IntptrTy, IntptrTy, NULL));
1273 AsanPoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1274 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1275 AsanUnpoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1276 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1279 void FunctionStackPoisoner::poisonRedZones(
1280 const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB, Value *ShadowBase,
1282 size_t ShadowRZSize = RedzoneSize() >> Mapping.Scale;
1283 assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
1284 Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
1285 Type *RZPtrTy = PointerType::get(RZTy, 0);
1287 Value *PoisonLeft = ConstantInt::get(RZTy,
1288 ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
1289 Value *PoisonMid = ConstantInt::get(RZTy,
1290 ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
1291 Value *PoisonRight = ConstantInt::get(RZTy,
1292 ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
1294 // poison the first red zone.
1295 IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
1297 // poison all other red zones.
1298 uint64_t Pos = RedzoneSize();
1299 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1300 AllocaInst *AI = AllocaVec[i];
1301 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1302 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1303 assert(AlignedSize - SizeInBytes < RedzoneSize());
1308 assert(ShadowBase->getType() == IntptrTy);
1309 if (SizeInBytes < AlignedSize) {
1310 // Poison the partial redzone at right
1311 Ptr = IRB.CreateAdd(
1312 ShadowBase, ConstantInt::get(IntptrTy,
1313 (Pos >> Mapping.Scale) - ShadowRZSize));
1314 size_t AddressableBytes = RedzoneSize() - (AlignedSize - SizeInBytes);
1315 uint32_t Poison = 0;
1317 PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
1319 1ULL << Mapping.Scale,
1320 kAsanStackPartialRedzoneMagic);
1322 ASan.TD->isLittleEndian()
1323 ? support::endian::byte_swap<uint32_t, support::little>(Poison)
1324 : support::endian::byte_swap<uint32_t, support::big>(Poison);
1326 Value *PartialPoison = ConstantInt::get(RZTy, Poison);
1327 IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1330 // Poison the full redzone at right.
1331 Ptr = IRB.CreateAdd(ShadowBase,
1332 ConstantInt::get(IntptrTy, Pos >> Mapping.Scale));
1333 bool LastAlloca = (i == AllocaVec.size() - 1);
1334 Value *Poison = LastAlloca ? PoisonRight : PoisonMid;
1335 IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1337 Pos += RedzoneSize();
1341 void FunctionStackPoisoner::poisonStack() {
1342 uint64_t LocalStackSize = TotalStackSize +
1343 (AllocaVec.size() + 1) * RedzoneSize();
1345 bool DoStackMalloc = ASan.CheckUseAfterReturn
1346 && LocalStackSize <= kMaxStackMallocSize;
1348 assert(AllocaVec.size() > 0);
1349 Instruction *InsBefore = AllocaVec[0];
1350 IRBuilder<> IRB(InsBefore);
1353 Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
1354 AllocaInst *MyAlloca =
1355 new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
1356 if (ClRealignStack && StackAlignment < RedzoneSize())
1357 StackAlignment = RedzoneSize();
1358 MyAlloca->setAlignment(StackAlignment);
1359 assert(MyAlloca->isStaticAlloca());
1360 Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
1361 Value *LocalStackBase = OrigStackBase;
1363 if (DoStackMalloc) {
1364 LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc,
1365 ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
1368 // This string will be parsed by the run-time (DescribeAddressIfStack).
1369 SmallString<2048> StackDescriptionStorage;
1370 raw_svector_ostream StackDescription(StackDescriptionStorage);
1371 StackDescription << AllocaVec.size() << " ";
1373 // Insert poison calls for lifetime intrinsics for alloca.
1374 bool HavePoisonedAllocas = false;
1375 for (size_t i = 0, n = AllocaPoisonCallVec.size(); i < n; i++) {
1376 const AllocaPoisonCall &APC = AllocaPoisonCallVec[i];
1377 IntrinsicInst *II = APC.InsBefore;
1378 AllocaInst *AI = findAllocaForValue(II->getArgOperand(1));
1380 IRBuilder<> IRB(II);
1381 poisonAlloca(AI, APC.Size, IRB, APC.DoPoison);
1382 HavePoisonedAllocas |= APC.DoPoison;
1385 uint64_t Pos = RedzoneSize();
1386 // Replace Alloca instructions with base+offset.
1387 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1388 AllocaInst *AI = AllocaVec[i];
1389 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1390 StringRef Name = AI->getName();
1391 StackDescription << Pos << " " << SizeInBytes << " "
1392 << Name.size() << " " << Name << " ";
1393 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1394 assert((AlignedSize % RedzoneSize()) == 0);
1395 Value *NewAllocaPtr = IRB.CreateIntToPtr(
1396 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
1398 replaceDbgDeclareForAlloca(AI, NewAllocaPtr, DIB);
1399 AI->replaceAllUsesWith(NewAllocaPtr);
1400 Pos += AlignedSize + RedzoneSize();
1402 assert(Pos == LocalStackSize);
1404 // The left-most redzone has enough space for at least 4 pointers.
1405 // Write the Magic value to redzone[0].
1406 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
1407 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
1409 // Write the frame description constant to redzone[1].
1410 Value *BasePlus1 = IRB.CreateIntToPtr(
1411 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, ASan.LongSize/8)),
1413 GlobalVariable *StackDescriptionGlobal =
1414 createPrivateGlobalForString(*F.getParent(), StackDescription.str());
1415 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal,
1417 IRB.CreateStore(Description, BasePlus1);
1418 // Write the PC to redzone[2].
1419 Value *BasePlus2 = IRB.CreateIntToPtr(
1420 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy,
1421 2 * ASan.LongSize/8)),
1423 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
1425 // Poison the stack redzones at the entry.
1426 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
1427 poisonRedZones(AllocaVec, IRB, ShadowBase, true);
1429 // Unpoison the stack before all ret instructions.
1430 for (size_t i = 0, n = RetVec.size(); i < n; i++) {
1431 Instruction *Ret = RetVec[i];
1432 IRBuilder<> IRBRet(Ret);
1433 // Mark the current frame as retired.
1434 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
1436 // Unpoison the stack.
1437 poisonRedZones(AllocaVec, IRBRet, ShadowBase, false);
1438 if (DoStackMalloc) {
1439 // In use-after-return mode, mark the whole stack frame unaddressable.
1440 IRBRet.CreateCall3(AsanStackFreeFunc, LocalStackBase,
1441 ConstantInt::get(IntptrTy, LocalStackSize),
1443 } else if (HavePoisonedAllocas) {
1444 // If we poisoned some allocas in llvm.lifetime analysis,
1445 // unpoison whole stack frame now.
1446 assert(LocalStackBase == OrigStackBase);
1447 poisonAlloca(LocalStackBase, LocalStackSize, IRBRet, false);
1451 // We are done. Remove the old unused alloca instructions.
1452 for (size_t i = 0, n = AllocaVec.size(); i < n; i++)
1453 AllocaVec[i]->eraseFromParent();
1456 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
1457 IRBuilder<> IRB, bool DoPoison) {
1458 // For now just insert the call to ASan runtime.
1459 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
1460 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
1461 IRB.CreateCall2(DoPoison ? AsanPoisonStackMemoryFunc
1462 : AsanUnpoisonStackMemoryFunc,
1466 // Handling llvm.lifetime intrinsics for a given %alloca:
1467 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
1468 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
1469 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
1470 // could be poisoned by previous llvm.lifetime.end instruction, as the
1471 // variable may go in and out of scope several times, e.g. in loops).
1472 // (3) if we poisoned at least one %alloca in a function,
1473 // unpoison the whole stack frame at function exit.
1475 AllocaInst *FunctionStackPoisoner::findAllocaForValue(Value *V) {
1476 if (AllocaInst *AI = dyn_cast<AllocaInst>(V))
1477 // We're intested only in allocas we can handle.
1478 return isInterestingAlloca(*AI) ? AI : 0;
1479 // See if we've already calculated (or started to calculate) alloca for a
1481 AllocaForValueMapTy::iterator I = AllocaForValue.find(V);
1482 if (I != AllocaForValue.end())
1484 // Store 0 while we're calculating alloca for value V to avoid
1485 // infinite recursion if the value references itself.
1486 AllocaForValue[V] = 0;
1487 AllocaInst *Res = 0;
1488 if (CastInst *CI = dyn_cast<CastInst>(V))
1489 Res = findAllocaForValue(CI->getOperand(0));
1490 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1491 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1492 Value *IncValue = PN->getIncomingValue(i);
1493 // Allow self-referencing phi-nodes.
1494 if (IncValue == PN) continue;
1495 AllocaInst *IncValueAI = findAllocaForValue(IncValue);
1496 // AI for incoming values should exist and should all be equal.
1497 if (IncValueAI == 0 || (Res != 0 && IncValueAI != Res))
1503 AllocaForValue[V] = Res;