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.
72 /// Ideally, every atomic store of application value should update the
73 /// corresponding shadow location in an atomic way. Unfortunately, atomic store
74 /// of two disjoint locations can not be done without severe slowdown.
76 /// Therefore, we implement an approximation that may err on the safe side.
77 /// In this implementation, every atomically accessed location in the program
78 /// may only change from (partially) uninitialized to fully initialized, but
79 /// not the other way around. We load the shadow _after_ the application load,
80 /// and we store the shadow _before_ the app store. Also, we always store clean
81 /// shadow (if the application store is atomic). This way, if the store-load
82 /// pair constitutes a happens-before arc, shadow store and load are correctly
83 /// ordered such that the load will get either the value that was stored, or
84 /// some later value (which is always clean).
86 /// This does not work very well with Compare-And-Swap (CAS) and
87 /// Read-Modify-Write (RMW) operations. To follow the above logic, CAS and RMW
88 /// must store the new shadow before the app operation, and load the shadow
89 /// after the app operation. Computers don't work this way. Current
90 /// implementation ignores the load aspect of CAS/RMW, always returning a clean
91 /// value. It implements the store part as a simple atomic store by storing a
94 //===----------------------------------------------------------------------===//
96 #define DEBUG_TYPE "msan"
98 #include "llvm/Transforms/Instrumentation.h"
99 #include "llvm/ADT/DepthFirstIterator.h"
100 #include "llvm/ADT/SmallString.h"
101 #include "llvm/ADT/SmallVector.h"
102 #include "llvm/ADT/Triple.h"
103 #include "llvm/IR/DataLayout.h"
104 #include "llvm/IR/Function.h"
105 #include "llvm/IR/IRBuilder.h"
106 #include "llvm/IR/InlineAsm.h"
107 #include "llvm/IR/InstVisitor.h"
108 #include "llvm/IR/IntrinsicInst.h"
109 #include "llvm/IR/LLVMContext.h"
110 #include "llvm/IR/MDBuilder.h"
111 #include "llvm/IR/Module.h"
112 #include "llvm/IR/Type.h"
113 #include "llvm/IR/ValueMap.h"
114 #include "llvm/Support/CommandLine.h"
115 #include "llvm/Support/Compiler.h"
116 #include "llvm/Support/Debug.h"
117 #include "llvm/Support/raw_ostream.h"
118 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
119 #include "llvm/Transforms/Utils/Local.h"
120 #include "llvm/Transforms/Utils/ModuleUtils.h"
121 #include "llvm/Transforms/Utils/SpecialCaseList.h"
123 using namespace llvm;
125 static const uint64_t kShadowMask32 = 1ULL << 31;
126 static const uint64_t kShadowMask64 = 1ULL << 46;
127 static const uint64_t kOriginOffset32 = 1ULL << 30;
128 static const uint64_t kOriginOffset64 = 1ULL << 45;
129 static const unsigned kMinOriginAlignment = 4;
130 static const unsigned kShadowTLSAlignment = 8;
132 /// \brief Track origins of uninitialized values.
134 /// Adds a section to MemorySanitizer report that points to the allocation
135 /// (stack or heap) the uninitialized bits came from originally.
136 static cl::opt<int> ClTrackOrigins("msan-track-origins",
137 cl::desc("Track origins (allocation sites) of poisoned memory"),
138 cl::Hidden, cl::init(0));
139 static cl::opt<bool> ClKeepGoing("msan-keep-going",
140 cl::desc("keep going after reporting a UMR"),
141 cl::Hidden, cl::init(false));
142 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
143 cl::desc("poison uninitialized stack variables"),
144 cl::Hidden, cl::init(true));
145 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
146 cl::desc("poison uninitialized stack variables with a call"),
147 cl::Hidden, cl::init(false));
148 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
149 cl::desc("poison uninitialized stack variables with the given patter"),
150 cl::Hidden, cl::init(0xff));
151 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
152 cl::desc("poison undef temps"),
153 cl::Hidden, cl::init(true));
155 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
156 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
157 cl::Hidden, cl::init(true));
159 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
160 cl::desc("exact handling of relational integer ICmp"),
161 cl::Hidden, cl::init(false));
163 // This flag controls whether we check the shadow of the address
164 // operand of load or store. Such bugs are very rare, since load from
165 // a garbage address typically results in SEGV, but still happen
166 // (e.g. only lower bits of address are garbage, or the access happens
167 // early at program startup where malloc-ed memory is more likely to
168 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
169 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
170 cl::desc("report accesses through a pointer which has poisoned shadow"),
171 cl::Hidden, cl::init(true));
173 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
174 cl::desc("print out instructions with default strict semantics"),
175 cl::Hidden, cl::init(false));
177 static cl::opt<std::string> ClBlacklistFile("msan-blacklist",
178 cl::desc("File containing the list of functions where MemorySanitizer "
179 "should not report bugs"), cl::Hidden);
181 // Experimental. Wraps all indirect calls in the instrumented code with
182 // a call to the given function. This is needed to assist the dynamic
183 // helper tool (MSanDR) to regain control on transition between instrumented and
184 // non-instrumented code.
185 static cl::opt<std::string> ClWrapIndirectCalls("msan-wrap-indirect-calls",
186 cl::desc("Wrap indirect calls with a given function"),
189 static cl::opt<bool> ClWrapIndirectCallsFast("msan-wrap-indirect-calls-fast",
190 cl::desc("Do not wrap indirect calls with target in the same module"),
191 cl::Hidden, cl::init(true));
195 /// \brief An instrumentation pass implementing detection of uninitialized
198 /// MemorySanitizer: instrument the code in module to find
199 /// uninitialized reads.
200 class MemorySanitizer : public FunctionPass {
202 MemorySanitizer(int TrackOrigins = 0,
203 StringRef BlacklistFile = StringRef())
205 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
208 BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile : BlacklistFile),
209 WrapIndirectCalls(!ClWrapIndirectCalls.empty()) {}
210 const char *getPassName() const override { return "MemorySanitizer"; }
211 bool runOnFunction(Function &F) override;
212 bool doInitialization(Module &M) override;
213 static char ID; // Pass identification, replacement for typeid.
216 void initializeCallbacks(Module &M);
218 /// \brief Track origins (allocation points) of uninitialized values.
221 const DataLayout *DL;
225 /// \brief Thread-local shadow storage for function parameters.
226 GlobalVariable *ParamTLS;
227 /// \brief Thread-local origin storage for function parameters.
228 GlobalVariable *ParamOriginTLS;
229 /// \brief Thread-local shadow storage for function return value.
230 GlobalVariable *RetvalTLS;
231 /// \brief Thread-local origin storage for function return value.
232 GlobalVariable *RetvalOriginTLS;
233 /// \brief Thread-local shadow storage for in-register va_arg function
234 /// parameters (x86_64-specific).
235 GlobalVariable *VAArgTLS;
236 /// \brief Thread-local shadow storage for va_arg overflow area
237 /// (x86_64-specific).
238 GlobalVariable *VAArgOverflowSizeTLS;
239 /// \brief Thread-local space used to pass origin value to the UMR reporting
241 GlobalVariable *OriginTLS;
243 GlobalVariable *MsandrModuleStart;
244 GlobalVariable *MsandrModuleEnd;
246 /// \brief The run-time callback to print a warning.
248 /// \brief Run-time helper that generates a new origin value for a stack
250 Value *MsanSetAllocaOrigin4Fn;
251 /// \brief Run-time helper that poisons stack on function entry.
252 Value *MsanPoisonStackFn;
253 /// \brief Run-time helper that records a store (or any event) of an
254 /// uninitialized value and returns an updated origin id encoding this info.
255 Value *MsanChainOriginFn;
256 /// \brief MSan runtime replacements for memmove, memcpy and memset.
257 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
259 /// \brief Address mask used in application-to-shadow address calculation.
260 /// ShadowAddr is computed as ApplicationAddr & ~ShadowMask.
262 /// \brief Offset of the origin shadow from the "normal" shadow.
263 /// OriginAddr is computed as (ShadowAddr + OriginOffset) & ~3ULL
264 uint64_t OriginOffset;
265 /// \brief Branch weights for error reporting.
266 MDNode *ColdCallWeights;
267 /// \brief Branch weights for origin store.
268 MDNode *OriginStoreWeights;
269 /// \brief Path to blacklist file.
270 SmallString<64> BlacklistFile;
271 /// \brief The blacklist.
272 std::unique_ptr<SpecialCaseList> BL;
273 /// \brief An empty volatile inline asm that prevents callback merge.
276 bool WrapIndirectCalls;
277 /// \brief Run-time wrapper for indirect calls.
278 Value *IndirectCallWrapperFn;
279 // Argument and return type of IndirectCallWrapperFn: void (*f)(void).
280 Type *AnyFunctionPtrTy;
282 friend struct MemorySanitizerVisitor;
283 friend struct VarArgAMD64Helper;
287 char MemorySanitizer::ID = 0;
288 INITIALIZE_PASS(MemorySanitizer, "msan",
289 "MemorySanitizer: detects uninitialized reads.",
292 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins,
293 StringRef BlacklistFile) {
294 return new MemorySanitizer(TrackOrigins, BlacklistFile);
297 /// \brief Create a non-const global initialized with the given string.
299 /// Creates a writable global for Str so that we can pass it to the
300 /// run-time lib. Runtime uses first 4 bytes of the string to store the
301 /// frame ID, so the string needs to be mutable.
302 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
304 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
305 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
306 GlobalValue::PrivateLinkage, StrConst, "");
310 /// \brief Insert extern declaration of runtime-provided functions and globals.
311 void MemorySanitizer::initializeCallbacks(Module &M) {
312 // Only do this once.
317 // Create the callback.
318 // FIXME: this function should have "Cold" calling conv,
319 // which is not yet implemented.
320 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
321 : "__msan_warning_noreturn";
322 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), NULL);
324 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
325 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
326 IRB.getInt8PtrTy(), IntptrTy, NULL);
327 MsanPoisonStackFn = M.getOrInsertFunction(
328 "__msan_poison_stack", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy, NULL);
329 MsanChainOriginFn = M.getOrInsertFunction(
330 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), NULL);
331 MemmoveFn = M.getOrInsertFunction(
332 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
333 IRB.getInt8PtrTy(), IntptrTy, NULL);
334 MemcpyFn = M.getOrInsertFunction(
335 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
337 MemsetFn = M.getOrInsertFunction(
338 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
342 RetvalTLS = new GlobalVariable(
343 M, ArrayType::get(IRB.getInt64Ty(), 8), false,
344 GlobalVariable::ExternalLinkage, 0, "__msan_retval_tls", 0,
345 GlobalVariable::InitialExecTLSModel);
346 RetvalOriginTLS = new GlobalVariable(
347 M, OriginTy, false, GlobalVariable::ExternalLinkage, 0,
348 "__msan_retval_origin_tls", 0, GlobalVariable::InitialExecTLSModel);
350 ParamTLS = new GlobalVariable(
351 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
352 GlobalVariable::ExternalLinkage, 0, "__msan_param_tls", 0,
353 GlobalVariable::InitialExecTLSModel);
354 ParamOriginTLS = new GlobalVariable(
355 M, ArrayType::get(OriginTy, 1000), false, GlobalVariable::ExternalLinkage,
356 0, "__msan_param_origin_tls", 0, GlobalVariable::InitialExecTLSModel);
358 VAArgTLS = new GlobalVariable(
359 M, ArrayType::get(IRB.getInt64Ty(), 1000), false,
360 GlobalVariable::ExternalLinkage, 0, "__msan_va_arg_tls", 0,
361 GlobalVariable::InitialExecTLSModel);
362 VAArgOverflowSizeTLS = new GlobalVariable(
363 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, 0,
364 "__msan_va_arg_overflow_size_tls", 0,
365 GlobalVariable::InitialExecTLSModel);
366 OriginTLS = new GlobalVariable(
367 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, 0,
368 "__msan_origin_tls", 0, GlobalVariable::InitialExecTLSModel);
370 // We insert an empty inline asm after __msan_report* to avoid callback merge.
371 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
372 StringRef(""), StringRef(""),
373 /*hasSideEffects=*/true);
375 if (WrapIndirectCalls) {
377 PointerType::getUnqual(FunctionType::get(IRB.getVoidTy(), false));
378 IndirectCallWrapperFn = M.getOrInsertFunction(
379 ClWrapIndirectCalls, AnyFunctionPtrTy, AnyFunctionPtrTy, NULL);
382 if (ClWrapIndirectCallsFast) {
383 MsandrModuleStart = new GlobalVariable(
384 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
385 0, "__executable_start");
386 MsandrModuleStart->setVisibility(GlobalVariable::HiddenVisibility);
387 MsandrModuleEnd = new GlobalVariable(
388 M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage,
390 MsandrModuleEnd->setVisibility(GlobalVariable::HiddenVisibility);
394 /// \brief Module-level initialization.
396 /// inserts a call to __msan_init to the module's constructor list.
397 bool MemorySanitizer::doInitialization(Module &M) {
398 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
401 DL = &DLP->getDataLayout();
403 BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
404 C = &(M.getContext());
405 unsigned PtrSize = DL->getPointerSizeInBits(/* AddressSpace */0);
408 ShadowMask = kShadowMask64;
409 OriginOffset = kOriginOffset64;
412 ShadowMask = kShadowMask32;
413 OriginOffset = kOriginOffset32;
416 report_fatal_error("unsupported pointer size");
421 IntptrTy = IRB.getIntPtrTy(DL);
422 OriginTy = IRB.getInt32Ty();
424 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
425 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
427 // Insert a call to __msan_init/__msan_track_origins into the module's CTORs.
428 appendToGlobalCtors(M, cast<Function>(M.getOrInsertFunction(
429 "__msan_init", IRB.getVoidTy(), NULL)), 0);
432 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
433 IRB.getInt32(TrackOrigins), "__msan_track_origins");
436 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
437 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
444 /// \brief A helper class that handles instrumentation of VarArg
445 /// functions on a particular platform.
447 /// Implementations are expected to insert the instrumentation
448 /// necessary to propagate argument shadow through VarArg function
449 /// calls. Visit* methods are called during an InstVisitor pass over
450 /// the function, and should avoid creating new basic blocks. A new
451 /// instance of this class is created for each instrumented function.
452 struct VarArgHelper {
453 /// \brief Visit a CallSite.
454 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
456 /// \brief Visit a va_start call.
457 virtual void visitVAStartInst(VAStartInst &I) = 0;
459 /// \brief Visit a va_copy call.
460 virtual void visitVACopyInst(VACopyInst &I) = 0;
462 /// \brief Finalize function instrumentation.
464 /// This method is called after visiting all interesting (see above)
465 /// instructions in a function.
466 virtual void finalizeInstrumentation() = 0;
468 virtual ~VarArgHelper() {}
471 struct MemorySanitizerVisitor;
474 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
475 MemorySanitizerVisitor &Visitor);
477 /// This class does all the work for a given function. Store and Load
478 /// instructions store and load corresponding shadow and origin
479 /// values. Most instructions propagate shadow from arguments to their
480 /// return values. Certain instructions (most importantly, BranchInst)
481 /// test their argument shadow and print reports (with a runtime call) if it's
483 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
486 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
487 ValueMap<Value*, Value*> ShadowMap, OriginMap;
488 std::unique_ptr<VarArgHelper> VAHelper;
490 // The following flags disable parts of MSan instrumentation based on
491 // blacklist contents and command-line options.
496 bool CheckReturnValue;
498 struct ShadowOriginAndInsertPoint {
501 Instruction *OrigIns;
502 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
503 : Shadow(S), Origin(O), OrigIns(I) { }
505 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
506 SmallVector<Instruction*, 16> StoreList;
507 SmallVector<CallSite, 16> IndirectCallList;
509 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
510 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
511 bool SanitizeFunction = !MS.BL->isIn(F) && F.getAttributes().hasAttribute(
512 AttributeSet::FunctionIndex,
513 Attribute::SanitizeMemory);
514 InsertChecks = SanitizeFunction;
515 LoadShadow = SanitizeFunction;
516 PoisonStack = SanitizeFunction && ClPoisonStack;
517 PoisonUndef = SanitizeFunction && ClPoisonUndef;
518 // FIXME: Consider using SpecialCaseList to specify a list of functions that
519 // must always return fully initialized values. For now, we hardcode "main".
520 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
522 DEBUG(if (!InsertChecks)
523 dbgs() << "MemorySanitizer is not inserting checks into '"
524 << F.getName() << "'\n");
527 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
528 if (MS.TrackOrigins <= 1) return V;
529 return IRB.CreateCall(MS.MsanChainOriginFn, V);
532 void materializeStores() {
533 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
534 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]);
537 Value *Val = I.getValueOperand();
538 Value *Addr = I.getPointerOperand();
539 Value *Shadow = I.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
540 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
543 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
544 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
547 if (ClCheckAccessAddress)
548 insertShadowCheck(Addr, &I);
551 I.setOrdering(addReleaseOrdering(I.getOrdering()));
553 if (MS.TrackOrigins) {
554 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
555 if (isa<StructType>(Shadow->getType())) {
556 IRB.CreateAlignedStore(updateOrigin(getOrigin(Val), IRB),
557 getOriginPtr(Addr, IRB), Alignment);
559 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
561 // TODO(eugenis): handle non-zero constant shadow by inserting an
562 // unconditional check (can not simply fail compilation as this could
563 // be in the dead code).
564 if (isa<Constant>(ConvertedShadow))
567 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
568 getCleanShadow(ConvertedShadow), "_mscmp");
569 Instruction *CheckTerm =
570 SplitBlockAndInsertIfThen(Cmp, &I, false, MS.OriginStoreWeights);
571 IRBuilder<> IRBNew(CheckTerm);
572 IRBNew.CreateAlignedStore(updateOrigin(getOrigin(Val), IRBNew),
573 getOriginPtr(Addr, IRBNew), Alignment);
579 void materializeChecks() {
580 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
581 Value *Shadow = InstrumentationList[i].Shadow;
582 Instruction *OrigIns = InstrumentationList[i].OrigIns;
583 IRBuilder<> IRB(OrigIns);
584 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
585 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
586 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
587 // See the comment in materializeStores().
588 if (isa<Constant>(ConvertedShadow))
590 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
591 getCleanShadow(ConvertedShadow), "_mscmp");
592 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
594 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
596 IRB.SetInsertPoint(CheckTerm);
597 if (MS.TrackOrigins) {
598 Value *Origin = InstrumentationList[i].Origin;
599 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
602 IRB.CreateCall(MS.WarningFn);
603 IRB.CreateCall(MS.EmptyAsm);
604 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
606 DEBUG(dbgs() << "DONE:\n" << F);
609 void materializeIndirectCalls() {
610 for (size_t i = 0, n = IndirectCallList.size(); i < n; i++) {
611 CallSite CS = IndirectCallList[i];
612 Instruction *I = CS.getInstruction();
613 BasicBlock *B = I->getParent();
615 Value *Fn0 = CS.getCalledValue();
616 Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
618 if (ClWrapIndirectCallsFast) {
619 // Check that call target is inside this module limits.
621 IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
622 Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
624 Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
625 IRB.CreateICmpUGE(Fn, End));
628 IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
630 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
631 NotInThisModule, NewFnPhi,
632 /* Unreachable */ false, MS.ColdCallWeights);
634 IRB.SetInsertPoint(CheckTerm);
635 // Slow path: call wrapper function to possibly transform the call
637 Value *NewFn = IRB.CreateBitCast(
638 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
640 NewFnPhi->addIncoming(Fn0, B);
641 NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
642 CS.setCalledFunction(NewFnPhi);
644 Value *NewFn = IRB.CreateBitCast(
645 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
646 CS.setCalledFunction(NewFn);
651 /// \brief Add MemorySanitizer instrumentation to a function.
652 bool runOnFunction() {
653 MS.initializeCallbacks(*F.getParent());
654 if (!MS.DL) return false;
656 // In the presence of unreachable blocks, we may see Phi nodes with
657 // incoming nodes from such blocks. Since InstVisitor skips unreachable
658 // blocks, such nodes will not have any shadow value associated with them.
659 // It's easier to remove unreachable blocks than deal with missing shadow.
660 removeUnreachableBlocks(F);
662 // Iterate all BBs in depth-first order and create shadow instructions
663 // for all instructions (where applicable).
664 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
665 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
669 // Finalize PHI nodes.
670 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
671 PHINode *PN = ShadowPHINodes[i];
672 PHINode *PNS = cast<PHINode>(getShadow(PN));
673 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
674 size_t NumValues = PN->getNumIncomingValues();
675 for (size_t v = 0; v < NumValues; v++) {
676 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
678 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
682 VAHelper->finalizeInstrumentation();
684 // Delayed instrumentation of StoreInst.
685 // This may add new checks to be inserted later.
688 // Insert shadow value checks.
691 // Wrap indirect calls.
692 materializeIndirectCalls();
697 /// \brief Compute the shadow type that corresponds to a given Value.
698 Type *getShadowTy(Value *V) {
699 return getShadowTy(V->getType());
702 /// \brief Compute the shadow type that corresponds to a given Type.
703 Type *getShadowTy(Type *OrigTy) {
704 if (!OrigTy->isSized()) {
707 // For integer type, shadow is the same as the original type.
708 // This may return weird-sized types like i1.
709 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
711 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
712 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
713 return VectorType::get(IntegerType::get(*MS.C, EltSize),
714 VT->getNumElements());
716 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
717 SmallVector<Type*, 4> Elements;
718 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
719 Elements.push_back(getShadowTy(ST->getElementType(i)));
720 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
721 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
724 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
725 return IntegerType::get(*MS.C, TypeSize);
728 /// \brief Flatten a vector type.
729 Type *getShadowTyNoVec(Type *ty) {
730 if (VectorType *vt = dyn_cast<VectorType>(ty))
731 return IntegerType::get(*MS.C, vt->getBitWidth());
735 /// \brief Convert a shadow value to it's flattened variant.
736 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
737 Type *Ty = V->getType();
738 Type *NoVecTy = getShadowTyNoVec(Ty);
739 if (Ty == NoVecTy) return V;
740 return IRB.CreateBitCast(V, NoVecTy);
743 /// \brief Compute the shadow address that corresponds to a given application
746 /// Shadow = Addr & ~ShadowMask.
747 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
750 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
751 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
752 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
755 /// \brief Compute the origin address that corresponds to a given application
758 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
759 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
761 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
762 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
764 IRB.CreateAdd(ShadowLong,
765 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
767 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
768 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
771 /// \brief Compute the shadow address for a given function argument.
773 /// Shadow = ParamTLS+ArgOffset.
774 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
776 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
777 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
778 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
782 /// \brief Compute the origin address for a given function argument.
783 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
785 if (!MS.TrackOrigins) return 0;
786 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
787 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
788 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
792 /// \brief Compute the shadow address for a retval.
793 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
794 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
795 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
799 /// \brief Compute the origin address for a retval.
800 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
801 // We keep a single origin for the entire retval. Might be too optimistic.
802 return MS.RetvalOriginTLS;
805 /// \brief Set SV to be the shadow value for V.
806 void setShadow(Value *V, Value *SV) {
807 assert(!ShadowMap.count(V) && "Values may only have one shadow");
811 /// \brief Set Origin to be the origin value for V.
812 void setOrigin(Value *V, Value *Origin) {
813 if (!MS.TrackOrigins) return;
814 assert(!OriginMap.count(V) && "Values may only have one origin");
815 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
816 OriginMap[V] = Origin;
819 /// \brief Create a clean shadow value for a given value.
821 /// Clean shadow (all zeroes) means all bits of the value are defined
823 Constant *getCleanShadow(Value *V) {
824 Type *ShadowTy = getShadowTy(V);
827 return Constant::getNullValue(ShadowTy);
830 /// \brief Create a dirty shadow of a given shadow type.
831 Constant *getPoisonedShadow(Type *ShadowTy) {
833 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
834 return Constant::getAllOnesValue(ShadowTy);
835 StructType *ST = cast<StructType>(ShadowTy);
836 SmallVector<Constant *, 4> Vals;
837 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
838 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
839 return ConstantStruct::get(ST, Vals);
842 /// \brief Create a dirty shadow for a given value.
843 Constant *getPoisonedShadow(Value *V) {
844 Type *ShadowTy = getShadowTy(V);
847 return getPoisonedShadow(ShadowTy);
850 /// \brief Create a clean (zero) origin.
851 Value *getCleanOrigin() {
852 return Constant::getNullValue(MS.OriginTy);
855 /// \brief Get the shadow value for a given Value.
857 /// This function either returns the value set earlier with setShadow,
858 /// or extracts if from ParamTLS (for function arguments).
859 Value *getShadow(Value *V) {
860 if (Instruction *I = dyn_cast<Instruction>(V)) {
861 // For instructions the shadow is already stored in the map.
862 Value *Shadow = ShadowMap[V];
864 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
866 assert(Shadow && "No shadow for a value");
870 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
871 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
872 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
876 if (Argument *A = dyn_cast<Argument>(V)) {
877 // For arguments we compute the shadow on demand and store it in the map.
878 Value **ShadowPtr = &ShadowMap[V];
881 Function *F = A->getParent();
882 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
883 unsigned ArgOffset = 0;
884 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
886 if (!AI->getType()->isSized()) {
887 DEBUG(dbgs() << "Arg is not sized\n");
890 unsigned Size = AI->hasByValAttr()
891 ? MS.DL->getTypeAllocSize(AI->getType()->getPointerElementType())
892 : MS.DL->getTypeAllocSize(AI->getType());
894 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
895 if (AI->hasByValAttr()) {
896 // ByVal pointer itself has clean shadow. We copy the actual
897 // argument shadow to the underlying memory.
898 // Figure out maximal valid memcpy alignment.
899 unsigned ArgAlign = AI->getParamAlignment();
901 Type *EltType = A->getType()->getPointerElementType();
902 ArgAlign = MS.DL->getABITypeAlignment(EltType);
904 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
905 Value *Cpy = EntryIRB.CreateMemCpy(
906 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
908 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
910 *ShadowPtr = getCleanShadow(V);
912 *ShadowPtr = EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
914 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
915 **ShadowPtr << "\n");
916 if (MS.TrackOrigins) {
917 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
918 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
921 ArgOffset += DataLayout::RoundUpAlignment(Size, kShadowTLSAlignment);
923 assert(*ShadowPtr && "Could not find shadow for an argument");
926 // For everything else the shadow is zero.
927 return getCleanShadow(V);
930 /// \brief Get the shadow for i-th argument of the instruction I.
931 Value *getShadow(Instruction *I, int i) {
932 return getShadow(I->getOperand(i));
935 /// \brief Get the origin for a value.
936 Value *getOrigin(Value *V) {
937 if (!MS.TrackOrigins) return 0;
938 if (isa<Instruction>(V) || isa<Argument>(V)) {
939 Value *Origin = OriginMap[V];
941 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
942 Origin = getCleanOrigin();
946 return getCleanOrigin();
949 /// \brief Get the origin for i-th argument of the instruction I.
950 Value *getOrigin(Instruction *I, int i) {
951 return getOrigin(I->getOperand(i));
954 /// \brief Remember the place where a shadow check should be inserted.
956 /// This location will be later instrumented with a check that will print a
957 /// UMR warning in runtime if the shadow value is not 0.
958 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
960 if (!InsertChecks) return;
962 Type *ShadowTy = Shadow->getType();
963 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
964 "Can only insert checks for integer and vector shadow types");
966 InstrumentationList.push_back(
967 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
970 /// \brief Remember the place where a shadow check should be inserted.
972 /// This location will be later instrumented with a check that will print a
973 /// UMR warning in runtime if the value is not fully defined.
974 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
976 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
978 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
979 insertShadowCheck(Shadow, Origin, OrigIns);
982 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
992 return AcquireRelease;
993 case SequentiallyConsistent:
994 return SequentiallyConsistent;
996 llvm_unreachable("Unknown ordering");
999 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1008 case AcquireRelease:
1009 return AcquireRelease;
1010 case SequentiallyConsistent:
1011 return SequentiallyConsistent;
1013 llvm_unreachable("Unknown ordering");
1016 // ------------------- Visitors.
1018 /// \brief Instrument LoadInst
1020 /// Loads the corresponding shadow and (optionally) origin.
1021 /// Optionally, checks that the load address is fully defined.
1022 void visitLoadInst(LoadInst &I) {
1023 assert(I.getType()->isSized() && "Load type must have size");
1024 IRBuilder<> IRB(I.getNextNode());
1025 Type *ShadowTy = getShadowTy(&I);
1026 Value *Addr = I.getPointerOperand();
1028 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1030 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1032 setShadow(&I, getCleanShadow(&I));
1035 if (ClCheckAccessAddress)
1036 insertShadowCheck(I.getPointerOperand(), &I);
1039 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1041 if (MS.TrackOrigins) {
1043 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
1045 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
1047 setOrigin(&I, getCleanOrigin());
1052 /// \brief Instrument StoreInst
1054 /// Stores the corresponding shadow and (optionally) origin.
1055 /// Optionally, checks that the store address is fully defined.
1056 void visitStoreInst(StoreInst &I) {
1057 StoreList.push_back(&I);
1060 void handleCASOrRMW(Instruction &I) {
1061 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1063 IRBuilder<> IRB(&I);
1064 Value *Addr = I.getOperand(0);
1065 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1067 if (ClCheckAccessAddress)
1068 insertShadowCheck(Addr, &I);
1070 // Only test the conditional argument of cmpxchg instruction.
1071 // The other argument can potentially be uninitialized, but we can not
1072 // detect this situation reliably without possible false positives.
1073 if (isa<AtomicCmpXchgInst>(I))
1074 insertShadowCheck(I.getOperand(1), &I);
1076 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1078 setShadow(&I, getCleanShadow(&I));
1081 void visitAtomicRMWInst(AtomicRMWInst &I) {
1083 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1086 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1088 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1091 // Vector manipulation.
1092 void visitExtractElementInst(ExtractElementInst &I) {
1093 insertShadowCheck(I.getOperand(1), &I);
1094 IRBuilder<> IRB(&I);
1095 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1097 setOrigin(&I, getOrigin(&I, 0));
1100 void visitInsertElementInst(InsertElementInst &I) {
1101 insertShadowCheck(I.getOperand(2), &I);
1102 IRBuilder<> IRB(&I);
1103 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1104 I.getOperand(2), "_msprop"));
1105 setOriginForNaryOp(I);
1108 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1109 insertShadowCheck(I.getOperand(2), &I);
1110 IRBuilder<> IRB(&I);
1111 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1112 I.getOperand(2), "_msprop"));
1113 setOriginForNaryOp(I);
1117 void visitSExtInst(SExtInst &I) {
1118 IRBuilder<> IRB(&I);
1119 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1120 setOrigin(&I, getOrigin(&I, 0));
1123 void visitZExtInst(ZExtInst &I) {
1124 IRBuilder<> IRB(&I);
1125 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1126 setOrigin(&I, getOrigin(&I, 0));
1129 void visitTruncInst(TruncInst &I) {
1130 IRBuilder<> IRB(&I);
1131 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1132 setOrigin(&I, getOrigin(&I, 0));
1135 void visitBitCastInst(BitCastInst &I) {
1136 IRBuilder<> IRB(&I);
1137 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1138 setOrigin(&I, getOrigin(&I, 0));
1141 void visitPtrToIntInst(PtrToIntInst &I) {
1142 IRBuilder<> IRB(&I);
1143 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1144 "_msprop_ptrtoint"));
1145 setOrigin(&I, getOrigin(&I, 0));
1148 void visitIntToPtrInst(IntToPtrInst &I) {
1149 IRBuilder<> IRB(&I);
1150 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1151 "_msprop_inttoptr"));
1152 setOrigin(&I, getOrigin(&I, 0));
1155 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1156 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1157 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1158 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1159 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1160 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1162 /// \brief Propagate shadow for bitwise AND.
1164 /// This code is exact, i.e. if, for example, a bit in the left argument
1165 /// is defined and 0, then neither the value not definedness of the
1166 /// corresponding bit in B don't affect the resulting shadow.
1167 void visitAnd(BinaryOperator &I) {
1168 IRBuilder<> IRB(&I);
1169 // "And" of 0 and a poisoned value results in unpoisoned value.
1170 // 1&1 => 1; 0&1 => 0; p&1 => p;
1171 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1172 // 1&p => p; 0&p => 0; p&p => p;
1173 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1174 Value *S1 = getShadow(&I, 0);
1175 Value *S2 = getShadow(&I, 1);
1176 Value *V1 = I.getOperand(0);
1177 Value *V2 = I.getOperand(1);
1178 if (V1->getType() != S1->getType()) {
1179 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1180 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1182 Value *S1S2 = IRB.CreateAnd(S1, S2);
1183 Value *V1S2 = IRB.CreateAnd(V1, S2);
1184 Value *S1V2 = IRB.CreateAnd(S1, V2);
1185 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1186 setOriginForNaryOp(I);
1189 void visitOr(BinaryOperator &I) {
1190 IRBuilder<> IRB(&I);
1191 // "Or" of 1 and a poisoned value results in unpoisoned value.
1192 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1193 // 1|0 => 1; 0|0 => 0; p|0 => p;
1194 // 1|p => 1; 0|p => p; p|p => p;
1195 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1196 Value *S1 = getShadow(&I, 0);
1197 Value *S2 = getShadow(&I, 1);
1198 Value *V1 = IRB.CreateNot(I.getOperand(0));
1199 Value *V2 = IRB.CreateNot(I.getOperand(1));
1200 if (V1->getType() != S1->getType()) {
1201 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1202 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1204 Value *S1S2 = IRB.CreateAnd(S1, S2);
1205 Value *V1S2 = IRB.CreateAnd(V1, S2);
1206 Value *S1V2 = IRB.CreateAnd(S1, V2);
1207 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1208 setOriginForNaryOp(I);
1211 /// \brief Default propagation of shadow and/or origin.
1213 /// This class implements the general case of shadow propagation, used in all
1214 /// cases where we don't know and/or don't care about what the operation
1215 /// actually does. It converts all input shadow values to a common type
1216 /// (extending or truncating as necessary), and bitwise OR's them.
1218 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1219 /// fully initialized), and less prone to false positives.
1221 /// This class also implements the general case of origin propagation. For a
1222 /// Nary operation, result origin is set to the origin of an argument that is
1223 /// not entirely initialized. If there is more than one such arguments, the
1224 /// rightmost of them is picked. It does not matter which one is picked if all
1225 /// arguments are initialized.
1226 template <bool CombineShadow>
1231 MemorySanitizerVisitor *MSV;
1234 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1235 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {}
1237 /// \brief Add a pair of shadow and origin values to the mix.
1238 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1239 if (CombineShadow) {
1244 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1245 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1249 if (MSV->MS.TrackOrigins) {
1254 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1255 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1256 MSV->getCleanShadow(FlatShadow));
1257 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1263 /// \brief Add an application value to the mix.
1264 Combiner &Add(Value *V) {
1265 Value *OpShadow = MSV->getShadow(V);
1266 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : 0;
1267 return Add(OpShadow, OpOrigin);
1270 /// \brief Set the current combined values as the given instruction's shadow
1272 void Done(Instruction *I) {
1273 if (CombineShadow) {
1275 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1276 MSV->setShadow(I, Shadow);
1278 if (MSV->MS.TrackOrigins) {
1280 MSV->setOrigin(I, Origin);
1285 typedef Combiner<true> ShadowAndOriginCombiner;
1286 typedef Combiner<false> OriginCombiner;
1288 /// \brief Propagate origin for arbitrary operation.
1289 void setOriginForNaryOp(Instruction &I) {
1290 if (!MS.TrackOrigins) return;
1291 IRBuilder<> IRB(&I);
1292 OriginCombiner OC(this, IRB);
1293 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1298 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1299 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1300 "Vector of pointers is not a valid shadow type");
1301 return Ty->isVectorTy() ?
1302 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1303 Ty->getPrimitiveSizeInBits();
1306 /// \brief Cast between two shadow types, extending or truncating as
1308 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1309 bool Signed = false) {
1310 Type *srcTy = V->getType();
1311 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1312 return IRB.CreateIntCast(V, dstTy, Signed);
1313 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1314 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1315 return IRB.CreateIntCast(V, dstTy, Signed);
1316 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1317 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1318 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1320 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1321 return IRB.CreateBitCast(V2, dstTy);
1322 // TODO: handle struct types.
1325 /// \brief Cast an application value to the type of its own shadow.
1326 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1327 Type *ShadowTy = getShadowTy(V);
1328 if (V->getType() == ShadowTy)
1330 if (V->getType()->isPtrOrPtrVectorTy())
1331 return IRB.CreatePtrToInt(V, ShadowTy);
1333 return IRB.CreateBitCast(V, ShadowTy);
1336 /// \brief Propagate shadow for arbitrary operation.
1337 void handleShadowOr(Instruction &I) {
1338 IRBuilder<> IRB(&I);
1339 ShadowAndOriginCombiner SC(this, IRB);
1340 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1345 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1346 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1347 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1348 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1349 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1350 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1351 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1353 void handleDiv(Instruction &I) {
1354 IRBuilder<> IRB(&I);
1355 // Strict on the second argument.
1356 insertShadowCheck(I.getOperand(1), &I);
1357 setShadow(&I, getShadow(&I, 0));
1358 setOrigin(&I, getOrigin(&I, 0));
1361 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1362 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1363 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1364 void visitURem(BinaryOperator &I) { handleDiv(I); }
1365 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1366 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1368 /// \brief Instrument == and != comparisons.
1370 /// Sometimes the comparison result is known even if some of the bits of the
1371 /// arguments are not.
1372 void handleEqualityComparison(ICmpInst &I) {
1373 IRBuilder<> IRB(&I);
1374 Value *A = I.getOperand(0);
1375 Value *B = I.getOperand(1);
1376 Value *Sa = getShadow(A);
1377 Value *Sb = getShadow(B);
1379 // Get rid of pointers and vectors of pointers.
1380 // For ints (and vectors of ints), types of A and Sa match,
1381 // and this is a no-op.
1382 A = IRB.CreatePointerCast(A, Sa->getType());
1383 B = IRB.CreatePointerCast(B, Sb->getType());
1385 // A == B <==> (C = A^B) == 0
1386 // A != B <==> (C = A^B) != 0
1388 Value *C = IRB.CreateXor(A, B);
1389 Value *Sc = IRB.CreateOr(Sa, Sb);
1390 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1391 // Result is defined if one of the following is true
1392 // * there is a defined 1 bit in C
1393 // * C is fully defined
1394 // Si = !(C & ~Sc) && Sc
1395 Value *Zero = Constant::getNullValue(Sc->getType());
1396 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1398 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1400 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1401 Si->setName("_msprop_icmp");
1403 setOriginForNaryOp(I);
1406 /// \brief Build the lowest possible value of V, taking into account V's
1407 /// uninitialized bits.
1408 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1411 // Split shadow into sign bit and other bits.
1412 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1413 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1414 // Maximise the undefined shadow bit, minimize other undefined bits.
1416 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1418 // Minimize undefined bits.
1419 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1423 /// \brief Build the highest possible value of V, taking into account V's
1424 /// uninitialized bits.
1425 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1428 // Split shadow into sign bit and other bits.
1429 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1430 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1431 // Minimise the undefined shadow bit, maximise other undefined bits.
1433 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1435 // Maximize undefined bits.
1436 return IRB.CreateOr(A, Sa);
1440 /// \brief Instrument relational comparisons.
1442 /// This function does exact shadow propagation for all relational
1443 /// comparisons of integers, pointers and vectors of those.
1444 /// FIXME: output seems suboptimal when one of the operands is a constant
1445 void handleRelationalComparisonExact(ICmpInst &I) {
1446 IRBuilder<> IRB(&I);
1447 Value *A = I.getOperand(0);
1448 Value *B = I.getOperand(1);
1449 Value *Sa = getShadow(A);
1450 Value *Sb = getShadow(B);
1452 // Get rid of pointers and vectors of pointers.
1453 // For ints (and vectors of ints), types of A and Sa match,
1454 // and this is a no-op.
1455 A = IRB.CreatePointerCast(A, Sa->getType());
1456 B = IRB.CreatePointerCast(B, Sb->getType());
1458 // Let [a0, a1] be the interval of possible values of A, taking into account
1459 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1460 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1461 bool IsSigned = I.isSigned();
1462 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1463 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1464 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1465 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1466 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1467 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1468 Value *Si = IRB.CreateXor(S1, S2);
1470 setOriginForNaryOp(I);
1473 /// \brief Instrument signed relational comparisons.
1475 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1476 /// propagating the highest bit of the shadow. Everything else is delegated
1477 /// to handleShadowOr().
1478 void handleSignedRelationalComparison(ICmpInst &I) {
1479 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1480 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1482 CmpInst::Predicate pre = I.getPredicate();
1483 if (constOp0 && constOp0->isNullValue() &&
1484 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1485 op = I.getOperand(1);
1486 } else if (constOp1 && constOp1->isNullValue() &&
1487 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1488 op = I.getOperand(0);
1491 IRBuilder<> IRB(&I);
1493 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1494 setShadow(&I, Shadow);
1495 setOrigin(&I, getOrigin(op));
1501 void visitICmpInst(ICmpInst &I) {
1502 if (!ClHandleICmp) {
1506 if (I.isEquality()) {
1507 handleEqualityComparison(I);
1511 assert(I.isRelational());
1512 if (ClHandleICmpExact) {
1513 handleRelationalComparisonExact(I);
1517 handleSignedRelationalComparison(I);
1521 assert(I.isUnsigned());
1522 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1523 handleRelationalComparisonExact(I);
1530 void visitFCmpInst(FCmpInst &I) {
1534 void handleShift(BinaryOperator &I) {
1535 IRBuilder<> IRB(&I);
1536 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1537 // Otherwise perform the same shift on S1.
1538 Value *S1 = getShadow(&I, 0);
1539 Value *S2 = getShadow(&I, 1);
1540 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1542 Value *V2 = I.getOperand(1);
1543 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1544 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1545 setOriginForNaryOp(I);
1548 void visitShl(BinaryOperator &I) { handleShift(I); }
1549 void visitAShr(BinaryOperator &I) { handleShift(I); }
1550 void visitLShr(BinaryOperator &I) { handleShift(I); }
1552 /// \brief Instrument llvm.memmove
1554 /// At this point we don't know if llvm.memmove will be inlined or not.
1555 /// If we don't instrument it and it gets inlined,
1556 /// our interceptor will not kick in and we will lose the memmove.
1557 /// If we instrument the call here, but it does not get inlined,
1558 /// we will memove the shadow twice: which is bad in case
1559 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1561 /// Similar situation exists for memcpy and memset.
1562 void visitMemMoveInst(MemMoveInst &I) {
1563 IRBuilder<> IRB(&I);
1566 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1567 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1568 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1569 I.eraseFromParent();
1572 // Similar to memmove: avoid copying shadow twice.
1573 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1574 // FIXME: consider doing manual inline for small constant sizes and proper
1576 void visitMemCpyInst(MemCpyInst &I) {
1577 IRBuilder<> IRB(&I);
1580 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1581 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1582 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1583 I.eraseFromParent();
1587 void visitMemSetInst(MemSetInst &I) {
1588 IRBuilder<> IRB(&I);
1591 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1592 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1593 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1594 I.eraseFromParent();
1597 void visitVAStartInst(VAStartInst &I) {
1598 VAHelper->visitVAStartInst(I);
1601 void visitVACopyInst(VACopyInst &I) {
1602 VAHelper->visitVACopyInst(I);
1605 enum IntrinsicKind {
1606 IK_DoesNotAccessMemory,
1611 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1612 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1613 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1614 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1615 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1616 const int UnknownModRefBehavior = IK_WritesMemory;
1617 #define GET_INTRINSIC_MODREF_BEHAVIOR
1618 #define ModRefBehavior IntrinsicKind
1619 #include "llvm/IR/Intrinsics.gen"
1620 #undef ModRefBehavior
1621 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1624 /// \brief Handle vector store-like intrinsics.
1626 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1627 /// has 1 pointer argument and 1 vector argument, returns void.
1628 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1629 IRBuilder<> IRB(&I);
1630 Value* Addr = I.getArgOperand(0);
1631 Value *Shadow = getShadow(&I, 1);
1632 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1634 // We don't know the pointer alignment (could be unaligned SSE store!).
1635 // Have to assume to worst case.
1636 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1638 if (ClCheckAccessAddress)
1639 insertShadowCheck(Addr, &I);
1641 // FIXME: use ClStoreCleanOrigin
1642 // FIXME: factor out common code from materializeStores
1643 if (MS.TrackOrigins)
1644 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1648 /// \brief Handle vector load-like intrinsics.
1650 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1651 /// has 1 pointer argument, returns a vector.
1652 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1653 IRBuilder<> IRB(&I);
1654 Value *Addr = I.getArgOperand(0);
1656 Type *ShadowTy = getShadowTy(&I);
1658 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1659 // We don't know the pointer alignment (could be unaligned SSE load!).
1660 // Have to assume to worst case.
1661 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1663 setShadow(&I, getCleanShadow(&I));
1666 if (ClCheckAccessAddress)
1667 insertShadowCheck(Addr, &I);
1669 if (MS.TrackOrigins) {
1671 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1673 setOrigin(&I, getCleanOrigin());
1678 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1680 /// Instrument intrinsics with any number of arguments of the same type,
1681 /// equal to the return type. The type should be simple (no aggregates or
1682 /// pointers; vectors are fine).
1683 /// Caller guarantees that this intrinsic does not access memory.
1684 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1685 Type *RetTy = I.getType();
1686 if (!(RetTy->isIntOrIntVectorTy() ||
1687 RetTy->isFPOrFPVectorTy() ||
1688 RetTy->isX86_MMXTy()))
1691 unsigned NumArgOperands = I.getNumArgOperands();
1693 for (unsigned i = 0; i < NumArgOperands; ++i) {
1694 Type *Ty = I.getArgOperand(i)->getType();
1699 IRBuilder<> IRB(&I);
1700 ShadowAndOriginCombiner SC(this, IRB);
1701 for (unsigned i = 0; i < NumArgOperands; ++i)
1702 SC.Add(I.getArgOperand(i));
1708 /// \brief Heuristically instrument unknown intrinsics.
1710 /// The main purpose of this code is to do something reasonable with all
1711 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1712 /// We recognize several classes of intrinsics by their argument types and
1713 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1714 /// sure that we know what the intrinsic does.
1716 /// We special-case intrinsics where this approach fails. See llvm.bswap
1717 /// handling as an example of that.
1718 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1719 unsigned NumArgOperands = I.getNumArgOperands();
1720 if (NumArgOperands == 0)
1723 Intrinsic::ID iid = I.getIntrinsicID();
1724 IntrinsicKind IK = getIntrinsicKind(iid);
1725 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1726 bool WritesMemory = IK == IK_WritesMemory;
1727 assert(!(OnlyReadsMemory && WritesMemory));
1729 if (NumArgOperands == 2 &&
1730 I.getArgOperand(0)->getType()->isPointerTy() &&
1731 I.getArgOperand(1)->getType()->isVectorTy() &&
1732 I.getType()->isVoidTy() &&
1734 // This looks like a vector store.
1735 return handleVectorStoreIntrinsic(I);
1738 if (NumArgOperands == 1 &&
1739 I.getArgOperand(0)->getType()->isPointerTy() &&
1740 I.getType()->isVectorTy() &&
1742 // This looks like a vector load.
1743 return handleVectorLoadIntrinsic(I);
1746 if (!OnlyReadsMemory && !WritesMemory)
1747 if (maybeHandleSimpleNomemIntrinsic(I))
1750 // FIXME: detect and handle SSE maskstore/maskload
1754 void handleBswap(IntrinsicInst &I) {
1755 IRBuilder<> IRB(&I);
1756 Value *Op = I.getArgOperand(0);
1757 Type *OpType = Op->getType();
1758 Function *BswapFunc = Intrinsic::getDeclaration(
1759 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1760 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1761 setOrigin(&I, getOrigin(Op));
1764 // \brief Instrument vector convert instrinsic.
1766 // This function instruments intrinsics like cvtsi2ss:
1767 // %Out = int_xxx_cvtyyy(%ConvertOp)
1769 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1770 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1771 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1772 // elements from \p CopyOp.
1773 // In most cases conversion involves floating-point value which may trigger a
1774 // hardware exception when not fully initialized. For this reason we require
1775 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1776 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1777 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1778 // return a fully initialized value.
1779 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1780 IRBuilder<> IRB(&I);
1781 Value *CopyOp, *ConvertOp;
1783 switch (I.getNumArgOperands()) {
1785 CopyOp = I.getArgOperand(0);
1786 ConvertOp = I.getArgOperand(1);
1789 ConvertOp = I.getArgOperand(0);
1793 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1796 // The first *NumUsedElements* elements of ConvertOp are converted to the
1797 // same number of output elements. The rest of the output is copied from
1798 // CopyOp, or (if not available) filled with zeroes.
1799 // Combine shadow for elements of ConvertOp that are used in this operation,
1800 // and insert a check.
1801 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1802 // int->any conversion.
1803 Value *ConvertShadow = getShadow(ConvertOp);
1804 Value *AggShadow = 0;
1805 if (ConvertOp->getType()->isVectorTy()) {
1806 AggShadow = IRB.CreateExtractElement(
1807 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
1808 for (int i = 1; i < NumUsedElements; ++i) {
1809 Value *MoreShadow = IRB.CreateExtractElement(
1810 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
1811 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
1814 AggShadow = ConvertShadow;
1816 assert(AggShadow->getType()->isIntegerTy());
1817 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
1819 // Build result shadow by zero-filling parts of CopyOp shadow that come from
1822 assert(CopyOp->getType() == I.getType());
1823 assert(CopyOp->getType()->isVectorTy());
1824 Value *ResultShadow = getShadow(CopyOp);
1825 Type *EltTy = ResultShadow->getType()->getVectorElementType();
1826 for (int i = 0; i < NumUsedElements; ++i) {
1827 ResultShadow = IRB.CreateInsertElement(
1828 ResultShadow, ConstantInt::getNullValue(EltTy),
1829 ConstantInt::get(IRB.getInt32Ty(), i));
1831 setShadow(&I, ResultShadow);
1832 setOrigin(&I, getOrigin(CopyOp));
1834 setShadow(&I, getCleanShadow(&I));
1838 // Given a scalar or vector, extract lower 64 bits (or less), and return all
1839 // zeroes if it is zero, and all ones otherwise.
1840 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
1841 if (S->getType()->isVectorTy())
1842 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
1843 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
1844 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1845 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
1848 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
1849 Type *T = S->getType();
1850 assert(T->isVectorTy());
1851 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1852 return IRB.CreateSExt(S2, T);
1855 // \brief Instrument vector shift instrinsic.
1857 // This function instruments intrinsics like int_x86_avx2_psll_w.
1858 // Intrinsic shifts %In by %ShiftSize bits.
1859 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
1860 // size, and the rest is ignored. Behavior is defined even if shift size is
1861 // greater than register (or field) width.
1862 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
1863 assert(I.getNumArgOperands() == 2);
1864 IRBuilder<> IRB(&I);
1865 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1866 // Otherwise perform the same shift on S1.
1867 Value *S1 = getShadow(&I, 0);
1868 Value *S2 = getShadow(&I, 1);
1869 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
1870 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
1871 Value *V1 = I.getOperand(0);
1872 Value *V2 = I.getOperand(1);
1873 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
1874 IRB.CreateBitCast(S1, V1->getType()), V2);
1875 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
1876 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1877 setOriginForNaryOp(I);
1880 void visitIntrinsicInst(IntrinsicInst &I) {
1881 switch (I.getIntrinsicID()) {
1882 case llvm::Intrinsic::bswap:
1885 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
1886 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
1887 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
1888 case llvm::Intrinsic::x86_avx512_cvtss2usi:
1889 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
1890 case llvm::Intrinsic::x86_avx512_cvttss2usi:
1891 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
1892 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
1893 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
1894 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
1895 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
1896 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
1897 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
1898 case llvm::Intrinsic::x86_sse2_cvtsd2si:
1899 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
1900 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
1901 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
1902 case llvm::Intrinsic::x86_sse2_cvtss2sd:
1903 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
1904 case llvm::Intrinsic::x86_sse2_cvttsd2si:
1905 case llvm::Intrinsic::x86_sse_cvtsi2ss:
1906 case llvm::Intrinsic::x86_sse_cvtsi642ss:
1907 case llvm::Intrinsic::x86_sse_cvtss2si64:
1908 case llvm::Intrinsic::x86_sse_cvtss2si:
1909 case llvm::Intrinsic::x86_sse_cvttss2si64:
1910 case llvm::Intrinsic::x86_sse_cvttss2si:
1911 handleVectorConvertIntrinsic(I, 1);
1913 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
1914 case llvm::Intrinsic::x86_sse2_cvtps2pd:
1915 case llvm::Intrinsic::x86_sse_cvtps2pi:
1916 case llvm::Intrinsic::x86_sse_cvttps2pi:
1917 handleVectorConvertIntrinsic(I, 2);
1919 case llvm::Intrinsic::x86_avx512_psll_dq:
1920 case llvm::Intrinsic::x86_avx512_psrl_dq:
1921 case llvm::Intrinsic::x86_avx2_psll_w:
1922 case llvm::Intrinsic::x86_avx2_psll_d:
1923 case llvm::Intrinsic::x86_avx2_psll_q:
1924 case llvm::Intrinsic::x86_avx2_pslli_w:
1925 case llvm::Intrinsic::x86_avx2_pslli_d:
1926 case llvm::Intrinsic::x86_avx2_pslli_q:
1927 case llvm::Intrinsic::x86_avx2_psll_dq:
1928 case llvm::Intrinsic::x86_avx2_psrl_w:
1929 case llvm::Intrinsic::x86_avx2_psrl_d:
1930 case llvm::Intrinsic::x86_avx2_psrl_q:
1931 case llvm::Intrinsic::x86_avx2_psra_w:
1932 case llvm::Intrinsic::x86_avx2_psra_d:
1933 case llvm::Intrinsic::x86_avx2_psrli_w:
1934 case llvm::Intrinsic::x86_avx2_psrli_d:
1935 case llvm::Intrinsic::x86_avx2_psrli_q:
1936 case llvm::Intrinsic::x86_avx2_psrai_w:
1937 case llvm::Intrinsic::x86_avx2_psrai_d:
1938 case llvm::Intrinsic::x86_avx2_psrl_dq:
1939 case llvm::Intrinsic::x86_sse2_psll_w:
1940 case llvm::Intrinsic::x86_sse2_psll_d:
1941 case llvm::Intrinsic::x86_sse2_psll_q:
1942 case llvm::Intrinsic::x86_sse2_pslli_w:
1943 case llvm::Intrinsic::x86_sse2_pslli_d:
1944 case llvm::Intrinsic::x86_sse2_pslli_q:
1945 case llvm::Intrinsic::x86_sse2_psll_dq:
1946 case llvm::Intrinsic::x86_sse2_psrl_w:
1947 case llvm::Intrinsic::x86_sse2_psrl_d:
1948 case llvm::Intrinsic::x86_sse2_psrl_q:
1949 case llvm::Intrinsic::x86_sse2_psra_w:
1950 case llvm::Intrinsic::x86_sse2_psra_d:
1951 case llvm::Intrinsic::x86_sse2_psrli_w:
1952 case llvm::Intrinsic::x86_sse2_psrli_d:
1953 case llvm::Intrinsic::x86_sse2_psrli_q:
1954 case llvm::Intrinsic::x86_sse2_psrai_w:
1955 case llvm::Intrinsic::x86_sse2_psrai_d:
1956 case llvm::Intrinsic::x86_sse2_psrl_dq:
1957 case llvm::Intrinsic::x86_mmx_psll_w:
1958 case llvm::Intrinsic::x86_mmx_psll_d:
1959 case llvm::Intrinsic::x86_mmx_psll_q:
1960 case llvm::Intrinsic::x86_mmx_pslli_w:
1961 case llvm::Intrinsic::x86_mmx_pslli_d:
1962 case llvm::Intrinsic::x86_mmx_pslli_q:
1963 case llvm::Intrinsic::x86_mmx_psrl_w:
1964 case llvm::Intrinsic::x86_mmx_psrl_d:
1965 case llvm::Intrinsic::x86_mmx_psrl_q:
1966 case llvm::Intrinsic::x86_mmx_psra_w:
1967 case llvm::Intrinsic::x86_mmx_psra_d:
1968 case llvm::Intrinsic::x86_mmx_psrli_w:
1969 case llvm::Intrinsic::x86_mmx_psrli_d:
1970 case llvm::Intrinsic::x86_mmx_psrli_q:
1971 case llvm::Intrinsic::x86_mmx_psrai_w:
1972 case llvm::Intrinsic::x86_mmx_psrai_d:
1973 handleVectorShiftIntrinsic(I, /* Variable */ false);
1975 case llvm::Intrinsic::x86_avx2_psllv_d:
1976 case llvm::Intrinsic::x86_avx2_psllv_d_256:
1977 case llvm::Intrinsic::x86_avx2_psllv_q:
1978 case llvm::Intrinsic::x86_avx2_psllv_q_256:
1979 case llvm::Intrinsic::x86_avx2_psrlv_d:
1980 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
1981 case llvm::Intrinsic::x86_avx2_psrlv_q:
1982 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
1983 case llvm::Intrinsic::x86_avx2_psrav_d:
1984 case llvm::Intrinsic::x86_avx2_psrav_d_256:
1985 handleVectorShiftIntrinsic(I, /* Variable */ true);
1988 // Byte shifts are not implemented.
1989 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
1990 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
1991 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
1992 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
1993 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
1994 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
1997 if (!handleUnknownIntrinsic(I))
1998 visitInstruction(I);
2003 void visitCallSite(CallSite CS) {
2004 Instruction &I = *CS.getInstruction();
2005 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2007 CallInst *Call = cast<CallInst>(&I);
2009 // For inline asm, do the usual thing: check argument shadow and mark all
2010 // outputs as clean. Note that any side effects of the inline asm that are
2011 // not immediately visible in its constraints are not handled.
2012 if (Call->isInlineAsm()) {
2013 visitInstruction(I);
2017 // Allow only tail calls with the same types, otherwise
2018 // we may have a false positive: shadow for a non-void RetVal
2019 // will get propagated to a void RetVal.
2020 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
2021 Call->setTailCall(false);
2023 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2025 // We are going to insert code that relies on the fact that the callee
2026 // will become a non-readonly function after it is instrumented by us. To
2027 // prevent this code from being optimized out, mark that function
2028 // non-readonly in advance.
2029 if (Function *Func = Call->getCalledFunction()) {
2030 // Clear out readonly/readnone attributes.
2032 B.addAttribute(Attribute::ReadOnly)
2033 .addAttribute(Attribute::ReadNone);
2034 Func->removeAttributes(AttributeSet::FunctionIndex,
2035 AttributeSet::get(Func->getContext(),
2036 AttributeSet::FunctionIndex,
2040 IRBuilder<> IRB(&I);
2042 if (MS.WrapIndirectCalls && !CS.getCalledFunction())
2043 IndirectCallList.push_back(CS);
2045 unsigned ArgOffset = 0;
2046 DEBUG(dbgs() << " CallSite: " << I << "\n");
2047 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2048 ArgIt != End; ++ArgIt) {
2050 unsigned i = ArgIt - CS.arg_begin();
2051 if (!A->getType()->isSized()) {
2052 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2057 // Compute the Shadow for arg even if it is ByVal, because
2058 // in that case getShadow() will copy the actual arg shadow to
2059 // __msan_param_tls.
2060 Value *ArgShadow = getShadow(A);
2061 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2062 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2063 " Shadow: " << *ArgShadow << "\n");
2064 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2065 assert(A->getType()->isPointerTy() &&
2066 "ByVal argument is not a pointer!");
2067 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2068 unsigned Alignment = CS.getParamAlignment(i + 1);
2069 Store = IRB.CreateMemCpy(ArgShadowBase,
2070 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2073 Size = MS.DL->getTypeAllocSize(A->getType());
2074 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2075 kShadowTLSAlignment);
2077 if (MS.TrackOrigins)
2078 IRB.CreateStore(getOrigin(A),
2079 getOriginPtrForArgument(A, IRB, ArgOffset));
2081 assert(Size != 0 && Store != 0);
2082 DEBUG(dbgs() << " Param:" << *Store << "\n");
2083 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
2085 DEBUG(dbgs() << " done with call args\n");
2088 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2089 if (FT->isVarArg()) {
2090 VAHelper->visitCallSite(CS, IRB);
2093 // Now, get the shadow for the RetVal.
2094 if (!I.getType()->isSized()) return;
2095 IRBuilder<> IRBBefore(&I);
2096 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2097 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2098 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2099 Instruction *NextInsn = 0;
2101 NextInsn = I.getNextNode();
2103 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2104 if (!NormalDest->getSinglePredecessor()) {
2105 // FIXME: this case is tricky, so we are just conservative here.
2106 // Perhaps we need to split the edge between this BB and NormalDest,
2107 // but a naive attempt to use SplitEdge leads to a crash.
2108 setShadow(&I, getCleanShadow(&I));
2109 setOrigin(&I, getCleanOrigin());
2112 NextInsn = NormalDest->getFirstInsertionPt();
2114 "Could not find insertion point for retval shadow load");
2116 IRBuilder<> IRBAfter(NextInsn);
2117 Value *RetvalShadow =
2118 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2119 kShadowTLSAlignment, "_msret");
2120 setShadow(&I, RetvalShadow);
2121 if (MS.TrackOrigins)
2122 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2125 void visitReturnInst(ReturnInst &I) {
2126 IRBuilder<> IRB(&I);
2127 Value *RetVal = I.getReturnValue();
2128 if (!RetVal) return;
2129 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2130 if (CheckReturnValue) {
2131 insertShadowCheck(RetVal, &I);
2132 Value *Shadow = getCleanShadow(RetVal);
2133 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2135 Value *Shadow = getShadow(RetVal);
2136 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2137 // FIXME: make it conditional if ClStoreCleanOrigin==0
2138 if (MS.TrackOrigins)
2139 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2143 void visitPHINode(PHINode &I) {
2144 IRBuilder<> IRB(&I);
2145 ShadowPHINodes.push_back(&I);
2146 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2148 if (MS.TrackOrigins)
2149 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2153 void visitAllocaInst(AllocaInst &I) {
2154 setShadow(&I, getCleanShadow(&I));
2155 IRBuilder<> IRB(I.getNextNode());
2156 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2157 if (PoisonStack && ClPoisonStackWithCall) {
2158 IRB.CreateCall2(MS.MsanPoisonStackFn,
2159 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2160 ConstantInt::get(MS.IntptrTy, Size));
2162 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2163 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2164 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2167 if (PoisonStack && MS.TrackOrigins) {
2168 setOrigin(&I, getCleanOrigin());
2169 SmallString<2048> StackDescriptionStorage;
2170 raw_svector_ostream StackDescription(StackDescriptionStorage);
2171 // We create a string with a description of the stack allocation and
2172 // pass it into __msan_set_alloca_origin.
2173 // It will be printed by the run-time if stack-originated UMR is found.
2174 // The first 4 bytes of the string are set to '----' and will be replaced
2175 // by __msan_va_arg_overflow_size_tls at the first call.
2176 StackDescription << "----" << I.getName() << "@" << F.getName();
2178 createPrivateNonConstGlobalForString(*F.getParent(),
2179 StackDescription.str());
2181 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2182 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2183 ConstantInt::get(MS.IntptrTy, Size),
2184 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2185 IRB.CreatePointerCast(&F, MS.IntptrTy));
2189 void visitSelectInst(SelectInst& I) {
2190 IRBuilder<> IRB(&I);
2191 // a = select b, c, d
2192 Value *B = I.getCondition();
2193 Value *C = I.getTrueValue();
2194 Value *D = I.getFalseValue();
2195 Value *Sb = getShadow(B);
2196 Value *Sc = getShadow(C);
2197 Value *Sd = getShadow(D);
2199 // Result shadow if condition shadow is 0.
2200 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2202 if (I.getType()->isAggregateType()) {
2203 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2204 // an extra "select". This results in much more compact IR.
2205 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2206 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2208 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2209 // If Sb (condition is poisoned), look for bits in c and d that are equal
2210 // and both unpoisoned.
2211 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2213 // Cast arguments to shadow-compatible type.
2214 C = CreateAppToShadowCast(IRB, C);
2215 D = CreateAppToShadowCast(IRB, D);
2217 // Result shadow if condition shadow is 1.
2218 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2220 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2222 if (MS.TrackOrigins) {
2223 // Origins are always i32, so any vector conditions must be flattened.
2224 // FIXME: consider tracking vector origins for app vectors?
2225 if (B->getType()->isVectorTy()) {
2226 Type *FlatTy = getShadowTyNoVec(B->getType());
2227 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2228 ConstantInt::getNullValue(FlatTy));
2229 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2230 ConstantInt::getNullValue(FlatTy));
2232 // a = select b, c, d
2233 // Oa = Sb ? Ob : (b ? Oc : Od)
2234 setOrigin(&I, IRB.CreateSelect(
2235 Sb, getOrigin(I.getCondition()),
2236 IRB.CreateSelect(B, getOrigin(C), getOrigin(D))));
2240 void visitLandingPadInst(LandingPadInst &I) {
2242 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2243 setShadow(&I, getCleanShadow(&I));
2244 setOrigin(&I, getCleanOrigin());
2247 void visitGetElementPtrInst(GetElementPtrInst &I) {
2251 void visitExtractValueInst(ExtractValueInst &I) {
2252 IRBuilder<> IRB(&I);
2253 Value *Agg = I.getAggregateOperand();
2254 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2255 Value *AggShadow = getShadow(Agg);
2256 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2257 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2258 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2259 setShadow(&I, ResShadow);
2260 setOriginForNaryOp(I);
2263 void visitInsertValueInst(InsertValueInst &I) {
2264 IRBuilder<> IRB(&I);
2265 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2266 Value *AggShadow = getShadow(I.getAggregateOperand());
2267 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2268 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2269 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2270 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2271 DEBUG(dbgs() << " Res: " << *Res << "\n");
2273 setOriginForNaryOp(I);
2276 void dumpInst(Instruction &I) {
2277 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2278 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2280 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2282 errs() << "QQQ " << I << "\n";
2285 void visitResumeInst(ResumeInst &I) {
2286 DEBUG(dbgs() << "Resume: " << I << "\n");
2287 // Nothing to do here.
2290 void visitInstruction(Instruction &I) {
2291 // Everything else: stop propagating and check for poisoned shadow.
2292 if (ClDumpStrictInstructions)
2294 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2295 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2296 insertShadowCheck(I.getOperand(i), &I);
2297 setShadow(&I, getCleanShadow(&I));
2298 setOrigin(&I, getCleanOrigin());
2302 /// \brief AMD64-specific implementation of VarArgHelper.
2303 struct VarArgAMD64Helper : public VarArgHelper {
2304 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2305 // See a comment in visitCallSite for more details.
2306 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2307 static const unsigned AMD64FpEndOffset = 176;
2310 MemorySanitizer &MS;
2311 MemorySanitizerVisitor &MSV;
2312 Value *VAArgTLSCopy;
2313 Value *VAArgOverflowSize;
2315 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2317 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2318 MemorySanitizerVisitor &MSV)
2319 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
2321 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2323 ArgKind classifyArgument(Value* arg) {
2324 // A very rough approximation of X86_64 argument classification rules.
2325 Type *T = arg->getType();
2326 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2327 return AK_FloatingPoint;
2328 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2329 return AK_GeneralPurpose;
2330 if (T->isPointerTy())
2331 return AK_GeneralPurpose;
2335 // For VarArg functions, store the argument shadow in an ABI-specific format
2336 // that corresponds to va_list layout.
2337 // We do this because Clang lowers va_arg in the frontend, and this pass
2338 // only sees the low level code that deals with va_list internals.
2339 // A much easier alternative (provided that Clang emits va_arg instructions)
2340 // would have been to associate each live instance of va_list with a copy of
2341 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2343 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2344 unsigned GpOffset = 0;
2345 unsigned FpOffset = AMD64GpEndOffset;
2346 unsigned OverflowOffset = AMD64FpEndOffset;
2347 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2348 ArgIt != End; ++ArgIt) {
2350 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2351 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2353 // ByVal arguments always go to the overflow area.
2354 assert(A->getType()->isPointerTy());
2355 Type *RealTy = A->getType()->getPointerElementType();
2356 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2357 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2358 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2359 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2360 ArgSize, kShadowTLSAlignment);
2362 ArgKind AK = classifyArgument(A);
2363 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2365 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2369 case AK_GeneralPurpose:
2370 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2373 case AK_FloatingPoint:
2374 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2378 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2379 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2380 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2382 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2385 Constant *OverflowSize =
2386 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2387 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2390 /// \brief Compute the shadow address for a given va_arg.
2391 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2393 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2394 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2395 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2399 void visitVAStartInst(VAStartInst &I) override {
2400 IRBuilder<> IRB(&I);
2401 VAStartInstrumentationList.push_back(&I);
2402 Value *VAListTag = I.getArgOperand(0);
2403 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2405 // Unpoison the whole __va_list_tag.
2406 // FIXME: magic ABI constants.
2407 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2408 /* size */24, /* alignment */8, false);
2411 void visitVACopyInst(VACopyInst &I) override {
2412 IRBuilder<> IRB(&I);
2413 Value *VAListTag = I.getArgOperand(0);
2414 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2416 // Unpoison the whole __va_list_tag.
2417 // FIXME: magic ABI constants.
2418 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2419 /* size */24, /* alignment */8, false);
2422 void finalizeInstrumentation() override {
2423 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2424 "finalizeInstrumentation called twice");
2425 if (!VAStartInstrumentationList.empty()) {
2426 // If there is a va_start in this function, make a backup copy of
2427 // va_arg_tls somewhere in the function entry block.
2428 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2429 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2431 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2433 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2434 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2437 // Instrument va_start.
2438 // Copy va_list shadow from the backup copy of the TLS contents.
2439 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2440 CallInst *OrigInst = VAStartInstrumentationList[i];
2441 IRBuilder<> IRB(OrigInst->getNextNode());
2442 Value *VAListTag = OrigInst->getArgOperand(0);
2444 Value *RegSaveAreaPtrPtr =
2446 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2447 ConstantInt::get(MS.IntptrTy, 16)),
2448 Type::getInt64PtrTy(*MS.C));
2449 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2450 Value *RegSaveAreaShadowPtr =
2451 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2452 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2453 AMD64FpEndOffset, 16);
2455 Value *OverflowArgAreaPtrPtr =
2457 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2458 ConstantInt::get(MS.IntptrTy, 8)),
2459 Type::getInt64PtrTy(*MS.C));
2460 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2461 Value *OverflowArgAreaShadowPtr =
2462 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2463 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2464 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2469 /// \brief A no-op implementation of VarArgHelper.
2470 struct VarArgNoOpHelper : public VarArgHelper {
2471 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2472 MemorySanitizerVisitor &MSV) {}
2474 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2476 void visitVAStartInst(VAStartInst &I) override {}
2478 void visitVACopyInst(VACopyInst &I) override {}
2480 void finalizeInstrumentation() override {}
2483 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2484 MemorySanitizerVisitor &Visitor) {
2485 // VarArg handling is only implemented on AMD64. False positives are possible
2486 // on other platforms.
2487 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2488 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2489 return new VarArgAMD64Helper(Func, Msan, Visitor);
2491 return new VarArgNoOpHelper(Func, Msan, Visitor);
2496 bool MemorySanitizer::runOnFunction(Function &F) {
2497 MemorySanitizerVisitor Visitor(F, *this);
2499 // Clear out readonly/readnone attributes.
2501 B.addAttribute(Attribute::ReadOnly)
2502 .addAttribute(Attribute::ReadNone);
2503 F.removeAttributes(AttributeSet::FunctionIndex,
2504 AttributeSet::get(F.getContext(),
2505 AttributeSet::FunctionIndex, B));
2507 return Visitor.runOnFunction();