1 //===-- MemorySanitizer.cpp - detector of uninitialized reads -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file is a part of MemorySanitizer, a detector of uninitialized
13 /// Status: early prototype.
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.
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).
39 /// The default value of shadow is 0, which means "clean" (not poisoned).
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 //===----------------------------------------------------------------------===//
48 #define DEBUG_TYPE "msan"
50 #include "llvm/Transforms/Instrumentation.h"
51 #include "BlackList.h"
52 #include "llvm/ADT/DepthFirstIterator.h"
53 #include "llvm/ADT/SmallString.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/ValueMap.h"
56 #include "llvm/DataLayout.h"
57 #include "llvm/Function.h"
58 #include "llvm/IRBuilder.h"
59 #include "llvm/InlineAsm.h"
60 #include "llvm/InstVisitor.h"
61 #include "llvm/IntrinsicInst.h"
62 #include "llvm/LLVMContext.h"
63 #include "llvm/MDBuilder.h"
64 #include "llvm/Module.h"
65 #include "llvm/Support/CommandLine.h"
66 #include "llvm/Support/Compiler.h"
67 #include "llvm/Support/Debug.h"
68 #include "llvm/Support/raw_ostream.h"
69 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
70 #include "llvm/Transforms/Utils/ModuleUtils.h"
71 #include "llvm/Type.h"
75 static const uint64_t kShadowMask32 = 1ULL << 31;
76 static const uint64_t kShadowMask64 = 1ULL << 46;
77 static const uint64_t kOriginOffset32 = 1ULL << 30;
78 static const uint64_t kOriginOffset64 = 1ULL << 45;
79 static const uint64_t kShadowTLSAlignment = 8;
81 // This is an important flag that makes the reports much more
82 // informative at the cost of greater slowdown. Not fully implemented
84 // FIXME: this should be a top-level clang flag, e.g.
85 // -fmemory-sanitizer-full.
86 static cl::opt<bool> ClTrackOrigins("msan-track-origins",
87 cl::desc("Track origins (allocation sites) of poisoned memory"),
88 cl::Hidden, cl::init(false));
89 static cl::opt<bool> ClKeepGoing("msan-keep-going",
90 cl::desc("keep going after reporting a UMR"),
91 cl::Hidden, cl::init(false));
92 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
93 cl::desc("poison uninitialized stack variables"),
94 cl::Hidden, cl::init(true));
95 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
96 cl::desc("poison uninitialized stack variables with a call"),
97 cl::Hidden, cl::init(false));
98 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
99 cl::desc("poison uninitialized stack variables with the given patter"),
100 cl::Hidden, cl::init(0xff));
102 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
103 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
104 cl::Hidden, cl::init(true));
106 static cl::opt<bool> ClStoreCleanOrigin("msan-store-clean-origin",
107 cl::desc("store origin for clean (fully initialized) values"),
108 cl::Hidden, cl::init(false));
110 // This flag controls whether we check the shadow of the address
111 // operand of load or store. Such bugs are very rare, since load from
112 // a garbage address typically results in SEGV, but still happen
113 // (e.g. only lower bits of address are garbage, or the access happens
114 // early at program startup where malloc-ed memory is more likely to
115 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
116 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
117 cl::desc("report accesses through a pointer which has poisoned shadow"),
118 cl::Hidden, cl::init(true));
120 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
121 cl::desc("print out instructions with default strict semantics"),
122 cl::Hidden, cl::init(false));
124 static cl::opt<std::string> ClBlackListFile("msan-blacklist",
125 cl::desc("File containing the list of functions where MemorySanitizer "
126 "should not report bugs"), cl::Hidden);
130 /// \brief An instrumentation pass implementing detection of uninitialized
133 /// MemorySanitizer: instrument the code in module to find
134 /// uninitialized reads.
135 class MemorySanitizer : public FunctionPass {
137 MemorySanitizer() : FunctionPass(ID), TD(0), WarningFn(0) { }
138 const char *getPassName() const { return "MemorySanitizer"; }
139 bool runOnFunction(Function &F);
140 bool doInitialization(Module &M);
141 static char ID; // Pass identification, replacement for typeid.
144 void initializeCallbacks(Module &M);
150 /// \brief Thread-local shadow storage for function parameters.
151 GlobalVariable *ParamTLS;
152 /// \brief Thread-local origin storage for function parameters.
153 GlobalVariable *ParamOriginTLS;
154 /// \brief Thread-local shadow storage for function return value.
155 GlobalVariable *RetvalTLS;
156 /// \brief Thread-local origin storage for function return value.
157 GlobalVariable *RetvalOriginTLS;
158 /// \brief Thread-local shadow storage for in-register va_arg function
159 /// parameters (x86_64-specific).
160 GlobalVariable *VAArgTLS;
161 /// \brief Thread-local shadow storage for va_arg overflow area
162 /// (x86_64-specific).
163 GlobalVariable *VAArgOverflowSizeTLS;
164 /// \brief Thread-local space used to pass origin value to the UMR reporting
166 GlobalVariable *OriginTLS;
168 /// \brief The run-time callback to print a warning.
170 /// \brief Run-time helper that copies origin info for a memory range.
171 Value *MsanCopyOriginFn;
172 /// \brief Run-time helper that generates a new origin value for a stack
174 Value *MsanSetAllocaOriginFn;
175 /// \brief Run-time helper that poisons stack on function entry.
176 Value *MsanPoisonStackFn;
177 /// \brief MSan runtime replacements for memmove, memcpy and memset.
178 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
180 /// \brief Address mask used in application-to-shadow address calculation.
181 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
183 /// \brief Offset of the origin shadow from the "normal" shadow.
184 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
185 uint64_t OriginOffset;
186 /// \brief Branch weights for error reporting.
187 MDNode *ColdCallWeights;
188 /// \brief Branch weights for origin store.
189 MDNode *OriginStoreWeights;
190 /// \brief The blacklist.
191 OwningPtr<BlackList> BL;
192 /// \brief An empty volatile inline asm that prevents callback merge.
195 friend struct MemorySanitizerVisitor;
196 friend struct VarArgAMD64Helper;
200 char MemorySanitizer::ID = 0;
201 INITIALIZE_PASS(MemorySanitizer, "msan",
202 "MemorySanitizer: detects uninitialized reads.",
205 FunctionPass *llvm::createMemorySanitizerPass() {
206 return new MemorySanitizer();
209 /// \brief Create a non-const global initialized with the given string.
211 /// Creates a writable global for Str so that we can pass it to the
212 /// run-time lib. Runtime uses first 4 bytes of the string to store the
213 /// frame ID, so the string needs to be mutable.
214 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
216 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
217 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
218 GlobalValue::PrivateLinkage, StrConst, "");
222 /// \brief Insert extern declaration of runtime-provided functions and globals.
223 void MemorySanitizer::initializeCallbacks(Module &M) {
224 // Only do this once.
229 // Create the callback.
230 // FIXME: this function should have "Cold" calling conv,
231 // which is not yet implemented.
232 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
233 : "__msan_warning_noreturn";
234 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
236 MsanCopyOriginFn = M.getOrInsertFunction(
237 "__msan_copy_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(),
238 IRB.getInt8PtrTy(), IntptrTy, NULL);
239 MsanSetAllocaOriginFn = M.getOrInsertFunction(
240 "__msan_set_alloca_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
241 IRB.getInt8PtrTy(), NULL);
242 MsanPoisonStackFn = M.getOrInsertFunction(
243 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
244 MemmoveFn = M.getOrInsertFunction(
245 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
247 MemcpyFn = M.getOrInsertFunction(
248 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
250 MemsetFn = M.getOrInsertFunction(
251 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
255 RetvalTLS = new GlobalVariable(
256 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
257 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
258 GlobalVariable::GeneralDynamicTLSModel);
259 RetvalOriginTLS = new GlobalVariable(
260 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
261 "__msan_retval_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
263 ParamTLS = new GlobalVariable(
264 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
265 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
266 GlobalVariable::GeneralDynamicTLSModel);
267 ParamOriginTLS = new GlobalVariable(
268 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
269 0, "__msan_param_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
271 VAArgTLS = new GlobalVariable(
272 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
273 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
274 GlobalVariable::GeneralDynamicTLSModel);
275 VAArgOverflowSizeTLS = new GlobalVariable(
276 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
277 "__msan_va_arg_overflow_size_tls", 0,
278 GlobalVariable::GeneralDynamicTLSModel);
279 OriginTLS = new GlobalVariable(
280 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
281 "__msan_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
283 // We insert an empty inline asm after __msan_report* to avoid callback merge.
284 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
285 StringRef(""), StringRef(""),
286 /*hasSideEffects=*/true);
289 /// \brief Module-level initialization.
291 /// inserts a call to __msan_init to the module's constructor list.
292 bool MemorySanitizer::doInitialization(Module &M) {
293 TD = getAnalysisIfAvailable<DataLayout>();
296 BL.reset(new BlackList(ClBlackListFile));
297 C = &(M.getContext());
298 unsigned PtrSize = TD->getPointerSizeInBits(/* AddressSpace */0);
301 ShadowMask = kShadowMask64;
302 OriginOffset = kOriginOffset64;
305 ShadowMask = kShadowMask32;
306 OriginOffset = kOriginOffset32;
309 report_fatal_error("unsupported pointer size");
314 IntptrTy = IRB.getIntPtrTy(TD);
315 OriginTy = IRB.getInt32Ty();
317 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
318 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
320 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
321 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
322 "__msan_init", IRB.getVoidTy(), NULL)), 0);
324 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
325 IRB.getInt32(ClTrackOrigins), "__msan_track_origins");
332 /// \brief A helper class that handles instrumentation of VarArg
333 /// functions on a particular platform.
335 /// Implementations are expected to insert the instrumentation
336 /// necessary to propagate argument shadow through VarArg function
337 /// calls. Visit* methods are called during an InstVisitor pass over
338 /// the function, and should avoid creating new basic blocks. A new
339 /// instance of this class is created for each instrumented function.
340 struct VarArgHelper {
341 /// \brief Visit a CallSite.
342 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
344 /// \brief Visit a va_start call.
345 virtual void visitVAStartInst(VAStartInst &I) = 0;
347 /// \brief Visit a va_copy call.
348 virtual void visitVACopyInst(VACopyInst &I) = 0;
350 /// \brief Finalize function instrumentation.
352 /// This method is called after visiting all interesting (see above)
353 /// instructions in a function.
354 virtual void finalizeInstrumentation() = 0;
356 virtual ~VarArgHelper() {}
359 struct MemorySanitizerVisitor;
362 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
363 MemorySanitizerVisitor &Visitor);
365 /// This class does all the work for a given function. Store and Load
366 /// instructions store and load corresponding shadow and origin
367 /// values. Most instructions propagate shadow from arguments to their
368 /// return values. Certain instructions (most importantly, BranchInst)
369 /// test their argument shadow and print reports (with a runtime call) if it's
371 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
374 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
375 ValueMap<Value*, Value*> ShadowMap, OriginMap;
377 OwningPtr<VarArgHelper> VAHelper;
379 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
380 // See a comment in visitCallSite for more details.
381 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
382 static const unsigned AMD64FpEndOffset = 176;
384 struct ShadowOriginAndInsertPoint {
387 Instruction *OrigIns;
388 ShadowOriginAndInsertPoint(Instruction *S, Instruction *O, Instruction *I)
389 : Shadow(S), Origin(O), OrigIns(I) { }
390 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
392 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
393 SmallVector<Instruction*, 16> StoreList;
395 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
396 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
397 InsertChecks = !MS.BL->isIn(F);
398 DEBUG(if (!InsertChecks)
399 dbgs() << "MemorySanitizer is not inserting checks into '"
400 << F.getName() << "'\n");
403 void materializeStores() {
404 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
405 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]);
408 Value *Val = I.getValueOperand();
409 Value *Addr = I.getPointerOperand();
410 Value *Shadow = getShadow(Val);
411 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
413 StoreInst *NewSI = IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
414 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
416 // If the store is volatile, add a check.
418 insertCheck(Val, &I);
419 if (ClCheckAccessAddress)
420 insertCheck(Addr, &I);
422 if (ClTrackOrigins) {
423 if (ClStoreCleanOrigin || isa<StructType>(Shadow->getType())) {
424 IRB.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRB), I.getAlignment());
426 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
428 Constant *Cst = dyn_cast_or_null<Constant>(ConvertedShadow);
429 // TODO(eugenis): handle non-zero constant shadow by inserting an
430 // unconditional check (can not simply fail compilation as this could
431 // be in the dead code).
435 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
436 getCleanShadow(ConvertedShadow), "_mscmp");
437 Instruction *CheckTerm =
438 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false, MS.OriginStoreWeights);
439 IRBuilder<> IRBNewBlock(CheckTerm);
440 IRBNewBlock.CreateAlignedStore(getOrigin(Val),
441 getOriginPtr(Addr, IRBNewBlock), I.getAlignment());
447 void materializeChecks() {
448 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
449 Instruction *Shadow = InstrumentationList[i].Shadow;
450 Instruction *OrigIns = InstrumentationList[i].OrigIns;
451 IRBuilder<> IRB(OrigIns);
452 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
453 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
454 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
455 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
456 getCleanShadow(ConvertedShadow), "_mscmp");
457 Instruction *CheckTerm =
458 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp),
459 /* Unreachable */ !ClKeepGoing,
462 IRB.SetInsertPoint(CheckTerm);
463 if (ClTrackOrigins) {
464 Instruction *Origin = InstrumentationList[i].Origin;
465 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
468 CallInst *Call = IRB.CreateCall(MS.WarningFn);
469 Call->setDebugLoc(OrigIns->getDebugLoc());
470 IRB.CreateCall(MS.EmptyAsm);
471 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
473 DEBUG(dbgs() << "DONE:\n" << F);
476 /// \brief Add MemorySanitizer instrumentation to a function.
477 bool runOnFunction() {
478 MS.initializeCallbacks(*F.getParent());
479 if (!MS.TD) return false;
480 // Iterate all BBs in depth-first order and create shadow instructions
481 // for all instructions (where applicable).
482 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
483 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
484 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
485 BasicBlock *BB = *DI;
489 // Finalize PHI nodes.
490 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
491 PHINode *PN = ShadowPHINodes[i];
492 PHINode *PNS = cast<PHINode>(getShadow(PN));
493 PHINode *PNO = ClTrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
494 size_t NumValues = PN->getNumIncomingValues();
495 for (size_t v = 0; v < NumValues; v++) {
496 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
498 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
502 VAHelper->finalizeInstrumentation();
504 // Delayed instrumentation of StoreInst.
505 // This may add new checks to be inserted later.
508 // Insert shadow value checks.
514 /// \brief Compute the shadow type that corresponds to a given Value.
515 Type *getShadowTy(Value *V) {
516 return getShadowTy(V->getType());
519 /// \brief Compute the shadow type that corresponds to a given Type.
520 Type *getShadowTy(Type *OrigTy) {
521 if (!OrigTy->isSized()) {
524 // For integer type, shadow is the same as the original type.
525 // This may return weird-sized types like i1.
526 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
528 if (VectorType *VT = dyn_cast<VectorType>(OrigTy))
529 return VectorType::getInteger(VT);
530 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
531 SmallVector<Type*, 4> Elements;
532 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
533 Elements.push_back(getShadowTy(ST->getElementType(i)));
534 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
535 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
538 uint32_t TypeSize = MS.TD->getTypeStoreSizeInBits(OrigTy);
539 return IntegerType::get(*MS.C, TypeSize);
542 /// \brief Flatten a vector type.
543 Type *getShadowTyNoVec(Type *ty) {
544 if (VectorType *vt = dyn_cast<VectorType>(ty))
545 return IntegerType::get(*MS.C, vt->getBitWidth());
549 /// \brief Convert a shadow value to it's flattened variant.
550 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
551 Type *Ty = V->getType();
552 Type *NoVecTy = getShadowTyNoVec(Ty);
553 if (Ty == NoVecTy) return V;
554 return IRB.CreateBitCast(V, NoVecTy);
557 /// \brief Compute the shadow address that corresponds to a given application
560 /// Shadow = Addr & ~ShadowMask.
561 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
564 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
565 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
566 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
569 /// \brief Compute the origin address that corresponds to a given application
572 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
573 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
575 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
576 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
578 IRB.CreateAdd(ShadowLong,
579 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
581 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
582 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
585 /// \brief Compute the shadow address for a given function argument.
587 /// Shadow = ParamTLS+ArgOffset.
588 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
590 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
591 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
592 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
596 /// \brief Compute the origin address for a given function argument.
597 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
599 if (!ClTrackOrigins) return 0;
600 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
601 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
602 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
606 /// \brief Compute the shadow address for a retval.
607 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
608 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
609 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
613 /// \brief Compute the origin address for a retval.
614 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
615 // We keep a single origin for the entire retval. Might be too optimistic.
616 return MS.RetvalOriginTLS;
619 /// \brief Set SV to be the shadow value for V.
620 void setShadow(Value *V, Value *SV) {
621 assert(!ShadowMap.count(V) && "Values may only have one shadow");
625 /// \brief Set Origin to be the origin value for V.
626 void setOrigin(Value *V, Value *Origin) {
627 if (!ClTrackOrigins) return;
628 assert(!OriginMap.count(V) && "Values may only have one origin");
629 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
630 OriginMap[V] = Origin;
633 /// \brief Create a clean shadow value for a given value.
635 /// Clean shadow (all zeroes) means all bits of the value are defined
637 Value *getCleanShadow(Value *V) {
638 Type *ShadowTy = getShadowTy(V);
641 return Constant::getNullValue(ShadowTy);
644 /// \brief Create a dirty shadow of a given shadow type.
645 Constant *getPoisonedShadow(Type *ShadowTy) {
647 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
648 return Constant::getAllOnesValue(ShadowTy);
649 StructType *ST = cast<StructType>(ShadowTy);
650 SmallVector<Constant *, 4> Vals;
651 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
652 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
653 return ConstantStruct::get(ST, Vals);
656 /// \brief Create a clean (zero) origin.
657 Value *getCleanOrigin() {
658 return Constant::getNullValue(MS.OriginTy);
661 /// \brief Get the shadow value for a given Value.
663 /// This function either returns the value set earlier with setShadow,
664 /// or extracts if from ParamTLS (for function arguments).
665 Value *getShadow(Value *V) {
666 if (Instruction *I = dyn_cast<Instruction>(V)) {
667 // For instructions the shadow is already stored in the map.
668 Value *Shadow = ShadowMap[V];
670 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
672 assert(Shadow && "No shadow for a value");
676 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
677 Value *AllOnes = getPoisonedShadow(getShadowTy(V));
678 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
682 if (Argument *A = dyn_cast<Argument>(V)) {
683 // For arguments we compute the shadow on demand and store it in the map.
684 Value **ShadowPtr = &ShadowMap[V];
687 Function *F = A->getParent();
688 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
689 unsigned ArgOffset = 0;
690 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
692 if (!AI->getType()->isSized()) {
693 DEBUG(dbgs() << "Arg is not sized\n");
696 unsigned Size = AI->hasByValAttr()
697 ? MS.TD->getTypeAllocSize(AI->getType()->getPointerElementType())
698 : MS.TD->getTypeAllocSize(AI->getType());
700 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
701 if (AI->hasByValAttr()) {
702 // ByVal pointer itself has clean shadow. We copy the actual
703 // argument shadow to the underlying memory.
704 Value *Cpy = EntryIRB.CreateMemCpy(
705 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
706 Base, Size, AI->getParamAlignment());
707 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
709 *ShadowPtr = getCleanShadow(V);
711 *ShadowPtr = EntryIRB.CreateLoad(Base);
713 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
714 **ShadowPtr << "\n");
715 if (ClTrackOrigins) {
716 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
717 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
720 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
722 assert(*ShadowPtr && "Could not find shadow for an argument");
725 // For everything else the shadow is zero.
726 return getCleanShadow(V);
729 /// \brief Get the shadow for i-th argument of the instruction I.
730 Value *getShadow(Instruction *I, int i) {
731 return getShadow(I->getOperand(i));
734 /// \brief Get the origin for a value.
735 Value *getOrigin(Value *V) {
736 if (!ClTrackOrigins) return 0;
737 if (isa<Instruction>(V) || isa<Argument>(V)) {
738 Value *Origin = OriginMap[V];
740 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
741 Origin = getCleanOrigin();
745 return getCleanOrigin();
748 /// \brief Get the origin for i-th argument of the instruction I.
749 Value *getOrigin(Instruction *I, int i) {
750 return getOrigin(I->getOperand(i));
753 /// \brief Remember the place where a shadow check should be inserted.
755 /// This location will be later instrumented with a check that will print a
756 /// UMR warning in runtime if the value is not fully defined.
757 void insertCheck(Value *Val, Instruction *OrigIns) {
759 if (!InsertChecks) return;
760 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
763 Type *ShadowTy = Shadow->getType();
764 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
765 "Can only insert checks for integer and vector shadow types");
767 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
768 InstrumentationList.push_back(
769 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
772 //------------------- Visitors.
774 /// \brief Instrument LoadInst
776 /// Loads the corresponding shadow and (optionally) origin.
777 /// Optionally, checks that the load address is fully defined.
778 void visitLoadInst(LoadInst &I) {
779 assert(I.getType()->isSized() && "Load type must have size");
781 Type *ShadowTy = getShadowTy(&I);
782 Value *Addr = I.getPointerOperand();
783 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
784 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
786 if (ClCheckAccessAddress)
787 insertCheck(I.getPointerOperand(), &I);
790 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), I.getAlignment()));
793 /// \brief Instrument StoreInst
795 /// Stores the corresponding shadow and (optionally) origin.
796 /// Optionally, checks that the store address is fully defined.
797 /// Volatile stores check that the value being stored is fully defined.
798 void visitStoreInst(StoreInst &I) {
799 StoreList.push_back(&I);
802 // Vector manipulation.
803 void visitExtractElementInst(ExtractElementInst &I) {
804 insertCheck(I.getOperand(1), &I);
806 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
808 setOrigin(&I, getOrigin(&I, 0));
811 void visitInsertElementInst(InsertElementInst &I) {
812 insertCheck(I.getOperand(2), &I);
814 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
815 I.getOperand(2), "_msprop"));
816 setOriginForNaryOp(I);
819 void visitShuffleVectorInst(ShuffleVectorInst &I) {
820 insertCheck(I.getOperand(2), &I);
822 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
823 I.getOperand(2), "_msprop"));
824 setOriginForNaryOp(I);
828 void visitSExtInst(SExtInst &I) {
830 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
831 setOrigin(&I, getOrigin(&I, 0));
834 void visitZExtInst(ZExtInst &I) {
836 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
837 setOrigin(&I, getOrigin(&I, 0));
840 void visitTruncInst(TruncInst &I) {
842 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
843 setOrigin(&I, getOrigin(&I, 0));
846 void visitBitCastInst(BitCastInst &I) {
848 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
849 setOrigin(&I, getOrigin(&I, 0));
852 void visitPtrToIntInst(PtrToIntInst &I) {
854 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
855 "_msprop_ptrtoint"));
856 setOrigin(&I, getOrigin(&I, 0));
859 void visitIntToPtrInst(IntToPtrInst &I) {
861 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
862 "_msprop_inttoptr"));
863 setOrigin(&I, getOrigin(&I, 0));
866 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
867 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
868 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
869 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
870 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
871 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
873 /// \brief Propagate shadow for bitwise AND.
875 /// This code is exact, i.e. if, for example, a bit in the left argument
876 /// is defined and 0, then neither the value not definedness of the
877 /// corresponding bit in B don't affect the resulting shadow.
878 void visitAnd(BinaryOperator &I) {
880 // "And" of 0 and a poisoned value results in unpoisoned value.
881 // 1&1 => 1; 0&1 => 0; p&1 => p;
882 // 1&0 => 0; 0&0 => 0; p&0 => 0;
883 // 1&p => p; 0&p => 0; p&p => p;
884 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
885 Value *S1 = getShadow(&I, 0);
886 Value *S2 = getShadow(&I, 1);
887 Value *V1 = I.getOperand(0);
888 Value *V2 = I.getOperand(1);
889 if (V1->getType() != S1->getType()) {
890 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
891 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
893 Value *S1S2 = IRB.CreateAnd(S1, S2);
894 Value *V1S2 = IRB.CreateAnd(V1, S2);
895 Value *S1V2 = IRB.CreateAnd(S1, V2);
896 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
897 setOriginForNaryOp(I);
900 void visitOr(BinaryOperator &I) {
902 // "Or" of 1 and a poisoned value results in unpoisoned value.
903 // 1|1 => 1; 0|1 => 1; p|1 => 1;
904 // 1|0 => 1; 0|0 => 0; p|0 => p;
905 // 1|p => 1; 0|p => p; p|p => p;
906 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
907 Value *S1 = getShadow(&I, 0);
908 Value *S2 = getShadow(&I, 1);
909 Value *V1 = IRB.CreateNot(I.getOperand(0));
910 Value *V2 = IRB.CreateNot(I.getOperand(1));
911 if (V1->getType() != S1->getType()) {
912 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
913 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
915 Value *S1S2 = IRB.CreateAnd(S1, S2);
916 Value *V1S2 = IRB.CreateAnd(V1, S2);
917 Value *S1V2 = IRB.CreateAnd(S1, V2);
918 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
919 setOriginForNaryOp(I);
922 /// \brief Default propagation of shadow and/or origin.
924 /// This class implements the general case of shadow propagation, used in all
925 /// cases where we don't know and/or don't care about what the operation
926 /// actually does. It converts all input shadow values to a common type
927 /// (extending or truncating as necessary), and bitwise OR's them.
929 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
930 /// fully initialized), and less prone to false positives.
932 /// This class also implements the general case of origin propagation. For a
933 /// Nary operation, result origin is set to the origin of an argument that is
934 /// not entirely initialized. If there is more than one such arguments, the
935 /// rightmost of them is picked. It does not matter which one is picked if all
936 /// arguments are initialized.
937 template <bool CombineShadow>
942 MemorySanitizerVisitor *MSV;
944 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
945 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {}
947 /// \brief Add a pair of shadow and origin values to the mix.
948 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
954 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
955 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
959 if (ClTrackOrigins) {
964 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
965 Value *Cond = IRB.CreateICmpNE(FlatShadow,
966 MSV->getCleanShadow(FlatShadow));
967 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
973 /// \brief Add an application value to the mix.
974 Combiner &Add(Value *V) {
975 Value *OpShadow = MSV->getShadow(V);
976 Value *OpOrigin = ClTrackOrigins ? MSV->getOrigin(V) : 0;
977 return Add(OpShadow, OpOrigin);
980 /// \brief Set the current combined values as the given instruction's shadow
982 void Done(Instruction *I) {
985 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
986 MSV->setShadow(I, Shadow);
988 if (ClTrackOrigins) {
990 MSV->setOrigin(I, Origin);
995 typedef Combiner<true> ShadowAndOriginCombiner;
996 typedef Combiner<false> OriginCombiner;
998 /// \brief Propagate origin for arbitrary operation.
999 void setOriginForNaryOp(Instruction &I) {
1000 if (!ClTrackOrigins) return;
1001 IRBuilder<> IRB(&I);
1002 OriginCombiner OC(this, IRB);
1003 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1008 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1009 return Ty->isVectorTy() ?
1010 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1011 Ty->getPrimitiveSizeInBits();
1014 /// \brief Cast between two shadow types, extending or truncating as
1016 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy) {
1017 Type *srcTy = V->getType();
1018 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1019 return IRB.CreateIntCast(V, dstTy, false);
1020 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1021 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1022 return IRB.CreateIntCast(V, dstTy, false);
1023 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1024 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1025 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1027 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), false);
1028 return IRB.CreateBitCast(V2, dstTy);
1029 // TODO: handle struct types.
1032 /// \brief Propagate shadow for arbitrary operation.
1033 void handleShadowOr(Instruction &I) {
1034 IRBuilder<> IRB(&I);
1035 ShadowAndOriginCombiner SC(this, IRB);
1036 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1041 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1042 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1043 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1044 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1045 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1046 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1047 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1049 void handleDiv(Instruction &I) {
1050 IRBuilder<> IRB(&I);
1051 // Strict on the second argument.
1052 insertCheck(I.getOperand(1), &I);
1053 setShadow(&I, getShadow(&I, 0));
1054 setOrigin(&I, getOrigin(&I, 0));
1057 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1058 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1059 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1060 void visitURem(BinaryOperator &I) { handleDiv(I); }
1061 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1062 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1064 /// \brief Instrument == and != comparisons.
1066 /// Sometimes the comparison result is known even if some of the bits of the
1067 /// arguments are not.
1068 void handleEqualityComparison(ICmpInst &I) {
1069 IRBuilder<> IRB(&I);
1070 Value *A = I.getOperand(0);
1071 Value *B = I.getOperand(1);
1072 Value *Sa = getShadow(A);
1073 Value *Sb = getShadow(B);
1074 if (A->getType()->isPointerTy())
1075 A = IRB.CreatePointerCast(A, MS.IntptrTy);
1076 if (B->getType()->isPointerTy())
1077 B = IRB.CreatePointerCast(B, MS.IntptrTy);
1078 // A == B <==> (C = A^B) == 0
1079 // A != B <==> (C = A^B) != 0
1081 Value *C = IRB.CreateXor(A, B);
1082 Value *Sc = IRB.CreateOr(Sa, Sb);
1083 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1084 // Result is defined if one of the following is true
1085 // * there is a defined 1 bit in C
1086 // * C is fully defined
1087 // Si = !(C & ~Sc) && Sc
1088 Value *Zero = Constant::getNullValue(Sc->getType());
1089 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1091 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1093 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1094 Si->setName("_msprop_icmp");
1096 setOriginForNaryOp(I);
1099 /// \brief Instrument signed relational comparisons.
1101 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1102 /// propagating the highest bit of the shadow. Everything else is delegated
1103 /// to handleShadowOr().
1104 void handleSignedRelationalComparison(ICmpInst &I) {
1105 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1106 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1108 CmpInst::Predicate pre = I.getPredicate();
1109 if (constOp0 && constOp0->isNullValue() &&
1110 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1111 op = I.getOperand(1);
1112 } else if (constOp1 && constOp1->isNullValue() &&
1113 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1114 op = I.getOperand(0);
1117 IRBuilder<> IRB(&I);
1119 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1120 setShadow(&I, Shadow);
1121 setOrigin(&I, getOrigin(op));
1127 void visitICmpInst(ICmpInst &I) {
1128 if (ClHandleICmp && I.isEquality())
1129 handleEqualityComparison(I);
1130 else if (ClHandleICmp && I.isSigned() && I.isRelational())
1131 handleSignedRelationalComparison(I);
1136 void visitFCmpInst(FCmpInst &I) {
1140 void handleShift(BinaryOperator &I) {
1141 IRBuilder<> IRB(&I);
1142 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1143 // Otherwise perform the same shift on S1.
1144 Value *S1 = getShadow(&I, 0);
1145 Value *S2 = getShadow(&I, 1);
1146 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1148 Value *V2 = I.getOperand(1);
1149 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1150 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1151 setOriginForNaryOp(I);
1154 void visitShl(BinaryOperator &I) { handleShift(I); }
1155 void visitAShr(BinaryOperator &I) { handleShift(I); }
1156 void visitLShr(BinaryOperator &I) { handleShift(I); }
1158 /// \brief Instrument llvm.memmove
1160 /// At this point we don't know if llvm.memmove will be inlined or not.
1161 /// If we don't instrument it and it gets inlined,
1162 /// our interceptor will not kick in and we will lose the memmove.
1163 /// If we instrument the call here, but it does not get inlined,
1164 /// we will memove the shadow twice: which is bad in case
1165 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1167 /// Similar situation exists for memcpy and memset.
1168 void visitMemMoveInst(MemMoveInst &I) {
1169 IRBuilder<> IRB(&I);
1172 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1173 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1174 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1175 I.eraseFromParent();
1178 // Similar to memmove: avoid copying shadow twice.
1179 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1180 // FIXME: consider doing manual inline for small constant sizes and proper
1182 void visitMemCpyInst(MemCpyInst &I) {
1183 IRBuilder<> IRB(&I);
1186 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1187 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1188 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1189 I.eraseFromParent();
1193 void visitMemSetInst(MemSetInst &I) {
1194 IRBuilder<> IRB(&I);
1197 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1198 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1199 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1200 I.eraseFromParent();
1203 void visitVAStartInst(VAStartInst &I) {
1204 VAHelper->visitVAStartInst(I);
1207 void visitVACopyInst(VACopyInst &I) {
1208 VAHelper->visitVACopyInst(I);
1211 void handleBswap(IntrinsicInst &I) {
1212 IRBuilder<> IRB(&I);
1213 Value *Op = I.getArgOperand(0);
1214 Type *OpType = Op->getType();
1215 Function *BswapFunc = Intrinsic::getDeclaration(
1216 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1217 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1218 setOrigin(&I, getOrigin(Op));
1221 void visitIntrinsicInst(IntrinsicInst &I) {
1222 switch (I.getIntrinsicID()) {
1223 case llvm::Intrinsic::bswap:
1224 handleBswap(I); break;
1226 visitInstruction(I); break;
1230 void visitCallSite(CallSite CS) {
1231 Instruction &I = *CS.getInstruction();
1232 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
1234 CallInst *Call = cast<CallInst>(&I);
1236 // For inline asm, do the usual thing: check argument shadow and mark all
1237 // outputs as clean. Note that any side effects of the inline asm that are
1238 // not immediately visible in its constraints are not handled.
1239 if (Call->isInlineAsm()) {
1240 visitInstruction(I);
1244 // Allow only tail calls with the same types, otherwise
1245 // we may have a false positive: shadow for a non-void RetVal
1246 // will get propagated to a void RetVal.
1247 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
1248 Call->setTailCall(false);
1250 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
1252 // We are going to insert code that relies on the fact that the callee
1253 // will become a non-readonly function after it is instrumented by us. To
1254 // prevent this code from being optimized out, mark that function
1255 // non-readonly in advance.
1256 if (Function *Func = Call->getCalledFunction()) {
1257 // Clear out readonly/readnone attributes.
1259 B.addAttribute(Attributes::ReadOnly)
1260 .addAttribute(Attributes::ReadNone);
1261 Func->removeAttribute(AttributeSet::FunctionIndex,
1262 Attributes::get(Func->getContext(), B));
1265 IRBuilder<> IRB(&I);
1266 unsigned ArgOffset = 0;
1267 DEBUG(dbgs() << " CallSite: " << I << "\n");
1268 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1269 ArgIt != End; ++ArgIt) {
1271 unsigned i = ArgIt - CS.arg_begin();
1272 if (!A->getType()->isSized()) {
1273 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
1278 // Compute the Shadow for arg even if it is ByVal, because
1279 // in that case getShadow() will copy the actual arg shadow to
1280 // __msan_param_tls.
1281 Value *ArgShadow = getShadow(A);
1282 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
1283 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
1284 " Shadow: " << *ArgShadow << "\n");
1285 if (CS.paramHasAttr(i + 1, Attributes::ByVal)) {
1286 assert(A->getType()->isPointerTy() &&
1287 "ByVal argument is not a pointer!");
1288 Size = MS.TD->getTypeAllocSize(A->getType()->getPointerElementType());
1289 unsigned Alignment = CS.getParamAlignment(i + 1);
1290 Store = IRB.CreateMemCpy(ArgShadowBase,
1291 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
1294 Size = MS.TD->getTypeAllocSize(A->getType());
1295 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
1296 kShadowTLSAlignment);
1299 IRB.CreateAlignedStore(getOrigin(A),
1300 getOriginPtrForArgument(A, IRB, ArgOffset),
1301 kShadowTLSAlignment);
1302 assert(Size != 0 && Store != 0);
1303 DEBUG(dbgs() << " Param:" << *Store << "\n");
1304 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
1306 DEBUG(dbgs() << " done with call args\n");
1309 cast<FunctionType>(CS.getCalledValue()->getType()-> getContainedType(0));
1310 if (FT->isVarArg()) {
1311 VAHelper->visitCallSite(CS, IRB);
1314 // Now, get the shadow for the RetVal.
1315 if (!I.getType()->isSized()) return;
1316 IRBuilder<> IRBBefore(&I);
1317 // Untill we have full dynamic coverage, make sure the retval shadow is 0.
1318 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
1319 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
1320 Instruction *NextInsn = 0;
1322 NextInsn = I.getNextNode();
1324 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
1325 if (!NormalDest->getSinglePredecessor()) {
1326 // FIXME: this case is tricky, so we are just conservative here.
1327 // Perhaps we need to split the edge between this BB and NormalDest,
1328 // but a naive attempt to use SplitEdge leads to a crash.
1329 setShadow(&I, getCleanShadow(&I));
1330 setOrigin(&I, getCleanOrigin());
1333 NextInsn = NormalDest->getFirstInsertionPt();
1335 "Could not find insertion point for retval shadow load");
1337 IRBuilder<> IRBAfter(NextInsn);
1338 Value *RetvalShadow =
1339 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
1340 kShadowTLSAlignment, "_msret");
1341 setShadow(&I, RetvalShadow);
1343 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
1346 void visitReturnInst(ReturnInst &I) {
1347 IRBuilder<> IRB(&I);
1348 if (Value *RetVal = I.getReturnValue()) {
1349 // Set the shadow for the RetVal.
1350 Value *Shadow = getShadow(RetVal);
1351 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
1352 DEBUG(dbgs() << "Return: " << *Shadow << "\n" << *ShadowPtr << "\n");
1353 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
1355 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
1359 void visitPHINode(PHINode &I) {
1360 IRBuilder<> IRB(&I);
1361 ShadowPHINodes.push_back(&I);
1362 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
1365 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
1369 void visitAllocaInst(AllocaInst &I) {
1370 setShadow(&I, getCleanShadow(&I));
1371 if (!ClPoisonStack) return;
1372 IRBuilder<> IRB(I.getNextNode());
1373 uint64_t Size = MS.TD->getTypeAllocSize(I.getAllocatedType());
1374 if (ClPoisonStackWithCall) {
1375 IRB.CreateCall2(MS.MsanPoisonStackFn,
1376 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1377 ConstantInt::get(MS.IntptrTy, Size));
1379 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
1380 IRB.CreateMemSet(ShadowBase, IRB.getInt8(ClPoisonStackPattern),
1381 Size, I.getAlignment());
1384 if (ClTrackOrigins) {
1385 setOrigin(&I, getCleanOrigin());
1386 SmallString<2048> StackDescriptionStorage;
1387 raw_svector_ostream StackDescription(StackDescriptionStorage);
1388 // We create a string with a description of the stack allocation and
1389 // pass it into __msan_set_alloca_origin.
1390 // It will be printed by the run-time if stack-originated UMR is found.
1391 // The first 4 bytes of the string are set to '----' and will be replaced
1392 // by __msan_va_arg_overflow_size_tls at the first call.
1393 StackDescription << "----" << I.getName() << "@" << F.getName();
1395 createPrivateNonConstGlobalForString(*F.getParent(),
1396 StackDescription.str());
1397 IRB.CreateCall3(MS.MsanSetAllocaOriginFn,
1398 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1399 ConstantInt::get(MS.IntptrTy, Size),
1400 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()));
1404 void visitSelectInst(SelectInst& I) {
1405 IRBuilder<> IRB(&I);
1406 setShadow(&I, IRB.CreateSelect(I.getCondition(),
1407 getShadow(I.getTrueValue()), getShadow(I.getFalseValue()),
1410 setOrigin(&I, IRB.CreateSelect(I.getCondition(),
1411 getOrigin(I.getTrueValue()), getOrigin(I.getFalseValue())));
1414 void visitLandingPadInst(LandingPadInst &I) {
1416 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
1417 setShadow(&I, getCleanShadow(&I));
1418 setOrigin(&I, getCleanOrigin());
1421 void visitGetElementPtrInst(GetElementPtrInst &I) {
1425 void visitExtractValueInst(ExtractValueInst &I) {
1426 IRBuilder<> IRB(&I);
1427 Value *Agg = I.getAggregateOperand();
1428 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
1429 Value *AggShadow = getShadow(Agg);
1430 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1431 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
1432 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
1433 setShadow(&I, ResShadow);
1434 setOrigin(&I, getCleanOrigin());
1437 void visitInsertValueInst(InsertValueInst &I) {
1438 IRBuilder<> IRB(&I);
1439 DEBUG(dbgs() << "InsertValue: " << I << "\n");
1440 Value *AggShadow = getShadow(I.getAggregateOperand());
1441 Value *InsShadow = getShadow(I.getInsertedValueOperand());
1442 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1443 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
1444 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
1445 DEBUG(dbgs() << " Res: " << *Res << "\n");
1447 setOrigin(&I, getCleanOrigin());
1450 void dumpInst(Instruction &I) {
1451 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1452 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
1454 errs() << "ZZZ " << I.getOpcodeName() << "\n";
1456 errs() << "QQQ " << I << "\n";
1459 void visitResumeInst(ResumeInst &I) {
1460 DEBUG(dbgs() << "Resume: " << I << "\n");
1461 // Nothing to do here.
1464 void visitInstruction(Instruction &I) {
1465 // Everything else: stop propagating and check for poisoned shadow.
1466 if (ClDumpStrictInstructions)
1468 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
1469 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
1470 insertCheck(I.getOperand(i), &I);
1471 setShadow(&I, getCleanShadow(&I));
1472 setOrigin(&I, getCleanOrigin());
1476 /// \brief AMD64-specific implementation of VarArgHelper.
1477 struct VarArgAMD64Helper : public VarArgHelper {
1478 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
1479 // See a comment in visitCallSite for more details.
1480 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
1481 static const unsigned AMD64FpEndOffset = 176;
1484 MemorySanitizer &MS;
1485 MemorySanitizerVisitor &MSV;
1486 Value *VAArgTLSCopy;
1487 Value *VAArgOverflowSize;
1489 SmallVector<CallInst*, 16> VAStartInstrumentationList;
1491 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
1492 MemorySanitizerVisitor &MSV)
1493 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
1495 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
1497 ArgKind classifyArgument(Value* arg) {
1498 // A very rough approximation of X86_64 argument classification rules.
1499 Type *T = arg->getType();
1500 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
1501 return AK_FloatingPoint;
1502 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
1503 return AK_GeneralPurpose;
1504 if (T->isPointerTy())
1505 return AK_GeneralPurpose;
1509 // For VarArg functions, store the argument shadow in an ABI-specific format
1510 // that corresponds to va_list layout.
1511 // We do this because Clang lowers va_arg in the frontend, and this pass
1512 // only sees the low level code that deals with va_list internals.
1513 // A much easier alternative (provided that Clang emits va_arg instructions)
1514 // would have been to associate each live instance of va_list with a copy of
1515 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
1517 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) {
1518 unsigned GpOffset = 0;
1519 unsigned FpOffset = AMD64GpEndOffset;
1520 unsigned OverflowOffset = AMD64FpEndOffset;
1521 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1522 ArgIt != End; ++ArgIt) {
1524 ArgKind AK = classifyArgument(A);
1525 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
1527 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
1531 case AK_GeneralPurpose:
1532 Base = getShadowPtrForVAArgument(A, IRB, GpOffset);
1535 case AK_FloatingPoint:
1536 Base = getShadowPtrForVAArgument(A, IRB, FpOffset);
1540 uint64_t ArgSize = MS.TD->getTypeAllocSize(A->getType());
1541 Base = getShadowPtrForVAArgument(A, IRB, OverflowOffset);
1542 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
1544 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
1546 Constant *OverflowSize =
1547 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
1548 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
1551 /// \brief Compute the shadow address for a given va_arg.
1552 Value *getShadowPtrForVAArgument(Value *A, IRBuilder<> &IRB,
1554 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
1555 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
1556 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(A), 0),
1560 void visitVAStartInst(VAStartInst &I) {
1561 IRBuilder<> IRB(&I);
1562 VAStartInstrumentationList.push_back(&I);
1563 Value *VAListTag = I.getArgOperand(0);
1564 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1566 // Unpoison the whole __va_list_tag.
1567 // FIXME: magic ABI constants.
1568 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1569 /* size */24, /* alignment */16, false);
1572 void visitVACopyInst(VACopyInst &I) {
1573 IRBuilder<> IRB(&I);
1574 Value *VAListTag = I.getArgOperand(0);
1575 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1577 // Unpoison the whole __va_list_tag.
1578 // FIXME: magic ABI constants.
1579 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1580 /* size */ 24, /* alignment */ 16, false);
1583 void finalizeInstrumentation() {
1584 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
1585 "finalizeInstrumentation called twice");
1586 if (!VAStartInstrumentationList.empty()) {
1587 // If there is a va_start in this function, make a backup copy of
1588 // va_arg_tls somewhere in the function entry block.
1589 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
1590 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
1592 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
1594 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
1595 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
1598 // Instrument va_start.
1599 // Copy va_list shadow from the backup copy of the TLS contents.
1600 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
1601 CallInst *OrigInst = VAStartInstrumentationList[i];
1602 IRBuilder<> IRB(OrigInst->getNextNode());
1603 Value *VAListTag = OrigInst->getArgOperand(0);
1605 Value *RegSaveAreaPtrPtr =
1607 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1608 ConstantInt::get(MS.IntptrTy, 16)),
1609 Type::getInt64PtrTy(*MS.C));
1610 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
1611 Value *RegSaveAreaShadowPtr =
1612 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
1613 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
1614 AMD64FpEndOffset, 16);
1616 Value *OverflowArgAreaPtrPtr =
1618 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1619 ConstantInt::get(MS.IntptrTy, 8)),
1620 Type::getInt64PtrTy(*MS.C));
1621 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
1622 Value *OverflowArgAreaShadowPtr =
1623 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
1625 getShadowPtrForVAArgument(VAArgTLSCopy, IRB, AMD64FpEndOffset);
1626 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
1631 VarArgHelper* CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
1632 MemorySanitizerVisitor &Visitor) {
1633 return new VarArgAMD64Helper(Func, Msan, Visitor);
1638 bool MemorySanitizer::runOnFunction(Function &F) {
1639 MemorySanitizerVisitor Visitor(F, *this);
1641 // Clear out readonly/readnone attributes.
1643 B.addAttribute(Attributes::ReadOnly)
1644 .addAttribute(Attributes::ReadNone);
1645 F.removeAttribute(AttributeSet::FunctionIndex,
1646 Attributes::get(F.getContext(), B));
1648 return Visitor.runOnFunction();