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).
49 /// MemorySanitizer can track origins (allocation points) of all uninitialized
50 /// values. This behavior is controlled with a flag (msan-track-origins) and is
51 /// disabled by default.
53 /// Origins are 4-byte values created and interpreted by the runtime library.
54 /// They are stored in a second shadow mapping, one 4-byte value for 4 bytes
55 /// of application memory. Propagation of origins is basically a bunch of
56 /// "select" instructions that pick the origin of a dirty argument, if an
57 /// instruction has one.
59 /// Every 4 aligned, consecutive bytes of application memory have one origin
60 /// value associated with them. If these bytes contain uninitialized data
61 /// coming from 2 different allocations, the last store wins. Because of this,
62 /// MemorySanitizer reports can show unrelated origins, but this is unlikely in
65 /// Origins are meaningless for fully initialized values, so MemorySanitizer
66 /// avoids storing origin to memory when a fully initialized value is stored.
67 /// This way it avoids needless overwritting origin of the 4-byte region on
68 /// a short (i.e. 1 byte) clean store, and it is also good for performance.
69 //===----------------------------------------------------------------------===//
71 #define DEBUG_TYPE "msan"
73 #include "llvm/Transforms/Instrumentation.h"
74 #include "BlackList.h"
75 #include "llvm/ADT/DepthFirstIterator.h"
76 #include "llvm/ADT/SmallString.h"
77 #include "llvm/ADT/SmallVector.h"
78 #include "llvm/ADT/ValueMap.h"
79 #include "llvm/DataLayout.h"
80 #include "llvm/Function.h"
81 #include "llvm/IRBuilder.h"
82 #include "llvm/InlineAsm.h"
83 #include "llvm/InstVisitor.h"
84 #include "llvm/IntrinsicInst.h"
85 #include "llvm/LLVMContext.h"
86 #include "llvm/MDBuilder.h"
87 #include "llvm/Module.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Compiler.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/ModuleUtils.h"
95 #include "llvm/Type.h"
99 static const uint64_t kShadowMask32 = 1ULL << 31;
100 static const uint64_t kShadowMask64 = 1ULL << 46;
101 static const uint64_t kOriginOffset32 = 1ULL << 30;
102 static const uint64_t kOriginOffset64 = 1ULL << 45;
103 static const uint64_t kShadowTLSAlignment = 8;
105 /// \brief Track origins of uninitialized values.
107 /// Adds a section to MemorySanitizer report that points to the allocation
108 /// (stack or heap) the uninitialized bits came from originally.
109 static cl::opt<bool> ClTrackOrigins("msan-track-origins",
110 cl::desc("Track origins (allocation sites) of poisoned memory"),
111 cl::Hidden, cl::init(false));
112 static cl::opt<bool> ClKeepGoing("msan-keep-going",
113 cl::desc("keep going after reporting a UMR"),
114 cl::Hidden, cl::init(false));
115 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
116 cl::desc("poison uninitialized stack variables"),
117 cl::Hidden, cl::init(true));
118 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
119 cl::desc("poison uninitialized stack variables with a call"),
120 cl::Hidden, cl::init(false));
121 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
122 cl::desc("poison uninitialized stack variables with the given patter"),
123 cl::Hidden, cl::init(0xff));
125 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
126 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
127 cl::Hidden, cl::init(true));
129 static cl::opt<bool> ClStoreCleanOrigin("msan-store-clean-origin",
130 cl::desc("store origin for clean (fully initialized) values"),
131 cl::Hidden, cl::init(false));
133 // This flag controls whether we check the shadow of the address
134 // operand of load or store. Such bugs are very rare, since load from
135 // a garbage address typically results in SEGV, but still happen
136 // (e.g. only lower bits of address are garbage, or the access happens
137 // early at program startup where malloc-ed memory is more likely to
138 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
139 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
140 cl::desc("report accesses through a pointer which has poisoned shadow"),
141 cl::Hidden, cl::init(true));
143 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
144 cl::desc("print out instructions with default strict semantics"),
145 cl::Hidden, cl::init(false));
147 static cl::opt<std::string> ClBlackListFile("msan-blacklist",
148 cl::desc("File containing the list of functions where MemorySanitizer "
149 "should not report bugs"), cl::Hidden);
153 /// \brief An instrumentation pass implementing detection of uninitialized
156 /// MemorySanitizer: instrument the code in module to find
157 /// uninitialized reads.
158 class MemorySanitizer : public FunctionPass {
160 MemorySanitizer(bool TrackOrigins = false)
162 TrackOrigins(TrackOrigins || ClTrackOrigins),
165 const char *getPassName() const { return "MemorySanitizer"; }
166 bool runOnFunction(Function &F);
167 bool doInitialization(Module &M);
168 static char ID; // Pass identification, replacement for typeid.
171 void initializeCallbacks(Module &M);
173 /// \brief Track origins (allocation points) of uninitialized values.
180 /// \brief Thread-local shadow storage for function parameters.
181 GlobalVariable *ParamTLS;
182 /// \brief Thread-local origin storage for function parameters.
183 GlobalVariable *ParamOriginTLS;
184 /// \brief Thread-local shadow storage for function return value.
185 GlobalVariable *RetvalTLS;
186 /// \brief Thread-local origin storage for function return value.
187 GlobalVariable *RetvalOriginTLS;
188 /// \brief Thread-local shadow storage for in-register va_arg function
189 /// parameters (x86_64-specific).
190 GlobalVariable *VAArgTLS;
191 /// \brief Thread-local shadow storage for va_arg overflow area
192 /// (x86_64-specific).
193 GlobalVariable *VAArgOverflowSizeTLS;
194 /// \brief Thread-local space used to pass origin value to the UMR reporting
196 GlobalVariable *OriginTLS;
198 /// \brief The run-time callback to print a warning.
200 /// \brief Run-time helper that copies origin info for a memory range.
201 Value *MsanCopyOriginFn;
202 /// \brief Run-time helper that generates a new origin value for a stack
204 Value *MsanSetAllocaOriginFn;
205 /// \brief Run-time helper that poisons stack on function entry.
206 Value *MsanPoisonStackFn;
207 /// \brief MSan runtime replacements for memmove, memcpy and memset.
208 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
210 /// \brief Address mask used in application-to-shadow address calculation.
211 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
213 /// \brief Offset of the origin shadow from the "normal" shadow.
214 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
215 uint64_t OriginOffset;
216 /// \brief Branch weights for error reporting.
217 MDNode *ColdCallWeights;
218 /// \brief Branch weights for origin store.
219 MDNode *OriginStoreWeights;
220 /// \brief The blacklist.
221 OwningPtr<BlackList> BL;
222 /// \brief An empty volatile inline asm that prevents callback merge.
225 friend struct MemorySanitizerVisitor;
226 friend struct VarArgAMD64Helper;
230 char MemorySanitizer::ID = 0;
231 INITIALIZE_PASS(MemorySanitizer, "msan",
232 "MemorySanitizer: detects uninitialized reads.",
235 FunctionPass *llvm::createMemorySanitizerPass(bool TrackOrigins) {
236 return new MemorySanitizer(TrackOrigins);
239 /// \brief Create a non-const global initialized with the given string.
241 /// Creates a writable global for Str so that we can pass it to the
242 /// run-time lib. Runtime uses first 4 bytes of the string to store the
243 /// frame ID, so the string needs to be mutable.
244 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
246 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
247 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
248 GlobalValue::PrivateLinkage, StrConst, "");
252 /// \brief Insert extern declaration of runtime-provided functions and globals.
253 void MemorySanitizer::initializeCallbacks(Module &M) {
254 // Only do this once.
259 // Create the callback.
260 // FIXME: this function should have "Cold" calling conv,
261 // which is not yet implemented.
262 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
263 : "__msan_warning_noreturn";
264 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
266 MsanCopyOriginFn = M.getOrInsertFunction(
267 "__msan_copy_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(),
268 IRB.getInt8PtrTy(), IntptrTy, NULL);
269 MsanSetAllocaOriginFn = M.getOrInsertFunction(
270 "__msan_set_alloca_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
271 IRB.getInt8PtrTy(), NULL);
272 MsanPoisonStackFn = M.getOrInsertFunction(
273 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
274 MemmoveFn = M.getOrInsertFunction(
275 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
276 IRB.getInt8PtrTy(), IntptrTy, NULL);
277 MemcpyFn = M.getOrInsertFunction(
278 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
280 MemsetFn = M.getOrInsertFunction(
281 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
285 RetvalTLS = new GlobalVariable(
286 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
287 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
288 GlobalVariable::GeneralDynamicTLSModel);
289 RetvalOriginTLS = new GlobalVariable(
290 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
291 "__msan_retval_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
293 ParamTLS = new GlobalVariable(
294 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
295 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
296 GlobalVariable::GeneralDynamicTLSModel);
297 ParamOriginTLS = new GlobalVariable(
298 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
299 0, "__msan_param_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
301 VAArgTLS = new GlobalVariable(
302 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
303 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
304 GlobalVariable::GeneralDynamicTLSModel);
305 VAArgOverflowSizeTLS = new GlobalVariable(
306 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
307 "__msan_va_arg_overflow_size_tls", 0,
308 GlobalVariable::GeneralDynamicTLSModel);
309 OriginTLS = new GlobalVariable(
310 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
311 "__msan_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
313 // We insert an empty inline asm after __msan_report* to avoid callback merge.
314 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
315 StringRef(""), StringRef(""),
316 /*hasSideEffects=*/true);
319 /// \brief Module-level initialization.
321 /// inserts a call to __msan_init to the module's constructor list.
322 bool MemorySanitizer::doInitialization(Module &M) {
323 TD = getAnalysisIfAvailable<DataLayout>();
326 BL.reset(new BlackList(ClBlackListFile));
327 C = &(M.getContext());
328 unsigned PtrSize = TD->getPointerSizeInBits(/* AddressSpace */0);
331 ShadowMask = kShadowMask64;
332 OriginOffset = kOriginOffset64;
335 ShadowMask = kShadowMask32;
336 OriginOffset = kOriginOffset32;
339 report_fatal_error("unsupported pointer size");
344 IntptrTy = IRB.getIntPtrTy(TD);
345 OriginTy = IRB.getInt32Ty();
347 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
348 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
350 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
351 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
352 "__msan_init", IRB.getVoidTy(), NULL)), 0);
354 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
355 IRB.getInt32(TrackOrigins), "__msan_track_origins");
362 /// \brief A helper class that handles instrumentation of VarArg
363 /// functions on a particular platform.
365 /// Implementations are expected to insert the instrumentation
366 /// necessary to propagate argument shadow through VarArg function
367 /// calls. Visit* methods are called during an InstVisitor pass over
368 /// the function, and should avoid creating new basic blocks. A new
369 /// instance of this class is created for each instrumented function.
370 struct VarArgHelper {
371 /// \brief Visit a CallSite.
372 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
374 /// \brief Visit a va_start call.
375 virtual void visitVAStartInst(VAStartInst &I) = 0;
377 /// \brief Visit a va_copy call.
378 virtual void visitVACopyInst(VACopyInst &I) = 0;
380 /// \brief Finalize function instrumentation.
382 /// This method is called after visiting all interesting (see above)
383 /// instructions in a function.
384 virtual void finalizeInstrumentation() = 0;
386 virtual ~VarArgHelper() {}
389 struct MemorySanitizerVisitor;
392 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
393 MemorySanitizerVisitor &Visitor);
395 /// This class does all the work for a given function. Store and Load
396 /// instructions store and load corresponding shadow and origin
397 /// values. Most instructions propagate shadow from arguments to their
398 /// return values. Certain instructions (most importantly, BranchInst)
399 /// test their argument shadow and print reports (with a runtime call) if it's
401 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
404 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
405 ValueMap<Value*, Value*> ShadowMap, OriginMap;
407 OwningPtr<VarArgHelper> VAHelper;
409 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
410 // See a comment in visitCallSite for more details.
411 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
412 static const unsigned AMD64FpEndOffset = 176;
414 struct ShadowOriginAndInsertPoint {
417 Instruction *OrigIns;
418 ShadowOriginAndInsertPoint(Instruction *S, Instruction *O, Instruction *I)
419 : Shadow(S), Origin(O), OrigIns(I) { }
420 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
422 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
423 SmallVector<Instruction*, 16> StoreList;
425 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
426 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
427 InsertChecks = !MS.BL->isIn(F);
428 DEBUG(if (!InsertChecks)
429 dbgs() << "MemorySanitizer is not inserting checks into '"
430 << F.getName() << "'\n");
433 void materializeStores() {
434 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
435 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]);
438 Value *Val = I.getValueOperand();
439 Value *Addr = I.getPointerOperand();
440 Value *Shadow = getShadow(Val);
441 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
444 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
445 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
447 // If the store is volatile, add a check.
449 insertCheck(Val, &I);
450 if (ClCheckAccessAddress)
451 insertCheck(Addr, &I);
453 if (MS.TrackOrigins) {
454 if (ClStoreCleanOrigin || isa<StructType>(Shadow->getType())) {
455 IRB.CreateStore(getOrigin(Val), getOriginPtr(Addr, IRB));
457 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
459 Constant *Cst = dyn_cast_or_null<Constant>(ConvertedShadow);
460 // TODO(eugenis): handle non-zero constant shadow by inserting an
461 // unconditional check (can not simply fail compilation as this could
462 // be in the dead code).
466 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
467 getCleanShadow(ConvertedShadow), "_mscmp");
468 Instruction *CheckTerm =
469 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false,
470 MS.OriginStoreWeights);
471 IRBuilder<> IRBNew(CheckTerm);
472 IRBNew.CreateStore(getOrigin(Val), getOriginPtr(Addr, IRBNew));
478 void materializeChecks() {
479 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
480 Instruction *Shadow = InstrumentationList[i].Shadow;
481 Instruction *OrigIns = InstrumentationList[i].OrigIns;
482 IRBuilder<> IRB(OrigIns);
483 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
484 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
485 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
486 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
487 getCleanShadow(ConvertedShadow), "_mscmp");
488 Instruction *CheckTerm =
489 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp),
490 /* Unreachable */ !ClKeepGoing,
493 IRB.SetInsertPoint(CheckTerm);
494 if (MS.TrackOrigins) {
495 Instruction *Origin = InstrumentationList[i].Origin;
496 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
499 CallInst *Call = IRB.CreateCall(MS.WarningFn);
500 Call->setDebugLoc(OrigIns->getDebugLoc());
501 IRB.CreateCall(MS.EmptyAsm);
502 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
504 DEBUG(dbgs() << "DONE:\n" << F);
507 /// \brief Add MemorySanitizer instrumentation to a function.
508 bool runOnFunction() {
509 MS.initializeCallbacks(*F.getParent());
510 if (!MS.TD) return false;
512 // In the presence of unreachable blocks, we may see Phi nodes with
513 // incoming nodes from such blocks. Since InstVisitor skips unreachable
514 // blocks, such nodes will not have any shadow value associated with them.
515 // It's easier to remove unreachable blocks than deal with missing shadow.
516 removeUnreachableBlocks(F);
518 // Iterate all BBs in depth-first order and create shadow instructions
519 // for all instructions (where applicable).
520 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
521 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
522 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
523 BasicBlock *BB = *DI;
527 // Finalize PHI nodes.
528 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
529 PHINode *PN = ShadowPHINodes[i];
530 PHINode *PNS = cast<PHINode>(getShadow(PN));
531 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
532 size_t NumValues = PN->getNumIncomingValues();
533 for (size_t v = 0; v < NumValues; v++) {
534 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
536 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
540 VAHelper->finalizeInstrumentation();
542 // Delayed instrumentation of StoreInst.
543 // This may add new checks to be inserted later.
546 // Insert shadow value checks.
552 /// \brief Compute the shadow type that corresponds to a given Value.
553 Type *getShadowTy(Value *V) {
554 return getShadowTy(V->getType());
557 /// \brief Compute the shadow type that corresponds to a given Type.
558 Type *getShadowTy(Type *OrigTy) {
559 if (!OrigTy->isSized()) {
562 // For integer type, shadow is the same as the original type.
563 // This may return weird-sized types like i1.
564 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
566 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
567 uint32_t EltSize = MS.TD->getTypeStoreSizeInBits(VT->getElementType());
568 return VectorType::get(IntegerType::get(*MS.C, EltSize),
569 VT->getNumElements());
571 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
572 SmallVector<Type*, 4> Elements;
573 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
574 Elements.push_back(getShadowTy(ST->getElementType(i)));
575 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
576 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
579 uint32_t TypeSize = MS.TD->getTypeStoreSizeInBits(OrigTy);
580 return IntegerType::get(*MS.C, TypeSize);
583 /// \brief Flatten a vector type.
584 Type *getShadowTyNoVec(Type *ty) {
585 if (VectorType *vt = dyn_cast<VectorType>(ty))
586 return IntegerType::get(*MS.C, vt->getBitWidth());
590 /// \brief Convert a shadow value to it's flattened variant.
591 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
592 Type *Ty = V->getType();
593 Type *NoVecTy = getShadowTyNoVec(Ty);
594 if (Ty == NoVecTy) return V;
595 return IRB.CreateBitCast(V, NoVecTy);
598 /// \brief Compute the shadow address that corresponds to a given application
601 /// Shadow = Addr & ~ShadowMask.
602 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
605 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
606 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
607 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
610 /// \brief Compute the origin address that corresponds to a given application
613 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
614 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
616 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
617 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
619 IRB.CreateAdd(ShadowLong,
620 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
622 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
623 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
626 /// \brief Compute the shadow address for a given function argument.
628 /// Shadow = ParamTLS+ArgOffset.
629 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
631 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
632 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
633 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
637 /// \brief Compute the origin address for a given function argument.
638 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
640 if (!MS.TrackOrigins) return 0;
641 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
642 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
643 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
647 /// \brief Compute the shadow address for a retval.
648 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
649 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
650 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
654 /// \brief Compute the origin address for a retval.
655 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
656 // We keep a single origin for the entire retval. Might be too optimistic.
657 return MS.RetvalOriginTLS;
660 /// \brief Set SV to be the shadow value for V.
661 void setShadow(Value *V, Value *SV) {
662 assert(!ShadowMap.count(V) && "Values may only have one shadow");
666 /// \brief Set Origin to be the origin value for V.
667 void setOrigin(Value *V, Value *Origin) {
668 if (!MS.TrackOrigins) return;
669 assert(!OriginMap.count(V) && "Values may only have one origin");
670 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
671 OriginMap[V] = Origin;
674 /// \brief Create a clean shadow value for a given value.
676 /// Clean shadow (all zeroes) means all bits of the value are defined
678 Value *getCleanShadow(Value *V) {
679 Type *ShadowTy = getShadowTy(V);
682 return Constant::getNullValue(ShadowTy);
685 /// \brief Create a dirty shadow of a given shadow type.
686 Constant *getPoisonedShadow(Type *ShadowTy) {
688 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
689 return Constant::getAllOnesValue(ShadowTy);
690 StructType *ST = cast<StructType>(ShadowTy);
691 SmallVector<Constant *, 4> Vals;
692 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
693 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
694 return ConstantStruct::get(ST, Vals);
697 /// \brief Create a clean (zero) origin.
698 Value *getCleanOrigin() {
699 return Constant::getNullValue(MS.OriginTy);
702 /// \brief Get the shadow value for a given Value.
704 /// This function either returns the value set earlier with setShadow,
705 /// or extracts if from ParamTLS (for function arguments).
706 Value *getShadow(Value *V) {
707 if (Instruction *I = dyn_cast<Instruction>(V)) {
708 // For instructions the shadow is already stored in the map.
709 Value *Shadow = ShadowMap[V];
711 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
713 assert(Shadow && "No shadow for a value");
717 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
718 Value *AllOnes = getPoisonedShadow(getShadowTy(V));
719 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
723 if (Argument *A = dyn_cast<Argument>(V)) {
724 // For arguments we compute the shadow on demand and store it in the map.
725 Value **ShadowPtr = &ShadowMap[V];
728 Function *F = A->getParent();
729 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
730 unsigned ArgOffset = 0;
731 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
733 if (!AI->getType()->isSized()) {
734 DEBUG(dbgs() << "Arg is not sized\n");
737 unsigned Size = AI->hasByValAttr()
738 ? MS.TD->getTypeAllocSize(AI->getType()->getPointerElementType())
739 : MS.TD->getTypeAllocSize(AI->getType());
741 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
742 if (AI->hasByValAttr()) {
743 // ByVal pointer itself has clean shadow. We copy the actual
744 // argument shadow to the underlying memory.
745 Value *Cpy = EntryIRB.CreateMemCpy(
746 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
747 Base, Size, AI->getParamAlignment());
748 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
750 *ShadowPtr = getCleanShadow(V);
752 *ShadowPtr = EntryIRB.CreateLoad(Base);
754 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
755 **ShadowPtr << "\n");
756 if (MS.TrackOrigins) {
757 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
758 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
761 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
763 assert(*ShadowPtr && "Could not find shadow for an argument");
766 // For everything else the shadow is zero.
767 return getCleanShadow(V);
770 /// \brief Get the shadow for i-th argument of the instruction I.
771 Value *getShadow(Instruction *I, int i) {
772 return getShadow(I->getOperand(i));
775 /// \brief Get the origin for a value.
776 Value *getOrigin(Value *V) {
777 if (!MS.TrackOrigins) return 0;
778 if (isa<Instruction>(V) || isa<Argument>(V)) {
779 Value *Origin = OriginMap[V];
781 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
782 Origin = getCleanOrigin();
786 return getCleanOrigin();
789 /// \brief Get the origin for i-th argument of the instruction I.
790 Value *getOrigin(Instruction *I, int i) {
791 return getOrigin(I->getOperand(i));
794 /// \brief Remember the place where a shadow check should be inserted.
796 /// This location will be later instrumented with a check that will print a
797 /// UMR warning in runtime if the value is not fully defined.
798 void insertCheck(Value *Val, Instruction *OrigIns) {
800 if (!InsertChecks) return;
801 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
804 Type *ShadowTy = Shadow->getType();
805 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
806 "Can only insert checks for integer and vector shadow types");
808 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
809 InstrumentationList.push_back(
810 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
813 // ------------------- Visitors.
815 /// \brief Instrument LoadInst
817 /// Loads the corresponding shadow and (optionally) origin.
818 /// Optionally, checks that the load address is fully defined.
819 void visitLoadInst(LoadInst &I) {
820 assert(I.getType()->isSized() && "Load type must have size");
822 Type *ShadowTy = getShadowTy(&I);
823 Value *Addr = I.getPointerOperand();
824 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
825 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
827 if (ClCheckAccessAddress)
828 insertCheck(I.getPointerOperand(), &I);
831 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
834 /// \brief Instrument StoreInst
836 /// Stores the corresponding shadow and (optionally) origin.
837 /// Optionally, checks that the store address is fully defined.
838 /// Volatile stores check that the value being stored is fully defined.
839 void visitStoreInst(StoreInst &I) {
840 StoreList.push_back(&I);
843 // Vector manipulation.
844 void visitExtractElementInst(ExtractElementInst &I) {
845 insertCheck(I.getOperand(1), &I);
847 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
849 setOrigin(&I, getOrigin(&I, 0));
852 void visitInsertElementInst(InsertElementInst &I) {
853 insertCheck(I.getOperand(2), &I);
855 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
856 I.getOperand(2), "_msprop"));
857 setOriginForNaryOp(I);
860 void visitShuffleVectorInst(ShuffleVectorInst &I) {
861 insertCheck(I.getOperand(2), &I);
863 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
864 I.getOperand(2), "_msprop"));
865 setOriginForNaryOp(I);
869 void visitSExtInst(SExtInst &I) {
871 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
872 setOrigin(&I, getOrigin(&I, 0));
875 void visitZExtInst(ZExtInst &I) {
877 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
878 setOrigin(&I, getOrigin(&I, 0));
881 void visitTruncInst(TruncInst &I) {
883 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
884 setOrigin(&I, getOrigin(&I, 0));
887 void visitBitCastInst(BitCastInst &I) {
889 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
890 setOrigin(&I, getOrigin(&I, 0));
893 void visitPtrToIntInst(PtrToIntInst &I) {
895 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
896 "_msprop_ptrtoint"));
897 setOrigin(&I, getOrigin(&I, 0));
900 void visitIntToPtrInst(IntToPtrInst &I) {
902 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
903 "_msprop_inttoptr"));
904 setOrigin(&I, getOrigin(&I, 0));
907 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
908 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
909 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
910 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
911 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
912 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
914 /// \brief Propagate shadow for bitwise AND.
916 /// This code is exact, i.e. if, for example, a bit in the left argument
917 /// is defined and 0, then neither the value not definedness of the
918 /// corresponding bit in B don't affect the resulting shadow.
919 void visitAnd(BinaryOperator &I) {
921 // "And" of 0 and a poisoned value results in unpoisoned value.
922 // 1&1 => 1; 0&1 => 0; p&1 => p;
923 // 1&0 => 0; 0&0 => 0; p&0 => 0;
924 // 1&p => p; 0&p => 0; p&p => p;
925 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
926 Value *S1 = getShadow(&I, 0);
927 Value *S2 = getShadow(&I, 1);
928 Value *V1 = I.getOperand(0);
929 Value *V2 = I.getOperand(1);
930 if (V1->getType() != S1->getType()) {
931 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
932 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
934 Value *S1S2 = IRB.CreateAnd(S1, S2);
935 Value *V1S2 = IRB.CreateAnd(V1, S2);
936 Value *S1V2 = IRB.CreateAnd(S1, V2);
937 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
938 setOriginForNaryOp(I);
941 void visitOr(BinaryOperator &I) {
943 // "Or" of 1 and a poisoned value results in unpoisoned value.
944 // 1|1 => 1; 0|1 => 1; p|1 => 1;
945 // 1|0 => 1; 0|0 => 0; p|0 => p;
946 // 1|p => 1; 0|p => p; p|p => p;
947 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
948 Value *S1 = getShadow(&I, 0);
949 Value *S2 = getShadow(&I, 1);
950 Value *V1 = IRB.CreateNot(I.getOperand(0));
951 Value *V2 = IRB.CreateNot(I.getOperand(1));
952 if (V1->getType() != S1->getType()) {
953 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
954 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
956 Value *S1S2 = IRB.CreateAnd(S1, S2);
957 Value *V1S2 = IRB.CreateAnd(V1, S2);
958 Value *S1V2 = IRB.CreateAnd(S1, V2);
959 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
960 setOriginForNaryOp(I);
963 /// \brief Default propagation of shadow and/or origin.
965 /// This class implements the general case of shadow propagation, used in all
966 /// cases where we don't know and/or don't care about what the operation
967 /// actually does. It converts all input shadow values to a common type
968 /// (extending or truncating as necessary), and bitwise OR's them.
970 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
971 /// fully initialized), and less prone to false positives.
973 /// This class also implements the general case of origin propagation. For a
974 /// Nary operation, result origin is set to the origin of an argument that is
975 /// not entirely initialized. If there is more than one such arguments, the
976 /// rightmost of them is picked. It does not matter which one is picked if all
977 /// arguments are initialized.
978 template <bool CombineShadow>
983 MemorySanitizerVisitor *MSV;
986 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
987 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {}
989 /// \brief Add a pair of shadow and origin values to the mix.
990 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
996 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
997 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1001 if (MSV->MS.TrackOrigins) {
1006 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1007 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1008 MSV->getCleanShadow(FlatShadow));
1009 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1015 /// \brief Add an application value to the mix.
1016 Combiner &Add(Value *V) {
1017 Value *OpShadow = MSV->getShadow(V);
1018 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : 0;
1019 return Add(OpShadow, OpOrigin);
1022 /// \brief Set the current combined values as the given instruction's shadow
1024 void Done(Instruction *I) {
1025 if (CombineShadow) {
1027 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1028 MSV->setShadow(I, Shadow);
1030 if (MSV->MS.TrackOrigins) {
1032 MSV->setOrigin(I, Origin);
1037 typedef Combiner<true> ShadowAndOriginCombiner;
1038 typedef Combiner<false> OriginCombiner;
1040 /// \brief Propagate origin for arbitrary operation.
1041 void setOriginForNaryOp(Instruction &I) {
1042 if (!MS.TrackOrigins) return;
1043 IRBuilder<> IRB(&I);
1044 OriginCombiner OC(this, IRB);
1045 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1050 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1051 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1052 "Vector of pointers is not a valid shadow type");
1053 return Ty->isVectorTy() ?
1054 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1055 Ty->getPrimitiveSizeInBits();
1058 /// \brief Cast between two shadow types, extending or truncating as
1060 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy) {
1061 Type *srcTy = V->getType();
1062 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1063 return IRB.CreateIntCast(V, dstTy, false);
1064 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1065 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1066 return IRB.CreateIntCast(V, dstTy, false);
1067 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1068 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1069 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1071 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), false);
1072 return IRB.CreateBitCast(V2, dstTy);
1073 // TODO: handle struct types.
1076 /// \brief Propagate shadow for arbitrary operation.
1077 void handleShadowOr(Instruction &I) {
1078 IRBuilder<> IRB(&I);
1079 ShadowAndOriginCombiner SC(this, IRB);
1080 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1085 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1086 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1087 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1088 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1089 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1090 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1091 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1093 void handleDiv(Instruction &I) {
1094 IRBuilder<> IRB(&I);
1095 // Strict on the second argument.
1096 insertCheck(I.getOperand(1), &I);
1097 setShadow(&I, getShadow(&I, 0));
1098 setOrigin(&I, getOrigin(&I, 0));
1101 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1102 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1103 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1104 void visitURem(BinaryOperator &I) { handleDiv(I); }
1105 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1106 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1108 /// \brief Instrument == and != comparisons.
1110 /// Sometimes the comparison result is known even if some of the bits of the
1111 /// arguments are not.
1112 void handleEqualityComparison(ICmpInst &I) {
1113 IRBuilder<> IRB(&I);
1114 Value *A = I.getOperand(0);
1115 Value *B = I.getOperand(1);
1116 Value *Sa = getShadow(A);
1117 Value *Sb = getShadow(B);
1118 if (A->getType()->isPointerTy())
1119 A = IRB.CreatePointerCast(A, MS.IntptrTy);
1120 if (B->getType()->isPointerTy())
1121 B = IRB.CreatePointerCast(B, MS.IntptrTy);
1122 // A == B <==> (C = A^B) == 0
1123 // A != B <==> (C = A^B) != 0
1125 Value *C = IRB.CreateXor(A, B);
1126 Value *Sc = IRB.CreateOr(Sa, Sb);
1127 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1128 // Result is defined if one of the following is true
1129 // * there is a defined 1 bit in C
1130 // * C is fully defined
1131 // Si = !(C & ~Sc) && Sc
1132 Value *Zero = Constant::getNullValue(Sc->getType());
1133 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1135 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1137 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1138 Si->setName("_msprop_icmp");
1140 setOriginForNaryOp(I);
1143 /// \brief Instrument signed relational comparisons.
1145 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1146 /// propagating the highest bit of the shadow. Everything else is delegated
1147 /// to handleShadowOr().
1148 void handleSignedRelationalComparison(ICmpInst &I) {
1149 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1150 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1152 CmpInst::Predicate pre = I.getPredicate();
1153 if (constOp0 && constOp0->isNullValue() &&
1154 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1155 op = I.getOperand(1);
1156 } else if (constOp1 && constOp1->isNullValue() &&
1157 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1158 op = I.getOperand(0);
1161 IRBuilder<> IRB(&I);
1163 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1164 setShadow(&I, Shadow);
1165 setOrigin(&I, getOrigin(op));
1171 void visitICmpInst(ICmpInst &I) {
1172 if (ClHandleICmp && I.isEquality())
1173 handleEqualityComparison(I);
1174 else if (ClHandleICmp && I.isSigned() && I.isRelational())
1175 handleSignedRelationalComparison(I);
1180 void visitFCmpInst(FCmpInst &I) {
1184 void handleShift(BinaryOperator &I) {
1185 IRBuilder<> IRB(&I);
1186 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1187 // Otherwise perform the same shift on S1.
1188 Value *S1 = getShadow(&I, 0);
1189 Value *S2 = getShadow(&I, 1);
1190 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1192 Value *V2 = I.getOperand(1);
1193 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1194 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1195 setOriginForNaryOp(I);
1198 void visitShl(BinaryOperator &I) { handleShift(I); }
1199 void visitAShr(BinaryOperator &I) { handleShift(I); }
1200 void visitLShr(BinaryOperator &I) { handleShift(I); }
1202 /// \brief Instrument llvm.memmove
1204 /// At this point we don't know if llvm.memmove will be inlined or not.
1205 /// If we don't instrument it and it gets inlined,
1206 /// our interceptor will not kick in and we will lose the memmove.
1207 /// If we instrument the call here, but it does not get inlined,
1208 /// we will memove the shadow twice: which is bad in case
1209 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1211 /// Similar situation exists for memcpy and memset.
1212 void visitMemMoveInst(MemMoveInst &I) {
1213 IRBuilder<> IRB(&I);
1216 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1217 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1218 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1219 I.eraseFromParent();
1222 // Similar to memmove: avoid copying shadow twice.
1223 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1224 // FIXME: consider doing manual inline for small constant sizes and proper
1226 void visitMemCpyInst(MemCpyInst &I) {
1227 IRBuilder<> IRB(&I);
1230 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1231 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1232 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1233 I.eraseFromParent();
1237 void visitMemSetInst(MemSetInst &I) {
1238 IRBuilder<> IRB(&I);
1241 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1242 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1243 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1244 I.eraseFromParent();
1247 void visitVAStartInst(VAStartInst &I) {
1248 VAHelper->visitVAStartInst(I);
1251 void visitVACopyInst(VACopyInst &I) {
1252 VAHelper->visitVACopyInst(I);
1255 enum IntrinsicKind {
1256 IK_DoesNotAccessMemory,
1261 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1262 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1263 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1264 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1265 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1266 const int UnknownModRefBehavior = IK_WritesMemory;
1267 #define GET_INTRINSIC_MODREF_BEHAVIOR
1268 #define ModRefBehavior IntrinsicKind
1269 #include "llvm/Intrinsics.gen"
1270 #undef ModRefBehavior
1271 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1274 /// \brief Handle vector store-like intrinsics.
1276 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1277 /// has 1 pointer argument and 1 vector argument, returns void.
1278 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1279 IRBuilder<> IRB(&I);
1280 Value* Addr = I.getArgOperand(0);
1281 Value *Shadow = getShadow(&I, 1);
1282 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1284 // We don't know the pointer alignment (could be unaligned SSE store!).
1285 // Have to assume to worst case.
1286 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1288 if (ClCheckAccessAddress)
1289 insertCheck(Addr, &I);
1291 // FIXME: use ClStoreCleanOrigin
1292 // FIXME: factor out common code from materializeStores
1293 if (MS.TrackOrigins)
1294 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1298 /// \brief Handle vector load-like intrinsics.
1300 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1301 /// has 1 pointer argument, returns a vector.
1302 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1303 IRBuilder<> IRB(&I);
1304 Value *Addr = I.getArgOperand(0);
1306 Type *ShadowTy = getShadowTy(&I);
1307 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1308 // We don't know the pointer alignment (could be unaligned SSE load!).
1309 // Have to assume to worst case.
1310 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1312 if (ClCheckAccessAddress)
1313 insertCheck(Addr, &I);
1315 if (MS.TrackOrigins)
1316 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1320 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1322 /// Instrument intrinsics with any number of arguments of the same type,
1323 /// equal to the return type. The type should be simple (no aggregates or
1324 /// pointers; vectors are fine).
1325 /// Caller guarantees that this intrinsic does not access memory.
1326 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1327 Type *RetTy = I.getType();
1328 if (!(RetTy->isIntOrIntVectorTy() ||
1329 RetTy->isFPOrFPVectorTy() ||
1330 RetTy->isX86_MMXTy()))
1333 unsigned NumArgOperands = I.getNumArgOperands();
1335 for (unsigned i = 0; i < NumArgOperands; ++i) {
1336 Type *Ty = I.getArgOperand(i)->getType();
1341 IRBuilder<> IRB(&I);
1342 ShadowAndOriginCombiner SC(this, IRB);
1343 for (unsigned i = 0; i < NumArgOperands; ++i)
1344 SC.Add(I.getArgOperand(i));
1350 /// \brief Heuristically instrument unknown intrinsics.
1352 /// The main purpose of this code is to do something reasonable with all
1353 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1354 /// We recognize several classes of intrinsics by their argument types and
1355 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1356 /// sure that we know what the intrinsic does.
1358 /// We special-case intrinsics where this approach fails. See llvm.bswap
1359 /// handling as an example of that.
1360 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1361 unsigned NumArgOperands = I.getNumArgOperands();
1362 if (NumArgOperands == 0)
1365 Intrinsic::ID iid = I.getIntrinsicID();
1366 IntrinsicKind IK = getIntrinsicKind(iid);
1367 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1368 bool WritesMemory = IK == IK_WritesMemory;
1369 assert(!(OnlyReadsMemory && WritesMemory));
1371 if (NumArgOperands == 2 &&
1372 I.getArgOperand(0)->getType()->isPointerTy() &&
1373 I.getArgOperand(1)->getType()->isVectorTy() &&
1374 I.getType()->isVoidTy() &&
1376 // This looks like a vector store.
1377 return handleVectorStoreIntrinsic(I);
1380 if (NumArgOperands == 1 &&
1381 I.getArgOperand(0)->getType()->isPointerTy() &&
1382 I.getType()->isVectorTy() &&
1384 // This looks like a vector load.
1385 return handleVectorLoadIntrinsic(I);
1388 if (!OnlyReadsMemory && !WritesMemory)
1389 if (maybeHandleSimpleNomemIntrinsic(I))
1392 // FIXME: detect and handle SSE maskstore/maskload
1396 void handleBswap(IntrinsicInst &I) {
1397 IRBuilder<> IRB(&I);
1398 Value *Op = I.getArgOperand(0);
1399 Type *OpType = Op->getType();
1400 Function *BswapFunc = Intrinsic::getDeclaration(
1401 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1402 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1403 setOrigin(&I, getOrigin(Op));
1406 void visitIntrinsicInst(IntrinsicInst &I) {
1407 switch (I.getIntrinsicID()) {
1408 case llvm::Intrinsic::bswap:
1412 if (!handleUnknownIntrinsic(I))
1413 visitInstruction(I);
1418 void visitCallSite(CallSite CS) {
1419 Instruction &I = *CS.getInstruction();
1420 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
1422 CallInst *Call = cast<CallInst>(&I);
1424 // For inline asm, do the usual thing: check argument shadow and mark all
1425 // outputs as clean. Note that any side effects of the inline asm that are
1426 // not immediately visible in its constraints are not handled.
1427 if (Call->isInlineAsm()) {
1428 visitInstruction(I);
1432 // Allow only tail calls with the same types, otherwise
1433 // we may have a false positive: shadow for a non-void RetVal
1434 // will get propagated to a void RetVal.
1435 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
1436 Call->setTailCall(false);
1438 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
1440 // We are going to insert code that relies on the fact that the callee
1441 // will become a non-readonly function after it is instrumented by us. To
1442 // prevent this code from being optimized out, mark that function
1443 // non-readonly in advance.
1444 if (Function *Func = Call->getCalledFunction()) {
1445 // Clear out readonly/readnone attributes.
1447 B.addAttribute(Attribute::ReadOnly)
1448 .addAttribute(Attribute::ReadNone);
1449 Func->removeAttribute(AttributeSet::FunctionIndex,
1450 Attribute::get(Func->getContext(), B));
1453 IRBuilder<> IRB(&I);
1454 unsigned ArgOffset = 0;
1455 DEBUG(dbgs() << " CallSite: " << I << "\n");
1456 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1457 ArgIt != End; ++ArgIt) {
1459 unsigned i = ArgIt - CS.arg_begin();
1460 if (!A->getType()->isSized()) {
1461 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
1466 // Compute the Shadow for arg even if it is ByVal, because
1467 // in that case getShadow() will copy the actual arg shadow to
1468 // __msan_param_tls.
1469 Value *ArgShadow = getShadow(A);
1470 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
1471 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
1472 " Shadow: " << *ArgShadow << "\n");
1473 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
1474 assert(A->getType()->isPointerTy() &&
1475 "ByVal argument is not a pointer!");
1476 Size = MS.TD->getTypeAllocSize(A->getType()->getPointerElementType());
1477 unsigned Alignment = CS.getParamAlignment(i + 1);
1478 Store = IRB.CreateMemCpy(ArgShadowBase,
1479 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
1482 Size = MS.TD->getTypeAllocSize(A->getType());
1483 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
1484 kShadowTLSAlignment);
1486 if (MS.TrackOrigins)
1487 IRB.CreateStore(getOrigin(A),
1488 getOriginPtrForArgument(A, IRB, ArgOffset));
1489 assert(Size != 0 && Store != 0);
1490 DEBUG(dbgs() << " Param:" << *Store << "\n");
1491 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
1493 DEBUG(dbgs() << " done with call args\n");
1496 cast<FunctionType>(CS.getCalledValue()->getType()-> getContainedType(0));
1497 if (FT->isVarArg()) {
1498 VAHelper->visitCallSite(CS, IRB);
1501 // Now, get the shadow for the RetVal.
1502 if (!I.getType()->isSized()) return;
1503 IRBuilder<> IRBBefore(&I);
1504 // Untill we have full dynamic coverage, make sure the retval shadow is 0.
1505 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
1506 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
1507 Instruction *NextInsn = 0;
1509 NextInsn = I.getNextNode();
1511 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
1512 if (!NormalDest->getSinglePredecessor()) {
1513 // FIXME: this case is tricky, so we are just conservative here.
1514 // Perhaps we need to split the edge between this BB and NormalDest,
1515 // but a naive attempt to use SplitEdge leads to a crash.
1516 setShadow(&I, getCleanShadow(&I));
1517 setOrigin(&I, getCleanOrigin());
1520 NextInsn = NormalDest->getFirstInsertionPt();
1522 "Could not find insertion point for retval shadow load");
1524 IRBuilder<> IRBAfter(NextInsn);
1525 Value *RetvalShadow =
1526 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
1527 kShadowTLSAlignment, "_msret");
1528 setShadow(&I, RetvalShadow);
1529 if (MS.TrackOrigins)
1530 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
1533 void visitReturnInst(ReturnInst &I) {
1534 IRBuilder<> IRB(&I);
1535 if (Value *RetVal = I.getReturnValue()) {
1536 // Set the shadow for the RetVal.
1537 Value *Shadow = getShadow(RetVal);
1538 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
1539 DEBUG(dbgs() << "Return: " << *Shadow << "\n" << *ShadowPtr << "\n");
1540 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
1541 if (MS.TrackOrigins)
1542 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
1546 void visitPHINode(PHINode &I) {
1547 IRBuilder<> IRB(&I);
1548 ShadowPHINodes.push_back(&I);
1549 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
1551 if (MS.TrackOrigins)
1552 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
1556 void visitAllocaInst(AllocaInst &I) {
1557 setShadow(&I, getCleanShadow(&I));
1558 if (!ClPoisonStack) return;
1559 IRBuilder<> IRB(I.getNextNode());
1560 uint64_t Size = MS.TD->getTypeAllocSize(I.getAllocatedType());
1561 if (ClPoisonStackWithCall) {
1562 IRB.CreateCall2(MS.MsanPoisonStackFn,
1563 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1564 ConstantInt::get(MS.IntptrTy, Size));
1566 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
1567 IRB.CreateMemSet(ShadowBase, IRB.getInt8(ClPoisonStackPattern),
1568 Size, I.getAlignment());
1571 if (MS.TrackOrigins) {
1572 setOrigin(&I, getCleanOrigin());
1573 SmallString<2048> StackDescriptionStorage;
1574 raw_svector_ostream StackDescription(StackDescriptionStorage);
1575 // We create a string with a description of the stack allocation and
1576 // pass it into __msan_set_alloca_origin.
1577 // It will be printed by the run-time if stack-originated UMR is found.
1578 // The first 4 bytes of the string are set to '----' and will be replaced
1579 // by __msan_va_arg_overflow_size_tls at the first call.
1580 StackDescription << "----" << I.getName() << "@" << F.getName();
1582 createPrivateNonConstGlobalForString(*F.getParent(),
1583 StackDescription.str());
1584 IRB.CreateCall3(MS.MsanSetAllocaOriginFn,
1585 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1586 ConstantInt::get(MS.IntptrTy, Size),
1587 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()));
1591 void visitSelectInst(SelectInst& I) {
1592 IRBuilder<> IRB(&I);
1593 setShadow(&I, IRB.CreateSelect(I.getCondition(),
1594 getShadow(I.getTrueValue()), getShadow(I.getFalseValue()),
1596 if (MS.TrackOrigins) {
1597 // Origins are always i32, so any vector conditions must be flattened.
1598 // FIXME: consider tracking vector origins for app vectors?
1599 Value *Cond = I.getCondition();
1600 if (Cond->getType()->isVectorTy()) {
1601 Value *ConvertedShadow = convertToShadowTyNoVec(Cond, IRB);
1602 Cond = IRB.CreateICmpNE(ConvertedShadow,
1603 getCleanShadow(ConvertedShadow), "_mso_select");
1605 setOrigin(&I, IRB.CreateSelect(Cond,
1606 getOrigin(I.getTrueValue()), getOrigin(I.getFalseValue())));
1610 void visitLandingPadInst(LandingPadInst &I) {
1612 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
1613 setShadow(&I, getCleanShadow(&I));
1614 setOrigin(&I, getCleanOrigin());
1617 void visitGetElementPtrInst(GetElementPtrInst &I) {
1621 void visitExtractValueInst(ExtractValueInst &I) {
1622 IRBuilder<> IRB(&I);
1623 Value *Agg = I.getAggregateOperand();
1624 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
1625 Value *AggShadow = getShadow(Agg);
1626 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1627 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
1628 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
1629 setShadow(&I, ResShadow);
1630 setOrigin(&I, getCleanOrigin());
1633 void visitInsertValueInst(InsertValueInst &I) {
1634 IRBuilder<> IRB(&I);
1635 DEBUG(dbgs() << "InsertValue: " << I << "\n");
1636 Value *AggShadow = getShadow(I.getAggregateOperand());
1637 Value *InsShadow = getShadow(I.getInsertedValueOperand());
1638 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1639 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
1640 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
1641 DEBUG(dbgs() << " Res: " << *Res << "\n");
1643 setOrigin(&I, getCleanOrigin());
1646 void dumpInst(Instruction &I) {
1647 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1648 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
1650 errs() << "ZZZ " << I.getOpcodeName() << "\n";
1652 errs() << "QQQ " << I << "\n";
1655 void visitResumeInst(ResumeInst &I) {
1656 DEBUG(dbgs() << "Resume: " << I << "\n");
1657 // Nothing to do here.
1660 void visitInstruction(Instruction &I) {
1661 // Everything else: stop propagating and check for poisoned shadow.
1662 if (ClDumpStrictInstructions)
1664 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
1665 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
1666 insertCheck(I.getOperand(i), &I);
1667 setShadow(&I, getCleanShadow(&I));
1668 setOrigin(&I, getCleanOrigin());
1672 /// \brief AMD64-specific implementation of VarArgHelper.
1673 struct VarArgAMD64Helper : public VarArgHelper {
1674 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
1675 // See a comment in visitCallSite for more details.
1676 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
1677 static const unsigned AMD64FpEndOffset = 176;
1680 MemorySanitizer &MS;
1681 MemorySanitizerVisitor &MSV;
1682 Value *VAArgTLSCopy;
1683 Value *VAArgOverflowSize;
1685 SmallVector<CallInst*, 16> VAStartInstrumentationList;
1687 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
1688 MemorySanitizerVisitor &MSV)
1689 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
1691 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
1693 ArgKind classifyArgument(Value* arg) {
1694 // A very rough approximation of X86_64 argument classification rules.
1695 Type *T = arg->getType();
1696 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
1697 return AK_FloatingPoint;
1698 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
1699 return AK_GeneralPurpose;
1700 if (T->isPointerTy())
1701 return AK_GeneralPurpose;
1705 // For VarArg functions, store the argument shadow in an ABI-specific format
1706 // that corresponds to va_list layout.
1707 // We do this because Clang lowers va_arg in the frontend, and this pass
1708 // only sees the low level code that deals with va_list internals.
1709 // A much easier alternative (provided that Clang emits va_arg instructions)
1710 // would have been to associate each live instance of va_list with a copy of
1711 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
1713 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) {
1714 unsigned GpOffset = 0;
1715 unsigned FpOffset = AMD64GpEndOffset;
1716 unsigned OverflowOffset = AMD64FpEndOffset;
1717 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1718 ArgIt != End; ++ArgIt) {
1720 ArgKind AK = classifyArgument(A);
1721 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
1723 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
1727 case AK_GeneralPurpose:
1728 Base = getShadowPtrForVAArgument(A, IRB, GpOffset);
1731 case AK_FloatingPoint:
1732 Base = getShadowPtrForVAArgument(A, IRB, FpOffset);
1736 uint64_t ArgSize = MS.TD->getTypeAllocSize(A->getType());
1737 Base = getShadowPtrForVAArgument(A, IRB, OverflowOffset);
1738 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
1740 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
1742 Constant *OverflowSize =
1743 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
1744 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
1747 /// \brief Compute the shadow address for a given va_arg.
1748 Value *getShadowPtrForVAArgument(Value *A, IRBuilder<> &IRB,
1750 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
1751 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
1752 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(A), 0),
1756 void visitVAStartInst(VAStartInst &I) {
1757 IRBuilder<> IRB(&I);
1758 VAStartInstrumentationList.push_back(&I);
1759 Value *VAListTag = I.getArgOperand(0);
1760 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1762 // Unpoison the whole __va_list_tag.
1763 // FIXME: magic ABI constants.
1764 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1765 /* size */24, /* alignment */16, false);
1768 void visitVACopyInst(VACopyInst &I) {
1769 IRBuilder<> IRB(&I);
1770 Value *VAListTag = I.getArgOperand(0);
1771 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1773 // Unpoison the whole __va_list_tag.
1774 // FIXME: magic ABI constants.
1775 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1776 /* size */ 24, /* alignment */ 16, false);
1779 void finalizeInstrumentation() {
1780 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
1781 "finalizeInstrumentation called twice");
1782 if (!VAStartInstrumentationList.empty()) {
1783 // If there is a va_start in this function, make a backup copy of
1784 // va_arg_tls somewhere in the function entry block.
1785 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
1786 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
1788 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
1790 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
1791 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
1794 // Instrument va_start.
1795 // Copy va_list shadow from the backup copy of the TLS contents.
1796 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
1797 CallInst *OrigInst = VAStartInstrumentationList[i];
1798 IRBuilder<> IRB(OrigInst->getNextNode());
1799 Value *VAListTag = OrigInst->getArgOperand(0);
1801 Value *RegSaveAreaPtrPtr =
1803 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1804 ConstantInt::get(MS.IntptrTy, 16)),
1805 Type::getInt64PtrTy(*MS.C));
1806 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
1807 Value *RegSaveAreaShadowPtr =
1808 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
1809 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
1810 AMD64FpEndOffset, 16);
1812 Value *OverflowArgAreaPtrPtr =
1814 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1815 ConstantInt::get(MS.IntptrTy, 8)),
1816 Type::getInt64PtrTy(*MS.C));
1817 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
1818 Value *OverflowArgAreaShadowPtr =
1819 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
1821 getShadowPtrForVAArgument(VAArgTLSCopy, IRB, AMD64FpEndOffset);
1822 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
1827 VarArgHelper* CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
1828 MemorySanitizerVisitor &Visitor) {
1829 return new VarArgAMD64Helper(Func, Msan, Visitor);
1834 bool MemorySanitizer::runOnFunction(Function &F) {
1835 MemorySanitizerVisitor Visitor(F, *this);
1837 // Clear out readonly/readnone attributes.
1839 B.addAttribute(Attribute::ReadOnly)
1840 .addAttribute(Attribute::ReadNone);
1841 F.removeAttribute(AttributeSet::FunctionIndex,
1842 Attribute::get(F.getContext(), B));
1844 return Visitor.runOnFunction();