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;
80 // This is an important flag that makes the reports much more
81 // informative at the cost of greater slowdown. Not fully implemented
83 // FIXME: this should be a top-level clang flag, e.g.
84 // -fmemory-sanitizer-full.
85 static cl::opt<bool> ClTrackOrigins("msan-track-origins",
86 cl::desc("Track origins (allocation sites) of poisoned memory"),
87 cl::Hidden, cl::init(false));
88 static cl::opt<bool> ClKeepGoing("msan-keep-going",
89 cl::desc("keep going after reporting a UMR"),
90 cl::Hidden, cl::init(false));
91 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
92 cl::desc("poison uninitialized stack variables"),
93 cl::Hidden, cl::init(true));
94 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
95 cl::desc("poison uninitialized stack variables with a call"),
96 cl::Hidden, cl::init(false));
97 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
98 cl::desc("poison uninitialized stack variables with the given patter"),
99 cl::Hidden, cl::init(0xff));
101 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
102 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
103 cl::Hidden, cl::init(true));
105 // This flag controls whether we check the shadow of the address
106 // operand of load or store. Such bugs are very rare, since load from
107 // a garbage address typically results in SEGV, but still happen
108 // (e.g. only lower bits of address are garbage, or the access happens
109 // early at program startup where malloc-ed memory is more likely to
110 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
111 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
112 cl::desc("report accesses through a pointer which has poisoned shadow"),
113 cl::Hidden, cl::init(true));
115 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
116 cl::desc("print out instructions with default strict semantics"),
117 cl::Hidden, cl::init(false));
119 static cl::opt<std::string> ClBlackListFile("msan-blacklist",
120 cl::desc("File containing the list of functions where MemorySanitizer "
121 "should not report bugs"), cl::Hidden);
125 /// \brief An instrumentation pass implementing detection of uninitialized
128 /// MemorySanitizer: instrument the code in module to find
129 /// uninitialized reads.
130 class MemorySanitizer : public FunctionPass {
132 MemorySanitizer() : FunctionPass(ID), TD(0) { }
133 const char *getPassName() const { return "MemorySanitizer"; }
134 bool runOnFunction(Function &F);
135 bool doInitialization(Module &M);
136 static char ID; // Pass identification, replacement for typeid.
143 /// \brief Thread-local shadow storage for function parameters.
144 GlobalVariable *ParamTLS;
145 /// \brief Thread-local origin storage for function parameters.
146 GlobalVariable *ParamOriginTLS;
147 /// \brief Thread-local shadow storage for function return value.
148 GlobalVariable *RetvalTLS;
149 /// \brief Thread-local origin storage for function return value.
150 GlobalVariable *RetvalOriginTLS;
151 /// \brief Thread-local shadow storage for in-register va_arg function
152 /// parameters (x86_64-specific).
153 GlobalVariable *VAArgTLS;
154 /// \brief Thread-local shadow storage for va_arg overflow area
155 /// (x86_64-specific).
156 GlobalVariable *VAArgOverflowSizeTLS;
157 /// \brief Thread-local space used to pass origin value to the UMR reporting
159 GlobalVariable *OriginTLS;
161 /// \brief The run-time callback to print a warning.
163 /// \brief Run-time helper that copies origin info for a memory range.
164 Value *MsanCopyOriginFn;
165 /// \brief Run-time helper that generates a new origin value for a stack
167 Value *MsanSetAllocaOriginFn;
168 /// \brief Run-time helper that poisons stack on function entry.
169 Value *MsanPoisonStackFn;
170 /// \brief MSan runtime replacements for memmove, memcpy and memset.
171 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
173 /// \brief Address mask used in application-to-shadow address calculation.
174 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
176 /// \brief Offset of the origin shadow from the "normal" shadow.
177 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
178 uint64_t OriginOffset;
179 /// \brief Branch weights for error reporting.
180 MDNode *ColdCallWeights;
181 /// \brief The blacklist.
182 OwningPtr<BlackList> BL;
183 /// \brief An empty volatile inline asm that prevents callback merge.
186 friend struct MemorySanitizerVisitor;
187 friend struct VarArgAMD64Helper;
191 char MemorySanitizer::ID = 0;
192 INITIALIZE_PASS(MemorySanitizer, "msan",
193 "MemorySanitizer: detects uninitialized reads.",
196 FunctionPass *llvm::createMemorySanitizerPass() {
197 return new MemorySanitizer();
200 /// \brief Create a non-const global initialized with the given string.
202 /// Creates a writable global for Str so that we can pass it to the
203 /// run-time lib. Runtime uses first 4 bytes of the string to store the
204 /// frame ID, so the string needs to be mutable.
205 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
207 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
208 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
209 GlobalValue::PrivateLinkage, StrConst, "");
212 /// \brief Module-level initialization.
214 /// Obtains pointers to the required runtime library functions, and
215 /// inserts a call to __msan_init to the module's constructor list.
216 bool MemorySanitizer::doInitialization(Module &M) {
217 TD = getAnalysisIfAvailable<DataLayout>();
220 BL.reset(new BlackList(ClBlackListFile));
221 C = &(M.getContext());
222 unsigned PtrSize = TD->getPointerSizeInBits(/* AddressSpace */0);
225 ShadowMask = kShadowMask64;
226 OriginOffset = kOriginOffset64;
229 ShadowMask = kShadowMask32;
230 OriginOffset = kOriginOffset32;
233 report_fatal_error("unsupported pointer size");
238 IntptrTy = IRB.getIntPtrTy(TD);
239 OriginTy = IRB.getInt32Ty();
241 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
243 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
244 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
245 "__msan_init", IRB.getVoidTy(), NULL)), 0);
247 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::LinkOnceODRLinkage,
248 IRB.getInt32(ClTrackOrigins), "__msan_track_origins");
250 // Create the callback.
251 // FIXME: this function should have "Cold" calling conv,
252 // which is not yet implemented.
253 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
254 : "__msan_warning_noreturn";
255 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
257 MsanCopyOriginFn = M.getOrInsertFunction(
258 "__msan_copy_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(),
259 IRB.getInt8PtrTy(), IntptrTy, NULL);
260 MsanSetAllocaOriginFn = M.getOrInsertFunction(
261 "__msan_set_alloca_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
262 IRB.getInt8PtrTy(), NULL);
263 MsanPoisonStackFn = M.getOrInsertFunction(
264 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
265 MemmoveFn = M.getOrInsertFunction(
266 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
268 MemcpyFn = M.getOrInsertFunction(
269 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
271 MemsetFn = M.getOrInsertFunction(
272 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
276 RetvalTLS = new GlobalVariable(
277 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
278 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
279 GlobalVariable::GeneralDynamicTLSModel);
280 RetvalOriginTLS = new GlobalVariable(
281 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
282 "__msan_retval_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
284 ParamTLS = new GlobalVariable(
285 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
286 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
287 GlobalVariable::GeneralDynamicTLSModel);
288 ParamOriginTLS = new GlobalVariable(
289 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
290 0, "__msan_param_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
292 VAArgTLS = new GlobalVariable(
293 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
294 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
295 GlobalVariable::GeneralDynamicTLSModel);
296 VAArgOverflowSizeTLS = new GlobalVariable(
297 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
298 "__msan_va_arg_overflow_size_tls", 0,
299 GlobalVariable::GeneralDynamicTLSModel);
300 OriginTLS = new GlobalVariable(
301 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
302 "__msan_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
304 // We insert an empty inline asm after __msan_report* to avoid callback merge.
305 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
306 StringRef(""), StringRef(""),
307 /*hasSideEffects=*/true);
313 /// \brief A helper class that handles instrumentation of VarArg
314 /// functions on a particular platform.
316 /// Implementations are expected to insert the instrumentation
317 /// necessary to propagate argument shadow through VarArg function
318 /// calls. Visit* methods are called during an InstVisitor pass over
319 /// the function, and should avoid creating new basic blocks. A new
320 /// instance of this class is created for each instrumented function.
321 struct VarArgHelper {
322 /// \brief Visit a CallSite.
323 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
325 /// \brief Visit a va_start call.
326 virtual void visitVAStartInst(VAStartInst &I) = 0;
328 /// \brief Visit a va_copy call.
329 virtual void visitVACopyInst(VACopyInst &I) = 0;
331 /// \brief Finalize function instrumentation.
333 /// This method is called after visiting all interesting (see above)
334 /// instructions in a function.
335 virtual void finalizeInstrumentation() = 0;
337 virtual ~VarArgHelper() {}
340 struct MemorySanitizerVisitor;
343 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
344 MemorySanitizerVisitor &Visitor);
346 /// This class does all the work for a given function. Store and Load
347 /// instructions store and load corresponding shadow and origin
348 /// values. Most instructions propagate shadow from arguments to their
349 /// return values. Certain instructions (most importantly, BranchInst)
350 /// test their argument shadow and print reports (with a runtime call) if it's
352 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
355 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
356 ValueMap<Value*, Value*> ShadowMap, OriginMap;
358 OwningPtr<VarArgHelper> VAHelper;
360 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
361 // See a comment in visitCallSite for more details.
362 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
363 static const unsigned AMD64FpEndOffset = 176;
365 struct ShadowOriginAndInsertPoint {
368 Instruction *OrigIns;
369 ShadowOriginAndInsertPoint(Instruction *S, Instruction *O, Instruction *I)
370 : Shadow(S), Origin(O), OrigIns(I) { }
371 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
373 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
375 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
376 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
377 InsertChecks = !MS.BL->isIn(F);
378 DEBUG(if (!InsertChecks)
379 dbgs() << "MemorySanitizer is not inserting checks into '"
380 << F.getName() << "'\n");
383 void materializeChecks() {
384 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
385 Instruction *Shadow = InstrumentationList[i].Shadow;
386 Instruction *OrigIns = InstrumentationList[i].OrigIns;
387 IRBuilder<> IRB(OrigIns);
388 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
389 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
390 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
391 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
392 getCleanShadow(ConvertedShadow), "_mscmp");
393 Instruction *CheckTerm =
394 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp),
395 /* Unreachable */ !ClKeepGoing,
398 IRB.SetInsertPoint(CheckTerm);
399 if (ClTrackOrigins) {
400 Instruction *Origin = InstrumentationList[i].Origin;
401 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
404 CallInst *Call = IRB.CreateCall(MS.WarningFn);
405 Call->setDebugLoc(OrigIns->getDebugLoc());
406 IRB.CreateCall(MS.EmptyAsm);
407 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
409 DEBUG(dbgs() << "DONE:\n" << F);
412 /// \brief Add MemorySanitizer instrumentation to a function.
413 bool runOnFunction() {
414 if (!MS.TD) return false;
415 // Iterate all BBs in depth-first order and create shadow instructions
416 // for all instructions (where applicable).
417 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
418 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
419 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
420 BasicBlock *BB = *DI;
424 // Finalize PHI nodes.
425 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
426 PHINode *PN = ShadowPHINodes[i];
427 PHINode *PNS = cast<PHINode>(getShadow(PN));
428 PHINode *PNO = ClTrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
429 size_t NumValues = PN->getNumIncomingValues();
430 for (size_t v = 0; v < NumValues; v++) {
431 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
433 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
437 VAHelper->finalizeInstrumentation();
444 /// \brief Compute the shadow type that corresponds to a given Value.
445 Type *getShadowTy(Value *V) {
446 return getShadowTy(V->getType());
449 /// \brief Compute the shadow type that corresponds to a given Type.
450 Type *getShadowTy(Type *OrigTy) {
451 if (!OrigTy->isSized()) {
454 // For integer type, shadow is the same as the original type.
455 // This may return weird-sized types like i1.
456 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
458 if (VectorType *VT = dyn_cast<VectorType>(OrigTy))
459 return VectorType::getInteger(VT);
460 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
461 SmallVector<Type*, 4> Elements;
462 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
463 Elements.push_back(getShadowTy(ST->getElementType(i)));
464 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
465 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
468 uint32_t TypeSize = MS.TD->getTypeStoreSizeInBits(OrigTy);
469 return IntegerType::get(*MS.C, TypeSize);
472 /// \brief Flatten a vector type.
473 Type *getShadowTyNoVec(Type *ty) {
474 if (VectorType *vt = dyn_cast<VectorType>(ty))
475 return IntegerType::get(*MS.C, vt->getBitWidth());
479 /// \brief Convert a shadow value to it's flattened variant.
480 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
481 Type *Ty = V->getType();
482 Type *NoVecTy = getShadowTyNoVec(Ty);
483 if (Ty == NoVecTy) return V;
484 return IRB.CreateBitCast(V, NoVecTy);
487 /// \brief Compute the shadow address that corresponds to a given application
490 /// Shadow = Addr & ~ShadowMask.
491 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
494 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
495 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
496 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
499 /// \brief Compute the origin address that corresponds to a given application
502 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
503 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
505 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
506 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
508 IRB.CreateAdd(ShadowLong,
509 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
511 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
512 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
515 /// \brief Compute the shadow address for a given function argument.
517 /// Shadow = ParamTLS+ArgOffset.
518 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
520 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
521 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
522 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
526 /// \brief Compute the origin address for a given function argument.
527 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
529 if (!ClTrackOrigins) return 0;
530 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
531 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
532 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
536 /// \brief Compute the shadow address for a retval.
537 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
538 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
539 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
543 /// \brief Compute the origin address for a retval.
544 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
545 // We keep a single origin for the entire retval. Might be too optimistic.
546 return MS.RetvalOriginTLS;
549 /// \brief Set SV to be the shadow value for V.
550 void setShadow(Value *V, Value *SV) {
551 assert(!ShadowMap.count(V) && "Values may only have one shadow");
555 /// \brief Set Origin to be the origin value for V.
556 void setOrigin(Value *V, Value *Origin) {
557 if (!ClTrackOrigins) return;
558 assert(!OriginMap.count(V) && "Values may only have one origin");
559 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
560 OriginMap[V] = Origin;
563 /// \brief Create a clean shadow value for a given value.
565 /// Clean shadow (all zeroes) means all bits of the value are defined
567 Value *getCleanShadow(Value *V) {
568 Type *ShadowTy = getShadowTy(V);
571 return Constant::getNullValue(ShadowTy);
574 /// \brief Create a dirty shadow of a given shadow type.
575 Constant *getPoisonedShadow(Type *ShadowTy) {
577 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
578 return Constant::getAllOnesValue(ShadowTy);
579 StructType *ST = cast<StructType>(ShadowTy);
580 SmallVector<Constant *, 4> Vals;
581 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
582 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
583 return ConstantStruct::get(ST, Vals);
586 /// \brief Create a clean (zero) origin.
587 Value *getCleanOrigin() {
588 return Constant::getNullValue(MS.OriginTy);
591 /// \brief Get the shadow value for a given Value.
593 /// This function either returns the value set earlier with setShadow,
594 /// or extracts if from ParamTLS (for function arguments).
595 Value *getShadow(Value *V) {
596 if (Instruction *I = dyn_cast<Instruction>(V)) {
597 // For instructions the shadow is already stored in the map.
598 Value *Shadow = ShadowMap[V];
600 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
602 assert(Shadow && "No shadow for a value");
606 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
607 Value *AllOnes = getPoisonedShadow(getShadowTy(V));
608 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
612 if (Argument *A = dyn_cast<Argument>(V)) {
613 // For arguments we compute the shadow on demand and store it in the map.
614 Value **ShadowPtr = &ShadowMap[V];
617 Function *F = A->getParent();
618 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
619 unsigned ArgOffset = 0;
620 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
622 if (!AI->getType()->isSized()) {
623 DEBUG(dbgs() << "Arg is not sized\n");
626 unsigned Size = AI->hasByValAttr()
627 ? MS.TD->getTypeAllocSize(AI->getType()->getPointerElementType())
628 : MS.TD->getTypeAllocSize(AI->getType());
630 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
631 if (AI->hasByValAttr()) {
632 // ByVal pointer itself has clean shadow. We copy the actual
633 // argument shadow to the underlying memory.
634 Value *Cpy = EntryIRB.CreateMemCpy(
635 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
636 Base, Size, AI->getParamAlignment());
637 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
639 *ShadowPtr = getCleanShadow(V);
641 *ShadowPtr = EntryIRB.CreateLoad(Base);
643 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
644 **ShadowPtr << "\n");
645 if (ClTrackOrigins) {
646 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
647 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
650 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
652 assert(*ShadowPtr && "Could not find shadow for an argument");
655 // For everything else the shadow is zero.
656 return getCleanShadow(V);
659 /// \brief Get the shadow for i-th argument of the instruction I.
660 Value *getShadow(Instruction *I, int i) {
661 return getShadow(I->getOperand(i));
664 /// \brief Get the origin for a value.
665 Value *getOrigin(Value *V) {
666 if (!ClTrackOrigins) return 0;
667 if (isa<Instruction>(V) || isa<Argument>(V)) {
668 Value *Origin = OriginMap[V];
670 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
671 Origin = getCleanOrigin();
675 return getCleanOrigin();
678 /// \brief Get the origin for i-th argument of the instruction I.
679 Value *getOrigin(Instruction *I, int i) {
680 return getOrigin(I->getOperand(i));
683 /// \brief Remember the place where a shadow check should be inserted.
685 /// This location will be later instrumented with a check that will print a
686 /// UMR warning in runtime if the value is not fully defined.
687 void insertCheck(Value *Val, Instruction *OrigIns) {
689 if (!InsertChecks) return;
690 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
693 Type *ShadowTy = Shadow->getType();
694 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
695 "Can only insert checks for integer and vector shadow types");
697 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
698 InstrumentationList.push_back(
699 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
702 //------------------- Visitors.
704 /// \brief Instrument LoadInst
706 /// Loads the corresponding shadow and (optionally) origin.
707 /// Optionally, checks that the load address is fully defined.
708 void visitLoadInst(LoadInst &I) {
709 assert(I.getType()->isSized() && "Load type must have size");
711 Type *ShadowTy = getShadowTy(&I);
712 Value *Addr = I.getPointerOperand();
713 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
714 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
716 if (ClCheckAccessAddress)
717 insertCheck(I.getPointerOperand(), &I);
720 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), I.getAlignment()));
723 /// \brief Instrument StoreInst
725 /// Stores the corresponding shadow and (optionally) origin.
726 /// Optionally, checks that the store address is fully defined.
727 /// Volatile stores check that the value being stored is fully defined.
728 void visitStoreInst(StoreInst &I) {
730 Value *Val = I.getValueOperand();
731 Value *Addr = I.getPointerOperand();
732 Value *Shadow = getShadow(Val);
733 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
735 StoreInst *NewSI = IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
736 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
738 // If the store is volatile, add a check.
740 insertCheck(Val, &I);
741 if (ClCheckAccessAddress)
742 insertCheck(Addr, &I);
745 IRB.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRB), I.getAlignment());
748 // Vector manipulation.
749 void visitExtractElementInst(ExtractElementInst &I) {
750 insertCheck(I.getOperand(1), &I);
752 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
754 setOrigin(&I, getOrigin(&I, 0));
757 void visitInsertElementInst(InsertElementInst &I) {
758 insertCheck(I.getOperand(2), &I);
760 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
761 I.getOperand(2), "_msprop"));
762 setOriginForNaryOp(I);
765 void visitShuffleVectorInst(ShuffleVectorInst &I) {
766 insertCheck(I.getOperand(2), &I);
768 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
769 I.getOperand(2), "_msprop"));
770 setOriginForNaryOp(I);
774 void visitSExtInst(SExtInst &I) {
776 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
777 setOrigin(&I, getOrigin(&I, 0));
780 void visitZExtInst(ZExtInst &I) {
782 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
783 setOrigin(&I, getOrigin(&I, 0));
786 void visitTruncInst(TruncInst &I) {
788 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
789 setOrigin(&I, getOrigin(&I, 0));
792 void visitBitCastInst(BitCastInst &I) {
794 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
795 setOrigin(&I, getOrigin(&I, 0));
798 void visitPtrToIntInst(PtrToIntInst &I) {
800 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
801 "_msprop_ptrtoint"));
802 setOrigin(&I, getOrigin(&I, 0));
805 void visitIntToPtrInst(IntToPtrInst &I) {
807 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
808 "_msprop_inttoptr"));
809 setOrigin(&I, getOrigin(&I, 0));
812 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
813 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
814 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
815 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
816 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
817 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
819 /// \brief Propagate shadow for bitwise AND.
821 /// This code is exact, i.e. if, for example, a bit in the left argument
822 /// is defined and 0, then neither the value not definedness of the
823 /// corresponding bit in B don't affect the resulting shadow.
824 void visitAnd(BinaryOperator &I) {
826 // "And" of 0 and a poisoned value results in unpoisoned value.
827 // 1&1 => 1; 0&1 => 0; p&1 => p;
828 // 1&0 => 0; 0&0 => 0; p&0 => 0;
829 // 1&p => p; 0&p => 0; p&p => p;
830 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
831 Value *S1 = getShadow(&I, 0);
832 Value *S2 = getShadow(&I, 1);
833 Value *V1 = I.getOperand(0);
834 Value *V2 = I.getOperand(1);
835 if (V1->getType() != S1->getType()) {
836 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
837 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
839 Value *S1S2 = IRB.CreateAnd(S1, S2);
840 Value *V1S2 = IRB.CreateAnd(V1, S2);
841 Value *S1V2 = IRB.CreateAnd(S1, V2);
842 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
843 setOriginForNaryOp(I);
846 void visitOr(BinaryOperator &I) {
848 // "Or" of 1 and a poisoned value results in unpoisoned value.
849 // 1|1 => 1; 0|1 => 1; p|1 => 1;
850 // 1|0 => 1; 0|0 => 0; p|0 => p;
851 // 1|p => 1; 0|p => p; p|p => p;
852 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
853 Value *S1 = getShadow(&I, 0);
854 Value *S2 = getShadow(&I, 1);
855 Value *V1 = IRB.CreateNot(I.getOperand(0));
856 Value *V2 = IRB.CreateNot(I.getOperand(1));
857 if (V1->getType() != S1->getType()) {
858 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
859 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
861 Value *S1S2 = IRB.CreateAnd(S1, S2);
862 Value *V1S2 = IRB.CreateAnd(V1, S2);
863 Value *S1V2 = IRB.CreateAnd(S1, V2);
864 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
865 setOriginForNaryOp(I);
868 /// \brief Propagate origin for an instruction.
870 /// This is a general case of origin propagation. For an Nary operation,
871 /// is set to the origin of an argument that is not entirely initialized.
872 /// If there is more than one such arguments, the rightmost of them is picked.
873 /// It does not matter which one is picked if all arguments are initialized.
874 void setOriginForNaryOp(Instruction &I) {
875 if (!ClTrackOrigins) return;
877 Value *Origin = getOrigin(&I, 0);
878 for (unsigned Op = 1, n = I.getNumOperands(); Op < n; ++Op) {
879 Value *S = convertToShadowTyNoVec(getShadow(&I, Op), IRB);
880 Origin = IRB.CreateSelect(IRB.CreateICmpNE(S, getCleanShadow(S)),
881 getOrigin(&I, Op), Origin);
883 setOrigin(&I, Origin);
886 /// \brief Propagate shadow for a binary operation.
888 /// Shadow = Shadow0 | Shadow1, all 3 must have the same type.
889 /// Bitwise OR is selected as an operation that will never lose even a bit of
891 void handleShadowOrBinary(Instruction &I) {
893 Value *Shadow0 = getShadow(&I, 0);
894 Value *Shadow1 = getShadow(&I, 1);
895 setShadow(&I, IRB.CreateOr(Shadow0, Shadow1, "_msprop"));
896 setOriginForNaryOp(I);
899 /// \brief Propagate shadow for arbitrary operation.
901 /// This is a general case of shadow propagation, used in all cases where we
902 /// don't know and/or care about what the operation actually does.
903 /// It converts all input shadow values to a common type (extending or
904 /// truncating as necessary), and bitwise OR's them.
906 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
907 /// fully initialized), and less prone to false positives.
908 // FIXME: is the casting actually correct?
909 // FIXME: merge this with handleShadowOrBinary.
910 void handleShadowOr(Instruction &I) {
912 Value *Shadow = getShadow(&I, 0);
913 for (unsigned Op = 1, n = I.getNumOperands(); Op < n; ++Op)
914 Shadow = IRB.CreateOr(
915 Shadow, IRB.CreateIntCast(getShadow(&I, Op), Shadow->getType(), false),
917 Shadow = IRB.CreateIntCast(Shadow, getShadowTy(&I), false);
918 setShadow(&I, Shadow);
919 setOriginForNaryOp(I);
922 void visitFAdd(BinaryOperator &I) { handleShadowOrBinary(I); }
923 void visitFSub(BinaryOperator &I) { handleShadowOrBinary(I); }
924 void visitFMul(BinaryOperator &I) { handleShadowOrBinary(I); }
925 void visitAdd(BinaryOperator &I) { handleShadowOrBinary(I); }
926 void visitSub(BinaryOperator &I) { handleShadowOrBinary(I); }
927 void visitXor(BinaryOperator &I) { handleShadowOrBinary(I); }
928 void visitMul(BinaryOperator &I) { handleShadowOrBinary(I); }
930 void handleDiv(Instruction &I) {
932 // Strict on the second argument.
933 insertCheck(I.getOperand(1), &I);
934 setShadow(&I, getShadow(&I, 0));
935 setOrigin(&I, getOrigin(&I, 0));
938 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
939 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
940 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
941 void visitURem(BinaryOperator &I) { handleDiv(I); }
942 void visitSRem(BinaryOperator &I) { handleDiv(I); }
943 void visitFRem(BinaryOperator &I) { handleDiv(I); }
945 /// \brief Instrument == and != comparisons.
947 /// Sometimes the comparison result is known even if some of the bits of the
948 /// arguments are not.
949 void handleEqualityComparison(ICmpInst &I) {
951 Value *A = I.getOperand(0);
952 Value *B = I.getOperand(1);
953 Value *Sa = getShadow(A);
954 Value *Sb = getShadow(B);
955 if (A->getType()->isPointerTy())
956 A = IRB.CreatePointerCast(A, MS.IntptrTy);
957 if (B->getType()->isPointerTy())
958 B = IRB.CreatePointerCast(B, MS.IntptrTy);
959 // A == B <==> (C = A^B) == 0
960 // A != B <==> (C = A^B) != 0
962 Value *C = IRB.CreateXor(A, B);
963 Value *Sc = IRB.CreateOr(Sa, Sb);
964 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
965 // Result is defined if one of the following is true
966 // * there is a defined 1 bit in C
967 // * C is fully defined
968 // Si = !(C & ~Sc) && Sc
969 Value *Zero = Constant::getNullValue(Sc->getType());
970 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
972 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
974 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
975 Si->setName("_msprop_icmp");
977 setOriginForNaryOp(I);
980 /// \brief Instrument signed relational comparisons.
982 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
983 /// propagating the highest bit of the shadow. Everything else is delegated
984 /// to handleShadowOr().
985 void handleSignedRelationalComparison(ICmpInst &I) {
986 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
987 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
989 CmpInst::Predicate pre = I.getPredicate();
990 if (constOp0 && constOp0->isNullValue() &&
991 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
992 op = I.getOperand(1);
993 } else if (constOp1 && constOp1->isNullValue() &&
994 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
995 op = I.getOperand(0);
1000 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1001 setShadow(&I, Shadow);
1002 setOrigin(&I, getOrigin(op));
1008 void visitICmpInst(ICmpInst &I) {
1009 if (ClHandleICmp && I.isEquality())
1010 handleEqualityComparison(I);
1011 else if (ClHandleICmp && I.isSigned() && I.isRelational())
1012 handleSignedRelationalComparison(I);
1017 void visitFCmpInst(FCmpInst &I) {
1021 void handleShift(BinaryOperator &I) {
1022 IRBuilder<> IRB(&I);
1023 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1024 // Otherwise perform the same shift on S1.
1025 Value *S1 = getShadow(&I, 0);
1026 Value *S2 = getShadow(&I, 1);
1027 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1029 Value *V2 = I.getOperand(1);
1030 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1031 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1032 setOriginForNaryOp(I);
1035 void visitShl(BinaryOperator &I) { handleShift(I); }
1036 void visitAShr(BinaryOperator &I) { handleShift(I); }
1037 void visitLShr(BinaryOperator &I) { handleShift(I); }
1039 /// \brief Instrument llvm.memmove
1041 /// At this point we don't know if llvm.memmove will be inlined or not.
1042 /// If we don't instrument it and it gets inlined,
1043 /// our interceptor will not kick in and we will lose the memmove.
1044 /// If we instrument the call here, but it does not get inlined,
1045 /// we will memove the shadow twice: which is bad in case
1046 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1048 /// Similar situation exists for memcpy and memset.
1049 void visitMemMoveInst(MemMoveInst &I) {
1050 IRBuilder<> IRB(&I);
1053 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1054 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1055 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1056 I.eraseFromParent();
1059 // Similar to memmove: avoid copying shadow twice.
1060 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1061 // FIXME: consider doing manual inline for small constant sizes and proper
1063 void visitMemCpyInst(MemCpyInst &I) {
1064 IRBuilder<> IRB(&I);
1067 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1068 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1069 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1070 I.eraseFromParent();
1074 void visitMemSetInst(MemSetInst &I) {
1075 IRBuilder<> IRB(&I);
1078 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1079 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1080 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1081 I.eraseFromParent();
1084 void visitVAStartInst(VAStartInst &I) {
1085 VAHelper->visitVAStartInst(I);
1088 void visitVACopyInst(VACopyInst &I) {
1089 VAHelper->visitVACopyInst(I);
1092 void visitCallSite(CallSite CS) {
1093 Instruction &I = *CS.getInstruction();
1094 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
1096 CallInst *Call = cast<CallInst>(&I);
1098 // For inline asm, do the usual thing: check argument shadow and mark all
1099 // outputs as clean. Note that any side effects of the inline asm that are
1100 // not immediately visible in its constraints are not handled.
1101 if (Call->isInlineAsm()) {
1102 visitInstruction(I);
1106 // Allow only tail calls with the same types, otherwise
1107 // we may have a false positive: shadow for a non-void RetVal
1108 // will get propagated to a void RetVal.
1109 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
1110 Call->setTailCall(false);
1111 if (isa<IntrinsicInst>(&I)) {
1112 // All intrinsics we care about are handled in corresponding visit*
1113 // methods. Add checks for the arguments, mark retval as clean.
1114 visitInstruction(I);
1118 IRBuilder<> IRB(&I);
1119 unsigned ArgOffset = 0;
1120 DEBUG(dbgs() << " CallSite: " << I << "\n");
1121 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1122 ArgIt != End; ++ArgIt) {
1124 unsigned i = ArgIt - CS.arg_begin();
1125 if (!A->getType()->isSized()) {
1126 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
1131 // Compute the Shadow for arg even if it is ByVal, because
1132 // in that case getShadow() will copy the actual arg shadow to
1133 // __msan_param_tls.
1134 Value *ArgShadow = getShadow(A);
1135 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
1136 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
1137 " Shadow: " << *ArgShadow << "\n");
1138 if (CS.paramHasAttr(i + 1, Attributes::ByVal)) {
1139 assert(A->getType()->isPointerTy() &&
1140 "ByVal argument is not a pointer!");
1141 Size = MS.TD->getTypeAllocSize(A->getType()->getPointerElementType());
1142 unsigned Alignment = CS.getParamAlignment(i + 1);
1143 Store = IRB.CreateMemCpy(ArgShadowBase,
1144 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
1147 Size = MS.TD->getTypeAllocSize(A->getType());
1148 Store = IRB.CreateStore(ArgShadow, ArgShadowBase);
1151 IRB.CreateStore(getOrigin(A),
1152 getOriginPtrForArgument(A, IRB, ArgOffset));
1153 assert(Size != 0 && Store != 0);
1154 DEBUG(dbgs() << " Param:" << *Store << "\n");
1155 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
1157 DEBUG(dbgs() << " done with call args\n");
1160 cast<FunctionType>(CS.getCalledValue()->getType()-> getContainedType(0));
1161 if (FT->isVarArg()) {
1162 VAHelper->visitCallSite(CS, IRB);
1165 // Now, get the shadow for the RetVal.
1166 if (!I.getType()->isSized()) return;
1167 IRBuilder<> IRBBefore(&I);
1168 // Untill we have full dynamic coverage, make sure the retval shadow is 0.
1169 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
1170 IRBBefore.CreateStore(getCleanShadow(&I), Base);
1171 Instruction *NextInsn = 0;
1173 NextInsn = I.getNextNode();
1175 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
1176 if (!NormalDest->getSinglePredecessor()) {
1177 // FIXME: this case is tricky, so we are just conservative here.
1178 // Perhaps we need to split the edge between this BB and NormalDest,
1179 // but a naive attempt to use SplitEdge leads to a crash.
1180 setShadow(&I, getCleanShadow(&I));
1181 setOrigin(&I, getCleanOrigin());
1184 NextInsn = NormalDest->getFirstInsertionPt();
1186 "Could not find insertion point for retval shadow load");
1188 IRBuilder<> IRBAfter(NextInsn);
1189 setShadow(&I, IRBAfter.CreateLoad(getShadowPtrForRetval(&I, IRBAfter),
1192 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
1195 void visitReturnInst(ReturnInst &I) {
1196 IRBuilder<> IRB(&I);
1197 if (Value *RetVal = I.getReturnValue()) {
1198 // Set the shadow for the RetVal.
1199 Value *Shadow = getShadow(RetVal);
1200 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
1201 DEBUG(dbgs() << "Return: " << *Shadow << "\n" << *ShadowPtr << "\n");
1202 IRB.CreateStore(Shadow, ShadowPtr);
1204 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
1208 void visitPHINode(PHINode &I) {
1209 IRBuilder<> IRB(&I);
1210 ShadowPHINodes.push_back(&I);
1211 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
1214 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
1218 void visitAllocaInst(AllocaInst &I) {
1219 setShadow(&I, getCleanShadow(&I));
1220 if (!ClPoisonStack) return;
1221 IRBuilder<> IRB(I.getNextNode());
1222 uint64_t Size = MS.TD->getTypeAllocSize(I.getAllocatedType());
1223 if (ClPoisonStackWithCall) {
1224 IRB.CreateCall2(MS.MsanPoisonStackFn,
1225 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1226 ConstantInt::get(MS.IntptrTy, Size));
1228 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
1229 IRB.CreateMemSet(ShadowBase, IRB.getInt8(ClPoisonStackPattern),
1230 Size, I.getAlignment());
1233 if (ClTrackOrigins) {
1234 setOrigin(&I, getCleanOrigin());
1235 SmallString<2048> StackDescriptionStorage;
1236 raw_svector_ostream StackDescription(StackDescriptionStorage);
1237 // We create a string with a description of the stack allocation and
1238 // pass it into __msan_set_alloca_origin.
1239 // It will be printed by the run-time if stack-originated UMR is found.
1240 // The first 4 bytes of the string are set to '----' and will be replaced
1241 // by __msan_va_arg_overflow_size_tls at the first call.
1242 StackDescription << "----" << I.getName() << "@" << F.getName();
1244 createPrivateNonConstGlobalForString(*F.getParent(),
1245 StackDescription.str());
1246 IRB.CreateCall3(MS.MsanSetAllocaOriginFn,
1247 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1248 ConstantInt::get(MS.IntptrTy, Size),
1249 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()));
1253 void visitSelectInst(SelectInst& I) {
1254 IRBuilder<> IRB(&I);
1255 setShadow(&I, IRB.CreateSelect(I.getCondition(),
1256 getShadow(I.getTrueValue()), getShadow(I.getFalseValue()),
1259 setOrigin(&I, IRB.CreateSelect(I.getCondition(),
1260 getOrigin(I.getTrueValue()), getOrigin(I.getFalseValue())));
1263 void visitLandingPadInst(LandingPadInst &I) {
1265 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
1266 setShadow(&I, getCleanShadow(&I));
1267 setOrigin(&I, getCleanOrigin());
1270 void visitGetElementPtrInst(GetElementPtrInst &I) {
1274 void visitExtractValueInst(ExtractValueInst &I) {
1275 IRBuilder<> IRB(&I);
1276 Value *Agg = I.getAggregateOperand();
1277 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
1278 Value *AggShadow = getShadow(Agg);
1279 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1280 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
1281 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
1282 setShadow(&I, ResShadow);
1283 setOrigin(&I, getCleanOrigin());
1286 void visitInsertValueInst(InsertValueInst &I) {
1287 IRBuilder<> IRB(&I);
1288 DEBUG(dbgs() << "InsertValue: " << I << "\n");
1289 Value *AggShadow = getShadow(I.getAggregateOperand());
1290 Value *InsShadow = getShadow(I.getInsertedValueOperand());
1291 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1292 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
1293 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
1294 DEBUG(dbgs() << " Res: " << *Res << "\n");
1296 setOrigin(&I, getCleanOrigin());
1299 void dumpInst(Instruction &I) {
1300 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1301 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
1303 errs() << "ZZZ " << I.getOpcodeName() << "\n";
1305 errs() << "QQQ " << I << "\n";
1308 void visitResumeInst(ResumeInst &I) {
1309 DEBUG(dbgs() << "Resume: " << I << "\n");
1310 // Nothing to do here.
1313 void visitInstruction(Instruction &I) {
1314 // Everything else: stop propagating and check for poisoned shadow.
1315 if (ClDumpStrictInstructions)
1317 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
1318 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
1319 insertCheck(I.getOperand(i), &I);
1320 setShadow(&I, getCleanShadow(&I));
1321 setOrigin(&I, getCleanOrigin());
1325 /// \brief AMD64-specific implementation of VarArgHelper.
1326 struct VarArgAMD64Helper : public VarArgHelper {
1327 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
1328 // See a comment in visitCallSite for more details.
1329 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
1330 static const unsigned AMD64FpEndOffset = 176;
1333 MemorySanitizer &MS;
1334 MemorySanitizerVisitor &MSV;
1335 Value *VAArgTLSCopy;
1336 Value *VAArgOverflowSize;
1338 SmallVector<CallInst*, 16> VAStartInstrumentationList;
1340 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
1341 MemorySanitizerVisitor &MSV)
1342 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
1344 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
1346 ArgKind classifyArgument(Value* arg) {
1347 // A very rough approximation of X86_64 argument classification rules.
1348 Type *T = arg->getType();
1349 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
1350 return AK_FloatingPoint;
1351 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
1352 return AK_GeneralPurpose;
1353 if (T->isPointerTy())
1354 return AK_GeneralPurpose;
1358 // For VarArg functions, store the argument shadow in an ABI-specific format
1359 // that corresponds to va_list layout.
1360 // We do this because Clang lowers va_arg in the frontend, and this pass
1361 // only sees the low level code that deals with va_list internals.
1362 // A much easier alternative (provided that Clang emits va_arg instructions)
1363 // would have been to associate each live instance of va_list with a copy of
1364 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
1366 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) {
1367 unsigned GpOffset = 0;
1368 unsigned FpOffset = AMD64GpEndOffset;
1369 unsigned OverflowOffset = AMD64FpEndOffset;
1370 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1371 ArgIt != End; ++ArgIt) {
1373 ArgKind AK = classifyArgument(A);
1374 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
1376 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
1380 case AK_GeneralPurpose:
1381 Base = getShadowPtrForVAArgument(A, IRB, GpOffset);
1384 case AK_FloatingPoint:
1385 Base = getShadowPtrForVAArgument(A, IRB, FpOffset);
1389 uint64_t ArgSize = MS.TD->getTypeAllocSize(A->getType());
1390 Base = getShadowPtrForVAArgument(A, IRB, OverflowOffset);
1391 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
1393 IRB.CreateStore(MSV.getShadow(A), Base);
1395 Constant *OverflowSize =
1396 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
1397 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
1400 /// \brief Compute the shadow address for a given va_arg.
1401 Value *getShadowPtrForVAArgument(Value *A, IRBuilder<> &IRB,
1403 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
1404 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
1405 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(A), 0),
1409 void visitVAStartInst(VAStartInst &I) {
1410 IRBuilder<> IRB(&I);
1411 VAStartInstrumentationList.push_back(&I);
1412 Value *VAListTag = I.getArgOperand(0);
1413 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1415 // Unpoison the whole __va_list_tag.
1416 // FIXME: magic ABI constants.
1417 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1418 /* size */24, /* alignment */16, false);
1421 void visitVACopyInst(VACopyInst &I) {
1422 IRBuilder<> IRB(&I);
1423 Value *VAListTag = I.getArgOperand(0);
1424 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1426 // Unpoison the whole __va_list_tag.
1427 // FIXME: magic ABI constants.
1428 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1429 /* size */ 24, /* alignment */ 16, false);
1432 void finalizeInstrumentation() {
1433 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
1434 "finalizeInstrumentation called twice");
1435 if (!VAStartInstrumentationList.empty()) {
1436 // If there is a va_start in this function, make a backup copy of
1437 // va_arg_tls somewhere in the function entry block.
1438 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
1439 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
1441 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
1443 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
1444 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
1447 // Instrument va_start.
1448 // Copy va_list shadow from the backup copy of the TLS contents.
1449 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
1450 CallInst *OrigInst = VAStartInstrumentationList[i];
1451 IRBuilder<> IRB(OrigInst->getNextNode());
1452 Value *VAListTag = OrigInst->getArgOperand(0);
1454 Value *RegSaveAreaPtrPtr =
1456 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1457 ConstantInt::get(MS.IntptrTy, 16)),
1458 Type::getInt64PtrTy(*MS.C));
1459 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
1460 Value *RegSaveAreaShadowPtr =
1461 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
1462 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
1463 AMD64FpEndOffset, 16);
1465 Value *OverflowArgAreaPtrPtr =
1467 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1468 ConstantInt::get(MS.IntptrTy, 8)),
1469 Type::getInt64PtrTy(*MS.C));
1470 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
1471 Value *OverflowArgAreaShadowPtr =
1472 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
1474 getShadowPtrForVAArgument(VAArgTLSCopy, IRB, AMD64FpEndOffset);
1475 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
1480 VarArgHelper* CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
1481 MemorySanitizerVisitor &Visitor) {
1482 return new VarArgAMD64Helper(Func, Msan, Visitor);
1487 bool MemorySanitizer::runOnFunction(Function &F) {
1488 MemorySanitizerVisitor Visitor(F, *this);
1490 // Clear out readonly/readnone attributes.
1492 B.addAttribute(Attributes::ReadOnly)
1493 .addAttribute(Attributes::ReadNone);
1494 F.removeAttribute(AttrListPtr::FunctionIndex,
1495 Attributes::get(F.getContext(), B));
1497 return Visitor.runOnFunction();