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 "BlackList.h"
51 #include "llvm/DataLayout.h"
52 #include "llvm/Function.h"
53 #include "llvm/InlineAsm.h"
54 #include "llvm/IntrinsicInst.h"
55 #include "llvm/IRBuilder.h"
56 #include "llvm/LLVMContext.h"
57 #include "llvm/MDBuilder.h"
58 #include "llvm/Module.h"
59 #include "llvm/Type.h"
60 #include "llvm/ADT/DepthFirstIterator.h"
61 #include "llvm/ADT/SmallString.h"
62 #include "llvm/ADT/SmallVector.h"
63 #include "llvm/ADT/ValueMap.h"
64 #include "llvm/Transforms/Instrumentation.h"
65 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
66 #include "llvm/Transforms/Utils/ModuleUtils.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Support/Compiler.h"
69 #include "llvm/Support/Debug.h"
70 #include "llvm/Support/InstVisitor.h"
71 #include "llvm/Support/raw_ostream.h"
72 #include "llvm/Transforms/Instrumentation.h"
73 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
74 #include "llvm/Transforms/Utils/ModuleUtils.h"
78 static const uint64_t kShadowMask32 = 1ULL << 31;
79 static const uint64_t kShadowMask64 = 1ULL << 46;
80 static const uint64_t kOriginOffset32 = 1ULL << 30;
81 static const uint64_t kOriginOffset64 = 1ULL << 45;
83 // This is an important flag that makes the reports much more
84 // informative at the cost of greater slowdown. Not fully implemented
86 // FIXME: this should be a top-level clang flag, e.g.
87 // -fmemory-sanitizer-full.
88 static cl::opt<bool> ClTrackOrigins("msan-track-origins",
89 cl::desc("Track origins (allocation sites) of poisoned memory"),
90 cl::Hidden, cl::init(false));
91 static cl::opt<bool> ClKeepGoing("msan-keep-going",
92 cl::desc("keep going after reporting a UMR"),
93 cl::Hidden, cl::init(false));
94 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
95 cl::desc("poison uninitialized stack variables"),
96 cl::Hidden, cl::init(true));
97 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
98 cl::desc("poison uninitialized stack variables with a call"),
99 cl::Hidden, cl::init(false));
100 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
101 cl::desc("poison uninitialized stack variables with the given patter"),
102 cl::Hidden, cl::init(0xff));
104 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
105 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
106 cl::Hidden, cl::init(true));
108 // This flag controls whether we check the shadow of the address
109 // operand of load or store. Such bugs are very rare, since load from
110 // a garbage address typically results in SEGV, but still happen
111 // (e.g. only lower bits of address are garbage, or the access happens
112 // early at program startup where malloc-ed memory is more likely to
113 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
114 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
115 cl::desc("report accesses through a pointer which has poisoned shadow"),
116 cl::Hidden, cl::init(true));
118 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
119 cl::desc("print out instructions with default strict semantics"),
120 cl::Hidden, cl::init(false));
122 static cl::opt<std::string> ClBlackListFile("msan-blacklist",
123 cl::desc("File containing the list of functions where MemorySanitizer "
124 "should not report bugs"), cl::Hidden);
128 /// \brief An instrumentation pass implementing detection of uninitialized
131 /// MemorySanitizer: instrument the code in module to find
132 /// uninitialized reads.
133 class MemorySanitizer : public FunctionPass {
135 MemorySanitizer() : FunctionPass(ID), TD(0) { }
136 const char *getPassName() const { return "MemorySanitizer"; }
137 bool runOnFunction(Function &F);
138 bool doInitialization(Module &M);
139 static char ID; // Pass identification, replacement for typeid.
146 /// \brief Thread-local shadow storage for function parameters.
147 GlobalVariable *ParamTLS;
148 /// \brief Thread-local origin storage for function parameters.
149 GlobalVariable *ParamOriginTLS;
150 /// \brief Thread-local shadow storage for function return value.
151 GlobalVariable *RetvalTLS;
152 /// \brief Thread-local origin storage for function return value.
153 GlobalVariable *RetvalOriginTLS;
154 /// \brief Thread-local shadow storage for in-register va_arg function
155 /// parameters (x86_64-specific).
156 GlobalVariable *VAArgTLS;
157 /// \brief Thread-local shadow storage for va_arg overflow area
158 /// (x86_64-specific).
159 GlobalVariable *VAArgOverflowSizeTLS;
160 /// \brief Thread-local space used to pass origin value to the UMR reporting
162 GlobalVariable *OriginTLS;
164 /// \brief The run-time callback to print a warning.
166 /// \brief Run-time helper that copies origin info for a memory range.
167 Value *MsanCopyOriginFn;
168 /// \brief Run-time helper that generates a new origin value for a stack
170 Value *MsanSetAllocaOriginFn;
171 /// \brief Run-time helper that poisons stack on function entry.
172 Value *MsanPoisonStackFn;
173 /// \brief The actual "memmove" function.
176 /// \brief Address mask used in application-to-shadow address calculation.
177 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
179 /// \brief Offset of the origin shadow from the "normal" shadow.
180 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
181 uint64_t OriginOffset;
182 /// \brief Branch weights for error reporting.
183 MDNode *ColdCallWeights;
184 /// \brief The blacklist.
185 OwningPtr<BlackList> BL;
187 friend class MemorySanitizerVisitor;
188 friend class VarArgAMD64Helper;
192 char MemorySanitizer::ID = 0;
193 INITIALIZE_PASS(MemorySanitizer, "msan",
194 "MemorySanitizer: detects uninitialized reads.",
197 FunctionPass *llvm::createMemorySanitizerPass() {
198 return new MemorySanitizer();
201 /// \brief Create a non-const global initialized with the given string.
203 /// Creates a writable global for Str so that we can pass it to the
204 /// run-time lib. Runtime uses first 4 bytes of the string to store the
205 /// frame ID, so the string needs to be mutable.
206 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
208 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
209 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
210 GlobalValue::PrivateLinkage, StrConst, "");
213 /// \brief Module-level initialization.
215 /// Obtains pointers to the required runtime library functions, and
216 /// inserts a call to __msan_init to the module's constructor list.
217 bool MemorySanitizer::doInitialization(Module &M) {
218 TD = getAnalysisIfAvailable<DataLayout>();
221 BL.reset(new BlackList(ClBlackListFile));
222 C = &(M.getContext());
223 unsigned PtrSize = TD->getPointerSizeInBits(/* AddressSpace */0);
226 ShadowMask = kShadowMask64;
227 OriginOffset = kOriginOffset64;
230 ShadowMask = kShadowMask32;
231 OriginOffset = kOriginOffset32;
234 report_fatal_error("unsupported pointer size");
239 IntptrTy = IRB.getIntPtrTy(TD);
240 OriginTy = IRB.getInt32Ty();
242 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
244 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
245 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
246 "__msan_init", IRB.getVoidTy(), NULL)), 0);
248 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::LinkOnceODRLinkage,
249 IRB.getInt32(ClTrackOrigins), "__msan_track_origins");
251 // Create the callback.
252 // FIXME: this function should have "Cold" calling conv,
253 // which is not yet implemented.
254 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
255 : "__msan_warning_noreturn";
256 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
258 MsanCopyOriginFn = M.getOrInsertFunction(
259 "__msan_copy_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(),
260 IRB.getInt8PtrTy(), IntptrTy, NULL);
261 MsanSetAllocaOriginFn = M.getOrInsertFunction(
262 "__msan_set_alloca_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
263 IRB.getInt8PtrTy(), NULL);
264 MsanPoisonStackFn = M.getOrInsertFunction(
265 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
266 MemmoveFn = M.getOrInsertFunction(
267 "memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
271 RetvalTLS = new GlobalVariable(
272 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
273 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
274 GlobalVariable::GeneralDynamicTLSModel);
275 RetvalOriginTLS = new GlobalVariable(
276 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
277 "__msan_retval_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
279 ParamTLS = new GlobalVariable(
280 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
281 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
282 GlobalVariable::GeneralDynamicTLSModel);
283 ParamOriginTLS = new GlobalVariable(
284 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
285 0, "__msan_param_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
287 VAArgTLS = new GlobalVariable(
288 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
289 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
290 GlobalVariable::GeneralDynamicTLSModel);
291 VAArgOverflowSizeTLS = new GlobalVariable(
292 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
293 "__msan_va_arg_overflow_size_tls", 0,
294 GlobalVariable::GeneralDynamicTLSModel);
295 OriginTLS = new GlobalVariable(
296 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
297 "__msan_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
303 /// \brief A helper class that handles instrumentation of VarArg
304 /// functions on a particular platform.
306 /// Implementations are expected to insert the instrumentation
307 /// necessary to propagate argument shadow through VarArg function
308 /// calls. Visit* methods are called during an InstVisitor pass over
309 /// the function, and should avoid creating new basic blocks. A new
310 /// instance of this class is created for each instrumented function.
311 struct VarArgHelper {
312 /// \brief Visit a CallSite.
313 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
315 /// \brief Visit a va_start call.
316 virtual void visitVAStartInst(VAStartInst &I) = 0;
318 /// \brief Visit a va_copy call.
319 virtual void visitVACopyInst(VACopyInst &I) = 0;
321 /// \brief Finalize function instrumentation.
323 /// This method is called after visiting all interesting (see above)
324 /// instructions in a function.
325 virtual void finalizeInstrumentation() = 0;
328 struct MemorySanitizerVisitor;
331 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
332 MemorySanitizerVisitor &Visitor);
334 /// This class does all the work for a given function. Store and Load
335 /// instructions store and load corresponding shadow and origin
336 /// values. Most instructions propagate shadow from arguments to their
337 /// return values. Certain instructions (most importantly, BranchInst)
338 /// test their argument shadow and print reports (with a runtime call) if it's
340 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
343 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
344 ValueMap<Value*, Value*> ShadowMap, OriginMap;
346 OwningPtr<VarArgHelper> VAHelper;
348 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
349 // See a comment in visitCallSite for more details.
350 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
351 static const unsigned AMD64FpEndOffset = 176;
353 struct ShadowOriginAndInsertPoint {
356 Instruction *OrigIns;
357 ShadowOriginAndInsertPoint(Instruction *S, Instruction *O, Instruction *I)
358 : Shadow(S), Origin(O), OrigIns(I) { }
359 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
361 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
363 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
364 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
365 InsertChecks = !MS.BL->isIn(F);
366 DEBUG(if (!InsertChecks)
367 dbgs() << "MemorySanitizer is not inserting checks into '"
368 << F.getName() << "'\n");
371 void materializeChecks() {
372 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
373 Instruction *Shadow = InstrumentationList[i].Shadow;
374 Instruction *OrigIns = InstrumentationList[i].OrigIns;
375 IRBuilder<> IRB(OrigIns);
376 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
377 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
378 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
379 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
380 getCleanShadow(ConvertedShadow), "_mscmp");
381 Instruction *CheckTerm =
382 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp),
383 /* Unreachable */ !ClKeepGoing,
386 IRB.SetInsertPoint(CheckTerm);
387 if (ClTrackOrigins) {
388 Instruction *Origin = InstrumentationList[i].Origin;
389 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
392 CallInst *Call = IRB.CreateCall(MS.WarningFn);
393 Call->setDebugLoc(OrigIns->getDebugLoc());
394 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
396 DEBUG(dbgs() << "DONE:\n" << F);
399 /// \brief Add MemorySanitizer instrumentation to a function.
400 bool runOnFunction() {
401 if (!MS.TD) return false;
402 // Iterate all BBs in depth-first order and create shadow instructions
403 // for all instructions (where applicable).
404 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
405 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
406 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
407 BasicBlock *BB = *DI;
411 // Finalize PHI nodes.
412 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
413 PHINode *PN = ShadowPHINodes[i];
414 PHINode *PNS = cast<PHINode>(getShadow(PN));
415 PHINode *PNO = ClTrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
416 size_t NumValues = PN->getNumIncomingValues();
417 for (size_t v = 0; v < NumValues; v++) {
418 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
420 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
424 VAHelper->finalizeInstrumentation();
431 /// \brief Compute the shadow type that corresponds to a given Value.
432 Type *getShadowTy(Value *V) {
433 return getShadowTy(V->getType());
436 /// \brief Compute the shadow type that corresponds to a given Type.
437 Type *getShadowTy(Type *OrigTy) {
438 if (!OrigTy->isSized()) {
441 // For integer type, shadow is the same as the original type.
442 // This may return weird-sized types like i1.
443 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
445 if (VectorType *VT = dyn_cast<VectorType>(OrigTy))
446 return VectorType::getInteger(VT);
447 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
448 SmallVector<Type*, 4> Elements;
449 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
450 Elements.push_back(getShadowTy(ST->getElementType(i)));
451 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
452 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
455 uint32_t TypeSize = MS.TD->getTypeStoreSizeInBits(OrigTy);
456 return IntegerType::get(*MS.C, TypeSize);
459 /// \brief Flatten a vector type.
460 Type *getShadowTyNoVec(Type *ty) {
461 if (VectorType *vt = dyn_cast<VectorType>(ty))
462 return IntegerType::get(*MS.C, vt->getBitWidth());
466 /// \brief Convert a shadow value to it's flattened variant.
467 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
468 Type *Ty = V->getType();
469 Type *NoVecTy = getShadowTyNoVec(Ty);
470 if (Ty == NoVecTy) return V;
471 return IRB.CreateBitCast(V, NoVecTy);
474 /// \brief Compute the shadow address that corresponds to a given application
477 /// Shadow = Addr & ~ShadowMask.
478 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
481 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
482 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
483 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
486 /// \brief Compute the origin address that corresponds to a given application
489 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
490 /// = Addr & (~ShadowMask & ~3ULL) + OriginOffset
491 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
493 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
494 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask & ~3ULL));
496 IRB.CreateAdd(ShadowLong,
497 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
498 return IRB.CreateIntToPtr(Add, PointerType::get(IRB.getInt32Ty(), 0));
501 /// \brief Compute the shadow address for a given function argument.
503 /// Shadow = ParamTLS+ArgOffset.
504 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
506 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
507 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
508 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
512 /// \brief Compute the origin address for a given function argument.
513 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
515 if (!ClTrackOrigins) return 0;
516 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
517 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
518 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
522 /// \brief Compute the shadow address for a retval.
523 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
524 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
525 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
529 /// \brief Compute the origin address for a retval.
530 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
531 // We keep a single origin for the entire retval. Might be too optimistic.
532 return MS.RetvalOriginTLS;
535 /// \brief Set SV to be the shadow value for V.
536 void setShadow(Value *V, Value *SV) {
537 assert(!ShadowMap.count(V) && "Values may only have one shadow");
541 /// \brief Set Origin to be the origin value for V.
542 void setOrigin(Value *V, Value *Origin) {
543 if (!ClTrackOrigins) return;
544 assert(!OriginMap.count(V) && "Values may only have one origin");
545 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
546 OriginMap[V] = Origin;
549 /// \brief Create a clean shadow value for a given value.
551 /// Clean shadow (all zeroes) means all bits of the value are defined
553 Value *getCleanShadow(Value *V) {
554 Type *ShadowTy = getShadowTy(V);
557 return Constant::getNullValue(ShadowTy);
560 /// \brief Create a dirty shadow of a given shadow type.
561 Constant *getPoisonedShadow(Type *ShadowTy) {
563 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
564 return Constant::getAllOnesValue(ShadowTy);
565 StructType *ST = cast<StructType>(ShadowTy);
566 SmallVector<Constant *, 4> Vals;
567 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
568 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
569 return ConstantStruct::get(ST, Vals);
572 /// \brief Create a clean (zero) origin.
573 Value *getCleanOrigin() {
574 return Constant::getNullValue(MS.OriginTy);
577 /// \brief Get the shadow value for a given Value.
579 /// This function either returns the value set earlier with setShadow,
580 /// or extracts if from ParamTLS (for function arguments).
581 Value *getShadow(Value *V) {
582 if (Instruction *I = dyn_cast<Instruction>(V)) {
583 // For instructions the shadow is already stored in the map.
584 Value *Shadow = ShadowMap[V];
586 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
587 assert(Shadow && "No shadow for a value");
591 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
592 Value *AllOnes = getPoisonedShadow(getShadowTy(V));
593 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
596 if (Argument *A = dyn_cast<Argument>(V)) {
597 // For arguments we compute the shadow on demand and store it in the map.
598 Value **ShadowPtr = &ShadowMap[V];
601 Function *F = A->getParent();
602 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
603 unsigned ArgOffset = 0;
604 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
606 if (!AI->getType()->isSized()) {
607 DEBUG(dbgs() << "Arg is not sized\n");
610 unsigned Size = AI->hasByValAttr()
611 ? MS.TD->getTypeAllocSize(AI->getType()->getPointerElementType())
612 : MS.TD->getTypeAllocSize(AI->getType());
614 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
615 if (AI->hasByValAttr()) {
616 // ByVal pointer itself has clean shadow. We copy the actual
617 // argument shadow to the underlying memory.
618 Value *Cpy = EntryIRB.CreateMemCpy(
619 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
620 Base, Size, AI->getParamAlignment());
621 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
622 *ShadowPtr = getCleanShadow(V);
624 *ShadowPtr = EntryIRB.CreateLoad(Base);
626 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
627 **ShadowPtr << "\n");
628 if (ClTrackOrigins) {
629 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
630 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
633 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
635 assert(*ShadowPtr && "Could not find shadow for an argument");
638 // For everything else the shadow is zero.
639 return getCleanShadow(V);
642 /// \brief Get the shadow for i-th argument of the instruction I.
643 Value *getShadow(Instruction *I, int i) {
644 return getShadow(I->getOperand(i));
647 /// \brief Get the origin for a value.
648 Value *getOrigin(Value *V) {
649 if (!ClTrackOrigins) return 0;
650 if (isa<Instruction>(V) || isa<Argument>(V)) {
651 Value *Origin = OriginMap[V];
653 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
654 Origin = getCleanOrigin();
658 return getCleanOrigin();
661 /// \brief Get the origin for i-th argument of the instruction I.
662 Value *getOrigin(Instruction *I, int i) {
663 return getOrigin(I->getOperand(i));
666 /// \brief Remember the place where a shadow check should be inserted.
668 /// This location will be later instrumented with a check that will print a
669 /// UMR warning in runtime if the value is not fully defined.
670 void insertCheck(Value *Val, Instruction *OrigIns) {
672 if (!InsertChecks) return;
673 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
675 Type *ShadowTy = Shadow->getType();
676 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
677 "Can only insert checks for integer and vector shadow types");
678 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
679 InstrumentationList.push_back(
680 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
683 //------------------- Visitors.
685 /// \brief Instrument LoadInst
687 /// Loads the corresponding shadow and (optionally) origin.
688 /// Optionally, checks that the load address is fully defined.
689 void visitLoadInst(LoadInst &I) {
690 Type *LoadTy = I.getType();
691 assert(LoadTy->isSized() && "Load type must have size");
693 Type *ShadowTy = getShadowTy(&I);
694 Value *Addr = I.getPointerOperand();
695 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
696 setShadow(&I, IRB.CreateLoad(ShadowPtr, "_msld"));
698 if (ClCheckAccessAddress)
699 insertCheck(I.getPointerOperand(), &I);
702 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
705 /// \brief Instrument StoreInst
707 /// Stores the corresponding shadow and (optionally) origin.
708 /// Optionally, checks that the store address is fully defined.
709 /// Volatile stores check that the value being stored is fully defined.
710 void visitStoreInst(StoreInst &I) {
712 Value *Val = I.getValueOperand();
713 Value *Addr = I.getPointerOperand();
714 Value *Shadow = getShadow(Val);
715 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
717 StoreInst *NewSI = IRB.CreateStore(Shadow, ShadowPtr);
718 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
719 // If the store is volatile, add a check.
721 insertCheck(Val, &I);
722 if (ClCheckAccessAddress)
723 insertCheck(Addr, &I);
726 IRB.CreateStore(getOrigin(Val), getOriginPtr(Addr, IRB));
730 void visitSExtInst(SExtInst &I) {
732 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
733 setOrigin(&I, getOrigin(&I, 0));
736 void visitZExtInst(ZExtInst &I) {
738 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
739 setOrigin(&I, getOrigin(&I, 0));
742 void visitTruncInst(TruncInst &I) {
744 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
745 setOrigin(&I, getOrigin(&I, 0));
748 void visitBitCastInst(BitCastInst &I) {
750 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
751 setOrigin(&I, getOrigin(&I, 0));
754 void visitPtrToIntInst(PtrToIntInst &I) {
756 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
757 "_msprop_ptrtoint"));
758 setOrigin(&I, getOrigin(&I, 0));
761 void visitIntToPtrInst(IntToPtrInst &I) {
763 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
764 "_msprop_inttoptr"));
765 setOrigin(&I, getOrigin(&I, 0));
768 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
769 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
770 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
771 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
772 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
773 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
775 /// \brief Propagate shadow for bitwise AND.
777 /// This code is exact, i.e. if, for example, a bit in the left argument
778 /// is defined and 0, then neither the value not definedness of the
779 /// corresponding bit in B don't affect the resulting shadow.
780 void visitAnd(BinaryOperator &I) {
782 // "And" of 0 and a poisoned value results in unpoisoned value.
783 // 1&1 => 1; 0&1 => 0; p&1 => p;
784 // 1&0 => 0; 0&0 => 0; p&0 => 0;
785 // 1&p => p; 0&p => 0; p&p => p;
786 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
787 Value *S1 = getShadow(&I, 0);
788 Value *S2 = getShadow(&I, 1);
789 Value *V1 = I.getOperand(0);
790 Value *V2 = I.getOperand(1);
791 if (V1->getType() != S1->getType()) {
792 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
793 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
795 Value *S1S2 = IRB.CreateAnd(S1, S2);
796 Value *V1S2 = IRB.CreateAnd(V1, S2);
797 Value *S1V2 = IRB.CreateAnd(S1, V2);
798 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
799 setOriginForNaryOp(I);
802 void visitOr(BinaryOperator &I) {
804 // "Or" of 1 and a poisoned value results in unpoisoned value.
805 // 1|1 => 1; 0|1 => 1; p|1 => 1;
806 // 1|0 => 1; 0|0 => 0; p|0 => p;
807 // 1|p => 1; 0|p => p; p|p => p;
808 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
809 Value *S1 = getShadow(&I, 0);
810 Value *S2 = getShadow(&I, 1);
811 Value *V1 = IRB.CreateNot(I.getOperand(0));
812 Value *V2 = IRB.CreateNot(I.getOperand(1));
813 if (V1->getType() != S1->getType()) {
814 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
815 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
817 Value *S1S2 = IRB.CreateAnd(S1, S2);
818 Value *V1S2 = IRB.CreateAnd(V1, S2);
819 Value *S1V2 = IRB.CreateAnd(S1, V2);
820 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
821 setOriginForNaryOp(I);
824 /// \brief Propagate origin for an instruction.
826 /// This is a general case of origin propagation. For an Nary operation,
827 /// is set to the origin of an argument that is not entirely initialized.
828 /// It does not matter which one is picked if all arguments are initialized.
829 void setOriginForNaryOp(Instruction &I) {
830 if (!ClTrackOrigins) return;
832 Value *Origin = getOrigin(&I, 0);
833 for (unsigned Op = 1, n = I.getNumOperands(); Op < n; ++Op) {
834 Value *S = convertToShadowTyNoVec(getShadow(&I, Op - 1), IRB);
835 Origin = IRB.CreateSelect(IRB.CreateICmpNE(S, getCleanShadow(S)),
836 Origin, getOrigin(&I, Op));
838 setOrigin(&I, Origin);
841 /// \brief Propagate shadow for a binary operation.
843 /// Shadow = Shadow0 | Shadow1, all 3 must have the same type.
844 /// Bitwise OR is selected as an operation that will never lose even a bit of
846 void handleShadowOrBinary(Instruction &I) {
848 Value *Shadow0 = getShadow(&I, 0);
849 Value *Shadow1 = getShadow(&I, 1);
850 setShadow(&I, IRB.CreateOr(Shadow0, Shadow1, "_msprop"));
851 setOriginForNaryOp(I);
854 /// \brief Propagate shadow for arbitrary operation.
856 /// This is a general case of shadow propagation, used in all cases where we
857 /// don't know and/or care about what the operation actually does.
858 /// It converts all input shadow values to a common type (extending or
859 /// truncating as necessary), and bitwise OR's them.
861 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
862 /// fully initialized), and less prone to false positives.
863 // FIXME: is the casting actually correct?
864 // FIXME: merge this with handleShadowOrBinary.
865 void handleShadowOr(Instruction &I) {
867 Value *Shadow = getShadow(&I, 0);
868 for (unsigned Op = 1, n = I.getNumOperands(); Op < n; ++Op)
869 Shadow = IRB.CreateOr(
870 Shadow, IRB.CreateIntCast(getShadow(&I, Op), Shadow->getType(), false),
872 Shadow = IRB.CreateIntCast(Shadow, getShadowTy(&I), false);
873 setShadow(&I, Shadow);
874 setOriginForNaryOp(I);
877 void visitFAdd(BinaryOperator &I) { handleShadowOrBinary(I); }
878 void visitFSub(BinaryOperator &I) { handleShadowOrBinary(I); }
879 void visitFMul(BinaryOperator &I) { handleShadowOrBinary(I); }
880 void visitAdd(BinaryOperator &I) { handleShadowOrBinary(I); }
881 void visitSub(BinaryOperator &I) { handleShadowOrBinary(I); }
882 void visitXor(BinaryOperator &I) { handleShadowOrBinary(I); }
883 void visitMul(BinaryOperator &I) { handleShadowOrBinary(I); }
885 void handleDiv(Instruction &I) {
887 // Strict on the second argument.
888 insertCheck(I.getOperand(1), &I);
889 setShadow(&I, getShadow(&I, 0));
890 setOrigin(&I, getOrigin(&I, 0));
893 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
894 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
895 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
896 void visitURem(BinaryOperator &I) { handleDiv(I); }
897 void visitSRem(BinaryOperator &I) { handleDiv(I); }
898 void visitFRem(BinaryOperator &I) { handleDiv(I); }
900 /// \brief Instrument == and != comparisons.
902 /// Sometimes the comparison result is known even if some of the bits of the
903 /// arguments are not.
904 void handleEqualityComparison(ICmpInst &I) {
906 Value *A = I.getOperand(0);
907 Value *B = I.getOperand(1);
908 Value *Sa = getShadow(A);
909 Value *Sb = getShadow(B);
910 if (A->getType()->isPointerTy())
911 A = IRB.CreatePointerCast(A, MS.IntptrTy);
912 if (B->getType()->isPointerTy())
913 B = IRB.CreatePointerCast(B, MS.IntptrTy);
914 // A == B <==> (C = A^B) == 0
915 // A != B <==> (C = A^B) != 0
917 Value *C = IRB.CreateXor(A, B);
918 Value *Sc = IRB.CreateOr(Sa, Sb);
919 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
920 // Result is defined if one of the following is true
921 // * there is a defined 1 bit in C
922 // * C is fully defined
923 // Si = !(C & ~Sc) && Sc
924 Value *Zero = Constant::getNullValue(Sc->getType());
925 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
927 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
929 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
930 Si->setName("_msprop_icmp");
932 setOriginForNaryOp(I);
935 void visitICmpInst(ICmpInst &I) {
936 if (ClHandleICmp && I.isEquality())
937 handleEqualityComparison(I);
942 void visitFCmpInst(FCmpInst &I) {
946 void handleShift(BinaryOperator &I) {
948 // If any of the S2 bits are poisoned, the whole thing is poisoned.
949 // Otherwise perform the same shift on S1.
950 Value *S1 = getShadow(&I, 0);
951 Value *S2 = getShadow(&I, 1);
952 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
954 Value *V2 = I.getOperand(1);
955 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
956 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
957 setOriginForNaryOp(I);
960 void visitShl(BinaryOperator &I) { handleShift(I); }
961 void visitAShr(BinaryOperator &I) { handleShift(I); }
962 void visitLShr(BinaryOperator &I) { handleShift(I); }
964 void visitMemSetInst(MemSetInst &I) {
966 Value *Ptr = I.getArgOperand(0);
967 Value *Val = I.getArgOperand(1);
968 Value *ShadowPtr = getShadowPtr(Ptr, Val->getType(), IRB);
969 Value *ShadowVal = getCleanShadow(Val);
970 Value *Size = I.getArgOperand(2);
971 unsigned Align = I.getAlignment();
972 bool isVolatile = I.isVolatile();
974 IRB.CreateMemSet(ShadowPtr, ShadowVal, Size, Align, isVolatile);
977 void visitMemCpyInst(MemCpyInst &I) {
979 Value *Dst = I.getArgOperand(0);
980 Value *Src = I.getArgOperand(1);
981 Type *ElementType = dyn_cast<PointerType>(Dst->getType())->getElementType();
982 Value *ShadowDst = getShadowPtr(Dst, ElementType, IRB);
983 Value *ShadowSrc = getShadowPtr(Src, ElementType, IRB);
984 Value *Size = I.getArgOperand(2);
985 unsigned Align = I.getAlignment();
986 bool isVolatile = I.isVolatile();
988 IRB.CreateMemCpy(ShadowDst, ShadowSrc, Size, Align, isVolatile);
990 IRB.CreateCall3(MS.MsanCopyOriginFn, Dst, Src, Size);
993 /// \brief Instrument llvm.memmove
995 /// At this point we don't know if llvm.memmove will be inlined or not.
996 /// If we don't instrument it and it gets inlined,
997 /// our interceptor will not kick in and we will lose the memmove.
998 /// If we instrument the call here, but it does not get inlined,
999 /// we will memove the shadow twice: which is bad in case
1000 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1002 /// Similar situation exists for memcpy and memset, but for those functions
1003 /// calling instrumentation twice does not lead to incorrect results.
1004 void visitMemMoveInst(MemMoveInst &I) {
1005 IRBuilder<> IRB(&I);
1008 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1009 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1010 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1011 I.eraseFromParent();
1014 void visitVAStartInst(VAStartInst &I) {
1015 VAHelper->visitVAStartInst(I);
1018 void visitVACopyInst(VACopyInst &I) {
1019 VAHelper->visitVACopyInst(I);
1022 void visitCallSite(CallSite CS) {
1023 Instruction &I = *CS.getInstruction();
1024 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
1026 // Allow only tail calls with the same types, otherwise
1027 // we may have a false positive: shadow for a non-void RetVal
1028 // will get propagated to a void RetVal.
1029 CallInst *Call = cast<CallInst>(&I);
1030 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
1031 Call->setTailCall(false);
1032 if (isa<IntrinsicInst>(&I)) {
1033 // All intrinsics we care about are handled in corresponding visit*
1034 // methods. Add checks for the arguments, mark retval as clean.
1035 visitInstruction(I);
1039 IRBuilder<> IRB(&I);
1040 unsigned ArgOffset = 0;
1041 DEBUG(dbgs() << " CallSite: " << I << "\n");
1042 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1043 ArgIt != End; ++ArgIt) {
1045 unsigned i = ArgIt - CS.arg_begin();
1046 if (!A->getType()->isSized()) {
1047 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
1052 // Compute the Shadow for arg even if it is ByVal, because
1053 // in that case getShadow() will copy the actual arg shadow to
1054 // __msan_param_tls.
1055 Value *ArgShadow = getShadow(A);
1056 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
1057 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
1058 " Shadow: " << *ArgShadow << "\n");
1059 if (CS.paramHasAttr(i + 1, Attributes::ByVal)) {
1060 assert(A->getType()->isPointerTy() &&
1061 "ByVal argument is not a pointer!");
1062 Size = MS.TD->getTypeAllocSize(A->getType()->getPointerElementType());
1063 unsigned Alignment = CS.getParamAlignment(i + 1);
1064 Store = IRB.CreateMemCpy(ArgShadowBase,
1065 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
1068 Size = MS.TD->getTypeAllocSize(A->getType());
1069 Store = IRB.CreateStore(ArgShadow, ArgShadowBase);
1072 IRB.CreateStore(getOrigin(A),
1073 getOriginPtrForArgument(A, IRB, ArgOffset));
1074 assert(Size != 0 && Store != 0);
1075 DEBUG(dbgs() << " Param:" << *Store << "\n");
1076 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
1078 DEBUG(dbgs() << " done with call args\n");
1081 cast<FunctionType>(CS.getCalledValue()->getType()-> getContainedType(0));
1082 if (FT->isVarArg()) {
1083 VAHelper->visitCallSite(CS, IRB);
1086 // Now, get the shadow for the RetVal.
1087 if (!I.getType()->isSized()) return;
1088 IRBuilder<> IRBBefore(&I);
1089 // Untill we have full dynamic coverage, make sure the retval shadow is 0.
1090 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
1091 IRBBefore.CreateStore(getCleanShadow(&I), Base);
1092 Instruction *NextInsn = 0;
1094 NextInsn = I.getNextNode();
1096 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
1097 if (!NormalDest->getSinglePredecessor()) {
1098 // FIXME: this case is tricky, so we are just conservative here.
1099 // Perhaps we need to split the edge between this BB and NormalDest,
1100 // but a naive attempt to use SplitEdge leads to a crash.
1101 setShadow(&I, getCleanShadow(&I));
1102 setOrigin(&I, getCleanOrigin());
1105 NextInsn = NormalDest->getFirstInsertionPt();
1107 "Could not find insertion point for retval shadow load");
1109 IRBuilder<> IRBAfter(NextInsn);
1110 setShadow(&I, IRBAfter.CreateLoad(getShadowPtrForRetval(&I, IRBAfter),
1113 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
1116 void visitReturnInst(ReturnInst &I) {
1117 IRBuilder<> IRB(&I);
1118 if (Value *RetVal = I.getReturnValue()) {
1119 // Set the shadow for the RetVal.
1120 Value *Shadow = getShadow(RetVal);
1121 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
1122 DEBUG(dbgs() << "Return: " << *Shadow << "\n" << *ShadowPtr << "\n");
1123 IRB.CreateStore(Shadow, ShadowPtr);
1125 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
1129 void visitPHINode(PHINode &I) {
1130 IRBuilder<> IRB(&I);
1131 ShadowPHINodes.push_back(&I);
1132 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
1135 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
1139 void visitAllocaInst(AllocaInst &I) {
1140 setShadow(&I, getCleanShadow(&I));
1141 if (!ClPoisonStack) return;
1142 IRBuilder<> IRB(I.getNextNode());
1143 uint64_t Size = MS.TD->getTypeAllocSize(I.getAllocatedType());
1144 if (ClPoisonStackWithCall) {
1145 IRB.CreateCall2(MS.MsanPoisonStackFn,
1146 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1147 ConstantInt::get(MS.IntptrTy, Size));
1149 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
1150 IRB.CreateMemSet(ShadowBase, IRB.getInt8(ClPoisonStackPattern),
1151 Size, I.getAlignment());
1154 if (ClTrackOrigins) {
1155 setOrigin(&I, getCleanOrigin());
1156 SmallString<2048> StackDescriptionStorage;
1157 raw_svector_ostream StackDescription(StackDescriptionStorage);
1158 // We create a string with a description of the stack allocation and
1159 // pass it into __msan_set_alloca_origin.
1160 // It will be printed by the run-time if stack-originated UMR is found.
1161 // The first 4 bytes of the string are set to '----' and will be replaced
1162 // by __msan_va_arg_overflow_size_tls at the first call.
1163 StackDescription << "----" << I.getName() << "@" << F.getName();
1165 createPrivateNonConstGlobalForString(*F.getParent(),
1166 StackDescription.str());
1167 IRB.CreateCall3(MS.MsanSetAllocaOriginFn,
1168 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1169 ConstantInt::get(MS.IntptrTy, Size),
1170 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()));
1174 void visitSelectInst(SelectInst& I) {
1175 IRBuilder<> IRB(&I);
1176 setShadow(&I, IRB.CreateSelect(I.getCondition(),
1177 getShadow(I.getTrueValue()), getShadow(I.getFalseValue()),
1180 setOrigin(&I, IRB.CreateSelect(I.getCondition(),
1181 getOrigin(I.getTrueValue()), getOrigin(I.getFalseValue())));
1184 void visitLandingPadInst(LandingPadInst &I) {
1186 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
1187 setShadow(&I, getCleanShadow(&I));
1188 setOrigin(&I, getCleanOrigin());
1191 void visitGetElementPtrInst(GetElementPtrInst &I) {
1195 void visitExtractValueInst(ExtractValueInst &I) {
1196 IRBuilder<> IRB(&I);
1197 Value *Agg = I.getAggregateOperand();
1198 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
1199 Value *AggShadow = getShadow(Agg);
1200 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1201 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
1202 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
1203 setShadow(&I, ResShadow);
1204 setOrigin(&I, getCleanOrigin());
1207 void visitInsertValueInst(InsertValueInst &I) {
1208 IRBuilder<> IRB(&I);
1209 DEBUG(dbgs() << "InsertValue: " << I << "\n");
1210 Value *AggShadow = getShadow(I.getAggregateOperand());
1211 Value *InsShadow = getShadow(I.getInsertedValueOperand());
1212 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1213 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
1214 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
1215 DEBUG(dbgs() << " Res: " << *Res << "\n");
1217 setOrigin(&I, getCleanOrigin());
1220 void dumpInst(Instruction &I) {
1221 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1222 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
1224 errs() << "ZZZ " << I.getOpcodeName() << "\n";
1226 errs() << "QQQ " << I << "\n";
1229 void visitResumeInst(ResumeInst &I) {
1230 DEBUG(dbgs() << "Resume: " << I << "\n");
1231 // Nothing to do here.
1234 void visitInstruction(Instruction &I) {
1235 // Everything else: stop propagating and check for poisoned shadow.
1236 if (ClDumpStrictInstructions)
1238 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
1239 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
1240 insertCheck(I.getOperand(i), &I);
1241 setShadow(&I, getCleanShadow(&I));
1242 setOrigin(&I, getCleanOrigin());
1246 /// \brief AMD64-specific implementation of VarArgHelper.
1247 struct VarArgAMD64Helper : public VarArgHelper {
1248 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
1249 // See a comment in visitCallSite for more details.
1250 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
1251 static const unsigned AMD64FpEndOffset = 176;
1254 MemorySanitizer &MS;
1255 MemorySanitizerVisitor &MSV;
1256 Value *VAArgTLSCopy;
1257 Value *VAArgOverflowSize;
1259 SmallVector<CallInst*, 16> VAStartInstrumentationList;
1261 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
1262 MemorySanitizerVisitor &MSV)
1263 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
1265 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
1267 ArgKind classifyArgument(Value* arg) {
1268 // A very rough approximation of X86_64 argument classification rules.
1269 Type *T = arg->getType();
1270 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
1271 return AK_FloatingPoint;
1272 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
1273 return AK_GeneralPurpose;
1274 if (T->isPointerTy())
1275 return AK_GeneralPurpose;
1279 // For VarArg functions, store the argument shadow in an ABI-specific format
1280 // that corresponds to va_list layout.
1281 // We do this because Clang lowers va_arg in the frontend, and this pass
1282 // only sees the low level code that deals with va_list internals.
1283 // A much easier alternative (provided that Clang emits va_arg instructions)
1284 // would have been to associate each live instance of va_list with a copy of
1285 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
1287 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) {
1288 unsigned GpOffset = 0;
1289 unsigned FpOffset = AMD64GpEndOffset;
1290 unsigned OverflowOffset = AMD64FpEndOffset;
1291 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1292 ArgIt != End; ++ArgIt) {
1294 ArgKind AK = classifyArgument(A);
1295 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
1297 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
1301 case AK_GeneralPurpose:
1302 Base = getShadowPtrForVAArgument(A, IRB, GpOffset);
1305 case AK_FloatingPoint:
1306 Base = getShadowPtrForVAArgument(A, IRB, FpOffset);
1310 uint64_t ArgSize = MS.TD->getTypeAllocSize(A->getType());
1311 Base = getShadowPtrForVAArgument(A, IRB, OverflowOffset);
1312 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
1314 IRB.CreateStore(MSV.getShadow(A), Base);
1316 Constant *OverflowSize =
1317 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
1318 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
1321 /// \brief Compute the shadow address for a given va_arg.
1322 Value *getShadowPtrForVAArgument(Value *A, IRBuilder<> &IRB,
1324 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
1325 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
1326 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(A), 0),
1330 void visitVAStartInst(VAStartInst &I) {
1331 IRBuilder<> IRB(&I);
1332 VAStartInstrumentationList.push_back(&I);
1333 Value *VAListTag = I.getArgOperand(0);
1334 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1336 // Unpoison the whole __va_list_tag.
1337 // FIXME: magic ABI constants.
1338 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1339 /* size */24, /* alignment */16, false);
1342 void visitVACopyInst(VACopyInst &I) {
1343 IRBuilder<> IRB(&I);
1344 Value *VAListTag = I.getArgOperand(0);
1345 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1347 // Unpoison the whole __va_list_tag.
1348 // FIXME: magic ABI constants.
1349 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1350 /* size */ 24, /* alignment */ 16, false);
1353 void finalizeInstrumentation() {
1354 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
1355 "finalizeInstrumentation called twice");
1356 if (!VAStartInstrumentationList.empty()) {
1357 // If there is a va_start in this function, make a backup copy of
1358 // va_arg_tls somewhere in the function entry block.
1359 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
1360 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
1362 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
1364 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
1365 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
1368 // Instrument va_start.
1369 // Copy va_list shadow from the backup copy of the TLS contents.
1370 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
1371 CallInst *OrigInst = VAStartInstrumentationList[i];
1372 IRBuilder<> IRB(OrigInst->getNextNode());
1373 Value *VAListTag = OrigInst->getArgOperand(0);
1375 Value *RegSaveAreaPtrPtr =
1377 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1378 ConstantInt::get(MS.IntptrTy, 16)),
1379 Type::getInt64PtrTy(*MS.C));
1380 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
1381 Value *RegSaveAreaShadowPtr =
1382 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
1383 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
1384 AMD64FpEndOffset, 16);
1386 Value *OverflowArgAreaPtrPtr =
1388 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1389 ConstantInt::get(MS.IntptrTy, 8)),
1390 Type::getInt64PtrTy(*MS.C));
1391 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
1392 Value *OverflowArgAreaShadowPtr =
1393 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
1395 getShadowPtrForVAArgument(VAArgTLSCopy, IRB, AMD64FpEndOffset);
1396 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
1401 VarArgHelper* CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
1402 MemorySanitizerVisitor &Visitor) {
1403 return new VarArgAMD64Helper(Func, Msan, Visitor);
1408 bool MemorySanitizer::runOnFunction(Function &F) {
1409 MemorySanitizerVisitor Visitor(F, *this);
1411 // Clear out readonly/readnone attributes.
1413 B.addAttribute(Attributes::ReadOnly)
1414 .addAttribute(Attributes::ReadNone);
1415 F.removeAttribute(AttrListPtr::FunctionIndex,
1416 Attributes::get(F.getContext(), B));
1418 return Visitor.runOnFunction();