1 //===-- MemorySanitizer.cpp - detector of uninitialized reads -------------===// 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 /// \file 10 /// This file is a part of MemorySanitizer, a detector of uninitialized 11 /// reads. 12 /// 13 /// Status: early prototype. 14 /// 15 /// The algorithm of the tool is similar to Memcheck 16 /// (http://goo.gl/QKbem). We associate a few shadow bits with every 17 /// byte of the application memory, poison the shadow of the malloc-ed 18 /// or alloca-ed memory, load the shadow bits on every memory read, 19 /// propagate the shadow bits through some of the arithmetic 20 /// instruction (including MOV), store the shadow bits on every memory 21 /// write, report a bug on some other instructions (e.g. JMP) if the 22 /// associated shadow is poisoned. 23 /// 24 /// But there are differences too. The first and the major one: 25 /// compiler instrumentation instead of binary instrumentation. This 26 /// gives us much better register allocation, possible compiler 27 /// optimizations and a fast start-up. But this brings the major issue 28 /// as well: msan needs to see all program events, including system 29 /// calls and reads/writes in system libraries, so we either need to 30 /// compile *everything* with msan or use a binary translation 31 /// component (e.g. DynamoRIO) to instrument pre-built libraries. 32 /// Another difference from Memcheck is that we use 8 shadow bits per 33 /// byte of application memory and use a direct shadow mapping. This 34 /// greatly simplifies the instrumentation code and avoids races on 35 /// shadow updates (Memcheck is single-threaded so races are not a 36 /// concern there. Memcheck uses 2 shadow bits per byte with a slow 37 /// path storage that uses 8 bits per byte). 38 /// 39 /// The default value of shadow is 0, which means "clean" (not poisoned). 40 /// 41 /// Every module initializer should call __msan_init to ensure that the 42 /// shadow memory is ready. On error, __msan_warning is called. Since 43 /// parameters and return values may be passed via registers, we have a 44 /// specialized thread-local shadow for return values 45 /// (__msan_retval_tls) and parameters (__msan_param_tls). 46 /// 47 /// Origin tracking. 48 /// 49 /// MemorySanitizer can track origins (allocation points) of all uninitialized 50 /// values. This behavior is controlled with a flag (msan-track-origins) and is 51 /// disabled by default. 52 /// 53 /// Origins are 4-byte values created and interpreted by the runtime library. 54 /// They are stored in a second shadow mapping, one 4-byte value for 4 bytes 55 /// of application memory. Propagation of origins is basically a bunch of 56 /// "select" instructions that pick the origin of a dirty argument, if an 57 /// instruction has one. 58 /// 59 /// Every 4 aligned, consecutive bytes of application memory have one origin 60 /// value associated with them. If these bytes contain uninitialized data 61 /// coming from 2 different allocations, the last store wins. Because of this, 62 /// MemorySanitizer reports can show unrelated origins, but this is unlikely in 63 /// practice. 64 /// 65 /// Origins are meaningless for fully initialized values, so MemorySanitizer 66 /// avoids storing origin to memory when a fully initialized value is stored. 67 /// This way it avoids needless overwritting origin of the 4-byte region on 68 /// a short (i.e. 1 byte) clean store, and it is also good for performance. 69 //===----------------------------------------------------------------------===// 70 71 #define DEBUG_TYPE "msan" 72 73 #include "llvm/Transforms/Instrumentation.h" 74 #include "llvm/ADT/DepthFirstIterator.h" 75 #include "llvm/ADT/SmallString.h" 76 #include "llvm/ADT/SmallVector.h" 77 #include "llvm/ADT/ValueMap.h" 78 #include "llvm/IR/DataLayout.h" 79 #include "llvm/IR/Function.h" 80 #include "llvm/IR/IRBuilder.h" 81 #include "llvm/IR/InlineAsm.h" 82 #include "llvm/IR/IntrinsicInst.h" 83 #include "llvm/IR/LLVMContext.h" 84 #include "llvm/IR/MDBuilder.h" 85 #include "llvm/IR/Module.h" 86 #include "llvm/IR/Type.h" 87 #include "llvm/InstVisitor.h" 88 #include "llvm/Support/CommandLine.h" 89 #include "llvm/Support/Compiler.h" 90 #include "llvm/Support/Debug.h" 91 #include "llvm/Support/raw_ostream.h" 92 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 93 #include "llvm/Transforms/Utils/BlackList.h" 94 #include "llvm/Transforms/Utils/Local.h" 95 #include "llvm/Transforms/Utils/ModuleUtils.h" 96 97 using namespace llvm; 98 99 static const uint64_t kShadowMask32 = 1ULL << 31; 100 static const uint64_t kShadowMask64 = 1ULL << 46; 101 static const uint64_t kOriginOffset32 = 1ULL << 30; 102 static const uint64_t kOriginOffset64 = 1ULL << 45; 103 static const unsigned kMinOriginAlignment = 4; 104 static const unsigned kShadowTLSAlignment = 8; 105 106 /// \brief Track origins of uninitialized values. 107 /// 108 /// Adds a section to MemorySanitizer report that points to the allocation 109 /// (stack or heap) the uninitialized bits came from originally. 110 static cl::opt<bool> ClTrackOrigins("msan-track-origins", 111 cl::desc("Track origins (allocation sites) of poisoned memory"), 112 cl::Hidden, cl::init(false)); 113 static cl::opt<bool> ClKeepGoing("msan-keep-going", 114 cl::desc("keep going after reporting a UMR"), 115 cl::Hidden, cl::init(false)); 116 static cl::opt<bool> ClPoisonStack("msan-poison-stack", 117 cl::desc("poison uninitialized stack variables"), 118 cl::Hidden, cl::init(true)); 119 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call", 120 cl::desc("poison uninitialized stack variables with a call"), 121 cl::Hidden, cl::init(false)); 122 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern", 123 cl::desc("poison uninitialized stack variables with the given patter"), 124 cl::Hidden, cl::init(0xff)); 125 126 static cl::opt<bool> ClHandleICmp("msan-handle-icmp", 127 cl::desc("propagate shadow through ICmpEQ and ICmpNE"), 128 cl::Hidden, cl::init(true)); 129 130 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact", 131 cl::desc("exact handling of relational integer ICmp"), 132 cl::Hidden, cl::init(false)); 133 134 static cl::opt<bool> ClStoreCleanOrigin("msan-store-clean-origin", 135 cl::desc("store origin for clean (fully initialized) values"), 136 cl::Hidden, cl::init(false)); 137 138 // This flag controls whether we check the shadow of the address 139 // operand of load or store. Such bugs are very rare, since load from 140 // a garbage address typically results in SEGV, but still happen 141 // (e.g. only lower bits of address are garbage, or the access happens 142 // early at program startup where malloc-ed memory is more likely to 143 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown. 144 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address", 145 cl::desc("report accesses through a pointer which has poisoned shadow"), 146 cl::Hidden, cl::init(true)); 147 148 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions", 149 cl::desc("print out instructions with default strict semantics"), 150 cl::Hidden, cl::init(false)); 151 152 static cl::opt<std::string> ClBlacklistFile("msan-blacklist", 153 cl::desc("File containing the list of functions where MemorySanitizer " 154 "should not report bugs"), cl::Hidden); 155 156 namespace { 157 158 /// \brief An instrumentation pass implementing detection of uninitialized 159 /// reads. 160 /// 161 /// MemorySanitizer: instrument the code in module to find 162 /// uninitialized reads. 163 class MemorySanitizer : public FunctionPass { 164 public: 165 MemorySanitizer(bool TrackOrigins = false, 166 StringRef BlacklistFile = StringRef()) 167 : FunctionPass(ID), 168 TrackOrigins(TrackOrigins || ClTrackOrigins), 169 TD(0), 170 WarningFn(0), 171 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile 172 : BlacklistFile) { } 173 const char *getPassName() const { return "MemorySanitizer"; } 174 bool runOnFunction(Function &F); 175 bool doInitialization(Module &M); 176 static char ID; // Pass identification, replacement for typeid. 177 178 private: 179 void initializeCallbacks(Module &M); 180 181 /// \brief Track origins (allocation points) of uninitialized values. 182 bool TrackOrigins; 183 184 DataLayout *TD; 185 LLVMContext *C; 186 Type *IntptrTy; 187 Type *OriginTy; 188 /// \brief Thread-local shadow storage for function parameters. 189 GlobalVariable *ParamTLS; 190 /// \brief Thread-local origin storage for function parameters. 191 GlobalVariable *ParamOriginTLS; 192 /// \brief Thread-local shadow storage for function return value. 193 GlobalVariable *RetvalTLS; 194 /// \brief Thread-local origin storage for function return value. 195 GlobalVariable *RetvalOriginTLS; 196 /// \brief Thread-local shadow storage for in-register va_arg function 197 /// parameters (x86_64-specific). 198 GlobalVariable *VAArgTLS; 199 /// \brief Thread-local shadow storage for va_arg overflow area 200 /// (x86_64-specific). 201 GlobalVariable *VAArgOverflowSizeTLS; 202 /// \brief Thread-local space used to pass origin value to the UMR reporting 203 /// function. 204 GlobalVariable *OriginTLS; 205 206 /// \brief The run-time callback to print a warning. 207 Value *WarningFn; 208 /// \brief Run-time helper that copies origin info for a memory range. 209 Value *MsanCopyOriginFn; 210 /// \brief Run-time helper that generates a new origin value for a stack 211 /// allocation. 212 Value *MsanSetAllocaOriginFn; 213 /// \brief Run-time helper that poisons stack on function entry. 214 Value *MsanPoisonStackFn; 215 /// \brief MSan runtime replacements for memmove, memcpy and memset. 216 Value *MemmoveFn, *MemcpyFn, *MemsetFn; 217 218 /// \brief Address mask used in application-to-shadow address calculation. 219 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask. 220 uint64_t ShadowMask; 221 /// \brief Offset of the origin shadow from the "normal" shadow. 222 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL 223 uint64_t OriginOffset; 224 /// \brief Branch weights for error reporting. 225 MDNode *ColdCallWeights; 226 /// \brief Branch weights for origin store. 227 MDNode *OriginStoreWeights; 228 /// \bried Path to blacklist file. 229 SmallString<64> BlacklistFile; 230 /// \brief The blacklist. 231 OwningPtr<BlackList> BL; 232 /// \brief An empty volatile inline asm that prevents callback merge. 233 InlineAsm *EmptyAsm; 234 235 friend struct MemorySanitizerVisitor; 236 friend struct VarArgAMD64Helper; 237 }; 238 } // namespace 239 240 char MemorySanitizer::ID = 0; 241 INITIALIZE_PASS(MemorySanitizer, "msan", 242 "MemorySanitizer: detects uninitialized reads.", 243 false, false) 244 245 FunctionPass *llvm::createMemorySanitizerPass(bool TrackOrigins, 246 StringRef BlacklistFile) { 247 return new MemorySanitizer(TrackOrigins, BlacklistFile); 248 } 249 250 /// \brief Create a non-const global initialized with the given string. 251 /// 252 /// Creates a writable global for Str so that we can pass it to the 253 /// run-time lib. Runtime uses first 4 bytes of the string to store the 254 /// frame ID, so the string needs to be mutable. 255 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M, 256 StringRef Str) { 257 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str); 258 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false, 259 GlobalValue::PrivateLinkage, StrConst, ""); 260 } 261 262 263 /// \brief Insert extern declaration of runtime-provided functions and globals. 264 void MemorySanitizer::initializeCallbacks(Module &M) { 265 // Only do this once. 266 if (WarningFn) 267 return; 268 269 IRBuilder<> IRB(*C); 270 // Create the callback. 271 // FIXME: this function should have "Cold" calling conv, 272 // which is not yet implemented. 273 StringRef WarningFnName = ClKeepGoing ? "__msan_warning" 274 : "__msan_warning_noreturn"; 275 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL); 276 277 MsanCopyOriginFn = M.getOrInsertFunction( 278 "__msan_copy_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), 279 IRB.getInt8PtrTy(), IntptrTy, NULL); 280 MsanSetAllocaOriginFn = M.getOrInsertFunction( 281 "__msan_set_alloca_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, 282 IRB.getInt8PtrTy(), NULL); 283 MsanPoisonStackFn = M.getOrInsertFunction( 284 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL); 285 MemmoveFn = M.getOrInsertFunction( 286 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 287 IRB.getInt8PtrTy(), IntptrTy, NULL); 288 MemcpyFn = M.getOrInsertFunction( 289 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 290 IntptrTy, NULL); 291 MemsetFn = M.getOrInsertFunction( 292 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(), 293 IntptrTy, NULL); 294 295 // Create globals. 296 RetvalTLS = new GlobalVariable( 297 M, ArrayType::get(IRB.getInt64Ty(), 8), false, 298 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0, 299 GlobalVariable::GeneralDynamicTLSModel); 300 RetvalOriginTLS = new GlobalVariable( 301 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0, 302 "__msan_retval_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel); 303 304 ParamTLS = new GlobalVariable( 305 M, ArrayType::get(IRB.getInt64Ty(), 1000), false, 306 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0, 307 GlobalVariable::GeneralDynamicTLSModel); 308 ParamOriginTLS = new GlobalVariable( 309 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage, 310 0, "__msan_param_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel); 311 312 VAArgTLS = new GlobalVariable( 313 M, ArrayType::get(IRB.getInt64Ty(), 1000), false, 314 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0, 315 GlobalVariable::GeneralDynamicTLSModel); 316 VAArgOverflowSizeTLS = new GlobalVariable( 317 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0, 318 "__msan_va_arg_overflow_size_tls", 0, 319 GlobalVariable::GeneralDynamicTLSModel); 320 OriginTLS = new GlobalVariable( 321 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0, 322 "__msan_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel); 323 324 // We insert an empty inline asm after __msan_report* to avoid callback merge. 325 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false), 326 StringRef(""), StringRef(""), 327 /*hasSideEffects=*/true); 328 } 329 330 /// \brief Module-level initialization. 331 /// 332 /// inserts a call to __msan_init to the module's constructor list. 333 bool MemorySanitizer::doInitialization(Module &M) { 334 TD = getAnalysisIfAvailable<DataLayout>(); 335 if (!TD) 336 return false; 337 BL.reset(new BlackList(BlacklistFile)); 338 C = &(M.getContext()); 339 unsigned PtrSize = TD->getPointerSizeInBits(/* AddressSpace */0); 340 switch (PtrSize) { 341 case 64: 342 ShadowMask = kShadowMask64; 343 OriginOffset = kOriginOffset64; 344 break; 345 case 32: 346 ShadowMask = kShadowMask32; 347 OriginOffset = kOriginOffset32; 348 break; 349 default: 350 report_fatal_error("unsupported pointer size"); 351 break; 352 } 353 354 IRBuilder<> IRB(*C); 355 IntptrTy = IRB.getIntPtrTy(TD); 356 OriginTy = IRB.getInt32Ty(); 357 358 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000); 359 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000); 360 361 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs. 362 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction( 363 "__msan_init", IRB.getVoidTy(), NULL)), 0); 364 365 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage, 366 IRB.getInt32(TrackOrigins), "__msan_track_origins"); 367 368 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage, 369 IRB.getInt32(ClKeepGoing), "__msan_keep_going"); 370 371 return true; 372 } 373 374 namespace { 375 376 /// \brief A helper class that handles instrumentation of VarArg 377 /// functions on a particular platform. 378 /// 379 /// Implementations are expected to insert the instrumentation 380 /// necessary to propagate argument shadow through VarArg function 381 /// calls. Visit* methods are called during an InstVisitor pass over 382 /// the function, and should avoid creating new basic blocks. A new 383 /// instance of this class is created for each instrumented function. 384 struct VarArgHelper { 385 /// \brief Visit a CallSite. 386 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0; 387 388 /// \brief Visit a va_start call. 389 virtual void visitVAStartInst(VAStartInst &I) = 0; 390 391 /// \brief Visit a va_copy call. 392 virtual void visitVACopyInst(VACopyInst &I) = 0; 393 394 /// \brief Finalize function instrumentation. 395 /// 396 /// This method is called after visiting all interesting (see above) 397 /// instructions in a function. 398 virtual void finalizeInstrumentation() = 0; 399 400 virtual ~VarArgHelper() {} 401 }; 402 403 struct MemorySanitizerVisitor; 404 405 VarArgHelper* 406 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan, 407 MemorySanitizerVisitor &Visitor); 408 409 /// This class does all the work for a given function. Store and Load 410 /// instructions store and load corresponding shadow and origin 411 /// values. Most instructions propagate shadow from arguments to their 412 /// return values. Certain instructions (most importantly, BranchInst) 413 /// test their argument shadow and print reports (with a runtime call) if it's 414 /// non-zero. 415 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> { 416 Function &F; 417 MemorySanitizer &MS; 418 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes; 419 ValueMap<Value*, Value*> ShadowMap, OriginMap; 420 bool InsertChecks; 421 bool LoadShadow; 422 OwningPtr<VarArgHelper> VAHelper; 423 424 struct ShadowOriginAndInsertPoint { 425 Instruction *Shadow; 426 Instruction *Origin; 427 Instruction *OrigIns; 428 ShadowOriginAndInsertPoint(Instruction *S, Instruction *O, Instruction *I) 429 : Shadow(S), Origin(O), OrigIns(I) { } 430 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { } 431 }; 432 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList; 433 SmallVector<Instruction*, 16> StoreList; 434 435 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS) 436 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) { 437 LoadShadow = InsertChecks = 438 !MS.BL->isIn(F) && 439 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 440 Attribute::SanitizeMemory); 441 442 DEBUG(if (!InsertChecks) 443 dbgs() << "MemorySanitizer is not inserting checks into '" 444 << F.getName() << "'\n"); 445 } 446 447 void materializeStores() { 448 for (size_t i = 0, n = StoreList.size(); i < n; i++) { 449 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]); 450 451 IRBuilder<> IRB(&I); 452 Value *Val = I.getValueOperand(); 453 Value *Addr = I.getPointerOperand(); 454 Value *Shadow = getShadow(Val); 455 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB); 456 457 StoreInst *NewSI = 458 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment()); 459 DEBUG(dbgs() << " STORE: " << *NewSI << "\n"); 460 (void)NewSI; 461 462 if (ClCheckAccessAddress) 463 insertCheck(Addr, &I); 464 465 if (MS.TrackOrigins) { 466 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment()); 467 if (ClStoreCleanOrigin || isa<StructType>(Shadow->getType())) { 468 IRB.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRB), 469 Alignment); 470 } else { 471 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB); 472 473 Constant *Cst = dyn_cast_or_null<Constant>(ConvertedShadow); 474 // TODO(eugenis): handle non-zero constant shadow by inserting an 475 // unconditional check (can not simply fail compilation as this could 476 // be in the dead code). 477 if (Cst) 478 continue; 479 480 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow, 481 getCleanShadow(ConvertedShadow), "_mscmp"); 482 Instruction *CheckTerm = 483 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false, 484 MS.OriginStoreWeights); 485 IRBuilder<> IRBNew(CheckTerm); 486 IRBNew.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRBNew), 487 Alignment); 488 } 489 } 490 } 491 } 492 493 void materializeChecks() { 494 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) { 495 Instruction *Shadow = InstrumentationList[i].Shadow; 496 Instruction *OrigIns = InstrumentationList[i].OrigIns; 497 IRBuilder<> IRB(OrigIns); 498 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n"); 499 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB); 500 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n"); 501 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow, 502 getCleanShadow(ConvertedShadow), "_mscmp"); 503 Instruction *CheckTerm = 504 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), 505 /* Unreachable */ !ClKeepGoing, 506 MS.ColdCallWeights); 507 508 IRB.SetInsertPoint(CheckTerm); 509 if (MS.TrackOrigins) { 510 Instruction *Origin = InstrumentationList[i].Origin; 511 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0), 512 MS.OriginTLS); 513 } 514 CallInst *Call = IRB.CreateCall(MS.WarningFn); 515 Call->setDebugLoc(OrigIns->getDebugLoc()); 516 IRB.CreateCall(MS.EmptyAsm); 517 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n"); 518 } 519 DEBUG(dbgs() << "DONE:\n" << F); 520 } 521 522 /// \brief Add MemorySanitizer instrumentation to a function. 523 bool runOnFunction() { 524 MS.initializeCallbacks(*F.getParent()); 525 if (!MS.TD) return false; 526 527 // In the presence of unreachable blocks, we may see Phi nodes with 528 // incoming nodes from such blocks. Since InstVisitor skips unreachable 529 // blocks, such nodes will not have any shadow value associated with them. 530 // It's easier to remove unreachable blocks than deal with missing shadow. 531 removeUnreachableBlocks(F); 532 533 // Iterate all BBs in depth-first order and create shadow instructions 534 // for all instructions (where applicable). 535 // For PHI nodes we create dummy shadow PHIs which will be finalized later. 536 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 537 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 538 BasicBlock *BB = *DI; 539 visit(*BB); 540 } 541 542 // Finalize PHI nodes. 543 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) { 544 PHINode *PN = ShadowPHINodes[i]; 545 PHINode *PNS = cast<PHINode>(getShadow(PN)); 546 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0; 547 size_t NumValues = PN->getNumIncomingValues(); 548 for (size_t v = 0; v < NumValues; v++) { 549 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v)); 550 if (PNO) 551 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v)); 552 } 553 } 554 555 VAHelper->finalizeInstrumentation(); 556 557 // Delayed instrumentation of StoreInst. 558 // This may add new checks to be inserted later. 559 materializeStores(); 560 561 // Insert shadow value checks. 562 materializeChecks(); 563 564 return true; 565 } 566 567 /// \brief Compute the shadow type that corresponds to a given Value. 568 Type *getShadowTy(Value *V) { 569 return getShadowTy(V->getType()); 570 } 571 572 /// \brief Compute the shadow type that corresponds to a given Type. 573 Type *getShadowTy(Type *OrigTy) { 574 if (!OrigTy->isSized()) { 575 return 0; 576 } 577 // For integer type, shadow is the same as the original type. 578 // This may return weird-sized types like i1. 579 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy)) 580 return IT; 581 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) { 582 uint32_t EltSize = MS.TD->getTypeSizeInBits(VT->getElementType()); 583 return VectorType::get(IntegerType::get(*MS.C, EltSize), 584 VT->getNumElements()); 585 } 586 if (StructType *ST = dyn_cast<StructType>(OrigTy)) { 587 SmallVector<Type*, 4> Elements; 588 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++) 589 Elements.push_back(getShadowTy(ST->getElementType(i))); 590 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked()); 591 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n"); 592 return Res; 593 } 594 uint32_t TypeSize = MS.TD->getTypeSizeInBits(OrigTy); 595 return IntegerType::get(*MS.C, TypeSize); 596 } 597 598 /// \brief Flatten a vector type. 599 Type *getShadowTyNoVec(Type *ty) { 600 if (VectorType *vt = dyn_cast<VectorType>(ty)) 601 return IntegerType::get(*MS.C, vt->getBitWidth()); 602 return ty; 603 } 604 605 /// \brief Convert a shadow value to it's flattened variant. 606 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) { 607 Type *Ty = V->getType(); 608 Type *NoVecTy = getShadowTyNoVec(Ty); 609 if (Ty == NoVecTy) return V; 610 return IRB.CreateBitCast(V, NoVecTy); 611 } 612 613 /// \brief Compute the shadow address that corresponds to a given application 614 /// address. 615 /// 616 /// Shadow = Addr & ~ShadowMask. 617 Value *getShadowPtr(Value *Addr, Type *ShadowTy, 618 IRBuilder<> &IRB) { 619 Value *ShadowLong = 620 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy), 621 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask)); 622 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0)); 623 } 624 625 /// \brief Compute the origin address that corresponds to a given application 626 /// address. 627 /// 628 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL 629 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) { 630 Value *ShadowLong = 631 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy), 632 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask)); 633 Value *Add = 634 IRB.CreateAdd(ShadowLong, 635 ConstantInt::get(MS.IntptrTy, MS.OriginOffset)); 636 Value *SecondAnd = 637 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL)); 638 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0)); 639 } 640 641 /// \brief Compute the shadow address for a given function argument. 642 /// 643 /// Shadow = ParamTLS+ArgOffset. 644 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB, 645 int ArgOffset) { 646 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy); 647 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset)); 648 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0), 649 "_msarg"); 650 } 651 652 /// \brief Compute the origin address for a given function argument. 653 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB, 654 int ArgOffset) { 655 if (!MS.TrackOrigins) return 0; 656 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy); 657 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset)); 658 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0), 659 "_msarg_o"); 660 } 661 662 /// \brief Compute the shadow address for a retval. 663 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) { 664 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy); 665 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0), 666 "_msret"); 667 } 668 669 /// \brief Compute the origin address for a retval. 670 Value *getOriginPtrForRetval(IRBuilder<> &IRB) { 671 // We keep a single origin for the entire retval. Might be too optimistic. 672 return MS.RetvalOriginTLS; 673 } 674 675 /// \brief Set SV to be the shadow value for V. 676 void setShadow(Value *V, Value *SV) { 677 assert(!ShadowMap.count(V) && "Values may only have one shadow"); 678 ShadowMap[V] = SV; 679 } 680 681 /// \brief Set Origin to be the origin value for V. 682 void setOrigin(Value *V, Value *Origin) { 683 if (!MS.TrackOrigins) return; 684 assert(!OriginMap.count(V) && "Values may only have one origin"); 685 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n"); 686 OriginMap[V] = Origin; 687 } 688 689 /// \brief Create a clean shadow value for a given value. 690 /// 691 /// Clean shadow (all zeroes) means all bits of the value are defined 692 /// (initialized). 693 Value *getCleanShadow(Value *V) { 694 Type *ShadowTy = getShadowTy(V); 695 if (!ShadowTy) 696 return 0; 697 return Constant::getNullValue(ShadowTy); 698 } 699 700 /// \brief Create a dirty shadow of a given shadow type. 701 Constant *getPoisonedShadow(Type *ShadowTy) { 702 assert(ShadowTy); 703 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) 704 return Constant::getAllOnesValue(ShadowTy); 705 StructType *ST = cast<StructType>(ShadowTy); 706 SmallVector<Constant *, 4> Vals; 707 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++) 708 Vals.push_back(getPoisonedShadow(ST->getElementType(i))); 709 return ConstantStruct::get(ST, Vals); 710 } 711 712 /// \brief Create a clean (zero) origin. 713 Value *getCleanOrigin() { 714 return Constant::getNullValue(MS.OriginTy); 715 } 716 717 /// \brief Get the shadow value for a given Value. 718 /// 719 /// This function either returns the value set earlier with setShadow, 720 /// or extracts if from ParamTLS (for function arguments). 721 Value *getShadow(Value *V) { 722 if (Instruction *I = dyn_cast<Instruction>(V)) { 723 // For instructions the shadow is already stored in the map. 724 Value *Shadow = ShadowMap[V]; 725 if (!Shadow) { 726 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent())); 727 (void)I; 728 assert(Shadow && "No shadow for a value"); 729 } 730 return Shadow; 731 } 732 if (UndefValue *U = dyn_cast<UndefValue>(V)) { 733 Value *AllOnes = getPoisonedShadow(getShadowTy(V)); 734 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n"); 735 (void)U; 736 return AllOnes; 737 } 738 if (Argument *A = dyn_cast<Argument>(V)) { 739 // For arguments we compute the shadow on demand and store it in the map. 740 Value **ShadowPtr = &ShadowMap[V]; 741 if (*ShadowPtr) 742 return *ShadowPtr; 743 Function *F = A->getParent(); 744 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI()); 745 unsigned ArgOffset = 0; 746 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end(); 747 AI != AE; ++AI) { 748 if (!AI->getType()->isSized()) { 749 DEBUG(dbgs() << "Arg is not sized\n"); 750 continue; 751 } 752 unsigned Size = AI->hasByValAttr() 753 ? MS.TD->getTypeAllocSize(AI->getType()->getPointerElementType()) 754 : MS.TD->getTypeAllocSize(AI->getType()); 755 if (A == AI) { 756 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset); 757 if (AI->hasByValAttr()) { 758 // ByVal pointer itself has clean shadow. We copy the actual 759 // argument shadow to the underlying memory. 760 Value *Cpy = EntryIRB.CreateMemCpy( 761 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), 762 Base, Size, AI->getParamAlignment()); 763 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n"); 764 (void)Cpy; 765 *ShadowPtr = getCleanShadow(V); 766 } else { 767 *ShadowPtr = EntryIRB.CreateLoad(Base); 768 } 769 DEBUG(dbgs() << " ARG: " << *AI << " ==> " << 770 **ShadowPtr << "\n"); 771 if (MS.TrackOrigins) { 772 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset); 773 setOrigin(A, EntryIRB.CreateLoad(OriginPtr)); 774 } 775 } 776 ArgOffset += DataLayout::RoundUpAlignment(Size, 8); 777 } 778 assert(*ShadowPtr && "Could not find shadow for an argument"); 779 return *ShadowPtr; 780 } 781 // For everything else the shadow is zero. 782 return getCleanShadow(V); 783 } 784 785 /// \brief Get the shadow for i-th argument of the instruction I. 786 Value *getShadow(Instruction *I, int i) { 787 return getShadow(I->getOperand(i)); 788 } 789 790 /// \brief Get the origin for a value. 791 Value *getOrigin(Value *V) { 792 if (!MS.TrackOrigins) return 0; 793 if (isa<Instruction>(V) || isa<Argument>(V)) { 794 Value *Origin = OriginMap[V]; 795 if (!Origin) { 796 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n"); 797 Origin = getCleanOrigin(); 798 } 799 return Origin; 800 } 801 return getCleanOrigin(); 802 } 803 804 /// \brief Get the origin for i-th argument of the instruction I. 805 Value *getOrigin(Instruction *I, int i) { 806 return getOrigin(I->getOperand(i)); 807 } 808 809 /// \brief Remember the place where a shadow check should be inserted. 810 /// 811 /// This location will be later instrumented with a check that will print a 812 /// UMR warning in runtime if the value is not fully defined. 813 void insertCheck(Value *Val, Instruction *OrigIns) { 814 assert(Val); 815 if (!InsertChecks) return; 816 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val)); 817 if (!Shadow) return; 818 #ifndef NDEBUG 819 Type *ShadowTy = Shadow->getType(); 820 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) && 821 "Can only insert checks for integer and vector shadow types"); 822 #endif 823 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val)); 824 InstrumentationList.push_back( 825 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns)); 826 } 827 828 // ------------------- Visitors. 829 830 /// \brief Instrument LoadInst 831 /// 832 /// Loads the corresponding shadow and (optionally) origin. 833 /// Optionally, checks that the load address is fully defined. 834 void visitLoadInst(LoadInst &I) { 835 assert(I.getType()->isSized() && "Load type must have size"); 836 IRBuilder<> IRB(&I); 837 Type *ShadowTy = getShadowTy(&I); 838 Value *Addr = I.getPointerOperand(); 839 if (LoadShadow) { 840 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB); 841 setShadow(&I, 842 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld")); 843 } else { 844 setShadow(&I, getCleanShadow(&I)); 845 } 846 847 if (ClCheckAccessAddress) 848 insertCheck(I.getPointerOperand(), &I); 849 850 if (MS.TrackOrigins) { 851 if (LoadShadow) { 852 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment()); 853 setOrigin(&I, 854 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment)); 855 } else { 856 setOrigin(&I, getCleanOrigin()); 857 } 858 } 859 } 860 861 /// \brief Instrument StoreInst 862 /// 863 /// Stores the corresponding shadow and (optionally) origin. 864 /// Optionally, checks that the store address is fully defined. 865 void visitStoreInst(StoreInst &I) { 866 StoreList.push_back(&I); 867 } 868 869 // Vector manipulation. 870 void visitExtractElementInst(ExtractElementInst &I) { 871 insertCheck(I.getOperand(1), &I); 872 IRBuilder<> IRB(&I); 873 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1), 874 "_msprop")); 875 setOrigin(&I, getOrigin(&I, 0)); 876 } 877 878 void visitInsertElementInst(InsertElementInst &I) { 879 insertCheck(I.getOperand(2), &I); 880 IRBuilder<> IRB(&I); 881 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1), 882 I.getOperand(2), "_msprop")); 883 setOriginForNaryOp(I); 884 } 885 886 void visitShuffleVectorInst(ShuffleVectorInst &I) { 887 insertCheck(I.getOperand(2), &I); 888 IRBuilder<> IRB(&I); 889 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1), 890 I.getOperand(2), "_msprop")); 891 setOriginForNaryOp(I); 892 } 893 894 // Casts. 895 void visitSExtInst(SExtInst &I) { 896 IRBuilder<> IRB(&I); 897 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop")); 898 setOrigin(&I, getOrigin(&I, 0)); 899 } 900 901 void visitZExtInst(ZExtInst &I) { 902 IRBuilder<> IRB(&I); 903 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop")); 904 setOrigin(&I, getOrigin(&I, 0)); 905 } 906 907 void visitTruncInst(TruncInst &I) { 908 IRBuilder<> IRB(&I); 909 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop")); 910 setOrigin(&I, getOrigin(&I, 0)); 911 } 912 913 void visitBitCastInst(BitCastInst &I) { 914 IRBuilder<> IRB(&I); 915 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I))); 916 setOrigin(&I, getOrigin(&I, 0)); 917 } 918 919 void visitPtrToIntInst(PtrToIntInst &I) { 920 IRBuilder<> IRB(&I); 921 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false, 922 "_msprop_ptrtoint")); 923 setOrigin(&I, getOrigin(&I, 0)); 924 } 925 926 void visitIntToPtrInst(IntToPtrInst &I) { 927 IRBuilder<> IRB(&I); 928 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false, 929 "_msprop_inttoptr")); 930 setOrigin(&I, getOrigin(&I, 0)); 931 } 932 933 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); } 934 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); } 935 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); } 936 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); } 937 void visitFPExtInst(CastInst& I) { handleShadowOr(I); } 938 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); } 939 940 /// \brief Propagate shadow for bitwise AND. 941 /// 942 /// This code is exact, i.e. if, for example, a bit in the left argument 943 /// is defined and 0, then neither the value not definedness of the 944 /// corresponding bit in B don't affect the resulting shadow. 945 void visitAnd(BinaryOperator &I) { 946 IRBuilder<> IRB(&I); 947 // "And" of 0 and a poisoned value results in unpoisoned value. 948 // 1&1 => 1; 0&1 => 0; p&1 => p; 949 // 1&0 => 0; 0&0 => 0; p&0 => 0; 950 // 1&p => p; 0&p => 0; p&p => p; 951 // S = (S1 & S2) | (V1 & S2) | (S1 & V2) 952 Value *S1 = getShadow(&I, 0); 953 Value *S2 = getShadow(&I, 1); 954 Value *V1 = I.getOperand(0); 955 Value *V2 = I.getOperand(1); 956 if (V1->getType() != S1->getType()) { 957 V1 = IRB.CreateIntCast(V1, S1->getType(), false); 958 V2 = IRB.CreateIntCast(V2, S2->getType(), false); 959 } 960 Value *S1S2 = IRB.CreateAnd(S1, S2); 961 Value *V1S2 = IRB.CreateAnd(V1, S2); 962 Value *S1V2 = IRB.CreateAnd(S1, V2); 963 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2))); 964 setOriginForNaryOp(I); 965 } 966 967 void visitOr(BinaryOperator &I) { 968 IRBuilder<> IRB(&I); 969 // "Or" of 1 and a poisoned value results in unpoisoned value. 970 // 1|1 => 1; 0|1 => 1; p|1 => 1; 971 // 1|0 => 1; 0|0 => 0; p|0 => p; 972 // 1|p => 1; 0|p => p; p|p => p; 973 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2) 974 Value *S1 = getShadow(&I, 0); 975 Value *S2 = getShadow(&I, 1); 976 Value *V1 = IRB.CreateNot(I.getOperand(0)); 977 Value *V2 = IRB.CreateNot(I.getOperand(1)); 978 if (V1->getType() != S1->getType()) { 979 V1 = IRB.CreateIntCast(V1, S1->getType(), false); 980 V2 = IRB.CreateIntCast(V2, S2->getType(), false); 981 } 982 Value *S1S2 = IRB.CreateAnd(S1, S2); 983 Value *V1S2 = IRB.CreateAnd(V1, S2); 984 Value *S1V2 = IRB.CreateAnd(S1, V2); 985 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2))); 986 setOriginForNaryOp(I); 987 } 988 989 /// \brief Default propagation of shadow and/or origin. 990 /// 991 /// This class implements the general case of shadow propagation, used in all 992 /// cases where we don't know and/or don't care about what the operation 993 /// actually does. It converts all input shadow values to a common type 994 /// (extending or truncating as necessary), and bitwise OR's them. 995 /// 996 /// This is much cheaper than inserting checks (i.e. requiring inputs to be 997 /// fully initialized), and less prone to false positives. 998 /// 999 /// This class also implements the general case of origin propagation. For a 1000 /// Nary operation, result origin is set to the origin of an argument that is 1001 /// not entirely initialized. If there is more than one such arguments, the 1002 /// rightmost of them is picked. It does not matter which one is picked if all 1003 /// arguments are initialized. 1004 template <bool CombineShadow> 1005 class Combiner { 1006 Value *Shadow; 1007 Value *Origin; 1008 IRBuilder<> &IRB; 1009 MemorySanitizerVisitor *MSV; 1010 1011 public: 1012 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) : 1013 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {} 1014 1015 /// \brief Add a pair of shadow and origin values to the mix. 1016 Combiner &Add(Value *OpShadow, Value *OpOrigin) { 1017 if (CombineShadow) { 1018 assert(OpShadow); 1019 if (!Shadow) 1020 Shadow = OpShadow; 1021 else { 1022 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType()); 1023 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop"); 1024 } 1025 } 1026 1027 if (MSV->MS.TrackOrigins) { 1028 assert(OpOrigin); 1029 if (!Origin) { 1030 Origin = OpOrigin; 1031 } else { 1032 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB); 1033 Value *Cond = IRB.CreateICmpNE(FlatShadow, 1034 MSV->getCleanShadow(FlatShadow)); 1035 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin); 1036 } 1037 } 1038 return *this; 1039 } 1040 1041 /// \brief Add an application value to the mix. 1042 Combiner &Add(Value *V) { 1043 Value *OpShadow = MSV->getShadow(V); 1044 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : 0; 1045 return Add(OpShadow, OpOrigin); 1046 } 1047 1048 /// \brief Set the current combined values as the given instruction's shadow 1049 /// and origin. 1050 void Done(Instruction *I) { 1051 if (CombineShadow) { 1052 assert(Shadow); 1053 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I)); 1054 MSV->setShadow(I, Shadow); 1055 } 1056 if (MSV->MS.TrackOrigins) { 1057 assert(Origin); 1058 MSV->setOrigin(I, Origin); 1059 } 1060 } 1061 }; 1062 1063 typedef Combiner<true> ShadowAndOriginCombiner; 1064 typedef Combiner<false> OriginCombiner; 1065 1066 /// \brief Propagate origin for arbitrary operation. 1067 void setOriginForNaryOp(Instruction &I) { 1068 if (!MS.TrackOrigins) return; 1069 IRBuilder<> IRB(&I); 1070 OriginCombiner OC(this, IRB); 1071 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI) 1072 OC.Add(OI->get()); 1073 OC.Done(&I); 1074 } 1075 1076 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) { 1077 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) && 1078 "Vector of pointers is not a valid shadow type"); 1079 return Ty->isVectorTy() ? 1080 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() : 1081 Ty->getPrimitiveSizeInBits(); 1082 } 1083 1084 /// \brief Cast between two shadow types, extending or truncating as 1085 /// necessary. 1086 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy) { 1087 Type *srcTy = V->getType(); 1088 if (dstTy->isIntegerTy() && srcTy->isIntegerTy()) 1089 return IRB.CreateIntCast(V, dstTy, false); 1090 if (dstTy->isVectorTy() && srcTy->isVectorTy() && 1091 dstTy->getVectorNumElements() == srcTy->getVectorNumElements()) 1092 return IRB.CreateIntCast(V, dstTy, false); 1093 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy); 1094 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy); 1095 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits)); 1096 Value *V2 = 1097 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), false); 1098 return IRB.CreateBitCast(V2, dstTy); 1099 // TODO: handle struct types. 1100 } 1101 1102 /// \brief Propagate shadow for arbitrary operation. 1103 void handleShadowOr(Instruction &I) { 1104 IRBuilder<> IRB(&I); 1105 ShadowAndOriginCombiner SC(this, IRB); 1106 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI) 1107 SC.Add(OI->get()); 1108 SC.Done(&I); 1109 } 1110 1111 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); } 1112 void visitFSub(BinaryOperator &I) { handleShadowOr(I); } 1113 void visitFMul(BinaryOperator &I) { handleShadowOr(I); } 1114 void visitAdd(BinaryOperator &I) { handleShadowOr(I); } 1115 void visitSub(BinaryOperator &I) { handleShadowOr(I); } 1116 void visitXor(BinaryOperator &I) { handleShadowOr(I); } 1117 void visitMul(BinaryOperator &I) { handleShadowOr(I); } 1118 1119 void handleDiv(Instruction &I) { 1120 IRBuilder<> IRB(&I); 1121 // Strict on the second argument. 1122 insertCheck(I.getOperand(1), &I); 1123 setShadow(&I, getShadow(&I, 0)); 1124 setOrigin(&I, getOrigin(&I, 0)); 1125 } 1126 1127 void visitUDiv(BinaryOperator &I) { handleDiv(I); } 1128 void visitSDiv(BinaryOperator &I) { handleDiv(I); } 1129 void visitFDiv(BinaryOperator &I) { handleDiv(I); } 1130 void visitURem(BinaryOperator &I) { handleDiv(I); } 1131 void visitSRem(BinaryOperator &I) { handleDiv(I); } 1132 void visitFRem(BinaryOperator &I) { handleDiv(I); } 1133 1134 /// \brief Instrument == and != comparisons. 1135 /// 1136 /// Sometimes the comparison result is known even if some of the bits of the 1137 /// arguments are not. 1138 void handleEqualityComparison(ICmpInst &I) { 1139 IRBuilder<> IRB(&I); 1140 Value *A = I.getOperand(0); 1141 Value *B = I.getOperand(1); 1142 Value *Sa = getShadow(A); 1143 Value *Sb = getShadow(B); 1144 1145 // Get rid of pointers and vectors of pointers. 1146 // For ints (and vectors of ints), types of A and Sa match, 1147 // and this is a no-op. 1148 A = IRB.CreatePointerCast(A, Sa->getType()); 1149 B = IRB.CreatePointerCast(B, Sb->getType()); 1150 1151 // A == B <==> (C = A^B) == 0 1152 // A != B <==> (C = A^B) != 0 1153 // Sc = Sa | Sb 1154 Value *C = IRB.CreateXor(A, B); 1155 Value *Sc = IRB.CreateOr(Sa, Sb); 1156 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now) 1157 // Result is defined if one of the following is true 1158 // * there is a defined 1 bit in C 1159 // * C is fully defined 1160 // Si = !(C & ~Sc) && Sc 1161 Value *Zero = Constant::getNullValue(Sc->getType()); 1162 Value *MinusOne = Constant::getAllOnesValue(Sc->getType()); 1163 Value *Si = 1164 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero), 1165 IRB.CreateICmpEQ( 1166 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero)); 1167 Si->setName("_msprop_icmp"); 1168 setShadow(&I, Si); 1169 setOriginForNaryOp(I); 1170 } 1171 1172 /// \brief Build the lowest possible value of V, taking into account V's 1173 /// uninitialized bits. 1174 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa, 1175 bool isSigned) { 1176 if (isSigned) { 1177 // Split shadow into sign bit and other bits. 1178 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1); 1179 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits); 1180 // Maximise the undefined shadow bit, minimize other undefined bits. 1181 return 1182 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit); 1183 } else { 1184 // Minimize undefined bits. 1185 return IRB.CreateAnd(A, IRB.CreateNot(Sa)); 1186 } 1187 } 1188 1189 /// \brief Build the highest possible value of V, taking into account V's 1190 /// uninitialized bits. 1191 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa, 1192 bool isSigned) { 1193 if (isSigned) { 1194 // Split shadow into sign bit and other bits. 1195 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1); 1196 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits); 1197 // Minimise the undefined shadow bit, maximise other undefined bits. 1198 return 1199 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits); 1200 } else { 1201 // Maximize undefined bits. 1202 return IRB.CreateOr(A, Sa); 1203 } 1204 } 1205 1206 /// \brief Instrument relational comparisons. 1207 /// 1208 /// This function does exact shadow propagation for all relational 1209 /// comparisons of integers, pointers and vectors of those. 1210 /// FIXME: output seems suboptimal when one of the operands is a constant 1211 void handleRelationalComparisonExact(ICmpInst &I) { 1212 IRBuilder<> IRB(&I); 1213 Value *A = I.getOperand(0); 1214 Value *B = I.getOperand(1); 1215 Value *Sa = getShadow(A); 1216 Value *Sb = getShadow(B); 1217 1218 // Get rid of pointers and vectors of pointers. 1219 // For ints (and vectors of ints), types of A and Sa match, 1220 // and this is a no-op. 1221 A = IRB.CreatePointerCast(A, Sa->getType()); 1222 B = IRB.CreatePointerCast(B, Sb->getType()); 1223 1224 // Let [a0, a1] be the interval of possible values of A, taking into account 1225 // its undefined bits. Let [b0, b1] be the interval of possible values of B. 1226 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0). 1227 bool IsSigned = I.isSigned(); 1228 Value *S1 = IRB.CreateICmp(I.getPredicate(), 1229 getLowestPossibleValue(IRB, A, Sa, IsSigned), 1230 getHighestPossibleValue(IRB, B, Sb, IsSigned)); 1231 Value *S2 = IRB.CreateICmp(I.getPredicate(), 1232 getHighestPossibleValue(IRB, A, Sa, IsSigned), 1233 getLowestPossibleValue(IRB, B, Sb, IsSigned)); 1234 Value *Si = IRB.CreateXor(S1, S2); 1235 setShadow(&I, Si); 1236 setOriginForNaryOp(I); 1237 } 1238 1239 /// \brief Instrument signed relational comparisons. 1240 /// 1241 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by 1242 /// propagating the highest bit of the shadow. Everything else is delegated 1243 /// to handleShadowOr(). 1244 void handleSignedRelationalComparison(ICmpInst &I) { 1245 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0)); 1246 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1)); 1247 Value* op = NULL; 1248 CmpInst::Predicate pre = I.getPredicate(); 1249 if (constOp0 && constOp0->isNullValue() && 1250 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) { 1251 op = I.getOperand(1); 1252 } else if (constOp1 && constOp1->isNullValue() && 1253 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) { 1254 op = I.getOperand(0); 1255 } 1256 if (op) { 1257 IRBuilder<> IRB(&I); 1258 Value* Shadow = 1259 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt"); 1260 setShadow(&I, Shadow); 1261 setOrigin(&I, getOrigin(op)); 1262 } else { 1263 handleShadowOr(I); 1264 } 1265 } 1266 1267 void visitICmpInst(ICmpInst &I) { 1268 if (!ClHandleICmp) { 1269 handleShadowOr(I); 1270 return; 1271 } 1272 if (I.isEquality()) { 1273 handleEqualityComparison(I); 1274 return; 1275 } 1276 1277 assert(I.isRelational()); 1278 if (ClHandleICmpExact) { 1279 handleRelationalComparisonExact(I); 1280 return; 1281 } 1282 if (I.isSigned()) { 1283 handleSignedRelationalComparison(I); 1284 return; 1285 } 1286 1287 assert(I.isUnsigned()); 1288 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) { 1289 handleRelationalComparisonExact(I); 1290 return; 1291 } 1292 1293 handleShadowOr(I); 1294 } 1295 1296 void visitFCmpInst(FCmpInst &I) { 1297 handleShadowOr(I); 1298 } 1299 1300 void handleShift(BinaryOperator &I) { 1301 IRBuilder<> IRB(&I); 1302 // If any of the S2 bits are poisoned, the whole thing is poisoned. 1303 // Otherwise perform the same shift on S1. 1304 Value *S1 = getShadow(&I, 0); 1305 Value *S2 = getShadow(&I, 1); 1306 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)), 1307 S2->getType()); 1308 Value *V2 = I.getOperand(1); 1309 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2); 1310 setShadow(&I, IRB.CreateOr(Shift, S2Conv)); 1311 setOriginForNaryOp(I); 1312 } 1313 1314 void visitShl(BinaryOperator &I) { handleShift(I); } 1315 void visitAShr(BinaryOperator &I) { handleShift(I); } 1316 void visitLShr(BinaryOperator &I) { handleShift(I); } 1317 1318 /// \brief Instrument llvm.memmove 1319 /// 1320 /// At this point we don't know if llvm.memmove will be inlined or not. 1321 /// If we don't instrument it and it gets inlined, 1322 /// our interceptor will not kick in and we will lose the memmove. 1323 /// If we instrument the call here, but it does not get inlined, 1324 /// we will memove the shadow twice: which is bad in case 1325 /// of overlapping regions. So, we simply lower the intrinsic to a call. 1326 /// 1327 /// Similar situation exists for memcpy and memset. 1328 void visitMemMoveInst(MemMoveInst &I) { 1329 IRBuilder<> IRB(&I); 1330 IRB.CreateCall3( 1331 MS.MemmoveFn, 1332 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()), 1333 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()), 1334 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)); 1335 I.eraseFromParent(); 1336 } 1337 1338 // Similar to memmove: avoid copying shadow twice. 1339 // This is somewhat unfortunate as it may slowdown small constant memcpys. 1340 // FIXME: consider doing manual inline for small constant sizes and proper 1341 // alignment. 1342 void visitMemCpyInst(MemCpyInst &I) { 1343 IRBuilder<> IRB(&I); 1344 IRB.CreateCall3( 1345 MS.MemcpyFn, 1346 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()), 1347 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()), 1348 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)); 1349 I.eraseFromParent(); 1350 } 1351 1352 // Same as memcpy. 1353 void visitMemSetInst(MemSetInst &I) { 1354 IRBuilder<> IRB(&I); 1355 IRB.CreateCall3( 1356 MS.MemsetFn, 1357 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()), 1358 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false), 1359 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)); 1360 I.eraseFromParent(); 1361 } 1362 1363 void visitVAStartInst(VAStartInst &I) { 1364 VAHelper->visitVAStartInst(I); 1365 } 1366 1367 void visitVACopyInst(VACopyInst &I) { 1368 VAHelper->visitVACopyInst(I); 1369 } 1370 1371 enum IntrinsicKind { 1372 IK_DoesNotAccessMemory, 1373 IK_OnlyReadsMemory, 1374 IK_WritesMemory 1375 }; 1376 1377 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) { 1378 const int DoesNotAccessMemory = IK_DoesNotAccessMemory; 1379 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory; 1380 const int OnlyReadsMemory = IK_OnlyReadsMemory; 1381 const int OnlyAccessesArgumentPointees = IK_WritesMemory; 1382 const int UnknownModRefBehavior = IK_WritesMemory; 1383 #define GET_INTRINSIC_MODREF_BEHAVIOR 1384 #define ModRefBehavior IntrinsicKind 1385 #include "llvm/IR/Intrinsics.gen" 1386 #undef ModRefBehavior 1387 #undef GET_INTRINSIC_MODREF_BEHAVIOR 1388 } 1389 1390 /// \brief Handle vector store-like intrinsics. 1391 /// 1392 /// Instrument intrinsics that look like a simple SIMD store: writes memory, 1393 /// has 1 pointer argument and 1 vector argument, returns void. 1394 bool handleVectorStoreIntrinsic(IntrinsicInst &I) { 1395 IRBuilder<> IRB(&I); 1396 Value* Addr = I.getArgOperand(0); 1397 Value *Shadow = getShadow(&I, 1); 1398 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB); 1399 1400 // We don't know the pointer alignment (could be unaligned SSE store!). 1401 // Have to assume to worst case. 1402 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1); 1403 1404 if (ClCheckAccessAddress) 1405 insertCheck(Addr, &I); 1406 1407 // FIXME: use ClStoreCleanOrigin 1408 // FIXME: factor out common code from materializeStores 1409 if (MS.TrackOrigins) 1410 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB)); 1411 return true; 1412 } 1413 1414 /// \brief Handle vector load-like intrinsics. 1415 /// 1416 /// Instrument intrinsics that look like a simple SIMD load: reads memory, 1417 /// has 1 pointer argument, returns a vector. 1418 bool handleVectorLoadIntrinsic(IntrinsicInst &I) { 1419 IRBuilder<> IRB(&I); 1420 Value *Addr = I.getArgOperand(0); 1421 1422 Type *ShadowTy = getShadowTy(&I); 1423 if (LoadShadow) { 1424 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB); 1425 // We don't know the pointer alignment (could be unaligned SSE load!). 1426 // Have to assume to worst case. 1427 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld")); 1428 } else { 1429 setShadow(&I, getCleanShadow(&I)); 1430 } 1431 1432 1433 if (ClCheckAccessAddress) 1434 insertCheck(Addr, &I); 1435 1436 if (MS.TrackOrigins) { 1437 if (LoadShadow) 1438 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB))); 1439 else 1440 setOrigin(&I, getCleanOrigin()); 1441 } 1442 return true; 1443 } 1444 1445 /// \brief Handle (SIMD arithmetic)-like intrinsics. 1446 /// 1447 /// Instrument intrinsics with any number of arguments of the same type, 1448 /// equal to the return type. The type should be simple (no aggregates or 1449 /// pointers; vectors are fine). 1450 /// Caller guarantees that this intrinsic does not access memory. 1451 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) { 1452 Type *RetTy = I.getType(); 1453 if (!(RetTy->isIntOrIntVectorTy() || 1454 RetTy->isFPOrFPVectorTy() || 1455 RetTy->isX86_MMXTy())) 1456 return false; 1457 1458 unsigned NumArgOperands = I.getNumArgOperands(); 1459 1460 for (unsigned i = 0; i < NumArgOperands; ++i) { 1461 Type *Ty = I.getArgOperand(i)->getType(); 1462 if (Ty != RetTy) 1463 return false; 1464 } 1465 1466 IRBuilder<> IRB(&I); 1467 ShadowAndOriginCombiner SC(this, IRB); 1468 for (unsigned i = 0; i < NumArgOperands; ++i) 1469 SC.Add(I.getArgOperand(i)); 1470 SC.Done(&I); 1471 1472 return true; 1473 } 1474 1475 /// \brief Heuristically instrument unknown intrinsics. 1476 /// 1477 /// The main purpose of this code is to do something reasonable with all 1478 /// random intrinsics we might encounter, most importantly - SIMD intrinsics. 1479 /// We recognize several classes of intrinsics by their argument types and 1480 /// ModRefBehaviour and apply special intrumentation when we are reasonably 1481 /// sure that we know what the intrinsic does. 1482 /// 1483 /// We special-case intrinsics where this approach fails. See llvm.bswap 1484 /// handling as an example of that. 1485 bool handleUnknownIntrinsic(IntrinsicInst &I) { 1486 unsigned NumArgOperands = I.getNumArgOperands(); 1487 if (NumArgOperands == 0) 1488 return false; 1489 1490 Intrinsic::ID iid = I.getIntrinsicID(); 1491 IntrinsicKind IK = getIntrinsicKind(iid); 1492 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory; 1493 bool WritesMemory = IK == IK_WritesMemory; 1494 assert(!(OnlyReadsMemory && WritesMemory)); 1495 1496 if (NumArgOperands == 2 && 1497 I.getArgOperand(0)->getType()->isPointerTy() && 1498 I.getArgOperand(1)->getType()->isVectorTy() && 1499 I.getType()->isVoidTy() && 1500 WritesMemory) { 1501 // This looks like a vector store. 1502 return handleVectorStoreIntrinsic(I); 1503 } 1504 1505 if (NumArgOperands == 1 && 1506 I.getArgOperand(0)->getType()->isPointerTy() && 1507 I.getType()->isVectorTy() && 1508 OnlyReadsMemory) { 1509 // This looks like a vector load. 1510 return handleVectorLoadIntrinsic(I); 1511 } 1512 1513 if (!OnlyReadsMemory && !WritesMemory) 1514 if (maybeHandleSimpleNomemIntrinsic(I)) 1515 return true; 1516 1517 // FIXME: detect and handle SSE maskstore/maskload 1518 return false; 1519 } 1520 1521 void handleBswap(IntrinsicInst &I) { 1522 IRBuilder<> IRB(&I); 1523 Value *Op = I.getArgOperand(0); 1524 Type *OpType = Op->getType(); 1525 Function *BswapFunc = Intrinsic::getDeclaration( 1526 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1)); 1527 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op))); 1528 setOrigin(&I, getOrigin(Op)); 1529 } 1530 1531 void visitIntrinsicInst(IntrinsicInst &I) { 1532 switch (I.getIntrinsicID()) { 1533 case llvm::Intrinsic::bswap: 1534 handleBswap(I); 1535 break; 1536 default: 1537 if (!handleUnknownIntrinsic(I)) 1538 visitInstruction(I); 1539 break; 1540 } 1541 } 1542 1543 void visitCallSite(CallSite CS) { 1544 Instruction &I = *CS.getInstruction(); 1545 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite"); 1546 if (CS.isCall()) { 1547 CallInst *Call = cast<CallInst>(&I); 1548 1549 // For inline asm, do the usual thing: check argument shadow and mark all 1550 // outputs as clean. Note that any side effects of the inline asm that are 1551 // not immediately visible in its constraints are not handled. 1552 if (Call->isInlineAsm()) { 1553 visitInstruction(I); 1554 return; 1555 } 1556 1557 // Allow only tail calls with the same types, otherwise 1558 // we may have a false positive: shadow for a non-void RetVal 1559 // will get propagated to a void RetVal. 1560 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType()) 1561 Call->setTailCall(false); 1562 1563 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere"); 1564 1565 // We are going to insert code that relies on the fact that the callee 1566 // will become a non-readonly function after it is instrumented by us. To 1567 // prevent this code from being optimized out, mark that function 1568 // non-readonly in advance. 1569 if (Function *Func = Call->getCalledFunction()) { 1570 // Clear out readonly/readnone attributes. 1571 AttrBuilder B; 1572 B.addAttribute(Attribute::ReadOnly) 1573 .addAttribute(Attribute::ReadNone); 1574 Func->removeAttributes(AttributeSet::FunctionIndex, 1575 AttributeSet::get(Func->getContext(), 1576 AttributeSet::FunctionIndex, 1577 B)); 1578 } 1579 } 1580 IRBuilder<> IRB(&I); 1581 unsigned ArgOffset = 0; 1582 DEBUG(dbgs() << " CallSite: " << I << "\n"); 1583 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end(); 1584 ArgIt != End; ++ArgIt) { 1585 Value *A = *ArgIt; 1586 unsigned i = ArgIt - CS.arg_begin(); 1587 if (!A->getType()->isSized()) { 1588 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n"); 1589 continue; 1590 } 1591 unsigned Size = 0; 1592 Value *Store = 0; 1593 // Compute the Shadow for arg even if it is ByVal, because 1594 // in that case getShadow() will copy the actual arg shadow to 1595 // __msan_param_tls. 1596 Value *ArgShadow = getShadow(A); 1597 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset); 1598 DEBUG(dbgs() << " Arg#" << i << ": " << *A << 1599 " Shadow: " << *ArgShadow << "\n"); 1600 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) { 1601 assert(A->getType()->isPointerTy() && 1602 "ByVal argument is not a pointer!"); 1603 Size = MS.TD->getTypeAllocSize(A->getType()->getPointerElementType()); 1604 unsigned Alignment = CS.getParamAlignment(i + 1); 1605 Store = IRB.CreateMemCpy(ArgShadowBase, 1606 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB), 1607 Size, Alignment); 1608 } else { 1609 Size = MS.TD->getTypeAllocSize(A->getType()); 1610 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase, 1611 kShadowTLSAlignment); 1612 } 1613 if (MS.TrackOrigins) 1614 IRB.CreateStore(getOrigin(A), 1615 getOriginPtrForArgument(A, IRB, ArgOffset)); 1616 (void)Store; 1617 assert(Size != 0 && Store != 0); 1618 DEBUG(dbgs() << " Param:" << *Store << "\n"); 1619 ArgOffset += DataLayout::RoundUpAlignment(Size, 8); 1620 } 1621 DEBUG(dbgs() << " done with call args\n"); 1622 1623 FunctionType *FT = 1624 cast<FunctionType>(CS.getCalledValue()->getType()-> getContainedType(0)); 1625 if (FT->isVarArg()) { 1626 VAHelper->visitCallSite(CS, IRB); 1627 } 1628 1629 // Now, get the shadow for the RetVal. 1630 if (!I.getType()->isSized()) return; 1631 IRBuilder<> IRBBefore(&I); 1632 // Untill we have full dynamic coverage, make sure the retval shadow is 0. 1633 Value *Base = getShadowPtrForRetval(&I, IRBBefore); 1634 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment); 1635 Instruction *NextInsn = 0; 1636 if (CS.isCall()) { 1637 NextInsn = I.getNextNode(); 1638 } else { 1639 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest(); 1640 if (!NormalDest->getSinglePredecessor()) { 1641 // FIXME: this case is tricky, so we are just conservative here. 1642 // Perhaps we need to split the edge between this BB and NormalDest, 1643 // but a naive attempt to use SplitEdge leads to a crash. 1644 setShadow(&I, getCleanShadow(&I)); 1645 setOrigin(&I, getCleanOrigin()); 1646 return; 1647 } 1648 NextInsn = NormalDest->getFirstInsertionPt(); 1649 assert(NextInsn && 1650 "Could not find insertion point for retval shadow load"); 1651 } 1652 IRBuilder<> IRBAfter(NextInsn); 1653 Value *RetvalShadow = 1654 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter), 1655 kShadowTLSAlignment, "_msret"); 1656 setShadow(&I, RetvalShadow); 1657 if (MS.TrackOrigins) 1658 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter))); 1659 } 1660 1661 void visitReturnInst(ReturnInst &I) { 1662 IRBuilder<> IRB(&I); 1663 if (Value *RetVal = I.getReturnValue()) { 1664 // Set the shadow for the RetVal. 1665 Value *Shadow = getShadow(RetVal); 1666 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB); 1667 DEBUG(dbgs() << "Return: " << *Shadow << "\n" << *ShadowPtr << "\n"); 1668 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment); 1669 if (MS.TrackOrigins) 1670 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB)); 1671 } 1672 } 1673 1674 void visitPHINode(PHINode &I) { 1675 IRBuilder<> IRB(&I); 1676 ShadowPHINodes.push_back(&I); 1677 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(), 1678 "_msphi_s")); 1679 if (MS.TrackOrigins) 1680 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(), 1681 "_msphi_o")); 1682 } 1683 1684 void visitAllocaInst(AllocaInst &I) { 1685 setShadow(&I, getCleanShadow(&I)); 1686 if (!ClPoisonStack) return; 1687 IRBuilder<> IRB(I.getNextNode()); 1688 uint64_t Size = MS.TD->getTypeAllocSize(I.getAllocatedType()); 1689 if (ClPoisonStackWithCall) { 1690 IRB.CreateCall2(MS.MsanPoisonStackFn, 1691 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()), 1692 ConstantInt::get(MS.IntptrTy, Size)); 1693 } else { 1694 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB); 1695 IRB.CreateMemSet(ShadowBase, IRB.getInt8(ClPoisonStackPattern), 1696 Size, I.getAlignment()); 1697 } 1698 1699 if (MS.TrackOrigins) { 1700 setOrigin(&I, getCleanOrigin()); 1701 SmallString<2048> StackDescriptionStorage; 1702 raw_svector_ostream StackDescription(StackDescriptionStorage); 1703 // We create a string with a description of the stack allocation and 1704 // pass it into __msan_set_alloca_origin. 1705 // It will be printed by the run-time if stack-originated UMR is found. 1706 // The first 4 bytes of the string are set to '----' and will be replaced 1707 // by __msan_va_arg_overflow_size_tls at the first call. 1708 StackDescription << "----" << I.getName() << "@" << F.getName(); 1709 Value *Descr = 1710 createPrivateNonConstGlobalForString(*F.getParent(), 1711 StackDescription.str()); 1712 IRB.CreateCall3(MS.MsanSetAllocaOriginFn, 1713 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()), 1714 ConstantInt::get(MS.IntptrTy, Size), 1715 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy())); 1716 } 1717 } 1718 1719 void visitSelectInst(SelectInst& I) { 1720 IRBuilder<> IRB(&I); 1721 setShadow(&I, IRB.CreateSelect(I.getCondition(), 1722 getShadow(I.getTrueValue()), getShadow(I.getFalseValue()), 1723 "_msprop")); 1724 if (MS.TrackOrigins) { 1725 // Origins are always i32, so any vector conditions must be flattened. 1726 // FIXME: consider tracking vector origins for app vectors? 1727 Value *Cond = I.getCondition(); 1728 if (Cond->getType()->isVectorTy()) { 1729 Value *ConvertedShadow = convertToShadowTyNoVec(Cond, IRB); 1730 Cond = IRB.CreateICmpNE(ConvertedShadow, 1731 getCleanShadow(ConvertedShadow), "_mso_select"); 1732 } 1733 setOrigin(&I, IRB.CreateSelect(Cond, 1734 getOrigin(I.getTrueValue()), getOrigin(I.getFalseValue()))); 1735 } 1736 } 1737 1738 void visitLandingPadInst(LandingPadInst &I) { 1739 // Do nothing. 1740 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1 1741 setShadow(&I, getCleanShadow(&I)); 1742 setOrigin(&I, getCleanOrigin()); 1743 } 1744 1745 void visitGetElementPtrInst(GetElementPtrInst &I) { 1746 handleShadowOr(I); 1747 } 1748 1749 void visitExtractValueInst(ExtractValueInst &I) { 1750 IRBuilder<> IRB(&I); 1751 Value *Agg = I.getAggregateOperand(); 1752 DEBUG(dbgs() << "ExtractValue: " << I << "\n"); 1753 Value *AggShadow = getShadow(Agg); 1754 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n"); 1755 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices()); 1756 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n"); 1757 setShadow(&I, ResShadow); 1758 setOrigin(&I, getCleanOrigin()); 1759 } 1760 1761 void visitInsertValueInst(InsertValueInst &I) { 1762 IRBuilder<> IRB(&I); 1763 DEBUG(dbgs() << "InsertValue: " << I << "\n"); 1764 Value *AggShadow = getShadow(I.getAggregateOperand()); 1765 Value *InsShadow = getShadow(I.getInsertedValueOperand()); 1766 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n"); 1767 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n"); 1768 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices()); 1769 DEBUG(dbgs() << " Res: " << *Res << "\n"); 1770 setShadow(&I, Res); 1771 setOrigin(&I, getCleanOrigin()); 1772 } 1773 1774 void dumpInst(Instruction &I) { 1775 if (CallInst *CI = dyn_cast<CallInst>(&I)) { 1776 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n"; 1777 } else { 1778 errs() << "ZZZ " << I.getOpcodeName() << "\n"; 1779 } 1780 errs() << "QQQ " << I << "\n"; 1781 } 1782 1783 void visitResumeInst(ResumeInst &I) { 1784 DEBUG(dbgs() << "Resume: " << I << "\n"); 1785 // Nothing to do here. 1786 } 1787 1788 void visitInstruction(Instruction &I) { 1789 // Everything else: stop propagating and check for poisoned shadow. 1790 if (ClDumpStrictInstructions) 1791 dumpInst(I); 1792 DEBUG(dbgs() << "DEFAULT: " << I << "\n"); 1793 for (size_t i = 0, n = I.getNumOperands(); i < n; i++) 1794 insertCheck(I.getOperand(i), &I); 1795 setShadow(&I, getCleanShadow(&I)); 1796 setOrigin(&I, getCleanOrigin()); 1797 } 1798 }; 1799 1800 /// \brief AMD64-specific implementation of VarArgHelper. 1801 struct VarArgAMD64Helper : public VarArgHelper { 1802 // An unfortunate workaround for asymmetric lowering of va_arg stuff. 1803 // See a comment in visitCallSite for more details. 1804 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7 1805 static const unsigned AMD64FpEndOffset = 176; 1806 1807 Function &F; 1808 MemorySanitizer &MS; 1809 MemorySanitizerVisitor &MSV; 1810 Value *VAArgTLSCopy; 1811 Value *VAArgOverflowSize; 1812 1813 SmallVector<CallInst*, 16> VAStartInstrumentationList; 1814 1815 VarArgAMD64Helper(Function &F, MemorySanitizer &MS, 1816 MemorySanitizerVisitor &MSV) 1817 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { } 1818 1819 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory }; 1820 1821 ArgKind classifyArgument(Value* arg) { 1822 // A very rough approximation of X86_64 argument classification rules. 1823 Type *T = arg->getType(); 1824 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy()) 1825 return AK_FloatingPoint; 1826 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64) 1827 return AK_GeneralPurpose; 1828 if (T->isPointerTy()) 1829 return AK_GeneralPurpose; 1830 return AK_Memory; 1831 } 1832 1833 // For VarArg functions, store the argument shadow in an ABI-specific format 1834 // that corresponds to va_list layout. 1835 // We do this because Clang lowers va_arg in the frontend, and this pass 1836 // only sees the low level code that deals with va_list internals. 1837 // A much easier alternative (provided that Clang emits va_arg instructions) 1838 // would have been to associate each live instance of va_list with a copy of 1839 // MSanParamTLS, and extract shadow on va_arg() call in the argument list 1840 // order. 1841 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) { 1842 unsigned GpOffset = 0; 1843 unsigned FpOffset = AMD64GpEndOffset; 1844 unsigned OverflowOffset = AMD64FpEndOffset; 1845 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end(); 1846 ArgIt != End; ++ArgIt) { 1847 Value *A = *ArgIt; 1848 ArgKind AK = classifyArgument(A); 1849 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset) 1850 AK = AK_Memory; 1851 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset) 1852 AK = AK_Memory; 1853 Value *Base; 1854 switch (AK) { 1855 case AK_GeneralPurpose: 1856 Base = getShadowPtrForVAArgument(A, IRB, GpOffset); 1857 GpOffset += 8; 1858 break; 1859 case AK_FloatingPoint: 1860 Base = getShadowPtrForVAArgument(A, IRB, FpOffset); 1861 FpOffset += 16; 1862 break; 1863 case AK_Memory: 1864 uint64_t ArgSize = MS.TD->getTypeAllocSize(A->getType()); 1865 Base = getShadowPtrForVAArgument(A, IRB, OverflowOffset); 1866 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8); 1867 } 1868 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment); 1869 } 1870 Constant *OverflowSize = 1871 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset); 1872 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS); 1873 } 1874 1875 /// \brief Compute the shadow address for a given va_arg. 1876 Value *getShadowPtrForVAArgument(Value *A, IRBuilder<> &IRB, 1877 int ArgOffset) { 1878 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy); 1879 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset)); 1880 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(A), 0), 1881 "_msarg"); 1882 } 1883 1884 void visitVAStartInst(VAStartInst &I) { 1885 IRBuilder<> IRB(&I); 1886 VAStartInstrumentationList.push_back(&I); 1887 Value *VAListTag = I.getArgOperand(0); 1888 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB); 1889 1890 // Unpoison the whole __va_list_tag. 1891 // FIXME: magic ABI constants. 1892 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()), 1893 /* size */24, /* alignment */8, false); 1894 } 1895 1896 void visitVACopyInst(VACopyInst &I) { 1897 IRBuilder<> IRB(&I); 1898 Value *VAListTag = I.getArgOperand(0); 1899 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB); 1900 1901 // Unpoison the whole __va_list_tag. 1902 // FIXME: magic ABI constants. 1903 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()), 1904 /* size */24, /* alignment */8, false); 1905 } 1906 1907 void finalizeInstrumentation() { 1908 assert(!VAArgOverflowSize && !VAArgTLSCopy && 1909 "finalizeInstrumentation called twice"); 1910 if (!VAStartInstrumentationList.empty()) { 1911 // If there is a va_start in this function, make a backup copy of 1912 // va_arg_tls somewhere in the function entry block. 1913 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI()); 1914 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS); 1915 Value *CopySize = 1916 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset), 1917 VAArgOverflowSize); 1918 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize); 1919 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8); 1920 } 1921 1922 // Instrument va_start. 1923 // Copy va_list shadow from the backup copy of the TLS contents. 1924 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) { 1925 CallInst *OrigInst = VAStartInstrumentationList[i]; 1926 IRBuilder<> IRB(OrigInst->getNextNode()); 1927 Value *VAListTag = OrigInst->getArgOperand(0); 1928 1929 Value *RegSaveAreaPtrPtr = 1930 IRB.CreateIntToPtr( 1931 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy), 1932 ConstantInt::get(MS.IntptrTy, 16)), 1933 Type::getInt64PtrTy(*MS.C)); 1934 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr); 1935 Value *RegSaveAreaShadowPtr = 1936 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB); 1937 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, 1938 AMD64FpEndOffset, 16); 1939 1940 Value *OverflowArgAreaPtrPtr = 1941 IRB.CreateIntToPtr( 1942 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy), 1943 ConstantInt::get(MS.IntptrTy, 8)), 1944 Type::getInt64PtrTy(*MS.C)); 1945 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr); 1946 Value *OverflowArgAreaShadowPtr = 1947 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB); 1948 Value *SrcPtr = 1949 getShadowPtrForVAArgument(VAArgTLSCopy, IRB, AMD64FpEndOffset); 1950 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16); 1951 } 1952 } 1953 }; 1954 1955 VarArgHelper* CreateVarArgHelper(Function &Func, MemorySanitizer &Msan, 1956 MemorySanitizerVisitor &Visitor) { 1957 return new VarArgAMD64Helper(Func, Msan, Visitor); 1958 } 1959 1960 } // namespace 1961 1962 bool MemorySanitizer::runOnFunction(Function &F) { 1963 MemorySanitizerVisitor Visitor(F, *this); 1964 1965 // Clear out readonly/readnone attributes. 1966 AttrBuilder B; 1967 B.addAttribute(Attribute::ReadOnly) 1968 .addAttribute(Attribute::ReadNone); 1969 F.removeAttributes(AttributeSet::FunctionIndex, 1970 AttributeSet::get(F.getContext(), 1971 AttributeSet::FunctionIndex, B)); 1972 1973 return Visitor.runOnFunction(); 1974 } 1975