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 "llvm/ADT/DepthFirstIterator.h"
75 #include "llvm/ADT/SmallString.h"
76 #include "llvm/ADT/SmallVector.h"
77 #include "llvm/ADT/ValueMap.h"
78 #include "llvm/IR/DataLayout.h"
79 #include "llvm/IR/Function.h"
80 #include "llvm/IR/IRBuilder.h"
81 #include "llvm/IR/InlineAsm.h"
82 #include "llvm/IR/IntrinsicInst.h"
83 #include "llvm/IR/LLVMContext.h"
84 #include "llvm/IR/MDBuilder.h"
85 #include "llvm/IR/Module.h"
86 #include "llvm/IR/Type.h"
87 #include "llvm/InstVisitor.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Compiler.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/BlackList.h"
94 #include "llvm/Transforms/Utils/Local.h"
95 #include "llvm/Transforms/Utils/ModuleUtils.h"
99 static const uint64_t kShadowMask32 = 1ULL << 31;
100 static const uint64_t kShadowMask64 = 1ULL << 46;
101 static const uint64_t kOriginOffset32 = 1ULL << 30;
102 static const uint64_t kOriginOffset64 = 1ULL << 45;
103 static const unsigned kMinOriginAlignment = 4;
104 static const unsigned kShadowTLSAlignment = 8;
106 /// \brief Track origins of uninitialized values.
108 /// Adds a section to MemorySanitizer report that points to the allocation
109 /// (stack or heap) the uninitialized bits came from originally.
110 static cl::opt<bool> ClTrackOrigins("msan-track-origins",
111 cl::desc("Track origins (allocation sites) of poisoned memory"),
112 cl::Hidden, cl::init(false));
113 static cl::opt<bool> ClKeepGoing("msan-keep-going",
114 cl::desc("keep going after reporting a UMR"),
115 cl::Hidden, cl::init(false));
116 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
117 cl::desc("poison uninitialized stack variables"),
118 cl::Hidden, cl::init(true));
119 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
120 cl::desc("poison uninitialized stack variables with a call"),
121 cl::Hidden, cl::init(false));
122 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
123 cl::desc("poison uninitialized stack variables with the given patter"),
124 cl::Hidden, cl::init(0xff));
125 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
126 cl::desc("poison undef temps"),
127 cl::Hidden, cl::init(true));
129 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
130 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
131 cl::Hidden, cl::init(true));
133 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
134 cl::desc("exact handling of relational integer ICmp"),
135 cl::Hidden, cl::init(false));
137 static cl::opt<bool> ClStoreCleanOrigin("msan-store-clean-origin",
138 cl::desc("store origin for clean (fully initialized) values"),
139 cl::Hidden, cl::init(false));
141 // This flag controls whether we check the shadow of the address
142 // operand of load or store. Such bugs are very rare, since load from
143 // a garbage address typically results in SEGV, but still happen
144 // (e.g. only lower bits of address are garbage, or the access happens
145 // early at program startup where malloc-ed memory is more likely to
146 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
147 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
148 cl::desc("report accesses through a pointer which has poisoned shadow"),
149 cl::Hidden, cl::init(true));
151 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
152 cl::desc("print out instructions with default strict semantics"),
153 cl::Hidden, cl::init(false));
155 static cl::opt<std::string> ClBlacklistFile("msan-blacklist",
156 cl::desc("File containing the list of functions where MemorySanitizer "
157 "should not report bugs"), cl::Hidden);
161 /// \brief An instrumentation pass implementing detection of uninitialized
164 /// MemorySanitizer: instrument the code in module to find
165 /// uninitialized reads.
166 class MemorySanitizer : public FunctionPass {
168 MemorySanitizer(bool TrackOrigins = false,
169 StringRef BlacklistFile = StringRef())
171 TrackOrigins(TrackOrigins || ClTrackOrigins),
174 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
176 const char *getPassName() const { return "MemorySanitizer"; }
177 bool runOnFunction(Function &F);
178 bool doInitialization(Module &M);
179 static char ID; // Pass identification, replacement for typeid.
182 void initializeCallbacks(Module &M);
184 /// \brief Track origins (allocation points) of uninitialized values.
191 /// \brief Thread-local shadow storage for function parameters.
192 GlobalVariable *ParamTLS;
193 /// \brief Thread-local origin storage for function parameters.
194 GlobalVariable *ParamOriginTLS;
195 /// \brief Thread-local shadow storage for function return value.
196 GlobalVariable *RetvalTLS;
197 /// \brief Thread-local origin storage for function return value.
198 GlobalVariable *RetvalOriginTLS;
199 /// \brief Thread-local shadow storage for in-register va_arg function
200 /// parameters (x86_64-specific).
201 GlobalVariable *VAArgTLS;
202 /// \brief Thread-local shadow storage for va_arg overflow area
203 /// (x86_64-specific).
204 GlobalVariable *VAArgOverflowSizeTLS;
205 /// \brief Thread-local space used to pass origin value to the UMR reporting
207 GlobalVariable *OriginTLS;
209 /// \brief The run-time callback to print a warning.
211 /// \brief Run-time helper that copies origin info for a memory range.
212 Value *MsanCopyOriginFn;
213 /// \brief Run-time helper that generates a new origin value for a stack
215 Value *MsanSetAllocaOriginFn;
216 /// \brief Run-time helper that poisons stack on function entry.
217 Value *MsanPoisonStackFn;
218 /// \brief MSan runtime replacements for memmove, memcpy and memset.
219 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
221 /// \brief Address mask used in application-to-shadow address calculation.
222 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
224 /// \brief Offset of the origin shadow from the "normal" shadow.
225 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
226 uint64_t OriginOffset;
227 /// \brief Branch weights for error reporting.
228 MDNode *ColdCallWeights;
229 /// \brief Branch weights for origin store.
230 MDNode *OriginStoreWeights;
231 /// \brief Path to blacklist file.
232 SmallString<64> BlacklistFile;
233 /// \brief The blacklist.
234 OwningPtr<BlackList> BL;
235 /// \brief An empty volatile inline asm that prevents callback merge.
238 friend struct MemorySanitizerVisitor;
239 friend struct VarArgAMD64Helper;
243 char MemorySanitizer::ID = 0;
244 INITIALIZE_PASS(MemorySanitizer, "msan",
245 "MemorySanitizer: detects uninitialized reads.",
248 FunctionPass *llvm::createMemorySanitizerPass(bool TrackOrigins,
249 StringRef BlacklistFile) {
250 return new MemorySanitizer(TrackOrigins, BlacklistFile);
253 /// \brief Create a non-const global initialized with the given string.
255 /// Creates a writable global for Str so that we can pass it to the
256 /// run-time lib. Runtime uses first 4 bytes of the string to store the
257 /// frame ID, so the string needs to be mutable.
258 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
260 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
261 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
262 GlobalValue::PrivateLinkage, StrConst, "");
266 /// \brief Insert extern declaration of runtime-provided functions and globals.
267 void MemorySanitizer::initializeCallbacks(Module &M) {
268 // Only do this once.
273 // Create the callback.
274 // FIXME: this function should have "Cold" calling conv,
275 // which is not yet implemented.
276 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
277 : "__msan_warning_noreturn";
278 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
280 MsanCopyOriginFn = M.getOrInsertFunction(
281 "__msan_copy_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(),
282 IRB.getInt8PtrTy(), IntptrTy, NULL);
283 MsanSetAllocaOriginFn = M.getOrInsertFunction(
284 "__msan_set_alloca_origin", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
285 IRB.getInt8PtrTy(), NULL);
286 MsanPoisonStackFn = M.getOrInsertFunction(
287 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
288 MemmoveFn = M.getOrInsertFunction(
289 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
290 IRB.getInt8PtrTy(), IntptrTy, NULL);
291 MemcpyFn = M.getOrInsertFunction(
292 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
294 MemsetFn = M.getOrInsertFunction(
295 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
299 RetvalTLS = new GlobalVariable(
300 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
301 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
302 GlobalVariable::GeneralDynamicTLSModel);
303 RetvalOriginTLS = new GlobalVariable(
304 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
305 "__msan_retval_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
307 ParamTLS = new GlobalVariable(
308 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
309 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
310 GlobalVariable::GeneralDynamicTLSModel);
311 ParamOriginTLS = new GlobalVariable(
312 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
313 0, "__msan_param_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
315 VAArgTLS = new GlobalVariable(
316 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
317 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
318 GlobalVariable::GeneralDynamicTLSModel);
319 VAArgOverflowSizeTLS = new GlobalVariable(
320 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
321 "__msan_va_arg_overflow_size_tls", 0,
322 GlobalVariable::GeneralDynamicTLSModel);
323 OriginTLS = new GlobalVariable(
324 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
325 "__msan_origin_tls", 0, GlobalVariable::GeneralDynamicTLSModel);
327 // We insert an empty inline asm after __msan_report* to avoid callback merge.
328 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
329 StringRef(""), StringRef(""),
330 /*hasSideEffects=*/true);
333 /// \brief Module-level initialization.
335 /// inserts a call to __msan_init to the module's constructor list.
336 bool MemorySanitizer::doInitialization(Module &M) {
337 TD = getAnalysisIfAvailable<DataLayout>();
340 BL.reset(new BlackList(BlacklistFile));
341 C = &(M.getContext());
342 unsigned PtrSize = TD->getPointerSizeInBits(/* AddressSpace */0);
345 ShadowMask = kShadowMask64;
346 OriginOffset = kOriginOffset64;
349 ShadowMask = kShadowMask32;
350 OriginOffset = kOriginOffset32;
353 report_fatal_error("unsupported pointer size");
358 IntptrTy = IRB.getIntPtrTy(TD);
359 OriginTy = IRB.getInt32Ty();
361 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
362 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
364 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
365 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
366 "__msan_init", IRB.getVoidTy(), NULL)), 0);
368 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
369 IRB.getInt32(TrackOrigins), "__msan_track_origins");
371 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
372 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
379 /// \brief A helper class that handles instrumentation of VarArg
380 /// functions on a particular platform.
382 /// Implementations are expected to insert the instrumentation
383 /// necessary to propagate argument shadow through VarArg function
384 /// calls. Visit* methods are called during an InstVisitor pass over
385 /// the function, and should avoid creating new basic blocks. A new
386 /// instance of this class is created for each instrumented function.
387 struct VarArgHelper {
388 /// \brief Visit a CallSite.
389 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
391 /// \brief Visit a va_start call.
392 virtual void visitVAStartInst(VAStartInst &I) = 0;
394 /// \brief Visit a va_copy call.
395 virtual void visitVACopyInst(VACopyInst &I) = 0;
397 /// \brief Finalize function instrumentation.
399 /// This method is called after visiting all interesting (see above)
400 /// instructions in a function.
401 virtual void finalizeInstrumentation() = 0;
403 virtual ~VarArgHelper() {}
406 struct MemorySanitizerVisitor;
409 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
410 MemorySanitizerVisitor &Visitor);
412 /// This class does all the work for a given function. Store and Load
413 /// instructions store and load corresponding shadow and origin
414 /// values. Most instructions propagate shadow from arguments to their
415 /// return values. Certain instructions (most importantly, BranchInst)
416 /// test their argument shadow and print reports (with a runtime call) if it's
418 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
421 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
422 ValueMap<Value*, Value*> ShadowMap, OriginMap;
425 OwningPtr<VarArgHelper> VAHelper;
427 struct ShadowOriginAndInsertPoint {
430 Instruction *OrigIns;
431 ShadowOriginAndInsertPoint(Instruction *S, Instruction *O, Instruction *I)
432 : Shadow(S), Origin(O), OrigIns(I) { }
433 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
435 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
436 SmallVector<Instruction*, 16> StoreList;
438 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
439 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
440 LoadShadow = InsertChecks =
442 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
443 Attribute::SanitizeMemory);
445 DEBUG(if (!InsertChecks)
446 dbgs() << "MemorySanitizer is not inserting checks into '"
447 << F.getName() << "'\n");
450 void materializeStores() {
451 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
452 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]);
455 Value *Val = I.getValueOperand();
456 Value *Addr = I.getPointerOperand();
457 Value *Shadow = getShadow(Val);
458 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
461 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
462 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
465 if (ClCheckAccessAddress)
466 insertCheck(Addr, &I);
468 if (MS.TrackOrigins) {
469 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
470 if (ClStoreCleanOrigin || isa<StructType>(Shadow->getType())) {
471 IRB.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRB),
474 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
476 Constant *Cst = dyn_cast_or_null<Constant>(ConvertedShadow);
477 // TODO(eugenis): handle non-zero constant shadow by inserting an
478 // unconditional check (can not simply fail compilation as this could
479 // be in the dead code).
483 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
484 getCleanShadow(ConvertedShadow), "_mscmp");
485 Instruction *CheckTerm =
486 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp), false,
487 MS.OriginStoreWeights);
488 IRBuilder<> IRBNew(CheckTerm);
489 IRBNew.CreateAlignedStore(getOrigin(Val), getOriginPtr(Addr, IRBNew),
496 void materializeChecks() {
497 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
498 Instruction *Shadow = InstrumentationList[i].Shadow;
499 Instruction *OrigIns = InstrumentationList[i].OrigIns;
500 IRBuilder<> IRB(OrigIns);
501 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
502 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
503 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
504 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
505 getCleanShadow(ConvertedShadow), "_mscmp");
506 Instruction *CheckTerm =
507 SplitBlockAndInsertIfThen(cast<Instruction>(Cmp),
508 /* Unreachable */ !ClKeepGoing,
511 IRB.SetInsertPoint(CheckTerm);
512 if (MS.TrackOrigins) {
513 Instruction *Origin = InstrumentationList[i].Origin;
514 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
517 CallInst *Call = IRB.CreateCall(MS.WarningFn);
518 Call->setDebugLoc(OrigIns->getDebugLoc());
519 IRB.CreateCall(MS.EmptyAsm);
520 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
522 DEBUG(dbgs() << "DONE:\n" << F);
525 /// \brief Add MemorySanitizer instrumentation to a function.
526 bool runOnFunction() {
527 MS.initializeCallbacks(*F.getParent());
528 if (!MS.TD) return false;
530 // In the presence of unreachable blocks, we may see Phi nodes with
531 // incoming nodes from such blocks. Since InstVisitor skips unreachable
532 // blocks, such nodes will not have any shadow value associated with them.
533 // It's easier to remove unreachable blocks than deal with missing shadow.
534 removeUnreachableBlocks(F);
536 // Iterate all BBs in depth-first order and create shadow instructions
537 // for all instructions (where applicable).
538 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
539 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
540 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
541 BasicBlock *BB = *DI;
545 // Finalize PHI nodes.
546 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
547 PHINode *PN = ShadowPHINodes[i];
548 PHINode *PNS = cast<PHINode>(getShadow(PN));
549 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
550 size_t NumValues = PN->getNumIncomingValues();
551 for (size_t v = 0; v < NumValues; v++) {
552 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
554 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
558 VAHelper->finalizeInstrumentation();
560 // Delayed instrumentation of StoreInst.
561 // This may add new checks to be inserted later.
564 // Insert shadow value checks.
570 /// \brief Compute the shadow type that corresponds to a given Value.
571 Type *getShadowTy(Value *V) {
572 return getShadowTy(V->getType());
575 /// \brief Compute the shadow type that corresponds to a given Type.
576 Type *getShadowTy(Type *OrigTy) {
577 if (!OrigTy->isSized()) {
580 // For integer type, shadow is the same as the original type.
581 // This may return weird-sized types like i1.
582 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
584 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
585 uint32_t EltSize = MS.TD->getTypeSizeInBits(VT->getElementType());
586 return VectorType::get(IntegerType::get(*MS.C, EltSize),
587 VT->getNumElements());
589 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
590 SmallVector<Type*, 4> Elements;
591 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
592 Elements.push_back(getShadowTy(ST->getElementType(i)));
593 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
594 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
597 uint32_t TypeSize = MS.TD->getTypeSizeInBits(OrigTy);
598 return IntegerType::get(*MS.C, TypeSize);
601 /// \brief Flatten a vector type.
602 Type *getShadowTyNoVec(Type *ty) {
603 if (VectorType *vt = dyn_cast<VectorType>(ty))
604 return IntegerType::get(*MS.C, vt->getBitWidth());
608 /// \brief Convert a shadow value to it's flattened variant.
609 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
610 Type *Ty = V->getType();
611 Type *NoVecTy = getShadowTyNoVec(Ty);
612 if (Ty == NoVecTy) return V;
613 return IRB.CreateBitCast(V, NoVecTy);
616 /// \brief Compute the shadow address that corresponds to a given application
619 /// Shadow = Addr & ~ShadowMask.
620 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
623 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
624 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
625 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
628 /// \brief Compute the origin address that corresponds to a given application
631 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
632 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
634 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
635 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
637 IRB.CreateAdd(ShadowLong,
638 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
640 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
641 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
644 /// \brief Compute the shadow address for a given function argument.
646 /// Shadow = ParamTLS+ArgOffset.
647 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
649 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
650 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
651 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
655 /// \brief Compute the origin address for a given function argument.
656 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
658 if (!MS.TrackOrigins) return 0;
659 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
660 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
661 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
665 /// \brief Compute the shadow address for a retval.
666 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
667 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
668 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
672 /// \brief Compute the origin address for a retval.
673 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
674 // We keep a single origin for the entire retval. Might be too optimistic.
675 return MS.RetvalOriginTLS;
678 /// \brief Set SV to be the shadow value for V.
679 void setShadow(Value *V, Value *SV) {
680 assert(!ShadowMap.count(V) && "Values may only have one shadow");
684 /// \brief Set Origin to be the origin value for V.
685 void setOrigin(Value *V, Value *Origin) {
686 if (!MS.TrackOrigins) return;
687 assert(!OriginMap.count(V) && "Values may only have one origin");
688 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
689 OriginMap[V] = Origin;
692 /// \brief Create a clean shadow value for a given value.
694 /// Clean shadow (all zeroes) means all bits of the value are defined
696 Constant *getCleanShadow(Value *V) {
697 Type *ShadowTy = getShadowTy(V);
700 return Constant::getNullValue(ShadowTy);
703 /// \brief Create a dirty shadow of a given shadow type.
704 Constant *getPoisonedShadow(Type *ShadowTy) {
706 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
707 return Constant::getAllOnesValue(ShadowTy);
708 StructType *ST = cast<StructType>(ShadowTy);
709 SmallVector<Constant *, 4> Vals;
710 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
711 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
712 return ConstantStruct::get(ST, Vals);
715 /// \brief Create a dirty shadow for a given value.
716 Constant *getPoisonedShadow(Value *V) {
717 Type *ShadowTy = getShadowTy(V);
720 return getPoisonedShadow(ShadowTy);
723 /// \brief Create a clean (zero) origin.
724 Value *getCleanOrigin() {
725 return Constant::getNullValue(MS.OriginTy);
728 /// \brief Get the shadow value for a given Value.
730 /// This function either returns the value set earlier with setShadow,
731 /// or extracts if from ParamTLS (for function arguments).
732 Value *getShadow(Value *V) {
733 if (Instruction *I = dyn_cast<Instruction>(V)) {
734 // For instructions the shadow is already stored in the map.
735 Value *Shadow = ShadowMap[V];
737 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
739 assert(Shadow && "No shadow for a value");
743 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
744 Value *AllOnes = ClPoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
745 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
749 if (Argument *A = dyn_cast<Argument>(V)) {
750 // For arguments we compute the shadow on demand and store it in the map.
751 Value **ShadowPtr = &ShadowMap[V];
754 Function *F = A->getParent();
755 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
756 unsigned ArgOffset = 0;
757 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
759 if (!AI->getType()->isSized()) {
760 DEBUG(dbgs() << "Arg is not sized\n");
763 unsigned Size = AI->hasByValAttr()
764 ? MS.TD->getTypeAllocSize(AI->getType()->getPointerElementType())
765 : MS.TD->getTypeAllocSize(AI->getType());
767 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
768 if (AI->hasByValAttr()) {
769 // ByVal pointer itself has clean shadow. We copy the actual
770 // argument shadow to the underlying memory.
771 Value *Cpy = EntryIRB.CreateMemCpy(
772 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
773 Base, Size, AI->getParamAlignment());
774 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
776 *ShadowPtr = getCleanShadow(V);
778 *ShadowPtr = EntryIRB.CreateLoad(Base);
780 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
781 **ShadowPtr << "\n");
782 if (MS.TrackOrigins) {
783 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
784 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
787 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
789 assert(*ShadowPtr && "Could not find shadow for an argument");
792 // For everything else the shadow is zero.
793 return getCleanShadow(V);
796 /// \brief Get the shadow for i-th argument of the instruction I.
797 Value *getShadow(Instruction *I, int i) {
798 return getShadow(I->getOperand(i));
801 /// \brief Get the origin for a value.
802 Value *getOrigin(Value *V) {
803 if (!MS.TrackOrigins) return 0;
804 if (isa<Instruction>(V) || isa<Argument>(V)) {
805 Value *Origin = OriginMap[V];
807 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
808 Origin = getCleanOrigin();
812 return getCleanOrigin();
815 /// \brief Get the origin for i-th argument of the instruction I.
816 Value *getOrigin(Instruction *I, int i) {
817 return getOrigin(I->getOperand(i));
820 /// \brief Remember the place where a shadow check should be inserted.
822 /// This location will be later instrumented with a check that will print a
823 /// UMR warning in runtime if the value is not fully defined.
824 void insertCheck(Value *Val, Instruction *OrigIns) {
826 if (!InsertChecks) return;
827 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
830 Type *ShadowTy = Shadow->getType();
831 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
832 "Can only insert checks for integer and vector shadow types");
834 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
835 InstrumentationList.push_back(
836 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
839 // ------------------- Visitors.
841 /// \brief Instrument LoadInst
843 /// Loads the corresponding shadow and (optionally) origin.
844 /// Optionally, checks that the load address is fully defined.
845 void visitLoadInst(LoadInst &I) {
846 assert(I.getType()->isSized() && "Load type must have size");
848 Type *ShadowTy = getShadowTy(&I);
849 Value *Addr = I.getPointerOperand();
851 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
853 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
855 setShadow(&I, getCleanShadow(&I));
858 if (ClCheckAccessAddress)
859 insertCheck(I.getPointerOperand(), &I);
861 if (MS.TrackOrigins) {
863 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
865 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
867 setOrigin(&I, getCleanOrigin());
872 /// \brief Instrument StoreInst
874 /// Stores the corresponding shadow and (optionally) origin.
875 /// Optionally, checks that the store address is fully defined.
876 void visitStoreInst(StoreInst &I) {
877 StoreList.push_back(&I);
880 // Vector manipulation.
881 void visitExtractElementInst(ExtractElementInst &I) {
882 insertCheck(I.getOperand(1), &I);
884 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
886 setOrigin(&I, getOrigin(&I, 0));
889 void visitInsertElementInst(InsertElementInst &I) {
890 insertCheck(I.getOperand(2), &I);
892 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
893 I.getOperand(2), "_msprop"));
894 setOriginForNaryOp(I);
897 void visitShuffleVectorInst(ShuffleVectorInst &I) {
898 insertCheck(I.getOperand(2), &I);
900 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
901 I.getOperand(2), "_msprop"));
902 setOriginForNaryOp(I);
906 void visitSExtInst(SExtInst &I) {
908 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
909 setOrigin(&I, getOrigin(&I, 0));
912 void visitZExtInst(ZExtInst &I) {
914 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
915 setOrigin(&I, getOrigin(&I, 0));
918 void visitTruncInst(TruncInst &I) {
920 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
921 setOrigin(&I, getOrigin(&I, 0));
924 void visitBitCastInst(BitCastInst &I) {
926 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
927 setOrigin(&I, getOrigin(&I, 0));
930 void visitPtrToIntInst(PtrToIntInst &I) {
932 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
933 "_msprop_ptrtoint"));
934 setOrigin(&I, getOrigin(&I, 0));
937 void visitIntToPtrInst(IntToPtrInst &I) {
939 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
940 "_msprop_inttoptr"));
941 setOrigin(&I, getOrigin(&I, 0));
944 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
945 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
946 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
947 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
948 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
949 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
951 /// \brief Propagate shadow for bitwise AND.
953 /// This code is exact, i.e. if, for example, a bit in the left argument
954 /// is defined and 0, then neither the value not definedness of the
955 /// corresponding bit in B don't affect the resulting shadow.
956 void visitAnd(BinaryOperator &I) {
958 // "And" of 0 and a poisoned value results in unpoisoned value.
959 // 1&1 => 1; 0&1 => 0; p&1 => p;
960 // 1&0 => 0; 0&0 => 0; p&0 => 0;
961 // 1&p => p; 0&p => 0; p&p => p;
962 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
963 Value *S1 = getShadow(&I, 0);
964 Value *S2 = getShadow(&I, 1);
965 Value *V1 = I.getOperand(0);
966 Value *V2 = I.getOperand(1);
967 if (V1->getType() != S1->getType()) {
968 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
969 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
971 Value *S1S2 = IRB.CreateAnd(S1, S2);
972 Value *V1S2 = IRB.CreateAnd(V1, S2);
973 Value *S1V2 = IRB.CreateAnd(S1, V2);
974 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
975 setOriginForNaryOp(I);
978 void visitOr(BinaryOperator &I) {
980 // "Or" of 1 and a poisoned value results in unpoisoned value.
981 // 1|1 => 1; 0|1 => 1; p|1 => 1;
982 // 1|0 => 1; 0|0 => 0; p|0 => p;
983 // 1|p => 1; 0|p => p; p|p => p;
984 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
985 Value *S1 = getShadow(&I, 0);
986 Value *S2 = getShadow(&I, 1);
987 Value *V1 = IRB.CreateNot(I.getOperand(0));
988 Value *V2 = IRB.CreateNot(I.getOperand(1));
989 if (V1->getType() != S1->getType()) {
990 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
991 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
993 Value *S1S2 = IRB.CreateAnd(S1, S2);
994 Value *V1S2 = IRB.CreateAnd(V1, S2);
995 Value *S1V2 = IRB.CreateAnd(S1, V2);
996 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
997 setOriginForNaryOp(I);
1000 /// \brief Default propagation of shadow and/or origin.
1002 /// This class implements the general case of shadow propagation, used in all
1003 /// cases where we don't know and/or don't care about what the operation
1004 /// actually does. It converts all input shadow values to a common type
1005 /// (extending or truncating as necessary), and bitwise OR's them.
1007 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1008 /// fully initialized), and less prone to false positives.
1010 /// This class also implements the general case of origin propagation. For a
1011 /// Nary operation, result origin is set to the origin of an argument that is
1012 /// not entirely initialized. If there is more than one such arguments, the
1013 /// rightmost of them is picked. It does not matter which one is picked if all
1014 /// arguments are initialized.
1015 template <bool CombineShadow>
1020 MemorySanitizerVisitor *MSV;
1023 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1024 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {}
1026 /// \brief Add a pair of shadow and origin values to the mix.
1027 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1028 if (CombineShadow) {
1033 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1034 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1038 if (MSV->MS.TrackOrigins) {
1043 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1044 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1045 MSV->getCleanShadow(FlatShadow));
1046 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1052 /// \brief Add an application value to the mix.
1053 Combiner &Add(Value *V) {
1054 Value *OpShadow = MSV->getShadow(V);
1055 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : 0;
1056 return Add(OpShadow, OpOrigin);
1059 /// \brief Set the current combined values as the given instruction's shadow
1061 void Done(Instruction *I) {
1062 if (CombineShadow) {
1064 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1065 MSV->setShadow(I, Shadow);
1067 if (MSV->MS.TrackOrigins) {
1069 MSV->setOrigin(I, Origin);
1074 typedef Combiner<true> ShadowAndOriginCombiner;
1075 typedef Combiner<false> OriginCombiner;
1077 /// \brief Propagate origin for arbitrary operation.
1078 void setOriginForNaryOp(Instruction &I) {
1079 if (!MS.TrackOrigins) return;
1080 IRBuilder<> IRB(&I);
1081 OriginCombiner OC(this, IRB);
1082 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1087 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1088 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1089 "Vector of pointers is not a valid shadow type");
1090 return Ty->isVectorTy() ?
1091 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1092 Ty->getPrimitiveSizeInBits();
1095 /// \brief Cast between two shadow types, extending or truncating as
1097 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy) {
1098 Type *srcTy = V->getType();
1099 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1100 return IRB.CreateIntCast(V, dstTy, false);
1101 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1102 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1103 return IRB.CreateIntCast(V, dstTy, false);
1104 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1105 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1106 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1108 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), false);
1109 return IRB.CreateBitCast(V2, dstTy);
1110 // TODO: handle struct types.
1113 /// \brief Propagate shadow for arbitrary operation.
1114 void handleShadowOr(Instruction &I) {
1115 IRBuilder<> IRB(&I);
1116 ShadowAndOriginCombiner SC(this, IRB);
1117 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1122 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1123 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1124 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1125 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1126 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1127 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1128 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1130 void handleDiv(Instruction &I) {
1131 IRBuilder<> IRB(&I);
1132 // Strict on the second argument.
1133 insertCheck(I.getOperand(1), &I);
1134 setShadow(&I, getShadow(&I, 0));
1135 setOrigin(&I, getOrigin(&I, 0));
1138 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1139 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1140 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1141 void visitURem(BinaryOperator &I) { handleDiv(I); }
1142 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1143 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1145 /// \brief Instrument == and != comparisons.
1147 /// Sometimes the comparison result is known even if some of the bits of the
1148 /// arguments are not.
1149 void handleEqualityComparison(ICmpInst &I) {
1150 IRBuilder<> IRB(&I);
1151 Value *A = I.getOperand(0);
1152 Value *B = I.getOperand(1);
1153 Value *Sa = getShadow(A);
1154 Value *Sb = getShadow(B);
1156 // Get rid of pointers and vectors of pointers.
1157 // For ints (and vectors of ints), types of A and Sa match,
1158 // and this is a no-op.
1159 A = IRB.CreatePointerCast(A, Sa->getType());
1160 B = IRB.CreatePointerCast(B, Sb->getType());
1162 // A == B <==> (C = A^B) == 0
1163 // A != B <==> (C = A^B) != 0
1165 Value *C = IRB.CreateXor(A, B);
1166 Value *Sc = IRB.CreateOr(Sa, Sb);
1167 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1168 // Result is defined if one of the following is true
1169 // * there is a defined 1 bit in C
1170 // * C is fully defined
1171 // Si = !(C & ~Sc) && Sc
1172 Value *Zero = Constant::getNullValue(Sc->getType());
1173 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1175 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1177 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1178 Si->setName("_msprop_icmp");
1180 setOriginForNaryOp(I);
1183 /// \brief Build the lowest possible value of V, taking into account V's
1184 /// uninitialized bits.
1185 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1188 // Split shadow into sign bit and other bits.
1189 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1190 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1191 // Maximise the undefined shadow bit, minimize other undefined bits.
1193 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1195 // Minimize undefined bits.
1196 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1200 /// \brief Build the highest possible value of V, taking into account V's
1201 /// uninitialized bits.
1202 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1205 // Split shadow into sign bit and other bits.
1206 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1207 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1208 // Minimise the undefined shadow bit, maximise other undefined bits.
1210 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1212 // Maximize undefined bits.
1213 return IRB.CreateOr(A, Sa);
1217 /// \brief Instrument relational comparisons.
1219 /// This function does exact shadow propagation for all relational
1220 /// comparisons of integers, pointers and vectors of those.
1221 /// FIXME: output seems suboptimal when one of the operands is a constant
1222 void handleRelationalComparisonExact(ICmpInst &I) {
1223 IRBuilder<> IRB(&I);
1224 Value *A = I.getOperand(0);
1225 Value *B = I.getOperand(1);
1226 Value *Sa = getShadow(A);
1227 Value *Sb = getShadow(B);
1229 // Get rid of pointers and vectors of pointers.
1230 // For ints (and vectors of ints), types of A and Sa match,
1231 // and this is a no-op.
1232 A = IRB.CreatePointerCast(A, Sa->getType());
1233 B = IRB.CreatePointerCast(B, Sb->getType());
1235 // Let [a0, a1] be the interval of possible values of A, taking into account
1236 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1237 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1238 bool IsSigned = I.isSigned();
1239 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1240 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1241 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1242 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1243 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1244 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1245 Value *Si = IRB.CreateXor(S1, S2);
1247 setOriginForNaryOp(I);
1250 /// \brief Instrument signed relational comparisons.
1252 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1253 /// propagating the highest bit of the shadow. Everything else is delegated
1254 /// to handleShadowOr().
1255 void handleSignedRelationalComparison(ICmpInst &I) {
1256 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1257 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1259 CmpInst::Predicate pre = I.getPredicate();
1260 if (constOp0 && constOp0->isNullValue() &&
1261 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1262 op = I.getOperand(1);
1263 } else if (constOp1 && constOp1->isNullValue() &&
1264 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1265 op = I.getOperand(0);
1268 IRBuilder<> IRB(&I);
1270 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1271 setShadow(&I, Shadow);
1272 setOrigin(&I, getOrigin(op));
1278 void visitICmpInst(ICmpInst &I) {
1279 if (!ClHandleICmp) {
1283 if (I.isEquality()) {
1284 handleEqualityComparison(I);
1288 assert(I.isRelational());
1289 if (ClHandleICmpExact) {
1290 handleRelationalComparisonExact(I);
1294 handleSignedRelationalComparison(I);
1298 assert(I.isUnsigned());
1299 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1300 handleRelationalComparisonExact(I);
1307 void visitFCmpInst(FCmpInst &I) {
1311 void handleShift(BinaryOperator &I) {
1312 IRBuilder<> IRB(&I);
1313 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1314 // Otherwise perform the same shift on S1.
1315 Value *S1 = getShadow(&I, 0);
1316 Value *S2 = getShadow(&I, 1);
1317 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1319 Value *V2 = I.getOperand(1);
1320 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1321 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1322 setOriginForNaryOp(I);
1325 void visitShl(BinaryOperator &I) { handleShift(I); }
1326 void visitAShr(BinaryOperator &I) { handleShift(I); }
1327 void visitLShr(BinaryOperator &I) { handleShift(I); }
1329 /// \brief Instrument llvm.memmove
1331 /// At this point we don't know if llvm.memmove will be inlined or not.
1332 /// If we don't instrument it and it gets inlined,
1333 /// our interceptor will not kick in and we will lose the memmove.
1334 /// If we instrument the call here, but it does not get inlined,
1335 /// we will memove the shadow twice: which is bad in case
1336 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1338 /// Similar situation exists for memcpy and memset.
1339 void visitMemMoveInst(MemMoveInst &I) {
1340 IRBuilder<> IRB(&I);
1343 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1344 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1345 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1346 I.eraseFromParent();
1349 // Similar to memmove: avoid copying shadow twice.
1350 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1351 // FIXME: consider doing manual inline for small constant sizes and proper
1353 void visitMemCpyInst(MemCpyInst &I) {
1354 IRBuilder<> IRB(&I);
1357 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1358 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1359 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1360 I.eraseFromParent();
1364 void visitMemSetInst(MemSetInst &I) {
1365 IRBuilder<> IRB(&I);
1368 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1369 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1370 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1371 I.eraseFromParent();
1374 void visitVAStartInst(VAStartInst &I) {
1375 VAHelper->visitVAStartInst(I);
1378 void visitVACopyInst(VACopyInst &I) {
1379 VAHelper->visitVACopyInst(I);
1382 enum IntrinsicKind {
1383 IK_DoesNotAccessMemory,
1388 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1389 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1390 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1391 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1392 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1393 const int UnknownModRefBehavior = IK_WritesMemory;
1394 #define GET_INTRINSIC_MODREF_BEHAVIOR
1395 #define ModRefBehavior IntrinsicKind
1396 #include "llvm/IR/Intrinsics.gen"
1397 #undef ModRefBehavior
1398 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1401 /// \brief Handle vector store-like intrinsics.
1403 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1404 /// has 1 pointer argument and 1 vector argument, returns void.
1405 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1406 IRBuilder<> IRB(&I);
1407 Value* Addr = I.getArgOperand(0);
1408 Value *Shadow = getShadow(&I, 1);
1409 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1411 // We don't know the pointer alignment (could be unaligned SSE store!).
1412 // Have to assume to worst case.
1413 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1415 if (ClCheckAccessAddress)
1416 insertCheck(Addr, &I);
1418 // FIXME: use ClStoreCleanOrigin
1419 // FIXME: factor out common code from materializeStores
1420 if (MS.TrackOrigins)
1421 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1425 /// \brief Handle vector load-like intrinsics.
1427 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1428 /// has 1 pointer argument, returns a vector.
1429 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1430 IRBuilder<> IRB(&I);
1431 Value *Addr = I.getArgOperand(0);
1433 Type *ShadowTy = getShadowTy(&I);
1435 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1436 // We don't know the pointer alignment (could be unaligned SSE load!).
1437 // Have to assume to worst case.
1438 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1440 setShadow(&I, getCleanShadow(&I));
1444 if (ClCheckAccessAddress)
1445 insertCheck(Addr, &I);
1447 if (MS.TrackOrigins) {
1449 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1451 setOrigin(&I, getCleanOrigin());
1456 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1458 /// Instrument intrinsics with any number of arguments of the same type,
1459 /// equal to the return type. The type should be simple (no aggregates or
1460 /// pointers; vectors are fine).
1461 /// Caller guarantees that this intrinsic does not access memory.
1462 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1463 Type *RetTy = I.getType();
1464 if (!(RetTy->isIntOrIntVectorTy() ||
1465 RetTy->isFPOrFPVectorTy() ||
1466 RetTy->isX86_MMXTy()))
1469 unsigned NumArgOperands = I.getNumArgOperands();
1471 for (unsigned i = 0; i < NumArgOperands; ++i) {
1472 Type *Ty = I.getArgOperand(i)->getType();
1477 IRBuilder<> IRB(&I);
1478 ShadowAndOriginCombiner SC(this, IRB);
1479 for (unsigned i = 0; i < NumArgOperands; ++i)
1480 SC.Add(I.getArgOperand(i));
1486 /// \brief Heuristically instrument unknown intrinsics.
1488 /// The main purpose of this code is to do something reasonable with all
1489 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1490 /// We recognize several classes of intrinsics by their argument types and
1491 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1492 /// sure that we know what the intrinsic does.
1494 /// We special-case intrinsics where this approach fails. See llvm.bswap
1495 /// handling as an example of that.
1496 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1497 unsigned NumArgOperands = I.getNumArgOperands();
1498 if (NumArgOperands == 0)
1501 Intrinsic::ID iid = I.getIntrinsicID();
1502 IntrinsicKind IK = getIntrinsicKind(iid);
1503 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1504 bool WritesMemory = IK == IK_WritesMemory;
1505 assert(!(OnlyReadsMemory && WritesMemory));
1507 if (NumArgOperands == 2 &&
1508 I.getArgOperand(0)->getType()->isPointerTy() &&
1509 I.getArgOperand(1)->getType()->isVectorTy() &&
1510 I.getType()->isVoidTy() &&
1512 // This looks like a vector store.
1513 return handleVectorStoreIntrinsic(I);
1516 if (NumArgOperands == 1 &&
1517 I.getArgOperand(0)->getType()->isPointerTy() &&
1518 I.getType()->isVectorTy() &&
1520 // This looks like a vector load.
1521 return handleVectorLoadIntrinsic(I);
1524 if (!OnlyReadsMemory && !WritesMemory)
1525 if (maybeHandleSimpleNomemIntrinsic(I))
1528 // FIXME: detect and handle SSE maskstore/maskload
1532 void handleBswap(IntrinsicInst &I) {
1533 IRBuilder<> IRB(&I);
1534 Value *Op = I.getArgOperand(0);
1535 Type *OpType = Op->getType();
1536 Function *BswapFunc = Intrinsic::getDeclaration(
1537 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1538 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1539 setOrigin(&I, getOrigin(Op));
1542 void visitIntrinsicInst(IntrinsicInst &I) {
1543 switch (I.getIntrinsicID()) {
1544 case llvm::Intrinsic::bswap:
1548 if (!handleUnknownIntrinsic(I))
1549 visitInstruction(I);
1554 void visitCallSite(CallSite CS) {
1555 Instruction &I = *CS.getInstruction();
1556 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
1558 CallInst *Call = cast<CallInst>(&I);
1560 // For inline asm, do the usual thing: check argument shadow and mark all
1561 // outputs as clean. Note that any side effects of the inline asm that are
1562 // not immediately visible in its constraints are not handled.
1563 if (Call->isInlineAsm()) {
1564 visitInstruction(I);
1568 // Allow only tail calls with the same types, otherwise
1569 // we may have a false positive: shadow for a non-void RetVal
1570 // will get propagated to a void RetVal.
1571 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
1572 Call->setTailCall(false);
1574 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
1576 // We are going to insert code that relies on the fact that the callee
1577 // will become a non-readonly function after it is instrumented by us. To
1578 // prevent this code from being optimized out, mark that function
1579 // non-readonly in advance.
1580 if (Function *Func = Call->getCalledFunction()) {
1581 // Clear out readonly/readnone attributes.
1583 B.addAttribute(Attribute::ReadOnly)
1584 .addAttribute(Attribute::ReadNone);
1585 Func->removeAttributes(AttributeSet::FunctionIndex,
1586 AttributeSet::get(Func->getContext(),
1587 AttributeSet::FunctionIndex,
1591 IRBuilder<> IRB(&I);
1592 unsigned ArgOffset = 0;
1593 DEBUG(dbgs() << " CallSite: " << I << "\n");
1594 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1595 ArgIt != End; ++ArgIt) {
1597 unsigned i = ArgIt - CS.arg_begin();
1598 if (!A->getType()->isSized()) {
1599 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
1604 // Compute the Shadow for arg even if it is ByVal, because
1605 // in that case getShadow() will copy the actual arg shadow to
1606 // __msan_param_tls.
1607 Value *ArgShadow = getShadow(A);
1608 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
1609 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
1610 " Shadow: " << *ArgShadow << "\n");
1611 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
1612 assert(A->getType()->isPointerTy() &&
1613 "ByVal argument is not a pointer!");
1614 Size = MS.TD->getTypeAllocSize(A->getType()->getPointerElementType());
1615 unsigned Alignment = CS.getParamAlignment(i + 1);
1616 Store = IRB.CreateMemCpy(ArgShadowBase,
1617 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
1620 Size = MS.TD->getTypeAllocSize(A->getType());
1621 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
1622 kShadowTLSAlignment);
1624 if (MS.TrackOrigins)
1625 IRB.CreateStore(getOrigin(A),
1626 getOriginPtrForArgument(A, IRB, ArgOffset));
1628 assert(Size != 0 && Store != 0);
1629 DEBUG(dbgs() << " Param:" << *Store << "\n");
1630 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
1632 DEBUG(dbgs() << " done with call args\n");
1635 cast<FunctionType>(CS.getCalledValue()->getType()-> getContainedType(0));
1636 if (FT->isVarArg()) {
1637 VAHelper->visitCallSite(CS, IRB);
1640 // Now, get the shadow for the RetVal.
1641 if (!I.getType()->isSized()) return;
1642 IRBuilder<> IRBBefore(&I);
1643 // Untill we have full dynamic coverage, make sure the retval shadow is 0.
1644 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
1645 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
1646 Instruction *NextInsn = 0;
1648 NextInsn = I.getNextNode();
1650 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
1651 if (!NormalDest->getSinglePredecessor()) {
1652 // FIXME: this case is tricky, so we are just conservative here.
1653 // Perhaps we need to split the edge between this BB and NormalDest,
1654 // but a naive attempt to use SplitEdge leads to a crash.
1655 setShadow(&I, getCleanShadow(&I));
1656 setOrigin(&I, getCleanOrigin());
1659 NextInsn = NormalDest->getFirstInsertionPt();
1661 "Could not find insertion point for retval shadow load");
1663 IRBuilder<> IRBAfter(NextInsn);
1664 Value *RetvalShadow =
1665 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
1666 kShadowTLSAlignment, "_msret");
1667 setShadow(&I, RetvalShadow);
1668 if (MS.TrackOrigins)
1669 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
1672 void visitReturnInst(ReturnInst &I) {
1673 IRBuilder<> IRB(&I);
1674 if (Value *RetVal = I.getReturnValue()) {
1675 // Set the shadow for the RetVal.
1676 Value *Shadow = getShadow(RetVal);
1677 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
1678 DEBUG(dbgs() << "Return: " << *Shadow << "\n" << *ShadowPtr << "\n");
1679 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
1680 if (MS.TrackOrigins)
1681 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
1685 void visitPHINode(PHINode &I) {
1686 IRBuilder<> IRB(&I);
1687 ShadowPHINodes.push_back(&I);
1688 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
1690 if (MS.TrackOrigins)
1691 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
1695 void visitAllocaInst(AllocaInst &I) {
1696 setShadow(&I, getCleanShadow(&I));
1697 if (!ClPoisonStack) return;
1698 IRBuilder<> IRB(I.getNextNode());
1699 uint64_t Size = MS.TD->getTypeAllocSize(I.getAllocatedType());
1700 if (ClPoisonStackWithCall) {
1701 IRB.CreateCall2(MS.MsanPoisonStackFn,
1702 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1703 ConstantInt::get(MS.IntptrTy, Size));
1705 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
1706 IRB.CreateMemSet(ShadowBase, IRB.getInt8(ClPoisonStackPattern),
1707 Size, I.getAlignment());
1710 if (MS.TrackOrigins) {
1711 setOrigin(&I, getCleanOrigin());
1712 SmallString<2048> StackDescriptionStorage;
1713 raw_svector_ostream StackDescription(StackDescriptionStorage);
1714 // We create a string with a description of the stack allocation and
1715 // pass it into __msan_set_alloca_origin.
1716 // It will be printed by the run-time if stack-originated UMR is found.
1717 // The first 4 bytes of the string are set to '----' and will be replaced
1718 // by __msan_va_arg_overflow_size_tls at the first call.
1719 StackDescription << "----" << I.getName() << "@" << F.getName();
1721 createPrivateNonConstGlobalForString(*F.getParent(),
1722 StackDescription.str());
1723 IRB.CreateCall3(MS.MsanSetAllocaOriginFn,
1724 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
1725 ConstantInt::get(MS.IntptrTy, Size),
1726 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()));
1730 void visitSelectInst(SelectInst& I) {
1731 IRBuilder<> IRB(&I);
1732 setShadow(&I, IRB.CreateSelect(I.getCondition(),
1733 getShadow(I.getTrueValue()), getShadow(I.getFalseValue()),
1735 if (MS.TrackOrigins) {
1736 // Origins are always i32, so any vector conditions must be flattened.
1737 // FIXME: consider tracking vector origins for app vectors?
1738 Value *Cond = I.getCondition();
1739 if (Cond->getType()->isVectorTy()) {
1740 Value *ConvertedShadow = convertToShadowTyNoVec(Cond, IRB);
1741 Cond = IRB.CreateICmpNE(ConvertedShadow,
1742 getCleanShadow(ConvertedShadow), "_mso_select");
1744 setOrigin(&I, IRB.CreateSelect(Cond,
1745 getOrigin(I.getTrueValue()), getOrigin(I.getFalseValue())));
1749 void visitLandingPadInst(LandingPadInst &I) {
1751 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
1752 setShadow(&I, getCleanShadow(&I));
1753 setOrigin(&I, getCleanOrigin());
1756 void visitGetElementPtrInst(GetElementPtrInst &I) {
1760 void visitExtractValueInst(ExtractValueInst &I) {
1761 IRBuilder<> IRB(&I);
1762 Value *Agg = I.getAggregateOperand();
1763 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
1764 Value *AggShadow = getShadow(Agg);
1765 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1766 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
1767 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
1768 setShadow(&I, ResShadow);
1769 setOrigin(&I, getCleanOrigin());
1772 void visitInsertValueInst(InsertValueInst &I) {
1773 IRBuilder<> IRB(&I);
1774 DEBUG(dbgs() << "InsertValue: " << I << "\n");
1775 Value *AggShadow = getShadow(I.getAggregateOperand());
1776 Value *InsShadow = getShadow(I.getInsertedValueOperand());
1777 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
1778 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
1779 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
1780 DEBUG(dbgs() << " Res: " << *Res << "\n");
1782 setOrigin(&I, getCleanOrigin());
1785 void dumpInst(Instruction &I) {
1786 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1787 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
1789 errs() << "ZZZ " << I.getOpcodeName() << "\n";
1791 errs() << "QQQ " << I << "\n";
1794 void visitResumeInst(ResumeInst &I) {
1795 DEBUG(dbgs() << "Resume: " << I << "\n");
1796 // Nothing to do here.
1799 void visitInstruction(Instruction &I) {
1800 // Everything else: stop propagating and check for poisoned shadow.
1801 if (ClDumpStrictInstructions)
1803 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
1804 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
1805 insertCheck(I.getOperand(i), &I);
1806 setShadow(&I, getCleanShadow(&I));
1807 setOrigin(&I, getCleanOrigin());
1811 /// \brief AMD64-specific implementation of VarArgHelper.
1812 struct VarArgAMD64Helper : public VarArgHelper {
1813 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
1814 // See a comment in visitCallSite for more details.
1815 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
1816 static const unsigned AMD64FpEndOffset = 176;
1819 MemorySanitizer &MS;
1820 MemorySanitizerVisitor &MSV;
1821 Value *VAArgTLSCopy;
1822 Value *VAArgOverflowSize;
1824 SmallVector<CallInst*, 16> VAStartInstrumentationList;
1826 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
1827 MemorySanitizerVisitor &MSV)
1828 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
1830 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
1832 ArgKind classifyArgument(Value* arg) {
1833 // A very rough approximation of X86_64 argument classification rules.
1834 Type *T = arg->getType();
1835 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
1836 return AK_FloatingPoint;
1837 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
1838 return AK_GeneralPurpose;
1839 if (T->isPointerTy())
1840 return AK_GeneralPurpose;
1844 // For VarArg functions, store the argument shadow in an ABI-specific format
1845 // that corresponds to va_list layout.
1846 // We do this because Clang lowers va_arg in the frontend, and this pass
1847 // only sees the low level code that deals with va_list internals.
1848 // A much easier alternative (provided that Clang emits va_arg instructions)
1849 // would have been to associate each live instance of va_list with a copy of
1850 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
1852 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) {
1853 unsigned GpOffset = 0;
1854 unsigned FpOffset = AMD64GpEndOffset;
1855 unsigned OverflowOffset = AMD64FpEndOffset;
1856 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
1857 ArgIt != End; ++ArgIt) {
1859 ArgKind AK = classifyArgument(A);
1860 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
1862 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
1866 case AK_GeneralPurpose:
1867 Base = getShadowPtrForVAArgument(A, IRB, GpOffset);
1870 case AK_FloatingPoint:
1871 Base = getShadowPtrForVAArgument(A, IRB, FpOffset);
1875 uint64_t ArgSize = MS.TD->getTypeAllocSize(A->getType());
1876 Base = getShadowPtrForVAArgument(A, IRB, OverflowOffset);
1877 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
1879 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
1881 Constant *OverflowSize =
1882 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
1883 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
1886 /// \brief Compute the shadow address for a given va_arg.
1887 Value *getShadowPtrForVAArgument(Value *A, IRBuilder<> &IRB,
1889 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
1890 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
1891 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(A), 0),
1895 void visitVAStartInst(VAStartInst &I) {
1896 IRBuilder<> IRB(&I);
1897 VAStartInstrumentationList.push_back(&I);
1898 Value *VAListTag = I.getArgOperand(0);
1899 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1901 // Unpoison the whole __va_list_tag.
1902 // FIXME: magic ABI constants.
1903 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1904 /* size */24, /* alignment */8, false);
1907 void visitVACopyInst(VACopyInst &I) {
1908 IRBuilder<> IRB(&I);
1909 Value *VAListTag = I.getArgOperand(0);
1910 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
1912 // Unpoison the whole __va_list_tag.
1913 // FIXME: magic ABI constants.
1914 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
1915 /* size */24, /* alignment */8, false);
1918 void finalizeInstrumentation() {
1919 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
1920 "finalizeInstrumentation called twice");
1921 if (!VAStartInstrumentationList.empty()) {
1922 // If there is a va_start in this function, make a backup copy of
1923 // va_arg_tls somewhere in the function entry block.
1924 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
1925 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
1927 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
1929 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
1930 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
1933 // Instrument va_start.
1934 // Copy va_list shadow from the backup copy of the TLS contents.
1935 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
1936 CallInst *OrigInst = VAStartInstrumentationList[i];
1937 IRBuilder<> IRB(OrigInst->getNextNode());
1938 Value *VAListTag = OrigInst->getArgOperand(0);
1940 Value *RegSaveAreaPtrPtr =
1942 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1943 ConstantInt::get(MS.IntptrTy, 16)),
1944 Type::getInt64PtrTy(*MS.C));
1945 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
1946 Value *RegSaveAreaShadowPtr =
1947 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
1948 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
1949 AMD64FpEndOffset, 16);
1951 Value *OverflowArgAreaPtrPtr =
1953 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
1954 ConstantInt::get(MS.IntptrTy, 8)),
1955 Type::getInt64PtrTy(*MS.C));
1956 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
1957 Value *OverflowArgAreaShadowPtr =
1958 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
1960 getShadowPtrForVAArgument(VAArgTLSCopy, IRB, AMD64FpEndOffset);
1961 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
1966 VarArgHelper* CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
1967 MemorySanitizerVisitor &Visitor) {
1968 return new VarArgAMD64Helper(Func, Msan, Visitor);
1973 bool MemorySanitizer::runOnFunction(Function &F) {
1974 MemorySanitizerVisitor Visitor(F, *this);
1976 // Clear out readonly/readnone attributes.
1978 B.addAttribute(Attribute::ReadOnly)
1979 .addAttribute(Attribute::ReadNone);
1980 F.removeAttributes(AttributeSet::FunctionIndex,
1981 AttributeSet::get(F.getContext(),
1982 AttributeSet::FunctionIndex, B));
1984 return Visitor.runOnFunction();