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/Cloning.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Transforms/Utils/ModuleUtils.h"
49 #include "llvm/Transforms/Utils/SpecialCaseList.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 *const kAsanModuleCtorName = "asan.module_ctor";
67 static const char *const kAsanModuleDtorName = "asan.module_dtor";
68 static const int kAsanCtorAndCtorPriority = 1;
69 static const char *const kAsanReportErrorTemplate = "__asan_report_";
70 static const char *const kAsanReportLoadN = "__asan_report_load_n";
71 static const char *const kAsanReportStoreN = "__asan_report_store_n";
72 static const char *const kAsanRegisterGlobalsName = "__asan_register_globals";
73 static const char *const kAsanUnregisterGlobalsName = "__asan_unregister_globals";
74 static const char *const kAsanPoisonGlobalsName = "__asan_before_dynamic_init";
75 static const char *const kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init";
76 static const char *const kAsanInitName = "__asan_init_v3";
77 static const char *const kAsanHandleNoReturnName = "__asan_handle_no_return";
78 static const char *const kAsanMappingOffsetName = "__asan_mapping_offset";
79 static const char *const kAsanMappingScaleName = "__asan_mapping_scale";
80 static const char *const kAsanStackMallocName = "__asan_stack_malloc";
81 static const char *const kAsanStackFreeName = "__asan_stack_free";
82 static const char *const kAsanGenPrefix = "__asan_gen_";
83 static const char *const kAsanPoisonStackMemoryName =
84 "__asan_poison_stack_memory";
85 static const char *const kAsanUnpoisonStackMemoryName =
86 "__asan_unpoison_stack_memory";
88 static const int kAsanStackLeftRedzoneMagic = 0xf1;
89 static const int kAsanStackMidRedzoneMagic = 0xf2;
90 static const int kAsanStackRightRedzoneMagic = 0xf3;
91 static const int kAsanStackPartialRedzoneMagic = 0xf4;
93 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
94 static const size_t kNumberOfAccessSizes = 5;
96 // Command-line flags.
98 // This flag may need to be replaced with -f[no-]asan-reads.
99 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
100 cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
101 static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
102 cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
103 static cl::opt<bool> ClInstrumentAtomics("asan-instrument-atomics",
104 cl::desc("instrument atomic instructions (rmw, cmpxchg)"),
105 cl::Hidden, cl::init(true));
106 static cl::opt<bool> ClAlwaysSlowPath("asan-always-slow-path",
107 cl::desc("use instrumentation with slow path for all accesses"),
108 cl::Hidden, cl::init(false));
109 // This flag limits the number of instructions to be instrumented
110 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
111 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
113 static cl::opt<int> ClMaxInsnsToInstrumentPerBB("asan-max-ins-per-bb",
115 cl::desc("maximal number of instructions to instrument in any given BB"),
117 // This flag may need to be replaced with -f[no]asan-stack.
118 static cl::opt<bool> ClStack("asan-stack",
119 cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
120 // This flag may need to be replaced with -f[no]asan-use-after-return.
121 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
122 cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
123 // This flag may need to be replaced with -f[no]asan-globals.
124 static cl::opt<bool> ClGlobals("asan-globals",
125 cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
126 static cl::opt<bool> ClInitializers("asan-initialization-order",
127 cl::desc("Handle C++ initializer order"), cl::Hidden, cl::init(false));
128 static cl::opt<bool> ClMemIntrin("asan-memintrin",
129 cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
130 static cl::opt<bool> ClRealignStack("asan-realign-stack",
131 cl::desc("Realign stack to 32"), cl::Hidden, cl::init(true));
132 static cl::opt<std::string> ClBlacklistFile("asan-blacklist",
133 cl::desc("File containing the list of objects to ignore "
134 "during instrumentation"), cl::Hidden);
136 // This is an experimental feature that will allow to choose between
137 // instrumented and non-instrumented code at link-time.
138 // If this option is on, just before instrumenting a function we create its
139 // clone; if the function is not changed by asan the clone is deleted.
140 // If we end up with a clone, we put the instrumented function into a section
141 // called "ASAN" and the uninstrumented function into a section called "NOASAN".
143 // This is still a prototype, we need to figure out a way to keep two copies of
144 // a function so that the linker can easily choose one of them.
145 static cl::opt<bool> ClKeepUninstrumented("asan-keep-uninstrumented-functions",
146 cl::desc("Keep uninstrumented copies of functions"),
147 cl::Hidden, cl::init(false));
149 // These flags allow to change the shadow mapping.
150 // The shadow mapping looks like
151 // Shadow = (Mem >> scale) + (1 << offset_log)
152 static cl::opt<int> ClMappingScale("asan-mapping-scale",
153 cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
154 static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
155 cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
156 static cl::opt<bool> ClShort64BitOffset("asan-short-64bit-mapping-offset",
157 cl::desc("Use short immediate constant as the mapping offset for 64bit"),
158 cl::Hidden, cl::init(true));
160 // Optimization flags. Not user visible, used mostly for testing
161 // and benchmarking the tool.
162 static cl::opt<bool> ClOpt("asan-opt",
163 cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
164 static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
165 cl::desc("Instrument the same temp just once"), cl::Hidden,
167 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
168 cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
170 static cl::opt<bool> ClCheckLifetime("asan-check-lifetime",
171 cl::desc("Use llvm.lifetime intrinsics to insert extra checks"),
172 cl::Hidden, cl::init(false));
175 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
177 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
178 cl::Hidden, cl::init(0));
179 static cl::opt<std::string> ClDebugFunc("asan-debug-func",
180 cl::Hidden, cl::desc("Debug func"));
181 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
182 cl::Hidden, cl::init(-1));
183 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
184 cl::Hidden, cl::init(-1));
187 /// A set of dynamically initialized globals extracted from metadata.
188 class SetOfDynamicallyInitializedGlobals {
190 void Init(Module& M) {
191 // Clang generates metadata identifying all dynamically initialized globals.
192 NamedMDNode *DynamicGlobals =
193 M.getNamedMetadata("llvm.asan.dynamically_initialized_globals");
196 for (int i = 0, n = DynamicGlobals->getNumOperands(); i < n; ++i) {
197 MDNode *MDN = DynamicGlobals->getOperand(i);
198 assert(MDN->getNumOperands() == 1);
199 Value *VG = MDN->getOperand(0);
200 // The optimizer may optimize away a global entirely, in which case we
201 // cannot instrument access to it.
204 DynInitGlobals.insert(cast<GlobalVariable>(VG));
207 bool Contains(GlobalVariable *G) { return DynInitGlobals.count(G) != 0; }
209 SmallSet<GlobalValue*, 32> DynInitGlobals;
212 /// This struct defines the shadow mapping using the rule:
213 /// shadow = (mem >> Scale) ADD-or-OR Offset.
214 struct ShadowMapping {
220 static ShadowMapping getShadowMapping(const Module &M, int LongSize,
221 bool ZeroBaseShadow) {
222 llvm::Triple TargetTriple(M.getTargetTriple());
223 bool IsAndroid = TargetTriple.getEnvironment() == llvm::Triple::Android;
224 bool IsMacOSX = TargetTriple.getOS() == llvm::Triple::MacOSX;
225 bool IsPPC64 = TargetTriple.getArch() == llvm::Triple::ppc64;
226 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
227 bool IsMIPS32 = TargetTriple.getArch() == llvm::Triple::mips ||
228 TargetTriple.getArch() == llvm::Triple::mipsel;
230 ShadowMapping Mapping;
232 // OR-ing shadow offset if more efficient (at least on x86),
233 // but on ppc64 we have to use add since the shadow offset is not neccesary
234 // 1/8-th of the address space.
235 Mapping.OrShadowOffset = !IsPPC64 && !ClShort64BitOffset;
237 Mapping.Offset = (IsAndroid || ZeroBaseShadow) ? 0 :
239 (IsMIPS32 ? kMIPS32_ShadowOffset32 : kDefaultShadowOffset32) :
240 IsPPC64 ? kPPC64_ShadowOffset64 : kDefaultShadowOffset64);
241 if (!ZeroBaseShadow && ClShort64BitOffset && IsX86_64 && !IsMacOSX) {
242 assert(LongSize == 64);
243 Mapping.Offset = kDefaultShort64bitShadowOffset;
245 if (!ZeroBaseShadow && ClMappingOffsetLog >= 0) {
246 // Zero offset log is the special case.
247 Mapping.Offset = (ClMappingOffsetLog == 0) ? 0 : 1ULL << ClMappingOffsetLog;
250 Mapping.Scale = kDefaultShadowScale;
251 if (ClMappingScale) {
252 Mapping.Scale = ClMappingScale;
258 static size_t RedzoneSizeForScale(int MappingScale) {
259 // Redzone used for stack and globals is at least 32 bytes.
260 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
261 return std::max(32U, 1U << MappingScale);
264 /// AddressSanitizer: instrument the code in module to find memory bugs.
265 struct AddressSanitizer : public FunctionPass {
266 AddressSanitizer(bool CheckInitOrder = true,
267 bool CheckUseAfterReturn = false,
268 bool CheckLifetime = false,
269 StringRef BlacklistFile = StringRef(),
270 bool ZeroBaseShadow = false)
272 CheckInitOrder(CheckInitOrder || ClInitializers),
273 CheckUseAfterReturn(CheckUseAfterReturn || ClUseAfterReturn),
274 CheckLifetime(CheckLifetime || ClCheckLifetime),
275 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
277 ZeroBaseShadow(ZeroBaseShadow) {}
278 virtual const char *getPassName() const {
279 return "AddressSanitizerFunctionPass";
281 void instrumentMop(Instruction *I);
282 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
283 Value *Addr, uint32_t TypeSize, bool IsWrite,
284 Value *SizeArgument);
285 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
286 Value *ShadowValue, uint32_t TypeSize);
287 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
288 bool IsWrite, size_t AccessSizeIndex,
289 Value *SizeArgument);
290 bool instrumentMemIntrinsic(MemIntrinsic *MI);
291 void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
293 Instruction *InsertBefore, bool IsWrite);
294 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
295 bool runOnFunction(Function &F);
296 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
297 void emitShadowMapping(Module &M, IRBuilder<> &IRB) const;
298 virtual bool doInitialization(Module &M);
299 static char ID; // Pass identification, replacement for typeid
302 void initializeCallbacks(Module &M);
304 bool ShouldInstrumentGlobal(GlobalVariable *G);
305 bool LooksLikeCodeInBug11395(Instruction *I);
306 void FindDynamicInitializers(Module &M);
309 bool CheckUseAfterReturn;
311 SmallString<64> BlacklistFile;
318 ShadowMapping Mapping;
319 Function *AsanCtorFunction;
320 Function *AsanInitFunction;
321 Function *AsanHandleNoReturnFunc;
322 OwningPtr<SpecialCaseList> BL;
323 // This array is indexed by AccessIsWrite and log2(AccessSize).
324 Function *AsanErrorCallback[2][kNumberOfAccessSizes];
325 // This array is indexed by AccessIsWrite.
326 Function *AsanErrorCallbackSized[2];
328 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
330 friend struct FunctionStackPoisoner;
333 class AddressSanitizerModule : public ModulePass {
335 AddressSanitizerModule(bool CheckInitOrder = true,
336 StringRef BlacklistFile = StringRef(),
337 bool ZeroBaseShadow = false)
339 CheckInitOrder(CheckInitOrder || ClInitializers),
340 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
342 ZeroBaseShadow(ZeroBaseShadow) {}
343 bool runOnModule(Module &M);
344 static char ID; // Pass identification, replacement for typeid
345 virtual const char *getPassName() const {
346 return "AddressSanitizerModule";
350 void initializeCallbacks(Module &M);
352 bool ShouldInstrumentGlobal(GlobalVariable *G);
353 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
354 size_t RedzoneSize() const {
355 return RedzoneSizeForScale(Mapping.Scale);
359 SmallString<64> BlacklistFile;
362 OwningPtr<SpecialCaseList> BL;
363 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
367 ShadowMapping Mapping;
368 Function *AsanPoisonGlobals;
369 Function *AsanUnpoisonGlobals;
370 Function *AsanRegisterGlobals;
371 Function *AsanUnregisterGlobals;
374 // Stack poisoning does not play well with exception handling.
375 // When an exception is thrown, we essentially bypass the code
376 // that unpoisones the stack. This is why the run-time library has
377 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
378 // stack in the interceptor. This however does not work inside the
379 // actual function which catches the exception. Most likely because the
380 // compiler hoists the load of the shadow value somewhere too high.
381 // This causes asan to report a non-existing bug on 453.povray.
382 // It sounds like an LLVM bug.
383 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
385 AddressSanitizer &ASan;
390 ShadowMapping Mapping;
392 SmallVector<AllocaInst*, 16> AllocaVec;
393 SmallVector<Instruction*, 8> RetVec;
394 uint64_t TotalStackSize;
395 unsigned StackAlignment;
397 Function *AsanStackMallocFunc, *AsanStackFreeFunc;
398 Function *AsanPoisonStackMemoryFunc, *AsanUnpoisonStackMemoryFunc;
400 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
401 struct AllocaPoisonCall {
402 IntrinsicInst *InsBefore;
406 SmallVector<AllocaPoisonCall, 8> AllocaPoisonCallVec;
408 // Maps Value to an AllocaInst from which the Value is originated.
409 typedef DenseMap<Value*, AllocaInst*> AllocaForValueMapTy;
410 AllocaForValueMapTy AllocaForValue;
412 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
413 : F(F), ASan(ASan), DIB(*F.getParent()), C(ASan.C),
414 IntptrTy(ASan.IntptrTy), IntptrPtrTy(PointerType::get(IntptrTy, 0)),
415 Mapping(ASan.Mapping),
416 TotalStackSize(0), StackAlignment(1 << Mapping.Scale) {}
418 bool runOnFunction() {
419 if (!ClStack) return false;
420 // Collect alloca, ret, lifetime instructions etc.
421 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
422 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
423 BasicBlock *BB = *DI;
426 if (AllocaVec.empty()) return false;
428 initializeCallbacks(*F.getParent());
438 // Finds all static Alloca instructions and puts
439 // poisoned red zones around all of them.
440 // Then unpoison everything back before the function returns.
443 // ----------------------- Visitors.
444 /// \brief Collect all Ret instructions.
445 void visitReturnInst(ReturnInst &RI) {
446 RetVec.push_back(&RI);
449 /// \brief Collect Alloca instructions we want (and can) handle.
450 void visitAllocaInst(AllocaInst &AI) {
451 if (!isInterestingAlloca(AI)) return;
453 StackAlignment = std::max(StackAlignment, AI.getAlignment());
454 AllocaVec.push_back(&AI);
455 uint64_t AlignedSize = getAlignedAllocaSize(&AI);
456 TotalStackSize += AlignedSize;
459 /// \brief Collect lifetime intrinsic calls to check for use-after-scope
461 void visitIntrinsicInst(IntrinsicInst &II) {
462 if (!ASan.CheckLifetime) return;
463 Intrinsic::ID ID = II.getIntrinsicID();
464 if (ID != Intrinsic::lifetime_start &&
465 ID != Intrinsic::lifetime_end)
467 // Found lifetime intrinsic, add ASan instrumentation if necessary.
468 ConstantInt *Size = dyn_cast<ConstantInt>(II.getArgOperand(0));
469 // If size argument is undefined, don't do anything.
470 if (Size->isMinusOne()) return;
471 // Check that size doesn't saturate uint64_t and can
472 // be stored in IntptrTy.
473 const uint64_t SizeValue = Size->getValue().getLimitedValue();
474 if (SizeValue == ~0ULL ||
475 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
477 // Find alloca instruction that corresponds to llvm.lifetime argument.
478 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1));
480 bool DoPoison = (ID == Intrinsic::lifetime_end);
481 AllocaPoisonCall APC = {&II, SizeValue, DoPoison};
482 AllocaPoisonCallVec.push_back(APC);
485 // ---------------------- Helpers.
486 void initializeCallbacks(Module &M);
488 // Check if we want (and can) handle this alloca.
489 bool isInterestingAlloca(AllocaInst &AI) {
490 return (!AI.isArrayAllocation() &&
491 AI.isStaticAlloca() &&
492 AI.getAlignment() <= RedzoneSize() &&
493 AI.getAllocatedType()->isSized());
496 size_t RedzoneSize() const {
497 return RedzoneSizeForScale(Mapping.Scale);
499 uint64_t getAllocaSizeInBytes(AllocaInst *AI) {
500 Type *Ty = AI->getAllocatedType();
501 uint64_t SizeInBytes = ASan.TD->getTypeAllocSize(Ty);
504 uint64_t getAlignedSize(uint64_t SizeInBytes) {
505 size_t RZ = RedzoneSize();
506 return ((SizeInBytes + RZ - 1) / RZ) * RZ;
508 uint64_t getAlignedAllocaSize(AllocaInst *AI) {
509 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
510 return getAlignedSize(SizeInBytes);
512 /// Finds alloca where the value comes from.
513 AllocaInst *findAllocaForValue(Value *V);
514 void poisonRedZones(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB,
515 Value *ShadowBase, bool DoPoison);
516 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> IRB, bool DoPoison);
521 char AddressSanitizer::ID = 0;
522 INITIALIZE_PASS(AddressSanitizer, "asan",
523 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
525 FunctionPass *llvm::createAddressSanitizerFunctionPass(
526 bool CheckInitOrder, bool CheckUseAfterReturn, bool CheckLifetime,
527 StringRef BlacklistFile, bool ZeroBaseShadow) {
528 return new AddressSanitizer(CheckInitOrder, CheckUseAfterReturn,
529 CheckLifetime, BlacklistFile, ZeroBaseShadow);
532 char AddressSanitizerModule::ID = 0;
533 INITIALIZE_PASS(AddressSanitizerModule, "asan-module",
534 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
535 "ModulePass", false, false)
536 ModulePass *llvm::createAddressSanitizerModulePass(
537 bool CheckInitOrder, StringRef BlacklistFile, bool ZeroBaseShadow) {
538 return new AddressSanitizerModule(CheckInitOrder, BlacklistFile,
542 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
543 size_t Res = countTrailingZeros(TypeSize / 8);
544 assert(Res < kNumberOfAccessSizes);
548 // Create a constant for Str so that we can pass it to the run-time lib.
549 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
550 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
551 GlobalVariable *GV = new GlobalVariable(M, StrConst->getType(), true,
552 GlobalValue::PrivateLinkage, StrConst,
554 GV->setUnnamedAddr(true); // Ok to merge these.
555 GV->setAlignment(1); // Strings may not be merged w/o setting align 1.
559 static bool GlobalWasGeneratedByAsan(GlobalVariable *G) {
560 return G->getName().find(kAsanGenPrefix) == 0;
563 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
565 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
566 if (Mapping.Offset == 0)
568 // (Shadow >> scale) | offset
569 if (Mapping.OrShadowOffset)
570 return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
572 return IRB.CreateAdd(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
575 void AddressSanitizer::instrumentMemIntrinsicParam(
576 Instruction *OrigIns,
577 Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
578 IRBuilder<> IRB(InsertBefore);
579 if (Size->getType() != IntptrTy)
580 Size = IRB.CreateIntCast(Size, IntptrTy, false);
581 // Check the first byte.
582 instrumentAddress(OrigIns, InsertBefore, Addr, 8, IsWrite, Size);
583 // Check the last byte.
584 IRB.SetInsertPoint(InsertBefore);
585 Value *SizeMinusOne = IRB.CreateSub(Size, ConstantInt::get(IntptrTy, 1));
586 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
587 Value *AddrLast = IRB.CreateAdd(AddrLong, SizeMinusOne);
588 instrumentAddress(OrigIns, InsertBefore, AddrLast, 8, IsWrite, Size);
591 // Instrument memset/memmove/memcpy
592 bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
593 Value *Dst = MI->getDest();
594 MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
595 Value *Src = MemTran ? MemTran->getSource() : 0;
596 Value *Length = MI->getLength();
598 Constant *ConstLength = dyn_cast<Constant>(Length);
599 Instruction *InsertBefore = MI;
601 if (ConstLength->isNullValue()) return false;
603 // The size is not a constant so it could be zero -- check at run-time.
604 IRBuilder<> IRB(InsertBefore);
606 Value *Cmp = IRB.CreateICmpNE(Length,
607 Constant::getNullValue(Length->getType()));
608 InsertBefore = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
611 instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
613 instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
617 // If I is an interesting memory access, return the PointerOperand
618 // and set IsWrite. Otherwise return NULL.
619 static Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite) {
620 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
621 if (!ClInstrumentReads) return NULL;
623 return LI->getPointerOperand();
625 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
626 if (!ClInstrumentWrites) return NULL;
628 return SI->getPointerOperand();
630 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
631 if (!ClInstrumentAtomics) return NULL;
633 return RMW->getPointerOperand();
635 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
636 if (!ClInstrumentAtomics) return NULL;
638 return XCHG->getPointerOperand();
643 void AddressSanitizer::instrumentMop(Instruction *I) {
644 bool IsWrite = false;
645 Value *Addr = isInterestingMemoryAccess(I, &IsWrite);
647 if (ClOpt && ClOptGlobals) {
648 if (GlobalVariable *G = dyn_cast<GlobalVariable>(Addr)) {
649 // If initialization order checking is disabled, a simple access to a
650 // dynamically initialized global is always valid.
653 // If a global variable does not have dynamic initialization we don't
654 // have to instrument it. However, if a global does not have initailizer
655 // at all, we assume it has dynamic initializer (in other TU).
656 if (G->hasInitializer() && !DynamicallyInitializedGlobals.Contains(G))
661 Type *OrigPtrTy = Addr->getType();
662 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
664 assert(OrigTy->isSized());
665 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
667 assert((TypeSize % 8) == 0);
669 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check.
670 if (TypeSize == 8 || TypeSize == 16 ||
671 TypeSize == 32 || TypeSize == 64 || TypeSize == 128)
672 return instrumentAddress(I, I, Addr, TypeSize, IsWrite, 0);
673 // Instrument unusual size (but still multiple of 8).
674 // We can not do it with a single check, so we do 1-byte check for the first
675 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
676 // to report the actual access size.
678 Value *LastByte = IRB.CreateIntToPtr(
679 IRB.CreateAdd(IRB.CreatePointerCast(Addr, IntptrTy),
680 ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
682 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
683 instrumentAddress(I, I, Addr, 8, IsWrite, Size);
684 instrumentAddress(I, I, LastByte, 8, IsWrite, Size);
687 // Validate the result of Module::getOrInsertFunction called for an interface
688 // function of AddressSanitizer. If the instrumented module defines a function
689 // with the same name, their prototypes must match, otherwise
690 // getOrInsertFunction returns a bitcast.
691 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
692 if (isa<Function>(FuncOrBitcast)) return cast<Function>(FuncOrBitcast);
693 FuncOrBitcast->dump();
694 report_fatal_error("trying to redefine an AddressSanitizer "
695 "interface function");
698 Instruction *AddressSanitizer::generateCrashCode(
699 Instruction *InsertBefore, Value *Addr,
700 bool IsWrite, size_t AccessSizeIndex, Value *SizeArgument) {
701 IRBuilder<> IRB(InsertBefore);
702 CallInst *Call = SizeArgument
703 ? IRB.CreateCall2(AsanErrorCallbackSized[IsWrite], Addr, SizeArgument)
704 : IRB.CreateCall(AsanErrorCallback[IsWrite][AccessSizeIndex], Addr);
706 // We don't do Call->setDoesNotReturn() because the BB already has
707 // UnreachableInst at the end.
708 // This EmptyAsm is required to avoid callback merge.
709 IRB.CreateCall(EmptyAsm);
713 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
716 size_t Granularity = 1 << Mapping.Scale;
717 // Addr & (Granularity - 1)
718 Value *LastAccessedByte = IRB.CreateAnd(
719 AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
720 // (Addr & (Granularity - 1)) + size - 1
721 if (TypeSize / 8 > 1)
722 LastAccessedByte = IRB.CreateAdd(
723 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
724 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
725 LastAccessedByte = IRB.CreateIntCast(
726 LastAccessedByte, ShadowValue->getType(), false);
727 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
728 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
731 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
732 Instruction *InsertBefore,
733 Value *Addr, uint32_t TypeSize,
734 bool IsWrite, Value *SizeArgument) {
735 IRBuilder<> IRB(InsertBefore);
736 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
738 Type *ShadowTy = IntegerType::get(
739 *C, std::max(8U, TypeSize >> Mapping.Scale));
740 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
741 Value *ShadowPtr = memToShadow(AddrLong, IRB);
742 Value *CmpVal = Constant::getNullValue(ShadowTy);
743 Value *ShadowValue = IRB.CreateLoad(
744 IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
746 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
747 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
748 size_t Granularity = 1 << Mapping.Scale;
749 TerminatorInst *CrashTerm = 0;
751 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
752 TerminatorInst *CheckTerm =
753 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
754 assert(dyn_cast<BranchInst>(CheckTerm)->isUnconditional());
755 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
756 IRB.SetInsertPoint(CheckTerm);
757 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
758 BasicBlock *CrashBlock =
759 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
760 CrashTerm = new UnreachableInst(*C, CrashBlock);
761 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
762 ReplaceInstWithInst(CheckTerm, NewTerm);
764 CrashTerm = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), true);
767 Instruction *Crash = generateCrashCode(
768 CrashTerm, AddrLong, IsWrite, AccessSizeIndex, SizeArgument);
769 Crash->setDebugLoc(OrigIns->getDebugLoc());
772 void AddressSanitizerModule::createInitializerPoisonCalls(
773 Module &M, GlobalValue *ModuleName) {
774 // We do all of our poisoning and unpoisoning within _GLOBAL__I_a.
775 Function *GlobalInit = M.getFunction("_GLOBAL__I_a");
776 // If that function is not present, this TU contains no globals, or they have
777 // all been optimized away
781 // Set up the arguments to our poison/unpoison functions.
782 IRBuilder<> IRB(GlobalInit->begin()->getFirstInsertionPt());
784 // Add a call to poison all external globals before the given function starts.
785 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
786 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
788 // Add calls to unpoison all globals before each return instruction.
789 for (Function::iterator I = GlobalInit->begin(), E = GlobalInit->end();
791 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator())) {
792 CallInst::Create(AsanUnpoisonGlobals, "", RI);
797 bool AddressSanitizerModule::ShouldInstrumentGlobal(GlobalVariable *G) {
798 Type *Ty = cast<PointerType>(G->getType())->getElementType();
799 DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
801 if (BL->isIn(*G)) return false;
802 if (!Ty->isSized()) return false;
803 if (!G->hasInitializer()) return false;
804 if (GlobalWasGeneratedByAsan(G)) return false; // Our own global.
805 // Touch only those globals that will not be defined in other modules.
806 // Don't handle ODR type linkages since other modules may be built w/o asan.
807 if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
808 G->getLinkage() != GlobalVariable::PrivateLinkage &&
809 G->getLinkage() != GlobalVariable::InternalLinkage)
811 // Two problems with thread-locals:
812 // - The address of the main thread's copy can't be computed at link-time.
813 // - Need to poison all copies, not just the main thread's one.
814 if (G->isThreadLocal())
816 // For now, just ignore this Alloca if the alignment is large.
817 if (G->getAlignment() > RedzoneSize()) return false;
819 // Ignore all the globals with the names starting with "\01L_OBJC_".
820 // Many of those are put into the .cstring section. The linker compresses
821 // that section by removing the spare \0s after the string terminator, so
822 // our redzones get broken.
823 if ((G->getName().find("\01L_OBJC_") == 0) ||
824 (G->getName().find("\01l_OBJC_") == 0)) {
825 DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
829 if (G->hasSection()) {
830 StringRef Section(G->getSection());
831 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
832 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
834 if ((Section.find("__OBJC,") == 0) ||
835 (Section.find("__DATA, __objc_") == 0)) {
836 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
839 // See http://code.google.com/p/address-sanitizer/issues/detail?id=32
840 // Constant CFString instances are compiled in the following way:
841 // -- the string buffer is emitted into
842 // __TEXT,__cstring,cstring_literals
843 // -- the constant NSConstantString structure referencing that buffer
844 // is placed into __DATA,__cfstring
845 // Therefore there's no point in placing redzones into __DATA,__cfstring.
846 // Moreover, it causes the linker to crash on OS X 10.7
847 if (Section.find("__DATA,__cfstring") == 0) {
848 DEBUG(dbgs() << "Ignoring CFString: " << *G);
856 void AddressSanitizerModule::initializeCallbacks(Module &M) {
858 // Declare our poisoning and unpoisoning functions.
859 AsanPoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
860 kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, NULL));
861 AsanPoisonGlobals->setLinkage(Function::ExternalLinkage);
862 AsanUnpoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
863 kAsanUnpoisonGlobalsName, IRB.getVoidTy(), NULL));
864 AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage);
865 // Declare functions that register/unregister globals.
866 AsanRegisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
867 kAsanRegisterGlobalsName, IRB.getVoidTy(),
868 IntptrTy, IntptrTy, NULL));
869 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
870 AsanUnregisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
871 kAsanUnregisterGlobalsName,
872 IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
873 AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage);
876 // This function replaces all global variables with new variables that have
877 // trailing redzones. It also creates a function that poisons
878 // redzones and inserts this function into llvm.global_ctors.
879 bool AddressSanitizerModule::runOnModule(Module &M) {
880 if (!ClGlobals) return false;
881 TD = getAnalysisIfAvailable<DataLayout>();
884 BL.reset(new SpecialCaseList(BlacklistFile));
885 if (BL->isIn(M)) return false;
886 C = &(M.getContext());
887 int LongSize = TD->getPointerSizeInBits();
888 IntptrTy = Type::getIntNTy(*C, LongSize);
889 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
890 initializeCallbacks(M);
891 DynamicallyInitializedGlobals.Init(M);
893 SmallVector<GlobalVariable *, 16> GlobalsToChange;
895 for (Module::GlobalListType::iterator G = M.global_begin(),
896 E = M.global_end(); G != E; ++G) {
897 if (ShouldInstrumentGlobal(G))
898 GlobalsToChange.push_back(G);
901 size_t n = GlobalsToChange.size();
902 if (n == 0) return false;
904 // A global is described by a structure
907 // size_t size_with_redzone;
909 // const char *module_name;
910 // size_t has_dynamic_init;
911 // We initialize an array of such structures and pass it to a run-time call.
912 StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
914 IntptrTy, IntptrTy, NULL);
915 SmallVector<Constant *, 16> Initializers(n), DynamicInit;
918 Function *CtorFunc = M.getFunction(kAsanModuleCtorName);
920 IRBuilder<> IRB(CtorFunc->getEntryBlock().getTerminator());
922 bool HasDynamicallyInitializedGlobals = false;
924 GlobalVariable *ModuleName = createPrivateGlobalForString(
925 M, M.getModuleIdentifier());
926 // We shouldn't merge same module names, as this string serves as unique
927 // module ID in runtime.
928 ModuleName->setUnnamedAddr(false);
930 for (size_t i = 0; i < n; i++) {
931 static const uint64_t kMaxGlobalRedzone = 1 << 18;
932 GlobalVariable *G = GlobalsToChange[i];
933 PointerType *PtrTy = cast<PointerType>(G->getType());
934 Type *Ty = PtrTy->getElementType();
935 uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
936 uint64_t MinRZ = RedzoneSize();
937 // MinRZ <= RZ <= kMaxGlobalRedzone
938 // and trying to make RZ to be ~ 1/4 of SizeInBytes.
939 uint64_t RZ = std::max(MinRZ,
940 std::min(kMaxGlobalRedzone,
941 (SizeInBytes / MinRZ / 4) * MinRZ));
942 uint64_t RightRedzoneSize = RZ;
944 if (SizeInBytes % MinRZ)
945 RightRedzoneSize += MinRZ - (SizeInBytes % MinRZ);
946 assert(((RightRedzoneSize + SizeInBytes) % MinRZ) == 0);
947 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
948 // Determine whether this global should be poisoned in initialization.
949 bool GlobalHasDynamicInitializer =
950 DynamicallyInitializedGlobals.Contains(G);
951 // Don't check initialization order if this global is blacklisted.
952 GlobalHasDynamicInitializer &= !BL->isIn(*G, "init");
954 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
955 Constant *NewInitializer = ConstantStruct::get(
956 NewTy, G->getInitializer(),
957 Constant::getNullValue(RightRedZoneTy), NULL);
959 GlobalVariable *Name = createPrivateGlobalForString(M, G->getName());
961 // Create a new global variable with enough space for a redzone.
962 GlobalVariable *NewGlobal = new GlobalVariable(
963 M, NewTy, G->isConstant(), G->getLinkage(),
964 NewInitializer, "", G, G->getThreadLocalMode());
965 NewGlobal->copyAttributesFrom(G);
966 NewGlobal->setAlignment(MinRZ);
969 Indices2[0] = IRB.getInt32(0);
970 Indices2[1] = IRB.getInt32(0);
972 G->replaceAllUsesWith(
973 ConstantExpr::getGetElementPtr(NewGlobal, Indices2, true));
974 NewGlobal->takeName(G);
975 G->eraseFromParent();
977 Initializers[i] = ConstantStruct::get(
979 ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
980 ConstantInt::get(IntptrTy, SizeInBytes),
981 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
982 ConstantExpr::getPointerCast(Name, IntptrTy),
983 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
984 ConstantInt::get(IntptrTy, GlobalHasDynamicInitializer),
987 // Populate the first and last globals declared in this TU.
988 if (CheckInitOrder && GlobalHasDynamicInitializer)
989 HasDynamicallyInitializedGlobals = true;
991 DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
994 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
995 GlobalVariable *AllGlobals = new GlobalVariable(
996 M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage,
997 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
999 // Create calls for poisoning before initializers run and unpoisoning after.
1000 if (CheckInitOrder && HasDynamicallyInitializedGlobals)
1001 createInitializerPoisonCalls(M, ModuleName);
1002 IRB.CreateCall2(AsanRegisterGlobals,
1003 IRB.CreatePointerCast(AllGlobals, IntptrTy),
1004 ConstantInt::get(IntptrTy, n));
1006 // We also need to unregister globals at the end, e.g. when a shared library
1008 Function *AsanDtorFunction = Function::Create(
1009 FunctionType::get(Type::getVoidTy(*C), false),
1010 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
1011 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
1012 IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB));
1013 IRB_Dtor.CreateCall2(AsanUnregisterGlobals,
1014 IRB.CreatePointerCast(AllGlobals, IntptrTy),
1015 ConstantInt::get(IntptrTy, n));
1016 appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndCtorPriority);
1022 void AddressSanitizer::initializeCallbacks(Module &M) {
1023 IRBuilder<> IRB(*C);
1024 // Create __asan_report* callbacks.
1025 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
1026 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
1027 AccessSizeIndex++) {
1028 // IsWrite and TypeSize are encoded in the function name.
1029 std::string FunctionName = std::string(kAsanReportErrorTemplate) +
1030 (AccessIsWrite ? "store" : "load") + itostr(1 << AccessSizeIndex);
1031 // If we are merging crash callbacks, they have two parameters.
1032 AsanErrorCallback[AccessIsWrite][AccessSizeIndex] =
1033 checkInterfaceFunction(M.getOrInsertFunction(
1034 FunctionName, IRB.getVoidTy(), IntptrTy, NULL));
1037 AsanErrorCallbackSized[0] = checkInterfaceFunction(M.getOrInsertFunction(
1038 kAsanReportLoadN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1039 AsanErrorCallbackSized[1] = checkInterfaceFunction(M.getOrInsertFunction(
1040 kAsanReportStoreN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1042 AsanHandleNoReturnFunc = checkInterfaceFunction(M.getOrInsertFunction(
1043 kAsanHandleNoReturnName, IRB.getVoidTy(), NULL));
1044 // We insert an empty inline asm after __asan_report* to avoid callback merge.
1045 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
1046 StringRef(""), StringRef(""),
1047 /*hasSideEffects=*/true);
1050 void AddressSanitizer::emitShadowMapping(Module &M, IRBuilder<> &IRB) const {
1051 // Tell the values of mapping offset and scale to the run-time.
1052 GlobalValue *asan_mapping_offset =
1053 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1054 ConstantInt::get(IntptrTy, Mapping.Offset),
1055 kAsanMappingOffsetName);
1056 // Read the global, otherwise it may be optimized away.
1057 IRB.CreateLoad(asan_mapping_offset, true);
1059 GlobalValue *asan_mapping_scale =
1060 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1061 ConstantInt::get(IntptrTy, Mapping.Scale),
1062 kAsanMappingScaleName);
1063 // Read the global, otherwise it may be optimized away.
1064 IRB.CreateLoad(asan_mapping_scale, true);
1068 bool AddressSanitizer::doInitialization(Module &M) {
1069 // Initialize the private fields. No one has accessed them before.
1070 TD = getAnalysisIfAvailable<DataLayout>();
1074 BL.reset(new SpecialCaseList(BlacklistFile));
1075 DynamicallyInitializedGlobals.Init(M);
1077 C = &(M.getContext());
1078 LongSize = TD->getPointerSizeInBits();
1079 IntptrTy = Type::getIntNTy(*C, LongSize);
1081 AsanCtorFunction = Function::Create(
1082 FunctionType::get(Type::getVoidTy(*C), false),
1083 GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
1084 BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
1085 // call __asan_init in the module ctor.
1086 IRBuilder<> IRB(ReturnInst::Create(*C, AsanCtorBB));
1087 AsanInitFunction = checkInterfaceFunction(
1088 M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
1089 AsanInitFunction->setLinkage(Function::ExternalLinkage);
1090 IRB.CreateCall(AsanInitFunction);
1092 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
1093 emitShadowMapping(M, IRB);
1095 appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority);
1099 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
1100 // For each NSObject descendant having a +load method, this method is invoked
1101 // by the ObjC runtime before any of the static constructors is called.
1102 // Therefore we need to instrument such methods with a call to __asan_init
1103 // at the beginning in order to initialize our runtime before any access to
1104 // the shadow memory.
1105 // We cannot just ignore these methods, because they may call other
1106 // instrumented functions.
1107 if (F.getName().find(" load]") != std::string::npos) {
1108 IRBuilder<> IRB(F.begin()->begin());
1109 IRB.CreateCall(AsanInitFunction);
1115 bool AddressSanitizer::runOnFunction(Function &F) {
1116 if (BL->isIn(F)) return false;
1117 if (&F == AsanCtorFunction) return false;
1118 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
1119 DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
1120 initializeCallbacks(*F.getParent());
1122 // If needed, insert __asan_init before checking for SanitizeAddress attr.
1123 maybeInsertAsanInitAtFunctionEntry(F);
1125 if (!F.hasFnAttribute(Attribute::SanitizeAddress))
1128 if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
1131 // We want to instrument every address only once per basic block (unless there
1132 // are calls between uses).
1133 SmallSet<Value*, 16> TempsToInstrument;
1134 SmallVector<Instruction*, 16> ToInstrument;
1135 SmallVector<Instruction*, 8> NoReturnCalls;
1139 // Fill the set of memory operations to instrument.
1140 for (Function::iterator FI = F.begin(), FE = F.end();
1142 TempsToInstrument.clear();
1143 int NumInsnsPerBB = 0;
1144 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1146 if (LooksLikeCodeInBug11395(BI)) return false;
1147 if (Value *Addr = isInterestingMemoryAccess(BI, &IsWrite)) {
1148 if (ClOpt && ClOptSameTemp) {
1149 if (!TempsToInstrument.insert(Addr))
1150 continue; // We've seen this temp in the current BB.
1152 } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
1155 if (isa<AllocaInst>(BI))
1159 // A call inside BB.
1160 TempsToInstrument.clear();
1161 if (CS.doesNotReturn())
1162 NoReturnCalls.push_back(CS.getInstruction());
1166 ToInstrument.push_back(BI);
1168 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB)
1173 Function *UninstrumentedDuplicate = 0;
1174 bool LikelyToInstrument =
1175 !NoReturnCalls.empty() || !ToInstrument.empty() || (NumAllocas > 0);
1176 if (ClKeepUninstrumented && LikelyToInstrument) {
1177 ValueToValueMapTy VMap;
1178 UninstrumentedDuplicate = CloneFunction(&F, VMap, false);
1179 UninstrumentedDuplicate->removeFnAttr(Attribute::SanitizeAddress);
1180 UninstrumentedDuplicate->setName("NOASAN_" + F.getName());
1181 F.getParent()->getFunctionList().push_back(UninstrumentedDuplicate);
1185 int NumInstrumented = 0;
1186 for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
1187 Instruction *Inst = ToInstrument[i];
1188 if (ClDebugMin < 0 || ClDebugMax < 0 ||
1189 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
1190 if (isInterestingMemoryAccess(Inst, &IsWrite))
1191 instrumentMop(Inst);
1193 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
1198 FunctionStackPoisoner FSP(F, *this);
1199 bool ChangedStack = FSP.runOnFunction();
1201 // We must unpoison the stack before every NoReturn call (throw, _exit, etc).
1202 // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37
1203 for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) {
1204 Instruction *CI = NoReturnCalls[i];
1205 IRBuilder<> IRB(CI);
1206 IRB.CreateCall(AsanHandleNoReturnFunc);
1209 bool res = NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty();
1210 DEBUG(dbgs() << "ASAN done instrumenting: " << res << " " << F << "\n");
1212 if (ClKeepUninstrumented) {
1214 // No instrumentation is done, no need for the duplicate.
1215 if (UninstrumentedDuplicate)
1216 UninstrumentedDuplicate->eraseFromParent();
1218 // The function was instrumented. We must have the duplicate.
1219 assert(UninstrumentedDuplicate);
1220 UninstrumentedDuplicate->setSection("NOASAN");
1221 assert(!F.hasSection());
1222 F.setSection("ASAN");
1229 static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
1230 if (ShadowRedzoneSize == 1) return PoisonByte;
1231 if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
1232 if (ShadowRedzoneSize == 4)
1233 return (PoisonByte << 24) + (PoisonByte << 16) +
1234 (PoisonByte << 8) + (PoisonByte);
1235 llvm_unreachable("ShadowRedzoneSize is either 1, 2 or 4");
1238 static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
1241 size_t ShadowGranularity,
1243 for (size_t i = 0; i < RZSize;
1244 i+= ShadowGranularity, Shadow++) {
1245 if (i + ShadowGranularity <= Size) {
1246 *Shadow = 0; // fully addressable
1247 } else if (i >= Size) {
1248 *Shadow = Magic; // unaddressable
1250 *Shadow = Size - i; // first Size-i bytes are addressable
1255 // Workaround for bug 11395: we don't want to instrument stack in functions
1256 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
1257 // FIXME: remove once the bug 11395 is fixed.
1258 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
1259 if (LongSize != 32) return false;
1260 CallInst *CI = dyn_cast<CallInst>(I);
1261 if (!CI || !CI->isInlineAsm()) return false;
1262 if (CI->getNumArgOperands() <= 5) return false;
1263 // We have inline assembly with quite a few arguments.
1267 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
1268 IRBuilder<> IRB(*C);
1269 AsanStackMallocFunc = checkInterfaceFunction(M.getOrInsertFunction(
1270 kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL));
1271 AsanStackFreeFunc = checkInterfaceFunction(M.getOrInsertFunction(
1272 kAsanStackFreeName, IRB.getVoidTy(),
1273 IntptrTy, IntptrTy, IntptrTy, NULL));
1274 AsanPoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1275 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1276 AsanUnpoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1277 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1280 void FunctionStackPoisoner::poisonRedZones(
1281 const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB, Value *ShadowBase,
1283 size_t ShadowRZSize = RedzoneSize() >> Mapping.Scale;
1284 assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
1285 Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
1286 Type *RZPtrTy = PointerType::get(RZTy, 0);
1288 Value *PoisonLeft = ConstantInt::get(RZTy,
1289 ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
1290 Value *PoisonMid = ConstantInt::get(RZTy,
1291 ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
1292 Value *PoisonRight = ConstantInt::get(RZTy,
1293 ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
1295 // poison the first red zone.
1296 IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
1298 // poison all other red zones.
1299 uint64_t Pos = RedzoneSize();
1300 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1301 AllocaInst *AI = AllocaVec[i];
1302 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1303 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1304 assert(AlignedSize - SizeInBytes < RedzoneSize());
1309 assert(ShadowBase->getType() == IntptrTy);
1310 if (SizeInBytes < AlignedSize) {
1311 // Poison the partial redzone at right
1312 Ptr = IRB.CreateAdd(
1313 ShadowBase, ConstantInt::get(IntptrTy,
1314 (Pos >> Mapping.Scale) - ShadowRZSize));
1315 size_t AddressableBytes = RedzoneSize() - (AlignedSize - SizeInBytes);
1316 uint32_t Poison = 0;
1318 PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
1320 1ULL << Mapping.Scale,
1321 kAsanStackPartialRedzoneMagic);
1323 ASan.TD->isLittleEndian()
1324 ? support::endian::byte_swap<uint32_t, support::little>(Poison)
1325 : support::endian::byte_swap<uint32_t, support::big>(Poison);
1327 Value *PartialPoison = ConstantInt::get(RZTy, Poison);
1328 IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1331 // Poison the full redzone at right.
1332 Ptr = IRB.CreateAdd(ShadowBase,
1333 ConstantInt::get(IntptrTy, Pos >> Mapping.Scale));
1334 bool LastAlloca = (i == AllocaVec.size() - 1);
1335 Value *Poison = LastAlloca ? PoisonRight : PoisonMid;
1336 IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1338 Pos += RedzoneSize();
1342 void FunctionStackPoisoner::poisonStack() {
1343 uint64_t LocalStackSize = TotalStackSize +
1344 (AllocaVec.size() + 1) * RedzoneSize();
1346 bool DoStackMalloc = ASan.CheckUseAfterReturn
1347 && LocalStackSize <= kMaxStackMallocSize;
1349 assert(AllocaVec.size() > 0);
1350 Instruction *InsBefore = AllocaVec[0];
1351 IRBuilder<> IRB(InsBefore);
1354 Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
1355 AllocaInst *MyAlloca =
1356 new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
1357 if (ClRealignStack && StackAlignment < RedzoneSize())
1358 StackAlignment = RedzoneSize();
1359 MyAlloca->setAlignment(StackAlignment);
1360 assert(MyAlloca->isStaticAlloca());
1361 Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
1362 Value *LocalStackBase = OrigStackBase;
1364 if (DoStackMalloc) {
1365 LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc,
1366 ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
1369 // This string will be parsed by the run-time (DescribeAddressIfStack).
1370 SmallString<2048> StackDescriptionStorage;
1371 raw_svector_ostream StackDescription(StackDescriptionStorage);
1372 StackDescription << AllocaVec.size() << " ";
1374 // Insert poison calls for lifetime intrinsics for alloca.
1375 bool HavePoisonedAllocas = false;
1376 for (size_t i = 0, n = AllocaPoisonCallVec.size(); i < n; i++) {
1377 const AllocaPoisonCall &APC = AllocaPoisonCallVec[i];
1378 IntrinsicInst *II = APC.InsBefore;
1379 AllocaInst *AI = findAllocaForValue(II->getArgOperand(1));
1381 IRBuilder<> IRB(II);
1382 poisonAlloca(AI, APC.Size, IRB, APC.DoPoison);
1383 HavePoisonedAllocas |= APC.DoPoison;
1386 uint64_t Pos = RedzoneSize();
1387 // Replace Alloca instructions with base+offset.
1388 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1389 AllocaInst *AI = AllocaVec[i];
1390 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1391 StringRef Name = AI->getName();
1392 StackDescription << Pos << " " << SizeInBytes << " "
1393 << Name.size() << " " << Name << " ";
1394 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1395 assert((AlignedSize % RedzoneSize()) == 0);
1396 Value *NewAllocaPtr = IRB.CreateIntToPtr(
1397 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
1399 replaceDbgDeclareForAlloca(AI, NewAllocaPtr, DIB);
1400 AI->replaceAllUsesWith(NewAllocaPtr);
1401 Pos += AlignedSize + RedzoneSize();
1403 assert(Pos == LocalStackSize);
1405 // The left-most redzone has enough space for at least 4 pointers.
1406 // Write the Magic value to redzone[0].
1407 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
1408 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
1410 // Write the frame description constant to redzone[1].
1411 Value *BasePlus1 = IRB.CreateIntToPtr(
1412 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, ASan.LongSize/8)),
1414 GlobalVariable *StackDescriptionGlobal =
1415 createPrivateGlobalForString(*F.getParent(), StackDescription.str());
1416 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal,
1418 IRB.CreateStore(Description, BasePlus1);
1419 // Write the PC to redzone[2].
1420 Value *BasePlus2 = IRB.CreateIntToPtr(
1421 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy,
1422 2 * ASan.LongSize/8)),
1424 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
1426 // Poison the stack redzones at the entry.
1427 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
1428 poisonRedZones(AllocaVec, IRB, ShadowBase, true);
1430 // Unpoison the stack before all ret instructions.
1431 for (size_t i = 0, n = RetVec.size(); i < n; i++) {
1432 Instruction *Ret = RetVec[i];
1433 IRBuilder<> IRBRet(Ret);
1434 // Mark the current frame as retired.
1435 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
1437 // Unpoison the stack.
1438 poisonRedZones(AllocaVec, IRBRet, ShadowBase, false);
1439 if (DoStackMalloc) {
1440 // In use-after-return mode, mark the whole stack frame unaddressable.
1441 IRBRet.CreateCall3(AsanStackFreeFunc, LocalStackBase,
1442 ConstantInt::get(IntptrTy, LocalStackSize),
1444 } else if (HavePoisonedAllocas) {
1445 // If we poisoned some allocas in llvm.lifetime analysis,
1446 // unpoison whole stack frame now.
1447 assert(LocalStackBase == OrigStackBase);
1448 poisonAlloca(LocalStackBase, LocalStackSize, IRBRet, false);
1452 // We are done. Remove the old unused alloca instructions.
1453 for (size_t i = 0, n = AllocaVec.size(); i < n; i++)
1454 AllocaVec[i]->eraseFromParent();
1457 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
1458 IRBuilder<> IRB, bool DoPoison) {
1459 // For now just insert the call to ASan runtime.
1460 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
1461 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
1462 IRB.CreateCall2(DoPoison ? AsanPoisonStackMemoryFunc
1463 : AsanUnpoisonStackMemoryFunc,
1467 // Handling llvm.lifetime intrinsics for a given %alloca:
1468 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
1469 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
1470 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
1471 // could be poisoned by previous llvm.lifetime.end instruction, as the
1472 // variable may go in and out of scope several times, e.g. in loops).
1473 // (3) if we poisoned at least one %alloca in a function,
1474 // unpoison the whole stack frame at function exit.
1476 AllocaInst *FunctionStackPoisoner::findAllocaForValue(Value *V) {
1477 if (AllocaInst *AI = dyn_cast<AllocaInst>(V))
1478 // We're intested only in allocas we can handle.
1479 return isInterestingAlloca(*AI) ? AI : 0;
1480 // See if we've already calculated (or started to calculate) alloca for a
1482 AllocaForValueMapTy::iterator I = AllocaForValue.find(V);
1483 if (I != AllocaForValue.end())
1485 // Store 0 while we're calculating alloca for value V to avoid
1486 // infinite recursion if the value references itself.
1487 AllocaForValue[V] = 0;
1488 AllocaInst *Res = 0;
1489 if (CastInst *CI = dyn_cast<CastInst>(V))
1490 Res = findAllocaForValue(CI->getOperand(0));
1491 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1492 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1493 Value *IncValue = PN->getIncomingValue(i);
1494 // Allow self-referencing phi-nodes.
1495 if (IncValue == PN) continue;
1496 AllocaInst *IncValueAI = findAllocaForValue(IncValue);
1497 // AI for incoming values should exist and should all be equal.
1498 if (IncValueAI == 0 || (Res != 0 && IncValueAI != Res))
1504 AllocaForValue[V] = Res;