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 kMinStackMallocSize = 1 << 6; // 64B
63 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
64 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
65 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
67 static const char *const kAsanModuleCtorName = "asan.module_ctor";
68 static const char *const kAsanModuleDtorName = "asan.module_dtor";
69 static const int kAsanCtorAndCtorPriority = 1;
70 static const char *const kAsanReportErrorTemplate = "__asan_report_";
71 static const char *const kAsanReportLoadN = "__asan_report_load_n";
72 static const char *const kAsanReportStoreN = "__asan_report_store_n";
73 static const char *const kAsanRegisterGlobalsName = "__asan_register_globals";
74 static const char *const kAsanUnregisterGlobalsName =
75 "__asan_unregister_globals";
76 static const char *const kAsanPoisonGlobalsName = "__asan_before_dynamic_init";
77 static const char *const kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init";
78 static const char *const kAsanInitName = "__asan_init_v3";
79 static const char *const kAsanHandleNoReturnName = "__asan_handle_no_return";
80 static const char *const kAsanMappingOffsetName = "__asan_mapping_offset";
81 static const char *const kAsanMappingScaleName = "__asan_mapping_scale";
82 static const int kMaxAsanStackMallocSizeClass = 10;
83 static const char *const kAsanStackMallocNameTemplate = "__asan_stack_malloc_";
84 static const char *const kAsanStackFreeNameTemplate = "__asan_stack_free_";
85 static const char *const kAsanGenPrefix = "__asan_gen_";
86 static const char *const kAsanPoisonStackMemoryName =
87 "__asan_poison_stack_memory";
88 static const char *const kAsanUnpoisonStackMemoryName =
89 "__asan_unpoison_stack_memory";
91 static const int kAsanStackLeftRedzoneMagic = 0xf1;
92 static const int kAsanStackMidRedzoneMagic = 0xf2;
93 static const int kAsanStackRightRedzoneMagic = 0xf3;
94 static const int kAsanStackPartialRedzoneMagic = 0xf4;
96 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
97 static const size_t kNumberOfAccessSizes = 5;
99 // Command-line flags.
101 // This flag may need to be replaced with -f[no-]asan-reads.
102 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
103 cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
104 static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
105 cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
106 static cl::opt<bool> ClInstrumentAtomics("asan-instrument-atomics",
107 cl::desc("instrument atomic instructions (rmw, cmpxchg)"),
108 cl::Hidden, cl::init(true));
109 static cl::opt<bool> ClAlwaysSlowPath("asan-always-slow-path",
110 cl::desc("use instrumentation with slow path for all accesses"),
111 cl::Hidden, cl::init(false));
112 // This flag limits the number of instructions to be instrumented
113 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
114 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
116 static cl::opt<int> ClMaxInsnsToInstrumentPerBB("asan-max-ins-per-bb",
118 cl::desc("maximal number of instructions to instrument in any given BB"),
120 // This flag may need to be replaced with -f[no]asan-stack.
121 static cl::opt<bool> ClStack("asan-stack",
122 cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
123 // This flag may need to be replaced with -f[no]asan-use-after-return.
124 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
125 cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
126 // This flag may need to be replaced with -f[no]asan-globals.
127 static cl::opt<bool> ClGlobals("asan-globals",
128 cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
129 static cl::opt<bool> ClInitializers("asan-initialization-order",
130 cl::desc("Handle C++ initializer order"), cl::Hidden, cl::init(false));
131 static cl::opt<bool> ClMemIntrin("asan-memintrin",
132 cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
133 static cl::opt<bool> ClRealignStack("asan-realign-stack",
134 cl::desc("Realign stack to 32"), cl::Hidden, cl::init(true));
135 static cl::opt<std::string> ClBlacklistFile("asan-blacklist",
136 cl::desc("File containing the list of objects to ignore "
137 "during instrumentation"), cl::Hidden);
139 // This is an experimental feature that will allow to choose between
140 // instrumented and non-instrumented code at link-time.
141 // If this option is on, just before instrumenting a function we create its
142 // clone; if the function is not changed by asan the clone is deleted.
143 // If we end up with a clone, we put the instrumented function into a section
144 // called "ASAN" and the uninstrumented function into a section called "NOASAN".
146 // This is still a prototype, we need to figure out a way to keep two copies of
147 // a function so that the linker can easily choose one of them.
148 static cl::opt<bool> ClKeepUninstrumented("asan-keep-uninstrumented-functions",
149 cl::desc("Keep uninstrumented copies of functions"),
150 cl::Hidden, cl::init(false));
152 // These flags allow to change the shadow mapping.
153 // The shadow mapping looks like
154 // Shadow = (Mem >> scale) + (1 << offset_log)
155 static cl::opt<int> ClMappingScale("asan-mapping-scale",
156 cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
157 static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
158 cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
159 static cl::opt<bool> ClShort64BitOffset("asan-short-64bit-mapping-offset",
160 cl::desc("Use short immediate constant as the mapping offset for 64bit"),
161 cl::Hidden, cl::init(true));
163 // Optimization flags. Not user visible, used mostly for testing
164 // and benchmarking the tool.
165 static cl::opt<bool> ClOpt("asan-opt",
166 cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
167 static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
168 cl::desc("Instrument the same temp just once"), cl::Hidden,
170 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
171 cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
173 static cl::opt<bool> ClCheckLifetime("asan-check-lifetime",
174 cl::desc("Use llvm.lifetime intrinsics to insert extra checks"),
175 cl::Hidden, cl::init(false));
178 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
180 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
181 cl::Hidden, cl::init(0));
182 static cl::opt<std::string> ClDebugFunc("asan-debug-func",
183 cl::Hidden, cl::desc("Debug func"));
184 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
185 cl::Hidden, cl::init(-1));
186 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
187 cl::Hidden, cl::init(-1));
190 /// A set of dynamically initialized globals extracted from metadata.
191 class SetOfDynamicallyInitializedGlobals {
193 void Init(Module& M) {
194 // Clang generates metadata identifying all dynamically initialized globals.
195 NamedMDNode *DynamicGlobals =
196 M.getNamedMetadata("llvm.asan.dynamically_initialized_globals");
199 for (int i = 0, n = DynamicGlobals->getNumOperands(); i < n; ++i) {
200 MDNode *MDN = DynamicGlobals->getOperand(i);
201 assert(MDN->getNumOperands() == 1);
202 Value *VG = MDN->getOperand(0);
203 // The optimizer may optimize away a global entirely, in which case we
204 // cannot instrument access to it.
207 DynInitGlobals.insert(cast<GlobalVariable>(VG));
210 bool Contains(GlobalVariable *G) { return DynInitGlobals.count(G) != 0; }
212 SmallSet<GlobalValue*, 32> DynInitGlobals;
215 /// This struct defines the shadow mapping using the rule:
216 /// shadow = (mem >> Scale) ADD-or-OR Offset.
217 struct ShadowMapping {
223 static ShadowMapping getShadowMapping(const Module &M, int LongSize,
224 bool ZeroBaseShadow) {
225 llvm::Triple TargetTriple(M.getTargetTriple());
226 bool IsAndroid = TargetTriple.getEnvironment() == llvm::Triple::Android;
227 bool IsMacOSX = TargetTriple.getOS() == llvm::Triple::MacOSX;
228 bool IsPPC64 = TargetTriple.getArch() == llvm::Triple::ppc64 ||
229 TargetTriple.getArch() == llvm::Triple::ppc64le;
230 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64;
231 bool IsMIPS32 = TargetTriple.getArch() == llvm::Triple::mips ||
232 TargetTriple.getArch() == llvm::Triple::mipsel;
234 ShadowMapping Mapping;
236 // OR-ing shadow offset if more efficient (at least on x86),
237 // but on ppc64 we have to use add since the shadow offset is not neccesary
238 // 1/8-th of the address space.
239 Mapping.OrShadowOffset = !IsPPC64 && !ClShort64BitOffset;
241 Mapping.Offset = (IsAndroid || ZeroBaseShadow) ? 0 :
243 (IsMIPS32 ? kMIPS32_ShadowOffset32 : kDefaultShadowOffset32) :
244 IsPPC64 ? kPPC64_ShadowOffset64 : kDefaultShadowOffset64);
245 if (!ZeroBaseShadow && ClShort64BitOffset && IsX86_64 && !IsMacOSX) {
246 assert(LongSize == 64);
247 Mapping.Offset = kDefaultShort64bitShadowOffset;
249 if (!ZeroBaseShadow && ClMappingOffsetLog >= 0) {
250 // Zero offset log is the special case.
251 Mapping.Offset = (ClMappingOffsetLog == 0) ? 0 : 1ULL << ClMappingOffsetLog;
254 Mapping.Scale = kDefaultShadowScale;
255 if (ClMappingScale) {
256 Mapping.Scale = ClMappingScale;
262 static size_t RedzoneSizeForScale(int MappingScale) {
263 // Redzone used for stack and globals is at least 32 bytes.
264 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
265 return std::max(32U, 1U << MappingScale);
268 /// AddressSanitizer: instrument the code in module to find memory bugs.
269 struct AddressSanitizer : public FunctionPass {
270 AddressSanitizer(bool CheckInitOrder = true,
271 bool CheckUseAfterReturn = false,
272 bool CheckLifetime = false,
273 StringRef BlacklistFile = StringRef(),
274 bool ZeroBaseShadow = false)
276 CheckInitOrder(CheckInitOrder || ClInitializers),
277 CheckUseAfterReturn(CheckUseAfterReturn || ClUseAfterReturn),
278 CheckLifetime(CheckLifetime || ClCheckLifetime),
279 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
281 ZeroBaseShadow(ZeroBaseShadow) {}
282 virtual const char *getPassName() const {
283 return "AddressSanitizerFunctionPass";
285 void instrumentMop(Instruction *I);
286 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
287 Value *Addr, uint32_t TypeSize, bool IsWrite,
288 Value *SizeArgument);
289 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
290 Value *ShadowValue, uint32_t TypeSize);
291 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
292 bool IsWrite, size_t AccessSizeIndex,
293 Value *SizeArgument);
294 bool instrumentMemIntrinsic(MemIntrinsic *MI);
295 void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
297 Instruction *InsertBefore, bool IsWrite);
298 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
299 bool runOnFunction(Function &F);
300 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
301 void emitShadowMapping(Module &M, IRBuilder<> &IRB) const;
302 virtual bool doInitialization(Module &M);
303 static char ID; // Pass identification, replacement for typeid
306 void initializeCallbacks(Module &M);
308 bool ShouldInstrumentGlobal(GlobalVariable *G);
309 bool LooksLikeCodeInBug11395(Instruction *I);
310 void FindDynamicInitializers(Module &M);
313 bool CheckUseAfterReturn;
315 SmallString<64> BlacklistFile;
322 ShadowMapping Mapping;
323 Function *AsanCtorFunction;
324 Function *AsanInitFunction;
325 Function *AsanHandleNoReturnFunc;
326 OwningPtr<SpecialCaseList> BL;
327 // This array is indexed by AccessIsWrite and log2(AccessSize).
328 Function *AsanErrorCallback[2][kNumberOfAccessSizes];
329 // This array is indexed by AccessIsWrite.
330 Function *AsanErrorCallbackSized[2];
332 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
334 friend struct FunctionStackPoisoner;
337 class AddressSanitizerModule : public ModulePass {
339 AddressSanitizerModule(bool CheckInitOrder = true,
340 StringRef BlacklistFile = StringRef(),
341 bool ZeroBaseShadow = false)
343 CheckInitOrder(CheckInitOrder || ClInitializers),
344 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
346 ZeroBaseShadow(ZeroBaseShadow) {}
347 bool runOnModule(Module &M);
348 static char ID; // Pass identification, replacement for typeid
349 virtual const char *getPassName() const {
350 return "AddressSanitizerModule";
354 void initializeCallbacks(Module &M);
356 bool ShouldInstrumentGlobal(GlobalVariable *G);
357 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
358 size_t RedzoneSize() const {
359 return RedzoneSizeForScale(Mapping.Scale);
363 SmallString<64> BlacklistFile;
366 OwningPtr<SpecialCaseList> BL;
367 SetOfDynamicallyInitializedGlobals DynamicallyInitializedGlobals;
371 ShadowMapping Mapping;
372 Function *AsanPoisonGlobals;
373 Function *AsanUnpoisonGlobals;
374 Function *AsanRegisterGlobals;
375 Function *AsanUnregisterGlobals;
378 // Stack poisoning does not play well with exception handling.
379 // When an exception is thrown, we essentially bypass the code
380 // that unpoisones the stack. This is why the run-time library has
381 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
382 // stack in the interceptor. This however does not work inside the
383 // actual function which catches the exception. Most likely because the
384 // compiler hoists the load of the shadow value somewhere too high.
385 // This causes asan to report a non-existing bug on 453.povray.
386 // It sounds like an LLVM bug.
387 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
389 AddressSanitizer &ASan;
394 ShadowMapping Mapping;
396 SmallVector<AllocaInst*, 16> AllocaVec;
397 SmallVector<Instruction*, 8> RetVec;
398 uint64_t TotalStackSize;
399 unsigned StackAlignment;
401 Function *AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1],
402 *AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1];
403 Function *AsanPoisonStackMemoryFunc, *AsanUnpoisonStackMemoryFunc;
405 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
406 struct AllocaPoisonCall {
407 IntrinsicInst *InsBefore;
411 SmallVector<AllocaPoisonCall, 8> AllocaPoisonCallVec;
413 // Maps Value to an AllocaInst from which the Value is originated.
414 typedef DenseMap<Value*, AllocaInst*> AllocaForValueMapTy;
415 AllocaForValueMapTy AllocaForValue;
417 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
418 : F(F), ASan(ASan), DIB(*F.getParent()), C(ASan.C),
419 IntptrTy(ASan.IntptrTy), IntptrPtrTy(PointerType::get(IntptrTy, 0)),
420 Mapping(ASan.Mapping),
421 TotalStackSize(0), StackAlignment(1 << Mapping.Scale) {}
423 bool runOnFunction() {
424 if (!ClStack) return false;
425 // Collect alloca, ret, lifetime instructions etc.
426 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
427 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
428 BasicBlock *BB = *DI;
431 if (AllocaVec.empty()) return false;
433 initializeCallbacks(*F.getParent());
443 // Finds all static Alloca instructions and puts
444 // poisoned red zones around all of them.
445 // Then unpoison everything back before the function returns.
448 // ----------------------- Visitors.
449 /// \brief Collect all Ret instructions.
450 void visitReturnInst(ReturnInst &RI) {
451 RetVec.push_back(&RI);
454 /// \brief Collect Alloca instructions we want (and can) handle.
455 void visitAllocaInst(AllocaInst &AI) {
456 if (!isInterestingAlloca(AI)) return;
458 StackAlignment = std::max(StackAlignment, AI.getAlignment());
459 AllocaVec.push_back(&AI);
460 uint64_t AlignedSize = getAlignedAllocaSize(&AI);
461 TotalStackSize += AlignedSize;
464 /// \brief Collect lifetime intrinsic calls to check for use-after-scope
466 void visitIntrinsicInst(IntrinsicInst &II) {
467 if (!ASan.CheckLifetime) return;
468 Intrinsic::ID ID = II.getIntrinsicID();
469 if (ID != Intrinsic::lifetime_start &&
470 ID != Intrinsic::lifetime_end)
472 // Found lifetime intrinsic, add ASan instrumentation if necessary.
473 ConstantInt *Size = dyn_cast<ConstantInt>(II.getArgOperand(0));
474 // If size argument is undefined, don't do anything.
475 if (Size->isMinusOne()) return;
476 // Check that size doesn't saturate uint64_t and can
477 // be stored in IntptrTy.
478 const uint64_t SizeValue = Size->getValue().getLimitedValue();
479 if (SizeValue == ~0ULL ||
480 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
482 // Find alloca instruction that corresponds to llvm.lifetime argument.
483 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1));
485 bool DoPoison = (ID == Intrinsic::lifetime_end);
486 AllocaPoisonCall APC = {&II, SizeValue, DoPoison};
487 AllocaPoisonCallVec.push_back(APC);
490 // ---------------------- Helpers.
491 void initializeCallbacks(Module &M);
493 // Check if we want (and can) handle this alloca.
494 bool isInterestingAlloca(AllocaInst &AI) const {
495 return (!AI.isArrayAllocation() &&
496 AI.isStaticAlloca() &&
497 AI.getAlignment() <= RedzoneSize() &&
498 AI.getAllocatedType()->isSized());
501 size_t RedzoneSize() const {
502 return RedzoneSizeForScale(Mapping.Scale);
504 uint64_t getAllocaSizeInBytes(AllocaInst *AI) const {
505 Type *Ty = AI->getAllocatedType();
506 uint64_t SizeInBytes = ASan.TD->getTypeAllocSize(Ty);
509 uint64_t getAlignedSize(uint64_t SizeInBytes) const {
510 size_t RZ = RedzoneSize();
511 return ((SizeInBytes + RZ - 1) / RZ) * RZ;
513 uint64_t getAlignedAllocaSize(AllocaInst *AI) const {
514 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
515 return getAlignedSize(SizeInBytes);
517 /// Finds alloca where the value comes from.
518 AllocaInst *findAllocaForValue(Value *V);
519 void poisonRedZones(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> &IRB,
520 Value *ShadowBase, bool DoPoison);
521 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
526 char AddressSanitizer::ID = 0;
527 INITIALIZE_PASS(AddressSanitizer, "asan",
528 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
530 FunctionPass *llvm::createAddressSanitizerFunctionPass(
531 bool CheckInitOrder, bool CheckUseAfterReturn, bool CheckLifetime,
532 StringRef BlacklistFile, bool ZeroBaseShadow) {
533 return new AddressSanitizer(CheckInitOrder, CheckUseAfterReturn,
534 CheckLifetime, BlacklistFile, ZeroBaseShadow);
537 char AddressSanitizerModule::ID = 0;
538 INITIALIZE_PASS(AddressSanitizerModule, "asan-module",
539 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
540 "ModulePass", false, false)
541 ModulePass *llvm::createAddressSanitizerModulePass(
542 bool CheckInitOrder, StringRef BlacklistFile, bool ZeroBaseShadow) {
543 return new AddressSanitizerModule(CheckInitOrder, BlacklistFile,
547 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
548 size_t Res = countTrailingZeros(TypeSize / 8);
549 assert(Res < kNumberOfAccessSizes);
553 // \brief Create a constant for Str so that we can pass it to the run-time lib.
554 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
555 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
556 GlobalVariable *GV = new GlobalVariable(M, StrConst->getType(), true,
557 GlobalValue::InternalLinkage, StrConst,
559 GV->setUnnamedAddr(true); // Ok to merge these.
560 GV->setAlignment(1); // Strings may not be merged w/o setting align 1.
564 static bool GlobalWasGeneratedByAsan(GlobalVariable *G) {
565 return G->getName().find(kAsanGenPrefix) == 0;
568 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
570 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
571 if (Mapping.Offset == 0)
573 // (Shadow >> scale) | offset
574 if (Mapping.OrShadowOffset)
575 return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
577 return IRB.CreateAdd(Shadow, ConstantInt::get(IntptrTy, Mapping.Offset));
580 void AddressSanitizer::instrumentMemIntrinsicParam(
581 Instruction *OrigIns,
582 Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
583 IRBuilder<> IRB(InsertBefore);
584 if (Size->getType() != IntptrTy)
585 Size = IRB.CreateIntCast(Size, IntptrTy, false);
586 // Check the first byte.
587 instrumentAddress(OrigIns, InsertBefore, Addr, 8, IsWrite, Size);
588 // Check the last byte.
589 IRB.SetInsertPoint(InsertBefore);
590 Value *SizeMinusOne = IRB.CreateSub(Size, ConstantInt::get(IntptrTy, 1));
591 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
592 Value *AddrLast = IRB.CreateAdd(AddrLong, SizeMinusOne);
593 instrumentAddress(OrigIns, InsertBefore, AddrLast, 8, IsWrite, Size);
596 // Instrument memset/memmove/memcpy
597 bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
598 Value *Dst = MI->getDest();
599 MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
600 Value *Src = MemTran ? MemTran->getSource() : 0;
601 Value *Length = MI->getLength();
603 Constant *ConstLength = dyn_cast<Constant>(Length);
604 Instruction *InsertBefore = MI;
606 if (ConstLength->isNullValue()) return false;
608 // The size is not a constant so it could be zero -- check at run-time.
609 IRBuilder<> IRB(InsertBefore);
611 Value *Cmp = IRB.CreateICmpNE(Length,
612 Constant::getNullValue(Length->getType()));
613 InsertBefore = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
616 instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
618 instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
622 // If I is an interesting memory access, return the PointerOperand
623 // and set IsWrite. Otherwise return NULL.
624 static Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite) {
625 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
626 if (!ClInstrumentReads) return NULL;
628 return LI->getPointerOperand();
630 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
631 if (!ClInstrumentWrites) return NULL;
633 return SI->getPointerOperand();
635 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
636 if (!ClInstrumentAtomics) return NULL;
638 return RMW->getPointerOperand();
640 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
641 if (!ClInstrumentAtomics) return NULL;
643 return XCHG->getPointerOperand();
648 void AddressSanitizer::instrumentMop(Instruction *I) {
649 bool IsWrite = false;
650 Value *Addr = isInterestingMemoryAccess(I, &IsWrite);
652 if (ClOpt && ClOptGlobals) {
653 if (GlobalVariable *G = dyn_cast<GlobalVariable>(Addr)) {
654 // If initialization order checking is disabled, a simple access to a
655 // dynamically initialized global is always valid.
658 // If a global variable does not have dynamic initialization we don't
659 // have to instrument it. However, if a global does not have initailizer
660 // at all, we assume it has dynamic initializer (in other TU).
661 if (G->hasInitializer() && !DynamicallyInitializedGlobals.Contains(G))
666 Type *OrigPtrTy = Addr->getType();
667 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
669 assert(OrigTy->isSized());
670 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
672 assert((TypeSize % 8) == 0);
674 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check.
675 if (TypeSize == 8 || TypeSize == 16 ||
676 TypeSize == 32 || TypeSize == 64 || TypeSize == 128)
677 return instrumentAddress(I, I, Addr, TypeSize, IsWrite, 0);
678 // Instrument unusual size (but still multiple of 8).
679 // We can not do it with a single check, so we do 1-byte check for the first
680 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
681 // to report the actual access size.
683 Value *LastByte = IRB.CreateIntToPtr(
684 IRB.CreateAdd(IRB.CreatePointerCast(Addr, IntptrTy),
685 ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
687 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
688 instrumentAddress(I, I, Addr, 8, IsWrite, Size);
689 instrumentAddress(I, I, LastByte, 8, IsWrite, Size);
692 // Validate the result of Module::getOrInsertFunction called for an interface
693 // function of AddressSanitizer. If the instrumented module defines a function
694 // with the same name, their prototypes must match, otherwise
695 // getOrInsertFunction returns a bitcast.
696 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
697 if (isa<Function>(FuncOrBitcast)) return cast<Function>(FuncOrBitcast);
698 FuncOrBitcast->dump();
699 report_fatal_error("trying to redefine an AddressSanitizer "
700 "interface function");
703 Instruction *AddressSanitizer::generateCrashCode(
704 Instruction *InsertBefore, Value *Addr,
705 bool IsWrite, size_t AccessSizeIndex, Value *SizeArgument) {
706 IRBuilder<> IRB(InsertBefore);
707 CallInst *Call = SizeArgument
708 ? IRB.CreateCall2(AsanErrorCallbackSized[IsWrite], Addr, SizeArgument)
709 : IRB.CreateCall(AsanErrorCallback[IsWrite][AccessSizeIndex], Addr);
711 // We don't do Call->setDoesNotReturn() because the BB already has
712 // UnreachableInst at the end.
713 // This EmptyAsm is required to avoid callback merge.
714 IRB.CreateCall(EmptyAsm);
718 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
721 size_t Granularity = 1 << Mapping.Scale;
722 // Addr & (Granularity - 1)
723 Value *LastAccessedByte = IRB.CreateAnd(
724 AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
725 // (Addr & (Granularity - 1)) + size - 1
726 if (TypeSize / 8 > 1)
727 LastAccessedByte = IRB.CreateAdd(
728 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
729 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
730 LastAccessedByte = IRB.CreateIntCast(
731 LastAccessedByte, ShadowValue->getType(), false);
732 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
733 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
736 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
737 Instruction *InsertBefore,
738 Value *Addr, uint32_t TypeSize,
739 bool IsWrite, Value *SizeArgument) {
740 IRBuilder<> IRB(InsertBefore);
741 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
743 Type *ShadowTy = IntegerType::get(
744 *C, std::max(8U, TypeSize >> Mapping.Scale));
745 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
746 Value *ShadowPtr = memToShadow(AddrLong, IRB);
747 Value *CmpVal = Constant::getNullValue(ShadowTy);
748 Value *ShadowValue = IRB.CreateLoad(
749 IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
751 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
752 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
753 size_t Granularity = 1 << Mapping.Scale;
754 TerminatorInst *CrashTerm = 0;
756 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
757 TerminatorInst *CheckTerm =
758 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false);
759 assert(dyn_cast<BranchInst>(CheckTerm)->isUnconditional());
760 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
761 IRB.SetInsertPoint(CheckTerm);
762 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
763 BasicBlock *CrashBlock =
764 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
765 CrashTerm = new UnreachableInst(*C, CrashBlock);
766 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
767 ReplaceInstWithInst(CheckTerm, NewTerm);
769 CrashTerm = SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), true);
772 Instruction *Crash = generateCrashCode(
773 CrashTerm, AddrLong, IsWrite, AccessSizeIndex, SizeArgument);
774 Crash->setDebugLoc(OrigIns->getDebugLoc());
777 void AddressSanitizerModule::createInitializerPoisonCalls(
778 Module &M, GlobalValue *ModuleName) {
779 // We do all of our poisoning and unpoisoning within _GLOBAL__I_a.
780 Function *GlobalInit = M.getFunction("_GLOBAL__I_a");
781 // If that function is not present, this TU contains no globals, or they have
782 // all been optimized away
786 // Set up the arguments to our poison/unpoison functions.
787 IRBuilder<> IRB(GlobalInit->begin()->getFirstInsertionPt());
789 // Add a call to poison all external globals before the given function starts.
790 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
791 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
793 // Add calls to unpoison all globals before each return instruction.
794 for (Function::iterator I = GlobalInit->begin(), E = GlobalInit->end();
796 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator())) {
797 CallInst::Create(AsanUnpoisonGlobals, "", RI);
802 bool AddressSanitizerModule::ShouldInstrumentGlobal(GlobalVariable *G) {
803 Type *Ty = cast<PointerType>(G->getType())->getElementType();
804 DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
806 if (BL->isIn(*G)) return false;
807 if (!Ty->isSized()) return false;
808 if (!G->hasInitializer()) return false;
809 if (GlobalWasGeneratedByAsan(G)) return false; // Our own global.
810 // Touch only those globals that will not be defined in other modules.
811 // Don't handle ODR type linkages since other modules may be built w/o asan.
812 if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
813 G->getLinkage() != GlobalVariable::PrivateLinkage &&
814 G->getLinkage() != GlobalVariable::InternalLinkage)
816 // Two problems with thread-locals:
817 // - The address of the main thread's copy can't be computed at link-time.
818 // - Need to poison all copies, not just the main thread's one.
819 if (G->isThreadLocal())
821 // For now, just ignore this Alloca if the alignment is large.
822 if (G->getAlignment() > RedzoneSize()) return false;
824 // Ignore all the globals with the names starting with "\01L_OBJC_".
825 // Many of those are put into the .cstring section. The linker compresses
826 // that section by removing the spare \0s after the string terminator, so
827 // our redzones get broken.
828 if ((G->getName().find("\01L_OBJC_") == 0) ||
829 (G->getName().find("\01l_OBJC_") == 0)) {
830 DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
834 if (G->hasSection()) {
835 StringRef Section(G->getSection());
836 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
837 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
839 if ((Section.find("__OBJC,") == 0) ||
840 (Section.find("__DATA, __objc_") == 0)) {
841 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
844 // See http://code.google.com/p/address-sanitizer/issues/detail?id=32
845 // Constant CFString instances are compiled in the following way:
846 // -- the string buffer is emitted into
847 // __TEXT,__cstring,cstring_literals
848 // -- the constant NSConstantString structure referencing that buffer
849 // is placed into __DATA,__cfstring
850 // Therefore there's no point in placing redzones into __DATA,__cfstring.
851 // Moreover, it causes the linker to crash on OS X 10.7
852 if (Section.find("__DATA,__cfstring") == 0) {
853 DEBUG(dbgs() << "Ignoring CFString: " << *G);
861 void AddressSanitizerModule::initializeCallbacks(Module &M) {
863 // Declare our poisoning and unpoisoning functions.
864 AsanPoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
865 kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, NULL));
866 AsanPoisonGlobals->setLinkage(Function::ExternalLinkage);
867 AsanUnpoisonGlobals = checkInterfaceFunction(M.getOrInsertFunction(
868 kAsanUnpoisonGlobalsName, IRB.getVoidTy(), NULL));
869 AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage);
870 // Declare functions that register/unregister globals.
871 AsanRegisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
872 kAsanRegisterGlobalsName, IRB.getVoidTy(),
873 IntptrTy, IntptrTy, NULL));
874 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
875 AsanUnregisterGlobals = checkInterfaceFunction(M.getOrInsertFunction(
876 kAsanUnregisterGlobalsName,
877 IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
878 AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage);
881 // This function replaces all global variables with new variables that have
882 // trailing redzones. It also creates a function that poisons
883 // redzones and inserts this function into llvm.global_ctors.
884 bool AddressSanitizerModule::runOnModule(Module &M) {
885 if (!ClGlobals) return false;
886 TD = getAnalysisIfAvailable<DataLayout>();
889 BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
890 if (BL->isIn(M)) return false;
891 C = &(M.getContext());
892 int LongSize = TD->getPointerSizeInBits();
893 IntptrTy = Type::getIntNTy(*C, LongSize);
894 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
895 initializeCallbacks(M);
896 DynamicallyInitializedGlobals.Init(M);
898 SmallVector<GlobalVariable *, 16> GlobalsToChange;
900 for (Module::GlobalListType::iterator G = M.global_begin(),
901 E = M.global_end(); G != E; ++G) {
902 if (ShouldInstrumentGlobal(G))
903 GlobalsToChange.push_back(G);
906 size_t n = GlobalsToChange.size();
907 if (n == 0) return false;
909 // A global is described by a structure
912 // size_t size_with_redzone;
914 // const char *module_name;
915 // size_t has_dynamic_init;
916 // We initialize an array of such structures and pass it to a run-time call.
917 StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
919 IntptrTy, IntptrTy, NULL);
920 SmallVector<Constant *, 16> Initializers(n), DynamicInit;
923 Function *CtorFunc = M.getFunction(kAsanModuleCtorName);
925 IRBuilder<> IRB(CtorFunc->getEntryBlock().getTerminator());
927 bool HasDynamicallyInitializedGlobals = false;
929 GlobalVariable *ModuleName = createPrivateGlobalForString(
930 M, M.getModuleIdentifier());
931 // We shouldn't merge same module names, as this string serves as unique
932 // module ID in runtime.
933 ModuleName->setUnnamedAddr(false);
935 for (size_t i = 0; i < n; i++) {
936 static const uint64_t kMaxGlobalRedzone = 1 << 18;
937 GlobalVariable *G = GlobalsToChange[i];
938 PointerType *PtrTy = cast<PointerType>(G->getType());
939 Type *Ty = PtrTy->getElementType();
940 uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
941 uint64_t MinRZ = RedzoneSize();
942 // MinRZ <= RZ <= kMaxGlobalRedzone
943 // and trying to make RZ to be ~ 1/4 of SizeInBytes.
944 uint64_t RZ = std::max(MinRZ,
945 std::min(kMaxGlobalRedzone,
946 (SizeInBytes / MinRZ / 4) * MinRZ));
947 uint64_t RightRedzoneSize = RZ;
949 if (SizeInBytes % MinRZ)
950 RightRedzoneSize += MinRZ - (SizeInBytes % MinRZ);
951 assert(((RightRedzoneSize + SizeInBytes) % MinRZ) == 0);
952 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
953 // Determine whether this global should be poisoned in initialization.
954 bool GlobalHasDynamicInitializer =
955 DynamicallyInitializedGlobals.Contains(G);
956 // Don't check initialization order if this global is blacklisted.
957 GlobalHasDynamicInitializer &= !BL->isIn(*G, "init");
959 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
960 Constant *NewInitializer = ConstantStruct::get(
961 NewTy, G->getInitializer(),
962 Constant::getNullValue(RightRedZoneTy), NULL);
964 GlobalVariable *Name = createPrivateGlobalForString(M, G->getName());
966 // Create a new global variable with enough space for a redzone.
967 GlobalValue::LinkageTypes Linkage = G->getLinkage();
968 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
969 Linkage = GlobalValue::InternalLinkage;
970 GlobalVariable *NewGlobal = new GlobalVariable(
971 M, NewTy, G->isConstant(), Linkage,
972 NewInitializer, "", G, G->getThreadLocalMode());
973 NewGlobal->copyAttributesFrom(G);
974 NewGlobal->setAlignment(MinRZ);
977 Indices2[0] = IRB.getInt32(0);
978 Indices2[1] = IRB.getInt32(0);
980 G->replaceAllUsesWith(
981 ConstantExpr::getGetElementPtr(NewGlobal, Indices2, true));
982 NewGlobal->takeName(G);
983 G->eraseFromParent();
985 Initializers[i] = ConstantStruct::get(
987 ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
988 ConstantInt::get(IntptrTy, SizeInBytes),
989 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
990 ConstantExpr::getPointerCast(Name, IntptrTy),
991 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
992 ConstantInt::get(IntptrTy, GlobalHasDynamicInitializer),
995 // Populate the first and last globals declared in this TU.
996 if (CheckInitOrder && GlobalHasDynamicInitializer)
997 HasDynamicallyInitializedGlobals = true;
999 DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
1002 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
1003 GlobalVariable *AllGlobals = new GlobalVariable(
1004 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
1005 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
1007 // Create calls for poisoning before initializers run and unpoisoning after.
1008 if (CheckInitOrder && HasDynamicallyInitializedGlobals)
1009 createInitializerPoisonCalls(M, ModuleName);
1010 IRB.CreateCall2(AsanRegisterGlobals,
1011 IRB.CreatePointerCast(AllGlobals, IntptrTy),
1012 ConstantInt::get(IntptrTy, n));
1014 // We also need to unregister globals at the end, e.g. when a shared library
1016 Function *AsanDtorFunction = Function::Create(
1017 FunctionType::get(Type::getVoidTy(*C), false),
1018 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
1019 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
1020 IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB));
1021 IRB_Dtor.CreateCall2(AsanUnregisterGlobals,
1022 IRB.CreatePointerCast(AllGlobals, IntptrTy),
1023 ConstantInt::get(IntptrTy, n));
1024 appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndCtorPriority);
1030 void AddressSanitizer::initializeCallbacks(Module &M) {
1031 IRBuilder<> IRB(*C);
1032 // Create __asan_report* callbacks.
1033 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
1034 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
1035 AccessSizeIndex++) {
1036 // IsWrite and TypeSize are encoded in the function name.
1037 std::string FunctionName = std::string(kAsanReportErrorTemplate) +
1038 (AccessIsWrite ? "store" : "load") + itostr(1 << AccessSizeIndex);
1039 // If we are merging crash callbacks, they have two parameters.
1040 AsanErrorCallback[AccessIsWrite][AccessSizeIndex] =
1041 checkInterfaceFunction(M.getOrInsertFunction(
1042 FunctionName, IRB.getVoidTy(), IntptrTy, NULL));
1045 AsanErrorCallbackSized[0] = checkInterfaceFunction(M.getOrInsertFunction(
1046 kAsanReportLoadN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1047 AsanErrorCallbackSized[1] = checkInterfaceFunction(M.getOrInsertFunction(
1048 kAsanReportStoreN, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1050 AsanHandleNoReturnFunc = checkInterfaceFunction(M.getOrInsertFunction(
1051 kAsanHandleNoReturnName, IRB.getVoidTy(), NULL));
1052 // We insert an empty inline asm after __asan_report* to avoid callback merge.
1053 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
1054 StringRef(""), StringRef(""),
1055 /*hasSideEffects=*/true);
1058 void AddressSanitizer::emitShadowMapping(Module &M, IRBuilder<> &IRB) const {
1059 // Tell the values of mapping offset and scale to the run-time.
1060 GlobalValue *asan_mapping_offset =
1061 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1062 ConstantInt::get(IntptrTy, Mapping.Offset),
1063 kAsanMappingOffsetName);
1064 // Read the global, otherwise it may be optimized away.
1065 IRB.CreateLoad(asan_mapping_offset, true);
1067 GlobalValue *asan_mapping_scale =
1068 new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
1069 ConstantInt::get(IntptrTy, Mapping.Scale),
1070 kAsanMappingScaleName);
1071 // Read the global, otherwise it may be optimized away.
1072 IRB.CreateLoad(asan_mapping_scale, true);
1076 bool AddressSanitizer::doInitialization(Module &M) {
1077 // Initialize the private fields. No one has accessed them before.
1078 TD = getAnalysisIfAvailable<DataLayout>();
1082 BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
1083 DynamicallyInitializedGlobals.Init(M);
1085 C = &(M.getContext());
1086 LongSize = TD->getPointerSizeInBits();
1087 IntptrTy = Type::getIntNTy(*C, LongSize);
1089 AsanCtorFunction = Function::Create(
1090 FunctionType::get(Type::getVoidTy(*C), false),
1091 GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
1092 BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
1093 // call __asan_init in the module ctor.
1094 IRBuilder<> IRB(ReturnInst::Create(*C, AsanCtorBB));
1095 AsanInitFunction = checkInterfaceFunction(
1096 M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
1097 AsanInitFunction->setLinkage(Function::ExternalLinkage);
1098 IRB.CreateCall(AsanInitFunction);
1100 Mapping = getShadowMapping(M, LongSize, ZeroBaseShadow);
1101 emitShadowMapping(M, IRB);
1103 appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority);
1107 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
1108 // For each NSObject descendant having a +load method, this method is invoked
1109 // by the ObjC runtime before any of the static constructors is called.
1110 // Therefore we need to instrument such methods with a call to __asan_init
1111 // at the beginning in order to initialize our runtime before any access to
1112 // the shadow memory.
1113 // We cannot just ignore these methods, because they may call other
1114 // instrumented functions.
1115 if (F.getName().find(" load]") != std::string::npos) {
1116 IRBuilder<> IRB(F.begin()->begin());
1117 IRB.CreateCall(AsanInitFunction);
1123 bool AddressSanitizer::runOnFunction(Function &F) {
1124 if (BL->isIn(F)) return false;
1125 if (&F == AsanCtorFunction) return false;
1126 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
1127 DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
1128 initializeCallbacks(*F.getParent());
1130 // If needed, insert __asan_init before checking for SanitizeAddress attr.
1131 maybeInsertAsanInitAtFunctionEntry(F);
1133 if (!F.hasFnAttribute(Attribute::SanitizeAddress))
1136 if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
1139 // We want to instrument every address only once per basic block (unless there
1140 // are calls between uses).
1141 SmallSet<Value*, 16> TempsToInstrument;
1142 SmallVector<Instruction*, 16> ToInstrument;
1143 SmallVector<Instruction*, 8> NoReturnCalls;
1147 // Fill the set of memory operations to instrument.
1148 for (Function::iterator FI = F.begin(), FE = F.end();
1150 TempsToInstrument.clear();
1151 int NumInsnsPerBB = 0;
1152 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1154 if (LooksLikeCodeInBug11395(BI)) return false;
1155 if (Value *Addr = isInterestingMemoryAccess(BI, &IsWrite)) {
1156 if (ClOpt && ClOptSameTemp) {
1157 if (!TempsToInstrument.insert(Addr))
1158 continue; // We've seen this temp in the current BB.
1160 } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
1163 if (isa<AllocaInst>(BI))
1167 // A call inside BB.
1168 TempsToInstrument.clear();
1169 if (CS.doesNotReturn())
1170 NoReturnCalls.push_back(CS.getInstruction());
1174 ToInstrument.push_back(BI);
1176 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB)
1181 Function *UninstrumentedDuplicate = 0;
1182 bool LikelyToInstrument =
1183 !NoReturnCalls.empty() || !ToInstrument.empty() || (NumAllocas > 0);
1184 if (ClKeepUninstrumented && LikelyToInstrument) {
1185 ValueToValueMapTy VMap;
1186 UninstrumentedDuplicate = CloneFunction(&F, VMap, false);
1187 UninstrumentedDuplicate->removeFnAttr(Attribute::SanitizeAddress);
1188 UninstrumentedDuplicate->setName("NOASAN_" + F.getName());
1189 F.getParent()->getFunctionList().push_back(UninstrumentedDuplicate);
1193 int NumInstrumented = 0;
1194 for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
1195 Instruction *Inst = ToInstrument[i];
1196 if (ClDebugMin < 0 || ClDebugMax < 0 ||
1197 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
1198 if (isInterestingMemoryAccess(Inst, &IsWrite))
1199 instrumentMop(Inst);
1201 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
1206 FunctionStackPoisoner FSP(F, *this);
1207 bool ChangedStack = FSP.runOnFunction();
1209 // We must unpoison the stack before every NoReturn call (throw, _exit, etc).
1210 // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37
1211 for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) {
1212 Instruction *CI = NoReturnCalls[i];
1213 IRBuilder<> IRB(CI);
1214 IRB.CreateCall(AsanHandleNoReturnFunc);
1217 bool res = NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty();
1218 DEBUG(dbgs() << "ASAN done instrumenting: " << res << " " << F << "\n");
1220 if (ClKeepUninstrumented) {
1222 // No instrumentation is done, no need for the duplicate.
1223 if (UninstrumentedDuplicate)
1224 UninstrumentedDuplicate->eraseFromParent();
1226 // The function was instrumented. We must have the duplicate.
1227 assert(UninstrumentedDuplicate);
1228 UninstrumentedDuplicate->setSection("NOASAN");
1229 assert(!F.hasSection());
1230 F.setSection("ASAN");
1237 static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
1238 if (ShadowRedzoneSize == 1) return PoisonByte;
1239 if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
1240 if (ShadowRedzoneSize == 4)
1241 return (PoisonByte << 24) + (PoisonByte << 16) +
1242 (PoisonByte << 8) + (PoisonByte);
1243 llvm_unreachable("ShadowRedzoneSize is either 1, 2 or 4");
1246 static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
1249 size_t ShadowGranularity,
1251 for (size_t i = 0; i < RZSize;
1252 i+= ShadowGranularity, Shadow++) {
1253 if (i + ShadowGranularity <= Size) {
1254 *Shadow = 0; // fully addressable
1255 } else if (i >= Size) {
1256 *Shadow = Magic; // unaddressable
1258 *Shadow = Size - i; // first Size-i bytes are addressable
1263 // Workaround for bug 11395: we don't want to instrument stack in functions
1264 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
1265 // FIXME: remove once the bug 11395 is fixed.
1266 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
1267 if (LongSize != 32) return false;
1268 CallInst *CI = dyn_cast<CallInst>(I);
1269 if (!CI || !CI->isInlineAsm()) return false;
1270 if (CI->getNumArgOperands() <= 5) return false;
1271 // We have inline assembly with quite a few arguments.
1275 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
1276 IRBuilder<> IRB(*C);
1277 for (int i = 0; i <= kMaxAsanStackMallocSizeClass; i++) {
1278 std::string Suffix = itostr(i);
1279 AsanStackMallocFunc[i] = checkInterfaceFunction(
1280 M.getOrInsertFunction(kAsanStackMallocNameTemplate + Suffix, IntptrTy,
1281 IntptrTy, IntptrTy, NULL));
1282 AsanStackFreeFunc[i] = checkInterfaceFunction(M.getOrInsertFunction(
1283 kAsanStackFreeNameTemplate + Suffix, IRB.getVoidTy(), IntptrTy,
1284 IntptrTy, IntptrTy, NULL));
1286 AsanPoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1287 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1288 AsanUnpoisonStackMemoryFunc = checkInterfaceFunction(M.getOrInsertFunction(
1289 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
1292 void FunctionStackPoisoner::poisonRedZones(
1293 const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> &IRB, Value *ShadowBase,
1295 size_t ShadowRZSize = RedzoneSize() >> Mapping.Scale;
1296 assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
1297 Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
1298 Type *RZPtrTy = PointerType::get(RZTy, 0);
1300 Value *PoisonLeft = ConstantInt::get(RZTy,
1301 ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
1302 Value *PoisonMid = ConstantInt::get(RZTy,
1303 ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
1304 Value *PoisonRight = ConstantInt::get(RZTy,
1305 ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
1307 // poison the first red zone.
1308 IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
1310 // poison all other red zones.
1311 uint64_t Pos = RedzoneSize();
1312 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1313 AllocaInst *AI = AllocaVec[i];
1314 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1315 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1316 assert(AlignedSize - SizeInBytes < RedzoneSize());
1321 assert(ShadowBase->getType() == IntptrTy);
1322 if (SizeInBytes < AlignedSize) {
1323 // Poison the partial redzone at right
1324 Ptr = IRB.CreateAdd(
1325 ShadowBase, ConstantInt::get(IntptrTy,
1326 (Pos >> Mapping.Scale) - ShadowRZSize));
1327 size_t AddressableBytes = RedzoneSize() - (AlignedSize - SizeInBytes);
1328 uint32_t Poison = 0;
1330 PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
1332 1ULL << Mapping.Scale,
1333 kAsanStackPartialRedzoneMagic);
1335 ASan.TD->isLittleEndian()
1336 ? support::endian::byte_swap<uint32_t, support::little>(Poison)
1337 : support::endian::byte_swap<uint32_t, support::big>(Poison);
1339 Value *PartialPoison = ConstantInt::get(RZTy, Poison);
1340 IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1343 // Poison the full redzone at right.
1344 Ptr = IRB.CreateAdd(ShadowBase,
1345 ConstantInt::get(IntptrTy, Pos >> Mapping.Scale));
1346 bool LastAlloca = (i == AllocaVec.size() - 1);
1347 Value *Poison = LastAlloca ? PoisonRight : PoisonMid;
1348 IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
1350 Pos += RedzoneSize();
1354 // Fake stack allocator (asan_fake_stack.h) has 11 size classes
1355 // for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
1356 static int StackMallocSizeClass(uint64_t LocalStackSize) {
1357 assert(LocalStackSize <= kMaxStackMallocSize);
1358 uint64_t MaxSize = kMinStackMallocSize;
1359 for (int i = 0; ; i++, MaxSize *= 2)
1360 if (LocalStackSize <= MaxSize)
1362 llvm_unreachable("impossible LocalStackSize");
1365 void FunctionStackPoisoner::poisonStack() {
1366 uint64_t LocalStackSize = TotalStackSize +
1367 (AllocaVec.size() + 1) * RedzoneSize();
1369 bool DoStackMalloc = ASan.CheckUseAfterReturn
1370 && LocalStackSize <= kMaxStackMallocSize;
1371 int StackMallocIdx = -1;
1373 assert(AllocaVec.size() > 0);
1374 Instruction *InsBefore = AllocaVec[0];
1375 IRBuilder<> IRB(InsBefore);
1378 Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
1379 AllocaInst *MyAlloca =
1380 new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
1381 if (ClRealignStack && StackAlignment < RedzoneSize())
1382 StackAlignment = RedzoneSize();
1383 MyAlloca->setAlignment(StackAlignment);
1384 assert(MyAlloca->isStaticAlloca());
1385 Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
1386 Value *LocalStackBase = OrigStackBase;
1388 if (DoStackMalloc) {
1389 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
1390 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
1391 LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc[StackMallocIdx],
1392 ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
1395 // This string will be parsed by the run-time (DescribeAddressIfStack).
1396 SmallString<2048> StackDescriptionStorage;
1397 raw_svector_ostream StackDescription(StackDescriptionStorage);
1398 StackDescription << AllocaVec.size() << " ";
1400 // Insert poison calls for lifetime intrinsics for alloca.
1401 bool HavePoisonedAllocas = false;
1402 for (size_t i = 0, n = AllocaPoisonCallVec.size(); i < n; i++) {
1403 const AllocaPoisonCall &APC = AllocaPoisonCallVec[i];
1404 IntrinsicInst *II = APC.InsBefore;
1405 AllocaInst *AI = findAllocaForValue(II->getArgOperand(1));
1407 IRBuilder<> IRB(II);
1408 poisonAlloca(AI, APC.Size, IRB, APC.DoPoison);
1409 HavePoisonedAllocas |= APC.DoPoison;
1412 uint64_t Pos = RedzoneSize();
1413 // Replace Alloca instructions with base+offset.
1414 for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
1415 AllocaInst *AI = AllocaVec[i];
1416 uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
1417 StringRef Name = AI->getName();
1418 StackDescription << Pos << " " << SizeInBytes << " "
1419 << Name.size() << " " << Name << " ";
1420 uint64_t AlignedSize = getAlignedAllocaSize(AI);
1421 assert((AlignedSize % RedzoneSize()) == 0);
1422 Value *NewAllocaPtr = IRB.CreateIntToPtr(
1423 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
1425 replaceDbgDeclareForAlloca(AI, NewAllocaPtr, DIB);
1426 AI->replaceAllUsesWith(NewAllocaPtr);
1427 Pos += AlignedSize + RedzoneSize();
1429 assert(Pos == LocalStackSize);
1431 // The left-most redzone has enough space for at least 4 pointers.
1432 // Write the Magic value to redzone[0].
1433 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
1434 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
1436 // Write the frame description constant to redzone[1].
1437 Value *BasePlus1 = IRB.CreateIntToPtr(
1438 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, ASan.LongSize/8)),
1440 GlobalVariable *StackDescriptionGlobal =
1441 createPrivateGlobalForString(*F.getParent(), StackDescription.str());
1442 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal,
1444 IRB.CreateStore(Description, BasePlus1);
1445 // Write the PC to redzone[2].
1446 Value *BasePlus2 = IRB.CreateIntToPtr(
1447 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy,
1448 2 * ASan.LongSize/8)),
1450 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
1452 // Poison the stack redzones at the entry.
1453 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
1454 poisonRedZones(AllocaVec, IRB, ShadowBase, true);
1456 // Unpoison the stack before all ret instructions.
1457 for (size_t i = 0, n = RetVec.size(); i < n; i++) {
1458 Instruction *Ret = RetVec[i];
1459 IRBuilder<> IRBRet(Ret);
1460 // Mark the current frame as retired.
1461 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
1463 // Unpoison the stack.
1464 poisonRedZones(AllocaVec, IRBRet, ShadowBase, false);
1465 if (DoStackMalloc) {
1466 assert(StackMallocIdx >= 0);
1467 // In use-after-return mode, mark the whole stack frame unaddressable.
1468 IRBRet.CreateCall3(AsanStackFreeFunc[StackMallocIdx], LocalStackBase,
1469 ConstantInt::get(IntptrTy, LocalStackSize),
1471 } else if (HavePoisonedAllocas) {
1472 // If we poisoned some allocas in llvm.lifetime analysis,
1473 // unpoison whole stack frame now.
1474 assert(LocalStackBase == OrigStackBase);
1475 poisonAlloca(LocalStackBase, LocalStackSize, IRBRet, false);
1479 // We are done. Remove the old unused alloca instructions.
1480 for (size_t i = 0, n = AllocaVec.size(); i < n; i++)
1481 AllocaVec[i]->eraseFromParent();
1484 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
1485 IRBuilder<> &IRB, bool DoPoison) {
1486 // For now just insert the call to ASan runtime.
1487 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
1488 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
1489 IRB.CreateCall2(DoPoison ? AsanPoisonStackMemoryFunc
1490 : AsanUnpoisonStackMemoryFunc,
1494 // Handling llvm.lifetime intrinsics for a given %alloca:
1495 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
1496 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
1497 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
1498 // could be poisoned by previous llvm.lifetime.end instruction, as the
1499 // variable may go in and out of scope several times, e.g. in loops).
1500 // (3) if we poisoned at least one %alloca in a function,
1501 // unpoison the whole stack frame at function exit.
1503 AllocaInst *FunctionStackPoisoner::findAllocaForValue(Value *V) {
1504 if (AllocaInst *AI = dyn_cast<AllocaInst>(V))
1505 // We're intested only in allocas we can handle.
1506 return isInterestingAlloca(*AI) ? AI : 0;
1507 // See if we've already calculated (or started to calculate) alloca for a
1509 AllocaForValueMapTy::iterator I = AllocaForValue.find(V);
1510 if (I != AllocaForValue.end())
1512 // Store 0 while we're calculating alloca for value V to avoid
1513 // infinite recursion if the value references itself.
1514 AllocaForValue[V] = 0;
1515 AllocaInst *Res = 0;
1516 if (CastInst *CI = dyn_cast<CastInst>(V))
1517 Res = findAllocaForValue(CI->getOperand(0));
1518 else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1519 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1520 Value *IncValue = PN->getIncomingValue(i);
1521 // Allow self-referencing phi-nodes.
1522 if (IncValue == PN) continue;
1523 AllocaInst *IncValueAI = findAllocaForValue(IncValue);
1524 // AI for incoming values should exist and should all be equal.
1525 if (IncValueAI == 0 || (Res != 0 && IncValueAI != Res))
1531 AllocaForValue[V] = Res;