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) { }
504 ShadowOriginAndInsertPoint() : Shadow(0), Origin(0), OrigIns(0) { }
506 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
507 SmallVector<Instruction*, 16> StoreList;
508 SmallVector<CallSite, 16> IndirectCallList;
510 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
511 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
512 bool SanitizeFunction = !MS.BL->isIn(F) && F.getAttributes().hasAttribute(
513 AttributeSet::FunctionIndex,
514 Attribute::SanitizeMemory);
515 InsertChecks = SanitizeFunction;
516 LoadShadow = SanitizeFunction;
517 PoisonStack = SanitizeFunction && ClPoisonStack;
518 PoisonUndef = SanitizeFunction && ClPoisonUndef;
519 // FIXME: Consider using SpecialCaseList to specify a list of functions that
520 // must always return fully initialized values. For now, we hardcode "main".
521 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
523 DEBUG(if (!InsertChecks)
524 dbgs() << "MemorySanitizer is not inserting checks into '"
525 << F.getName() << "'\n");
528 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
529 if (MS.TrackOrigins <= 1) return V;
530 return IRB.CreateCall(MS.MsanChainOriginFn, V);
533 void materializeStores() {
534 for (size_t i = 0, n = StoreList.size(); i < n; i++) {
535 StoreInst& I = *dyn_cast<StoreInst>(StoreList[i]);
538 Value *Val = I.getValueOperand();
539 Value *Addr = I.getPointerOperand();
540 Value *Shadow = I.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
541 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
544 IRB.CreateAlignedStore(Shadow, ShadowPtr, I.getAlignment());
545 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
548 if (ClCheckAccessAddress)
549 insertShadowCheck(Addr, &I);
552 I.setOrdering(addReleaseOrdering(I.getOrdering()));
554 if (MS.TrackOrigins) {
555 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
556 if (isa<StructType>(Shadow->getType())) {
557 IRB.CreateAlignedStore(updateOrigin(getOrigin(Val), IRB),
558 getOriginPtr(Addr, IRB), Alignment);
560 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
562 // TODO(eugenis): handle non-zero constant shadow by inserting an
563 // unconditional check (can not simply fail compilation as this could
564 // be in the dead code).
565 if (isa<Constant>(ConvertedShadow))
568 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
569 getCleanShadow(ConvertedShadow), "_mscmp");
570 Instruction *CheckTerm =
571 SplitBlockAndInsertIfThen(Cmp, &I, false, MS.OriginStoreWeights);
572 IRBuilder<> IRBNew(CheckTerm);
573 IRBNew.CreateAlignedStore(updateOrigin(getOrigin(Val), IRBNew),
574 getOriginPtr(Addr, IRBNew), Alignment);
580 void materializeChecks() {
581 for (size_t i = 0, n = InstrumentationList.size(); i < n; i++) {
582 Value *Shadow = InstrumentationList[i].Shadow;
583 Instruction *OrigIns = InstrumentationList[i].OrigIns;
584 IRBuilder<> IRB(OrigIns);
585 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
586 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
587 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
588 // See the comment in materializeStores().
589 if (isa<Constant>(ConvertedShadow))
591 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
592 getCleanShadow(ConvertedShadow), "_mscmp");
593 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
595 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
597 IRB.SetInsertPoint(CheckTerm);
598 if (MS.TrackOrigins) {
599 Value *Origin = InstrumentationList[i].Origin;
600 IRB.CreateStore(Origin ? (Value*)Origin : (Value*)IRB.getInt32(0),
603 CallInst *Call = IRB.CreateCall(MS.WarningFn);
604 Call->setDebugLoc(OrigIns->getDebugLoc());
605 IRB.CreateCall(MS.EmptyAsm);
606 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
608 DEBUG(dbgs() << "DONE:\n" << F);
611 void materializeIndirectCalls() {
612 for (size_t i = 0, n = IndirectCallList.size(); i < n; i++) {
613 CallSite CS = IndirectCallList[i];
614 Instruction *I = CS.getInstruction();
615 BasicBlock *B = I->getParent();
617 Value *Fn0 = CS.getCalledValue();
618 Value *Fn = IRB.CreateBitCast(Fn0, MS.AnyFunctionPtrTy);
620 if (ClWrapIndirectCallsFast) {
621 // Check that call target is inside this module limits.
623 IRB.CreateBitCast(MS.MsandrModuleStart, MS.AnyFunctionPtrTy);
624 Value *End = IRB.CreateBitCast(MS.MsandrModuleEnd, MS.AnyFunctionPtrTy);
626 Value *NotInThisModule = IRB.CreateOr(IRB.CreateICmpULT(Fn, Start),
627 IRB.CreateICmpUGE(Fn, End));
630 IRB.CreatePHI(Fn0->getType(), 2, "msandr.indirect_target");
632 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
633 NotInThisModule, NewFnPhi,
634 /* Unreachable */ false, MS.ColdCallWeights);
636 IRB.SetInsertPoint(CheckTerm);
637 // Slow path: call wrapper function to possibly transform the call
639 Value *NewFn = IRB.CreateBitCast(
640 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
642 NewFnPhi->addIncoming(Fn0, B);
643 NewFnPhi->addIncoming(NewFn, dyn_cast<Instruction>(NewFn)->getParent());
644 CS.setCalledFunction(NewFnPhi);
646 Value *NewFn = IRB.CreateBitCast(
647 IRB.CreateCall(MS.IndirectCallWrapperFn, Fn), Fn0->getType());
648 CS.setCalledFunction(NewFn);
653 /// \brief Add MemorySanitizer instrumentation to a function.
654 bool runOnFunction() {
655 MS.initializeCallbacks(*F.getParent());
656 if (!MS.DL) return false;
658 // In the presence of unreachable blocks, we may see Phi nodes with
659 // incoming nodes from such blocks. Since InstVisitor skips unreachable
660 // blocks, such nodes will not have any shadow value associated with them.
661 // It's easier to remove unreachable blocks than deal with missing shadow.
662 removeUnreachableBlocks(F);
664 // Iterate all BBs in depth-first order and create shadow instructions
665 // for all instructions (where applicable).
666 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
667 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
668 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
669 BasicBlock *BB = *DI;
673 // Finalize PHI nodes.
674 for (size_t i = 0, n = ShadowPHINodes.size(); i < n; i++) {
675 PHINode *PN = ShadowPHINodes[i];
676 PHINode *PNS = cast<PHINode>(getShadow(PN));
677 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : 0;
678 size_t NumValues = PN->getNumIncomingValues();
679 for (size_t v = 0; v < NumValues; v++) {
680 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
682 PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
686 VAHelper->finalizeInstrumentation();
688 // Delayed instrumentation of StoreInst.
689 // This may add new checks to be inserted later.
692 // Insert shadow value checks.
695 // Wrap indirect calls.
696 materializeIndirectCalls();
701 /// \brief Compute the shadow type that corresponds to a given Value.
702 Type *getShadowTy(Value *V) {
703 return getShadowTy(V->getType());
706 /// \brief Compute the shadow type that corresponds to a given Type.
707 Type *getShadowTy(Type *OrigTy) {
708 if (!OrigTy->isSized()) {
711 // For integer type, shadow is the same as the original type.
712 // This may return weird-sized types like i1.
713 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
715 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
716 uint32_t EltSize = MS.DL->getTypeSizeInBits(VT->getElementType());
717 return VectorType::get(IntegerType::get(*MS.C, EltSize),
718 VT->getNumElements());
720 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
721 SmallVector<Type*, 4> Elements;
722 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
723 Elements.push_back(getShadowTy(ST->getElementType(i)));
724 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
725 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
728 uint32_t TypeSize = MS.DL->getTypeSizeInBits(OrigTy);
729 return IntegerType::get(*MS.C, TypeSize);
732 /// \brief Flatten a vector type.
733 Type *getShadowTyNoVec(Type *ty) {
734 if (VectorType *vt = dyn_cast<VectorType>(ty))
735 return IntegerType::get(*MS.C, vt->getBitWidth());
739 /// \brief Convert a shadow value to it's flattened variant.
740 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
741 Type *Ty = V->getType();
742 Type *NoVecTy = getShadowTyNoVec(Ty);
743 if (Ty == NoVecTy) return V;
744 return IRB.CreateBitCast(V, NoVecTy);
747 /// \brief Compute the shadow address that corresponds to a given application
750 /// Shadow = Addr & ~ShadowMask.
751 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
754 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
755 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
756 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
759 /// \brief Compute the origin address that corresponds to a given application
762 /// OriginAddr = (ShadowAddr + OriginOffset) & ~3ULL
763 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB) {
765 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
766 ConstantInt::get(MS.IntptrTy, ~MS.ShadowMask));
768 IRB.CreateAdd(ShadowLong,
769 ConstantInt::get(MS.IntptrTy, MS.OriginOffset));
771 IRB.CreateAnd(Add, ConstantInt::get(MS.IntptrTy, ~3ULL));
772 return IRB.CreateIntToPtr(SecondAnd, PointerType::get(IRB.getInt32Ty(), 0));
775 /// \brief Compute the shadow address for a given function argument.
777 /// Shadow = ParamTLS+ArgOffset.
778 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
780 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
781 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
782 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
786 /// \brief Compute the origin address for a given function argument.
787 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
789 if (!MS.TrackOrigins) return 0;
790 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
791 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
792 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
796 /// \brief Compute the shadow address for a retval.
797 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
798 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
799 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
803 /// \brief Compute the origin address for a retval.
804 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
805 // We keep a single origin for the entire retval. Might be too optimistic.
806 return MS.RetvalOriginTLS;
809 /// \brief Set SV to be the shadow value for V.
810 void setShadow(Value *V, Value *SV) {
811 assert(!ShadowMap.count(V) && "Values may only have one shadow");
815 /// \brief Set Origin to be the origin value for V.
816 void setOrigin(Value *V, Value *Origin) {
817 if (!MS.TrackOrigins) return;
818 assert(!OriginMap.count(V) && "Values may only have one origin");
819 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
820 OriginMap[V] = Origin;
823 /// \brief Create a clean shadow value for a given value.
825 /// Clean shadow (all zeroes) means all bits of the value are defined
827 Constant *getCleanShadow(Value *V) {
828 Type *ShadowTy = getShadowTy(V);
831 return Constant::getNullValue(ShadowTy);
834 /// \brief Create a dirty shadow of a given shadow type.
835 Constant *getPoisonedShadow(Type *ShadowTy) {
837 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
838 return Constant::getAllOnesValue(ShadowTy);
839 StructType *ST = cast<StructType>(ShadowTy);
840 SmallVector<Constant *, 4> Vals;
841 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
842 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
843 return ConstantStruct::get(ST, Vals);
846 /// \brief Create a dirty shadow for a given value.
847 Constant *getPoisonedShadow(Value *V) {
848 Type *ShadowTy = getShadowTy(V);
851 return getPoisonedShadow(ShadowTy);
854 /// \brief Create a clean (zero) origin.
855 Value *getCleanOrigin() {
856 return Constant::getNullValue(MS.OriginTy);
859 /// \brief Get the shadow value for a given Value.
861 /// This function either returns the value set earlier with setShadow,
862 /// or extracts if from ParamTLS (for function arguments).
863 Value *getShadow(Value *V) {
864 if (Instruction *I = dyn_cast<Instruction>(V)) {
865 // For instructions the shadow is already stored in the map.
866 Value *Shadow = ShadowMap[V];
868 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
870 assert(Shadow && "No shadow for a value");
874 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
875 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
876 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
880 if (Argument *A = dyn_cast<Argument>(V)) {
881 // For arguments we compute the shadow on demand and store it in the map.
882 Value **ShadowPtr = &ShadowMap[V];
885 Function *F = A->getParent();
886 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
887 unsigned ArgOffset = 0;
888 for (Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
890 if (!AI->getType()->isSized()) {
891 DEBUG(dbgs() << "Arg is not sized\n");
894 unsigned Size = AI->hasByValAttr()
895 ? MS.DL->getTypeAllocSize(AI->getType()->getPointerElementType())
896 : MS.DL->getTypeAllocSize(AI->getType());
898 Value *Base = getShadowPtrForArgument(AI, EntryIRB, ArgOffset);
899 if (AI->hasByValAttr()) {
900 // ByVal pointer itself has clean shadow. We copy the actual
901 // argument shadow to the underlying memory.
902 // Figure out maximal valid memcpy alignment.
903 unsigned ArgAlign = AI->getParamAlignment();
905 Type *EltType = A->getType()->getPointerElementType();
906 ArgAlign = MS.DL->getABITypeAlignment(EltType);
908 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
909 Value *Cpy = EntryIRB.CreateMemCpy(
910 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
912 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
914 *ShadowPtr = getCleanShadow(V);
916 *ShadowPtr = EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
918 DEBUG(dbgs() << " ARG: " << *AI << " ==> " <<
919 **ShadowPtr << "\n");
920 if (MS.TrackOrigins) {
921 Value* OriginPtr = getOriginPtrForArgument(AI, EntryIRB, ArgOffset);
922 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
925 ArgOffset += DataLayout::RoundUpAlignment(Size, kShadowTLSAlignment);
927 assert(*ShadowPtr && "Could not find shadow for an argument");
930 // For everything else the shadow is zero.
931 return getCleanShadow(V);
934 /// \brief Get the shadow for i-th argument of the instruction I.
935 Value *getShadow(Instruction *I, int i) {
936 return getShadow(I->getOperand(i));
939 /// \brief Get the origin for a value.
940 Value *getOrigin(Value *V) {
941 if (!MS.TrackOrigins) return 0;
942 if (isa<Instruction>(V) || isa<Argument>(V)) {
943 Value *Origin = OriginMap[V];
945 DEBUG(dbgs() << "NO ORIGIN: " << *V << "\n");
946 Origin = getCleanOrigin();
950 return getCleanOrigin();
953 /// \brief Get the origin for i-th argument of the instruction I.
954 Value *getOrigin(Instruction *I, int i) {
955 return getOrigin(I->getOperand(i));
958 /// \brief Remember the place where a shadow check should be inserted.
960 /// This location will be later instrumented with a check that will print a
961 /// UMR warning in runtime if the shadow value is not 0.
962 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
964 if (!InsertChecks) return;
966 Type *ShadowTy = Shadow->getType();
967 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
968 "Can only insert checks for integer and vector shadow types");
970 InstrumentationList.push_back(
971 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
974 /// \brief Remember the place where a shadow check should be inserted.
976 /// This location will be later instrumented with a check that will print a
977 /// UMR warning in runtime if the value is not fully defined.
978 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
980 Instruction *Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
982 Instruction *Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
983 insertShadowCheck(Shadow, Origin, OrigIns);
986 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
996 return AcquireRelease;
997 case SequentiallyConsistent:
998 return SequentiallyConsistent;
1000 llvm_unreachable("Unknown ordering");
1003 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1012 case AcquireRelease:
1013 return AcquireRelease;
1014 case SequentiallyConsistent:
1015 return SequentiallyConsistent;
1017 llvm_unreachable("Unknown ordering");
1020 // ------------------- Visitors.
1022 /// \brief Instrument LoadInst
1024 /// Loads the corresponding shadow and (optionally) origin.
1025 /// Optionally, checks that the load address is fully defined.
1026 void visitLoadInst(LoadInst &I) {
1027 assert(I.getType()->isSized() && "Load type must have size");
1028 IRBuilder<> IRB(I.getNextNode());
1029 Type *ShadowTy = getShadowTy(&I);
1030 Value *Addr = I.getPointerOperand();
1032 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1034 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1036 setShadow(&I, getCleanShadow(&I));
1039 if (ClCheckAccessAddress)
1040 insertShadowCheck(I.getPointerOperand(), &I);
1043 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1045 if (MS.TrackOrigins) {
1047 unsigned Alignment = std::max(kMinOriginAlignment, I.getAlignment());
1049 IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB), Alignment));
1051 setOrigin(&I, getCleanOrigin());
1056 /// \brief Instrument StoreInst
1058 /// Stores the corresponding shadow and (optionally) origin.
1059 /// Optionally, checks that the store address is fully defined.
1060 void visitStoreInst(StoreInst &I) {
1061 StoreList.push_back(&I);
1064 void handleCASOrRMW(Instruction &I) {
1065 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1067 IRBuilder<> IRB(&I);
1068 Value *Addr = I.getOperand(0);
1069 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1071 if (ClCheckAccessAddress)
1072 insertShadowCheck(Addr, &I);
1074 // Only test the conditional argument of cmpxchg instruction.
1075 // The other argument can potentially be uninitialized, but we can not
1076 // detect this situation reliably without possible false positives.
1077 if (isa<AtomicCmpXchgInst>(I))
1078 insertShadowCheck(I.getOperand(1), &I);
1080 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1082 setShadow(&I, getCleanShadow(&I));
1085 void visitAtomicRMWInst(AtomicRMWInst &I) {
1087 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1090 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1092 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1095 // Vector manipulation.
1096 void visitExtractElementInst(ExtractElementInst &I) {
1097 insertShadowCheck(I.getOperand(1), &I);
1098 IRBuilder<> IRB(&I);
1099 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1101 setOrigin(&I, getOrigin(&I, 0));
1104 void visitInsertElementInst(InsertElementInst &I) {
1105 insertShadowCheck(I.getOperand(2), &I);
1106 IRBuilder<> IRB(&I);
1107 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1108 I.getOperand(2), "_msprop"));
1109 setOriginForNaryOp(I);
1112 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1113 insertShadowCheck(I.getOperand(2), &I);
1114 IRBuilder<> IRB(&I);
1115 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1116 I.getOperand(2), "_msprop"));
1117 setOriginForNaryOp(I);
1121 void visitSExtInst(SExtInst &I) {
1122 IRBuilder<> IRB(&I);
1123 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1124 setOrigin(&I, getOrigin(&I, 0));
1127 void visitZExtInst(ZExtInst &I) {
1128 IRBuilder<> IRB(&I);
1129 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1130 setOrigin(&I, getOrigin(&I, 0));
1133 void visitTruncInst(TruncInst &I) {
1134 IRBuilder<> IRB(&I);
1135 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1136 setOrigin(&I, getOrigin(&I, 0));
1139 void visitBitCastInst(BitCastInst &I) {
1140 IRBuilder<> IRB(&I);
1141 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1142 setOrigin(&I, getOrigin(&I, 0));
1145 void visitPtrToIntInst(PtrToIntInst &I) {
1146 IRBuilder<> IRB(&I);
1147 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1148 "_msprop_ptrtoint"));
1149 setOrigin(&I, getOrigin(&I, 0));
1152 void visitIntToPtrInst(IntToPtrInst &I) {
1153 IRBuilder<> IRB(&I);
1154 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1155 "_msprop_inttoptr"));
1156 setOrigin(&I, getOrigin(&I, 0));
1159 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1160 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1161 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1162 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1163 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1164 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1166 /// \brief Propagate shadow for bitwise AND.
1168 /// This code is exact, i.e. if, for example, a bit in the left argument
1169 /// is defined and 0, then neither the value not definedness of the
1170 /// corresponding bit in B don't affect the resulting shadow.
1171 void visitAnd(BinaryOperator &I) {
1172 IRBuilder<> IRB(&I);
1173 // "And" of 0 and a poisoned value results in unpoisoned value.
1174 // 1&1 => 1; 0&1 => 0; p&1 => p;
1175 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1176 // 1&p => p; 0&p => 0; p&p => p;
1177 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1178 Value *S1 = getShadow(&I, 0);
1179 Value *S2 = getShadow(&I, 1);
1180 Value *V1 = I.getOperand(0);
1181 Value *V2 = I.getOperand(1);
1182 if (V1->getType() != S1->getType()) {
1183 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1184 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1186 Value *S1S2 = IRB.CreateAnd(S1, S2);
1187 Value *V1S2 = IRB.CreateAnd(V1, S2);
1188 Value *S1V2 = IRB.CreateAnd(S1, V2);
1189 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1190 setOriginForNaryOp(I);
1193 void visitOr(BinaryOperator &I) {
1194 IRBuilder<> IRB(&I);
1195 // "Or" of 1 and a poisoned value results in unpoisoned value.
1196 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1197 // 1|0 => 1; 0|0 => 0; p|0 => p;
1198 // 1|p => 1; 0|p => p; p|p => p;
1199 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1200 Value *S1 = getShadow(&I, 0);
1201 Value *S2 = getShadow(&I, 1);
1202 Value *V1 = IRB.CreateNot(I.getOperand(0));
1203 Value *V2 = IRB.CreateNot(I.getOperand(1));
1204 if (V1->getType() != S1->getType()) {
1205 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1206 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1208 Value *S1S2 = IRB.CreateAnd(S1, S2);
1209 Value *V1S2 = IRB.CreateAnd(V1, S2);
1210 Value *S1V2 = IRB.CreateAnd(S1, V2);
1211 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1212 setOriginForNaryOp(I);
1215 /// \brief Default propagation of shadow and/or origin.
1217 /// This class implements the general case of shadow propagation, used in all
1218 /// cases where we don't know and/or don't care about what the operation
1219 /// actually does. It converts all input shadow values to a common type
1220 /// (extending or truncating as necessary), and bitwise OR's them.
1222 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1223 /// fully initialized), and less prone to false positives.
1225 /// This class also implements the general case of origin propagation. For a
1226 /// Nary operation, result origin is set to the origin of an argument that is
1227 /// not entirely initialized. If there is more than one such arguments, the
1228 /// rightmost of them is picked. It does not matter which one is picked if all
1229 /// arguments are initialized.
1230 template <bool CombineShadow>
1235 MemorySanitizerVisitor *MSV;
1238 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1239 Shadow(0), Origin(0), IRB(IRB), MSV(MSV) {}
1241 /// \brief Add a pair of shadow and origin values to the mix.
1242 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1243 if (CombineShadow) {
1248 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1249 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1253 if (MSV->MS.TrackOrigins) {
1258 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1259 Value *Cond = IRB.CreateICmpNE(FlatShadow,
1260 MSV->getCleanShadow(FlatShadow));
1261 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1267 /// \brief Add an application value to the mix.
1268 Combiner &Add(Value *V) {
1269 Value *OpShadow = MSV->getShadow(V);
1270 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : 0;
1271 return Add(OpShadow, OpOrigin);
1274 /// \brief Set the current combined values as the given instruction's shadow
1276 void Done(Instruction *I) {
1277 if (CombineShadow) {
1279 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1280 MSV->setShadow(I, Shadow);
1282 if (MSV->MS.TrackOrigins) {
1284 MSV->setOrigin(I, Origin);
1289 typedef Combiner<true> ShadowAndOriginCombiner;
1290 typedef Combiner<false> OriginCombiner;
1292 /// \brief Propagate origin for arbitrary operation.
1293 void setOriginForNaryOp(Instruction &I) {
1294 if (!MS.TrackOrigins) return;
1295 IRBuilder<> IRB(&I);
1296 OriginCombiner OC(this, IRB);
1297 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1302 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1303 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1304 "Vector of pointers is not a valid shadow type");
1305 return Ty->isVectorTy() ?
1306 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1307 Ty->getPrimitiveSizeInBits();
1310 /// \brief Cast between two shadow types, extending or truncating as
1312 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1313 bool Signed = false) {
1314 Type *srcTy = V->getType();
1315 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1316 return IRB.CreateIntCast(V, dstTy, Signed);
1317 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1318 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1319 return IRB.CreateIntCast(V, dstTy, Signed);
1320 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1321 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1322 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1324 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1325 return IRB.CreateBitCast(V2, dstTy);
1326 // TODO: handle struct types.
1329 /// \brief Propagate shadow for arbitrary operation.
1330 void handleShadowOr(Instruction &I) {
1331 IRBuilder<> IRB(&I);
1332 ShadowAndOriginCombiner SC(this, IRB);
1333 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1338 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1339 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1340 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1341 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1342 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1343 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1344 void visitMul(BinaryOperator &I) { handleShadowOr(I); }
1346 void handleDiv(Instruction &I) {
1347 IRBuilder<> IRB(&I);
1348 // Strict on the second argument.
1349 insertShadowCheck(I.getOperand(1), &I);
1350 setShadow(&I, getShadow(&I, 0));
1351 setOrigin(&I, getOrigin(&I, 0));
1354 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1355 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1356 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1357 void visitURem(BinaryOperator &I) { handleDiv(I); }
1358 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1359 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1361 /// \brief Instrument == and != comparisons.
1363 /// Sometimes the comparison result is known even if some of the bits of the
1364 /// arguments are not.
1365 void handleEqualityComparison(ICmpInst &I) {
1366 IRBuilder<> IRB(&I);
1367 Value *A = I.getOperand(0);
1368 Value *B = I.getOperand(1);
1369 Value *Sa = getShadow(A);
1370 Value *Sb = getShadow(B);
1372 // Get rid of pointers and vectors of pointers.
1373 // For ints (and vectors of ints), types of A and Sa match,
1374 // and this is a no-op.
1375 A = IRB.CreatePointerCast(A, Sa->getType());
1376 B = IRB.CreatePointerCast(B, Sb->getType());
1378 // A == B <==> (C = A^B) == 0
1379 // A != B <==> (C = A^B) != 0
1381 Value *C = IRB.CreateXor(A, B);
1382 Value *Sc = IRB.CreateOr(Sa, Sb);
1383 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1384 // Result is defined if one of the following is true
1385 // * there is a defined 1 bit in C
1386 // * C is fully defined
1387 // Si = !(C & ~Sc) && Sc
1388 Value *Zero = Constant::getNullValue(Sc->getType());
1389 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1391 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1393 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1394 Si->setName("_msprop_icmp");
1396 setOriginForNaryOp(I);
1399 /// \brief Build the lowest possible value of V, taking into account V's
1400 /// uninitialized bits.
1401 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1404 // Split shadow into sign bit and other bits.
1405 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1406 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1407 // Maximise the undefined shadow bit, minimize other undefined bits.
1409 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1411 // Minimize undefined bits.
1412 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1416 /// \brief Build the highest possible value of V, taking into account V's
1417 /// uninitialized bits.
1418 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1421 // Split shadow into sign bit and other bits.
1422 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1423 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1424 // Minimise the undefined shadow bit, maximise other undefined bits.
1426 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1428 // Maximize undefined bits.
1429 return IRB.CreateOr(A, Sa);
1433 /// \brief Instrument relational comparisons.
1435 /// This function does exact shadow propagation for all relational
1436 /// comparisons of integers, pointers and vectors of those.
1437 /// FIXME: output seems suboptimal when one of the operands is a constant
1438 void handleRelationalComparisonExact(ICmpInst &I) {
1439 IRBuilder<> IRB(&I);
1440 Value *A = I.getOperand(0);
1441 Value *B = I.getOperand(1);
1442 Value *Sa = getShadow(A);
1443 Value *Sb = getShadow(B);
1445 // Get rid of pointers and vectors of pointers.
1446 // For ints (and vectors of ints), types of A and Sa match,
1447 // and this is a no-op.
1448 A = IRB.CreatePointerCast(A, Sa->getType());
1449 B = IRB.CreatePointerCast(B, Sb->getType());
1451 // Let [a0, a1] be the interval of possible values of A, taking into account
1452 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1453 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1454 bool IsSigned = I.isSigned();
1455 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1456 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1457 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1458 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1459 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1460 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1461 Value *Si = IRB.CreateXor(S1, S2);
1463 setOriginForNaryOp(I);
1466 /// \brief Instrument signed relational comparisons.
1468 /// Handle (x<0) and (x>=0) comparisons (essentially, sign bit tests) by
1469 /// propagating the highest bit of the shadow. Everything else is delegated
1470 /// to handleShadowOr().
1471 void handleSignedRelationalComparison(ICmpInst &I) {
1472 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1473 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1475 CmpInst::Predicate pre = I.getPredicate();
1476 if (constOp0 && constOp0->isNullValue() &&
1477 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE)) {
1478 op = I.getOperand(1);
1479 } else if (constOp1 && constOp1->isNullValue() &&
1480 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) {
1481 op = I.getOperand(0);
1484 IRBuilder<> IRB(&I);
1486 IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op), "_msprop_icmpslt");
1487 setShadow(&I, Shadow);
1488 setOrigin(&I, getOrigin(op));
1494 void visitICmpInst(ICmpInst &I) {
1495 if (!ClHandleICmp) {
1499 if (I.isEquality()) {
1500 handleEqualityComparison(I);
1504 assert(I.isRelational());
1505 if (ClHandleICmpExact) {
1506 handleRelationalComparisonExact(I);
1510 handleSignedRelationalComparison(I);
1514 assert(I.isUnsigned());
1515 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1516 handleRelationalComparisonExact(I);
1523 void visitFCmpInst(FCmpInst &I) {
1527 void handleShift(BinaryOperator &I) {
1528 IRBuilder<> IRB(&I);
1529 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1530 // Otherwise perform the same shift on S1.
1531 Value *S1 = getShadow(&I, 0);
1532 Value *S2 = getShadow(&I, 1);
1533 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1535 Value *V2 = I.getOperand(1);
1536 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1537 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1538 setOriginForNaryOp(I);
1541 void visitShl(BinaryOperator &I) { handleShift(I); }
1542 void visitAShr(BinaryOperator &I) { handleShift(I); }
1543 void visitLShr(BinaryOperator &I) { handleShift(I); }
1545 /// \brief Instrument llvm.memmove
1547 /// At this point we don't know if llvm.memmove will be inlined or not.
1548 /// If we don't instrument it and it gets inlined,
1549 /// our interceptor will not kick in and we will lose the memmove.
1550 /// If we instrument the call here, but it does not get inlined,
1551 /// we will memove the shadow twice: which is bad in case
1552 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1554 /// Similar situation exists for memcpy and memset.
1555 void visitMemMoveInst(MemMoveInst &I) {
1556 IRBuilder<> IRB(&I);
1559 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1560 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1561 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1562 I.eraseFromParent();
1565 // Similar to memmove: avoid copying shadow twice.
1566 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1567 // FIXME: consider doing manual inline for small constant sizes and proper
1569 void visitMemCpyInst(MemCpyInst &I) {
1570 IRBuilder<> IRB(&I);
1573 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1574 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1575 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1576 I.eraseFromParent();
1580 void visitMemSetInst(MemSetInst &I) {
1581 IRBuilder<> IRB(&I);
1584 IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1585 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1586 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false));
1587 I.eraseFromParent();
1590 void visitVAStartInst(VAStartInst &I) {
1591 VAHelper->visitVAStartInst(I);
1594 void visitVACopyInst(VACopyInst &I) {
1595 VAHelper->visitVACopyInst(I);
1598 enum IntrinsicKind {
1599 IK_DoesNotAccessMemory,
1604 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1605 const int DoesNotAccessMemory = IK_DoesNotAccessMemory;
1606 const int OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1607 const int OnlyReadsMemory = IK_OnlyReadsMemory;
1608 const int OnlyAccessesArgumentPointees = IK_WritesMemory;
1609 const int UnknownModRefBehavior = IK_WritesMemory;
1610 #define GET_INTRINSIC_MODREF_BEHAVIOR
1611 #define ModRefBehavior IntrinsicKind
1612 #include "llvm/IR/Intrinsics.gen"
1613 #undef ModRefBehavior
1614 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1617 /// \brief Handle vector store-like intrinsics.
1619 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1620 /// has 1 pointer argument and 1 vector argument, returns void.
1621 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1622 IRBuilder<> IRB(&I);
1623 Value* Addr = I.getArgOperand(0);
1624 Value *Shadow = getShadow(&I, 1);
1625 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1627 // We don't know the pointer alignment (could be unaligned SSE store!).
1628 // Have to assume to worst case.
1629 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1631 if (ClCheckAccessAddress)
1632 insertShadowCheck(Addr, &I);
1634 // FIXME: use ClStoreCleanOrigin
1635 // FIXME: factor out common code from materializeStores
1636 if (MS.TrackOrigins)
1637 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB));
1641 /// \brief Handle vector load-like intrinsics.
1643 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1644 /// has 1 pointer argument, returns a vector.
1645 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1646 IRBuilder<> IRB(&I);
1647 Value *Addr = I.getArgOperand(0);
1649 Type *ShadowTy = getShadowTy(&I);
1651 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1652 // We don't know the pointer alignment (could be unaligned SSE load!).
1653 // Have to assume to worst case.
1654 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1656 setShadow(&I, getCleanShadow(&I));
1659 if (ClCheckAccessAddress)
1660 insertShadowCheck(Addr, &I);
1662 if (MS.TrackOrigins) {
1664 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB)));
1666 setOrigin(&I, getCleanOrigin());
1671 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1673 /// Instrument intrinsics with any number of arguments of the same type,
1674 /// equal to the return type. The type should be simple (no aggregates or
1675 /// pointers; vectors are fine).
1676 /// Caller guarantees that this intrinsic does not access memory.
1677 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1678 Type *RetTy = I.getType();
1679 if (!(RetTy->isIntOrIntVectorTy() ||
1680 RetTy->isFPOrFPVectorTy() ||
1681 RetTy->isX86_MMXTy()))
1684 unsigned NumArgOperands = I.getNumArgOperands();
1686 for (unsigned i = 0; i < NumArgOperands; ++i) {
1687 Type *Ty = I.getArgOperand(i)->getType();
1692 IRBuilder<> IRB(&I);
1693 ShadowAndOriginCombiner SC(this, IRB);
1694 for (unsigned i = 0; i < NumArgOperands; ++i)
1695 SC.Add(I.getArgOperand(i));
1701 /// \brief Heuristically instrument unknown intrinsics.
1703 /// The main purpose of this code is to do something reasonable with all
1704 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
1705 /// We recognize several classes of intrinsics by their argument types and
1706 /// ModRefBehaviour and apply special intrumentation when we are reasonably
1707 /// sure that we know what the intrinsic does.
1709 /// We special-case intrinsics where this approach fails. See llvm.bswap
1710 /// handling as an example of that.
1711 bool handleUnknownIntrinsic(IntrinsicInst &I) {
1712 unsigned NumArgOperands = I.getNumArgOperands();
1713 if (NumArgOperands == 0)
1716 Intrinsic::ID iid = I.getIntrinsicID();
1717 IntrinsicKind IK = getIntrinsicKind(iid);
1718 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
1719 bool WritesMemory = IK == IK_WritesMemory;
1720 assert(!(OnlyReadsMemory && WritesMemory));
1722 if (NumArgOperands == 2 &&
1723 I.getArgOperand(0)->getType()->isPointerTy() &&
1724 I.getArgOperand(1)->getType()->isVectorTy() &&
1725 I.getType()->isVoidTy() &&
1727 // This looks like a vector store.
1728 return handleVectorStoreIntrinsic(I);
1731 if (NumArgOperands == 1 &&
1732 I.getArgOperand(0)->getType()->isPointerTy() &&
1733 I.getType()->isVectorTy() &&
1735 // This looks like a vector load.
1736 return handleVectorLoadIntrinsic(I);
1739 if (!OnlyReadsMemory && !WritesMemory)
1740 if (maybeHandleSimpleNomemIntrinsic(I))
1743 // FIXME: detect and handle SSE maskstore/maskload
1747 void handleBswap(IntrinsicInst &I) {
1748 IRBuilder<> IRB(&I);
1749 Value *Op = I.getArgOperand(0);
1750 Type *OpType = Op->getType();
1751 Function *BswapFunc = Intrinsic::getDeclaration(
1752 F.getParent(), Intrinsic::bswap, ArrayRef<Type*>(&OpType, 1));
1753 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
1754 setOrigin(&I, getOrigin(Op));
1757 // \brief Instrument vector convert instrinsic.
1759 // This function instruments intrinsics like cvtsi2ss:
1760 // %Out = int_xxx_cvtyyy(%ConvertOp)
1762 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
1763 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
1764 // number \p Out elements, and (if has 2 arguments) copies the rest of the
1765 // elements from \p CopyOp.
1766 // In most cases conversion involves floating-point value which may trigger a
1767 // hardware exception when not fully initialized. For this reason we require
1768 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
1769 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
1770 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
1771 // return a fully initialized value.
1772 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
1773 IRBuilder<> IRB(&I);
1774 Value *CopyOp, *ConvertOp;
1776 switch (I.getNumArgOperands()) {
1778 CopyOp = I.getArgOperand(0);
1779 ConvertOp = I.getArgOperand(1);
1782 ConvertOp = I.getArgOperand(0);
1786 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
1789 // The first *NumUsedElements* elements of ConvertOp are converted to the
1790 // same number of output elements. The rest of the output is copied from
1791 // CopyOp, or (if not available) filled with zeroes.
1792 // Combine shadow for elements of ConvertOp that are used in this operation,
1793 // and insert a check.
1794 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
1795 // int->any conversion.
1796 Value *ConvertShadow = getShadow(ConvertOp);
1797 Value *AggShadow = 0;
1798 if (ConvertOp->getType()->isVectorTy()) {
1799 AggShadow = IRB.CreateExtractElement(
1800 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
1801 for (int i = 1; i < NumUsedElements; ++i) {
1802 Value *MoreShadow = IRB.CreateExtractElement(
1803 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
1804 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
1807 AggShadow = ConvertShadow;
1809 assert(AggShadow->getType()->isIntegerTy());
1810 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
1812 // Build result shadow by zero-filling parts of CopyOp shadow that come from
1815 assert(CopyOp->getType() == I.getType());
1816 assert(CopyOp->getType()->isVectorTy());
1817 Value *ResultShadow = getShadow(CopyOp);
1818 Type *EltTy = ResultShadow->getType()->getVectorElementType();
1819 for (int i = 0; i < NumUsedElements; ++i) {
1820 ResultShadow = IRB.CreateInsertElement(
1821 ResultShadow, ConstantInt::getNullValue(EltTy),
1822 ConstantInt::get(IRB.getInt32Ty(), i));
1824 setShadow(&I, ResultShadow);
1825 setOrigin(&I, getOrigin(CopyOp));
1827 setShadow(&I, getCleanShadow(&I));
1831 // Given a scalar or vector, extract lower 64 bits (or less), and return all
1832 // zeroes if it is zero, and all ones otherwise.
1833 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
1834 if (S->getType()->isVectorTy())
1835 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
1836 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
1837 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1838 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
1841 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
1842 Type *T = S->getType();
1843 assert(T->isVectorTy());
1844 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
1845 return IRB.CreateSExt(S2, T);
1848 // \brief Instrument vector shift instrinsic.
1850 // This function instruments intrinsics like int_x86_avx2_psll_w.
1851 // Intrinsic shifts %In by %ShiftSize bits.
1852 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
1853 // size, and the rest is ignored. Behavior is defined even if shift size is
1854 // greater than register (or field) width.
1855 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
1856 assert(I.getNumArgOperands() == 2);
1857 IRBuilder<> IRB(&I);
1858 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1859 // Otherwise perform the same shift on S1.
1860 Value *S1 = getShadow(&I, 0);
1861 Value *S2 = getShadow(&I, 1);
1862 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
1863 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
1864 Value *V1 = I.getOperand(0);
1865 Value *V2 = I.getOperand(1);
1866 Value *Shift = IRB.CreateCall2(I.getCalledValue(),
1867 IRB.CreateBitCast(S1, V1->getType()), V2);
1868 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
1869 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1870 setOriginForNaryOp(I);
1873 void visitIntrinsicInst(IntrinsicInst &I) {
1874 switch (I.getIntrinsicID()) {
1875 case llvm::Intrinsic::bswap:
1878 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
1879 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
1880 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
1881 case llvm::Intrinsic::x86_avx512_cvtss2usi:
1882 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
1883 case llvm::Intrinsic::x86_avx512_cvttss2usi:
1884 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
1885 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
1886 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
1887 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
1888 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
1889 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
1890 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
1891 case llvm::Intrinsic::x86_sse2_cvtsd2si:
1892 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
1893 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
1894 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
1895 case llvm::Intrinsic::x86_sse2_cvtss2sd:
1896 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
1897 case llvm::Intrinsic::x86_sse2_cvttsd2si:
1898 case llvm::Intrinsic::x86_sse_cvtsi2ss:
1899 case llvm::Intrinsic::x86_sse_cvtsi642ss:
1900 case llvm::Intrinsic::x86_sse_cvtss2si64:
1901 case llvm::Intrinsic::x86_sse_cvtss2si:
1902 case llvm::Intrinsic::x86_sse_cvttss2si64:
1903 case llvm::Intrinsic::x86_sse_cvttss2si:
1904 handleVectorConvertIntrinsic(I, 1);
1906 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
1907 case llvm::Intrinsic::x86_sse2_cvtps2pd:
1908 case llvm::Intrinsic::x86_sse_cvtps2pi:
1909 case llvm::Intrinsic::x86_sse_cvttps2pi:
1910 handleVectorConvertIntrinsic(I, 2);
1912 case llvm::Intrinsic::x86_avx512_psll_dq:
1913 case llvm::Intrinsic::x86_avx512_psrl_dq:
1914 case llvm::Intrinsic::x86_avx2_psll_w:
1915 case llvm::Intrinsic::x86_avx2_psll_d:
1916 case llvm::Intrinsic::x86_avx2_psll_q:
1917 case llvm::Intrinsic::x86_avx2_pslli_w:
1918 case llvm::Intrinsic::x86_avx2_pslli_d:
1919 case llvm::Intrinsic::x86_avx2_pslli_q:
1920 case llvm::Intrinsic::x86_avx2_psll_dq:
1921 case llvm::Intrinsic::x86_avx2_psrl_w:
1922 case llvm::Intrinsic::x86_avx2_psrl_d:
1923 case llvm::Intrinsic::x86_avx2_psrl_q:
1924 case llvm::Intrinsic::x86_avx2_psra_w:
1925 case llvm::Intrinsic::x86_avx2_psra_d:
1926 case llvm::Intrinsic::x86_avx2_psrli_w:
1927 case llvm::Intrinsic::x86_avx2_psrli_d:
1928 case llvm::Intrinsic::x86_avx2_psrli_q:
1929 case llvm::Intrinsic::x86_avx2_psrai_w:
1930 case llvm::Intrinsic::x86_avx2_psrai_d:
1931 case llvm::Intrinsic::x86_avx2_psrl_dq:
1932 case llvm::Intrinsic::x86_sse2_psll_w:
1933 case llvm::Intrinsic::x86_sse2_psll_d:
1934 case llvm::Intrinsic::x86_sse2_psll_q:
1935 case llvm::Intrinsic::x86_sse2_pslli_w:
1936 case llvm::Intrinsic::x86_sse2_pslli_d:
1937 case llvm::Intrinsic::x86_sse2_pslli_q:
1938 case llvm::Intrinsic::x86_sse2_psll_dq:
1939 case llvm::Intrinsic::x86_sse2_psrl_w:
1940 case llvm::Intrinsic::x86_sse2_psrl_d:
1941 case llvm::Intrinsic::x86_sse2_psrl_q:
1942 case llvm::Intrinsic::x86_sse2_psra_w:
1943 case llvm::Intrinsic::x86_sse2_psra_d:
1944 case llvm::Intrinsic::x86_sse2_psrli_w:
1945 case llvm::Intrinsic::x86_sse2_psrli_d:
1946 case llvm::Intrinsic::x86_sse2_psrli_q:
1947 case llvm::Intrinsic::x86_sse2_psrai_w:
1948 case llvm::Intrinsic::x86_sse2_psrai_d:
1949 case llvm::Intrinsic::x86_sse2_psrl_dq:
1950 case llvm::Intrinsic::x86_mmx_psll_w:
1951 case llvm::Intrinsic::x86_mmx_psll_d:
1952 case llvm::Intrinsic::x86_mmx_psll_q:
1953 case llvm::Intrinsic::x86_mmx_pslli_w:
1954 case llvm::Intrinsic::x86_mmx_pslli_d:
1955 case llvm::Intrinsic::x86_mmx_pslli_q:
1956 case llvm::Intrinsic::x86_mmx_psrl_w:
1957 case llvm::Intrinsic::x86_mmx_psrl_d:
1958 case llvm::Intrinsic::x86_mmx_psrl_q:
1959 case llvm::Intrinsic::x86_mmx_psra_w:
1960 case llvm::Intrinsic::x86_mmx_psra_d:
1961 case llvm::Intrinsic::x86_mmx_psrli_w:
1962 case llvm::Intrinsic::x86_mmx_psrli_d:
1963 case llvm::Intrinsic::x86_mmx_psrli_q:
1964 case llvm::Intrinsic::x86_mmx_psrai_w:
1965 case llvm::Intrinsic::x86_mmx_psrai_d:
1966 handleVectorShiftIntrinsic(I, /* Variable */ false);
1968 case llvm::Intrinsic::x86_avx2_psllv_d:
1969 case llvm::Intrinsic::x86_avx2_psllv_d_256:
1970 case llvm::Intrinsic::x86_avx2_psllv_q:
1971 case llvm::Intrinsic::x86_avx2_psllv_q_256:
1972 case llvm::Intrinsic::x86_avx2_psrlv_d:
1973 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
1974 case llvm::Intrinsic::x86_avx2_psrlv_q:
1975 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
1976 case llvm::Intrinsic::x86_avx2_psrav_d:
1977 case llvm::Intrinsic::x86_avx2_psrav_d_256:
1978 handleVectorShiftIntrinsic(I, /* Variable */ true);
1981 // Byte shifts are not implemented.
1982 // case llvm::Intrinsic::x86_avx512_psll_dq_bs:
1983 // case llvm::Intrinsic::x86_avx512_psrl_dq_bs:
1984 // case llvm::Intrinsic::x86_avx2_psll_dq_bs:
1985 // case llvm::Intrinsic::x86_avx2_psrl_dq_bs:
1986 // case llvm::Intrinsic::x86_sse2_psll_dq_bs:
1987 // case llvm::Intrinsic::x86_sse2_psrl_dq_bs:
1990 if (!handleUnknownIntrinsic(I))
1991 visitInstruction(I);
1996 void visitCallSite(CallSite CS) {
1997 Instruction &I = *CS.getInstruction();
1998 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2000 CallInst *Call = cast<CallInst>(&I);
2002 // For inline asm, do the usual thing: check argument shadow and mark all
2003 // outputs as clean. Note that any side effects of the inline asm that are
2004 // not immediately visible in its constraints are not handled.
2005 if (Call->isInlineAsm()) {
2006 visitInstruction(I);
2010 // Allow only tail calls with the same types, otherwise
2011 // we may have a false positive: shadow for a non-void RetVal
2012 // will get propagated to a void RetVal.
2013 if (Call->isTailCall() && Call->getType() != Call->getParent()->getType())
2014 Call->setTailCall(false);
2016 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2018 // We are going to insert code that relies on the fact that the callee
2019 // will become a non-readonly function after it is instrumented by us. To
2020 // prevent this code from being optimized out, mark that function
2021 // non-readonly in advance.
2022 if (Function *Func = Call->getCalledFunction()) {
2023 // Clear out readonly/readnone attributes.
2025 B.addAttribute(Attribute::ReadOnly)
2026 .addAttribute(Attribute::ReadNone);
2027 Func->removeAttributes(AttributeSet::FunctionIndex,
2028 AttributeSet::get(Func->getContext(),
2029 AttributeSet::FunctionIndex,
2033 IRBuilder<> IRB(&I);
2035 if (MS.WrapIndirectCalls && !CS.getCalledFunction())
2036 IndirectCallList.push_back(CS);
2038 unsigned ArgOffset = 0;
2039 DEBUG(dbgs() << " CallSite: " << I << "\n");
2040 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2041 ArgIt != End; ++ArgIt) {
2043 unsigned i = ArgIt - CS.arg_begin();
2044 if (!A->getType()->isSized()) {
2045 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2050 // Compute the Shadow for arg even if it is ByVal, because
2051 // in that case getShadow() will copy the actual arg shadow to
2052 // __msan_param_tls.
2053 Value *ArgShadow = getShadow(A);
2054 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2055 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2056 " Shadow: " << *ArgShadow << "\n");
2057 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2058 assert(A->getType()->isPointerTy() &&
2059 "ByVal argument is not a pointer!");
2060 Size = MS.DL->getTypeAllocSize(A->getType()->getPointerElementType());
2061 unsigned Alignment = CS.getParamAlignment(i + 1);
2062 Store = IRB.CreateMemCpy(ArgShadowBase,
2063 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2066 Size = MS.DL->getTypeAllocSize(A->getType());
2067 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2068 kShadowTLSAlignment);
2070 if (MS.TrackOrigins)
2071 IRB.CreateStore(getOrigin(A),
2072 getOriginPtrForArgument(A, IRB, ArgOffset));
2074 assert(Size != 0 && Store != 0);
2075 DEBUG(dbgs() << " Param:" << *Store << "\n");
2076 ArgOffset += DataLayout::RoundUpAlignment(Size, 8);
2078 DEBUG(dbgs() << " done with call args\n");
2081 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2082 if (FT->isVarArg()) {
2083 VAHelper->visitCallSite(CS, IRB);
2086 // Now, get the shadow for the RetVal.
2087 if (!I.getType()->isSized()) return;
2088 IRBuilder<> IRBBefore(&I);
2089 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2090 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2091 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2092 Instruction *NextInsn = 0;
2094 NextInsn = I.getNextNode();
2096 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2097 if (!NormalDest->getSinglePredecessor()) {
2098 // FIXME: this case is tricky, so we are just conservative here.
2099 // Perhaps we need to split the edge between this BB and NormalDest,
2100 // but a naive attempt to use SplitEdge leads to a crash.
2101 setShadow(&I, getCleanShadow(&I));
2102 setOrigin(&I, getCleanOrigin());
2105 NextInsn = NormalDest->getFirstInsertionPt();
2107 "Could not find insertion point for retval shadow load");
2109 IRBuilder<> IRBAfter(NextInsn);
2110 Value *RetvalShadow =
2111 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2112 kShadowTLSAlignment, "_msret");
2113 setShadow(&I, RetvalShadow);
2114 if (MS.TrackOrigins)
2115 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2118 void visitReturnInst(ReturnInst &I) {
2119 IRBuilder<> IRB(&I);
2120 Value *RetVal = I.getReturnValue();
2121 if (!RetVal) return;
2122 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2123 if (CheckReturnValue) {
2124 insertShadowCheck(RetVal, &I);
2125 Value *Shadow = getCleanShadow(RetVal);
2126 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2128 Value *Shadow = getShadow(RetVal);
2129 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2130 // FIXME: make it conditional if ClStoreCleanOrigin==0
2131 if (MS.TrackOrigins)
2132 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2136 void visitPHINode(PHINode &I) {
2137 IRBuilder<> IRB(&I);
2138 ShadowPHINodes.push_back(&I);
2139 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2141 if (MS.TrackOrigins)
2142 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2146 void visitAllocaInst(AllocaInst &I) {
2147 setShadow(&I, getCleanShadow(&I));
2148 IRBuilder<> IRB(I.getNextNode());
2149 uint64_t Size = MS.DL->getTypeAllocSize(I.getAllocatedType());
2150 if (PoisonStack && ClPoisonStackWithCall) {
2151 IRB.CreateCall2(MS.MsanPoisonStackFn,
2152 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2153 ConstantInt::get(MS.IntptrTy, Size));
2155 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2156 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2157 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2160 if (PoisonStack && MS.TrackOrigins) {
2161 setOrigin(&I, getCleanOrigin());
2162 SmallString<2048> StackDescriptionStorage;
2163 raw_svector_ostream StackDescription(StackDescriptionStorage);
2164 // We create a string with a description of the stack allocation and
2165 // pass it into __msan_set_alloca_origin.
2166 // It will be printed by the run-time if stack-originated UMR is found.
2167 // The first 4 bytes of the string are set to '----' and will be replaced
2168 // by __msan_va_arg_overflow_size_tls at the first call.
2169 StackDescription << "----" << I.getName() << "@" << F.getName();
2171 createPrivateNonConstGlobalForString(*F.getParent(),
2172 StackDescription.str());
2174 IRB.CreateCall4(MS.MsanSetAllocaOrigin4Fn,
2175 IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2176 ConstantInt::get(MS.IntptrTy, Size),
2177 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2178 IRB.CreatePointerCast(&F, MS.IntptrTy));
2182 void visitSelectInst(SelectInst& I) {
2183 IRBuilder<> IRB(&I);
2184 // a = select b, c, d
2185 Value *S = IRB.CreateSelect(I.getCondition(), getShadow(I.getTrueValue()),
2186 getShadow(I.getFalseValue()));
2187 if (I.getType()->isAggregateType()) {
2188 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2189 // an extra "select". This results in much more compact IR.
2190 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2191 S = IRB.CreateSelect(getShadow(I.getCondition()),
2192 getPoisonedShadow(getShadowTy(I.getType())), S,
2193 "_msprop_select_agg");
2195 // Sa = (sext Sb) | (select b, Sc, Sd)
2196 S = IRB.CreateOr(S, CreateShadowCast(IRB, getShadow(I.getCondition()),
2197 S->getType(), true),
2201 if (MS.TrackOrigins) {
2202 // Origins are always i32, so any vector conditions must be flattened.
2203 // FIXME: consider tracking vector origins for app vectors?
2204 Value *Cond = I.getCondition();
2205 Value *CondShadow = getShadow(Cond);
2206 if (Cond->getType()->isVectorTy()) {
2207 Type *FlatTy = getShadowTyNoVec(Cond->getType());
2208 Cond = IRB.CreateICmpNE(IRB.CreateBitCast(Cond, FlatTy),
2209 ConstantInt::getNullValue(FlatTy));
2210 CondShadow = IRB.CreateICmpNE(IRB.CreateBitCast(CondShadow, FlatTy),
2211 ConstantInt::getNullValue(FlatTy));
2213 // a = select b, c, d
2214 // Oa = Sb ? Ob : (b ? Oc : Od)
2215 setOrigin(&I, IRB.CreateSelect(
2216 CondShadow, getOrigin(I.getCondition()),
2217 IRB.CreateSelect(Cond, getOrigin(I.getTrueValue()),
2218 getOrigin(I.getFalseValue()))));
2222 void visitLandingPadInst(LandingPadInst &I) {
2224 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2225 setShadow(&I, getCleanShadow(&I));
2226 setOrigin(&I, getCleanOrigin());
2229 void visitGetElementPtrInst(GetElementPtrInst &I) {
2233 void visitExtractValueInst(ExtractValueInst &I) {
2234 IRBuilder<> IRB(&I);
2235 Value *Agg = I.getAggregateOperand();
2236 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2237 Value *AggShadow = getShadow(Agg);
2238 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2239 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2240 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2241 setShadow(&I, ResShadow);
2242 setOriginForNaryOp(I);
2245 void visitInsertValueInst(InsertValueInst &I) {
2246 IRBuilder<> IRB(&I);
2247 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2248 Value *AggShadow = getShadow(I.getAggregateOperand());
2249 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2250 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2251 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2252 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2253 DEBUG(dbgs() << " Res: " << *Res << "\n");
2255 setOriginForNaryOp(I);
2258 void dumpInst(Instruction &I) {
2259 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2260 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2262 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2264 errs() << "QQQ " << I << "\n";
2267 void visitResumeInst(ResumeInst &I) {
2268 DEBUG(dbgs() << "Resume: " << I << "\n");
2269 // Nothing to do here.
2272 void visitInstruction(Instruction &I) {
2273 // Everything else: stop propagating and check for poisoned shadow.
2274 if (ClDumpStrictInstructions)
2276 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2277 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2278 insertShadowCheck(I.getOperand(i), &I);
2279 setShadow(&I, getCleanShadow(&I));
2280 setOrigin(&I, getCleanOrigin());
2284 /// \brief AMD64-specific implementation of VarArgHelper.
2285 struct VarArgAMD64Helper : public VarArgHelper {
2286 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2287 // See a comment in visitCallSite for more details.
2288 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2289 static const unsigned AMD64FpEndOffset = 176;
2292 MemorySanitizer &MS;
2293 MemorySanitizerVisitor &MSV;
2294 Value *VAArgTLSCopy;
2295 Value *VAArgOverflowSize;
2297 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2299 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2300 MemorySanitizerVisitor &MSV)
2301 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(0), VAArgOverflowSize(0) { }
2303 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2305 ArgKind classifyArgument(Value* arg) {
2306 // A very rough approximation of X86_64 argument classification rules.
2307 Type *T = arg->getType();
2308 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2309 return AK_FloatingPoint;
2310 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2311 return AK_GeneralPurpose;
2312 if (T->isPointerTy())
2313 return AK_GeneralPurpose;
2317 // For VarArg functions, store the argument shadow in an ABI-specific format
2318 // that corresponds to va_list layout.
2319 // We do this because Clang lowers va_arg in the frontend, and this pass
2320 // only sees the low level code that deals with va_list internals.
2321 // A much easier alternative (provided that Clang emits va_arg instructions)
2322 // would have been to associate each live instance of va_list with a copy of
2323 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2325 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2326 unsigned GpOffset = 0;
2327 unsigned FpOffset = AMD64GpEndOffset;
2328 unsigned OverflowOffset = AMD64FpEndOffset;
2329 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2330 ArgIt != End; ++ArgIt) {
2332 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2333 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2335 // ByVal arguments always go to the overflow area.
2336 assert(A->getType()->isPointerTy());
2337 Type *RealTy = A->getType()->getPointerElementType();
2338 uint64_t ArgSize = MS.DL->getTypeAllocSize(RealTy);
2339 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2340 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2341 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2342 ArgSize, kShadowTLSAlignment);
2344 ArgKind AK = classifyArgument(A);
2345 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2347 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2351 case AK_GeneralPurpose:
2352 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2355 case AK_FloatingPoint:
2356 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2360 uint64_t ArgSize = MS.DL->getTypeAllocSize(A->getType());
2361 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2362 OverflowOffset += DataLayout::RoundUpAlignment(ArgSize, 8);
2364 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2367 Constant *OverflowSize =
2368 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2369 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2372 /// \brief Compute the shadow address for a given va_arg.
2373 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2375 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2376 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2377 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2381 void visitVAStartInst(VAStartInst &I) override {
2382 IRBuilder<> IRB(&I);
2383 VAStartInstrumentationList.push_back(&I);
2384 Value *VAListTag = I.getArgOperand(0);
2385 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2387 // Unpoison the whole __va_list_tag.
2388 // FIXME: magic ABI constants.
2389 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2390 /* size */24, /* alignment */8, false);
2393 void visitVACopyInst(VACopyInst &I) override {
2394 IRBuilder<> IRB(&I);
2395 Value *VAListTag = I.getArgOperand(0);
2396 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2398 // Unpoison the whole __va_list_tag.
2399 // FIXME: magic ABI constants.
2400 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2401 /* size */24, /* alignment */8, false);
2404 void finalizeInstrumentation() override {
2405 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2406 "finalizeInstrumentation called twice");
2407 if (!VAStartInstrumentationList.empty()) {
2408 // If there is a va_start in this function, make a backup copy of
2409 // va_arg_tls somewhere in the function entry block.
2410 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2411 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2413 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2415 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2416 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2419 // Instrument va_start.
2420 // Copy va_list shadow from the backup copy of the TLS contents.
2421 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2422 CallInst *OrigInst = VAStartInstrumentationList[i];
2423 IRBuilder<> IRB(OrigInst->getNextNode());
2424 Value *VAListTag = OrigInst->getArgOperand(0);
2426 Value *RegSaveAreaPtrPtr =
2428 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2429 ConstantInt::get(MS.IntptrTy, 16)),
2430 Type::getInt64PtrTy(*MS.C));
2431 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2432 Value *RegSaveAreaShadowPtr =
2433 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2434 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2435 AMD64FpEndOffset, 16);
2437 Value *OverflowArgAreaPtrPtr =
2439 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2440 ConstantInt::get(MS.IntptrTy, 8)),
2441 Type::getInt64PtrTy(*MS.C));
2442 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2443 Value *OverflowArgAreaShadowPtr =
2444 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2445 Value *SrcPtr = IRB.CreateConstGEP1_32(VAArgTLSCopy, AMD64FpEndOffset);
2446 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2451 /// \brief A no-op implementation of VarArgHelper.
2452 struct VarArgNoOpHelper : public VarArgHelper {
2453 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
2454 MemorySanitizerVisitor &MSV) {}
2456 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
2458 void visitVAStartInst(VAStartInst &I) override {}
2460 void visitVACopyInst(VACopyInst &I) override {}
2462 void finalizeInstrumentation() override {}
2465 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
2466 MemorySanitizerVisitor &Visitor) {
2467 // VarArg handling is only implemented on AMD64. False positives are possible
2468 // on other platforms.
2469 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
2470 if (TargetTriple.getArch() == llvm::Triple::x86_64)
2471 return new VarArgAMD64Helper(Func, Msan, Visitor);
2473 return new VarArgNoOpHelper(Func, Msan, Visitor);
2478 bool MemorySanitizer::runOnFunction(Function &F) {
2479 MemorySanitizerVisitor Visitor(F, *this);
2481 // Clear out readonly/readnone attributes.
2483 B.addAttribute(Attribute::ReadOnly)
2484 .addAttribute(Attribute::ReadNone);
2485 F.removeAttributes(AttributeSet::FunctionIndex,
2486 AttributeSet::get(F.getContext(),
2487 AttributeSet::FunctionIndex, B));
2489 return Visitor.runOnFunction();