1 //===-- AddressSanitizer.cpp - memory error detector ------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 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 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DepthFirstIterator.h" 19 #include "llvm/ADT/SetVector.h" 20 #include "llvm/ADT/SmallSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/ADT/StringExtras.h" 24 #include "llvm/ADT/Triple.h" 25 #include "llvm/Analysis/MemoryBuiltins.h" 26 #include "llvm/Analysis/TargetLibraryInfo.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/IR/CallSite.h" 29 #include "llvm/IR/DIBuilder.h" 30 #include "llvm/IR/DataLayout.h" 31 #include "llvm/IR/Dominators.h" 32 #include "llvm/IR/Function.h" 33 #include "llvm/IR/IRBuilder.h" 34 #include "llvm/IR/InlineAsm.h" 35 #include "llvm/IR/InstVisitor.h" 36 #include "llvm/IR/IntrinsicInst.h" 37 #include "llvm/IR/LLVMContext.h" 38 #include "llvm/IR/MDBuilder.h" 39 #include "llvm/IR/Module.h" 40 #include "llvm/IR/Type.h" 41 #include "llvm/MC/MCSectionMachO.h" 42 #include "llvm/Support/CommandLine.h" 43 #include "llvm/Support/DataTypes.h" 44 #include "llvm/Support/Debug.h" 45 #include "llvm/Support/Endian.h" 46 #include "llvm/Support/SwapByteOrder.h" 47 #include "llvm/Support/raw_ostream.h" 48 #include "llvm/Transforms/Instrumentation.h" 49 #include "llvm/Transforms/Scalar.h" 50 #include "llvm/Transforms/Utils/ASanStackFrameLayout.h" 51 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 52 #include "llvm/Transforms/Utils/Cloning.h" 53 #include "llvm/Transforms/Utils/Local.h" 54 #include "llvm/Transforms/Utils/ModuleUtils.h" 55 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 56 #include <algorithm> 57 #include <iomanip> 58 #include <limits> 59 #include <sstream> 60 #include <string> 61 #include <system_error> 62 63 using namespace llvm; 64 65 #define DEBUG_TYPE "asan" 66 67 static const uint64_t kDefaultShadowScale = 3; 68 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29; 69 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44; 70 static const uint64_t kDynamicShadowSentinel = ~(uint64_t)0; 71 static const uint64_t kIOSShadowOffset32 = 1ULL << 30; 72 static const uint64_t kIOSSimShadowOffset32 = 1ULL << 30; 73 static const uint64_t kIOSSimShadowOffset64 = kDefaultShadowOffset64; 74 static const uint64_t kSmallX86_64ShadowOffset = 0x7FFF8000; // < 2G. 75 static const uint64_t kLinuxKasan_ShadowOffset64 = 0xdffffc0000000000; 76 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 41; 77 static const uint64_t kSystemZ_ShadowOffset64 = 1ULL << 52; 78 static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa0000; 79 static const uint64_t kMIPS64_ShadowOffset64 = 1ULL << 37; 80 static const uint64_t kAArch64_ShadowOffset64 = 1ULL << 36; 81 static const uint64_t kFreeBSD_ShadowOffset32 = 1ULL << 30; 82 static const uint64_t kFreeBSD_ShadowOffset64 = 1ULL << 46; 83 static const uint64_t kWindowsShadowOffset32 = 3ULL << 28; 84 // The shadow memory space is dynamically allocated. 85 static const uint64_t kWindowsShadowOffset64 = kDynamicShadowSentinel; 86 87 static const size_t kMinStackMallocSize = 1 << 6; // 64B 88 static const size_t kMaxStackMallocSize = 1 << 16; // 64K 89 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3; 90 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E; 91 92 static const char *const kAsanModuleCtorName = "asan.module_ctor"; 93 static const char *const kAsanModuleDtorName = "asan.module_dtor"; 94 static const uint64_t kAsanCtorAndDtorPriority = 1; 95 static const char *const kAsanReportErrorTemplate = "__asan_report_"; 96 static const char *const kAsanRegisterGlobalsName = "__asan_register_globals"; 97 static const char *const kAsanUnregisterGlobalsName = 98 "__asan_unregister_globals"; 99 static const char *const kAsanRegisterImageGlobalsName = 100 "__asan_register_image_globals"; 101 static const char *const kAsanUnregisterImageGlobalsName = 102 "__asan_unregister_image_globals"; 103 static const char *const kAsanPoisonGlobalsName = "__asan_before_dynamic_init"; 104 static const char *const kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init"; 105 static const char *const kAsanInitName = "__asan_init"; 106 static const char *const kAsanVersionCheckName = 107 "__asan_version_mismatch_check_v8"; 108 static const char *const kAsanPtrCmp = "__sanitizer_ptr_cmp"; 109 static const char *const kAsanPtrSub = "__sanitizer_ptr_sub"; 110 static const char *const kAsanHandleNoReturnName = "__asan_handle_no_return"; 111 static const int kMaxAsanStackMallocSizeClass = 10; 112 static const char *const kAsanStackMallocNameTemplate = "__asan_stack_malloc_"; 113 static const char *const kAsanStackFreeNameTemplate = "__asan_stack_free_"; 114 static const char *const kAsanGenPrefix = "__asan_gen_"; 115 static const char *const kODRGenPrefix = "__odr_asan_gen_"; 116 static const char *const kSanCovGenPrefix = "__sancov_gen_"; 117 static const char *const kAsanSetShadowPrefix = "__asan_set_shadow_"; 118 static const char *const kAsanPoisonStackMemoryName = 119 "__asan_poison_stack_memory"; 120 static const char *const kAsanUnpoisonStackMemoryName = 121 "__asan_unpoison_stack_memory"; 122 static const char *const kAsanGlobalsRegisteredFlagName = 123 "__asan_globals_registered"; 124 125 static const char *const kAsanOptionDetectUseAfterReturn = 126 "__asan_option_detect_stack_use_after_return"; 127 128 static const char *const kAsanShadowMemoryDynamicAddress = 129 "__asan_shadow_memory_dynamic_address"; 130 131 static const char *const kAsanAllocaPoison = "__asan_alloca_poison"; 132 static const char *const kAsanAllocasUnpoison = "__asan_allocas_unpoison"; 133 134 // Accesses sizes are powers of two: 1, 2, 4, 8, 16. 135 static const size_t kNumberOfAccessSizes = 5; 136 137 static const unsigned kAllocaRzSize = 32; 138 139 // Command-line flags. 140 static cl::opt<bool> ClEnableKasan( 141 "asan-kernel", cl::desc("Enable KernelAddressSanitizer instrumentation"), 142 cl::Hidden, cl::init(false)); 143 static cl::opt<bool> ClRecover( 144 "asan-recover", 145 cl::desc("Enable recovery mode (continue-after-error)."), 146 cl::Hidden, cl::init(false)); 147 148 // This flag may need to be replaced with -f[no-]asan-reads. 149 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads", 150 cl::desc("instrument read instructions"), 151 cl::Hidden, cl::init(true)); 152 static cl::opt<bool> ClInstrumentWrites( 153 "asan-instrument-writes", cl::desc("instrument write instructions"), 154 cl::Hidden, cl::init(true)); 155 static cl::opt<bool> ClInstrumentAtomics( 156 "asan-instrument-atomics", 157 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden, 158 cl::init(true)); 159 static cl::opt<bool> ClAlwaysSlowPath( 160 "asan-always-slow-path", 161 cl::desc("use instrumentation with slow path for all accesses"), cl::Hidden, 162 cl::init(false)); 163 static cl::opt<bool> ClForceDynamicShadow( 164 "asan-force-dynamic-shadow", 165 cl::desc("Load shadow address into a local variable for each function"), 166 cl::Hidden, cl::init(false)); 167 168 // This flag limits the number of instructions to be instrumented 169 // in any given BB. Normally, this should be set to unlimited (INT_MAX), 170 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary 171 // set it to 10000. 172 static cl::opt<int> ClMaxInsnsToInstrumentPerBB( 173 "asan-max-ins-per-bb", cl::init(10000), 174 cl::desc("maximal number of instructions to instrument in any given BB"), 175 cl::Hidden); 176 // This flag may need to be replaced with -f[no]asan-stack. 177 static cl::opt<bool> ClStack("asan-stack", cl::desc("Handle stack memory"), 178 cl::Hidden, cl::init(true)); 179 static cl::opt<uint32_t> ClMaxInlinePoisoningSize( 180 "asan-max-inline-poisoning-size", 181 cl::desc( 182 "Inline shadow poisoning for blocks up to the given size in bytes."), 183 cl::Hidden, cl::init(64)); 184 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return", 185 cl::desc("Check stack-use-after-return"), 186 cl::Hidden, cl::init(true)); 187 static cl::opt<bool> ClUseAfterScope("asan-use-after-scope", 188 cl::desc("Check stack-use-after-scope"), 189 cl::Hidden, cl::init(false)); 190 // This flag may need to be replaced with -f[no]asan-globals. 191 static cl::opt<bool> ClGlobals("asan-globals", 192 cl::desc("Handle global objects"), cl::Hidden, 193 cl::init(true)); 194 static cl::opt<bool> ClInitializers("asan-initialization-order", 195 cl::desc("Handle C++ initializer order"), 196 cl::Hidden, cl::init(true)); 197 static cl::opt<bool> ClInvalidPointerPairs( 198 "asan-detect-invalid-pointer-pair", 199 cl::desc("Instrument <, <=, >, >=, - with pointer operands"), cl::Hidden, 200 cl::init(false)); 201 static cl::opt<unsigned> ClRealignStack( 202 "asan-realign-stack", 203 cl::desc("Realign stack to the value of this flag (power of two)"), 204 cl::Hidden, cl::init(32)); 205 static cl::opt<int> ClInstrumentationWithCallsThreshold( 206 "asan-instrumentation-with-call-threshold", 207 cl::desc( 208 "If the function being instrumented contains more than " 209 "this number of memory accesses, use callbacks instead of " 210 "inline checks (-1 means never use callbacks)."), 211 cl::Hidden, cl::init(7000)); 212 static cl::opt<std::string> ClMemoryAccessCallbackPrefix( 213 "asan-memory-access-callback-prefix", 214 cl::desc("Prefix for memory access callbacks"), cl::Hidden, 215 cl::init("__asan_")); 216 static cl::opt<bool> 217 ClInstrumentDynamicAllocas("asan-instrument-dynamic-allocas", 218 cl::desc("instrument dynamic allocas"), 219 cl::Hidden, cl::init(true)); 220 static cl::opt<bool> ClSkipPromotableAllocas( 221 "asan-skip-promotable-allocas", 222 cl::desc("Do not instrument promotable allocas"), cl::Hidden, 223 cl::init(true)); 224 225 // These flags allow to change the shadow mapping. 226 // The shadow mapping looks like 227 // Shadow = (Mem >> scale) + offset 228 static cl::opt<int> ClMappingScale("asan-mapping-scale", 229 cl::desc("scale of asan shadow mapping"), 230 cl::Hidden, cl::init(0)); 231 static cl::opt<unsigned long long> ClMappingOffset( 232 "asan-mapping-offset", 233 cl::desc("offset of asan shadow mapping [EXPERIMENTAL]"), cl::Hidden, 234 cl::init(0)); 235 236 // Optimization flags. Not user visible, used mostly for testing 237 // and benchmarking the tool. 238 static cl::opt<bool> ClOpt("asan-opt", cl::desc("Optimize instrumentation"), 239 cl::Hidden, cl::init(true)); 240 static cl::opt<bool> ClOptSameTemp( 241 "asan-opt-same-temp", cl::desc("Instrument the same temp just once"), 242 cl::Hidden, cl::init(true)); 243 static cl::opt<bool> ClOptGlobals("asan-opt-globals", 244 cl::desc("Don't instrument scalar globals"), 245 cl::Hidden, cl::init(true)); 246 static cl::opt<bool> ClOptStack( 247 "asan-opt-stack", cl::desc("Don't instrument scalar stack variables"), 248 cl::Hidden, cl::init(false)); 249 250 static cl::opt<bool> ClDynamicAllocaStack( 251 "asan-stack-dynamic-alloca", 252 cl::desc("Use dynamic alloca to represent stack variables"), cl::Hidden, 253 cl::init(true)); 254 255 static cl::opt<uint32_t> ClForceExperiment( 256 "asan-force-experiment", 257 cl::desc("Force optimization experiment (for testing)"), cl::Hidden, 258 cl::init(0)); 259 260 static cl::opt<bool> 261 ClUsePrivateAliasForGlobals("asan-use-private-alias", 262 cl::desc("Use private aliases for global" 263 " variables"), 264 cl::Hidden, cl::init(false)); 265 266 static cl::opt<bool> 267 ClUseMachOGlobalsSection("asan-globals-live-support", 268 cl::desc("Use linker features to support dead " 269 "code stripping of globals " 270 "(Mach-O only)"), 271 cl::Hidden, cl::init(true)); 272 273 // Debug flags. 274 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden, 275 cl::init(0)); 276 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"), 277 cl::Hidden, cl::init(0)); 278 static cl::opt<std::string> ClDebugFunc("asan-debug-func", cl::Hidden, 279 cl::desc("Debug func")); 280 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"), 281 cl::Hidden, cl::init(-1)); 282 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug max inst"), 283 cl::Hidden, cl::init(-1)); 284 285 STATISTIC(NumInstrumentedReads, "Number of instrumented reads"); 286 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes"); 287 STATISTIC(NumOptimizedAccessesToGlobalVar, 288 "Number of optimized accesses to global vars"); 289 STATISTIC(NumOptimizedAccessesToStackVar, 290 "Number of optimized accesses to stack vars"); 291 292 namespace { 293 /// Frontend-provided metadata for source location. 294 struct LocationMetadata { 295 StringRef Filename; 296 int LineNo; 297 int ColumnNo; 298 299 LocationMetadata() : Filename(), LineNo(0), ColumnNo(0) {} 300 301 bool empty() const { return Filename.empty(); } 302 303 void parse(MDNode *MDN) { 304 assert(MDN->getNumOperands() == 3); 305 MDString *DIFilename = cast<MDString>(MDN->getOperand(0)); 306 Filename = DIFilename->getString(); 307 LineNo = 308 mdconst::extract<ConstantInt>(MDN->getOperand(1))->getLimitedValue(); 309 ColumnNo = 310 mdconst::extract<ConstantInt>(MDN->getOperand(2))->getLimitedValue(); 311 } 312 }; 313 314 /// Frontend-provided metadata for global variables. 315 class GlobalsMetadata { 316 public: 317 struct Entry { 318 Entry() : SourceLoc(), Name(), IsDynInit(false), IsBlacklisted(false) {} 319 LocationMetadata SourceLoc; 320 StringRef Name; 321 bool IsDynInit; 322 bool IsBlacklisted; 323 }; 324 325 GlobalsMetadata() : inited_(false) {} 326 327 void reset() { 328 inited_ = false; 329 Entries.clear(); 330 } 331 332 void init(Module &M) { 333 assert(!inited_); 334 inited_ = true; 335 NamedMDNode *Globals = M.getNamedMetadata("llvm.asan.globals"); 336 if (!Globals) return; 337 for (auto MDN : Globals->operands()) { 338 // Metadata node contains the global and the fields of "Entry". 339 assert(MDN->getNumOperands() == 5); 340 auto *GV = mdconst::extract_or_null<GlobalVariable>(MDN->getOperand(0)); 341 // The optimizer may optimize away a global entirely. 342 if (!GV) continue; 343 // We can already have an entry for GV if it was merged with another 344 // global. 345 Entry &E = Entries[GV]; 346 if (auto *Loc = cast_or_null<MDNode>(MDN->getOperand(1))) 347 E.SourceLoc.parse(Loc); 348 if (auto *Name = cast_or_null<MDString>(MDN->getOperand(2))) 349 E.Name = Name->getString(); 350 ConstantInt *IsDynInit = 351 mdconst::extract<ConstantInt>(MDN->getOperand(3)); 352 E.IsDynInit |= IsDynInit->isOne(); 353 ConstantInt *IsBlacklisted = 354 mdconst::extract<ConstantInt>(MDN->getOperand(4)); 355 E.IsBlacklisted |= IsBlacklisted->isOne(); 356 } 357 } 358 359 /// Returns metadata entry for a given global. 360 Entry get(GlobalVariable *G) const { 361 auto Pos = Entries.find(G); 362 return (Pos != Entries.end()) ? Pos->second : Entry(); 363 } 364 365 private: 366 bool inited_; 367 DenseMap<GlobalVariable *, Entry> Entries; 368 }; 369 370 /// This struct defines the shadow mapping using the rule: 371 /// shadow = (mem >> Scale) ADD-or-OR Offset. 372 struct ShadowMapping { 373 int Scale; 374 uint64_t Offset; 375 bool OrShadowOffset; 376 }; 377 378 static ShadowMapping getShadowMapping(Triple &TargetTriple, int LongSize, 379 bool IsKasan) { 380 bool IsAndroid = TargetTriple.isAndroid(); 381 bool IsIOS = TargetTriple.isiOS() || TargetTriple.isWatchOS(); 382 bool IsFreeBSD = TargetTriple.isOSFreeBSD(); 383 bool IsLinux = TargetTriple.isOSLinux(); 384 bool IsPPC64 = TargetTriple.getArch() == llvm::Triple::ppc64 || 385 TargetTriple.getArch() == llvm::Triple::ppc64le; 386 bool IsSystemZ = TargetTriple.getArch() == llvm::Triple::systemz; 387 bool IsX86 = TargetTriple.getArch() == llvm::Triple::x86; 388 bool IsX86_64 = TargetTriple.getArch() == llvm::Triple::x86_64; 389 bool IsMIPS32 = TargetTriple.getArch() == llvm::Triple::mips || 390 TargetTriple.getArch() == llvm::Triple::mipsel; 391 bool IsMIPS64 = TargetTriple.getArch() == llvm::Triple::mips64 || 392 TargetTriple.getArch() == llvm::Triple::mips64el; 393 bool IsAArch64 = TargetTriple.getArch() == llvm::Triple::aarch64; 394 bool IsWindows = TargetTriple.isOSWindows(); 395 396 ShadowMapping Mapping; 397 398 if (LongSize == 32) { 399 // Android is always PIE, which means that the beginning of the address 400 // space is always available. 401 if (IsAndroid) 402 Mapping.Offset = 0; 403 else if (IsMIPS32) 404 Mapping.Offset = kMIPS32_ShadowOffset32; 405 else if (IsFreeBSD) 406 Mapping.Offset = kFreeBSD_ShadowOffset32; 407 else if (IsIOS) 408 // If we're targeting iOS and x86, the binary is built for iOS simulator. 409 Mapping.Offset = IsX86 ? kIOSSimShadowOffset32 : kIOSShadowOffset32; 410 else if (IsWindows) 411 Mapping.Offset = kWindowsShadowOffset32; 412 else 413 Mapping.Offset = kDefaultShadowOffset32; 414 } else { // LongSize == 64 415 if (IsPPC64) 416 Mapping.Offset = kPPC64_ShadowOffset64; 417 else if (IsSystemZ) 418 Mapping.Offset = kSystemZ_ShadowOffset64; 419 else if (IsFreeBSD) 420 Mapping.Offset = kFreeBSD_ShadowOffset64; 421 else if (IsLinux && IsX86_64) { 422 if (IsKasan) 423 Mapping.Offset = kLinuxKasan_ShadowOffset64; 424 else 425 Mapping.Offset = kSmallX86_64ShadowOffset; 426 } else if (IsWindows && IsX86_64) { 427 Mapping.Offset = kWindowsShadowOffset64; 428 } else if (IsMIPS64) 429 Mapping.Offset = kMIPS64_ShadowOffset64; 430 else if (IsIOS) 431 // If we're targeting iOS and x86, the binary is built for iOS simulator. 432 // We are using dynamic shadow offset on the 64-bit devices. 433 Mapping.Offset = 434 IsX86_64 ? kIOSSimShadowOffset64 : kDynamicShadowSentinel; 435 else if (IsAArch64) 436 Mapping.Offset = kAArch64_ShadowOffset64; 437 else 438 Mapping.Offset = kDefaultShadowOffset64; 439 } 440 441 if (ClForceDynamicShadow) { 442 Mapping.Offset = kDynamicShadowSentinel; 443 } 444 445 Mapping.Scale = kDefaultShadowScale; 446 if (ClMappingScale.getNumOccurrences() > 0) { 447 Mapping.Scale = ClMappingScale; 448 } 449 450 if (ClMappingOffset.getNumOccurrences() > 0) { 451 Mapping.Offset = ClMappingOffset; 452 } 453 454 // OR-ing shadow offset if more efficient (at least on x86) if the offset 455 // is a power of two, but on ppc64 we have to use add since the shadow 456 // offset is not necessary 1/8-th of the address space. On SystemZ, 457 // we could OR the constant in a single instruction, but it's more 458 // efficient to load it once and use indexed addressing. 459 Mapping.OrShadowOffset = !IsAArch64 && !IsPPC64 && !IsSystemZ 460 && !(Mapping.Offset & (Mapping.Offset - 1)) 461 && Mapping.Offset != kDynamicShadowSentinel; 462 463 return Mapping; 464 } 465 466 static size_t RedzoneSizeForScale(int MappingScale) { 467 // Redzone used for stack and globals is at least 32 bytes. 468 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively. 469 return std::max(32U, 1U << MappingScale); 470 } 471 472 /// AddressSanitizer: instrument the code in module to find memory bugs. 473 struct AddressSanitizer : public FunctionPass { 474 explicit AddressSanitizer(bool CompileKernel = false, bool Recover = false, 475 bool UseAfterScope = false) 476 : FunctionPass(ID), CompileKernel(CompileKernel || ClEnableKasan), 477 Recover(Recover || ClRecover), 478 UseAfterScope(UseAfterScope || ClUseAfterScope), 479 LocalDynamicShadow(nullptr) { 480 initializeAddressSanitizerPass(*PassRegistry::getPassRegistry()); 481 } 482 StringRef getPassName() const override { 483 return "AddressSanitizerFunctionPass"; 484 } 485 void getAnalysisUsage(AnalysisUsage &AU) const override { 486 AU.addRequired<DominatorTreeWrapperPass>(); 487 AU.addRequired<TargetLibraryInfoWrapperPass>(); 488 } 489 uint64_t getAllocaSizeInBytes(const AllocaInst &AI) const { 490 uint64_t ArraySize = 1; 491 if (AI.isArrayAllocation()) { 492 const ConstantInt *CI = dyn_cast<ConstantInt>(AI.getArraySize()); 493 assert(CI && "non-constant array size"); 494 ArraySize = CI->getZExtValue(); 495 } 496 Type *Ty = AI.getAllocatedType(); 497 uint64_t SizeInBytes = 498 AI.getModule()->getDataLayout().getTypeAllocSize(Ty); 499 return SizeInBytes * ArraySize; 500 } 501 /// Check if we want (and can) handle this alloca. 502 bool isInterestingAlloca(const AllocaInst &AI); 503 504 /// If it is an interesting memory access, return the PointerOperand 505 /// and set IsWrite/Alignment. Otherwise return nullptr. 506 /// MaybeMask is an output parameter for the mask Value, if we're looking at a 507 /// masked load/store. 508 Value *isInterestingMemoryAccess(Instruction *I, bool *IsWrite, 509 uint64_t *TypeSize, unsigned *Alignment, 510 Value **MaybeMask = nullptr); 511 void instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis, Instruction *I, 512 bool UseCalls, const DataLayout &DL); 513 void instrumentPointerComparisonOrSubtraction(Instruction *I); 514 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore, 515 Value *Addr, uint32_t TypeSize, bool IsWrite, 516 Value *SizeArgument, bool UseCalls, uint32_t Exp); 517 void instrumentUnusualSizeOrAlignment(Instruction *I, Value *Addr, 518 uint32_t TypeSize, bool IsWrite, 519 Value *SizeArgument, bool UseCalls, 520 uint32_t Exp); 521 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong, 522 Value *ShadowValue, uint32_t TypeSize); 523 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr, 524 bool IsWrite, size_t AccessSizeIndex, 525 Value *SizeArgument, uint32_t Exp); 526 void instrumentMemIntrinsic(MemIntrinsic *MI); 527 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB); 528 bool runOnFunction(Function &F) override; 529 bool maybeInsertAsanInitAtFunctionEntry(Function &F); 530 void maybeInsertDynamicShadowAtFunctionEntry(Function &F); 531 void markEscapedLocalAllocas(Function &F); 532 bool doInitialization(Module &M) override; 533 bool doFinalization(Module &M) override; 534 static char ID; // Pass identification, replacement for typeid 535 536 DominatorTree &getDominatorTree() const { return *DT; } 537 538 private: 539 void initializeCallbacks(Module &M); 540 541 bool LooksLikeCodeInBug11395(Instruction *I); 542 bool GlobalIsLinkerInitialized(GlobalVariable *G); 543 bool isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, Value *Addr, 544 uint64_t TypeSize) const; 545 546 /// Helper to cleanup per-function state. 547 struct FunctionStateRAII { 548 AddressSanitizer *Pass; 549 FunctionStateRAII(AddressSanitizer *Pass) : Pass(Pass) { 550 assert(Pass->ProcessedAllocas.empty() && 551 "last pass forgot to clear cache"); 552 assert(!Pass->LocalDynamicShadow); 553 } 554 ~FunctionStateRAII() { 555 Pass->LocalDynamicShadow = nullptr; 556 Pass->ProcessedAllocas.clear(); 557 } 558 }; 559 560 LLVMContext *C; 561 Triple TargetTriple; 562 int LongSize; 563 bool CompileKernel; 564 bool Recover; 565 bool UseAfterScope; 566 Type *IntptrTy; 567 ShadowMapping Mapping; 568 DominatorTree *DT; 569 Function *AsanCtorFunction = nullptr; 570 Function *AsanInitFunction = nullptr; 571 Function *AsanHandleNoReturnFunc; 572 Function *AsanPtrCmpFunction, *AsanPtrSubFunction; 573 // This array is indexed by AccessIsWrite, Experiment and log2(AccessSize). 574 Function *AsanErrorCallback[2][2][kNumberOfAccessSizes]; 575 Function *AsanMemoryAccessCallback[2][2][kNumberOfAccessSizes]; 576 // This array is indexed by AccessIsWrite and Experiment. 577 Function *AsanErrorCallbackSized[2][2]; 578 Function *AsanMemoryAccessCallbackSized[2][2]; 579 Function *AsanMemmove, *AsanMemcpy, *AsanMemset; 580 InlineAsm *EmptyAsm; 581 Value *LocalDynamicShadow; 582 GlobalsMetadata GlobalsMD; 583 DenseMap<const AllocaInst *, bool> ProcessedAllocas; 584 585 friend struct FunctionStackPoisoner; 586 }; 587 588 class AddressSanitizerModule : public ModulePass { 589 public: 590 explicit AddressSanitizerModule(bool CompileKernel = false, 591 bool Recover = false) 592 : ModulePass(ID), CompileKernel(CompileKernel || ClEnableKasan), 593 Recover(Recover || ClRecover) {} 594 bool runOnModule(Module &M) override; 595 static char ID; // Pass identification, replacement for typeid 596 StringRef getPassName() const override { return "AddressSanitizerModule"; } 597 598 private: 599 void initializeCallbacks(Module &M); 600 601 bool InstrumentGlobals(IRBuilder<> &IRB, Module &M); 602 bool ShouldInstrumentGlobal(GlobalVariable *G); 603 bool ShouldUseMachOGlobalsSection() const; 604 StringRef getGlobalMetadataSection() const; 605 void poisonOneInitializer(Function &GlobalInit, GlobalValue *ModuleName); 606 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName); 607 size_t MinRedzoneSizeForGlobal() const { 608 return RedzoneSizeForScale(Mapping.Scale); 609 } 610 611 GlobalsMetadata GlobalsMD; 612 bool CompileKernel; 613 bool Recover; 614 Type *IntptrTy; 615 LLVMContext *C; 616 Triple TargetTriple; 617 ShadowMapping Mapping; 618 Function *AsanPoisonGlobals; 619 Function *AsanUnpoisonGlobals; 620 Function *AsanRegisterGlobals; 621 Function *AsanUnregisterGlobals; 622 Function *AsanRegisterImageGlobals; 623 Function *AsanUnregisterImageGlobals; 624 }; 625 626 // Stack poisoning does not play well with exception handling. 627 // When an exception is thrown, we essentially bypass the code 628 // that unpoisones the stack. This is why the run-time library has 629 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire 630 // stack in the interceptor. This however does not work inside the 631 // actual function which catches the exception. Most likely because the 632 // compiler hoists the load of the shadow value somewhere too high. 633 // This causes asan to report a non-existing bug on 453.povray. 634 // It sounds like an LLVM bug. 635 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> { 636 Function &F; 637 AddressSanitizer &ASan; 638 DIBuilder DIB; 639 LLVMContext *C; 640 Type *IntptrTy; 641 Type *IntptrPtrTy; 642 ShadowMapping Mapping; 643 644 SmallVector<AllocaInst *, 16> AllocaVec; 645 SmallVector<AllocaInst *, 16> StaticAllocasToMoveUp; 646 SmallVector<Instruction *, 8> RetVec; 647 unsigned StackAlignment; 648 649 Function *AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1], 650 *AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1]; 651 Function *AsanSetShadowFunc[0x100] = {}; 652 Function *AsanPoisonStackMemoryFunc, *AsanUnpoisonStackMemoryFunc; 653 Function *AsanAllocaPoisonFunc, *AsanAllocasUnpoisonFunc; 654 655 // Stores a place and arguments of poisoning/unpoisoning call for alloca. 656 struct AllocaPoisonCall { 657 IntrinsicInst *InsBefore; 658 AllocaInst *AI; 659 uint64_t Size; 660 bool DoPoison; 661 }; 662 SmallVector<AllocaPoisonCall, 8> DynamicAllocaPoisonCallVec; 663 SmallVector<AllocaPoisonCall, 8> StaticAllocaPoisonCallVec; 664 665 SmallVector<AllocaInst *, 1> DynamicAllocaVec; 666 SmallVector<IntrinsicInst *, 1> StackRestoreVec; 667 AllocaInst *DynamicAllocaLayout = nullptr; 668 IntrinsicInst *LocalEscapeCall = nullptr; 669 670 // Maps Value to an AllocaInst from which the Value is originated. 671 typedef DenseMap<Value *, AllocaInst *> AllocaForValueMapTy; 672 AllocaForValueMapTy AllocaForValue; 673 674 bool HasNonEmptyInlineAsm = false; 675 bool HasReturnsTwiceCall = false; 676 std::unique_ptr<CallInst> EmptyInlineAsm; 677 678 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan) 679 : F(F), 680 ASan(ASan), 681 DIB(*F.getParent(), /*AllowUnresolved*/ false), 682 C(ASan.C), 683 IntptrTy(ASan.IntptrTy), 684 IntptrPtrTy(PointerType::get(IntptrTy, 0)), 685 Mapping(ASan.Mapping), 686 StackAlignment(1 << Mapping.Scale), 687 EmptyInlineAsm(CallInst::Create(ASan.EmptyAsm)) {} 688 689 bool runOnFunction() { 690 if (!ClStack) return false; 691 // Collect alloca, ret, lifetime instructions etc. 692 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) visit(*BB); 693 694 if (AllocaVec.empty() && DynamicAllocaVec.empty()) return false; 695 696 initializeCallbacks(*F.getParent()); 697 698 processDynamicAllocas(); 699 processStaticAllocas(); 700 701 if (ClDebugStack) { 702 DEBUG(dbgs() << F); 703 } 704 return true; 705 } 706 707 // Finds all Alloca instructions and puts 708 // poisoned red zones around all of them. 709 // Then unpoison everything back before the function returns. 710 void processStaticAllocas(); 711 void processDynamicAllocas(); 712 713 void createDynamicAllocasInitStorage(); 714 715 // ----------------------- Visitors. 716 /// \brief Collect all Ret instructions. 717 void visitReturnInst(ReturnInst &RI) { RetVec.push_back(&RI); } 718 719 /// \brief Collect all Resume instructions. 720 void visitResumeInst(ResumeInst &RI) { RetVec.push_back(&RI); } 721 722 /// \brief Collect all CatchReturnInst instructions. 723 void visitCleanupReturnInst(CleanupReturnInst &CRI) { RetVec.push_back(&CRI); } 724 725 void unpoisonDynamicAllocasBeforeInst(Instruction *InstBefore, 726 Value *SavedStack) { 727 IRBuilder<> IRB(InstBefore); 728 Value *DynamicAreaPtr = IRB.CreatePtrToInt(SavedStack, IntptrTy); 729 // When we insert _asan_allocas_unpoison before @llvm.stackrestore, we 730 // need to adjust extracted SP to compute the address of the most recent 731 // alloca. We have a special @llvm.get.dynamic.area.offset intrinsic for 732 // this purpose. 733 if (!isa<ReturnInst>(InstBefore)) { 734 Function *DynamicAreaOffsetFunc = Intrinsic::getDeclaration( 735 InstBefore->getModule(), Intrinsic::get_dynamic_area_offset, 736 {IntptrTy}); 737 738 Value *DynamicAreaOffset = IRB.CreateCall(DynamicAreaOffsetFunc, {}); 739 740 DynamicAreaPtr = IRB.CreateAdd(IRB.CreatePtrToInt(SavedStack, IntptrTy), 741 DynamicAreaOffset); 742 } 743 744 IRB.CreateCall(AsanAllocasUnpoisonFunc, 745 {IRB.CreateLoad(DynamicAllocaLayout), DynamicAreaPtr}); 746 } 747 748 // Unpoison dynamic allocas redzones. 749 void unpoisonDynamicAllocas() { 750 for (auto &Ret : RetVec) 751 unpoisonDynamicAllocasBeforeInst(Ret, DynamicAllocaLayout); 752 753 for (auto &StackRestoreInst : StackRestoreVec) 754 unpoisonDynamicAllocasBeforeInst(StackRestoreInst, 755 StackRestoreInst->getOperand(0)); 756 } 757 758 // Deploy and poison redzones around dynamic alloca call. To do this, we 759 // should replace this call with another one with changed parameters and 760 // replace all its uses with new address, so 761 // addr = alloca type, old_size, align 762 // is replaced by 763 // new_size = (old_size + additional_size) * sizeof(type) 764 // tmp = alloca i8, new_size, max(align, 32) 765 // addr = tmp + 32 (first 32 bytes are for the left redzone). 766 // Additional_size is added to make new memory allocation contain not only 767 // requested memory, but also left, partial and right redzones. 768 void handleDynamicAllocaCall(AllocaInst *AI); 769 770 /// \brief Collect Alloca instructions we want (and can) handle. 771 void visitAllocaInst(AllocaInst &AI) { 772 if (!ASan.isInterestingAlloca(AI)) { 773 if (AI.isStaticAlloca()) { 774 // Skip over allocas that are present *before* the first instrumented 775 // alloca, we don't want to move those around. 776 if (AllocaVec.empty()) 777 return; 778 779 StaticAllocasToMoveUp.push_back(&AI); 780 } 781 return; 782 } 783 784 StackAlignment = std::max(StackAlignment, AI.getAlignment()); 785 if (!AI.isStaticAlloca()) 786 DynamicAllocaVec.push_back(&AI); 787 else 788 AllocaVec.push_back(&AI); 789 } 790 791 /// \brief Collect lifetime intrinsic calls to check for use-after-scope 792 /// errors. 793 void visitIntrinsicInst(IntrinsicInst &II) { 794 Intrinsic::ID ID = II.getIntrinsicID(); 795 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(&II); 796 if (ID == Intrinsic::localescape) LocalEscapeCall = &II; 797 if (!ASan.UseAfterScope) 798 return; 799 if (ID != Intrinsic::lifetime_start && ID != Intrinsic::lifetime_end) 800 return; 801 // Found lifetime intrinsic, add ASan instrumentation if necessary. 802 ConstantInt *Size = dyn_cast<ConstantInt>(II.getArgOperand(0)); 803 // If size argument is undefined, don't do anything. 804 if (Size->isMinusOne()) return; 805 // Check that size doesn't saturate uint64_t and can 806 // be stored in IntptrTy. 807 const uint64_t SizeValue = Size->getValue().getLimitedValue(); 808 if (SizeValue == ~0ULL || 809 !ConstantInt::isValueValidForType(IntptrTy, SizeValue)) 810 return; 811 // Find alloca instruction that corresponds to llvm.lifetime argument. 812 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1)); 813 if (!AI || !ASan.isInterestingAlloca(*AI)) 814 return; 815 bool DoPoison = (ID == Intrinsic::lifetime_end); 816 AllocaPoisonCall APC = {&II, AI, SizeValue, DoPoison}; 817 if (AI->isStaticAlloca()) 818 StaticAllocaPoisonCallVec.push_back(APC); 819 else if (ClInstrumentDynamicAllocas) 820 DynamicAllocaPoisonCallVec.push_back(APC); 821 } 822 823 void visitCallSite(CallSite CS) { 824 Instruction *I = CS.getInstruction(); 825 if (CallInst *CI = dyn_cast<CallInst>(I)) { 826 HasNonEmptyInlineAsm |= 827 CI->isInlineAsm() && !CI->isIdenticalTo(EmptyInlineAsm.get()); 828 HasReturnsTwiceCall |= CI->canReturnTwice(); 829 } 830 } 831 832 // ---------------------- Helpers. 833 void initializeCallbacks(Module &M); 834 835 bool doesDominateAllExits(const Instruction *I) const { 836 for (auto Ret : RetVec) { 837 if (!ASan.getDominatorTree().dominates(I, Ret)) return false; 838 } 839 return true; 840 } 841 842 /// Finds alloca where the value comes from. 843 AllocaInst *findAllocaForValue(Value *V); 844 845 // Copies bytes from ShadowBytes into shadow memory for indexes where 846 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that 847 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten. 848 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes, 849 IRBuilder<> &IRB, Value *ShadowBase); 850 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes, 851 size_t Begin, size_t End, IRBuilder<> &IRB, 852 Value *ShadowBase); 853 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask, 854 ArrayRef<uint8_t> ShadowBytes, size_t Begin, 855 size_t End, IRBuilder<> &IRB, Value *ShadowBase); 856 857 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison); 858 859 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L, 860 bool Dynamic); 861 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue, 862 Instruction *ThenTerm, Value *ValueIfFalse); 863 }; 864 865 } // anonymous namespace 866 867 char AddressSanitizer::ID = 0; 868 INITIALIZE_PASS_BEGIN( 869 AddressSanitizer, "asan", 870 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false, 871 false) 872 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 873 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 874 INITIALIZE_PASS_END( 875 AddressSanitizer, "asan", 876 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false, 877 false) 878 FunctionPass *llvm::createAddressSanitizerFunctionPass(bool CompileKernel, 879 bool Recover, 880 bool UseAfterScope) { 881 assert(!CompileKernel || Recover); 882 return new AddressSanitizer(CompileKernel, Recover, UseAfterScope); 883 } 884 885 char AddressSanitizerModule::ID = 0; 886 INITIALIZE_PASS( 887 AddressSanitizerModule, "asan-module", 888 "AddressSanitizer: detects use-after-free and out-of-bounds bugs." 889 "ModulePass", 890 false, false) 891 ModulePass *llvm::createAddressSanitizerModulePass(bool CompileKernel, 892 bool Recover) { 893 assert(!CompileKernel || Recover); 894 return new AddressSanitizerModule(CompileKernel, Recover); 895 } 896 897 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) { 898 size_t Res = countTrailingZeros(TypeSize / 8); 899 assert(Res < kNumberOfAccessSizes); 900 return Res; 901 } 902 903 // \brief Create a constant for Str so that we can pass it to the run-time lib. 904 static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str, 905 bool AllowMerging) { 906 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str); 907 // We use private linkage for module-local strings. If they can be merged 908 // with another one, we set the unnamed_addr attribute. 909 GlobalVariable *GV = 910 new GlobalVariable(M, StrConst->getType(), true, 911 GlobalValue::PrivateLinkage, StrConst, kAsanGenPrefix); 912 if (AllowMerging) GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); 913 GV->setAlignment(1); // Strings may not be merged w/o setting align 1. 914 return GV; 915 } 916 917 /// \brief Create a global describing a source location. 918 static GlobalVariable *createPrivateGlobalForSourceLoc(Module &M, 919 LocationMetadata MD) { 920 Constant *LocData[] = { 921 createPrivateGlobalForString(M, MD.Filename, true), 922 ConstantInt::get(Type::getInt32Ty(M.getContext()), MD.LineNo), 923 ConstantInt::get(Type::getInt32Ty(M.getContext()), MD.ColumnNo), 924 }; 925 auto LocStruct = ConstantStruct::getAnon(LocData); 926 auto GV = new GlobalVariable(M, LocStruct->getType(), true, 927 GlobalValue::PrivateLinkage, LocStruct, 928 kAsanGenPrefix); 929 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); 930 return GV; 931 } 932 933 /// \brief Check if \p G has been created by a trusted compiler pass. 934 static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) { 935 // Do not instrument asan globals. 936 if (G->getName().startswith(kAsanGenPrefix) || 937 G->getName().startswith(kSanCovGenPrefix) || 938 G->getName().startswith(kODRGenPrefix)) 939 return true; 940 941 // Do not instrument gcov counter arrays. 942 if (G->getName() == "__llvm_gcov_ctr") 943 return true; 944 945 return false; 946 } 947 948 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) { 949 // Shadow >> scale 950 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale); 951 if (Mapping.Offset == 0) return Shadow; 952 // (Shadow >> scale) | offset 953 Value *ShadowBase; 954 if (LocalDynamicShadow) 955 ShadowBase = LocalDynamicShadow; 956 else 957 ShadowBase = ConstantInt::get(IntptrTy, Mapping.Offset); 958 if (Mapping.OrShadowOffset) 959 return IRB.CreateOr(Shadow, ShadowBase); 960 else 961 return IRB.CreateAdd(Shadow, ShadowBase); 962 } 963 964 // Instrument memset/memmove/memcpy 965 void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) { 966 IRBuilder<> IRB(MI); 967 if (isa<MemTransferInst>(MI)) { 968 IRB.CreateCall( 969 isa<MemMoveInst>(MI) ? AsanMemmove : AsanMemcpy, 970 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 971 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()), 972 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 973 } else if (isa<MemSetInst>(MI)) { 974 IRB.CreateCall( 975 AsanMemset, 976 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 977 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false), 978 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 979 } 980 MI->eraseFromParent(); 981 } 982 983 /// Check if we want (and can) handle this alloca. 984 bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) { 985 auto PreviouslySeenAllocaInfo = ProcessedAllocas.find(&AI); 986 987 if (PreviouslySeenAllocaInfo != ProcessedAllocas.end()) 988 return PreviouslySeenAllocaInfo->getSecond(); 989 990 bool IsInteresting = 991 (AI.getAllocatedType()->isSized() && 992 // alloca() may be called with 0 size, ignore it. 993 ((!AI.isStaticAlloca()) || getAllocaSizeInBytes(AI) > 0) && 994 // We are only interested in allocas not promotable to registers. 995 // Promotable allocas are common under -O0. 996 (!ClSkipPromotableAllocas || !isAllocaPromotable(&AI)) && 997 // inalloca allocas are not treated as static, and we don't want 998 // dynamic alloca instrumentation for them as well. 999 !AI.isUsedWithInAlloca()); 1000 1001 ProcessedAllocas[&AI] = IsInteresting; 1002 return IsInteresting; 1003 } 1004 1005 Value *AddressSanitizer::isInterestingMemoryAccess(Instruction *I, 1006 bool *IsWrite, 1007 uint64_t *TypeSize, 1008 unsigned *Alignment, 1009 Value **MaybeMask) { 1010 // Skip memory accesses inserted by another instrumentation. 1011 if (I->getMetadata("nosanitize")) return nullptr; 1012 1013 // Do not instrument the load fetching the dynamic shadow address. 1014 if (LocalDynamicShadow == I) 1015 return nullptr; 1016 1017 Value *PtrOperand = nullptr; 1018 const DataLayout &DL = I->getModule()->getDataLayout(); 1019 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1020 if (!ClInstrumentReads) return nullptr; 1021 *IsWrite = false; 1022 *TypeSize = DL.getTypeStoreSizeInBits(LI->getType()); 1023 *Alignment = LI->getAlignment(); 1024 PtrOperand = LI->getPointerOperand(); 1025 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1026 if (!ClInstrumentWrites) return nullptr; 1027 *IsWrite = true; 1028 *TypeSize = DL.getTypeStoreSizeInBits(SI->getValueOperand()->getType()); 1029 *Alignment = SI->getAlignment(); 1030 PtrOperand = SI->getPointerOperand(); 1031 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { 1032 if (!ClInstrumentAtomics) return nullptr; 1033 *IsWrite = true; 1034 *TypeSize = DL.getTypeStoreSizeInBits(RMW->getValOperand()->getType()); 1035 *Alignment = 0; 1036 PtrOperand = RMW->getPointerOperand(); 1037 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) { 1038 if (!ClInstrumentAtomics) return nullptr; 1039 *IsWrite = true; 1040 *TypeSize = DL.getTypeStoreSizeInBits(XCHG->getCompareOperand()->getType()); 1041 *Alignment = 0; 1042 PtrOperand = XCHG->getPointerOperand(); 1043 } else if (auto CI = dyn_cast<CallInst>(I)) { 1044 auto *F = dyn_cast<Function>(CI->getCalledValue()); 1045 if (F && (F->getName().startswith("llvm.masked.load.") || 1046 F->getName().startswith("llvm.masked.store."))) { 1047 unsigned OpOffset = 0; 1048 if (F->getName().startswith("llvm.masked.store.")) { 1049 // Masked store has an initial operand for the value. 1050 OpOffset = 1; 1051 *IsWrite = true; 1052 } else { 1053 *IsWrite = false; 1054 } 1055 // Only instrument if the mask is constant for now. 1056 if (isa<ConstantVector>(CI->getOperand(2 + OpOffset))) { 1057 auto BasePtr = CI->getOperand(0 + OpOffset); 1058 auto Ty = cast<PointerType>(BasePtr->getType())->getElementType(); 1059 *TypeSize = DL.getTypeStoreSizeInBits(Ty); 1060 if (auto AlignmentConstant = 1061 dyn_cast<ConstantInt>(CI->getOperand(1 + OpOffset))) 1062 *Alignment = (unsigned)AlignmentConstant->getZExtValue(); 1063 else 1064 *Alignment = 1; // No alignment guarantees. We probably got Undef 1065 if (MaybeMask) 1066 *MaybeMask = CI->getOperand(2 + OpOffset); 1067 PtrOperand = BasePtr; 1068 } 1069 } 1070 } 1071 1072 // Do not instrument acesses from different address spaces; we cannot deal 1073 // with them. 1074 if (PtrOperand) { 1075 Type *PtrTy = cast<PointerType>(PtrOperand->getType()->getScalarType()); 1076 if (PtrTy->getPointerAddressSpace() != 0) 1077 return nullptr; 1078 } 1079 1080 // Treat memory accesses to promotable allocas as non-interesting since they 1081 // will not cause memory violations. This greatly speeds up the instrumented 1082 // executable at -O0. 1083 if (ClSkipPromotableAllocas) 1084 if (auto AI = dyn_cast_or_null<AllocaInst>(PtrOperand)) 1085 return isInterestingAlloca(*AI) ? AI : nullptr; 1086 1087 return PtrOperand; 1088 } 1089 1090 static bool isPointerOperand(Value *V) { 1091 return V->getType()->isPointerTy() || isa<PtrToIntInst>(V); 1092 } 1093 1094 // This is a rough heuristic; it may cause both false positives and 1095 // false negatives. The proper implementation requires cooperation with 1096 // the frontend. 1097 static bool isInterestingPointerComparisonOrSubtraction(Instruction *I) { 1098 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(I)) { 1099 if (!Cmp->isRelational()) return false; 1100 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 1101 if (BO->getOpcode() != Instruction::Sub) return false; 1102 } else { 1103 return false; 1104 } 1105 return isPointerOperand(I->getOperand(0)) && 1106 isPointerOperand(I->getOperand(1)); 1107 } 1108 1109 bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) { 1110 // If a global variable does not have dynamic initialization we don't 1111 // have to instrument it. However, if a global does not have initializer 1112 // at all, we assume it has dynamic initializer (in other TU). 1113 return G->hasInitializer() && !GlobalsMD.get(G).IsDynInit; 1114 } 1115 1116 void AddressSanitizer::instrumentPointerComparisonOrSubtraction( 1117 Instruction *I) { 1118 IRBuilder<> IRB(I); 1119 Function *F = isa<ICmpInst>(I) ? AsanPtrCmpFunction : AsanPtrSubFunction; 1120 Value *Param[2] = {I->getOperand(0), I->getOperand(1)}; 1121 for (Value *&i : Param) { 1122 if (i->getType()->isPointerTy()) 1123 i = IRB.CreatePointerCast(i, IntptrTy); 1124 } 1125 IRB.CreateCall(F, Param); 1126 } 1127 1128 static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I, 1129 Value *Addr, unsigned Alignment, 1130 unsigned Granularity, uint32_t TypeSize, 1131 bool IsWrite, Value *SizeArgument, 1132 bool UseCalls, uint32_t Exp) { 1133 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check 1134 // if the data is properly aligned. 1135 if ((TypeSize == 8 || TypeSize == 16 || TypeSize == 32 || TypeSize == 64 || 1136 TypeSize == 128) && 1137 (Alignment >= Granularity || Alignment == 0 || Alignment >= TypeSize / 8)) 1138 return Pass->instrumentAddress(I, I, Addr, TypeSize, IsWrite, nullptr, 1139 UseCalls, Exp); 1140 Pass->instrumentUnusualSizeOrAlignment(I, Addr, TypeSize, IsWrite, nullptr, 1141 UseCalls, Exp); 1142 } 1143 1144 static void instrumentMaskedLoadOrStore(AddressSanitizer *Pass, 1145 const DataLayout &DL, Type *IntptrTy, 1146 ConstantVector *Mask, Instruction *I, 1147 Value *Addr, unsigned Alignment, 1148 unsigned Granularity, uint32_t TypeSize, 1149 bool IsWrite, Value *SizeArgument, 1150 bool UseCalls, uint32_t Exp) { 1151 auto *VTy = cast<PointerType>(Addr->getType())->getElementType(); 1152 uint64_t ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType()); 1153 unsigned Num = VTy->getVectorNumElements(); 1154 auto Zero = ConstantInt::get(IntptrTy, 0); 1155 for (unsigned Idx = 0; Idx < Num; ++Idx) { 1156 // dyn_cast as we might get UndefValue 1157 auto Masked = dyn_cast<ConstantInt>(Mask->getOperand(Idx)); 1158 if (Masked && Masked->isAllOnesValue()) { 1159 IRBuilder<> IRB(I); 1160 auto InstrumentedAddress = 1161 IRB.CreateGEP(Addr, {Zero, ConstantInt::get(IntptrTy, Idx)}); 1162 doInstrumentAddress(Pass, I, InstrumentedAddress, Alignment, Granularity, 1163 ElemTypeSize, IsWrite, SizeArgument, UseCalls, Exp); 1164 } 1165 } 1166 } 1167 1168 void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis, 1169 Instruction *I, bool UseCalls, 1170 const DataLayout &DL) { 1171 bool IsWrite = false; 1172 unsigned Alignment = 0; 1173 uint64_t TypeSize = 0; 1174 Value *MaybeMask = nullptr; 1175 Value *Addr = 1176 isInterestingMemoryAccess(I, &IsWrite, &TypeSize, &Alignment, &MaybeMask); 1177 assert(Addr); 1178 1179 // Optimization experiments. 1180 // The experiments can be used to evaluate potential optimizations that remove 1181 // instrumentation (assess false negatives). Instead of completely removing 1182 // some instrumentation, you set Exp to a non-zero value (mask of optimization 1183 // experiments that want to remove instrumentation of this instruction). 1184 // If Exp is non-zero, this pass will emit special calls into runtime 1185 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls 1186 // make runtime terminate the program in a special way (with a different 1187 // exit status). Then you run the new compiler on a buggy corpus, collect 1188 // the special terminations (ideally, you don't see them at all -- no false 1189 // negatives) and make the decision on the optimization. 1190 uint32_t Exp = ClForceExperiment; 1191 1192 if (ClOpt && ClOptGlobals) { 1193 // If initialization order checking is disabled, a simple access to a 1194 // dynamically initialized global is always valid. 1195 GlobalVariable *G = dyn_cast<GlobalVariable>(GetUnderlyingObject(Addr, DL)); 1196 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) && 1197 isSafeAccess(ObjSizeVis, Addr, TypeSize)) { 1198 NumOptimizedAccessesToGlobalVar++; 1199 return; 1200 } 1201 } 1202 1203 if (ClOpt && ClOptStack) { 1204 // A direct inbounds access to a stack variable is always valid. 1205 if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) && 1206 isSafeAccess(ObjSizeVis, Addr, TypeSize)) { 1207 NumOptimizedAccessesToStackVar++; 1208 return; 1209 } 1210 } 1211 1212 if (IsWrite) 1213 NumInstrumentedWrites++; 1214 else 1215 NumInstrumentedReads++; 1216 1217 unsigned Granularity = 1 << Mapping.Scale; 1218 if (MaybeMask) { 1219 auto Mask = cast<ConstantVector>(MaybeMask); 1220 instrumentMaskedLoadOrStore(this, DL, IntptrTy, Mask, I, Addr, Alignment, 1221 Granularity, TypeSize, IsWrite, nullptr, 1222 UseCalls, Exp); 1223 } else { 1224 doInstrumentAddress(this, I, Addr, Alignment, Granularity, TypeSize, 1225 IsWrite, nullptr, UseCalls, Exp); 1226 } 1227 } 1228 1229 Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore, 1230 Value *Addr, bool IsWrite, 1231 size_t AccessSizeIndex, 1232 Value *SizeArgument, 1233 uint32_t Exp) { 1234 IRBuilder<> IRB(InsertBefore); 1235 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(IRB.getInt32Ty(), Exp); 1236 CallInst *Call = nullptr; 1237 if (SizeArgument) { 1238 if (Exp == 0) 1239 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][0], 1240 {Addr, SizeArgument}); 1241 else 1242 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][1], 1243 {Addr, SizeArgument, ExpVal}); 1244 } else { 1245 if (Exp == 0) 1246 Call = 1247 IRB.CreateCall(AsanErrorCallback[IsWrite][0][AccessSizeIndex], Addr); 1248 else 1249 Call = IRB.CreateCall(AsanErrorCallback[IsWrite][1][AccessSizeIndex], 1250 {Addr, ExpVal}); 1251 } 1252 1253 // We don't do Call->setDoesNotReturn() because the BB already has 1254 // UnreachableInst at the end. 1255 // This EmptyAsm is required to avoid callback merge. 1256 IRB.CreateCall(EmptyAsm, {}); 1257 return Call; 1258 } 1259 1260 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong, 1261 Value *ShadowValue, 1262 uint32_t TypeSize) { 1263 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale; 1264 // Addr & (Granularity - 1) 1265 Value *LastAccessedByte = 1266 IRB.CreateAnd(AddrLong, ConstantInt::get(IntptrTy, Granularity - 1)); 1267 // (Addr & (Granularity - 1)) + size - 1 1268 if (TypeSize / 8 > 1) 1269 LastAccessedByte = IRB.CreateAdd( 1270 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1)); 1271 // (uint8_t) ((Addr & (Granularity-1)) + size - 1) 1272 LastAccessedByte = 1273 IRB.CreateIntCast(LastAccessedByte, ShadowValue->getType(), false); 1274 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue 1275 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue); 1276 } 1277 1278 void AddressSanitizer::instrumentAddress(Instruction *OrigIns, 1279 Instruction *InsertBefore, Value *Addr, 1280 uint32_t TypeSize, bool IsWrite, 1281 Value *SizeArgument, bool UseCalls, 1282 uint32_t Exp) { 1283 IRBuilder<> IRB(InsertBefore); 1284 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); 1285 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize); 1286 1287 if (UseCalls) { 1288 if (Exp == 0) 1289 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex], 1290 AddrLong); 1291 else 1292 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex], 1293 {AddrLong, ConstantInt::get(IRB.getInt32Ty(), Exp)}); 1294 return; 1295 } 1296 1297 Type *ShadowTy = 1298 IntegerType::get(*C, std::max(8U, TypeSize >> Mapping.Scale)); 1299 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0); 1300 Value *ShadowPtr = memToShadow(AddrLong, IRB); 1301 Value *CmpVal = Constant::getNullValue(ShadowTy); 1302 Value *ShadowValue = 1303 IRB.CreateLoad(IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy)); 1304 1305 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal); 1306 size_t Granularity = 1ULL << Mapping.Scale; 1307 TerminatorInst *CrashTerm = nullptr; 1308 1309 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) { 1310 // We use branch weights for the slow path check, to indicate that the slow 1311 // path is rarely taken. This seems to be the case for SPEC benchmarks. 1312 TerminatorInst *CheckTerm = SplitBlockAndInsertIfThen( 1313 Cmp, InsertBefore, false, MDBuilder(*C).createBranchWeights(1, 100000)); 1314 assert(cast<BranchInst>(CheckTerm)->isUnconditional()); 1315 BasicBlock *NextBB = CheckTerm->getSuccessor(0); 1316 IRB.SetInsertPoint(CheckTerm); 1317 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize); 1318 if (Recover) { 1319 CrashTerm = SplitBlockAndInsertIfThen(Cmp2, CheckTerm, false); 1320 } else { 1321 BasicBlock *CrashBlock = 1322 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB); 1323 CrashTerm = new UnreachableInst(*C, CrashBlock); 1324 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2); 1325 ReplaceInstWithInst(CheckTerm, NewTerm); 1326 } 1327 } else { 1328 CrashTerm = SplitBlockAndInsertIfThen(Cmp, InsertBefore, !Recover); 1329 } 1330 1331 Instruction *Crash = generateCrashCode(CrashTerm, AddrLong, IsWrite, 1332 AccessSizeIndex, SizeArgument, Exp); 1333 Crash->setDebugLoc(OrigIns->getDebugLoc()); 1334 } 1335 1336 // Instrument unusual size or unusual alignment. 1337 // We can not do it with a single check, so we do 1-byte check for the first 1338 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able 1339 // to report the actual access size. 1340 void AddressSanitizer::instrumentUnusualSizeOrAlignment( 1341 Instruction *I, Value *Addr, uint32_t TypeSize, bool IsWrite, 1342 Value *SizeArgument, bool UseCalls, uint32_t Exp) { 1343 IRBuilder<> IRB(I); 1344 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8); 1345 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); 1346 if (UseCalls) { 1347 if (Exp == 0) 1348 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][0], 1349 {AddrLong, Size}); 1350 else 1351 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][1], 1352 {AddrLong, Size, ConstantInt::get(IRB.getInt32Ty(), Exp)}); 1353 } else { 1354 Value *LastByte = IRB.CreateIntToPtr( 1355 IRB.CreateAdd(AddrLong, ConstantInt::get(IntptrTy, TypeSize / 8 - 1)), 1356 Addr->getType()); 1357 instrumentAddress(I, I, Addr, 8, IsWrite, Size, false, Exp); 1358 instrumentAddress(I, I, LastByte, 8, IsWrite, Size, false, Exp); 1359 } 1360 } 1361 1362 void AddressSanitizerModule::poisonOneInitializer(Function &GlobalInit, 1363 GlobalValue *ModuleName) { 1364 // Set up the arguments to our poison/unpoison functions. 1365 IRBuilder<> IRB(&GlobalInit.front(), 1366 GlobalInit.front().getFirstInsertionPt()); 1367 1368 // Add a call to poison all external globals before the given function starts. 1369 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy); 1370 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr); 1371 1372 // Add calls to unpoison all globals before each return instruction. 1373 for (auto &BB : GlobalInit.getBasicBlockList()) 1374 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB.getTerminator())) 1375 CallInst::Create(AsanUnpoisonGlobals, "", RI); 1376 } 1377 1378 void AddressSanitizerModule::createInitializerPoisonCalls( 1379 Module &M, GlobalValue *ModuleName) { 1380 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors"); 1381 1382 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer()); 1383 for (Use &OP : CA->operands()) { 1384 if (isa<ConstantAggregateZero>(OP)) continue; 1385 ConstantStruct *CS = cast<ConstantStruct>(OP); 1386 1387 // Must have a function or null ptr. 1388 if (Function *F = dyn_cast<Function>(CS->getOperand(1))) { 1389 if (F->getName() == kAsanModuleCtorName) continue; 1390 ConstantInt *Priority = dyn_cast<ConstantInt>(CS->getOperand(0)); 1391 // Don't instrument CTORs that will run before asan.module_ctor. 1392 if (Priority->getLimitedValue() <= kAsanCtorAndDtorPriority) continue; 1393 poisonOneInitializer(*F, ModuleName); 1394 } 1395 } 1396 } 1397 1398 bool AddressSanitizerModule::ShouldInstrumentGlobal(GlobalVariable *G) { 1399 Type *Ty = G->getValueType(); 1400 DEBUG(dbgs() << "GLOBAL: " << *G << "\n"); 1401 1402 if (GlobalsMD.get(G).IsBlacklisted) return false; 1403 if (!Ty->isSized()) return false; 1404 if (!G->hasInitializer()) return false; 1405 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals. 1406 // Touch only those globals that will not be defined in other modules. 1407 // Don't handle ODR linkage types and COMDATs since other modules may be built 1408 // without ASan. 1409 if (G->getLinkage() != GlobalVariable::ExternalLinkage && 1410 G->getLinkage() != GlobalVariable::PrivateLinkage && 1411 G->getLinkage() != GlobalVariable::InternalLinkage) 1412 return false; 1413 if (G->hasComdat()) return false; 1414 // Two problems with thread-locals: 1415 // - The address of the main thread's copy can't be computed at link-time. 1416 // - Need to poison all copies, not just the main thread's one. 1417 if (G->isThreadLocal()) return false; 1418 // For now, just ignore this Global if the alignment is large. 1419 if (G->getAlignment() > MinRedzoneSizeForGlobal()) return false; 1420 1421 if (G->hasSection()) { 1422 StringRef Section = G->getSection(); 1423 1424 // Globals from llvm.metadata aren't emitted, do not instrument them. 1425 if (Section == "llvm.metadata") return false; 1426 // Do not instrument globals from special LLVM sections. 1427 if (Section.find("__llvm") != StringRef::npos || Section.find("__LLVM") != StringRef::npos) return false; 1428 1429 // Do not instrument function pointers to initialization and termination 1430 // routines: dynamic linker will not properly handle redzones. 1431 if (Section.startswith(".preinit_array") || 1432 Section.startswith(".init_array") || 1433 Section.startswith(".fini_array")) { 1434 return false; 1435 } 1436 1437 // Callbacks put into the CRT initializer/terminator sections 1438 // should not be instrumented. 1439 // See https://code.google.com/p/address-sanitizer/issues/detail?id=305 1440 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx 1441 if (Section.startswith(".CRT")) { 1442 DEBUG(dbgs() << "Ignoring a global initializer callback: " << *G << "\n"); 1443 return false; 1444 } 1445 1446 if (TargetTriple.isOSBinFormatMachO()) { 1447 StringRef ParsedSegment, ParsedSection; 1448 unsigned TAA = 0, StubSize = 0; 1449 bool TAAParsed; 1450 std::string ErrorCode = MCSectionMachO::ParseSectionSpecifier( 1451 Section, ParsedSegment, ParsedSection, TAA, TAAParsed, StubSize); 1452 assert(ErrorCode.empty() && "Invalid section specifier."); 1453 1454 // Ignore the globals from the __OBJC section. The ObjC runtime assumes 1455 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to 1456 // them. 1457 if (ParsedSegment == "__OBJC" || 1458 (ParsedSegment == "__DATA" && ParsedSection.startswith("__objc_"))) { 1459 DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n"); 1460 return false; 1461 } 1462 // See http://code.google.com/p/address-sanitizer/issues/detail?id=32 1463 // Constant CFString instances are compiled in the following way: 1464 // -- the string buffer is emitted into 1465 // __TEXT,__cstring,cstring_literals 1466 // -- the constant NSConstantString structure referencing that buffer 1467 // is placed into __DATA,__cfstring 1468 // Therefore there's no point in placing redzones into __DATA,__cfstring. 1469 // Moreover, it causes the linker to crash on OS X 10.7 1470 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") { 1471 DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n"); 1472 return false; 1473 } 1474 // The linker merges the contents of cstring_literals and removes the 1475 // trailing zeroes. 1476 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) { 1477 DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n"); 1478 return false; 1479 } 1480 } 1481 } 1482 1483 return true; 1484 } 1485 1486 // On Mach-O platforms, we emit global metadata in a separate section of the 1487 // binary in order to allow the linker to properly dead strip. This is only 1488 // supported on recent versions of ld64. 1489 bool AddressSanitizerModule::ShouldUseMachOGlobalsSection() const { 1490 if (!ClUseMachOGlobalsSection) 1491 return false; 1492 1493 if (!TargetTriple.isOSBinFormatMachO()) 1494 return false; 1495 1496 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(10, 11)) 1497 return true; 1498 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(9)) 1499 return true; 1500 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(2)) 1501 return true; 1502 1503 return false; 1504 } 1505 1506 StringRef AddressSanitizerModule::getGlobalMetadataSection() const { 1507 switch (TargetTriple.getObjectFormat()) { 1508 case Triple::COFF: return ".ASAN$GL"; 1509 case Triple::ELF: return "asan_globals"; 1510 case Triple::MachO: return "__DATA,__asan_globals,regular"; 1511 default: break; 1512 } 1513 llvm_unreachable("unsupported object format"); 1514 } 1515 1516 void AddressSanitizerModule::initializeCallbacks(Module &M) { 1517 IRBuilder<> IRB(*C); 1518 1519 // Declare our poisoning and unpoisoning functions. 1520 AsanPoisonGlobals = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1521 kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy, nullptr)); 1522 AsanPoisonGlobals->setLinkage(Function::ExternalLinkage); 1523 AsanUnpoisonGlobals = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1524 kAsanUnpoisonGlobalsName, IRB.getVoidTy(), nullptr)); 1525 AsanUnpoisonGlobals->setLinkage(Function::ExternalLinkage); 1526 1527 // Declare functions that register/unregister globals. 1528 AsanRegisterGlobals = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1529 kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 1530 AsanRegisterGlobals->setLinkage(Function::ExternalLinkage); 1531 AsanUnregisterGlobals = checkSanitizerInterfaceFunction( 1532 M.getOrInsertFunction(kAsanUnregisterGlobalsName, IRB.getVoidTy(), 1533 IntptrTy, IntptrTy, nullptr)); 1534 AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage); 1535 1536 // Declare the functions that find globals in a shared object and then invoke 1537 // the (un)register function on them. 1538 AsanRegisterImageGlobals = checkSanitizerInterfaceFunction( 1539 M.getOrInsertFunction(kAsanRegisterImageGlobalsName, 1540 IRB.getVoidTy(), IntptrTy, nullptr)); 1541 AsanRegisterImageGlobals->setLinkage(Function::ExternalLinkage); 1542 1543 AsanUnregisterImageGlobals = checkSanitizerInterfaceFunction( 1544 M.getOrInsertFunction(kAsanUnregisterImageGlobalsName, 1545 IRB.getVoidTy(), IntptrTy, nullptr)); 1546 AsanUnregisterImageGlobals->setLinkage(Function::ExternalLinkage); 1547 } 1548 1549 // This function replaces all global variables with new variables that have 1550 // trailing redzones. It also creates a function that poisons 1551 // redzones and inserts this function into llvm.global_ctors. 1552 bool AddressSanitizerModule::InstrumentGlobals(IRBuilder<> &IRB, Module &M) { 1553 GlobalsMD.init(M); 1554 1555 SmallVector<GlobalVariable *, 16> GlobalsToChange; 1556 1557 for (auto &G : M.globals()) { 1558 if (ShouldInstrumentGlobal(&G)) GlobalsToChange.push_back(&G); 1559 } 1560 1561 size_t n = GlobalsToChange.size(); 1562 if (n == 0) return false; 1563 1564 auto &DL = M.getDataLayout(); 1565 bool UseComdatMetadata = TargetTriple.isOSBinFormatCOFF(); 1566 bool UseMachOGlobalsSection = ShouldUseMachOGlobalsSection(); 1567 bool UseMetadataArray = !(UseComdatMetadata || UseMachOGlobalsSection); 1568 1569 // A global is described by a structure 1570 // size_t beg; 1571 // size_t size; 1572 // size_t size_with_redzone; 1573 // const char *name; 1574 // const char *module_name; 1575 // size_t has_dynamic_init; 1576 // void *source_location; 1577 // size_t odr_indicator; 1578 // We initialize an array of such structures and pass it to a run-time call. 1579 StructType *GlobalStructTy = 1580 StructType::get(IntptrTy, IntptrTy, IntptrTy, IntptrTy, IntptrTy, 1581 IntptrTy, IntptrTy, IntptrTy, nullptr); 1582 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(GlobalStructTy); 1583 assert((isPowerOf2_32(SizeOfGlobalStruct) || 1584 !TargetTriple.isOSBinFormatCOFF()) && 1585 "global metadata will not be padded appropriately"); 1586 SmallVector<Constant *, 16> Initializers(UseMetadataArray ? n : 0); 1587 1588 // On recent Mach-O platforms, use a structure which binds the liveness of 1589 // the global variable to the metadata struct. Keep the list of "Liveness" GV 1590 // created to be added to llvm.compiler.used 1591 StructType *LivenessTy = nullptr; 1592 if (UseMachOGlobalsSection) 1593 LivenessTy = StructType::get(IntptrTy, IntptrTy, nullptr); 1594 SmallVector<GlobalValue *, 16> LivenessGlobals( 1595 UseMachOGlobalsSection ? n : 0); 1596 1597 bool HasDynamicallyInitializedGlobals = false; 1598 1599 // We shouldn't merge same module names, as this string serves as unique 1600 // module ID in runtime. 1601 GlobalVariable *ModuleName = createPrivateGlobalForString( 1602 M, M.getModuleIdentifier(), /*AllowMerging*/ false); 1603 1604 for (size_t i = 0; i < n; i++) { 1605 static const uint64_t kMaxGlobalRedzone = 1 << 18; 1606 GlobalVariable *G = GlobalsToChange[i]; 1607 1608 auto MD = GlobalsMD.get(G); 1609 StringRef NameForGlobal = G->getName(); 1610 // Create string holding the global name (use global name from metadata 1611 // if it's available, otherwise just write the name of global variable). 1612 GlobalVariable *Name = createPrivateGlobalForString( 1613 M, MD.Name.empty() ? NameForGlobal : MD.Name, 1614 /*AllowMerging*/ true); 1615 1616 Type *Ty = G->getValueType(); 1617 uint64_t SizeInBytes = DL.getTypeAllocSize(Ty); 1618 uint64_t MinRZ = MinRedzoneSizeForGlobal(); 1619 // MinRZ <= RZ <= kMaxGlobalRedzone 1620 // and trying to make RZ to be ~ 1/4 of SizeInBytes. 1621 uint64_t RZ = std::max( 1622 MinRZ, std::min(kMaxGlobalRedzone, (SizeInBytes / MinRZ / 4) * MinRZ)); 1623 uint64_t RightRedzoneSize = RZ; 1624 // Round up to MinRZ 1625 if (SizeInBytes % MinRZ) RightRedzoneSize += MinRZ - (SizeInBytes % MinRZ); 1626 assert(((RightRedzoneSize + SizeInBytes) % MinRZ) == 0); 1627 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize); 1628 1629 StructType *NewTy = StructType::get(Ty, RightRedZoneTy, nullptr); 1630 Constant *NewInitializer = 1631 ConstantStruct::get(NewTy, G->getInitializer(), 1632 Constant::getNullValue(RightRedZoneTy), nullptr); 1633 1634 // Create a new global variable with enough space for a redzone. 1635 GlobalValue::LinkageTypes Linkage = G->getLinkage(); 1636 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage) 1637 Linkage = GlobalValue::InternalLinkage; 1638 GlobalVariable *NewGlobal = 1639 new GlobalVariable(M, NewTy, G->isConstant(), Linkage, NewInitializer, 1640 "", G, G->getThreadLocalMode()); 1641 NewGlobal->copyAttributesFrom(G); 1642 NewGlobal->setAlignment(MinRZ); 1643 1644 // Move null-terminated C strings to "__asan_cstring" section on Darwin. 1645 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() && 1646 G->isConstant()) { 1647 auto Seq = dyn_cast<ConstantDataSequential>(G->getInitializer()); 1648 if (Seq && Seq->isCString()) 1649 NewGlobal->setSection("__TEXT,__asan_cstring,regular"); 1650 } 1651 1652 // Transfer the debug info. The payload starts at offset zero so we can 1653 // copy the debug info over as is. 1654 SmallVector<DIGlobalVariable *, 1> GVs; 1655 G->getDebugInfo(GVs); 1656 for (auto *GV : GVs) 1657 NewGlobal->addDebugInfo(GV); 1658 1659 Value *Indices2[2]; 1660 Indices2[0] = IRB.getInt32(0); 1661 Indices2[1] = IRB.getInt32(0); 1662 1663 G->replaceAllUsesWith( 1664 ConstantExpr::getGetElementPtr(NewTy, NewGlobal, Indices2, true)); 1665 NewGlobal->takeName(G); 1666 G->eraseFromParent(); 1667 G = NewGlobal; 1668 1669 if (UseComdatMetadata) { 1670 // Get or create a COMDAT for G so that we can use it with our metadata. 1671 Comdat *C = G->getComdat(); 1672 if (!C) { 1673 if (!G->hasName()) { 1674 // If G is unnamed, it must be internal. Give it an artificial name 1675 // so we can put it in a comdat. 1676 assert(G->hasLocalLinkage()); 1677 G->setName(Twine(kAsanGenPrefix) + "_anon_global"); 1678 } 1679 C = M.getOrInsertComdat(G->getName()); 1680 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. 1681 if (TargetTriple.isOSBinFormatCOFF()) 1682 C->setSelectionKind(Comdat::NoDuplicates); 1683 G->setComdat(C); 1684 } 1685 } 1686 1687 Constant *SourceLoc; 1688 if (!MD.SourceLoc.empty()) { 1689 auto SourceLocGlobal = createPrivateGlobalForSourceLoc(M, MD.SourceLoc); 1690 SourceLoc = ConstantExpr::getPointerCast(SourceLocGlobal, IntptrTy); 1691 } else { 1692 SourceLoc = ConstantInt::get(IntptrTy, 0); 1693 } 1694 1695 Constant *ODRIndicator = ConstantExpr::getNullValue(IRB.getInt8PtrTy()); 1696 GlobalValue *InstrumentedGlobal = NewGlobal; 1697 1698 bool CanUsePrivateAliases = 1699 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO(); 1700 if (CanUsePrivateAliases && ClUsePrivateAliasForGlobals) { 1701 // Create local alias for NewGlobal to avoid crash on ODR between 1702 // instrumented and non-instrumented libraries. 1703 auto *GA = GlobalAlias::create(GlobalValue::InternalLinkage, 1704 NameForGlobal + M.getName(), NewGlobal); 1705 1706 // With local aliases, we need to provide another externally visible 1707 // symbol __odr_asan_XXX to detect ODR violation. 1708 auto *ODRIndicatorSym = 1709 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage, 1710 Constant::getNullValue(IRB.getInt8Ty()), 1711 kODRGenPrefix + NameForGlobal, nullptr, 1712 NewGlobal->getThreadLocalMode()); 1713 1714 // Set meaningful attributes for indicator symbol. 1715 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility()); 1716 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass()); 1717 ODRIndicatorSym->setAlignment(1); 1718 ODRIndicator = ODRIndicatorSym; 1719 InstrumentedGlobal = GA; 1720 } 1721 1722 Constant *Initializer = ConstantStruct::get( 1723 GlobalStructTy, 1724 ConstantExpr::getPointerCast(InstrumentedGlobal, IntptrTy), 1725 ConstantInt::get(IntptrTy, SizeInBytes), 1726 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize), 1727 ConstantExpr::getPointerCast(Name, IntptrTy), 1728 ConstantExpr::getPointerCast(ModuleName, IntptrTy), 1729 ConstantInt::get(IntptrTy, MD.IsDynInit), SourceLoc, 1730 ConstantExpr::getPointerCast(ODRIndicator, IntptrTy), nullptr); 1731 1732 if (ClInitializers && MD.IsDynInit) HasDynamicallyInitializedGlobals = true; 1733 1734 DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n"); 1735 1736 // If we aren't using separate metadata globals, add it to the initializer 1737 // list and continue. 1738 if (UseMetadataArray) { 1739 Initializers[i] = Initializer; 1740 continue; 1741 } 1742 1743 // Create a separate metadata global and put it in the appropriate ASan 1744 // global registration section. 1745 GlobalVariable *Metadata = new GlobalVariable( 1746 M, GlobalStructTy, false, GlobalVariable::InternalLinkage, 1747 Initializer, Twine("__asan_global_") + 1748 GlobalValue::getRealLinkageName(G->getName())); 1749 Metadata->setSection(getGlobalMetadataSection()); 1750 1751 // The MSVC linker always inserts padding when linking incrementally. We 1752 // cope with that by aligning each struct to its size, which must be a power 1753 // of two. 1754 if (TargetTriple.isOSBinFormatCOFF()) 1755 Metadata->setAlignment(SizeOfGlobalStruct); 1756 else 1757 Metadata->setAlignment(1); // Don't leave padding in between. 1758 1759 // On platforms that support comdats, put the metadata and the 1760 // instrumented global in the same group. This ensures that the metadata 1761 // is discarded if the instrumented global is discarded. 1762 if (UseComdatMetadata) { 1763 assert(G->hasComdat()); 1764 Metadata->setComdat(G->getComdat()); 1765 continue; 1766 } 1767 assert(UseMachOGlobalsSection); 1768 1769 // On recent Mach-O platforms, we emit the global metadata in a way that 1770 // allows the linker to properly strip dead globals. 1771 auto LivenessBinder = ConstantStruct::get( 1772 LivenessTy, Initializer->getAggregateElement(0u), 1773 ConstantExpr::getPointerCast(Metadata, IntptrTy), nullptr); 1774 GlobalVariable *Liveness = new GlobalVariable( 1775 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder, 1776 Twine("__asan_binder_") + G->getName()); 1777 Liveness->setSection("__DATA,__asan_liveness,regular,live_support"); 1778 LivenessGlobals[i] = Liveness; 1779 } 1780 1781 // Create calls for poisoning before initializers run and unpoisoning after. 1782 if (HasDynamicallyInitializedGlobals) 1783 createInitializerPoisonCalls(M, ModuleName); 1784 1785 // Platforms with a dedicated metadata section don't need to emit any more 1786 // code. 1787 if (UseComdatMetadata) 1788 return true; 1789 1790 GlobalVariable *AllGlobals = nullptr; 1791 GlobalVariable *RegisteredFlag = nullptr; 1792 1793 if (UseMachOGlobalsSection) { 1794 // RegisteredFlag serves two purposes. First, we can pass it to dladdr() 1795 // to look up the loaded image that contains it. Second, we can store in it 1796 // whether registration has already occurred, to prevent duplicate 1797 // registration. 1798 // 1799 // common linkage ensures that there is only one global per shared library. 1800 RegisteredFlag = new GlobalVariable( 1801 M, IntptrTy, false, GlobalVariable::CommonLinkage, 1802 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName); 1803 1804 // Update llvm.compiler.used, adding the new liveness globals. This is 1805 // needed so that during LTO these variables stay alive. The alternative 1806 // would be to have the linker handling the LTO symbols, but libLTO 1807 // current API does not expose access to the section for each symbol. 1808 if (!LivenessGlobals.empty()) 1809 appendToCompilerUsed(M, LivenessGlobals); 1810 } else if (UseMetadataArray) { 1811 // On platforms that don't have a custom metadata section, we emit an array 1812 // of global metadata structures. 1813 ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n); 1814 AllGlobals = new GlobalVariable( 1815 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage, 1816 ConstantArray::get(ArrayOfGlobalStructTy, Initializers), ""); 1817 } 1818 1819 // Create a call to register the globals with the runtime. 1820 if (UseMachOGlobalsSection) { 1821 IRB.CreateCall(AsanRegisterImageGlobals, 1822 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)}); 1823 } else { 1824 IRB.CreateCall(AsanRegisterGlobals, 1825 {IRB.CreatePointerCast(AllGlobals, IntptrTy), 1826 ConstantInt::get(IntptrTy, n)}); 1827 } 1828 1829 // We also need to unregister globals at the end, e.g., when a shared library 1830 // gets closed. 1831 Function *AsanDtorFunction = 1832 Function::Create(FunctionType::get(Type::getVoidTy(*C), false), 1833 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M); 1834 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction); 1835 IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB)); 1836 1837 if (UseMachOGlobalsSection) { 1838 IRB_Dtor.CreateCall(AsanUnregisterImageGlobals, 1839 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)}); 1840 } else { 1841 IRB_Dtor.CreateCall(AsanUnregisterGlobals, 1842 {IRB.CreatePointerCast(AllGlobals, IntptrTy), 1843 ConstantInt::get(IntptrTy, n)}); 1844 } 1845 1846 appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndDtorPriority); 1847 1848 DEBUG(dbgs() << M); 1849 return true; 1850 } 1851 1852 bool AddressSanitizerModule::runOnModule(Module &M) { 1853 C = &(M.getContext()); 1854 int LongSize = M.getDataLayout().getPointerSizeInBits(); 1855 IntptrTy = Type::getIntNTy(*C, LongSize); 1856 TargetTriple = Triple(M.getTargetTriple()); 1857 Mapping = getShadowMapping(TargetTriple, LongSize, CompileKernel); 1858 initializeCallbacks(M); 1859 1860 bool Changed = false; 1861 1862 // TODO(glider): temporarily disabled globals instrumentation for KASan. 1863 if (ClGlobals && !CompileKernel) { 1864 Function *CtorFunc = M.getFunction(kAsanModuleCtorName); 1865 assert(CtorFunc); 1866 IRBuilder<> IRB(CtorFunc->getEntryBlock().getTerminator()); 1867 Changed |= InstrumentGlobals(IRB, M); 1868 } 1869 1870 return Changed; 1871 } 1872 1873 void AddressSanitizer::initializeCallbacks(Module &M) { 1874 IRBuilder<> IRB(*C); 1875 // Create __asan_report* callbacks. 1876 // IsWrite, TypeSize and Exp are encoded in the function name. 1877 for (int Exp = 0; Exp < 2; Exp++) { 1878 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) { 1879 const std::string TypeStr = AccessIsWrite ? "store" : "load"; 1880 const std::string ExpStr = Exp ? "exp_" : ""; 1881 const std::string SuffixStr = CompileKernel ? "N" : "_n"; 1882 const std::string EndingStr = Recover ? "_noabort" : ""; 1883 Type *ExpType = Exp ? Type::getInt32Ty(*C) : nullptr; 1884 AsanErrorCallbackSized[AccessIsWrite][Exp] = 1885 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1886 kAsanReportErrorTemplate + ExpStr + TypeStr + SuffixStr + EndingStr, 1887 IRB.getVoidTy(), IntptrTy, IntptrTy, ExpType, nullptr)); 1888 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] = 1889 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1890 ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr, 1891 IRB.getVoidTy(), IntptrTy, IntptrTy, ExpType, nullptr)); 1892 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes; 1893 AccessSizeIndex++) { 1894 const std::string Suffix = TypeStr + itostr(1ULL << AccessSizeIndex); 1895 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] = 1896 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1897 kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr, 1898 IRB.getVoidTy(), IntptrTy, ExpType, nullptr)); 1899 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] = 1900 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1901 ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr, 1902 IRB.getVoidTy(), IntptrTy, ExpType, nullptr)); 1903 } 1904 } 1905 } 1906 1907 const std::string MemIntrinCallbackPrefix = 1908 CompileKernel ? std::string("") : ClMemoryAccessCallbackPrefix; 1909 AsanMemmove = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1910 MemIntrinCallbackPrefix + "memmove", IRB.getInt8PtrTy(), 1911 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy, nullptr)); 1912 AsanMemcpy = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1913 MemIntrinCallbackPrefix + "memcpy", IRB.getInt8PtrTy(), 1914 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy, nullptr)); 1915 AsanMemset = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1916 MemIntrinCallbackPrefix + "memset", IRB.getInt8PtrTy(), 1917 IRB.getInt8PtrTy(), IRB.getInt32Ty(), IntptrTy, nullptr)); 1918 1919 AsanHandleNoReturnFunc = checkSanitizerInterfaceFunction( 1920 M.getOrInsertFunction(kAsanHandleNoReturnName, IRB.getVoidTy(), nullptr)); 1921 1922 AsanPtrCmpFunction = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1923 kAsanPtrCmp, IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 1924 AsanPtrSubFunction = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 1925 kAsanPtrSub, IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 1926 // We insert an empty inline asm after __asan_report* to avoid callback merge. 1927 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false), 1928 StringRef(""), StringRef(""), 1929 /*hasSideEffects=*/true); 1930 } 1931 1932 // virtual 1933 bool AddressSanitizer::doInitialization(Module &M) { 1934 // Initialize the private fields. No one has accessed them before. 1935 1936 GlobalsMD.init(M); 1937 1938 C = &(M.getContext()); 1939 LongSize = M.getDataLayout().getPointerSizeInBits(); 1940 IntptrTy = Type::getIntNTy(*C, LongSize); 1941 TargetTriple = Triple(M.getTargetTriple()); 1942 1943 if (!CompileKernel) { 1944 std::tie(AsanCtorFunction, AsanInitFunction) = 1945 createSanitizerCtorAndInitFunctions( 1946 M, kAsanModuleCtorName, kAsanInitName, 1947 /*InitArgTypes=*/{}, /*InitArgs=*/{}, kAsanVersionCheckName); 1948 appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndDtorPriority); 1949 } 1950 Mapping = getShadowMapping(TargetTriple, LongSize, CompileKernel); 1951 return true; 1952 } 1953 1954 bool AddressSanitizer::doFinalization(Module &M) { 1955 GlobalsMD.reset(); 1956 return false; 1957 } 1958 1959 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) { 1960 // For each NSObject descendant having a +load method, this method is invoked 1961 // by the ObjC runtime before any of the static constructors is called. 1962 // Therefore we need to instrument such methods with a call to __asan_init 1963 // at the beginning in order to initialize our runtime before any access to 1964 // the shadow memory. 1965 // We cannot just ignore these methods, because they may call other 1966 // instrumented functions. 1967 if (F.getName().find(" load]") != std::string::npos) { 1968 IRBuilder<> IRB(&F.front(), F.front().begin()); 1969 IRB.CreateCall(AsanInitFunction, {}); 1970 return true; 1971 } 1972 return false; 1973 } 1974 1975 void AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) { 1976 // Generate code only when dynamic addressing is needed. 1977 if (Mapping.Offset != kDynamicShadowSentinel) 1978 return; 1979 1980 IRBuilder<> IRB(&F.front().front()); 1981 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal( 1982 kAsanShadowMemoryDynamicAddress, IntptrTy); 1983 LocalDynamicShadow = IRB.CreateLoad(GlobalDynamicAddress); 1984 } 1985 1986 void AddressSanitizer::markEscapedLocalAllocas(Function &F) { 1987 // Find the one possible call to llvm.localescape and pre-mark allocas passed 1988 // to it as uninteresting. This assumes we haven't started processing allocas 1989 // yet. This check is done up front because iterating the use list in 1990 // isInterestingAlloca would be algorithmically slower. 1991 assert(ProcessedAllocas.empty() && "must process localescape before allocas"); 1992 1993 // Try to get the declaration of llvm.localescape. If it's not in the module, 1994 // we can exit early. 1995 if (!F.getParent()->getFunction("llvm.localescape")) return; 1996 1997 // Look for a call to llvm.localescape call in the entry block. It can't be in 1998 // any other block. 1999 for (Instruction &I : F.getEntryBlock()) { 2000 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I); 2001 if (II && II->getIntrinsicID() == Intrinsic::localescape) { 2002 // We found a call. Mark all the allocas passed in as uninteresting. 2003 for (Value *Arg : II->arg_operands()) { 2004 AllocaInst *AI = dyn_cast<AllocaInst>(Arg->stripPointerCasts()); 2005 assert(AI && AI->isStaticAlloca() && 2006 "non-static alloca arg to localescape"); 2007 ProcessedAllocas[AI] = false; 2008 } 2009 break; 2010 } 2011 } 2012 } 2013 2014 bool AddressSanitizer::runOnFunction(Function &F) { 2015 if (&F == AsanCtorFunction) return false; 2016 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false; 2017 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName()) return false; 2018 if (F.getName().startswith("__asan_")) return false; 2019 2020 bool FunctionModified = false; 2021 2022 // If needed, insert __asan_init before checking for SanitizeAddress attr. 2023 // This function needs to be called even if the function body is not 2024 // instrumented. 2025 if (maybeInsertAsanInitAtFunctionEntry(F)) 2026 FunctionModified = true; 2027 2028 // Leave if the function doesn't need instrumentation. 2029 if (!F.hasFnAttribute(Attribute::SanitizeAddress)) return FunctionModified; 2030 2031 DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n"); 2032 2033 initializeCallbacks(*F.getParent()); 2034 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2035 2036 FunctionStateRAII CleanupObj(this); 2037 2038 maybeInsertDynamicShadowAtFunctionEntry(F); 2039 2040 // We can't instrument allocas used with llvm.localescape. Only static allocas 2041 // can be passed to that intrinsic. 2042 markEscapedLocalAllocas(F); 2043 2044 // We want to instrument every address only once per basic block (unless there 2045 // are calls between uses). 2046 SmallSet<Value *, 16> TempsToInstrument; 2047 SmallVector<Instruction *, 16> ToInstrument; 2048 SmallVector<Instruction *, 8> NoReturnCalls; 2049 SmallVector<BasicBlock *, 16> AllBlocks; 2050 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts; 2051 int NumAllocas = 0; 2052 bool IsWrite; 2053 unsigned Alignment; 2054 uint64_t TypeSize; 2055 const TargetLibraryInfo *TLI = 2056 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 2057 2058 // Fill the set of memory operations to instrument. 2059 for (auto &BB : F) { 2060 AllBlocks.push_back(&BB); 2061 TempsToInstrument.clear(); 2062 int NumInsnsPerBB = 0; 2063 for (auto &Inst : BB) { 2064 if (LooksLikeCodeInBug11395(&Inst)) return false; 2065 if (Value *Addr = isInterestingMemoryAccess(&Inst, &IsWrite, &TypeSize, 2066 &Alignment)) { 2067 if (ClOpt && ClOptSameTemp) { 2068 if (!TempsToInstrument.insert(Addr).second) 2069 continue; // We've seen this temp in the current BB. 2070 } 2071 } else if (ClInvalidPointerPairs && 2072 isInterestingPointerComparisonOrSubtraction(&Inst)) { 2073 PointerComparisonsOrSubtracts.push_back(&Inst); 2074 continue; 2075 } else if (isa<MemIntrinsic>(Inst)) { 2076 // ok, take it. 2077 } else { 2078 if (isa<AllocaInst>(Inst)) NumAllocas++; 2079 CallSite CS(&Inst); 2080 if (CS) { 2081 // A call inside BB. 2082 TempsToInstrument.clear(); 2083 if (CS.doesNotReturn()) NoReturnCalls.push_back(CS.getInstruction()); 2084 } 2085 if (CallInst *CI = dyn_cast<CallInst>(&Inst)) 2086 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI); 2087 continue; 2088 } 2089 ToInstrument.push_back(&Inst); 2090 NumInsnsPerBB++; 2091 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break; 2092 } 2093 } 2094 2095 bool UseCalls = 2096 CompileKernel || 2097 (ClInstrumentationWithCallsThreshold >= 0 && 2098 ToInstrument.size() > (unsigned)ClInstrumentationWithCallsThreshold); 2099 const DataLayout &DL = F.getParent()->getDataLayout(); 2100 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext(), 2101 /*RoundToAlign=*/true); 2102 2103 // Instrument. 2104 int NumInstrumented = 0; 2105 for (auto Inst : ToInstrument) { 2106 if (ClDebugMin < 0 || ClDebugMax < 0 || 2107 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) { 2108 if (isInterestingMemoryAccess(Inst, &IsWrite, &TypeSize, &Alignment)) 2109 instrumentMop(ObjSizeVis, Inst, UseCalls, 2110 F.getParent()->getDataLayout()); 2111 else 2112 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst)); 2113 } 2114 NumInstrumented++; 2115 } 2116 2117 FunctionStackPoisoner FSP(F, *this); 2118 bool ChangedStack = FSP.runOnFunction(); 2119 2120 // We must unpoison the stack before every NoReturn call (throw, _exit, etc). 2121 // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37 2122 for (auto CI : NoReturnCalls) { 2123 IRBuilder<> IRB(CI); 2124 IRB.CreateCall(AsanHandleNoReturnFunc, {}); 2125 } 2126 2127 for (auto Inst : PointerComparisonsOrSubtracts) { 2128 instrumentPointerComparisonOrSubtraction(Inst); 2129 NumInstrumented++; 2130 } 2131 2132 if (NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty()) 2133 FunctionModified = true; 2134 2135 DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " " 2136 << F << "\n"); 2137 2138 return FunctionModified; 2139 } 2140 2141 // Workaround for bug 11395: we don't want to instrument stack in functions 2142 // with large assembly blobs (32-bit only), otherwise reg alloc may crash. 2143 // FIXME: remove once the bug 11395 is fixed. 2144 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) { 2145 if (LongSize != 32) return false; 2146 CallInst *CI = dyn_cast<CallInst>(I); 2147 if (!CI || !CI->isInlineAsm()) return false; 2148 if (CI->getNumArgOperands() <= 5) return false; 2149 // We have inline assembly with quite a few arguments. 2150 return true; 2151 } 2152 2153 void FunctionStackPoisoner::initializeCallbacks(Module &M) { 2154 IRBuilder<> IRB(*C); 2155 for (int i = 0; i <= kMaxAsanStackMallocSizeClass; i++) { 2156 std::string Suffix = itostr(i); 2157 AsanStackMallocFunc[i] = checkSanitizerInterfaceFunction( 2158 M.getOrInsertFunction(kAsanStackMallocNameTemplate + Suffix, IntptrTy, 2159 IntptrTy, nullptr)); 2160 AsanStackFreeFunc[i] = checkSanitizerInterfaceFunction( 2161 M.getOrInsertFunction(kAsanStackFreeNameTemplate + Suffix, 2162 IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 2163 } 2164 if (ASan.UseAfterScope) { 2165 AsanPoisonStackMemoryFunc = checkSanitizerInterfaceFunction( 2166 M.getOrInsertFunction(kAsanPoisonStackMemoryName, IRB.getVoidTy(), 2167 IntptrTy, IntptrTy, nullptr)); 2168 AsanUnpoisonStackMemoryFunc = checkSanitizerInterfaceFunction( 2169 M.getOrInsertFunction(kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), 2170 IntptrTy, IntptrTy, nullptr)); 2171 } 2172 2173 for (size_t Val : {0x00, 0xf1, 0xf2, 0xf3, 0xf5, 0xf8}) { 2174 std::ostringstream Name; 2175 Name << kAsanSetShadowPrefix; 2176 Name << std::setw(2) << std::setfill('0') << std::hex << Val; 2177 AsanSetShadowFunc[Val] = 2178 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 2179 Name.str(), IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 2180 } 2181 2182 AsanAllocaPoisonFunc = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 2183 kAsanAllocaPoison, IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 2184 AsanAllocasUnpoisonFunc = 2185 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 2186 kAsanAllocasUnpoison, IRB.getVoidTy(), IntptrTy, IntptrTy, nullptr)); 2187 } 2188 2189 void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask, 2190 ArrayRef<uint8_t> ShadowBytes, 2191 size_t Begin, size_t End, 2192 IRBuilder<> &IRB, 2193 Value *ShadowBase) { 2194 if (Begin >= End) 2195 return; 2196 2197 const size_t LargestStoreSizeInBytes = 2198 std::min<size_t>(sizeof(uint64_t), ASan.LongSize / 8); 2199 2200 const bool IsLittleEndian = F.getParent()->getDataLayout().isLittleEndian(); 2201 2202 // Poison given range in shadow using larges store size with out leading and 2203 // trailing zeros in ShadowMask. Zeros never change, so they need neither 2204 // poisoning nor up-poisoning. Still we don't mind if some of them get into a 2205 // middle of a store. 2206 for (size_t i = Begin; i < End;) { 2207 if (!ShadowMask[i]) { 2208 assert(!ShadowBytes[i]); 2209 ++i; 2210 continue; 2211 } 2212 2213 size_t StoreSizeInBytes = LargestStoreSizeInBytes; 2214 // Fit store size into the range. 2215 while (StoreSizeInBytes > End - i) 2216 StoreSizeInBytes /= 2; 2217 2218 // Minimize store size by trimming trailing zeros. 2219 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) { 2220 while (j <= StoreSizeInBytes / 2) 2221 StoreSizeInBytes /= 2; 2222 } 2223 2224 uint64_t Val = 0; 2225 for (size_t j = 0; j < StoreSizeInBytes; j++) { 2226 if (IsLittleEndian) 2227 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j); 2228 else 2229 Val = (Val << 8) | ShadowBytes[i + j]; 2230 } 2231 2232 Value *Ptr = IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i)); 2233 Value *Poison = IRB.getIntN(StoreSizeInBytes * 8, Val); 2234 IRB.CreateAlignedStore( 2235 Poison, IRB.CreateIntToPtr(Ptr, Poison->getType()->getPointerTo()), 1); 2236 2237 i += StoreSizeInBytes; 2238 } 2239 } 2240 2241 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask, 2242 ArrayRef<uint8_t> ShadowBytes, 2243 IRBuilder<> &IRB, Value *ShadowBase) { 2244 copyToShadow(ShadowMask, ShadowBytes, 0, ShadowMask.size(), IRB, ShadowBase); 2245 } 2246 2247 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask, 2248 ArrayRef<uint8_t> ShadowBytes, 2249 size_t Begin, size_t End, 2250 IRBuilder<> &IRB, Value *ShadowBase) { 2251 assert(ShadowMask.size() == ShadowBytes.size()); 2252 size_t Done = Begin; 2253 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) { 2254 if (!ShadowMask[i]) { 2255 assert(!ShadowBytes[i]); 2256 continue; 2257 } 2258 uint8_t Val = ShadowBytes[i]; 2259 if (!AsanSetShadowFunc[Val]) 2260 continue; 2261 2262 // Skip same values. 2263 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) { 2264 } 2265 2266 if (j - i >= ClMaxInlinePoisoningSize) { 2267 copyToShadowInline(ShadowMask, ShadowBytes, Done, i, IRB, ShadowBase); 2268 IRB.CreateCall(AsanSetShadowFunc[Val], 2269 {IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i)), 2270 ConstantInt::get(IntptrTy, j - i)}); 2271 Done = j; 2272 } 2273 } 2274 2275 copyToShadowInline(ShadowMask, ShadowBytes, Done, End, IRB, ShadowBase); 2276 } 2277 2278 // Fake stack allocator (asan_fake_stack.h) has 11 size classes 2279 // for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass 2280 static int StackMallocSizeClass(uint64_t LocalStackSize) { 2281 assert(LocalStackSize <= kMaxStackMallocSize); 2282 uint64_t MaxSize = kMinStackMallocSize; 2283 for (int i = 0;; i++, MaxSize *= 2) 2284 if (LocalStackSize <= MaxSize) return i; 2285 llvm_unreachable("impossible LocalStackSize"); 2286 } 2287 2288 PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond, 2289 Value *ValueIfTrue, 2290 Instruction *ThenTerm, 2291 Value *ValueIfFalse) { 2292 PHINode *PHI = IRB.CreatePHI(IntptrTy, 2); 2293 BasicBlock *CondBlock = cast<Instruction>(Cond)->getParent(); 2294 PHI->addIncoming(ValueIfFalse, CondBlock); 2295 BasicBlock *ThenBlock = ThenTerm->getParent(); 2296 PHI->addIncoming(ValueIfTrue, ThenBlock); 2297 return PHI; 2298 } 2299 2300 Value *FunctionStackPoisoner::createAllocaForLayout( 2301 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) { 2302 AllocaInst *Alloca; 2303 if (Dynamic) { 2304 Alloca = IRB.CreateAlloca(IRB.getInt8Ty(), 2305 ConstantInt::get(IRB.getInt64Ty(), L.FrameSize), 2306 "MyAlloca"); 2307 } else { 2308 Alloca = IRB.CreateAlloca(ArrayType::get(IRB.getInt8Ty(), L.FrameSize), 2309 nullptr, "MyAlloca"); 2310 assert(Alloca->isStaticAlloca()); 2311 } 2312 assert((ClRealignStack & (ClRealignStack - 1)) == 0); 2313 size_t FrameAlignment = std::max(L.FrameAlignment, (size_t)ClRealignStack); 2314 Alloca->setAlignment(FrameAlignment); 2315 return IRB.CreatePointerCast(Alloca, IntptrTy); 2316 } 2317 2318 void FunctionStackPoisoner::createDynamicAllocasInitStorage() { 2319 BasicBlock &FirstBB = *F.begin(); 2320 IRBuilder<> IRB(dyn_cast<Instruction>(FirstBB.begin())); 2321 DynamicAllocaLayout = IRB.CreateAlloca(IntptrTy, nullptr); 2322 IRB.CreateStore(Constant::getNullValue(IntptrTy), DynamicAllocaLayout); 2323 DynamicAllocaLayout->setAlignment(32); 2324 } 2325 2326 void FunctionStackPoisoner::processDynamicAllocas() { 2327 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) { 2328 assert(DynamicAllocaPoisonCallVec.empty()); 2329 return; 2330 } 2331 2332 // Insert poison calls for lifetime intrinsics for dynamic allocas. 2333 for (const auto &APC : DynamicAllocaPoisonCallVec) { 2334 assert(APC.InsBefore); 2335 assert(APC.AI); 2336 assert(ASan.isInterestingAlloca(*APC.AI)); 2337 assert(!APC.AI->isStaticAlloca()); 2338 2339 IRBuilder<> IRB(APC.InsBefore); 2340 poisonAlloca(APC.AI, APC.Size, IRB, APC.DoPoison); 2341 // Dynamic allocas will be unpoisoned unconditionally below in 2342 // unpoisonDynamicAllocas. 2343 // Flag that we need unpoison static allocas. 2344 } 2345 2346 // Handle dynamic allocas. 2347 createDynamicAllocasInitStorage(); 2348 for (auto &AI : DynamicAllocaVec) 2349 handleDynamicAllocaCall(AI); 2350 unpoisonDynamicAllocas(); 2351 } 2352 2353 void FunctionStackPoisoner::processStaticAllocas() { 2354 if (AllocaVec.empty()) { 2355 assert(StaticAllocaPoisonCallVec.empty()); 2356 return; 2357 } 2358 2359 int StackMallocIdx = -1; 2360 DebugLoc EntryDebugLocation; 2361 if (auto SP = F.getSubprogram()) 2362 EntryDebugLocation = DebugLoc::get(SP->getScopeLine(), 0, SP); 2363 2364 Instruction *InsBefore = AllocaVec[0]; 2365 IRBuilder<> IRB(InsBefore); 2366 IRB.SetCurrentDebugLocation(EntryDebugLocation); 2367 2368 // Make sure non-instrumented allocas stay in the entry block. Otherwise, 2369 // debug info is broken, because only entry-block allocas are treated as 2370 // regular stack slots. 2371 auto InsBeforeB = InsBefore->getParent(); 2372 assert(InsBeforeB == &F.getEntryBlock()); 2373 for (auto *AI : StaticAllocasToMoveUp) 2374 if (AI->getParent() == InsBeforeB) 2375 AI->moveBefore(InsBefore); 2376 2377 // If we have a call to llvm.localescape, keep it in the entry block. 2378 if (LocalEscapeCall) LocalEscapeCall->moveBefore(InsBefore); 2379 2380 SmallVector<ASanStackVariableDescription, 16> SVD; 2381 SVD.reserve(AllocaVec.size()); 2382 for (AllocaInst *AI : AllocaVec) { 2383 ASanStackVariableDescription D = {AI->getName().data(), 2384 ASan.getAllocaSizeInBytes(*AI), 2385 0, 2386 AI->getAlignment(), 2387 AI, 2388 0, 2389 0}; 2390 SVD.push_back(D); 2391 } 2392 2393 // Minimal header size (left redzone) is 4 pointers, 2394 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms. 2395 size_t MinHeaderSize = ASan.LongSize / 2; 2396 const ASanStackFrameLayout &L = 2397 ComputeASanStackFrameLayout(SVD, 1ULL << Mapping.Scale, MinHeaderSize); 2398 2399 // Build AllocaToSVDMap for ASanStackVariableDescription lookup. 2400 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap; 2401 for (auto &Desc : SVD) 2402 AllocaToSVDMap[Desc.AI] = &Desc; 2403 2404 // Update SVD with information from lifetime intrinsics. 2405 for (const auto &APC : StaticAllocaPoisonCallVec) { 2406 assert(APC.InsBefore); 2407 assert(APC.AI); 2408 assert(ASan.isInterestingAlloca(*APC.AI)); 2409 assert(APC.AI->isStaticAlloca()); 2410 2411 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI]; 2412 Desc.LifetimeSize = Desc.Size; 2413 if (const DILocation *FnLoc = EntryDebugLocation.get()) { 2414 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) { 2415 if (LifetimeLoc->getFile() == FnLoc->getFile()) 2416 if (unsigned Line = LifetimeLoc->getLine()) 2417 Desc.Line = std::min(Desc.Line ? Desc.Line : Line, Line); 2418 } 2419 } 2420 } 2421 2422 auto DescriptionString = ComputeASanStackFrameDescription(SVD); 2423 DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n"); 2424 uint64_t LocalStackSize = L.FrameSize; 2425 bool DoStackMalloc = ClUseAfterReturn && !ASan.CompileKernel && 2426 LocalStackSize <= kMaxStackMallocSize; 2427 bool DoDynamicAlloca = ClDynamicAllocaStack; 2428 // Don't do dynamic alloca or stack malloc if: 2429 // 1) There is inline asm: too often it makes assumptions on which registers 2430 // are available. 2431 // 2) There is a returns_twice call (typically setjmp), which is 2432 // optimization-hostile, and doesn't play well with introduced indirect 2433 // register-relative calculation of local variable addresses. 2434 DoDynamicAlloca &= !HasNonEmptyInlineAsm && !HasReturnsTwiceCall; 2435 DoStackMalloc &= !HasNonEmptyInlineAsm && !HasReturnsTwiceCall; 2436 2437 Value *StaticAlloca = 2438 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, false); 2439 2440 Value *FakeStack; 2441 Value *LocalStackBase; 2442 2443 if (DoStackMalloc) { 2444 // void *FakeStack = __asan_option_detect_stack_use_after_return 2445 // ? __asan_stack_malloc_N(LocalStackSize) 2446 // : nullptr; 2447 // void *LocalStackBase = (FakeStack) ? FakeStack : alloca(LocalStackSize); 2448 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal( 2449 kAsanOptionDetectUseAfterReturn, IRB.getInt32Ty()); 2450 Value *UseAfterReturnIsEnabled = 2451 IRB.CreateICmpNE(IRB.CreateLoad(OptionDetectUseAfterReturn), 2452 Constant::getNullValue(IRB.getInt32Ty())); 2453 Instruction *Term = 2454 SplitBlockAndInsertIfThen(UseAfterReturnIsEnabled, InsBefore, false); 2455 IRBuilder<> IRBIf(Term); 2456 IRBIf.SetCurrentDebugLocation(EntryDebugLocation); 2457 StackMallocIdx = StackMallocSizeClass(LocalStackSize); 2458 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass); 2459 Value *FakeStackValue = 2460 IRBIf.CreateCall(AsanStackMallocFunc[StackMallocIdx], 2461 ConstantInt::get(IntptrTy, LocalStackSize)); 2462 IRB.SetInsertPoint(InsBefore); 2463 IRB.SetCurrentDebugLocation(EntryDebugLocation); 2464 FakeStack = createPHI(IRB, UseAfterReturnIsEnabled, FakeStackValue, Term, 2465 ConstantInt::get(IntptrTy, 0)); 2466 2467 Value *NoFakeStack = 2468 IRB.CreateICmpEQ(FakeStack, Constant::getNullValue(IntptrTy)); 2469 Term = SplitBlockAndInsertIfThen(NoFakeStack, InsBefore, false); 2470 IRBIf.SetInsertPoint(Term); 2471 IRBIf.SetCurrentDebugLocation(EntryDebugLocation); 2472 Value *AllocaValue = 2473 DoDynamicAlloca ? createAllocaForLayout(IRBIf, L, true) : StaticAlloca; 2474 IRB.SetInsertPoint(InsBefore); 2475 IRB.SetCurrentDebugLocation(EntryDebugLocation); 2476 LocalStackBase = createPHI(IRB, NoFakeStack, AllocaValue, Term, FakeStack); 2477 } else { 2478 // void *FakeStack = nullptr; 2479 // void *LocalStackBase = alloca(LocalStackSize); 2480 FakeStack = ConstantInt::get(IntptrTy, 0); 2481 LocalStackBase = 2482 DoDynamicAlloca ? createAllocaForLayout(IRB, L, true) : StaticAlloca; 2483 } 2484 2485 // Replace Alloca instructions with base+offset. 2486 for (const auto &Desc : SVD) { 2487 AllocaInst *AI = Desc.AI; 2488 Value *NewAllocaPtr = IRB.CreateIntToPtr( 2489 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Desc.Offset)), 2490 AI->getType()); 2491 replaceDbgDeclareForAlloca(AI, NewAllocaPtr, DIB, /*Deref=*/true); 2492 AI->replaceAllUsesWith(NewAllocaPtr); 2493 } 2494 2495 // The left-most redzone has enough space for at least 4 pointers. 2496 // Write the Magic value to redzone[0]. 2497 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy); 2498 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic), 2499 BasePlus0); 2500 // Write the frame description constant to redzone[1]. 2501 Value *BasePlus1 = IRB.CreateIntToPtr( 2502 IRB.CreateAdd(LocalStackBase, 2503 ConstantInt::get(IntptrTy, ASan.LongSize / 8)), 2504 IntptrPtrTy); 2505 GlobalVariable *StackDescriptionGlobal = 2506 createPrivateGlobalForString(*F.getParent(), DescriptionString, 2507 /*AllowMerging*/ true); 2508 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal, IntptrTy); 2509 IRB.CreateStore(Description, BasePlus1); 2510 // Write the PC to redzone[2]. 2511 Value *BasePlus2 = IRB.CreateIntToPtr( 2512 IRB.CreateAdd(LocalStackBase, 2513 ConstantInt::get(IntptrTy, 2 * ASan.LongSize / 8)), 2514 IntptrPtrTy); 2515 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2); 2516 2517 const auto &ShadowAfterScope = GetShadowBytesAfterScope(SVD, L); 2518 2519 // Poison the stack red zones at the entry. 2520 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB); 2521 // As mask we must use most poisoned case: red zones and after scope. 2522 // As bytes we can use either the same or just red zones only. 2523 copyToShadow(ShadowAfterScope, ShadowAfterScope, IRB, ShadowBase); 2524 2525 if (!StaticAllocaPoisonCallVec.empty()) { 2526 const auto &ShadowInScope = GetShadowBytes(SVD, L); 2527 2528 // Poison static allocas near lifetime intrinsics. 2529 for (const auto &APC : StaticAllocaPoisonCallVec) { 2530 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI]; 2531 assert(Desc.Offset % L.Granularity == 0); 2532 size_t Begin = Desc.Offset / L.Granularity; 2533 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity; 2534 2535 IRBuilder<> IRB(APC.InsBefore); 2536 copyToShadow(ShadowAfterScope, 2537 APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End, 2538 IRB, ShadowBase); 2539 } 2540 } 2541 2542 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0); 2543 SmallVector<uint8_t, 64> ShadowAfterReturn; 2544 2545 // (Un)poison the stack before all ret instructions. 2546 for (auto Ret : RetVec) { 2547 IRBuilder<> IRBRet(Ret); 2548 // Mark the current frame as retired. 2549 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic), 2550 BasePlus0); 2551 if (DoStackMalloc) { 2552 assert(StackMallocIdx >= 0); 2553 // if FakeStack != 0 // LocalStackBase == FakeStack 2554 // // In use-after-return mode, poison the whole stack frame. 2555 // if StackMallocIdx <= 4 2556 // // For small sizes inline the whole thing: 2557 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize); 2558 // **SavedFlagPtr(FakeStack) = 0 2559 // else 2560 // __asan_stack_free_N(FakeStack, LocalStackSize) 2561 // else 2562 // <This is not a fake stack; unpoison the redzones> 2563 Value *Cmp = 2564 IRBRet.CreateICmpNE(FakeStack, Constant::getNullValue(IntptrTy)); 2565 TerminatorInst *ThenTerm, *ElseTerm; 2566 SplitBlockAndInsertIfThenElse(Cmp, Ret, &ThenTerm, &ElseTerm); 2567 2568 IRBuilder<> IRBPoison(ThenTerm); 2569 if (StackMallocIdx <= 4) { 2570 int ClassSize = kMinStackMallocSize << StackMallocIdx; 2571 ShadowAfterReturn.resize(ClassSize / L.Granularity, 2572 kAsanStackUseAfterReturnMagic); 2573 copyToShadow(ShadowAfterReturn, ShadowAfterReturn, IRBPoison, 2574 ShadowBase); 2575 Value *SavedFlagPtrPtr = IRBPoison.CreateAdd( 2576 FakeStack, 2577 ConstantInt::get(IntptrTy, ClassSize - ASan.LongSize / 8)); 2578 Value *SavedFlagPtr = IRBPoison.CreateLoad( 2579 IRBPoison.CreateIntToPtr(SavedFlagPtrPtr, IntptrPtrTy)); 2580 IRBPoison.CreateStore( 2581 Constant::getNullValue(IRBPoison.getInt8Ty()), 2582 IRBPoison.CreateIntToPtr(SavedFlagPtr, IRBPoison.getInt8PtrTy())); 2583 } else { 2584 // For larger frames call __asan_stack_free_*. 2585 IRBPoison.CreateCall( 2586 AsanStackFreeFunc[StackMallocIdx], 2587 {FakeStack, ConstantInt::get(IntptrTy, LocalStackSize)}); 2588 } 2589 2590 IRBuilder<> IRBElse(ElseTerm); 2591 copyToShadow(ShadowAfterScope, ShadowClean, IRBElse, ShadowBase); 2592 } else { 2593 copyToShadow(ShadowAfterScope, ShadowClean, IRBRet, ShadowBase); 2594 } 2595 } 2596 2597 // We are done. Remove the old unused alloca instructions. 2598 for (auto AI : AllocaVec) AI->eraseFromParent(); 2599 } 2600 2601 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size, 2602 IRBuilder<> &IRB, bool DoPoison) { 2603 // For now just insert the call to ASan runtime. 2604 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy); 2605 Value *SizeArg = ConstantInt::get(IntptrTy, Size); 2606 IRB.CreateCall( 2607 DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc, 2608 {AddrArg, SizeArg}); 2609 } 2610 2611 // Handling llvm.lifetime intrinsics for a given %alloca: 2612 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca. 2613 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect 2614 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory 2615 // could be poisoned by previous llvm.lifetime.end instruction, as the 2616 // variable may go in and out of scope several times, e.g. in loops). 2617 // (3) if we poisoned at least one %alloca in a function, 2618 // unpoison the whole stack frame at function exit. 2619 2620 AllocaInst *FunctionStackPoisoner::findAllocaForValue(Value *V) { 2621 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) 2622 // We're interested only in allocas we can handle. 2623 return ASan.isInterestingAlloca(*AI) ? AI : nullptr; 2624 // See if we've already calculated (or started to calculate) alloca for a 2625 // given value. 2626 AllocaForValueMapTy::iterator I = AllocaForValue.find(V); 2627 if (I != AllocaForValue.end()) return I->second; 2628 // Store 0 while we're calculating alloca for value V to avoid 2629 // infinite recursion if the value references itself. 2630 AllocaForValue[V] = nullptr; 2631 AllocaInst *Res = nullptr; 2632 if (CastInst *CI = dyn_cast<CastInst>(V)) 2633 Res = findAllocaForValue(CI->getOperand(0)); 2634 else if (PHINode *PN = dyn_cast<PHINode>(V)) { 2635 for (Value *IncValue : PN->incoming_values()) { 2636 // Allow self-referencing phi-nodes. 2637 if (IncValue == PN) continue; 2638 AllocaInst *IncValueAI = findAllocaForValue(IncValue); 2639 // AI for incoming values should exist and should all be equal. 2640 if (IncValueAI == nullptr || (Res != nullptr && IncValueAI != Res)) 2641 return nullptr; 2642 Res = IncValueAI; 2643 } 2644 } else if (GetElementPtrInst *EP = dyn_cast<GetElementPtrInst>(V)) { 2645 Res = findAllocaForValue(EP->getPointerOperand()); 2646 } else { 2647 DEBUG(dbgs() << "Alloca search canceled on unknown instruction: " << *V << "\n"); 2648 } 2649 if (Res) AllocaForValue[V] = Res; 2650 return Res; 2651 } 2652 2653 void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) { 2654 IRBuilder<> IRB(AI); 2655 2656 const unsigned Align = std::max(kAllocaRzSize, AI->getAlignment()); 2657 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1; 2658 2659 Value *Zero = Constant::getNullValue(IntptrTy); 2660 Value *AllocaRzSize = ConstantInt::get(IntptrTy, kAllocaRzSize); 2661 Value *AllocaRzMask = ConstantInt::get(IntptrTy, AllocaRedzoneMask); 2662 2663 // Since we need to extend alloca with additional memory to locate 2664 // redzones, and OldSize is number of allocated blocks with 2665 // ElementSize size, get allocated memory size in bytes by 2666 // OldSize * ElementSize. 2667 const unsigned ElementSize = 2668 F.getParent()->getDataLayout().getTypeAllocSize(AI->getAllocatedType()); 2669 Value *OldSize = 2670 IRB.CreateMul(IRB.CreateIntCast(AI->getArraySize(), IntptrTy, false), 2671 ConstantInt::get(IntptrTy, ElementSize)); 2672 2673 // PartialSize = OldSize % 32 2674 Value *PartialSize = IRB.CreateAnd(OldSize, AllocaRzMask); 2675 2676 // Misalign = kAllocaRzSize - PartialSize; 2677 Value *Misalign = IRB.CreateSub(AllocaRzSize, PartialSize); 2678 2679 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0; 2680 Value *Cond = IRB.CreateICmpNE(Misalign, AllocaRzSize); 2681 Value *PartialPadding = IRB.CreateSelect(Cond, Misalign, Zero); 2682 2683 // AdditionalChunkSize = Align + PartialPadding + kAllocaRzSize 2684 // Align is added to locate left redzone, PartialPadding for possible 2685 // partial redzone and kAllocaRzSize for right redzone respectively. 2686 Value *AdditionalChunkSize = IRB.CreateAdd( 2687 ConstantInt::get(IntptrTy, Align + kAllocaRzSize), PartialPadding); 2688 2689 Value *NewSize = IRB.CreateAdd(OldSize, AdditionalChunkSize); 2690 2691 // Insert new alloca with new NewSize and Align params. 2692 AllocaInst *NewAlloca = IRB.CreateAlloca(IRB.getInt8Ty(), NewSize); 2693 NewAlloca->setAlignment(Align); 2694 2695 // NewAddress = Address + Align 2696 Value *NewAddress = IRB.CreateAdd(IRB.CreatePtrToInt(NewAlloca, IntptrTy), 2697 ConstantInt::get(IntptrTy, Align)); 2698 2699 // Insert __asan_alloca_poison call for new created alloca. 2700 IRB.CreateCall(AsanAllocaPoisonFunc, {NewAddress, OldSize}); 2701 2702 // Store the last alloca's address to DynamicAllocaLayout. We'll need this 2703 // for unpoisoning stuff. 2704 IRB.CreateStore(IRB.CreatePtrToInt(NewAlloca, IntptrTy), DynamicAllocaLayout); 2705 2706 Value *NewAddressPtr = IRB.CreateIntToPtr(NewAddress, AI->getType()); 2707 2708 // Replace all uses of AddessReturnedByAlloca with NewAddressPtr. 2709 AI->replaceAllUsesWith(NewAddressPtr); 2710 2711 // We are done. Erase old alloca from parent. 2712 AI->eraseFromParent(); 2713 } 2714 2715 // isSafeAccess returns true if Addr is always inbounds with respect to its 2716 // base object. For example, it is a field access or an array access with 2717 // constant inbounds index. 2718 bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, 2719 Value *Addr, uint64_t TypeSize) const { 2720 SizeOffsetType SizeOffset = ObjSizeVis.compute(Addr); 2721 if (!ObjSizeVis.bothKnown(SizeOffset)) return false; 2722 uint64_t Size = SizeOffset.first.getZExtValue(); 2723 int64_t Offset = SizeOffset.second.getSExtValue(); 2724 // Three checks are required to ensure safety: 2725 // . Offset >= 0 (since the offset is given from the base ptr) 2726 // . Size >= Offset (unsigned) 2727 // . Size - Offset >= NeededSize (unsigned) 2728 return Offset >= 0 && Size >= uint64_t(Offset) && 2729 Size - uint64_t(Offset) >= TypeSize / 8; 2730 } 2731